[RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[MORITA Kazutaka]

Hi all,

This patch adds a block driver for Sheepdog distributed storage
system.  Please consider for inclusion.

Sheepdog is a distributed storage system for QEMU.  It provides highly
available block level storage volumes to VMs like Amazon EBS.

Sheepdog features are:
- No node in the cluster is special (no metadata node, no control
  node, etc)
- Linear scalability in performance and capacity
- No single point of failure
- Autonomous management (zero configuration)
- Useful volume management support such as snapshot and cloning
- Thin provisioning
- Autonomous load balancing

The more details are available at the project site [1] and my previous
post about sheepdog [2].

We have implemented the essential part of sheepdog features, and
believe the API between Sheepdog and QEMU is finalized.

Any comments or suggestions would be greatly appreciated.


Here are examples:

$ qemu-img create -f sheepdog vol1 256G     # create images

$ qemu --drive format=sheepdog,file=vol1    # start up a VM

$ qemu-img snapshot -c name sheepdog:vol1   # create a snapshot

$ qemu-img snapshot -l sheepdog:vol1        # list snapshots
ID        TAG                 VM SIZE                DATE       VM CLOCK
1                                   0 2010-05-06 02:29:29   00:00:00.000
2                                   0 2010-05-06 02:29:55   00:00:00.000

$ qemu --drive format=sheepdog,file=vol1:1  # start up from a snapshot

$ qemu-img create -b sheepdog:vol1:1 -f sheepdog vol2   # clone images


Thanks,

Kazutaka

[1] http://www.osrg.net/sheepdog/

[2] http://lists.nongnu.org/archive/html/qemu-devel/2009-10/msg01773.html


MORITA Kazutaka (2):
  close all the block drivers before the qemu process exits
  block: add sheepdog driver for distributed storage support

 Makefile         |    2 +-
 block.c          |   14 +-
 block.h          |    1 +
 block/sheepdog.c | 1828 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 monitor.c        |    1 +
 vl.c             |    1 +
 6 files changed, 1845 insertions(+), 2 deletions(-)
 create mode 100644 block/sheepdog.c

[RFC PATCH 1/2] close all the block drivers before the qemu process exits

[MORITA Kazutaka]

This patch calls the close handler of the block driver before the qemu
process exits.

This is necessary because the sheepdog block driver releases the lock
of VM images in the close handler.

Signed-off-by: MORITA Kazutaka <morita.kazutaka-Zyj7fXuS5i5L9jVzuh4AOg@public.gmane.org>
---
 block.c   |   11 +++++++++++
 block.h   |    1 +
 monitor.c |    1 +
 vl.c      |    1 +
 4 files changed, 14 insertions(+), 0 deletions(-)

diff --git a/block.c b/block.c
index 7326bfe..a606820 100644
--- a/block.c
+++ b/block.c
@@ -526,6 +526,17 @@ void bdrv_close(BlockDriverState *bs)
     }
 }
 
+void bdrv_close_all(void)
+{
+    BlockDriverState *bs, *n;
+
+    for (bs = bdrv_first, n = bs->next; bs; bs = n, n = bs ? bs->next : NULL) {
+        if (bs && bs->drv && bs->drv->bdrv_close) {
+            bs->drv->bdrv_close(bs);
+        }
+    }
+}
+
 void bdrv_delete(BlockDriverState *bs)
 {
     BlockDriverState **pbs;
diff --git a/block.h b/block.h
index fa51ddf..1a1293a 100644
--- a/block.h
+++ b/block.h
@@ -123,6 +123,7 @@ BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
 /* Ensure contents are flushed to disk.  */
 void bdrv_flush(BlockDriverState *bs);
 void bdrv_flush_all(void);
+void bdrv_close_all(void);
 
 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
 	int *pnum);
diff --git a/monitor.c b/monitor.c
index 17e59f5..44bfe83 100644
--- a/monitor.c
+++ b/monitor.c
@@ -845,6 +845,7 @@ static void do_info_cpu_stats(Monitor *mon)
  */
 static void do_quit(Monitor *mon, const QDict *qdict, QObject **ret_data)
 {
+    bdrv_close_all();
     exit(0);
 }
 
diff --git a/vl.c b/vl.c
index 77677e8..65160ed 100644
--- a/vl.c
+++ b/vl.c
@@ -4205,6 +4205,7 @@ static void main_loop(void)
             vm_stop(r);
         }
     }
+    bdrv_close_all();
     pause_all_vcpus();
 }
 
-- 
1.5.6.5

Re: [Qemu-devel] [RFC PATCH 1/2] close all the block drivers before the qemu process exits

[Christoph Hellwig]

On Wed, May 12, 2010 at 07:46:52PM +0900, MORITA Kazutaka wrote:
> This patch calls the close handler of the block driver before the qemu
> process exits.
> 
> This is necessary because the sheepdog block driver releases the lock
> of VM images in the close handler.
> 
> Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>

Looks good in principle, except that bdrv_first is gone and has been
replaced with a real list in the meantime, so this won't even apply.

Re: [Qemu-devel] [RFC PATCH 1/2] close all the block drivers before the qemu process exits

[MORITA Kazutaka]

On 2010/05/12 23:01, Christoph Hellwig wrote:
> On Wed, May 12, 2010 at 07:46:52PM +0900, MORITA Kazutaka wrote:
>> This patch calls the close handler of the block driver before the qemu
>> process exits.
>>
>> This is necessary because the sheepdog block driver releases the lock
>> of VM images in the close handler.
>>
>> Signed-off-by: MORITA Kazutaka <morita.kazutaka-Zyj7fXuS5i5L9jVzuh4AOg@public.gmane.org>
> 
> Looks good in principle, except that bdrv_first is gone and has been
> replaced with a real list in the meantime, so this won't even apply.
> 

Thank you for your comment.
I'll rebase and resend the updated version in the next days.

Thanks,

Kazutaka

Re: [RFC PATCH 1/2] close all the block drivers before the qemu process exits

[Avi Kivity]

On 05/12/2010 01:46 PM, MORITA Kazutaka wrote:
> This patch calls the close handler of the block driver before the qemu
> process exits.
>
> This is necessary because the sheepdog block driver releases the lock
> of VM images in the close handler.
>
>    

How do you handle abnormal termination?

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.

Re: [RFC PATCH 1/2] close all the block drivers before the qemu process exits

[MORITA Kazutaka]

On 2010/05/12 23:28, Avi Kivity wrote:
> On 05/12/2010 01:46 PM, MORITA Kazutaka wrote:
>> This patch calls the close handler of the block driver before the qemu
>> process exits.
>>
>> This is necessary because the sheepdog block driver releases the lock
>> of VM images in the close handler.
>>
>>    
> 
> How do you handle abnormal termination?
> 

In the case, we need to release the lock manually, unfortunately.
Sheepdog admin tool has a command to do that.

Thanks,

Kazutaka

Re: [RFC PATCH 1/2] close all the block drivers before the qemu process exits

[MORITA Kazutaka]

At Thu, 13 May 2010 05:16:35 +0900,
MORITA Kazutaka wrote:
> 
> On 2010/05/12 23:28, Avi Kivity wrote:
> > On 05/12/2010 01:46 PM, MORITA Kazutaka wrote:
> >> This patch calls the close handler of the block driver before the qemu
> >> process exits.
> >>
> >> This is necessary because the sheepdog block driver releases the lock
> >> of VM images in the close handler.
> >>
> >>    
> > 
> > How do you handle abnormal termination?
> > 
> 
> In the case, we need to release the lock manually, unfortunately.
> Sheepdog admin tool has a command to do that.
> 

More precisely, if qemu fails down with its host machine, we detect
the qemu failure and release the lock.  It is because Sheepdog
currently assumes that all the qemu processes are in the sheepdog
cluster, and remember that where they are running.  When machine
failures happen, sheepdog release the lock of the VMs on the machines
(We uses corosync to check which machines are alive or not).

If the qemu process exits abnormally and its host machine is alive,
the sheepdog daemon on the host needs to detect the qemu failure.
However, the feature is not implemented yet.  We think of checking a
socket connection between qemu and sheepdog daemon to detect the
failure.  Currently, we need to release the lock manually from the
admin tool in this case.

Thanks,

Kazutaka

[RFC PATCH v2 0/3] Sheepdog: distributed storage system for QEMU

[MORITA Kazutaka]

Hi all,

This patch adds a block driver for Sheepdog distributed storage
system.

Changes from v1 to v2 are:

 - rebase onto git://repo.or.cz/qemu/kevin.git block
 - modify the sheepdog driver as a protocol driver
 - add new patch to call the snapshot handler of the protocol

One issue still remains; qemu-img parses command line options with
`create_options' of the format handler, so we cannot use protocol
specific options.

In this version, sheepdog needs to be used as a format driver when we
want to use sheepdog specific options.

e.g. Create clone image vol2 from vol1

  $ qemu-img create -b sheepdog:vol1:1 -f sheepdog vol2


Thanks,

Kazutaka


MORITA Kazutaka (3):
  close all the block drivers before the qemu process exits
  block: call the snapshot handlers of the protocol drivers
  block: add sheepdog driver for distributed storage support

 Makefile.objs    |    2 +-
 block.c          |   57 ++-
 block.h          |    1 +
 block/sheepdog.c | 1831 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 vl.c             |    1 +
 5 files changed, 1873 insertions(+), 19 deletions(-)
 create mode 100644 block/sheepdog.c

[RFC PATCH v3 0/3] Sheepdog: distributed storage system for QEMU

[MORITA Kazutaka]

Hi all,

This patch adds a block driver for Sheepdog distributed storage
system.

Changes from v2 to v3 are:

 - add drv->bdrv_close() and drv->bdrv_open() before and after
   bdrv_snapshot_goto() call of the protocol.
 - address the review comments on the sheepdog driver code.
   I'll send the details in the reply of the review mail.

Changes from v1 to v2 are:

 - rebase onto git://repo.or.cz/qemu/kevin.git block
 - modify the sheepdog driver as a protocol driver
 - add new patch to call the snapshot handler of the protocol

If this patchset is okay, I'll work on the image creation problem of the
protocol driver.

Thanks,

Kazutaka


MORITA Kazutaka (3):
  close all the block drivers before the qemu process exits
  block: call the snapshot handlers of the protocol drivers
  block: add sheepdog driver for distributed storage support

 Makefile.objs    |    2 +-
 block.c          |   70 ++-
 block.h          |    1 +
 block/sheepdog.c | 1845 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 vl.c             |    1 +
 5 files changed, 1900 insertions(+), 19 deletions(-)
 create mode 100644 block/sheepdog.c

[RFC PATCH v3 1/3] close all the block drivers before the qemu process exits

[MORITA Kazutaka]

This patch calls the close handler of the block driver before the qemu
process exits.

This is necessary because the sheepdog block driver releases the lock
of VM images in the close handler.

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 block.c |    9 +++++++++
 block.h |    1 +
 vl.c    |    1 +
 3 files changed, 11 insertions(+), 0 deletions(-)

diff --git a/block.c b/block.c
index bfe46e3..f3bf3f2 100644
--- a/block.c
+++ b/block.c
@@ -636,6 +636,15 @@ void bdrv_close(BlockDriverState *bs)
     }
 }
 
+void bdrv_close_all(void)
+{
+    BlockDriverState *bs;
+
+    QTAILQ_FOREACH(bs, &bdrv_states, list) {
+        bdrv_close(bs);
+    }
+}
+
 void bdrv_delete(BlockDriverState *bs)
 {
     /* remove from list, if necessary */
diff --git a/block.h b/block.h
index 278259c..531e802 100644
--- a/block.h
+++ b/block.h
@@ -121,6 +121,7 @@ BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
 /* Ensure contents are flushed to disk.  */
 void bdrv_flush(BlockDriverState *bs);
 void bdrv_flush_all(void);
+void bdrv_close_all(void);
 
 int bdrv_has_zero_init(BlockDriverState *bs);
 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
diff --git a/vl.c b/vl.c
index 85bcc84..5ce7807 100644
--- a/vl.c
+++ b/vl.c
@@ -2007,6 +2007,7 @@ static void main_loop(void)
             exit(0);
         }
     }
+    bdrv_close_all();
     pause_all_vcpus();
 }
 
-- 
1.5.6.5

[RFC PATCH v3 2/3] block: call the snapshot handlers of the protocol drivers

[MORITA Kazutaka]

When snapshot handlers are not defined in the format driver, it is
better to call the ones of the protocol driver.  This enables us to
implement snapshot support in the protocol driver.

We need to call bdrv_close() and bdrv_open() handlers of the format
driver before and after bdrv_snapshot_goto() call of the protocol.  It is
because the contents of the block driver state may need to be changed
after loading vmstate.

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 block.c |   61 +++++++++++++++++++++++++++++++++++++++++++------------------
 1 files changed, 43 insertions(+), 18 deletions(-)

diff --git a/block.c b/block.c
index f3bf3f2..c987e57 100644
--- a/block.c
+++ b/block.c
@@ -1683,9 +1683,11 @@ int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_save_vmstate)
-        return -ENOTSUP;
-    return drv->bdrv_save_vmstate(bs, buf, pos, size);
+    if (drv->bdrv_save_vmstate)
+        return drv->bdrv_save_vmstate(bs, buf, pos, size);
+    if (bs->file)
+        return bdrv_save_vmstate(bs->file, buf, pos, size);
+    return -ENOTSUP;
 }
 
 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
@@ -1694,9 +1696,11 @@ int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_load_vmstate)
-        return -ENOTSUP;
-    return drv->bdrv_load_vmstate(bs, buf, pos, size);
+    if (drv->bdrv_load_vmstate)
+        return drv->bdrv_load_vmstate(bs, buf, pos, size);
+    if (bs->file)
+        return bdrv_load_vmstate(bs->file, buf, pos, size);
+    return -ENOTSUP;
 }
 
 void bdrv_debug_event(BlockDriverState *bs, BlkDebugEvent event)
@@ -1720,20 +1724,37 @@ int bdrv_snapshot_create(BlockDriverState *bs,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_create)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_create(bs, sn_info);
+    if (drv->bdrv_snapshot_create)
+        return drv->bdrv_snapshot_create(bs, sn_info);
+    if (bs->file)
+        return bdrv_snapshot_create(bs->file, sn_info);
+    return -ENOTSUP;
 }
 
 int bdrv_snapshot_goto(BlockDriverState *bs,
                        const char *snapshot_id)
 {
     BlockDriver *drv = bs->drv;
+    int ret, open_ret;
+
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_goto)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_goto(bs, snapshot_id);
+    if (drv->bdrv_snapshot_goto)
+        return drv->bdrv_snapshot_goto(bs, snapshot_id);
+
+    if (bs->file) {
+        drv->bdrv_close(bs);
+        ret = bdrv_snapshot_goto(bs->file, snapshot_id);
+        open_ret = drv->bdrv_open(bs, bs->open_flags);
+        if (open_ret < 0) {
+            bdrv_delete(bs);
+            bs->drv = NULL;
+            return open_ret;
+        }
+        return ret;
+    }
+
+    return -ENOTSUP;
 }
 
 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
@@ -1741,9 +1762,11 @@ int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_delete)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_delete(bs, snapshot_id);
+    if (drv->bdrv_snapshot_delete)
+        return drv->bdrv_snapshot_delete(bs, snapshot_id);
+    if (bs->file)
+        return bdrv_snapshot_delete(bs->file, snapshot_id);
+    return -ENOTSUP;
 }
 
 int bdrv_snapshot_list(BlockDriverState *bs,
@@ -1752,9 +1775,11 @@ int bdrv_snapshot_list(BlockDriverState *bs,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_list)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_list(bs, psn_info);
+    if (drv->bdrv_snapshot_list)
+        return drv->bdrv_snapshot_list(bs, psn_info);
+    if (bs->file)
+        return bdrv_snapshot_list(bs->file, psn_info);
+    return -ENOTSUP;
 }
 
 #define NB_SUFFIXES 4
-- 
1.5.6.5

Re: [Qemu-devel] [RFC PATCH v3 2/3] block: call the snapshot handlers of the protocol drivers

[Kevin Wolf]

Am 17.05.2010 12:19, schrieb MORITA Kazutaka:
> When snapshot handlers are not defined in the format driver, it is
> better to call the ones of the protocol driver.  This enables us to
> implement snapshot support in the protocol driver.
> 
> We need to call bdrv_close() and bdrv_open() handlers of the format
> driver before and after bdrv_snapshot_goto() call of the protocol.  It is
> because the contents of the block driver state may need to be changed
> after loading vmstate.
> 
> Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
> ---
>  block.c |   61 +++++++++++++++++++++++++++++++++++++++++++------------------
>  1 files changed, 43 insertions(+), 18 deletions(-)
> 
> diff --git a/block.c b/block.c
> index f3bf3f2..c987e57 100644
> --- a/block.c
> +++ b/block.c
> @@ -1683,9 +1683,11 @@ int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
>      BlockDriver *drv = bs->drv;
>      if (!drv)
>          return -ENOMEDIUM;
> -    if (!drv->bdrv_save_vmstate)
> -        return -ENOTSUP;
> -    return drv->bdrv_save_vmstate(bs, buf, pos, size);
> +    if (drv->bdrv_save_vmstate)
> +        return drv->bdrv_save_vmstate(bs, buf, pos, size);
> +    if (bs->file)
> +        return bdrv_save_vmstate(bs->file, buf, pos, size);
> +    return -ENOTSUP;
>  }
>  
>  int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
> @@ -1694,9 +1696,11 @@ int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
>      BlockDriver *drv = bs->drv;
>      if (!drv)
>          return -ENOMEDIUM;
> -    if (!drv->bdrv_load_vmstate)
> -        return -ENOTSUP;
> -    return drv->bdrv_load_vmstate(bs, buf, pos, size);
> +    if (drv->bdrv_load_vmstate)
> +        return drv->bdrv_load_vmstate(bs, buf, pos, size);
> +    if (bs->file)
> +        return bdrv_load_vmstate(bs->file, buf, pos, size);
> +    return -ENOTSUP;
>  }
>  
>  void bdrv_debug_event(BlockDriverState *bs, BlkDebugEvent event)
> @@ -1720,20 +1724,37 @@ int bdrv_snapshot_create(BlockDriverState *bs,
>      BlockDriver *drv = bs->drv;
>      if (!drv)
>          return -ENOMEDIUM;
> -    if (!drv->bdrv_snapshot_create)
> -        return -ENOTSUP;
> -    return drv->bdrv_snapshot_create(bs, sn_info);
> +    if (drv->bdrv_snapshot_create)
> +        return drv->bdrv_snapshot_create(bs, sn_info);
> +    if (bs->file)
> +        return bdrv_snapshot_create(bs->file, sn_info);
> +    return -ENOTSUP;
>  }
>  
>  int bdrv_snapshot_goto(BlockDriverState *bs,
>                         const char *snapshot_id)
>  {
>      BlockDriver *drv = bs->drv;
> +    int ret, open_ret;
> +
>      if (!drv)
>          return -ENOMEDIUM;
> -    if (!drv->bdrv_snapshot_goto)
> -        return -ENOTSUP;
> -    return drv->bdrv_snapshot_goto(bs, snapshot_id);
> +    if (drv->bdrv_snapshot_goto)
> +        return drv->bdrv_snapshot_goto(bs, snapshot_id);
> +
> +    if (bs->file) {
> +        drv->bdrv_close(bs);
> +        ret = bdrv_snapshot_goto(bs->file, snapshot_id);
> +        open_ret = drv->bdrv_open(bs, bs->open_flags);
> +        if (open_ret < 0) {
> +            bdrv_delete(bs);

I think you mean bs->file here.

Kevin

> +            bs->drv = NULL;
> +            return open_ret;
> +        }
> +        return ret;
> +    }
> +
> +    return -ENOTSUP;
>  }
>  
>  int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)

Re: [RFC PATCH v3 2/3] block: call the snapshot handlers of the protocol drivers

[MORITA Kazutaka]

At Mon, 17 May 2010 13:08:08 +0200,
Kevin Wolf wrote:
> 
> Am 17.05.2010 12:19, schrieb MORITA Kazutaka:
> >  
> >  int bdrv_snapshot_goto(BlockDriverState *bs,
> >                         const char *snapshot_id)
> >  {
> >      BlockDriver *drv = bs->drv;
> > +    int ret, open_ret;
> > +
> >      if (!drv)
> >          return -ENOMEDIUM;
> > -    if (!drv->bdrv_snapshot_goto)
> > -        return -ENOTSUP;
> > -    return drv->bdrv_snapshot_goto(bs, snapshot_id);
> > +    if (drv->bdrv_snapshot_goto)
> > +        return drv->bdrv_snapshot_goto(bs, snapshot_id);
> > +
> > +    if (bs->file) {
> > +        drv->bdrv_close(bs);
> > +        ret = bdrv_snapshot_goto(bs->file, snapshot_id);
> > +        open_ret = drv->bdrv_open(bs, bs->open_flags);
> > +        if (open_ret < 0) {
> > +            bdrv_delete(bs);
> 
> I think you mean bs->file here.
> 
> Kevin

This is an error of re-opening the format driver, so what we should
delete here is not bs->file but bs, isn't it?  If we failed to open bs
here, the drive doesn't seem to work anymore.

Regards,

Kazutaka

> > +            bs->drv = NULL;
> > +            return open_ret;
> > +        }
> > +        return ret;
> > +    }
> > +
> > +    return -ENOTSUP;
> >  }

Re: [Qemu-devel] Re: [RFC PATCH v3 2/3] block: call the snapshot handlers of the protocol drivers

[Kevin Wolf]

Am 17.05.2010 14:19, schrieb MORITA Kazutaka:
> At Mon, 17 May 2010 13:08:08 +0200,
> Kevin Wolf wrote:
>>
>> Am 17.05.2010 12:19, schrieb MORITA Kazutaka:
>>>  
>>>  int bdrv_snapshot_goto(BlockDriverState *bs,
>>>                         const char *snapshot_id)
>>>  {
>>>      BlockDriver *drv = bs->drv;
>>> +    int ret, open_ret;
>>> +
>>>      if (!drv)
>>>          return -ENOMEDIUM;
>>> -    if (!drv->bdrv_snapshot_goto)
>>> -        return -ENOTSUP;
>>> -    return drv->bdrv_snapshot_goto(bs, snapshot_id);
>>> +    if (drv->bdrv_snapshot_goto)
>>> +        return drv->bdrv_snapshot_goto(bs, snapshot_id);
>>> +
>>> +    if (bs->file) {
>>> +        drv->bdrv_close(bs);
>>> +        ret = bdrv_snapshot_goto(bs->file, snapshot_id);
>>> +        open_ret = drv->bdrv_open(bs, bs->open_flags);
>>> +        if (open_ret < 0) {
>>> +            bdrv_delete(bs);
>>
>> I think you mean bs->file here.
>>
>> Kevin
> 
> This is an error of re-opening the format driver, so what we should
> delete here is not bs->file but bs, isn't it?  If we failed to open bs
> here, the drive doesn't seem to work anymore.

But bdrv_delete means basically free it. This is almost guaranteed to
lead to crashes because that BlockDriverState is still in use in other
places.

One additional case of use after free is in the very next line:

>>> +            bs->drv = NULL;

You can't do that when bs is freed, obviously. But I think just setting
bs->drv to NULL without bdrv_deleting it before is the better way. It
will fail any requests (with -ENOMEDIUM), but can't produce crashes.
This is also what bdrv_commit does in such situations.

In this state, we don't access the underlying file any more, so we could
delete bs->file - this is why I thought you actually meant to do that.

Kevin

Re: [Qemu-devel] Re: [RFC PATCH v3 2/3] block: call the snapshot handlers of the protocol drivers

[MORITA Kazutaka]

On Mon, May 17, 2010 at 9:20 PM, Kevin Wolf <kwolf@redhat.com> wrote:
> Am 17.05.2010 14:19, schrieb MORITA Kazutaka:
>> At Mon, 17 May 2010 13:08:08 +0200,
>> Kevin Wolf wrote:
>>>
>>> Am 17.05.2010 12:19, schrieb MORITA Kazutaka:
>>>>
>>>>  int bdrv_snapshot_goto(BlockDriverState *bs,
>>>>                         const char *snapshot_id)
>>>>  {
>>>>      BlockDriver *drv = bs->drv;
>>>> +    int ret, open_ret;
>>>> +
>>>>      if (!drv)
>>>>          return -ENOMEDIUM;
>>>> -    if (!drv->bdrv_snapshot_goto)
>>>> -        return -ENOTSUP;
>>>> -    return drv->bdrv_snapshot_goto(bs, snapshot_id);
>>>> +    if (drv->bdrv_snapshot_goto)
>>>> +        return drv->bdrv_snapshot_goto(bs, snapshot_id);
>>>> +
>>>> +    if (bs->file) {
>>>> +        drv->bdrv_close(bs);
>>>> +        ret = bdrv_snapshot_goto(bs->file, snapshot_id);
>>>> +        open_ret = drv->bdrv_open(bs, bs->open_flags);
>>>> +        if (open_ret < 0) {
>>>> +            bdrv_delete(bs);
>>>
>>> I think you mean bs->file here.
>>>
>>> Kevin
>>
>> This is an error of re-opening the format driver, so what we should
>> delete here is not bs->file but bs, isn't it?  If we failed to open bs
>> here, the drive doesn't seem to work anymore.
>
> But bdrv_delete means basically free it. This is almost guaranteed to
> lead to crashes because that BlockDriverState is still in use in other
> places.
>
> One additional case of use after free is in the very next line:
>
>>>> +            bs->drv = NULL;
>
> You can't do that when bs is freed, obviously. But I think just setting
> bs->drv to NULL without bdrv_deleting it before is the better way. It
> will fail any requests (with -ENOMEDIUM), but can't produce crashes.
> This is also what bdrv_commit does in such situations.
>
> In this state, we don't access the underlying file any more, so we could
> delete bs->file - this is why I thought you actually meant to do that.
>

I'm sorry for the confusion.  I understand what we should do here.
I'll fix it for the next post.

Thanks,

Kazutaka

[RFC PATCH v3 3/3] block: add sheepdog driver for distributed storage support

[MORITA Kazutaka]

Sheepdog is a distributed storage system for QEMU. It provides highly
available block level storage volumes to VMs like Amazon EBS.  This
patch adds a qemu block driver for Sheepdog.

Sheepdog features are:
- No node in the cluster is special (no metadata node, no control
  node, etc)
- Linear scalability in performance and capacity
- No single point of failure
- Autonomous management (zero configuration)
- Useful volume management support such as snapshot and cloning
- Thin provisioning
- Autonomous load balancing

The more details are available at the project site:
    http://www.osrg.net/sheepdog/

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 Makefile.objs    |    2 +-
 block/sheepdog.c | 1845 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 1846 insertions(+), 1 deletions(-)
 create mode 100644 block/sheepdog.c

diff --git a/Makefile.objs b/Makefile.objs
index ecdd53e..6edbc57 100644
--- a/Makefile.objs
+++ b/Makefile.objs
@@ -14,7 +14,7 @@ block-obj-$(CONFIG_LINUX_AIO) += linux-aio.o
 
 block-nested-y += raw.o cow.o qcow.o vdi.o vmdk.o cloop.o dmg.o bochs.o vpc.o vvfat.o
 block-nested-y += qcow2.o qcow2-refcount.o qcow2-cluster.o qcow2-snapshot.o
-block-nested-y += parallels.o nbd.o blkdebug.o
+block-nested-y += parallels.o nbd.o blkdebug.o sheepdog.o
 block-nested-$(CONFIG_WIN32) += raw-win32.o
 block-nested-$(CONFIG_POSIX) += raw-posix.o
 block-nested-$(CONFIG_CURL) += curl.o
diff --git a/block/sheepdog.c b/block/sheepdog.c
new file mode 100644
index 0000000..4672f00
--- /dev/null
+++ b/block/sheepdog.c
@@ -0,0 +1,1845 @@
+/*
+ * Copyright (C) 2009-2010 Nippon Telegraph and Telephone Corporation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+#include <netdb.h>
+#include <netinet/tcp.h>
+
+#include "qemu-common.h"
+#include "qemu-error.h"
+#include "block_int.h"
+
+#define SD_PROTO_VER 0x01
+
+#define SD_DEFAULT_ADDR "localhost:7000"
+
+#define SD_OP_CREATE_AND_WRITE_OBJ  0x01
+#define SD_OP_READ_OBJ       0x02
+#define SD_OP_WRITE_OBJ      0x03
+
+#define SD_OP_NEW_VDI        0x11
+#define SD_OP_LOCK_VDI       0x12
+#define SD_OP_RELEASE_VDI    0x13
+#define SD_OP_GET_VDI_INFO   0x14
+#define SD_OP_READ_VDIS      0x15
+
+#define SD_FLAG_CMD_WRITE    0x01
+#define SD_FLAG_CMD_COW      0x02
+
+#define SD_RES_SUCCESS       0x00 /* Success */
+#define SD_RES_UNKNOWN       0x01 /* Unknown error */
+#define SD_RES_NO_OBJ        0x02 /* No object found */
+#define SD_RES_EIO           0x03 /* I/O error */
+#define SD_RES_VDI_EXIST     0x04 /* Vdi exists already */
+#define SD_RES_INVALID_PARMS 0x05 /* Invalid parameters */
+#define SD_RES_SYSTEM_ERROR  0x06 /* System error */
+#define SD_RES_VDI_LOCKED    0x07 /* Vdi is locked */
+#define SD_RES_NO_VDI        0x08 /* No vdi found */
+#define SD_RES_NO_BASE_VDI   0x09 /* No base vdi found */
+#define SD_RES_VDI_READ      0x0A /* Cannot read requested vdi */
+#define SD_RES_VDI_WRITE     0x0B /* Cannot write requested vdi */
+#define SD_RES_BASE_VDI_READ 0x0C /* Cannot read base vdi */
+#define SD_RES_BASE_VDI_WRITE   0x0D /* Cannot write base vdi */
+#define SD_RES_NO_TAG        0x0E /* Requested tag is not found */
+#define SD_RES_STARTUP       0x0F /* Sheepdog is on starting up */
+#define SD_RES_VDI_NOT_LOCKED   0x10 /* Vdi is not locked */
+#define SD_RES_SHUTDOWN      0x11 /* Sheepdog is shutting down */
+#define SD_RES_NO_MEM        0x12 /* Cannot allocate memory */
+#define SD_RES_FULL_VDI      0x13 /* we already have the maximum vdis */
+#define SD_RES_VER_MISMATCH  0x14 /* Protocol version mismatch */
+#define SD_RES_NO_SPACE      0x15 /* Server has no room for new objects */
+#define SD_RES_WAIT_FOR_FORMAT  0x16 /* Sheepdog is waiting for a format operation */
+#define SD_RES_WAIT_FOR_JOIN    0x17 /* Sheepdog is waiting for other nodes joining */
+#define SD_RES_JOIN_FAILED   0x18 /* Target node had failed to join sheepdog */
+
+/*
+ * Object ID rules
+ *
+ *  0 - 19 (20 bits): data object space
+ * 20 - 31 (12 bits): reserved data object space
+ * 32 - 55 (24 bits): vdi object space
+ * 56 - 59 ( 4 bits): reserved vdi object space
+ * 60 - 63 ( 4 bits): object type indentifier space
+ */
+
+#define VDI_SPACE_SHIFT   32
+#define VDI_BIT (UINT64_C(1) << 63)
+#define VMSTATE_BIT (UINT64_C(1) << 62)
+#define MAX_DATA_OBJS (1ULL << 20)
+#define MAX_CHILDREN 1024
+#define SD_MAX_VDI_LEN 256
+#define SD_NR_VDIS   (1U << 24)
+#define SD_DATA_OBJ_SIZE (UINT64_C(1) << 22)
+
+#define SD_INODE_SIZE (sizeof(SheepdogInode))
+#define CURRENT_VDI_ID 0
+
+typedef struct SheepdogReq {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t	opcode_specific[8];
+} SheepdogReq;
+
+typedef struct SheepdogRsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t	opcode_specific[7];
+} SheepdogRsp;
+
+typedef struct SheepdogObjReq {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint64_t        oid;
+	uint64_t        cow_oid;
+	uint32_t        copies;
+	uint32_t        rsvd;
+	uint64_t        offset;
+} SheepdogObjReq;
+
+typedef struct SheepdogObjRsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t        copies;
+	uint32_t        pad[6];
+} SheepdogObjRsp;
+
+typedef struct SheepdogVdiReq {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint64_t	vdi_size;
+	uint32_t        base_vdi_id;
+	uint32_t        copies;
+	uint32_t        snapid;
+	uint32_t        pad[3];
+} SheepdogVdiReq;
+
+typedef struct SheepdogVdiRsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t        rsvd;
+	uint32_t        vdi_id;
+	uint32_t        pad[5];
+} SheepdogVdiRsp;
+
+typedef struct SheepdogInode {
+	char name[SD_MAX_VDI_LEN];
+	uint64_t ctime;
+	uint64_t snap_ctime;
+	uint64_t vm_clock_nsec;
+	uint64_t vdi_size;
+	uint64_t vm_state_size;
+	uint16_t copy_policy;
+	uint8_t  nr_copies;
+	uint8_t  block_size_shift;
+	uint32_t snap_id;
+	uint32_t vdi_id;
+	uint32_t parent_vdi_id;
+	uint32_t child_vdi_id[MAX_CHILDREN];
+	uint32_t data_vdi_id[MAX_DATA_OBJS];
+} SheepdogInode;
+
+/*
+ * 64 bit FNV-1a non-zero initial basis
+ */
+#define FNV1A_64_INIT ((uint64_t)0xcbf29ce484222325ULL)
+
+/*
+ * 64 bit Fowler/Noll/Vo FNV-1a hash code
+ */
+static inline uint64_t fnv_64a_buf(void *buf, size_t len, uint64_t hval)
+{
+        unsigned char *bp = (unsigned char *) buf;
+        unsigned char *be = bp + len;
+        while (bp < be) {
+                hval ^= (uint64_t) *bp++;
+                hval += (hval << 1) + (hval << 4) + (hval << 5) +
+                        (hval << 7) + (hval << 8) + (hval << 40);
+        }
+        return hval;
+}
+
+static inline int is_data_obj_writeable(SheepdogInode *inode, unsigned int idx)
+{
+	return inode->vdi_id == inode->data_vdi_id[idx];
+}
+
+static inline int is_data_obj(uint64_t oid)
+{
+	return !(VDI_BIT & oid);
+}
+
+static inline uint64_t data_oid_to_idx(uint64_t oid)
+{
+	return oid & (MAX_DATA_OBJS - 1);
+}
+
+static inline uint64_t vid_to_vdi_oid(uint32_t vid)
+{
+	return VDI_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT);
+}
+
+static inline uint64_t vid_to_vmstate_oid(uint32_t vid, uint32_t idx)
+{
+	return VMSTATE_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
+}
+
+static inline uint64_t vid_to_data_oid(uint32_t vid, uint32_t idx)
+{
+	return ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
+}
+
+#undef dprintf
+#ifdef DEBUG_SDOG
+#define dprintf(fmt, args...)						\
+do {									\
+	fprintf(stdout, "%s %d: " fmt, __func__, __LINE__, ##args);	\
+} while (0)
+#else
+#define dprintf(fmt, args...)
+#endif
+
+#define min_t(type, x, y) ({			\
+	type __min1 = (x);			\
+	type __min2 = (y);			\
+	__min1 < __min2 ? __min1: __min2; })
+
+#define max_t(type, x, y) ({			\
+	type __max1 = (x);			\
+	type __max2 = (y);			\
+	__max1 > __max2 ? __max1: __max2; })
+
+typedef struct SheepdogAIOCB SheepdogAIOCB;
+
+typedef struct AIOReq {
+	SheepdogAIOCB *aiocb;
+	unsigned int iov_offset;
+
+	uint64_t oid;
+	uint64_t base_oid;
+	uint64_t offset;
+	unsigned int data_len;
+	uint8_t flags;
+	uint32_t id;
+
+	QLIST_ENTRY(AIOReq) outstanding_aio_siblings;
+	QLIST_ENTRY(AIOReq) aioreq_siblings;
+} AIOReq;
+
+enum AIOCBState {
+	AIOCB_WRITE_UDATA,
+	AIOCB_READ_UDATA,
+};
+
+struct SheepdogAIOCB {
+	BlockDriverAIOCB common;
+
+	QEMUIOVector *qiov;
+
+	int64_t sector_num;
+	int nb_sectors;
+
+	int ret;
+	enum AIOCBState aiocb_type;
+
+	QEMUBH *bh;
+	void (*aio_done_func)(SheepdogAIOCB *);
+
+	int canceled;
+
+	QLIST_HEAD(aioreq_head, AIOReq) aioreq_head;
+};
+
+typedef struct BDRVSheepdogState {
+	SheepdogInode inode;
+
+	uint32_t min_dirty_data_idx;
+	uint32_t max_dirty_data_idx;
+
+	char name[SD_MAX_VDI_LEN];
+	int is_current;
+
+	char *addr;
+	int fd;
+
+	uint32_t aioreq_seq_num;
+	QLIST_HEAD(outstanding_aio_head, AIOReq) outstanding_aio_head;
+} BDRVSheepdogState;
+
+static const char * sd_strerror(int err)
+{
+	int i;
+
+	static const struct {
+		int err;
+		const char *desc;
+	} errors[] = {
+		{SD_RES_SUCCESS, "Success"},
+		{SD_RES_UNKNOWN, "Unknown error"},
+		{SD_RES_NO_OBJ, "No object found"},
+		{SD_RES_EIO, "I/O error"},
+		{SD_RES_VDI_EXIST, "VDI exists already"},
+		{SD_RES_INVALID_PARMS, "Invalid parameters"},
+		{SD_RES_SYSTEM_ERROR, "System error"},
+		{SD_RES_VDI_LOCKED, "VDI is already locked"},
+		{SD_RES_NO_VDI, "No vdi found"},
+		{SD_RES_NO_BASE_VDI, "No base VDI found"},
+		{SD_RES_VDI_READ, "Failed read the requested VDI"},
+		{SD_RES_VDI_WRITE, "Failed to write the requested VDI"},
+		{SD_RES_BASE_VDI_READ, "Failed to read the base VDI"},
+		{SD_RES_BASE_VDI_WRITE, "Failed to write the base VDI"},
+		{SD_RES_NO_TAG, "Failed to find the requested tag"},
+		{SD_RES_STARTUP, "The system is still booting"},
+		{SD_RES_VDI_NOT_LOCKED, "VDI isn't locked"},
+		{SD_RES_SHUTDOWN, "The system is shutting down"},
+		{SD_RES_NO_MEM, "Out of memory on the server"},
+		{SD_RES_FULL_VDI, "We already have the maximum vdis"},
+		{SD_RES_VER_MISMATCH, "Protocol version mismatch"},
+		{SD_RES_NO_SPACE, "Server has no space for new objects"},
+		{SD_RES_WAIT_FOR_FORMAT, "Sheepdog is waiting for a format operation"},
+		{SD_RES_WAIT_FOR_JOIN, "Sheepdog is waiting for other nodes joining"},
+		{SD_RES_JOIN_FAILED, "Target node had failed to join sheepdog"},
+	};
+
+	for (i = 0; i < ARRAY_SIZE(errors); ++i) {
+		if (errors[i].err == err) {
+			return errors[i].desc;
+		}
+	}
+
+	return "Invalid error code";
+}
+
+static inline AIOReq *alloc_aio_req(BDRVSheepdogState *s,
+				    SheepdogAIOCB *acb,
+				    uint64_t oid, unsigned int data_len,
+				    uint64_t offset, uint8_t flags,
+				    uint64_t base_oid,
+				    unsigned int iov_offset)
+{
+	AIOReq *aio_req;
+
+	aio_req = qemu_malloc(sizeof(*aio_req));
+	aio_req->aiocb = acb;
+	aio_req->iov_offset = iov_offset;
+	aio_req->oid = oid;
+	aio_req->base_oid = base_oid;
+	aio_req->offset = offset;
+	aio_req->data_len = data_len;
+	aio_req->flags = flags;
+	aio_req->id = s->aioreq_seq_num++;
+
+	QLIST_INSERT_HEAD(&s->outstanding_aio_head, aio_req,
+			  outstanding_aio_siblings);
+	QLIST_INSERT_HEAD(&acb->aioreq_head, aio_req, aioreq_siblings);
+
+	return aio_req;
+}
+
+static inline int free_aio_req(BDRVSheepdogState *s, AIOReq *aio_req)
+{
+	SheepdogAIOCB *acb = aio_req->aiocb;
+	QLIST_REMOVE(aio_req, outstanding_aio_siblings);
+	QLIST_REMOVE(aio_req, aioreq_siblings);
+	qemu_free(aio_req);
+
+	return !QLIST_EMPTY(&acb->aioreq_head);
+}
+
+static void sd_finish_aiocb(SheepdogAIOCB *acb)
+{
+	if (!acb->canceled) {
+		acb->common.cb(acb->common.opaque, acb->ret);
+	}
+	qemu_aio_release(acb);
+}
+
+static void sd_aio_cancel(BlockDriverAIOCB *blockacb)
+{
+	SheepdogAIOCB *acb = (SheepdogAIOCB *)blockacb;
+
+	acb->canceled = 1;
+}
+
+static AIOPool sd_aio_pool = {
+	.aiocb_size = sizeof(SheepdogAIOCB),
+	.cancel = sd_aio_cancel,
+};
+
+static SheepdogAIOCB *sd_aio_setup(BlockDriverState *bs, QEMUIOVector *qiov,
+				   int64_t sector_num, int nb_sectors,
+				   BlockDriverCompletionFunc *cb,
+				   void *opaque)
+{
+	SheepdogAIOCB *acb;
+
+	acb = qemu_aio_get(&sd_aio_pool, bs, cb, opaque);
+
+	acb->qiov = qiov;
+
+	acb->sector_num = sector_num;
+	acb->nb_sectors = nb_sectors;
+
+	acb->aio_done_func = NULL;
+	acb->canceled = 0;
+	acb->bh = NULL;
+	acb->ret = 0;
+	QLIST_INIT(&acb->aioreq_head);
+	return acb;
+}
+
+static int sd_schedule_bh(QEMUBHFunc *cb, SheepdogAIOCB *acb)
+{
+	if (acb->bh) {
+		error_report("bug: %d %d\n", acb->aiocb_type, acb->aiocb_type);
+		return -EIO;
+	}
+
+	acb->bh = qemu_bh_new(cb, acb);
+	if (!acb->bh) {
+		error_report("oom: %d %d\n", acb->aiocb_type, acb->aiocb_type);
+		return -EIO;
+	}
+
+	qemu_bh_schedule(acb->bh);
+
+	return 0;
+}
+
+static int do_send_recv(int sockfd, struct iovec *iov, int len, int offset,
+			int write)
+{
+	struct msghdr msg;
+	int ret, diff;
+
+	memset(&msg, 0, sizeof(msg));
+	msg.msg_iov = iov;
+	msg.msg_iovlen = 1;
+
+	len += offset;
+
+	while (iov->iov_len < len) {
+		len -= iov->iov_len;
+
+		iov++;
+		msg.msg_iovlen++;
+	}
+
+	diff = iov->iov_len - len;
+	iov->iov_len -= diff;
+
+	while (msg.msg_iov->iov_len <= offset) {
+		offset -= msg.msg_iov->iov_len;
+
+		msg.msg_iov++;
+		msg.msg_iovlen--;
+	}
+
+	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base + offset;
+	msg.msg_iov->iov_len -= offset;
+
+	if (write) {
+		ret = sendmsg(sockfd, &msg, 0);
+	} else {
+		ret = recvmsg(sockfd, &msg, MSG_WAITALL);
+	}
+
+	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base - offset;
+	msg.msg_iov->iov_len += offset;
+
+	iov->iov_len += diff;
+	return ret;
+}
+
+static int connect_to_sdog(const char *addr)
+{
+	char buf[64];
+	char hbuf[NI_MAXHOST], sbuf[NI_MAXSERV];
+	char name[256], *p;
+	int fd, ret;
+	struct addrinfo hints, *res, *res0;
+	int port = 0;
+
+	if (!addr) {
+		addr = SD_DEFAULT_ADDR;
+	}
+
+	strcpy(name, addr);
+
+	p = name;
+	while (*p) {
+		if (*p == ':') {
+			*p++ = '\0';
+			break;
+		} else {
+			p++;
+		}
+	}
+
+	if (*p == '\0') {
+		error_report("cannot find a port number, %s\n", name);
+		return -1;
+	}
+	port = strtol(p, NULL, 10);
+	if (port == 0) {
+		error_report("invalid port number, %s\n", p);
+		return -1;
+	}
+
+	memset(&hints, 0, sizeof(hints));
+	snprintf(buf, sizeof(buf), "%d", port);
+
+	hints.ai_socktype = SOCK_STREAM;
+
+	ret = getaddrinfo(name, buf, &hints, &res0);
+	if (ret) {
+		error_report("unable to get address info %s, %m\n", name);
+		return -1;
+	}
+
+	for (res = res0; res; res = res->ai_next) {
+		ret = getnameinfo(res->ai_addr, res->ai_addrlen,
+				  hbuf, sizeof(hbuf), sbuf, sizeof(sbuf),
+				  NI_NUMERICHOST | NI_NUMERICSERV);
+		if (ret) {
+			continue;
+		}
+
+		fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol);
+		if (fd < 0) {
+			continue;
+		}
+
+reconnect:
+		ret = connect(fd, res->ai_addr, res->ai_addrlen);
+		if (ret < 0) {
+			if (errno == EINTR) {
+				goto reconnect;
+			}
+			break;
+		}
+
+		dprintf("connected to %s:%d\n", name, port);
+		goto success;
+	}
+	fd = -1;
+	error_report("failed connect to %s:%d\n", name, port);
+success:
+	freeaddrinfo(res0);
+	return fd;
+}
+
+static int do_readv_writev(int sockfd, struct iovec *iov, int len,
+			   int iov_offset, int write)
+{
+	int ret;
+again:
+	ret = do_send_recv(sockfd, iov, len, iov_offset, write);
+	if (ret < 0) {
+		if (errno == EINTR || errno == EAGAIN) {
+			goto again;
+		}
+		error_report("failed to recv a rsp, %m\n");
+		return 1;
+	}
+
+	iov_offset += ret;
+	len -= ret;
+	if (len) {
+		goto again;
+	}
+
+	return 0;
+}
+
+static int do_readv(int sockfd, struct iovec *iov, int len, int iov_offset)
+{
+	return do_readv_writev(sockfd, iov, len, iov_offset, 0);
+}
+
+static int do_writev(int sockfd, struct iovec *iov, int len, int iov_offset)
+{
+	return do_readv_writev(sockfd, iov, len, iov_offset, 1);
+}
+
+static int do_read_write(int sockfd, void *buf, int len, int write)
+{
+	struct iovec iov;
+
+	iov.iov_base = buf;
+	iov.iov_len = len;
+
+	return do_readv_writev(sockfd, &iov, len, 0, write);
+}
+
+static int do_read(int sockfd, void *buf, int len)
+{
+	return do_read_write(sockfd, buf, len, 0);
+}
+
+static int do_write(int sockfd, void *buf, int len)
+{
+	return do_read_write(sockfd, buf, len, 1);
+}
+
+static int send_req(int sockfd, SheepdogReq *hdr, void *data,
+		    unsigned int *wlen)
+{
+	int ret;
+	struct iovec iov[2];
+
+	iov[0].iov_base = hdr;
+	iov[0].iov_len = sizeof(*hdr);
+
+	if (*wlen) {
+		iov[1].iov_base = data;
+		iov[1].iov_len = *wlen;
+	}
+
+	ret = do_writev(sockfd, iov, sizeof(*hdr) + *wlen, 0);
+	if (ret) {
+		error_report("failed to send a req, %m\n");
+		ret = -1;
+	}
+
+	return ret;
+}
+
+static int do_req(int sockfd, SheepdogReq *hdr, void *data,
+		  unsigned int *wlen, unsigned int *rlen)
+{
+	int ret;
+
+	ret = send_req(sockfd, hdr, data, wlen);
+	if (ret) {
+		ret = -1;
+		goto out;
+	}
+
+	ret = do_read(sockfd, hdr, sizeof(*hdr));
+	if (ret) {
+		error_report("failed to get a rsp, %m\n");
+		ret = -1;
+		goto out;
+	}
+
+	if (*rlen > hdr->data_length) {
+		*rlen = hdr->data_length;
+	}
+
+	if (*rlen) {
+		ret = do_read(sockfd, data, *rlen);
+		if (ret) {
+			error_report("failed to get the data, %m\n");
+			ret = -1;
+			goto out;
+		}
+	}
+	ret = 0;
+out:
+	return ret;
+}
+
+static int add_aio_request(BDRVSheepdogState *s, AIOReq *aio_req,
+			   struct iovec *iov, int niov, int create,
+			   enum AIOCBState aiocb_type);
+
+static void send_pending_req(BDRVSheepdogState *s, uint64_t oid, uint32_t id)
+{
+	AIOReq *aio_req, *next;
+	SheepdogAIOCB *acb;
+	int ret;
+
+	QLIST_FOREACH_SAFE(aio_req, &s->outstanding_aio_head,
+			   outstanding_aio_siblings, next) {
+		if (id == aio_req->id) {
+			continue;
+		}
+		if (aio_req->oid != oid) {
+			continue;
+		}
+
+		acb = aio_req->aiocb;
+		ret = add_aio_request(s, aio_req, acb->qiov->iov,
+				      acb->qiov->niov, 0, acb->aiocb_type);
+		if (ret < 0) {
+			error_report("add_aio_request is faled\n");
+			free_aio_req(s, aio_req);
+			if (QLIST_EMPTY(&acb->aioreq_head)) {
+				sd_finish_aiocb(acb);
+			}
+		}
+	}
+}
+
+static void aio_read_response(void *opaque)
+{
+	SheepdogObjReq hdr;
+	SheepdogObjRsp *rsp = (SheepdogObjRsp *)&hdr;
+	BDRVSheepdogState *s = (BDRVSheepdogState *)opaque;
+	int fd = s->fd;
+	int ret;
+	AIOReq *aio_req = NULL;
+	SheepdogAIOCB *acb;
+	int rest;
+	unsigned long idx;
+
+	if (QLIST_EMPTY(&s->outstanding_aio_head)) {
+		return;
+	}
+
+	ret = do_read(fd, (void *)rsp, sizeof(*rsp));
+	if (ret) {
+		error_report("failed to get the header, %m\n");
+		return;
+	}
+
+	QLIST_FOREACH(aio_req, &s->outstanding_aio_head, outstanding_aio_siblings) {
+		if (aio_req->id == rsp->id) {
+			break;
+		}
+	}
+	if (!aio_req) {
+		error_report("cannot find aio_req %x\n", rsp->id);
+		return;
+	}
+
+	acb = aio_req->aiocb;
+
+	switch (acb->aiocb_type) {
+	case AIOCB_WRITE_UDATA:
+		if (!is_data_obj(aio_req->oid)) {
+			break;
+		}
+		idx = data_oid_to_idx(aio_req->oid);
+
+		if (s->inode.data_vdi_id[idx] != s->inode.vdi_id) {
+			s->inode.data_vdi_id[idx] = s->inode.vdi_id;
+			s->max_dirty_data_idx = max_t(uint32_t, idx,
+						      s->max_dirty_data_idx);
+			s->min_dirty_data_idx = min_t(uint32_t, idx,
+						      s->min_dirty_data_idx);
+
+			send_pending_req(s, vid_to_data_oid(s->inode.vdi_id, idx),
+					 rsp->id);
+		}
+		break;
+	case AIOCB_READ_UDATA:
+		ret = do_readv(fd, acb->qiov->iov, rsp->data_length,
+			       aio_req->iov_offset);
+		if (ret) {
+			error_report("failed to get the data, %m\n");
+			return;
+		}
+		break;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		acb->ret = -EIO;
+		error_report("%s\n", sd_strerror(rsp->result));
+	}
+
+	rest = free_aio_req(s, aio_req);
+	if (!rest) {
+		acb->aio_done_func(acb);
+	}
+}
+
+static int aio_flush_request(void *opaque)
+{
+	BDRVSheepdogState *s = (BDRVSheepdogState *)opaque;
+
+	return !QLIST_EMPTY(&s->outstanding_aio_head);
+}
+
+static int set_nonblocking(int fd)
+{
+	int ret;
+
+	ret = fcntl(fd, F_GETFL);
+	if (ret < 0) {
+		error_report("can't fcntl (F_GETFL), %m\n");
+		close(fd);
+	} else {
+		ret = fcntl(fd, F_SETFL, ret | O_NONBLOCK);
+		if (ret < 0) {
+			error_report("can't fcntl (O_NONBLOCK), %m\n");
+		}
+	}
+
+	return ret;
+}
+
+static int set_nodelay(int fd)
+{
+	int ret, opt;
+
+	opt = 1;
+	ret = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &opt, sizeof(opt));
+	return ret;
+}
+
+/*
+ * Return a socket discriptor to read/write objects.
+ * We cannot use this discriptor for other operations because
+ * the block driver may be on waiting response from the server.
+ */
+static int get_sheep_fd(BDRVSheepdogState *s)
+{
+	int ret, fd;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		error_report("%m\n");
+		return -1;
+	}
+
+	ret = set_nonblocking(fd);
+	if (ret) {
+		error_report("%m\n");
+		close(fd);
+		return -1;
+	}
+
+	ret = set_nodelay(fd);
+	if (ret) {
+		error_report("%m\n");
+		close(fd);
+		return -1;
+	}
+
+	qemu_aio_set_fd_handler(fd, aio_read_response, NULL, aio_flush_request,
+				NULL, s);
+	s->fd = fd;
+
+	return fd;
+}
+
+static int parse_vdiname(BDRVSheepdogState *s, const char *filename,
+			 char *vdi, int vdi_len, uint32_t *snapid)
+{
+	char *p, *q;
+	int nr_sep;
+
+	p = q = strdup(filename);
+
+	if (!p) {
+		return 1;
+	}
+
+	nr_sep = 0;
+	while (*p) {
+		if (*p == ':') {
+			nr_sep++;
+		}
+		if (nr_sep == 2) {
+			break;
+		}
+		p++;
+	}
+
+	if (nr_sep == 2) {
+		*p++ = '\0';
+	} else {
+		p = q;
+	}
+
+	strncpy(vdi, p, vdi_len);
+
+	p = strchr(vdi, ':');
+	if (p) {
+		*p++ = '\0';
+		*snapid = strtol(p, NULL, 10);
+	} else {
+		*snapid = CURRENT_VDI_ID; /* search current vdi */
+	}
+
+	if (nr_sep == 2) {
+		s->addr = q;
+	} else {
+		free(q);
+		s->addr = NULL;
+	}
+
+	return 0;
+}
+
+static int find_vdi_name(BDRVSheepdogState *s, char *filename, uint32_t snapid,
+			 uint32_t *vid, int for_snapshot)
+{
+	int ret, fd;
+	SheepdogVdiReq hdr;
+	SheepdogVdiRsp *rsp = (SheepdogVdiRsp *)&hdr;
+	unsigned int wlen, rlen = 0;
+	char buf[SD_MAX_VDI_LEN];
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		return -1;
+	}
+
+	memset(&hdr, 0, sizeof(hdr));
+	snprintf(buf, sizeof(buf), "%s", filename);
+	if (for_snapshot) {
+		hdr.opcode = SD_OP_GET_VDI_INFO;
+	} else {
+		hdr.opcode = SD_OP_LOCK_VDI;
+	}
+	wlen = SD_MAX_VDI_LEN;
+	hdr.proto_ver = SD_PROTO_VER;
+	hdr.data_length = SD_MAX_VDI_LEN;
+	hdr.snapid = snapid;
+	hdr.flags = SD_FLAG_CMD_WRITE;
+
+	ret = do_req(fd, (SheepdogReq *)&hdr, buf, &wlen, &rlen);
+	if (ret) {
+		ret = -1;
+		goto out;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		error_report("%s, %s\n", sd_strerror(rsp->result), filename);
+		ret = -1;
+		goto out;
+	}
+	*vid = rsp->vdi_id;
+
+	ret = 0;
+out:
+	close(fd);
+	return ret;
+}
+
+static int add_aio_request(BDRVSheepdogState *s, AIOReq *aio_req,
+			   struct iovec *iov, int niov, int create,
+			   enum AIOCBState aiocb_type)
+{
+	int nr_copies = s->inode.nr_copies;
+	SheepdogObjReq hdr;
+	unsigned int wlen;
+	int ret, opt;
+	uint64_t oid = aio_req->oid;
+	unsigned int datalen = aio_req->data_len;
+	uint64_t offset = aio_req->offset;
+	uint8_t flags = aio_req->flags;
+	uint64_t old_oid = aio_req->base_oid;
+
+	if (!nr_copies) {
+		error_report("bug\n");
+	}
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	if (aiocb_type == AIOCB_READ_UDATA) {
+		wlen = 0;
+		hdr.opcode = SD_OP_READ_OBJ;
+		hdr.flags = flags;
+	} else if (create) {
+		wlen = datalen;
+		hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
+		hdr.flags = SD_FLAG_CMD_WRITE | flags;
+	} else {
+		wlen = datalen;
+		hdr.opcode = SD_OP_WRITE_OBJ;
+		hdr.flags = SD_FLAG_CMD_WRITE | flags;
+	}
+
+	hdr.oid = oid;
+	hdr.cow_oid = old_oid;
+	hdr.copies = s->inode.nr_copies;
+
+	hdr.data_length = datalen;
+	hdr.offset = offset;
+
+	hdr.id = aio_req->id;
+
+	opt = 1;
+	setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
+
+	ret = do_write(s->fd, &hdr, sizeof(hdr));
+	if (ret) {
+		error_report("failed to send a req, %m\n");
+		return -EIO;
+	}
+
+	if (wlen) {
+		ret = do_writev(s->fd, iov, wlen, aio_req->iov_offset);
+		if (ret) {
+			error_report("failed to send a data, %m\n");
+			return -EIO;
+		}
+	}
+        opt = 0;
+        setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
+
+	return 0;
+}
+
+static int read_write_object(int fd, char *buf, uint64_t oid, int copies,
+			     unsigned int datalen, uint64_t offset,
+			     int write, int create)
+{
+	SheepdogObjReq hdr;
+	SheepdogObjRsp *rsp = (SheepdogObjRsp *)&hdr;
+	unsigned int wlen, rlen;
+	int ret;
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	if (write) {
+		wlen = datalen;
+		rlen = 0;
+		hdr.flags = SD_FLAG_CMD_WRITE;
+		if (create) {
+			hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
+		} else {
+			hdr.opcode = SD_OP_WRITE_OBJ;
+		}
+	} else {
+		wlen = 0;
+		rlen = datalen;
+		hdr.opcode = SD_OP_READ_OBJ;
+	}
+	hdr.oid = oid;
+	hdr.data_length = datalen;
+	hdr.offset = offset;
+	hdr.copies = copies;
+
+	ret = do_req(fd, (SheepdogReq *)&hdr, buf, &wlen, &rlen);
+	if (ret) {
+		error_report("failed to send a request to the sheep\n");
+		return -1;
+	}
+
+	switch (rsp->result) {
+	case SD_RES_SUCCESS:
+		return 0;
+	default:
+		error_report("%s\n", sd_strerror(rsp->result));
+		return -1;
+	}
+}
+
+static int read_object(int fd, char *buf, uint64_t oid, int copies,
+		       unsigned int datalen, uint64_t offset)
+{
+	return read_write_object(fd, buf, oid, copies, datalen, offset, 0, 0);
+}
+
+static int write_object(int fd, char *buf, uint64_t oid, int copies,
+			unsigned int datalen, uint64_t offset, int create)
+{
+	return read_write_object(fd, buf, oid, copies, datalen, offset, 1, create);
+}
+
+/* TODO: error cleanups */
+static int sd_open(BlockDriverState *bs, const char *filename, int flags)
+{
+	int ret, fd;
+	uint32_t vid = 0;
+	BDRVSheepdogState *s = bs->opaque;
+	char vdi[256];
+	uint32_t snapid;
+	int for_snapshot = 0;
+	char *buf;
+
+	strstart(filename, "sheepdog:", (const char **)&filename);
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+
+	memset(vdi, 0, sizeof(vdi));
+	if (parse_vdiname(s, filename, vdi, sizeof(vdi), &snapid) < 0) {
+		goto out;
+	}
+	s->fd = get_sheep_fd(s);
+	if (s->fd < 0) {
+		return -1;
+	}
+
+	if (snapid != CURRENT_VDI_ID) {
+		for_snapshot = 1;
+	}
+
+	ret = find_vdi_name(s, vdi, snapid, &vid, for_snapshot);
+	if (ret) {
+		goto out;
+	}
+
+	if (snapid) {
+		dprintf("%" PRIx32 " non current inode was open.\n", vid);
+	} else {
+		s->is_current = 1;
+	}
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		error_report("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), 0, SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret) {
+		goto out;
+	}
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+	s->min_dirty_data_idx = UINT32_MAX;
+	s->max_dirty_data_idx = 0;
+
+	bs->total_sectors = s->inode.vdi_size >> 9;
+	strncpy(s->name, vdi, sizeof(s->name));
+	qemu_free(buf);
+
+	QLIST_INIT(&s->outstanding_aio_head);
+	return 0;
+out:
+	qemu_free(buf);
+	return -1;
+}
+
+static int do_sd_create(const char *addr, char *filename, char *tag,
+			int64_t total_sectors, uint32_t base_vid,
+			uint32_t *vdi_id, int snapshot)
+{
+	SheepdogVdiReq hdr;
+	SheepdogVdiRsp *rsp = (SheepdogVdiRsp *)&hdr;
+	int fd, ret;
+	unsigned int wlen, rlen = 0;
+	char buf[SD_MAX_VDI_LEN];
+
+	fd = connect_to_sdog(addr);
+	if (fd < 0) {
+		return -1;
+	}
+
+	strncpy(buf, filename, SD_MAX_VDI_LEN);
+
+	memset(&hdr, 0, sizeof(hdr));
+	hdr.opcode = SD_OP_NEW_VDI;
+	hdr.base_vdi_id = base_vid;
+
+	wlen = SD_MAX_VDI_LEN;
+
+	hdr.flags = SD_FLAG_CMD_WRITE;
+	hdr.snapid = snapshot;
+
+	hdr.data_length = wlen;
+	hdr.vdi_size = total_sectors * 512;
+
+	ret = do_req(fd, (SheepdogReq *)&hdr, buf, &wlen, &rlen);
+
+	close(fd);
+
+	if (ret) {
+		return -1;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		error_report("%s, %s\n", sd_strerror(rsp->result), filename);
+		return -1;
+	}
+
+	if (vdi_id) {
+		*vdi_id = rsp->vdi_id;
+	}
+
+	return 0;
+}
+
+static int sd_create(const char *filename, QEMUOptionParameter *options)
+{
+	int ret;
+	uint32_t vid = 0;
+	int64_t total_sectors = 0;
+	char *backing_file = NULL;
+
+	strstart(filename, "sheepdog:", (const char **)&filename);
+
+	while (options && options->name) {
+		if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
+			total_sectors = options->value.n / 512;
+		} else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
+			backing_file = options->value.s;
+		}
+		options++;
+	}
+
+	if (backing_file) {
+		BlockDriverState bs;
+		char vdi[SD_MAX_VDI_LEN];
+		uint32_t snapid;
+
+		strstart(backing_file, "sheepdog:", (const char **)&backing_file);
+		memset(&bs, 0, sizeof(bs));
+
+		bs.opaque = qemu_malloc(sizeof(BDRVSheepdogState));
+
+		ret = sd_open(&bs, backing_file, 0);
+		if (ret < 0) {
+			return -1;
+		}
+
+		if (parse_vdiname(bs.opaque, backing_file, vdi, sizeof(vdi), &snapid) < 0) {
+			return -1;
+		}
+
+		/* cannot clone from a current inode */
+		if (snapid == CURRENT_VDI_ID) {
+			return -1;
+		}
+
+		ret = find_vdi_name(bs.opaque, vdi, snapid, &vid, 1);
+		if (ret) {
+			return -1;
+		}
+	}
+
+	return do_sd_create(NULL, (char *)filename, NULL, total_sectors, vid,
+			    NULL, 0);
+}
+
+static void sd_close(BlockDriverState *bs)
+{
+	BDRVSheepdogState *s = bs->opaque;
+	SheepdogVdiReq hdr;
+	SheepdogVdiRsp *rsp = (SheepdogVdiRsp *)&hdr;
+	unsigned int wlen, rlen = 0;
+	int fd, ret;
+
+	dprintf("%s\n", s->name);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		return;
+	}
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	hdr.opcode = SD_OP_RELEASE_VDI;
+	wlen = strlen(s->name) + 1;
+	hdr.data_length = wlen;
+	hdr.flags = SD_FLAG_CMD_WRITE;
+
+	ret = do_req(fd, (SheepdogReq *)&hdr, s->name, &wlen, &rlen);
+
+	close(fd);
+
+	if (!ret && rsp->result != SD_RES_SUCCESS &&
+	    rsp->result != SD_RES_VDI_NOT_LOCKED) {
+		error_report("%s, %s\n", sd_strerror(rsp->result), s->name);
+	}
+
+	close(s->fd);
+	free(s->addr);
+}
+
+static void sd_write_done(SheepdogAIOCB *acb)
+{
+	int ret;
+	BDRVSheepdogState *s = acb->common.bs->opaque;
+	struct iovec iov;
+	AIOReq *aio_req;
+	uint32_t offset, data_len, mn, mx;
+
+	mn = s->min_dirty_data_idx;
+	mx = s->max_dirty_data_idx;
+	if (mn <= mx) {
+		offset = sizeof(s->inode) - sizeof(s->inode.data_vdi_id) +
+			mn * sizeof(s->inode.data_vdi_id[0]);
+		data_len = (mx - mn + 1) * sizeof(s->inode.data_vdi_id[0]);
+
+		s->min_dirty_data_idx = UINT32_MAX;
+		s->max_dirty_data_idx = 0;
+
+		iov.iov_base = &s->inode;
+		iov.iov_len = sizeof(s->inode);
+		aio_req = alloc_aio_req(s, acb, vid_to_vdi_oid(s->inode.vdi_id),
+					data_len, offset, 0, 0, offset);
+		if (!aio_req) {
+			error_report("too many requests\n");
+			acb->ret = -EIO;
+			goto out;
+		}
+		ret = add_aio_request(s, aio_req, &iov, 1, 0, AIOCB_WRITE_UDATA);
+		if (ret) {
+			free_aio_req(s, aio_req);
+			acb->ret = -EIO;
+			goto out;
+		}
+
+		acb->aio_done_func = sd_finish_aiocb;
+		acb->aiocb_type = AIOCB_WRITE_UDATA;
+		return;
+	}
+out:
+	sd_finish_aiocb(acb);
+}
+
+static int sd_create_branch(BDRVSheepdogState *s)
+{
+	int ret, fd;
+	uint32_t vid;
+	char *buf;
+
+	dprintf("%" PRIx32 " is not current.\n", s->inode.vdi_id);
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+
+	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
+			   s->inode.vdi_id, &vid, 1);
+	if (ret) {
+		goto out;
+	}
+
+	dprintf("%" PRIx32 " is created.\n", vid);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		error_report("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
+			  SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret < 0) {
+		goto out;
+	}
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+
+	s->is_current = 1;
+	ret = 0;
+	dprintf("%" PRIx32 " was newly created.\n", s->inode.vdi_id);
+
+out:
+	qemu_free(buf);
+
+	return ret;
+}
+
+static void sd_readv_writev_bh_cb(void *p)
+{
+	SheepdogAIOCB *acb = p;
+	int ret = 0;
+	unsigned long len, done = 0, total = acb->nb_sectors * 512;
+	unsigned long idx = acb->sector_num * 512 / SD_DATA_OBJ_SIZE;
+	uint64_t oid;
+	uint64_t offset = (acb->sector_num * 512) % SD_DATA_OBJ_SIZE;
+	BDRVSheepdogState *s = acb->common.bs->opaque;
+	SheepdogInode *inode = &s->inode;
+	AIOReq *aio_req;
+
+	qemu_bh_delete(acb->bh);
+	acb->bh = NULL;
+
+	if (acb->aiocb_type == AIOCB_WRITE_UDATA && !s->is_current) {
+		ret = sd_create_branch(s);
+		if (ret) {
+			acb->ret = -EIO;
+			goto out;
+		}
+	}
+
+	while (done != total) {
+		uint8_t flags = 0;
+		uint64_t old_oid = 0;
+		int create = 0;
+
+		oid = vid_to_data_oid(inode->data_vdi_id[idx], idx);
+
+		len = min_t(unsigned long, total - done, SD_DATA_OBJ_SIZE - offset);
+
+		if (!inode->data_vdi_id[idx]) {
+			if (acb->aiocb_type == AIOCB_READ_UDATA) {
+				goto done;
+			}
+
+			create = 1;
+		} else if (acb->aiocb_type == AIOCB_WRITE_UDATA
+			   && !is_data_obj_writeable(inode, idx)) {
+			create = 1;
+			old_oid = oid;
+			flags = SD_FLAG_CMD_COW;
+		}
+
+		if (create) {
+			dprintf("update ino (%" PRIu32") %"
+				PRIu64 " %" PRIu64 " %" PRIu64 "\n",
+				inode->vdi_id, oid,
+				vid_to_data_oid(inode->data_vdi_id[idx], idx), idx);
+			oid = vid_to_data_oid(inode->vdi_id, idx);
+			dprintf("new oid %lx\n", oid);
+		}
+
+		aio_req = alloc_aio_req(s, acb, oid, len, offset, flags,
+					old_oid, done);
+		if (!aio_req) {
+			error_report("too many requests\n");
+			acb->ret = -EIO;
+			goto out;
+		}
+
+		if (create) {
+			AIOReq *areq;
+			QLIST_FOREACH(areq, &s->outstanding_aio_head,
+				      outstanding_aio_siblings) {
+				if (areq == aio_req) {
+					continue;
+				}
+				if (areq->oid == oid) {
+					aio_req->flags = 0;
+					aio_req->base_oid = 0;
+					goto done;
+				}
+			}
+		}
+
+		ret = add_aio_request(s, aio_req, acb->qiov->iov, acb->qiov->niov,
+				      create, acb->aiocb_type);
+		if (ret < 0) {
+			error_report("add_aio_request is faled\n");
+			free_aio_req(s, aio_req);
+			acb->ret = -EIO;
+			goto out;
+		}
+	done:
+		offset = 0;
+		idx++;
+		done += len;
+	}
+out:
+	if (QLIST_EMPTY(&acb->aioreq_head)) {
+		sd_finish_aiocb(acb);
+	}
+}
+
+static BlockDriverAIOCB *sd_aio_writev(BlockDriverState *bs,
+				       int64_t sector_num,
+				       QEMUIOVector *qiov,
+				       int nb_sectors,
+				       BlockDriverCompletionFunc *cb,
+				       void *opaque)
+{
+	SheepdogAIOCB *acb;
+
+	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
+	acb->aio_done_func = sd_write_done;
+	acb->aiocb_type = AIOCB_WRITE_UDATA;
+
+	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
+	return &acb->common;
+}
+
+static BlockDriverAIOCB *sd_aio_readv(BlockDriverState *bs,
+				      int64_t sector_num,
+				      QEMUIOVector *qiov,
+				      int nb_sectors,
+				      BlockDriverCompletionFunc *cb,
+				      void *opaque)
+{
+	SheepdogAIOCB *acb;
+	int i;
+
+	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
+	acb->aiocb_type = AIOCB_READ_UDATA;
+	acb->aio_done_func = sd_finish_aiocb;
+
+	/*
+	 * TODO: we can do better; we don't need to initialize
+	 * blindly.
+	 */
+	for (i = 0; i < qiov->niov; i++) {
+		memset(qiov->iov[i].iov_base, 0, qiov->iov[i].iov_len);
+	}
+
+	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
+	return &acb->common;
+}
+
+static int sd_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info)
+{
+	BDRVSheepdogState *s = bs->opaque;
+	int ret, fd;
+	uint32_t new_vid;
+	SheepdogInode *inode;
+	unsigned int datalen;
+	uint64_t offset;
+
+	dprintf("sn_info: name %s id_str %s s: name %s vm_state_size %d "
+		"is_current %d\n", sn_info->name, sn_info->id_str,
+		s->name, sn_info->vm_state_size, s->is_current);
+
+	if (!s->is_current) {
+		error_report("You can't create a snapshot of "
+			"a non current VDI, %s (%" PRIu32 ").\n",
+			s->name, s->inode.vdi_id);
+
+		return -1;
+	}
+
+	dprintf("%s %s\n", sn_info->name, sn_info->id_str);
+
+	s->inode.vm_state_size = sn_info->vm_state_size;
+	s->inode.vm_clock_nsec = sn_info->vm_clock_nsec;
+	offset = 0;
+	/* we don't need to read entire object */
+	datalen = SD_INODE_SIZE - sizeof(s->inode.data_vdi_id);
+
+	/* refresh inode. */
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	ret = write_object(fd, (char *)&s->inode, vid_to_vdi_oid(s->inode.vdi_id),
+			   s->inode.nr_copies, datalen, offset, 0);
+	if (ret < 0) {
+		error_report("failed to write snapshot's inode.\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
+			   s->inode.vdi_id, &new_vid, 1);
+	if (ret < 0) {
+		error_report("failed to create inode for snapshot. %m\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	inode = (SheepdogInode *)qemu_malloc(datalen);
+
+	ret = read_object(fd, (char *)inode, vid_to_vdi_oid(new_vid),
+			  s->inode.nr_copies, datalen, offset);
+
+	close(fd);
+
+	if (ret < 0) {
+		error_report("failed to read new inode info. %m\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	memcpy(&s->inode, inode, datalen);
+	dprintf("s->inode: name %s snap_id %x oid %x\n",
+		s->inode.name, s->inode.snap_id, s->inode.vdi_id);
+
+cleanup:
+	close(fd);
+	return ret;
+}
+
+static int sd_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
+{
+	BDRVSheepdogState *s = bs->opaque;
+	BDRVSheepdogState *old_s;
+	char vdi[SD_MAX_VDI_LEN];
+	char *buf = NULL;
+	uint32_t vid;
+	uint32_t snapid = 0;
+	int ret = -ENOENT, fd;
+
+	old_s = qemu_malloc(sizeof(BDRVSheepdogState));
+
+	memcpy(old_s, s, sizeof(BDRVSheepdogState));
+
+	snapid = strtol(snapshot_id, NULL, 10);
+	if (!snapid) {
+		error_report("Invalid snapshot_id\n");
+		goto out;
+	}
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	strncpy(vdi, s->name, sizeof(vdi));
+	ret = find_vdi_name(s, vdi, snapid, &vid, 1);
+	if (ret) {
+		error_report("Failed to find_vdi_name\n");
+		ret = -ENOENT;
+		goto out;
+	}
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		error_report("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
+			  SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret) {
+		ret = -ENOENT;
+		goto out;
+	}
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+
+	if (!s->inode.vm_state_size) {
+		error_report("Invalid snapshot\n");
+		ret = -ENOENT;
+		goto out;
+	}
+
+	s->is_current = 0;
+
+	qemu_free(buf);
+	qemu_free(old_s);
+
+	return 0;
+out:
+	/* recover bdrv_sd_state */
+	memcpy(s, old_s, sizeof(BDRVSheepdogState));
+	qemu_free(buf);
+	qemu_free(old_s);
+
+	error_report("failed to open. recover old bdrv_sd_state.\n");
+
+	return ret;
+}
+
+static int sd_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
+{
+	/* FIXME: Delete specified snapshot id.  */
+	return 0;
+}
+
+#define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
+#define BITS_PER_BYTE		8
+#define BITS_TO_LONGS(nr)	DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(long))
+#define DECLARE_BITMAP(name,bits) \
+	unsigned long name[BITS_TO_LONGS(bits)]
+
+#define BITS_PER_LONG (BITS_PER_BYTE * sizeof(long))
+
+static inline int test_bit(unsigned int nr, const unsigned long *addr)
+{
+	return ((1UL << (nr % BITS_PER_LONG)) &
+		(((unsigned long *)addr)[nr / BITS_PER_LONG])) != 0;
+}
+
+static int sd_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab)
+{
+	BDRVSheepdogState *s = bs->opaque;
+	SheepdogReq req;
+	int i, fd, nr = 1024, ret, max = BITS_TO_LONGS(SD_NR_VDIS) * sizeof(long);
+	QEMUSnapshotInfo *sn_tab = NULL;
+	unsigned wlen, rlen;
+	int found = 0;
+	static SheepdogInode inode;
+	unsigned long *vdi_inuse;
+	unsigned int start_nr;
+
+	vdi_inuse = qemu_malloc(max);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		goto out;
+	}
+
+	rlen = max;
+	wlen = 0;
+
+	memset(&req, 0, sizeof(req));
+
+	req.opcode = SD_OP_READ_VDIS;
+	req.data_length = max;
+
+	ret = do_req(fd, (SheepdogReq *)&req, vdi_inuse, &wlen, &rlen);
+
+	close(fd);
+	if (ret) {
+		goto out;
+	}
+
+	sn_tab = qemu_mallocz(nr * sizeof(*sn_tab));
+
+	start_nr = fnv_64a_buf(s->name, strlen(s->name), FNV1A_64_INIT) & (SD_NR_VDIS - 1);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		error_report("failed to connect\n");
+		goto out;
+	}
+
+	/* TODO: round up */
+	for (i = start_nr; i < SD_NR_VDIS && found < nr; i++) {
+		if (!test_bit(i, vdi_inuse)) {
+			break;
+		}
+
+		/* we don't need to read entire object */
+		ret = read_object(fd, (char *)&inode, vid_to_vdi_oid(i),
+				  0, SD_INODE_SIZE - sizeof(inode.data_vdi_id), 0);
+
+		if (ret) {
+			continue;
+		}
+
+		if (!strcmp(inode.name, s->name) && inode.snap_ctime) {
+			sn_tab[found].date_sec = inode.snap_ctime >> 32;
+			sn_tab[found].date_nsec = inode.snap_ctime & 0xffffffff;
+			sn_tab[found].vm_state_size = inode.vm_state_size;
+			sn_tab[found].vm_clock_nsec = inode.vm_clock_nsec;
+
+			snprintf(sn_tab[found].id_str, sizeof(sn_tab[found].id_str), "%u",
+				 inode.snap_id);
+			found++;
+		}
+	}
+
+	close(fd);
+out:
+	*psn_tab = sn_tab;
+
+	qemu_free(vdi_inuse);
+
+	return found;
+}
+
+static int do_load_save_vmstate(BDRVSheepdogState *s, uint8_t *data,
+				int64_t pos, int size, int load)
+{
+	int fd, create;
+	int ret = 0;
+	unsigned int data_len;
+	uint64_t vmstate_oid;
+	uint32_t vdi_index;
+	uint64_t offset;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	while (size) {
+		vdi_index = pos / SD_DATA_OBJ_SIZE;
+		offset = pos % SD_DATA_OBJ_SIZE;
+
+		data_len = min_t(unsigned int, size, SD_DATA_OBJ_SIZE);
+
+		vmstate_oid = vid_to_vmstate_oid(s->inode.vdi_id, vdi_index);
+
+		create = (offset == 0);
+		if (load) {
+			ret = read_object(fd, (char *)data, vmstate_oid,
+					  s->inode.nr_copies, data_len, offset);
+		} else {
+			ret = write_object(fd, (char *)data, vmstate_oid,
+					   s->inode.nr_copies, data_len, offset, create);
+		}
+
+		if (ret < 0) {
+			error_report("failed to save vmstate %m\n");
+			ret = -EIO;
+			goto cleanup;
+		}
+
+		pos += data_len;
+		size -= data_len;
+		ret += data_len;
+	}
+cleanup:
+	close(fd);
+	return ret;
+}
+
+static int sd_save_vmstate(BlockDriverState *bs, const uint8_t *data,
+			   int64_t pos, int size)
+{
+	BDRVSheepdogState *s = bs->opaque;
+
+	return do_load_save_vmstate(s, (uint8_t *)data, pos, size, 0);
+}
+
+static int sd_load_vmstate(BlockDriverState *bs, uint8_t *data,
+			   int64_t pos, int size)
+{
+	BDRVSheepdogState *s = bs->opaque;
+
+	return do_load_save_vmstate(s, data, pos, size, 1);
+}
+
+
+static QEMUOptionParameter sd_create_options[] = {
+	{
+		.name = BLOCK_OPT_SIZE,
+		.type = OPT_SIZE,
+		.help = "Virtual disk size"
+	},
+	{
+		.name = BLOCK_OPT_BACKING_FILE,
+		.type = OPT_STRING,
+		.help = "File name of a base image"
+	},
+	{ NULL }
+};
+
+BlockDriver bdrv_sheepdog = {
+	.format_name    = "sheepdog",
+	.protocol_name  = "sheepdog",
+	.instance_size  = sizeof(BDRVSheepdogState),
+	.bdrv_file_open = sd_open,
+	.bdrv_close     = sd_close,
+	.bdrv_create    = sd_create,
+
+	.bdrv_aio_readv     = sd_aio_readv,
+	.bdrv_aio_writev    = sd_aio_writev,
+
+	.bdrv_snapshot_create   = sd_snapshot_create,
+	.bdrv_snapshot_goto     = sd_snapshot_goto,
+	.bdrv_snapshot_delete   = sd_snapshot_delete,
+	.bdrv_snapshot_list     = sd_snapshot_list,
+
+	.bdrv_save_vmstate  = sd_save_vmstate,
+	.bdrv_load_vmstate  = sd_load_vmstate,
+
+	.create_options = sd_create_options,
+};
+
+static void bdrv_sheepdog_init(void)
+{
+	bdrv_register(&bdrv_sheepdog);
+}
+block_init(bdrv_sheepdog_init);
-- 
1.5.6.5

[RFC PATCH 2/2] block: add sheepdog driver for distributed storage support

[MORITA Kazutaka]

Sheepdog is a distributed storage system for QEMU. It provides highly
available block level storage volumes to VMs like Amazon EBS.  This
patch adds a qemu block driver for Sheepdog.

Sheepdog features are:
- No node in the cluster is special (no metadata node, no control
  node, etc)
- Linear scalability in performance and capacity
- No single point of failure
- Autonomous management (zero configuration)
- Useful volume management support such as snapshot and cloning
- Thin provisioning
- Autonomous load balancing

The more details are available at the project site:
    http://www.osrg.net/sheepdog/

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 Makefile         |    2 +-
 block.c          |    3 +-
 block/sheepdog.c | 1828 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 1831 insertions(+), 2 deletions(-)
 create mode 100644 block/sheepdog.c

diff --git a/Makefile b/Makefile
index c1fa08c..d03cda1 100644
--- a/Makefile
+++ b/Makefile
@@ -97,7 +97,7 @@ block-obj-$(CONFIG_LINUX_AIO) += linux-aio.o
 
 block-nested-y += cow.o qcow.o vdi.o vmdk.o cloop.o dmg.o bochs.o vpc.o vvfat.o
 block-nested-y += qcow2.o qcow2-refcount.o qcow2-cluster.o qcow2-snapshot.o
-block-nested-y += parallels.o nbd.o
+block-nested-y += parallels.o nbd.o sheepdog.o
 block-nested-$(CONFIG_WIN32) += raw-win32.o
 block-nested-$(CONFIG_POSIX) += raw-posix.o
 block-nested-$(CONFIG_CURL) += curl.o
diff --git a/block.c b/block.c
index a606820..ab00f3f 100644
--- a/block.c
+++ b/block.c
@@ -307,7 +307,8 @@ static BlockDriver *find_image_format(const char *filename)
 
     drv = find_protocol(filename);
     /* no need to test disk image formats for vvfat */
-    if (drv && strcmp(drv->format_name, "vvfat") == 0)
+    if (drv && (!strcmp(drv->format_name, "vvfat") ||
+        !strcmp(drv->format_name, "sheepdog")))
         return drv;
 
     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
diff --git a/block/sheepdog.c b/block/sheepdog.c
new file mode 100644
index 0000000..7c07a52
--- /dev/null
+++ b/block/sheepdog.c
@@ -0,0 +1,1828 @@
+/*
+ * Copyright (C) 2009-2010 Nippon Telegraph and Telephone Corporation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+#include <netdb.h>
+#include <netinet/tcp.h>
+
+#include "qemu-common.h"
+#include "block_int.h"
+
+#define SD_PROTO_VER 0x01
+
+#define SD_DEFAULT_ADDR "localhost:7000"
+
+#define SD_OP_CREATE_AND_WRITE_OBJ  0x01
+#define SD_OP_READ_OBJ       0x02
+#define SD_OP_WRITE_OBJ      0x03
+
+#define SD_OP_NEW_VDI        0x11
+#define SD_OP_LOCK_VDI       0x12
+#define SD_OP_RELEASE_VDI    0x13
+#define SD_OP_GET_VDI_INFO   0x14
+#define SD_OP_READ_VDIS      0x15
+
+#define SD_FLAG_CMD_WRITE    0x01
+#define SD_FLAG_CMD_COW      0x02
+
+#define SD_RES_SUCCESS       0x00 /* Success */
+#define SD_RES_UNKNOWN       0x01 /* Unknown error */
+#define SD_RES_NO_OBJ        0x02 /* No object found */
+#define SD_RES_EIO           0x03 /* I/O error */
+#define SD_RES_VDI_EXIST     0x04 /* Vdi exists already */
+#define SD_RES_INVALID_PARMS 0x05 /* Invalid parameters */
+#define SD_RES_SYSTEM_ERROR  0x06 /* System error */
+#define SD_RES_VDI_LOCKED    0x07 /* Vdi is locked */
+#define SD_RES_NO_VDI        0x08 /* No vdi found */
+#define SD_RES_NO_BASE_VDI   0x09 /* No base vdi found */
+#define SD_RES_VDI_READ      0x0A /* Cannot read requested vdi */
+#define SD_RES_VDI_WRITE     0x0B /* Cannot write requested vdi */
+#define SD_RES_BASE_VDI_READ 0x0C /* Cannot read base vdi */
+#define SD_RES_BASE_VDI_WRITE   0x0D /* Cannot write base vdi */
+#define SD_RES_NO_TAG        0x0E /* Requested tag is not found */
+#define SD_RES_STARTUP       0x0F /* Sheepdog is on starting up */
+#define SD_RES_VDI_NOT_LOCKED   0x10 /* Vdi is not locked */
+#define SD_RES_SHUTDOWN      0x11 /* Sheepdog is shutting down */
+#define SD_RES_NO_MEM        0x12 /* Cannot allocate memory */
+#define SD_RES_FULL_VDI      0x13 /* we already have the maximum vdis */
+#define SD_RES_VER_MISMATCH  0x14 /* Protocol version mismatch */
+#define SD_RES_NO_SPACE      0x15 /* Server has no room for new objects */
+#define SD_RES_WAIT_FOR_FORMAT  0x16 /* Sheepdog is waiting for a format operation */
+#define SD_RES_WAIT_FOR_JOIN    0x17 /* Sheepdog is waiting for other nodes joining */
+#define SD_RES_JOIN_FAILED   0x18 /* Target node had failed to join sheepdog */
+
+/*
+ * Object ID rules
+ *
+ *  0 - 19 (20 bits): data object space
+ * 20 - 31 (12 bits): reserved data object space
+ * 32 - 55 (24 bits): vdi object space
+ * 56 - 59 ( 4 bits): reserved vdi object space
+ * 60 - 63 ( 4 bits): object type indentifier space
+ */
+
+#define VDI_SPACE_SHIFT   32
+#define VDI_BIT (UINT64_C(1) << 63)
+#define VMSTATE_BIT (UINT64_C(1) << 62)
+#define MAX_DATA_OBJS (1ULL << 20)
+#define MAX_CHILDREN 1024
+#define SD_MAX_VDI_LEN 256
+#define SD_NR_VDIS   (1U << 24)
+#define SD_DATA_OBJ_SIZE (UINT64_C(1) << 22)
+
+#define SD_INODE_SIZE (sizeof(struct sd_inode))
+#define CURRENT_VDI_ID 0
+
+struct sd_req {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t	opcode_specific[8];
+};
+
+struct sd_rsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t	opcode_specific[7];
+};
+
+struct sd_obj_req {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint64_t        oid;
+	uint64_t        cow_oid;
+	uint32_t        copies;
+	uint32_t        rsvd;
+	uint64_t        offset;
+};
+
+struct sd_obj_rsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t        copies;
+	uint32_t        pad[6];
+};
+
+struct sd_vdi_req {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint64_t	vdi_size;
+	uint32_t        base_vdi_id;
+	uint32_t        copies;
+	uint32_t        snapid;
+	uint32_t        pad[3];
+};
+
+struct sd_vdi_rsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t        rsvd;
+	uint32_t        vdi_id;
+	uint32_t        pad[5];
+};
+
+struct sd_inode {
+	char name[SD_MAX_VDI_LEN];
+	uint64_t ctime;
+	uint64_t snap_ctime;
+	uint64_t vm_clock_nsec;
+	uint64_t vdi_size;
+	uint64_t vm_state_size;
+	uint16_t copy_policy;
+	uint8_t  nr_copies;
+	uint8_t  block_size_shift;
+	uint32_t snap_id;
+	uint32_t vdi_id;
+	uint32_t parent_vdi_id;
+	uint32_t child_vdi_id[MAX_CHILDREN];
+	uint32_t data_vdi_id[MAX_DATA_OBJS];
+};
+
+/*
+ * 64 bit FNV-1a non-zero initial basis
+ */
+#define FNV1A_64_INIT ((uint64_t)0xcbf29ce484222325ULL)
+
+/*
+ * 64 bit Fowler/Noll/Vo FNV-1a hash code
+ */
+static inline uint64_t fnv_64a_buf(void *buf, size_t len, uint64_t hval)
+{
+        unsigned char *bp = (unsigned char *) buf;
+        unsigned char *be = bp + len;
+        while (bp < be) {
+                hval ^= (uint64_t) *bp++;
+                hval += (hval << 1) + (hval << 4) + (hval << 5) +
+                        (hval << 7) + (hval << 8) + (hval << 40);
+        }
+        return hval;
+}
+
+static inline int is_data_obj_writeable(struct sd_inode *inode, unsigned int idx)
+{
+	return inode->vdi_id == inode->data_vdi_id[idx];
+}
+
+static inline int is_data_obj(uint64_t oid)
+{
+	return !(VDI_BIT & oid);
+}
+
+static inline uint64_t data_oid_to_idx(uint64_t oid)
+{
+	return oid & (MAX_DATA_OBJS - 1);
+}
+
+static inline uint64_t vid_to_vdi_oid(uint32_t vid)
+{
+	return VDI_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT);
+}
+
+static inline uint64_t vid_to_vmstate_oid(uint32_t vid, uint32_t idx)
+{
+	return VMSTATE_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
+}
+
+static inline uint64_t vid_to_data_oid(uint32_t vid, uint32_t idx)
+{
+	return ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
+}
+
+#undef eprintf
+#define eprintf(fmt, args...)						\
+do {									\
+	fprintf(stderr, "%s %d: " fmt, __func__, __LINE__, ##args);	\
+} while (0)
+
+#undef dprintf
+#ifdef DEBUG_SDOG
+#define dprintf(fmt, args...)						\
+do {									\
+	fprintf(stdout, "%s %d: " fmt, __func__, __LINE__, ##args);	\
+} while (0)
+#else
+#define dprintf(fmt, args...)
+#endif
+
+#define min_t(type, x, y) ({			\
+	type __min1 = (x);			\
+	type __min2 = (y);			\
+	__min1 < __min2 ? __min1: __min2; })
+
+struct aio_req {
+	struct sd_aiocb *aiocb;
+	unsigned int iov_offset;
+
+	uint64_t oid;
+	uint64_t base_oid;
+	uint64_t offset;
+	unsigned int data_len;
+	uint8_t flags;
+
+	QLIST_ENTRY(aio_req) pending_siblings;
+	QLIST_ENTRY(aio_req) aioreq_siblings;
+};
+
+enum aiocb_state {
+	AIOCB_WRITE_UDATA,
+	AIOCB_READ_UDATA,
+};
+
+struct sd_aiocb {
+	BlockDriverAIOCB common;
+
+	QEMUIOVector *qiov;
+
+	int64_t sector_num;
+	int nb_sectors;
+
+	int ret;
+	enum aiocb_state aiocb_type;
+
+	QEMUBH *bh;
+	void (*aio_done_func)(struct sd_aiocb *);
+
+	int canceled;
+
+	QLIST_HEAD(aioreq_head, aio_req) aioreq_head;
+};
+
+#define MAX_AIO_REQS 4096
+
+struct bdrv_sd_state {
+	struct sd_inode inode;
+
+	int nr_dirty_data_oids;
+	uint32_t dirty_data_oids[MAX_AIO_REQS];
+
+	char name[SD_MAX_VDI_LEN];
+	int is_current;
+
+	char *addr;
+	int fd;
+
+	struct aio_req aio_req_list[MAX_AIO_REQS];
+	struct aio_req *aio_req_free[MAX_AIO_REQS];
+	int nr_aio_req_free;
+
+	QLIST_HEAD(pending_head, aio_req) pending_head;
+};
+
+static const char * sd_strerror(int err)
+{
+	int i;
+
+	static const struct {
+		int err;
+		const char *desc;
+	} errors[] = {
+		{SD_RES_SUCCESS, "Success"},
+		{SD_RES_UNKNOWN, "Unknown error"},
+		{SD_RES_NO_OBJ, "No object found"},
+		{SD_RES_EIO, "I/O error"},
+		{SD_RES_VDI_EXIST, "VDI exists already"},
+		{SD_RES_INVALID_PARMS, "Invalid parameters"},
+		{SD_RES_SYSTEM_ERROR, "System error"},
+		{SD_RES_VDI_LOCKED, "VDI is already locked"},
+		{SD_RES_NO_VDI, "No vdi found"},
+		{SD_RES_NO_BASE_VDI, "No base VDI found"},
+		{SD_RES_VDI_READ, "Failed read the requested VDI"},
+		{SD_RES_VDI_WRITE, "Failed to write the requested VDI"},
+		{SD_RES_BASE_VDI_READ, "Failed to read the base VDI"},
+		{SD_RES_BASE_VDI_WRITE, "Failed to write the base VDI"},
+		{SD_RES_NO_TAG, "Failed to find the requested tag"},
+		{SD_RES_STARTUP, "The system is still booting"},
+		{SD_RES_VDI_NOT_LOCKED, "VDI isn't locked"},
+		{SD_RES_SHUTDOWN, "The system is shutting down"},
+		{SD_RES_NO_MEM, "Out of memory on the server"},
+		{SD_RES_FULL_VDI, "We already have the maximum vdis"},
+		{SD_RES_VER_MISMATCH, "Protocol version mismatch"},
+		{SD_RES_NO_SPACE, "Server has no space for new objects"},
+		{SD_RES_WAIT_FOR_FORMAT, "Sheepdog is waiting for a format operation"},
+		{SD_RES_WAIT_FOR_JOIN, "Sheepdog is waiting for other nodes joining"},
+		{SD_RES_JOIN_FAILED, "Target node had failed to join sheepdog"},
+	};
+
+	for (i = 0; i < ARRAY_SIZE(errors); ++i)
+		if (errors[i].err == err)
+			return errors[i].desc;
+
+	return "Invalid error code";
+}
+
+static inline int before(uint32_t seq1, uint32_t seq2)
+{
+        return (int32_t)(seq1 - seq2) < 0;
+}
+
+static inline int after(uint32_t seq1, uint32_t seq2)
+{
+	return (int32_t)(seq2 - seq1) < 0;
+}
+
+static inline struct aio_req *alloc_aio_req(struct bdrv_sd_state *s,
+					    struct sd_aiocb *acb,
+					    uint64_t oid, unsigned int data_len,
+					    uint64_t offset, uint8_t flags,
+					    uint64_t base_oid,
+					    unsigned int iov_offset)
+{
+	struct aio_req *aio_req;
+
+	if (!s->nr_aio_req_free)
+		return NULL;
+
+	aio_req = s->aio_req_free[--s->nr_aio_req_free];
+	aio_req->aiocb = acb;
+	aio_req->iov_offset = iov_offset;
+	aio_req->oid = oid;
+	aio_req->base_oid = base_oid;
+	aio_req->offset = offset;
+	aio_req->data_len = data_len;
+	aio_req->flags = flags;
+
+	QLIST_INSERT_HEAD(&s->pending_head, aio_req, pending_siblings);
+	QLIST_INSERT_HEAD(&acb->aioreq_head, aio_req, aioreq_siblings);
+
+	return aio_req;
+}
+
+static inline int free_aio_req(struct bdrv_sd_state *s, struct aio_req *aio_req)
+{
+	struct sd_aiocb *acb = aio_req->aiocb;
+	QLIST_REMOVE(aio_req, pending_siblings);
+	QLIST_REMOVE(aio_req, aioreq_siblings);
+	aio_req->aiocb = NULL;
+	s->aio_req_free[s->nr_aio_req_free++] = aio_req;
+
+	return !QLIST_EMPTY(&acb->aioreq_head);
+}
+
+static inline int nr_outstanding_aio_req(struct bdrv_sd_state *s)
+{
+	return MAX_AIO_REQS - s->nr_aio_req_free;
+}
+
+static inline int get_id_from_req(struct bdrv_sd_state *s, struct aio_req *aio_req)
+{
+	return aio_req - s->aio_req_list;
+}
+
+static inline struct aio_req *get_req_from_id(struct bdrv_sd_state *s, int id)
+{
+	return s->aio_req_list + id;
+}
+
+static void sd_finish_aiocb(struct sd_aiocb *acb)
+{
+	if (!acb->canceled)
+		acb->common.cb(acb->common.opaque, acb->ret);
+	qemu_aio_release(acb);
+}
+
+static void sd_aio_cancel(BlockDriverAIOCB *blockacb)
+{
+	struct sd_aiocb *acb = (struct sd_aiocb *)blockacb;
+
+	acb->canceled = 1;
+}
+
+static AIOPool sd_aio_pool = {
+	.aiocb_size = sizeof(struct sd_aiocb),
+	.cancel = sd_aio_cancel,
+};
+
+static struct sd_aiocb *sd_aio_setup(BlockDriverState *bs, QEMUIOVector *qiov,
+				     int64_t sector_num, int nb_sectors,
+				     BlockDriverCompletionFunc *cb,
+				     void *opaque)
+{
+	struct sd_aiocb *acb;
+
+	acb = qemu_aio_get(&sd_aio_pool, bs, cb, opaque);
+
+	acb->qiov = qiov;
+
+	acb->sector_num = sector_num;
+	acb->nb_sectors = nb_sectors;
+
+	acb->aio_done_func = NULL;
+	acb->canceled = 0;
+	acb->bh = NULL;
+	acb->ret = 0;
+	QLIST_INIT(&acb->aioreq_head);
+	return acb;
+}
+
+static int sd_schedule_bh(QEMUBHFunc *cb, struct sd_aiocb *acb)
+{
+	if (acb->bh) {
+		eprintf("bug: %d %d\n", acb->aiocb_type, acb->aiocb_type);
+		return -EIO;
+	}
+
+	acb->bh = qemu_bh_new(cb, acb);
+	if (!acb->bh) {
+		eprintf("oom: %d %d\n", acb->aiocb_type, acb->aiocb_type);
+		return -EIO;
+	}
+
+	qemu_bh_schedule(acb->bh);
+
+	return 0;
+}
+
+static int do_send_recv(int sockfd, struct iovec *iov, int len, int offset,
+			int write)
+{
+	struct msghdr msg;
+	int ret, diff;
+
+	memset(&msg, 0, sizeof(msg));
+	msg.msg_iov = iov;
+	msg.msg_iovlen = 1;
+
+	len += offset;
+
+	while (iov->iov_len < len) {
+		len -= iov->iov_len;
+
+		iov++;
+		msg.msg_iovlen++;
+	}
+
+	diff = iov->iov_len - len;
+	iov->iov_len -= diff;
+
+	while (msg.msg_iov->iov_len <= offset) {
+		offset -= msg.msg_iov->iov_len;
+
+		msg.msg_iov++;
+		msg.msg_iovlen--;
+	}
+
+	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base + offset;
+	msg.msg_iov->iov_len -= offset;
+
+	if (write)
+		ret = sendmsg(sockfd, &msg, 0);
+	else
+		ret = recvmsg(sockfd, &msg, MSG_WAITALL);
+
+	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base - offset;
+	msg.msg_iov->iov_len += offset;
+
+	iov->iov_len += diff;
+	return ret;
+}
+
+static int connect_to_sdog(const char *addr)
+{
+	char buf[64];
+	char hbuf[NI_MAXHOST], sbuf[NI_MAXSERV];
+	char name[256], *p;
+	int fd, ret;
+	struct addrinfo hints, *res, *res0;
+	int port = 0;
+
+	if (!addr)
+		addr = SD_DEFAULT_ADDR;
+
+	strcpy(name, addr);
+
+	p = name;
+	while (*p) {
+		if (*p == ':') {
+			*p++ = '\0';
+			break;
+		} else
+			p++;
+	}
+
+	if (*p == '\0') {
+		eprintf("cannot find a port number, %s\n", name);
+		return -1;
+	}
+	port = strtol(p, NULL, 10);
+	if (port == 0) {
+		eprintf("invalid port number, %s\n", p);
+		return -1;
+	}
+
+	memset(&hints, 0, sizeof(hints));
+	snprintf(buf, sizeof(buf), "%d", port);
+
+	hints.ai_socktype = SOCK_STREAM;
+
+	ret = getaddrinfo(name, buf, &hints, &res0);
+	if (ret) {
+		eprintf("unable to get address info %s, %m\n", name);
+		return -1;
+	}
+
+	for (res = res0; res; res = res->ai_next) {
+		ret = getnameinfo(res->ai_addr, res->ai_addrlen,
+				  hbuf, sizeof(hbuf), sbuf, sizeof(sbuf),
+				  NI_NUMERICHOST | NI_NUMERICSERV);
+		if (ret)
+			continue;
+
+		fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol);
+		if (fd < 0)
+			continue;
+
+reconnect:
+		ret = connect(fd, res->ai_addr, res->ai_addrlen);
+		if (ret < 0) {
+			if (errno == EINTR)
+				goto reconnect;
+			break;
+		}
+
+		dprintf("connected to %s:%d\n", name, port);
+		goto success;
+	}
+	fd = -1;
+	eprintf("failed connect to %s:%d\n", name, port);
+success:
+	freeaddrinfo(res0);
+	return fd;
+}
+
+static int do_readv_writev(int sockfd, struct iovec *iov, int len,
+			   int iov_offset, int write)
+{
+	int ret;
+again:
+	ret = do_send_recv(sockfd, iov, len, iov_offset, write);
+	if (ret < 0) {
+		if (errno == EINTR || errno == EAGAIN)
+			goto again;
+		eprintf("failed to recv a rsp, %m\n");
+		return 1;
+	}
+
+	iov_offset += ret;
+	len -= ret;
+	if (len)
+		goto again;
+
+	return 0;
+}
+
+static int do_readv(int sockfd, struct iovec *iov, int len, int iov_offset)
+{
+	return do_readv_writev(sockfd, iov, len, iov_offset, 0);
+}
+
+static int do_writev(int sockfd, struct iovec *iov, int len, int iov_offset)
+{
+	return do_readv_writev(sockfd, iov, len, iov_offset, 1);
+}
+
+static int do_read_write(int sockfd, void *buf, int len, int write)
+{
+	struct iovec iov;
+
+	iov.iov_base = buf;
+	iov.iov_len = len;
+
+	return do_readv_writev(sockfd, &iov, len, 0, write);
+}
+
+static int do_read(int sockfd, void *buf, int len)
+{
+	return do_read_write(sockfd, buf, len, 0);
+}
+
+static int do_write(int sockfd, void *buf, int len)
+{
+	return do_read_write(sockfd, buf, len, 1);
+}
+
+static int send_req(int sockfd, struct sd_req *hdr, void *data,
+		    unsigned int *wlen)
+{
+	int ret;
+	struct iovec iov[2];
+
+	iov[0].iov_base = hdr;
+	iov[0].iov_len = sizeof(*hdr);
+
+	if (*wlen) {
+		iov[1].iov_base = data;
+		iov[1].iov_len = *wlen;
+	}
+
+	ret = do_writev(sockfd, iov, sizeof(*hdr) + *wlen, 0);
+	if (ret) {
+		eprintf("failed to send a req, %m\n");
+		ret = -1;
+	}
+
+	return ret;
+}
+
+static int do_req(int sockfd, struct sd_req *hdr, void *data,
+		  unsigned int *wlen, unsigned int *rlen)
+{
+	int ret;
+
+	ret = send_req(sockfd, hdr, data, wlen);
+	if (ret) {
+		ret = -1;
+		goto out;
+	}
+
+	ret = do_read(sockfd, hdr, sizeof(*hdr));
+	if (ret) {
+		eprintf("failed to get a rsp, %m\n");
+		ret = -1;
+		goto out;
+	}
+
+	if (*rlen > hdr->data_length)
+		*rlen = hdr->data_length;
+
+	if (*rlen) {
+		ret = do_read(sockfd, data, *rlen);
+		if (ret) {
+			eprintf("failed to get the data, %m\n");
+			ret = -1;
+			goto out;
+		}
+	}
+	ret = 0;
+out:
+	return ret;
+}
+
+static int add_aio_request(struct bdrv_sd_state *s, struct aio_req *aio_req,
+			   struct iovec *iov, int niov, int create,
+			   enum aiocb_state aiocb_type);
+
+static void send_pending_req(struct bdrv_sd_state *s, uint64_t oid, uint32_t id)
+{
+	struct aio_req *aio_req, *next;
+	struct sd_aiocb *acb;
+	int ret;
+
+	QLIST_FOREACH_SAFE(aio_req, &s->pending_head, pending_siblings, next) {
+		if (id == get_id_from_req(s, aio_req))
+			continue;
+		if (aio_req->oid != oid)
+			continue;
+
+		acb = aio_req->aiocb;
+		ret = add_aio_request(s, aio_req, acb->qiov->iov,
+				      acb->qiov->niov, 0, acb->aiocb_type);
+		if (ret < 0) {
+			eprintf("add_aio_request is faled\n");
+			free_aio_req(s, aio_req);
+			if (QLIST_EMPTY(&acb->aioreq_head))
+				sd_finish_aiocb(acb);
+		}
+	}
+}
+
+static void aio_read_response(void *opaque)
+{
+	struct sd_obj_req hdr;
+	struct sd_obj_rsp *rsp = (struct sd_obj_rsp *)&hdr;
+	struct bdrv_sd_state *s = (struct bdrv_sd_state *)opaque;
+	int fd = s->fd;
+	int ret;
+	struct aio_req *aio_req;
+	struct sd_aiocb *acb;
+	int rest;
+	unsigned long idx;
+
+	if (!nr_outstanding_aio_req(s))
+		return;
+
+	ret = do_read(fd, (void *)rsp, sizeof(*rsp));
+	if (ret) {
+		eprintf("failed to get the header, %m\n");
+		return;
+	}
+
+	aio_req = get_req_from_id(s, rsp->id);
+	acb = aio_req->aiocb;
+
+	switch (acb->aiocb_type) {
+	case AIOCB_WRITE_UDATA:
+		if (!is_data_obj(aio_req->oid))
+			break;
+		idx = data_oid_to_idx(aio_req->oid);
+
+		if (s->inode.data_vdi_id[idx] != s->inode.vdi_id) {
+			s->inode.data_vdi_id[idx] = s->inode.vdi_id;
+			s->dirty_data_oids[s->nr_dirty_data_oids++] = idx;
+
+			send_pending_req(s, vid_to_data_oid(s->inode.vdi_id, idx),
+					 rsp->id);
+		}
+		break;
+	case AIOCB_READ_UDATA:
+		ret = do_readv(fd, acb->qiov->iov, rsp->data_length,
+			       aio_req->iov_offset);
+		if (ret) {
+			eprintf("failed to get the data, %m\n");
+			return;
+		}
+		break;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		acb->ret = -EIO;
+		eprintf("%s\n", sd_strerror(rsp->result));
+	}
+
+	rest = free_aio_req(s, aio_req);
+	if (!rest)
+		acb->aio_done_func(acb);
+}
+
+static int aio_flush_request(void *opaque)
+{
+	return nr_outstanding_aio_req((struct bdrv_sd_state *)opaque);
+}
+
+static int set_nonblocking(int fd)
+{
+	int ret;
+
+	ret = fcntl(fd, F_GETFL);
+	if (ret < 0) {
+		eprintf("can't fcntl (F_GETFL), %m\n");
+		close(fd);
+	} else {
+		ret = fcntl(fd, F_SETFL, ret | O_NONBLOCK);
+		if (ret < 0)
+			eprintf("can't fcntl (O_NONBLOCK), %m\n");
+	}
+
+	return ret;
+}
+
+static int set_nodelay(int fd)
+{
+	int ret, opt;
+
+	opt = 1;
+	ret = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &opt, sizeof(opt));
+	return ret;
+}
+
+static int get_sheep_fd(struct bdrv_sd_state *s)
+{
+	int ret, fd;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("%m\n");
+		return -1;
+	}
+
+	ret = set_nonblocking(fd);
+	if (ret) {
+		eprintf("%m\n");
+		close(fd);
+		return -1;
+	}
+
+	ret = set_nodelay(fd);
+	if (ret) {
+		eprintf("%m\n");
+		close(fd);
+		return -1;
+	}
+
+	qemu_aio_set_fd_handler(fd, aio_read_response, NULL, aio_flush_request,
+				NULL, s);
+	s->fd = fd;
+
+	return fd;
+}
+
+static int parse_vdiname(struct bdrv_sd_state *s, const char *filename,
+			 char *vdi, int vdi_len, uint32_t *snapid)
+{
+	char *p, *q;
+	int nr_sep;
+
+	strstart(filename, "sheepdog:", (const char **)&filename);
+
+	p = q = strdup(filename);
+
+	if (!p)
+		return 1;
+
+	nr_sep = 0;
+	while (*p) {
+		if (*p == ':')
+			nr_sep++;
+		if (nr_sep == 2)
+			break;
+		p++;
+	}
+
+	if (nr_sep == 2)
+		*p++ = '\0';
+	else
+		p = q;
+
+	strncpy(vdi, p, vdi_len);
+
+	p = strchr(vdi, ':');
+	if (p) {
+		*p++ = '\0';
+		*snapid = strtol(p, NULL, 10);
+	} else
+		*snapid = CURRENT_VDI_ID; /* search current vdi */
+
+	if (nr_sep == 2)
+		s->addr = q;
+	else {
+		free(q);
+		s->addr = NULL;
+	}
+
+	return 0;
+}
+
+static int find_vdi_name(struct bdrv_sd_state *s, char *filename, uint32_t snapid,
+			 uint32_t *vid, int for_snapshot)
+{
+	int ret, fd;
+	struct sd_vdi_req hdr;
+	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
+	unsigned int wlen, rlen = 0;
+	char buf[SD_MAX_VDI_LEN];
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0)
+		return -1;
+
+	memset(&hdr, 0, sizeof(hdr));
+	snprintf(buf, sizeof(buf), "%s", filename);
+	if (for_snapshot)
+		hdr.opcode = SD_OP_GET_VDI_INFO;
+	else
+		hdr.opcode = SD_OP_LOCK_VDI;
+	wlen = SD_MAX_VDI_LEN;
+	hdr.proto_ver = SD_PROTO_VER;
+	hdr.data_length = SD_MAX_VDI_LEN;
+	hdr.snapid = snapid;
+	hdr.flags = SD_FLAG_CMD_WRITE;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
+	if (ret) {
+		ret = -1;
+		goto out;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		eprintf("%s, %s\n", sd_strerror(rsp->result), filename);
+		ret = -1;
+		goto out;
+	}
+	*vid = rsp->vdi_id;
+
+	ret = 0;
+out:
+	close(fd);
+	return ret;
+}
+
+static int add_aio_request(struct bdrv_sd_state *s, struct aio_req *aio_req,
+			   struct iovec *iov, int niov, int create,
+			   enum aiocb_state aiocb_type)
+{
+	int nr_copies = s->inode.nr_copies;
+	struct sd_obj_req hdr;
+	unsigned int wlen;
+	int ret, opt;
+	uint64_t oid = aio_req->oid;
+	unsigned int datalen = aio_req->data_len;
+	uint64_t offset = aio_req->offset;
+	uint8_t flags = aio_req->flags;
+	uint64_t old_oid = aio_req->base_oid;
+
+	if (!nr_copies)
+		eprintf("bug\n");
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	if (aiocb_type == AIOCB_READ_UDATA) {
+		wlen = 0;
+		hdr.opcode = SD_OP_READ_OBJ;
+		hdr.flags = flags;
+	} else if (create) {
+		wlen = datalen;
+		hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
+		hdr.flags = SD_FLAG_CMD_WRITE | flags;
+	} else {
+		wlen = datalen;
+		hdr.opcode = SD_OP_WRITE_OBJ;
+		hdr.flags = SD_FLAG_CMD_WRITE | flags;
+	}
+
+	hdr.oid = oid;
+	hdr.cow_oid = old_oid;
+	hdr.copies = s->inode.nr_copies;
+
+	hdr.data_length = datalen;
+	hdr.offset = offset;
+
+	hdr.id = get_id_from_req(s, aio_req);
+
+	opt = 1;
+	setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
+
+	ret = do_write(s->fd, &hdr, sizeof(hdr));
+	if (ret) {
+		eprintf("failed to send a req, %m\n");
+		return -EIO;
+	}
+
+	if (wlen) {
+		ret = do_writev(s->fd, iov, wlen, aio_req->iov_offset);
+		if (ret) {
+			eprintf("failed to send a data, %m\n");
+			return -EIO;
+		}
+	}
+        opt = 0;
+        setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
+
+	return 0;
+}
+
+static int read_write_object(int fd, char *buf, uint64_t oid, int copies,
+			     unsigned int datalen, uint64_t offset,
+			     int write, int create)
+{
+	struct sd_obj_req hdr;
+	struct sd_obj_rsp *rsp = (struct sd_obj_rsp *)&hdr;
+	unsigned int wlen, rlen;
+	int ret;
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	if (write) {
+		wlen = datalen;
+		rlen = 0;
+		hdr.flags = SD_FLAG_CMD_WRITE;
+		if (create)
+			hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
+		else
+			hdr.opcode = SD_OP_WRITE_OBJ;
+	} else {
+		wlen = 0;
+		rlen = datalen;
+		hdr.opcode = SD_OP_READ_OBJ;
+	}
+	hdr.oid = oid;
+	hdr.data_length = datalen;
+	hdr.offset = offset;
+	hdr.copies = copies;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
+	if (ret) {
+		eprintf("failed to send a request to the sheep\n");
+		return -1;
+	}
+
+	switch (rsp->result) {
+	case SD_RES_SUCCESS:
+		return 0;
+	default:
+		eprintf("%s\n", sd_strerror(rsp->result));
+		return -1;
+	}
+}
+
+static int read_object(int fd, char *buf, uint64_t oid, int copies,
+		       unsigned int datalen, uint64_t offset)
+{
+	return read_write_object(fd, buf, oid, copies, datalen, offset, 0, 0);
+}
+
+static int write_object(int fd, char *buf, uint64_t oid, int copies,
+			unsigned int datalen, uint64_t offset, int create)
+{
+	return read_write_object(fd, buf, oid, copies, datalen, offset, 1, create);
+}
+
+/* TODO: error cleanups */
+static int sd_open(BlockDriverState *bs, const char *filename, int flags)
+{
+	int ret, i, fd;
+	uint32_t vid = 0;
+	struct bdrv_sd_state *s = bs->opaque;
+	char vdi[256];
+	uint32_t snapid;
+	int for_snapshot = 0;
+	char *buf;
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	if (!buf) {
+		eprintf("Failed to allocate memory\n");
+		return -1;
+	}
+
+	for (i = 0; i < MAX_AIO_REQS; i++) {
+		s->aio_req_free[i] = &s->aio_req_list[i];
+		s->aio_req_list[i].aiocb = NULL;
+	}
+	s->nr_aio_req_free = MAX_AIO_REQS;
+
+	if (strstart(filename, "sheepdog:", NULL))
+		for_snapshot = 1;
+
+	memset(vdi, 0, sizeof(vdi));
+	if (parse_vdiname(s, filename, vdi, sizeof(vdi), &snapid) < 0)
+		goto out;
+	s->fd = get_sheep_fd(s);
+	if (s->fd < 0)
+		return -1;
+
+	ret = find_vdi_name(s, vdi, snapid, &vid, for_snapshot);
+	if (ret)
+		goto out;
+
+	if (snapid)
+		dprintf("%" PRIx32 " non current inode was open.\n", vid);
+	else
+		s->is_current = 1;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), 0, SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret)
+		goto out;
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+	s->nr_dirty_data_oids = 0;
+
+	bs->total_sectors = s->inode.vdi_size >> 9;
+	strncpy(s->name, vdi, sizeof(s->name));
+	qemu_free(buf);
+
+	QLIST_INIT(&s->pending_head);
+	return 0;
+out:
+	qemu_free(buf);
+	return -1;
+}
+
+static int do_sd_create(const char *addr, char *filename, char *tag,
+			int64_t total_sectors, uint32_t base_vid,
+			uint32_t *vdi_id, int snapshot)
+{
+	struct sd_vdi_req hdr;
+	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
+	int fd, ret;
+	unsigned int wlen, rlen = 0;
+	char buf[SD_MAX_VDI_LEN];
+
+	fd = connect_to_sdog(addr);
+	if (fd < 0)
+		return -1;
+
+	strncpy(buf, filename, SD_MAX_VDI_LEN);
+
+	memset(&hdr, 0, sizeof(hdr));
+	hdr.opcode = SD_OP_NEW_VDI;
+	hdr.base_vdi_id = base_vid;
+
+	wlen = SD_MAX_VDI_LEN;
+
+	hdr.flags = SD_FLAG_CMD_WRITE;
+	hdr.snapid = snapshot;
+
+	hdr.data_length = wlen;
+	hdr.vdi_size = total_sectors * 512;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
+
+	close(fd);
+
+	if (ret)
+		return -1;
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		eprintf("%s, %s\n", sd_strerror(rsp->result), filename);
+		return -1;
+	}
+
+	if (vdi_id)
+		*vdi_id = rsp->vdi_id;
+
+	return 0;
+}
+
+static int sd_create(const char *filename, QEMUOptionParameter *options)
+{
+	int ret;
+	uint32_t vid = 0;
+	int64_t total_sectors = 0;
+	char *backing_file = NULL;
+
+	while (options && options->name) {
+		if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
+			total_sectors = options->value.n / 512;
+		} else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
+			backing_file = options->value.s;
+		}
+		options++;
+	}
+
+	if (backing_file) {
+		BlockDriverState bs;
+		char vdi[SD_MAX_VDI_LEN];
+		uint32_t snapid;
+
+		memset(&bs, 0, sizeof(bs));
+
+		bs.opaque = qemu_malloc(sizeof(struct bdrv_sd_state));
+		if (!bs.opaque)
+			return -1;
+
+		ret = sd_open(&bs, backing_file, 0);
+		if (ret < 0)
+			return -1;
+
+		if (parse_vdiname(bs.opaque, backing_file, vdi, sizeof(vdi), &snapid) < 0)
+			return -1;
+
+		/* cannot clone from a current inode */
+		if (snapid == CURRENT_VDI_ID)
+			return -1;
+
+		ret = find_vdi_name(bs.opaque, vdi, snapid, &vid, 1);
+		if (ret)
+			return -1;
+	}
+
+	return do_sd_create(NULL, (char *)filename, NULL, total_sectors, vid,
+			    NULL, 0);
+}
+
+static void sd_close(BlockDriverState *bs)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	struct sd_vdi_req hdr;
+	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
+	unsigned int wlen, rlen = 0;
+	int fd, ret;
+
+	dprintf("%s\n", s->name);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0)
+		return;
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	hdr.opcode = SD_OP_RELEASE_VDI;
+	wlen = strlen(s->name) + 1;
+	hdr.data_length = wlen;
+	hdr.flags = SD_FLAG_CMD_WRITE;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, s->name, &wlen, &rlen);
+
+	close(fd);
+
+	if (!ret && rsp->result != SD_RES_SUCCESS)
+		eprintf("%s, %s\n", sd_strerror(rsp->result), s->name);
+
+	close(s->fd);
+	free(s->addr);
+}
+
+static void sd_write_done(struct sd_aiocb *acb)
+{
+	int ret, i;
+	struct bdrv_sd_state *s = acb->common.bs->opaque;
+	struct iovec iov;
+	struct aio_req *aio_req;
+	uint32_t offset, data_len, mn, mx;
+
+	if (s->nr_dirty_data_oids) {
+		mn = mx = s->dirty_data_oids[0];
+		for (i = 0; i < s->nr_dirty_data_oids; i++) {
+			if (mn > s->dirty_data_oids[i])
+				mn = s->dirty_data_oids[i];
+			if (mx < s->dirty_data_oids[i])
+				mx = s->dirty_data_oids[i];
+		}
+		offset = sizeof(s->inode) - sizeof(s->inode.data_vdi_id) +
+			mn * sizeof(s->inode.data_vdi_id[0]);
+		data_len = (mx - mn + 1) * sizeof(s->inode.data_vdi_id[0]);
+		s->nr_dirty_data_oids = 0;
+
+		iov.iov_base = &s->inode;
+		iov.iov_len = sizeof(s->inode);
+		aio_req = alloc_aio_req(s, acb, vid_to_vdi_oid(s->inode.vdi_id),
+					data_len, offset, 0, 0, offset);
+		if (!aio_req) {
+			eprintf("too many requests\n");
+			acb->ret = -EIO;
+			goto out;
+		}
+		ret = add_aio_request(s, aio_req, &iov, 1, 0, AIOCB_WRITE_UDATA);
+		if (ret) {
+			free_aio_req(s, aio_req);
+			acb->ret = -EIO;
+			goto out;
+		}
+
+		acb->aio_done_func = sd_finish_aiocb;
+		acb->aiocb_type = AIOCB_WRITE_UDATA;
+		return;
+	}
+out:
+	sd_finish_aiocb(acb);
+}
+
+static int sd_create_branch(struct bdrv_sd_state *s)
+{
+	int ret, fd;
+	uint32_t vid;
+	char *buf;
+
+	dprintf("%" PRIx32 " is not current.\n", s->inode.vdi_id);
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	if (!buf)
+		return -1;
+
+	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
+			   s->inode.vdi_id, &vid, 1);
+	if (ret)
+		goto out;
+
+	dprintf("%" PRIx32 " is created.\n", vid);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
+			  SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret < 0)
+		goto out;
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+
+	s->is_current = 1;
+	ret = 0;
+	dprintf("%" PRIx32 " was newly created.\n", s->inode.vdi_id);
+
+out:
+	qemu_free(buf);
+
+	return ret;
+}
+
+static void sd_readv_writev_bh_cb(void *p)
+{
+	struct sd_aiocb *acb = p;
+	int ret = 0;
+	unsigned long len, done = 0, total = acb->nb_sectors * 512;
+	unsigned long idx = acb->sector_num * 512 / SD_DATA_OBJ_SIZE;
+	uint64_t oid;
+	uint64_t offset = (acb->sector_num * 512) % SD_DATA_OBJ_SIZE;
+	struct bdrv_sd_state *s = acb->common.bs->opaque;
+	struct sd_inode *inode = &s->inode;
+	struct aio_req *aio_req;
+
+	qemu_bh_delete(acb->bh);
+	acb->bh = NULL;
+
+	if (acb->aiocb_type == AIOCB_WRITE_UDATA && !s->is_current) {
+		ret = sd_create_branch(s);
+		if (ret) {
+			acb->ret = -EIO;
+			goto out;
+		}
+	}
+
+	while (done != total) {
+		uint8_t flags = 0;
+		uint64_t old_oid = 0;
+		int create = 0;
+
+		oid = vid_to_data_oid(inode->data_vdi_id[idx], idx);
+
+		len = min_t(unsigned long, total - done, SD_DATA_OBJ_SIZE - offset);
+
+		if (!inode->data_vdi_id[idx]) {
+			if (acb->aiocb_type == AIOCB_READ_UDATA)
+				goto done;
+
+			create = 1;
+		} else if (acb->aiocb_type == AIOCB_WRITE_UDATA
+			   && !is_data_obj_writeable(inode, idx)) {
+			create = 1;
+			old_oid = oid;
+			flags = SD_FLAG_CMD_COW;
+		}
+
+		if (create) {
+			dprintf("update ino (%" PRIu32") %"
+				PRIu64 " %" PRIu64 " %" PRIu64 "\n",
+				inode->vdi_id, oid,
+				vid_to_data_oid(inode->data_vdi_id[idx], idx), idx);
+			oid = vid_to_data_oid(inode->vdi_id, idx);
+			dprintf("new oid %lx\n", oid);
+		}
+
+		aio_req = alloc_aio_req(s, acb, oid, len, offset, flags,
+					old_oid, done);
+		if (!aio_req) {
+			eprintf("too many requests\n");
+			acb->ret = -EIO;
+			goto out;
+		}
+
+		if (create) {
+			struct aio_req *areq;
+			QLIST_FOREACH(areq, &s->pending_head, pending_siblings) {
+				if (get_id_from_req(s, areq) == get_id_from_req(s, aio_req))
+					continue;
+				if (areq->oid == oid) {
+					aio_req->flags = 0;
+					aio_req->base_oid = 0;
+					goto done;
+				}
+			}
+		}
+
+		ret = add_aio_request(s, aio_req, acb->qiov->iov, acb->qiov->niov,
+				      create, acb->aiocb_type);
+		if (ret < 0) {
+			eprintf("add_aio_request is faled\n");
+			free_aio_req(s, aio_req);
+			acb->ret = -EIO;
+			goto out;
+		}
+	done:
+		offset = 0;
+		idx++;
+		done += len;
+	}
+out:
+	if (QLIST_EMPTY(&acb->aioreq_head))
+		sd_finish_aiocb(acb);
+}
+
+static BlockDriverAIOCB *sd_aio_writev(BlockDriverState *bs,
+				       int64_t sector_num,
+				       QEMUIOVector *qiov,
+				       int nb_sectors,
+				       BlockDriverCompletionFunc *cb,
+				       void *opaque)
+{
+	struct sd_aiocb *acb;
+
+	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
+	acb->aio_done_func = sd_write_done;
+	acb->aiocb_type = AIOCB_WRITE_UDATA;
+
+	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
+	return &acb->common;
+}
+
+static BlockDriverAIOCB *sd_aio_readv(BlockDriverState *bs,
+				      int64_t sector_num,
+				      QEMUIOVector *qiov,
+				      int nb_sectors,
+				      BlockDriverCompletionFunc *cb,
+				      void *opaque)
+{
+	struct sd_aiocb *acb;
+	int i;
+
+	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
+	acb->aiocb_type = AIOCB_READ_UDATA;
+	acb->aio_done_func = sd_finish_aiocb;
+
+	/*
+	 * TODO: we can do better; we don't need to initialize
+	 * blindly.
+	 */
+	for (i = 0; i < qiov->niov; i++)
+		memset(qiov->iov[i].iov_base, 0, qiov->iov[i].iov_len);
+
+	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
+	return &acb->common;
+}
+
+static int sd_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	int ret, fd;
+	uint32_t new_vid;
+	struct sd_inode *inode;
+	unsigned int datalen;
+	uint64_t offset;
+
+	dprintf("sn_info: name %s id_str %s s: name %s vm_state_size %d "
+		"is_current %d\n", sn_info->name, sn_info->id_str,
+		s->name, sn_info->vm_state_size, s->is_current);
+
+	if (!s->is_current) {
+		eprintf("You can't create a snapshot of "
+			"a non current VDI, %s (%" PRIu32 ").\n",
+			s->name, s->inode.vdi_id);
+
+		return -1;
+	}
+
+	dprintf("%s %s\n", sn_info->name, sn_info->id_str);
+
+	s->inode.vm_state_size = sn_info->vm_state_size;
+	s->inode.vm_clock_nsec = sn_info->vm_clock_nsec;
+	offset = 0;
+	/* we don't need to read entire object */
+	datalen = SD_INODE_SIZE - sizeof(s->inode.data_vdi_id);
+
+	/* refresh inode. */
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	ret = write_object(fd, (char *)&s->inode, vid_to_vdi_oid(s->inode.vdi_id),
+			   s->inode.nr_copies, datalen, offset, 0);
+	if (ret < 0) {
+		eprintf("failed to write snapshot's inode.\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
+			   s->inode.vdi_id, &new_vid, 1);
+	if (ret < 0) {
+		eprintf("failed to create inode for snapshot. %m\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	inode = (struct sd_inode *)qemu_malloc(datalen);
+	if (!inode) {
+		eprintf("failed to allocate memory for inode. %m\n");
+		goto cleanup;
+	}
+
+	ret = read_object(fd, (char *)inode, vid_to_vdi_oid(new_vid),
+			  s->inode.nr_copies, datalen, offset);
+
+	close(fd);
+
+	if (ret < 0) {
+		eprintf("failed to read new inode info. %m\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	memcpy(&s->inode, inode, datalen);
+	dprintf("s->inode: name %s snap_id %x oid %x\n",
+		s->inode.name, s->inode.snap_id, s->inode.vdi_id);
+
+cleanup:
+	close(fd);
+	return ret;
+}
+
+static int sd_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	struct bdrv_sd_state *old_s;
+	char vdi[SD_MAX_VDI_LEN];
+	char *buf = NULL;
+	uint32_t vid;
+	uint32_t snapid = 0;
+	int ret = -ENOENT, fd;
+
+	old_s = qemu_malloc(sizeof(struct bdrv_sd_state));
+	if (!old_s) {
+		eprintf("failed to allocate memory for old state. %m\n");
+		goto out;
+	}
+
+	memcpy(old_s, s, sizeof(struct bdrv_sd_state));
+
+	snapid = strtol(snapshot_id, NULL, 10);
+	if (!snapid) {
+		eprintf("Invalid snapshot_id\n");
+		goto out;
+	}
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	if (!buf) {
+		eprintf("Failed to allocate memory\n");
+		goto out;
+	}
+	strncpy(vdi, s->name, sizeof(vdi));
+	ret = find_vdi_name(s, vdi, snapid, &vid, 1);
+	if (ret) {
+		eprintf("Failed to find_vdi_name\n");
+		ret = -ENOENT;
+		goto out;
+	}
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
+			  SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret) {
+		ret = -ENOENT;
+		goto out;
+	}
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+
+	if (!s->inode.vm_state_size) {
+		eprintf("Invalid snapshot\n");
+		ret = -ENOENT;
+		goto out;
+	}
+
+	s->is_current = 0;
+
+	qemu_free(buf);
+	qemu_free(old_s);
+
+	return 0;
+out:
+	/* recover bdrv_sd_state */
+	memcpy(s, old_s, sizeof(struct bdrv_sd_state));
+	qemu_free(buf);
+	qemu_free(old_s);
+
+	eprintf("failed to open. recover old bdrv_sd_state.\n");
+
+	return ret;
+}
+
+static int sd_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
+{
+	/* FIXME: Delete specified snapshot id.  */
+	return 0;
+}
+
+#define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
+#define BITS_PER_BYTE		8
+#define BITS_TO_LONGS(nr)	DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(long))
+#define DECLARE_BITMAP(name,bits) \
+	unsigned long name[BITS_TO_LONGS(bits)]
+
+#define BITS_PER_LONG (BITS_PER_BYTE * sizeof(long))
+
+static inline int test_bit(unsigned int nr, const unsigned long *addr)
+{
+	return ((1UL << (nr % BITS_PER_LONG)) &
+		(((unsigned long *)addr)[nr / BITS_PER_LONG])) != 0;
+}
+
+static int sd_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	struct sd_req req;
+	int i, fd, nr = 1024, ret, max = BITS_TO_LONGS(SD_NR_VDIS) * sizeof(long);
+	QEMUSnapshotInfo *sn_tab = NULL;
+	unsigned wlen, rlen;
+	int found = 0;
+	static struct sd_inode inode;
+	unsigned long *vdi_inuse;
+	unsigned int start_nr;
+
+	vdi_inuse = qemu_malloc(max);
+	if (!vdi_inuse)
+		return 0;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0)
+		goto out;
+
+	rlen = max;
+	wlen = 0;
+
+	memset(&req, 0, sizeof(req));
+
+	req.opcode = SD_OP_READ_VDIS;
+	req.data_length = max;
+
+	ret = do_req(fd, (struct sd_req *)&req, vdi_inuse, &wlen, &rlen);
+
+	close(fd);
+	if (ret)
+		goto out;
+
+	sn_tab = qemu_mallocz(nr * sizeof(*sn_tab));
+	if (!sn_tab)
+		goto out;
+
+	start_nr = fnv_64a_buf(s->name, strlen(s->name), FNV1A_64_INIT) & (SD_NR_VDIS - 1);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	/* TODO: round up */
+	for (i = start_nr; i < SD_NR_VDIS && found < nr; i++) {
+		if (!test_bit(i, vdi_inuse))
+			break;
+
+		/* we don't need to read entire object */
+		ret = read_object(fd, (char *)&inode, vid_to_vdi_oid(i),
+				  0, SD_INODE_SIZE - sizeof(inode.data_vdi_id), 0);
+
+		if (ret)
+			continue;
+
+		if (!strcmp(inode.name, s->name) && inode.snap_ctime) {
+			sn_tab[found].date_sec = inode.snap_ctime >> 32;
+			sn_tab[found].date_nsec = inode.snap_ctime & 0xffffffff;
+			sn_tab[found].vm_state_size = inode.vm_state_size;
+			sn_tab[found].vm_clock_nsec = inode.vm_clock_nsec;
+
+			snprintf(sn_tab[found].id_str, sizeof(sn_tab[found].id_str), "%u",
+				 inode.snap_id);
+			found++;
+		}
+	}
+
+	close(fd);
+out:
+	*psn_tab = sn_tab;
+
+	qemu_free(vdi_inuse);
+
+	return found;
+}
+
+static int do_load_save_vmstate(struct bdrv_sd_state *s, uint8_t *data,
+				int64_t pos, int size, int load)
+{
+	int fd, create;
+	int ret = 0;
+	unsigned int data_len;
+	uint64_t vmstate_oid;
+	uint32_t vdi_index;
+	uint64_t offset;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	while (size) {
+		vdi_index = pos / SD_DATA_OBJ_SIZE;
+		offset = pos % SD_DATA_OBJ_SIZE;
+
+		data_len = min_t(unsigned int, size, SD_DATA_OBJ_SIZE);
+
+		vmstate_oid = vid_to_vmstate_oid(s->inode.vdi_id, vdi_index);
+
+		create = (offset == 0);
+		if (load)
+			ret = read_object(fd, (char *)data, vmstate_oid,
+					  s->inode.nr_copies, data_len, offset);
+		else
+			ret = write_object(fd, (char *)data, vmstate_oid,
+					   s->inode.nr_copies, data_len, offset, create);
+
+		if (ret < 0) {
+			eprintf("failed to save vmstate %m\n");
+			ret = -EIO;
+			goto cleanup;
+		}
+
+		pos += data_len;
+		size -= data_len;
+		ret += data_len;
+	}
+cleanup:
+	close(fd);
+	return ret;
+}
+
+static int sd_save_vmstate(BlockDriverState *bs, const uint8_t *data,
+			   int64_t pos, int size)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+
+	return do_load_save_vmstate(s, (uint8_t *)data, pos, size, 0);
+}
+
+static int sd_load_vmstate(BlockDriverState *bs, uint8_t *data,
+			   int64_t pos, int size)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+
+	return do_load_save_vmstate(s, data, pos, size, 1);
+}
+
+
+static QEMUOptionParameter sd_create_options[] = {
+	{
+		.name = BLOCK_OPT_SIZE,
+		.type = OPT_SIZE,
+		.help = "Virtual disk size"
+	},
+	{
+		.name = BLOCK_OPT_BACKING_FILE,
+		.type = OPT_STRING,
+		.help = "File name of a base image"
+	},
+	{ NULL }
+};
+
+BlockDriver bdrv_sheepdog = {
+	.format_name = "sheepdog",
+	.protocol_name = "sheepdog",
+	.instance_size = sizeof(struct bdrv_sd_state),
+	.bdrv_open = sd_open,
+	.bdrv_close = sd_close,
+	.bdrv_create = sd_create,
+
+	.bdrv_aio_readv = sd_aio_readv,
+	.bdrv_aio_writev = sd_aio_writev,
+
+	.bdrv_snapshot_create = sd_snapshot_create,
+	.bdrv_snapshot_goto = sd_snapshot_goto,
+	.bdrv_snapshot_delete   = sd_snapshot_delete,
+	.bdrv_snapshot_list = sd_snapshot_list,
+
+	.bdrv_save_vmstate    = sd_save_vmstate,
+	.bdrv_load_vmstate    = sd_load_vmstate,
+
+	.create_options = sd_create_options,
+
+};
+
+static void bdrv_sheepdog_init(void)
+{
+	bdrv_register(&bdrv_sheepdog);
+}
+block_init(bdrv_sheepdog_init);
-- 
1.5.6.5

Re: [Qemu-devel] [RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[Kevin Wolf]

Am 12.05.2010 12:46, schrieb MORITA Kazutaka:
> Hi all,
> 
> This patch adds a block driver for Sheepdog distributed storage
> system.  Please consider for inclusion.
> 
> Sheepdog is a distributed storage system for QEMU.  It provides highly
> available block level storage volumes to VMs like Amazon EBS.
> 
> Sheepdog features are:
> - No node in the cluster is special (no metadata node, no control
>   node, etc)
> - Linear scalability in performance and capacity
> - No single point of failure
> - Autonomous management (zero configuration)
> - Useful volume management support such as snapshot and cloning
> - Thin provisioning
> - Autonomous load balancing
> 
> The more details are available at the project site [1] and my previous
> post about sheepdog [2].
> 
> We have implemented the essential part of sheepdog features, and
> believe the API between Sheepdog and QEMU is finalized.
> 
> Any comments or suggestions would be greatly appreciated.

These patches don't apply, neither on git master nor on the block
branch. Please rebase them on git://repo.or.cz/qemu/kevin.git block for
the next submission.

I'll have a closer look at your code later, but one thing I noticed is
that the new block driver is something in between a protocol and a
format driver (just like vvfat, which should stop doing so, too). I
think it ought to be a real protocol with the raw format driver on top
(or any other format - I don't see a reason why this should be
restricted to raw).

The one thing that is unusual about it as a protocol driver is that it
supports snapshots. However, while it is the first one, supporting
snapshots in protocols is a thing that could be generally useful to
support (for example thinking of a LVM protocol, which was discussed in
the past).

So in block.c we could check if the format driver supports snapshots,
and if it doesn't we try again with the underlying protocol. Not sure
yet what we would do when both format and protocol do support snapshots
(qcow2 on sheepdog/LVM/...), but that's a detail.

Kevin

Re: [Qemu-devel] [RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[MORITA Kazutaka]

On 2010/05/12 20:38, Kevin Wolf wrote:
> Am 12.05.2010 12:46, schrieb MORITA Kazutaka:
>> Hi all,
>>
>> This patch adds a block driver for Sheepdog distributed storage
>> system.  Please consider for inclusion.
>>
>> Sheepdog is a distributed storage system for QEMU.  It provides highly
>> available block level storage volumes to VMs like Amazon EBS.
>>
>> Sheepdog features are:
>> - No node in the cluster is special (no metadata node, no control
>>   node, etc)
>> - Linear scalability in performance and capacity
>> - No single point of failure
>> - Autonomous management (zero configuration)
>> - Useful volume management support such as snapshot and cloning
>> - Thin provisioning
>> - Autonomous load balancing
>>
>> The more details are available at the project site [1] and my previous
>> post about sheepdog [2].
>>
>> We have implemented the essential part of sheepdog features, and
>> believe the API between Sheepdog and QEMU is finalized.
>>
>> Any comments or suggestions would be greatly appreciated.
> 
> These patches don't apply, neither on git master nor on the block
> branch. Please rebase them on git://repo.or.cz/qemu/kevin.git block for
> the next submission.
> 

Ok, I'll rebase them and resend later. Sorry for inconvenience.

> I'll have a closer look at your code later, but one thing I noticed is
> that the new block driver is something in between a protocol and a
> format driver (just like vvfat, which should stop doing so, too). I
> think it ought to be a real protocol with the raw format driver on top
> (or any other format - I don't see a reason why this should be
> restricted to raw).
> 
> The one thing that is unusual about it as a protocol driver is that it
> supports snapshots. However, while it is the first one, supporting
> snapshots in protocols is a thing that could be generally useful to
> support (for example thinking of a LVM protocol, which was discussed in
> the past).
> 

I agreed.  I'll modify the sheepdog driver patch as a protocol driver one,
and remove unnecessary format check from my patch.

> So in block.c we could check if the format driver supports snapshots,
> and if it doesn't we try again with the underlying protocol. Not sure
> yet what we would do when both format and protocol do support snapshots
> (qcow2 on sheepdog/LVM/...), but that's a detail.
> 

Thanks,

Kazutaka

Re: [Qemu-devel] [RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[MORITA Kazutaka]

At Thu, 13 May 2010 04:46:46 +0900,
MORITA Kazutaka wrote:
> 
> On 2010/05/12 20:38, Kevin Wolf wrote:
> > I'll have a closer look at your code later, but one thing I noticed is
> > that the new block driver is something in between a protocol and a
> > format driver (just like vvfat, which should stop doing so, too). I
> > think it ought to be a real protocol with the raw format driver on top
> > (or any other format - I don't see a reason why this should be
> > restricted to raw).
> > 
> > The one thing that is unusual about it as a protocol driver is that it
> > supports snapshots. However, while it is the first one, supporting
> > snapshots in protocols is a thing that could be generally useful to
> > support (for example thinking of a LVM protocol, which was discussed in
> > the past).
> > 
> 
> I agreed.  I'll modify the sheepdog driver patch as a protocol driver one,
> and remove unnecessary format check from my patch.
> 
> > So in block.c we could check if the format driver supports snapshots,
> > and if it doesn't we try again with the underlying protocol. Not sure
> > yet what we would do when both format and protocol do support snapshots
> > (qcow2 on sheepdog/LVM/...), but that's a detail.
> > 
> 

To support snapshot in a protocol, I'd like to call the hander of the
protocol driver in the following functions in block.c:

    bdrv_snapshot_create
    bdrv_snapshot_goto
    bdrv_snapshot_delete
    bdrv_snapshot_list
    bdrv_save_vmstate
    bdrv_load_vmstate

Is it okay?

In the case both format and protocol drivers support snapshots, I
think it is better to call the format driver handler.  Because qcow2
is well known as a snapshot support format, so when users use qcow2,
they expect to get snapshot with qcow2.

There is another problem to make the sheepdog driver be a protocol;
how to deal with protocol specific create_options?

For example, sheepdog supports cloning images as a format driver:

  $ qemu-img create -f sheepdog dst -b sheepdog:src

But if the sheepdog driver is a protocol, error will occur.

  $ qemu-img create sheepdog:dst -b sheepdog:src
  Unknown option 'backing_file'
  qemu-img: Backing file not supported for file format 'raw'

It is because the raw format doesn't support a backing_file option.
To support the protocol specific create_options, if the format driver
cannot parse some of the arguments, the protocol driver need to parse
them.

If my suggestions are okay, I'd like to prepare the patches.

Regards,

Kazutaka

Re: [Qemu-devel] [RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[Kevin Wolf]

Am 13.05.2010 16:03, schrieb MORITA Kazutaka:
> To support snapshot in a protocol, I'd like to call the hander of the
> protocol driver in the following functions in block.c:
> 
>     bdrv_snapshot_create
>     bdrv_snapshot_goto
>     bdrv_snapshot_delete
>     bdrv_snapshot_list
>     bdrv_save_vmstate
>     bdrv_load_vmstate
> 
> Is it okay?

Yes, I think this is the way to go.

> In the case both format and protocol drivers support snapshots, I
> think it is better to call the format driver handler.  Because qcow2
> is well known as a snapshot support format, so when users use qcow2,
> they expect to get snapshot with qcow2.

I agree.

> There is another problem to make the sheepdog driver be a protocol;
> how to deal with protocol specific create_options?
> 
> For example, sheepdog supports cloning images as a format driver:
> 
>   $ qemu-img create -f sheepdog dst -b sheepdog:src
> 
> But if the sheepdog driver is a protocol, error will occur.
> 
>   $ qemu-img create sheepdog:dst -b sheepdog:src
>   Unknown option 'backing_file'
>   qemu-img: Backing file not supported for file format 'raw'
> 
> It is because the raw format doesn't support a backing_file option.
> To support the protocol specific create_options, if the format driver
> cannot parse some of the arguments, the protocol driver need to parse
> them.

That's actually a good point. Yes, I think it makes a lot of sense to
allow parameters to be passed to the protocol driver.

Also, I've never tried to create an image over a protocol other than
file. As far as I know, raw is the only format for which it should work
right now (at least in theory). As we're going forward, I'm planning to
convert the other drivers, too.

Kevin

Re: [Qemu-devel] [RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[MORITA Kazutaka]

At Fri, 14 May 2010 10:32:26 +0200,
Kevin Wolf wrote:
> 
> Am 13.05.2010 16:03, schrieb MORITA Kazutaka:
> > To support snapshot in a protocol, I'd like to call the hander of the
> > protocol driver in the following functions in block.c:
> > 
> >     bdrv_snapshot_create
> >     bdrv_snapshot_goto
> >     bdrv_snapshot_delete
> >     bdrv_snapshot_list
> >     bdrv_save_vmstate
> >     bdrv_load_vmstate
> > 
> > Is it okay?
> 
> Yes, I think this is the way to go.
> 
Done.

> > In the case both format and protocol drivers support snapshots, I
> > think it is better to call the format driver handler.  Because qcow2
> > is well known as a snapshot support format, so when users use qcow2,
> > they expect to get snapshot with qcow2.
> 
> I agree.
> 
Done.

> > There is another problem to make the sheepdog driver be a protocol;
> > how to deal with protocol specific create_options?
> > 
> > For example, sheepdog supports cloning images as a format driver:
> > 
> >   $ qemu-img create -f sheepdog dst -b sheepdog:src
> > 
> > But if the sheepdog driver is a protocol, error will occur.
> > 
> >   $ qemu-img create sheepdog:dst -b sheepdog:src
> >   Unknown option 'backing_file'
> >   qemu-img: Backing file not supported for file format 'raw'
> > 
> > It is because the raw format doesn't support a backing_file option.
> > To support the protocol specific create_options, if the format driver
> > cannot parse some of the arguments, the protocol driver need to parse
> > them.
> 
> That's actually a good point. Yes, I think it makes a lot of sense to
> allow parameters to be passed to the protocol driver.
> 

Okay. But it seemed to require many changes to the qemu-img parser, so I didn't
do it in the patchset I sent just now.

> Also, I've never tried to create an image over a protocol other than
> file. As far as I know, raw is the only format for which it should work
> right now (at least in theory). As we're going forward, I'm planning to
> convert the other drivers, too.
> 

I see. Thank you for the explanations.


Regards,

Kazutaka

Re: [Qemu-devel] [RFC PATCH 0/2] Sheepdog: distributed storage system for QEMU

[Kevin Wolf]

Am 14.05.2010 11:54, schrieb MORITA Kazutaka:
>>> There is another problem to make the sheepdog driver be a protocol;
>>> how to deal with protocol specific create_options?
>>>
>>> For example, sheepdog supports cloning images as a format driver:
>>>
>>>   $ qemu-img create -f sheepdog dst -b sheepdog:src
>>>
>>> But if the sheepdog driver is a protocol, error will occur.
>>>
>>>   $ qemu-img create sheepdog:dst -b sheepdog:src
>>>   Unknown option 'backing_file'
>>>   qemu-img: Backing file not supported for file format 'raw'
>>>
>>> It is because the raw format doesn't support a backing_file option.
>>> To support the protocol specific create_options, if the format driver
>>> cannot parse some of the arguments, the protocol driver need to parse
>>> them.
>>
>> That's actually a good point. Yes, I think it makes a lot of sense to
>> allow parameters to be passed to the protocol driver.
>>
> 
> Okay. But it seemed to require many changes to the qemu-img parser, so I didn't
> do it in the patchset I sent just now.

Yes, it probably requires a bit more than just five lines. We should
have it in the final version of the patch series, but we seem to agree
there and image creation is separate enough, so we can ignore it for the
discussion of the other parts.

Kevin

[RFC PATCH v2 1/3] close all the block drivers before the qemu process exits

[MORITA Kazutaka]

This patch calls the close handler of the block driver before the qemu
process exits.

This is necessary because the sheepdog block driver releases the lock
of VM images in the close handler.

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 block.c |    9 +++++++++
 block.h |    1 +
 vl.c    |    1 +
 3 files changed, 11 insertions(+), 0 deletions(-)

diff --git a/block.c b/block.c
index c134c2b..988a94a 100644
--- a/block.c
+++ b/block.c
@@ -641,6 +641,15 @@ void bdrv_close(BlockDriverState *bs)
     }
 }
 
+void bdrv_close_all(void)
+{
+    BlockDriverState *bs;
+
+    QTAILQ_FOREACH(bs, &bdrv_states, list) {
+        bdrv_close(bs);
+    }
+}
+
 void bdrv_delete(BlockDriverState *bs)
 {
     /* remove from list, if necessary */
diff --git a/block.h b/block.h
index 278259c..531e802 100644
--- a/block.h
+++ b/block.h
@@ -121,6 +121,7 @@ BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
 /* Ensure contents are flushed to disk.  */
 void bdrv_flush(BlockDriverState *bs);
 void bdrv_flush_all(void);
+void bdrv_close_all(void);
 
 int bdrv_has_zero_init(BlockDriverState *bs);
 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
diff --git a/vl.c b/vl.c
index 85bcc84..5ce7807 100644
--- a/vl.c
+++ b/vl.c
@@ -2007,6 +2007,7 @@ static void main_loop(void)
             exit(0);
         }
     }
+    bdrv_close_all();
     pause_all_vcpus();
 }
 
-- 
1.5.6.5

[RFC PATCH v2 2/3] block: call the snapshot handlers of the protocol drivers

[MORITA Kazutaka]

When snapshot handlers of the format driver is not defined, it is
better to call the ones of the protocol driver.

This enables us to implement snapshot support in the protocol driver.

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 block.c |   48 ++++++++++++++++++++++++++++++------------------
 1 files changed, 30 insertions(+), 18 deletions(-)

diff --git a/block.c b/block.c
index 988a94a..d1866be 100644
--- a/block.c
+++ b/block.c
@@ -1689,9 +1689,11 @@ int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_save_vmstate)
-        return -ENOTSUP;
-    return drv->bdrv_save_vmstate(bs, buf, pos, size);
+    if (drv->bdrv_save_vmstate)
+        return drv->bdrv_save_vmstate(bs, buf, pos, size);
+    if (bs->file)
+        return bdrv_save_vmstate(bs->file, buf, pos, size);
+    return -ENOTSUP;
 }
 
 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
@@ -1700,9 +1702,11 @@ int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_load_vmstate)
-        return -ENOTSUP;
-    return drv->bdrv_load_vmstate(bs, buf, pos, size);
+    if (drv->bdrv_load_vmstate)
+        return drv->bdrv_load_vmstate(bs, buf, pos, size);
+    if (bs->file)
+        return bdrv_load_vmstate(bs->file, buf, pos, size);
+    return -ENOTSUP;
 }
 
 void bdrv_debug_event(BlockDriverState *bs, BlkDebugEvent event)
@@ -1726,9 +1730,11 @@ int bdrv_snapshot_create(BlockDriverState *bs,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_create)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_create(bs, sn_info);
+    if (drv->bdrv_snapshot_create)
+        return drv->bdrv_snapshot_create(bs, sn_info);
+    if (bs->file)
+        return bdrv_snapshot_create(bs->file, sn_info);
+    return -ENOTSUP;
 }
 
 int bdrv_snapshot_goto(BlockDriverState *bs,
@@ -1737,9 +1743,11 @@ int bdrv_snapshot_goto(BlockDriverState *bs,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_goto)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_goto(bs, snapshot_id);
+    if (drv->bdrv_snapshot_goto)
+        return drv->bdrv_snapshot_goto(bs, snapshot_id);
+    if (bs->file)
+        return bdrv_snapshot_goto(bs->file, snapshot_id);
+    return -ENOTSUP;
 }
 
 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
@@ -1747,9 +1755,11 @@ int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_delete)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_delete(bs, snapshot_id);
+    if (drv->bdrv_snapshot_delete)
+        return drv->bdrv_snapshot_delete(bs, snapshot_id);
+    if (bs->file)
+        return bdrv_snapshot_delete(bs->file, snapshot_id);
+    return -ENOTSUP;
 }
 
 int bdrv_snapshot_list(BlockDriverState *bs,
@@ -1758,9 +1768,11 @@ int bdrv_snapshot_list(BlockDriverState *bs,
     BlockDriver *drv = bs->drv;
     if (!drv)
         return -ENOMEDIUM;
-    if (!drv->bdrv_snapshot_list)
-        return -ENOTSUP;
-    return drv->bdrv_snapshot_list(bs, psn_info);
+    if (drv->bdrv_snapshot_list)
+        return drv->bdrv_snapshot_list(bs, psn_info);
+    if (bs->file)
+        return bdrv_snapshot_list(bs->file, psn_info);
+    return -ENOTSUP;
 }
 
 #define NB_SUFFIXES 4
-- 
1.5.6.5

Re: [RFC PATCH v2 2/3] block: call the snapshot handlers of the protocol drivers

[Kevin Wolf]

Am 14.05.2010 11:51, schrieb MORITA Kazutaka:
> When snapshot handlers of the format driver is not defined, it is
> better to call the ones of the protocol driver.
> 
> This enables us to implement snapshot support in the protocol driver.
> 
> Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>

>  int bdrv_snapshot_goto(BlockDriverState *bs,
> @@ -1737,9 +1743,11 @@ int bdrv_snapshot_goto(BlockDriverState *bs,
>      BlockDriver *drv = bs->drv;
>      if (!drv)
>          return -ENOMEDIUM;
> -    if (!drv->bdrv_snapshot_goto)
> -        return -ENOTSUP;
> -    return drv->bdrv_snapshot_goto(bs, snapshot_id);
> +    if (drv->bdrv_snapshot_goto)
> +        return drv->bdrv_snapshot_goto(bs, snapshot_id);
> +    if (bs->file)
> +        return bdrv_snapshot_goto(bs->file, snapshot_id);
> +    return -ENOTSUP;
>  }

During lunch one more thing came to my mind...

For bdrv_snapshot_goto it's probably not that easy for formats other
than raw. Usually the drivers keep some state in memory, and with this
code they won't re-read it after the snapshot has been loaded - which
will lead to an inconsistent state very easily.

I think the right thing to do here is:

1. Call drv->bdrv_close() so that the format driver cleans up, but the
underlying protocol stays open
2. Load the snapshot on the protocol level
3. Call drv->bdrv_open() to load the new state

The other functions in this patch should be okay without re-opening
because they don't change the current state.

Kevin

[RFC PATCH v2 3/3] block: add sheepdog driver for distributed storage support

[MORITA Kazutaka]

Sheepdog is a distributed storage system for QEMU. It provides highly
available block level storage volumes to VMs like Amazon EBS.  This
patch adds a qemu block driver for Sheepdog.

Sheepdog features are:
- No node in the cluster is special (no metadata node, no control
  node, etc)
- Linear scalability in performance and capacity
- No single point of failure
- Autonomous management (zero configuration)
- Useful volume management support such as snapshot and cloning
- Thin provisioning
- Autonomous load balancing

The more details are available at the project site:
    http://www.osrg.net/sheepdog/

Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
---
 Makefile.objs    |    2 +-
 block/sheepdog.c | 1831 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 1832 insertions(+), 1 deletions(-)
 create mode 100644 block/sheepdog.c

diff --git a/Makefile.objs b/Makefile.objs
index ecdd53e..6edbc57 100644
--- a/Makefile.objs
+++ b/Makefile.objs
@@ -14,7 +14,7 @@ block-obj-$(CONFIG_LINUX_AIO) += linux-aio.o
 
 block-nested-y += raw.o cow.o qcow.o vdi.o vmdk.o cloop.o dmg.o bochs.o vpc.o vvfat.o
 block-nested-y += qcow2.o qcow2-refcount.o qcow2-cluster.o qcow2-snapshot.o
-block-nested-y += parallels.o nbd.o blkdebug.o
+block-nested-y += parallels.o nbd.o blkdebug.o sheepdog.o
 block-nested-$(CONFIG_WIN32) += raw-win32.o
 block-nested-$(CONFIG_POSIX) += raw-posix.o
 block-nested-$(CONFIG_CURL) += curl.o
diff --git a/block/sheepdog.c b/block/sheepdog.c
new file mode 100644
index 0000000..adf3a71
--- /dev/null
+++ b/block/sheepdog.c
@@ -0,0 +1,1831 @@
+/*
+ * Copyright (C) 2009-2010 Nippon Telegraph and Telephone Corporation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+#include <netdb.h>
+#include <netinet/tcp.h>
+
+#include "qemu-common.h"
+#include "block_int.h"
+
+#define SD_PROTO_VER 0x01
+
+#define SD_DEFAULT_ADDR "localhost:7000"
+
+#define SD_OP_CREATE_AND_WRITE_OBJ  0x01
+#define SD_OP_READ_OBJ       0x02
+#define SD_OP_WRITE_OBJ      0x03
+
+#define SD_OP_NEW_VDI        0x11
+#define SD_OP_LOCK_VDI       0x12
+#define SD_OP_RELEASE_VDI    0x13
+#define SD_OP_GET_VDI_INFO   0x14
+#define SD_OP_READ_VDIS      0x15
+
+#define SD_FLAG_CMD_WRITE    0x01
+#define SD_FLAG_CMD_COW      0x02
+
+#define SD_RES_SUCCESS       0x00 /* Success */
+#define SD_RES_UNKNOWN       0x01 /* Unknown error */
+#define SD_RES_NO_OBJ        0x02 /* No object found */
+#define SD_RES_EIO           0x03 /* I/O error */
+#define SD_RES_VDI_EXIST     0x04 /* Vdi exists already */
+#define SD_RES_INVALID_PARMS 0x05 /* Invalid parameters */
+#define SD_RES_SYSTEM_ERROR  0x06 /* System error */
+#define SD_RES_VDI_LOCKED    0x07 /* Vdi is locked */
+#define SD_RES_NO_VDI        0x08 /* No vdi found */
+#define SD_RES_NO_BASE_VDI   0x09 /* No base vdi found */
+#define SD_RES_VDI_READ      0x0A /* Cannot read requested vdi */
+#define SD_RES_VDI_WRITE     0x0B /* Cannot write requested vdi */
+#define SD_RES_BASE_VDI_READ 0x0C /* Cannot read base vdi */
+#define SD_RES_BASE_VDI_WRITE   0x0D /* Cannot write base vdi */
+#define SD_RES_NO_TAG        0x0E /* Requested tag is not found */
+#define SD_RES_STARTUP       0x0F /* Sheepdog is on starting up */
+#define SD_RES_VDI_NOT_LOCKED   0x10 /* Vdi is not locked */
+#define SD_RES_SHUTDOWN      0x11 /* Sheepdog is shutting down */
+#define SD_RES_NO_MEM        0x12 /* Cannot allocate memory */
+#define SD_RES_FULL_VDI      0x13 /* we already have the maximum vdis */
+#define SD_RES_VER_MISMATCH  0x14 /* Protocol version mismatch */
+#define SD_RES_NO_SPACE      0x15 /* Server has no room for new objects */
+#define SD_RES_WAIT_FOR_FORMAT  0x16 /* Sheepdog is waiting for a format operation */
+#define SD_RES_WAIT_FOR_JOIN    0x17 /* Sheepdog is waiting for other nodes joining */
+#define SD_RES_JOIN_FAILED   0x18 /* Target node had failed to join sheepdog */
+
+/*
+ * Object ID rules
+ *
+ *  0 - 19 (20 bits): data object space
+ * 20 - 31 (12 bits): reserved data object space
+ * 32 - 55 (24 bits): vdi object space
+ * 56 - 59 ( 4 bits): reserved vdi object space
+ * 60 - 63 ( 4 bits): object type indentifier space
+ */
+
+#define VDI_SPACE_SHIFT   32
+#define VDI_BIT (UINT64_C(1) << 63)
+#define VMSTATE_BIT (UINT64_C(1) << 62)
+#define MAX_DATA_OBJS (1ULL << 20)
+#define MAX_CHILDREN 1024
+#define SD_MAX_VDI_LEN 256
+#define SD_NR_VDIS   (1U << 24)
+#define SD_DATA_OBJ_SIZE (UINT64_C(1) << 22)
+
+#define SD_INODE_SIZE (sizeof(struct sd_inode))
+#define CURRENT_VDI_ID 0
+
+struct sd_req {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t	opcode_specific[8];
+};
+
+struct sd_rsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t	opcode_specific[7];
+};
+
+struct sd_obj_req {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint64_t        oid;
+	uint64_t        cow_oid;
+	uint32_t        copies;
+	uint32_t        rsvd;
+	uint64_t        offset;
+};
+
+struct sd_obj_rsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t        copies;
+	uint32_t        pad[6];
+};
+
+struct sd_vdi_req {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint64_t	vdi_size;
+	uint32_t        base_vdi_id;
+	uint32_t        copies;
+	uint32_t        snapid;
+	uint32_t        pad[3];
+};
+
+struct sd_vdi_rsp {
+	uint8_t		proto_ver;
+	uint8_t		opcode;
+	uint16_t	flags;
+	uint32_t	epoch;
+	uint32_t        id;
+	uint32_t        data_length;
+	uint32_t        result;
+	uint32_t        rsvd;
+	uint32_t        vdi_id;
+	uint32_t        pad[5];
+};
+
+struct sd_inode {
+	char name[SD_MAX_VDI_LEN];
+	uint64_t ctime;
+	uint64_t snap_ctime;
+	uint64_t vm_clock_nsec;
+	uint64_t vdi_size;
+	uint64_t vm_state_size;
+	uint16_t copy_policy;
+	uint8_t  nr_copies;
+	uint8_t  block_size_shift;
+	uint32_t snap_id;
+	uint32_t vdi_id;
+	uint32_t parent_vdi_id;
+	uint32_t child_vdi_id[MAX_CHILDREN];
+	uint32_t data_vdi_id[MAX_DATA_OBJS];
+};
+
+/*
+ * 64 bit FNV-1a non-zero initial basis
+ */
+#define FNV1A_64_INIT ((uint64_t)0xcbf29ce484222325ULL)
+
+/*
+ * 64 bit Fowler/Noll/Vo FNV-1a hash code
+ */
+static inline uint64_t fnv_64a_buf(void *buf, size_t len, uint64_t hval)
+{
+        unsigned char *bp = (unsigned char *) buf;
+        unsigned char *be = bp + len;
+        while (bp < be) {
+                hval ^= (uint64_t) *bp++;
+                hval += (hval << 1) + (hval << 4) + (hval << 5) +
+                        (hval << 7) + (hval << 8) + (hval << 40);
+        }
+        return hval;
+}
+
+static inline int is_data_obj_writeable(struct sd_inode *inode, unsigned int idx)
+{
+	return inode->vdi_id == inode->data_vdi_id[idx];
+}
+
+static inline int is_data_obj(uint64_t oid)
+{
+	return !(VDI_BIT & oid);
+}
+
+static inline uint64_t data_oid_to_idx(uint64_t oid)
+{
+	return oid & (MAX_DATA_OBJS - 1);
+}
+
+static inline uint64_t vid_to_vdi_oid(uint32_t vid)
+{
+	return VDI_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT);
+}
+
+static inline uint64_t vid_to_vmstate_oid(uint32_t vid, uint32_t idx)
+{
+	return VMSTATE_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
+}
+
+static inline uint64_t vid_to_data_oid(uint32_t vid, uint32_t idx)
+{
+	return ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
+}
+
+#undef eprintf
+#define eprintf(fmt, args...)						\
+do {									\
+	fprintf(stderr, "%s %d: " fmt, __func__, __LINE__, ##args);	\
+} while (0)
+
+#undef dprintf
+#ifdef DEBUG_SDOG
+#define dprintf(fmt, args...)						\
+do {									\
+	fprintf(stdout, "%s %d: " fmt, __func__, __LINE__, ##args);	\
+} while (0)
+#else
+#define dprintf(fmt, args...)
+#endif
+
+#define min_t(type, x, y) ({			\
+	type __min1 = (x);			\
+	type __min2 = (y);			\
+	__min1 < __min2 ? __min1: __min2; })
+
+struct aio_req {
+	struct sd_aiocb *aiocb;
+	unsigned int iov_offset;
+
+	uint64_t oid;
+	uint64_t base_oid;
+	uint64_t offset;
+	unsigned int data_len;
+	uint8_t flags;
+
+	QLIST_ENTRY(aio_req) pending_siblings;
+	QLIST_ENTRY(aio_req) aioreq_siblings;
+};
+
+enum aiocb_state {
+	AIOCB_WRITE_UDATA,
+	AIOCB_READ_UDATA,
+};
+
+struct sd_aiocb {
+	BlockDriverAIOCB common;
+
+	QEMUIOVector *qiov;
+
+	int64_t sector_num;
+	int nb_sectors;
+
+	int ret;
+	enum aiocb_state aiocb_type;
+
+	QEMUBH *bh;
+	void (*aio_done_func)(struct sd_aiocb *);
+
+	int canceled;
+
+	QLIST_HEAD(aioreq_head, aio_req) aioreq_head;
+};
+
+#define MAX_AIO_REQS 4096
+
+struct bdrv_sd_state {
+	struct sd_inode inode;
+
+	int nr_dirty_data_oids;
+	uint32_t dirty_data_oids[MAX_AIO_REQS];
+
+	char name[SD_MAX_VDI_LEN];
+	int is_current;
+
+	char *addr;
+	int fd;
+
+	struct aio_req aio_req_list[MAX_AIO_REQS];
+	struct aio_req *aio_req_free[MAX_AIO_REQS];
+	int nr_aio_req_free;
+
+	QLIST_HEAD(pending_head, aio_req) pending_head;
+};
+
+static const char * sd_strerror(int err)
+{
+	int i;
+
+	static const struct {
+		int err;
+		const char *desc;
+	} errors[] = {
+		{SD_RES_SUCCESS, "Success"},
+		{SD_RES_UNKNOWN, "Unknown error"},
+		{SD_RES_NO_OBJ, "No object found"},
+		{SD_RES_EIO, "I/O error"},
+		{SD_RES_VDI_EXIST, "VDI exists already"},
+		{SD_RES_INVALID_PARMS, "Invalid parameters"},
+		{SD_RES_SYSTEM_ERROR, "System error"},
+		{SD_RES_VDI_LOCKED, "VDI is already locked"},
+		{SD_RES_NO_VDI, "No vdi found"},
+		{SD_RES_NO_BASE_VDI, "No base VDI found"},
+		{SD_RES_VDI_READ, "Failed read the requested VDI"},
+		{SD_RES_VDI_WRITE, "Failed to write the requested VDI"},
+		{SD_RES_BASE_VDI_READ, "Failed to read the base VDI"},
+		{SD_RES_BASE_VDI_WRITE, "Failed to write the base VDI"},
+		{SD_RES_NO_TAG, "Failed to find the requested tag"},
+		{SD_RES_STARTUP, "The system is still booting"},
+		{SD_RES_VDI_NOT_LOCKED, "VDI isn't locked"},
+		{SD_RES_SHUTDOWN, "The system is shutting down"},
+		{SD_RES_NO_MEM, "Out of memory on the server"},
+		{SD_RES_FULL_VDI, "We already have the maximum vdis"},
+		{SD_RES_VER_MISMATCH, "Protocol version mismatch"},
+		{SD_RES_NO_SPACE, "Server has no space for new objects"},
+		{SD_RES_WAIT_FOR_FORMAT, "Sheepdog is waiting for a format operation"},
+		{SD_RES_WAIT_FOR_JOIN, "Sheepdog is waiting for other nodes joining"},
+		{SD_RES_JOIN_FAILED, "Target node had failed to join sheepdog"},
+	};
+
+	for (i = 0; i < ARRAY_SIZE(errors); ++i)
+		if (errors[i].err == err)
+			return errors[i].desc;
+
+	return "Invalid error code";
+}
+
+static inline int before(uint32_t seq1, uint32_t seq2)
+{
+        return (int32_t)(seq1 - seq2) < 0;
+}
+
+static inline int after(uint32_t seq1, uint32_t seq2)
+{
+	return (int32_t)(seq2 - seq1) < 0;
+}
+
+static inline struct aio_req *alloc_aio_req(struct bdrv_sd_state *s,
+					    struct sd_aiocb *acb,
+					    uint64_t oid, unsigned int data_len,
+					    uint64_t offset, uint8_t flags,
+					    uint64_t base_oid,
+					    unsigned int iov_offset)
+{
+	struct aio_req *aio_req;
+
+	if (!s->nr_aio_req_free)
+		return NULL;
+
+	aio_req = s->aio_req_free[--s->nr_aio_req_free];
+	aio_req->aiocb = acb;
+	aio_req->iov_offset = iov_offset;
+	aio_req->oid = oid;
+	aio_req->base_oid = base_oid;
+	aio_req->offset = offset;
+	aio_req->data_len = data_len;
+	aio_req->flags = flags;
+
+	QLIST_INSERT_HEAD(&s->pending_head, aio_req, pending_siblings);
+	QLIST_INSERT_HEAD(&acb->aioreq_head, aio_req, aioreq_siblings);
+
+	return aio_req;
+}
+
+static inline int free_aio_req(struct bdrv_sd_state *s, struct aio_req *aio_req)
+{
+	struct sd_aiocb *acb = aio_req->aiocb;
+	QLIST_REMOVE(aio_req, pending_siblings);
+	QLIST_REMOVE(aio_req, aioreq_siblings);
+	aio_req->aiocb = NULL;
+	s->aio_req_free[s->nr_aio_req_free++] = aio_req;
+
+	return !QLIST_EMPTY(&acb->aioreq_head);
+}
+
+static inline int nr_outstanding_aio_req(struct bdrv_sd_state *s)
+{
+	return MAX_AIO_REQS - s->nr_aio_req_free;
+}
+
+static inline int get_id_from_req(struct bdrv_sd_state *s, struct aio_req *aio_req)
+{
+	return aio_req - s->aio_req_list;
+}
+
+static inline struct aio_req *get_req_from_id(struct bdrv_sd_state *s, int id)
+{
+	return s->aio_req_list + id;
+}
+
+static void sd_finish_aiocb(struct sd_aiocb *acb)
+{
+	if (!acb->canceled)
+		acb->common.cb(acb->common.opaque, acb->ret);
+	qemu_aio_release(acb);
+}
+
+static void sd_aio_cancel(BlockDriverAIOCB *blockacb)
+{
+	struct sd_aiocb *acb = (struct sd_aiocb *)blockacb;
+
+	acb->canceled = 1;
+}
+
+static AIOPool sd_aio_pool = {
+	.aiocb_size = sizeof(struct sd_aiocb),
+	.cancel = sd_aio_cancel,
+};
+
+static struct sd_aiocb *sd_aio_setup(BlockDriverState *bs, QEMUIOVector *qiov,
+				     int64_t sector_num, int nb_sectors,
+				     BlockDriverCompletionFunc *cb,
+				     void *opaque)
+{
+	struct sd_aiocb *acb;
+
+	acb = qemu_aio_get(&sd_aio_pool, bs, cb, opaque);
+
+	acb->qiov = qiov;
+
+	acb->sector_num = sector_num;
+	acb->nb_sectors = nb_sectors;
+
+	acb->aio_done_func = NULL;
+	acb->canceled = 0;
+	acb->bh = NULL;
+	acb->ret = 0;
+	QLIST_INIT(&acb->aioreq_head);
+	return acb;
+}
+
+static int sd_schedule_bh(QEMUBHFunc *cb, struct sd_aiocb *acb)
+{
+	if (acb->bh) {
+		eprintf("bug: %d %d\n", acb->aiocb_type, acb->aiocb_type);
+		return -EIO;
+	}
+
+	acb->bh = qemu_bh_new(cb, acb);
+	if (!acb->bh) {
+		eprintf("oom: %d %d\n", acb->aiocb_type, acb->aiocb_type);
+		return -EIO;
+	}
+
+	qemu_bh_schedule(acb->bh);
+
+	return 0;
+}
+
+static int do_send_recv(int sockfd, struct iovec *iov, int len, int offset,
+			int write)
+{
+	struct msghdr msg;
+	int ret, diff;
+
+	memset(&msg, 0, sizeof(msg));
+	msg.msg_iov = iov;
+	msg.msg_iovlen = 1;
+
+	len += offset;
+
+	while (iov->iov_len < len) {
+		len -= iov->iov_len;
+
+		iov++;
+		msg.msg_iovlen++;
+	}
+
+	diff = iov->iov_len - len;
+	iov->iov_len -= diff;
+
+	while (msg.msg_iov->iov_len <= offset) {
+		offset -= msg.msg_iov->iov_len;
+
+		msg.msg_iov++;
+		msg.msg_iovlen--;
+	}
+
+	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base + offset;
+	msg.msg_iov->iov_len -= offset;
+
+	if (write)
+		ret = sendmsg(sockfd, &msg, 0);
+	else
+		ret = recvmsg(sockfd, &msg, MSG_WAITALL);
+
+	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base - offset;
+	msg.msg_iov->iov_len += offset;
+
+	iov->iov_len += diff;
+	return ret;
+}
+
+static int connect_to_sdog(const char *addr)
+{
+	char buf[64];
+	char hbuf[NI_MAXHOST], sbuf[NI_MAXSERV];
+	char name[256], *p;
+	int fd, ret;
+	struct addrinfo hints, *res, *res0;
+	int port = 0;
+
+	if (!addr)
+		addr = SD_DEFAULT_ADDR;
+
+	strcpy(name, addr);
+
+	p = name;
+	while (*p) {
+		if (*p == ':') {
+			*p++ = '\0';
+			break;
+		} else
+			p++;
+	}
+
+	if (*p == '\0') {
+		eprintf("cannot find a port number, %s\n", name);
+		return -1;
+	}
+	port = strtol(p, NULL, 10);
+	if (port == 0) {
+		eprintf("invalid port number, %s\n", p);
+		return -1;
+	}
+
+	memset(&hints, 0, sizeof(hints));
+	snprintf(buf, sizeof(buf), "%d", port);
+
+	hints.ai_socktype = SOCK_STREAM;
+
+	ret = getaddrinfo(name, buf, &hints, &res0);
+	if (ret) {
+		eprintf("unable to get address info %s, %m\n", name);
+		return -1;
+	}
+
+	for (res = res0; res; res = res->ai_next) {
+		ret = getnameinfo(res->ai_addr, res->ai_addrlen,
+				  hbuf, sizeof(hbuf), sbuf, sizeof(sbuf),
+				  NI_NUMERICHOST | NI_NUMERICSERV);
+		if (ret)
+			continue;
+
+		fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol);
+		if (fd < 0)
+			continue;
+
+reconnect:
+		ret = connect(fd, res->ai_addr, res->ai_addrlen);
+		if (ret < 0) {
+			if (errno == EINTR)
+				goto reconnect;
+			break;
+		}
+
+		dprintf("connected to %s:%d\n", name, port);
+		goto success;
+	}
+	fd = -1;
+	eprintf("failed connect to %s:%d\n", name, port);
+success:
+	freeaddrinfo(res0);
+	return fd;
+}
+
+static int do_readv_writev(int sockfd, struct iovec *iov, int len,
+			   int iov_offset, int write)
+{
+	int ret;
+again:
+	ret = do_send_recv(sockfd, iov, len, iov_offset, write);
+	if (ret < 0) {
+		if (errno == EINTR || errno == EAGAIN)
+			goto again;
+		eprintf("failed to recv a rsp, %m\n");
+		return 1;
+	}
+
+	iov_offset += ret;
+	len -= ret;
+	if (len)
+		goto again;
+
+	return 0;
+}
+
+static int do_readv(int sockfd, struct iovec *iov, int len, int iov_offset)
+{
+	return do_readv_writev(sockfd, iov, len, iov_offset, 0);
+}
+
+static int do_writev(int sockfd, struct iovec *iov, int len, int iov_offset)
+{
+	return do_readv_writev(sockfd, iov, len, iov_offset, 1);
+}
+
+static int do_read_write(int sockfd, void *buf, int len, int write)
+{
+	struct iovec iov;
+
+	iov.iov_base = buf;
+	iov.iov_len = len;
+
+	return do_readv_writev(sockfd, &iov, len, 0, write);
+}
+
+static int do_read(int sockfd, void *buf, int len)
+{
+	return do_read_write(sockfd, buf, len, 0);
+}
+
+static int do_write(int sockfd, void *buf, int len)
+{
+	return do_read_write(sockfd, buf, len, 1);
+}
+
+static int send_req(int sockfd, struct sd_req *hdr, void *data,
+		    unsigned int *wlen)
+{
+	int ret;
+	struct iovec iov[2];
+
+	iov[0].iov_base = hdr;
+	iov[0].iov_len = sizeof(*hdr);
+
+	if (*wlen) {
+		iov[1].iov_base = data;
+		iov[1].iov_len = *wlen;
+	}
+
+	ret = do_writev(sockfd, iov, sizeof(*hdr) + *wlen, 0);
+	if (ret) {
+		eprintf("failed to send a req, %m\n");
+		ret = -1;
+	}
+
+	return ret;
+}
+
+static int do_req(int sockfd, struct sd_req *hdr, void *data,
+		  unsigned int *wlen, unsigned int *rlen)
+{
+	int ret;
+
+	ret = send_req(sockfd, hdr, data, wlen);
+	if (ret) {
+		ret = -1;
+		goto out;
+	}
+
+	ret = do_read(sockfd, hdr, sizeof(*hdr));
+	if (ret) {
+		eprintf("failed to get a rsp, %m\n");
+		ret = -1;
+		goto out;
+	}
+
+	if (*rlen > hdr->data_length)
+		*rlen = hdr->data_length;
+
+	if (*rlen) {
+		ret = do_read(sockfd, data, *rlen);
+		if (ret) {
+			eprintf("failed to get the data, %m\n");
+			ret = -1;
+			goto out;
+		}
+	}
+	ret = 0;
+out:
+	return ret;
+}
+
+static int add_aio_request(struct bdrv_sd_state *s, struct aio_req *aio_req,
+			   struct iovec *iov, int niov, int create,
+			   enum aiocb_state aiocb_type);
+
+static void send_pending_req(struct bdrv_sd_state *s, uint64_t oid, uint32_t id)
+{
+	struct aio_req *aio_req, *next;
+	struct sd_aiocb *acb;
+	int ret;
+
+	QLIST_FOREACH_SAFE(aio_req, &s->pending_head, pending_siblings, next) {
+		if (id == get_id_from_req(s, aio_req))
+			continue;
+		if (aio_req->oid != oid)
+			continue;
+
+		acb = aio_req->aiocb;
+		ret = add_aio_request(s, aio_req, acb->qiov->iov,
+				      acb->qiov->niov, 0, acb->aiocb_type);
+		if (ret < 0) {
+			eprintf("add_aio_request is faled\n");
+			free_aio_req(s, aio_req);
+			if (QLIST_EMPTY(&acb->aioreq_head))
+				sd_finish_aiocb(acb);
+		}
+	}
+}
+
+static void aio_read_response(void *opaque)
+{
+	struct sd_obj_req hdr;
+	struct sd_obj_rsp *rsp = (struct sd_obj_rsp *)&hdr;
+	struct bdrv_sd_state *s = (struct bdrv_sd_state *)opaque;
+	int fd = s->fd;
+	int ret;
+	struct aio_req *aio_req;
+	struct sd_aiocb *acb;
+	int rest;
+	unsigned long idx;
+
+	if (!nr_outstanding_aio_req(s))
+		return;
+
+	ret = do_read(fd, (void *)rsp, sizeof(*rsp));
+	if (ret) {
+		eprintf("failed to get the header, %m\n");
+		return;
+	}
+
+	aio_req = get_req_from_id(s, rsp->id);
+	acb = aio_req->aiocb;
+
+	switch (acb->aiocb_type) {
+	case AIOCB_WRITE_UDATA:
+		if (!is_data_obj(aio_req->oid))
+			break;
+		idx = data_oid_to_idx(aio_req->oid);
+
+		if (s->inode.data_vdi_id[idx] != s->inode.vdi_id) {
+			s->inode.data_vdi_id[idx] = s->inode.vdi_id;
+			s->dirty_data_oids[s->nr_dirty_data_oids++] = idx;
+
+			send_pending_req(s, vid_to_data_oid(s->inode.vdi_id, idx),
+					 rsp->id);
+		}
+		break;
+	case AIOCB_READ_UDATA:
+		ret = do_readv(fd, acb->qiov->iov, rsp->data_length,
+			       aio_req->iov_offset);
+		if (ret) {
+			eprintf("failed to get the data, %m\n");
+			return;
+		}
+		break;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		acb->ret = -EIO;
+		eprintf("%s\n", sd_strerror(rsp->result));
+	}
+
+	rest = free_aio_req(s, aio_req);
+	if (!rest)
+		acb->aio_done_func(acb);
+}
+
+static int aio_flush_request(void *opaque)
+{
+	return nr_outstanding_aio_req((struct bdrv_sd_state *)opaque);
+}
+
+static int set_nonblocking(int fd)
+{
+	int ret;
+
+	ret = fcntl(fd, F_GETFL);
+	if (ret < 0) {
+		eprintf("can't fcntl (F_GETFL), %m\n");
+		close(fd);
+	} else {
+		ret = fcntl(fd, F_SETFL, ret | O_NONBLOCK);
+		if (ret < 0)
+			eprintf("can't fcntl (O_NONBLOCK), %m\n");
+	}
+
+	return ret;
+}
+
+static int set_nodelay(int fd)
+{
+	int ret, opt;
+
+	opt = 1;
+	ret = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &opt, sizeof(opt));
+	return ret;
+}
+
+static int get_sheep_fd(struct bdrv_sd_state *s)
+{
+	int ret, fd;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("%m\n");
+		return -1;
+	}
+
+	ret = set_nonblocking(fd);
+	if (ret) {
+		eprintf("%m\n");
+		close(fd);
+		return -1;
+	}
+
+	ret = set_nodelay(fd);
+	if (ret) {
+		eprintf("%m\n");
+		close(fd);
+		return -1;
+	}
+
+	qemu_aio_set_fd_handler(fd, aio_read_response, NULL, aio_flush_request,
+				NULL, s);
+	s->fd = fd;
+
+	return fd;
+}
+
+static int parse_vdiname(struct bdrv_sd_state *s, const char *filename,
+			 char *vdi, int vdi_len, uint32_t *snapid)
+{
+	char *p, *q;
+	int nr_sep;
+
+	p = q = strdup(filename);
+
+	if (!p)
+		return 1;
+
+	nr_sep = 0;
+	while (*p) {
+		if (*p == ':')
+			nr_sep++;
+		if (nr_sep == 2)
+			break;
+		p++;
+	}
+
+	if (nr_sep == 2)
+		*p++ = '\0';
+	else
+		p = q;
+
+	strncpy(vdi, p, vdi_len);
+
+	p = strchr(vdi, ':');
+	if (p) {
+		*p++ = '\0';
+		*snapid = strtol(p, NULL, 10);
+	} else
+		*snapid = CURRENT_VDI_ID; /* search current vdi */
+
+	if (nr_sep == 2)
+		s->addr = q;
+	else {
+		free(q);
+		s->addr = NULL;
+	}
+
+	return 0;
+}
+
+static int find_vdi_name(struct bdrv_sd_state *s, char *filename, uint32_t snapid,
+			 uint32_t *vid, int for_snapshot)
+{
+	int ret, fd;
+	struct sd_vdi_req hdr;
+	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
+	unsigned int wlen, rlen = 0;
+	char buf[SD_MAX_VDI_LEN];
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0)
+		return -1;
+
+	memset(&hdr, 0, sizeof(hdr));
+	snprintf(buf, sizeof(buf), "%s", filename);
+	if (for_snapshot)
+		hdr.opcode = SD_OP_GET_VDI_INFO;
+	else
+		hdr.opcode = SD_OP_LOCK_VDI;
+	wlen = SD_MAX_VDI_LEN;
+	hdr.proto_ver = SD_PROTO_VER;
+	hdr.data_length = SD_MAX_VDI_LEN;
+	hdr.snapid = snapid;
+	hdr.flags = SD_FLAG_CMD_WRITE;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
+	if (ret) {
+		ret = -1;
+		goto out;
+	}
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		eprintf("%s, %s\n", sd_strerror(rsp->result), filename);
+		ret = -1;
+		goto out;
+	}
+	*vid = rsp->vdi_id;
+
+	ret = 0;
+out:
+	close(fd);
+	return ret;
+}
+
+static int add_aio_request(struct bdrv_sd_state *s, struct aio_req *aio_req,
+			   struct iovec *iov, int niov, int create,
+			   enum aiocb_state aiocb_type)
+{
+	int nr_copies = s->inode.nr_copies;
+	struct sd_obj_req hdr;
+	unsigned int wlen;
+	int ret, opt;
+	uint64_t oid = aio_req->oid;
+	unsigned int datalen = aio_req->data_len;
+	uint64_t offset = aio_req->offset;
+	uint8_t flags = aio_req->flags;
+	uint64_t old_oid = aio_req->base_oid;
+
+	if (!nr_copies)
+		eprintf("bug\n");
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	if (aiocb_type == AIOCB_READ_UDATA) {
+		wlen = 0;
+		hdr.opcode = SD_OP_READ_OBJ;
+		hdr.flags = flags;
+	} else if (create) {
+		wlen = datalen;
+		hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
+		hdr.flags = SD_FLAG_CMD_WRITE | flags;
+	} else {
+		wlen = datalen;
+		hdr.opcode = SD_OP_WRITE_OBJ;
+		hdr.flags = SD_FLAG_CMD_WRITE | flags;
+	}
+
+	hdr.oid = oid;
+	hdr.cow_oid = old_oid;
+	hdr.copies = s->inode.nr_copies;
+
+	hdr.data_length = datalen;
+	hdr.offset = offset;
+
+	hdr.id = get_id_from_req(s, aio_req);
+
+	opt = 1;
+	setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
+
+	ret = do_write(s->fd, &hdr, sizeof(hdr));
+	if (ret) {
+		eprintf("failed to send a req, %m\n");
+		return -EIO;
+	}
+
+	if (wlen) {
+		ret = do_writev(s->fd, iov, wlen, aio_req->iov_offset);
+		if (ret) {
+			eprintf("failed to send a data, %m\n");
+			return -EIO;
+		}
+	}
+        opt = 0;
+        setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
+
+	return 0;
+}
+
+static int read_write_object(int fd, char *buf, uint64_t oid, int copies,
+			     unsigned int datalen, uint64_t offset,
+			     int write, int create)
+{
+	struct sd_obj_req hdr;
+	struct sd_obj_rsp *rsp = (struct sd_obj_rsp *)&hdr;
+	unsigned int wlen, rlen;
+	int ret;
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	if (write) {
+		wlen = datalen;
+		rlen = 0;
+		hdr.flags = SD_FLAG_CMD_WRITE;
+		if (create)
+			hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
+		else
+			hdr.opcode = SD_OP_WRITE_OBJ;
+	} else {
+		wlen = 0;
+		rlen = datalen;
+		hdr.opcode = SD_OP_READ_OBJ;
+	}
+	hdr.oid = oid;
+	hdr.data_length = datalen;
+	hdr.offset = offset;
+	hdr.copies = copies;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
+	if (ret) {
+		eprintf("failed to send a request to the sheep\n");
+		return -1;
+	}
+
+	switch (rsp->result) {
+	case SD_RES_SUCCESS:
+		return 0;
+	default:
+		eprintf("%s\n", sd_strerror(rsp->result));
+		return -1;
+	}
+}
+
+static int read_object(int fd, char *buf, uint64_t oid, int copies,
+		       unsigned int datalen, uint64_t offset)
+{
+	return read_write_object(fd, buf, oid, copies, datalen, offset, 0, 0);
+}
+
+static int write_object(int fd, char *buf, uint64_t oid, int copies,
+			unsigned int datalen, uint64_t offset, int create)
+{
+	return read_write_object(fd, buf, oid, copies, datalen, offset, 1, create);
+}
+
+/* TODO: error cleanups */
+static int sd_open(BlockDriverState *bs, const char *filename, int flags)
+{
+	int ret, i, fd;
+	uint32_t vid = 0;
+	struct bdrv_sd_state *s = bs->opaque;
+	char vdi[256];
+	uint32_t snapid;
+	int for_snapshot = 0;
+	char *buf;
+
+	strstart(filename, "sheepdog:", (const char **)&filename);
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	if (!buf) {
+		eprintf("Failed to allocate memory\n");
+		return -1;
+	}
+
+	for (i = 0; i < MAX_AIO_REQS; i++) {
+		s->aio_req_free[i] = &s->aio_req_list[i];
+		s->aio_req_list[i].aiocb = NULL;
+	}
+	s->nr_aio_req_free = MAX_AIO_REQS;
+
+	memset(vdi, 0, sizeof(vdi));
+	if (parse_vdiname(s, filename, vdi, sizeof(vdi), &snapid) < 0)
+		goto out;
+	s->fd = get_sheep_fd(s);
+	if (s->fd < 0)
+		return -1;
+
+	if (snapid != CURRENT_VDI_ID)
+		for_snapshot = 1;
+
+	ret = find_vdi_name(s, vdi, snapid, &vid, for_snapshot);
+	if (ret)
+		goto out;
+
+	if (snapid)
+		dprintf("%" PRIx32 " non current inode was open.\n", vid);
+	else
+		s->is_current = 1;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), 0, SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret)
+		goto out;
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+	s->nr_dirty_data_oids = 0;
+
+	bs->total_sectors = s->inode.vdi_size >> 9;
+	strncpy(s->name, vdi, sizeof(s->name));
+	qemu_free(buf);
+
+	QLIST_INIT(&s->pending_head);
+	return 0;
+out:
+	qemu_free(buf);
+	return -1;
+}
+
+static int do_sd_create(const char *addr, char *filename, char *tag,
+			int64_t total_sectors, uint32_t base_vid,
+			uint32_t *vdi_id, int snapshot)
+{
+	struct sd_vdi_req hdr;
+	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
+	int fd, ret;
+	unsigned int wlen, rlen = 0;
+	char buf[SD_MAX_VDI_LEN];
+
+	fd = connect_to_sdog(addr);
+	if (fd < 0)
+		return -1;
+
+	strncpy(buf, filename, SD_MAX_VDI_LEN);
+
+	memset(&hdr, 0, sizeof(hdr));
+	hdr.opcode = SD_OP_NEW_VDI;
+	hdr.base_vdi_id = base_vid;
+
+	wlen = SD_MAX_VDI_LEN;
+
+	hdr.flags = SD_FLAG_CMD_WRITE;
+	hdr.snapid = snapshot;
+
+	hdr.data_length = wlen;
+	hdr.vdi_size = total_sectors * 512;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
+
+	close(fd);
+
+	if (ret)
+		return -1;
+
+	if (rsp->result != SD_RES_SUCCESS) {
+		eprintf("%s, %s\n", sd_strerror(rsp->result), filename);
+		return -1;
+	}
+
+	if (vdi_id)
+		*vdi_id = rsp->vdi_id;
+
+	return 0;
+}
+
+static int sd_create(const char *filename, QEMUOptionParameter *options)
+{
+	int ret;
+	uint32_t vid = 0;
+	int64_t total_sectors = 0;
+	char *backing_file = NULL;
+
+	strstart(filename, "sheepdog:", (const char **)&filename);
+
+	while (options && options->name) {
+		if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
+			total_sectors = options->value.n / 512;
+		} else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
+			backing_file = options->value.s;
+		}
+		options++;
+	}
+
+	if (backing_file) {
+		BlockDriverState bs;
+		char vdi[SD_MAX_VDI_LEN];
+		uint32_t snapid;
+
+		strstart(backing_file, "sheepdog:", (const char **)&backing_file);
+		memset(&bs, 0, sizeof(bs));
+
+		bs.opaque = qemu_malloc(sizeof(struct bdrv_sd_state));
+		if (!bs.opaque)
+			return -1;
+
+		ret = sd_open(&bs, backing_file, 0);
+		if (ret < 0)
+			return -1;
+
+		if (parse_vdiname(bs.opaque, backing_file, vdi, sizeof(vdi), &snapid) < 0)
+			return -1;
+
+		/* cannot clone from a current inode */
+		if (snapid == CURRENT_VDI_ID)
+			return -1;
+
+		ret = find_vdi_name(bs.opaque, vdi, snapid, &vid, 1);
+		if (ret)
+			return -1;
+	}
+
+	return do_sd_create(NULL, (char *)filename, NULL, total_sectors, vid,
+			    NULL, 0);
+}
+
+static void sd_close(BlockDriverState *bs)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	struct sd_vdi_req hdr;
+	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
+	unsigned int wlen, rlen = 0;
+	int fd, ret;
+
+	dprintf("%s\n", s->name);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0)
+		return;
+
+	memset(&hdr, 0, sizeof(hdr));
+
+	hdr.opcode = SD_OP_RELEASE_VDI;
+	wlen = strlen(s->name) + 1;
+	hdr.data_length = wlen;
+	hdr.flags = SD_FLAG_CMD_WRITE;
+
+	ret = do_req(fd, (struct sd_req *)&hdr, s->name, &wlen, &rlen);
+
+	close(fd);
+
+	if (!ret && rsp->result != SD_RES_SUCCESS &&
+	    rsp->result != SD_RES_VDI_NOT_LOCKED)
+		eprintf("%s, %s\n", sd_strerror(rsp->result), s->name);
+
+	close(s->fd);
+	free(s->addr);
+}
+
+static void sd_write_done(struct sd_aiocb *acb)
+{
+	int ret, i;
+	struct bdrv_sd_state *s = acb->common.bs->opaque;
+	struct iovec iov;
+	struct aio_req *aio_req;
+	uint32_t offset, data_len, mn, mx;
+
+	if (s->nr_dirty_data_oids) {
+		mn = mx = s->dirty_data_oids[0];
+		for (i = 0; i < s->nr_dirty_data_oids; i++) {
+			if (mn > s->dirty_data_oids[i])
+				mn = s->dirty_data_oids[i];
+			if (mx < s->dirty_data_oids[i])
+				mx = s->dirty_data_oids[i];
+		}
+		offset = sizeof(s->inode) - sizeof(s->inode.data_vdi_id) +
+			mn * sizeof(s->inode.data_vdi_id[0]);
+		data_len = (mx - mn + 1) * sizeof(s->inode.data_vdi_id[0]);
+		s->nr_dirty_data_oids = 0;
+
+		iov.iov_base = &s->inode;
+		iov.iov_len = sizeof(s->inode);
+		aio_req = alloc_aio_req(s, acb, vid_to_vdi_oid(s->inode.vdi_id),
+					data_len, offset, 0, 0, offset);
+		if (!aio_req) {
+			eprintf("too many requests\n");
+			acb->ret = -EIO;
+			goto out;
+		}
+		ret = add_aio_request(s, aio_req, &iov, 1, 0, AIOCB_WRITE_UDATA);
+		if (ret) {
+			free_aio_req(s, aio_req);
+			acb->ret = -EIO;
+			goto out;
+		}
+
+		acb->aio_done_func = sd_finish_aiocb;
+		acb->aiocb_type = AIOCB_WRITE_UDATA;
+		return;
+	}
+out:
+	sd_finish_aiocb(acb);
+}
+
+static int sd_create_branch(struct bdrv_sd_state *s)
+{
+	int ret, fd;
+	uint32_t vid;
+	char *buf;
+
+	dprintf("%" PRIx32 " is not current.\n", s->inode.vdi_id);
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	if (!buf)
+		return -1;
+
+	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
+			   s->inode.vdi_id, &vid, 1);
+	if (ret)
+		goto out;
+
+	dprintf("%" PRIx32 " is created.\n", vid);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
+			  SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret < 0)
+		goto out;
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+
+	s->is_current = 1;
+	ret = 0;
+	dprintf("%" PRIx32 " was newly created.\n", s->inode.vdi_id);
+
+out:
+	qemu_free(buf);
+
+	return ret;
+}
+
+static void sd_readv_writev_bh_cb(void *p)
+{
+	struct sd_aiocb *acb = p;
+	int ret = 0;
+	unsigned long len, done = 0, total = acb->nb_sectors * 512;
+	unsigned long idx = acb->sector_num * 512 / SD_DATA_OBJ_SIZE;
+	uint64_t oid;
+	uint64_t offset = (acb->sector_num * 512) % SD_DATA_OBJ_SIZE;
+	struct bdrv_sd_state *s = acb->common.bs->opaque;
+	struct sd_inode *inode = &s->inode;
+	struct aio_req *aio_req;
+
+	qemu_bh_delete(acb->bh);
+	acb->bh = NULL;
+
+	if (acb->aiocb_type == AIOCB_WRITE_UDATA && !s->is_current) {
+		ret = sd_create_branch(s);
+		if (ret) {
+			acb->ret = -EIO;
+			goto out;
+		}
+	}
+
+	while (done != total) {
+		uint8_t flags = 0;
+		uint64_t old_oid = 0;
+		int create = 0;
+
+		oid = vid_to_data_oid(inode->data_vdi_id[idx], idx);
+
+		len = min_t(unsigned long, total - done, SD_DATA_OBJ_SIZE - offset);
+
+		if (!inode->data_vdi_id[idx]) {
+			if (acb->aiocb_type == AIOCB_READ_UDATA)
+				goto done;
+
+			create = 1;
+		} else if (acb->aiocb_type == AIOCB_WRITE_UDATA
+			   && !is_data_obj_writeable(inode, idx)) {
+			create = 1;
+			old_oid = oid;
+			flags = SD_FLAG_CMD_COW;
+		}
+
+		if (create) {
+			dprintf("update ino (%" PRIu32") %"
+				PRIu64 " %" PRIu64 " %" PRIu64 "\n",
+				inode->vdi_id, oid,
+				vid_to_data_oid(inode->data_vdi_id[idx], idx), idx);
+			oid = vid_to_data_oid(inode->vdi_id, idx);
+			dprintf("new oid %lx\n", oid);
+		}
+
+		aio_req = alloc_aio_req(s, acb, oid, len, offset, flags,
+					old_oid, done);
+		if (!aio_req) {
+			eprintf("too many requests\n");
+			acb->ret = -EIO;
+			goto out;
+		}
+
+		if (create) {
+			struct aio_req *areq;
+			QLIST_FOREACH(areq, &s->pending_head, pending_siblings) {
+				if (get_id_from_req(s, areq) == get_id_from_req(s, aio_req))
+					continue;
+				if (areq->oid == oid) {
+					aio_req->flags = 0;
+					aio_req->base_oid = 0;
+					goto done;
+				}
+			}
+		}
+
+		ret = add_aio_request(s, aio_req, acb->qiov->iov, acb->qiov->niov,
+				      create, acb->aiocb_type);
+		if (ret < 0) {
+			eprintf("add_aio_request is faled\n");
+			free_aio_req(s, aio_req);
+			acb->ret = -EIO;
+			goto out;
+		}
+	done:
+		offset = 0;
+		idx++;
+		done += len;
+	}
+out:
+	if (QLIST_EMPTY(&acb->aioreq_head))
+		sd_finish_aiocb(acb);
+}
+
+static BlockDriverAIOCB *sd_aio_writev(BlockDriverState *bs,
+				       int64_t sector_num,
+				       QEMUIOVector *qiov,
+				       int nb_sectors,
+				       BlockDriverCompletionFunc *cb,
+				       void *opaque)
+{
+	struct sd_aiocb *acb;
+
+	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
+	acb->aio_done_func = sd_write_done;
+	acb->aiocb_type = AIOCB_WRITE_UDATA;
+
+	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
+	return &acb->common;
+}
+
+static BlockDriverAIOCB *sd_aio_readv(BlockDriverState *bs,
+				      int64_t sector_num,
+				      QEMUIOVector *qiov,
+				      int nb_sectors,
+				      BlockDriverCompletionFunc *cb,
+				      void *opaque)
+{
+	struct sd_aiocb *acb;
+	int i;
+
+	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
+	acb->aiocb_type = AIOCB_READ_UDATA;
+	acb->aio_done_func = sd_finish_aiocb;
+
+	/*
+	 * TODO: we can do better; we don't need to initialize
+	 * blindly.
+	 */
+	for (i = 0; i < qiov->niov; i++)
+		memset(qiov->iov[i].iov_base, 0, qiov->iov[i].iov_len);
+
+	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
+	return &acb->common;
+}
+
+static int sd_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	int ret, fd;
+	uint32_t new_vid;
+	struct sd_inode *inode;
+	unsigned int datalen;
+	uint64_t offset;
+
+	dprintf("sn_info: name %s id_str %s s: name %s vm_state_size %d "
+		"is_current %d\n", sn_info->name, sn_info->id_str,
+		s->name, sn_info->vm_state_size, s->is_current);
+
+	if (!s->is_current) {
+		eprintf("You can't create a snapshot of "
+			"a non current VDI, %s (%" PRIu32 ").\n",
+			s->name, s->inode.vdi_id);
+
+		return -1;
+	}
+
+	dprintf("%s %s\n", sn_info->name, sn_info->id_str);
+
+	s->inode.vm_state_size = sn_info->vm_state_size;
+	s->inode.vm_clock_nsec = sn_info->vm_clock_nsec;
+	offset = 0;
+	/* we don't need to read entire object */
+	datalen = SD_INODE_SIZE - sizeof(s->inode.data_vdi_id);
+
+	/* refresh inode. */
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	ret = write_object(fd, (char *)&s->inode, vid_to_vdi_oid(s->inode.vdi_id),
+			   s->inode.nr_copies, datalen, offset, 0);
+	if (ret < 0) {
+		eprintf("failed to write snapshot's inode.\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
+			   s->inode.vdi_id, &new_vid, 1);
+	if (ret < 0) {
+		eprintf("failed to create inode for snapshot. %m\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	inode = (struct sd_inode *)qemu_malloc(datalen);
+	if (!inode) {
+		eprintf("failed to allocate memory for inode. %m\n");
+		goto cleanup;
+	}
+
+	ret = read_object(fd, (char *)inode, vid_to_vdi_oid(new_vid),
+			  s->inode.nr_copies, datalen, offset);
+
+	close(fd);
+
+	if (ret < 0) {
+		eprintf("failed to read new inode info. %m\n");
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	memcpy(&s->inode, inode, datalen);
+	dprintf("s->inode: name %s snap_id %x oid %x\n",
+		s->inode.name, s->inode.snap_id, s->inode.vdi_id);
+
+cleanup:
+	close(fd);
+	return ret;
+}
+
+static int sd_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	struct bdrv_sd_state *old_s;
+	char vdi[SD_MAX_VDI_LEN];
+	char *buf = NULL;
+	uint32_t vid;
+	uint32_t snapid = 0;
+	int ret = -ENOENT, fd;
+
+	old_s = qemu_malloc(sizeof(struct bdrv_sd_state));
+	if (!old_s) {
+		eprintf("failed to allocate memory for old state. %m\n");
+		goto out;
+	}
+
+	memcpy(old_s, s, sizeof(struct bdrv_sd_state));
+
+	snapid = strtol(snapshot_id, NULL, 10);
+	if (!snapid) {
+		eprintf("Invalid snapshot_id\n");
+		goto out;
+	}
+
+	buf = qemu_malloc(SD_INODE_SIZE);
+	if (!buf) {
+		eprintf("Failed to allocate memory\n");
+		goto out;
+	}
+	strncpy(vdi, s->name, sizeof(vdi));
+	ret = find_vdi_name(s, vdi, snapid, &vid, 1);
+	if (ret) {
+		eprintf("Failed to find_vdi_name\n");
+		ret = -ENOENT;
+		goto out;
+	}
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
+			  SD_INODE_SIZE, 0);
+
+	close(fd);
+
+	if (ret) {
+		ret = -ENOENT;
+		goto out;
+	}
+
+	memcpy(&s->inode, buf, sizeof(s->inode));
+
+	if (!s->inode.vm_state_size) {
+		eprintf("Invalid snapshot\n");
+		ret = -ENOENT;
+		goto out;
+	}
+
+	s->is_current = 0;
+
+	qemu_free(buf);
+	qemu_free(old_s);
+
+	return 0;
+out:
+	/* recover bdrv_sd_state */
+	memcpy(s, old_s, sizeof(struct bdrv_sd_state));
+	qemu_free(buf);
+	qemu_free(old_s);
+
+	eprintf("failed to open. recover old bdrv_sd_state.\n");
+
+	return ret;
+}
+
+static int sd_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
+{
+	/* FIXME: Delete specified snapshot id.  */
+	return 0;
+}
+
+#define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
+#define BITS_PER_BYTE		8
+#define BITS_TO_LONGS(nr)	DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(long))
+#define DECLARE_BITMAP(name,bits) \
+	unsigned long name[BITS_TO_LONGS(bits)]
+
+#define BITS_PER_LONG (BITS_PER_BYTE * sizeof(long))
+
+static inline int test_bit(unsigned int nr, const unsigned long *addr)
+{
+	return ((1UL << (nr % BITS_PER_LONG)) &
+		(((unsigned long *)addr)[nr / BITS_PER_LONG])) != 0;
+}
+
+static int sd_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+	struct sd_req req;
+	int i, fd, nr = 1024, ret, max = BITS_TO_LONGS(SD_NR_VDIS) * sizeof(long);
+	QEMUSnapshotInfo *sn_tab = NULL;
+	unsigned wlen, rlen;
+	int found = 0;
+	static struct sd_inode inode;
+	unsigned long *vdi_inuse;
+	unsigned int start_nr;
+
+	vdi_inuse = qemu_malloc(max);
+	if (!vdi_inuse)
+		return 0;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0)
+		goto out;
+
+	rlen = max;
+	wlen = 0;
+
+	memset(&req, 0, sizeof(req));
+
+	req.opcode = SD_OP_READ_VDIS;
+	req.data_length = max;
+
+	ret = do_req(fd, (struct sd_req *)&req, vdi_inuse, &wlen, &rlen);
+
+	close(fd);
+	if (ret)
+		goto out;
+
+	sn_tab = qemu_mallocz(nr * sizeof(*sn_tab));
+	if (!sn_tab)
+		goto out;
+
+	start_nr = fnv_64a_buf(s->name, strlen(s->name), FNV1A_64_INIT) & (SD_NR_VDIS - 1);
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		eprintf("failed to connect\n");
+		goto out;
+	}
+
+	/* TODO: round up */
+	for (i = start_nr; i < SD_NR_VDIS && found < nr; i++) {
+		if (!test_bit(i, vdi_inuse))
+			break;
+
+		/* we don't need to read entire object */
+		ret = read_object(fd, (char *)&inode, vid_to_vdi_oid(i),
+				  0, SD_INODE_SIZE - sizeof(inode.data_vdi_id), 0);
+
+		if (ret)
+			continue;
+
+		if (!strcmp(inode.name, s->name) && inode.snap_ctime) {
+			sn_tab[found].date_sec = inode.snap_ctime >> 32;
+			sn_tab[found].date_nsec = inode.snap_ctime & 0xffffffff;
+			sn_tab[found].vm_state_size = inode.vm_state_size;
+			sn_tab[found].vm_clock_nsec = inode.vm_clock_nsec;
+
+			snprintf(sn_tab[found].id_str, sizeof(sn_tab[found].id_str), "%u",
+				 inode.snap_id);
+			found++;
+		}
+	}
+
+	close(fd);
+out:
+	*psn_tab = sn_tab;
+
+	qemu_free(vdi_inuse);
+
+	return found;
+}
+
+static int do_load_save_vmstate(struct bdrv_sd_state *s, uint8_t *data,
+				int64_t pos, int size, int load)
+{
+	int fd, create;
+	int ret = 0;
+	unsigned int data_len;
+	uint64_t vmstate_oid;
+	uint32_t vdi_index;
+	uint64_t offset;
+
+	fd = connect_to_sdog(s->addr);
+	if (fd < 0) {
+		ret = -EIO;
+		goto cleanup;
+	}
+
+	while (size) {
+		vdi_index = pos / SD_DATA_OBJ_SIZE;
+		offset = pos % SD_DATA_OBJ_SIZE;
+
+		data_len = min_t(unsigned int, size, SD_DATA_OBJ_SIZE);
+
+		vmstate_oid = vid_to_vmstate_oid(s->inode.vdi_id, vdi_index);
+
+		create = (offset == 0);
+		if (load)
+			ret = read_object(fd, (char *)data, vmstate_oid,
+					  s->inode.nr_copies, data_len, offset);
+		else
+			ret = write_object(fd, (char *)data, vmstate_oid,
+					   s->inode.nr_copies, data_len, offset, create);
+
+		if (ret < 0) {
+			eprintf("failed to save vmstate %m\n");
+			ret = -EIO;
+			goto cleanup;
+		}
+
+		pos += data_len;
+		size -= data_len;
+		ret += data_len;
+	}
+cleanup:
+	close(fd);
+	return ret;
+}
+
+static int sd_save_vmstate(BlockDriverState *bs, const uint8_t *data,
+			   int64_t pos, int size)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+
+	return do_load_save_vmstate(s, (uint8_t *)data, pos, size, 0);
+}
+
+static int sd_load_vmstate(BlockDriverState *bs, uint8_t *data,
+			   int64_t pos, int size)
+{
+	struct bdrv_sd_state *s = bs->opaque;
+
+	return do_load_save_vmstate(s, data, pos, size, 1);
+}
+
+
+static QEMUOptionParameter sd_create_options[] = {
+	{
+		.name = BLOCK_OPT_SIZE,
+		.type = OPT_SIZE,
+		.help = "Virtual disk size"
+	},
+	{
+		.name = BLOCK_OPT_BACKING_FILE,
+		.type = OPT_STRING,
+		.help = "File name of a base image"
+	},
+	{ NULL }
+};
+
+BlockDriver bdrv_sheepdog = {
+	.format_name = "sheepdog",
+	.protocol_name = "sheepdog",
+	.instance_size = sizeof(struct bdrv_sd_state),
+	.bdrv_file_open = sd_open,
+	.bdrv_close = sd_close,
+	.bdrv_create = sd_create,
+
+	.bdrv_aio_readv = sd_aio_readv,
+	.bdrv_aio_writev = sd_aio_writev,
+
+	.bdrv_snapshot_create = sd_snapshot_create,
+	.bdrv_snapshot_goto = sd_snapshot_goto,
+	.bdrv_snapshot_delete = sd_snapshot_delete,
+	.bdrv_snapshot_list = sd_snapshot_list,
+
+	.bdrv_save_vmstate = sd_save_vmstate,
+	.bdrv_load_vmstate = sd_load_vmstate,
+
+	.create_options = sd_create_options,
+};
+
+static void bdrv_sheepdog_init(void)
+{
+	bdrv_register(&bdrv_sheepdog);
+}
+block_init(bdrv_sheepdog_init);
-- 
1.5.6.5

Re: [RFC PATCH v2 3/3] block: add sheepdog driver for distributed storage support

[Kevin Wolf]

Am 14.05.2010 11:51, schrieb MORITA Kazutaka:
> Sheepdog is a distributed storage system for QEMU. It provides highly
> available block level storage volumes to VMs like Amazon EBS.  This
> patch adds a qemu block driver for Sheepdog.
> 
> Sheepdog features are:
> - No node in the cluster is special (no metadata node, no control
>   node, etc)
> - Linear scalability in performance and capacity
> - No single point of failure
> - Autonomous management (zero configuration)
> - Useful volume management support such as snapshot and cloning
> - Thin provisioning
> - Autonomous load balancing
> 
> The more details are available at the project site:
>     http://www.osrg.net/sheepdog/
> 
> Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>

Once we solved the image creation thing, I think I'm going to be happy
with the block interface. Of course, this is something that doesn't even
directly affect the driver code, just the way it is used.

I have no clue about the Sheepdog protocol, so I'm just trying to
comment on some general details.

> ---
>  Makefile.objs    |    2 +-
>  block/sheepdog.c | 1831 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 1832 insertions(+), 1 deletions(-)
>  create mode 100644 block/sheepdog.c
> 
> diff --git a/Makefile.objs b/Makefile.objs
> index ecdd53e..6edbc57 100644
> --- a/Makefile.objs
> +++ b/Makefile.objs
> @@ -14,7 +14,7 @@ block-obj-$(CONFIG_LINUX_AIO) += linux-aio.o
>  
>  block-nested-y += raw.o cow.o qcow.o vdi.o vmdk.o cloop.o dmg.o bochs.o vpc.o vvfat.o
>  block-nested-y += qcow2.o qcow2-refcount.o qcow2-cluster.o qcow2-snapshot.o
> -block-nested-y += parallels.o nbd.o blkdebug.o
> +block-nested-y += parallels.o nbd.o blkdebug.o sheepdog.o
>  block-nested-$(CONFIG_WIN32) += raw-win32.o
>  block-nested-$(CONFIG_POSIX) += raw-posix.o
>  block-nested-$(CONFIG_CURL) += curl.o
> diff --git a/block/sheepdog.c b/block/sheepdog.c
> new file mode 100644
> index 0000000..adf3a71
> --- /dev/null
> +++ b/block/sheepdog.c
> @@ -0,0 +1,1831 @@
> +/*
> + * Copyright (C) 2009-2010 Nippon Telegraph and Telephone Corporation.
> + *
> + * This program is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU General Public License version
> + * 2 as published by the Free Software Foundation.
> + *
> + * You should have received a copy of the GNU General Public License
> + * along with this program. If not, see <http://www.gnu.org/licenses/>.
> + */
> +#include <netdb.h>
> +#include <netinet/tcp.h>
> +
> +#include "qemu-common.h"
> +#include "block_int.h"
> +
> +#define SD_PROTO_VER 0x01
> +
> +#define SD_DEFAULT_ADDR "localhost:7000"
> +
> +#define SD_OP_CREATE_AND_WRITE_OBJ  0x01
> +#define SD_OP_READ_OBJ       0x02
> +#define SD_OP_WRITE_OBJ      0x03
> +
> +#define SD_OP_NEW_VDI        0x11
> +#define SD_OP_LOCK_VDI       0x12
> +#define SD_OP_RELEASE_VDI    0x13
> +#define SD_OP_GET_VDI_INFO   0x14
> +#define SD_OP_READ_VDIS      0x15
> +
> +#define SD_FLAG_CMD_WRITE    0x01
> +#define SD_FLAG_CMD_COW      0x02
> +
> +#define SD_RES_SUCCESS       0x00 /* Success */
> +#define SD_RES_UNKNOWN       0x01 /* Unknown error */
> +#define SD_RES_NO_OBJ        0x02 /* No object found */
> +#define SD_RES_EIO           0x03 /* I/O error */
> +#define SD_RES_VDI_EXIST     0x04 /* Vdi exists already */
> +#define SD_RES_INVALID_PARMS 0x05 /* Invalid parameters */
> +#define SD_RES_SYSTEM_ERROR  0x06 /* System error */
> +#define SD_RES_VDI_LOCKED    0x07 /* Vdi is locked */
> +#define SD_RES_NO_VDI        0x08 /* No vdi found */
> +#define SD_RES_NO_BASE_VDI   0x09 /* No base vdi found */
> +#define SD_RES_VDI_READ      0x0A /* Cannot read requested vdi */
> +#define SD_RES_VDI_WRITE     0x0B /* Cannot write requested vdi */
> +#define SD_RES_BASE_VDI_READ 0x0C /* Cannot read base vdi */
> +#define SD_RES_BASE_VDI_WRITE   0x0D /* Cannot write base vdi */
> +#define SD_RES_NO_TAG        0x0E /* Requested tag is not found */
> +#define SD_RES_STARTUP       0x0F /* Sheepdog is on starting up */
> +#define SD_RES_VDI_NOT_LOCKED   0x10 /* Vdi is not locked */
> +#define SD_RES_SHUTDOWN      0x11 /* Sheepdog is shutting down */
> +#define SD_RES_NO_MEM        0x12 /* Cannot allocate memory */
> +#define SD_RES_FULL_VDI      0x13 /* we already have the maximum vdis */
> +#define SD_RES_VER_MISMATCH  0x14 /* Protocol version mismatch */
> +#define SD_RES_NO_SPACE      0x15 /* Server has no room for new objects */
> +#define SD_RES_WAIT_FOR_FORMAT  0x16 /* Sheepdog is waiting for a format operation */
> +#define SD_RES_WAIT_FOR_JOIN    0x17 /* Sheepdog is waiting for other nodes joining */
> +#define SD_RES_JOIN_FAILED   0x18 /* Target node had failed to join sheepdog */
> +
> +/*
> + * Object ID rules
> + *
> + *  0 - 19 (20 bits): data object space
> + * 20 - 31 (12 bits): reserved data object space
> + * 32 - 55 (24 bits): vdi object space
> + * 56 - 59 ( 4 bits): reserved vdi object space
> + * 60 - 63 ( 4 bits): object type indentifier space
> + */
> +
> +#define VDI_SPACE_SHIFT   32
> +#define VDI_BIT (UINT64_C(1) << 63)
> +#define VMSTATE_BIT (UINT64_C(1) << 62)
> +#define MAX_DATA_OBJS (1ULL << 20)
> +#define MAX_CHILDREN 1024
> +#define SD_MAX_VDI_LEN 256
> +#define SD_NR_VDIS   (1U << 24)
> +#define SD_DATA_OBJ_SIZE (UINT64_C(1) << 22)
> +
> +#define SD_INODE_SIZE (sizeof(struct sd_inode))
> +#define CURRENT_VDI_ID 0
> +
> +struct sd_req {
> +	uint8_t		proto_ver;
> +	uint8_t		opcode;
> +	uint16_t	flags;
> +	uint32_t	epoch;
> +	uint32_t        id;
> +	uint32_t        data_length;
> +	uint32_t	opcode_specific[8];
> +};

CODING_STYLE says that structs should be typedefed and their names
should be in CamelCase. So something like this:

typedef struct SheepdogReq {
    ...
} SheepdogReq;

(Or, if your prefer, SDReq; but with things like SDAIOCB I think it
becomes hard to read)

> +
> +struct sd_rsp {
> +	uint8_t		proto_ver;
> +	uint8_t		opcode;
> +	uint16_t	flags;
> +	uint32_t	epoch;
> +	uint32_t        id;
> +	uint32_t        data_length;
> +	uint32_t        result;
> +	uint32_t	opcode_specific[7];
> +};
> +
> +struct sd_obj_req {
> +	uint8_t		proto_ver;
> +	uint8_t		opcode;
> +	uint16_t	flags;
> +	uint32_t	epoch;
> +	uint32_t        id;
> +	uint32_t        data_length;
> +	uint64_t        oid;
> +	uint64_t        cow_oid;
> +	uint32_t        copies;
> +	uint32_t        rsvd;
> +	uint64_t        offset;
> +};
> +
> +struct sd_obj_rsp {
> +	uint8_t		proto_ver;
> +	uint8_t		opcode;
> +	uint16_t	flags;
> +	uint32_t	epoch;
> +	uint32_t        id;
> +	uint32_t        data_length;
> +	uint32_t        result;
> +	uint32_t        copies;
> +	uint32_t        pad[6];
> +};
> +
> +struct sd_vdi_req {
> +	uint8_t		proto_ver;
> +	uint8_t		opcode;
> +	uint16_t	flags;
> +	uint32_t	epoch;
> +	uint32_t        id;
> +	uint32_t        data_length;
> +	uint64_t	vdi_size;
> +	uint32_t        base_vdi_id;
> +	uint32_t        copies;
> +	uint32_t        snapid;
> +	uint32_t        pad[3];
> +};
> +
> +struct sd_vdi_rsp {
> +	uint8_t		proto_ver;
> +	uint8_t		opcode;
> +	uint16_t	flags;
> +	uint32_t	epoch;
> +	uint32_t        id;
> +	uint32_t        data_length;
> +	uint32_t        result;
> +	uint32_t        rsvd;
> +	uint32_t        vdi_id;
> +	uint32_t        pad[5];
> +};
> +
> +struct sd_inode {
> +	char name[SD_MAX_VDI_LEN];
> +	uint64_t ctime;
> +	uint64_t snap_ctime;
> +	uint64_t vm_clock_nsec;
> +	uint64_t vdi_size;
> +	uint64_t vm_state_size;
> +	uint16_t copy_policy;
> +	uint8_t  nr_copies;
> +	uint8_t  block_size_shift;
> +	uint32_t snap_id;
> +	uint32_t vdi_id;
> +	uint32_t parent_vdi_id;
> +	uint32_t child_vdi_id[MAX_CHILDREN];
> +	uint32_t data_vdi_id[MAX_DATA_OBJS];
> +};
> +
> +/*
> + * 64 bit FNV-1a non-zero initial basis
> + */
> +#define FNV1A_64_INIT ((uint64_t)0xcbf29ce484222325ULL)
> +
> +/*
> + * 64 bit Fowler/Noll/Vo FNV-1a hash code
> + */
> +static inline uint64_t fnv_64a_buf(void *buf, size_t len, uint64_t hval)
> +{
> +        unsigned char *bp = (unsigned char *) buf;
> +        unsigned char *be = bp + len;
> +        while (bp < be) {
> +                hval ^= (uint64_t) *bp++;
> +                hval += (hval << 1) + (hval << 4) + (hval << 5) +
> +                        (hval << 7) + (hval << 8) + (hval << 40);
> +        }
> +        return hval;
> +}
> +
> +static inline int is_data_obj_writeable(struct sd_inode *inode, unsigned int idx)
> +{
> +	return inode->vdi_id == inode->data_vdi_id[idx];
> +}
> +
> +static inline int is_data_obj(uint64_t oid)
> +{
> +	return !(VDI_BIT & oid);
> +}
> +
> +static inline uint64_t data_oid_to_idx(uint64_t oid)
> +{
> +	return oid & (MAX_DATA_OBJS - 1);
> +}
> +
> +static inline uint64_t vid_to_vdi_oid(uint32_t vid)
> +{
> +	return VDI_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT);
> +}
> +
> +static inline uint64_t vid_to_vmstate_oid(uint32_t vid, uint32_t idx)
> +{
> +	return VMSTATE_BIT | ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
> +}
> +
> +static inline uint64_t vid_to_data_oid(uint32_t vid, uint32_t idx)
> +{
> +	return ((uint64_t)vid << VDI_SPACE_SHIFT) | idx;
> +}
> +
> +#undef eprintf
> +#define eprintf(fmt, args...)						\
> +do {									\
> +	fprintf(stderr, "%s %d: " fmt, __func__, __LINE__, ##args);	\
> +} while (0)

What about using error_report() instead of fprintf? Though it should be
the same currently.

> +
> +#undef dprintf
> +#ifdef DEBUG_SDOG
> +#define dprintf(fmt, args...)						\
> +do {									\
> +	fprintf(stdout, "%s %d: " fmt, __func__, __LINE__, ##args);	\
> +} while (0)
> +#else
> +#define dprintf(fmt, args...)
> +#endif
> +
> +#define min_t(type, x, y) ({			\
> +	type __min1 = (x);			\
> +	type __min2 = (y);			\
> +	__min1 < __min2 ? __min1: __min2; })
> +
> +struct aio_req {
> +	struct sd_aiocb *aiocb;
> +	unsigned int iov_offset;
> +
> +	uint64_t oid;
> +	uint64_t base_oid;
> +	uint64_t offset;
> +	unsigned int data_len;
> +	uint8_t flags;
> +
> +	QLIST_ENTRY(aio_req) pending_siblings;
> +	QLIST_ENTRY(aio_req) aioreq_siblings;
> +};
> +
> +enum aiocb_state {
> +	AIOCB_WRITE_UDATA,
> +	AIOCB_READ_UDATA,
> +};
> +
> +struct sd_aiocb {
> +	BlockDriverAIOCB common;
> +
> +	QEMUIOVector *qiov;
> +
> +	int64_t sector_num;
> +	int nb_sectors;
> +
> +	int ret;
> +	enum aiocb_state aiocb_type;
> +
> +	QEMUBH *bh;
> +	void (*aio_done_func)(struct sd_aiocb *);
> +
> +	int canceled;
> +
> +	QLIST_HEAD(aioreq_head, aio_req) aioreq_head;
> +};
> +
> +#define MAX_AIO_REQS 4096
> +
> +struct bdrv_sd_state {
> +	struct sd_inode inode;
> +
> +	int nr_dirty_data_oids;
> +	uint32_t dirty_data_oids[MAX_AIO_REQS];
> +
> +	char name[SD_MAX_VDI_LEN];
> +	int is_current;
> +
> +	char *addr;
> +	int fd;
> +
> +	struct aio_req aio_req_list[MAX_AIO_REQS];
> +	struct aio_req *aio_req_free[MAX_AIO_REQS];
> +	int nr_aio_req_free;
> +
> +	QLIST_HEAD(pending_head, aio_req) pending_head;
> +};
> +
> +static const char * sd_strerror(int err)
> +{
> +	int i;
> +
> +	static const struct {
> +		int err;
> +		const char *desc;
> +	} errors[] = {
> +		{SD_RES_SUCCESS, "Success"},
> +		{SD_RES_UNKNOWN, "Unknown error"},
> +		{SD_RES_NO_OBJ, "No object found"},
> +		{SD_RES_EIO, "I/O error"},
> +		{SD_RES_VDI_EXIST, "VDI exists already"},
> +		{SD_RES_INVALID_PARMS, "Invalid parameters"},
> +		{SD_RES_SYSTEM_ERROR, "System error"},
> +		{SD_RES_VDI_LOCKED, "VDI is already locked"},
> +		{SD_RES_NO_VDI, "No vdi found"},
> +		{SD_RES_NO_BASE_VDI, "No base VDI found"},
> +		{SD_RES_VDI_READ, "Failed read the requested VDI"},
> +		{SD_RES_VDI_WRITE, "Failed to write the requested VDI"},
> +		{SD_RES_BASE_VDI_READ, "Failed to read the base VDI"},
> +		{SD_RES_BASE_VDI_WRITE, "Failed to write the base VDI"},
> +		{SD_RES_NO_TAG, "Failed to find the requested tag"},
> +		{SD_RES_STARTUP, "The system is still booting"},
> +		{SD_RES_VDI_NOT_LOCKED, "VDI isn't locked"},
> +		{SD_RES_SHUTDOWN, "The system is shutting down"},
> +		{SD_RES_NO_MEM, "Out of memory on the server"},
> +		{SD_RES_FULL_VDI, "We already have the maximum vdis"},
> +		{SD_RES_VER_MISMATCH, "Protocol version mismatch"},
> +		{SD_RES_NO_SPACE, "Server has no space for new objects"},
> +		{SD_RES_WAIT_FOR_FORMAT, "Sheepdog is waiting for a format operation"},
> +		{SD_RES_WAIT_FOR_JOIN, "Sheepdog is waiting for other nodes joining"},
> +		{SD_RES_JOIN_FAILED, "Target node had failed to join sheepdog"},
> +	};
> +
> +	for (i = 0; i < ARRAY_SIZE(errors); ++i)
> +		if (errors[i].err == err)
> +			return errors[i].desc;

CODING_STYLE requires braces here.

> +
> +	return "Invalid error code";
> +}
> +
> +static inline int before(uint32_t seq1, uint32_t seq2)
> +{
> +        return (int32_t)(seq1 - seq2) < 0;
> +}
> +
> +static inline int after(uint32_t seq1, uint32_t seq2)
> +{
> +	return (int32_t)(seq2 - seq1) < 0;
> +}

These functions look strange... Is the difference to seq1 < seq2 that
the cast introduces intentional? (after(0x0, 0xabcdefff) == 1)

If yes, why is this useful? This needs a comment. If no, why even bother
to have this function instead of directly using < or > ?

> +
> +static inline struct aio_req *alloc_aio_req(struct bdrv_sd_state *s,
> +					    struct sd_aiocb *acb,
> +					    uint64_t oid, unsigned int data_len,
> +					    uint64_t offset, uint8_t flags,
> +					    uint64_t base_oid,
> +					    unsigned int iov_offset)
> +{
> +	struct aio_req *aio_req;
> +
> +	if (!s->nr_aio_req_free)
> +		return NULL;
> +
> +	aio_req = s->aio_req_free[--s->nr_aio_req_free];
> +	aio_req->aiocb = acb;
> +	aio_req->iov_offset = iov_offset;
> +	aio_req->oid = oid;
> +	aio_req->base_oid = base_oid;
> +	aio_req->offset = offset;
> +	aio_req->data_len = data_len;
> +	aio_req->flags = flags;
> +
> +	QLIST_INSERT_HEAD(&s->pending_head, aio_req, pending_siblings);
> +	QLIST_INSERT_HEAD(&acb->aioreq_head, aio_req, aioreq_siblings);
> +
> +	return aio_req;
> +}
> +
> +static inline int free_aio_req(struct bdrv_sd_state *s, struct aio_req *aio_req)
> +{
> +	struct sd_aiocb *acb = aio_req->aiocb;
> +	QLIST_REMOVE(aio_req, pending_siblings);
> +	QLIST_REMOVE(aio_req, aioreq_siblings);
> +	aio_req->aiocb = NULL;
> +	s->aio_req_free[s->nr_aio_req_free++] = aio_req;
> +
> +	return !QLIST_EMPTY(&acb->aioreq_head);
> +}
> +
> +static inline int nr_outstanding_aio_req(struct bdrv_sd_state *s)
> +{
> +	return MAX_AIO_REQS - s->nr_aio_req_free;
> +}
> +
> +static inline int get_id_from_req(struct bdrv_sd_state *s, struct aio_req *aio_req)
> +{
> +	return aio_req - s->aio_req_list;
> +}
> +
> +static inline struct aio_req *get_req_from_id(struct bdrv_sd_state *s, int id)
> +{
> +	return s->aio_req_list + id;
> +}
> +
> +static void sd_finish_aiocb(struct sd_aiocb *acb)
> +{
> +	if (!acb->canceled)
> +		acb->common.cb(acb->common.opaque, acb->ret);
> +	qemu_aio_release(acb);
> +}
> +
> +static void sd_aio_cancel(BlockDriverAIOCB *blockacb)
> +{
> +	struct sd_aiocb *acb = (struct sd_aiocb *)blockacb;
> +
> +	acb->canceled = 1;
> +}
> +
> +static AIOPool sd_aio_pool = {
> +	.aiocb_size = sizeof(struct sd_aiocb),
> +	.cancel = sd_aio_cancel,
> +};
> +
> +static struct sd_aiocb *sd_aio_setup(BlockDriverState *bs, QEMUIOVector *qiov,
> +				     int64_t sector_num, int nb_sectors,
> +				     BlockDriverCompletionFunc *cb,
> +				     void *opaque)
> +{
> +	struct sd_aiocb *acb;
> +
> +	acb = qemu_aio_get(&sd_aio_pool, bs, cb, opaque);
> +
> +	acb->qiov = qiov;
> +
> +	acb->sector_num = sector_num;
> +	acb->nb_sectors = nb_sectors;
> +
> +	acb->aio_done_func = NULL;
> +	acb->canceled = 0;
> +	acb->bh = NULL;
> +	acb->ret = 0;
> +	QLIST_INIT(&acb->aioreq_head);
> +	return acb;
> +}
> +
> +static int sd_schedule_bh(QEMUBHFunc *cb, struct sd_aiocb *acb)
> +{
> +	if (acb->bh) {
> +		eprintf("bug: %d %d\n", acb->aiocb_type, acb->aiocb_type);
> +		return -EIO;
> +	}
> +
> +	acb->bh = qemu_bh_new(cb, acb);
> +	if (!acb->bh) {
> +		eprintf("oom: %d %d\n", acb->aiocb_type, acb->aiocb_type);
> +		return -EIO;
> +	}
> +
> +	qemu_bh_schedule(acb->bh);
> +
> +	return 0;
> +}
> +
> +static int do_send_recv(int sockfd, struct iovec *iov, int len, int offset,
> +			int write)
> +{
> +	struct msghdr msg;
> +	int ret, diff;
> +
> +	memset(&msg, 0, sizeof(msg));
> +	msg.msg_iov = iov;
> +	msg.msg_iovlen = 1;
> +
> +	len += offset;
> +
> +	while (iov->iov_len < len) {
> +		len -= iov->iov_len;
> +
> +		iov++;
> +		msg.msg_iovlen++;
> +	}
> +
> +	diff = iov->iov_len - len;
> +	iov->iov_len -= diff;
> +
> +	while (msg.msg_iov->iov_len <= offset) {
> +		offset -= msg.msg_iov->iov_len;
> +
> +		msg.msg_iov++;
> +		msg.msg_iovlen--;
> +	}
> +
> +	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base + offset;
> +	msg.msg_iov->iov_len -= offset;
> +
> +	if (write)
> +		ret = sendmsg(sockfd, &msg, 0);
> +	else
> +		ret = recvmsg(sockfd, &msg, MSG_WAITALL);
> +
> +	msg.msg_iov->iov_base = (char *) msg.msg_iov->iov_base - offset;
> +	msg.msg_iov->iov_len += offset;
> +
> +	iov->iov_len += diff;
> +	return ret;
> +}
> +
> +static int connect_to_sdog(const char *addr)
> +{
> +	char buf[64];
> +	char hbuf[NI_MAXHOST], sbuf[NI_MAXSERV];
> +	char name[256], *p;
> +	int fd, ret;
> +	struct addrinfo hints, *res, *res0;
> +	int port = 0;
> +
> +	if (!addr)
> +		addr = SD_DEFAULT_ADDR;
> +
> +	strcpy(name, addr);
> +
> +	p = name;
> +	while (*p) {
> +		if (*p == ':') {
> +			*p++ = '\0';
> +			break;
> +		} else
> +			p++;
> +	}
> +
> +	if (*p == '\0') {
> +		eprintf("cannot find a port number, %s\n", name);
> +		return -1;
> +	}
> +	port = strtol(p, NULL, 10);
> +	if (port == 0) {
> +		eprintf("invalid port number, %s\n", p);
> +		return -1;
> +	}
> +
> +	memset(&hints, 0, sizeof(hints));
> +	snprintf(buf, sizeof(buf), "%d", port);
> +
> +	hints.ai_socktype = SOCK_STREAM;
> +
> +	ret = getaddrinfo(name, buf, &hints, &res0);
> +	if (ret) {
> +		eprintf("unable to get address info %s, %m\n", name);
> +		return -1;
> +	}
> +
> +	for (res = res0; res; res = res->ai_next) {
> +		ret = getnameinfo(res->ai_addr, res->ai_addrlen,
> +				  hbuf, sizeof(hbuf), sbuf, sizeof(sbuf),
> +				  NI_NUMERICHOST | NI_NUMERICSERV);
> +		if (ret)
> +			continue;
> +
> +		fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol);
> +		if (fd < 0)
> +			continue;
> +
> +reconnect:
> +		ret = connect(fd, res->ai_addr, res->ai_addrlen);
> +		if (ret < 0) {
> +			if (errno == EINTR)
> +				goto reconnect;
> +			break;
> +		}
> +
> +		dprintf("connected to %s:%d\n", name, port);
> +		goto success;
> +	}
> +	fd = -1;
> +	eprintf("failed connect to %s:%d\n", name, port);
> +success:
> +	freeaddrinfo(res0);
> +	return fd;
> +}
> +
> +static int do_readv_writev(int sockfd, struct iovec *iov, int len,
> +			   int iov_offset, int write)
> +{
> +	int ret;
> +again:
> +	ret = do_send_recv(sockfd, iov, len, iov_offset, write);
> +	if (ret < 0) {
> +		if (errno == EINTR || errno == EAGAIN)
> +			goto again;
> +		eprintf("failed to recv a rsp, %m\n");
> +		return 1;
> +	}
> +
> +	iov_offset += ret;
> +	len -= ret;
> +	if (len)
> +		goto again;
> +
> +	return 0;
> +}
> +
> +static int do_readv(int sockfd, struct iovec *iov, int len, int iov_offset)
> +{
> +	return do_readv_writev(sockfd, iov, len, iov_offset, 0);
> +}
> +
> +static int do_writev(int sockfd, struct iovec *iov, int len, int iov_offset)
> +{
> +	return do_readv_writev(sockfd, iov, len, iov_offset, 1);
> +}
> +
> +static int do_read_write(int sockfd, void *buf, int len, int write)
> +{
> +	struct iovec iov;
> +
> +	iov.iov_base = buf;
> +	iov.iov_len = len;
> +
> +	return do_readv_writev(sockfd, &iov, len, 0, write);
> +}
> +
> +static int do_read(int sockfd, void *buf, int len)
> +{
> +	return do_read_write(sockfd, buf, len, 0);
> +}
> +
> +static int do_write(int sockfd, void *buf, int len)
> +{
> +	return do_read_write(sockfd, buf, len, 1);
> +}
> +
> +static int send_req(int sockfd, struct sd_req *hdr, void *data,
> +		    unsigned int *wlen)
> +{
> +	int ret;
> +	struct iovec iov[2];
> +
> +	iov[0].iov_base = hdr;
> +	iov[0].iov_len = sizeof(*hdr);
> +
> +	if (*wlen) {
> +		iov[1].iov_base = data;
> +		iov[1].iov_len = *wlen;
> +	}
> +
> +	ret = do_writev(sockfd, iov, sizeof(*hdr) + *wlen, 0);
> +	if (ret) {
> +		eprintf("failed to send a req, %m\n");
> +		ret = -1;
> +	}
> +
> +	return ret;
> +}
> +
> +static int do_req(int sockfd, struct sd_req *hdr, void *data,
> +		  unsigned int *wlen, unsigned int *rlen)
> +{
> +	int ret;
> +
> +	ret = send_req(sockfd, hdr, data, wlen);
> +	if (ret) {
> +		ret = -1;
> +		goto out;
> +	}
> +
> +	ret = do_read(sockfd, hdr, sizeof(*hdr));
> +	if (ret) {
> +		eprintf("failed to get a rsp, %m\n");
> +		ret = -1;
> +		goto out;
> +	}
> +
> +	if (*rlen > hdr->data_length)
> +		*rlen = hdr->data_length;
> +
> +	if (*rlen) {
> +		ret = do_read(sockfd, data, *rlen);
> +		if (ret) {
> +			eprintf("failed to get the data, %m\n");
> +			ret = -1;
> +			goto out;
> +		}
> +	}
> +	ret = 0;
> +out:
> +	return ret;
> +}
> +
> +static int add_aio_request(struct bdrv_sd_state *s, struct aio_req *aio_req,
> +			   struct iovec *iov, int niov, int create,
> +			   enum aiocb_state aiocb_type);
> +
> +static void send_pending_req(struct bdrv_sd_state *s, uint64_t oid, uint32_t id)
> +{
> +	struct aio_req *aio_req, *next;
> +	struct sd_aiocb *acb;
> +	int ret;
> +
> +	QLIST_FOREACH_SAFE(aio_req, &s->pending_head, pending_siblings, next) {
> +		if (id == get_id_from_req(s, aio_req))
> +			continue;
> +		if (aio_req->oid != oid)
> +			continue;
> +
> +		acb = aio_req->aiocb;
> +		ret = add_aio_request(s, aio_req, acb->qiov->iov,
> +				      acb->qiov->niov, 0, acb->aiocb_type);
> +		if (ret < 0) {
> +			eprintf("add_aio_request is faled\n");
> +			free_aio_req(s, aio_req);
> +			if (QLIST_EMPTY(&acb->aioreq_head))
> +				sd_finish_aiocb(acb);
> +		}
> +	}
> +}
> +
> +static void aio_read_response(void *opaque)
> +{
> +	struct sd_obj_req hdr;
> +	struct sd_obj_rsp *rsp = (struct sd_obj_rsp *)&hdr;
> +	struct bdrv_sd_state *s = (struct bdrv_sd_state *)opaque;
> +	int fd = s->fd;
> +	int ret;
> +	struct aio_req *aio_req;
> +	struct sd_aiocb *acb;
> +	int rest;
> +	unsigned long idx;
> +
> +	if (!nr_outstanding_aio_req(s))
> +		return;
> +
> +	ret = do_read(fd, (void *)rsp, sizeof(*rsp));
> +	if (ret) {
> +		eprintf("failed to get the header, %m\n");
> +		return;
> +	}
> +
> +	aio_req = get_req_from_id(s, rsp->id);
> +	acb = aio_req->aiocb;
> +
> +	switch (acb->aiocb_type) {
> +	case AIOCB_WRITE_UDATA:
> +		if (!is_data_obj(aio_req->oid))
> +			break;
> +		idx = data_oid_to_idx(aio_req->oid);
> +
> +		if (s->inode.data_vdi_id[idx] != s->inode.vdi_id) {
> +			s->inode.data_vdi_id[idx] = s->inode.vdi_id;
> +			s->dirty_data_oids[s->nr_dirty_data_oids++] = idx;
> +
> +			send_pending_req(s, vid_to_data_oid(s->inode.vdi_id, idx),
> +					 rsp->id);
> +		}
> +		break;
> +	case AIOCB_READ_UDATA:
> +		ret = do_readv(fd, acb->qiov->iov, rsp->data_length,
> +			       aio_req->iov_offset);
> +		if (ret) {
> +			eprintf("failed to get the data, %m\n");
> +			return;
> +		}
> +		break;
> +	}
> +
> +	if (rsp->result != SD_RES_SUCCESS) {
> +		acb->ret = -EIO;
> +		eprintf("%s\n", sd_strerror(rsp->result));
> +	}
> +
> +	rest = free_aio_req(s, aio_req);
> +	if (!rest)
> +		acb->aio_done_func(acb);
> +}
> +
> +static int aio_flush_request(void *opaque)
> +{
> +	return nr_outstanding_aio_req((struct bdrv_sd_state *)opaque);
> +}
> +
> +static int set_nonblocking(int fd)
> +{
> +	int ret;
> +
> +	ret = fcntl(fd, F_GETFL);
> +	if (ret < 0) {
> +		eprintf("can't fcntl (F_GETFL), %m\n");
> +		close(fd);
> +	} else {
> +		ret = fcntl(fd, F_SETFL, ret | O_NONBLOCK);
> +		if (ret < 0)
> +			eprintf("can't fcntl (O_NONBLOCK), %m\n");
> +	}
> +
> +	return ret;
> +}
> +
> +static int set_nodelay(int fd)
> +{
> +	int ret, opt;
> +
> +	opt = 1;
> +	ret = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &opt, sizeof(opt));
> +	return ret;
> +}
> +
> +static int get_sheep_fd(struct bdrv_sd_state *s)
> +{
> +	int ret, fd;
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0) {
> +		eprintf("%m\n");
> +		return -1;
> +	}
> +
> +	ret = set_nonblocking(fd);
> +	if (ret) {
> +		eprintf("%m\n");
> +		close(fd);
> +		return -1;
> +	}
> +
> +	ret = set_nodelay(fd);
> +	if (ret) {
> +		eprintf("%m\n");
> +		close(fd);
> +		return -1;
> +	}
> +
> +	qemu_aio_set_fd_handler(fd, aio_read_response, NULL, aio_flush_request,
> +				NULL, s);
> +	s->fd = fd;
> +
> +	return fd;
> +}
> +
> +static int parse_vdiname(struct bdrv_sd_state *s, const char *filename,
> +			 char *vdi, int vdi_len, uint32_t *snapid)
> +{
> +	char *p, *q;
> +	int nr_sep;
> +
> +	p = q = strdup(filename);
> +
> +	if (!p)
> +		return 1;
> +
> +	nr_sep = 0;
> +	while (*p) {
> +		if (*p == ':')
> +			nr_sep++;
> +		if (nr_sep == 2)
> +			break;
> +		p++;
> +	}
> +
> +	if (nr_sep == 2)
> +		*p++ = '\0';
> +	else
> +		p = q;
> +
> +	strncpy(vdi, p, vdi_len);
> +
> +	p = strchr(vdi, ':');
> +	if (p) {
> +		*p++ = '\0';
> +		*snapid = strtol(p, NULL, 10);
> +	} else
> +		*snapid = CURRENT_VDI_ID; /* search current vdi */
> +
> +	if (nr_sep == 2)
> +		s->addr = q;
> +	else {
> +		free(q);
> +		s->addr = NULL;
> +	}
> +
> +	return 0;
> +}
> +
> +static int find_vdi_name(struct bdrv_sd_state *s, char *filename, uint32_t snapid,
> +			 uint32_t *vid, int for_snapshot)
> +{
> +	int ret, fd;
> +	struct sd_vdi_req hdr;
> +	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
> +	unsigned int wlen, rlen = 0;
> +	char buf[SD_MAX_VDI_LEN];
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0)
> +		return -1;
> +
> +	memset(&hdr, 0, sizeof(hdr));
> +	snprintf(buf, sizeof(buf), "%s", filename);
> +	if (for_snapshot)
> +		hdr.opcode = SD_OP_GET_VDI_INFO;
> +	else
> +		hdr.opcode = SD_OP_LOCK_VDI;
> +	wlen = SD_MAX_VDI_LEN;
> +	hdr.proto_ver = SD_PROTO_VER;
> +	hdr.data_length = SD_MAX_VDI_LEN;
> +	hdr.snapid = snapid;
> +	hdr.flags = SD_FLAG_CMD_WRITE;
> +
> +	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
> +	if (ret) {
> +		ret = -1;
> +		goto out;
> +	}
> +
> +	if (rsp->result != SD_RES_SUCCESS) {
> +		eprintf("%s, %s\n", sd_strerror(rsp->result), filename);
> +		ret = -1;
> +		goto out;
> +	}
> +	*vid = rsp->vdi_id;
> +
> +	ret = 0;
> +out:
> +	close(fd);
> +	return ret;
> +}
> +
> +static int add_aio_request(struct bdrv_sd_state *s, struct aio_req *aio_req,
> +			   struct iovec *iov, int niov, int create,
> +			   enum aiocb_state aiocb_type)
> +{
> +	int nr_copies = s->inode.nr_copies;
> +	struct sd_obj_req hdr;
> +	unsigned int wlen;
> +	int ret, opt;
> +	uint64_t oid = aio_req->oid;
> +	unsigned int datalen = aio_req->data_len;
> +	uint64_t offset = aio_req->offset;
> +	uint8_t flags = aio_req->flags;
> +	uint64_t old_oid = aio_req->base_oid;
> +
> +	if (!nr_copies)
> +		eprintf("bug\n");
> +
> +	memset(&hdr, 0, sizeof(hdr));
> +
> +	if (aiocb_type == AIOCB_READ_UDATA) {
> +		wlen = 0;
> +		hdr.opcode = SD_OP_READ_OBJ;
> +		hdr.flags = flags;
> +	} else if (create) {
> +		wlen = datalen;
> +		hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
> +		hdr.flags = SD_FLAG_CMD_WRITE | flags;
> +	} else {
> +		wlen = datalen;
> +		hdr.opcode = SD_OP_WRITE_OBJ;
> +		hdr.flags = SD_FLAG_CMD_WRITE | flags;
> +	}
> +
> +	hdr.oid = oid;
> +	hdr.cow_oid = old_oid;
> +	hdr.copies = s->inode.nr_copies;
> +
> +	hdr.data_length = datalen;
> +	hdr.offset = offset;
> +
> +	hdr.id = get_id_from_req(s, aio_req);
> +
> +	opt = 1;
> +	setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
> +
> +	ret = do_write(s->fd, &hdr, sizeof(hdr));
> +	if (ret) {
> +		eprintf("failed to send a req, %m\n");
> +		return -EIO;
> +	}
> +
> +	if (wlen) {
> +		ret = do_writev(s->fd, iov, wlen, aio_req->iov_offset);
> +		if (ret) {
> +			eprintf("failed to send a data, %m\n");
> +			return -EIO;
> +		}
> +	}
> +        opt = 0;
> +        setsockopt(s->fd, SOL_TCP, TCP_CORK, &opt, sizeof(opt));
> +
> +	return 0;
> +}
> +
> +static int read_write_object(int fd, char *buf, uint64_t oid, int copies,
> +			     unsigned int datalen, uint64_t offset,
> +			     int write, int create)
> +{
> +	struct sd_obj_req hdr;
> +	struct sd_obj_rsp *rsp = (struct sd_obj_rsp *)&hdr;
> +	unsigned int wlen, rlen;
> +	int ret;
> +
> +	memset(&hdr, 0, sizeof(hdr));
> +
> +	if (write) {
> +		wlen = datalen;
> +		rlen = 0;
> +		hdr.flags = SD_FLAG_CMD_WRITE;
> +		if (create)
> +			hdr.opcode = SD_OP_CREATE_AND_WRITE_OBJ;
> +		else
> +			hdr.opcode = SD_OP_WRITE_OBJ;
> +	} else {
> +		wlen = 0;
> +		rlen = datalen;
> +		hdr.opcode = SD_OP_READ_OBJ;
> +	}
> +	hdr.oid = oid;
> +	hdr.data_length = datalen;
> +	hdr.offset = offset;
> +	hdr.copies = copies;
> +
> +	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
> +	if (ret) {
> +		eprintf("failed to send a request to the sheep\n");
> +		return -1;
> +	}
> +
> +	switch (rsp->result) {
> +	case SD_RES_SUCCESS:
> +		return 0;
> +	default:
> +		eprintf("%s\n", sd_strerror(rsp->result));
> +		return -1;
> +	}
> +}
> +
> +static int read_object(int fd, char *buf, uint64_t oid, int copies,
> +		       unsigned int datalen, uint64_t offset)
> +{
> +	return read_write_object(fd, buf, oid, copies, datalen, offset, 0, 0);
> +}
> +
> +static int write_object(int fd, char *buf, uint64_t oid, int copies,
> +			unsigned int datalen, uint64_t offset, int create)
> +{
> +	return read_write_object(fd, buf, oid, copies, datalen, offset, 1, create);
> +}
> +
> +/* TODO: error cleanups */
> +static int sd_open(BlockDriverState *bs, const char *filename, int flags)
> +{
> +	int ret, i, fd;
> +	uint32_t vid = 0;
> +	struct bdrv_sd_state *s = bs->opaque;
> +	char vdi[256];
> +	uint32_t snapid;
> +	int for_snapshot = 0;
> +	char *buf;
> +
> +	strstart(filename, "sheepdog:", (const char **)&filename);
> +
> +	buf = qemu_malloc(SD_INODE_SIZE);
> +	if (!buf) {
> +		eprintf("Failed to allocate memory\n");
> +		return -1;
> +	}
> +
> +	for (i = 0; i < MAX_AIO_REQS; i++) {
> +		s->aio_req_free[i] = &s->aio_req_list[i];
> +		s->aio_req_list[i].aiocb = NULL;
> +	}
> +	s->nr_aio_req_free = MAX_AIO_REQS;
> +
> +	memset(vdi, 0, sizeof(vdi));
> +	if (parse_vdiname(s, filename, vdi, sizeof(vdi), &snapid) < 0)
> +		goto out;
> +	s->fd = get_sheep_fd(s);
> +	if (s->fd < 0)
> +		return -1;
> +
> +	if (snapid != CURRENT_VDI_ID)
> +		for_snapshot = 1;
> +
> +	ret = find_vdi_name(s, vdi, snapid, &vid, for_snapshot);
> +	if (ret)
> +		goto out;
> +
> +	if (snapid)
> +		dprintf("%" PRIx32 " non current inode was open.\n", vid);
> +	else
> +		s->is_current = 1;
> +
> +	fd = connect_to_sdog(s->addr);

I wonder why you need to open another connection here instead of using
s->fd. This pattern repeats at least in the snapshot functions, so I'm
sure it's there for a reason. Maybe add a comment?

> +	if (fd < 0) {
> +		eprintf("failed to connect\n");
> +		goto out;
> +	}
> +
> +	ret = read_object(fd, buf, vid_to_vdi_oid(vid), 0, SD_INODE_SIZE, 0);
> +
> +	close(fd);
> +
> +	if (ret)
> +		goto out;
> +
> +	memcpy(&s->inode, buf, sizeof(s->inode));
> +	s->nr_dirty_data_oids = 0;
> +
> +	bs->total_sectors = s->inode.vdi_size >> 9;
> +	strncpy(s->name, vdi, sizeof(s->name));
> +	qemu_free(buf);
> +
> +	QLIST_INIT(&s->pending_head);
> +	return 0;
> +out:
> +	qemu_free(buf);
> +	return -1;
> +}
> +
> +static int do_sd_create(const char *addr, char *filename, char *tag,
> +			int64_t total_sectors, uint32_t base_vid,
> +			uint32_t *vdi_id, int snapshot)
> +{
> +	struct sd_vdi_req hdr;
> +	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
> +	int fd, ret;
> +	unsigned int wlen, rlen = 0;
> +	char buf[SD_MAX_VDI_LEN];
> +
> +	fd = connect_to_sdog(addr);
> +	if (fd < 0)
> +		return -1;
> +
> +	strncpy(buf, filename, SD_MAX_VDI_LEN);
> +
> +	memset(&hdr, 0, sizeof(hdr));
> +	hdr.opcode = SD_OP_NEW_VDI;
> +	hdr.base_vdi_id = base_vid;
> +
> +	wlen = SD_MAX_VDI_LEN;
> +
> +	hdr.flags = SD_FLAG_CMD_WRITE;
> +	hdr.snapid = snapshot;
> +
> +	hdr.data_length = wlen;
> +	hdr.vdi_size = total_sectors * 512;
> +
> +	ret = do_req(fd, (struct sd_req *)&hdr, buf, &wlen, &rlen);
> +
> +	close(fd);
> +
> +	if (ret)
> +		return -1;
> +
> +	if (rsp->result != SD_RES_SUCCESS) {
> +		eprintf("%s, %s\n", sd_strerror(rsp->result), filename);
> +		return -1;
> +	}
> +
> +	if (vdi_id)
> +		*vdi_id = rsp->vdi_id;
> +
> +	return 0;
> +}
> +
> +static int sd_create(const char *filename, QEMUOptionParameter *options)
> +{
> +	int ret;
> +	uint32_t vid = 0;
> +	int64_t total_sectors = 0;
> +	char *backing_file = NULL;
> +
> +	strstart(filename, "sheepdog:", (const char **)&filename);
> +
> +	while (options && options->name) {
> +		if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
> +			total_sectors = options->value.n / 512;
> +		} else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
> +			backing_file = options->value.s;
> +		}
> +		options++;
> +	}
> +
> +	if (backing_file) {
> +		BlockDriverState bs;
> +		char vdi[SD_MAX_VDI_LEN];
> +		uint32_t snapid;
> +
> +		strstart(backing_file, "sheepdog:", (const char **)&backing_file);
> +		memset(&bs, 0, sizeof(bs));
> +
> +		bs.opaque = qemu_malloc(sizeof(struct bdrv_sd_state));
> +		if (!bs.opaque)
> +			return -1;
> +
> +		ret = sd_open(&bs, backing_file, 0);
> +		if (ret < 0)
> +			return -1;
> +
> +		if (parse_vdiname(bs.opaque, backing_file, vdi, sizeof(vdi), &snapid) < 0)
> +			return -1;
> +
> +		/* cannot clone from a current inode */
> +		if (snapid == CURRENT_VDI_ID)
> +			return -1;
> +
> +		ret = find_vdi_name(bs.opaque, vdi, snapid, &vid, 1);
> +		if (ret)
> +			return -1;
> +	}
> +
> +	return do_sd_create(NULL, (char *)filename, NULL, total_sectors, vid,
> +			    NULL, 0);
> +}
> +
> +static void sd_close(BlockDriverState *bs)
> +{
> +	struct bdrv_sd_state *s = bs->opaque;
> +	struct sd_vdi_req hdr;
> +	struct sd_vdi_rsp *rsp = (struct sd_vdi_rsp *)&hdr;
> +	unsigned int wlen, rlen = 0;
> +	int fd, ret;
> +
> +	dprintf("%s\n", s->name);
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0)
> +		return;
> +
> +	memset(&hdr, 0, sizeof(hdr));
> +
> +	hdr.opcode = SD_OP_RELEASE_VDI;
> +	wlen = strlen(s->name) + 1;
> +	hdr.data_length = wlen;
> +	hdr.flags = SD_FLAG_CMD_WRITE;
> +
> +	ret = do_req(fd, (struct sd_req *)&hdr, s->name, &wlen, &rlen);
> +
> +	close(fd);
> +
> +	if (!ret && rsp->result != SD_RES_SUCCESS &&
> +	    rsp->result != SD_RES_VDI_NOT_LOCKED)
> +		eprintf("%s, %s\n", sd_strerror(rsp->result), s->name);
> +
> +	close(s->fd);
> +	free(s->addr);
> +}
> +
> +static void sd_write_done(struct sd_aiocb *acb)
> +{
> +	int ret, i;
> +	struct bdrv_sd_state *s = acb->common.bs->opaque;
> +	struct iovec iov;
> +	struct aio_req *aio_req;
> +	uint32_t offset, data_len, mn, mx;
> +
> +	if (s->nr_dirty_data_oids) {
> +		mn = mx = s->dirty_data_oids[0];
> +		for (i = 0; i < s->nr_dirty_data_oids; i++) {
> +			if (mn > s->dirty_data_oids[i])
> +				mn = s->dirty_data_oids[i];
> +			if (mx < s->dirty_data_oids[i])
> +				mx = s->dirty_data_oids[i];
> +		}
> +		offset = sizeof(s->inode) - sizeof(s->inode.data_vdi_id) +
> +			mn * sizeof(s->inode.data_vdi_id[0]);
> +		data_len = (mx - mn + 1) * sizeof(s->inode.data_vdi_id[0]);
> +		s->nr_dirty_data_oids = 0;
> +
> +		iov.iov_base = &s->inode;
> +		iov.iov_len = sizeof(s->inode);
> +		aio_req = alloc_aio_req(s, acb, vid_to_vdi_oid(s->inode.vdi_id),
> +					data_len, offset, 0, 0, offset);
> +		if (!aio_req) {
> +			eprintf("too many requests\n");
> +			acb->ret = -EIO;
> +			goto out;
> +		}

Randomly failing requests is probably not a good idea. The guest might
decide that the disk/file system is broken and stop using it. Can't you
use a list like in AIOPool, so you can dynamically add new requests as
needed?

> +		ret = add_aio_request(s, aio_req, &iov, 1, 0, AIOCB_WRITE_UDATA);
> +		if (ret) {
> +			free_aio_req(s, aio_req);
> +			acb->ret = -EIO;
> +			goto out;
> +		}
> +
> +		acb->aio_done_func = sd_finish_aiocb;
> +		acb->aiocb_type = AIOCB_WRITE_UDATA;
> +		return;
> +	}
> +out:
> +	sd_finish_aiocb(acb);
> +}
> +
> +static int sd_create_branch(struct bdrv_sd_state *s)
> +{
> +	int ret, fd;
> +	uint32_t vid;
> +	char *buf;
> +
> +	dprintf("%" PRIx32 " is not current.\n", s->inode.vdi_id);
> +
> +	buf = qemu_malloc(SD_INODE_SIZE);
> +	if (!buf)
> +		return -1;
> +
> +	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
> +			   s->inode.vdi_id, &vid, 1);
> +	if (ret)
> +		goto out;
> +
> +	dprintf("%" PRIx32 " is created.\n", vid);
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0) {
> +		eprintf("failed to connect\n");
> +		goto out;
> +	}
> +
> +	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
> +			  SD_INODE_SIZE, 0);
> +
> +	close(fd);
> +
> +	if (ret < 0)
> +		goto out;
> +
> +	memcpy(&s->inode, buf, sizeof(s->inode));
> +
> +	s->is_current = 1;
> +	ret = 0;
> +	dprintf("%" PRIx32 " was newly created.\n", s->inode.vdi_id);
> +
> +out:
> +	qemu_free(buf);
> +
> +	return ret;
> +}
> +
> +static void sd_readv_writev_bh_cb(void *p)
> +{
> +	struct sd_aiocb *acb = p;
> +	int ret = 0;
> +	unsigned long len, done = 0, total = acb->nb_sectors * 512;
> +	unsigned long idx = acb->sector_num * 512 / SD_DATA_OBJ_SIZE;
> +	uint64_t oid;
> +	uint64_t offset = (acb->sector_num * 512) % SD_DATA_OBJ_SIZE;
> +	struct bdrv_sd_state *s = acb->common.bs->opaque;
> +	struct sd_inode *inode = &s->inode;
> +	struct aio_req *aio_req;
> +
> +	qemu_bh_delete(acb->bh);
> +	acb->bh = NULL;
> +
> +	if (acb->aiocb_type == AIOCB_WRITE_UDATA && !s->is_current) {
> +		ret = sd_create_branch(s);
> +		if (ret) {
> +			acb->ret = -EIO;
> +			goto out;
> +		}
> +	}
> +
> +	while (done != total) {
> +		uint8_t flags = 0;
> +		uint64_t old_oid = 0;
> +		int create = 0;
> +
> +		oid = vid_to_data_oid(inode->data_vdi_id[idx], idx);
> +
> +		len = min_t(unsigned long, total - done, SD_DATA_OBJ_SIZE - offset);
> +
> +		if (!inode->data_vdi_id[idx]) {
> +			if (acb->aiocb_type == AIOCB_READ_UDATA)
> +				goto done;
> +
> +			create = 1;
> +		} else if (acb->aiocb_type == AIOCB_WRITE_UDATA
> +			   && !is_data_obj_writeable(inode, idx)) {
> +			create = 1;
> +			old_oid = oid;
> +			flags = SD_FLAG_CMD_COW;
> +		}
> +
> +		if (create) {
> +			dprintf("update ino (%" PRIu32") %"
> +				PRIu64 " %" PRIu64 " %" PRIu64 "\n",
> +				inode->vdi_id, oid,
> +				vid_to_data_oid(inode->data_vdi_id[idx], idx), idx);
> +			oid = vid_to_data_oid(inode->vdi_id, idx);
> +			dprintf("new oid %lx\n", oid);
> +		}
> +
> +		aio_req = alloc_aio_req(s, acb, oid, len, offset, flags,
> +					old_oid, done);
> +		if (!aio_req) {
> +			eprintf("too many requests\n");
> +			acb->ret = -EIO;
> +			goto out;
> +		}
> +
> +		if (create) {
> +			struct aio_req *areq;
> +			QLIST_FOREACH(areq, &s->pending_head, pending_siblings) {
> +				if (get_id_from_req(s, areq) == get_id_from_req(s, aio_req))
> +					continue;
> +				if (areq->oid == oid) {
> +					aio_req->flags = 0;
> +					aio_req->base_oid = 0;
> +					goto done;
> +				}
> +			}
> +		}
> +
> +		ret = add_aio_request(s, aio_req, acb->qiov->iov, acb->qiov->niov,
> +				      create, acb->aiocb_type);
> +		if (ret < 0) {
> +			eprintf("add_aio_request is faled\n");
> +			free_aio_req(s, aio_req);
> +			acb->ret = -EIO;
> +			goto out;
> +		}
> +	done:
> +		offset = 0;
> +		idx++;
> +		done += len;
> +	}
> +out:
> +	if (QLIST_EMPTY(&acb->aioreq_head))
> +		sd_finish_aiocb(acb);
> +}
> +
> +static BlockDriverAIOCB *sd_aio_writev(BlockDriverState *bs,
> +				       int64_t sector_num,
> +				       QEMUIOVector *qiov,
> +				       int nb_sectors,
> +				       BlockDriverCompletionFunc *cb,
> +				       void *opaque)
> +{
> +	struct sd_aiocb *acb;
> +
> +	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
> +	acb->aio_done_func = sd_write_done;
> +	acb->aiocb_type = AIOCB_WRITE_UDATA;
> +
> +	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
> +	return &acb->common;
> +}
> +
> +static BlockDriverAIOCB *sd_aio_readv(BlockDriverState *bs,
> +				      int64_t sector_num,
> +				      QEMUIOVector *qiov,
> +				      int nb_sectors,
> +				      BlockDriverCompletionFunc *cb,
> +				      void *opaque)
> +{
> +	struct sd_aiocb *acb;
> +	int i;
> +
> +	acb = sd_aio_setup(bs, qiov, sector_num, nb_sectors, cb, opaque);
> +	acb->aiocb_type = AIOCB_READ_UDATA;
> +	acb->aio_done_func = sd_finish_aiocb;
> +
> +	/*
> +	 * TODO: we can do better; we don't need to initialize
> +	 * blindly.
> +	 */
> +	for (i = 0; i < qiov->niov; i++)
> +		memset(qiov->iov[i].iov_base, 0, qiov->iov[i].iov_len);
> +
> +	sd_schedule_bh(sd_readv_writev_bh_cb, acb);
> +	return &acb->common;
> +}
> +
> +static int sd_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info)
> +{
> +	struct bdrv_sd_state *s = bs->opaque;
> +	int ret, fd;
> +	uint32_t new_vid;
> +	struct sd_inode *inode;
> +	unsigned int datalen;
> +	uint64_t offset;
> +
> +	dprintf("sn_info: name %s id_str %s s: name %s vm_state_size %d "
> +		"is_current %d\n", sn_info->name, sn_info->id_str,
> +		s->name, sn_info->vm_state_size, s->is_current);
> +
> +	if (!s->is_current) {
> +		eprintf("You can't create a snapshot of "
> +			"a non current VDI, %s (%" PRIu32 ").\n",
> +			s->name, s->inode.vdi_id);
> +
> +		return -1;
> +	}
> +
> +	dprintf("%s %s\n", sn_info->name, sn_info->id_str);
> +
> +	s->inode.vm_state_size = sn_info->vm_state_size;
> +	s->inode.vm_clock_nsec = sn_info->vm_clock_nsec;
> +	offset = 0;
> +	/* we don't need to read entire object */
> +	datalen = SD_INODE_SIZE - sizeof(s->inode.data_vdi_id);
> +
> +	/* refresh inode. */
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0) {
> +		ret = -EIO;
> +		goto cleanup;
> +	}
> +
> +	ret = write_object(fd, (char *)&s->inode, vid_to_vdi_oid(s->inode.vdi_id),
> +			   s->inode.nr_copies, datalen, offset, 0);
> +	if (ret < 0) {
> +		eprintf("failed to write snapshot's inode.\n");
> +		ret = -EIO;
> +		goto cleanup;
> +	}
> +
> +	ret = do_sd_create(s->addr, s->name, NULL, s->inode.vdi_size >> 9,
> +			   s->inode.vdi_id, &new_vid, 1);
> +	if (ret < 0) {
> +		eprintf("failed to create inode for snapshot. %m\n");
> +		ret = -EIO;
> +		goto cleanup;
> +	}
> +
> +	inode = (struct sd_inode *)qemu_malloc(datalen);
> +	if (!inode) {
> +		eprintf("failed to allocate memory for inode. %m\n");
> +		goto cleanup;
> +	}
> +
> +	ret = read_object(fd, (char *)inode, vid_to_vdi_oid(new_vid),
> +			  s->inode.nr_copies, datalen, offset);
> +
> +	close(fd);
> +
> +	if (ret < 0) {
> +		eprintf("failed to read new inode info. %m\n");
> +		ret = -EIO;
> +		goto cleanup;
> +	}
> +
> +	memcpy(&s->inode, inode, datalen);
> +	dprintf("s->inode: name %s snap_id %x oid %x\n",
> +		s->inode.name, s->inode.snap_id, s->inode.vdi_id);
> +
> +cleanup:
> +	close(fd);
> +	return ret;
> +}
> +
> +static int sd_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
> +{
> +	struct bdrv_sd_state *s = bs->opaque;
> +	struct bdrv_sd_state *old_s;
> +	char vdi[SD_MAX_VDI_LEN];
> +	char *buf = NULL;
> +	uint32_t vid;
> +	uint32_t snapid = 0;
> +	int ret = -ENOENT, fd;
> +
> +	old_s = qemu_malloc(sizeof(struct bdrv_sd_state));
> +	if (!old_s) {

qemu_malloc never returns NULL.

> +		eprintf("failed to allocate memory for old state. %m\n");
> +		goto out;
> +	}
> +
> +	memcpy(old_s, s, sizeof(struct bdrv_sd_state));
> +
> +	snapid = strtol(snapshot_id, NULL, 10);
> +	if (!snapid) {
> +		eprintf("Invalid snapshot_id\n");
> +		goto out;
> +	}
> +
> +	buf = qemu_malloc(SD_INODE_SIZE);
> +	if (!buf) {
> +		eprintf("Failed to allocate memory\n");
> +		goto out;
> +	}
> +	strncpy(vdi, s->name, sizeof(vdi));
> +	ret = find_vdi_name(s, vdi, snapid, &vid, 1);
> +	if (ret) {
> +		eprintf("Failed to find_vdi_name\n");
> +		ret = -ENOENT;
> +		goto out;
> +	}
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0) {
> +		eprintf("failed to connect\n");
> +		goto out;
> +	}
> +
> +	ret = read_object(fd, buf, vid_to_vdi_oid(vid), s->inode.nr_copies,
> +			  SD_INODE_SIZE, 0);
> +
> +	close(fd);
> +
> +	if (ret) {
> +		ret = -ENOENT;
> +		goto out;
> +	}
> +
> +	memcpy(&s->inode, buf, sizeof(s->inode));
> +
> +	if (!s->inode.vm_state_size) {
> +		eprintf("Invalid snapshot\n");
> +		ret = -ENOENT;
> +		goto out;
> +	}
> +
> +	s->is_current = 0;
> +
> +	qemu_free(buf);
> +	qemu_free(old_s);
> +
> +	return 0;
> +out:
> +	/* recover bdrv_sd_state */
> +	memcpy(s, old_s, sizeof(struct bdrv_sd_state));
> +	qemu_free(buf);
> +	qemu_free(old_s);
> +
> +	eprintf("failed to open. recover old bdrv_sd_state.\n");
> +
> +	return ret;
> +}
> +
> +static int sd_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
> +{
> +	/* FIXME: Delete specified snapshot id.  */
> +	return 0;
> +}

Ok, obviously there's something missing. ;-)

> +
> +#define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
> +#define BITS_PER_BYTE		8
> +#define BITS_TO_LONGS(nr)	DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(long))
> +#define DECLARE_BITMAP(name,bits) \
> +	unsigned long name[BITS_TO_LONGS(bits)]
> +
> +#define BITS_PER_LONG (BITS_PER_BYTE * sizeof(long))
> +
> +static inline int test_bit(unsigned int nr, const unsigned long *addr)
> +{
> +	return ((1UL << (nr % BITS_PER_LONG)) &
> +		(((unsigned long *)addr)[nr / BITS_PER_LONG])) != 0;
> +}
> +
> +static int sd_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab)
> +{
> +	struct bdrv_sd_state *s = bs->opaque;
> +	struct sd_req req;
> +	int i, fd, nr = 1024, ret, max = BITS_TO_LONGS(SD_NR_VDIS) * sizeof(long);
> +	QEMUSnapshotInfo *sn_tab = NULL;
> +	unsigned wlen, rlen;
> +	int found = 0;
> +	static struct sd_inode inode;
> +	unsigned long *vdi_inuse;
> +	unsigned int start_nr;
> +
> +	vdi_inuse = qemu_malloc(max);
> +	if (!vdi_inuse)
> +		return 0;
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0)
> +		goto out;
> +
> +	rlen = max;
> +	wlen = 0;
> +
> +	memset(&req, 0, sizeof(req));
> +
> +	req.opcode = SD_OP_READ_VDIS;
> +	req.data_length = max;
> +
> +	ret = do_req(fd, (struct sd_req *)&req, vdi_inuse, &wlen, &rlen);
> +
> +	close(fd);
> +	if (ret)
> +		goto out;
> +
> +	sn_tab = qemu_mallocz(nr * sizeof(*sn_tab));
> +	if (!sn_tab)
> +		goto out;
> +
> +	start_nr = fnv_64a_buf(s->name, strlen(s->name), FNV1A_64_INIT) & (SD_NR_VDIS - 1);
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0) {
> +		eprintf("failed to connect\n");
> +		goto out;
> +	}
> +
> +	/* TODO: round up */
> +	for (i = start_nr; i < SD_NR_VDIS && found < nr; i++) {
> +		if (!test_bit(i, vdi_inuse))
> +			break;
> +
> +		/* we don't need to read entire object */
> +		ret = read_object(fd, (char *)&inode, vid_to_vdi_oid(i),
> +				  0, SD_INODE_SIZE - sizeof(inode.data_vdi_id), 0);
> +
> +		if (ret)
> +			continue;
> +
> +		if (!strcmp(inode.name, s->name) && inode.snap_ctime) {
> +			sn_tab[found].date_sec = inode.snap_ctime >> 32;
> +			sn_tab[found].date_nsec = inode.snap_ctime & 0xffffffff;
> +			sn_tab[found].vm_state_size = inode.vm_state_size;
> +			sn_tab[found].vm_clock_nsec = inode.vm_clock_nsec;
> +
> +			snprintf(sn_tab[found].id_str, sizeof(sn_tab[found].id_str), "%u",
> +				 inode.snap_id);
> +			found++;
> +		}
> +	}
> +
> +	close(fd);
> +out:
> +	*psn_tab = sn_tab;
> +
> +	qemu_free(vdi_inuse);
> +
> +	return found;
> +}
> +
> +static int do_load_save_vmstate(struct bdrv_sd_state *s, uint8_t *data,
> +				int64_t pos, int size, int load)
> +{
> +	int fd, create;
> +	int ret = 0;
> +	unsigned int data_len;
> +	uint64_t vmstate_oid;
> +	uint32_t vdi_index;
> +	uint64_t offset;
> +
> +	fd = connect_to_sdog(s->addr);
> +	if (fd < 0) {
> +		ret = -EIO;
> +		goto cleanup;
> +	}
> +
> +	while (size) {
> +		vdi_index = pos / SD_DATA_OBJ_SIZE;
> +		offset = pos % SD_DATA_OBJ_SIZE;
> +
> +		data_len = min_t(unsigned int, size, SD_DATA_OBJ_SIZE);
> +
> +		vmstate_oid = vid_to_vmstate_oid(s->inode.vdi_id, vdi_index);
> +
> +		create = (offset == 0);
> +		if (load)
> +			ret = read_object(fd, (char *)data, vmstate_oid,
> +					  s->inode.nr_copies, data_len, offset);
> +		else
> +			ret = write_object(fd, (char *)data, vmstate_oid,
> +					   s->inode.nr_copies, data_len, offset, create);
> +
> +		if (ret < 0) {
> +			eprintf("failed to save vmstate %m\n");
> +			ret = -EIO;
> +			goto cleanup;
> +		}
> +
> +		pos += data_len;
> +		size -= data_len;
> +		ret += data_len;
> +	}
> +cleanup:
> +	close(fd);
> +	return ret;
> +}
> +
> +static int sd_save_vmstate(BlockDriverState *bs, const uint8_t *data,
> +			   int64_t pos, int size)
> +{
> +	struct bdrv_sd_state *s = bs->opaque;
> +
> +	return do_load_save_vmstate(s, (uint8_t *)data, pos, size, 0);
> +}
> +
> +static int sd_load_vmstate(BlockDriverState *bs, uint8_t *data,
> +			   int64_t pos, int size)
> +{
> +	struct bdrv_sd_state *s = bs->opaque;
> +
> +	return do_load_save_vmstate(s, data, pos, size, 1);
> +}
> +
> +
> +static QEMUOptionParameter sd_create_options[] = {
> +	{
> +		.name = BLOCK_OPT_SIZE,
> +		.type = OPT_SIZE,
> +		.help = "Virtual disk size"
> +	},
> +	{
> +		.name = BLOCK_OPT_BACKING_FILE,
> +		.type = OPT_STRING,
> +		.help = "File name of a base image"
> +	},
> +	{ NULL }
> +};
> +
> +BlockDriver bdrv_sheepdog = {
> +	.format_name = "sheepdog",
> +	.protocol_name = "sheepdog",
> +	.instance_size = sizeof(struct bdrv_sd_state),
> +	.bdrv_file_open = sd_open,
> +	.bdrv_close = sd_close,
> +	.bdrv_create = sd_create,
> +
> +	.bdrv_aio_readv = sd_aio_readv,
> +	.bdrv_aio_writev = sd_aio_writev,
> +
> +	.bdrv_snapshot_create = sd_snapshot_create,
> +	.bdrv_snapshot_goto = sd_snapshot_goto,
> +	.bdrv_snapshot_delete = sd_snapshot_delete,
> +	.bdrv_snapshot_list = sd_snapshot_list,
> +
> +	.bdrv_save_vmstate = sd_save_vmstate,
> +	.bdrv_load_vmstate = sd_load_vmstate,
> +
> +	.create_options = sd_create_options,
> +};

Please align the = to the same column, at least in each block.

Kevin

Re: [RFC PATCH v2 3/3] block: add sheepdog driver for distributed storage support

[MORITA Kazutaka]

Hi,

Thank you very much for the reviewing!

At Fri, 14 May 2010 13:08:06 +0200,
Kevin Wolf wrote:

> > +
> > +struct sd_req {
> > +	uint8_t		proto_ver;
> > +	uint8_t		opcode;
> > +	uint16_t	flags;
> > +	uint32_t	epoch;
> > +	uint32_t        id;
> > +	uint32_t        data_length;
> > +	uint32_t	opcode_specific[8];
> > +};
> 
> CODING_STYLE says that structs should be typedefed and their names
> should be in CamelCase. So something like this:
> 
> typedef struct SheepdogReq {
>     ...
> } SheepdogReq;
> 
> (Or, if your prefer, SDReq; but with things like SDAIOCB I think it
> becomes hard to read)
> 

I see. I use Sheepdog as a prefix, like SheepdogReq.


> > +/*
> > +
> > +#undef eprintf
> > +#define eprintf(fmt, args...)						\
> > +do {									\
> > +	fprintf(stderr, "%s %d: " fmt, __func__, __LINE__, ##args);	\
> > +} while (0)
> 
> What about using error_report() instead of fprintf? Though it should be
> the same currently.
> 

Yes, using common helper functions is better.  I replaced all the
printf.


> > +
> > +	for (i = 0; i < ARRAY_SIZE(errors); ++i)
> > +		if (errors[i].err == err)
> > +			return errors[i].desc;
> 
> CODING_STYLE requires braces here.
> 

I fixed all the missing braces.


> > +
> > +	return "Invalid error code";
> > +}
> > +
> > +static inline int before(uint32_t seq1, uint32_t seq2)
> > +{
> > +        return (int32_t)(seq1 - seq2) < 0;
> > +}
> > +
> > +static inline int after(uint32_t seq1, uint32_t seq2)
> > +{
> > +	return (int32_t)(seq2 - seq1) < 0;
> > +}
> 
> These functions look strange... Is the difference to seq1 < seq2 that
> the cast introduces intentional? (after(0x0, 0xabcdefff) == 1)
> 
> If yes, why is this useful? This needs a comment. If no, why even bother
> to have this function instead of directly using < or > ?
> 

These functions are used to compare sequential numbers which can be
wrap-around. For example, linux/net/tcp.h in the linux kernel.

Anyway, sheepdog doesn't use these functions, so I removed them.


> > +	if (snapid)
> > +		dprintf("%" PRIx32 " non current inode was open.\n", vid);
> > +	else
> > +		s->is_current = 1;
> > +
> > +	fd = connect_to_sdog(s->addr);
> 
> I wonder why you need to open another connection here instead of using
> s->fd. This pattern repeats at least in the snapshot functions, so I'm
> sure it's there for a reason. Maybe add a comment?
> 

We can use s->fd only for aio read/write operations.  It is because
the block driver may be during waiting response from the server, so we
cannot send other requests to the discriptor to avoid receiving wrong
data.

I added the comment to get_sheep_fd().


> > +
> > +		iov.iov_base = &s->inode;
> > +		iov.iov_len = sizeof(s->inode);
> > +		aio_req = alloc_aio_req(s, acb, vid_to_vdi_oid(s->inode.vdi_id),
> > +					data_len, offset, 0, 0, offset);
> > +		if (!aio_req) {
> > +			eprintf("too many requests\n");
> > +			acb->ret = -EIO;
> > +			goto out;
> > +		}
> 
> Randomly failing requests is probably not a good idea. The guest might
> decide that the disk/file system is broken and stop using it. Can't you
> use a list like in AIOPool, so you can dynamically add new requests as
> needed?
> 

I agree.  In the v3 patch, AIO requests are allocated dynamically, and
all the requests are linked to the outstanding_aio_head in the
BDRVSheepdogState.


> > +
> > +static int sd_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
> > +{
> > +	struct bdrv_sd_state *s = bs->opaque;
> > +	struct bdrv_sd_state *old_s;
> > +	char vdi[SD_MAX_VDI_LEN];
> > +	char *buf = NULL;
> > +	uint32_t vid;
> > +	uint32_t snapid = 0;
> > +	int ret = -ENOENT, fd;
> > +
> > +	old_s = qemu_malloc(sizeof(struct bdrv_sd_state));
> > +	if (!old_s) {
> 
> qemu_malloc never returns NULL.
> 

I removed all the NULL checks.


> > +
> > +BlockDriver bdrv_sheepdog = {
> > +	.format_name = "sheepdog",
> > +	.protocol_name = "sheepdog",
> > +	.instance_size = sizeof(struct bdrv_sd_state),
> > +	.bdrv_file_open = sd_open,
> > +	.bdrv_close = sd_close,
> > +	.bdrv_create = sd_create,
> > +
> > +	.bdrv_aio_readv = sd_aio_readv,
> > +	.bdrv_aio_writev = sd_aio_writev,
> > +
> > +	.bdrv_snapshot_create = sd_snapshot_create,
> > +	.bdrv_snapshot_goto = sd_snapshot_goto,
> > +	.bdrv_snapshot_delete = sd_snapshot_delete,
> > +	.bdrv_snapshot_list = sd_snapshot_list,
> > +
> > +	.bdrv_save_vmstate = sd_save_vmstate,
> > +	.bdrv_load_vmstate = sd_load_vmstate,
> > +
> > +	.create_options = sd_create_options,
> > +};
> 
> Please align the = to the same column, at least in each block.
> 

I have aligned in the v3 patch.


Thanks,

Kazutaka