[PATCH for-5.0] xen-block: Fix double qlist remove

[PATCH for-5.0] xen-block: Fix double qlist remove

[Anthony PERARD]

Commit a31ca6801c02 ("qemu/queue.h: clear linked list pointers on
remove") revealed that a request was removed twice from a list, once
in xen_block_finish_request() and a second time in
xen_block_release_request() when both function are called from
xen_block_complete_aio(). But also, the `requests_inflight' counter is
decreased twice, and thus became negative.

This is a bug that was introduced in bfd0d6366043, where a `finished'
list was removed.

This patch simply re-add the `finish' parameter of
xen_block_release_request() so that we can distinguish when we need to
remove a request from the inflight list and when not.

Fixes: bfd0d6366043 ("xen-block: improve response latency")
Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>
---
 hw/block/dataplane/xen-block.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/hw/block/dataplane/xen-block.c b/hw/block/dataplane/xen-block.c
index 288a87a814ad..6cc089fc561f 100644
--- a/hw/block/dataplane/xen-block.c
+++ b/hw/block/dataplane/xen-block.c
@@ -123,15 +123,19 @@ static void xen_block_finish_request(XenBlockRequest *request)
     dataplane->requests_inflight--;
 }
 
-static void xen_block_release_request(XenBlockRequest *request)
+static void xen_block_release_request(XenBlockRequest *request, bool finish)
 {
     XenBlockDataPlane *dataplane = request->dataplane;
 
-    QLIST_REMOVE(request, list);
+    if (!finish) {
+        QLIST_REMOVE(request, list);
+    }
     reset_request(request);
     request->dataplane = dataplane;
     QLIST_INSERT_HEAD(&dataplane->freelist, request, list);
-    dataplane->requests_inflight--;
+    if (!finish) {
+        dataplane->requests_inflight--;
+    }
 }
 
 /*
@@ -316,7 +320,7 @@ static void xen_block_complete_aio(void *opaque, int ret)
             error_report_err(local_err);
         }
     }
-    xen_block_release_request(request);
+    xen_block_release_request(request, true);
 
     if (dataplane->more_work) {
         qemu_bh_schedule(dataplane->bh);
@@ -585,7 +589,7 @@ static bool xen_block_handle_requests(XenBlockDataPlane *dataplane)
                     error_report_err(local_err);
                 }
             }
-            xen_block_release_request(request);
+            xen_block_release_request(request, false);
             continue;
         }
 
-- 
Anthony PERARD

RE: [PATCH for-5.0] xen-block: Fix double qlist remove

[Paul Durrant]

> -----Original Message-----
> From: Anthony PERARD <anthony.perard@citrix.com>
> Sent: 02 April 2020 14:08
> To: qemu-devel@nongnu.org
> Cc: qemu-stable@nongnu.org; Anthony PERARD <anthony.perard@citrix.com>; Stefano Stabellini
> <sstabellini@kernel.org>; Paul Durrant <paul@xen.org>; Stefan Hajnoczi <stefanha@redhat.com>; Kevin
> Wolf <kwolf@redhat.com>; Max Reitz <mreitz@redhat.com>; xen-devel@lists.xenproject.org; qemu-
> block@nongnu.org
> Subject: [PATCH for-5.0] xen-block: Fix double qlist remove
> 
> Commit a31ca6801c02 ("qemu/queue.h: clear linked list pointers on
> remove") revealed that a request was removed twice from a list, once
> in xen_block_finish_request() and a second time in
> xen_block_release_request() when both function are called from
> xen_block_complete_aio(). But also, the `requests_inflight' counter is
> decreased twice, and thus became negative.
> 
> This is a bug that was introduced in bfd0d6366043, where a `finished'
> list was removed.
> 
> This patch simply re-add the `finish' parameter of
> xen_block_release_request() so that we can distinguish when we need to
> remove a request from the inflight list and when not.
> 
> Fixes: bfd0d6366043 ("xen-block: improve response latency")
> Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>

It looks to me like it would just be more straightforward to simply drop the QLIST_REMOVE and requests_inflight-- from
xen_block_release_request() and simply insist that xen_block_finish_request() is called in all cases (which I think means adding one
extra call to it in xen_block_handle_requests()).

  Paul

> ---
>  hw/block/dataplane/xen-block.c | 14 +++++++++-----
>  1 file changed, 9 insertions(+), 5 deletions(-)
> 
> diff --git a/hw/block/dataplane/xen-block.c b/hw/block/dataplane/xen-block.c
> index 288a87a814ad..6cc089fc561f 100644
> --- a/hw/block/dataplane/xen-block.c
> +++ b/hw/block/dataplane/xen-block.c
> @@ -123,15 +123,19 @@ static void xen_block_finish_request(XenBlockRequest *request)
>      dataplane->requests_inflight--;
>  }
> 
> -static void xen_block_release_request(XenBlockRequest *request)
> +static void xen_block_release_request(XenBlockRequest *request, bool finish)
>  {
>      XenBlockDataPlane *dataplane = request->dataplane;
> 
> -    QLIST_REMOVE(request, list);
> +    if (!finish) {
> +        QLIST_REMOVE(request, list);
> +    }
>      reset_request(request);
>      request->dataplane = dataplane;
>      QLIST_INSERT_HEAD(&dataplane->freelist, request, list);
> -    dataplane->requests_inflight--;
> +    if (!finish) {
> +        dataplane->requests_inflight--;
> +    }
>  }
> 
>  /*
> @@ -316,7 +320,7 @@ static void xen_block_complete_aio(void *opaque, int ret)
>              error_report_err(local_err);
>          }
>      }
> -    xen_block_release_request(request);
> +    xen_block_release_request(request, true);
> 
>      if (dataplane->more_work) {
>          qemu_bh_schedule(dataplane->bh);
> @@ -585,7 +589,7 @@ static bool xen_block_handle_requests(XenBlockDataPlane *dataplane)
>                      error_report_err(local_err);
>                  }
>              }
> -            xen_block_release_request(request);
> +            xen_block_release_request(request, false);
>              continue;
>          }
> 
> --
> Anthony PERARD

Re: [PATCH for-5.0] xen-block: Fix double qlist remove

[Anthony PERARD]

On Thu, Apr 02, 2020 at 03:27:22PM +0100, Paul Durrant wrote:
> > -----Original Message-----
> > From: Anthony PERARD <anthony.perard@citrix.com>
> > Sent: 02 April 2020 14:08
> > To: qemu-devel@nongnu.org
> > Cc: qemu-stable@nongnu.org; Anthony PERARD <anthony.perard@citrix.com>; Stefano Stabellini
> > <sstabellini@kernel.org>; Paul Durrant <paul@xen.org>; Stefan Hajnoczi <stefanha@redhat.com>; Kevin
> > Wolf <kwolf@redhat.com>; Max Reitz <mreitz@redhat.com>; xen-devel@lists.xenproject.org; qemu-
> > block@nongnu.org
> > Subject: [PATCH for-5.0] xen-block: Fix double qlist remove
> > 
> > Commit a31ca6801c02 ("qemu/queue.h: clear linked list pointers on
> > remove") revealed that a request was removed twice from a list, once
> > in xen_block_finish_request() and a second time in
> > xen_block_release_request() when both function are called from
> > xen_block_complete_aio(). But also, the `requests_inflight' counter is
> > decreased twice, and thus became negative.
> > 
> > This is a bug that was introduced in bfd0d6366043, where a `finished'
> > list was removed.
> > 
> > This patch simply re-add the `finish' parameter of
> > xen_block_release_request() so that we can distinguish when we need to
> > remove a request from the inflight list and when not.
> > 
> > Fixes: bfd0d6366043 ("xen-block: improve response latency")
> > Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>
> 
> It looks to me like it would just be more straightforward to simply drop the QLIST_REMOVE and requests_inflight-- from
> xen_block_release_request() and simply insist that xen_block_finish_request() is called in all cases (which I think means adding one
> extra call to it in xen_block_handle_requests()).

I'm thinking of going further than that. I've notice another bug, in
case of error in xen_block_do_aio(), xen_block_finish_request() is
called without ever calling send_response() or release_request(). I
think that mean a leak of request.

So, I'm thinking of creating a function that would do finish_request(),
send_response(), release_request(), has I believe those operations needs
to be done together anyway.

I'll rework the patch.

-- 
Anthony PERARD

[PATCH v2 for-5.0] xen-block: Fix double qlist remove and request leak

[Anthony PERARD]

Commit a31ca6801c02 ("qemu/queue.h: clear linked list pointers on
remove") revealed that a request was removed twice from a list, once
in xen_block_finish_request() and a second time in
xen_block_release_request() when both function are called from
xen_block_complete_aio(). But also, the `requests_inflight' counter is
decreased twice, and thus became negative.

This is a bug that was introduced in bfd0d6366043, where a `finished'
list was removed.

That commit also introduced a leak of request in xen_block_do_aio().
That function calls xen_block_finish_request() but the request is
never released after that.

To fix both issue, we do two changes:
- we squash finish_request() and release_request() together as we want
  to remove a request from 'inflight' list to add it to 'freelist'.
- before releasing a request, we need to let now the result to the
  other end, thus we should call xen_block_send_response() before
  releasing a request.

The first change fix the double QLIST_REMOVE() as we remove the extra
call. The second change makes the leak go away because if we want to
call finish_request(), we need to call a function that do all of
finish, send response, and release.

Fixes: bfd0d6366043 ("xen-block: improve response latency")
Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>
---
 hw/block/dataplane/xen-block.c | 48 ++++++++++++----------------------
 1 file changed, 16 insertions(+), 32 deletions(-)

diff --git a/hw/block/dataplane/xen-block.c b/hw/block/dataplane/xen-block.c
index 288a87a814ad..5f8f15778ba5 100644
--- a/hw/block/dataplane/xen-block.c
+++ b/hw/block/dataplane/xen-block.c
@@ -64,6 +64,8 @@ struct XenBlockDataPlane {
     AioContext *ctx;
 };
 
+static int xen_block_send_response(XenBlockRequest *request);
+
 static void reset_request(XenBlockRequest *request)
 {
     memset(&request->req, 0, sizeof(request->req));
@@ -115,23 +117,26 @@ static XenBlockRequest *xen_block_start_request(XenBlockDataPlane *dataplane)
     return request;
 }
 
-static void xen_block_finish_request(XenBlockRequest *request)
+static void xen_block_complete_request(XenBlockRequest *request)
 {
     XenBlockDataPlane *dataplane = request->dataplane;
 
-    QLIST_REMOVE(request, list);
-    dataplane->requests_inflight--;
-}
+    if (xen_block_send_response(request)) {
+        Error *local_err = NULL;
 
-static void xen_block_release_request(XenBlockRequest *request)
-{
-    XenBlockDataPlane *dataplane = request->dataplane;
+        xen_device_notify_event_channel(dataplane->xendev,
+                                        dataplane->event_channel,
+                                        &local_err);
+        if (local_err) {
+            error_report_err(local_err);
+        }
+    }
 
     QLIST_REMOVE(request, list);
+    dataplane->requests_inflight--;
     reset_request(request);
     request->dataplane = dataplane;
     QLIST_INSERT_HEAD(&dataplane->freelist, request, list);
-    dataplane->requests_inflight--;
 }
 
 /*
@@ -246,7 +251,6 @@ static int xen_block_copy_request(XenBlockRequest *request)
 }
 
 static int xen_block_do_aio(XenBlockRequest *request);
-static int xen_block_send_response(XenBlockRequest *request);
 
 static void xen_block_complete_aio(void *opaque, int ret)
 {
@@ -286,7 +290,6 @@ static void xen_block_complete_aio(void *opaque, int ret)
     }
 
     request->status = request->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
-    xen_block_finish_request(request);
 
     switch (request->req.operation) {
     case BLKIF_OP_WRITE:
@@ -306,17 +309,8 @@ static void xen_block_complete_aio(void *opaque, int ret)
     default:
         break;
     }
-    if (xen_block_send_response(request)) {
-        Error *local_err = NULL;
 
-        xen_device_notify_event_channel(dataplane->xendev,
-                                        dataplane->event_channel,
-                                        &local_err);
-        if (local_err) {
-            error_report_err(local_err);
-        }
-    }
-    xen_block_release_request(request);
+    xen_block_complete_request(request);
 
     if (dataplane->more_work) {
         qemu_bh_schedule(dataplane->bh);
@@ -420,8 +414,8 @@ static int xen_block_do_aio(XenBlockRequest *request)
     return 0;
 
 err:
-    xen_block_finish_request(request);
     request->status = BLKIF_RSP_ERROR;
+    xen_block_complete_request(request);
     return -1;
 }
 
@@ -575,17 +569,7 @@ static bool xen_block_handle_requests(XenBlockDataPlane *dataplane)
                 break;
             };
 
-            if (xen_block_send_response(request)) {
-                Error *local_err = NULL;
-
-                xen_device_notify_event_channel(dataplane->xendev,
-                                                dataplane->event_channel,
-                                                &local_err);
-                if (local_err) {
-                    error_report_err(local_err);
-                }
-            }
-            xen_block_release_request(request);
+            xen_block_complete_request(request);
             continue;
         }
 
-- 
Anthony PERARD

RE: [PATCH v2 for-5.0] xen-block: Fix double qlist remove and request leak

[Paul Durrant]

> -----Original Message-----
> From: Anthony PERARD <anthony.perard@citrix.com>
> Sent: 06 April 2020 15:02
> To: qemu-devel@nongnu.org
> Cc: qemu-stable@nongnu.org; Anthony PERARD <anthony.perard@citrix.com>; Stefano Stabellini
> <sstabellini@kernel.org>; Paul Durrant <paul@xen.org>; Stefan Hajnoczi <stefanha@redhat.com>; Kevin
> Wolf <kwolf@redhat.com>; Max Reitz <mreitz@redhat.com>; xen-devel@lists.xenproject.org; qemu-
> block@nongnu.org
> Subject: [PATCH v2 for-5.0] xen-block: Fix double qlist remove and request leak
> 
> Commit a31ca6801c02 ("qemu/queue.h: clear linked list pointers on
> remove") revealed that a request was removed twice from a list, once
> in xen_block_finish_request() and a second time in
> xen_block_release_request() when both function are called from
> xen_block_complete_aio(). But also, the `requests_inflight' counter is
> decreased twice, and thus became negative.
> 
> This is a bug that was introduced in bfd0d6366043

NIT: I guess you should quote the patch title here as well.

> , where a `finished'
> list was removed.
> 
> That commit also introduced a leak of request in xen_block_do_aio().
> That function calls xen_block_finish_request() but the request is
> never released after that.
> 
> To fix both issue, we do two changes:
> - we squash finish_request() and release_request() together as we want
>   to remove a request from 'inflight' list to add it to 'freelist'.
> - before releasing a request, we need to let now the result to the
>   other end,

"we need to let the other end know the result"

> thus we should call xen_block_send_response() before
>   releasing a request.
> 
> The first change fix the double QLIST_REMOVE() as we remove the extra

s/fix/fixes

> call. The second change makes the leak go away because if we want to
> call finish_request(), we need to call a function that do all of

s/do/does

> finish, send response, and release.
> 
> Fixes: bfd0d6366043 ("xen-block: improve response latency")
> Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>

The code looks ok, so with the cosmetic fixes...

Reviewed-by: Paul Durrant <paul@xen.org>

Re: [PATCH v2 for-5.0] xen-block: Fix double qlist remove and request leak

[Max Reitz]

[-- Attachment #1.1: Type: text/plain, Size: 1804 bytes --]

On 06.04.20 16:02, Anthony PERARD wrote:
> Commit a31ca6801c02 ("qemu/queue.h: clear linked list pointers on
> remove") revealed that a request was removed twice from a list, once
> in xen_block_finish_request() and a second time in
> xen_block_release_request() when both function are called from
> xen_block_complete_aio(). But also, the `requests_inflight' counter is
> decreased twice, and thus became negative.
> 
> This is a bug that was introduced in bfd0d6366043, where a `finished'
> list was removed.
> 
> That commit also introduced a leak of request in xen_block_do_aio().
> That function calls xen_block_finish_request() but the request is
> never released after that.
> 
> To fix both issue, we do two changes:
> - we squash finish_request() and release_request() together as we want
>   to remove a request from 'inflight' list to add it to 'freelist'.
> - before releasing a request, we need to let now the result to the
>   other end, thus we should call xen_block_send_response() before
>   releasing a request.
> 
> The first change fix the double QLIST_REMOVE() as we remove the extra
> call. The second change makes the leak go away because if we want to
> call finish_request(), we need to call a function that do all of
> finish, send response, and release.
> 
> Fixes: bfd0d6366043 ("xen-block: improve response latency")
> Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>
> ---
>  hw/block/dataplane/xen-block.c | 48 ++++++++++++----------------------
>  1 file changed, 16 insertions(+), 32 deletions(-)

I’m going to send a pull request today anyway, so I hope you won’t mind
and let me take this patch to my branch (with Paul’s suggestions
incorporated):

https://git.xanclic.moe/XanClic/qemu/commits/branch/block

Max


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]