Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

> If you have time to write up some lines about steps required to test this,
> that'd be nice, it'll help people to test this stuff.

To quickly test, I compiled the package and just replaced the tapdisk
binary from my "normal" blktap install with the newly compiled one.

Then you need to setup a RBD image named 'test' in the default 'rbd'
pool. You also need to setup a proper ceph.conf and keyring file on
the client (since librbd will use those for the parameters). The
keyring must contain the 'client.admin' key

Then in the config file, use something like
"tap2:tapdisk:rbd:xxx,xvda1,w"  the 'xxx' part is currently ignored
...


Cheers,

    Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Nathan O'Sullivan]

I've been testing this on Ubuntu 12.04.02 64-bit with kernel 3.2.0-48 
and ceph 0.61.4

With rbd cache disabled, it works well enough in initial testing.

However when rbd cache is enabled with:
[client]
rbd_cache = true

the tapdisk process crashes if I do this in the domU:
dd if=/dev/xvda bs=1M > /dev/null


I grabbed the tapdisk stacktrace with gdb:

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7f1677186700 (LWP 6507)]
0x00007f167d21857c in free () from /lib/x86_64-linux-gnu/libc.so.6
(gdb) bt
#0  0x00007f167d21857c in free () from /lib/x86_64-linux-gnu/libc.so.6
#1  0x00007f167daab84b in 
ceph::buffer::raw_posix_aligned::~raw_posix_aligned() () from 
/usr/lib/librados.so.2
#2  0x00007f167daa6f6e in ceph::buffer::ptr::release() () from 
/usr/lib/librados.so.2
#3  0x00007f167d5711c7 in std::_List_base<ceph::buffer::ptr, 
std::allocator<ceph::buffer::ptr> >::_M_clear() () from /usr/lib/librbd.so.1
#4  0x00007f167d5a8ffe in ObjectCacher::trim(long, long) () from 
/usr/lib/librbd.so.1
#5  0x00007f167d5b7e60 in ObjectCacher::_readx(ObjectCacher::OSDRead*, 
ObjectCacher::ObjectSet*, Context*, bool) () from /usr/lib/librbd.so.1
#6  0x00007f167d5bd620 in ObjectCacher::C_RetryRead::finish(int) () from 
/usr/lib/librbd.so.1
#7  0x00007f167d57281a in Context::complete(int) () from 
/usr/lib/librbd.so.1
#8  0x00007f167d5b8f65 in finish_contexts(CephContext*, 
std::list<Context*, std::allocator<Context*> >&, int) () from 
/usr/lib/librbd.so.1
#9  0x00007f167d5ae705 in ObjectCacher::bh_read_finish(long, sobject_t, 
long, unsigned long, ceph::buffer::list&, int, bool) ()
    from /usr/lib/librbd.so.1
#10 0x00007f167d5bc32f in ObjectCacher::C_ReadFinish::finish(int) () 
from /usr/lib/librbd.so.1
#11 0x00007f167d57281a in Context::complete(int) () from 
/usr/lib/librbd.so.1
#12 0x00007f167d5a31f5 in librbd::C_Request::finish(int) () from 
/usr/lib/librbd.so.1
#13 0x00007f167d5a1c14 in librbd::context_cb(void*, void*) () from 
/usr/lib/librbd.so.1
#14 0x00007f167d90f56d in librados::C_AioComplete::finish(int) () from 
/usr/lib/librados.so.2
#15 0x00007f167d97bb00 in Finisher::finisher_thread_entry() () from 
/usr/lib/librados.so.2
#16 0x00007f167c812e9a in start_thread () from 
/lib/x86_64-linux-gnu/libpthread.so.0
#17 0x00007f167d288ccd in clone () from /lib/x86_64-linux-gnu/libc.so.6
#18 0x0000000000000000 in ?? ()


Regards
Nathan

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

> I've been testing this on Ubuntu 12.04.02 64-bit with kernel 3.2.0-48 and
> ceph 0.61.4

Thanks for testing :)

> However when rbd cache is enabled with:
> [client]
> rbd_cache = true
>
> the tapdisk process crashes if I do this in the domU:
> dd if=/dev/xvda bs=1M > /dev/null

Interesting. I'm currently away for I'll try to setup a test and see
if I can reproduce the issue locally.

I never really tried with the cache enabled.

Cheers,

   Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi again,

>> However when rbd cache is enabled with:
>> [client]
>> rbd_cache = true
>>
>> the tapdisk process crashes if I do this in the domU:
>> dd if=/dev/xvda bs=1M > /dev/null

I tested this locally and couldn't reproduce the issue.

Doing reads doesn't do anything bad AFAICT.
Doing writes OTOH seems to leak memory (or at least use much more
memory than the configured cache size).

I also rechecked the code and I don't see anything wrong with it.
AFAICT with or without cache shouldn't change anything so the issue
might be in librbd itself.

Cheers,

    Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Nathan O'Sullivan]

I've installed debug symbols, perhaps that will give a better idea what 
is going on?

#0  __GI___libc_free (mem=0x7f5160650000) at malloc.c:2970
#1  0x00007f515f3ac84b in ~raw_posix_aligned (this=0x7f513c418f20, 
__in_chrg=<optimised out>) at common/buffer.cc:152
#2  ceph::buffer::raw_posix_aligned::~raw_posix_aligned (this=<optimised 
out>, __in_chrg=<optimised out>) at common/buffer.cc:155
#3  0x00007f515f3a7f6e in ceph::buffer::ptr::release 
(this=0x7f513801d600) at common/buffer.cc:328
#4  0x00007f515ee721c7 in ~ptr (this=0x7f513801d600, 
__in_chrg=<optimised out>) at ./include/buffer.h:159
#5  destroy (__p=0x7f513801d600, this=<optimised out>) at 
/usr/include/c++/4.6/ext/new_allocator.h:118
#6  std::_List_base<ceph::buffer::ptr, std::allocator<ceph::buffer::ptr> 
 >::_M_clear (this=0x15e3908) at /usr/include/c++/4.6/bits/list.tcc:78
#7  0x00007f515eea9ffe in ~_List_base (this=0x15e3908, 
__in_chrg=<optimised out>) at /usr/include/c++/4.6/bits/stl_list.h:372
#8  ~list (this=0x15e3908, __in_chrg=<optimised out>) at 
/usr/include/c++/4.6/bits/stl_list.h:429
#9  ~list (this=0x15e3908, __in_chrg=<optimised out>) at 
./include/buffer.h:304
#10 ~BufferHead (this=0x15e38c0, __in_chrg=<optimised out>) at 
osdc/ObjectCacher.h:84
#11 ObjectCacher::trim (this=0x1594a00, max_bytes=33554432, max_ob=42) 
at osdc/ObjectCacher.cc:949
#12 0x00007f515eeb8e60 in ObjectCacher::_readx (this=<optimised out>, 
rd=0x15f1f70, oset=0x1595110, onfinish=0x1591280, external_call=false) 
at osdc/ObjectCacher.cc:1240
#13 0x00007f515eebe620 in ObjectCacher::C_RetryRead::finish 
(this=0x15c3c30, r=<optimised out>) at osdc/ObjectCacher.h:554
#14 0x00007f515ee7381a in Context::complete (this=0x15c3c30, 
r=<optimised out>) at ./include/Context.h:41
#15 0x00007f515eeb9f65 in finish_contexts (cct=0x155cc30, finished=..., 
result=0) at ./include/Context.h:78
#16 0x00007f515eeaf705 in ObjectCacher::bh_read_finish (this=<optimised 
out>, poolid=<optimised out>, oid=..., start=983040, length=131072, 
bl=..., r=0, trust_enoent=true)
     at osdc/ObjectCacher.cc:773
#17 0x00007f515eebd32f in ObjectCacher::C_ReadFinish::finish 
(this=0x15ced30, r=0) at osdc/ObjectCacher.h:478
#18 0x00007f515ee7381a in Context::complete (this=0x15ced30, 
r=<optimised out>) at ./include/Context.h:41
#19 0x00007f515eea41f5 in librbd::C_Request::finish (this=0x159dfd0, 
r=0) at librbd/LibrbdWriteback.cc:55
#20 0x00007f515eea2c14 in librbd::context_cb (c=<optimised out>, 
arg=<optimised out>) at librbd/LibrbdWriteback.cc:35
#21 0x00007f515f21056d in librados::C_AioComplete::finish 
(this=<optimised out>, r=<optimised out>) at 
./librados/AioCompletionImpl.h:171
#22 0x00007f515f27cb00 in Finisher::finisher_thread_entry 
(this=0x1576d98) at common/Finisher.cc:56
#23 0x00007f515e113e9a in start_thread (arg=0x7f5158a87700) at 
pthread_create.c:308
#24 0x00007f515eb89ccd in clone () at 
../sysdeps/unix/sysv/linux/x86_64/clone.S:112
#25 0x0000000000000000 in ?? ()



On 1/07/2013 7:57 PM, Sylvain Munaut wrote:
> Hi again,
>
>>> However when rbd cache is enabled with:
>>> [client]
>>> rbd_cache = true
>>>
>>> the tapdisk process crashes if I do this in the domU:
>>> dd if=/dev/xvda bs=1M > /dev/null
> I tested this locally and couldn't reproduce the issue.
>
> Doing reads doesn't do anything bad AFAICT.
> Doing writes OTOH seems to leak memory (or at least use much more
> memory than the configured cache size).
>
> I also rechecked the code and I don't see anything wrong with it.
> AFAICT with or without cache shouldn't change anything so the issue
> might be in librbd itself.
>
> Cheers,
>
>      Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,


Yes the procedure didn't change.

If you're on debian I could also sent your prebuilt .deb for blktap
and for a patched xen version that includes userspace RBD support.

If you have any issue, I can be found on ceph's IRC under 'tnt' nick.


Cheers,

   Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> Yes the procedure didn't change.
> 
> If you're on debian I could also sent your prebuilt .deb for blktap
> and for a patched xen version that includes userspace RBD support.
> 

It's working great so far. I just pulled the source and built it then copied blktap in.

For some reason I already had a tapdisk in /usr/sbin, as well as the one in /usr/bin, which confused the issue for a while. I must have installed something manually but I don't remember what.

Xen also includes tap-ctl:

blktap-utils: /usr/sbin/tap-ctl
xen-utils-4.1: /usr/lib/xen-4.1/bin/tap-ctl

and I removed the one from xen and linked it to the one in /usr/sbin. I did that before I found the other tapdisk in /usr/sbin so I'm not sure if that step was necessary.

Any chance this will be rolled into the main blktap sources?

> If you have any issue, I can be found on ceph's IRC under 'tnt' nick.
> 

Even though I have been on the internet since 94, I never got the hang of IRC... always found the stream of information a little overwhelming.

Thanks

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,


> It's working great so far. I just pulled the source and built it then copied blktap in.

Good to hear :)

I've been using it more and more recently and it'll been good for me
too, even with live migrations.


> For some reason I already had a tapdisk in /usr/sbin, as well as the one in /usr/bin, which confused the issue for a while. I must have installed something manually but I don't remember what.

What distribution are you using ?


> Any chance this will be rolled into the main blktap sources?

I'd like to ... but I ave no idea how or even who to contact for that
... blktap is so fragmented ...

You have blktap2 which is in the man Xen tree. But that's not what's
used in debian (it's not installed / compiled)

You have the so called blktap2.5 which is what's on github and what I
have based my stuff on. It's also what's shipped with debian as
blktap-utils I think.
I also think Citrix have their own version based off blktap2.5 as well.

And soon there will be blktap3 in the official Xen tree.

I want to at least get it merged in blktap3 but since that code is not
ready (or even merged) yet, it's a bit early for that. That's also
probably Xen 4.4 or Xen 4.5 stuff and so won't hit debian for a while.


Cheers,

   Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> > For some reason I already had a tapdisk in /usr/sbin, as well as the one in
> > /usr/bin, which confused the issue for a while. I must have installed
> > something manually but I don't remember what.
> 
> What distribution are you using ?
> 

Debian Wheezy

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

> I think I saw an announcement recently on xen-devel that blktap3 development has been stopped..

Oh :(

In the mail it speaks about QEMU but is it possible to use the QEMU
driver model when booting PV domains ? (and not PVHVM).

Cheers,

    Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[George Dunlap]

On Mon, Aug 5, 2013 at 1:03 PM, Sylvain Munaut
<s.munaut@whatever-company.com> wrote:
>> I think I saw an announcement recently on xen-devel that blktap3 development has been stopped..
>
> Oh :(
>
> In the mail it speaks about QEMU but is it possible to use the QEMU
> driver model when booting PV domains ? (and not PVHVM).

Yes; qemu knows how to be a Xen PV block back-end.

One of the reasons for stopping work on blktap3 (AIUI) was that it
should in theory have performance characteristics similar to blktap3,
and tends to get newer protocols like ceph "for free" (i.e.,
implemented by someone else).

 -George

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi George,


> Yes; qemu knows how to be a Xen PV block back-end.

Very interesting. Is there documentation about this somewhere ?
I had a look some time ago and it was really not very clear.

Things like what Xen version support this. And with which features (
indirect descriptors, persistent grants, discard, flush, ...) and/or
which limitation.


> One of the reasons for stopping work on blktap3 (AIUI) was that it
> should in theory have performance characteristics similar to blktap3,

And did anyone check the theory currently ? :)


> and tends to get newer protocols like ceph "for free" (i.e.,
> implemented by someone else).

Yes I can definitely see the appeal.


Cheers,

    Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[George Dunlap]

On 05/08/13 14:55, Sylvain Munaut wrote:
> Hi George,
>
>
>> Yes; qemu knows how to be a Xen PV block back-end.
> Very interesting. Is there documentation about this somewhere ?
> I had a look some time ago and it was really not very clear.
>
> Things like what Xen version support this. And with which features (
> indirect descriptors, persistent grants, discard, flush, ...) and/or
> which limitation.

I don't think this is documented anywhere; you'll need to ask the 
experts.  Stefano? Roger? Wei?

>
>
>> One of the reasons for stopping work on blktap3 (AIUI) was that it
>> should in theory have performance characteristics similar to blktap3,
> And did anyone check the theory currently ? :)

I say "in theory" because they are using the same basic architecture: a 
normal process running in dom0, with no special kernel support.  If 
there were a performance difference, it would be something that should 
(in theory) be able to be optimized.

I don't think we have comparisons between qdisk (which is what we call 
qemu-as-pv-backend in Xen) and blktap3 (and since blktap3 wasn't 
finished they wouldn't mean much anyway); but I think qdisk compares 
reasonably with blkback.

  -George

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Wei Liu]

On Mon, Aug 05, 2013 at 03:04:47PM +0100, George Dunlap wrote:
> On 05/08/13 14:55, Sylvain Munaut wrote:
> >Hi George,
> >
> >
> >>Yes; qemu knows how to be a Xen PV block back-end.
> >Very interesting. Is there documentation about this somewhere ?
> >I had a look some time ago and it was really not very clear.
> >
> >Things like what Xen version support this. And with which features (
> >indirect descriptors, persistent grants, discard, flush, ...) and/or
> >which limitation.
> 
> I don't think this is documented anywhere; you'll need to ask the
> experts.  Stefano? Roger? Wei?
> 

These are Linux features not Xen ones AFAICT. In theory they are not
bound to specific Xen versions.

For the network part I don't think new features depend on any specific
hypercall. However for block Roger and Stefano seem to introduce
new hypercalls for certain features (I might be wrong though).


Wei.

> >
> >
> >>One of the reasons for stopping work on blktap3 (AIUI) was that it
> >>should in theory have performance characteristics similar to blktap3,
> >And did anyone check the theory currently ? :)
> 
> I say "in theory" because they are using the same basic
> architecture: a normal process running in dom0, with no special
> kernel support.  If there were a performance difference, it would be
> something that should (in theory) be able to be optimized.
> 
> I don't think we have comparisons between qdisk (which is what we
> call qemu-as-pv-backend in Xen) and blktap3 (and since blktap3
> wasn't finished they wouldn't mean much anyway); but I think qdisk
> compares reasonably with blkback.
> 
>  -George

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[George Dunlap]

On 05/08/13 16:18, Wei Liu wrote:
> On Mon, Aug 05, 2013 at 03:04:47PM +0100, George Dunlap wrote:
>> On 05/08/13 14:55, Sylvain Munaut wrote:
>>> Hi George,
>>>
>>>
>>>> Yes; qemu knows how to be a Xen PV block back-end.
>>> Very interesting. Is there documentation about this somewhere ?
>>> I had a look some time ago and it was really not very clear.
>>>
>>> Things like what Xen version support this. And with which features (
>>> indirect descriptors, persistent grants, discard, flush, ...) and/or
>>> which limitation.
>> I don't think this is documented anywhere; you'll need to ask the
>> experts.  Stefano? Roger? Wei?
>>
> These are Linux features not Xen ones AFAICT. In theory they are not
> bound to specific Xen versions.
>
> For the network part I don't think new features depend on any specific
> hypercall. However for block Roger and Stefano seem to introduce
> new hypercalls for certain features (I might be wrong though).

We're talking about qemu; so the toolstack needs to know how to set up 
qdisk, and I think qdisk would need to be programmed to use, for 
example, persistent grants, yes?

  -G

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Wei Liu]

On Mon, Aug 05, 2013 at 04:20:20PM +0100, George Dunlap wrote:
> On 05/08/13 16:18, Wei Liu wrote:
> >On Mon, Aug 05, 2013 at 03:04:47PM +0100, George Dunlap wrote:
> >>On 05/08/13 14:55, Sylvain Munaut wrote:
> >>>Hi George,
> >>>
> >>>
> >>>>Yes; qemu knows how to be a Xen PV block back-end.
> >>>Very interesting. Is there documentation about this somewhere ?
> >>>I had a look some time ago and it was really not very clear.
> >>>
> >>>Things like what Xen version support this. And with which features (
> >>>indirect descriptors, persistent grants, discard, flush, ...) and/or
> >>>which limitation.
> >>I don't think this is documented anywhere; you'll need to ask the
> >>experts.  Stefano? Roger? Wei?
> >>
> >These are Linux features not Xen ones AFAICT. In theory they are not
> >bound to specific Xen versions.
> >
> >For the network part I don't think new features depend on any specific
> >hypercall. However for block Roger and Stefano seem to introduce
> >new hypercalls for certain features (I might be wrong though).
> 
> We're talking about qemu; so the toolstack needs to know how to set
> up qdisk, and I think qdisk would need to be programmed to use, for
> example, persistent grants, yes?
> 

I don't think toolstack needs to involve in this. At least for the
network part FE and BE negotiate what features to use. The general idea
is that new feature will always be of benifit to enable so we make use
of them whenever possible. Certain features do have sysfs entries to
configure but that's not coded into libxl.

I cannot speak for block drivers, but grepping the source code I don't
think you can configure persistent grants via libxl either.


Wei.

>  -G

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> Yes the procedure didn't change.
> 
> If you're on debian I could also sent your prebuilt .deb for blktap
> and for a patched xen version that includes userspace RBD support.
> 
> If you have any issue, I can be found on ceph's IRC under 'tnt' nick.
> 

I've had a few occasions where tapdisk has segfaulted:

tapdisk[9180]: segfault at 7f7e3a5c8c10 ip 00007f7e387532d4 sp 00007f7e3a5c8c10 error 4 in libpthread-2.13.so[7f7e38748000+17000]
tapdisk:9180 blocked for more than 120 seconds.
tapdisk         D ffff88043fc13540     0  9180      1 0x00000000

and then like:

end_request: I/O error, dev tdc, sector 472008

I can't be sure but I suspect that when this happened either one OSD was offline, or the cluster lost quorum briefly.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

> I've had a few occasions where tapdisk has segfaulted:
>
> tapdisk[9180]: segfault at 7f7e3a5c8c10 ip 00007f7e387532d4 sp 00007f7e3a5c8c10 error 4 in libpthread-2.13.so[7f7e38748000+17000]
> tapdisk:9180 blocked for more than 120 seconds.
> tapdisk         D ffff88043fc13540     0  9180      1 0x00000000
>
> and then like:
>
> end_request: I/O error, dev tdc, sector 472008
>
> I can't be sure but I suspect that when this happened either one OSD was offline, or the cluster lost quorum briefly.

Interesting. There might be an issue if a request ends in error, I'll
have to check that.
I'll have a look on monday.

Cheers,

    Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> Hi,
> 
> > I've had a few occasions where tapdisk has segfaulted:
> >
> > tapdisk[9180]: segfault at 7f7e3a5c8c10 ip 00007f7e387532d4 sp
> 00007f7e3a5c8c10 error 4 in libpthread-2.13.so[7f7e38748000+17000]
> > tapdisk:9180 blocked for more than 120 seconds.
> > tapdisk         D ffff88043fc13540     0  9180      1 0x00000000
> >
> > and then like:
> >
> > end_request: I/O error, dev tdc, sector 472008
> >
> > I can't be sure but I suspect that when this happened either one OSD was
> > offline, or the cluster lost quorum briefly.
> 
> Interesting. There might be an issue if a request ends in error, I'll
> have to check that.
> I'll have a look on monday.
> 

You say in tdrbd_finish_aiocb:

        while (1) {
                /* POSIX says write will be atomic or blocking */
                rv = write(prv->pipe_fds[1], (void*)&req, sizeof(req));

but from what I've read in "man 7 pipe", the statement about being atomic only applies if the pipe is open in non-blocking mode, and you open it with a call to pipe() (same as pipe2(,0)) and you never call fcntl to change it. This would be consistent with the random crashes I'm seeing - I thought they were related to transient errors but my ceph cluster has been perfectly stable for a few days now and it's still happening.

What do you think?

Thanks

James

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> >
> > Hi,
> >
> > > I've had a few occasions where tapdisk has segfaulted:
> > >
> > > tapdisk[9180]: segfault at 7f7e3a5c8c10 ip 00007f7e387532d4 sp
> > 00007f7e3a5c8c10 error 4 in libpthread-2.13.so[7f7e38748000+17000]
> > > tapdisk:9180 blocked for more than 120 seconds.
> > > tapdisk         D ffff88043fc13540     0  9180      1 0x00000000
> > >
> > > and then like:
> > >
> > > end_request: I/O error, dev tdc, sector 472008
> > >
> > > I can't be sure but I suspect that when this happened either one OSD was
> > > offline, or the cluster lost quorum briefly.
> >
> > Interesting. There might be an issue if a request ends in error, I'll
> > have to check that.
> > I'll have a look on monday.
> >
> 
> You say in tdrbd_finish_aiocb:
> 
>         while (1) {
>                 /* POSIX says write will be atomic or blocking */
>                 rv = write(prv->pipe_fds[1], (void*)&req, sizeof(req));
> 
> but from what I've read in "man 7 pipe", the statement about being atomic
> only applies if the pipe is open in non-blocking mode, and you open it with a
> call to pipe() (same as pipe2(,0)) and you never call fcntl to change it. This
> would be consistent with the random crashes I'm seeing - I thought they
> were related to transient errors but my ceph cluster has been perfectly
> stable for a few days now and it's still happening.
> 
> What do you think?
> 

Actually maybe not. What I was reading only applies for large number of bytes written to the pipe, and even then I got confused by the double negatives. Sorry for the noise.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

>> > > tapdisk[9180]: segfault at 7f7e3a5c8c10 ip 00007f7e387532d4 sp
>> > 00007f7e3a5c8c10 error 4 in libpthread-2.13.so[7f7e38748000+17000]
>> > > tapdisk:9180 blocked for more than 120 seconds.
>> > > tapdisk         D ffff88043fc13540     0  9180      1 0x00000000

You can try generating a core file by changing the ulimit on the running process

http://superuser.com/questions/404239/setting-ulimit-on-a-running-process

A backtrace would be useful :)


> Actually maybe not. What I was reading only applies for large number of bytes written to the pipe, and even then I got confused by the double negatives. Sorry for the noise.

Yes, as you discovered but size < PIPE_BUF, they should be atomic even
in non-blocking mode. But I could still add assert() there to make
sure it is.


I did find a bug where it could "leak" requests which may lead to
hang. But it shouldn't crash ...

Here's an (untested yet) patch in the rbd error path:


diff --git a/drivers/block-rbd.c b/drivers/block-rbd.c
index 68fbed7..ab2d2c5 100644
--- a/drivers/block-rbd.c
+++ b/drivers/block-rbd.c
@@ -560,6 +560,9 @@ err:
        if (c)
                rbd_aio_release(c);

+       list_move(&req->queue, &prv->reqs_free);
+       prv->reqs_free_count++;
+
        return rv;
 }


Cheers,

     Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> >> > > tapdisk[9180]: segfault at 7f7e3a5c8c10 ip 00007f7e387532d4 sp
> >> > 00007f7e3a5c8c10 error 4 in libpthread-2.13.so[7f7e38748000+17000]
> >> > > tapdisk:9180 blocked for more than 120 seconds.
> >> > > tapdisk         D ffff88043fc13540     0  9180      1 0x00000000
> 
> You can try generating a core file by changing the ulimit on the running
> process
> 
> A backtrace would be useful :)
> 

I found it was actually dumping core in /, but gdb doesn't seem to work nicely and all I get is this:

warning: Can't read pathname for load map: Input/output error.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
Cannot find new threads: generic error
Core was generated by `tapdisk'.
Program terminated with signal 11, Segmentation fault.
#0  pthread_cond_wait@@GLIBC_2.3.2 () at ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_cond_wait.S:163
163     ../nptl/sysdeps/unix/sysv/linux/x86_64/pthread_cond_wait.S: No such file or directory.

Even when I attach to a running process.

One VM segfaults on startup, pretty much everytime except never when I attach strace to it, meaning it's probably a race condition and may not actually be in your code...

> 
> > Actually maybe not. What I was reading only applies for large number of
> > bytes written to the pipe, and even then I got confused by the double
> > negatives. Sorry for the noise.
> 
> Yes, as you discovered but size < PIPE_BUF, they should be atomic even
> in non-blocking mode. But I could still add assert() there to make
> sure it is.

Nah I got that completely backwards. I see now you are only passing a pointer so yes it should never be non-atomic.

> I did find a bug where it could "leak" requests which may lead to
> hang. But it shouldn't crash ...
> 
> Here's an (untested yet) patch in the rbd error path:
> 

I'll try that later this morning when I get a minute.

I've done the poor-mans-debugger thing and riddled the code with printf's but as far as I can determine every routine starts and ends. My thinking at the moment is that it's either a race (the VM's most likely to crash have multiple disks), or a buffer overflow that trips it up either immediately, or later.

I have definitely observed multiple VM's crash when something in ceph hiccup's (eg I bring a mon up or down), if that helps.

I also followed through the rbd_aio_release idea on the weekend - I can see that if the read returns failure it means the callback was never called so the release is then the responsibility of the caller.

Thanks

James

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> Here's an (untested yet) patch in the rbd error path:
> 
> diff --git a/drivers/block-rbd.c b/drivers/block-rbd.c
> index 68fbed7..ab2d2c5 100644
> --- a/drivers/block-rbd.c
> +++ b/drivers/block-rbd.c
> @@ -560,6 +560,9 @@ err:
>         if (c)
>                 rbd_aio_release(c);
> 
> +       list_move(&req->queue, &prv->reqs_free);
> +       prv->reqs_free_count++;
> +
>         return rv;
>  }
> 

FWIW, I can confirm via printf's that this error path is never hit in at least some of the crashes I'm seeing.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

> FWIW, I can confirm via printf's that this error path is never hit in at least some of the crashes I'm seeing.

Ok thanks.

Are you using cache btw ?

Cheers,

    Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> > FWIW, I can confirm via printf's that this error path is never hit in at least
> some of the crashes I'm seeing.
> 
> Ok thanks.
> 
> Are you using cache btw ?
> 

I hope not. How could I tell? It's not something I've explicitly enabled.

Thanks

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

> I hope not. How could I tell? It's not something I've explicitly enabled.

It's disabled by default.

So you'd have to have enabled it either in ceph.conf  or directly in
the device path in the xen config. (option is 'rbd cache',
http://ceph.com/docs/next/rbd/rbd-config-ref/ )

Cheers,

    Sylvain

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Frederik Thuysbaert]

Hi,

I have been testing this a while now, and just finished testing your 
untested patch. The rbd caching problem still persists.

The system I am testing on has the following characteristics:

Dom0:
     - Linux xen-001 3.2.0-4-amd64 #1 SMP Debian 3.2.46-1 x86_64
     - Most recent git checkout of blktap rbd branch

DomU:
     - Same kernel as dom0
     - Root (xvda1) is a logical volume on dom0
     - xvda2 is a Rados Block Device format 1

Let me start by saying that the errors only occur with RBD client 
caching ON.
I will give the error messages of both dom0 and domU before and after I 
applied the patch.

Actions in domU to trigger errors:

~# mkfs.xfs -f /dev/xvda2
~# mount /dev/xvda2 /mnt
~# bonnie -u 0 -g 0 /mnt


Error messages:

BEFORE patch:

Without RBD cache:

dom0: no errors
domU: no errors

With RBD cache:

dom0: no errors

domU:
Aug 13 18:18:33 debian-vm-101 kernel: [   37.960475] lost page write due 
to I/O error on xvda2
Aug 13 18:18:33 debian-vm-101 kernel: [   37.960488] lost page write due 
to I/O error on xvda2
Aug 13 18:18:33 debian-vm-101 kernel: [   37.960501] lost page write due 
to I/O error on xvda2
...
Aug 13 18:18:52 debian-vm-101 kernel: [   56.394645] XFS (xvda2): 
xfs_do_force_shutdown(0x2) called from line 1007 of file 
/build/linux-s5x2oE/linux-3.2.46/fs/xfs/xfs_log.c.  Return address = 
0xffffffffa013ced5
Aug 13 18:19:19 debian-vm-101 kernel: [   83.941539] XFS (xvda2): 
xfs_log_force: error 5 returned.
Aug 13 18:19:19 debian-vm-101 kernel: [   83.941565] XFS (xvda2): 
xfs_log_force: error 5 returned.
...

AFTER patch:

Without RBD cache:

dom0: no errors
domU: no errors

With RBD cache:

dom0:
Aug 13 16:40:49 xen-001 kernel: [   94.954734] tapdisk[3075]: segfault 
at 7f749ee86da0 ip 00007f749d060776 sp 00007f748ea7a460 error 7 in 
libpthread-2.13.so[7f749d059000+17000]


domU:
Same as before patch.



I would like to add that I have the time to test this, we are happy to 
help you in any way possible. However, since I am no C developer, I 
won't be able to do much more than testing.


Regards

Frederik


On 13-08-13 11:20, Sylvain Munaut wrote:
> Hi,
>
>> I hope not. How could I tell? It's not something I've explicitly enabled.
> It's disabled by default.
>
> So you'd have to have enabled it either in ceph.conf  or directly in
> the device path in the xen config. (option is 'rbd cache',
> http://ceph.com/docs/next/rbd/rbd-config-ref/ )
>
> Cheers,
>
>      Sylvain
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

> I have been testing this a while now, and just finished testing your
> untested patch. The rbd caching problem still persists.

Yes, I wouldn't expect to change anything for caching. But I still
don't understand why caching would change anything at all ... all of
it should be handled within the librbd lib.


Note that I would recommend against caching anyway. The blktap layer
doesn't pass through the FLUSH commands and so this make it completely
unsafe because the VM will think things are commited to disk durably
even though they are not ...



> I will give the error messages of both dom0 and domU before and after I
> applied the patch.

It's actually strange that it changes anything at all.

Can you try adding a ERROR("HERE\n");  in that error path processing
and check syslog to see if it's triggered at all ?

A traceback would be great if you can get a core file. And possibly
compile tapdisk with debug symbols.


Cheers,

    Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

Just noticed email subject "qemu-1.4.0 and onwards, linux kernel 3.2.x, ceph-RBD, heavy I/O leads to kernel_hung_tasks_timout_secs message and unresponsive qemu-process, [Qemu-devel] [Bug 1207686]" where Sage noted that he has seen a completion called twice in the logs the OP posted. If that is actually happening (and not just an artefact of logging ring buffer overflowing or something) then I think that could easily cause a segfault in tapdisk rbd.

I'll try and see if I can log when that happens.

James

> -----Original Message-----
> From: Sylvain Munaut [mailto:s.munaut@whatever-company.com]
> Sent: Tuesday, 13 August 2013 7:20 PM
> To: James Harper
> Cc: Pasi Kärkkäinen; ceph-devel@vger.kernel.org; xen-devel@lists.xen.org
> Subject: Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to
> test ? :p
> 
> Hi,
> 
> > I hope not. How could I tell? It's not something I've explicitly enabled.
> 
> It's disabled by default.
> 
> So you'd have to have enabled it either in ceph.conf  or directly in
> the device path in the xen config. (option is 'rbd cache',
> http://ceph.com/docs/next/rbd/rbd-config-ref/ )
> 
> Cheers,
> 
>     Sylvain
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

I think I have a separate problem too - tapdisk will segfault almost immediately upon starting but seemingly only for Linux PV DomU's. Once it has started doing this I have to wait a few hours to a day before it starts working again. My Windows DomU's appear to be able to start normally though.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

On Wed, Aug 14, 2013 at 1:39 AM, James Harper
<james.harper@bendigoit.com.au> wrote:
> I think I have a separate problem too - tapdisk will segfault almost immediately upon starting but seemingly only for Linux PV DomU's. Once it has started doing this I have to wait a few hours to a day before it starts working again. My Windows DomU's appear to be able to start normally though.

What about other blktap driver ? like using blktap raw driver, does
that work without issue ?

Cheers,

    Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> On Wed, Aug 14, 2013 at 1:39 AM, James Harper
> <james.harper@bendigoit.com.au> wrote:
> > I think I have a separate problem too - tapdisk will segfault almost
> immediately upon starting but seemingly only for Linux PV DomU's. Once it
> has started doing this I have to wait a few hours to a day before it starts
> working again. My Windows DomU's appear to be able to start normally
> though.
> 
> What about other blktap driver ? like using blktap raw driver, does
> that work without issue ?
> 

What's the syntax for that? I use tap2:tapdisk:rbd for rbd, but I don't know how to specify raw and anything I try just says it doesn't understand

Thanks

James

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> >
> > On Wed, Aug 14, 2013 at 1:39 AM, James Harper
> > <james.harper@bendigoit.com.au> wrote:
> > > I think I have a separate problem too - tapdisk will segfault almost
> > immediately upon starting but seemingly only for Linux PV DomU's. Once it
> > has started doing this I have to wait a few hours to a day before it starts
> > working again. My Windows DomU's appear to be able to start normally
> > though.
> >
> > What about other blktap driver ? like using blktap raw driver, does
> > that work without issue ?
> >
> 
> What's the syntax for that? I use tap2:tapdisk:rbd for rbd, but I don't know
> how to specify raw and anything I try just says it doesn't understand
> 

I just tested with tap2:aio and that worked (had an old image of the VM on lvm still so just tested with that). Switching back to rbd and it crashes every time, just as postgres is starting in the vm. Booting into single user mode, waiting 30 seconds, then letting the boot continue it still crashes at the same point so I think it's not a timing thing - maybe postgres has a disk access pattern that is triggering the bug?

Putting printf's in seems to make the problem go away sometimes, so it's hard to debug.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Frederik Thuysbaert]

On 13-08-13 17:39, Sylvain Munaut wrote:
> Hi,
>
>> I have been testing this a while now, and just finished testing your
>> untested patch. The rbd caching problem still persists.
> Yes, I wouldn't expect to change anything for caching. But I still
> don't understand why caching would change anything at all ... all of
> it should be handled within the librbd lib.
>
>
> Note that I would recommend against caching anyway. The blktap layer
> doesn't pass through the FLUSH commands and so this make it completely
> unsafe because the VM will think things are commited to disk durably
> even though they are not ...
>
>
>
>> I will give the error messages of both dom0 and domU before and after I
>> applied the patch.
> It's actually strange that it changes anything at all.
>
> Can you try adding a ERROR("HERE\n");  in that error path processing
> and check syslog to see if it's triggered at all ?
I did this, and can confirm that it is not triggered.
>
> A traceback would be great if you can get a core file. And possibly
> compile tapdisk with debug symbols.
I'm not quite sure what u mean, can u give some more information on how 
I do this? I compiled tapdisk with ./configure CFLAGS=-g, but I'm not 
sure this is what u meant.
>
> Cheers,
>
>      Sylvain
Regards

- Frederik

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Frederik Thuysbaert]

On 13-08-13 17:39, Sylvain Munaut wrote:
>
> It's actually strange that it changes anything at all.
>
> Can you try adding a ERROR("HERE\n");  in that error path processing
> and check syslog to see if it's triggered at all ?
>
> A traceback would be great if you can get a core file. And possibly
> compile tapdisk with debug symbols.
When halting the domU after the errors, I get the following in dom0 syslog:

Aug 14 10:43:57 xen-001 kernel: [ 5041.338756] INFO: task tapdisk:9690 
blocked for more than 120 seconds.
Aug 14 10:43:57 xen-001 kernel: [ 5041.338817] "echo 0 > 
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
Aug 14 10:43:57 xen-001 kernel: [ 5041.338903] tapdisk         D 
ffff8800bf213780     0  9690      1 0x00000000
Aug 14 10:43:57 xen-001 kernel: [ 5041.338908]  ffff8800b4b0e730 
0000000000000246 ffff880000000000 ffffffff8160d020
Aug 14 10:43:57 xen-001 kernel: [ 5041.338912]  0000000000013780 
ffff8800b4ebffd8 ffff8800b4ebffd8 ffff8800b4b0e730
Aug 14 10:43:57 xen-001 kernel: [ 5041.338916]  ffff8800b4d36190 
0000000181199c37 ffff8800b5798c00 ffff8800b5798c00
Aug 14 10:43:57 xen-001 kernel: [ 5041.338921] Call Trace:
Aug 14 10:43:57 xen-001 kernel: [ 5041.338929] [<ffffffffa0308411>] ? 
blktap_device_destroy_sync+0x85/0x9b [blktap]
Aug 14 10:43:57 xen-001 kernel: [ 5041.338936] [<ffffffff8105fadf>] ? 
add_wait_queue+0x3c/0x3c
Aug 14 10:43:57 xen-001 kernel: [ 5041.338940] [<ffffffffa0307444>] ? 
blktap_ring_release+0x10/0x2d [blktap]
Aug 14 10:43:57 xen-001 kernel: [ 5041.338945] [<ffffffff810fb141>] ? 
fput+0xf9/0x1a1
Aug 14 10:43:57 xen-001 kernel: [ 5041.338949] [<ffffffff810f8e6c>] ? 
filp_close+0x62/0x6a
Aug 14 10:43:57 xen-001 kernel: [ 5041.338954] [<ffffffff81049831>] ? 
put_files_struct+0x60/0xad
Aug 14 10:43:57 xen-001 kernel: [ 5041.338958] [<ffffffff81049e38>] ? 
do_exit+0x292/0x713
Aug 14 10:43:57 xen-001 kernel: [ 5041.338961] [<ffffffff8104a539>] ? 
do_group_exit+0x74/0x9e
Aug 14 10:43:57 xen-001 kernel: [ 5041.338965] [<ffffffff81055f94>] ? 
get_signal_to_deliver+0x46d/0x48f
Aug 14 10:43:57 xen-001 kernel: [ 5041.338970] [<ffffffff81347759>] ? 
force_sig_info_fault+0x5b/0x63
Aug 14 10:43:57 xen-001 kernel: [ 5041.338975] [<ffffffff8100de27>] ? 
do_signal+0x38/0x610
Aug 14 10:43:57 xen-001 kernel: [ 5041.338979] [<ffffffff81070deb>] ? 
arch_local_irq_restore+0x7/0x8
Aug 14 10:43:57 xen-001 kernel: [ 5041.338983] [<ffffffff8134eb77>] ? 
_raw_spin_unlock_irqrestore+0xe/0xf
Aug 14 10:43:57 xen-001 kernel: [ 5041.338987] [<ffffffff8103f944>] ? 
wake_up_new_task+0xb9/0xc2
Aug 14 10:43:57 xen-001 kernel: [ 5041.338992] [<ffffffff8106f987>] ? 
sys_futex+0x120/0x151
Aug 14 10:43:57 xen-001 kernel: [ 5041.338995] [<ffffffff8100e435>] ? 
do_notify_resume+0x25/0x68
Aug 14 10:43:57 xen-001 kernel: [ 5041.338999] [<ffffffff8134ef3c>] ? 
retint_signal+0x48/0x8c
...
Aug 14 10:44:17 xen-001 tap-ctl: tap-err:tap_ctl_connect: couldn't 
connect to /var/run/blktap-control/ctl9478: 111

>
>
> Cheers,
>
>      Sylvain
Regards

- Frederik

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

> I just tested with tap2:aio and that worked (had an old image of the VM on lvm still so just tested with that). Switching back to rbd and it crashes every time, just as postgres is starting in the vm. Booting into single user mode, waiting 30 seconds, then letting the boot continue it still crashes at the same point so I think it's not a timing thing - maybe postgres has a disk access pattern that is triggering the bug?

Mmm, that's really interesting.

Could you try to disable request merging ? Just give option
max_merge_size=0 in the tap2 disk description. Something like
'tap2:tapdisk:rbd:rbd/test:max_merge_size=0,xvda2,w'

Cheers,

     Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> Hi,
> 
> > I just tested with tap2:aio and that worked (had an old image of the VM on
> lvm still so just tested with that). Switching back to rbd and it crashes every
> time, just as postgres is starting in the vm. Booting into single user mode,
> waiting 30 seconds, then letting the boot continue it still crashes at the same
> point so I think it's not a timing thing - maybe postgres has a disk access
> pattern that is triggering the bug?
> 
> Mmm, that's really interesting.
> 
> Could you try to disable request merging ? Just give option
> max_merge_size=0 in the tap2 disk description. Something like
> 'tap2:tapdisk:rbd:rbd/test:max_merge_size=0,xvda2,w'
> 

Just as suddenly the problem went away and I can no longer reproduce the crash on startup. Very frustrating. Most likely it still crashed during heavy use but that can take days.

I've just upgraded librbd to dumpling (from cuttlefish) on that one server and will see what it's doing by morning. I'll disable merging when I can reproduce it next.

Thanks

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi Frederik,

>> A traceback would be great if you can get a core file. And possibly
>> compile tapdisk with debug symbols.
>
> I'm not quite sure what u mean, can u give some more information on how I do
> this? I compiled tapdisk with ./configure CFLAGS=-g, but I'm not sure this
> is what u meant.

Yes, ./configure CFLAGS=-g LDFLAGS=-g  is a good start.

Then when it crashes, if will leave a 'core' time somewhere. (not sure
where, maybe in / or in /tmp)
If it doesn't you may have to enable it. When the process is running,
use this on the tapdisk PID :

http://superuser.com/questions/404239/setting-ulimit-on-a-running-process

Then once you have a core file, you can use gdb along with the tapdisk
executable to generate a meaningful backtrace of where the crash
happenned :

See for ex http://publib.boulder.ibm.com/httpserv/ihsdiag/get_backtrace.html
for how to do it.


> When halting the domU after the errors, I get the following in dom0 syslog:

It's not really unexpected. If tapdisk crashes the IO ring is going to
be left hanging and god knows what weird behaviour will happen ...


Cheers,

    Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> >
> > Hi,
> >
> > > I just tested with tap2:aio and that worked (had an old image of the VM
> on
> > lvm still so just tested with that). Switching back to rbd and it crashes every
> > time, just as postgres is starting in the vm. Booting into single user mode,
> > waiting 30 seconds, then letting the boot continue it still crashes at the
> same
> > point so I think it's not a timing thing - maybe postgres has a disk access
> > pattern that is triggering the bug?
> >
> > Mmm, that's really interesting.
> >
> > Could you try to disable request merging ? Just give option
> > max_merge_size=0 in the tap2 disk description. Something like
> > 'tap2:tapdisk:rbd:rbd/test:max_merge_size=0,xvda2,w'
> >
> 
> Just as suddenly the problem went away and I can no longer reproduce the
> crash on startup. Very frustrating. Most likely it still crashed during heavy use
> but that can take days.
> 
> I've just upgraded librbd to dumpling (from cuttlefish) on that one server and
> will see what it's doing by morning. I'll disable merging when I can reproduce
> it next.
> 

I just had a crash since upgrading to dumpling, and will disable merging tonight.

James

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> I just had a crash since upgrading to dumpling, and will disable merging
> tonight.
> 

Still crashes with merging disabled.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Frederik Thuysbaert]

Hi Sylvain,

>> I'm not quite sure what u mean, can u give some more information on how I do
>> this? I compiled tapdisk with ./configure CFLAGS=-g, but I'm not sure this
>> is what u meant.
>
> Yes, ./configure CFLAGS=-g LDFLAGS=-g  is a good start.
>
>...
>
> Then once you have a core file, you can use gdb along with the tapdisk
> executable to generate a meaningful backtrace of where the crash
>

I did 2 runs, with a cold reboot in between just to be sure. I don't 
think I'm getting a lot of valuable information, but I will post it 
anyway. The reason for the cold reboot was a 'Cannot access memory at 
address ...' in gdb after the first frame, I thought it could help.

Here's what I got:

try 1:
Core was generated by `tapdisk'.
Program terminated with signal 11, Segmentation fault.
#0  0x00007fb42d2082d7 in pthread_cond_wait@@GLIBC_2.3.2 () from 
/lib/x86_64-linux-gnu/libpthread.so.0
(gdb) bt
#0  0x00007fb42d2082d7 in pthread_cond_wait@@GLIBC_2.3.2 () from 
/lib/x86_64-linux-gnu/libpthread.so.0
Cannot access memory at address 0x7fb42f081c38
(gdb) frame 0
#0  0x00007fb42d2082d7 in pthread_cond_wait@@GLIBC_2.3.2 () from 
/lib/x86_64-linux-gnu/libpthread.so.0
(gdb) list
77	}
78	
79	int
80	main(int argc, char *argv[])
81	{
82		char *control;
83		int c, err, nodaemon;
84		FILE *out;
85	
86		control  = NULL;
(gdb) info locals
No symbol table info available.

try 2:
Core was generated by `tapdisk'.
Program terminated with signal 11, Segmentation fault.
#0  0x00007fe05a721e6b in poll () from /lib/x86_64-linux-gnu/libc.so.6
(gdb) bt
#0  0x00007fe05a721e6b in poll () from /lib/x86_64-linux-gnu/libc.so.6
Cannot access memory at address 0x7fe05c2ba518
(gdb) frame 0
#0  0x00007fe05a721e6b in poll () from /lib/x86_64-linux-gnu/libc.so.6
(gdb) list
77	}
78	
79	int
80	main(int argc, char *argv[])
81	{
82		char *control;
83		int c, err, nodaemon;
84		FILE *out;
85	
86		control  = NULL;
(gdb) info locals
No symbol table info available.

Regards,

- Frederik

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

Sylvain,

Are you still working on this in any way?

It's been working great for me but seems to use an excessive amount of memory, like 300MB per process. Is that expected?

Thanks

James

> -----Original Message-----
> From: ceph-devel-owner@vger.kernel.org [mailto:ceph-devel-
> owner@vger.kernel.org] On Behalf Of Sylvain Munaut
> Sent: Saturday, 20 April 2013 12:41 AM
> To: Pasi Kärkkäinen
> Cc: ceph-devel@vger.kernel.org; xen-devel@lists.xen.org
> Subject: Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to
> test ? :p
> 
> > If you have time to write up some lines about steps required to test this,
> > that'd be nice, it'll help people to test this stuff.
> 
> To quickly test, I compiled the package and just replaced the tapdisk
> binary from my "normal" blktap install with the newly compiled one.
> 
> Then you need to setup a RBD image named 'test' in the default 'rbd'
> pool. You also need to setup a proper ceph.conf and keyring file on
> the client (since librbd will use those for the parameters). The
> keyring must contain the 'client.admin' key
> 
> Then in the config file, use something like
> "tap2:tapdisk:rbd:xxx,xvda1,w"  the 'xxx' part is currently ignored
> ...
> 
> 
> Cheers,
> 
>     Sylvain
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi James,

> Are you still working on this in any way?

Well I'm using it, but I haven't worked on it. I never was able to
reproduce any issue with it locally ...
In prod, I do run it with cache disabled though since I never took the
time to check using the cache was safe in the various failure modes.

Is 300 MB normal ? Well, that probably depends on your settings (cache
enabled / size / ...). But in anycase I'd guess the memory comes from
a librbd itself. It's not like I do much allocation myself :p

Cheers,

   Sylvain

RE: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[James Harper]

> 
> Hi James,
> 
> > Are you still working on this in any way?
> 
> Well I'm using it, but I haven't worked on it. I never was able to
> reproduce any issue with it locally ...
> In prod, I do run it with cache disabled though since I never took the
> time to check using the cache was safe in the various failure modes.
> 
> Is 300 MB normal ? Well, that probably depends on your settings (cache
> enabled / size / ...). But in anycase I'd guess the memory comes from
> a librbd itself. It's not like I do much allocation myself :p
> 

What sort of memory are your instances using? I haven't turned on any caching so I assume it's disabled.

I increased the stack size to 8M to work around the crash I was having, but lowering that to 2MB doesn't have any significant impact on memory usage.

James

Re: [Xen-devel] Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

[Sylvain Munaut]

Hi,

> What sort of memory are your instances using?

I just had a look. Around 120 Mb. Which indeed is a bit higher that I'd like.


> I haven't turned on any caching so I assume it's disabled.

Yes.


Cheers,

    Sylvain