Xen devel - Apr 2013 - Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

Hi,

I''ve been working on getting a working blktap driver allowing to
access ceph RBD block devices without relying on the RBD kernel driver
and it finally got to a point where, it works and is testable.

Some of the advantages are:
 - Easier to update to newer RBD version
 - Allows functionality only available in the userspace RBD library
(write cache, layering, ...)
 - Less issue when you have OSD as domU on the same dom0
 - Contains crash to user space :p (they shouldn''t happen, but ...)

It''s still an early prototype, but if you want to give it a shot and
give feedback.

You can find the code there https://github.com/smunaut/blktap/tree/rbd
 (rbd branch).

Currently the username, poolname and image name are hardcoded ...
(look for FIXME in the code). I''ll get to that next, once I figured
the best format for arguments.

Cheers,

    Sylvain

On Thu, Apr 18, 2013 at 05:05:29PM +0200, Sylvain Munaut
wrote:
> Hi,
>
Hi,
 
> I''ve been working on getting a working blktap driver allowing to
> access ceph RBD block devices without relying on the RBD kernel driver
> and it finally got to a point where, it works and is testable.
> 
Great! Ceph distributed block storage is cool.
> Some of the advantages are:
>  - Easier to update to newer RBD version
>  - Allows functionality only available in the userspace RBD library
> (write cache, layering, ...)
>  - Less issue when you have OSD as domU on the same dom0
>  - Contains crash to user space :p (they shouldn''t happen, but
...)
> 
> It''s still an early prototype, but if you want to give it a shot
and
> give feedback.
> 
> You can find the code there https://github.com/smunaut/blktap/tree/rbd
>  (rbd branch).
> 
> Currently the username, poolname and image name are hardcoded ...
> (look for FIXME in the code). I''ll get to that next, once I
figured
> the best format for arguments.
> 
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.

Thanks,

-- Pasi

> 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

I''m about to start trying this out. Has anything changed since this
email http://www.mail-archive.com/ceph-devel@vger.kernel.org/msg13984.html ?

Thanks

James

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

> 
> 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

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

> 
> > 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

On Mon, Aug 05, 2013 at 01:01:35PM +0200, Sylvain Munaut
wrote:
> 
> > 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.
>
Yep, XenServer is using blktap2.5. 

Also the Centos-6 Xen packages have blktap2.5 patched in.
 
> 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.
> 
I think I saw an announcement recently on xen-devel that blktap3 development has
been stopped..


-- Pasi

> 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

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

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

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

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

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

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

> 
> 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

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

> 
> 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

> >
> > 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

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

> >> > > 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

> 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

> 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

> 
> > 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

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

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

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

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

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

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

> 
> 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

> 
> >
> > 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

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

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

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

> 
> 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

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

> >
> > 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

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

James