mbox series

[v3,0/6] ceph: don't request caps for idle open files

Message ID 20200228115550.6904-1-zyan@redhat.com (mailing list archive)
Headers show
Series ceph: don't request caps for idle open files | expand

Message

Yan, Zheng Feb. 28, 2020, 11:55 a.m. UTC
This series make cephfs client not request caps for open files that
idle for a long time. For the case that one active client and multiple
standby clients open the same file, this increase the possibility that
mds issues exclusive caps to the active client.

Yan, Zheng (4):
  ceph: always renew caps if mds_wanted is insufficient
  ceph: consider inode's last read/write when calculating wanted caps
  ceph: simplify calling of ceph_get_fmode()
  ceph: remove delay check logic from ceph_check_caps()

 fs/ceph/caps.c               | 324 +++++++++++++++--------------------
 fs/ceph/file.c               |  39 ++---
 fs/ceph/inode.c              |  19 +-
 fs/ceph/ioctl.c              |   2 +
 fs/ceph/mds_client.c         |   5 -
 fs/ceph/super.h              |  35 ++--
 include/linux/ceph/ceph_fs.h |   1 +
 7 files changed, 188 insertions(+), 237 deletions(-)

changes since v2
 - make __ceph_caps_file_wanted more readable
 - add patch 5 and 6, which fix hung write during testing patch 1~4

Comments

Jeffrey Layton Feb. 28, 2020, 1:01 p.m. UTC | #1
On Fri, 2020-02-28 at 19:55 +0800, Yan, Zheng wrote:
> This series make cephfs client not request caps for open files that
> idle for a long time. For the case that one active client and multiple
> standby clients open the same file, this increase the possibility that
> mds issues exclusive caps to the active client.
> 
> Yan, Zheng (4):
>   ceph: always renew caps if mds_wanted is insufficient
>   ceph: consider inode's last read/write when calculating wanted caps
>   ceph: simplify calling of ceph_get_fmode()
>   ceph: remove delay check logic from ceph_check_caps()
> 
>  fs/ceph/caps.c               | 324 +++++++++++++++--------------------
>  fs/ceph/file.c               |  39 ++---
>  fs/ceph/inode.c              |  19 +-
>  fs/ceph/ioctl.c              |   2 +
>  fs/ceph/mds_client.c         |   5 -
>  fs/ceph/super.h              |  35 ++--
>  include/linux/ceph/ceph_fs.h |   1 +
>  7 files changed, 188 insertions(+), 237 deletions(-)
> 
> changes since v2
>  - make __ceph_caps_file_wanted more readable
>  - add patch 5 and 6, which fix hung write during testing patch 1~4
> 

Thanks Zheng. This looks good to me -- merged into testing branch with
some small revisions to the changelogs. Let me know if I made any
mistakes there.
Jeffrey Layton March 2, 2020, 7:53 p.m. UTC | #2
On Fri, 2020-02-28 at 19:55 +0800, Yan, Zheng wrote:
> This series make cephfs client not request caps for open files that
> idle for a long time. For the case that one active client and multiple
> standby clients open the same file, this increase the possibility that
> mds issues exclusive caps to the active client.
> 
> Yan, Zheng (4):
>   ceph: always renew caps if mds_wanted is insufficient
>   ceph: consider inode's last read/write when calculating wanted caps
>   ceph: simplify calling of ceph_get_fmode()
>   ceph: remove delay check logic from ceph_check_caps()
> 
>  fs/ceph/caps.c               | 324 +++++++++++++++--------------------
>  fs/ceph/file.c               |  39 ++---
>  fs/ceph/inode.c              |  19 +-
>  fs/ceph/ioctl.c              |   2 +
>  fs/ceph/mds_client.c         |   5 -
>  fs/ceph/super.h              |  35 ++--
>  include/linux/ceph/ceph_fs.h |   1 +
>  7 files changed, 188 insertions(+), 237 deletions(-)
> 
> changes since v2
>  - make __ceph_caps_file_wanted more readable
>  - add patch 5 and 6, which fix hung write during testing patch 1~4
> 

This patch series causes some serious slowdown in the async dirops
patches that I've not yet fully tracked down, and I suspect that they
may also be the culprit in these bugs:

    https://tracker.ceph.com/issues/44381
    https://tracker.ceph.com/issues/44382

I'm going to drop this series from the testing branch for now, until we
can track down the issue.

Thanks,
Yan, Zheng March 3, 2020, 4:23 p.m. UTC | #3
On 3/3/20 3:53 AM, Jeff Layton wrote:
> On Fri, 2020-02-28 at 19:55 +0800, Yan, Zheng wrote:
>> This series make cephfs client not request caps for open files that
>> idle for a long time. For the case that one active client and multiple
>> standby clients open the same file, this increase the possibility that
>> mds issues exclusive caps to the active client.
>>
>> Yan, Zheng (4):
>>    ceph: always renew caps if mds_wanted is insufficient
>>    ceph: consider inode's last read/write when calculating wanted caps
>>    ceph: simplify calling of ceph_get_fmode()
>>    ceph: remove delay check logic from ceph_check_caps()
>>
>>   fs/ceph/caps.c               | 324 +++++++++++++++--------------------
>>   fs/ceph/file.c               |  39 ++---
>>   fs/ceph/inode.c              |  19 +-
>>   fs/ceph/ioctl.c              |   2 +
>>   fs/ceph/mds_client.c         |   5 -
>>   fs/ceph/super.h              |  35 ++--
>>   include/linux/ceph/ceph_fs.h |   1 +
>>   7 files changed, 188 insertions(+), 237 deletions(-)
>>
>> changes since v2
>>   - make __ceph_caps_file_wanted more readable
>>   - add patch 5 and 6, which fix hung write during testing patch 1~4
>>
> 
> This patch series causes some serious slowdown in the async dirops
> patches that I've not yet fully tracked down, and I suspect that they
> may also be the culprit in these bugs:
> 

slow down which tests?

>      https://tracker.ceph.com/issues/44381

this is because I forgot to check if inode is snap when queue delayed 
check. But it can't explain slow down.

>      https://tracker.ceph.com/issues/44382
> 
> I'm going to drop this series from the testing branch for now, until we
> can track down the issue.
> 
> Thanks,
>
Jeffrey Layton March 3, 2020, 8:17 p.m. UTC | #4
On Wed, 2020-03-04 at 00:23 +0800, Yan, Zheng wrote:
> On 3/3/20 3:53 AM, Jeff Layton wrote:
> > On Fri, 2020-02-28 at 19:55 +0800, Yan, Zheng wrote:
> > > This series make cephfs client not request caps for open files that
> > > idle for a long time. For the case that one active client and multiple
> > > standby clients open the same file, this increase the possibility that
> > > mds issues exclusive caps to the active client.
> > > 
> > > Yan, Zheng (4):
> > >    ceph: always renew caps if mds_wanted is insufficient
> > >    ceph: consider inode's last read/write when calculating wanted caps
> > >    ceph: simplify calling of ceph_get_fmode()
> > >    ceph: remove delay check logic from ceph_check_caps()
> > > 
> > >   fs/ceph/caps.c               | 324 +++++++++++++++--------------------
> > >   fs/ceph/file.c               |  39 ++---
> > >   fs/ceph/inode.c              |  19 +-
> > >   fs/ceph/ioctl.c              |   2 +
> > >   fs/ceph/mds_client.c         |   5 -
> > >   fs/ceph/super.h              |  35 ++--
> > >   include/linux/ceph/ceph_fs.h |   1 +
> > >   7 files changed, 188 insertions(+), 237 deletions(-)
> > > 
> > > changes since v2
> > >   - make __ceph_caps_file_wanted more readable
> > >   - add patch 5 and 6, which fix hung write during testing patch 1~4
> > > 
> > 
> > This patch series causes some serious slowdown in the async dirops
> > patches that I've not yet fully tracked down, and I suspect that they
> > may also be the culprit in these bugs:
> > 
> 
> slow down which tests?
> 

Most of the simple tests I was doing to sanity check async dirops.
Basically, this script was not seeing speed gain with async dirops
enabled:

-----------------8<-------------------
#!/bin/sh

MOUNTPOINT=/mnt/cephfs
TESTDIR=$MOUNTPOINT/test-dirops.$$

mkdir $TESTDIR
stat $TESTDIR
echo "Creating files in $TESTDIR"
time for i in `seq 1 10000`; do
    echo "foobarbaz" > $TESTDIR/$i
done
echo; echo "sync"
time sync
echo "Starting rm"
time rm -f $TESTDIR/*
echo; echo "rmdir"
time rmdir $TESTDIR
echo; echo "sync"
time sync
-----------------8<-------------------

It mostly seemed like it was just not getting caps in some cases.
Cranking up dynamic_debug seemed to make the problem go away, which led
me to believe there was probably a race condition in there.

At this point, I've gone ahead and merged the async dirops patches into
testing, so if you could rebase this on top of the current testing
branch and repost, I'll test them out again.

> >      https://tracker.ceph.com/issues/44381
> 
> this is because I forgot to check if inode is snap when queue delayed 
> check. But it can't explain slow down.
>

Ok, good to know.

> >      https://tracker.ceph.com/issues/44382
> > 
> > I'm going to drop this series from the testing branch for now, until we
> > can track down the issue.
> > 
> > 

Thanks,
Yan, Zheng March 4, 2020, 1:41 p.m. UTC | #5
On Wed, Mar 4, 2020 at 5:48 AM Jeff Layton <jlayton@kernel.org> wrote:
>
> On Wed, 2020-03-04 at 00:23 +0800, Yan, Zheng wrote:
> > On 3/3/20 3:53 AM, Jeff Layton wrote:
> > > On Fri, 2020-02-28 at 19:55 +0800, Yan, Zheng wrote:
> > > > This series make cephfs client not request caps for open files that
> > > > idle for a long time. For the case that one active client and multiple
> > > > standby clients open the same file, this increase the possibility that
> > > > mds issues exclusive caps to the active client.
> > > >
> > > > Yan, Zheng (4):
> > > >    ceph: always renew caps if mds_wanted is insufficient
> > > >    ceph: consider inode's last read/write when calculating wanted caps
> > > >    ceph: simplify calling of ceph_get_fmode()
> > > >    ceph: remove delay check logic from ceph_check_caps()
> > > >
> > > >   fs/ceph/caps.c               | 324 +++++++++++++++--------------------
> > > >   fs/ceph/file.c               |  39 ++---
> > > >   fs/ceph/inode.c              |  19 +-
> > > >   fs/ceph/ioctl.c              |   2 +
> > > >   fs/ceph/mds_client.c         |   5 -
> > > >   fs/ceph/super.h              |  35 ++--
> > > >   include/linux/ceph/ceph_fs.h |   1 +
> > > >   7 files changed, 188 insertions(+), 237 deletions(-)
> > > >
> > > > changes since v2
> > > >   - make __ceph_caps_file_wanted more readable
> > > >   - add patch 5 and 6, which fix hung write during testing patch 1~4
> > > >
> > >
> > > This patch series causes some serious slowdown in the async dirops
> > > patches that I've not yet fully tracked down, and I suspect that they
> > > may also be the culprit in these bugs:
> > >
> >
> > slow down which tests?
> >
>
> Most of the simple tests I was doing to sanity check async dirops.
> Basically, this script was not seeing speed gain with async dirops
> enabled:
>
> -----------------8<-------------------
> #!/bin/sh
>
> MOUNTPOINT=/mnt/cephfs
> TESTDIR=$MOUNTPOINT/test-dirops.$$
>
> mkdir $TESTDIR
> stat $TESTDIR
> echo "Creating files in $TESTDIR"
> time for i in `seq 1 10000`; do
>     echo "foobarbaz" > $TESTDIR/$i
> done
> echo; echo "sync"
> time sync
> echo "Starting rm"
> time rm -f $TESTDIR/*
> echo; echo "rmdir"
> time rmdir $TESTDIR
> echo; echo "sync"
> time sync
> -----------------8<-------------------
>
> It mostly seemed like it was just not getting caps in some cases.
> Cranking up dynamic_debug seemed to make the problem go away, which led
> me to believe there was probably a race condition in there.
>

should be fixed by patch "mds: update dentry lease for async create".
I saw dput() from ceph_mdsc_release_request() pruned dentry and
cleared dir's complete flags.  The issue happens if client gets safe
replies quickly.

I also found that mds  revoke Fsx when fragmenting dirfrags. Fixed by
following PR
https://github.com/ceph/ceph/pull/33719

> At this point, I've gone ahead and merged the async dirops patches into
> testing, so if you could rebase this on top of the current testing
> branch and repost, I'll test them out again.
>
> > >      https://tracker.ceph.com/issues/44381
> >
> > this is because I forgot to check if inode is snap when queue delayed
> > check. But it can't explain slow down.
> >
>
> Ok, good to know.
>
> > >      https://tracker.ceph.com/issues/44382
> > >
> > > I'm going to drop this series from the testing branch for now, until we
> > > can track down the issue.
> > >
> > >
>
> Thanks,
> --
> Jeff Layton <jlayton@kernel.org>
>