lkml.org 
[lkml]   [2019]   [Mar]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH] ceph: Fix a memory leak in ci->i_head_snapc
Date
"Yan, Zheng" <ukernel@gmail.com> writes:

> On Fri, Mar 15, 2019 at 7:13 PM Luis Henriques <lhenriques@suse.com> wrote:
>>
>> I'm occasionally seeing a kmemleak warning in xfstest generic/013:
>>
>> unreferenced object 0xffff8881fccca940 (size 32):
>> comm "kworker/0:1", pid 12, jiffies 4295005883 (age 130.648s)
>> hex dump (first 32 bytes):
>> 01 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 ................
>> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
>> backtrace:
>> [<00000000d741a1ea>] build_snap_context+0x5b/0x2a0
>> [<0000000021a00533>] rebuild_snap_realms+0x27/0x90
>> [<00000000ac538600>] rebuild_snap_realms+0x42/0x90
>> [<000000000e955fac>] ceph_update_snap_trace+0x2ee/0x610
>> [<00000000a9550416>] ceph_handle_snap+0x317/0x5f3
>> [<00000000fc287b83>] dispatch+0x362/0x176c
>> [<00000000a312c741>] ceph_con_workfn+0x9ce/0x2cf0
>> [<000000004168e3a9>] process_one_work+0x1d4/0x400
>> [<000000002188e9e7>] worker_thread+0x2d/0x3c0
>> [<00000000b593e4b3>] kthread+0x112/0x130
>> [<00000000a8587dca>] ret_from_fork+0x35/0x40
>> [<00000000ba1c9c1d>] 0xffffffffffffffff
>>
>> It looks like it is possible that we miss a flush_ack from the MDS when,
>> for example, umounting the filesystem. In that case, we can simply drop
>> the reference to the ceph_snap_context obtained in ceph_queue_cap_snap().
>>
>> Link: https://tracker.ceph.com/issues/38224
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Luis Henriques <lhenriques@suse.com>
>> ---
>> fs/ceph/caps.c | 7 +++++++
>> 1 file changed, 7 insertions(+)
>>
>> diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
>> index 36a8dc699448..208f4dc6f574 100644
>> --- a/fs/ceph/caps.c
>> +++ b/fs/ceph/caps.c
>> @@ -1054,6 +1054,7 @@ int ceph_is_any_caps(struct inode *inode)
>> static void drop_inode_snap_realm(struct ceph_inode_info *ci)
>> {
>> struct ceph_snap_realm *realm = ci->i_snap_realm;
>> +
>> spin_lock(&realm->inodes_with_caps_lock);
>> list_del_init(&ci->i_snap_realm_item);
>> ci->i_snap_realm_counter++;
>> @@ -1063,6 +1064,12 @@ static void drop_inode_snap_realm(struct ceph_inode_info *ci)
>> spin_unlock(&realm->inodes_with_caps_lock);
>> ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
>> realm);
>> + /*
>> + * ci->i_head_snapc should be NULL, but we may still be waiting for a
>> + * flush_ack from the MDS. In that case, we still hold a ref for the
>> + * ceph_snap_context and we need to drop it.
>> + */
>> + ceph_put_snap_context(ci->i_head_snapc);
>> }
>>
>> /*
>
> This does not seem right. i_head_snapc is cleared when
> (ci->i_wrbuffer_ref_head == 0 && ci->i_dirty_caps == 0 &&
> ci->i_flushing_caps == 0) . Nothing do with dropping ci->i_snap_realm.
> Did you see 'reconnect denied' during the test? If you did, the fix
> should be in iterate_session_caps()
>

No, I didn't saw any 'reconnect denied' in the test. The test actually
seems to execute fine, except from the memory leak.

It's very difficult to reproduce this issue, but last time I managed to
get this memory leak to trigger I actually had some debugging code in
drop_inode_snap_realm, something like:

if (ci->i_head_snapc)
printk("i_head_snapc: 0x%px\n", ci->i_head_snapc);

This printk was only executed when the bug triggered (during a
filesystem umount) and the address shown was the same as in the kmemleak
warning.

After spending some time looking, I assumed this to be a missing call to
handle_cap_flush_ack, which would do the i_head_snapc cleanup.

Cheers,
--
Luis

\
 
 \ /
  Last update: 2019-03-18 11:35    [W:2.538 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site