Skip to contentSkip to conversation

uncollected nfsd open owners

DISCUSSION NeilBrown 3 participants 6 messages Oct 25, 2019 – Oct 28, 2019

Subsystems (2)
lore
Share & tools
lore

NeilBrown

author
#0 lore raw

Hi,
I have a coredump from a machine that was running as an NFS server. nfs4_laundromat was trying to expire a client, and in particular was cleaning up the ->cl_openowners.
As there were 6.5 million of these, it took rather longer than the softlockup timer thought was acceptable, and hence the core dump.

Those open owners that I looked at had empty so_stateids lists, so I would normally expect them to be on the close_lru and to be removed fairly soon. But they weren't (only 32 openowners on close_lru).

The only explanation I can think of for this is that maybe an OPEN request successfully got through nfs4_process_open1(), thus creating an open owner, but failed to get to or through nfs4_process_open2(), and so didn't add a stateid. I *think* this can leave an openowner that is unused but will never be cleaned up (until the client is expired, which might be too late).

Is this possible? If so, how should we handle those openowners which never had a stateid?
In 3.0 (which it the kernel were I saw this) I could probably just put the openowner on the close_lru when it is created.
In more recent kernels, it seems to be assumed that openowners are only on close_lru if they have a oo_last_closed_stid. Would we need a separate "never used lru", or should they just be destroyed as soon as the open fails?

Also, should we put a cond_resched() in some or all of those loops in __destroy_client() ??

Thanks for your help,
NeilBrown

#1 lore raw
Quoted 17 lines · NeilBrown

I have a coredump from a machine that was running as an NFS server. nfs4_laundromat was trying to expire a client, and in particular was cleaning up the ->cl_openowners.
As there were 6.5 million of these, it took rather longer than the softlockup timer thought was acceptable, and hence the core dump.

Those open owners that I looked at had empty so_stateids lists, so I would normally expect them to be on the close_lru and to be removed fairly soon. But they weren't (only 32 openowners on close_lru).

The only explanation I can think of for this is that maybe an OPEN request successfully got through nfs4_process_open1(), thus creating an open owner, but failed to get to or through nfs4_process_open2(), and so didn't add a stateid. I *think* this can leave an openowner that is unused but will never be cleaned up (until the client is expired, which might be too late).

Quoted 8 lines

Is this possible? If so, how should we handle those openowners which never had a stateid?
In 3.0 (which it the kernel were I saw this) I could probably just put the openowner on the close_lru when it is created.
In more recent kernels, it seems to be assumed that openowners are only on close_lru if they have a oo_last_closed_stid. Would we need a separate "never used lru", or should they just be destroyed as soon as the open fails?

Hopefully we can just throw the new openowner away when the open fails.

But it looks like the new openowner is visible on global data structures by then, so we need to be sure somebody else isn't about to use it.

Quoted 2 lines

Also, should we put a cond_resched() in some or all of those loops in __destroy_client() ??

Looks like it helped find a bug in this case....

Destroying a client that has a ton of active state should be an unusual situation.

I don't know, maybe? I'm sure this isn't the only spinlock-protected kernel code where we don't have a strict bound on a loop, what's been the practice elsewhere? Worst case, the realtime code allows preempting spinlocks, right?

Might be nice to have some sort of limits on the number of objects (like stateowners) that can be created. But it's a pain when we get one of those limits wrong. (See
git log -L :nfsd4_get_drc_mem:fs/nfsd/nfs4state.c.)

--b.

#2 lore raw
Quoted 27 lines · J. Bruce Fields
Quoted 17 lines · NeilBrown

I have a coredump from a machine that was running as an NFS server. nfs4_laundromat was trying to expire a client, and in particular was cleaning up the ->cl_openowners.
As there were 6.5 million of these, it took rather longer than the softlockup timer thought was acceptable, and hence the core dump.

Those open owners that I looked at had empty so_stateids lists, so I would normally expect them to be on the close_lru and to be removed fairly soon. But they weren't (only 32 openowners on close_lru).

The only explanation I can think of for this is that maybe an OPEN request successfully got through nfs4_process_open1(), thus creating an open owner, but failed to get to or through nfs4_process_open2(), and so didn't add a stateid. I *think* this can leave an openowner that is unused but will never be cleaned up (until the client is expired, which might be too late).

Quoted 8 lines

Is this possible? If so, how should we handle those openowners which never had a stateid?
In 3.0 (which it the kernel were I saw this) I could probably just put the openowner on the close_lru when it is created.
In more recent kernels, it seems to be assumed that openowners are only on close_lru if they have a oo_last_closed_stid. Would we need a separate "never used lru", or should they just be destroyed as soon as the open fails?

Quoted 4 lines

Hopefully we can just throw the new openowner away when the open fails.

But it looks like the new openowner is visible on global data structures by then, so we need to be sure somebody else isn't about to use it.

But, also, if this has only been seen on 3.0, it may have been fixed already. It sounds like kind of a familiar problem, but I didn't spot a relevant commit on a quick look through the logs.

--b.

author
#3 lore raw
Quoted 28 lines · J. Bruce Fields
Quoted 27 lines · J. Bruce Fields
Quoted 25 lines · NeilBrown

I have a coredump from a machine that was running as an NFS server. nfs4_laundromat was trying to expire a client, and in particular was cleaning up the ->cl_openowners.
As there were 6.5 million of these, it took rather longer than the softlockup timer thought was acceptable, and hence the core dump.

Those open owners that I looked at had empty so_stateids lists, so I would normally expect them to be on the close_lru and to be removed fairly soon. But they weren't (only 32 openowners on close_lru).

The only explanation I can think of for this is that maybe an OPEN request successfully got through nfs4_process_open1(), thus creating an open owner, but failed to get to or through nfs4_process_open2(), and so didn't add a stateid. I *think* this can leave an openowner that is unused but will never be cleaned up (until the client is expired, which might be too late).

Is this possible? If so, how should we handle those openowners which never had a stateid?
In 3.0 (which it the kernel were I saw this) I could probably just put the openowner on the close_lru when it is created.
In more recent kernels, it seems to be assumed that openowners are only on close_lru if they have a oo_last_closed_stid. Would we need a separate "never used lru", or should they just be destroyed as soon as the open fails?

Quoted 9 lines
Quoted 4 lines

Hopefully we can just throw the new openowner away when the open fails.

But it looks like the new openowner is visible on global data structures by then, so we need to be sure somebody else isn't about to use it.

But, also, if this has only been seen on 3.0, it may have been fixed already. It sounds like kind of a familiar problem, but I didn't spot a relevant commit on a quick look through the logs.

I guess I shouldn't expect you to remember something from 8 years ago. This seems a perfect fit for what I see:

commit d29b20cd589128a599e5045d4effc2d7dbc388f5

Author: J. Bruce Fields <[email protected]>
Date:   Thu Oct 13 15:12:59 2011 -0400

nfsd4: clean up open owners on OPEN failure

If process_open1() creates a new open owner, but the open later fails, the current code will leave the open owner around. It won't be on the close_lru list, and the client isn't expected to send a CLOSE, so it will hang around as long as the client does.

Similarly, if process_open1() removes an existing open owner from the close lru, anticipating that an open owner that previously had no associated stateid's now will, but the open subsequently fails, then we'll again be left with the same leak.

Fix both problems.

I wonder if this is safe to backport ... 3.2 is fairly close to 3.0, but there have been lots of changes to the code since then ... maybe there are more bug fixes.
I'll work something out.

Thanks,
NeilBrown

#4 lore raw
Quoted 58 lines · NeilBrown
Quoted 28 lines · J. Bruce Fields
Quoted 27 lines · J. Bruce Fields
Quoted 25 lines · NeilBrown

I have a coredump from a machine that was running as an NFS server. nfs4_laundromat was trying to expire a client, and in particular was cleaning up the ->cl_openowners.
As there were 6.5 million of these, it took rather longer than the softlockup timer thought was acceptable, and hence the core dump.

Those open owners that I looked at had empty so_stateids lists, so I would normally expect them to be on the close_lru and to be removed fairly soon. But they weren't (only 32 openowners on close_lru).

The only explanation I can think of for this is that maybe an OPEN request successfully got through nfs4_process_open1(), thus creating an open owner, but failed to get to or through nfs4_process_open2(), and so didn't add a stateid. I *think* this can leave an openowner that is unused but will never be cleaned up (until the client is expired, which might be too late).

Is this possible? If so, how should we handle those openowners which never had a stateid?
In 3.0 (which it the kernel were I saw this) I could probably just put the openowner on the close_lru when it is created.
In more recent kernels, it seems to be assumed that openowners are only on close_lru if they have a oo_last_closed_stid. Would we need a separate "never used lru", or should they just be destroyed as soon as the open fails?

Quoted 9 lines
Quoted 4 lines

Hopefully we can just throw the new openowner away when the open fails.

But it looks like the new openowner is visible on global data structures by then, so we need to be sure somebody else isn't about to use it.

But, also, if this has only been seen on 3.0, it may have been fixed already. It sounds like kind of a familiar problem, but I didn't spot a relevant commit on a quick look through the logs.

I guess I shouldn't expect you to remember something from 8 years ago. This seems a perfect fit for what I see:

commit d29b20cd589128a599e5045d4effc2d7dbc388f5

Author: J. Bruce Fields <[email protected]>
Date:   Thu Oct 13 15:12:59 2011 -0400

nfsd4: clean up open owners on OPEN failure

If process_open1() creates a new open owner, but the open later fails, the current code will leave the open owner around. It won't be on the close_lru list, and the client isn't expected to send a CLOSE, so it will hang around as long as the client does.

Similarly, if process_open1() removes an existing open owner from the close lru, anticipating that an open owner that previously had no associated stateid's now will, but the open subsequently fails, then we'll again be left with the same leak.

Quoted 6 lines

Fix both problems.

I wonder if this is safe to backport ... 3.2 is fairly close to 3.0, but there have been lots of changes to the code since then ... maybe there are more bug fixes.
I'll work something out.

This is all before the breakup of the big state lock, so you don't have that to worry about, at least.

Anyway, not even remembering that commit, I'm afraid I'm no more expert on that era of the history that you at this point....

--b.

author
#5 lore raw
In reply to J. Bruce Fields
Quoted 6 lines · J. Bruce Fields
Quoted 3 lines

Also, should we put a cond_resched() in some or all of those loops in __destroy_client() ??

Looks like it helped find a bug in this case....

Quoted 7 lines

Destroying a client that has a ton of active state should be an unusual situation.

I don't know, maybe? I'm sure this isn't the only spinlock-protected kernel code where we don't have a strict bound on a loop, what's been the practice elsewhere? Worst case, the realtime code allows preempting spinlocks, right?

git grep cond_resched_lock

But most of __destroy_client isn't protected by a spinlock....

I dunno - maybe it doesn't matter.

Quoted 4 lines

Might be nice to have some sort of limits on the number of objects (like stateowners) that can be created. But it's a pain when we get one of those limits wrong. (See
git log -L :nfsd4_get_drc_mem:fs/nfsd/nfs4state.c.)

Grin...

Quoted 1 line

NeilBrown