Re: [PATCH 0/9 v3] locks: avoid thundering-herd wake-ups
From: Jeff Layton
Date: Tue Oct 30 2018 - 08:05:03 EST
On Mon, 2018-10-29 at 08:38 -0400, Jeff Layton wrote:
> On Mon, 2018-10-29 at 12:56 +1100, NeilBrown wrote:
> > On Fri, Oct 26 2018, Jeff Layton wrote:
> >
> > > On Wed, 2018-10-24 at 09:43 +1100, NeilBrown wrote:
> > > > This took longer that I had wanted, due to various reasons - sorry.
> > > > And I'm now posting it in a merge window, which is not ideal. I don't
> > > > expect it to be included in this merge window and I won't be at all
> > > > impatient for review, but I didn't want to delay it further.
> > > >
> > > > Testing found some problems, particularly the need to use
> > > > locks_copy_lock in NFS. And there was a small thinko in there that
> > > > effectively removed all the speed gains :-(
> > > >
> > > > But this version:
> > > > - shows excellent scalability with lots of threads on lots of CPUs
> > > > contending on a single file
> > > > - avoids the problems that Bruce found
> > > > - seems to work.
> > > >
> > > > Thanks,
> > > > NeilBrown
> > > >
> > > >
> > > > ---
> > > >
> > > > NeilBrown (9):
> > > > fs/locks: rename some lists and pointers.
> > > > fs/locks: split out __locks_wake_up_blocks().
> > > > NFS: use locks_copy_lock() to copy locks.
> > > > fs/locks: allow a lock request to block other requests.
> > > > fs/locks: always delete_block after waiting.
> > > > fs/locks: change all *_conflict() functions to return bool.
> > > > fs/locks: create a tree of dependent requests.
> > > > locks: merge posix_unblock_lock() and locks_delete_block()
> > > > VFS: locks: remove unnecessary white space.
> > > >
> > > >
> > > > fs/cifs/file.c | 4 -
> > > > fs/lockd/svclock.c | 2
> > > > fs/locks.c | 231 ++++++++++++++++++++++-----------------
> > > > fs/nfs/nfs4proc.c | 6 +
> > > > fs/nfsd/nfs4state.c | 6 +
> > > > include/linux/fs.h | 11 +-
> > > > include/trace/events/filelock.h | 16 +--
> > > > 7 files changed, 153 insertions(+), 123 deletions(-)
> > > >
> > > > --
> > > > Signature
> > > >
> > >
> > >
> > > I built a kernel with these patches and ran the cthon04 lock tests and
> > > got this on lock test 1 after a long hang (the test passed though):
> >
> > I've been looking deeper into this, and I cannot see how this can
> > happen.
> >
> > This is an unlock request, happening when a file is closed.
> > locks_delete_block() will only be called from locks_lock_inode_wait()
> > after posix_lock_inode() (or possible flock_lock_inode()) returns
> > FILE_LOCK_DEFERRED.
> >
> > But these only return FILE_LOCK_DEFERRED when fl_type != F_UNLCK.
> > (or possibly if FL_ACCESS and FL_SLEEP are both set - that would be
> > weird).
> >
> > So this shouldn't happen - an unlock request should never result in
> > locks_delete_block() being called.
> > But if it can, I'll need to change do_flock() in gfs2/file.c to use a
> > properly initialized 'struct file_lock', rather than a manifest
> > constant. Maybe I should do that anyway.
> >
> > Any ideas? I'll try running connectathon myself later and see if I can
> > reproduce it.
> >
> > Thanks,
> > NeilBrown
> >
> > >
> > > [ 1694.787367] BUG: unable to handle kernel NULL pointer dereference at 000000000000002c
> > > [ 1694.789546] PGD 118ff0067 P4D 118ff0067 PUD 135915067 PMD 0
> > > [ 1694.790772] Oops: 0000 [#1] SMP NOPTI
> > > [ 1694.791749] CPU: 7 PID: 1514 Comm: tlocklfs Not tainted 4.19.0+ #56
> > > [ 1694.792876] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20180531_142017-buildhw-08.phx2.fedoraproject.org-1.fc28 04/01/2014
> > > [ 1694.795179] RIP: 0010:__locks_delete_block+0x14/0x90
> > > [ 1694.796203] Code: 01 a1 e9 9f 4f d8 ff 0f 1f 44 00 00 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 8b 05 29 9d 58 01 55 53 48 89 fb 85 c0 75 5a <48> 8b 43 20 48 85 c0 74 20 48 8b 53 18 48 89 10 48 85 d2 74 04 48
> > > [ 1694.799277] RSP: 0018:ffff9d21c1f63cb8 EFLAGS: 00010202
> > > [ 1694.800374] RAX: 0000000000000001 RBX: 000000000000000c RCX: aaaaaaaaaaaaaaad
> > > [ 1694.801682] RDX: 0000000000000001 RSI: ffffffff9f7b0c38 RDI: 0000000000000246
> > > [ 1694.802996] RBP: 000000000000000c R08: 0000000000000000 R09: 0000000000000001
> > > [ 1694.804317] R10: 0000000000000001 R11: ffffffffa0bdc188 R12: ffff9d21c1f63dd8
> > > [ 1694.805633] R13: ffff9d21c1f63e00 R14: ffffffff9f3241a8 R15: ffff8d0b5aef72e0
> > > [ 1694.806941] FS: 00007efc8699c740(0000) GS:ffff8d0b7ba00000(0000) knlGS:0000000000000000
> > > [ 1694.808380] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > [ 1694.809550] CR2: 000000000000002c CR3: 000000011e0d8000 CR4: 00000000000006e0
> > > [ 1694.810888] Call Trace:
> > > [ 1694.811692] __locks_wake_up_blocks+0x2d/0x80
> > > [ 1694.812713] locks_delete_block+0x1d/0x40
> > > [ 1694.813691] locks_lock_inode_wait+0x9c/0x1c0
> > > [ 1694.814731] nfs4_proc_lock+0x120/0x440 [nfsv4]
> > > [ 1694.815786] ? nfs_put_lock_context+0x25/0x80 [nfs]
> > > [ 1694.816866] ? do_unlk+0x98/0xe0 [nfs]
> > > [ 1694.817818] locks_remove_posix+0xba/0x1d0
> > > [ 1694.818811] ? _cond_resched+0x15/0x30
> > > [ 1694.819768] ? wait_on_commit+0x38/0xb0 [nfs]
> > > [ 1694.820787] ? process_echoes+0x60/0x60
> > > [ 1694.821752] ? __nfs_commit_inode+0xc2/0x1c0 [nfs]
> > > [ 1694.822819] filp_close+0x56/0x70
> > > [ 1694.823712] __x64_sys_close+0x1e/0x50
> > > [ 1694.824661] do_syscall_64+0x60/0x1f0
> > > [ 1694.825599] entry_SYSCALL_64_after_hwframe+0x49/0xbe
> > > [ 1694.826731] RIP: 0033:0x7efc8616c0a4
> > > [ 1694.827673] Code: eb 89 e8 af f6 01 00 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 8b 05 aa e7 2c 00 48 63 ff 85 c0 75 13 b8 03 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 44 f3 c3 66 90 48 83 ec 18 48 89 7c 24 08 e8
> > > [ 1694.830929] RSP: 002b:00007ffc70beb7b8 EFLAGS: 00000246 ORIG_RAX: 0000000000000003
> > > [ 1694.832371] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007efc8616c0a4
> > > [ 1694.833784] RDX: 0000000000000000 RSI: 00007efc864378a0 RDI: 0000000000000009
> > > [ 1694.835183] RBP: 00007ffc70beb7d0 R08: 00007efc864378a0 R09: 00007efc8699c740
> > > [ 1694.836560] R10: 00000000000006b4 R11: 0000000000000246 R12: 0000000000401000
> > > [ 1694.837941] R13: 00007ffc70beb990 R14: 0000000000000000 R15: 0000000000000000
> > > [ 1694.839322] Modules linked in: rpcsec_gss_krb5 nfsv4 dns_resolver nfs fscache nfsd auth_rpcgss nfs_acl lockd grace ip6t_rpfilter ip6t_REJECT nf_reject_ipv6 xt_conntrack ip_set nfnetlink ebtable_nat ebtable_broute bridge stp llc ip6table_nat nf_nat_ipv6 ip6table_mangle ip6table_raw ip6table_security iptable_nat nf_nat_ipv4 nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 iptable_mangle iptable_raw iptable_security ebtable_filter ebtables ip6table_filter ip6_tables sunrpc joydev virtio_balloon i2c_piix4 pcspkr edac_mce_amd xfs libcrc32c serio_raw virtio_blk qxl drm_kms_helper virtio_net ttm net_failover virtio_console failover drm ata_generic pata_acpi floppy qemu_fw_cfg
> > > [ 1694.849736] CR2: 000000000000002c
> > > [ 1694.850813] ---[ end trace da2f469c62deb564 ]---
> > >
> > > --
> > > Jeff Layton <jlayton@xxxxxxxxxx>
>
> Yes. It crashed here:
>
> (gdb) list *(__locks_delete_block+0x14)
> 0xffffffff81391374 is in __locks_delete_block
> (./include/linux/list.h:693).
> 688 n->pprev = LIST_POISON2;
> 689 }
> 690
> 691 static inline void hlist_del_init(struct hlist_node *n)
> 692 {
> 693 if (!hlist_unhashed(n)) {
> 694 __hlist_del(n);
> 695 INIT_HLIST_NODE(n);
> 696 }
> 697 }
>
> ...and that should be the address of fl->fl_link.
>
> I think the issue is probably in locks_remove_posix. It creates a lock
> request on the stack, and doesn't seem to initialize fl_link. That used
> to be ok, but patch #5 in the series changes that.
>
> The following patch seems to fix the problem in a quick test. I'll plan
> to run some more tests later. It may be a day or two before I can get to
> it though.
>
> -----------------------8<---------------------------
>
> [PATCH] locks: initialize list heads in locks_remove_posix lock request
>
> Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> ---
> fs/locks.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/fs/locks.c b/fs/locks.c
> index 6b9da320d9a0..60019e146839 100644
> --- a/fs/locks.c
> +++ b/fs/locks.c
> @@ -2535,6 +2535,7 @@ void locks_remove_posix(struct file *filp, fl_owner_t owner)
> lock.fl_file = filp;
> lock.fl_ops = NULL;
> lock.fl_lmops = NULL;
> + locks_init_lock_heads(&lock);
>
> error = vfs_lock_file(filp, F_SETLK, &lock, NULL);
>
I ran a bunch of tests on top of your series + this patch and it seemed
to do fine. I'm guessing maybe you just got "lucky" and your kernel
happened to end up with the fl_link value set in such a way that
hlist_unhashed returned true there?
In any case, this seems to fix the issue I was seeing. The only question
I have is whether this will be more expensive than doing something more
clever like checking for FL_SLEEP in locks_remove_block.
I've gone ahead and pushed the set to my locks-next branch to start
getting more testing. If all goes well, I'll plan to send Linus a PR for
v4.21 (or v5.1).
Thanks again for doing this!
--
Jeff Layton <jlayton@xxxxxxxxxx>