fs: turn iprune_mutex into rwsem
authorNick Piggin <npiggin@suse.de>
Tue, 22 Sep 2009 23:43:50 +0000 (16:43 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 23 Sep 2009 14:39:29 +0000 (07:39 -0700)
commit88e0fbc452ed94393bf89585c2b90edb94749b45
tree374b7c8f397609da7a7b22ddee77a9c70956fcb1
parent70867453092297be9afb2249e712a1f960ec0a09
fs: turn iprune_mutex into rwsem

We have had a report of bad memory allocation latency during DVD-RAM (UDF)
writing.  This is causing the user's desktop session to become unusable.

Jan tracked the cause of this down to UDF inode reclaim blocking:

gnome-screens D ffff810006d1d598     0 20686      1
 ffff810006d1d508 0000000000000082 ffff810037db6718 0000000000000800
 ffff810006d1d488 ffffffff807e4280 ffffffff807e4280 ffff810006d1a580
 ffff8100bccbc140 ffff810006d1a8c0 0000000006d1d4e8 ffff810006d1a8c0
Call Trace:
 [<ffffffff804477f3>] io_schedule+0x63/0xa5
 [<ffffffff802c2587>] sync_buffer+0x3b/0x3f
 [<ffffffff80447d2a>] __wait_on_bit+0x47/0x79
 [<ffffffff80447dc6>] out_of_line_wait_on_bit+0x6a/0x77
 [<ffffffff802c24f6>] __wait_on_buffer+0x1f/0x21
 [<ffffffff802c442a>] __bread+0x70/0x86
 [<ffffffff88de9ec7>] :udf:udf_tread+0x38/0x3a
 [<ffffffff88de0fcf>] :udf:udf_update_inode+0x4d/0x68c
 [<ffffffff88de26e1>] :udf:udf_write_inode+0x1d/0x2b
 [<ffffffff802bcf85>] __writeback_single_inode+0x1c0/0x394
 [<ffffffff802bd205>] write_inode_now+0x7d/0xc4
 [<ffffffff88de2e76>] :udf:udf_clear_inode+0x3d/0x53
 [<ffffffff802b39ae>] clear_inode+0xc2/0x11b
 [<ffffffff802b3ab1>] dispose_list+0x5b/0x102
 [<ffffffff802b3d35>] shrink_icache_memory+0x1dd/0x213
 [<ffffffff8027ede3>] shrink_slab+0xe3/0x158
 [<ffffffff8027fbab>] try_to_free_pages+0x177/0x232
 [<ffffffff8027a578>] __alloc_pages+0x1fa/0x392
 [<ffffffff802951fa>] alloc_page_vma+0x176/0x189
 [<ffffffff802822d8>] __do_fault+0x10c/0x417
 [<ffffffff80284232>] handle_mm_fault+0x466/0x940
 [<ffffffff8044b922>] do_page_fault+0x676/0xabf

This blocks with iprune_mutex held, which then blocks other reclaimers:

X             D ffff81009d47c400     0 17285  14831
 ffff8100844f3728 0000000000000086 0000000000000000 ffff81000000e288
 ffff81000000da00 ffffffff807e4280 ffffffff807e4280 ffff81009d47c400
 ffffffff805ff890 ffff81009d47c740 00000000844f3808 ffff81009d47c740
Call Trace:
 [<ffffffff80447f8c>] __mutex_lock_slowpath+0x72/0xa9
 [<ffffffff80447e1a>] mutex_lock+0x1e/0x22
 [<ffffffff802b3ba1>] shrink_icache_memory+0x49/0x213
 [<ffffffff8027ede3>] shrink_slab+0xe3/0x158
 [<ffffffff8027fbab>] try_to_free_pages+0x177/0x232
 [<ffffffff8027a578>] __alloc_pages+0x1fa/0x392
 [<ffffffff8029507f>] alloc_pages_current+0xd1/0xd6
 [<ffffffff80279ac0>] __get_free_pages+0xe/0x4d
 [<ffffffff802ae1b7>] __pollwait+0x5e/0xdf
 [<ffffffff8860f2b4>] :nvidia:nv_kern_poll+0x2e/0x73
 [<ffffffff802ad949>] do_select+0x308/0x506
 [<ffffffff802adced>] core_sys_select+0x1a6/0x254
 [<ffffffff802ae0b7>] sys_select+0xb5/0x157

Now I think the main problem is having the filesystem block (and do IO) in
inode reclaim.  The problem is that this doesn't get accounted well and
penalizes a random allocator with a big latency spike caused by work
generated from elsewhere.

I think the best idea would be to avoid this.  By design if possible, or
by deferring the hard work to an asynchronous context.  If the latter,
then the fs would probably want to throttle creation of new work with
queue size of the deferred work, but let's not get into those details.

Anyway, the other obvious thing we looked at is the iprune_mutex which is
causing the cascading blocking.  We could turn this into an rwsem to
improve concurrency.  It is unreasonable to totally ban all potentially
slow or blocking operations in inode reclaim, so I think this is a cheap
way to get a small improvement.

This doesn't solve the whole problem of course.  The process doing inode
reclaim will still take the latency hit, and concurrent processes may end
up contending on filesystem locks.  So fs developers should keep these
problems in mind.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jan Kara <jack@ucw.cz>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/inode.c