From 88e0fbc452ed94393bf89585c2b90edb94749b45 Mon Sep 17 00:00:00 2001 From: Nick Piggin Date: Tue, 22 Sep 2009 16:43:50 -0700 Subject: 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: [] io_schedule+0x63/0xa5 [] sync_buffer+0x3b/0x3f [] __wait_on_bit+0x47/0x79 [] out_of_line_wait_on_bit+0x6a/0x77 [] __wait_on_buffer+0x1f/0x21 [] __bread+0x70/0x86 [] :udf:udf_tread+0x38/0x3a [] :udf:udf_update_inode+0x4d/0x68c [] :udf:udf_write_inode+0x1d/0x2b [] __writeback_single_inode+0x1c0/0x394 [] write_inode_now+0x7d/0xc4 [] :udf:udf_clear_inode+0x3d/0x53 [] clear_inode+0xc2/0x11b [] dispose_list+0x5b/0x102 [] shrink_icache_memory+0x1dd/0x213 [] shrink_slab+0xe3/0x158 [] try_to_free_pages+0x177/0x232 [] __alloc_pages+0x1fa/0x392 [] alloc_page_vma+0x176/0x189 [] __do_fault+0x10c/0x417 [] handle_mm_fault+0x466/0x940 [] 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: [] __mutex_lock_slowpath+0x72/0xa9 [] mutex_lock+0x1e/0x22 [] shrink_icache_memory+0x49/0x213 [] shrink_slab+0xe3/0x158 [] try_to_free_pages+0x177/0x232 [] __alloc_pages+0x1fa/0x392 [] alloc_pages_current+0xd1/0xd6 [] __get_free_pages+0xe/0x4d [] __pollwait+0x5e/0xdf [] :nvidia:nv_kern_poll+0x2e/0x73 [] do_select+0x308/0x506 [] core_sys_select+0x1a6/0x254 [] 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 Cc: Jan Kara Cc: Al Viro Cc: Christoph Hellwig Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/inode.c | 19 ++++++++++++------- 1 file changed, 12 insertions(+), 7 deletions(-) (limited to 'fs') diff --git a/fs/inode.c b/fs/inode.c index f5ff71cb3e2..76582b06ab9 100644 --- a/fs/inode.c +++ b/fs/inode.c @@ -14,6 +14,7 @@ #include #include #include +#include #include #include #include @@ -87,14 +88,18 @@ static struct hlist_head *inode_hashtable __read_mostly; DEFINE_SPINLOCK(inode_lock); /* - * iprune_mutex provides exclusion between the kswapd or try_to_free_pages + * iprune_sem provides exclusion between the kswapd or try_to_free_pages * icache shrinking path, and the umount path. Without this exclusion, * by the time prune_icache calls iput for the inode whose pages it has * been invalidating, or by the time it calls clear_inode & destroy_inode * from its final dispose_list, the struct super_block they refer to * (for inode->i_sb->s_op) may already have been freed and reused. + * + * We make this an rwsem because the fastpath is icache shrinking. In + * some cases a filesystem may be doing a significant amount of work in + * its inode reclaim code, so this should improve parallelism. */ -static DEFINE_MUTEX(iprune_mutex); +static DECLARE_RWSEM(iprune_sem); /* * Statistics gathering.. @@ -381,7 +386,7 @@ static int invalidate_list(struct list_head *head, struct list_head *dispose) /* * We can reschedule here without worrying about the list's * consistency because the per-sb list of inodes must not - * change during umount anymore, and because iprune_mutex keeps + * change during umount anymore, and because iprune_sem keeps * shrink_icache_memory() away. */ cond_resched_lock(&inode_lock); @@ -420,7 +425,7 @@ int invalidate_inodes(struct super_block *sb) int busy; LIST_HEAD(throw_away); - mutex_lock(&iprune_mutex); + down_write(&iprune_sem); spin_lock(&inode_lock); inotify_unmount_inodes(&sb->s_inodes); fsnotify_unmount_inodes(&sb->s_inodes); @@ -428,7 +433,7 @@ int invalidate_inodes(struct super_block *sb) spin_unlock(&inode_lock); dispose_list(&throw_away); - mutex_unlock(&iprune_mutex); + up_write(&iprune_sem); return busy; } @@ -467,7 +472,7 @@ static void prune_icache(int nr_to_scan) int nr_scanned; unsigned long reap = 0; - mutex_lock(&iprune_mutex); + down_read(&iprune_sem); spin_lock(&inode_lock); for (nr_scanned = 0; nr_scanned < nr_to_scan; nr_scanned++) { struct inode *inode; @@ -509,7 +514,7 @@ static void prune_icache(int nr_to_scan) spin_unlock(&inode_lock); dispose_list(&freeable); - mutex_unlock(&iprune_mutex); + up_read(&iprune_sem); } /* -- cgit v1.2.3-18-g5258