On Fri, Aug 14, 2009 at 03:58:47PM -0700, Andrew Morton wrote:
> On Fri, 14 Aug 2009 17:25:05 +0200
> Nick Piggin <npiggin@xxxxxxx> wrote:
> > 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.
>
> Yes. Why does UDF do all that stuff in ->clear_inode()? Other
> filesystems have very simple, non-blocking, non-IO-doing
> ->clear_inode() implementations. This sounds like a design problem
> within UDF.
BTW. some filesystems do some work, and we can do blocking memory
allocations, take blocking locks etc, and a fair bit of work like
waiting for inode to sync or throwing away pagecache and assoc buffers.
UDF probably is just doing more work than most and also is more likely
to be used on desktop and noticed than like ocfs2 or something like
that doing a bit of work there.
OK, if you like the patch then I've fixed up the changelog and comment
(yes it was helpful). Maybe it could sit in -mm for a while?
Thanks,
--
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@xxxxxxx>
---
fs/inode.c | 19 ++++++++++++-------
1 file changed, 12 insertions(+), 7 deletions(-)
Index: linux-2.6/fs/inode.c
===================================================================
--- linux-2.6.orig/fs/inode.c
+++ linux-2.6/fs/inode.c
@@ -14,6 +14,7 @@
#include <linux/module.h>
#include <linux/backing-dev.h>
#include <linux/wait.h>
+#include <linux/rwsem.h>
#include <linux/hash.h>
#include <linux/swap.h>
#include <linux/security.h>
@@ -87,14 +88,18 @@ static struct hlist_head *inode_hashtabl
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..
@@ -375,7 +380,7 @@ static int invalidate_list(struct list_h
/*
* 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);
@@ -414,7 +419,7 @@ int invalidate_inodes(struct super_block
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);
@@ -422,7 +427,7 @@ int invalidate_inodes(struct super_block
spin_unlock(&inode_lock);
dispose_list(&throw_away);
- mutex_unlock(&iprune_mutex);
+ up_write(&iprune_sem);
return busy;
}
@@ -461,7 +466,7 @@ 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;
@@ -503,7 +508,7 @@ static void prune_icache(int nr_to_scan)
spin_unlock(&inode_lock);
dispose_list(&freeable);
- mutex_unlock(&iprune_mutex);
+ up_read(&iprune_sem);
}
/*
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
