Re: [RFC PATCH] fs: ext4: don't trap kswapd and allocating tasks on ext4 inode IO

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue 16-05-17 11:41:05, Johannes Weiner wrote:
> On Tue, May 16, 2017 at 04:36:45PM +0200, Jan Kara wrote:
> > On Mon 15-05-17 11:46:34, Johannes Weiner wrote:
> > > We have observed across several workloads situations where kswapd and
> > > direct reclaimers get stuck in the inode shrinker of the ext4 / mount,
> > > causing allocation latencies across tasks in the system, while there
> > > are dozens of gigabytes of clean page cache covering multiple disks.
> > > 
> > > The stack traces of such an instance looks like this:
> > > 
> > > [<ffffffff812b3225>] jbd2_log_wait_commit+0x95/0x110
> > > [<ffffffff812b4f29>] jbd2_complete_transaction+0x59/0x90
> > > [<ffffffff812668da>] ext4_evict_inode+0x2da/0x480
> > > [<ffffffff811f2230>] evict+0xc0/0x190
> > > [<ffffffff811f2339>] dispose_list+0x39/0x50
> > > [<ffffffff811f323b>] prune_icache_sb+0x4b/0x60
> > > [<ffffffff811dba71>] super_cache_scan+0x141/0x190
> > > [<ffffffff8116e755>] shrink_slab+0x235/0x440
> > > [<ffffffff81172b48>] shrink_zone+0x268/0x2d0
> > > [<ffffffff81172f04>] do_try_to_free_pages+0x164/0x410
> > > [<ffffffff81173265>] try_to_free_pages+0xb5/0x160
> > > [<ffffffff811656b6>] __alloc_pages_nodemask+0x636/0xb30
> > > [<ffffffff811acac8>] alloc_pages_current+0x88/0x120
> > > [<ffffffff816d4e46>] skb_page_frag_refill+0xc6/0xf0
> > > [<ffffffff816d4e8d>] sk_page_frag_refill+0x1d/0x80
> > > [<ffffffff8173f86b>] tcp_sendmsg+0x28b/0xb10
> > > [<ffffffff81769727>] inet_sendmsg+0x67/0xa0
> > > [<ffffffff816d0488>] sock_sendmsg+0x38/0x50
> > > [<ffffffff816d0518>] sock_write_iter+0x78/0xd0
> > > [<ffffffff811d774e>] do_iter_readv_writev+0x5e/0xa0
> > > [<ffffffff811d8468>] do_readv_writev+0x178/0x210
> > > [<ffffffff811d871c>] vfs_writev+0x3c/0x50
> > > [<ffffffff811d8782>] do_writev+0x52/0xd0
> > > [<ffffffff811d9810>] SyS_writev+0x10/0x20
> > > [<ffffffff81002910>] do_syscall_64+0x50/0xa0
> > > [<ffffffff817eed3c>] return_from_SYSCALL_64+0x0/0x6a
> > > [<ffffffffffffffff>] 0xffffffffffffffff
> > > 
> > > The inode shrinker has provisions to skip any inodes that require
> > > writeback, to avoid tarpitting the entire system behind a single
> > > object when there are many other pools to recycle memory from. But
> > > that logic doesn't cover the situation where an ext4 inode is clean
> > > but journaled and tied to a commit that yet needs to hit the platter.
> > > 
> > > Add a superblock operation that lets the generic inode shrinker query
> > > the filesystem whether evicting a given inode will require any IO; add
> > > an ext4 implementation that checks whether the journal is caught up to
> > > the commit id associated with the inode.
> > > 
> > > Fixes: 2d859db3e4a8 ("ext4: fix data corruption in inodes with journalled data")
> > > Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
> > 
> > OK. I have to say I'm somewhat surprised you use data journalling on some
> > of your files / filesystems but whatever - maybe these are long symlink
> > after all which would make sense.
> 
> The filesystem is actually mounted data=ordered and we didn't catch
> anyone in userspace enabling journaling on individual inodes. So we
> assumed this must be from symlinks.

OK.

> > And I'm actually doubly surprised you can see these stack traces as
> > these days inode_lru_isolate() checks inode->i_data.nrpages and
> > uncommitted pages cannot be evicted from pagecache
> > (ext4_releasepage() will refuse to free them) so I don't see how
> > such inode can get to dispose_list(). But maybe the inode doesn't
> > really have any pages and i_datasync_tid just happens to be set to
> > the current transaction because it is initialized that way and we
> > are evicting inode that was recently read from disk.
> 
> Hm, we're running 4.6, but that already has the nrpages check in
> inode_lru_isolate(). There couldn't be any pages in those inodes by
> the time the shrinker gets to them.
> 
> > Anyway if you add: "&& inode->i_data.nrpages" to the test in
> > ext4_evict_inode() do the stalls go away?
> 
> Want me to still test this?

Can you try attached patch? I'd like to confirm the theory before merging
this... Thanks!

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
>From e87281dee65589e07b9251ad98191c1e6c488870 Mon Sep 17 00:00:00 2001
From: Jan Kara <jack@xxxxxxx>
Date: Tue, 16 May 2017 17:56:36 +0200
Subject: [PATCH] ext4: Avoid unnecessary stalls in ext4_evict_inode()

These days inode reclaim calls evict_inode() only when it has no pages
in the mapping. In that case it is not necessary to wait for transaction
commit in ext4_evict_inode() as there can be no pages waiting to be
committed. So avoid unnecessary transaction waiting in that case.

We still have to keep the check for the case where ext4_evict_inode()
gets called from other paths (e.g. umount) where inode still can have
some page cache pages.

Reported-by: Johannes Weiner <hannes@xxxxxxxxxxx>
Signed-off-by: Jan Kara <jack@xxxxxxx>
---
 fs/ext4/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 5834c4d76be8..3aef67ca18ac 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -213,7 +213,8 @@ void ext4_evict_inode(struct inode *inode)
 		 */
 		if (inode->i_ino != EXT4_JOURNAL_INO &&
 		    ext4_should_journal_data(inode) &&
-		    (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
+		    (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
+		    inode->i_data.nrpages) {
 			journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
 			tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
 
-- 
2.12.0


[Index of Archives]     [Reiser Filesystem Development]     [Ceph FS]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite National Park]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]     [Linux Media]

  Powered by Linux