>
> On Wed 18-11-15 22:36:51, Jan Kara wrote:
> > On Tue 17-11-15 13:47:50, Namjae Jeon wrote:
> > > > On Mon 09-11-15 14:21:11, Namjae Jeon wrote:
> > > > > >
> > > > > > On Wed, Oct 21, 2015 at 06:44:10PM +0900, Namjae Jeon wrote:
> > > > > > > > Interestingly we're not seeing these memory leaks on the truncate
> > > > > > > > path, so I suspect the issue is in how collapse range is clearing
> > > > > > > > pages from the page cache, especially pages that were freshly written
> > > > > > > > to the journal by the commit but which hadn't yet been writtten to
> > > > > > > > disk and then marked as complete so we can allow the relevant
> > > > > > > > transaction to be checkpointed. (Although we're not leaking the
> > > > > > > > journal head structures, but only the buffer heads, so the story most
> > > > > > > > be a bit more complicated than that.)
> > > > > > >
> > > > > > > Okay, Thanks for sharing your view and points !!
> > > > > > >
> > > > > > > Currently I can reproduce memory leak issue without collase/insert/zero range.
> > > > > > > conditions like the following.(collase/insert/zero range are disable with -I -C -z option
> and
> > > > add -y
> > > > > > option instead of -W)
> > > > > > > 1. small size parition(1GB)
> > > > > > > 2. run fsx with these options "./fsx -N 30000 -o 128000 -l 500000 -r 4096 -t 512 -w 512
> -Z -R
> > > > -y -
> > > > > > I -C -z testfile"
> > > > > > > And same result with generic/091 is showing (buffer_head leak)
> > > > > > >
> > > > > > > So I am starting to find root-cause base on your points.
> > > > > > > I will share the result or the patch.
> > > > > >
> > > > > > Thanks, that's very interesting data point. So this makes it appear
> > > > > > that the problem *is* probably with how we deal with checkpointing
> > > > > > buffers after the pages get discarded using either a truncate or a
> > > > > > collapse_range, since the 'y' option causes a lot fsync's, and hence
> > > > > > commits, some of which are happening after a truncate command.
> > > > > >
> > > > > > Thanks for a taking a look at this. I really appreciate it.
> > > > > >
> > > > > > Cheers,
> > > > >
> > > > >
> > > > > Hi Ted,
> > > > >
> > > > > Could you review this patch?
> > > > >
> > > > > Thanks!
> > > > > ---------------------------------------------------------------------------------
> > > > > Subject: [PATCH] jbd2: try to free buffers from truncated page after
> > > > > checkpoint
> > > > >
> > > > > when ext4 is mounted in data=journal mode, and truncate operation
> > > > > such as settatr(size), collopse, insert and zero range are used, there are
> > > > > are many truncated pages with NULL page->mapping. Such truncated pages
> > > > > pile up quickly due to truncate_pagecache on data pages associated with journal.
> > > > > As page->mapping is NULL for such truncated pages, they are not freed
> > > > > by drop cache(3) or umount. As a result, MemFree in /proc/meminfo decreases
> > > > > quickly and active buffer_head slab objects grow in /proc/slabinfo.
> > > > > This patch attempts to free buffers from such pages at the end of jbd2
> > > > > checkpoint, if pages do not have any busy buffers and NULL mapping.
> > > >
> > > > Hum, why such pages didn't get freed by release_buffer_page() call
> > > > happening when processing transaction's forget list? Because the idea is
> > > > that such pages should be discarded at that point...
> > > Hi Jan,
> > >
> > > When I checked this code, release_buffer_page is not called
> > > when buffer_jbddirty is true. Such buffers with JBD_Dirty are added
> > > to new checkpoint.
> > >
> > > if (buffer_jbddirty(bh)) {
> > > JBUFFER_TRACE(jh, "add to new checkpointing trans");
> > > __jbd2_journal_insert_checkpoint(jh, commit_transaction);
> > > if (is_journal_aborted(journal))
> > > clear_buffer_jbddirty(bh);
> > > } else {
> > > J_ASSERT_BH(bh, !buffer_dirty(bh));
> > > /*
> > > * The buffer on BJ_Forget list and not jbddirty means
> > > * it has been freed by this transaction and hence it
> > > * could not have been reallocated until this
> > > * transaction has committed. *BUT* it could be
> > > * reallocated once we have written all the data to
> > > * disk and before we process the buffer on BJ_Forget
> > > * list.
> > > */
> > > if (!jh->b_next_transaction)
> > > try_to_free = 1;
> > > }
> > > JBUFFER_TRACE(jh, "refile or unfile buffer");
> > > __jbd2_journal_refile_buffer(jh);
> > > jbd_unlock_bh_state(bh);
> > >
> > > Next buffer was unfiled by __jbd2_journal_refile_buffer, JBD_Dirty cleared and
> > > BH_Dirty was set in same function. Later it must have been written back as
> > > BH_Dirty was cleared.
> > > And ext4_wait_for_tail_page_commit-> __ext4_journalled_invalidatepage ->
> > > journal_unmap_buffer zaps the buffer:
> > >
> > > if (!buffer_dirty(bh)) {
> > > /* bdflush has written it. We can drop it now */
> > > goto zap_buffer;
> > > }
> > >
> > > next, truncate_pagecache is called, which clears the page mapping.
> > > Eventually, remove checkpoint is called, but such page with NULL mapping was
> > > not freed. So, I had added release_buffer_page at the end of remove checkpoint
> > > to attempt to free such free buffer pages. Please let me know your opinion.
> >
> > OK, thanks for the detailed analysis. But when the buffer gets truncated,
> > jbd2_journal_invalidatepage() either removes the buffer from the
> > transaction (obviously didn't happen here) or it sets buffer_freed and
> > buffer_jbddirty should get cleared when processing the BJ_Forget list. So
> > why that didn't happen? Can you have a look into what
> > jbd2_journal_invalidatepage() did to buffer in that page?
> >
> > Generally truncated buffers should not enter checkpoint list since writing
> > them is just a waste of disk bandwidth...
>
> I thought a bit more about this and my guess is that
> jbd2_journal_invalidatepage() fails to invalidate the partial page and
> returns EBUSY. However we should see warnings about that from
> ext4_journalled_invalidatepage(). Can you see them? This would perfectly
> match Ted's observation that the problem happens only for fallocate
> operations but not for truncate which does the dance with
> ext4_wait_for_tail_page_commit() but fallocate operations don't do it...
>
> Now if I'm right and this is indeed the path we are hitting, we need to put
> more thought into how we handle truncation of partial pages in data=journal
> mode.
Actually, I can quickly create memleak with simple testcase like:
while(1000)
{
Random seek(max 500M) -> write 1M -> fsync ->truncate(random size)
}
As there is significant memleak in this situation also, I started focusing
on this simple testcase only. When use this testcase, I never saw EBUSY path
in jbd2_journal_invalidatepage() -> journal_unmap_buffer().
On the other hand(), for memleak pages jbd2_journal_invalidatepage() ->
journal_unmap_buffer() was almost bailing out every time from
if (!buffer_dirty(bh)) {
/* bdflush has written it. We can drop it now */
goto zap_buffer;
}
As per debugging, NULL mapping page or buffer is created in below scenario:
Write(or Kjournald) -> jbd2_journal_commit_transaction
-> BH_JBDDirty buffer is added to forget list.
-> Buffer is added to new checkpoint,
additional reference count is taken on b_jcount which
keeps buffer busy until remove checkpoint.
-> Buffer is unfiled, removed from forget list, BH_JBDDirty
is cleared and BH_Dirty is set, it is exposed to VM.
As BH_Dirty is set, buffer is written by ext4_writepage by flusher.
ext4_writepage finally clears BH_Dirty.
Now, ftruncate is called, almost every time code is returned from below path:
Ftruncate -> jbd2_journal_invalidatepage -> journal_unmap_buffer
if (!buffer_dirty(bh)) {
/* bdflush has written it. We can drop it now */
goto zap_buffer;
}
page was not released by jbd2_journal_invalidatepage, mainly because b_jcount
was 2 on entry of jbd2_journal_invalidatepage (addional reference by insert
checkpoint above), buffer state was sane.
At exit time jbd2_journal_invalidatepage, b_jcount is still 1.
Next, truncate_pagecache is called which makes page mapping NULL.
Later, __jbd2_journal_remove_checkpoint is called, it makes b_jcount zero and
remove BH_JBD from buffer. Such pages with NULL mapping pages are not freed.
Previous patch fixed this leak if try release_page_buffer at the end of
checkpoint.
>
> BTW: The ext4_force_commit() call in fallocate operations is definitely
> racy as pages can get dirtied and journalled in the running transaction
> anytime while we wait for transaction commit (which I suspect is happening
> in your testcase)... So that needs some more thought as well.
Need to check more on this. I can reproduce memleak without fsync
(i.e. ext4_force_commit), though it is less.
Thanks!
>
> Honza
> --
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
