Re: [RFC 2/2] x86_64: expand kernel stack to 16K

[Dave Chinner <david@fromorbit.com>]

On Wed, May 28, 2014 at 12:06:58PM -0400, Johannes Weiner wrote:
> On Wed, May 28, 2014 at 07:13:45PM +1000, Dave Chinner wrote:
> > On Wed, May 28, 2014 at 06:37:38PM +1000, Dave Chinner wrote:
> > > [ cc XFS list ]
> > 
> > [and now there is a complete copy on the XFs list, I'll add my 2c]
> > 
> > > On Wed, May 28, 2014 at 03:53:59PM +0900, Minchan Kim wrote:
> > > > While I play inhouse patches with much memory pressure on qemu-kvm,
> > > > 3.14 kernel was randomly crashed. The reason was kernel stack overflow.
> > > > 
> > > > When I investigated the problem, the callstack was a little bit deeper
> > > > by involve with reclaim functions but not direct reclaim path.
> > > > 
> > > > I tried to diet stack size of some functions related with alloc/reclaim
> > > > so did a hundred of byte but overflow was't disappeard so that I encounter
> > > > overflow by another deeper callstack on reclaim/allocator path.
> > 
> > That's a no win situation. The stack overruns through ->writepage
> > we've been seeing with XFS over the past *4 years* are much larger
> > than a few bytes. The worst case stack usage on a virtio block
> > device was about 10.5KB of stack usage.
> > 
> > And, like this one, it came from the flusher thread as well. The
> > difference was that the allocation that triggered the reclaim path
> > you've reported occurred when 5k of the stack had already been
> > used...
> > 
> > > > Of course, we might sweep every sites we have found for reducing
> > > > stack usage but I'm not sure how long it saves the world(surely,
> > > > lots of developer start to add nice features which will use stack
> > > > agains) and if we consider another more complex feature in I/O layer
> > > > and/or reclaim path, it might be better to increase stack size(
> > > > meanwhile, stack usage on 64bit machine was doubled compared to 32bit
> > > > while it have sticked to 8K. Hmm, it's not a fair to me and arm64
> > > > already expaned to 16K. )
> > 
> > Yup, that's all been pointed out previously. 8k stacks were never
> > large enough to fit the linux IO architecture on x86-64, but nobody
> > outside filesystem and IO developers has been willing to accept that
> > argument as valid, despite regular stack overruns and filesystem
> > having to add workaround after workaround to prevent stack overruns.
> > 
> > That's why stuff like this appears in various filesystem's
> > ->writepage:
> > 
> >         /*
> >          * Refuse to write the page out if we are called from reclaim context.
> >          *
> >          * This avoids stack overflows when called from deeply used stacks in
> >          * random callers for direct reclaim or memcg reclaim.  We explicitly
> >          * allow reclaim from kswapd as the stack usage there is relatively low.
> >          *
> >          * This should never happen except in the case of a VM regression so
> >          * warn about it.
> >          */
> >         if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
> >                         PF_MEMALLOC))
> >                 goto redirty;
> > 
> > That still doesn't guarantee us enough stack space to do writeback,
> > though, because memory allocation can occur when reading in metadata
> > needed to do delayed allocation, and so we could trigger GFP_NOFS
> > memory allocation from the flusher thread with 4-5k of stack already
> > consumed, so that would still overrun teh stack.
> > 
> > So, a couple of years ago we started defering half the writeback
> > stack usage to a worker thread (commit c999a22 "xfs: introduce an
> > allocation workqueue"), under the assumption that the worst stack
> > usage when we call memory allocation is around 3-3.5k of stack used.
> > We thought that would be safe, but the stack trace you've posted
> > shows that alloc_page(GFP_NOFS) can consume upwards of 5k of stack,
> > which means we're still screwed despite all the workarounds we have
> > in place.
> 
> The allocation and reclaim stack itself is only 2k per the stacktrace
> below.  What got us in this particular case is that we engaged a
> complicated block layer setup from within the allocation context in
> order to swap out a page.

The report does not have a complicated block layer setup - it's just
a swap device on a virtio device. There's no MD, no raid, no complex
transport and protocol layer, etc. It's about as simple as it gets.

> In the past we disabled filesystem ->writepage from within the
> allocation context and deferred it to kswapd for stack reasons (see
> the WARN_ON_ONCE and the comment in your above quote), but I think we
> have to go further and do the same for even swap_writepage():

I don't think that solves the problem. I've seen plenty of near
stack overflows that were caused by >3k of stack being used because
of memory allocation/reclaim overhead and then scheduling.
usage and another 1k of stack scheduling waiting.

If we have a subsystem that can put >3k on the stack at arbitrary
locations, then we really only have <5k of stack available for
callers. And when the generic code typically consumes 1-2k of stack
before we get to filesystem specific methods, we only have 3-4k of
stack left for the worst case storage path stack usage. With the
block layer and driver layers requiring 2.5-3k because they can do
memory allocation and schedule, that leaves very little for the
layers in the middle, which is arguably the most algorithmically
complex layer of the storage stack.....

> > > > I guess this topic was discussed several time so there might be
> > > > strong reason not to increase kernel stack size on x86_64, for me not
> > > > knowing so Ccing x86_64 maintainers, other MM guys and virtio
> > > > maintainers.
> > > >
> > > >          Depth    Size   Location    (51 entries)
> > > > 
> > > >    0)     7696      16   lookup_address+0x28/0x30
> > > >    1)     7680      16   _lookup_address_cpa.isra.3+0x3b/0x40
> > > >    2)     7664      24   __change_page_attr_set_clr+0xe0/0xb50
> > > >    3)     7640     392   kernel_map_pages+0x6c/0x120
> > > >    4)     7248     256   get_page_from_freelist+0x489/0x920
> > > >    5)     6992     352   __alloc_pages_nodemask+0x5e1/0xb20
> > > >    6)     6640       8   alloc_pages_current+0x10f/0x1f0
> > > >    7)     6632     168   new_slab+0x2c5/0x370
> > > >    8)     6464       8   __slab_alloc+0x3a9/0x501
> > > >    9)     6456      80   __kmalloc+0x1cb/0x200
> > > >   10)     6376     376   vring_add_indirect+0x36/0x200
> > > >   11)     6000     144   virtqueue_add_sgs+0x2e2/0x320
> > > >   12)     5856     288   __virtblk_add_req+0xda/0x1b0
> > > >   13)     5568      96   virtio_queue_rq+0xd3/0x1d0
> > > >   14)     5472     128   __blk_mq_run_hw_queue+0x1ef/0x440
> > > >   15)     5344      16   blk_mq_run_hw_queue+0x35/0x40
> > > >   16)     5328      96   blk_mq_insert_requests+0xdb/0x160
> > > >   17)     5232     112   blk_mq_flush_plug_list+0x12b/0x140
> > > >   18)     5120     112   blk_flush_plug_list+0xc7/0x220
> > > >   19)     5008      64   io_schedule_timeout+0x88/0x100
> > > >   20)     4944     128   mempool_alloc+0x145/0x170
> > > >   21)     4816      96   bio_alloc_bioset+0x10b/0x1d0
> > > >   22)     4720      48   get_swap_bio+0x30/0x90
> > > >   23)     4672     160   __swap_writepage+0x150/0x230
> > > >   24)     4512      32   swap_writepage+0x42/0x90
> 
> Without swap IO from the allocation context, the stack would have
> ended here, which would have been easily survivable.  And left the
> writeout work to kswapd, which has a much shallower stack than this:

Sure, but this is just playing whack-a-stack. We can keep slapping
band-aids and restrictions on code and make the code more complex,
constrainted, convouted and slower, or we can just increase the
stack size....

> > > >   25)     4480     320   shrink_page_list+0x676/0xa80
> > > >   26)     4160     208   shrink_inactive_list+0x262/0x4e0
> > > >   27)     3952     304   shrink_lruvec+0x3e1/0x6a0
> > > >   28)     3648      80   shrink_zone+0x3f/0x110
> > > >   29)     3568     128   do_try_to_free_pages+0x156/0x4c0
> > > >   30)     3440     208   try_to_free_pages+0xf7/0x1e0
> > > >   31)     3232     352   __alloc_pages_nodemask+0x783/0xb20
> > > >   32)     2880       8   alloc_pages_current+0x10f/0x1f0
> > > >   33)     2872     200   __page_cache_alloc+0x13f/0x160
> > > >   34)     2672      80   find_or_create_page+0x4c/0xb0
> > > >   35)     2592      80   ext4_mb_load_buddy+0x1e9/0x370
> > > >   36)     2512     176   ext4_mb_regular_allocator+0x1b7/0x460
> > > >   37)     2336     128   ext4_mb_new_blocks+0x458/0x5f0
> > > >   38)     2208     256   ext4_ext_map_blocks+0x70b/0x1010
> > > >   39)     1952     160   ext4_map_blocks+0x325/0x530
> > > >   40)     1792     384   ext4_writepages+0x6d1/0xce0
> > > >   41)     1408      16   do_writepages+0x23/0x40
> > > >   42)     1392      96   __writeback_single_inode+0x45/0x2e0
> > > >   43)     1296     176   writeback_sb_inodes+0x2ad/0x500
> > > >   44)     1120      80   __writeback_inodes_wb+0x9e/0xd0
> > > >   45)     1040     160   wb_writeback+0x29b/0x350
> > > >   46)      880     208   bdi_writeback_workfn+0x11c/0x480
> > > >   47)      672     144   process_one_work+0x1d2/0x570
> > > >   48)      528     112   worker_thread+0x116/0x370
> > > >   49)      416     240   kthread+0xf3/0x110
> > > >   50)      176     176   ret_from_fork+0x7c/0xb0
> > 
> > Impressive: 3 nested allocations - GFP_NOFS, GFP_NOIO and then
> > GFP_ATOMIC before the stack goes boom. XFS usually only needs 2...
> 
> Do they also usually involve swap_writepage()?

No.  Have a look at this recent thread when Dave Jones reported
trinity was busting the stack.

http://oss.sgi.com/archives/xfs/2014-02/msg00325.html

What happens when a shrinker issues IO:

http://oss.sgi.com/archives/xfs/2014-02/msg00361.html

Yes, there was an XFS problem in there that was fixed (by moving
work to a workqueue!) but the point is that swap is not the only
path through memory allocation that can consume huge amounts of
stack. That above trace also points out a path through the scheduler
of close to 1k of stack usage. That gets worse -
wait_for_completion() typically requires 1.5k of stack....

Contributing is the new blk-mq layer, which from the above stack
trace still hasn't been fixed:

http://oss.sgi.com/archives/xfs/2014-02/msg00355.html

and a lot of the stack usage is because of saved registers on each
function call:

http://oss.sgi.com/archives/xfs/2014-02/msg00470.html

And here's a good set of examples of the amount of stack certain
functions can require:

http://oss.sgi.com/archives/xfs/2014-02/msg00365.html

Am I the only person who sees a widespread problem here?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs