Re: [PATCH] mm,ksm: fix endless looping in allocating memory when ksm enable

[Hugh Dickins <hughd@google.com>]

On Mon, 19 Sep 2016, Andrew Morton wrote:
> On Sun, 18 Sep 2016 10:26:10 +0800 zhongjiang <zhongjiang@huawei.com> wrote:
> 
> > I hit the following issue when run a OOM case of the LTP and
> > ksm enable.
> > 
> > Call trace:
> > [<ffffffc000086a88>] __switch_to+0x74/0x8c
> > [<ffffffc000a1bae0>] __schedule+0x23c/0x7bc
> > [<ffffffc000a1c09c>] schedule+0x3c/0x94
> > [<ffffffc000a1eb84>] rwsem_down_write_failed+0x214/0x350
> > [<ffffffc000a1e32c>] down_write+0x64/0x80
> > [<ffffffc00021f794>] __ksm_exit+0x90/0x19c
> > [<ffffffc0000be650>] mmput+0x118/0x11c
> > [<ffffffc0000c3ec4>] do_exit+0x2dc/0xa74
> > [<ffffffc0000c46f8>] do_group_exit+0x4c/0xe4
> > [<ffffffc0000d0f34>] get_signal+0x444/0x5e0
> > [<ffffffc000089fcc>] do_signal+0x1d8/0x450
> > [<ffffffc00008a35c>] do_notify_resume+0x70/0x78
> > 
> > it will leads to a hung task because the exiting task cannot get the
> > mmap sem for write. but the root cause is that the ksmd holds it for
> > read while allocateing memory which just takes ages to complete.
> > and ksmd  will loop in the following path.
> > 
> >  scan_get_next_rmap_item
> >           down_read
> >                 get_next_rmap_item
> >                         alloc_rmap_item   #ksmd will loop permanently.
> > 
> > we fix it by changing the GFP to allow the allocation sometimes fail, and
> > we're not at all interested in hearing abot that.
> 
> It would be better if the changelog were to describe *why* this is
> harmless.  I assume that if the allocation fails,
> scan_get_next_rmap_item() will bale out and ksmd just gives up and
> takes a sleep?

Exactly.  (If that sleep time has been configured to 0, so be it.)
Michal asked for the same reassurance, I expect a new version will
be coming.

> 
> Also, did you instead consider changing scan_get_next_rmap_item() to
> simply not hold mmap_sem for so long?  Scan a megabyte or so then drop
> mmap_sem for a while, then scan some more?  The whole thing is driven by
> ksm.scan_address so handling the races should be simple.

It already does that, configurable intervals: the "endless looping in
allocating memory" is not at the ksm.c level, but inside page_alloc.c:
the __GFP_NORETRY being to get it out of there and back to ksm.c,
which then does the right thing on failure.

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>