Re: [PATCH] mm: fix oom work when memory is under pressure

[Michal Hocko <mhocko@suse.cz>]

On Fri 16-09-16 15:13:56, Hugh Dickins wrote:
> On Wed, 14 Sep 2016, Michal Hocko wrote:
> > On Wed 14-09-16 10:42:19, Michal Hocko wrote:
> > > [Let's CC Hugh]
> > 
> > now for real...
> > 
> > > 
> > > On Wed 14-09-16 15:13:50, zhong jiang wrote:
> > > [...]
> > > >   hi, Michal
> > > > 
> > > >   Recently, I hit the same issue when run a OOM case of the LTP and ksm enable.
> > > >  
> > > > [  601.937145] Call trace:
> > > > [  601.939600] [<ffffffc000086a88>] __switch_to+0x74/0x8c
> > > > [  601.944760] [<ffffffc000a1bae0>] __schedule+0x23c/0x7bc
> > > > [  601.950007] [<ffffffc000a1c09c>] schedule+0x3c/0x94
> > > > [  601.954907] [<ffffffc000a1eb84>] rwsem_down_write_failed+0x214/0x350
> > > > [  601.961289] [<ffffffc000a1e32c>] down_write+0x64/0x80
> > > > [  601.966363] [<ffffffc00021f794>] __ksm_exit+0x90/0x19c
> > > > [  601.971523] [<ffffffc0000be650>] mmput+0x118/0x11c
> > > > [  601.976335] [<ffffffc0000c3ec4>] do_exit+0x2dc/0xa74
> > > > [  601.981321] [<ffffffc0000c46f8>] do_group_exit+0x4c/0xe4
> > > > [  601.986656] [<ffffffc0000d0f34>] get_signal+0x444/0x5e0
> > > > [  601.991904] [<ffffffc000089fcc>] do_signal+0x1d8/0x450
> > > > [  601.997065] [<ffffffc00008a35c>] do_notify_resume+0x70/0x78
> > > 
> > > So this is a hung task triggering because the exiting task cannot get
> > > the mmap sem for write because the ksmd holds it for read while
> > > allocating memory which just takes ages to complete, right?
> > > 
> > > > 
> > > > The root case is that ksmd hold the read lock. and the lock is not released.
> > > >  scan_get_next_rmap_item
> > > >          down_read
> > > >                    get_next_rmap_item
> > > >                              alloc_rmap_item     #ksmd will loop permanently.
> > > > 
> > > > How do you see this kind of situation ? or  let the issue alone.
> > > 
> > > I am not familiar with the ksmd code so it is hard for me to judge but
> > > one thing to do would be __GFP_NORETRY which would force a bail out from
> > > the allocation rather than looping for ever. A quick look tells me that
> > > the allocation failure here is quite easy to handle. There might be
> > > others...
> 
> Yes, very good suggestion in this case: the ksmd code does exactly the
> right thing when that allocation fails, but was too stupid to use an
> allocation mode which might fail - and it can allocate rather a lot of
> slots along that path, so it will be good to let it break out there.
> 
> Thank you, Zhongjiang, please send akpm a fully signed-off patch, tagged
> for stable, with your explanation above (which was a lot more helpful
> to me than what you wrote in your other mail of Sept 13th).  But please
> make it GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN (and break that line

agreed

> before 80 cols): the allocation will sometimes fail, and we're not at
> all interested in hearing about that.
> 
> Michal, how would you feel about this or a separate patch adding
> __GFP_HIGH to the allocation in ksm's alloc_stable_node()?  That
> allocation could cause the same problem, but it is much less common
> (so less important to do anything about it), and differs from the
> rmap_item case in that if it succeeds, it will usually free a page;
> whereas if it fails, the fallback (two break_cow()s) may want to
> allocate a couple of pages.  So __GFP_HIGH makes more sense for it
> than __GFP_NORETRY: but perhaps we prefer not to add __GFP_HIGHs?

I am not familiar with the ksmd code enough to have a strong opinion
here. __GFP_HIGH should be imho used only when really necessary but as
you point out and comment in cmp_and_merge_page explain
			/*
			 * If we fail to insert the page into the stable tree,
			 * we will have 2 virtual addresses that are pointing
			 * to a ksm page left outside the stable tree,
			 * in which case we need to break_cow on both.
			 */
this can actually save some memory if succeed. So I will leave the
decision to you. I have no experience in how much this path can actually
eat and whether the flag actually makes much difference.

Thanks!
-- 
Michal Hocko
SUSE Labs

--
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>