Re: Linux 5.0-rc1 (test results)

[Michal Hocko]

On Wed 09-01-19 00:16:59, Guo Ren wrote:
> Thx Michal,
> 
> On Tue, Jan 08, 2019 at 04:40:31PM +0100, Michal Hocko wrote:
> > On Tue 08-01-19 17:51:07, Guo Ren wrote:
> > [...]
> > > static inline pte_t *pte_alloc_one_kernel(struct mm_struct *mm)
> > > {
> > > 	pte_t *pte;
> > > 	unsigned long i;
> > > 
> > > 	pte = (pte_t *) __get_free_page(GFP_KERNEL | __GFP_RETRY_MAYFAIL);
> > > 						     ^^^^^^^^^^^^^^^^^^^
> > > 						     It's necessary ?
> > > 						     x86 & arm don't use
> > > 						     it.
> > > 	if (!pte)
> > > 		return NULL;
> > 
> > That depends on whether you want OOM killer to be triggered for these
> > allocations. If you add the flag then the allocation bails out with a
> > failure rather than kill an oom victim.
> 
> Yes, in page_alloc.c: 
> 	if (gfp_mask & __GFP_RETRY_MAYFAIL)
> 		goto out;
> 	...
> 	if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
>             ^^^^^^^^^^^^^ OOM kill victim
> 	...
> 		if (gfp_mask & __GFP_NOFAIL)
> 			page = __alloc_pages_cpuset_fallback(gfp_mask, order,
> 					ALLOC_NO_WATERMARKS, ac);
> 	}

Btw. we have that nice documentation in gfp.h. I would encourage you to
read through that rather than try to imply the semantic from the
implementation.

> Seems it could affect the behavior of the system which is out of memory.
> OOM killer could help to get_page for current. So keep the same as x86 &
> arm here. I'll remove __GFP_RETRY_MAYFAIL in patch.

OK

-- 
Michal Hocko
SUSE Labs