Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1662021 > unrolled thread

Re: [RFC PATCH 2/4] mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [RFC PATCH 2/4] mm, tree wide: replace __GFP_REPEAT by  __GFP_RETRY_MAYFAIL with more useful semantic Michal Hocko <mhocko@kernel.org> - 2017-06-09 09:40 +0200

#1662021 — Re: [RFC PATCH 2/4] mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic

On Wed 07-06-17 10:10:36, Wei Yang wrote:
[...]
> Hmm... Let me be more specific. With two factors, costly or not, flag set or
> not, we have four combinations. Here it is classified into two categories.
> 
> 1. __GFP_RETRY_MAYFAIL not set
> 
> Brief description on behavior:
>     costly: pick up the shortcut, so no OOM
>     !costly: no shortcut and will OOM I think
> 
> Impact from this patch set:
>     No.

true

> My personal understanding:
>     The allocation without __GFP_RETRY_MAYFAIL is not effected by this patch
>     set.  Since !costly allocation will trigger OOM, this is the reason why
>     "small allocations never fail _practically_", as mentioned in
>     https://lwn.net/Articles/723317/.
> 
> 
> 3. __GFP_RETRY_MAYFAIL set
> 
> Brief description on behavior:
>     costly/!costly: no shortcut here and no OOM invoked
> 
> Impact from this patch set:
>     For those allocations with __GFP_RETRY_MAYFAIL, OOM is not invoked for
>     both.

yes

> My personal understanding:
>     This is the semantic you are willing to introduce in this patch set. By
>     cutting off the OOM invoke when __GFP_RETRY_MAYFAIL is set, you makes this
>     a middle situation between NOFAIL and NORETRY.

yes

>     page_alloc will try some luck to get some free pages without disturb other
>     part of the system. By doing so, the never fail allocation for !costly
>     pages will be "fixed". If I understand correctly, you are willing to make
>     this the default behavior in the future?

I do not think we can make this a default in a foreseeable future
unfortunately. That's why I've made it a gfp modifier in the first
place. I assume many users will opt in by using the flag. In future we
can even help by adding a highlevel GFP_$FOO flag but I am worried that
this would just add to the explosion of existing highlevel gfp masks
(e.g. do we want GFP_NOFS_MAY_FAIL, GFP_USER_MAY_FAIL,
GFP_USER_HIGH_MOVABLE_MAYFAIL etc...)
-- 
Michal Hocko
SUSE Labs

[toc] | [standalone]

Back to top | Article view | linux.kernel

csiph-web