Re: [PATCH v2 2/4] mm/vmalloc: add support for __GFP_NOFAIL

[Michal Hocko]

On Fri 26-11-21 15:50:15, Vlastimil Babka wrote:
> On 11/24/21 06:23, NeilBrown wrote:
> >> 
> >> I forget why radix_tree_preload used a cpu-local store rather than a
> >> per-task one.
> >> 
> >> Plus "what order pages would you like" and "on which node" and "in
> >> which zone", etc...
> > 
> > "what order" - only order-0 I hope.  I'd hazard a guess that 90% of
> > current NOFAIL allocations only need one page (providing slub is used -
> > slab seems to insist on high-order pages sometimes).
> 
> Yeah AFAIK SLUB can prefer higher orders than SLAB, but also allows fallback
> to smallest order that's enough (thus 0 unless the objects are larger than a
> page).
> 
> > "which node" - whichever.  Unless __GFP_HARDWALL is set, alloc_page()
> > will fall-back to "whichever" anyway, and NOFAIL with HARDWALL is
> > probably a poor choice.
> > "which zone" - NORMAL.  I cannot find any NOFAIL allocations that want
> > DMA.  fs/ntfs asks for __GFP_HIGHMEM with NOFAIL, but that that doesn't
> > *requre* highmem.
> > 
> > Of course, before designing this interface too precisely we should check
> > if anyone can use it.  From a quick through the some of the 100-ish
> > users of __GFP_NOFAIL I'd guess that mempools would help - the
> > preallocation should happen at init-time, not request-time.  Maybe if we
> > made mempools even more light weight .... though that risks allocating a
> > lot of memory that will never get used.
> > 
> > This brings me back to the idea that
> >     alloc_page(wait and reclaim allowed)
> > should only fail on OOM_KILL.  That way kernel threads are safe, and
> > user-threads are free to return ENOMEM knowing it won't get to
> 
> Hm I thought that's already pretty much the case of the "too small to fail"
> of today. IIRC there's exactly that gotcha that OOM KILL can result in such
> allocation failure. But I believe that approach is rather fragile. If you
> encounter such an allocation not checking the resulting page != NULL, you
> can only guess which one is true:
> 
> - the author simply forgot to check at all
> - the author relied on "too small to fail" without realizing the gotcha
> - at the time of writing the code was verified that it can be only run in
> kernel thread context, not user and
>   - it is still true
>   - it stopped being true at some later point
>   - might be hard to even decide which is the case
> 
> IIRC at some point we tried to abolish the "too small to fail" rule because
> of this, but Linus denied that. But the opposite - make it hard guarantee in
> all cases - also didn't happen, so...

Yeah. IMHO we should treat each missing check for allocation failure
(except for GFP_NOFAIL) as a bug regardless the practical implementation
that say that small allocations do not fail. Because they can fail and
we should never subscribe to official support implicit non-fail
semantic.
-- 
Michal Hocko
SUSE Labs