Re: [PATCH v2] mm, page_alloc: disallow __GFP_COMP in alloc_pages_exact()

[Takashi Iwai <tiwai@suse.de>]

On Thu, 14 Mar 2019 18:37:06 +0100,
Hugh Dickins wrote:
> 
> On Thu, 14 Mar 2019, Takashi Iwai wrote:
> > On Thu, 14 Mar 2019 14:29:33 +0100,
> > Michal Hocko wrote:
> > > 
> > > On Thu 14-03-19 14:15:38, Takashi Iwai wrote:
> > > > On Thu, 14 Mar 2019 13:09:39 +0100,
> > > > Michal Hocko wrote:
> > > > > 
> > > > > On Thu 14-03-19 12:56:43, Takashi Iwai wrote:
> > > > > > On Thu, 14 Mar 2019 12:36:26 +0100,
> > > > > > Michal Hocko wrote:
> > > > > > > 
> > > > > > > On Thu 14-03-19 11:30:03, Vlastimil Babka wrote:
> > > [...]
> > > > > > > > I initially went with 2 as well, as you can see from v1 :) but then I looked at
> > > > > > > > the commit [2] mentioned in [1] and I think ALSA legitimaly uses __GFP_COMP so
> > > > > > > > that the pages are then mapped to userspace. Breaking that didn't seem good.
> > > > > > > 
> > > > > > > It used the flag legitimately before because they were allocating
> > > > > > > compound pages but now they don't so this is just a conversion bug.
> > > > > > 
> > > > > > We still use __GFP_COMP for allocation of the sound buffers that are
> > > > > > also mmapped to user-space.  The mentioned commit above [2] was
> > > > > > reverted later.
> > > > > 
> > > > > Yes, I understand that part. __GFP_COMP makes sense on a comound page.
> > > > > But if you are using alloc_pages_exact then the flag doesn't make sense
> > > > > because split out should already do what you want. Unless I am missing
> > > > > something.
> > > > 
> > > > The __GFP_COMP was taken as a sort of workaround for the problem wrt
> > > > mmap I already forgot.  If it can be eliminated, it's all good.
> > > 
> > > Without __GFP_COMP you would get tail pages which are not setup properly
> > > AFAIU. With alloc_pages_exact you should get an "array" of head pages
> > > which are properly reference counted. But I might misunderstood the
> > > original problem which __GFP_COMP tried to solve.
> > 
> > I only vaguely remember that it was about a Bad Page error for the
> > reserved pages, but forgot the all details, sorry.
> > 
> > Hugh, could you confirm whether we still need __GFP_COMP in the sound
> > buffer allocations?  FWIW, it's the change introduced by the ancient
> > commit f3d48f0373c1.
> 
> I'm not confident in finding all "the sound buffer allocations".
> Where you're using alloc_pages_exact() for them, you do not need
> __GFP_COMP, and should not pass it.

It was my fault attempt to convert to alloc_pages_exact() and hitting
the incompatibility with __GFP_COMP, so it was reverted in the end.

> But if there are other places
> where you use one of those page allocators with an "order" argument
> non-zero, and map that buffer into userspace (without any split_page()),
> there you would still need the __GFP_COMP - zap_pte_range() and others
> do the wrong thing on tail ptes if the non-zero-order page has neither
> been set up as compound nor split into zero-order pages.

Hm, what if we allocate the whole pages via alloc_pages_exact() (but
without __GFP_COMP)?  Can we mmap them properly to user-space like
before, or it won't work as-is?


thanks,

Takashi