Re: [PATCH 2/3] mm: add find_alloc_contig_pages() interface

[Michal Hocko]

On Wed 02-05-18 14:13:32, Mike Kravetz wrote:
> On 04/21/2018 09:16 AM, Vlastimil Babka wrote:
> > On 04/17/2018 04:09 AM, Mike Kravetz wrote:
> >> find_alloc_contig_pages() is a new interface that attempts to locate
> >> and allocate a contiguous range of pages.  It is provided as a more
> >> convenient interface than alloc_contig_range() which is currently
> >> used by CMA and gigantic huge pages.
> >>
> >> When attempting to allocate a range of pages, migration is employed
> >> if possible.  There is no guarantee that the routine will succeed.
> >> So, the user must be prepared for failure and have a fall back plan.
> >>
> >> Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> > 
> > Hi, just two quick observations, maybe discussion pointers for the
> > LSF/MM session:
> > - it's weird that find_alloc_contig_pages() takes an order, and
> > free_contig_pages() takes a nr_pages. I suspect the interface would be
> > more future-proof with both using nr_pages? Perhaps also minimum
> > alignment for the allocation side? Order is fine for hugetlb, but what
> > about other potential users?
> 
> Agreed, and I am changing this to nr_pages and adding alignment.
> 
> > - contig_alloc_migratetype_ok() says that MIGRATE_CMA blocks are OK to
> > allocate from. This silently assumes that everything allocated by this
> > will be migratable itself, or it might eat CMA reserves. Is it the case?
> > Also you then call alloc_contig_range() with MIGRATE_MOVABLE, so it will
> > skip/fail on MIGRATE_CMA anyway IIRC.
> 
> When looking closer at the code, alloc_contig_range currently has comments
> saying migratetype must be MIGRATE_MOVABLE or MIGRATE_CMA.  However, this
> is not checked/enforced anywhere in the code (that I can see).  The
> migratetype passed to alloc_contig_range() will be used to set the migrate
> type of all pageblocks in the range.  If there is an error, one side effect
> is that some pageblocks may have their migrate type changed to migratetype.
> Depending on how far we got before hitting the error, the number of pageblocks
> changed is unknown.  This actually can happen at the lower level routine
> start_isolate_page_range().
> 
> My first thought was to make start_isolate_page_range/set_migratetype_isolate
> check that the migrate type of a pageblock was migratetype before isolating.
> This would work for CMA, and I could make it work for the new allocator.
> However, offline_pages also calls start_isolate_page_range and I believe we
> do not want to enforce such a rule (all pageblocks must be of the same migrate
> type) for memory hotplug/offline?
> 
> Should we be concerned at all about this potential changing of migrate type
> on error?  The only way I can think to avoid this is to save the original
> migrate type before isolation.

This is more a question to Vlastimil, Joonsoo. But my understanding is
that it doesn't matter. MIGRATE_MOVABLE will not block other
allocations. So we seem to need it only for MIGRATE_CMA. The later
should die sooner or later hopefully so this awful kludge should just
die with it.
-- 
Michal Hocko
SUSE Labs