On 05.08.22 16:09, Arnd Bergmann wrote:
> On Thu, Aug 4, 2022 at 12:07 PM Li Chen <me@linux.beauty> wrote:
>
>> Apart from our cases, I heard there are some other cases where cma_alloc
>> failed even non-cma system memory has enough memory because pages in
>> CMA memory are pinned and cannot move out of CMA. There are some fixes like
>> 1. move these memory out of CMA before pinned
>> 2. only allow non-long-time pinned memory allocation from CMA.
>>
>> But these two solutions are not merged into the mainline yet.
>
> Right, I think this has come up before, not sure why it wasn't implemented.
> My feeling is that 2. cannot work because you don't know if memory will be
> pinned in the future at the time of allocation, but 1. should be doable.
We disallow longterm pinning of CMA memory already and migrate it out of
the CMA region. If migration fails, we reject pinning.
See
9a4e9f3b2d73 ("mm: update get_user_pages_longterm to migrate pages
allocated from CMA region")
and recent
1c563432588d ("mm: fix is_pinnable_page against a cma page")
It's worth nothing that is_pinnable_page() will be renamed to
is_longterm_pinnable_page() soon to express what it actually means.
Note that some FOLL_GET users (vmsplice, O_DIRECT) still have to be
converted to FOLL_PIN, and especially also set FOLL_LONGTERM (vmsplice).
--
Thanks,
David / dhildenb
