Hi Sebastian, David,

Sorry for the long delay in getting back to this, and thanks a lot for
the review.


On 7/9/26 21:15, Sebastian Andrzej Siewior wrote:
> On 2026-07-09 12:04:33 [+0200], David Hildenbrand (Arm) wrote:
>> On 7/7/26 14:59, Wandun Chen wrote:
>>> From: Wandun Chen <[email protected]>
>>>
>>> The region covered by mlock[all] may contain CMA pages. cma_alloc installs
>>
>> What about ZONE_MOVABLE where memory is supposed to be migratable?
> 
> Would it be bad if the pages would not be movable anymore? Does this
> effect just memory-hotplug or something else, too?

ZONE_MOVABLE does not need the same treatment as CMA.
1. On RT compact_unevictable_allowed is 0 by default, pages in ZONE_MOVABLE are
   never migrated by compaction, so no latency spike, no problem.

2. For the scenario of memory offlining within ZONE_MOVABLE, because of offline
   is an explicit administrator action. Any jitter/latency is expected and
   acceptable, the operator chose to offline memory and would not do so
   during an RT-critical phase.

3. cma_alloc() runs at runtime, triggered by drivers as part of normal 
operation,
   and isolation path hardcodes ISOLATE_UNEVICTABLE, so it migrates the mlocked
   pages and installs migration entries. The fault-and-wait latency hits the RT
   task unpredictably during normal operation, which is not expected, and is
   exactly the spike this patch targets.

> 
>> Also, what about drivers that mmap() CMA memory to user space, and
>> __mm_populate()->populate_vma_page_range() would actually try mlocking them, 
>> and
>> they actually must remain on CMA areas?
> 
> It should be safe to skip those. They belong to device and they
> shouldn't be affected by anything including getting swapped out.

Agreed. Only LRU folios are ever isolated for migration here
(isolate_folio_to_list() -> folio_isolate_lru()), and device-owned CMA
buffers are not on the LRU, so they are left in place. I'll double-check
the mmap paths and make sure this holds, and state it explicitly in v3.


Best regards,
Wandun

> 
> Sebastian


Reply via email to