[PATCH 1/2] mm/cma: provide option to opt out from exposing pages on activation failure

[David Hildenbrand]

On 06.01.22 13:01, Hari Bathini wrote:
> 
> 
> On 22/12/21 12:18 am, David Hildenbrand wrote:
>> On 20.12.21 20:34, Hari Bathini wrote:
>>> Commit 072355c1cf2d ("mm/cma: expose all pages to the buddy if
>>> activation of an area fails") started exposing all pages to buddy
>>> allocator on CMA activation failure. But there can be CMA users that
>>> want to handle the reserved memory differently on CMA allocation
>>> failure. Provide an option to opt out from exposing pages to buddy
>>> for such cases.
> 
> Hi David,
> 
> Sorry, I could not get back on this sooner. I went out on vacation
> and missed this.
> .
> 
>>
>> Can you elaborate why that is important and what the target user can
>> actually do with it?
> Previously, firmware-assisted dump [1] used to reserve memory that it 
> needs for booting a capture kernel & offloading /proc/vmcore.
> This memory is reserved, basically blocked from being used by
> production kernel, to ensure kernel crash context is not lost on
> booting into a capture kernel from this memory chunk.
> 
> But [2] started using CMA instead to let the memory be used at least
> in some cases as long as this memory is not going to have kernel pages. 
> So, the intention in using CMA was to keep the memory unused if CMA
> activation fails and only let it be used for some purpose, if at all,
> if CMA activation succeeds. But [3] breaks that assumption reporting
> weird errors on vmcore captured with fadump, when CMA activation fails.
> 
> To answer the question, fadump does not want the memory to be used for
> kernel pages, if CMA activation fails...

Okay, so what you want is a reserved region, and if possible, let CMA
use that memory for other (movable allocation) purposes until you
actually need that area and free it up by using CMA. If CMA cannot use
the region because of zone issues, you just want that region to stay
reserved.

I guess the biggest different to other CMA users is that it can make use
of the memory even if not allocated via CMA -- because it's going to
make use of the the physical memory range indirectly via a HW facility,
not via any "struct page" access.


I wonder if we can make the terminology a bit clearer, the freeing part
is a bit confusing, because init_cma_reserved_pageblock() essentially
also frees pages, just to the MIGRATE_CMA lists ... what you want is to
treat it like a simple memblock allocation/reservation on error.

What about:
* cma->reserve_pages_on_error that defaults to false
* void __init cma_reserve_pages_on_error(struct cma *cma)


-- 
Thanks,

David / dhildenb