Re: [PATCH RFC 0/2] dma-pool: allow user to disable atomic pool

[Tom Lendacky]

On 8/5/21 1:54 AM, Baoquan He wrote:
> On 06/24/21 at 11:47am, Robin Murphy wrote:
>> On 2021-06-24 10:29, Baoquan He wrote:
>>> On 06/24/21 at 08:40am, Christoph Hellwig wrote:
>>>> So reduce the amount allocated.  But the pool is needed for proper
>>>> operation on systems with memory encryption.  And please add the right
>>>> maintainer or at least mailing list for the code you're touching next
>>>> time.
>>>
>>> Oh, I thoutht it's memory issue only, should have run
>>> ./scripts/get_maintainer.pl. sorry.
>>>
>>> About reducing the amount allocated, it may not help. Because on x86_64,
>>> kdump kernel doesn't put any page of memory into buddy allocator of DMA
>>> zone. Means it will defenitely OOM for atomic_pool_dma initialization.
>>>
>>> Wondering in which case or on which device the atomic pool is needed on
>>> AMD system with mem encrytion enabled. As we can see, the OOM will
>>> happen too in kdump kernel on Intel system, even though it's not
>>> necessary.
> 
> Sorry for very late response, and thank both for your comments.
> 
>>
>> Hmm, I think the Kconfig reshuffle has actually left a slight wrinkle here.
>> For DMA_DIRECT_REMAP=y we can assume an atomic pool is always needed, since
>> that was the original behaviour anyway. However the implications of
>> AMD_MEM_ENCRYPT=y are different - even if support is enabled, it still
>> should only be relevant if mem_encrypt_active(), so it probably does make
>> sense to have an additional runtime gate on that.
> 
>>
>> From a quick scan, use of dma_alloc_from_pool() already depends on
>> force_dma_unencrypted() so that's probably fine already, but I think we'd
>> need a bit of extra protection around dma_free_from_pool() to prevent
>> gen_pool_has_addr() dereferencing NULL if the pools are uninitialised, even
>> with your proposed patch as it is. Presumably nothing actually called
>> dma_direct_free() when you tested this?
> 
> Yes, enforcing the conditional check of force_dma_unencrypted() around
> dma_free_from_pool sounds reasonable, just as we have done in
> dma_alloc_from_pool().
> 
> I have tested this patchset on normal x86_64 systems and one amd system
> with SME support, disabling atomic pool can fix the issue that there's no
> managed pages in dma zone then requesting page from dma zone will cause
> allocation failure. And even disabling atomic pool in 1st kernel didn't
> cause any problem on one AMD EPYC system which supports SME. I am not
> expert of DMA area, wondering how atomic pool is supposed to do in
> SME/SEV system. 

I think the atomic pool is used by the NVMe driver. My understanding is
that driver will do a dma_alloc_coherent() from interrupt context, so it
needs to use GFP_ATOMIC. The pool was created because dma_alloc_coherent()
would perform a set_memory_decrypted() call, which can sleep. The pool
eliminates that issue (David can correct me if I got that wrong).

Thanks,
Tom

> 
> Besides, even though atomic pool is disabled, slub page for allocation
> of dma-kmalloc also triggers page allocation failure. So I change to
> take another way to fix them, please check v2 post. The atomic pool
> disabling an be a good to have change.
> 
> Thanks
> Baoquan
>