Re: [PATCH v5 01/13] mm: add zone device coherent type memory support

["Sierra Guiza, Alejandro (Alex)" <alex.sierra@amd.com>]

On 6/23/2022 1:21 PM, David Hildenbrand wrote:
> On 23.06.22 20:20, Sierra Guiza, Alejandro (Alex) wrote:
>> On 6/23/2022 2:57 AM, David Hildenbrand wrote:
>>> On 23.06.22 01:16, Sierra Guiza, Alejandro (Alex) wrote:
>>>> On 6/21/2022 11:16 AM, David Hildenbrand wrote:
>>>>> On 21.06.22 18:08, Sierra Guiza, Alejandro (Alex) wrote:
>>>>>> On 6/21/2022 7:25 AM, David Hildenbrand wrote:
>>>>>>> On 21.06.22 13:55, Alistair Popple wrote:
>>>>>>>> David Hildenbrand<david@redhat.com>  writes:
>>>>>>>>
>>>>>>>>> On 21.06.22 13:25, Felix Kuehling wrote:
>>>>>>>>>> Am 6/17/22 um 23:19 schrieb David Hildenbrand:
>>>>>>>>>>> On 17.06.22 21:27, Sierra Guiza, Alejandro (Alex) wrote:
>>>>>>>>>>>> On 6/17/2022 12:33 PM, David Hildenbrand wrote:
>>>>>>>>>>>>> On 17.06.22 19:20, Sierra Guiza, Alejandro (Alex) wrote:
>>>>>>>>>>>>>> On 6/17/2022 4:40 AM, David Hildenbrand wrote:
>>>>>>>>>>>>>>> On 31.05.22 22:00, Alex Sierra wrote:
>>>>>>>>>>>>>>>> Device memory that is cache coherent from device and CPU point of view.
>>>>>>>>>>>>>>>> This is used on platforms that have an advanced system bus (like CAPI
>>>>>>>>>>>>>>>> or CXL). Any page of a process can be migrated to such memory. However,
>>>>>>>>>>>>>>>> no one should be allowed to pin such memory so that it can always be
>>>>>>>>>>>>>>>> evicted.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Signed-off-by: Alex Sierra<alex.sierra@amd.com>
>>>>>>>>>>>>>>>> Acked-by: Felix Kuehling<Felix.Kuehling@amd.com>
>>>>>>>>>>>>>>>> Reviewed-by: Alistair Popple<apopple@nvidia.com>
>>>>>>>>>>>>>>>> [hch: rebased ontop of the refcount changes,
>>>>>>>>>>>>>>>>              removed is_dev_private_or_coherent_page]
>>>>>>>>>>>>>>>> Signed-off-by: Christoph Hellwig<hch@lst.de>
>>>>>>>>>>>>>>>> ---
>>>>>>>>>>>>>>>>         include/linux/memremap.h | 19 +++++++++++++++++++
>>>>>>>>>>>>>>>>         mm/memcontrol.c          |  7 ++++---
>>>>>>>>>>>>>>>>         mm/memory-failure.c      |  8 ++++++--
>>>>>>>>>>>>>>>>         mm/memremap.c            | 10 ++++++++++
>>>>>>>>>>>>>>>>         mm/migrate_device.c      | 16 +++++++---------
>>>>>>>>>>>>>>>>         mm/rmap.c                |  5 +++--
>>>>>>>>>>>>>>>>         6 files changed, 49 insertions(+), 16 deletions(-)
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> diff --git a/include/linux/memremap.h b/include/linux/memremap.h
>>>>>>>>>>>>>>>> index 8af304f6b504..9f752ebed613 100644
>>>>>>>>>>>>>>>> --- a/include/linux/memremap.h
>>>>>>>>>>>>>>>> +++ b/include/linux/memremap.h
>>>>>>>>>>>>>>>> @@ -41,6 +41,13 @@ struct vmem_altmap {
>>>>>>>>>>>>>>>>          * A more complete discussion of unaddressable memory may be found in
>>>>>>>>>>>>>>>>          * include/linux/hmm.h and Documentation/vm/hmm.rst.
>>>>>>>>>>>>>>>>          *
>>>>>>>>>>>>>>>> + * MEMORY_DEVICE_COHERENT:
>>>>>>>>>>>>>>>> + * Device memory that is cache coherent from device and CPU point of view. This
>>>>>>>>>>>>>>>> + * is used on platforms that have an advanced system bus (like CAPI or CXL). A
>>>>>>>>>>>>>>>> + * driver can hotplug the device memory using ZONE_DEVICE and with that memory
>>>>>>>>>>>>>>>> + * type. Any page of a process can be migrated to such memory. However no one
>>>>>>>>>>>>>>> Any page might not be right, I'm pretty sure. ... just thinking about special pages
>>>>>>>>>>>>>>> like vdso, shared zeropage, ... pinned pages ...
>>>>>>>>>>>>> Well, you cannot migrate long term pages, that's what I meant :)
>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> + * should be allowed to pin such memory so that it can always be evicted.
>>>>>>>>>>>>>>>> + *
>>>>>>>>>>>>>>>>          * MEMORY_DEVICE_FS_DAX:
>>>>>>>>>>>>>>>>          * Host memory that has similar access semantics as System RAM i.e. DMA
>>>>>>>>>>>>>>>>          * coherent and supports page pinning. In support of coordinating page
>>>>>>>>>>>>>>>> @@ -61,6 +68,7 @@ struct vmem_altmap {
>>>>>>>>>>>>>>>>         enum memory_type {
>>>>>>>>>>>>>>>>         	/* 0 is reserved to catch uninitialized type fields */
>>>>>>>>>>>>>>>>         	MEMORY_DEVICE_PRIVATE = 1,
>>>>>>>>>>>>>>>> +	MEMORY_DEVICE_COHERENT,
>>>>>>>>>>>>>>>>         	MEMORY_DEVICE_FS_DAX,
>>>>>>>>>>>>>>>>         	MEMORY_DEVICE_GENERIC,
>>>>>>>>>>>>>>>>         	MEMORY_DEVICE_PCI_P2PDMA,
>>>>>>>>>>>>>>>> @@ -143,6 +151,17 @@ static inline bool folio_is_device_private(const struct folio *folio)
>>>>>>>>>>>>>>> In general, this LGTM, and it should be correct with PageAnonExclusive I think.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> However, where exactly is pinning forbidden?
>>>>>>>>>>>>>> Long-term pinning is forbidden since it would interfere with the device
>>>>>>>>>>>>>> memory manager owning the
>>>>>>>>>>>>>> device-coherent pages (e.g. evictions in TTM). However, normal pinning
>>>>>>>>>>>>>> is allowed on this device type.
>>>>>>>>>>>>> I don't see updates to folio_is_pinnable() in this patch.
>>>>>>>>>>>> Device coherent type pages should return true here, as they are pinnable
>>>>>>>>>>>> pages.
>>>>>>>>>>> That function is only called for long-term pinnings in try_grab_folio().
>>>>>>>>>>>
>>>>>>>>>>>>> So wouldn't try_grab_folio() simply pin these pages? What am I missing?
>>>>>>>>>>>> As far as I understand this return NULL for long term pin pages.
>>>>>>>>>>>> Otherwise they get refcount incremented.
>>>>>>>>>>> I don't follow.
>>>>>>>>>>>
>>>>>>>>>>> You're saying
>>>>>>>>>>>
>>>>>>>>>>> a) folio_is_pinnable() returns true for device coherent pages
>>>>>>>>>>>
>>>>>>>>>>> and that
>>>>>>>>>>>
>>>>>>>>>>> b) device coherent pages don't get long-term pinned
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Yet, the code says
>>>>>>>>>>>
>>>>>>>>>>> struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
>>>>>>>>>>> {
>>>>>>>>>>> 	if (flags & FOLL_GET)
>>>>>>>>>>> 		return try_get_folio(page, refs);
>>>>>>>>>>> 	else if (flags & FOLL_PIN) {
>>>>>>>>>>> 		struct folio *folio;
>>>>>>>>>>>
>>>>>>>>>>> 		/*
>>>>>>>>>>> 		 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
>>>>>>>>>>> 		 * right zone, so fail and let the caller fall back to the slow
>>>>>>>>>>> 		 * path.
>>>>>>>>>>> 		 */
>>>>>>>>>>> 		if (unlikely((flags & FOLL_LONGTERM) &&
>>>>>>>>>>> 			     !is_pinnable_page(page)))
>>>>>>>>>>> 			return NULL;
>>>>>>>>>>> 		...
>>>>>>>>>>> 		return folio;
>>>>>>>>>>> 	}
>>>>>>>>>>> }
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> What prevents these pages from getting long-term pinned as stated in this patch?
>>>>>>>>>> Long-term pinning is handled by __gup_longterm_locked, which migrates
>>>>>>>>>> pages returned by __get_user_pages_locked that cannot be long-term
>>>>>>>>>> pinned. try_grab_folio is OK to grab the pages. Anything that can't be
>>>>>>>>>> long-term pinned will be migrated afterwards, and
>>>>>>>>>> __get_user_pages_locked will be retried. The migration of
>>>>>>>>>> DEVICE_COHERENT pages was implemented by Alistair in patch 5/13
>>>>>>>>>> ("mm/gup: migrate device coherent pages when pinning instead of failing").
>>>>>>>>> Thanks.
>>>>>>>>>
>>>>>>>>> __gup_longterm_locked()->check_and_migrate_movable_pages()
>>>>>>>>>
>>>>>>>>> Which checks folio_is_pinnable() and doesn't do anything if set.
>>>>>>>>>
>>>>>>>>> Sorry to be dense here, but I don't see how what's stated in this patch
>>>>>>>>> works without adjusting folio_is_pinnable().
>>>>>>>> Ugh, I think you might be right about try_grab_folio().
>>>>>>>>
>>>>>>>> We didn't update folio_is_pinnable() to include device coherent pages
>>>>>>>> because device coherent pages are pinnable. It is really just
>>>>>>>> FOLL_LONGTERM that we want to prevent here.
>>>>>>>>
>>>>>>>> For normal PUP that is done by my change in
>>>>>>>> check_and_migrate_movable_pages() which migrates pages being pinned with
>>>>>>>> FOLL_LONGTERM. But I think I incorrectly assumed we would take the
>>>>>>>> pte_devmap() path in gup_pte_range(), which we don't for coherent pages.
>>>>>>>> So I think the check in try_grab_folio() needs to be:
>>>>>>> I think I said it already (and I might be wrong without reading the
>>>>>>> code), but folio_is_pinnable() is *only* called for long-term pinnings.
>>>>>>>
>>>>>>> It should actually be called folio_is_longterm_pinnable().
>>>>>>>
>>>>>>> That's where that check should go, no?
>>>>>> David, I think you're right. We didn't catch this since the LONGTERM gup
>>>>>> test we added to hmm-test only calls to pin_user_pages. Apparently
>>>>>> try_grab_folio is called only from fast callers (ex.
>>>>>> pin_user_pages_fast/get_user_pages_fast). I have added a conditional
>>>>>> similar to what Alistair has proposed to return null on LONGTERM &&
>>>>>> (coherent_pages || folio_is_pinnable) at try_grab_folio. Also a new gup
>>>>>> test was added with LONGTERM set that calls pin_user_pages_fast.
>>>>>> Returning null under this condition it does causes the migration from
>>>>>> dev to system memory.
>>>>>>
>>>>> Why can't coherent memory simply put its checks into
>>>>> folio_is_pinnable()? I don't get it why we have to do things differently
>>>>> here.
>>>>>
>>>>>> Actually, Im having different problems with a call to PageAnonExclusive
>>>>>> from try_to_migrate_one during page fault from a HMM test that first
>>>>>> migrate pages to device private and forks to mark as COW these pages.
>>>>>> Apparently is catching the first BUG VM_BUG_ON_PGFLAGS(!PageAnon(page),
>>>>>> page)
>>>>> With or without this series? A backtrace would be great.
>>>> Here's the back trace. This happens in a hmm-test added in this patch
>>>> series. However, I have tried to isolate this BUG by just adding the COW
>>>> test with private device memory only. This is only present as follows.
>>>> Allocate anonymous mem->Migrate to private device memory->fork->try to
>>>> access to parent's anonymous memory (which will suppose to trigger a
>>>> page fault and migration to system mem). Just for the record, if the
>>>> child is terminated before the parent's memory is accessed, this problem
>>>> is not present.
>>> The only usage of PageAnonExclusive() in try_to_migrate_one() is:
>>>
>>> anon_exclusive = folio_test_anon(folio) &&
>>> 		 PageAnonExclusive(subpage);
>>>
>>> Which can only possibly fail if subpage is not actually part of the folio.
>>>
>>>
>>> I see some controversial code in the the if (folio_is_zone_device(folio)) case later:
>>>
>>> 			 * The assignment to subpage above was computed from a
>>> 			 * swap PTE which results in an invalid pointer.
>>> 			 * Since only PAGE_SIZE pages can currently be
>>> 			 * migrated, just set it to page. This will need to be
>>> 			 * changed when hugepage migrations to device private
>>> 			 * memory are supported.
>>> 			 */
>>> 			subpage = &folio->page;
>>>
>>> There we have our invalid pointer hint.
>>>
>>> I don't see how it could have worked if the child quit, though? Maybe
>>> just pure luck?
>>>
>>>
>>> Does the following fix your issue:
>> Yes, it fixed the issue. Thanks. Should we include this patch in this
>> patch series or separated?
>>
>> Regards,
>> Alex Sierra
> I'll send it right away "officially" so we can get it into 5.19. Can I
> add your tested-by?

Of course.

Alex Sierra

>
>