Skip to contentSkip to conversation

cma_alloc(), add sleep-and-retry for temporary page pinning

DISCUSSION [email protected] 2 participants 5 messages Aug 11, 2020 – Sep 11, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw
Quoted 15 lines · Andrew Morton
Quoted 10 lines · On Wed, 5 Aug 2020 19:56:21 -0700 Chris Goldsworthy <[email protected]>…

On mobile devices, failure to allocate from a CMA area constitutes a functional failure. Sometimes during CMA allocations, we have observed that pages in a CMA area allocated through alloc_pages(), that we're trying to migrate away to make room for a CMA allocation, are temporarily pinned. This temporary pinning can occur when a process that owns the pinned page is being forked (the example is explained further in the commit text). This patch addresses this issue by adding a sleep-and-retry loop in cma_alloc() . There's another example we know of similar to the above that occurs during exit_mmap() (in zap_pte_range() specifically), but I need to determine if this is still relevant today.

Quoted 6 lines

Sounds fairly serious but boy, we're late for 5.9.

I can queue it for 5.10 with a cc:stable so that it gets backported into earlier kernels a couple of months from now, if we think the seriousness justifies backporting(?).

Queuing this seems like the best way to proceed, if we were to pick up this patch. I think we can forgo back-porting this, as this is something that will only be needed as vendors such as our selves start using Google's Generic Kernel Image (we've carried this patch in our tree for over four years).

Quoted 4 lines

And... it really is a sad little patch, isn't it? Instead of fixing the problem, it reduces the problem's probability by 5x. Can't we do better than this?

I have one alternative in mind. I have been able to review the exit_mmap() case, so before proceeding, let's do a breakdown of the problem: we can categorize the pinning issue we're trying to address here as being one of
(1) incrementing _refcount and getting context-switched out before incrementing _mapcount (applies to forking a process / copy_one_pte()), and
(2) decrementing _mapcount and getting context-switched out before decrementing _refcount (applies to tearing down a process / exit_mmap()). So, one alternative would be to insert preempt_disable/enable() calls at affected sites. So, for the copy_one_pte() pinning case, we could do the following inside of copy_one_pte():

        if (page) {
+               preempt_disable();
                get_page(page);
                page_dup_rmap(page, false);
+               preempt_enable();
                rss[mm_counter(page)]++;
        }

I'm not sure if this approach would be acceptable for the exit_mmap() pinning case (applicable when CONFIG_MMU_GATHER_NO_GATHER=y). For the purposes of this discussion, we can look at two function calls inside of exit_mmap(), in the order they're called in, to show how the pinning is occuring:

1. Calling unmap_vmas(): this unmaps the pages in each VMA for an exiting task, using zap_pte_range() - zap_pte_range() reduces the _mapcount for each page in a VMA, using page_remove_rmap(). After calling page_remove_rmap(), the page is placed into a list in __tlb_remove_page(). This list of pages will be used when flushing TLB entries later on during the process teardown.

2. Calling tlb_finish_mmu(): This is will flush the TLB entries associated with pages, before calling put_page() on them, using the previously collected pages from __tlb_remove_page() - the call flow
is
tlb_flush_mmu() > tlb_flush_mmu() > tlb_flush_mmu_free() > tlb_batch_pages_flush() > free_pages_and_swap_cache() > release_pages(), where release_pages() is described as a "batched put_page()"

The preempt_disable/enable() approach would entail doing the following inside of exit_mmap():

+ preempt_disable();
unmap_vmas(&tlb, vma, 0, -1); free_pgtables(&tlb, vma, FIRST_USER_ADDRESS,

USER_PGTABLES_CEILING);
        tlb_finish_mmu(&tlb, 0, -1);

+ preempt_enable();

I'm not sure doing this is feasible, given how long it could take to do the process teardown.

The good thing about this patch is that it has been stable in our kernel for four years (though for some SoCs we increased the retry counts). One thing to stress is that there are other instances of CMA page pinning, that this patch isn't attempting to address. Please let me know if you're okay with queuing this for the 5.10 merge window - if you are, I can add an option to configure the number of retries, and will resend the patch once the 5.9 merge window closes.

Thanks,

Chris.

Signature · 3 lines
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#1 lore raw
Quoted 96 lines · On 2020-08-11 15:20, [email protected] wrote:
Quoted 15 lines · Andrew Morton
Quoted 10 lines · On Wed, 5 Aug 2020 19:56:21 -0700 Chris Goldsworthy <[email protected]>…

On mobile devices, failure to allocate from a CMA area constitutes a functional failure. Sometimes during CMA allocations, we have observed that pages in a CMA area allocated through alloc_pages(), that we're trying to migrate away to make room for a CMA allocation, are temporarily pinned. This temporary pinning can occur when a process that owns the pinned page is being forked (the example is explained further in the commit text). This patch addresses this issue by adding a sleep-and-retry loop in cma_alloc() . There's another example we know of similar to the above that occurs during exit_mmap() (in zap_pte_range() specifically), but I need to determine if this is still relevant today.

Quoted 6 lines

Sounds fairly serious but boy, we're late for 5.9.

I can queue it for 5.10 with a cc:stable so that it gets backported into earlier kernels a couple of months from now, if we think the seriousness justifies backporting(?).

Queuing this seems like the best way to proceed, if we were to pick up this patch.
I think we can forgo back-porting this, as this is something that will only be needed as vendors such as our selves start using Google's Generic Kernel Image (we've carried this patch in our tree for over four years).

Quoted 4 lines

And... it really is a sad little patch, isn't it? Instead of fixing the problem, it reduces the problem's probability by 5x. Can't we do better than this?

I have one alternative in mind. I have been able to review the exit_mmap() case, so before proceeding, let's do a breakdown of the problem: we can categorize the pinning issue we're trying to address here as being one of
(1) incrementing _refcount and getting context-switched out before incrementing _mapcount (applies to forking a process / copy_one_pte()), and
(2) decrementing _mapcount and getting context-switched out before decrementing _refcount (applies to tearing down a process / exit_mmap()). So, one alternative would be to insert preempt_disable/enable() calls at affected sites. So, for the copy_one_pte() pinning case, we could do the following inside of copy_one_pte():

        if (page) {
(quoted diff) +2 -0
+               preempt_disable();
                get_page(page);
                page_dup_rmap(page, false);
+               preempt_enable();
                rss[mm_counter(page)]++;
        }

I'm not sure if this approach would be acceptable for the exit_mmap() pinning case (applicable when CONFIG_MMU_GATHER_NO_GATHER=y). For the purposes of this discussion, we can look at two function calls inside of exit_mmap(), in the order they're called in, to show how the pinning is occuring:

1. Calling unmap_vmas(): this unmaps the pages in each VMA for an exiting task, using zap_pte_range() - zap_pte_range() reduces the _mapcount for each page in a VMA, using page_remove_rmap(). After calling page_remove_rmap(), the page is placed into a list in __tlb_remove_page(). This list of pages will be used when flushing TLB entries later on during the process teardown.

2. Calling tlb_finish_mmu(): This is will flush the TLB entries associated with pages, before calling put_page() on them, using the previously collected pages from __tlb_remove_page() - the call flow
is
tlb_flush_mmu() > tlb_flush_mmu() > tlb_flush_mmu_free() > tlb_batch_pages_flush() > free_pages_and_swap_cache() > release_pages(), where release_pages() is described as a "batched put_page()"

The preempt_disable/enable() approach would entail doing the following inside of exit_mmap():

+ preempt_disable();
unmap_vmas(&tlb, vma, 0, -1); free_pgtables(&tlb, vma, FIRST_USER_ADDRESS,

USER_PGTABLES_CEILING);
        tlb_finish_mmu(&tlb, 0, -1);

+ preempt_enable();

I'm not sure doing this is feasible, given how long it could take to do the process teardown.

The good thing about this patch is that it has been stable in our kernel for four years (though for some SoCs we increased the retry counts). One thing to stress is that there are other instances of CMA page pinning, that this patch isn't attempting to address. Please let me know if you're okay with queuing this for the 5.10 merge window - if you are, I can add an option to configure the number of retries, and will resend the patch once the 5.9 merge window closes.

Quoted 3 lines

Thanks,

Chris.

Hi Andrew,

Have you been able to give the patch any further consideration?

Thanks,

Chris.

Signature · 3 lines
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
maintainer
#2 lore raw
In reply to [email protected]
Quoted 22 lines · On Tue, 11 Aug 2020 15:20:47 -0700 [email protected] wrote:
Quoted 20 lines · Andrew Morton
Quoted 16 lines · On Wed, 5 Aug 2020 19:56:21 -0700 Chris Goldsworthy <[email protected]>…

On mobile devices, failure to allocate from a CMA area constitutes a functional failure. Sometimes during CMA allocations, we have observed
that pages in a CMA area allocated through alloc_pages(), that we're trying
to migrate away to make room for a CMA allocation, are temporarily pinned.
This temporary pinning can occur when a process that owns the pinned page
is being forked (the example is explained further in the commit text). This patch addresses this issue by adding a sleep-and-retry loop in cma_alloc() . There's another example we know of similar to the above that
occurs during exit_mmap() (in zap_pte_range() specifically), but I need to
determine if this is still relevant today.

Quoted 8 lines
Quoted 6 lines

Sounds fairly serious but boy, we're late for 5.9.

I can queue it for 5.10 with a cc:stable so that it gets backported into earlier kernels a couple of months from now, if we think the seriousness justifies backporting(?).

Queuing this seems like the best way to proceed

I'd really prefer not. It's very hacky and it isn't a fix - it's a likelihood-reducer.

Quoted 21 lines
Quoted 4 lines

And... it really is a sad little patch, isn't it? Instead of fixing the problem, it reduces the problem's probability by 5x. Can't we do better than this?

I have one alternative in mind. I have been able to review the exit_mmap()
case, so before proceeding, let's do a breakdown of the problem: we can categorize the pinning issue we're trying to address here as being one of
(1) incrementing _refcount and getting context-switched out before incrementing _mapcount (applies to forking a process / copy_one_pte()), and
(2) decrementing _mapcount and getting context-switched out before decrementing _refcount (applies to tearing down a process / exit_mmap()).
So, one alternative would be to insert preempt_disable/enable() calls at affected sites. So, for the copy_one_pte() pinning case, we could do the following inside of copy_one_pte():

         if (page) {

Review

Quoted 6 lines
(quoted diff) +2 -0
+               preempt_disable();
                 get_page(page);
                 page_dup_rmap(page, false);
+               preempt_enable();
                 rss[mm_counter(page)]++;
         }

This would make the retry loop much more reliable, and preempt_disable() is fast. Such additions should be clearly commented (a bare preempt_disable() explains nothing). Perhaps by wrapping the prerempt_disable() in a suitably-named wrapper function.

But it's still really unpleasant.

Quoted 3 lines

The good thing about this patch is that it has been stable in our kernel for four years (though for some SoCs we increased the retry counts).

That's discouraging!

Quoted 4 lines

One
thing to stress is that there are other instances of CMA page pinning, that
this patch isn't attempting to address.

Oh. How severe are these?

Quoted 6 lines

Please let me know if you're
okay
with queuing this for the 5.10 merge window - if you are, I can add an option to configure the number of retries, and will resend the patch once
the 5.9 merge window closes.

Well. Why not wait infinitely? Because there are other sources of CMA page pinning, I guess.

Could we take a sleeping lock on the exit_mmap() path and on the migration path? So that the migration path automatically waits for the exact amount of time?

#3 lore raw
Quoted 8 lines · Andrew Morton
Quoted 4 lines · On Tue, 11 Aug 2020 15:20:47 -0700 [email protected] wrote:

One
thing to stress is that there are other instances of CMA page pinning, that
this patch isn't attempting to address.

Oh. How severe are these?

Hey Andrew,

- get_user_pages() will pin pages (I think that's a 100% guarantee but
I'd need to double check that). There isn't a workaround for this as far as I know.
- One issue we've encountered is when the pages for buffer heads are
stuck on an LRU list (see the call to buffer_busy() in drop_buffers() https://elixir.bootlin.com/linux/v5.8.8/source/fs/buffer.c#L3225). We deal with this by allowing the buffers to be dropped, if the reason buffer_busy() returns true is because the page is on an LRU list.

Quoted 6 lines

Well. Why not wait infinitely? Because there are other sources of CMA page pinning, I guess.

Could we take a sleeping lock on the exit_mmap() path and on the migration path? So that the migration path automatically waits for the exact amount of time?

Retrying indefinitely whilst alloc_contig_range() returns -EBUSY is actually a viable approach. From the perspective of how long it takes to perform a CMA allocation, it is preferable compared to the lock-based method, in terms of how long it would take to do a CMA allocation. With our current method, we attempt allocations across an entire CMA region before sleeping and retrying. With the lock-based method, we'd be sleeping on a per-page basis - this could lead to a situation where we spend a great deal of time waiting for a particular page A to be freed, that lies in the subset of the CMA region we're allocating form. We could have instead given up on allocating that subset of the CMA region (because page A is pinned), and have moved on to a different subset of the CMA region, and have successfully allocated that subset, whilst page A is still pinned.

I have a patch ready that does this indefinite-retrying, that will be sent in reply to this e-mail.

Regards,

Chris.

--
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

#4 lore raw
In reply to Andrew Morton
Quoted 8 lines · Andrew Morton
Quoted 4 lines · On Tue, 11 Aug 2020 15:20:47 -0700 [email protected] wrote:

One
thing to stress is that there are other instances of CMA page pinning, that
this patch isn't attempting to address.

Oh. How severe are these?

Hey Andrew,

- get_user_pages() will pin pages (I think that's a 100% guarantee but
I'd need to double check that). There isn't a workaround for this as far as I know.
- One issue we've encountered is when the pages for buffer heads are
stuck on an LRU list (see the call to buffer_busy() in drop_buffers() https://elixir.bootlin.com/linux/v5.8.8/source/fs/buffer.c#L3225). We deal with this by allowing the buffers to be dropped, if the reason buffer_busy() returns true is because the page is on an LRU list.

Quoted 6 lines

Well. Why not wait infinitely? Because there are other sources of CMA page pinning, I guess.

Could we take a sleeping lock on the exit_mmap() path and on the migration path? So that the migration path automatically waits for the exact amount of time?

Retrying indefinitely whilst alloc_contig_range() returns -EBUSY is actually a viable approach. From the perspective of how long it takes to perform a CMA allocation, it is preferable compared to the lock-based method, in terms of how long it would take to do a CMA allocation. With our current method, we attempt allocations across an entire CMA region before sleeping and retrying. With the lock-based method, we'd be sleeping on a per-page basis - this could lead to a situation where we spend a great deal of time waiting for a particular page A to be freed, that lies in the subset of the CMA region we're allocating form. We could have instead given up on allocating that subset of the CMA region (because page A is pinned), and have moved on to a different subset of the CMA region, and have successfully allocated that subset, whilst page A is still pinned.

I have a patch ready that does this indefinite-retrying, that will be sent in reply to this e-mail.

Regards,

Chris.

--
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project