Re: [PATCH v4] mm: cma: indefinitely retry allocations in cma_alloc

From: Chris Goldsworthy
Date: Fri Oct 09 2020 - 16:27:14 EST

On 2020-09-28 22:59, Christoph Hellwig wrote:
On Mon, Sep 28, 2020 at 01:30:27PM -0700, Chris Goldsworthy wrote:
CMA allocations will fail if 'pinned' pages are in a CMA area, since we
cannot migrate pinned pages. The _refcount of a struct page being greater
than _mapcount for that page can cause pinning for anonymous pages. This
is because try_to_unmap(), which (1) is called in the CMA allocation path,
and (2) decrements both _refcount and _mapcount for a page, will stop
unmapping a page from VMAs once the _mapcount for a page reaches 0. This
implies that after try_to_unmap() has finished successfully for a page
where _recount > _mapcount, that _refcount will be greater than 0. Later
in the CMA allocation path in migrate_page_move_mapping(), we will have one
more reference count than intended for anonymous pages, meaning the
allocation will fail for that page.

If a process ends up causing _refcount > _mapcount for a page (by either
incrementing _recount or decrementing _mapcount), such that the process is
context switched out after modifying one refcount but before modifying the
other, the page will be temporarily pinned.

One example of where _refcount can be greater than _mapcount is inside of
zap_pte_range(), which is called for all the entries of a PMD when a
process is exiting, to unmap the process's memory. Inside of
zap_pte_range(), after unammping a page with page_remove_rmap(), we have
that _recount > _mapcount. _refcount can only be decremented after a TLB
flush is performed for the page - this doesn't occur until enough pages
have been batched together for flushing. The flush can either occur inside
of zap_pte_range() (during the same invocation or a later one), or if there
aren't enough pages collected by the time we unmap all of the pages in a
process, the flush will occur in tlb_finish_mmu() in exit_mmap(). After
the flush has occurred, tlb_batch_pages_flush() will decrement the
references on the flushed pages.

Another such example like the above is inside of copy_one_pte(), which is
called during a fork. For PTEs for which pte_present(pte) == true,
copy_one_pte() will increment the _refcount field followed by the
_mapcount field of a page.

So, inside of cma_alloc(), add the option of letting users pass in
__GFP_NOFAIL to indicate that we should retry CMA allocations indefinitely,
in the event that alloc_contig_range() returns -EBUSY after having scanned
a whole CMA-region bitmap.

And who is going to use this? AS-is this just seems to add code that
isn't actually used and thus actually tested. (In addition to beeing
a relly bad idea as discussed before)

Hi Christoph,

That had slipped my mind - what we would have submitted would have been a modified /drivers/dma-heap/heaps/cma_heap.c, which would have created a "linux,cma-nofail" heap, that when allocated from, passes GFP_NOFAIL to cma_alloc(). But, since this retry approach (finite and infinite) has effectively been nacked, I've gone back to the drawing board to find either (1) a lock based approach to solving this (as posed by Andrew Morton here: https://lkml.org/lkml/2020/8/21/1490), or (2) using preempt_disable() calls.

Thanks,

Chris.

--- a/kernel/dma/contiguous.c
+++ b/kernel/dma/contiguous.c
@@ -196,7 +196,7 @@ struct page *dma_alloc_from_contiguous(struct device *dev, size_t count,
if (align > CONFIG_CMA_ALIGNMENT)
align = CONFIG_CMA_ALIGNMENT;

- return cma_alloc(dev_get_cma_area(dev), count, align, no_warn);
+ return cma_alloc(dev_get_cma_area(dev), count, align, no_warn ? __GFP_NOWARN : 0);

Also don't add pointlessly overlong lines.

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