On 2020-08-21 15:01, Andrew Morton wrote:
On Tue, 11 Aug 2020 15:20:47 -0700 cgold...@codeaurora.org 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.

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

Reply via email to