On Wed, Jun 24, 2020 at 12:21 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Wed, Jun 24, 2020 at 08:14:17PM +0100, Chris Wilson wrote: > > A general rule of thumb is that shrinkers should be fast and effective. > > They are called from direct reclaim at the most incovenient of times when > > the caller is waiting for a page. If we attempt to reclaim a page being > > pinned for active dma [pin_user_pages()], we will incur far greater > > latency than a normal anonymous page mapped multiple times. Worse the > > page may be in use indefinitely by the HW and unable to be reclaimed > > in a timely manner. > > A pinned page can't be migrated, discarded or swapped by definition - > it would cause data corruption. > > So, how do things even get here and/or work today at all? I think the > explanation is missing something important. The __remove_mapping() will try to freeze page count if the count is expected otherwise just not discard the page. I'm not quite sure why the check is done that late, my wild guess is to check the refcount at the last minute so there might be a chance the pin gets released right before it. But I noticed a bug in __remove_ampping() for THP since THP's dma pinned count is recorded in the tail page's hpage_pinned_refcount instead of refcount. So, the refcount freeze might be successful for pinned THP. Chris's patch could solve this issue too, but I'm not sure if it is worth backing earlier once dma pinned page is met. If it is worth, the follow-up question is why not just skip such page in scan phase? > > Jason >
