On 16/07/2019 16:37, Chris Wilson wrote:
Quoting Tvrtko Ursulin (2019-07-16 16:25:22)
On 16/07/2019 13:49, Chris Wilson wrote:
Following a try_to_unmap() we may want to remove the userptr and so call put_pages(). However, try_to_unmap() acquires the page lock and so we must avoid recursively locking the pages ourselves -- which means that we cannot safely acquire the lock around set_page_dirty(). Since we can't be sure of the lock, we have to risk skip dirtying the page, or else risk calling set_page_dirty() without a lock and so risk fs corruption.
So if trylock randomly fail we get data corruption in whatever data set application is working on, which is what the original patch was trying to avoid? Are we able to detect the backing store type so at least we don't risk skipping set_page_dirty with anonymous/shmemfs?
page->mapping???
Would page->mapping work? What is it telling us?
We still have the issue that if there is a mapping we should be taking the lock, and we may have both a mapping and be inside try_to_unmap().
Is this a problem? On a path with mappings we trylock and so solve the set_dirty_locked and recursive deadlock issues, and with no mappings with always dirty the page and avoid data corruption.
I think it's lose-lose. The only way to win is not to userptr :)
It's looking more and more like this indeed.
Regards,
Tvrtko