Alistair Popple <apopple@xxxxxxxxxx> writes: > Yang Shi <shy828301@xxxxxxxxx> writes: > >> On Mon, Sep 26, 2022 at 2:37 AM Alistair Popple <apopple@xxxxxxxxxx> wrote: >>> >>> >>> Huang Ying <ying.huang@xxxxxxxxx> writes: >>> >>> > This is a preparation patch to batch the page unmapping and moving >>> > for the normal pages and THP. >>> > >>> > In this patch, unmap_and_move() is split to migrate_page_unmap() and >>> > migrate_page_move(). So, we can batch _unmap() and _move() in >>> > different loops later. To pass some information between unmap and >>> > move, the original unused newpage->mapping and newpage->private are >>> > used. >>> >>> This looks like it could cause a deadlock between two threads migrating >>> the same pages if force == true && mode != MIGRATE_ASYNC as >>> migrate_page_unmap() will call lock_page() while holding the lock on >>> other pages in the list. Therefore the two threads could deadlock if the >>> pages are in a different order. >> >> It seems unlikely to me since the page has to be isolated from lru >> before migration. The isolating from lru is atomic, so the two threads >> unlikely see the same pages on both lists. > > Oh thanks! That is a good point and I agree since lru isolation is > atomic the two threads won't see the same pages. migrate_vma_setup() > does LRU isolation after locking the page which is why the potential > exists there. We could potentially switch that around but given > ZONE_DEVICE pages aren't on an lru it wouldn't help much. > >> But there might be other cases which may incur deadlock, for example, >> filesystem writeback IIUC. Some filesystems may lock a bunch of pages >> then write them back in a batch. The same pages may be on the >> migration list and they are also dirty and seen by writeback. I'm not >> sure whether I miss something that could prevent such a deadlock from >> happening. > > I'm not overly familiar with that area but I would assume any filesystem > code doing this would already have to deal with deadlock potential. Thank you very much for pointing this out. I think the deadlock is a real issue. Anyway, we shouldn't forbid other places in kernel to lock 2 pages at the same time. The simplest solution is to batch page migration only if mode == MIGRATE_ASYNC. Then we may consider to fall back to non-batch mode if mode != MIGRATE_ASYNC and trylock page fails. Best Regards, Huang, Ying [snip]
