Matthew Wilcox <willy@xxxxxxxxxxxxx> writes:
> On Wed, Mar 23, 2022 at 11:29:12AM +1100, Alistair Popple wrote:
>> Matthew Wilcox <willy@xxxxxxxxxxxxx> writes:
>> > It's not been reported before, but it seems obvious why it triggers.
>> > I think it's ffa65753c431 "mm/migrate.c: rework migration_entry_wait()
>> > to not take a pageref".
>>
>> It's obvious the BUG_ON() is because the page isn't locked, but it's less
>> obvious to me at least why we have a migration entry pointing to an unlocked
>> page.
>>
>> > There's a lot of inlining going on, so let's write this out:
>> >
>> > __handle_mm_fault
>> > handle_pte_fault
>> > do_swap_page():
>> >
>> > entry = pte_to_swp_entry(vmf->orig_pte);
>> > if (unlikely(non_swap_entry(entry))) {
>> > if (is_migration_entry(entry)) {
>> > migration_entry_wait(vma->vm_mm, vmf->pmd,
>> > vmf->address);
>> >
>> > We don't (and can't) have the underlying page locked here. We're just
>> > waiting for it to finish migration.
>>
>> Why can't the page be locked here? Waiting for migration to finish is equivalent
>> to waiting for the page to be unlocked. __unmap_and_move() follows this rough
>> sequence:
>>
>> __unmap_and_move()
>> if (!trylock_page(page)) {
>> if (!force || mode `= MIGRATE_ASYNC)
>> goto out;
>>
>> /*
>> * It's not safe for direct compaction to call lock_page.
>> * For example, during page readahead pages are added locked
>> * to the LRU. Later, when the IO completes the pages are
>> * marked uptodate and unlocked. However, the queueing
>> * could be merging multiple pages for one bio (e.g.
>> * mpage_readahead). If an allocation happens for the
>> * second or third page, the process can end up locking
>> * the same page twice and deadlocking. Rather than
>> * trying to be clever about what pages can be locked,
>> * avoid the use of lock_page for direct compaction
>> * altogether.
>> */
>> if (current->flags & PF_MEMALLOC)
>> goto out;
>>
>> lock_page(page);
>> }
>>
>> [...]
>>
>> if (unlikely(!trylock_page(newpage)))
>> goto out_unlock;
>>
>> [...]
>>
>> } else if (page_mapped(page)) {
>> /* Establish migration ptes */
>> VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
>> page);
>> try_to_migrate(page, 0);
>> page_was_mapped = true;
>> }
>>
>> if (!page_mapped(page))
>> rc = move_to_new_page(newpage, page, mode);
>>
>> if (page_was_mapped)
>> remove_migration_ptes(page,
>> rc =' MIGRATEPAGE_SUCCESS ? newpage : page, false);
>>
>> Both the source and destination pages (page and newpage respectively) are locked
>> prior to installing migration entries in try_to_migrate() and unlocked after
>> migration entries are removed. In order to remove a migration entry the PTL is
>> required. Therefore while holding the PTL for a migration entry it should be
>> impossible to observe it pointing to an unlocked page.
>
> Hm, right. We check the PTE initially, then take the PTL in
> __migration_entry_wait() and then re-check that it's still a migration
> entry. I can't think of how we're seeing this ...
Right. The only thing I can think of is if we somehow didn't find all the
migration entries to remove once migration is complete. There have been a few
changes to page_vma_mapped_walk() but I haven't spotted anything yet that would
cause this.
