Re: [v2 PATCH] mm: thp: check total_mapcount instead of page_mapcount

From: Yang Shi
Date: Fri May 21 2021 - 20:37:16 EST


On Fri, May 21, 2021 at 4:17 PM Hugh Dickins <hughd@xxxxxxxxxx> wrote:
>
> On Fri, 21 May 2021, Yang Shi wrote:
> > On Fri, May 21, 2021 at 10:16 AM Yang Shi <shy828301@xxxxxxxxx> wrote:
> > > On Thu, May 20, 2021 at 10:06 PM Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> > > > On Thu, 13 May 2021, Yang Shi wrote:
> > > >
> > > > > When debugging the bug reported by Wang Yugui [1], try_to_unmap() may
> > > > > return false positive for PTE-mapped THP since page_mapcount() is used
> > > > > to check if the THP is unmapped, but it just checks compound mapount and
> > > > > head page's mapcount. If the THP is PTE-mapped and head page is not
> > > > > mapped, it may return false positive.
> > > > >
> > > > > Use total_mapcount() instead of page_mapcount() for try_to_unmap() and
> > > > > do so for the VM_BUG_ON_PAGE in split_huge_page_to_list as well.
> > > > >
> > > > > This changed the semantic of try_to_unmap(), but I don't see there is
> > > > > any usecase that expects try_to_unmap() just unmap one subpage of a huge
> > > > > page. So using page_mapcount() seems like a bug.
> > > > >
> > > > > [1] https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@xxxxxxxxxxxx/
> > > > >
> > > > > Signed-off-by: Yang Shi <shy828301@xxxxxxxxx>
> > > >
> > > > I don't object to this patch, I've no reason to NAK it; but I'll
> > > > point out a few deficiencies which might make you want to revisit it.
> > > >
> > > > > ---
> > > > > v2: Removed dead code and updated the comment of try_to_unmap() per Zi
> > > > > Yan.
> > > > >
> > > > > mm/huge_memory.c | 11 +----------
> > > > > mm/rmap.c | 10 ++++++----
> > > > > 2 files changed, 7 insertions(+), 14 deletions(-)
> > > > >
> > > > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > > > > index 63ed6b25deaa..3b08b9ba1578 100644
> > > > > --- a/mm/huge_memory.c
> > > > > +++ b/mm/huge_memory.c
> > > > > @@ -2348,7 +2348,6 @@ static void unmap_page(struct page *page)
> > > > > ttu_flags |= TTU_SPLIT_FREEZE;
> > > > >
> > > > > unmap_success = try_to_unmap(page, ttu_flags);
> > > > > - VM_BUG_ON_PAGE(!unmap_success, page);
> > > >
> > > > The unused variable unmap_success has already been reported and
> > > > dealt with. But I couldn't tell what you intended: why change
> > > > try_to_unmap()'s output, if you then ignore it?
> > >
> > > Because some other callers of try_to_unmap() check the output.
>
> memory-failure.c and one in vmscan.c: most callers are not interested,
> so replacing a quick atomic_read by a scan of 512 struct pages is not
> necessarily a good idea. But I am exaggerating, it's not usually that
> bad: probably most of those callers will rarely encounter a THP.
>
> > >
> > > >
> > > > > }
> > > > >
> > > > > static void remap_page(struct page *page, unsigned int nr)
> > > > > @@ -2718,7 +2717,7 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
> > > > > }
> > > > >
> > > > > unmap_page(head);
> > > > > - VM_BUG_ON_PAGE(compound_mapcount(head), head);
> > > > > + VM_BUG_ON_PAGE(total_mapcount(head), head);
> > > >
> > > > And having forced try_to_unmap() to do the expensive-on-a-THP
> > > > total_mapcount() calculation, you now repeat it here. Better
> > > > to stick with the previous VM_BUG_ON_PAGE(!unmap_success).
> > > >
> > > > Or better a VM_WARN_ONCE(), accompanied by dump_page()s as before,
> > > > to get some perhaps useful info out, which this patch has deleted.
> > > > Probably better inside unmap_page() than cluttering up here.
> > >
> > > Moving the BUG or WARN into unmap_page() looks fine to me. IIUC,
> > > VM_BUG_ON_PAGE or VM_WARN_ON_PAGE does call dump_page(), so dumping
> > > something useful is not deleted.
>
> Yes, you're right, I was forgetting that. The original DEBUG_VM block
> does do a separate dump_page() for head and tail (if different), but
> I don't think you need to replicate that now: (a) it was more a mark
> of frustration at not getting enough info to solve the problem (and
> if we really want more info, it's a dump of 512 struct pages needed);
> and (b) I think Matthew and others have enhanced dump_page() meanwhile,
> to issue some head info when dumping a tail; and (c) most splitters
> pass the head anyway, IIRC there's only one or two who pass a tail.

Yes, deferred split queue is one of them. We may just get rid of it by
passing the head page to reduce the exception case.

>
> >
> > I misspelled the function name. There is *NOT* VM_WARN_ON_PAGE(), the
> > name is VM_WARN_ON_ONCE_PAGE(). We may need to add VM_WARN_ON_PAGE()
> > since I'd like this warning to be printed every time when it is met.
>
> I get very confused by all those variants too; and then indeed very
> often the one variant you think you want turns out not to exist yet.
>
> But I'm not sure that printing it every time will be good: isn't it
> quite likely that the situation is long-lasting, and that the page
> is on the anon deferred or shmem unused list, and will keep coming
> back to report the same unhelpful info, just cluttering up the logs?
> I'd be more comfortable with _ONCE myself.
>
> But I'd also suggest dropping the VM_ part of it: we have no idea
> how often this happens on non-DEBUG_VM machines in production,
> and a pattern might be revealed there which would help to solve it.

If we use the non-VM_ version we have to call dump_page() explicitly
and duplicate what VM_WARN_ON_ONCE_PAGE() does since WARN_* don't dump
struct page at all.

>
> >
> > >
> > > >
> > > > VM_WARN_ONCE() because nothing in this patch fixes whatever Wang
> > > > Yugui is suffering from; and (aside from the BUG()) it's harmless,
> > > > because there are other ways in which the page_ref_freeze() can fail,
> > > > and that is allowed for. We would like to know when this problem
> > > > occurs: there is something wrong, but no reason to crash.
> > >
> > > Yes, it fixes nothing. I didn't figure out why try_to_unmap() failed.
> > > I agree BUG_ON could be relaxed.
>
> I have to confess, I have known of several reasons why one or other
> of those BUG()s can be hit. Sigh. I'll rearrange priorities, research
> back through what work I've done and not had time to upstream, filter
> out the patches relevant to this "mysterious failure to unmap a THP".
>
> I did not speak up earlier because I did not notice any obvious
> connection between Wang Yugui's occurrence, and anything we have
> fixed; and (as usual) there's other work I should be doing. But
> it's getting to feel dishonest, having this conversation without
> showing our fixes: I'd better sort those out and post (but that
> will take me more than a week I expect).

That would be great. I tried to figure out the problem, but the bug
report doesn't show a pattern, and I didn't get too much by staring at
the code.

>
> Not saying that you should stop with this patch at all:
> no, I can easily rebase on top of what you end up with here.

Thanks.

>
> > >
> > > >
> > > > >
> > > > > /* block interrupt reentry in xa_lock and spinlock */
> > > > > local_irq_disable();
> > > > > @@ -2758,14 +2757,6 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
> > > > > __split_huge_page(page, list, end);
> > > > > ret = 0;
> > > > > } else {
> > > > > - if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
> > > > > - pr_alert("total_mapcount: %u, page_count(): %u\n",
> > > > > - mapcount, count);
> > > > > - if (PageTail(page))
> > > > > - dump_page(head, NULL);
> > > > > - dump_page(page, "total_mapcount(head) > 0");
> > > > > - BUG();
> > > > > - }
> > > >
> > > > This has always looked ugly (as if Kirill had hit an unsolved case),
> > > > so it is nice to remove it; but you're losing the dump_page() info,
> > > > and not really gaining anything more than a cosmetic cleanup.
> > >
> > > As I mentioned above, IIUC VM_BUG_ON_PAGE and VM_WARN_ON_PAGE do call
> > > dump_page().
> > >
> > > >
> > > > > spin_unlock(&ds_queue->split_queue_lock);
> > > > > fail: if (mapping)
> > > > > xa_unlock(&mapping->i_pages);
> > > > > diff --git a/mm/rmap.c b/mm/rmap.c
> > > > > index 693a610e181d..f52825b1330d 100644
> > > > > --- a/mm/rmap.c
> > > > > +++ b/mm/rmap.c
> > > > > @@ -1742,12 +1742,14 @@ static int page_not_mapped(struct page *page)
> > > > > }
> > > > >
> > > > > /**
> > > > > - * try_to_unmap - try to remove all page table mappings to a page
> > > > > - * @page: the page to get unmapped
> > > > > + * try_to_unmap - try to remove all page table mappings to a page and the
> > > > > + * compound page it belongs to
> > > > > + * @page: the page or the subpages of compound page to get unmapped
> > > > > * @flags: action and flags
> > > > > *
> > > > > * Tries to remove all the page table entries which are mapping this
> > > > > - * page, used in the pageout path. Caller must hold the page lock.
> > > > > + * page and the compound page it belongs to, used in the pageout path.
> > > > > + * Caller must hold the page lock.
> > > > > *
> > > > > * If unmap is successful, return true. Otherwise, false.
> > > > > */
> > > > > @@ -1777,7 +1779,7 @@ bool try_to_unmap(struct page *page, enum ttu_flags flags)
> > > > > else
> > > > > rmap_walk(page, &rwc);
> > > > >
> > > > > - return !page_mapcount(page) ? true : false;
> > > > > + return !total_mapcount(page) ? true : false;
> > > >
> > > > That always made me wince: "return !total_mapcount(page);" surely.
> > >
> > > But page_mapcount() seems not correct, it may return false positive,
> > > right? Or it is harmless?
>
> I wasn't disagreeing with your use of total_mapcount() in that comment,
> just with your propagating the "return boolean() ? true : false" style.
>
> > >
> > > And I actually spotted a few other places which should use
> > > total_mapcount() but using page_mapcount() instead, for example, some
> > > madvise code check if the page is shared by using page_mapcount(),
> > > however it may return false negative (double mapped THP, but head page
> > > is not PTE-mapped, just like what Wang Yugui reported). It is not
> > > fatal, but not expected behavior. I understand total_mapcount() is
> > > expensive, so is it a trade-off between cost and correctness or just
> > > overlooked the false negative case in the first place? I can't tell.
>
> I haven't looked through those other places. I do agree with you that
> total_mapcount() or page_mapped() is likely to be more correct in most
> instances, but it's often more costly than it's worth (on THPs).

Kind of. But some of them are not a hot path, for example, madvise.

>
> And you've reminded me that some months back, Andrea pointed out how
> total_mapcount() can race with something, and give the wrong answer:
> he was arguing against relying on it (but has patches with additional
> locking to make it more reliable - though that gets more complicated).
>
> In Google prod actually, we have mostly avoided using total_mapcount(),
> merely because I feared the cost and couldn't convince myself of its
> safety - I'm not saying I ever saw it go wrong, I just didn't feel
> comfortable relying on it. Maybe that's part of what I need to post
> after sifting through.

I'm supposed it should be safe (but of course more costly) as long as
the page table lock is held. Very few places call page_mapcount() or
total_mapcount() *without* page table lock held, compaction is an
example.

>
> Hugh
>
> > >
> > > >
> > > > Or slightly better, "return !page_mapped(page);", since at least that
> > > > one breaks out as soon as it sees a mapcount. Though I guess I'm
> > > > being silly there, since that case should never occur, so both
> > > > total_mapcount() and page_mapped() scan through all pages.
> > > >
> > > > Or better, change try_to_unmap() to void: most callers ignore its
> > > > return value anyway, and make their own decisions; the remaining
> > > > few could be changed to do the same. Though again, I may be
> > > > being silly, since the expensive THP case is not the common case.
> > >
> > > I'd say half callers ignore its return value. But I think it should be
> > > worth doing. At least we could remove half unnecessary
> > > total_mapcount() or page_mapped() call.
> > >
> > > Thanks a lot for all the suggestions, will incorporate them in the new version.
> > >
> > > >
> > > > > }
> > > > >
> > > > > /**
> > > > > --
> > > > > 2.26.2