On Thu, May 25, 2023 at 5:28 PM Jiaqi Yan <jiaqiyan@xxxxxxxxxx> wrote:
>
> On Mon, May 22, 2023 at 7:43 PM HORIGUCHI NAOYA(堀口 直也)
> <naoya.horiguchi@xxxxxxx> wrote:
> >
> > On Mon, May 22, 2023 at 11:22:49AM -0700, Jiaqi Yan wrote:
> > > On Sun, May 21, 2023 at 9:50 PM HORIGUCHI NAOYA(堀口 直也)
> > > <naoya.horiguchi@xxxxxxx> wrote:
> > > >
> > > > On Fri, May 19, 2023 at 03:42:14PM -0700, Mike Kravetz wrote:
> > > > > On 05/19/23 13:54, Jiaqi Yan wrote:
> > > > > > On Wed, May 17, 2023 at 4:53 PM Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On 05/17/23 16:09, Jiaqi Yan wrote:
> > > > > > > > Adds the functionality to search a subpage's corresponding raw_hwp_page
> > > > > > > > in hugetlb page's HWPOISON list. This functionality can also tell if a
> > > > > > > > subpage is a raw HWPOISON page.
> > > > > > > >
> > > > > > > > Exports this functionality to be immediately used in the read operation
> > > > > > > > for hugetlbfs.
> > > > > > > >
> > > > > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@xxxxxxxxxx>
> > > > > > > > ---
> > > > > > > > include/linux/mm.h | 23 +++++++++++++++++++++++
> > > > > > > > mm/memory-failure.c | 26 ++++++++++++++++----------
> > > > > > > > 2 files changed, 39 insertions(+), 10 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > > > > > > > index 27ce77080c79..f191a4119719 100644
> > > > > > > > --- a/include/linux/mm.h
> > > > > > > > +++ b/include/linux/mm.h
> > > > > > >
> > > > > > > Any reason why you decided to add the following to linux/mm.h instead of
> > > > > > > linux/hugetlb.h? Since it is hugetlb specific I would have thought
> > > > > > > hugetlb.h was more appropriate.
> > > > > > >
> > > > > > > > @@ -3683,6 +3683,29 @@ enum mf_action_page_type {
> > > > > > > > */
> > > > > > > > extern const struct attribute_group memory_failure_attr_group;
> > > > > > > >
> > > > > > > > +#ifdef CONFIG_HUGETLB_PAGE
> > > > > > > > +/*
> > > > > > > > + * Struct raw_hwp_page represents information about "raw error page",
> > > > > > > > + * constructing singly linked list from ->_hugetlb_hwpoison field of folio.
> > > > > > > > + */
> > > > > > > > +struct raw_hwp_page {
> > > > > > > > + struct llist_node node;
> > > > > > > > + struct page *page;
> > > > > > > > +};
> > > > > > > > +
> > > > > > > > +static inline struct llist_head *raw_hwp_list_head(struct folio *folio)
> > > > > > > > +{
> > > > > > > > + return (struct llist_head *)&folio->_hugetlb_hwpoison;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > +/*
> > > > > > > > + * Given @subpage, a raw page in a hugepage, find its location in @folio's
> > > > > > > > + * _hugetlb_hwpoison list. Return NULL if @subpage is not in the list.
> > > > > > > > + */
> > > > > > > > +struct raw_hwp_page *find_raw_hwp_page(struct folio *folio,
> > > > > > > > + struct page *subpage);
> > > > > > > > +#endif
> > > > > > > > +
> > > > > > > > #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
> > > > > > > > extern void clear_huge_page(struct page *page,
> > > > > > > > unsigned long addr_hint,
> > > > > > > > diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> > > > > > > > index 5b663eca1f29..c49e6c2d1f07 100644
> > > > > > > > --- a/mm/memory-failure.c
> > > > > > > > +++ b/mm/memory-failure.c
> > > > > > > > @@ -1818,18 +1818,24 @@ EXPORT_SYMBOL_GPL(mf_dax_kill_procs);
> > > > > > > > #endif /* CONFIG_FS_DAX */
> > > > > > > >
> > > > > > > > #ifdef CONFIG_HUGETLB_PAGE
> > > > > > > > -/*
> > > > > > > > - * Struct raw_hwp_page represents information about "raw error page",
> > > > > > > > - * constructing singly linked list from ->_hugetlb_hwpoison field of folio.
> > > > > > > > - */
> > > > > > > > -struct raw_hwp_page {
> > > > > > > > - struct llist_node node;
> > > > > > > > - struct page *page;
> > > > > > > > -};
> > > > > > > >
> > > > > > > > -static inline struct llist_head *raw_hwp_list_head(struct folio *folio)
> > > > > > > > +struct raw_hwp_page *find_raw_hwp_page(struct folio *folio,
> > > > > > > > + struct page *subpage)
> > > > > > > > {
> > > > > > > > - return (struct llist_head *)&folio->_hugetlb_hwpoison;
> > > > > > > > + struct llist_node *t, *tnode;
> > > > > > > > + struct llist_head *raw_hwp_head = raw_hwp_list_head(folio);
> > > > > > > > + struct raw_hwp_page *hwp_page = NULL;
> > > > > > > > + struct raw_hwp_page *p;
> > > > > > > > +
> > > > > > > > + llist_for_each_safe(tnode, t, raw_hwp_head->first) {
> > > > > > >
> > > > > > > IIUC, in rare error cases a hugetlb page can be poisoned WITHOUT a
> > > > > > > raw_hwp_list. This is indicated by the hugetlb page specific flag
> > > > > > > RawHwpUnreliable or folio_test_hugetlb_raw_hwp_unreliable().
> > > > > > >
> > > > > > > Looks like this routine does not consider that case. Seems like it should
> > > > > > > always return the passed subpage if folio_test_hugetlb_raw_hwp_unreliable()
> > > > > > > is true?
> > > > > >
> > > > > > Thanks for catching this. I wonder should this routine consider
> > > > > > RawHwpUnreliable or should the caller do.
> > > > > >
> > > > > > find_raw_hwp_page now returns raw_hwp_page* in the llist entry to
> > > > > > caller (valid one at the moment), but once RawHwpUnreliable is set,
> > > > > > all the raw_hwp_page in the llist will be kfree(), and the returned
> > > > > > value becomes dangling pointer to caller (if the caller holds that
> > > > > > caller long enough). Maybe returning a bool would be safer to the
> > > > > > caller? If the routine returns bool, then checking RawHwpUnreliable
> > > > > > can definitely be within the routine.
> > > > >
> > > > > I think the check for RawHwpUnreliable should be within this routine.
> > > > > Looking closer at the code, I do not see any way to synchronize this.
> > > > > It looks like manipulation in the memory-failure code would be
> > > > > synchronized via the mf_mutex. However, I do not see how traversal and
> > > > > freeing of the raw_hwp_list called from __update_and_free_hugetlb_folio
> > > > > is synchronized against memory-failure code modifying the list.
> > > > >
> > > > > Naoya, can you provide some thoughts?
> > > >
> > > > Thanks for elaborating the issue. I think that making find_raw_hwp_page() and
> > > > folio_clear_hugetlb_hwpoison() do their works within mf_mutex can be one solution.
> > > > try_memory_failure_hugetlb(), one of the callers of folio_clear_hugetlb_hwpoison(),
> > > > already calls it within mf_mutex, so some wrapper might be needed to implement
> > > > calling path from __update_and_free_hugetlb_folio() to take mf_mutex.
> > > >
> > > > It might be a concern that mf_mutex is a big lock to cover overall hwpoison
> > > > subsystem, but I think that the impact is not so big if the changed code paths
> > > > take mf_mutex only after checking hwpoisoned hugepage. Maybe using folio_lock
> > > > to synchronize accesses to the raw_hwp_list could be possible, but currently
> > > > __get_huge_page_for_hwpoison() calls folio_set_hugetlb_hwpoison() without
> > > > folio_lock, so this approach needs update on locking rule and it sounds more
> > > > error-prone to me.
> > >
> > > Thanks Naoya, since memory_failure is the main user of raw_hwp_list, I
> > > agree mf_mutex could help to sync its two del_all operations (one from
> > > try_memory_failure_hugetlb and one from
> > > __update_and_free_hugetlb_folio).
> > >
> > > I still want to ask a perhaps stupid question, somewhat related to how
> > > to implement find_raw_hwp_page() correctly. It seems
> > > llist_for_each_safe should only be used to traverse after list entries
> > > already *deleted* via llist_del_all. But the llist_for_each_safe calls
> > > in memory_failure today is used *directly* on the raw_hwp_list. This
> > > is quite different from other users of llist_for_each_safe (for
> > > example, kernel/bpf/memalloc.c).
> >
> > Oh, I don't noticed that when writing the original code. (I just chose
> > llist_for_each_safe because I just wanted struct llist_node as a singly
> > linked list.)
>
> And maybe because you can avoid doing INIT_LIST_HEAD (which seems
> doable in folio_set_hugetlb_hwpoison if hugepage is hwpoison-ed for
> the first time)?
>
> >
> > > Why is it correct? I guess mostly
> > > because they are sync-ed under mf_mutex (except the missing coverage
> > > on __update_and_free_hugetlb_folio)?
> >
> > Yes, and there seems no good reason to use the macro llist_for_each_safe
> > here. I think it's OK to switch to simpler one like list_for_each which
> > should be supposed to be called directly. To do this, struct raw_hwp_page
> > need to have @node (typed struct list_head instead of struct llist_node).
Hi Naoya, a maybe-stupid question on list vs llist: _hugetlb_hwpoison
in folio is a void*. struct list_head is composed of two pointers to
list_node (prev and next) so folio just can't hold a list_head in the
_hugetlb_hwpoison field, right? llist_head on the other hand only
contains one pointer to llist_node first. I wonder if this is one of
the reason you picked llist instead of list in the first place.
The reason I ask is while I was testing my refactor draft, I
constantly see the refcount of the 3rd subpage in the folio got
corrupted. I am not sure about the exact reason but it feels to me
related to the above reason.
>
> I will start to work on a separate patch to switch to list_head, and
> make sure operations from __update_and_free_hugetlb_folio and
> memory_failure are serialized (hopefully without intro new locks and
> just use mf_mutex).
>
> >
> > Thanks,
> > Naoya Horiguchi
