On Wed, Nov 02, 2022 at 11:24:57AM -0700, James Houghton wrote: > On Sun, Oct 30, 2022 at 2:29 PM Peter Xu <peterx@xxxxxxxxxx> wrote: > > > > RCU makes sure the pte_t* won't go away from under us. Please refer to the > > comment above huge_pte_offset() for more information. > > > > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx> > > --- > > mm/hugetlb.c | 6 +++++- > > 1 file changed, 5 insertions(+), 1 deletion(-) > > > > diff --git a/mm/hugetlb.c b/mm/hugetlb.c > > index 9869c12e6460..85214095fb85 100644 > > --- a/mm/hugetlb.c > > +++ b/mm/hugetlb.c > > @@ -6229,10 +6229,12 @@ struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, > > if (WARN_ON_ONCE(flags & FOLL_PIN)) > > return NULL; > > > > + /* For huge_pte_offset() */ > > + rcu_read_lock(); > > retry: > > pte = huge_pte_offset(mm, haddr, huge_page_size(h)); > > if (!pte) > > - return NULL; > > + goto out_rcu; > > > > ptl = huge_pte_lock(h, mm, pte); > > Just to make sure -- this huge_pte_lock doesn't count as "blocking" > (for the purposes of what is allowed in an RCU read-side critical > section), right? If so, great! Yeah I think spinlock should be fine, iiuc it'll be fine as long as we don't proactively yield with any form of sleeping locks. For RT sleepable spinlock should also be fine in this case, as explicitly mentioned in the RCU docs: b. What about the -rt patchset? If readers would need to block in an non-rt kernel, you need SRCU. If readers would block in a -rt kernel, but not in a non-rt kernel, SRCU is not necessary. (The -rt patchset turns spinlocks into sleeplocks, hence this distinction.) > But I think we need to call `rcu_read_unlock` before entering > `__migration_entry_wait_huge`, as that function really can block. Right, let me revisit this after I figure out how to do with the hugetlb_fault() path first, as you commented in the other patch. Actually here I really think we should just remove the migration chunk and return with page==NULL, since I really don't think follow_page_mask should block at all.. then for !sleep cases (FOLL_NOWAIT) or follow_page we'll return the NULL upwards early, while for generic GUP (__get_user_pages) we'll just wait in the upcoming faultin_page(). That's afaict what we do with non-hugetlb memories too (after the recent removal of FOLL_MIGRATE in 4a0499782a). -- Peter Xu
