On Wed, Jul 5, 2023 at 9:10 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote: > > On Wed, Jul 5, 2023 at 1:08 AM David Hildenbrand <david@xxxxxxxxxx> wrote: > > > > On 05.07.23 08:37, Suren Baghdasaryan wrote: > > > When forking a child process, parent write-protects an anonymous page > > > and COW-shares it with the child being forked using copy_present_pte(). > > > Parent's TLB is flushed right before we drop the parent's mmap_lock in > > > dup_mmap(). If we get a write-fault before that TLB flush in the parent, > > > and we end up replacing that anonymous page in the parent process in > > > do_wp_page() (because, COW-shared with the child), this might lead to > > > some stale writable TLB entries targeting the wrong (old) page. > > > Similar issue happened in the past with userfaultfd (see flush_tlb_page() > > > call inside do_wp_page()). > > > Lock VMAs of the parent process when forking a child, which prevents > > > concurrent page faults during fork operation and avoids this issue. > > > This fix can potentially regress some fork-heavy workloads. Kernel build > > > time did not show noticeable regression on a 56-core machine while a > > > stress test mapping 10000 VMAs and forking 5000 times in a tight loop > > > shows ~5% regression. If such fork time regression is unacceptable, > > > disabling CONFIG_PER_VMA_LOCK should restore its performance. Further > > > optimizations are possible if this regression proves to be problematic. > > > > Out of interest, did you also populate page tables / pages for some of these > > VMAs, or is this primarily looping over 10000 VMAs that don't actually copy any > > page tables? > > I did not populate the page tables, therefore this represents the > worst case scenario (the share of time used to lock the VMAs is > maximized). > > > > > > > > > Suggested-by: David Hildenbrand <david@xxxxxxxxxx> > > > Reported-by: Jiri Slaby <jirislaby@xxxxxxxxxx> > > > Closes: https://lore.kernel.org/all/dbdef34c-3a07-5951-e1ae-e9c6e3cdf51b@xxxxxxxxxx/ > > > Reported-by: Holger Hoffstätte <holger@xxxxxxxxxxxxxxxxxxxxxx> > > > Closes: https://lore.kernel.org/all/b198d649-f4bf-b971-31d0-e8433ec2a34c@xxxxxxxxxxxxxxxxxxxxxx/ > > > Reported-by: Jacob Young <jacobly.alt@xxxxxxxxx> > > > Closes: https://bugzilla.kernel.org/show_bug.cgi?id=217624 > > > Fixes: 0bff0aaea03e ("x86/mm: try VMA lock-based page fault handling first") > > > Cc: stable@xxxxxxxxxxxxxxx > > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx> > > > --- > > > kernel/fork.c | 1 + > > > 1 file changed, 1 insertion(+) > > > > > > diff --git a/kernel/fork.c b/kernel/fork.c > > > index b85814e614a5..d2e12b6d2b18 100644 > > > --- a/kernel/fork.c > > > +++ b/kernel/fork.c > > > @@ -686,6 +686,7 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm, > > > for_each_vma(old_vmi, mpnt) { > > > struct file *file; > > > > > > + vma_start_write(mpnt); > > > if (mpnt->vm_flags & VM_DONTCOPY) { > > > vm_stat_account(mm, mpnt->vm_flags, -vma_pages(mpnt)); > > > continue; > > > > After the mmap_write_lock_killable(), there will still be a period where page > > faults can happen. Essentially, page faults can happen for a VMA until we lock that VMA. > > > > I cannot immediately name something that is broken allowing for that, and this change > > should fix the issue at hand, but exotic things like > > > > flush_cache_dup_mm(oldmm); > > > > make me wonder if we really want to allow for that or if there is some other corner case > > in fork() handling that really doesn't expect concurrent page faults (and, thereby, page > > table modifications) with fork. > > > > For example, documentation/core-api/cachetlb.rst says > > > > 2) ``void flush_cache_dup_mm(struct mm_struct *mm)`` > > > > This interface flushes an entire user address space from > > the caches. That is, after running, there will be no cache > > lines associated with 'mm'. > > > > This interface is used to handle whole address space > > page table operations such as what happens during fork. > > > > This option is separate from flush_cache_mm to allow some > > optimizations for VIPT caches. > > > > I see. So, we really need to lock all VMAs before > flush_cache_dup_mm(). Makes sense. I'll post an update to this patch > shortly. v3 of the patchset with this fix is posted at https://lore.kernel.org/all/20230705171213.2843068-1-surenb@xxxxxxxxxx/ > Thanks, > Suren. > > > > > An alternative that requires another VMA walk would be > > > > diff --git a/kernel/fork.c b/kernel/fork.c > > index 41c964104b58..0f182d3f049b 100644 > > --- a/kernel/fork.c > > +++ b/kernel/fork.c > > @@ -662,6 +662,13 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm, > > retval = -EINTR; > > goto fail_uprobe_end; > > } > > + > > + /* Disallow any page faults early by locking all VMAs. */ > > + if (IS_ENABLED(CONFIG_PER_VMA_LOCK)) { > > + for_each_vma(old_vmi, mpnt) > > + vma_start_write(mpnt); > > + vma_iter_init(old_vmi, old_mm, 0); > > + } > > flush_cache_dup_mm(oldmm); > > uprobe_dup_mmap(oldmm, mm); > > /* > > -- > > 2.41.0 > > > > > > Unless there are other thoughts, I guess you change is fine regarding the problem > > at hand. Not so sure regarding any other corner cases, that's why I'm spelling it out. > > > > > > Acked-by: David Hildenbrand <david@xxxxxxxxxx> > > > > -- > > Cheers, > > > > David / dhildenb > >
