fork: lock VMAs of the parent process when forking

Patch series "Avoid memory corruption caused by per-VMA locks", v4.

A memory corruption was reported in [1] with bisection pointing to the
patch [2] enabling per-VMA locks for x86.  Based on the reproducer
provided in [1] we suspect this is caused by the lack of VMA locking while
forking a child process.

Patch 1/2 in the series implements proper VMA locking during fork.  I
tested the fix locally using the reproducer and was unable to reproduce
the memory corruption problem.

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 ~7% 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.

Patch 2/2 disables per-VMA locks until the fix is tested and verified.


This patch (of 2):

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 ~7%
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.

Link: https://lkml.kernel.org/r/[email protected]
Link: https://lkml.kernel.org/r/[email protected]
Fixes: 0bff0aa ("x86/mm: try VMA lock-based page fault handling first")
Signed-off-by: Suren Baghdasaryan <[email protected]>
Suggested-by: David Hildenbrand <[email protected]>
Reported-by: Jiri Slaby <[email protected]>
Closes: https://lore.kernel.org/all/[email protected]/
Reported-by: Holger Hoffstätte <[email protected]>
Closes: https://lore.kernel.org/all/[email protected]/
Reported-by: Jacob Young <[email protected]>
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=3D217624
Reviewed-by: Liam R. Howlett <[email protected]>
Acked-by: David Hildenbrand <[email protected]>
Tested-by: Holger Hoffsttte <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

kernel/fork.c

@@ -662,6 +662,12 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
 		retval = -EINTR;
 		goto fail_uprobe_end;
 	}
+#ifdef CONFIG_PER_VMA_LOCK
+	/* Disallow any page faults before calling flush_cache_dup_mm */
+	for_each_vma(old_vmi, mpnt)
+		vma_start_write(mpnt);
+	vma_iter_set(&old_vmi, 0);
+#endif
 	flush_cache_dup_mm(oldmm);
 	uprobe_dup_mmap(oldmm, mm);
 	/*