[PATCH v6 6/6] mm: handle userfaults under VMA lock

[Suren Baghdasaryan <surenb@google.com>]

Enable handle_userfault to operate under VMA lock by releasing VMA lock
instead of mmap_lock and retrying. Note that FAULT_FLAG_RETRY_NOWAIT
should never be used when handling faults under per-VMA lock protection
because that would break the assumption that lock is dropped on retry.

Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Acked-by: Peter Xu <peterx@redhat.com>
---
 fs/userfaultfd.c   | 34 ++++++++++++++--------------------
 include/linux/mm.h | 24 ++++++++++++++++++++++++
 mm/memory.c        |  9 ---------
 3 files changed, 38 insertions(+), 29 deletions(-)

diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 7cecd49e078b..21a546eaf9f7 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -277,17 +277,16 @@ static inline struct uffd_msg userfault_msg(unsigned long address,
  * hugepmd ranges.
  */
 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
-					 struct vm_area_struct *vma,
-					 unsigned long address,
-					 unsigned long flags,
-					 unsigned long reason)
+					      struct vm_fault *vmf,
+					      unsigned long reason)
 {
+	struct vm_area_struct *vma = vmf->vma;
 	pte_t *ptep, pte;
 	bool ret = true;
 
-	mmap_assert_locked(ctx->mm);
+	assert_fault_locked(vmf);
 
-	ptep = hugetlb_walk(vma, address, vma_mmu_pagesize(vma));
+	ptep = hugetlb_walk(vma, vmf->address, vma_mmu_pagesize(vma));
 	if (!ptep)
 		goto out;
 
@@ -308,10 +307,8 @@ static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 }
 #else
 static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
-					 struct vm_area_struct *vma,
-					 unsigned long address,
-					 unsigned long flags,
-					 unsigned long reason)
+					      struct vm_fault *vmf,
+					      unsigned long reason)
 {
 	return false;	/* should never get here */
 }
@@ -325,11 +322,11 @@ static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
  * threads.
  */
 static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
-					 unsigned long address,
-					 unsigned long flags,
+					 struct vm_fault *vmf,
 					 unsigned long reason)
 {
 	struct mm_struct *mm = ctx->mm;
+	unsigned long address = vmf->address;
 	pgd_t *pgd;
 	p4d_t *p4d;
 	pud_t *pud;
@@ -338,7 +335,7 @@ static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx,
 	pte_t ptent;
 	bool ret = true;
 
-	mmap_assert_locked(mm);
+	assert_fault_locked(vmf);
 
 	pgd = pgd_offset(mm, address);
 	if (!pgd_present(*pgd))
@@ -440,7 +437,7 @@ vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 	 * Coredumping runs without mmap_lock so we can only check that
 	 * the mmap_lock is held, if PF_DUMPCORE was not set.
 	 */
-	mmap_assert_locked(mm);
+	assert_fault_locked(vmf);
 
 	ctx = vma->vm_userfaultfd_ctx.ctx;
 	if (!ctx)
@@ -556,15 +553,12 @@ vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason)
 	spin_unlock_irq(&ctx->fault_pending_wqh.lock);
 
 	if (!is_vm_hugetlb_page(vma))
-		must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags,
-						  reason);
+		must_wait = userfaultfd_must_wait(ctx, vmf, reason);
 	else
-		must_wait = userfaultfd_huge_must_wait(ctx, vma,
-						       vmf->address,
-						       vmf->flags, reason);
+		must_wait = userfaultfd_huge_must_wait(ctx, vmf, reason);
 	if (is_vm_hugetlb_page(vma))
 		hugetlb_vma_unlock_read(vma);
-	mmap_read_unlock(mm);
+	release_fault_lock(vmf);
 
 	if (likely(must_wait && !READ_ONCE(ctx->released))) {
 		wake_up_poll(&ctx->fd_wqh, EPOLLIN);
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 54ab11214f4f..2794225b2d42 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -705,6 +705,17 @@ static inline bool vma_try_start_write(struct vm_area_struct *vma)
 	return true;
 }
 
+static inline void vma_assert_locked(struct vm_area_struct *vma)
+{
+	int mm_lock_seq;
+
+	if (__is_vma_write_locked(vma, &mm_lock_seq))
+		return;
+
+	lockdep_assert_held(&vma->vm_lock->lock);
+	VM_BUG_ON_VMA(!rwsem_is_locked(&vma->vm_lock->lock), vma);
+}
+
 static inline void vma_assert_write_locked(struct vm_area_struct *vma)
 {
 	int mm_lock_seq;
@@ -728,6 +739,14 @@ static inline void release_fault_lock(struct vm_fault *vmf)
 		mmap_read_unlock(vmf->vma->vm_mm);
 }
 
+static inline void assert_fault_locked(struct vm_fault *vmf)
+{
+	if (vmf->flags & FAULT_FLAG_VMA_LOCK)
+		vma_assert_locked(vmf->vma);
+	else
+		mmap_assert_locked(vmf->vma->vm_mm);
+}
+
 struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
 					  unsigned long address);
 
@@ -748,6 +767,11 @@ static inline void release_fault_lock(struct vm_fault *vmf)
 	mmap_read_unlock(vmf->vma->vm_mm);
 }
 
+static inline void assert_fault_locked(struct vm_fault *vmf)
+{
+	mmap_assert_locked(vmf->vma->vm_mm);
+}
+
 #endif /* CONFIG_PER_VMA_LOCK */
 
 /*
diff --git a/mm/memory.c b/mm/memory.c
index bb0f68a73b0c..d9f36f9392a9 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -5407,15 +5407,6 @@ struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
 	if (!vma_start_read(vma))
 		goto inval;
 
-	/*
-	 * Due to the possibility of userfault handler dropping mmap_lock, avoid
-	 * it for now and fall back to page fault handling under mmap_lock.
-	 */
-	if (userfaultfd_armed(vma)) {
-		vma_end_read(vma);
-		goto inval;
-	}
-
 	/* Check since vm_start/vm_end might change before we lock the VMA */
 	if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
 		vma_end_read(vma);
-- 
2.41.0.255.g8b1d071c50-goog