[PATCH v2 2/6] mm: handle swap page faults under VMA lock if page is uncontended

[Suren Baghdasaryan]

When page fault is handled under VMA lock protection, all swap page
faults are retried with mmap_lock because folio_lock_or_retry
implementation has to drop and reacquire mmap_lock if folio could
not be immediately locked.
Instead of retrying all swapped page faults, retry only when folio
locking fails.
Note that the only time do_swap_page calls synchronous swap_readpage
is when SWP_SYNCHRONOUS_IO is set, which is only set for
QUEUE_FLAG_SYNCHRONOUS devices: brd, zram and nvdimms (both btt and
pmem). Therefore we don't sleep in this path, and there's no need to
drop the mmap or per-vma lock.
Drivers implementing ops->migrate_to_ram might still rely on mmap_lock,
therefore fall back to mmap_lock in this case.

Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
---
 mm/filemap.c |  6 ++++++
 mm/memory.c  | 14 +++++++++-----
 2 files changed, 15 insertions(+), 5 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index b4c9bd368b7e..7cb0a3776a07 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1706,6 +1706,8 @@ static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
  *     mmap_lock has been released (mmap_read_unlock(), unless flags had both
  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
  *     which case mmap_lock is still held.
+ *     If flags had FAULT_FLAG_VMA_LOCK set, meaning the operation is performed
+ *     with VMA lock only, the VMA lock is still held.
  *
  * If neither ALLOW_RETRY nor KILLABLE are set, will always return true
  * with the folio locked and the mmap_lock unperturbed.
@@ -1713,6 +1715,10 @@ static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait)
 bool __folio_lock_or_retry(struct folio *folio, struct mm_struct *mm,
 			 unsigned int flags)
 {
+	/* Can't do this if not holding mmap_lock */
+	if (flags & FAULT_FLAG_VMA_LOCK)
+		return false;
+
 	if (fault_flag_allow_retry_first(flags)) {
 		/*
 		 * CAUTION! In this case, mmap_lock is not released
diff --git a/mm/memory.c b/mm/memory.c
index f69fbc251198..41f45819a923 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3711,11 +3711,6 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
 	if (!pte_unmap_same(vmf))
 		goto out;
 
-	if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
-		ret = VM_FAULT_RETRY;
-		goto out;
-	}
-
 	entry = pte_to_swp_entry(vmf->orig_pte);
 	if (unlikely(non_swap_entry(entry))) {
 		if (is_migration_entry(entry)) {
@@ -3725,6 +3720,15 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
 			vmf->page = pfn_swap_entry_to_page(entry);
 			ret = remove_device_exclusive_entry(vmf);
 		} else if (is_device_private_entry(entry)) {
+			if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
+				/*
+				 * migrate_to_ram is not yet ready to operate
+				 * under VMA lock.
+				 */
+				ret |= VM_FAULT_RETRY;
+				goto out;
+			}
+
 			vmf->page = pfn_swap_entry_to_page(entry);
 			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
 					vmf->address, &vmf->ptl);
-- 
2.41.0.162.gfafddb0af9-goog