svn commit: r331072 - user/markj/vm-playground/sys/kern

[Mark Johnston]

Author: markj
Date: Fri Mar 16 20:04:52 2018
New Revision: 331072
URL: https://svnweb.freebsd.org/changeset/base/331072

Log:
  Elide the object lock in vfs_vmio_truncate() in the common case.
  
  Rather attempting to free invalid pages, accelerate their reclamation
  by moving them near the head of the inactive queue. With this change,
  bufspace threads don't touch the object lock.

Modified:
  user/markj/vm-playground/sys/kern/vfs_bio.c

Modified: user/markj/vm-playground/sys/kern/vfs_bio.c
==============================================================================
--- user/markj/vm-playground/sys/kern/vfs_bio.c	Fri Mar 16 19:26:58 2018	(r331071)
+++ user/markj/vm-playground/sys/kern/vfs_bio.c	Fri Mar 16 20:04:52 2018	(r331072)
@@ -2915,24 +2915,14 @@ vfs_vmio_unwire(struct buf *bp, vm_page_t m)
 
 	vm_page_lock(m);
 	if (vm_page_unwire_noq(m)) {
-		/*
-		 * Determine if the page should be freed before adding
-		 * it to the inactive queue.
-		 */
-		if (m->valid == 0) {
-			freed = !vm_page_busied(m);
-			if (freed)
-				vm_page_free(m);
-		} else if ((bp->b_flags & B_DIRECT) != 0)
-			freed = vm_page_try_to_free(m);
-		else
-			freed = false;
-		if (!freed) {
+		if ((bp->b_flags & B_DIRECT) == 0 || !vm_page_try_to_free(m)) {
 			/*
-			 * If the page is unlikely to be reused, let the
-			 * VM know.  Otherwise, maintain LRU.
+			 * Use a racy check of the valid bits to determine
+			 * whether we can accelerate reclamation of the page.
+			 * This allows elision of the object write lock in the
+			 * common case.
 			 */
-			if ((bp->b_flags & B_NOREUSE) != 0)
+			if (m->valid == 0 || (bp->b_flags & B_NOREUSE) != 0)
 				vm_page_deactivate_noreuse(m);
 			else if (m->queue == PQ_ACTIVE)
 				vm_page_reference(m);
@@ -3021,7 +3011,12 @@ vfs_vmio_truncate(struct buf *bp, int desiredpages)
 		    (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages);
 	} else
 		BUF_CHECK_UNMAPPED(bp);
-	obj = bp->b_bufobj->bo_object;
+
+	/*
+	 * The object lock is needed only if we attempt to free the buffer's
+	 * pages.
+	 */
+	obj = (bp->b_flags & B_DIRECT) != 0 ? bp->b_bufobj->bo_object : NULL;
 	if (obj != NULL)
 		VM_OBJECT_WLOCK(obj);
 	for (i = desiredpages; i < bp->b_npages; i++) {