Pull request openzfs#4858

config/kernel-blk-queue-unplug.m4

@@ -0,0 +1,23 @@
+dnl #
+dnl # 2.6.32-2.6.35 API - The BIO_RW_UNPLUG enum can be used as a hint
+dnl # to unplug the queue.
+dnl #
+AC_DEFUN([ZFS_AC_KERNEL_BLK_QUEUE_HAVE_BIO_RW_UNPLUG], [
+	AC_MSG_CHECKING([whether the BIO_RW_UNPLUG enum is available])
+	tmp_flags="$EXTRA_KCFLAGS"
+	EXTRA_KCFLAGS="${NO_UNUSED_BUT_SET_VARIABLE}"
+	ZFS_LINUX_TRY_COMPILE([
+		#include <linux/blkdev.h>
+	],[
+		extern enum bio_rw_flags rw;
+
+		rw = BIO_RW_UNPLUG;
+	],[
+		AC_MSG_RESULT(yes)
+		AC_DEFINE(HAVE_BLK_QUEUE_HAVE_BIO_RW_UNPLUG, 1,
+		          [BIO_RW_UNPLUG is available])
+	],[
+		AC_MSG_RESULT(no)
+	])
+	EXTRA_KCFLAGS="$tmp_flags"
+])

config/kernel.m4

@@ -28,6 +28,7 @@ AC_DEFUN([ZFS_AC_CONFIG_KERNEL], [
 	ZFS_AC_KERNEL_BLK_QUEUE_FLUSH
 	ZFS_AC_KERNEL_BLK_QUEUE_MAX_HW_SECTORS
 	ZFS_AC_KERNEL_BLK_QUEUE_MAX_SEGMENTS
+	ZFS_AC_KERNEL_BLK_QUEUE_HAVE_BIO_RW_UNPLUG
 	ZFS_AC_KERNEL_GET_DISK_RO
 	ZFS_AC_KERNEL_GET_GENDISK
 	ZFS_AC_KERNEL_DISCARD_GRANULARITY

module/zfs/spa.c

@@ -6767,7 +6767,6 @@ EXPORT_SYMBOL(spa_open);
 EXPORT_SYMBOL(spa_open_rewind);
 EXPORT_SYMBOL(spa_get_stats);
 EXPORT_SYMBOL(spa_create);
-EXPORT_SYMBOL(spa_import_rootpool);
 EXPORT_SYMBOL(spa_import);
 EXPORT_SYMBOL(spa_tryimport);
 EXPORT_SYMBOL(spa_destroy);

module/zfs/vdev_disk.c

@@ -43,10 +43,8 @@ static void *zfs_vdev_holder = VDEV_HOLDER;
  * Virtual device vector for disks.
  */
 typedef struct dio_request {
-	struct completion	dr_comp;	/* Completion for sync IO */
 	zio_t			*dr_zio;	/* Parent ZIO */
 	atomic_t		dr_ref;		/* References */
-	int			dr_wait;	/* Wait for IO */
 	int			dr_error;	/* Bio error */
 	int			dr_bio_count;	/* Count of bio's */
 	struct bio		*dr_bio[0];	/* Attached bio's */
@@ -365,7 +363,6 @@ vdev_disk_dio_alloc(int bio_count)
 	dr = kmem_zalloc(sizeof (dio_request_t) +
 	    sizeof (struct bio *) * bio_count, KM_SLEEP);
 	if (dr) {
-		init_completion(&dr->dr_comp);
 		atomic_set(&dr->dr_ref, 0);
 		dr->dr_bio_count = bio_count;
 		dr->dr_error = 0;
@@ -428,7 +425,6 @@ BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, error)
 {
 	dio_request_t *dr = bio->bi_private;
 	int rc;
-	int wait;
 
 	if (dr->dr_error == 0) {
 #ifdef HAVE_1ARG_BIO_END_IO_T
@@ -441,13 +437,8 @@ BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, error)
 #endif
 	}
 
-	wait = dr->dr_wait;
 	/* Drop reference aquired by __vdev_disk_physio */
 	rc = vdev_disk_dio_put(dr);
-
-	/* Wake up synchronous waiter this is the last outstanding bio */
-	if (wait && rc == 1)
-		complete(&dr->dr_comp);
 }
 
 static inline unsigned long
@@ -514,7 +505,7 @@ vdev_submit_bio(int rw, struct bio *bio)
 
 static int
 __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
-    size_t kbuf_size, uint64_t kbuf_offset, int flags, int wait)
+    size_t kbuf_size, uint64_t kbuf_offset, int flags)
 {
 	dio_request_t *dr;
 	uint64_t zio_offset;
@@ -537,7 +528,6 @@ __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
 
 	rw = flags;
 	dr->dr_zio = zio;
-	dr->dr_wait = wait;
 
 	/*
 	 * When the IO size exceeds the maximum bio size for the request
@@ -612,20 +602,6 @@ __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
 		if (dr->dr_bio[i])
 			vdev_submit_bio(rw, dr->dr_bio[i]);
 
-	/*
-	 * On synchronous blocking requests we wait for all bio the completion
-	 * callbacks to run.  We will be woken when the last callback runs
-	 * for this dio.  We are responsible for putting the last dio_request
-	 * reference will in turn put back the last bio references.  The
-	 * only synchronous consumer is vdev_disk_read_rootlabel() all other
-	 * IO originating from vdev_disk_io_start() is asynchronous.
-	 */
-	if (wait) {
-		wait_for_completion(&dr->dr_comp);
-		error = dr->dr_error;
-		ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
-	}
-
 	if (kbuf_ptr)
 		abd_put(zio_data);
 
@@ -639,7 +615,7 @@ vdev_disk_physio(struct block_device *bdev, caddr_t kbuf,
     size_t size, uint64_t offset, int flags)
 {
 	bio_set_flags_failfast(bdev, &flags);
-	return (__vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags, 1));
+	return (__vdev_disk_physio(bdev, NULL, kbuf, size, offset, flags));
 }
 
 BIO_END_IO_PROTO(vdev_disk_io_flush_completion, bio, rc)
@@ -691,7 +667,6 @@ vdev_disk_io_start(zio_t *zio)
 {
 	vdev_t *v = zio->io_vd;
 	vdev_disk_t *vd = v->vdev_tsd;
-	zio_priority_t pri = zio->io_priority;
 	int flags, error;
 
 	switch (zio->io_type) {
@@ -731,17 +706,23 @@ vdev_disk_io_start(zio_t *zio)
 		zio_execute(zio);
 		return;
 	case ZIO_TYPE_WRITE:
-		if ((pri == ZIO_PRIORITY_SYNC_WRITE) && (v->vdev_nonrot))
-			flags = WRITE_SYNC;
-		else
-			flags = WRITE;
+#if defined(HAVE_BLK_QUEUE_HAVE_BIO_RW_UNPLUG)
+		flags = WRITE | (1 << BIO_RW_UNPLUG);
+#elif defined(REQ_UNPLUG)
+		flags = WRITE | REQ_UNPLUG;
+#else
+		flags = WRITE;
+#endif
 		break;
 
 	case ZIO_TYPE_READ:
-		if ((pri == ZIO_PRIORITY_SYNC_READ) && (v->vdev_nonrot))
-			flags = READ_SYNC;
-		else
-			flags = READ;
+#if defined(HAVE_BLK_QUEUE_HAVE_BIO_RW_UNPLUG)
+		flags = READ | (1 << BIO_RW_UNPLUG);
+#elif defined(REQ_UNPLUG)
+		flags = READ | REQ_UNPLUG;
+#else
+		flags = READ;
+#endif
 		break;
 
 	default:
@@ -751,7 +732,7 @@ vdev_disk_io_start(zio_t *zio)
 	}
 
 	error = __vdev_disk_physio(vd->vd_bdev, zio, NULL,
-	    zio->io_size, zio->io_offset, flags, 0);
+	    zio->io_size, zio->io_offset, flags);
 	if (error) {
 		zio->io_error = error;
 		zio_interrupt(zio);