X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fvdev_disk.c;h=aba3c4ab57ef284df3641bb3661312eef8cb2419;hb=4c0d8e50b99b4f3b4a9b7bc67ac7fc4e406f5755;hp=51062ef23529d3489e954b7a9d1f270603b37803;hpb=a69052be7f9a4008e2b09578e9db5fdebc186111;p=zfs.git

diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c
index 51062ef..aba3c4a 100644
--- a/module/zfs/vdev_disk.c
+++ b/module/zfs/vdev_disk.c
@@ -33,6 +33,8 @@
 #include <sys/zio.h>
 #include <sys/sunldi.h>
 
+char *zfs_vdev_scheduler = VDEV_SCHEDULER;
+
 /*
  * Virtual device vector for disks.
  */
@@ -102,6 +104,43 @@ vdev_disk_error(zio_t *zio)
 #endif
 }
 
+/*
+ * Use the Linux 'noop' elevator for zfs managed block devices.  This
+ * strikes the ideal balance by allowing the zfs elevator to do all
+ * request ordering and prioritization.  While allowing the Linux
+ * elevator to do the maximum front/back merging allowed by the
+ * physical device.  This yields the largest possible requests for
+ * the device with the lowest total overhead.
+ *
+ * Unfortunately we cannot directly call the elevator_switch() function
+ * because it is not exported from the block layer.  This means we have
+ * to use the sysfs interface and a user space upcall.  Pools will be
+ * automatically imported on module load so we must do this at device
+ * open time from the kernel.
+ */
+static int
+vdev_elevator_switch(vdev_t *v, char *elevator, char *device)
+{
+	char sh_path[] = "/bin/sh";
+	char sh_cmd[128];
+	char *argv[] = { sh_path, "-c", sh_cmd };
+	char *envp[] = { NULL };
+	int error;
+
+	if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4))
+		return (0);
+
+	sprintf(sh_cmd, "%s \"%s\" >/sys/block/%s/queue/scheduler",
+	    "/bin/echo", elevator, device);
+
+	error = call_usermodehelper(sh_path, argv, envp, 1);
+	if (error)
+		printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
+		    elevator, v->vdev_path, device, error);
+
+	return (error);
+}
+
 static int
 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
 {
@@ -167,6 +206,10 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
 	/* Based on the minimum sector size set the block size */
 	*ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
 
+	/* Try to set the io scheduler elevator algorithm */
+	(void) vdev_elevator_switch(v, zfs_vdev_scheduler,
+	    bdev->bd_disk->disk_name);
+
 	return 0;
 }
 
@@ -220,6 +263,27 @@ vdev_disk_dio_free(dio_request_t *dr)
 	          sizeof(struct bio *) * dr->dr_bio_count);
 }
 
+static int
+vdev_disk_dio_is_sync(dio_request_t *dr)
+{
+#ifdef HAVE_BIO_RW_SYNC
+	/* BIO_RW_SYNC preferred interface from 2.6.12-2.6.29 */
+        return (dr->dr_rw & (1 << BIO_RW_SYNC));
+#else
+# ifdef HAVE_BIO_RW_SYNCIO
+	/* BIO_RW_SYNCIO preferred interface from 2.6.30-2.6.35 */
+        return (dr->dr_rw & (1 << BIO_RW_SYNCIO));
+# else
+#  ifdef HAVE_REQ_SYNC
+	/* REQ_SYNC preferred interface from 2.6.36-2.6.xx */
+        return (dr->dr_rw & REQ_SYNC);
+#  else
+#   error "Unable to determine bio sync flag"
+#  endif /* HAVE_REQ_SYNC */
+# endif /* HAVE_BIO_RW_SYNC */
+#endif /* HAVE_BIO_RW_SYNCIO */
+}
+
 static void
 vdev_disk_dio_get(dio_request_t *dr)
 {
@@ -284,7 +348,7 @@ BIO_END_IO_PROTO(vdev_disk_physio_completion, bio, size, error)
 	rc = vdev_disk_dio_put(dr);
 
 	/* Wake up synchronous waiter this is the last outstanding bio */
-	if ((rc == 1) && (dr->dr_rw & (1 << DIO_RW_SYNCIO)))
+	if ((rc == 1) && vdev_disk_dio_is_sync(dr))
 		complete(&dr->dr_comp);
 
 	BIO_END_IO_RETURN(0);
@@ -339,6 +403,8 @@ __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr,
 	int bio_size, bio_count = 16;
 	int i = 0, error = 0, block_size;
 
+	ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size);
+
 retry:
 	dr = vdev_disk_dio_alloc(bio_count);
 	if (dr == NULL)
@@ -421,7 +487,7 @@ retry:
 	 * only synchronous consumer is vdev_disk_read_rootlabel() all other
 	 * IO originating from vdev_disk_io_start() is asynchronous.
 	 */
-	if (dr->dr_rw & (1 << DIO_RW_SYNCIO)) {
+	if (vdev_disk_dio_is_sync(dr)) {
 		wait_for_completion(&dr->dr_comp);
 		error = dr->dr_error;
 		ASSERT3S(atomic_read(&dr->dr_ref), ==, 1);
@@ -679,3 +745,6 @@ vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
 
 	return 0;
 }
+
+module_param(zfs_vdev_scheduler, charp, 0644);
+MODULE_PARM_DESC(zfs_vdev_scheduler, "IO Scheduler (noop)");