X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fvdev_disk.c;h=5cb3fe279c9de21c01f0f1f4c5c6a35f1429fe3f;hb=40a39e1103bbbd6908bc1b5cf4af953c7e4706ad;hp=86a089d12ddcfaa7c76bfbb5eb2b7a86a80f0bac;hpb=675de5aa379bfceb0258a20c4ccd1414e4798ee6;p=zfs.git

diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c
index 86a089d..5cb3fe2 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 +104,60 @@ 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.
+ */
+#define SET_SCHEDULER_CMD \
+	"exec 0</dev/null " \
+	"     1>/sys/block/%s/queue/scheduler " \
+	"     2>/dev/null; " \
+	"echo %s"
+
+static int
+vdev_elevator_switch(vdev_t *v, char *elevator)
+{
+	vdev_disk_t *vd = v->vdev_tsd;
+	struct block_device *bdev = vd->vd_bdev;
+	struct request_queue *q = bdev_get_queue(bdev);
+	char *device = bdev->bd_disk->disk_name;
+	char *argv[] = { "/bin/sh", "-c", NULL, NULL };
+	char *envp[] = { NULL };
+	int error;
+
+	/* Skip devices which are not whole disks (partitions) */
+	if (!v->vdev_wholedisk)
+		return (0);
+
+	/* Skip devices without schedulers (loop, ram, dm, etc) */
+	if (!q->elevator || !blk_queue_stackable(q))
+		return (0);
+
+	/* Leave existing scheduler when set to "none" */
+	if (!strncmp(elevator, "none", 4) && (strlen(elevator) == 4))
+		return (0);
+
+	argv[2] = kmem_asprintf(SET_SCHEDULER_CMD, device, elevator);
+	error = call_usermodehelper(argv[0], argv, envp, 1);
+	if (error)
+		printk("ZFS: Unable to set \"%s\" scheduler for %s (%s): %d\n",
+		       elevator, v->vdev_path, device, error);
+
+	strfree(argv[2]);
+
+	return (error);
+}
+
 static int
 vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
 {
@@ -167,6 +223,9 @@ 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);
+
 	return 0;
 }
 
@@ -360,6 +419,8 @@ __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)
@@ -700,3 +761,6 @@ 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)");