X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fvdev_disk.c;h=cbf0a4d4fb0896c74f34ceeb4899d2f91675eea5;hb=36f86f73f68548f46eb3229c8adf583d59fa9988;hp=eee03d08055d46291ac011b4118213b9c32446f9;hpb=b5a28807cdec3c05aa69cbe4689cd914dc94783a;p=zfs.git

diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c
index eee03d0..cbf0a4d 100644
--- a/module/zfs/vdev_disk.c
+++ b/module/zfs/vdev_disk.c
@@ -111,19 +111,7 @@ vdev_disk_error(zio_t *zio)
  * 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)
 {
@@ -131,12 +119,17 @@ vdev_elevator_switch(vdev_t *v, char *elevator)
 	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)
+	/*
+	 * Skip devices which are not whole disks (partitions).
+	 * Device-mapper devices are excepted since they may be whole
+	 * disks despite the vdev_wholedisk flag, in which case we can
+	 * and should switch the elevator. If the device-mapper device
+	 * does not have an elevator (i.e. dm-raid, dm-crypt, etc.) the
+	 * "Skip devices without schedulers" check below will fail.
+	 */
+	if (!v->vdev_wholedisk && strncmp(device, "dm-", 3) != 0)
 		return (0);
 
 	/* Skip devices without schedulers (loop, ram, dm, etc) */
@@ -147,14 +140,33 @@ vdev_elevator_switch(vdev_t *v, char *elevator)
 	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);
+#ifdef HAVE_ELEVATOR_CHANGE
+	error = elevator_change(q, elevator);
+#else
+	/* For pre-2.6.36 kernels elevator_change() is not available.
+	 * Therefore we fall back to using a usermodehelper to echo the
+	 * elevator into sysfs;  This requires /bin/echo and sysfs to be
+	 * mounted which may not be true early in the boot process.
+	 */
+# define SET_SCHEDULER_CMD \
+	"exec 0</dev/null " \
+	"     1>/sys/block/%s/queue/scheduler " \
+	"     2>/dev/null; " \
+	"echo %s"
+
+	{
+		char *argv[] = { "/bin/sh", "-c", NULL, NULL };
+		char *envp[] = { NULL };
+
+		argv[2] = kmem_asprintf(SET_SCHEDULER_CMD, device, elevator);
+		error = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
+		strfree(argv[2]);
+	}
+#endif /* HAVE_ELEVATOR_CHANGE */
 	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);
 }
 
@@ -224,7 +236,8 @@ vdev_disk_rrpart(const char *path, int mode, vdev_disk_t *vd)
 }
 
 static int
-vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
+vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
+    uint64_t *ashift)
 {
 	struct block_device *bdev = ERR_PTR(-ENXIO);
 	vdev_disk_t *vd;
@@ -236,7 +249,7 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
 		return EINVAL;
 	}
 
-	vd = kmem_zalloc(sizeof(vdev_disk_t), KM_SLEEP);
+	vd = kmem_zalloc(sizeof(vdev_disk_t), KM_PUSHPAGE);
 	if (vd == NULL)
 		return ENOMEM;
 
@@ -288,6 +301,9 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift)
 	/* Physical volume size in bytes */
 	*psize = bdev_capacity(bdev);
 
+	/* TODO: report possible expansion size */
+	*max_psize = *psize;
+
 	/* Based on the minimum sector size set the block size */
 	*ashift = highbit(MAX(block_size, SPA_MINBLOCKSIZE)) - 1;
 
@@ -320,7 +336,7 @@ vdev_disk_dio_alloc(int bio_count)
 	int i;
 
 	dr = kmem_zalloc(sizeof(dio_request_t) +
-	                 sizeof(struct bio *) * bio_count, KM_SLEEP);
+	                 sizeof(struct bio *) * bio_count, KM_PUSHPAGE);
 	if (dr) {
 		init_completion(&dr->dr_comp);
 		atomic_set(&dr->dr_ref, 0);
@@ -524,7 +540,6 @@ retry:
 		if (dr->dr_bio_count == i) {
 			vdev_disk_dio_free(dr);
 			bio_count *= 2;
-			printk("WARNING: Resized bio's/dio to %d\n",bio_count);
 			goto retry;
 		}
 
@@ -789,7 +804,7 @@ vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
 	}
 
 	size = P2ALIGN_TYPED(s, sizeof(vdev_label_t), uint64_t);
-	label = vmem_alloc(sizeof(vdev_label_t), KM_SLEEP);
+	label = vmem_alloc(sizeof(vdev_label_t), KM_PUSHPAGE);
 
 	for (i = 0; i < VDEV_LABELS; i++) {
 	        uint64_t offset, state, txg = 0;