X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fvdev_disk.c;h=cbf0a4d4fb0896c74f34ceeb4899d2f91675eea5;hb=36f86f73f68548f46eb3229c8adf583d59fa9988;hp=86a089d12ddcfaa7c76bfbb5eb2b7a86a80f0bac;hpb=675de5aa379bfceb0258a20c4ccd1414e4798ee6;p=zfs.git diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c index 86a089d..cbf0a4d 100644 --- a/module/zfs/vdev_disk.c +++ b/module/zfs/vdev_disk.c @@ -33,6 +33,8 @@ #include #include +char *zfs_vdev_scheduler = VDEV_SCHEDULER; + /* * Virtual device vector for disks. */ @@ -85,10 +87,10 @@ bdev_capacity(struct block_device *bdev) /* The partition capacity referenced by the block device */ if (part) - return part->nr_sects; + return (part->nr_sects << 9); /* Otherwise assume the full device capacity */ - return get_capacity(bdev->bd_disk); + return (get_capacity(bdev->bd_disk) << 9); } static void @@ -102,10 +104,142 @@ 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. + */ static int -vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift) +vdev_elevator_switch(vdev_t *v, char *elevator) { - struct block_device *bdev; + 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; + int error; + + /* + * 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) */ + 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); + +#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/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); + + return (error); +} + +/* + * Expanding a whole disk vdev involves invoking BLKRRPART on the + * whole disk device. This poses a problem, because BLKRRPART will + * return EBUSY if one of the disk's partitions is open. That's why + * we have to do it here, just before opening the data partition. + * Unfortunately, BLKRRPART works by dropping all partitions and + * recreating them, which means that for a short time window, all + * /dev/sdxN device files disappear (until udev recreates them). + * This means two things: + * - When we open the data partition just after a BLKRRPART, we + * can't do it using the normal device file path because of the + * obvious race condition with udev. Instead, we use reliable + * kernel APIs to get a handle to the new partition device from + * the whole disk device. + * - Because vdev_disk_open() initially needs to find the device + * using its path, multiple vdev_disk_open() invocations in + * short succession on the same disk with BLKRRPARTs in the + * middle have a high probability of failure (because of the + * race condition with udev). A typical situation where this + * might happen is when the zpool userspace tool does a + * TRYIMPORT immediately followed by an IMPORT. For this + * reason, we only invoke BLKRRPART in the module when strictly + * necessary (zpool online -e case), and rely on userspace to + * do it when possible. + */ +static struct block_device * +vdev_disk_rrpart(const char *path, int mode, vdev_disk_t *vd) +{ +#if defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) + struct block_device *bdev, *result = ERR_PTR(-ENXIO); + struct gendisk *disk; + int error, partno; + + bdev = vdev_bdev_open(path, vdev_bdev_mode(mode), vd); + if (IS_ERR(bdev)) + return bdev; + + disk = get_gendisk(bdev->bd_dev, &partno); + vdev_bdev_close(bdev, vdev_bdev_mode(mode)); + + if (disk) { + bdev = bdget(disk_devt(disk)); + if (bdev) { + error = blkdev_get(bdev, vdev_bdev_mode(mode), vd); + if (error == 0) + error = ioctl_by_bdev(bdev, BLKRRPART, 0); + vdev_bdev_close(bdev, vdev_bdev_mode(mode)); + } + + bdev = bdget_disk(disk, partno); + if (bdev) { + error = blkdev_get(bdev, + vdev_bdev_mode(mode) | FMODE_EXCL, vd); + if (error == 0) + result = bdev; + } + put_disk(disk); + } + + return result; +#else + return ERR_PTR(-EOPNOTSUPP); +#endif /* defined(HAVE_3ARG_BLKDEV_GET) && defined(HAVE_GET_GENDISK) */ +} + +static int +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; int mode, block_size; @@ -115,7 +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; @@ -134,7 +268,10 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift) * level vdev validation. */ mode = spa_mode(v->vdev_spa); - bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd); + if (v->vdev_wholedisk && v->vdev_expanding) + bdev = vdev_disk_rrpart(v->vdev_path, mode, vd); + if (IS_ERR(bdev)) + bdev = vdev_bdev_open(v->vdev_path, vdev_bdev_mode(mode), vd); if (IS_ERR(bdev)) { kmem_free(vd, sizeof(vdev_disk_t)); return -PTR_ERR(bdev); @@ -162,11 +299,17 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *ashift) v->vdev_nowritecache = B_FALSE; /* Physical volume size in bytes */ - *psize = bdev_capacity(bdev) * block_size; + *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; + /* Try to set the io scheduler elevator algorithm */ + (void) vdev_elevator_switch(v, zfs_vdev_scheduler); + return 0; } @@ -193,7 +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); @@ -358,7 +501,9 @@ __vdev_disk_physio(struct block_device *bdev, zio_t *zio, caddr_t kbuf_ptr, caddr_t bio_ptr; uint64_t bio_offset; int bio_size, bio_count = 16; - int i = 0, error = 0, block_size; + int i = 0, error = 0; + + ASSERT3U(kbuf_offset + kbuf_size, <=, bdev->bd_inode->i_size); retry: dr = vdev_disk_dio_alloc(bio_count); @@ -370,7 +515,6 @@ retry: dr->dr_zio = zio; dr->dr_rw = flags; - block_size = vdev_bdev_block_size(bdev); /* * When the IO size exceeds the maximum bio size for the request @@ -396,7 +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; } @@ -411,7 +554,7 @@ retry: vdev_disk_dio_get(dr); dr->dr_bio[i]->bi_bdev = bdev; - dr->dr_bio[i]->bi_sector = bio_offset / block_size; + dr->dr_bio[i]->bi_sector = bio_offset >> 9; dr->dr_bio[i]->bi_rw = dr->dr_rw; dr->dr_bio[i]->bi_end_io = vdev_disk_physio_completion; dr->dr_bio[i]->bi_private = dr; @@ -499,7 +642,7 @@ vdev_disk_io_flush(struct block_device *bdev, zio_t *zio) bio->bi_private = zio; bio->bi_bdev = bdev; zio->io_delay = jiffies_64; - submit_bio(WRITE_BARRIER, bio); + submit_bio(VDEV_WRITE_FLUSH_FUA, bio); return 0; } @@ -654,14 +797,14 @@ vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config) if (IS_ERR(bdev)) return -PTR_ERR(bdev); - s = bdev_capacity(bdev) * vdev_bdev_block_size(bdev); + s = bdev_capacity(bdev); if (s == 0) { vdev_bdev_close(bdev, vdev_bdev_mode(FREAD)); return EIO; } 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; @@ -700,3 +843,6 @@ 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, "I/O scheduler");