Illumos #1948: zpool list should show more detailed pool info

[mirror_zfs.git] / module / zfs / vdev_disk.c

diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c
index 28a4861abd7260c2c1feb7246fbcfe5d64fd71c4..ffb2980d28758bbc37cea17d17656f1619da77bf 100644 (file)
--- a/module/zfs/vdev_disk.c
+++ b/module/zfs/vdev_disk.c
@@ -158,10 +158,76 @@ vdev_elevator_switch(vdev_t *v, char *elevator)
        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 *ashift)
+vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
+    uint64_t *ashift)
 {
-       struct block_device *bdev;
+       struct block_device *bdev = ERR_PTR(-ENXIO);
        vdev_disk_t *vd;
        int mode, block_size;
 
@@ -171,7 +237,7 @@ 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;
 
@@ -190,7 +256,10 @@ 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);
@@ -220,6 +289,9 @@ 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;
 
@@ -252,7 +324,7 @@ 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);
@@ -721,7 +793,7 @@ 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;