These changes are the raw update to qemu-2.6.
[kvmfornfv.git] / qemu / block / mirror.c
index fc4d8f5..039f481 100644 (file)
  *
  */
 
+#include "qemu/osdep.h"
 #include "trace.h"
 #include "block/blockjob.h"
 #include "block/block_int.h"
+#include "sysemu/block-backend.h"
+#include "qapi/error.h"
 #include "qapi/qmp/qerror.h"
 #include "qemu/ratelimit.h"
 #include "qemu/bitmap.h"
+#include "qemu/error-report.h"
 
 #define SLICE_TIME    100000000ULL /* ns */
 #define MAX_IN_FLIGHT 16
@@ -44,7 +48,6 @@ typedef struct MirrorBlockJob {
     BlockdevOnError on_source_error, on_target_error;
     bool synced;
     bool should_complete;
-    int64_t sector_num;
     int64_t granularity;
     size_t buf_size;
     int64_t bdev_length;
@@ -60,6 +63,9 @@ typedef struct MirrorBlockJob {
     int sectors_in_flight;
     int ret;
     bool unmap;
+    bool waiting_for_io;
+    int target_cluster_sectors;
+    int max_iov;
 } MirrorBlockJob;
 
 typedef struct MirrorOp {
@@ -102,7 +108,7 @@ static void mirror_iteration_done(MirrorOp *op, int ret)
 
     sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
     chunk_num = op->sector_num / sectors_per_chunk;
-    nb_chunks = op->nb_sectors / sectors_per_chunk;
+    nb_chunks = DIV_ROUND_UP(op->nb_sectors, sectors_per_chunk);
     bitmap_clear(s->in_flight_bitmap, chunk_num, nb_chunks);
     if (ret >= 0) {
         if (s->cow_bitmap) {
@@ -112,13 +118,9 @@ static void mirror_iteration_done(MirrorOp *op, int ret)
     }
 
     qemu_iovec_destroy(&op->qiov);
-    g_slice_free(MirrorOp, op);
+    g_free(op);
 
-    /* Enter coroutine when it is not sleeping.  The coroutine sleeps to
-     * rate-limit itself.  The coroutine will eventually resume since there is
-     * a sleep timeout so don't wake it early.
-     */
-    if (s->common.busy) {
+    if (s->waiting_for_io) {
         qemu_coroutine_enter(s->common.co, NULL);
     }
 }
@@ -159,107 +161,97 @@ static void mirror_read_complete(void *opaque, int ret)
                     mirror_write_complete, op);
 }
 
-static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
+static inline void mirror_clip_sectors(MirrorBlockJob *s,
+                                       int64_t sector_num,
+                                       int *nb_sectors)
 {
-    BlockDriverState *source = s->common.bs;
-    int nb_sectors, sectors_per_chunk, nb_chunks;
-    int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector;
-    uint64_t delay_ns = 0;
-    MirrorOp *op;
-    int pnum;
-    int64_t ret;
-
-    s->sector_num = hbitmap_iter_next(&s->hbi);
-    if (s->sector_num < 0) {
-        bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
-        s->sector_num = hbitmap_iter_next(&s->hbi);
-        trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
-        assert(s->sector_num >= 0);
-    }
-
-    hbitmap_next_sector = s->sector_num;
-    sector_num = s->sector_num;
-    sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
-    end = s->bdev_length / BDRV_SECTOR_SIZE;
-
-    /* Extend the QEMUIOVector to include all adjacent blocks that will
-     * be copied in this operation.
-     *
-     * We have to do this if we have no backing file yet in the destination,
-     * and the cluster size is very large.  Then we need to do COW ourselves.
-     * The first time a cluster is copied, copy it entirely.  Note that,
-     * because both the granularity and the cluster size are powers of two,
-     * the number of sectors to copy cannot exceed one cluster.
-     *
-     * We also want to extend the QEMUIOVector to include more adjacent
-     * dirty blocks if possible, to limit the number of I/O operations and
-     * run efficiently even with a small granularity.
-     */
-    nb_chunks = 0;
-    nb_sectors = 0;
-    next_sector = sector_num;
-    next_chunk = sector_num / sectors_per_chunk;
+    *nb_sectors = MIN(*nb_sectors,
+                      s->bdev_length / BDRV_SECTOR_SIZE - sector_num);
+}
 
-    /* Wait for I/O to this cluster (from a previous iteration) to be done.  */
-    while (test_bit(next_chunk, s->in_flight_bitmap)) {
-        trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
-        qemu_coroutine_yield();
+/* Round sector_num and/or nb_sectors to target cluster if COW is needed, and
+ * return the offset of the adjusted tail sector against original. */
+static int mirror_cow_align(MirrorBlockJob *s,
+                            int64_t *sector_num,
+                            int *nb_sectors)
+{
+    bool need_cow;
+    int ret = 0;
+    int chunk_sectors = s->granularity >> BDRV_SECTOR_BITS;
+    int64_t align_sector_num = *sector_num;
+    int align_nb_sectors = *nb_sectors;
+    int max_sectors = chunk_sectors * s->max_iov;
+
+    need_cow = !test_bit(*sector_num / chunk_sectors, s->cow_bitmap);
+    need_cow |= !test_bit((*sector_num + *nb_sectors - 1) / chunk_sectors,
+                          s->cow_bitmap);
+    if (need_cow) {
+        bdrv_round_to_clusters(s->target, *sector_num, *nb_sectors,
+                               &align_sector_num, &align_nb_sectors);
+    }
+
+    if (align_nb_sectors > max_sectors) {
+        align_nb_sectors = max_sectors;
+        if (need_cow) {
+            align_nb_sectors = QEMU_ALIGN_DOWN(align_nb_sectors,
+                                               s->target_cluster_sectors);
+        }
     }
+    /* Clipping may result in align_nb_sectors unaligned to chunk boundary, but
+     * that doesn't matter because it's already the end of source image. */
+    mirror_clip_sectors(s, align_sector_num, &align_nb_sectors);
 
-    do {
-        int added_sectors, added_chunks;
-
-        if (!bdrv_get_dirty(source, s->dirty_bitmap, next_sector) ||
-            test_bit(next_chunk, s->in_flight_bitmap)) {
-            assert(nb_sectors > 0);
-            break;
-        }
+    ret = align_sector_num + align_nb_sectors - (*sector_num + *nb_sectors);
+    *sector_num = align_sector_num;
+    *nb_sectors = align_nb_sectors;
+    assert(ret >= 0);
+    return ret;
+}
 
-        added_sectors = sectors_per_chunk;
-        if (s->cow_bitmap && !test_bit(next_chunk, s->cow_bitmap)) {
-            bdrv_round_to_clusters(s->target,
-                                   next_sector, added_sectors,
-                                   &next_sector, &added_sectors);
+static inline void mirror_wait_for_io(MirrorBlockJob *s)
+{
+    assert(!s->waiting_for_io);
+    s->waiting_for_io = true;
+    qemu_coroutine_yield();
+    s->waiting_for_io = false;
+}
 
-            /* On the first iteration, the rounding may make us copy
-             * sectors before the first dirty one.
-             */
-            if (next_sector < sector_num) {
-                assert(nb_sectors == 0);
-                sector_num = next_sector;
-                next_chunk = next_sector / sectors_per_chunk;
-            }
-        }
+/* Submit async read while handling COW.
+ * Returns: nb_sectors if no alignment is necessary, or
+ *          (new_end - sector_num) if tail is rounded up or down due to
+ *          alignment or buffer limit.
+ */
+static int mirror_do_read(MirrorBlockJob *s, int64_t sector_num,
+                          int nb_sectors)
+{
+    BlockDriverState *source = s->common.bs;
+    int sectors_per_chunk, nb_chunks;
+    int ret = nb_sectors;
+    MirrorOp *op;
 
-        added_sectors = MIN(added_sectors, end - (sector_num + nb_sectors));
-        added_chunks = (added_sectors + sectors_per_chunk - 1) / sectors_per_chunk;
+    sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
 
-        /* When doing COW, it may happen that there is not enough space for
-         * a full cluster.  Wait if that is the case.
-         */
-        while (nb_chunks == 0 && s->buf_free_count < added_chunks) {
-            trace_mirror_yield_buf_busy(s, nb_chunks, s->in_flight);
-            qemu_coroutine_yield();
-        }
-        if (s->buf_free_count < nb_chunks + added_chunks) {
-            trace_mirror_break_buf_busy(s, nb_chunks, s->in_flight);
-            break;
-        }
+    /* We can only handle as much as buf_size at a time. */
+    nb_sectors = MIN(s->buf_size >> BDRV_SECTOR_BITS, nb_sectors);
+    assert(nb_sectors);
 
-        /* We have enough free space to copy these sectors.  */
-        bitmap_set(s->in_flight_bitmap, next_chunk, added_chunks);
+    if (s->cow_bitmap) {
+        ret += mirror_cow_align(s, &sector_num, &nb_sectors);
+    }
+    assert(nb_sectors << BDRV_SECTOR_BITS <= s->buf_size);
+    /* The sector range must meet granularity because:
+     * 1) Caller passes in aligned values;
+     * 2) mirror_cow_align is used only when target cluster is larger. */
+    assert(!(sector_num % sectors_per_chunk));
+    nb_chunks = DIV_ROUND_UP(nb_sectors, sectors_per_chunk);
 
-        nb_sectors += added_sectors;
-        nb_chunks += added_chunks;
-        next_sector += added_sectors;
-        next_chunk += added_chunks;
-        if (!s->synced && s->common.speed) {
-            delay_ns = ratelimit_calculate_delay(&s->limit, added_sectors);
-        }
-    } while (delay_ns == 0 && next_sector < end);
+    while (s->buf_free_count < nb_chunks) {
+        trace_mirror_yield_in_flight(s, sector_num, s->in_flight);
+        mirror_wait_for_io(s);
+    }
 
     /* Allocate a MirrorOp that is used as an AIO callback.  */
-    op = g_slice_new(MirrorOp);
+    op = g_new(MirrorOp, 1);
     op->s = s;
     op->sector_num = sector_num;
     op->nb_sectors = nb_sectors;
@@ -268,47 +260,158 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
      * from s->buf_free.
      */
     qemu_iovec_init(&op->qiov, nb_chunks);
-    next_sector = sector_num;
     while (nb_chunks-- > 0) {
         MirrorBuffer *buf = QSIMPLEQ_FIRST(&s->buf_free);
-        size_t remaining = (nb_sectors * BDRV_SECTOR_SIZE) - op->qiov.size;
+        size_t remaining = nb_sectors * BDRV_SECTOR_SIZE - op->qiov.size;
 
         QSIMPLEQ_REMOVE_HEAD(&s->buf_free, next);
         s->buf_free_count--;
         qemu_iovec_add(&op->qiov, buf, MIN(s->granularity, remaining));
-
-        /* Advance the HBitmapIter in parallel, so that we do not examine
-         * the same sector twice.
-         */
-        if (next_sector > hbitmap_next_sector
-            && bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
-            hbitmap_next_sector = hbitmap_iter_next(&s->hbi);
-        }
-
-        next_sector += sectors_per_chunk;
     }
 
-    bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num, nb_sectors);
-
     /* Copy the dirty cluster.  */
     s->in_flight++;
     s->sectors_in_flight += nb_sectors;
     trace_mirror_one_iteration(s, sector_num, nb_sectors);
 
-    ret = bdrv_get_block_status_above(source, NULL, sector_num,
-                                      nb_sectors, &pnum);
-    if (ret < 0 || pnum < nb_sectors ||
-            (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) {
-        bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
-                       mirror_read_complete, op);
-    } else if (ret & BDRV_BLOCK_ZERO) {
+    bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors,
+                   mirror_read_complete, op);
+    return ret;
+}
+
+static void mirror_do_zero_or_discard(MirrorBlockJob *s,
+                                      int64_t sector_num,
+                                      int nb_sectors,
+                                      bool is_discard)
+{
+    MirrorOp *op;
+
+    /* Allocate a MirrorOp that is used as an AIO callback. The qiov is zeroed
+     * so the freeing in mirror_iteration_done is nop. */
+    op = g_new0(MirrorOp, 1);
+    op->s = s;
+    op->sector_num = sector_num;
+    op->nb_sectors = nb_sectors;
+
+    s->in_flight++;
+    s->sectors_in_flight += nb_sectors;
+    if (is_discard) {
+        bdrv_aio_discard(s->target, sector_num, op->nb_sectors,
+                         mirror_write_complete, op);
+    } else {
         bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors,
                               s->unmap ? BDRV_REQ_MAY_UNMAP : 0,
                               mirror_write_complete, op);
-    } else {
-        assert(!(ret & BDRV_BLOCK_DATA));
-        bdrv_aio_discard(s->target, sector_num, op->nb_sectors,
-                         mirror_write_complete, op);
+    }
+}
+
+static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s)
+{
+    BlockDriverState *source = s->common.bs;
+    int64_t sector_num, first_chunk;
+    uint64_t delay_ns = 0;
+    /* At least the first dirty chunk is mirrored in one iteration. */
+    int nb_chunks = 1;
+    int64_t end = s->bdev_length / BDRV_SECTOR_SIZE;
+    int sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS;
+
+    sector_num = hbitmap_iter_next(&s->hbi);
+    if (sector_num < 0) {
+        bdrv_dirty_iter_init(s->dirty_bitmap, &s->hbi);
+        sector_num = hbitmap_iter_next(&s->hbi);
+        trace_mirror_restart_iter(s, bdrv_get_dirty_count(s->dirty_bitmap));
+        assert(sector_num >= 0);
+    }
+
+    first_chunk = sector_num / sectors_per_chunk;
+    while (test_bit(first_chunk, s->in_flight_bitmap)) {
+        trace_mirror_yield_in_flight(s, first_chunk, s->in_flight);
+        mirror_wait_for_io(s);
+    }
+
+    /* Find the number of consective dirty chunks following the first dirty
+     * one, and wait for in flight requests in them. */
+    while (nb_chunks * sectors_per_chunk < (s->buf_size >> BDRV_SECTOR_BITS)) {
+        int64_t hbitmap_next;
+        int64_t next_sector = sector_num + nb_chunks * sectors_per_chunk;
+        int64_t next_chunk = next_sector / sectors_per_chunk;
+        if (next_sector >= end ||
+            !bdrv_get_dirty(source, s->dirty_bitmap, next_sector)) {
+            break;
+        }
+        if (test_bit(next_chunk, s->in_flight_bitmap)) {
+            break;
+        }
+
+        hbitmap_next = hbitmap_iter_next(&s->hbi);
+        if (hbitmap_next > next_sector || hbitmap_next < 0) {
+            /* The bitmap iterator's cache is stale, refresh it */
+            bdrv_set_dirty_iter(&s->hbi, next_sector);
+            hbitmap_next = hbitmap_iter_next(&s->hbi);
+        }
+        assert(hbitmap_next == next_sector);
+        nb_chunks++;
+    }
+
+    /* Clear dirty bits before querying the block status, because
+     * calling bdrv_get_block_status_above could yield - if some blocks are
+     * marked dirty in this window, we need to know.
+     */
+    bdrv_reset_dirty_bitmap(s->dirty_bitmap, sector_num,
+                            nb_chunks * sectors_per_chunk);
+    bitmap_set(s->in_flight_bitmap, sector_num / sectors_per_chunk, nb_chunks);
+    while (nb_chunks > 0 && sector_num < end) {
+        int ret;
+        int io_sectors;
+        BlockDriverState *file;
+        enum MirrorMethod {
+            MIRROR_METHOD_COPY,
+            MIRROR_METHOD_ZERO,
+            MIRROR_METHOD_DISCARD
+        } mirror_method = MIRROR_METHOD_COPY;
+
+        assert(!(sector_num % sectors_per_chunk));
+        ret = bdrv_get_block_status_above(source, NULL, sector_num,
+                                          nb_chunks * sectors_per_chunk,
+                                          &io_sectors, &file);
+        if (ret < 0) {
+            io_sectors = nb_chunks * sectors_per_chunk;
+        }
+
+        io_sectors -= io_sectors % sectors_per_chunk;
+        if (io_sectors < sectors_per_chunk) {
+            io_sectors = sectors_per_chunk;
+        } else if (ret >= 0 && !(ret & BDRV_BLOCK_DATA)) {
+            int64_t target_sector_num;
+            int target_nb_sectors;
+            bdrv_round_to_clusters(s->target, sector_num, io_sectors,
+                                   &target_sector_num, &target_nb_sectors);
+            if (target_sector_num == sector_num &&
+                target_nb_sectors == io_sectors) {
+                mirror_method = ret & BDRV_BLOCK_ZERO ?
+                                    MIRROR_METHOD_ZERO :
+                                    MIRROR_METHOD_DISCARD;
+            }
+        }
+
+        mirror_clip_sectors(s, sector_num, &io_sectors);
+        switch (mirror_method) {
+        case MIRROR_METHOD_COPY:
+            io_sectors = mirror_do_read(s, sector_num, io_sectors);
+            break;
+        case MIRROR_METHOD_ZERO:
+            mirror_do_zero_or_discard(s, sector_num, io_sectors, false);
+            break;
+        case MIRROR_METHOD_DISCARD:
+            mirror_do_zero_or_discard(s, sector_num, io_sectors, true);
+            break;
+        default:
+            abort();
+        }
+        assert(io_sectors);
+        sector_num += io_sectors;
+        nb_chunks -= DIV_ROUND_UP(io_sectors, sectors_per_chunk);
+        delay_ns += ratelimit_calculate_delay(&s->limit, io_sectors);
     }
     return delay_ns;
 }
@@ -333,7 +436,7 @@ static void mirror_free_init(MirrorBlockJob *s)
 static void mirror_drain(MirrorBlockJob *s)
 {
     while (s->in_flight > 0) {
-        qemu_coroutine_yield();
+        mirror_wait_for_io(s);
     }
 }
 
@@ -346,6 +449,11 @@ static void mirror_exit(BlockJob *job, void *opaque)
     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
     MirrorExitData *data = opaque;
     AioContext *replace_aio_context = NULL;
+    BlockDriverState *src = s->common.bs;
+
+    /* Make sure that the source BDS doesn't go away before we called
+     * block_job_completed(). */
+    bdrv_ref(src);
 
     if (s->to_replace) {
         replace_aio_context = bdrv_get_aio_context(s->to_replace);
@@ -357,18 +465,22 @@ static void mirror_exit(BlockJob *job, void *opaque)
         if (s->to_replace) {
             to_replace = s->to_replace;
         }
+
+        /* This was checked in mirror_start_job(), but meanwhile one of the
+         * nodes could have been newly attached to a BlockBackend. */
+        if (to_replace->blk && s->target->blk) {
+            error_report("block job: Can't create node with two BlockBackends");
+            data->ret = -EINVAL;
+            goto out;
+        }
+
         if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
             bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
         }
-        bdrv_swap(s->target, to_replace);
-        if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
-            /* drop the bs loop chain formed by the swap: break the loop then
-             * trigger the unref from the top one */
-            BlockDriverState *p = s->base->backing_hd;
-            bdrv_set_backing_hd(s->base, NULL);
-            bdrv_unref(p);
-        }
+        bdrv_replace_in_backing_chain(to_replace, s->target);
     }
+
+out:
     if (s->to_replace) {
         bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
         error_free(s->replace_blocker);
@@ -378,9 +490,15 @@ static void mirror_exit(BlockJob *job, void *opaque)
         aio_context_release(replace_aio_context);
     }
     g_free(s->replaces);
+    bdrv_op_unblock_all(s->target, s->common.blocker);
     bdrv_unref(s->target);
     block_job_completed(&s->common, data->ret);
     g_free(data);
+    bdrv_drained_end(src);
+    if (qemu_get_aio_context() == bdrv_get_aio_context(src)) {
+        aio_enable_external(iohandler_get_aio_context());
+    }
+    bdrv_unref(src);
 }
 
 static void coroutine_fn mirror_run(void *opaque)
@@ -395,6 +513,7 @@ static void coroutine_fn mirror_run(void *opaque)
                                  checking for a NULL string */
     int ret = 0;
     int n;
+    int target_cluster_size = BDRV_SECTOR_SIZE;
 
     if (block_job_is_cancelled(&s->common)) {
         goto immediate_exit;
@@ -424,16 +543,16 @@ static void coroutine_fn mirror_run(void *opaque)
      */
     bdrv_get_backing_filename(s->target, backing_filename,
                               sizeof(backing_filename));
-    if (backing_filename[0] && !s->target->backing_hd) {
-        ret = bdrv_get_info(s->target, &bdi);
-        if (ret < 0) {
-            goto immediate_exit;
-        }
-        if (s->granularity < bdi.cluster_size) {
-            s->buf_size = MAX(s->buf_size, bdi.cluster_size);
-            s->cow_bitmap = bitmap_new(length);
-        }
+    if (!bdrv_get_info(s->target, &bdi) && bdi.cluster_size) {
+        target_cluster_size = bdi.cluster_size;
     }
+    if (backing_filename[0] && !s->target->backing
+        && s->granularity < target_cluster_size) {
+        s->buf_size = MAX(s->buf_size, target_cluster_size);
+        s->cow_bitmap = bitmap_new(length);
+    }
+    s->target_cluster_sectors = target_cluster_size >> BDRV_SECTOR_BITS;
+    s->max_iov = MIN(s->common.bs->bl.max_iov, s->target->bl.max_iov);
 
     end = s->bdev_length / BDRV_SECTOR_SIZE;
     s->buf = qemu_try_blockalign(bs, s->buf_size);
@@ -448,6 +567,8 @@ static void coroutine_fn mirror_run(void *opaque)
     if (!s->is_none_mode) {
         /* First part, loop on the sectors and initialize the dirty bitmap.  */
         BlockDriverState *base = s->base;
+        bool mark_all_dirty = s->base == NULL && !bdrv_has_zero_init(s->target);
+
         for (sector_num = 0; sector_num < end; ) {
             /* Just to make sure we are not exceeding int limit. */
             int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS,
@@ -470,7 +591,7 @@ static void coroutine_fn mirror_run(void *opaque)
             }
 
             assert(n > 0);
-            if (ret == 1) {
+            if (ret == 1 || mark_all_dirty) {
                 bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n);
             }
             sector_num += n;
@@ -506,7 +627,7 @@ static void coroutine_fn mirror_run(void *opaque)
             if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||
                 (cnt == 0 && s->in_flight > 0)) {
                 trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);
-                qemu_coroutine_yield();
+                mirror_wait_for_io(s);
                 continue;
             } else if (cnt != 0) {
                 delay_ns = mirror_iteration(s);
@@ -549,7 +670,7 @@ static void coroutine_fn mirror_run(void *opaque)
              * mirror_populate runs.
              */
             trace_mirror_before_drain(s, cnt);
-            bdrv_drain(bs);
+            bdrv_co_drain(bs);
             cnt = bdrv_get_dirty_count(s->dirty_bitmap);
         }
 
@@ -589,10 +710,21 @@ immediate_exit:
     g_free(s->cow_bitmap);
     g_free(s->in_flight_bitmap);
     bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
-    bdrv_iostatus_disable(s->target);
+    if (s->target->blk) {
+        blk_iostatus_disable(s->target->blk);
+    }
 
     data = g_malloc(sizeof(*data));
     data->ret = ret;
+    /* Before we switch to target in mirror_exit, make sure data doesn't
+     * change. */
+    bdrv_drained_begin(s->common.bs);
+    if (qemu_get_aio_context() == bdrv_get_aio_context(bs)) {
+        /* FIXME: virtio host notifiers run on iohandler_ctx, therefore the
+         * above bdrv_drained_end isn't enough to quiesce it. This is ugly, we
+         * need a block layer API change to achieve this. */
+        aio_disable_external(iohandler_get_aio_context());
+    }
     block_job_defer_to_main_loop(&s->common, mirror_exit, data);
 }
 
@@ -611,7 +743,9 @@ static void mirror_iostatus_reset(BlockJob *job)
 {
     MirrorBlockJob *s = container_of(job, MirrorBlockJob, common);
 
-    bdrv_iostatus_reset(s->target);
+    if (s->target->blk) {
+        blk_iostatus_reset(s->target->blk);
+    }
 }
 
 static void mirror_complete(BlockJob *job, Error **errp)
@@ -620,14 +754,13 @@ static void mirror_complete(BlockJob *job, Error **errp)
     Error *local_err = NULL;
     int ret;
 
-    ret = bdrv_open_backing_file(s->target, NULL, &local_err);
+    ret = bdrv_open_backing_file(s->target, NULL, "backing", &local_err);
     if (ret < 0) {
         error_propagate(errp, local_err);
         return;
     }
     if (!s->synced) {
-        error_setg(errp, QERR_BLOCK_JOB_NOT_READY,
-                   bdrv_get_device_name(job->bs));
+        error_setg(errp, QERR_BLOCK_JOB_NOT_READY, job->id);
         return;
     }
 
@@ -635,9 +768,9 @@ static void mirror_complete(BlockJob *job, Error **errp)
     if (s->replaces) {
         AioContext *replace_aio_context;
 
-        s->to_replace = check_to_replace_node(s->replaces, &local_err);
+        s->to_replace = bdrv_find_node(s->replaces);
         if (!s->to_replace) {
-            error_propagate(errp, local_err);
+            error_setg(errp, "Node name '%s' not found", s->replaces);
             return;
         }
 
@@ -686,6 +819,7 @@ static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
                              bool is_none_mode, BlockDriverState *base)
 {
     MirrorBlockJob *s;
+    BlockDriverState *replaced_bs;
 
     if (granularity == 0) {
         granularity = bdrv_get_default_bitmap_granularity(target);
@@ -695,7 +829,7 @@ static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
 
     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
-        !bdrv_iostatus_is_enabled(bs)) {
+        (!bs->blk || !blk_iostatus_is_enabled(bs->blk))) {
         error_setg(errp, QERR_INVALID_PARAMETER, "on-source-error");
         return;
     }
@@ -709,6 +843,21 @@ static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
         buf_size = DEFAULT_MIRROR_BUF_SIZE;
     }
 
+    /* We can't support this case as long as the block layer can't handle
+     * multiple BlockBackends per BlockDriverState. */
+    if (replaces) {
+        replaced_bs = bdrv_lookup_bs(replaces, replaces, errp);
+        if (replaced_bs == NULL) {
+            return;
+        }
+    } else {
+        replaced_bs = bs;
+    }
+    if (replaced_bs->blk && target->blk) {
+        error_setg(errp, "Can't create node with two BlockBackends");
+        return;
+    }
+
     s = block_job_create(driver, bs, speed, cb, opaque, errp);
     if (!s) {
         return;
@@ -727,12 +876,16 @@ static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
     s->dirty_bitmap = bdrv_create_dirty_bitmap(bs, granularity, NULL, errp);
     if (!s->dirty_bitmap) {
         g_free(s->replaces);
-        block_job_release(bs);
+        block_job_unref(&s->common);
         return;
     }
-    bdrv_set_enable_write_cache(s->target, true);
-    bdrv_set_on_error(s->target, on_target_error, on_target_error);
-    bdrv_iostatus_enable(s->target);
+
+    bdrv_op_block_all(s->target, s->common.blocker);
+
+    if (s->target->blk) {
+        blk_set_on_error(s->target->blk, on_target_error, on_target_error);
+        blk_iostatus_enable(s->target->blk);
+    }
     s->common.co = qemu_coroutine_create(mirror_run);
     trace_mirror_start(bs, s, s->common.co, opaque);
     qemu_coroutine_enter(s->common.co, s);
@@ -755,7 +908,7 @@ void mirror_start(BlockDriverState *bs, BlockDriverState *target,
         return;
     }
     is_none_mode = mode == MIRROR_SYNC_MODE_NONE;
-    base = mode == MIRROR_SYNC_MODE_TOP ? bs->backing_hd : NULL;
+    base = mode == MIRROR_SYNC_MODE_TOP ? backing_bs(bs) : NULL;
     mirror_start_job(bs, target, replaces,
                      speed, granularity, buf_size,
                      on_source_error, on_target_error, unmap, cb, opaque, errp,