diff options
author | Max Reitz <mreitz@redhat.com> | 2014-10-24 15:57:32 +0200 |
---|---|---|
committer | Stefan Hajnoczi <stefanha@redhat.com> | 2014-11-03 11:41:48 +0000 |
commit | 94054183daffaa41cd77ced9301c01a01027923a (patch) | |
tree | 75c6d4efe55d181cee75d7e3eaaf59ecbf1e9aa5 | |
parent | 491d27e2af4f6e157c4b29d43269c5cb0d191171 (diff) |
qcow2: Optimize bdrv_make_empty()
bdrv_make_empty() is currently only called if the current image
represents an external snapshot that has been committed to its base
image; it is therefore unlikely to have internal snapshots. In this
case, bdrv_make_empty() can be greatly sped up by emptying the L1 and
refcount table (while having the dirty flag set, which only works for
compat=1.1) and creating a trivial refcount structure.
If there are snapshots or for compat=0.10, fall back to the simple
implementation (discard all clusters).
[Applied s/clusters/cluster/ typo fix suggested by Eric Blake
--Stefan]
Signed-off-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-id: 1414159063-25977-4-git-send-email-mreitz@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
-rw-r--r-- | block/blkdebug.c | 2 | ||||
-rw-r--r-- | block/qcow2.c | 165 | ||||
-rw-r--r-- | include/block/block.h | 2 |
3 files changed, 168 insertions, 1 deletions
diff --git a/block/blkdebug.c b/block/blkdebug.c index e046b920fb..862d93b59b 100644 --- a/block/blkdebug.c +++ b/block/blkdebug.c @@ -195,6 +195,8 @@ static const char *event_names[BLKDBG_EVENT_MAX] = { [BLKDBG_PWRITEV] = "pwritev", [BLKDBG_PWRITEV_ZERO] = "pwritev_zero", [BLKDBG_PWRITEV_DONE] = "pwritev_done", + + [BLKDBG_EMPTY_IMAGE_PREPARE] = "empty_image_prepare", }; static int get_event_by_name(const char *name, BlkDebugEvent *event) diff --git a/block/qcow2.c b/block/qcow2.c index bf871d58f4..7ec7830fac 100644 --- a/block/qcow2.c +++ b/block/qcow2.c @@ -2230,12 +2230,175 @@ fail: return ret; } +static int make_completely_empty(BlockDriverState *bs) +{ + BDRVQcowState *s = bs->opaque; + int ret, l1_clusters; + int64_t offset; + uint64_t *new_reftable = NULL; + uint64_t rt_entry, l1_size2; + struct { + uint64_t l1_offset; + uint64_t reftable_offset; + uint32_t reftable_clusters; + } QEMU_PACKED l1_ofs_rt_ofs_cls; + + ret = qcow2_cache_empty(bs, s->l2_table_cache); + if (ret < 0) { + goto fail; + } + + ret = qcow2_cache_empty(bs, s->refcount_block_cache); + if (ret < 0) { + goto fail; + } + + /* Refcounts will be broken utterly */ + ret = qcow2_mark_dirty(bs); + if (ret < 0) { + goto fail; + } + + BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE); + + l1_clusters = DIV_ROUND_UP(s->l1_size, s->cluster_size / sizeof(uint64_t)); + l1_size2 = (uint64_t)s->l1_size * sizeof(uint64_t); + + /* After this call, neither the in-memory nor the on-disk refcount + * information accurately describe the actual references */ + + ret = bdrv_write_zeroes(bs->file, s->l1_table_offset / BDRV_SECTOR_SIZE, + l1_clusters * s->cluster_sectors, 0); + if (ret < 0) { + goto fail_broken_refcounts; + } + memset(s->l1_table, 0, l1_size2); + + BLKDBG_EVENT(bs->file, BLKDBG_EMPTY_IMAGE_PREPARE); + + /* Overwrite enough clusters at the beginning of the sectors to place + * the refcount table, a refcount block and the L1 table in; this may + * overwrite parts of the existing refcount and L1 table, which is not + * an issue because the dirty flag is set, complete data loss is in fact + * desired and partial data loss is consequently fine as well */ + ret = bdrv_write_zeroes(bs->file, s->cluster_size / BDRV_SECTOR_SIZE, + (2 + l1_clusters) * s->cluster_size / + BDRV_SECTOR_SIZE, 0); + /* This call (even if it failed overall) may have overwritten on-disk + * refcount structures; in that case, the in-memory refcount information + * will probably differ from the on-disk information which makes the BDS + * unusable */ + if (ret < 0) { + goto fail_broken_refcounts; + } + + BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE); + BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_UPDATE); + + /* "Create" an empty reftable (one cluster) directly after the image + * header and an empty L1 table three clusters after the image header; + * the cluster between those two will be used as the first refblock */ + cpu_to_be64w(&l1_ofs_rt_ofs_cls.l1_offset, 3 * s->cluster_size); + cpu_to_be64w(&l1_ofs_rt_ofs_cls.reftable_offset, s->cluster_size); + cpu_to_be32w(&l1_ofs_rt_ofs_cls.reftable_clusters, 1); + ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_table_offset), + &l1_ofs_rt_ofs_cls, sizeof(l1_ofs_rt_ofs_cls)); + if (ret < 0) { + goto fail_broken_refcounts; + } + + s->l1_table_offset = 3 * s->cluster_size; + + new_reftable = g_try_new0(uint64_t, s->cluster_size / sizeof(uint64_t)); + if (!new_reftable) { + ret = -ENOMEM; + goto fail_broken_refcounts; + } + + s->refcount_table_offset = s->cluster_size; + s->refcount_table_size = s->cluster_size / sizeof(uint64_t); + + g_free(s->refcount_table); + s->refcount_table = new_reftable; + new_reftable = NULL; + + /* Now the in-memory refcount information again corresponds to the on-disk + * information (reftable is empty and no refblocks (the refblock cache is + * empty)); however, this means some clusters (e.g. the image header) are + * referenced, but not refcounted, but the normal qcow2 code assumes that + * the in-memory information is always correct */ + + BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC); + + /* Enter the first refblock into the reftable */ + rt_entry = cpu_to_be64(2 * s->cluster_size); + ret = bdrv_pwrite_sync(bs->file, s->cluster_size, + &rt_entry, sizeof(rt_entry)); + if (ret < 0) { + goto fail_broken_refcounts; + } + s->refcount_table[0] = 2 * s->cluster_size; + + s->free_cluster_index = 0; + assert(3 + l1_clusters <= s->refcount_block_size); + offset = qcow2_alloc_clusters(bs, 3 * s->cluster_size + l1_size2); + if (offset < 0) { + ret = offset; + goto fail_broken_refcounts; + } else if (offset > 0) { + error_report("First cluster in emptied image is in use"); + abort(); + } + + /* Now finally the in-memory information corresponds to the on-disk + * structures and is correct */ + ret = qcow2_mark_clean(bs); + if (ret < 0) { + goto fail; + } + + ret = bdrv_truncate(bs->file, (3 + l1_clusters) * s->cluster_size); + if (ret < 0) { + goto fail; + } + + return 0; + +fail_broken_refcounts: + /* The BDS is unusable at this point. If we wanted to make it usable, we + * would have to call qcow2_refcount_close(), qcow2_refcount_init(), + * qcow2_check_refcounts(), qcow2_refcount_close() and qcow2_refcount_init() + * again. However, because the functions which could have caused this error + * path to be taken are used by those functions as well, it's very likely + * that that sequence will fail as well. Therefore, just eject the BDS. */ + bs->drv = NULL; + +fail: + g_free(new_reftable); + return ret; +} + static int qcow2_make_empty(BlockDriverState *bs) { - int ret = 0; + BDRVQcowState *s = bs->opaque; uint64_t start_sector; int sector_step = INT_MAX / BDRV_SECTOR_SIZE; + int l1_clusters, ret = 0; + + l1_clusters = DIV_ROUND_UP(s->l1_size, s->cluster_size / sizeof(uint64_t)); + + if (s->qcow_version >= 3 && !s->snapshots && + 3 + l1_clusters <= s->refcount_block_size) { + /* The following function only works for qcow2 v3 images (it requires + * the dirty flag) and only as long as there are no snapshots (because + * it completely empties the image). Furthermore, the L1 table and three + * additional clusters (image header, refcount table, one refcount + * block) have to fit inside one refcount block. */ + return make_completely_empty(bs); + } + /* This fallback code simply discards every active cluster; this is slow, + * but works in all cases */ for (start_sector = 0; start_sector < bs->total_sectors; start_sector += sector_step) { diff --git a/include/block/block.h b/include/block/block.h index 341054dcf7..b1f4385f03 100644 --- a/include/block/block.h +++ b/include/block/block.h @@ -498,6 +498,8 @@ typedef enum { BLKDBG_PWRITEV_ZERO, BLKDBG_PWRITEV_DONE, + BLKDBG_EMPTY_IMAGE_PREPARE, + BLKDBG_EVENT_MAX, } BlkDebugEvent; |