diff mbox

[RFC,V6,08/33] qcow2: Add qcow2_dedup_store_new_hashes.

Message ID 1360153926-9492-9-git-send-email-benoit@irqsave.net
State New
Headers show

Commit Message

Benoît Canet Feb. 6, 2013, 12:31 p.m. UTC
Signed-off-by: Benoit Canet <benoit@irqsave.net>
---
 block/qcow2-dedup.c |  325 ++++++++++++++++++++++++++++++++++++++++++++++++++-
 block/qcow2.h       |    5 +
 2 files changed, 329 insertions(+), 1 deletion(-)

Comments

Eric Blake Feb. 6, 2013, 5:56 p.m. UTC | #1
On 02/06/2013 05:31 AM, Benoît Canet wrote:
> Signed-off-by: Benoit Canet <benoit@irqsave.net>
> ---
>  block/qcow2-dedup.c |  325 ++++++++++++++++++++++++++++++++++++++++++++++++++-
>  block/qcow2.h       |    5 +
>  2 files changed, 329 insertions(+), 1 deletion(-)
> 

> +
> +    /* Remember that this QCowHashNoderepresent the first occurence of the

s/QCowHashNoderepresent/QCowHashNode represents/
s/occurence/occurrence/

> +     * cluste so we will be able to clear QCOW_OFLAG_COPIED from the L2 table

s/cluste/cluster/
Stefan Hajnoczi Feb. 6, 2013, 9:47 p.m. UTC | #2
On Wed, Feb 06, 2013 at 01:31:41PM +0100, Benoît Canet wrote:
> @@ -291,7 +297,11 @@ static int qcow2_clear_l2_copied_flag_if_needed(BlockDriverState *bs,
>      /* remember that we dont't need to clear QCOW_OFLAG_COPIED again */
>      hash_node->first_logical_sect &= first_logical_sect;
>  
> -    return 0;
> +    /* clear the QCOW_FLAG_FIRST flag from disk */
> +    return qcow2_dedup_read_write_hash(bs, &hash_node->hash,
> +                                       &hash_node->first_logical_sect,
> +                                       hash_node->physical_sect,
> +                                       true);

The order of metadata updates needs to be thought through, especially
what happens on power failure partway through.  I don't understand the
dedup code well enough to decide whether this call is safe.

Hopefully towards the end of this patch series I'll be able to think
through this.

>  }
>  
>  /* This function deduplicate a cluster
> @@ -553,3 +563,316 @@ exit:
>  
>      return deduped_clusters_nr * s->cluster_sectors - begining_index;
>  }
> +
> +
> +/* Create a deduplication table hash block, write it's offset to disk and
> + * reference it in the RAM deduplication table
> + *
> + * sync this to disk and get the dedup cluster cache entry
> + *
> + * @index: index in the RAM deduplication table
> + * @ret:   offset on success, negative on error
> + */
> +static uint64_t qcow2_create_block(BlockDriverState *bs,
> +                                               int32_t index)

qcow2_create_block() is a vague name.  qcow2_create_dedup_block()?

> +{
> +    BDRVQcowState *s = bs->opaque;
> +    int64_t offset;
> +    uint64_t data64;
> +    int ret = 0;
> +
> +    /* allocate a new dedup table hash block */
> +    offset = qcow2_alloc_clusters(bs, s->hash_block_size);
> +
> +    if (offset < 0) {
> +        return offset;
> +    }
> +
> +    ret = qcow2_cache_flush(bs, s->refcount_block_cache);
> +    if (ret < 0) {
> +        goto free_fail;
> +    }

The hash block needs to be initialized and on disk before it gets linked
into the dedup L1 table.  Otherwise the L1 table entry would point to a
hash table with undefined values.

> +static inline bool qcow2_has_dedup_block(BlockDriverState *bs,
> +                                         uint32_t index)
> +{
> +    BDRVQcowState *s = bs->opaque;
> +    return s->dedup_table[index] == 0 ? false : true;

Zero is false and non-zeor is true, so the following is equivalent:

return s->dedup_table[index];

> +}
> +
> +static inline void qcow2_write_hash_to_block_and_dirty(BlockDriverState *bs,
> +                                                       uint8_t *block,
> +                                                       QCowHash *hash,
> +                                                       int offset,
> +                                                       uint64_t *logical_sect)

Nothing gets written to disk here, so "write" shouldn't be in the name.
How about calling it qcow2_set_hash_block_entry()?

> +{
> +    BDRVQcowState *s = bs->opaque;
> +    uint64_t first;
> +    first = cpu_to_be64(*logical_sect);
> +    memcpy(block + offset, hash->data, HASH_LENGTH);
> +    memcpy(block + offset + HASH_LENGTH, &first, 8);
> +    qcow2_cache_entry_mark_dirty(s->dedup_cluster_cache, block);
> +}
> +
> +static inline uint64_t qcow2_read_hash_from_block(uint8_t *block,
> +                                                  QCowHash *hash,
> +                                                  int offset)
> +{
> +    uint64_t first;
> +    memcpy(hash->data, block + offset, HASH_LENGTH);
> +    memcpy(&first, block + offset + HASH_LENGTH, 8);
> +    return be64_to_cpu(first);
> +}
> +
> +/* Read/write a given hash and cluster_sect from/to the dedup table
> + *
> + * This function doesn't flush the dedup cache to disk
> + *
> + * @hash:                     the hash to read or store
> + * @first_logical_sect:       logical sector of the QCOW_FLAG_OCOPIED cluster

Why mention QCOW_FLAG_OCOPIED here?  Not sure if this holds true after
the first logical sector contents are changed.  Probably best not to
mention QCOW_FLAG_OCOPIED.

> + * @physical_sect:            sector of the cluster in QCOW2 file (in sectors)
> + * @write:                    true to write, false to read
> + * @ret:                      0 on succes, errno on error

s/succes/success/
s/errno/-errno/

> +/* This function store a hash information to disk and RAM
> + *
> + * @hash:           the QCowHash to process
> + * @logical_sect:   the logical sector of the cluster seen by the guest
> + * @physical_sect:  the physical sector of the stored cluster
> + * @ret:            0 on success, negative on error
> + */
> +static int qcow2_store_hash(BlockDriverState *bs,
> +                            QCowHash *hash,
> +                            uint64_t logical_sect,
> +                            uint64_t physical_sect)
> +{
> +    BDRVQcowState *s = bs->opaque;
> +    QCowHashNode *hash_node;
> +
> +    hash_node = g_tree_lookup(s->dedup_tree_by_hash, hash);
> +
> +    /* no hash node found for this hash */
> +    if (!hash_node) {
> +        return 0;
> +    }
> +
> +    /* the hash node information are already completed */
> +    if (!is_hash_node_empty(hash_node)) {
> +        return 0;
> +    }
> +
> +    /* Remember that this QCowHashNoderepresent the first occurence of the

s/QCowHashNoderepresent/QCowHashNode represents/
s/occurence/occurrence/

> +     * cluste so we will be able to clear QCOW_OFLAG_COPIED from the L2 table

s/cluste/cluster/
diff mbox

Patch

diff --git a/block/qcow2-dedup.c b/block/qcow2-dedup.c
index 5901749..a424af8 100644
--- a/block/qcow2-dedup.c
+++ b/block/qcow2-dedup.c
@@ -29,6 +29,12 @@ 
 #include "qemu-common.h"
 #include "qcow2.h"
 
+static int qcow2_dedup_read_write_hash(BlockDriverState *bs,
+                                       QCowHash *hash,
+                                       uint64_t *first_logical_sect,
+                                       uint64_t physical_sect,
+                                       bool write);
+
 /*
  * Prepare a buffer containing all the required data required to compute cluster
  * sized deduplication hashes.
@@ -291,7 +297,11 @@  static int qcow2_clear_l2_copied_flag_if_needed(BlockDriverState *bs,
     /* remember that we dont't need to clear QCOW_OFLAG_COPIED again */
     hash_node->first_logical_sect &= first_logical_sect;
 
-    return 0;
+    /* clear the QCOW_FLAG_FIRST flag from disk */
+    return qcow2_dedup_read_write_hash(bs, &hash_node->hash,
+                                       &hash_node->first_logical_sect,
+                                       hash_node->physical_sect,
+                                       true);
 }
 
 /* This function deduplicate a cluster
@@ -553,3 +563,316 @@  exit:
 
     return deduped_clusters_nr * s->cluster_sectors - begining_index;
 }
+
+
+/* Create a deduplication table hash block, write it's offset to disk and
+ * reference it in the RAM deduplication table
+ *
+ * sync this to disk and get the dedup cluster cache entry
+ *
+ * @index: index in the RAM deduplication table
+ * @ret:   offset on success, negative on error
+ */
+static uint64_t qcow2_create_block(BlockDriverState *bs,
+                                               int32_t index)
+{
+    BDRVQcowState *s = bs->opaque;
+    int64_t offset;
+    uint64_t data64;
+    int ret = 0;
+
+    /* allocate a new dedup table hash block */
+    offset = qcow2_alloc_clusters(bs, s->hash_block_size);
+
+    if (offset < 0) {
+        return offset;
+    }
+
+    ret = qcow2_cache_flush(bs, s->refcount_block_cache);
+    if (ret < 0) {
+        goto free_fail;
+    }
+
+    /* write the new block offset in the dedup table L1 */
+    data64 = cpu_to_be64(offset);
+    ret = bdrv_pwrite_sync(bs->file,
+                           s->dedup_table_offset +
+                           index * sizeof(uint64_t),
+                           &data64, sizeof(data64));
+
+    if (ret < 0) {
+        goto free_fail;
+    }
+
+    s->dedup_table[index] = offset;
+
+    return offset;
+
+free_fail:
+    qcow2_free_clusters(bs, offset, s->hash_block_size);
+    return ret;
+}
+
+static int qcow2_create_and_get_block(BlockDriverState *bs,
+                                      uint32_t index,
+                                      uint8_t **block)
+{
+    BDRVQcowState *s = bs->opaque;
+    int ret = 0;
+    int64_t offset;
+
+    offset = qcow2_create_block(bs, index);
+
+    if (offset < 0) {
+        return offset;
+    }
+
+
+    /* get an empty cluster from the dedup cache */
+    ret = qcow2_cache_get_empty(bs, s->dedup_cluster_cache,
+                                offset,
+                                (void **) block);
+
+    if (ret < 0) {
+        return ret;
+    }
+
+    /* clear it */
+    memset(*block, 0, s->hash_block_size);
+
+    return 0;
+}
+
+static inline bool qcow2_has_dedup_block(BlockDriverState *bs,
+                                         uint32_t index)
+{
+    BDRVQcowState *s = bs->opaque;
+    return s->dedup_table[index] == 0 ? false : true;
+}
+
+static inline void qcow2_write_hash_to_block_and_dirty(BlockDriverState *bs,
+                                                       uint8_t *block,
+                                                       QCowHash *hash,
+                                                       int offset,
+                                                       uint64_t *logical_sect)
+{
+    BDRVQcowState *s = bs->opaque;
+    uint64_t first;
+    first = cpu_to_be64(*logical_sect);
+    memcpy(block + offset, hash->data, HASH_LENGTH);
+    memcpy(block + offset + HASH_LENGTH, &first, 8);
+    qcow2_cache_entry_mark_dirty(s->dedup_cluster_cache, block);
+}
+
+static inline uint64_t qcow2_read_hash_from_block(uint8_t *block,
+                                                  QCowHash *hash,
+                                                  int offset)
+{
+    uint64_t first;
+    memcpy(hash->data, block + offset, HASH_LENGTH);
+    memcpy(&first, block + offset + HASH_LENGTH, 8);
+    return be64_to_cpu(first);
+}
+
+/* Read/write a given hash and cluster_sect from/to the dedup table
+ *
+ * This function doesn't flush the dedup cache to disk
+ *
+ * @hash:                     the hash to read or store
+ * @first_logical_sect:       logical sector of the QCOW_FLAG_OCOPIED cluster
+ * @physical_sect:            sector of the cluster in QCOW2 file (in sectors)
+ * @write:                    true to write, false to read
+ * @ret:                      0 on succes, errno on error
+ */
+static int qcow2_dedup_read_write_hash(BlockDriverState *bs,
+                                       QCowHash *hash,
+                                       uint64_t *first_logical_sect,
+                                       uint64_t physical_sect,
+                                       bool write)
+{
+    BDRVQcowState *s = bs->opaque;
+    uint8_t *block = NULL;
+    int ret = 0;
+    int64_t cluster_number;
+    uint32_t index_in_dedup_table;
+    int offset_in_block;
+    int nb_hash_in_block = s->hash_block_size / (HASH_LENGTH + 8);
+
+    cluster_number = physical_sect / s->cluster_sectors;
+    index_in_dedup_table = cluster_number / nb_hash_in_block;
+
+    if (s->dedup_table_size <= index_in_dedup_table) {
+        return -ENOSPC;
+    }
+
+    /* if we must read and there is nothing to read return a null hash */
+    if (!qcow2_has_dedup_block(bs, index_in_dedup_table) && !write) {
+        memset(hash->data, 0, HASH_LENGTH);
+        *first_logical_sect = 0;
+        return 0;
+    }
+
+    if (qcow2_has_dedup_block(bs, index_in_dedup_table)) {
+        ret = qcow2_cache_get(bs,
+                              s->dedup_cluster_cache,
+                              s->dedup_table[index_in_dedup_table],
+                              (void **) &block);
+    } else {
+        ret = qcow2_create_and_get_block(bs,
+                                         index_in_dedup_table,
+                                         &block);
+    }
+
+    if (ret < 0) {
+        return ret;
+    }
+
+    offset_in_block = (cluster_number % nb_hash_in_block) *
+                      (HASH_LENGTH + 8);
+
+    if (write)  {
+        qcow2_write_hash_to_block_and_dirty(bs,
+                                            block,
+                                            hash,
+                                            offset_in_block,
+                                            first_logical_sect);
+    } else  {
+        *first_logical_sect = qcow2_read_hash_from_block(block,
+                                                         hash,
+                                                         offset_in_block);
+    }
+
+    qcow2_cache_put(bs, s->dedup_cluster_cache, (void **) &block);
+
+    return 0;
+}
+
+static inline bool is_hash_node_empty(QCowHashNode *hash_node)
+{
+    return hash_node->physical_sect & QCOW_FLAG_EMPTY;
+}
+
+static void qcow2_remove_hash_node(BlockDriverState *bs,
+                                   QCowHashNode *hash_node)
+{
+    BDRVQcowState *s = bs->opaque;
+    g_tree_remove(s->dedup_tree_by_sect, &hash_node->physical_sect);
+    g_tree_remove(s->dedup_tree_by_hash, &hash_node->hash);
+}
+
+/* This function removes a hash_node from the trees given a physical sector
+ *
+ * @physical_sect: The physical sector of the cluster corresponding to the hash
+ */
+static void qcow2_remove_hash_node_by_sector(BlockDriverState *bs,
+                                            uint64_t physical_sect)
+{
+    BDRVQcowState *s = bs->opaque;
+    QCowHashNode *hash_node;
+
+    hash_node = g_tree_lookup(s->dedup_tree_by_sect, &physical_sect);
+
+    if (!hash_node) {
+        return;
+    }
+
+    qcow2_remove_hash_node(bs, hash_node);
+}
+
+/* This function store a hash information to disk and RAM
+ *
+ * @hash:           the QCowHash to process
+ * @logical_sect:   the logical sector of the cluster seen by the guest
+ * @physical_sect:  the physical sector of the stored cluster
+ * @ret:            0 on success, negative on error
+ */
+static int qcow2_store_hash(BlockDriverState *bs,
+                            QCowHash *hash,
+                            uint64_t logical_sect,
+                            uint64_t physical_sect)
+{
+    BDRVQcowState *s = bs->opaque;
+    QCowHashNode *hash_node;
+
+    hash_node = g_tree_lookup(s->dedup_tree_by_hash, hash);
+
+    /* no hash node found for this hash */
+    if (!hash_node) {
+        return 0;
+    }
+
+    /* the hash node information are already completed */
+    if (!is_hash_node_empty(hash_node)) {
+        return 0;
+    }
+
+    /* Remember that this QCowHashNoderepresent the first occurence of the
+     * cluste so we will be able to clear QCOW_OFLAG_COPIED from the L2 table
+     * entry when refcount will go > 1.
+     */
+    logical_sect = logical_sect | QCOW_FLAG_FIRST;
+
+    /* remove stale hash node pointing to this physical sector from the trees */
+    qcow2_remove_hash_node_by_sector(bs, physical_sect);
+
+    /* fill the missing fields of the hash node */
+    hash_node->physical_sect = physical_sect;
+    hash_node->first_logical_sect = logical_sect;
+
+    /* insert the hash node in the second tree: it's already in the first one */
+    g_tree_insert(s->dedup_tree_by_sect, &hash_node->physical_sect, hash_node);
+
+    /* write the hash to disk */
+    return qcow2_dedup_read_write_hash(bs,
+                                       hash,
+                                       &logical_sect,
+                                       physical_sect,
+                                       true);
+}
+
+/* This function store the hashes of the clusters which are not duplicated
+ *
+ * @ds:            The deduplication state
+ * @count:         the number of dedup hash to process
+ * @logical_sect:  logical offset of the first cluster (in sectors)
+ * @physical_sect: offset of the first cluster (in sectors)
+ * @ret:           0 on succes, errno on error
+ */
+int qcow2_dedup_store_new_hashes(BlockDriverState *bs,
+                                 QCowDedupState *ds,
+                                 int count,
+                                 uint64_t logical_sect,
+                                 uint64_t physical_sect)
+{
+    int ret = 0;
+    int i = 0;
+    BDRVQcowState *s = bs->opaque;
+    QCowHashElement *dedup_hash, *next_dedup_hash;
+
+    /* round values on cluster boundaries for easier cluster deletion */
+    logical_sect = logical_sect & ~(s->cluster_sectors - 1);
+    physical_sect = physical_sect & ~(s->cluster_sectors - 1);
+
+    QTAILQ_FOREACH_SAFE(dedup_hash, &ds->undedupables, next, next_dedup_hash) {
+
+        ret = qcow2_store_hash(bs,
+                               &dedup_hash->hash,
+                               logical_sect + i * s->cluster_sectors,
+                               physical_sect + i * s->cluster_sectors);
+
+        QTAILQ_REMOVE(&ds->undedupables, dedup_hash, next);
+        g_free(dedup_hash);
+
+        if (ret < 0) {
+            break;
+        }
+
+        i++;
+
+        if (i == count) {
+            break;
+        }
+    }
+
+    return ret;
+}
diff --git a/block/qcow2.h b/block/qcow2.h
index 46a5800..3b076db 100644
--- a/block/qcow2.h
+++ b/block/qcow2.h
@@ -468,5 +468,10 @@  int qcow2_dedup(BlockDriverState *bs,
                 uint64_t sector_num,
                 uint8_t *data,
                 int data_nr);
+int qcow2_dedup_store_new_hashes(BlockDriverState *bs,
+                                 QCowDedupState *ds,
+                                 int count,
+                                 uint64_t logical_sect,
+                                 uint64_t physical_sect);
 
 #endif