Re: [Qemu-devel] [PATCH v3 3/6] vmdk: New functions to assist allocating multiple clusters

[Fam Zheng <famz@redhat.com>]

On Sat, 04/01 20:14, Ashijeet Acharya wrote:
> Move the cluster tables loading code out of the existing
> get_cluster_offset() function to avoid code duplication and implement it
> in separate get_cluster_table() and vmdk_L2load() functions.
> 
> Introduce two new helper functions handle_alloc() and
> vmdk_alloc_cluster_offset(). handle_alloc() helps to allocate multiple
> clusters at once starting from a given offset on disk and performs COW
> if necessary for first and last allocated clusters.
> vmdk_alloc_cluster_offset() helps to return the offset of the first of
> the many newly allocated clusters. Also, provide proper documentation
> for both.
> 
> Signed-off-by: Ashijeet Acharya <ashijeetacharya@gmail.com>
> ---
>  block/vmdk.c | 337 ++++++++++++++++++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 308 insertions(+), 29 deletions(-)
> 
> diff --git a/block/vmdk.c b/block/vmdk.c
> index 73ae786..e5a289d 100644
> --- a/block/vmdk.c
> +++ b/block/vmdk.c
> @@ -136,6 +136,7 @@ typedef struct VmdkMetaData {
>      unsigned int l2_offset;
>      int valid;
>      uint32_t *l2_cache_entry;
> +    uint32_t nb_clusters;
>  } VmdkMetaData;
>  
>  typedef struct VmdkGrainMarker {
> @@ -254,6 +255,14 @@ static inline uint64_t vmdk_find_offset_in_cluster(VmdkExtent *extent,
>      return extent_relative_offset % cluster_size;
>  }
>  
> +static inline uint64_t size_to_clusters(VmdkExtent *extent, uint64_t size)
> +{
> +    uint64_t cluster_size, round_off_size;
> +    cluster_size = extent->cluster_sectors * BDRV_SECTOR_SIZE;
> +    round_off_size = cluster_size - (size % cluster_size);
> +    return DIV_ROUND_UP(size + round_off_size, BDRV_SECTOR_SIZE * 128) - 1;

What is (BDRV_SECTOR_SIZE * 128)? Do you mean extent->cluster_size?  And the
function doesn't make sense up to me.

Just un-inline this to

    DIV_ROUND_UP(size,
                 extent->cluster_sectors << BDRV_SECTOR_BITS) - 1

in the calling site and be done with it.

> +}
> +
>  static uint32_t vmdk_read_cid(BlockDriverState *bs, int parent)
>  {
>      char *desc;
> @@ -1028,6 +1037,133 @@ static void vmdk_refresh_limits(BlockDriverState *bs, Error **errp)
>      }
>  }
>  
> +static int vmdk_L2update(VmdkExtent *extent, VmdkMetaData *m_data,
> +                         uint32_t offset)
> +{
> +    offset = cpu_to_le32(offset);
> +    /* update L2 table */
> +    if (bdrv_pwrite_sync(extent->file,
> +                ((int64_t)m_data->l2_offset * 512)
> +                    + (m_data->l2_index * sizeof(offset)),
> +                &offset, sizeof(offset)) < 0) {
> +        return VMDK_ERROR;
> +    }
> +    /* update backup L2 table */
> +    if (extent->l1_backup_table_offset != 0) {
> +        m_data->l2_offset = extent->l1_backup_table[m_data->l1_index];
> +        if (bdrv_pwrite_sync(extent->file,
> +                    ((int64_t)m_data->l2_offset * 512)
> +                        + (m_data->l2_index * sizeof(offset)),
> +                    &offset, sizeof(offset)) < 0) {
> +            return VMDK_ERROR;
> +        }
> +    }
> +    if (m_data->l2_cache_entry) {
> +        *m_data->l2_cache_entry = offset;
> +    }
> +
> +    return VMDK_OK;
> +}
> +
> +/*
> + * vmdk_l2load
> + *
> + * Loads a new L2 table into memory. If the table is in the cache, the cache

Not a native speaker, but s/Loads/Load/ feels more nature and consistent with
other comments.

> + * is used; otherwise the L2 table is loaded from the image file.
> + *
> + * Returns:
> + *   VMDK_OK:       on success
> + *   VMDK_ERROR:    in error cases
> + */
> +static int vmdk_l2load(VmdkExtent *extent, uint64_t offset, int l2_offset,
> +                       uint32_t **new_l2_table, int *new_l2_index)
> +{
> +    int min_index, i, j;
> +    uint32_t *l2_table;
> +    uint32_t min_count;
> +
> +    for (i = 0; i < L2_CACHE_SIZE; i++) {
> +        if (l2_offset == extent->l2_cache_offsets[i]) {
> +            /* increment the hit count */
> +            if (++extent->l2_cache_counts[i] == UINT32_MAX) {
> +                for (j = 0; j < L2_CACHE_SIZE; j++) {
> +                    extent->l2_cache_counts[j] >>= 1;
> +                }
> +            }
> +            l2_table = extent->l2_cache + (i * extent->l2_size);
> +            goto found;
> +        }
> +    }
> +    /* not found: load a new entry in the least used one */
> +    min_index = 0;
> +    min_count = UINT32_MAX;
> +    for (i = 0; i < L2_CACHE_SIZE; i++) {
> +        if (extent->l2_cache_counts[i] < min_count) {
> +            min_count = extent->l2_cache_counts[i];
> +            min_index = i;
> +        }
> +    }
> +    l2_table = extent->l2_cache + (min_index * extent->l2_size);
> +    if (bdrv_pread(extent->file,
> +                (int64_t)l2_offset * 512,
> +                l2_table,
> +                extent->l2_size * sizeof(uint32_t)
> +            ) != extent->l2_size * sizeof(uint32_t)) {
> +        return VMDK_ERROR;
> +    }
> +
> +    extent->l2_cache_offsets[min_index] = l2_offset;
> +    extent->l2_cache_counts[min_index] = 1;
> +found:
> +    *new_l2_index = ((offset >> 9) / extent->cluster_sectors) % extent->l2_size;
> +    *new_l2_table = l2_table;
> +
> +    return VMDK_OK;
> +}
> +
> +/*
> + * get_cluster_table
> + *
> + * for a given offset, load (and allocate if needed) the l2 table.
> + *
> + * Returns:
> + *   VMDK_OK:        on success
> + *
> + *   VMDK_UNALLOC:   if cluster is not mapped
> + *
> + *   VMDK_ERROR:     in error cases
> + */
> +static int get_cluster_table(VmdkExtent *extent, uint64_t offset,
> +                             int *new_l1_index, int *new_l2_offset,
> +                             int *new_l2_index, uint32_t **new_l2_table)
> +{
> +    int l1_index, l2_offset, l2_index;
> +    uint32_t *l2_table;
> +    int ret;
> +
> +    offset -= (extent->end_sector - extent->sectors) * SECTOR_SIZE;
> +    l1_index = (offset >> 9) / extent->l1_entry_sectors;
> +    if (l1_index >= extent->l1_size) {
> +        return VMDK_ERROR;
> +    }
> +    l2_offset = extent->l1_table[l1_index];
> +    if (!l2_offset) {
> +        return VMDK_UNALLOC;
> +    }
> +
> +    ret = vmdk_l2load(extent, offset, l2_offset, &l2_table, &l2_index);
> +    if (ret < 0) {
> +        return ret;
> +    }
> +
> +    *new_l1_index = l1_index;
> +    *new_l2_offset = l2_offset;
> +    *new_l2_index = l2_index;
> +    *new_l2_table = l2_table;
> +
> +    return VMDK_OK;
> +}
> +

Can you move this hunk into patch 4 and put it before this patch? It will make
reviewing a bit easier. (Yes, this patch is already big.)

>  /*
>   * vmdk_perform_cow
>   *
> @@ -1115,29 +1251,168 @@ exit:
>      return ret;
>  }
>  
> -static int vmdk_L2update(VmdkExtent *extent, VmdkMetaData *m_data,
> -                         uint32_t offset)
> +/*
> + * handle_alloc
> + *
> + * Allocates new clusters for an area that either is yet unallocated or needs a

Similar to vmdk_l2load, s/Allocates/Allocate/

> + * copy on write. If *cluster_offset is non_zero, clusters are only allocated if
> + * the new allocation can match the specified host offset.
> + *
> + * Returns:
> + *   VMDK_OK:       if new clusters were allocated, *bytes may be decreased if
> + *                  the new allocation doesn't cover all of the requested area.
> + *                  *cluster_offset is updated to contain the offset of the
> + *                  first newly allocated cluster.
> + *
> + *   VMDK_UNALLOC:  if no clusters could be allocated. *cluster_offset is left
> + *                  unchanged.
> + *
> + *   VMDK_ERROR:    in error cases
> + */
> +static int handle_alloc(BlockDriverState *bs, VmdkExtent *extent,
> +                        uint64_t offset, uint64_t *cluster_offset,
> +                        int64_t *bytes, VmdkMetaData *m_data,
> +                        bool allocate, uint32_t *total_alloc_clusters)
>  {
> -    offset = cpu_to_le32(offset);
> -    /* update L2 table */
> -    if (bdrv_pwrite_sync(extent->file,
> -                ((int64_t)m_data->l2_offset * 512)
> -                    + (m_data->l2_index * sizeof(offset)),
> -                &offset, sizeof(offset)) < 0) {
> -        return VMDK_ERROR;
> +    int l1_index, l2_offset, l2_index;
> +    uint32_t *l2_table;
> +    uint32_t cluster_sector;
> +    uint32_t nb_clusters;
> +    bool zeroed = false;
> +    uint64_t skip_start_bytes, skip_end_bytes;
> +    int ret;
> +
> +    ret = get_cluster_table(extent, offset, &l1_index, &l2_offset,
> +                            &l2_index, &l2_table);
> +    if (ret < 0) {
> +        return ret;
>      }
> -    /* update backup L2 table */
> -    if (extent->l1_backup_table_offset != 0) {
> -        m_data->l2_offset = extent->l1_backup_table[m_data->l1_index];
> -        if (bdrv_pwrite_sync(extent->file,
> -                    ((int64_t)m_data->l2_offset * 512)
> -                        + (m_data->l2_index * sizeof(offset)),
> -                    &offset, sizeof(offset)) < 0) {
> -            return VMDK_ERROR;
> +
> +    cluster_sector = le32_to_cpu(l2_table[l2_index]);
> +
> +    skip_start_bytes = vmdk_find_offset_in_cluster(extent, offset);
> +    /* Calculate the number of clusters to look for. Here it will return one
> +     * cluster less than the actual value calculated as we may need to perfrom
> +     * COW for the last one. */
> +    nb_clusters = size_to_clusters(extent, skip_start_bytes + *bytes);
> +
> +    nb_clusters = MIN(nb_clusters, extent->l2_size - l2_index);
> +    assert(nb_clusters <= INT_MAX);
> +
> +    /* update bytes according to final nb_clusters value */
> +    if (nb_clusters != 0) {
> +        *bytes = ((nb_clusters * extent->cluster_sectors) << 9)
> +                                            - skip_start_bytes;
> +    } else {
> +        nb_clusters = 1;
> +    }
> +    *total_alloc_clusters += nb_clusters;
> +    skip_end_bytes = skip_start_bytes + MIN(*bytes,
> +                     extent->cluster_sectors * BDRV_SECTOR_SIZE
> +                                    - skip_start_bytes);
> +
> +    if (extent->has_zero_grain && cluster_sector == VMDK_GTE_ZEROED) {
> +        zeroed = true;
> +    }
> +
> +    if (!cluster_sector || zeroed) {
> +        if (!allocate) {
> +            return zeroed ? VMDK_ZEROED : VMDK_UNALLOC;
> +        }
> +
> +        cluster_sector = extent->next_cluster_sector;
> +        extent->next_cluster_sector += extent->cluster_sectors
> +                                                * nb_clusters;
> +
> +        ret = vmdk_perform_cow(bs, extent, cluster_sector * BDRV_SECTOR_SIZE,
> +                               offset, skip_start_bytes,
> +                               skip_end_bytes);
> +        if (ret < 0) {
> +            return ret;
> +        }
> +        if (m_data) {
> +            m_data->valid = 1;
> +            m_data->l1_index = l1_index;
> +            m_data->l2_index = l2_index;
> +            m_data->l2_offset = l2_offset;
> +            m_data->l2_cache_entry = &l2_table[l2_index];
> +            m_data->nb_clusters = nb_clusters;
>          }
>      }
> -    if (m_data->l2_cache_entry) {
> -        *m_data->l2_cache_entry = offset;
> +    *cluster_offset = cluster_sector << BDRV_SECTOR_BITS;
> +    return VMDK_OK;
> +}
> +
> +/*
> + * vmdk_alloc_cluster_offset

Maybe just name it "vmdk_alloc_clusters", which sounds better to me? Because the
clusters are what we allocate here, it's rather
"vmdk_alloc_clusters_and_get_offset" but we probably don't want it that long.

> + *
> + * For a given offset on the virtual disk, find the cluster offset in vmdk
> + * file. If the offset is not found, allocate a new cluster.
> + *
> + * If the cluster is newly allocated, m_data->nb_clusters is set to the number
> + * of contiguous clusters that have been allocated. In this case, the other
> + * fields of m_data are valid and contain information about the first allocated
> + * cluster.
> + *
> + * Returns:
> + *
> + *   VMDK_OK:           on success and @cluster_offset was set
> + *
> + *   VMDK_UNALLOC:      if no clusters were allocated and @cluster_offset is
> + *                      set to zero
> + *
> + *   VMDK_ERROR:        in error cases

Thank you for adding the function documentations!

> + */
> +static int vmdk_alloc_cluster_offset(BlockDriverState *bs,
> +                                     VmdkExtent *extent,
> +                                     VmdkMetaData *m_data, uint64_t offset,
> +                                     bool allocate, uint64_t *cluster_offset,
> +                                     int64_t bytes,
> +                                     uint32_t *total_alloc_clusters)
> +{
> +    uint64_t start, remaining;
> +    uint64_t new_cluster_offset;
> +    int64_t n_bytes;
> +    int ret;
> +
> +    if (extent->flat) {
> +        *cluster_offset = extent->flat_start_offset;
> +        return VMDK_OK;
> +    }
> +
> +    start = offset;
> +    remaining = bytes;
> +    new_cluster_offset = 0;
> +    *cluster_offset = 0;
> +    n_bytes = 0;
> +    if (m_data) {
> +        m_data->valid = 0;
> +    }
> +
> +    /* due to L2 table margins all bytes may not get allocated at once */
> +    while (true) {
> +
> +        if (!*cluster_offset) {
> +            *cluster_offset = new_cluster_offset;
> +        }
> +
> +        start              += n_bytes;
> +        remaining          -= n_bytes;

Here, in the first iteration, remaining == bytes and n_bytes == 0.

> +        new_cluster_offset += n_bytes;
> +
> +        if (remaining == 0) {
> +            break;
> +        }
> +
> +        n_bytes = remaining;

Then n_bytes becomes bytes;

In the second iteration, remaining is always 0 because of "remaining -=
n_bytes". What's the point of the while loop?

> +
> +        ret = handle_alloc(bs, extent, start, &new_cluster_offset, &n_bytes,
> +                           m_data, allocate, total_alloc_clusters);
> +
> +        if (ret < 0) {
> +            return ret;
> +
> +        }
>      }
>  
>      return VMDK_OK;
> @@ -1567,6 +1842,7 @@ static int vmdk_pwritev(BlockDriverState *bs, uint64_t offset,
>      uint64_t cluster_offset;
>      uint64_t bytes_done = 0;
>      VmdkMetaData m_data;
> +    uint32_t total_alloc_clusters = 0;
>  
>      if (DIV_ROUND_UP(offset, BDRV_SECTOR_SIZE) > bs->total_sectors) {
>          error_report("Wrong offset: offset=0x%" PRIx64
> @@ -1584,10 +1860,10 @@ static int vmdk_pwritev(BlockDriverState *bs, uint64_t offset,
>          n_bytes = MIN(bytes, extent->cluster_sectors * BDRV_SECTOR_SIZE
>                               - offset_in_cluster);
>  
> -        ret = get_cluster_offset(bs, extent, &m_data, offset,
> -                                 !(extent->compressed || zeroed),
> -                                 &cluster_offset, offset_in_cluster,
> -                                 offset_in_cluster + n_bytes);
> +        ret = vmdk_alloc_cluster_offset(bs, extent, &m_data, offset,
> +                                        !(extent->compressed || zeroed),
> +                                        &cluster_offset, n_bytes,
> +                                        &total_alloc_clusters);
>          if (extent->compressed) {
>              if (ret == VMDK_OK) {
>                  /* Refuse write to allocated cluster for streamOptimized */
> @@ -1596,19 +1872,22 @@ static int vmdk_pwritev(BlockDriverState *bs, uint64_t offset,
>                  return -EIO;
>              } else {
>                  /* allocate */
> -                ret = get_cluster_offset(bs, extent, &m_data, offset,
> -                                         true, &cluster_offset, 0, 0);
> +                ret = vmdk_alloc_cluster_offset(bs, extent, &m_data, offset,
> +                                        true, &cluster_offset, n_bytes,
> +                                        &total_alloc_clusters);

Parameter list is no longer aligned now.

>              }
>          }
>          if (ret == VMDK_ERROR) {
>              return -EINVAL;
>          }
> +
>          if (zeroed) {
>              /* Do zeroed write, buf is ignored */
> -            if (extent->has_zero_grain &&
> -                    offset_in_cluster == 0 &&
> -                    n_bytes >= extent->cluster_sectors * BDRV_SECTOR_SIZE) {
> -                n_bytes = extent->cluster_sectors * BDRV_SECTOR_SIZE;
> +            if (extent->has_zero_grain && offset_in_cluster == 0 &&
> +                    n_bytes >= extent->cluster_sectors * BDRV_SECTOR_SIZE *
> +                        total_alloc_clusters) {
> +                n_bytes = extent->cluster_sectors * BDRV_SECTOR_SIZE *
> +                                        total_alloc_clusters;
>                  if (!zero_dry_run) {
>                      /* update L2 tables */
>                      if (vmdk_L2update(extent, &m_data, VMDK_GTE_ZEROED)
> -- 
> 2.6.2
>