Re: [PATCH v2 4/9] exfat: add support for multi-cluster allocation

[CharSyam]

hi, Namjae.

The critical issue is that this hunk drops the hole-care walk in
  exfat_map_cluster():

                  *clu = new_clu.dir;
  +               *count = new_clu.size;

  -               inode->i_blocks +=
  -                       exfat_cluster_to_sectors(sbi,
num_to_be_allocated) >> 9;
  -
  -               /*
  -                * Move *clu pointer along FAT chains (hole care) because the
  -                * caller of this function expect *clu to be the last cluster.
  -                * This only works when num_to_be_allocated >= 2,
  -                * *clu = (the first cluster of the allocated chain) =>
  -                * (the last cluster of ...)
  -                */
  -               if (exfat_cluster_walk(sb, clu, num_to_be_allocated
- 1, ei->flags))
  -                       return -EIO;
  -               *count = 1;
  +               inode->i_blocks += exfat_cluster_to_sectors(sbi,
new_clu.size);

  The removed comment is explicit: the caller expects *clu to be the
  cluster at the requested clu_offset, not necessarily the first cluster
  of the newly allocated chain. The old code walked
  num_to_be_allocated - 1 steps so that *clu ended up at clu_offset.

  With this change, *clu is left as new_clu.dir (the first newly
  allocated cluster) and *count is set to new_clu.size. That is only
  correct for the sequential-extension case where
  clu_offset == num_clusters. For clu_offset > num_clusters, the caller
  will compute the physical sector from the wrong cluster:

        phys = exfat_cluster_to_sector(sbi, cluster) + sec_offset;

  This also interacts badly with the new contig=true allocation mode.
  If exfat_alloc_cluster() stops early at a non-contiguous boundary,
  new_clu.size can be smaller than num_to_be_allocated, so the allocated
  run may not even reach clu_offset while exfat_map_cluster() still
  returns success.

  Could you keep the old hole-care behavior and extend it to the
  multi-cluster case? After allocation, the code should advance from
  new_clu.dir to the cluster corresponding to clu_offset, set *count to
  the remaining number of clusters from that point, and handle partial
  contiguous allocation by either continuing allocation or unwinding it
  before returning an error.