implement-allocation-of-pre-zeroed-blocks

   1 ext4: implement allocation of pre-zeroed blocks
   2
   3 From: Jan Kara <jack@suse.com>
   4
   5 DAX page fault path needs to get blocks that are pre-zeroed to avoid
   6 races when two concurrent page faults happen in the same block of a
   7 file. Implement support for this in ext4_map_blocks().
   8
   9 Signed-off-by: Jan Kara <jack@suse.com>
  10 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  11 ---
  12  fs/ext4/ext4.h              |  4 ++++
  13  fs/ext4/extents.c           |  8 ++++++++
  14  fs/ext4/inode.c             | 20 +++++++++++++++++---
  15  include/trace/events/ext4.h |  3 ++-
  16  4 files changed, 31 insertions(+), 4 deletions(-)
  17
  18 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  19 index 785737a342e6..0a11a34d54c9 100644
  20 --- a/fs/ext4/ext4.h
  21 +++ b/fs/ext4/ext4.h
  22 @@ -553,6 +553,10 @@ enum {
  23  #define EXT4_GET_BLOCKS_KEEP_SIZE              0x0080
  24         /* Convert written extents to unwritten */
  25  #define EXT4_GET_BLOCKS_CONVERT_UNWRITTEN      0x0100
  26 +       /* Write zeros to newly created written extents */
  27 +#define EXT4_GET_BLOCKS_ZERO                   0x0200
  28 +#define EXT4_GET_BLOCKS_CREATE_ZERO            (EXT4_GET_BLOCKS_CREATE |\
  29 +                                       EXT4_GET_BLOCKS_ZERO)
  30
  31  /*
  32   * The bit position of these flags must not overlap with any of the
  33 diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
  34 index 273bf40cca23..8e0d853a8044 100644
  35 --- a/fs/ext4/extents.c
  36 +++ b/fs/ext4/extents.c
  37 @@ -4040,6 +4040,14 @@ ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
  38         }
  39         /* IO end_io complete, convert the filled extent to written */
  40         if (flags & EXT4_GET_BLOCKS_CONVERT) {
  41 +               if (flags & EXT4_GET_BLOCKS_ZERO) {
  42 +                       if (allocated > map->m_len)
  43 +                               allocated = map->m_len;
  44 +                       err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
  45 +                                                allocated);
  46 +                       if (err < 0)
  47 +                               goto out2;
  48 +               }
  49                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
  50                                                            ppath);
  51                 if (ret >= 0) {
  52 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
  53 index 5e5795bde6b4..73c6214084d4 100644
  54 --- a/fs/ext4/inode.c
  55 +++ b/fs/ext4/inode.c
  56 @@ -643,7 +643,7 @@ found:
  57                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
  58                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
  59                         if (ext4_es_is_written(&es))
  60 -                               goto has_zeroout;
  61 +                               goto out_sem;
  62                 }
  63                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
  64                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
  65 @@ -654,11 +654,24 @@ found:
  66                         status |= EXTENT_STATUS_DELAYED;
  67                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
  68                                             map->m_pblk, status);
  69 -               if (ret < 0)
  70 +               if (ret < 0) {
  71                         retval = ret;
  72 +                       goto out_sem;
  73 +               }
  74 +
  75 +               if (flags & EXT4_GET_BLOCKS_ZERO &&
  76 +                   map->m_flags & EXT4_MAP_MAPPED &&
  77 +                   map->m_flags & EXT4_MAP_NEW) {
  78 +                       ret = ext4_issue_zeroout(inode, map->m_lblk,
  79 +                                                map->m_pblk, map->m_len);
  80 +                       if (ret) {
  81 +                               retval = ret;
  82 +                               goto out_sem;
  83 +                       }
  84 +               }
  85         }
  86
  87 -has_zeroout:
  88 +out_sem:
  89         up_write((&EXT4_I(inode)->i_data_sem));
  90         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
  91                 ret = check_block_validity(inode, map);
  92 @@ -3041,6 +3054,7 @@ int ext4_get_block_dax(struct inode *inode, sector_t iblock,
  93                    struct buffer_head *bh_result, int create)
  94  {
  95         int flags = EXT4_GET_BLOCKS_PRE_IO | EXT4_GET_BLOCKS_UNWRIT_EXT;
  96 +
  97         if (create)
  98                 flags |= EXT4_GET_BLOCKS_CREATE;
  99         ext4_debug("ext4_get_block_dax: inode %lu, create flag %d\n",
 100 diff --git a/include/trace/events/ext4.h b/include/trace/events/ext4.h
 101 index 5f2ace56efc5..4e4b2fa78609 100644
 102 --- a/include/trace/events/ext4.h
 103 +++ b/include/trace/events/ext4.h
 104 @@ -42,7 +42,8 @@ struct extent_status;
 105         { EXT4_GET_BLOCKS_CONVERT,              "CONVERT" },            \
 106         { EXT4_GET_BLOCKS_METADATA_NOFAIL,      "METADATA_NOFAIL" },    \
 107         { EXT4_GET_BLOCKS_NO_NORMALIZE,         "NO_NORMALIZE" },       \
 108 -       { EXT4_GET_BLOCKS_KEEP_SIZE,            "KEEP_SIZE" })
 109 +       { EXT4_GET_BLOCKS_KEEP_SIZE,            "KEEP_SIZE" },          \
 110 +       { EXT4_GET_BLOCKS_ZERO,                 "ZERO" })
 111
 112  #define show_mflags(flags) __print_flags(flags, "",    \
 113         { EXT4_MAP_NEW,         "N" },                  \
 114 --
 115 2.1.4
 116
 117 --
 118 To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
 119 the body of a message to majordomo@vger.kernel.org
 120 More majordomo info at  http://vger.kernel.org/majordomo-info.html
 121