cache-extent-hole-in-extent-status-tree-for-ext4_da_map_blocks

   1 ext4: cache extent hole in extent status tree for ext4_da_map_blocks()
   2
   3 From: Zheng Liu <wenqing.lz@taobao.com>
   4
   5 Currently extent status tree doesn't cache extent hole when a write
   6 looks up in extent tree to make sure whether a block has been allocated
   7 or not.  In this case, we don't put extent hole in extent cache because
   8 later this extent might be removed and a new delayed extent might be
   9 added back.  But it will cause a defect when we do a lot of writes.  If
  10 we don't put extent hole in extent cache, the following writes also need
  11 to access extent tree to look at whether or not a block has been
  12 allocated.  It brings a cache miss.  This commit fixes this defect.
  13 Also if the inode doesn't have any extent, this extent hole will be
  14 cached as well.
  15
  16 Cc: Andreas Dilger <adilger.kernel@dilger.ca>
  17 Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
  18 Signed-off-by: Jan Kara <jack@suse.cz>
  19 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  20 ---
  21  fs/ext4/ext4.h              |  4 +---
  22  fs/ext4/extents.c           | 31 ++++++++++++++++---------------
  23  fs/ext4/inode.c             |  6 ++----
  24  include/trace/events/ext4.h |  3 +--
  25  4 files changed, 20 insertions(+), 24 deletions(-)
  26
  27 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  28 index 7b3f3b1..98da4cd 100644
  29 --- a/fs/ext4/ext4.h
  30 +++ b/fs/ext4/ext4.h
  31 @@ -556,10 +556,8 @@ enum {
  32  #define EXT4_GET_BLOCKS_KEEP_SIZE              0x0080
  33         /* Do not take i_data_sem locking in ext4_map_blocks */
  34  #define EXT4_GET_BLOCKS_NO_LOCK                        0x0100
  35 -       /* Do not put hole in extent cache */
  36 -#define EXT4_GET_BLOCKS_NO_PUT_HOLE            0x0200
  37         /* Convert written extents to unwritten */
  38 -#define EXT4_GET_BLOCKS_CONVERT_UNWRITTEN      0x0400
  39 +#define EXT4_GET_BLOCKS_CONVERT_UNWRITTEN      0x0200
  40
  41  /*
  42   * The bit position of these flags must not overlap with any of the
  43 diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
  44 index 7ef2f11..1ee24d7 100644
  45 --- a/fs/ext4/extents.c
  46 +++ b/fs/ext4/extents.c
  47 @@ -2306,16 +2306,16 @@ ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
  48                                 ext4_lblk_t block)
  49  {
  50         int depth = ext_depth(inode);
  51 -       unsigned long len = 0;
  52 -       ext4_lblk_t lblock = 0;
  53 +       ext4_lblk_t len;
  54 +       ext4_lblk_t lblock;
  55         struct ext4_extent *ex;
  56 +       struct extent_status es;
  57
  58         ex = path[depth].p_ext;
  59         if (ex == NULL) {
  60 -               /*
  61 -                * there is no extent yet, so gap is [0;-] and we
  62 -                * don't cache it
  63 -                */
  64 +               /* there is no extent yet, so gap is [0;-] */
  65 +               lblock = 0;
  66 +               len = EXT_MAX_BLOCKS;
  67                 ext_debug("cache gap(whole file):");
  68         } else if (block < le32_to_cpu(ex->ee_block)) {
  69                 lblock = block;
  70 @@ -2324,9 +2324,6 @@ ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
  71                                 block,
  72                                 le32_to_cpu(ex->ee_block),
  73                                  ext4_ext_get_actual_len(ex));
  74 -               if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
  75 -                       ext4_es_insert_extent(inode, lblock, len, ~0,
  76 -                                             EXTENT_STATUS_HOLE);
  77         } else if (block >= le32_to_cpu(ex->ee_block)
  78                         + ext4_ext_get_actual_len(ex)) {
  79                 ext4_lblk_t next;
  80 @@ -2340,14 +2337,19 @@ ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
  81                                 block);
  82                 BUG_ON(next == lblock);
  83                 len = next - lblock;
  84 -               if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
  85 -                       ext4_es_insert_extent(inode, lblock, len, ~0,
  86 -                                             EXTENT_STATUS_HOLE);
  87         } else {
  88                 BUG();
  89         }
  90
  91 -       ext_debug(" -> %u:%lu\n", lblock, len);
  92 +       ext4_es_find_delayed_extent_range(inode, lblock, lblock + len - 1, &es);
  93 +       if (es.es_len) {
  94 +               /* There's delayed extent containing lblock? */
  95 +               if (es.es_lblk <= lblock)
  96 +                       return;
  97 +               len = min(es.es_lblk - lblock, len);
  98 +       }
  99 +       ext_debug(" -> %u:%u\n", lblock, len);
 100 +       ext4_es_insert_extent(inode, lblock, len, ~0, EXTENT_STATUS_HOLE);
 101  }
 102
 103  /*
 104 @@ -4368,8 +4370,7 @@ int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
 105                  * put just found gap into cache to speed up
 106                  * subsequent requests
 107                  */
 108 -               if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
 109 -                       ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
 110 +               ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
 111                 goto out2;
 112         }
 113
 114 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
 115 index ddd8e43..7bd272f 100644
 116 --- a/fs/ext4/inode.c
 117 +++ b/fs/ext4/inode.c
 118 @@ -1432,11 +1432,9 @@ static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
 119         if (ext4_has_inline_data(inode))
 120                 retval = 0;
 121         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 122 -               retval = ext4_ext_map_blocks(NULL, inode, map,
 123 -                                            EXT4_GET_BLOCKS_NO_PUT_HOLE);
 124 +               retval = ext4_ext_map_blocks(NULL, inode, map, 0);
 125         else
 126 -               retval = ext4_ind_map_blocks(NULL, inode, map,
 127 -                                            EXT4_GET_BLOCKS_NO_PUT_HOLE);
 128 +               retval = ext4_ind_map_blocks(NULL, inode, map, 0);
 129
 130  add_delayed:
 131         if (retval == 0) {
 132 diff --git a/include/trace/events/ext4.h b/include/trace/events/ext4.h
 133 index 06343e1..d71b00d3 100644
 134 --- a/include/trace/events/ext4.h
 135 +++ b/include/trace/events/ext4.h
 136 @@ -43,8 +43,7 @@ struct extent_status;
 137         { EXT4_GET_BLOCKS_METADATA_NOFAIL,      "METADATA_NOFAIL" },    \
 138         { EXT4_GET_BLOCKS_NO_NORMALIZE,         "NO_NORMALIZE" },       \
 139         { EXT4_GET_BLOCKS_KEEP_SIZE,            "KEEP_SIZE" },          \
 140 -       { EXT4_GET_BLOCKS_NO_LOCK,              "NO_LOCK" },            \
 141 -       { EXT4_GET_BLOCKS_NO_PUT_HOLE,          "NO_PUT_HOLE" })
 142 +       { EXT4_GET_BLOCKS_NO_LOCK,              "NO_LOCK" })
 143
 144  #define show_mflags(flags) __print_flags(flags, "",    \
 145         { EXT4_MAP_NEW,         "N" },                  \