xfs: factor buffer reading from xfs_dir2_leaf_getdents
authorDave Chinner <dchinner@redhat.com>
Fri, 22 Jun 2012 08:50:15 +0000 (18:50 +1000)
committerBen Myers <bpm@sgi.com>
Sun, 1 Jul 2012 19:50:08 +0000 (14:50 -0500)
The buffer reading code in xfs_dir2_leaf_getdents is complex and difficult to
follow due to the readahead and all the context is carries. it is also badly
indented and so difficult to read. Factor it out into a separate function to
make it easier to understand and optimise in future patches.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ben Myers <bpm@sgi.com>
fs/xfs/xfs_dir2_leaf.c

index 69accf6..0b29625 100644 (file)
@@ -759,6 +759,218 @@ xfs_dir2_leaf_compact_x1(
        *highstalep = highstale;
 }
 
+struct xfs_dir2_leaf_map_info {
+       xfs_extlen_t    map_blocks;     /* number of fsbs in map */
+       xfs_dablk_t     map_off;        /* last mapped file offset */
+       int             map_size;       /* total entries in *map */
+       int             map_valid;      /* valid entries in *map */
+       int             nmap;           /* mappings to ask xfs_bmapi */
+       xfs_dir2_db_t   curdb;          /* db for current block */
+       int             ra_current;     /* number of read-ahead blks */
+       int             ra_index;       /* *map index for read-ahead */
+       int             ra_offset;      /* map entry offset for ra */
+       int             ra_want;        /* readahead count wanted */
+       struct xfs_bmbt_irec map[];     /* map vector for blocks */
+};
+
+STATIC int
+xfs_dir2_leaf_readbuf(
+       struct xfs_inode        *dp,
+       size_t                  bufsize,
+       struct xfs_dir2_leaf_map_info *mip,
+       xfs_dir2_off_t          *curoff,
+       struct xfs_buf          **bpp)
+{
+       struct xfs_mount        *mp = dp->i_mount;
+       struct xfs_buf          *bp = *bpp;
+       struct xfs_bmbt_irec    *map = mip->map;
+       int                     error = 0;
+       int                     length;
+       int                     i;
+       int                     j;
+
+       /*
+        * If we have a buffer, we need to release it and
+        * take it out of the mapping.
+        */
+
+       if (bp) {
+               xfs_trans_brelse(NULL, bp);
+               bp = NULL;
+               mip->map_blocks -= mp->m_dirblkfsbs;
+               /*
+                * Loop to get rid of the extents for the
+                * directory block.
+                */
+               for (i = mp->m_dirblkfsbs; i > 0; ) {
+                       j = min_t(int, map->br_blockcount, i);
+                       map->br_blockcount -= j;
+                       map->br_startblock += j;
+                       map->br_startoff += j;
+                       /*
+                        * If mapping is done, pitch it from
+                        * the table.
+                        */
+                       if (!map->br_blockcount && --mip->map_valid)
+                               memmove(&map[0], &map[1],
+                                       sizeof(map[0]) * mip->map_valid);
+                       i -= j;
+               }
+       }
+
+       /*
+        * Recalculate the readahead blocks wanted.
+        */
+       mip->ra_want = howmany(bufsize + mp->m_dirblksize,
+                              mp->m_sb.sb_blocksize) - 1;
+       ASSERT(mip->ra_want >= 0);
+
+       /*
+        * If we don't have as many as we want, and we haven't
+        * run out of data blocks, get some more mappings.
+        */
+       if (1 + mip->ra_want > mip->map_blocks &&
+           mip->map_off < xfs_dir2_byte_to_da(mp, XFS_DIR2_LEAF_OFFSET)) {
+               /*
+                * Get more bmaps, fill in after the ones
+                * we already have in the table.
+                */
+               mip->nmap = mip->map_size - mip->map_valid;
+               error = xfs_bmapi_read(dp, mip->map_off,
+                               xfs_dir2_byte_to_da(mp, XFS_DIR2_LEAF_OFFSET) -
+                                                               mip->map_off,
+                               &map[mip->map_valid], &mip->nmap, 0);
+
+               /*
+                * Don't know if we should ignore this or try to return an
+                * error.  The trouble with returning errors is that readdir
+                * will just stop without actually passing the error through.
+                */
+               if (error)
+                       goto out;       /* XXX */
+
+               /*
+                * If we got all the mappings we asked for, set the final map
+                * offset based on the last bmap value received.  Otherwise,
+                * we've reached the end.
+                */
+               if (mip->nmap == mip->map_size - mip->map_valid) {
+                       i = mip->map_valid + mip->nmap - 1;
+                       mip->map_off = map[i].br_startoff + map[i].br_blockcount;
+               } else
+                       mip->map_off = xfs_dir2_byte_to_da(mp,
+                                                       XFS_DIR2_LEAF_OFFSET);
+
+               /*
+                * Look for holes in the mapping, and eliminate them.  Count up
+                * the valid blocks.
+                */
+               for (i = mip->map_valid; i < mip->map_valid + mip->nmap; ) {
+                       if (map[i].br_startblock == HOLESTARTBLOCK) {
+                               mip->nmap--;
+                               length = mip->map_valid + mip->nmap - i;
+                               if (length)
+                                       memmove(&map[i], &map[i + 1],
+                                               sizeof(map[i]) * length);
+                       } else {
+                               mip->map_blocks += map[i].br_blockcount;
+                               i++;
+                       }
+               }
+               mip->map_valid += mip->nmap;
+       }
+
+       /*
+        * No valid mappings, so no more data blocks.
+        */
+       if (!mip->map_valid) {
+               *curoff = xfs_dir2_da_to_byte(mp, mip->map_off);
+               goto out;
+       }
+
+       /*
+        * Read the directory block starting at the first mapping.
+        */
+       mip->curdb = xfs_dir2_da_to_db(mp, map->br_startoff);
+       error = xfs_da_read_buf(NULL, dp, map->br_startoff,
+                       map->br_blockcount >= mp->m_dirblkfsbs ?
+                           XFS_FSB_TO_DADDR(mp, map->br_startblock) : -1,
+                       &bp, XFS_DATA_FORK);
+
+       /*
+        * Should just skip over the data block instead of giving up.
+        */
+       if (error)
+               goto out;       /* XXX */
+
+       /*
+        * Adjust the current amount of read-ahead: we just read a block that
+        * was previously ra.
+        */
+       if (mip->ra_current)
+               mip->ra_current -= mp->m_dirblkfsbs;
+
+       /*
+        * Do we need more readahead?
+        */
+       for (mip->ra_index = mip->ra_offset = i = 0;
+            mip->ra_want > mip->ra_current && i < mip->map_blocks;
+            i += mp->m_dirblkfsbs) {
+               ASSERT(mip->ra_index < mip->map_valid);
+               /*
+                * Read-ahead a contiguous directory block.
+                */
+               if (i > mip->ra_current &&
+                   map[mip->ra_index].br_blockcount >= mp->m_dirblkfsbs) {
+                       xfs_buf_readahead(mp->m_ddev_targp,
+                               XFS_FSB_TO_DADDR(mp,
+                                       map[mip->ra_index].br_startblock +
+                                                       mip->ra_offset),
+                               (int)BTOBB(mp->m_dirblksize));
+                       mip->ra_current = i;
+               }
+
+               /*
+                * Read-ahead a non-contiguous directory block.  This doesn't
+                * use our mapping, but this is a very rare case.
+                */
+               else if (i > mip->ra_current) {
+                       xfs_da_reada_buf(NULL, dp,
+                                       map[mip->ra_index].br_startoff +
+                                                       mip->ra_offset,
+                                       XFS_DATA_FORK);
+                       mip->ra_current = i;
+               }
+
+               /*
+                * Advance offset through the mapping table.
+                */
+               for (j = 0; j < mp->m_dirblkfsbs; j++) {
+                       /*
+                        * The rest of this extent but not more than a dir
+                        * block.
+                        */
+                       length = min_t(int, mp->m_dirblkfsbs,
+                                       map[mip->ra_index].br_blockcount -
+                                                       mip->ra_offset);
+                       j += length;
+                       mip->ra_offset += length;
+
+                       /*
+                        * Advance to the next mapping if this one is used up.
+                        */
+                       if (mip->ra_offset == map[mip->ra_index].br_blockcount) {
+                               mip->ra_offset = 0;
+                               mip->ra_index++;
+                       }
+               }
+       }
+
+out:
+       *bpp = bp;
+       return error;
+}
+
 /*
  * Getdents (readdir) for leaf and node directories.
  * This reads the data blocks only, so is the same for both forms.
@@ -771,30 +983,18 @@ xfs_dir2_leaf_getdents(
        xfs_off_t               *offset,
        filldir_t               filldir)
 {
-       struct xfs_buf          *bp;            /* data block buffer */
-       int                     byteoff;        /* offset in current block */
-       xfs_dir2_db_t           curdb;          /* db for current block */
-       xfs_dir2_off_t          curoff;         /* current overall offset */
+       struct xfs_buf          *bp = NULL;     /* data block buffer */
        xfs_dir2_data_hdr_t     *hdr;           /* data block header */
        xfs_dir2_data_entry_t   *dep;           /* data entry */
        xfs_dir2_data_unused_t  *dup;           /* unused entry */
        int                     error = 0;      /* error return value */
-       int                     i;              /* temporary loop index */
-       int                     j;              /* temporary loop index */
        int                     length;         /* temporary length value */
-       xfs_bmbt_irec_t         *map;           /* map vector for blocks */
-       xfs_extlen_t            map_blocks;     /* number of fsbs in map */
-       xfs_dablk_t             map_off;        /* last mapped file offset */
-       int                     map_size;       /* total entries in *map */
-       int                     map_valid;      /* valid entries in *map */
        xfs_mount_t             *mp;            /* filesystem mount point */
+       int                     byteoff;        /* offset in current block */
+       xfs_dir2_off_t          curoff;         /* current overall offset */
        xfs_dir2_off_t          newoff;         /* new curoff after new blk */
-       int                     nmap;           /* mappings to ask xfs_bmapi */
        char                    *ptr = NULL;    /* pointer to current data */
-       int                     ra_current;     /* number of read-ahead blks */
-       int                     ra_index;       /* *map index for read-ahead */
-       int                     ra_offset;      /* map entry offset for ra */
-       int                     ra_want;        /* readahead count wanted */
+       struct xfs_dir2_leaf_map_info *map_info;
 
        /*
         * If the offset is at or past the largest allowed value,
@@ -810,10 +1010,12 @@ xfs_dir2_leaf_getdents(
         * buffer size, the directory block size, and the filesystem
         * block size.
         */
-       map_size = howmany(bufsize + mp->m_dirblksize, mp->m_sb.sb_blocksize);
-       map = kmem_alloc(map_size * sizeof(*map), KM_SLEEP);
-       map_valid = ra_index = ra_offset = ra_current = map_blocks = 0;
-       bp = NULL;
+       length = howmany(bufsize + mp->m_dirblksize,
+                                    mp->m_sb.sb_blocksize);
+       map_info = kmem_zalloc(offsetof(struct xfs_dir2_leaf_map_info, map) +
+                               (length * sizeof(struct xfs_bmbt_irec)),
+                              KM_SLEEP);
+       map_info->map_size = length;
 
        /*
         * Inside the loop we keep the main offset value as a byte offset
@@ -825,7 +1027,9 @@ xfs_dir2_leaf_getdents(
         * Force this conversion through db so we truncate the offset
         * down to get the start of the data block.
         */
-       map_off = xfs_dir2_db_to_da(mp, xfs_dir2_byte_to_db(mp, curoff));
+       map_info->map_off = xfs_dir2_db_to_da(mp,
+                                             xfs_dir2_byte_to_db(mp, curoff));
+
        /*
         * Loop over directory entries until we reach the end offset.
         * Get more blocks and readahead as necessary.
@@ -836,190 +1040,16 @@ xfs_dir2_leaf_getdents(
                 * current buffer, need to get another one.
                 */
                if (!bp || ptr >= (char *)bp->b_addr + mp->m_dirblksize) {
-                       /*
-                        * If we have a buffer, we need to release it and
-                        * take it out of the mapping.
-                        */
-                       if (bp) {
-                               xfs_trans_brelse(NULL, bp);
-                               bp = NULL;
-                               map_blocks -= mp->m_dirblkfsbs;
-                               /*
-                                * Loop to get rid of the extents for the
-                                * directory block.
-                                */
-                               for (i = mp->m_dirblkfsbs; i > 0; ) {
-                                       j = MIN((int)map->br_blockcount, i);
-                                       map->br_blockcount -= j;
-                                       map->br_startblock += j;
-                                       map->br_startoff += j;
-                                       /*
-                                        * If mapping is done, pitch it from
-                                        * the table.
-                                        */
-                                       if (!map->br_blockcount && --map_valid)
-                                               memmove(&map[0], &map[1],
-                                                       sizeof(map[0]) *
-                                                       map_valid);
-                                       i -= j;
-                               }
-                       }
-                       /*
-                        * Recalculate the readahead blocks wanted.
-                        */
-                       ra_want = howmany(bufsize + mp->m_dirblksize,
-                                         mp->m_sb.sb_blocksize) - 1;
-                       ASSERT(ra_want >= 0);
 
-                       /*
-                        * If we don't have as many as we want, and we haven't
-                        * run out of data blocks, get some more mappings.
-                        */
-                       if (1 + ra_want > map_blocks &&
-                           map_off <
-                           xfs_dir2_byte_to_da(mp, XFS_DIR2_LEAF_OFFSET)) {
-                               /*
-                                * Get more bmaps, fill in after the ones
-                                * we already have in the table.
-                                */
-                               nmap = map_size - map_valid;
-                               error = xfs_bmapi_read(dp, map_off,
-                                       xfs_dir2_byte_to_da(mp,
-                                               XFS_DIR2_LEAF_OFFSET) - map_off,
-                                       &map[map_valid], &nmap, 0);
-                               /*
-                                * Don't know if we should ignore this or
-                                * try to return an error.
-                                * The trouble with returning errors
-                                * is that readdir will just stop without
-                                * actually passing the error through.
-                                */
-                               if (error)
-                                       break;  /* XXX */
-                               /*
-                                * If we got all the mappings we asked for,
-                                * set the final map offset based on the
-                                * last bmap value received.
-                                * Otherwise, we've reached the end.
-                                */
-                               if (nmap == map_size - map_valid)
-                                       map_off =
-                                       map[map_valid + nmap - 1].br_startoff +
-                                       map[map_valid + nmap - 1].br_blockcount;
-                               else
-                                       map_off =
-                                               xfs_dir2_byte_to_da(mp,
-                                                       XFS_DIR2_LEAF_OFFSET);
-                               /*
-                                * Look for holes in the mapping, and
-                                * eliminate them.  Count up the valid blocks.
-                                */
-                               for (i = map_valid; i < map_valid + nmap; ) {
-                                       if (map[i].br_startblock ==
-                                           HOLESTARTBLOCK) {
-                                               nmap--;
-                                               length = map_valid + nmap - i;
-                                               if (length)
-                                                       memmove(&map[i],
-                                                               &map[i + 1],
-                                                               sizeof(map[i]) *
-                                                               length);
-                                       } else {
-                                               map_blocks +=
-                                                       map[i].br_blockcount;
-                                               i++;
-                                       }
-                               }
-                               map_valid += nmap;
-                       }
-                       /*
-                        * No valid mappings, so no more data blocks.
-                        */
-                       if (!map_valid) {
-                               curoff = xfs_dir2_da_to_byte(mp, map_off);
+                       error = xfs_dir2_leaf_readbuf(dp, bufsize, map_info,
+                                                     &curoff, &bp);
+                       if (error || !map_info->map_valid)
                                break;
-                       }
-                       /*
-                        * Read the directory block starting at the first
-                        * mapping.
-                        */
-                       curdb = xfs_dir2_da_to_db(mp, map->br_startoff);
-                       error = xfs_da_read_buf(NULL, dp, map->br_startoff,
-                               map->br_blockcount >= mp->m_dirblkfsbs ?
-                                   XFS_FSB_TO_DADDR(mp, map->br_startblock) :
-                                   -1,
-                               &bp, XFS_DATA_FORK);
-                       /*
-                        * Should just skip over the data block instead
-                        * of giving up.
-                        */
-                       if (error)
-                               break;  /* XXX */
-                       /*
-                        * Adjust the current amount of read-ahead: we just
-                        * read a block that was previously ra.
-                        */
-                       if (ra_current)
-                               ra_current -= mp->m_dirblkfsbs;
-                       /*
-                        * Do we need more readahead?
-                        */
-                       for (ra_index = ra_offset = i = 0;
-                            ra_want > ra_current && i < map_blocks;
-                            i += mp->m_dirblkfsbs) {
-                               ASSERT(ra_index < map_valid);
-                               /*
-                                * Read-ahead a contiguous directory block.
-                                */
-                               if (i > ra_current &&
-                                   map[ra_index].br_blockcount >=
-                                   mp->m_dirblkfsbs) {
-                                       xfs_buf_readahead(mp->m_ddev_targp,
-                                               XFS_FSB_TO_DADDR(mp,
-                                                  map[ra_index].br_startblock +
-                                                  ra_offset),
-                                               (int)BTOBB(mp->m_dirblksize));
-                                       ra_current = i;
-                               }
-                               /*
-                                * Read-ahead a non-contiguous directory block.
-                                * This doesn't use our mapping, but this
-                                * is a very rare case.
-                                */
-                               else if (i > ra_current) {
-                                       (void)xfs_da_reada_buf(NULL, dp,
-                                               map[ra_index].br_startoff +
-                                               ra_offset, XFS_DATA_FORK);
-                                       ra_current = i;
-                               }
-                               /*
-                                * Advance offset through the mapping table.
-                                */
-                               for (j = 0; j < mp->m_dirblkfsbs; j++) {
-                                       /*
-                                        * The rest of this extent but not
-                                        * more than a dir block.
-                                        */
-                                       length = MIN(mp->m_dirblkfsbs,
-                                               (int)(map[ra_index].br_blockcount -
-                                               ra_offset));
-                                       j += length;
-                                       ra_offset += length;
-                                       /*
-                                        * Advance to the next mapping if
-                                        * this one is used up.
-                                        */
-                                       if (ra_offset ==
-                                           map[ra_index].br_blockcount) {
-                                               ra_offset = 0;
-                                               ra_index++;
-                                       }
-                               }
-                       }
+
                        /*
                         * Having done a read, we need to set a new offset.
                         */
-                       newoff = xfs_dir2_db_off_to_byte(mp, curdb, 0);
+                       newoff = xfs_dir2_db_off_to_byte(mp, map_info->curdb, 0);
                        /*
                         * Start of the current block.
                         */
@@ -1030,7 +1060,7 @@ xfs_dir2_leaf_getdents(
                         */
                        else if (curoff > newoff)
                                ASSERT(xfs_dir2_byte_to_db(mp, curoff) ==
-                                      curdb);
+                                      map_info->curdb);
                        hdr = bp->b_addr;
                        xfs_dir2_data_check(dp, bp);
                        /*
@@ -1113,7 +1143,7 @@ xfs_dir2_leaf_getdents(
                *offset = XFS_DIR2_MAX_DATAPTR & 0x7fffffff;
        else
                *offset = xfs_dir2_byte_to_dataptr(mp, curoff) & 0x7fffffff;
-       kmem_free(map);
+       kmem_free(map_info);
        if (bp)
                xfs_trans_brelse(NULL, bp);
        return error;