write support for preallocated blocks
authorAmit Arora <aarora@in.ibm.com>
Wed, 18 Jul 2007 01:42:38 +0000 (21:42 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Wed, 18 Jul 2007 01:42:38 +0000 (21:42 -0400)
This patch adds write support to the uninitialized extents that get
created when a preallocation is done using fallocate(). It takes care of
splitting the extents into multiple (upto three) extents and merging the
new split extents with neighbouring ones, if possible.

Signed-off-by: Amit Arora <aarora@in.ibm.com>
fs/ext4/extents.c
include/linux/ext4_fs_extents.h

index ba25832..ded3d46 100644 (file)
@@ -1141,6 +1141,53 @@ ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
 }
 
 /*
+ * This function tries to merge the "ex" extent to the next extent in the tree.
+ * It always tries to merge towards right. If you want to merge towards
+ * left, pass "ex - 1" as argument instead of "ex".
+ * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
+ * 1 if they got merged.
+ */
+int ext4_ext_try_to_merge(struct inode *inode,
+                         struct ext4_ext_path *path,
+                         struct ext4_extent *ex)
+{
+       struct ext4_extent_header *eh;
+       unsigned int depth, len;
+       int merge_done = 0;
+       int uninitialized = 0;
+
+       depth = ext_depth(inode);
+       BUG_ON(path[depth].p_hdr == NULL);
+       eh = path[depth].p_hdr;
+
+       while (ex < EXT_LAST_EXTENT(eh)) {
+               if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
+                       break;
+               /* merge with next extent! */
+               if (ext4_ext_is_uninitialized(ex))
+                       uninitialized = 1;
+               ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
+                               + ext4_ext_get_actual_len(ex + 1));
+               if (uninitialized)
+                       ext4_ext_mark_uninitialized(ex);
+
+               if (ex + 1 < EXT_LAST_EXTENT(eh)) {
+                       len = (EXT_LAST_EXTENT(eh) - ex - 1)
+                               * sizeof(struct ext4_extent);
+                       memmove(ex + 1, ex + 2, len);
+               }
+               eh->eh_entries = cpu_to_le16(le16_to_cpu(eh->eh_entries) - 1);
+               merge_done = 1;
+               WARN_ON(eh->eh_entries == 0);
+               if (!eh->eh_entries)
+                       ext4_error(inode->i_sb, "ext4_ext_try_to_merge",
+                          "inode#%lu, eh->eh_entries = 0!", inode->i_ino);
+       }
+
+       return merge_done;
+}
+
+/*
  * check if a portion of the "newext" extent overlaps with an
  * existing extent.
  *
@@ -1328,25 +1375,7 @@ has_space:
 
 merge:
        /* try to merge extents to the right */
-       while (nearex < EXT_LAST_EXTENT(eh)) {
-               if (!ext4_can_extents_be_merged(inode, nearex, nearex + 1))
-                       break;
-               /* merge with next extent! */
-               if (ext4_ext_is_uninitialized(nearex))
-                       uninitialized = 1;
-               nearex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(nearex)
-                                       + ext4_ext_get_actual_len(nearex + 1));
-               if (uninitialized)
-                       ext4_ext_mark_uninitialized(nearex);
-
-               if (nearex + 1 < EXT_LAST_EXTENT(eh)) {
-                       len = (EXT_LAST_EXTENT(eh) - nearex - 1)
-                                       * sizeof(struct ext4_extent);
-                       memmove(nearex + 1, nearex + 2, len);
-               }
-               eh->eh_entries = cpu_to_le16(le16_to_cpu(eh->eh_entries)-1);
-               BUG_ON(eh->eh_entries == 0);
-       }
+       ext4_ext_try_to_merge(inode, path, nearex);
 
        /* try to merge extents to the left */
 
@@ -2012,15 +2041,158 @@ void ext4_ext_release(struct super_block *sb)
 #endif
 }
 
+/*
+ * This function is called by ext4_ext_get_blocks() if someone tries to write
+ * to an uninitialized extent. It may result in splitting the uninitialized
+ * extent into multiple extents (upto three - one initialized and two
+ * uninitialized).
+ * There are three possibilities:
+ *   a> There is no split required: Entire extent should be initialized
+ *   b> Splits in two extents: Write is happening at either end of the extent
+ *   c> Splits in three extents: Somone is writing in middle of the extent
+ */
+int ext4_ext_convert_to_initialized(handle_t *handle, struct inode *inode,
+                                       struct ext4_ext_path *path,
+                                       ext4_fsblk_t iblock,
+                                       unsigned long max_blocks)
+{
+       struct ext4_extent *ex, newex;
+       struct ext4_extent *ex1 = NULL;
+       struct ext4_extent *ex2 = NULL;
+       struct ext4_extent *ex3 = NULL;
+       struct ext4_extent_header *eh;
+       unsigned int allocated, ee_block, ee_len, depth;
+       ext4_fsblk_t newblock;
+       int err = 0;
+       int ret = 0;
+
+       depth = ext_depth(inode);
+       eh = path[depth].p_hdr;
+       ex = path[depth].p_ext;
+       ee_block = le32_to_cpu(ex->ee_block);
+       ee_len = ext4_ext_get_actual_len(ex);
+       allocated = ee_len - (iblock - ee_block);
+       newblock = iblock - ee_block + ext_pblock(ex);
+       ex2 = ex;
+
+       /* ex1: ee_block to iblock - 1 : uninitialized */
+       if (iblock > ee_block) {
+               ex1 = ex;
+               ex1->ee_len = cpu_to_le16(iblock - ee_block);
+               ext4_ext_mark_uninitialized(ex1);
+               ex2 = &newex;
+       }
+       /*
+        * for sanity, update the length of the ex2 extent before
+        * we insert ex3, if ex1 is NULL. This is to avoid temporary
+        * overlap of blocks.
+        */
+       if (!ex1 && allocated > max_blocks)
+               ex2->ee_len = cpu_to_le16(max_blocks);
+       /* ex3: to ee_block + ee_len : uninitialised */
+       if (allocated > max_blocks) {
+               unsigned int newdepth;
+               ex3 = &newex;
+               ex3->ee_block = cpu_to_le32(iblock + max_blocks);
+               ext4_ext_store_pblock(ex3, newblock + max_blocks);
+               ex3->ee_len = cpu_to_le16(allocated - max_blocks);
+               ext4_ext_mark_uninitialized(ex3);
+               err = ext4_ext_insert_extent(handle, inode, path, ex3);
+               if (err)
+                       goto out;
+               /*
+                * The depth, and hence eh & ex might change
+                * as part of the insert above.
+                */
+               newdepth = ext_depth(inode);
+               if (newdepth != depth) {
+                       depth = newdepth;
+                       path = ext4_ext_find_extent(inode, iblock, NULL);
+                       if (IS_ERR(path)) {
+                               err = PTR_ERR(path);
+                               path = NULL;
+                               goto out;
+                       }
+                       eh = path[depth].p_hdr;
+                       ex = path[depth].p_ext;
+                       if (ex2 != &newex)
+                               ex2 = ex;
+               }
+               allocated = max_blocks;
+       }
+       /*
+        * If there was a change of depth as part of the
+        * insertion of ex3 above, we need to update the length
+        * of the ex1 extent again here
+        */
+       if (ex1 && ex1 != ex) {
+               ex1 = ex;
+               ex1->ee_len = cpu_to_le16(iblock - ee_block);
+               ext4_ext_mark_uninitialized(ex1);
+               ex2 = &newex;
+       }
+       /* ex2: iblock to iblock + maxblocks-1 : initialised */
+       ex2->ee_block = cpu_to_le32(iblock);
+       ex2->ee_start = cpu_to_le32(newblock);
+       ext4_ext_store_pblock(ex2, newblock);
+       ex2->ee_len = cpu_to_le16(allocated);
+       if (ex2 != ex)
+               goto insert;
+       err = ext4_ext_get_access(handle, inode, path + depth);
+       if (err)
+               goto out;
+       /*
+        * New (initialized) extent starts from the first block
+        * in the current extent. i.e., ex2 == ex
+        * We have to see if it can be merged with the extent
+        * on the left.
+        */
+       if (ex2 > EXT_FIRST_EXTENT(eh)) {
+               /*
+                * To merge left, pass "ex2 - 1" to try_to_merge(),
+                * since it merges towards right _only_.
+                */
+               ret = ext4_ext_try_to_merge(inode, path, ex2 - 1);
+               if (ret) {
+                       err = ext4_ext_correct_indexes(handle, inode, path);
+                       if (err)
+                               goto out;
+                       depth = ext_depth(inode);
+                       ex2--;
+               }
+       }
+       /*
+        * Try to Merge towards right. This might be required
+        * only when the whole extent is being written to.
+        * i.e. ex2 == ex and ex3 == NULL.
+        */
+       if (!ex3) {
+               ret = ext4_ext_try_to_merge(inode, path, ex2);
+               if (ret) {
+                       err = ext4_ext_correct_indexes(handle, inode, path);
+                       if (err)
+                               goto out;
+               }
+       }
+       /* Mark modified extent as dirty */
+       err = ext4_ext_dirty(handle, inode, path + depth);
+       goto out;
+insert:
+       err = ext4_ext_insert_extent(handle, inode, path, &newex);
+out:
+       return err ? err : allocated;
+}
+
 int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
                        ext4_fsblk_t iblock,
                        unsigned long max_blocks, struct buffer_head *bh_result,
                        int create, int extend_disksize)
 {
        struct ext4_ext_path *path = NULL;
+       struct ext4_extent_header *eh;
        struct ext4_extent newex, *ex;
        ext4_fsblk_t goal, newblock;
-       int err = 0, depth;
+       int err = 0, depth, ret;
        unsigned long allocated = 0;
 
        __clear_bit(BH_New, &bh_result->b_state);
@@ -2033,8 +2205,10 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
        if (goal) {
                if (goal == EXT4_EXT_CACHE_GAP) {
                        if (!create) {
-                               /* block isn't allocated yet and
-                                * user doesn't want to allocate it */
+                               /*
+                                * block isn't allocated yet and
+                                * user doesn't want to allocate it
+                                */
                                goto out2;
                        }
                        /* we should allocate requested block */
@@ -2068,6 +2242,7 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
         * this is why assert can't be put in ext4_ext_find_extent()
         */
        BUG_ON(path[depth].p_ext == NULL && depth != 0);
+       eh = path[depth].p_hdr;
 
        ex = path[depth].p_ext;
        if (ex) {
@@ -2076,13 +2251,9 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
                unsigned short ee_len;
 
                /*
-                * Allow future support for preallocated extents to be added
-                * as an RO_COMPAT feature:
                 * Uninitialized extents are treated as holes, except that
-                * we avoid (fail) allocating new blocks during a write.
+                * we split out initialized portions during a write.
                 */
-               if (le16_to_cpu(ex->ee_len) > EXT_MAX_LEN)
-                       goto out2;
                ee_len = ext4_ext_get_actual_len(ex);
                /* if found extent covers block, simply return it */
                if (iblock >= ee_block && iblock < ee_block + ee_len) {
@@ -2091,12 +2262,27 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
                        allocated = ee_len - (iblock - ee_block);
                        ext_debug("%d fit into %lu:%d -> %llu\n", (int) iblock,
                                        ee_block, ee_len, newblock);
+
                        /* Do not put uninitialized extent in the cache */
-                       if (!ext4_ext_is_uninitialized(ex))
+                       if (!ext4_ext_is_uninitialized(ex)) {
                                ext4_ext_put_in_cache(inode, ee_block,
                                                        ee_len, ee_start,
                                                        EXT4_EXT_CACHE_EXTENT);
-                       goto out;
+                               goto out;
+                       }
+                       if (create == EXT4_CREATE_UNINITIALIZED_EXT)
+                               goto out;
+                       if (!create)
+                               goto out2;
+
+                       ret = ext4_ext_convert_to_initialized(handle, inode,
+                                                               path, iblock,
+                                                               max_blocks);
+                       if (ret <= 0)
+                               goto out2;
+                       else
+                               allocated = ret;
+                       goto outnew;
                }
        }
 
@@ -2105,8 +2291,10 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
         * we couldn't try to create block if create flag is zero
         */
        if (!create) {
-               /* put just found gap into cache to speed up
-                * subsequent requests */
+               /*
+                * put just found gap into cache to speed up
+                * subsequent requests
+                */
                ext4_ext_put_gap_in_cache(inode, path, iblock);
                goto out2;
        }
@@ -2152,6 +2340,7 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
 
        /* previous routine could use block we allocated */
        newblock = ext_pblock(&newex);
+outnew:
        __set_bit(BH_New, &bh_result->b_state);
 
        /* Cache only when it is _not_ an uninitialized extent */
@@ -2221,7 +2410,8 @@ void ext4_ext_truncate(struct inode * inode, struct page *page)
        err = ext4_ext_remove_space(inode, last_block);
 
        /* In a multi-transaction truncate, we only make the final
-        * transaction synchronous. */
+        * transaction synchronous.
+        */
        if (IS_SYNC(inode))
                handle->h_sync = 1;
 
index e3d5afc..edf49ec 100644 (file)
@@ -205,6 +205,9 @@ static inline int ext4_ext_get_actual_len(struct ext4_extent *ext)
 
 extern int ext4_extent_tree_init(handle_t *, struct inode *);
 extern int ext4_ext_calc_credits_for_insert(struct inode *, struct ext4_ext_path *);
+extern int ext4_ext_try_to_merge(struct inode *inode,
+                                struct ext4_ext_path *path,
+                                struct ext4_extent *);
 extern unsigned int ext4_ext_check_overlap(struct inode *, struct ext4_extent *, struct ext4_ext_path *);
 extern int ext4_ext_insert_extent(handle_t *, struct inode *, struct ext4_ext_path *, struct ext4_extent *);
 extern int ext4_ext_walk_space(struct inode *, unsigned long, unsigned long, ext_prepare_callback, void *);