fcntl: return -EFAULT if copy_to_user fails
[safe/jmp/linux-2.6] / fs / ocfs2 / suballoc.c
index 667d622..f4c2a9e 100644 (file)
 
 #define OCFS2_MAX_TO_STEAL             1024
 
+struct ocfs2_suballoc_result {
+       u64             sr_bg_blkno;    /* The bg we allocated from.  Set
+                                          to 0 when a block group is
+                                          contiguous. */
+       u64             sr_blkno;       /* The first allocated block */
+       unsigned int    sr_bit_offset;  /* The bit in the bg */
+       unsigned int    sr_bits;        /* How many bits we claimed */
+};
+
 static inline void ocfs2_debug_bg(struct ocfs2_group_desc *bg);
 static inline void ocfs2_debug_suballoc_inode(struct ocfs2_dinode *fe);
 static inline u16 ocfs2_find_victim_chain(struct ocfs2_chain_list *cl);
@@ -60,6 +69,7 @@ static int ocfs2_block_group_fill(handle_t *handle,
                                  struct inode *alloc_inode,
                                  struct buffer_head *bg_bh,
                                  u64 group_blkno,
+                                 unsigned int group_clusters,
                                  u16 my_chain,
                                  struct ocfs2_chain_list *cl);
 static int ocfs2_block_group_alloc(struct ocfs2_super *osb,
@@ -73,20 +83,17 @@ static int ocfs2_cluster_group_search(struct inode *inode,
                                      struct buffer_head *group_bh,
                                      u32 bits_wanted, u32 min_bits,
                                      u64 max_block,
-                                     u16 *bit_off, u16 *bits_found);
+                                     struct ocfs2_suballoc_result *res);
 static int ocfs2_block_group_search(struct inode *inode,
                                    struct buffer_head *group_bh,
                                    u32 bits_wanted, u32 min_bits,
                                    u64 max_block,
-                                   u16 *bit_off, u16 *bits_found);
-static int ocfs2_claim_suballoc_bits(struct ocfs2_super *osb,
-                                    struct ocfs2_alloc_context *ac,
+                                   struct ocfs2_suballoc_result *res);
+static int ocfs2_claim_suballoc_bits(struct ocfs2_alloc_context *ac,
                                     handle_t *handle,
                                     u32 bits_wanted,
                                     u32 min_bits,
-                                    u16 *bit_off,
-                                    unsigned int *num_bits,
-                                    u64 *bg_blkno);
+                                    struct ocfs2_suballoc_result *res);
 static int ocfs2_test_bg_bit_allocatable(struct buffer_head *bg_bh,
                                         int nr);
 static inline int ocfs2_block_group_set_bits(handle_t *handle,
@@ -326,14 +333,38 @@ out:
        return rc;
 }
 
+static void ocfs2_bg_discontig_add_extent(struct ocfs2_super *osb,
+                                         struct ocfs2_group_desc *bg,
+                                         struct ocfs2_chain_list *cl,
+                                         u64 p_blkno, u32 clusters)
+{
+       struct ocfs2_extent_list *el = &bg->bg_list;
+       struct ocfs2_extent_rec *rec;
+
+       BUG_ON(!ocfs2_supports_discontig_bg(osb));
+       if (!el->l_next_free_rec)
+               el->l_count = cpu_to_le16(ocfs2_extent_recs_per_gd(osb->sb));
+       rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec)];
+       rec->e_blkno = cpu_to_le64(p_blkno);
+       rec->e_cpos = cpu_to_le32(le16_to_cpu(bg->bg_bits) /
+                                 le16_to_cpu(cl->cl_bpc));
+       rec->e_leaf_clusters = cpu_to_le32(clusters);
+       le16_add_cpu(&bg->bg_bits, clusters * le16_to_cpu(cl->cl_bpc));
+       le16_add_cpu(&bg->bg_free_bits_count,
+                    clusters * le16_to_cpu(cl->cl_bpc));
+       le16_add_cpu(&el->l_next_free_rec, 1);
+}
+
 static int ocfs2_block_group_fill(handle_t *handle,
                                  struct inode *alloc_inode,
                                  struct buffer_head *bg_bh,
                                  u64 group_blkno,
+                                 unsigned int group_clusters,
                                  u16 my_chain,
                                  struct ocfs2_chain_list *cl)
 {
        int status = 0;
+       struct ocfs2_super *osb = OCFS2_SB(alloc_inode->i_sb);
        struct ocfs2_group_desc *bg = (struct ocfs2_group_desc *) bg_bh->b_data;
        struct super_block * sb = alloc_inode->i_sb;
 
@@ -360,12 +391,18 @@ static int ocfs2_block_group_fill(handle_t *handle,
        memset(bg, 0, sb->s_blocksize);
        strcpy(bg->bg_signature, OCFS2_GROUP_DESC_SIGNATURE);
        bg->bg_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation);
-       bg->bg_size = cpu_to_le16(ocfs2_group_bitmap_size(sb));
-       bg->bg_bits = cpu_to_le16(ocfs2_bits_per_group(cl));
+       bg->bg_size = cpu_to_le16(ocfs2_group_bitmap_size(sb, 1,
+                                               osb->s_feature_incompat));
        bg->bg_chain = cpu_to_le16(my_chain);
        bg->bg_next_group = cl->cl_recs[my_chain].c_blkno;
        bg->bg_parent_dinode = cpu_to_le64(OCFS2_I(alloc_inode)->ip_blkno);
        bg->bg_blkno = cpu_to_le64(group_blkno);
+       if (group_clusters == le16_to_cpu(cl->cl_cpg))
+               bg->bg_bits = cpu_to_le16(ocfs2_bits_per_group(cl));
+       else
+               ocfs2_bg_discontig_add_extent(osb, bg, cl, group_blkno,
+                                             group_clusters);
+
        /* set the 1st bit in the bitmap to account for the descriptor block */
        ocfs2_set_bit(0, (unsigned long *)bg->bg_bitmap);
        bg->bg_free_bits_count = cpu_to_le16(le16_to_cpu(bg->bg_bits) - 1);
@@ -396,6 +433,238 @@ static inline u16 ocfs2_find_smallest_chain(struct ocfs2_chain_list *cl)
        return best;
 }
 
+static struct buffer_head *
+ocfs2_block_group_alloc_contig(struct ocfs2_super *osb, handle_t *handle,
+                              struct inode *alloc_inode,
+                              struct ocfs2_alloc_context *ac,
+                              struct ocfs2_chain_list *cl)
+{
+       int status;
+       u32 bit_off, num_bits;
+       u64 bg_blkno;
+       struct buffer_head *bg_bh;
+       unsigned int alloc_rec = ocfs2_find_smallest_chain(cl);
+
+       status = ocfs2_claim_clusters(handle, ac,
+                                     le16_to_cpu(cl->cl_cpg), &bit_off,
+                                     &num_bits);
+       if (status < 0) {
+               if (status != -ENOSPC)
+                       mlog_errno(status);
+               goto bail;
+       }
+
+       /* setup the group */
+       bg_blkno = ocfs2_clusters_to_blocks(osb->sb, bit_off);
+       mlog(0, "new descriptor, record %u, at block %llu\n",
+            alloc_rec, (unsigned long long)bg_blkno);
+
+       bg_bh = sb_getblk(osb->sb, bg_blkno);
+       if (!bg_bh) {
+               status = -EIO;
+               mlog_errno(status);
+               goto bail;
+       }
+       ocfs2_set_new_buffer_uptodate(INODE_CACHE(alloc_inode), bg_bh);
+
+       status = ocfs2_block_group_fill(handle, alloc_inode, bg_bh,
+                                       bg_blkno, num_bits, alloc_rec, cl);
+       if (status < 0) {
+               brelse(bg_bh);
+               mlog_errno(status);
+       }
+
+bail:
+       return status ? ERR_PTR(status) : bg_bh;
+}
+
+static int ocfs2_block_group_claim_bits(struct ocfs2_super *osb,
+                                       handle_t *handle,
+                                       struct ocfs2_alloc_context *ac,
+                                       unsigned int min_bits,
+                                       u32 *bit_off, u32 *num_bits)
+{
+       int status = 0;
+
+       while (min_bits) {
+               status = ocfs2_claim_clusters(handle, ac, min_bits,
+                                             bit_off, num_bits);
+               if (status != -ENOSPC)
+                       break;
+
+               min_bits >>= 1;
+       }
+
+       return status;
+}
+
+static int ocfs2_block_group_grow_discontig(handle_t *handle,
+                                           struct inode *alloc_inode,
+                                           struct buffer_head *bg_bh,
+                                           struct ocfs2_alloc_context *ac,
+                                           struct ocfs2_chain_list *cl,
+                                           unsigned int min_bits)
+{
+       int status;
+       struct ocfs2_super *osb = OCFS2_SB(alloc_inode->i_sb);
+       struct ocfs2_group_desc *bg =
+               (struct ocfs2_group_desc *)bg_bh->b_data;
+       unsigned int needed = le16_to_cpu(cl->cl_cpg) -
+                        le16_to_cpu(bg->bg_bits) / le16_to_cpu(cl->cl_bpc);
+       u32 p_cpos, clusters;
+       u64 p_blkno;
+       struct ocfs2_extent_list *el = &bg->bg_list;
+
+       status = ocfs2_journal_access_gd(handle,
+                                        INODE_CACHE(alloc_inode),
+                                        bg_bh,
+                                        OCFS2_JOURNAL_ACCESS_CREATE);
+       if (status < 0) {
+               mlog_errno(status);
+               goto bail;
+       }
+
+       while ((needed > 0) && (le16_to_cpu(el->l_next_free_rec) <
+                               le16_to_cpu(el->l_count))) {
+               if (min_bits > needed)
+                       min_bits = needed;
+               status = ocfs2_block_group_claim_bits(osb, handle, ac,
+                                                     min_bits, &p_cpos,
+                                                     &clusters);
+               if (status < 0) {
+                       if (status != -ENOSPC)
+                               mlog_errno(status);
+                       goto bail;
+               }
+               p_blkno = ocfs2_clusters_to_blocks(osb->sb, p_cpos);
+               ocfs2_bg_discontig_add_extent(osb, bg, cl, p_blkno,
+                                             clusters);
+
+               min_bits = clusters;
+               needed = le16_to_cpu(cl->cl_cpg) -
+                        le16_to_cpu(bg->bg_bits) / le16_to_cpu(cl->cl_bpc);
+       }
+
+       if (needed > 0) {
+               /*
+                * We have used up all the extent rec but can't fill up
+                * the cpg. So bail out.
+                */
+               status = -ENOSPC;
+               goto bail;
+       }
+
+       ocfs2_journal_dirty(handle, bg_bh);
+
+bail:
+       return status;
+}
+
+static void ocfs2_bg_alloc_cleanup(handle_t *handle,
+                                  struct ocfs2_alloc_context *cluster_ac,
+                                  struct inode *alloc_inode,
+                                  struct buffer_head *bg_bh)
+{
+       int i, ret;
+       struct ocfs2_group_desc *bg;
+       struct ocfs2_extent_list *el;
+       struct ocfs2_extent_rec *rec;
+
+       if (!bg_bh)
+               return;
+
+       bg = (struct ocfs2_group_desc *)bg_bh->b_data;
+       el = &bg->bg_list;
+       for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
+               rec = &el->l_recs[i];
+               ret = ocfs2_free_clusters(handle, cluster_ac->ac_inode,
+                                         cluster_ac->ac_bh,
+                                         le64_to_cpu(rec->e_blkno),
+                                         le32_to_cpu(rec->e_leaf_clusters));
+               if (ret)
+                       mlog_errno(ret);
+               /* Try all the clusters to free */
+       }
+
+       ocfs2_remove_from_cache(INODE_CACHE(alloc_inode), bg_bh);
+       brelse(bg_bh);
+}
+
+static struct buffer_head *
+ocfs2_block_group_alloc_discontig(handle_t *handle,
+                                 struct inode *alloc_inode,
+                                 struct ocfs2_alloc_context *ac,
+                                 struct ocfs2_chain_list *cl)
+{
+       int status;
+       u32 bit_off, num_bits;
+       u64 bg_blkno;
+       unsigned int min_bits = le16_to_cpu(cl->cl_cpg) >> 1;
+       struct buffer_head *bg_bh = NULL;
+       unsigned int alloc_rec = ocfs2_find_smallest_chain(cl);
+       struct ocfs2_super *osb = OCFS2_SB(alloc_inode->i_sb);
+
+       if (!ocfs2_supports_discontig_bg(osb)) {
+               status = -ENOSPC;
+               goto bail;
+       }
+
+       status = ocfs2_extend_trans(handle,
+                                   ocfs2_calc_bg_discontig_credits(osb->sb));
+       if (status) {
+               mlog_errno(status);
+               goto bail;
+       }
+
+       /*
+        * We're going to be grabbing from multiple cluster groups.
+        * We don't have enough credits to relink them all, and the
+        * cluster groups will be staying in cache for the duration of
+        * this operation.
+        */
+       ac->ac_allow_chain_relink = 0;
+
+       /* Claim the first region */
+       status = ocfs2_block_group_claim_bits(osb, handle, ac, min_bits,
+                                             &bit_off, &num_bits);
+       if (status < 0) {
+               if (status != -ENOSPC)
+                       mlog_errno(status);
+               goto bail;
+       }
+       min_bits = num_bits;
+
+       /* setup the group */
+       bg_blkno = ocfs2_clusters_to_blocks(osb->sb, bit_off);
+       mlog(0, "new descriptor, record %u, at block %llu\n",
+            alloc_rec, (unsigned long long)bg_blkno);
+
+       bg_bh = sb_getblk(osb->sb, bg_blkno);
+       if (!bg_bh) {
+               status = -EIO;
+               mlog_errno(status);
+               goto bail;
+       }
+       ocfs2_set_new_buffer_uptodate(INODE_CACHE(alloc_inode), bg_bh);
+
+       status = ocfs2_block_group_fill(handle, alloc_inode, bg_bh,
+                                       bg_blkno, num_bits, alloc_rec, cl);
+       if (status < 0) {
+               mlog_errno(status);
+               goto bail;
+       }
+
+       status = ocfs2_block_group_grow_discontig(handle, alloc_inode,
+                                                 bg_bh, ac, cl, min_bits);
+       if (status)
+               mlog_errno(status);
+
+bail:
+       if (status)
+               ocfs2_bg_alloc_cleanup(handle, ac, alloc_inode, bg_bh);
+       return status ? ERR_PTR(status) : bg_bh;
+}
+
 /*
  * We expect the block group allocator to already be locked.
  */
@@ -411,9 +680,7 @@ static int ocfs2_block_group_alloc(struct ocfs2_super *osb,
        struct ocfs2_chain_list *cl;
        struct ocfs2_alloc_context *ac = NULL;
        handle_t *handle = NULL;
-       u32 bit_off, num_bits;
        u16 alloc_rec;
-       u64 bg_blkno;
        struct buffer_head *bg_bh = NULL;
        struct ocfs2_group_desc *bg;
 
@@ -446,44 +713,20 @@ static int ocfs2_block_group_alloc(struct ocfs2_super *osb,
                     (unsigned long long)*last_alloc_group);
                ac->ac_last_group = *last_alloc_group;
        }
-       status = ocfs2_claim_clusters(osb,
-                                     handle,
-                                     ac,
-                                     le16_to_cpu(cl->cl_cpg),
-                                     &bit_off,
-                                     &num_bits);
-       if (status < 0) {
+
+       bg_bh = ocfs2_block_group_alloc_contig(osb, handle, alloc_inode,
+                                              ac, cl);
+       if (IS_ERR(bg_bh) && (PTR_ERR(bg_bh) == -ENOSPC))
+               bg_bh = ocfs2_block_group_alloc_discontig(handle,
+                                                         alloc_inode,
+                                                         ac, cl);
+       if (IS_ERR(bg_bh)) {
+               status = PTR_ERR(bg_bh);
+               bg_bh = NULL;
                if (status != -ENOSPC)
                        mlog_errno(status);
                goto bail;
        }
-
-       alloc_rec = ocfs2_find_smallest_chain(cl);
-
-       /* setup the group */
-       bg_blkno = ocfs2_clusters_to_blocks(osb->sb, bit_off);
-       mlog(0, "new descriptor, record %u, at block %llu\n",
-            alloc_rec, (unsigned long long)bg_blkno);
-
-       bg_bh = sb_getblk(osb->sb, bg_blkno);
-       if (!bg_bh) {
-               status = -EIO;
-               mlog_errno(status);
-               goto bail;
-       }
-       ocfs2_set_new_buffer_uptodate(INODE_CACHE(alloc_inode), bg_bh);
-
-       status = ocfs2_block_group_fill(handle,
-                                       alloc_inode,
-                                       bg_bh,
-                                       bg_blkno,
-                                       alloc_rec,
-                                       cl);
-       if (status < 0) {
-               mlog_errno(status);
-               goto bail;
-       }
-
        bg = (struct ocfs2_group_desc *) bg_bh->b_data;
 
        status = ocfs2_journal_access_di(handle, INODE_CACHE(alloc_inode),
@@ -493,10 +736,12 @@ static int ocfs2_block_group_alloc(struct ocfs2_super *osb,
                goto bail;
        }
 
+       alloc_rec = le16_to_cpu(bg->bg_chain);
        le32_add_cpu(&cl->cl_recs[alloc_rec].c_free,
                     le16_to_cpu(bg->bg_free_bits_count));
-       le32_add_cpu(&cl->cl_recs[alloc_rec].c_total, le16_to_cpu(bg->bg_bits));
-       cl->cl_recs[alloc_rec].c_blkno  = cpu_to_le64(bg_blkno);
+       le32_add_cpu(&cl->cl_recs[alloc_rec].c_total,
+                    le16_to_cpu(bg->bg_bits));
+       cl->cl_recs[alloc_rec].c_blkno  = cpu_to_le64(bg->bg_blkno);
        if (le16_to_cpu(cl->cl_next_free_rec) < le16_to_cpu(cl->cl_count))
                le16_add_cpu(&cl->cl_next_free_rec, 1);
 
@@ -1024,8 +1269,7 @@ static int ocfs2_block_group_find_clear_bits(struct ocfs2_super *osb,
                                             struct buffer_head *bg_bh,
                                             unsigned int bits_wanted,
                                             unsigned int total_bits,
-                                            u16 *bit_off,
-                                            u16 *bits_found)
+                                            struct ocfs2_suballoc_result *res)
 {
        void *bitmap;
        u16 best_offset, best_size;
@@ -1069,14 +1313,9 @@ static int ocfs2_block_group_find_clear_bits(struct ocfs2_super *osb,
                }
        }
 
-       /* XXX: I think the first clause is equivalent to the second
-        *      - jlbec */
-       if (found == bits_wanted) {
-               *bit_off = start - found;
-               *bits_found = found;
-       } else if (best_size) {
-               *bit_off = best_offset;
-               *bits_found = best_size;
+       if (best_size) {
+               res->sr_bit_offset = best_offset;
+               res->sr_bits = best_size;
        } else {
                status = -ENOSPC;
                /* No error log here -- see the comment above
@@ -1232,14 +1471,13 @@ static int ocfs2_cluster_group_search(struct inode *inode,
                                      struct buffer_head *group_bh,
                                      u32 bits_wanted, u32 min_bits,
                                      u64 max_block,
-                                     u16 *bit_off, u16 *bits_found)
+                                     struct ocfs2_suballoc_result *res)
 {
        int search = -ENOSPC;
        int ret;
        u64 blkoff;
        struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *) group_bh->b_data;
        struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
-       u16 tmp_off, tmp_found;
        unsigned int max_bits, gd_cluster_off;
 
        BUG_ON(!ocfs2_is_cluster_bitmap(inode));
@@ -1266,15 +1504,15 @@ static int ocfs2_cluster_group_search(struct inode *inode,
 
                ret = ocfs2_block_group_find_clear_bits(OCFS2_SB(inode->i_sb),
                                                        group_bh, bits_wanted,
-                                                       max_bits,
-                                                       &tmp_off, &tmp_found);
+                                                       max_bits, res);
                if (ret)
                        return ret;
 
                if (max_block) {
                        blkoff = ocfs2_clusters_to_blocks(inode->i_sb,
                                                          gd_cluster_off +
-                                                         tmp_off + tmp_found);
+                                                         res->sr_bit_offset +
+                                                         res->sr_bits);
                        mlog(0, "Checking %llu against %llu\n",
                             (unsigned long long)blkoff,
                             (unsigned long long)max_block);
@@ -1286,16 +1524,14 @@ static int ocfs2_cluster_group_search(struct inode *inode,
                 * return success, but we still want to return
                 * -ENOSPC unless it found the minimum number
                 * of bits. */
-               if (min_bits <= tmp_found) {
-                       *bit_off = tmp_off;
-                       *bits_found = tmp_found;
+               if (min_bits <= res->sr_bits)
                        search = 0; /* success */
-               } else if (tmp_found) {
+               else if (res->sr_bits) {
                        /*
                         * Don't show bits which we'll be returning
                         * for allocation to the local alloc bitmap.
                         */
-                       ocfs2_local_alloc_seen_free_bits(osb, tmp_found);
+                       ocfs2_local_alloc_seen_free_bits(osb, res->sr_bits);
                }
        }
 
@@ -1306,7 +1542,7 @@ static int ocfs2_block_group_search(struct inode *inode,
                                    struct buffer_head *group_bh,
                                    u32 bits_wanted, u32 min_bits,
                                    u64 max_block,
-                                   u16 *bit_off, u16 *bits_found)
+                                   struct ocfs2_suballoc_result *res)
 {
        int ret = -ENOSPC;
        u64 blkoff;
@@ -1319,10 +1555,10 @@ static int ocfs2_block_group_search(struct inode *inode,
                ret = ocfs2_block_group_find_clear_bits(OCFS2_SB(inode->i_sb),
                                                        group_bh, bits_wanted,
                                                        le16_to_cpu(bg->bg_bits),
-                                                       bit_off, bits_found);
+                                                       res);
                if (!ret && max_block) {
-                       blkoff = le64_to_cpu(bg->bg_blkno) + *bit_off +
-                               *bits_found;
+                       blkoff = le64_to_cpu(bg->bg_blkno) +
+                               res->sr_bit_offset + res->sr_bits;
                        mlog(0, "Checking %llu against %llu\n",
                             (unsigned long long)blkoff,
                             (unsigned long long)max_block);
@@ -1361,24 +1597,70 @@ out:
        return ret;
 }
 
+static int ocfs2_bg_discontig_fix_by_rec(struct ocfs2_suballoc_result *res,
+                                        struct ocfs2_extent_rec *rec,
+                                        struct ocfs2_chain_list *cl)
+{
+       unsigned int bpc = le16_to_cpu(cl->cl_bpc);
+       unsigned int bitoff = le32_to_cpu(rec->e_cpos) * bpc;
+       unsigned int bitcount = le32_to_cpu(rec->e_leaf_clusters) * bpc;
+
+       if (res->sr_bit_offset < bitoff)
+               return 0;
+       if (res->sr_bit_offset >= (bitoff + bitcount))
+               return 0;
+       res->sr_blkno = le64_to_cpu(rec->e_blkno) +
+               (res->sr_bit_offset - bitoff);
+       if ((res->sr_bit_offset + res->sr_bits) > (bitoff + bitcount))
+               res->sr_bits = (bitoff + bitcount) - res->sr_bit_offset;
+       return 1;
+}
+
+static void ocfs2_bg_discontig_fix_result(struct ocfs2_alloc_context *ac,
+                                         struct ocfs2_group_desc *bg,
+                                         struct ocfs2_suballoc_result *res)
+{
+       int i;
+       u64 bg_blkno = res->sr_bg_blkno;  /* Save off */
+       struct ocfs2_extent_rec *rec;
+       struct ocfs2_dinode *di = (struct ocfs2_dinode *)ac->ac_bh->b_data;
+       struct ocfs2_chain_list *cl = &di->id2.i_chain;
+
+       if (ocfs2_is_cluster_bitmap(ac->ac_inode)) {
+               res->sr_blkno = 0;
+               return;
+       }
+
+       res->sr_blkno = res->sr_bg_blkno + res->sr_bit_offset;
+       res->sr_bg_blkno = 0;  /* Clear it for contig block groups */
+       if (!ocfs2_supports_discontig_bg(OCFS2_SB(ac->ac_inode->i_sb)) ||
+           !bg->bg_list.l_next_free_rec)
+               return;
+
+       for (i = 0; i < le16_to_cpu(bg->bg_list.l_next_free_rec); i++) {
+               rec = &bg->bg_list.l_recs[i];
+               if (ocfs2_bg_discontig_fix_by_rec(res, rec, cl)) {
+                       res->sr_bg_blkno = bg_blkno;  /* Restore */
+                       break;
+               }
+       }
+}
+
 static int ocfs2_search_one_group(struct ocfs2_alloc_context *ac,
                                  handle_t *handle,
                                  u32 bits_wanted,
                                  u32 min_bits,
-                                 u16 *bit_off,
-                                 unsigned int *num_bits,
-                                 u64 gd_blkno,
+                                 struct ocfs2_suballoc_result *res,
                                  u16 *bits_left)
 {
        int ret;
-       u16 found;
        struct buffer_head *group_bh = NULL;
        struct ocfs2_group_desc *gd;
        struct ocfs2_dinode *di = (struct ocfs2_dinode *)ac->ac_bh->b_data;
        struct inode *alloc_inode = ac->ac_inode;
 
-       ret = ocfs2_read_group_descriptor(alloc_inode, di, gd_blkno,
-                                         &group_bh);
+       ret = ocfs2_read_group_descriptor(alloc_inode, di,
+                                         res->sr_bg_blkno, &group_bh);
        if (ret < 0) {
                mlog_errno(ret);
                return ret;
@@ -1386,17 +1668,18 @@ static int ocfs2_search_one_group(struct ocfs2_alloc_context *ac,
 
        gd = (struct ocfs2_group_desc *) group_bh->b_data;
        ret = ac->ac_group_search(alloc_inode, group_bh, bits_wanted, min_bits,
-                                 ac->ac_max_block, bit_off, &found);
+                                 ac->ac_max_block, res);
        if (ret < 0) {
                if (ret != -ENOSPC)
                        mlog_errno(ret);
                goto out;
        }
 
-       *num_bits = found;
+       if (!ret)
+               ocfs2_bg_discontig_fix_result(ac, gd, res);
 
        ret = ocfs2_alloc_dinode_update_counts(alloc_inode, handle, ac->ac_bh,
-                                              *num_bits,
+                                              res->sr_bits,
                                               le16_to_cpu(gd->bg_chain));
        if (ret < 0) {
                mlog_errno(ret);
@@ -1404,7 +1687,7 @@ static int ocfs2_search_one_group(struct ocfs2_alloc_context *ac,
        }
 
        ret = ocfs2_block_group_set_bits(handle, alloc_inode, gd, group_bh,
-                                        *bit_off, *num_bits);
+                                        res->sr_bit_offset, res->sr_bits);
        if (ret < 0)
                mlog_errno(ret);
 
@@ -1420,13 +1703,11 @@ static int ocfs2_search_chain(struct ocfs2_alloc_context *ac,
                              handle_t *handle,
                              u32 bits_wanted,
                              u32 min_bits,
-                             u16 *bit_off,
-                             unsigned int *num_bits,
-                             u64 *bg_blkno,
+                             struct ocfs2_suballoc_result *res,
                              u16 *bits_left)
 {
        int status;
-       u16 chain, tmp_bits;
+       u16 chain;
        u32 tmp_used;
        u64 next_group;
        struct inode *alloc_inode = ac->ac_inode;
@@ -1455,8 +1736,8 @@ static int ocfs2_search_chain(struct ocfs2_alloc_context *ac,
         * the 1st group with any empty bits. */
        while ((status = ac->ac_group_search(alloc_inode, group_bh,
                                             bits_wanted, min_bits,
-                                            ac->ac_max_block, bit_off,
-                                            &tmp_bits)) == -ENOSPC) {
+                                            ac->ac_max_block,
+                                            res)) == -ENOSPC) {
                if (!bg->bg_next_group)
                        break;
 
@@ -1481,11 +1762,14 @@ static int ocfs2_search_chain(struct ocfs2_alloc_context *ac,
        }
 
        mlog(0, "alloc succeeds: we give %u bits from block group %llu\n",
-            tmp_bits, (unsigned long long)le64_to_cpu(bg->bg_blkno));
+            res->sr_bits, (unsigned long long)le64_to_cpu(bg->bg_blkno));
+
+       res->sr_bg_blkno = le64_to_cpu(bg->bg_blkno);
 
-       *num_bits = tmp_bits;
+       BUG_ON(res->sr_bits == 0);
+       if (!status)
+               ocfs2_bg_discontig_fix_result(ac, bg, res);
 
-       BUG_ON(*num_bits == 0);
 
        /*
         * Keep track of previous block descriptor read. When
@@ -1502,7 +1786,7 @@ static int ocfs2_search_chain(struct ocfs2_alloc_context *ac,
         */
        if (ac->ac_allow_chain_relink &&
            (prev_group_bh) &&
-           (ocfs2_block_group_reasonably_empty(bg, *num_bits))) {
+           (ocfs2_block_group_reasonably_empty(bg, res->sr_bits))) {
                status = ocfs2_relink_block_group(handle, alloc_inode,
                                                  ac->ac_bh, group_bh,
                                                  prev_group_bh, chain);
@@ -1524,25 +1808,24 @@ static int ocfs2_search_chain(struct ocfs2_alloc_context *ac,
        }
 
        tmp_used = le32_to_cpu(fe->id1.bitmap1.i_used);
-       fe->id1.bitmap1.i_used = cpu_to_le32(*num_bits + tmp_used);
-       le32_add_cpu(&cl->cl_recs[chain].c_free, -(*num_bits));
+       fe->id1.bitmap1.i_used = cpu_to_le32(res->sr_bits + tmp_used);
+       le32_add_cpu(&cl->cl_recs[chain].c_free, -res->sr_bits);
        ocfs2_journal_dirty(handle, ac->ac_bh);
 
        status = ocfs2_block_group_set_bits(handle,
                                            alloc_inode,
                                            bg,
                                            group_bh,
-                                           *bit_off,
-                                           *num_bits);
+                                           res->sr_bit_offset,
+                                           res->sr_bits);
        if (status < 0) {
                mlog_errno(status);
                goto bail;
        }
 
-       mlog(0, "Allocated %u bits from suballocator %llu\n", *num_bits,
+       mlog(0, "Allocated %u bits from suballocator %llu\n", res->sr_bits,
             (unsigned long long)le64_to_cpu(fe->i_blkno));
 
-       *bg_blkno = le64_to_cpu(bg->bg_blkno);
        *bits_left = le16_to_cpu(bg->bg_free_bits_count);
 bail:
        brelse(group_bh);
@@ -1553,19 +1836,15 @@ bail:
 }
 
 /* will give out up to bits_wanted contiguous bits. */
-static int ocfs2_claim_suballoc_bits(struct ocfs2_super *osb,
-                                    struct ocfs2_alloc_context *ac,
+static int ocfs2_claim_suballoc_bits(struct ocfs2_alloc_context *ac,
                                     handle_t *handle,
                                     u32 bits_wanted,
                                     u32 min_bits,
-                                    u16 *bit_off,
-                                    unsigned int *num_bits,
-                                    u64 *bg_blkno)
+                                    struct ocfs2_suballoc_result *res)
 {
        int status;
        u16 victim, i;
        u16 bits_left = 0;
-       u64 hint_blkno = ac->ac_last_group;
        struct ocfs2_chain_list *cl;
        struct ocfs2_dinode *fe;
 
@@ -1583,7 +1862,8 @@ static int ocfs2_claim_suballoc_bits(struct ocfs2_super *osb,
 
        if (le32_to_cpu(fe->id1.bitmap1.i_used) >=
            le32_to_cpu(fe->id1.bitmap1.i_total)) {
-               ocfs2_error(osb->sb, "Chain allocator dinode %llu has %u used "
+               ocfs2_error(ac->ac_inode->i_sb,
+                           "Chain allocator dinode %llu has %u used "
                            "bits but only %u total.",
                            (unsigned long long)le64_to_cpu(fe->i_blkno),
                            le32_to_cpu(fe->id1.bitmap1.i_used),
@@ -1592,22 +1872,16 @@ static int ocfs2_claim_suballoc_bits(struct ocfs2_super *osb,
                goto bail;
        }
 
-       if (hint_blkno) {
+       res->sr_bg_blkno = ac->ac_last_group;
+       if (res->sr_bg_blkno) {
                /* Attempt to short-circuit the usual search mechanism
                 * by jumping straight to the most recently used
                 * allocation group. This helps us mantain some
                 * contiguousness across allocations. */
                status = ocfs2_search_one_group(ac, handle, bits_wanted,
-                                               min_bits, bit_off, num_bits,
-                                               hint_blkno, &bits_left);
-               if (!status) {
-                       /* Be careful to update *bg_blkno here as the
-                        * caller is expecting it to be filled in, and
-                        * ocfs2_search_one_group() won't do that for
-                        * us. */
-                       *bg_blkno = hint_blkno;
+                                               min_bits, res, &bits_left);
+               if (!status)
                        goto set_hint;
-               }
                if (status < 0 && status != -ENOSPC) {
                        mlog_errno(status);
                        goto bail;
@@ -1620,8 +1894,8 @@ static int ocfs2_claim_suballoc_bits(struct ocfs2_super *osb,
        ac->ac_chain = victim;
        ac->ac_allow_chain_relink = 1;
 
-       status = ocfs2_search_chain(ac, handle, bits_wanted, min_bits, bit_off,
-                                   num_bits, bg_blkno, &bits_left);
+       status = ocfs2_search_chain(ac, handle, bits_wanted, min_bits,
+                                   res, &bits_left);
        if (!status)
                goto set_hint;
        if (status < 0 && status != -ENOSPC) {
@@ -1645,8 +1919,7 @@ static int ocfs2_claim_suballoc_bits(struct ocfs2_super *osb,
 
                ac->ac_chain = i;
                status = ocfs2_search_chain(ac, handle, bits_wanted, min_bits,
-                                           bit_off, num_bits, bg_blkno,
-                                           &bits_left);
+                                           res, &bits_left);
                if (!status)
                        break;
                if (status < 0 && status != -ENOSPC) {
@@ -1663,7 +1936,7 @@ set_hint:
                if (bits_left < min_bits)
                        ac->ac_last_group = 0;
                else
-                       ac->ac_last_group = *bg_blkno;
+                       ac->ac_last_group = res->sr_bg_blkno;
        }
 
 bail:
@@ -1671,37 +1944,37 @@ bail:
        return status;
 }
 
-int ocfs2_claim_metadata(struct ocfs2_super *osb,
-                        handle_t *handle,
+int ocfs2_claim_metadata(handle_t *handle,
                         struct ocfs2_alloc_context *ac,
                         u32 bits_wanted,
+                        u64 *suballoc_loc,
                         u16 *suballoc_bit_start,
                         unsigned int *num_bits,
                         u64 *blkno_start)
 {
        int status;
-       u64 bg_blkno;
+       struct ocfs2_suballoc_result res = { .sr_blkno = 0, };
 
        BUG_ON(!ac);
        BUG_ON(ac->ac_bits_wanted < (ac->ac_bits_given + bits_wanted));
        BUG_ON(ac->ac_which != OCFS2_AC_USE_META);
 
-       status = ocfs2_claim_suballoc_bits(osb,
-                                          ac,
+       status = ocfs2_claim_suballoc_bits(ac,
                                           handle,
                                           bits_wanted,
                                           1,
-                                          suballoc_bit_start,
-                                          num_bits,
-                                          &bg_blkno);
+                                          &res);
        if (status < 0) {
                mlog_errno(status);
                goto bail;
        }
-       atomic_inc(&osb->alloc_stats.bg_allocs);
+       atomic_inc(&OCFS2_SB(ac->ac_inode->i_sb)->alloc_stats.bg_allocs);
 
-       *blkno_start = bg_blkno + (u64) *suballoc_bit_start;
-       ac->ac_bits_given += (*num_bits);
+       *suballoc_loc = res.sr_bg_blkno;
+       *suballoc_bit_start = res.sr_bit_offset;
+       *blkno_start = res.sr_blkno;
+       ac->ac_bits_given += res.sr_bits;
+       *num_bits = res.sr_bits;
        status = 0;
 bail:
        mlog_exit(status);
@@ -1709,10 +1982,10 @@ bail:
 }
 
 static void ocfs2_init_inode_ac_group(struct inode *dir,
-                                     struct buffer_head *parent_fe_bh,
+                                     struct buffer_head *parent_di_bh,
                                      struct ocfs2_alloc_context *ac)
 {
-       struct ocfs2_dinode *fe = (struct ocfs2_dinode *)parent_fe_bh->b_data;
+       struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_di_bh->b_data;
        /*
         * Try to allocate inodes from some specific group.
         *
@@ -1726,10 +1999,14 @@ static void ocfs2_init_inode_ac_group(struct inode *dir,
        if (OCFS2_I(dir)->ip_last_used_group &&
            OCFS2_I(dir)->ip_last_used_slot == ac->ac_alloc_slot)
                ac->ac_last_group = OCFS2_I(dir)->ip_last_used_group;
-       else if (le16_to_cpu(fe->i_suballoc_slot) == ac->ac_alloc_slot)
-               ac->ac_last_group = ocfs2_which_suballoc_group(
-                                       le64_to_cpu(fe->i_blkno),
-                                       le16_to_cpu(fe->i_suballoc_bit));
+       else if (le16_to_cpu(di->i_suballoc_slot) == ac->ac_alloc_slot) {
+               if (di->i_suballoc_loc)
+                       ac->ac_last_group = le64_to_cpu(di->i_suballoc_loc);
+               else
+                       ac->ac_last_group = ocfs2_which_suballoc_group(
+                                       le64_to_cpu(di->i_blkno),
+                                       le16_to_cpu(di->i_suballoc_bit));
+       }
 }
 
 static inline void ocfs2_save_inode_ac_group(struct inode *dir,
@@ -1739,17 +2016,16 @@ static inline void ocfs2_save_inode_ac_group(struct inode *dir,
        OCFS2_I(dir)->ip_last_used_slot = ac->ac_alloc_slot;
 }
 
-int ocfs2_claim_new_inode(struct ocfs2_super *osb,
-                         handle_t *handle,
+int ocfs2_claim_new_inode(handle_t *handle,
                          struct inode *dir,
                          struct buffer_head *parent_fe_bh,
                          struct ocfs2_alloc_context *ac,
+                         u64 *suballoc_loc,
                          u16 *suballoc_bit,
                          u64 *fe_blkno)
 {
        int status;
-       unsigned int num_bits;
-       u64 bg_blkno;
+       struct ocfs2_suballoc_result res;
 
        mlog_entry_void();
 
@@ -1760,23 +2036,22 @@ int ocfs2_claim_new_inode(struct ocfs2_super *osb,
 
        ocfs2_init_inode_ac_group(dir, parent_fe_bh, ac);
 
-       status = ocfs2_claim_suballoc_bits(osb,
-                                          ac,
+       status = ocfs2_claim_suballoc_bits(ac,
                                           handle,
                                           1,
                                           1,
-                                          suballoc_bit,
-                                          &num_bits,
-                                          &bg_blkno);
+                                          &res);
        if (status < 0) {
                mlog_errno(status);
                goto bail;
        }
-       atomic_inc(&osb->alloc_stats.bg_allocs);
+       atomic_inc(&OCFS2_SB(ac->ac_inode->i_sb)->alloc_stats.bg_allocs);
 
-       BUG_ON(num_bits != 1);
+       BUG_ON(res.sr_bits != 1);
 
-       *fe_blkno = bg_blkno + (u64) (*suballoc_bit);
+       *suballoc_loc = res.sr_bg_blkno;
+       *suballoc_bit = res.sr_bit_offset;
+       *fe_blkno = res.sr_blkno;
        ac->ac_bits_given++;
        ocfs2_save_inode_ac_group(dir, ac);
        status = 0;
@@ -1846,8 +2121,7 @@ static inline void ocfs2_block_to_cluster_group(struct inode *inode,
  * contig. allocation, set to '1' to indicate we can deal with extents
  * of any size.
  */
-int __ocfs2_claim_clusters(struct ocfs2_super *osb,
-                          handle_t *handle,
+int __ocfs2_claim_clusters(handle_t *handle,
                           struct ocfs2_alloc_context *ac,
                           u32 min_clusters,
                           u32 max_clusters,
@@ -1856,8 +2130,8 @@ int __ocfs2_claim_clusters(struct ocfs2_super *osb,
 {
        int status;
        unsigned int bits_wanted = max_clusters;
-       u64 bg_blkno = 0;
-       u16 bg_bit_off;
+       struct ocfs2_suballoc_result res = { .sr_blkno = 0, };
+       struct ocfs2_super *osb = OCFS2_SB(ac->ac_inode->i_sb);
 
        mlog_entry_void();
 
@@ -1891,20 +2165,19 @@ int __ocfs2_claim_clusters(struct ocfs2_super *osb,
                if (bits_wanted > (osb->bitmap_cpg - 1))
                        bits_wanted = osb->bitmap_cpg - 1;
 
-               status = ocfs2_claim_suballoc_bits(osb,
-                                                  ac,
+               status = ocfs2_claim_suballoc_bits(ac,
                                                   handle,
                                                   bits_wanted,
                                                   min_clusters,
-                                                  &bg_bit_off,
-                                                  num_clusters,
-                                                  &bg_blkno);
+                                                  &res);
                if (!status) {
+                       BUG_ON(res.sr_blkno); /* cluster alloc can't set */
                        *cluster_start =
                                ocfs2_desc_bitmap_to_cluster_off(ac->ac_inode,
-                                                                bg_blkno,
-                                                                bg_bit_off);
+                                                                res.sr_bg_blkno,
+                                                                res.sr_bit_offset);
                        atomic_inc(&osb->alloc_stats.bitmap_data);
+                       *num_clusters = res.sr_bits;
                }
        }
        if (status < 0) {
@@ -1920,8 +2193,7 @@ bail:
        return status;
 }
 
-int ocfs2_claim_clusters(struct ocfs2_super *osb,
-                        handle_t *handle,
+int ocfs2_claim_clusters(handle_t *handle,
                         struct ocfs2_alloc_context *ac,
                         u32 min_clusters,
                         u32 *cluster_start,
@@ -1929,7 +2201,7 @@ int ocfs2_claim_clusters(struct ocfs2_super *osb,
 {
        unsigned int bits_wanted = ac->ac_bits_wanted - ac->ac_bits_given;
 
-       return __ocfs2_claim_clusters(osb, handle, ac, min_clusters,
+       return __ocfs2_claim_clusters(handle, ac, min_clusters,
                                      bits_wanted, cluster_start, num_clusters);
 }
 
@@ -2081,6 +2353,8 @@ int ocfs2_free_dinode(handle_t *handle,
        u16 bit = le16_to_cpu(di->i_suballoc_bit);
        u64 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
 
+       if (di->i_suballoc_loc)
+               bg_blkno = le64_to_cpu(di->i_suballoc_loc);
        return ocfs2_free_suballoc_bits(handle, inode_alloc_inode,
                                        inode_alloc_bh, bit, bg_blkno, 1);
 }
@@ -2350,7 +2624,7 @@ static int ocfs2_test_suballoc_bit(struct ocfs2_super *osb,
                                   struct buffer_head *alloc_bh, u64 blkno,
                                   u16 bit, int *res)
 {
-       struct ocfs2_dinode *alloc_fe;
+       struct ocfs2_dinode *alloc_di;
        struct ocfs2_group_desc *group;
        struct buffer_head *group_bh = NULL;
        u64 bg_blkno;
@@ -2359,17 +2633,20 @@ static int ocfs2_test_suballoc_bit(struct ocfs2_super *osb,
        mlog_entry("blkno: %llu bit: %u\n", (unsigned long long)blkno,
                   (unsigned int)bit);
 
-       alloc_fe = (struct ocfs2_dinode *)alloc_bh->b_data;
-       if ((bit + 1) > ocfs2_bits_per_group(&alloc_fe->id2.i_chain)) {
+       alloc_di = (struct ocfs2_dinode *)alloc_bh->b_data;
+       if ((bit + 1) > ocfs2_bits_per_group(&alloc_di->id2.i_chain)) {
                mlog(ML_ERROR, "suballoc bit %u out of range of %u\n",
                     (unsigned int)bit,
-                    ocfs2_bits_per_group(&alloc_fe->id2.i_chain));
+                    ocfs2_bits_per_group(&alloc_di->id2.i_chain));
                status = -EINVAL;
                goto bail;
        }
 
-       bg_blkno = ocfs2_which_suballoc_group(blkno, bit);
-       status = ocfs2_read_group_descriptor(suballoc, alloc_fe, bg_blkno,
+       if (alloc_di->i_suballoc_loc)
+               bg_blkno = le64_to_cpu(alloc_di->i_suballoc_loc);
+       else
+               bg_blkno = ocfs2_which_suballoc_group(blkno, bit);
+       status = ocfs2_read_group_descriptor(suballoc, alloc_di, bg_blkno,
                                             &group_bh);
        if (status < 0) {
                mlog(ML_ERROR, "read group %llu failed %d\n",