LU-7132 osd-ldiskfs: speedup rewrites

[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_io.c

diff --git a/lustre/osd-ldiskfs/osd_io.c b/lustre/osd-ldiskfs/osd_io.c
index 01b574b..98da270 100644 (file)
--- a/lustre/osd-ldiskfs/osd_io.c
+++ b/lustre/osd-ldiskfs/osd_io.c
@@ -1039,29 +1039,57 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
         RETURN(rc);
 }
 
-/* Check if a block is allocated or not */
-static int osd_is_mapped(struct inode *inode, u64 offset)
-{
-       sector_t (*fs_bmap)(struct address_space *, sector_t);
+struct osd_fextent {
+       sector_t        start;
+       sector_t        end;
+       unsigned int    mapped:1;
+};
 
-       fs_bmap = inode->i_mapping->a_ops->bmap;
+static int osd_is_mapped(struct dt_object *dt, __u64 offset,
+                        struct osd_fextent *cached_extent)
+{
+       struct inode *inode = osd_dt_obj(dt)->oo_inode;
+       sector_t block = offset >> inode->i_blkbits;
+       sector_t start;
+       struct fiemap_extent_info fei = { 0 };
+       struct fiemap_extent fe = { 0 };
+       mm_segment_t saved_fs;
+       int rc;
 
-       /* We can't know if we are overwriting or not */
-       if (unlikely(fs_bmap == NULL))
-               return 0;
+       if (block >= cached_extent->start && block < cached_extent->end)
+               return cached_extent->mapped;
 
        if (i_size_read(inode) == 0)
                return 0;
 
        /* Beyond EOF, must not be mapped */
-       if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
-           (offset >> inode->i_blkbits))
+       if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
                return 0;
 
-       if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
+       fei.fi_extents_max = 1;
+       fei.fi_extents_start = &fe;
+
+       saved_fs = get_fs();
+       set_fs(get_ds());
+       rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
+       set_fs(saved_fs);
+       if (rc != 0)
                return 0;
 
-       return 1;
+       start = fe.fe_logical >> inode->i_blkbits;
+
+       if (start > block) {
+               cached_extent->start = block;
+               cached_extent->end = start;
+               cached_extent->mapped = 0;
+       } else {
+               cached_extent->start = start;
+               cached_extent->end = (fe.fe_logical + fe.fe_length) >>
+                                     inode->i_blkbits;
+               cached_extent->mapped = 1;
+       }
+
+       return cached_extent->mapped;
 }
 
 static int osd_declare_write_commit(const struct lu_env *env,
@@ -1081,6 +1109,7 @@ static int osd_declare_write_commit(const struct lu_env *env,
        int                      credits = 0;
        bool                     ignore_quota = false;
        long long                quota_space = 0;
+       struct osd_fextent       extent = { 0 };
        ENTRY;
 
         LASSERT(handle != NULL);
@@ -1095,7 +1124,7 @@ static int osd_declare_write_commit(const struct lu_env *env,
                    lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
                        extents++;
 
-               if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
+               if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
                        quota_space += PAGE_CACHE_SIZE;
 
                /* ignore quota for the whole request if any page is from
@@ -1186,6 +1215,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
         loff_t isize;
         int rc = 0, i;
+       struct osd_fextent extent = { 0 };
 
         LASSERT(inode);
 
@@ -1198,7 +1228,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 
         for (i = 0; i < npages; i++) {
                if (lnb[i].lnb_rc == -ENOSPC &&
-                   osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
+                   osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
                        /* Allow the write to proceed if overwriting an
                         * existing block */
                        lnb[i].lnb_rc = 0;