LU-2049 grant: add support for OBD_CONNECT_GRANT_PARAM

[fs/lustre-release.git] / lustre / osd-zfs / osd_handler.c

diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c
index 4f78957..9c19163 100644 (file)
--- a/lustre/osd-zfs/osd_handler.c
+++ b/lustre/osd-zfs/osd_handler.c
@@ -214,9 +214,8 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
                struct osd_device *osd = osd_dt_dev(d);
                /* dmu will call commit callback with error code during abort */
                if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
-                       CERROR("%s: failed to start transaction due to ENOSPC. "
-                              "Metadata overhead is underestimated or "
-                              "grant_ratio is too low.\n", osd->od_svname);
+                       CERROR("%s: failed to start transaction due to ENOSPC"
+                              "\n", osd->od_svname);
                else
                        CERROR("%s: can't assign tx: rc = %d\n",
                               osd->od_svname, rc);
@@ -474,14 +473,13 @@ static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs)
         * for internal files to be created/unlinked when space is tight.
         */
        CLASSERT(OSD_STATFS_RESERVED_SIZE > 0);
-       if (likely(osfs->os_blocks >= OSD_STATFS_RESERVED_SIZE))
+       reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
+       if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
                reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
-       else
-               reserved = OSD_STATFS_RESERVED_SIZE >> bshift;
 
        osfs->os_blocks -= reserved;
-       osfs->os_bfree  -= MIN(reserved, osfs->os_bfree);
-       osfs->os_bavail -= MIN(reserved, osfs->os_bavail);
+       osfs->os_bfree  -= min(reserved, osfs->os_bfree);
+       osfs->os_bavail -= min(reserved, osfs->os_bavail);
 
        /*
         * The availobjs value returned from dmu_objset_space() is largely
@@ -563,7 +561,7 @@ static void osd_conf_get(const struct lu_env *env,
         */
        param->ddp_max_name_len = MAXNAMELEN;
        param->ddp_max_nlink    = 1 << 31; /* it's 8byte on a disk */
-       param->ddp_block_shift  = 12; /* XXX */
+       param->ddp_symlink_max  = PATH_MAX;
        param->ddp_mount_type   = LDD_MT_ZFS;
 
        param->ddp_mntopts      = MNTOPT_USERXATTR;
@@ -574,20 +572,22 @@ static void osd_conf_get(const struct lu_env *env,
        /* for maxbytes, report same value as ZPL */
        param->ddp_maxbytes     = MAX_LFS_FILESIZE;
 
-       /* Default reserved fraction of the available space that should be kept
-        * for error margin. Unfortunately, there are many factors that can
-        * impact the overhead with zfs, so let's be very cautious for now and
-        * reserve 20% of the available space which is not given out as grant.
-        * This tunable can be changed on a live system via procfs if needed. */
-       param->ddp_grant_reserved = 20;
-
        /* inodes are dynamically allocated, so we report the per-inode space
         * consumption to upper layers. This static value is not really accurate
         * and we should use the same logic as in udmu_objset_statfs() to
         * estimate the real size consumed by an object */
        param->ddp_inodespace = OSD_DNODE_EST_COUNT;
-       /* per-fragment overhead to be used by the client code */
-       param->ddp_grant_frag = osd_blk_insert_cost(osd);
+       /* Although ZFS isn't an extent-based filesystem, the metadata overhead
+        * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should
+        * not be accounted for every single new block insertion.
+        * Instead, the maximum extent size is set to the number of blocks that
+        * can fit into a single contiguous indirect block. There would be some
+        * cases where this crosses indirect blocks, but it also won't have 7
+        * new levels of indirect blocks in that case either, so it will still
+        * have enough reserved space for the extra indirect block */
+       param->ddp_max_extent_blks =
+               (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT));
+       param->ddp_extent_tax = osd_blk_insert_cost(osd);
 }
 
 /*