LU-4865 zfs: grow block size by write pattern

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

diff --git a/lustre/osd-zfs/osd_io.c b/lustre/osd-zfs/osd_io.c
index 89d6d20..eed35f5 100644 (file)
--- a/lustre/osd-zfs/osd_io.c
+++ b/lustre/osd-zfs/osd_io.c
@@ -27,7 +27,7 @@
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -685,8 +685,8 @@ retry:
 /**
  * Policy to grow ZFS block size by write pattern.
  * For sequential write, it grows block size gradually until it reaches the
- * maximum blocksize the dataset can support. Otherwise, it will just use
- * the maximum block size.
+ * maximum blocksize the dataset can support. Otherwise, it will pick a
+ * a block size by the writing region of this I/O.
  */
 static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
                              uint64_t start, uint64_t end)
@@ -696,6 +696,7 @@ static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
        dnode_t                 *dn;
        uint32_t                 blksz;
        int                      rc = 0;
+
        ENTRY;
 
        DB_DNODE_ENTER(db);
@@ -704,24 +705,40 @@ static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
        if (dn->dn_maxblkid > 0) /* can't change block size */
                GOTO(out, rc);
 
-       blksz = dn->dn_datablksz;
-       if (blksz >= osd->od_max_blksz)
+       if (dn->dn_datablksz >= osd->od_max_blksz)
                GOTO(out, rc);
 
+       down_write(&obj->oo_guard);
+
+       blksz = dn->dn_datablksz;
+       if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
+               GOTO(out_unlock, rc);
+
        /* now ZFS can support up to 16MB block size, and if the write
         * is sequential, it just increases the block size gradually */
        if (start <= blksz) { /* sequential */
                blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
-               if (!is_power_of_2(blksz))
-                       blksz = size_roundup_power2(blksz);
-       } else { /* otherwise, use maximum block size */
-               blksz = osd->od_max_blksz;
+       } else { /* sparse, pick a block size by write region */
+               blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
+                                       end - start);
        }
 
-       if (blksz > dn->dn_datablksz)
+       if (!is_power_of_2(blksz))
+               blksz = size_roundup_power2(blksz);
+
+       if (blksz > dn->dn_datablksz) {
                rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
                                               blksz, 0, oh->ot_tx);
+               LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
+               if (rc < 0)
+                       CDEBUG(D_INODE, "object "DFID": change block size"
+                              "%u -> %u error rc = %d\n",
+                              PFID(lu_object_fid(&obj->oo_dt.do_lu)),
+                              dn->dn_datablksz, blksz, rc);
+       }
        EXIT;
+out_unlock:
+       up_write(&obj->oo_guard);
 out:
        DB_DNODE_EXIT(db);
        return rc;
@@ -746,13 +763,9 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
        oh = container_of0(th, struct osd_thandle, ot_super);
 
        /* adjust block size. Assume the buffers are sorted. */
-       rc = osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
-                               lnb[npages - 1].lnb_file_offset +
-                               lnb[npages - 1].lnb_len);
-       if (rc < 0) /* ignore the error */
-               CDEBUG(D_INODE, "obj "DFID": change block size error rc=%d\n",
-                      PFID(lu_object_fid(&dt->do_lu)), rc);
-
+       (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
+                                lnb[npages - 1].lnb_file_offset +
+                                lnb[npages - 1].lnb_len);
        for (i = 0; i < npages; i++) {
                CDEBUG(D_INODE, "write %u bytes at %u\n",
                        (unsigned) lnb[i].lnb_len,
@@ -959,6 +972,20 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
                                 false));
 }
 
+static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
+                      __u64 start, __u64 end, enum lu_ladvise_type advice)
+{
+       int     rc;
+       ENTRY;
+
+       switch (advice) {
+       default:
+               rc = -ENOTSUPP;
+               break;
+       }
+
+       RETURN(rc);
+}
 
 struct dt_body_operations osd_body_ops = {
        .dbo_read                       = osd_read,
@@ -972,4 +999,5 @@ struct dt_body_operations osd_body_ops = {
        .dbo_read_prep                  = osd_read_prep,
        .dbo_declare_punch              = osd_declare_punch,
        .dbo_punch                      = osd_punch,
+       .dbo_ladvise                    = osd_ladvise,
 };