LU-3489 osd-zfs: varsize index support

LFSCK needs indices with keys/values of different size.
add a support for this to osd-zfs. key/value size can't
vary within a single index object so far.

Signed-off-by: Alex Zhuravlev <alexey.zhuravlev@intel.com>
Change-Id: I6faa730ce1f778563c1fcf43d3b50c5b62c899d9
Reviewed-on: http://review.whamcloud.com/7152
Tested-by: Jenkins
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Nathaniel Clark <nathaniel.l.clark@intel.com>
Reviewed-by: Fan Yong <fan.yong@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lustre/osd-zfs/osd_index.c b/lustre/osd-zfs/osd_index.c
index 29df71d..29b0d68 100644 (file)
--- a/lustre/osd-zfs/osd_index.c
+++ b/lustre/osd-zfs/osd_index.c
@@ -1046,20 +1046,43 @@ static struct dt_index_operations osd_dir_ops = {
 
 /*
  * Primitives for index files using binary keys.
- * XXX: only 64-bit keys are supported for now.
  */
 
+static int osd_prepare_key(struct osd_object *o, __u64 *dst,
+                          const struct dt_key *src)
+{
+       int size;
+
+       LASSERT(dst);
+       LASSERT(src);
+
+       /* align keysize to 64bit */
+       size = (o->oo_keysize + sizeof(__u64) - 1) / sizeof(__u64);
+       size *= sizeof(__u64);
+
+       LASSERT(size <= MAXNAMELEN);
+
+       if (unlikely(size > o->oo_keysize))
+               memset(dst + o->oo_keysize, 0, size - o->oo_keysize);
+       memcpy(dst, (const char *)src, o->oo_keysize);
+
+       return size;
+}
+
 static int osd_index_lookup(const struct lu_env *env, struct dt_object *dt,
                        struct dt_rec *rec, const struct dt_key *key,
                        struct lustre_capa *capa)
 {
        struct osd_object *obj = osd_dt_obj(dt);
        struct osd_device *osd = osd_obj2dev(obj);
+       __u64             *k = osd_oti_get(env)->oti_key64;
        int                rc;
        ENTRY;
 
+       rc = osd_prepare_key(obj, k, key);
+
        rc = -zap_lookup_uint64(osd->od_objset.os, obj->oo_db->db_object,
-                               (const __u64 *)key, 1, 8, obj->oo_recsize,
+                               k, rc, obj->oo_recusize, obj->oo_recsize,
                                (void *)rec);
        RETURN(rc == 0 ? 1 : rc);
 }
@@ -1097,6 +1120,7 @@ static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
        struct osd_object  *obj = osd_dt_obj(dt);
        struct osd_device  *osd = osd_obj2dev(obj);
        struct osd_thandle *oh;
+       __u64              *k = osd_oti_get(env)->oti_key64;
        int                 rc;
        ENTRY;
 
@@ -1107,9 +1131,11 @@ static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
 
        oh = container_of0(th, struct osd_thandle, ot_super);
 
+       rc = osd_prepare_key(obj, k, key);
+
        /* Insert (key,oid) into ZAP */
        rc = -zap_add_uint64(osd->od_objset.os, obj->oo_db->db_object,
-                            (const __u64 *)key, 1, 8, obj->oo_recsize,
+                            k, rc, obj->oo_recusize, obj->oo_recsize,
                             (void *)rec, oh->ot_tx);
        RETURN(rc);
 }
@@ -1141,6 +1167,7 @@ static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
        struct osd_object  *obj = osd_dt_obj(dt);
        struct osd_device  *osd = osd_obj2dev(obj);
        struct osd_thandle *oh;
+       __u64              *k = osd_oti_get(env)->oti_key64;
        int                 rc;
        ENTRY;
 
@@ -1148,9 +1175,11 @@ static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
        LASSERT(th != NULL);
        oh = container_of0(th, struct osd_thandle, ot_super);
 
+       rc = osd_prepare_key(obj, k, key);
+
        /* Remove binary key from the ZAP */
        rc = -zap_remove_uint64(osd->od_objset.os, obj->oo_db->db_object,
-                               (const __u64 *)key, 1, oh->ot_tx);
+                               k, rc, oh->ot_tx);
        RETURN(rc);
 }
 
@@ -1165,8 +1194,11 @@ static int osd_index_it_get(const struct lu_env *env, struct dt_it *di,
        LASSERT(it);
        LASSERT(it->ozi_zc);
 
-       /* XXX: API is broken at the moment */
-       LASSERT(*((const __u64 *)key) == 0);
+       /*
+        * XXX: we need a binary version of zap_cursor_move_to_key()
+        *      to implement this API */
+       if (*((const __u64 *)key) != 0)
+               CERROR("NOT IMPLEMETED YET (move to %Lx)\n", *((__u64 *)key));
 
        zap_cursor_fini(it->ozi_zc);
        memset(it->ozi_zc, 0, sizeof(*it->ozi_zc));
@@ -1204,6 +1236,7 @@ static struct dt_key *osd_index_it_key(const struct lu_env *env,
                                       const struct dt_it *di)
 {
        struct osd_zap_it *it = (struct osd_zap_it *)di;
+       struct osd_object *obj = it->ozi_obj;
        zap_attribute_t   *za = &osd_oti_get(env)->oti_za;
        int                rc = 0;
        ENTRY;
@@ -1214,7 +1247,7 @@ static struct dt_key *osd_index_it_key(const struct lu_env *env,
                RETURN(ERR_PTR(rc));
 
        /* the binary key is stored in the name */
-       it->ozi_key = *((__u64 *)za->za_name);
+       memcpy(&it->ozi_key, za->za_name, obj->oo_keysize);
 
        RETURN((struct dt_key *)&it->ozi_key);
 }
@@ -1222,8 +1255,9 @@ static struct dt_key *osd_index_it_key(const struct lu_env *env,
 static int osd_index_it_key_size(const struct lu_env *env,
                                const struct dt_it *di)
 {
-       /* we only support 64-bit binary keys for the time being */
-       RETURN(sizeof(__u64));
+       struct osd_zap_it *it = (struct osd_zap_it *)di;
+       struct osd_object *obj = it->ozi_obj;
+       RETURN(obj->oo_keysize);
 }
 
 static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
@@ -1233,6 +1267,7 @@ static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
        struct osd_zap_it *it = (struct osd_zap_it *)di;
        struct osd_object *obj = it->ozi_obj;
        struct osd_device *osd = osd_obj2dev(obj);
+       __u64             *k = osd_oti_get(env)->oti_key64;
        int                rc;
        ENTRY;
 
@@ -1241,9 +1276,11 @@ static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
        if (rc)
                RETURN(rc);
 
+       rc = osd_prepare_key(obj, k, (const struct dt_key *)za->za_name);
+
        rc = -zap_lookup_uint64(osd->od_objset.os, obj->oo_db->db_object,
-                               (const __u64 *)za->za_name, 1, 8,
-                               obj->oo_recsize, (void *)rec);
+                               k, rc, obj->oo_recusize, obj->oo_recsize,
+                               (void *)rec);
        RETURN(rc);
 }
 
@@ -1339,10 +1376,9 @@ int osd_index_try(const struct lu_env *env, struct dt_object *dt,
                if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
                        RETURN(-EINVAL);
 
-               /* Although the zap_*_uint64() primitives support large keys, we
-                * limit ourselves to 64-bit keys for now */
-               if (feat->dif_keysize_max != sizeof(__u64) ||
-                   feat->dif_keysize_min != sizeof(__u64))
+               if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
+                       RETURN(-E2BIG);
+               if (feat->dif_keysize_max != feat->dif_keysize_min)
                        RETURN(-EINVAL);
 
                /* As for the record size, it should be a multiple of 8 bytes
@@ -1350,11 +1386,18 @@ int osd_index_try(const struct lu_env *env, struct dt_object *dt,
                 */
                if (feat->dif_recsize_max > ZAP_MAXVALUELEN)
                        RETURN(-E2BIG);
-               if (feat->dif_recsize_max != feat->dif_recsize_min ||
-                   (feat->dif_recsize_max & (sizeof(__u64) - 1)))
+               if (feat->dif_recsize_max != feat->dif_recsize_min)
                        RETURN(-EINVAL);
 
-               obj->oo_recsize = feat->dif_recsize_max / sizeof(__u64);
+               obj->oo_keysize = feat->dif_keysize_max;
+               obj->oo_recsize = feat->dif_recsize_max;
+               obj->oo_recusize = 1;
+
+               /* ZFS prefers to work with array of 64bits */
+               if ((obj->oo_recsize & 7) == 0) {
+                       obj->oo_recsize >>= 3;
+                       obj->oo_recusize = 8;
+               }
                dt->do_index_ops = &osd_index_ops;
        }
 

diff --git a/lustre/osd-zfs/osd_internal.h b/lustre/osd-zfs/osd_internal.h
index d217a56..0e202d2 100644 (file)
--- a/lustre/osd-zfs/osd_internal.h
+++ b/lustre/osd-zfs/osd_internal.h
@@ -104,7 +104,7 @@ struct osd_zap_it {
         * 3 - real records */
        unsigned                 ozi_pos:3;
        union {
-               char             ozi_name[NAME_MAX + 1]; /* file name for dir */
+               char             ozi_name[MAXNAMELEN]; /* file name for dir */
                __u64            ozi_key; /* binary key for index files */
        };
 };
@@ -167,7 +167,10 @@ struct osd_thread_info {
        };
 
        char                     oti_str[64];
-       char                     oti_key[MAXNAMELEN + 1];
+       union {
+               char             oti_key[MAXNAMELEN + 1];
+               __u64            oti_key64[(MAXNAMELEN + 1)/sizeof(__u64)];
+       };
        struct lustre_mdt_attrs oti_mdt_attrs;
 
        struct lu_attr           oti_la;
@@ -303,7 +306,9 @@ struct osd_object {
        uint64_t                 oo_xattr;
 
        /* record size for index file */
-       int                      oo_recsize;
+       unsigned char            oo_keysize;
+       unsigned char            oo_recsize;
+       unsigned char            oo_recusize;   /* unit size */
 };
 
 int osd_statfs(const struct lu_env *, struct dt_device *, struct obd_statfs *);