Whamcloud - gitweb
LU-13004 target: use KIOV for out_handle
[fs/lustre-release.git] / lustre / target / out_handler.c
index 7781ff3..6f5dd68 100644 (file)
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2013, 2016, Intel Corporation.
+ * Copyright (c) 2013, 2017, Intel Corporation.
  *
  * lustre/target/out_handler.c
  *
@@ -56,7 +56,6 @@ static void out_reconstruct(const struct lu_env *env, struct dt_device *dt,
               dt_obd_name(dt), reply, index, 0);
 
        object_update_result_insert(reply, NULL, 0, index, 0);
-       return;
 }
 
 typedef void (*out_reconstruct_t)(const struct lu_env *env,
@@ -217,14 +216,13 @@ static int out_attr_get(struct tgt_session_info *tsi)
                RETURN(-ENOENT);
        }
 
-       dt_read_lock(env, obj, MOR_TGT_CHILD);
+       dt_read_lock(env, obj, DT_TGT_CHILD);
        rc = dt_attr_get(env, obj, la);
        if (rc)
                GOTO(out_unlock, rc);
 
        obdo->o_valid = 0;
        obdo_from_la(obdo, la, la->la_valid);
-       lustre_set_wire_obdo(NULL, obdo, obdo);
 
 out_unlock:
        dt_read_unlock(env, obj);
@@ -267,7 +265,7 @@ static int out_xattr_get(struct tgt_session_info *tsi)
                RETURN(PTR_ERR(name));
        }
 
-       update_result = object_update_result_get(reply, 0, NULL);
+       update_result = object_update_result_get(reply, idx, NULL);
        if (update_result == NULL) {
                CERROR("%s: empty name for xattr get: rc = %d\n",
                       tgt_name(tsi->tsi_tgt), -EPROTO);
@@ -275,10 +273,12 @@ static int out_xattr_get(struct tgt_session_info *tsi)
        }
 
        lbuf->lb_len = (int)tti->tti_u.update.tti_update->ou_result_size;
-       lbuf->lb_buf = update_result->our_data;
        if (lbuf->lb_len == 0)
                lbuf->lb_buf = NULL;
-       dt_read_lock(env, obj, MOR_TGT_CHILD);
+       else
+               lbuf->lb_buf = update_result->our_data;
+
+       dt_read_lock(env, obj, DT_TGT_CHILD);
        rc = dt_xattr_get(env, obj, lbuf, name);
        dt_read_unlock(env, obj);
        if (rc <= 0) {
@@ -288,15 +288,70 @@ static int out_xattr_get(struct tgt_session_info *tsi)
        } else if (lbuf->lb_buf) {
                lbuf->lb_len = rc;
        }
-
-       CDEBUG(D_INFO, "%s: "DFID" get xattr %s len %d: rc = %d\n",
+       CDEBUG(D_INFO, "%s: "DFID" get xattr %s len %d\n",
               tgt_name(tsi->tsi_tgt), PFID(lu_object_fid(&obj->do_lu)),
-              name, (int)lbuf->lb_len, rc);
+              name, rc);
+
+       GOTO(out, rc);
 
+out:
        object_update_result_insert(reply, lbuf->lb_buf, lbuf->lb_len, idx, rc);
        RETURN(0);
 }
 
+static int out_xattr_list(struct tgt_session_info *tsi)
+{
+       const struct lu_env *env = tsi->tsi_env;
+       struct tgt_thread_info *tti = tgt_th_info(env);
+       struct lu_buf *lbuf = &tti->tti_buf;
+       struct object_update_reply *reply = tti->tti_u.update.tti_update_reply;
+       struct dt_object *obj = tti->tti_u.update.tti_dt_object;
+       struct object_update_result *update_result;
+       int idx = tti->tti_u.update.tti_update_reply_index;
+       int rc;
+
+       ENTRY;
+
+       if (!lu_object_exists(&obj->do_lu)) {
+               set_bit(LU_OBJECT_HEARD_BANSHEE,
+                       &obj->do_lu.lo_header->loh_flags);
+               RETURN(-ENOENT);
+       }
+
+       update_result = object_update_result_get(reply, 0, NULL);
+       if (!update_result) {
+               rc = -EPROTO;
+               CERROR("%s: empty buf for xattr list: rc = %d\n",
+                      tgt_name(tsi->tsi_tgt), rc);
+               RETURN(rc);
+       }
+
+       lbuf->lb_len = (int)tti->tti_u.update.tti_update->ou_result_size;
+       lbuf->lb_buf = update_result->our_data;
+       if (lbuf->lb_len == 0)
+               lbuf->lb_buf = 0;
+
+       dt_read_lock(env, obj, DT_TGT_CHILD);
+       rc = dt_xattr_list(env, obj, lbuf);
+       dt_read_unlock(env, obj);
+       if (rc <= 0) {
+               lbuf->lb_len = 0;
+               if (unlikely(!rc))
+                       rc = -ENODATA;
+       } else if (lbuf->lb_buf) {
+               lbuf->lb_len = rc;
+       }
+
+       CDEBUG(D_INFO, "%s: "DFID" list xattr len %d\n",
+              tgt_name(tsi->tsi_tgt), PFID(lu_object_fid(&obj->do_lu)), rc);
+
+       /* Since we directly use update_result->our_data as the lbuf->lb_buf,
+        * then use NULL for result_insert to avoid unnecessary memory copy. */
+       object_update_result_insert(reply, NULL, lbuf->lb_len, idx, rc);
+
+       RETURN(0);
+}
+
 static int out_index_lookup(struct tgt_session_info *tsi)
 {
        const struct lu_env     *env = tsi->tsi_env;
@@ -321,7 +376,7 @@ static int out_index_lookup(struct tgt_session_info *tsi)
                RETURN(PTR_ERR(name));
        }
 
-       dt_read_lock(env, obj, MOR_TGT_CHILD);
+       dt_read_lock(env, obj, DT_TGT_CHILD);
        if (!dt_try_as_dir(env, obj))
                GOTO(out_unlock, rc = -ENOTDIR);
 
@@ -512,6 +567,11 @@ static int out_index_insert(struct tgt_session_info *tsi)
                                 tti->tti_tea.ta_handle,
                                 tti->tti_u.update.tti_update_reply,
                                 tti->tti_u.update.tti_update_reply_index);
+
+       CDEBUG(D_INFO, "%s: "DFID" index insert %s: rc = %d\n",
+              tgt_name(tsi->tsi_tgt), PFID(lu_object_fid(&obj->do_lu)),
+              name, rc);
+
        RETURN(rc);
 }
 
@@ -655,8 +715,7 @@ static int out_read(struct tgt_session_info *tsi)
        orr = (struct out_read_reply *)update_result->our_data;
 
        nbufs = (size + OUT_BULK_BUFFER_SIZE - 1) / OUT_BULK_BUFFER_SIZE;
-       OBD_ALLOC(rdbuf, sizeof(struct lu_rdbuf) +
-                        nbufs * sizeof(rdbuf->rb_bufs[0]));
+       OBD_ALLOC(rdbuf, sizeof(*rdbuf) + nbufs * sizeof(rdbuf->rb_bufs[0]));
        if (rdbuf == NULL)
                GOTO(out, rc = -ENOMEM);
 
@@ -672,7 +731,7 @@ static int out_read(struct tgt_session_info *tsi)
                        GOTO(out_free, rc = -ENOMEM);
 
                rdbuf->rb_bufs[i].lb_len = read_size;
-               dt_read_lock(env, obj, MOR_TGT_CHILD);
+               dt_read_lock(env, obj, DT_TGT_CHILD);
                rc = dt_read(env, obj, &rdbuf->rb_bufs[i], &pos);
                dt_read_unlock(env, obj);
 
@@ -701,7 +760,7 @@ out_free:
                                 rdbuf->rb_bufs[i].lb_len);
                }
        }
-       OBD_FREE(rdbuf, sizeof(struct lu_rdbuf) +
+       OBD_FREE(rdbuf, sizeof(*rdbuf) +
                        nbufs * sizeof(rdbuf->rb_bufs[0]));
 out:
        /* Insert read buffer */
@@ -728,33 +787,35 @@ static int out_noop(struct tgt_session_info *tsi)
 }
 
 static struct tgt_handler out_update_ops[] = {
-       DEF_OUT_HNDL(OUT_CREATE, "out_create", MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_CREATE, "out_create", IS_MUTABLE | HAS_REPLY,
                     out_create),
-       DEF_OUT_HNDL(OUT_DESTROY, "out_create", MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_DESTROY, "out_create", IS_MUTABLE | HAS_REPLY,
                     out_destroy),
-       DEF_OUT_HNDL(OUT_REF_ADD, "out_ref_add", MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_REF_ADD, "out_ref_add", IS_MUTABLE | HAS_REPLY,
                     out_ref_add),
-       DEF_OUT_HNDL(OUT_REF_DEL, "out_ref_del", MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_REF_DEL, "out_ref_del", IS_MUTABLE | HAS_REPLY,
                     out_ref_del),
-       DEF_OUT_HNDL(OUT_ATTR_SET, "out_attr_set",  MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_ATTR_SET, "out_attr_set",  IS_MUTABLE | HAS_REPLY,
                     out_attr_set),
-       DEF_OUT_HNDL(OUT_ATTR_GET, "out_attr_get",  HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_ATTR_GET, "out_attr_get",  HAS_REPLY,
                     out_attr_get),
-       DEF_OUT_HNDL(OUT_XATTR_SET, "out_xattr_set", MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_XATTR_SET, "out_xattr_set", IS_MUTABLE | HAS_REPLY,
                     out_xattr_set),
-       DEF_OUT_HNDL(OUT_XATTR_DEL, "out_xattr_del", MUTABOR | HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_XATTR_DEL, "out_xattr_del", IS_MUTABLE | HAS_REPLY,
                     out_xattr_del),
-       DEF_OUT_HNDL(OUT_XATTR_GET, "out_xattr_get", HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_XATTR_GET, "out_xattr_get", HAS_REPLY,
                     out_xattr_get),
-       DEF_OUT_HNDL(OUT_INDEX_LOOKUP, "out_index_lookup", HABEO_REFERO,
+       DEF_OUT_HNDL(OUT_INDEX_LOOKUP, "out_index_lookup", HAS_REPLY,
                     out_index_lookup),
        DEF_OUT_HNDL(OUT_INDEX_INSERT, "out_index_insert",
-                    MUTABOR | HABEO_REFERO, out_index_insert),
+                    IS_MUTABLE | HAS_REPLY, out_index_insert),
        DEF_OUT_HNDL(OUT_INDEX_DELETE, "out_index_delete",
-                    MUTABOR | HABEO_REFERO, out_index_delete),
-       DEF_OUT_HNDL(OUT_WRITE, "out_write", MUTABOR | HABEO_REFERO, out_write),
-       DEF_OUT_HNDL(OUT_READ, "out_read", HABEO_REFERO, out_read),
-       DEF_OUT_HNDL(OUT_NOOP, "out_noop", HABEO_REFERO, out_noop),
+                    IS_MUTABLE | HAS_REPLY, out_index_delete),
+       DEF_OUT_HNDL(OUT_WRITE, "out_write", IS_MUTABLE | HAS_REPLY, out_write),
+       DEF_OUT_HNDL(OUT_READ, "out_read", HAS_REPLY, out_read),
+       DEF_OUT_HNDL(OUT_NOOP, "out_noop", HAS_REPLY, out_noop),
+       DEF_OUT_HNDL(OUT_XATTR_LIST, "out_xattr_list", HAS_REPLY,
+                    out_xattr_list),
 };
 
 static struct tgt_handler *out_handler_find(__u32 opc)
@@ -808,17 +869,7 @@ static int out_trans_stop(const struct lu_env *env,
        rc = dt_trans_stop(env, ta->ta_handle->th_dev, ta->ta_handle);
        for (i = 0; i < ta->ta_argno; i++) {
                if (ta->ta_args[i]->object != NULL) {
-                       struct dt_object *obj = ta->ta_args[i]->object;
-
-                       /* If the object is being created during this
-                        * transaction, we need to remove them from the
-                        * cache immediately, because a few layers are
-                        * missing in OUT handler, i.e. the object might
-                        * not be initialized in all layers */
-                       if (ta->ta_args[i]->exec_fn == out_tx_create_exec)
-                               set_bit(LU_OBJECT_HEARD_BANSHEE,
-                                       &obj->do_lu.lo_header->loh_flags);
-                       lu_object_put(env, &ta->ta_args[i]->object->do_lu);
+                       dt_object_put(env, ta->ta_args[i]->object);
                        ta->ta_args[i]->object = NULL;
                }
        }
@@ -913,7 +964,6 @@ int out_handle(struct tgt_session_info *tsi)
        struct object_update            *update;
        struct object_update_reply      *reply;
        struct ptlrpc_bulk_desc         *desc = NULL;
-       struct l_wait_info              lwi;
        void                            **update_bufs;
        int                             current_batchid = -1;
        __u32                           update_buf_count;
@@ -954,17 +1004,23 @@ int out_handle(struct tgt_session_info *tsi)
                update_bufs[0] = ouh->ouh_inline_data;
        } else {
                struct out_update_buffer *tmp;
+               int page_count = 0;
 
                oub = req_capsule_client_get(pill, &RMF_OUT_UPDATE_BUF);
                if (oub == NULL)
                        GOTO(out_free, rc = err_serious(-EPROTO));
 
-               desc = ptlrpc_prep_bulk_exp(pill->rc_req, update_buf_count,
+               for (i = 0; i < update_buf_count; i++)
+                       /* First *and* last might be partial pages, hence +1 */
+                       page_count += DIV_ROUND_UP(oub[i].oub_size,
+                                                  PAGE_SIZE) + 1;
+
+               desc = ptlrpc_prep_bulk_exp(pill->rc_req, page_count,
                                            PTLRPC_BULK_OPS_COUNT,
                                            PTLRPC_BULK_GET_SINK |
-                                           PTLRPC_BULK_BUF_KVEC,
+                                           PTLRPC_BULK_BUF_KIOV,
                                            MDS_BULK_PORTAL,
-                                           &ptlrpc_bulk_kvec_ops);
+                                           &ptlrpc_bulk_kiov_nopin_ops);
                if (desc == NULL)
                        GOTO(out_free, rc = err_serious(-ENOMEM));
 
@@ -973,7 +1029,7 @@ int out_handle(struct tgt_session_info *tsi)
                        if (tmp->oub_size >= OUT_MAXREQSIZE)
                                GOTO(out_free, rc = err_serious(-EPROTO));
 
-                       OBD_ALLOC(update_bufs[i], tmp->oub_size);
+                       OBD_ALLOC_LARGE(update_bufs[i], tmp->oub_size);
                        if (update_bufs[i] == NULL)
                                GOTO(out_free, rc = err_serious(-ENOMEM));
 
@@ -986,7 +1042,7 @@ int out_handle(struct tgt_session_info *tsi)
                if (rc != 0)
                        GOTO(out_free, rc = err_serious(rc));
 
-               rc = target_bulk_io(pill->rc_req->rq_export, desc, &lwi);
+               rc = target_bulk_io(pill->rc_req->rq_export, desc);
                if (rc < 0)
                        GOTO(out_free, rc = err_serious(rc));
        }
@@ -1085,12 +1141,13 @@ int out_handle(struct tgt_session_info *tsi)
                                GOTO(out, rc = PTR_ERR(dt_obj));
 
                        if (dt->dd_record_fid_accessed) {
-                               lfsck_pack_rfa(&tti->tti_lr,
+                               struct lfsck_req_local *lrl = &tti->tti_lrl;
+
+                               lfsck_pack_rfa(lrl,
                                               lu_object_fid(&dt_obj->do_lu),
-                                              LE_FID_ACCESSED,
+                                              LEL_FID_ACCESSED,
                                               LFSCK_TYPE_LAYOUT);
-                               tgt_lfsck_in_notify(env, dt, &tti->tti_lr,
-                                                   NULL);
+                               tgt_lfsck_in_notify_local(env, dt, lrl, NULL);
                        }
 
                        tti->tti_u.update.tti_dt_object = dt_obj;
@@ -1105,17 +1162,20 @@ int out_handle(struct tgt_session_info *tsi)
                        }
 
                        /* Check resend case only for modifying RPC */
-                       if (h->th_flags & MUTABOR) {
+                       if (h->th_flags & IS_MUTABLE) {
                                struct ptlrpc_request *req = tgt_ses_req(tsi);
 
                                if (out_check_resent(env, dt, dt_obj, req,
                                                     out_reconstruct, reply,
                                                     reply_index))
                                        GOTO(next, rc = 0);
+
+                               if (dt->dd_rdonly)
+                                       GOTO(next, rc = -EROFS);
                        }
 
                        /* start transaction for modification RPC only */
-                       if (h->th_flags & MUTABOR && current_batchid == -1) {
+                       if (h->th_flags & IS_MUTABLE && current_batchid == -1) {
                                current_batchid = update->ou_batchid;
                                rc = out_tx_start(env, dt, ta, tsi->tsi_exp);
                                if (rc != 0)
@@ -1128,7 +1188,7 @@ int out_handle(struct tgt_session_info *tsi)
                        /* Stop the current update transaction, if the update
                         * has different batchid, or read-only update */
                        if (((current_batchid != update->ou_batchid) ||
-                            !(h->th_flags & MUTABOR)) &&
+                            !(h->th_flags & IS_MUTABLE)) &&
                             ta->ta_handle != NULL) {
                                rc = out_tx_end(env, ta, rc);
                                current_batchid = -1;
@@ -1136,7 +1196,7 @@ int out_handle(struct tgt_session_info *tsi)
                                        GOTO(next, rc);
 
                                /* start a new transaction if needed */
-                               if (h->th_flags & MUTABOR) {
+                               if (h->th_flags & IS_MUTABLE) {
                                        rc = out_tx_start(env, dt, ta,
                                                          tsi->tsi_exp);
                                        if (rc != 0)
@@ -1150,7 +1210,7 @@ int out_handle(struct tgt_session_info *tsi)
                        rc = h->th_act(tsi);
 next:
                        reply_index++;
-                       lu_object_put(env, &dt_obj->do_lu);
+                       dt_object_put(env, dt_obj);
                        if (rc < 0)
                                GOTO(out, rc);
                }
@@ -1167,7 +1227,8 @@ out_free:
                if (oub != NULL) {
                        for (i = 0; i < update_buf_count; i++, oub++) {
                                if (update_bufs[i] != NULL)
-                                       OBD_FREE(update_bufs[i], oub->oub_size);
+                                       OBD_FREE_LARGE(update_bufs[i],
+                                                      oub->oub_size);
                        }
                }
 
@@ -1181,7 +1242,7 @@ out_free:
 }
 
 struct tgt_handler tgt_out_handlers[] = {
-TGT_UPDATE_HDL(MUTABOR,        OUT_UPDATE,     out_handle),
+TGT_UPDATE_HDL(IS_MUTABLE,     OUT_UPDATE,     out_handle),
 };
 EXPORT_SYMBOL(tgt_out_handlers);