Whamcloud - gitweb
LU-1061 agl: cl_locks_prune() waits for the last user
[fs/lustre-release.git] / lustre / osc / osc_request.c
index b9499b0..2c4cef7 100644 (file)
 #include <lustre_param.h>
 #include "osc_internal.h"
 
-static quota_interface_t *quota_interface = NULL;
-extern quota_interface_t osc_quota_interface;
-
 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
 static int brw_interpret(const struct lu_env *env,
                          struct ptlrpc_request *req, void *data, int rc);
+static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli,
+                            int ptlrpc);
 int osc_cleanup(struct obd_device *obd);
 
 /* Pack OSC object metadata for disk storage (LE byte order). */
@@ -403,7 +402,7 @@ int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
         /* do mds to ost setattr asynchronously */
         if (!rqset) {
                 /* Do not wait for response. */
-                ptlrpcd_add_req(req, PSCOPE_OTHER);
+                ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
         } else {
                 req->rq_interpret_reply =
                         (ptlrpc_interpterer_t)osc_setattr_interpret;
@@ -415,7 +414,7 @@ int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
                 sa->sa_cookie = cookie;
 
                 if (rqset == PTLRPCD_SET)
-                        ptlrpcd_add_req(req, PSCOPE_OTHER);
+                        ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
                 else
                         ptlrpc_set_add_req(rqset, req);
         }
@@ -554,7 +553,7 @@ int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
         sa->sa_upcall = upcall;
         sa->sa_cookie = cookie;
         if (rqset == PTLRPCD_SET)
-                ptlrpcd_add_req(req, PSCOPE_OTHER);
+                ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
         else
                 ptlrpc_set_add_req(rqset, req);
 
@@ -768,7 +767,7 @@ static int osc_destroy(struct obd_export *exp, struct obdo *oa,
         }
 
         /* Do not wait for response */
-        ptlrpcd_add_req(req, PSCOPE_OTHER);
+        ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
         RETURN(0);
 }
 
@@ -1219,14 +1218,16 @@ static int check_write_rcs(struct ptlrpc_request *req,
 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 {
         if (p1->flag != p2->flag) {
-                unsigned mask = ~(OBD_BRW_FROM_GRANT|
-                                  OBD_BRW_NOCACHE|OBD_BRW_SYNC|OBD_BRW_ASYNC);
+                unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
+                                  OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
 
                 /* warn if we try to combine flags that we don't know to be
                  * safe to combine */
-                if ((p1->flag & mask) != (p2->flag & mask))
-                        CERROR("is it ok to have flags 0x%x and 0x%x in the "
-                               "same brw?\n", p1->flag, p2->flag);
+                if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
+                        CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
+                              "report this at http://bugs.whamcloud.com/\n",
+                              p1->flag, p2->flag);
+                }
                 return 0;
         }
 
@@ -1266,7 +1267,7 @@ static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
                 cksum++;
 
-        return cksum;
+        return fini_checksum(cksum, cksum_type);
 }
 
 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
@@ -1484,10 +1485,6 @@ static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
                 return 0;
         }
 
-        /* If this is mmaped file - it can be changed at any time */
-        if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
-                return 1;
-
         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
                                        oa->o_flags : 0);
         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
@@ -1544,7 +1541,6 @@ static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
                 RETURN(-EPROTO);
         }
 
-#ifdef HAVE_QUOTA_SUPPORT
         /* set/clear over quota flag for a uid/gid */
         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
@@ -1553,10 +1549,8 @@ static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
                        body->oa.o_flags);
-                lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
-                             body->oa.o_flags);
+                osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
         }
-#endif
 
         osc_update_grant(cli, body);
 
@@ -1973,7 +1967,6 @@ static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
                          int cmd)
 {
-        int optimal;
         ENTRY;
 
         if (lop->lop_num_pending == 0)
@@ -1994,8 +1987,7 @@ static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
                 RETURN(1);
         }
-        /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
-        optimal = cli->cl_max_pages_per_rpc;
+
         if (cmd & OBD_BRW_WRITE) {
                 /* trigger a write rpc stream as long as there are dirtiers
                  * waiting for space.  as they're waiting, they're not going to
@@ -2004,13 +1996,8 @@ static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
                         RETURN(1);
                 }
-                /* +16 to avoid triggering rpcs that would want to include pages
-                 * that are being queued but which can't be made ready until
-                 * the queuer finishes with the page. this is a wart for
-                 * llite::commit_write() */
-                optimal += 16;
         }
-        if (lop->lop_num_pending >= optimal)
+        if (lop->lop_num_pending >= cli->cl_max_pages_per_rpc)
                 RETURN(1);
 
         RETURN(0);
@@ -2201,9 +2188,9 @@ static void osc_ap_completion(const struct lu_env *env,
         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
                                                 oap->oap_cmd, oa, rc);
 
-        /* ll_ap_completion (from llite) drops PG_locked. so, a new
-         * I/O on the page could start, but OSC calls it under lock
-         * and thus we can add oap back to pending safely */
+        /* cl_page_completion() drops PG_locked. so, a new I/O on the page could
+         * start, but OSC calls it under lock and thus we can add oap back to
+         * pending safely */
         if (rc)
                 /* upper layer wants to leave the page on pending queue */
                 osc_oap_to_pending(oap);
@@ -2212,6 +2199,18 @@ static void osc_ap_completion(const struct lu_env *env,
         EXIT;
 }
 
+static int brw_queue_work(const struct lu_env *env, void *data)
+{
+        struct client_obd *cli = data;
+
+        CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
+
+        client_obd_list_lock(&cli->cl_loi_list_lock);
+        osc_check_rpcs0(env, cli, 1);
+        client_obd_list_unlock(&cli->cl_loi_list_lock);
+        RETURN(0);
+}
+
 static int brw_interpret(const struct lu_env *env,
                          struct ptlrpc_request *req, void *data, int rc)
 {
@@ -2223,20 +2222,9 @@ static int brw_interpret(const struct lu_env *env,
         rc = osc_brw_fini_request(req, rc);
         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
         if (osc_recoverable_error(rc)) {
-                /* Only retry once for mmaped files since the mmaped page
-                 * might be modified at anytime. We have to retry at least
-                 * once in case there WAS really a corruption of the page
-                 * on the network, that was not caused by mmap() modifying
-                 * the page. Bug11742 */
-                if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
-                    aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
-                    aa->aa_oa->o_flags & OBD_FL_MMAP) {
-                        rc = 0;
-                } else {
-                        rc = osc_brw_redo_request(req, aa);
-                        if (rc == 0)
-                                RETURN(0);
-                }
+                rc = osc_brw_redo_request(req, aa);
+                if (rc == 0)
+                        RETURN(0);
         }
 
         if (aa->aa_ocapa) {
@@ -2245,7 +2233,6 @@ static int brw_interpret(const struct lu_env *env,
         }
 
         cli = aa->aa_cli;
-
         client_obd_list_lock(&cli->cl_loi_list_lock);
 
         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
@@ -2273,12 +2260,14 @@ static int brw_interpret(const struct lu_env *env,
                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
         }
         osc_wake_cache_waiters(cli);
-        osc_check_rpcs(env, cli);
+        osc_check_rpcs0(env, cli, 1);
         client_obd_list_unlock(&cli->cl_loi_list_lock);
+
         if (!async)
                 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
                                   req->rq_bulk->bd_nob_transferred);
         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
+        ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
 
         RETURN(rc);
 }
@@ -2293,10 +2282,8 @@ static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
         struct osc_brw_async_args *aa;
         struct obdo *oa = NULL;
         const struct obd_async_page_ops *ops = NULL;
-        void *caller_data = NULL;
         struct osc_async_page *oap;
         struct osc_async_page *tmp;
-        struct ost_body *body;
         struct cl_req *clerq = NULL;
         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
         struct ldlm_lock *lock = NULL;
@@ -2323,7 +2310,6 @@ static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
                 struct cl_page *page = osc_oap2cl_page(oap);
                 if (ops == NULL) {
                         ops = oap->oap_caller_ops;
-                        caller_data = oap->oap_caller_data;
 
                         clerq = cl_req_alloc(env, page, crt,
                                              1 /* only 1-object rpcs for
@@ -2372,7 +2358,6 @@ static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
          * later setattr before earlier BRW (as determined by the request xid),
          * the OST will not use BRW timestamps.  Sadly, there is no obvious
          * way to do this in a single call.  bug 10150 */
-        body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
         cl_req_attr_set(env, clerq, &crattr,
                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
 
@@ -2426,8 +2411,8 @@ out:
  */
 static int
 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
-                 struct lov_oinfo *loi,
-                 int cmd, struct loi_oap_pages *lop)
+                 struct lov_oinfo *loi, int cmd,
+                 struct loi_oap_pages *lop, pdl_policy_t pol)
 {
         struct ptlrpc_request *req;
         obd_count page_count = 0;
@@ -2447,10 +2432,14 @@ osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
          * with ASYNC_HP. We have to send out them as soon as possible. */
         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
                 if (oap->oap_async_flags & ASYNC_HP)
-                        cfs_list_move(&oap->oap_pending_item, &lop->lop_pending);
+                        cfs_list_move(&oap->oap_pending_item, &rpc_list);
+                else if (!(oap->oap_brw_flags & OBD_BRW_SYNC))
+                        /* only do this for writeback pages. */
+                        cfs_list_move_tail(&oap->oap_pending_item, &rpc_list);
                 if (++page_count >= cli->cl_max_pages_per_rpc)
                         break;
         }
+        cfs_list_splice_init(&rpc_list, &lop->lop_pending);
         page_count = 0;
 
         /* first we find the pages we're allowed to work with */
@@ -2536,27 +2525,6 @@ osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
                 }
                 if (oap == NULL)
                         break;
-                /*
-                 * Page submitted for IO has to be locked. Either by
-                 * ->ap_make_ready() or by higher layers.
-                 */
-#if defined(__KERNEL__) && defined(__linux__)
-                {
-                        struct cl_page *page;
-
-                        page = osc_oap2cl_page(oap);
-
-                        if (page->cp_type == CPT_CACHEABLE &&
-                            !(PageLocked(oap->oap_page) &&
-                              (CheckWriteback(oap->oap_page, cmd)))) {
-                                CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
-                                       oap->oap_page,
-                                       (long)oap->oap_page->flags,
-                                       oap->oap_async_flags);
-                                LBUG();
-                        }
-                }
-#endif
 
                 /* take the page out of our book-keeping */
                 cfs_list_del_init(&oap->oap_pending_item);
@@ -2642,7 +2610,6 @@ osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
         }
-        ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
 
         client_obd_list_lock(&cli->cl_loi_list_lock);
 
@@ -2671,7 +2638,19 @@ osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
 
         req->rq_interpret_reply = brw_interpret;
-        ptlrpcd_add_req(req, PSCOPE_BRW);
+
+        /* XXX: Maybe the caller can check the RPC bulk descriptor to see which
+         *      CPU/NUMA node the majority of pages were allocated on, and try
+         *      to assign the async RPC to the CPU core (PDL_POLICY_PREFERRED)
+         *      to reduce cross-CPU memory traffic.
+         *
+         *      But on the other hand, we expect that multiple ptlrpcd threads
+         *      and the initial write sponsor can run in parallel, especially
+         *      when data checksum is enabled, which is CPU-bound operation and
+         *      single ptlrpcd thread cannot process in time. So more ptlrpcd
+         *      threads sharing BRW load (with PDL_POLICY_ROUND) seems better.
+         */
+        ptlrpcd_add_req(req, pol, -1);
         RETURN(1);
 }
 
@@ -2745,12 +2724,15 @@ static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
 }
 
 /* called with the loi list lock held */
-void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
+static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli, int ptlrpc)
 {
         struct lov_oinfo *loi;
         int rc = 0, race_counter = 0;
+        pdl_policy_t pol;
         ENTRY;
 
+        pol = ptlrpc ? PDL_POLICY_SAME : PDL_POLICY_ROUND;
+
         while ((loi = osc_next_loi(cli)) != NULL) {
                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
 
@@ -2765,7 +2747,7 @@ void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
                  * do io on writes while there are cache waiters */
                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
-                                              &loi->loi_write_lop);
+                                              &loi->loi_write_lop, pol);
                         if (rc < 0) {
                                 CERROR("Write request failed with %d\n", rc);
 
@@ -2795,7 +2777,7 @@ void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
                 }
                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
-                                              &loi->loi_read_lop);
+                                              &loi->loi_read_lop, pol);
                         if (rc < 0)
                                 CERROR("Read request failed with %d\n", rc);
 
@@ -2826,7 +2808,11 @@ void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
                 if (race_counter == 10)
                         break;
         }
-        EXIT;
+}
+
+void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
+{
+        osc_check_rpcs0(env, cli, 0);
 }
 
 /* we're trying to queue a page in the osc so we're subject to the
@@ -3004,7 +2990,7 @@ int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
                 qid[USRQUOTA] = attr.cat_uid;
                 qid[GRPQUOTA] = attr.cat_gid;
                 if (rc == 0 &&
-                    lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
+                    osc_quota_chkdq(cli, qid) == NO_QUOTA)
                         rc = -EDQUOT;
                 if (rc)
                         RETURN(rc);
@@ -3035,13 +3021,19 @@ int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
                 }
         }
 
-        osc_oap_to_pending(oap);
-        loi_list_maint(cli, loi);
-
         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
                   cmd);
 
-        osc_check_rpcs(env, cli);
+        osc_oap_to_pending(oap);
+        loi_list_maint(cli, loi);
+        if (!osc_max_rpc_in_flight(cli, loi) &&
+            lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
+                LASSERT(cli->cl_writeback_work != NULL);
+                rc = ptlrpcd_queue_work(cli->cl_writeback_work);
+
+                CDEBUG(D_CACHE, "Queued writeback work for client obd %p/%d.\n",
+                       cli, rc);
+        }
         client_obd_list_unlock(&cli->cl_loi_list_lock);
 
         RETURN(0);
@@ -3210,7 +3202,7 @@ static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
 
 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
                             obd_enqueue_update_f upcall, void *cookie,
-                            int *flags, int rc)
+                            int *flags, int agl, int rc)
 {
         int intent = *flags & LDLM_FL_HAS_INTENT;
         ENTRY;
@@ -3228,7 +3220,8 @@ static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
                 }
         }
 
-        if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
+        if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
+            (rc == 0)) {
                 *flags |= LDLM_FL_LVB_READY;
                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
@@ -3246,6 +3239,9 @@ static int osc_enqueue_interpret(const struct lu_env *env,
         struct ldlm_lock *lock;
         struct lustre_handle handle;
         __u32 mode;
+        struct ost_lvb *lvb;
+        __u32 lvb_len;
+        int *flags = aa->oa_flags;
 
         /* Make a local copy of a lock handle and a mode, because aa->oa_*
          * might be freed anytime after lock upcall has been called. */
@@ -3265,13 +3261,20 @@ static int osc_enqueue_interpret(const struct lu_env *env,
         /* Let CP AST to grant the lock first. */
         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
 
+        if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
+                lvb = NULL;
+                lvb_len = 0;
+        } else {
+                lvb = aa->oa_lvb;
+                lvb_len = sizeof(*aa->oa_lvb);
+        }
+
         /* Complete obtaining the lock procedure. */
         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
-                                   mode, aa->oa_flags, aa->oa_lvb,
-                                   sizeof(*aa->oa_lvb), &handle, rc);
+                                   mode, flags, lvb, lvb_len, &handle, rc);
         /* Complete osc stuff. */
-        rc = osc_enqueue_fini(req, aa->oa_lvb,
-                              aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
+        rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
+                              flags, aa->oa_agl, rc);
 
         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
 
@@ -3295,8 +3298,9 @@ void osc_update_enqueue(struct lustre_handle *lov_lockhp,
                         struct lov_oinfo *loi, int flags,
                         struct ost_lvb *lvb, __u32 mode, int rc)
 {
+        struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
+
         if (rc == ELDLM_OK) {
-                struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
                 __u64 tmp;
 
                 LASSERT(lock != NULL);
@@ -3317,13 +3321,21 @@ void osc_update_enqueue(struct lustre_handle *lov_lockhp,
                                    lock->l_policy_data.l_extent.end);
                 }
                 ldlm_lock_allow_match(lock);
-                LDLM_LOCK_PUT(lock);
         } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
+                LASSERT(lock != NULL);
                 loi->loi_lvb = *lvb;
+                ldlm_lock_allow_match(lock);
                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
                        " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
                 rc = ELDLM_OK;
         }
+
+        if (lock != NULL) {
+                if (rc != ELDLM_OK)
+                        ldlm_lock_fail_match(lock);
+
+                LDLM_LOCK_PUT(lock);
+        }
 }
 EXPORT_SYMBOL(osc_update_enqueue);
 
@@ -3342,11 +3354,12 @@ int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
                      obd_enqueue_update_f upcall, void *cookie,
                      struct ldlm_enqueue_info *einfo,
                      struct lustre_handle *lockh,
-                     struct ptlrpc_request_set *rqset, int async)
+                     struct ptlrpc_request_set *rqset, int async, int agl)
 {
         struct obd_device *obd = exp->exp_obd;
         struct ptlrpc_request *req = NULL;
         int intent = *flags & LDLM_FL_HAS_INTENT;
+        int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
         ldlm_mode_t mode;
         int rc;
         ENTRY;
@@ -3380,13 +3393,20 @@ int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
         mode = einfo->ei_mode;
         if (einfo->ei_mode == LCK_PR)
                 mode |= LCK_PW;
-        mode = ldlm_lock_match(obd->obd_namespace,
-                               *flags | LDLM_FL_LVB_READY, res_id,
+        mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
                                einfo->ei_type, policy, mode, lockh, 0);
         if (mode) {
                 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
 
-                if (osc_set_lock_data_with_check(matched, einfo)) {
+                if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
+                        /* For AGL, if enqueue RPC is sent but the lock is not
+                         * granted, then skip to process this strpe.
+                         * Return -ECANCELED to tell the caller. */
+                        ldlm_lock_decref(lockh, mode);
+                        LDLM_LOCK_PUT(matched);
+                        RETURN(-ECANCELED);
+                } else if (osc_set_lock_data_with_check(matched, einfo)) {
+                        *flags |= LDLM_FL_LVB_READY;
                         /* addref the lock only if not async requests and PW
                          * lock is matched whereas we asked for PR. */
                         if (!rqset && einfo->ei_mode != mode)
@@ -3400,16 +3420,17 @@ int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
                         /* We already have a lock, and it's referenced */
                         (*upcall)(cookie, ELDLM_OK);
 
-                        /* For async requests, decref the lock. */
                         if (einfo->ei_mode != mode)
                                 ldlm_lock_decref(lockh, LCK_PW);
                         else if (rqset)
+                                /* For async requests, decref the lock. */
                                 ldlm_lock_decref(lockh, einfo->ei_mode);
                         LDLM_LOCK_PUT(matched);
                         RETURN(ELDLM_OK);
-                } else
+                } else {
                         ldlm_lock_decref(lockh, mode);
-                LDLM_LOCK_PUT(matched);
+                        LDLM_LOCK_PUT(matched);
+                }
         }
 
  no_match:
@@ -3448,11 +3469,12 @@ int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
                         aa->oa_cookie = cookie;
                         aa->oa_lvb    = lvb;
                         aa->oa_lockh  = lockh;
+                        aa->oa_agl    = !!agl;
 
                         req->rq_interpret_reply =
                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
                         if (rqset == PTLRPCD_SET)
-                                ptlrpcd_add_req(req, PSCOPE_OTHER);
+                                ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
                         else
                                 ptlrpc_set_add_req(rqset, req);
                 } else if (intent) {
@@ -3461,7 +3483,7 @@ int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
                 RETURN(rc);
         }
 
-        rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
+        rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
         if (intent)
                 ptlrpc_req_finished(req);
 
@@ -3483,7 +3505,7 @@ static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
                               &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
                               oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
                               oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
-                              rqset, rqset != NULL);
+                              rqset, rqset != NULL, 0);
         RETURN(rc);
 }
 
@@ -3886,8 +3908,7 @@ static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                                                data->ioc_offset);
                 GOTO(out, err);
         case OBD_IOC_POLL_QUOTACHECK:
-                err = lquota_poll_check(quota_interface, exp,
-                                        (struct if_quotacheck *)karg);
+                err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
                 GOTO(out, err);
         case OBD_IOC_PING_TARGET:
                 err = ptlrpc_obd_ping(obd);
@@ -4152,7 +4173,7 @@ static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
                 ptlrpc_set_add_req(set, req);
                 ptlrpc_check_set(NULL, set);
         } else
-                ptlrpcd_add_req(req, PSCOPE_OTHER);
+                ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 
         RETURN(0);
 }
@@ -4447,6 +4468,7 @@ static int osc_cancel_for_recovery(struct ldlm_lock *lock)
 
 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 {
+        struct client_obd *cli = &obd->u.cli;
         int rc;
         ENTRY;
 
@@ -4456,11 +4478,18 @@ int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
                 RETURN(rc);
 
         rc = client_obd_setup(obd, lcfg);
-        if (rc) {
-                ptlrpcd_decref();
-        } else {
+        if (rc == 0) {
+                void *handler;
+                handler = ptlrpcd_alloc_work(cli->cl_import,
+                                             brw_queue_work, cli);
+                if (!IS_ERR(handler))
+                        cli->cl_writeback_work = handler;
+                else
+                        rc = PTR_ERR(handler);
+        }
+
+        if (rc == 0) {
                 struct lprocfs_static_vars lvars = { 0 };
-                struct client_obd *cli = &obd->u.cli;
 
                 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
                 lprocfs_osc_init_vars(&lvars);
@@ -4487,6 +4516,8 @@ int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
                 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
         }
 
+        if (rc)
+                ptlrpcd_decref();
         RETURN(rc);
 }
 
@@ -4508,6 +4539,7 @@ static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
                 break;
         }
         case OBD_CLEANUP_EXPORTS: {
+                struct client_obd *cli = &obd->u.cli;
                 /* LU-464
                  * for echo client, export may be on zombie list, wait for
                  * zombie thread to cull it, because cli.cl_import will be
@@ -4518,7 +4550,13 @@ static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
                  *   client_disconnect_export()
                  */
                 obd_zombie_barrier();
+                if (cli->cl_writeback_work) {
+                        ptlrpcd_destroy_work(cli->cl_writeback_work);
+                        cli->cl_writeback_work = NULL;
+                }
                 obd_cleanup_client_import(obd);
+                ptlrpc_lprocfs_unregister_obd(obd);
+                lprocfs_obd_cleanup(obd);
                 rc = obd_llog_finish(obd, 0);
                 if (rc != 0)
                         CERROR("failed to cleanup llogging subsystems\n");
@@ -4533,11 +4571,9 @@ int osc_cleanup(struct obd_device *obd)
         int rc;
 
         ENTRY;
-        ptlrpc_lprocfs_unregister_obd(obd);
-        lprocfs_obd_cleanup(obd);
 
         /* free memory of osc quota cache */
-        lquota_cleanup(quota_interface, obd);
+        osc_quota_cleanup(obd);
 
         rc = client_obd_cleanup(obd);
 
@@ -4606,6 +4642,9 @@ struct obd_ops osc_obd_ops = {
         .o_llog_init            = osc_llog_init,
         .o_llog_finish          = osc_llog_finish,
         .o_process_config       = osc_process_config,
+        .o_quotactl             = osc_quotactl,
+        .o_quotacheck           = osc_quotacheck,
+        .o_quota_adjust_qunit   = osc_quota_adjust_qunit,
 };
 
 extern struct lu_kmem_descr osc_caches[];
@@ -4621,22 +4660,16 @@ int __init osc_init(void)
         /* print an address of _any_ initialized kernel symbol from this
          * module, to allow debugging with gdb that doesn't support data
          * symbols from modules.*/
-        CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
+        CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
 
         rc = lu_kmem_init(osc_caches);
 
         lprocfs_osc_init_vars(&lvars);
 
-        cfs_request_module("lquota");
-        quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
-        lquota_init(quota_interface);
-        init_obd_quota_ops(quota_interface, &osc_obd_ops);
-
+        osc_quota_init();
         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
                                  LUSTRE_OSC_NAME, &osc_device_type);
         if (rc) {
-                if (quota_interface)
-                        PORTAL_SYMBOL_PUT(osc_quota_interface);
                 lu_kmem_fini(osc_caches);
                 RETURN(rc);
         }
@@ -4658,10 +4691,7 @@ static void /*__exit*/ osc_exit(void)
 {
         lu_device_type_fini(&osc_device_type);
 
-        lquota_exit(quota_interface);
-        if (quota_interface)
-                PORTAL_SYMBOL_PUT(osc_quota_interface);
-
+        osc_quota_exit();
         class_unregister_type(LUSTRE_OSC_NAME);
         lu_kmem_fini(osc_caches);
 }