Whamcloud - gitweb
b=22755 Don't consume grant twice on recoverable resend
[fs/lustre-release.git] / lustre / osc / osc_request.c
index 7144f0b..a01df94 100644 (file)
@@ -864,6 +864,17 @@ static unsigned long rpcs_in_flight(struct client_obd *cli)
         return cli->cl_r_in_flight + cli->cl_w_in_flight;
 }
 
+int osc_wake_sync_fs(struct client_obd *cli)
+{
+        ENTRY;
+        if (cfs_list_empty(&cli->cl_loi_sync_fs_list) &&
+            cli->cl_sf_wait.started) {
+                cli->cl_sf_wait.sfw_upcall(cli->cl_sf_wait.sfw_oi, 0);
+                cli->cl_sf_wait.started = 0;
+        }
+        RETURN(0);
+}
+
 /* caller must hold loi_list_lock */
 void osc_wake_cache_waiters(struct client_obd *cli)
 {
@@ -953,6 +964,10 @@ static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
         oa->o_grant = cli->cl_avail_grant / 4;
         cli->cl_avail_grant -= oa->o_grant;
         client_obd_list_unlock(&cli->cl_loi_list_lock);
+        if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
+                oa->o_valid |= OBD_MD_FLFLAGS;
+                oa->o_flags = 0;
+        }
         oa->o_flags |= OBD_FL_SHRINK_GRANT;
         osc_update_next_shrink(cli);
 }
@@ -1003,6 +1018,10 @@ int osc_shrink_grant_to_target(struct client_obd *cli, long target)
         body->oa.o_grant = cli->cl_avail_grant - target;
         cli->cl_avail_grant = target;
         client_obd_list_unlock(&cli->cl_loi_list_lock);
+        if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
+                body->oa.o_valid |= OBD_MD_FLFLAGS;
+                body->oa.o_flags = 0;
+        }
         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
         osc_update_next_shrink(cli);
 
@@ -1240,7 +1259,8 @@ static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
                                 struct lov_stripe_md *lsm, obd_count page_count,
                                 struct brw_page **pga,
                                 struct ptlrpc_request **reqp,
-                                struct obd_capa *ocapa, int reserve)
+                                struct obd_capa *ocapa, int reserve,
+                                int resend)
 {
         struct ptlrpc_request   *req;
         struct ptlrpc_bulk_desc *desc;
@@ -1356,6 +1376,14 @@ static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
 
         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
+        if (resend) {
+                if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
+                        body->oa.o_valid |= OBD_MD_FLFLAGS;
+                        body->oa.o_flags = 0;
+                }
+                body->oa.o_flags |= OBD_FL_RECOV_RESEND;
+        }
+
         if (osc_should_shrink_grant(cli))
                 osc_shrink_grant_local(cli, &body->oa);
 
@@ -1658,7 +1686,7 @@ static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
 
 restart_bulk:
         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
-                                  page_count, pga, &req, ocapa, 0);
+                                  page_count, pga, &req, ocapa, 0, resends);
         if (rc != 0)
                 return (rc);
 
@@ -1711,7 +1739,7 @@ int osc_brw_redo_request(struct ptlrpc_request *request,
                                   aa->aa_cli, aa->aa_oa,
                                   NULL /* lsm unused by osc currently */,
                                   aa->aa_page_count, aa->aa_ppga,
-                                  &new_req, aa->aa_ocapa, 0);
+                                  &new_req, aa->aa_ocapa, 0, 1);
         if (rc)
                 RETURN(rc);
 
@@ -1924,6 +1952,24 @@ static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
 }
 
+static int lop_makes_syncfs_rpc(struct loi_oap_pages *lop)
+{
+        struct osc_async_page *oap;
+        ENTRY;
+
+        if (cfs_list_empty(&lop->lop_urgent))
+                RETURN(0);
+
+        oap = cfs_list_entry(lop->lop_urgent.next,
+                             struct osc_async_page, oap_urgent_item);
+
+        if (oap->oap_async_flags & ASYNC_SYNCFS) {
+                CDEBUG(D_CACHE, "syncfs request forcing RPC\n");
+                RETURN(1);
+        }
+
+        RETURN(0);
+}
 
 /* This maintains the lists of pending pages to read/write for a given object
  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
@@ -2012,10 +2058,19 @@ void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
         } else {
-                on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
-                on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
-                        lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
-                        lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
+                if (lop_makes_syncfs_rpc(&loi->loi_write_lop)) {
+                        on_list(&loi->loi_sync_fs_item,
+                                &cli->cl_loi_sync_fs_list,
+                                loi->loi_write_lop.lop_num_pending);
+                } else {
+                        on_list(&loi->loi_hp_ready_item,
+                                &cli->cl_loi_hp_ready_list, 0);
+                        on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
+                                lop_makes_rpc(cli, &loi->loi_write_lop,
+                                              OBD_BRW_WRITE)||
+                                lop_makes_rpc(cli, &loi->loi_read_lop,
+                                              OBD_BRW_READ));
+                }
         }
 
         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
@@ -2103,6 +2158,34 @@ static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
                 ar->ar_force_sync = 0;
 }
 
+static int osc_add_to_lop_urgent(struct loi_oap_pages *lop,
+                                 struct osc_async_page *oap,
+                                 obd_flag async_flags)
+{
+
+        /* If true, then already present in lop urgent */
+        if (!cfs_list_empty(&oap->oap_urgent_item)) {
+                CWARN("Request to add duplicate oap_urgent for flag = %d\n",
+                       oap->oap_async_flags);
+                return 1;
+        }
+
+        /* item from sync_fs, to avoid duplicates check the existing flags */
+        if (async_flags & ASYNC_SYNCFS) {
+                cfs_list_add_tail(&oap->oap_urgent_item,
+                                  &lop->lop_urgent);
+                return 0;
+        }
+
+        if (oap->oap_async_flags & ASYNC_HP)
+                cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
+        else if (oap->oap_async_flags & ASYNC_URGENT ||
+                 async_flags & ASYNC_URGENT)
+                cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
+
+        return 0;
+}
+
 void osc_oap_to_pending(struct osc_async_page *oap)
 {
         struct loi_oap_pages *lop;
@@ -2112,10 +2195,7 @@ void osc_oap_to_pending(struct osc_async_page *oap)
         else
                 lop = &oap->oap_loi->loi_read_lop;
 
-        if (oap->oap_async_flags & ASYNC_HP)
-                cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
-        else if (oap->oap_async_flags & ASYNC_URGENT)
-                cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
+        osc_add_to_lop_urgent(lop, oap, 0);
         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
 }
@@ -2231,6 +2311,7 @@ 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_wake_sync_fs(cli);
         osc_check_rpcs(env, cli);
         client_obd_list_unlock(&cli->cl_loi_list_lock);
         if (!async)
@@ -2315,7 +2396,7 @@ static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
 
         sort_brw_pages(pga, page_count);
         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
-                                  pga, &req, crattr.cra_capa, 1);
+                                  pga, &req, crattr.cra_capa, 1, 0);
         if (rc != 0) {
                 CERROR("prep_req failed: %d\n", rc);
                 GOTO(out, req = ERR_PTR(rc));
@@ -2563,7 +2644,7 @@ osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
         }
 
         osc_wake_cache_waiters(cli);
-
+        osc_wake_sync_fs(cli);
         loi_list_maint(cli, loi);
 
         client_obd_list_unlock(&cli->cl_loi_list_lock);
@@ -2656,6 +2737,9 @@ struct lov_oinfo *osc_next_loi(struct client_obd *cli)
         if (!cfs_list_empty(&cli->cl_loi_ready_list))
                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
                                       struct lov_oinfo, loi_ready_item));
+        if (!cfs_list_empty(&cli->cl_loi_sync_fs_list))
+                RETURN(cfs_list_entry(cli->cl_loi_sync_fs_list.next,
+                                      struct lov_oinfo, loi_sync_fs_item));
 
         /* then if we have cache waiters, return all objects with queued
          * writes.  This is especially important when many small files
@@ -2771,6 +2855,8 @@ void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
                         cfs_list_del_init(&loi->loi_write_item);
                 if (!cfs_list_empty(&loi->loi_read_item))
                         cfs_list_del_init(&loi->loi_read_item);
+                if (!cfs_list_empty(&loi->loi_sync_fs_item))
+                        cfs_list_del_init(&loi->loi_sync_fs_item);
 
                 loi_list_maint(cli, loi);
 
@@ -3035,16 +3121,27 @@ int osc_set_async_flags_base(struct client_obd *cli,
         if ((oap->oap_async_flags & async_flags) == async_flags)
                 RETURN(0);
 
+        /* XXX: This introduces a tiny insignificant race for the case if this
+         * loi already had other urgent items.
+         */
+        if (SETTING(oap->oap_async_flags, async_flags, ASYNC_SYNCFS) &&
+            cfs_list_empty(&oap->oap_rpc_item) &&
+            cfs_list_empty(&oap->oap_urgent_item)) {
+                osc_add_to_lop_urgent(lop, oap, ASYNC_SYNCFS);
+                flags |= ASYNC_SYNCFS;
+                cfs_spin_lock(&oap->oap_lock);
+                oap->oap_async_flags |= flags;
+                cfs_spin_unlock(&oap->oap_lock);
+                loi_list_maint(cli, loi);
+                RETURN(0);
+        }
+
         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
                 flags |= ASYNC_READY;
 
         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
             cfs_list_empty(&oap->oap_rpc_item)) {
-                if (oap->oap_async_flags & ASYNC_HP)
-                        cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
-                else
-                        cfs_list_add_tail(&oap->oap_urgent_item,
-                                          &lop->lop_urgent);
+                osc_add_to_lop_urgent(lop, oap, ASYNC_URGENT);
                 flags |= ASYNC_URGENT;
                 loi_list_maint(cli, loi);
         }
@@ -3091,7 +3188,8 @@ int osc_teardown_async_page(struct obd_export *exp,
         if (!cfs_list_empty(&oap->oap_urgent_item)) {
                 cfs_list_del_init(&oap->oap_urgent_item);
                 cfs_spin_lock(&oap->oap_lock);
-                oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
+                oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP |
+                                          ASYNC_SYNCFS);
                 cfs_spin_unlock(&oap->oap_lock);
         }
         if (!cfs_list_empty(&oap->oap_pending_item)) {
@@ -4513,6 +4611,46 @@ int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
         return(rc);
 }
 
+static int osc_sync_fs(struct obd_export *exp, struct obd_info *oinfo,
+                       int wait)
+{
+        struct obd_device *obd = class_exp2obd(exp);
+        struct client_obd *cli;
+        struct lov_oinfo *loi;
+        struct lov_oinfo *tloi;
+        struct osc_async_page *oap;
+        struct osc_async_page *toap;
+        struct loi_oap_pages *lop;
+        struct lu_env *env;
+        int refcheck;
+        int rc = 0;
+        ENTRY;
+
+        env = cl_env_get(&refcheck);
+        if (IS_ERR(env))
+                RETURN(PTR_ERR(env));
+
+        cli = &obd->u.cli;
+        client_obd_list_lock(&cli->cl_loi_list_lock);
+        cli->cl_sf_wait.sfw_oi = oinfo;
+        cli->cl_sf_wait.sfw_upcall = oinfo->oi_cb_up;
+        cli->cl_sf_wait.started = 1;
+        /* creating cl_loi_sync_fs list */
+        cfs_list_for_each_entry_safe(loi, tloi, &cli->cl_loi_write_list,
+                                     loi_write_item) {
+                lop = &loi->loi_write_lop;
+                cfs_list_for_each_entry_safe(oap, toap, &lop->lop_pending,
+                                             oap_pending_item)
+                        osc_set_async_flags_base(cli, loi, oap, ASYNC_SYNCFS);
+        }
+        osc_check_rpcs(env, cli);
+        osc_wake_sync_fs(cli);
+        client_obd_list_unlock(&cli->cl_loi_list_lock);
+        cl_env_put(env, &refcheck);
+
+        RETURN(rc);
+}
+
 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
 {
         return osc_process_config_base(obd, buf);
@@ -4555,6 +4693,7 @@ 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_sync_fs              = osc_sync_fs,
 };
 
 extern struct lu_kmem_descr osc_caches[];