LU-2139 osc: Use SOFT_SYNC to urge server commit

[fs/lustre-release.git] / lustre / osc / osc_request.c

diff --git a/lustre/osc/osc_request.c b/lustre/osc/osc_request.c
index 44ff796..2aa6b30 100644 (file)
--- a/lustre/osc/osc_request.c
+++ b/lustre/osc/osc_request.c
@@ -838,13 +838,16 @@ static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
                CERROR("dirty %lu - %lu > dirty_max %lu\n",
                       cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
                oa->o_undirty = 0;
-       } else if (unlikely(cfs_atomic_read(&obd_dirty_pages) -
+       } else if (unlikely(cfs_atomic_read(&obd_unstable_pages) +
+                           cfs_atomic_read(&obd_dirty_pages) -
                            cfs_atomic_read(&obd_dirty_transit_pages) >
                            (long)(obd_max_dirty_pages + 1))) {
                /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
                 * not covered by a lock thus they may safely race and trip
                 * this CERROR() unless we add in a small fudge factor (+1). */
-               CERROR("dirty %d - %d > system dirty_max %d\n",
+               CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
+                      cli->cl_import->imp_obd->obd_name,
+                      cfs_atomic_read(&obd_unstable_pages),
                       cfs_atomic_read(&obd_dirty_pages),
                       cfs_atomic_read(&obd_dirty_transit_pages),
                       obd_max_dirty_pages);
@@ -1169,8 +1172,9 @@ 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|OBD_BRW_NOQUOTA);
+               unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
+                                 OBD_BRW_SYNC       | OBD_BRW_ASYNC   |
+                                 OBD_BRW_NOQUOTA    | OBD_BRW_SOFT_SYNC);
 
                 /* warn if we try to combine flags that we don't know to be
                  * safe to combine */
@@ -1758,6 +1762,7 @@ static int osc_brw_redo_request(struct ptlrpc_request *request,
         aa->aa_resends++;
         new_req->rq_interpret_reply = request->rq_interpret_reply;
         new_req->rq_async_args = request->rq_async_args;
+       new_req->rq_commit_cb = request->rq_commit_cb;
        /* cap resend delay to the current request timeout, this is similar to
         * what ptlrpc does (see after_reply()) */
        if (aa->aa_resends > new_req->rq_timeout)
@@ -2051,6 +2056,23 @@ static int brw_interpret(const struct lu_env *env,
        RETURN(rc);
 }
 
+static void brw_commit(struct ptlrpc_request *req)
+{
+       spin_lock(&req->rq_lock);
+       /* If osc_inc_unstable_pages (via osc_extent_finish) races with
+        * this called via the rq_commit_cb, I need to ensure
+        * osc_dec_unstable_pages is still called. Otherwise unstable
+        * pages may be leaked. */
+       if (req->rq_unstable) {
+               spin_unlock(&req->rq_lock);
+               osc_dec_unstable_pages(req);
+               spin_lock(&req->rq_lock);
+       } else {
+               req->rq_committed = 1;
+       }
+       spin_unlock(&req->rq_lock);
+}
+
 /**
  * Build an RPC by the list of extent @ext_list. The caller must ensure
  * that the total pages in this list are NOT over max pages per RPC.
@@ -2162,7 +2184,9 @@ int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
                GOTO(out, rc);
        }
 
+       req->rq_commit_cb = brw_commit;
        req->rq_interpret_reply = brw_interpret;
+
        if (mem_tight != 0)
                req->rq_memalloc = 1;