- ldiskfs osd doesn't wait for write io to complete in osd_do_bio(), instead

  it does in osd_trans_stop() to allow concurrent data IO and IO to journal

diff --git a/lustre/dmu-osd/osd_handler.c b/lustre/dmu-osd/osd_handler.c
index 0692bde..26305c5 100644 (file)
--- a/lustre/dmu-osd/osd_handler.c
+++ b/lustre/dmu-osd/osd_handler.c
@@ -193,7 +193,7 @@ static int   osd_index_try     (const struct lu_env *env,
 static void  osd_conf_get      (const struct lu_env *env,
                                 const struct dt_device *dev,
                                 struct dt_device_param *param);
-static void  osd_trans_stop    (const struct lu_env *env,
+static int   osd_trans_stop    (const struct lu_env *env,
                                 struct thandle *th);
 static int   osd_object_is_root(const struct osd_object *obj);
 
@@ -201,7 +201,6 @@ static struct thandle *osd_trans_create(const struct lu_env *env,
                                        struct dt_device *dt);
 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
                            struct thandle *th);
-static void osd_trans_stop(const struct lu_env *env, struct thandle *th);
 
 static struct osd_object  *osd_obj          (const struct lu_object *o);
 static struct osd_device  *osd_dev          (const struct lu_device *d);
@@ -832,7 +831,7 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
 /*
  * Concurrency: shouldn't matter.
  */
-static void osd_trans_stop(const struct lu_env *env, struct thandle *th)
+static int osd_trans_stop(const struct lu_env *env, struct thandle *th)
 {
         struct osd_device  *osd = osd_dt_dev(th->th_dev);
         struct osd_thandle *oh;
@@ -848,7 +847,8 @@ static void osd_trans_stop(const struct lu_env *env, struct thandle *th)
         udmu_tx_commit(oh->ot_tx);
         if (0 && oh->ot_sync)
                 udmu_wait_synced(&osd->od_objset, oh->ot_tx);
-        EXIT;
+
+        RETURN(result);
 }
 
 /*

diff --git a/lustre/include/dt_object.h b/lustre/include/dt_object.h
index d5fef26..75425a4 100644 (file)
--- a/lustre/include/dt_object.h
+++ b/lustre/include/dt_object.h
@@ -99,7 +99,7 @@ struct dt_device_operations {
         /**
          * Finish previously started transaction.
          */
-        void  (*dt_trans_stop)(const struct lu_env *env,
+        int   (*dt_trans_stop)(const struct lu_env *env,
                                struct thandle *th);
         /**
          * Return fid of root index object.
@@ -715,7 +715,7 @@ static inline int dt_trans_start(const struct lu_env *env,
         return d->dd_ops->dt_trans_start(env, d, th);
 }
 
-static inline void dt_trans_stop(const struct lu_env *env,
+static inline int dt_trans_stop(const struct lu_env *env,
                                  struct dt_device *d,
                                  struct thandle *th)
 {

diff --git a/lustre/osd/osd_handler.c b/lustre/osd/osd_handler.c
index 38ed948..ad6dbe5 100644 (file)
--- a/lustre/osd/osd_handler.c
+++ b/lustre/osd/osd_handler.c
@@ -183,7 +183,7 @@ static int   osd_it_ea_key_size(const struct lu_env *env,
 static void  osd_conf_get      (const struct lu_env *env,
                                 const struct dt_device *dev,
                                 struct dt_device_param *param);
-static void  osd_trans_stop    (const struct lu_env *env,
+static int   osd_trans_stop    (const struct lu_env *env,
                                 struct thandle *th);
 static int   osd_object_is_root(const struct osd_object *obj);
 
@@ -778,11 +778,12 @@ out:
 /*
  * Concurrency: shouldn't matter.
  */
-static void osd_trans_stop(const struct lu_env *env, struct thandle *th)
+static int osd_trans_stop(const struct lu_env *env, struct thandle *th)
 {
-        int result;
-        struct osd_thandle *oh;
+        int                     result;
+        struct osd_thandle     *oh;
         struct osd_thread_info *oti = osd_oti_get(env);
+        struct filter_iobuf    *iobuf = &oti->oti_iobuf;
 
         ENTRY;
 
@@ -822,7 +823,19 @@ static void osd_trans_stop(const struct lu_env *env, struct thandle *th)
         } else {
                 OBD_FREE_PTR(oh);
         }
-        EXIT;
+
+        /* as we want IO to journal and data IO be concurrent, we don't block
+         * awaiting data IO completion in osd_do_bio(), instead we wait here
+         * once transaction is submitted to the journal.
+         *
+         * IMPORTANT: we have to wait till any IO submited by the thread is
+         * completed otherwise iobuf may be corrupted by different request
+         */
+        wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
+        if (!result)
+                result = iobuf->dr_error;
+
+        RETURN(result);
 }
 
 /*

diff --git a/lustre/osd/osd_io.c b/lustre/osd/osd_io.c
index 79df0b7..1752e54 100644 (file)
--- a/lustre/osd/osd_io.c
+++ b/lustre/osd/osd_io.c
@@ -328,7 +328,12 @@ static int osd_do_bio(struct inode *inode, struct filter_iobuf *iobuf, int rw)
         }
 
  out:
-        wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
+        /* in order to achieve better IO throughput, we don't wait for writes
+         * completion here. instead we proceed with transaction commit in
+         * parallel and wait for IO completion once transaction is stopped
+         * see osd_trans_stop() for more details -bzzz */
+        if (rw == OBD_BRW_WRITE)
+                wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 
         if (rc == 0)
                 rc = iobuf->dr_error;
@@ -540,6 +545,7 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 
                 /* preceding filemap_write_and_wait() should have clean pages */
 #if 0
+                /* XXX */
                 if (fo->fo_writethrough_cache)
                         clear_page_dirty_for_io(lb[i].page);
 #endif