Whamcloud - gitweb
LU-4198 clio: AIO support for direct IO
[fs/lustre-release.git] / lustre / osc / osc_cache.c
index 8cf0158..8122fd1 100644 (file)
@@ -617,7 +617,8 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
        RETURN(rc);
 }
 
-static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
+static inline bool
+overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
 {
        return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 }
@@ -932,8 +933,6 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
                           enum osc_extent_state state)
 {
        struct osc_object *obj = ext->oe_obj;
-       struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
-                                                 LWI_ON_SIGNAL_NOOP, NULL);
        int rc = 0;
        ENTRY;
 
@@ -955,18 +954,19 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
                osc_extent_release(env, ext);
 
        /* wait for the extent until its state becomes @state */
-       rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
-       if (rc == -ETIMEDOUT) {
+       rc = wait_event_idle_timeout(ext->oe_waitq, extent_wait_cb(ext, state),
+                                    cfs_time_seconds(600));
+       if (rc == 0) {
                OSC_EXTENT_DUMP(D_ERROR, ext,
                        "%s: wait ext to %u timedout, recovery in progress?\n",
                        cli_name(osc_cli(obj)), state);
 
-               lwi = LWI_INTR(NULL, NULL);
-               rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
-                                 &lwi);
+               wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
        }
-       if (rc == 0 && ext->oe_rc < 0)
+       if (ext->oe_rc < 0)
                rc = ext->oe_rc;
+       else
+               rc = 0;
        RETURN(rc);
 }
 
@@ -1328,7 +1328,7 @@ static int osc_refresh_count(const struct lu_env *env,
                return 0;
        else if (cl_offset(obj, index + 1) > kms)
                /* catch sub-page write at end of file */
-               return kms % PAGE_SIZE;
+               return kms & ~PAGE_MASK;
        else
                return PAGE_SIZE;
 }
@@ -1428,11 +1428,6 @@ static void osc_release_write_grant(struct client_obd *cli,
        pga->flag &= ~OBD_BRW_FROM_GRANT;
        atomic_long_dec(&obd_dirty_pages);
        cli->cl_dirty_pages--;
-       if (pga->flag & OBD_BRW_NOCACHE) {
-               pga->flag &= ~OBD_BRW_NOCACHE;
-               atomic_long_dec(&obd_dirty_transit_pages);
-               cli->cl_dirty_transit--;
-       }
        EXIT;
 }
 
@@ -1545,7 +1540,7 @@ static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
  */
 static int osc_enter_cache_try(struct client_obd *cli,
                               struct osc_async_page *oap,
-                              int bytes, int transient)
+                              int bytes)
 {
        int rc;
 
@@ -1559,11 +1554,6 @@ static int osc_enter_cache_try(struct client_obd *cli,
                if (atomic_long_add_return(1, &obd_dirty_pages) <=
                    obd_max_dirty_pages) {
                        osc_consume_write_grant(cli, &oap->oap_brw_page);
-                       if (transient) {
-                               cli->cl_dirty_transit++;
-                               atomic_long_inc(&obd_dirty_transit_pages);
-                               oap->oap_brw_flags |= OBD_BRW_NOCACHE;
-                       }
                        rc = 1;
                        goto out;
                } else
@@ -1597,13 +1587,9 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
        struct osc_object       *osc = oap->oap_obj;
        struct lov_oinfo        *loi = osc->oo_oinfo;
        struct osc_cache_waiter  ocw;
-       struct l_wait_info       lwi;
        int                      rc = -EDQUOT;
        ENTRY;
 
-       lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
-                              NULL, LWI_ON_SIGNAL_NOOP, NULL);
-
        OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
 
        spin_lock(&cli->cl_loi_list_lock);
@@ -1619,7 +1605,7 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
 
        /* Hopefully normal case - cache space and write credits available */
        if (list_empty(&cli->cl_cache_waiters) &&
-           osc_enter_cache_try(cli, oap, bytes, 0)) {
+           osc_enter_cache_try(cli, oap, bytes)) {
                OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
                GOTO(out, rc = 0);
        }
@@ -1643,13 +1629,19 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
                CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
                       cli_name(cli), &ocw, oap);
 
-               rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
+               rc = wait_event_idle_timeout(ocw.ocw_waitq,
+                                            ocw_granted(cli, &ocw),
+                                            cfs_time_seconds(AT_OFF ?
+                                                             obd_timeout :
+                                                             at_max));
 
                spin_lock(&cli->cl_loi_list_lock);
 
-               if (rc < 0) {
+               if (rc <= 0) {
                        /* l_wait_event is interrupted by signal or timed out */
                        list_del_init(&ocw.ocw_entry);
+                       if (rc == 0)
+                               rc = -ETIMEDOUT;
                        break;
                }
                LASSERT(list_empty(&ocw.ocw_entry));
@@ -1657,7 +1649,7 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
 
                if (rc != -EDQUOT)
                        break;
-               if (osc_enter_cache_try(cli, oap, bytes, 0)) {
+               if (osc_enter_cache_try(cli, oap, bytes)) {
                        rc = 0;
                        break;
                }
@@ -1707,7 +1699,7 @@ void osc_wake_cache_waiters(struct client_obd *cli)
 
                ocw->ocw_rc = -EDQUOT;
 
-               if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
+               if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant))
                        ocw->ocw_rc = 0;
 
                if (ocw->ocw_rc == 0 ||
@@ -1927,6 +1919,31 @@ static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
        return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
 }
 
+static inline bool
+can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
+{
+       if (ext->oe_no_merge || in_rpc->oe_no_merge)
+               return false;
+
+       if (ext->oe_srvlock != in_rpc->oe_srvlock)
+               return false;
+
+       if (ext->oe_ndelay != in_rpc->oe_ndelay)
+               return false;
+
+       if (!ext->oe_grants != !in_rpc->oe_grants)
+               return false;
+
+       if (ext->oe_dio != in_rpc->oe_dio)
+               return false;
+
+       /* It's possible to have overlap on DIO */
+       if (in_rpc->oe_dio && overlapped(ext, in_rpc))
+               return false;
+
+       return true;
+}
+
 /**
  * Try to add extent to one RPC. We need to think about the following things:
  * - # of pages must not be over max_pages_per_rpc
@@ -1938,9 +1955,6 @@ static int try_to_add_extent_for_io(struct client_obd *cli,
 {
        struct osc_extent *tmp;
        unsigned int chunk_count;
-       struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
-                                                     struct osc_async_page,
-                                                     oap_pending_item);
        ENTRY;
 
        EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
@@ -1969,30 +1983,10 @@ static int try_to_add_extent_for_io(struct client_obd *cli,
                RETURN(0);
 
        list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
-               struct osc_async_page *oap2;
-               oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
-                                       oap_pending_item);
                EASSERT(tmp->oe_owner == current, tmp);
-#if 0
-               if (overlapped(tmp, ext)) {
-                       OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
-                       EASSERT(0, ext);
-               }
-#endif
-               if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
-                       CDEBUG(D_CACHE, "Do not permit different types of IO "
-                              "in one RPC\n");
-                       RETURN(0);
-               }
 
-               if (tmp->oe_srvlock != ext->oe_srvlock ||
-                   !tmp->oe_grants != !ext->oe_grants ||
-                   tmp->oe_ndelay != ext->oe_ndelay ||
-                   tmp->oe_no_merge || ext->oe_no_merge)
+               if (!can_merge(ext, tmp))
                        RETURN(0);
-
-               /* remove break for strict check */
-               break;
        }
 
        data->erd_max_extents--;
@@ -2086,7 +2080,7 @@ osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
                   struct osc_object *osc)
 __must_hold(osc)
 {
-       struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
+       LIST_HEAD(rpclist);
        struct osc_extent *ext;
        struct osc_extent *tmp;
        struct osc_extent *first = NULL;
@@ -2162,7 +2156,7 @@ __must_hold(osc)
 {
        struct osc_extent *ext;
        struct osc_extent *next;
-       struct list_head rpclist = LIST_HEAD_INIT(rpclist);
+       LIST_HEAD(rpclist);
        struct extent_rpc_data data = {
                .erd_rpc_list   = &rpclist,
                .erd_page_count = 0,
@@ -2356,13 +2350,14 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
 EXPORT_SYMBOL(osc_prep_async_page);
 
 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
-                      struct osc_page *ops)
+                      struct osc_page *ops, cl_commit_cbt cb)
 {
        struct osc_io *oio = osc_env_io(env);
        struct osc_extent     *ext = NULL;
        struct osc_async_page *oap = &ops->ops_oap;
        struct client_obd     *cli = oap->oap_cli;
        struct osc_object     *osc = oap->oap_obj;
+       struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
        pgoff_t index;
        unsigned int tmp;
        unsigned int grants = 0;
@@ -2440,7 +2435,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
                /* it doesn't need any grant to dirty this page */
                spin_lock(&cli->cl_loi_list_lock);
-               rc = osc_enter_cache_try(cli, oap, grants, 0);
+               rc = osc_enter_cache_try(cli, oap, grants);
                if (rc == 0) { /* try failed */
                        grants = 0;
                        need_release = 1;
@@ -2481,7 +2476,14 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
                rc = 0;
                if (grants == 0) {
-                       /* we haven't allocated grant for this page. */
+                       /* We haven't allocated grant for this page, and we
+                        * must not hold a page lock while we do enter_cache,
+                        * so we must mark dirty & unlock any pages in the
+                        * write commit pagevec. */
+                       if (pagevec_count(pvec)) {
+                               cb(env, io, pvec);
+                               pagevec_reinit(pvec);
+                       }
                        rc = osc_enter_cache(env, cli, oap, tmp);
                        if (rc == 0)
                                grants = tmp;
@@ -2756,6 +2758,7 @@ int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
        ext->oe_obj = obj;
        ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
        ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
+       ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
        ext->oe_nr_pages = page_count;
        ext->oe_mppr = mppr;
        list_splice_init(list, &ext->oe_pages);
@@ -2787,7 +2790,7 @@ int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
        struct osc_extent *ext;
        struct osc_extent *waiting = NULL;
        pgoff_t index;
-       struct list_head list = LIST_HEAD_INIT(list);
+       LIST_HEAD(list);
        int result = 0;
        bool partial;
        ENTRY;
@@ -3005,7 +3008,7 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
                              pgoff_t start, pgoff_t end, int hp, int discard)
 {
        struct osc_extent *ext;
-       struct list_head discard_list = LIST_HEAD_INIT(discard_list);
+       LIST_HEAD(discard_list);
        bool unplug = false;
        int result = 0;
        ENTRY;