Whamcloud - gitweb
LU-9679 modules: Use LIST_HEAD for declaring list_heads
[fs/lustre-release.git] / lustre / osc / osc_cache.c
index 67187c5..891d560 100644 (file)
@@ -220,6 +220,7 @@ static int osc_extent_sanity_check0(struct osc_extent *ext,
                        GOTO(out, rc = 60);
                if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
                        GOTO(out, rc = 65);
+               /* fallthrough */
        default:
                if (atomic_read(&ext->oe_users) > 0)
                        GOTO(out, rc = 70);
@@ -584,6 +585,7 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
                         * osc_cache_truncate_start(). */
                        osc_extent_state_set(ext, OES_TRUNC);
                        ext->oe_trunc_pending = 0;
+                       osc_object_unlock(obj);
                } else {
                        int grant = 0;
 
@@ -596,8 +598,6 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
                                grant += cli->cl_grant_extent_tax;
                        if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
                                grant += cli->cl_grant_extent_tax;
-                       if (grant > 0)
-                               osc_unreserve_grant(cli, 0, grant);
 
                        if (ext->oe_urgent)
                                list_move_tail(&ext->oe_link,
@@ -606,8 +606,10 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
                                list_move_tail(&ext->oe_link,
                                               &obj->oo_full_exts);
                        }
+                       osc_object_unlock(obj);
+                       if (grant > 0)
+                               osc_unreserve_grant(cli, 0, grant);
                }
-               osc_object_unlock(obj);
 
                osc_io_unplug_async(env, cli, obj);
        }
@@ -959,9 +961,8 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *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));
+               rc = 0;
        }
        if (rc == 0 && ext->oe_rc < 0)
                rc = ext->oe_rc;
@@ -1403,7 +1404,6 @@ static void osc_consume_write_grant(struct client_obd *cli,
 {
        assert_spin_locked(&cli->cl_loi_list_lock);
        LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
-       atomic_long_inc(&obd_dirty_pages);
        cli->cl_dirty_pages++;
        pga->flag |= OBD_BRW_FROM_GRANT;
        CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
@@ -1471,13 +1471,20 @@ static void __osc_unreserve_grant(struct client_obd *cli,
        }
 }
 
-static void osc_unreserve_grant(struct client_obd *cli,
-                               unsigned int reserved, unsigned int unused)
+static void osc_unreserve_grant_nolock(struct client_obd *cli,
+                                      unsigned int reserved,
+                                      unsigned int unused)
 {
-       spin_lock(&cli->cl_loi_list_lock);
        __osc_unreserve_grant(cli, reserved, unused);
        if (unused > 0)
                osc_wake_cache_waiters(cli);
+}
+
+static void osc_unreserve_grant(struct client_obd *cli,
+                               unsigned int reserved, unsigned int unused)
+{
+       spin_lock(&cli->cl_loi_list_lock);
+       osc_unreserve_grant_nolock(cli, reserved, unused);
        spin_unlock(&cli->cl_loi_list_lock);
 }
 
@@ -1547,19 +1554,23 @@ static int osc_enter_cache_try(struct client_obd *cli,
        if (rc < 0)
                return 0;
 
-       if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
-           1 + atomic_long_read(&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;
-       } else {
-               __osc_unreserve_grant(cli, bytes, bytes);
-               rc = 0;
+       if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
+               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
+                       atomic_long_dec(&obd_dirty_pages);
        }
+       __osc_unreserve_grant(cli, bytes, bytes);
+
+out:
        return rc;
 }
 
@@ -2074,7 +2085,7 @@ 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;
@@ -2150,7 +2161,7 @@ __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,
@@ -2333,9 +2344,6 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
        oap->oap_obj_off = offset;
        LASSERT(!(offset & ~PAGE_MASK));
 
-       if (cfs_capable(CFS_CAP_SYS_RESOURCE))
-               oap->oap_brw_flags = OBD_BRW_NOQUOTA;
-
        INIT_LIST_HEAD(&oap->oap_pending_item);
        INIT_LIST_HEAD(&oap->oap_rpc_item);
 
@@ -2347,13 +2355,14 @@ 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;
@@ -2375,7 +2384,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
        /* Set the OBD_BRW_SRVLOCK before the page is queued. */
        brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
-       if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
+       if (oio->oi_cap_sys_resource) {
                brw_flags |= OBD_BRW_NOQUOTA;
                cmd |= OBD_BRW_NOQUOTA;
        }
@@ -2432,7 +2441,6 @@ 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);
-               spin_unlock(&cli->cl_loi_list_lock);
                if (rc == 0) { /* try failed */
                        grants = 0;
                        need_release = 1;
@@ -2446,10 +2454,11 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
                        } else {
                                OSC_EXTENT_DUMP(D_CACHE, ext,
                                                "expanded for %lu.\n", index);
-                               osc_unreserve_grant(cli, grants, tmp);
+                               osc_unreserve_grant_nolock(cli, grants, tmp);
                                grants = 0;
                        }
                }
+               spin_unlock(&cli->cl_loi_list_lock);
                rc = 0;
        } else if (ext != NULL) {
                /* index is located outside of active extent */
@@ -2472,7 +2481,14 @@ 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;
@@ -2778,7 +2794,7 @@ 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;
@@ -2996,7 +3012,7 @@ 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;
@@ -3029,10 +3045,25 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
                                        list_move_tail(&ext->oe_link, list);
                                unplug = true;
                        } else {
+                               struct client_obd *cli = osc_cli(obj);
+                               int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
+                               pgoff_t align_by = (1 << pcc_bits);
+                               pgoff_t a_start = round_down(start, align_by);
+                               pgoff_t a_end = round_up(end, align_by);
+
+                               /* overflow case */
+                               if (end && !a_end)
+                                       a_end = CL_PAGE_EOF;
                                /* the only discarder is lock cancelling, so
-                                * [start, end] must contain this extent */
-                               EASSERT(ext->oe_start >= start &&
-                                       ext->oe_max_end <= end, ext);
+                                * [start, end], aligned by chunk size, must
+                                * contain this extent */
+                               LASSERTF(ext->oe_start >= a_start &&
+                                        ext->oe_end <= a_end,
+                                        "ext [%lu, %lu] reg [%lu, %lu] "
+                                        "orig [%lu %lu] align %lu bits "
+                                        "%d\n", ext->oe_start, ext->oe_end,
+                                        a_start, a_end, start, end,
+                                        align_by, pcc_bits);
                                osc_extent_state_set(ext, OES_LOCKING);
                                ext->oe_owner = current;
                                list_move_tail(&ext->oe_link,