client_obd_list_unlock(&cli->cl_loi_list_lock);
CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
+
+}
+
+static void osc_update_next_shrink(struct client_obd *cli)
+{
+ int time = GRANT_SHRINK_INTERVAL;
+ cli->cl_next_shrink_grant = cfs_time_shift(time);
+ CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
+ cli->cl_next_shrink_grant);
}
/* caller must hold loi_list_lock */
CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
CFS_PAGE_SIZE, pga, pga->pg);
LASSERT(cli->cl_avail_grant >= 0);
+ osc_update_next_shrink(cli);
}
/* the companion to osc_consume_write_grant, called when a brw has completed.
EXIT;
}
-static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
+static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
{
client_obd_list_lock(&cli->cl_loi_list_lock);
- cli->cl_avail_grant = ocd->ocd_grant;
+ CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
+ if (body->oa.o_valid & OBD_MD_FLGRANT)
+ cli->cl_avail_grant += body->oa.o_grant;
+ /* waiters are woken in brw_interpret */
client_obd_list_unlock(&cli->cl_loi_list_lock);
+}
- CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld\n",
- cli->cl_avail_grant, cli->cl_lost_grant);
- LASSERT(cli->cl_avail_grant >= 0);
+static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
+ void *key, obd_count vallen, void *val,
+ struct ptlrpc_request_set *set);
+
+static int osc_shrink_grant_interpret(const struct lu_env *env,
+ struct ptlrpc_request *req,
+ void *aa, int rc)
+{
+ struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
+ struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
+ struct ost_body *body;
+
+ if (rc != 0) {
+ client_obd_list_lock(&cli->cl_loi_list_lock);
+ cli->cl_avail_grant += oa->o_grant;
+ client_obd_list_unlock(&cli->cl_loi_list_lock);
+ GOTO(out, rc);
+ }
+
+ body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ LASSERT(body);
+ osc_update_grant(cli, body);
+out:
+ OBD_FREE_PTR(oa);
+ return rc;
}
-static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
+static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
{
client_obd_list_lock(&cli->cl_loi_list_lock);
- CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
- if (body->oa.o_valid & OBD_MD_FLGRANT)
+ oa->o_grant = cli->cl_avail_grant / 4;
+ cli->cl_avail_grant -= oa->o_grant;
+ client_obd_list_unlock(&cli->cl_loi_list_lock);
+ oa->o_flags |= OBD_FL_SHRINK_GRANT;
+ osc_update_next_shrink(cli);
+}
+
+static int osc_shrink_grant(struct client_obd *cli)
+{
+ int rc = 0;
+ struct ost_body *body;
+ ENTRY;
+
+ OBD_ALLOC_PTR(body);
+ if (!body)
+ RETURN(-ENOMEM);
+
+ osc_announce_cached(cli, &body->oa, 0);
+ osc_shrink_grant_local(cli, &body->oa);
+ rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
+ sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
+ sizeof(*body), body, NULL);
+ if (rc) {
+ client_obd_list_lock(&cli->cl_loi_list_lock);
cli->cl_avail_grant += body->oa.o_grant;
- /* waiters are woken in brw_interpret */
+ client_obd_list_unlock(&cli->cl_loi_list_lock);
+ }
+ if (body)
+ OBD_FREE_PTR(body);
+ RETURN(rc);
+}
+
+#define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
+static int osc_should_shrink_grant(struct client_obd *client)
+{
+ cfs_time_t time = cfs_time_current();
+ cfs_time_t next_shrink = client->cl_next_shrink_grant;
+ if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
+ if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
+ client->cl_avail_grant > GRANT_SHRINK_LIMIT)
+ return 1;
+ else
+ osc_update_next_shrink(client);
+ }
+ return 0;
+}
+
+static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
+{
+ struct client_obd *client;
+
+ list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
+ if (osc_should_shrink_grant(client))
+ osc_shrink_grant(client);
+ }
+ return 0;
+}
+
+static int osc_add_shrink_grant(struct client_obd *client)
+{
+ int rc;
+
+ rc = ptlrpc_add_timeout_client(GRANT_SHRINK_INTERVAL,
+ TIMEOUT_GRANT,
+ osc_grant_shrink_grant_cb, NULL,
+ &client->cl_grant_shrink_list);
+ if (rc) {
+ CERROR("add grant client %s error %d\n",
+ client->cl_import->imp_obd->obd_name, rc);
+ return rc;
+ }
+ CDEBUG(D_CACHE, "add grant client %s \n",
+ client->cl_import->imp_obd->obd_name);
+ osc_update_next_shrink(client);
+ return 0;
+}
+
+static int osc_del_shrink_grant(struct client_obd *client)
+{
+ return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
+ TIMEOUT_GRANT);
+}
+
+static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
+{
+ client_obd_list_lock(&cli->cl_loi_list_lock);
+ cli->cl_avail_grant = ocd->ocd_grant;
client_obd_list_unlock(&cli->cl_loi_list_lock);
+
+ if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
+ list_empty(&cli->cl_grant_shrink_list))
+ osc_add_shrink_grant(cli);
+
+ CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
+ cli->cl_avail_grant, cli->cl_lost_grant);
+ LASSERT(cli->cl_avail_grant >= 0);
}
/* We assume that the reason this OSC got a short read is because it read
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);
+ unsigned mask = ~(OBD_BRW_FROM_GRANT|
+ OBD_BRW_NOCACHE|OBD_BRW_SYNC);
/* warn if we try to combine flags that we don't know to be
* safe to combine */
(void *)(niobuf - niocount));
osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
+ if (osc_should_shrink_grant(cli))
+ osc_shrink_grant_local(cli, &body->oa);
/* size[REQ_REC_OFF] still sizeof (*body) */
if (opc == OST_WRITE) {
struct obdo *saved_oa = NULL;
struct brw_page **ppga, **orig;
struct obd_import *imp = class_exp2cliimp(exp);
- struct client_obd *cli = &imp->imp_obd->u.cli;
+ struct client_obd *cli;
int rc, page_count_orig;
ENTRY;
+ LASSERT((imp != NULL) && (imp->imp_obd != NULL));
+ cli = &imp->imp_obd->u.cli;
+
if (cmd & OBD_BRW_CHECK) {
/* The caller just wants to know if there's a chance that this
* I/O can succeed */
- if (imp == NULL || imp->imp_invalid)
+ if (imp->imp_invalid)
RETURN(-EIO);
RETURN(0);
}
RETURN(0);
}
+static int lop_makes_hprpc(struct loi_oap_pages *lop)
+{
+ struct osc_async_page *oap;
+ ENTRY;
+
+ if (list_empty(&lop->lop_urgent))
+ RETURN(0);
+
+ oap = list_entry(lop->lop_urgent.next,
+ struct osc_async_page, oap_urgent_item);
+
+ if (oap->oap_async_flags & ASYNC_HP) {
+ CDEBUG(D_CACHE, "hp request forcing RPC\n");
+ RETURN(1);
+ }
+
+ RETURN(0);
+}
+
static void on_list(struct list_head *item, struct list_head *list,
int should_be_on)
{
* can find pages to build into rpcs quickly */
void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
{
- on_list(&loi->loi_cli_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_hprpc(&loi->loi_write_lop) ||
+ lop_makes_hprpc(&loi->loi_read_lop)) {
+ /* HP rpc */
+ 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));
+ }
on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
loi->loi_write_lop.lop_num_pending);
else
lop = &oap->oap_loi->loi_read_lop;
- if (oap->oap_async_flags & ASYNC_URGENT)
+ if (oap->oap_async_flags & ASYNC_HP)
list_add(&oap->oap_urgent_item, &lop->lop_urgent);
+ else if (oap->oap_async_flags & ASYNC_URGENT)
+ list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
}
struct cl_object *clob = NULL;
ENTRY;
+ /* If there are HP OAPs we need to handle at least 1 of them,
+ * move it the beginning of the pending list for that. */
+ if (!list_empty(&lop->lop_urgent)) {
+ oap = list_entry(lop->lop_urgent.next,
+ struct osc_async_page, oap_urgent_item);
+ if (oap->oap_async_flags & ASYNC_HP)
+ list_move(&oap->oap_pending_item, &lop->lop_pending);
+ }
+
/* first we find the pages we're allowed to work with */
list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
oap_pending_item) {
ops = oap->oap_caller_ops;
- LASSERT(oap->oap_magic == OAP_MAGIC);
+ LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
+ "magic 0x%x\n", oap, oap->oap_magic);
if (clob == NULL) {
/* pin object in memory, so that completion call-backs
#define LOI_DEBUG(LOI, STR, args...) \
CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
- !list_empty(&(LOI)->loi_cli_item), \
+ !list_empty(&(LOI)->loi_ready_item) || \
+ !list_empty(&(LOI)->loi_hp_ready_item), \
(LOI)->loi_write_lop.lop_num_pending, \
!list_empty(&(LOI)->loi_write_lop.lop_urgent), \
(LOI)->loi_read_lop.lop_num_pending, \
struct lov_oinfo *osc_next_loi(struct client_obd *cli)
{
ENTRY;
- /* first return all objects which we already know to have
- * pages ready to be stuffed into rpcs */
+
+ /* First return objects that have blocked locks so that they
+ * will be flushed quickly and other clients can get the lock,
+ * then objects which have pages ready to be stuffed into RPCs */
+ if (!list_empty(&cli->cl_loi_hp_ready_list))
+ RETURN(list_entry(cli->cl_loi_hp_ready_list.next,
+ struct lov_oinfo, loi_hp_ready_item));
if (!list_empty(&cli->cl_loi_ready_list))
RETURN(list_entry(cli->cl_loi_ready_list.next,
- struct lov_oinfo, loi_cli_item));
+ struct lov_oinfo, loi_ready_item));
/* then if we have cache waiters, return all objects with queued
* writes. This is especially important when many small files
RETURN(NULL);
}
+static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
+{
+ struct osc_async_page *oap;
+ int hprpc = 0;
+
+ if (!list_empty(&loi->loi_write_lop.lop_urgent)) {
+ oap = list_entry(loi->loi_write_lop.lop_urgent.next,
+ struct osc_async_page, oap_urgent_item);
+ hprpc = !!(oap->oap_async_flags & ASYNC_HP);
+ }
+
+ if (!hprpc && !list_empty(&loi->loi_read_lop.lop_urgent)) {
+ oap = list_entry(loi->loi_write_lop.lop_urgent.next,
+ struct osc_async_page, oap_urgent_item);
+ hprpc = !!(oap->oap_async_flags & ASYNC_HP);
+ }
+
+ return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
+}
+
/* called with the loi list lock held */
void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
{
while ((loi = osc_next_loi(cli)) != NULL) {
LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
- if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
+ if (osc_max_rpc_in_flight(cli, loi))
break;
/* attempt some read/write balancing by alternating between
/* attempt some inter-object balancing by issueing rpcs
* for each object in turn */
- if (!list_empty(&loi->loi_cli_item))
- list_del_init(&loi->loi_cli_item);
+ if (!list_empty(&loi->loi_hp_ready_item))
+ list_del_init(&loi->loi_hp_ready_item);
+ if (!list_empty(&loi->loi_ready_item))
+ list_del_init(&loi->loi_ready_item);
if (!list_empty(&loi->loi_write_item))
list_del_init(&loi->loi_write_item);
if (!list_empty(&loi->loi_read_item))
if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
oap->oap_async_flags |= ASYNC_READY;
- if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
- if (list_empty(&oap->oap_rpc_item)) {
+ if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
+ list_empty(&oap->oap_rpc_item)) {
+ if (oap->oap_async_flags & ASYNC_HP)
list_add(&oap->oap_urgent_item, &lop->lop_urgent);
- loi_list_maint(cli, loi);
- }
+ else
+ list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
+ oap->oap_async_flags |= ASYNC_URGENT;
+ loi_list_maint(cli, loi);
}
LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
if (!list_empty(&oap->oap_urgent_item)) {
list_del_init(&oap->oap_urgent_item);
- oap->oap_async_flags &= ~ASYNC_URGENT;
+ oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
}
if (!list_empty(&oap->oap_pending_item)) {
list_del_init(&oap->oap_pending_item);
/* Complete osc stuff. */
rc = osc_enqueue_fini(req, aa->oa_lvb,
aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
+
+ OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
+
/* Release the lock for async request. */
if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
/*
RETURN(0);
}
- if (!set)
+ if (!set && !KEY_IS(KEY_GRANT_SHRINK))
RETURN(-EINVAL);
/* We pass all other commands directly to OST. Since nobody calls osc
Even if something bad goes through, we'd get a -EINVAL from OST
anyway. */
-
- req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
- if (req == NULL)
+ if (KEY_IS(KEY_GRANT_SHRINK))
+ req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
+ else
+ req = ptlrpc_request_alloc(imp, &RQF_OST_SET_INFO);
+
+ if (req == NULL)
RETURN(-ENOMEM);
req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
LASSERT_MDS_GROUP(oscc->oscc_oa.o_gr);
req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
- }
+ } else if (KEY_IS(KEY_GRANT_SHRINK)) {
+ struct osc_grant_args *aa;
+ struct obdo *oa;
- ptlrpc_request_set_replen(req);
- ptlrpc_set_add_req(set, req);
- ptlrpc_check_set(NULL, set);
-
- RETURN(0);
+ CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
+ aa = ptlrpc_req_async_args(req);
+ OBD_ALLOC_PTR(oa);
+ if (!oa) {
+ ptlrpc_req_finished(req);
+ RETURN(-ENOMEM);
+ }
+ *oa = ((struct ost_body *)val)->oa;
+ aa->aa_oa = oa;
+ req->rq_interpret_reply = osc_shrink_grant_interpret;
+ }
+
+ ptlrpc_request_set_replen(req);
+ if (!KEY_IS(KEY_GRANT_SHRINK)) {
+ LASSERT(set != NULL);
+ ptlrpc_set_add_req(set, req);
+ ptlrpc_check_set(NULL, set);
+ } else
+ ptlrpcd_add_req(req, PSCOPE_OTHER);
+
+ RETURN(0);
}
}
rc = client_disconnect_export(exp);
+ /**
+ * Initially we put del_shrink_grant before disconnect_export, but it
+ * causes the following problem if setup (connect) and cleanup
+ * (disconnect) are tangled together.
+ * connect p1 disconnect p2
+ * ptlrpc_connect_import
+ * ............... class_manual_cleanup
+ * osc_disconnect
+ * del_shrink_grant
+ * ptlrpc_connect_interrupt
+ * init_grant_shrink
+ * add this client to shrink list
+ * cleanup_osc
+ * Bang! pinger trigger the shrink.
+ * So the osc should be disconnected from the shrink list, after we
+ * are sure the import has been destroyed. BUG18662
+ */
+ if (obd->u.cli.cl_import == NULL)
+ osc_del_shrink_grant(&obd->u.cli);
return rc;
}
ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
OST_MAXREQSIZE,
ptlrpc_add_rqs_to_pool);
+
+ CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
+ sema_init(&cli->cl_grant_sem, 1);
}
RETURN(rc);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff