return cli->cl_r_in_flight + cli->cl_w_in_flight;
}
+int osc_wake_sync_fs(struct client_obd *cli)
+{
+ int rc = 0;
+ ENTRY;
+ if (cfs_list_empty(&cli->cl_loi_sync_fs_list) &&
+ cli->cl_sf_wait.started) {
+ cli->cl_sf_wait.sfw_upcall(cli->cl_sf_wait.sfw_oi, rc);
+ cli->cl_sf_wait.started = 0;
+ CDEBUG(D_CACHE, "sync_fs_loi list is empty\n");
+ }
+ RETURN(rc);
+}
+
/* caller must hold loi_list_lock */
void osc_wake_cache_waiters(struct client_obd *cli)
{
LASSERT(body);
osc_update_grant(cli, body);
out:
- OBD_FREE_PTR(oa);
+ OBDO_FREE(oa);
return rc;
}
{
cfs_time_t time = cfs_time_current();
cfs_time_t next_shrink = client->cl_next_shrink_grant;
+
+ if ((client->cl_import->imp_connect_data.ocd_connect_flags &
+ OBD_CONNECT_GRANT_SHRINK) == 0)
+ return 0;
+
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)
cli->cl_avail_grant = ocd->ocd_grant;
else
cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
+
+ if (cli->cl_avail_grant < 0) {
+ CWARN("%s: available grant < 0, the OSS is probably not running"
+ " with patch from bug20278 (%ld) \n",
+ cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
+ /* workaround for 1.6 servers which do not have
+ * the patch from bug20278 */
+ cli->cl_avail_grant = ocd->ocd_grant;
+ }
+
client_obd_list_unlock(&cli->cl_loi_list_lock);
- CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
+ CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
+ cli->cl_import->imp_obd->obd_name,
cli->cl_avail_grant, cli->cl_lost_grant);
- LASSERT(cli->cl_avail_grant >= 0);
if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
cfs_list_empty(&cli->cl_grant_shrink_list))
{
if (p1->flag != p2->flag) {
unsigned mask = ~(OBD_BRW_FROM_GRANT|
- OBD_BRW_NOCACHE|OBD_BRW_SYNC);
+ OBD_BRW_NOCACHE|OBD_BRW_SYNC|OBD_BRW_ASYNC);
/* warn if we try to combine flags that we don't know to be
* safe to combine */
return 0;
}
+ /* If this is mmaped file - it can be changed at any time */
+ if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
+ return 1;
+
if (oa->o_valid & OBD_MD_FLFLAGS)
cksum_type = cksum_type_unpack(oa->o_flags);
else
}
#endif
+ osc_update_grant(cli, body);
+
if (rc < 0)
RETURN(rc);
if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
client_cksum = aa->aa_oa->o_cksum; /* save for later */
- osc_update_grant(cli, body);
-
if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
if (rc > 0) {
CERROR("Unexpected +ve rc %d\n", rc);
ENTRY;
if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
- CERROR("too many resend retries, returning error\n");
+ CERROR("too many resent retries, returning error\n");
RETURN(-EIO);
}
osc_release_write_grant(cli, &oap->oap_brw_page, sent);
}
+static int lop_makes_syncfs_rpc(struct loi_oap_pages *lop)
+{
+ struct osc_async_page *oap;
+ ENTRY;
+
+ if (cfs_list_empty(&lop->lop_urgent))
+ RETURN(0);
+
+ oap = cfs_list_entry(lop->lop_urgent.next,
+ struct osc_async_page, oap_urgent_item);
+
+ if (oap->oap_async_flags & ASYNC_SYNCFS) {
+ CDEBUG(D_CACHE, "syncfs request forcing RPC\n");
+ RETURN(1);
+ }
+
+ RETURN(0);
+}
/* This maintains the lists of pending pages to read/write for a given object
* (lop). This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
if (cmd & OBD_BRW_WRITE) {
/* trigger a write rpc stream as long as there are dirtiers
* waiting for space. as they're waiting, they're not going to
- * create more pages to coallesce with what's waiting.. */
+ * create more pages to coalesce with what's waiting.. */
if (!cfs_list_empty(&cli->cl_cache_waiters)) {
CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
RETURN(1);
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));
+ if (lop_makes_syncfs_rpc(&loi->loi_write_lop)) {
+ on_list(&loi->loi_sync_fs_item,
+ &cli->cl_loi_sync_fs_list,
+ loi->loi_write_lop.lop_num_pending);
+ } 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,
rc = osc_brw_fini_request(req, rc);
CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
if (osc_recoverable_error(rc)) {
- rc = osc_brw_redo_request(req, aa);
- if (rc == 0)
- RETURN(0);
+ /* Only retry once for mmaped files since the mmaped page
+ * might be modified at anytime. We have to retry at least
+ * once in case there WAS really a corruption of the page
+ * on the network, that was not caused by mmap() modifying
+ * the page. Bug11742 */
+ if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
+ aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
+ aa->aa_oa->o_flags & OBD_FL_MMAP) {
+ rc = 0;
+ } else {
+ rc = osc_brw_redo_request(req, aa);
+ if (rc == 0)
+ RETURN(0);
+ }
}
if (aa->aa_ocapa) {
}
OBDO_FREE(aa->aa_oa);
} else { /* from async_internal() */
- int i;
+ obd_count i;
for (i = 0; i < aa->aa_page_count; i++)
osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
-
- if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
- OBDO_FREE(aa->aa_oa);
}
osc_wake_cache_waiters(cli);
+ osc_wake_sync_fs(cli);
osc_check_rpcs(env, cli);
client_obd_list_unlock(&cli->cl_loi_list_lock);
if (!async)
cl_req_completion(env, aa->aa_clerq, rc);
osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
+
RETURN(rc);
}
enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
struct ldlm_lock *lock = NULL;
struct cl_req_attr crattr;
- int i, rc;
+ int i, rc, mpflag = 0;
ENTRY;
LASSERT(!cfs_list_empty(rpc_list));
+ if (cmd & OBD_BRW_MEMALLOC)
+ mpflag = cfs_memory_pressure_get_and_set();
+
memset(&crattr, 0, sizeof crattr);
OBD_ALLOC(pga, sizeof(*pga) * page_count);
if (pga == NULL)
GOTO(out, req = ERR_PTR(rc));
}
+ if (cmd & OBD_BRW_MEMALLOC)
+ req->rq_memalloc = 1;
+
/* Need to update the timestamps after the request is built in case
* we race with setattr (locally or in queue at OST). If OST gets
* later setattr before earlier BRW (as determined by the request xid),
CFS_INIT_LIST_HEAD(rpc_list);
aa->aa_clerq = clerq;
out:
+ if (cmd & OBD_BRW_MEMALLOC)
+ cfs_memory_pressure_restore(mpflag);
+
capa_put(crattr.cra_capa);
if (IS_ERR(req)) {
if (oa)
* \param cmd OBD_BRW_* macroses
* \param lop pending pages
*
- * \return zero if pages successfully add to send queue.
- * \return not zere if error occurring.
+ * \return zero if no page added to send queue.
+ * \return 1 if pages successfully added to send queue.
+ * \return negative on errors.
*/
static int
osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
CFS_LIST_HEAD(tmp_list);
unsigned int ending_offset;
unsigned starting_offset = 0;
- int srvlock = 0;
+ int srvlock = 0, mem_tight = 0;
struct cl_object *clob = NULL;
ENTRY;
* to be canceled, the pages covered by the lock will be sent out
* with ASYNC_HP. We have to send out them as soon as possible. */
cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
- if (oap->oap_async_flags & ASYNC_HP)
+ if (oap->oap_async_flags & ASYNC_HP)
cfs_list_move(&oap->oap_pending_item, &tmp_list);
else
cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
* until completion unlocks it. commit_write submits a page
* as not ready because its unlock will happen unconditionally
* as the call returns. if we race with commit_write giving
- * us that page we dont' want to create a hole in the page
+ * us that page we don't want to create a hole in the page
* stream, so we stop and leave the rpc to be fired by
* another dirtier or kupdated interval (the not ready page
* will still be on the dirty list). we could call in
/* now put the page back in our accounting */
cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
+ if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
+ mem_tight = 1;
if (page_count == 0)
srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
if (++page_count >= cli->cl_max_pages_per_rpc)
}
osc_wake_cache_waiters(cli);
-
+ osc_wake_sync_fs(cli);
loi_list_maint(cli, loi);
client_obd_list_unlock(&cli->cl_loi_list_lock);
RETURN(0);
}
- req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
+ req = osc_build_req(env, cli, &rpc_list, page_count,
+ mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
if (IS_ERR(req)) {
LASSERT(cfs_list_empty(&rpc_list));
loi_list_maint(cli, loi);
if (!cfs_list_empty(&cli->cl_loi_ready_list))
RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
struct lov_oinfo, loi_ready_item));
+ if (!cfs_list_empty(&cli->cl_loi_sync_fs_list))
+ RETURN(cfs_list_entry(cli->cl_loi_sync_fs_list.next,
+ struct lov_oinfo, loi_sync_fs_item));
/* then if we have cache waiters, return all objects with queued
* writes. This is especially important when many small files
race_counter++;
}
- /* attempt some inter-object balancing by issueing rpcs
+ /* attempt some inter-object balancing by issuing rpcs
* for each object in turn */
if (!cfs_list_empty(&loi->loi_hp_ready_item))
cfs_list_del_init(&loi->loi_hp_ready_item);
cfs_list_del_init(&loi->loi_write_item);
if (!cfs_list_empty(&loi->loi_read_item))
cfs_list_del_init(&loi->loi_read_item);
+ if (!cfs_list_empty(&loi->loi_sync_fs_item))
+ cfs_list_del_init(&loi->loi_sync_fs_item);
loi_list_maint(cli, loi);
oap->oap_count = count;
oap->oap_brw_flags = brw_flags;
/* Give a hint to OST that requests are coming from kswapd - bug19529 */
- if (libcfs_memory_pressure_get())
+ if (cfs_memory_pressure_get())
oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
cfs_spin_lock(&oap->oap_lock);
oap->oap_async_flags = async_flags;
if ((oap->oap_async_flags & async_flags) == async_flags)
RETURN(0);
+ /* XXX: This introduces a tiny insignificant race for the case if this
+ * loi already had other urgent items.
+ */
+ if (SETTING(oap->oap_async_flags, async_flags, ASYNC_SYNCFS) &&
+ cfs_list_empty(&oap->oap_rpc_item) &&
+ cfs_list_empty(&oap->oap_urgent_item)) {
+ cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
+ flags |= ASYNC_SYNCFS;
+ cfs_spin_lock(&oap->oap_lock);
+ oap->oap_async_flags |= flags;
+ cfs_spin_unlock(&oap->oap_lock);
+ loi_list_maint(cli, loi);
+ RETURN(0);
+ }
+
if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
flags |= ASYNC_READY;
if (!cfs_list_empty(&oap->oap_urgent_item)) {
cfs_list_del_init(&oap->oap_urgent_item);
cfs_spin_lock(&oap->oap_lock);
- oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
+ oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP |
+ ASYNC_SYNCFS);
cfs_spin_unlock(&oap->oap_lock);
}
if (!cfs_list_empty(&oap->oap_pending_item)) {
* osc_enqueue_fini(). */
ldlm_lock_addref(&handle, mode);
+ /* Let CP AST to grant the lock first. */
+ OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
+
/* Complete obtaining the lock procedure. */
rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
mode, aa->oa_flags, aa->oa_lvb,
RETURN(-ENOMEM);
rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
- if (rc)
+ if (rc) {
+ ptlrpc_request_free(req);
RETURN(rc);
+ }
req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
sizeof *lvb);
CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
aa = ptlrpc_req_async_args(req);
- OBD_ALLOC_PTR(oa);
+ OBDO_ALLOC(oa);
if (!oa) {
ptlrpc_req_finished(req);
RETURN(-ENOMEM);
RETURN(rc);
}
+/**
+ * Determine whether the lock can be canceled before replaying the lock
+ * during recovery, see bug16774 for detailed information.
+ *
+ * \retval zero the lock can't be canceled
+ * \retval other ok to cancel
+ */
+static int osc_cancel_for_recovery(struct ldlm_lock *lock)
+{
+ check_res_locked(lock->l_resource);
+
+ /*
+ * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
+ *
+ * XXX as a future improvement, we can also cancel unused write lock
+ * if it doesn't have dirty data and active mmaps.
+ */
+ if (lock->l_resource->lr_type == LDLM_EXTENT &&
+ (lock->l_granted_mode == LCK_PR ||
+ lock->l_granted_mode == LCK_CR) &&
+ (osc_dlm_lock_pageref(lock) == 0))
+ RETURN(1);
+
+ RETURN(0);
+}
+
int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
{
int rc;
CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
cfs_sema_init(&cli->cl_grant_sem, 1);
+
+ ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
}
RETURN(rc);
return(rc);
}
+static int osc_sync_fs(struct obd_device *obd, struct obd_info *oinfo,
+ int wait)
+{
+ struct client_obd *cli;
+ struct lov_oinfo *loi;
+ struct lov_oinfo *tloi;
+ struct osc_async_page *oap;
+ struct osc_async_page *toap;
+ struct loi_oap_pages *lop;
+ struct lu_env *env;
+ int refcheck;
+ int rc = 0;
+ ENTRY;
+
+ env = cl_env_get(&refcheck);
+ if (IS_ERR(env))
+ RETURN(PTR_ERR(env));
+
+ cli = &obd->u.cli;
+ client_obd_list_lock(&cli->cl_loi_list_lock);
+ cli->cl_sf_wait.sfw_oi = oinfo;
+ cli->cl_sf_wait.sfw_upcall = oinfo->oi_cb_up;
+ cli->cl_sf_wait.started = 1;
+ /* creating cl_loi_sync_fs list */
+ cfs_list_for_each_entry_safe(loi, tloi, &cli->cl_loi_write_list,
+ loi_write_item) {
+ lop = &loi->loi_write_lop;
+ cfs_list_for_each_entry_safe(oap, toap, &lop->lop_pending,
+ oap_pending_item)
+ osc_set_async_flags_base(cli, loi, oap, ASYNC_SYNCFS);
+ }
+
+ osc_check_rpcs(env, cli);
+ osc_wake_sync_fs(cli);
+ client_obd_list_unlock(&cli->cl_loi_list_lock);
+ cl_env_put(env, &refcheck);
+ RETURN(rc);
+}
+
static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
{
return osc_process_config_base(obd, buf);
.o_llog_init = osc_llog_init,
.o_llog_finish = osc_llog_finish,
.o_process_config = osc_process_config,
+ .o_sync_fs = osc_sync_fs,
};
extern struct lu_kmem_descr osc_caches[];
git://git.whamcloud.com / fs / lustre-release.git / blobdiff