This patched covered a couple of things:
- remove global router-checker data
In current LNet we can have both global RC data and per-peer
RC data, the later is only for asymmetrical router failure so far,
however we can use per-peer RC data even asym_router_failure is
disabled, it's just increase a little memory overhead but will
simplify router checker.
Also, I'm thinking that asym_router_failure is a nice feature and
it should be enabled by default in the future.
- Code cleanup for router checker, again, it's still an intermediate
patch for LNet SMP improvements, we need many code cleanup before
we change locking mechanism of LNet.
Signed-off-by: Liang Zhen <liang@whamcloud.com>
Change-Id: I830e706ce5099f15f122b521f74b6dfcde443592
Reviewed-on: http://review.whamcloud.com/2930
Tested-by: Hudson
Reviewed-by: Doug Oucharek <doug@whamcloud.com>
Tested-by: Maloo <whamcloud.maloo@gmail.com>
Reviewed-by: Bobi Jam <bobijam@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
};
/* Router Checker states */
};
/* Router Checker states */
-#define LNET_RC_STATE_SHUTDOWN 0 /* not started */
-#define LNET_RC_STATE_RUNNING 1 /* started up OK */
-#define LNET_RC_STATE_STOPTHREAD 2 /* telling thread to stop */
-#define LNET_RC_STATE_UNLINKING 3 /* unlinking RC MD */
-#define LNET_RC_STATE_UNLINKED 4 /* RC's MD has been unlinked */
+#define LNET_RC_STATE_SHUTDOWN 0 /* not started */
+#define LNET_RC_STATE_RUNNING 1 /* started up OK */
+#define LNET_RC_STATE_STOPPING 2 /* telling thread to stop */
#ifdef __KERNEL__
cfs_semaphore_t ln_rc_signal; /* serialise startup/shutdown */
#endif
#ifdef __KERNEL__
cfs_semaphore_t ln_rc_signal; /* serialise startup/shutdown */
#endif
- int ln_rc_state; /* router checker startup/shutdown state */
- lnet_handle_eq_t ln_rc_eqh; /* router checker's event queue */
- lnet_handle_md_t ln_rc_mdh;
- cfs_list_t ln_zombie_rcd;
+ /* router checker startup/shutdown state */
+ int ln_rc_state;
+ /* router checker's event queue */
+ lnet_handle_eq_t ln_rc_eqh;
+ /* rcd still pending on net */
+ cfs_list_t ln_rcd_deathrow;
+ /* rcd ready for free */
+ cfs_list_t ln_rcd_zombie;
lnet_counters_t ln_counters;
lnet_counters_t ln_counters;
the_lnet.ln_init = 1;
LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
the_lnet.ln_init = 1;
LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
- CFS_INIT_LIST_HEAD(&the_lnet.ln_zombie_rcd);
+ CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
+ CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
#ifdef __KERNEL__
/* All LNDs apart from the LOLND are in separate modules. They
#ifdef __KERNEL__
/* All LNDs apart from the LOLND are in separate modules. They
-lnet_do_notify (lnet_peer_t *lp)
+lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
- lnet_ni_t *ni = lp->lp_ni;
int alive;
int notifylnd;
int alive;
int notifylnd;
/* Notify only in 1 thread at any time to ensure ordered notification.
* NB individual events can be missed; the only guarantee is that you
* always get the most recent news */
/* Notify only in 1 thread at any time to ensure ordered notification.
* NB individual events can be missed; the only guarantee is that you
* always get the most recent news */
- if (lp->lp_notifying) {
- LNET_UNLOCK();
if (lp->lp_rtr_refcount == 0) {
if (lp->lp_rcd != NULL) {
cfs_list_add(&lp->lp_rcd->rcd_list,
if (lp->lp_rtr_refcount == 0) {
if (lp->lp_rcd != NULL) {
cfs_list_add(&lp->lp_rcd->rcd_list,
- &the_lnet.ln_zombie_rcd);
+ &the_lnet.ln_rcd_deathrow);
lnet_peer_t *lp;
lnet_nid_t nid;
lnet_peer_t *lp;
lnet_nid_t nid;
- if (event->unlinked) {
- if (rcd != NULL) {
- LNetInvalidateHandle(&rcd->rcd_mdh);
- return;
- }
- /* The router checker thread has unlinked the default rc_md
- * and exited. */
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING);
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKED;
-#ifdef __KERNEL__
- cfs_up(&the_lnet.ln_rc_signal);
-#endif
- return;
- }
+ if (event->unlinked) {
+ LNetInvalidateHandle(&rcd->rcd_mdh);
+ return;
+ }
LASSERT (event->type == LNET_EVENT_SEND ||
event->type == LNET_EVENT_REPLY);
LASSERT (event->type == LNET_EVENT_SEND ||
event->type == LNET_EVENT_REPLY);
void
lnet_destroy_rc_data (lnet_rc_data_t *rcd)
{
void
lnet_destroy_rc_data (lnet_rc_data_t *rcd)
{
- LASSERT (cfs_list_empty(&rcd->rcd_list));
- /* detached from network */
- LASSERT (LNetHandleIsInvalid(rcd->rcd_mdh));
+ LASSERT(cfs_list_empty(&rcd->rcd_list));
+ /* detached from network */
+ LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
- LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
- LIBCFS_FREE(rcd, sizeof(*rcd));
- return;
+ if (rcd->rcd_pinginfo != NULL)
+ LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
+
+ LIBCFS_FREE(rcd, sizeof(*rcd));
-lnet_create_rc_data (void)
+lnet_create_rc_data_locked(lnet_peer_t *gateway)
- int i;
- int rc;
- lnet_ping_info_t *pi;
- lnet_rc_data_t *rcd;
+ lnet_rc_data_t *rcd = NULL;
+ lnet_ping_info_t *pi;
+ int rc;
+ int i;
- LIBCFS_ALLOC(rcd, sizeof(*rcd));
- if (rcd == NULL)
- return NULL;
- LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
- if (pi == NULL) {
- LIBCFS_FREE(rcd, sizeof(*rcd));
- return NULL;
- }
+ LIBCFS_ALLOC(rcd, sizeof(*rcd));
+ if (rcd == NULL)
+ goto out;
+
+ LNetInvalidateHandle(&rcd->rcd_mdh);
+ CFS_INIT_LIST_HEAD(&rcd->rcd_list);
+
+ LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
+ if (pi == NULL)
+ goto out;
memset(pi, 0, LNET_PINGINFO_SIZE);
for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
memset(pi, 0, LNET_PINGINFO_SIZE);
for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
}
rcd->rcd_pinginfo = pi;
pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
}
rcd->rcd_pinginfo = pi;
- LNetInvalidateHandle(&rcd->rcd_mdh);
- CFS_INIT_LIST_HEAD(&rcd->rcd_list);
LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
rc = LNetMDBind((lnet_md_t){.start = pi,
LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
rc = LNetMDBind((lnet_md_t){.start = pi,
&rcd->rcd_mdh);
if (rc < 0) {
CERROR("Can't bind MD: %d\n", rc);
&rcd->rcd_mdh);
if (rc < 0) {
CERROR("Can't bind MD: %d\n", rc);
- lnet_destroy_rc_data(rcd);
- return NULL;
- }
- LASSERT (rc == 0);
- return rcd;
+ goto out;
+ }
+ LASSERT(rc == 0);
+
+ LNET_LOCK();
+ /* router table changed or someone has created rcd for this gateway */
+ if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
+ LNET_UNLOCK();
+ goto out;
+ }
+
+ gateway->lp_rcd = rcd;
+ return rcd;
+
+ out:
+ if (rcd != NULL) {
+ if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
+ rc = LNetMDUnlink(rcd->rcd_mdh);
+ LASSERT(rc == 0);
+ }
+ lnet_destroy_rc_data(rcd);
+ }
+
+ LNET_LOCK();
+ return gateway->lp_rcd;
static void
lnet_ping_router_locked (lnet_peer_t *rtr)
{
static void
lnet_ping_router_locked (lnet_peer_t *rtr)
{
lnet_rc_data_t *rcd = NULL;
cfs_time_t now = cfs_time_current();
int secs;
lnet_rc_data_t *rcd = NULL;
cfs_time_t now = cfs_time_current();
int secs;
cfs_time_after(now, rtr->lp_ping_deadline))
lnet_notify_locked(rtr, 1, 0, now);
cfs_time_after(now, rtr->lp_ping_deadline))
lnet_notify_locked(rtr, 1, 0, now);
- if (avoid_asym_router_failure && rtr->lp_rcd == NULL)
- newrcd = 1;
-
- LNET_UNLOCK();
-
- /* Run any outstanding notifications */
- lnet_do_notify(rtr);
+ /* Run any outstanding notifications */
+ lnet_ni_notify_locked(rtr->lp_ni, rtr);
- if (newrcd)
- rcd = lnet_create_rc_data();
+ if (!lnet_isrouter(rtr) ||
+ the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
+ /* router table changed or router checker is shutting down */
+ lnet_peer_decref_locked(rtr);
+ return;
+ }
+ rcd = rtr->lp_rcd != NULL ?
+ rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
- if (!lnet_isrouter(rtr)) {
- lnet_peer_decref_locked(rtr);
- if (rcd != NULL)
- cfs_list_add(&rcd->rcd_list, &the_lnet.ln_zombie_rcd);
- return; /* router table changed! */
- }
-
- if (rcd != NULL) {
- LASSERT (rtr->lp_rcd == NULL);
- rtr->lp_rcd = rcd;
- }
+ if (rcd == NULL)
+ return;
secs = lnet_router_check_interval(rtr);
secs = lnet_router_check_interval(rtr);
rtr->lp_ping_notsent = 1;
rtr->lp_ping_timestamp = now;
rtr->lp_ping_notsent = 1;
rtr->lp_ping_timestamp = now;
- mdh = (rtr->lp_rcd == NULL) ? the_lnet.ln_rc_mdh :
- rtr->lp_rcd->rcd_mdh;
- if (rtr->lp_ping_deadline == 0)
- rtr->lp_ping_deadline = cfs_time_shift(router_ping_timeout);
+ mdh = rcd->rcd_mdh;
+
+ if (rtr->lp_ping_deadline == 0) {
+ rtr->lp_ping_deadline = \
+ cfs_time_shift(router_ping_timeout);
+ }
int
lnet_router_checker_start(void)
{
int
lnet_router_checker_start(void)
{
- static lnet_ping_info_t pinginfo;
-
- lnet_md_t md;
int rc;
int eqsz;
#ifndef __KERNEL__
int rc;
int eqsz;
#ifndef __KERNEL__
- memset(&md, 0, sizeof(md));
- md.user_ptr = NULL;
- md.start = &pinginfo;
- md.length = sizeof(pinginfo);
- md.options = LNET_MD_TRUNCATE;
- md.threshold = LNET_MD_THRESH_INF;
- md.eq_handle = the_lnet.ln_rc_eqh;
- rc = LNetMDBind(md, LNET_UNLINK, &the_lnet.ln_rc_mdh);
- if (rc < 0) {
- CERROR("Can't bind MD: %d\n", rc);
- rc = LNetEQFree(the_lnet.ln_rc_eqh);
- LASSERT (rc == 0);
- return -ENOMEM;
- }
- LASSERT (rc == 0);
-
the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
#ifdef __KERNEL__
rc = cfs_create_thread(lnet_router_checker, NULL, 0);
if (rc < 0) {
CERROR("Can't start router checker thread: %d\n", rc);
the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
#ifdef __KERNEL__
rc = cfs_create_thread(lnet_router_checker, NULL, 0);
if (rc < 0) {
CERROR("Can't start router checker thread: %d\n", rc);
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
- rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
- LASSERT (rc == 0);
/* block until event callback signals exit */
cfs_down(&the_lnet.ln_rc_signal);
rc = LNetEQFree(the_lnet.ln_rc_eqh);
/* block until event callback signals exit */
cfs_down(&the_lnet.ln_rc_signal);
rc = LNetEQFree(the_lnet.ln_rc_eqh);
return;
LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
return;
LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
- the_lnet.ln_rc_state = LNET_RC_STATE_STOPTHREAD;
+ the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
- /* block until event callback signals exit */
- cfs_down(&the_lnet.ln_rc_signal);
+ /* block until event callback signals exit */
+ cfs_down(&the_lnet.ln_rc_signal);
- while (the_lnet.ln_rc_state != LNET_RC_STATE_UNLINKED) {
- lnet_router_checker();
- cfs_pause(cfs_time_seconds(1));
- }
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED);
+ LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
rc = LNetEQFree(the_lnet.ln_rc_eqh);
LASSERT (rc == 0);
rc = LNetEQFree(the_lnet.ln_rc_eqh);
LASSERT (rc == 0);
- the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
-#if defined(__KERNEL__) && defined(LNET_ROUTER)
-
-lnet_prune_zombie_rcd (int wait_unlink)
+lnet_prune_rc_data(int wait_unlink)
- lnet_rc_data_t *rcd;
- lnet_rc_data_t *tmp;
- cfs_list_t free_rcd;
- int i;
- __u64 version;
+ lnet_rc_data_t *rcd;
+ lnet_rc_data_t *tmp;
+ lnet_peer_t *lp;
+ cfs_list_t head;
+ int i = 2;
- CFS_INIT_LIST_HEAD(&free_rcd);
+ if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
+ cfs_list_empty(&the_lnet.ln_rcd_deathrow) &&
+ cfs_list_empty(&the_lnet.ln_rcd_zombie)))
+ return;
- LNET_LOCK();
-rescan:
- version = the_lnet.ln_routers_version;
- cfs_list_for_each_entry_safe (rcd, tmp, &the_lnet.ln_zombie_rcd,
- rcd_list) {
- if (LNetHandleIsInvalid(rcd->rcd_mdh)) {
- cfs_list_del(&rcd->rcd_list);
- cfs_list_add(&rcd->rcd_list, &free_rcd);
- continue;
- }
+ CFS_INIT_LIST_HEAD(&head);
- LNetMDUnlink(rcd->rcd_mdh);
+ if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
+ /* router checker is stopping, prune all */
+ cfs_list_for_each_entry(lp, &the_lnet.ln_routers,
+ lp_rtr_list) {
+ if (lp->lp_rcd == NULL)
+ continue;
- LNET_LOCK();
- if (version != the_lnet.ln_routers_version)
- goto rescan;
+ LASSERT(cfs_list_empty(&lp->lp_rcd->rcd_list));
+ cfs_list_add(&lp->lp_rcd->rcd_list,
+ &the_lnet.ln_rcd_deathrow);
+ lp->lp_rcd = NULL;
+ }
+ }
+
+ /* unlink all RCDs on deathrow list */
+ cfs_list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
+
+ if (!cfs_list_empty(&head)) {
+ LNET_UNLOCK();
+
+ cfs_list_for_each_entry(rcd, &head, rcd_list)
+ LNetMDUnlink(rcd->rcd_mdh);
+
+ LNET_LOCK();
- i = 2;
- while (wait_unlink && !cfs_list_empty(&the_lnet.ln_zombie_rcd)) {
- rcd = cfs_list_entry(the_lnet.ln_zombie_rcd.next,
- lnet_rc_data_t, rcd_list);
- if (LNetHandleIsInvalid(rcd->rcd_mdh)) {
- cfs_list_del(&rcd->rcd_list);
- cfs_list_add(&rcd->rcd_list, &free_rcd);
- continue;
+ cfs_list_splice_init(&head, &the_lnet.ln_rcd_zombie);
+
+ /* release all zombie RCDs */
+ while (!cfs_list_empty(&the_lnet.ln_rcd_zombie)) {
+ cfs_list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
+ rcd_list) {
+ if (!LNetHandleIsInvalid(rcd->rcd_mdh))
+ cfs_list_move(&rcd->rcd_list, &head);
+ wait_unlink = wait_unlink &&
+ !cfs_list_empty(&the_lnet.ln_rcd_zombie);
- LNetMDUnlink(rcd->rcd_mdh);
- i++;
- CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
- "Waiting for rc buffers to unlink\n");
- cfs_pause(cfs_time_seconds(1));
+ while (!cfs_list_empty(&head)) {
+ rcd = cfs_list_entry(head.next,
+ lnet_rc_data_t, rcd_list);
+ cfs_list_del_init(&rcd->rcd_list);
+ lnet_destroy_rc_data(rcd);
+ }
+ if (!wait_unlink)
+ break;
+ i++;
+ CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
+ "Waiting for rc buffers to unlink\n");
+ cfs_pause(cfs_time_seconds(1) / 4);
- while (!cfs_list_empty(&free_rcd)) {
- rcd = cfs_list_entry(free_rcd.next, lnet_rc_data_t, rcd_list);
- cfs_list_del_init(&rcd->rcd_list);
- lnet_destroy_rc_data(rcd);
- }
- return;
+
+#if defined(__KERNEL__) && defined(LNET_ROUTER)
+
static int
lnet_router_checker(void *arg)
{
static int
lnet_router_checker(void *arg)
{
lnet_peer_t *rtr;
cfs_list_t *entry;
lnet_peer_t *rtr;
cfs_list_t *entry;
if (the_lnet.ln_routing)
lnet_update_ni_status();
if (the_lnet.ln_routing)
lnet_update_ni_status();
- lnet_prune_zombie_rcd(0); /* don't wait for UNLINK */
+ lnet_prune_rc_data(0); /* don't wait for UNLINK */
/* Call cfs_pause() here always adds 1 to load average
* because kernel counts # active tasks as nr_running
/* Call cfs_pause() here always adds 1 to load average
* because kernel counts # active tasks as nr_running
- LNET_LOCK();
-
- cfs_list_for_each (entry, &the_lnet.ln_routers) {
- rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
-
- if (rtr->lp_rcd == NULL)
- continue;
-
- LASSERT (cfs_list_empty(&rtr->lp_rcd->rcd_list));
- cfs_list_add(&rtr->lp_rcd->rcd_list, &the_lnet.ln_zombie_rcd);
- rtr->lp_rcd = NULL;
- }
-
- LNET_UNLOCK();
-
- lnet_prune_zombie_rcd(1); /* wait for UNLINK */
+ LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
- LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD);
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
-
- rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
- LASSERT (rc == 0);
+ lnet_prune_rc_data(1); /* wait for UNLINK */
+ the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
+ cfs_up(&the_lnet.ln_rc_signal);
/* The unlink event callback will signal final completion */
return 0;
}
/* The unlink event callback will signal final completion */
return 0;
}
lnet_notify_locked(lp, ni == NULL, alive, when);
lnet_notify_locked(lp, ni == NULL, alive, when);
- LNET_UNLOCK();
-
- lnet_do_notify(lp);
-
- LNET_LOCK();
+ lnet_ni_notify_locked(ni, lp);
lnet_peer_decref_locked(lp);
lnet_peer_decref_locked(lp);
- if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD) {
- the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
- rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
- LASSERT (rc == 0);
- }
+ if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
+ lnet_prune_rc_data(0); /* unlink all rcd and nowait */
/* consume all pending events */
while (1) {
/* consume all pending events */
while (1) {
- if (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED ||
- the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING) {
+ if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
+ lnet_prune_rc_data(1); /* release rcd */
+ the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
git://git.whamcloud.com / fs / lustre-release.git / commitdiff