#include <linux/obd_class.h>
#include "ptlrpc_internal.h"
+#define PINGER_RATE 3 /* how many pings we'll do in obd_timeout period */
+
static DECLARE_MUTEX(pinger_sem);
static struct list_head pinger_imports = LIST_HEAD_INIT(pinger_imports);
+int ptlrpc_ping(struct obd_import *imp)
+{
+ struct ptlrpc_request *req;
+ int rc = 0;
+ ENTRY;
+
+ req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 0, NULL, NULL);
+ if (req) {
+ DEBUG_REQ(D_HA, req, "pinging %s->%s",
+ imp->imp_obd->obd_uuid.uuid,
+ imp->imp_target_uuid.uuid);
+ req->rq_no_resend = req->rq_no_delay = 1;
+ req->rq_replen = lustre_msg_size(0, NULL);
+ req->rq_timeout = obd_timeout / PINGER_RATE;
+ ptlrpcd_add_req(req);
+ } else {
+ CERROR("OOM trying to ping %s->%s\n",
+ imp->imp_obd->obd_uuid.uuid,
+ imp->imp_target_uuid.uuid);
+ rc = -ENOMEM;
+ }
+
+ RETURN(rc);
+}
+
#ifdef __KERNEL__
-static struct ptlrpc_thread *pinger_thread = NULL;
+static inline int ptlrpc_next_ping(struct obd_import *imp)
+{
+ return jiffies + (obd_timeout / PINGER_RATE * HZ);
+}
+
+static inline int ptlrpc_next_reconnect(struct obd_import *imp)
+{
+ if (imp->imp_server_timeout)
+ return jiffies + (obd_timeout / 2 * HZ);
+ else
+ return jiffies + (obd_timeout * HZ);
+}
+
+static atomic_t suspend_timeouts = ATOMIC_INIT(0);
+static wait_queue_head_t suspend_timeouts_waitq;
+
+void ptlrpc_deactivate_timeouts(void)
+{
+ CDEBUG(D_HA, "deactivate timeouts\n");
+ atomic_inc(&suspend_timeouts);
+}
+
+void ptlrpc_activate_timeouts(void)
+{
+ CDEBUG(D_HA, "activate timeouts\n");
+ LASSERT(atomic_read(&suspend_timeouts) > 0);
+ if (atomic_dec_and_test(&suspend_timeouts))
+ wake_up(&suspend_timeouts_waitq);
+}
+
+int ptlrpc_check_suspend(void)
+{
+ if (atomic_read(&suspend_timeouts))
+ return 1;
+ return 0;
+}
+
+int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
+{
+ struct l_wait_info lwi;
+
+ if (atomic_read(&suspend_timeouts)) {
+ DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
+ atomic_read(&suspend_timeouts));
+ lwi = LWI_INTR(NULL, NULL);
+ l_wait_event(suspend_timeouts_waitq,
+ atomic_read(&suspend_timeouts) == 0, &lwi);
+ DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
+ return 1;
+ }
+ return 0;
+}
+
+static void ptlrpc_pinger_process_import(struct obd_import *imp,
+ unsigned long this_ping)
+{
+ unsigned long flags;
+ int force, level;
+
+ spin_lock_irqsave(&imp->imp_lock, flags);
+ level = imp->imp_state;
+ force = imp->imp_force_verify;
+ if (force)
+ imp->imp_force_verify = 0;
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
+
+ if (imp->imp_next_ping - this_ping > obd_timeout * HZ) {
+ CWARN("wrong ping time %lu (current %lu)\n",
+ imp->imp_next_ping, this_ping);
+ imp->imp_next_ping = ptlrpc_next_reconnect(imp);
+ }
+
+ if (imp->imp_next_ping > this_ping && force == 0)
+ return;
+
+ if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
+ /* wait at least a timeout before trying recovery again */
+ imp->imp_next_ping = ptlrpc_next_reconnect(imp);
+ ptlrpc_initiate_recovery(imp);
+ } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov) {
+ CDEBUG(D_HA, "not pinging %s (in recovery "
+ " or recovery disabled: %s)\n",
+ imp->imp_target_uuid.uuid,
+ ptlrpc_import_state_name(level));
+ } else if (imp->imp_pingable || force) {
+ imp->imp_next_ping = ptlrpc_next_ping(imp);
+ ptlrpc_ping(imp);
+ }
+}
static int ptlrpc_pinger_main(void *arg)
{
RECALC_SIGPENDING;
SIGNAL_MASK_UNLOCK(current, flags);
- THREAD_NAME(current->comm, "%s", data->name);
+ LASSERTF(strlen(data->name) < sizeof(current->comm),
+ "name %d > len %d\n",
+ (int)strlen(data->name), (int)sizeof(current->comm));
+ THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
unlock_kernel();
/* Record that the thread is running */
long time_to_next_ping;
struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
NULL, NULL);
- struct ptlrpc_request_set *set;
- struct ptlrpc_request *req;
struct list_head *iter;
- wait_queue_t set_wait;
- int rc;
- set = ptlrpc_prep_set();
+ time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
down(&pinger_sem);
list_for_each(iter, &pinger_imports) {
struct obd_import *imp =
list_entry(iter, struct obd_import,
imp_pinger_chain);
- int generation, level;
- unsigned long flags;
-
- if (imp->imp_next_ping <= this_ping) {
- /* Add a ping. */
- spin_lock_irqsave(&imp->imp_lock, flags);
- generation = imp->imp_generation;
- level = imp->imp_state;
- spin_unlock_irqrestore(&imp->imp_lock, flags);
-
- if (level != LUSTRE_IMP_FULL) {
- CDEBUG(D_HA,
- "not pinging %s (in recovery)\n",
- imp->imp_target_uuid.uuid);
- continue;
- }
-
- req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL,
- NULL);
- if (!req) {
- CERROR("OOM trying to ping\n");
- break;
- }
- req->rq_no_resend = 1;
- req->rq_replen = lustre_msg_size(0, NULL);
- req->rq_send_state = LUSTRE_IMP_FULL;
- req->rq_phase = RQ_PHASE_RPC;
- req->rq_import_generation = generation;
- ptlrpc_set_add_req(set, req);
- } else {
- CDEBUG(D_HA, "don't need to ping %s (%lu > "
- "%lu)\n", imp->imp_target_uuid.uuid,
- imp->imp_next_ping, this_ping);
- }
- }
- up(&pinger_sem);
- /* Might be empty, that's OK. */
- if (set->set_remaining == 0)
- CDEBUG(D_HA, "nothing to ping\n");
- list_for_each(iter, &set->set_requests) {
- struct ptlrpc_request *req =
- list_entry(iter, struct ptlrpc_request,
- rq_set_chain);
- DEBUG_REQ(D_HA, req, "pinging %s->%s",
- req->rq_import->imp_obd->obd_uuid.uuid,
- req->rq_import->imp_target_uuid.uuid);
- (void)ptl_send_rpc(req);
- }
+ ptlrpc_pinger_process_import(imp, this_ping);
- /* Have to wait on both the thread's queue and the set's. */
- init_waitqueue_entry(&set_wait, current);
- add_wait_queue(&set->set_waitq, &set_wait);
- rc = l_wait_event(thread->t_ctl_waitq,
- thread->t_flags & SVC_STOPPING ||
- ptlrpc_check_set(set),
- &lwi);
- remove_wait_queue(&set->set_waitq, &set_wait);
- CDEBUG(D_HA, "ping complete (%lu)\n", jiffies);
-
- if (thread->t_flags & SVC_STOPPING) {
- thread->t_flags &= ~SVC_STOPPING;
- list_for_each(iter, &set->set_requests) {
- req = list_entry(iter, struct ptlrpc_request,
- rq_set_chain);
- if (!req->rq_replied)
- ptlrpc_unregister_reply(req);
- }
- ptlrpc_set_destroy(set);
- EXIT;
- break;
- }
+ CDEBUG(D_OTHER, "%s: pingable %d, next_ping %lu(%lu)\n",
+ imp->imp_target_uuid.uuid,
+ imp->imp_pingable, imp->imp_next_ping, jiffies);
- /* Expire all the requests that didn't come back. */
- down(&pinger_sem);
- list_for_each(iter, &set->set_requests) {
- req = list_entry(iter, struct ptlrpc_request,
- rq_set_chain);
-
- if (req->rq_replied)
- continue;
-
- req->rq_phase = RQ_PHASE_COMPLETE;
- set->set_remaining--;
- /* If it was disconnected, don't sweat it. */
- if (list_empty(&req->rq_import->imp_pinger_chain)) {
- ptlrpc_unregister_reply(req);
- continue;
- }
-
- ptlrpc_expire_one_request(req);
+ if (imp->imp_pingable && imp->imp_next_ping &&
+ imp->imp_next_ping - jiffies < time_to_next_ping &&
+ imp->imp_next_ping > jiffies)
+ time_to_next_ping = imp->imp_next_ping - jiffies;
}
up(&pinger_sem);
- ptlrpc_set_destroy(set);
/* Wait until the next ping time, or until we're stopped. */
- time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
this_ping + (obd_timeout * HZ));
if (time_to_next_ping > 0) {
return 0;
}
+static struct ptlrpc_thread *pinger_thread = NULL;
+
int ptlrpc_start_pinger(void)
{
struct l_wait_info lwi = { 0 };
#endif
ENTRY;
+ LASSERT(obd_timeout > PINGER_RATE);
+
if (pinger_thread != NULL)
RETURN(-EALREADY);
if (pinger_thread == NULL)
RETURN(-ENOMEM);
init_waitqueue_head(&pinger_thread->t_ctl_waitq);
+ init_waitqueue_head(&suspend_timeouts_waitq);
d.name = "ll_ping";
d.thread = pinger_thread;
void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
{
down(&pinger_sem);
- imp->imp_next_ping = jiffies + (obd_timeout * HZ);
+ imp->imp_next_ping = ptlrpc_next_ping(imp);
up(&pinger_sem);
}
list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
class_import_get(imp);
-#ifdef ENABLE_PINGER
- pinger_thread->t_flags |= SVC_EVENT;
- wake_up(&pinger_thread->t_ctl_waitq);
-#endif
+ ptlrpc_pinger_wake_up();
up(&pinger_sem);
RETURN(0);
RETURN(0);
}
-#else
+void ptlrpc_pinger_wake_up()
+{
+#ifdef ENABLE_PINGER
+ pinger_thread->t_flags |= SVC_EVENT;
+ wake_up(&pinger_thread->t_ctl_waitq);
+#endif
+}
+
+#else /* !__KERNEL__ */
+
/* XXX
* the current implementation of pinger in liblustre is not optimized
*/
int pd_recursion;
unsigned long pd_this_ping;
unsigned long pd_next_ping;
- struct ptlrpc_request_set *pd_set;
+ int pd_force_check;
} pinger_args;
static int pinger_check_rpcs(void *arg)
{
unsigned long curtime = time(NULL);
- struct ptlrpc_request *req;
- struct ptlrpc_request_set *set;
struct list_head *iter;
struct pinger_data *pd = &pinger_args;
- int rc;
/* prevent recursion */
if (pd->pd_recursion++) {
CDEBUG(D_HA, "pinger: recursion! quit\n");
- LASSERT(pd->pd_set);
pd->pd_recursion--;
return 0;
}
/* have we reached ping point? */
- if (!pd->pd_set && pd->pd_next_ping > curtime) {
+ if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
pd->pd_recursion--;
return 0;
}
- /* if we have rpc_set already, continue processing it */
- if (pd->pd_set) {
- LASSERT(pd->pd_this_ping);
- set = pd->pd_set;
- goto do_check_set;
- }
+ if (pd->pd_force_check)
+ pd->pd_force_check = 0;
pd->pd_this_ping = curtime;
- pd->pd_set = ptlrpc_prep_set();
- set = pd->pd_set;
/* add rpcs into set */
down(&pinger_sem);
struct obd_import *imp =
list_entry(iter, struct obd_import,
imp_pinger_chain);
- int generation, level;
+ int level, force;
unsigned long flags;
- if (imp->imp_next_ping <= pd->pd_this_ping) {
- /* Add a ping. */
- spin_lock_irqsave(&imp->imp_lock, flags);
- generation = imp->imp_generation;
- level = imp->imp_state;
- spin_unlock_irqrestore(&imp->imp_lock, flags);
-
- if (level != LUSTRE_IMP_FULL) {
- CDEBUG(D_HA,
- "not pinging %s (in recovery)\n",
- imp->imp_target_uuid.uuid);
- continue;
- }
- req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL,
- NULL);
- if (!req) {
- CERROR("out of memory\n");
- break;
+ spin_lock_irqsave(&imp->imp_lock, flags);
+ level = imp->imp_state;
+ force = imp->imp_force_verify;
+ if (force)
+ imp->imp_force_verify = 0;
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
+
+ if (imp->imp_next_ping <= pd->pd_this_ping || force) {
+ if (level == LUSTRE_IMP_DISCON) {
+ /* wait at least a timeout before
+ trying recovery again. */
+ unsigned long timeout = obd_timeout;
+ if (imp->imp_server_timeout)
+ timeout = obd_timeout / 2;
+ imp->imp_next_ping = time(NULL) +
+ (timeout * HZ);
+ ptlrpc_initiate_recovery(imp);
+ }
+ else if (level != LUSTRE_IMP_FULL ||
+ imp->imp_obd->obd_no_recov) {
+ CDEBUG(D_HA,
+ "not pinging %s (in recovery "
+ " or recovery disabled: %s)\n",
+ imp->imp_target_uuid.uuid,
+ ptlrpc_import_state_name(level));
+ }
+ else if (imp->imp_pingable || force) {
+ ptlrpc_ping(imp);
}
- req->rq_no_resend = 1;
- req->rq_replen = lustre_msg_size(0, NULL);
- req->rq_send_state = LUSTRE_IMP_FULL;
- req->rq_phase = RQ_PHASE_RPC;
- req->rq_import_generation = generation;
- ptlrpc_set_add_req(set, req);
+
} else {
- CDEBUG(D_HA, "don't need to ping %s (%lu > "
- "%lu)\n", imp->imp_target_uuid.uuid,
- imp->imp_next_ping, pd->pd_this_ping);
+ if (imp->imp_pingable) {
+ CDEBUG(D_HA, "don't need to ping %s "
+ "(%lu > %lu)\n",
+ imp->imp_target_uuid.uuid,
+ imp->imp_next_ping, pd->pd_this_ping);
+ }
}
}
- pd->pd_this_ping = curtime;
- up(&pinger_sem);
-
- /* Might be empty, that's OK. */
- if (set->set_remaining == 0)
- CDEBUG(D_HA, "nothing to ping\n");
-
- list_for_each(iter, &set->set_requests) {
- struct ptlrpc_request *req =
- list_entry(iter, struct ptlrpc_request,
- rq_set_chain);
- DEBUG_REQ(D_HA, req, "pinging %s->%s",
- req->rq_import->imp_obd->obd_uuid.uuid,
- req->rq_import->imp_target_uuid.uuid);
- (void)ptl_send_rpc(req);
- }
-
-do_check_set:
- rc = ptlrpc_check_set(set);
- /* not finished, and we are not expired, simply return */
- if (!rc && curtime < pd->pd_this_ping + obd_timeout) {
- CDEBUG(D_HA, "not finished, but also not expired\n");
- pd->pd_recursion--;
- return 0;
- }
-
- /* Expire all the requests that didn't come back. */
- down(&pinger_sem);
- list_for_each(iter, &set->set_requests) {
- req = list_entry(iter, struct ptlrpc_request,
- rq_set_chain);
-
- if (req->rq_replied)
- continue;
-
- req->rq_phase = RQ_PHASE_COMPLETE;
- set->set_remaining--;
- /* If it was disconnected, don't sweat it. */
- if (list_empty(&req->rq_import->imp_pinger_chain)) {
- ptlrpc_unregister_reply(req);
- continue;
- }
-
- CDEBUG(D_HA, "pinger initiate expire_one_request\n");
- ptlrpc_expire_one_request(req);
- }
up(&pinger_sem);
- ptlrpc_set_destroy(set);
- pd->pd_set = NULL;
-
- pd->pd_next_ping = pd->pd_this_ping + obd_timeout;
- pd->pd_this_ping = 0; /* XXX for debug */
+ pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
CDEBUG(D_HA, "finished a round ping\n");
pd->pd_recursion--;
{
down(&pinger_sem);
imp->imp_next_ping = time(NULL) + obd_timeout;
- if (pinger_args.pd_set == NULL &&
- pinger_args.pd_next_ping > imp->imp_next_ping) {
+ if (pinger_args.pd_next_ping > imp->imp_next_ping) {
CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
imp->imp_next_ping, time(NULL));
pinger_args.pd_next_ping = imp->imp_next_ping;
RETURN(0);
}
+void ptlrpc_pinger_wake_up()
+{
+ pinger_args.pd_force_check = 1;
+}
#endif /* !__KERNEL__ */