*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
* Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2010, 2014, Intel Corporation.
+ * Copyright (c) 2010, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*/
#define DEBUG_SUBSYSTEM S_RPC
+#include <linux/kthread.h>
#include <obd_support.h>
#include <obd_class.h>
#include <lustre_net.h>
/* The following are visible and mutable through /sys/module/ptlrpc */
int test_req_buffer_pressure = 0;
-CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
- "set non-zero to put pressure on request buffer pools");
-CFS_MODULE_PARM(at_min, "i", int, 0644,
- "Adaptive timeout minimum (sec)");
-CFS_MODULE_PARM(at_max, "i", int, 0644,
- "Adaptive timeout maximum (sec)");
-CFS_MODULE_PARM(at_history, "i", int, 0644,
- "Adaptive timeouts remember the slowest event that took place "
- "within this period (sec)");
-CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
- "How soon before an RPC deadline to send an early reply");
-CFS_MODULE_PARM(at_extra, "i", int, 0644,
- "How much extra time to give with each early reply");
-
+module_param(test_req_buffer_pressure, int, 0444);
+MODULE_PARM_DESC(test_req_buffer_pressure, "set non-zero to put pressure on request buffer pools");
+module_param(at_min, int, 0644);
+MODULE_PARM_DESC(at_min, "Adaptive timeout minimum (sec)");
+module_param(at_max, int, 0644);
+MODULE_PARM_DESC(at_max, "Adaptive timeout maximum (sec)");
+module_param(at_history, int, 0644);
+MODULE_PARM_DESC(at_history,
+ "Adaptive timeouts remember the slowest event that took place within this period (sec)");
+module_param(at_early_margin, int, 0644);
+MODULE_PARM_DESC(at_early_margin, "How soon before an RPC deadline to send an early reply");
+module_param(at_extra, int, 0644);
+MODULE_PARM_DESC(at_extra, "How much extra time to give with each early reply");
/* forward ref */
static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
* Puts a lock and its mode into reply state assotiated to request reply.
*/
void
-ptlrpc_save_lock(struct ptlrpc_request *req,
- struct lustre_handle *lock, int mode, int no_ack)
+ptlrpc_save_lock(struct ptlrpc_request *req, struct lustre_handle *lock,
+ int mode, bool no_ack, bool convert_lock)
{
- struct ptlrpc_reply_state *rs = req->rq_reply_state;
- int idx;
+ struct ptlrpc_reply_state *rs = req->rq_reply_state;
+ int idx;
- LASSERT(rs != NULL);
- LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
+ LASSERT(rs != NULL);
+ LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
- if (req->rq_export->exp_disconnected) {
- ldlm_lock_decref(lock, mode);
- } else {
- idx = rs->rs_nlocks++;
- rs->rs_locks[idx] = *lock;
- rs->rs_modes[idx] = mode;
- rs->rs_difficult = 1;
- rs->rs_no_ack = !!no_ack;
- }
+ idx = rs->rs_nlocks++;
+ rs->rs_locks[idx] = *lock;
+ rs->rs_modes[idx] = mode;
+ rs->rs_difficult = 1;
+ rs->rs_no_ack = no_ack;
+ rs->rs_convert_lock = convert_lock;
}
EXPORT_SYMBOL(ptlrpc_save_lock);
int hrt_id; /* thread ID */
spinlock_t hrt_lock;
wait_queue_head_t hrt_waitq;
- struct list_head hrt_queue; /* RS queue */
+ struct list_head hrt_queue;
struct ptlrpc_hr_partition *hrt_partition;
};
/*
* User wants to increase number of threads with for
- * each CPU core/HT, most likely the factor is larger then
+ * each CPU core/HT, most likely the factor is larger than
* one thread/core because service threads are supposed to
* be blocked by lock or wait for IO.
*/
* have too many threads no matter how many cores/HTs
* there are.
*/
- if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
+ if (cpumask_weight(topology_sibling_cpumask(smp_processor_id())) > 1) {
+ /* weight is # of HTs */
/* depress thread factor for hyper-thread */
factor = factor - (factor >> 1) + (factor >> 3);
}
weight = cfs_cpt_weight(svc->srv_cptable, 0);
- LASSERT(weight > 0);
for (; factor > 0 && weight > 0; factor--, weight -= fade)
nthrs += min(weight, fade) * factor;
if (array->paa_reqs_count == NULL)
goto failed;
- cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, svcpt);
+ setup_timer(&svcpt->scp_at_timer, ptlrpc_at_timer,
+ (unsigned long)svcpt);
+
/* At SOW, service time should be quick; 10s seems generous. If client
* timeout is less than this, we'll be sending an early reply. */
at_init(&svcpt->scp_at_estimate, 10, 0);
*/
struct ptlrpc_service *
ptlrpc_register_service(struct ptlrpc_service_conf *conf,
+ struct kset *parent,
struct proc_dir_entry *proc_entry)
{
struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
list_add(&service->srv_list, &ptlrpc_all_services);
mutex_unlock(&ptlrpc_all_services_mutex);
+ if (parent) {
+ rc = ptlrpc_sysfs_register_service(parent, service);
+ if (rc)
+ GOTO(failed, rc);
+ }
+
if (proc_entry != NULL)
ptlrpc_lprocfs_register_service(proc_entry, service);
spin_lock(&svcpt->scp_lock);
/*
* now all reqs including the embedded req has been
- * disposed, schedule request buffer for re-use.
+ * disposed, schedule request buffer for re-use
+ * or free it to drain some in excess.
*/
LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
- list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
+ if (svcpt->scp_nrqbds_posted >=
+ svc->srv_nbuf_per_group &&
+ !test_req_buffer_pressure) {
+ /* like in ptlrpc_free_rqbd() */
+ svcpt->scp_nrqbds_total--;
+ OBD_FREE_LARGE(rqbd->rqbd_buffer,
+ svc->srv_buf_size);
+ OBD_FREE_PTR(rqbd);
+ } else {
+ list_add_tail(&rqbd->rqbd_list,
+ &svcpt->scp_rqbd_idle);
+ }
}
spin_unlock(&svcpt->scp_lock);
}
/**
- * to finish a active request: stop sending more early replies, and release
+ * to finish an active request: stop sending more early replies, and release
* the request. should be called after we finished handling the request.
*/
static void ptlrpc_server_finish_active_request(
* we better wait for 3. */
exp->exp_obd->obd_eviction_timer =
cfs_time_current_sec() + 3 * PING_INTERVAL;
- CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
+ CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
exp->exp_obd->obd_name,
obd_export_nid2str(oldest_exp), oldest_time);
}
} else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
!obd->obd_recovering) {
DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
- LPU64" without recovery",
+ "%llu without recovery",
lustre_msg_get_transno(req->rq_reqmsg));
class_fail_export(req->rq_export);
rc = -ENODEV;
__s32 next;
if (array->paa_count == 0) {
- cfs_timer_disarm(&svcpt->scp_at_timer);
+ del_timer(&svcpt->scp_at_timer);
return;
}
/* Set timer for closest deadline */
- next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
+ next = (__s32)(array->paa_deadline - ktime_get_real_seconds() -
at_early_margin);
if (next <= 0) {
ptlrpc_at_timer((unsigned long)svcpt);
} else {
- cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
+ mod_timer(&svcpt->scp_at_timer, cfs_time_shift(next));
CDEBUG(D_INFO, "armed %s at %+ds\n",
svcpt->scp_service->srv_name, next);
}
spin_lock(&svcpt->scp_at_lock);
LASSERT(list_empty(&req->rq_timed_list));
- index = (unsigned long)req->rq_deadline % array->paa_size;
+ div_u64_rem(req->rq_deadline, array->paa_size, &index);
if (array->paa_reqs_count[index] > 0) {
/* latest rpcs will have the latest deadlines in the list,
* so search backward. */
struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
struct ptlrpc_request *reqcopy;
struct lustre_msg *reqmsg;
- cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
- time_t newdl;
+ time64_t olddl = req->rq_deadline - ktime_get_real_seconds();
+ time64_t newdl;
int rc;
ENTRY;
/* deadline is when the client expects us to reply, margin is the
difference between clients' and servers' expectations */
DEBUG_REQ(D_ADAPTTO, req,
- "%ssending early reply (deadline %+lds, margin %+lds) for "
+ "%ssending early reply (deadline %+llds, margin %+llds) for "
"%d+%d", AT_OFF ? "AT off - not " : "",
- olddl, olddl - at_get(&svcpt->scp_at_estimate),
+ (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
at_get(&svcpt->scp_at_estimate), at_extra);
if (AT_OFF)
RETURN(0);
if (olddl < 0) {
- DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
+ DEBUG_REQ(D_WARNING, req, "Already past deadline (%+llds), "
"not sending early reply. Consider increasing "
- "at_early_margin (%d)?", olddl, at_early_margin);
+ "at_early_margin (%d)?", (s64)olddl, at_early_margin);
/* Return an error so we're not re-added to the timed list. */
RETURN(-ETIMEDOUT);
if (req->rq_export &&
lustre_msg_get_flags(req->rq_reqmsg) &
(MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
+ struct obd_device *obd_exp = req->rq_export->exp_obd;
+
/* During recovery, we don't want to send too many early
* replies, but on the other hand we want to make sure the
* client has enough time to resend if the rpc is lost. So
* during the recovery period send at least 4 early replies,
* spacing them every at_extra if we can. at_estimate should
- * always equal this fixed value during recovery. */
+ * always equal this fixed value during recovery.
+ */
/* Don't account request processing time into AT history
* during recovery, it is not service time we need but
- * includes also waiting time for recovering clients */
- newdl = cfs_time_current_sec() + min(at_extra,
- req->rq_export->exp_obd->obd_recovery_timeout / 4);
+ * includes also waiting time for recovering clients
+ */
+ newdl = min_t(time64_t, at_extra,
+ obd_exp->obd_recovery_timeout / 4) +
+ ktime_get_real_seconds();
} else {
/* We want to extend the request deadline by at_extra seconds,
* so we set our service estimate to reflect how much time has
* account for network latency). See ptlrpc_at_recv_early_reply
*/
at_measured(&svcpt->scp_at_estimate, at_extra +
- cfs_time_current_sec() -
+ ktime_get_real_seconds() -
req->rq_arrival_time.tv_sec);
newdl = req->rq_arrival_time.tv_sec +
at_get(&svcpt->scp_at_estimate);
/* Check to see if we've actually increased the deadline -
* we may be past adaptive_max */
if (req->rq_deadline >= newdl) {
- DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
- "(%ld/%ld), not sending early reply\n",
- olddl, newdl - cfs_time_current_sec());
+ DEBUG_REQ(D_WARNING, req, "Couldn't add any time (%lld/%lld), not sending early reply\n",
+ (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
RETURN(-ETIMEDOUT);
}
struct ptlrpc_request *rq, *n;
struct list_head work_list;
__u32 index, count;
- time_t deadline;
- time_t now = cfs_time_current_sec();
+ time64_t deadline;
+ time64_t now = ktime_get_real_seconds();
cfs_duration_t delay;
int first, counter = 0;
ENTRY;
server will take. Send early replies to everyone expiring soon. */
INIT_LIST_HEAD(&work_list);
deadline = -1;
- index = (unsigned long)array->paa_deadline % array->paa_size;
+ div_u64_rem(array->paa_deadline, array->paa_size, &index);
count = array->paa_count;
while (count > 0) {
count -= array->paa_reqs_count[index];
LCONSOLE_WARN("%s: This server is not able to keep up with "
"request traffic (cpu-bound).\n",
svcpt->scp_service->srv_name);
- CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
- "delay="CFS_DURATION_T"(jiff)\n",
+ CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, delay=%ld(jiff)\n",
counter, svcpt->scp_nreqs_incoming,
svcpt->scp_nreqs_active,
at_get(&svcpt->scp_at_estimate), delay);
}
/**
- * Put the request to the export list if the request may become
- * a high priority one.
+ * Check if a request should be assigned with a high priority.
+ *
+ * \retval < 0: error occurred
+ * 0: normal RPC request
+ * +1: high priority request
*/
static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
struct ptlrpc_request *req)
{
- struct list_head *list;
- int rc, hp = 0;
-
+ int rc = 0;
ENTRY;
- if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
+ if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
if (rc < 0)
RETURN(rc);
+
LASSERT(rc == 0);
}
- if (req->rq_export) {
- if (req->rq_ops) {
- /* Perform request specific check. We should do this
- * check before the request is added into exp_hp_rpcs
- * list otherwise it may hit swab race at LU-1044. */
- if (req->rq_ops->hpreq_check) {
- rc = req->rq_ops->hpreq_check(req);
- /**
- * XXX: Out of all current
- * ptlrpc_hpreq_ops::hpreq_check(), only
- * ldlm_cancel_hpreq_check() can return an
- * error code; other functions assert in
- * similar places, which seems odd.
- * What also does not seem right is that
- * handlers for those RPCs do not assert
- * on the same checks, but rather handle the
- * error cases. e.g. see ost_rw_hpreq_check(),
- * and ost_brw_read(), ost_brw_write().
- */
- if (rc < 0)
- RETURN(rc);
- LASSERT(rc == 0 || rc == 1);
- hp = rc;
- }
- list = &req->rq_export->exp_hp_rpcs;
- } else {
- list = &req->rq_export->exp_reg_rpcs;
- }
- /* do search for duplicated xid and the adding to the list
- * atomically */
- spin_lock_bh(&req->rq_export->exp_rpc_lock);
- rc = ptlrpc_server_check_resend_in_progress(req);
- if (rc < 0) {
- spin_unlock_bh(&req->rq_export->exp_rpc_lock);
- RETURN(rc);
+ if (req->rq_export != NULL && req->rq_ops != NULL) {
+ /* Perform request specific check. We should do this
+ * check before the request is added into exp_hp_rpcs
+ * list otherwise it may hit swab race at LU-1044. */
+ if (req->rq_ops->hpreq_check != NULL) {
+ rc = req->rq_ops->hpreq_check(req);
+ if (rc == -ESTALE) {
+ req->rq_status = rc;
+ ptlrpc_error(req);
+ }
+ /** can only return error,
+ * 0 for normal request,
+ * or 1 for high priority request */
+ LASSERT(rc <= 1);
}
- list_add(&req->rq_exp_list, list);
- spin_unlock_bh(&req->rq_export->exp_rpc_lock);
}
- ptlrpc_nrs_req_initialize(svcpt, req, !!hp);
-
- RETURN(hp);
+ RETURN(rc);
}
/** Remove the request from the export list. */
static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
struct ptlrpc_request *req)
{
- int rc;
+ int rc;
+ bool hp;
ENTRY;
rc = ptlrpc_server_hpreq_init(svcpt, req);
if (rc < 0)
RETURN(rc);
+ hp = rc > 0;
+ ptlrpc_nrs_req_initialize(svcpt, req, hp);
+
+ if (req->rq_export != NULL) {
+ struct obd_export *exp = req->rq_export;
+
+ /* do search for duplicated xid and the adding to the list
+ * atomically */
+ spin_lock_bh(&exp->exp_rpc_lock);
+ rc = ptlrpc_server_check_resend_in_progress(req);
+ if (rc < 0) {
+ spin_unlock_bh(&exp->exp_rpc_lock);
+
+ ptlrpc_nrs_req_finalize(req);
+ RETURN(rc);
+ }
+
+ if (hp || req->rq_ops != NULL)
+ list_add(&req->rq_exp_list, &exp->exp_hp_rpcs);
+ else
+ list_add(&req->rq_exp_list, &exp->exp_reg_rpcs);
+ spin_unlock_bh(&exp->exp_rpc_lock);
+ }
+
/* the current thread is not the processing thread for this request
* since that, but request is in exp_hp_list and can be find there.
* Remove all relations between request and old thread. */
req->rq_svc_thread = NULL;
req->rq_session.lc_thread = NULL;
- ptlrpc_nrs_req_add(svcpt, req, !!rc);
+ ptlrpc_nrs_req_add(svcpt, req, hp);
RETURN(0);
}
rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
if (rc != 0) {
CERROR("error unpacking request: ptl %d from %s "
- "x"LPU64"\n", svc->srv_req_portal,
+ "x%llu\n", svc->srv_req_portal,
libcfs_id2str(req->rq_peer), req->rq_xid);
goto err_req;
}
rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
if (rc) {
CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
- LPU64"\n", svc->srv_req_portal,
+ "%llu\n", svc->srv_req_portal,
libcfs_id2str(req->rq_peer), req->rq_xid);
goto err_req;
}
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_REQ_OPC) &&
lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val) {
- CERROR("drop incoming rpc opc %u, x"LPU64"\n",
+ CERROR("drop incoming rpc opc %u, x%llu\n",
cfs_fail_val, req->rq_xid);
goto err_req;
}
goto err_req;
}
- switch(lustre_msg_get_opc(req->rq_reqmsg)) {
- case MDS_WRITEPAGE:
- case OST_WRITE:
- req->rq_bulk_write = 1;
- break;
- case MDS_READPAGE:
- case OST_READ:
- case MGS_CONFIG_READ:
- req->rq_bulk_read = 1;
- break;
- }
+ switch (lustre_msg_get_opc(req->rq_reqmsg)) {
+ case MDS_WRITEPAGE:
+ case OST_WRITE:
+ case OUT_UPDATE:
+ req->rq_bulk_write = 1;
+ break;
+ case MDS_READPAGE:
+ case OST_READ:
+ case MGS_CONFIG_READ:
+ req->rq_bulk_read = 1;
+ break;
+ }
- CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
+ CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
req->rq_export = class_conn2export(
lustre_msg_get_handle(req->rq_reqmsg));
}
/* req_in handling should/must be fast */
- if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
- DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
- cfs_time_sub(cfs_time_current_sec(),
- req->rq_arrival_time.tv_sec));
+ if (ktime_get_real_seconds() - req->rq_arrival_time.tv_sec > 5)
+ DEBUG_REQ(D_WARNING, req, "Slow req_in handling %llds",
+ (s64)(ktime_get_real_seconds() -
+ req->rq_arrival_time.tv_sec));
/* Set rpc server deadline and add it to the timed list */
deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
struct ptlrpc_thread *thread)
{
- struct ptlrpc_service *svc = svcpt->scp_service;
- struct ptlrpc_request *request;
- struct timeval work_start;
- struct timeval work_end;
- long timediff;
- int fail_opc = 0;
+ struct ptlrpc_service *svc = svcpt->scp_service;
+ struct ptlrpc_request *request;
+ ktime_t work_start;
+ ktime_t work_end;
+ ktime_t arrived;
+ s64 timediff_usecs;
+ s64 arrived_usecs;
+ int fail_opc = 0;
ENTRY;
if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
libcfs_debug_dumplog();
- do_gettimeofday(&work_start);
- timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
+ work_start = ktime_get_real();
+ arrived = timespec64_to_ktime(request->rq_arrival_time);
+ timediff_usecs = ktime_us_delta(work_start, arrived);
if (likely(svc->srv_stats != NULL)) {
lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
- timediff);
+ timediff_usecs);
lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
svcpt->scp_nreqs_incoming);
lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
if (likely(request->rq_export)) {
if (unlikely(ptlrpc_check_req(request)))
goto put_conn;
- ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
+ ptlrpc_update_export_timer(request->rq_export,
+ timediff_usecs >> 19);
}
/* Discard requests queued for longer than the deadline.
The deadline is increased if we send an early reply. */
- if (cfs_time_current_sec() > request->rq_deadline) {
- DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
- ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
+ if (ktime_get_real_seconds() > request->rq_deadline) {
+ DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s: deadline %lld:%llds ago\n",
libcfs_id2str(request->rq_peer),
- cfs_time_sub(request->rq_deadline,
- request->rq_arrival_time.tv_sec),
- cfs_time_sub(cfs_time_current_sec(),
- request->rq_deadline));
+ request->rq_deadline -
+ request->rq_arrival_time.tv_sec,
+ ktime_get_real_seconds() - request->rq_deadline);
goto put_conn;
}
CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
- "%s:%s+%d:%d:x"LPU64":%s:%d\n", current_comm(),
+ "%s:%s+%d:%d:x%llu:%s:%d\n", current_comm(),
(request->rq_export ?
(char *)request->rq_export->exp_client_uuid.uuid : "0"),
(request->rq_export ?
if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
- CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
+ CDEBUG(D_NET, "got req %llu\n", request->rq_xid);
/* re-assign request and sesson thread to the current one */
request->rq_svc_thread = thread;
ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
put_conn:
- if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
- DEBUG_REQ(D_WARNING, request, "Request took longer "
- "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
- " client may timeout.",
- cfs_time_sub(request->rq_deadline,
- request->rq_arrival_time.tv_sec),
- cfs_time_sub(cfs_time_current_sec(),
- request->rq_deadline));
- }
-
- do_gettimeofday(&work_end);
- timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
- CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
- "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
- "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
+ if (unlikely(ktime_get_real_seconds() > request->rq_deadline)) {
+ DEBUG_REQ(D_WARNING, request, "Request took longer than estimated (%lld:%llds); client may timeout.",
+ request->rq_deadline -
+ request->rq_arrival_time.tv_sec,
+ ktime_get_real_seconds() - request->rq_deadline);
+ }
+
+ work_end = ktime_get_real();
+ timediff_usecs = ktime_us_delta(work_end, work_start);
+ arrived_usecs = ktime_us_delta(work_end, arrived);
+ CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc %s:%s+%d:%d:x%llu:%s:%d Request procesed in %lldus (%lldus total) trans %llu rc %d/%d\n",
current_comm(),
(request->rq_export ?
(char *)request->rq_export->exp_client_uuid.uuid : "0"),
request->rq_xid,
libcfs_id2str(request->rq_peer),
lustre_msg_get_opc(request->rq_reqmsg),
- timediff,
- cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
+ timediff_usecs,
+ arrived_usecs,
(request->rq_repmsg ?
lustre_msg_get_transno(request->rq_repmsg) :
request->rq_transno),
LASSERT(opc < LUSTRE_MAX_OPCODES);
lprocfs_counter_add(svc->srv_stats,
opc + EXTRA_MAX_OPCODES,
- timediff);
+ timediff_usecs);
}
}
if (unlikely(request->rq_early_count)) {
DEBUG_REQ(D_ADAPTTO, request,
- "sent %d early replies before finishing in "
- CFS_DURATION_T"s",
+ "sent %d early replies before finishing in %llds",
request->rq_early_count,
- cfs_time_sub(work_end.tv_sec,
- request->rq_arrival_time.tv_sec));
+ arrived_usecs / USEC_PER_SEC);
}
ptlrpc_server_finish_active_request(svcpt, request);
{
struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
struct ptlrpc_service *svc = svcpt->scp_service;
- struct obd_export *exp;
- int nlocks;
- int been_handled;
- ENTRY;
+ struct obd_export *exp;
+ int nlocks;
+ int been_handled;
+ ENTRY;
exp = rs->rs_export;
LASSERT(rs->rs_scheduled);
LASSERT(list_empty(&rs->rs_list));
- spin_lock(&exp->exp_lock);
- /* Noop if removed already */
- list_del_init(&rs->rs_exp_list);
- spin_unlock(&exp->exp_lock);
-
- /* The disk commit callback holds exp_uncommitted_replies_lock while it
- * iterates over newly committed replies, removing them from
- * exp_uncommitted_replies. It then drops this lock and schedules the
- * replies it found for handling here.
- *
- * We can avoid contention for exp_uncommitted_replies_lock between the
- * HRT threads and further commit callbacks by checking rs_committed
- * which is set in the commit callback while it holds both
- * rs_lock and exp_uncommitted_reples.
- *
- * If we see rs_committed clear, the commit callback _may_ not have
- * handled this reply yet and we race with it to grab
- * exp_uncommitted_replies_lock before removing the reply from
- * exp_uncommitted_replies. Note that if we lose the race and the
- * reply has already been removed, list_del_init() is a noop.
- *
- * If we see rs_committed set, we know the commit callback is handling,
- * or has handled this reply since store reordering might allow us to
- * see rs_committed set out of sequence. But since this is done
- * holding rs_lock, we can be sure it has all completed once we hold
- * rs_lock, which we do right next.
- */
+ /* The disk commit callback holds exp_uncommitted_replies_lock while it
+ * iterates over newly committed replies, removing them from
+ * exp_uncommitted_replies. It then drops this lock and schedules the
+ * replies it found for handling here.
+ *
+ * We can avoid contention for exp_uncommitted_replies_lock between the
+ * HRT threads and further commit callbacks by checking rs_committed
+ * which is set in the commit callback while it holds both
+ * rs_lock and exp_uncommitted_reples.
+ *
+ * If we see rs_committed clear, the commit callback _may_ not have
+ * handled this reply yet and we race with it to grab
+ * exp_uncommitted_replies_lock before removing the reply from
+ * exp_uncommitted_replies. Note that if we lose the race and the
+ * reply has already been removed, list_del_init() is a noop.
+ *
+ * If we see rs_committed set, we know the commit callback is handling,
+ * or has handled this reply since store reordering might allow us to
+ * see rs_committed set out of sequence. But since this is done
+ * holding rs_lock, we can be sure it has all completed once we hold
+ * rs_lock, which we do right next.
+ */
if (!rs->rs_committed) {
+ /* if rs was commited, no need to convert locks, don't check
+ * rs_committed here because rs may never be added into
+ * exp_uncommitted_replies and this flag never be set, see
+ * target_send_reply() */
+ if (rs->rs_convert_lock &&
+ rs->rs_transno > exp->exp_last_committed) {
+ struct ldlm_lock *lock;
+ struct ldlm_lock *ack_locks[RS_MAX_LOCKS] = { NULL };
+
+ spin_lock(&rs->rs_lock);
+ if (rs->rs_convert_lock &&
+ rs->rs_transno > exp->exp_last_committed) {
+ nlocks = rs->rs_nlocks;
+ while (nlocks-- > 0) {
+ /*
+ * NB don't assume rs is always handled
+ * by the same service thread (see
+ * ptlrpc_hr_select, so REP-ACK hr may
+ * race with trans commit, while the
+ * latter will release locks, get locks
+ * here early to convert to COS mode
+ * safely.
+ */
+ lock = ldlm_handle2lock(
+ &rs->rs_locks[nlocks]);
+ LASSERT(lock);
+ ack_locks[nlocks] = lock;
+ rs->rs_modes[nlocks] = LCK_COS;
+ }
+ nlocks = rs->rs_nlocks;
+ rs->rs_convert_lock = 0;
+ /* clear rs_scheduled so that commit callback
+ * can schedule again */
+ rs->rs_scheduled = 0;
+ spin_unlock(&rs->rs_lock);
+
+ while (nlocks-- > 0) {
+ lock = ack_locks[nlocks];
+ ldlm_lock_downgrade(lock, LCK_COS);
+ LDLM_LOCK_PUT(lock);
+ }
+ RETURN(0);
+ }
+ spin_unlock(&rs->rs_lock);
+ }
+
spin_lock(&exp->exp_uncommitted_replies_lock);
list_del_init(&rs->rs_obd_list);
spin_unlock(&exp->exp_uncommitted_replies_lock);
}
+ spin_lock(&exp->exp_lock);
+ /* Noop if removed already */
+ list_del_init(&rs->rs_exp_list);
+ spin_unlock(&exp->exp_lock);
+
spin_lock(&rs->rs_lock);
- been_handled = rs->rs_handled;
- rs->rs_handled = 1;
+ been_handled = rs->rs_handled;
+ rs->rs_handled = 1;
- nlocks = rs->rs_nlocks; /* atomic "steal", but */
- rs->rs_nlocks = 0; /* locks still on rs_locks! */
+ nlocks = rs->rs_nlocks; /* atomic "steal", but */
+ rs->rs_nlocks = 0; /* locks still on rs_locks! */
- if (nlocks == 0 && !been_handled) {
- /* If we see this, we should already have seen the warning
- * in mds_steal_ack_locks() */
- CDEBUG(D_HA, "All locks stolen from rs %p x"LPD64".t"LPD64
+ if (nlocks == 0 && !been_handled) {
+ /* If we see this, we should already have seen the warning
+ * in mds_steal_ack_locks() */
+ CDEBUG(D_HA, "All locks stolen from rs %p x%lld.t%lld"
" o%d NID %s\n",
rs,
rs->rs_xid, rs->rs_transno, rs->rs_opc,
libcfs_nid2str(exp->exp_connection->c_peer.nid));
- }
+ }
- if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
+ if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
spin_unlock(&rs->rs_lock);
if (!been_handled && rs->rs_on_net) {
}
rs->rs_scheduled = 0;
+ rs->rs_convert_lock = 0;
if (!rs->rs_on_net) {
/* Off the net */
goto out_srv_fini;
}
- /* Alloc reply state structure for this one */
- OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
- if (!rs) {
- rc = -ENOMEM;
- goto out_srv_fini;
- }
+ /* Alloc reply state structure for this one */
+ OBD_ALLOC_LARGE(rs, svc->srv_max_reply_size);
+ if (!rs) {
+ rc = -ENOMEM;
+ goto out_srv_fini;
+ }
spin_lock(&svcpt->scp_lock);
struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
struct list_head replies;
- char threadname[20];
int rc;
INIT_LIST_HEAD(&replies);
- snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
- hrp->hrp_cpt, hrt->hrt_id);
unshare_fs_struct();
rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
if (rc != 0) {
+ char threadname[20];
+
+ snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
+ hrp->hrp_cpt, hrt->hrt_id);
CWARN("Failed to bind %s on CPT %d of CPT table %p: rc = %d\n",
threadname, hrp->hrp_cpt, ptlrpc_hr.hr_cpt_table, rc);
}
struct ptlrpc_hr_partition *hrp;
struct ptlrpc_hr_thread *hrt;
int rc;
+ int cpt;
int i;
- int j;
int weight;
ENTRY;
init_waitqueue_head(&ptlrpc_hr.hr_waitq);
- weight = cfs_cpu_ht_nsiblings(0);
+ weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
- cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
- hrp->hrp_cpt = i;
+ cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
+ hrp->hrp_cpt = cpt;
atomic_set(&hrp->hrp_nstarted, 0);
atomic_set(&hrp->hrp_nstopped, 0);
- hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, i);
+ hrp->hrp_nthrs = cfs_cpt_weight(ptlrpc_hr.hr_cpt_table, cpt);
hrp->hrp_nthrs /= weight;
+ if (hrp->hrp_nthrs == 0)
+ hrp->hrp_nthrs = 1;
- LASSERT(hrp->hrp_nthrs > 0);
- OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
+ OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, cpt,
hrp->hrp_nthrs * sizeof(*hrt));
if (hrp->hrp_thrs == NULL)
GOTO(out, rc = -ENOMEM);
- for (j = 0; j < hrp->hrp_nthrs; j++) {
- hrt = &hrp->hrp_thrs[j];
+ for (i = 0; i < hrp->hrp_nthrs; i++) {
+ hrt = &hrp->hrp_thrs[i];
- hrt->hrt_id = j;
+ hrt->hrt_id = i;
hrt->hrt_partition = hrp;
init_waitqueue_head(&hrt->hrt_waitq);
spin_lock_init(&hrt->hrt_lock);
void ptlrpc_hr_fini(void)
{
struct ptlrpc_hr_partition *hrp;
- int i;
+ int cpt;
if (ptlrpc_hr.hr_partitions == NULL)
return;
ptlrpc_stop_hr_threads();
- cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+ cfs_percpt_for_each(hrp, cpt, ptlrpc_hr.hr_partitions) {
if (hrp->hrp_thrs != NULL) {
OBD_FREE(hrp->hrp_thrs,
hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
/* early disarm AT timer... */
ptlrpc_service_for_each_part(svcpt, i, svc) {
if (svcpt->scp_service != NULL)
- cfs_timer_disarm(&svcpt->scp_at_timer);
+ del_timer(&svcpt->scp_at_timer);
}
}
break;
/* In case somebody rearmed this in the meantime */
- cfs_timer_disarm(&svcpt->scp_at_timer);
+ del_timer(&svcpt->scp_at_timer);
array = &svcpt->scp_at_array;
if (array->paa_reqs_array != NULL) {
ptlrpc_service_nrs_cleanup(service);
ptlrpc_lprocfs_unregister_service(service);
+ ptlrpc_sysfs_unregister_service(service);
ptlrpc_service_free(service);
* to be shot, so it's intentionally non-aggressive. */
static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
{
- struct ptlrpc_request *request = NULL;
- struct timeval right_now;
- long timediff;
+ struct ptlrpc_request *request = NULL;
+ struct timespec64 right_now;
+ struct timespec64 timediff;
- do_gettimeofday(&right_now);
+ ktime_get_real_ts64(&right_now);
spin_lock(&svcpt->scp_req_lock);
/* How long has the next entry been waiting? */
return 0;
}
- timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
+ timediff = timespec64_sub(right_now, request->rq_arrival_time);
spin_unlock(&svcpt->scp_req_lock);
- if ((timediff / ONE_MILLION) >
+ if ((timediff.tv_sec) >
(AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
- CERROR("%s: unhealthy - request has been waiting %lds\n",
- svcpt->scp_service->srv_name, timediff / ONE_MILLION);
+ CERROR("%s: unhealthy - request has been waiting %llds\n",
+ svcpt->scp_service->srv_name, (s64)timediff.tv_sec);
return -1;
}