*
* 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, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*/
#define DEBUG_SUBSYSTEM S_RPC
+
#include <linux/kthread.h>
+#include <linux/ratelimit.h>
+
#include <obd_support.h>
#include <obd_class.h>
#include <lustre_net.h>
#include <lu_object.h>
-#include <lnet/types.h>
+#include <uapi/linux/lnet/lnet-types.h>
#include "ptlrpc_internal.h"
+#include <linux/delay.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);
static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
/** Holds a list of all PTLRPC services */
-struct list_head ptlrpc_all_services;
+LIST_HEAD(ptlrpc_all_services);
/** Used to protect the \e ptlrpc_all_services list */
struct mutex ptlrpc_all_services_mutex;
return rqbd;
}
-static void
-ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
+static void ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
{
struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
OBD_FREE_PTR(rqbd);
}
-static int
-ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
+static int ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
{
- struct ptlrpc_service *svc = svcpt->scp_service;
- struct ptlrpc_request_buffer_desc *rqbd;
- int rc = 0;
- int i;
+ struct ptlrpc_service *svc = svcpt->scp_service;
+ struct ptlrpc_request_buffer_desc *rqbd;
+ int rc = 0;
+ int i;
if (svcpt->scp_rqbd_allocating)
goto try_post;
spin_unlock(&svcpt->scp_lock);
- for (i = 0; i < svc->srv_nbuf_per_group; i++) {
- /* NB: another thread might have recycled enough rqbds, we
- * need to make sure it wouldn't over-allocate, see LU-1212. */
- if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group)
+ for (i = 0; i < svc->srv_nbuf_per_group; i++) {
+ /*
+ * NB: another thread might have recycled enough rqbds, we
+ * need to make sure it wouldn't over-allocate, see LU-1212.
+ */
+ if (svcpt->scp_nrqbds_posted >= svc->srv_nbuf_per_group ||
+ (svc->srv_nrqbds_max != 0 &&
+ svcpt->scp_nrqbds_total > svc->srv_nrqbds_max))
break;
rqbd = ptlrpc_alloc_rqbd(svcpt);
- if (rqbd == NULL) {
- CERROR("%s: Can't allocate request buffer\n",
- svc->srv_name);
- rc = -ENOMEM;
- break;
- }
+ if (rqbd == NULL) {
+ CERROR("%s: Can't allocate request buffer\n",
+ svc->srv_name);
+ rc = -ENOMEM;
+ break;
+ }
}
spin_lock(&svcpt->scp_lock);
* Part of Rep-Ack logic.
* 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)
-{
- struct ptlrpc_reply_state *rs = req->rq_reply_state;
- int idx;
-
- 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;
- }
+void 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;
+
+ LASSERT(rs != NULL);
+ LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
+
+ 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;
};
struct cfs_cpt_table *hr_cpt_table;
/** controller sleep waitq */
wait_queue_head_t hr_waitq;
- unsigned int hr_stopping;
+ unsigned int hr_stopping;
/** roundrobin rotor for non-affinity service */
unsigned int hr_rotor;
/* partition data */
*/
static void rs_batch_init(struct rs_batch *b)
{
- memset(b, 0, sizeof *b);
+ memset(b, 0, sizeof(*b));
INIT_LIST_HEAD(&b->rsb_replies);
}
/**
* Choose an hr thread to dispatch requests to.
*/
-static struct ptlrpc_hr_thread *
-ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
+static
+struct ptlrpc_hr_thread *ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
{
struct ptlrpc_hr_partition *hrp;
unsigned int rotor;
void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
{
struct ptlrpc_hr_thread *hrt;
+
ENTRY;
LASSERT(list_empty(&rs->rs_list));
EXIT;
}
-void
-ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
+void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs)
{
ENTRY;
assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
assert_spin_locked(&rs->rs_lock);
- LASSERT (rs->rs_difficult);
+ LASSERT(rs->rs_difficult);
rs->rs_scheduled_ever = 1; /* flag any notification attempt */
if (rs->rs_scheduled) { /* being set up or already notified */
void ptlrpc_commit_replies(struct obd_export *exp)
{
- struct ptlrpc_reply_state *rs, *nxt;
- DECLARE_RS_BATCH(batch);
- ENTRY;
+ struct ptlrpc_reply_state *rs, *nxt;
+ DECLARE_RS_BATCH(batch);
- rs_batch_init(&batch);
- /* Find any replies that have been committed and get their service
- * to attend to complete them. */
+ ENTRY;
+
+ rs_batch_init(&batch);
+ /*
+ * Find any replies that have been committed and get their service
+ * to attend to complete them.
+ */
- /* CAVEAT EMPTOR: spinlock ordering!!! */
+ /* CAVEAT EMPTOR: spinlock ordering!!! */
spin_lock(&exp->exp_uncommitted_replies_lock);
list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
- rs_obd_list) {
- LASSERT (rs->rs_difficult);
- /* VBR: per-export last_committed */
- LASSERT(rs->rs_export);
- if (rs->rs_transno <= exp->exp_last_committed) {
+ rs_obd_list) {
+ LASSERT(rs->rs_difficult);
+ /* VBR: per-export last_committed */
+ LASSERT(rs->rs_export);
+ if (rs->rs_transno <= exp->exp_last_committed) {
list_del_init(&rs->rs_obd_list);
- rs_batch_add(&batch, rs);
- }
- }
+ rs_batch_add(&batch, rs);
+ }
+ }
spin_unlock(&exp->exp_uncommitted_replies_lock);
rs_batch_fini(&batch);
EXIT;
}
-static int
-ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
+static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt)
{
struct ptlrpc_request_buffer_desc *rqbd;
- int rc;
- int posted = 0;
+ int rc;
+ int posted = 0;
for (;;) {
spin_lock(&svcpt->scp_lock);
rqbd = list_entry(svcpt->scp_rqbd_idle.next,
struct ptlrpc_request_buffer_desc,
rqbd_list);
- list_del(&rqbd->rqbd_list);
/* assume we will post successfully */
svcpt->scp_nrqbds_posted++;
- list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
+ list_move(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
spin_unlock(&svcpt->scp_lock);
spin_lock(&svcpt->scp_lock);
svcpt->scp_nrqbds_posted--;
- list_del(&rqbd->rqbd_list);
- list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
+ list_move_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
- /* Don't complain if no request buffers are posted right now; LNET
- * won't drop requests because we set the portal lazy! */
+ /*
+ * Don't complain if no request buffers are posted right now; LNET
+ * won't drop requests because we set the portal lazy!
+ */
spin_unlock(&svcpt->scp_lock);
return -1;
}
-static void ptlrpc_at_timer(unsigned long castmeharder)
+static void ptlrpc_at_timer(cfs_timer_cb_arg_t data)
{
struct ptlrpc_service_part *svcpt;
- svcpt = (struct ptlrpc_service_part *)castmeharder;
+ svcpt = cfs_from_timer(svcpt, data, scp_at_timer);
svcpt->scp_at_check = 1;
- svcpt->scp_at_checktime = cfs_time_current();
+ svcpt->scp_at_checktime = ktime_get();
wake_up(&svcpt->scp_waitq);
}
-static void
-ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
- struct ptlrpc_service_conf *conf)
+static void ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
+ struct ptlrpc_service_conf *conf)
{
- struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
- unsigned init;
- unsigned total;
- unsigned nthrs;
- int weight;
+ struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
+ unsigned int init;
+ unsigned int total;
+ unsigned int nthrs;
+ int weight;
/*
* Common code for estimating & validating threads number.
init = PTLRPC_NTHRS_INIT + (svc->srv_ops.so_hpreq_handler != NULL);
init = max_t(int, init, tc->tc_nthrs_init);
- /* NB: please see comments in lustre_lnet.h for definition
- * details of these members */
+ /*
+ * NB: please see comments in lustre_lnet.h for definition
+ * details of these members
+ */
LASSERT(tc->tc_nthrs_max != 0);
if (tc->tc_nthrs_user != 0) {
- /* In case there is a reason to test a service with many
+ /*
+ * In case there is a reason to test a service with many
* threads, we give a less strict check here, it can
- * be up to 8 * nthrs_max */
+ * be up to 8 * nthrs_max
+ */
total = min(tc->tc_nthrs_max * 8, tc->tc_nthrs_user);
nthrs = total / svc->srv_ncpts;
init = max(init, nthrs);
total = tc->tc_nthrs_max;
if (tc->tc_nthrs_base == 0) {
- /* don't care about base threads number per partition,
- * this is most for non-affinity service */
+ /*
+ * don't care about base threads number per partition,
+ * this is most for non-affinity service
+ */
nthrs = total / svc->srv_ncpts;
goto out;
}
if (svc->srv_ncpts == 1) {
int i;
- /* NB: Increase the base number if it's single partition
+ /*
+ * NB: Increase the base number if it's single partition
* and total number of cores/HTs is larger or equal to 4.
- * result will always < 2 * nthrs_base */
+ * result will always < 2 * nthrs_base
+ */
weight = cfs_cpt_weight(svc->srv_cptable, CFS_CPT_ANY);
for (i = 1; (weight >> (i + 1)) != 0 && /* >= 4 cores/HTs */
(tc->tc_nthrs_base >> i) != 0; i++)
* have too many threads no matter how many cores/HTs
* there are.
*/
- if (cfs_cpu_ht_nsiblings(0) > 1) { /* weight is # of HTs */
+ preempt_disable();
+ 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);
}
+ preempt_enable();
weight = cfs_cpt_weight(svc->srv_cptable, 0);
- LASSERT(weight > 0);
for (; factor > 0 && weight > 0; factor--, weight -= fade)
nthrs += min(weight, fade) * factor;
svc->srv_nthrs_cpt_init = init;
if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
- CDEBUG(D_OTHER, "%s: This service may have more threads (%d) "
- "than the given soft limit (%d)\n",
+ CDEBUG(D_OTHER,
+ "%s: This service may have more threads (%d) than the given soft limit (%d)\n",
svc->srv_name, nthrs * svc->srv_ncpts,
tc->tc_nthrs_max);
}
/**
* Initialize percpt data for a service
*/
-static int
-ptlrpc_service_part_init(struct ptlrpc_service *svc,
- struct ptlrpc_service_part *svcpt, int cpt)
+static int ptlrpc_service_part_init(struct ptlrpc_service *svc,
+ struct ptlrpc_service_part *svcpt, int cpt)
{
- struct ptlrpc_at_array *array;
- int size;
- int index;
- int rc;
+ struct ptlrpc_at_array *array;
+ int size;
+ int index;
+ int rc;
svcpt->scp_cpt = cpt;
INIT_LIST_HEAD(&svcpt->scp_threads);
/* rqbd and incoming request queue */
spin_lock_init(&svcpt->scp_lock);
+ mutex_init(&svcpt->scp_mutex);
INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
INIT_LIST_HEAD(&svcpt->scp_req_incoming);
if (array->paa_reqs_count == NULL)
goto failed;
- cfs_timer_init(&svcpt->scp_at_timer, ptlrpc_at_timer, 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. */
+ cfs_timer_setup(&svcpt->scp_at_timer, ptlrpc_at_timer,
+ (unsigned long)svcpt, 0);
+
+ /*
+ * 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);
/* assign this before call ptlrpc_grow_req_bufs */
svcpt->scp_service = svc;
/* Now allocate the request buffers, but don't post them now */
rc = ptlrpc_grow_req_bufs(svcpt, 0);
- /* We shouldn't be under memory pressure at startup, so
- * fail if we can't allocate all our buffers at this time. */
+ /*
+ * We shouldn't be under memory pressure at startup, so
+ * fail if we can't allocate all our buffers at this time.
+ */
if (rc != 0)
goto failed;
* This includes starting serving threads , allocating and posting rqbds and
* so on.
*/
-struct ptlrpc_service *
-ptlrpc_register_service(struct ptlrpc_service_conf *conf,
- struct proc_dir_entry *proc_entry)
+struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf,
+ struct kset *parent,
+ struct dentry *debugfs_entry)
{
- struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
- struct ptlrpc_service *service;
- struct ptlrpc_service_part *svcpt;
- struct cfs_cpt_table *cptable;
- __u32 *cpts = NULL;
- int ncpts;
- int cpt;
- int rc;
- int i;
+ struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
+ struct ptlrpc_service *service;
+ struct ptlrpc_service_part *svcpt;
+ struct cfs_cpt_table *cptable;
+ __u32 *cpts = NULL;
+ int ncpts;
+ int cpt;
+ int rc;
+ int i;
+
ENTRY;
LASSERT(conf->psc_buf.bc_nbufs > 0);
if (cptable == NULL)
cptable = cfs_cpt_table;
- if (!conf->psc_thr.tc_cpu_affinity) {
+ if (conf->psc_thr.tc_cpu_bind > 1) {
+ CERROR("%s: Invalid cpu bind value %d, only 1 or 0 allowed\n",
+ conf->psc_name, conf->psc_thr.tc_cpu_bind);
+ RETURN(ERR_PTR(-EINVAL));
+ }
+
+ if (!cconf->cc_affinity) {
ncpts = 1;
} else {
ncpts = cfs_cpt_number(cptable);
service->srv_cptable = cptable;
service->srv_cpts = cpts;
service->srv_ncpts = ncpts;
+ service->srv_cpt_bind = conf->psc_thr.tc_cpu_bind;
service->srv_cpt_bits = 0; /* it's zero already, easy to read... */
while ((1 << service->srv_cpt_bits) < cfs_cpt_number(cptable))
/* buffer configuration */
service->srv_nbuf_per_group = test_req_buffer_pressure ?
1 : conf->psc_buf.bc_nbufs;
+ /* do not limit max number of rqbds by default */
+ service->srv_nrqbds_max = 0;
+
service->srv_max_req_size = conf->psc_buf.bc_req_max_size +
SPTLRPC_MAX_PAYLOAD;
service->srv_buf_size = conf->psc_buf.bc_buf_size;
service->srv_rep_portal = conf->psc_buf.bc_rep_portal;
service->srv_req_portal = conf->psc_buf.bc_req_portal;
+ /* With slab/alloc_pages buffer size will be rounded up to 2^n */
+ if (service->srv_buf_size & (service->srv_buf_size - 1)) {
+ int round = size_roundup_power2(service->srv_buf_size);
+
+ service->srv_buf_size = round;
+ }
+
/* Increase max reply size to next power of two */
service->srv_max_reply_size = 1;
while (service->srv_max_reply_size <
service->srv_ops = conf->psc_ops;
for (i = 0; i < ncpts; i++) {
- if (!conf->psc_thr.tc_cpu_affinity)
+ if (!cconf->cc_affinity)
cpt = CFS_CPT_ANY;
else
cpt = cpts != NULL ? cpts[i] : i;
list_add(&service->srv_list, &ptlrpc_all_services);
mutex_unlock(&ptlrpc_all_services_mutex);
- if (proc_entry != NULL)
- ptlrpc_lprocfs_register_service(proc_entry, service);
+ if (parent) {
+ rc = ptlrpc_sysfs_register_service(parent, service);
+ if (rc)
+ GOTO(failed, rc);
+ }
+
+ if (debugfs_entry != NULL)
+ ptlrpc_ldebugfs_register_service(debugfs_entry, service);
rc = ptlrpc_service_nrs_setup(service);
if (rc != 0)
LASSERT(atomic_read(&req->rq_refcount) == 0);
LASSERT(list_empty(&req->rq_timed_list));
- /* DEBUG_REQ() assumes the reply state of a request with a valid
- * ref will not be destroyed until that reference is dropped. */
+ /*
+ * DEBUG_REQ() assumes the reply state of a request with a valid
+ * ref will not be destroyed until that reference is dropped.
+ */
ptlrpc_req_drop_rs(req);
sptlrpc_svc_ctx_decref(req);
if (req != &req->rq_rqbd->rqbd_req) {
- /* NB request buffers use an embedded
+ /*
+ * NB request buffers use an embedded
* req if the incoming req unlinked the
- * MD; this isn't one of them! */
+ * MD; this isn't one of them!
+ */
ptlrpc_request_cache_free(req);
}
}
if (req->rq_at_linked) {
spin_lock(&svcpt->scp_at_lock);
- /* recheck with lock, in case it's unlinked by
- * ptlrpc_at_check_timed() */
+ /*
+ * recheck with lock, in case it's unlinked by
+ * ptlrpc_at_check_timed()
+ */
if (likely(req->rq_at_linked))
ptlrpc_at_remove_timed(req);
spin_unlock(&svcpt->scp_at_lock);
refcount = --(rqbd->rqbd_refcount);
if (refcount == 0) {
/* request buffer is now idle: add to history */
- list_del(&rqbd->rqbd_list);
-
- list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
+ list_move_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
svcpt->scp_hist_nrqbds++;
- /* cull some history?
- * I expect only about 1 or 2 rqbds need to be recycled here */
+ /*
+ * cull some history?
+ * I expect only about 1 or 2 rqbds need to be recycled here
+ */
while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
rqbd = list_entry(svcpt->scp_hist_rqbds.next,
struct ptlrpc_request_buffer_desc,
list_del(&rqbd->rqbd_list);
svcpt->scp_hist_nrqbds--;
- /* remove rqbd's reqs from svc's req history while
- * I've got the service lock */
+ /*
+ * remove rqbd's reqs from svc's req history while
+ * I've got the service lock
+ */
list_for_each(tmp, &rqbd->rqbd_reqs) {
req = list_entry(tmp, struct ptlrpc_request,
rq_list);
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 ||
+ (svc->srv_nrqbds_max != 0 &&
+ svcpt->scp_nrqbds_total > svc->srv_nrqbds_max) ||
+ 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);
}
}
+static void ptlrpc_add_exp_list_nolock(struct ptlrpc_request *req,
+ struct obd_export *export, bool hp)
+{
+ __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
+
+ if (hp)
+ list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
+ else
+ list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
+ if (tag && export->exp_used_slots)
+ set_bit(tag - 1, export->exp_used_slots);
+}
+
+static void ptlrpc_del_exp_list(struct ptlrpc_request *req)
+{
+ __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
+
+ spin_lock(&req->rq_export->exp_rpc_lock);
+ list_del_init(&req->rq_exp_list);
+ if (tag && !req->rq_obsolete && req->rq_export->exp_used_slots)
+ clear_bit(tag - 1, req->rq_export->exp_used_slots);
+ spin_unlock(&req->rq_export->exp_rpc_lock);
+}
+
/** Change request export and move hp request from old export to new */
void ptlrpc_request_change_export(struct ptlrpc_request *req,
struct obd_export *export)
if (req->rq_export != NULL) {
LASSERT(!list_empty(&req->rq_exp_list));
/* remove rq_exp_list from last export */
- spin_lock_bh(&req->rq_export->exp_rpc_lock);
- list_del_init(&req->rq_exp_list);
- spin_unlock_bh(&req->rq_export->exp_rpc_lock);
- /* export has one reference already, so it`s safe to
+ ptlrpc_del_exp_list(req);
+ /* export has one reference already, so it's safe to
* add req to export queue here and get another
- * reference for request later */
- spin_lock_bh(&export->exp_rpc_lock);
- if (req->rq_ops != NULL) /* hp request */
- list_add(&req->rq_exp_list, &export->exp_hp_rpcs);
- else
- list_add(&req->rq_exp_list, &export->exp_reg_rpcs);
- spin_unlock_bh(&export->exp_rpc_lock);
+ * reference for request later
+ */
+ spin_lock(&export->exp_rpc_lock);
+ ptlrpc_add_exp_list_nolock(req, export, req->rq_ops != NULL);
+ spin_unlock(&export->exp_rpc_lock);
class_export_rpc_dec(req->rq_export);
class_export_put(req->rq_export);
/* request takes one export refcount */
req->rq_export = class_export_get(export);
class_export_rpc_inc(export);
-
- return;
}
/**
* This function is only called when some export receives a message (i.e.,
* the network is up.)
*/
-void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
+void ptlrpc_update_export_timer(struct obd_export *exp, time64_t extra_delay)
{
- struct obd_export *oldest_exp;
- time_t oldest_time, new_time;
+ struct obd_export *oldest_exp;
+ time64_t oldest_time, new_time;
- ENTRY;
+ ENTRY;
- LASSERT(exp);
+ LASSERT(exp);
- /* Compensate for slow machines, etc, by faking our request time
- into the future. Although this can break the strict time-ordering
- of the list, we can be really lazy here - we don't have to evict
- at the exact right moment. Eventually, all silent exports
- will make it to the top of the list. */
+ /*
+ * Compensate for slow machines, etc, by faking our request time
+ * into the future. Although this can break the strict time-ordering
+ * of the list, we can be really lazy here - we don't have to evict
+ * at the exact right moment. Eventually, all silent exports
+ * will make it to the top of the list.
+ */
- /* Do not pay attention on 1sec or smaller renewals. */
- new_time = cfs_time_current_sec() + extra_delay;
- if (exp->exp_last_request_time + 1 /*second */ >= new_time)
- RETURN_EXIT;
+ /* Do not pay attention on 1sec or smaller renewals. */
+ new_time = ktime_get_real_seconds() + extra_delay;
+ if (exp->exp_last_request_time + 1 /*second */ >= new_time)
+ RETURN_EXIT;
- exp->exp_last_request_time = new_time;
+ exp->exp_last_request_time = new_time;
- /* exports may get disconnected from the chain even though the
- export has references, so we must keep the spin lock while
- manipulating the lists */
+ /*
+ * exports may get disconnected from the chain even though the
+ * export has references, so we must keep the spin lock while
+ * manipulating the lists
+ */
spin_lock(&exp->exp_obd->obd_dev_lock);
if (list_empty(&exp->exp_obd_chain_timed)) {
oldest_time = oldest_exp->exp_last_request_time;
spin_unlock(&exp->exp_obd->obd_dev_lock);
- if (exp->exp_obd->obd_recovering) {
- /* be nice to everyone during recovery */
- EXIT;
- return;
- }
-
- /* Note - racing to start/reset the obd_eviction timer is safe */
- if (exp->exp_obd->obd_eviction_timer == 0) {
- /* Check if the oldest entry is expired. */
- if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
- extra_delay)) {
- /* We need a second timer, in case the net was down and
- * it just came back. Since the pinger may skip every
- * other PING_INTERVAL (see note in ptlrpc_pinger_main),
- * 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",
- exp->exp_obd->obd_name,
- obd_export_nid2str(oldest_exp), oldest_time);
- }
- } else {
- if (cfs_time_current_sec() >
- (exp->exp_obd->obd_eviction_timer + extra_delay)) {
- /* The evictor won't evict anyone who we've heard from
- * recently, so we don't have to check before we start
- * it. */
- if (!ping_evictor_wake(exp))
- exp->exp_obd->obd_eviction_timer = 0;
- }
- }
-
- EXIT;
+ if (exp->exp_obd->obd_recovering) {
+ /* be nice to everyone during recovery */
+ EXIT;
+ return;
+ }
+
+ /* Note - racing to start/reset the obd_eviction timer is safe */
+ if (exp->exp_obd->obd_eviction_timer == 0) {
+ /* Check if the oldest entry is expired. */
+ if (ktime_get_real_seconds() >
+ oldest_time + PING_EVICT_TIMEOUT + extra_delay) {
+ /*
+ * We need a second timer, in case the net was down and
+ * it just came back. Since the pinger may skip every
+ * other PING_INTERVAL (see note in ptlrpc_pinger_main),
+ * we better wait for 3.
+ */
+ exp->exp_obd->obd_eviction_timer =
+ ktime_get_real_seconds() + 3 * PING_INTERVAL;
+ CDEBUG(D_HA, "%s: Think about evicting %s from %lld\n",
+ exp->exp_obd->obd_name,
+ obd_export_nid2str(oldest_exp), oldest_time);
+ }
+ } else {
+ if (ktime_get_real_seconds() >
+ (exp->exp_obd->obd_eviction_timer + extra_delay)) {
+ /*
+ * The evictor won't evict anyone who we've heard from
+ * recently, so we don't have to check before we start
+ * it.
+ */
+ if (!ping_evictor_wake(exp))
+ exp->exp_obd->obd_eviction_timer = 0;
+ }
+ }
+
+ EXIT;
}
/**
return -EEXIST;
}
if (unlikely(obd == NULL || obd->obd_fail)) {
- /* Failing over, don't handle any more reqs,
- * send error response instead. */
+ /*
+ * Failing over, don't handle any more reqs,
+ * send error response instead.
+ */
CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
req, (obd != NULL) ? obd->obd_name : "unknown");
- rc = -ENODEV;
- } else if (lustre_msg_get_flags(req->rq_reqmsg) &
- (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
+ rc = -ENODEV;
+ } else if (lustre_msg_get_flags(req->rq_reqmsg) &
+ (MSG_REPLAY | MSG_REQ_REPLAY_DONE) &&
!obd->obd_recovering) {
- DEBUG_REQ(D_ERROR, req,
- "Invalid replay without recovery");
- class_fail_export(req->rq_export);
- rc = -ENODEV;
- } else if (lustre_msg_get_transno(req->rq_reqmsg) != 0 &&
+ DEBUG_REQ(D_ERROR, req,
+ "Invalid replay without recovery");
+ class_fail_export(req->rq_export);
+ rc = -ENODEV;
+ } 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",
- lustre_msg_get_transno(req->rq_reqmsg));
- class_fail_export(req->rq_export);
- rc = -ENODEV;
- }
+ DEBUG_REQ(D_ERROR, req,
+ "Invalid req with transno %llu without recovery",
+ lustre_msg_get_transno(req->rq_reqmsg));
+ class_fail_export(req->rq_export);
+ rc = -ENODEV;
+ }
- if (unlikely(rc < 0)) {
- req->rq_status = rc;
- ptlrpc_error(req);
- }
- return rc;
+ if (unlikely(rc < 0)) {
+ req->rq_status = rc;
+ ptlrpc_error(req);
+ }
+ return rc;
}
static void ptlrpc_at_set_timer(struct ptlrpc_service_part *svcpt)
{
struct ptlrpc_at_array *array = &svcpt->scp_at_array;
- __s32 next;
+ time64_t 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() -
- at_early_margin);
+ next = array->paa_deadline - ktime_get_real_seconds() -
+ at_early_margin;
if (next <= 0) {
- ptlrpc_at_timer((unsigned long)svcpt);
+ ptlrpc_at_timer(cfs_timer_cb_arg(svcpt, scp_at_timer));
} else {
- cfs_timer_arm(&svcpt->scp_at_timer, cfs_time_shift(next));
- CDEBUG(D_INFO, "armed %s at %+ds\n",
+ mod_timer(&svcpt->scp_at_timer,
+ jiffies + nsecs_to_jiffies(next * NSEC_PER_SEC));
+ CDEBUG(D_INFO, "armed %s at %+llds\n",
svcpt->scp_service->srv_name, next);
}
}
{
struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
struct ptlrpc_at_array *array = &svcpt->scp_at_array;
- struct ptlrpc_request *rq = NULL;
- __u32 index;
+ struct ptlrpc_request *rq = NULL;
+ __u32 index;
- if (AT_OFF)
- return(0);
+ if (AT_OFF)
+ return(0);
- if (req->rq_no_reply)
- return 0;
+ if (req->rq_no_reply)
+ return 0;
- if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
- return(-ENOSYS);
+ if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
+ return(-ENOSYS);
spin_lock(&svcpt->scp_at_lock);
LASSERT(list_empty(&req->rq_timed_list));
- index = (unsigned long)req->rq_deadline % array->paa_size;
- if (array->paa_reqs_count[index] > 0) {
- /* latest rpcs will have the latest deadlines in the list,
- * so search backward. */
- list_for_each_entry_reverse(rq,
- &array->paa_reqs_array[index],
- rq_timed_list) {
- if (req->rq_deadline >= rq->rq_deadline) {
+ 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.
+ */
+ list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
+ rq_timed_list) {
+ if (req->rq_deadline >= rq->rq_deadline) {
list_add(&req->rq_timed_list,
- &rq->rq_timed_list);
- break;
- }
- }
- }
+ &rq->rq_timed_list);
+ break;
+ }
+ }
+ }
- /* Add the request at the head of the list */
+ /* Add the request at the head of the list */
if (list_empty(&req->rq_timed_list))
- list_add(&req->rq_timed_list,
- &array->paa_reqs_array[index]);
+ list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
spin_lock(&req->rq_lock);
req->rq_at_linked = 1;
return 0;
}
-static void
-ptlrpc_at_remove_timed(struct ptlrpc_request *req)
+static void ptlrpc_at_remove_timed(struct ptlrpc_request *req)
{
struct ptlrpc_at_array *array;
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;
RETURN(1);
}
- /* 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 "
- "%d+%d", AT_OFF ? "AT off - not " : "",
- olddl, olddl - at_get(&svcpt->scp_at_estimate),
+ /*
+ * 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 %+llds, margin %+llds) for %d+%d",
+ AT_OFF ? "AT off - not " : "",
+ (s64)olddl, (s64)(olddl - at_get(&svcpt->scp_at_estimate)),
at_get(&svcpt->scp_at_estimate), at_extra);
- if (AT_OFF)
- RETURN(0);
+ if (AT_OFF)
+ RETURN(0);
- if (olddl < 0) {
- DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
- "not sending early reply. Consider increasing "
- "at_early_margin (%d)?", olddl, at_early_margin);
+ if (olddl < 0) {
+ /* below message is checked in replay-ost-single.sh test_9 */
+ DEBUG_REQ(D_WARNING, req,
+ "Already past deadline (%+llds), not sending early reply. Consider increasing at_early_margin (%d)?",
+ (s64)olddl, at_early_margin);
- /* Return an error so we're not re-added to the timed list. */
- RETURN(-ETIMEDOUT);
- }
+ /* Return an error so we're not re-added to the timed list. */
+ RETURN(-ETIMEDOUT);
+ }
- if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
- DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
- "but no AT support");
- RETURN(-ENOSYS);
- }
+ if ((lustre_msghdr_get_flags(req->rq_reqmsg) &
+ MSGHDR_AT_SUPPORT) == 0) {
+ DEBUG_REQ(D_INFO, req,
+ "Wanted to ask client for more time, but no AT support");
+ RETURN(-ENOSYS);
+ }
if (req->rq_export &&
lustre_msg_get_flags(req->rq_reqmsg) &
(MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
- /* During recovery, we don't want to send too many early
+ 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. */
- /* Don't account request processing time into AT history
+ * 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,
+ /*
+ * We want to extend the request deadline by at_extra seconds,
* so we set our service estimate to reflect how much time has
* passed since this request arrived plus an additional
* at_extra seconds. The client will calculate the new deadline
* 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 */
+ /*
+ * 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,
+ "Could not add any time (%lld/%lld), not sending early reply",
+ (s64)olddl, (s64)(newdl - ktime_get_real_seconds()));
RETURN(-ETIMEDOUT);
}
if (!reqmsg)
GOTO(out_free, rc = -ENOMEM);
- *reqcopy = *req;
- reqcopy->rq_reply_state = NULL;
- reqcopy->rq_rep_swab_mask = 0;
- reqcopy->rq_pack_bulk = 0;
- reqcopy->rq_pack_udesc = 0;
- reqcopy->rq_packed_final = 0;
- sptlrpc_svc_ctx_addref(reqcopy);
- /* We only need the reqmsg for the magic */
- reqcopy->rq_reqmsg = reqmsg;
- memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
+ *reqcopy = *req;
+ reqcopy->rq_reply_state = NULL;
+ reqcopy->rq_rep_swab_mask = 0;
+ reqcopy->rq_pack_bulk = 0;
+ reqcopy->rq_pack_udesc = 0;
+ reqcopy->rq_packed_final = 0;
+ sptlrpc_svc_ctx_addref(reqcopy);
+ /* We only need the reqmsg for the magic */
+ reqcopy->rq_reqmsg = reqmsg;
+ memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
/*
* tgt_brw_read() and tgt_brw_write() may have decided not to reply.
GOTO(out, rc = -ETIMEDOUT);
LASSERT(atomic_read(&req->rq_refcount));
- /** if it is last refcount then early reply isn't needed */
+ /* if it is last refcount then early reply isn't needed */
if (atomic_read(&req->rq_refcount) == 1) {
- DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
- "abort sending early reply\n");
+ DEBUG_REQ(D_ADAPTTO, reqcopy,
+ "Normal reply already sent, abort early reply");
GOTO(out, rc = -EINVAL);
}
- /* Connection ref */
- reqcopy->rq_export = class_conn2export(
- lustre_msg_get_handle(reqcopy->rq_reqmsg));
- if (reqcopy->rq_export == NULL)
- GOTO(out, rc = -ENODEV);
+ /* Connection ref */
+ reqcopy->rq_export = class_conn2export(
+ lustre_msg_get_handle(reqcopy->rq_reqmsg));
+ if (reqcopy->rq_export == NULL)
+ GOTO(out, rc = -ENODEV);
- /* RPC ref */
+ /* RPC ref */
class_export_rpc_inc(reqcopy->rq_export);
- if (reqcopy->rq_export->exp_obd &&
- reqcopy->rq_export->exp_obd->obd_fail)
- GOTO(out_put, rc = -ENODEV);
+ if (reqcopy->rq_export->exp_obd &&
+ reqcopy->rq_export->exp_obd->obd_fail)
+ GOTO(out_put, rc = -ENODEV);
- rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
- if (rc)
- GOTO(out_put, rc);
+ rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
+ if (rc)
+ GOTO(out_put, rc);
- rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
+ rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
if (!rc) {
/* Adjust our own deadline to what we told the client */
req->rq_deadline = newdl;
req->rq_early_count++; /* number sent, server side */
} else {
- DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
+ DEBUG_REQ(D_ERROR, req, "Early reply send failed: rc = %d", rc);
}
- /* Free the (early) reply state from lustre_pack_reply.
- (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
- ptlrpc_req_drop_rs(reqcopy);
+ /*
+ * Free the (early) reply state from lustre_pack_reply.
+ * (ptlrpc_send_reply takes it's own rs ref, so this is safe here)
+ */
+ ptlrpc_req_drop_rs(reqcopy);
out_put:
class_export_rpc_dec(reqcopy->rq_export);
RETURN(rc);
}
-/* Send early replies to everybody expiring within at_early_margin
- asking for at_extra time */
+/*
+ * Send early replies to everybody expiring within at_early_margin
+ * asking for at_extra time
+ */
static int ptlrpc_at_check_timed(struct ptlrpc_service_part *svcpt)
{
struct ptlrpc_at_array *array = &svcpt->scp_at_array;
- struct ptlrpc_request *rq, *n;
- struct list_head work_list;
- __u32 index, count;
- time_t deadline;
- time_t now = cfs_time_current_sec();
- cfs_duration_t delay;
- int first, counter = 0;
- ENTRY;
+ struct ptlrpc_request *rq, *n;
+ LIST_HEAD(work_list);
+ __u32 index, count;
+ time64_t deadline;
+ time64_t now = ktime_get_real_seconds();
+ s64 delay;
+ int first, counter = 0;
+ ENTRY;
spin_lock(&svcpt->scp_at_lock);
if (svcpt->scp_at_check == 0) {
spin_unlock(&svcpt->scp_at_lock);
RETURN(0);
}
- delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
+ delay = ktime_ms_delta(ktime_get(), svcpt->scp_at_checktime);
svcpt->scp_at_check = 0;
if (array->paa_count == 0) {
RETURN(0);
}
- /* We're close to a timeout, and we don't know how much longer the
- server will take. Send early replies to everyone expiring soon. */
- INIT_LIST_HEAD(&work_list);
+ /*
+ * We're close to a timeout, and we don't know how much longer the
+ * server will take. Send early replies to everyone expiring soon.
+ */
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];
break;
}
- ptlrpc_at_remove_timed(rq);
/**
* ptlrpc_server_drop_request() may drop
* refcount to 0 already. Let's check this and
* don't add entry to work_list
*/
- if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
+ if (likely(atomic_inc_not_zero(&rq->rq_refcount))) {
+ ptlrpc_at_remove_timed(rq);
list_add(&rq->rq_timed_list, &work_list);
+ } else {
+ ptlrpc_at_remove_timed(rq);
+ }
+
counter++;
}
spin_unlock(&svcpt->scp_at_lock);
- CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
- "replies\n", first, at_extra, counter);
- if (first < 0) {
- /* We're already past request deadlines before we even get a
- chance to send early replies */
- LCONSOLE_WARN("%s: This server is not able to keep up with "
- "request traffic (cpu-bound).\n",
+ CDEBUG(D_ADAPTTO,
+ "timeout in %+ds, asking for %d secs on %d early replies\n",
+ first, at_extra, counter);
+ if (first < 0) {
+ /*
+ * We're already past request deadlines before we even get a
+ * chance to send early replies
+ */
+ 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=%lld\n",
counter, svcpt->scp_nreqs_incoming,
svcpt->scp_nreqs_active,
at_get(&svcpt->scp_at_estimate), delay);
- }
+ }
- /* we took additional refcount so entries can't be deleted from list, no
- * locking is needed */
+ /*
+ * we took additional refcount so entries can't be deleted from list, no
+ * locking is needed
+ */
while (!list_empty(&work_list)) {
rq = list_entry(work_list.next, struct ptlrpc_request,
- rq_timed_list);
+ rq_timed_list);
list_del_init(&rq->rq_timed_list);
- if (ptlrpc_at_send_early_reply(rq) == 0)
- ptlrpc_at_add_timed(rq);
+ if (ptlrpc_at_send_early_reply(rq) == 0)
+ ptlrpc_at_add_timed(rq);
- ptlrpc_server_drop_request(rq);
- }
+ ptlrpc_server_drop_request(rq);
+ }
RETURN(1); /* return "did_something" for liblustre */
}
-/* Check if we are already handling earlier incarnation of this request.
- * Called under &req->rq_export->exp_rpc_lock locked */
-static int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
+/*
+ * Check if we are already handling earlier incarnation of this request.
+ * Called under &req->rq_export->exp_rpc_lock locked
+ */
+static struct ptlrpc_request*
+ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
{
- struct ptlrpc_request *tmp = NULL;
+ struct ptlrpc_request *tmp = NULL;
if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
(atomic_read(&req->rq_export->exp_rpc_count) == 0))
- return 0;
+ return NULL;
- /* bulk request are aborted upon reconnect, don't try to
- * find a match */
+ /*
+ * bulk request are aborted upon reconnect, don't try to
+ * find a match
+ */
if (req->rq_bulk_write || req->rq_bulk_read)
- return 0;
+ return NULL;
- /* This list should not be longer than max_requests in
+ /*
+ * This list should not be longer than max_requests in
* flights on the client, so it is not all that long.
* Also we only hit this codepath in case of a resent
- * request which makes it even more rarely hit */
+ * request which makes it even more rarely hit
+ */
list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs,
rq_exp_list) {
/* Found duplicate one */
if (tmp->rq_xid == req->rq_xid)
goto found;
}
- return 0;
+ return NULL;
found:
DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
DEBUG_REQ(D_HA, tmp, "Request being processed");
- return -EBUSY;
+ return tmp;
+}
+
+#ifdef HAVE_SERVER_SUPPORT
+static void ptlrpc_server_mark_obsolete(struct ptlrpc_request *req)
+{
+ req->rq_obsolete = 1;
}
+static void
+ptlrpc_server_mark_in_progress_obsolete(struct ptlrpc_request *req)
+{
+ struct ptlrpc_request *tmp = NULL;
+ __u16 tag;
+
+ if (!tgt_is_increasing_xid_client(req->rq_export) ||
+ req->rq_export->exp_used_slots == NULL)
+ return;
+
+ tag = lustre_msg_get_tag(req->rq_reqmsg);
+ if (tag == 0)
+ return;
+
+ if (!test_bit(tag - 1, req->rq_export->exp_used_slots))
+ return;
+
+ /* This list should not be longer than max_requests in
+ * flights on the client, so it is not all that long.
+ * Also we only hit this codepath in case of a resent
+ * request which makes it even more rarely hit */
+ list_for_each_entry(tmp, &req->rq_export->exp_reg_rpcs, rq_exp_list) {
+ if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
+ req->rq_xid > tmp->rq_xid)
+ ptlrpc_server_mark_obsolete(tmp);
+
+ }
+ list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs, rq_exp_list) {
+ if (tag == lustre_msg_get_tag(tmp->rq_reqmsg) &&
+ req->rq_xid > tmp->rq_xid)
+ ptlrpc_server_mark_obsolete(tmp);
+ }
+}
+#endif
+
/**
* Check if a request should be assigned with a high priority.
*
- * \retval < 0: error occurred
+ * \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)
{
- int rc;
- ENTRY;
+ int rc = 0;
+ ENTRY;
if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL) {
rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
if (rc < 0)
}
if (req->rq_export != NULL && req->rq_ops != NULL) {
- /* Perform request specific check. We should do this
+ /*
+ * 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. */
+ * list otherwise it may hit swab race at LU-1044.
+ */
if (req->rq_ops->hpreq_check != NULL) {
rc = req->rq_ops->hpreq_check(req);
- LASSERT(rc <= 1); /* can only return error, 0, or 1 */
+ 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);
}
}
{
ENTRY;
if (req->rq_export) {
- /* refresh lock timeout again so that client has more
- * room to send lock cancel RPC. */
+ /*
+ * refresh lock timeout again so that client has more
+ * room to send lock cancel RPC.
+ */
if (req->rq_ops && req->rq_ops->hpreq_fini)
req->rq_ops->hpreq_fini(req);
- spin_lock_bh(&req->rq_export->exp_rpc_lock);
- list_del_init(&req->rq_exp_list);
- spin_unlock_bh(&req->rq_export->exp_rpc_lock);
+ ptlrpc_del_exp_list(req);
}
EXIT;
}
{
int opc = lustre_msg_get_opc(req->rq_reqmsg);
- /* Check for export to let only reconnects for not yet evicted
- * export to become a HP rpc. */
+ /*
+ * Check for export to let only reconnects for not yet evicted
+ * export to become a HP rpc.
+ */
if ((req->rq_export != NULL) &&
(opc == OBD_PING || opc == MDS_CONNECT || opc == OST_CONNECT))
req->rq_ops = &ptlrpc_hpreq_common;
{
int rc;
bool hp;
+ struct ptlrpc_request *orig;
+
ENTRY;
rc = ptlrpc_server_hpreq_init(svcpt, req);
hp = rc > 0;
ptlrpc_nrs_req_initialize(svcpt, req, hp);
- if (req->rq_export != NULL) {
+ while (req->rq_export != NULL) {
struct obd_export *exp = req->rq_export;
- /* do search for duplicated xid and the adding to the list
- * atomically */
+ /*
+ * 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) {
+#ifdef HAVE_SERVER_SUPPORT
+ ptlrpc_server_mark_in_progress_obsolete(req);
+#endif
+ orig = ptlrpc_server_check_resend_in_progress(req);
+ if (orig && OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
+ spin_unlock_bh(&exp->exp_rpc_lock);
+
+ OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
+ msleep(4 * MSEC_PER_SEC);
+ continue;
+ }
+
+ if (orig && likely(atomic_inc_not_zero(&orig->rq_refcount))) {
+ bool linked;
+
spin_unlock_bh(&exp->exp_rpc_lock);
+ /*
+ * When the client resend request and the server has
+ * the previous copy of it, we need to update deadlines,
+ * to be sure that the client and the server have equal
+ * request deadlines.
+ */
+
+ spin_lock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
+ linked = orig->rq_at_linked;
+ if (likely(linked))
+ ptlrpc_at_remove_timed(orig);
+ spin_unlock(&orig->rq_rqbd->rqbd_svcpt->scp_at_lock);
+ orig->rq_deadline = req->rq_deadline;
+ if (likely(linked))
+ ptlrpc_at_add_timed(orig);
+ ptlrpc_server_drop_request(orig);
ptlrpc_nrs_req_finalize(req);
- RETURN(rc);
+
+ /* don't mark slot unused for resend in progress */
+ req->rq_obsolete = 1;
+
+ RETURN(-EBUSY);
}
- 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);
+ ptlrpc_add_exp_list_nolock(req, exp, hp || req->rq_ops != NULL);
+
spin_unlock_bh(&exp->exp_rpc_lock);
+ break;
}
- /* the current thread is not the processing thread for this request
+ /*
+ * 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. */
+ * Remove all relations between request and old thread.
+ */
req->rq_svc_thread->t_env->le_ses = NULL;
req->rq_svc_thread = NULL;
req->rq_session.lc_thread = NULL;
bool force)
{
int running = svcpt->scp_nthrs_running;
+
if (unlikely(svcpt->scp_service->srv_req_portal == MDS_REQUEST_PORTAL &&
CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))) {
/* leave just 1 thread for normal RPCs */
* \see ptlrpc_server_allow_normal
* \see ptlrpc_server_allow high
*/
-static inline bool
-ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, bool force)
+static inline
+bool ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt,
+ bool force)
{
return ptlrpc_server_high_pending(svcpt, force) ||
ptlrpc_server_normal_pending(svcpt, force);
ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
{
struct ptlrpc_request *req = NULL;
+
ENTRY;
spin_lock(&svcpt->scp_req_lock);
* All incoming requests pass through here before getting into
* ptlrpc_server_handle_req later on.
*/
-static int
-ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
- struct ptlrpc_thread *thread)
-{
- struct ptlrpc_service *svc = svcpt->scp_service;
- struct ptlrpc_request *req;
- __u32 deadline;
- int rc;
+static int ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt,
+ struct ptlrpc_thread *thread)
+{
+ struct ptlrpc_service *svc = svcpt->scp_service;
+ struct ptlrpc_request *req;
+ __u32 deadline;
+ int rc;
+
ENTRY;
spin_lock(&svcpt->scp_lock);
struct ptlrpc_request, rq_list);
list_del_init(&req->rq_list);
svcpt->scp_nreqs_incoming--;
- /* Consider this still a "queued" request as far as stats are
- * concerned */
+ /*
+ * Consider this still a "queued" request as far as stats are
+ * concerned
+ */
spin_unlock(&svcpt->scp_lock);
- /* go through security check/transform */
- rc = sptlrpc_svc_unwrap_request(req);
- switch (rc) {
- case SECSVC_OK:
- break;
- case SECSVC_COMPLETE:
- target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
- goto err_req;
- case SECSVC_DROP:
- goto err_req;
- default:
- LBUG();
- }
-
- /*
- * for null-flavored rpc, msg has been unpacked by sptlrpc, although
- * redo it wouldn't be harmful.
- */
- if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
- 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,
- 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,
- 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",
- cfs_fail_val, req->rq_xid);
- goto err_req;
- }
-
- rc = -EINVAL;
- if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
- CERROR("wrong packet type received (type=%u) from %s\n",
- lustre_msg_get_type(req->rq_reqmsg),
- libcfs_id2str(req->rq_peer));
- goto err_req;
- }
+ /* go through security check/transform */
+ rc = sptlrpc_svc_unwrap_request(req);
+ switch (rc) {
+ case SECSVC_OK:
+ break;
+ case SECSVC_COMPLETE:
+ target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
+ goto err_req;
+ case SECSVC_DROP:
+ goto err_req;
+ default:
+ LBUG();
+ }
+
+ /*
+ * for null-flavored rpc, msg has been unpacked by sptlrpc, although
+ * redo it wouldn't be harmful.
+ */
+ if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
+ rc = ptlrpc_unpack_req_msg(req, req->rq_reqlen);
+ if (rc != 0) {
+ CERROR("error unpacking request: ptl %d from %s 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 %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%llu\n",
+ cfs_fail_val, req->rq_xid);
+ goto err_req;
+ }
+
+ rc = -EINVAL;
+ if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
+ CERROR("wrong packet type received (type=%u) from %s\n",
+ lustre_msg_get_type(req->rq_reqmsg),
+ libcfs_id2str(req->rq_peer));
+ goto err_req;
+ }
switch (lustre_msg_get_opc(req->rq_reqmsg)) {
case MDS_WRITEPAGE:
break;
}
- CDEBUG(D_RPCTRACE, "got req x"LPU64"\n", req->rq_xid);
-
- req->rq_export = class_conn2export(
- lustre_msg_get_handle(req->rq_reqmsg));
- if (req->rq_export) {
- rc = ptlrpc_check_req(req);
- if (rc == 0) {
- rc = sptlrpc_target_export_check(req->rq_export, req);
- if (rc)
- DEBUG_REQ(D_ERROR, req, "DROPPING req with "
- "illegal security flavor,");
- }
-
- if (rc)
- goto err_req;
- ptlrpc_update_export_timer(req->rq_export, 0);
- }
-
- /* 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));
-
- /* Set rpc server deadline and add it to the timed list */
- deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
- MSGHDR_AT_SUPPORT) ?
- /* The max time the client expects us to take */
- lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
-
- req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
- if (unlikely(deadline == 0)) {
- DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
- goto err_req;
- }
+ CDEBUG(D_RPCTRACE, "got req x%llu\n", req->rq_xid);
+
+ req->rq_export = class_conn2export(
+ lustre_msg_get_handle(req->rq_reqmsg));
+ if (req->rq_export) {
+ rc = ptlrpc_check_req(req);
+ if (rc == 0) {
+ rc = sptlrpc_target_export_check(req->rq_export, req);
+ if (rc)
+ DEBUG_REQ(D_ERROR, req,
+ "DROPPING req with illegal security flavor");
+ }
+
+ if (rc)
+ goto err_req;
+ ptlrpc_update_export_timer(req->rq_export, 0);
+ }
+
+ /* req_in handling should/must be fast */
+ 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) &
+ MSGHDR_AT_SUPPORT) ?
+ /* The max time the client expects us to take */
+ lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
+
+ req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
+ if (unlikely(deadline == 0)) {
+ DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
+ goto err_req;
+ }
/* Skip early reply */
if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
req->rq_svc_thread = thread;
if (thread != NULL) {
- /* initialize request session, it is needed for request
- * processing by target */
+ /*
+ * initialize request session, it is needed for request
+ * processing by target
+ */
rc = lu_context_init(&req->rq_session, LCT_SERVER_SESSION |
LCT_NOREF);
if (rc) {
* Main incoming request handling logic.
* Calls handler function from service to do actual processing.
*/
-static int
-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;
+static int ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
+ struct ptlrpc_thread *thread)
+{
+ 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 (request == NULL)
RETURN(0);
- if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
- fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
- else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
- fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
+ fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
+ else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
+ fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
- if (unlikely(fail_opc)) {
+ if (unlikely(fail_opc)) {
if (request->rq_export && request->rq_ops)
OBD_FAIL_TIMEOUT(fail_opc, 4);
}
- ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
+ ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
- if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
+ 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);
- lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
+ lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
+ timediff_usecs);
+ lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
svcpt->scp_nreqs_incoming);
lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
svcpt->scp_nreqs_active);
lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
at_get(&svcpt->scp_at_estimate));
- }
+ }
if (likely(request->rq_export)) {
if (unlikely(ptlrpc_check_req(request)))
goto put_conn;
- ptlrpc_update_export_timer(request->rq_export, timediff >> 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",
- 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));
- 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(),
+ ptlrpc_update_export_timer(request->rq_export,
+ div_u64(timediff_usecs,
+ USEC_PER_SEC / 2));
+ }
+
+ /*
+ * Discard requests queued for longer than the deadline.
+ * The deadline is increased if we send an early reply.
+ */
+ if (ktime_get_real_seconds() > request->rq_deadline) {
+ DEBUG_REQ(D_ERROR, request,
+ "Dropping timed-out request from %s: deadline %lld/%llds ago",
+ libcfs_id2str(request->rq_peer),
+ request->rq_deadline -
+ request->rq_arrival_time.tv_sec,
+ ktime_get_real_seconds() - request->rq_deadline);
+ goto put_conn;
+ }
+
+ CDEBUG(D_RPCTRACE,
+ "Handling RPC req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s\n",
+ request, current_comm(),
(request->rq_export ?
(char *)request->rq_export->exp_client_uuid.uuid : "0"),
(request->rq_export ?
- atomic_read(&request->rq_export->exp_refcount) : -99),
+ refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
libcfs_id2str(request->rq_peer),
- lustre_msg_get_opc(request->rq_reqmsg));
+ lustre_msg_get_opc(request->rq_reqmsg),
+ lustre_msg_get_jobid(request->rq_reqmsg) ?: "");
- if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
- CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
+ 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",
- current_comm(),
- (request->rq_export ?
- (char *)request->rq_export->exp_client_uuid.uuid : "0"),
- (request->rq_export ?
- atomic_read(&request->rq_export->exp_refcount) : -99),
- lustre_msg_get_status(request->rq_reqmsg),
- 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),
- (request->rq_repmsg ?
- lustre_msg_get_transno(request->rq_repmsg) :
- request->rq_transno),
- request->rq_status,
- (request->rq_repmsg ?
- lustre_msg_get_status(request->rq_repmsg) : -999));
- if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
- __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
- int opc = opcode_offset(op);
- if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
- LASSERT(opc < LUSTRE_MAX_OPCODES);
- lprocfs_counter_add(svc->srv_stats,
- opc + EXTRA_MAX_OPCODES,
- timediff);
- }
- }
- if (unlikely(request->rq_early_count)) {
- DEBUG_REQ(D_ADAPTTO, request,
- "sent %d early replies before finishing in "
- CFS_DURATION_T"s",
- request->rq_early_count,
- cfs_time_sub(work_end.tv_sec,
- request->rq_arrival_time.tv_sec));
- }
+ 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 req@%p pname:cluuid+ref:pid:xid:nid:opc:job %s:%s+%d:%d:x%llu:%s:%d:%s Request processed in %lldus (%lldus total) trans %llu rc %d/%d\n",
+ request, current_comm(),
+ (request->rq_export ?
+ (char *)request->rq_export->exp_client_uuid.uuid : "0"),
+ (request->rq_export ?
+ refcount_read(&request->rq_export->exp_handle.h_ref) : -99),
+ lustre_msg_get_status(request->rq_reqmsg),
+ request->rq_xid,
+ libcfs_id2str(request->rq_peer),
+ lustre_msg_get_opc(request->rq_reqmsg),
+ lustre_msg_get_jobid(request->rq_reqmsg) ?: "",
+ timediff_usecs,
+ arrived_usecs,
+ (request->rq_repmsg ?
+ lustre_msg_get_transno(request->rq_repmsg) :
+ request->rq_transno),
+ request->rq_status,
+ (request->rq_repmsg ?
+ lustre_msg_get_status(request->rq_repmsg) : -999));
+ if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
+ __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
+ int opc = opcode_offset(op);
+
+ if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
+ LASSERT(opc < LUSTRE_MAX_OPCODES);
+ lprocfs_counter_add(svc->srv_stats,
+ opc + EXTRA_MAX_OPCODES,
+ timediff_usecs);
+ }
+ }
+ if (unlikely(request->rq_early_count)) {
+ DEBUG_REQ(D_ADAPTTO, request,
+ "sent %d early replies before finishing in %llds",
+ request->rq_early_count,
+ div_u64(arrived_usecs, USEC_PER_SEC));
+ }
ptlrpc_server_finish_active_request(svcpt, request);
/**
* An internal function to process a single reply state object.
*/
-static int
-ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
+static int ptlrpc_handle_rs(struct ptlrpc_reply_state *rs)
{
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 ptlrpc_service *svc = svcpt->scp_service;
+ 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_mode_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
- " o%d NID %s\n",
- rs,
- rs->rs_xid, rs->rs_transno, rs->rs_opc,
+ 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 */
spin_unlock(&rs->rs_lock);
- class_export_put (exp);
+ class_export_put(exp);
rs->rs_export = NULL;
ptlrpc_rs_decref(rs);
if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
}
-static void
-ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
+static void ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
{
int avail = svcpt->scp_nrqbds_posted;
int low_water = test_req_buffer_pressure ? 0 :
svcpt->scp_service->srv_nbuf_per_group / 2;
- /* NB I'm not locking; just looking. */
+ /* NB I'm not locking; just looking. */
- /* CAVEAT EMPTOR: We might be allocating buffers here because we've
- * allowed the request history to grow out of control. We could put a
- * sanity check on that here and cull some history if we need the
- * space. */
+ /*
+ * CAVEAT EMPTOR: We might be allocating buffers here because we've
+ * allowed the request history to grow out of control. We could put a
+ * sanity check on that here and cull some history if we need the
+ * space.
+ */
- if (avail <= low_water)
+ if (avail <= low_water)
ptlrpc_grow_req_bufs(svcpt, 1);
if (svcpt->scp_service->srv_stats) {
}
}
-static int
-ptlrpc_retry_rqbds(void *arg)
-{
- struct ptlrpc_service_part *svcpt = (struct ptlrpc_service_part *)arg;
-
- svcpt->scp_rqbd_timeout = 0;
- return -ETIMEDOUT;
-}
-
-static inline int
-ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
+static inline int ptlrpc_threads_enough(struct ptlrpc_service_part *svcpt)
{
return svcpt->scp_nreqs_active <
svcpt->scp_nthrs_running - 1 -
* user can call it w/o any lock but need to hold
* ptlrpc_service_part::scp_lock to get reliable result
*/
-static inline int
-ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
+static inline int ptlrpc_threads_increasable(struct ptlrpc_service_part *svcpt)
{
return svcpt->scp_nthrs_running +
svcpt->scp_nthrs_starting <
/**
* too many requests and allowed to create more threads
*/
-static inline int
-ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
+static inline int ptlrpc_threads_need_create(struct ptlrpc_service_part *svcpt)
{
return !ptlrpc_threads_enough(svcpt) &&
ptlrpc_threads_increasable(svcpt);
}
-static inline int
-ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
+static inline int ptlrpc_thread_stopping(struct ptlrpc_thread *thread)
{
return thread_is_stopping(thread) ||
thread->t_svcpt->scp_service->srv_is_stopping;
}
-static inline int
-ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
+/* stop the highest numbered thread if there are too many threads running */
+static inline bool ptlrpc_thread_should_stop(struct ptlrpc_thread *thread)
+{
+ struct ptlrpc_service_part *svcpt = thread->t_svcpt;
+
+ return thread->t_id >= svcpt->scp_service->srv_nthrs_cpt_limit &&
+ thread->t_id == svcpt->scp_thr_nextid - 1;
+}
+
+static void ptlrpc_stop_thread(struct ptlrpc_thread *thread)
+{
+ CDEBUG(D_INFO, "Stopping thread %s #%u\n",
+ thread->t_svcpt->scp_service->srv_thread_name, thread->t_id);
+ thread_add_flags(thread, SVC_STOPPING);
+}
+
+static inline void ptlrpc_thread_stop(struct ptlrpc_thread *thread)
+{
+ struct ptlrpc_service_part *svcpt = thread->t_svcpt;
+
+ spin_lock(&svcpt->scp_lock);
+ if (ptlrpc_thread_should_stop(thread)) {
+ ptlrpc_stop_thread(thread);
+ svcpt->scp_thr_nextid--;
+ }
+ spin_unlock(&svcpt->scp_lock);
+}
+
+static inline int ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
{
return !list_empty(&svcpt->scp_rqbd_idle) &&
svcpt->scp_rqbd_timeout == 0;
return svcpt->scp_at_check;
}
+/*
+ * If a thread runs too long or spends to much time on a single request,
+ * we want to know about it, so we set up a delayed work item as a watchdog.
+ * If it fires, we display a stack trace of the delayed thread,
+ * providing we aren't rate-limited
+ *
+ * Watchdog stack traces are limited to 3 per 'libcfs_watchdog_ratelimit'
+ * seconds
+ */
+static struct ratelimit_state watchdog_limit;
+
+static void ptlrpc_watchdog_fire(struct work_struct *w)
+{
+ struct ptlrpc_thread *thread = container_of(w, struct ptlrpc_thread,
+ t_watchdog.work);
+ u64 ms_lapse = ktime_ms_delta(ktime_get(), thread->t_touched);
+ u32 ms_frac = do_div(ms_lapse, MSEC_PER_SEC);
+
+ /* ___ratelimit() returns true if the action is NOT ratelimited */
+ if (__ratelimit(&watchdog_limit)) {
+ /* below message is checked in sanity-quota.sh test_6,18 */
+ LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. The thread might be hung, or it might only be slow and will resume later. Dumping the stack trace for debugging purposes:\n",
+ thread->t_task->comm, thread->t_task->pid,
+ ms_lapse, ms_frac);
+
+ libcfs_debug_dumpstack(thread->t_task);
+ } else {
+ /* below message is checked in sanity-quota.sh test_6,18 */
+ LCONSOLE_WARN("%s: service thread pid %u was inactive for %llu.%03u seconds. Watchdog stack traces are limited to 3 per %u seconds, skipping this one.\n",
+ thread->t_task->comm, thread->t_task->pid,
+ ms_lapse, ms_frac, libcfs_watchdog_ratelimit);
+ }
+}
+
+static void ptlrpc_watchdog_init(struct delayed_work *work, time_t time)
+{
+ INIT_DELAYED_WORK(work, ptlrpc_watchdog_fire);
+ schedule_delayed_work(work, cfs_time_seconds(time));
+}
+
+static void ptlrpc_watchdog_disable(struct delayed_work *work)
+{
+ cancel_delayed_work_sync(work);
+}
+
+static void ptlrpc_watchdog_touch(struct delayed_work *work, time_t time)
+{
+ struct ptlrpc_thread *thread = container_of(&work->work,
+ struct ptlrpc_thread,
+ t_watchdog.work);
+ thread->t_touched = ktime_get();
+ mod_delayed_work(system_wq, work, cfs_time_seconds(time));
+}
+
/**
* requests wait on preprocessing
* user can call it w/o any lock but need to hold
ptlrpc_wait_event(struct ptlrpc_service_part *svcpt,
struct ptlrpc_thread *thread)
{
- /* Don't exit while there are replies to be handled */
- struct l_wait_info lwi = LWI_TIMEOUT(svcpt->scp_rqbd_timeout,
- ptlrpc_retry_rqbds, svcpt);
-
- lc_watchdog_disable(thread->t_watchdog);
+ ptlrpc_watchdog_disable(&thread->t_watchdog);
cond_resched();
- l_wait_event_exclusive_head(svcpt->scp_waitq,
- ptlrpc_thread_stopping(thread) ||
- ptlrpc_server_request_incoming(svcpt) ||
- ptlrpc_server_request_pending(svcpt, false) ||
- ptlrpc_rqbd_pending(svcpt) ||
- ptlrpc_at_check(svcpt), &lwi);
+ if (svcpt->scp_rqbd_timeout == 0)
+ /* Don't exit while there are replies to be handled */
+ wait_event_idle_exclusive_lifo(
+ svcpt->scp_waitq,
+ ptlrpc_thread_stopping(thread) ||
+ ptlrpc_server_request_incoming(svcpt) ||
+ ptlrpc_server_request_pending(svcpt, false) ||
+ ptlrpc_rqbd_pending(svcpt) ||
+ ptlrpc_at_check(svcpt));
+ else if (wait_event_idle_exclusive_lifo_timeout(
+ svcpt->scp_waitq,
+ ptlrpc_thread_stopping(thread) ||
+ ptlrpc_server_request_incoming(svcpt) ||
+ ptlrpc_server_request_pending(svcpt, false) ||
+ ptlrpc_rqbd_pending(svcpt) ||
+ ptlrpc_at_check(svcpt),
+ svcpt->scp_rqbd_timeout) == 0)
+ svcpt->scp_rqbd_timeout = 0;
if (ptlrpc_thread_stopping(thread))
return -EINTR;
- lc_watchdog_touch(thread->t_watchdog,
- ptlrpc_server_get_timeout(svcpt));
+ ptlrpc_watchdog_touch(&thread->t_watchdog,
+ ptlrpc_server_get_timeout(svcpt));
return 0;
}
*/
static int ptlrpc_main(void *arg)
{
- struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
- struct ptlrpc_service_part *svcpt = thread->t_svcpt;
- struct ptlrpc_service *svc = svcpt->scp_service;
- struct ptlrpc_reply_state *rs;
+ struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
+ struct ptlrpc_service_part *svcpt = thread->t_svcpt;
+ struct ptlrpc_service *svc = svcpt->scp_service;
+ struct ptlrpc_reply_state *rs;
struct group_info *ginfo = NULL;
struct lu_env *env;
int counter = 0, rc = 0;
+
ENTRY;
+ thread->t_task = current;
thread->t_pid = current_pid();
unshare_fs_struct();
- /* NB: we will call cfs_cpt_bind() for all threads, because we
- * might want to run lustre server only on a subset of system CPUs,
- * in that case ->scp_cpt is CFS_CPT_ANY */
- rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
- if (rc != 0) {
- CWARN("%s: failed to bind %s on CPT %d\n",
- svc->srv_name, thread->t_name, svcpt->scp_cpt);
+ if (svc->srv_cpt_bind) {
+ rc = cfs_cpt_bind(svc->srv_cptable, svcpt->scp_cpt);
+ if (rc != 0) {
+ CWARN("%s: failed to bind %s on CPT %d\n",
+ svc->srv_name, thread->t_name, svcpt->scp_cpt);
+ }
}
ginfo = groups_alloc(0);
- if (!ginfo) {
- rc = -ENOMEM;
- goto out;
- }
+ if (!ginfo)
+ GOTO(out, rc = -ENOMEM);
set_current_groups(ginfo);
put_group_info(ginfo);
if (svc->srv_ops.so_thr_init != NULL) {
rc = svc->srv_ops.so_thr_init(thread);
- if (rc)
- goto out;
- }
+ if (rc)
+ GOTO(out, rc);
+ }
- OBD_ALLOC_PTR(env);
- if (env == NULL) {
- rc = -ENOMEM;
- goto out_srv_fini;
- }
+ OBD_ALLOC_PTR(env);
+ if (env == NULL)
+ GOTO(out_srv_fini, rc = -ENOMEM);
+ rc = lu_env_add(env);
+ if (rc)
+ GOTO(out_env, rc);
- rc = lu_context_init(&env->le_ctx,
- svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
- if (rc)
- goto out_srv_fini;
+ rc = lu_context_init(&env->le_ctx,
+ svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
+ if (rc)
+ GOTO(out_env_remove, rc);
- thread->t_env = env;
- env->le_ctx.lc_thread = thread;
- env->le_ctx.lc_cookie = 0x6;
+ thread->t_env = env;
+ env->le_ctx.lc_thread = thread;
+ env->le_ctx.lc_cookie = 0x6;
while (!list_empty(&svcpt->scp_rqbd_idle)) {
rc = ptlrpc_server_post_idle_rqbds(svcpt);
CERROR("Failed to post rqbd for %s on CPT %d: %d\n",
svc->srv_name, svcpt->scp_cpt, rc);
- goto out_srv_fini;
+ GOTO(out_ctx_fini, rc);
}
- /* 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)
+ GOTO(out_ctx_fini, rc = -ENOMEM);
spin_lock(&svcpt->scp_lock);
LASSERT(svcpt->scp_nthrs_starting == 1);
svcpt->scp_nthrs_starting--;
- /* SVC_STOPPING may already be set here if someone else is trying
+ /*
+ * SVC_STOPPING may already be set here if someone else is trying
* to stop the service while this new thread has been dynamically
* forked. We still set SVC_RUNNING to let our creator know that
- * we are now running, however we will exit as soon as possible */
+ * we are now running, however we will exit as soon as possible
+ */
thread_add_flags(thread, SVC_RUNNING);
svcpt->scp_nthrs_running++;
spin_unlock(&svcpt->scp_lock);
/* wake up our creator in case he's still waiting. */
wake_up(&thread->t_ctl_waitq);
- thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
- NULL, NULL);
+ thread->t_touched = ktime_get();
+ ptlrpc_watchdog_init(&thread->t_watchdog,
+ ptlrpc_server_get_timeout(svcpt));
spin_lock(&svcpt->scp_rep_lock);
list_add(&rs->rs_list, &svcpt->scp_rep_idle);
if (ptlrpc_threads_need_create(svcpt)) {
/* Ignore return code - we tried... */
ptlrpc_start_thread(svcpt, 0);
- }
+ }
/* reset le_ses to initial state */
env->le_ses = NULL;
+ /* Refill the context before execution to make sure
+ * all thread keys are allocated */
+ lu_env_refill(env);
/* Process all incoming reqs before handling any */
if (ptlrpc_server_request_incoming(svcpt)) {
lu_context_enter(&env->le_ctx);
lu_context_enter(&env->le_ctx);
ptlrpc_server_handle_request(svcpt, thread);
lu_context_exit(&env->le_ctx);
- }
+ }
if (ptlrpc_rqbd_pending(svcpt) &&
ptlrpc_server_post_idle_rqbds(svcpt) < 0) {
- /* I just failed to repost request buffers.
+ /*
+ * I just failed to repost request buffers.
* Wait for a timeout (unless something else
- * happens) before I try again */
+ * happens) before I try again
+ */
svcpt->scp_rqbd_timeout = cfs_time_seconds(1) / 10;
CDEBUG(D_RPCTRACE, "Posted buffers: %d\n",
svcpt->scp_nrqbds_posted);
- }
- }
+ }
+ /*
+ * If the number of threads has been tuned downward and this
+ * thread should be stopped, then stop in reverse order so the
+ * the threads always have contiguous thread index values.
+ */
+ if (unlikely(ptlrpc_thread_should_stop(thread)))
+ ptlrpc_thread_stop(thread);
+ }
- lc_watchdog_delete(thread->t_watchdog);
- thread->t_watchdog = NULL;
+ ptlrpc_watchdog_disable(&thread->t_watchdog);
+out_ctx_fini:
+ lu_context_fini(&env->le_ctx);
+out_env_remove:
+ lu_env_remove(env);
+out_env:
+ OBD_FREE_PTR(env);
out_srv_fini:
- /*
- * deconstruct service specific state created by ptlrpc_start_thread()
- */
+ /* deconstruct service thread state created by ptlrpc_start_thread() */
if (svc->srv_ops.so_thr_done != NULL)
svc->srv_ops.so_thr_done(thread);
-
- if (env != NULL) {
- lu_context_fini(&env->le_ctx);
- OBD_FREE_PTR(env);
- }
out:
- CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
- thread, thread->t_pid, thread->t_id, rc);
-
+ CDEBUG(D_RPCTRACE, "%s: service thread [%p:%u] %d exiting: rc = %d\n",
+ thread->t_name, thread, thread->t_pid, thread->t_id, rc);
spin_lock(&svcpt->scp_lock);
if (thread_test_and_clear_flags(thread, SVC_STARTING))
svcpt->scp_nthrs_starting--;
*/
static int ptlrpc_hr_main(void *arg)
{
- 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;
+ struct ptlrpc_hr_thread *hrt = (struct ptlrpc_hr_thread *)arg;
+ struct ptlrpc_hr_partition *hrp = hrt->hrt_partition;
+ LIST_HEAD(replies);
+ struct lu_env *env;
+ int rc;
+
+ OBD_ALLOC_PTR(env);
+ if (env == NULL)
+ RETURN(-ENOMEM);
- 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);
}
+ rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD |
+ LCT_REMEMBER | LCT_NOREF);
+ if (rc)
+ GOTO(out_env, rc);
+
+ rc = lu_env_add(env);
+ if (rc)
+ GOTO(out_ctx_fini, rc);
+
atomic_inc(&hrp->hrp_nstarted);
wake_up(&ptlrpc_hr.hr_waitq);
struct ptlrpc_reply_state,
rs_list);
list_del_init(&rs->rs_list);
+ /* refill keys if needed */
+ lu_env_refill(env);
+ lu_context_enter(&env->le_ctx);
ptlrpc_handle_rs(rs);
+ lu_context_exit(&env->le_ctx);
}
}
atomic_inc(&hrp->hrp_nstopped);
wake_up(&ptlrpc_hr.hr_waitq);
+ lu_env_remove(env);
+out_ctx_fini:
+ lu_context_fini(&env->le_ctx);
+out_env:
+ OBD_FREE_PTR(env);
return 0;
}
static void ptlrpc_stop_hr_threads(void)
{
- struct ptlrpc_hr_partition *hrp;
- int i;
- int j;
+ struct ptlrpc_hr_partition *hrp;
+ int i;
+ int j;
ptlrpc_hr.hr_stopping = 1;
static int ptlrpc_start_hr_threads(void)
{
- struct ptlrpc_hr_partition *hrp;
- int i;
- int j;
+ struct ptlrpc_hr_partition *hrp;
+ int i;
+ int j;
+
ENTRY;
cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
atomic_read(&hrp->hrp_nstarted) == j);
if (rc < 0) {
- CERROR("cannot start reply handler thread %d:%d: "
- "rc = %d\n", i, j, rc);
+ CERROR("cannot start reply handler thread %d:%d: rc = %d\n",
+ i, j, rc);
ptlrpc_stop_hr_threads();
RETURN(rc);
}
static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
{
- struct l_wait_info lwi = { 0 };
- struct ptlrpc_thread *thread;
- struct list_head zombie;
+ struct ptlrpc_thread *thread;
+ LIST_HEAD(zombie);
ENTRY;
CDEBUG(D_INFO, "Stopping threads for service %s\n",
svcpt->scp_service->srv_name);
- INIT_LIST_HEAD(&zombie);
spin_lock(&svcpt->scp_lock);
/* let the thread know that we would like it to stop asap */
- list_for_each_entry(thread, &svcpt->scp_threads, t_link) {
- CDEBUG(D_INFO, "Stopping thread %s #%u\n",
- svcpt->scp_service->srv_thread_name, thread->t_id);
- thread_add_flags(thread, SVC_STOPPING);
- }
+ list_for_each_entry(thread, &svcpt->scp_threads, t_link)
+ ptlrpc_stop_thread(thread);
wake_up_all(&svcpt->scp_waitq);
thread = list_entry(svcpt->scp_threads.next,
struct ptlrpc_thread, t_link);
if (thread_is_stopped(thread)) {
- list_del(&thread->t_link);
- list_add(&thread->t_link, &zombie);
+ list_move(&thread->t_link, &zombie);
continue;
}
spin_unlock(&svcpt->scp_lock);
CDEBUG(D_INFO, "waiting for stopping-thread %s #%u\n",
svcpt->scp_service->srv_thread_name, thread->t_id);
- l_wait_event(thread->t_ctl_waitq,
- thread_is_stopped(thread), &lwi);
+ wait_event_idle(thread->t_ctl_waitq,
+ thread_is_stopped(thread));
spin_lock(&svcpt->scp_lock);
}
void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
{
struct ptlrpc_service_part *svcpt;
- int i;
+ int i;
+
ENTRY;
ptlrpc_service_for_each_part(svcpt, i, svc) {
int ptlrpc_start_threads(struct ptlrpc_service *svc)
{
- int rc = 0;
- int i;
- int j;
+ int rc = 0;
+ int i;
+ int j;
+
ENTRY;
/* We require 2 threads min, see note in ptlrpc_server_handle_request */
int ptlrpc_start_thread(struct ptlrpc_service_part *svcpt, int wait)
{
- struct l_wait_info lwi = { 0 };
- struct ptlrpc_thread *thread;
- struct ptlrpc_service *svc;
- struct task_struct *task;
- int rc;
+ struct ptlrpc_thread *thread;
+ struct ptlrpc_service *svc;
+ struct task_struct *task;
+ int rc;
+
ENTRY;
LASSERT(svcpt != NULL);
}
if (svcpt->scp_nthrs_starting != 0) {
- /* serialize starting because some modules (obdfilter)
- * might require unique and contiguous t_id */
+ /*
+ * serialize starting because some modules (obdfilter)
+ * might require unique and contiguous t_id
+ */
LASSERT(svcpt->scp_nthrs_starting == 1);
spin_unlock(&svcpt->scp_lock);
OBD_FREE_PTR(thread);
spin_lock(&svcpt->scp_lock);
--svcpt->scp_nthrs_starting;
if (thread_is_stopping(thread)) {
- /* this ptlrpc_thread is being hanled
+ /*
+ * this ptlrpc_thread is being hanled
* by ptlrpc_svcpt_stop_threads now
*/
thread_add_flags(thread, SVC_STOPPED);
spin_unlock(&svcpt->scp_lock);
OBD_FREE_PTR(thread);
}
- RETURN(rc);
- }
+ RETURN(rc);
+ }
if (!wait)
RETURN(0);
- l_wait_event(thread->t_ctl_waitq,
- thread_is_running(thread) || thread_is_stopped(thread),
- &lwi);
+ wait_event_idle(thread->t_ctl_waitq,
+ thread_is_running(thread) || thread_is_stopped(thread));
- rc = thread_is_stopped(thread) ? thread->t_id : 0;
- RETURN(rc);
+ rc = thread_is_stopped(thread) ? thread->t_id : 0;
+ RETURN(rc);
}
int ptlrpc_hr_init(void)
{
- struct ptlrpc_hr_partition *hrp;
- struct ptlrpc_hr_thread *hrt;
- int rc;
- int i;
- int j;
- int weight;
+ struct ptlrpc_hr_partition *hrp;
+ struct ptlrpc_hr_thread *hrt;
+ int rc;
+ int cpt;
+ int i;
+ int weight;
+
ENTRY;
memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
if (ptlrpc_hr.hr_partitions == NULL)
RETURN(-ENOMEM);
+ ratelimit_state_init(&watchdog_limit,
+ cfs_time_seconds(libcfs_watchdog_ratelimit), 3);
+
init_waitqueue_head(&ptlrpc_hr.hr_waitq);
- weight = cfs_cpu_ht_nsiblings(0);
+ preempt_disable();
+ weight = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
+ preempt_enable();
- 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;
+ struct ptlrpc_hr_partition *hrp;
+ 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]));
static void ptlrpc_wait_replies(struct ptlrpc_service_part *svcpt)
{
while (1) {
- int rc;
- struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10),
- NULL, NULL);
-
- rc = l_wait_event(svcpt->scp_waitq,
- atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
- if (rc == 0)
+ if (wait_event_idle_timeout(
+ svcpt->scp_waitq,
+ atomic_read(&svcpt->scp_nreps_difficult) == 0,
+ cfs_time_seconds(10)) > 0)
break;
CWARN("Unexpectedly long timeout %s %p\n",
svcpt->scp_service->srv_name, svcpt->scp_service);
static void
ptlrpc_service_del_atimer(struct ptlrpc_service *svc)
{
- struct ptlrpc_service_part *svcpt;
- int i;
+ struct ptlrpc_service_part *svcpt;
+ int i;
/* 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);
}
}
static void
ptlrpc_service_unlink_rqbd(struct ptlrpc_service *svc)
{
- struct ptlrpc_service_part *svcpt;
+ struct ptlrpc_service_part *svcpt;
struct ptlrpc_request_buffer_desc *rqbd;
- struct l_wait_info lwi;
- int rc;
- int i;
+ int rc;
+ int i;
- /* All history will be culled when the next request buffer is
- * freed in ptlrpc_service_purge_all() */
+ /*
+ * All history will be culled when the next request buffer is
+ * freed in ptlrpc_service_purge_all()
+ */
svc->srv_hist_nrqbds_cpt_max = 0;
rc = LNetClearLazyPortal(svc->srv_req_portal);
if (svcpt->scp_service == NULL)
break;
- /* Unlink all the request buffers. This forces a 'final'
- * event with its 'unlink' flag set for each posted rqbd */
+ /*
+ * Unlink all the request buffers. This forces a 'final'
+ * event with its 'unlink' flag set for each posted rqbd
+ */
list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
rqbd_list) {
rc = LNetMDUnlink(rqbd->rqbd_md_h);
if (svcpt->scp_service == NULL)
break;
- /* Wait for the network to release any buffers
- * it's currently filling */
+ /*
+ * Wait for the network to release any buffers
+ * it's currently filling
+ */
spin_lock(&svcpt->scp_lock);
while (svcpt->scp_nrqbds_posted != 0) {
+ int seconds = LONG_UNLINK;
+
spin_unlock(&svcpt->scp_lock);
- /* Network access will complete in finite time but
+ /*
+ * Network access will complete in finite time but
* the HUGE timeout lets us CWARN for visibility
- * of sluggish NALs */
- lwi = LWI_TIMEOUT_INTERVAL(
- cfs_time_seconds(LONG_UNLINK),
- cfs_time_seconds(1), NULL, NULL);
- rc = l_wait_event(svcpt->scp_waitq,
- svcpt->scp_nrqbds_posted == 0, &lwi);
- if (rc == -ETIMEDOUT) {
- CWARN("Service %s waiting for "
- "request buffers\n",
+ * of sluggish NALs
+ */
+ while (seconds > 0 &&
+ wait_event_idle_timeout(
+ svcpt->scp_waitq,
+ svcpt->scp_nrqbds_posted == 0,
+ cfs_time_seconds(1)) == 0)
+ seconds -= 1;
+ if (seconds == 0) {
+ CWARN("Service %s waiting for request buffers\n",
svcpt->scp_service->srv_name);
}
spin_lock(&svcpt->scp_lock);
static void
ptlrpc_service_purge_all(struct ptlrpc_service *svc)
{
- struct ptlrpc_service_part *svcpt;
- struct ptlrpc_request_buffer_desc *rqbd;
- struct ptlrpc_request *req;
- struct ptlrpc_reply_state *rs;
- int i;
+ struct ptlrpc_service_part *svcpt;
+ struct ptlrpc_request_buffer_desc *rqbd;
+ struct ptlrpc_request *req;
+ struct ptlrpc_reply_state *rs;
+ int i;
ptlrpc_service_for_each_part(svcpt, i, svc) {
if (svcpt->scp_service == NULL)
}
spin_unlock(&svcpt->scp_rep_lock);
- /* purge the request queue. NB No new replies (rqbds
+ /*
+ * purge the request queue. NB No new replies (rqbds
* all unlinked) and no service threads, so I'm the only
- * thread noodling the request queue now */
+ * thread noodling the request queue now
+ */
while (!list_empty(&svcpt->scp_req_incoming)) {
req = list_entry(svcpt->scp_req_incoming.next,
struct ptlrpc_request, rq_list);
LASSERT(list_empty(&svcpt->scp_rqbd_posted));
LASSERT(svcpt->scp_nreqs_incoming == 0);
LASSERT(svcpt->scp_nreqs_active == 0);
- /* history should have been culled by
- * ptlrpc_server_finish_request */
+ /*
+ * history should have been culled by
+ * ptlrpc_server_finish_request
+ */
LASSERT(svcpt->scp_hist_nrqbds == 0);
- /* Now free all the request buffers since nothing
- * references them any more... */
+ /*
+ * Now free all the request buffers since nothing
+ * references them any more...
+ */
while (!list_empty(&svcpt->scp_rqbd_idle)) {
rqbd = list_entry(svcpt->scp_rqbd_idle.next,
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);
*
* Right now, it just checks to make sure that requests aren't languishing
* in the queue. We'll use this health check to govern whether a node needs
- * to be shot, so it's intentionally non-aggressive. */
+ * 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;
}