* Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2012, Whamcloud, Inc.
+ * Copyright (c) 2010, 2014, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*/
#define DEBUG_SUBSYSTEM S_RPC
-#ifndef __KERNEL__
-#include <liblustre.h>
-#endif
#include <obd_support.h>
#include <obd_class.h>
#include <lustre_net.h>
/* forward ref */
static int ptlrpc_server_post_idle_rqbds(struct ptlrpc_service_part *svcpt);
-static void ptlrpc_hpreq_fini(struct ptlrpc_request *req);
+static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req);
+static void ptlrpc_at_remove_timed(struct ptlrpc_request *req);
-static CFS_LIST_HEAD(ptlrpc_all_services);
-cfs_spinlock_t ptlrpc_all_services_lock;
+/** Holds a list of all PTLRPC services */
+struct list_head ptlrpc_all_services;
+/** Used to protect the \e ptlrpc_all_services list */
+struct mutex ptlrpc_all_services_mutex;
-struct ptlrpc_request_buffer_desc *
+static struct ptlrpc_request_buffer_desc *
ptlrpc_alloc_rqbd(struct ptlrpc_service_part *svcpt)
{
struct ptlrpc_service *svc = svcpt->scp_service;
rqbd->rqbd_refcount = 0;
rqbd->rqbd_cbid.cbid_fn = request_in_callback;
rqbd->rqbd_cbid.cbid_arg = rqbd;
- CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
+ INIT_LIST_HEAD(&rqbd->rqbd_reqs);
OBD_CPT_ALLOC_LARGE(rqbd->rqbd_buffer, svc->srv_cptable,
svcpt->scp_cpt, svc->srv_buf_size);
if (rqbd->rqbd_buffer == NULL) {
return NULL;
}
- cfs_spin_lock(&svcpt->scp_lock);
- cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
+ spin_lock(&svcpt->scp_lock);
+ list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
svcpt->scp_nrqbds_total++;
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
return rqbd;
}
-void
+static void
ptlrpc_free_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
{
struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
LASSERT(rqbd->rqbd_refcount == 0);
- LASSERT(cfs_list_empty(&rqbd->rqbd_reqs));
+ LASSERT(list_empty(&rqbd->rqbd_reqs));
- cfs_spin_lock(&svcpt->scp_lock);
- cfs_list_del(&rqbd->rqbd_list);
+ spin_lock(&svcpt->scp_lock);
+ list_del(&rqbd->rqbd_list);
svcpt->scp_nrqbds_total--;
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
OBD_FREE_LARGE(rqbd->rqbd_buffer, svcpt->scp_service->srv_buf_size);
OBD_FREE_PTR(rqbd);
}
-int
+static int
ptlrpc_grow_req_bufs(struct ptlrpc_service_part *svcpt, int post)
{
struct ptlrpc_service *svc = svcpt->scp_service;
int rc = 0;
int i;
+ if (svcpt->scp_rqbd_allocating)
+ goto try_post;
+
+ spin_lock(&svcpt->scp_lock);
+ /* check again with lock */
+ if (svcpt->scp_rqbd_allocating) {
+ /* NB: we might allow more than one thread in the future */
+ LASSERT(svcpt->scp_rqbd_allocating == 1);
+ spin_unlock(&svcpt->scp_lock);
+ goto try_post;
+ }
+
+ svcpt->scp_rqbd_allocating++;
+ spin_unlock(&svcpt->scp_lock);
+
+
for (i = 0; i < svc->srv_nbuf_per_group; i++) {
- /* NB: another thread might be doing this as well, we need to
- * make sure that it wouldn't over-allocate, see LU-1212. */
+ /* 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)
break;
}
}
+ spin_lock(&svcpt->scp_lock);
+
+ LASSERT(svcpt->scp_rqbd_allocating == 1);
+ svcpt->scp_rqbd_allocating--;
+
+ spin_unlock(&svcpt->scp_lock);
+
CDEBUG(D_RPCTRACE,
"%s: allocate %d new %d-byte reqbufs (%d/%d left), rc = %d\n",
svc->srv_name, i, svc->srv_buf_size, svcpt->scp_nrqbds_posted,
svcpt->scp_nrqbds_total, rc);
+ try_post:
if (post && rc == 0)
rc = ptlrpc_server_post_idle_rqbds(svcpt);
rs->rs_no_ack = !!no_ack;
}
}
+EXPORT_SYMBOL(ptlrpc_save_lock);
-#ifdef __KERNEL__
-#define HRT_RUNNING 0
-#define HRT_STOPPING 1
+struct ptlrpc_hr_partition;
struct ptlrpc_hr_thread {
- cfs_spinlock_t hrt_lock;
- unsigned long hrt_flags;
- cfs_waitq_t hrt_wait;
- cfs_list_t hrt_queue;
- cfs_completion_t hrt_completion;
+ int hrt_id; /* thread ID */
+ spinlock_t hrt_lock;
+ wait_queue_head_t hrt_waitq;
+ struct list_head hrt_queue; /* RS queue */
+ struct ptlrpc_hr_partition *hrt_partition;
+};
+
+struct ptlrpc_hr_partition {
+ /* # of started threads */
+ atomic_t hrp_nstarted;
+ /* # of stopped threads */
+ atomic_t hrp_nstopped;
+ /* cpu partition id */
+ int hrp_cpt;
+ /* round-robin rotor for choosing thread */
+ int hrp_rotor;
+ /* total number of threads on this partition */
+ int hrp_nthrs;
+ /* threads table */
+ struct ptlrpc_hr_thread *hrp_thrs;
};
+#define HRT_RUNNING 0
+#define HRT_STOPPING 1
+
struct ptlrpc_hr_service {
- int hr_index;
- int hr_n_threads;
- int hr_size;
- struct ptlrpc_hr_thread hr_threads[0];
+ /* CPU partition table, it's just cfs_cpt_table for now */
+ struct cfs_cpt_table *hr_cpt_table;
+ /** controller sleep waitq */
+ wait_queue_head_t hr_waitq;
+ unsigned int hr_stopping;
+ /** roundrobin rotor for non-affinity service */
+ unsigned int hr_rotor;
+ /* partition data */
+ struct ptlrpc_hr_partition **hr_partitions;
};
struct rs_batch {
- cfs_list_t rsb_replies;
+ struct list_head rsb_replies;
unsigned int rsb_n_replies;
struct ptlrpc_service_part *rsb_svcpt;
};
-/**
- * A pointer to per-node reply handling service.
- */
-static struct ptlrpc_hr_service *ptlrpc_hr = NULL;
+/** reply handling service. */
+static struct ptlrpc_hr_service ptlrpc_hr;
/**
* maximum mumber of replies scheduled in one batch
*/
static void rs_batch_init(struct rs_batch *b)
{
- memset(b, 0, sizeof *b);
- CFS_INIT_LIST_HEAD(&b->rsb_replies);
+ memset(b, 0, sizeof *b);
+ INIT_LIST_HEAD(&b->rsb_replies);
}
/**
* Choose an hr thread to dispatch requests to.
*/
-static unsigned int get_hr_thread_index(struct ptlrpc_hr_service *hr)
+static struct ptlrpc_hr_thread *
+ptlrpc_hr_select(struct ptlrpc_service_part *svcpt)
{
- unsigned int idx;
+ struct ptlrpc_hr_partition *hrp;
+ unsigned int rotor;
+
+ if (svcpt->scp_cpt >= 0 &&
+ svcpt->scp_service->srv_cptable == ptlrpc_hr.hr_cpt_table) {
+ /* directly match partition */
+ hrp = ptlrpc_hr.hr_partitions[svcpt->scp_cpt];
- /* Concurrent modification of hr_index w/o any spinlock
- protection is harmless as long as the result fits
- [0..(hr_n_threads-1)] range and each thread gets near equal
- load. */
- idx = hr->hr_index;
- hr->hr_index = (idx >= hr->hr_n_threads - 1) ? 0 : idx + 1;
- return idx;
+ } else {
+ rotor = ptlrpc_hr.hr_rotor++;
+ rotor %= cfs_cpt_number(ptlrpc_hr.hr_cpt_table);
+
+ hrp = ptlrpc_hr.hr_partitions[rotor];
+ }
+
+ rotor = hrp->hrp_rotor++;
+ return &hrp->hrp_thrs[rotor % hrp->hrp_nthrs];
}
/**
*/
static void rs_batch_dispatch(struct rs_batch *b)
{
- if (b->rsb_n_replies != 0) {
- struct ptlrpc_hr_service *hr = ptlrpc_hr;
- int idx;
+ if (b->rsb_n_replies != 0) {
+ struct ptlrpc_hr_thread *hrt;
- idx = get_hr_thread_index(hr);
+ hrt = ptlrpc_hr_select(b->rsb_svcpt);
- cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
- cfs_list_splice_init(&b->rsb_replies,
- &hr->hr_threads[idx].hrt_queue);
- cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
- cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
- b->rsb_n_replies = 0;
- }
+ spin_lock(&hrt->hrt_lock);
+ list_splice_init(&b->rsb_replies, &hrt->hrt_queue);
+ spin_unlock(&hrt->hrt_lock);
+
+ wake_up(&hrt->hrt_waitq);
+ b->rsb_n_replies = 0;
+ }
}
/**
if (svcpt != b->rsb_svcpt || b->rsb_n_replies >= MAX_SCHEDULED) {
if (b->rsb_svcpt != NULL) {
rs_batch_dispatch(b);
- cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
+ spin_unlock(&b->rsb_svcpt->scp_rep_lock);
}
- cfs_spin_lock(&svcpt->scp_rep_lock);
+ spin_lock(&svcpt->scp_rep_lock);
b->rsb_svcpt = svcpt;
- }
- cfs_spin_lock(&rs->rs_lock);
- rs->rs_scheduled_ever = 1;
- if (rs->rs_scheduled == 0) {
- cfs_list_move(&rs->rs_list, &b->rsb_replies);
- rs->rs_scheduled = 1;
- b->rsb_n_replies++;
- }
- rs->rs_committed = 1;
- cfs_spin_unlock(&rs->rs_lock);
+ }
+ spin_lock(&rs->rs_lock);
+ rs->rs_scheduled_ever = 1;
+ if (rs->rs_scheduled == 0) {
+ list_move(&rs->rs_list, &b->rsb_replies);
+ rs->rs_scheduled = 1;
+ b->rsb_n_replies++;
+ }
+ rs->rs_committed = 1;
+ spin_unlock(&rs->rs_lock);
}
/**
{
if (b->rsb_svcpt != NULL) {
rs_batch_dispatch(b);
- cfs_spin_unlock(&b->rsb_svcpt->scp_rep_lock);
+ spin_unlock(&b->rsb_svcpt->scp_rep_lock);
}
}
#define DECLARE_RS_BATCH(b) struct rs_batch b
-#else /* __KERNEL__ */
-
-#define rs_batch_init(b) do{}while(0)
-#define rs_batch_fini(b) do{}while(0)
-#define rs_batch_add(b, r) ptlrpc_schedule_difficult_reply(r)
-#define DECLARE_RS_BATCH(b)
-
-#endif /* __KERNEL__ */
/**
* Put reply state into a queue for processing because we received
*/
void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs)
{
-#ifdef __KERNEL__
- struct ptlrpc_hr_service *hr = ptlrpc_hr;
- int idx;
- ENTRY;
+ struct ptlrpc_hr_thread *hrt;
+ ENTRY;
- LASSERT(cfs_list_empty(&rs->rs_list));
+ LASSERT(list_empty(&rs->rs_list));
- idx = get_hr_thread_index(hr);
- cfs_spin_lock(&hr->hr_threads[idx].hrt_lock);
- cfs_list_add_tail(&rs->rs_list, &hr->hr_threads[idx].hrt_queue);
- cfs_spin_unlock(&hr->hr_threads[idx].hrt_lock);
- cfs_waitq_signal(&hr->hr_threads[idx].hrt_wait);
- EXIT;
-#else
- cfs_list_add_tail(&rs->rs_list, &rs->rs_svcpt->scp_rep_queue);
-#endif
+ hrt = ptlrpc_hr_select(rs->rs_svcpt);
+
+ spin_lock(&hrt->hrt_lock);
+ list_add_tail(&rs->rs_list, &hrt->hrt_queue);
+ spin_unlock(&hrt->hrt_lock);
+
+ wake_up(&hrt->hrt_waitq);
+ EXIT;
}
void
{
ENTRY;
- LASSERT_SPIN_LOCKED(&rs->rs_svcpt->scp_rep_lock);
- LASSERT_SPIN_LOCKED(&rs->rs_lock);
- LASSERT (rs->rs_difficult);
- rs->rs_scheduled_ever = 1; /* flag any notification attempt */
+ assert_spin_locked(&rs->rs_svcpt->scp_rep_lock);
+ assert_spin_locked(&rs->rs_lock);
+ LASSERT (rs->rs_difficult);
+ rs->rs_scheduled_ever = 1; /* flag any notification attempt */
- if (rs->rs_scheduled) { /* being set up or already notified */
- EXIT;
- return;
- }
+ if (rs->rs_scheduled) { /* being set up or already notified */
+ EXIT;
+ return;
+ }
- rs->rs_scheduled = 1;
- cfs_list_del_init(&rs->rs_list);
- ptlrpc_dispatch_difficult_reply(rs);
- EXIT;
+ rs->rs_scheduled = 1;
+ list_del_init(&rs->rs_list);
+ ptlrpc_dispatch_difficult_reply(rs);
+ EXIT;
}
+EXPORT_SYMBOL(ptlrpc_schedule_difficult_reply);
void ptlrpc_commit_replies(struct obd_export *exp)
{
* to attend to complete them. */
/* CAVEAT EMPTOR: spinlock ordering!!! */
- cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
- cfs_list_for_each_entry_safe(rs, nxt, &exp->exp_uncommitted_replies,
+ 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) {
- cfs_list_del_init(&rs->rs_obd_list);
+ list_del_init(&rs->rs_obd_list);
rs_batch_add(&batch, rs);
}
}
- cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
- rs_batch_fini(&batch);
- EXIT;
+ spin_unlock(&exp->exp_uncommitted_replies_lock);
+ rs_batch_fini(&batch);
+ EXIT;
}
static int
int posted = 0;
for (;;) {
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
- if (cfs_list_empty(&svcpt->scp_rqbd_idle)) {
- cfs_spin_unlock(&svcpt->scp_lock);
+ if (list_empty(&svcpt->scp_rqbd_idle)) {
+ spin_unlock(&svcpt->scp_lock);
return posted;
}
- rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
+ rqbd = list_entry(svcpt->scp_rqbd_idle.next,
struct ptlrpc_request_buffer_desc,
rqbd_list);
- cfs_list_del(&rqbd->rqbd_list);
+ list_del(&rqbd->rqbd_list);
/* assume we will post successfully */
svcpt->scp_nrqbds_posted++;
- cfs_list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
+ list_add(&rqbd->rqbd_list, &svcpt->scp_rqbd_posted);
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
rc = ptlrpc_register_rqbd(rqbd);
if (rc != 0)
posted = 1;
}
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
svcpt->scp_nrqbds_posted--;
- cfs_list_del(&rqbd->rqbd_list);
- cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
+ list_del(&rqbd->rqbd_list);
+ list_add_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! */
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
return -1;
}
svcpt->scp_at_check = 1;
svcpt->scp_at_checktime = cfs_time_current();
- cfs_waitq_signal(&svcpt->scp_waitq);
+ wake_up(&svcpt->scp_waitq);
}
static void
ptlrpc_server_nthreads_check(struct ptlrpc_service *svc,
struct ptlrpc_service_conf *conf)
{
-#ifdef __KERNEL__
struct ptlrpc_service_thr_conf *tc = &conf->psc_thr;
unsigned init;
unsigned total;
svc->srv_nthrs_cpt_init = init;
if (nthrs * svc->srv_ncpts > tc->tc_nthrs_max) {
- LCONSOLE_WARN("%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);
+ 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);
}
-#endif
}
/**
int rc;
svcpt->scp_cpt = cpt;
- CFS_INIT_LIST_HEAD(&svcpt->scp_threads);
+ INIT_LIST_HEAD(&svcpt->scp_threads);
/* rqbd and incoming request queue */
- cfs_spin_lock_init(&svcpt->scp_lock);
- CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
- CFS_INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
- CFS_INIT_LIST_HEAD(&svcpt->scp_req_incoming);
- cfs_waitq_init(&svcpt->scp_waitq);
+ spin_lock_init(&svcpt->scp_lock);
+ INIT_LIST_HEAD(&svcpt->scp_rqbd_idle);
+ INIT_LIST_HEAD(&svcpt->scp_rqbd_posted);
+ INIT_LIST_HEAD(&svcpt->scp_req_incoming);
+ init_waitqueue_head(&svcpt->scp_waitq);
/* history request & rqbd list */
- CFS_INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
- CFS_INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
+ INIT_LIST_HEAD(&svcpt->scp_hist_reqs);
+ INIT_LIST_HEAD(&svcpt->scp_hist_rqbds);
/* acitve requests and hp requests */
- cfs_spin_lock_init(&svcpt->scp_req_lock);
- CFS_INIT_LIST_HEAD(&svcpt->scp_req_pending);
- CFS_INIT_LIST_HEAD(&svcpt->scp_hreq_pending);
+ spin_lock_init(&svcpt->scp_req_lock);
/* reply states */
- cfs_spin_lock_init(&svcpt->scp_rep_lock);
- CFS_INIT_LIST_HEAD(&svcpt->scp_rep_active);
-#ifndef __KERNEL__
- CFS_INIT_LIST_HEAD(&svcpt->scp_rep_queue);
-#endif
- CFS_INIT_LIST_HEAD(&svcpt->scp_rep_idle);
- cfs_waitq_init(&svcpt->scp_rep_waitq);
- cfs_atomic_set(&svcpt->scp_nreps_difficult, 0);
+ spin_lock_init(&svcpt->scp_rep_lock);
+ INIT_LIST_HEAD(&svcpt->scp_rep_active);
+ INIT_LIST_HEAD(&svcpt->scp_rep_idle);
+ init_waitqueue_head(&svcpt->scp_rep_waitq);
+ atomic_set(&svcpt->scp_nreps_difficult, 0);
/* adaptive timeout */
- cfs_spin_lock_init(&svcpt->scp_at_lock);
+ spin_lock_init(&svcpt->scp_at_lock);
array = &svcpt->scp_at_array;
size = at_est2timeout(at_max);
/* allocate memory for scp_at_array (ptlrpc_at_array) */
OBD_CPT_ALLOC(array->paa_reqs_array,
- svc->srv_cptable, cpt, sizeof(cfs_list_t) * size);
+ svc->srv_cptable, cpt, sizeof(struct list_head) * size);
if (array->paa_reqs_array == NULL)
return -ENOMEM;
for (index = 0; index < size; index++)
- CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
+ INIT_LIST_HEAD(&array->paa_reqs_array[index]);
OBD_CPT_ALLOC(array->paa_reqs_count,
svc->srv_cptable, cpt, sizeof(__u32) * size);
if (array->paa_reqs_array != NULL) {
OBD_FREE(array->paa_reqs_array,
- sizeof(cfs_list_t) * array->paa_size);
+ sizeof(struct list_head) * array->paa_size);
array->paa_reqs_array = NULL;
}
*/
struct ptlrpc_service *
ptlrpc_register_service(struct ptlrpc_service_conf *conf,
- cfs_proc_dir_entry_t *proc_entry)
+ struct proc_dir_entry *proc_entry)
{
struct ptlrpc_service_cpt_conf *cconf = &conf->psc_cpt;
struct ptlrpc_service *service;
if (rc <= 0) {
CERROR("%s: failed to parse CPT array %s: %d\n",
conf->psc_name, cconf->cc_pattern, rc);
+ if (cpts != NULL)
+ OBD_FREE(cpts, sizeof(*cpts) * ncpts);
RETURN(ERR_PTR(rc < 0 ? rc : -EINVAL));
}
ncpts = rc;
service->srv_cpt_bits++;
/* public members */
- cfs_spin_lock_init(&service->srv_lock);
+ spin_lock_init(&service->srv_lock);
service->srv_name = conf->psc_name;
service->srv_watchdog_factor = conf->psc_watchdog_factor;
- CFS_INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
+ INIT_LIST_HEAD(&service->srv_list); /* for safty of cleanup */
/* buffer configuration */
- service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 :
- max(conf->psc_buf.bc_nbufs /
- service->srv_ncpts, 1U);
+ service->srv_nbuf_per_group = test_req_buffer_pressure ?
+ 1 : conf->psc_buf.bc_nbufs;
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;
rc = LNetSetLazyPortal(service->srv_req_portal);
LASSERT(rc == 0);
- cfs_spin_lock (&ptlrpc_all_services_lock);
- cfs_list_add (&service->srv_list, &ptlrpc_all_services);
- cfs_spin_unlock (&ptlrpc_all_services_lock);
+ mutex_lock(&ptlrpc_all_services_mutex);
+ 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 (proc_entry != NULL)
- ptlrpc_lprocfs_register_service(proc_entry, service);
+ rc = ptlrpc_service_nrs_setup(service);
+ if (rc != 0)
+ GOTO(failed, rc);
- CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
- service->srv_name, service->srv_req_portal);
+ CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
+ service->srv_name, service->srv_req_portal);
-#ifdef __KERNEL__
rc = ptlrpc_start_threads(service);
if (rc != 0) {
CERROR("Failed to start threads for service %s: %d\n",
service->srv_name, rc);
GOTO(failed, rc);
}
-#endif
RETURN(service);
failed:
ptlrpc_unregister_service(service);
RETURN(ERR_PTR(rc));
}
+EXPORT_SYMBOL(ptlrpc_register_service);
/**
* to actually free the request, must be called without holding svc_lock.
*/
static void ptlrpc_server_free_request(struct ptlrpc_request *req)
{
- LASSERT(cfs_atomic_read(&req->rq_refcount) == 0);
- LASSERT(cfs_list_empty(&req->rq_timed_list));
+ 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. */
- ptlrpc_req_drop_rs(req);
+ /* 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);
+ sptlrpc_svc_ctx_decref(req);
- if (req != &req->rq_rqbd->rqbd_req) {
- /* NB request buffers use an embedded
- * req if the incoming req unlinked the
- * MD; this isn't one of them! */
- OBD_FREE(req, sizeof(*req));
- }
+ if (req != &req->rq_rqbd->rqbd_req) {
+ /* NB request buffers use an embedded
+ * req if the incoming req unlinked the
+ * MD; this isn't one of them! */
+ ptlrpc_request_cache_free(req);
+ }
}
/**
struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
struct ptlrpc_service_part *svcpt = rqbd->rqbd_svcpt;
struct ptlrpc_service *svc = svcpt->scp_service;
- int refcount;
- cfs_list_t *tmp;
- cfs_list_t *nxt;
+ int refcount;
+ struct list_head *tmp;
+ struct list_head *nxt;
- if (!cfs_atomic_dec_and_test(&req->rq_refcount))
- return;
+ if (!atomic_dec_and_test(&req->rq_refcount))
+ return;
- if (req->rq_at_linked) {
- struct ptlrpc_at_array *array = &svcpt->scp_at_array;
- __u32 index = req->rq_at_index;
+ if (req->rq_session.lc_state == LCS_ENTERED) {
+ lu_context_exit(&req->rq_session);
+ lu_context_fini(&req->rq_session);
+ }
- cfs_spin_lock(&svcpt->scp_at_lock);
+ if (req->rq_at_linked) {
+ spin_lock(&svcpt->scp_at_lock);
+ /* 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);
+ }
- LASSERT(!cfs_list_empty(&req->rq_timed_list));
- cfs_list_del_init(&req->rq_timed_list);
- cfs_spin_lock(&req->rq_lock);
- req->rq_at_linked = 0;
- cfs_spin_unlock(&req->rq_lock);
- array->paa_reqs_count[index]--;
- array->paa_count--;
+ LASSERT(list_empty(&req->rq_timed_list));
- cfs_spin_unlock(&svcpt->scp_at_lock);
- } else {
- LASSERT(cfs_list_empty(&req->rq_timed_list));
+ /* finalize request */
+ if (req->rq_export) {
+ class_export_put(req->rq_export);
+ req->rq_export = NULL;
}
- /* finalize request */
- if (req->rq_export) {
- class_export_put(req->rq_export);
- req->rq_export = NULL;
- }
-
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
- cfs_list_add(&req->rq_list, &rqbd->rqbd_reqs);
+ list_add(&req->rq_list, &rqbd->rqbd_reqs);
- refcount = --(rqbd->rqbd_refcount);
- if (refcount == 0) {
- /* request buffer is now idle: add to history */
- cfs_list_del(&rqbd->rqbd_list);
+ refcount = --(rqbd->rqbd_refcount);
+ if (refcount == 0) {
+ /* request buffer is now idle: add to history */
+ list_del(&rqbd->rqbd_list);
- cfs_list_add_tail(&rqbd->rqbd_list, &svcpt->scp_hist_rqbds);
+ list_add_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 */
while (svcpt->scp_hist_nrqbds > svc->srv_hist_nrqbds_cpt_max) {
- rqbd = cfs_list_entry(svcpt->scp_hist_rqbds.next,
- struct ptlrpc_request_buffer_desc,
- rqbd_list);
+ rqbd = list_entry(svcpt->scp_hist_rqbds.next,
+ struct ptlrpc_request_buffer_desc,
+ rqbd_list);
- cfs_list_del(&rqbd->rqbd_list);
+ 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 */
- cfs_list_for_each(tmp, &rqbd->rqbd_reqs) {
- req = cfs_list_entry(tmp, struct ptlrpc_request,
- rq_list);
- /* Track the highest culled req seq */
+ /* 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);
+ /* Track the highest culled req seq */
if (req->rq_history_seq >
svcpt->scp_hist_seq_culled) {
svcpt->scp_hist_seq_culled =
req->rq_history_seq;
}
- cfs_list_del(&req->rq_history_list);
+ list_del(&req->rq_history_list);
}
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
- cfs_list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
- req = cfs_list_entry(rqbd->rqbd_reqs.next,
- struct ptlrpc_request,
- rq_list);
- cfs_list_del(&req->rq_list);
- ptlrpc_server_free_request(req);
- }
+ list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
+ req = list_entry(rqbd->rqbd_reqs.next,
+ struct ptlrpc_request,
+ rq_list);
+ list_del(&req->rq_list);
+ ptlrpc_server_free_request(req);
+ }
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
/*
* now all reqs including the embedded req has been
* disposed, schedule request buffer for re-use.
*/
- LASSERT(cfs_atomic_read(&rqbd->rqbd_req.rq_refcount) ==
- 0);
- cfs_list_add_tail(&rqbd->rqbd_list,
- &svcpt->scp_rqbd_idle);
+ LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
+ list_add_tail(&rqbd->rqbd_list, &svcpt->scp_rqbd_idle);
}
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
} else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
/* If we are low on memory, we are not interested in history */
- cfs_list_del(&req->rq_list);
- cfs_list_del_init(&req->rq_history_list);
+ list_del(&req->rq_list);
+ list_del_init(&req->rq_history_list);
+
+ /* Track the highest culled req seq */
+ if (req->rq_history_seq > svcpt->scp_hist_seq_culled)
+ svcpt->scp_hist_seq_culled = req->rq_history_seq;
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
ptlrpc_server_free_request(req);
} else {
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_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
+ * 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);
+
+ 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;
}
/**
* to finish a request: stop sending more early replies, and release
- * the request. should be called after we finished handling the request.
+ * the request.
*/
static void ptlrpc_server_finish_request(struct ptlrpc_service_part *svcpt,
struct ptlrpc_request *req)
{
- ptlrpc_hpreq_fini(req);
+ ptlrpc_server_hpreq_fini(req);
+
+ ptlrpc_server_drop_request(req);
+}
- cfs_spin_lock(&svcpt->scp_req_lock);
+/**
+ * to finish an active request: stop sending more early replies, and release
+ * the request. should be called after we finished handling the request.
+ */
+static void ptlrpc_server_finish_active_request(
+ struct ptlrpc_service_part *svcpt,
+ struct ptlrpc_request *req)
+{
+ spin_lock(&svcpt->scp_req_lock);
+ ptlrpc_nrs_req_stop_nolock(req);
svcpt->scp_nreqs_active--;
if (req->rq_hp)
svcpt->scp_nhreqs_active--;
- cfs_spin_unlock(&svcpt->scp_req_lock);
+ spin_unlock(&svcpt->scp_req_lock);
- ptlrpc_server_drop_request(req);
+ ptlrpc_nrs_req_finalize(req);
+
+ if (req->rq_export != NULL)
+ class_export_rpc_dec(req->rq_export);
+
+ ptlrpc_server_finish_request(svcpt, req);
}
/**
* This function is only called when some export receives a message (i.e.,
* the network is up.)
*/
-static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
+void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
{
struct obd_export *oldest_exp;
time_t oldest_time, new_time;
RETURN_EXIT;
exp->exp_last_request_time = new_time;
- CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
- exp->exp_client_uuid.uuid,
- exp->exp_last_request_time, exp);
-
- /* exports may get disconnected from the chain even though the
- export has references, so we must keep the spin lock while
- manipulating the lists */
- cfs_spin_lock(&exp->exp_obd->obd_dev_lock);
-
- if (cfs_list_empty(&exp->exp_obd_chain_timed)) {
- /* this one is not timed */
- cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
- RETURN_EXIT;
- }
- cfs_list_move_tail(&exp->exp_obd_chain_timed,
- &exp->exp_obd->obd_exports_timed);
+ /* 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);
- oldest_exp = cfs_list_entry(exp->exp_obd->obd_exports_timed.next,
- struct obd_export, exp_obd_chain_timed);
- oldest_time = oldest_exp->exp_last_request_time;
- cfs_spin_unlock(&exp->exp_obd->obd_dev_lock);
+ if (list_empty(&exp->exp_obd_chain_timed)) {
+ /* this one is not timed */
+ spin_unlock(&exp->exp_obd->obd_dev_lock);
+ RETURN_EXIT;
+ }
+
+ list_move_tail(&exp->exp_obd_chain_timed,
+ &exp->exp_obd->obd_exports_timed);
+
+ oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
+ struct obd_export, 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 */
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,
+ exp->exp_obd->obd_name,
obd_export_nid2str(oldest_exp), oldest_time);
}
} else {
*/
static int ptlrpc_check_req(struct ptlrpc_request *req)
{
- int rc = 0;
+ struct obd_device *obd = req->rq_export->exp_obd;
+ int rc = 0;
- if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
- req->rq_export->exp_conn_cnt)) {
- DEBUG_REQ(D_ERROR, req,
- "DROPPING req from old connection %d < %d",
- lustre_msg_get_conn_cnt(req->rq_reqmsg),
- req->rq_export->exp_conn_cnt);
- return -EEXIST;
- }
- if (unlikely(req->rq_export->exp_obd &&
- req->rq_export->exp_obd->obd_fail)) {
- /* Failing over, don't handle any more reqs, send
- error response instead. */
- CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
- req, req->rq_export->exp_obd->obd_name);
+ if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
+ req->rq_export->exp_conn_cnt)) {
+ DEBUG_REQ(D_RPCTRACE, req,
+ "DROPPING req from old connection %d < %d",
+ lustre_msg_get_conn_cnt(req->rq_reqmsg),
+ req->rq_export->exp_conn_cnt);
+ return -EEXIST;
+ }
+ if (unlikely(obd == NULL || obd->obd_fail)) {
+ /* 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) &&
- !(req->rq_export->exp_obd->obd_recovering)) {
+ !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 &&
- !(req->rq_export->exp_obd->obd_recovering)) {
+ !obd->obd_recovering) {
DEBUG_REQ(D_ERROR, req, "Invalid req with transno "
LPU64" without recovery",
lustre_msg_get_transno(req->rq_reqmsg));
if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
return(-ENOSYS);
- cfs_spin_lock(&svcpt->scp_at_lock);
- LASSERT(cfs_list_empty(&req->rq_timed_list));
+ 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. */
- cfs_list_for_each_entry_reverse(rq,
+ list_for_each_entry_reverse(rq,
&array->paa_reqs_array[index],
rq_timed_list) {
if (req->rq_deadline >= rq->rq_deadline) {
- cfs_list_add(&req->rq_timed_list,
+ list_add(&req->rq_timed_list,
&rq->rq_timed_list);
break;
}
}
/* Add the request at the head of the list */
- if (cfs_list_empty(&req->rq_timed_list))
- cfs_list_add(&req->rq_timed_list,
+ if (list_empty(&req->rq_timed_list))
+ list_add(&req->rq_timed_list,
&array->paa_reqs_array[index]);
- cfs_spin_lock(&req->rq_lock);
- req->rq_at_linked = 1;
- cfs_spin_unlock(&req->rq_lock);
- req->rq_at_index = index;
- array->paa_reqs_count[index]++;
- array->paa_count++;
- if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
- array->paa_deadline = req->rq_deadline;
+ spin_lock(&req->rq_lock);
+ req->rq_at_linked = 1;
+ spin_unlock(&req->rq_lock);
+ req->rq_at_index = index;
+ array->paa_reqs_count[index]++;
+ array->paa_count++;
+ if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
+ array->paa_deadline = req->rq_deadline;
ptlrpc_at_set_timer(svcpt);
}
- cfs_spin_unlock(&svcpt->scp_at_lock);
+ spin_unlock(&svcpt->scp_at_lock);
return 0;
}
+static void
+ptlrpc_at_remove_timed(struct ptlrpc_request *req)
+{
+ struct ptlrpc_at_array *array;
+
+ array = &req->rq_rqbd->rqbd_svcpt->scp_at_array;
+
+ /* NB: must call with hold svcpt::scp_at_lock */
+ LASSERT(!list_empty(&req->rq_timed_list));
+ list_del_init(&req->rq_timed_list);
+
+ spin_lock(&req->rq_lock);
+ req->rq_at_linked = 0;
+ spin_unlock(&req->rq_lock);
+
+ array->paa_reqs_count[req->rq_at_index]--;
+ array->paa_count--;
+}
+
+/*
+ * Attempt to extend the request deadline by sending an early reply to the
+ * client.
+ */
static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
{
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;
- int rc;
- ENTRY;
+ struct ptlrpc_request *reqcopy;
+ struct lustre_msg *reqmsg;
+ cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
+ time_t newdl;
+ int rc;
+
+ ENTRY;
+
+ if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_RECONNECT)) {
+ /* don't send early reply */
+ RETURN(1);
+ }
/* deadline is when the client expects us to reply, margin is the
difference between clients' and servers' expectations */
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
- * 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. */
- at_measured(&svcpt->scp_at_estimate, min(at_extra,
- req->rq_export->exp_obd->obd_recovery_timeout / 4));
+ 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
+ * 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
+ * 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);
} else {
- /* Fake our processing time into the future to ask the clients
- * for some extra amount of time */
+ /* 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
+ * based on this service estimate (plus some additional time to
+ * account for network latency). See ptlrpc_at_recv_early_reply
+ */
at_measured(&svcpt->scp_at_estimate, at_extra +
cfs_time_current_sec() -
req->rq_arrival_time.tv_sec);
+ newdl = req->rq_arrival_time.tv_sec +
+ at_get(&svcpt->scp_at_estimate);
+ }
- /* Check to see if we've actually increased the deadline -
- * we may be past adaptive_max */
- if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
- at_get(&svcpt->scp_at_estimate)) {
- DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
- "(%ld/%ld), not sending early reply\n",
- olddl, req->rq_arrival_time.tv_sec +
- at_get(&svcpt->scp_at_estimate) -
- cfs_time_current_sec());
- RETURN(-ETIMEDOUT);
- }
+ /* 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());
+ RETURN(-ETIMEDOUT);
}
- newdl = cfs_time_current_sec() + at_get(&svcpt->scp_at_estimate);
- OBD_ALLOC(reqcopy, sizeof *reqcopy);
- if (reqcopy == NULL)
- RETURN(-ENOMEM);
- OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
- if (!reqmsg) {
- OBD_FREE(reqcopy, sizeof *reqcopy);
- RETURN(-ENOMEM);
- }
+ reqcopy = ptlrpc_request_cache_alloc(GFP_NOFS);
+ if (reqcopy == NULL)
+ RETURN(-ENOMEM);
+ OBD_ALLOC_LARGE(reqmsg, req->rq_reqlen);
+ if (!reqmsg)
+ GOTO(out_free, rc = -ENOMEM);
*reqcopy = *req;
reqcopy->rq_reply_state = NULL;
reqcopy->rq_reqmsg = reqmsg;
memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
- LASSERT(cfs_atomic_read(&req->rq_refcount));
- /** if it is last refcount then early reply isn't needed */
- if (cfs_atomic_read(&req->rq_refcount) == 1) {
- DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
- "abort sending early reply\n");
- GOTO(out, rc = -EINVAL);
- }
+ /*
+ * tgt_brw_read() and tgt_brw_write() may have decided not to reply.
+ * Without this check, we would fail the rq_no_reply assertion in
+ * ptlrpc_send_reply().
+ */
+ if (reqcopy->rq_no_reply)
+ GOTO(out, rc = -ETIMEDOUT);
+
+ LASSERT(atomic_read(&req->rq_refcount));
+ /** 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");
+ GOTO(out, rc = -EINVAL);
+ }
/* Connection ref */
reqcopy->rq_export = class_conn2export(
GOTO(out, rc = -ENODEV);
/* RPC ref */
- class_export_rpc_get(reqcopy->rq_export);
+ 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);
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);
- }
+ 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);
+ }
/* 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_put(reqcopy->rq_export);
- class_export_put(reqcopy->rq_export);
+ class_export_rpc_dec(reqcopy->rq_export);
+ class_export_put(reqcopy->rq_export);
out:
- sptlrpc_svc_ctx_decref(reqcopy);
- OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
- OBD_FREE(reqcopy, sizeof *reqcopy);
- RETURN(rc);
+ sptlrpc_svc_ctx_decref(reqcopy);
+ OBD_FREE_LARGE(reqmsg, req->rq_reqlen);
+out_free:
+ ptlrpc_request_cache_free(reqcopy);
+ RETURN(rc);
}
/* Send early replies to everybody expiring within at_early_margin
{
struct ptlrpc_at_array *array = &svcpt->scp_at_array;
struct ptlrpc_request *rq, *n;
- cfs_list_t work_list;
+ struct list_head work_list;
__u32 index, count;
time_t deadline;
time_t now = cfs_time_current_sec();
int first, counter = 0;
ENTRY;
- cfs_spin_lock(&svcpt->scp_at_lock);
+ spin_lock(&svcpt->scp_at_lock);
if (svcpt->scp_at_check == 0) {
- cfs_spin_unlock(&svcpt->scp_at_lock);
+ spin_unlock(&svcpt->scp_at_lock);
RETURN(0);
}
delay = cfs_time_sub(cfs_time_current(), svcpt->scp_at_checktime);
svcpt->scp_at_check = 0;
if (array->paa_count == 0) {
- cfs_spin_unlock(&svcpt->scp_at_lock);
+ spin_unlock(&svcpt->scp_at_lock);
RETURN(0);
}
if (first > at_early_margin) {
/* We've still got plenty of time. Reset the timer. */
ptlrpc_at_set_timer(svcpt);
- cfs_spin_unlock(&svcpt->scp_at_lock);
+ spin_unlock(&svcpt->scp_at_lock);
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. */
- CFS_INIT_LIST_HEAD(&work_list);
- deadline = -1;
- index = (unsigned long)array->paa_deadline % array->paa_size;
- count = array->paa_count;
- while (count > 0) {
- count -= array->paa_reqs_count[index];
- cfs_list_for_each_entry_safe(rq, n,
- &array->paa_reqs_array[index],
- rq_timed_list) {
- if (rq->rq_deadline <= now + at_early_margin) {
- cfs_list_del_init(&rq->rq_timed_list);
- /**
- * ptlrpc_server_drop_request() may drop
- * refcount to 0 already. Let's check this and
- * don't add entry to work_list
- */
- if (likely(cfs_atomic_inc_not_zero(&rq->rq_refcount)))
- cfs_list_add(&rq->rq_timed_list, &work_list);
- counter++;
- array->paa_reqs_count[index]--;
- array->paa_count--;
- cfs_spin_lock(&rq->rq_lock);
- rq->rq_at_linked = 0;
- cfs_spin_unlock(&rq->rq_lock);
- continue;
- }
-
- /* update the earliest deadline */
- if (deadline == -1 || rq->rq_deadline < deadline)
- deadline = rq->rq_deadline;
+ /* 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);
+ deadline = -1;
+ index = (unsigned long)array->paa_deadline % array->paa_size;
+ count = array->paa_count;
+ while (count > 0) {
+ count -= array->paa_reqs_count[index];
+ list_for_each_entry_safe(rq, n,
+ &array->paa_reqs_array[index],
+ rq_timed_list) {
+ if (rq->rq_deadline > now + at_early_margin) {
+ /* update the earliest deadline */
+ if (deadline == -1 ||
+ rq->rq_deadline < deadline)
+ deadline = rq->rq_deadline;
+ break;
+ }
- 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)))
+ list_add(&rq->rq_timed_list, &work_list);
+ counter++;
+ }
- if (++index >= array->paa_size)
- index = 0;
- }
- array->paa_deadline = deadline;
+ if (++index >= array->paa_size)
+ index = 0;
+ }
+ array->paa_deadline = deadline;
/* we have a new earliest deadline, restart the timer */
ptlrpc_at_set_timer(svcpt);
- cfs_spin_unlock(&svcpt->scp_at_lock);
+ 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);
/* we took additional refcount so entries can't be deleted from list, no
* locking is needed */
- while (!cfs_list_empty(&work_list)) {
- rq = cfs_list_entry(work_list.next, struct ptlrpc_request,
+ while (!list_empty(&work_list)) {
+ rq = list_entry(work_list.next, struct ptlrpc_request,
rq_timed_list);
- cfs_list_del_init(&rq->rq_timed_list);
+ list_del_init(&rq->rq_timed_list);
if (ptlrpc_at_send_early_reply(rq) == 0)
ptlrpc_at_add_timed(rq);
RETURN(1); /* return "did_something" for liblustre */
}
-/**
- * Put the request to the export list if the request may become
- * a high priority one.
- */
-static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
- 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 int ptlrpc_server_check_resend_in_progress(struct ptlrpc_request *req)
{
- int rc = 0;
- ENTRY;
+ struct ptlrpc_request *tmp = NULL;
- if (svc->srv_ops.so_hpreq_handler) {
- rc = svc->srv_ops.so_hpreq_handler(req);
- if (rc)
- RETURN(rc);
- }
- if (req->rq_export && req->rq_ops) {
- /* Perform request specific check. We should do this check
- * before the request is added into exp_hp_rpcs list otherwise
- * it may hit swab race at LU-1044. */
- if (req->rq_ops->hpreq_check)
- rc = req->rq_ops->hpreq_check(req);
-
- cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
- cfs_list_add(&req->rq_exp_list,
- &req->rq_export->exp_hp_rpcs);
- cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
- }
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ||
+ (atomic_read(&req->rq_export->exp_rpc_count) == 0))
+ return 0;
- RETURN(rc);
-}
+ /* bulk request are aborted upon reconnect, don't try to
+ * find a match */
+ if (req->rq_bulk_write || req->rq_bulk_read)
+ return 0;
-/** Remove the request from the export list. */
-static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
-{
- ENTRY;
- if (req->rq_export && req->rq_ops) {
- /* refresh lock timeout again so that client has more
- * room to send lock cancel RPC. */
- if (req->rq_ops->hpreq_fini)
- req->rq_ops->hpreq_fini(req);
-
- cfs_spin_lock_bh(&req->rq_export->exp_rpc_lock);
- cfs_list_del_init(&req->rq_exp_list);
- cfs_spin_unlock_bh(&req->rq_export->exp_rpc_lock);
- }
- EXIT;
+ /* 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) {
+ /* Found duplicate one */
+ if (tmp->rq_xid == req->rq_xid)
+ goto found;
+ }
+ list_for_each_entry(tmp, &req->rq_export->exp_hp_rpcs,
+ rq_exp_list) {
+ /* Found duplicate one */
+ if (tmp->rq_xid == req->rq_xid)
+ goto found;
+ }
+ return 0;
+
+found:
+ DEBUG_REQ(D_HA, req, "Found duplicate req in processing");
+ DEBUG_REQ(D_HA, tmp, "Request being processed");
+ return -EBUSY;
}
/**
- * Make the request a high priority one.
- *
- * All the high priority requests are queued in a separate FIFO
- * ptlrpc_service_part::scp_hpreq_pending list which is parallel to
- * ptlrpc_service_part::scp_req_pending list but has a higher priority
- * for handling.
- *
- * \see ptlrpc_server_handle_request().
+ * Put the request to the export list if the request may become
+ * a high priority one.
*/
-static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service_part *svcpt,
- struct ptlrpc_request *req)
+static int ptlrpc_server_hpreq_init(struct ptlrpc_service_part *svcpt,
+ struct ptlrpc_request *req)
{
- ENTRY;
+ struct list_head *list;
+ int rc, hp = 0;
+
+ ENTRY;
- cfs_spin_lock(&req->rq_lock);
- if (req->rq_hp == 0) {
- int opc = lustre_msg_get_opc(req->rq_reqmsg);
+ if (svcpt->scp_service->srv_ops.so_hpreq_handler) {
+ rc = svcpt->scp_service->srv_ops.so_hpreq_handler(req);
+ if (rc < 0)
+ RETURN(rc);
+ LASSERT(rc == 0);
+ }
+ if (req->rq_export) {
+ if (req->rq_ops) {
+ /* Perform request specific check. We should do this
+ * check before the request is added into exp_hp_rpcs
+ * list otherwise it may hit swab race at LU-1044. */
+ if (req->rq_ops->hpreq_check) {
+ rc = req->rq_ops->hpreq_check(req);
+ /**
+ * XXX: Out of all current
+ * ptlrpc_hpreq_ops::hpreq_check(), only
+ * ldlm_cancel_hpreq_check() can return an
+ * error code; other functions assert in
+ * similar places, which seems odd.
+ * What also does not seem right is that
+ * handlers for those RPCs do not assert
+ * on the same checks, but rather handle the
+ * error cases. e.g. see ost_rw_hpreq_check(),
+ * and ost_brw_read(), ost_brw_write().
+ */
+ if (rc < 0)
+ RETURN(rc);
+ LASSERT(rc == 0 || rc == 1);
+ hp = rc;
+ }
+ list = &req->rq_export->exp_hp_rpcs;
+ } else {
+ list = &req->rq_export->exp_reg_rpcs;
+ }
- /* Add to the high priority queue. */
- cfs_list_move_tail(&req->rq_list, &svcpt->scp_hreq_pending);
- req->rq_hp = 1;
- if (opc != OBD_PING)
- DEBUG_REQ(D_RPCTRACE, req, "high priority req");
- }
- cfs_spin_unlock(&req->rq_lock);
- EXIT;
+ /* do search for duplicated xid and the adding to the list
+ * atomically */
+ spin_lock_bh(&req->rq_export->exp_rpc_lock);
+ rc = ptlrpc_server_check_resend_in_progress(req);
+ if (rc < 0) {
+ spin_unlock_bh(&req->rq_export->exp_rpc_lock);
+ RETURN(rc);
+ }
+ list_add(&req->rq_exp_list, list);
+ spin_unlock_bh(&req->rq_export->exp_rpc_lock);
+ }
+
+ ptlrpc_nrs_req_initialize(svcpt, req, !!hp);
+
+ RETURN(hp);
}
-/**
- * \see ptlrpc_hpreq_reorder_nolock
- */
-void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
+/** Remove the request from the export list. */
+static void ptlrpc_server_hpreq_fini(struct ptlrpc_request *req)
{
- struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
ENTRY;
-
- cfs_spin_lock(&svcpt->scp_req_lock);
- /* It may happen that the request is already taken for the processing
- * but still in the export list, or the request is not in the request
- * queue but in the export list already, do not add it into the
- * HP list. */
- if (!cfs_list_empty(&req->rq_list))
- ptlrpc_hpreq_reorder_nolock(svcpt, req);
- cfs_spin_unlock(&svcpt->scp_req_lock);
+ if (req->rq_export) {
+ /* 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);
+ }
EXIT;
}
-/** Check if the request is a high priority one. */
-static int ptlrpc_server_hpreq_check(struct ptlrpc_service *svc,
- struct ptlrpc_request *req)
+static int ptlrpc_hpreq_check(struct ptlrpc_request *req)
{
- ENTRY;
+ return 1;
+}
- /* Check by request opc. */
- if (OBD_PING == lustre_msg_get_opc(req->rq_reqmsg))
- RETURN(1);
+static struct ptlrpc_hpreq_ops ptlrpc_hpreq_common = {
+ .hpreq_check = ptlrpc_hpreq_check,
+};
+
+/* Hi-Priority RPC check by RPC operation code. */
+int ptlrpc_hpreq_handler(struct ptlrpc_request *req)
+{
+ int opc = lustre_msg_get_opc(req->rq_reqmsg);
- RETURN(ptlrpc_hpreq_init(svc, req));
+ /* 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;
+
+ return 0;
}
+EXPORT_SYMBOL(ptlrpc_hpreq_handler);
-/** Check if a request is a high priority one. */
static int ptlrpc_server_request_add(struct ptlrpc_service_part *svcpt,
struct ptlrpc_request *req)
{
int rc;
ENTRY;
- rc = ptlrpc_server_hpreq_check(svcpt->scp_service, req);
+ rc = ptlrpc_server_hpreq_init(svcpt, req);
if (rc < 0)
RETURN(rc);
- cfs_spin_lock(&svcpt->scp_req_lock);
-
- if (rc)
- ptlrpc_hpreq_reorder_nolock(svcpt, req);
- else
- cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_pending);
+ /* the current thread is not the processing thread for this request
+ * since that, but request is in exp_hp_list and can be find there.
+ * Remove all relations between request and old thread. */
+ req->rq_svc_thread->t_env->le_ses = NULL;
+ req->rq_svc_thread = NULL;
+ req->rq_session.lc_thread = NULL;
- cfs_spin_unlock(&svcpt->scp_req_lock);
+ ptlrpc_nrs_req_add(svcpt, req, !!rc);
RETURN(0);
}
* User can call it w/o any lock but need to hold
* ptlrpc_service_part::scp_req_lock to get reliable result
*/
-static int ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
- int force)
+static bool ptlrpc_server_allow_high(struct ptlrpc_service_part *svcpt,
+ bool force)
{
+ int running = svcpt->scp_nthrs_running;
+
+ if (!nrs_svcpt_has_hp(svcpt))
+ return false;
+
if (force)
- return 1;
+ return true;
- if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
- return 0;
+ if (ptlrpc_nrs_req_throttling_nolock(svcpt, true))
+ return false;
+
+ 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 */
+ running = PTLRPC_NTHRS_INIT;
+ if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
+ running += 1;
+ }
- return cfs_list_empty(&svcpt->scp_req_pending) ||
+ if (svcpt->scp_nreqs_active >= running - 1)
+ return false;
+
+ if (svcpt->scp_nhreqs_active == 0)
+ return true;
+
+ return !ptlrpc_nrs_req_pending_nolock(svcpt, false) ||
svcpt->scp_hreq_count < svcpt->scp_service->srv_hpreq_ratio;
}
-static int ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
- int force)
+static bool ptlrpc_server_high_pending(struct ptlrpc_service_part *svcpt,
+ bool force)
{
return ptlrpc_server_allow_high(svcpt, force) &&
- !cfs_list_empty(&svcpt->scp_hreq_pending);
+ ptlrpc_nrs_req_pending_nolock(svcpt, true);
}
/**
* User can call it w/o any lock but need to hold
* ptlrpc_service_part::scp_req_lock to get reliable result
*/
-static int ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
- int force)
-{
-#ifndef __KERNEL__
- if (1) /* always allow to handle normal request for liblustre */
- return 1;
-#endif
- if (force ||
- svcpt->scp_nreqs_active < svcpt->scp_nthrs_running - 2)
- return 1;
-
- if (svcpt->scp_nreqs_active >= svcpt->scp_nthrs_running - 1)
- return 0;
+static bool ptlrpc_server_allow_normal(struct ptlrpc_service_part *svcpt,
+ 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 */
+ running = PTLRPC_NTHRS_INIT;
+ if (svcpt->scp_service->srv_ops.so_hpreq_handler != NULL)
+ running += 1;
+ }
+
+ if (force)
+ return true;
- return svcpt->scp_nhreqs_active > 0 ||
- svcpt->scp_service->srv_ops.so_hpreq_handler == NULL;
+ if (ptlrpc_nrs_req_throttling_nolock(svcpt, false))
+ return false;
+
+ if (svcpt->scp_nreqs_active < running - 2)
+ return true;
+
+ if (svcpt->scp_nreqs_active >= running - 1)
+ return false;
+
+ return svcpt->scp_nhreqs_active > 0 || !nrs_svcpt_has_hp(svcpt);
}
-static int ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
- int force)
+static bool ptlrpc_server_normal_pending(struct ptlrpc_service_part *svcpt,
+ bool force)
{
return ptlrpc_server_allow_normal(svcpt, force) &&
- !cfs_list_empty(&svcpt->scp_req_pending);
+ ptlrpc_nrs_req_pending_nolock(svcpt, false);
}
/**
* \see ptlrpc_server_allow_normal
* \see ptlrpc_server_allow high
*/
-static inline int
-ptlrpc_server_request_pending(struct ptlrpc_service_part *svcpt, int 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);
* Returns a pointer to fetched request.
*/
static struct ptlrpc_request *
-ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, int force)
+ptlrpc_server_request_get(struct ptlrpc_service_part *svcpt, bool force)
{
- struct ptlrpc_request *req;
+ struct ptlrpc_request *req = NULL;
ENTRY;
+ spin_lock(&svcpt->scp_req_lock);
+
if (ptlrpc_server_high_pending(svcpt, force)) {
- req = cfs_list_entry(svcpt->scp_hreq_pending.next,
- struct ptlrpc_request, rq_list);
- svcpt->scp_hreq_count++;
- RETURN(req);
+ req = ptlrpc_nrs_req_get_nolock(svcpt, true, force);
+ if (req != NULL) {
+ svcpt->scp_hreq_count++;
+ goto got_request;
+ }
}
if (ptlrpc_server_normal_pending(svcpt, force)) {
- req = cfs_list_entry(svcpt->scp_req_pending.next,
- struct ptlrpc_request, rq_list);
- svcpt->scp_hreq_count = 0;
- RETURN(req);
+ req = ptlrpc_nrs_req_get_nolock(svcpt, false, force);
+ if (req != NULL) {
+ svcpt->scp_hreq_count = 0;
+ goto got_request;
+ }
}
+
+ spin_unlock(&svcpt->scp_req_lock);
RETURN(NULL);
+
+got_request:
+ svcpt->scp_nreqs_active++;
+ if (req->rq_hp)
+ svcpt->scp_nhreqs_active++;
+
+ spin_unlock(&svcpt->scp_req_lock);
+
+ if (likely(req->rq_export))
+ class_export_rpc_inc(req->rq_export);
+
+ RETURN(req);
}
/**
* ptlrpc_server_handle_req later on.
*/
static int
-ptlrpc_server_handle_req_in(struct ptlrpc_service_part *svcpt)
+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;
int rc;
ENTRY;
- cfs_spin_lock(&svcpt->scp_lock);
- if (cfs_list_empty(&svcpt->scp_req_incoming)) {
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
+ if (list_empty(&svcpt->scp_req_incoming)) {
+ spin_unlock(&svcpt->scp_lock);
RETURN(0);
}
- req = cfs_list_entry(svcpt->scp_req_incoming.next,
+ req = list_entry(svcpt->scp_req_incoming.next,
struct ptlrpc_request, rq_list);
- cfs_list_del_init(&req->rq_list);
+ list_del_init(&req->rq_list);
svcpt->scp_nreqs_incoming--;
/* Consider this still a "queued" request as far as stats are
* concerned */
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
/* go through security check/transform */
rc = sptlrpc_svc_unwrap_request(req);
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;
}
- ptlrpc_at_add_timed(req);
+ /* Skip early reply */
+ if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_RESEND))
+ req->rq_deadline += obd_timeout;
+
+ req->rq_svc_thread = thread;
+ if (thread != NULL) {
+ /* 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) {
+ CERROR("%s: failure to initialize session: rc = %d\n",
+ thread->t_name, rc);
+ goto err_req;
+ }
+ req->rq_session.lc_thread = thread;
+ lu_context_enter(&req->rq_session);
+ thread->t_env->le_ses = &req->rq_session;
+ }
+
+ ptlrpc_at_add_timed(req);
- /* Move it over to the request processing queue */
+ /* Move it over to the request processing queue */
rc = ptlrpc_server_request_add(svcpt, req);
- if (rc) {
- ptlrpc_hpreq_fini(req);
+ if (rc)
GOTO(err_req, rc);
- }
- cfs_waitq_signal(&svcpt->scp_waitq);
+
+ wake_up(&svcpt->scp_waitq);
RETURN(1);
err_req:
- cfs_spin_lock(&svcpt->scp_req_lock);
- svcpt->scp_nreqs_active++;
- cfs_spin_unlock(&svcpt->scp_req_lock);
ptlrpc_server_finish_request(svcpt, req);
RETURN(1);
ptlrpc_server_handle_request(struct ptlrpc_service_part *svcpt,
struct ptlrpc_thread *thread)
{
- struct ptlrpc_service *svc = svcpt->scp_service;
- struct obd_export *export = NULL;
- struct ptlrpc_request *request;
- struct timeval work_start;
- struct timeval work_end;
- long timediff;
- int rc;
- int fail_opc = 0;
- ENTRY;
+ 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;
+
+ ENTRY;
- cfs_spin_lock(&svcpt->scp_req_lock);
-#ifndef __KERNEL__
- /* !@%$# liblustre only has 1 thread */
- if (cfs_atomic_read(&svcpt->scp_nreps_difficult) != 0) {
- cfs_spin_unlock(&svcpt->scp_req_lock);
+ request = ptlrpc_server_request_get(svcpt, false);
+ if (request == NULL)
RETURN(0);
- }
-#endif
- request = ptlrpc_server_request_get(svcpt, 0);
- if (request == NULL) {
- cfs_spin_unlock(&svcpt->scp_req_lock);
- RETURN(0);
- }
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
if (unlikely(fail_opc)) {
- if (request->rq_export && request->rq_ops) {
- cfs_spin_unlock(&svcpt->scp_req_lock);
-
+ if (request->rq_export && request->rq_ops)
OBD_FAIL_TIMEOUT(fail_opc, 4);
-
- cfs_spin_lock(&svcpt->scp_req_lock);
- request = ptlrpc_server_request_get(svcpt, 0);
- if (request == NULL) {
- cfs_spin_unlock(&svcpt->scp_req_lock);
- RETURN(0);
- }
- }
}
- cfs_list_del_init(&request->rq_list);
- svcpt->scp_nreqs_active++;
- if (request->rq_hp)
- svcpt->scp_nhreqs_active++;
-
- cfs_spin_unlock(&svcpt->scp_req_lock);
-
ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
- if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
- libcfs_debug_dumplog();
+ if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
+ libcfs_debug_dumplog();
- cfs_gettimeofday(&work_start);
- timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
- if (likely(svc->srv_stats != NULL)) {
+ do_gettimeofday(&work_start);
+ timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
+ 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,
at_get(&svcpt->scp_at_estimate));
}
- rc = lu_context_init(&request->rq_session, LCT_SESSION | LCT_NOREF);
- if (rc) {
- CERROR("Failure to initialize session: %d\n", rc);
- goto out_req;
- }
- request->rq_session.lc_thread = thread;
- request->rq_session.lc_cookie = 0x5;
- lu_context_enter(&request->rq_session);
-
- CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
-
- request->rq_svc_thread = thread;
- if (thread)
- request->rq_svc_thread->t_env->le_ses = &request->rq_session;
-
- if (likely(request->rq_export)) {
- if (unlikely(ptlrpc_check_req(request)))
- goto put_conn;
+ if (likely(request->rq_export)) {
+ if (unlikely(ptlrpc_check_req(request)))
+ goto put_conn;
ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
- export = class_export_rpc_get(request->rq_export);
}
/* Discard requests queued for longer than the deadline.
request->rq_arrival_time.tv_sec),
cfs_time_sub(cfs_time_current_sec(),
request->rq_deadline));
- goto put_rpc_export;
+ goto put_conn;
}
- CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
- "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
- (request->rq_export ?
- (char *)request->rq_export->exp_client_uuid.uuid : "0"),
- (request->rq_export ?
- cfs_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));
+ CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
+ "%s:%s+%d:%d:x"LPU64":%s:%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));
if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
CFS_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, cfs_fail_val);
- rc = svc->srv_ops.so_req_handler(request);
+ CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
+
+ /* re-assign request and sesson thread to the current one */
+ request->rq_svc_thread = thread;
+ if (thread != NULL) {
+ LASSERT(request->rq_session.lc_thread == NULL);
+ request->rq_session.lc_thread = thread;
+ thread->t_env->le_ses = &request->rq_session;
+ }
+ svc->srv_ops.so_req_handler(request);
- ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
+ ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
-put_rpc_export:
- if (export != NULL)
- class_export_rpc_put(export);
put_conn:
- lu_context_exit(&request->rq_session);
- lu_context_fini(&request->rq_session);
-
- if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
- DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
- "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
- " client may timeout.",
- request->rq_xid, cfs_time_sub(request->rq_deadline,
- request->rq_arrival_time.tv_sec),
- cfs_time_sub(cfs_time_current_sec(),
- request->rq_deadline));
- }
+ 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));
+ }
- cfs_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",
- cfs_curproc_comm(),
- (request->rq_export ?
- (char *)request->rq_export->exp_client_uuid.uuid : "0"),
+ 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 ?
- cfs_atomic_read(&request->rq_export->exp_refcount) : -99),
+ atomic_read(&request->rq_export->exp_refcount) : -99),
lustre_msg_get_status(request->rq_reqmsg),
request->rq_xid,
libcfs_id2str(request->rq_peer),
request->rq_arrival_time.tv_sec));
}
-out_req:
- ptlrpc_server_finish_request(svcpt, request);
+ ptlrpc_server_finish_active_request(svcpt, request);
RETURN(1);
}
int been_handled;
ENTRY;
- exp = rs->rs_export;
+ exp = rs->rs_export;
- LASSERT (rs->rs_difficult);
- LASSERT (rs->rs_scheduled);
- LASSERT (cfs_list_empty(&rs->rs_list));
+ LASSERT(rs->rs_difficult);
+ LASSERT(rs->rs_scheduled);
+ LASSERT(list_empty(&rs->rs_list));
- cfs_spin_lock (&exp->exp_lock);
- /* Noop if removed already */
- cfs_list_del_init (&rs->rs_exp_list);
- cfs_spin_unlock (&exp->exp_lock);
+ 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
* 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) {
- cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
- cfs_list_del_init(&rs->rs_obd_list);
- cfs_spin_unlock(&exp->exp_uncommitted_replies_lock);
- }
+ if (!rs->rs_committed) {
+ spin_lock(&exp->exp_uncommitted_replies_lock);
+ list_del_init(&rs->rs_obd_list);
+ spin_unlock(&exp->exp_uncommitted_replies_lock);
+ }
- cfs_spin_lock(&rs->rs_lock);
+ spin_lock(&rs->rs_lock);
been_handled = rs->rs_handled;
rs->rs_handled = 1;
if (nlocks == 0 && !been_handled) {
/* If we see this, we should already have seen the warning
* in mds_steal_ack_locks() */
- CWARN("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,
- libcfs_nid2str(exp->exp_connection->c_peer.nid));
+ 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,
+ libcfs_nid2str(exp->exp_connection->c_peer.nid));
}
if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
- cfs_spin_unlock(&rs->rs_lock);
+ spin_unlock(&rs->rs_lock);
- if (!been_handled && rs->rs_on_net) {
- LNetMDUnlink(rs->rs_md_h);
- /* Ignore return code; we're racing with
- * completion... */
- }
+ if (!been_handled && rs->rs_on_net) {
+ LNetMDUnlink(rs->rs_md_h);
+ /* Ignore return code; we're racing with completion */
+ }
- while (nlocks-- > 0)
- ldlm_lock_decref(&rs->rs_locks[nlocks],
- rs->rs_modes[nlocks]);
+ while (nlocks-- > 0)
+ ldlm_lock_decref(&rs->rs_locks[nlocks],
+ rs->rs_modes[nlocks]);
- cfs_spin_lock(&rs->rs_lock);
- }
+ spin_lock(&rs->rs_lock);
+ }
- rs->rs_scheduled = 0;
+ rs->rs_scheduled = 0;
- if (!rs->rs_on_net) {
- /* Off the net */
- cfs_spin_unlock(&rs->rs_lock);
+ if (!rs->rs_on_net) {
+ /* Off the net */
+ spin_unlock(&rs->rs_lock);
- class_export_put (exp);
- rs->rs_export = NULL;
- ptlrpc_rs_decref (rs);
- if (cfs_atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
+ class_export_put (exp);
+ rs->rs_export = NULL;
+ ptlrpc_rs_decref(rs);
+ if (atomic_dec_and_test(&svcpt->scp_nreps_difficult) &&
svc->srv_is_stopping)
- cfs_waitq_broadcast(&svcpt->scp_waitq);
+ wake_up_all(&svcpt->scp_waitq);
RETURN(1);
}
/* still on the net; callback will schedule */
- cfs_spin_unlock(&rs->rs_lock);
+ spin_unlock(&rs->rs_lock);
RETURN(1);
}
-#ifndef __KERNEL__
-
-/**
- * Check whether given service has a reply available for processing
- * and process it.
- *
- * \param svc a ptlrpc service
- * \retval 0 no replies processed
- * \retval 1 one reply processed
- */
-static int
-ptlrpc_server_handle_reply(struct ptlrpc_service_part *svcpt)
-{
- struct ptlrpc_reply_state *rs = NULL;
- ENTRY;
-
- cfs_spin_lock(&svcpt->scp_rep_lock);
- if (!cfs_list_empty(&svcpt->scp_rep_queue)) {
- rs = cfs_list_entry(svcpt->scp_rep_queue.prev,
- struct ptlrpc_reply_state,
- rs_list);
- cfs_list_del_init(&rs->rs_list);
- }
- cfs_spin_unlock(&svcpt->scp_rep_lock);
- if (rs != NULL)
- ptlrpc_handle_rs(rs);
- RETURN(rs != NULL);
-}
-
-/* FIXME make use of timeout later */
-int
-liblustre_check_services (void *arg)
-{
- int did_something = 0;
- int rc;
- cfs_list_t *tmp, *nxt;
- ENTRY;
-
- /* I'm relying on being single threaded, not to have to lock
- * ptlrpc_all_services etc */
- cfs_list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
- struct ptlrpc_service *svc =
- cfs_list_entry (tmp, struct ptlrpc_service, srv_list);
- struct ptlrpc_service_part *svcpt;
-
- LASSERT(svc->srv_ncpts == 1);
- svcpt = svc->srv_parts[0];
-
- if (svcpt->scp_nthrs_running != 0) /* I've recursed */
- continue;
-
- /* service threads can block for bulk, so this limits us
- * (arbitrarily) to recursing 1 stack frame per service.
- * Note that the problem with recursion is that we have to
- * unwind completely before our caller can resume. */
-
- svcpt->scp_nthrs_running++;
-
- do {
- rc = ptlrpc_server_handle_req_in(svcpt);
- rc |= ptlrpc_server_handle_reply(svcpt);
- rc |= ptlrpc_at_check_timed(svcpt);
- rc |= ptlrpc_server_handle_request(svcpt, NULL);
- rc |= (ptlrpc_server_post_idle_rqbds(svcpt) > 0);
- did_something |= rc;
- } while (rc);
-
- svcpt->scp_nthrs_running--;
- }
-
- RETURN(did_something);
-}
-#define ptlrpc_stop_all_threads(s) do {} while (0)
-
-#else /* __KERNEL__ */
static void
ptlrpc_check_rqbd_pool(struct ptlrpc_service_part *svcpt)
static inline int
ptlrpc_rqbd_pending(struct ptlrpc_service_part *svcpt)
{
- return !cfs_list_empty(&svcpt->scp_rqbd_idle) &&
+ return !list_empty(&svcpt->scp_rqbd_idle) &&
svcpt->scp_rqbd_timeout == 0;
}
static inline int
ptlrpc_server_request_incoming(struct ptlrpc_service_part *svcpt)
{
- return !cfs_list_empty(&svcpt->scp_req_incoming);
+ return !list_empty(&svcpt->scp_req_incoming);
}
static __attribute__((__noinline__)) int
lc_watchdog_disable(thread->t_watchdog);
- cfs_cond_resched();
+ cond_resched();
l_wait_event_exclusive_head(svcpt->scp_waitq,
ptlrpc_thread_stopping(thread) ||
ptlrpc_server_request_incoming(svcpt) ||
- ptlrpc_server_request_pending(svcpt, 0) ||
+ ptlrpc_server_request_pending(svcpt, false) ||
ptlrpc_rqbd_pending(svcpt) ||
ptlrpc_at_check(svcpt), &lwi);
struct ptlrpc_service_part *svcpt = thread->t_svcpt;
struct ptlrpc_service *svc = svcpt->scp_service;
struct ptlrpc_reply_state *rs;
-#ifdef WITH_GROUP_INFO
- cfs_group_info_t *ginfo = NULL;
-#endif
- struct lu_env *env;
- int counter = 0, rc = 0;
- ENTRY;
+ struct group_info *ginfo = NULL;
+ struct lu_env *env;
+ int counter = 0, rc = 0;
+ ENTRY;
- thread->t_pid = cfs_curproc_pid();
- cfs_daemonize_ctxt(thread->t_name);
+ 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,
svc->srv_name, thread->t_name, svcpt->scp_cpt);
}
-#ifdef WITH_GROUP_INFO
- ginfo = cfs_groups_alloc(0);
- if (!ginfo) {
- rc = -ENOMEM;
- goto out;
- }
+ ginfo = groups_alloc(0);
+ if (!ginfo) {
+ rc = -ENOMEM;
+ goto out;
+ }
- cfs_set_current_groups(ginfo);
- cfs_put_group_info(ginfo);
-#endif
+ set_current_groups(ginfo);
+ put_group_info(ginfo);
if (svc->srv_ops.so_thr_init != NULL) {
rc = svc->srv_ops.so_thr_init(thread);
env->le_ctx.lc_thread = thread;
env->le_ctx.lc_cookie = 0x6;
- while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
+ while (!list_empty(&svcpt->scp_rqbd_idle)) {
rc = ptlrpc_server_post_idle_rqbds(svcpt);
if (rc >= 0)
continue;
goto out_srv_fini;
}
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
LASSERT(thread_is_starting(thread));
thread_clear_flags(thread, SVC_STARTING);
* we are now running, however we will exit as soon as possible */
thread_add_flags(thread, SVC_RUNNING);
svcpt->scp_nthrs_running++;
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
/* wake up our creator in case he's still waiting. */
- cfs_waitq_signal(&thread->t_ctl_waitq);
+ wake_up(&thread->t_ctl_waitq);
thread->t_watchdog = lc_watchdog_add(ptlrpc_server_get_timeout(svcpt),
NULL, NULL);
- cfs_spin_lock(&svcpt->scp_rep_lock);
- cfs_list_add(&rs->rs_list, &svcpt->scp_rep_idle);
- cfs_waitq_signal(&svcpt->scp_rep_waitq);
- cfs_spin_unlock(&svcpt->scp_rep_lock);
+ spin_lock(&svcpt->scp_rep_lock);
+ list_add(&rs->rs_list, &svcpt->scp_rep_idle);
+ wake_up(&svcpt->scp_rep_waitq);
+ spin_unlock(&svcpt->scp_rep_lock);
CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
svcpt->scp_nthrs_running);
ptlrpc_start_thread(svcpt, 0);
}
+ /* reset le_ses to initial state */
+ env->le_ses = NULL;
/* Process all incoming reqs before handling any */
if (ptlrpc_server_request_incoming(svcpt)) {
- ptlrpc_server_handle_req_in(svcpt);
+ lu_context_enter(&env->le_ctx);
+ ptlrpc_server_handle_req_in(svcpt, thread);
+ lu_context_exit(&env->le_ctx);
+
/* but limit ourselves in case of flood */
if (counter++ < 100)
continue;
if (ptlrpc_at_check(svcpt))
ptlrpc_at_check_timed(svcpt);
- if (ptlrpc_server_request_pending(svcpt, 0)) {
+ if (ptlrpc_server_request_pending(svcpt, false)) {
lu_context_enter(&env->le_ctx);
ptlrpc_server_handle_request(svcpt, thread);
lu_context_exit(&env->le_ctx);
CDEBUG(D_RPCTRACE, "service thread [ %p : %u ] %d exiting: rc %d\n",
thread, thread->t_pid, thread->t_id, rc);
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
if (thread_test_and_clear_flags(thread, SVC_STARTING))
svcpt->scp_nthrs_starting--;
thread->t_id = rc;
thread_add_flags(thread, SVC_STOPPED);
- cfs_waitq_signal(&thread->t_ctl_waitq);
- cfs_spin_unlock(&svcpt->scp_lock);
+ wake_up(&thread->t_ctl_waitq);
+ spin_unlock(&svcpt->scp_lock);
return rc;
}
-struct ptlrpc_hr_args {
- int thread_index;
- int cpu_index;
- struct ptlrpc_hr_service *hrs;
-};
-
-static int hrt_dont_sleep(struct ptlrpc_hr_thread *t,
- cfs_list_t *replies)
+static int hrt_dont_sleep(struct ptlrpc_hr_thread *hrt,
+ struct list_head *replies)
{
- int result;
+ int result;
+
+ spin_lock(&hrt->hrt_lock);
- cfs_spin_lock(&t->hrt_lock);
- cfs_list_splice_init(&t->hrt_queue, replies);
- result = cfs_test_bit(HRT_STOPPING, &t->hrt_flags) ||
- !cfs_list_empty(replies);
- cfs_spin_unlock(&t->hrt_lock);
- return result;
+ list_splice_init(&hrt->hrt_queue, replies);
+ result = ptlrpc_hr.hr_stopping || !list_empty(replies);
+
+ spin_unlock(&hrt->hrt_lock);
+ return result;
}
/**
*/
static int ptlrpc_hr_main(void *arg)
{
- struct ptlrpc_hr_args * hr_args = arg;
- struct ptlrpc_hr_service *hr = hr_args->hrs;
- struct ptlrpc_hr_thread *t = &hr->hr_threads[hr_args->thread_index];
- char threadname[20];
- CFS_LIST_HEAD(replies);
-
- snprintf(threadname, sizeof(threadname),
- "ptlrpc_hr_%d", hr_args->thread_index);
-
- cfs_daemonize_ctxt(threadname);
-#if defined(CONFIG_NUMA) && defined(HAVE_NODE_TO_CPUMASK)
- cfs_set_cpus_allowed(cfs_current(),
- node_to_cpumask(cpu_to_node(hr_args->cpu_index)));
-#endif
- cfs_set_bit(HRT_RUNNING, &t->hrt_flags);
- cfs_waitq_signal(&t->hrt_wait);
-
- while (!cfs_test_bit(HRT_STOPPING, &t->hrt_flags)) {
-
- l_wait_condition(t->hrt_wait, hrt_dont_sleep(t, &replies));
- while (!cfs_list_empty(&replies)) {
- struct ptlrpc_reply_state *rs;
-
- rs = cfs_list_entry(replies.prev,
- struct ptlrpc_reply_state,
- rs_list);
- cfs_list_del_init(&rs->rs_list);
- ptlrpc_handle_rs(rs);
- }
- }
+ 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;
- cfs_clear_bit(HRT_RUNNING, &t->hrt_flags);
- cfs_complete(&t->hrt_completion);
+ INIT_LIST_HEAD(&replies);
+ snprintf(threadname, sizeof(threadname), "ptlrpc_hr%02d_%03d",
+ hrp->hrp_cpt, hrt->hrt_id);
+ unshare_fs_struct();
- return 0;
-}
+ rc = cfs_cpt_bind(ptlrpc_hr.hr_cpt_table, hrp->hrp_cpt);
+ if (rc != 0) {
+ 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);
+ }
-static int ptlrpc_start_hr_thread(struct ptlrpc_hr_service *hr, int n, int cpu)
-{
- struct ptlrpc_hr_thread *t = &hr->hr_threads[n];
- struct ptlrpc_hr_args args;
- int rc;
- ENTRY;
+ atomic_inc(&hrp->hrp_nstarted);
+ wake_up(&ptlrpc_hr.hr_waitq);
- args.thread_index = n;
- args.cpu_index = cpu;
- args.hrs = hr;
+ while (!ptlrpc_hr.hr_stopping) {
+ l_wait_condition(hrt->hrt_waitq, hrt_dont_sleep(hrt, &replies));
- rc = cfs_create_thread(ptlrpc_hr_main, (void*)&args, CFS_DAEMON_FLAGS);
- if (rc < 0) {
- cfs_complete(&t->hrt_completion);
- GOTO(out, rc);
- }
- l_wait_condition(t->hrt_wait, cfs_test_bit(HRT_RUNNING, &t->hrt_flags));
- RETURN(0);
- out:
- return rc;
-}
+ while (!list_empty(&replies)) {
+ struct ptlrpc_reply_state *rs;
-static void ptlrpc_stop_hr_thread(struct ptlrpc_hr_thread *t)
-{
- ENTRY;
+ rs = list_entry(replies.prev,
+ struct ptlrpc_reply_state,
+ rs_list);
+ list_del_init(&rs->rs_list);
+ ptlrpc_handle_rs(rs);
+ }
+ }
- cfs_set_bit(HRT_STOPPING, &t->hrt_flags);
- cfs_waitq_signal(&t->hrt_wait);
- cfs_wait_for_completion(&t->hrt_completion);
+ atomic_inc(&hrp->hrp_nstopped);
+ wake_up(&ptlrpc_hr.hr_waitq);
- EXIT;
+ return 0;
}
-static void ptlrpc_stop_hr_threads(struct ptlrpc_hr_service *hrs)
+static void ptlrpc_stop_hr_threads(void)
{
- int n;
- ENTRY;
+ struct ptlrpc_hr_partition *hrp;
+ int i;
+ int j;
- for (n = 0; n < hrs->hr_n_threads; n++)
- ptlrpc_stop_hr_thread(&hrs->hr_threads[n]);
+ ptlrpc_hr.hr_stopping = 1;
- EXIT;
+ cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+ if (hrp->hrp_thrs == NULL)
+ continue; /* uninitialized */
+ for (j = 0; j < hrp->hrp_nthrs; j++)
+ wake_up_all(&hrp->hrp_thrs[j].hrt_waitq);
+ }
+
+ cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+ if (hrp->hrp_thrs == NULL)
+ continue; /* uninitialized */
+ wait_event(ptlrpc_hr.hr_waitq,
+ atomic_read(&hrp->hrp_nstopped) ==
+ atomic_read(&hrp->hrp_nstarted));
+ }
}
-static int ptlrpc_start_hr_threads(struct ptlrpc_hr_service *hr)
+static int ptlrpc_start_hr_threads(void)
{
- int rc = -ENOMEM;
- int n, cpu, threads_started = 0;
- ENTRY;
+ struct ptlrpc_hr_partition *hrp;
+ int i;
+ int j;
+ ENTRY;
+
+ cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+ int rc = 0;
+
+ for (j = 0; j < hrp->hrp_nthrs; j++) {
+ struct ptlrpc_hr_thread *hrt = &hrp->hrp_thrs[j];
+ struct task_struct *task;
+
+ task = kthread_run(ptlrpc_hr_main,
+ &hrp->hrp_thrs[j],
+ "ptlrpc_hr%02d_%03d",
+ hrp->hrp_cpt,
+ hrt->hrt_id);
+ if (IS_ERR(task)) {
+ rc = PTR_ERR(task);
+ break;
+ }
+ }
- LASSERT(hr != NULL);
- LASSERT(hr->hr_n_threads > 0);
+ wait_event(ptlrpc_hr.hr_waitq,
+ atomic_read(&hrp->hrp_nstarted) == j);
- for (n = 0, cpu = 0; n < hr->hr_n_threads; n++) {
-#if defined(CONFIG_SMP) && defined(HAVE_NODE_TO_CPUMASK)
- while (!cpu_online(cpu)) {
- cpu++;
- if (cpu >= cfs_num_possible_cpus())
- cpu = 0;
- }
-#endif
- rc = ptlrpc_start_hr_thread(hr, n, cpu);
- if (rc != 0)
- break;
- threads_started++;
- cpu++;
- }
- if (threads_started == 0) {
- CERROR("No reply handling threads started\n");
- RETURN(-ESRCH);
- }
- if (threads_started < hr->hr_n_threads) {
- CWARN("Started only %d reply handling threads from %d\n",
- threads_started, hr->hr_n_threads);
- hr->hr_n_threads = threads_started;
- }
- RETURN(0);
+ if (rc < 0) {
+ CERROR("cannot start reply handler thread %d:%d: "
+ "rc = %d\n", i, j, rc);
+ ptlrpc_stop_hr_threads();
+ RETURN(rc);
+ }
+ }
+
+ RETURN(0);
}
static void ptlrpc_svcpt_stop_threads(struct ptlrpc_service_part *svcpt)
{
struct l_wait_info lwi = { 0 };
struct ptlrpc_thread *thread;
- CFS_LIST_HEAD (zombie);
+ struct list_head zombie;
ENTRY;
CDEBUG(D_INFO, "Stopping threads for service %s\n",
svcpt->scp_service->srv_name);
- cfs_spin_lock(&svcpt->scp_lock);
+ 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",
thread_add_flags(thread, SVC_STOPPING);
}
- cfs_waitq_broadcast(&svcpt->scp_waitq);
+ wake_up_all(&svcpt->scp_waitq);
- while (!cfs_list_empty(&svcpt->scp_threads)) {
- thread = cfs_list_entry(svcpt->scp_threads.next,
+ while (!list_empty(&svcpt->scp_threads)) {
+ thread = list_entry(svcpt->scp_threads.next,
struct ptlrpc_thread, t_link);
if (thread_is_stopped(thread)) {
- cfs_list_del(&thread->t_link);
- cfs_list_add(&thread->t_link, &zombie);
+ list_del(&thread->t_link);
+ list_add(&thread->t_link, &zombie);
continue;
}
- cfs_spin_unlock(&svcpt->scp_lock);
+ 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);
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
}
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
- while (!cfs_list_empty(&zombie)) {
- thread = cfs_list_entry(zombie.next,
+ while (!list_empty(&zombie)) {
+ thread = list_entry(zombie.next,
struct ptlrpc_thread, t_link);
- cfs_list_del(&thread->t_link);
+ list_del(&thread->t_link);
OBD_FREE_PTR(thread);
}
EXIT;
{
struct l_wait_info lwi = { 0 };
struct ptlrpc_thread *thread;
- struct ptlrpc_service *svc = svcpt->scp_service;
+ struct ptlrpc_service *svc;
+ struct task_struct *task;
int rc;
ENTRY;
LASSERT(svcpt != NULL);
+ svc = svcpt->scp_service;
+
CDEBUG(D_RPCTRACE, "%s[%d] started %d min %d max %d\n",
svc->srv_name, svcpt->scp_cpt, svcpt->scp_nthrs_running,
svc->srv_nthrs_cpt_init, svc->srv_nthrs_cpt_limit);
OBD_CPT_ALLOC_PTR(thread, svc->srv_cptable, svcpt->scp_cpt);
if (thread == NULL)
RETURN(-ENOMEM);
- cfs_waitq_init(&thread->t_ctl_waitq);
+ init_waitqueue_head(&thread->t_ctl_waitq);
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
if (!ptlrpc_threads_increasable(svcpt)) {
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
OBD_FREE_PTR(thread);
RETURN(-EMFILE);
}
/* serialize starting because some modules (obdfilter)
* might require unique and contiguous t_id */
LASSERT(svcpt->scp_nthrs_starting == 1);
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
OBD_FREE_PTR(thread);
if (wait) {
CDEBUG(D_INFO, "Waiting for creating thread %s #%d\n",
svc->srv_thread_name, svcpt->scp_thr_nextid);
- cfs_schedule();
+ schedule();
goto again;
}
thread_add_flags(thread, SVC_STARTING);
thread->t_svcpt = svcpt;
- cfs_list_add(&thread->t_link, &svcpt->scp_threads);
- cfs_spin_unlock(&svcpt->scp_lock);
+ list_add(&thread->t_link, &svcpt->scp_threads);
+ spin_unlock(&svcpt->scp_lock);
if (svcpt->scp_cpt >= 0) {
snprintf(thread->t_name, PTLRPC_THR_NAME_LEN, "%s%02d_%03d",
}
CDEBUG(D_RPCTRACE, "starting thread '%s'\n", thread->t_name);
- /*
- * CLONE_VM and CLONE_FILES just avoid a needless copy, because we
- * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
- */
- rc = cfs_create_thread(ptlrpc_main, thread, CFS_DAEMON_FLAGS);
- if (rc < 0) {
- CERROR("cannot start thread '%s': rc %d\n",
+ task = kthread_run(ptlrpc_main, thread, "%s", thread->t_name);
+ if (IS_ERR(task)) {
+ rc = PTR_ERR(task);
+ CERROR("cannot start thread '%s': rc = %d\n",
thread->t_name, rc);
- cfs_spin_lock(&svcpt->scp_lock);
- cfs_list_del(&thread->t_link);
+ spin_lock(&svcpt->scp_lock);
--svcpt->scp_nthrs_starting;
- cfs_spin_unlock(&svcpt->scp_lock);
-
- OBD_FREE(thread, sizeof(*thread));
+ if (thread_is_stopping(thread)) {
+ /* this ptlrpc_thread is being hanled
+ * by ptlrpc_svcpt_stop_threads now
+ */
+ thread_add_flags(thread, SVC_STOPPED);
+ wake_up(&thread->t_ctl_waitq);
+ spin_unlock(&svcpt->scp_lock);
+ } else {
+ list_del(&thread->t_link);
+ spin_unlock(&svcpt->scp_lock);
+ OBD_FREE_PTR(thread);
+ }
RETURN(rc);
}
int ptlrpc_hr_init(void)
{
- int i;
- int n_cpus = cfs_num_online_cpus();
- struct ptlrpc_hr_service *hr;
- int size;
- int rc;
- ENTRY;
+ struct ptlrpc_hr_partition *hrp;
+ struct ptlrpc_hr_thread *hrt;
+ int rc;
+ int i;
+ int j;
+ int weight;
+ ENTRY;
- LASSERT(ptlrpc_hr == NULL);
+ memset(&ptlrpc_hr, 0, sizeof(ptlrpc_hr));
+ ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
- size = offsetof(struct ptlrpc_hr_service, hr_threads[n_cpus]);
- OBD_ALLOC(hr, size);
- if (hr == NULL)
- RETURN(-ENOMEM);
- for (i = 0; i < n_cpus; i++) {
- struct ptlrpc_hr_thread *t = &hr->hr_threads[i];
+ ptlrpc_hr.hr_partitions = cfs_percpt_alloc(ptlrpc_hr.hr_cpt_table,
+ sizeof(*hrp));
+ if (ptlrpc_hr.hr_partitions == NULL)
+ RETURN(-ENOMEM);
- cfs_spin_lock_init(&t->hrt_lock);
- cfs_waitq_init(&t->hrt_wait);
- CFS_INIT_LIST_HEAD(&t->hrt_queue);
- cfs_init_completion(&t->hrt_completion);
- }
- hr->hr_n_threads = n_cpus;
- hr->hr_size = size;
- ptlrpc_hr = hr;
+ init_waitqueue_head(&ptlrpc_hr.hr_waitq);
- rc = ptlrpc_start_hr_threads(hr);
- if (rc) {
- OBD_FREE(hr, hr->hr_size);
- ptlrpc_hr = NULL;
- }
- RETURN(rc);
+ weight = cfs_cpu_ht_nsiblings(0);
+
+ cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+ hrp->hrp_cpt = i;
+
+ 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 /= weight;
+
+ LASSERT(hrp->hrp_nthrs > 0);
+ OBD_CPT_ALLOC(hrp->hrp_thrs, ptlrpc_hr.hr_cpt_table, i,
+ 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];
+
+ hrt->hrt_id = j;
+ hrt->hrt_partition = hrp;
+ init_waitqueue_head(&hrt->hrt_waitq);
+ spin_lock_init(&hrt->hrt_lock);
+ INIT_LIST_HEAD(&hrt->hrt_queue);
+ }
+ }
+
+ rc = ptlrpc_start_hr_threads();
+out:
+ if (rc != 0)
+ ptlrpc_hr_fini();
+ RETURN(rc);
}
void ptlrpc_hr_fini(void)
{
- if (ptlrpc_hr != NULL) {
- ptlrpc_stop_hr_threads(ptlrpc_hr);
- OBD_FREE(ptlrpc_hr, ptlrpc_hr->hr_size);
- ptlrpc_hr = NULL;
- }
+ struct ptlrpc_hr_partition *hrp;
+ int i;
+
+ if (ptlrpc_hr.hr_partitions == NULL)
+ return;
+
+ ptlrpc_stop_hr_threads();
+
+ cfs_percpt_for_each(hrp, i, ptlrpc_hr.hr_partitions) {
+ if (hrp->hrp_thrs != NULL) {
+ OBD_FREE(hrp->hrp_thrs,
+ hrp->hrp_nthrs * sizeof(hrp->hrp_thrs[0]));
+ }
+ }
+
+ cfs_percpt_free(ptlrpc_hr.hr_partitions);
+ ptlrpc_hr.hr_partitions = NULL;
}
-#endif /* __KERNEL__ */
/**
* Wait until all already scheduled replies are processed.
NULL, NULL);
rc = l_wait_event(svcpt->scp_waitq,
- cfs_atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
+ atomic_read(&svcpt->scp_nreps_difficult) == 0, &lwi);
if (rc == 0)
break;
CWARN("Unexpectedly long timeout %s %p\n",
/* Unlink all the request buffers. This forces a 'final'
* event with its 'unlink' flag set for each posted rqbd */
- cfs_list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
+ list_for_each_entry(rqbd, &svcpt->scp_rqbd_posted,
rqbd_list) {
rc = LNetMDUnlink(rqbd->rqbd_md_h);
LASSERT(rc == 0 || rc == -ENOENT);
/* Wait for the network to release any buffers
* it's currently filling */
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
while (svcpt->scp_nrqbds_posted != 0) {
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
/* Network access will complete in finite time but
* the HUGE timeout lets us CWARN for visibility
* of sluggish NALs */
"request buffers\n",
svcpt->scp_service->srv_name);
}
- cfs_spin_lock(&svcpt->scp_lock);
+ spin_lock(&svcpt->scp_lock);
}
- cfs_spin_unlock(&svcpt->scp_lock);
+ spin_unlock(&svcpt->scp_lock);
}
}
if (svcpt->scp_service == NULL)
break;
- cfs_spin_lock(&svcpt->scp_rep_lock);
- while (!cfs_list_empty(&svcpt->scp_rep_active)) {
- rs = cfs_list_entry(svcpt->scp_rep_active.next,
+ spin_lock(&svcpt->scp_rep_lock);
+ while (!list_empty(&svcpt->scp_rep_active)) {
+ rs = list_entry(svcpt->scp_rep_active.next,
struct ptlrpc_reply_state, rs_list);
- cfs_spin_lock(&rs->rs_lock);
+ spin_lock(&rs->rs_lock);
ptlrpc_schedule_difficult_reply(rs);
- cfs_spin_unlock(&rs->rs_lock);
+ spin_unlock(&rs->rs_lock);
}
- cfs_spin_unlock(&svcpt->scp_rep_lock);
+ spin_unlock(&svcpt->scp_rep_lock);
/* 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 */
- while (!cfs_list_empty(&svcpt->scp_req_incoming)) {
- req = cfs_list_entry(svcpt->scp_req_incoming.next,
+ while (!list_empty(&svcpt->scp_req_incoming)) {
+ req = list_entry(svcpt->scp_req_incoming.next,
struct ptlrpc_request, rq_list);
- cfs_list_del(&req->rq_list);
+ list_del(&req->rq_list);
svcpt->scp_nreqs_incoming--;
- svcpt->scp_nreqs_active++;
ptlrpc_server_finish_request(svcpt, req);
}
- while (ptlrpc_server_request_pending(svcpt, 1)) {
- req = ptlrpc_server_request_get(svcpt, 1);
- cfs_list_del(&req->rq_list);
- svcpt->scp_nreqs_active++;
- ptlrpc_hpreq_fini(req);
- ptlrpc_server_finish_request(svcpt, req);
+ while (ptlrpc_server_request_pending(svcpt, true)) {
+ req = ptlrpc_server_request_get(svcpt, true);
+ ptlrpc_server_finish_active_request(svcpt, req);
}
- LASSERT(cfs_list_empty(&svcpt->scp_rqbd_posted));
+ 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
/* Now free all the request buffers since nothing
* references them any more... */
- while (!cfs_list_empty(&svcpt->scp_rqbd_idle)) {
- rqbd = cfs_list_entry(svcpt->scp_rqbd_idle.next,
+ while (!list_empty(&svcpt->scp_rqbd_idle)) {
+ rqbd = list_entry(svcpt->scp_rqbd_idle.next,
struct ptlrpc_request_buffer_desc,
rqbd_list);
ptlrpc_free_rqbd(rqbd);
}
ptlrpc_wait_replies(svcpt);
- while (!cfs_list_empty(&svcpt->scp_rep_idle)) {
- rs = cfs_list_entry(svcpt->scp_rep_idle.next,
+ while (!list_empty(&svcpt->scp_rep_idle)) {
+ rs = list_entry(svcpt->scp_rep_idle.next,
struct ptlrpc_reply_state,
rs_list);
- cfs_list_del(&rs->rs_list);
+ list_del(&rs->rs_list);
OBD_FREE_LARGE(rs, svc->srv_max_reply_size);
}
}
if (array->paa_reqs_array != NULL) {
OBD_FREE(array->paa_reqs_array,
- sizeof(cfs_list_t) * array->paa_size);
+ sizeof(struct list_head) * array->paa_size);
array->paa_reqs_array = NULL;
}
service->srv_is_stopping = 1;
- cfs_spin_lock(&ptlrpc_all_services_lock);
- cfs_list_del_init(&service->srv_list);
- cfs_spin_unlock(&ptlrpc_all_services_lock);
-
- ptlrpc_lprocfs_unregister_service(service);
+ mutex_lock(&ptlrpc_all_services_mutex);
+ list_del_init(&service->srv_list);
+ mutex_unlock(&ptlrpc_all_services_mutex);
ptlrpc_service_del_atimer(service);
ptlrpc_stop_all_threads(service);
ptlrpc_service_unlink_rqbd(service);
ptlrpc_service_purge_all(service);
+ ptlrpc_service_nrs_cleanup(service);
+
+ ptlrpc_lprocfs_unregister_service(service);
+
ptlrpc_service_free(service);
RETURN(0);
}
+EXPORT_SYMBOL(ptlrpc_unregister_service);
/**
* Returns 0 if the service is healthy.
* 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. */
-int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
+static int ptlrpc_svcpt_health_check(struct ptlrpc_service_part *svcpt)
{
- struct ptlrpc_request *request;
+ struct ptlrpc_request *request = NULL;
struct timeval right_now;
long timediff;
- cfs_gettimeofday(&right_now);
-
- cfs_spin_lock(&svcpt->scp_req_lock);
- if (!ptlrpc_server_request_pending(svcpt, 1)) {
- cfs_spin_unlock(&svcpt->scp_req_lock);
- return 0;
- }
+ do_gettimeofday(&right_now);
+ spin_lock(&svcpt->scp_req_lock);
/* How long has the next entry been waiting? */
- if (cfs_list_empty(&svcpt->scp_req_pending)) {
- request = cfs_list_entry(svcpt->scp_hreq_pending.next,
- struct ptlrpc_request, rq_list);
- } else {
- request = cfs_list_entry(svcpt->scp_req_pending.next,
- struct ptlrpc_request, rq_list);
+ if (ptlrpc_server_high_pending(svcpt, true))
+ request = ptlrpc_nrs_req_peek_nolock(svcpt, true);
+ else if (ptlrpc_server_normal_pending(svcpt, true))
+ request = ptlrpc_nrs_req_peek_nolock(svcpt, false);
+
+ if (request == NULL) {
+ spin_unlock(&svcpt->scp_req_lock);
+ return 0;
}
timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
- cfs_spin_unlock(&svcpt->scp_req_lock);
+ spin_unlock(&svcpt->scp_req_lock);
if ((timediff / ONE_MILLION) >
(AT_OFF ? obd_timeout * 3 / 2 : at_max)) {
struct ptlrpc_service_part *svcpt;
int i;
- if (svc == NULL || svc->srv_parts == NULL)
+ if (svc == NULL)
return 0;
ptlrpc_service_for_each_part(svcpt, i, svc) {
}
return 0;
}
+EXPORT_SYMBOL(ptlrpc_service_health_check);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff