module_param(ldlm_num_threads, int, 0444);
MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
+static unsigned int ldlm_cpu_bind = 1;
+module_param(ldlm_cpu_bind, uint, 0444);
+MODULE_PARM_DESC(ldlm_cpu_bind,
+ "bind DLM service threads to particular CPU partitions");
+
static char *ldlm_cpts;
module_param(ldlm_cpts, charp, 0444);
MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
* All access to it should be under waiting_locks_spinlock.
*/
static LIST_HEAD(waiting_locks_list);
-static void waiting_locks_callback(cfs_timer_cb_arg_t unused);
+static void waiting_locks_callback(TIMER_DATA_TYPE unused);
static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
enum elt_state {
}
/* This is called from within a timer interrupt and cannot schedule */
-static void waiting_locks_callback(cfs_timer_cb_arg_t unused)
+static void waiting_locks_callback(TIMER_DATA_TYPE unused)
{
struct ldlm_lock *lock;
int need_dump = 0;
/**
* Perform lock cleanup if AST reply came with error.
*/
-static int ldlm_handle_ast_error(struct ldlm_lock *lock,
+static int ldlm_handle_ast_error(const struct lu_env *env,
+ struct ldlm_lock *lock,
struct ptlrpc_request *req, int rc,
const char *ast_type)
{
/* update lvbo to return proper attributes.
* see bug 23174 */
ldlm_resource_getref(res);
- ldlm_lvbo_update(res, lock, NULL, 1);
+ ldlm_lvbo_update(env, res, lock, NULL, 1);
ldlm_resource_putref(res);
}
ldlm_lock_cancel(lock);
}
static int ldlm_cb_interpret(const struct lu_env *env,
- struct ptlrpc_request *req, void *data, int rc)
+ struct ptlrpc_request *req, void *args, int rc)
{
- struct ldlm_cb_async_args *ca = data;
- struct ldlm_lock *lock = ca->ca_lock;
- struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
- ENTRY;
+ struct ldlm_cb_async_args *ca = args;
+ struct ldlm_lock *lock = ca->ca_lock;
+ struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
- LASSERT(lock != NULL);
+ ENTRY;
+
+ LASSERT(lock != NULL);
switch (arg->type) {
case LDLM_GL_CALLBACK:
* -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
*/
if (unlikely(arg->gl_interpret_reply)) {
- rc = arg->gl_interpret_reply(env, req, data, rc);
+ rc = arg->gl_interpret_reply(env, req, args, rc);
} else if (rc == -ELDLM_NO_LOCK_DATA) {
LDLM_DEBUG(lock, "lost race - client has a lock but no "
"inode");
- ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
+ ldlm_lvbo_update(env, lock->l_resource, lock, NULL, 1);
} else if (rc != 0) {
- rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
+ rc = ldlm_handle_ast_error(env, lock, req,
+ rc, "glimpse");
} else {
- rc = ldlm_lvbo_update(lock->l_resource, lock, req, 1);
+ rc = ldlm_lvbo_update(env, lock->l_resource,
+ lock, req, 1);
}
break;
case LDLM_BL_CALLBACK:
if (rc != 0)
- rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
+ rc = ldlm_handle_ast_error(env, lock, req,
+ rc, "blocking");
break;
case LDLM_CP_CALLBACK:
if (rc != 0)
- rc = ldlm_handle_ast_error(lock, req, rc, "completion");
+ rc = ldlm_handle_ast_error(env, lock, req,
+ rc, "completion");
break;
default:
LDLM_ERROR(lock, "invalid opcode for lock callback %d",
ldlm_lock2desc(lock, &body->lock_desc);
if (lvb_len > 0) {
void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
+ const struct lu_env *env = NULL;
- lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
+ if (req->rq_svc_thread)
+ env = req->rq_svc_thread->t_env;
+
+ lvb_len = ldlm_lvbo_fill(env, lock, lvb, &lvb_len);
if (lvb_len < 0) {
/* We still need to send the RPC to wake up the blocked
* enqueue thread on the client.
ca->ca_set_arg = arg;
ca->ca_lock = lock;
- /* server namespace, doesn't need lock */
- req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
- ldlm_lvbo_size(lock));
- ptlrpc_request_set_replen(req);
+ /* server namespace, doesn't need lock */
+ req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
+ ldlm_lvbo_size(lock));
+ ptlrpc_request_set_replen(req);
- req->rq_send_state = LUSTRE_IMP_FULL;
- /* ptlrpc_request_alloc_pack already set timeout */
- if (AT_OFF)
- req->rq_timeout = ldlm_get_rq_timeout();
+ req->rq_send_state = LUSTRE_IMP_FULL;
+ /* ptlrpc_request_alloc_pack already set timeout */
+ if (AT_OFF)
+ req->rq_timeout = ldlm_get_rq_timeout();
req->rq_interpret_reply = ldlm_cb_interpret;
- if (lock->l_export && lock->l_export->exp_nid_stats &&
- lock->l_export->exp_nid_stats->nid_ldlm_stats)
- lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
- LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
+ if (lock->l_export && lock->l_export->exp_nid_stats) {
+ struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
+
+ lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
+ LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
+ }
rc = ldlm_ast_fini(req, arg, lock, 0);
void *cookie = NULL;
int rc = 0;
struct ldlm_resource *res = NULL;
+ const struct lu_env *env = req->rq_svc_thread->t_env;
ENTRY;
LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
res = lock->l_resource;
if (!(flags & LDLM_FL_REPLAY)) {
/* non-replayed lock, delayed lvb init may need to be done */
- rc = ldlm_lvbo_init(res);
+ rc = ldlm_lvbo_init(env, res);
if (rc < 0) {
LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
GOTO(out, rc);
req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
RCL_SERVER, ldlm_lvbo_size(lock));
- if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
- GOTO(out, rc = -ENOMEM);
+ if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
+ GOTO(out, rc = -ENOMEM);
- rc = req_capsule_server_pack(&req->rq_pill);
- if (rc)
- GOTO(out, rc);
- }
+ rc = req_capsule_server_pack(&req->rq_pill);
+ if (rc)
+ GOTO(out, rc);
+ }
- err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
+ err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
if (err) {
if ((int)err < 0)
rc = (int)err;
GOTO(out, err);
}
- dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
+ dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
- ldlm_lock2desc(lock, &dlm_rep->lock_desc);
- ldlm_lock2handle(lock, &dlm_rep->lock_handle);
+ ldlm_lock2desc(lock, &dlm_rep->lock_desc);
+ ldlm_lock2handle(lock, &dlm_rep->lock_handle);
if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
"(err=%d, rc=%d)", err, rc);
- if (rc == 0) {
- if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
- RCL_SERVER) &&
- ldlm_lvbo_size(lock) > 0) {
- void *buf;
- int buflen;
-
- buf = req_capsule_server_get(&req->rq_pill,
- &RMF_DLM_LVB);
- LASSERTF(buf != NULL, "req %p, lock %p\n",
- req, lock);
- buflen = req_capsule_get_size(&req->rq_pill,
- &RMF_DLM_LVB, RCL_SERVER);
- /* non-replayed lock, delayed lvb init may
- * need to be occur now */
- if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
- buflen = ldlm_lvbo_fill(lock, buf,
- buflen);
- if (buflen >= 0)
- req_capsule_shrink(
+ if (rc == 0 &&
+ req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
+ RCL_SERVER) &&
+ ldlm_lvbo_size(lock) > 0) {
+ void *buf;
+ int buflen;
+
+retry:
+ buf = req_capsule_server_get(&req->rq_pill,
+ &RMF_DLM_LVB);
+ LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
+ buflen = req_capsule_get_size(&req->rq_pill,
+ &RMF_DLM_LVB, RCL_SERVER);
+ /* non-replayed lock, delayed lvb init may
+ * need to be occur now
+ */
+ if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
+ int rc2;
+
+ rc2 = ldlm_lvbo_fill(env, lock, buf, &buflen);
+ if (rc2 >= 0) {
+ req_capsule_shrink(&req->rq_pill,
+ &RMF_DLM_LVB,
+ rc2, RCL_SERVER);
+ } else if (rc2 == -ERANGE) {
+ rc2 = req_capsule_server_grow(
&req->rq_pill,
- &RMF_DLM_LVB,
- buflen, RCL_SERVER);
- else
- rc = buflen;
- } else if (flags & LDLM_FL_REPLAY) {
- /* no LVB resend upon replay */
- if (buflen > 0)
+ &RMF_DLM_LVB, buflen);
+ if (!rc2) {
+ goto retry;
+ } else {
+ /* if we can't grow the buffer,
+ * it's ok to return empty lvb
+ * to client.
+ */
req_capsule_shrink(
&req->rq_pill,
- &RMF_DLM_LVB,
- 0, RCL_SERVER);
- else
- rc = buflen;
+ &RMF_DLM_LVB, 0,
+ RCL_SERVER);
+ }
} else {
- rc = buflen;
+ rc = rc2;
}
+ } else if (flags & LDLM_FL_REPLAY) {
+ /* no LVB resend upon replay */
+ if (buflen > 0)
+ req_capsule_shrink(&req->rq_pill,
+ &RMF_DLM_LVB,
+ 0, RCL_SERVER);
+ else
+ rc = buflen;
+ } else {
+ rc = buflen;
}
}
return rc;
}
+/* Clear the blocking lock, the race is possible between ldlm_handle_convert0()
+ * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
+ */
+void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
+{
+ if (lock->l_blocking_lock) {
+ LDLM_LOCK_RELEASE(lock->l_blocking_lock);
+ lock->l_blocking_lock = NULL;
+ }
+}
+
+/* A lock can be converted to new ibits or mode and should be considered
+ * as new lock. Clear all states related to a previous blocking AST
+ * processing so new conflicts will cause new blocking ASTs.
+ *
+ * This is used during lock convert below and lock downgrade to COS mode in
+ * ldlm_lock_mode_downgrade().
+ */
+void ldlm_clear_blocking_data(struct ldlm_lock *lock)
+{
+ ldlm_clear_ast_sent(lock);
+ lock->l_bl_ast_run = 0;
+ ldlm_clear_blocking_lock(lock);
+}
+
/**
* Main LDLM entry point for server code to process lock conversion requests.
*/
ldlm_clear_cbpending(lock);
lock->l_policy_data.l_inodebits.cancel_bits = 0;
ldlm_inodebits_drop(lock, bits & ~new);
- lock->l_bl_ast_run = 0;
- ldlm_clear_ast_sent(lock);
+
+ ldlm_clear_blocking_data(lock);
unlock_res_and_lock(lock);
ldlm_reprocess_all(lock->l_resource);
const struct ldlm_request *dlm_req,
int first, enum lustre_at_flags flags)
{
+ const struct lu_env *env = req->rq_svc_thread->t_env;
struct ldlm_resource *res, *pres = NULL;
struct ldlm_lock *lock;
int i, count, done = 0;
LDLM_RESOURCE_DELREF(pres);
ldlm_resource_putref(pres);
}
- if (res != NULL) {
- ldlm_resource_getref(res);
- LDLM_RESOURCE_ADDREF(res);
+ if (res != NULL) {
+ ldlm_resource_getref(res);
+ LDLM_RESOURCE_ADDREF(res);
if (!ldlm_is_discard_data(lock))
- ldlm_lvbo_update(res, lock, NULL, 1);
- }
- pres = res;
- }
+ ldlm_lvbo_update(env, res, lock,
+ NULL, 1);
+ }
+ pres = res;
+ }
if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
lock->l_blast_sent != 0) {
}
lock_res_and_lock(lock);
+
+ if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
+ &lock->l_resource->lr_name)) {
+ ldlm_resource_unlink_lock(lock);
+ unlock_res_and_lock(lock);
+ rc = ldlm_lock_change_resource(ns, lock,
+ &dlm_req->lock_desc.l_resource.lr_name);
+ if (rc < 0) {
+ LDLM_ERROR(lock, "Failed to allocate resource");
+ GOTO(out, rc);
+ }
+ LDLM_DEBUG(lock, "completion AST, new resource");
+ lock_res_and_lock(lock);
+ }
+
if (ldlm_is_destroyed(lock) ||
lock->l_granted_mode == lock->l_req_mode) {
/* bug 11300: the lock has already been granted */
LDLM_DEBUG(lock, "completion AST, new policy data");
}
- ldlm_resource_unlink_lock(lock);
- if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
- &lock->l_resource->lr_name,
- sizeof(lock->l_resource->lr_name)) != 0) {
- unlock_res_and_lock(lock);
- rc = ldlm_lock_change_resource(ns, lock,
- &dlm_req->lock_desc.l_resource.lr_name);
- if (rc < 0) {
- LDLM_ERROR(lock, "Failed to allocate resource");
- GOTO(out, rc);
- }
- LDLM_DEBUG(lock, "completion AST, new resource");
- CERROR("change resource!\n");
- lock_res_and_lock(lock);
- }
+ ldlm_resource_unlink_lock(lock);
if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
/* BL_AST locks are not needed in LRU.
.tc_nthrs_base = LDLM_NTHRS_BASE,
.tc_nthrs_max = LDLM_NTHRS_MAX,
.tc_nthrs_user = ldlm_num_threads,
- .tc_cpu_affinity = 1,
+ .tc_cpu_bind = ldlm_cpu_bind,
.tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
},
.psc_cpt = {
.cc_pattern = ldlm_cpts,
+ .cc_affinity = true,
},
.psc_ops = {
.so_req_handler = ldlm_callback_handler,
.tc_nthrs_base = LDLM_NTHRS_BASE,
.tc_nthrs_max = LDLM_NTHRS_MAX,
.tc_nthrs_user = ldlm_num_threads,
- .tc_cpu_affinity = 1,
+ .tc_cpu_bind = ldlm_cpu_bind,
.tc_ctx_tags = LCT_MD_THREAD | \
LCT_DT_THREAD | \
LCT_CL_THREAD,
},
.psc_cpt = {
.cc_pattern = ldlm_cpts,
+ .cc_affinity = true,
},
.psc_ops = {
.so_req_handler = ldlm_cancel_handler,
ldlm_lock_slab = kmem_cache_create("ldlm_locks",
sizeof(struct ldlm_lock), 0,
- SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
+ SLAB_HWCACHE_ALIGN, NULL);
if (ldlm_lock_slab == NULL)
goto out_resource;