1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
26 #define DEBUG_SUBSYSTEM S_RPC
29 #include <liblustre.h>
31 #include <obd_class.h>
32 #include <lustre_net.h>
33 #include <lustre_sec.h>
34 #include "ptlrpc_internal.h"
36 lnet_handle_eq_t ptlrpc_eq_h;
39 * Client's outgoing request callback
41 void request_out_callback(lnet_event_t *ev)
43 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
44 struct ptlrpc_request *req = cbid->cbid_arg;
47 LASSERT (ev->type == LNET_EVENT_SEND ||
48 ev->type == LNET_EVENT_UNLINK);
49 LASSERT (ev->unlinked);
51 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
52 "type %d, status %d", ev->type, ev->status);
54 sptlrpc_request_out_callback(req);
56 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
58 /* Failed send: make it seem like the reply timed out, just
59 * like failing sends in client.c does currently... */
61 spin_lock(&req->rq_lock);
63 spin_unlock(&req->rq_lock);
65 ptlrpc_wake_client_req(req);
68 /* these balance the references in ptl_send_rpc() */
69 atomic_dec(&req->rq_import->imp_inflight);
70 ptlrpc_req_finished(req);
76 * Client's incoming reply callback
78 void reply_in_callback(lnet_event_t *ev)
80 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
81 struct ptlrpc_request *req = cbid->cbid_arg;
84 LASSERT (ev->type == LNET_EVENT_PUT ||
85 ev->type == LNET_EVENT_UNLINK);
86 LASSERT (ev->unlinked);
87 LASSERT (ev->md.start == req->rq_repbuf);
88 LASSERT (ev->offset == 0);
89 LASSERT (ev->mlength <= req->rq_repbuf_len);
91 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
92 "type %d, status %d", ev->type, ev->status);
94 spin_lock(&req->rq_lock);
96 LASSERT (req->rq_receiving_reply);
97 req->rq_receiving_reply = 0;
99 if (ev->type == LNET_EVENT_PUT && ev->status == 0) {
101 req->rq_nob_received = ev->mlength;
104 /* NB don't unlock till after wakeup; req can disappear under us
105 * since we don't have our own ref */
106 ptlrpc_wake_client_req(req);
108 spin_unlock(&req->rq_lock);
113 * Client's bulk has been written/read
115 void client_bulk_callback (lnet_event_t *ev)
117 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
118 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
121 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
122 ev->type == LNET_EVENT_PUT) ||
123 (desc->bd_type == BULK_GET_SOURCE &&
124 ev->type == LNET_EVENT_GET) ||
125 ev->type == LNET_EVENT_UNLINK);
126 LASSERT (ev->unlinked);
128 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
129 "event type %d, status %d, desc %p\n",
130 ev->type, ev->status, desc);
132 spin_lock(&desc->bd_lock);
134 LASSERT(desc->bd_network_rw);
135 desc->bd_network_rw = 0;
137 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
138 desc->bd_success = 1;
139 desc->bd_nob_transferred = ev->mlength;
142 sptlrpc_enc_pool_put_pages(desc);
144 /* NB don't unlock till after wakeup; desc can disappear under us
146 ptlrpc_wake_client_req(desc->bd_req);
148 spin_unlock(&desc->bd_lock);
153 * Server's incoming request callback
155 void request_in_callback(lnet_event_t *ev)
157 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
158 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
159 struct ptlrpc_service *service = rqbd->rqbd_service;
160 struct ptlrpc_request *req;
163 LASSERT (ev->type == LNET_EVENT_PUT ||
164 ev->type == LNET_EVENT_UNLINK);
165 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
166 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
167 rqbd->rqbd_buffer + service->srv_buf_size);
169 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
170 "event type %d, status %d, service %s\n",
171 ev->type, ev->status, service->srv_name);
174 /* If this is the last request message to fit in the
175 * request buffer we can use the request object embedded in
176 * rqbd. Note that if we failed to allocate a request,
177 * we'd have to re-post the rqbd, which we can't do in this
179 req = &rqbd->rqbd_req;
180 memset(req, 0, sizeof (*req));
182 LASSERT (ev->type == LNET_EVENT_PUT);
183 if (ev->status != 0) {
184 /* We moaned above already... */
187 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
189 CERROR("Can't allocate incoming request descriptor: "
190 "Dropping %s RPC from %s\n",
192 libcfs_id2str(ev->initiator));
197 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
198 * flags are reset and scalars are zero. We only set the message
199 * size to non-zero if this was a successful receive. */
200 req->rq_xid = ev->match_bits;
201 req->rq_reqbuf = ev->md.start + ev->offset;
202 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
203 req->rq_reqdata_len = ev->mlength;
204 do_gettimeofday(&req->rq_arrival_time);
205 req->rq_peer = ev->initiator;
206 req->rq_self = ev->target.nid;
208 req->rq_phase = RQ_PHASE_NEW;
210 req->rq_uid = ev->uid;
213 spin_lock(&service->srv_lock);
215 req->rq_history_seq = service->srv_request_seq++;
216 list_add_tail(&req->rq_history_list, &service->srv_request_history);
219 service->srv_nrqbd_receiving--;
220 CDEBUG(D_RPCTRACE,"Buffer complete: %d buffers still posted\n",
221 service->srv_nrqbd_receiving);
223 /* Normally, don't complain about 0 buffers posted; LNET won't
224 * drop incoming reqs since we set the portal lazy */
225 if (test_req_buffer_pressure &&
226 ev->type != LNET_EVENT_UNLINK &&
227 service->srv_nrqbd_receiving == 0)
228 CWARN("All %s request buffers busy\n",
231 /* req takes over the network's ref on rqbd */
233 /* req takes a ref on rqbd */
234 rqbd->rqbd_refcount++;
237 list_add_tail(&req->rq_list, &service->srv_request_queue);
238 service->srv_n_queued_reqs++;
240 /* NB everything can disappear under us once the request
241 * has been queued and we unlock, so do the wake now... */
242 cfs_waitq_signal(&service->srv_waitq);
244 spin_unlock(&service->srv_lock);
249 * Server's outgoing reply callback
251 void reply_out_callback(lnet_event_t *ev)
253 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
254 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
255 struct ptlrpc_service *svc = rs->rs_service;
258 LASSERT (ev->type == LNET_EVENT_SEND ||
259 ev->type == LNET_EVENT_ACK ||
260 ev->type == LNET_EVENT_UNLINK);
262 if (!rs->rs_difficult) {
263 /* 'Easy' replies have no further processing so I drop the
264 * net's ref on 'rs' */
265 LASSERT (ev->unlinked);
266 ptlrpc_rs_decref(rs);
267 atomic_dec (&svc->srv_outstanding_replies);
272 LASSERT (rs->rs_on_net);
275 /* Last network callback. The net's ref on 'rs' stays put
276 * until ptlrpc_server_handle_reply() is done with it */
277 spin_lock(&svc->srv_lock);
279 ptlrpc_schedule_difficult_reply (rs);
280 spin_unlock(&svc->srv_lock);
287 * Server's bulk completion callback
289 void server_bulk_callback (lnet_event_t *ev)
291 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
292 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
295 LASSERT (ev->type == LNET_EVENT_SEND ||
296 ev->type == LNET_EVENT_UNLINK ||
297 (desc->bd_type == BULK_PUT_SOURCE &&
298 ev->type == LNET_EVENT_ACK) ||
299 (desc->bd_type == BULK_GET_SINK &&
300 ev->type == LNET_EVENT_REPLY));
302 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
303 "event type %d, status %d, desc %p\n",
304 ev->type, ev->status, desc);
306 spin_lock(&desc->bd_lock);
308 if ((ev->type == LNET_EVENT_ACK ||
309 ev->type == LNET_EVENT_REPLY) &&
311 /* We heard back from the peer, so even if we get this
312 * before the SENT event (oh yes we can), we know we
313 * read/wrote the peer buffer and how much... */
314 desc->bd_success = 1;
315 desc->bd_nob_transferred = ev->mlength;
319 /* This is the last callback no matter what... */
320 desc->bd_network_rw = 0;
321 cfs_waitq_signal(&desc->bd_waitq);
324 spin_unlock(&desc->bd_lock);
328 static void ptlrpc_master_callback(lnet_event_t *ev)
330 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
331 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
333 /* Honestly, it's best to find out early. */
334 LASSERT (cbid->cbid_arg != LP_POISON);
335 LASSERT (callback == request_out_callback ||
336 callback == reply_in_callback ||
337 callback == client_bulk_callback ||
338 callback == request_in_callback ||
339 callback == reply_out_callback ||
340 callback == server_bulk_callback);
345 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
346 lnet_process_id_t *peer, lnet_nid_t *self)
352 int portals_compatibility;
358 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
360 peer->pid = LUSTRE_SRV_LNET_PID;
362 /* Choose the matching UUID that's closest */
363 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
364 dist = LNetDist(dst_nid, &src_nid, &order);
368 if (dist == 0) { /* local! use loopback LND */
369 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
374 LASSERT (order >= 0);
377 (dist == best_dist && order < best_order)) {
381 if (portals_compatibility > 1) {
382 /* Strong portals compatibility: Zero the nid's
383 * NET, so if I'm reading new config logs, or
384 * getting configured by (new) lconf I can
385 * still talk to old servers. */
386 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
387 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
395 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
397 CERROR("No NID found for %s\n", uuid->uuid);
401 void ptlrpc_ni_fini(void)
404 struct l_wait_info lwi;
408 /* Wait for the event queue to become idle since there may still be
409 * messages in flight with pending events (i.e. the fire-and-forget
410 * messages == client requests and "non-difficult" server
413 for (retries = 0;; retries++) {
414 rc = LNetEQFree(ptlrpc_eq_h);
425 CWARN("Event queue still busy\n");
428 cfs_waitq_init(&waitq);
429 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
430 l_wait_event(waitq, 0, &lwi);
437 lnet_pid_t ptl_get_pid(void)
444 pid = LUSTRE_SRV_LNET_PID;
449 int ptlrpc_ni_init(void)
455 CDEBUG(D_NET, "My pid is: %x\n", pid);
457 /* We're not passing any limits yet... */
458 rc = LNetNIInit(pid);
460 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
464 /* CAVEAT EMPTOR: how we process portals events is _radically_
465 * different depending on... */
467 /* kernel portals calls our master callback when events are added to
468 * the event queue. In fact lustre never pulls events off this queue,
469 * so it's only sized for some debug history. */
470 rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
472 /* liblustre calls the master callback when it removes events from the
473 * event queue. The event queue has to be big enough not to drop
475 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
480 CERROR ("Failed to allocate event queue: %d\n", rc);
487 CFS_LIST_HEAD(liblustre_wait_callbacks);
488 void *liblustre_services_callback;
491 liblustre_register_wait_callback (int (*fn)(void *arg), void *arg)
493 struct liblustre_wait_callback *llwc;
495 OBD_ALLOC(llwc, sizeof(*llwc));
496 LASSERT (llwc != NULL);
499 llwc->llwc_arg = arg;
500 list_add_tail(&llwc->llwc_list, &liblustre_wait_callbacks);
506 liblustre_deregister_wait_callback (void *opaque)
508 struct liblustre_wait_callback *llwc = opaque;
510 list_del(&llwc->llwc_list);
511 OBD_FREE(llwc, sizeof(*llwc));
515 liblustre_check_events (int timeout)
522 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
526 LASSERT (rc == -EOVERFLOW || rc == 1);
528 /* liblustre: no asynch callback so we can't affort to miss any
530 if (rc == -EOVERFLOW) {
531 CERROR ("Dropped an event!!!\n");
535 ptlrpc_master_callback (&ev);
539 int liblustre_waiting = 0;
542 liblustre_wait_event (int timeout)
544 struct list_head *tmp;
545 struct liblustre_wait_callback *llwc;
546 int found_something = 0;
548 /* single threaded recursion check... */
549 liblustre_waiting = 1;
552 /* Deal with all pending events */
553 while (liblustre_check_events(0))
556 /* Give all registered callbacks a bite at the cherry */
557 list_for_each(tmp, &liblustre_wait_callbacks) {
558 llwc = list_entry(tmp, struct liblustre_wait_callback,
561 if (llwc->llwc_fn(llwc->llwc_arg))
565 if (found_something || timeout == 0)
568 /* Nothing so far, but I'm allowed to block... */
569 found_something = liblustre_check_events(timeout);
570 if (!found_something) /* still nothing */
571 break; /* I timed out */
574 liblustre_waiting = 0;
576 return found_something;
579 #endif /* __KERNEL__ */
581 int ptlrpc_init_portals(void)
583 int rc = ptlrpc_ni_init();
586 CERROR("network initialisation failed\n");
590 liblustre_services_callback =
591 liblustre_register_wait_callback(&liblustre_check_services, NULL);
596 void ptlrpc_exit_portals(void)
599 liblustre_deregister_wait_callback(liblustre_services_callback);