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 Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_RPC
26 #include <linux/module.h>
28 #include <liblustre.h>
30 #include <linux/obd_class.h>
31 #include <linux/lustre_net.h>
32 #include "ptlrpc_internal.h"
34 #if !defined(__KERNEL__) && CRAY_PORTALS
35 /* forward ref in events.c */
36 static void cray_portals_callback(ptl_event_t *ev);
40 struct ptlrpc_ni ptlrpc_interfaces[8];
41 int ptlrpc_ninterfaces;
44 * Client's outgoing request callback
46 void request_out_callback(ptl_event_t *ev)
48 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
49 struct ptlrpc_request *req = cbid->cbid_arg;
53 LASSERT (ev->type == PTL_EVENT_SEND_END ||
54 ev->type == PTL_EVENT_UNLINK);
55 LASSERT (ev->unlinked);
57 DEBUG_REQ((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, req,
58 "type %d, status %d", ev->type, ev->ni_fail_type);
60 if (ev->type == PTL_EVENT_UNLINK ||
61 ev->ni_fail_type != PTL_NI_OK) {
63 /* Failed send: make it seem like the reply timed out, just
64 * like failing sends in client.c does currently... */
66 spin_lock_irqsave(&req->rq_lock, flags);
68 spin_unlock_irqrestore(&req->rq_lock, flags);
70 ptlrpc_wake_client_req(req);
73 /* this balances the atomic_inc in ptl_send_rpc() */
74 ptlrpc_req_finished(req);
79 * Client's incoming reply callback
81 void reply_in_callback(ptl_event_t *ev)
83 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
84 struct ptlrpc_request *req = cbid->cbid_arg;
88 LASSERT (ev->type == PTL_EVENT_PUT_END ||
89 ev->type == PTL_EVENT_UNLINK);
90 LASSERT (ev->unlinked);
91 LASSERT (ev->md.start == req->rq_repmsg);
92 LASSERT (ev->offset == 0);
93 LASSERT (ev->mlength <= req->rq_replen);
95 DEBUG_REQ((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR, req,
96 "type %d, status %d", ev->type, ev->ni_fail_type);
98 spin_lock_irqsave (&req->rq_lock, flags);
100 LASSERT (req->rq_receiving_reply);
101 req->rq_receiving_reply = 0;
103 if (ev->type == PTL_EVENT_PUT_END &&
104 ev->ni_fail_type == PTL_NI_OK) {
106 req->rq_nob_received = ev->mlength;
109 /* NB don't unlock till after wakeup; req can disappear under us
110 * since we don't have our own ref */
111 ptlrpc_wake_client_req(req);
113 spin_unlock_irqrestore (&req->rq_lock, flags);
118 * Client's bulk has been written/read
120 void client_bulk_callback (ptl_event_t *ev)
122 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
123 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
127 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
128 ev->type == PTL_EVENT_PUT_END) ||
129 (desc->bd_type == BULK_GET_SOURCE &&
130 ev->type == PTL_EVENT_GET_END) ||
131 ev->type == PTL_EVENT_UNLINK);
132 LASSERT (ev->unlinked);
134 CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR,
135 "event type %d, status %d, desc %p\n",
136 ev->type, ev->ni_fail_type, desc);
138 spin_lock_irqsave (&desc->bd_lock, flags);
140 LASSERT(desc->bd_network_rw);
141 desc->bd_network_rw = 0;
143 if (ev->type != PTL_EVENT_UNLINK &&
144 ev->ni_fail_type == PTL_NI_OK) {
145 desc->bd_success = 1;
146 desc->bd_nob_transferred = ev->mlength;
149 /* NB don't unlock till after wakeup; desc can disappear under us
151 ptlrpc_wake_client_req(desc->bd_req);
153 spin_unlock_irqrestore (&desc->bd_lock, flags);
158 * Server's incoming request callback
160 void request_in_callback(ptl_event_t *ev)
162 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
163 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
164 struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
165 struct ptlrpc_service *service = srv_ni->sni_service;
166 struct ptlrpc_request *req;
167 char str[PTL_NALFMT_SIZE];
171 LASSERT (ev->type == PTL_EVENT_PUT_END ||
172 ev->type == PTL_EVENT_UNLINK);
173 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
174 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
175 rqbd->rqbd_buffer + service->srv_buf_size);
177 CDEBUG((ev->ni_fail_type == PTL_OK) ? D_NET : D_ERROR,
178 "event type %d, status %d, service %s\n",
179 ev->type, ev->ni_fail_type, service->srv_name);
182 /* If this is the last request message to fit in the
183 * request buffer we can use the request object embedded in
184 * rqbd. Note that if we failed to allocate a request,
185 * we'd have to re-post the rqbd, which we can't do in this
187 req = &rqbd->rqbd_req;
188 memset(req, 0, sizeof (*req));
190 LASSERT (ev->type == PTL_EVENT_PUT_END);
191 if (ev->ni_fail_type != PTL_NI_OK) {
192 /* We moaned above already... */
195 OBD_ALLOC_GFP(req, sizeof(*req), GFP_ATOMIC);
197 CERROR("Can't allocate incoming request descriptor: "
198 "Dropping %s RPC from %s\n",
200 portals_id2str(srv_ni->sni_ni->pni_number,
201 ev->initiator, str));
206 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
207 * flags are reset and scalars are zero. We only set the message
208 * size to non-zero if this was a successful receive. */
209 req->rq_xid = ev->match_bits;
210 req->rq_reqmsg = ev->md.start + ev->offset;
211 if (ev->type == PTL_EVENT_PUT_END &&
212 ev->ni_fail_type == PTL_NI_OK)
213 req->rq_reqlen = ev->mlength;
214 do_gettimeofday(&req->rq_arrival_time);
215 req->rq_peer.peer_id = ev->initiator;
216 req->rq_peer.peer_ni = rqbd->rqbd_srv_ni->sni_ni;
217 ptlrpc_id2str(&req->rq_peer, req->rq_peerstr);
219 req->rq_phase = RQ_PHASE_NEW;
221 spin_lock_irqsave (&service->srv_lock, flags);
223 req->rq_history_seq = service->srv_request_seq++;
224 list_add_tail(&req->rq_history_list, &service->srv_request_history);
227 srv_ni->sni_nrqbd_receiving--;
228 if (ev->type != PTL_EVENT_UNLINK &&
229 srv_ni->sni_nrqbd_receiving == 0) {
230 /* This service is off-air on this interface because
231 * all its request buffers are busy. Portals will
232 * start dropping incoming requests until more buffers
233 * get posted. NB don't moan if it's because we're
234 * tearing down the service. */
235 CWARN("All %s %s request buffers busy\n",
236 service->srv_name, srv_ni->sni_ni->pni_name);
238 /* req takes over the network's ref on rqbd */
240 /* req takes a ref on rqbd */
241 rqbd->rqbd_refcount++;
244 list_add_tail(&req->rq_list, &service->srv_request_queue);
245 service->srv_n_queued_reqs++;
247 /* NB everything can disappear under us once the request
248 * has been queued and we unlock, so do the wake now... */
249 wake_up(&service->srv_waitq);
251 spin_unlock_irqrestore(&service->srv_lock, flags);
256 * Server's outgoing reply callback
258 void reply_out_callback(ptl_event_t *ev)
260 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
261 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
262 struct ptlrpc_srv_ni *sni = rs->rs_srv_ni;
263 struct ptlrpc_service *svc = sni->sni_service;
267 LASSERT (ev->type == PTL_EVENT_SEND_END ||
268 ev->type == PTL_EVENT_ACK ||
269 ev->type == PTL_EVENT_UNLINK);
271 if (!rs->rs_difficult) {
272 /* 'Easy' replies have no further processing so I drop the
273 * net's ref on 'rs' */
274 LASSERT (ev->unlinked);
275 ptlrpc_rs_decref(rs);
276 atomic_dec (&svc->srv_outstanding_replies);
281 LASSERT (rs->rs_on_net);
284 /* Last network callback. The net's ref on 'rs' stays put
285 * until ptlrpc_server_handle_reply() is done with it */
286 spin_lock_irqsave (&svc->srv_lock, flags);
288 ptlrpc_schedule_difficult_reply (rs);
289 spin_unlock_irqrestore (&svc->srv_lock, flags);
296 * Server's bulk completion callback
298 void server_bulk_callback (ptl_event_t *ev)
300 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
301 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
305 LASSERT (ev->type == PTL_EVENT_SEND_END ||
306 ev->type == PTL_EVENT_UNLINK ||
307 (desc->bd_type == BULK_PUT_SOURCE &&
308 ev->type == PTL_EVENT_ACK) ||
309 (desc->bd_type == BULK_GET_SINK &&
310 ev->type == PTL_EVENT_REPLY_END));
312 CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR,
313 "event type %d, status %d, desc %p\n",
314 ev->type, ev->ni_fail_type, desc);
316 spin_lock_irqsave (&desc->bd_lock, flags);
318 if ((ev->type == PTL_EVENT_ACK ||
319 ev->type == PTL_EVENT_REPLY_END) &&
320 ev->ni_fail_type == PTL_NI_OK) {
321 /* We heard back from the peer, so even if we get this
322 * before the SENT event (oh yes we can), we know we
323 * read/wrote the peer buffer and how much... */
324 desc->bd_success = 1;
325 desc->bd_nob_transferred = ev->mlength;
329 /* This is the last callback no matter what... */
330 desc->bd_network_rw = 0;
331 wake_up(&desc->bd_waitq);
334 spin_unlock_irqrestore (&desc->bd_lock, flags);
338 static void ptlrpc_master_callback(ptl_event_t *ev)
340 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
341 void (*callback)(ptl_event_t *ev) = cbid->cbid_fn;
343 /* Honestly, it's best to find out early. */
344 LASSERT (cbid->cbid_arg != LP_POISON);
345 LASSERT (callback == request_out_callback ||
346 callback == reply_in_callback ||
347 callback == client_bulk_callback ||
348 callback == request_in_callback ||
349 callback == reply_out_callback ||
350 callback == server_bulk_callback);
355 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid, struct ptlrpc_peer *peer)
357 struct ptlrpc_ni *pni;
361 char str[PTL_NALFMT_SIZE];
362 int rc = lustre_uuid_to_peer(uuid->uuid,
363 &peer_nal, &peer_nid);
367 for (i = 0; i < ptlrpc_ninterfaces; i++) {
368 pni = &ptlrpc_interfaces[i];
370 if (pni->pni_number == peer_nal) {
371 peer->peer_id.nid = peer_nid;
372 peer->peer_id.pid = LUSTRE_SRV_PTL_PID;
378 CERROR("Can't find ptlrpc interface for NAL %x, NID %s\n",
379 peer_nal, portals_nid2str(peer_nal, peer_nid, str));
383 void ptlrpc_ni_fini(struct ptlrpc_ni *pni)
385 wait_queue_head_t waitq;
386 struct l_wait_info lwi;
390 /* Wait for the event queue to become idle since there may still be
391 * messages in flight with pending events (i.e. the fire-and-forget
392 * messages == client requests and "non-difficult" server
395 for (retries = 0;; retries++) {
396 rc = PtlEQFree(pni->pni_eq_h);
402 PtlNIFini(pni->pni_ni_h);
407 CWARN("Event queue for %s still busy\n",
411 init_waitqueue_head(&waitq);
412 lwi = LWI_TIMEOUT(2*HZ, NULL, NULL);
413 l_wait_event(waitq, 0, &lwi);
420 ptl_pid_t ptl_get_pid(void)
427 pid = LUSTRE_SRV_PTL_PID;
432 int ptlrpc_ni_init(int number, char *name, struct ptlrpc_ni *pni)
441 /* We're not passing any limits yet... */
442 rc = PtlNIInit(number, pid, NULL, NULL, &nih);
443 if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
444 CDEBUG (D_NET, "Can't init network interface %s: %d\n",
449 CDEBUG(D_NET, "My pid is: %x\n", ptl_get_pid());
451 PtlSnprintHandle(str, sizeof(str), nih);
452 CDEBUG (D_NET, "init %x %s: %s\n", number, name, str);
454 pni->pni_name = name;
455 pni->pni_number = number;
458 pni->pni_eq_h = PTL_INVALID_HANDLE;
460 /* CAVEAT EMPTOR: how we process portals events is _radically_
461 * different depending on... */
463 /* kernel portals calls our master callback when events are added to
464 * the event queue. In fact lustre never pulls events off this queue,
465 * so it's only sized for some debug history. */
466 rc = PtlEQAlloc(pni->pni_ni_h, 1024, ptlrpc_master_callback,
469 /* liblustre calls the master callback when it removes events from the
470 * event queue. The event queue has to be big enough not to drop
473 /* cray portals implements a non-standard callback to notify us there
474 * are buffered events even when the app is not doing a filesystem
476 rc = PtlEQAlloc(pni->pni_ni_h, 10240, cray_portals_callback,
479 rc = PtlEQAlloc(pni->pni_ni_h, 10240, PTL_EQ_HANDLER_NONE,
484 GOTO (fail, rc = -ENOMEM);
488 CERROR ("Failed to initialise network interface %s: %d\n",
491 /* OK to do complete teardown since we invalidated the handles above */
492 ptlrpc_ni_fini (pni);
497 LIST_HEAD(liblustre_wait_callbacks);
498 void *liblustre_services_callback;
501 liblustre_register_wait_callback (int (*fn)(void *arg), void *arg)
503 struct liblustre_wait_callback *llwc;
505 OBD_ALLOC(llwc, sizeof(*llwc));
506 LASSERT (llwc != NULL);
509 llwc->llwc_arg = arg;
510 list_add_tail(&llwc->llwc_list, &liblustre_wait_callbacks);
516 liblustre_deregister_wait_callback (void *opaque)
518 struct liblustre_wait_callback *llwc = opaque;
520 list_del(&llwc->llwc_list);
521 OBD_FREE(llwc, sizeof(*llwc));
525 liblustre_check_events (int timeout)
532 rc = PtlEQPoll(&ptlrpc_interfaces[0].pni_eq_h, 1, timeout * 1000,
534 if (rc == PTL_EQ_EMPTY)
537 LASSERT (rc == PTL_EQ_DROPPED || rc == PTL_OK);
539 /* liblustre: no asynch callback so we can't affort to miss any
541 if (rc == PTL_EQ_DROPPED) {
542 CERROR ("Dropped an event!!!\n");
546 ptlrpc_master_callback (&ev);
550 int liblustre_waiting = 0;
553 liblustre_wait_event (int timeout)
555 struct list_head *tmp;
556 struct liblustre_wait_callback *llwc;
557 int found_something = 0;
559 /* single threaded recursion check... */
560 liblustre_waiting = 1;
563 /* Deal with all pending events */
564 while (liblustre_check_events(0))
567 /* Give all registered callbacks a bite at the cherry */
568 list_for_each(tmp, &liblustre_wait_callbacks) {
569 llwc = list_entry(tmp, struct liblustre_wait_callback,
572 if (llwc->llwc_fn(llwc->llwc_arg))
576 if (found_something || timeout == 0)
579 /* Nothing so far, but I'm allowed to block... */
580 found_something = liblustre_check_events(timeout);
581 if (!found_something) /* still nothing */
582 break; /* I timed out */
585 liblustre_waiting = 0;
587 return found_something;
591 static void cray_portals_callback(ptl_event_t *ev)
593 /* We get a callback from the client Cray portals implementation
594 * whenever anyone calls PtlEQPoll(), and an event queue with a
595 * callback handler has outstanding events.
597 * If it's not liblustre calling PtlEQPoll(), this lets us know we
598 * have outstanding events which we handle with
599 * liblustre_wait_event().
601 * Otherwise, we're already eagerly consuming events and we'd
602 * handle events out of order if we recursed. */
603 if (!liblustre_waiting)
604 liblustre_wait_event(0);
607 #endif /* __KERNEL__ */
609 int ptlrpc_default_nal(void)
611 if (ptlrpc_ninterfaces == 0)
614 return (ptlrpc_interfaces[0].pni_number);
617 int ptlrpc_init_portals(void)
619 /* Add new portals network interfaces here.
620 * Order is irrelevent! */
627 {SOCKNAL, "socknal"},
629 {OPENIBNAL, "openibnal"},
636 {CRAY_KB_ERNAL, "cray_kb_ernal"},
642 LASSERT(ptlrpc_ninterfaces == 0);
644 for (i = 0; i < sizeof (ptl_nis) / sizeof (ptl_nis[0]); i++) {
645 LASSERT(ptlrpc_ninterfaces < (sizeof(ptlrpc_interfaces) /
646 sizeof(ptlrpc_interfaces[0])));
648 rc = ptlrpc_ni_init(ptl_nis[i].number, ptl_nis[i].name,
649 &ptlrpc_interfaces[ptlrpc_ninterfaces]);
651 ptlrpc_ninterfaces++;
654 if (ptlrpc_ninterfaces == 0) {
655 CERROR("network initialisation failed: is a NAL module "
660 liblustre_services_callback =
661 liblustre_register_wait_callback(&liblustre_check_services, NULL);
666 void ptlrpc_exit_portals(void)
669 liblustre_deregister_wait_callback(liblustre_services_callback);
671 while (ptlrpc_ninterfaces > 0)
672 ptlrpc_ni_fini (&ptlrpc_interfaces[--ptlrpc_ninterfaces]);