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_repbuf);
92 LASSERT (ev->offset == 0);
93 LASSERT (ev->mlength <= req->rq_repbuf_len);
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_reqbuf = ev->md.start + ev->offset;
211 if (ev->type == PTL_EVENT_PUT_END &&
212 ev->ni_fail_type == PTL_NI_OK)
213 req->rq_reqbuf_len = 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);
220 req->rq_uid = ev->uid;
223 spin_lock_irqsave (&service->srv_lock, flags);
226 srv_ni->sni_nrqbd_receiving--;
227 if (ev->type != PTL_EVENT_UNLINK &&
228 srv_ni->sni_nrqbd_receiving == 0) {
229 /* This service is off-air on this interface because
230 * all its request buffers are busy. Portals will
231 * start dropping incoming requests until more buffers
232 * get posted. NB don't moan if it's because we're
233 * tearing down the service. */
234 CWARN("All %s %s request buffers busy\n",
235 service->srv_name, srv_ni->sni_ni->pni_name);
237 /* req takes over the network's ref on rqbd */
239 /* req takes a ref on rqbd */
240 rqbd->rqbd_refcount++;
243 list_add_tail(&req->rq_list, &service->srv_request_queue);
244 service->srv_n_queued_reqs++;
246 /* NB everything can disappear under us once the request
247 * has been queued and we unlock, so do the wake now... */
248 wake_up(&service->srv_waitq);
250 spin_unlock_irqrestore(&service->srv_lock, flags);
255 * Server's outgoing reply callback
257 void reply_out_callback(ptl_event_t *ev)
259 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
260 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
261 struct ptlrpc_srv_ni *sni = rs->rs_srv_ni;
262 struct ptlrpc_service *svc = sni->sni_service;
266 LASSERT (ev->type == PTL_EVENT_SEND_END ||
267 ev->type == PTL_EVENT_ACK ||
268 ev->type == PTL_EVENT_UNLINK);
270 if (!rs->rs_difficult) {
271 /* I'm totally responsible for freeing "easy" replies */
272 LASSERT (ev->unlinked);
273 lustre_free_reply_state (rs);
274 atomic_dec (&svc->srv_outstanding_replies);
279 LASSERT (rs->rs_on_net);
282 /* Last network callback */
283 spin_lock_irqsave (&svc->srv_lock, flags);
285 ptlrpc_schedule_difficult_reply (rs);
286 spin_unlock_irqrestore (&svc->srv_lock, flags);
293 * Server's bulk completion callback
295 void server_bulk_callback (ptl_event_t *ev)
297 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
298 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
302 LASSERT (ev->type == PTL_EVENT_SEND_END ||
303 ev->type == PTL_EVENT_UNLINK ||
304 (desc->bd_type == BULK_PUT_SOURCE &&
305 ev->type == PTL_EVENT_ACK) ||
306 (desc->bd_type == BULK_GET_SINK &&
307 ev->type == PTL_EVENT_REPLY_END));
309 CDEBUG((ev->ni_fail_type == PTL_NI_OK) ? D_NET : D_ERROR,
310 "event type %d, status %d, desc %p\n",
311 ev->type, ev->ni_fail_type, desc);
313 spin_lock_irqsave (&desc->bd_lock, flags);
315 if ((ev->type == PTL_EVENT_ACK ||
316 ev->type == PTL_EVENT_REPLY_END) &&
317 ev->ni_fail_type == PTL_NI_OK) {
318 /* We heard back from the peer, so even if we get this
319 * before the SENT event (oh yes we can), we know we
320 * read/wrote the peer buffer and how much... */
321 desc->bd_success = 1;
322 desc->bd_nob_transferred = ev->mlength;
326 /* This is the last callback no matter what... */
327 desc->bd_network_rw = 0;
328 wake_up(&desc->bd_waitq);
331 spin_unlock_irqrestore (&desc->bd_lock, flags);
335 static void ptlrpc_master_callback(ptl_event_t *ev)
337 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
338 void (*callback)(ptl_event_t *ev) = cbid->cbid_fn;
340 /* Honestly, it's best to find out early. */
341 LASSERT (cbid->cbid_arg != LP_POISON);
342 LASSERT (callback == request_out_callback ||
343 callback == reply_in_callback ||
344 callback == client_bulk_callback ||
345 callback == request_in_callback ||
346 callback == reply_out_callback ||
347 callback == server_bulk_callback);
352 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid, struct ptlrpc_peer *peer)
354 struct ptlrpc_ni *pni;
358 char str[PTL_NALFMT_SIZE];
359 int rc = lustre_uuid_to_peer(uuid->uuid,
360 &peer_nal, &peer_nid);
364 for (i = 0; i < ptlrpc_ninterfaces; i++) {
365 pni = &ptlrpc_interfaces[i];
368 if (pni->pni_number == peer_nal) {
370 /* compatible nals but may be from different bridges */
371 if (NALID_FROM_IFACE(pni->pni_number) ==
372 NALID_FROM_IFACE(peer_nal)) {
374 peer->peer_id.nid = peer_nid;
375 peer->peer_id.pid = LUSTRE_SRV_PTL_PID;
381 CERROR("Can't find ptlrpc interface for NAL %x, NID %s\n",
382 peer_nal, portals_nid2str(peer_nal, peer_nid, str));
386 void ptlrpc_ni_fini(struct ptlrpc_ni *pni)
388 wait_queue_head_t waitq;
389 struct l_wait_info lwi;
393 /* Wait for the event queue to become idle since there may still be
394 * messages in flight with pending events (i.e. the fire-and-forget
395 * messages == client requests and "non-difficult" server
398 for (retries = 0;; retries++) {
399 rc = PtlEQFree(pni->pni_eq_h);
405 PtlNIFini(pni->pni_ni_h);
410 CWARN("Event queue for %s still busy\n",
414 init_waitqueue_head(&waitq);
415 lwi = LWI_TIMEOUT(2*HZ, NULL, NULL);
416 l_wait_event(waitq, 0, &lwi);
423 ptl_pid_t ptl_get_pid(void)
430 /* hack to keep pid in range accepted by ernal */
432 if (pid == LUSTRE_SRV_PTL_PID)
436 pid = LUSTRE_SRV_PTL_PID;
441 int ptlrpc_ni_init(int number, char *name, struct ptlrpc_ni *pni)
450 /* We're not passing any limits yet... */
451 rc = PtlNIInit(number, pid, NULL, NULL, &nih);
452 if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
453 CDEBUG (D_NET, "Can't init network interface %s: %d\n",
458 CDEBUG(D_NET, "My pid is: %x\n", ptl_get_pid());
460 PtlSnprintHandle(str, sizeof(str), nih);
461 CDEBUG (D_NET, "init %x %s: %s\n", number, name, str);
463 pni->pni_name = name;
464 pni->pni_number = number;
467 pni->pni_eq_h = PTL_INVALID_HANDLE;
469 /* CAVEAT EMPTOR: how we process portals events is _radically_
470 * different depending on... */
472 /* kernel portals calls our master callback when events are added to
473 * the event queue. In fact lustre never pulls events off this queue,
474 * so it's only sized for some debug history. */
475 rc = PtlEQAlloc(pni->pni_ni_h, 1024, ptlrpc_master_callback,
478 /* liblustre calls the master callback when it removes events from the
479 * event queue. The event queue has to be big enough not to drop
482 /* cray portals implements a non-standard callback to notify us there
483 * are buffered events even when the app is not doing a filesystem
485 rc = PtlEQAlloc(pni->pni_ni_h, 10240, cray_portals_callback,
488 rc = PtlEQAlloc(pni->pni_ni_h, 10240, PTL_EQ_HANDLER_NONE,
493 GOTO (fail, rc = -ENOMEM);
497 CERROR ("Failed to initialise network interface %s: %d\n",
500 /* OK to do complete teardown since we invalidated the handles above */
501 ptlrpc_ni_fini (pni);
506 LIST_HEAD(liblustre_wait_callbacks);
507 void *liblustre_services_callback;
510 liblustre_register_wait_callback (int (*fn)(void *arg), void *arg)
512 struct liblustre_wait_callback *llwc;
514 OBD_ALLOC(llwc, sizeof(*llwc));
515 LASSERT (llwc != NULL);
518 llwc->llwc_arg = arg;
519 list_add_tail(&llwc->llwc_list, &liblustre_wait_callbacks);
525 liblustre_deregister_wait_callback (void *opaque)
527 struct liblustre_wait_callback *llwc = opaque;
529 list_del(&llwc->llwc_list);
530 OBD_FREE(llwc, sizeof(*llwc));
534 liblustre_check_events (int timeout)
541 rc = PtlEQPoll(&ptlrpc_interfaces[0].pni_eq_h, 1, timeout * 1000,
543 if (rc == PTL_EQ_EMPTY)
546 LASSERT (rc == PTL_EQ_DROPPED || rc == PTL_OK);
548 /* liblustre: no asynch callback so we can't affort to miss any
550 if (rc == PTL_EQ_DROPPED) {
551 CERROR ("Dropped an event!!!\n");
555 ptlrpc_master_callback (&ev);
559 int liblustre_waiting = 0;
562 liblustre_wait_event (int timeout)
564 struct list_head *tmp;
565 struct liblustre_wait_callback *llwc;
566 int found_something = 0;
568 /* single threaded recursion check... */
569 liblustre_waiting = 1;
572 /* Deal with all pending events */
573 while (liblustre_check_events(0))
576 /* Give all registered callbacks a bite at the cherry */
577 list_for_each(tmp, &liblustre_wait_callbacks) {
578 llwc = list_entry(tmp, struct liblustre_wait_callback,
581 if (llwc->llwc_fn(llwc->llwc_arg))
585 if (found_something || timeout == 0)
588 /* Nothing so far, but I'm allowed to block... */
589 found_something = liblustre_check_events(timeout);
590 if (!found_something) /* still nothing */
591 break; /* I timed out */
594 liblustre_waiting = 0;
596 return found_something;
600 static void cray_portals_callback(ptl_event_t *ev)
602 /* We get a callback from the client Cray portals implementation
603 * whenever anyone calls PtlEQPoll(), and an event queue with a
604 * callback handler has outstanding events.
606 * If it's not liblustre calling PtlEQPoll(), this lets us know we
607 * have outstanding events which we handle with
608 * liblustre_wait_event().
610 * Otherwise, we're already eagerly consuming events and we'd
611 * handle events out of order if we recursed. */
612 if (!liblustre_waiting)
613 liblustre_wait_event(0);
616 #endif /* __KERNEL__ */
618 int ptlrpc_default_nal(void)
620 if (ptlrpc_ninterfaces == 0)
623 return (ptlrpc_interfaces[0].pni_number);
626 int ptlrpc_init_portals(void)
628 /* Add new portals network interfaces here.
629 * Order is irrelevent! */
636 {SOCKNAL, "socknal"},
638 {OPENIBNAL, "openibnal"},
645 {CRAY_KERN_NAL, "cray_kern_nal"},
646 {CRAY_QK_NAL, "cray_qk_nal"},
647 {CRAY_USER_NAL, "cray_user_nal"},
653 LASSERT(ptlrpc_ninterfaces == 0);
655 for (i = 0; i < sizeof (ptl_nis) / sizeof (ptl_nis[0]); i++) {
656 LASSERT(ptlrpc_ninterfaces < (sizeof(ptlrpc_interfaces) /
657 sizeof(ptlrpc_interfaces[0])));
659 rc = ptlrpc_ni_init(ptl_nis[i].number, ptl_nis[i].name,
660 &ptlrpc_interfaces[ptlrpc_ninterfaces]);
662 ptlrpc_ninterfaces++;
665 if (ptlrpc_ninterfaces == 0) {
666 CERROR("network initialisation failed: is a NAL module "
671 liblustre_services_callback =
672 liblustre_register_wait_callback(&liblustre_check_services, NULL);
677 void ptlrpc_exit_portals(void)
680 liblustre_deregister_wait_callback(liblustre_services_callback);
682 while (ptlrpc_ninterfaces > 0)
683 ptlrpc_ni_fini (&ptlrpc_interfaces[--ptlrpc_ninterfaces]);