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>
32 # include <linux/kernel.h>
35 #include <obd_class.h>
36 #include <lustre_net.h>
37 #include <lustre_sec.h>
38 #include "ptlrpc_internal.h"
40 lnet_handle_eq_t ptlrpc_eq_h;
43 * Client's outgoing request callback
45 void request_out_callback(lnet_event_t *ev)
47 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
48 struct ptlrpc_request *req = cbid->cbid_arg;
51 LASSERT (ev->type == LNET_EVENT_SEND ||
52 ev->type == LNET_EVENT_UNLINK);
53 LASSERT (ev->unlinked);
55 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
56 "type %d, status %d", ev->type, ev->status);
58 sptlrpc_request_out_callback(req);
60 if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
62 /* Failed send: make it seem like the reply timed out, just
63 * like failing sends in client.c does currently... */
65 spin_lock(&req->rq_lock);
67 spin_unlock(&req->rq_lock);
69 ptlrpc_wake_client_req(req);
72 ptlrpc_req_finished(req);
78 * Client's incoming reply callback
80 void reply_in_callback(lnet_event_t *ev)
82 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
83 struct ptlrpc_request *req = cbid->cbid_arg;
86 LASSERT (ev->type == LNET_EVENT_PUT ||
87 ev->type == LNET_EVENT_UNLINK);
88 LASSERT (ev->unlinked);
89 LASSERT (ev->md.start == req->rq_repbuf);
90 LASSERT (ev->offset == 0);
91 LASSERT (ev->mlength <= req->rq_repbuf_len);
93 DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
94 "type %d, status %d", ev->type, ev->status);
96 spin_lock(&req->rq_lock);
98 LASSERT (req->rq_receiving_reply);
99 req->rq_receiving_reply = 0;
101 if (ev->type == LNET_EVENT_PUT && ev->status == 0) {
103 req->rq_nob_received = ev->mlength;
106 /* NB don't unlock till after wakeup; req can disappear under us
107 * since we don't have our own ref */
108 ptlrpc_wake_client_req(req);
110 spin_unlock(&req->rq_lock);
115 * Client's bulk has been written/read
117 void client_bulk_callback (lnet_event_t *ev)
119 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
120 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
123 LASSERT ((desc->bd_type == BULK_PUT_SINK &&
124 ev->type == LNET_EVENT_PUT) ||
125 (desc->bd_type == BULK_GET_SOURCE &&
126 ev->type == LNET_EVENT_GET) ||
127 ev->type == LNET_EVENT_UNLINK);
128 LASSERT (ev->unlinked);
130 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
131 "event type %d, status %d, desc %p\n",
132 ev->type, ev->status, desc);
134 spin_lock(&desc->bd_lock);
136 LASSERT(desc->bd_network_rw);
137 desc->bd_network_rw = 0;
139 if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
140 desc->bd_success = 1;
141 desc->bd_nob_transferred = ev->mlength;
142 desc->bd_sender = ev->sender;
145 sptlrpc_enc_pool_put_pages(desc);
147 /* NB don't unlock till after wakeup; desc can disappear under us
149 ptlrpc_wake_client_req(desc->bd_req);
151 spin_unlock(&desc->bd_lock);
156 * Server's incoming request callback
158 void request_in_callback(lnet_event_t *ev)
160 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
161 struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
162 struct ptlrpc_service *service = rqbd->rqbd_service;
163 struct ptlrpc_request *req;
166 LASSERT (ev->type == LNET_EVENT_PUT ||
167 ev->type == LNET_EVENT_UNLINK);
168 LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
169 LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
170 rqbd->rqbd_buffer + service->srv_buf_size);
172 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
173 "event type %d, status %d, service %s\n",
174 ev->type, ev->status, service->srv_name);
177 /* If this is the last request message to fit in the
178 * request buffer we can use the request object embedded in
179 * rqbd. Note that if we failed to allocate a request,
180 * we'd have to re-post the rqbd, which we can't do in this
182 req = &rqbd->rqbd_req;
183 memset(req, 0, sizeof (*req));
185 LASSERT (ev->type == LNET_EVENT_PUT);
186 if (ev->status != 0) {
187 /* We moaned above already... */
190 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
192 CERROR("Can't allocate incoming request descriptor: "
193 "Dropping %s RPC from %s\n",
195 libcfs_id2str(ev->initiator));
200 /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
201 * flags are reset and scalars are zero. We only set the message
202 * size to non-zero if this was a successful receive. */
203 req->rq_xid = ev->match_bits;
204 req->rq_reqbuf = ev->md.start + ev->offset;
205 if (ev->type == LNET_EVENT_PUT && ev->status == 0)
206 req->rq_reqdata_len = ev->mlength;
207 do_gettimeofday(&req->rq_arrival_time);
208 req->rq_peer = ev->initiator;
209 req->rq_self = ev->target.nid;
211 req->rq_phase = RQ_PHASE_NEW;
213 req->rq_uid = ev->uid;
216 CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
218 spin_lock(&service->srv_lock);
220 req->rq_history_seq = service->srv_request_seq++;
221 list_add_tail(&req->rq_history_list, &service->srv_request_history);
224 service->srv_nrqbd_receiving--;
225 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
226 service->srv_nrqbd_receiving);
228 /* Normally, don't complain about 0 buffers posted; LNET won't
229 * drop incoming reqs since we set the portal lazy */
230 if (test_req_buffer_pressure &&
231 ev->type != LNET_EVENT_UNLINK &&
232 service->srv_nrqbd_receiving == 0)
233 CWARN("All %s request buffers busy\n",
236 /* req takes over the network's ref on rqbd */
238 /* req takes a ref on rqbd */
239 rqbd->rqbd_refcount++;
242 list_add_tail(&req->rq_list, &service->srv_request_queue);
243 service->srv_n_queued_reqs++;
245 /* NB everything can disappear under us once the request
246 * has been queued and we unlock, so do the wake now... */
247 cfs_waitq_signal(&service->srv_waitq);
249 spin_unlock(&service->srv_lock);
254 * Server's outgoing reply callback
256 void reply_out_callback(lnet_event_t *ev)
258 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
259 struct ptlrpc_reply_state *rs = cbid->cbid_arg;
260 struct ptlrpc_service *svc = rs->rs_service;
263 LASSERT (ev->type == LNET_EVENT_SEND ||
264 ev->type == LNET_EVENT_ACK ||
265 ev->type == LNET_EVENT_UNLINK);
267 if (!rs->rs_difficult) {
268 /* 'Easy' replies have no further processing so I drop the
269 * net's ref on 'rs' */
270 LASSERT (ev->unlinked);
271 ptlrpc_rs_decref(rs);
272 atomic_dec (&svc->srv_outstanding_replies);
277 LASSERT (rs->rs_on_net);
280 /* Last network callback. The net's ref on 'rs' stays put
281 * until ptlrpc_server_handle_reply() is done with it */
282 spin_lock(&svc->srv_lock);
284 ptlrpc_schedule_difficult_reply (rs);
285 spin_unlock(&svc->srv_lock);
292 * Server's bulk completion callback
294 void server_bulk_callback (lnet_event_t *ev)
296 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
297 struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
300 LASSERT (ev->type == LNET_EVENT_SEND ||
301 ev->type == LNET_EVENT_UNLINK ||
302 (desc->bd_type == BULK_PUT_SOURCE &&
303 ev->type == LNET_EVENT_ACK) ||
304 (desc->bd_type == BULK_GET_SINK &&
305 ev->type == LNET_EVENT_REPLY));
307 CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
308 "event type %d, status %d, desc %p\n",
309 ev->type, ev->status, desc);
311 spin_lock(&desc->bd_lock);
313 if ((ev->type == LNET_EVENT_ACK ||
314 ev->type == LNET_EVENT_REPLY) &&
316 /* We heard back from the peer, so even if we get this
317 * before the SENT event (oh yes we can), we know we
318 * read/wrote the peer buffer and how much... */
319 desc->bd_success = 1;
320 desc->bd_nob_transferred = ev->mlength;
321 desc->bd_sender = ev->sender;
325 /* This is the last callback no matter what... */
326 desc->bd_network_rw = 0;
327 cfs_waitq_signal(&desc->bd_waitq);
330 spin_unlock(&desc->bd_lock);
334 static void ptlrpc_master_callback(lnet_event_t *ev)
336 struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
337 void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
339 /* Honestly, it's best to find out early. */
340 LASSERT (cbid->cbid_arg != LP_POISON);
341 LASSERT (callback == request_out_callback ||
342 callback == reply_in_callback ||
343 callback == client_bulk_callback ||
344 callback == request_in_callback ||
345 callback == reply_out_callback ||
346 callback == server_bulk_callback);
351 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
352 lnet_process_id_t *peer, lnet_nid_t *self)
355 __u32 best_order = 0;
358 int portals_compatibility;
364 portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
366 peer->pid = LUSTRE_SRV_LNET_PID;
368 /* Choose the matching UUID that's closest */
369 while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
370 dist = LNetDist(dst_nid, &src_nid, &order);
374 if (dist == 0) { /* local! use loopback LND */
375 peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
382 (dist == best_dist && order < best_order)) {
386 if (portals_compatibility > 1) {
387 /* Strong portals compatibility: Zero the nid's
388 * NET, so if I'm reading new config logs, or
389 * getting configured by (new) lconf I can
390 * still talk to old servers. */
391 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
392 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
400 CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
402 CERROR("No NID found for %s\n", uuid->uuid);
406 void ptlrpc_ni_fini(void)
409 struct l_wait_info lwi;
413 /* Wait for the event queue to become idle since there may still be
414 * messages in flight with pending events (i.e. the fire-and-forget
415 * messages == client requests and "non-difficult" server
418 for (retries = 0;; retries++) {
419 rc = LNetEQFree(ptlrpc_eq_h);
430 CWARN("Event queue still busy\n");
433 cfs_waitq_init(&waitq);
434 lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
435 l_wait_event(waitq, 0, &lwi);
442 lnet_pid_t ptl_get_pid(void)
449 pid = LUSTRE_SRV_LNET_PID;
454 int ptlrpc_ni_init(void)
460 CDEBUG(D_NET, "My pid is: %x\n", pid);
462 /* We're not passing any limits yet... */
463 rc = LNetNIInit(pid);
465 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
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 = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
477 /* liblustre calls the master callback when it removes events from the
478 * event queue. The event queue has to be big enough not to drop
480 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
485 CERROR ("Failed to allocate event queue: %d\n", rc);
492 CFS_LIST_HEAD(liblustre_wait_callbacks);
493 CFS_LIST_HEAD(liblustre_idle_callbacks);
494 void *liblustre_services_callback;
497 liblustre_register_waitidle_callback (struct list_head *callback_list,
499 int (*fn)(void *arg), void *arg)
501 struct liblustre_wait_callback *llwc;
503 OBD_ALLOC(llwc, sizeof(*llwc));
504 LASSERT (llwc != NULL);
506 llwc->llwc_name = name;
508 llwc->llwc_arg = arg;
509 list_add_tail(&llwc->llwc_list, callback_list);
515 liblustre_deregister_waitidle_callback (void *opaque)
517 struct liblustre_wait_callback *llwc = opaque;
519 list_del(&llwc->llwc_list);
520 OBD_FREE(llwc, sizeof(*llwc));
524 liblustre_register_wait_callback (const char *name,
525 int (*fn)(void *arg), void *arg)
527 return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
532 liblustre_deregister_wait_callback (void *opaque)
534 liblustre_deregister_waitidle_callback(opaque);
538 liblustre_register_idle_callback (const char *name,
539 int (*fn)(void *arg), void *arg)
541 return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
546 liblustre_deregister_idle_callback (void *opaque)
548 liblustre_deregister_waitidle_callback(opaque);
552 liblustre_check_events (int timeout)
559 rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
563 LASSERT (rc == -EOVERFLOW || rc == 1);
565 /* liblustre: no asynch callback so we can't affort to miss any
567 if (rc == -EOVERFLOW) {
568 CERROR ("Dropped an event!!!\n");
572 ptlrpc_master_callback (&ev);
576 int liblustre_waiting = 0;
579 liblustre_wait_event (int timeout)
581 struct list_head *tmp;
582 struct liblustre_wait_callback *llwc;
583 int found_something = 0;
585 /* single threaded recursion check... */
586 liblustre_waiting = 1;
589 /* Deal with all pending events */
590 while (liblustre_check_events(0))
593 /* Give all registered callbacks a bite at the cherry */
594 list_for_each(tmp, &liblustre_wait_callbacks) {
595 llwc = list_entry(tmp, struct liblustre_wait_callback,
598 if (llwc->llwc_fn(llwc->llwc_arg))
602 if (found_something || timeout == 0)
605 /* Nothing so far, but I'm allowed to block... */
606 found_something = liblustre_check_events(timeout);
607 if (!found_something) /* still nothing */
608 break; /* I timed out */
611 liblustre_waiting = 0;
613 return found_something;
617 liblustre_wait_idle(void)
619 static int recursed = 0;
621 struct list_head *tmp;
622 struct liblustre_wait_callback *llwc;
629 liblustre_wait_event(0);
633 list_for_each(tmp, &liblustre_idle_callbacks) {
634 llwc = list_entry(tmp, struct liblustre_wait_callback,
637 if (!llwc->llwc_fn(llwc->llwc_arg)) {
648 #endif /* __KERNEL__ */
650 int ptlrpc_init_portals(void)
652 int rc = ptlrpc_ni_init();
655 CERROR("network initialisation failed\n");
659 liblustre_services_callback =
660 liblustre_register_wait_callback("liblustre_check_services",
661 &liblustre_check_services,
664 rc = ptlrpcd_addref();
668 CERROR("rpcd initialisation failed\n");
670 liblustre_deregister_wait_callback(liblustre_services_callback);
676 void ptlrpc_exit_portals(void)
679 liblustre_deregister_wait_callback(liblustre_services_callback);