1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/ptllnd/ptllnd.c
38 * Author: PJ Kirner <pjkirner@clusterfs.com>
45 .lnd_startup = kptllnd_startup,
46 .lnd_shutdown = kptllnd_shutdown,
47 .lnd_ctl = kptllnd_ctl,
48 .lnd_query = kptllnd_query,
49 .lnd_send = kptllnd_send,
50 .lnd_recv = kptllnd_recv,
51 .lnd_eager_recv = kptllnd_eager_recv,
54 kptl_data_t kptllnd_data;
57 kptllnd_ptlid2str(ptl_process_id_t id)
59 static char strs[64][32];
65 spin_lock_irqsave(&kptllnd_data.kptl_ptlid2str_lock, flags);
67 if (idx >= sizeof(strs)/sizeof(strs[0]))
69 spin_unlock_irqrestore(&kptllnd_data.kptl_ptlid2str_lock, flags);
71 snprintf(str, sizeof(strs[0]), FMT_PTLID, id.pid, id.nid);
76 kptllnd_assert_wire_constants (void)
78 /* Wire protocol assertions generated by 'wirecheck'
79 * running on Linux fedora 2.6.11-co-0.6.4 #1 Mon Jun 19 05:36:13 UTC 2006 i686 i686 i386 GNU
80 * with gcc version 4.1.1 20060525 (Red Hat 4.1.1-1) */
84 CLASSERT (PTL_RESERVED_MATCHBITS == 0x100);
85 CLASSERT (LNET_MSG_MATCHBITS == 0);
86 CLASSERT (PTLLND_MSG_MAGIC == 0x50746C4E);
87 CLASSERT (PTLLND_MSG_VERSION == 0x04);
88 CLASSERT (PTLLND_RDMA_OK == 0x00);
89 CLASSERT (PTLLND_RDMA_FAIL == 0x01);
90 CLASSERT (PTLLND_MSG_TYPE_INVALID == 0x00);
91 CLASSERT (PTLLND_MSG_TYPE_PUT == 0x01);
92 CLASSERT (PTLLND_MSG_TYPE_GET == 0x02);
93 CLASSERT (PTLLND_MSG_TYPE_IMMEDIATE == 0x03);
94 CLASSERT (PTLLND_MSG_TYPE_NOOP == 0x04);
95 CLASSERT (PTLLND_MSG_TYPE_HELLO == 0x05);
96 CLASSERT (PTLLND_MSG_TYPE_NAK == 0x06);
98 /* Checks for struct kptl_msg_t */
99 CLASSERT ((int)sizeof(kptl_msg_t) == 136);
100 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_magic) == 0);
101 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_magic) == 4);
102 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_version) == 4);
103 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_version) == 2);
104 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_type) == 6);
105 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_type) == 1);
106 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_credits) == 7);
107 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_credits) == 1);
108 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_nob) == 8);
109 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_nob) == 4);
110 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_cksum) == 12);
111 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_cksum) == 4);
112 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_srcnid) == 16);
113 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_srcnid) == 8);
114 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_srcstamp) == 24);
115 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_srcstamp) == 8);
116 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_dstnid) == 32);
117 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_dstnid) == 8);
118 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_dststamp) == 40);
119 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_dststamp) == 8);
120 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_srcpid) == 48);
121 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_srcpid) == 4);
122 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_dstpid) == 52);
123 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_dstpid) == 4);
124 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_u.immediate) == 56);
125 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_u.immediate) == 72);
126 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_u.rdma) == 56);
127 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_u.rdma) == 80);
128 CLASSERT ((int)offsetof(kptl_msg_t, ptlm_u.hello) == 56);
129 CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_u.hello) == 12);
131 /* Checks for struct kptl_immediate_msg_t */
132 CLASSERT ((int)sizeof(kptl_immediate_msg_t) == 72);
133 CLASSERT ((int)offsetof(kptl_immediate_msg_t, kptlim_hdr) == 0);
134 CLASSERT ((int)sizeof(((kptl_immediate_msg_t *)0)->kptlim_hdr) == 72);
135 CLASSERT ((int)offsetof(kptl_immediate_msg_t, kptlim_payload[13]) == 85);
136 CLASSERT ((int)sizeof(((kptl_immediate_msg_t *)0)->kptlim_payload[13]) == 1);
138 /* Checks for struct kptl_rdma_msg_t */
139 CLASSERT ((int)sizeof(kptl_rdma_msg_t) == 80);
140 CLASSERT ((int)offsetof(kptl_rdma_msg_t, kptlrm_hdr) == 0);
141 CLASSERT ((int)sizeof(((kptl_rdma_msg_t *)0)->kptlrm_hdr) == 72);
142 CLASSERT ((int)offsetof(kptl_rdma_msg_t, kptlrm_matchbits) == 72);
143 CLASSERT ((int)sizeof(((kptl_rdma_msg_t *)0)->kptlrm_matchbits) == 8);
145 /* Checks for struct kptl_hello_msg_t */
146 CLASSERT ((int)sizeof(kptl_hello_msg_t) == 12);
147 CLASSERT ((int)offsetof(kptl_hello_msg_t, kptlhm_matchbits) == 0);
148 CLASSERT ((int)sizeof(((kptl_hello_msg_t *)0)->kptlhm_matchbits) == 8);
149 CLASSERT ((int)offsetof(kptl_hello_msg_t, kptlhm_max_msg_size) == 8);
150 CLASSERT ((int)sizeof(((kptl_hello_msg_t *)0)->kptlhm_max_msg_size) == 4);
153 const char *kptllnd_evtype2str(int type)
155 #define DO_TYPE(x) case x: return #x;
158 DO_TYPE(PTL_EVENT_GET_START);
159 DO_TYPE(PTL_EVENT_GET_END);
160 DO_TYPE(PTL_EVENT_PUT_START);
161 DO_TYPE(PTL_EVENT_PUT_END);
162 DO_TYPE(PTL_EVENT_REPLY_START);
163 DO_TYPE(PTL_EVENT_REPLY_END);
164 DO_TYPE(PTL_EVENT_ACK);
165 DO_TYPE(PTL_EVENT_SEND_START);
166 DO_TYPE(PTL_EVENT_SEND_END);
167 DO_TYPE(PTL_EVENT_UNLINK);
169 return "<unknown event type>";
174 const char *kptllnd_msgtype2str(int type)
176 #define DO_TYPE(x) case x: return #x;
179 DO_TYPE(PTLLND_MSG_TYPE_INVALID);
180 DO_TYPE(PTLLND_MSG_TYPE_PUT);
181 DO_TYPE(PTLLND_MSG_TYPE_GET);
182 DO_TYPE(PTLLND_MSG_TYPE_IMMEDIATE);
183 DO_TYPE(PTLLND_MSG_TYPE_HELLO);
184 DO_TYPE(PTLLND_MSG_TYPE_NOOP);
185 DO_TYPE(PTLLND_MSG_TYPE_NAK);
187 return "<unknown msg type>";
192 const char *kptllnd_errtype2str(int type)
194 #define DO_TYPE(x) case x: return #x;
199 DO_TYPE(PTL_NO_SPACE);
200 DO_TYPE(PTL_ME_IN_USE);
201 DO_TYPE(PTL_NAL_FAILED);
202 DO_TYPE(PTL_NO_INIT);
203 DO_TYPE(PTL_IFACE_DUP);
204 DO_TYPE(PTL_IFACE_INVALID);
205 DO_TYPE(PTL_HANDLE_INVALID);
206 DO_TYPE(PTL_MD_INVALID);
207 DO_TYPE(PTL_ME_INVALID);
208 DO_TYPE(PTL_PROCESS_INVALID);
209 DO_TYPE(PTL_PT_INDEX_INVALID);
210 DO_TYPE(PTL_SR_INDEX_INVALID);
211 DO_TYPE(PTL_EQ_INVALID);
212 DO_TYPE(PTL_EQ_DROPPED);
213 DO_TYPE(PTL_EQ_EMPTY);
214 DO_TYPE(PTL_MD_NO_UPDATE);
216 DO_TYPE(PTL_AC_INDEX_INVALID);
217 DO_TYPE(PTL_MD_ILLEGAL);
218 DO_TYPE(PTL_ME_LIST_TOO_LONG);
219 DO_TYPE(PTL_MD_IN_USE);
220 DO_TYPE(PTL_NI_INVALID);
221 DO_TYPE(PTL_PID_INVALID);
222 DO_TYPE(PTL_PT_FULL);
223 DO_TYPE(PTL_VAL_FAILED);
224 DO_TYPE(PTL_NOT_IMPLEMENTED);
226 DO_TYPE(PTL_EQ_IN_USE);
227 DO_TYPE(PTL_PID_IN_USE);
228 DO_TYPE(PTL_INV_EQ_SIZE);
231 return "<unknown event type>";
237 kptllnd_cksum (void *ptr, int nob)
243 sum = ((sum << 1) | (sum >> 31)) + *c++;
245 /* ensure I don't return 0 (== no checksum) */
246 return (sum == 0) ? 1 : sum;
250 kptllnd_init_msg(kptl_msg_t *msg, int type,
251 lnet_process_id_t target, int body_nob)
253 msg->ptlm_type = type;
254 msg->ptlm_nob = (offsetof(kptl_msg_t, ptlm_u) + body_nob + 7) & ~7;
255 msg->ptlm_dstpid = target.pid;
256 msg->ptlm_dstnid = target.nid;
257 msg->ptlm_srcpid = the_lnet.ln_pid;
258 msg->ptlm_srcnid = kptllnd_ptl2lnetnid(target.nid,
259 kptllnd_data.kptl_portals_id.nid);
261 LASSERT(msg->ptlm_nob <= *kptllnd_tunables.kptl_max_msg_size);
265 kptllnd_msg_pack(kptl_msg_t *msg, kptl_peer_t *peer)
267 msg->ptlm_magic = PTLLND_MSG_MAGIC;
268 msg->ptlm_version = PTLLND_MSG_VERSION;
269 /* msg->ptlm_type Filled in kptllnd_init_msg() */
270 msg->ptlm_credits = peer->peer_outstanding_credits;
271 /* msg->ptlm_nob Filled in kptllnd_init_msg() */
273 /* msg->ptlm_{src|dst}[pn]id Filled in kptllnd_init_msg */
274 msg->ptlm_srcstamp = peer->peer_myincarnation;
275 msg->ptlm_dststamp = peer->peer_incarnation;
277 if (*kptllnd_tunables.kptl_checksum) {
278 /* NB ptlm_cksum zero while computing cksum */
279 msg->ptlm_cksum = kptllnd_cksum(msg,
280 offsetof(kptl_msg_t, ptlm_u));
285 kptllnd_msg_unpack(kptl_msg_t *msg, int nob)
287 const int hdr_size = offsetof(kptl_msg_t, ptlm_u);
292 /* 6 bytes are enough to have received magic + version */
294 CERROR("Very Short message: %d\n", nob);
299 * Determine if we need to flip
301 if (msg->ptlm_magic == PTLLND_MSG_MAGIC) {
303 } else if (msg->ptlm_magic == __swab32(PTLLND_MSG_MAGIC)) {
306 CERROR("Bad magic: %08x\n", msg->ptlm_magic);
310 msg_version = flip ? __swab16(msg->ptlm_version) : msg->ptlm_version;
312 if (msg_version != PTLLND_MSG_VERSION) {
313 CERROR("Bad version: got %04x expected %04x\n",
314 (__u32)msg_version, PTLLND_MSG_VERSION);
318 if (nob < hdr_size) {
319 CERROR("Short message: got %d, wanted at least %d\n",
324 /* checksum must be computed with
325 * 1) ptlm_cksum zero and
326 * 2) BEFORE anything gets modified/flipped
328 msg_cksum = flip ? __swab32(msg->ptlm_cksum) : msg->ptlm_cksum;
330 if (msg_cksum != 0 &&
331 msg_cksum != kptllnd_cksum(msg, hdr_size)) {
332 CERROR("Bad checksum\n");
336 msg->ptlm_version = msg_version;
337 msg->ptlm_cksum = msg_cksum;
340 /* These two are 1 byte long so we don't swap them
341 But check this assumtion*/
342 CLASSERT (sizeof(msg->ptlm_type) == 1);
343 CLASSERT (sizeof(msg->ptlm_credits) == 1);
344 /* src & dst stamps are opaque cookies */
345 __swab32s(&msg->ptlm_nob);
346 __swab64s(&msg->ptlm_srcnid);
347 __swab64s(&msg->ptlm_dstnid);
348 __swab32s(&msg->ptlm_srcpid);
349 __swab32s(&msg->ptlm_dstpid);
352 if (msg->ptlm_nob != nob) {
353 CERROR("msg_nob corrupt: got 0x%08x, wanted %08x\n",
358 switch(msg->ptlm_type)
360 case PTLLND_MSG_TYPE_PUT:
361 case PTLLND_MSG_TYPE_GET:
362 if (nob < hdr_size + sizeof(kptl_rdma_msg_t)) {
363 CERROR("Short rdma request: got %d, want %d\n",
364 nob, hdr_size + (int)sizeof(kptl_rdma_msg_t));
369 __swab64s(&msg->ptlm_u.rdma.kptlrm_matchbits);
371 if (msg->ptlm_u.rdma.kptlrm_matchbits < PTL_RESERVED_MATCHBITS) {
372 CERROR("Bad matchbits "LPX64"\n",
373 msg->ptlm_u.rdma.kptlrm_matchbits);
378 case PTLLND_MSG_TYPE_IMMEDIATE:
379 if (nob < offsetof(kptl_msg_t,
380 ptlm_u.immediate.kptlim_payload)) {
381 CERROR("Short immediate: got %d, want %d\n", nob,
382 (int)offsetof(kptl_msg_t,
383 ptlm_u.immediate.kptlim_payload));
389 case PTLLND_MSG_TYPE_NOOP:
390 case PTLLND_MSG_TYPE_NAK:
394 case PTLLND_MSG_TYPE_HELLO:
395 if (nob < hdr_size + sizeof(kptl_hello_msg_t)) {
396 CERROR("Short hello: got %d want %d\n",
397 nob, hdr_size + (int)sizeof(kptl_hello_msg_t));
401 __swab64s(&msg->ptlm_u.hello.kptlhm_matchbits);
402 __swab32s(&msg->ptlm_u.hello.kptlhm_max_msg_size);
407 CERROR("Bad message type: 0x%02x\n", (__u32)msg->ptlm_type);
415 kptllnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
417 kptl_net_t *net = ni->ni_data;
418 struct libcfs_ioctl_data *data = arg;
421 CDEBUG(D_NET, ">>> kptllnd_ctl cmd=%u arg=%p\n", cmd, arg);
424 * Validate that the context block is actually
425 * pointing to this interface
427 LASSERT (ni == net->net_ni);
430 case IOC_LIBCFS_DEL_PEER: {
431 lnet_process_id_t id;
433 id.nid = data->ioc_nid;
434 id.pid = data->ioc_u32[1];
436 rc = kptllnd_peer_del(id);
440 case IOC_LIBCFS_GET_PEER: {
441 lnet_process_id_t id = {.nid = LNET_NID_ANY,
442 .pid = LNET_PID_ANY};
443 __u64 incarnation = 0;
444 __u64 next_matchbits = 0;
445 __u64 last_matchbits_seen = 0;
452 int outstanding_credits = 0;
454 rc = kptllnd_get_peer_info(data->ioc_count, &id,
456 &refcount, &incarnation,
457 &next_matchbits, &last_matchbits_seen,
459 &credits, &outstanding_credits);
461 data->ioc_nid = id.nid;
462 data->ioc_net = state;
463 data->ioc_flags = sent_hello;
464 data->ioc_count = refcount;
465 data->ioc_u64[0] = incarnation;
466 data->ioc_u32[0] = (__u32)next_matchbits;
467 data->ioc_u32[1] = (__u32)(next_matchbits >> 32);
468 data->ioc_u32[2] = (__u32)last_matchbits_seen;
469 data->ioc_u32[3] = (__u32)(last_matchbits_seen >> 32);
470 data->ioc_u32[4] = id.pid;
471 data->ioc_u32[5] = (nsendq << 16) | nactiveq;
472 data->ioc_u32[6] = (credits << 16) | outstanding_credits;
480 CDEBUG(D_NET, "<<< kptllnd_ctl rc=%d\n", rc);
485 kptllnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
487 kptl_net_t *net = ni->ni_data;
488 kptl_peer_t *peer = NULL;
489 lnet_process_id_t id = {.nid = nid, .pid = LUSTRE_SRV_LNET_PID};
492 /* NB: kptllnd_find_target connects to peer if necessary */
493 if (kptllnd_find_target(net, id, &peer) != 0)
496 spin_lock_irqsave(&peer->peer_lock, flags);
497 if (peer->peer_last_alive != 0)
498 *when = peer->peer_last_alive;
499 spin_unlock_irqrestore(&peer->peer_lock, flags);
500 kptllnd_peer_decref(peer);
505 kptllnd_base_shutdown (void)
510 lnet_process_id_t process_id;
512 read_lock(&kptllnd_data.kptl_net_rw_lock);
513 LASSERT (list_empty(&kptllnd_data.kptl_nets));
514 read_unlock(&kptllnd_data.kptl_net_rw_lock);
516 switch (kptllnd_data.kptl_init) {
520 case PTLLND_INIT_ALL:
521 case PTLLND_INIT_DATA:
523 kptllnd_rx_buffer_pool_fini(&kptllnd_data.kptl_rx_buffer_pool);
524 LASSERT (list_empty(&kptllnd_data.kptl_sched_rxq));
525 LASSERT (list_empty(&kptllnd_data.kptl_sched_rxbq));
527 /* lock to interleave cleanly with peer birth/death */
528 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
529 LASSERT (kptllnd_data.kptl_shutdown == 0);
530 kptllnd_data.kptl_shutdown = 1; /* phase 1 == destroy peers */
531 /* no new peers possible now */
532 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
534 /* nuke all existing peers */
535 process_id.nid = LNET_NID_ANY;
536 process_id.pid = LNET_PID_ANY;
537 kptllnd_peer_del(process_id);
539 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
541 LASSERT (kptllnd_data.kptl_n_active_peers == 0);
544 while (kptllnd_data.kptl_npeers != 0) {
546 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
547 "Waiting for %d peers to terminate\n",
548 kptllnd_data.kptl_npeers);
550 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
553 cfs_pause(cfs_time_seconds(1));
555 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock,
559 LASSERT (list_empty(&kptllnd_data.kptl_closing_peers));
560 LASSERT (list_empty(&kptllnd_data.kptl_zombie_peers));
561 LASSERT (kptllnd_data.kptl_peers != NULL);
562 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++)
563 LASSERT (list_empty (&kptllnd_data.kptl_peers[i]));
565 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
566 CDEBUG(D_NET, "All peers deleted\n");
568 /* Shutdown phase 2: kill the daemons... */
569 kptllnd_data.kptl_shutdown = 2;
573 while (atomic_read (&kptllnd_data.kptl_nthreads) != 0) {
574 /* Wake up all threads*/
575 wake_up_all(&kptllnd_data.kptl_sched_waitq);
576 wake_up_all(&kptllnd_data.kptl_watchdog_waitq);
579 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
580 "Waiting for %d threads to terminate\n",
581 atomic_read(&kptllnd_data.kptl_nthreads));
582 cfs_pause(cfs_time_seconds(1));
585 CDEBUG(D_NET, "All Threads stopped\n");
586 LASSERT(list_empty(&kptllnd_data.kptl_sched_txq));
588 kptllnd_cleanup_tx_descs();
590 /* Nothing here now, but libcfs might soon require
591 * us to explicitly destroy wait queues and semaphores
592 * that would be done here */
596 case PTLLND_INIT_NOTHING:
597 CDEBUG(D_NET, "PTLLND_INIT_NOTHING\n");
601 if (!PtlHandleIsEqual(kptllnd_data.kptl_eqh, PTL_INVALID_HANDLE)) {
602 prc = PtlEQFree(kptllnd_data.kptl_eqh);
604 CERROR("Error %s(%d) freeing portals EQ\n",
605 kptllnd_errtype2str(prc), prc);
608 if (!PtlHandleIsEqual(kptllnd_data.kptl_nih, PTL_INVALID_HANDLE)) {
609 prc = PtlNIFini(kptllnd_data.kptl_nih);
611 CERROR("Error %s(%d) finalizing portals NI\n",
612 kptllnd_errtype2str(prc), prc);
615 LASSERT (atomic_read(&kptllnd_data.kptl_ntx) == 0);
616 LASSERT (list_empty(&kptllnd_data.kptl_idle_txs));
618 if (kptllnd_data.kptl_rx_cache != NULL)
619 cfs_mem_cache_destroy(kptllnd_data.kptl_rx_cache);
621 if (kptllnd_data.kptl_peers != NULL)
622 LIBCFS_FREE(kptllnd_data.kptl_peers,
623 sizeof (struct list_head) *
624 kptllnd_data.kptl_peer_hash_size);
626 if (kptllnd_data.kptl_nak_msg != NULL)
627 LIBCFS_FREE(kptllnd_data.kptl_nak_msg,
628 offsetof(kptl_msg_t, ptlm_u));
630 memset(&kptllnd_data, 0, sizeof(kptllnd_data));
636 kptllnd_base_startup (void)
642 lnet_process_id_t target;
645 if (*kptllnd_tunables.kptl_max_procs_per_node < 1) {
646 CERROR("max_procs_per_node must be >= 1\n");
650 if (*kptllnd_tunables.kptl_peertxcredits > PTLLND_MSG_MAX_CREDITS) {
651 CERROR("peercredits must be <= %d\n", PTLLND_MSG_MAX_CREDITS);
655 *kptllnd_tunables.kptl_max_msg_size &= ~7;
656 if (*kptllnd_tunables.kptl_max_msg_size < PTLLND_MIN_BUFFER_SIZE)
657 *kptllnd_tunables.kptl_max_msg_size = PTLLND_MIN_BUFFER_SIZE;
659 CLASSERT ((PTLLND_MIN_BUFFER_SIZE & 7) == 0);
660 CLASSERT (sizeof(kptl_msg_t) <= PTLLND_MIN_BUFFER_SIZE);
662 /* Zero pointers, flags etc; put everything into a known state. */
663 memset (&kptllnd_data, 0, sizeof (kptllnd_data));
665 LIBCFS_ALLOC(kptllnd_data.kptl_nak_msg, offsetof(kptl_msg_t, ptlm_u));
666 if (kptllnd_data.kptl_nak_msg == NULL) {
667 CERROR("Can't allocate NAK msg\n");
670 memset(kptllnd_data.kptl_nak_msg, 0, offsetof(kptl_msg_t, ptlm_u));
672 kptllnd_data.kptl_eqh = PTL_INVALID_HANDLE;
673 kptllnd_data.kptl_nih = PTL_INVALID_HANDLE;
675 rwlock_init(&kptllnd_data.kptl_net_rw_lock);
676 INIT_LIST_HEAD(&kptllnd_data.kptl_nets);
678 /* Setup the sched locks/lists/waitq */
679 spin_lock_init(&kptllnd_data.kptl_sched_lock);
680 init_waitqueue_head(&kptllnd_data.kptl_sched_waitq);
681 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_txq);
682 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_rxq);
683 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_rxbq);
685 /* Init kptl_ptlid2str_lock before any call to kptllnd_ptlid2str */
686 spin_lock_init(&kptllnd_data.kptl_ptlid2str_lock);
688 /* Setup the tx locks/lists */
689 spin_lock_init(&kptllnd_data.kptl_tx_lock);
690 INIT_LIST_HEAD(&kptllnd_data.kptl_idle_txs);
691 atomic_set(&kptllnd_data.kptl_ntx, 0);
693 /* Uptick the module reference count */
696 kptllnd_data.kptl_expected_peers =
697 *kptllnd_tunables.kptl_max_nodes *
698 *kptllnd_tunables.kptl_max_procs_per_node;
701 * Initialize the Network interface instance
702 * We use the default because we don't have any
703 * way to choose a better interface.
704 * Requested and actual limits are ignored.
707 #ifdef _USING_LUSTRE_PORTALS_
712 *kptllnd_tunables.kptl_pid, NULL, NULL,
713 &kptllnd_data.kptl_nih);
716 * Note: PTL_IFACE_DUP simply means that the requested
717 * interface was already inited and that we're sharing it.
720 if (ptl_rc != PTL_OK && ptl_rc != PTL_IFACE_DUP) {
721 CERROR ("PtlNIInit: error %s(%d)\n",
722 kptllnd_errtype2str(ptl_rc), ptl_rc);
727 /* NB eq size irrelevant if using a callback */
728 ptl_rc = PtlEQAlloc(kptllnd_data.kptl_nih,
730 kptllnd_eq_callback, /* handler callback */
731 &kptllnd_data.kptl_eqh); /* output handle */
732 if (ptl_rc != PTL_OK) {
733 CERROR("PtlEQAlloc failed %s(%d)\n",
734 kptllnd_errtype2str(ptl_rc), ptl_rc);
739 /* Fetch the lower NID */
740 ptl_rc = PtlGetId(kptllnd_data.kptl_nih,
741 &kptllnd_data.kptl_portals_id);
742 if (ptl_rc != PTL_OK) {
743 CERROR ("PtlGetID: error %s(%d)\n",
744 kptllnd_errtype2str(ptl_rc), ptl_rc);
749 if (kptllnd_data.kptl_portals_id.pid != *kptllnd_tunables.kptl_pid) {
750 /* The kernel ptllnd must have the expected PID */
751 CERROR("Unexpected PID: %u (%u expected)\n",
752 kptllnd_data.kptl_portals_id.pid,
753 *kptllnd_tunables.kptl_pid);
758 /* Initialized the incarnation - it must be for-all-time unique, even
759 * accounting for the fact that we increment it when we disconnect a
760 * peer that's using it */
761 do_gettimeofday(&tv);
762 kptllnd_data.kptl_incarnation = (((__u64)tv.tv_sec) * 1000000) +
764 CDEBUG(D_NET, "Incarnation="LPX64"\n", kptllnd_data.kptl_incarnation);
766 target.nid = LNET_NID_ANY;
767 target.pid = LNET_PID_ANY; /* NB target for NAK doesn't matter */
768 kptllnd_init_msg(kptllnd_data.kptl_nak_msg, PTLLND_MSG_TYPE_NAK, target, 0);
769 kptllnd_data.kptl_nak_msg->ptlm_magic = PTLLND_MSG_MAGIC;
770 kptllnd_data.kptl_nak_msg->ptlm_version = PTLLND_MSG_VERSION;
771 kptllnd_data.kptl_nak_msg->ptlm_srcpid = the_lnet.ln_pid;
772 kptllnd_data.kptl_nak_msg->ptlm_srcstamp = kptllnd_data.kptl_incarnation;
774 rwlock_init(&kptllnd_data.kptl_peer_rw_lock);
775 init_waitqueue_head(&kptllnd_data.kptl_watchdog_waitq);
776 atomic_set(&kptllnd_data.kptl_needs_ptltrace, 0);
777 INIT_LIST_HEAD(&kptllnd_data.kptl_closing_peers);
778 INIT_LIST_HEAD(&kptllnd_data.kptl_zombie_peers);
780 /* Allocate and setup the peer hash table */
781 kptllnd_data.kptl_peer_hash_size =
782 *kptllnd_tunables.kptl_peer_hash_table_size;
783 LIBCFS_ALLOC(kptllnd_data.kptl_peers,
784 sizeof(struct list_head) *
785 kptllnd_data.kptl_peer_hash_size);
786 if (kptllnd_data.kptl_peers == NULL) {
787 CERROR("Failed to allocate space for peer hash table size=%d\n",
788 kptllnd_data.kptl_peer_hash_size);
792 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++)
793 INIT_LIST_HEAD(&kptllnd_data.kptl_peers[i]);
795 kptllnd_rx_buffer_pool_init(&kptllnd_data.kptl_rx_buffer_pool);
797 kptllnd_data.kptl_rx_cache =
798 cfs_mem_cache_create("ptllnd_rx",
800 *kptllnd_tunables.kptl_max_msg_size,
803 if (kptllnd_data.kptl_rx_cache == NULL) {
804 CERROR("Can't create slab for RX descriptors\n");
809 /* lists/ptrs/locks initialised */
810 kptllnd_data.kptl_init = PTLLND_INIT_DATA;
812 /*****************************************************/
814 rc = kptllnd_setup_tx_descs();
816 CERROR("Can't pre-allocate %d TX descriptors: %d\n",
817 *kptllnd_tunables.kptl_ntx, rc);
821 /* Start the scheduler threads for handling incoming requests. No need
822 * to advance the state because this will be automatically cleaned up
823 * now that PTLLND_INIT_DATA state has been entered */
824 CDEBUG(D_NET, "starting %d scheduler threads\n", PTLLND_N_SCHED);
825 for (i = 0; i < PTLLND_N_SCHED; i++) {
826 rc = kptllnd_thread_start(kptllnd_scheduler, (void *)((long)i));
828 CERROR("Can't spawn scheduler[%d]: %d\n", i, rc);
833 rc = kptllnd_thread_start(kptllnd_watchdog, NULL);
835 CERROR("Can't spawn watchdog: %d\n", rc);
839 /* Ensure that 'rxb_nspare' buffers can be off the net (being emptied)
840 * and we will still have enough buffers posted for all our peers */
841 spares = *kptllnd_tunables.kptl_rxb_nspare *
842 ((*kptllnd_tunables.kptl_rxb_npages * PAGE_SIZE)/
843 *kptllnd_tunables.kptl_max_msg_size);
845 /* reserve and post the buffers */
846 rc = kptllnd_rx_buffer_pool_reserve(&kptllnd_data.kptl_rx_buffer_pool,
847 kptllnd_data.kptl_expected_peers +
850 CERROR("Can't reserve RX Buffer pool: %d\n", rc);
854 /* flag everything initialised */
855 kptllnd_data.kptl_init = PTLLND_INIT_ALL;
857 /*****************************************************/
859 if (*kptllnd_tunables.kptl_checksum)
860 CWARN("Checksumming enabled\n");
862 CDEBUG(D_NET, "<<< kptllnd_base_startup SUCCESS\n");
866 CERROR("kptllnd_base_startup failed: %d\n", rc);
867 kptllnd_base_shutdown();
872 kptllnd_startup (lnet_ni_t *ni)
877 LASSERT (ni->ni_lnd == &kptllnd_lnd);
879 if (kptllnd_data.kptl_init == PTLLND_INIT_NOTHING) {
880 rc = kptllnd_base_startup();
885 LIBCFS_ALLOC(net, sizeof(*net));
888 CERROR("Can't allocate kptl_net_t\n");
892 memset(net, 0, sizeof(*net));
895 ni->ni_maxtxcredits = *kptllnd_tunables.kptl_credits;
896 ni->ni_peertxcredits = *kptllnd_tunables.kptl_peertxcredits;
897 ni->ni_peerrtrcredits = *kptllnd_tunables.kptl_peerrtrcredits;
898 ni->ni_nid = kptllnd_ptl2lnetnid(ni->ni_nid,
899 kptllnd_data.kptl_portals_id.nid);
900 CDEBUG(D_NET, "ptl id=%s, lnet id=%s\n",
901 kptllnd_ptlid2str(kptllnd_data.kptl_portals_id),
902 libcfs_nid2str(ni->ni_nid));
904 /* NB LNET_NIDNET(ptlm_srcnid) of NAK doesn't matter in case of
906 kptllnd_data.kptl_nak_msg->ptlm_srcnid = ni->ni_nid;
908 atomic_set(&net->net_refcount, 1);
909 write_lock(&kptllnd_data.kptl_net_rw_lock);
910 list_add_tail(&net->net_list, &kptllnd_data.kptl_nets);
911 write_unlock(&kptllnd_data.kptl_net_rw_lock);
915 kptllnd_shutdown(ni);
920 kptllnd_shutdown (lnet_ni_t *ni)
922 kptl_net_t *net = ni->ni_data;
926 LASSERT (kptllnd_data.kptl_init == PTLLND_INIT_ALL);
928 CDEBUG(D_MALLOC, "before LND cleanup: kmem %d\n",
929 atomic_read (&libcfs_kmemory));
934 LASSERT (ni == net->net_ni);
935 LASSERT (!net->net_shutdown);
936 LASSERT (!list_empty(&net->net_list));
937 LASSERT (atomic_read(&net->net_refcount) != 0);
941 write_lock(&kptllnd_data.kptl_net_rw_lock);
942 kptllnd_net_decref(net);
943 list_del_init(&net->net_list);
944 write_unlock(&kptllnd_data.kptl_net_rw_lock);
946 /* Can't nuke peers here - they are shared among all NIs */
947 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
948 net->net_shutdown = 1; /* Order with peer creation */
949 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
952 while (atomic_read(&net->net_refcount) != 0) {
954 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
955 "Waiting for %d references to drop\n",
956 atomic_read(&net->net_refcount));
958 cfs_pause(cfs_time_seconds(1));
961 LIBCFS_FREE(net, sizeof(*net));
963 /* NB no locking since I don't race with writers */
964 if (list_empty(&kptllnd_data.kptl_nets))
965 kptllnd_base_shutdown();
966 CDEBUG(D_MALLOC, "after LND cleanup: kmem %d\n",
967 atomic_read (&libcfs_kmemory));
972 kptllnd_module_init (void)
976 kptllnd_assert_wire_constants();
978 rc = kptllnd_tunables_init();
982 kptllnd_init_ptltrace();
984 lnet_register_lnd(&kptllnd_lnd);
990 kptllnd_module_fini (void)
992 lnet_unregister_lnd(&kptllnd_lnd);
993 kptllnd_tunables_fini();
996 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
997 MODULE_DESCRIPTION("Kernel Portals LND v1.00");
998 MODULE_LICENSE("GPL");
1000 module_init(kptllnd_module_init);
1001 module_exit(kptllnd_module_fini);