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, int body_nob)
252 msg->ptlm_type = type;
253 msg->ptlm_nob = (offsetof(kptl_msg_t, ptlm_u) + body_nob + 7) & ~7;
255 LASSERT(msg->ptlm_nob <= *kptllnd_tunables.kptl_max_msg_size);
259 kptllnd_msg_pack(kptl_msg_t *msg, kptl_peer_t *peer)
261 msg->ptlm_magic = PTLLND_MSG_MAGIC;
262 msg->ptlm_version = PTLLND_MSG_VERSION;
263 /* msg->ptlm_type Filled in kptllnd_init_msg() */
264 msg->ptlm_credits = peer->peer_outstanding_credits;
265 /* msg->ptlm_nob Filled in kptllnd_init_msg() */
267 msg->ptlm_srcnid = kptllnd_data.kptl_ni->ni_nid;
268 msg->ptlm_srcstamp = peer->peer_myincarnation;
269 msg->ptlm_dstnid = peer->peer_id.nid;
270 msg->ptlm_dststamp = peer->peer_incarnation;
271 msg->ptlm_srcpid = the_lnet.ln_pid;
272 msg->ptlm_dstpid = peer->peer_id.pid;
274 if (*kptllnd_tunables.kptl_checksum) {
275 /* NB ptlm_cksum zero while computing cksum */
276 msg->ptlm_cksum = kptllnd_cksum(msg,
277 offsetof(kptl_msg_t, ptlm_u));
282 kptllnd_msg_unpack(kptl_msg_t *msg, int nob)
284 const int hdr_size = offsetof(kptl_msg_t, ptlm_u);
289 /* 6 bytes are enough to have received magic + version */
291 CERROR("Very Short message: %d\n", nob);
296 * Determine if we need to flip
298 if (msg->ptlm_magic == PTLLND_MSG_MAGIC) {
300 } else if (msg->ptlm_magic == __swab32(PTLLND_MSG_MAGIC)) {
303 CERROR("Bad magic: %08x\n", msg->ptlm_magic);
307 msg_version = flip ? __swab16(msg->ptlm_version) : msg->ptlm_version;
309 if (msg_version != PTLLND_MSG_VERSION) {
310 CERROR("Bad version: got %04x expected %04x\n",
311 (__u32)msg_version, PTLLND_MSG_VERSION);
315 if (nob < hdr_size) {
316 CERROR("Short message: got %d, wanted at least %d\n",
321 /* checksum must be computed with
322 * 1) ptlm_cksum zero and
323 * 2) BEFORE anything gets modified/flipped
325 msg_cksum = flip ? __swab32(msg->ptlm_cksum) : msg->ptlm_cksum;
327 if (msg_cksum != 0 &&
328 msg_cksum != kptllnd_cksum(msg, hdr_size)) {
329 CERROR("Bad checksum\n");
333 msg->ptlm_version = msg_version;
334 msg->ptlm_cksum = msg_cksum;
337 /* These two are 1 byte long so we don't swap them
338 But check this assumtion*/
339 CLASSERT (sizeof(msg->ptlm_type) == 1);
340 CLASSERT (sizeof(msg->ptlm_credits) == 1);
341 /* src & dst stamps are opaque cookies */
342 __swab32s(&msg->ptlm_nob);
343 __swab64s(&msg->ptlm_srcnid);
344 __swab64s(&msg->ptlm_dstnid);
345 __swab32s(&msg->ptlm_srcpid);
346 __swab32s(&msg->ptlm_dstpid);
349 if (msg->ptlm_nob != nob) {
350 CERROR("msg_nob corrupt: got 0x%08x, wanted %08x\n",
355 switch(msg->ptlm_type)
357 case PTLLND_MSG_TYPE_PUT:
358 case PTLLND_MSG_TYPE_GET:
359 if (nob < hdr_size + sizeof(kptl_rdma_msg_t)) {
360 CERROR("Short rdma request: got %d, want %d\n",
361 nob, hdr_size + (int)sizeof(kptl_rdma_msg_t));
366 __swab64s(&msg->ptlm_u.rdma.kptlrm_matchbits);
368 if (msg->ptlm_u.rdma.kptlrm_matchbits < PTL_RESERVED_MATCHBITS) {
369 CERROR("Bad matchbits "LPX64"\n",
370 msg->ptlm_u.rdma.kptlrm_matchbits);
375 case PTLLND_MSG_TYPE_IMMEDIATE:
376 if (nob < offsetof(kptl_msg_t,
377 ptlm_u.immediate.kptlim_payload)) {
378 CERROR("Short immediate: got %d, want %d\n", nob,
379 (int)offsetof(kptl_msg_t,
380 ptlm_u.immediate.kptlim_payload));
386 case PTLLND_MSG_TYPE_NOOP:
387 case PTLLND_MSG_TYPE_NAK:
391 case PTLLND_MSG_TYPE_HELLO:
392 if (nob < hdr_size + sizeof(kptl_hello_msg_t)) {
393 CERROR("Short hello: got %d want %d\n",
394 nob, hdr_size + (int)sizeof(kptl_hello_msg_t));
398 __swab64s(&msg->ptlm_u.hello.kptlhm_matchbits);
399 __swab32s(&msg->ptlm_u.hello.kptlhm_max_msg_size);
404 CERROR("Bad message type: 0x%02x\n", (__u32)msg->ptlm_type);
412 kptllnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
414 struct libcfs_ioctl_data *data = arg;
417 CDEBUG(D_NET, ">>> kptllnd_ctl cmd=%u arg=%p\n", cmd, arg);
420 * Validate that the context block is actually
421 * pointing to this interface
423 LASSERT (ni == kptllnd_data.kptl_ni);
426 case IOC_LIBCFS_DEL_PEER: {
427 lnet_process_id_t id;
429 id.nid = data->ioc_nid;
430 id.pid = data->ioc_u32[1];
432 rc = kptllnd_peer_del(id);
436 case IOC_LIBCFS_GET_PEER: {
437 lnet_process_id_t id = {.nid = LNET_NID_ANY,
438 .pid = LNET_PID_ANY};
439 __u64 incarnation = 0;
440 __u64 next_matchbits = 0;
441 __u64 last_matchbits_seen = 0;
448 int outstanding_credits = 0;
450 rc = kptllnd_get_peer_info(data->ioc_count, &id,
452 &refcount, &incarnation,
453 &next_matchbits, &last_matchbits_seen,
455 &credits, &outstanding_credits);
457 data->ioc_nid = id.nid;
458 data->ioc_net = state;
459 data->ioc_flags = sent_hello;
460 data->ioc_count = refcount;
461 data->ioc_u64[0] = incarnation;
462 data->ioc_u32[0] = (__u32)next_matchbits;
463 data->ioc_u32[1] = (__u32)(next_matchbits >> 32);
464 data->ioc_u32[2] = (__u32)last_matchbits_seen;
465 data->ioc_u32[3] = (__u32)(last_matchbits_seen >> 32);
466 data->ioc_u32[4] = id.pid;
467 data->ioc_u32[5] = (nsendq << 16) | nactiveq;
468 data->ioc_u32[6] = (credits << 16) | outstanding_credits;
476 CDEBUG(D_NET, "<<< kptllnd_ctl rc=%d\n", rc);
481 kptllnd_query (lnet_ni_t *ni, lnet_nid_t nid, time_t *when)
483 kptl_peer_t *peer = NULL;
484 lnet_process_id_t id = {.nid = nid, .pid = LUSTRE_SRV_LNET_PID};
487 /* NB: kptllnd_find_target connects to peer if necessary */
488 if (kptllnd_find_target(&peer, id) != 0)
491 spin_lock_irqsave(&peer->peer_lock, flags);
492 if (peer->peer_last_alive != 0)
493 *when = cfs_time_current_sec() -
494 cfs_duration_sec(cfs_time_current() -
495 peer->peer_last_alive);
496 spin_unlock_irqrestore(&peer->peer_lock, flags);
497 kptllnd_peer_decref(peer);
502 kptllnd_startup (lnet_ni_t *ni)
510 LASSERT (ni->ni_lnd == &kptllnd_lnd);
512 if (kptllnd_data.kptl_init != PTLLND_INIT_NOTHING) {
513 CERROR("Only 1 instance supported\n");
517 if (*kptllnd_tunables.kptl_max_procs_per_node < 1) {
518 CERROR("max_procs_per_node must be >= 1\n");
522 if (*kptllnd_tunables.kptl_peertxcredits > PTLLND_MSG_MAX_CREDITS) {
523 CERROR("peercredits must be <= %d\n", PTLLND_MSG_MAX_CREDITS);
527 *kptllnd_tunables.kptl_max_msg_size &= ~7;
528 if (*kptllnd_tunables.kptl_max_msg_size < PTLLND_MIN_BUFFER_SIZE)
529 *kptllnd_tunables.kptl_max_msg_size = PTLLND_MIN_BUFFER_SIZE;
531 CLASSERT ((PTLLND_MIN_BUFFER_SIZE & 7) == 0);
532 CLASSERT (sizeof(kptl_msg_t) <= PTLLND_MIN_BUFFER_SIZE);
535 * zero pointers, flags etc
536 * put everything into a known state.
538 memset (&kptllnd_data, 0, sizeof (kptllnd_data));
539 kptllnd_data.kptl_eqh = PTL_INVALID_HANDLE;
540 kptllnd_data.kptl_nih = PTL_INVALID_HANDLE;
543 * Setup the sched locks/lists/waitq
545 spin_lock_init(&kptllnd_data.kptl_sched_lock);
546 init_waitqueue_head(&kptllnd_data.kptl_sched_waitq);
547 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_txq);
548 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_rxq);
549 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_rxbq);
551 /* init kptl_ptlid2str_lock before any call to kptllnd_ptlid2str */
552 spin_lock_init(&kptllnd_data.kptl_ptlid2str_lock);
555 * Setup the tx locks/lists
557 spin_lock_init(&kptllnd_data.kptl_tx_lock);
558 INIT_LIST_HEAD(&kptllnd_data.kptl_idle_txs);
559 atomic_set(&kptllnd_data.kptl_ntx, 0);
562 * Uptick the module reference count
567 * Setup pointers between the ni and context data block
569 kptllnd_data.kptl_ni = ni;
570 ni->ni_data = &kptllnd_data;
575 ni->ni_maxtxcredits = *kptllnd_tunables.kptl_credits;
576 ni->ni_peertxcredits = *kptllnd_tunables.kptl_peertxcredits;
577 ni->ni_peerrtrcredits = *kptllnd_tunables.kptl_peerrtrcredits;
579 kptllnd_data.kptl_expected_peers =
580 *kptllnd_tunables.kptl_max_nodes *
581 *kptllnd_tunables.kptl_max_procs_per_node;
584 * Initialize the Network interface instance
585 * We use the default because we don't have any
586 * way to choose a better interface.
587 * Requested and actual limits are ignored.
590 #ifdef _USING_LUSTRE_PORTALS_
595 *kptllnd_tunables.kptl_pid, NULL, NULL,
596 &kptllnd_data.kptl_nih);
599 * Note: PTL_IFACE_DUP simply means that the requested
600 * interface was already inited and that we're sharing it.
603 if (ptl_rc != PTL_OK && ptl_rc != PTL_IFACE_DUP) {
604 CERROR ("PtlNIInit: error %s(%d)\n",
605 kptllnd_errtype2str(ptl_rc), ptl_rc);
610 /* NB eq size irrelevant if using a callback */
611 ptl_rc = PtlEQAlloc(kptllnd_data.kptl_nih,
613 kptllnd_eq_callback, /* handler callback */
614 &kptllnd_data.kptl_eqh); /* output handle */
615 if (ptl_rc != PTL_OK) {
616 CERROR("PtlEQAlloc failed %s(%d)\n",
617 kptllnd_errtype2str(ptl_rc), ptl_rc);
623 * Fetch the lower NID
625 ptl_rc = PtlGetId(kptllnd_data.kptl_nih,
626 &kptllnd_data.kptl_portals_id);
627 if (ptl_rc != PTL_OK) {
628 CERROR ("PtlGetID: error %s(%d)\n",
629 kptllnd_errtype2str(ptl_rc), ptl_rc);
634 if (kptllnd_data.kptl_portals_id.pid != *kptllnd_tunables.kptl_pid) {
635 /* The kernel ptllnd must have the expected PID */
636 CERROR("Unexpected PID: %u (%u expected)\n",
637 kptllnd_data.kptl_portals_id.pid,
638 *kptllnd_tunables.kptl_pid);
643 ni->ni_nid = kptllnd_ptl2lnetnid(kptllnd_data.kptl_portals_id.nid);
645 CDEBUG(D_NET, "ptl id=%s, lnet id=%s\n",
646 kptllnd_ptlid2str(kptllnd_data.kptl_portals_id),
647 libcfs_nid2str(ni->ni_nid));
649 /* Initialized the incarnation - it must be for-all-time unique, even
650 * accounting for the fact that we increment it when we disconnect a
651 * peer that's using it */
652 do_gettimeofday(&tv);
653 kptllnd_data.kptl_incarnation = (((__u64)tv.tv_sec) * 1000000) +
655 CDEBUG(D_NET, "Incarnation="LPX64"\n", kptllnd_data.kptl_incarnation);
658 * Allocate and setup the peer hash table
660 rwlock_init(&kptllnd_data.kptl_peer_rw_lock);
661 init_waitqueue_head(&kptllnd_data.kptl_watchdog_waitq);
662 atomic_set(&kptllnd_data.kptl_needs_ptltrace, 0);
663 INIT_LIST_HEAD(&kptllnd_data.kptl_closing_peers);
664 INIT_LIST_HEAD(&kptllnd_data.kptl_zombie_peers);
666 kptllnd_data.kptl_peer_hash_size =
667 *kptllnd_tunables.kptl_peer_hash_table_size;
668 LIBCFS_ALLOC(kptllnd_data.kptl_peers,
669 (kptllnd_data.kptl_peer_hash_size *
670 sizeof(struct list_head)));
671 if (kptllnd_data.kptl_peers == NULL) {
672 CERROR("Failed to allocate space for peer hash table size=%d\n",
673 kptllnd_data.kptl_peer_hash_size);
677 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++)
678 INIT_LIST_HEAD(&kptllnd_data.kptl_peers[i]);
680 LIBCFS_ALLOC(kptllnd_data.kptl_nak_msg, offsetof(kptl_msg_t, ptlm_u));
681 if (kptllnd_data.kptl_nak_msg == NULL) {
682 CERROR("Can't allocate NAK msg\n");
686 memset(kptllnd_data.kptl_nak_msg, 0, offsetof(kptl_msg_t, ptlm_u));
687 kptllnd_init_msg(kptllnd_data.kptl_nak_msg, PTLLND_MSG_TYPE_NAK, 0);
688 kptllnd_data.kptl_nak_msg->ptlm_magic = PTLLND_MSG_MAGIC;
689 kptllnd_data.kptl_nak_msg->ptlm_version = PTLLND_MSG_VERSION;
690 kptllnd_data.kptl_nak_msg->ptlm_srcpid = the_lnet.ln_pid;
691 kptllnd_data.kptl_nak_msg->ptlm_srcnid = ni->ni_nid;
692 kptllnd_data.kptl_nak_msg->ptlm_srcstamp = kptllnd_data.kptl_incarnation;
693 kptllnd_data.kptl_nak_msg->ptlm_dstpid = LNET_PID_ANY;
694 kptllnd_data.kptl_nak_msg->ptlm_dstnid = LNET_NID_ANY;
696 kptllnd_rx_buffer_pool_init(&kptllnd_data.kptl_rx_buffer_pool);
698 kptllnd_data.kptl_rx_cache =
699 cfs_mem_cache_create("ptllnd_rx",
701 *kptllnd_tunables.kptl_max_msg_size,
704 if (kptllnd_data.kptl_rx_cache == NULL) {
705 CERROR("Can't create slab for RX descriptors\n");
710 /* lists/ptrs/locks initialised */
711 kptllnd_data.kptl_init = PTLLND_INIT_DATA;
713 /*****************************************************/
715 rc = kptllnd_setup_tx_descs();
717 CERROR("Can't pre-allocate %d TX descriptors: %d\n",
718 *kptllnd_tunables.kptl_ntx, rc);
722 /* Start the scheduler threads for handling incoming requests. No need
723 * to advance the state because this will be automatically cleaned up
724 * now that PTLNAT_INIT_DATA state has been entered */
725 CDEBUG(D_NET, "starting %d scheduler threads\n", PTLLND_N_SCHED);
726 for (i = 0; i < PTLLND_N_SCHED; i++) {
727 rc = kptllnd_thread_start(kptllnd_scheduler, (void *)((long)i));
729 CERROR("Can't spawn scheduler[%d]: %d\n", i, rc);
734 rc = kptllnd_thread_start(kptllnd_watchdog, NULL);
736 CERROR("Can't spawn watchdog: %d\n", rc);
740 /* Ensure that 'rxb_nspare' buffers can be off the net (being emptied)
741 * and we will still have enough buffers posted for all our peers */
742 spares = *kptllnd_tunables.kptl_rxb_nspare *
743 ((*kptllnd_tunables.kptl_rxb_npages * PAGE_SIZE)/
744 *kptllnd_tunables.kptl_max_msg_size);
746 /* reserve and post the buffers */
747 rc = kptllnd_rx_buffer_pool_reserve(&kptllnd_data.kptl_rx_buffer_pool,
748 kptllnd_data.kptl_expected_peers +
751 CERROR("Can't reserve RX Buffer pool: %d\n", rc);
755 /* flag everything initialised */
756 kptllnd_data.kptl_init = PTLLND_INIT_ALL;
758 /*****************************************************/
760 if (*kptllnd_tunables.kptl_checksum)
761 CWARN("Checksumming enabled\n");
763 CDEBUG(D_NET, "<<< kptllnd_startup SUCCESS\n");
767 CDEBUG(D_NET, "kptllnd_startup failed rc=%d\n", rc);
768 kptllnd_shutdown(ni);
773 kptllnd_shutdown (lnet_ni_t *ni)
777 lnet_process_id_t process_id;
780 CDEBUG(D_MALLOC, "before LND cleanup: kmem %d\n",
781 atomic_read (&libcfs_kmemory));
783 LASSERT (ni == kptllnd_data.kptl_ni);
785 switch (kptllnd_data.kptl_init) {
789 case PTLLND_INIT_ALL:
790 case PTLLND_INIT_DATA:
792 kptllnd_rx_buffer_pool_fini(&kptllnd_data.kptl_rx_buffer_pool);
793 LASSERT (list_empty(&kptllnd_data.kptl_sched_rxq));
794 LASSERT (list_empty(&kptllnd_data.kptl_sched_rxbq));
796 /* Hold peertable lock to interleave cleanly with peer birth/death */
797 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
799 LASSERT (kptllnd_data.kptl_shutdown == 0);
800 kptllnd_data.kptl_shutdown = 1; /* phase 1 == destroy peers */
802 /* no new peers possible now */
803 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
806 /* nuke all existing peers */
807 process_id.nid = LNET_NID_ANY;
808 process_id.pid = LNET_PID_ANY;
809 kptllnd_peer_del(process_id);
811 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
813 LASSERT (kptllnd_data.kptl_n_active_peers == 0);
816 while (kptllnd_data.kptl_npeers != 0) {
818 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
819 "Waiting for %d peers to terminate\n",
820 kptllnd_data.kptl_npeers);
822 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
825 cfs_pause(cfs_time_seconds(1));
827 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock,
831 LASSERT(list_empty(&kptllnd_data.kptl_closing_peers));
832 LASSERT(list_empty(&kptllnd_data.kptl_zombie_peers));
833 LASSERT (kptllnd_data.kptl_peers != NULL);
834 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++)
835 LASSERT (list_empty (&kptllnd_data.kptl_peers[i]));
837 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
838 CDEBUG(D_NET, "All peers deleted\n");
840 /* Shutdown phase 2: kill the daemons... */
841 kptllnd_data.kptl_shutdown = 2;
845 while (atomic_read (&kptllnd_data.kptl_nthreads) != 0) {
846 /* Wake up all threads*/
847 wake_up_all(&kptllnd_data.kptl_sched_waitq);
848 wake_up_all(&kptllnd_data.kptl_watchdog_waitq);
851 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
852 "Waiting for %d threads to terminate\n",
853 atomic_read(&kptllnd_data.kptl_nthreads));
854 cfs_pause(cfs_time_seconds(1));
857 CDEBUG(D_NET, "All Threads stopped\n");
858 LASSERT(list_empty(&kptllnd_data.kptl_sched_txq));
860 kptllnd_cleanup_tx_descs();
862 /* Nothing here now, but libcfs might soon require
863 * us to explicitly destroy wait queues and semaphores
864 * that would be done here */
868 case PTLLND_INIT_NOTHING:
869 CDEBUG(D_NET, "PTLLND_INIT_NOTHING\n");
873 if (!PtlHandleIsEqual(kptllnd_data.kptl_eqh, PTL_INVALID_HANDLE)) {
874 prc = PtlEQFree(kptllnd_data.kptl_eqh);
876 CERROR("Error %s(%d) freeing portals EQ\n",
877 kptllnd_errtype2str(prc), prc);
880 if (!PtlHandleIsEqual(kptllnd_data.kptl_nih, PTL_INVALID_HANDLE)) {
881 prc = PtlNIFini(kptllnd_data.kptl_nih);
883 CERROR("Error %s(%d) finalizing portals NI\n",
884 kptllnd_errtype2str(prc), prc);
887 LASSERT (atomic_read(&kptllnd_data.kptl_ntx) == 0);
888 LASSERT (list_empty(&kptllnd_data.kptl_idle_txs));
890 if (kptllnd_data.kptl_rx_cache != NULL)
891 cfs_mem_cache_destroy(kptllnd_data.kptl_rx_cache);
893 if (kptllnd_data.kptl_peers != NULL)
894 LIBCFS_FREE (kptllnd_data.kptl_peers,
895 sizeof (struct list_head) *
896 kptllnd_data.kptl_peer_hash_size);
898 if (kptllnd_data.kptl_nak_msg != NULL)
899 LIBCFS_FREE (kptllnd_data.kptl_nak_msg,
900 offsetof(kptl_msg_t, ptlm_u));
902 memset(&kptllnd_data, 0, sizeof(kptllnd_data));
904 CDEBUG(D_MALLOC, "after LND cleanup: kmem %d\n",
905 atomic_read (&libcfs_kmemory));
911 kptllnd_module_init (void)
915 kptllnd_assert_wire_constants();
917 rc = kptllnd_tunables_init();
921 kptllnd_init_ptltrace();
923 lnet_register_lnd(&kptllnd_lnd);
929 kptllnd_module_fini (void)
931 lnet_unregister_lnd(&kptllnd_lnd);
932 kptllnd_tunables_fini();
935 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
936 MODULE_DESCRIPTION("Kernel Portals LND v1.00");
937 MODULE_LICENSE("GPL");
939 module_init(kptllnd_module_init);
940 module_exit(kptllnd_module_fini);