2 * -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
3 * vim:expandtab:shiftwidth=8:tabstop=8:
7 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 only,
11 * as published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License version 2 for more details (a copy is included
17 * in the LICENSE file that accompanied this code).
19 * You should have received a copy of the GNU General Public License
20 * version 2 along with this program; If not, see [sun.com URL with a
23 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
24 * CA 95054 USA or visit www.sun.com if you need additional information or
30 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
31 * Use is subject to license terms.
34 * This file is part of Lustre, http://www.lustre.org/
35 * Lustre is a trademark of Sun Microsystems, Inc.
37 * lnet/klnds/ptllnd/ptllnd.c
39 * Author: PJ Kirner <pjkirner@clusterfs.com>
46 .lnd_startup = kptllnd_startup,
47 .lnd_shutdown = kptllnd_shutdown,
48 .lnd_ctl = kptllnd_ctl,
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_startup (lnet_ni_t *ni)
489 LASSERT (ni->ni_lnd == &kptllnd_lnd);
491 if (kptllnd_data.kptl_init != PTLLND_INIT_NOTHING) {
492 CERROR("Only 1 instance supported\n");
496 if (*kptllnd_tunables.kptl_max_procs_per_node < 1) {
497 CERROR("max_procs_per_node must be >= 1\n");
501 /* kptl_msg_t::ptlm_credits is only a __u8 */
502 if (*kptllnd_tunables.kptl_peercredits > 255) {
503 CERROR("kptl_peercredits must be <= 255\n");
507 *kptllnd_tunables.kptl_max_msg_size &= ~7;
508 if (*kptllnd_tunables.kptl_max_msg_size < PTLLND_MIN_BUFFER_SIZE)
509 *kptllnd_tunables.kptl_max_msg_size = PTLLND_MIN_BUFFER_SIZE;
511 CLASSERT ((PTLLND_MIN_BUFFER_SIZE & 7) == 0);
512 CLASSERT (sizeof(kptl_msg_t) <= PTLLND_MIN_BUFFER_SIZE);
515 * zero pointers, flags etc
516 * put everything into a known state.
518 memset (&kptllnd_data, 0, sizeof (kptllnd_data));
519 kptllnd_data.kptl_eqh = PTL_INVALID_HANDLE;
520 kptllnd_data.kptl_nih = PTL_INVALID_HANDLE;
523 * Setup the sched locks/lists/waitq
525 spin_lock_init(&kptllnd_data.kptl_sched_lock);
526 init_waitqueue_head(&kptllnd_data.kptl_sched_waitq);
527 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_txq);
528 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_rxq);
529 INIT_LIST_HEAD(&kptllnd_data.kptl_sched_rxbq);
531 /* init kptl_ptlid2str_lock before any call to kptllnd_ptlid2str */
532 spin_lock_init(&kptllnd_data.kptl_ptlid2str_lock);
535 * Setup the tx locks/lists
537 spin_lock_init(&kptllnd_data.kptl_tx_lock);
538 INIT_LIST_HEAD(&kptllnd_data.kptl_idle_txs);
539 atomic_set(&kptllnd_data.kptl_ntx, 0);
542 * Uptick the module reference count
547 * Setup pointers between the ni and context data block
549 kptllnd_data.kptl_ni = ni;
550 ni->ni_data = &kptllnd_data;
555 ni->ni_maxtxcredits = *kptllnd_tunables.kptl_credits;
556 ni->ni_peertxcredits = *kptllnd_tunables.kptl_peercredits;
558 kptllnd_data.kptl_expected_peers =
559 *kptllnd_tunables.kptl_max_nodes *
560 *kptllnd_tunables.kptl_max_procs_per_node;
563 * Initialize the Network interface instance
564 * We use the default because we don't have any
565 * way to choose a better interface.
566 * Requested and actual limits are ignored.
569 #ifdef _USING_LUSTRE_PORTALS_
574 *kptllnd_tunables.kptl_pid, NULL, NULL,
575 &kptllnd_data.kptl_nih);
578 * Note: PTL_IFACE_DUP simply means that the requested
579 * interface was already inited and that we're sharing it.
582 if (ptl_rc != PTL_OK && ptl_rc != PTL_IFACE_DUP) {
583 CERROR ("PtlNIInit: error %s(%d)\n",
584 kptllnd_errtype2str(ptl_rc), ptl_rc);
589 /* NB eq size irrelevant if using a callback */
590 ptl_rc = PtlEQAlloc(kptllnd_data.kptl_nih,
592 kptllnd_eq_callback, /* handler callback */
593 &kptllnd_data.kptl_eqh); /* output handle */
594 if (ptl_rc != PTL_OK) {
595 CERROR("PtlEQAlloc failed %s(%d)\n",
596 kptllnd_errtype2str(ptl_rc), ptl_rc);
602 * Fetch the lower NID
604 ptl_rc = PtlGetId(kptllnd_data.kptl_nih,
605 &kptllnd_data.kptl_portals_id);
606 if (ptl_rc != PTL_OK) {
607 CERROR ("PtlGetID: error %s(%d)\n",
608 kptllnd_errtype2str(ptl_rc), ptl_rc);
613 if (kptllnd_data.kptl_portals_id.pid != *kptllnd_tunables.kptl_pid) {
614 /* The kernel ptllnd must have the expected PID */
615 CERROR("Unexpected PID: %u (%u expected)\n",
616 kptllnd_data.kptl_portals_id.pid,
617 *kptllnd_tunables.kptl_pid);
622 ni->ni_nid = kptllnd_ptl2lnetnid(kptllnd_data.kptl_portals_id.nid);
624 CDEBUG(D_NET, "ptl id=%s, lnet id=%s\n",
625 kptllnd_ptlid2str(kptllnd_data.kptl_portals_id),
626 libcfs_nid2str(ni->ni_nid));
628 /* Initialized the incarnation - it must be for-all-time unique, even
629 * accounting for the fact that we increment it when we disconnect a
630 * peer that's using it */
631 do_gettimeofday(&tv);
632 kptllnd_data.kptl_incarnation = (((__u64)tv.tv_sec) * 1000000) +
634 CDEBUG(D_NET, "Incarnation="LPX64"\n", kptllnd_data.kptl_incarnation);
637 * Allocate and setup the peer hash table
639 rwlock_init(&kptllnd_data.kptl_peer_rw_lock);
640 init_waitqueue_head(&kptllnd_data.kptl_watchdog_waitq);
641 INIT_LIST_HEAD(&kptllnd_data.kptl_closing_peers);
642 INIT_LIST_HEAD(&kptllnd_data.kptl_zombie_peers);
644 kptllnd_data.kptl_peer_hash_size =
645 *kptllnd_tunables.kptl_peer_hash_table_size;
646 LIBCFS_ALLOC(kptllnd_data.kptl_peers,
647 (kptllnd_data.kptl_peer_hash_size *
648 sizeof(struct list_head)));
649 if (kptllnd_data.kptl_peers == NULL) {
650 CERROR("Failed to allocate space for peer hash table size=%d\n",
651 kptllnd_data.kptl_peer_hash_size);
655 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++)
656 INIT_LIST_HEAD(&kptllnd_data.kptl_peers[i]);
658 LIBCFS_ALLOC(kptllnd_data.kptl_nak_msg, offsetof(kptl_msg_t, ptlm_u));
659 if (kptllnd_data.kptl_nak_msg == NULL) {
660 CERROR("Can't allocate NAK msg\n");
664 memset(kptllnd_data.kptl_nak_msg, 0, offsetof(kptl_msg_t, ptlm_u));
665 kptllnd_init_msg(kptllnd_data.kptl_nak_msg, PTLLND_MSG_TYPE_NAK, 0);
666 kptllnd_data.kptl_nak_msg->ptlm_magic = PTLLND_MSG_MAGIC;
667 kptllnd_data.kptl_nak_msg->ptlm_version = PTLLND_MSG_VERSION;
668 kptllnd_data.kptl_nak_msg->ptlm_srcpid = the_lnet.ln_pid;
669 kptllnd_data.kptl_nak_msg->ptlm_srcnid = ni->ni_nid;
670 kptllnd_data.kptl_nak_msg->ptlm_srcstamp = kptllnd_data.kptl_incarnation;
671 kptllnd_data.kptl_nak_msg->ptlm_dstpid = LNET_PID_ANY;
672 kptllnd_data.kptl_nak_msg->ptlm_dstnid = LNET_NID_ANY;
674 kptllnd_rx_buffer_pool_init(&kptllnd_data.kptl_rx_buffer_pool);
676 kptllnd_data.kptl_rx_cache =
677 cfs_mem_cache_create("ptllnd_rx",
679 *kptllnd_tunables.kptl_max_msg_size,
682 if (kptllnd_data.kptl_rx_cache == NULL) {
683 CERROR("Can't create slab for RX descriptors\n");
688 /* lists/ptrs/locks initialised */
689 kptllnd_data.kptl_init = PTLLND_INIT_DATA;
691 /*****************************************************/
693 rc = kptllnd_setup_tx_descs();
695 CERROR("Can't pre-allocate %d TX descriptors: %d\n",
696 *kptllnd_tunables.kptl_ntx, rc);
700 /* Start the scheduler threads for handling incoming requests. No need
701 * to advance the state because this will be automatically cleaned up
702 * now that PTLNAT_INIT_DATA state has been entered */
703 CDEBUG(D_NET, "starting %d scheduler threads\n", PTLLND_N_SCHED);
704 for (i = 0; i < PTLLND_N_SCHED; i++) {
705 rc = kptllnd_thread_start(kptllnd_scheduler, (void *)((long)i));
707 CERROR("Can't spawn scheduler[%d]: %d\n", i, rc);
712 rc = kptllnd_thread_start(kptllnd_watchdog, NULL);
714 CERROR("Can't spawn watchdog: %d\n", rc);
718 /* Ensure that 'rxb_nspare' buffers can be off the net (being emptied)
719 * and we will still have enough buffers posted for all our peers */
720 spares = *kptllnd_tunables.kptl_rxb_nspare *
721 ((*kptllnd_tunables.kptl_rxb_npages * PAGE_SIZE)/
722 *kptllnd_tunables.kptl_max_msg_size);
724 /* reserve and post the buffers */
725 rc = kptllnd_rx_buffer_pool_reserve(&kptllnd_data.kptl_rx_buffer_pool,
726 kptllnd_data.kptl_expected_peers +
729 CERROR("Can't reserve RX Buffer pool: %d\n", rc);
733 /* flag everything initialised */
734 kptllnd_data.kptl_init = PTLLND_INIT_ALL;
736 /*****************************************************/
738 if (*kptllnd_tunables.kptl_checksum)
739 CWARN("Checksumming enabled\n");
741 CDEBUG(D_NET, "<<< kptllnd_startup SUCCESS\n");
745 CDEBUG(D_NET, "kptllnd_startup failed rc=%d\n", rc);
746 kptllnd_shutdown(ni);
751 kptllnd_shutdown (lnet_ni_t *ni)
755 lnet_process_id_t process_id;
758 CDEBUG(D_MALLOC, "before LND cleanup: kmem %d\n",
759 atomic_read (&libcfs_kmemory));
761 LASSERT (ni == kptllnd_data.kptl_ni);
763 switch (kptllnd_data.kptl_init) {
767 case PTLLND_INIT_ALL:
768 case PTLLND_INIT_DATA:
770 kptllnd_rx_buffer_pool_fini(&kptllnd_data.kptl_rx_buffer_pool);
771 LASSERT (list_empty(&kptllnd_data.kptl_sched_rxq));
772 LASSERT (list_empty(&kptllnd_data.kptl_sched_rxbq));
774 /* Hold peertable lock to interleave cleanly with peer birth/death */
775 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
777 LASSERT (kptllnd_data.kptl_shutdown == 0);
778 kptllnd_data.kptl_shutdown = 1; /* phase 1 == destroy peers */
780 /* no new peers possible now */
781 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
784 /* nuke all existing peers */
785 process_id.nid = LNET_NID_ANY;
786 process_id.pid = LNET_PID_ANY;
787 kptllnd_peer_del(process_id);
789 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
791 LASSERT (kptllnd_data.kptl_n_active_peers == 0);
794 while (kptllnd_data.kptl_npeers != 0) {
796 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
797 "Waiting for %d peers to terminate\n",
798 kptllnd_data.kptl_npeers);
800 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
803 cfs_pause(cfs_time_seconds(1));
805 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock,
809 LASSERT(list_empty(&kptllnd_data.kptl_closing_peers));
810 LASSERT(list_empty(&kptllnd_data.kptl_zombie_peers));
811 LASSERT (kptllnd_data.kptl_peers != NULL);
812 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++)
813 LASSERT (list_empty (&kptllnd_data.kptl_peers[i]));
815 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
816 CDEBUG(D_NET, "All peers deleted\n");
818 /* Shutdown phase 2: kill the daemons... */
819 kptllnd_data.kptl_shutdown = 2;
823 while (atomic_read (&kptllnd_data.kptl_nthreads) != 0) {
824 /* Wake up all threads*/
825 wake_up_all(&kptllnd_data.kptl_sched_waitq);
826 wake_up_all(&kptllnd_data.kptl_watchdog_waitq);
829 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
830 "Waiting for %d threads to terminate\n",
831 atomic_read(&kptllnd_data.kptl_nthreads));
832 cfs_pause(cfs_time_seconds(1));
835 CDEBUG(D_NET, "All Threads stopped\n");
836 LASSERT(list_empty(&kptllnd_data.kptl_sched_txq));
838 kptllnd_cleanup_tx_descs();
840 /* Nothing here now, but libcfs might soon require
841 * us to explicitly destroy wait queues and semaphores
842 * that would be done here */
846 case PTLLND_INIT_NOTHING:
847 CDEBUG(D_NET, "PTLLND_INIT_NOTHING\n");
851 if (!PtlHandleIsEqual(kptllnd_data.kptl_eqh, PTL_INVALID_HANDLE)) {
852 prc = PtlEQFree(kptllnd_data.kptl_eqh);
854 CERROR("Error %s(%d) freeing portals EQ\n",
855 kptllnd_errtype2str(prc), prc);
858 if (!PtlHandleIsEqual(kptllnd_data.kptl_nih, PTL_INVALID_HANDLE)) {
859 prc = PtlNIFini(kptllnd_data.kptl_nih);
861 CERROR("Error %s(%d) finalizing portals NI\n",
862 kptllnd_errtype2str(prc), prc);
865 LASSERT (atomic_read(&kptllnd_data.kptl_ntx) == 0);
866 LASSERT (list_empty(&kptllnd_data.kptl_idle_txs));
868 if (kptllnd_data.kptl_rx_cache != NULL)
869 cfs_mem_cache_destroy(kptllnd_data.kptl_rx_cache);
871 if (kptllnd_data.kptl_peers != NULL)
872 LIBCFS_FREE (kptllnd_data.kptl_peers,
873 sizeof (struct list_head) *
874 kptllnd_data.kptl_peer_hash_size);
876 if (kptllnd_data.kptl_nak_msg != NULL)
877 LIBCFS_FREE (kptllnd_data.kptl_nak_msg,
878 offsetof(kptl_msg_t, ptlm_u));
880 memset(&kptllnd_data, 0, sizeof(kptllnd_data));
882 CDEBUG(D_MALLOC, "after LND cleanup: kmem %d\n",
883 atomic_read (&libcfs_kmemory));
889 kptllnd_module_init (void)
893 kptllnd_assert_wire_constants();
895 rc = kptllnd_tunables_init();
899 kptllnd_init_ptltrace();
901 lnet_register_lnd(&kptllnd_lnd);
907 kptllnd_module_fini (void)
909 lnet_unregister_lnd(&kptllnd_lnd);
910 kptllnd_tunables_fini();
913 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
914 MODULE_DESCRIPTION("Kernel Portals LND v1.00");
915 MODULE_LICENSE("GPL");
917 module_init(kptllnd_module_init);
918 module_exit(kptllnd_module_fini);