1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2005 Cluster File Systems, Inc. All rights reserved.
5 * Author: PJ Kirner <pjkirner@clusterfs.com>
6 * E Barton <eeb@bartonsoftware.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * This file is confidential source code owned by Cluster File Systems.
12 * No viewing, modification, compilation, redistribution, or any other
13 * form of use is permitted except through a signed license agreement.
15 * If you have not signed such an agreement, then you have no rights to
16 * this file. Please destroy it immediately and contact CFS.
21 #include <libcfs/list.h>
24 kptllnd_count_queue(struct list_head *q)
37 kptllnd_get_peer_info(int index,
38 lnet_process_id_t *id,
39 int *state, int *sent_hello,
40 int *refcount, __u64 *incarnation,
41 __u64 *next_matchbits, __u64 *last_matchbits_seen,
42 int *nsendq, int *nactiveq,
43 int *credits, int *outstanding_credits)
45 rwlock_t *g_lock = &kptllnd_data.kptl_peer_rw_lock;
47 struct list_head *ptmp;
52 read_lock_irqsave(g_lock, flags);
54 for (i = 0; i < kptllnd_data.kptl_peer_hash_size; i++) {
56 list_for_each (ptmp, &kptllnd_data.kptl_peers[i]) {
57 peer = list_entry(ptmp, kptl_peer_t, peer_list);
63 *state = peer->peer_state;
64 *sent_hello = peer->peer_sent_hello;
65 *refcount = atomic_read(&peer->peer_refcount);
66 *incarnation = peer->peer_incarnation;
68 spin_lock(&peer->peer_lock);
70 *next_matchbits = peer->peer_next_matchbits;
71 *last_matchbits_seen = peer->peer_last_matchbits_seen;
72 *credits = peer->peer_credits;
73 *outstanding_credits = peer->peer_outstanding_credits;
75 *nsendq = kptllnd_count_queue(&peer->peer_sendq);
76 *nactiveq = kptllnd_count_queue(&peer->peer_activeq);
78 spin_unlock(&peer->peer_lock);
86 read_unlock_irqrestore(g_lock, flags);
91 kptllnd_peer_add_peertable_locked (kptl_peer_t *peer)
93 LASSERT (kptllnd_data.kptl_n_active_peers <
94 kptllnd_data.kptl_expected_peers);
96 LASSERT (peer->peer_state == PEER_STATE_WAITING_HELLO ||
97 peer->peer_state == PEER_STATE_ACTIVE);
99 kptllnd_data.kptl_n_active_peers++;
100 atomic_inc(&peer->peer_refcount); /* +1 ref for the list */
102 /* NB add to HEAD of peer list for MRU order!
103 * (see kptllnd_cull_peertable) */
104 list_add(&peer->peer_list, kptllnd_nid2peerlist(peer->peer_id.nid));
108 kptllnd_cull_peertable_locked (lnet_process_id_t pid)
110 /* I'm about to add a new peer with this portals ID to the peer table,
111 * so (a) this peer should not exist already and (b) I want to leave at
112 * most (max_procs_per_nid - 1) peers with this NID in the table. */
113 struct list_head *peers = kptllnd_nid2peerlist(pid.nid);
114 int cull_count = *kptllnd_tunables.kptl_max_procs_per_node;
116 struct list_head *tmp;
117 struct list_head *nxt;
121 list_for_each_safe (tmp, nxt, peers) {
122 /* NB I rely on kptllnd_peer_add_peertable_locked to add peers
124 peer = list_entry(tmp, kptl_peer_t, peer_list);
126 if (peer->peer_id.nid != pid.nid)
129 LASSERT (peer->peer_id.pid != pid.pid);
133 if (count < cull_count) /* recent (don't cull) */
136 CDEBUG(D_NET, "Cull %s(%s)\n",
137 libcfs_id2str(peer->peer_id),
138 kptllnd_ptlid2str(peer->peer_ptlid));
140 kptllnd_peer_close_locked(peer, 0);
145 kptllnd_peer_allocate (lnet_process_id_t lpid, ptl_process_id_t ppid)
150 LIBCFS_ALLOC(peer, sizeof (*peer));
152 CERROR("Can't create peer %s (%s)\n",
154 kptllnd_ptlid2str(ppid));
158 memset(peer, 0, sizeof(*peer)); /* zero flags etc */
160 INIT_LIST_HEAD (&peer->peer_sendq);
161 INIT_LIST_HEAD (&peer->peer_activeq);
162 spin_lock_init (&peer->peer_lock);
164 peer->peer_state = PEER_STATE_ALLOCATED;
165 peer->peer_error = 0;
166 peer->peer_last_alive = cfs_time_current();
167 peer->peer_id = lpid;
168 peer->peer_ptlid = ppid;
169 peer->peer_credits = 1; /* enough for HELLO */
170 peer->peer_next_matchbits = PTL_RESERVED_MATCHBITS;
171 peer->peer_outstanding_credits = *kptllnd_tunables.kptl_peercredits - 1;
172 peer->peer_sent_credits = 1; /* HELLO credit is implicit */
173 peer->peer_max_msg_size = PTLLND_MIN_BUFFER_SIZE; /* until we know better */
175 atomic_set(&peer->peer_refcount, 1); /* 1 ref for caller */
177 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
179 peer->peer_myincarnation = kptllnd_data.kptl_incarnation;
181 /* Only increase # peers under lock, to guarantee we dont grow it
183 if (kptllnd_data.kptl_shutdown) {
184 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
186 LIBCFS_FREE(peer, sizeof(*peer));
190 kptllnd_data.kptl_npeers++;
191 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
197 kptllnd_peer_destroy (kptl_peer_t *peer)
201 CDEBUG(D_NET, "Peer=%p\n", peer);
203 LASSERT (!in_interrupt());
204 LASSERT (atomic_read(&peer->peer_refcount) == 0);
205 LASSERT (peer->peer_state == PEER_STATE_ALLOCATED ||
206 peer->peer_state == PEER_STATE_ZOMBIE);
207 LASSERT (list_empty(&peer->peer_sendq));
208 LASSERT (list_empty(&peer->peer_activeq));
210 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
212 if (peer->peer_state == PEER_STATE_ZOMBIE)
213 list_del(&peer->peer_list);
215 kptllnd_data.kptl_npeers--;
217 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
219 LIBCFS_FREE (peer, sizeof (*peer));
223 kptllnd_peer_cancel_txs(kptl_peer_t *peer)
225 struct list_head sendq;
226 struct list_head activeq;
227 struct list_head *tmp;
228 struct list_head *nxt;
232 /* atomically grab all the peer's tx-es... */
234 spin_lock_irqsave(&peer->peer_lock, flags);
236 list_add(&sendq, &peer->peer_sendq);
237 list_del_init(&peer->peer_sendq);
238 list_for_each (tmp, &sendq) {
239 tx = list_entry(tmp, kptl_tx_t, tx_list);
243 list_add(&activeq, &peer->peer_activeq);
244 list_del_init(&peer->peer_activeq);
245 list_for_each (tmp, &activeq) {
246 tx = list_entry(tmp, kptl_tx_t, tx_list);
250 spin_unlock_irqrestore(&peer->peer_lock, flags);
252 /* ...then drop the peer's ref on them at leasure. This will get
253 * kptllnd_tx_fini() to abort outstanding comms if necessary. */
255 list_for_each_safe (tmp, nxt, &sendq) {
256 tx = list_entry(tmp, kptl_tx_t, tx_list);
257 list_del(&tx->tx_list);
258 tx->tx_status = -EIO;
259 kptllnd_tx_decref(tx);
262 list_for_each_safe (tmp, nxt, &activeq) {
263 tx = list_entry(tmp, kptl_tx_t, tx_list);
264 list_del(&tx->tx_list);
265 tx->tx_status = -EIO;
266 kptllnd_tx_decref(tx);
271 kptllnd_peer_alive (kptl_peer_t *peer)
273 /* This is racy, but everyone's only writing cfs_time_current() */
274 peer->peer_last_alive = cfs_time_current();
279 kptllnd_peer_notify (kptl_peer_t *peer)
282 time_t last_alive = 0;
285 spin_lock_irqsave(&peer->peer_lock, flags);
287 if (peer->peer_error != 0) {
288 error = peer->peer_error;
289 peer->peer_error = 0;
291 last_alive = cfs_time_current_sec() -
292 cfs_duration_sec(cfs_time_current() -
293 peer->peer_last_alive);
296 spin_unlock_irqrestore(&peer->peer_lock, flags);
299 lnet_notify (kptllnd_data.kptl_ni, peer->peer_id.nid, 0,
304 kptllnd_handle_closing_peers ()
308 struct list_head *tmp;
309 struct list_head *nxt;
312 /* Check with a read lock first to avoid blocking anyone */
314 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
315 idle = list_empty(&kptllnd_data.kptl_closing_peers);
316 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
321 /* Scan the closing peers and cancel their txs.
322 * NB only safe while there is only a single watchdog */
324 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
326 list_for_each_safe (tmp, nxt, &kptllnd_data.kptl_closing_peers) {
327 peer = list_entry (tmp, kptl_peer_t, peer_list);
329 LASSERT (peer->peer_state == PEER_STATE_CLOSING);
331 list_del(&peer->peer_list);
332 list_add_tail(&peer->peer_list,
333 &kptllnd_data.kptl_zombie_peers);
334 peer->peer_state = PEER_STATE_ZOMBIE;
336 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
338 kptllnd_peer_notify(peer);
339 kptllnd_peer_cancel_txs(peer);
340 kptllnd_peer_decref(peer);
342 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
345 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
349 kptllnd_peer_close_locked(kptl_peer_t *peer, int why)
351 switch (peer->peer_state) {
355 case PEER_STATE_WAITING_HELLO:
356 case PEER_STATE_ACTIVE:
357 /* Ensure new peers see a new incarnation of me */
358 LASSERT(peer->peer_myincarnation <= kptllnd_data.kptl_incarnation);
359 if (peer->peer_myincarnation == kptllnd_data.kptl_incarnation)
360 kptllnd_data.kptl_incarnation++;
362 /* Removing from peer table */
363 kptllnd_data.kptl_n_active_peers--;
364 LASSERT (kptllnd_data.kptl_n_active_peers >= 0);
366 list_del(&peer->peer_list);
367 kptllnd_peer_unreserve_buffers();
369 peer->peer_error = why; /* stash 'why' only on first close */
371 /* Schedule for immediate attention, taking peer table's ref */
372 list_add_tail(&peer->peer_list,
373 &kptllnd_data.kptl_closing_peers);
374 wake_up(&kptllnd_data.kptl_watchdog_waitq);
377 case PEER_STATE_ZOMBIE:
378 /* Schedule for attention at next timeout */
379 kptllnd_peer_addref(peer);
380 list_del(&peer->peer_list);
381 list_add_tail(&peer->peer_list,
382 &kptllnd_data.kptl_closing_peers);
385 case PEER_STATE_CLOSING:
389 peer->peer_state = PEER_STATE_CLOSING;
393 kptllnd_peer_close(kptl_peer_t *peer, int why)
397 write_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
398 kptllnd_peer_close_locked(peer, why);
399 write_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
403 kptllnd_peer_del(lnet_process_id_t id)
405 struct list_head *ptmp;
406 struct list_head *pnxt;
415 * Find the single bucket we are supposed to look at or if nid is a
416 * wildcard (LNET_NID_ANY) then look at all of the buckets
418 if (id.nid != LNET_NID_ANY) {
419 struct list_head *l = kptllnd_nid2peerlist(id.nid);
421 lo = hi = l - kptllnd_data.kptl_peers;
423 if (id.pid != LNET_PID_ANY)
427 hi = kptllnd_data.kptl_peer_hash_size - 1;
431 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
433 for (i = lo; i <= hi; i++) {
434 list_for_each_safe (ptmp, pnxt, &kptllnd_data.kptl_peers[i]) {
435 peer = list_entry (ptmp, kptl_peer_t, peer_list);
437 if (!(id.nid == LNET_NID_ANY ||
438 (peer->peer_id.nid == id.nid &&
439 (id.pid == LNET_PID_ANY ||
440 peer->peer_id.pid == id.pid))))
443 kptllnd_peer_addref(peer); /* 1 ref for me... */
445 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
448 kptllnd_peer_close(peer, 0);
449 kptllnd_peer_decref(peer); /* ...until here */
451 rc = 0; /* matched something */
453 /* start again now I've dropped the lock */
458 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
464 kptllnd_post_tx(kptl_peer_t *peer, kptl_tx_t *tx, int nfrag)
466 /* CAVEAT EMPTOR: I take over caller's ref on 'tx' */
467 ptl_handle_md_t rdma_mdh = PTL_INVALID_HANDLE;
468 ptl_handle_md_t msg_mdh = PTL_INVALID_HANDLE;
474 LASSERT (!tx->tx_idle);
475 LASSERT (!tx->tx_active);
476 LASSERT (PtlHandleIsEqual(tx->tx_msg_mdh, PTL_INVALID_HANDLE));
477 LASSERT (PtlHandleIsEqual(tx->tx_rdma_mdh, PTL_INVALID_HANDLE));
478 LASSERT (tx->tx_type == TX_TYPE_SMALL_MESSAGE ||
479 tx->tx_type == TX_TYPE_PUT_REQUEST ||
480 tx->tx_type == TX_TYPE_GET_REQUEST);
482 kptllnd_set_tx_peer(tx, peer);
484 if (tx->tx_type == TX_TYPE_PUT_REQUEST ||
485 tx->tx_type == TX_TYPE_GET_REQUEST) {
487 spin_lock_irqsave(&peer->peer_lock, flags);
489 /* Assume 64-bit matchbits can't wrap */
490 LASSERT (peer->peer_next_matchbits >= PTL_RESERVED_MATCHBITS);
491 tx->tx_msg->ptlm_u.rdma.kptlrm_matchbits =
492 peer->peer_next_matchbits++;
494 spin_unlock_irqrestore(&peer->peer_lock, flags);
496 prc = PtlMEAttach(kptllnd_data.kptl_nih,
497 *kptllnd_tunables.kptl_portal,
499 tx->tx_msg->ptlm_u.rdma.kptlrm_matchbits,
505 CERROR("PtlMEAttach(%s) failed: %d\n",
506 libcfs_id2str(peer->peer_id), prc);
510 prc = PtlMDAttach(meh, tx->tx_rdma_md, PTL_UNLINK, &rdma_mdh);
512 CERROR("PtlMDAttach(%s) failed: %d\n",
513 libcfs_id2str(tx->tx_peer->peer_id), prc);
514 prc = PtlMEUnlink(meh);
515 LASSERT(prc == PTL_OK);
516 rdma_mdh = PTL_INVALID_HANDLE;
520 /* I'm not racing with the event callback here. It's a bug if
521 * there's an event on the MD I just attached before I actually
522 * send the RDMA request message which the event callback
523 * catches by asserting 'rdma_mdh' is valid. */
526 memset(&md, 0, sizeof(md));
528 md.threshold = tx->tx_acked ? 2 : 1; /* SEND END + ACK? */
529 md.options = PTL_MD_OP_PUT |
530 PTL_MD_LUSTRE_COMPLETION_SEMANTICS |
531 PTL_MD_EVENT_START_DISABLE;
532 md.user_ptr = &tx->tx_msg_eventarg;
533 md.eq_handle = kptllnd_data.kptl_eqh;
536 md.start = tx->tx_msg;
537 md.length = tx->tx_msg->ptlm_nob;
540 LASSERT (tx->tx_frags->iov[0].iov_base == (void *)tx->tx_msg);
542 md.start = tx->tx_frags;
544 md.options |= PTL_MD_IOVEC;
547 prc = PtlMDBind(kptllnd_data.kptl_nih, md, PTL_UNLINK, &msg_mdh);
549 msg_mdh = PTL_INVALID_HANDLE;
553 spin_lock_irqsave(&peer->peer_lock, flags);
555 tx->tx_deadline = jiffies + (*kptllnd_tunables.kptl_timeout * HZ);
557 tx->tx_rdma_mdh = rdma_mdh;
558 tx->tx_msg_mdh = msg_mdh;
560 /* Ensure HELLO is sent first */
561 if (tx->tx_msg->ptlm_type == PTLLND_MSG_TYPE_HELLO)
562 list_add(&tx->tx_list, &peer->peer_sendq);
564 list_add_tail(&tx->tx_list, &peer->peer_sendq);
566 spin_unlock_irqrestore(&peer->peer_lock, flags);
570 spin_lock_irqsave(&peer->peer_lock, flags);
572 tx->tx_status = -EIO;
573 tx->tx_rdma_mdh = rdma_mdh;
574 tx->tx_msg_mdh = msg_mdh;
576 spin_unlock_irqrestore(&peer->peer_lock, flags);
578 kptllnd_tx_decref(tx);
582 kptllnd_peer_check_sends (kptl_peer_t *peer)
589 LASSERT(!in_interrupt());
591 spin_lock_irqsave(&peer->peer_lock, flags);
593 if (list_empty(&peer->peer_sendq) &&
594 peer->peer_outstanding_credits >= PTLLND_CREDIT_HIGHWATER &&
595 peer->peer_credits != 0) {
597 /* post a NOOP to return credits */
598 spin_unlock_irqrestore(&peer->peer_lock, flags);
600 tx = kptllnd_get_idle_tx(TX_TYPE_SMALL_MESSAGE);
602 CERROR("Can't return credits to %s: can't allocate descriptor\n",
603 libcfs_id2str(peer->peer_id));
605 kptllnd_init_msg(tx->tx_msg, PTLLND_MSG_TYPE_NOOP, 0);
606 kptllnd_post_tx(peer, tx, 0);
609 spin_lock_irqsave(&peer->peer_lock, flags);
612 while (!list_empty(&peer->peer_sendq)) {
613 tx = list_entry (peer->peer_sendq.next, kptl_tx_t, tx_list);
615 LASSERT (tx->tx_active);
616 LASSERT (!PtlHandleIsEqual(tx->tx_msg_mdh, PTL_INVALID_HANDLE));
617 LASSERT (tx->tx_type == TX_TYPE_SMALL_MESSAGE ||
618 !PtlHandleIsEqual(tx->tx_rdma_mdh, PTL_INVALID_HANDLE));
620 LASSERT (peer->peer_outstanding_credits >= 0);
621 LASSERT (peer->peer_sent_credits >= 0);
622 LASSERT (peer->peer_sent_credits +
623 peer->peer_outstanding_credits <=
624 *kptllnd_tunables.kptl_peercredits);
625 LASSERT (peer->peer_credits >= 0);
627 /* Ensure HELLO is sent first */
628 if (!peer->peer_sent_hello) {
629 if (tx->tx_msg->ptlm_type != PTLLND_MSG_TYPE_HELLO)
631 peer->peer_sent_hello = 1;
634 if (peer->peer_credits == 0) {
635 CDEBUG(D_NETTRACE, "%s[%d/%d+%d]: no credits for %p\n",
636 libcfs_id2str(peer->peer_id),
638 peer->peer_outstanding_credits,
639 peer->peer_sent_credits, tx);
643 /* Don't use the last credit unless I've got credits to
645 if (peer->peer_credits == 1 &&
646 peer->peer_outstanding_credits == 0) {
647 CDEBUG(D_NETTRACE, "%s[%d/%d+%d]: "
648 "not using last credit for %p\n",
649 libcfs_id2str(peer->peer_id),
651 peer->peer_outstanding_credits,
652 peer->peer_sent_credits, tx);
656 list_del(&tx->tx_list);
658 /* Discard any NOOP I queued if I'm not at the high-water mark
659 * any more or more messages have been queued */
660 if (tx->tx_msg->ptlm_type == PTLLND_MSG_TYPE_NOOP &&
661 (!list_empty(&peer->peer_sendq) ||
662 peer->peer_outstanding_credits < PTLLND_CREDIT_HIGHWATER)) {
666 spin_unlock_irqrestore(&peer->peer_lock, flags);
668 CDEBUG(D_NET, "%s: redundant noop\n",
669 libcfs_id2str(peer->peer_id));
670 kptllnd_tx_decref(tx);
672 spin_lock_irqsave(&peer->peer_lock, flags);
676 /* fill last-minute msg header fields */
677 kptllnd_msg_pack(tx->tx_msg, peer);
679 peer->peer_sent_credits += peer->peer_outstanding_credits;
680 peer->peer_outstanding_credits = 0;
681 peer->peer_credits--;
683 CDEBUG(D_NETTRACE, "%s[%d/%d+%d]: %s tx=%p nob=%d cred=%d\n",
684 libcfs_id2str(peer->peer_id), peer->peer_credits,
685 peer->peer_outstanding_credits, peer->peer_sent_credits,
686 kptllnd_msgtype2str(tx->tx_msg->ptlm_type),
687 tx, tx->tx_msg->ptlm_nob,
688 tx->tx_msg->ptlm_credits);
690 list_add_tail(&tx->tx_list, &peer->peer_activeq);
692 kptllnd_tx_addref(tx); /* 1 ref for me... */
694 spin_unlock_irqrestore(&peer->peer_lock, flags);
696 tx->tx_tposted = jiffies; /* going on the wire */
698 rc = PtlPut (tx->tx_msg_mdh,
699 tx->tx_acked ? PTL_ACK_REQ : PTL_NOACK_REQ,
701 *kptllnd_tunables.kptl_portal,
705 0); /* header data */
707 CERROR("PtlPut %s error %d\n",
708 libcfs_id2str(peer->peer_id), rc);
710 /* Nuke everything (including this tx) */
711 kptllnd_peer_close(peer, -EIO);
715 kptllnd_tx_decref(tx); /* drop my ref */
717 spin_lock_irqsave(&peer->peer_lock, flags);
720 spin_unlock_irqrestore(&peer->peer_lock, flags);
724 kptllnd_find_timed_out_tx(kptl_peer_t *peer)
727 struct list_head *tmp;
730 spin_lock_irqsave(&peer->peer_lock, flags);
732 list_for_each(tmp, &peer->peer_sendq) {
733 tx = list_entry(peer->peer_sendq.next, kptl_tx_t, tx_list);
735 if (time_after_eq(jiffies, tx->tx_deadline)) {
736 kptllnd_tx_addref(tx);
737 spin_unlock_irqrestore(&peer->peer_lock, flags);
742 list_for_each(tmp, &peer->peer_activeq) {
743 tx = list_entry(peer->peer_activeq.next, kptl_tx_t, tx_list);
745 if (time_after_eq(jiffies, tx->tx_deadline)) {
746 kptllnd_tx_addref(tx);
747 spin_unlock_irqrestore(&peer->peer_lock, flags);
752 spin_unlock_irqrestore(&peer->peer_lock, flags);
758 kptllnd_peer_check_bucket (int idx)
760 struct list_head *peers = &kptllnd_data.kptl_peers[idx];
761 struct list_head *ptmp;
768 CDEBUG(D_NET, "Bucket=%d\n", idx);
771 /* NB. Shared lock while I just look */
772 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
774 list_for_each (ptmp, peers) {
775 peer = list_entry (ptmp, kptl_peer_t, peer_list);
777 CDEBUG(D_NET, "Peer=%s Credits=%d Outstanding=%d Send=%d\n",
778 libcfs_id2str(peer->peer_id), peer->peer_credits,
779 peer->peer_outstanding_credits, peer->peer_sent_credits);
781 /* In case we have enough credits to return via a
782 * NOOP, but there were no non-blocking tx descs
783 * free to do it last time... */
784 kptllnd_peer_check_sends(peer);
786 tx = kptllnd_find_timed_out_tx(peer);
790 kptllnd_peer_addref(peer); /* 1 ref for me... */
792 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock,
795 spin_lock_irqsave(&peer->peer_lock, flags);
796 nsend = kptllnd_count_queue(&peer->peer_sendq);
797 nactive = kptllnd_count_queue(&peer->peer_activeq);
798 spin_unlock_irqrestore(&peer->peer_lock, flags);
800 LCONSOLE_ERROR("Timing out %s: %s\n",
801 libcfs_id2str(peer->peer_id),
802 (tx->tx_tposted == 0) ?
803 "no free peer buffers" : "please check Portals");
805 CERROR("%s timed out: cred %d outstanding %d, sent %d, "
806 "sendq %d, activeq %d Tx %p %s (%s%s%s) status %d "
807 "%sposted %lu T/O %ds\n",
808 libcfs_id2str(peer->peer_id), peer->peer_credits,
809 peer->peer_outstanding_credits, peer->peer_sent_credits,
810 nsend, nactive, tx, kptllnd_tx_typestr(tx->tx_type),
811 tx->tx_active ? "A" : "",
812 PtlHandleIsEqual(tx->tx_msg_mdh, PTL_INVALID_HANDLE) ?
814 PtlHandleIsEqual(tx->tx_rdma_mdh, PTL_INVALID_HANDLE) ?
817 (tx->tx_tposted == 0) ? "not " : "",
818 (tx->tx_tposted == 0) ? 0UL : (jiffies - tx->tx_tposted),
819 *kptllnd_tunables.kptl_timeout);
821 kptllnd_dump_ptltrace();
823 kptllnd_tx_decref(tx);
825 kptllnd_peer_close(peer, -ETIMEDOUT);
826 kptllnd_peer_decref(peer); /* ...until here */
828 /* start again now I've dropped the lock */
832 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
836 kptllnd_id2peer_locked (lnet_process_id_t id)
838 struct list_head *peers = kptllnd_nid2peerlist(id.nid);
839 struct list_head *tmp;
842 list_for_each (tmp, peers) {
844 peer = list_entry (tmp, kptl_peer_t, peer_list);
846 LASSERT(peer->peer_state == PEER_STATE_WAITING_HELLO ||
847 peer->peer_state == PEER_STATE_ACTIVE);
849 if (peer->peer_id.nid != id.nid ||
850 peer->peer_id.pid != id.pid)
853 kptllnd_peer_addref(peer);
855 CDEBUG(D_NET, "%s -> %s (%d)\n",
857 kptllnd_ptlid2str(peer->peer_ptlid),
858 atomic_read (&peer->peer_refcount));
866 kptllnd_peertable_overflow_msg(char *str, lnet_process_id_t id)
868 LCONSOLE_ERROR("%s %s overflows the peer table[%d]: "
869 "messages may be dropped\n",
870 str, libcfs_id2str(id),
871 kptllnd_data.kptl_n_active_peers);
872 LCONSOLE_ERROR("Please correct by increasing "
873 "'max_nodes' or 'max_procs_per_node'\n");
877 kptllnd_get_last_seen_matchbits_locked(lnet_process_id_t lpid)
880 struct list_head *tmp;
882 /* Find the last matchbits I saw this new peer using. Note..
883 A. This peer cannot be in the peer table - she's new!
884 B. If I can't find the peer in the closing/zombie peers, all
885 matchbits are safe because all refs to the (old) peer have gone
886 so all txs have completed so there's no risk of matchbit
890 LASSERT(kptllnd_id2peer_locked(lpid) == NULL);
892 /* peer's last matchbits can't change after it comes out of the peer
893 * table, so first match is fine */
895 list_for_each (tmp, &kptllnd_data.kptl_closing_peers) {
896 peer = list_entry (tmp, kptl_peer_t, peer_list);
898 if (peer->peer_id.nid == lpid.nid &&
899 peer->peer_id.pid == lpid.pid)
900 return peer->peer_last_matchbits_seen;
903 list_for_each (tmp, &kptllnd_data.kptl_zombie_peers) {
904 peer = list_entry (tmp, kptl_peer_t, peer_list);
906 if (peer->peer_id.nid == lpid.nid &&
907 peer->peer_id.pid == lpid.pid)
908 return peer->peer_last_matchbits_seen;
911 return PTL_RESERVED_MATCHBITS;
915 kptllnd_peer_handle_hello (ptl_process_id_t initiator,
918 rwlock_t *g_lock = &kptllnd_data.kptl_peer_rw_lock;
920 kptl_peer_t *new_peer;
921 lnet_process_id_t lpid;
925 __u64 safe_matchbits;
926 __u64 last_matchbits_seen;
928 lpid.nid = msg->ptlm_srcnid;
929 lpid.pid = msg->ptlm_srcpid;
931 CDEBUG(D_NET, "hello from %s(%s)\n",
932 libcfs_id2str(lpid), kptllnd_ptlid2str(initiator));
934 if (initiator.pid != kptllnd_data.kptl_portals_id.pid &&
935 (msg->ptlm_srcpid & LNET_PID_USERFLAG) == 0) {
936 /* If the peer's PID isn't _the_ ptllnd kernel pid, she must be
937 * userspace. Refuse the connection if she hasn't set the
938 * correct flag in her PID... */
939 CERROR("Userflag not set in hello from %s (%s)\n",
940 libcfs_id2str(lpid), kptllnd_ptlid2str(initiator));
944 /* kptlhm_matchbits are the highest matchbits my peer may have used to
945 * RDMA to me. I ensure I never register buffers for RDMA that could
946 * match any she used */
947 safe_matchbits = msg->ptlm_u.hello.kptlhm_matchbits + 1;
949 if (safe_matchbits < PTL_RESERVED_MATCHBITS) {
950 CERROR("Illegal matchbits "LPX64" in HELLO from %s\n",
951 safe_matchbits, libcfs_id2str(lpid));
955 if (msg->ptlm_u.hello.kptlhm_max_msg_size < PTLLND_MIN_BUFFER_SIZE) {
956 CERROR("%s: max message size %d < MIN %d",
958 msg->ptlm_u.hello.kptlhm_max_msg_size,
959 *kptllnd_tunables.kptl_max_msg_size);
963 if (msg->ptlm_credits <= 1) {
964 CERROR("Need more than 1+%d credits from %s\n",
965 msg->ptlm_credits, libcfs_id2str(lpid));
969 write_lock_irqsave(g_lock, flags);
971 peer = kptllnd_id2peer_locked(lpid);
973 if (peer->peer_state == PEER_STATE_WAITING_HELLO) {
974 /* Completing HELLO handshake */
975 LASSERT(peer->peer_incarnation == 0);
977 if (msg->ptlm_dststamp != 0 &&
978 msg->ptlm_dststamp != peer->peer_myincarnation) {
979 write_unlock_irqrestore(g_lock, flags);
981 CERROR("Ignoring HELLO from %s: unexpected "
982 "dststamp "LPX64" ("LPX64" wanted)\n",
985 peer->peer_myincarnation);
986 kptllnd_peer_decref(peer);
990 /* Concurrent initiation or response to my HELLO */
991 peer->peer_state = PEER_STATE_ACTIVE;
992 peer->peer_incarnation = msg->ptlm_srcstamp;
993 peer->peer_next_matchbits = safe_matchbits;
994 peer->peer_max_msg_size =
995 msg->ptlm_u.hello.kptlhm_max_msg_size;
997 write_unlock_irqrestore(g_lock, flags);
1001 if (msg->ptlm_dststamp != 0 &&
1002 msg->ptlm_dststamp <= peer->peer_myincarnation) {
1003 write_unlock_irqrestore(g_lock, flags);
1005 CERROR("Ignoring stale HELLO from %s: "
1006 "dststamp "LPX64" (current "LPX64")\n",
1007 libcfs_id2str(lpid),
1009 peer->peer_myincarnation);
1010 kptllnd_peer_decref(peer);
1014 /* Brand new connection attempt: remove old incarnation */
1015 kptllnd_peer_close_locked(peer, 0);
1018 kptllnd_cull_peertable_locked(lpid);
1020 write_unlock_irqrestore(g_lock, flags);
1023 CDEBUG(D_NET, "Peer %s (%s) reconnecting:"
1024 " stamp "LPX64"("LPX64")\n",
1025 libcfs_id2str(lpid), kptllnd_ptlid2str(initiator),
1026 msg->ptlm_srcstamp, peer->peer_incarnation);
1028 kptllnd_peer_decref(peer);
1031 hello_tx = kptllnd_get_idle_tx(TX_TYPE_SMALL_MESSAGE);
1032 if (hello_tx == NULL) {
1033 CERROR("Unable to allocate HELLO message for %s\n",
1034 libcfs_id2str(lpid));
1038 kptllnd_init_msg(hello_tx->tx_msg, PTLLND_MSG_TYPE_HELLO,
1039 sizeof(kptl_hello_msg_t));
1041 new_peer = kptllnd_peer_allocate(lpid, initiator);
1042 if (new_peer == NULL) {
1043 kptllnd_tx_decref(hello_tx);
1047 rc = kptllnd_peer_reserve_buffers();
1049 kptllnd_peer_decref(new_peer);
1050 kptllnd_tx_decref(hello_tx);
1052 CERROR("Failed to reserve buffers for %s\n",
1053 libcfs_id2str(lpid));
1057 write_lock_irqsave(g_lock, flags);
1059 peer = kptllnd_id2peer_locked(lpid);
1061 if (peer->peer_state == PEER_STATE_WAITING_HELLO) {
1062 /* An outgoing message instantiated 'peer' for me */
1063 CWARN("Outgoing instantiated peer %s\n", libcfs_id2str(lpid));
1064 LASSERT(peer->peer_incarnation == 0);
1066 peer->peer_state = PEER_STATE_ACTIVE;
1067 peer->peer_incarnation = msg->ptlm_srcstamp;
1068 peer->peer_next_matchbits = safe_matchbits;
1069 peer->peer_max_msg_size =
1070 msg->ptlm_u.hello.kptlhm_max_msg_size;
1072 LASSERT (peer->peer_state == PEER_STATE_ACTIVE);
1073 /* WOW! Somehow this peer completed the HELLO
1074 * handshake while I slept. I guess I could have slept
1075 * while it rebooted and sent a new HELLO, so I'll fail
1077 CWARN("Wow! peer %s\n", libcfs_id2str(lpid));
1078 kptllnd_peer_decref(peer);
1082 write_unlock_irqrestore(g_lock, flags);
1084 kptllnd_peer_unreserve_buffers();
1085 kptllnd_peer_decref(new_peer);
1086 kptllnd_tx_decref(hello_tx);
1090 if (kptllnd_data.kptl_n_active_peers ==
1091 kptllnd_data.kptl_expected_peers) {
1092 /* peer table full */
1093 write_unlock_irqrestore(g_lock, flags);
1095 kptllnd_peertable_overflow_msg("Connection from ", lpid);
1097 rc = kptllnd_reserve_buffers(1); /* HELLO headroom */
1099 CERROR("Refusing connection from %s\n",
1100 libcfs_id2str(lpid));
1101 kptllnd_peer_unreserve_buffers();
1102 kptllnd_peer_decref(new_peer);
1103 kptllnd_tx_decref(hello_tx);
1107 write_lock_irqsave(g_lock, flags);
1108 kptllnd_data.kptl_expected_peers++;
1112 last_matchbits_seen = kptllnd_get_last_seen_matchbits_locked(lpid);
1114 hello_tx->tx_msg->ptlm_u.hello.kptlhm_matchbits = last_matchbits_seen;
1115 hello_tx->tx_msg->ptlm_u.hello.kptlhm_max_msg_size =
1116 *kptllnd_tunables.kptl_max_msg_size;
1118 new_peer->peer_state = PEER_STATE_ACTIVE;
1119 new_peer->peer_incarnation = msg->ptlm_srcstamp;
1120 new_peer->peer_next_matchbits = safe_matchbits;
1121 new_peer->peer_last_matchbits_seen = last_matchbits_seen;
1122 new_peer->peer_max_msg_size = msg->ptlm_u.hello.kptlhm_max_msg_size;
1124 kptllnd_peer_add_peertable_locked(new_peer);
1126 write_unlock_irqrestore(g_lock, flags);
1128 /* NB someone else could get in now and post a message before I post
1129 * the HELLO, but post_tx/check_sends take care of that! */
1131 CDEBUG(D_NETTRACE, "%s: post response hello %p\n",
1132 libcfs_id2str(new_peer->peer_id), hello_tx);
1134 kptllnd_post_tx(new_peer, hello_tx, 0);
1135 kptllnd_peer_check_sends(new_peer);
1141 kptllnd_tx_launch(kptl_peer_t *peer, kptl_tx_t *tx, int nfrag)
1143 kptllnd_post_tx(peer, tx, nfrag);
1144 kptllnd_peer_check_sends(peer);
1148 kptllnd_find_target(kptl_peer_t **peerp, lnet_process_id_t target)
1150 rwlock_t *g_lock = &kptllnd_data.kptl_peer_rw_lock;
1151 ptl_process_id_t ptl_id;
1152 kptl_peer_t *new_peer;
1153 kptl_tx_t *hello_tx;
1154 unsigned long flags;
1156 __u64 last_matchbits_seen;
1158 /* I expect to find the peer, so I only take a read lock... */
1159 read_lock_irqsave(g_lock, flags);
1160 *peerp = kptllnd_id2peer_locked(target);
1161 read_unlock_irqrestore(g_lock, flags);
1166 if ((target.pid & LNET_PID_USERFLAG) != 0) {
1167 CWARN("Refusing to create a new connection to %s "
1168 "(non-kernel peer)\n", libcfs_id2str(target));
1169 return -EHOSTUNREACH;
1172 /* The new peer is a kernel ptllnd, and kernel ptllnds all have
1173 * the same portals PID */
1174 ptl_id.nid = kptllnd_lnet2ptlnid(target.nid);
1175 ptl_id.pid = kptllnd_data.kptl_portals_id.pid;
1177 hello_tx = kptllnd_get_idle_tx(TX_TYPE_SMALL_MESSAGE);
1178 if (hello_tx == NULL) {
1179 CERROR("Unable to allocate connect message for %s\n",
1180 libcfs_id2str(target));
1184 kptllnd_init_msg(hello_tx->tx_msg, PTLLND_MSG_TYPE_HELLO,
1185 sizeof(kptl_hello_msg_t));
1187 new_peer = kptllnd_peer_allocate(target, ptl_id);
1188 if (new_peer == NULL) {
1193 rc = kptllnd_peer_reserve_buffers();
1197 write_lock_irqsave(g_lock, flags);
1199 *peerp = kptllnd_id2peer_locked(target);
1200 if (*peerp != NULL) {
1201 write_unlock_irqrestore(g_lock, flags);
1205 kptllnd_cull_peertable_locked(target);
1207 if (kptllnd_data.kptl_n_active_peers ==
1208 kptllnd_data.kptl_expected_peers) {
1209 /* peer table full */
1210 write_unlock_irqrestore(g_lock, flags);
1212 kptllnd_peertable_overflow_msg("Connection to ", target);
1214 rc = kptllnd_reserve_buffers(1); /* HELLO headroom */
1216 CERROR("Can't create connection to %s\n",
1217 libcfs_id2str(target));
1221 write_lock_irqsave(g_lock, flags);
1222 kptllnd_data.kptl_expected_peers++;
1226 last_matchbits_seen = kptllnd_get_last_seen_matchbits_locked(target);
1228 hello_tx->tx_msg->ptlm_u.hello.kptlhm_matchbits = last_matchbits_seen;
1229 hello_tx->tx_msg->ptlm_u.hello.kptlhm_max_msg_size =
1230 *kptllnd_tunables.kptl_max_msg_size;
1232 new_peer->peer_state = PEER_STATE_WAITING_HELLO;
1233 new_peer->peer_last_matchbits_seen = last_matchbits_seen;
1235 kptllnd_peer_add_peertable_locked(new_peer);
1237 write_unlock_irqrestore(g_lock, flags);
1239 /* NB someone else could get in now and post a message before I post
1240 * the HELLO, but post_tx/check_sends take care of that! */
1242 CDEBUG(D_NETTRACE, "%s: post initial hello %p\n",
1243 libcfs_id2str(new_peer->peer_id), hello_tx);
1245 kptllnd_post_tx(new_peer, hello_tx, 0);
1246 kptllnd_peer_check_sends(new_peer);
1252 kptllnd_peer_unreserve_buffers();
1254 kptllnd_peer_decref(new_peer);
1256 kptllnd_tx_decref(hello_tx);