2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by Eric Barton <eric@bartonsoftware.com>
7 * Author: James Shimek <jshimek@cray.com>
8 * Author: Nic Henke <nic@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #include <linux/nmi.h>
31 /* this is useful when needed to debug wire corruption. */
33 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
41 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
42 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
45 } else if (len >= 16) {
47 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
48 prefix, ptr, *(ptr), *(ptr + 1));
52 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
61 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
63 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
64 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
65 msg->gnm_magic, msg->gnm_version,
66 msg->gnm_type, msg->gnm_srcnid,
67 msg->gnm_connstamp, msg->gnm_seq,
68 msg->gnm_cksum, msg->gnm_payload_cksum,
69 msg->gnm_payload_len);
73 kgnilnd_schedule_device(kgn_device_t *dev)
75 short already_live = 0;
77 /* we'll only want to wake if the scheduler thread
78 * has come around and set ready to zero */
79 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
82 wake_up_all(&dev->gnd_waitq);
87 void kgnilnd_schedule_device_timer(unsigned long arg)
89 kgn_device_t *dev = (kgn_device_t *) arg;
91 kgnilnd_schedule_device(dev);
95 kgnilnd_device_callback(__u32 devid, __u64 arg)
98 int index = (int) arg;
100 if (index >= kgnilnd_data.kgn_ndevs) {
101 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
102 * interrupt context. */
103 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
108 dev = &kgnilnd_data.kgn_devices[index];
109 /* just basic sanity */
110 if (dev->gnd_id == devid) {
111 kgnilnd_schedule_device(dev);
113 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
118 /* sched_intent values:
119 * < 0 : do not reschedule under any circumstances
120 * == 0: reschedule if someone marked him WANTS_SCHED
121 * > 0 : force a reschedule */
122 /* Return code 0 means it did not schedule the conn, 1
123 * means it successfully scheduled the conn.
127 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
131 /* move back to IDLE but save previous state.
132 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
133 * let the xchg there handle any racing callers to get it
134 * onto gnd_ready_conns */
136 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
137 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
138 conn_sched == GNILND_CONN_PROCESS,
139 "conn %p after process in bad state: %d\n",
142 if (sched_intent >= 0) {
143 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
144 return kgnilnd_schedule_conn_refheld(conn, 1);
150 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
154 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld)
156 kgn_device_t *dev = conn->gnc_device;
160 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
161 /* we only care about the last person who marked want_sched since they
162 * are most likely the culprit
164 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
165 conn->gnc_sched_line = line;
166 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
167 * the chance to put it onto gnd_ready_conns.
168 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
169 * the conn in process_conns will take care of moving it back to
170 * SCHED when it is done processing */
172 if (sched == GNILND_CONN_IDLE) {
173 /* if the conn is already scheduled, we've already requested
174 * the scheduler thread wakeup */
176 /* Add a reference to the conn if we are not holding a reference
177 * already from the exisiting scheduler. We now use the same
178 * reference if we need to reschedule a conn while in a scheduler
181 kgnilnd_conn_addref(conn);
183 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
186 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
188 spin_lock(&dev->gnd_lock);
189 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
190 spin_unlock(&dev->gnd_lock);
191 set_mb(conn->gnc_last_sched_ask, jiffies);
194 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
198 /* make sure thread(s) going to process conns - but let it make
199 * separate decision from conn schedule */
200 kgnilnd_schedule_device(dev);
205 kgnilnd_schedule_dgram(kgn_device_t *dev)
209 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
210 if (wake != GNILND_DGRAM_SCHED) {
211 wake_up(&dev->gnd_dgram_waitq);
213 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
218 kgnilnd_free_tx(kgn_tx_t *tx)
220 /* taken from kgnilnd_tx_add_state_locked */
222 LASSERTF((tx->tx_list_p == NULL &&
223 tx->tx_list_state == GNILND_TX_ALLOCD) &&
224 list_empty(&tx->tx_list),
225 "tx %p with bad state %s (list_p %p) tx_list %s\n",
226 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
227 list_empty(&tx->tx_list) ? "empty" : "not empty");
229 atomic_dec(&kgnilnd_data.kgn_ntx);
231 /* we only allocate this if we need to */
232 if (tx->tx_phys != NULL) {
233 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
234 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
235 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
238 /* Only free the buffer if we used it */
239 if (tx->tx_buffer_copy != NULL) {
240 vfree(tx->tx_buffer_copy);
241 tx->tx_buffer_copy = NULL;
242 CDEBUG(D_MALLOC, "vfreed buffer2\n");
245 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
247 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
248 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
252 kgnilnd_alloc_tx (void)
256 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
259 tx = kmem_cache_alloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
261 CERROR("failed to allocate tx\n");
264 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
267 /* need this memset, cache alloc'd memory is not cleared */
268 memset(tx, 0, sizeof(*tx));
270 /* setup everything here to minimize time under the lock */
271 tx->tx_buftype = GNILND_BUF_NONE;
272 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
273 INIT_LIST_HEAD(&tx->tx_list);
274 INIT_LIST_HEAD(&tx->tx_map_list);
275 tx->tx_list_state = GNILND_TX_ALLOCD;
277 atomic_inc(&kgnilnd_data.kgn_ntx);
282 /* csum_fold needs to be run on the return value before shipping over the wire */
283 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
285 /* we don't use offset as every one is passing a buffer reference that already
286 * includes the offset into the base address -
287 * see kgnilnd_setup_virt_buffer and kgnilnd_setup_immediate_buffer */
289 kgnilnd_cksum(void *ptr, size_t nob)
293 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
295 /* don't use magic 'no checksum' value */
299 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
306 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov,
307 unsigned int offset, unsigned int nob, int dump_blob)
313 unsigned int fraglen;
319 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
320 kiov, nkiov, offset, nob, dump_blob);
322 /* if loops changes, please change kgnilnd_setup_phys_buffer */
324 while (offset >= kiov->kiov_len) {
325 offset -= kiov->kiov_len;
331 /* ignore nob here, if nob < (kiov_len - offset), kiov == 1 */
332 odd = (unsigned long) (kiov[0].kiov_len - offset) & 1;
334 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
335 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
337 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
338 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
340 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
341 odd, kiov[0].kiov_len, offset, nob);
343 for (i = 0; i < nkiov; i++) {
344 pages[i] = kiov[i].kiov_page;
347 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
349 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
351 /* return zero to avoid killing tx - we'll just get warning on console
352 * when remote end sees zero checksum */
355 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
357 tmpck = _kgnilnd_cksum(0, (void *) addr + kiov[0].kiov_offset + offset, nob);
361 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
362 (void *)addr + kiov[0].kiov_offset + offset, nob);
364 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
365 cksum, tmpck, addr, kiov[0].kiov_offset, nob, offset);
369 fraglen = min(kiov->kiov_len - offset, nob);
371 /* make dang sure we don't send a bogus checksum if somehow we get
372 * an odd length fragment on anything but the last entry in a kiov -
373 * we know from kgnilnd_setup_rdma_buffer that we can't have non
374 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
375 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
376 "odd fraglen %u on nkiov %d, nob %u kiov_len %u offset %u kiov 0x%p\n",
377 fraglen, nkiov, nob, kiov->kiov_len, offset, kiov);
379 addr = (void *)kmap(kiov->kiov_page) + kiov->kiov_offset + offset;
380 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
383 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
384 cksum, tmpck, kiov->kiov_page, kiov->kiov_offset, addr,
390 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
392 kunmap(kiov->kiov_page);
399 /* iov must not run out before end of data */
400 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
405 retsum = csum_fold(cksum);
407 /* don't use magic 'no checksum' value */
411 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
417 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
419 msg->gnm_magic = GNILND_MSG_MAGIC;
420 msg->gnm_version = GNILND_MSG_VERSION;
421 msg->gnm_type = type;
422 msg->gnm_payload_len = 0;
423 msg->gnm_srcnid = source;
424 /* gnm_connstamp gets set when FMA is sent */
425 /* gnm_srcnid is set on creation via function argument
426 * The right interface/net and nid is passed in when the message
432 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
434 kgn_tx_t *tx = kgnilnd_alloc_tx();
437 kgnilnd_init_msg(&tx->tx_msg, type, source);
439 CERROR("couldn't allocate new tx type %s!\n",
440 kgnilnd_msgtype2str(type));
447 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
453 case GNILND_MSG_GET_REQ:
454 case GNILND_MSG_GET_DONE:
455 nak_type = GNILND_MSG_GET_NAK;
457 case GNILND_MSG_PUT_REQ:
458 case GNILND_MSG_PUT_ACK:
459 case GNILND_MSG_PUT_DONE:
460 nak_type = GNILND_MSG_PUT_NAK;
462 case GNILND_MSG_PUT_REQ_REV:
463 case GNILND_MSG_PUT_DONE_REV:
464 nak_type = GNILND_MSG_PUT_NAK_REV;
466 case GNILND_MSG_GET_REQ_REV:
467 case GNILND_MSG_GET_ACK_REV:
468 case GNILND_MSG_GET_DONE_REV:
469 nak_type = GNILND_MSG_GET_NAK_REV;
472 CERROR("invalid msg type %s (%d)\n",
473 kgnilnd_msgtype2str(rx_type), rx_type);
476 /* only allow NAK on error and truncate to zero */
477 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
478 error, conn, cookie);
480 tx = kgnilnd_new_tx_msg(nak_type, source);
482 CNETERR("can't get TX to NAK RDMA to %s\n",
483 libcfs_nid2str(conn->gnc_peer->gnp_nid));
487 tx->tx_msg.gnm_u.completion.gncm_retval = error;
488 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
489 kgnilnd_queue_tx(conn, tx);
493 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
494 struct kvec *iov, lnet_kiov_t *kiov,
495 unsigned int offset, unsigned int nob)
497 kgn_msg_t *msg = &tx->tx_msg;
500 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
501 * gni_smsg_send to send that as the payload */
503 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
506 tx->tx_buffer = NULL;
507 } else if (kiov != NULL) {
509 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
510 niov = ((nob + offset + kiov->kiov_offset + PAGE_SIZE - 1) /
514 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
515 "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n",
516 niov, msg, kiov, iov, offset, nob);
518 while (offset >= kiov->kiov_len) {
519 offset -= kiov->kiov_len;
524 for (i = 0; i < niov; i++) {
525 /* We can't have a kiov_offset on anything but the first entry,
526 * otherwise we'll have a hole at the end of the mapping as we only map
528 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
529 * than kiov_len, we will also have a whole at the end of that page
530 * which isn't allowed */
531 if ((kiov[i].kiov_offset != 0 && i > 0) ||
532 (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_SIZE && i < niov - 1)) {
533 CNETERR("Can't make payload contiguous in I/O VM:"
534 "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
535 i, offset, nob, kiov->kiov_offset, kiov->kiov_len);
538 tx->tx_imm_pages[i] = kiov[i].kiov_page;
541 /* hijack tx_phys for the later unmap */
543 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
545 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) + kiov[0].kiov_offset + offset;
546 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
547 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
548 nob, kiov, tx->tx_buffer);
550 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
551 if (tx->tx_phys == NULL) {
552 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
556 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
557 /* make sure we take into account the kiov offset as the start of the buffer */
558 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].kiov_offset + offset;
559 GNIDBG_TX(D_NET, tx, "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
560 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
562 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
566 /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
567 * could "flatten" the payload into a single contiguous buffer ready
568 * for sending direct over an FMA if we ever needed to. */
572 while (offset >= iov->iov_len) {
573 offset -= iov->iov_len;
579 if (nob > iov->iov_len - offset) {
580 CERROR("Can't handle multiple vaddr fragments\n");
584 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
586 tx->tx_buftype = GNILND_BUF_IMMEDIATE;
590 /* checksum payload early - it shouldn't be changing after lnd_send */
591 if (*kgnilnd_tunables.kgn_checksum >= 2) {
592 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
593 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
594 msg->gnm_payload_cksum += 0xe00e;
596 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
597 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
601 msg->gnm_payload_cksum = 0;
608 kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
609 unsigned int niov, struct kvec *iov,
610 unsigned int offset, unsigned int nob)
615 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
617 while (offset >= iov->iov_len) {
618 offset -= iov->iov_len;
624 if (nob > iov->iov_len - offset) {
625 CERROR("Can't handle multiple vaddr fragments\n");
629 tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
631 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
636 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
637 unsigned int offset, unsigned int nob)
639 gni_mem_segment_t *phys;
641 unsigned int fraglen;
643 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
647 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
649 /* only allocate this if we are going to use it */
650 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
652 if (tx->tx_phys == NULL) {
653 CERROR("failed to allocate tx_phys\n");
658 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
659 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
661 /* if loops changes, please change kgnilnd_cksum_kiov
662 * and kgnilnd_setup_immediate_buffer */
664 while (offset >= kiov->kiov_len) {
665 offset -= kiov->kiov_len;
671 /* at this point, kiov points to the first page that we'll actually map
672 * now that we've seeked into the koiv for offset and dropped any
673 * leading pages that fall entirely within the offset */
674 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
677 /* kiov_offset is start of 'valid' buffer, so index offset past that */
678 tx->tx_buffer = (void *)((unsigned long)(kiov->kiov_offset + offset));
681 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
682 tx, tx->tx_buffer, kiov, kiov->kiov_offset, nkiov, offset);
685 fraglen = min(kiov->kiov_len - offset, nob);
687 /* We can't have a kiov_offset on anything but the first entry,
688 * otherwise we'll have a hole at the end of the mapping as we only map
689 * whole pages. Only the first page is allowed to have an offset -
690 * we'll add that into tx->tx_buffer and that will get used when we
691 * map in the segments (see kgnilnd_map_buffer).
692 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
693 * than kiov_len, we will also have a whole at the end of that page
694 * which isn't allowed */
695 if ((phys != tx->tx_phys) &&
696 ((kiov->kiov_offset != 0) ||
697 ((kiov->kiov_len < PAGE_SIZE) && (nob > kiov->kiov_len)))) {
698 CERROR("Can't make payload contiguous in I/O VM:"
699 "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
700 (int)(phys - tx->tx_phys),
701 offset, nob, kiov->kiov_offset, kiov->kiov_len);
706 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
707 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
712 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
717 CDEBUG(D_BUFFS, "page 0x%p kiov_offset %u kiov_len %u nob %u "
718 "nkiov %u offset %u\n",
719 kiov->kiov_page, kiov->kiov_offset, kiov->kiov_len, nob, nkiov, offset);
721 phys->address = page_to_phys(kiov->kiov_page);
728 /* iov must not run out before end of data */
729 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
733 tx->tx_phys_npages = phys - tx->tx_phys;
738 if (tx->tx_phys != NULL) {
739 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
740 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
741 sizeof(*tx->tx_phys), tx->tx_phys);
748 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
749 struct kvec *iov, lnet_kiov_t *kiov,
750 unsigned int offset, unsigned int nob)
754 LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
755 " offset %d, nob %d, niov %d\n"
756 , iov, kiov, tx, offset, nob, niov);
759 rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
761 rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
766 /* kgnilnd_parse_lnet_rdma()
767 * lntmsg - message passed in from lnet.
768 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
769 * nob - actual number of bytes to in this message.
770 * put_len - It is possible for PUTs to have a different length than the
771 * length stored in lntmsg->msg_len since LNET can adjust this
772 * length based on it's buffer size and offset.
773 * lnet_try_match_md() sets the mlength that we use to do the RDMA
777 kgnilnd_parse_lnet_rdma(lnet_msg_t *lntmsg, unsigned int *niov,
778 unsigned int *offset, unsigned int *nob,
779 lnet_kiov_t **kiov, int put_len)
781 /* GETs are weird, see kgnilnd_send */
782 if (lntmsg->msg_type == LNET_MSG_GET) {
783 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
786 *kiov = lntmsg->msg_md->md_iov.kiov;
788 *niov = lntmsg->msg_md->md_niov;
789 *nob = lntmsg->msg_md->md_length;
792 *kiov = lntmsg->msg_kiov;
793 *niov = lntmsg->msg_niov;
795 *offset = lntmsg->msg_offset;
800 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
802 unsigned int niov, offset, nob;
804 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
805 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
807 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
808 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
809 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
810 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
811 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
812 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
813 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
815 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
816 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
817 tx->tx_msg.gnm_payload_cksum = 0;
820 if (*kgnilnd_tunables.kgn_checksum < 3) {
821 tx->tx_msg.gnm_payload_cksum = 0;
825 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
827 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
831 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
833 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
835 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
839 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
840 tx->tx_msg.gnm_payload_cksum += 0xd00d;
844 /* kgnilnd_verify_rdma_cksum()
845 * tx - PUT_DONE/GET_DONE matched tx.
846 * rx_cksum - received checksum to compare against.
847 * put_len - see kgnilnd_parse_lnet_rdma comments.
850 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
854 unsigned int niov, offset, nob;
856 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
857 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
859 /* we can only match certain requests */
860 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
861 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
862 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
863 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
864 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
865 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
866 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
868 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
869 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
874 if (*kgnilnd_tunables.kgn_checksum >= 3) {
875 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
876 "no RDMA payload checksum when enabled");
881 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
883 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
886 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
888 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
891 if (cksum != rx_cksum) {
892 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
893 "Bad RDMA payload checksum (%x expected %x); "
894 "kiov 0x%p niov %d nob %u offset %u",
895 cksum, rx_cksum, kiov, niov, nob, offset);
896 switch (dump_on_err) {
899 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
901 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
904 /* fall through to dump log */
906 libcfs_debug_dumplog();
912 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
918 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
922 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
923 "already mapped!", NULL);
925 spin_lock(&dev->gnd_map_lock);
926 switch (tx->tx_buftype) {
928 GNIDBG_TX(D_EMERG, tx,
929 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
930 spin_unlock(&dev->gnd_map_lock);
934 case GNILND_BUF_PHYS_MAPPED:
935 bytes = tx->tx_phys_npages * PAGE_SIZE;
936 dev->gnd_map_nphys++;
937 dev->gnd_map_physnop += tx->tx_phys_npages;
940 case GNILND_BUF_VIRT_MAPPED:
942 dev->gnd_map_nvirt++;
943 dev->gnd_map_virtnob += tx->tx_nob;
947 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
948 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
949 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
950 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
951 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
954 atomic_inc(&dev->gnd_n_mdd);
955 atomic64_add(bytes, &dev->gnd_nbytes_map);
957 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
960 /* we only get here in the valid cases */
961 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
962 dev->gnd_map_version++;
963 spin_unlock(&dev->gnd_map_lock);
967 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
971 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
972 "not mapped!", NULL);
973 spin_lock(&dev->gnd_map_lock);
975 switch (tx->tx_buftype) {
977 GNIDBG_TX(D_EMERG, tx,
978 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
979 spin_unlock(&dev->gnd_map_lock);
983 case GNILND_BUF_PHYS_UNMAPPED:
984 bytes = tx->tx_phys_npages * PAGE_SIZE;
985 dev->gnd_map_nphys--;
986 dev->gnd_map_physnop -= tx->tx_phys_npages;
989 case GNILND_BUF_VIRT_UNMAPPED:
991 dev->gnd_map_nvirt--;
992 dev->gnd_map_virtnob -= tx->tx_nob;
996 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
997 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
998 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
999 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
1000 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
1001 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
1002 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1005 atomic_dec(&dev->gnd_n_mdd);
1006 atomic64_sub(bytes, &dev->gnd_nbytes_map);
1008 /* we only get here in the valid cases */
1009 list_del_init(&tx->tx_map_list);
1010 dev->gnd_map_version++;
1011 spin_unlock(&dev->gnd_map_lock);
1015 kgnilnd_map_buffer(kgn_tx_t *tx)
1017 kgn_conn_t *conn = tx->tx_conn;
1018 kgn_device_t *dev = conn->gnc_device;
1019 __u32 flags = GNI_MEM_READWRITE;
1022 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
1023 * be called concurrently as there are internal locks that protect
1024 * any data structures or HW resources. We just need to ensure
1025 * that our concurrency doesn't result in the kgn_device_t
1026 * getting nuked while we are in here */
1028 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
1029 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
1031 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1034 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1035 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1038 switch (tx->tx_buftype) {
1042 case GNILND_BUF_NONE:
1043 case GNILND_BUF_IMMEDIATE:
1044 case GNILND_BUF_IMMEDIATE_KIOV:
1045 case GNILND_BUF_PHYS_MAPPED:
1046 case GNILND_BUF_VIRT_MAPPED:
1049 case GNILND_BUF_PHYS_UNMAPPED:
1050 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1051 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1052 tx->tx_phys, tx->tx_phys_npages, NULL,
1053 GNI_MEM_PHYS_SEGMENTS | flags,
1055 /* could race with other uses of the map counts, but this is ok
1056 * - this needs to turn into a non-fatal error soon to allow
1057 * GART resource, etc starvation handling */
1058 if (rrc != GNI_RC_SUCCESS) {
1059 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1060 "phys %u pp %u, virt %u nob %llu",
1061 tx->tx_phys_npages, dev->gnd_id,
1062 dev->gnd_map_nphys, dev->gnd_map_physnop,
1063 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1064 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1067 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1068 kgnilnd_mem_add_map_list(dev, tx);
1071 case GNILND_BUF_VIRT_UNMAPPED:
1072 rrc = kgnilnd_mem_register(dev->gnd_handle,
1073 (__u64)tx->tx_buffer, tx->tx_nob,
1074 NULL, flags, &tx->tx_map_key);
1075 if (rrc != GNI_RC_SUCCESS) {
1076 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1077 "phys %u pp %u, virt %u nob %llu",
1078 tx->tx_nob, dev->gnd_id,
1079 dev->gnd_map_nphys, dev->gnd_map_physnop,
1080 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1081 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1084 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1085 kgnilnd_mem_add_map_list(dev, tx);
1086 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1087 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1088 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1089 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1090 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1098 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1100 kgn_conn_t *conn = tx->tx_conn;
1101 kgn_mdd_purgatory_t *gmp;
1103 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1104 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1105 " asserting to avoid data corruption\n");
1106 if (tx->tx_buffer_copy)
1107 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1109 gmp->gmp_map_key = tx->tx_map_key;
1111 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1113 /* ensure that we don't have a blank purgatory - indicating the
1114 * conn is not already on purgatory lists - we'd never recover these
1115 * MDD if that were the case */
1116 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1117 "conn 0x%p->%s with NULL purgatory",
1118 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1120 /* link 'er up! - only place we really need to lock for
1121 * concurrent access */
1122 spin_lock(&conn->gnc_list_lock);
1123 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1124 spin_unlock(&conn->gnc_list_lock);
1128 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1132 int hold_timeout = 0;
1134 /* code below relies on +1 relationship ... */
1135 CLASSERT(GNILND_BUF_PHYS_MAPPED == (GNILND_BUF_PHYS_UNMAPPED + 1));
1136 CLASSERT(GNILND_BUF_VIRT_MAPPED == (GNILND_BUF_VIRT_UNMAPPED + 1));
1138 switch (tx->tx_buftype) {
1142 case GNILND_BUF_NONE:
1143 case GNILND_BUF_IMMEDIATE:
1144 case GNILND_BUF_PHYS_UNMAPPED:
1145 case GNILND_BUF_VIRT_UNMAPPED:
1147 case GNILND_BUF_IMMEDIATE_KIOV:
1148 if (tx->tx_phys != NULL) {
1149 vunmap(tx->tx_phys);
1150 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1151 kunmap(tx->tx_imm_pages[0]);
1153 /* clear to prevent kgnilnd_free_tx from thinking
1154 * this is a RDMA descriptor */
1158 case GNILND_BUF_PHYS_MAPPED:
1159 case GNILND_BUF_VIRT_MAPPED:
1160 LASSERT(tx->tx_conn != NULL);
1162 dev = tx->tx_conn->gnc_device;
1164 /* only want to hold if we are closing conn without
1165 * verified peer notification - the theory is that
1166 * a TX error can be communicated in all other cases */
1167 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1168 error != -GNILND_NOPURG &&
1169 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1170 kgnilnd_add_purgatory_tx(tx);
1172 /* The timeout we give to kgni is a deadman stop only.
1173 * we are setting high to ensure we don't have the kgni timer
1174 * fire before ours fires _and_ is handled */
1175 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1177 GNIDBG_TX(D_NET, tx,
1178 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1179 tx->tx_conn->gnc_device, hold_timeout,
1180 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1182 if (tx->tx_buffer_copy != NULL) {
1183 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1184 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1185 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1186 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1188 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1189 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1193 kgnilnd_mem_del_map_list(dev, tx);
1199 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1201 lnet_msg_t *lntmsg0, *lntmsg1;
1202 int status0, status1;
1203 lnet_ni_t *ni = NULL;
1204 kgn_conn_t *conn = tx->tx_conn;
1206 LASSERT(!in_interrupt());
1208 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1209 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1212 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1213 !kgnilnd_conn_clean_errno(completion)) {
1214 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1215 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1216 completion, tx, conn ?
1217 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1218 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1219 kgnilnd_tx_state2str(tx->tx_list_state),
1220 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1223 /* The error codes determine if we hold onto the MDD */
1224 kgnilnd_unmap_buffer(tx, completion);
1226 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1227 * we play nice with the error codes to avoid delivering a failed
1228 * REQUEST and then a REPLY event as well */
1230 /* return -EIO to lnet - it is the magic value for failed sends */
1231 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1233 status1 = completion;
1235 status0 = status1 = completion;
1238 tx->tx_buftype = GNILND_BUF_NONE;
1239 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1241 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1242 * set NULL when we are a tx without a conn */
1244 ni = conn->gnc_peer->gnp_net->gnn_ni;
1246 spin_lock(&conn->gnc_tx_lock);
1248 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1249 (volatile unsigned long *)&conn->gnc_tx_bits),
1250 "conn %p tx %p bit %d already cleared\n",
1251 conn, tx, tx->tx_id.txe_idx);
1253 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1254 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1256 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1257 spin_unlock(&conn->gnc_tx_lock);
1260 kgnilnd_free_tx(tx);
1262 /* finalize AFTER freeing lnet msgs */
1264 /* warning - we should hold no locks here - calling lnet_finalize
1265 * could free up lnet credits, resulting in a call chain back into
1266 * the LND via kgnilnd_send and friends */
1268 lnet_finalize(ni, lntmsg0, status0);
1270 if (lntmsg1 != NULL) {
1271 lnet_finalize(ni, lntmsg1, status1);
1276 kgnilnd_txlist_done(struct list_head *txlist, int error)
1279 int err_printed = 0;
1281 if (list_empty(txlist))
1284 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1285 /* only print the first error */
1287 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1288 list_del_init(&tx->tx_list);
1289 kgnilnd_tx_done(tx, error);
1294 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1298 spin_lock(&conn->gnc_tx_lock);
1300 /* ID zero is NOT ALLOWED!!! */
1303 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1304 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1305 if (id == GNILND_MAX_MSG_ID) {
1306 if (conn->gnc_next_tx != 1) {
1307 /* we only searched from next_tx to end and didn't find
1308 * one, so search again from start */
1309 conn->gnc_next_tx = 1;
1312 /* couldn't find one! */
1313 spin_unlock(&conn->gnc_tx_lock);
1317 /* bump next_tx to prevent immediate reuse */
1318 conn->gnc_next_tx = id + 1;
1320 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1321 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1322 "tx 0x%p already at id %d\n",
1323 conn->gnc_tx_ref_table[id], id);
1325 /* delay these until we have a valid ID - prevents bad clear of the bit
1326 * in kgnilnd_tx_done */
1328 tx->tx_id.txe_cqid = conn->gnc_cqid;
1330 tx->tx_id.txe_idx = id;
1331 conn->gnc_tx_ref_table[id] = tx;
1333 /* Using jiffies to help differentiate against TX reuse - with
1334 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1335 * if we are sending to the same node faster than 256000/sec.
1336 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1338 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1340 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1342 spin_unlock(&conn->gnc_tx_lock);
1347 kgnilnd_tx_should_retry(kgn_conn_t *conn, kgn_tx_t *tx)
1349 int max_retrans = *kgnilnd_tunables.kgn_max_retransmits;
1351 int log_retrans_level;
1353 /* I need kgni credits to send this. Replace tx at the head of the
1354 * fmaq and I'll get rescheduled when credits appear */
1357 conn->gnc_tx_retrans++;
1358 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0) ||
1359 (tx->tx_retrans > (max_retrans / 2)));
1360 log_retrans_level = tx->tx_retrans < (max_retrans / 2) ? D_NET : D_NETERROR;
1362 /* Decision time - either error, warn or just retransmit */
1364 /* we don't care about TX timeout - it could be that the network is slower
1365 * or throttled. We'll keep retranmitting - so if the network is so slow
1366 * that we fill up our mailbox, we'll keep trying to resend that msg
1367 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1368 * that he hasn't send us any traffic in return */
1370 if (tx->tx_retrans > max_retrans) {
1371 /* this means we are not backing off the retransmits
1372 * in a healthy manner and are likely chewing up the
1373 * CPU cycles quite badly */
1374 GNIDBG_TOMSG(D_ERROR, &tx->tx_msg,
1375 "SOFTWARE BUG: too many retransmits (%d) for tx id %x "
1377 tx->tx_retrans, tx->tx_id, conn,
1378 libcfs_nid2str(conn->gnc_peer->gnp_nid));
1380 /* yes - double errors to help debug this condition */
1381 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg, "connection dead. "
1382 "unable to send to %s for %lu secs (%d tries)",
1383 libcfs_nid2str(tx->tx_conn->gnc_peer->gnp_nid),
1384 cfs_duration_sec(jiffies - tx->tx_cred_wait),
1387 kgnilnd_close_conn(conn, -ETIMEDOUT);
1389 /* caller should terminate */
1392 /* some reasonable throttling of the debug message */
1394 unsigned long now = jiffies;
1395 /* XXX Nic: Mystical TX debug here... */
1396 GNIDBG_SMSG_CREDS(log_retrans_level, conn);
1397 GNIDBG_TOMSG(log_retrans_level, &tx->tx_msg,
1398 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1399 " last_msg %uus/%uus last_cq %uus/%uus",
1400 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1401 tx->tx_id, tx->tx_retrans,
1402 jiffies_to_usecs(now - tx->tx_cred_wait),
1403 jiffies_to_usecs(now - conn->gnc_last_tx),
1404 jiffies_to_usecs(now - conn->gnc_last_rx),
1405 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1406 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1408 /* caller should retry */
1413 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1414 * to avoid bad ordering with state_lock */
1417 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1418 spinlock_t *state_lock, kgn_tx_list_state_t state)
1420 kgn_conn_t *conn = tx->tx_conn;
1421 kgn_msg_t *msg = &tx->tx_msg;
1424 unsigned long newest_last_rx, timeout;
1427 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1428 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1430 "msg 0x%p type %d wrong payload size %d\n",
1431 msg, msg->gnm_type, immediatenob);
1433 /* make sure we catch all the cases where we'd send on a dirty old mbox
1434 * but allow case for sending CLOSE. Since this check is within the CQ
1435 * mutex barrier and the close message is only sent through
1436 * kgnilnd_send_conn_close the last message out the door will be the
1439 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1440 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1441 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1442 /* Return -ETIME, we are closing the connection already so we dont want to
1443 * have this tx hit the wire. The tx will be killed by the calling function.
1444 * Once the EP is marked dirty the close message will be the last
1445 * thing to hit the wire */
1450 timeout = cfs_time_seconds(conn->gnc_timeout);
1452 newest_last_rx = GNILND_LASTRX(conn);
1454 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1455 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1458 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1459 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1460 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1461 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1462 cfs_duration_sec(now - newest_last_rx),
1463 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1464 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1465 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1469 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1470 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1471 * creating the message this allows the message to contain the correct LNET NID/NET needed
1472 * instead of the one that the peer/conn uses for sending the data.
1474 msg->gnm_connstamp = conn->gnc_my_connstamp;
1475 msg->gnm_payload_len = immediatenob;
1476 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1478 /* always init here - kgn_checksum is a /sys module tunable
1479 * and can be flipped at any point, even between msg init and sending */
1481 if (*kgnilnd_tunables.kgn_checksum) {
1482 /* We must set here and not in kgnilnd_init_msg,
1483 * we could resend this msg many times
1484 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1485 * through init_msg again */
1486 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1487 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1488 msg->gnm_cksum += 0xf00f;
1492 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1493 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1494 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1495 tx->tx_id.txe_idx, immediate, immediatenob);
1497 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1498 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1500 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1501 msg, sizeof(*msg), immediate,
1503 tx->tx_id.txe_smsg_id);
1507 case GNI_RC_SUCCESS:
1508 atomic_inc(&conn->gnc_tx_seq);
1509 conn->gnc_last_tx = jiffies;
1510 /* no locking here as LIVE isn't a list */
1511 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1513 /* this needs to be checked under lock as it might be freed from a completion
1516 if (msg->gnm_type == GNILND_MSG_NOOP) {
1517 set_mb(conn->gnc_last_noop_sent, jiffies);
1520 /* serialize with seeing CQ events for completion on this, as well as
1522 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1523 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1525 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1526 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1527 kgnilnd_peer_alive(conn->gnc_peer);
1528 GNIDBG_SMSG_CREDS(D_NET, conn);
1531 case GNI_RC_NOT_DONE:
1532 /* XXX Nic: We need to figure out how to track this
1533 * - there are bound to be good reasons for it,
1534 * but we want to know when it happens */
1535 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1536 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1537 /* We'll handle this error inline - makes the calling logic much more
1540 /* If no lock, caller doesn't want us to retry */
1541 if (state_lock == NULL) {
1545 retry_send = kgnilnd_tx_should_retry(conn, tx);
1547 /* add to head of list for the state and retries */
1548 spin_lock(state_lock);
1549 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1550 spin_unlock(state_lock);
1552 /* We only reschedule for a certain number of retries, then
1553 * we will wait for the CQ events indicating a release of SMSG
1555 if (tx->tx_retrans < (*kgnilnd_tunables.kgn_max_retransmits/4)) {
1556 kgnilnd_schedule_conn(conn);
1559 /* CQ event coming in signifies either TX completed or
1560 * RX receive. Either of these *could* free up credits
1561 * in the SMSG mbox and we should try sending again */
1562 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1563 tx->tx_conn->gnc_cqid);
1564 /* use +ve return code to let upper layers know they
1565 * should stop looping on sends */
1572 /* handle bad retcode gracefully */
1573 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1574 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1579 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1581 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1582 spinlock_t *state_lock, kgn_tx_list_state_t state)
1584 kgn_device_t *dev = tx->tx_conn->gnc_device;
1585 unsigned long timestamp;
1588 timestamp = jiffies;
1589 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1590 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1591 /* delay in jiffies - we are really concerned only with things that
1592 * result in a schedule() or really holding this off for long times .
1593 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1594 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1596 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1602 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1604 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1605 spinlock_t *state_lock, kgn_tx_list_state_t state)
1607 kgn_conn_t *conn = tx->tx_conn;
1608 kgn_device_t *dev = conn->gnc_device;
1609 unsigned long timestamp;
1612 timestamp = jiffies;
1614 /* technically we are doing bad things with the read_lock on the peer_conn
1615 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1616 * for the mutex. I bet lockdep is gonna flag this one though... */
1618 /* there are a few cases where we don't want the immediate send - like
1619 * when we are in the scheduler thread and it'd harm the latency of
1620 * getting messages up to LNet */
1622 /* rmb for gnd_ready */
1624 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1626 atomic_inc(&conn->gnc_device->gnd_fast_block);
1627 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1628 /* dont hit HW during quiesce */
1630 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1631 /* dont hit HW if stale EPs and conns left to close */
1634 atomic_inc(&conn->gnc_device->gnd_fast_try);
1635 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1636 &conn->gnc_smsg_mutex);
1641 /* we got the mutex and weren't blocked */
1643 /* delay in jiffies - we are really concerned only with things that
1644 * result in a schedule() or really holding this off for long times .
1645 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1646 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1648 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1649 tx->tx_qtime = jiffies;
1650 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1651 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1652 /* _nolock unlocks the mutex for us */
1658 /* lets us know if we can push this RDMA through now */
1660 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1664 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1666 if (bytes_left < 0) {
1667 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1668 atomic_inc(&dev->gnd_rdmaq_nstalls);
1671 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1672 dev->gnd_rdmaq_deadline);
1673 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1674 /* we never del this timer - at worst it schedules us.. */
1681 /* this adds a TX to the queue pending throttling authorization before
1682 * we allow our remote peer to launch a PUT at us */
1684 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1688 /* we cannot go into send_mapped_tx from here as we are holding locks
1689 * and mem registration might end up allocating memory in kgni.
1690 * That said, we'll push this as far as we can into the queue process */
1691 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1694 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1695 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1696 /* lets us know how delayed RDMA is */
1697 tx->tx_qtime = jiffies;
1698 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1700 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1701 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1702 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1703 /* lets us know how delayed mapping is */
1704 tx->tx_qtime = jiffies;
1705 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1708 /* make sure we wake up sched to run this */
1709 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1712 /* push TX through state machine */
1714 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1719 /* set the tx_id here, we delay it until we have an actual conn
1721 * in some cases, the tx_id is already set to provide for things
1722 * like RDMA completion cookies, etc */
1723 if (tx->tx_id.txe_idx == 0) {
1724 rc = kgnilnd_set_tx_id(tx, conn);
1726 kgnilnd_tx_done(tx, rc);
1731 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1732 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1734 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1736 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1737 kgnilnd_tx_done(tx, rc);
1741 switch (tx->tx_msg.gnm_type) {
1742 case GNILND_MSG_PUT_ACK:
1743 case GNILND_MSG_GET_REQ:
1744 case GNILND_MSG_PUT_REQ_REV:
1745 case GNILND_MSG_GET_ACK_REV:
1746 /* hijacking time! If this messages will authorize our peer to
1747 * send his dirty little bytes in an RDMA, we need to get permission */
1748 kgnilnd_queue_rdma(conn, tx);
1750 case GNILND_MSG_IMMEDIATE:
1751 /* try to send right now, can help reduce latency */
1752 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1755 /* it was sent, break out of switch to avoid default case of queueing */
1758 /* needs to queue to try again, so fall through to default case */
1759 case GNILND_MSG_NOOP:
1760 /* Just make sure this goes out first for this conn */
1762 /* fall through... */
1764 spin_lock(&conn->gnc_list_lock);
1765 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1766 tx->tx_qtime = jiffies;
1767 spin_unlock(&conn->gnc_list_lock);
1768 kgnilnd_schedule_conn(conn);
1773 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
1776 kgn_peer_t *new_peer = NULL;
1777 kgn_conn_t *conn = NULL;
1783 /* If I get here, I've committed to send, so I complete the tx with
1784 * failure on any problems */
1786 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1787 "tx already has connection %p", tx->tx_conn);
1789 /* do all of the peer & conn searching in one swoop - this avoids
1790 * nastiness when dropping locks and needing to maintain a sane state
1791 * in the face of stack reset or something else nuking peers & conns */
1793 /* I expect to find him, so only take a read lock */
1794 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1796 peer = kgnilnd_find_peer_locked(target->nid);
1798 conn = kgnilnd_find_conn_locked(peer);
1799 /* this could be NULL during quiesce */
1801 /* Connection exists; queue message on it */
1802 kgnilnd_queue_tx(conn, tx);
1803 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1807 /* don't create a connection if the peer is marked down */
1808 if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
1809 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1815 /* creating peer or conn; I'll need a write lock... */
1816 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1818 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1820 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1822 /* NB - this will not block during normal operations -
1823 * the only writer of this is in the startup/shutdown path. */
1824 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1830 /* ignore previous peer entirely - we cycled the lock, so we
1831 * will create new peer and at worst drop it if peer is still
1833 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1835 up_read(&kgnilnd_data.kgn_net_rw_sem);
1839 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1840 up_read(&kgnilnd_data.kgn_net_rw_sem);
1842 /* search for peer again now that we have the lock
1843 * if we don't find it, add our new one to the list */
1844 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1846 /* don't create a connection if the peer is not up */
1847 if (peer->gnp_down != GNILND_RCA_NODE_UP) {
1848 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1853 conn = kgnilnd_find_or_create_conn_locked(peer);
1855 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1856 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1861 /* oh hey, found a conn now... magical */
1862 kgnilnd_queue_tx(conn, tx);
1864 /* no conn, must be trying to connect - so we queue for now */
1865 tx->tx_qtime = jiffies;
1866 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1868 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1871 kgnilnd_tx_done(tx, rc);
1876 kgnilnd_rdma(kgn_tx_t *tx, int type,
1877 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1879 kgn_conn_t *conn = tx->tx_conn;
1880 unsigned long timestamp;
1881 gni_post_type_t post_type;
1884 unsigned int desc_nob = nob;
1885 void *desc_buffer = tx->tx_buffer;
1886 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1887 LASSERTF(kgnilnd_tx_mapped(tx),
1888 "unmapped tx %p\n", tx);
1889 LASSERTF(conn != NULL,
1890 "NULL conn on tx %p, naughty, naughty\n", tx);
1891 LASSERTF(nob <= sink->gnrd_nob,
1892 "nob %u > sink->gnrd_nob %d (%p)\n",
1893 nob, sink->gnrd_nob, sink);
1894 LASSERTF(nob <= tx->tx_nob,
1895 "nob %d > tx(%p)->tx_nob %d\n",
1896 nob, tx, tx->tx_nob);
1899 case GNILND_MSG_GET_DONE:
1900 case GNILND_MSG_PUT_DONE:
1901 post_type = GNI_POST_RDMA_PUT;
1903 case GNILND_MSG_GET_DONE_REV:
1904 case GNILND_MSG_PUT_DONE_REV:
1905 post_type = GNI_POST_RDMA_GET;
1908 CERROR("invalid msg type %s (%d)\n",
1909 kgnilnd_msgtype2str(type), type);
1912 if (post_type == GNI_POST_RDMA_GET) {
1913 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1914 * aligned. If the local buffer is not aligned correctly using the copy buffer
1915 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1916 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1917 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1918 * making sure the length and addr are aligned and copying the data into the correct location
1919 * once the transfer has completed.
1921 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1922 (sink->gnrd_addr & 3) ||
1925 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1927 kgnilnd_admin_addref(kgnilnd_data.kgn_rev_offset);
1929 if ((nob + tx->tx_offset) & 3) {
1930 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1931 kgnilnd_admin_addref(kgnilnd_data.kgn_rev_length);
1933 desc_nob = (nob + tx->tx_offset);
1936 if (tx->tx_buffer_copy == NULL) {
1937 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1938 * and require at most we allocate a few extra bytes. */
1939 tx->tx_buffer_copy = vmalloc(desc_nob);
1941 if (!tx->tx_buffer_copy) {
1942 /* allocation of buffer failed nak the rdma */
1943 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1944 kgnilnd_tx_done(tx, -EFAULT);
1947 kgnilnd_admin_addref(kgnilnd_data.kgn_rev_copy_buff);
1948 rc = kgnilnd_mem_register(conn->gnc_device->gnd_handle, (__u64)tx->tx_buffer_copy, desc_nob, NULL, GNI_MEM_READWRITE, &tx->tx_buffer_copy_map_key);
1949 if (rc != GNI_RC_SUCCESS) {
1950 /* Registration Failed nak rdma and kill the tx. */
1951 vfree(tx->tx_buffer_copy);
1952 tx->tx_buffer_copy = NULL;
1953 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1954 kgnilnd_tx_done(tx, -EFAULT);
1958 desc_map_key = tx->tx_buffer_copy_map_key;
1959 desc_buffer = tx->tx_buffer_copy;
1963 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1964 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1965 tx->tx_rdma_desc.type = post_type;
1966 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1967 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1968 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1969 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1970 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1971 tx->tx_rdma_desc.length = desc_nob;
1972 tx->tx_nob_rdma = nob;
1973 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1974 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1975 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1976 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1977 /* prep final completion message */
1978 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1979 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1980 /* send actual size RDMA'd in retval */
1981 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1983 kgnilnd_compute_rdma_cksum(tx, nob);
1986 kgnilnd_queue_tx(conn, tx);
1990 /* Don't lie (CLOSE == RDMA idle) */
1991 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1992 tx, conn, conn->gnc_close_sent);
1994 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1995 "0x%x cookie:%#llx",
1996 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1998 /* set CQ dedicated for RDMA */
1999 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
2001 timestamp = jiffies;
2002 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
2003 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2004 /* delay in jiffies - we are really concerned only with things that
2005 * result in a schedule() or really holding this off for long times .
2006 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2007 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2009 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2011 if (rrc == GNI_RC_ERROR_RESOURCE) {
2012 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2013 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2014 kgnilnd_unmap_buffer(tx, 0);
2016 if (tx->tx_buffer_copy != NULL) {
2017 vfree(tx->tx_buffer_copy);
2018 tx->tx_buffer_copy = NULL;
2021 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2022 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2024 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2025 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2029 spin_lock(&conn->gnc_list_lock);
2030 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2031 tx->tx_qtime = jiffies;
2032 spin_unlock(&conn->gnc_list_lock);
2033 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2034 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2036 /* XXX Nic: is this a place we should handle more errors for
2037 * robustness sake */
2038 LASSERT(rrc == GNI_RC_SUCCESS);
2043 kgnilnd_alloc_rx(void)
2047 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2049 CERROR("failed to allocate rx\n");
2052 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2055 /* no memset to zero, we'll always fill all members */
2059 /* release is to just free connection resources
2060 * we use this for the eager path after copying */
2062 kgnilnd_release_msg(kgn_conn_t *conn)
2065 unsigned long timestamp;
2067 CDEBUG(D_NET, "consuming %p\n", conn);
2069 timestamp = jiffies;
2070 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2071 /* delay in jiffies - we are really concerned only with things that
2072 * result in a schedule() or really holding this off for long times .
2073 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2074 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2076 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2077 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2079 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2080 GNIDBG_SMSG_CREDS(D_NET, conn);
2086 kgnilnd_consume_rx(kgn_rx_t *rx)
2088 kgn_conn_t *conn = rx->grx_conn;
2089 kgn_msg_t *rxmsg = rx->grx_msg;
2091 /* if we are eager, free the cache alloc'd msg */
2092 if (unlikely(rx->grx_eager)) {
2093 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2094 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2096 /* release ref from eager_recv */
2097 kgnilnd_conn_decref(conn);
2099 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2100 kgnilnd_release_msg(conn);
2103 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2104 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2111 kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
2113 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
2114 int type = lntmsg->msg_type;
2115 lnet_process_id_t target = lntmsg->msg_target;
2116 int target_is_router = lntmsg->msg_target_is_router;
2117 int routing = lntmsg->msg_routing;
2118 unsigned int niov = lntmsg->msg_niov;
2119 struct kvec *iov = lntmsg->msg_iov;
2120 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2121 unsigned int offset = lntmsg->msg_offset;
2122 unsigned int nob = lntmsg->msg_len;
2123 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2124 kgn_net_t *net = ni->ni_data;
2128 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2130 /* NB 'private' is different depending on what we're sending.... */
2131 LASSERT(!in_interrupt());
2133 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2134 type, nob, niov, libcfs_id2str(target));
2136 LASSERTF(nob == 0 || niov > 0,
2137 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2138 LASSERTF(niov <= LNET_MAX_IOV,
2139 "lntmsg %p niov %d\n", lntmsg, niov);
2141 /* payload is either all vaddrs or all pages */
2142 LASSERTF(!(kiov != NULL && iov != NULL),
2143 "lntmsg %p kiov %p iov %p\n", lntmsg, kiov, iov);
2146 mpflag = cfs_memory_pressure_get_and_set();
2150 CERROR("lntmsg %p with unexpected type %d\n",
2155 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2163 if (routing || target_is_router)
2164 break; /* send IMMEDIATE */
2166 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2167 * check the eventual sink buffer here - if small enough, remote
2168 * end is perfectly capable of returning data in short message -
2169 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2170 * for IMMEDIATE messages which will have it send a real reply instead
2171 * of doing kgnilnd_recv to have the RDMA continued */
2172 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2175 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2176 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2178 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2184 /* slightly different options as we might actually have a GET with a
2185 * MD_KIOV set but a non-NULL md_iov.iov */
2186 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
2187 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2188 lntmsg->msg_md->md_iov.iov, NULL,
2189 0, lntmsg->msg_md->md_length);
2191 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2192 NULL, lntmsg->msg_md->md_iov.kiov,
2193 0, lntmsg->msg_md->md_length);
2195 CERROR("unable to setup buffer: %d\n", rc);
2196 kgnilnd_tx_done(tx, rc);
2201 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2202 if (tx->tx_lntmsg[1] == NULL) {
2203 CERROR("Can't create reply for GET to %s\n",
2204 libcfs_nid2str(target.nid));
2205 kgnilnd_tx_done(tx, rc);
2210 tx->tx_lntmsg[0] = lntmsg;
2211 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2212 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2214 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2216 /* rest of tx_msg is setup just before it is sent */
2217 kgnilnd_launch_tx(tx, net, &target);
2219 case LNET_MSG_REPLY:
2221 /* to save on MDDs, we'll handle short kiov by vmap'ing
2222 * and sending via SMSG */
2223 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2226 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2227 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2229 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2236 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2238 kgnilnd_tx_done(tx, rc);
2243 tx->tx_lntmsg[0] = lntmsg;
2244 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2245 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2247 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2249 /* rest of tx_msg is setup just before it is sent */
2250 kgnilnd_launch_tx(tx, net, &target);
2254 /* send IMMEDIATE */
2256 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2257 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2259 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2265 rc = kgnilnd_setup_immediate_buffer(tx, niov, iov, kiov, offset, nob);
2267 kgnilnd_tx_done(tx, rc);
2271 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2272 tx->tx_lntmsg[0] = lntmsg;
2273 kgnilnd_launch_tx(tx, net, &target);
2276 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2278 cfs_memory_pressure_restore(mpflag);
2283 kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen)
2285 kgn_conn_t *conn = rx->grx_conn;
2286 kgn_msg_t *rxmsg = rx->grx_msg;
2287 unsigned int niov = lntmsg->msg_niov;
2288 struct kvec *iov = lntmsg->msg_iov;
2289 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2290 unsigned int offset = lntmsg->msg_offset;
2291 unsigned int nob = lntmsg->msg_len;
2296 switch (rxmsg->gnm_type) {
2297 case GNILND_MSG_PUT_REQ_REV:
2298 done_type = GNILND_MSG_PUT_DONE_REV;
2301 case GNILND_MSG_GET_REQ:
2302 done_type = GNILND_MSG_GET_DONE;
2305 CERROR("invalid msg type %s (%d)\n",
2306 kgnilnd_msgtype2str(rxmsg->gnm_type),
2311 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2315 rc = kgnilnd_set_tx_id(tx, conn);
2319 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2323 tx->tx_lntmsg[0] = lntmsg;
2324 tx->tx_getinfo = rxmsg->gnm_u.get;
2326 /* we only queue from kgnilnd_recv - we might get called from other contexts
2327 * and we don't want to block the mutex in those cases */
2329 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2330 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2331 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2332 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2337 kgnilnd_tx_done(tx, rc);
2338 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2340 lnet_finalize(ni, lntmsg, rc);
2344 kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
2347 kgn_rx_t *rx = private;
2348 kgn_conn_t *conn = rx->grx_conn;
2349 kgn_msg_t *rxmsg = rx->grx_msg;
2350 kgn_msg_t *eagermsg = NULL;
2351 kgn_peer_t *peer = NULL;
2352 kgn_conn_t *found_conn = NULL;
2354 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2355 conn, rxmsg, lntmsg);
2357 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2358 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2359 rxmsg->gnm_payload_len);
2362 /* Grab a read lock so the connection doesnt disappear on us
2363 * while we look it up
2365 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2367 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2369 found_conn = kgnilnd_find_conn_locked(peer);
2372 /* Verify the connection found is the same one that the message
2373 * is supposed to be using, if it is not output an error message
2376 if (!peer || !found_conn
2377 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2378 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2379 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2380 peer, conn, found_conn);
2382 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2383 " from %s", rxmsg->gnm_connstamp,
2384 found_conn->gnc_peer_connstamp,
2385 libcfs_nid2str(peer->gnp_nid));
2390 /* add conn ref to ensure it doesn't go away until all eager
2391 * messages processed */
2392 kgnilnd_conn_addref(conn);
2394 /* Now that we have verified the connection is valid and added a
2395 * reference we can remove the read_lock on the peer_conn_lock */
2396 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2398 /* we have no credits or buffers for this message, so copy it
2399 * somewhere for a later kgnilnd_recv */
2400 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2401 *kgnilnd_tunables.kgn_eager_credits) {
2402 CERROR("Out of eager credits to %s\n",
2403 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2407 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2409 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2410 if (eagermsg == NULL) {
2411 kgnilnd_conn_decref(conn);
2412 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2413 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2417 /* copy msg and payload */
2418 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2419 rx->grx_msg = eagermsg;
2422 /* stash this for lnet_finalize on cancel-on-conn-close */
2423 rx->grx_lntmsg = lntmsg;
2425 /* keep the same rx_t, it just has a new grx_msg now */
2426 *new_private = private;
2428 /* release SMSG buffer */
2429 kgnilnd_release_msg(conn);
2435 kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
2436 int delayed, unsigned int niov,
2437 struct kvec *iov, lnet_kiov_t *kiov,
2438 unsigned int offset, unsigned int mlen, unsigned int rlen)
2440 kgn_rx_t *rx = private;
2441 kgn_conn_t *conn = rx->grx_conn;
2442 kgn_msg_t *rxmsg = rx->grx_msg;
2448 LASSERT(!in_interrupt());
2449 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2450 /* Either all pages or all vaddrs */
2451 LASSERTF(!(kiov != NULL && iov != NULL), "kiov %p iov %p\n",
2454 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2455 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2456 conn, rxmsg, lntmsg,
2457 niov, kiov, iov, offset, mlen, rlen);
2459 /* we need to lock here as recv can be called from any context */
2460 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2461 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2462 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2464 /* someone closed the conn after we copied this out, nuke it */
2465 kgnilnd_consume_rx(rx);
2466 lnet_finalize(ni, lntmsg, conn->gnc_error);
2469 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2471 switch (rxmsg->gnm_type) {
2473 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2474 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2475 conn, rx, rxmsg, lntmsg, niov, kiov, iov, offset, mlen, rlen);
2478 case GNILND_MSG_IMMEDIATE:
2479 if (mlen > rxmsg->gnm_payload_len) {
2480 GNIDBG_MSG(D_ERROR, rxmsg,
2481 "Immediate message from %s too big: %d > %d",
2482 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2483 rxmsg->gnm_payload_len);
2485 kgnilnd_consume_rx(rx);
2489 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2490 * right after the kgn_msg_t header - so just 'cute' way of saying
2491 * rxmsg + sizeof(kgn_msg_t) */
2493 /* check payload checksum if sent */
2495 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2496 !rxmsg->gnm_payload_cksum &&
2497 rxmsg->gnm_payload_len != 0)
2498 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2500 if (rxmsg->gnm_payload_cksum != 0) {
2501 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2502 * which is what is used to calculate the cksum on the TX side */
2503 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2505 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2506 GNIDBG_MSG(D_NETERROR, rxmsg,
2507 "Bad payload checksum (%x expected %x)",
2508 pload_cksum, rxmsg->gnm_payload_cksum);
2509 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2511 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2512 &rxmsg[1], rxmsg->gnm_payload_len);
2513 /* fall through to dump */
2515 libcfs_debug_dumplog();
2521 /* checksum problems are fatal, kill the conn */
2522 kgnilnd_consume_rx(rx);
2523 kgnilnd_close_conn(conn, rc);
2529 lnet_copy_flat2kiov(
2531 *kgnilnd_tunables.kgn_max_immediate,
2532 &rxmsg[1], 0, mlen);
2536 *kgnilnd_tunables.kgn_max_immediate,
2537 &rxmsg[1], 0, mlen);
2539 kgnilnd_consume_rx(rx);
2540 lnet_finalize(ni, lntmsg, 0);
2543 case GNILND_MSG_PUT_REQ:
2544 /* LNET wants to truncate or drop transaction, sending NAK */
2546 kgnilnd_consume_rx(rx);
2547 lnet_finalize(ni, lntmsg, 0);
2549 /* only error if lntmsg == NULL, otherwise we are just
2550 * short circuiting the rdma process of 0 bytes */
2551 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2552 lntmsg == NULL ? -ENOENT : 0,
2553 rxmsg->gnm_u.get.gngm_cookie,
2557 /* sending ACK with sink buff. info */
2558 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2560 kgnilnd_consume_rx(rx);
2564 rc = kgnilnd_set_tx_id(tx, conn);
2566 GOTO(nak_put_req, rc);
2569 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2571 GOTO(nak_put_req, rc);
2574 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2575 rxmsg->gnm_u.putreq.gnprm_cookie;
2576 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2577 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2578 (__u64)((unsigned long)tx->tx_buffer);
2579 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2581 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2582 tx->tx_qtime = jiffies;
2583 /* we only queue from kgnilnd_recv - we might get called from other contexts
2584 * and we don't want to block the mutex in those cases */
2586 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2587 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2588 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2589 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2591 kgnilnd_consume_rx(rx);
2595 /* make sure we send an error back when the PUT fails */
2596 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2597 kgnilnd_tx_done(tx, rc);
2598 kgnilnd_consume_rx(rx);
2600 /* return magic LNet network error */
2602 case GNILND_MSG_GET_REQ_REV:
2603 /* LNET wants to truncate or drop transaction, sending NAK */
2605 kgnilnd_consume_rx(rx);
2606 lnet_finalize(ni, lntmsg, 0);
2608 /* only error if lntmsg == NULL, otherwise we are just
2609 * short circuiting the rdma process of 0 bytes */
2610 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2611 lntmsg == NULL ? -ENOENT : 0,
2612 rxmsg->gnm_u.get.gngm_cookie,
2616 /* lntmsg can be null when parsing a LNET_GET */
2617 if (lntmsg != NULL) {
2618 /* sending ACK with sink buff. info */
2619 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2621 kgnilnd_consume_rx(rx);
2625 rc = kgnilnd_set_tx_id(tx, conn);
2627 GOTO(nak_get_req_rev, rc);
2630 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2632 GOTO(nak_get_req_rev, rc);
2635 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2636 rxmsg->gnm_u.putreq.gnprm_cookie;
2637 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2638 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2639 (__u64)((unsigned long)tx->tx_buffer);
2640 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2642 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2644 /* we only queue from kgnilnd_recv - we might get called from other contexts
2645 * and we don't want to block the mutex in those cases */
2647 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2648 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2649 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2650 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2653 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2655 rxmsg->gnm_u.get.gngm_cookie,
2659 kgnilnd_consume_rx(rx);
2663 /* make sure we send an error back when the GET fails */
2664 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2665 kgnilnd_tx_done(tx, rc);
2666 kgnilnd_consume_rx(rx);
2668 /* return magic LNet network error */
2672 case GNILND_MSG_PUT_REQ_REV:
2673 /* LNET wants to truncate or drop transaction, sending NAK */
2675 kgnilnd_consume_rx(rx);
2676 lnet_finalize(ni, lntmsg, 0);
2678 /* only error if lntmsg == NULL, otherwise we are just
2679 * short circuiting the rdma process of 0 bytes */
2680 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2681 lntmsg == NULL ? -ENOENT : 0,
2682 rxmsg->gnm_u.get.gngm_cookie,
2687 if (lntmsg != NULL) {
2689 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2692 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2694 rxmsg->gnm_u.get.gngm_cookie,
2697 kgnilnd_consume_rx(rx);
2699 case GNILND_MSG_GET_REQ:
2700 if (lntmsg != NULL) {
2702 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2705 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2707 rxmsg->gnm_u.get.gngm_cookie,
2710 kgnilnd_consume_rx(rx);
2716 /* needs write_lock on kgn_peer_conn_lock held */
2718 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2720 unsigned long timeout, keepalive;
2721 unsigned long now = jiffies;
2722 unsigned long newest_last_rx;
2725 /* given that we found this conn hanging off a peer, it better damned
2726 * well be connected */
2727 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2728 "conn 0x%p->%s with bad state%s\n", conn,
2729 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2731 kgnilnd_conn_state2str(conn));
2733 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2734 "rx_diff %lu tx_diff %lu\n",
2735 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2736 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2737 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2738 cfs_duration_sec(now - conn->gnc_last_tx));
2740 timeout = cfs_time_seconds(conn->gnc_timeout);
2741 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2743 /* just in case our lack of RX msg processing is gumming up the works - give the
2744 * remove an extra chance */
2746 newest_last_rx = GNILND_LASTRX(conn);
2748 if (time_after_eq(now, newest_last_rx + timeout)) {
2749 uint32_t level = D_CONSOLE|D_NETERROR;
2751 if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_DOWN) {
2754 GNIDBG_CONN(level, conn,
2755 "No gnilnd traffic received from %s for %lu "
2756 "seconds, terminating connection. Is node down? ",
2757 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2758 cfs_duration_sec(now - newest_last_rx));
2762 /* we don't timeout on last_tx stalls - we are going to trust the
2763 * underlying network to let us know when sends are failing.
2764 * At worst, the peer will timeout our RX stamp and drop the connection
2765 * at that point. We'll then see his CLOSE or at worst his RX
2766 * stamp stop and drop the connection on our end */
2768 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2769 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2770 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2771 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2772 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2774 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2775 conn->gnc_last_noop_cq,
2776 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2777 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2778 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2779 set_mb(conn->gnc_last_noop_want, jiffies);
2780 atomic_inc(&conn->gnc_reaper_noop);
2781 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2784 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2787 kgnilnd_queue_tx(conn, tx);
2793 /* needs write_lock on kgn_peer_conn_lock held */
2795 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2796 struct list_head *souls)
2798 unsigned long timeout;
2799 kgn_conn_t *conn, *connN = NULL;
2805 short releaseconn = 0;
2806 unsigned long first_rx = 0;
2807 int purgatory_conn_cnt = 0;
2809 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2810 peer, libcfs_nid2str(peer->gnp_nid),
2811 peer->gnp_reconnect_interval);
2813 timeout = cfs_time_seconds(MAX(*kgnilnd_tunables.kgn_timeout,
2814 GNILND_MIN_TIMEOUT));
2816 conn = kgnilnd_find_conn_locked(peer);
2818 /* if there is a valid conn, check the queues for timeouts */
2819 rc = kgnilnd_check_conn_timeouts_locked(conn);
2821 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2822 /* simulate a RX CLOSE after the timeout but before
2823 * the scheduler thread gets it */
2824 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2825 conn->gnc_peer_error = -ETIMEDOUT;
2827 /* Once we mark closed, any of the scheduler threads could
2828 * get it and move through before we hit the fail loc code */
2829 kgnilnd_close_conn_locked(conn, rc);
2831 /* first_rx is used to decide when to release a conn from purgatory.
2833 first_rx = conn->gnc_first_rx;
2837 /* now regardless of starting new conn, find tx on peer queue that
2838 * are old and smell bad - do this first so we don't trigger
2839 * reconnect on empty queue if we timeout all */
2840 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2841 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2843 LCONSOLE_INFO("could not send to %s due to connection"
2844 " setup failure after %lu seconds\n",
2845 libcfs_nid2str(peer->gnp_nid),
2846 cfs_duration_sec(jiffies - tx->tx_qtime));
2848 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2850 list_add_tail(&tx->tx_list, todie);
2855 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2856 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2857 count, peer, libcfs_nid2str(peer->gnp_nid));
2858 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2859 if (list_empty(&peer->gnp_tx_queue) ||
2860 peer->gnp_connecting == GNILND_PEER_KILL) {
2861 /* we pass down todie to use a common function - but we know there are
2863 kgnilnd_cancel_peer_connect_locked(peer, todie);
2867 /* Don't reconnect if we are still trying to clear out old conns.
2868 * This prevents us sending traffic on the new mbox before ensuring we are done
2869 * with the old one */
2870 reconnect = (peer->gnp_down == GNILND_RCA_NODE_UP) &&
2871 (atomic_read(&peer->gnp_dirty_eps) == 0);
2873 /* fast reconnect after a timeout */
2874 to_reconn = !conn &&
2875 (peer->gnp_last_errno == -ETIMEDOUT) &&
2876 *kgnilnd_tunables.kgn_fast_reconn;
2878 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2879 * to be sent, we'll check the reconnect interval and fire up a new
2880 * connection request */
2883 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2884 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2885 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2887 CDEBUG(D_NET, "starting connect to %s\n",
2888 libcfs_nid2str(peer->gnp_nid));
2889 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE, "Peer was idle and we"
2890 "have a write_lock, state issue %d\n", peer->gnp_connecting);
2892 peer->gnp_connecting = GNILND_PEER_CONNECT;
2893 kgnilnd_peer_addref(peer); /* extra ref for connd */
2895 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2896 list_add_tail(&peer->gnp_connd_list,
2897 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2898 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2900 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2903 /* fail_loc to allow us to delay release of purgatory */
2904 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2907 /* This check allows us to verify that the new conn is actually being used. This allows us to
2908 * pull the old conns out of purgatory if they have actually seen traffic.
2909 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2912 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2913 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2914 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2918 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2919 /* check for purgatory timeouts */
2920 if (conn->gnc_in_purgatory) {
2921 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2922 * that way the next time we check it we can detach it from purgatory
2925 if (conn->gnc_state != GNILND_CONN_DONE) {
2926 /* Skip over conns that are currently not DONE. If they arent already scheduled
2927 * for completion something in the state machine is broken.
2932 /* We only detach a conn that is in purgatory if we have received a close message,
2933 * we have a new valid connection that has successfully received data, or an admin
2934 * command tells us we need to detach.
2937 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2938 unsigned long waiting;
2940 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2942 /* C.E: The remote peer is expected to close the
2943 * connection (see kgnilnd_check_conn_timeouts)
2944 * via the reaper thread and nuke out the MDD and
2945 * FMA resources after conn->gnc_timeout has expired
2946 * without an FMA RX */
2947 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2948 " held resources\n",
2949 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2950 cfs_duration_sec(waiting));
2952 kgnilnd_detach_purgatory_locked(conn, souls);
2954 purgatory_conn_cnt++;
2959 /* If we have too many connections in purgatory we could run out of
2960 * resources. Limit the number of connections to a tunable number,
2961 * clean up to the minimum all in one fell swoop... there are
2962 * situations where dvs will retry tx's and we can eat up several
2963 * hundread connection requests at once.
2965 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2966 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2968 if (conn->gnc_in_purgatory &&
2969 conn->gnc_state == GNILND_CONN_DONE) {
2970 CDEBUG(D_NET, "Dropping Held resource due to"
2971 " resource limits being hit\n");
2972 kgnilnd_detach_purgatory_locked(conn, souls);
2974 if (purgatory_conn_cnt-- <
2975 *kgnilnd_tunables.kgn_max_purgatory)
2985 kgnilnd_reaper_check(int idx)
2987 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2988 struct list_head *ctmp, *ctmpN;
2989 struct list_head geriatrics;
2990 struct list_head souls;
2992 INIT_LIST_HEAD(&geriatrics);
2993 INIT_LIST_HEAD(&souls);
2995 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2997 list_for_each_safe(ctmp, ctmpN, peers) {
2998 kgn_peer_t *peer = NULL;
3000 /* don't timeout stuff if the network is mucked or shutting down */
3001 if (kgnilnd_check_hw_quiesce()) {
3004 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
3006 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
3009 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3011 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3012 kgnilnd_release_purgatory_list(&souls);
3016 kgnilnd_update_reaper_timeout(long timeout)
3018 LASSERT(timeout > 0);
3020 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3022 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3023 kgnilnd_data.kgn_new_min_timeout = timeout;
3025 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3029 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3031 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3035 kgnilnd_reaper(void *arg)
3040 unsigned long next_check_time = jiffies;
3041 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3042 struct timer_list timer;
3045 cfs_block_allsigs();
3047 /* all gnilnd threads need to run fairly urgently */
3048 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3049 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3051 while (!kgnilnd_data.kgn_shutdown) {
3052 /* I wake up every 'p' seconds to check for timeouts on some
3053 * more peers. I try to check every connection 'n' times
3054 * within the global minimum of all keepalive and timeout
3055 * intervals, to ensure I attend to every connection within
3056 * (n+1)/n times its timeout intervals. */
3057 const int p = GNILND_REAPER_THREAD_WAKE;
3058 const int n = GNILND_REAPER_NCHECKS;
3060 /* to quiesce or to not quiesce, that is the question */
3061 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3062 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3063 KGNILND_SPIN_QUIESCE;
3064 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3067 /* careful with the jiffy wrap... */
3068 timeout = (long)(next_check_time - jiffies);
3071 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3072 TASK_INTERRUPTIBLE);
3073 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3074 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3076 mod_timer(&timer, (long) jiffies + timeout);
3078 /* check flag variables before committing */
3079 if (!kgnilnd_data.kgn_shutdown &&
3080 !kgnilnd_data.kgn_quiesce_trigger) {
3081 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3082 timeout, cfs_duration_sec(timeout));
3084 CDEBUG(D_INFO, "awake after schedule\n");
3087 del_singleshot_timer_sync(&timer);
3088 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3089 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3093 /* new_min_timeout is set from the conn timeouts and keepalive
3094 * this should end up with a min timeout of
3095 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3096 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3097 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3098 CDEBUG(D_NET, "Set new min timeout %ld\n",
3099 current_min_timeout);
3102 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3104 /* Compute how many table entries to check now so I get round
3105 * the whole table fast enough given that I do this at fixed
3106 * intervals of 'p' seconds) */
3107 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3108 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3109 chunk = (chunk * n * p) /
3110 kgnilnd_data.kgn_new_min_timeout;
3113 for (i = 0; i < chunk; i++) {
3114 kgnilnd_reaper_check(hash_index);
3115 hash_index = (hash_index + 1) %
3116 *kgnilnd_tunables.kgn_peer_hash_size;
3118 next_check_time = (long) jiffies + cfs_time_seconds(p);
3119 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3121 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3124 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3126 kgnilnd_thread_fini();
3131 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3132 unsigned niov, offset, nob;
3134 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
3135 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3138 lnet_copy_flat2kiov(
3141 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3143 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3150 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3153 gni_post_descriptor_t *desc;
3155 kgn_tx_ev_id_t ev_id;
3157 int should_retry, rc;
3158 long num_processed = 0;
3159 kgn_conn_t *conn = NULL;
3160 kgn_tx_t *tx = NULL;
3161 kgn_rdma_desc_t *rdesc;
3166 /* make sure we don't keep looping if we need to reset */
3167 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3168 return num_processed;
3170 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3172 /* we didn't get the mutex, so return that there is still work
3176 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3177 /* a bit gross - but we need a good way to test for
3178 * delayed RDMA completions and the easiest way to do
3179 * that is to delay the RDMA CQ events */
3180 rrc = GNI_RC_NOT_DONE;
3182 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3185 if (rrc == GNI_RC_NOT_DONE) {
3186 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3187 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3188 dev->gnd_id, num_processed);
3189 return num_processed;
3191 dev->gnd_sched_alive = jiffies;
3194 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3195 "this is bad, somehow our credits didn't protect us"
3196 " from CQ overrun\n");
3197 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3198 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3199 event_data, GNI_CQ_GET_TYPE(event_data));
3201 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3203 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3205 /* XXX Nic: Need better error handling here... */
3206 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3207 (rrc == GNI_RC_TRANSACTION_ERROR),
3210 ev_id.txe_cookie = desc->post_id;
3212 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3214 if (conn == NULL || tx == NULL) {
3215 /* either conn or tx was already nuked and this is a "late"
3216 * completion, so drop it */
3220 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3221 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3222 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3223 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3224 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3226 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3228 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3229 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3230 tx->tx_msg.gnm_u.completion.gncm_retval);
3234 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3235 if (tx->tx_buffer_copy != NULL)
3236 kgnilnd_recv_bte_get(tx);
3237 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3240 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3241 if (tx->tx_buffer_copy != NULL)
3242 kgnilnd_recv_bte_get(tx);
3243 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3246 /* remove from rdmaq */
3247 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3248 spin_lock(&conn->gnc_list_lock);
3249 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3250 spin_unlock(&conn->gnc_list_lock);
3251 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3253 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3254 event_data = 1LL << 48;
3258 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3259 atomic_inc(&dev->gnd_rdma_ntx);
3260 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3261 /* transaction succeeded, add into fmaq */
3262 kgnilnd_queue_tx(conn, tx);
3263 kgnilnd_peer_alive(conn->gnc_peer);
3265 /* drop ref from kgnilnd_validate_tx_ev_id */
3266 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3267 kgnilnd_conn_decref(conn);
3272 /* fall through to the TRANSACTION_ERROR case */
3275 /* get stringified version for log messages */
3276 kgnilnd_cq_error_str(event_data, &err_str, 256);
3277 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3279 /* make sure we are not off in the weeds with this tx */
3280 if (tx->tx_retrans >
3281 *kgnilnd_tunables.kgn_max_retransmits) {
3282 GNIDBG_TX(D_NETERROR, tx,
3283 "giving up on TX, too many retries", NULL);
3287 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3288 should_retry ? "transient" : "unrecoverable", err_str);
3290 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3291 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3292 rdesc = &tx->tx_putinfo.gnpam_desc;
3293 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3294 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3296 rdesc = &tx->tx_getinfo.gngm_desc;
3297 rnob = tx->tx_lntmsg[0]->msg_len;
3298 rcookie = tx->tx_getinfo.gngm_cookie;
3303 tx->tx_msg.gnm_type,
3307 kgnilnd_nak_rdma(conn,
3308 tx->tx_msg.gnm_type,
3311 tx->tx_msg.gnm_srcnid);
3312 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3313 kgnilnd_close_conn(conn, -ECOMM);
3316 /* drop ref from kgnilnd_validate_tx_ev_id */
3317 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3318 kgnilnd_conn_decref(conn);
3323 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3327 kgn_tx_ev_id_t ev_id;
3328 kgn_tx_t *tx = NULL;
3329 kgn_conn_t *conn = NULL;
3330 int queued_fma, saw_reply, rc;
3331 long num_processed = 0;
3334 /* make sure we don't keep looping if we need to reset */
3335 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3336 return num_processed;
3339 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3341 /* we didn't get the mutex, so return that there is still work
3346 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3347 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3349 if (rrc == GNI_RC_NOT_DONE) {
3351 "SMSG send CQ %d not ready (data %#llx) "
3352 "processed %ld\n", dev->gnd_id, event_data,
3354 return num_processed;
3357 dev->gnd_sched_alive = jiffies;
3360 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3361 "this is bad, somehow our credits didn't "
3362 "protect us from CQ overrun\n");
3363 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3364 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3365 event_data, GNI_CQ_GET_TYPE(event_data));
3367 /* if SMSG couldn't handle an error, time for conn to die */
3368 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3371 /* need to take the write_lock to ensure atomicity
3372 * on the conn state if we need to close it */
3373 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3374 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3376 /* Conn was destroyed? */
3378 "SMSG CQID lookup %#llx failed\n",
3379 GNI_CQ_GET_INST_ID(event_data));
3380 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3384 kgnilnd_cq_error_str(event_data, &err_str, 256);
3385 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3386 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3388 kgnilnd_close_conn_locked(conn, -ECOMM);
3390 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3392 /* no need to process rest of this tx -
3393 * it is getting canceled */
3397 /* fall through to GNI_RC_SUCCESS case */
3398 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3400 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3401 if (conn == NULL || tx == NULL) {
3402 /* either conn or tx was already nuked and this is a "late"
3403 * completion, so drop it */
3407 tx->tx_conn->gnc_last_tx_cq = jiffies;
3408 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3409 set_mb(conn->gnc_last_noop_cq, jiffies);
3412 /* lock tx_list_state and tx_state */
3413 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3414 spin_lock(&tx->tx_conn->gnc_list_lock);
3416 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3417 "state not GNILND_TX_LIVE_FMAQ", NULL);
3418 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3419 "not waiting for completion", NULL);
3421 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3424 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3426 /* This will trigger other FMA sends that were
3427 * pending this completion */
3428 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3430 /* we either did not expect reply or we already got it */
3431 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3433 spin_unlock(&tx->tx_conn->gnc_list_lock);
3434 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3437 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3439 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3440 kgnilnd_schedule_conn(conn);
3443 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3444 * If saw_reply is true we know that the tx is safe to use as the other thread
3445 * is already finished with it.
3449 /* no longer need to track on the live_fmaq */
3450 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3452 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3453 /* we already got reply & were waiting for
3454 * completion of initial send */
3455 /* to initiate RDMA transaction */
3456 GNIDBG_TX(D_NET, tx,
3457 "Pending RDMA 0x%p type 0x%02x",
3458 tx->tx_msg.gnm_type);
3459 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3460 rc = kgnilnd_send_mapped_tx(tx, 0);
3461 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3463 /* we are done with this tx */
3464 GNIDBG_TX(D_NET, tx,
3465 "Done with tx type 0x%02x",
3466 tx->tx_msg.gnm_type);
3467 kgnilnd_tx_done(tx, tx->tx_rc);
3471 /* drop ref from kgnilnd_validate_tx_ev_id */
3472 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3473 kgnilnd_conn_decref(conn);
3475 /* if we are waiting for a REPLY, we'll handle the tx then */
3476 } /* end for loop */
3480 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3485 long num_processed = 0;
3486 struct list_head *conns;
3487 struct list_head *tmp;
3491 /* make sure we don't keep looping if we need to reset */
3492 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3493 return num_processed;
3496 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3498 /* we didn't get the mutex, so return that there is still work
3502 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3503 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3505 if (rrc == GNI_RC_NOT_DONE) {
3506 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3508 dev->gnd_id, event_data, num_processed);
3509 return num_processed;
3511 dev->gnd_sched_alive = jiffies;
3514 /* this is the only CQ that can really handle transient
3516 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3517 rrc = cfs_fail_val ? cfs_fail_val
3518 : GNI_RC_ERROR_RESOURCE;
3519 if (rrc == GNI_RC_ERROR_RESOURCE) {
3520 /* set overrun too */
3521 event_data |= (1UL << 63);
3522 LASSERTF(GNI_CQ_OVERRUN(event_data),
3523 "(1UL << 63) is no longer the bit to"
3524 "set to indicate CQ_OVERRUN\n");
3527 /* sender should get error event too and take care
3528 of failed transaction by re-transmitting */
3529 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3530 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3534 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3535 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3536 conn = kgnilnd_cqid2conn_locked(
3537 GNI_CQ_GET_INST_ID(event_data));
3539 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3540 "failed, dropping event %#llx\n",
3541 GNI_CQ_GET_INST_ID(event_data),
3544 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3546 GNI_CQ_GET_INST_ID(event_data),
3547 conn, conn->gnc_peer ?
3548 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3551 conn->gnc_last_rx_cq = jiffies;
3553 /* stash first rx so we can clear out purgatory.
3555 if (conn->gnc_first_rx == 0) {
3556 conn->gnc_first_rx = jiffies;
3558 kgnilnd_peer_alive(conn->gnc_peer);
3559 kgnilnd_schedule_conn(conn);
3561 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3565 /* FMA CQ has overflowed: check ALL conns */
3566 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3567 "conns on device %d\n", dev->gnd_id);
3569 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3571 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3572 conns = &kgnilnd_data.kgn_conns[rc];
3574 list_for_each(tmp, conns) {
3575 conn = list_entry(tmp, kgn_conn_t,
3578 if (conn->gnc_device == dev) {
3579 kgnilnd_schedule_conn(conn);
3580 conn->gnc_last_rx_cq = jiffies;
3584 /* don't block write lockers for too long... */
3585 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3590 /* try_map_if_full should only be used when processing TX from list of
3591 * backlog TX waiting on mappings to free up
3594 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3595 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3598 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3600 /* slight bit of race if multiple people calling, but at worst we'll have
3601 * order altered just a bit... which would not be determenistic anyways */
3602 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3604 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3606 /* We know that we have a GART reservation that should guarantee forward progress.
3607 * This means we don't need to take any extraordinary efforts if we are failing
3608 * mappings here - even if we are holding a very small number of these. */
3610 if (try_map_if_full || (rc == 0)) {
3611 rc = kgnilnd_map_buffer(tx);
3614 /* rc should be 0 if we mapped successfully here, if non-zero
3615 * we are queueing */
3617 /* if try_map_if_full set, they handle requeuing */
3618 if (unlikely(try_map_if_full)) {
3621 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3622 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3623 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3624 /* make sure we wake up sched to run this */
3625 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3626 /* return 0 as this is now queued for later sending */
3631 switch (tx->tx_msg.gnm_type) {
3635 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3636 * remote node where the RDMA will be started
3637 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3638 * be satisified. The rest of the errors are "hard" errors that require
3639 * upper layers to handle themselves.
3640 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3641 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3642 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3643 * GNILND_MSG_GET_DONE_REV.
3645 case GNILND_MSG_GET_REQ:
3646 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3647 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3648 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3649 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3650 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3651 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3652 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3654 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3655 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3657 case GNILND_MSG_PUT_ACK:
3658 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3659 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3660 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3661 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3663 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3664 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3667 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3668 case GNILND_MSG_PUT_DONE:
3669 case GNILND_MSG_PUT_REQ:
3670 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3671 &tx->tx_putinfo.gnpam_desc,
3672 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3673 tx->tx_putinfo.gnpam_dst_cookie);
3674 RETURN(try_map_if_full ? rc : 0);
3676 case GNILND_MSG_GET_DONE:
3677 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3678 &tx->tx_getinfo.gngm_desc,
3679 tx->tx_lntmsg[0]->msg_len,
3680 tx->tx_getinfo.gngm_cookie);
3681 RETURN(try_map_if_full ? rc : 0);
3683 case GNILND_MSG_PUT_REQ_REV:
3684 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3685 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3686 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3687 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3688 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3689 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3690 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3692 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3694 case GNILND_MSG_PUT_DONE_REV:
3695 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3696 &tx->tx_getinfo.gngm_desc,
3698 tx->tx_getinfo.gngm_cookie);
3699 RETURN(try_map_if_full ? rc : 0);
3701 case GNILND_MSG_GET_ACK_REV:
3702 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3703 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3704 /* LNET_GETS are a special case for parse */
3705 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3706 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3708 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3709 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3711 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3712 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3714 case GNILND_MSG_GET_DONE_REV:
3715 case GNILND_MSG_GET_REQ_REV:
3716 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3717 &tx->tx_putinfo.gnpam_desc,
3718 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3719 tx->tx_putinfo.gnpam_dst_cookie);
3720 RETURN(try_map_if_full ? rc : 0);
3728 kgnilnd_process_fmaq(kgn_conn_t *conn)
3731 kgn_tx_t *tx = NULL;
3732 void *buffer = NULL;
3733 unsigned int nob = 0;
3736 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3737 * However I will be rescheduled by an FMA completion event
3738 * when I eventually get some.
3739 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3740 * But it doesn't matter if I try to send a "real" message just
3741 * as I start closing because I'll get scheduled to send the
3744 /* Short circuit if the ep_handle is null we cant send anyway. */
3745 if (conn->gnc_ephandle == NULL)
3748 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3750 spin_lock(&conn->gnc_list_lock);
3752 if (list_empty(&conn->gnc_fmaq)) {
3753 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3755 spin_unlock(&conn->gnc_list_lock);
3757 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3758 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3759 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3760 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3761 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3763 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3764 conn->gnc_last_noop_cq,
3765 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3766 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3767 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3768 atomic_inc(&conn->gnc_sched_noop);
3769 set_mb(conn->gnc_last_noop_want, jiffies);
3771 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3774 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3778 rc = kgnilnd_set_tx_id(tx, conn);
3780 kgnilnd_tx_done(tx, rc);
3786 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3787 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3788 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3789 more_to_do = !list_empty(&conn->gnc_fmaq);
3790 spin_unlock(&conn->gnc_list_lock);
3793 /* if there is no real TX or no NOOP to send, bail */
3798 if (!tx->tx_retrans)
3799 tx->tx_cred_wait = jiffies;
3801 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3802 "tx with zero id", NULL);
3804 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3805 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3806 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3810 switch (tx->tx_msg.gnm_type) {
3814 case GNILND_MSG_NOOP:
3815 case GNILND_MSG_CLOSE:
3816 case GNILND_MSG_IMMEDIATE:
3817 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3818 buffer = tx->tx_buffer;
3822 case GNILND_MSG_GET_DONE:
3823 case GNILND_MSG_PUT_DONE:
3824 case GNILND_MSG_PUT_DONE_REV:
3825 case GNILND_MSG_GET_DONE_REV:
3826 case GNILND_MSG_PUT_NAK:
3827 case GNILND_MSG_GET_NAK:
3828 case GNILND_MSG_GET_NAK_REV:
3829 case GNILND_MSG_PUT_NAK_REV:
3830 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3833 case GNILND_MSG_PUT_REQ:
3834 case GNILND_MSG_GET_REQ_REV:
3835 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3837 case GNILND_MSG_PUT_ACK:
3838 case GNILND_MSG_PUT_REQ_REV:
3839 case GNILND_MSG_GET_ACK_REV:
3840 case GNILND_MSG_GET_REQ:
3841 /* This is really only to handle the retransmit of SMSG once these
3842 * two messages are setup in send_mapped_tx */
3843 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3847 if (likely(rc == 0)) {
3848 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3852 /* don't explicitly reschedule here - we are short credits and will rely on
3853 * kgnilnd_sendmsg to resched the conn if need be */
3855 } else if (rc < 0) {
3856 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3857 * almost certainly a software bug, but lets play nice with the other kids */
3858 kgnilnd_tx_done(tx, rc);
3859 /* just for fun, kick peer in arse - resetting conn might help to correct
3860 * this almost certainly buggy software caused return code */
3861 kgnilnd_close_conn(conn, rc);
3865 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3866 kgnilnd_schedule_conn(conn);
3871 kgnilnd_process_rdmaq(kgn_device_t *dev)
3876 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3880 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3881 unsigned long dead_bump;
3884 /* if we think we need to adjust, take lock to serialize and recheck */
3885 spin_lock(&dev->gnd_rdmaq_lock);
3886 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3887 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3889 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3891 /* roll the bucket forward */
3892 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3894 if (kgnilnd_data.kgn_rdmaq_override &&
3895 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3896 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3901 /* roll current outstanding forward to make sure we carry outstanding
3902 * committment forward
3903 * new_ok starts out as the whole interval value
3904 * - first subtract bytes_out from last interval, as that would push us over
3905 * strict limits for this interval
3906 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3908 * there is a small race here if someone is actively processing mappings and
3909 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3910 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3911 * - if this gives us problems in the future, we could use a read/write lock
3912 * to protect the resetting of these values */
3913 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3914 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3916 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3917 "current out %ld\n",
3918 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3919 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3921 spin_unlock(&dev->gnd_rdmaq_lock);
3924 spin_lock(&dev->gnd_rdmaq_lock);
3925 while (!list_empty(&dev->gnd_rdmaq)) {
3928 /* make sure we break out early on quiesce */
3929 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3930 /* always break with lock held - we unlock outside loop */
3934 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3935 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3938 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3939 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3940 /* if conn is dying, mark tx in tx_ref_table for
3941 * kgnilnd_complete_closed_conn to finish up */
3942 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3944 /* tx was moved to DYING, get next */
3947 spin_unlock(&dev->gnd_rdmaq_lock);
3949 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3950 spin_lock(&dev->gnd_rdmaq_lock);
3953 /* no ticket! add back to head */
3954 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3955 /* clear found_work so scheduler threads wait for timer */
3959 /* TX is GO for launch */
3960 tx->tx_qtime = jiffies;
3961 kgnilnd_send_mapped_tx(tx, 0);
3965 spin_unlock(&dev->gnd_rdmaq_lock);
3971 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3973 __swab64s(&d->gnrd_key.qword1);
3974 __swab64s(&d->gnrd_key.qword2);
3975 __swab64s(&d->gnrd_addr);
3976 __swab32s(&d->gnrd_nob);
3979 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3980 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3983 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3985 kgn_tx_ev_id_t ev_id;
3988 /* we use the cookie from the original TX, so we can find the match
3989 * by parsing that and using the txe_idx */
3990 ev_id.txe_cookie = cookie;
3992 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3995 /* check tx to make sure kgni didn't eat it */
3996 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3997 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3999 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
4000 (tx->tx_id.txe_cookie = cookie)),
4001 "conn 0x%p->%s tx_ref_table hosed: wanted "
4002 "txe_cookie %#llx txe_idx %d "
4003 "found tx %p cookie %#llx txe_idx %d\n",
4004 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4005 cookie, ev_id.txe_idx,
4006 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
4008 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4009 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4010 "Unexpected TX type (%x, %x or %x) "
4011 "or state (%x, expected +%x) "
4012 "matched reply from %s\n",
4013 tx->tx_msg.gnm_type, type1, type2,
4014 tx->tx_state, GNILND_TX_WAITING_REPLY,
4015 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4017 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
4018 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4024 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4027 kgn_conn_t *conn = tx->tx_conn;
4028 __u64 nob = tx->tx_nob;
4029 __u32 physnop = tx->tx_phys_npages;
4030 int id = tx->tx_id.txe_smsg_id;
4031 int buftype = tx->tx_buftype;
4032 gni_mem_handle_t hndl;
4033 hndl.qword1 = tx->tx_map_key.qword1;
4034 hndl.qword2 = tx->tx_map_key.qword2;
4036 spin_lock(&conn->gnc_list_lock);
4038 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4039 "not waiting for reply", NULL);
4042 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4044 if (rc == -EFAULT) {
4045 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
4046 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4048 if(*kgnilnd_tunables.kgn_efault_lbug) {
4049 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4050 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4052 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4053 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4054 kgnilnd_tx_state2str(tx->tx_list_state),
4055 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4060 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4061 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4062 /* sample under lock as follow on steps require gnc_list_lock
4063 * - or call kgnilnd_tx_done which requires no locks held over
4064 * call to lnet_finalize */
4067 spin_unlock(&conn->gnc_list_lock);
4070 kgnilnd_tx_done(tx, tx->tx_rc);
4075 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4078 kgn_conn_t *conn = tx->tx_conn;
4080 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4081 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4083 /* the gncm_retval is passed in for PUTs */
4084 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4085 msg->gnm_u.completion.gncm_retval);
4087 kgnilnd_complete_tx(tx, rc);
4091 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4099 kgn_peer_t *peer = conn->gnc_peer;
4102 __u16 tmp_cksum = 0, msg_cksum = 0;
4103 int repost = 1, saw_complete;
4104 unsigned long timestamp, newest_last_rx, timeout;
4108 /* Short circuit if the ep_handle is null.
4109 * It's likely that its about to be closed as stale.
4111 if (conn->gnc_ephandle == NULL)
4114 timestamp = jiffies;
4115 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4116 /* delay in jiffies - we are really concerned only with things that
4117 * result in a schedule() or really holding this off for long times .
4118 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4119 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4121 /* Resample current time as we have no idea how long it took to get the mutex */
4122 timestamp = jiffies;
4124 /* We check here when the last time we received an rx, we do this before
4125 * we call getnext in case the thread has been blocked for a while. If we
4126 * havent received an rx since our timeout value we close the connection
4127 * as we should assume the other side has closed the connection. This will
4128 * stop us from sending replies to a mailbox that is already in purgatory.
4131 timeout = cfs_time_seconds(conn->gnc_timeout);
4132 newest_last_rx = GNILND_LASTRX(conn);
4134 /* Error injection to validate that timestamp checking works and closing the conn */
4135 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4136 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4139 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4140 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4141 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4142 cfs_duration_sec(timestamp - newest_last_rx),
4143 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4144 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4146 kgnilnd_close_conn(conn, rc);
4150 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4152 if (rrc == GNI_RC_NOT_DONE) {
4153 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4154 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4158 /* Instead of asserting when we get mailbox corruption lets attempt to
4159 * close the conn and recover. We can put the conn/mailbox into
4160 * purgatory and let purgatory deal with the problem. If we see
4161 * this NETTERROR reported on production systems in large amounts
4162 * we will need to revisit the state machine to see if we can tighten
4163 * it up further to improve data protection.
4166 if (rrc == GNI_RC_INVALID_STATE) {
4167 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4168 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4169 "detected closing conn %p from peer %s\n", conn,
4170 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4172 kgnilnd_close_conn(conn, rc);
4176 LASSERTF(rrc == GNI_RC_SUCCESS,
4177 "bad rc %d on conn %p from peer %s\n",
4178 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4180 msg = (kgn_msg_t *)prefix;
4182 rx = kgnilnd_alloc_rx();
4184 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4185 kgnilnd_release_msg(conn);
4186 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4187 conn, libcfs_nid2str(peer->gnp_nid));
4191 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4193 timestamp = conn->gnc_last_rx;
4194 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4195 atomic_inc(&conn->gnc_rx_seq);
4197 conn->gnc_last_rx = jiffies;
4198 /* stash first rx so we can clear out purgatory
4200 if (conn->gnc_first_rx == 0)
4201 conn->gnc_first_rx = jiffies;
4203 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4204 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4205 kgnilnd_peer_alive(conn->gnc_peer);
4208 rx->grx_conn = conn;
4210 rx->grx_received = current_kernel_time();
4212 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4215 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4221 kgnilnd_net_decref(net);
4224 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4225 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4227 /* XXX Nic: Do we need to swab cksum */
4228 if (msg->gnm_cksum != 0) {
4229 msg_cksum = msg->gnm_cksum;
4231 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4233 if (tmp_cksum != msg_cksum) {
4234 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4235 tmp_cksum, msg_cksum);
4236 kgnilnd_dump_msg(D_BUFFS, msg);
4241 /* restore checksum for future debug messages */
4242 msg->gnm_cksum = tmp_cksum;
4244 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4245 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4246 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4247 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4252 __swab32s(&msg->gnm_magic);
4253 __swab16s(&msg->gnm_version);
4254 __swab16s(&msg->gnm_type);
4255 __swab64s(&msg->gnm_srcnid);
4256 __swab64s(&msg->gnm_connstamp);
4257 __swab32s(&msg->gnm_seq);
4259 /* NB message type checked below; NOT here... */
4260 switch (msg->gnm_type) {
4261 case GNILND_MSG_GET_ACK_REV:
4262 case GNILND_MSG_PUT_ACK:
4263 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4266 case GNILND_MSG_PUT_REQ_REV:
4267 case GNILND_MSG_GET_REQ:
4268 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4276 if (msg->gnm_version != GNILND_MSG_VERSION) {
4277 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4278 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4283 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4284 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4285 libcfs_nid2str(msg->gnm_srcnid),
4286 libcfs_nid2str(peer->gnp_nid));
4291 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4292 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4293 " expected) from %s",
4294 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4295 libcfs_nid2str(peer->gnp_nid));
4300 if (msg->gnm_seq != seq) {
4301 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4302 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4307 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4309 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4310 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4311 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4312 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4313 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4314 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4315 /* double check state with lock held */
4316 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4317 /* only error if we are not already closing */
4318 if (conn->gnc_peer_error == -ETIMEDOUT) {
4319 unsigned long now = jiffies;
4320 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4322 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4323 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4324 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4326 cfs_duration_sec(now - timestamp),
4327 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4328 atomic_read(&conn->gnc_tx_seq),
4329 cfs_duration_sec(now - conn->gnc_last_tx),
4330 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4331 cfs_duration_sec(now - conn->gnc_last_noop_want),
4332 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4333 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4334 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4335 cfs_duration_sec(now - conn->gnc_last_sched_do),
4336 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4338 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4340 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4344 if (conn->gnc_close_recvd) {
4345 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4346 kgnilnd_msgtype2str(msg->gnm_type),
4347 msg->gnm_type, conn->gnc_close_recvd,
4348 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4353 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4354 /* XXX Nic: log message received on bad connection state */
4358 switch (msg->gnm_type) {
4359 case GNILND_MSG_NOOP:
4360 /* Nothing to do; just a keepalive */
4363 case GNILND_MSG_IMMEDIATE:
4364 /* only get SMSG payload for IMMEDIATE */
4365 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4366 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4367 msg->gnm_srcnid, rx, 0);
4370 case GNILND_MSG_GET_REQ_REV:
4371 case GNILND_MSG_PUT_REQ:
4372 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4373 msg->gnm_srcnid, rx, 1);
4376 case GNILND_MSG_GET_NAK_REV:
4377 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4378 msg->gnm_u.completion.gncm_cookie);
4382 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4384 case GNILND_MSG_PUT_NAK:
4385 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4386 msg->gnm_u.completion.gncm_cookie);
4390 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4392 case GNILND_MSG_PUT_ACK:
4393 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4394 msg->gnm_u.putack.gnpam_src_cookie);
4398 /* store putack data for later: deferred rdma or re-try */
4399 tx->tx_putinfo = msg->gnm_u.putack;
4402 spin_lock(&tx->tx_conn->gnc_list_lock);
4404 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4405 "not waiting for reply", NULL);
4407 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4409 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4410 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4411 /* sample under lock as follow on steps require gnc_list_lock
4412 * - or call kgnilnd_tx_done which requires no locks held over
4413 * call to lnet_finalize */
4416 /* cannot launch rdma if still waiting for fma-msg completion */
4417 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4418 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4419 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4421 spin_unlock(&tx->tx_conn->gnc_list_lock);
4424 rc = kgnilnd_send_mapped_tx(tx, 0);
4426 kgnilnd_tx_done(tx, rc);
4429 case GNILND_MSG_GET_ACK_REV:
4430 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4431 msg->gnm_u.putack.gnpam_src_cookie);
4435 /* store putack data for later: deferred rdma or re-try */
4436 tx->tx_putinfo = msg->gnm_u.putack;
4438 spin_lock(&tx->tx_conn->gnc_list_lock);
4440 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4441 "not waiting for reply", NULL);
4443 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4445 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4446 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4447 /* sample under lock as follow on steps require gnc_list_lock
4448 * - or call kgnilnd_tx_done which requires no locks held over
4449 * call to lnet_finalize */
4452 /* cannot launch rdma if still waiting for fma-msg completion */
4453 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4454 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4455 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4457 spin_unlock(&tx->tx_conn->gnc_list_lock);
4460 rc = kgnilnd_send_mapped_tx(tx, 0);
4462 kgnilnd_tx_done(tx, rc);
4465 case GNILND_MSG_PUT_DONE:
4466 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4467 msg->gnm_u.completion.gncm_cookie);
4471 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4472 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4473 "bad tx buftype %d", tx->tx_buftype);
4475 kgnilnd_finalize_rx_done(tx, msg);
4477 case GNILND_MSG_PUT_REQ_REV:
4478 case GNILND_MSG_GET_REQ:
4479 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4480 msg->gnm_srcnid, rx, 1);
4484 case GNILND_MSG_GET_NAK:
4485 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4486 msg->gnm_u.completion.gncm_cookie);
4490 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4491 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4492 "bad tx buftype %d", tx->tx_buftype);
4494 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4497 case GNILND_MSG_GET_DONE:
4498 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4499 msg->gnm_u.completion.gncm_cookie);
4503 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4504 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4505 "bad tx buftype %d", tx->tx_buftype);
4507 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4508 msg->gnm_u.completion.gncm_retval);
4510 kgnilnd_finalize_rx_done(tx, msg);
4512 case GNILND_MSG_GET_DONE_REV:
4513 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4514 msg->gnm_u.completion.gncm_cookie);
4518 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4519 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4520 "bad tx buftype %d", tx->tx_buftype);
4522 kgnilnd_finalize_rx_done(tx, msg);
4525 case GNILND_MSG_PUT_DONE_REV:
4526 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4527 msg->gnm_u.completion.gncm_cookie);
4532 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4533 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4534 "bad tx buftype %d", tx->tx_buftype);
4536 kgnilnd_finalize_rx_done(tx, msg);
4538 case GNILND_MSG_PUT_NAK_REV:
4539 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4540 msg->gnm_u.completion.gncm_cookie);
4545 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4546 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4547 "bad tx buftype %d", tx->tx_buftype);
4549 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4554 if (rc < 0) /* protocol/comms error */
4555 kgnilnd_close_conn(conn, rc);
4557 if (repost && rx != NULL) {
4558 kgnilnd_consume_rx(rx);
4561 /* we got an event so assume more there and call for reschedule */
4563 kgnilnd_schedule_conn(conn);
4567 /* Do the failure injections that we need to affect conn processing in the following function.
4568 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4569 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4572 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4576 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4580 /* short circuit out when not set */
4581 if (likely(!cfs_fail_loc)) {
4585 /* failure injection to test for stack reset clean ups */
4586 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4587 /* we can't rely on busy loops being nice enough to get the
4588 * stack reset triggered - it'd just spin on this conn */
4589 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4592 GOTO(did_fail_loc, rc);
4595 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4596 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4598 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4599 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4602 GOTO(did_fail_loc, rc);
4606 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4607 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4608 * will be closed by kgnilnd_complete_closed_conn.
4610 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4611 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4614 GOTO(did_fail_loc, rc);
4617 /* this one is a bit gross - we can't hold the mutex from process_conns
4618 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4619 * and moving onto finish_connect
4620 * so, we'll just set the rc - kgnilnd_process_conns will clear
4621 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4622 * and effectively getting this thread to sleep */
4623 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4626 GOTO(did_fail_loc, rc);
4634 kgnilnd_send_conn_close(kgn_conn_t *conn)
4638 /* we are closing the conn - we will try to send the CLOSE msg
4639 * but will not wait for anything else to flush */
4641 /* send the close if not already done so or received one */
4642 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4643 /* set close_sent regardless of the success of the
4644 * CLOSE message. We are going to try once and then
4645 * kick him out of the sandbox */
4646 conn->gnc_close_sent = 1;
4649 /* EP might be null already if remote side initiated a new connection.
4650 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4651 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4653 if (conn->gnc_ephandle != NULL) {
4656 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4658 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4659 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4660 tx->tx_qtime = jiffies;
4662 if (tx->tx_id.txe_idx == 0) {
4663 rc = kgnilnd_set_tx_id(tx, conn);
4665 kgnilnd_tx_done(tx, rc);
4669 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4672 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4673 kgnilnd_tx_done(tx, -EAGAIN);
4675 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4677 /* It wasnt sent and we dont care. */
4678 kgnilnd_tx_done(tx, rc);
4686 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4687 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4688 conn->gnc_state = GNILND_CONN_CLOSED;
4689 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4690 /* mark this conn as CLOSED now that we processed it
4691 * do after TX, so we can use CLOSING in asserts */
4695 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4696 /* simulate a RX CLOSE after the timeout but before
4697 * the scheduler thread gets it */
4698 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4699 conn->gnc_peer_error = -ETIMEDOUT;
4701 /* schedule to allow potential CLOSE and get the complete phase run */
4702 kgnilnd_schedule_conn(conn);
4706 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4711 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4712 int log_retrans, log_retrans_level;
4713 static int last_map_version;
4716 spin_lock(&dev->gnd_lock);
4717 if (list_empty(&dev->gnd_map_tx)) {
4718 /* if the list is empty make sure we dont have a timer running */
4719 del_singleshot_timer_sync(&dev->gnd_map_timer);
4720 spin_unlock(&dev->gnd_lock);
4724 dev->gnd_sched_alive = jiffies;
4726 /* we'll retry as fast as possible up to 25% of the limit, then we start
4727 * backing off until our map version changes - indicating we unmapped
4729 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4730 if (likely(dev->gnd_map_attempt == 0) ||
4731 time_after_eq(jiffies, dev->gnd_next_map) ||
4732 last_map_version != dev->gnd_map_version) {
4734 /* if this is our first attempt at mapping set last mapped to current
4735 * jiffies so we can timeout our attempt correctly.
4737 if (dev->gnd_map_attempt == 0)
4738 dev->gnd_last_map = jiffies;
4740 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4741 spin_unlock(&dev->gnd_lock);
4745 /* delete the previous timer if it exists */
4746 del_singleshot_timer_sync(&dev->gnd_map_timer);
4747 /* stash the last map version to let us know when a good one was seen */
4748 last_map_version = dev->gnd_map_version;
4750 /* we need to to take the lock and continually refresh the head of the list as
4751 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4752 * allowing them to squeeze in */
4754 while (!list_empty(&dev->gnd_map_tx)) {
4755 /* make sure we break out early on quiesce */
4756 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4757 /* always break with lock held - we unlock outside loop */
4761 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4763 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4766 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4767 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4768 /* if conn is dying, mark tx in tx_ref_table for
4769 * kgnilnd_complete_closed_conn to finish up */
4770 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4773 /* tx was moved to DYING, get next */
4777 spin_unlock(&dev->gnd_lock);
4778 rc = kgnilnd_send_mapped_tx(tx, 1);
4780 /* We made it! skip error handling.. */
4782 /* OK to continue on +ve errors as it won't get seen until
4783 * this function is called again - we operate on a copy of the original
4784 * list and not the live list */
4785 spin_lock(&dev->gnd_lock);
4786 /* reset map attempts back to zero we successfully
4787 * mapped so we can reset our timers */
4788 dev->gnd_map_attempt = 0;
4790 } else if (rc == -EAGAIN) {
4791 spin_lock(&dev->gnd_lock);
4792 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4793 spin_unlock(&dev->gnd_lock);
4794 GOTO(get_out_mapped, rc);
4795 } else if (rc != -ENOMEM) {
4796 /* carp, failure we can't handle */
4797 kgnilnd_tx_done(tx, rc);
4798 spin_lock(&dev->gnd_lock);
4799 /* reset map attempts back to zero we dont know what happened but it
4800 * wasnt a failed mapping
4802 dev->gnd_map_attempt = 0;
4806 /* time to handle the retry cases.. lock so we dont have 2 threads
4807 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4809 spin_lock(&dev->gnd_lock);
4810 dev->gnd_map_attempt++;
4811 if (dev->gnd_map_attempt < fast_remaps) {
4812 /* do nothing we just want it to go as fast as possible.
4813 * just set gnd_next_map to current jiffies so it will process
4814 * as fast as possible.
4816 dev->gnd_next_map = jiffies;
4818 /* Retry based on GNILND_MAP_RETRY_RATE */
4819 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4822 /* only log occasionally once we've retried fast_remaps */
4823 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4824 ((dev->gnd_map_attempt % fast_remaps) == 0);
4825 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4827 /* make sure we are not off in the weeds with this tx */
4828 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4829 GNIDBG_TX(D_NETERROR, tx,
4830 "giving up on TX, too many retries", NULL);
4831 spin_unlock(&dev->gnd_lock);
4832 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4833 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4834 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4836 tx->tx_putinfo.gnpam_dst_cookie,
4837 tx->tx_msg.gnm_srcnid);
4839 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4841 tx->tx_getinfo.gngm_cookie,
4842 tx->tx_msg.gnm_srcnid);
4844 kgnilnd_tx_done(tx, -ENOMEM);
4845 GOTO(get_out_mapped, rc);
4847 GNIDBG_TX(log_retrans_level, tx,
4848 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4850 "nq_map %d mdd# %d/%d GART %ld",
4851 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4852 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4853 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4854 atomic_read(&dev->gnd_nq_map),
4855 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4856 atomic64_read(&dev->gnd_nbytes_map));
4859 /* we need to stop processing the rest of the list, so add it back in */
4860 /* set timer to wake device when we need to schedule this tx */
4861 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4862 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4863 spin_unlock(&dev->gnd_lock);
4864 GOTO(get_out_mapped, rc);
4866 spin_unlock(&dev->gnd_lock);
4872 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4877 int error_inject = 0;
4881 spin_lock(&dev->gnd_lock);
4882 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4883 dev->gnd_sched_alive = jiffies;
4887 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4888 /* break with lock held */
4892 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4893 list_del_init(&conn->gnc_schedlist);
4894 spin_unlock(&dev->gnd_lock);
4896 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4898 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4899 conn_sched != GNILND_CONN_PROCESS,
4900 "conn %p on ready list but in bad state: %d\n",
4903 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4904 conn, kgnilnd_conn_state2str(conn));
4907 set_mb(conn->gnc_last_sched_do, jiffies);
4909 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4911 /* based on intent see if we should run again. */
4912 rc = kgnilnd_schedule_process_conn(conn, intent);
4914 /* drop ref from gnd_ready_conns */
4915 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4916 down_write(&dev->gnd_conn_sem);
4917 kgnilnd_conn_decref(conn);
4918 up_write(&dev->gnd_conn_sem);
4919 } else if (rc != 1) {
4920 kgnilnd_conn_decref(conn);
4922 /* clear this so that scheduler thread doesn't spin */
4924 /* break with lock held... */
4925 spin_lock(&dev->gnd_lock);
4929 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4930 down_write(&dev->gnd_conn_sem);
4932 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4933 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4934 /* If there are tx's currently in use in another
4935 * thread we dont want to complete the close
4936 * yet. Cycle this conn back through
4938 kgnilnd_schedule_conn(conn);
4940 kgnilnd_complete_closed_conn(conn);
4942 up_write(&dev->gnd_conn_sem);
4943 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4944 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4945 /* serialize SMSG CQs with ep_bind and smsg_release */
4946 down_write(&dev->gnd_conn_sem);
4947 kgnilnd_destroy_conn_ep(conn);
4948 up_write(&dev->gnd_conn_sem);
4949 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4950 /* if we need to do some CLOSE sending, etc done here do it */
4951 down_write(&dev->gnd_conn_sem);
4952 kgnilnd_send_conn_close(conn);
4953 kgnilnd_check_fma_rx(conn);
4954 up_write(&dev->gnd_conn_sem);
4955 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4956 /* start moving traffic if the old conns are cleared out */
4957 down_read(&dev->gnd_conn_sem);
4958 kgnilnd_check_fma_rx(conn);
4959 kgnilnd_process_fmaq(conn);
4960 up_read(&dev->gnd_conn_sem);
4963 rc = kgnilnd_schedule_process_conn(conn, 0);
4965 /* drop ref from gnd_ready_conns */
4966 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4967 down_write(&dev->gnd_conn_sem);
4968 kgnilnd_conn_decref(conn);
4969 up_write(&dev->gnd_conn_sem);
4970 } else if (rc != 1) {
4971 kgnilnd_conn_decref(conn);
4974 /* check list again with lock held */
4975 spin_lock(&dev->gnd_lock);
4978 /* If we are short circuiting due to timing we want to be scheduled
4979 * as soon as possible.
4981 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4984 spin_unlock(&dev->gnd_lock);
4990 kgnilnd_scheduler(void *arg)
4992 int threadno = (long)arg;
4995 unsigned long deadline = 0;
4998 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
5000 cfs_block_allsigs();
5002 /* all gnilnd threads need to run fairly urgently */
5003 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
5004 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5005 while (!kgnilnd_data.kgn_shutdown) {
5007 /* Safe: kgn_shutdown only set when quiescent */
5009 /* to quiesce or to not quiesce, that is the question */
5011 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5012 KGNILND_SPIN_QUIESCE;
5015 /* tracking for when thread goes AWOL */
5016 dev->gnd_sched_alive = jiffies;
5018 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5019 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5020 /* let folks know we are up and kicking
5021 * - they can use this for latency savings, etc
5022 * - only change if IRQ, if IDLE leave alone as that
5023 * schedule_device calls to put us back to IRQ */
5024 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5026 down_read(&dev->gnd_conn_sem);
5027 /* always check these - they are super low cost */
5028 found_work += kgnilnd_check_fma_send_cq(dev);
5029 found_work += kgnilnd_check_fma_rcv_cq(dev);
5031 /* rdma CQ doesn't care about eps */
5032 found_work += kgnilnd_check_rdma_cq(dev);
5034 /* move some RDMA ? */
5035 found_work += kgnilnd_process_rdmaq(dev);
5037 /* map some pending RDMA requests ? */
5038 found_work += kgnilnd_process_mapped_tx(dev);
5040 /* the EP for a conn is not destroyed until all the references
5041 * to it are gone, so these checks should be safe
5042 * even if run in parallel with the CQ checking functions
5043 * _AND_ a thread that processes the CLOSED->DONE
5047 up_read(&dev->gnd_conn_sem);
5049 /* process all conns ready now */
5050 found_work += kgnilnd_process_conns(dev, deadline);
5052 /* do an eager check to avoid the IRQ disabling in
5053 * prepare_to_wait and friends */
5056 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5057 time_before(jiffies, deadline)) {
5059 if ((busy_loops % 10) == 0) {
5060 /* tickle heartbeat and watchdog to ensure our
5061 * piggishness doesn't turn into heartbeat failure */
5062 touch_nmi_watchdog();
5068 /* if we got here, found_work was zero or busy_loops means we
5069 * need to take a break. We'll clear gnd_ready but we'll check
5070 * one last time if there is an IRQ that needs processing */
5072 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5074 /* the first time this will go LOOP -> IDLE and let us do one final check
5075 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5076 * - this might allow other threads to block us for a bit if they
5077 * try to get the mutex, but that is good as we'd need to wake
5078 * up soon to handle the CQ or other processing anyways */
5080 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5082 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5083 time_after_eq(jiffies, deadline)) {
5085 "yeilding: found_work %d busy_loops %d\n",
5086 found_work, busy_loops);
5088 /* use yield if we are bailing due to busy_loops
5089 * - this will ensure we wake up soonish. This closes
5090 * a race with kgnilnd_device_callback - where it'd
5091 * not call wake_up() because gnd_ready == 1, but then
5092 * we come down and schedule() because of busy_loops.
5093 * We'd not be woken up until something poked our waitq
5094 * again. yield() ensures we wake up without another
5095 * waitq poke in that case */
5096 atomic_inc(&dev->gnd_n_yield);
5097 kgnilnd_data.kgn_last_condresched = jiffies;
5099 CDEBUG(D_INFO, "awake after yeild\n");
5100 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5101 } else if (found_work == GNILND_DEV_IDLE) {
5102 /* busy_loops is low and there is nothing to do,
5103 * go to sleep and wait for a waitq poke */
5105 "scheduling: found_work %d busy_loops %d\n",
5106 found_work, busy_loops);
5107 atomic_inc(&dev->gnd_n_schedule);
5108 kgnilnd_data.kgn_last_scheduled = jiffies;
5110 CDEBUG(D_INFO, "awake after schedule\n");
5111 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5113 finish_wait(&dev->gnd_waitq, &wait);
5116 kgnilnd_thread_fini();