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 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
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 = kgnilnd_vzalloc(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 kgnilnd_vfree(tx->tx_buffer_copy,
1953 tx->tx_buffer_copy = NULL;
1954 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1955 kgnilnd_tx_done(tx, -EFAULT);
1959 desc_map_key = tx->tx_buffer_copy_map_key;
1960 desc_buffer = tx->tx_buffer_copy;
1964 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1965 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1966 tx->tx_rdma_desc.type = post_type;
1967 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1968 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1969 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1970 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1971 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1972 tx->tx_rdma_desc.length = desc_nob;
1973 tx->tx_nob_rdma = nob;
1974 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1975 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1976 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1977 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1978 /* prep final completion message */
1979 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1980 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1981 /* send actual size RDMA'd in retval */
1982 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1984 kgnilnd_compute_rdma_cksum(tx, nob);
1987 kgnilnd_queue_tx(conn, tx);
1991 /* Don't lie (CLOSE == RDMA idle) */
1992 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1993 tx, conn, conn->gnc_close_sent);
1995 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1996 "0x%x cookie:%#llx",
1997 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1999 /* set CQ dedicated for RDMA */
2000 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
2002 timestamp = jiffies;
2003 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
2004 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2005 /* delay in jiffies - we are really concerned only with things that
2006 * result in a schedule() or really holding this off for long times .
2007 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2008 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2010 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2012 if (rrc == GNI_RC_ERROR_RESOURCE) {
2013 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2014 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2015 kgnilnd_unmap_buffer(tx, 0);
2017 if (tx->tx_buffer_copy != NULL) {
2018 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
2019 tx->tx_buffer_copy = NULL;
2022 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2023 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2025 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2026 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2030 spin_lock(&conn->gnc_list_lock);
2031 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2032 tx->tx_qtime = jiffies;
2033 spin_unlock(&conn->gnc_list_lock);
2034 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2035 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2037 /* XXX Nic: is this a place we should handle more errors for
2038 * robustness sake */
2039 LASSERT(rrc == GNI_RC_SUCCESS);
2044 kgnilnd_alloc_rx(void)
2048 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2050 CERROR("failed to allocate rx\n");
2053 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2056 /* no memset to zero, we'll always fill all members */
2060 /* release is to just free connection resources
2061 * we use this for the eager path after copying */
2063 kgnilnd_release_msg(kgn_conn_t *conn)
2066 unsigned long timestamp;
2068 CDEBUG(D_NET, "consuming %p\n", conn);
2070 timestamp = jiffies;
2071 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2072 /* delay in jiffies - we are really concerned only with things that
2073 * result in a schedule() or really holding this off for long times .
2074 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2075 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2077 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2078 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2080 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2081 GNIDBG_SMSG_CREDS(D_NET, conn);
2087 kgnilnd_consume_rx(kgn_rx_t *rx)
2089 kgn_conn_t *conn = rx->grx_conn;
2090 kgn_msg_t *rxmsg = rx->grx_msg;
2092 /* if we are eager, free the cache alloc'd msg */
2093 if (unlikely(rx->grx_eager)) {
2094 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2095 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2097 /* release ref from eager_recv */
2098 kgnilnd_conn_decref(conn);
2100 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2101 kgnilnd_release_msg(conn);
2104 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2105 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2112 kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
2114 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
2115 int type = lntmsg->msg_type;
2116 lnet_process_id_t target = lntmsg->msg_target;
2117 int target_is_router = lntmsg->msg_target_is_router;
2118 int routing = lntmsg->msg_routing;
2119 unsigned int niov = lntmsg->msg_niov;
2120 struct kvec *iov = lntmsg->msg_iov;
2121 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2122 unsigned int offset = lntmsg->msg_offset;
2123 unsigned int nob = lntmsg->msg_len;
2124 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2125 kgn_net_t *net = ni->ni_data;
2129 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2131 /* NB 'private' is different depending on what we're sending.... */
2132 LASSERT(!in_interrupt());
2134 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2135 type, nob, niov, libcfs_id2str(target));
2137 LASSERTF(nob == 0 || niov > 0,
2138 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2139 LASSERTF(niov <= LNET_MAX_IOV,
2140 "lntmsg %p niov %d\n", lntmsg, niov);
2142 /* payload is either all vaddrs or all pages */
2143 LASSERTF(!(kiov != NULL && iov != NULL),
2144 "lntmsg %p kiov %p iov %p\n", lntmsg, kiov, iov);
2147 mpflag = cfs_memory_pressure_get_and_set();
2151 CERROR("lntmsg %p with unexpected type %d\n",
2156 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2164 if (routing || target_is_router)
2165 break; /* send IMMEDIATE */
2167 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2168 * check the eventual sink buffer here - if small enough, remote
2169 * end is perfectly capable of returning data in short message -
2170 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2171 * for IMMEDIATE messages which will have it send a real reply instead
2172 * of doing kgnilnd_recv to have the RDMA continued */
2173 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2176 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2177 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2179 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2185 /* slightly different options as we might actually have a GET with a
2186 * MD_KIOV set but a non-NULL md_iov.iov */
2187 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
2188 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2189 lntmsg->msg_md->md_iov.iov, NULL,
2190 0, lntmsg->msg_md->md_length);
2192 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2193 NULL, lntmsg->msg_md->md_iov.kiov,
2194 0, lntmsg->msg_md->md_length);
2196 CERROR("unable to setup buffer: %d\n", rc);
2197 kgnilnd_tx_done(tx, rc);
2202 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2203 if (tx->tx_lntmsg[1] == NULL) {
2204 CERROR("Can't create reply for GET to %s\n",
2205 libcfs_nid2str(target.nid));
2206 kgnilnd_tx_done(tx, rc);
2211 tx->tx_lntmsg[0] = lntmsg;
2212 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2213 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2215 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2217 /* rest of tx_msg is setup just before it is sent */
2218 kgnilnd_launch_tx(tx, net, &target);
2220 case LNET_MSG_REPLY:
2222 /* to save on MDDs, we'll handle short kiov by vmap'ing
2223 * and sending via SMSG */
2224 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2227 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2228 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2230 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2237 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2239 kgnilnd_tx_done(tx, rc);
2244 tx->tx_lntmsg[0] = lntmsg;
2245 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2246 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2248 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2250 /* rest of tx_msg is setup just before it is sent */
2251 kgnilnd_launch_tx(tx, net, &target);
2255 /* send IMMEDIATE */
2257 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2258 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2260 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2266 rc = kgnilnd_setup_immediate_buffer(tx, niov, iov, kiov, offset, nob);
2268 kgnilnd_tx_done(tx, rc);
2272 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2273 tx->tx_lntmsg[0] = lntmsg;
2274 kgnilnd_launch_tx(tx, net, &target);
2277 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2279 cfs_memory_pressure_restore(mpflag);
2284 kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen)
2286 kgn_conn_t *conn = rx->grx_conn;
2287 kgn_msg_t *rxmsg = rx->grx_msg;
2288 unsigned int niov = lntmsg->msg_niov;
2289 struct kvec *iov = lntmsg->msg_iov;
2290 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2291 unsigned int offset = lntmsg->msg_offset;
2292 unsigned int nob = lntmsg->msg_len;
2297 switch (rxmsg->gnm_type) {
2298 case GNILND_MSG_PUT_REQ_REV:
2299 done_type = GNILND_MSG_PUT_DONE_REV;
2302 case GNILND_MSG_GET_REQ:
2303 done_type = GNILND_MSG_GET_DONE;
2306 CERROR("invalid msg type %s (%d)\n",
2307 kgnilnd_msgtype2str(rxmsg->gnm_type),
2312 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2316 rc = kgnilnd_set_tx_id(tx, conn);
2320 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2324 tx->tx_lntmsg[0] = lntmsg;
2325 tx->tx_getinfo = rxmsg->gnm_u.get;
2327 /* we only queue from kgnilnd_recv - we might get called from other contexts
2328 * and we don't want to block the mutex in those cases */
2330 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2331 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2332 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2333 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2338 kgnilnd_tx_done(tx, rc);
2339 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2341 lnet_finalize(ni, lntmsg, rc);
2345 kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
2348 kgn_rx_t *rx = private;
2349 kgn_conn_t *conn = rx->grx_conn;
2350 kgn_msg_t *rxmsg = rx->grx_msg;
2351 kgn_msg_t *eagermsg = NULL;
2352 kgn_peer_t *peer = NULL;
2353 kgn_conn_t *found_conn = NULL;
2355 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2356 conn, rxmsg, lntmsg);
2358 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2359 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2360 rxmsg->gnm_payload_len);
2363 /* Grab a read lock so the connection doesnt disappear on us
2364 * while we look it up
2366 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2368 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2370 found_conn = kgnilnd_find_conn_locked(peer);
2373 /* Verify the connection found is the same one that the message
2374 * is supposed to be using, if it is not output an error message
2377 if (!peer || !found_conn
2378 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2379 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2380 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2381 peer, conn, found_conn);
2383 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2384 " from %s", rxmsg->gnm_connstamp,
2385 found_conn->gnc_peer_connstamp,
2386 libcfs_nid2str(peer->gnp_nid));
2391 /* add conn ref to ensure it doesn't go away until all eager
2392 * messages processed */
2393 kgnilnd_conn_addref(conn);
2395 /* Now that we have verified the connection is valid and added a
2396 * reference we can remove the read_lock on the peer_conn_lock */
2397 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2399 /* we have no credits or buffers for this message, so copy it
2400 * somewhere for a later kgnilnd_recv */
2401 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2402 *kgnilnd_tunables.kgn_eager_credits) {
2403 CERROR("Out of eager credits to %s\n",
2404 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2408 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2410 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2411 if (eagermsg == NULL) {
2412 kgnilnd_conn_decref(conn);
2413 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2414 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2418 /* copy msg and payload */
2419 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2420 rx->grx_msg = eagermsg;
2423 /* stash this for lnet_finalize on cancel-on-conn-close */
2424 rx->grx_lntmsg = lntmsg;
2426 /* keep the same rx_t, it just has a new grx_msg now */
2427 *new_private = private;
2429 /* release SMSG buffer */
2430 kgnilnd_release_msg(conn);
2436 kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
2437 int delayed, unsigned int niov,
2438 struct kvec *iov, lnet_kiov_t *kiov,
2439 unsigned int offset, unsigned int mlen, unsigned int rlen)
2441 kgn_rx_t *rx = private;
2442 kgn_conn_t *conn = rx->grx_conn;
2443 kgn_msg_t *rxmsg = rx->grx_msg;
2449 LASSERT(!in_interrupt());
2450 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2451 /* Either all pages or all vaddrs */
2452 LASSERTF(!(kiov != NULL && iov != NULL), "kiov %p iov %p\n",
2455 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2456 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2457 conn, rxmsg, lntmsg,
2458 niov, kiov, iov, offset, mlen, rlen);
2460 /* we need to lock here as recv can be called from any context */
2461 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2462 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2463 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2465 /* someone closed the conn after we copied this out, nuke it */
2466 kgnilnd_consume_rx(rx);
2467 lnet_finalize(ni, lntmsg, conn->gnc_error);
2470 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2472 switch (rxmsg->gnm_type) {
2474 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2475 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2476 conn, rx, rxmsg, lntmsg, niov, kiov, iov, offset, mlen, rlen);
2479 case GNILND_MSG_IMMEDIATE:
2480 if (mlen > rxmsg->gnm_payload_len) {
2481 GNIDBG_MSG(D_ERROR, rxmsg,
2482 "Immediate message from %s too big: %d > %d",
2483 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2484 rxmsg->gnm_payload_len);
2486 kgnilnd_consume_rx(rx);
2490 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2491 * right after the kgn_msg_t header - so just 'cute' way of saying
2492 * rxmsg + sizeof(kgn_msg_t) */
2494 /* check payload checksum if sent */
2496 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2497 !rxmsg->gnm_payload_cksum &&
2498 rxmsg->gnm_payload_len != 0)
2499 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2501 if (rxmsg->gnm_payload_cksum != 0) {
2502 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2503 * which is what is used to calculate the cksum on the TX side */
2504 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2506 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2507 GNIDBG_MSG(D_NETERROR, rxmsg,
2508 "Bad payload checksum (%x expected %x)",
2509 pload_cksum, rxmsg->gnm_payload_cksum);
2510 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2512 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2513 &rxmsg[1], rxmsg->gnm_payload_len);
2514 /* fall through to dump */
2516 libcfs_debug_dumplog();
2522 /* checksum problems are fatal, kill the conn */
2523 kgnilnd_consume_rx(rx);
2524 kgnilnd_close_conn(conn, rc);
2530 lnet_copy_flat2kiov(
2532 *kgnilnd_tunables.kgn_max_immediate,
2533 &rxmsg[1], 0, mlen);
2537 *kgnilnd_tunables.kgn_max_immediate,
2538 &rxmsg[1], 0, mlen);
2540 kgnilnd_consume_rx(rx);
2541 lnet_finalize(ni, lntmsg, 0);
2544 case GNILND_MSG_PUT_REQ:
2545 /* LNET wants to truncate or drop transaction, sending NAK */
2547 kgnilnd_consume_rx(rx);
2548 lnet_finalize(ni, lntmsg, 0);
2550 /* only error if lntmsg == NULL, otherwise we are just
2551 * short circuiting the rdma process of 0 bytes */
2552 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2553 lntmsg == NULL ? -ENOENT : 0,
2554 rxmsg->gnm_u.get.gngm_cookie,
2558 /* sending ACK with sink buff. info */
2559 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2561 kgnilnd_consume_rx(rx);
2565 rc = kgnilnd_set_tx_id(tx, conn);
2567 GOTO(nak_put_req, rc);
2570 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2572 GOTO(nak_put_req, rc);
2575 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2576 rxmsg->gnm_u.putreq.gnprm_cookie;
2577 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2578 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2579 (__u64)((unsigned long)tx->tx_buffer);
2580 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2582 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2583 tx->tx_qtime = jiffies;
2584 /* we only queue from kgnilnd_recv - we might get called from other contexts
2585 * and we don't want to block the mutex in those cases */
2587 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2588 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2589 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2590 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2592 kgnilnd_consume_rx(rx);
2596 /* make sure we send an error back when the PUT fails */
2597 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2598 kgnilnd_tx_done(tx, rc);
2599 kgnilnd_consume_rx(rx);
2601 /* return magic LNet network error */
2603 case GNILND_MSG_GET_REQ_REV:
2604 /* LNET wants to truncate or drop transaction, sending NAK */
2606 kgnilnd_consume_rx(rx);
2607 lnet_finalize(ni, lntmsg, 0);
2609 /* only error if lntmsg == NULL, otherwise we are just
2610 * short circuiting the rdma process of 0 bytes */
2611 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2612 lntmsg == NULL ? -ENOENT : 0,
2613 rxmsg->gnm_u.get.gngm_cookie,
2617 /* lntmsg can be null when parsing a LNET_GET */
2618 if (lntmsg != NULL) {
2619 /* sending ACK with sink buff. info */
2620 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2622 kgnilnd_consume_rx(rx);
2626 rc = kgnilnd_set_tx_id(tx, conn);
2628 GOTO(nak_get_req_rev, rc);
2631 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2633 GOTO(nak_get_req_rev, rc);
2636 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2637 rxmsg->gnm_u.putreq.gnprm_cookie;
2638 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2639 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2640 (__u64)((unsigned long)tx->tx_buffer);
2641 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2643 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2645 /* we only queue from kgnilnd_recv - we might get called from other contexts
2646 * and we don't want to block the mutex in those cases */
2648 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2649 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2650 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2651 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2654 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2656 rxmsg->gnm_u.get.gngm_cookie,
2660 kgnilnd_consume_rx(rx);
2664 /* make sure we send an error back when the GET fails */
2665 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2666 kgnilnd_tx_done(tx, rc);
2667 kgnilnd_consume_rx(rx);
2669 /* return magic LNet network error */
2673 case GNILND_MSG_PUT_REQ_REV:
2674 /* LNET wants to truncate or drop transaction, sending NAK */
2676 kgnilnd_consume_rx(rx);
2677 lnet_finalize(ni, lntmsg, 0);
2679 /* only error if lntmsg == NULL, otherwise we are just
2680 * short circuiting the rdma process of 0 bytes */
2681 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2682 lntmsg == NULL ? -ENOENT : 0,
2683 rxmsg->gnm_u.get.gngm_cookie,
2688 if (lntmsg != NULL) {
2690 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2693 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2695 rxmsg->gnm_u.get.gngm_cookie,
2698 kgnilnd_consume_rx(rx);
2700 case GNILND_MSG_GET_REQ:
2701 if (lntmsg != NULL) {
2703 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2706 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2708 rxmsg->gnm_u.get.gngm_cookie,
2711 kgnilnd_consume_rx(rx);
2717 /* needs write_lock on kgn_peer_conn_lock held */
2719 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2721 unsigned long timeout, keepalive;
2722 unsigned long now = jiffies;
2723 unsigned long newest_last_rx;
2726 /* given that we found this conn hanging off a peer, it better damned
2727 * well be connected */
2728 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2729 "conn 0x%p->%s with bad state%s\n", conn,
2730 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2732 kgnilnd_conn_state2str(conn));
2734 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2735 "rx_diff %lu tx_diff %lu\n",
2736 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2737 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2738 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2739 cfs_duration_sec(now - conn->gnc_last_tx));
2741 timeout = cfs_time_seconds(conn->gnc_timeout);
2742 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2744 /* just in case our lack of RX msg processing is gumming up the works - give the
2745 * remove an extra chance */
2747 newest_last_rx = GNILND_LASTRX(conn);
2749 if (time_after_eq(now, newest_last_rx + timeout)) {
2750 uint32_t level = D_CONSOLE|D_NETERROR;
2752 if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_DOWN) {
2755 GNIDBG_CONN(level, conn,
2756 "No gnilnd traffic received from %s for %lu "
2757 "seconds, terminating connection. Is node down? ",
2758 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2759 cfs_duration_sec(now - newest_last_rx));
2763 /* we don't timeout on last_tx stalls - we are going to trust the
2764 * underlying network to let us know when sends are failing.
2765 * At worst, the peer will timeout our RX stamp and drop the connection
2766 * at that point. We'll then see his CLOSE or at worst his RX
2767 * stamp stop and drop the connection on our end */
2769 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2770 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2771 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2772 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2773 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2775 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2776 conn->gnc_last_noop_cq,
2777 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2778 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2779 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2780 set_mb(conn->gnc_last_noop_want, jiffies);
2781 atomic_inc(&conn->gnc_reaper_noop);
2782 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2785 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2788 kgnilnd_queue_tx(conn, tx);
2794 /* needs write_lock on kgn_peer_conn_lock held */
2796 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2797 struct list_head *souls)
2799 unsigned long timeout;
2800 kgn_conn_t *conn, *connN = NULL;
2806 short releaseconn = 0;
2807 unsigned long first_rx = 0;
2808 int purgatory_conn_cnt = 0;
2810 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2811 peer, libcfs_nid2str(peer->gnp_nid),
2812 peer->gnp_reconnect_interval);
2814 timeout = cfs_time_seconds(MAX(*kgnilnd_tunables.kgn_timeout,
2815 GNILND_MIN_TIMEOUT));
2817 conn = kgnilnd_find_conn_locked(peer);
2819 /* if there is a valid conn, check the queues for timeouts */
2820 rc = kgnilnd_check_conn_timeouts_locked(conn);
2822 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2823 /* simulate a RX CLOSE after the timeout but before
2824 * the scheduler thread gets it */
2825 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2826 conn->gnc_peer_error = -ETIMEDOUT;
2828 /* Once we mark closed, any of the scheduler threads could
2829 * get it and move through before we hit the fail loc code */
2830 kgnilnd_close_conn_locked(conn, rc);
2832 /* first_rx is used to decide when to release a conn from purgatory.
2834 first_rx = conn->gnc_first_rx;
2838 /* now regardless of starting new conn, find tx on peer queue that
2839 * are old and smell bad - do this first so we don't trigger
2840 * reconnect on empty queue if we timeout all */
2841 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2842 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2844 LCONSOLE_INFO("could not send to %s due to connection"
2845 " setup failure after %lu seconds\n",
2846 libcfs_nid2str(peer->gnp_nid),
2847 cfs_duration_sec(jiffies - tx->tx_qtime));
2849 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2851 list_add_tail(&tx->tx_list, todie);
2856 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2857 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2858 count, peer, libcfs_nid2str(peer->gnp_nid));
2859 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2860 if (list_empty(&peer->gnp_tx_queue) ||
2861 peer->gnp_connecting == GNILND_PEER_KILL) {
2862 /* we pass down todie to use a common function - but we know there are
2864 kgnilnd_cancel_peer_connect_locked(peer, todie);
2868 /* Don't reconnect if we are still trying to clear out old conns.
2869 * This prevents us sending traffic on the new mbox before ensuring we are done
2870 * with the old one */
2871 reconnect = (peer->gnp_down == GNILND_RCA_NODE_UP) &&
2872 (atomic_read(&peer->gnp_dirty_eps) == 0);
2874 /* fast reconnect after a timeout */
2875 to_reconn = !conn &&
2876 (peer->gnp_last_errno == -ETIMEDOUT) &&
2877 *kgnilnd_tunables.kgn_fast_reconn;
2879 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2880 * to be sent, we'll check the reconnect interval and fire up a new
2881 * connection request */
2884 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2885 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2886 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2888 CDEBUG(D_NET, "starting connect to %s\n",
2889 libcfs_nid2str(peer->gnp_nid));
2890 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE, "Peer was idle and we"
2891 "have a write_lock, state issue %d\n", peer->gnp_connecting);
2893 peer->gnp_connecting = GNILND_PEER_CONNECT;
2894 kgnilnd_peer_addref(peer); /* extra ref for connd */
2896 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2897 list_add_tail(&peer->gnp_connd_list,
2898 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2899 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2901 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2904 /* fail_loc to allow us to delay release of purgatory */
2905 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2908 /* This check allows us to verify that the new conn is actually being used. This allows us to
2909 * pull the old conns out of purgatory if they have actually seen traffic.
2910 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2913 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2914 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2915 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2919 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2920 /* check for purgatory timeouts */
2921 if (conn->gnc_in_purgatory) {
2922 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2923 * that way the next time we check it we can detach it from purgatory
2926 if (conn->gnc_state != GNILND_CONN_DONE) {
2927 /* Skip over conns that are currently not DONE. If they arent already scheduled
2928 * for completion something in the state machine is broken.
2933 /* We only detach a conn that is in purgatory if we have received a close message,
2934 * we have a new valid connection that has successfully received data, or an admin
2935 * command tells us we need to detach.
2938 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2939 unsigned long waiting;
2941 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2943 /* C.E: The remote peer is expected to close the
2944 * connection (see kgnilnd_check_conn_timeouts)
2945 * via the reaper thread and nuke out the MDD and
2946 * FMA resources after conn->gnc_timeout has expired
2947 * without an FMA RX */
2948 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2949 " held resources\n",
2950 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2951 cfs_duration_sec(waiting));
2953 kgnilnd_detach_purgatory_locked(conn, souls);
2955 purgatory_conn_cnt++;
2960 /* If we have too many connections in purgatory we could run out of
2961 * resources. Limit the number of connections to a tunable number,
2962 * clean up to the minimum all in one fell swoop... there are
2963 * situations where dvs will retry tx's and we can eat up several
2964 * hundread connection requests at once.
2966 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2967 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2969 if (conn->gnc_in_purgatory &&
2970 conn->gnc_state == GNILND_CONN_DONE) {
2971 CDEBUG(D_NET, "Dropping Held resource due to"
2972 " resource limits being hit\n");
2973 kgnilnd_detach_purgatory_locked(conn, souls);
2975 if (purgatory_conn_cnt-- <
2976 *kgnilnd_tunables.kgn_max_purgatory)
2986 kgnilnd_reaper_check(int idx)
2988 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2989 struct list_head *ctmp, *ctmpN;
2990 struct list_head geriatrics;
2991 struct list_head souls;
2993 INIT_LIST_HEAD(&geriatrics);
2994 INIT_LIST_HEAD(&souls);
2996 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2998 list_for_each_safe(ctmp, ctmpN, peers) {
2999 kgn_peer_t *peer = NULL;
3001 /* don't timeout stuff if the network is mucked or shutting down */
3002 if (kgnilnd_check_hw_quiesce()) {
3005 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
3007 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
3010 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3012 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3013 kgnilnd_release_purgatory_list(&souls);
3017 kgnilnd_update_reaper_timeout(long timeout)
3019 LASSERT(timeout > 0);
3021 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3023 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3024 kgnilnd_data.kgn_new_min_timeout = timeout;
3026 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3030 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3032 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3036 kgnilnd_reaper(void *arg)
3041 unsigned long next_check_time = jiffies;
3042 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3043 struct timer_list timer;
3046 cfs_block_allsigs();
3048 /* all gnilnd threads need to run fairly urgently */
3049 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3050 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3052 while (!kgnilnd_data.kgn_shutdown) {
3053 /* I wake up every 'p' seconds to check for timeouts on some
3054 * more peers. I try to check every connection 'n' times
3055 * within the global minimum of all keepalive and timeout
3056 * intervals, to ensure I attend to every connection within
3057 * (n+1)/n times its timeout intervals. */
3058 const int p = GNILND_REAPER_THREAD_WAKE;
3059 const int n = GNILND_REAPER_NCHECKS;
3061 /* to quiesce or to not quiesce, that is the question */
3062 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3063 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3064 KGNILND_SPIN_QUIESCE;
3065 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3068 /* careful with the jiffy wrap... */
3069 timeout = (long)(next_check_time - jiffies);
3072 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3073 TASK_INTERRUPTIBLE);
3074 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3075 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3077 mod_timer(&timer, (long) jiffies + timeout);
3079 /* check flag variables before committing */
3080 if (!kgnilnd_data.kgn_shutdown &&
3081 !kgnilnd_data.kgn_quiesce_trigger) {
3082 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3083 timeout, cfs_duration_sec(timeout));
3085 CDEBUG(D_INFO, "awake after schedule\n");
3088 del_singleshot_timer_sync(&timer);
3089 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3090 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3094 /* new_min_timeout is set from the conn timeouts and keepalive
3095 * this should end up with a min timeout of
3096 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3097 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3098 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3099 CDEBUG(D_NET, "Set new min timeout %ld\n",
3100 current_min_timeout);
3103 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3105 /* Compute how many table entries to check now so I get round
3106 * the whole table fast enough given that I do this at fixed
3107 * intervals of 'p' seconds) */
3108 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3109 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3110 chunk = (chunk * n * p) /
3111 kgnilnd_data.kgn_new_min_timeout;
3114 for (i = 0; i < chunk; i++) {
3115 kgnilnd_reaper_check(hash_index);
3116 hash_index = (hash_index + 1) %
3117 *kgnilnd_tunables.kgn_peer_hash_size;
3119 next_check_time = (long) jiffies + cfs_time_seconds(p);
3120 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3122 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3125 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3127 kgnilnd_thread_fini();
3132 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3133 unsigned niov, offset, nob;
3135 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
3136 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3139 lnet_copy_flat2kiov(
3142 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3144 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3151 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3154 gni_post_descriptor_t *desc;
3156 kgn_tx_ev_id_t ev_id;
3158 int should_retry, rc;
3159 long num_processed = 0;
3160 kgn_conn_t *conn = NULL;
3161 kgn_tx_t *tx = NULL;
3162 kgn_rdma_desc_t *rdesc;
3167 /* make sure we don't keep looping if we need to reset */
3168 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3169 return num_processed;
3171 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3173 /* we didn't get the mutex, so return that there is still work
3177 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3178 /* a bit gross - but we need a good way to test for
3179 * delayed RDMA completions and the easiest way to do
3180 * that is to delay the RDMA CQ events */
3181 rrc = GNI_RC_NOT_DONE;
3183 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3186 if (rrc == GNI_RC_NOT_DONE) {
3187 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3188 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3189 dev->gnd_id, num_processed);
3190 return num_processed;
3192 dev->gnd_sched_alive = jiffies;
3195 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3196 "this is bad, somehow our credits didn't protect us"
3197 " from CQ overrun\n");
3198 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3199 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3200 event_data, GNI_CQ_GET_TYPE(event_data));
3202 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3204 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3206 /* XXX Nic: Need better error handling here... */
3207 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3208 (rrc == GNI_RC_TRANSACTION_ERROR),
3211 ev_id.txe_cookie = desc->post_id;
3213 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3215 if (conn == NULL || tx == NULL) {
3216 /* either conn or tx was already nuked and this is a "late"
3217 * completion, so drop it */
3221 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3222 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3223 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3224 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3225 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3227 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3229 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3230 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3231 tx->tx_msg.gnm_u.completion.gncm_retval);
3235 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3236 if (tx->tx_buffer_copy != NULL)
3237 kgnilnd_recv_bte_get(tx);
3238 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3241 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3242 if (tx->tx_buffer_copy != NULL)
3243 kgnilnd_recv_bte_get(tx);
3244 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3247 /* remove from rdmaq */
3248 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3249 spin_lock(&conn->gnc_list_lock);
3250 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3251 spin_unlock(&conn->gnc_list_lock);
3252 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3254 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3255 event_data = 1LL << 48;
3259 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3260 atomic_inc(&dev->gnd_rdma_ntx);
3261 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3262 /* transaction succeeded, add into fmaq */
3263 kgnilnd_queue_tx(conn, tx);
3264 kgnilnd_peer_alive(conn->gnc_peer);
3266 /* drop ref from kgnilnd_validate_tx_ev_id */
3267 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3268 kgnilnd_conn_decref(conn);
3273 /* fall through to the TRANSACTION_ERROR case */
3276 /* get stringified version for log messages */
3277 kgnilnd_cq_error_str(event_data, &err_str, 256);
3278 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3280 /* make sure we are not off in the weeds with this tx */
3281 if (tx->tx_retrans >
3282 *kgnilnd_tunables.kgn_max_retransmits) {
3283 GNIDBG_TX(D_NETERROR, tx,
3284 "giving up on TX, too many retries", NULL);
3288 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3289 should_retry ? "transient" : "unrecoverable", err_str);
3291 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3292 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3293 rdesc = &tx->tx_putinfo.gnpam_desc;
3294 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3295 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3297 rdesc = &tx->tx_getinfo.gngm_desc;
3298 rnob = tx->tx_lntmsg[0]->msg_len;
3299 rcookie = tx->tx_getinfo.gngm_cookie;
3304 tx->tx_msg.gnm_type,
3308 kgnilnd_nak_rdma(conn,
3309 tx->tx_msg.gnm_type,
3312 tx->tx_msg.gnm_srcnid);
3313 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3314 kgnilnd_close_conn(conn, -ECOMM);
3317 /* drop ref from kgnilnd_validate_tx_ev_id */
3318 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3319 kgnilnd_conn_decref(conn);
3324 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3328 kgn_tx_ev_id_t ev_id;
3329 kgn_tx_t *tx = NULL;
3330 kgn_conn_t *conn = NULL;
3331 int queued_fma, saw_reply, rc;
3332 long num_processed = 0;
3335 /* make sure we don't keep looping if we need to reset */
3336 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3337 return num_processed;
3340 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3342 /* we didn't get the mutex, so return that there is still work
3347 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3348 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3350 if (rrc == GNI_RC_NOT_DONE) {
3352 "SMSG send CQ %d not ready (data %#llx) "
3353 "processed %ld\n", dev->gnd_id, event_data,
3355 return num_processed;
3358 dev->gnd_sched_alive = jiffies;
3361 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3362 "this is bad, somehow our credits didn't "
3363 "protect us from CQ overrun\n");
3364 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3365 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3366 event_data, GNI_CQ_GET_TYPE(event_data));
3368 /* if SMSG couldn't handle an error, time for conn to die */
3369 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3372 /* need to take the write_lock to ensure atomicity
3373 * on the conn state if we need to close it */
3374 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3375 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3377 /* Conn was destroyed? */
3379 "SMSG CQID lookup %#llx failed\n",
3380 GNI_CQ_GET_INST_ID(event_data));
3381 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3385 kgnilnd_cq_error_str(event_data, &err_str, 256);
3386 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3387 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3389 kgnilnd_close_conn_locked(conn, -ECOMM);
3391 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3393 /* no need to process rest of this tx -
3394 * it is getting canceled */
3398 /* fall through to GNI_RC_SUCCESS case */
3399 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3401 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3402 if (conn == NULL || tx == NULL) {
3403 /* either conn or tx was already nuked and this is a "late"
3404 * completion, so drop it */
3408 tx->tx_conn->gnc_last_tx_cq = jiffies;
3409 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3410 set_mb(conn->gnc_last_noop_cq, jiffies);
3413 /* lock tx_list_state and tx_state */
3414 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3415 spin_lock(&tx->tx_conn->gnc_list_lock);
3417 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3418 "state not GNILND_TX_LIVE_FMAQ", NULL);
3419 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3420 "not waiting for completion", NULL);
3422 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3425 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3427 /* This will trigger other FMA sends that were
3428 * pending this completion */
3429 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3431 /* we either did not expect reply or we already got it */
3432 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3434 spin_unlock(&tx->tx_conn->gnc_list_lock);
3435 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3438 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3440 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3441 kgnilnd_schedule_conn(conn);
3444 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3445 * If saw_reply is true we know that the tx is safe to use as the other thread
3446 * is already finished with it.
3450 /* no longer need to track on the live_fmaq */
3451 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3453 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3454 /* we already got reply & were waiting for
3455 * completion of initial send */
3456 /* to initiate RDMA transaction */
3457 GNIDBG_TX(D_NET, tx,
3458 "Pending RDMA 0x%p type 0x%02x",
3459 tx->tx_msg.gnm_type);
3460 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3461 rc = kgnilnd_send_mapped_tx(tx, 0);
3462 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3464 /* we are done with this tx */
3465 GNIDBG_TX(D_NET, tx,
3466 "Done with tx type 0x%02x",
3467 tx->tx_msg.gnm_type);
3468 kgnilnd_tx_done(tx, tx->tx_rc);
3472 /* drop ref from kgnilnd_validate_tx_ev_id */
3473 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3474 kgnilnd_conn_decref(conn);
3476 /* if we are waiting for a REPLY, we'll handle the tx then */
3477 } /* end for loop */
3481 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3486 long num_processed = 0;
3487 struct list_head *conns;
3488 struct list_head *tmp;
3492 /* make sure we don't keep looping if we need to reset */
3493 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3494 return num_processed;
3497 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3499 /* we didn't get the mutex, so return that there is still work
3503 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3504 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3506 if (rrc == GNI_RC_NOT_DONE) {
3507 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3509 dev->gnd_id, event_data, num_processed);
3510 return num_processed;
3512 dev->gnd_sched_alive = jiffies;
3515 /* this is the only CQ that can really handle transient
3517 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3518 rrc = cfs_fail_val ? cfs_fail_val
3519 : GNI_RC_ERROR_RESOURCE;
3520 if (rrc == GNI_RC_ERROR_RESOURCE) {
3521 /* set overrun too */
3522 event_data |= (1UL << 63);
3523 LASSERTF(GNI_CQ_OVERRUN(event_data),
3524 "(1UL << 63) is no longer the bit to"
3525 "set to indicate CQ_OVERRUN\n");
3528 /* sender should get error event too and take care
3529 of failed transaction by re-transmitting */
3530 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3531 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3535 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3536 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3537 conn = kgnilnd_cqid2conn_locked(
3538 GNI_CQ_GET_INST_ID(event_data));
3540 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3541 "failed, dropping event %#llx\n",
3542 GNI_CQ_GET_INST_ID(event_data),
3545 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3547 GNI_CQ_GET_INST_ID(event_data),
3548 conn, conn->gnc_peer ?
3549 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3552 conn->gnc_last_rx_cq = jiffies;
3554 /* stash first rx so we can clear out purgatory.
3556 if (conn->gnc_first_rx == 0) {
3557 conn->gnc_first_rx = jiffies;
3559 kgnilnd_peer_alive(conn->gnc_peer);
3560 kgnilnd_schedule_conn(conn);
3562 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3566 /* FMA CQ has overflowed: check ALL conns */
3567 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3568 "conns on device %d\n", dev->gnd_id);
3570 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3572 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3573 conns = &kgnilnd_data.kgn_conns[rc];
3575 list_for_each(tmp, conns) {
3576 conn = list_entry(tmp, kgn_conn_t,
3579 if (conn->gnc_device == dev) {
3580 kgnilnd_schedule_conn(conn);
3581 conn->gnc_last_rx_cq = jiffies;
3585 /* don't block write lockers for too long... */
3586 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3591 /* try_map_if_full should only be used when processing TX from list of
3592 * backlog TX waiting on mappings to free up
3595 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3596 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3599 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3601 /* slight bit of race if multiple people calling, but at worst we'll have
3602 * order altered just a bit... which would not be determenistic anyways */
3603 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3605 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3607 /* We know that we have a GART reservation that should guarantee forward progress.
3608 * This means we don't need to take any extraordinary efforts if we are failing
3609 * mappings here - even if we are holding a very small number of these. */
3611 if (try_map_if_full || (rc == 0)) {
3612 rc = kgnilnd_map_buffer(tx);
3615 /* rc should be 0 if we mapped successfully here, if non-zero
3616 * we are queueing */
3618 /* if try_map_if_full set, they handle requeuing */
3619 if (unlikely(try_map_if_full)) {
3622 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3623 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3624 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3625 /* make sure we wake up sched to run this */
3626 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3627 /* return 0 as this is now queued for later sending */
3632 switch (tx->tx_msg.gnm_type) {
3636 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3637 * remote node where the RDMA will be started
3638 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3639 * be satisified. The rest of the errors are "hard" errors that require
3640 * upper layers to handle themselves.
3641 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3642 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3643 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3644 * GNILND_MSG_GET_DONE_REV.
3646 case GNILND_MSG_GET_REQ:
3647 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3648 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3649 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3650 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3651 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3652 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3653 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3655 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3656 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3658 case GNILND_MSG_PUT_ACK:
3659 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3660 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3661 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3662 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3664 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3665 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3668 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3669 case GNILND_MSG_PUT_DONE:
3670 case GNILND_MSG_PUT_REQ:
3671 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3672 &tx->tx_putinfo.gnpam_desc,
3673 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3674 tx->tx_putinfo.gnpam_dst_cookie);
3675 RETURN(try_map_if_full ? rc : 0);
3677 case GNILND_MSG_GET_DONE:
3678 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3679 &tx->tx_getinfo.gngm_desc,
3680 tx->tx_lntmsg[0]->msg_len,
3681 tx->tx_getinfo.gngm_cookie);
3682 RETURN(try_map_if_full ? rc : 0);
3684 case GNILND_MSG_PUT_REQ_REV:
3685 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3686 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3687 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3688 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3689 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3690 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3691 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3693 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3695 case GNILND_MSG_PUT_DONE_REV:
3696 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3697 &tx->tx_getinfo.gngm_desc,
3699 tx->tx_getinfo.gngm_cookie);
3700 RETURN(try_map_if_full ? rc : 0);
3702 case GNILND_MSG_GET_ACK_REV:
3703 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3704 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3705 /* LNET_GETS are a special case for parse */
3706 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3707 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3709 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3710 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3712 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3713 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3715 case GNILND_MSG_GET_DONE_REV:
3716 case GNILND_MSG_GET_REQ_REV:
3717 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3718 &tx->tx_putinfo.gnpam_desc,
3719 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3720 tx->tx_putinfo.gnpam_dst_cookie);
3721 RETURN(try_map_if_full ? rc : 0);
3729 kgnilnd_process_fmaq(kgn_conn_t *conn)
3732 kgn_tx_t *tx = NULL;
3733 void *buffer = NULL;
3734 unsigned int nob = 0;
3737 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3738 * However I will be rescheduled by an FMA completion event
3739 * when I eventually get some.
3740 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3741 * But it doesn't matter if I try to send a "real" message just
3742 * as I start closing because I'll get scheduled to send the
3745 /* Short circuit if the ep_handle is null we cant send anyway. */
3746 if (conn->gnc_ephandle == NULL)
3749 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3751 spin_lock(&conn->gnc_list_lock);
3753 if (list_empty(&conn->gnc_fmaq)) {
3754 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3756 spin_unlock(&conn->gnc_list_lock);
3758 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3759 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3760 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3761 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3762 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3764 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3765 conn->gnc_last_noop_cq,
3766 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3767 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3768 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3769 atomic_inc(&conn->gnc_sched_noop);
3770 set_mb(conn->gnc_last_noop_want, jiffies);
3772 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3775 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3779 rc = kgnilnd_set_tx_id(tx, conn);
3781 kgnilnd_tx_done(tx, rc);
3787 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3788 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3789 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3790 more_to_do = !list_empty(&conn->gnc_fmaq);
3791 spin_unlock(&conn->gnc_list_lock);
3794 /* if there is no real TX or no NOOP to send, bail */
3799 if (!tx->tx_retrans)
3800 tx->tx_cred_wait = jiffies;
3802 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3803 "tx with zero id", NULL);
3805 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3806 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3807 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3811 switch (tx->tx_msg.gnm_type) {
3815 case GNILND_MSG_NOOP:
3816 case GNILND_MSG_CLOSE:
3817 case GNILND_MSG_IMMEDIATE:
3818 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3819 buffer = tx->tx_buffer;
3823 case GNILND_MSG_GET_DONE:
3824 case GNILND_MSG_PUT_DONE:
3825 case GNILND_MSG_PUT_DONE_REV:
3826 case GNILND_MSG_GET_DONE_REV:
3827 case GNILND_MSG_PUT_NAK:
3828 case GNILND_MSG_GET_NAK:
3829 case GNILND_MSG_GET_NAK_REV:
3830 case GNILND_MSG_PUT_NAK_REV:
3831 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3834 case GNILND_MSG_PUT_REQ:
3835 case GNILND_MSG_GET_REQ_REV:
3836 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3838 case GNILND_MSG_PUT_ACK:
3839 case GNILND_MSG_PUT_REQ_REV:
3840 case GNILND_MSG_GET_ACK_REV:
3841 case GNILND_MSG_GET_REQ:
3842 /* This is really only to handle the retransmit of SMSG once these
3843 * two messages are setup in send_mapped_tx */
3844 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3848 if (likely(rc == 0)) {
3849 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3853 /* don't explicitly reschedule here - we are short credits and will rely on
3854 * kgnilnd_sendmsg to resched the conn if need be */
3856 } else if (rc < 0) {
3857 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3858 * almost certainly a software bug, but lets play nice with the other kids */
3859 kgnilnd_tx_done(tx, rc);
3860 /* just for fun, kick peer in arse - resetting conn might help to correct
3861 * this almost certainly buggy software caused return code */
3862 kgnilnd_close_conn(conn, rc);
3866 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3867 kgnilnd_schedule_conn(conn);
3872 kgnilnd_process_rdmaq(kgn_device_t *dev)
3877 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3881 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3882 unsigned long dead_bump;
3885 /* if we think we need to adjust, take lock to serialize and recheck */
3886 spin_lock(&dev->gnd_rdmaq_lock);
3887 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3888 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3890 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3892 /* roll the bucket forward */
3893 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3895 if (kgnilnd_data.kgn_rdmaq_override &&
3896 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3897 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3902 /* roll current outstanding forward to make sure we carry outstanding
3903 * committment forward
3904 * new_ok starts out as the whole interval value
3905 * - first subtract bytes_out from last interval, as that would push us over
3906 * strict limits for this interval
3907 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3909 * there is a small race here if someone is actively processing mappings and
3910 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3911 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3912 * - if this gives us problems in the future, we could use a read/write lock
3913 * to protect the resetting of these values */
3914 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3915 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3917 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3918 "current out %ld\n",
3919 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3920 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3922 spin_unlock(&dev->gnd_rdmaq_lock);
3925 spin_lock(&dev->gnd_rdmaq_lock);
3926 while (!list_empty(&dev->gnd_rdmaq)) {
3929 /* make sure we break out early on quiesce */
3930 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3931 /* always break with lock held - we unlock outside loop */
3935 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3936 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3939 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3940 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3941 /* if conn is dying, mark tx in tx_ref_table for
3942 * kgnilnd_complete_closed_conn to finish up */
3943 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3945 /* tx was moved to DYING, get next */
3948 spin_unlock(&dev->gnd_rdmaq_lock);
3950 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3951 spin_lock(&dev->gnd_rdmaq_lock);
3954 /* no ticket! add back to head */
3955 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3956 /* clear found_work so scheduler threads wait for timer */
3960 /* TX is GO for launch */
3961 tx->tx_qtime = jiffies;
3962 kgnilnd_send_mapped_tx(tx, 0);
3966 spin_unlock(&dev->gnd_rdmaq_lock);
3972 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3974 __swab64s(&d->gnrd_key.qword1);
3975 __swab64s(&d->gnrd_key.qword2);
3976 __swab64s(&d->gnrd_addr);
3977 __swab32s(&d->gnrd_nob);
3980 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3981 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3984 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3986 kgn_tx_ev_id_t ev_id;
3989 /* we use the cookie from the original TX, so we can find the match
3990 * by parsing that and using the txe_idx */
3991 ev_id.txe_cookie = cookie;
3993 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3996 /* check tx to make sure kgni didn't eat it */
3997 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3998 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
4000 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
4001 (tx->tx_id.txe_cookie = cookie)),
4002 "conn 0x%p->%s tx_ref_table hosed: wanted "
4003 "txe_cookie %#llx txe_idx %d "
4004 "found tx %p cookie %#llx txe_idx %d\n",
4005 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4006 cookie, ev_id.txe_idx,
4007 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
4009 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4010 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4011 "Unexpected TX type (%x, %x or %x) "
4012 "or state (%x, expected +%x) "
4013 "matched reply from %s\n",
4014 tx->tx_msg.gnm_type, type1, type2,
4015 tx->tx_state, GNILND_TX_WAITING_REPLY,
4016 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4018 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
4019 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4025 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4028 kgn_conn_t *conn = tx->tx_conn;
4029 __u64 nob = tx->tx_nob;
4030 __u32 physnop = tx->tx_phys_npages;
4031 int id = tx->tx_id.txe_smsg_id;
4032 int buftype = tx->tx_buftype;
4033 gni_mem_handle_t hndl;
4034 hndl.qword1 = tx->tx_map_key.qword1;
4035 hndl.qword2 = tx->tx_map_key.qword2;
4037 spin_lock(&conn->gnc_list_lock);
4039 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4040 "not waiting for reply", NULL);
4043 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4045 if (rc == -EFAULT) {
4046 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
4047 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4049 if(*kgnilnd_tunables.kgn_efault_lbug) {
4050 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4051 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4053 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4054 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4055 kgnilnd_tx_state2str(tx->tx_list_state),
4056 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4061 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4062 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4063 /* sample under lock as follow on steps require gnc_list_lock
4064 * - or call kgnilnd_tx_done which requires no locks held over
4065 * call to lnet_finalize */
4068 spin_unlock(&conn->gnc_list_lock);
4071 kgnilnd_tx_done(tx, tx->tx_rc);
4076 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4079 kgn_conn_t *conn = tx->tx_conn;
4081 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4082 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4084 /* the gncm_retval is passed in for PUTs */
4085 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4086 msg->gnm_u.completion.gncm_retval);
4088 kgnilnd_complete_tx(tx, rc);
4092 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4100 kgn_peer_t *peer = conn->gnc_peer;
4103 __u16 tmp_cksum = 0, msg_cksum = 0;
4104 int repost = 1, saw_complete;
4105 unsigned long timestamp, newest_last_rx, timeout;
4109 /* Short circuit if the ep_handle is null.
4110 * It's likely that its about to be closed as stale.
4112 if (conn->gnc_ephandle == NULL)
4115 timestamp = jiffies;
4116 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4117 /* delay in jiffies - we are really concerned only with things that
4118 * result in a schedule() or really holding this off for long times .
4119 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4120 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4122 /* Resample current time as we have no idea how long it took to get the mutex */
4123 timestamp = jiffies;
4125 /* We check here when the last time we received an rx, we do this before
4126 * we call getnext in case the thread has been blocked for a while. If we
4127 * havent received an rx since our timeout value we close the connection
4128 * as we should assume the other side has closed the connection. This will
4129 * stop us from sending replies to a mailbox that is already in purgatory.
4132 timeout = cfs_time_seconds(conn->gnc_timeout);
4133 newest_last_rx = GNILND_LASTRX(conn);
4135 /* Error injection to validate that timestamp checking works and closing the conn */
4136 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4137 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4140 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4141 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4142 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4143 cfs_duration_sec(timestamp - newest_last_rx),
4144 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4145 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4147 kgnilnd_close_conn(conn, rc);
4151 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4153 if (rrc == GNI_RC_NOT_DONE) {
4154 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4155 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4159 /* Instead of asserting when we get mailbox corruption lets attempt to
4160 * close the conn and recover. We can put the conn/mailbox into
4161 * purgatory and let purgatory deal with the problem. If we see
4162 * this NETTERROR reported on production systems in large amounts
4163 * we will need to revisit the state machine to see if we can tighten
4164 * it up further to improve data protection.
4167 if (rrc == GNI_RC_INVALID_STATE) {
4168 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4169 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4170 "detected closing conn %p from peer %s\n", conn,
4171 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4173 kgnilnd_close_conn(conn, rc);
4177 LASSERTF(rrc == GNI_RC_SUCCESS,
4178 "bad rc %d on conn %p from peer %s\n",
4179 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4181 msg = (kgn_msg_t *)prefix;
4183 rx = kgnilnd_alloc_rx();
4185 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4186 kgnilnd_release_msg(conn);
4187 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4188 conn, libcfs_nid2str(peer->gnp_nid));
4192 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4194 timestamp = conn->gnc_last_rx;
4195 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4196 atomic_inc(&conn->gnc_rx_seq);
4198 conn->gnc_last_rx = jiffies;
4199 /* stash first rx so we can clear out purgatory
4201 if (conn->gnc_first_rx == 0)
4202 conn->gnc_first_rx = jiffies;
4204 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4205 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4206 kgnilnd_peer_alive(conn->gnc_peer);
4209 rx->grx_conn = conn;
4211 rx->grx_received = current_kernel_time();
4213 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4216 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4222 kgnilnd_net_decref(net);
4225 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4226 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4228 /* XXX Nic: Do we need to swab cksum */
4229 if (msg->gnm_cksum != 0) {
4230 msg_cksum = msg->gnm_cksum;
4232 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4234 if (tmp_cksum != msg_cksum) {
4235 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4236 tmp_cksum, msg_cksum);
4237 kgnilnd_dump_msg(D_BUFFS, msg);
4242 /* restore checksum for future debug messages */
4243 msg->gnm_cksum = tmp_cksum;
4245 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4246 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4247 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4248 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4253 __swab32s(&msg->gnm_magic);
4254 __swab16s(&msg->gnm_version);
4255 __swab16s(&msg->gnm_type);
4256 __swab64s(&msg->gnm_srcnid);
4257 __swab64s(&msg->gnm_connstamp);
4258 __swab32s(&msg->gnm_seq);
4260 /* NB message type checked below; NOT here... */
4261 switch (msg->gnm_type) {
4262 case GNILND_MSG_GET_ACK_REV:
4263 case GNILND_MSG_PUT_ACK:
4264 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4267 case GNILND_MSG_PUT_REQ_REV:
4268 case GNILND_MSG_GET_REQ:
4269 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4277 if (msg->gnm_version != GNILND_MSG_VERSION) {
4278 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4279 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4284 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4285 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4286 libcfs_nid2str(msg->gnm_srcnid),
4287 libcfs_nid2str(peer->gnp_nid));
4292 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4293 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4294 " expected) from %s",
4295 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4296 libcfs_nid2str(peer->gnp_nid));
4301 if (msg->gnm_seq != seq) {
4302 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4303 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4308 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4310 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4311 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4312 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4313 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4314 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4315 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4316 /* double check state with lock held */
4317 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4318 /* only error if we are not already closing */
4319 if (conn->gnc_peer_error == -ETIMEDOUT) {
4320 unsigned long now = jiffies;
4321 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4323 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4324 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4325 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4327 cfs_duration_sec(now - timestamp),
4328 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4329 atomic_read(&conn->gnc_tx_seq),
4330 cfs_duration_sec(now - conn->gnc_last_tx),
4331 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4332 cfs_duration_sec(now - conn->gnc_last_noop_want),
4333 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4334 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4335 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4336 cfs_duration_sec(now - conn->gnc_last_sched_do),
4337 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4339 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4341 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4345 if (conn->gnc_close_recvd) {
4346 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4347 kgnilnd_msgtype2str(msg->gnm_type),
4348 msg->gnm_type, conn->gnc_close_recvd,
4349 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4354 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4355 /* XXX Nic: log message received on bad connection state */
4359 switch (msg->gnm_type) {
4360 case GNILND_MSG_NOOP:
4361 /* Nothing to do; just a keepalive */
4364 case GNILND_MSG_IMMEDIATE:
4365 /* only get SMSG payload for IMMEDIATE */
4366 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4367 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4368 msg->gnm_srcnid, rx, 0);
4371 case GNILND_MSG_GET_REQ_REV:
4372 case GNILND_MSG_PUT_REQ:
4373 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4374 msg->gnm_srcnid, rx, 1);
4377 case GNILND_MSG_GET_NAK_REV:
4378 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4379 msg->gnm_u.completion.gncm_cookie);
4383 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4385 case GNILND_MSG_PUT_NAK:
4386 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4387 msg->gnm_u.completion.gncm_cookie);
4391 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4393 case GNILND_MSG_PUT_ACK:
4394 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4395 msg->gnm_u.putack.gnpam_src_cookie);
4399 /* store putack data for later: deferred rdma or re-try */
4400 tx->tx_putinfo = msg->gnm_u.putack;
4403 spin_lock(&tx->tx_conn->gnc_list_lock);
4405 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4406 "not waiting for reply", NULL);
4408 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4410 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4411 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4412 /* sample under lock as follow on steps require gnc_list_lock
4413 * - or call kgnilnd_tx_done which requires no locks held over
4414 * call to lnet_finalize */
4417 /* cannot launch rdma if still waiting for fma-msg completion */
4418 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4419 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4420 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4422 spin_unlock(&tx->tx_conn->gnc_list_lock);
4425 rc = kgnilnd_send_mapped_tx(tx, 0);
4427 kgnilnd_tx_done(tx, rc);
4430 case GNILND_MSG_GET_ACK_REV:
4431 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4432 msg->gnm_u.putack.gnpam_src_cookie);
4436 /* store putack data for later: deferred rdma or re-try */
4437 tx->tx_putinfo = msg->gnm_u.putack;
4439 spin_lock(&tx->tx_conn->gnc_list_lock);
4441 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4442 "not waiting for reply", NULL);
4444 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4446 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4447 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4448 /* sample under lock as follow on steps require gnc_list_lock
4449 * - or call kgnilnd_tx_done which requires no locks held over
4450 * call to lnet_finalize */
4453 /* cannot launch rdma if still waiting for fma-msg completion */
4454 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4455 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4456 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4458 spin_unlock(&tx->tx_conn->gnc_list_lock);
4461 rc = kgnilnd_send_mapped_tx(tx, 0);
4463 kgnilnd_tx_done(tx, rc);
4466 case GNILND_MSG_PUT_DONE:
4467 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4468 msg->gnm_u.completion.gncm_cookie);
4472 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4473 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4474 "bad tx buftype %d", tx->tx_buftype);
4476 kgnilnd_finalize_rx_done(tx, msg);
4478 case GNILND_MSG_PUT_REQ_REV:
4479 case GNILND_MSG_GET_REQ:
4480 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4481 msg->gnm_srcnid, rx, 1);
4485 case GNILND_MSG_GET_NAK:
4486 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4487 msg->gnm_u.completion.gncm_cookie);
4491 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4492 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4493 "bad tx buftype %d", tx->tx_buftype);
4495 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4498 case GNILND_MSG_GET_DONE:
4499 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4500 msg->gnm_u.completion.gncm_cookie);
4504 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4505 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4506 "bad tx buftype %d", tx->tx_buftype);
4508 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4509 msg->gnm_u.completion.gncm_retval);
4511 kgnilnd_finalize_rx_done(tx, msg);
4513 case GNILND_MSG_GET_DONE_REV:
4514 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4515 msg->gnm_u.completion.gncm_cookie);
4519 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4520 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4521 "bad tx buftype %d", tx->tx_buftype);
4523 kgnilnd_finalize_rx_done(tx, msg);
4526 case GNILND_MSG_PUT_DONE_REV:
4527 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4528 msg->gnm_u.completion.gncm_cookie);
4533 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4534 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4535 "bad tx buftype %d", tx->tx_buftype);
4537 kgnilnd_finalize_rx_done(tx, msg);
4539 case GNILND_MSG_PUT_NAK_REV:
4540 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4541 msg->gnm_u.completion.gncm_cookie);
4546 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4547 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4548 "bad tx buftype %d", tx->tx_buftype);
4550 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4555 if (rc < 0) /* protocol/comms error */
4556 kgnilnd_close_conn(conn, rc);
4558 if (repost && rx != NULL) {
4559 kgnilnd_consume_rx(rx);
4562 /* we got an event so assume more there and call for reschedule */
4564 kgnilnd_schedule_conn(conn);
4568 /* Do the failure injections that we need to affect conn processing in the following function.
4569 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4570 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4573 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4577 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4581 /* short circuit out when not set */
4582 if (likely(!cfs_fail_loc)) {
4586 /* failure injection to test for stack reset clean ups */
4587 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4588 /* we can't rely on busy loops being nice enough to get the
4589 * stack reset triggered - it'd just spin on this conn */
4590 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4593 GOTO(did_fail_loc, rc);
4596 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4597 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4599 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4600 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4603 GOTO(did_fail_loc, rc);
4607 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4608 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4609 * will be closed by kgnilnd_complete_closed_conn.
4611 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4612 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4615 GOTO(did_fail_loc, rc);
4618 /* this one is a bit gross - we can't hold the mutex from process_conns
4619 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4620 * and moving onto finish_connect
4621 * so, we'll just set the rc - kgnilnd_process_conns will clear
4622 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4623 * and effectively getting this thread to sleep */
4624 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4627 GOTO(did_fail_loc, rc);
4635 kgnilnd_send_conn_close(kgn_conn_t *conn)
4639 /* we are closing the conn - we will try to send the CLOSE msg
4640 * but will not wait for anything else to flush */
4642 /* send the close if not already done so or received one */
4643 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4644 /* set close_sent regardless of the success of the
4645 * CLOSE message. We are going to try once and then
4646 * kick him out of the sandbox */
4647 conn->gnc_close_sent = 1;
4650 /* EP might be null already if remote side initiated a new connection.
4651 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4652 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4654 if (conn->gnc_ephandle != NULL) {
4657 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4659 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4660 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4661 tx->tx_qtime = jiffies;
4663 if (tx->tx_id.txe_idx == 0) {
4664 rc = kgnilnd_set_tx_id(tx, conn);
4666 kgnilnd_tx_done(tx, rc);
4670 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4673 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4674 kgnilnd_tx_done(tx, -EAGAIN);
4676 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4678 /* It wasnt sent and we dont care. */
4679 kgnilnd_tx_done(tx, rc);
4687 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4688 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4689 conn->gnc_state = GNILND_CONN_CLOSED;
4690 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4691 /* mark this conn as CLOSED now that we processed it
4692 * do after TX, so we can use CLOSING in asserts */
4696 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4697 /* simulate a RX CLOSE after the timeout but before
4698 * the scheduler thread gets it */
4699 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4700 conn->gnc_peer_error = -ETIMEDOUT;
4702 /* schedule to allow potential CLOSE and get the complete phase run */
4703 kgnilnd_schedule_conn(conn);
4707 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4712 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4713 int log_retrans, log_retrans_level;
4714 static int last_map_version;
4717 spin_lock(&dev->gnd_lock);
4718 if (list_empty(&dev->gnd_map_tx)) {
4719 /* if the list is empty make sure we dont have a timer running */
4720 del_singleshot_timer_sync(&dev->gnd_map_timer);
4721 spin_unlock(&dev->gnd_lock);
4725 dev->gnd_sched_alive = jiffies;
4727 /* we'll retry as fast as possible up to 25% of the limit, then we start
4728 * backing off until our map version changes - indicating we unmapped
4730 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4731 if (likely(dev->gnd_map_attempt == 0) ||
4732 time_after_eq(jiffies, dev->gnd_next_map) ||
4733 last_map_version != dev->gnd_map_version) {
4735 /* if this is our first attempt at mapping set last mapped to current
4736 * jiffies so we can timeout our attempt correctly.
4738 if (dev->gnd_map_attempt == 0)
4739 dev->gnd_last_map = jiffies;
4741 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4742 spin_unlock(&dev->gnd_lock);
4746 /* delete the previous timer if it exists */
4747 del_singleshot_timer_sync(&dev->gnd_map_timer);
4748 /* stash the last map version to let us know when a good one was seen */
4749 last_map_version = dev->gnd_map_version;
4751 /* we need to to take the lock and continually refresh the head of the list as
4752 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4753 * allowing them to squeeze in */
4755 while (!list_empty(&dev->gnd_map_tx)) {
4756 /* make sure we break out early on quiesce */
4757 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4758 /* always break with lock held - we unlock outside loop */
4762 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4764 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4767 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4768 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4769 /* if conn is dying, mark tx in tx_ref_table for
4770 * kgnilnd_complete_closed_conn to finish up */
4771 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4774 /* tx was moved to DYING, get next */
4778 spin_unlock(&dev->gnd_lock);
4779 rc = kgnilnd_send_mapped_tx(tx, 1);
4781 /* We made it! skip error handling.. */
4783 /* OK to continue on +ve errors as it won't get seen until
4784 * this function is called again - we operate on a copy of the original
4785 * list and not the live list */
4786 spin_lock(&dev->gnd_lock);
4787 /* reset map attempts back to zero we successfully
4788 * mapped so we can reset our timers */
4789 dev->gnd_map_attempt = 0;
4791 } else if (rc == -EAGAIN) {
4792 spin_lock(&dev->gnd_lock);
4793 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4794 spin_unlock(&dev->gnd_lock);
4795 GOTO(get_out_mapped, rc);
4796 } else if (rc != -ENOMEM) {
4797 /* carp, failure we can't handle */
4798 kgnilnd_tx_done(tx, rc);
4799 spin_lock(&dev->gnd_lock);
4800 /* reset map attempts back to zero we dont know what happened but it
4801 * wasnt a failed mapping
4803 dev->gnd_map_attempt = 0;
4807 /* time to handle the retry cases.. lock so we dont have 2 threads
4808 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4810 spin_lock(&dev->gnd_lock);
4811 dev->gnd_map_attempt++;
4812 if (dev->gnd_map_attempt < fast_remaps) {
4813 /* do nothing we just want it to go as fast as possible.
4814 * just set gnd_next_map to current jiffies so it will process
4815 * as fast as possible.
4817 dev->gnd_next_map = jiffies;
4819 /* Retry based on GNILND_MAP_RETRY_RATE */
4820 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4823 /* only log occasionally once we've retried fast_remaps */
4824 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4825 ((dev->gnd_map_attempt % fast_remaps) == 0);
4826 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4828 /* make sure we are not off in the weeds with this tx */
4829 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4830 GNIDBG_TX(D_NETERROR, tx,
4831 "giving up on TX, too many retries", NULL);
4832 spin_unlock(&dev->gnd_lock);
4833 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4834 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4835 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4837 tx->tx_putinfo.gnpam_dst_cookie,
4838 tx->tx_msg.gnm_srcnid);
4840 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4842 tx->tx_getinfo.gngm_cookie,
4843 tx->tx_msg.gnm_srcnid);
4845 kgnilnd_tx_done(tx, -ENOMEM);
4846 GOTO(get_out_mapped, rc);
4848 GNIDBG_TX(log_retrans_level, tx,
4849 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4851 "nq_map %d mdd# %d/%d GART %ld",
4852 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4853 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4854 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4855 atomic_read(&dev->gnd_nq_map),
4856 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4857 atomic64_read(&dev->gnd_nbytes_map));
4860 /* we need to stop processing the rest of the list, so add it back in */
4861 /* set timer to wake device when we need to schedule this tx */
4862 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4863 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4864 spin_unlock(&dev->gnd_lock);
4865 GOTO(get_out_mapped, rc);
4867 spin_unlock(&dev->gnd_lock);
4873 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4878 int error_inject = 0;
4882 spin_lock(&dev->gnd_lock);
4883 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4884 dev->gnd_sched_alive = jiffies;
4888 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4889 /* break with lock held */
4893 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4894 list_del_init(&conn->gnc_schedlist);
4895 spin_unlock(&dev->gnd_lock);
4897 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4899 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4900 conn_sched != GNILND_CONN_PROCESS,
4901 "conn %p on ready list but in bad state: %d\n",
4904 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4905 conn, kgnilnd_conn_state2str(conn));
4908 set_mb(conn->gnc_last_sched_do, jiffies);
4910 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4912 /* based on intent see if we should run again. */
4913 rc = kgnilnd_schedule_process_conn(conn, intent);
4915 /* drop ref from gnd_ready_conns */
4916 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4917 down_write(&dev->gnd_conn_sem);
4918 kgnilnd_conn_decref(conn);
4919 up_write(&dev->gnd_conn_sem);
4920 } else if (rc != 1) {
4921 kgnilnd_conn_decref(conn);
4923 /* clear this so that scheduler thread doesn't spin */
4925 /* break with lock held... */
4926 spin_lock(&dev->gnd_lock);
4930 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4931 down_write(&dev->gnd_conn_sem);
4933 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4934 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4935 /* If there are tx's currently in use in another
4936 * thread we dont want to complete the close
4937 * yet. Cycle this conn back through
4939 kgnilnd_schedule_conn(conn);
4941 kgnilnd_complete_closed_conn(conn);
4943 up_write(&dev->gnd_conn_sem);
4944 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4945 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4946 /* serialize SMSG CQs with ep_bind and smsg_release */
4947 down_write(&dev->gnd_conn_sem);
4948 kgnilnd_destroy_conn_ep(conn);
4949 up_write(&dev->gnd_conn_sem);
4950 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4951 /* if we need to do some CLOSE sending, etc done here do it */
4952 down_write(&dev->gnd_conn_sem);
4953 kgnilnd_send_conn_close(conn);
4954 kgnilnd_check_fma_rx(conn);
4955 up_write(&dev->gnd_conn_sem);
4956 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4957 /* start moving traffic if the old conns are cleared out */
4958 down_read(&dev->gnd_conn_sem);
4959 kgnilnd_check_fma_rx(conn);
4960 kgnilnd_process_fmaq(conn);
4961 up_read(&dev->gnd_conn_sem);
4964 rc = kgnilnd_schedule_process_conn(conn, 0);
4966 /* drop ref from gnd_ready_conns */
4967 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4968 down_write(&dev->gnd_conn_sem);
4969 kgnilnd_conn_decref(conn);
4970 up_write(&dev->gnd_conn_sem);
4971 } else if (rc != 1) {
4972 kgnilnd_conn_decref(conn);
4975 /* check list again with lock held */
4976 spin_lock(&dev->gnd_lock);
4979 /* If we are short circuiting due to timing we want to be scheduled
4980 * as soon as possible.
4982 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4985 spin_unlock(&dev->gnd_lock);
4991 kgnilnd_scheduler(void *arg)
4993 int threadno = (long)arg;
4996 unsigned long deadline = 0;
4999 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
5001 cfs_block_allsigs();
5003 /* all gnilnd threads need to run fairly urgently */
5004 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
5005 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5006 while (!kgnilnd_data.kgn_shutdown) {
5008 /* Safe: kgn_shutdown only set when quiescent */
5010 /* to quiesce or to not quiesce, that is the question */
5012 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5013 KGNILND_SPIN_QUIESCE;
5016 /* tracking for when thread goes AWOL */
5017 dev->gnd_sched_alive = jiffies;
5019 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5020 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5021 /* let folks know we are up and kicking
5022 * - they can use this for latency savings, etc
5023 * - only change if IRQ, if IDLE leave alone as that
5024 * schedule_device calls to put us back to IRQ */
5025 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5027 down_read(&dev->gnd_conn_sem);
5028 /* always check these - they are super low cost */
5029 found_work += kgnilnd_check_fma_send_cq(dev);
5030 found_work += kgnilnd_check_fma_rcv_cq(dev);
5032 /* rdma CQ doesn't care about eps */
5033 found_work += kgnilnd_check_rdma_cq(dev);
5035 /* move some RDMA ? */
5036 found_work += kgnilnd_process_rdmaq(dev);
5038 /* map some pending RDMA requests ? */
5039 found_work += kgnilnd_process_mapped_tx(dev);
5041 /* the EP for a conn is not destroyed until all the references
5042 * to it are gone, so these checks should be safe
5043 * even if run in parallel with the CQ checking functions
5044 * _AND_ a thread that processes the CLOSED->DONE
5048 up_read(&dev->gnd_conn_sem);
5050 /* process all conns ready now */
5051 found_work += kgnilnd_process_conns(dev, deadline);
5053 /* do an eager check to avoid the IRQ disabling in
5054 * prepare_to_wait and friends */
5057 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5058 time_before(jiffies, deadline)) {
5060 if ((busy_loops % 10) == 0) {
5061 /* tickle heartbeat and watchdog to ensure our
5062 * piggishness doesn't turn into heartbeat failure */
5063 touch_nmi_watchdog();
5069 /* if we got here, found_work was zero or busy_loops means we
5070 * need to take a break. We'll clear gnd_ready but we'll check
5071 * one last time if there is an IRQ that needs processing */
5073 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5075 /* the first time this will go LOOP -> IDLE and let us do one final check
5076 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5077 * - this might allow other threads to block us for a bit if they
5078 * try to get the mutex, but that is good as we'd need to wake
5079 * up soon to handle the CQ or other processing anyways */
5081 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5083 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5084 time_after_eq(jiffies, deadline)) {
5086 "yeilding: found_work %d busy_loops %d\n",
5087 found_work, busy_loops);
5089 /* use yield if we are bailing due to busy_loops
5090 * - this will ensure we wake up soonish. This closes
5091 * a race with kgnilnd_device_callback - where it'd
5092 * not call wake_up() because gnd_ready == 1, but then
5093 * we come down and schedule() because of busy_loops.
5094 * We'd not be woken up until something poked our waitq
5095 * again. yield() ensures we wake up without another
5096 * waitq poke in that case */
5097 atomic_inc(&dev->gnd_n_yield);
5098 kgnilnd_data.kgn_last_condresched = jiffies;
5100 CDEBUG(D_INFO, "awake after yeild\n");
5101 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5102 } else if (found_work == GNILND_DEV_IDLE) {
5103 /* busy_loops is low and there is nothing to do,
5104 * go to sleep and wait for a waitq poke */
5106 "scheduling: found_work %d busy_loops %d\n",
5107 found_work, busy_loops);
5108 atomic_inc(&dev->gnd_n_schedule);
5109 kgnilnd_data.kgn_last_scheduled = jiffies;
5111 CDEBUG(D_INFO, "awake after schedule\n");
5112 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5114 finish_wait(&dev->gnd_waitq, &wait);
5117 kgnilnd_thread_fini();