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>
29 #include <linux/pagemap.h>
31 #include <libcfs/linux/linux-mem.h>
35 /* this is useful when needed to debug wire corruption. */
37 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
45 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
46 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
49 } else if (len >= 16) {
51 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
52 prefix, ptr, *(ptr), *(ptr + 1));
56 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
65 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
67 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
68 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
69 msg->gnm_magic, msg->gnm_version,
70 msg->gnm_type, msg->gnm_srcnid,
71 msg->gnm_connstamp, msg->gnm_seq,
72 msg->gnm_cksum, msg->gnm_payload_cksum,
73 msg->gnm_payload_len);
77 kgnilnd_schedule_device(kgn_device_t *dev)
79 short already_live = 0;
81 /* we'll only want to wake if the scheduler thread
82 * has come around and set ready to zero */
83 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
86 wake_up_all(&dev->gnd_waitq);
90 void kgnilnd_schedule_device_timer(cfs_timer_cb_arg_t data)
92 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_map_timer);
94 kgnilnd_schedule_device(dev);
97 void kgnilnd_schedule_device_timer_rd(cfs_timer_cb_arg_t data)
99 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_rdmaq_timer);
101 kgnilnd_schedule_device(dev);
105 kgnilnd_device_callback(__u32 devid, __u64 arg)
108 int index = (int) arg;
110 if (index >= kgnilnd_data.kgn_ndevs) {
111 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
112 * interrupt context. */
113 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
118 dev = &kgnilnd_data.kgn_devices[index];
119 /* just basic sanity */
120 if (dev->gnd_id == devid) {
121 kgnilnd_schedule_device(dev);
123 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
128 /* sched_intent values:
129 * < 0 : do not reschedule under any circumstances
130 * == 0: reschedule if someone marked him WANTS_SCHED
131 * > 0 : force a reschedule */
132 /* Return code 0 means it did not schedule the conn, 1
133 * means it successfully scheduled the conn.
137 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
141 /* move back to IDLE but save previous state.
142 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
143 * let the xchg there handle any racing callers to get it
144 * onto gnd_ready_conns */
146 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
147 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
148 conn_sched == GNILND_CONN_PROCESS,
149 "conn %p after process in bad state: %d\n",
152 if (sched_intent >= 0) {
153 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
154 return kgnilnd_schedule_conn_refheld(conn, 1);
160 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
164 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
166 kgn_device_t *dev = conn->gnc_device;
170 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
171 /* we only care about the last person who marked want_sched since they
172 * are most likely the culprit
174 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
175 conn->gnc_sched_line = line;
176 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
177 * the chance to put it onto gnd_ready_conns.
178 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
179 * the conn in process_conns will take care of moving it back to
180 * SCHED when it is done processing */
182 if (sched == GNILND_CONN_IDLE) {
183 /* if the conn is already scheduled, we've already requested
184 * the scheduler thread wakeup */
186 /* Add a reference to the conn if we are not holding a reference
187 * already from the exisiting scheduler. We now use the same
188 * reference if we need to reschedule a conn while in a scheduler
191 kgnilnd_conn_addref(conn);
193 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
196 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
198 spin_lock(&dev->gnd_lock);
199 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
201 spin_unlock(&dev->gnd_lock);
202 set_mb(conn->gnc_last_sched_ask, jiffies);
205 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
209 /* make sure thread(s) going to process conns - but let it make
210 * separate decision from conn schedule */
212 kgnilnd_schedule_device(dev);
217 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
219 kgn_device_t *dev = conn->gnc_device;
221 spin_lock(&dev->gnd_lock);
222 if (list_empty(&conn->gnc_delaylist)) {
223 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
226 spin_unlock(&dev->gnd_lock);
228 kgnilnd_schedule_device(dev);
233 kgnilnd_schedule_dgram(kgn_device_t *dev)
237 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
238 if (wake != GNILND_DGRAM_SCHED) {
239 wake_up(&dev->gnd_dgram_waitq);
241 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
246 kgnilnd_free_tx(kgn_tx_t *tx)
248 /* taken from kgnilnd_tx_add_state_locked */
250 LASSERTF((tx->tx_list_p == NULL &&
251 tx->tx_list_state == GNILND_TX_ALLOCD) &&
252 list_empty(&tx->tx_list),
253 "tx %p with bad state %s (list_p %p) tx_list %s\n",
254 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
255 list_empty(&tx->tx_list) ? "empty" : "not empty");
257 atomic_dec(&kgnilnd_data.kgn_ntx);
259 /* we only allocate this if we need to */
260 if (tx->tx_phys != NULL) {
261 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
262 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
263 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
266 /* Only free the buffer if we used it */
267 if (tx->tx_buffer_copy != NULL) {
268 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
269 tx->tx_buffer_copy = NULL;
270 CDEBUG(D_MALLOC, "vfreed buffer2\n");
273 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
275 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
276 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
280 kgnilnd_alloc_tx (void)
284 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
287 tx = kmem_cache_zalloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
289 CERROR("failed to allocate tx\n");
292 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
295 /* setup everything here to minimize time under the lock */
296 tx->tx_buftype = GNILND_BUF_NONE;
297 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
298 INIT_LIST_HEAD(&tx->tx_list);
299 INIT_LIST_HEAD(&tx->tx_map_list);
300 tx->tx_list_state = GNILND_TX_ALLOCD;
302 atomic_inc(&kgnilnd_data.kgn_ntx);
307 /* csum_fold needs to be run on the return value before shipping over the wire */
308 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
310 /* we don't use offset as every one is passing a buffer reference that already
311 * includes the offset into the base address.
314 kgnilnd_cksum(void *ptr, size_t nob)
318 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
320 /* don't use magic 'no checksum' value */
324 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
331 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
332 unsigned int offset, unsigned int nob, int dump_blob)
338 unsigned int fraglen;
344 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
345 kiov, nkiov, offset, nob, dump_blob);
347 /* if loops changes, please change kgnilnd_setup_phys_buffer */
349 while (offset >= kiov->bv_len) {
350 offset -= kiov->bv_len;
356 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
357 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
359 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
360 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
362 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
363 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
365 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
366 odd, kiov[0].bv_len, offset, nob);
368 for (i = 0; i < nkiov; i++) {
369 pages[i] = kiov[i].bv_page;
372 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
374 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
376 /* return zero to avoid killing tx - we'll just get warning on console
377 * when remote end sees zero checksum */
380 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
382 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
387 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
388 (void *)addr + kiov[0].bv_offset +
391 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
392 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
396 fraglen = min(kiov->bv_len - offset, nob);
398 /* make dang sure we don't send a bogus checksum if somehow we get
399 * an odd length fragment on anything but the last entry in a kiov -
400 * we know from kgnilnd_setup_rdma_buffer that we can't have non
401 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
402 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
403 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
404 fraglen, nkiov, nob, kiov->bv_len,
407 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
409 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
412 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
413 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
414 addr, fraglen, offset);
419 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
421 kunmap(kiov->bv_page);
428 /* iov must not run out before end of data */
429 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
434 retsum = csum_fold(cksum);
436 /* don't use magic 'no checksum' value */
440 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
446 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
448 msg->gnm_magic = GNILND_MSG_MAGIC;
449 msg->gnm_version = GNILND_MSG_VERSION;
450 msg->gnm_type = type;
451 msg->gnm_payload_len = 0;
452 msg->gnm_srcnid = source;
453 /* gnm_connstamp gets set when FMA is sent */
454 /* gnm_srcnid is set on creation via function argument
455 * The right interface/net and nid is passed in when the message
461 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
463 kgn_tx_t *tx = kgnilnd_alloc_tx();
466 kgnilnd_init_msg(&tx->tx_msg, type, source);
468 CERROR("couldn't allocate new tx type %s!\n",
469 kgnilnd_msgtype2str(type));
476 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
482 case GNILND_MSG_GET_REQ:
483 case GNILND_MSG_GET_DONE:
484 nak_type = GNILND_MSG_GET_NAK;
486 case GNILND_MSG_PUT_REQ:
487 case GNILND_MSG_PUT_ACK:
488 case GNILND_MSG_PUT_DONE:
489 nak_type = GNILND_MSG_PUT_NAK;
491 case GNILND_MSG_PUT_REQ_REV:
492 case GNILND_MSG_PUT_DONE_REV:
493 nak_type = GNILND_MSG_PUT_NAK_REV;
495 case GNILND_MSG_GET_REQ_REV:
496 case GNILND_MSG_GET_ACK_REV:
497 case GNILND_MSG_GET_DONE_REV:
498 nak_type = GNILND_MSG_GET_NAK_REV;
501 CERROR("invalid msg type %s (%d)\n",
502 kgnilnd_msgtype2str(rx_type), rx_type);
505 /* only allow NAK on error and truncate to zero */
506 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
507 error, conn, cookie);
509 tx = kgnilnd_new_tx_msg(nak_type, source);
511 CNETERR("can't get TX to NAK RDMA to %s\n",
512 libcfs_nid2str(conn->gnc_peer->gnp_nid));
516 tx->tx_msg.gnm_u.completion.gncm_retval = error;
517 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
518 kgnilnd_queue_tx(conn, tx);
522 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
523 struct bio_vec *kiov,
524 unsigned int offset, unsigned int nob)
526 kgn_msg_t *msg = &tx->tx_msg;
529 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
530 * gni_smsg_send to send that as the payload */
532 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
535 tx->tx_buffer = NULL;
538 if (niov && niov > (nob >> PAGE_SHIFT))
539 niov = DIV_ROUND_UP(nob + offset + kiov->bv_offset,
542 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
543 "bad niov %d msg %p kiov %p offset %d nob%d\n",
544 niov, msg, kiov, offset, nob);
546 while (offset >= kiov->bv_len) {
547 offset -= kiov->bv_len;
552 for (i = 0; i < niov; i++) {
553 /* We can't have a bv_offset on anything but the first
554 * entry, otherwise we'll have a hole at the end of the
555 * mapping as we only map whole pages.
556 * Also, if we have a bv_len < PAGE_SIZE but we need to
557 * map more than bv_len, we will also have a whole at
558 * the end of that page which isn't allowed
560 if ((kiov[i].bv_offset != 0 && i > 0) ||
561 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
563 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
564 i, offset, nob, kiov->bv_offset,
568 tx->tx_imm_pages[i] = kiov[i].bv_page;
571 /* hijack tx_phys for the later unmap */
573 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
575 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
576 kiov[0].bv_offset + offset;
577 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
578 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
579 nob, kiov, tx->tx_buffer);
581 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
582 if (tx->tx_phys == NULL) {
583 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
587 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
588 /* make sure we take into account the kiov offset as the
589 * start of the buffer
591 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
594 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
595 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
597 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
602 /* checksum payload early - it shouldn't be changing after lnd_send */
603 if (*kgnilnd_tunables.kgn_checksum >= 2) {
604 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
605 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
606 msg->gnm_payload_cksum += 0xe00e;
608 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
609 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
613 msg->gnm_payload_cksum = 0;
620 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
621 unsigned int offset, unsigned int nob)
623 gni_mem_segment_t *phys;
625 unsigned int fraglen;
627 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
631 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
633 /* only allocate this if we are going to use it */
634 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
636 if (tx->tx_phys == NULL) {
637 CERROR("failed to allocate tx_phys\n");
642 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
643 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
645 /* if loops changes, please change kgnilnd_cksum_kiov
646 * and kgnilnd_setup_immediate_buffer */
648 while (offset >= kiov->bv_len) {
649 offset -= kiov->bv_len;
655 /* at this point, kiov points to the first page that we'll actually map
656 * now that we've seeked into the koiv for offset and dropped any
657 * leading pages that fall entirely within the offset */
658 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
661 /* bv_offset is start of 'valid' buffer, so index offset past that */
662 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
665 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
666 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
669 fraglen = min(kiov->bv_len - offset, nob);
671 /* We can't have a bv_offset on anything but the first entry,
672 * otherwise we'll have a hole at the end of the mapping as we
673 * only map whole pages. Only the first page is allowed to
674 * have an offset - we'll add that into tx->tx_buffer and that
675 * will get used when we map in the segments (see
676 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
677 * but we need to map more than bv_len, we will also have a
678 * whole at the end of that page which isn't allowed
680 if ((phys != tx->tx_phys) &&
681 ((kiov->bv_offset != 0) ||
682 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
683 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
684 (int)(phys - tx->tx_phys),
685 offset, nob, kiov->bv_offset, kiov->bv_len);
690 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
691 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
696 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
702 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
703 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
706 phys->address = page_to_phys(kiov->bv_page);
713 /* iov must not run out before end of data */
714 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
718 tx->tx_phys_npages = phys - tx->tx_phys;
723 if (tx->tx_phys != NULL) {
724 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
725 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
726 sizeof(*tx->tx_phys), tx->tx_phys);
733 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
734 struct bio_vec *kiov,
735 unsigned int offset, unsigned int nob)
737 return kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
740 /* kgnilnd_parse_lnet_rdma()
741 * lntmsg - message passed in from lnet.
742 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
743 * nob - actual number of bytes to in this message.
744 * put_len - It is possible for PUTs to have a different length than the
745 * length stored in lntmsg->msg_len since LNET can adjust this
746 * length based on it's buffer size and offset.
747 * lnet_try_match_md() sets the mlength that we use to do the RDMA
751 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
752 unsigned int *offset, unsigned int *nob,
753 struct bio_vec **kiov, int put_len)
755 /* GETs are weird, see kgnilnd_send */
756 if (lntmsg->msg_type == LNET_MSG_GET) {
757 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
760 *kiov = lntmsg->msg_md->md_kiov;
762 *niov = lntmsg->msg_md->md_niov;
763 *nob = lntmsg->msg_md->md_length;
766 *kiov = lntmsg->msg_kiov;
767 *niov = lntmsg->msg_niov;
769 *offset = lntmsg->msg_offset;
774 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
776 unsigned int niov, offset, nob;
777 struct bio_vec *kiov;
778 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
779 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
781 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
782 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
783 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
784 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
785 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
786 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
787 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
789 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
790 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
791 tx->tx_msg.gnm_payload_cksum = 0;
794 if (*kgnilnd_tunables.kgn_checksum < 3) {
795 tx->tx_msg.gnm_payload_cksum = 0;
799 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
801 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
805 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
807 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
809 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
813 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
814 tx->tx_msg.gnm_payload_cksum += 0xd00d;
818 /* kgnilnd_verify_rdma_cksum()
819 * tx - PUT_DONE/GET_DONE matched tx.
820 * rx_cksum - received checksum to compare against.
821 * put_len - see kgnilnd_parse_lnet_rdma comments.
824 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
828 unsigned int niov, offset, nob;
829 struct bio_vec *kiov;
830 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
831 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
833 /* we can only match certain requests */
834 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
835 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
836 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
837 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
838 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
839 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
840 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
842 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
843 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
848 if (*kgnilnd_tunables.kgn_checksum >= 3) {
849 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
850 "no RDMA payload checksum when enabled");
855 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
857 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
860 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
862 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
865 if (cksum != rx_cksum) {
866 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
867 "Bad RDMA payload checksum (%x expected %x); "
868 "kiov 0x%p niov %d nob %u offset %u",
869 cksum, rx_cksum, kiov, niov, nob, offset);
870 switch (dump_on_err) {
873 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
875 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
880 libcfs_debug_dumplog();
886 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
892 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
896 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
897 "already mapped!", NULL);
899 spin_lock(&dev->gnd_map_lock);
900 switch (tx->tx_buftype) {
902 GNIDBG_TX(D_EMERG, tx,
903 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
904 spin_unlock(&dev->gnd_map_lock);
908 case GNILND_BUF_PHYS_MAPPED:
909 bytes = tx->tx_phys_npages * PAGE_SIZE;
910 dev->gnd_map_nphys++;
911 dev->gnd_map_physnop += tx->tx_phys_npages;
915 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
916 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
917 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
918 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
919 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
922 atomic_inc(&dev->gnd_n_mdd);
923 atomic64_add(bytes, &dev->gnd_nbytes_map);
925 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
928 /* we only get here in the valid cases */
929 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
930 dev->gnd_map_version++;
931 spin_unlock(&dev->gnd_map_lock);
935 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
939 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
940 "not mapped!", NULL);
941 spin_lock(&dev->gnd_map_lock);
943 switch (tx->tx_buftype) {
945 GNIDBG_TX(D_EMERG, tx,
946 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
947 spin_unlock(&dev->gnd_map_lock);
951 case GNILND_BUF_PHYS_UNMAPPED:
952 bytes = tx->tx_phys_npages * PAGE_SIZE;
953 dev->gnd_map_nphys--;
954 dev->gnd_map_physnop -= tx->tx_phys_npages;
958 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
959 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
960 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
961 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
962 "bytes_out negative! %lld\n",
963 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
964 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
965 bytes, (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
968 atomic_dec(&dev->gnd_n_mdd);
969 atomic64_sub(bytes, &dev->gnd_nbytes_map);
971 /* we only get here in the valid cases */
972 list_del_init(&tx->tx_map_list);
973 dev->gnd_map_version++;
974 spin_unlock(&dev->gnd_map_lock);
978 kgnilnd_map_buffer(kgn_tx_t *tx)
980 kgn_conn_t *conn = tx->tx_conn;
981 kgn_device_t *dev = conn->gnc_device;
982 __u32 flags = GNI_MEM_READWRITE;
985 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
986 * be called concurrently as there are internal locks that protect
987 * any data structures or HW resources. We just need to ensure
988 * that our concurrency doesn't result in the kgn_device_t
989 * getting nuked while we are in here */
991 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
992 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
994 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
997 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
998 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1001 switch (tx->tx_buftype) {
1005 case GNILND_BUF_NONE:
1006 case GNILND_BUF_IMMEDIATE:
1007 case GNILND_BUF_IMMEDIATE_KIOV:
1008 case GNILND_BUF_PHYS_MAPPED:
1011 case GNILND_BUF_PHYS_UNMAPPED:
1012 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1013 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1014 tx->tx_phys, tx->tx_phys_npages, NULL,
1015 GNI_MEM_PHYS_SEGMENTS | flags,
1017 /* could race with other uses of the map counts, but this is ok
1018 * - this needs to turn into a non-fatal error soon to allow
1019 * GART resource, etc starvation handling */
1020 if (rrc != GNI_RC_SUCCESS) {
1021 GNIDBG_TX(D_NET, tx,
1022 "Can't map %d pages: dev %d phys %u pp %u",
1023 tx->tx_phys_npages, dev->gnd_id,
1024 dev->gnd_map_nphys, dev->gnd_map_physnop);
1025 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1028 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1029 kgnilnd_mem_add_map_list(dev, tx);
1035 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1037 kgn_conn_t *conn = tx->tx_conn;
1038 kgn_mdd_purgatory_t *gmp;
1040 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1041 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1042 " asserting to avoid data corruption\n");
1043 if (tx->tx_buffer_copy)
1044 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1046 gmp->gmp_map_key = tx->tx_map_key;
1048 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1050 /* ensure that we don't have a blank purgatory - indicating the
1051 * conn is not already on purgatory lists - we'd never recover these
1052 * MDD if that were the case */
1053 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1054 "conn 0x%p->%s with NULL purgatory",
1055 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1057 /* link 'er up! - only place we really need to lock for
1058 * concurrent access */
1059 spin_lock(&conn->gnc_list_lock);
1060 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1061 spin_unlock(&conn->gnc_list_lock);
1065 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1069 int hold_timeout = 0;
1071 /* code below relies on +1 relationship ... */
1072 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1073 (GNILND_BUF_PHYS_UNMAPPED + 1));
1075 switch (tx->tx_buftype) {
1079 case GNILND_BUF_NONE:
1080 case GNILND_BUF_IMMEDIATE:
1081 case GNILND_BUF_PHYS_UNMAPPED:
1083 case GNILND_BUF_IMMEDIATE_KIOV:
1084 if (tx->tx_phys != NULL) {
1085 vunmap(tx->tx_phys);
1086 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1087 kunmap(tx->tx_imm_pages[0]);
1089 /* clear to prevent kgnilnd_free_tx from thinking
1090 * this is a RDMA descriptor */
1094 case GNILND_BUF_PHYS_MAPPED:
1095 LASSERT(tx->tx_conn != NULL);
1097 dev = tx->tx_conn->gnc_device;
1099 /* only want to hold if we are closing conn without
1100 * verified peer notification - the theory is that
1101 * a TX error can be communicated in all other cases */
1102 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1103 error != -GNILND_NOPURG &&
1104 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1105 kgnilnd_add_purgatory_tx(tx);
1107 /* The timeout we give to kgni is a deadman stop only.
1108 * we are setting high to ensure we don't have the kgni timer
1109 * fire before ours fires _and_ is handled */
1110 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1112 GNIDBG_TX(D_NET, tx,
1113 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1114 tx->tx_conn->gnc_device, hold_timeout,
1115 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1117 if (tx->tx_buffer_copy != NULL) {
1118 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1119 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1120 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1121 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1123 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1124 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1128 kgnilnd_mem_del_map_list(dev, tx);
1134 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1136 struct lnet_msg *lntmsg0, *lntmsg1;
1137 int status0, status1;
1138 struct lnet_ni *ni = NULL;
1139 kgn_conn_t *conn = tx->tx_conn;
1141 LASSERT(!in_interrupt());
1143 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1144 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1147 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1148 !kgnilnd_conn_clean_errno(completion)) {
1149 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1150 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1151 completion, tx, conn ?
1152 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1153 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1154 kgnilnd_tx_state2str(tx->tx_list_state),
1155 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1158 /* The error codes determine if we hold onto the MDD */
1159 kgnilnd_unmap_buffer(tx, completion);
1161 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1162 * we play nice with the error codes to avoid delivering a failed
1163 * REQUEST and then a REPLY event as well */
1165 /* return -EIO to lnet - it is the magic value for failed sends */
1166 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1168 status1 = completion;
1170 status0 = status1 = completion;
1173 tx->tx_buftype = GNILND_BUF_NONE;
1174 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1176 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1177 * set NULL when we are a tx without a conn */
1179 ni = conn->gnc_peer->gnp_net->gnn_ni;
1181 spin_lock(&conn->gnc_tx_lock);
1183 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1184 (volatile unsigned long *)&conn->gnc_tx_bits),
1185 "conn %p tx %p bit %d already cleared\n",
1186 conn, tx, tx->tx_id.txe_idx);
1188 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1189 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1191 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1192 spin_unlock(&conn->gnc_tx_lock);
1195 kgnilnd_free_tx(tx);
1197 /* finalize AFTER freeing lnet msgs */
1199 /* warning - we should hold no locks here - calling lnet_finalize
1200 * could free up lnet credits, resulting in a call chain back into
1201 * the LND via kgnilnd_send and friends */
1203 lnet_finalize(lntmsg0, status0);
1205 if (lntmsg1 != NULL) {
1206 lnet_finalize(lntmsg1, status1);
1211 kgnilnd_txlist_done(struct list_head *txlist, int error)
1214 int err_printed = 0;
1216 if (list_empty(txlist))
1219 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1220 /* only print the first error */
1222 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1223 list_del_init(&tx->tx_list);
1224 kgnilnd_tx_done(tx, error);
1229 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1233 spin_lock(&conn->gnc_tx_lock);
1235 /* ID zero is NOT ALLOWED!!! */
1238 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1239 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1240 if (id == GNILND_MAX_MSG_ID) {
1241 if (conn->gnc_next_tx != 1) {
1242 /* we only searched from next_tx to end and didn't find
1243 * one, so search again from start */
1244 conn->gnc_next_tx = 1;
1247 /* couldn't find one! */
1248 spin_unlock(&conn->gnc_tx_lock);
1252 /* bump next_tx to prevent immediate reuse */
1253 conn->gnc_next_tx = id + 1;
1255 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1256 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1257 "tx 0x%p already at id %d\n",
1258 conn->gnc_tx_ref_table[id], id);
1260 /* delay these until we have a valid ID - prevents bad clear of the bit
1261 * in kgnilnd_tx_done */
1263 tx->tx_id.txe_cqid = conn->gnc_cqid;
1265 tx->tx_id.txe_idx = id;
1266 conn->gnc_tx_ref_table[id] = tx;
1268 /* Using jiffies to help differentiate against TX reuse - with
1269 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1270 * if we are sending to the same node faster than 256000/sec.
1271 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1273 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1275 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1277 spin_unlock(&conn->gnc_tx_lock);
1282 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1286 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1288 /* we don't care about TX timeout - it could be that the network is slower
1289 * or throttled. We'll keep retranmitting - so if the network is so slow
1290 * that we fill up our mailbox, we'll keep trying to resend that msg
1291 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1292 * that he hasn't send us any traffic in return */
1294 /* some reasonable throttling of the debug message */
1296 unsigned long now = jiffies;
1297 /* XXX Nic: Mystical TX debug here... */
1298 /* We expect retransmissions so only log when D_NET is enabled */
1299 GNIDBG_SMSG_CREDS(D_NET, conn);
1300 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1301 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1302 " last_msg %uus/%uus last_cq %uus/%uus",
1303 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1304 tx->tx_id, tx->tx_retrans,
1305 jiffies_to_usecs(now - tx->tx_cred_wait),
1306 jiffies_to_usecs(now - conn->gnc_last_tx),
1307 jiffies_to_usecs(now - conn->gnc_last_rx),
1308 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1309 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1313 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1314 * to avoid bad ordering with state_lock */
1317 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1318 spinlock_t *state_lock, kgn_tx_list_state_t state)
1320 kgn_conn_t *conn = tx->tx_conn;
1321 kgn_msg_t *msg = &tx->tx_msg;
1323 unsigned long newest_last_rx, timeout;
1326 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1327 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1329 "msg 0x%p type %d wrong payload size %d\n",
1330 msg, msg->gnm_type, immediatenob);
1332 /* make sure we catch all the cases where we'd send on a dirty old mbox
1333 * but allow case for sending CLOSE. Since this check is within the CQ
1334 * mutex barrier and the close message is only sent through
1335 * kgnilnd_send_conn_close the last message out the door will be the
1338 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1339 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1340 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1341 /* Return -ETIME, we are closing the connection already so we dont want to
1342 * have this tx hit the wire. The tx will be killed by the calling function.
1343 * Once the EP is marked dirty the close message will be the last
1344 * thing to hit the wire */
1349 timeout = cfs_time_seconds(conn->gnc_timeout);
1351 newest_last_rx = GNILND_LASTRX(conn);
1353 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1354 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1357 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1358 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1359 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1360 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1361 cfs_duration_sec(now - newest_last_rx),
1362 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1363 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1364 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1368 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1369 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1370 * creating the message this allows the message to contain the correct LNET NID/NET needed
1371 * instead of the one that the peer/conn uses for sending the data.
1373 msg->gnm_connstamp = conn->gnc_my_connstamp;
1374 msg->gnm_payload_len = immediatenob;
1375 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1377 /* always init here - kgn_checksum is a /sys module tunable
1378 * and can be flipped at any point, even between msg init and sending */
1380 if (*kgnilnd_tunables.kgn_checksum) {
1381 /* We must set here and not in kgnilnd_init_msg,
1382 * we could resend this msg many times
1383 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1384 * through init_msg again */
1385 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1386 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1387 msg->gnm_cksum += 0xf00f;
1391 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1392 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1393 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1394 tx->tx_id.txe_idx, immediate, immediatenob);
1396 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1397 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1399 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1400 msg, sizeof(*msg), immediate,
1402 tx->tx_id.txe_smsg_id);
1406 case GNI_RC_SUCCESS:
1407 atomic_inc(&conn->gnc_tx_seq);
1408 conn->gnc_last_tx = jiffies;
1409 /* no locking here as LIVE isn't a list */
1410 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1412 /* this needs to be checked under lock as it might be freed from a completion
1415 if (msg->gnm_type == GNILND_MSG_NOOP) {
1416 set_mb(conn->gnc_last_noop_sent, jiffies);
1419 /* serialize with seeing CQ events for completion on this, as well as
1421 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1422 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1424 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1425 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1426 kgnilnd_peer_alive(conn->gnc_peer);
1427 GNIDBG_SMSG_CREDS(D_NET, conn);
1430 case GNI_RC_NOT_DONE:
1431 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1432 * 1: out of mbox credits
1433 * 2: out of mbox payload credits
1434 * 3: On Aries out of dla credits
1436 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1437 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1438 /* We'll handle this error inline - makes the calling logic much more
1441 /* If no lock, caller doesn't want us to retry */
1442 if (state_lock == NULL) {
1446 /* I need kgni credits to send this. Replace tx at the head of the
1447 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1448 * and bump retrans counts since we are requeueing the tx.
1452 conn->gnc_tx_retrans++;
1454 kgnilnd_tx_log_retrans(conn, tx);
1455 /* add to head of list for the state and retries */
1456 spin_lock(state_lock);
1457 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1458 spin_unlock(state_lock);
1460 /* We only reschedule for a certain number of retries, then
1461 * we will wait for the CQ events indicating a release of SMSG
1463 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1464 kgnilnd_schedule_conn(conn);
1467 /* CQ event coming in signifies either TX completed or
1468 * RX receive. Either of these *could* free up credits
1469 * in the SMSG mbox and we should try sending again */
1470 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1471 tx->tx_conn->gnc_cqid);
1472 kgnilnd_schedule_delay_conn(conn);
1473 /* use +ve return code to let upper layers know they
1474 * should stop looping on sends */
1478 /* handle bad retcode gracefully */
1479 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1480 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1485 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1487 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1488 spinlock_t *state_lock, kgn_tx_list_state_t state)
1490 kgn_device_t *dev = tx->tx_conn->gnc_device;
1491 unsigned long timestamp;
1494 timestamp = jiffies;
1495 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1496 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1497 /* delay in jiffies - we are really concerned only with things that
1498 * result in a schedule() or really holding this off for long times .
1499 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1500 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1502 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1508 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1510 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1511 spinlock_t *state_lock, kgn_tx_list_state_t state)
1513 kgn_conn_t *conn = tx->tx_conn;
1514 kgn_device_t *dev = conn->gnc_device;
1515 unsigned long timestamp;
1518 timestamp = jiffies;
1520 /* technically we are doing bad things with the read_lock on the peer_conn
1521 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1522 * for the mutex. I bet lockdep is gonna flag this one though... */
1524 /* there are a few cases where we don't want the immediate send - like
1525 * when we are in the scheduler thread and it'd harm the latency of
1526 * getting messages up to LNet */
1528 /* rmb for gnd_ready */
1530 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1532 atomic_inc(&conn->gnc_device->gnd_fast_block);
1533 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1534 /* dont hit HW during quiesce */
1536 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1537 /* dont hit HW if stale EPs and conns left to close */
1540 atomic_inc(&conn->gnc_device->gnd_fast_try);
1541 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1542 &conn->gnc_smsg_mutex);
1547 /* we got the mutex and weren't blocked */
1549 /* delay in jiffies - we are really concerned only with things that
1550 * result in a schedule() or really holding this off for long times .
1551 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1552 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1554 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1555 tx->tx_qtime = jiffies;
1556 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1557 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1558 /* _nolock unlocks the mutex for us */
1564 /* lets us know if we can push this RDMA through now */
1566 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1570 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1572 if (bytes_left < 0) {
1573 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1574 atomic_inc(&dev->gnd_rdmaq_nstalls);
1577 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1578 dev->gnd_rdmaq_deadline);
1579 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1580 /* we never del this timer - at worst it schedules us.. */
1587 /* this adds a TX to the queue pending throttling authorization before
1588 * we allow our remote peer to launch a PUT at us */
1590 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1594 /* we cannot go into send_mapped_tx from here as we are holding locks
1595 * and mem registration might end up allocating memory in kgni.
1596 * That said, we'll push this as far as we can into the queue process */
1597 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1600 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1601 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1602 /* lets us know how delayed RDMA is */
1603 tx->tx_qtime = jiffies;
1604 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1606 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1607 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1608 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1609 /* lets us know how delayed mapping is */
1610 tx->tx_qtime = jiffies;
1611 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1614 /* make sure we wake up sched to run this */
1615 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1618 /* push TX through state machine */
1620 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1625 /* set the tx_id here, we delay it until we have an actual conn
1627 * in some cases, the tx_id is already set to provide for things
1628 * like RDMA completion cookies, etc */
1629 if (tx->tx_id.txe_idx == 0) {
1630 rc = kgnilnd_set_tx_id(tx, conn);
1632 kgnilnd_tx_done(tx, rc);
1637 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1638 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1640 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1642 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1643 kgnilnd_tx_done(tx, rc);
1647 switch (tx->tx_msg.gnm_type) {
1648 case GNILND_MSG_PUT_ACK:
1649 case GNILND_MSG_GET_REQ:
1650 case GNILND_MSG_PUT_REQ_REV:
1651 case GNILND_MSG_GET_ACK_REV:
1652 /* hijacking time! If this messages will authorize our peer to
1653 * send his dirty little bytes in an RDMA, we need to get permission */
1654 kgnilnd_queue_rdma(conn, tx);
1656 case GNILND_MSG_IMMEDIATE:
1657 /* try to send right now, can help reduce latency */
1658 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1661 /* it was sent, break out of switch to avoid default case of queueing */
1664 /* needs to queue to try again, so... */
1665 /* fall through... */
1666 case GNILND_MSG_NOOP:
1667 /* Just make sure this goes out first for this conn */
1669 /* fall through... */
1671 spin_lock(&conn->gnc_list_lock);
1672 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1673 tx->tx_qtime = jiffies;
1674 spin_unlock(&conn->gnc_list_lock);
1675 kgnilnd_schedule_conn(conn);
1680 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1683 kgn_peer_t *new_peer = NULL;
1684 kgn_conn_t *conn = NULL;
1690 /* If I get here, I've committed to send, so I complete the tx with
1691 * failure on any problems */
1693 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1694 "tx already has connection %p", tx->tx_conn);
1696 /* do all of the peer & conn searching in one swoop - this avoids
1697 * nastiness when dropping locks and needing to maintain a sane state
1698 * in the face of stack reset or something else nuking peers & conns */
1700 /* I expect to find him, so only take a read lock */
1701 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1703 peer = kgnilnd_find_peer_locked(target->nid);
1705 conn = kgnilnd_find_conn_locked(peer);
1706 /* this could be NULL during quiesce */
1708 /* Connection exists; queue message on it */
1709 kgnilnd_queue_tx(conn, tx);
1710 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1714 /* don't create a connection if the peer is marked down */
1715 if (peer->gnp_state != GNILND_PEER_UP) {
1716 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1722 /* creating peer or conn; I'll need a write lock... */
1723 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1725 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1727 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1729 /* NB - this will not block during normal operations -
1730 * the only writer of this is in the startup/shutdown path. */
1731 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1737 /* ignore previous peer entirely - we cycled the lock, so we
1738 * will create new peer and at worst drop it if peer is still
1740 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1742 up_read(&kgnilnd_data.kgn_net_rw_sem);
1746 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1747 up_read(&kgnilnd_data.kgn_net_rw_sem);
1749 /* search for peer again now that we have the lock
1750 * if we don't find it, add our new one to the list */
1751 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1753 /* don't create a connection if the peer is not up */
1754 if (peer->gnp_state != GNILND_PEER_UP) {
1755 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1760 conn = kgnilnd_find_or_create_conn_locked(peer);
1762 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1763 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1768 /* oh hey, found a conn now... magical */
1769 kgnilnd_queue_tx(conn, tx);
1771 /* no conn, must be trying to connect - so we queue for now */
1772 tx->tx_qtime = jiffies;
1773 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1775 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1778 kgnilnd_tx_done(tx, rc);
1783 kgnilnd_rdma(kgn_tx_t *tx, int type,
1784 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1786 kgn_conn_t *conn = tx->tx_conn;
1787 unsigned long timestamp;
1788 gni_post_type_t post_type;
1791 unsigned int desc_nob = nob;
1792 void *desc_buffer = tx->tx_buffer;
1793 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1794 LASSERTF(kgnilnd_tx_mapped(tx),
1795 "unmapped tx %p\n", tx);
1796 LASSERTF(conn != NULL,
1797 "NULL conn on tx %p, naughty, naughty\n", tx);
1798 LASSERTF(nob <= sink->gnrd_nob,
1799 "nob %u > sink->gnrd_nob %d (%p)\n",
1800 nob, sink->gnrd_nob, sink);
1801 LASSERTF(nob <= tx->tx_nob,
1802 "nob %d > tx(%p)->tx_nob %d\n",
1803 nob, tx, tx->tx_nob);
1806 case GNILND_MSG_GET_DONE:
1807 case GNILND_MSG_PUT_DONE:
1808 post_type = GNI_POST_RDMA_PUT;
1810 case GNILND_MSG_GET_DONE_REV:
1811 case GNILND_MSG_PUT_DONE_REV:
1812 post_type = GNI_POST_RDMA_GET;
1815 CERROR("invalid msg type %s (%d)\n",
1816 kgnilnd_msgtype2str(type), type);
1819 if (post_type == GNI_POST_RDMA_GET) {
1820 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1821 * aligned. If the local buffer is not aligned correctly using the copy buffer
1822 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1823 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1824 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1825 * making sure the length and addr are aligned and copying the data into the correct location
1826 * once the transfer has completed.
1828 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1829 (sink->gnrd_addr & 3) ||
1832 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1834 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1836 if ((nob + tx->tx_offset) & 3) {
1837 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1838 atomic_inc(&kgnilnd_data.kgn_rev_length);
1840 desc_nob = (nob + tx->tx_offset);
1843 if (tx->tx_buffer_copy == NULL) {
1844 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1845 * and require at most we allocate a few extra bytes. */
1846 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1848 if (!tx->tx_buffer_copy) {
1849 /* allocation of buffer failed nak the rdma */
1850 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1851 kgnilnd_tx_done(tx, -EFAULT);
1854 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1855 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);
1856 if (rc != GNI_RC_SUCCESS) {
1857 /* Registration Failed nak rdma and kill the tx. */
1858 kgnilnd_vfree(tx->tx_buffer_copy,
1860 tx->tx_buffer_copy = NULL;
1861 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1862 kgnilnd_tx_done(tx, -EFAULT);
1866 desc_map_key = tx->tx_buffer_copy_map_key;
1867 desc_buffer = tx->tx_buffer_copy;
1871 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1872 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1873 tx->tx_rdma_desc.type = post_type;
1874 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1875 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1876 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1877 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1878 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1879 tx->tx_rdma_desc.length = desc_nob;
1880 tx->tx_nob_rdma = nob;
1881 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1882 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1883 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1884 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1885 /* prep final completion message */
1886 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1887 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1888 /* send actual size RDMA'd in retval */
1889 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1891 kgnilnd_compute_rdma_cksum(tx, nob);
1894 kgnilnd_queue_tx(conn, tx);
1898 /* Don't lie (CLOSE == RDMA idle) */
1899 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1900 tx, conn, conn->gnc_close_sent);
1902 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1903 "0x%x cookie:%#llx",
1904 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1906 /* set CQ dedicated for RDMA */
1907 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1909 timestamp = jiffies;
1910 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1911 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1912 /* delay in jiffies - we are really concerned only with things that
1913 * result in a schedule() or really holding this off for long times .
1914 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1915 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1917 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
1919 if (rrc == GNI_RC_ERROR_RESOURCE) {
1920 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1921 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1922 kgnilnd_unmap_buffer(tx, 0);
1924 if (tx->tx_buffer_copy != NULL) {
1925 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
1926 tx->tx_buffer_copy = NULL;
1929 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1930 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
1932 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1933 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1937 spin_lock(&conn->gnc_list_lock);
1938 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
1939 tx->tx_qtime = jiffies;
1940 spin_unlock(&conn->gnc_list_lock);
1941 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1942 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1944 /* XXX Nic: is this a place we should handle more errors for
1945 * robustness sake */
1946 LASSERT(rrc == GNI_RC_SUCCESS);
1951 kgnilnd_alloc_rx(void)
1955 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
1957 CERROR("failed to allocate rx\n");
1960 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
1963 /* no memset to zero, we'll always fill all members */
1967 /* release is to just free connection resources
1968 * we use this for the eager path after copying */
1970 kgnilnd_release_msg(kgn_conn_t *conn)
1973 unsigned long timestamp;
1975 CDEBUG(D_NET, "consuming %p\n", conn);
1977 timestamp = jiffies;
1978 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1979 /* delay in jiffies - we are really concerned only with things that
1980 * result in a schedule() or really holding this off for long times .
1981 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1982 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1984 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
1985 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1987 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
1988 GNIDBG_SMSG_CREDS(D_NET, conn);
1990 kgnilnd_schedule_conn(conn);
1994 kgnilnd_consume_rx(kgn_rx_t *rx)
1996 kgn_conn_t *conn = rx->grx_conn;
1997 kgn_msg_t *rxmsg = rx->grx_msg;
1999 /* if we are eager, free the cache alloc'd msg */
2000 if (unlikely(rx->grx_eager)) {
2001 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2002 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2004 /* release ref from eager_recv */
2005 kgnilnd_conn_decref(conn);
2007 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2008 kgnilnd_release_msg(conn);
2011 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2012 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2017 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2019 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2020 int type = lntmsg->msg_type;
2021 struct lnet_process_id target = lntmsg->msg_target;
2022 int target_is_router = lntmsg->msg_target_is_router;
2023 int routing = lntmsg->msg_routing;
2024 unsigned int niov = lntmsg->msg_niov;
2025 struct bio_vec *kiov = lntmsg->msg_kiov;
2026 unsigned int offset = lntmsg->msg_offset;
2027 unsigned int nob = lntmsg->msg_len;
2028 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2029 kgn_net_t *net = ni->ni_data;
2032 /* '1' for consistency with code that checks !mpflag to restore */
2033 unsigned int mpflag = 1;
2034 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2036 /* NB 'private' is different depending on what we're sending.... */
2037 LASSERT(!in_interrupt());
2039 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2040 type, nob, niov, libcfs_id2str(target));
2042 LASSERTF(nob == 0 || niov > 0,
2043 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2044 LASSERTF(niov <= LNET_MAX_IOV,
2045 "lntmsg %p niov %d\n", lntmsg, niov);
2048 mpflag = memalloc_noreclaim_save();
2052 CERROR("lntmsg %p with unexpected type %d\n",
2057 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2065 if (routing || target_is_router)
2066 break; /* send IMMEDIATE */
2068 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2069 * check the eventual sink buffer here - if small enough, remote
2070 * end is perfectly capable of returning data in short message -
2071 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2072 * for IMMEDIATE messages which will have it send a real reply instead
2073 * of doing kgnilnd_recv to have the RDMA continued */
2074 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2077 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2078 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2080 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2086 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2087 lntmsg->msg_md->md_kiov,
2088 0, lntmsg->msg_md->md_length);
2090 CERROR("unable to setup buffer: %d\n", rc);
2091 kgnilnd_tx_done(tx, rc);
2096 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2097 if (tx->tx_lntmsg[1] == NULL) {
2098 CERROR("Can't create reply for GET to %s\n",
2099 libcfs_nid2str(target.nid));
2100 kgnilnd_tx_done(tx, rc);
2105 tx->tx_lntmsg[0] = lntmsg;
2106 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2107 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2109 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2111 /* rest of tx_msg is setup just before it is sent */
2112 kgnilnd_launch_tx(tx, net, &target);
2114 case LNET_MSG_REPLY:
2116 /* to save on MDDs, we'll handle short kiov by vmap'ing
2117 * and sending via SMSG */
2118 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2121 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2122 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2124 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2131 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2134 kgnilnd_tx_done(tx, rc);
2139 tx->tx_lntmsg[0] = lntmsg;
2140 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2141 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2143 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2145 /* rest of tx_msg is setup just before it is sent */
2146 kgnilnd_launch_tx(tx, net, &target);
2150 /* send IMMEDIATE */
2152 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2153 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2155 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2161 rc = kgnilnd_setup_immediate_buffer(tx, niov, kiov, offset, nob);
2163 kgnilnd_tx_done(tx, rc);
2167 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2168 tx->tx_lntmsg[0] = lntmsg;
2169 kgnilnd_launch_tx(tx, net, &target);
2172 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2174 memalloc_noreclaim_restore(mpflag);
2179 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2181 kgn_conn_t *conn = rx->grx_conn;
2182 kgn_msg_t *rxmsg = rx->grx_msg;
2183 unsigned int niov = lntmsg->msg_niov;
2184 struct bio_vec *kiov = lntmsg->msg_kiov;
2185 unsigned int offset = lntmsg->msg_offset;
2186 unsigned int nob = lntmsg->msg_len;
2191 switch (rxmsg->gnm_type) {
2192 case GNILND_MSG_PUT_REQ_REV:
2193 done_type = GNILND_MSG_PUT_DONE_REV;
2196 case GNILND_MSG_GET_REQ:
2197 done_type = GNILND_MSG_GET_DONE;
2200 CERROR("invalid msg type %s (%d)\n",
2201 kgnilnd_msgtype2str(rxmsg->gnm_type),
2206 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2210 rc = kgnilnd_set_tx_id(tx, conn);
2214 rc = kgnilnd_setup_rdma_buffer(tx, niov, kiov, offset, nob);
2218 tx->tx_lntmsg[0] = lntmsg;
2219 tx->tx_getinfo = rxmsg->gnm_u.get;
2221 /* we only queue from kgnilnd_recv - we might get called from other contexts
2222 * and we don't want to block the mutex in those cases */
2224 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2225 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2226 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2227 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2232 kgnilnd_tx_done(tx, rc);
2233 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2235 lnet_finalize(lntmsg, rc);
2239 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2242 kgn_rx_t *rx = private;
2243 kgn_conn_t *conn = rx->grx_conn;
2244 kgn_msg_t *rxmsg = rx->grx_msg;
2245 kgn_msg_t *eagermsg = NULL;
2246 kgn_peer_t *peer = NULL;
2247 kgn_conn_t *found_conn = NULL;
2249 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2250 conn, rxmsg, lntmsg);
2252 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2253 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2254 rxmsg->gnm_payload_len);
2257 /* Grab a read lock so the connection doesnt disappear on us
2258 * while we look it up
2260 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2262 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2264 found_conn = kgnilnd_find_conn_locked(peer);
2267 /* Verify the connection found is the same one that the message
2268 * is supposed to be using, if it is not output an error message
2271 if (!peer || !found_conn
2272 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2273 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2274 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2275 peer, conn, found_conn);
2277 CERROR("Unexpected connstamp %#llx(%#llx expected) from %s\n",
2278 rxmsg->gnm_connstamp,
2279 found_conn->gnc_peer_connstamp,
2280 libcfs_nid2str(peer->gnp_nid));
2285 /* add conn ref to ensure it doesn't go away until all eager
2286 * messages processed */
2287 kgnilnd_conn_addref(conn);
2289 /* Now that we have verified the connection is valid and added a
2290 * reference we can remove the read_lock on the peer_conn_lock */
2291 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2293 /* we have no credits or buffers for this message, so copy it
2294 * somewhere for a later kgnilnd_recv */
2295 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2296 *kgnilnd_tunables.kgn_eager_credits) {
2297 CERROR("Out of eager credits to %s\n",
2298 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2302 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2304 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2305 if (eagermsg == NULL) {
2306 kgnilnd_conn_decref(conn);
2307 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2308 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2312 /* copy msg and payload */
2313 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2314 rx->grx_msg = eagermsg;
2317 /* stash this for lnet_finalize on cancel-on-conn-close */
2318 rx->grx_lntmsg = lntmsg;
2320 /* keep the same rx_t, it just has a new grx_msg now */
2321 *new_private = private;
2323 /* release SMSG buffer */
2324 kgnilnd_release_msg(conn);
2330 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2331 int delayed, unsigned int niov,
2332 struct bio_vec *kiov,
2333 unsigned int offset, unsigned int mlen, unsigned int rlen)
2335 kgn_rx_t *rx = private;
2336 kgn_conn_t *conn = rx->grx_conn;
2337 kgn_msg_t *rxmsg = rx->grx_msg;
2343 LASSERT(!in_interrupt());
2344 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2346 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2347 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2348 conn, rxmsg, lntmsg,
2349 niov, kiov, offset, mlen, rlen);
2351 /* we need to lock here as recv can be called from any context */
2352 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2353 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2354 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2356 /* someone closed the conn after we copied this out, nuke it */
2357 kgnilnd_consume_rx(rx);
2358 lnet_finalize(lntmsg, conn->gnc_error);
2361 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2363 switch (rxmsg->gnm_type) {
2365 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2366 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2367 conn, rx, rxmsg, lntmsg, niov, kiov, offset, mlen, rlen);
2370 case GNILND_MSG_IMMEDIATE:
2371 if (mlen > rxmsg->gnm_payload_len) {
2372 GNIDBG_MSG(D_ERROR, rxmsg,
2373 "Immediate message from %s too big: %d > %d",
2374 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2375 rxmsg->gnm_payload_len);
2377 kgnilnd_consume_rx(rx);
2381 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2382 * right after the kgn_msg_t header - so just 'cute' way of saying
2383 * rxmsg + sizeof(kgn_msg_t) */
2385 /* check payload checksum if sent */
2387 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2388 !rxmsg->gnm_payload_cksum &&
2389 rxmsg->gnm_payload_len != 0)
2390 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2392 if (rxmsg->gnm_payload_cksum != 0) {
2393 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2394 * which is what is used to calculate the cksum on the TX side */
2395 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2397 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2398 GNIDBG_MSG(D_NETERROR, rxmsg,
2399 "Bad payload checksum (%x expected %x)",
2400 pload_cksum, rxmsg->gnm_payload_cksum);
2401 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2403 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2404 &rxmsg[1], rxmsg->gnm_payload_len);
2407 libcfs_debug_dumplog();
2413 /* checksum problems are fatal, kill the conn */
2414 kgnilnd_consume_rx(rx);
2415 kgnilnd_close_conn(conn, rc);
2420 lnet_copy_flat2kiov(
2422 *kgnilnd_tunables.kgn_max_immediate,
2423 &rxmsg[1], 0, mlen);
2425 kgnilnd_consume_rx(rx);
2426 lnet_finalize(lntmsg, 0);
2429 case GNILND_MSG_PUT_REQ:
2430 /* LNET wants to truncate or drop transaction, sending NAK */
2432 kgnilnd_consume_rx(rx);
2433 lnet_finalize(lntmsg, 0);
2435 /* only error if lntmsg == NULL, otherwise we are just
2436 * short circuiting the rdma process of 0 bytes */
2437 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2438 lntmsg == NULL ? -ENOENT : 0,
2439 rxmsg->gnm_u.get.gngm_cookie,
2443 /* sending ACK with sink buff. info */
2444 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2446 kgnilnd_consume_rx(rx);
2450 rc = kgnilnd_set_tx_id(tx, conn);
2452 GOTO(nak_put_req, rc);
2455 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2456 kiov, offset, mlen);
2458 GOTO(nak_put_req, rc);
2461 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2462 rxmsg->gnm_u.putreq.gnprm_cookie;
2463 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2464 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2465 (__u64)((unsigned long)tx->tx_buffer);
2466 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2468 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2469 tx->tx_qtime = jiffies;
2470 /* we only queue from kgnilnd_recv - we might get called from other contexts
2471 * and we don't want to block the mutex in those cases */
2473 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2474 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2475 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2476 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2478 kgnilnd_consume_rx(rx);
2482 /* make sure we send an error back when the PUT fails */
2483 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2484 kgnilnd_tx_done(tx, rc);
2485 kgnilnd_consume_rx(rx);
2487 /* return magic LNet network error */
2489 case GNILND_MSG_GET_REQ_REV:
2490 /* LNET wants to truncate or drop transaction, sending NAK */
2492 kgnilnd_consume_rx(rx);
2493 lnet_finalize(lntmsg, 0);
2495 /* only error if lntmsg == NULL, otherwise we are just
2496 * short circuiting the rdma process of 0 bytes */
2497 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2498 lntmsg == NULL ? -ENOENT : 0,
2499 rxmsg->gnm_u.get.gngm_cookie,
2503 /* lntmsg can be null when parsing a LNET_GET */
2504 if (lntmsg != NULL) {
2505 /* sending ACK with sink buff. info */
2506 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2508 kgnilnd_consume_rx(rx);
2512 rc = kgnilnd_set_tx_id(tx, conn);
2514 GOTO(nak_get_req_rev, rc);
2516 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2517 kiov, offset, mlen);
2519 GOTO(nak_get_req_rev, rc);
2521 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2522 rxmsg->gnm_u.putreq.gnprm_cookie;
2523 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2524 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2525 (__u64)((unsigned long)tx->tx_buffer);
2526 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2528 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2530 /* we only queue from kgnilnd_recv - we might get called from other contexts
2531 * and we don't want to block the mutex in those cases */
2533 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2534 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2535 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2536 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2539 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2541 rxmsg->gnm_u.get.gngm_cookie,
2545 kgnilnd_consume_rx(rx);
2549 /* make sure we send an error back when the GET fails */
2550 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2551 kgnilnd_tx_done(tx, rc);
2552 kgnilnd_consume_rx(rx);
2554 /* return magic LNet network error */
2558 case GNILND_MSG_PUT_REQ_REV:
2559 /* LNET wants to truncate or drop transaction, sending NAK */
2561 kgnilnd_consume_rx(rx);
2562 lnet_finalize(lntmsg, 0);
2564 /* only error if lntmsg == NULL, otherwise we are just
2565 * short circuiting the rdma process of 0 bytes */
2566 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2567 lntmsg == NULL ? -ENOENT : 0,
2568 rxmsg->gnm_u.get.gngm_cookie,
2573 if (lntmsg != NULL) {
2575 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2578 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2580 rxmsg->gnm_u.get.gngm_cookie,
2583 kgnilnd_consume_rx(rx);
2585 case GNILND_MSG_GET_REQ:
2586 if (lntmsg != NULL) {
2588 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2591 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2593 rxmsg->gnm_u.get.gngm_cookie,
2596 kgnilnd_consume_rx(rx);
2602 /* needs write_lock on kgn_peer_conn_lock held */
2604 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2606 unsigned long timeout, keepalive;
2607 unsigned long now = jiffies;
2608 unsigned long newest_last_rx;
2611 /* given that we found this conn hanging off a peer, it better damned
2612 * well be connected */
2613 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2614 "conn 0x%p->%s with bad state%s\n", conn,
2615 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2617 kgnilnd_conn_state2str(conn));
2619 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2620 "rx_diff %lu tx_diff %lu\n",
2621 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2622 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2623 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2624 cfs_duration_sec(now - conn->gnc_last_tx));
2626 timeout = cfs_time_seconds(conn->gnc_timeout);
2627 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2629 /* just in case our lack of RX msg processing is gumming up the works - give the
2630 * remove an extra chance */
2632 newest_last_rx = GNILND_LASTRX(conn);
2634 if (time_after_eq(now, newest_last_rx + timeout)) {
2635 uint32_t level = D_CONSOLE|D_NETERROR;
2637 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2640 GNIDBG_CONN(level, conn,
2641 "No gnilnd traffic received from %s for %lu "
2642 "seconds, terminating connection. Is node down? ",
2643 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2644 cfs_duration_sec(now - newest_last_rx));
2648 /* we don't timeout on last_tx stalls - we are going to trust the
2649 * underlying network to let us know when sends are failing.
2650 * At worst, the peer will timeout our RX stamp and drop the connection
2651 * at that point. We'll then see his CLOSE or at worst his RX
2652 * stamp stop and drop the connection on our end */
2654 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2655 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2656 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2657 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2658 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2660 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2661 conn->gnc_last_noop_cq,
2662 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2663 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2664 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2665 set_mb(conn->gnc_last_noop_want, jiffies);
2666 atomic_inc(&conn->gnc_reaper_noop);
2667 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2670 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2673 kgnilnd_queue_tx(conn, tx);
2679 /* needs write_lock on kgn_peer_conn_lock held */
2681 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2682 struct list_head *souls)
2684 unsigned long timeout;
2685 kgn_conn_t *conn, *connN = NULL;
2691 short releaseconn = 0;
2692 unsigned long first_rx = 0;
2693 int purgatory_conn_cnt = 0;
2695 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2696 peer, libcfs_nid2str(peer->gnp_nid),
2697 peer->gnp_reconnect_interval);
2699 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2700 GNILND_MIN_TIMEOUT));
2702 conn = kgnilnd_find_conn_locked(peer);
2704 /* if there is a valid conn, check the queues for timeouts */
2705 rc = kgnilnd_check_conn_timeouts_locked(conn);
2707 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2708 /* simulate a RX CLOSE after the timeout but before
2709 * the scheduler thread gets it */
2710 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2711 conn->gnc_peer_error = -ETIMEDOUT;
2714 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2716 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2717 CDEBUG(D_WARNING, "%s conn timed out, will "
2718 "reconnect upon request from peer\n",
2719 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2721 /* Once we mark closed, any of the scheduler threads could
2722 * get it and move through before we hit the fail loc code */
2723 kgnilnd_close_conn_locked(conn, rc);
2725 /* first_rx is used to decide when to release a conn from purgatory.
2727 first_rx = conn->gnc_first_rx;
2731 /* now regardless of starting new conn, find tx on peer queue that
2732 * are old and smell bad - do this first so we don't trigger
2733 * reconnect on empty queue if we timeout all */
2734 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2735 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2737 LCONSOLE_INFO("could not send to %s due to connection"
2738 " setup failure after %lu seconds\n",
2739 libcfs_nid2str(peer->gnp_nid),
2740 cfs_duration_sec(jiffies - tx->tx_qtime));
2742 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2744 list_add_tail(&tx->tx_list, todie);
2749 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2750 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2751 count, peer, libcfs_nid2str(peer->gnp_nid));
2752 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2753 if (list_empty(&peer->gnp_tx_queue) ||
2754 peer->gnp_connecting == GNILND_PEER_KILL) {
2755 /* we pass down todie to use a common function - but we know there are
2757 kgnilnd_cancel_peer_connect_locked(peer, todie);
2761 /* Don't reconnect if we are still trying to clear out old conns.
2762 * This prevents us sending traffic on the new mbox before ensuring we are done
2763 * with the old one */
2764 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2765 (atomic_read(&peer->gnp_dirty_eps) == 0);
2767 /* fast reconnect after a timeout */
2768 to_reconn = !conn &&
2769 (peer->gnp_last_errno == -ETIMEDOUT) &&
2770 *kgnilnd_tunables.kgn_fast_reconn;
2772 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2773 * to be sent, we'll check the reconnect interval and fire up a new
2774 * connection request */
2777 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2778 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2779 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2781 CDEBUG(D_NET, "starting connect to %s\n",
2782 libcfs_nid2str(peer->gnp_nid));
2783 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2784 "Peer was idle and we have a write_lock, state issue %d\n",
2785 peer->gnp_connecting);
2787 peer->gnp_connecting = GNILND_PEER_CONNECT;
2788 kgnilnd_peer_addref(peer); /* extra ref for connd */
2790 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2791 list_add_tail(&peer->gnp_connd_list,
2792 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2793 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2795 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2798 /* fail_loc to allow us to delay release of purgatory */
2799 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2802 /* This check allows us to verify that the new conn is actually being used. This allows us to
2803 * pull the old conns out of purgatory if they have actually seen traffic.
2804 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2807 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2808 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2809 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2813 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2814 /* check for purgatory timeouts */
2815 if (conn->gnc_in_purgatory) {
2816 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2817 * that way the next time we check it we can detach it from purgatory
2820 if (conn->gnc_state != GNILND_CONN_DONE) {
2821 /* Skip over conns that are currently not DONE. If they arent already scheduled
2822 * for completion something in the state machine is broken.
2827 /* We only detach a conn that is in purgatory if we have received a close message,
2828 * we have a new valid connection that has successfully received data, or an admin
2829 * command tells us we need to detach.
2832 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2833 unsigned long waiting;
2835 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2837 /* C.E: The remote peer is expected to close the
2838 * connection (see kgnilnd_check_conn_timeouts)
2839 * via the reaper thread and nuke out the MDD and
2840 * FMA resources after conn->gnc_timeout has expired
2841 * without an FMA RX */
2842 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2843 " held resources\n",
2844 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2845 cfs_duration_sec(waiting));
2847 kgnilnd_detach_purgatory_locked(conn, souls);
2849 purgatory_conn_cnt++;
2854 /* If we have too many connections in purgatory we could run out of
2855 * resources. Limit the number of connections to a tunable number,
2856 * clean up to the minimum all in one fell swoop... there are
2857 * situations where dvs will retry tx's and we can eat up several
2858 * hundread connection requests at once.
2860 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2861 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2863 if (conn->gnc_in_purgatory &&
2864 conn->gnc_state == GNILND_CONN_DONE) {
2865 CDEBUG(D_NET, "Dropping Held resource due to"
2866 " resource limits being hit\n");
2867 kgnilnd_detach_purgatory_locked(conn, souls);
2869 if (purgatory_conn_cnt-- <
2870 *kgnilnd_tunables.kgn_max_purgatory)
2878 kgnilnd_reaper_check(int idx)
2880 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2881 struct list_head *ctmp, *ctmpN;
2882 LIST_HEAD(geriatrics);
2885 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2887 list_for_each_safe(ctmp, ctmpN, peers) {
2888 kgn_peer_t *peer = NULL;
2890 /* don't timeout stuff if the network is mucked or shutting down */
2891 if (kgnilnd_check_hw_quiesce()) {
2894 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2896 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
2899 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2901 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
2902 kgnilnd_release_purgatory_list(&souls);
2906 kgnilnd_update_reaper_timeout(long timeout)
2908 LASSERT(timeout > 0);
2910 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2912 if (timeout < kgnilnd_data.kgn_new_min_timeout)
2913 kgnilnd_data.kgn_new_min_timeout = timeout;
2915 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2919 kgnilnd_reaper_poke_with_stick(cfs_timer_cb_arg_t arg)
2921 wake_up(&kgnilnd_data.kgn_reaper_waitq);
2925 kgnilnd_reaper(void *arg)
2930 unsigned long next_check_time = jiffies;
2931 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
2932 struct timer_list timer;
2935 /* all gnilnd threads need to run fairly urgently */
2936 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2937 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2939 while (!kgnilnd_data.kgn_shutdown) {
2940 /* I wake up every 'p' seconds to check for timeouts on some
2941 * more peers. I try to check every connection 'n' times
2942 * within the global minimum of all keepalive and timeout
2943 * intervals, to ensure I attend to every connection within
2944 * (n+1)/n times its timeout intervals. */
2945 const int p = GNILND_REAPER_THREAD_WAKE;
2946 const int n = GNILND_REAPER_NCHECKS;
2948 /* to quiesce or to not quiesce, that is the question */
2949 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2950 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2951 KGNILND_SPIN_QUIESCE;
2952 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2955 /* careful with the jiffy wrap... */
2956 timeout = (long)(next_check_time - jiffies);
2959 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
2960 TASK_INTERRUPTIBLE);
2961 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2962 cfs_timer_setup(&timer, kgnilnd_reaper_poke_with_stick,
2963 next_check_time, 0);
2964 mod_timer(&timer, (long) jiffies + timeout);
2966 /* check flag variables before committing */
2967 if (!kgnilnd_data.kgn_shutdown &&
2968 !kgnilnd_data.kgn_quiesce_trigger) {
2969 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2970 timeout, cfs_duration_sec(timeout));
2972 CDEBUG(D_INFO, "awake after schedule\n");
2975 del_singleshot_timer_sync(&timer);
2976 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2977 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
2981 /* new_min_timeout is set from the conn timeouts and keepalive
2982 * this should end up with a min timeout of
2983 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
2984 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
2985 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
2986 CDEBUG(D_NET, "Set new min timeout %ld\n",
2987 current_min_timeout);
2990 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2992 /* Compute how many table entries to check now so I get round
2993 * the whole table fast enough given that I do this at fixed
2994 * intervals of 'p' seconds) */
2995 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
2996 if (kgnilnd_data.kgn_new_min_timeout > n * p)
2997 chunk = (chunk * n * p) /
2998 kgnilnd_data.kgn_new_min_timeout;
3001 for (i = 0; i < chunk; i++) {
3002 kgnilnd_reaper_check(hash_index);
3003 hash_index = (hash_index + 1) %
3004 *kgnilnd_tunables.kgn_peer_hash_size;
3006 next_check_time = (long) jiffies + cfs_time_seconds(p);
3007 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3009 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3012 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3014 kgnilnd_thread_fini();
3019 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3020 unsigned niov, offset, nob;
3021 struct bio_vec *kiov;
3022 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3023 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3026 lnet_copy_flat2kiov(
3029 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3031 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3038 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3041 gni_post_descriptor_t *desc;
3043 kgn_tx_ev_id_t ev_id;
3045 int should_retry, rc;
3046 long num_processed = 0;
3047 kgn_conn_t *conn = NULL;
3048 kgn_tx_t *tx = NULL;
3049 kgn_rdma_desc_t *rdesc;
3054 /* make sure we don't keep looping if we need to reset */
3055 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3056 return num_processed;
3058 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3060 /* we didn't get the mutex, so return that there is still work
3064 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3065 /* a bit gross - but we need a good way to test for
3066 * delayed RDMA completions and the easiest way to do
3067 * that is to delay the RDMA CQ events */
3068 rrc = GNI_RC_NOT_DONE;
3070 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3073 if (rrc == GNI_RC_NOT_DONE) {
3074 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3075 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3076 dev->gnd_id, num_processed);
3077 return num_processed;
3079 dev->gnd_sched_alive = jiffies;
3082 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3083 "this is bad, somehow our credits didn't protect us"
3084 " from CQ overrun\n");
3085 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3086 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3087 event_data, GNI_CQ_GET_TYPE(event_data));
3089 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3091 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3093 /* XXX Nic: Need better error handling here... */
3094 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3095 (rrc == GNI_RC_TRANSACTION_ERROR),
3098 ev_id.txe_cookie = desc->post_id;
3100 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3102 if (conn == NULL || tx == NULL) {
3103 /* either conn or tx was already nuked and this is a "late"
3104 * completion, so drop it */
3108 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3109 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3110 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3111 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3112 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3114 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3116 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3117 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3118 tx->tx_msg.gnm_u.completion.gncm_retval);
3122 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3123 if (tx->tx_buffer_copy != NULL)
3124 kgnilnd_recv_bte_get(tx);
3125 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3128 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3129 if (tx->tx_buffer_copy != NULL)
3130 kgnilnd_recv_bte_get(tx);
3131 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3134 /* remove from rdmaq */
3135 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3136 spin_lock(&conn->gnc_list_lock);
3137 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3138 spin_unlock(&conn->gnc_list_lock);
3139 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3141 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3142 event_data = 1LL << 48;
3146 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3147 atomic_inc(&dev->gnd_rdma_ntx);
3148 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3149 /* transaction succeeded, add into fmaq */
3150 kgnilnd_queue_tx(conn, tx);
3151 kgnilnd_peer_alive(conn->gnc_peer);
3153 /* drop ref from kgnilnd_validate_tx_ev_id */
3154 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3155 kgnilnd_conn_decref(conn);
3160 /* fall through to the TRANSACTION_ERROR case */
3163 /* get stringified version for log messages */
3164 kgnilnd_cq_error_str(event_data, &err_str, 256);
3165 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3167 /* make sure we are not off in the weeds with this tx */
3168 if (tx->tx_retrans >
3169 *kgnilnd_tunables.kgn_max_retransmits) {
3170 GNIDBG_TX(D_NETERROR, tx,
3171 "giving up on TX, too many retries", NULL);
3175 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3176 should_retry ? "transient" : "unrecoverable", err_str);
3178 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3179 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3180 rdesc = &tx->tx_putinfo.gnpam_desc;
3181 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3182 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3184 rdesc = &tx->tx_getinfo.gngm_desc;
3185 rnob = tx->tx_lntmsg[0]->msg_len;
3186 rcookie = tx->tx_getinfo.gngm_cookie;
3191 tx->tx_msg.gnm_type,
3195 kgnilnd_nak_rdma(conn,
3196 tx->tx_msg.gnm_type,
3199 tx->tx_msg.gnm_srcnid);
3200 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3201 kgnilnd_close_conn(conn, -ECOMM);
3204 /* drop ref from kgnilnd_validate_tx_ev_id */
3205 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3206 kgnilnd_conn_decref(conn);
3211 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3215 kgn_tx_ev_id_t ev_id;
3216 kgn_tx_t *tx = NULL;
3217 kgn_conn_t *conn = NULL;
3218 int queued_fma, saw_reply, rc;
3219 long num_processed = 0;
3220 struct list_head *ctmp, *ctmpN;
3223 /* make sure we don't keep looping if we need to reset */
3224 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3225 return num_processed;
3228 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3230 /* we didn't get the mutex, so return that there is still work
3235 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3236 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3238 if (rrc == GNI_RC_NOT_DONE) {
3240 "SMSG send CQ %d not ready (data %#llx) "
3241 "processed %ld\n", dev->gnd_id, event_data,
3244 if (num_processed > 0) {
3245 spin_lock(&dev->gnd_lock);
3246 if (!list_empty(&dev->gnd_delay_conns)) {
3247 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3248 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3249 list_del_init(&conn->gnc_delaylist);
3250 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3251 kgnilnd_schedule_conn_nolock(conn);
3253 spin_unlock(&dev->gnd_lock);
3254 kgnilnd_schedule_device(dev);
3256 spin_unlock(&dev->gnd_lock);
3259 return num_processed;
3262 dev->gnd_sched_alive = jiffies;
3265 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3266 "this is bad, somehow our credits didn't "
3267 "protect us from CQ overrun\n");
3268 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3269 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3270 event_data, GNI_CQ_GET_TYPE(event_data));
3272 /* if SMSG couldn't handle an error, time for conn to die */
3273 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3276 /* need to take the write_lock to ensure atomicity
3277 * on the conn state if we need to close it */
3278 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3279 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3281 /* Conn was destroyed? */
3283 "SMSG CQID lookup %#llx failed\n",
3284 GNI_CQ_GET_INST_ID(event_data));
3285 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3289 kgnilnd_cq_error_str(event_data, &err_str, 256);
3290 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3291 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3293 kgnilnd_close_conn_locked(conn, -ECOMM);
3295 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3297 /* no need to process rest of this tx -
3298 * it is getting canceled */
3302 /* fall through to GNI_RC_SUCCESS case */
3303 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3305 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3306 if (conn == NULL || tx == NULL) {
3307 /* either conn or tx was already nuked and this is a "late"
3308 * completion, so drop it */
3312 tx->tx_conn->gnc_last_tx_cq = jiffies;
3313 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3314 set_mb(conn->gnc_last_noop_cq, jiffies);
3317 /* lock tx_list_state and tx_state */
3318 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3319 spin_lock(&tx->tx_conn->gnc_list_lock);
3321 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3322 "state not GNILND_TX_LIVE_FMAQ", NULL);
3323 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3324 "not waiting for completion", NULL);
3326 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3329 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3331 /* This will trigger other FMA sends that were
3332 * pending this completion */
3333 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3335 /* we either did not expect reply or we already got it */
3336 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3338 spin_unlock(&tx->tx_conn->gnc_list_lock);
3339 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3342 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3344 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3345 kgnilnd_schedule_conn(conn);
3348 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3349 * If saw_reply is true we know that the tx is safe to use as the other thread
3350 * is already finished with it.
3354 /* no longer need to track on the live_fmaq */
3355 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3357 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3358 /* we already got reply & were waiting for
3359 * completion of initial send */
3360 /* to initiate RDMA transaction */
3361 GNIDBG_TX(D_NET, tx,
3362 "Pending RDMA 0x%p type 0x%02x",
3363 tx->tx_msg.gnm_type);
3364 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3365 rc = kgnilnd_send_mapped_tx(tx, 0);
3366 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3368 /* we are done with this tx */
3369 GNIDBG_TX(D_NET, tx,
3370 "Done with tx type 0x%02x",
3371 tx->tx_msg.gnm_type);
3372 kgnilnd_tx_done(tx, tx->tx_rc);
3376 /* drop ref from kgnilnd_validate_tx_ev_id */
3377 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3378 kgnilnd_conn_decref(conn);
3380 /* if we are waiting for a REPLY, we'll handle the tx then */
3381 } /* end for loop */
3385 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3390 long num_processed = 0;
3391 struct list_head *conns;
3392 struct list_head *tmp;
3396 /* make sure we don't keep looping if we need to reset */
3397 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3398 return num_processed;
3401 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3403 /* we didn't get the mutex, so return that there is still work
3407 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3408 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3410 if (rrc == GNI_RC_NOT_DONE) {
3411 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3413 dev->gnd_id, event_data, num_processed);
3414 return num_processed;
3416 dev->gnd_sched_alive = jiffies;
3419 /* this is the only CQ that can really handle transient
3421 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3422 rrc = cfs_fail_val ? cfs_fail_val
3423 : GNI_RC_ERROR_RESOURCE;
3424 if (rrc == GNI_RC_ERROR_RESOURCE) {
3425 /* set overrun too */
3426 event_data |= (1UL << 63);
3427 LASSERTF(GNI_CQ_OVERRUN(event_data),
3428 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3431 /* sender should get error event too and take care
3432 of failed transaction by re-transmitting */
3433 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3434 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3438 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3439 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3440 conn = kgnilnd_cqid2conn_locked(
3441 GNI_CQ_GET_INST_ID(event_data));
3443 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3444 "failed, dropping event %#llx\n",
3445 GNI_CQ_GET_INST_ID(event_data),
3448 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3450 GNI_CQ_GET_INST_ID(event_data),
3451 conn, conn->gnc_peer ?
3452 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3455 conn->gnc_last_rx_cq = jiffies;
3457 /* stash first rx so we can clear out purgatory.
3459 if (conn->gnc_first_rx == 0) {
3460 conn->gnc_first_rx = jiffies;
3462 kgnilnd_peer_alive(conn->gnc_peer);
3463 kgnilnd_schedule_conn(conn);
3465 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3469 /* FMA CQ has overflowed: check ALL conns */
3470 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3471 "conns on device %d\n", dev->gnd_id);
3473 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3475 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3476 conns = &kgnilnd_data.kgn_conns[rc];
3478 list_for_each(tmp, conns) {
3479 conn = list_entry(tmp, kgn_conn_t,
3482 if (conn->gnc_device == dev) {
3483 kgnilnd_schedule_conn(conn);
3484 conn->gnc_last_rx_cq = jiffies;
3488 /* don't block write lockers for too long... */
3489 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3494 /* try_map_if_full should only be used when processing TX from list of
3495 * backlog TX waiting on mappings to free up
3498 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3499 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3502 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3504 /* slight bit of race if multiple people calling, but at worst we'll have
3505 * order altered just a bit... which would not be determenistic anyways */
3506 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3508 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3510 /* We know that we have a GART reservation that should guarantee forward progress.
3511 * This means we don't need to take any extraordinary efforts if we are failing
3512 * mappings here - even if we are holding a very small number of these. */
3514 if (try_map_if_full || (rc == 0)) {
3515 rc = kgnilnd_map_buffer(tx);
3518 /* rc should be 0 if we mapped successfully here, if non-zero
3519 * we are queueing */
3521 /* if try_map_if_full set, they handle requeuing */
3522 if (unlikely(try_map_if_full)) {
3525 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3526 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3527 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3528 /* make sure we wake up sched to run this */
3529 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3530 /* return 0 as this is now queued for later sending */
3535 switch (tx->tx_msg.gnm_type) {
3539 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3540 * remote node where the RDMA will be started
3541 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3542 * be satisified. The rest of the errors are "hard" errors that require
3543 * upper layers to handle themselves.
3544 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3545 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3546 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3547 * GNILND_MSG_GET_DONE_REV.
3549 case GNILND_MSG_GET_REQ:
3550 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3551 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3552 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3553 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3554 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3555 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3556 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3558 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3559 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3561 case GNILND_MSG_PUT_ACK:
3562 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3563 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3564 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3565 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3567 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3568 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3571 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3572 case GNILND_MSG_PUT_DONE:
3573 case GNILND_MSG_PUT_REQ:
3574 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3575 &tx->tx_putinfo.gnpam_desc,
3576 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3577 tx->tx_putinfo.gnpam_dst_cookie);
3578 RETURN(try_map_if_full ? rc : 0);
3580 case GNILND_MSG_GET_DONE:
3581 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3582 &tx->tx_getinfo.gngm_desc,
3583 tx->tx_lntmsg[0]->msg_len,
3584 tx->tx_getinfo.gngm_cookie);
3585 RETURN(try_map_if_full ? rc : 0);
3587 case GNILND_MSG_PUT_REQ_REV:
3588 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3589 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3590 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3591 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3592 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3593 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3594 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3596 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3598 case GNILND_MSG_PUT_DONE_REV:
3599 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3600 &tx->tx_getinfo.gngm_desc,
3602 tx->tx_getinfo.gngm_cookie);
3603 RETURN(try_map_if_full ? rc : 0);
3605 case GNILND_MSG_GET_ACK_REV:
3606 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3607 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3608 /* LNET_GETS are a special case for parse */
3609 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3610 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3612 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3613 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3615 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3616 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3618 case GNILND_MSG_GET_DONE_REV:
3619 case GNILND_MSG_GET_REQ_REV:
3620 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3621 &tx->tx_putinfo.gnpam_desc,
3622 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3623 tx->tx_putinfo.gnpam_dst_cookie);
3624 RETURN(try_map_if_full ? rc : 0);
3632 kgnilnd_process_fmaq(kgn_conn_t *conn)
3635 kgn_tx_t *tx = NULL;
3636 void *buffer = NULL;
3637 unsigned int nob = 0;
3640 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3641 * However I will be rescheduled by an FMA completion event
3642 * when I eventually get some.
3643 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3644 * But it doesn't matter if I try to send a "real" message just
3645 * as I start closing because I'll get scheduled to send the
3648 /* Short circuit if the ep_handle is null we cant send anyway. */
3649 if (conn->gnc_ephandle == NULL)
3652 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3654 spin_lock(&conn->gnc_list_lock);
3656 if (list_empty(&conn->gnc_fmaq)) {
3657 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3659 spin_unlock(&conn->gnc_list_lock);
3661 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3662 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3663 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3664 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3665 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3667 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3668 conn->gnc_last_noop_cq,
3669 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3670 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3671 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3672 atomic_inc(&conn->gnc_sched_noop);
3673 set_mb(conn->gnc_last_noop_want, jiffies);
3675 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3678 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3682 rc = kgnilnd_set_tx_id(tx, conn);
3684 kgnilnd_tx_done(tx, rc);
3690 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3691 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3692 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3693 more_to_do = !list_empty(&conn->gnc_fmaq);
3694 spin_unlock(&conn->gnc_list_lock);
3697 /* if there is no real TX or no NOOP to send, bail */
3702 if (!tx->tx_retrans)
3703 tx->tx_cred_wait = jiffies;
3705 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3706 "tx with zero id", NULL);
3708 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3709 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3710 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3714 switch (tx->tx_msg.gnm_type) {
3718 case GNILND_MSG_NOOP:
3719 case GNILND_MSG_CLOSE:
3720 case GNILND_MSG_IMMEDIATE:
3721 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3722 buffer = tx->tx_buffer;
3726 case GNILND_MSG_GET_DONE:
3727 case GNILND_MSG_PUT_DONE:
3728 case GNILND_MSG_PUT_DONE_REV:
3729 case GNILND_MSG_GET_DONE_REV:
3730 case GNILND_MSG_PUT_NAK:
3731 case GNILND_MSG_GET_NAK:
3732 case GNILND_MSG_GET_NAK_REV:
3733 case GNILND_MSG_PUT_NAK_REV:
3734 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3737 case GNILND_MSG_PUT_REQ:
3738 case GNILND_MSG_GET_REQ_REV:
3739 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3741 case GNILND_MSG_PUT_ACK:
3742 case GNILND_MSG_PUT_REQ_REV:
3743 case GNILND_MSG_GET_ACK_REV:
3744 case GNILND_MSG_GET_REQ:
3745 /* This is really only to handle the retransmit of SMSG once these
3746 * two messages are setup in send_mapped_tx */
3747 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3751 if (likely(rc == 0)) {
3752 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3756 /* don't explicitly reschedule here - we are short credits and will rely on
3757 * kgnilnd_sendmsg to resched the conn if need be */
3759 } else if (rc < 0) {
3760 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3761 * almost certainly a software bug, but lets play nice with the other kids */
3762 kgnilnd_tx_done(tx, rc);
3763 /* just for fun, kick peer in arse - resetting conn might help to correct
3764 * this almost certainly buggy software caused return code */
3765 kgnilnd_close_conn(conn, rc);
3769 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3770 kgnilnd_schedule_conn(conn);
3775 kgnilnd_process_rdmaq(kgn_device_t *dev)
3780 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3784 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3785 unsigned long dead_bump;
3788 /* if we think we need to adjust, take lock to serialize and recheck */
3789 spin_lock(&dev->gnd_rdmaq_lock);
3790 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3791 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3793 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3795 /* roll the bucket forward */
3796 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3798 if (kgnilnd_data.kgn_rdmaq_override &&
3799 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3800 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3805 /* roll current outstanding forward to make sure we carry outstanding
3806 * committment forward
3807 * new_ok starts out as the whole interval value
3808 * - first subtract bytes_out from last interval, as that would push us over
3809 * strict limits for this interval
3810 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3812 * there is a small race here if someone is actively processing mappings and
3813 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3814 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3815 * - if this gives us problems in the future, we could use a read/write lock
3816 * to protect the resetting of these values */
3817 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3818 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3820 CDEBUG(D_NET, "resetting rdmaq bytes to %lld, deadline +%lu -> %lu, current out %lld\n",
3821 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3822 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
3824 spin_unlock(&dev->gnd_rdmaq_lock);
3827 spin_lock(&dev->gnd_rdmaq_lock);
3828 while (!list_empty(&dev->gnd_rdmaq)) {
3831 /* make sure we break out early on quiesce */
3832 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3833 /* always break with lock held - we unlock outside loop */
3837 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3838 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3841 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3842 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3843 /* if conn is dying, mark tx in tx_ref_table for
3844 * kgnilnd_complete_closed_conn to finish up */
3845 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3847 /* tx was moved to DYING, get next */
3850 spin_unlock(&dev->gnd_rdmaq_lock);
3852 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3853 spin_lock(&dev->gnd_rdmaq_lock);
3856 /* no ticket! add back to head */
3857 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3858 /* clear found_work so scheduler threads wait for timer */
3862 /* TX is GO for launch */
3863 tx->tx_qtime = jiffies;
3864 kgnilnd_send_mapped_tx(tx, 0);
3868 spin_unlock(&dev->gnd_rdmaq_lock);
3874 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3876 __swab64s(&d->gnrd_key.qword1);
3877 __swab64s(&d->gnrd_key.qword2);
3878 __swab64s(&d->gnrd_addr);
3879 __swab32s(&d->gnrd_nob);
3882 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3883 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3886 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3888 kgn_tx_ev_id_t ev_id;
3891 /* we use the cookie from the original TX, so we can find the match
3892 * by parsing that and using the txe_idx */
3893 ev_id.txe_cookie = cookie;
3895 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3898 /* check tx to make sure kgni didn't eat it */
3899 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3900 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3902 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
3903 (tx->tx_id.txe_cookie = cookie)),
3904 "conn 0x%p->%s tx_ref_table hosed: wanted "
3905 "txe_cookie %#llx txe_idx %d "
3906 "found tx %p cookie %#llx txe_idx %d\n",
3907 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
3908 cookie, ev_id.txe_idx,
3909 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
3911 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
3912 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
3913 "Unexpected TX type (%x, %x or %x) "
3914 "or state (%x, expected +%x) "
3915 "matched reply from %s\n",
3916 tx->tx_msg.gnm_type, type1, type2,
3917 tx->tx_state, GNILND_TX_WAITING_REPLY,
3918 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3920 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
3921 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
3927 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
3930 kgn_conn_t *conn = tx->tx_conn;
3931 __u64 nob = tx->tx_nob;
3932 __u32 physnop = tx->tx_phys_npages;
3933 int id = tx->tx_id.txe_smsg_id;
3934 int buftype = tx->tx_buftype;
3935 gni_mem_handle_t hndl;
3936 hndl.qword1 = tx->tx_map_key.qword1;
3937 hndl.qword2 = tx->tx_map_key.qword2;
3939 spin_lock(&conn->gnc_list_lock);
3941 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
3942 "not waiting for reply", NULL);
3945 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
3947 if (rc == -EFAULT) {
3948 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
3949 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
3951 if(*kgnilnd_tunables.kgn_efault_lbug) {
3952 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
3953 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
3955 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
3956 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
3957 kgnilnd_tx_state2str(tx->tx_list_state),
3958 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
3963 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
3964 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3965 /* sample under lock as follow on steps require gnc_list_lock
3966 * - or call kgnilnd_tx_done which requires no locks held over
3967 * call to lnet_finalize */
3970 spin_unlock(&conn->gnc_list_lock);
3973 kgnilnd_tx_done(tx, tx->tx_rc);
3978 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
3981 kgn_conn_t *conn = tx->tx_conn;
3983 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
3984 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
3986 /* the gncm_retval is passed in for PUTs */
3987 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
3988 msg->gnm_u.completion.gncm_retval);
3990 kgnilnd_complete_tx(tx, rc);
3994 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4002 kgn_peer_t *peer = conn->gnc_peer;
4005 __u16 tmp_cksum = 0, msg_cksum = 0;
4006 int repost = 1, saw_complete;
4007 unsigned long timestamp, newest_last_rx, timeout;
4011 /* Short circuit if the ep_handle is null.
4012 * It's likely that its about to be closed as stale.
4014 if (conn->gnc_ephandle == NULL)
4017 timestamp = jiffies;
4018 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4019 /* delay in jiffies - we are really concerned only with things that
4020 * result in a schedule() or really holding this off for long times .
4021 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4022 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4024 /* Resample current time as we have no idea how long it took to get the mutex */
4025 timestamp = jiffies;
4027 /* We check here when the last time we received an rx, we do this before
4028 * we call getnext in case the thread has been blocked for a while. If we
4029 * havent received an rx since our timeout value we close the connection
4030 * as we should assume the other side has closed the connection. This will
4031 * stop us from sending replies to a mailbox that is already in purgatory.
4034 timeout = cfs_time_seconds(conn->gnc_timeout);
4035 newest_last_rx = GNILND_LASTRX(conn);
4037 /* Error injection to validate that timestamp checking works and closing the conn */
4038 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4039 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4042 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4043 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4044 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4045 cfs_duration_sec(timestamp - newest_last_rx),
4046 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4047 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4049 kgnilnd_close_conn(conn, rc);
4053 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4055 if (rrc == GNI_RC_NOT_DONE) {
4056 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4057 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4061 /* Instead of asserting when we get mailbox corruption lets attempt to
4062 * close the conn and recover. We can put the conn/mailbox into
4063 * purgatory and let purgatory deal with the problem. If we see
4064 * this NETTERROR reported on production systems in large amounts
4065 * we will need to revisit the state machine to see if we can tighten
4066 * it up further to improve data protection.
4069 if (rrc == GNI_RC_INVALID_STATE) {
4070 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4071 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4072 "detected closing conn %p from peer %s\n", conn,
4073 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4075 kgnilnd_close_conn(conn, rc);
4079 LASSERTF(rrc == GNI_RC_SUCCESS,
4080 "bad rc %d on conn %p from peer %s\n",
4081 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4083 msg = (kgn_msg_t *)prefix;
4085 rx = kgnilnd_alloc_rx();
4087 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4088 kgnilnd_release_msg(conn);
4089 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4090 conn, libcfs_nid2str(peer->gnp_nid));
4094 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4096 timestamp = conn->gnc_last_rx;
4097 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4098 atomic_inc(&conn->gnc_rx_seq);
4100 conn->gnc_last_rx = jiffies;
4101 /* stash first rx so we can clear out purgatory
4103 if (conn->gnc_first_rx == 0)
4104 conn->gnc_first_rx = jiffies;
4106 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4107 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4108 kgnilnd_peer_alive(conn->gnc_peer);
4111 rx->grx_conn = conn;
4113 ktime_get_ts64(&rx->grx_received);
4115 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4118 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4124 kgnilnd_net_decref(net);
4127 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4128 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4130 /* XXX Nic: Do we need to swab cksum */
4131 if (msg->gnm_cksum != 0) {
4132 msg_cksum = msg->gnm_cksum;
4134 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4136 if (tmp_cksum != msg_cksum) {
4137 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4138 tmp_cksum, msg_cksum);
4139 kgnilnd_dump_msg(D_BUFFS, msg);
4144 /* restore checksum for future debug messages */
4145 msg->gnm_cksum = tmp_cksum;
4147 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4148 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4149 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4150 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4155 __swab32s(&msg->gnm_magic);
4156 __swab16s(&msg->gnm_version);
4157 __swab16s(&msg->gnm_type);
4158 __swab64s(&msg->gnm_srcnid);
4159 __swab64s(&msg->gnm_connstamp);
4160 __swab32s(&msg->gnm_seq);
4162 /* NB message type checked below; NOT here... */
4163 switch (msg->gnm_type) {
4164 case GNILND_MSG_GET_ACK_REV:
4165 case GNILND_MSG_PUT_ACK:
4166 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4169 case GNILND_MSG_PUT_REQ_REV:
4170 case GNILND_MSG_GET_REQ:
4171 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4179 if (msg->gnm_version != GNILND_MSG_VERSION) {
4180 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4181 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4186 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4187 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4188 libcfs_nid2str(msg->gnm_srcnid),
4189 libcfs_nid2str(peer->gnp_nid));
4194 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4195 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4196 " expected) from %s",
4197 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4198 libcfs_nid2str(peer->gnp_nid));
4203 if (msg->gnm_seq != seq) {
4204 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4205 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4210 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4212 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4213 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4214 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4215 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4216 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4217 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4218 /* double check state with lock held */
4219 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4220 /* only error if we are not already closing */
4221 if (conn->gnc_peer_error == -ETIMEDOUT) {
4222 unsigned long now = jiffies;
4223 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4225 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4226 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4227 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4229 cfs_duration_sec(now - timestamp),
4230 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4231 atomic_read(&conn->gnc_tx_seq),
4232 cfs_duration_sec(now - conn->gnc_last_tx),
4233 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4234 cfs_duration_sec(now - conn->gnc_last_noop_want),
4235 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4236 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4237 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4238 cfs_duration_sec(now - conn->gnc_last_sched_do),
4239 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4241 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4243 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4247 if (conn->gnc_close_recvd) {
4248 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4249 kgnilnd_msgtype2str(msg->gnm_type),
4250 msg->gnm_type, conn->gnc_close_recvd,
4251 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4256 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4257 /* XXX Nic: log message received on bad connection state */
4261 switch (msg->gnm_type) {
4262 case GNILND_MSG_NOOP:
4263 /* Nothing to do; just a keepalive */
4266 case GNILND_MSG_IMMEDIATE:
4267 /* only get SMSG payload for IMMEDIATE */
4268 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4269 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4270 msg->gnm_srcnid, rx, 0);
4273 case GNILND_MSG_GET_REQ_REV:
4274 case GNILND_MSG_PUT_REQ:
4275 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4276 msg->gnm_srcnid, rx, 1);
4279 case GNILND_MSG_GET_NAK_REV:
4280 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4281 msg->gnm_u.completion.gncm_cookie);
4285 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4287 case GNILND_MSG_PUT_NAK:
4288 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4289 msg->gnm_u.completion.gncm_cookie);
4293 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4295 case GNILND_MSG_PUT_ACK:
4296 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4297 msg->gnm_u.putack.gnpam_src_cookie);
4301 /* store putack data for later: deferred rdma or re-try */
4302 tx->tx_putinfo = msg->gnm_u.putack;
4305 spin_lock(&tx->tx_conn->gnc_list_lock);
4307 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4308 "not waiting for reply", NULL);
4310 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4312 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4313 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4314 /* sample under lock as follow on steps require gnc_list_lock
4315 * - or call kgnilnd_tx_done which requires no locks held over
4316 * call to lnet_finalize */
4319 /* cannot launch rdma if still waiting for fma-msg completion */
4320 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4321 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4322 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4324 spin_unlock(&tx->tx_conn->gnc_list_lock);
4327 rc = kgnilnd_send_mapped_tx(tx, 0);
4329 kgnilnd_tx_done(tx, rc);
4332 case GNILND_MSG_GET_ACK_REV:
4333 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4334 msg->gnm_u.putack.gnpam_src_cookie);
4338 /* store putack data for later: deferred rdma or re-try */
4339 tx->tx_putinfo = msg->gnm_u.putack;
4341 spin_lock(&tx->tx_conn->gnc_list_lock);
4343 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4344 "not waiting for reply", NULL);
4346 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4348 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4349 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4350 /* sample under lock as follow on steps require gnc_list_lock
4351 * - or call kgnilnd_tx_done which requires no locks held over
4352 * call to lnet_finalize */
4355 /* cannot launch rdma if still waiting for fma-msg completion */
4356 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4357 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4358 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4360 spin_unlock(&tx->tx_conn->gnc_list_lock);
4363 rc = kgnilnd_send_mapped_tx(tx, 0);
4365 kgnilnd_tx_done(tx, rc);
4368 case GNILND_MSG_PUT_DONE:
4369 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4370 msg->gnm_u.completion.gncm_cookie);
4374 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4375 "bad tx buftype %d", tx->tx_buftype);
4377 kgnilnd_finalize_rx_done(tx, msg);
4379 case GNILND_MSG_PUT_REQ_REV:
4380 case GNILND_MSG_GET_REQ:
4381 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4382 msg->gnm_srcnid, rx, 1);
4386 case GNILND_MSG_GET_NAK:
4387 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4388 msg->gnm_u.completion.gncm_cookie);
4392 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4393 "bad tx buftype %d", tx->tx_buftype);
4395 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4398 case GNILND_MSG_GET_DONE:
4399 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4400 msg->gnm_u.completion.gncm_cookie);
4404 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4405 "bad tx buftype %d", tx->tx_buftype);
4407 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4408 msg->gnm_u.completion.gncm_retval);
4410 kgnilnd_finalize_rx_done(tx, msg);
4412 case GNILND_MSG_GET_DONE_REV:
4413 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4414 msg->gnm_u.completion.gncm_cookie);
4418 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4419 "bad tx buftype %d", tx->tx_buftype);
4421 kgnilnd_finalize_rx_done(tx, msg);
4424 case GNILND_MSG_PUT_DONE_REV:
4425 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4426 msg->gnm_u.completion.gncm_cookie);
4431 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4432 "bad tx buftype %d", tx->tx_buftype);
4434 kgnilnd_finalize_rx_done(tx, msg);
4436 case GNILND_MSG_PUT_NAK_REV:
4437 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4438 msg->gnm_u.completion.gncm_cookie);
4443 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4444 "bad tx buftype %d", tx->tx_buftype);
4446 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4451 if (rc < 0) /* protocol/comms error */
4452 kgnilnd_close_conn(conn, rc);
4454 if (repost && rx != NULL) {
4455 kgnilnd_consume_rx(rx);
4458 /* we got an event so assume more there and call for reschedule */
4460 kgnilnd_schedule_conn(conn);
4464 /* Do the failure injections that we need to affect conn processing in the following function.
4465 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4466 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4469 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4473 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4477 /* short circuit out when not set */
4478 if (likely(!cfs_fail_loc)) {
4482 /* failure injection to test for stack reset clean ups */
4483 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4484 /* we can't rely on busy loops being nice enough to get the
4485 * stack reset triggered - it'd just spin on this conn */
4486 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4489 GOTO(did_fail_loc, rc);
4492 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4493 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4495 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4496 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4499 GOTO(did_fail_loc, rc);
4503 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4504 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4505 * will be closed by kgnilnd_complete_closed_conn.
4507 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4508 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4511 GOTO(did_fail_loc, rc);
4514 /* this one is a bit gross - we can't hold the mutex from process_conns
4515 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4516 * and moving onto finish_connect
4517 * so, we'll just set the rc - kgnilnd_process_conns will clear
4518 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4519 * and effectively getting this thread to sleep */
4520 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4523 GOTO(did_fail_loc, rc);
4531 kgnilnd_send_conn_close(kgn_conn_t *conn)
4535 /* we are closing the conn - we will try to send the CLOSE msg
4536 * but will not wait for anything else to flush */
4538 /* send the close if not already done so or received one */
4539 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4540 /* set close_sent regardless of the success of the
4541 * CLOSE message. We are going to try once and then
4542 * kick him out of the sandbox */
4543 conn->gnc_close_sent = 1;
4546 /* EP might be null already if remote side initiated a new connection.
4547 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4548 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4550 if (conn->gnc_ephandle != NULL) {
4553 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4555 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4556 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4557 tx->tx_qtime = jiffies;
4559 if (tx->tx_id.txe_idx == 0) {
4560 rc = kgnilnd_set_tx_id(tx, conn);
4562 kgnilnd_tx_done(tx, rc);
4566 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4569 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4570 kgnilnd_tx_done(tx, -EAGAIN);
4572 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4574 /* It wasnt sent and we dont care. */
4575 kgnilnd_tx_done(tx, rc);
4583 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4584 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4585 conn->gnc_state = GNILND_CONN_CLOSED;
4586 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4587 /* mark this conn as CLOSED now that we processed it
4588 * do after TX, so we can use CLOSING in asserts */
4592 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4593 /* simulate a RX CLOSE after the timeout but before
4594 * the scheduler thread gets it */
4595 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4596 conn->gnc_peer_error = -ETIMEDOUT;
4598 /* schedule to allow potential CLOSE and get the complete phase run */
4599 kgnilnd_schedule_conn(conn);
4603 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4608 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4609 int log_retrans, log_retrans_level;
4610 static int last_map_version;
4613 spin_lock(&dev->gnd_lock);
4614 if (list_empty(&dev->gnd_map_tx)) {
4615 /* if the list is empty make sure we dont have a timer running */
4616 del_singleshot_timer_sync(&dev->gnd_map_timer);
4617 spin_unlock(&dev->gnd_lock);
4621 dev->gnd_sched_alive = jiffies;
4623 /* we'll retry as fast as possible up to 25% of the limit, then we start
4624 * backing off until our map version changes - indicating we unmapped
4626 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4627 if (likely(dev->gnd_map_attempt == 0) ||
4628 time_after_eq(jiffies, dev->gnd_next_map) ||
4629 last_map_version != dev->gnd_map_version) {
4631 /* if this is our first attempt at mapping set last mapped to current
4632 * jiffies so we can timeout our attempt correctly.
4634 if (dev->gnd_map_attempt == 0)
4635 dev->gnd_last_map = jiffies;
4637 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4638 spin_unlock(&dev->gnd_lock);
4642 /* delete the previous timer if it exists */
4643 del_singleshot_timer_sync(&dev->gnd_map_timer);
4644 /* stash the last map version to let us know when a good one was seen */
4645 last_map_version = dev->gnd_map_version;
4647 /* we need to to take the lock and continually refresh the head of the list as
4648 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4649 * allowing them to squeeze in */
4651 while (!list_empty(&dev->gnd_map_tx)) {
4652 /* make sure we break out early on quiesce */
4653 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4654 /* always break with lock held - we unlock outside loop */
4658 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4660 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4663 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4664 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4665 /* if conn is dying, mark tx in tx_ref_table for
4666 * kgnilnd_complete_closed_conn to finish up */
4667 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4670 /* tx was moved to DYING, get next */
4674 spin_unlock(&dev->gnd_lock);
4675 rc = kgnilnd_send_mapped_tx(tx, 1);
4677 /* We made it! skip error handling.. */
4679 /* OK to continue on +ve errors as it won't get seen until
4680 * this function is called again - we operate on a copy of the original
4681 * list and not the live list */
4682 spin_lock(&dev->gnd_lock);
4683 /* reset map attempts back to zero we successfully
4684 * mapped so we can reset our timers */
4685 dev->gnd_map_attempt = 0;
4687 } else if (rc == -EAGAIN) {
4688 spin_lock(&dev->gnd_lock);
4689 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4690 spin_unlock(&dev->gnd_lock);
4691 GOTO(get_out_mapped, rc);
4692 } else if (rc != -ENOMEM) {
4693 /* carp, failure we can't handle */
4694 kgnilnd_tx_done(tx, rc);
4695 spin_lock(&dev->gnd_lock);
4696 /* reset map attempts back to zero we dont know what happened but it
4697 * wasnt a failed mapping
4699 dev->gnd_map_attempt = 0;
4703 /* time to handle the retry cases.. lock so we dont have 2 threads
4704 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4706 spin_lock(&dev->gnd_lock);
4707 dev->gnd_map_attempt++;
4708 if (dev->gnd_map_attempt < fast_remaps) {
4709 /* do nothing we just want it to go as fast as possible.
4710 * just set gnd_next_map to current jiffies so it will process
4711 * as fast as possible.
4713 dev->gnd_next_map = jiffies;
4715 /* Retry based on GNILND_MAP_RETRY_RATE */
4716 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4719 /* only log occasionally once we've retried fast_remaps */
4720 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4721 ((dev->gnd_map_attempt % fast_remaps) == 0);
4722 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4724 /* make sure we are not off in the weeds with this tx */
4725 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4726 GNIDBG_TX(D_NETERROR, tx,
4727 "giving up on TX, too many retries", NULL);
4728 spin_unlock(&dev->gnd_lock);
4729 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4730 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4731 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4733 tx->tx_putinfo.gnpam_dst_cookie,
4734 tx->tx_msg.gnm_srcnid);
4736 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4738 tx->tx_getinfo.gngm_cookie,
4739 tx->tx_msg.gnm_srcnid);
4741 kgnilnd_tx_done(tx, -ENOMEM);
4742 GOTO(get_out_mapped, rc);
4744 GNIDBG_TX(log_retrans_level, tx,
4745 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4746 "nq_map %d mdd# %d/%d GART %ld",
4747 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4748 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4749 atomic_read(&dev->gnd_nq_map),
4750 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4751 atomic64_read(&dev->gnd_nbytes_map));
4754 /* we need to stop processing the rest of the list, so add it back in */
4755 /* set timer to wake device when we need to schedule this tx */
4756 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4757 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4758 spin_unlock(&dev->gnd_lock);
4759 GOTO(get_out_mapped, rc);
4761 spin_unlock(&dev->gnd_lock);
4767 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4772 int error_inject = 0;
4776 spin_lock(&dev->gnd_lock);
4777 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4778 dev->gnd_sched_alive = jiffies;
4782 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4783 /* break with lock held */
4787 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4788 list_del_init(&conn->gnc_schedlist);
4790 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4793 if (!list_empty(&conn->gnc_delaylist))
4794 list_del_init(&conn->gnc_delaylist);
4795 spin_unlock(&dev->gnd_lock);
4797 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4799 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4800 conn_sched != GNILND_CONN_PROCESS,
4801 "conn %p on ready list but in bad state: %d\n",
4804 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4805 conn, kgnilnd_conn_state2str(conn));
4808 set_mb(conn->gnc_last_sched_do, jiffies);
4810 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4812 /* based on intent see if we should run again. */
4813 rc = kgnilnd_schedule_process_conn(conn, intent);
4815 /* drop ref from gnd_ready_conns */
4816 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4817 down_write(&dev->gnd_conn_sem);
4818 kgnilnd_conn_decref(conn);
4819 up_write(&dev->gnd_conn_sem);
4820 } else if (rc != 1) {
4821 kgnilnd_conn_decref(conn);
4823 /* clear this so that scheduler thread doesn't spin */
4825 /* break with lock held... */
4826 spin_lock(&dev->gnd_lock);
4830 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4831 down_write(&dev->gnd_conn_sem);
4833 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4834 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4835 /* If there are tx's currently in use in another
4836 * thread we dont want to complete the close
4837 * yet. Cycle this conn back through
4839 kgnilnd_schedule_conn(conn);
4841 kgnilnd_complete_closed_conn(conn);
4843 up_write(&dev->gnd_conn_sem);
4844 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4845 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4846 /* serialize SMSG CQs with ep_bind and smsg_release */
4847 down_write(&dev->gnd_conn_sem);
4848 kgnilnd_destroy_conn_ep(conn);
4849 up_write(&dev->gnd_conn_sem);
4850 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4851 /* if we need to do some CLOSE sending, etc done here do it */
4852 down_write(&dev->gnd_conn_sem);
4853 kgnilnd_send_conn_close(conn);
4854 kgnilnd_check_fma_rx(conn);
4855 up_write(&dev->gnd_conn_sem);
4856 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4857 /* start moving traffic if the old conns are cleared out */
4858 down_read(&dev->gnd_conn_sem);
4859 kgnilnd_check_fma_rx(conn);
4860 kgnilnd_process_fmaq(conn);
4861 up_read(&dev->gnd_conn_sem);
4864 rc = kgnilnd_schedule_process_conn(conn, 0);
4866 /* drop ref from gnd_ready_conns */
4867 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4868 down_write(&dev->gnd_conn_sem);
4869 kgnilnd_conn_decref(conn);
4870 up_write(&dev->gnd_conn_sem);
4871 } else if (rc != 1) {
4872 kgnilnd_conn_decref(conn);
4875 /* check list again with lock held */
4876 spin_lock(&dev->gnd_lock);
4879 /* If we are short circuiting due to timing we want to be scheduled
4880 * as soon as possible.
4882 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4885 spin_unlock(&dev->gnd_lock);
4891 kgnilnd_scheduler(void *arg)
4893 int threadno = (long)arg;
4896 unsigned long deadline = 0;
4899 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
4901 /* all gnilnd threads need to run fairly urgently */
4902 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
4903 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
4904 while (!kgnilnd_data.kgn_shutdown) {
4906 /* Safe: kgn_shutdown only set when quiescent */
4908 /* to quiesce or to not quiesce, that is the question */
4910 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4911 KGNILND_SPIN_QUIESCE;
4914 /* tracking for when thread goes AWOL */
4915 dev->gnd_sched_alive = jiffies;
4917 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
4918 (*kgnilnd_tunables.kgn_sched_timeout + 1));
4919 /* let folks know we are up and kicking
4920 * - they can use this for latency savings, etc
4921 * - only change if IRQ, if IDLE leave alone as that
4922 * schedule_device calls to put us back to IRQ */
4923 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
4925 down_read(&dev->gnd_conn_sem);
4926 /* always check these - they are super low cost */
4927 found_work += kgnilnd_check_fma_send_cq(dev);
4928 found_work += kgnilnd_check_fma_rcv_cq(dev);
4930 /* rdma CQ doesn't care about eps */
4931 found_work += kgnilnd_check_rdma_cq(dev);
4933 /* move some RDMA ? */
4934 found_work += kgnilnd_process_rdmaq(dev);
4936 /* map some pending RDMA requests ? */
4937 found_work += kgnilnd_process_mapped_tx(dev);
4939 /* the EP for a conn is not destroyed until all the references
4940 * to it are gone, so these checks should be safe
4941 * even if run in parallel with the CQ checking functions
4942 * _AND_ a thread that processes the CLOSED->DONE
4946 up_read(&dev->gnd_conn_sem);
4948 /* process all conns ready now */
4949 found_work += kgnilnd_process_conns(dev, deadline);
4951 /* do an eager check to avoid the IRQ disabling in
4952 * prepare_to_wait and friends */
4955 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
4956 time_before(jiffies, deadline)) {
4958 if ((busy_loops % 10) == 0) {
4959 /* tickle heartbeat and watchdog to ensure our
4960 * piggishness doesn't turn into heartbeat failure */
4961 touch_nmi_watchdog();
4967 /* if we got here, found_work was zero or busy_loops means we
4968 * need to take a break. We'll clear gnd_ready but we'll check
4969 * one last time if there is an IRQ that needs processing */
4971 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
4973 /* the first time this will go LOOP -> IDLE and let us do one final check
4974 * during which we might get an IRQ, then IDLE->IDLE and schedule()
4975 * - this might allow other threads to block us for a bit if they
4976 * try to get the mutex, but that is good as we'd need to wake
4977 * up soon to handle the CQ or other processing anyways */
4979 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
4981 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
4982 time_after_eq(jiffies, deadline)) {
4984 "yeilding: found_work %d busy_loops %d\n",
4985 found_work, busy_loops);
4987 /* use yield if we are bailing due to busy_loops
4988 * - this will ensure we wake up soonish. This closes
4989 * a race with kgnilnd_device_callback - where it'd
4990 * not call wake_up() because gnd_ready == 1, but then
4991 * we come down and schedule() because of busy_loops.
4992 * We'd not be woken up until something poked our waitq
4993 * again. yield() ensures we wake up without another
4994 * waitq poke in that case */
4995 atomic_inc(&dev->gnd_n_yield);
4996 kgnilnd_data.kgn_last_condresched = jiffies;
4998 CDEBUG(D_INFO, "awake after yeild\n");
4999 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5000 } else if (found_work == GNILND_DEV_IDLE) {
5001 /* busy_loops is low and there is nothing to do,
5002 * go to sleep and wait for a waitq poke */
5004 "scheduling: found_work %d busy_loops %d\n",
5005 found_work, busy_loops);
5006 atomic_inc(&dev->gnd_n_schedule);
5007 kgnilnd_data.kgn_last_scheduled = jiffies;
5009 CDEBUG(D_INFO, "awake after schedule\n");
5010 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5012 finish_wait(&dev->gnd_waitq, &wait);
5015 kgnilnd_thread_fini();