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(&dev->gnd_waitq);
89 void kgnilnd_schedule_device_timer(cfs_timer_cb_arg_t data)
91 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_map_timer);
93 kgnilnd_schedule_device(dev);
96 void kgnilnd_schedule_device_timer_rd(cfs_timer_cb_arg_t data)
98 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_rdmaq_timer);
100 kgnilnd_schedule_device(dev);
104 kgnilnd_device_callback(__u32 devid, __u64 arg)
107 int index = (int) arg;
109 if (index >= kgnilnd_data.kgn_ndevs) {
110 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
111 * interrupt context. */
112 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
117 dev = &kgnilnd_data.kgn_devices[index];
118 /* just basic sanity */
119 if (dev->gnd_id == devid) {
120 kgnilnd_schedule_device(dev);
122 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
127 /* sched_intent values:
128 * < 0 : do not reschedule under any circumstances
129 * == 0: reschedule if someone marked him WANTS_SCHED
130 * > 0 : force a reschedule */
131 /* Return code 0 means it did not schedule the conn, 1
132 * means it successfully scheduled the conn.
136 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
140 /* move back to IDLE but save previous state.
141 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
142 * let the xchg there handle any racing callers to get it
143 * onto gnd_ready_conns */
145 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
146 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
147 conn_sched == GNILND_CONN_PROCESS,
148 "conn %p after process in bad state: %d\n",
151 if (sched_intent >= 0) {
152 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
153 return kgnilnd_schedule_conn_refheld(conn, 1);
159 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
163 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
165 kgn_device_t *dev = conn->gnc_device;
169 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
170 /* we only care about the last person who marked want_sched since they
171 * are most likely the culprit
173 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
174 conn->gnc_sched_line = line;
175 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
176 * the chance to put it onto gnd_ready_conns.
177 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
178 * the conn in process_conns will take care of moving it back to
179 * SCHED when it is done processing */
181 if (sched == GNILND_CONN_IDLE) {
182 /* if the conn is already scheduled, we've already requested
183 * the scheduler thread wakeup */
185 /* Add a reference to the conn if we are not holding a reference
186 * already from the exisiting scheduler. We now use the same
187 * reference if we need to reschedule a conn while in a scheduler
190 kgnilnd_conn_addref(conn);
192 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
195 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
197 spin_lock(&dev->gnd_lock);
198 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
200 spin_unlock(&dev->gnd_lock);
201 set_mb(conn->gnc_last_sched_ask, jiffies);
204 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
208 /* make sure thread(s) going to process conns - but let it make
209 * separate decision from conn schedule */
211 kgnilnd_schedule_device(dev);
216 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
218 kgn_device_t *dev = conn->gnc_device;
220 spin_lock(&dev->gnd_lock);
221 if (list_empty(&conn->gnc_delaylist)) {
222 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
225 spin_unlock(&dev->gnd_lock);
227 kgnilnd_schedule_device(dev);
232 kgnilnd_schedule_dgram(kgn_device_t *dev)
236 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
237 if (wake != GNILND_DGRAM_SCHED) {
238 wake_up(&dev->gnd_dgram_waitq);
240 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
245 kgnilnd_free_tx(kgn_tx_t *tx)
247 /* taken from kgnilnd_tx_add_state_locked */
249 LASSERTF((tx->tx_list_p == NULL &&
250 tx->tx_list_state == GNILND_TX_ALLOCD) &&
251 list_empty(&tx->tx_list),
252 "tx %p with bad state %s (list_p %p) tx_list %s\n",
253 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
254 list_empty(&tx->tx_list) ? "empty" : "not empty");
256 atomic_dec(&kgnilnd_data.kgn_ntx);
258 /* we only allocate this if we need to */
259 if (tx->tx_phys != NULL) {
260 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
261 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
262 GNILND_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
265 /* Only free the buffer if we used it */
266 if (tx->tx_buffer_copy != NULL) {
267 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
268 tx->tx_buffer_copy = NULL;
269 CDEBUG(D_MALLOC, "vfreed buffer2\n");
272 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
274 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
275 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
279 kgnilnd_alloc_tx (void)
283 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
286 tx = kmem_cache_zalloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
288 CERROR("failed to allocate tx\n");
291 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
294 /* setup everything here to minimize time under the lock */
295 tx->tx_buftype = GNILND_BUF_NONE;
296 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
297 INIT_LIST_HEAD(&tx->tx_list);
298 INIT_LIST_HEAD(&tx->tx_map_list);
299 tx->tx_list_state = GNILND_TX_ALLOCD;
301 atomic_inc(&kgnilnd_data.kgn_ntx);
306 /* csum_fold needs to be run on the return value before shipping over the wire */
307 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
309 /* we don't use offset as every one is passing a buffer reference that already
310 * includes the offset into the base address.
313 kgnilnd_cksum(void *ptr, size_t nob)
317 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
319 /* don't use magic 'no checksum' value */
323 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
330 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
331 unsigned int offset, unsigned int nob, int dump_blob)
337 unsigned int fraglen;
343 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
344 kiov, nkiov, offset, nob, dump_blob);
346 /* if loops changes, please change kgnilnd_setup_phys_buffer */
348 while (offset >= kiov->bv_len) {
349 offset -= kiov->bv_len;
355 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
356 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
358 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
359 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
361 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
362 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
364 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
365 odd, kiov[0].bv_len, offset, nob);
367 for (i = 0; i < nkiov; i++) {
368 pages[i] = kiov[i].bv_page;
371 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
373 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
375 /* return zero to avoid killing tx - we'll just get warning on console
376 * when remote end sees zero checksum */
379 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
381 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
386 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
387 (void *)addr + kiov[0].bv_offset +
390 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
391 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
395 fraglen = min(kiov->bv_len - offset, nob);
397 /* make dang sure we don't send a bogus checksum if somehow we get
398 * an odd length fragment on anything but the last entry in a kiov -
399 * we know from kgnilnd_setup_rdma_buffer that we can't have non
400 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
401 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
402 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
403 fraglen, nkiov, nob, kiov->bv_len,
406 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
408 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
411 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
412 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
413 addr, fraglen, offset);
418 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
420 kunmap(kiov->bv_page);
427 /* iov must not run out before end of data */
428 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
433 retsum = csum_fold(cksum);
435 /* don't use magic 'no checksum' value */
439 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
445 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
447 msg->gnm_magic = GNILND_MSG_MAGIC;
448 msg->gnm_version = GNILND_MSG_VERSION;
449 msg->gnm_type = type;
450 msg->gnm_payload_len = 0;
451 msg->gnm_srcnid = source;
452 /* gnm_connstamp gets set when FMA is sent */
453 /* gnm_srcnid is set on creation via function argument
454 * The right interface/net and nid is passed in when the message
460 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
462 kgn_tx_t *tx = kgnilnd_alloc_tx();
465 kgnilnd_init_msg(&tx->tx_msg, type, source);
467 CERROR("couldn't allocate new tx type %s!\n",
468 kgnilnd_msgtype2str(type));
475 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
481 case GNILND_MSG_GET_REQ:
482 case GNILND_MSG_GET_DONE:
483 nak_type = GNILND_MSG_GET_NAK;
485 case GNILND_MSG_PUT_REQ:
486 case GNILND_MSG_PUT_ACK:
487 case GNILND_MSG_PUT_DONE:
488 nak_type = GNILND_MSG_PUT_NAK;
490 case GNILND_MSG_PUT_REQ_REV:
491 case GNILND_MSG_PUT_DONE_REV:
492 nak_type = GNILND_MSG_PUT_NAK_REV;
494 case GNILND_MSG_GET_REQ_REV:
495 case GNILND_MSG_GET_ACK_REV:
496 case GNILND_MSG_GET_DONE_REV:
497 nak_type = GNILND_MSG_GET_NAK_REV;
500 CERROR("invalid msg type %s (%d)\n",
501 kgnilnd_msgtype2str(rx_type), rx_type);
504 /* only allow NAK on error and truncate to zero */
505 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
506 error, conn, cookie);
508 tx = kgnilnd_new_tx_msg(nak_type, source);
510 CNETERR("can't get TX to NAK RDMA to %s\n",
511 libcfs_nid2str(conn->gnc_peer->gnp_nid));
515 tx->tx_msg.gnm_u.completion.gncm_retval = error;
516 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
517 kgnilnd_queue_tx(conn, tx);
521 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
522 struct bio_vec *kiov,
523 unsigned int offset, unsigned int nob)
525 kgn_msg_t *msg = &tx->tx_msg;
528 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
529 * gni_smsg_send to send that as the payload */
531 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
534 tx->tx_buffer = NULL;
537 if (niov && niov > (nob >> PAGE_SHIFT))
538 niov = DIV_ROUND_UP(nob + offset + kiov->bv_offset,
541 while (offset >= kiov->bv_len) {
542 offset -= kiov->bv_len;
548 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
549 "bad niov %d msg %p kiov %p offset %d nob%d\n",
550 niov, msg, kiov, offset, nob);
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 GNILND_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) == GNILND_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... */
1666 case GNILND_MSG_NOOP:
1667 /* Just make sure this goes out first for this conn */
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_processid *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(lnet_nid_to_nid4(&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(ntohl(target->nid.nid_addr[0]));
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, lnet_nid_to_nid4(&target->nid),
1743 up_read(&kgnilnd_data.kgn_net_rw_sem);
1747 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1748 up_read(&kgnilnd_data.kgn_net_rw_sem);
1750 /* search for peer again now that we have the lock
1751 * if we don't find it, add our new one to the list */
1752 kgnilnd_add_peer_locked(lnet_nid_to_nid4(&target->nid), new_peer,
1755 /* don't create a connection if the peer is not up */
1756 if (peer->gnp_state != GNILND_PEER_UP) {
1757 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1762 conn = kgnilnd_find_or_create_conn_locked(peer);
1764 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1765 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1770 /* oh hey, found a conn now... magical */
1771 kgnilnd_queue_tx(conn, tx);
1773 /* no conn, must be trying to connect - so we queue for now */
1774 tx->tx_qtime = jiffies;
1775 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1777 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1780 kgnilnd_tx_done(tx, rc);
1785 kgnilnd_rdma(kgn_tx_t *tx, int type,
1786 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1788 kgn_conn_t *conn = tx->tx_conn;
1789 unsigned long timestamp;
1790 gni_post_type_t post_type;
1793 unsigned int desc_nob = nob;
1794 void *desc_buffer = tx->tx_buffer;
1795 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1796 LASSERTF(kgnilnd_tx_mapped(tx),
1797 "unmapped tx %p\n", tx);
1798 LASSERTF(conn != NULL,
1799 "NULL conn on tx %p, naughty, naughty\n", tx);
1800 LASSERTF(nob <= sink->gnrd_nob,
1801 "nob %u > sink->gnrd_nob %d (%p)\n",
1802 nob, sink->gnrd_nob, sink);
1803 LASSERTF(nob <= tx->tx_nob,
1804 "nob %d > tx(%p)->tx_nob %d\n",
1805 nob, tx, tx->tx_nob);
1808 case GNILND_MSG_GET_DONE:
1809 case GNILND_MSG_PUT_DONE:
1810 post_type = GNI_POST_RDMA_PUT;
1812 case GNILND_MSG_GET_DONE_REV:
1813 case GNILND_MSG_PUT_DONE_REV:
1814 post_type = GNI_POST_RDMA_GET;
1817 CERROR("invalid msg type %s (%d)\n",
1818 kgnilnd_msgtype2str(type), type);
1821 if (post_type == GNI_POST_RDMA_GET) {
1822 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1823 * aligned. If the local buffer is not aligned correctly using the copy buffer
1824 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1825 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1826 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1827 * making sure the length and addr are aligned and copying the data into the correct location
1828 * once the transfer has completed.
1830 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1831 (sink->gnrd_addr & 3) ||
1834 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1836 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1838 if ((nob + tx->tx_offset) & 3) {
1839 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1840 atomic_inc(&kgnilnd_data.kgn_rev_length);
1842 desc_nob = (nob + tx->tx_offset);
1845 if (tx->tx_buffer_copy == NULL) {
1846 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1847 * and require at most we allocate a few extra bytes. */
1848 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1850 if (!tx->tx_buffer_copy) {
1851 /* allocation of buffer failed nak the rdma */
1852 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1853 kgnilnd_tx_done(tx, -EFAULT);
1856 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1857 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);
1858 if (rc != GNI_RC_SUCCESS) {
1859 /* Registration Failed nak rdma and kill the tx. */
1860 kgnilnd_vfree(tx->tx_buffer_copy,
1862 tx->tx_buffer_copy = NULL;
1863 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1864 kgnilnd_tx_done(tx, -EFAULT);
1868 desc_map_key = tx->tx_buffer_copy_map_key;
1869 desc_buffer = tx->tx_buffer_copy;
1873 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1874 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1875 tx->tx_rdma_desc.type = post_type;
1876 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1877 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1878 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1879 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1880 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1881 tx->tx_rdma_desc.length = desc_nob;
1882 tx->tx_nob_rdma = nob;
1883 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1884 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1885 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1886 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1887 /* prep final completion message */
1888 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1889 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1890 /* send actual size RDMA'd in retval */
1891 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1893 kgnilnd_compute_rdma_cksum(tx, nob);
1896 kgnilnd_queue_tx(conn, tx);
1900 /* Don't lie (CLOSE == RDMA idle) */
1901 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1902 tx, conn, conn->gnc_close_sent);
1904 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1905 "0x%x cookie:%#llx",
1906 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1908 /* set CQ dedicated for RDMA */
1909 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1911 timestamp = jiffies;
1912 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1913 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1914 /* delay in jiffies - we are really concerned only with things that
1915 * result in a schedule() or really holding this off for long times .
1916 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1917 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1919 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
1921 if (rrc == GNI_RC_ERROR_RESOURCE) {
1922 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1923 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1924 kgnilnd_unmap_buffer(tx, 0);
1926 if (tx->tx_buffer_copy != NULL) {
1927 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
1928 tx->tx_buffer_copy = NULL;
1931 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1932 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
1934 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1935 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1939 spin_lock(&conn->gnc_list_lock);
1940 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
1941 tx->tx_qtime = jiffies;
1942 spin_unlock(&conn->gnc_list_lock);
1943 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1944 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1946 /* XXX Nic: is this a place we should handle more errors for
1947 * robustness sake */
1948 LASSERT(rrc == GNI_RC_SUCCESS);
1953 kgnilnd_alloc_rx(void)
1957 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
1959 CERROR("failed to allocate rx\n");
1962 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
1965 /* no memset to zero, we'll always fill all members */
1969 /* release is to just free connection resources
1970 * we use this for the eager path after copying */
1972 kgnilnd_release_msg(kgn_conn_t *conn)
1975 unsigned long timestamp;
1977 CDEBUG(D_NET, "consuming %p\n", conn);
1979 timestamp = jiffies;
1980 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1981 /* delay in jiffies - we are really concerned only with things that
1982 * result in a schedule() or really holding this off for long times .
1983 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1984 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1986 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
1987 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1989 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
1990 GNIDBG_SMSG_CREDS(D_NET, conn);
1992 kgnilnd_schedule_conn(conn);
1996 kgnilnd_consume_rx(kgn_rx_t *rx)
1998 kgn_conn_t *conn = rx->grx_conn;
1999 kgn_msg_t *rxmsg = rx->grx_msg;
2001 /* if we are eager, free the cache alloc'd msg */
2002 if (unlikely(rx->grx_eager)) {
2003 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2004 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2006 /* release ref from eager_recv */
2007 kgnilnd_conn_decref(conn);
2009 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2010 kgnilnd_release_msg(conn);
2013 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2014 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2019 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2021 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2022 int type = lntmsg->msg_type;
2023 struct lnet_processid *target = &lntmsg->msg_target;
2024 int target_is_router = lntmsg->msg_target_is_router;
2025 int routing = lntmsg->msg_routing;
2026 unsigned int niov = lntmsg->msg_niov;
2027 struct bio_vec *kiov = lntmsg->msg_kiov;
2028 unsigned int offset = lntmsg->msg_offset;
2029 unsigned int nob = lntmsg->msg_len;
2030 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2031 kgn_net_t *net = ni->ni_data;
2034 /* '1' for consistency with code that checks !mpflag to restore */
2035 unsigned int mpflag = 1;
2036 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2038 /* NB 'private' is different depending on what we're sending.... */
2039 LASSERT(!in_interrupt());
2041 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2042 type, nob, niov, libcfs_idstr(target));
2044 LASSERTF(nob == 0 || niov > 0,
2045 "lntmsg %p nob %d niov %d\n", lntmsg, nob, 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,
2079 lnet_nid_to_nid4(&ni->ni_nid));
2081 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV,
2082 lnet_nid_to_nid4(&ni->ni_nid));
2088 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2089 lntmsg->msg_md->md_kiov,
2090 0, lntmsg->msg_md->md_length);
2092 CERROR("unable to setup buffer: %d\n", rc);
2093 kgnilnd_tx_done(tx, rc);
2098 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2099 if (tx->tx_lntmsg[1] == NULL) {
2100 CERROR("Can't create reply for GET to %s\n",
2101 libcfs_nidstr(&target->nid));
2102 kgnilnd_tx_done(tx, rc);
2107 tx->tx_lntmsg[0] = lntmsg;
2108 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2109 lnet_hdr_to_nid4(hdr, &tx->tx_msg.gnm_u.get.gngm_hdr);
2111 lnet_hdr_to_nid4(hdr,
2112 &tx->tx_msg.gnm_u.putreq.gnprm_hdr);
2114 /* rest of tx_msg is setup just before it is sent */
2115 kgnilnd_launch_tx(tx, net, target);
2117 case LNET_MSG_REPLY:
2119 /* to save on MDDs, we'll handle short kiov by vmap'ing
2120 * and sending via SMSG */
2121 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2124 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2125 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ,
2126 lnet_nid_to_nid4(&ni->ni_nid));
2128 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV,
2129 lnet_nid_to_nid4(&ni->ni_nid));
2136 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2139 kgnilnd_tx_done(tx, rc);
2144 tx->tx_lntmsg[0] = lntmsg;
2145 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2146 lnet_hdr_to_nid4(hdr,
2147 &tx->tx_msg.gnm_u.putreq.gnprm_hdr);
2149 lnet_hdr_to_nid4(hdr, &tx->tx_msg.gnm_u.get.gngm_hdr);
2151 /* rest of tx_msg is setup just before it is sent */
2152 kgnilnd_launch_tx(tx, net, target);
2156 /* send IMMEDIATE */
2158 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2159 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2161 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE,
2162 lnet_nid_to_nid4(&ni->ni_nid));
2168 rc = kgnilnd_setup_immediate_buffer(tx, niov, kiov, offset, nob);
2170 kgnilnd_tx_done(tx, rc);
2174 lnet_hdr_to_nid4(hdr, &tx->tx_msg.gnm_u.immediate.gnim_hdr);
2175 tx->tx_lntmsg[0] = lntmsg;
2176 kgnilnd_launch_tx(tx, net, target);
2179 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2181 memalloc_noreclaim_restore(mpflag);
2186 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2188 kgn_conn_t *conn = rx->grx_conn;
2189 kgn_msg_t *rxmsg = rx->grx_msg;
2190 unsigned int niov = lntmsg->msg_niov;
2191 struct bio_vec *kiov = lntmsg->msg_kiov;
2192 unsigned int offset = lntmsg->msg_offset;
2193 unsigned int nob = lntmsg->msg_len;
2198 switch (rxmsg->gnm_type) {
2199 case GNILND_MSG_PUT_REQ_REV:
2200 done_type = GNILND_MSG_PUT_DONE_REV;
2203 case GNILND_MSG_GET_REQ:
2204 done_type = GNILND_MSG_GET_DONE;
2207 CERROR("invalid msg type %s (%d)\n",
2208 kgnilnd_msgtype2str(rxmsg->gnm_type),
2213 tx = kgnilnd_new_tx_msg(done_type, lnet_nid_to_nid4(&ni->ni_nid));
2217 rc = kgnilnd_set_tx_id(tx, conn);
2221 rc = kgnilnd_setup_rdma_buffer(tx, niov, kiov, offset, nob);
2225 tx->tx_lntmsg[0] = lntmsg;
2226 tx->tx_getinfo = rxmsg->gnm_u.get;
2228 /* we only queue from kgnilnd_recv - we might get called from other contexts
2229 * and we don't want to block the mutex in those cases */
2231 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2232 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2233 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2234 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2239 kgnilnd_tx_done(tx, rc);
2240 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie,
2241 lnet_nid_to_nid4(&ni->ni_nid));
2243 lnet_finalize(lntmsg, rc);
2247 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2250 kgn_rx_t *rx = private;
2251 kgn_conn_t *conn = rx->grx_conn;
2252 kgn_msg_t *rxmsg = rx->grx_msg;
2253 kgn_msg_t *eagermsg = NULL;
2254 kgn_peer_t *peer = NULL;
2255 kgn_conn_t *found_conn = NULL;
2257 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2258 conn, rxmsg, lntmsg);
2260 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2261 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2262 rxmsg->gnm_payload_len);
2265 /* Grab a read lock so the connection doesnt disappear on us
2266 * while we look it up
2268 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2270 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2272 found_conn = kgnilnd_find_conn_locked(peer);
2275 /* Verify the connection found is the same one that the message
2276 * is supposed to be using, if it is not output an error message
2279 if (!peer || !found_conn
2280 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2281 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2282 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2283 peer, conn, found_conn);
2285 CERROR("Unexpected connstamp %#llx(%#llx expected) from %s\n",
2286 rxmsg->gnm_connstamp,
2287 found_conn->gnc_peer_connstamp,
2288 libcfs_nid2str(peer->gnp_nid));
2293 /* add conn ref to ensure it doesn't go away until all eager
2294 * messages processed */
2295 kgnilnd_conn_addref(conn);
2297 /* Now that we have verified the connection is valid and added a
2298 * reference we can remove the read_lock on the peer_conn_lock */
2299 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2301 /* we have no credits or buffers for this message, so copy it
2302 * somewhere for a later kgnilnd_recv */
2303 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2304 *kgnilnd_tunables.kgn_eager_credits) {
2305 CERROR("Out of eager credits to %s\n",
2306 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2310 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2312 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2313 if (eagermsg == NULL) {
2314 kgnilnd_conn_decref(conn);
2315 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2316 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2320 /* copy msg and payload */
2321 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2322 rx->grx_msg = eagermsg;
2325 /* stash this for lnet_finalize on cancel-on-conn-close */
2326 rx->grx_lntmsg = lntmsg;
2328 /* keep the same rx_t, it just has a new grx_msg now */
2329 *new_private = private;
2331 /* release SMSG buffer */
2332 kgnilnd_release_msg(conn);
2338 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2339 int delayed, unsigned int niov,
2340 struct bio_vec *kiov,
2341 unsigned int offset, unsigned int mlen, unsigned int rlen)
2343 kgn_rx_t *rx = private;
2344 kgn_conn_t *conn = rx->grx_conn;
2345 kgn_msg_t *rxmsg = rx->grx_msg;
2351 LASSERT(!in_interrupt());
2352 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2354 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2355 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2356 conn, rxmsg, lntmsg,
2357 niov, kiov, offset, mlen, rlen);
2359 /* we need to lock here as recv can be called from any context */
2360 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2361 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2362 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2364 /* someone closed the conn after we copied this out, nuke it */
2365 kgnilnd_consume_rx(rx);
2366 lnet_finalize(lntmsg, conn->gnc_error);
2369 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2371 switch (rxmsg->gnm_type) {
2373 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2374 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2375 conn, rx, rxmsg, lntmsg, niov, kiov, offset, mlen, rlen);
2378 case GNILND_MSG_IMMEDIATE:
2379 if (mlen > rxmsg->gnm_payload_len) {
2380 GNIDBG_MSG(D_ERROR, rxmsg,
2381 "Immediate message from %s too big: %d > %d",
2382 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2383 rxmsg->gnm_payload_len);
2385 kgnilnd_consume_rx(rx);
2389 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2390 * right after the kgn_msg_t header - so just 'cute' way of saying
2391 * rxmsg + sizeof(kgn_msg_t) */
2393 /* check payload checksum if sent */
2395 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2396 !rxmsg->gnm_payload_cksum &&
2397 rxmsg->gnm_payload_len != 0)
2398 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2400 if (rxmsg->gnm_payload_cksum != 0) {
2401 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2402 * which is what is used to calculate the cksum on the TX side */
2403 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2405 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2406 GNIDBG_MSG(D_NETERROR, rxmsg,
2407 "Bad payload checksum (%x expected %x)",
2408 pload_cksum, rxmsg->gnm_payload_cksum);
2409 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2411 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2412 &rxmsg[1], rxmsg->gnm_payload_len);
2415 libcfs_debug_dumplog();
2421 /* checksum problems are fatal, kill the conn */
2422 kgnilnd_consume_rx(rx);
2423 kgnilnd_close_conn(conn, rc);
2428 lnet_copy_flat2kiov(
2430 *kgnilnd_tunables.kgn_max_immediate,
2431 &rxmsg[1], 0, mlen);
2433 kgnilnd_consume_rx(rx);
2434 lnet_finalize(lntmsg, 0);
2437 case GNILND_MSG_PUT_REQ:
2438 /* LNET wants to truncate or drop transaction, sending NAK */
2440 kgnilnd_consume_rx(rx);
2441 lnet_finalize(lntmsg, 0);
2443 /* only error if lntmsg == NULL, otherwise we are just
2444 * short circuiting the rdma process of 0 bytes */
2445 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2446 lntmsg == NULL ? -ENOENT : 0,
2447 rxmsg->gnm_u.get.gngm_cookie,
2448 lnet_nid_to_nid4(&ni->ni_nid));
2451 /* sending ACK with sink buff. info */
2452 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK,
2453 lnet_nid_to_nid4(&ni->ni_nid));
2455 kgnilnd_consume_rx(rx);
2459 rc = kgnilnd_set_tx_id(tx, conn);
2461 GOTO(nak_put_req, rc);
2464 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2465 kiov, offset, mlen);
2467 GOTO(nak_put_req, rc);
2470 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2471 rxmsg->gnm_u.putreq.gnprm_cookie;
2472 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2473 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2474 (__u64)((unsigned long)tx->tx_buffer);
2475 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2477 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2478 tx->tx_qtime = jiffies;
2479 /* we only queue from kgnilnd_recv - we might get called from other contexts
2480 * and we don't want to block the mutex in those cases */
2482 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2483 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2484 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2485 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2487 kgnilnd_consume_rx(rx);
2491 /* make sure we send an error back when the PUT fails */
2492 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc,
2493 rxmsg->gnm_u.get.gngm_cookie,
2494 lnet_nid_to_nid4(&ni->ni_nid));
2495 kgnilnd_tx_done(tx, rc);
2496 kgnilnd_consume_rx(rx);
2498 /* return magic LNet network error */
2500 case GNILND_MSG_GET_REQ_REV:
2501 /* LNET wants to truncate or drop transaction, sending NAK */
2503 kgnilnd_consume_rx(rx);
2504 lnet_finalize(lntmsg, 0);
2506 /* only error if lntmsg == NULL, otherwise we are just
2507 * short circuiting the rdma process of 0 bytes */
2508 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2509 lntmsg == NULL ? -ENOENT : 0,
2510 rxmsg->gnm_u.get.gngm_cookie,
2511 lnet_nid_to_nid4(&ni->ni_nid));
2514 /* lntmsg can be null when parsing a LNET_GET */
2515 if (lntmsg != NULL) {
2516 /* sending ACK with sink buff. info */
2517 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV,
2518 lnet_nid_to_nid4(&ni->ni_nid));
2520 kgnilnd_consume_rx(rx);
2524 rc = kgnilnd_set_tx_id(tx, conn);
2526 GOTO(nak_get_req_rev, rc);
2528 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2529 kiov, offset, mlen);
2531 GOTO(nak_get_req_rev, rc);
2533 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2534 rxmsg->gnm_u.putreq.gnprm_cookie;
2535 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2536 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2537 (__u64)((unsigned long)tx->tx_buffer);
2538 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2540 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2542 /* we only queue from kgnilnd_recv - we might get called from other contexts
2543 * and we don't want to block the mutex in those cases */
2545 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2546 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2547 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2548 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2551 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2553 rxmsg->gnm_u.get.gngm_cookie,
2554 lnet_nid_to_nid4(&ni->ni_nid));
2557 kgnilnd_consume_rx(rx);
2561 /* make sure we send an error back when the GET fails */
2562 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc,
2563 rxmsg->gnm_u.get.gngm_cookie,
2564 lnet_nid_to_nid4(&ni->ni_nid));
2565 kgnilnd_tx_done(tx, rc);
2566 kgnilnd_consume_rx(rx);
2568 /* return magic LNet network error */
2572 case GNILND_MSG_PUT_REQ_REV:
2573 /* LNET wants to truncate or drop transaction, sending NAK */
2575 kgnilnd_consume_rx(rx);
2576 lnet_finalize(lntmsg, 0);
2578 /* only error if lntmsg == NULL, otherwise we are just
2579 * short circuiting the rdma process of 0 bytes */
2580 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2581 lntmsg == NULL ? -ENOENT : 0,
2582 rxmsg->gnm_u.get.gngm_cookie,
2583 lnet_nid_to_nid4(&ni->ni_nid));
2587 if (lntmsg != NULL) {
2589 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2592 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2594 rxmsg->gnm_u.get.gngm_cookie,
2595 lnet_nid_to_nid4(&ni->ni_nid));
2597 kgnilnd_consume_rx(rx);
2599 case GNILND_MSG_GET_REQ:
2600 if (lntmsg != NULL) {
2602 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2605 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2607 rxmsg->gnm_u.get.gngm_cookie,
2608 lnet_nid_to_nid4(&ni->ni_nid));
2610 kgnilnd_consume_rx(rx);
2616 /* needs write_lock on kgn_peer_conn_lock held */
2618 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2620 unsigned long timeout, keepalive;
2621 unsigned long now = jiffies;
2622 unsigned long newest_last_rx;
2625 /* given that we found this conn hanging off a peer, it better damned
2626 * well be connected */
2627 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2628 "conn 0x%p->%s with bad state%s\n", conn,
2629 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2631 kgnilnd_conn_state2str(conn));
2633 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2634 "rx_diff %lu tx_diff %lu\n",
2635 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2636 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2637 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2638 cfs_duration_sec(now - conn->gnc_last_tx));
2640 timeout = cfs_time_seconds(conn->gnc_timeout);
2641 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2643 /* just in case our lack of RX msg processing is gumming up the works - give the
2644 * remove an extra chance */
2646 newest_last_rx = GNILND_LASTRX(conn);
2648 if (time_after_eq(now, newest_last_rx + timeout)) {
2649 uint32_t level = D_CONSOLE|D_NETERROR;
2651 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2654 GNIDBG_CONN(level, conn,
2655 "No gnilnd traffic received from %s for %lu "
2656 "seconds, terminating connection. Is node down? ",
2657 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2658 cfs_duration_sec(now - newest_last_rx));
2662 /* we don't timeout on last_tx stalls - we are going to trust the
2663 * underlying network to let us know when sends are failing.
2664 * At worst, the peer will timeout our RX stamp and drop the connection
2665 * at that point. We'll then see his CLOSE or at worst his RX
2666 * stamp stop and drop the connection on our end */
2668 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2669 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2670 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2671 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2672 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2674 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2675 conn->gnc_last_noop_cq,
2676 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2677 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2678 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2679 set_mb(conn->gnc_last_noop_want, jiffies);
2680 atomic_inc(&conn->gnc_reaper_noop);
2681 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2684 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP,
2685 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid));
2688 kgnilnd_queue_tx(conn, tx);
2694 /* needs write_lock on kgn_peer_conn_lock held */
2696 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2697 struct list_head *souls)
2699 unsigned long timeout;
2700 kgn_conn_t *conn, *connN = NULL;
2706 short releaseconn = 0;
2707 unsigned long first_rx = 0;
2708 int purgatory_conn_cnt = 0;
2710 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2711 peer, libcfs_nid2str(peer->gnp_nid),
2712 peer->gnp_reconnect_interval);
2714 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2715 GNILND_MIN_TIMEOUT));
2717 conn = kgnilnd_find_conn_locked(peer);
2719 /* if there is a valid conn, check the queues for timeouts */
2720 rc = kgnilnd_check_conn_timeouts_locked(conn);
2722 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2723 /* simulate a RX CLOSE after the timeout but before
2724 * the scheduler thread gets it */
2725 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2726 conn->gnc_peer_error = -ETIMEDOUT;
2729 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2731 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2732 CDEBUG(D_WARNING, "%s conn timed out, will "
2733 "reconnect upon request from peer\n",
2734 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2736 /* Once we mark closed, any of the scheduler threads could
2737 * get it and move through before we hit the fail loc code */
2738 kgnilnd_close_conn_locked(conn, rc);
2740 /* first_rx is used to decide when to release a conn from purgatory.
2742 first_rx = conn->gnc_first_rx;
2746 /* now regardless of starting new conn, find tx on peer queue that
2747 * are old and smell bad - do this first so we don't trigger
2748 * reconnect on empty queue if we timeout all */
2749 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2750 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2752 LCONSOLE_INFO("could not send to %s due to connection"
2753 " setup failure after %lu seconds\n",
2754 libcfs_nid2str(peer->gnp_nid),
2755 cfs_duration_sec(jiffies - tx->tx_qtime));
2757 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2759 list_add_tail(&tx->tx_list, todie);
2764 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2765 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2766 count, peer, libcfs_nid2str(peer->gnp_nid));
2767 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2768 if (list_empty(&peer->gnp_tx_queue) ||
2769 peer->gnp_connecting == GNILND_PEER_KILL) {
2770 /* we pass down todie to use a common function - but we know there are
2772 kgnilnd_cancel_peer_connect_locked(peer, todie);
2776 /* Don't reconnect if we are still trying to clear out old conns.
2777 * This prevents us sending traffic on the new mbox before ensuring we are done
2778 * with the old one */
2779 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2780 (atomic_read(&peer->gnp_dirty_eps) == 0);
2782 /* fast reconnect after a timeout */
2783 to_reconn = !conn &&
2784 (peer->gnp_last_errno == -ETIMEDOUT) &&
2785 *kgnilnd_tunables.kgn_fast_reconn;
2787 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2788 * to be sent, we'll check the reconnect interval and fire up a new
2789 * connection request */
2792 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2793 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2794 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2796 CDEBUG(D_NET, "starting connect to %s\n",
2797 libcfs_nid2str(peer->gnp_nid));
2798 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2799 "Peer was idle and we have a write_lock, state issue %d\n",
2800 peer->gnp_connecting);
2802 peer->gnp_connecting = GNILND_PEER_CONNECT;
2803 kgnilnd_peer_addref(peer); /* extra ref for connd */
2805 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2806 list_add_tail(&peer->gnp_connd_list,
2807 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2808 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2810 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2813 /* fail_loc to allow us to delay release of purgatory */
2814 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2817 /* This check allows us to verify that the new conn is actually being used. This allows us to
2818 * pull the old conns out of purgatory if they have actually seen traffic.
2819 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2822 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2823 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2824 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2828 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2829 /* check for purgatory timeouts */
2830 if (conn->gnc_in_purgatory) {
2831 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2832 * that way the next time we check it we can detach it from purgatory
2835 if (conn->gnc_state != GNILND_CONN_DONE) {
2836 /* Skip over conns that are currently not DONE. If they arent already scheduled
2837 * for completion something in the state machine is broken.
2842 /* We only detach a conn that is in purgatory if we have received a close message,
2843 * we have a new valid connection that has successfully received data, or an admin
2844 * command tells us we need to detach.
2847 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2848 unsigned long waiting;
2850 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2852 /* C.E: The remote peer is expected to close the
2853 * connection (see kgnilnd_check_conn_timeouts)
2854 * via the reaper thread and nuke out the MDD and
2855 * FMA resources after conn->gnc_timeout has expired
2856 * without an FMA RX */
2857 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2858 " held resources\n",
2859 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2860 cfs_duration_sec(waiting));
2862 kgnilnd_detach_purgatory_locked(conn, souls);
2864 purgatory_conn_cnt++;
2869 /* If we have too many connections in purgatory we could run out of
2870 * resources. Limit the number of connections to a tunable number,
2871 * clean up to the minimum all in one fell swoop... there are
2872 * situations where dvs will retry tx's and we can eat up several
2873 * hundread connection requests at once.
2875 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2876 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2878 if (conn->gnc_in_purgatory &&
2879 conn->gnc_state == GNILND_CONN_DONE) {
2880 CDEBUG(D_NET, "Dropping Held resource due to"
2881 " resource limits being hit\n");
2882 kgnilnd_detach_purgatory_locked(conn, souls);
2884 if (purgatory_conn_cnt-- <
2885 *kgnilnd_tunables.kgn_max_purgatory)
2893 kgnilnd_reaper_check(int idx)
2895 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2896 struct list_head *ctmp, *ctmpN;
2897 LIST_HEAD(geriatrics);
2900 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2902 list_for_each_safe(ctmp, ctmpN, peers) {
2903 kgn_peer_t *peer = NULL;
2905 /* don't timeout stuff if the network is mucked or shutting down */
2906 if (kgnilnd_check_hw_quiesce()) {
2909 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2911 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
2914 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2916 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
2917 kgnilnd_release_purgatory_list(&souls);
2921 kgnilnd_update_reaper_timeout(long timeout)
2923 LASSERT(timeout > 0);
2925 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2927 if (timeout < kgnilnd_data.kgn_new_min_timeout)
2928 kgnilnd_data.kgn_new_min_timeout = timeout;
2930 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2934 kgnilnd_reaper_poke_with_stick(cfs_timer_cb_arg_t arg)
2936 wake_up(&kgnilnd_data.kgn_reaper_waitq);
2940 kgnilnd_reaper(void *arg)
2945 unsigned long next_check_time = jiffies;
2946 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
2947 struct timer_list timer;
2950 /* all gnilnd threads need to run fairly urgently */
2951 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2952 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2954 while (!kgnilnd_data.kgn_shutdown) {
2955 /* I wake up every 'p' seconds to check for timeouts on some
2956 * more peers. I try to check every connection 'n' times
2957 * within the global minimum of all keepalive and timeout
2958 * intervals, to ensure I attend to every connection within
2959 * (n+1)/n times its timeout intervals. */
2960 const int p = GNILND_REAPER_THREAD_WAKE;
2961 const int n = GNILND_REAPER_NCHECKS;
2963 /* to quiesce or to not quiesce, that is the question */
2964 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2965 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2966 KGNILND_SPIN_QUIESCE;
2967 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2970 /* careful with the jiffy wrap... */
2971 timeout = (long)(next_check_time - jiffies);
2974 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
2975 TASK_INTERRUPTIBLE);
2976 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2977 cfs_timer_setup(&timer, kgnilnd_reaper_poke_with_stick,
2978 next_check_time, 0);
2979 mod_timer(&timer, (long) jiffies + timeout);
2981 /* check flag variables before committing */
2982 if (!kgnilnd_data.kgn_shutdown &&
2983 !kgnilnd_data.kgn_quiesce_trigger) {
2984 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2985 timeout, cfs_duration_sec(timeout));
2987 CDEBUG(D_INFO, "awake after schedule\n");
2990 del_singleshot_timer_sync(&timer);
2991 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2992 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
2996 /* new_min_timeout is set from the conn timeouts and keepalive
2997 * this should end up with a min timeout of
2998 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
2999 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3000 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3001 CDEBUG(D_NET, "Set new min timeout %ld\n",
3002 current_min_timeout);
3005 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3007 /* Compute how many table entries to check now so I get round
3008 * the whole table fast enough given that I do this at fixed
3009 * intervals of 'p' seconds) */
3010 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3011 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3012 chunk = (chunk * n * p) /
3013 kgnilnd_data.kgn_new_min_timeout;
3016 for (i = 0; i < chunk; i++) {
3017 kgnilnd_reaper_check(hash_index);
3018 hash_index = (hash_index + 1) %
3019 *kgnilnd_tunables.kgn_peer_hash_size;
3021 next_check_time = (long) jiffies + cfs_time_seconds(p);
3022 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3024 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3027 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3029 kgnilnd_thread_fini();
3034 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3035 unsigned niov, offset, nob;
3036 struct bio_vec *kiov;
3037 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3038 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3041 lnet_copy_flat2kiov(
3044 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3046 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3053 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3056 gni_post_descriptor_t *desc;
3058 kgn_tx_ev_id_t ev_id;
3060 int should_retry, rc;
3061 long num_processed = 0;
3062 kgn_conn_t *conn = NULL;
3063 kgn_tx_t *tx = NULL;
3064 kgn_rdma_desc_t *rdesc;
3069 /* make sure we don't keep looping if we need to reset */
3070 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3071 return num_processed;
3073 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3075 /* we didn't get the mutex, so return that there is still work
3079 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3080 /* a bit gross - but we need a good way to test for
3081 * delayed RDMA completions and the easiest way to do
3082 * that is to delay the RDMA CQ events */
3083 rrc = GNI_RC_NOT_DONE;
3085 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3088 if (rrc == GNI_RC_NOT_DONE) {
3089 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3090 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3091 dev->gnd_id, num_processed);
3092 return num_processed;
3094 dev->gnd_sched_alive = jiffies;
3097 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3098 "this is bad, somehow our credits didn't protect us"
3099 " from CQ overrun\n");
3100 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3101 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3102 event_data, GNI_CQ_GET_TYPE(event_data));
3104 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3106 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3108 /* XXX Nic: Need better error handling here... */
3109 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3110 (rrc == GNI_RC_TRANSACTION_ERROR),
3113 ev_id.txe_cookie = desc->post_id;
3115 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3117 if (conn == NULL || tx == NULL) {
3118 /* either conn or tx was already nuked and this is a "late"
3119 * completion, so drop it */
3123 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3124 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3125 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3126 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3127 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3129 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3131 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3132 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3133 tx->tx_msg.gnm_u.completion.gncm_retval);
3137 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3138 if (tx->tx_buffer_copy != NULL)
3139 kgnilnd_recv_bte_get(tx);
3140 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3143 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3144 if (tx->tx_buffer_copy != NULL)
3145 kgnilnd_recv_bte_get(tx);
3146 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3149 /* remove from rdmaq */
3150 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3151 spin_lock(&conn->gnc_list_lock);
3152 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3153 spin_unlock(&conn->gnc_list_lock);
3154 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3156 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3157 event_data = 1LL << 48;
3161 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3162 atomic_inc(&dev->gnd_rdma_ntx);
3163 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3164 /* transaction succeeded, add into fmaq */
3165 kgnilnd_queue_tx(conn, tx);
3166 kgnilnd_peer_alive(conn->gnc_peer);
3168 /* drop ref from kgnilnd_validate_tx_ev_id */
3169 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3170 kgnilnd_conn_decref(conn);
3175 /* fall through to the TRANSACTION_ERROR case */
3178 /* get stringified version for log messages */
3179 kgnilnd_cq_error_str(event_data, &err_str, 256);
3180 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3182 /* make sure we are not off in the weeds with this tx */
3183 if (tx->tx_retrans >
3184 *kgnilnd_tunables.kgn_max_retransmits) {
3185 GNIDBG_TX(D_NETERROR, tx,
3186 "giving up on TX, too many retries", NULL);
3190 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3191 should_retry ? "transient" : "unrecoverable", err_str);
3193 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3194 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3195 rdesc = &tx->tx_putinfo.gnpam_desc;
3196 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3197 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3199 rdesc = &tx->tx_getinfo.gngm_desc;
3200 rnob = tx->tx_lntmsg[0]->msg_len;
3201 rcookie = tx->tx_getinfo.gngm_cookie;
3206 tx->tx_msg.gnm_type,
3210 kgnilnd_nak_rdma(conn,
3211 tx->tx_msg.gnm_type,
3214 tx->tx_msg.gnm_srcnid);
3215 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3216 kgnilnd_close_conn(conn, -ECOMM);
3219 /* drop ref from kgnilnd_validate_tx_ev_id */
3220 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3221 kgnilnd_conn_decref(conn);
3226 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3230 kgn_tx_ev_id_t ev_id;
3231 kgn_tx_t *tx = NULL;
3232 kgn_conn_t *conn = NULL;
3233 int queued_fma, saw_reply, rc;
3234 long num_processed = 0;
3235 struct list_head *ctmp, *ctmpN;
3238 /* make sure we don't keep looping if we need to reset */
3239 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3240 return num_processed;
3243 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3245 /* we didn't get the mutex, so return that there is still work
3250 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3251 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3253 if (rrc == GNI_RC_NOT_DONE) {
3255 "SMSG send CQ %d not ready (data %#llx) "
3256 "processed %ld\n", dev->gnd_id, event_data,
3259 if (num_processed > 0) {
3260 spin_lock(&dev->gnd_lock);
3261 if (!list_empty(&dev->gnd_delay_conns)) {
3262 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3263 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3264 list_del_init(&conn->gnc_delaylist);
3265 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3266 kgnilnd_schedule_conn_nolock(conn);
3268 spin_unlock(&dev->gnd_lock);
3269 kgnilnd_schedule_device(dev);
3271 spin_unlock(&dev->gnd_lock);
3274 return num_processed;
3277 dev->gnd_sched_alive = jiffies;
3280 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3281 "this is bad, somehow our credits didn't "
3282 "protect us from CQ overrun\n");
3283 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3284 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3285 event_data, GNI_CQ_GET_TYPE(event_data));
3287 /* if SMSG couldn't handle an error, time for conn to die */
3288 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3291 /* need to take the write_lock to ensure atomicity
3292 * on the conn state if we need to close it */
3293 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3294 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3296 /* Conn was destroyed? */
3298 "SMSG CQID lookup %#llx failed\n",
3299 GNI_CQ_GET_INST_ID(event_data));
3300 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3304 kgnilnd_cq_error_str(event_data, &err_str, 256);
3305 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3306 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3308 kgnilnd_close_conn_locked(conn, -ECOMM);
3310 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3312 /* no need to process rest of this tx -
3313 * it is getting canceled */
3317 /* fall through to GNI_RC_SUCCESS case */
3318 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3320 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3321 if (conn == NULL || tx == NULL) {
3322 /* either conn or tx was already nuked and this is a "late"
3323 * completion, so drop it */
3327 tx->tx_conn->gnc_last_tx_cq = jiffies;
3328 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3329 set_mb(conn->gnc_last_noop_cq, jiffies);
3332 /* lock tx_list_state and tx_state */
3333 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3334 spin_lock(&tx->tx_conn->gnc_list_lock);
3336 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3337 "state not GNILND_TX_LIVE_FMAQ", NULL);
3338 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3339 "not waiting for completion", NULL);
3341 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3344 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3346 /* This will trigger other FMA sends that were
3347 * pending this completion */
3348 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3350 /* we either did not expect reply or we already got it */
3351 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3353 spin_unlock(&tx->tx_conn->gnc_list_lock);
3354 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3357 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3359 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3360 kgnilnd_schedule_conn(conn);
3363 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3364 * If saw_reply is true we know that the tx is safe to use as the other thread
3365 * is already finished with it.
3369 /* no longer need to track on the live_fmaq */
3370 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3372 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3373 /* we already got reply & were waiting for
3374 * completion of initial send */
3375 /* to initiate RDMA transaction */
3376 GNIDBG_TX(D_NET, tx,
3377 "Pending RDMA 0x%p type 0x%02x",
3378 tx->tx_msg.gnm_type);
3379 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3380 rc = kgnilnd_send_mapped_tx(tx, 0);
3381 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3383 /* we are done with this tx */
3384 GNIDBG_TX(D_NET, tx,
3385 "Done with tx type 0x%02x",
3386 tx->tx_msg.gnm_type);
3387 kgnilnd_tx_done(tx, tx->tx_rc);
3391 /* drop ref from kgnilnd_validate_tx_ev_id */
3392 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3393 kgnilnd_conn_decref(conn);
3395 /* if we are waiting for a REPLY, we'll handle the tx then */
3396 } /* end for loop */
3400 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3405 long num_processed = 0;
3406 struct list_head *conns;
3407 struct list_head *tmp;
3411 /* make sure we don't keep looping if we need to reset */
3412 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3413 return num_processed;
3416 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3418 /* we didn't get the mutex, so return that there is still work
3422 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3423 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3425 if (rrc == GNI_RC_NOT_DONE) {
3426 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3428 dev->gnd_id, event_data, num_processed);
3429 return num_processed;
3431 dev->gnd_sched_alive = jiffies;
3434 /* this is the only CQ that can really handle transient
3436 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3437 rrc = cfs_fail_val ? cfs_fail_val
3438 : GNI_RC_ERROR_RESOURCE;
3439 if (rrc == GNI_RC_ERROR_RESOURCE) {
3440 /* set overrun too */
3441 event_data |= (1UL << 63);
3442 LASSERTF(GNI_CQ_OVERRUN(event_data),
3443 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3446 /* sender should get error event too and take care
3447 of failed transaction by re-transmitting */
3448 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3449 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3453 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3454 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3455 conn = kgnilnd_cqid2conn_locked(
3456 GNI_CQ_GET_INST_ID(event_data));
3458 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3459 "failed, dropping event %#llx\n",
3460 GNI_CQ_GET_INST_ID(event_data),
3463 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3465 GNI_CQ_GET_INST_ID(event_data),
3466 conn, conn->gnc_peer ?
3467 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3470 conn->gnc_last_rx_cq = jiffies;
3472 /* stash first rx so we can clear out purgatory.
3474 if (conn->gnc_first_rx == 0) {
3475 conn->gnc_first_rx = jiffies;
3477 kgnilnd_peer_alive(conn->gnc_peer);
3478 kgnilnd_schedule_conn(conn);
3480 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3484 /* FMA CQ has overflowed: check ALL conns */
3485 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3486 "conns on device %d\n", dev->gnd_id);
3488 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3490 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3491 conns = &kgnilnd_data.kgn_conns[rc];
3493 list_for_each(tmp, conns) {
3494 conn = list_entry(tmp, kgn_conn_t,
3497 if (conn->gnc_device == dev) {
3498 kgnilnd_schedule_conn(conn);
3499 conn->gnc_last_rx_cq = jiffies;
3503 /* don't block write lockers for too long... */
3504 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3509 /* try_map_if_full should only be used when processing TX from list of
3510 * backlog TX waiting on mappings to free up
3513 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3514 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3517 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3519 /* slight bit of race if multiple people calling, but at worst we'll have
3520 * order altered just a bit... which would not be determenistic anyways */
3521 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3523 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3525 /* We know that we have a GART reservation that should guarantee forward progress.
3526 * This means we don't need to take any extraordinary efforts if we are failing
3527 * mappings here - even if we are holding a very small number of these. */
3529 if (try_map_if_full || (rc == 0)) {
3530 rc = kgnilnd_map_buffer(tx);
3533 /* rc should be 0 if we mapped successfully here, if non-zero
3534 * we are queueing */
3536 /* if try_map_if_full set, they handle requeuing */
3537 if (unlikely(try_map_if_full)) {
3540 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3541 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3542 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3543 /* make sure we wake up sched to run this */
3544 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3545 /* return 0 as this is now queued for later sending */
3550 switch (tx->tx_msg.gnm_type) {
3554 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3555 * remote node where the RDMA will be started
3556 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3557 * be satisified. The rest of the errors are "hard" errors that require
3558 * upper layers to handle themselves.
3559 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3560 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3561 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3562 * GNILND_MSG_GET_DONE_REV.
3564 case GNILND_MSG_GET_REQ:
3565 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3566 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3567 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3568 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3569 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3570 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3571 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3573 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3574 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3576 case GNILND_MSG_PUT_ACK:
3577 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3578 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3579 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3580 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3582 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3583 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3586 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3587 case GNILND_MSG_PUT_DONE:
3588 case GNILND_MSG_PUT_REQ:
3589 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3590 &tx->tx_putinfo.gnpam_desc,
3591 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3592 tx->tx_putinfo.gnpam_dst_cookie);
3593 RETURN(try_map_if_full ? rc : 0);
3595 case GNILND_MSG_GET_DONE:
3596 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3597 &tx->tx_getinfo.gngm_desc,
3598 tx->tx_lntmsg[0]->msg_len,
3599 tx->tx_getinfo.gngm_cookie);
3600 RETURN(try_map_if_full ? rc : 0);
3602 case GNILND_MSG_PUT_REQ_REV:
3603 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3604 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3605 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3606 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3607 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3608 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3609 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3611 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3613 case GNILND_MSG_PUT_DONE_REV:
3614 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3615 &tx->tx_getinfo.gngm_desc,
3617 tx->tx_getinfo.gngm_cookie);
3618 RETURN(try_map_if_full ? rc : 0);
3620 case GNILND_MSG_GET_ACK_REV:
3621 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3622 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3623 /* LNET_GETS are a special case for parse */
3624 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3625 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3627 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3628 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3630 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3631 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3633 case GNILND_MSG_GET_DONE_REV:
3634 case GNILND_MSG_GET_REQ_REV:
3635 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3636 &tx->tx_putinfo.gnpam_desc,
3637 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3638 tx->tx_putinfo.gnpam_dst_cookie);
3639 RETURN(try_map_if_full ? rc : 0);
3647 kgnilnd_process_fmaq(kgn_conn_t *conn)
3650 kgn_tx_t *tx = NULL;
3651 void *buffer = NULL;
3652 unsigned int nob = 0;
3655 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3656 * However I will be rescheduled by an FMA completion event
3657 * when I eventually get some.
3658 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3659 * But it doesn't matter if I try to send a "real" message just
3660 * as I start closing because I'll get scheduled to send the
3663 /* Short circuit if the ep_handle is null we cant send anyway. */
3664 if (conn->gnc_ephandle == NULL)
3667 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3669 spin_lock(&conn->gnc_list_lock);
3671 if (list_empty(&conn->gnc_fmaq)) {
3672 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3674 spin_unlock(&conn->gnc_list_lock);
3676 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3677 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3678 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3679 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3680 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3682 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3683 conn->gnc_last_noop_cq,
3684 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3685 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3686 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3687 atomic_inc(&conn->gnc_sched_noop);
3688 set_mb(conn->gnc_last_noop_want, jiffies);
3690 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3693 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP,
3694 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid));
3698 rc = kgnilnd_set_tx_id(tx, conn);
3700 kgnilnd_tx_done(tx, rc);
3706 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3707 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3708 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3709 more_to_do = !list_empty(&conn->gnc_fmaq);
3710 spin_unlock(&conn->gnc_list_lock);
3713 /* if there is no real TX or no NOOP to send, bail */
3718 if (!tx->tx_retrans)
3719 tx->tx_cred_wait = jiffies;
3721 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3722 "tx with zero id", NULL);
3724 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3725 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3726 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3730 switch (tx->tx_msg.gnm_type) {
3734 case GNILND_MSG_NOOP:
3735 case GNILND_MSG_CLOSE:
3736 case GNILND_MSG_IMMEDIATE:
3737 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3738 buffer = tx->tx_buffer;
3742 case GNILND_MSG_GET_DONE:
3743 case GNILND_MSG_PUT_DONE:
3744 case GNILND_MSG_PUT_DONE_REV:
3745 case GNILND_MSG_GET_DONE_REV:
3746 case GNILND_MSG_PUT_NAK:
3747 case GNILND_MSG_GET_NAK:
3748 case GNILND_MSG_GET_NAK_REV:
3749 case GNILND_MSG_PUT_NAK_REV:
3750 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3753 case GNILND_MSG_PUT_REQ:
3754 case GNILND_MSG_GET_REQ_REV:
3755 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3757 case GNILND_MSG_PUT_ACK:
3758 case GNILND_MSG_PUT_REQ_REV:
3759 case GNILND_MSG_GET_ACK_REV:
3760 case GNILND_MSG_GET_REQ:
3761 /* This is really only to handle the retransmit of SMSG once these
3762 * two messages are setup in send_mapped_tx */
3763 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3767 if (likely(rc == 0)) {
3768 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3772 /* don't explicitly reschedule here - we are short credits and will rely on
3773 * kgnilnd_sendmsg to resched the conn if need be */
3775 } else if (rc < 0) {
3776 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3777 * almost certainly a software bug, but lets play nice with the other kids */
3778 kgnilnd_tx_done(tx, rc);
3779 /* just for fun, kick peer in arse - resetting conn might help to correct
3780 * this almost certainly buggy software caused return code */
3781 kgnilnd_close_conn(conn, rc);
3785 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3786 kgnilnd_schedule_conn(conn);
3791 kgnilnd_process_rdmaq(kgn_device_t *dev)
3796 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3800 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3801 unsigned long dead_bump;
3804 /* if we think we need to adjust, take lock to serialize and recheck */
3805 spin_lock(&dev->gnd_rdmaq_lock);
3806 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3807 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3809 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3811 /* roll the bucket forward */
3812 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3814 if (kgnilnd_data.kgn_rdmaq_override &&
3815 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3816 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3821 /* roll current outstanding forward to make sure we carry outstanding
3822 * committment forward
3823 * new_ok starts out as the whole interval value
3824 * - first subtract bytes_out from last interval, as that would push us over
3825 * strict limits for this interval
3826 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3828 * there is a small race here if someone is actively processing mappings and
3829 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3830 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3831 * - if this gives us problems in the future, we could use a read/write lock
3832 * to protect the resetting of these values */
3833 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3834 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3836 CDEBUG(D_NET, "resetting rdmaq bytes to %lld, deadline +%lu -> %lu, current out %lld\n",
3837 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3838 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
3840 spin_unlock(&dev->gnd_rdmaq_lock);
3843 spin_lock(&dev->gnd_rdmaq_lock);
3844 while (!list_empty(&dev->gnd_rdmaq)) {
3847 /* make sure we break out early on quiesce */
3848 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3849 /* always break with lock held - we unlock outside loop */
3853 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3854 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3857 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3858 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3859 /* if conn is dying, mark tx in tx_ref_table for
3860 * kgnilnd_complete_closed_conn to finish up */
3861 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3863 /* tx was moved to DYING, get next */
3866 spin_unlock(&dev->gnd_rdmaq_lock);
3868 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3869 spin_lock(&dev->gnd_rdmaq_lock);
3872 /* no ticket! add back to head */
3873 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3874 /* clear found_work so scheduler threads wait for timer */
3878 /* TX is GO for launch */
3879 tx->tx_qtime = jiffies;
3880 kgnilnd_send_mapped_tx(tx, 0);
3884 spin_unlock(&dev->gnd_rdmaq_lock);
3890 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3892 __swab64s(&d->gnrd_key.qword1);
3893 __swab64s(&d->gnrd_key.qword2);
3894 __swab64s(&d->gnrd_addr);
3895 __swab32s(&d->gnrd_nob);
3898 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3899 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3902 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3904 kgn_tx_ev_id_t ev_id;
3907 /* we use the cookie from the original TX, so we can find the match
3908 * by parsing that and using the txe_idx */
3909 ev_id.txe_cookie = cookie;
3911 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3914 /* check tx to make sure kgni didn't eat it */
3915 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3916 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3918 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
3919 (tx->tx_id.txe_cookie = cookie)),
3920 "conn 0x%p->%s tx_ref_table hosed: wanted "
3921 "txe_cookie %#llx txe_idx %d "
3922 "found tx %p cookie %#llx txe_idx %d\n",
3923 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
3924 cookie, ev_id.txe_idx,
3925 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
3927 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
3928 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
3929 "Unexpected TX type (%x, %x or %x) "
3930 "or state (%x, expected +%x) "
3931 "matched reply from %s\n",
3932 tx->tx_msg.gnm_type, type1, type2,
3933 tx->tx_state, GNILND_TX_WAITING_REPLY,
3934 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3936 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
3937 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
3943 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
3946 kgn_conn_t *conn = tx->tx_conn;
3947 __u64 nob = tx->tx_nob;
3948 __u32 physnop = tx->tx_phys_npages;
3949 int id = tx->tx_id.txe_smsg_id;
3950 int buftype = tx->tx_buftype;
3951 gni_mem_handle_t hndl;
3952 hndl.qword1 = tx->tx_map_key.qword1;
3953 hndl.qword2 = tx->tx_map_key.qword2;
3955 spin_lock(&conn->gnc_list_lock);
3957 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
3958 "not waiting for reply", NULL);
3961 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
3963 if (rc == -EFAULT) {
3964 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
3965 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
3967 if(*kgnilnd_tunables.kgn_efault_lbug) {
3968 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
3969 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
3971 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
3972 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
3973 kgnilnd_tx_state2str(tx->tx_list_state),
3974 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
3979 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
3980 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3981 /* sample under lock as follow on steps require gnc_list_lock
3982 * - or call kgnilnd_tx_done which requires no locks held over
3983 * call to lnet_finalize */
3986 spin_unlock(&conn->gnc_list_lock);
3989 kgnilnd_tx_done(tx, tx->tx_rc);
3994 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
3997 kgn_conn_t *conn = tx->tx_conn;
3999 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4000 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4002 /* the gncm_retval is passed in for PUTs */
4003 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4004 msg->gnm_u.completion.gncm_retval);
4006 kgnilnd_complete_tx(tx, rc);
4010 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4018 kgn_peer_t *peer = conn->gnc_peer;
4021 __u16 tmp_cksum = 0, msg_cksum = 0;
4022 int repost = 1, saw_complete;
4023 unsigned long timestamp, newest_last_rx, timeout;
4025 struct lnet_hdr hdr;
4026 struct lnet_nid srcnid;
4029 /* Short circuit if the ep_handle is null.
4030 * It's likely that its about to be closed as stale.
4032 if (conn->gnc_ephandle == NULL)
4035 timestamp = jiffies;
4036 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4037 /* delay in jiffies - we are really concerned only with things that
4038 * result in a schedule() or really holding this off for long times .
4039 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4040 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4042 /* Resample current time as we have no idea how long it took to get the mutex */
4043 timestamp = jiffies;
4045 /* We check here when the last time we received an rx, we do this before
4046 * we call getnext in case the thread has been blocked for a while. If we
4047 * havent received an rx since our timeout value we close the connection
4048 * as we should assume the other side has closed the connection. This will
4049 * stop us from sending replies to a mailbox that is already in purgatory.
4052 timeout = cfs_time_seconds(conn->gnc_timeout);
4053 newest_last_rx = GNILND_LASTRX(conn);
4055 /* Error injection to validate that timestamp checking works and closing the conn */
4056 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4057 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4060 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4061 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4062 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4063 cfs_duration_sec(timestamp - newest_last_rx),
4064 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4065 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4067 kgnilnd_close_conn(conn, rc);
4071 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4073 if (rrc == GNI_RC_NOT_DONE) {
4074 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4075 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4079 /* Instead of asserting when we get mailbox corruption lets attempt to
4080 * close the conn and recover. We can put the conn/mailbox into
4081 * purgatory and let purgatory deal with the problem. If we see
4082 * this NETTERROR reported on production systems in large amounts
4083 * we will need to revisit the state machine to see if we can tighten
4084 * it up further to improve data protection.
4087 if (rrc == GNI_RC_INVALID_STATE) {
4088 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4089 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4090 "detected closing conn %p from peer %s\n", conn,
4091 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4093 kgnilnd_close_conn(conn, rc);
4097 LASSERTF(rrc == GNI_RC_SUCCESS,
4098 "bad rc %d on conn %p from peer %s\n",
4099 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4101 msg = (kgn_msg_t *)prefix;
4103 rx = kgnilnd_alloc_rx();
4105 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4106 kgnilnd_release_msg(conn);
4107 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4108 conn, libcfs_nid2str(peer->gnp_nid));
4112 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4114 timestamp = conn->gnc_last_rx;
4115 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4116 atomic_inc(&conn->gnc_rx_seq);
4118 conn->gnc_last_rx = jiffies;
4119 /* stash first rx so we can clear out purgatory
4121 if (conn->gnc_first_rx == 0)
4122 conn->gnc_first_rx = jiffies;
4124 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4125 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4126 kgnilnd_peer_alive(conn->gnc_peer);
4129 rx->grx_conn = conn;
4131 ktime_get_ts64(&rx->grx_received);
4133 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4136 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4142 kgnilnd_net_decref(net);
4145 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4146 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4148 /* XXX Nic: Do we need to swab cksum */
4149 if (msg->gnm_cksum != 0) {
4150 msg_cksum = msg->gnm_cksum;
4152 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4154 if (tmp_cksum != msg_cksum) {
4155 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4156 tmp_cksum, msg_cksum);
4157 kgnilnd_dump_msg(D_BUFFS, msg);
4162 /* restore checksum for future debug messages */
4163 msg->gnm_cksum = tmp_cksum;
4165 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4166 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4167 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4168 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4173 __swab32s(&msg->gnm_magic);
4174 __swab16s(&msg->gnm_version);
4175 __swab16s(&msg->gnm_type);
4176 __swab64s(&msg->gnm_srcnid);
4177 __swab64s(&msg->gnm_connstamp);
4178 __swab32s(&msg->gnm_seq);
4180 /* NB message type checked below; NOT here... */
4181 switch (msg->gnm_type) {
4182 case GNILND_MSG_GET_ACK_REV:
4183 case GNILND_MSG_PUT_ACK:
4184 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4187 case GNILND_MSG_PUT_REQ_REV:
4188 case GNILND_MSG_GET_REQ:
4189 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4197 if (msg->gnm_version != GNILND_MSG_VERSION) {
4198 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4199 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4204 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4205 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4206 libcfs_nid2str(msg->gnm_srcnid),
4207 libcfs_nid2str(peer->gnp_nid));
4212 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4213 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4214 " expected) from %s",
4215 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4216 libcfs_nid2str(peer->gnp_nid));
4221 if (msg->gnm_seq != seq) {
4222 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4223 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4228 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4230 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4231 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4232 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4233 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4234 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4235 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4236 /* double check state with lock held */
4237 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4238 /* only error if we are not already closing */
4239 if (conn->gnc_peer_error == -ETIMEDOUT) {
4240 unsigned long now = jiffies;
4241 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4243 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4244 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4245 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4247 cfs_duration_sec(now - timestamp),
4248 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4249 atomic_read(&conn->gnc_tx_seq),
4250 cfs_duration_sec(now - conn->gnc_last_tx),
4251 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4252 cfs_duration_sec(now - conn->gnc_last_noop_want),
4253 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4254 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4255 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4256 cfs_duration_sec(now - conn->gnc_last_sched_do),
4257 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4259 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4261 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4265 if (conn->gnc_close_recvd) {
4266 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4267 kgnilnd_msgtype2str(msg->gnm_type),
4268 msg->gnm_type, conn->gnc_close_recvd,
4269 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4274 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4275 /* XXX Nic: log message received on bad connection state */
4279 switch (msg->gnm_type) {
4280 case GNILND_MSG_NOOP:
4281 /* Nothing to do; just a keepalive */
4284 case GNILND_MSG_IMMEDIATE:
4285 /* only get SMSG payload for IMMEDIATE */
4286 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4287 lnet_hdr_from_nid4(&hdr, &msg->gnm_u.immediate.gnim_hdr);
4288 lnet_nid4_to_nid(msg->gnm_srcnid, &srcnid);
4289 rc = lnet_parse(net->gnn_ni, &hdr, &srcnid, rx, 0);
4292 case GNILND_MSG_GET_REQ_REV:
4293 case GNILND_MSG_PUT_REQ:
4294 lnet_hdr_from_nid4(&hdr, &msg->gnm_u.putreq.gnprm_hdr);
4295 lnet_nid4_to_nid(msg->gnm_srcnid, &srcnid);
4296 rc = lnet_parse(net->gnn_ni, &hdr, &srcnid, rx, 1);
4299 case GNILND_MSG_GET_NAK_REV:
4300 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4301 msg->gnm_u.completion.gncm_cookie);
4305 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4307 case GNILND_MSG_PUT_NAK:
4308 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4309 msg->gnm_u.completion.gncm_cookie);
4313 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4315 case GNILND_MSG_PUT_ACK:
4316 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4317 msg->gnm_u.putack.gnpam_src_cookie);
4321 /* store putack data for later: deferred rdma or re-try */
4322 tx->tx_putinfo = msg->gnm_u.putack;
4325 spin_lock(&tx->tx_conn->gnc_list_lock);
4327 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4328 "not waiting for reply", NULL);
4330 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4332 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4333 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4334 /* sample under lock as follow on steps require gnc_list_lock
4335 * - or call kgnilnd_tx_done which requires no locks held over
4336 * call to lnet_finalize */
4339 /* cannot launch rdma if still waiting for fma-msg completion */
4340 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4341 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4342 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4344 spin_unlock(&tx->tx_conn->gnc_list_lock);
4347 rc = kgnilnd_send_mapped_tx(tx, 0);
4349 kgnilnd_tx_done(tx, rc);
4352 case GNILND_MSG_GET_ACK_REV:
4353 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4354 msg->gnm_u.putack.gnpam_src_cookie);
4358 /* store putack data for later: deferred rdma or re-try */
4359 tx->tx_putinfo = msg->gnm_u.putack;
4361 spin_lock(&tx->tx_conn->gnc_list_lock);
4363 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4364 "not waiting for reply", NULL);
4366 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4368 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4369 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4370 /* sample under lock as follow on steps require gnc_list_lock
4371 * - or call kgnilnd_tx_done which requires no locks held over
4372 * call to lnet_finalize */
4375 /* cannot launch rdma if still waiting for fma-msg completion */
4376 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4377 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4378 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4380 spin_unlock(&tx->tx_conn->gnc_list_lock);
4383 rc = kgnilnd_send_mapped_tx(tx, 0);
4385 kgnilnd_tx_done(tx, rc);
4388 case GNILND_MSG_PUT_DONE:
4389 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4390 msg->gnm_u.completion.gncm_cookie);
4394 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4395 "bad tx buftype %d", tx->tx_buftype);
4397 kgnilnd_finalize_rx_done(tx, msg);
4399 case GNILND_MSG_PUT_REQ_REV:
4400 case GNILND_MSG_GET_REQ:
4401 lnet_hdr_from_nid4(&hdr, &msg->gnm_u.get.gngm_hdr);
4402 lnet_nid4_to_nid(msg->gnm_srcnid, &srcnid);
4403 rc = lnet_parse(net->gnn_ni, &hdr, &srcnid, rx, 1);
4407 case GNILND_MSG_GET_NAK:
4408 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4409 msg->gnm_u.completion.gncm_cookie);
4413 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4414 "bad tx buftype %d", tx->tx_buftype);
4416 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4419 case GNILND_MSG_GET_DONE:
4420 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4421 msg->gnm_u.completion.gncm_cookie);
4425 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4426 "bad tx buftype %d", tx->tx_buftype);
4428 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4429 msg->gnm_u.completion.gncm_retval);
4431 kgnilnd_finalize_rx_done(tx, msg);
4433 case GNILND_MSG_GET_DONE_REV:
4434 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4435 msg->gnm_u.completion.gncm_cookie);
4439 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4440 "bad tx buftype %d", tx->tx_buftype);
4442 kgnilnd_finalize_rx_done(tx, msg);
4445 case GNILND_MSG_PUT_DONE_REV:
4446 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4447 msg->gnm_u.completion.gncm_cookie);
4452 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4453 "bad tx buftype %d", tx->tx_buftype);
4455 kgnilnd_finalize_rx_done(tx, msg);
4457 case GNILND_MSG_PUT_NAK_REV:
4458 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4459 msg->gnm_u.completion.gncm_cookie);
4464 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4465 "bad tx buftype %d", tx->tx_buftype);
4467 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4472 if (rc < 0) /* protocol/comms error */
4473 kgnilnd_close_conn(conn, rc);
4475 if (repost && rx != NULL) {
4476 kgnilnd_consume_rx(rx);
4479 /* we got an event so assume more there and call for reschedule */
4481 kgnilnd_schedule_conn(conn);
4485 /* Do the failure injections that we need to affect conn processing in the following function.
4486 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4487 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4490 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4494 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4498 /* short circuit out when not set */
4499 if (likely(!cfs_fail_loc)) {
4503 /* failure injection to test for stack reset clean ups */
4504 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4505 /* we can't rely on busy loops being nice enough to get the
4506 * stack reset triggered - it'd just spin on this conn */
4507 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4510 GOTO(did_fail_loc, rc);
4513 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4514 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4516 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4517 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4520 GOTO(did_fail_loc, rc);
4524 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4525 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4526 * will be closed by kgnilnd_complete_closed_conn.
4528 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4529 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4532 GOTO(did_fail_loc, rc);
4535 /* this one is a bit gross - we can't hold the mutex from process_conns
4536 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4537 * and moving onto finish_connect
4538 * so, we'll just set the rc - kgnilnd_process_conns will clear
4539 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4540 * and effectively getting this thread to sleep */
4541 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4544 GOTO(did_fail_loc, rc);
4552 kgnilnd_send_conn_close(kgn_conn_t *conn)
4556 /* we are closing the conn - we will try to send the CLOSE msg
4557 * but will not wait for anything else to flush */
4559 /* send the close if not already done so or received one */
4560 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4561 /* set close_sent regardless of the success of the
4562 * CLOSE message. We are going to try once and then
4563 * kick him out of the sandbox */
4564 conn->gnc_close_sent = 1;
4567 /* EP might be null already if remote side initiated a new connection.
4568 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4569 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4571 if (conn->gnc_ephandle != NULL) {
4574 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE,
4575 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid));
4577 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4578 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4579 tx->tx_qtime = jiffies;
4581 if (tx->tx_id.txe_idx == 0) {
4582 rc = kgnilnd_set_tx_id(tx, conn);
4584 kgnilnd_tx_done(tx, rc);
4588 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4591 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4592 kgnilnd_tx_done(tx, -EAGAIN);
4594 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4596 /* It wasnt sent and we dont care. */
4597 kgnilnd_tx_done(tx, rc);
4605 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4606 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4607 conn->gnc_state = GNILND_CONN_CLOSED;
4608 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4609 /* mark this conn as CLOSED now that we processed it
4610 * do after TX, so we can use CLOSING in asserts */
4614 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4615 /* simulate a RX CLOSE after the timeout but before
4616 * the scheduler thread gets it */
4617 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4618 conn->gnc_peer_error = -ETIMEDOUT;
4620 /* schedule to allow potential CLOSE and get the complete phase run */
4621 kgnilnd_schedule_conn(conn);
4625 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4630 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4631 int log_retrans, log_retrans_level;
4632 static int last_map_version;
4635 spin_lock(&dev->gnd_lock);
4636 if (list_empty(&dev->gnd_map_tx)) {
4637 /* if the list is empty make sure we dont have a timer running */
4638 del_singleshot_timer_sync(&dev->gnd_map_timer);
4639 spin_unlock(&dev->gnd_lock);
4643 dev->gnd_sched_alive = jiffies;
4645 /* we'll retry as fast as possible up to 25% of the limit, then we start
4646 * backing off until our map version changes - indicating we unmapped
4648 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4649 if (likely(dev->gnd_map_attempt == 0) ||
4650 time_after_eq(jiffies, dev->gnd_next_map) ||
4651 last_map_version != dev->gnd_map_version) {
4653 /* if this is our first attempt at mapping set last mapped to current
4654 * jiffies so we can timeout our attempt correctly.
4656 if (dev->gnd_map_attempt == 0)
4657 dev->gnd_last_map = jiffies;
4659 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4660 spin_unlock(&dev->gnd_lock);
4664 /* delete the previous timer if it exists */
4665 del_singleshot_timer_sync(&dev->gnd_map_timer);
4666 /* stash the last map version to let us know when a good one was seen */
4667 last_map_version = dev->gnd_map_version;
4669 /* we need to to take the lock and continually refresh the head of the list as
4670 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4671 * allowing them to squeeze in */
4673 while (!list_empty(&dev->gnd_map_tx)) {
4674 /* make sure we break out early on quiesce */
4675 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4676 /* always break with lock held - we unlock outside loop */
4680 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4682 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4685 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4686 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4687 /* if conn is dying, mark tx in tx_ref_table for
4688 * kgnilnd_complete_closed_conn to finish up */
4689 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4692 /* tx was moved to DYING, get next */
4696 spin_unlock(&dev->gnd_lock);
4697 rc = kgnilnd_send_mapped_tx(tx, 1);
4699 /* We made it! skip error handling.. */
4701 /* OK to continue on +ve errors as it won't get seen until
4702 * this function is called again - we operate on a copy of the original
4703 * list and not the live list */
4704 spin_lock(&dev->gnd_lock);
4705 /* reset map attempts back to zero we successfully
4706 * mapped so we can reset our timers */
4707 dev->gnd_map_attempt = 0;
4709 } else if (rc == -EAGAIN) {
4710 spin_lock(&dev->gnd_lock);
4711 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4712 spin_unlock(&dev->gnd_lock);
4713 GOTO(get_out_mapped, rc);
4714 } else if (rc != -ENOMEM) {
4715 /* carp, failure we can't handle */
4716 kgnilnd_tx_done(tx, rc);
4717 spin_lock(&dev->gnd_lock);
4718 /* reset map attempts back to zero we dont know what happened but it
4719 * wasnt a failed mapping
4721 dev->gnd_map_attempt = 0;
4725 /* time to handle the retry cases.. lock so we dont have 2 threads
4726 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4728 spin_lock(&dev->gnd_lock);
4729 dev->gnd_map_attempt++;
4730 if (dev->gnd_map_attempt < fast_remaps) {
4731 /* do nothing we just want it to go as fast as possible.
4732 * just set gnd_next_map to current jiffies so it will process
4733 * as fast as possible.
4735 dev->gnd_next_map = jiffies;
4737 /* Retry based on GNILND_MAP_RETRY_RATE */
4738 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4741 /* only log occasionally once we've retried fast_remaps */
4742 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4743 ((dev->gnd_map_attempt % fast_remaps) == 0);
4744 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4746 /* make sure we are not off in the weeds with this tx */
4747 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4748 GNIDBG_TX(D_NETERROR, tx,
4749 "giving up on TX, too many retries", NULL);
4750 spin_unlock(&dev->gnd_lock);
4751 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4752 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4753 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4755 tx->tx_putinfo.gnpam_dst_cookie,
4756 tx->tx_msg.gnm_srcnid);
4758 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4760 tx->tx_getinfo.gngm_cookie,
4761 tx->tx_msg.gnm_srcnid);
4763 kgnilnd_tx_done(tx, -ENOMEM);
4764 GOTO(get_out_mapped, rc);
4766 GNIDBG_TX(log_retrans_level, tx,
4767 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4768 "nq_map %d mdd# %d/%d GART %ld",
4769 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4770 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4771 atomic_read(&dev->gnd_nq_map),
4772 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4773 atomic64_read(&dev->gnd_nbytes_map));
4776 /* we need to stop processing the rest of the list, so add it back in */
4777 /* set timer to wake device when we need to schedule this tx */
4778 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4779 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4780 spin_unlock(&dev->gnd_lock);
4781 GOTO(get_out_mapped, rc);
4783 spin_unlock(&dev->gnd_lock);
4789 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4794 int error_inject = 0;
4798 spin_lock(&dev->gnd_lock);
4799 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4800 dev->gnd_sched_alive = jiffies;
4804 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4805 /* break with lock held */
4809 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4810 list_del_init(&conn->gnc_schedlist);
4812 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4815 if (!list_empty(&conn->gnc_delaylist))
4816 list_del_init(&conn->gnc_delaylist);
4817 spin_unlock(&dev->gnd_lock);
4819 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4821 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4822 conn_sched != GNILND_CONN_PROCESS,
4823 "conn %p on ready list but in bad state: %d\n",
4826 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4827 conn, kgnilnd_conn_state2str(conn));
4830 set_mb(conn->gnc_last_sched_do, jiffies);
4832 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4834 /* based on intent see if we should run again. */
4835 rc = kgnilnd_schedule_process_conn(conn, intent);
4837 /* drop ref from gnd_ready_conns */
4838 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4839 down_write(&dev->gnd_conn_sem);
4840 kgnilnd_conn_decref(conn);
4841 up_write(&dev->gnd_conn_sem);
4842 } else if (rc != 1) {
4843 kgnilnd_conn_decref(conn);
4845 /* clear this so that scheduler thread doesn't spin */
4847 /* break with lock held... */
4848 spin_lock(&dev->gnd_lock);
4852 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4853 down_write(&dev->gnd_conn_sem);
4855 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4856 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4857 /* If there are tx's currently in use in another
4858 * thread we dont want to complete the close
4859 * yet. Cycle this conn back through
4861 kgnilnd_schedule_conn(conn);
4863 kgnilnd_complete_closed_conn(conn);
4865 up_write(&dev->gnd_conn_sem);
4866 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4867 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4868 /* serialize SMSG CQs with ep_bind and smsg_release */
4869 down_write(&dev->gnd_conn_sem);
4870 kgnilnd_destroy_conn_ep(conn);
4871 up_write(&dev->gnd_conn_sem);
4872 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4873 /* if we need to do some CLOSE sending, etc done here do it */
4874 down_write(&dev->gnd_conn_sem);
4875 kgnilnd_send_conn_close(conn);
4876 kgnilnd_check_fma_rx(conn);
4877 up_write(&dev->gnd_conn_sem);
4878 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4879 /* start moving traffic if the old conns are cleared out */
4880 down_read(&dev->gnd_conn_sem);
4881 kgnilnd_check_fma_rx(conn);
4882 kgnilnd_process_fmaq(conn);
4883 up_read(&dev->gnd_conn_sem);
4886 rc = kgnilnd_schedule_process_conn(conn, 0);
4888 /* drop ref from gnd_ready_conns */
4889 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4890 down_write(&dev->gnd_conn_sem);
4891 kgnilnd_conn_decref(conn);
4892 up_write(&dev->gnd_conn_sem);
4893 } else if (rc != 1) {
4894 kgnilnd_conn_decref(conn);
4897 /* check list again with lock held */
4898 spin_lock(&dev->gnd_lock);
4901 /* If we are short circuiting due to timing we want to be scheduled
4902 * as soon as possible.
4904 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4907 spin_unlock(&dev->gnd_lock);
4913 kgnilnd_scheduler(void *arg)
4915 int threadno = (long)arg;
4918 unsigned long deadline = 0;
4921 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
4923 /* all gnilnd threads need to run fairly urgently */
4924 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
4925 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
4926 while (!kgnilnd_data.kgn_shutdown) {
4928 /* Safe: kgn_shutdown only set when quiescent */
4930 /* to quiesce or to not quiesce, that is the question */
4932 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4933 KGNILND_SPIN_QUIESCE;
4936 /* tracking for when thread goes AWOL */
4937 dev->gnd_sched_alive = jiffies;
4939 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
4940 (*kgnilnd_tunables.kgn_sched_timeout + 1));
4941 /* let folks know we are up and kicking
4942 * - they can use this for latency savings, etc
4943 * - only change if IRQ, if IDLE leave alone as that
4944 * schedule_device calls to put us back to IRQ */
4945 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
4947 down_read(&dev->gnd_conn_sem);
4948 /* always check these - they are super low cost */
4949 found_work += kgnilnd_check_fma_send_cq(dev);
4950 found_work += kgnilnd_check_fma_rcv_cq(dev);
4952 /* rdma CQ doesn't care about eps */
4953 found_work += kgnilnd_check_rdma_cq(dev);
4955 /* move some RDMA ? */
4956 found_work += kgnilnd_process_rdmaq(dev);
4958 /* map some pending RDMA requests ? */
4959 found_work += kgnilnd_process_mapped_tx(dev);
4961 /* the EP for a conn is not destroyed until all the references
4962 * to it are gone, so these checks should be safe
4963 * even if run in parallel with the CQ checking functions
4964 * _AND_ a thread that processes the CLOSED->DONE
4968 up_read(&dev->gnd_conn_sem);
4970 /* process all conns ready now */
4971 found_work += kgnilnd_process_conns(dev, deadline);
4973 /* do an eager check to avoid the IRQ disabling in
4974 * prepare_to_wait and friends */
4977 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
4978 time_before(jiffies, deadline)) {
4980 if ((busy_loops % 10) == 0) {
4981 /* tickle heartbeat and watchdog to ensure our
4982 * piggishness doesn't turn into heartbeat failure */
4983 touch_nmi_watchdog();
4989 /* if we got here, found_work was zero or busy_loops means we
4990 * need to take a break. We'll clear gnd_ready but we'll check
4991 * one last time if there is an IRQ that needs processing */
4993 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
4995 /* the first time this will go LOOP -> IDLE and let us do one final check
4996 * during which we might get an IRQ, then IDLE->IDLE and schedule()
4997 * - this might allow other threads to block us for a bit if they
4998 * try to get the mutex, but that is good as we'd need to wake
4999 * up soon to handle the CQ or other processing anyways */
5001 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5003 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5004 time_after_eq(jiffies, deadline)) {
5006 "yeilding: found_work %d busy_loops %d\n",
5007 found_work, busy_loops);
5009 /* use yield if we are bailing due to busy_loops
5010 * - this will ensure we wake up soonish. This closes
5011 * a race with kgnilnd_device_callback - where it'd
5012 * not call wake_up() because gnd_ready == 1, but then
5013 * we come down and schedule() because of busy_loops.
5014 * We'd not be woken up until something poked our waitq
5015 * again. yield() ensures we wake up without another
5016 * waitq poke in that case */
5017 atomic_inc(&dev->gnd_n_yield);
5018 kgnilnd_data.kgn_last_condresched = jiffies;
5020 CDEBUG(D_INFO, "awake after yeild\n");
5021 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5022 } else if (found_work == GNILND_DEV_IDLE) {
5023 /* busy_loops is low and there is nothing to do,
5024 * go to sleep and wait for a waitq poke */
5026 "scheduling: found_work %d busy_loops %d\n",
5027 found_work, busy_loops);
5028 atomic_inc(&dev->gnd_n_schedule);
5029 kgnilnd_data.kgn_last_scheduled = jiffies;
5031 CDEBUG(D_INFO, "awake after schedule\n");
5032 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5034 finish_wait(&dev->gnd_waitq, &wait);
5037 kgnilnd_thread_fini();