2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by Eric Barton <eric@bartonsoftware.com>
7 * Author: James Shimek <jshimek@cray.com>
8 * Author: Nic Henke <nic@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #include <linux/nmi.h>
29 #include <linux/pagemap.h>
31 #include <libcfs/linux/linux-mem.h>
35 /* this is useful when needed to debug wire corruption. */
37 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
45 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
46 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
49 } else if (len >= 16) {
51 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
52 prefix, ptr, *(ptr), *(ptr + 1));
56 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
65 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
67 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
68 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
69 msg->gnm_magic, msg->gnm_version,
70 msg->gnm_type, msg->gnm_srcnid,
71 msg->gnm_connstamp, msg->gnm_seq,
72 msg->gnm_cksum, msg->gnm_payload_cksum,
73 msg->gnm_payload_len);
77 kgnilnd_schedule_device(kgn_device_t *dev)
79 short already_live = 0;
81 /* we'll only want to wake if the scheduler thread
82 * has come around and set ready to zero */
83 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
86 wake_up_all(&dev->gnd_waitq);
90 void kgnilnd_schedule_device_timer(cfs_timer_cb_arg_t data)
92 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_map_timer);
94 kgnilnd_schedule_device(dev);
97 void kgnilnd_schedule_device_timer_rd(cfs_timer_cb_arg_t data)
99 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_rdmaq_timer);
101 kgnilnd_schedule_device(dev);
105 kgnilnd_device_callback(__u32 devid, __u64 arg)
108 int index = (int) arg;
110 if (index >= kgnilnd_data.kgn_ndevs) {
111 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
112 * interrupt context. */
113 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
118 dev = &kgnilnd_data.kgn_devices[index];
119 /* just basic sanity */
120 if (dev->gnd_id == devid) {
121 kgnilnd_schedule_device(dev);
123 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
128 /* sched_intent values:
129 * < 0 : do not reschedule under any circumstances
130 * == 0: reschedule if someone marked him WANTS_SCHED
131 * > 0 : force a reschedule */
132 /* Return code 0 means it did not schedule the conn, 1
133 * means it successfully scheduled the conn.
137 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
141 /* move back to IDLE but save previous state.
142 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
143 * let the xchg there handle any racing callers to get it
144 * onto gnd_ready_conns */
146 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
147 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
148 conn_sched == GNILND_CONN_PROCESS,
149 "conn %p after process in bad state: %d\n",
152 if (sched_intent >= 0) {
153 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
154 return kgnilnd_schedule_conn_refheld(conn, 1);
160 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
164 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
166 kgn_device_t *dev = conn->gnc_device;
170 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
171 /* we only care about the last person who marked want_sched since they
172 * are most likely the culprit
174 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
175 conn->gnc_sched_line = line;
176 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
177 * the chance to put it onto gnd_ready_conns.
178 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
179 * the conn in process_conns will take care of moving it back to
180 * SCHED when it is done processing */
182 if (sched == GNILND_CONN_IDLE) {
183 /* if the conn is already scheduled, we've already requested
184 * the scheduler thread wakeup */
186 /* Add a reference to the conn if we are not holding a reference
187 * already from the exisiting scheduler. We now use the same
188 * reference if we need to reschedule a conn while in a scheduler
191 kgnilnd_conn_addref(conn);
193 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
196 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
198 spin_lock(&dev->gnd_lock);
199 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
201 spin_unlock(&dev->gnd_lock);
202 set_mb(conn->gnc_last_sched_ask, jiffies);
205 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
209 /* make sure thread(s) going to process conns - but let it make
210 * separate decision from conn schedule */
212 kgnilnd_schedule_device(dev);
217 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
219 kgn_device_t *dev = conn->gnc_device;
221 spin_lock(&dev->gnd_lock);
222 if (list_empty(&conn->gnc_delaylist)) {
223 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
226 spin_unlock(&dev->gnd_lock);
228 kgnilnd_schedule_device(dev);
233 kgnilnd_schedule_dgram(kgn_device_t *dev)
237 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
238 if (wake != GNILND_DGRAM_SCHED) {
239 wake_up(&dev->gnd_dgram_waitq);
241 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
246 kgnilnd_free_tx(kgn_tx_t *tx)
248 /* taken from kgnilnd_tx_add_state_locked */
250 LASSERTF((tx->tx_list_p == NULL &&
251 tx->tx_list_state == GNILND_TX_ALLOCD) &&
252 list_empty(&tx->tx_list),
253 "tx %p with bad state %s (list_p %p) tx_list %s\n",
254 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
255 list_empty(&tx->tx_list) ? "empty" : "not empty");
257 atomic_dec(&kgnilnd_data.kgn_ntx);
259 /* we only allocate this if we need to */
260 if (tx->tx_phys != NULL) {
261 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
262 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
263 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
266 /* Only free the buffer if we used it */
267 if (tx->tx_buffer_copy != NULL) {
268 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
269 tx->tx_buffer_copy = NULL;
270 CDEBUG(D_MALLOC, "vfreed buffer2\n");
273 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
275 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
276 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
280 kgnilnd_alloc_tx (void)
284 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
287 tx = kmem_cache_zalloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
289 CERROR("failed to allocate tx\n");
292 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
295 /* setup everything here to minimize time under the lock */
296 tx->tx_buftype = GNILND_BUF_NONE;
297 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
298 INIT_LIST_HEAD(&tx->tx_list);
299 INIT_LIST_HEAD(&tx->tx_map_list);
300 tx->tx_list_state = GNILND_TX_ALLOCD;
302 atomic_inc(&kgnilnd_data.kgn_ntx);
307 /* csum_fold needs to be run on the return value before shipping over the wire */
308 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
310 /* we don't use offset as every one is passing a buffer reference that already
311 * includes the offset into the base address.
314 kgnilnd_cksum(void *ptr, size_t nob)
318 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
320 /* don't use magic 'no checksum' value */
324 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
331 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
332 unsigned int offset, unsigned int nob, int dump_blob)
338 unsigned int fraglen;
344 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
345 kiov, nkiov, offset, nob, dump_blob);
347 /* if loops changes, please change kgnilnd_setup_phys_buffer */
349 while (offset >= kiov->bv_len) {
350 offset -= kiov->bv_len;
356 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
357 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
359 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
360 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
362 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
363 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
365 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
366 odd, kiov[0].bv_len, offset, nob);
368 for (i = 0; i < nkiov; i++) {
369 pages[i] = kiov[i].bv_page;
372 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
374 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
376 /* return zero to avoid killing tx - we'll just get warning on console
377 * when remote end sees zero checksum */
380 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
382 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
387 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
388 (void *)addr + kiov[0].bv_offset +
391 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
392 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
396 fraglen = min(kiov->bv_len - offset, nob);
398 /* make dang sure we don't send a bogus checksum if somehow we get
399 * an odd length fragment on anything but the last entry in a kiov -
400 * we know from kgnilnd_setup_rdma_buffer that we can't have non
401 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
402 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
403 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
404 fraglen, nkiov, nob, kiov->bv_len,
407 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
409 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
412 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
413 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
414 addr, fraglen, offset);
419 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
421 kunmap(kiov->bv_page);
428 /* iov must not run out before end of data */
429 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
434 retsum = csum_fold(cksum);
436 /* don't use magic 'no checksum' value */
440 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
446 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
448 msg->gnm_magic = GNILND_MSG_MAGIC;
449 msg->gnm_version = GNILND_MSG_VERSION;
450 msg->gnm_type = type;
451 msg->gnm_payload_len = 0;
452 msg->gnm_srcnid = source;
453 /* gnm_connstamp gets set when FMA is sent */
454 /* gnm_srcnid is set on creation via function argument
455 * The right interface/net and nid is passed in when the message
461 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
463 kgn_tx_t *tx = kgnilnd_alloc_tx();
466 kgnilnd_init_msg(&tx->tx_msg, type, source);
468 CERROR("couldn't allocate new tx type %s!\n",
469 kgnilnd_msgtype2str(type));
476 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
482 case GNILND_MSG_GET_REQ:
483 case GNILND_MSG_GET_DONE:
484 nak_type = GNILND_MSG_GET_NAK;
486 case GNILND_MSG_PUT_REQ:
487 case GNILND_MSG_PUT_ACK:
488 case GNILND_MSG_PUT_DONE:
489 nak_type = GNILND_MSG_PUT_NAK;
491 case GNILND_MSG_PUT_REQ_REV:
492 case GNILND_MSG_PUT_DONE_REV:
493 nak_type = GNILND_MSG_PUT_NAK_REV;
495 case GNILND_MSG_GET_REQ_REV:
496 case GNILND_MSG_GET_ACK_REV:
497 case GNILND_MSG_GET_DONE_REV:
498 nak_type = GNILND_MSG_GET_NAK_REV;
501 CERROR("invalid msg type %s (%d)\n",
502 kgnilnd_msgtype2str(rx_type), rx_type);
505 /* only allow NAK on error and truncate to zero */
506 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
507 error, conn, cookie);
509 tx = kgnilnd_new_tx_msg(nak_type, source);
511 CNETERR("can't get TX to NAK RDMA to %s\n",
512 libcfs_nid2str(conn->gnc_peer->gnp_nid));
516 tx->tx_msg.gnm_u.completion.gncm_retval = error;
517 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
518 kgnilnd_queue_tx(conn, tx);
522 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
523 struct bio_vec *kiov,
524 unsigned int offset, unsigned int nob)
526 kgn_msg_t *msg = &tx->tx_msg;
529 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
530 * gni_smsg_send to send that as the payload */
532 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
535 tx->tx_buffer = NULL;
538 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
539 niov = round_up(nob + offset + kiov->bv_offset,
543 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
544 "bad niov %d msg %p kiov %p offset %d nob%d\n",
545 niov, msg, kiov, offset, nob);
547 while (offset >= kiov->bv_len) {
548 offset -= kiov->bv_len;
553 for (i = 0; i < niov; i++) {
554 /* We can't have a bv_offset on anything but the first
555 * entry, otherwise we'll have a hole at the end of the
556 * mapping as we only map whole pages.
557 * Also, if we have a bv_len < PAGE_SIZE but we need to
558 * map more than bv_len, we will also have a whole at
559 * the end of that page which isn't allowed
561 if ((kiov[i].bv_offset != 0 && i > 0) ||
562 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
564 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
565 i, offset, nob, kiov->bv_offset,
569 tx->tx_imm_pages[i] = kiov[i].bv_page;
572 /* hijack tx_phys for the later unmap */
574 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
576 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
577 kiov[0].bv_offset + offset;
578 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
579 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
580 nob, kiov, tx->tx_buffer);
582 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
583 if (tx->tx_phys == NULL) {
584 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
588 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
589 /* make sure we take into account the kiov offset as the
590 * start of the buffer
592 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
595 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
596 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
598 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
603 /* checksum payload early - it shouldn't be changing after lnd_send */
604 if (*kgnilnd_tunables.kgn_checksum >= 2) {
605 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
606 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
607 msg->gnm_payload_cksum += 0xe00e;
609 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
610 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
614 msg->gnm_payload_cksum = 0;
621 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
622 unsigned int offset, unsigned int nob)
624 gni_mem_segment_t *phys;
626 unsigned int fraglen;
628 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
632 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
634 /* only allocate this if we are going to use it */
635 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
637 if (tx->tx_phys == NULL) {
638 CERROR("failed to allocate tx_phys\n");
643 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
644 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
646 /* if loops changes, please change kgnilnd_cksum_kiov
647 * and kgnilnd_setup_immediate_buffer */
649 while (offset >= kiov->bv_len) {
650 offset -= kiov->bv_len;
656 /* at this point, kiov points to the first page that we'll actually map
657 * now that we've seeked into the koiv for offset and dropped any
658 * leading pages that fall entirely within the offset */
659 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
662 /* bv_offset is start of 'valid' buffer, so index offset past that */
663 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
666 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
667 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
670 fraglen = min(kiov->bv_len - offset, nob);
672 /* We can't have a bv_offset on anything but the first entry,
673 * otherwise we'll have a hole at the end of the mapping as we
674 * only map whole pages. Only the first page is allowed to
675 * have an offset - we'll add that into tx->tx_buffer and that
676 * will get used when we map in the segments (see
677 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
678 * but we need to map more than bv_len, we will also have a
679 * whole at the end of that page which isn't allowed
681 if ((phys != tx->tx_phys) &&
682 ((kiov->bv_offset != 0) ||
683 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
684 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
685 (int)(phys - tx->tx_phys),
686 offset, nob, kiov->bv_offset, kiov->bv_len);
691 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
692 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
697 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
703 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
704 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
707 phys->address = page_to_phys(kiov->bv_page);
714 /* iov must not run out before end of data */
715 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
719 tx->tx_phys_npages = phys - tx->tx_phys;
724 if (tx->tx_phys != NULL) {
725 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
726 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
727 sizeof(*tx->tx_phys), tx->tx_phys);
734 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
735 struct bio_vec *kiov,
736 unsigned int offset, unsigned int nob)
738 return kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
741 /* kgnilnd_parse_lnet_rdma()
742 * lntmsg - message passed in from lnet.
743 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
744 * nob - actual number of bytes to in this message.
745 * put_len - It is possible for PUTs to have a different length than the
746 * length stored in lntmsg->msg_len since LNET can adjust this
747 * length based on it's buffer size and offset.
748 * lnet_try_match_md() sets the mlength that we use to do the RDMA
752 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
753 unsigned int *offset, unsigned int *nob,
754 struct bio_vec **kiov, int put_len)
756 /* GETs are weird, see kgnilnd_send */
757 if (lntmsg->msg_type == LNET_MSG_GET) {
758 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
761 *kiov = lntmsg->msg_md->md_kiov;
763 *niov = lntmsg->msg_md->md_niov;
764 *nob = lntmsg->msg_md->md_length;
767 *kiov = lntmsg->msg_kiov;
768 *niov = lntmsg->msg_niov;
770 *offset = lntmsg->msg_offset;
775 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
777 unsigned int niov, offset, nob;
778 struct bio_vec *kiov;
779 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
780 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
782 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
783 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
784 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
785 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
786 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
787 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
788 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
790 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
791 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
792 tx->tx_msg.gnm_payload_cksum = 0;
795 if (*kgnilnd_tunables.kgn_checksum < 3) {
796 tx->tx_msg.gnm_payload_cksum = 0;
800 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
802 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
806 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
808 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
810 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
814 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
815 tx->tx_msg.gnm_payload_cksum += 0xd00d;
819 /* kgnilnd_verify_rdma_cksum()
820 * tx - PUT_DONE/GET_DONE matched tx.
821 * rx_cksum - received checksum to compare against.
822 * put_len - see kgnilnd_parse_lnet_rdma comments.
825 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
829 unsigned int niov, offset, nob;
830 struct bio_vec *kiov;
831 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
832 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
834 /* we can only match certain requests */
835 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
836 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
837 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
838 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
839 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
840 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
841 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
843 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
844 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
849 if (*kgnilnd_tunables.kgn_checksum >= 3) {
850 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
851 "no RDMA payload checksum when enabled");
856 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
858 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
861 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
863 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
866 if (cksum != rx_cksum) {
867 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
868 "Bad RDMA payload checksum (%x expected %x); "
869 "kiov 0x%p niov %d nob %u offset %u",
870 cksum, rx_cksum, kiov, niov, nob, offset);
871 switch (dump_on_err) {
874 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
876 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
881 libcfs_debug_dumplog();
887 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
893 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
897 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
898 "already mapped!", NULL);
900 spin_lock(&dev->gnd_map_lock);
901 switch (tx->tx_buftype) {
903 GNIDBG_TX(D_EMERG, tx,
904 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
905 spin_unlock(&dev->gnd_map_lock);
909 case GNILND_BUF_PHYS_MAPPED:
910 bytes = tx->tx_phys_npages * PAGE_SIZE;
911 dev->gnd_map_nphys++;
912 dev->gnd_map_physnop += tx->tx_phys_npages;
916 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
917 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
918 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
919 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
920 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
923 atomic_inc(&dev->gnd_n_mdd);
924 atomic64_add(bytes, &dev->gnd_nbytes_map);
926 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
929 /* we only get here in the valid cases */
930 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
931 dev->gnd_map_version++;
932 spin_unlock(&dev->gnd_map_lock);
936 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
940 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
941 "not mapped!", NULL);
942 spin_lock(&dev->gnd_map_lock);
944 switch (tx->tx_buftype) {
946 GNIDBG_TX(D_EMERG, tx,
947 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
948 spin_unlock(&dev->gnd_map_lock);
952 case GNILND_BUF_PHYS_UNMAPPED:
953 bytes = tx->tx_phys_npages * PAGE_SIZE;
954 dev->gnd_map_nphys--;
955 dev->gnd_map_physnop -= tx->tx_phys_npages;
959 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
960 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
961 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
962 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
963 "bytes_out negative! %lld\n",
964 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
965 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
966 bytes, (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
969 atomic_dec(&dev->gnd_n_mdd);
970 atomic64_sub(bytes, &dev->gnd_nbytes_map);
972 /* we only get here in the valid cases */
973 list_del_init(&tx->tx_map_list);
974 dev->gnd_map_version++;
975 spin_unlock(&dev->gnd_map_lock);
979 kgnilnd_map_buffer(kgn_tx_t *tx)
981 kgn_conn_t *conn = tx->tx_conn;
982 kgn_device_t *dev = conn->gnc_device;
983 __u32 flags = GNI_MEM_READWRITE;
986 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
987 * be called concurrently as there are internal locks that protect
988 * any data structures or HW resources. We just need to ensure
989 * that our concurrency doesn't result in the kgn_device_t
990 * getting nuked while we are in here */
992 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
993 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
995 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
998 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
999 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1002 switch (tx->tx_buftype) {
1006 case GNILND_BUF_NONE:
1007 case GNILND_BUF_IMMEDIATE:
1008 case GNILND_BUF_IMMEDIATE_KIOV:
1009 case GNILND_BUF_PHYS_MAPPED:
1012 case GNILND_BUF_PHYS_UNMAPPED:
1013 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1014 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1015 tx->tx_phys, tx->tx_phys_npages, NULL,
1016 GNI_MEM_PHYS_SEGMENTS | flags,
1018 /* could race with other uses of the map counts, but this is ok
1019 * - this needs to turn into a non-fatal error soon to allow
1020 * GART resource, etc starvation handling */
1021 if (rrc != GNI_RC_SUCCESS) {
1022 GNIDBG_TX(D_NET, tx,
1023 "Can't map %d pages: dev %d phys %u pp %u",
1024 tx->tx_phys_npages, dev->gnd_id,
1025 dev->gnd_map_nphys, dev->gnd_map_physnop);
1026 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1029 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1030 kgnilnd_mem_add_map_list(dev, tx);
1036 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1038 kgn_conn_t *conn = tx->tx_conn;
1039 kgn_mdd_purgatory_t *gmp;
1041 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1042 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1043 " asserting to avoid data corruption\n");
1044 if (tx->tx_buffer_copy)
1045 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1047 gmp->gmp_map_key = tx->tx_map_key;
1049 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1051 /* ensure that we don't have a blank purgatory - indicating the
1052 * conn is not already on purgatory lists - we'd never recover these
1053 * MDD if that were the case */
1054 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1055 "conn 0x%p->%s with NULL purgatory",
1056 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1058 /* link 'er up! - only place we really need to lock for
1059 * concurrent access */
1060 spin_lock(&conn->gnc_list_lock);
1061 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1062 spin_unlock(&conn->gnc_list_lock);
1066 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1070 int hold_timeout = 0;
1072 /* code below relies on +1 relationship ... */
1073 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1074 (GNILND_BUF_PHYS_UNMAPPED + 1));
1076 switch (tx->tx_buftype) {
1080 case GNILND_BUF_NONE:
1081 case GNILND_BUF_IMMEDIATE:
1082 case GNILND_BUF_PHYS_UNMAPPED:
1084 case GNILND_BUF_IMMEDIATE_KIOV:
1085 if (tx->tx_phys != NULL) {
1086 vunmap(tx->tx_phys);
1087 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1088 kunmap(tx->tx_imm_pages[0]);
1090 /* clear to prevent kgnilnd_free_tx from thinking
1091 * this is a RDMA descriptor */
1095 case GNILND_BUF_PHYS_MAPPED:
1096 LASSERT(tx->tx_conn != NULL);
1098 dev = tx->tx_conn->gnc_device;
1100 /* only want to hold if we are closing conn without
1101 * verified peer notification - the theory is that
1102 * a TX error can be communicated in all other cases */
1103 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1104 error != -GNILND_NOPURG &&
1105 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1106 kgnilnd_add_purgatory_tx(tx);
1108 /* The timeout we give to kgni is a deadman stop only.
1109 * we are setting high to ensure we don't have the kgni timer
1110 * fire before ours fires _and_ is handled */
1111 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1113 GNIDBG_TX(D_NET, tx,
1114 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1115 tx->tx_conn->gnc_device, hold_timeout,
1116 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1118 if (tx->tx_buffer_copy != NULL) {
1119 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1120 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1121 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1122 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1124 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1125 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1129 kgnilnd_mem_del_map_list(dev, tx);
1135 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1137 struct lnet_msg *lntmsg0, *lntmsg1;
1138 int status0, status1;
1139 struct lnet_ni *ni = NULL;
1140 kgn_conn_t *conn = tx->tx_conn;
1142 LASSERT(!in_interrupt());
1144 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1145 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1148 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1149 !kgnilnd_conn_clean_errno(completion)) {
1150 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1151 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1152 completion, tx, conn ?
1153 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1154 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1155 kgnilnd_tx_state2str(tx->tx_list_state),
1156 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1159 /* The error codes determine if we hold onto the MDD */
1160 kgnilnd_unmap_buffer(tx, completion);
1162 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1163 * we play nice with the error codes to avoid delivering a failed
1164 * REQUEST and then a REPLY event as well */
1166 /* return -EIO to lnet - it is the magic value for failed sends */
1167 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1169 status1 = completion;
1171 status0 = status1 = completion;
1174 tx->tx_buftype = GNILND_BUF_NONE;
1175 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1177 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1178 * set NULL when we are a tx without a conn */
1180 ni = conn->gnc_peer->gnp_net->gnn_ni;
1182 spin_lock(&conn->gnc_tx_lock);
1184 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1185 (volatile unsigned long *)&conn->gnc_tx_bits),
1186 "conn %p tx %p bit %d already cleared\n",
1187 conn, tx, tx->tx_id.txe_idx);
1189 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1190 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1192 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1193 spin_unlock(&conn->gnc_tx_lock);
1196 kgnilnd_free_tx(tx);
1198 /* finalize AFTER freeing lnet msgs */
1200 /* warning - we should hold no locks here - calling lnet_finalize
1201 * could free up lnet credits, resulting in a call chain back into
1202 * the LND via kgnilnd_send and friends */
1204 lnet_finalize(lntmsg0, status0);
1206 if (lntmsg1 != NULL) {
1207 lnet_finalize(lntmsg1, status1);
1212 kgnilnd_txlist_done(struct list_head *txlist, int error)
1215 int err_printed = 0;
1217 if (list_empty(txlist))
1220 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1221 /* only print the first error */
1223 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1224 list_del_init(&tx->tx_list);
1225 kgnilnd_tx_done(tx, error);
1230 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1234 spin_lock(&conn->gnc_tx_lock);
1236 /* ID zero is NOT ALLOWED!!! */
1239 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1240 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1241 if (id == GNILND_MAX_MSG_ID) {
1242 if (conn->gnc_next_tx != 1) {
1243 /* we only searched from next_tx to end and didn't find
1244 * one, so search again from start */
1245 conn->gnc_next_tx = 1;
1248 /* couldn't find one! */
1249 spin_unlock(&conn->gnc_tx_lock);
1253 /* bump next_tx to prevent immediate reuse */
1254 conn->gnc_next_tx = id + 1;
1256 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1257 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1258 "tx 0x%p already at id %d\n",
1259 conn->gnc_tx_ref_table[id], id);
1261 /* delay these until we have a valid ID - prevents bad clear of the bit
1262 * in kgnilnd_tx_done */
1264 tx->tx_id.txe_cqid = conn->gnc_cqid;
1266 tx->tx_id.txe_idx = id;
1267 conn->gnc_tx_ref_table[id] = tx;
1269 /* Using jiffies to help differentiate against TX reuse - with
1270 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1271 * if we are sending to the same node faster than 256000/sec.
1272 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1274 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1276 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1278 spin_unlock(&conn->gnc_tx_lock);
1283 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1287 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1289 /* we don't care about TX timeout - it could be that the network is slower
1290 * or throttled. We'll keep retranmitting - so if the network is so slow
1291 * that we fill up our mailbox, we'll keep trying to resend that msg
1292 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1293 * that he hasn't send us any traffic in return */
1295 /* some reasonable throttling of the debug message */
1297 unsigned long now = jiffies;
1298 /* XXX Nic: Mystical TX debug here... */
1299 /* We expect retransmissions so only log when D_NET is enabled */
1300 GNIDBG_SMSG_CREDS(D_NET, conn);
1301 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1302 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1303 " last_msg %uus/%uus last_cq %uus/%uus",
1304 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1305 tx->tx_id, tx->tx_retrans,
1306 jiffies_to_usecs(now - tx->tx_cred_wait),
1307 jiffies_to_usecs(now - conn->gnc_last_tx),
1308 jiffies_to_usecs(now - conn->gnc_last_rx),
1309 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1310 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1314 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1315 * to avoid bad ordering with state_lock */
1318 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1319 spinlock_t *state_lock, kgn_tx_list_state_t state)
1321 kgn_conn_t *conn = tx->tx_conn;
1322 kgn_msg_t *msg = &tx->tx_msg;
1324 unsigned long newest_last_rx, timeout;
1327 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1328 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1330 "msg 0x%p type %d wrong payload size %d\n",
1331 msg, msg->gnm_type, immediatenob);
1333 /* make sure we catch all the cases where we'd send on a dirty old mbox
1334 * but allow case for sending CLOSE. Since this check is within the CQ
1335 * mutex barrier and the close message is only sent through
1336 * kgnilnd_send_conn_close the last message out the door will be the
1339 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1340 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1341 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1342 /* Return -ETIME, we are closing the connection already so we dont want to
1343 * have this tx hit the wire. The tx will be killed by the calling function.
1344 * Once the EP is marked dirty the close message will be the last
1345 * thing to hit the wire */
1350 timeout = cfs_time_seconds(conn->gnc_timeout);
1352 newest_last_rx = GNILND_LASTRX(conn);
1354 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1355 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1358 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1359 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1360 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1361 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1362 cfs_duration_sec(now - newest_last_rx),
1363 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1364 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1365 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1369 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1370 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1371 * creating the message this allows the message to contain the correct LNET NID/NET needed
1372 * instead of the one that the peer/conn uses for sending the data.
1374 msg->gnm_connstamp = conn->gnc_my_connstamp;
1375 msg->gnm_payload_len = immediatenob;
1376 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1378 /* always init here - kgn_checksum is a /sys module tunable
1379 * and can be flipped at any point, even between msg init and sending */
1381 if (*kgnilnd_tunables.kgn_checksum) {
1382 /* We must set here and not in kgnilnd_init_msg,
1383 * we could resend this msg many times
1384 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1385 * through init_msg again */
1386 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1387 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1388 msg->gnm_cksum += 0xf00f;
1392 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1393 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1394 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1395 tx->tx_id.txe_idx, immediate, immediatenob);
1397 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1398 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1400 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1401 msg, sizeof(*msg), immediate,
1403 tx->tx_id.txe_smsg_id);
1407 case GNI_RC_SUCCESS:
1408 atomic_inc(&conn->gnc_tx_seq);
1409 conn->gnc_last_tx = jiffies;
1410 /* no locking here as LIVE isn't a list */
1411 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1413 /* this needs to be checked under lock as it might be freed from a completion
1416 if (msg->gnm_type == GNILND_MSG_NOOP) {
1417 set_mb(conn->gnc_last_noop_sent, jiffies);
1420 /* serialize with seeing CQ events for completion on this, as well as
1422 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1423 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1425 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1426 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1427 kgnilnd_peer_alive(conn->gnc_peer);
1428 GNIDBG_SMSG_CREDS(D_NET, conn);
1431 case GNI_RC_NOT_DONE:
1432 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1433 * 1: out of mbox credits
1434 * 2: out of mbox payload credits
1435 * 3: On Aries out of dla credits
1437 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1438 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1439 /* We'll handle this error inline - makes the calling logic much more
1442 /* If no lock, caller doesn't want us to retry */
1443 if (state_lock == NULL) {
1447 /* I need kgni credits to send this. Replace tx at the head of the
1448 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1449 * and bump retrans counts since we are requeueing the tx.
1453 conn->gnc_tx_retrans++;
1455 kgnilnd_tx_log_retrans(conn, tx);
1456 /* add to head of list for the state and retries */
1457 spin_lock(state_lock);
1458 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1459 spin_unlock(state_lock);
1461 /* We only reschedule for a certain number of retries, then
1462 * we will wait for the CQ events indicating a release of SMSG
1464 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1465 kgnilnd_schedule_conn(conn);
1468 /* CQ event coming in signifies either TX completed or
1469 * RX receive. Either of these *could* free up credits
1470 * in the SMSG mbox and we should try sending again */
1471 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1472 tx->tx_conn->gnc_cqid);
1473 kgnilnd_schedule_delay_conn(conn);
1474 /* use +ve return code to let upper layers know they
1475 * should stop looping on sends */
1479 /* handle bad retcode gracefully */
1480 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1481 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1486 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1488 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1489 spinlock_t *state_lock, kgn_tx_list_state_t state)
1491 kgn_device_t *dev = tx->tx_conn->gnc_device;
1492 unsigned long timestamp;
1495 timestamp = jiffies;
1496 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1497 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1498 /* delay in jiffies - we are really concerned only with things that
1499 * result in a schedule() or really holding this off for long times .
1500 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1501 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1503 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1509 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1511 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1512 spinlock_t *state_lock, kgn_tx_list_state_t state)
1514 kgn_conn_t *conn = tx->tx_conn;
1515 kgn_device_t *dev = conn->gnc_device;
1516 unsigned long timestamp;
1519 timestamp = jiffies;
1521 /* technically we are doing bad things with the read_lock on the peer_conn
1522 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1523 * for the mutex. I bet lockdep is gonna flag this one though... */
1525 /* there are a few cases where we don't want the immediate send - like
1526 * when we are in the scheduler thread and it'd harm the latency of
1527 * getting messages up to LNet */
1529 /* rmb for gnd_ready */
1531 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1533 atomic_inc(&conn->gnc_device->gnd_fast_block);
1534 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1535 /* dont hit HW during quiesce */
1537 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1538 /* dont hit HW if stale EPs and conns left to close */
1541 atomic_inc(&conn->gnc_device->gnd_fast_try);
1542 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1543 &conn->gnc_smsg_mutex);
1548 /* we got the mutex and weren't blocked */
1550 /* delay in jiffies - we are really concerned only with things that
1551 * result in a schedule() or really holding this off for long times .
1552 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1553 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1555 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1556 tx->tx_qtime = jiffies;
1557 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1558 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1559 /* _nolock unlocks the mutex for us */
1565 /* lets us know if we can push this RDMA through now */
1567 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1571 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1573 if (bytes_left < 0) {
1574 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1575 atomic_inc(&dev->gnd_rdmaq_nstalls);
1578 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1579 dev->gnd_rdmaq_deadline);
1580 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1581 /* we never del this timer - at worst it schedules us.. */
1588 /* this adds a TX to the queue pending throttling authorization before
1589 * we allow our remote peer to launch a PUT at us */
1591 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1595 /* we cannot go into send_mapped_tx from here as we are holding locks
1596 * and mem registration might end up allocating memory in kgni.
1597 * That said, we'll push this as far as we can into the queue process */
1598 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1601 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1602 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1603 /* lets us know how delayed RDMA is */
1604 tx->tx_qtime = jiffies;
1605 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1607 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1608 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1609 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1610 /* lets us know how delayed mapping is */
1611 tx->tx_qtime = jiffies;
1612 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1615 /* make sure we wake up sched to run this */
1616 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1619 /* push TX through state machine */
1621 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1626 /* set the tx_id here, we delay it until we have an actual conn
1628 * in some cases, the tx_id is already set to provide for things
1629 * like RDMA completion cookies, etc */
1630 if (tx->tx_id.txe_idx == 0) {
1631 rc = kgnilnd_set_tx_id(tx, conn);
1633 kgnilnd_tx_done(tx, rc);
1638 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1639 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1641 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1643 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1644 kgnilnd_tx_done(tx, rc);
1648 switch (tx->tx_msg.gnm_type) {
1649 case GNILND_MSG_PUT_ACK:
1650 case GNILND_MSG_GET_REQ:
1651 case GNILND_MSG_PUT_REQ_REV:
1652 case GNILND_MSG_GET_ACK_REV:
1653 /* hijacking time! If this messages will authorize our peer to
1654 * send his dirty little bytes in an RDMA, we need to get permission */
1655 kgnilnd_queue_rdma(conn, tx);
1657 case GNILND_MSG_IMMEDIATE:
1658 /* try to send right now, can help reduce latency */
1659 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1662 /* it was sent, break out of switch to avoid default case of queueing */
1665 /* needs to queue to try again, so... */
1666 /* fall through... */
1667 case GNILND_MSG_NOOP:
1668 /* Just make sure this goes out first for this conn */
1670 /* fall through... */
1672 spin_lock(&conn->gnc_list_lock);
1673 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1674 tx->tx_qtime = jiffies;
1675 spin_unlock(&conn->gnc_list_lock);
1676 kgnilnd_schedule_conn(conn);
1681 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1684 kgn_peer_t *new_peer = NULL;
1685 kgn_conn_t *conn = NULL;
1691 /* If I get here, I've committed to send, so I complete the tx with
1692 * failure on any problems */
1694 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1695 "tx already has connection %p", tx->tx_conn);
1697 /* do all of the peer & conn searching in one swoop - this avoids
1698 * nastiness when dropping locks and needing to maintain a sane state
1699 * in the face of stack reset or something else nuking peers & conns */
1701 /* I expect to find him, so only take a read lock */
1702 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1704 peer = kgnilnd_find_peer_locked(target->nid);
1706 conn = kgnilnd_find_conn_locked(peer);
1707 /* this could be NULL during quiesce */
1709 /* Connection exists; queue message on it */
1710 kgnilnd_queue_tx(conn, tx);
1711 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1715 /* don't create a connection if the peer is marked down */
1716 if (peer->gnp_state != GNILND_PEER_UP) {
1717 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1723 /* creating peer or conn; I'll need a write lock... */
1724 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1726 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1728 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1730 /* NB - this will not block during normal operations -
1731 * the only writer of this is in the startup/shutdown path. */
1732 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1738 /* ignore previous peer entirely - we cycled the lock, so we
1739 * will create new peer and at worst drop it if peer is still
1741 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
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(target->nid, new_peer, &peer);
1754 /* don't create a connection if the peer is not up */
1755 if (peer->gnp_state != GNILND_PEER_UP) {
1756 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1761 conn = kgnilnd_find_or_create_conn_locked(peer);
1763 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1764 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1769 /* oh hey, found a conn now... magical */
1770 kgnilnd_queue_tx(conn, tx);
1772 /* no conn, must be trying to connect - so we queue for now */
1773 tx->tx_qtime = jiffies;
1774 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1776 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1779 kgnilnd_tx_done(tx, rc);
1784 kgnilnd_rdma(kgn_tx_t *tx, int type,
1785 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1787 kgn_conn_t *conn = tx->tx_conn;
1788 unsigned long timestamp;
1789 gni_post_type_t post_type;
1792 unsigned int desc_nob = nob;
1793 void *desc_buffer = tx->tx_buffer;
1794 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1795 LASSERTF(kgnilnd_tx_mapped(tx),
1796 "unmapped tx %p\n", tx);
1797 LASSERTF(conn != NULL,
1798 "NULL conn on tx %p, naughty, naughty\n", tx);
1799 LASSERTF(nob <= sink->gnrd_nob,
1800 "nob %u > sink->gnrd_nob %d (%p)\n",
1801 nob, sink->gnrd_nob, sink);
1802 LASSERTF(nob <= tx->tx_nob,
1803 "nob %d > tx(%p)->tx_nob %d\n",
1804 nob, tx, tx->tx_nob);
1807 case GNILND_MSG_GET_DONE:
1808 case GNILND_MSG_PUT_DONE:
1809 post_type = GNI_POST_RDMA_PUT;
1811 case GNILND_MSG_GET_DONE_REV:
1812 case GNILND_MSG_PUT_DONE_REV:
1813 post_type = GNI_POST_RDMA_GET;
1816 CERROR("invalid msg type %s (%d)\n",
1817 kgnilnd_msgtype2str(type), type);
1820 if (post_type == GNI_POST_RDMA_GET) {
1821 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1822 * aligned. If the local buffer is not aligned correctly using the copy buffer
1823 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1824 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1825 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1826 * making sure the length and addr are aligned and copying the data into the correct location
1827 * once the transfer has completed.
1829 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1830 (sink->gnrd_addr & 3) ||
1833 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1835 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1837 if ((nob + tx->tx_offset) & 3) {
1838 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1839 atomic_inc(&kgnilnd_data.kgn_rev_length);
1841 desc_nob = (nob + tx->tx_offset);
1844 if (tx->tx_buffer_copy == NULL) {
1845 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1846 * and require at most we allocate a few extra bytes. */
1847 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1849 if (!tx->tx_buffer_copy) {
1850 /* allocation of buffer failed nak the rdma */
1851 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1852 kgnilnd_tx_done(tx, -EFAULT);
1855 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1856 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);
1857 if (rc != GNI_RC_SUCCESS) {
1858 /* Registration Failed nak rdma and kill the tx. */
1859 kgnilnd_vfree(tx->tx_buffer_copy,
1861 tx->tx_buffer_copy = NULL;
1862 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1863 kgnilnd_tx_done(tx, -EFAULT);
1867 desc_map_key = tx->tx_buffer_copy_map_key;
1868 desc_buffer = tx->tx_buffer_copy;
1872 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1873 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1874 tx->tx_rdma_desc.type = post_type;
1875 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1876 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1877 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1878 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1879 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1880 tx->tx_rdma_desc.length = desc_nob;
1881 tx->tx_nob_rdma = nob;
1882 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1883 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1884 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1885 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1886 /* prep final completion message */
1887 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1888 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1889 /* send actual size RDMA'd in retval */
1890 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1892 kgnilnd_compute_rdma_cksum(tx, nob);
1895 kgnilnd_queue_tx(conn, tx);
1899 /* Don't lie (CLOSE == RDMA idle) */
1900 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1901 tx, conn, conn->gnc_close_sent);
1903 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1904 "0x%x cookie:%#llx",
1905 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1907 /* set CQ dedicated for RDMA */
1908 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1910 timestamp = jiffies;
1911 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1912 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1913 /* delay in jiffies - we are really concerned only with things that
1914 * result in a schedule() or really holding this off for long times .
1915 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1916 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1918 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
1920 if (rrc == GNI_RC_ERROR_RESOURCE) {
1921 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1922 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1923 kgnilnd_unmap_buffer(tx, 0);
1925 if (tx->tx_buffer_copy != NULL) {
1926 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
1927 tx->tx_buffer_copy = NULL;
1930 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1931 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
1933 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1934 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1938 spin_lock(&conn->gnc_list_lock);
1939 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
1940 tx->tx_qtime = jiffies;
1941 spin_unlock(&conn->gnc_list_lock);
1942 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1943 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1945 /* XXX Nic: is this a place we should handle more errors for
1946 * robustness sake */
1947 LASSERT(rrc == GNI_RC_SUCCESS);
1952 kgnilnd_alloc_rx(void)
1956 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
1958 CERROR("failed to allocate rx\n");
1961 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
1964 /* no memset to zero, we'll always fill all members */
1968 /* release is to just free connection resources
1969 * we use this for the eager path after copying */
1971 kgnilnd_release_msg(kgn_conn_t *conn)
1974 unsigned long timestamp;
1976 CDEBUG(D_NET, "consuming %p\n", conn);
1978 timestamp = jiffies;
1979 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1980 /* delay in jiffies - we are really concerned only with things that
1981 * result in a schedule() or really holding this off for long times .
1982 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1983 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1985 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
1986 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1988 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
1989 GNIDBG_SMSG_CREDS(D_NET, conn);
1991 kgnilnd_schedule_conn(conn);
1995 kgnilnd_consume_rx(kgn_rx_t *rx)
1997 kgn_conn_t *conn = rx->grx_conn;
1998 kgn_msg_t *rxmsg = rx->grx_msg;
2000 /* if we are eager, free the cache alloc'd msg */
2001 if (unlikely(rx->grx_eager)) {
2002 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2003 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2005 /* release ref from eager_recv */
2006 kgnilnd_conn_decref(conn);
2008 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2009 kgnilnd_release_msg(conn);
2012 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2013 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2018 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2020 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2021 int type = lntmsg->msg_type;
2022 struct lnet_process_id target = lntmsg->msg_target;
2023 int target_is_router = lntmsg->msg_target_is_router;
2024 int routing = lntmsg->msg_routing;
2025 unsigned int niov = lntmsg->msg_niov;
2026 struct bio_vec *kiov = lntmsg->msg_kiov;
2027 unsigned int offset = lntmsg->msg_offset;
2028 unsigned int nob = lntmsg->msg_len;
2029 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2030 kgn_net_t *net = ni->ni_data;
2033 /* '1' for consistency with code that checks !mpflag to restore */
2034 unsigned int mpflag = 1;
2035 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2037 /* NB 'private' is different depending on what we're sending.... */
2038 LASSERT(!in_interrupt());
2040 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2041 type, nob, niov, libcfs_id2str(target));
2043 LASSERTF(nob == 0 || niov > 0,
2044 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2045 LASSERTF(niov <= LNET_MAX_IOV,
2046 "lntmsg %p niov %d\n", lntmsg, niov);
2049 mpflag = memalloc_noreclaim_save();
2053 CERROR("lntmsg %p with unexpected type %d\n",
2058 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2066 if (routing || target_is_router)
2067 break; /* send IMMEDIATE */
2069 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2070 * check the eventual sink buffer here - if small enough, remote
2071 * end is perfectly capable of returning data in short message -
2072 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2073 * for IMMEDIATE messages which will have it send a real reply instead
2074 * of doing kgnilnd_recv to have the RDMA continued */
2075 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2078 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2079 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2081 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2087 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2088 lntmsg->msg_md->md_kiov,
2089 0, lntmsg->msg_md->md_length);
2091 CERROR("unable to setup buffer: %d\n", rc);
2092 kgnilnd_tx_done(tx, rc);
2097 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2098 if (tx->tx_lntmsg[1] == NULL) {
2099 CERROR("Can't create reply for GET to %s\n",
2100 libcfs_nid2str(target.nid));
2101 kgnilnd_tx_done(tx, rc);
2106 tx->tx_lntmsg[0] = lntmsg;
2107 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2108 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2110 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2112 /* rest of tx_msg is setup just before it is sent */
2113 kgnilnd_launch_tx(tx, net, &target);
2115 case LNET_MSG_REPLY:
2117 /* to save on MDDs, we'll handle short kiov by vmap'ing
2118 * and sending via SMSG */
2119 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2122 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2123 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2125 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2132 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2135 kgnilnd_tx_done(tx, rc);
2140 tx->tx_lntmsg[0] = lntmsg;
2141 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2142 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2144 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2146 /* rest of tx_msg is setup just before it is sent */
2147 kgnilnd_launch_tx(tx, net, &target);
2151 /* send IMMEDIATE */
2153 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2154 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2156 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2162 rc = kgnilnd_setup_immediate_buffer(tx, niov, kiov, offset, nob);
2164 kgnilnd_tx_done(tx, rc);
2168 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2169 tx->tx_lntmsg[0] = lntmsg;
2170 kgnilnd_launch_tx(tx, net, &target);
2173 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2175 memalloc_noreclaim_restore(mpflag);
2180 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2182 kgn_conn_t *conn = rx->grx_conn;
2183 kgn_msg_t *rxmsg = rx->grx_msg;
2184 unsigned int niov = lntmsg->msg_niov;
2185 struct bio_vec *kiov = lntmsg->msg_kiov;
2186 unsigned int offset = lntmsg->msg_offset;
2187 unsigned int nob = lntmsg->msg_len;
2192 switch (rxmsg->gnm_type) {
2193 case GNILND_MSG_PUT_REQ_REV:
2194 done_type = GNILND_MSG_PUT_DONE_REV;
2197 case GNILND_MSG_GET_REQ:
2198 done_type = GNILND_MSG_GET_DONE;
2201 CERROR("invalid msg type %s (%d)\n",
2202 kgnilnd_msgtype2str(rxmsg->gnm_type),
2207 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2211 rc = kgnilnd_set_tx_id(tx, conn);
2215 rc = kgnilnd_setup_rdma_buffer(tx, niov, kiov, offset, nob);
2219 tx->tx_lntmsg[0] = lntmsg;
2220 tx->tx_getinfo = rxmsg->gnm_u.get;
2222 /* we only queue from kgnilnd_recv - we might get called from other contexts
2223 * and we don't want to block the mutex in those cases */
2225 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2226 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2227 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2228 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2233 kgnilnd_tx_done(tx, rc);
2234 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2236 lnet_finalize(lntmsg, rc);
2240 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2243 kgn_rx_t *rx = private;
2244 kgn_conn_t *conn = rx->grx_conn;
2245 kgn_msg_t *rxmsg = rx->grx_msg;
2246 kgn_msg_t *eagermsg = NULL;
2247 kgn_peer_t *peer = NULL;
2248 kgn_conn_t *found_conn = NULL;
2250 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2251 conn, rxmsg, lntmsg);
2253 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2254 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2255 rxmsg->gnm_payload_len);
2258 /* Grab a read lock so the connection doesnt disappear on us
2259 * while we look it up
2261 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2263 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2265 found_conn = kgnilnd_find_conn_locked(peer);
2268 /* Verify the connection found is the same one that the message
2269 * is supposed to be using, if it is not output an error message
2272 if (!peer || !found_conn
2273 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2274 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2275 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2276 peer, conn, found_conn);
2278 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2279 " from %s", rxmsg->gnm_connstamp,
2280 found_conn->gnc_peer_connstamp,
2281 libcfs_nid2str(peer->gnp_nid));
2286 /* add conn ref to ensure it doesn't go away until all eager
2287 * messages processed */
2288 kgnilnd_conn_addref(conn);
2290 /* Now that we have verified the connection is valid and added a
2291 * reference we can remove the read_lock on the peer_conn_lock */
2292 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2294 /* we have no credits or buffers for this message, so copy it
2295 * somewhere for a later kgnilnd_recv */
2296 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2297 *kgnilnd_tunables.kgn_eager_credits) {
2298 CERROR("Out of eager credits to %s\n",
2299 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2303 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2305 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2306 if (eagermsg == NULL) {
2307 kgnilnd_conn_decref(conn);
2308 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2309 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2313 /* copy msg and payload */
2314 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2315 rx->grx_msg = eagermsg;
2318 /* stash this for lnet_finalize on cancel-on-conn-close */
2319 rx->grx_lntmsg = lntmsg;
2321 /* keep the same rx_t, it just has a new grx_msg now */
2322 *new_private = private;
2324 /* release SMSG buffer */
2325 kgnilnd_release_msg(conn);
2331 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2332 int delayed, unsigned int niov,
2333 struct bio_vec *kiov,
2334 unsigned int offset, unsigned int mlen, unsigned int rlen)
2336 kgn_rx_t *rx = private;
2337 kgn_conn_t *conn = rx->grx_conn;
2338 kgn_msg_t *rxmsg = rx->grx_msg;
2344 LASSERT(!in_interrupt());
2345 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2347 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2348 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2349 conn, rxmsg, lntmsg,
2350 niov, kiov, offset, mlen, rlen);
2352 /* we need to lock here as recv can be called from any context */
2353 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2354 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2355 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2357 /* someone closed the conn after we copied this out, nuke it */
2358 kgnilnd_consume_rx(rx);
2359 lnet_finalize(lntmsg, conn->gnc_error);
2362 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2364 switch (rxmsg->gnm_type) {
2366 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2367 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2368 conn, rx, rxmsg, lntmsg, niov, kiov, offset, mlen, rlen);
2371 case GNILND_MSG_IMMEDIATE:
2372 if (mlen > rxmsg->gnm_payload_len) {
2373 GNIDBG_MSG(D_ERROR, rxmsg,
2374 "Immediate message from %s too big: %d > %d",
2375 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2376 rxmsg->gnm_payload_len);
2378 kgnilnd_consume_rx(rx);
2382 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2383 * right after the kgn_msg_t header - so just 'cute' way of saying
2384 * rxmsg + sizeof(kgn_msg_t) */
2386 /* check payload checksum if sent */
2388 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2389 !rxmsg->gnm_payload_cksum &&
2390 rxmsg->gnm_payload_len != 0)
2391 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2393 if (rxmsg->gnm_payload_cksum != 0) {
2394 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2395 * which is what is used to calculate the cksum on the TX side */
2396 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2398 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2399 GNIDBG_MSG(D_NETERROR, rxmsg,
2400 "Bad payload checksum (%x expected %x)",
2401 pload_cksum, rxmsg->gnm_payload_cksum);
2402 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2404 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2405 &rxmsg[1], rxmsg->gnm_payload_len);
2408 libcfs_debug_dumplog();
2414 /* checksum problems are fatal, kill the conn */
2415 kgnilnd_consume_rx(rx);
2416 kgnilnd_close_conn(conn, rc);
2421 lnet_copy_flat2kiov(
2423 *kgnilnd_tunables.kgn_max_immediate,
2424 &rxmsg[1], 0, mlen);
2426 kgnilnd_consume_rx(rx);
2427 lnet_finalize(lntmsg, 0);
2430 case GNILND_MSG_PUT_REQ:
2431 /* LNET wants to truncate or drop transaction, sending NAK */
2433 kgnilnd_consume_rx(rx);
2434 lnet_finalize(lntmsg, 0);
2436 /* only error if lntmsg == NULL, otherwise we are just
2437 * short circuiting the rdma process of 0 bytes */
2438 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2439 lntmsg == NULL ? -ENOENT : 0,
2440 rxmsg->gnm_u.get.gngm_cookie,
2444 /* sending ACK with sink buff. info */
2445 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2447 kgnilnd_consume_rx(rx);
2451 rc = kgnilnd_set_tx_id(tx, conn);
2453 GOTO(nak_put_req, rc);
2456 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2457 kiov, offset, mlen);
2459 GOTO(nak_put_req, rc);
2462 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2463 rxmsg->gnm_u.putreq.gnprm_cookie;
2464 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2465 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2466 (__u64)((unsigned long)tx->tx_buffer);
2467 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2469 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2470 tx->tx_qtime = jiffies;
2471 /* we only queue from kgnilnd_recv - we might get called from other contexts
2472 * and we don't want to block the mutex in those cases */
2474 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2475 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2476 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2477 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2479 kgnilnd_consume_rx(rx);
2483 /* make sure we send an error back when the PUT fails */
2484 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2485 kgnilnd_tx_done(tx, rc);
2486 kgnilnd_consume_rx(rx);
2488 /* return magic LNet network error */
2490 case GNILND_MSG_GET_REQ_REV:
2491 /* LNET wants to truncate or drop transaction, sending NAK */
2493 kgnilnd_consume_rx(rx);
2494 lnet_finalize(lntmsg, 0);
2496 /* only error if lntmsg == NULL, otherwise we are just
2497 * short circuiting the rdma process of 0 bytes */
2498 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2499 lntmsg == NULL ? -ENOENT : 0,
2500 rxmsg->gnm_u.get.gngm_cookie,
2504 /* lntmsg can be null when parsing a LNET_GET */
2505 if (lntmsg != NULL) {
2506 /* sending ACK with sink buff. info */
2507 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2509 kgnilnd_consume_rx(rx);
2513 rc = kgnilnd_set_tx_id(tx, conn);
2515 GOTO(nak_get_req_rev, rc);
2517 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2518 kiov, offset, mlen);
2520 GOTO(nak_get_req_rev, rc);
2522 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2523 rxmsg->gnm_u.putreq.gnprm_cookie;
2524 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2525 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2526 (__u64)((unsigned long)tx->tx_buffer);
2527 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2529 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2531 /* we only queue from kgnilnd_recv - we might get called from other contexts
2532 * and we don't want to block the mutex in those cases */
2534 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2535 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2536 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2537 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2540 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2542 rxmsg->gnm_u.get.gngm_cookie,
2546 kgnilnd_consume_rx(rx);
2550 /* make sure we send an error back when the GET fails */
2551 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2552 kgnilnd_tx_done(tx, rc);
2553 kgnilnd_consume_rx(rx);
2555 /* return magic LNet network error */
2559 case GNILND_MSG_PUT_REQ_REV:
2560 /* LNET wants to truncate or drop transaction, sending NAK */
2562 kgnilnd_consume_rx(rx);
2563 lnet_finalize(lntmsg, 0);
2565 /* only error if lntmsg == NULL, otherwise we are just
2566 * short circuiting the rdma process of 0 bytes */
2567 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2568 lntmsg == NULL ? -ENOENT : 0,
2569 rxmsg->gnm_u.get.gngm_cookie,
2574 if (lntmsg != NULL) {
2576 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2579 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2581 rxmsg->gnm_u.get.gngm_cookie,
2584 kgnilnd_consume_rx(rx);
2586 case GNILND_MSG_GET_REQ:
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,
2597 kgnilnd_consume_rx(rx);
2603 /* needs write_lock on kgn_peer_conn_lock held */
2605 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2607 unsigned long timeout, keepalive;
2608 unsigned long now = jiffies;
2609 unsigned long newest_last_rx;
2612 /* given that we found this conn hanging off a peer, it better damned
2613 * well be connected */
2614 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2615 "conn 0x%p->%s with bad state%s\n", conn,
2616 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2618 kgnilnd_conn_state2str(conn));
2620 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2621 "rx_diff %lu tx_diff %lu\n",
2622 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2623 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2624 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2625 cfs_duration_sec(now - conn->gnc_last_tx));
2627 timeout = cfs_time_seconds(conn->gnc_timeout);
2628 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2630 /* just in case our lack of RX msg processing is gumming up the works - give the
2631 * remove an extra chance */
2633 newest_last_rx = GNILND_LASTRX(conn);
2635 if (time_after_eq(now, newest_last_rx + timeout)) {
2636 uint32_t level = D_CONSOLE|D_NETERROR;
2638 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2641 GNIDBG_CONN(level, conn,
2642 "No gnilnd traffic received from %s for %lu "
2643 "seconds, terminating connection. Is node down? ",
2644 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2645 cfs_duration_sec(now - newest_last_rx));
2649 /* we don't timeout on last_tx stalls - we are going to trust the
2650 * underlying network to let us know when sends are failing.
2651 * At worst, the peer will timeout our RX stamp and drop the connection
2652 * at that point. We'll then see his CLOSE or at worst his RX
2653 * stamp stop and drop the connection on our end */
2655 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2656 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2657 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2658 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2659 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2661 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2662 conn->gnc_last_noop_cq,
2663 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2664 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2665 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2666 set_mb(conn->gnc_last_noop_want, jiffies);
2667 atomic_inc(&conn->gnc_reaper_noop);
2668 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2671 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2674 kgnilnd_queue_tx(conn, tx);
2680 /* needs write_lock on kgn_peer_conn_lock held */
2682 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2683 struct list_head *souls)
2685 unsigned long timeout;
2686 kgn_conn_t *conn, *connN = NULL;
2692 short releaseconn = 0;
2693 unsigned long first_rx = 0;
2694 int purgatory_conn_cnt = 0;
2696 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2697 peer, libcfs_nid2str(peer->gnp_nid),
2698 peer->gnp_reconnect_interval);
2700 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2701 GNILND_MIN_TIMEOUT));
2703 conn = kgnilnd_find_conn_locked(peer);
2705 /* if there is a valid conn, check the queues for timeouts */
2706 rc = kgnilnd_check_conn_timeouts_locked(conn);
2708 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2709 /* simulate a RX CLOSE after the timeout but before
2710 * the scheduler thread gets it */
2711 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2712 conn->gnc_peer_error = -ETIMEDOUT;
2715 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2717 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2718 CDEBUG(D_WARNING, "%s conn timed out, will "
2719 "reconnect upon request from peer\n",
2720 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2722 /* Once we mark closed, any of the scheduler threads could
2723 * get it and move through before we hit the fail loc code */
2724 kgnilnd_close_conn_locked(conn, rc);
2726 /* first_rx is used to decide when to release a conn from purgatory.
2728 first_rx = conn->gnc_first_rx;
2732 /* now regardless of starting new conn, find tx on peer queue that
2733 * are old and smell bad - do this first so we don't trigger
2734 * reconnect on empty queue if we timeout all */
2735 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2736 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2738 LCONSOLE_INFO("could not send to %s due to connection"
2739 " setup failure after %lu seconds\n",
2740 libcfs_nid2str(peer->gnp_nid),
2741 cfs_duration_sec(jiffies - tx->tx_qtime));
2743 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2745 list_add_tail(&tx->tx_list, todie);
2750 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2751 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2752 count, peer, libcfs_nid2str(peer->gnp_nid));
2753 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2754 if (list_empty(&peer->gnp_tx_queue) ||
2755 peer->gnp_connecting == GNILND_PEER_KILL) {
2756 /* we pass down todie to use a common function - but we know there are
2758 kgnilnd_cancel_peer_connect_locked(peer, todie);
2762 /* Don't reconnect if we are still trying to clear out old conns.
2763 * This prevents us sending traffic on the new mbox before ensuring we are done
2764 * with the old one */
2765 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2766 (atomic_read(&peer->gnp_dirty_eps) == 0);
2768 /* fast reconnect after a timeout */
2769 to_reconn = !conn &&
2770 (peer->gnp_last_errno == -ETIMEDOUT) &&
2771 *kgnilnd_tunables.kgn_fast_reconn;
2773 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2774 * to be sent, we'll check the reconnect interval and fire up a new
2775 * connection request */
2778 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2779 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2780 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2782 CDEBUG(D_NET, "starting connect to %s\n",
2783 libcfs_nid2str(peer->gnp_nid));
2784 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2785 "Peer was idle and we have a write_lock, state issue %d\n",
2786 peer->gnp_connecting);
2788 peer->gnp_connecting = GNILND_PEER_CONNECT;
2789 kgnilnd_peer_addref(peer); /* extra ref for connd */
2791 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2792 list_add_tail(&peer->gnp_connd_list,
2793 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2794 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2796 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2799 /* fail_loc to allow us to delay release of purgatory */
2800 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2803 /* This check allows us to verify that the new conn is actually being used. This allows us to
2804 * pull the old conns out of purgatory if they have actually seen traffic.
2805 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2808 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2809 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2810 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2814 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2815 /* check for purgatory timeouts */
2816 if (conn->gnc_in_purgatory) {
2817 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2818 * that way the next time we check it we can detach it from purgatory
2821 if (conn->gnc_state != GNILND_CONN_DONE) {
2822 /* Skip over conns that are currently not DONE. If they arent already scheduled
2823 * for completion something in the state machine is broken.
2828 /* We only detach a conn that is in purgatory if we have received a close message,
2829 * we have a new valid connection that has successfully received data, or an admin
2830 * command tells us we need to detach.
2833 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2834 unsigned long waiting;
2836 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2838 /* C.E: The remote peer is expected to close the
2839 * connection (see kgnilnd_check_conn_timeouts)
2840 * via the reaper thread and nuke out the MDD and
2841 * FMA resources after conn->gnc_timeout has expired
2842 * without an FMA RX */
2843 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2844 " held resources\n",
2845 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2846 cfs_duration_sec(waiting));
2848 kgnilnd_detach_purgatory_locked(conn, souls);
2850 purgatory_conn_cnt++;
2855 /* If we have too many connections in purgatory we could run out of
2856 * resources. Limit the number of connections to a tunable number,
2857 * clean up to the minimum all in one fell swoop... there are
2858 * situations where dvs will retry tx's and we can eat up several
2859 * hundread connection requests at once.
2861 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2862 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2864 if (conn->gnc_in_purgatory &&
2865 conn->gnc_state == GNILND_CONN_DONE) {
2866 CDEBUG(D_NET, "Dropping Held resource due to"
2867 " resource limits being hit\n");
2868 kgnilnd_detach_purgatory_locked(conn, souls);
2870 if (purgatory_conn_cnt-- <
2871 *kgnilnd_tunables.kgn_max_purgatory)
2879 kgnilnd_reaper_check(int idx)
2881 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2882 struct list_head *ctmp, *ctmpN;
2883 LIST_HEAD(geriatrics);
2886 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2888 list_for_each_safe(ctmp, ctmpN, peers) {
2889 kgn_peer_t *peer = NULL;
2891 /* don't timeout stuff if the network is mucked or shutting down */
2892 if (kgnilnd_check_hw_quiesce()) {
2895 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2897 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
2900 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2902 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
2903 kgnilnd_release_purgatory_list(&souls);
2907 kgnilnd_update_reaper_timeout(long timeout)
2909 LASSERT(timeout > 0);
2911 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2913 if (timeout < kgnilnd_data.kgn_new_min_timeout)
2914 kgnilnd_data.kgn_new_min_timeout = timeout;
2916 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2920 kgnilnd_reaper_poke_with_stick(cfs_timer_cb_arg_t arg)
2922 wake_up(&kgnilnd_data.kgn_reaper_waitq);
2926 kgnilnd_reaper(void *arg)
2931 unsigned long next_check_time = jiffies;
2932 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
2933 struct timer_list timer;
2936 /* all gnilnd threads need to run fairly urgently */
2937 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2938 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2940 while (!kgnilnd_data.kgn_shutdown) {
2941 /* I wake up every 'p' seconds to check for timeouts on some
2942 * more peers. I try to check every connection 'n' times
2943 * within the global minimum of all keepalive and timeout
2944 * intervals, to ensure I attend to every connection within
2945 * (n+1)/n times its timeout intervals. */
2946 const int p = GNILND_REAPER_THREAD_WAKE;
2947 const int n = GNILND_REAPER_NCHECKS;
2949 /* to quiesce or to not quiesce, that is the question */
2950 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2951 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2952 KGNILND_SPIN_QUIESCE;
2953 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2956 /* careful with the jiffy wrap... */
2957 timeout = (long)(next_check_time - jiffies);
2960 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
2961 TASK_INTERRUPTIBLE);
2962 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2963 cfs_timer_setup(&timer, kgnilnd_reaper_poke_with_stick,
2964 next_check_time, 0);
2965 mod_timer(&timer, (long) jiffies + timeout);
2967 /* check flag variables before committing */
2968 if (!kgnilnd_data.kgn_shutdown &&
2969 !kgnilnd_data.kgn_quiesce_trigger) {
2970 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2971 timeout, cfs_duration_sec(timeout));
2973 CDEBUG(D_INFO, "awake after schedule\n");
2976 del_singleshot_timer_sync(&timer);
2977 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2978 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
2982 /* new_min_timeout is set from the conn timeouts and keepalive
2983 * this should end up with a min timeout of
2984 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
2985 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
2986 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
2987 CDEBUG(D_NET, "Set new min timeout %ld\n",
2988 current_min_timeout);
2991 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2993 /* Compute how many table entries to check now so I get round
2994 * the whole table fast enough given that I do this at fixed
2995 * intervals of 'p' seconds) */
2996 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
2997 if (kgnilnd_data.kgn_new_min_timeout > n * p)
2998 chunk = (chunk * n * p) /
2999 kgnilnd_data.kgn_new_min_timeout;
3002 for (i = 0; i < chunk; i++) {
3003 kgnilnd_reaper_check(hash_index);
3004 hash_index = (hash_index + 1) %
3005 *kgnilnd_tunables.kgn_peer_hash_size;
3007 next_check_time = (long) jiffies + cfs_time_seconds(p);
3008 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3010 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3013 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3015 kgnilnd_thread_fini();
3020 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3021 unsigned niov, offset, nob;
3022 struct bio_vec *kiov;
3023 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3024 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3027 lnet_copy_flat2kiov(
3030 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3032 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3039 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3042 gni_post_descriptor_t *desc;
3044 kgn_tx_ev_id_t ev_id;
3046 int should_retry, rc;
3047 long num_processed = 0;
3048 kgn_conn_t *conn = NULL;
3049 kgn_tx_t *tx = NULL;
3050 kgn_rdma_desc_t *rdesc;
3055 /* make sure we don't keep looping if we need to reset */
3056 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3057 return num_processed;
3059 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3061 /* we didn't get the mutex, so return that there is still work
3065 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3066 /* a bit gross - but we need a good way to test for
3067 * delayed RDMA completions and the easiest way to do
3068 * that is to delay the RDMA CQ events */
3069 rrc = GNI_RC_NOT_DONE;
3071 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3074 if (rrc == GNI_RC_NOT_DONE) {
3075 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3076 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3077 dev->gnd_id, num_processed);
3078 return num_processed;
3080 dev->gnd_sched_alive = jiffies;
3083 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3084 "this is bad, somehow our credits didn't protect us"
3085 " from CQ overrun\n");
3086 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3087 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3088 event_data, GNI_CQ_GET_TYPE(event_data));
3090 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3092 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3094 /* XXX Nic: Need better error handling here... */
3095 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3096 (rrc == GNI_RC_TRANSACTION_ERROR),
3099 ev_id.txe_cookie = desc->post_id;
3101 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3103 if (conn == NULL || tx == NULL) {
3104 /* either conn or tx was already nuked and this is a "late"
3105 * completion, so drop it */
3109 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3110 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3111 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3112 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3113 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3115 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3117 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3118 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3119 tx->tx_msg.gnm_u.completion.gncm_retval);
3123 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3124 if (tx->tx_buffer_copy != NULL)
3125 kgnilnd_recv_bte_get(tx);
3126 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3129 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3130 if (tx->tx_buffer_copy != NULL)
3131 kgnilnd_recv_bte_get(tx);
3132 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3135 /* remove from rdmaq */
3136 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3137 spin_lock(&conn->gnc_list_lock);
3138 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3139 spin_unlock(&conn->gnc_list_lock);
3140 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3142 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3143 event_data = 1LL << 48;
3147 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3148 atomic_inc(&dev->gnd_rdma_ntx);
3149 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3150 /* transaction succeeded, add into fmaq */
3151 kgnilnd_queue_tx(conn, tx);
3152 kgnilnd_peer_alive(conn->gnc_peer);
3154 /* drop ref from kgnilnd_validate_tx_ev_id */
3155 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3156 kgnilnd_conn_decref(conn);
3161 /* fall through to the TRANSACTION_ERROR case */
3164 /* get stringified version for log messages */
3165 kgnilnd_cq_error_str(event_data, &err_str, 256);
3166 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3168 /* make sure we are not off in the weeds with this tx */
3169 if (tx->tx_retrans >
3170 *kgnilnd_tunables.kgn_max_retransmits) {
3171 GNIDBG_TX(D_NETERROR, tx,
3172 "giving up on TX, too many retries", NULL);
3176 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3177 should_retry ? "transient" : "unrecoverable", err_str);
3179 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3180 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3181 rdesc = &tx->tx_putinfo.gnpam_desc;
3182 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3183 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3185 rdesc = &tx->tx_getinfo.gngm_desc;
3186 rnob = tx->tx_lntmsg[0]->msg_len;
3187 rcookie = tx->tx_getinfo.gngm_cookie;
3192 tx->tx_msg.gnm_type,
3196 kgnilnd_nak_rdma(conn,
3197 tx->tx_msg.gnm_type,
3200 tx->tx_msg.gnm_srcnid);
3201 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3202 kgnilnd_close_conn(conn, -ECOMM);
3205 /* drop ref from kgnilnd_validate_tx_ev_id */
3206 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3207 kgnilnd_conn_decref(conn);
3212 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3216 kgn_tx_ev_id_t ev_id;
3217 kgn_tx_t *tx = NULL;
3218 kgn_conn_t *conn = NULL;
3219 int queued_fma, saw_reply, rc;
3220 long num_processed = 0;
3221 struct list_head *ctmp, *ctmpN;
3224 /* make sure we don't keep looping if we need to reset */
3225 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3226 return num_processed;
3229 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3231 /* we didn't get the mutex, so return that there is still work
3236 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3237 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3239 if (rrc == GNI_RC_NOT_DONE) {
3241 "SMSG send CQ %d not ready (data %#llx) "
3242 "processed %ld\n", dev->gnd_id, event_data,
3245 if (num_processed > 0) {
3246 spin_lock(&dev->gnd_lock);
3247 if (!list_empty(&dev->gnd_delay_conns)) {
3248 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3249 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3250 list_del_init(&conn->gnc_delaylist);
3251 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3252 kgnilnd_schedule_conn_nolock(conn);
3254 spin_unlock(&dev->gnd_lock);
3255 kgnilnd_schedule_device(dev);
3257 spin_unlock(&dev->gnd_lock);
3260 return num_processed;
3263 dev->gnd_sched_alive = jiffies;
3266 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3267 "this is bad, somehow our credits didn't "
3268 "protect us from CQ overrun\n");
3269 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3270 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3271 event_data, GNI_CQ_GET_TYPE(event_data));
3273 /* if SMSG couldn't handle an error, time for conn to die */
3274 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3277 /* need to take the write_lock to ensure atomicity
3278 * on the conn state if we need to close it */
3279 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3280 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3282 /* Conn was destroyed? */
3284 "SMSG CQID lookup %#llx failed\n",
3285 GNI_CQ_GET_INST_ID(event_data));
3286 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3290 kgnilnd_cq_error_str(event_data, &err_str, 256);
3291 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3292 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3294 kgnilnd_close_conn_locked(conn, -ECOMM);
3296 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3298 /* no need to process rest of this tx -
3299 * it is getting canceled */
3303 /* fall through to GNI_RC_SUCCESS case */
3304 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3306 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3307 if (conn == NULL || tx == NULL) {
3308 /* either conn or tx was already nuked and this is a "late"
3309 * completion, so drop it */
3313 tx->tx_conn->gnc_last_tx_cq = jiffies;
3314 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3315 set_mb(conn->gnc_last_noop_cq, jiffies);
3318 /* lock tx_list_state and tx_state */
3319 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3320 spin_lock(&tx->tx_conn->gnc_list_lock);
3322 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3323 "state not GNILND_TX_LIVE_FMAQ", NULL);
3324 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3325 "not waiting for completion", NULL);
3327 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3330 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3332 /* This will trigger other FMA sends that were
3333 * pending this completion */
3334 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3336 /* we either did not expect reply or we already got it */
3337 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3339 spin_unlock(&tx->tx_conn->gnc_list_lock);
3340 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3343 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3345 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3346 kgnilnd_schedule_conn(conn);
3349 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3350 * If saw_reply is true we know that the tx is safe to use as the other thread
3351 * is already finished with it.
3355 /* no longer need to track on the live_fmaq */
3356 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3358 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3359 /* we already got reply & were waiting for
3360 * completion of initial send */
3361 /* to initiate RDMA transaction */
3362 GNIDBG_TX(D_NET, tx,
3363 "Pending RDMA 0x%p type 0x%02x",
3364 tx->tx_msg.gnm_type);
3365 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3366 rc = kgnilnd_send_mapped_tx(tx, 0);
3367 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3369 /* we are done with this tx */
3370 GNIDBG_TX(D_NET, tx,
3371 "Done with tx type 0x%02x",
3372 tx->tx_msg.gnm_type);
3373 kgnilnd_tx_done(tx, tx->tx_rc);
3377 /* drop ref from kgnilnd_validate_tx_ev_id */
3378 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3379 kgnilnd_conn_decref(conn);
3381 /* if we are waiting for a REPLY, we'll handle the tx then */
3382 } /* end for loop */
3386 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3391 long num_processed = 0;
3392 struct list_head *conns;
3393 struct list_head *tmp;
3397 /* make sure we don't keep looping if we need to reset */
3398 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3399 return num_processed;
3402 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3404 /* we didn't get the mutex, so return that there is still work
3408 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3409 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3411 if (rrc == GNI_RC_NOT_DONE) {
3412 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3414 dev->gnd_id, event_data, num_processed);
3415 return num_processed;
3417 dev->gnd_sched_alive = jiffies;
3420 /* this is the only CQ that can really handle transient
3422 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3423 rrc = cfs_fail_val ? cfs_fail_val
3424 : GNI_RC_ERROR_RESOURCE;
3425 if (rrc == GNI_RC_ERROR_RESOURCE) {
3426 /* set overrun too */
3427 event_data |= (1UL << 63);
3428 LASSERTF(GNI_CQ_OVERRUN(event_data),
3429 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3432 /* sender should get error event too and take care
3433 of failed transaction by re-transmitting */
3434 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3435 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3439 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3440 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3441 conn = kgnilnd_cqid2conn_locked(
3442 GNI_CQ_GET_INST_ID(event_data));
3444 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3445 "failed, dropping event %#llx\n",
3446 GNI_CQ_GET_INST_ID(event_data),
3449 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3451 GNI_CQ_GET_INST_ID(event_data),
3452 conn, conn->gnc_peer ?
3453 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3456 conn->gnc_last_rx_cq = jiffies;
3458 /* stash first rx so we can clear out purgatory.
3460 if (conn->gnc_first_rx == 0) {
3461 conn->gnc_first_rx = jiffies;
3463 kgnilnd_peer_alive(conn->gnc_peer);
3464 kgnilnd_schedule_conn(conn);
3466 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3470 /* FMA CQ has overflowed: check ALL conns */
3471 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3472 "conns on device %d\n", dev->gnd_id);
3474 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3476 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3477 conns = &kgnilnd_data.kgn_conns[rc];
3479 list_for_each(tmp, conns) {
3480 conn = list_entry(tmp, kgn_conn_t,
3483 if (conn->gnc_device == dev) {
3484 kgnilnd_schedule_conn(conn);
3485 conn->gnc_last_rx_cq = jiffies;
3489 /* don't block write lockers for too long... */
3490 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3495 /* try_map_if_full should only be used when processing TX from list of
3496 * backlog TX waiting on mappings to free up
3499 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3500 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3503 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3505 /* slight bit of race if multiple people calling, but at worst we'll have
3506 * order altered just a bit... which would not be determenistic anyways */
3507 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3509 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3511 /* We know that we have a GART reservation that should guarantee forward progress.
3512 * This means we don't need to take any extraordinary efforts if we are failing
3513 * mappings here - even if we are holding a very small number of these. */
3515 if (try_map_if_full || (rc == 0)) {
3516 rc = kgnilnd_map_buffer(tx);
3519 /* rc should be 0 if we mapped successfully here, if non-zero
3520 * we are queueing */
3522 /* if try_map_if_full set, they handle requeuing */
3523 if (unlikely(try_map_if_full)) {
3526 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3527 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3528 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3529 /* make sure we wake up sched to run this */
3530 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3531 /* return 0 as this is now queued for later sending */
3536 switch (tx->tx_msg.gnm_type) {
3540 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3541 * remote node where the RDMA will be started
3542 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3543 * be satisified. The rest of the errors are "hard" errors that require
3544 * upper layers to handle themselves.
3545 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3546 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3547 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3548 * GNILND_MSG_GET_DONE_REV.
3550 case GNILND_MSG_GET_REQ:
3551 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3552 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3553 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3554 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3555 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3556 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3557 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3559 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3560 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3562 case GNILND_MSG_PUT_ACK:
3563 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3564 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3565 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3566 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3568 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3569 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3572 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3573 case GNILND_MSG_PUT_DONE:
3574 case GNILND_MSG_PUT_REQ:
3575 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3576 &tx->tx_putinfo.gnpam_desc,
3577 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3578 tx->tx_putinfo.gnpam_dst_cookie);
3579 RETURN(try_map_if_full ? rc : 0);
3581 case GNILND_MSG_GET_DONE:
3582 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3583 &tx->tx_getinfo.gngm_desc,
3584 tx->tx_lntmsg[0]->msg_len,
3585 tx->tx_getinfo.gngm_cookie);
3586 RETURN(try_map_if_full ? rc : 0);
3588 case GNILND_MSG_PUT_REQ_REV:
3589 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3590 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3591 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3592 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3593 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3594 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3595 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3597 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3599 case GNILND_MSG_PUT_DONE_REV:
3600 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3601 &tx->tx_getinfo.gngm_desc,
3603 tx->tx_getinfo.gngm_cookie);
3604 RETURN(try_map_if_full ? rc : 0);
3606 case GNILND_MSG_GET_ACK_REV:
3607 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3608 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3609 /* LNET_GETS are a special case for parse */
3610 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3611 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3613 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3614 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3616 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3617 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3619 case GNILND_MSG_GET_DONE_REV:
3620 case GNILND_MSG_GET_REQ_REV:
3621 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3622 &tx->tx_putinfo.gnpam_desc,
3623 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3624 tx->tx_putinfo.gnpam_dst_cookie);
3625 RETURN(try_map_if_full ? rc : 0);
3633 kgnilnd_process_fmaq(kgn_conn_t *conn)
3636 kgn_tx_t *tx = NULL;
3637 void *buffer = NULL;
3638 unsigned int nob = 0;
3641 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3642 * However I will be rescheduled by an FMA completion event
3643 * when I eventually get some.
3644 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3645 * But it doesn't matter if I try to send a "real" message just
3646 * as I start closing because I'll get scheduled to send the
3649 /* Short circuit if the ep_handle is null we cant send anyway. */
3650 if (conn->gnc_ephandle == NULL)
3653 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3655 spin_lock(&conn->gnc_list_lock);
3657 if (list_empty(&conn->gnc_fmaq)) {
3658 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3660 spin_unlock(&conn->gnc_list_lock);
3662 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3663 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3664 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3665 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3666 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3668 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3669 conn->gnc_last_noop_cq,
3670 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3671 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3672 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3673 atomic_inc(&conn->gnc_sched_noop);
3674 set_mb(conn->gnc_last_noop_want, jiffies);
3676 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3679 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3683 rc = kgnilnd_set_tx_id(tx, conn);
3685 kgnilnd_tx_done(tx, rc);
3691 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3692 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3693 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3694 more_to_do = !list_empty(&conn->gnc_fmaq);
3695 spin_unlock(&conn->gnc_list_lock);
3698 /* if there is no real TX or no NOOP to send, bail */
3703 if (!tx->tx_retrans)
3704 tx->tx_cred_wait = jiffies;
3706 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3707 "tx with zero id", NULL);
3709 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3710 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3711 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3715 switch (tx->tx_msg.gnm_type) {
3719 case GNILND_MSG_NOOP:
3720 case GNILND_MSG_CLOSE:
3721 case GNILND_MSG_IMMEDIATE:
3722 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3723 buffer = tx->tx_buffer;
3727 case GNILND_MSG_GET_DONE:
3728 case GNILND_MSG_PUT_DONE:
3729 case GNILND_MSG_PUT_DONE_REV:
3730 case GNILND_MSG_GET_DONE_REV:
3731 case GNILND_MSG_PUT_NAK:
3732 case GNILND_MSG_GET_NAK:
3733 case GNILND_MSG_GET_NAK_REV:
3734 case GNILND_MSG_PUT_NAK_REV:
3735 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3738 case GNILND_MSG_PUT_REQ:
3739 case GNILND_MSG_GET_REQ_REV:
3740 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3742 case GNILND_MSG_PUT_ACK:
3743 case GNILND_MSG_PUT_REQ_REV:
3744 case GNILND_MSG_GET_ACK_REV:
3745 case GNILND_MSG_GET_REQ:
3746 /* This is really only to handle the retransmit of SMSG once these
3747 * two messages are setup in send_mapped_tx */
3748 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3752 if (likely(rc == 0)) {
3753 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3757 /* don't explicitly reschedule here - we are short credits and will rely on
3758 * kgnilnd_sendmsg to resched the conn if need be */
3760 } else if (rc < 0) {
3761 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3762 * almost certainly a software bug, but lets play nice with the other kids */
3763 kgnilnd_tx_done(tx, rc);
3764 /* just for fun, kick peer in arse - resetting conn might help to correct
3765 * this almost certainly buggy software caused return code */
3766 kgnilnd_close_conn(conn, rc);
3770 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3771 kgnilnd_schedule_conn(conn);
3776 kgnilnd_process_rdmaq(kgn_device_t *dev)
3781 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3785 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3786 unsigned long dead_bump;
3789 /* if we think we need to adjust, take lock to serialize and recheck */
3790 spin_lock(&dev->gnd_rdmaq_lock);
3791 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3792 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3794 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3796 /* roll the bucket forward */
3797 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3799 if (kgnilnd_data.kgn_rdmaq_override &&
3800 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3801 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3806 /* roll current outstanding forward to make sure we carry outstanding
3807 * committment forward
3808 * new_ok starts out as the whole interval value
3809 * - first subtract bytes_out from last interval, as that would push us over
3810 * strict limits for this interval
3811 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3813 * there is a small race here if someone is actively processing mappings and
3814 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3815 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3816 * - if this gives us problems in the future, we could use a read/write lock
3817 * to protect the resetting of these values */
3818 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3819 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3821 CDEBUG(D_NET, "resetting rdmaq bytes to %lld, deadline +%lu -> %lu, current out %lld\n",
3822 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3823 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
3825 spin_unlock(&dev->gnd_rdmaq_lock);
3828 spin_lock(&dev->gnd_rdmaq_lock);
3829 while (!list_empty(&dev->gnd_rdmaq)) {
3832 /* make sure we break out early on quiesce */
3833 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3834 /* always break with lock held - we unlock outside loop */
3838 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3839 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3842 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3843 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3844 /* if conn is dying, mark tx in tx_ref_table for
3845 * kgnilnd_complete_closed_conn to finish up */
3846 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3848 /* tx was moved to DYING, get next */
3851 spin_unlock(&dev->gnd_rdmaq_lock);
3853 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3854 spin_lock(&dev->gnd_rdmaq_lock);
3857 /* no ticket! add back to head */
3858 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3859 /* clear found_work so scheduler threads wait for timer */
3863 /* TX is GO for launch */
3864 tx->tx_qtime = jiffies;
3865 kgnilnd_send_mapped_tx(tx, 0);
3869 spin_unlock(&dev->gnd_rdmaq_lock);
3875 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3877 __swab64s(&d->gnrd_key.qword1);
3878 __swab64s(&d->gnrd_key.qword2);
3879 __swab64s(&d->gnrd_addr);
3880 __swab32s(&d->gnrd_nob);
3883 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3884 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3887 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3889 kgn_tx_ev_id_t ev_id;
3892 /* we use the cookie from the original TX, so we can find the match
3893 * by parsing that and using the txe_idx */
3894 ev_id.txe_cookie = cookie;
3896 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3899 /* check tx to make sure kgni didn't eat it */
3900 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3901 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3903 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
3904 (tx->tx_id.txe_cookie = cookie)),
3905 "conn 0x%p->%s tx_ref_table hosed: wanted "
3906 "txe_cookie %#llx txe_idx %d "
3907 "found tx %p cookie %#llx txe_idx %d\n",
3908 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
3909 cookie, ev_id.txe_idx,
3910 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
3912 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
3913 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
3914 "Unexpected TX type (%x, %x or %x) "
3915 "or state (%x, expected +%x) "
3916 "matched reply from %s\n",
3917 tx->tx_msg.gnm_type, type1, type2,
3918 tx->tx_state, GNILND_TX_WAITING_REPLY,
3919 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3921 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
3922 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
3928 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
3931 kgn_conn_t *conn = tx->tx_conn;
3932 __u64 nob = tx->tx_nob;
3933 __u32 physnop = tx->tx_phys_npages;
3934 int id = tx->tx_id.txe_smsg_id;
3935 int buftype = tx->tx_buftype;
3936 gni_mem_handle_t hndl;
3937 hndl.qword1 = tx->tx_map_key.qword1;
3938 hndl.qword2 = tx->tx_map_key.qword2;
3940 spin_lock(&conn->gnc_list_lock);
3942 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
3943 "not waiting for reply", NULL);
3946 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
3948 if (rc == -EFAULT) {
3949 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
3950 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
3952 if(*kgnilnd_tunables.kgn_efault_lbug) {
3953 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
3954 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
3956 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
3957 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
3958 kgnilnd_tx_state2str(tx->tx_list_state),
3959 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
3964 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
3965 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3966 /* sample under lock as follow on steps require gnc_list_lock
3967 * - or call kgnilnd_tx_done which requires no locks held over
3968 * call to lnet_finalize */
3971 spin_unlock(&conn->gnc_list_lock);
3974 kgnilnd_tx_done(tx, tx->tx_rc);
3979 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
3982 kgn_conn_t *conn = tx->tx_conn;
3984 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
3985 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
3987 /* the gncm_retval is passed in for PUTs */
3988 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
3989 msg->gnm_u.completion.gncm_retval);
3991 kgnilnd_complete_tx(tx, rc);
3995 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4003 kgn_peer_t *peer = conn->gnc_peer;
4006 __u16 tmp_cksum = 0, msg_cksum = 0;
4007 int repost = 1, saw_complete;
4008 unsigned long timestamp, newest_last_rx, timeout;
4012 /* Short circuit if the ep_handle is null.
4013 * It's likely that its about to be closed as stale.
4015 if (conn->gnc_ephandle == NULL)
4018 timestamp = jiffies;
4019 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4020 /* delay in jiffies - we are really concerned only with things that
4021 * result in a schedule() or really holding this off for long times .
4022 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4023 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4025 /* Resample current time as we have no idea how long it took to get the mutex */
4026 timestamp = jiffies;
4028 /* We check here when the last time we received an rx, we do this before
4029 * we call getnext in case the thread has been blocked for a while. If we
4030 * havent received an rx since our timeout value we close the connection
4031 * as we should assume the other side has closed the connection. This will
4032 * stop us from sending replies to a mailbox that is already in purgatory.
4035 timeout = cfs_time_seconds(conn->gnc_timeout);
4036 newest_last_rx = GNILND_LASTRX(conn);
4038 /* Error injection to validate that timestamp checking works and closing the conn */
4039 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4040 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4043 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4044 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4045 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4046 cfs_duration_sec(timestamp - newest_last_rx),
4047 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4048 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4050 kgnilnd_close_conn(conn, rc);
4054 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4056 if (rrc == GNI_RC_NOT_DONE) {
4057 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4058 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4062 /* Instead of asserting when we get mailbox corruption lets attempt to
4063 * close the conn and recover. We can put the conn/mailbox into
4064 * purgatory and let purgatory deal with the problem. If we see
4065 * this NETTERROR reported on production systems in large amounts
4066 * we will need to revisit the state machine to see if we can tighten
4067 * it up further to improve data protection.
4070 if (rrc == GNI_RC_INVALID_STATE) {
4071 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4072 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4073 "detected closing conn %p from peer %s\n", conn,
4074 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4076 kgnilnd_close_conn(conn, rc);
4080 LASSERTF(rrc == GNI_RC_SUCCESS,
4081 "bad rc %d on conn %p from peer %s\n",
4082 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4084 msg = (kgn_msg_t *)prefix;
4086 rx = kgnilnd_alloc_rx();
4088 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4089 kgnilnd_release_msg(conn);
4090 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4091 conn, libcfs_nid2str(peer->gnp_nid));
4095 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4097 timestamp = conn->gnc_last_rx;
4098 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4099 atomic_inc(&conn->gnc_rx_seq);
4101 conn->gnc_last_rx = jiffies;
4102 /* stash first rx so we can clear out purgatory
4104 if (conn->gnc_first_rx == 0)
4105 conn->gnc_first_rx = jiffies;
4107 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4108 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4109 kgnilnd_peer_alive(conn->gnc_peer);
4112 rx->grx_conn = conn;
4114 ktime_get_ts64(&rx->grx_received);
4116 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4119 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4125 kgnilnd_net_decref(net);
4128 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4129 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4131 /* XXX Nic: Do we need to swab cksum */
4132 if (msg->gnm_cksum != 0) {
4133 msg_cksum = msg->gnm_cksum;
4135 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4137 if (tmp_cksum != msg_cksum) {
4138 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4139 tmp_cksum, msg_cksum);
4140 kgnilnd_dump_msg(D_BUFFS, msg);
4145 /* restore checksum for future debug messages */
4146 msg->gnm_cksum = tmp_cksum;
4148 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4149 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4150 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4151 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4156 __swab32s(&msg->gnm_magic);
4157 __swab16s(&msg->gnm_version);
4158 __swab16s(&msg->gnm_type);
4159 __swab64s(&msg->gnm_srcnid);
4160 __swab64s(&msg->gnm_connstamp);
4161 __swab32s(&msg->gnm_seq);
4163 /* NB message type checked below; NOT here... */
4164 switch (msg->gnm_type) {
4165 case GNILND_MSG_GET_ACK_REV:
4166 case GNILND_MSG_PUT_ACK:
4167 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4170 case GNILND_MSG_PUT_REQ_REV:
4171 case GNILND_MSG_GET_REQ:
4172 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4180 if (msg->gnm_version != GNILND_MSG_VERSION) {
4181 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4182 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4187 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4188 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4189 libcfs_nid2str(msg->gnm_srcnid),
4190 libcfs_nid2str(peer->gnp_nid));
4195 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4196 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4197 " expected) from %s",
4198 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4199 libcfs_nid2str(peer->gnp_nid));
4204 if (msg->gnm_seq != seq) {
4205 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4206 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4211 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4213 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4214 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4215 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4216 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4217 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4218 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4219 /* double check state with lock held */
4220 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4221 /* only error if we are not already closing */
4222 if (conn->gnc_peer_error == -ETIMEDOUT) {
4223 unsigned long now = jiffies;
4224 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4226 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4227 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4228 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4230 cfs_duration_sec(now - timestamp),
4231 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4232 atomic_read(&conn->gnc_tx_seq),
4233 cfs_duration_sec(now - conn->gnc_last_tx),
4234 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4235 cfs_duration_sec(now - conn->gnc_last_noop_want),
4236 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4237 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4238 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4239 cfs_duration_sec(now - conn->gnc_last_sched_do),
4240 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4242 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4244 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4248 if (conn->gnc_close_recvd) {
4249 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4250 kgnilnd_msgtype2str(msg->gnm_type),
4251 msg->gnm_type, conn->gnc_close_recvd,
4252 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4257 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4258 /* XXX Nic: log message received on bad connection state */
4262 switch (msg->gnm_type) {
4263 case GNILND_MSG_NOOP:
4264 /* Nothing to do; just a keepalive */
4267 case GNILND_MSG_IMMEDIATE:
4268 /* only get SMSG payload for IMMEDIATE */
4269 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4270 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4271 msg->gnm_srcnid, rx, 0);
4274 case GNILND_MSG_GET_REQ_REV:
4275 case GNILND_MSG_PUT_REQ:
4276 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4277 msg->gnm_srcnid, rx, 1);
4280 case GNILND_MSG_GET_NAK_REV:
4281 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4282 msg->gnm_u.completion.gncm_cookie);
4286 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4288 case GNILND_MSG_PUT_NAK:
4289 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4290 msg->gnm_u.completion.gncm_cookie);
4294 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4296 case GNILND_MSG_PUT_ACK:
4297 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4298 msg->gnm_u.putack.gnpam_src_cookie);
4302 /* store putack data for later: deferred rdma or re-try */
4303 tx->tx_putinfo = msg->gnm_u.putack;
4306 spin_lock(&tx->tx_conn->gnc_list_lock);
4308 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4309 "not waiting for reply", NULL);
4311 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4313 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4314 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4315 /* sample under lock as follow on steps require gnc_list_lock
4316 * - or call kgnilnd_tx_done which requires no locks held over
4317 * call to lnet_finalize */
4320 /* cannot launch rdma if still waiting for fma-msg completion */
4321 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4322 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4323 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4325 spin_unlock(&tx->tx_conn->gnc_list_lock);
4328 rc = kgnilnd_send_mapped_tx(tx, 0);
4330 kgnilnd_tx_done(tx, rc);
4333 case GNILND_MSG_GET_ACK_REV:
4334 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4335 msg->gnm_u.putack.gnpam_src_cookie);
4339 /* store putack data for later: deferred rdma or re-try */
4340 tx->tx_putinfo = msg->gnm_u.putack;
4342 spin_lock(&tx->tx_conn->gnc_list_lock);
4344 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4345 "not waiting for reply", NULL);
4347 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4349 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4350 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4351 /* sample under lock as follow on steps require gnc_list_lock
4352 * - or call kgnilnd_tx_done which requires no locks held over
4353 * call to lnet_finalize */
4356 /* cannot launch rdma if still waiting for fma-msg completion */
4357 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4358 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4359 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4361 spin_unlock(&tx->tx_conn->gnc_list_lock);
4364 rc = kgnilnd_send_mapped_tx(tx, 0);
4366 kgnilnd_tx_done(tx, rc);
4369 case GNILND_MSG_PUT_DONE:
4370 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4371 msg->gnm_u.completion.gncm_cookie);
4375 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4376 "bad tx buftype %d", tx->tx_buftype);
4378 kgnilnd_finalize_rx_done(tx, msg);
4380 case GNILND_MSG_PUT_REQ_REV:
4381 case GNILND_MSG_GET_REQ:
4382 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4383 msg->gnm_srcnid, rx, 1);
4387 case GNILND_MSG_GET_NAK:
4388 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4389 msg->gnm_u.completion.gncm_cookie);
4393 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4394 "bad tx buftype %d", tx->tx_buftype);
4396 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4399 case GNILND_MSG_GET_DONE:
4400 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4401 msg->gnm_u.completion.gncm_cookie);
4405 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4406 "bad tx buftype %d", tx->tx_buftype);
4408 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4409 msg->gnm_u.completion.gncm_retval);
4411 kgnilnd_finalize_rx_done(tx, msg);
4413 case GNILND_MSG_GET_DONE_REV:
4414 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4415 msg->gnm_u.completion.gncm_cookie);
4419 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4420 "bad tx buftype %d", tx->tx_buftype);
4422 kgnilnd_finalize_rx_done(tx, msg);
4425 case GNILND_MSG_PUT_DONE_REV:
4426 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4427 msg->gnm_u.completion.gncm_cookie);
4432 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4433 "bad tx buftype %d", tx->tx_buftype);
4435 kgnilnd_finalize_rx_done(tx, msg);
4437 case GNILND_MSG_PUT_NAK_REV:
4438 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4439 msg->gnm_u.completion.gncm_cookie);
4444 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4445 "bad tx buftype %d", tx->tx_buftype);
4447 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4452 if (rc < 0) /* protocol/comms error */
4453 kgnilnd_close_conn(conn, rc);
4455 if (repost && rx != NULL) {
4456 kgnilnd_consume_rx(rx);
4459 /* we got an event so assume more there and call for reschedule */
4461 kgnilnd_schedule_conn(conn);
4465 /* Do the failure injections that we need to affect conn processing in the following function.
4466 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4467 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4470 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4474 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4478 /* short circuit out when not set */
4479 if (likely(!cfs_fail_loc)) {
4483 /* failure injection to test for stack reset clean ups */
4484 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4485 /* we can't rely on busy loops being nice enough to get the
4486 * stack reset triggered - it'd just spin on this conn */
4487 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4490 GOTO(did_fail_loc, rc);
4493 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4494 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4496 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4497 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4500 GOTO(did_fail_loc, rc);
4504 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4505 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4506 * will be closed by kgnilnd_complete_closed_conn.
4508 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4509 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4512 GOTO(did_fail_loc, rc);
4515 /* this one is a bit gross - we can't hold the mutex from process_conns
4516 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4517 * and moving onto finish_connect
4518 * so, we'll just set the rc - kgnilnd_process_conns will clear
4519 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4520 * and effectively getting this thread to sleep */
4521 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4524 GOTO(did_fail_loc, rc);
4532 kgnilnd_send_conn_close(kgn_conn_t *conn)
4536 /* we are closing the conn - we will try to send the CLOSE msg
4537 * but will not wait for anything else to flush */
4539 /* send the close if not already done so or received one */
4540 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4541 /* set close_sent regardless of the success of the
4542 * CLOSE message. We are going to try once and then
4543 * kick him out of the sandbox */
4544 conn->gnc_close_sent = 1;
4547 /* EP might be null already if remote side initiated a new connection.
4548 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4549 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4551 if (conn->gnc_ephandle != NULL) {
4554 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4556 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4557 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4558 tx->tx_qtime = jiffies;
4560 if (tx->tx_id.txe_idx == 0) {
4561 rc = kgnilnd_set_tx_id(tx, conn);
4563 kgnilnd_tx_done(tx, rc);
4567 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4570 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4571 kgnilnd_tx_done(tx, -EAGAIN);
4573 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4575 /* It wasnt sent and we dont care. */
4576 kgnilnd_tx_done(tx, rc);
4584 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4585 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4586 conn->gnc_state = GNILND_CONN_CLOSED;
4587 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4588 /* mark this conn as CLOSED now that we processed it
4589 * do after TX, so we can use CLOSING in asserts */
4593 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4594 /* simulate a RX CLOSE after the timeout but before
4595 * the scheduler thread gets it */
4596 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4597 conn->gnc_peer_error = -ETIMEDOUT;
4599 /* schedule to allow potential CLOSE and get the complete phase run */
4600 kgnilnd_schedule_conn(conn);
4604 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4609 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4610 int log_retrans, log_retrans_level;
4611 static int last_map_version;
4614 spin_lock(&dev->gnd_lock);
4615 if (list_empty(&dev->gnd_map_tx)) {
4616 /* if the list is empty make sure we dont have a timer running */
4617 del_singleshot_timer_sync(&dev->gnd_map_timer);
4618 spin_unlock(&dev->gnd_lock);
4622 dev->gnd_sched_alive = jiffies;
4624 /* we'll retry as fast as possible up to 25% of the limit, then we start
4625 * backing off until our map version changes - indicating we unmapped
4627 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4628 if (likely(dev->gnd_map_attempt == 0) ||
4629 time_after_eq(jiffies, dev->gnd_next_map) ||
4630 last_map_version != dev->gnd_map_version) {
4632 /* if this is our first attempt at mapping set last mapped to current
4633 * jiffies so we can timeout our attempt correctly.
4635 if (dev->gnd_map_attempt == 0)
4636 dev->gnd_last_map = jiffies;
4638 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4639 spin_unlock(&dev->gnd_lock);
4643 /* delete the previous timer if it exists */
4644 del_singleshot_timer_sync(&dev->gnd_map_timer);
4645 /* stash the last map version to let us know when a good one was seen */
4646 last_map_version = dev->gnd_map_version;
4648 /* we need to to take the lock and continually refresh the head of the list as
4649 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4650 * allowing them to squeeze in */
4652 while (!list_empty(&dev->gnd_map_tx)) {
4653 /* make sure we break out early on quiesce */
4654 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4655 /* always break with lock held - we unlock outside loop */
4659 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4661 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4664 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4665 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4666 /* if conn is dying, mark tx in tx_ref_table for
4667 * kgnilnd_complete_closed_conn to finish up */
4668 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4671 /* tx was moved to DYING, get next */
4675 spin_unlock(&dev->gnd_lock);
4676 rc = kgnilnd_send_mapped_tx(tx, 1);
4678 /* We made it! skip error handling.. */
4680 /* OK to continue on +ve errors as it won't get seen until
4681 * this function is called again - we operate on a copy of the original
4682 * list and not the live list */
4683 spin_lock(&dev->gnd_lock);
4684 /* reset map attempts back to zero we successfully
4685 * mapped so we can reset our timers */
4686 dev->gnd_map_attempt = 0;
4688 } else if (rc == -EAGAIN) {
4689 spin_lock(&dev->gnd_lock);
4690 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4691 spin_unlock(&dev->gnd_lock);
4692 GOTO(get_out_mapped, rc);
4693 } else if (rc != -ENOMEM) {
4694 /* carp, failure we can't handle */
4695 kgnilnd_tx_done(tx, rc);
4696 spin_lock(&dev->gnd_lock);
4697 /* reset map attempts back to zero we dont know what happened but it
4698 * wasnt a failed mapping
4700 dev->gnd_map_attempt = 0;
4704 /* time to handle the retry cases.. lock so we dont have 2 threads
4705 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4707 spin_lock(&dev->gnd_lock);
4708 dev->gnd_map_attempt++;
4709 if (dev->gnd_map_attempt < fast_remaps) {
4710 /* do nothing we just want it to go as fast as possible.
4711 * just set gnd_next_map to current jiffies so it will process
4712 * as fast as possible.
4714 dev->gnd_next_map = jiffies;
4716 /* Retry based on GNILND_MAP_RETRY_RATE */
4717 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4720 /* only log occasionally once we've retried fast_remaps */
4721 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4722 ((dev->gnd_map_attempt % fast_remaps) == 0);
4723 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4725 /* make sure we are not off in the weeds with this tx */
4726 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4727 GNIDBG_TX(D_NETERROR, tx,
4728 "giving up on TX, too many retries", NULL);
4729 spin_unlock(&dev->gnd_lock);
4730 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4731 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4732 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4734 tx->tx_putinfo.gnpam_dst_cookie,
4735 tx->tx_msg.gnm_srcnid);
4737 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4739 tx->tx_getinfo.gngm_cookie,
4740 tx->tx_msg.gnm_srcnid);
4742 kgnilnd_tx_done(tx, -ENOMEM);
4743 GOTO(get_out_mapped, rc);
4745 GNIDBG_TX(log_retrans_level, tx,
4746 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4747 "nq_map %d mdd# %d/%d GART %ld",
4748 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4749 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4750 atomic_read(&dev->gnd_nq_map),
4751 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4752 atomic64_read(&dev->gnd_nbytes_map));
4755 /* we need to stop processing the rest of the list, so add it back in */
4756 /* set timer to wake device when we need to schedule this tx */
4757 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4758 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4759 spin_unlock(&dev->gnd_lock);
4760 GOTO(get_out_mapped, rc);
4762 spin_unlock(&dev->gnd_lock);
4768 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4773 int error_inject = 0;
4777 spin_lock(&dev->gnd_lock);
4778 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4779 dev->gnd_sched_alive = jiffies;
4783 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4784 /* break with lock held */
4788 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4789 list_del_init(&conn->gnc_schedlist);
4791 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4794 if (!list_empty(&conn->gnc_delaylist))
4795 list_del_init(&conn->gnc_delaylist);
4796 spin_unlock(&dev->gnd_lock);
4798 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4800 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4801 conn_sched != GNILND_CONN_PROCESS,
4802 "conn %p on ready list but in bad state: %d\n",
4805 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4806 conn, kgnilnd_conn_state2str(conn));
4809 set_mb(conn->gnc_last_sched_do, jiffies);
4811 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4813 /* based on intent see if we should run again. */
4814 rc = kgnilnd_schedule_process_conn(conn, intent);
4816 /* drop ref from gnd_ready_conns */
4817 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4818 down_write(&dev->gnd_conn_sem);
4819 kgnilnd_conn_decref(conn);
4820 up_write(&dev->gnd_conn_sem);
4821 } else if (rc != 1) {
4822 kgnilnd_conn_decref(conn);
4824 /* clear this so that scheduler thread doesn't spin */
4826 /* break with lock held... */
4827 spin_lock(&dev->gnd_lock);
4831 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4832 down_write(&dev->gnd_conn_sem);
4834 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4835 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4836 /* If there are tx's currently in use in another
4837 * thread we dont want to complete the close
4838 * yet. Cycle this conn back through
4840 kgnilnd_schedule_conn(conn);
4842 kgnilnd_complete_closed_conn(conn);
4844 up_write(&dev->gnd_conn_sem);
4845 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4846 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4847 /* serialize SMSG CQs with ep_bind and smsg_release */
4848 down_write(&dev->gnd_conn_sem);
4849 kgnilnd_destroy_conn_ep(conn);
4850 up_write(&dev->gnd_conn_sem);
4851 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4852 /* if we need to do some CLOSE sending, etc done here do it */
4853 down_write(&dev->gnd_conn_sem);
4854 kgnilnd_send_conn_close(conn);
4855 kgnilnd_check_fma_rx(conn);
4856 up_write(&dev->gnd_conn_sem);
4857 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4858 /* start moving traffic if the old conns are cleared out */
4859 down_read(&dev->gnd_conn_sem);
4860 kgnilnd_check_fma_rx(conn);
4861 kgnilnd_process_fmaq(conn);
4862 up_read(&dev->gnd_conn_sem);
4865 rc = kgnilnd_schedule_process_conn(conn, 0);
4867 /* drop ref from gnd_ready_conns */
4868 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4869 down_write(&dev->gnd_conn_sem);
4870 kgnilnd_conn_decref(conn);
4871 up_write(&dev->gnd_conn_sem);
4872 } else if (rc != 1) {
4873 kgnilnd_conn_decref(conn);
4876 /* check list again with lock held */
4877 spin_lock(&dev->gnd_lock);
4880 /* If we are short circuiting due to timing we want to be scheduled
4881 * as soon as possible.
4883 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4886 spin_unlock(&dev->gnd_lock);
4892 kgnilnd_scheduler(void *arg)
4894 int threadno = (long)arg;
4897 unsigned long deadline = 0;
4900 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
4902 /* all gnilnd threads need to run fairly urgently */
4903 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
4904 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
4905 while (!kgnilnd_data.kgn_shutdown) {
4907 /* Safe: kgn_shutdown only set when quiescent */
4909 /* to quiesce or to not quiesce, that is the question */
4911 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4912 KGNILND_SPIN_QUIESCE;
4915 /* tracking for when thread goes AWOL */
4916 dev->gnd_sched_alive = jiffies;
4918 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
4919 (*kgnilnd_tunables.kgn_sched_timeout + 1));
4920 /* let folks know we are up and kicking
4921 * - they can use this for latency savings, etc
4922 * - only change if IRQ, if IDLE leave alone as that
4923 * schedule_device calls to put us back to IRQ */
4924 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
4926 down_read(&dev->gnd_conn_sem);
4927 /* always check these - they are super low cost */
4928 found_work += kgnilnd_check_fma_send_cq(dev);
4929 found_work += kgnilnd_check_fma_rcv_cq(dev);
4931 /* rdma CQ doesn't care about eps */
4932 found_work += kgnilnd_check_rdma_cq(dev);
4934 /* move some RDMA ? */
4935 found_work += kgnilnd_process_rdmaq(dev);
4937 /* map some pending RDMA requests ? */
4938 found_work += kgnilnd_process_mapped_tx(dev);
4940 /* the EP for a conn is not destroyed until all the references
4941 * to it are gone, so these checks should be safe
4942 * even if run in parallel with the CQ checking functions
4943 * _AND_ a thread that processes the CLOSED->DONE
4947 up_read(&dev->gnd_conn_sem);
4949 /* process all conns ready now */
4950 found_work += kgnilnd_process_conns(dev, deadline);
4952 /* do an eager check to avoid the IRQ disabling in
4953 * prepare_to_wait and friends */
4956 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
4957 time_before(jiffies, deadline)) {
4959 if ((busy_loops % 10) == 0) {
4960 /* tickle heartbeat and watchdog to ensure our
4961 * piggishness doesn't turn into heartbeat failure */
4962 touch_nmi_watchdog();
4968 /* if we got here, found_work was zero or busy_loops means we
4969 * need to take a break. We'll clear gnd_ready but we'll check
4970 * one last time if there is an IRQ that needs processing */
4972 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
4974 /* the first time this will go LOOP -> IDLE and let us do one final check
4975 * during which we might get an IRQ, then IDLE->IDLE and schedule()
4976 * - this might allow other threads to block us for a bit if they
4977 * try to get the mutex, but that is good as we'd need to wake
4978 * up soon to handle the CQ or other processing anyways */
4980 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
4982 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
4983 time_after_eq(jiffies, deadline)) {
4985 "yeilding: found_work %d busy_loops %d\n",
4986 found_work, busy_loops);
4988 /* use yield if we are bailing due to busy_loops
4989 * - this will ensure we wake up soonish. This closes
4990 * a race with kgnilnd_device_callback - where it'd
4991 * not call wake_up() because gnd_ready == 1, but then
4992 * we come down and schedule() because of busy_loops.
4993 * We'd not be woken up until something poked our waitq
4994 * again. yield() ensures we wake up without another
4995 * waitq poke in that case */
4996 atomic_inc(&dev->gnd_n_yield);
4997 kgnilnd_data.kgn_last_condresched = jiffies;
4999 CDEBUG(D_INFO, "awake after yeild\n");
5000 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5001 } else if (found_work == GNILND_DEV_IDLE) {
5002 /* busy_loops is low and there is nothing to do,
5003 * go to sleep and wait for a waitq poke */
5005 "scheduling: found_work %d busy_loops %d\n",
5006 found_work, busy_loops);
5007 atomic_inc(&dev->gnd_n_schedule);
5008 kgnilnd_data.kgn_last_scheduled = jiffies;
5010 CDEBUG(D_INFO, "awake after schedule\n");
5011 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5013 finish_wait(&dev->gnd_waitq, &wait);
5016 kgnilnd_thread_fini();