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(unsigned long arg)
92 kgn_device_t *dev = (kgn_device_t *) arg;
94 kgnilnd_schedule_device(dev);
98 kgnilnd_device_callback(__u32 devid, __u64 arg)
101 int index = (int) arg;
103 if (index >= kgnilnd_data.kgn_ndevs) {
104 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
105 * interrupt context. */
106 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
111 dev = &kgnilnd_data.kgn_devices[index];
112 /* just basic sanity */
113 if (dev->gnd_id == devid) {
114 kgnilnd_schedule_device(dev);
116 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
121 /* sched_intent values:
122 * < 0 : do not reschedule under any circumstances
123 * == 0: reschedule if someone marked him WANTS_SCHED
124 * > 0 : force a reschedule */
125 /* Return code 0 means it did not schedule the conn, 1
126 * means it successfully scheduled the conn.
130 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
134 /* move back to IDLE but save previous state.
135 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
136 * let the xchg there handle any racing callers to get it
137 * onto gnd_ready_conns */
139 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
140 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
141 conn_sched == GNILND_CONN_PROCESS,
142 "conn %p after process in bad state: %d\n",
145 if (sched_intent >= 0) {
146 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
147 return kgnilnd_schedule_conn_refheld(conn, 1);
153 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
157 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
159 kgn_device_t *dev = conn->gnc_device;
163 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
164 /* we only care about the last person who marked want_sched since they
165 * are most likely the culprit
167 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
168 conn->gnc_sched_line = line;
169 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
170 * the chance to put it onto gnd_ready_conns.
171 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
172 * the conn in process_conns will take care of moving it back to
173 * SCHED when it is done processing */
175 if (sched == GNILND_CONN_IDLE) {
176 /* if the conn is already scheduled, we've already requested
177 * the scheduler thread wakeup */
179 /* Add a reference to the conn if we are not holding a reference
180 * already from the exisiting scheduler. We now use the same
181 * reference if we need to reschedule a conn while in a scheduler
184 kgnilnd_conn_addref(conn);
186 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
189 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
191 spin_lock(&dev->gnd_lock);
192 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
194 spin_unlock(&dev->gnd_lock);
195 set_mb(conn->gnc_last_sched_ask, jiffies);
198 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
202 /* make sure thread(s) going to process conns - but let it make
203 * separate decision from conn schedule */
205 kgnilnd_schedule_device(dev);
210 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
212 kgn_device_t *dev = conn->gnc_device;
214 spin_lock(&dev->gnd_lock);
215 if (list_empty(&conn->gnc_delaylist)) {
216 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
219 spin_unlock(&dev->gnd_lock);
221 kgnilnd_schedule_device(dev);
226 kgnilnd_schedule_dgram(kgn_device_t *dev)
230 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
231 if (wake != GNILND_DGRAM_SCHED) {
232 wake_up(&dev->gnd_dgram_waitq);
234 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
239 kgnilnd_free_tx(kgn_tx_t *tx)
241 /* taken from kgnilnd_tx_add_state_locked */
243 LASSERTF((tx->tx_list_p == NULL &&
244 tx->tx_list_state == GNILND_TX_ALLOCD) &&
245 list_empty(&tx->tx_list),
246 "tx %p with bad state %s (list_p %p) tx_list %s\n",
247 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
248 list_empty(&tx->tx_list) ? "empty" : "not empty");
250 atomic_dec(&kgnilnd_data.kgn_ntx);
252 /* we only allocate this if we need to */
253 if (tx->tx_phys != NULL) {
254 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
255 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
256 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
259 /* Only free the buffer if we used it */
260 if (tx->tx_buffer_copy != NULL) {
261 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
262 tx->tx_buffer_copy = NULL;
263 CDEBUG(D_MALLOC, "vfreed buffer2\n");
266 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
268 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
269 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
273 kgnilnd_alloc_tx (void)
277 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
280 tx = kmem_cache_zalloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
282 CERROR("failed to allocate tx\n");
285 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
288 /* setup everything here to minimize time under the lock */
289 tx->tx_buftype = GNILND_BUF_NONE;
290 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
291 INIT_LIST_HEAD(&tx->tx_list);
292 INIT_LIST_HEAD(&tx->tx_map_list);
293 tx->tx_list_state = GNILND_TX_ALLOCD;
295 atomic_inc(&kgnilnd_data.kgn_ntx);
300 /* csum_fold needs to be run on the return value before shipping over the wire */
301 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
303 /* we don't use offset as every one is passing a buffer reference that already
304 * includes the offset into the base address.
307 kgnilnd_cksum(void *ptr, size_t nob)
311 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
313 /* don't use magic 'no checksum' value */
317 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
324 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
325 unsigned int offset, unsigned int nob, int dump_blob)
331 unsigned int fraglen;
337 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
338 kiov, nkiov, offset, nob, dump_blob);
340 /* if loops changes, please change kgnilnd_setup_phys_buffer */
342 while (offset >= kiov->bv_len) {
343 offset -= kiov->bv_len;
349 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
350 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
352 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
353 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
355 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
356 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
358 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
359 odd, kiov[0].bv_len, offset, nob);
361 for (i = 0; i < nkiov; i++) {
362 pages[i] = kiov[i].bv_page;
365 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
367 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
369 /* return zero to avoid killing tx - we'll just get warning on console
370 * when remote end sees zero checksum */
373 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
375 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
380 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
381 (void *)addr + kiov[0].bv_offset +
384 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
385 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
389 fraglen = min(kiov->bv_len - offset, nob);
391 /* make dang sure we don't send a bogus checksum if somehow we get
392 * an odd length fragment on anything but the last entry in a kiov -
393 * we know from kgnilnd_setup_rdma_buffer that we can't have non
394 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
395 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
396 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
397 fraglen, nkiov, nob, kiov->bv_len,
400 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
402 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
405 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
406 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
407 addr, fraglen, offset);
412 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
414 kunmap(kiov->bv_page);
421 /* iov must not run out before end of data */
422 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
427 retsum = csum_fold(cksum);
429 /* don't use magic 'no checksum' value */
433 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
439 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
441 msg->gnm_magic = GNILND_MSG_MAGIC;
442 msg->gnm_version = GNILND_MSG_VERSION;
443 msg->gnm_type = type;
444 msg->gnm_payload_len = 0;
445 msg->gnm_srcnid = source;
446 /* gnm_connstamp gets set when FMA is sent */
447 /* gnm_srcnid is set on creation via function argument
448 * The right interface/net and nid is passed in when the message
454 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
456 kgn_tx_t *tx = kgnilnd_alloc_tx();
459 kgnilnd_init_msg(&tx->tx_msg, type, source);
461 CERROR("couldn't allocate new tx type %s!\n",
462 kgnilnd_msgtype2str(type));
469 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
475 case GNILND_MSG_GET_REQ:
476 case GNILND_MSG_GET_DONE:
477 nak_type = GNILND_MSG_GET_NAK;
479 case GNILND_MSG_PUT_REQ:
480 case GNILND_MSG_PUT_ACK:
481 case GNILND_MSG_PUT_DONE:
482 nak_type = GNILND_MSG_PUT_NAK;
484 case GNILND_MSG_PUT_REQ_REV:
485 case GNILND_MSG_PUT_DONE_REV:
486 nak_type = GNILND_MSG_PUT_NAK_REV;
488 case GNILND_MSG_GET_REQ_REV:
489 case GNILND_MSG_GET_ACK_REV:
490 case GNILND_MSG_GET_DONE_REV:
491 nak_type = GNILND_MSG_GET_NAK_REV;
494 CERROR("invalid msg type %s (%d)\n",
495 kgnilnd_msgtype2str(rx_type), rx_type);
498 /* only allow NAK on error and truncate to zero */
499 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
500 error, conn, cookie);
502 tx = kgnilnd_new_tx_msg(nak_type, source);
504 CNETERR("can't get TX to NAK RDMA to %s\n",
505 libcfs_nid2str(conn->gnc_peer->gnp_nid));
509 tx->tx_msg.gnm_u.completion.gncm_retval = error;
510 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
511 kgnilnd_queue_tx(conn, tx);
515 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
516 struct bio_vec *kiov,
517 unsigned int offset, unsigned int nob)
519 kgn_msg_t *msg = &tx->tx_msg;
522 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
523 * gni_smsg_send to send that as the payload */
525 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
528 tx->tx_buffer = NULL;
531 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
532 niov = round_up(nob + offset + kiov->bv_offset,
536 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
537 "bad niov %d msg %p kiov %p offset %d nob%d\n",
538 niov, msg, kiov, offset, nob);
540 while (offset >= kiov->bv_len) {
541 offset -= kiov->bv_len;
546 for (i = 0; i < niov; i++) {
547 /* We can't have a bv_offset on anything but the first
548 * entry, otherwise we'll have a hole at the end of the
549 * mapping as we only map whole pages.
550 * Also, if we have a bv_len < PAGE_SIZE but we need to
551 * map more than bv_len, we will also have a whole at
552 * the end of that page which isn't allowed
554 if ((kiov[i].bv_offset != 0 && i > 0) ||
555 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
557 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
558 i, offset, nob, kiov->bv_offset,
562 tx->tx_imm_pages[i] = kiov[i].bv_page;
565 /* hijack tx_phys for the later unmap */
567 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
569 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
570 kiov[0].bv_offset + offset;
571 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
572 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
573 nob, kiov, tx->tx_buffer);
575 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
576 if (tx->tx_phys == NULL) {
577 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
581 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
582 /* make sure we take into account the kiov offset as the
583 * start of the buffer
585 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
588 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
589 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
591 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
596 /* checksum payload early - it shouldn't be changing after lnd_send */
597 if (*kgnilnd_tunables.kgn_checksum >= 2) {
598 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
599 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
600 msg->gnm_payload_cksum += 0xe00e;
602 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
603 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
607 msg->gnm_payload_cksum = 0;
614 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
615 unsigned int offset, unsigned int nob)
617 gni_mem_segment_t *phys;
619 unsigned int fraglen;
621 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
625 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
627 /* only allocate this if we are going to use it */
628 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
630 if (tx->tx_phys == NULL) {
631 CERROR("failed to allocate tx_phys\n");
636 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
637 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
639 /* if loops changes, please change kgnilnd_cksum_kiov
640 * and kgnilnd_setup_immediate_buffer */
642 while (offset >= kiov->bv_len) {
643 offset -= kiov->bv_len;
649 /* at this point, kiov points to the first page that we'll actually map
650 * now that we've seeked into the koiv for offset and dropped any
651 * leading pages that fall entirely within the offset */
652 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
655 /* bv_offset is start of 'valid' buffer, so index offset past that */
656 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
659 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
660 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
663 fraglen = min(kiov->bv_len - offset, nob);
665 /* We can't have a bv_offset on anything but the first entry,
666 * otherwise we'll have a hole at the end of the mapping as we
667 * only map whole pages. Only the first page is allowed to
668 * have an offset - we'll add that into tx->tx_buffer and that
669 * will get used when we map in the segments (see
670 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
671 * but we need to map more than bv_len, we will also have a
672 * whole at the end of that page which isn't allowed
674 if ((phys != tx->tx_phys) &&
675 ((kiov->bv_offset != 0) ||
676 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
677 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
678 (int)(phys - tx->tx_phys),
679 offset, nob, kiov->bv_offset, kiov->bv_len);
684 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
685 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
690 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
696 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
697 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
700 phys->address = page_to_phys(kiov->bv_page);
707 /* iov must not run out before end of data */
708 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
712 tx->tx_phys_npages = phys - tx->tx_phys;
717 if (tx->tx_phys != NULL) {
718 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
719 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
720 sizeof(*tx->tx_phys), tx->tx_phys);
727 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
728 struct bio_vec *kiov,
729 unsigned int offset, unsigned int nob)
731 return kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
734 /* kgnilnd_parse_lnet_rdma()
735 * lntmsg - message passed in from lnet.
736 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
737 * nob - actual number of bytes to in this message.
738 * put_len - It is possible for PUTs to have a different length than the
739 * length stored in lntmsg->msg_len since LNET can adjust this
740 * length based on it's buffer size and offset.
741 * lnet_try_match_md() sets the mlength that we use to do the RDMA
745 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
746 unsigned int *offset, unsigned int *nob,
747 struct bio_vec **kiov, int put_len)
749 /* GETs are weird, see kgnilnd_send */
750 if (lntmsg->msg_type == LNET_MSG_GET) {
751 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
754 *kiov = lntmsg->msg_md->md_kiov;
756 *niov = lntmsg->msg_md->md_niov;
757 *nob = lntmsg->msg_md->md_length;
760 *kiov = lntmsg->msg_kiov;
761 *niov = lntmsg->msg_niov;
763 *offset = lntmsg->msg_offset;
768 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
770 unsigned int niov, offset, nob;
771 struct bio_vec *kiov;
772 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
773 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
775 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
776 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
777 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
778 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
779 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
780 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
781 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
783 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
784 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
785 tx->tx_msg.gnm_payload_cksum = 0;
788 if (*kgnilnd_tunables.kgn_checksum < 3) {
789 tx->tx_msg.gnm_payload_cksum = 0;
793 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
795 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
799 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
801 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
803 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
807 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
808 tx->tx_msg.gnm_payload_cksum += 0xd00d;
812 /* kgnilnd_verify_rdma_cksum()
813 * tx - PUT_DONE/GET_DONE matched tx.
814 * rx_cksum - received checksum to compare against.
815 * put_len - see kgnilnd_parse_lnet_rdma comments.
818 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
822 unsigned int niov, offset, nob;
823 struct bio_vec *kiov;
824 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
825 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
827 /* we can only match certain requests */
828 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
829 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
830 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
831 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
832 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
833 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
834 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
836 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
837 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
842 if (*kgnilnd_tunables.kgn_checksum >= 3) {
843 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
844 "no RDMA payload checksum when enabled");
849 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
851 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
854 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
856 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
859 if (cksum != rx_cksum) {
860 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
861 "Bad RDMA payload checksum (%x expected %x); "
862 "kiov 0x%p niov %d nob %u offset %u",
863 cksum, rx_cksum, kiov, niov, nob, offset);
864 switch (dump_on_err) {
867 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
869 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
872 /* fall through to dump log */
874 libcfs_debug_dumplog();
880 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
886 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
890 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
891 "already mapped!", NULL);
893 spin_lock(&dev->gnd_map_lock);
894 switch (tx->tx_buftype) {
896 GNIDBG_TX(D_EMERG, tx,
897 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
898 spin_unlock(&dev->gnd_map_lock);
902 case GNILND_BUF_PHYS_MAPPED:
903 bytes = tx->tx_phys_npages * PAGE_SIZE;
904 dev->gnd_map_nphys++;
905 dev->gnd_map_physnop += tx->tx_phys_npages;
908 case GNILND_BUF_VIRT_MAPPED:
910 dev->gnd_map_nvirt++;
911 dev->gnd_map_virtnob += tx->tx_nob;
915 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
916 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
917 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
918 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
919 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
922 atomic_inc(&dev->gnd_n_mdd);
923 atomic64_add(bytes, &dev->gnd_nbytes_map);
925 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
928 /* we only get here in the valid cases */
929 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
930 dev->gnd_map_version++;
931 spin_unlock(&dev->gnd_map_lock);
935 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
939 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
940 "not mapped!", NULL);
941 spin_lock(&dev->gnd_map_lock);
943 switch (tx->tx_buftype) {
945 GNIDBG_TX(D_EMERG, tx,
946 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
947 spin_unlock(&dev->gnd_map_lock);
951 case GNILND_BUF_PHYS_UNMAPPED:
952 bytes = tx->tx_phys_npages * PAGE_SIZE;
953 dev->gnd_map_nphys--;
954 dev->gnd_map_physnop -= tx->tx_phys_npages;
957 case GNILND_BUF_VIRT_UNMAPPED:
959 dev->gnd_map_nvirt--;
960 dev->gnd_map_virtnob -= tx->tx_nob;
964 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
965 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
966 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
967 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
968 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
969 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
970 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
973 atomic_dec(&dev->gnd_n_mdd);
974 atomic64_sub(bytes, &dev->gnd_nbytes_map);
976 /* we only get here in the valid cases */
977 list_del_init(&tx->tx_map_list);
978 dev->gnd_map_version++;
979 spin_unlock(&dev->gnd_map_lock);
983 kgnilnd_map_buffer(kgn_tx_t *tx)
985 kgn_conn_t *conn = tx->tx_conn;
986 kgn_device_t *dev = conn->gnc_device;
987 __u32 flags = GNI_MEM_READWRITE;
990 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
991 * be called concurrently as there are internal locks that protect
992 * any data structures or HW resources. We just need to ensure
993 * that our concurrency doesn't result in the kgn_device_t
994 * getting nuked while we are in here */
996 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
997 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
999 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1002 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1003 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1006 switch (tx->tx_buftype) {
1010 case GNILND_BUF_NONE:
1011 case GNILND_BUF_IMMEDIATE:
1012 case GNILND_BUF_IMMEDIATE_KIOV:
1013 case GNILND_BUF_PHYS_MAPPED:
1014 case GNILND_BUF_VIRT_MAPPED:
1017 case GNILND_BUF_PHYS_UNMAPPED:
1018 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1019 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1020 tx->tx_phys, tx->tx_phys_npages, NULL,
1021 GNI_MEM_PHYS_SEGMENTS | flags,
1023 /* could race with other uses of the map counts, but this is ok
1024 * - this needs to turn into a non-fatal error soon to allow
1025 * GART resource, etc starvation handling */
1026 if (rrc != GNI_RC_SUCCESS) {
1027 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1028 "phys %u pp %u, virt %u nob %llu",
1029 tx->tx_phys_npages, dev->gnd_id,
1030 dev->gnd_map_nphys, dev->gnd_map_physnop,
1031 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1032 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1035 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1036 kgnilnd_mem_add_map_list(dev, tx);
1039 case GNILND_BUF_VIRT_UNMAPPED:
1040 rrc = kgnilnd_mem_register(dev->gnd_handle,
1041 (__u64)tx->tx_buffer, tx->tx_nob,
1042 NULL, flags, &tx->tx_map_key);
1043 if (rrc != GNI_RC_SUCCESS) {
1044 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1045 "phys %u pp %u, virt %u nob %llu",
1046 tx->tx_nob, dev->gnd_id,
1047 dev->gnd_map_nphys, dev->gnd_map_physnop,
1048 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1049 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1052 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1053 kgnilnd_mem_add_map_list(dev, tx);
1054 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1055 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1056 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1057 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1058 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1066 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1068 kgn_conn_t *conn = tx->tx_conn;
1069 kgn_mdd_purgatory_t *gmp;
1071 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1072 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1073 " asserting to avoid data corruption\n");
1074 if (tx->tx_buffer_copy)
1075 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1077 gmp->gmp_map_key = tx->tx_map_key;
1079 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1081 /* ensure that we don't have a blank purgatory - indicating the
1082 * conn is not already on purgatory lists - we'd never recover these
1083 * MDD if that were the case */
1084 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1085 "conn 0x%p->%s with NULL purgatory",
1086 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1088 /* link 'er up! - only place we really need to lock for
1089 * concurrent access */
1090 spin_lock(&conn->gnc_list_lock);
1091 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1092 spin_unlock(&conn->gnc_list_lock);
1096 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1100 int hold_timeout = 0;
1102 /* code below relies on +1 relationship ... */
1103 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1104 (GNILND_BUF_PHYS_UNMAPPED + 1));
1105 BUILD_BUG_ON(GNILND_BUF_VIRT_MAPPED !=
1106 (GNILND_BUF_VIRT_UNMAPPED + 1));
1108 switch (tx->tx_buftype) {
1112 case GNILND_BUF_NONE:
1113 case GNILND_BUF_IMMEDIATE:
1114 case GNILND_BUF_PHYS_UNMAPPED:
1115 case GNILND_BUF_VIRT_UNMAPPED:
1117 case GNILND_BUF_IMMEDIATE_KIOV:
1118 if (tx->tx_phys != NULL) {
1119 vunmap(tx->tx_phys);
1120 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1121 kunmap(tx->tx_imm_pages[0]);
1123 /* clear to prevent kgnilnd_free_tx from thinking
1124 * this is a RDMA descriptor */
1128 case GNILND_BUF_PHYS_MAPPED:
1129 case GNILND_BUF_VIRT_MAPPED:
1130 LASSERT(tx->tx_conn != NULL);
1132 dev = tx->tx_conn->gnc_device;
1134 /* only want to hold if we are closing conn without
1135 * verified peer notification - the theory is that
1136 * a TX error can be communicated in all other cases */
1137 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1138 error != -GNILND_NOPURG &&
1139 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1140 kgnilnd_add_purgatory_tx(tx);
1142 /* The timeout we give to kgni is a deadman stop only.
1143 * we are setting high to ensure we don't have the kgni timer
1144 * fire before ours fires _and_ is handled */
1145 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1147 GNIDBG_TX(D_NET, tx,
1148 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1149 tx->tx_conn->gnc_device, hold_timeout,
1150 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1152 if (tx->tx_buffer_copy != NULL) {
1153 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1154 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1155 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1156 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1158 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1159 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1163 kgnilnd_mem_del_map_list(dev, tx);
1169 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1171 struct lnet_msg *lntmsg0, *lntmsg1;
1172 int status0, status1;
1173 struct lnet_ni *ni = NULL;
1174 kgn_conn_t *conn = tx->tx_conn;
1176 LASSERT(!in_interrupt());
1178 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1179 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1182 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1183 !kgnilnd_conn_clean_errno(completion)) {
1184 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1185 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1186 completion, tx, conn ?
1187 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1188 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1189 kgnilnd_tx_state2str(tx->tx_list_state),
1190 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1193 /* The error codes determine if we hold onto the MDD */
1194 kgnilnd_unmap_buffer(tx, completion);
1196 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1197 * we play nice with the error codes to avoid delivering a failed
1198 * REQUEST and then a REPLY event as well */
1200 /* return -EIO to lnet - it is the magic value for failed sends */
1201 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1203 status1 = completion;
1205 status0 = status1 = completion;
1208 tx->tx_buftype = GNILND_BUF_NONE;
1209 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1211 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1212 * set NULL when we are a tx without a conn */
1214 ni = conn->gnc_peer->gnp_net->gnn_ni;
1216 spin_lock(&conn->gnc_tx_lock);
1218 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1219 (volatile unsigned long *)&conn->gnc_tx_bits),
1220 "conn %p tx %p bit %d already cleared\n",
1221 conn, tx, tx->tx_id.txe_idx);
1223 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1224 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1226 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1227 spin_unlock(&conn->gnc_tx_lock);
1230 kgnilnd_free_tx(tx);
1232 /* finalize AFTER freeing lnet msgs */
1234 /* warning - we should hold no locks here - calling lnet_finalize
1235 * could free up lnet credits, resulting in a call chain back into
1236 * the LND via kgnilnd_send and friends */
1238 lnet_finalize(lntmsg0, status0);
1240 if (lntmsg1 != NULL) {
1241 lnet_finalize(lntmsg1, status1);
1246 kgnilnd_txlist_done(struct list_head *txlist, int error)
1249 int err_printed = 0;
1251 if (list_empty(txlist))
1254 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1255 /* only print the first error */
1257 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1258 list_del_init(&tx->tx_list);
1259 kgnilnd_tx_done(tx, error);
1264 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1268 spin_lock(&conn->gnc_tx_lock);
1270 /* ID zero is NOT ALLOWED!!! */
1273 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1274 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1275 if (id == GNILND_MAX_MSG_ID) {
1276 if (conn->gnc_next_tx != 1) {
1277 /* we only searched from next_tx to end and didn't find
1278 * one, so search again from start */
1279 conn->gnc_next_tx = 1;
1282 /* couldn't find one! */
1283 spin_unlock(&conn->gnc_tx_lock);
1287 /* bump next_tx to prevent immediate reuse */
1288 conn->gnc_next_tx = id + 1;
1290 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1291 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1292 "tx 0x%p already at id %d\n",
1293 conn->gnc_tx_ref_table[id], id);
1295 /* delay these until we have a valid ID - prevents bad clear of the bit
1296 * in kgnilnd_tx_done */
1298 tx->tx_id.txe_cqid = conn->gnc_cqid;
1300 tx->tx_id.txe_idx = id;
1301 conn->gnc_tx_ref_table[id] = tx;
1303 /* Using jiffies to help differentiate against TX reuse - with
1304 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1305 * if we are sending to the same node faster than 256000/sec.
1306 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1308 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1310 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1312 spin_unlock(&conn->gnc_tx_lock);
1317 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1321 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1323 /* we don't care about TX timeout - it could be that the network is slower
1324 * or throttled. We'll keep retranmitting - so if the network is so slow
1325 * that we fill up our mailbox, we'll keep trying to resend that msg
1326 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1327 * that he hasn't send us any traffic in return */
1329 /* some reasonable throttling of the debug message */
1331 unsigned long now = jiffies;
1332 /* XXX Nic: Mystical TX debug here... */
1333 /* We expect retransmissions so only log when D_NET is enabled */
1334 GNIDBG_SMSG_CREDS(D_NET, conn);
1335 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1336 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1337 " last_msg %uus/%uus last_cq %uus/%uus",
1338 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1339 tx->tx_id, tx->tx_retrans,
1340 jiffies_to_usecs(now - tx->tx_cred_wait),
1341 jiffies_to_usecs(now - conn->gnc_last_tx),
1342 jiffies_to_usecs(now - conn->gnc_last_rx),
1343 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1344 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1348 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1349 * to avoid bad ordering with state_lock */
1352 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1353 spinlock_t *state_lock, kgn_tx_list_state_t state)
1355 kgn_conn_t *conn = tx->tx_conn;
1356 kgn_msg_t *msg = &tx->tx_msg;
1358 unsigned long newest_last_rx, timeout;
1361 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1362 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1364 "msg 0x%p type %d wrong payload size %d\n",
1365 msg, msg->gnm_type, immediatenob);
1367 /* make sure we catch all the cases where we'd send on a dirty old mbox
1368 * but allow case for sending CLOSE. Since this check is within the CQ
1369 * mutex barrier and the close message is only sent through
1370 * kgnilnd_send_conn_close the last message out the door will be the
1373 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1374 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1375 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1376 /* Return -ETIME, we are closing the connection already so we dont want to
1377 * have this tx hit the wire. The tx will be killed by the calling function.
1378 * Once the EP is marked dirty the close message will be the last
1379 * thing to hit the wire */
1384 timeout = cfs_time_seconds(conn->gnc_timeout);
1386 newest_last_rx = GNILND_LASTRX(conn);
1388 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1389 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1392 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1393 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1394 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1395 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1396 cfs_duration_sec(now - newest_last_rx),
1397 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1398 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1399 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1403 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1404 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1405 * creating the message this allows the message to contain the correct LNET NID/NET needed
1406 * instead of the one that the peer/conn uses for sending the data.
1408 msg->gnm_connstamp = conn->gnc_my_connstamp;
1409 msg->gnm_payload_len = immediatenob;
1410 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1412 /* always init here - kgn_checksum is a /sys module tunable
1413 * and can be flipped at any point, even between msg init and sending */
1415 if (*kgnilnd_tunables.kgn_checksum) {
1416 /* We must set here and not in kgnilnd_init_msg,
1417 * we could resend this msg many times
1418 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1419 * through init_msg again */
1420 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1421 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1422 msg->gnm_cksum += 0xf00f;
1426 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1427 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1428 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1429 tx->tx_id.txe_idx, immediate, immediatenob);
1431 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1432 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1434 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1435 msg, sizeof(*msg), immediate,
1437 tx->tx_id.txe_smsg_id);
1441 case GNI_RC_SUCCESS:
1442 atomic_inc(&conn->gnc_tx_seq);
1443 conn->gnc_last_tx = jiffies;
1444 /* no locking here as LIVE isn't a list */
1445 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1447 /* this needs to be checked under lock as it might be freed from a completion
1450 if (msg->gnm_type == GNILND_MSG_NOOP) {
1451 set_mb(conn->gnc_last_noop_sent, jiffies);
1454 /* serialize with seeing CQ events for completion on this, as well as
1456 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1457 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1459 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1460 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1461 kgnilnd_peer_alive(conn->gnc_peer);
1462 GNIDBG_SMSG_CREDS(D_NET, conn);
1465 case GNI_RC_NOT_DONE:
1466 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1467 * 1: out of mbox credits
1468 * 2: out of mbox payload credits
1469 * 3: On Aries out of dla credits
1471 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1472 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1473 /* We'll handle this error inline - makes the calling logic much more
1476 /* If no lock, caller doesn't want us to retry */
1477 if (state_lock == NULL) {
1481 /* I need kgni credits to send this. Replace tx at the head of the
1482 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1483 * and bump retrans counts since we are requeueing the tx.
1487 conn->gnc_tx_retrans++;
1489 kgnilnd_tx_log_retrans(conn, tx);
1490 /* add to head of list for the state and retries */
1491 spin_lock(state_lock);
1492 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1493 spin_unlock(state_lock);
1495 /* We only reschedule for a certain number of retries, then
1496 * we will wait for the CQ events indicating a release of SMSG
1498 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1499 kgnilnd_schedule_conn(conn);
1502 /* CQ event coming in signifies either TX completed or
1503 * RX receive. Either of these *could* free up credits
1504 * in the SMSG mbox and we should try sending again */
1505 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1506 tx->tx_conn->gnc_cqid);
1507 kgnilnd_schedule_delay_conn(conn);
1508 /* use +ve return code to let upper layers know they
1509 * should stop looping on sends */
1513 /* handle bad retcode gracefully */
1514 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1515 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1520 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1522 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1523 spinlock_t *state_lock, kgn_tx_list_state_t state)
1525 kgn_device_t *dev = tx->tx_conn->gnc_device;
1526 unsigned long timestamp;
1529 timestamp = jiffies;
1530 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1531 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1532 /* delay in jiffies - we are really concerned only with things that
1533 * result in a schedule() or really holding this off for long times .
1534 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1535 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1537 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1543 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1545 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1546 spinlock_t *state_lock, kgn_tx_list_state_t state)
1548 kgn_conn_t *conn = tx->tx_conn;
1549 kgn_device_t *dev = conn->gnc_device;
1550 unsigned long timestamp;
1553 timestamp = jiffies;
1555 /* technically we are doing bad things with the read_lock on the peer_conn
1556 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1557 * for the mutex. I bet lockdep is gonna flag this one though... */
1559 /* there are a few cases where we don't want the immediate send - like
1560 * when we are in the scheduler thread and it'd harm the latency of
1561 * getting messages up to LNet */
1563 /* rmb for gnd_ready */
1565 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1567 atomic_inc(&conn->gnc_device->gnd_fast_block);
1568 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1569 /* dont hit HW during quiesce */
1571 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1572 /* dont hit HW if stale EPs and conns left to close */
1575 atomic_inc(&conn->gnc_device->gnd_fast_try);
1576 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1577 &conn->gnc_smsg_mutex);
1582 /* we got the mutex and weren't blocked */
1584 /* delay in jiffies - we are really concerned only with things that
1585 * result in a schedule() or really holding this off for long times .
1586 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1587 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1589 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1590 tx->tx_qtime = jiffies;
1591 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1592 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1593 /* _nolock unlocks the mutex for us */
1599 /* lets us know if we can push this RDMA through now */
1601 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1605 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1607 if (bytes_left < 0) {
1608 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1609 atomic_inc(&dev->gnd_rdmaq_nstalls);
1612 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1613 dev->gnd_rdmaq_deadline);
1614 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1615 /* we never del this timer - at worst it schedules us.. */
1622 /* this adds a TX to the queue pending throttling authorization before
1623 * we allow our remote peer to launch a PUT at us */
1625 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1629 /* we cannot go into send_mapped_tx from here as we are holding locks
1630 * and mem registration might end up allocating memory in kgni.
1631 * That said, we'll push this as far as we can into the queue process */
1632 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1635 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1636 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1637 /* lets us know how delayed RDMA is */
1638 tx->tx_qtime = jiffies;
1639 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1641 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1642 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1643 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1644 /* lets us know how delayed mapping is */
1645 tx->tx_qtime = jiffies;
1646 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1649 /* make sure we wake up sched to run this */
1650 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1653 /* push TX through state machine */
1655 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1660 /* set the tx_id here, we delay it until we have an actual conn
1662 * in some cases, the tx_id is already set to provide for things
1663 * like RDMA completion cookies, etc */
1664 if (tx->tx_id.txe_idx == 0) {
1665 rc = kgnilnd_set_tx_id(tx, conn);
1667 kgnilnd_tx_done(tx, rc);
1672 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1673 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1675 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1677 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1678 kgnilnd_tx_done(tx, rc);
1682 switch (tx->tx_msg.gnm_type) {
1683 case GNILND_MSG_PUT_ACK:
1684 case GNILND_MSG_GET_REQ:
1685 case GNILND_MSG_PUT_REQ_REV:
1686 case GNILND_MSG_GET_ACK_REV:
1687 /* hijacking time! If this messages will authorize our peer to
1688 * send his dirty little bytes in an RDMA, we need to get permission */
1689 kgnilnd_queue_rdma(conn, tx);
1691 case GNILND_MSG_IMMEDIATE:
1692 /* try to send right now, can help reduce latency */
1693 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1696 /* it was sent, break out of switch to avoid default case of queueing */
1699 /* needs to queue to try again, so fall through to default case */
1700 case GNILND_MSG_NOOP:
1701 /* Just make sure this goes out first for this conn */
1703 /* fall through... */
1705 spin_lock(&conn->gnc_list_lock);
1706 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1707 tx->tx_qtime = jiffies;
1708 spin_unlock(&conn->gnc_list_lock);
1709 kgnilnd_schedule_conn(conn);
1714 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1717 kgn_peer_t *new_peer = NULL;
1718 kgn_conn_t *conn = NULL;
1724 /* If I get here, I've committed to send, so I complete the tx with
1725 * failure on any problems */
1727 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1728 "tx already has connection %p", tx->tx_conn);
1730 /* do all of the peer & conn searching in one swoop - this avoids
1731 * nastiness when dropping locks and needing to maintain a sane state
1732 * in the face of stack reset or something else nuking peers & conns */
1734 /* I expect to find him, so only take a read lock */
1735 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1737 peer = kgnilnd_find_peer_locked(target->nid);
1739 conn = kgnilnd_find_conn_locked(peer);
1740 /* this could be NULL during quiesce */
1742 /* Connection exists; queue message on it */
1743 kgnilnd_queue_tx(conn, tx);
1744 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1748 /* don't create a connection if the peer is marked down */
1749 if (peer->gnp_state != GNILND_PEER_UP) {
1750 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1756 /* creating peer or conn; I'll need a write lock... */
1757 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1759 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1761 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1763 /* NB - this will not block during normal operations -
1764 * the only writer of this is in the startup/shutdown path. */
1765 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1771 /* ignore previous peer entirely - we cycled the lock, so we
1772 * will create new peer and at worst drop it if peer is still
1774 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1776 up_read(&kgnilnd_data.kgn_net_rw_sem);
1780 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1781 up_read(&kgnilnd_data.kgn_net_rw_sem);
1783 /* search for peer again now that we have the lock
1784 * if we don't find it, add our new one to the list */
1785 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1787 /* don't create a connection if the peer is not up */
1788 if (peer->gnp_state != GNILND_PEER_UP) {
1789 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1794 conn = kgnilnd_find_or_create_conn_locked(peer);
1796 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1797 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1802 /* oh hey, found a conn now... magical */
1803 kgnilnd_queue_tx(conn, tx);
1805 /* no conn, must be trying to connect - so we queue for now */
1806 tx->tx_qtime = jiffies;
1807 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1809 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1812 kgnilnd_tx_done(tx, rc);
1817 kgnilnd_rdma(kgn_tx_t *tx, int type,
1818 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1820 kgn_conn_t *conn = tx->tx_conn;
1821 unsigned long timestamp;
1822 gni_post_type_t post_type;
1825 unsigned int desc_nob = nob;
1826 void *desc_buffer = tx->tx_buffer;
1827 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1828 LASSERTF(kgnilnd_tx_mapped(tx),
1829 "unmapped tx %p\n", tx);
1830 LASSERTF(conn != NULL,
1831 "NULL conn on tx %p, naughty, naughty\n", tx);
1832 LASSERTF(nob <= sink->gnrd_nob,
1833 "nob %u > sink->gnrd_nob %d (%p)\n",
1834 nob, sink->gnrd_nob, sink);
1835 LASSERTF(nob <= tx->tx_nob,
1836 "nob %d > tx(%p)->tx_nob %d\n",
1837 nob, tx, tx->tx_nob);
1840 case GNILND_MSG_GET_DONE:
1841 case GNILND_MSG_PUT_DONE:
1842 post_type = GNI_POST_RDMA_PUT;
1844 case GNILND_MSG_GET_DONE_REV:
1845 case GNILND_MSG_PUT_DONE_REV:
1846 post_type = GNI_POST_RDMA_GET;
1849 CERROR("invalid msg type %s (%d)\n",
1850 kgnilnd_msgtype2str(type), type);
1853 if (post_type == GNI_POST_RDMA_GET) {
1854 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1855 * aligned. If the local buffer is not aligned correctly using the copy buffer
1856 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1857 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1858 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1859 * making sure the length and addr are aligned and copying the data into the correct location
1860 * once the transfer has completed.
1862 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1863 (sink->gnrd_addr & 3) ||
1866 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1868 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1870 if ((nob + tx->tx_offset) & 3) {
1871 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1872 atomic_inc(&kgnilnd_data.kgn_rev_length);
1874 desc_nob = (nob + tx->tx_offset);
1877 if (tx->tx_buffer_copy == NULL) {
1878 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1879 * and require at most we allocate a few extra bytes. */
1880 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1882 if (!tx->tx_buffer_copy) {
1883 /* allocation of buffer failed nak the rdma */
1884 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1885 kgnilnd_tx_done(tx, -EFAULT);
1888 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1889 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);
1890 if (rc != GNI_RC_SUCCESS) {
1891 /* Registration Failed nak rdma and kill the tx. */
1892 kgnilnd_vfree(tx->tx_buffer_copy,
1894 tx->tx_buffer_copy = NULL;
1895 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1896 kgnilnd_tx_done(tx, -EFAULT);
1900 desc_map_key = tx->tx_buffer_copy_map_key;
1901 desc_buffer = tx->tx_buffer_copy;
1905 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1906 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1907 tx->tx_rdma_desc.type = post_type;
1908 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1909 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1910 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1911 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1912 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1913 tx->tx_rdma_desc.length = desc_nob;
1914 tx->tx_nob_rdma = nob;
1915 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1916 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1917 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1918 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1919 /* prep final completion message */
1920 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1921 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1922 /* send actual size RDMA'd in retval */
1923 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1925 kgnilnd_compute_rdma_cksum(tx, nob);
1928 kgnilnd_queue_tx(conn, tx);
1932 /* Don't lie (CLOSE == RDMA idle) */
1933 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1934 tx, conn, conn->gnc_close_sent);
1936 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1937 "0x%x cookie:%#llx",
1938 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1940 /* set CQ dedicated for RDMA */
1941 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1943 timestamp = jiffies;
1944 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1945 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1946 /* delay in jiffies - we are really concerned only with things that
1947 * result in a schedule() or really holding this off for long times .
1948 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1949 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1951 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
1953 if (rrc == GNI_RC_ERROR_RESOURCE) {
1954 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1955 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1956 kgnilnd_unmap_buffer(tx, 0);
1958 if (tx->tx_buffer_copy != NULL) {
1959 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
1960 tx->tx_buffer_copy = NULL;
1963 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1964 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
1966 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1967 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1971 spin_lock(&conn->gnc_list_lock);
1972 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
1973 tx->tx_qtime = jiffies;
1974 spin_unlock(&conn->gnc_list_lock);
1975 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1976 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1978 /* XXX Nic: is this a place we should handle more errors for
1979 * robustness sake */
1980 LASSERT(rrc == GNI_RC_SUCCESS);
1985 kgnilnd_alloc_rx(void)
1989 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
1991 CERROR("failed to allocate rx\n");
1994 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
1997 /* no memset to zero, we'll always fill all members */
2001 /* release is to just free connection resources
2002 * we use this for the eager path after copying */
2004 kgnilnd_release_msg(kgn_conn_t *conn)
2007 unsigned long timestamp;
2009 CDEBUG(D_NET, "consuming %p\n", conn);
2011 timestamp = jiffies;
2012 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2013 /* delay in jiffies - we are really concerned only with things that
2014 * result in a schedule() or really holding this off for long times .
2015 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2016 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2018 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2019 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2021 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2022 GNIDBG_SMSG_CREDS(D_NET, conn);
2024 kgnilnd_schedule_conn(conn);
2028 kgnilnd_consume_rx(kgn_rx_t *rx)
2030 kgn_conn_t *conn = rx->grx_conn;
2031 kgn_msg_t *rxmsg = rx->grx_msg;
2033 /* if we are eager, free the cache alloc'd msg */
2034 if (unlikely(rx->grx_eager)) {
2035 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2036 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2038 /* release ref from eager_recv */
2039 kgnilnd_conn_decref(conn);
2041 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2042 kgnilnd_release_msg(conn);
2045 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2046 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2051 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2053 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2054 int type = lntmsg->msg_type;
2055 struct lnet_process_id target = lntmsg->msg_target;
2056 int target_is_router = lntmsg->msg_target_is_router;
2057 int routing = lntmsg->msg_routing;
2058 unsigned int niov = lntmsg->msg_niov;
2059 struct bio_vec *kiov = lntmsg->msg_kiov;
2060 unsigned int offset = lntmsg->msg_offset;
2061 unsigned int nob = lntmsg->msg_len;
2062 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2063 kgn_net_t *net = ni->ni_data;
2066 /* '1' for consistency with code that checks !mpflag to restore */
2067 unsigned int mpflag = 1;
2068 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2070 /* NB 'private' is different depending on what we're sending.... */
2071 LASSERT(!in_interrupt());
2073 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2074 type, nob, niov, libcfs_id2str(target));
2076 LASSERTF(nob == 0 || niov > 0,
2077 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2078 LASSERTF(niov <= LNET_MAX_IOV,
2079 "lntmsg %p niov %d\n", lntmsg, niov);
2082 mpflag = memalloc_noreclaim_save();
2086 CERROR("lntmsg %p with unexpected type %d\n",
2091 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2099 if (routing || target_is_router)
2100 break; /* send IMMEDIATE */
2102 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2103 * check the eventual sink buffer here - if small enough, remote
2104 * end is perfectly capable of returning data in short message -
2105 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2106 * for IMMEDIATE messages which will have it send a real reply instead
2107 * of doing kgnilnd_recv to have the RDMA continued */
2108 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2111 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2112 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2114 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2120 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2121 lntmsg->msg_md->md_kiov,
2122 0, lntmsg->msg_md->md_length);
2124 CERROR("unable to setup buffer: %d\n", rc);
2125 kgnilnd_tx_done(tx, rc);
2130 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2131 if (tx->tx_lntmsg[1] == NULL) {
2132 CERROR("Can't create reply for GET to %s\n",
2133 libcfs_nid2str(target.nid));
2134 kgnilnd_tx_done(tx, rc);
2139 tx->tx_lntmsg[0] = lntmsg;
2140 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2141 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2143 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2145 /* rest of tx_msg is setup just before it is sent */
2146 kgnilnd_launch_tx(tx, net, &target);
2148 case LNET_MSG_REPLY:
2150 /* to save on MDDs, we'll handle short kiov by vmap'ing
2151 * and sending via SMSG */
2152 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2155 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2156 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2158 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2165 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2168 kgnilnd_tx_done(tx, rc);
2173 tx->tx_lntmsg[0] = lntmsg;
2174 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2175 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2177 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2179 /* rest of tx_msg is setup just before it is sent */
2180 kgnilnd_launch_tx(tx, net, &target);
2184 /* send IMMEDIATE */
2186 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2187 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2189 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2195 rc = kgnilnd_setup_immediate_buffer(tx, niov, NULL, kiov, offset, nob);
2197 kgnilnd_tx_done(tx, rc);
2201 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2202 tx->tx_lntmsg[0] = lntmsg;
2203 kgnilnd_launch_tx(tx, net, &target);
2206 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2208 memalloc_noreclaim_restore(mpflag);
2213 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2215 kgn_conn_t *conn = rx->grx_conn;
2216 kgn_msg_t *rxmsg = rx->grx_msg;
2217 unsigned int niov = lntmsg->msg_niov;
2218 struct bio_vec *kiov = lntmsg->msg_kiov;
2219 unsigned int offset = lntmsg->msg_offset;
2220 unsigned int nob = lntmsg->msg_len;
2225 switch (rxmsg->gnm_type) {
2226 case GNILND_MSG_PUT_REQ_REV:
2227 done_type = GNILND_MSG_PUT_DONE_REV;
2230 case GNILND_MSG_GET_REQ:
2231 done_type = GNILND_MSG_GET_DONE;
2234 CERROR("invalid msg type %s (%d)\n",
2235 kgnilnd_msgtype2str(rxmsg->gnm_type),
2240 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2244 rc = kgnilnd_set_tx_id(tx, conn);
2248 rc = kgnilnd_setup_rdma_buffer(tx, niov, kiov, offset, nob);
2252 tx->tx_lntmsg[0] = lntmsg;
2253 tx->tx_getinfo = rxmsg->gnm_u.get;
2255 /* we only queue from kgnilnd_recv - we might get called from other contexts
2256 * and we don't want to block the mutex in those cases */
2258 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2259 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2260 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2261 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2266 kgnilnd_tx_done(tx, rc);
2267 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2269 lnet_finalize(lntmsg, rc);
2273 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2276 kgn_rx_t *rx = private;
2277 kgn_conn_t *conn = rx->grx_conn;
2278 kgn_msg_t *rxmsg = rx->grx_msg;
2279 kgn_msg_t *eagermsg = NULL;
2280 kgn_peer_t *peer = NULL;
2281 kgn_conn_t *found_conn = NULL;
2283 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2284 conn, rxmsg, lntmsg);
2286 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2287 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2288 rxmsg->gnm_payload_len);
2291 /* Grab a read lock so the connection doesnt disappear on us
2292 * while we look it up
2294 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2296 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2298 found_conn = kgnilnd_find_conn_locked(peer);
2301 /* Verify the connection found is the same one that the message
2302 * is supposed to be using, if it is not output an error message
2305 if (!peer || !found_conn
2306 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2307 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2308 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2309 peer, conn, found_conn);
2311 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2312 " from %s", rxmsg->gnm_connstamp,
2313 found_conn->gnc_peer_connstamp,
2314 libcfs_nid2str(peer->gnp_nid));
2319 /* add conn ref to ensure it doesn't go away until all eager
2320 * messages processed */
2321 kgnilnd_conn_addref(conn);
2323 /* Now that we have verified the connection is valid and added a
2324 * reference we can remove the read_lock on the peer_conn_lock */
2325 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2327 /* we have no credits or buffers for this message, so copy it
2328 * somewhere for a later kgnilnd_recv */
2329 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2330 *kgnilnd_tunables.kgn_eager_credits) {
2331 CERROR("Out of eager credits to %s\n",
2332 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2336 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2338 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2339 if (eagermsg == NULL) {
2340 kgnilnd_conn_decref(conn);
2341 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2342 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2346 /* copy msg and payload */
2347 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2348 rx->grx_msg = eagermsg;
2351 /* stash this for lnet_finalize on cancel-on-conn-close */
2352 rx->grx_lntmsg = lntmsg;
2354 /* keep the same rx_t, it just has a new grx_msg now */
2355 *new_private = private;
2357 /* release SMSG buffer */
2358 kgnilnd_release_msg(conn);
2364 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2365 int delayed, unsigned int niov,
2366 struct bio_vec *kiov,
2367 unsigned int offset, unsigned int mlen, unsigned int rlen)
2369 kgn_rx_t *rx = private;
2370 kgn_conn_t *conn = rx->grx_conn;
2371 kgn_msg_t *rxmsg = rx->grx_msg;
2377 LASSERT(!in_interrupt());
2378 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2380 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2381 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2382 conn, rxmsg, lntmsg,
2383 niov, kiov, offset, mlen, rlen);
2385 /* we need to lock here as recv can be called from any context */
2386 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2387 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2388 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2390 /* someone closed the conn after we copied this out, nuke it */
2391 kgnilnd_consume_rx(rx);
2392 lnet_finalize(lntmsg, conn->gnc_error);
2395 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2397 switch (rxmsg->gnm_type) {
2399 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2400 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2401 conn, rx, rxmsg, lntmsg, niov, kiov, offset, mlen, rlen);
2404 case GNILND_MSG_IMMEDIATE:
2405 if (mlen > rxmsg->gnm_payload_len) {
2406 GNIDBG_MSG(D_ERROR, rxmsg,
2407 "Immediate message from %s too big: %d > %d",
2408 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2409 rxmsg->gnm_payload_len);
2411 kgnilnd_consume_rx(rx);
2415 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2416 * right after the kgn_msg_t header - so just 'cute' way of saying
2417 * rxmsg + sizeof(kgn_msg_t) */
2419 /* check payload checksum if sent */
2421 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2422 !rxmsg->gnm_payload_cksum &&
2423 rxmsg->gnm_payload_len != 0)
2424 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2426 if (rxmsg->gnm_payload_cksum != 0) {
2427 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2428 * which is what is used to calculate the cksum on the TX side */
2429 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2431 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2432 GNIDBG_MSG(D_NETERROR, rxmsg,
2433 "Bad payload checksum (%x expected %x)",
2434 pload_cksum, rxmsg->gnm_payload_cksum);
2435 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2437 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2438 &rxmsg[1], rxmsg->gnm_payload_len);
2439 /* fall through to dump */
2441 libcfs_debug_dumplog();
2447 /* checksum problems are fatal, kill the conn */
2448 kgnilnd_consume_rx(rx);
2449 kgnilnd_close_conn(conn, rc);
2454 lnet_copy_flat2kiov(
2456 *kgnilnd_tunables.kgn_max_immediate,
2457 &rxmsg[1], 0, mlen);
2459 kgnilnd_consume_rx(rx);
2460 lnet_finalize(lntmsg, 0);
2463 case GNILND_MSG_PUT_REQ:
2464 /* LNET wants to truncate or drop transaction, sending NAK */
2466 kgnilnd_consume_rx(rx);
2467 lnet_finalize(lntmsg, 0);
2469 /* only error if lntmsg == NULL, otherwise we are just
2470 * short circuiting the rdma process of 0 bytes */
2471 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2472 lntmsg == NULL ? -ENOENT : 0,
2473 rxmsg->gnm_u.get.gngm_cookie,
2477 /* sending ACK with sink buff. info */
2478 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2480 kgnilnd_consume_rx(rx);
2484 rc = kgnilnd_set_tx_id(tx, conn);
2486 GOTO(nak_put_req, rc);
2489 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2490 kiov, offset, mlen);
2492 GOTO(nak_put_req, rc);
2495 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2496 rxmsg->gnm_u.putreq.gnprm_cookie;
2497 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2498 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2499 (__u64)((unsigned long)tx->tx_buffer);
2500 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2502 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2503 tx->tx_qtime = jiffies;
2504 /* we only queue from kgnilnd_recv - we might get called from other contexts
2505 * and we don't want to block the mutex in those cases */
2507 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2508 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2509 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2510 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2512 kgnilnd_consume_rx(rx);
2516 /* make sure we send an error back when the PUT fails */
2517 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2518 kgnilnd_tx_done(tx, rc);
2519 kgnilnd_consume_rx(rx);
2521 /* return magic LNet network error */
2523 case GNILND_MSG_GET_REQ_REV:
2524 /* LNET wants to truncate or drop transaction, sending NAK */
2526 kgnilnd_consume_rx(rx);
2527 lnet_finalize(lntmsg, 0);
2529 /* only error if lntmsg == NULL, otherwise we are just
2530 * short circuiting the rdma process of 0 bytes */
2531 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2532 lntmsg == NULL ? -ENOENT : 0,
2533 rxmsg->gnm_u.get.gngm_cookie,
2537 /* lntmsg can be null when parsing a LNET_GET */
2538 if (lntmsg != NULL) {
2539 /* sending ACK with sink buff. info */
2540 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2542 kgnilnd_consume_rx(rx);
2546 rc = kgnilnd_set_tx_id(tx, conn);
2548 GOTO(nak_get_req_rev, rc);
2550 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2551 kiov, offset, mlen);
2553 GOTO(nak_get_req_rev, rc);
2555 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2556 rxmsg->gnm_u.putreq.gnprm_cookie;
2557 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2558 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2559 (__u64)((unsigned long)tx->tx_buffer);
2560 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2562 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2564 /* we only queue from kgnilnd_recv - we might get called from other contexts
2565 * and we don't want to block the mutex in those cases */
2567 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2568 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2569 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2570 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2573 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2575 rxmsg->gnm_u.get.gngm_cookie,
2579 kgnilnd_consume_rx(rx);
2583 /* make sure we send an error back when the GET fails */
2584 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2585 kgnilnd_tx_done(tx, rc);
2586 kgnilnd_consume_rx(rx);
2588 /* return magic LNet network error */
2592 case GNILND_MSG_PUT_REQ_REV:
2593 /* LNET wants to truncate or drop transaction, sending NAK */
2595 kgnilnd_consume_rx(rx);
2596 lnet_finalize(lntmsg, 0);
2598 /* only error if lntmsg == NULL, otherwise we are just
2599 * short circuiting the rdma process of 0 bytes */
2600 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2601 lntmsg == NULL ? -ENOENT : 0,
2602 rxmsg->gnm_u.get.gngm_cookie,
2607 if (lntmsg != NULL) {
2609 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2612 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2614 rxmsg->gnm_u.get.gngm_cookie,
2617 kgnilnd_consume_rx(rx);
2619 case GNILND_MSG_GET_REQ:
2620 if (lntmsg != NULL) {
2622 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2625 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2627 rxmsg->gnm_u.get.gngm_cookie,
2630 kgnilnd_consume_rx(rx);
2636 /* needs write_lock on kgn_peer_conn_lock held */
2638 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2640 unsigned long timeout, keepalive;
2641 unsigned long now = jiffies;
2642 unsigned long newest_last_rx;
2645 /* given that we found this conn hanging off a peer, it better damned
2646 * well be connected */
2647 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2648 "conn 0x%p->%s with bad state%s\n", conn,
2649 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2651 kgnilnd_conn_state2str(conn));
2653 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2654 "rx_diff %lu tx_diff %lu\n",
2655 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2656 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2657 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2658 cfs_duration_sec(now - conn->gnc_last_tx));
2660 timeout = cfs_time_seconds(conn->gnc_timeout);
2661 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2663 /* just in case our lack of RX msg processing is gumming up the works - give the
2664 * remove an extra chance */
2666 newest_last_rx = GNILND_LASTRX(conn);
2668 if (time_after_eq(now, newest_last_rx + timeout)) {
2669 uint32_t level = D_CONSOLE|D_NETERROR;
2671 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2674 GNIDBG_CONN(level, conn,
2675 "No gnilnd traffic received from %s for %lu "
2676 "seconds, terminating connection. Is node down? ",
2677 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2678 cfs_duration_sec(now - newest_last_rx));
2682 /* we don't timeout on last_tx stalls - we are going to trust the
2683 * underlying network to let us know when sends are failing.
2684 * At worst, the peer will timeout our RX stamp and drop the connection
2685 * at that point. We'll then see his CLOSE or at worst his RX
2686 * stamp stop and drop the connection on our end */
2688 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2689 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2690 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2691 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2692 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2694 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2695 conn->gnc_last_noop_cq,
2696 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2697 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2698 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2699 set_mb(conn->gnc_last_noop_want, jiffies);
2700 atomic_inc(&conn->gnc_reaper_noop);
2701 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2704 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2707 kgnilnd_queue_tx(conn, tx);
2713 /* needs write_lock on kgn_peer_conn_lock held */
2715 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2716 struct list_head *souls)
2718 unsigned long timeout;
2719 kgn_conn_t *conn, *connN = NULL;
2725 short releaseconn = 0;
2726 unsigned long first_rx = 0;
2727 int purgatory_conn_cnt = 0;
2729 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2730 peer, libcfs_nid2str(peer->gnp_nid),
2731 peer->gnp_reconnect_interval);
2733 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2734 GNILND_MIN_TIMEOUT));
2736 conn = kgnilnd_find_conn_locked(peer);
2738 /* if there is a valid conn, check the queues for timeouts */
2739 rc = kgnilnd_check_conn_timeouts_locked(conn);
2741 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2742 /* simulate a RX CLOSE after the timeout but before
2743 * the scheduler thread gets it */
2744 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2745 conn->gnc_peer_error = -ETIMEDOUT;
2748 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2750 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2751 CDEBUG(D_WARNING, "%s conn timed out, will "
2752 "reconnect upon request from peer\n",
2753 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2755 /* Once we mark closed, any of the scheduler threads could
2756 * get it and move through before we hit the fail loc code */
2757 kgnilnd_close_conn_locked(conn, rc);
2759 /* first_rx is used to decide when to release a conn from purgatory.
2761 first_rx = conn->gnc_first_rx;
2765 /* now regardless of starting new conn, find tx on peer queue that
2766 * are old and smell bad - do this first so we don't trigger
2767 * reconnect on empty queue if we timeout all */
2768 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2769 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2771 LCONSOLE_INFO("could not send to %s due to connection"
2772 " setup failure after %lu seconds\n",
2773 libcfs_nid2str(peer->gnp_nid),
2774 cfs_duration_sec(jiffies - tx->tx_qtime));
2776 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2778 list_add_tail(&tx->tx_list, todie);
2783 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2784 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2785 count, peer, libcfs_nid2str(peer->gnp_nid));
2786 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2787 if (list_empty(&peer->gnp_tx_queue) ||
2788 peer->gnp_connecting == GNILND_PEER_KILL) {
2789 /* we pass down todie to use a common function - but we know there are
2791 kgnilnd_cancel_peer_connect_locked(peer, todie);
2795 /* Don't reconnect if we are still trying to clear out old conns.
2796 * This prevents us sending traffic on the new mbox before ensuring we are done
2797 * with the old one */
2798 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2799 (atomic_read(&peer->gnp_dirty_eps) == 0);
2801 /* fast reconnect after a timeout */
2802 to_reconn = !conn &&
2803 (peer->gnp_last_errno == -ETIMEDOUT) &&
2804 *kgnilnd_tunables.kgn_fast_reconn;
2806 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2807 * to be sent, we'll check the reconnect interval and fire up a new
2808 * connection request */
2811 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2812 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2813 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2815 CDEBUG(D_NET, "starting connect to %s\n",
2816 libcfs_nid2str(peer->gnp_nid));
2817 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2818 "Peer was idle and we have a write_lock, state issue %d\n",
2819 peer->gnp_connecting);
2821 peer->gnp_connecting = GNILND_PEER_CONNECT;
2822 kgnilnd_peer_addref(peer); /* extra ref for connd */
2824 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2825 list_add_tail(&peer->gnp_connd_list,
2826 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2827 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2829 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2832 /* fail_loc to allow us to delay release of purgatory */
2833 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2836 /* This check allows us to verify that the new conn is actually being used. This allows us to
2837 * pull the old conns out of purgatory if they have actually seen traffic.
2838 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2841 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2842 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2843 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2847 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2848 /* check for purgatory timeouts */
2849 if (conn->gnc_in_purgatory) {
2850 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2851 * that way the next time we check it we can detach it from purgatory
2854 if (conn->gnc_state != GNILND_CONN_DONE) {
2855 /* Skip over conns that are currently not DONE. If they arent already scheduled
2856 * for completion something in the state machine is broken.
2861 /* We only detach a conn that is in purgatory if we have received a close message,
2862 * we have a new valid connection that has successfully received data, or an admin
2863 * command tells us we need to detach.
2866 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2867 unsigned long waiting;
2869 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2871 /* C.E: The remote peer is expected to close the
2872 * connection (see kgnilnd_check_conn_timeouts)
2873 * via the reaper thread and nuke out the MDD and
2874 * FMA resources after conn->gnc_timeout has expired
2875 * without an FMA RX */
2876 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2877 " held resources\n",
2878 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2879 cfs_duration_sec(waiting));
2881 kgnilnd_detach_purgatory_locked(conn, souls);
2883 purgatory_conn_cnt++;
2888 /* If we have too many connections in purgatory we could run out of
2889 * resources. Limit the number of connections to a tunable number,
2890 * clean up to the minimum all in one fell swoop... there are
2891 * situations where dvs will retry tx's and we can eat up several
2892 * hundread connection requests at once.
2894 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2895 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2897 if (conn->gnc_in_purgatory &&
2898 conn->gnc_state == GNILND_CONN_DONE) {
2899 CDEBUG(D_NET, "Dropping Held resource due to"
2900 " resource limits being hit\n");
2901 kgnilnd_detach_purgatory_locked(conn, souls);
2903 if (purgatory_conn_cnt-- <
2904 *kgnilnd_tunables.kgn_max_purgatory)
2912 kgnilnd_reaper_check(int idx)
2914 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2915 struct list_head *ctmp, *ctmpN;
2916 LIST_HEAD(geriatrics);
2919 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2921 list_for_each_safe(ctmp, ctmpN, peers) {
2922 kgn_peer_t *peer = NULL;
2924 /* don't timeout stuff if the network is mucked or shutting down */
2925 if (kgnilnd_check_hw_quiesce()) {
2928 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2930 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
2933 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2935 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
2936 kgnilnd_release_purgatory_list(&souls);
2940 kgnilnd_update_reaper_timeout(long timeout)
2942 LASSERT(timeout > 0);
2944 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2946 if (timeout < kgnilnd_data.kgn_new_min_timeout)
2947 kgnilnd_data.kgn_new_min_timeout = timeout;
2949 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2953 kgnilnd_reaper_poke_with_stick(unsigned long arg)
2955 wake_up(&kgnilnd_data.kgn_reaper_waitq);
2959 kgnilnd_reaper(void *arg)
2964 unsigned long next_check_time = jiffies;
2965 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
2966 struct timer_list timer;
2969 /* all gnilnd threads need to run fairly urgently */
2970 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2971 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2973 while (!kgnilnd_data.kgn_shutdown) {
2974 /* I wake up every 'p' seconds to check for timeouts on some
2975 * more peers. I try to check every connection 'n' times
2976 * within the global minimum of all keepalive and timeout
2977 * intervals, to ensure I attend to every connection within
2978 * (n+1)/n times its timeout intervals. */
2979 const int p = GNILND_REAPER_THREAD_WAKE;
2980 const int n = GNILND_REAPER_NCHECKS;
2982 /* to quiesce or to not quiesce, that is the question */
2983 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2984 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2985 KGNILND_SPIN_QUIESCE;
2986 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2989 /* careful with the jiffy wrap... */
2990 timeout = (long)(next_check_time - jiffies);
2993 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
2994 TASK_INTERRUPTIBLE);
2995 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2996 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
2998 mod_timer(&timer, (long) jiffies + timeout);
3000 /* check flag variables before committing */
3001 if (!kgnilnd_data.kgn_shutdown &&
3002 !kgnilnd_data.kgn_quiesce_trigger) {
3003 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3004 timeout, cfs_duration_sec(timeout));
3006 CDEBUG(D_INFO, "awake after schedule\n");
3009 del_singleshot_timer_sync(&timer);
3010 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3011 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3015 /* new_min_timeout is set from the conn timeouts and keepalive
3016 * this should end up with a min timeout of
3017 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3018 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3019 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3020 CDEBUG(D_NET, "Set new min timeout %ld\n",
3021 current_min_timeout);
3024 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3026 /* Compute how many table entries to check now so I get round
3027 * the whole table fast enough given that I do this at fixed
3028 * intervals of 'p' seconds) */
3029 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3030 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3031 chunk = (chunk * n * p) /
3032 kgnilnd_data.kgn_new_min_timeout;
3035 for (i = 0; i < chunk; i++) {
3036 kgnilnd_reaper_check(hash_index);
3037 hash_index = (hash_index + 1) %
3038 *kgnilnd_tunables.kgn_peer_hash_size;
3040 next_check_time = (long) jiffies + cfs_time_seconds(p);
3041 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3043 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3046 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3048 kgnilnd_thread_fini();
3053 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3054 unsigned niov, offset, nob;
3055 struct bio_vec *kiov;
3056 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3057 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3060 lnet_copy_flat2kiov(
3063 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3065 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3072 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3075 gni_post_descriptor_t *desc;
3077 kgn_tx_ev_id_t ev_id;
3079 int should_retry, rc;
3080 long num_processed = 0;
3081 kgn_conn_t *conn = NULL;
3082 kgn_tx_t *tx = NULL;
3083 kgn_rdma_desc_t *rdesc;
3088 /* make sure we don't keep looping if we need to reset */
3089 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3090 return num_processed;
3092 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3094 /* we didn't get the mutex, so return that there is still work
3098 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3099 /* a bit gross - but we need a good way to test for
3100 * delayed RDMA completions and the easiest way to do
3101 * that is to delay the RDMA CQ events */
3102 rrc = GNI_RC_NOT_DONE;
3104 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3107 if (rrc == GNI_RC_NOT_DONE) {
3108 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3109 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3110 dev->gnd_id, num_processed);
3111 return num_processed;
3113 dev->gnd_sched_alive = jiffies;
3116 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3117 "this is bad, somehow our credits didn't protect us"
3118 " from CQ overrun\n");
3119 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3120 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3121 event_data, GNI_CQ_GET_TYPE(event_data));
3123 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3125 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3127 /* XXX Nic: Need better error handling here... */
3128 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3129 (rrc == GNI_RC_TRANSACTION_ERROR),
3132 ev_id.txe_cookie = desc->post_id;
3134 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3136 if (conn == NULL || tx == NULL) {
3137 /* either conn or tx was already nuked and this is a "late"
3138 * completion, so drop it */
3142 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3143 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3144 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3145 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3146 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3148 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3150 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3151 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3152 tx->tx_msg.gnm_u.completion.gncm_retval);
3156 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3157 if (tx->tx_buffer_copy != NULL)
3158 kgnilnd_recv_bte_get(tx);
3159 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3162 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3163 if (tx->tx_buffer_copy != NULL)
3164 kgnilnd_recv_bte_get(tx);
3165 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3168 /* remove from rdmaq */
3169 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3170 spin_lock(&conn->gnc_list_lock);
3171 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3172 spin_unlock(&conn->gnc_list_lock);
3173 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3175 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3176 event_data = 1LL << 48;
3180 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3181 atomic_inc(&dev->gnd_rdma_ntx);
3182 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3183 /* transaction succeeded, add into fmaq */
3184 kgnilnd_queue_tx(conn, tx);
3185 kgnilnd_peer_alive(conn->gnc_peer);
3187 /* drop ref from kgnilnd_validate_tx_ev_id */
3188 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3189 kgnilnd_conn_decref(conn);
3194 /* fall through to the TRANSACTION_ERROR case */
3197 /* get stringified version for log messages */
3198 kgnilnd_cq_error_str(event_data, &err_str, 256);
3199 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3201 /* make sure we are not off in the weeds with this tx */
3202 if (tx->tx_retrans >
3203 *kgnilnd_tunables.kgn_max_retransmits) {
3204 GNIDBG_TX(D_NETERROR, tx,
3205 "giving up on TX, too many retries", NULL);
3209 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3210 should_retry ? "transient" : "unrecoverable", err_str);
3212 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3213 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3214 rdesc = &tx->tx_putinfo.gnpam_desc;
3215 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3216 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3218 rdesc = &tx->tx_getinfo.gngm_desc;
3219 rnob = tx->tx_lntmsg[0]->msg_len;
3220 rcookie = tx->tx_getinfo.gngm_cookie;
3225 tx->tx_msg.gnm_type,
3229 kgnilnd_nak_rdma(conn,
3230 tx->tx_msg.gnm_type,
3233 tx->tx_msg.gnm_srcnid);
3234 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3235 kgnilnd_close_conn(conn, -ECOMM);
3238 /* drop ref from kgnilnd_validate_tx_ev_id */
3239 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3240 kgnilnd_conn_decref(conn);
3245 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3249 kgn_tx_ev_id_t ev_id;
3250 kgn_tx_t *tx = NULL;
3251 kgn_conn_t *conn = NULL;
3252 int queued_fma, saw_reply, rc;
3253 long num_processed = 0;
3254 struct list_head *ctmp, *ctmpN;
3257 /* make sure we don't keep looping if we need to reset */
3258 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3259 return num_processed;
3262 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3264 /* we didn't get the mutex, so return that there is still work
3269 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3270 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3272 if (rrc == GNI_RC_NOT_DONE) {
3274 "SMSG send CQ %d not ready (data %#llx) "
3275 "processed %ld\n", dev->gnd_id, event_data,
3278 if (num_processed > 0) {
3279 spin_lock(&dev->gnd_lock);
3280 if (!list_empty(&dev->gnd_delay_conns)) {
3281 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3282 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3283 list_del_init(&conn->gnc_delaylist);
3284 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3285 kgnilnd_schedule_conn_nolock(conn);
3287 spin_unlock(&dev->gnd_lock);
3288 kgnilnd_schedule_device(dev);
3290 spin_unlock(&dev->gnd_lock);
3293 return num_processed;
3296 dev->gnd_sched_alive = jiffies;
3299 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3300 "this is bad, somehow our credits didn't "
3301 "protect us from CQ overrun\n");
3302 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3303 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3304 event_data, GNI_CQ_GET_TYPE(event_data));
3306 /* if SMSG couldn't handle an error, time for conn to die */
3307 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3310 /* need to take the write_lock to ensure atomicity
3311 * on the conn state if we need to close it */
3312 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3313 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3315 /* Conn was destroyed? */
3317 "SMSG CQID lookup %#llx failed\n",
3318 GNI_CQ_GET_INST_ID(event_data));
3319 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3323 kgnilnd_cq_error_str(event_data, &err_str, 256);
3324 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3325 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3327 kgnilnd_close_conn_locked(conn, -ECOMM);
3329 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3331 /* no need to process rest of this tx -
3332 * it is getting canceled */
3336 /* fall through to GNI_RC_SUCCESS case */
3337 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3339 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3340 if (conn == NULL || tx == NULL) {
3341 /* either conn or tx was already nuked and this is a "late"
3342 * completion, so drop it */
3346 tx->tx_conn->gnc_last_tx_cq = jiffies;
3347 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3348 set_mb(conn->gnc_last_noop_cq, jiffies);
3351 /* lock tx_list_state and tx_state */
3352 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3353 spin_lock(&tx->tx_conn->gnc_list_lock);
3355 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3356 "state not GNILND_TX_LIVE_FMAQ", NULL);
3357 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3358 "not waiting for completion", NULL);
3360 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3363 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3365 /* This will trigger other FMA sends that were
3366 * pending this completion */
3367 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3369 /* we either did not expect reply or we already got it */
3370 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3372 spin_unlock(&tx->tx_conn->gnc_list_lock);
3373 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3376 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3378 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3379 kgnilnd_schedule_conn(conn);
3382 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3383 * If saw_reply is true we know that the tx is safe to use as the other thread
3384 * is already finished with it.
3388 /* no longer need to track on the live_fmaq */
3389 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3391 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3392 /* we already got reply & were waiting for
3393 * completion of initial send */
3394 /* to initiate RDMA transaction */
3395 GNIDBG_TX(D_NET, tx,
3396 "Pending RDMA 0x%p type 0x%02x",
3397 tx->tx_msg.gnm_type);
3398 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3399 rc = kgnilnd_send_mapped_tx(tx, 0);
3400 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3402 /* we are done with this tx */
3403 GNIDBG_TX(D_NET, tx,
3404 "Done with tx type 0x%02x",
3405 tx->tx_msg.gnm_type);
3406 kgnilnd_tx_done(tx, tx->tx_rc);
3410 /* drop ref from kgnilnd_validate_tx_ev_id */
3411 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3412 kgnilnd_conn_decref(conn);
3414 /* if we are waiting for a REPLY, we'll handle the tx then */
3415 } /* end for loop */
3419 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3424 long num_processed = 0;
3425 struct list_head *conns;
3426 struct list_head *tmp;
3430 /* make sure we don't keep looping if we need to reset */
3431 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3432 return num_processed;
3435 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3437 /* we didn't get the mutex, so return that there is still work
3441 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3442 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3444 if (rrc == GNI_RC_NOT_DONE) {
3445 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3447 dev->gnd_id, event_data, num_processed);
3448 return num_processed;
3450 dev->gnd_sched_alive = jiffies;
3453 /* this is the only CQ that can really handle transient
3455 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3456 rrc = cfs_fail_val ? cfs_fail_val
3457 : GNI_RC_ERROR_RESOURCE;
3458 if (rrc == GNI_RC_ERROR_RESOURCE) {
3459 /* set overrun too */
3460 event_data |= (1UL << 63);
3461 LASSERTF(GNI_CQ_OVERRUN(event_data),
3462 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3465 /* sender should get error event too and take care
3466 of failed transaction by re-transmitting */
3467 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3468 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3472 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3473 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3474 conn = kgnilnd_cqid2conn_locked(
3475 GNI_CQ_GET_INST_ID(event_data));
3477 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3478 "failed, dropping event %#llx\n",
3479 GNI_CQ_GET_INST_ID(event_data),
3482 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3484 GNI_CQ_GET_INST_ID(event_data),
3485 conn, conn->gnc_peer ?
3486 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3489 conn->gnc_last_rx_cq = jiffies;
3491 /* stash first rx so we can clear out purgatory.
3493 if (conn->gnc_first_rx == 0) {
3494 conn->gnc_first_rx = jiffies;
3496 kgnilnd_peer_alive(conn->gnc_peer);
3497 kgnilnd_schedule_conn(conn);
3499 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3503 /* FMA CQ has overflowed: check ALL conns */
3504 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3505 "conns on device %d\n", dev->gnd_id);
3507 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3509 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3510 conns = &kgnilnd_data.kgn_conns[rc];
3512 list_for_each(tmp, conns) {
3513 conn = list_entry(tmp, kgn_conn_t,
3516 if (conn->gnc_device == dev) {
3517 kgnilnd_schedule_conn(conn);
3518 conn->gnc_last_rx_cq = jiffies;
3522 /* don't block write lockers for too long... */
3523 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3528 /* try_map_if_full should only be used when processing TX from list of
3529 * backlog TX waiting on mappings to free up
3532 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3533 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3536 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3538 /* slight bit of race if multiple people calling, but at worst we'll have
3539 * order altered just a bit... which would not be determenistic anyways */
3540 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3542 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3544 /* We know that we have a GART reservation that should guarantee forward progress.
3545 * This means we don't need to take any extraordinary efforts if we are failing
3546 * mappings here - even if we are holding a very small number of these. */
3548 if (try_map_if_full || (rc == 0)) {
3549 rc = kgnilnd_map_buffer(tx);
3552 /* rc should be 0 if we mapped successfully here, if non-zero
3553 * we are queueing */
3555 /* if try_map_if_full set, they handle requeuing */
3556 if (unlikely(try_map_if_full)) {
3559 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3560 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3561 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3562 /* make sure we wake up sched to run this */
3563 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3564 /* return 0 as this is now queued for later sending */
3569 switch (tx->tx_msg.gnm_type) {
3573 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3574 * remote node where the RDMA will be started
3575 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3576 * be satisified. The rest of the errors are "hard" errors that require
3577 * upper layers to handle themselves.
3578 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3579 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3580 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3581 * GNILND_MSG_GET_DONE_REV.
3583 case GNILND_MSG_GET_REQ:
3584 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3585 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3586 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3587 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3588 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3589 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3590 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3592 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3593 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3595 case GNILND_MSG_PUT_ACK:
3596 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3597 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3598 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3599 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3601 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3602 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3605 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3606 case GNILND_MSG_PUT_DONE:
3607 case GNILND_MSG_PUT_REQ:
3608 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3609 &tx->tx_putinfo.gnpam_desc,
3610 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3611 tx->tx_putinfo.gnpam_dst_cookie);
3612 RETURN(try_map_if_full ? rc : 0);
3614 case GNILND_MSG_GET_DONE:
3615 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3616 &tx->tx_getinfo.gngm_desc,
3617 tx->tx_lntmsg[0]->msg_len,
3618 tx->tx_getinfo.gngm_cookie);
3619 RETURN(try_map_if_full ? rc : 0);
3621 case GNILND_MSG_PUT_REQ_REV:
3622 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3623 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3624 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3625 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3626 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3627 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3628 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3630 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3632 case GNILND_MSG_PUT_DONE_REV:
3633 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3634 &tx->tx_getinfo.gngm_desc,
3636 tx->tx_getinfo.gngm_cookie);
3637 RETURN(try_map_if_full ? rc : 0);
3639 case GNILND_MSG_GET_ACK_REV:
3640 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3641 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3642 /* LNET_GETS are a special case for parse */
3643 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3644 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3646 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3647 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3649 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3650 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3652 case GNILND_MSG_GET_DONE_REV:
3653 case GNILND_MSG_GET_REQ_REV:
3654 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3655 &tx->tx_putinfo.gnpam_desc,
3656 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3657 tx->tx_putinfo.gnpam_dst_cookie);
3658 RETURN(try_map_if_full ? rc : 0);
3666 kgnilnd_process_fmaq(kgn_conn_t *conn)
3669 kgn_tx_t *tx = NULL;
3670 void *buffer = NULL;
3671 unsigned int nob = 0;
3674 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3675 * However I will be rescheduled by an FMA completion event
3676 * when I eventually get some.
3677 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3678 * But it doesn't matter if I try to send a "real" message just
3679 * as I start closing because I'll get scheduled to send the
3682 /* Short circuit if the ep_handle is null we cant send anyway. */
3683 if (conn->gnc_ephandle == NULL)
3686 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3688 spin_lock(&conn->gnc_list_lock);
3690 if (list_empty(&conn->gnc_fmaq)) {
3691 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3693 spin_unlock(&conn->gnc_list_lock);
3695 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3696 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3697 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3698 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3699 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3701 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3702 conn->gnc_last_noop_cq,
3703 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3704 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3705 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3706 atomic_inc(&conn->gnc_sched_noop);
3707 set_mb(conn->gnc_last_noop_want, jiffies);
3709 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3712 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3716 rc = kgnilnd_set_tx_id(tx, conn);
3718 kgnilnd_tx_done(tx, rc);
3724 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3725 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3726 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3727 more_to_do = !list_empty(&conn->gnc_fmaq);
3728 spin_unlock(&conn->gnc_list_lock);
3731 /* if there is no real TX or no NOOP to send, bail */
3736 if (!tx->tx_retrans)
3737 tx->tx_cred_wait = jiffies;
3739 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3740 "tx with zero id", NULL);
3742 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3743 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3744 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3748 switch (tx->tx_msg.gnm_type) {
3752 case GNILND_MSG_NOOP:
3753 case GNILND_MSG_CLOSE:
3754 case GNILND_MSG_IMMEDIATE:
3755 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3756 buffer = tx->tx_buffer;
3760 case GNILND_MSG_GET_DONE:
3761 case GNILND_MSG_PUT_DONE:
3762 case GNILND_MSG_PUT_DONE_REV:
3763 case GNILND_MSG_GET_DONE_REV:
3764 case GNILND_MSG_PUT_NAK:
3765 case GNILND_MSG_GET_NAK:
3766 case GNILND_MSG_GET_NAK_REV:
3767 case GNILND_MSG_PUT_NAK_REV:
3768 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3771 case GNILND_MSG_PUT_REQ:
3772 case GNILND_MSG_GET_REQ_REV:
3773 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3775 case GNILND_MSG_PUT_ACK:
3776 case GNILND_MSG_PUT_REQ_REV:
3777 case GNILND_MSG_GET_ACK_REV:
3778 case GNILND_MSG_GET_REQ:
3779 /* This is really only to handle the retransmit of SMSG once these
3780 * two messages are setup in send_mapped_tx */
3781 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3785 if (likely(rc == 0)) {
3786 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3790 /* don't explicitly reschedule here - we are short credits and will rely on
3791 * kgnilnd_sendmsg to resched the conn if need be */
3793 } else if (rc < 0) {
3794 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3795 * almost certainly a software bug, but lets play nice with the other kids */
3796 kgnilnd_tx_done(tx, rc);
3797 /* just for fun, kick peer in arse - resetting conn might help to correct
3798 * this almost certainly buggy software caused return code */
3799 kgnilnd_close_conn(conn, rc);
3803 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3804 kgnilnd_schedule_conn(conn);
3809 kgnilnd_process_rdmaq(kgn_device_t *dev)
3814 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3818 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3819 unsigned long dead_bump;
3822 /* if we think we need to adjust, take lock to serialize and recheck */
3823 spin_lock(&dev->gnd_rdmaq_lock);
3824 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3825 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3827 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3829 /* roll the bucket forward */
3830 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3832 if (kgnilnd_data.kgn_rdmaq_override &&
3833 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3834 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3839 /* roll current outstanding forward to make sure we carry outstanding
3840 * committment forward
3841 * new_ok starts out as the whole interval value
3842 * - first subtract bytes_out from last interval, as that would push us over
3843 * strict limits for this interval
3844 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3846 * there is a small race here if someone is actively processing mappings and
3847 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3848 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3849 * - if this gives us problems in the future, we could use a read/write lock
3850 * to protect the resetting of these values */
3851 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3852 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3854 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3855 "current out %ld\n",
3856 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3857 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3859 spin_unlock(&dev->gnd_rdmaq_lock);
3862 spin_lock(&dev->gnd_rdmaq_lock);
3863 while (!list_empty(&dev->gnd_rdmaq)) {
3866 /* make sure we break out early on quiesce */
3867 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3868 /* always break with lock held - we unlock outside loop */
3872 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3873 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3876 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3877 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3878 /* if conn is dying, mark tx in tx_ref_table for
3879 * kgnilnd_complete_closed_conn to finish up */
3880 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3882 /* tx was moved to DYING, get next */
3885 spin_unlock(&dev->gnd_rdmaq_lock);
3887 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3888 spin_lock(&dev->gnd_rdmaq_lock);
3891 /* no ticket! add back to head */
3892 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3893 /* clear found_work so scheduler threads wait for timer */
3897 /* TX is GO for launch */
3898 tx->tx_qtime = jiffies;
3899 kgnilnd_send_mapped_tx(tx, 0);
3903 spin_unlock(&dev->gnd_rdmaq_lock);
3909 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3911 __swab64s(&d->gnrd_key.qword1);
3912 __swab64s(&d->gnrd_key.qword2);
3913 __swab64s(&d->gnrd_addr);
3914 __swab32s(&d->gnrd_nob);
3917 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3918 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3921 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3923 kgn_tx_ev_id_t ev_id;
3926 /* we use the cookie from the original TX, so we can find the match
3927 * by parsing that and using the txe_idx */
3928 ev_id.txe_cookie = cookie;
3930 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3933 /* check tx to make sure kgni didn't eat it */
3934 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3935 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3937 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
3938 (tx->tx_id.txe_cookie = cookie)),
3939 "conn 0x%p->%s tx_ref_table hosed: wanted "
3940 "txe_cookie %#llx txe_idx %d "
3941 "found tx %p cookie %#llx txe_idx %d\n",
3942 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
3943 cookie, ev_id.txe_idx,
3944 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
3946 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
3947 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
3948 "Unexpected TX type (%x, %x or %x) "
3949 "or state (%x, expected +%x) "
3950 "matched reply from %s\n",
3951 tx->tx_msg.gnm_type, type1, type2,
3952 tx->tx_state, GNILND_TX_WAITING_REPLY,
3953 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3955 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
3956 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
3962 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
3965 kgn_conn_t *conn = tx->tx_conn;
3966 __u64 nob = tx->tx_nob;
3967 __u32 physnop = tx->tx_phys_npages;
3968 int id = tx->tx_id.txe_smsg_id;
3969 int buftype = tx->tx_buftype;
3970 gni_mem_handle_t hndl;
3971 hndl.qword1 = tx->tx_map_key.qword1;
3972 hndl.qword2 = tx->tx_map_key.qword2;
3974 spin_lock(&conn->gnc_list_lock);
3976 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
3977 "not waiting for reply", NULL);
3980 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
3982 if (rc == -EFAULT) {
3983 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
3984 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
3986 if(*kgnilnd_tunables.kgn_efault_lbug) {
3987 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
3988 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
3990 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
3991 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
3992 kgnilnd_tx_state2str(tx->tx_list_state),
3993 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
3998 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
3999 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4000 /* sample under lock as follow on steps require gnc_list_lock
4001 * - or call kgnilnd_tx_done which requires no locks held over
4002 * call to lnet_finalize */
4005 spin_unlock(&conn->gnc_list_lock);
4008 kgnilnd_tx_done(tx, tx->tx_rc);
4013 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4016 kgn_conn_t *conn = tx->tx_conn;
4018 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4019 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4021 /* the gncm_retval is passed in for PUTs */
4022 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4023 msg->gnm_u.completion.gncm_retval);
4025 kgnilnd_complete_tx(tx, rc);
4029 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4037 kgn_peer_t *peer = conn->gnc_peer;
4040 __u16 tmp_cksum = 0, msg_cksum = 0;
4041 int repost = 1, saw_complete;
4042 unsigned long timestamp, newest_last_rx, timeout;
4046 /* Short circuit if the ep_handle is null.
4047 * It's likely that its about to be closed as stale.
4049 if (conn->gnc_ephandle == NULL)
4052 timestamp = jiffies;
4053 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4054 /* delay in jiffies - we are really concerned only with things that
4055 * result in a schedule() or really holding this off for long times .
4056 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4057 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4059 /* Resample current time as we have no idea how long it took to get the mutex */
4060 timestamp = jiffies;
4062 /* We check here when the last time we received an rx, we do this before
4063 * we call getnext in case the thread has been blocked for a while. If we
4064 * havent received an rx since our timeout value we close the connection
4065 * as we should assume the other side has closed the connection. This will
4066 * stop us from sending replies to a mailbox that is already in purgatory.
4069 timeout = cfs_time_seconds(conn->gnc_timeout);
4070 newest_last_rx = GNILND_LASTRX(conn);
4072 /* Error injection to validate that timestamp checking works and closing the conn */
4073 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4074 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4077 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4078 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4079 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4080 cfs_duration_sec(timestamp - newest_last_rx),
4081 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4082 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4084 kgnilnd_close_conn(conn, rc);
4088 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4090 if (rrc == GNI_RC_NOT_DONE) {
4091 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4092 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4096 /* Instead of asserting when we get mailbox corruption lets attempt to
4097 * close the conn and recover. We can put the conn/mailbox into
4098 * purgatory and let purgatory deal with the problem. If we see
4099 * this NETTERROR reported on production systems in large amounts
4100 * we will need to revisit the state machine to see if we can tighten
4101 * it up further to improve data protection.
4104 if (rrc == GNI_RC_INVALID_STATE) {
4105 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4106 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4107 "detected closing conn %p from peer %s\n", conn,
4108 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4110 kgnilnd_close_conn(conn, rc);
4114 LASSERTF(rrc == GNI_RC_SUCCESS,
4115 "bad rc %d on conn %p from peer %s\n",
4116 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4118 msg = (kgn_msg_t *)prefix;
4120 rx = kgnilnd_alloc_rx();
4122 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4123 kgnilnd_release_msg(conn);
4124 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4125 conn, libcfs_nid2str(peer->gnp_nid));
4129 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4131 timestamp = conn->gnc_last_rx;
4132 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4133 atomic_inc(&conn->gnc_rx_seq);
4135 conn->gnc_last_rx = jiffies;
4136 /* stash first rx so we can clear out purgatory
4138 if (conn->gnc_first_rx == 0)
4139 conn->gnc_first_rx = jiffies;
4141 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4142 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4143 kgnilnd_peer_alive(conn->gnc_peer);
4146 rx->grx_conn = conn;
4148 rx->grx_received = current_kernel_time();
4150 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4153 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4159 kgnilnd_net_decref(net);
4162 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4163 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4165 /* XXX Nic: Do we need to swab cksum */
4166 if (msg->gnm_cksum != 0) {
4167 msg_cksum = msg->gnm_cksum;
4169 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4171 if (tmp_cksum != msg_cksum) {
4172 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4173 tmp_cksum, msg_cksum);
4174 kgnilnd_dump_msg(D_BUFFS, msg);
4179 /* restore checksum for future debug messages */
4180 msg->gnm_cksum = tmp_cksum;
4182 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4183 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4184 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4185 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4190 __swab32s(&msg->gnm_magic);
4191 __swab16s(&msg->gnm_version);
4192 __swab16s(&msg->gnm_type);
4193 __swab64s(&msg->gnm_srcnid);
4194 __swab64s(&msg->gnm_connstamp);
4195 __swab32s(&msg->gnm_seq);
4197 /* NB message type checked below; NOT here... */
4198 switch (msg->gnm_type) {
4199 case GNILND_MSG_GET_ACK_REV:
4200 case GNILND_MSG_PUT_ACK:
4201 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4204 case GNILND_MSG_PUT_REQ_REV:
4205 case GNILND_MSG_GET_REQ:
4206 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4214 if (msg->gnm_version != GNILND_MSG_VERSION) {
4215 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4216 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4221 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4222 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4223 libcfs_nid2str(msg->gnm_srcnid),
4224 libcfs_nid2str(peer->gnp_nid));
4229 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4230 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4231 " expected) from %s",
4232 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4233 libcfs_nid2str(peer->gnp_nid));
4238 if (msg->gnm_seq != seq) {
4239 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4240 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4245 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4247 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4248 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4249 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4250 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4251 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4252 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4253 /* double check state with lock held */
4254 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4255 /* only error if we are not already closing */
4256 if (conn->gnc_peer_error == -ETIMEDOUT) {
4257 unsigned long now = jiffies;
4258 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4260 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4261 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4262 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4264 cfs_duration_sec(now - timestamp),
4265 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4266 atomic_read(&conn->gnc_tx_seq),
4267 cfs_duration_sec(now - conn->gnc_last_tx),
4268 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4269 cfs_duration_sec(now - conn->gnc_last_noop_want),
4270 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4271 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4272 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4273 cfs_duration_sec(now - conn->gnc_last_sched_do),
4274 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4276 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4278 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4282 if (conn->gnc_close_recvd) {
4283 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4284 kgnilnd_msgtype2str(msg->gnm_type),
4285 msg->gnm_type, conn->gnc_close_recvd,
4286 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4291 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4292 /* XXX Nic: log message received on bad connection state */
4296 switch (msg->gnm_type) {
4297 case GNILND_MSG_NOOP:
4298 /* Nothing to do; just a keepalive */
4301 case GNILND_MSG_IMMEDIATE:
4302 /* only get SMSG payload for IMMEDIATE */
4303 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4304 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4305 msg->gnm_srcnid, rx, 0);
4308 case GNILND_MSG_GET_REQ_REV:
4309 case GNILND_MSG_PUT_REQ:
4310 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4311 msg->gnm_srcnid, rx, 1);
4314 case GNILND_MSG_GET_NAK_REV:
4315 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4316 msg->gnm_u.completion.gncm_cookie);
4320 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4322 case GNILND_MSG_PUT_NAK:
4323 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4324 msg->gnm_u.completion.gncm_cookie);
4328 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4330 case GNILND_MSG_PUT_ACK:
4331 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4332 msg->gnm_u.putack.gnpam_src_cookie);
4336 /* store putack data for later: deferred rdma or re-try */
4337 tx->tx_putinfo = msg->gnm_u.putack;
4340 spin_lock(&tx->tx_conn->gnc_list_lock);
4342 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4343 "not waiting for reply", NULL);
4345 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4347 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4348 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4349 /* sample under lock as follow on steps require gnc_list_lock
4350 * - or call kgnilnd_tx_done which requires no locks held over
4351 * call to lnet_finalize */
4354 /* cannot launch rdma if still waiting for fma-msg completion */
4355 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4356 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4357 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4359 spin_unlock(&tx->tx_conn->gnc_list_lock);
4362 rc = kgnilnd_send_mapped_tx(tx, 0);
4364 kgnilnd_tx_done(tx, rc);
4367 case GNILND_MSG_GET_ACK_REV:
4368 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4369 msg->gnm_u.putack.gnpam_src_cookie);
4373 /* store putack data for later: deferred rdma or re-try */
4374 tx->tx_putinfo = msg->gnm_u.putack;
4376 spin_lock(&tx->tx_conn->gnc_list_lock);
4378 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4379 "not waiting for reply", NULL);
4381 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4383 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4384 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4385 /* sample under lock as follow on steps require gnc_list_lock
4386 * - or call kgnilnd_tx_done which requires no locks held over
4387 * call to lnet_finalize */
4390 /* cannot launch rdma if still waiting for fma-msg completion */
4391 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4392 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4393 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4395 spin_unlock(&tx->tx_conn->gnc_list_lock);
4398 rc = kgnilnd_send_mapped_tx(tx, 0);
4400 kgnilnd_tx_done(tx, rc);
4403 case GNILND_MSG_PUT_DONE:
4404 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4405 msg->gnm_u.completion.gncm_cookie);
4409 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4410 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4411 "bad tx buftype %d", tx->tx_buftype);
4413 kgnilnd_finalize_rx_done(tx, msg);
4415 case GNILND_MSG_PUT_REQ_REV:
4416 case GNILND_MSG_GET_REQ:
4417 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4418 msg->gnm_srcnid, rx, 1);
4422 case GNILND_MSG_GET_NAK:
4423 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4424 msg->gnm_u.completion.gncm_cookie);
4428 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4429 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4430 "bad tx buftype %d", tx->tx_buftype);
4432 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4435 case GNILND_MSG_GET_DONE:
4436 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4437 msg->gnm_u.completion.gncm_cookie);
4441 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4442 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4443 "bad tx buftype %d", tx->tx_buftype);
4445 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4446 msg->gnm_u.completion.gncm_retval);
4448 kgnilnd_finalize_rx_done(tx, msg);
4450 case GNILND_MSG_GET_DONE_REV:
4451 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4452 msg->gnm_u.completion.gncm_cookie);
4456 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4457 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4458 "bad tx buftype %d", tx->tx_buftype);
4460 kgnilnd_finalize_rx_done(tx, msg);
4463 case GNILND_MSG_PUT_DONE_REV:
4464 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4465 msg->gnm_u.completion.gncm_cookie);
4470 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4471 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4472 "bad tx buftype %d", tx->tx_buftype);
4474 kgnilnd_finalize_rx_done(tx, msg);
4476 case GNILND_MSG_PUT_NAK_REV:
4477 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4478 msg->gnm_u.completion.gncm_cookie);
4483 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4484 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4485 "bad tx buftype %d", tx->tx_buftype);
4487 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4492 if (rc < 0) /* protocol/comms error */
4493 kgnilnd_close_conn(conn, rc);
4495 if (repost && rx != NULL) {
4496 kgnilnd_consume_rx(rx);
4499 /* we got an event so assume more there and call for reschedule */
4501 kgnilnd_schedule_conn(conn);
4505 /* Do the failure injections that we need to affect conn processing in the following function.
4506 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4507 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4510 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4514 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4518 /* short circuit out when not set */
4519 if (likely(!cfs_fail_loc)) {
4523 /* failure injection to test for stack reset clean ups */
4524 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4525 /* we can't rely on busy loops being nice enough to get the
4526 * stack reset triggered - it'd just spin on this conn */
4527 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4530 GOTO(did_fail_loc, rc);
4533 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4534 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4536 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4537 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4540 GOTO(did_fail_loc, rc);
4544 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4545 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4546 * will be closed by kgnilnd_complete_closed_conn.
4548 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4549 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4552 GOTO(did_fail_loc, rc);
4555 /* this one is a bit gross - we can't hold the mutex from process_conns
4556 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4557 * and moving onto finish_connect
4558 * so, we'll just set the rc - kgnilnd_process_conns will clear
4559 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4560 * and effectively getting this thread to sleep */
4561 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4564 GOTO(did_fail_loc, rc);
4572 kgnilnd_send_conn_close(kgn_conn_t *conn)
4576 /* we are closing the conn - we will try to send the CLOSE msg
4577 * but will not wait for anything else to flush */
4579 /* send the close if not already done so or received one */
4580 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4581 /* set close_sent regardless of the success of the
4582 * CLOSE message. We are going to try once and then
4583 * kick him out of the sandbox */
4584 conn->gnc_close_sent = 1;
4587 /* EP might be null already if remote side initiated a new connection.
4588 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4589 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4591 if (conn->gnc_ephandle != NULL) {
4594 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4596 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4597 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4598 tx->tx_qtime = jiffies;
4600 if (tx->tx_id.txe_idx == 0) {
4601 rc = kgnilnd_set_tx_id(tx, conn);
4603 kgnilnd_tx_done(tx, rc);
4607 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4610 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4611 kgnilnd_tx_done(tx, -EAGAIN);
4613 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4615 /* It wasnt sent and we dont care. */
4616 kgnilnd_tx_done(tx, rc);
4624 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4625 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4626 conn->gnc_state = GNILND_CONN_CLOSED;
4627 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4628 /* mark this conn as CLOSED now that we processed it
4629 * do after TX, so we can use CLOSING in asserts */
4633 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4634 /* simulate a RX CLOSE after the timeout but before
4635 * the scheduler thread gets it */
4636 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4637 conn->gnc_peer_error = -ETIMEDOUT;
4639 /* schedule to allow potential CLOSE and get the complete phase run */
4640 kgnilnd_schedule_conn(conn);
4644 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4649 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4650 int log_retrans, log_retrans_level;
4651 static int last_map_version;
4654 spin_lock(&dev->gnd_lock);
4655 if (list_empty(&dev->gnd_map_tx)) {
4656 /* if the list is empty make sure we dont have a timer running */
4657 del_singleshot_timer_sync(&dev->gnd_map_timer);
4658 spin_unlock(&dev->gnd_lock);
4662 dev->gnd_sched_alive = jiffies;
4664 /* we'll retry as fast as possible up to 25% of the limit, then we start
4665 * backing off until our map version changes - indicating we unmapped
4667 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4668 if (likely(dev->gnd_map_attempt == 0) ||
4669 time_after_eq(jiffies, dev->gnd_next_map) ||
4670 last_map_version != dev->gnd_map_version) {
4672 /* if this is our first attempt at mapping set last mapped to current
4673 * jiffies so we can timeout our attempt correctly.
4675 if (dev->gnd_map_attempt == 0)
4676 dev->gnd_last_map = jiffies;
4678 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4679 spin_unlock(&dev->gnd_lock);
4683 /* delete the previous timer if it exists */
4684 del_singleshot_timer_sync(&dev->gnd_map_timer);
4685 /* stash the last map version to let us know when a good one was seen */
4686 last_map_version = dev->gnd_map_version;
4688 /* we need to to take the lock and continually refresh the head of the list as
4689 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4690 * allowing them to squeeze in */
4692 while (!list_empty(&dev->gnd_map_tx)) {
4693 /* make sure we break out early on quiesce */
4694 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4695 /* always break with lock held - we unlock outside loop */
4699 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4701 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4704 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4705 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4706 /* if conn is dying, mark tx in tx_ref_table for
4707 * kgnilnd_complete_closed_conn to finish up */
4708 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4711 /* tx was moved to DYING, get next */
4715 spin_unlock(&dev->gnd_lock);
4716 rc = kgnilnd_send_mapped_tx(tx, 1);
4718 /* We made it! skip error handling.. */
4720 /* OK to continue on +ve errors as it won't get seen until
4721 * this function is called again - we operate on a copy of the original
4722 * list and not the live list */
4723 spin_lock(&dev->gnd_lock);
4724 /* reset map attempts back to zero we successfully
4725 * mapped so we can reset our timers */
4726 dev->gnd_map_attempt = 0;
4728 } else if (rc == -EAGAIN) {
4729 spin_lock(&dev->gnd_lock);
4730 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4731 spin_unlock(&dev->gnd_lock);
4732 GOTO(get_out_mapped, rc);
4733 } else if (rc != -ENOMEM) {
4734 /* carp, failure we can't handle */
4735 kgnilnd_tx_done(tx, rc);
4736 spin_lock(&dev->gnd_lock);
4737 /* reset map attempts back to zero we dont know what happened but it
4738 * wasnt a failed mapping
4740 dev->gnd_map_attempt = 0;
4744 /* time to handle the retry cases.. lock so we dont have 2 threads
4745 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4747 spin_lock(&dev->gnd_lock);
4748 dev->gnd_map_attempt++;
4749 if (dev->gnd_map_attempt < fast_remaps) {
4750 /* do nothing we just want it to go as fast as possible.
4751 * just set gnd_next_map to current jiffies so it will process
4752 * as fast as possible.
4754 dev->gnd_next_map = jiffies;
4756 /* Retry based on GNILND_MAP_RETRY_RATE */
4757 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4760 /* only log occasionally once we've retried fast_remaps */
4761 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4762 ((dev->gnd_map_attempt % fast_remaps) == 0);
4763 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4765 /* make sure we are not off in the weeds with this tx */
4766 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4767 GNIDBG_TX(D_NETERROR, tx,
4768 "giving up on TX, too many retries", NULL);
4769 spin_unlock(&dev->gnd_lock);
4770 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4771 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4772 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4774 tx->tx_putinfo.gnpam_dst_cookie,
4775 tx->tx_msg.gnm_srcnid);
4777 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4779 tx->tx_getinfo.gngm_cookie,
4780 tx->tx_msg.gnm_srcnid);
4782 kgnilnd_tx_done(tx, -ENOMEM);
4783 GOTO(get_out_mapped, rc);
4785 GNIDBG_TX(log_retrans_level, tx,
4786 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4788 "nq_map %d mdd# %d/%d GART %ld",
4789 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4790 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4791 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4792 atomic_read(&dev->gnd_nq_map),
4793 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4794 atomic64_read(&dev->gnd_nbytes_map));
4797 /* we need to stop processing the rest of the list, so add it back in */
4798 /* set timer to wake device when we need to schedule this tx */
4799 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4800 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4801 spin_unlock(&dev->gnd_lock);
4802 GOTO(get_out_mapped, rc);
4804 spin_unlock(&dev->gnd_lock);
4810 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4815 int error_inject = 0;
4819 spin_lock(&dev->gnd_lock);
4820 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4821 dev->gnd_sched_alive = jiffies;
4825 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4826 /* break with lock held */
4830 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4831 list_del_init(&conn->gnc_schedlist);
4833 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4836 if (!list_empty(&conn->gnc_delaylist))
4837 list_del_init(&conn->gnc_delaylist);
4838 spin_unlock(&dev->gnd_lock);
4840 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4842 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4843 conn_sched != GNILND_CONN_PROCESS,
4844 "conn %p on ready list but in bad state: %d\n",
4847 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4848 conn, kgnilnd_conn_state2str(conn));
4851 set_mb(conn->gnc_last_sched_do, jiffies);
4853 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4855 /* based on intent see if we should run again. */
4856 rc = kgnilnd_schedule_process_conn(conn, intent);
4858 /* drop ref from gnd_ready_conns */
4859 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4860 down_write(&dev->gnd_conn_sem);
4861 kgnilnd_conn_decref(conn);
4862 up_write(&dev->gnd_conn_sem);
4863 } else if (rc != 1) {
4864 kgnilnd_conn_decref(conn);
4866 /* clear this so that scheduler thread doesn't spin */
4868 /* break with lock held... */
4869 spin_lock(&dev->gnd_lock);
4873 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4874 down_write(&dev->gnd_conn_sem);
4876 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4877 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4878 /* If there are tx's currently in use in another
4879 * thread we dont want to complete the close
4880 * yet. Cycle this conn back through
4882 kgnilnd_schedule_conn(conn);
4884 kgnilnd_complete_closed_conn(conn);
4886 up_write(&dev->gnd_conn_sem);
4887 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4888 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4889 /* serialize SMSG CQs with ep_bind and smsg_release */
4890 down_write(&dev->gnd_conn_sem);
4891 kgnilnd_destroy_conn_ep(conn);
4892 up_write(&dev->gnd_conn_sem);
4893 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4894 /* if we need to do some CLOSE sending, etc done here do it */
4895 down_write(&dev->gnd_conn_sem);
4896 kgnilnd_send_conn_close(conn);
4897 kgnilnd_check_fma_rx(conn);
4898 up_write(&dev->gnd_conn_sem);
4899 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4900 /* start moving traffic if the old conns are cleared out */
4901 down_read(&dev->gnd_conn_sem);
4902 kgnilnd_check_fma_rx(conn);
4903 kgnilnd_process_fmaq(conn);
4904 up_read(&dev->gnd_conn_sem);
4907 rc = kgnilnd_schedule_process_conn(conn, 0);
4909 /* drop ref from gnd_ready_conns */
4910 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4911 down_write(&dev->gnd_conn_sem);
4912 kgnilnd_conn_decref(conn);
4913 up_write(&dev->gnd_conn_sem);
4914 } else if (rc != 1) {
4915 kgnilnd_conn_decref(conn);
4918 /* check list again with lock held */
4919 spin_lock(&dev->gnd_lock);
4922 /* If we are short circuiting due to timing we want to be scheduled
4923 * as soon as possible.
4925 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4928 spin_unlock(&dev->gnd_lock);
4934 kgnilnd_scheduler(void *arg)
4936 int threadno = (long)arg;
4939 unsigned long deadline = 0;
4942 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
4944 /* all gnilnd threads need to run fairly urgently */
4945 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
4946 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
4947 while (!kgnilnd_data.kgn_shutdown) {
4949 /* Safe: kgn_shutdown only set when quiescent */
4951 /* to quiesce or to not quiesce, that is the question */
4953 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4954 KGNILND_SPIN_QUIESCE;
4957 /* tracking for when thread goes AWOL */
4958 dev->gnd_sched_alive = jiffies;
4960 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
4961 (*kgnilnd_tunables.kgn_sched_timeout + 1));
4962 /* let folks know we are up and kicking
4963 * - they can use this for latency savings, etc
4964 * - only change if IRQ, if IDLE leave alone as that
4965 * schedule_device calls to put us back to IRQ */
4966 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
4968 down_read(&dev->gnd_conn_sem);
4969 /* always check these - they are super low cost */
4970 found_work += kgnilnd_check_fma_send_cq(dev);
4971 found_work += kgnilnd_check_fma_rcv_cq(dev);
4973 /* rdma CQ doesn't care about eps */
4974 found_work += kgnilnd_check_rdma_cq(dev);
4976 /* move some RDMA ? */
4977 found_work += kgnilnd_process_rdmaq(dev);
4979 /* map some pending RDMA requests ? */
4980 found_work += kgnilnd_process_mapped_tx(dev);
4982 /* the EP for a conn is not destroyed until all the references
4983 * to it are gone, so these checks should be safe
4984 * even if run in parallel with the CQ checking functions
4985 * _AND_ a thread that processes the CLOSED->DONE
4989 up_read(&dev->gnd_conn_sem);
4991 /* process all conns ready now */
4992 found_work += kgnilnd_process_conns(dev, deadline);
4994 /* do an eager check to avoid the IRQ disabling in
4995 * prepare_to_wait and friends */
4998 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
4999 time_before(jiffies, deadline)) {
5001 if ((busy_loops % 10) == 0) {
5002 /* tickle heartbeat and watchdog to ensure our
5003 * piggishness doesn't turn into heartbeat failure */
5004 touch_nmi_watchdog();
5010 /* if we got here, found_work was zero or busy_loops means we
5011 * need to take a break. We'll clear gnd_ready but we'll check
5012 * one last time if there is an IRQ that needs processing */
5014 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5016 /* the first time this will go LOOP -> IDLE and let us do one final check
5017 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5018 * - this might allow other threads to block us for a bit if they
5019 * try to get the mutex, but that is good as we'd need to wake
5020 * up soon to handle the CQ or other processing anyways */
5022 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5024 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5025 time_after_eq(jiffies, deadline)) {
5027 "yeilding: found_work %d busy_loops %d\n",
5028 found_work, busy_loops);
5030 /* use yield if we are bailing due to busy_loops
5031 * - this will ensure we wake up soonish. This closes
5032 * a race with kgnilnd_device_callback - where it'd
5033 * not call wake_up() because gnd_ready == 1, but then
5034 * we come down and schedule() because of busy_loops.
5035 * We'd not be woken up until something poked our waitq
5036 * again. yield() ensures we wake up without another
5037 * waitq poke in that case */
5038 atomic_inc(&dev->gnd_n_yield);
5039 kgnilnd_data.kgn_last_condresched = jiffies;
5041 CDEBUG(D_INFO, "awake after yeild\n");
5042 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5043 } else if (found_work == GNILND_DEV_IDLE) {
5044 /* busy_loops is low and there is nothing to do,
5045 * go to sleep and wait for a waitq poke */
5047 "scheduling: found_work %d busy_loops %d\n",
5048 found_work, busy_loops);
5049 atomic_inc(&dev->gnd_n_schedule);
5050 kgnilnd_data.kgn_last_scheduled = jiffies;
5052 CDEBUG(D_INFO, "awake after schedule\n");
5053 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5055 finish_wait(&dev->gnd_waitq, &wait);
5058 kgnilnd_thread_fini();