2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by Eric Barton <eric@bartonsoftware.com>
7 * Author: James Shimek <jshimek@cray.com>
8 * Author: Nic Henke <nic@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #include <linux/nmi.h>
31 /* this is useful when needed to debug wire corruption. */
33 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
41 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
42 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
45 } else if (len >= 16) {
47 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
48 prefix, ptr, *(ptr), *(ptr + 1));
52 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
61 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
63 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
64 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
65 msg->gnm_magic, msg->gnm_version,
66 msg->gnm_type, msg->gnm_srcnid,
67 msg->gnm_connstamp, msg->gnm_seq,
68 msg->gnm_cksum, msg->gnm_payload_cksum,
69 msg->gnm_payload_len);
73 kgnilnd_schedule_device(kgn_device_t *dev)
75 short already_live = 0;
77 /* we'll only want to wake if the scheduler thread
78 * has come around and set ready to zero */
79 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
82 wake_up_all(&dev->gnd_waitq);
87 void kgnilnd_schedule_device_timer(unsigned long arg)
89 kgn_device_t *dev = (kgn_device_t *) arg;
91 kgnilnd_schedule_device(dev);
95 kgnilnd_device_callback(__u32 devid, __u64 arg)
98 int index = (int) arg;
100 if (index >= kgnilnd_data.kgn_ndevs) {
101 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
102 * interrupt context. */
103 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
108 dev = &kgnilnd_data.kgn_devices[index];
109 /* just basic sanity */
110 if (dev->gnd_id == devid) {
111 kgnilnd_schedule_device(dev);
113 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
118 /* sched_intent values:
119 * < 0 : do not reschedule under any circumstances
120 * == 0: reschedule if someone marked him WANTS_SCHED
121 * > 0 : force a reschedule */
122 /* Return code 0 means it did not schedule the conn, 1
123 * means it successfully scheduled the conn.
127 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
131 /* move back to IDLE but save previous state.
132 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
133 * let the xchg there handle any racing callers to get it
134 * onto gnd_ready_conns */
136 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
137 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
138 conn_sched == GNILND_CONN_PROCESS,
139 "conn %p after process in bad state: %d\n",
142 if (sched_intent >= 0) {
143 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
144 return kgnilnd_schedule_conn_refheld(conn, 1);
150 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
154 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld)
156 kgn_device_t *dev = conn->gnc_device;
160 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
161 /* we only care about the last person who marked want_sched since they
162 * are most likely the culprit
164 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
165 conn->gnc_sched_line = line;
166 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
167 * the chance to put it onto gnd_ready_conns.
168 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
169 * the conn in process_conns will take care of moving it back to
170 * SCHED when it is done processing */
172 if (sched == GNILND_CONN_IDLE) {
173 /* if the conn is already scheduled, we've already requested
174 * the scheduler thread wakeup */
176 /* Add a reference to the conn if we are not holding a reference
177 * already from the exisiting scheduler. We now use the same
178 * reference if we need to reschedule a conn while in a scheduler
181 kgnilnd_conn_addref(conn);
183 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
186 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
188 spin_lock(&dev->gnd_lock);
189 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
190 spin_unlock(&dev->gnd_lock);
191 set_mb(conn->gnc_last_sched_ask, jiffies);
194 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
198 /* make sure thread(s) going to process conns - but let it make
199 * separate decision from conn schedule */
200 kgnilnd_schedule_device(dev);
205 kgnilnd_schedule_dgram(kgn_device_t *dev)
209 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
210 if (wake != GNILND_DGRAM_SCHED) {
211 wake_up(&dev->gnd_dgram_waitq);
213 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
218 kgnilnd_free_tx(kgn_tx_t *tx)
220 /* taken from kgnilnd_tx_add_state_locked */
222 LASSERTF((tx->tx_list_p == NULL &&
223 tx->tx_list_state == GNILND_TX_ALLOCD) &&
224 list_empty(&tx->tx_list),
225 "tx %p with bad state %s (list_p %p) tx_list %s\n",
226 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
227 list_empty(&tx->tx_list) ? "empty" : "not empty");
229 atomic_dec(&kgnilnd_data.kgn_ntx);
231 /* we only allocate this if we need to */
232 if (tx->tx_phys != NULL) {
233 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
234 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
235 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
238 /* Only free the buffer if we used it */
239 if (tx->tx_buffer_copy != NULL) {
240 vfree(tx->tx_buffer_copy);
241 tx->tx_buffer_copy = NULL;
242 CDEBUG(D_MALLOC, "vfreed buffer2\n");
245 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
247 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
248 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
252 kgnilnd_alloc_tx (void)
256 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
259 tx = kmem_cache_alloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
261 CERROR("failed to allocate tx\n");
264 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
267 /* need this memset, cache alloc'd memory is not cleared */
268 memset(tx, 0, sizeof(*tx));
270 /* setup everything here to minimize time under the lock */
271 tx->tx_buftype = GNILND_BUF_NONE;
272 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
273 INIT_LIST_HEAD(&tx->tx_list);
274 INIT_LIST_HEAD(&tx->tx_map_list);
275 tx->tx_list_state = GNILND_TX_ALLOCD;
277 atomic_inc(&kgnilnd_data.kgn_ntx);
282 /* csum_fold needs to be run on the return value before shipping over the wire */
283 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
285 /* we don't use offset as every one is passing a buffer reference that already
286 * includes the offset into the base address -
287 * see kgnilnd_setup_virt_buffer and kgnilnd_setup_immediate_buffer */
289 kgnilnd_cksum(void *ptr, size_t nob)
293 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
295 /* don't use magic 'no checksum' value */
299 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
306 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov,
307 unsigned int offset, unsigned int nob, int dump_blob)
313 unsigned int fraglen;
319 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
320 kiov, nkiov, offset, nob, dump_blob);
322 /* if loops changes, please change kgnilnd_setup_phys_buffer */
324 while (offset >= kiov->kiov_len) {
325 offset -= kiov->kiov_len;
331 /* ignore nob here, if nob < (kiov_len - offset), kiov == 1 */
332 odd = (unsigned long) (kiov[0].kiov_len - offset) & 1;
334 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
335 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
337 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
338 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
340 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
341 odd, kiov[0].kiov_len, offset, nob);
343 for (i = 0; i < nkiov; i++) {
344 pages[i] = kiov[i].kiov_page;
347 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
349 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
351 /* return zero to avoid killing tx - we'll just get warning on console
352 * when remote end sees zero checksum */
355 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
357 tmpck = _kgnilnd_cksum(0, (void *) addr + kiov[0].kiov_offset + offset, nob);
361 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
362 (void *)addr + kiov[0].kiov_offset + offset, nob);
364 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
365 cksum, tmpck, addr, kiov[0].kiov_offset, nob, offset);
369 fraglen = min(kiov->kiov_len - offset, nob);
371 /* make dang sure we don't send a bogus checksum if somehow we get
372 * an odd length fragment on anything but the last entry in a kiov -
373 * we know from kgnilnd_setup_rdma_buffer that we can't have non
374 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
375 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
376 "odd fraglen %u on nkiov %d, nob %u kiov_len %u offset %u kiov 0x%p\n",
377 fraglen, nkiov, nob, kiov->kiov_len, offset, kiov);
379 addr = (void *)kmap(kiov->kiov_page) + kiov->kiov_offset + offset;
380 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
383 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
384 cksum, tmpck, kiov->kiov_page, kiov->kiov_offset, addr,
390 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
392 kunmap(kiov->kiov_page);
399 /* iov must not run out before end of data */
400 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
405 retsum = csum_fold(cksum);
407 /* don't use magic 'no checksum' value */
411 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
417 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
419 msg->gnm_magic = GNILND_MSG_MAGIC;
420 msg->gnm_version = GNILND_MSG_VERSION;
421 msg->gnm_type = type;
422 msg->gnm_payload_len = 0;
423 msg->gnm_srcnid = source;
424 /* gnm_connstamp gets set when FMA is sent */
425 /* gnm_srcnid is set on creation via function argument
426 * The right interface/net and nid is passed in when the message
432 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
434 kgn_tx_t *tx = kgnilnd_alloc_tx();
437 kgnilnd_init_msg(&tx->tx_msg, type, source);
439 CERROR("couldn't allocate new tx type %s!\n",
440 kgnilnd_msgtype2str(type));
447 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
453 case GNILND_MSG_GET_REQ:
454 case GNILND_MSG_GET_DONE:
455 nak_type = GNILND_MSG_GET_NAK;
457 case GNILND_MSG_PUT_REQ:
458 case GNILND_MSG_PUT_ACK:
459 case GNILND_MSG_PUT_DONE:
460 nak_type = GNILND_MSG_PUT_NAK;
462 case GNILND_MSG_PUT_REQ_REV:
463 case GNILND_MSG_PUT_DONE_REV:
464 nak_type = GNILND_MSG_PUT_NAK_REV;
466 case GNILND_MSG_GET_REQ_REV:
467 case GNILND_MSG_GET_ACK_REV:
468 case GNILND_MSG_GET_DONE_REV:
469 nak_type = GNILND_MSG_GET_NAK_REV;
472 CERROR("invalid msg type %s (%d)\n",
473 kgnilnd_msgtype2str(rx_type), rx_type);
476 /* only allow NAK on error and truncate to zero */
477 LASSERTF(error <= 0, "error %d conn 0x%p, cookie "LPU64"\n",
478 error, conn, cookie);
480 tx = kgnilnd_new_tx_msg(nak_type, source);
482 CNETERR("can't get TX to NAK RDMA to %s\n",
483 libcfs_nid2str(conn->gnc_peer->gnp_nid));
487 tx->tx_msg.gnm_u.completion.gncm_retval = error;
488 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
489 kgnilnd_queue_tx(conn, tx);
493 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *iov,
494 lnet_kiov_t *kiov, unsigned int offset, unsigned int nob)
497 kgn_msg_t *msg = &tx->tx_msg;
500 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
501 * gni_smsg_send to send that as the payload */
503 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
506 tx->tx_buffer = NULL;
507 } else if (kiov != NULL) {
509 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
510 niov = ((nob + offset + PAGE_SIZE - 1) / PAGE_SIZE);
513 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
514 "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n",
515 niov, msg, kiov, iov, offset, nob);
517 while (offset >= kiov->kiov_len) {
518 offset -= kiov->kiov_len;
523 for (i = 0; i < niov; i++) {
524 /* We can't have a kiov_offset on anything but the first entry,
525 * otherwise we'll have a hole at the end of the mapping as we only map
527 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
528 * than kiov_len, we will also have a whole at the end of that page
529 * which isn't allowed */
530 if ((kiov[i].kiov_offset != 0 && i > 0) ||
531 (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_SIZE && i < niov - 1)) {
532 CNETERR("Can't make payload contiguous in I/O VM:"
533 "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
534 i, offset, nob, kiov->kiov_offset, kiov->kiov_len);
537 tx->tx_imm_pages[i] = kiov[i].kiov_page;
540 /* hijack tx_phys for the later unmap */
542 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
544 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) + kiov[0].kiov_offset + offset;
545 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
546 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
547 nob, kiov, tx->tx_buffer);
549 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
550 if (tx->tx_phys == NULL) {
551 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
555 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
556 /* make sure we take into account the kiov offset as the start of the buffer */
557 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].kiov_offset + offset;
558 GNIDBG_TX(D_NET, tx, "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
559 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
561 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
565 /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
566 * could "flatten" the payload into a single contiguous buffer ready
567 * for sending direct over an FMA if we ever needed to. */
571 while (offset >= iov->iov_len) {
572 offset -= iov->iov_len;
578 if (nob > iov->iov_len - offset) {
579 CERROR("Can't handle multiple vaddr fragments\n");
583 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
585 tx->tx_buftype = GNILND_BUF_IMMEDIATE;
589 /* checksum payload early - it shouldn't be changing after lnd_send */
590 if (*kgnilnd_tunables.kgn_checksum >= 2) {
591 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
592 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
593 msg->gnm_payload_cksum += 0xe00e;
595 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
596 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
600 msg->gnm_payload_cksum = 0;
607 kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
608 unsigned int niov, struct iovec *iov,
609 unsigned int offset, unsigned int nob)
614 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
616 while (offset >= iov->iov_len) {
617 offset -= iov->iov_len;
623 if (nob > iov->iov_len - offset) {
624 CERROR("Can't handle multiple vaddr fragments\n");
628 tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
630 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
635 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
636 unsigned int offset, unsigned int nob)
638 gni_mem_segment_t *phys;
640 unsigned int fraglen;
642 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
646 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
648 /* only allocate this if we are going to use it */
649 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
651 if (tx->tx_phys == NULL) {
652 CERROR("failed to allocate tx_phys\n");
657 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
658 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
660 /* if loops changes, please change kgnilnd_cksum_kiov
661 * and kgnilnd_setup_immediate_buffer */
663 while (offset >= kiov->kiov_len) {
664 offset -= kiov->kiov_len;
670 /* at this point, kiov points to the first page that we'll actually map
671 * now that we've seeked into the koiv for offset and dropped any
672 * leading pages that fall entirely within the offset */
673 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
676 /* kiov_offset is start of 'valid' buffer, so index offset past that */
677 tx->tx_buffer = (void *)((unsigned long)(kiov->kiov_offset + offset));
680 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
681 tx, tx->tx_buffer, kiov, kiov->kiov_offset, nkiov, offset);
684 fraglen = min(kiov->kiov_len - offset, nob);
686 /* We can't have a kiov_offset on anything but the first entry,
687 * otherwise we'll have a hole at the end of the mapping as we only map
688 * whole pages. Only the first page is allowed to have an offset -
689 * we'll add that into tx->tx_buffer and that will get used when we
690 * map in the segments (see kgnilnd_map_buffer).
691 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
692 * than kiov_len, we will also have a whole at the end of that page
693 * which isn't allowed */
694 if ((phys != tx->tx_phys) &&
695 ((kiov->kiov_offset != 0) ||
696 ((kiov->kiov_len < PAGE_SIZE) && (nob > kiov->kiov_len)))) {
697 CERROR("Can't make payload contiguous in I/O VM:"
698 "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
699 (int)(phys - tx->tx_phys),
700 offset, nob, kiov->kiov_offset, kiov->kiov_len);
705 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
706 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
711 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
716 CDEBUG(D_BUFFS, "page 0x%p kiov_offset %u kiov_len %u nob %u "
717 "nkiov %u offset %u\n",
718 kiov->kiov_page, kiov->kiov_offset, kiov->kiov_len, nob, nkiov, offset);
720 phys->address = page_to_phys(kiov->kiov_page);
727 /* iov must not run out before end of data */
728 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
732 tx->tx_phys_npages = phys - tx->tx_phys;
737 if (tx->tx_phys != NULL) {
738 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
739 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
740 sizeof(*tx->tx_phys), tx->tx_phys);
747 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
748 struct iovec *iov, lnet_kiov_t *kiov,
749 unsigned int offset, unsigned int nob)
753 LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
754 " offset %d, nob %d, niov %d\n"
755 , iov, kiov, tx, offset, nob, niov);
758 rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
760 rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
765 /* kgnilnd_parse_lnet_rdma()
766 * lntmsg - message passed in from lnet.
767 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
768 * nob - actual number of bytes to in this message.
769 * put_len - It is possible for PUTs to have a different length than the
770 * length stored in lntmsg->msg_len since LNET can adjust this
771 * length based on it's buffer size and offset.
772 * lnet_try_match_md() sets the mlength that we use to do the RDMA
776 kgnilnd_parse_lnet_rdma(lnet_msg_t *lntmsg, unsigned int *niov,
777 unsigned int *offset, unsigned int *nob,
778 lnet_kiov_t **kiov, int put_len)
780 /* GETs are weird, see kgnilnd_send */
781 if (lntmsg->msg_type == LNET_MSG_GET) {
782 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
785 *kiov = lntmsg->msg_md->md_iov.kiov;
787 *niov = lntmsg->msg_md->md_niov;
788 *nob = lntmsg->msg_md->md_length;
791 *kiov = lntmsg->msg_kiov;
792 *niov = lntmsg->msg_niov;
794 *offset = lntmsg->msg_offset;
799 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
801 unsigned int niov, offset, nob;
803 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
804 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
806 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
807 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
808 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
809 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
810 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
811 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
812 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
814 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
815 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
816 tx->tx_msg.gnm_payload_cksum = 0;
819 if (*kgnilnd_tunables.kgn_checksum < 3) {
820 tx->tx_msg.gnm_payload_cksum = 0;
824 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
826 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
830 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
832 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
834 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
838 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
839 tx->tx_msg.gnm_payload_cksum += 0xd00d;
843 /* kgnilnd_verify_rdma_cksum()
844 * tx - PUT_DONE/GET_DONE matched tx.
845 * rx_cksum - received checksum to compare against.
846 * put_len - see kgnilnd_parse_lnet_rdma comments.
849 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
853 unsigned int niov, offset, nob;
855 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
856 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
858 /* we can only match certain requests */
859 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
860 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
861 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
862 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
863 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
864 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
865 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
867 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
868 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
873 if (*kgnilnd_tunables.kgn_checksum >= 3) {
874 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
875 "no RDMA payload checksum when enabled");
880 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
882 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
885 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
887 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
890 if (cksum != rx_cksum) {
891 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
892 "Bad RDMA payload checksum (%x expected %x); "
893 "kiov 0x%p niov %d nob %u offset %u",
894 cksum, rx_cksum, kiov, niov, nob, offset);
895 switch (dump_on_err) {
898 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
900 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
903 /* fall through to dump log */
905 libcfs_debug_dumplog();
911 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
917 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
921 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
922 "already mapped!", NULL);
924 spin_lock(&dev->gnd_map_lock);
925 switch (tx->tx_buftype) {
927 GNIDBG_TX(D_EMERG, tx,
928 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
929 spin_unlock(&dev->gnd_map_lock);
933 case GNILND_BUF_PHYS_MAPPED:
934 bytes = tx->tx_phys_npages * PAGE_SIZE;
935 dev->gnd_map_nphys++;
936 dev->gnd_map_physnop += tx->tx_phys_npages;
939 case GNILND_BUF_VIRT_MAPPED:
941 dev->gnd_map_nvirt++;
942 dev->gnd_map_virtnob += tx->tx_nob;
946 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
947 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
948 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
949 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to "LPD64"",
950 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
953 atomic_inc(&dev->gnd_n_mdd);
954 atomic64_add(bytes, &dev->gnd_nbytes_map);
956 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
959 /* we only get here in the valid cases */
960 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
961 dev->gnd_map_version++;
962 spin_unlock(&dev->gnd_map_lock);
966 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
970 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
971 "not mapped!", NULL);
972 spin_lock(&dev->gnd_map_lock);
974 switch (tx->tx_buftype) {
976 GNIDBG_TX(D_EMERG, tx,
977 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
978 spin_unlock(&dev->gnd_map_lock);
982 case GNILND_BUF_PHYS_UNMAPPED:
983 bytes = tx->tx_phys_npages * PAGE_SIZE;
984 dev->gnd_map_nphys--;
985 dev->gnd_map_physnop -= tx->tx_phys_npages;
988 case GNILND_BUF_VIRT_UNMAPPED:
990 dev->gnd_map_nvirt--;
991 dev->gnd_map_virtnob -= tx->tx_nob;
995 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
996 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
997 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
998 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
999 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
1000 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to "LPD64"",
1001 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1004 atomic_dec(&dev->gnd_n_mdd);
1005 atomic64_sub(bytes, &dev->gnd_nbytes_map);
1007 /* we only get here in the valid cases */
1008 list_del_init(&tx->tx_map_list);
1009 dev->gnd_map_version++;
1010 spin_unlock(&dev->gnd_map_lock);
1014 kgnilnd_map_buffer(kgn_tx_t *tx)
1016 kgn_conn_t *conn = tx->tx_conn;
1017 kgn_device_t *dev = conn->gnc_device;
1018 __u32 flags = GNI_MEM_READWRITE;
1021 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
1022 * be called concurrently as there are internal locks that protect
1023 * any data structures or HW resources. We just need to ensure
1024 * that our concurrency doesn't result in the kgn_device_t
1025 * getting nuked while we are in here */
1027 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
1028 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
1030 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1033 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1034 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1037 switch (tx->tx_buftype) {
1041 case GNILND_BUF_NONE:
1042 case GNILND_BUF_IMMEDIATE:
1043 case GNILND_BUF_IMMEDIATE_KIOV:
1044 case GNILND_BUF_PHYS_MAPPED:
1045 case GNILND_BUF_VIRT_MAPPED:
1048 case GNILND_BUF_PHYS_UNMAPPED:
1049 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1050 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1051 tx->tx_phys, tx->tx_phys_npages, NULL,
1052 GNI_MEM_PHYS_SEGMENTS | flags,
1054 /* could race with other uses of the map counts, but this is ok
1055 * - this needs to turn into a non-fatal error soon to allow
1056 * GART resource, etc starvation handling */
1057 if (rrc != GNI_RC_SUCCESS) {
1058 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1059 "phys %u pp %u, virt %u nob "LPU64"",
1060 tx->tx_phys_npages, dev->gnd_id,
1061 dev->gnd_map_nphys, dev->gnd_map_physnop,
1062 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1063 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1066 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1067 kgnilnd_mem_add_map_list(dev, tx);
1070 case GNILND_BUF_VIRT_UNMAPPED:
1071 rrc = kgnilnd_mem_register(dev->gnd_handle,
1072 (__u64)tx->tx_buffer, tx->tx_nob,
1073 NULL, flags, &tx->tx_map_key);
1074 if (rrc != GNI_RC_SUCCESS) {
1075 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1076 "phys %u pp %u, virt %u nob "LPU64"",
1077 tx->tx_nob, dev->gnd_id,
1078 dev->gnd_map_nphys, dev->gnd_map_physnop,
1079 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1080 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1083 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1084 kgnilnd_mem_add_map_list(dev, tx);
1085 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1086 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1087 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1088 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1089 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1097 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1099 kgn_conn_t *conn = tx->tx_conn;
1100 kgn_mdd_purgatory_t *gmp;
1102 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1103 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1104 " asserting to avoid data corruption\n");
1105 if (tx->tx_buffer_copy)
1106 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1108 gmp->gmp_map_key = tx->tx_map_key;
1110 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1112 /* ensure that we don't have a blank purgatory - indicating the
1113 * conn is not already on purgatory lists - we'd never recover these
1114 * MDD if that were the case */
1115 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1116 "conn 0x%p->%s with NULL purgatory",
1117 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1119 /* link 'er up! - only place we really need to lock for
1120 * concurrent access */
1121 spin_lock(&conn->gnc_list_lock);
1122 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1123 spin_unlock(&conn->gnc_list_lock);
1127 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1131 int hold_timeout = 0;
1133 /* code below relies on +1 relationship ... */
1134 CLASSERT(GNILND_BUF_PHYS_MAPPED == (GNILND_BUF_PHYS_UNMAPPED + 1));
1135 CLASSERT(GNILND_BUF_VIRT_MAPPED == (GNILND_BUF_VIRT_UNMAPPED + 1));
1137 switch (tx->tx_buftype) {
1141 case GNILND_BUF_NONE:
1142 case GNILND_BUF_IMMEDIATE:
1143 case GNILND_BUF_PHYS_UNMAPPED:
1144 case GNILND_BUF_VIRT_UNMAPPED:
1146 case GNILND_BUF_IMMEDIATE_KIOV:
1147 if (tx->tx_phys != NULL) {
1148 vunmap(tx->tx_phys);
1149 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1150 kunmap(tx->tx_imm_pages[0]);
1152 /* clear to prevent kgnilnd_free_tx from thinking
1153 * this is a RDMA descriptor */
1157 case GNILND_BUF_PHYS_MAPPED:
1158 case GNILND_BUF_VIRT_MAPPED:
1159 LASSERT(tx->tx_conn != NULL);
1161 dev = tx->tx_conn->gnc_device;
1163 /* only want to hold if we are closing conn without
1164 * verified peer notification - the theory is that
1165 * a TX error can be communicated in all other cases */
1166 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1167 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1168 kgnilnd_add_purgatory_tx(tx);
1170 /* The timeout we give to kgni is a deadman stop only.
1171 * we are setting high to ensure we don't have the kgni timer
1172 * fire before ours fires _and_ is handled */
1173 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1175 GNIDBG_TX(D_NET, tx,
1176 "dev %p delaying MDD release for %dms key "LPX64"."LPX64"",
1177 tx->tx_conn->gnc_device, hold_timeout,
1178 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1180 if (tx->tx_buffer_copy != NULL) {
1181 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1182 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1183 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1184 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1186 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1187 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1191 kgnilnd_mem_del_map_list(dev, tx);
1197 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1199 lnet_msg_t *lntmsg0, *lntmsg1;
1200 int status0, status1;
1201 lnet_ni_t *ni = NULL;
1202 kgn_conn_t *conn = tx->tx_conn;
1204 LASSERT(!in_interrupt());
1206 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1207 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1210 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1211 !kgnilnd_conn_clean_errno(completion)) {
1212 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1213 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1214 completion, tx, conn ?
1215 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1216 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1217 kgnilnd_tx_state2str(tx->tx_list_state),
1218 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1221 /* The error codes determine if we hold onto the MDD */
1222 kgnilnd_unmap_buffer(tx, completion);
1224 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1225 * we play nice with the error codes to avoid delivering a failed
1226 * REQUEST and then a REPLY event as well */
1228 /* return -EIO to lnet - it is the magic value for failed sends */
1229 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1231 status1 = completion;
1233 status0 = status1 = completion;
1236 tx->tx_buftype = GNILND_BUF_NONE;
1237 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1239 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1240 * set NULL when we are a tx without a conn */
1242 ni = conn->gnc_peer->gnp_net->gnn_ni;
1244 spin_lock(&conn->gnc_tx_lock);
1246 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1247 (volatile unsigned long *)&conn->gnc_tx_bits),
1248 "conn %p tx %p bit %d already cleared\n",
1249 conn, tx, tx->tx_id.txe_idx);
1251 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1252 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1254 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1255 spin_unlock(&conn->gnc_tx_lock);
1258 kgnilnd_free_tx(tx);
1260 /* finalize AFTER freeing lnet msgs */
1262 /* warning - we should hold no locks here - calling lnet_finalize
1263 * could free up lnet credits, resulting in a call chain back into
1264 * the LND via kgnilnd_send and friends */
1266 lnet_finalize(ni, lntmsg0, status0);
1268 if (lntmsg1 != NULL) {
1269 lnet_finalize(ni, lntmsg1, status1);
1274 kgnilnd_txlist_done(struct list_head *txlist, int error)
1277 int err_printed = 0;
1279 if (list_empty(txlist))
1282 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1283 /* only print the first error */
1285 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1286 list_del_init(&tx->tx_list);
1287 kgnilnd_tx_done(tx, error);
1292 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1296 spin_lock(&conn->gnc_tx_lock);
1298 /* ID zero is NOT ALLOWED!!! */
1301 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1302 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1303 if (id == GNILND_MAX_MSG_ID) {
1304 if (conn->gnc_next_tx != 1) {
1305 /* we only searched from next_tx to end and didn't find
1306 * one, so search again from start */
1307 conn->gnc_next_tx = 1;
1310 /* couldn't find one! */
1311 spin_unlock(&conn->gnc_tx_lock);
1315 /* bump next_tx to prevent immediate reuse */
1316 conn->gnc_next_tx = id + 1;
1318 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1319 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1320 "tx 0x%p already at id %d\n",
1321 conn->gnc_tx_ref_table[id], id);
1323 /* delay these until we have a valid ID - prevents bad clear of the bit
1324 * in kgnilnd_tx_done */
1326 tx->tx_id.txe_cqid = conn->gnc_cqid;
1328 tx->tx_id.txe_idx = id;
1329 conn->gnc_tx_ref_table[id] = tx;
1331 /* Using jiffies to help differentiate against TX reuse - with
1332 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1333 * if we are sending to the same node faster than 256000/sec.
1334 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1336 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1338 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1340 spin_unlock(&conn->gnc_tx_lock);
1345 kgnilnd_tx_should_retry(kgn_conn_t *conn, kgn_tx_t *tx)
1347 int max_retrans = *kgnilnd_tunables.kgn_max_retransmits;
1349 int log_retrans_level;
1351 /* I need kgni credits to send this. Replace tx at the head of the
1352 * fmaq and I'll get rescheduled when credits appear */
1355 conn->gnc_tx_retrans++;
1356 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0) ||
1357 (tx->tx_retrans > (max_retrans / 2)));
1358 log_retrans_level = tx->tx_retrans < (max_retrans / 2) ? D_NET : D_NETERROR;
1360 /* Decision time - either error, warn or just retransmit */
1362 /* we don't care about TX timeout - it could be that the network is slower
1363 * or throttled. We'll keep retranmitting - so if the network is so slow
1364 * that we fill up our mailbox, we'll keep trying to resend that msg
1365 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1366 * that he hasn't send us any traffic in return */
1368 if (tx->tx_retrans > max_retrans) {
1369 /* this means we are not backing off the retransmits
1370 * in a healthy manner and are likely chewing up the
1371 * CPU cycles quite badly */
1372 GNIDBG_TOMSG(D_ERROR, &tx->tx_msg,
1373 "SOFTWARE BUG: too many retransmits (%d) for tx id %x "
1375 tx->tx_retrans, tx->tx_id, conn,
1376 libcfs_nid2str(conn->gnc_peer->gnp_nid));
1378 /* yes - double errors to help debug this condition */
1379 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg, "connection dead. "
1380 "unable to send to %s for %lu secs (%d tries)",
1381 libcfs_nid2str(tx->tx_conn->gnc_peer->gnp_nid),
1382 cfs_duration_sec(jiffies - tx->tx_cred_wait),
1385 kgnilnd_close_conn(conn, -ETIMEDOUT);
1387 /* caller should terminate */
1390 /* some reasonable throttling of the debug message */
1392 unsigned long now = jiffies;
1393 /* XXX Nic: Mystical TX debug here... */
1394 GNIDBG_SMSG_CREDS(log_retrans_level, conn);
1395 GNIDBG_TOMSG(log_retrans_level, &tx->tx_msg,
1396 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1397 " last_msg %uus/%uus last_cq %uus/%uus",
1398 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1399 tx->tx_id, tx->tx_retrans,
1400 jiffies_to_usecs(now - tx->tx_cred_wait),
1401 jiffies_to_usecs(now - conn->gnc_last_tx),
1402 jiffies_to_usecs(now - conn->gnc_last_rx),
1403 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1404 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1406 /* caller should retry */
1411 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1412 * to avoid bad ordering with state_lock */
1415 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1416 spinlock_t *state_lock, kgn_tx_list_state_t state)
1418 kgn_conn_t *conn = tx->tx_conn;
1419 kgn_msg_t *msg = &tx->tx_msg;
1422 unsigned long newest_last_rx, timeout;
1425 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1426 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1428 "msg 0x%p type %d wrong payload size %d\n",
1429 msg, msg->gnm_type, immediatenob);
1431 /* make sure we catch all the cases where we'd send on a dirty old mbox
1432 * but allow case for sending CLOSE. Since this check is within the CQ
1433 * mutex barrier and the close message is only sent through
1434 * kgnilnd_send_conn_close the last message out the door will be the
1437 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1438 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1439 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1440 /* Return -ETIME, we are closing the connection already so we dont want to
1441 * have this tx hit the wire. The tx will be killed by the calling function.
1442 * Once the EP is marked dirty the close message will be the last
1443 * thing to hit the wire */
1448 timeout = cfs_time_seconds(conn->gnc_timeout);
1450 newest_last_rx = GNILND_LASTRX(conn);
1452 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1453 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1456 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1457 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1458 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1459 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1460 cfs_duration_sec(now - newest_last_rx),
1461 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1462 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1463 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1467 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1468 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1469 * creating the message this allows the message to contain the correct LNET NID/NET needed
1470 * instead of the one that the peer/conn uses for sending the data.
1472 msg->gnm_connstamp = conn->gnc_my_connstamp;
1473 msg->gnm_payload_len = immediatenob;
1474 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1476 /* always init here - kgn_checksum is a /sys module tunable
1477 * and can be flipped at any point, even between msg init and sending */
1479 if (*kgnilnd_tunables.kgn_checksum) {
1480 /* We must set here and not in kgnilnd_init_msg,
1481 * we could resend this msg many times
1482 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1483 * through init_msg again */
1484 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1485 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1486 msg->gnm_cksum += 0xf00f;
1490 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1491 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1492 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1493 tx->tx_id.txe_idx, immediate, immediatenob);
1495 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1496 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1498 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1499 msg, sizeof(*msg), immediate,
1501 tx->tx_id.txe_smsg_id);
1505 case GNI_RC_SUCCESS:
1506 atomic_inc(&conn->gnc_tx_seq);
1507 conn->gnc_last_tx = jiffies;
1508 /* no locking here as LIVE isn't a list */
1509 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1511 /* this needs to be checked under lock as it might be freed from a completion
1514 if (msg->gnm_type == GNILND_MSG_NOOP) {
1515 set_mb(conn->gnc_last_noop_sent, jiffies);
1518 /* serialize with seeing CQ events for completion on this, as well as
1520 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1521 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1523 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1524 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1525 kgnilnd_peer_alive(conn->gnc_peer);
1526 GNIDBG_SMSG_CREDS(D_NET, conn);
1529 case GNI_RC_NOT_DONE:
1530 /* XXX Nic: We need to figure out how to track this
1531 * - there are bound to be good reasons for it,
1532 * but we want to know when it happens */
1533 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1534 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1535 /* We'll handle this error inline - makes the calling logic much more
1538 /* If no lock, caller doesn't want us to retry */
1539 if (state_lock == NULL) {
1543 retry_send = kgnilnd_tx_should_retry(conn, tx);
1545 /* add to head of list for the state and retries */
1546 spin_lock(state_lock);
1547 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1548 spin_unlock(state_lock);
1550 /* We only reschedule for a certain number of retries, then
1551 * we will wait for the CQ events indicating a release of SMSG
1553 if (tx->tx_retrans < (*kgnilnd_tunables.kgn_max_retransmits/4)) {
1554 kgnilnd_schedule_conn(conn);
1557 /* CQ event coming in signifies either TX completed or
1558 * RX receive. Either of these *could* free up credits
1559 * in the SMSG mbox and we should try sending again */
1560 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1561 tx->tx_conn->gnc_cqid);
1562 /* use +ve return code to let upper layers know they
1563 * should stop looping on sends */
1570 /* handle bad retcode gracefully */
1571 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1572 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1577 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1579 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1580 spinlock_t *state_lock, kgn_tx_list_state_t state)
1582 kgn_device_t *dev = tx->tx_conn->gnc_device;
1583 unsigned long timestamp;
1586 timestamp = jiffies;
1587 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1588 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1589 /* delay in jiffies - we are really concerned only with things that
1590 * result in a schedule() or really holding this off for long times .
1591 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1592 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1594 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1600 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1602 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1603 spinlock_t *state_lock, kgn_tx_list_state_t state)
1605 kgn_conn_t *conn = tx->tx_conn;
1606 kgn_device_t *dev = conn->gnc_device;
1607 unsigned long timestamp;
1610 timestamp = jiffies;
1612 /* technically we are doing bad things with the read_lock on the peer_conn
1613 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1614 * for the mutex. I bet lockdep is gonna flag this one though... */
1616 /* there are a few cases where we don't want the immediate send - like
1617 * when we are in the scheduler thread and it'd harm the latency of
1618 * getting messages up to LNet */
1620 /* rmb for gnd_ready */
1622 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1624 atomic_inc(&conn->gnc_device->gnd_fast_block);
1625 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1626 /* dont hit HW during quiesce */
1628 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1629 /* dont hit HW if stale EPs and conns left to close */
1632 atomic_inc(&conn->gnc_device->gnd_fast_try);
1633 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1634 &conn->gnc_smsg_mutex);
1639 /* we got the mutex and weren't blocked */
1641 /* delay in jiffies - we are really concerned only with things that
1642 * result in a schedule() or really holding this off for long times .
1643 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1644 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1646 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1647 tx->tx_qtime = jiffies;
1648 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1649 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1650 /* _nolock unlocks the mutex for us */
1656 /* lets us know if we can push this RDMA through now */
1658 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1662 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1664 if (bytes_left < 0) {
1665 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1666 atomic_inc(&dev->gnd_rdmaq_nstalls);
1669 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1670 dev->gnd_rdmaq_deadline);
1671 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1672 /* we never del this timer - at worst it schedules us.. */
1679 /* this adds a TX to the queue pending throttling authorization before
1680 * we allow our remote peer to launch a PUT at us */
1682 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1686 /* we cannot go into send_mapped_tx from here as we are holding locks
1687 * and mem registration might end up allocating memory in kgni.
1688 * That said, we'll push this as far as we can into the queue process */
1689 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1692 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1693 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1694 /* lets us know how delayed RDMA is */
1695 tx->tx_qtime = jiffies;
1696 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1698 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1699 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1700 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1701 /* lets us know how delayed mapping is */
1702 tx->tx_qtime = jiffies;
1703 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1706 /* make sure we wake up sched to run this */
1707 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1710 /* push TX through state machine */
1712 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1717 /* set the tx_id here, we delay it until we have an actual conn
1719 * in some cases, the tx_id is already set to provide for things
1720 * like RDMA completion cookies, etc */
1721 if (tx->tx_id.txe_idx == 0) {
1722 rc = kgnilnd_set_tx_id(tx, conn);
1724 kgnilnd_tx_done(tx, rc);
1729 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1730 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1732 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1734 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1735 kgnilnd_tx_done(tx, rc);
1739 switch (tx->tx_msg.gnm_type) {
1740 case GNILND_MSG_PUT_ACK:
1741 case GNILND_MSG_GET_REQ:
1742 case GNILND_MSG_PUT_REQ_REV:
1743 case GNILND_MSG_GET_ACK_REV:
1744 /* hijacking time! If this messages will authorize our peer to
1745 * send his dirty little bytes in an RDMA, we need to get permission */
1746 kgnilnd_queue_rdma(conn, tx);
1748 case GNILND_MSG_IMMEDIATE:
1749 /* try to send right now, can help reduce latency */
1750 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1753 /* it was sent, break out of switch to avoid default case of queueing */
1756 /* needs to queue to try again, so fall through to default case */
1757 case GNILND_MSG_NOOP:
1758 /* Just make sure this goes out first for this conn */
1760 /* fall through... */
1762 spin_lock(&conn->gnc_list_lock);
1763 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1764 tx->tx_qtime = jiffies;
1765 spin_unlock(&conn->gnc_list_lock);
1766 kgnilnd_schedule_conn(conn);
1771 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
1774 kgn_peer_t *new_peer = NULL;
1775 kgn_conn_t *conn = NULL;
1781 /* If I get here, I've committed to send, so I complete the tx with
1782 * failure on any problems */
1784 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1785 "tx already has connection %p", tx->tx_conn);
1787 /* do all of the peer & conn searching in one swoop - this avoids
1788 * nastiness when dropping locks and needing to maintain a sane state
1789 * in the face of stack reset or something else nuking peers & conns */
1791 /* I expect to find him, so only take a read lock */
1792 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1794 peer = kgnilnd_find_peer_locked(target->nid);
1796 conn = kgnilnd_find_conn_locked(peer);
1797 /* this could be NULL during quiesce */
1799 /* Connection exists; queue message on it */
1800 kgnilnd_queue_tx(conn, tx);
1801 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1805 /* don't create a connection if the peer is marked down */
1806 if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
1807 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1813 /* creating peer or conn; I'll need a write lock... */
1814 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1816 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1818 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1820 /* NB - this will not block during normal operations -
1821 * the only writer of this is in the startup/shutdown path. */
1822 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1828 /* ignore previous peer entirely - we cycled the lock, so we
1829 * will create new peer and at worst drop it if peer is still
1831 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1833 up_read(&kgnilnd_data.kgn_net_rw_sem);
1837 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1838 up_read(&kgnilnd_data.kgn_net_rw_sem);
1840 /* search for peer again now that we have the lock
1841 * if we don't find it, add our new one to the list */
1842 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1844 /* don't create a connection if the peer is not up */
1845 if (peer->gnp_down != GNILND_RCA_NODE_UP) {
1846 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1851 conn = kgnilnd_find_or_create_conn_locked(peer);
1853 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1854 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1859 /* oh hey, found a conn now... magical */
1860 kgnilnd_queue_tx(conn, tx);
1862 /* no conn, must be trying to connect - so we queue for now */
1863 tx->tx_qtime = jiffies;
1864 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1866 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1869 kgnilnd_tx_done(tx, rc);
1874 kgnilnd_rdma(kgn_tx_t *tx, int type,
1875 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1877 kgn_conn_t *conn = tx->tx_conn;
1878 unsigned long timestamp;
1879 gni_post_type_t post_type;
1882 unsigned int desc_nob = nob;
1883 void *desc_buffer = tx->tx_buffer;
1884 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1885 LASSERTF(kgnilnd_tx_mapped(tx),
1886 "unmapped tx %p\n", tx);
1887 LASSERTF(conn != NULL,
1888 "NULL conn on tx %p, naughty, naughty\n", tx);
1889 LASSERTF(nob <= sink->gnrd_nob,
1890 "nob %u > sink->gnrd_nob %d (%p)\n",
1891 nob, sink->gnrd_nob, sink);
1892 LASSERTF(nob <= tx->tx_nob,
1893 "nob %d > tx(%p)->tx_nob %d\n",
1894 nob, tx, tx->tx_nob);
1897 case GNILND_MSG_GET_DONE:
1898 case GNILND_MSG_PUT_DONE:
1899 post_type = GNI_POST_RDMA_PUT;
1901 case GNILND_MSG_GET_DONE_REV:
1902 case GNILND_MSG_PUT_DONE_REV:
1903 post_type = GNI_POST_RDMA_GET;
1906 CERROR("invalid msg type %s (%d)\n",
1907 kgnilnd_msgtype2str(type), type);
1910 if (post_type == GNI_POST_RDMA_GET) {
1911 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1912 * aligned. If the local buffer is not aligned correctly using the copy buffer
1913 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1914 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1915 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1916 * making sure the length and addr are aligned and copying the data into the correct location
1917 * once the transfer has completed.
1919 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1920 (sink->gnrd_addr & 3) ||
1923 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1925 kgnilnd_admin_addref(kgnilnd_data.kgn_rev_offset);
1927 if ((nob + tx->tx_offset) & 3) {
1928 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1929 kgnilnd_admin_addref(kgnilnd_data.kgn_rev_length);
1931 desc_nob = (nob + tx->tx_offset);
1934 if (tx->tx_buffer_copy == NULL) {
1935 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1936 * and require at most we allocate a few extra bytes. */
1937 tx->tx_buffer_copy = vmalloc(desc_nob);
1939 if (!tx->tx_buffer_copy) {
1940 /* allocation of buffer failed nak the rdma */
1941 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1942 kgnilnd_tx_done(tx, -EFAULT);
1945 kgnilnd_admin_addref(kgnilnd_data.kgn_rev_copy_buff);
1946 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);
1947 if (rc != GNI_RC_SUCCESS) {
1948 /* Registration Failed nak rdma and kill the tx. */
1949 vfree(tx->tx_buffer_copy);
1950 tx->tx_buffer_copy = NULL;
1951 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1952 kgnilnd_tx_done(tx, -EFAULT);
1956 desc_map_key = tx->tx_buffer_copy_map_key;
1957 desc_buffer = tx->tx_buffer_copy;
1961 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1962 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1963 tx->tx_rdma_desc.type = post_type;
1964 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1965 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1966 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1967 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1968 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1969 tx->tx_rdma_desc.length = desc_nob;
1970 tx->tx_nob_rdma = nob;
1971 if (*kgnilnd_tunables.kgn_bte_dlvr_mode)
1972 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_dlvr_mode;
1973 /* prep final completion message */
1974 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1975 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1976 /* send actual size RDMA'd in retval */
1977 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1979 kgnilnd_compute_rdma_cksum(tx, nob);
1982 kgnilnd_queue_tx(conn, tx);
1986 /* Don't lie (CLOSE == RDMA idle) */
1987 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1988 tx, conn, conn->gnc_close_sent);
1990 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1991 "0x%x cookie:"LPX64,
1992 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1994 /* set CQ dedicated for RDMA */
1995 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1997 timestamp = jiffies;
1998 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1999 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2000 /* delay in jiffies - we are really concerned only with things that
2001 * result in a schedule() or really holding this off for long times .
2002 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2003 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2005 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2007 if (rrc == GNI_RC_ERROR_RESOURCE) {
2008 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2009 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2010 kgnilnd_unmap_buffer(tx, 0);
2012 if (tx->tx_buffer_copy != NULL) {
2013 vfree(tx->tx_buffer_copy);
2014 tx->tx_buffer_copy = NULL;
2017 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2018 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2020 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2021 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2025 spin_lock(&conn->gnc_list_lock);
2026 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2027 tx->tx_qtime = jiffies;
2028 spin_unlock(&conn->gnc_list_lock);
2029 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2030 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2032 /* XXX Nic: is this a place we should handle more errors for
2033 * robustness sake */
2034 LASSERT(rrc == GNI_RC_SUCCESS);
2039 kgnilnd_alloc_rx(void)
2043 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2045 CERROR("failed to allocate rx\n");
2048 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2051 /* no memset to zero, we'll always fill all members */
2055 /* release is to just free connection resources
2056 * we use this for the eager path after copying */
2058 kgnilnd_release_msg(kgn_conn_t *conn)
2061 unsigned long timestamp;
2063 CDEBUG(D_NET, "consuming %p\n", conn);
2065 timestamp = jiffies;
2066 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2067 /* delay in jiffies - we are really concerned only with things that
2068 * result in a schedule() or really holding this off for long times .
2069 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2070 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2072 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2073 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2075 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2076 GNIDBG_SMSG_CREDS(D_NET, conn);
2082 kgnilnd_consume_rx(kgn_rx_t *rx)
2084 kgn_conn_t *conn = rx->grx_conn;
2085 kgn_msg_t *rxmsg = rx->grx_msg;
2087 /* if we are eager, free the cache alloc'd msg */
2088 if (unlikely(rx->grx_eager)) {
2089 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2090 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2092 /* release ref from eager_recv */
2093 kgnilnd_conn_decref(conn);
2095 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2096 kgnilnd_release_msg(conn);
2099 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2100 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2107 kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
2109 lnet_hdr_t *hdr = &lntmsg->msg_hdr;
2110 int type = lntmsg->msg_type;
2111 lnet_process_id_t target = lntmsg->msg_target;
2112 int target_is_router = lntmsg->msg_target_is_router;
2113 int routing = lntmsg->msg_routing;
2114 unsigned int niov = lntmsg->msg_niov;
2115 struct iovec *iov = lntmsg->msg_iov;
2116 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2117 unsigned int offset = lntmsg->msg_offset;
2118 unsigned int nob = lntmsg->msg_len;
2119 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2120 kgn_net_t *net = ni->ni_data;
2124 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2126 /* NB 'private' is different depending on what we're sending.... */
2127 LASSERT(!in_interrupt());
2129 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2130 type, nob, niov, libcfs_id2str(target));
2132 LASSERTF(nob == 0 || niov > 0,
2133 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2134 LASSERTF(niov <= LNET_MAX_IOV,
2135 "lntmsg %p niov %d\n", lntmsg, niov);
2137 /* payload is either all vaddrs or all pages */
2138 LASSERTF(!(kiov != NULL && iov != NULL),
2139 "lntmsg %p kiov %p iov %p\n", lntmsg, kiov, iov);
2142 mpflag = cfs_memory_pressure_get_and_set();
2146 CERROR("lntmsg %p with unexpected type %d\n",
2151 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2159 if (routing || target_is_router)
2160 break; /* send IMMEDIATE */
2162 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2163 * check the eventual sink buffer here - if small enough, remote
2164 * end is perfectly capable of returning data in short message -
2165 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2166 * for IMMEDIATE messages which will have it send a real reply instead
2167 * of doing kgnilnd_recv to have the RDMA continued */
2168 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2171 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2172 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2174 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2180 /* slightly different options as we might actually have a GET with a
2181 * MD_KIOV set but a non-NULL md_iov.iov */
2182 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
2183 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2184 lntmsg->msg_md->md_iov.iov, NULL,
2185 0, lntmsg->msg_md->md_length);
2187 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2188 NULL, lntmsg->msg_md->md_iov.kiov,
2189 0, lntmsg->msg_md->md_length);
2191 CERROR("unable to setup buffer: %d\n", rc);
2192 kgnilnd_tx_done(tx, rc);
2197 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2198 if (tx->tx_lntmsg[1] == NULL) {
2199 CERROR("Can't create reply for GET to %s\n",
2200 libcfs_nid2str(target.nid));
2201 kgnilnd_tx_done(tx, rc);
2206 tx->tx_lntmsg[0] = lntmsg;
2207 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2208 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2210 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2212 /* rest of tx_msg is setup just before it is sent */
2213 kgnilnd_launch_tx(tx, net, &target);
2215 case LNET_MSG_REPLY:
2217 /* to save on MDDs, we'll handle short kiov by vmap'ing
2218 * and sending via SMSG */
2219 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2222 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2223 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2225 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2232 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2234 kgnilnd_tx_done(tx, rc);
2239 tx->tx_lntmsg[0] = lntmsg;
2240 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2241 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2243 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2245 /* rest of tx_msg is setup just before it is sent */
2246 kgnilnd_launch_tx(tx, net, &target);
2250 /* send IMMEDIATE */
2252 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2253 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2255 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2261 rc = kgnilnd_setup_immediate_buffer(tx, niov, iov, kiov, offset, nob);
2263 kgnilnd_tx_done(tx, rc);
2267 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2268 tx->tx_lntmsg[0] = lntmsg;
2269 kgnilnd_launch_tx(tx, net, &target);
2272 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2274 cfs_memory_pressure_restore(mpflag);
2279 kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen)
2281 kgn_conn_t *conn = rx->grx_conn;
2282 kgn_msg_t *rxmsg = rx->grx_msg;
2283 unsigned int niov = lntmsg->msg_niov;
2284 struct iovec *iov = lntmsg->msg_iov;
2285 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2286 unsigned int offset = lntmsg->msg_offset;
2287 unsigned int nob = lntmsg->msg_len;
2292 switch (rxmsg->gnm_type) {
2293 case GNILND_MSG_PUT_REQ_REV:
2294 done_type = GNILND_MSG_PUT_DONE_REV;
2297 case GNILND_MSG_GET_REQ:
2298 done_type = GNILND_MSG_GET_DONE;
2301 CERROR("invalid msg type %s (%d)\n",
2302 kgnilnd_msgtype2str(rxmsg->gnm_type),
2307 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2311 rc = kgnilnd_set_tx_id(tx, conn);
2315 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2319 tx->tx_lntmsg[0] = lntmsg;
2320 tx->tx_getinfo = rxmsg->gnm_u.get;
2322 /* we only queue from kgnilnd_recv - we might get called from other contexts
2323 * and we don't want to block the mutex in those cases */
2325 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2326 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2327 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2328 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2333 kgnilnd_tx_done(tx, rc);
2334 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2336 lnet_finalize(ni, lntmsg, rc);
2340 kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
2343 kgn_rx_t *rx = private;
2344 kgn_conn_t *conn = rx->grx_conn;
2345 kgn_msg_t *rxmsg = rx->grx_msg;
2346 kgn_msg_t *eagermsg = NULL;
2347 kgn_peer_t *peer = NULL;
2348 kgn_conn_t *found_conn = NULL;
2350 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2351 conn, rxmsg, lntmsg);
2353 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2354 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2355 rxmsg->gnm_payload_len);
2358 /* Grab a read lock so the connection doesnt disappear on us
2359 * while we look it up
2361 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2363 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2365 found_conn = kgnilnd_find_conn_locked(peer);
2368 /* Verify the connection found is the same one that the message
2369 * is supposed to be using, if it is not output an error message
2372 if (!peer || !found_conn
2373 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2374 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2375 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2376 peer, conn, found_conn);
2378 CERROR("Unexpected connstamp "LPX64"("LPX64" expected)"
2379 " from %s", rxmsg->gnm_connstamp,
2380 found_conn->gnc_peer_connstamp,
2381 libcfs_nid2str(peer->gnp_nid));
2386 /* add conn ref to ensure it doesn't go away until all eager
2387 * messages processed */
2388 kgnilnd_conn_addref(conn);
2390 /* Now that we have verified the connection is valid and added a
2391 * reference we can remove the read_lock on the peer_conn_lock */
2392 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2394 /* we have no credits or buffers for this message, so copy it
2395 * somewhere for a later kgnilnd_recv */
2396 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2397 *kgnilnd_tunables.kgn_eager_credits) {
2398 CERROR("Out of eager credits to %s\n",
2399 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2403 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2405 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2406 if (eagermsg == NULL) {
2407 kgnilnd_conn_decref(conn);
2408 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2409 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2413 /* copy msg and payload */
2414 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2415 rx->grx_msg = eagermsg;
2418 /* stash this for lnet_finalize on cancel-on-conn-close */
2419 rx->grx_lntmsg = lntmsg;
2421 /* keep the same rx_t, it just has a new grx_msg now */
2422 *new_private = private;
2424 /* release SMSG buffer */
2425 kgnilnd_release_msg(conn);
2431 kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
2432 int delayed, unsigned int niov,
2433 struct iovec *iov, lnet_kiov_t *kiov,
2434 unsigned int offset, unsigned int mlen, unsigned int rlen)
2436 kgn_rx_t *rx = private;
2437 kgn_conn_t *conn = rx->grx_conn;
2438 kgn_msg_t *rxmsg = rx->grx_msg;
2444 LASSERT(!in_interrupt());
2445 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2446 /* Either all pages or all vaddrs */
2447 LASSERTF(!(kiov != NULL && iov != NULL), "kiov %p iov %p\n",
2450 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2451 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2452 conn, rxmsg, lntmsg,
2453 niov, kiov, iov, offset, mlen, rlen);
2455 /* we need to lock here as recv can be called from any context */
2456 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2457 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2458 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2460 /* someone closed the conn after we copied this out, nuke it */
2461 kgnilnd_consume_rx(rx);
2462 lnet_finalize(ni, lntmsg, conn->gnc_error);
2465 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2467 switch (rxmsg->gnm_type) {
2469 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2470 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2471 conn, rx, rxmsg, lntmsg, niov, kiov, iov, offset, mlen, rlen);
2474 case GNILND_MSG_IMMEDIATE:
2475 if (mlen > rxmsg->gnm_payload_len) {
2476 GNIDBG_MSG(D_ERROR, rxmsg,
2477 "Immediate message from %s too big: %d > %d",
2478 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2479 rxmsg->gnm_payload_len);
2481 kgnilnd_consume_rx(rx);
2485 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2486 * right after the kgn_msg_t header - so just 'cute' way of saying
2487 * rxmsg + sizeof(kgn_msg_t) */
2489 /* check payload checksum if sent */
2491 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2492 !rxmsg->gnm_payload_cksum &&
2493 rxmsg->gnm_payload_len != 0)
2494 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2496 if (rxmsg->gnm_payload_cksum != 0) {
2497 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2498 * which is what is used to calculate the cksum on the TX side */
2499 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2501 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2502 GNIDBG_MSG(D_NETERROR, rxmsg,
2503 "Bad payload checksum (%x expected %x)",
2504 pload_cksum, rxmsg->gnm_payload_cksum);
2505 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2507 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2508 &rxmsg[1], rxmsg->gnm_payload_len);
2509 /* fall through to dump */
2511 libcfs_debug_dumplog();
2517 /* checksum problems are fatal, kill the conn */
2518 kgnilnd_consume_rx(rx);
2519 kgnilnd_close_conn(conn, rc);
2525 lnet_copy_flat2kiov(
2527 *kgnilnd_tunables.kgn_max_immediate,
2528 &rxmsg[1], 0, mlen);
2532 *kgnilnd_tunables.kgn_max_immediate,
2533 &rxmsg[1], 0, mlen);
2535 kgnilnd_consume_rx(rx);
2536 lnet_finalize(ni, lntmsg, 0);
2539 case GNILND_MSG_PUT_REQ:
2540 /* LNET wants to truncate or drop transaction, sending NAK */
2542 kgnilnd_consume_rx(rx);
2543 lnet_finalize(ni, lntmsg, 0);
2545 /* only error if lntmsg == NULL, otherwise we are just
2546 * short circuiting the rdma process of 0 bytes */
2547 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2548 lntmsg == NULL ? -ENOENT : 0,
2549 rxmsg->gnm_u.get.gngm_cookie,
2553 /* sending ACK with sink buff. info */
2554 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2556 kgnilnd_consume_rx(rx);
2560 rc = kgnilnd_set_tx_id(tx, conn);
2562 GOTO(nak_put_req, rc);
2565 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2567 GOTO(nak_put_req, rc);
2570 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2571 rxmsg->gnm_u.putreq.gnprm_cookie;
2572 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2573 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2574 (__u64)((unsigned long)tx->tx_buffer);
2575 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2577 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2578 tx->tx_qtime = jiffies;
2579 /* we only queue from kgnilnd_recv - we might get called from other contexts
2580 * and we don't want to block the mutex in those cases */
2582 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2583 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2584 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2585 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2587 kgnilnd_consume_rx(rx);
2591 /* make sure we send an error back when the PUT fails */
2592 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2593 kgnilnd_tx_done(tx, rc);
2594 kgnilnd_consume_rx(rx);
2596 /* return magic LNet network error */
2598 case GNILND_MSG_GET_REQ_REV:
2599 /* LNET wants to truncate or drop transaction, sending NAK */
2601 kgnilnd_consume_rx(rx);
2602 lnet_finalize(ni, lntmsg, 0);
2604 /* only error if lntmsg == NULL, otherwise we are just
2605 * short circuiting the rdma process of 0 bytes */
2606 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2607 lntmsg == NULL ? -ENOENT : 0,
2608 rxmsg->gnm_u.get.gngm_cookie,
2612 /* lntmsg can be null when parsing a LNET_GET */
2613 if (lntmsg != NULL) {
2614 /* sending ACK with sink buff. info */
2615 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2617 kgnilnd_consume_rx(rx);
2621 rc = kgnilnd_set_tx_id(tx, conn);
2623 GOTO(nak_get_req_rev, rc);
2626 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2628 GOTO(nak_get_req_rev, rc);
2631 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2632 rxmsg->gnm_u.putreq.gnprm_cookie;
2633 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2634 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2635 (__u64)((unsigned long)tx->tx_buffer);
2636 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2638 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2640 /* we only queue from kgnilnd_recv - we might get called from other contexts
2641 * and we don't want to block the mutex in those cases */
2643 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2644 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2645 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2646 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2649 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2651 rxmsg->gnm_u.get.gngm_cookie,
2655 kgnilnd_consume_rx(rx);
2659 /* make sure we send an error back when the GET fails */
2660 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2661 kgnilnd_tx_done(tx, rc);
2662 kgnilnd_consume_rx(rx);
2664 /* return magic LNet network error */
2668 case GNILND_MSG_PUT_REQ_REV:
2669 /* LNET wants to truncate or drop transaction, sending NAK */
2671 kgnilnd_consume_rx(rx);
2672 lnet_finalize(ni, lntmsg, 0);
2674 /* only error if lntmsg == NULL, otherwise we are just
2675 * short circuiting the rdma process of 0 bytes */
2676 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2677 lntmsg == NULL ? -ENOENT : 0,
2678 rxmsg->gnm_u.get.gngm_cookie,
2683 if (lntmsg != NULL) {
2685 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2688 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2690 rxmsg->gnm_u.get.gngm_cookie,
2693 kgnilnd_consume_rx(rx);
2695 case GNILND_MSG_GET_REQ:
2696 if (lntmsg != NULL) {
2698 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2701 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2703 rxmsg->gnm_u.get.gngm_cookie,
2706 kgnilnd_consume_rx(rx);
2712 /* needs write_lock on kgn_peer_conn_lock held */
2714 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2716 unsigned long timeout, keepalive;
2717 unsigned long now = jiffies;
2718 unsigned long newest_last_rx;
2721 /* given that we found this conn hanging off a peer, it better damned
2722 * well be connected */
2723 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2724 "conn 0x%p->%s with bad state%s\n", conn,
2725 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2727 kgnilnd_conn_state2str(conn));
2729 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2730 "rx_diff %lu tx_diff %lu\n",
2731 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2732 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2733 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2734 cfs_duration_sec(now - conn->gnc_last_tx));
2736 timeout = cfs_time_seconds(conn->gnc_timeout);
2737 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2739 /* just in case our lack of RX msg processing is gumming up the works - give the
2740 * remove an extra chance */
2742 newest_last_rx = GNILND_LASTRX(conn);
2744 if (time_after_eq(now, newest_last_rx + timeout)) {
2745 uint32_t level = D_CONSOLE|D_NETERROR;
2747 if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_DOWN) {
2750 GNIDBG_CONN(level, conn,
2751 "No gnilnd traffic received from %s for %lu "
2752 "seconds, terminating connection. Is node down? ",
2753 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2754 cfs_duration_sec(now - newest_last_rx));
2758 /* we don't timeout on last_tx stalls - we are going to trust the
2759 * underlying network to let us know when sends are failing.
2760 * At worst, the peer will timeout our RX stamp and drop the connection
2761 * at that point. We'll then see his CLOSE or at worst his RX
2762 * stamp stop and drop the connection on our end */
2764 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2765 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2766 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2767 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2768 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2770 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2771 conn->gnc_last_noop_cq,
2772 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2773 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2774 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2775 set_mb(conn->gnc_last_noop_want, jiffies);
2776 atomic_inc(&conn->gnc_reaper_noop);
2777 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2780 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2783 kgnilnd_queue_tx(conn, tx);
2789 /* needs write_lock on kgn_peer_conn_lock held */
2791 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2792 struct list_head *souls)
2794 unsigned long timeout;
2795 kgn_conn_t *conn, *connN = NULL;
2801 short releaseconn = 0;
2802 unsigned long first_rx = 0;
2803 int purgatory_conn_cnt = 0;
2805 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2806 peer, libcfs_nid2str(peer->gnp_nid),
2807 peer->gnp_reconnect_interval);
2809 timeout = cfs_time_seconds(MAX(*kgnilnd_tunables.kgn_timeout,
2810 GNILND_MIN_TIMEOUT));
2812 conn = kgnilnd_find_conn_locked(peer);
2814 /* if there is a valid conn, check the queues for timeouts */
2815 rc = kgnilnd_check_conn_timeouts_locked(conn);
2817 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2818 /* simulate a RX CLOSE after the timeout but before
2819 * the scheduler thread gets it */
2820 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2821 conn->gnc_peer_error = -ETIMEDOUT;
2823 /* Once we mark closed, any of the scheduler threads could
2824 * get it and move through before we hit the fail loc code */
2825 kgnilnd_close_conn_locked(conn, rc);
2827 /* first_rx is used to decide when to release a conn from purgatory.
2829 first_rx = conn->gnc_first_rx;
2833 /* now regardless of starting new conn, find tx on peer queue that
2834 * are old and smell bad - do this first so we don't trigger
2835 * reconnect on empty queue if we timeout all */
2836 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2837 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2839 LCONSOLE_INFO("could not send to %s due to connection"
2840 " setup failure after %lu seconds\n",
2841 libcfs_nid2str(peer->gnp_nid),
2842 cfs_duration_sec(jiffies - tx->tx_qtime));
2844 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2846 list_add_tail(&tx->tx_list, todie);
2851 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2852 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2853 count, peer, libcfs_nid2str(peer->gnp_nid));
2854 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2855 if (list_empty(&peer->gnp_tx_queue) ||
2856 peer->gnp_connecting == GNILND_PEER_KILL) {
2857 /* we pass down todie to use a common function - but we know there are
2859 kgnilnd_cancel_peer_connect_locked(peer, todie);
2863 /* Don't reconnect if we are still trying to clear out old conns.
2864 * This prevents us sending traffic on the new mbox before ensuring we are done
2865 * with the old one */
2866 reconnect = (peer->gnp_down == GNILND_RCA_NODE_UP) &&
2867 (atomic_read(&peer->gnp_dirty_eps) == 0);
2869 /* fast reconnect after a timeout */
2870 to_reconn = !conn &&
2871 (peer->gnp_last_errno == -ETIMEDOUT) &&
2872 *kgnilnd_tunables.kgn_fast_reconn;
2874 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2875 * to be sent, we'll check the reconnect interval and fire up a new
2876 * connection request */
2879 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2880 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2881 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2883 CDEBUG(D_NET, "starting connect to %s\n",
2884 libcfs_nid2str(peer->gnp_nid));
2885 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE, "Peer was idle and we"
2886 "have a write_lock, state issue %d\n", peer->gnp_connecting);
2888 peer->gnp_connecting = GNILND_PEER_CONNECT;
2889 kgnilnd_peer_addref(peer); /* extra ref for connd */
2891 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2892 list_add_tail(&peer->gnp_connd_list,
2893 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2894 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2896 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2899 /* fail_loc to allow us to delay release of purgatory */
2900 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2903 /* This check allows us to verify that the new conn is actually being used. This allows us to
2904 * pull the old conns out of purgatory if they have actually seen traffic.
2905 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2908 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2909 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2910 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2914 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2915 /* check for purgatory timeouts */
2916 if (conn->gnc_in_purgatory) {
2917 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2918 * that way the next time we check it we can detach it from purgatory
2921 if (conn->gnc_state != GNILND_CONN_DONE) {
2922 /* Skip over conns that are currently not DONE. If they arent already scheduled
2923 * for completion something in the state machine is broken.
2928 /* We only detach a conn that is in purgatory if we have received a close message,
2929 * we have a new valid connection that has successfully received data, or an admin
2930 * command tells us we need to detach.
2933 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2934 unsigned long waiting;
2936 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2938 /* C.E: The remote peer is expected to close the
2939 * connection (see kgnilnd_check_conn_timeouts)
2940 * via the reaper thread and nuke out the MDD and
2941 * FMA resources after conn->gnc_timeout has expired
2942 * without an FMA RX */
2943 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2944 " held resources\n",
2945 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2946 cfs_duration_sec(waiting));
2948 kgnilnd_detach_purgatory_locked(conn, souls);
2950 purgatory_conn_cnt++;
2955 /* If we have too many connections in purgatory we could run out of
2956 * resources. Limit the number of connections to a tunable number,
2957 * clean up to the minimum all in one fell swoop... there are
2958 * situations where dvs will retry tx's and we can eat up several
2959 * hundread connection requests at once.
2961 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2962 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2964 if (conn->gnc_in_purgatory &&
2965 conn->gnc_state == GNILND_CONN_DONE) {
2966 CDEBUG(D_NET, "Dropping Held resource due to"
2967 " resource limits being hit\n");
2968 kgnilnd_detach_purgatory_locked(conn, souls);
2970 if (purgatory_conn_cnt-- <
2971 *kgnilnd_tunables.kgn_max_purgatory)
2981 kgnilnd_reaper_check(int idx)
2983 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2984 struct list_head *ctmp, *ctmpN;
2985 struct list_head geriatrics;
2986 struct list_head souls;
2988 INIT_LIST_HEAD(&geriatrics);
2989 INIT_LIST_HEAD(&souls);
2991 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2993 list_for_each_safe(ctmp, ctmpN, peers) {
2994 kgn_peer_t *peer = NULL;
2996 /* don't timeout stuff if the network is mucked or shutting down */
2997 if (kgnilnd_check_hw_quiesce()) {
3000 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
3002 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
3005 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3007 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3008 kgnilnd_release_purgatory_list(&souls);
3012 kgnilnd_update_reaper_timeout(long timeout)
3014 LASSERT(timeout > 0);
3016 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3018 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3019 kgnilnd_data.kgn_new_min_timeout = timeout;
3021 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3025 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3027 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3031 kgnilnd_reaper(void *arg)
3036 unsigned long next_check_time = jiffies;
3037 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3038 struct timer_list timer;
3041 cfs_block_allsigs();
3043 /* all gnilnd threads need to run fairly urgently */
3044 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3045 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3047 while (!kgnilnd_data.kgn_shutdown) {
3048 /* I wake up every 'p' seconds to check for timeouts on some
3049 * more peers. I try to check every connection 'n' times
3050 * within the global minimum of all keepalive and timeout
3051 * intervals, to ensure I attend to every connection within
3052 * (n+1)/n times its timeout intervals. */
3053 const int p = GNILND_REAPER_THREAD_WAKE;
3054 const int n = GNILND_REAPER_NCHECKS;
3056 /* to quiesce or to not quiesce, that is the question */
3057 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3058 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3059 KGNILND_SPIN_QUIESCE;
3060 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3063 /* careful with the jiffy wrap... */
3064 timeout = (long)(next_check_time - jiffies);
3067 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3068 TASK_INTERRUPTIBLE);
3069 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3070 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3072 mod_timer(&timer, (long) jiffies + timeout);
3074 /* check flag variables before comitting */
3075 if (!kgnilnd_data.kgn_shutdown &&
3076 !kgnilnd_data.kgn_quiesce_trigger) {
3077 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3078 timeout, cfs_duration_sec(timeout));
3080 CDEBUG(D_INFO, "awake after schedule\n");
3083 del_singleshot_timer_sync(&timer);
3084 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3085 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3089 /* new_min_timeout is set from the conn timeouts and keepalive
3090 * this should end up with a min timeout of
3091 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3092 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3093 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3094 CDEBUG(D_NET, "Set new min timeout %ld\n",
3095 current_min_timeout);
3098 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3100 /* Compute how many table entries to check now so I get round
3101 * the whole table fast enough given that I do this at fixed
3102 * intervals of 'p' seconds) */
3103 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3104 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3105 chunk = (chunk * n * p) /
3106 kgnilnd_data.kgn_new_min_timeout;
3109 for (i = 0; i < chunk; i++) {
3110 kgnilnd_reaper_check(hash_index);
3111 hash_index = (hash_index + 1) %
3112 *kgnilnd_tunables.kgn_peer_hash_size;
3114 next_check_time = (long) jiffies + cfs_time_seconds(p);
3115 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3117 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3120 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3122 kgnilnd_thread_fini();
3127 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3128 unsigned niov, offset, nob;
3130 lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
3131 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3134 lnet_copy_flat2kiov(
3137 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3139 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3146 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3149 gni_post_descriptor_t *desc;
3151 kgn_tx_ev_id_t ev_id;
3153 int should_retry, rc;
3154 long num_processed = 0;
3155 kgn_conn_t *conn = NULL;
3156 kgn_tx_t *tx = NULL;
3157 kgn_rdma_desc_t *rdesc;
3162 /* make sure we don't keep looping if we need to reset */
3163 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3164 return num_processed;
3166 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3168 /* we didn't get the mutex, so return that there is still work
3172 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3173 /* a bit gross - but we need a good way to test for
3174 * delayed RDMA completions and the easiest way to do
3175 * that is to delay the RDMA CQ events */
3176 rrc = GNI_RC_NOT_DONE;
3178 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3181 if (rrc == GNI_RC_NOT_DONE) {
3182 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3183 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3184 dev->gnd_id, num_processed);
3185 return num_processed;
3187 dev->gnd_sched_alive = jiffies;
3190 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3191 "this is bad, somehow our credits didn't protect us"
3192 " from CQ overrun\n");
3193 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3194 "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc,
3195 event_data, GNI_CQ_GET_TYPE(event_data));
3197 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3199 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3201 /* XXX Nic: Need better error handling here... */
3202 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3203 (rrc == GNI_RC_TRANSACTION_ERROR),
3206 ev_id.txe_cookie = desc->post_id;
3208 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3210 if (conn == NULL || tx == NULL) {
3211 /* either conn or tx was already nuked and this is a "late"
3212 * completion, so drop it */
3216 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3217 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3218 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3219 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3220 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3222 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3224 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3225 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3226 tx->tx_msg.gnm_u.completion.gncm_retval);
3230 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3231 if (tx->tx_buffer_copy != NULL)
3232 kgnilnd_recv_bte_get(tx);
3233 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3236 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3237 if (tx->tx_buffer_copy != NULL)
3238 kgnilnd_recv_bte_get(tx);
3239 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3242 /* remove from rdmaq */
3243 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3244 spin_lock(&conn->gnc_list_lock);
3245 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3246 spin_unlock(&conn->gnc_list_lock);
3247 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3249 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3250 atomic_inc(&dev->gnd_rdma_ntx);
3251 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3252 /* transaction succeeded, add into fmaq */
3253 kgnilnd_queue_tx(conn, tx);
3254 kgnilnd_peer_alive(conn->gnc_peer);
3256 /* drop ref from kgnilnd_validate_tx_ev_id */
3257 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3258 kgnilnd_conn_decref(conn);
3263 /* fall through to the TRANSACTION_ERROR case */
3266 /* get stringified version for log messages */
3267 kgnilnd_cq_error_str(event_data, &err_str, 256);
3268 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3270 /* make sure we are not off in the weeds with this tx */
3271 if (tx->tx_retrans >
3272 *kgnilnd_tunables.kgn_max_retransmits) {
3273 GNIDBG_TX(D_NETERROR, tx,
3274 "giving up on TX, too many retries", NULL);
3278 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3279 should_retry ? "transient" : "unrecoverable", err_str);
3281 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3282 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3283 rdesc = &tx->tx_putinfo.gnpam_desc;
3284 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3285 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3287 rdesc = &tx->tx_getinfo.gngm_desc;
3288 rnob = tx->tx_lntmsg[0]->msg_len;
3289 rcookie = tx->tx_getinfo.gngm_cookie;
3294 tx->tx_msg.gnm_type,
3298 kgnilnd_nak_rdma(conn,
3299 tx->tx_msg.gnm_type,
3302 tx->tx_msg.gnm_srcnid);
3303 kgnilnd_tx_done(tx, -EFAULT);
3304 kgnilnd_close_conn(conn, -ECOMM);
3307 /* drop ref from kgnilnd_validate_tx_ev_id */
3308 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3309 kgnilnd_conn_decref(conn);
3314 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3318 kgn_tx_ev_id_t ev_id;
3319 kgn_tx_t *tx = NULL;
3320 kgn_conn_t *conn = NULL;
3321 int queued_fma, saw_reply, rc;
3322 long num_processed = 0;
3325 /* make sure we don't keep looping if we need to reset */
3326 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3327 return num_processed;
3330 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3332 /* we didn't get the mutex, so return that there is still work
3337 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3338 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3340 if (rrc == GNI_RC_NOT_DONE) {
3342 "SMSG send CQ %d not ready (data "LPX64") "
3343 "processed %ld\n", dev->gnd_id, event_data,
3345 return num_processed;
3348 dev->gnd_sched_alive = jiffies;
3351 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3352 "this is bad, somehow our credits didn't "
3353 "protect us from CQ overrun\n");
3354 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3355 "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc,
3356 event_data, GNI_CQ_GET_TYPE(event_data));
3358 /* if SMSG couldn't handle an error, time for conn to die */
3359 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3362 /* need to take the write_lock to ensure atomicity
3363 * on the conn state if we need to close it */
3364 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3365 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3367 /* Conn was destroyed? */
3369 "SMSG CQID lookup "LPX64" failed\n",
3370 GNI_CQ_GET_INST_ID(event_data));
3371 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3375 kgnilnd_cq_error_str(event_data, &err_str, 256);
3376 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3377 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3379 kgnilnd_close_conn_locked(conn, -ECOMM);
3381 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3383 /* no need to process rest of this tx -
3384 * it is getting canceled */
3388 /* fall through to GNI_RC_SUCCESS case */
3389 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3391 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3392 if (conn == NULL || tx == NULL) {
3393 /* either conn or tx was already nuked and this is a "late"
3394 * completion, so drop it */
3398 tx->tx_conn->gnc_last_tx_cq = jiffies;
3399 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3400 set_mb(conn->gnc_last_noop_cq, jiffies);
3403 /* lock tx_list_state and tx_state */
3404 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3405 spin_lock(&tx->tx_conn->gnc_list_lock);
3407 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3408 "state not GNILND_TX_LIVE_FMAQ", NULL);
3409 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3410 "not waiting for completion", NULL);
3412 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3415 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3417 /* This will trigger other FMA sends that were
3418 * pending this completion */
3419 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3421 /* we either did not expect reply or we already got it */
3422 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3424 spin_unlock(&tx->tx_conn->gnc_list_lock);
3425 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3428 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3430 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3431 kgnilnd_schedule_conn(conn);
3434 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3435 * If saw_reply is true we know that the tx is safe to use as the other thread
3436 * is already finished with it.
3440 /* no longer need to track on the live_fmaq */
3441 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3443 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3444 /* we already got reply & were waiting for
3445 * completion of initial send */
3446 /* to initiate RDMA transaction */
3447 GNIDBG_TX(D_NET, tx,
3448 "Pending RDMA 0x%p type 0x%02x",
3449 tx->tx_msg.gnm_type);
3450 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3451 rc = kgnilnd_send_mapped_tx(tx, 0);
3452 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3454 /* we are done with this tx */
3455 GNIDBG_TX(D_NET, tx,
3456 "Done with tx type 0x%02x",
3457 tx->tx_msg.gnm_type);
3458 kgnilnd_tx_done(tx, tx->tx_rc);
3462 /* drop ref from kgnilnd_validate_tx_ev_id */
3463 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3464 kgnilnd_conn_decref(conn);
3466 /* if we are waiting for a REPLY, we'll handle the tx then */
3467 } /* end for loop */
3471 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3476 long num_processed = 0;
3477 struct list_head *conns;
3478 struct list_head *tmp;
3482 /* make sure we don't keep looping if we need to reset */
3483 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3484 return num_processed;
3487 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3489 /* we didn't get the mutex, so return that there is still work
3493 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3494 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3496 if (rrc == GNI_RC_NOT_DONE) {
3497 CDEBUG(D_INFO, "SMSG RX CQ %d empty data "LPX64" "
3499 dev->gnd_id, event_data, num_processed);
3500 return num_processed;
3502 dev->gnd_sched_alive = jiffies;
3505 /* this is the only CQ that can really handle transient
3507 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3508 rrc = cfs_fail_val ? cfs_fail_val
3509 : GNI_RC_ERROR_RESOURCE;
3510 if (rrc == GNI_RC_ERROR_RESOURCE) {
3511 /* set overrun too */
3512 event_data |= (1UL << 63);
3513 LASSERTF(GNI_CQ_OVERRUN(event_data),
3514 "(1UL << 63) is no longer the bit to"
3515 "set to indicate CQ_OVERRUN\n");
3518 /* sender should get error event too and take care
3519 of failed transaction by re-transmitting */
3520 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3521 CDEBUG(D_NET, "SMSG RX CQ error "LPX64"\n", event_data);
3525 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3526 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3527 conn = kgnilnd_cqid2conn_locked(
3528 GNI_CQ_GET_INST_ID(event_data));
3530 CDEBUG(D_NET, "SMSG RX CQID lookup "LPU64" "
3531 "failed, dropping event "LPX64"\n",
3532 GNI_CQ_GET_INST_ID(event_data),
3535 CDEBUG(D_NET, "SMSG RX: CQID "LPU64" "
3537 GNI_CQ_GET_INST_ID(event_data),
3538 conn, conn->gnc_peer ?
3539 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3542 conn->gnc_last_rx_cq = jiffies;
3544 /* stash first rx so we can clear out purgatory.
3546 if (conn->gnc_first_rx == 0) {
3547 conn->gnc_first_rx = jiffies;
3549 kgnilnd_peer_alive(conn->gnc_peer);
3550 kgnilnd_schedule_conn(conn);
3552 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3556 /* FMA CQ has overflowed: check ALL conns */
3557 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3558 "conns on device %d\n", dev->gnd_id);
3560 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3562 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3563 conns = &kgnilnd_data.kgn_conns[rc];
3565 list_for_each(tmp, conns) {
3566 conn = list_entry(tmp, kgn_conn_t,
3569 if (conn->gnc_device == dev) {
3570 kgnilnd_schedule_conn(conn);
3571 conn->gnc_last_rx_cq = jiffies;
3575 /* don't block write lockers for too long... */
3576 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3581 /* try_map_if_full should only be used when processing TX from list of
3582 * backlog TX waiting on mappings to free up
3585 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3586 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3589 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3591 /* slight bit of race if multiple people calling, but at worst we'll have
3592 * order altered just a bit... which would not be determenistic anyways */
3593 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3595 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3597 /* We know that we have a GART reservation that should guarantee forward progress.
3598 * This means we don't need to take any extraordinary efforts if we are failing
3599 * mappings here - even if we are holding a very small number of these. */
3601 if (try_map_if_full || (rc == 0)) {
3602 rc = kgnilnd_map_buffer(tx);
3605 /* rc should be 0 if we mapped successfully here, if non-zero
3606 * we are queueing */
3608 /* if try_map_if_full set, they handle requeuing */
3609 if (unlikely(try_map_if_full)) {
3612 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3613 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3614 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3615 /* make sure we wake up sched to run this */
3616 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3617 /* return 0 as this is now queued for later sending */
3622 switch (tx->tx_msg.gnm_type) {
3626 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3627 * remote node where the RDMA will be started
3628 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3629 * be satisified. The rest of the errors are "hard" errors that require
3630 * upper layers to handle themselves.
3631 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3632 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3633 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3634 * GNILND_MSG_GET_DONE_REV.
3636 case GNILND_MSG_GET_REQ:
3637 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3638 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3639 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3640 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3641 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3642 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3643 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3645 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3646 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3648 case GNILND_MSG_PUT_ACK:
3649 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3650 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3651 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3652 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3654 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3655 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3658 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3659 case GNILND_MSG_PUT_DONE:
3660 case GNILND_MSG_PUT_REQ:
3661 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3662 &tx->tx_putinfo.gnpam_desc,
3663 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3664 tx->tx_putinfo.gnpam_dst_cookie);
3665 RETURN(try_map_if_full ? rc : 0);
3667 case GNILND_MSG_GET_DONE:
3668 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3669 &tx->tx_getinfo.gngm_desc,
3670 tx->tx_lntmsg[0]->msg_len,
3671 tx->tx_getinfo.gngm_cookie);
3672 RETURN(try_map_if_full ? rc : 0);
3674 case GNILND_MSG_PUT_REQ_REV:
3675 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3676 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3677 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3678 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3679 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3680 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3681 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3683 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3685 case GNILND_MSG_PUT_DONE_REV:
3686 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3687 &tx->tx_getinfo.gngm_desc,
3689 tx->tx_getinfo.gngm_cookie);
3690 RETURN(try_map_if_full ? rc : 0);
3692 case GNILND_MSG_GET_ACK_REV:
3693 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3694 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3695 /* LNET_GETS are a special case for parse */
3696 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3697 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3699 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3700 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3702 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3703 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3705 case GNILND_MSG_GET_DONE_REV:
3706 case GNILND_MSG_GET_REQ_REV:
3707 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3708 &tx->tx_putinfo.gnpam_desc,
3709 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3710 tx->tx_putinfo.gnpam_dst_cookie);
3711 RETURN(try_map_if_full ? rc : 0);
3719 kgnilnd_process_fmaq(kgn_conn_t *conn)
3722 kgn_tx_t *tx = NULL;
3723 void *buffer = NULL;
3724 unsigned int nob = 0;
3727 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3728 * However I will be rescheduled by an FMA completion event
3729 * when I eventually get some.
3730 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3731 * But it doesn't matter if I try to send a "real" message just
3732 * as I start closing because I'll get scheduled to send the
3735 /* Short circuit if the ep_handle is null we cant send anyway. */
3736 if (conn->gnc_ephandle == NULL)
3739 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3741 spin_lock(&conn->gnc_list_lock);
3743 if (list_empty(&conn->gnc_fmaq)) {
3744 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3746 spin_unlock(&conn->gnc_list_lock);
3748 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3749 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3750 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3751 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3752 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3754 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3755 conn->gnc_last_noop_cq,
3756 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3757 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3758 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3759 atomic_inc(&conn->gnc_sched_noop);
3760 set_mb(conn->gnc_last_noop_want, jiffies);
3762 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3765 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3769 rc = kgnilnd_set_tx_id(tx, conn);
3771 kgnilnd_tx_done(tx, rc);
3777 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3778 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3779 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3780 more_to_do = !list_empty(&conn->gnc_fmaq);
3781 spin_unlock(&conn->gnc_list_lock);
3784 /* if there is no real TX or no NOOP to send, bail */
3789 if (!tx->tx_retrans)
3790 tx->tx_cred_wait = jiffies;
3792 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3793 "tx with zero id", NULL);
3795 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie "LPX64"\n",
3796 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3797 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3801 switch (tx->tx_msg.gnm_type) {
3805 case GNILND_MSG_NOOP:
3806 case GNILND_MSG_CLOSE:
3807 case GNILND_MSG_IMMEDIATE:
3808 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3809 buffer = tx->tx_buffer;
3813 case GNILND_MSG_GET_DONE:
3814 case GNILND_MSG_PUT_DONE:
3815 case GNILND_MSG_PUT_DONE_REV:
3816 case GNILND_MSG_GET_DONE_REV:
3817 case GNILND_MSG_PUT_NAK:
3818 case GNILND_MSG_GET_NAK:
3819 case GNILND_MSG_GET_NAK_REV:
3820 case GNILND_MSG_PUT_NAK_REV:
3821 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3824 case GNILND_MSG_PUT_REQ:
3825 case GNILND_MSG_GET_REQ_REV:
3826 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3828 case GNILND_MSG_PUT_ACK:
3829 case GNILND_MSG_PUT_REQ_REV:
3830 case GNILND_MSG_GET_ACK_REV:
3831 case GNILND_MSG_GET_REQ:
3832 /* This is really only to handle the retransmit of SMSG once these
3833 * two messages are setup in send_mapped_tx */
3834 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3838 if (likely(rc == 0)) {
3839 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3843 /* don't explicitly reschedule here - we are short credits and will rely on
3844 * kgnilnd_sendmsg to resched the conn if need be */
3846 } else if (rc < 0) {
3847 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3848 * almost certainly a software bug, but lets play nice with the other kids */
3849 kgnilnd_tx_done(tx, rc);
3850 /* just for fun, kick peer in arse - resetting conn might help to correct
3851 * this almost certainly buggy software caused return code */
3852 kgnilnd_close_conn(conn, rc);
3856 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3857 kgnilnd_schedule_conn(conn);
3862 kgnilnd_process_rdmaq(kgn_device_t *dev)
3867 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3871 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3872 unsigned long dead_bump;
3875 /* if we think we need to adjust, take lock to serialize and recheck */
3876 spin_lock(&dev->gnd_rdmaq_lock);
3877 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3878 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3880 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3882 /* roll the bucket forward */
3883 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3885 if (kgnilnd_data.kgn_rdmaq_override &&
3886 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3887 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3892 /* roll current outstanding forward to make sure we carry outstanding
3893 * committment forward
3894 * new_ok starts out as the whole interval value
3895 * - first subtract bytes_out from last interval, as that would push us over
3896 * strict limits for this interval
3897 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3899 * there is a small race here if someone is actively processing mappings and
3900 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3901 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3902 * - if this gives us problems in the future, we could use a read/write lock
3903 * to protect the resetting of these values */
3904 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3905 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3907 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3908 "current out %ld\n",
3909 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3910 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3912 spin_unlock(&dev->gnd_rdmaq_lock);
3915 spin_lock(&dev->gnd_rdmaq_lock);
3916 while (!list_empty(&dev->gnd_rdmaq)) {
3919 /* make sure we break out early on quiesce */
3920 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3921 /* always break with lock held - we unlock outside loop */
3925 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3926 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3929 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3930 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3931 /* if conn is dying, mark tx in tx_ref_table for
3932 * kgnilnd_complete_closed_conn to finish up */
3933 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3935 /* tx was moved to DYING, get next */
3938 spin_unlock(&dev->gnd_rdmaq_lock);
3940 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3941 spin_lock(&dev->gnd_rdmaq_lock);
3944 /* no ticket! add back to head */
3945 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3946 /* clear found_work so scheduler threads wait for timer */
3950 /* TX is GO for launch */
3951 tx->tx_qtime = jiffies;
3952 kgnilnd_send_mapped_tx(tx, 0);
3956 spin_unlock(&dev->gnd_rdmaq_lock);
3962 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3964 __swab64s(&d->gnrd_key.qword1);
3965 __swab64s(&d->gnrd_key.qword2);
3966 __swab64s(&d->gnrd_addr);
3967 __swab32s(&d->gnrd_nob);
3970 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3971 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3974 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3976 kgn_tx_ev_id_t ev_id;
3979 /* we use the cookie from the original TX, so we can find the match
3980 * by parsing that and using the txe_idx */
3981 ev_id.txe_cookie = cookie;
3983 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3986 /* check tx to make sure kgni didn't eat it */
3987 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3988 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3990 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
3991 (tx->tx_id.txe_cookie = cookie)),
3992 "conn 0x%p->%s tx_ref_table hosed: wanted "
3993 "txe_cookie "LPX64" txe_idx %d "
3994 "found tx %p cookie "LPX64" txe_idx %d\n",
3995 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
3996 cookie, ev_id.txe_idx,
3997 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
3999 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4000 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4001 "Unexpected TX type (%x, %x or %x) "
4002 "or state (%x, expected +%x) "
4003 "matched reply from %s\n",
4004 tx->tx_msg.gnm_type, type1, type2,
4005 tx->tx_state, GNILND_TX_WAITING_REPLY,
4006 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4008 CWARN("Unmatched reply %02x, or %02x/"LPX64" from %s\n",
4009 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4015 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4018 kgn_conn_t *conn = tx->tx_conn;
4019 __u64 nob = tx->tx_nob;
4020 __u32 physnop = tx->tx_phys_npages;
4021 int id = tx->tx_id.txe_smsg_id;
4022 int buftype = tx->tx_buftype;
4023 gni_mem_handle_t hndl;
4024 hndl.qword1 = tx->tx_map_key.qword1;
4025 hndl.qword2 = tx->tx_map_key.qword2;
4027 spin_lock(&conn->gnc_list_lock);
4029 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4030 "not waiting for reply", NULL);
4033 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4035 if (rc == -EFAULT) {
4036 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16"LPF64"u physnop %8d buffertype %#8x MemHandle "LPX64"."LPX64"x\n",
4037 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4039 if(*kgnilnd_tunables.kgn_efault_lbug) {
4040 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4041 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4043 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4044 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4045 kgnilnd_tx_state2str(tx->tx_list_state),
4046 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4051 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4052 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4053 /* sample under lock as follow on steps require gnc_list_lock
4054 * - or call kgnilnd_tx_done which requires no locks held over
4055 * call to lnet_finalize */
4058 spin_unlock(&conn->gnc_list_lock);
4061 kgnilnd_tx_done(tx, tx->tx_rc);
4066 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4069 kgn_conn_t *conn = tx->tx_conn;
4071 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4072 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4074 /* the gncm_retval is passed in for PUTs */
4075 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4076 msg->gnm_u.completion.gncm_retval);
4078 kgnilnd_complete_tx(tx, rc);
4082 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4090 kgn_peer_t *peer = conn->gnc_peer;
4093 __u16 tmp_cksum = 0, msg_cksum = 0;
4094 int repost = 1, saw_complete;
4095 unsigned long timestamp, newest_last_rx, timeout;
4099 /* Short circuit if the ep_handle is null.
4100 * It's likely that its about to be closed as stale.
4102 if (conn->gnc_ephandle == NULL)
4105 timestamp = jiffies;
4106 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4107 /* delay in jiffies - we are really concerned only with things that
4108 * result in a schedule() or really holding this off for long times .
4109 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4110 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4112 /* Resample current time as we have no idea how long it took to get the mutex */
4113 timestamp = jiffies;
4115 /* We check here when the last time we received an rx, we do this before
4116 * we call getnext in case the thread has been blocked for a while. If we
4117 * havent received an rx since our timeout value we close the connection
4118 * as we should assume the other side has closed the connection. This will
4119 * stop us from sending replies to a mailbox that is already in purgatory.
4122 timeout = cfs_time_seconds(conn->gnc_timeout);
4123 newest_last_rx = GNILND_LASTRX(conn);
4125 /* Error injection to validate that timestamp checking works and closing the conn */
4126 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4127 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4130 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4131 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4132 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4133 cfs_duration_sec(timestamp - newest_last_rx),
4134 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4135 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4137 kgnilnd_close_conn(conn, rc);
4141 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4143 if (rrc == GNI_RC_NOT_DONE) {
4144 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4145 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4149 /* Instead of asserting when we get mailbox corruption lets attempt to
4150 * close the conn and recover. We can put the conn/mailbox into
4151 * purgatory and let purgatory deal with the problem. If we see
4152 * this NETTERROR reported on production systems in large amounts
4153 * we will need to revisit the state machine to see if we can tighten
4154 * it up further to improve data protection.
4157 if (rrc == GNI_RC_INVALID_STATE) {
4158 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4159 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4160 "detected closing conn %p from peer %s\n", conn,
4161 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4163 kgnilnd_close_conn(conn, rc);
4167 LASSERTF(rrc == GNI_RC_SUCCESS,
4168 "bad rc %d on conn %p from peer %s\n",
4169 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4171 msg = (kgn_msg_t *)prefix;
4173 rx = kgnilnd_alloc_rx();
4175 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4176 kgnilnd_release_msg(conn);
4177 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4178 conn, libcfs_nid2str(peer->gnp_nid));
4182 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4184 timestamp = conn->gnc_last_rx;
4185 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4186 atomic_inc(&conn->gnc_rx_seq);
4188 conn->gnc_last_rx = jiffies;
4189 /* stash first rx so we can clear out purgatory
4191 if (conn->gnc_first_rx == 0)
4192 conn->gnc_first_rx = jiffies;
4194 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4195 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4196 kgnilnd_peer_alive(conn->gnc_peer);
4199 rx->grx_conn = conn;
4201 rx->grx_received = current_kernel_time();
4203 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4206 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4212 kgnilnd_net_decref(net);
4215 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4216 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4218 /* XXX Nic: Do we need to swab cksum */
4219 if (msg->gnm_cksum != 0) {
4220 msg_cksum = msg->gnm_cksum;
4222 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4224 if (tmp_cksum != msg_cksum) {
4225 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4226 tmp_cksum, msg_cksum);
4227 kgnilnd_dump_msg(D_BUFFS, msg);
4232 /* restore checksum for future debug messages */
4233 msg->gnm_cksum = tmp_cksum;
4235 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4236 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4237 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4238 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4243 __swab32s(&msg->gnm_magic);
4244 __swab16s(&msg->gnm_version);
4245 __swab16s(&msg->gnm_type);
4246 __swab64s(&msg->gnm_srcnid);
4247 __swab64s(&msg->gnm_connstamp);
4248 __swab32s(&msg->gnm_seq);
4250 /* NB message type checked below; NOT here... */
4251 switch (msg->gnm_type) {
4252 case GNILND_MSG_GET_ACK_REV:
4253 case GNILND_MSG_PUT_ACK:
4254 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4257 case GNILND_MSG_PUT_REQ_REV:
4258 case GNILND_MSG_GET_REQ:
4259 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4267 if (msg->gnm_version != GNILND_MSG_VERSION) {
4268 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4269 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4274 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4275 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4276 libcfs_nid2str(msg->gnm_srcnid),
4277 libcfs_nid2str(peer->gnp_nid));
4282 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4283 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp "LPX64"("LPX64
4284 " expected) from %s",
4285 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4286 libcfs_nid2str(peer->gnp_nid));
4291 if (msg->gnm_seq != seq) {
4292 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4293 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4298 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4300 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4301 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4302 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4303 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4304 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4305 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4306 /* double check state with lock held */
4307 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4308 /* only error if we are not already closing */
4309 if (conn->gnc_peer_error == -ETIMEDOUT) {
4310 unsigned long now = jiffies;
4311 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4313 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4314 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4315 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4317 cfs_duration_sec(now - timestamp),
4318 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4319 atomic_read(&conn->gnc_tx_seq),
4320 cfs_duration_sec(now - conn->gnc_last_tx),
4321 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4322 cfs_duration_sec(now - conn->gnc_last_noop_want),
4323 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4324 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4325 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4326 cfs_duration_sec(now - conn->gnc_last_sched_do),
4327 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4329 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4331 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4335 if (conn->gnc_close_recvd) {
4336 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4337 kgnilnd_msgtype2str(msg->gnm_type),
4338 msg->gnm_type, conn->gnc_close_recvd,
4339 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4344 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4345 /* XXX Nic: log message received on bad connection state */
4349 switch (msg->gnm_type) {
4350 case GNILND_MSG_NOOP:
4351 /* Nothing to do; just a keepalive */
4354 case GNILND_MSG_IMMEDIATE:
4355 /* only get SMSG payload for IMMEDIATE */
4356 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4357 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4358 msg->gnm_srcnid, rx, 0);
4361 case GNILND_MSG_GET_REQ_REV:
4362 case GNILND_MSG_PUT_REQ:
4363 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4364 msg->gnm_srcnid, rx, 1);
4367 case GNILND_MSG_GET_NAK_REV:
4368 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4369 msg->gnm_u.completion.gncm_cookie);
4373 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4375 case GNILND_MSG_PUT_NAK:
4376 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4377 msg->gnm_u.completion.gncm_cookie);
4381 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4383 case GNILND_MSG_PUT_ACK:
4384 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4385 msg->gnm_u.putack.gnpam_src_cookie);
4389 /* store putack data for later: deferred rdma or re-try */
4390 tx->tx_putinfo = msg->gnm_u.putack;
4393 spin_lock(&tx->tx_conn->gnc_list_lock);
4395 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4396 "not waiting for reply", NULL);
4398 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4400 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4401 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4402 /* sample under lock as follow on steps require gnc_list_lock
4403 * - or call kgnilnd_tx_done which requires no locks held over
4404 * call to lnet_finalize */
4407 /* cannot launch rdma if still waiting for fma-msg completion */
4408 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4409 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4410 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4412 spin_unlock(&tx->tx_conn->gnc_list_lock);
4415 rc = kgnilnd_send_mapped_tx(tx, 0);
4417 kgnilnd_tx_done(tx, rc);
4420 case GNILND_MSG_GET_ACK_REV:
4421 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4422 msg->gnm_u.putack.gnpam_src_cookie);
4426 /* store putack data for later: deferred rdma or re-try */
4427 tx->tx_putinfo = msg->gnm_u.putack;
4429 spin_lock(&tx->tx_conn->gnc_list_lock);
4431 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4432 "not waiting for reply", NULL);
4434 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4436 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4437 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4438 /* sample under lock as follow on steps require gnc_list_lock
4439 * - or call kgnilnd_tx_done which requires no locks held over
4440 * call to lnet_finalize */
4443 /* cannot launch rdma if still waiting for fma-msg completion */
4444 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4445 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4446 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4448 spin_unlock(&tx->tx_conn->gnc_list_lock);
4451 rc = kgnilnd_send_mapped_tx(tx, 0);
4453 kgnilnd_tx_done(tx, rc);
4456 case GNILND_MSG_PUT_DONE:
4457 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4458 msg->gnm_u.completion.gncm_cookie);
4462 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4463 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4464 "bad tx buftype %d", tx->tx_buftype);
4466 kgnilnd_finalize_rx_done(tx, msg);
4468 case GNILND_MSG_PUT_REQ_REV:
4469 case GNILND_MSG_GET_REQ:
4470 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4471 msg->gnm_srcnid, rx, 1);
4475 case GNILND_MSG_GET_NAK:
4476 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4477 msg->gnm_u.completion.gncm_cookie);
4481 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4482 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4483 "bad tx buftype %d", tx->tx_buftype);
4485 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4488 case GNILND_MSG_GET_DONE:
4489 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4490 msg->gnm_u.completion.gncm_cookie);
4494 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4495 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4496 "bad tx buftype %d", tx->tx_buftype);
4498 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4499 msg->gnm_u.completion.gncm_retval);
4501 kgnilnd_finalize_rx_done(tx, msg);
4503 case GNILND_MSG_GET_DONE_REV:
4504 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4505 msg->gnm_u.completion.gncm_cookie);
4509 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4510 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4511 "bad tx buftype %d", tx->tx_buftype);
4513 kgnilnd_finalize_rx_done(tx, msg);
4516 case GNILND_MSG_PUT_DONE_REV:
4517 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4518 msg->gnm_u.completion.gncm_cookie);
4523 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4524 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4525 "bad tx buftype %d", tx->tx_buftype);
4527 kgnilnd_finalize_rx_done(tx, msg);
4529 case GNILND_MSG_PUT_NAK_REV:
4530 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4531 msg->gnm_u.completion.gncm_cookie);
4536 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4537 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4538 "bad tx buftype %d", tx->tx_buftype);
4540 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4545 if (rc < 0) /* protocol/comms error */
4546 kgnilnd_close_conn(conn, rc);
4548 if (repost && rx != NULL) {
4549 kgnilnd_consume_rx(rx);
4552 /* we got an event so assume more there and call for reschedule */
4554 kgnilnd_schedule_conn(conn);
4558 /* Do the failure injections that we need to affect conn processing in the following function.
4559 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4560 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4563 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4567 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4571 /* short circuit out when not set */
4572 if (likely(!cfs_fail_loc)) {
4576 /* failure injection to test for stack reset clean ups */
4577 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4578 /* we can't rely on busy loops being nice enough to get the
4579 * stack reset triggered - it'd just spin on this conn */
4580 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4583 GOTO(did_fail_loc, rc);
4586 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4587 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4589 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4590 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4593 GOTO(did_fail_loc, rc);
4597 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4598 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4599 * will be closed by kgnilnd_complete_closed_conn.
4601 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4602 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4605 GOTO(did_fail_loc, rc);
4608 /* this one is a bit gross - we can't hold the mutex from process_conns
4609 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4610 * and moving onto finish_connect
4611 * so, we'll just set the rc - kgnilnd_process_conns will clear
4612 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4613 * and effectively getting this thread to sleep */
4614 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4617 GOTO(did_fail_loc, rc);
4625 kgnilnd_send_conn_close(kgn_conn_t *conn)
4629 /* we are closing the conn - we will try to send the CLOSE msg
4630 * but will not wait for anything else to flush */
4632 /* send the close if not already done so or received one */
4633 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4634 /* set close_sent regardless of the success of the
4635 * CLOSE message. We are going to try once and then
4636 * kick him out of the sandbox */
4637 conn->gnc_close_sent = 1;
4640 /* EP might be null already if remote side initiated a new connection.
4641 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4642 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4644 if (conn->gnc_ephandle != NULL) {
4647 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4649 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4650 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4651 tx->tx_qtime = jiffies;
4653 if (tx->tx_id.txe_idx == 0) {
4654 rc = kgnilnd_set_tx_id(tx, conn);
4656 kgnilnd_tx_done(tx, rc);
4660 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4663 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4664 kgnilnd_tx_done(tx, -EAGAIN);
4666 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4668 /* It wasnt sent and we dont care. */
4669 kgnilnd_tx_done(tx, rc);
4677 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4678 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4679 conn->gnc_state = GNILND_CONN_CLOSED;
4680 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4681 /* mark this conn as CLOSED now that we processed it
4682 * do after TX, so we can use CLOSING in asserts */
4686 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4687 /* simulate a RX CLOSE after the timeout but before
4688 * the scheduler thread gets it */
4689 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4690 conn->gnc_peer_error = -ETIMEDOUT;
4692 /* schedule to allow potential CLOSE and get the complete phase run */
4693 kgnilnd_schedule_conn(conn);
4697 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4702 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4703 int log_retrans, log_retrans_level;
4704 static int last_map_version;
4707 spin_lock(&dev->gnd_lock);
4708 if (list_empty(&dev->gnd_map_tx)) {
4709 /* if the list is empty make sure we dont have a timer running */
4710 del_singleshot_timer_sync(&dev->gnd_map_timer);
4711 spin_unlock(&dev->gnd_lock);
4715 dev->gnd_sched_alive = jiffies;
4717 /* we'll retry as fast as possible up to 25% of the limit, then we start
4718 * backing off until our map version changes - indicating we unmapped
4720 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4721 if (likely(dev->gnd_map_attempt == 0) ||
4722 time_after_eq(jiffies, dev->gnd_next_map) ||
4723 last_map_version != dev->gnd_map_version) {
4725 /* if this is our first attempt at mapping set last mapped to current
4726 * jiffies so we can timeout our attempt correctly.
4728 if (dev->gnd_map_attempt == 0)
4729 dev->gnd_last_map = jiffies;
4731 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4732 spin_unlock(&dev->gnd_lock);
4736 /* delete the previous timer if it exists */
4737 del_singleshot_timer_sync(&dev->gnd_map_timer);
4738 /* stash the last map version to let us know when a good one was seen */
4739 last_map_version = dev->gnd_map_version;
4741 /* we need to to take the lock and continually refresh the head of the list as
4742 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4743 * allowing them to squeeze in */
4745 while (!list_empty(&dev->gnd_map_tx)) {
4746 /* make sure we break out early on quiesce */
4747 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4748 /* always break with lock held - we unlock outside loop */
4752 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4754 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4757 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4758 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4759 /* if conn is dying, mark tx in tx_ref_table for
4760 * kgnilnd_complete_closed_conn to finish up */
4761 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4764 /* tx was moved to DYING, get next */
4768 spin_unlock(&dev->gnd_lock);
4769 rc = kgnilnd_send_mapped_tx(tx, 1);
4771 /* We made it! skip error handling.. */
4773 /* OK to continue on +ve errors as it won't get seen until
4774 * this function is called again - we operate on a copy of the original
4775 * list and not the live list */
4776 spin_lock(&dev->gnd_lock);
4777 /* reset map attempts back to zero we successfully
4778 * mapped so we can reset our timers */
4779 dev->gnd_map_attempt = 0;
4781 } else if (rc == -EAGAIN) {
4782 spin_lock(&dev->gnd_lock);
4783 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4784 spin_unlock(&dev->gnd_lock);
4785 GOTO(get_out_mapped, rc);
4786 } else if (rc != -ENOMEM) {
4787 /* carp, failure we can't handle */
4788 kgnilnd_tx_done(tx, rc);
4789 spin_lock(&dev->gnd_lock);
4790 /* reset map attempts back to zero we dont know what happened but it
4791 * wasnt a failed mapping
4793 dev->gnd_map_attempt = 0;
4797 /* time to handle the retry cases.. lock so we dont have 2 threads
4798 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4800 spin_lock(&dev->gnd_lock);
4801 dev->gnd_map_attempt++;
4802 if (dev->gnd_map_attempt < fast_remaps) {
4803 /* do nothing we just want it to go as fast as possible.
4804 * just set gnd_next_map to current jiffies so it will process
4805 * as fast as possible.
4807 dev->gnd_next_map = jiffies;
4809 /* Retry based on GNILND_MAP_RETRY_RATE */
4810 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4813 /* only log occasionally once we've retried fast_remaps */
4814 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4815 ((dev->gnd_map_attempt % fast_remaps) == 0);
4816 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4818 /* make sure we are not off in the weeds with this tx */
4819 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4820 GNIDBG_TX(D_NETERROR, tx,
4821 "giving up on TX, too many retries", NULL);
4822 spin_unlock(&dev->gnd_lock);
4823 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4824 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4825 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4827 tx->tx_putinfo.gnpam_dst_cookie,
4828 tx->tx_msg.gnm_srcnid);
4830 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4832 tx->tx_getinfo.gngm_cookie,
4833 tx->tx_msg.gnm_srcnid);
4835 kgnilnd_tx_done(tx, -ENOMEM);
4836 GOTO(get_out_mapped, rc);
4838 GNIDBG_TX(log_retrans_level, tx,
4839 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4841 "nq_map %d mdd# %d/%d GART %ld",
4842 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4843 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4844 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4845 atomic_read(&dev->gnd_nq_map),
4846 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4847 atomic64_read(&dev->gnd_nbytes_map));
4850 /* we need to stop processing the rest of the list, so add it back in */
4851 /* set timer to wake device when we need to schedule this tx */
4852 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4853 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4854 spin_unlock(&dev->gnd_lock);
4855 GOTO(get_out_mapped, rc);
4857 spin_unlock(&dev->gnd_lock);
4863 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4868 int error_inject = 0;
4872 spin_lock(&dev->gnd_lock);
4873 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4874 dev->gnd_sched_alive = jiffies;
4878 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4879 /* break with lock held */
4883 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4884 list_del_init(&conn->gnc_schedlist);
4885 spin_unlock(&dev->gnd_lock);
4887 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4889 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4890 conn_sched != GNILND_CONN_PROCESS,
4891 "conn %p on ready list but in bad state: %d\n",
4894 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4895 conn, kgnilnd_conn_state2str(conn));
4898 set_mb(conn->gnc_last_sched_do, jiffies);
4900 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4902 /* based on intent see if we should run again. */
4903 rc = kgnilnd_schedule_process_conn(conn, intent);
4905 /* drop ref from gnd_ready_conns */
4906 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4907 down_write(&dev->gnd_conn_sem);
4908 kgnilnd_conn_decref(conn);
4909 up_write(&dev->gnd_conn_sem);
4910 } else if (rc != 1) {
4911 kgnilnd_conn_decref(conn);
4913 /* clear this so that scheduler thread doesn't spin */
4915 /* break with lock held... */
4916 spin_lock(&dev->gnd_lock);
4920 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4921 down_write(&dev->gnd_conn_sem);
4923 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4924 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4925 /* If there are tx's currently in use in another
4926 * thread we dont want to complete the close
4927 * yet. Cycle this conn back through
4929 kgnilnd_schedule_conn(conn);
4931 kgnilnd_complete_closed_conn(conn);
4933 up_write(&dev->gnd_conn_sem);
4934 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4935 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4936 /* serialize SMSG CQs with ep_bind and smsg_release */
4937 down_write(&dev->gnd_conn_sem);
4938 kgnilnd_destroy_conn_ep(conn);
4939 up_write(&dev->gnd_conn_sem);
4940 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4941 /* if we need to do some CLOSE sending, etc done here do it */
4942 down_write(&dev->gnd_conn_sem);
4943 kgnilnd_send_conn_close(conn);
4944 kgnilnd_check_fma_rx(conn);
4945 up_write(&dev->gnd_conn_sem);
4946 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4947 /* start moving traffic if the old conns are cleared out */
4948 down_read(&dev->gnd_conn_sem);
4949 kgnilnd_check_fma_rx(conn);
4950 kgnilnd_process_fmaq(conn);
4951 up_read(&dev->gnd_conn_sem);
4954 rc = kgnilnd_schedule_process_conn(conn, 0);
4956 /* drop ref from gnd_ready_conns */
4957 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4958 down_write(&dev->gnd_conn_sem);
4959 kgnilnd_conn_decref(conn);
4960 up_write(&dev->gnd_conn_sem);
4961 } else if (rc != 1) {
4962 kgnilnd_conn_decref(conn);
4965 /* check list again with lock held */
4966 spin_lock(&dev->gnd_lock);
4969 /* If we are short circuiting due to timing we want to be scheduled
4970 * as soon as possible.
4972 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4975 spin_unlock(&dev->gnd_lock);
4981 kgnilnd_scheduler(void *arg)
4983 int threadno = (long)arg;
4986 unsigned long deadline = 0;
4989 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
4991 cfs_block_allsigs();
4993 /* all gnilnd threads need to run fairly urgently */
4994 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
4995 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
4996 while (!kgnilnd_data.kgn_shutdown) {
4998 /* Safe: kgn_shutdown only set when quiescent */
5000 /* to quiesce or to not quiesce, that is the question */
5002 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5003 KGNILND_SPIN_QUIESCE;
5006 /* tracking for when thread goes AWOL */
5007 dev->gnd_sched_alive = jiffies;
5009 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5010 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5011 /* let folks know we are up and kicking
5012 * - they can use this for latency savings, etc
5013 * - only change if IRQ, if IDLE leave alone as that
5014 * schedule_device calls to put us back to IRQ */
5015 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5017 down_read(&dev->gnd_conn_sem);
5018 /* always check these - they are super low cost */
5019 found_work += kgnilnd_check_fma_send_cq(dev);
5020 found_work += kgnilnd_check_fma_rcv_cq(dev);
5022 /* rdma CQ doesn't care about eps */
5023 found_work += kgnilnd_check_rdma_cq(dev);
5025 /* move some RDMA ? */
5026 found_work += kgnilnd_process_rdmaq(dev);
5028 /* map some pending RDMA requests ? */
5029 found_work += kgnilnd_process_mapped_tx(dev);
5031 /* the EP for a conn is not destroyed until all the references
5032 * to it are gone, so these checks should be safe
5033 * even if run in parallel with the CQ checking functions
5034 * _AND_ a thread that processes the CLOSED->DONE
5038 up_read(&dev->gnd_conn_sem);
5040 /* process all conns ready now */
5041 found_work += kgnilnd_process_conns(dev, deadline);
5043 /* do an eager check to avoid the IRQ disabling in
5044 * prepare_to_wait and friends */
5047 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5048 time_before(jiffies, deadline)) {
5050 if ((busy_loops % 10) == 0) {
5051 /* tickle heartbeat and watchdog to ensure our
5052 * piggishness doesn't turn into heartbeat failure */
5053 touch_nmi_watchdog();
5059 /* if we got here, found_work was zero or busy_loops means we
5060 * need to take a break. We'll clear gnd_ready but we'll check
5061 * one last time if there is an IRQ that needs processing */
5063 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5065 /* the first time this will go LOOP -> IDLE and let us do one final check
5066 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5067 * - this might allow other threads to block us for a bit if they
5068 * try to get the mutex, but that is good as we'd need to wake
5069 * up soon to handle the CQ or other processing anyways */
5071 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5073 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5074 time_after_eq(jiffies, deadline)) {
5076 "yeilding: found_work %d busy_loops %d\n",
5077 found_work, busy_loops);
5079 /* use yield if we are bailing due to busy_loops
5080 * - this will ensure we wake up soonish. This closes
5081 * a race with kgnilnd_device_callback - where it'd
5082 * not call wake_up() because gnd_ready == 1, but then
5083 * we come down and schedule() because of busy_loops.
5084 * We'd not be woken up until something poked our waitq
5085 * again. yield() ensures we wake up without another
5086 * waitq poke in that case */
5087 atomic_inc(&dev->gnd_n_yield);
5088 kgnilnd_data.kgn_last_condresched = jiffies;
5090 CDEBUG(D_INFO, "awake after yeild\n");
5091 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5092 } else if (found_work == GNILND_DEV_IDLE) {
5093 /* busy_loops is low and there is nothing to do,
5094 * go to sleep and wait for a waitq poke */
5096 "scheduling: found_work %d busy_loops %d\n",
5097 found_work, busy_loops);
5098 atomic_inc(&dev->gnd_n_schedule);
5099 kgnilnd_data.kgn_last_scheduled = jiffies;
5101 CDEBUG(D_INFO, "awake after schedule\n");
5102 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5104 finish_wait(&dev->gnd_waitq, &wait);
5107 kgnilnd_thread_fini();