2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by Eric Barton <eric@bartonsoftware.com>
7 * Author: James Shimek <jshimek@cray.com>
8 * Author: Nic Henke <nic@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #include <linux/nmi.h>
29 #include <linux/pagemap.h>
32 /* this is useful when needed to debug wire corruption. */
34 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
42 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
43 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
46 } else if (len >= 16) {
48 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
49 prefix, ptr, *(ptr), *(ptr + 1));
53 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
62 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
64 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
65 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
66 msg->gnm_magic, msg->gnm_version,
67 msg->gnm_type, msg->gnm_srcnid,
68 msg->gnm_connstamp, msg->gnm_seq,
69 msg->gnm_cksum, msg->gnm_payload_cksum,
70 msg->gnm_payload_len);
74 kgnilnd_schedule_device(kgn_device_t *dev)
76 short already_live = 0;
78 /* we'll only want to wake if the scheduler thread
79 * has come around and set ready to zero */
80 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
83 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, int lock_held)
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);
191 spin_unlock(&dev->gnd_lock);
192 set_mb(conn->gnc_last_sched_ask, jiffies);
195 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
199 /* make sure thread(s) going to process conns - but let it make
200 * separate decision from conn schedule */
202 kgnilnd_schedule_device(dev);
207 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
209 kgn_device_t *dev = conn->gnc_device;
211 spin_lock(&dev->gnd_lock);
212 if (list_empty(&conn->gnc_delaylist)) {
213 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
216 spin_unlock(&dev->gnd_lock);
218 kgnilnd_schedule_device(dev);
223 kgnilnd_schedule_dgram(kgn_device_t *dev)
227 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
228 if (wake != GNILND_DGRAM_SCHED) {
229 wake_up(&dev->gnd_dgram_waitq);
231 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
236 kgnilnd_free_tx(kgn_tx_t *tx)
238 /* taken from kgnilnd_tx_add_state_locked */
240 LASSERTF((tx->tx_list_p == NULL &&
241 tx->tx_list_state == GNILND_TX_ALLOCD) &&
242 list_empty(&tx->tx_list),
243 "tx %p with bad state %s (list_p %p) tx_list %s\n",
244 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
245 list_empty(&tx->tx_list) ? "empty" : "not empty");
247 atomic_dec(&kgnilnd_data.kgn_ntx);
249 /* we only allocate this if we need to */
250 if (tx->tx_phys != NULL) {
251 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
252 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
253 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
256 /* Only free the buffer if we used it */
257 if (tx->tx_buffer_copy != NULL) {
258 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
259 tx->tx_buffer_copy = NULL;
260 CDEBUG(D_MALLOC, "vfreed buffer2\n");
263 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
265 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
266 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
270 kgnilnd_alloc_tx (void)
274 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
277 tx = kmem_cache_alloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
279 CERROR("failed to allocate tx\n");
282 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
285 /* need this memset, cache alloc'd memory is not cleared */
286 memset(tx, 0, sizeof(*tx));
288 /* setup everything here to minimize time under the lock */
289 tx->tx_buftype = GNILND_BUF_NONE;
290 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
291 INIT_LIST_HEAD(&tx->tx_list);
292 INIT_LIST_HEAD(&tx->tx_map_list);
293 tx->tx_list_state = GNILND_TX_ALLOCD;
295 atomic_inc(&kgnilnd_data.kgn_ntx);
300 /* csum_fold needs to be run on the return value before shipping over the wire */
301 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
303 /* we don't use offset as every one is passing a buffer reference that already
304 * includes the offset into the base address -
305 * see kgnilnd_setup_virt_buffer and kgnilnd_setup_immediate_buffer */
307 kgnilnd_cksum(void *ptr, size_t nob)
311 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
313 /* don't use magic 'no checksum' value */
317 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
324 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
325 unsigned int offset, unsigned int nob, int dump_blob)
331 unsigned int fraglen;
337 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
338 kiov, nkiov, offset, nob, dump_blob);
340 /* if loops changes, please change kgnilnd_setup_phys_buffer */
342 while (offset >= kiov->bv_len) {
343 offset -= kiov->bv_len;
349 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
350 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
352 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
353 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
355 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
356 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
358 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
359 odd, kiov[0].bv_len, offset, nob);
361 for (i = 0; i < nkiov; i++) {
362 pages[i] = kiov[i].bv_page;
365 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
367 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
369 /* return zero to avoid killing tx - we'll just get warning on console
370 * when remote end sees zero checksum */
373 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
375 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
380 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
381 (void *)addr + kiov[0].bv_offset +
384 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
385 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
389 fraglen = min(kiov->bv_len - offset, nob);
391 /* make dang sure we don't send a bogus checksum if somehow we get
392 * an odd length fragment on anything but the last entry in a kiov -
393 * we know from kgnilnd_setup_rdma_buffer that we can't have non
394 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
395 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
396 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
397 fraglen, nkiov, nob, kiov->bv_len,
400 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
402 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
405 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
406 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
407 addr, fraglen, offset);
412 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
414 kunmap(kiov->bv_page);
421 /* iov must not run out before end of data */
422 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
427 retsum = csum_fold(cksum);
429 /* don't use magic 'no checksum' value */
433 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
439 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
441 msg->gnm_magic = GNILND_MSG_MAGIC;
442 msg->gnm_version = GNILND_MSG_VERSION;
443 msg->gnm_type = type;
444 msg->gnm_payload_len = 0;
445 msg->gnm_srcnid = source;
446 /* gnm_connstamp gets set when FMA is sent */
447 /* gnm_srcnid is set on creation via function argument
448 * The right interface/net and nid is passed in when the message
454 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
456 kgn_tx_t *tx = kgnilnd_alloc_tx();
459 kgnilnd_init_msg(&tx->tx_msg, type, source);
461 CERROR("couldn't allocate new tx type %s!\n",
462 kgnilnd_msgtype2str(type));
469 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
475 case GNILND_MSG_GET_REQ:
476 case GNILND_MSG_GET_DONE:
477 nak_type = GNILND_MSG_GET_NAK;
479 case GNILND_MSG_PUT_REQ:
480 case GNILND_MSG_PUT_ACK:
481 case GNILND_MSG_PUT_DONE:
482 nak_type = GNILND_MSG_PUT_NAK;
484 case GNILND_MSG_PUT_REQ_REV:
485 case GNILND_MSG_PUT_DONE_REV:
486 nak_type = GNILND_MSG_PUT_NAK_REV;
488 case GNILND_MSG_GET_REQ_REV:
489 case GNILND_MSG_GET_ACK_REV:
490 case GNILND_MSG_GET_DONE_REV:
491 nak_type = GNILND_MSG_GET_NAK_REV;
494 CERROR("invalid msg type %s (%d)\n",
495 kgnilnd_msgtype2str(rx_type), rx_type);
498 /* only allow NAK on error and truncate to zero */
499 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
500 error, conn, cookie);
502 tx = kgnilnd_new_tx_msg(nak_type, source);
504 CNETERR("can't get TX to NAK RDMA to %s\n",
505 libcfs_nid2str(conn->gnc_peer->gnp_nid));
509 tx->tx_msg.gnm_u.completion.gncm_retval = error;
510 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
511 kgnilnd_queue_tx(conn, tx);
515 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
516 struct kvec *iov, struct bio_vec *kiov,
517 unsigned int offset, unsigned int nob)
519 kgn_msg_t *msg = &tx->tx_msg;
522 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
523 * gni_smsg_send to send that as the payload */
525 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
528 tx->tx_buffer = NULL;
529 } else if (kiov != NULL) {
531 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
532 niov = round_up(nob + offset + kiov->bv_offset,
536 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
537 "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n",
538 niov, msg, kiov, iov, offset, nob);
540 while (offset >= kiov->bv_len) {
541 offset -= kiov->bv_len;
546 for (i = 0; i < niov; i++) {
547 /* We can't have a bv_offset on anything but the first
548 * entry, otherwise we'll have a hole at the end of the
549 * mapping as we only map whole pages.
550 * Also, if we have a bv_len < PAGE_SIZE but we need to
551 * map more than bv_len, we will also have a whole at
552 * the end of that page which isn't allowed
554 if ((kiov[i].bv_offset != 0 && i > 0) ||
555 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
557 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
558 i, offset, nob, kiov->bv_offset,
562 tx->tx_imm_pages[i] = kiov[i].bv_page;
565 /* hijack tx_phys for the later unmap */
567 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
569 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
570 kiov[0].bv_offset + offset;
571 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
572 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
573 nob, kiov, tx->tx_buffer);
575 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
576 if (tx->tx_phys == NULL) {
577 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
581 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
582 /* make sure we take into account the kiov offset as the
583 * start of the buffer
585 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
588 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
589 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
591 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
595 /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
596 * could "flatten" the payload into a single contiguous buffer ready
597 * for sending direct over an FMA if we ever needed to. */
601 while (offset >= iov->iov_len) {
602 offset -= iov->iov_len;
608 if (nob > iov->iov_len - offset) {
609 CERROR("Can't handle multiple vaddr fragments\n");
613 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
615 tx->tx_buftype = GNILND_BUF_IMMEDIATE;
619 /* checksum payload early - it shouldn't be changing after lnd_send */
620 if (*kgnilnd_tunables.kgn_checksum >= 2) {
621 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
622 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
623 msg->gnm_payload_cksum += 0xe00e;
625 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
626 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
630 msg->gnm_payload_cksum = 0;
637 kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
638 unsigned int niov, struct kvec *iov,
639 unsigned int offset, unsigned int nob)
644 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
646 while (offset >= iov->iov_len) {
647 offset -= iov->iov_len;
653 if (nob > iov->iov_len - offset) {
654 CERROR("Can't handle multiple vaddr fragments\n");
658 tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
660 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
665 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
666 unsigned int offset, unsigned int nob)
668 gni_mem_segment_t *phys;
670 unsigned int fraglen;
672 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
676 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
678 /* only allocate this if we are going to use it */
679 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
681 if (tx->tx_phys == NULL) {
682 CERROR("failed to allocate tx_phys\n");
687 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
688 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
690 /* if loops changes, please change kgnilnd_cksum_kiov
691 * and kgnilnd_setup_immediate_buffer */
693 while (offset >= kiov->bv_len) {
694 offset -= kiov->bv_len;
700 /* at this point, kiov points to the first page that we'll actually map
701 * now that we've seeked into the koiv for offset and dropped any
702 * leading pages that fall entirely within the offset */
703 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
706 /* bv_offset is start of 'valid' buffer, so index offset past that */
707 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
710 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
711 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
714 fraglen = min(kiov->bv_len - offset, nob);
716 /* We can't have a bv_offset on anything but the first entry,
717 * otherwise we'll have a hole at the end of the mapping as we
718 * only map whole pages. Only the first page is allowed to
719 * have an offset - we'll add that into tx->tx_buffer and that
720 * will get used when we map in the segments (see
721 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
722 * but we need to map more than bv_len, we will also have a
723 * whole at the end of that page which isn't allowed
725 if ((phys != tx->tx_phys) &&
726 ((kiov->bv_offset != 0) ||
727 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
728 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
729 (int)(phys - tx->tx_phys),
730 offset, nob, kiov->bv_offset, kiov->bv_len);
735 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
736 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
741 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
747 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
748 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
751 phys->address = page_to_phys(kiov->bv_page);
758 /* iov must not run out before end of data */
759 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
763 tx->tx_phys_npages = phys - tx->tx_phys;
768 if (tx->tx_phys != NULL) {
769 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
770 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
771 sizeof(*tx->tx_phys), tx->tx_phys);
778 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
779 struct kvec *iov, struct bio_vec *kiov,
780 unsigned int offset, unsigned int nob)
784 LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
785 " offset %d, nob %d, niov %d\n"
786 , iov, kiov, tx, offset, nob, niov);
789 rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
791 rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
796 /* kgnilnd_parse_lnet_rdma()
797 * lntmsg - message passed in from lnet.
798 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
799 * nob - actual number of bytes to in this message.
800 * put_len - It is possible for PUTs to have a different length than the
801 * length stored in lntmsg->msg_len since LNET can adjust this
802 * length based on it's buffer size and offset.
803 * lnet_try_match_md() sets the mlength that we use to do the RDMA
807 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
808 unsigned int *offset, unsigned int *nob,
809 struct bio_vec **kiov, int put_len)
811 /* GETs are weird, see kgnilnd_send */
812 if (lntmsg->msg_type == LNET_MSG_GET) {
813 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
816 *kiov = lntmsg->msg_md->md_iov.kiov;
818 *niov = lntmsg->msg_md->md_niov;
819 *nob = lntmsg->msg_md->md_length;
822 *kiov = lntmsg->msg_kiov;
823 *niov = lntmsg->msg_niov;
825 *offset = lntmsg->msg_offset;
830 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
832 unsigned int niov, offset, nob;
833 struct bio_vec *kiov;
834 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
835 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
837 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
838 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
839 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
840 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
841 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
842 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
843 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
845 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
846 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
847 tx->tx_msg.gnm_payload_cksum = 0;
850 if (*kgnilnd_tunables.kgn_checksum < 3) {
851 tx->tx_msg.gnm_payload_cksum = 0;
855 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
857 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
861 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
863 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
865 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
869 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
870 tx->tx_msg.gnm_payload_cksum += 0xd00d;
874 /* kgnilnd_verify_rdma_cksum()
875 * tx - PUT_DONE/GET_DONE matched tx.
876 * rx_cksum - received checksum to compare against.
877 * put_len - see kgnilnd_parse_lnet_rdma comments.
880 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
884 unsigned int niov, offset, nob;
885 struct bio_vec *kiov;
886 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
887 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
889 /* we can only match certain requests */
890 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
891 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
892 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
893 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
894 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
895 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
896 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
898 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
899 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
904 if (*kgnilnd_tunables.kgn_checksum >= 3) {
905 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
906 "no RDMA payload checksum when enabled");
911 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
913 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
916 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
918 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
921 if (cksum != rx_cksum) {
922 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
923 "Bad RDMA payload checksum (%x expected %x); "
924 "kiov 0x%p niov %d nob %u offset %u",
925 cksum, rx_cksum, kiov, niov, nob, offset);
926 switch (dump_on_err) {
929 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
931 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
934 /* fall through to dump log */
936 libcfs_debug_dumplog();
942 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
948 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
952 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
953 "already mapped!", NULL);
955 spin_lock(&dev->gnd_map_lock);
956 switch (tx->tx_buftype) {
958 GNIDBG_TX(D_EMERG, tx,
959 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
960 spin_unlock(&dev->gnd_map_lock);
964 case GNILND_BUF_PHYS_MAPPED:
965 bytes = tx->tx_phys_npages * PAGE_SIZE;
966 dev->gnd_map_nphys++;
967 dev->gnd_map_physnop += tx->tx_phys_npages;
970 case GNILND_BUF_VIRT_MAPPED:
972 dev->gnd_map_nvirt++;
973 dev->gnd_map_virtnob += tx->tx_nob;
977 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
978 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
979 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
980 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
981 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
984 atomic_inc(&dev->gnd_n_mdd);
985 atomic64_add(bytes, &dev->gnd_nbytes_map);
987 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
990 /* we only get here in the valid cases */
991 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
992 dev->gnd_map_version++;
993 spin_unlock(&dev->gnd_map_lock);
997 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
1001 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
1002 "not mapped!", NULL);
1003 spin_lock(&dev->gnd_map_lock);
1005 switch (tx->tx_buftype) {
1007 GNIDBG_TX(D_EMERG, tx,
1008 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
1009 spin_unlock(&dev->gnd_map_lock);
1013 case GNILND_BUF_PHYS_UNMAPPED:
1014 bytes = tx->tx_phys_npages * PAGE_SIZE;
1015 dev->gnd_map_nphys--;
1016 dev->gnd_map_physnop -= tx->tx_phys_npages;
1019 case GNILND_BUF_VIRT_UNMAPPED:
1021 dev->gnd_map_nvirt--;
1022 dev->gnd_map_virtnob -= tx->tx_nob;
1026 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1027 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1028 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
1029 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
1030 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
1031 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
1032 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1035 atomic_dec(&dev->gnd_n_mdd);
1036 atomic64_sub(bytes, &dev->gnd_nbytes_map);
1038 /* we only get here in the valid cases */
1039 list_del_init(&tx->tx_map_list);
1040 dev->gnd_map_version++;
1041 spin_unlock(&dev->gnd_map_lock);
1045 kgnilnd_map_buffer(kgn_tx_t *tx)
1047 kgn_conn_t *conn = tx->tx_conn;
1048 kgn_device_t *dev = conn->gnc_device;
1049 __u32 flags = GNI_MEM_READWRITE;
1052 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
1053 * be called concurrently as there are internal locks that protect
1054 * any data structures or HW resources. We just need to ensure
1055 * that our concurrency doesn't result in the kgn_device_t
1056 * getting nuked while we are in here */
1058 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
1059 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
1061 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1064 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1065 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1068 switch (tx->tx_buftype) {
1072 case GNILND_BUF_NONE:
1073 case GNILND_BUF_IMMEDIATE:
1074 case GNILND_BUF_IMMEDIATE_KIOV:
1075 case GNILND_BUF_PHYS_MAPPED:
1076 case GNILND_BUF_VIRT_MAPPED:
1079 case GNILND_BUF_PHYS_UNMAPPED:
1080 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1081 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1082 tx->tx_phys, tx->tx_phys_npages, NULL,
1083 GNI_MEM_PHYS_SEGMENTS | flags,
1085 /* could race with other uses of the map counts, but this is ok
1086 * - this needs to turn into a non-fatal error soon to allow
1087 * GART resource, etc starvation handling */
1088 if (rrc != GNI_RC_SUCCESS) {
1089 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1090 "phys %u pp %u, virt %u nob %llu",
1091 tx->tx_phys_npages, dev->gnd_id,
1092 dev->gnd_map_nphys, dev->gnd_map_physnop,
1093 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1094 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1097 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1098 kgnilnd_mem_add_map_list(dev, tx);
1101 case GNILND_BUF_VIRT_UNMAPPED:
1102 rrc = kgnilnd_mem_register(dev->gnd_handle,
1103 (__u64)tx->tx_buffer, tx->tx_nob,
1104 NULL, flags, &tx->tx_map_key);
1105 if (rrc != GNI_RC_SUCCESS) {
1106 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1107 "phys %u pp %u, virt %u nob %llu",
1108 tx->tx_nob, dev->gnd_id,
1109 dev->gnd_map_nphys, dev->gnd_map_physnop,
1110 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1111 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1114 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1115 kgnilnd_mem_add_map_list(dev, tx);
1116 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1117 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1118 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1119 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1120 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1128 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1130 kgn_conn_t *conn = tx->tx_conn;
1131 kgn_mdd_purgatory_t *gmp;
1133 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1134 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1135 " asserting to avoid data corruption\n");
1136 if (tx->tx_buffer_copy)
1137 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1139 gmp->gmp_map_key = tx->tx_map_key;
1141 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1143 /* ensure that we don't have a blank purgatory - indicating the
1144 * conn is not already on purgatory lists - we'd never recover these
1145 * MDD if that were the case */
1146 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1147 "conn 0x%p->%s with NULL purgatory",
1148 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1150 /* link 'er up! - only place we really need to lock for
1151 * concurrent access */
1152 spin_lock(&conn->gnc_list_lock);
1153 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1154 spin_unlock(&conn->gnc_list_lock);
1158 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1162 int hold_timeout = 0;
1164 /* code below relies on +1 relationship ... */
1165 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1166 (GNILND_BUF_PHYS_UNMAPPED + 1));
1167 BUILD_BUG_ON(GNILND_BUF_VIRT_MAPPED !=
1168 (GNILND_BUF_VIRT_UNMAPPED + 1));
1170 switch (tx->tx_buftype) {
1174 case GNILND_BUF_NONE:
1175 case GNILND_BUF_IMMEDIATE:
1176 case GNILND_BUF_PHYS_UNMAPPED:
1177 case GNILND_BUF_VIRT_UNMAPPED:
1179 case GNILND_BUF_IMMEDIATE_KIOV:
1180 if (tx->tx_phys != NULL) {
1181 vunmap(tx->tx_phys);
1182 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1183 kunmap(tx->tx_imm_pages[0]);
1185 /* clear to prevent kgnilnd_free_tx from thinking
1186 * this is a RDMA descriptor */
1190 case GNILND_BUF_PHYS_MAPPED:
1191 case GNILND_BUF_VIRT_MAPPED:
1192 LASSERT(tx->tx_conn != NULL);
1194 dev = tx->tx_conn->gnc_device;
1196 /* only want to hold if we are closing conn without
1197 * verified peer notification - the theory is that
1198 * a TX error can be communicated in all other cases */
1199 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1200 error != -GNILND_NOPURG &&
1201 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1202 kgnilnd_add_purgatory_tx(tx);
1204 /* The timeout we give to kgni is a deadman stop only.
1205 * we are setting high to ensure we don't have the kgni timer
1206 * fire before ours fires _and_ is handled */
1207 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1209 GNIDBG_TX(D_NET, tx,
1210 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1211 tx->tx_conn->gnc_device, hold_timeout,
1212 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1214 if (tx->tx_buffer_copy != NULL) {
1215 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1216 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1217 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1218 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1220 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1221 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1225 kgnilnd_mem_del_map_list(dev, tx);
1231 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1233 struct lnet_msg *lntmsg0, *lntmsg1;
1234 int status0, status1;
1235 struct lnet_ni *ni = NULL;
1236 kgn_conn_t *conn = tx->tx_conn;
1238 LASSERT(!in_interrupt());
1240 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1241 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1244 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1245 !kgnilnd_conn_clean_errno(completion)) {
1246 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1247 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1248 completion, tx, conn ?
1249 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1250 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1251 kgnilnd_tx_state2str(tx->tx_list_state),
1252 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1255 /* The error codes determine if we hold onto the MDD */
1256 kgnilnd_unmap_buffer(tx, completion);
1258 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1259 * we play nice with the error codes to avoid delivering a failed
1260 * REQUEST and then a REPLY event as well */
1262 /* return -EIO to lnet - it is the magic value for failed sends */
1263 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1265 status1 = completion;
1267 status0 = status1 = completion;
1270 tx->tx_buftype = GNILND_BUF_NONE;
1271 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1273 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1274 * set NULL when we are a tx without a conn */
1276 ni = conn->gnc_peer->gnp_net->gnn_ni;
1278 spin_lock(&conn->gnc_tx_lock);
1280 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1281 (volatile unsigned long *)&conn->gnc_tx_bits),
1282 "conn %p tx %p bit %d already cleared\n",
1283 conn, tx, tx->tx_id.txe_idx);
1285 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1286 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1288 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1289 spin_unlock(&conn->gnc_tx_lock);
1292 kgnilnd_free_tx(tx);
1294 /* finalize AFTER freeing lnet msgs */
1296 /* warning - we should hold no locks here - calling lnet_finalize
1297 * could free up lnet credits, resulting in a call chain back into
1298 * the LND via kgnilnd_send and friends */
1300 lnet_finalize(lntmsg0, status0);
1302 if (lntmsg1 != NULL) {
1303 lnet_finalize(lntmsg1, status1);
1308 kgnilnd_txlist_done(struct list_head *txlist, int error)
1311 int err_printed = 0;
1313 if (list_empty(txlist))
1316 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1317 /* only print the first error */
1319 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1320 list_del_init(&tx->tx_list);
1321 kgnilnd_tx_done(tx, error);
1326 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1330 spin_lock(&conn->gnc_tx_lock);
1332 /* ID zero is NOT ALLOWED!!! */
1335 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1336 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1337 if (id == GNILND_MAX_MSG_ID) {
1338 if (conn->gnc_next_tx != 1) {
1339 /* we only searched from next_tx to end and didn't find
1340 * one, so search again from start */
1341 conn->gnc_next_tx = 1;
1344 /* couldn't find one! */
1345 spin_unlock(&conn->gnc_tx_lock);
1349 /* bump next_tx to prevent immediate reuse */
1350 conn->gnc_next_tx = id + 1;
1352 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1353 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1354 "tx 0x%p already at id %d\n",
1355 conn->gnc_tx_ref_table[id], id);
1357 /* delay these until we have a valid ID - prevents bad clear of the bit
1358 * in kgnilnd_tx_done */
1360 tx->tx_id.txe_cqid = conn->gnc_cqid;
1362 tx->tx_id.txe_idx = id;
1363 conn->gnc_tx_ref_table[id] = tx;
1365 /* Using jiffies to help differentiate against TX reuse - with
1366 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1367 * if we are sending to the same node faster than 256000/sec.
1368 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1370 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1372 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1374 spin_unlock(&conn->gnc_tx_lock);
1379 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1383 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1385 /* we don't care about TX timeout - it could be that the network is slower
1386 * or throttled. We'll keep retranmitting - so if the network is so slow
1387 * that we fill up our mailbox, we'll keep trying to resend that msg
1388 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1389 * that he hasn't send us any traffic in return */
1391 /* some reasonable throttling of the debug message */
1393 unsigned long now = jiffies;
1394 /* XXX Nic: Mystical TX debug here... */
1395 /* We expect retransmissions so only log when D_NET is enabled */
1396 GNIDBG_SMSG_CREDS(D_NET, conn);
1397 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1398 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1399 " last_msg %uus/%uus last_cq %uus/%uus",
1400 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1401 tx->tx_id, tx->tx_retrans,
1402 jiffies_to_usecs(now - tx->tx_cred_wait),
1403 jiffies_to_usecs(now - conn->gnc_last_tx),
1404 jiffies_to_usecs(now - conn->gnc_last_rx),
1405 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1406 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1410 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1411 * to avoid bad ordering with state_lock */
1414 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1415 spinlock_t *state_lock, kgn_tx_list_state_t state)
1417 kgn_conn_t *conn = tx->tx_conn;
1418 kgn_msg_t *msg = &tx->tx_msg;
1420 unsigned long newest_last_rx, timeout;
1423 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1424 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1426 "msg 0x%p type %d wrong payload size %d\n",
1427 msg, msg->gnm_type, immediatenob);
1429 /* make sure we catch all the cases where we'd send on a dirty old mbox
1430 * but allow case for sending CLOSE. Since this check is within the CQ
1431 * mutex barrier and the close message is only sent through
1432 * kgnilnd_send_conn_close the last message out the door will be the
1435 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1436 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1437 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1438 /* Return -ETIME, we are closing the connection already so we dont want to
1439 * have this tx hit the wire. The tx will be killed by the calling function.
1440 * Once the EP is marked dirty the close message will be the last
1441 * thing to hit the wire */
1446 timeout = cfs_time_seconds(conn->gnc_timeout);
1448 newest_last_rx = GNILND_LASTRX(conn);
1450 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1451 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1454 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1455 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1456 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1457 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1458 cfs_duration_sec(now - newest_last_rx),
1459 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1460 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1461 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1465 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1466 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1467 * creating the message this allows the message to contain the correct LNET NID/NET needed
1468 * instead of the one that the peer/conn uses for sending the data.
1470 msg->gnm_connstamp = conn->gnc_my_connstamp;
1471 msg->gnm_payload_len = immediatenob;
1472 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1474 /* always init here - kgn_checksum is a /sys module tunable
1475 * and can be flipped at any point, even between msg init and sending */
1477 if (*kgnilnd_tunables.kgn_checksum) {
1478 /* We must set here and not in kgnilnd_init_msg,
1479 * we could resend this msg many times
1480 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1481 * through init_msg again */
1482 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1483 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1484 msg->gnm_cksum += 0xf00f;
1488 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1489 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1490 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1491 tx->tx_id.txe_idx, immediate, immediatenob);
1493 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1494 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1496 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1497 msg, sizeof(*msg), immediate,
1499 tx->tx_id.txe_smsg_id);
1503 case GNI_RC_SUCCESS:
1504 atomic_inc(&conn->gnc_tx_seq);
1505 conn->gnc_last_tx = jiffies;
1506 /* no locking here as LIVE isn't a list */
1507 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1509 /* this needs to be checked under lock as it might be freed from a completion
1512 if (msg->gnm_type == GNILND_MSG_NOOP) {
1513 set_mb(conn->gnc_last_noop_sent, jiffies);
1516 /* serialize with seeing CQ events for completion on this, as well as
1518 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1519 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1521 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1522 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1523 kgnilnd_peer_alive(conn->gnc_peer);
1524 GNIDBG_SMSG_CREDS(D_NET, conn);
1527 case GNI_RC_NOT_DONE:
1528 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1529 * 1: out of mbox credits
1530 * 2: out of mbox payload credits
1531 * 3: On Aries out of dla credits
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 /* I need kgni credits to send this. Replace tx at the head of the
1544 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1545 * and bump retrans counts since we are requeueing the tx.
1549 conn->gnc_tx_retrans++;
1551 kgnilnd_tx_log_retrans(conn, tx);
1552 /* add to head of list for the state and retries */
1553 spin_lock(state_lock);
1554 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1555 spin_unlock(state_lock);
1557 /* We only reschedule for a certain number of retries, then
1558 * we will wait for the CQ events indicating a release of SMSG
1560 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1561 kgnilnd_schedule_conn(conn);
1564 /* CQ event coming in signifies either TX completed or
1565 * RX receive. Either of these *could* free up credits
1566 * in the SMSG mbox and we should try sending again */
1567 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1568 tx->tx_conn->gnc_cqid);
1569 kgnilnd_schedule_delay_conn(conn);
1570 /* use +ve return code to let upper layers know they
1571 * should stop looping on sends */
1575 /* handle bad retcode gracefully */
1576 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1577 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1582 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1584 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1585 spinlock_t *state_lock, kgn_tx_list_state_t state)
1587 kgn_device_t *dev = tx->tx_conn->gnc_device;
1588 unsigned long timestamp;
1591 timestamp = jiffies;
1592 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1593 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1594 /* delay in jiffies - we are really concerned only with things that
1595 * result in a schedule() or really holding this off for long times .
1596 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1597 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1599 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1605 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1607 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1608 spinlock_t *state_lock, kgn_tx_list_state_t state)
1610 kgn_conn_t *conn = tx->tx_conn;
1611 kgn_device_t *dev = conn->gnc_device;
1612 unsigned long timestamp;
1615 timestamp = jiffies;
1617 /* technically we are doing bad things with the read_lock on the peer_conn
1618 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1619 * for the mutex. I bet lockdep is gonna flag this one though... */
1621 /* there are a few cases where we don't want the immediate send - like
1622 * when we are in the scheduler thread and it'd harm the latency of
1623 * getting messages up to LNet */
1625 /* rmb for gnd_ready */
1627 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1629 atomic_inc(&conn->gnc_device->gnd_fast_block);
1630 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1631 /* dont hit HW during quiesce */
1633 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1634 /* dont hit HW if stale EPs and conns left to close */
1637 atomic_inc(&conn->gnc_device->gnd_fast_try);
1638 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1639 &conn->gnc_smsg_mutex);
1644 /* we got the mutex and weren't blocked */
1646 /* delay in jiffies - we are really concerned only with things that
1647 * result in a schedule() or really holding this off for long times .
1648 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1649 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1651 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1652 tx->tx_qtime = jiffies;
1653 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1654 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1655 /* _nolock unlocks the mutex for us */
1661 /* lets us know if we can push this RDMA through now */
1663 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1667 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1669 if (bytes_left < 0) {
1670 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1671 atomic_inc(&dev->gnd_rdmaq_nstalls);
1674 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1675 dev->gnd_rdmaq_deadline);
1676 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1677 /* we never del this timer - at worst it schedules us.. */
1684 /* this adds a TX to the queue pending throttling authorization before
1685 * we allow our remote peer to launch a PUT at us */
1687 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1691 /* we cannot go into send_mapped_tx from here as we are holding locks
1692 * and mem registration might end up allocating memory in kgni.
1693 * That said, we'll push this as far as we can into the queue process */
1694 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1697 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1698 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1699 /* lets us know how delayed RDMA is */
1700 tx->tx_qtime = jiffies;
1701 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1703 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1704 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1705 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1706 /* lets us know how delayed mapping is */
1707 tx->tx_qtime = jiffies;
1708 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1711 /* make sure we wake up sched to run this */
1712 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1715 /* push TX through state machine */
1717 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1722 /* set the tx_id here, we delay it until we have an actual conn
1724 * in some cases, the tx_id is already set to provide for things
1725 * like RDMA completion cookies, etc */
1726 if (tx->tx_id.txe_idx == 0) {
1727 rc = kgnilnd_set_tx_id(tx, conn);
1729 kgnilnd_tx_done(tx, rc);
1734 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1735 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1737 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1739 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1740 kgnilnd_tx_done(tx, rc);
1744 switch (tx->tx_msg.gnm_type) {
1745 case GNILND_MSG_PUT_ACK:
1746 case GNILND_MSG_GET_REQ:
1747 case GNILND_MSG_PUT_REQ_REV:
1748 case GNILND_MSG_GET_ACK_REV:
1749 /* hijacking time! If this messages will authorize our peer to
1750 * send his dirty little bytes in an RDMA, we need to get permission */
1751 kgnilnd_queue_rdma(conn, tx);
1753 case GNILND_MSG_IMMEDIATE:
1754 /* try to send right now, can help reduce latency */
1755 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1758 /* it was sent, break out of switch to avoid default case of queueing */
1761 /* needs to queue to try again, so fall through to default case */
1762 case GNILND_MSG_NOOP:
1763 /* Just make sure this goes out first for this conn */
1765 /* fall through... */
1767 spin_lock(&conn->gnc_list_lock);
1768 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1769 tx->tx_qtime = jiffies;
1770 spin_unlock(&conn->gnc_list_lock);
1771 kgnilnd_schedule_conn(conn);
1776 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1779 kgn_peer_t *new_peer = NULL;
1780 kgn_conn_t *conn = NULL;
1786 /* If I get here, I've committed to send, so I complete the tx with
1787 * failure on any problems */
1789 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1790 "tx already has connection %p", tx->tx_conn);
1792 /* do all of the peer & conn searching in one swoop - this avoids
1793 * nastiness when dropping locks and needing to maintain a sane state
1794 * in the face of stack reset or something else nuking peers & conns */
1796 /* I expect to find him, so only take a read lock */
1797 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1799 peer = kgnilnd_find_peer_locked(target->nid);
1801 conn = kgnilnd_find_conn_locked(peer);
1802 /* this could be NULL during quiesce */
1804 /* Connection exists; queue message on it */
1805 kgnilnd_queue_tx(conn, tx);
1806 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1810 /* don't create a connection if the peer is marked down */
1811 if (peer->gnp_state != GNILND_PEER_UP) {
1812 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1818 /* creating peer or conn; I'll need a write lock... */
1819 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1821 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1823 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1825 /* NB - this will not block during normal operations -
1826 * the only writer of this is in the startup/shutdown path. */
1827 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1833 /* ignore previous peer entirely - we cycled the lock, so we
1834 * will create new peer and at worst drop it if peer is still
1836 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1838 up_read(&kgnilnd_data.kgn_net_rw_sem);
1842 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1843 up_read(&kgnilnd_data.kgn_net_rw_sem);
1845 /* search for peer again now that we have the lock
1846 * if we don't find it, add our new one to the list */
1847 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1849 /* don't create a connection if the peer is not up */
1850 if (peer->gnp_state != GNILND_PEER_UP) {
1851 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1856 conn = kgnilnd_find_or_create_conn_locked(peer);
1858 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1859 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1864 /* oh hey, found a conn now... magical */
1865 kgnilnd_queue_tx(conn, tx);
1867 /* no conn, must be trying to connect - so we queue for now */
1868 tx->tx_qtime = jiffies;
1869 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1871 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1874 kgnilnd_tx_done(tx, rc);
1879 kgnilnd_rdma(kgn_tx_t *tx, int type,
1880 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1882 kgn_conn_t *conn = tx->tx_conn;
1883 unsigned long timestamp;
1884 gni_post_type_t post_type;
1887 unsigned int desc_nob = nob;
1888 void *desc_buffer = tx->tx_buffer;
1889 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1890 LASSERTF(kgnilnd_tx_mapped(tx),
1891 "unmapped tx %p\n", tx);
1892 LASSERTF(conn != NULL,
1893 "NULL conn on tx %p, naughty, naughty\n", tx);
1894 LASSERTF(nob <= sink->gnrd_nob,
1895 "nob %u > sink->gnrd_nob %d (%p)\n",
1896 nob, sink->gnrd_nob, sink);
1897 LASSERTF(nob <= tx->tx_nob,
1898 "nob %d > tx(%p)->tx_nob %d\n",
1899 nob, tx, tx->tx_nob);
1902 case GNILND_MSG_GET_DONE:
1903 case GNILND_MSG_PUT_DONE:
1904 post_type = GNI_POST_RDMA_PUT;
1906 case GNILND_MSG_GET_DONE_REV:
1907 case GNILND_MSG_PUT_DONE_REV:
1908 post_type = GNI_POST_RDMA_GET;
1911 CERROR("invalid msg type %s (%d)\n",
1912 kgnilnd_msgtype2str(type), type);
1915 if (post_type == GNI_POST_RDMA_GET) {
1916 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1917 * aligned. If the local buffer is not aligned correctly using the copy buffer
1918 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1919 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1920 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1921 * making sure the length and addr are aligned and copying the data into the correct location
1922 * once the transfer has completed.
1924 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1925 (sink->gnrd_addr & 3) ||
1928 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1930 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1932 if ((nob + tx->tx_offset) & 3) {
1933 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1934 atomic_inc(&kgnilnd_data.kgn_rev_length);
1936 desc_nob = (nob + tx->tx_offset);
1939 if (tx->tx_buffer_copy == NULL) {
1940 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1941 * and require at most we allocate a few extra bytes. */
1942 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1944 if (!tx->tx_buffer_copy) {
1945 /* allocation of buffer failed nak the rdma */
1946 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1947 kgnilnd_tx_done(tx, -EFAULT);
1950 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1951 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);
1952 if (rc != GNI_RC_SUCCESS) {
1953 /* Registration Failed nak rdma and kill the tx. */
1954 kgnilnd_vfree(tx->tx_buffer_copy,
1956 tx->tx_buffer_copy = NULL;
1957 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1958 kgnilnd_tx_done(tx, -EFAULT);
1962 desc_map_key = tx->tx_buffer_copy_map_key;
1963 desc_buffer = tx->tx_buffer_copy;
1967 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1968 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1969 tx->tx_rdma_desc.type = post_type;
1970 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1971 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1972 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1973 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1974 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1975 tx->tx_rdma_desc.length = desc_nob;
1976 tx->tx_nob_rdma = nob;
1977 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1978 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1979 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1980 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1981 /* prep final completion message */
1982 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1983 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1984 /* send actual size RDMA'd in retval */
1985 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1987 kgnilnd_compute_rdma_cksum(tx, nob);
1990 kgnilnd_queue_tx(conn, tx);
1994 /* Don't lie (CLOSE == RDMA idle) */
1995 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1996 tx, conn, conn->gnc_close_sent);
1998 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1999 "0x%x cookie:%#llx",
2000 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
2002 /* set CQ dedicated for RDMA */
2003 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
2005 timestamp = jiffies;
2006 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
2007 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2008 /* delay in jiffies - we are really concerned only with things that
2009 * result in a schedule() or really holding this off for long times .
2010 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2011 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2013 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2015 if (rrc == GNI_RC_ERROR_RESOURCE) {
2016 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2017 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2018 kgnilnd_unmap_buffer(tx, 0);
2020 if (tx->tx_buffer_copy != NULL) {
2021 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
2022 tx->tx_buffer_copy = NULL;
2025 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2026 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2028 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2029 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2033 spin_lock(&conn->gnc_list_lock);
2034 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2035 tx->tx_qtime = jiffies;
2036 spin_unlock(&conn->gnc_list_lock);
2037 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2038 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2040 /* XXX Nic: is this a place we should handle more errors for
2041 * robustness sake */
2042 LASSERT(rrc == GNI_RC_SUCCESS);
2047 kgnilnd_alloc_rx(void)
2051 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2053 CERROR("failed to allocate rx\n");
2056 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2059 /* no memset to zero, we'll always fill all members */
2063 /* release is to just free connection resources
2064 * we use this for the eager path after copying */
2066 kgnilnd_release_msg(kgn_conn_t *conn)
2069 unsigned long timestamp;
2071 CDEBUG(D_NET, "consuming %p\n", conn);
2073 timestamp = jiffies;
2074 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2075 /* delay in jiffies - we are really concerned only with things that
2076 * result in a schedule() or really holding this off for long times .
2077 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2078 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2080 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2081 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2083 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2084 GNIDBG_SMSG_CREDS(D_NET, conn);
2086 kgnilnd_schedule_conn(conn);
2090 kgnilnd_consume_rx(kgn_rx_t *rx)
2092 kgn_conn_t *conn = rx->grx_conn;
2093 kgn_msg_t *rxmsg = rx->grx_msg;
2095 /* if we are eager, free the cache alloc'd msg */
2096 if (unlikely(rx->grx_eager)) {
2097 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2098 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2100 /* release ref from eager_recv */
2101 kgnilnd_conn_decref(conn);
2103 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2104 kgnilnd_release_msg(conn);
2107 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2108 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2113 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2115 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2116 int type = lntmsg->msg_type;
2117 struct lnet_process_id target = lntmsg->msg_target;
2118 int target_is_router = lntmsg->msg_target_is_router;
2119 int routing = lntmsg->msg_routing;
2120 unsigned int niov = lntmsg->msg_niov;
2121 struct kvec *iov = lntmsg->msg_iov;
2122 struct bio_vec *kiov = lntmsg->msg_kiov;
2123 unsigned int offset = lntmsg->msg_offset;
2124 unsigned int nob = lntmsg->msg_len;
2125 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2126 kgn_net_t *net = ni->ni_data;
2130 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2132 /* NB 'private' is different depending on what we're sending.... */
2133 LASSERT(!in_interrupt());
2135 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2136 type, nob, niov, libcfs_id2str(target));
2138 LASSERTF(nob == 0 || niov > 0,
2139 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2140 LASSERTF(niov <= LNET_MAX_IOV,
2141 "lntmsg %p niov %d\n", lntmsg, niov);
2143 /* payload is either all vaddrs or all pages */
2144 LASSERTF(!(kiov != NULL && iov != NULL),
2145 "lntmsg %p kiov %p iov %p\n", lntmsg, kiov, iov);
2148 mpflag = cfs_memory_pressure_get_and_set();
2152 CERROR("lntmsg %p with unexpected type %d\n",
2157 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2165 if (routing || target_is_router)
2166 break; /* send IMMEDIATE */
2168 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2169 * check the eventual sink buffer here - if small enough, remote
2170 * end is perfectly capable of returning data in short message -
2171 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2172 * for IMMEDIATE messages which will have it send a real reply instead
2173 * of doing kgnilnd_recv to have the RDMA continued */
2174 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2177 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2178 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2180 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2186 /* slightly different options as we might actually have a GET with a
2187 * MD_KIOV set but a non-NULL md_iov.iov */
2188 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
2189 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2190 lntmsg->msg_md->md_iov.iov, NULL,
2191 0, lntmsg->msg_md->md_length);
2193 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2194 NULL, lntmsg->msg_md->md_iov.kiov,
2195 0, lntmsg->msg_md->md_length);
2197 CERROR("unable to setup buffer: %d\n", rc);
2198 kgnilnd_tx_done(tx, rc);
2203 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2204 if (tx->tx_lntmsg[1] == NULL) {
2205 CERROR("Can't create reply for GET to %s\n",
2206 libcfs_nid2str(target.nid));
2207 kgnilnd_tx_done(tx, rc);
2212 tx->tx_lntmsg[0] = lntmsg;
2213 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2214 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2216 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2218 /* rest of tx_msg is setup just before it is sent */
2219 kgnilnd_launch_tx(tx, net, &target);
2221 case LNET_MSG_REPLY:
2223 /* to save on MDDs, we'll handle short kiov by vmap'ing
2224 * and sending via SMSG */
2225 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2228 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2229 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2231 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2238 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2240 kgnilnd_tx_done(tx, rc);
2245 tx->tx_lntmsg[0] = lntmsg;
2246 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2247 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2249 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2251 /* rest of tx_msg is setup just before it is sent */
2252 kgnilnd_launch_tx(tx, net, &target);
2256 /* send IMMEDIATE */
2258 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2259 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2261 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2267 rc = kgnilnd_setup_immediate_buffer(tx, niov, iov, kiov, offset, nob);
2269 kgnilnd_tx_done(tx, rc);
2273 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2274 tx->tx_lntmsg[0] = lntmsg;
2275 kgnilnd_launch_tx(tx, net, &target);
2278 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2280 cfs_memory_pressure_restore(mpflag);
2285 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2287 kgn_conn_t *conn = rx->grx_conn;
2288 kgn_msg_t *rxmsg = rx->grx_msg;
2289 unsigned int niov = lntmsg->msg_niov;
2290 struct kvec *iov = lntmsg->msg_iov;
2291 struct bio_vec *kiov = lntmsg->msg_kiov;
2292 unsigned int offset = lntmsg->msg_offset;
2293 unsigned int nob = lntmsg->msg_len;
2298 switch (rxmsg->gnm_type) {
2299 case GNILND_MSG_PUT_REQ_REV:
2300 done_type = GNILND_MSG_PUT_DONE_REV;
2303 case GNILND_MSG_GET_REQ:
2304 done_type = GNILND_MSG_GET_DONE;
2307 CERROR("invalid msg type %s (%d)\n",
2308 kgnilnd_msgtype2str(rxmsg->gnm_type),
2313 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2317 rc = kgnilnd_set_tx_id(tx, conn);
2321 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2325 tx->tx_lntmsg[0] = lntmsg;
2326 tx->tx_getinfo = rxmsg->gnm_u.get;
2328 /* we only queue from kgnilnd_recv - we might get called from other contexts
2329 * and we don't want to block the mutex in those cases */
2331 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2332 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2333 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2334 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2339 kgnilnd_tx_done(tx, rc);
2340 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2342 lnet_finalize(lntmsg, rc);
2346 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2349 kgn_rx_t *rx = private;
2350 kgn_conn_t *conn = rx->grx_conn;
2351 kgn_msg_t *rxmsg = rx->grx_msg;
2352 kgn_msg_t *eagermsg = NULL;
2353 kgn_peer_t *peer = NULL;
2354 kgn_conn_t *found_conn = NULL;
2356 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2357 conn, rxmsg, lntmsg);
2359 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2360 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2361 rxmsg->gnm_payload_len);
2364 /* Grab a read lock so the connection doesnt disappear on us
2365 * while we look it up
2367 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2369 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2371 found_conn = kgnilnd_find_conn_locked(peer);
2374 /* Verify the connection found is the same one that the message
2375 * is supposed to be using, if it is not output an error message
2378 if (!peer || !found_conn
2379 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2380 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2381 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2382 peer, conn, found_conn);
2384 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2385 " from %s", rxmsg->gnm_connstamp,
2386 found_conn->gnc_peer_connstamp,
2387 libcfs_nid2str(peer->gnp_nid));
2392 /* add conn ref to ensure it doesn't go away until all eager
2393 * messages processed */
2394 kgnilnd_conn_addref(conn);
2396 /* Now that we have verified the connection is valid and added a
2397 * reference we can remove the read_lock on the peer_conn_lock */
2398 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2400 /* we have no credits or buffers for this message, so copy it
2401 * somewhere for a later kgnilnd_recv */
2402 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2403 *kgnilnd_tunables.kgn_eager_credits) {
2404 CERROR("Out of eager credits to %s\n",
2405 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2409 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2411 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2412 if (eagermsg == NULL) {
2413 kgnilnd_conn_decref(conn);
2414 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2415 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2419 /* copy msg and payload */
2420 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2421 rx->grx_msg = eagermsg;
2424 /* stash this for lnet_finalize on cancel-on-conn-close */
2425 rx->grx_lntmsg = lntmsg;
2427 /* keep the same rx_t, it just has a new grx_msg now */
2428 *new_private = private;
2430 /* release SMSG buffer */
2431 kgnilnd_release_msg(conn);
2437 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2438 int delayed, unsigned int niov,
2439 struct kvec *iov, struct bio_vec *kiov,
2440 unsigned int offset, unsigned int mlen, unsigned int rlen)
2442 kgn_rx_t *rx = private;
2443 kgn_conn_t *conn = rx->grx_conn;
2444 kgn_msg_t *rxmsg = rx->grx_msg;
2450 LASSERT(!in_interrupt());
2451 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2452 /* Either all pages or all vaddrs */
2453 LASSERTF(!(kiov != NULL && iov != NULL), "kiov %p iov %p\n",
2456 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2457 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2458 conn, rxmsg, lntmsg,
2459 niov, kiov, iov, offset, mlen, rlen);
2461 /* we need to lock here as recv can be called from any context */
2462 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2463 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2464 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2466 /* someone closed the conn after we copied this out, nuke it */
2467 kgnilnd_consume_rx(rx);
2468 lnet_finalize(lntmsg, conn->gnc_error);
2471 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2473 switch (rxmsg->gnm_type) {
2475 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2476 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2477 conn, rx, rxmsg, lntmsg, niov, kiov, iov, offset, mlen, rlen);
2480 case GNILND_MSG_IMMEDIATE:
2481 if (mlen > rxmsg->gnm_payload_len) {
2482 GNIDBG_MSG(D_ERROR, rxmsg,
2483 "Immediate message from %s too big: %d > %d",
2484 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2485 rxmsg->gnm_payload_len);
2487 kgnilnd_consume_rx(rx);
2491 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2492 * right after the kgn_msg_t header - so just 'cute' way of saying
2493 * rxmsg + sizeof(kgn_msg_t) */
2495 /* check payload checksum if sent */
2497 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2498 !rxmsg->gnm_payload_cksum &&
2499 rxmsg->gnm_payload_len != 0)
2500 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2502 if (rxmsg->gnm_payload_cksum != 0) {
2503 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2504 * which is what is used to calculate the cksum on the TX side */
2505 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2507 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2508 GNIDBG_MSG(D_NETERROR, rxmsg,
2509 "Bad payload checksum (%x expected %x)",
2510 pload_cksum, rxmsg->gnm_payload_cksum);
2511 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2513 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2514 &rxmsg[1], rxmsg->gnm_payload_len);
2515 /* fall through to dump */
2517 libcfs_debug_dumplog();
2523 /* checksum problems are fatal, kill the conn */
2524 kgnilnd_consume_rx(rx);
2525 kgnilnd_close_conn(conn, rc);
2531 lnet_copy_flat2kiov(
2533 *kgnilnd_tunables.kgn_max_immediate,
2534 &rxmsg[1], 0, mlen);
2538 *kgnilnd_tunables.kgn_max_immediate,
2539 &rxmsg[1], 0, mlen);
2541 kgnilnd_consume_rx(rx);
2542 lnet_finalize(lntmsg, 0);
2545 case GNILND_MSG_PUT_REQ:
2546 /* LNET wants to truncate or drop transaction, sending NAK */
2548 kgnilnd_consume_rx(rx);
2549 lnet_finalize(lntmsg, 0);
2551 /* only error if lntmsg == NULL, otherwise we are just
2552 * short circuiting the rdma process of 0 bytes */
2553 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2554 lntmsg == NULL ? -ENOENT : 0,
2555 rxmsg->gnm_u.get.gngm_cookie,
2559 /* sending ACK with sink buff. info */
2560 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2562 kgnilnd_consume_rx(rx);
2566 rc = kgnilnd_set_tx_id(tx, conn);
2568 GOTO(nak_put_req, rc);
2571 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2573 GOTO(nak_put_req, rc);
2576 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2577 rxmsg->gnm_u.putreq.gnprm_cookie;
2578 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2579 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2580 (__u64)((unsigned long)tx->tx_buffer);
2581 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2583 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2584 tx->tx_qtime = jiffies;
2585 /* we only queue from kgnilnd_recv - we might get called from other contexts
2586 * and we don't want to block the mutex in those cases */
2588 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2589 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2590 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2591 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2593 kgnilnd_consume_rx(rx);
2597 /* make sure we send an error back when the PUT fails */
2598 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2599 kgnilnd_tx_done(tx, rc);
2600 kgnilnd_consume_rx(rx);
2602 /* return magic LNet network error */
2604 case GNILND_MSG_GET_REQ_REV:
2605 /* LNET wants to truncate or drop transaction, sending NAK */
2607 kgnilnd_consume_rx(rx);
2608 lnet_finalize(lntmsg, 0);
2610 /* only error if lntmsg == NULL, otherwise we are just
2611 * short circuiting the rdma process of 0 bytes */
2612 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2613 lntmsg == NULL ? -ENOENT : 0,
2614 rxmsg->gnm_u.get.gngm_cookie,
2618 /* lntmsg can be null when parsing a LNET_GET */
2619 if (lntmsg != NULL) {
2620 /* sending ACK with sink buff. info */
2621 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2623 kgnilnd_consume_rx(rx);
2627 rc = kgnilnd_set_tx_id(tx, conn);
2629 GOTO(nak_get_req_rev, rc);
2632 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2634 GOTO(nak_get_req_rev, rc);
2637 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2638 rxmsg->gnm_u.putreq.gnprm_cookie;
2639 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2640 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2641 (__u64)((unsigned long)tx->tx_buffer);
2642 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2644 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2646 /* we only queue from kgnilnd_recv - we might get called from other contexts
2647 * and we don't want to block the mutex in those cases */
2649 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2650 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2651 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2652 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2655 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2657 rxmsg->gnm_u.get.gngm_cookie,
2661 kgnilnd_consume_rx(rx);
2665 /* make sure we send an error back when the GET fails */
2666 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2667 kgnilnd_tx_done(tx, rc);
2668 kgnilnd_consume_rx(rx);
2670 /* return magic LNet network error */
2674 case GNILND_MSG_PUT_REQ_REV:
2675 /* LNET wants to truncate or drop transaction, sending NAK */
2677 kgnilnd_consume_rx(rx);
2678 lnet_finalize(lntmsg, 0);
2680 /* only error if lntmsg == NULL, otherwise we are just
2681 * short circuiting the rdma process of 0 bytes */
2682 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2683 lntmsg == NULL ? -ENOENT : 0,
2684 rxmsg->gnm_u.get.gngm_cookie,
2689 if (lntmsg != NULL) {
2691 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2694 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2696 rxmsg->gnm_u.get.gngm_cookie,
2699 kgnilnd_consume_rx(rx);
2701 case GNILND_MSG_GET_REQ:
2702 if (lntmsg != NULL) {
2704 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2707 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2709 rxmsg->gnm_u.get.gngm_cookie,
2712 kgnilnd_consume_rx(rx);
2718 /* needs write_lock on kgn_peer_conn_lock held */
2720 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2722 unsigned long timeout, keepalive;
2723 unsigned long now = jiffies;
2724 unsigned long newest_last_rx;
2727 /* given that we found this conn hanging off a peer, it better damned
2728 * well be connected */
2729 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2730 "conn 0x%p->%s with bad state%s\n", conn,
2731 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2733 kgnilnd_conn_state2str(conn));
2735 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2736 "rx_diff %lu tx_diff %lu\n",
2737 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2738 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2739 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2740 cfs_duration_sec(now - conn->gnc_last_tx));
2742 timeout = cfs_time_seconds(conn->gnc_timeout);
2743 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2745 /* just in case our lack of RX msg processing is gumming up the works - give the
2746 * remove an extra chance */
2748 newest_last_rx = GNILND_LASTRX(conn);
2750 if (time_after_eq(now, newest_last_rx + timeout)) {
2751 uint32_t level = D_CONSOLE|D_NETERROR;
2753 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2756 GNIDBG_CONN(level, conn,
2757 "No gnilnd traffic received from %s for %lu "
2758 "seconds, terminating connection. Is node down? ",
2759 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2760 cfs_duration_sec(now - newest_last_rx));
2764 /* we don't timeout on last_tx stalls - we are going to trust the
2765 * underlying network to let us know when sends are failing.
2766 * At worst, the peer will timeout our RX stamp and drop the connection
2767 * at that point. We'll then see his CLOSE or at worst his RX
2768 * stamp stop and drop the connection on our end */
2770 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2771 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2772 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2773 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2774 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2776 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2777 conn->gnc_last_noop_cq,
2778 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2779 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2780 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2781 set_mb(conn->gnc_last_noop_want, jiffies);
2782 atomic_inc(&conn->gnc_reaper_noop);
2783 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2786 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2789 kgnilnd_queue_tx(conn, tx);
2795 /* needs write_lock on kgn_peer_conn_lock held */
2797 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2798 struct list_head *souls)
2800 unsigned long timeout;
2801 kgn_conn_t *conn, *connN = NULL;
2807 short releaseconn = 0;
2808 unsigned long first_rx = 0;
2809 int purgatory_conn_cnt = 0;
2811 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2812 peer, libcfs_nid2str(peer->gnp_nid),
2813 peer->gnp_reconnect_interval);
2815 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2816 GNILND_MIN_TIMEOUT));
2818 conn = kgnilnd_find_conn_locked(peer);
2820 /* if there is a valid conn, check the queues for timeouts */
2821 rc = kgnilnd_check_conn_timeouts_locked(conn);
2823 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2824 /* simulate a RX CLOSE after the timeout but before
2825 * the scheduler thread gets it */
2826 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2827 conn->gnc_peer_error = -ETIMEDOUT;
2830 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2832 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2833 CDEBUG(D_WARNING, "%s conn timed out, will "
2834 "reconnect upon request from peer\n",
2835 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2837 /* Once we mark closed, any of the scheduler threads could
2838 * get it and move through before we hit the fail loc code */
2839 kgnilnd_close_conn_locked(conn, rc);
2841 /* first_rx is used to decide when to release a conn from purgatory.
2843 first_rx = conn->gnc_first_rx;
2847 /* now regardless of starting new conn, find tx on peer queue that
2848 * are old and smell bad - do this first so we don't trigger
2849 * reconnect on empty queue if we timeout all */
2850 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2851 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2853 LCONSOLE_INFO("could not send to %s due to connection"
2854 " setup failure after %lu seconds\n",
2855 libcfs_nid2str(peer->gnp_nid),
2856 cfs_duration_sec(jiffies - tx->tx_qtime));
2858 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2860 list_add_tail(&tx->tx_list, todie);
2865 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2866 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2867 count, peer, libcfs_nid2str(peer->gnp_nid));
2868 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2869 if (list_empty(&peer->gnp_tx_queue) ||
2870 peer->gnp_connecting == GNILND_PEER_KILL) {
2871 /* we pass down todie to use a common function - but we know there are
2873 kgnilnd_cancel_peer_connect_locked(peer, todie);
2877 /* Don't reconnect if we are still trying to clear out old conns.
2878 * This prevents us sending traffic on the new mbox before ensuring we are done
2879 * with the old one */
2880 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2881 (atomic_read(&peer->gnp_dirty_eps) == 0);
2883 /* fast reconnect after a timeout */
2884 to_reconn = !conn &&
2885 (peer->gnp_last_errno == -ETIMEDOUT) &&
2886 *kgnilnd_tunables.kgn_fast_reconn;
2888 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2889 * to be sent, we'll check the reconnect interval and fire up a new
2890 * connection request */
2893 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2894 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2895 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2897 CDEBUG(D_NET, "starting connect to %s\n",
2898 libcfs_nid2str(peer->gnp_nid));
2899 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2900 "Peer was idle and we have a write_lock, state issue %d\n",
2901 peer->gnp_connecting);
2903 peer->gnp_connecting = GNILND_PEER_CONNECT;
2904 kgnilnd_peer_addref(peer); /* extra ref for connd */
2906 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2907 list_add_tail(&peer->gnp_connd_list,
2908 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2909 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2911 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2914 /* fail_loc to allow us to delay release of purgatory */
2915 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2918 /* This check allows us to verify that the new conn is actually being used. This allows us to
2919 * pull the old conns out of purgatory if they have actually seen traffic.
2920 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2923 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2924 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2925 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2929 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2930 /* check for purgatory timeouts */
2931 if (conn->gnc_in_purgatory) {
2932 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2933 * that way the next time we check it we can detach it from purgatory
2936 if (conn->gnc_state != GNILND_CONN_DONE) {
2937 /* Skip over conns that are currently not DONE. If they arent already scheduled
2938 * for completion something in the state machine is broken.
2943 /* We only detach a conn that is in purgatory if we have received a close message,
2944 * we have a new valid connection that has successfully received data, or an admin
2945 * command tells us we need to detach.
2948 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2949 unsigned long waiting;
2951 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2953 /* C.E: The remote peer is expected to close the
2954 * connection (see kgnilnd_check_conn_timeouts)
2955 * via the reaper thread and nuke out the MDD and
2956 * FMA resources after conn->gnc_timeout has expired
2957 * without an FMA RX */
2958 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2959 " held resources\n",
2960 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2961 cfs_duration_sec(waiting));
2963 kgnilnd_detach_purgatory_locked(conn, souls);
2965 purgatory_conn_cnt++;
2970 /* If we have too many connections in purgatory we could run out of
2971 * resources. Limit the number of connections to a tunable number,
2972 * clean up to the minimum all in one fell swoop... there are
2973 * situations where dvs will retry tx's and we can eat up several
2974 * hundread connection requests at once.
2976 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2977 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2979 if (conn->gnc_in_purgatory &&
2980 conn->gnc_state == GNILND_CONN_DONE) {
2981 CDEBUG(D_NET, "Dropping Held resource due to"
2982 " resource limits being hit\n");
2983 kgnilnd_detach_purgatory_locked(conn, souls);
2985 if (purgatory_conn_cnt-- <
2986 *kgnilnd_tunables.kgn_max_purgatory)
2994 kgnilnd_reaper_check(int idx)
2996 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2997 struct list_head *ctmp, *ctmpN;
2998 LIST_HEAD(geriatrics);
3001 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3003 list_for_each_safe(ctmp, ctmpN, peers) {
3004 kgn_peer_t *peer = NULL;
3006 /* don't timeout stuff if the network is mucked or shutting down */
3007 if (kgnilnd_check_hw_quiesce()) {
3010 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
3012 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
3015 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3017 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3018 kgnilnd_release_purgatory_list(&souls);
3022 kgnilnd_update_reaper_timeout(long timeout)
3024 LASSERT(timeout > 0);
3026 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3028 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3029 kgnilnd_data.kgn_new_min_timeout = timeout;
3031 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3035 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3037 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3041 kgnilnd_reaper(void *arg)
3046 unsigned long next_check_time = jiffies;
3047 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3048 struct timer_list timer;
3051 /* all gnilnd threads need to run fairly urgently */
3052 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3053 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3055 while (!kgnilnd_data.kgn_shutdown) {
3056 /* I wake up every 'p' seconds to check for timeouts on some
3057 * more peers. I try to check every connection 'n' times
3058 * within the global minimum of all keepalive and timeout
3059 * intervals, to ensure I attend to every connection within
3060 * (n+1)/n times its timeout intervals. */
3061 const int p = GNILND_REAPER_THREAD_WAKE;
3062 const int n = GNILND_REAPER_NCHECKS;
3064 /* to quiesce or to not quiesce, that is the question */
3065 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3066 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3067 KGNILND_SPIN_QUIESCE;
3068 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3071 /* careful with the jiffy wrap... */
3072 timeout = (long)(next_check_time - jiffies);
3075 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3076 TASK_INTERRUPTIBLE);
3077 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3078 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3080 mod_timer(&timer, (long) jiffies + timeout);
3082 /* check flag variables before committing */
3083 if (!kgnilnd_data.kgn_shutdown &&
3084 !kgnilnd_data.kgn_quiesce_trigger) {
3085 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3086 timeout, cfs_duration_sec(timeout));
3088 CDEBUG(D_INFO, "awake after schedule\n");
3091 del_singleshot_timer_sync(&timer);
3092 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3093 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3097 /* new_min_timeout is set from the conn timeouts and keepalive
3098 * this should end up with a min timeout of
3099 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3100 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3101 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3102 CDEBUG(D_NET, "Set new min timeout %ld\n",
3103 current_min_timeout);
3106 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3108 /* Compute how many table entries to check now so I get round
3109 * the whole table fast enough given that I do this at fixed
3110 * intervals of 'p' seconds) */
3111 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3112 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3113 chunk = (chunk * n * p) /
3114 kgnilnd_data.kgn_new_min_timeout;
3117 for (i = 0; i < chunk; i++) {
3118 kgnilnd_reaper_check(hash_index);
3119 hash_index = (hash_index + 1) %
3120 *kgnilnd_tunables.kgn_peer_hash_size;
3122 next_check_time = (long) jiffies + cfs_time_seconds(p);
3123 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3125 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3128 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3130 kgnilnd_thread_fini();
3135 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3136 unsigned niov, offset, nob;
3137 struct bio_vec *kiov;
3138 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3139 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3142 lnet_copy_flat2kiov(
3145 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3147 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3154 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3157 gni_post_descriptor_t *desc;
3159 kgn_tx_ev_id_t ev_id;
3161 int should_retry, rc;
3162 long num_processed = 0;
3163 kgn_conn_t *conn = NULL;
3164 kgn_tx_t *tx = NULL;
3165 kgn_rdma_desc_t *rdesc;
3170 /* make sure we don't keep looping if we need to reset */
3171 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3172 return num_processed;
3174 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3176 /* we didn't get the mutex, so return that there is still work
3180 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3181 /* a bit gross - but we need a good way to test for
3182 * delayed RDMA completions and the easiest way to do
3183 * that is to delay the RDMA CQ events */
3184 rrc = GNI_RC_NOT_DONE;
3186 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3189 if (rrc == GNI_RC_NOT_DONE) {
3190 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3191 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3192 dev->gnd_id, num_processed);
3193 return num_processed;
3195 dev->gnd_sched_alive = jiffies;
3198 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3199 "this is bad, somehow our credits didn't protect us"
3200 " from CQ overrun\n");
3201 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3202 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3203 event_data, GNI_CQ_GET_TYPE(event_data));
3205 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3207 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3209 /* XXX Nic: Need better error handling here... */
3210 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3211 (rrc == GNI_RC_TRANSACTION_ERROR),
3214 ev_id.txe_cookie = desc->post_id;
3216 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3218 if (conn == NULL || tx == NULL) {
3219 /* either conn or tx was already nuked and this is a "late"
3220 * completion, so drop it */
3224 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3225 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3226 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3227 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3228 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3230 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3232 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3233 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3234 tx->tx_msg.gnm_u.completion.gncm_retval);
3238 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3239 if (tx->tx_buffer_copy != NULL)
3240 kgnilnd_recv_bte_get(tx);
3241 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3244 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3245 if (tx->tx_buffer_copy != NULL)
3246 kgnilnd_recv_bte_get(tx);
3247 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3250 /* remove from rdmaq */
3251 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3252 spin_lock(&conn->gnc_list_lock);
3253 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3254 spin_unlock(&conn->gnc_list_lock);
3255 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3257 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3258 event_data = 1LL << 48;
3262 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3263 atomic_inc(&dev->gnd_rdma_ntx);
3264 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3265 /* transaction succeeded, add into fmaq */
3266 kgnilnd_queue_tx(conn, tx);
3267 kgnilnd_peer_alive(conn->gnc_peer);
3269 /* drop ref from kgnilnd_validate_tx_ev_id */
3270 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3271 kgnilnd_conn_decref(conn);
3276 /* fall through to the TRANSACTION_ERROR case */
3279 /* get stringified version for log messages */
3280 kgnilnd_cq_error_str(event_data, &err_str, 256);
3281 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3283 /* make sure we are not off in the weeds with this tx */
3284 if (tx->tx_retrans >
3285 *kgnilnd_tunables.kgn_max_retransmits) {
3286 GNIDBG_TX(D_NETERROR, tx,
3287 "giving up on TX, too many retries", NULL);
3291 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3292 should_retry ? "transient" : "unrecoverable", err_str);
3294 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3295 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3296 rdesc = &tx->tx_putinfo.gnpam_desc;
3297 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3298 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3300 rdesc = &tx->tx_getinfo.gngm_desc;
3301 rnob = tx->tx_lntmsg[0]->msg_len;
3302 rcookie = tx->tx_getinfo.gngm_cookie;
3307 tx->tx_msg.gnm_type,
3311 kgnilnd_nak_rdma(conn,
3312 tx->tx_msg.gnm_type,
3315 tx->tx_msg.gnm_srcnid);
3316 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3317 kgnilnd_close_conn(conn, -ECOMM);
3320 /* drop ref from kgnilnd_validate_tx_ev_id */
3321 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3322 kgnilnd_conn_decref(conn);
3327 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3331 kgn_tx_ev_id_t ev_id;
3332 kgn_tx_t *tx = NULL;
3333 kgn_conn_t *conn = NULL;
3334 int queued_fma, saw_reply, rc;
3335 long num_processed = 0;
3336 struct list_head *ctmp, *ctmpN;
3339 /* make sure we don't keep looping if we need to reset */
3340 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3341 return num_processed;
3344 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3346 /* we didn't get the mutex, so return that there is still work
3351 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3352 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3354 if (rrc == GNI_RC_NOT_DONE) {
3356 "SMSG send CQ %d not ready (data %#llx) "
3357 "processed %ld\n", dev->gnd_id, event_data,
3360 if (num_processed > 0) {
3361 spin_lock(&dev->gnd_lock);
3362 if (!list_empty(&dev->gnd_delay_conns)) {
3363 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3364 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3365 list_del_init(&conn->gnc_delaylist);
3366 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3367 kgnilnd_schedule_conn_nolock(conn);
3369 spin_unlock(&dev->gnd_lock);
3370 kgnilnd_schedule_device(dev);
3372 spin_unlock(&dev->gnd_lock);
3375 return num_processed;
3378 dev->gnd_sched_alive = jiffies;
3381 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3382 "this is bad, somehow our credits didn't "
3383 "protect us from CQ overrun\n");
3384 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3385 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3386 event_data, GNI_CQ_GET_TYPE(event_data));
3388 /* if SMSG couldn't handle an error, time for conn to die */
3389 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3392 /* need to take the write_lock to ensure atomicity
3393 * on the conn state if we need to close it */
3394 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3395 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3397 /* Conn was destroyed? */
3399 "SMSG CQID lookup %#llx failed\n",
3400 GNI_CQ_GET_INST_ID(event_data));
3401 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3405 kgnilnd_cq_error_str(event_data, &err_str, 256);
3406 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3407 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3409 kgnilnd_close_conn_locked(conn, -ECOMM);
3411 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3413 /* no need to process rest of this tx -
3414 * it is getting canceled */
3418 /* fall through to GNI_RC_SUCCESS case */
3419 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3421 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3422 if (conn == NULL || tx == NULL) {
3423 /* either conn or tx was already nuked and this is a "late"
3424 * completion, so drop it */
3428 tx->tx_conn->gnc_last_tx_cq = jiffies;
3429 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3430 set_mb(conn->gnc_last_noop_cq, jiffies);
3433 /* lock tx_list_state and tx_state */
3434 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3435 spin_lock(&tx->tx_conn->gnc_list_lock);
3437 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3438 "state not GNILND_TX_LIVE_FMAQ", NULL);
3439 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3440 "not waiting for completion", NULL);
3442 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3445 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3447 /* This will trigger other FMA sends that were
3448 * pending this completion */
3449 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3451 /* we either did not expect reply or we already got it */
3452 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3454 spin_unlock(&tx->tx_conn->gnc_list_lock);
3455 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3458 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3460 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3461 kgnilnd_schedule_conn(conn);
3464 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3465 * If saw_reply is true we know that the tx is safe to use as the other thread
3466 * is already finished with it.
3470 /* no longer need to track on the live_fmaq */
3471 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3473 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3474 /* we already got reply & were waiting for
3475 * completion of initial send */
3476 /* to initiate RDMA transaction */
3477 GNIDBG_TX(D_NET, tx,
3478 "Pending RDMA 0x%p type 0x%02x",
3479 tx->tx_msg.gnm_type);
3480 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3481 rc = kgnilnd_send_mapped_tx(tx, 0);
3482 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3484 /* we are done with this tx */
3485 GNIDBG_TX(D_NET, tx,
3486 "Done with tx type 0x%02x",
3487 tx->tx_msg.gnm_type);
3488 kgnilnd_tx_done(tx, tx->tx_rc);
3492 /* drop ref from kgnilnd_validate_tx_ev_id */
3493 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3494 kgnilnd_conn_decref(conn);
3496 /* if we are waiting for a REPLY, we'll handle the tx then */
3497 } /* end for loop */
3501 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3506 long num_processed = 0;
3507 struct list_head *conns;
3508 struct list_head *tmp;
3512 /* make sure we don't keep looping if we need to reset */
3513 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3514 return num_processed;
3517 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3519 /* we didn't get the mutex, so return that there is still work
3523 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3524 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3526 if (rrc == GNI_RC_NOT_DONE) {
3527 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3529 dev->gnd_id, event_data, num_processed);
3530 return num_processed;
3532 dev->gnd_sched_alive = jiffies;
3535 /* this is the only CQ that can really handle transient
3537 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3538 rrc = cfs_fail_val ? cfs_fail_val
3539 : GNI_RC_ERROR_RESOURCE;
3540 if (rrc == GNI_RC_ERROR_RESOURCE) {
3541 /* set overrun too */
3542 event_data |= (1UL << 63);
3543 LASSERTF(GNI_CQ_OVERRUN(event_data),
3544 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3547 /* sender should get error event too and take care
3548 of failed transaction by re-transmitting */
3549 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3550 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3554 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3555 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3556 conn = kgnilnd_cqid2conn_locked(
3557 GNI_CQ_GET_INST_ID(event_data));
3559 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3560 "failed, dropping event %#llx\n",
3561 GNI_CQ_GET_INST_ID(event_data),
3564 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3566 GNI_CQ_GET_INST_ID(event_data),
3567 conn, conn->gnc_peer ?
3568 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3571 conn->gnc_last_rx_cq = jiffies;
3573 /* stash first rx so we can clear out purgatory.
3575 if (conn->gnc_first_rx == 0) {
3576 conn->gnc_first_rx = jiffies;
3578 kgnilnd_peer_alive(conn->gnc_peer);
3579 kgnilnd_schedule_conn(conn);
3581 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3585 /* FMA CQ has overflowed: check ALL conns */
3586 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3587 "conns on device %d\n", dev->gnd_id);
3589 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3591 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3592 conns = &kgnilnd_data.kgn_conns[rc];
3594 list_for_each(tmp, conns) {
3595 conn = list_entry(tmp, kgn_conn_t,
3598 if (conn->gnc_device == dev) {
3599 kgnilnd_schedule_conn(conn);
3600 conn->gnc_last_rx_cq = jiffies;
3604 /* don't block write lockers for too long... */
3605 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3610 /* try_map_if_full should only be used when processing TX from list of
3611 * backlog TX waiting on mappings to free up
3614 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3615 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3618 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3620 /* slight bit of race if multiple people calling, but at worst we'll have
3621 * order altered just a bit... which would not be determenistic anyways */
3622 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3624 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3626 /* We know that we have a GART reservation that should guarantee forward progress.
3627 * This means we don't need to take any extraordinary efforts if we are failing
3628 * mappings here - even if we are holding a very small number of these. */
3630 if (try_map_if_full || (rc == 0)) {
3631 rc = kgnilnd_map_buffer(tx);
3634 /* rc should be 0 if we mapped successfully here, if non-zero
3635 * we are queueing */
3637 /* if try_map_if_full set, they handle requeuing */
3638 if (unlikely(try_map_if_full)) {
3641 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3642 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3643 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3644 /* make sure we wake up sched to run this */
3645 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3646 /* return 0 as this is now queued for later sending */
3651 switch (tx->tx_msg.gnm_type) {
3655 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3656 * remote node where the RDMA will be started
3657 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3658 * be satisified. The rest of the errors are "hard" errors that require
3659 * upper layers to handle themselves.
3660 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3661 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3662 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3663 * GNILND_MSG_GET_DONE_REV.
3665 case GNILND_MSG_GET_REQ:
3666 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3667 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3668 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3669 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3670 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3671 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3672 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3674 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3675 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3677 case GNILND_MSG_PUT_ACK:
3678 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3679 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3680 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3681 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3683 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3684 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3687 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3688 case GNILND_MSG_PUT_DONE:
3689 case GNILND_MSG_PUT_REQ:
3690 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3691 &tx->tx_putinfo.gnpam_desc,
3692 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3693 tx->tx_putinfo.gnpam_dst_cookie);
3694 RETURN(try_map_if_full ? rc : 0);
3696 case GNILND_MSG_GET_DONE:
3697 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3698 &tx->tx_getinfo.gngm_desc,
3699 tx->tx_lntmsg[0]->msg_len,
3700 tx->tx_getinfo.gngm_cookie);
3701 RETURN(try_map_if_full ? rc : 0);
3703 case GNILND_MSG_PUT_REQ_REV:
3704 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3705 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3706 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3707 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3708 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3709 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3710 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3712 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3714 case GNILND_MSG_PUT_DONE_REV:
3715 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3716 &tx->tx_getinfo.gngm_desc,
3718 tx->tx_getinfo.gngm_cookie);
3719 RETURN(try_map_if_full ? rc : 0);
3721 case GNILND_MSG_GET_ACK_REV:
3722 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3723 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3724 /* LNET_GETS are a special case for parse */
3725 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3726 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3728 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3729 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3731 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3732 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3734 case GNILND_MSG_GET_DONE_REV:
3735 case GNILND_MSG_GET_REQ_REV:
3736 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3737 &tx->tx_putinfo.gnpam_desc,
3738 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3739 tx->tx_putinfo.gnpam_dst_cookie);
3740 RETURN(try_map_if_full ? rc : 0);
3748 kgnilnd_process_fmaq(kgn_conn_t *conn)
3751 kgn_tx_t *tx = NULL;
3752 void *buffer = NULL;
3753 unsigned int nob = 0;
3756 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3757 * However I will be rescheduled by an FMA completion event
3758 * when I eventually get some.
3759 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3760 * But it doesn't matter if I try to send a "real" message just
3761 * as I start closing because I'll get scheduled to send the
3764 /* Short circuit if the ep_handle is null we cant send anyway. */
3765 if (conn->gnc_ephandle == NULL)
3768 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3770 spin_lock(&conn->gnc_list_lock);
3772 if (list_empty(&conn->gnc_fmaq)) {
3773 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3775 spin_unlock(&conn->gnc_list_lock);
3777 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3778 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3779 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3780 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3781 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3783 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3784 conn->gnc_last_noop_cq,
3785 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3786 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3787 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3788 atomic_inc(&conn->gnc_sched_noop);
3789 set_mb(conn->gnc_last_noop_want, jiffies);
3791 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3794 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3798 rc = kgnilnd_set_tx_id(tx, conn);
3800 kgnilnd_tx_done(tx, rc);
3806 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3807 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3808 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3809 more_to_do = !list_empty(&conn->gnc_fmaq);
3810 spin_unlock(&conn->gnc_list_lock);
3813 /* if there is no real TX or no NOOP to send, bail */
3818 if (!tx->tx_retrans)
3819 tx->tx_cred_wait = jiffies;
3821 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3822 "tx with zero id", NULL);
3824 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3825 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3826 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3830 switch (tx->tx_msg.gnm_type) {
3834 case GNILND_MSG_NOOP:
3835 case GNILND_MSG_CLOSE:
3836 case GNILND_MSG_IMMEDIATE:
3837 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3838 buffer = tx->tx_buffer;
3842 case GNILND_MSG_GET_DONE:
3843 case GNILND_MSG_PUT_DONE:
3844 case GNILND_MSG_PUT_DONE_REV:
3845 case GNILND_MSG_GET_DONE_REV:
3846 case GNILND_MSG_PUT_NAK:
3847 case GNILND_MSG_GET_NAK:
3848 case GNILND_MSG_GET_NAK_REV:
3849 case GNILND_MSG_PUT_NAK_REV:
3850 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3853 case GNILND_MSG_PUT_REQ:
3854 case GNILND_MSG_GET_REQ_REV:
3855 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3857 case GNILND_MSG_PUT_ACK:
3858 case GNILND_MSG_PUT_REQ_REV:
3859 case GNILND_MSG_GET_ACK_REV:
3860 case GNILND_MSG_GET_REQ:
3861 /* This is really only to handle the retransmit of SMSG once these
3862 * two messages are setup in send_mapped_tx */
3863 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3867 if (likely(rc == 0)) {
3868 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3872 /* don't explicitly reschedule here - we are short credits and will rely on
3873 * kgnilnd_sendmsg to resched the conn if need be */
3875 } else if (rc < 0) {
3876 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3877 * almost certainly a software bug, but lets play nice with the other kids */
3878 kgnilnd_tx_done(tx, rc);
3879 /* just for fun, kick peer in arse - resetting conn might help to correct
3880 * this almost certainly buggy software caused return code */
3881 kgnilnd_close_conn(conn, rc);
3885 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3886 kgnilnd_schedule_conn(conn);
3891 kgnilnd_process_rdmaq(kgn_device_t *dev)
3896 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3900 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3901 unsigned long dead_bump;
3904 /* if we think we need to adjust, take lock to serialize and recheck */
3905 spin_lock(&dev->gnd_rdmaq_lock);
3906 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3907 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3909 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3911 /* roll the bucket forward */
3912 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3914 if (kgnilnd_data.kgn_rdmaq_override &&
3915 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3916 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3921 /* roll current outstanding forward to make sure we carry outstanding
3922 * committment forward
3923 * new_ok starts out as the whole interval value
3924 * - first subtract bytes_out from last interval, as that would push us over
3925 * strict limits for this interval
3926 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3928 * there is a small race here if someone is actively processing mappings and
3929 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3930 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3931 * - if this gives us problems in the future, we could use a read/write lock
3932 * to protect the resetting of these values */
3933 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3934 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3936 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3937 "current out %ld\n",
3938 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3939 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3941 spin_unlock(&dev->gnd_rdmaq_lock);
3944 spin_lock(&dev->gnd_rdmaq_lock);
3945 while (!list_empty(&dev->gnd_rdmaq)) {
3948 /* make sure we break out early on quiesce */
3949 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3950 /* always break with lock held - we unlock outside loop */
3954 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3955 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3958 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3959 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3960 /* if conn is dying, mark tx in tx_ref_table for
3961 * kgnilnd_complete_closed_conn to finish up */
3962 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3964 /* tx was moved to DYING, get next */
3967 spin_unlock(&dev->gnd_rdmaq_lock);
3969 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3970 spin_lock(&dev->gnd_rdmaq_lock);
3973 /* no ticket! add back to head */
3974 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3975 /* clear found_work so scheduler threads wait for timer */
3979 /* TX is GO for launch */
3980 tx->tx_qtime = jiffies;
3981 kgnilnd_send_mapped_tx(tx, 0);
3985 spin_unlock(&dev->gnd_rdmaq_lock);
3991 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3993 __swab64s(&d->gnrd_key.qword1);
3994 __swab64s(&d->gnrd_key.qword2);
3995 __swab64s(&d->gnrd_addr);
3996 __swab32s(&d->gnrd_nob);
3999 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
4000 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
4003 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
4005 kgn_tx_ev_id_t ev_id;
4008 /* we use the cookie from the original TX, so we can find the match
4009 * by parsing that and using the txe_idx */
4010 ev_id.txe_cookie = cookie;
4012 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
4015 /* check tx to make sure kgni didn't eat it */
4016 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
4017 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
4019 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
4020 (tx->tx_id.txe_cookie = cookie)),
4021 "conn 0x%p->%s tx_ref_table hosed: wanted "
4022 "txe_cookie %#llx txe_idx %d "
4023 "found tx %p cookie %#llx txe_idx %d\n",
4024 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4025 cookie, ev_id.txe_idx,
4026 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
4028 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4029 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4030 "Unexpected TX type (%x, %x or %x) "
4031 "or state (%x, expected +%x) "
4032 "matched reply from %s\n",
4033 tx->tx_msg.gnm_type, type1, type2,
4034 tx->tx_state, GNILND_TX_WAITING_REPLY,
4035 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4037 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
4038 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4044 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4047 kgn_conn_t *conn = tx->tx_conn;
4048 __u64 nob = tx->tx_nob;
4049 __u32 physnop = tx->tx_phys_npages;
4050 int id = tx->tx_id.txe_smsg_id;
4051 int buftype = tx->tx_buftype;
4052 gni_mem_handle_t hndl;
4053 hndl.qword1 = tx->tx_map_key.qword1;
4054 hndl.qword2 = tx->tx_map_key.qword2;
4056 spin_lock(&conn->gnc_list_lock);
4058 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4059 "not waiting for reply", NULL);
4062 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4064 if (rc == -EFAULT) {
4065 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
4066 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4068 if(*kgnilnd_tunables.kgn_efault_lbug) {
4069 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4070 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4072 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4073 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4074 kgnilnd_tx_state2str(tx->tx_list_state),
4075 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4080 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4081 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4082 /* sample under lock as follow on steps require gnc_list_lock
4083 * - or call kgnilnd_tx_done which requires no locks held over
4084 * call to lnet_finalize */
4087 spin_unlock(&conn->gnc_list_lock);
4090 kgnilnd_tx_done(tx, tx->tx_rc);
4095 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4098 kgn_conn_t *conn = tx->tx_conn;
4100 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4101 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4103 /* the gncm_retval is passed in for PUTs */
4104 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4105 msg->gnm_u.completion.gncm_retval);
4107 kgnilnd_complete_tx(tx, rc);
4111 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4119 kgn_peer_t *peer = conn->gnc_peer;
4122 __u16 tmp_cksum = 0, msg_cksum = 0;
4123 int repost = 1, saw_complete;
4124 unsigned long timestamp, newest_last_rx, timeout;
4128 /* Short circuit if the ep_handle is null.
4129 * It's likely that its about to be closed as stale.
4131 if (conn->gnc_ephandle == NULL)
4134 timestamp = jiffies;
4135 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4136 /* delay in jiffies - we are really concerned only with things that
4137 * result in a schedule() or really holding this off for long times .
4138 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4139 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4141 /* Resample current time as we have no idea how long it took to get the mutex */
4142 timestamp = jiffies;
4144 /* We check here when the last time we received an rx, we do this before
4145 * we call getnext in case the thread has been blocked for a while. If we
4146 * havent received an rx since our timeout value we close the connection
4147 * as we should assume the other side has closed the connection. This will
4148 * stop us from sending replies to a mailbox that is already in purgatory.
4151 timeout = cfs_time_seconds(conn->gnc_timeout);
4152 newest_last_rx = GNILND_LASTRX(conn);
4154 /* Error injection to validate that timestamp checking works and closing the conn */
4155 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4156 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4159 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4160 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4161 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4162 cfs_duration_sec(timestamp - newest_last_rx),
4163 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4164 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4166 kgnilnd_close_conn(conn, rc);
4170 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4172 if (rrc == GNI_RC_NOT_DONE) {
4173 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4174 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4178 /* Instead of asserting when we get mailbox corruption lets attempt to
4179 * close the conn and recover. We can put the conn/mailbox into
4180 * purgatory and let purgatory deal with the problem. If we see
4181 * this NETTERROR reported on production systems in large amounts
4182 * we will need to revisit the state machine to see if we can tighten
4183 * it up further to improve data protection.
4186 if (rrc == GNI_RC_INVALID_STATE) {
4187 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4188 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4189 "detected closing conn %p from peer %s\n", conn,
4190 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4192 kgnilnd_close_conn(conn, rc);
4196 LASSERTF(rrc == GNI_RC_SUCCESS,
4197 "bad rc %d on conn %p from peer %s\n",
4198 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4200 msg = (kgn_msg_t *)prefix;
4202 rx = kgnilnd_alloc_rx();
4204 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4205 kgnilnd_release_msg(conn);
4206 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4207 conn, libcfs_nid2str(peer->gnp_nid));
4211 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4213 timestamp = conn->gnc_last_rx;
4214 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4215 atomic_inc(&conn->gnc_rx_seq);
4217 conn->gnc_last_rx = jiffies;
4218 /* stash first rx so we can clear out purgatory
4220 if (conn->gnc_first_rx == 0)
4221 conn->gnc_first_rx = jiffies;
4223 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4224 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4225 kgnilnd_peer_alive(conn->gnc_peer);
4228 rx->grx_conn = conn;
4230 rx->grx_received = current_kernel_time();
4232 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4235 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4241 kgnilnd_net_decref(net);
4244 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4245 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4247 /* XXX Nic: Do we need to swab cksum */
4248 if (msg->gnm_cksum != 0) {
4249 msg_cksum = msg->gnm_cksum;
4251 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4253 if (tmp_cksum != msg_cksum) {
4254 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4255 tmp_cksum, msg_cksum);
4256 kgnilnd_dump_msg(D_BUFFS, msg);
4261 /* restore checksum for future debug messages */
4262 msg->gnm_cksum = tmp_cksum;
4264 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4265 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4266 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4267 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4272 __swab32s(&msg->gnm_magic);
4273 __swab16s(&msg->gnm_version);
4274 __swab16s(&msg->gnm_type);
4275 __swab64s(&msg->gnm_srcnid);
4276 __swab64s(&msg->gnm_connstamp);
4277 __swab32s(&msg->gnm_seq);
4279 /* NB message type checked below; NOT here... */
4280 switch (msg->gnm_type) {
4281 case GNILND_MSG_GET_ACK_REV:
4282 case GNILND_MSG_PUT_ACK:
4283 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4286 case GNILND_MSG_PUT_REQ_REV:
4287 case GNILND_MSG_GET_REQ:
4288 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4296 if (msg->gnm_version != GNILND_MSG_VERSION) {
4297 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4298 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4303 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4304 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4305 libcfs_nid2str(msg->gnm_srcnid),
4306 libcfs_nid2str(peer->gnp_nid));
4311 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4312 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4313 " expected) from %s",
4314 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4315 libcfs_nid2str(peer->gnp_nid));
4320 if (msg->gnm_seq != seq) {
4321 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4322 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4327 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4329 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4330 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4331 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4332 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4333 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4334 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4335 /* double check state with lock held */
4336 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4337 /* only error if we are not already closing */
4338 if (conn->gnc_peer_error == -ETIMEDOUT) {
4339 unsigned long now = jiffies;
4340 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4342 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4343 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4344 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4346 cfs_duration_sec(now - timestamp),
4347 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4348 atomic_read(&conn->gnc_tx_seq),
4349 cfs_duration_sec(now - conn->gnc_last_tx),
4350 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4351 cfs_duration_sec(now - conn->gnc_last_noop_want),
4352 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4353 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4354 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4355 cfs_duration_sec(now - conn->gnc_last_sched_do),
4356 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4358 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4360 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4364 if (conn->gnc_close_recvd) {
4365 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4366 kgnilnd_msgtype2str(msg->gnm_type),
4367 msg->gnm_type, conn->gnc_close_recvd,
4368 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4373 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4374 /* XXX Nic: log message received on bad connection state */
4378 switch (msg->gnm_type) {
4379 case GNILND_MSG_NOOP:
4380 /* Nothing to do; just a keepalive */
4383 case GNILND_MSG_IMMEDIATE:
4384 /* only get SMSG payload for IMMEDIATE */
4385 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4386 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4387 msg->gnm_srcnid, rx, 0);
4390 case GNILND_MSG_GET_REQ_REV:
4391 case GNILND_MSG_PUT_REQ:
4392 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4393 msg->gnm_srcnid, rx, 1);
4396 case GNILND_MSG_GET_NAK_REV:
4397 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4398 msg->gnm_u.completion.gncm_cookie);
4402 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4404 case GNILND_MSG_PUT_NAK:
4405 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4406 msg->gnm_u.completion.gncm_cookie);
4410 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4412 case GNILND_MSG_PUT_ACK:
4413 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4414 msg->gnm_u.putack.gnpam_src_cookie);
4418 /* store putack data for later: deferred rdma or re-try */
4419 tx->tx_putinfo = msg->gnm_u.putack;
4422 spin_lock(&tx->tx_conn->gnc_list_lock);
4424 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4425 "not waiting for reply", NULL);
4427 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4429 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4430 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4431 /* sample under lock as follow on steps require gnc_list_lock
4432 * - or call kgnilnd_tx_done which requires no locks held over
4433 * call to lnet_finalize */
4436 /* cannot launch rdma if still waiting for fma-msg completion */
4437 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4438 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4439 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4441 spin_unlock(&tx->tx_conn->gnc_list_lock);
4444 rc = kgnilnd_send_mapped_tx(tx, 0);
4446 kgnilnd_tx_done(tx, rc);
4449 case GNILND_MSG_GET_ACK_REV:
4450 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4451 msg->gnm_u.putack.gnpam_src_cookie);
4455 /* store putack data for later: deferred rdma or re-try */
4456 tx->tx_putinfo = msg->gnm_u.putack;
4458 spin_lock(&tx->tx_conn->gnc_list_lock);
4460 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4461 "not waiting for reply", NULL);
4463 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4465 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4466 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4467 /* sample under lock as follow on steps require gnc_list_lock
4468 * - or call kgnilnd_tx_done which requires no locks held over
4469 * call to lnet_finalize */
4472 /* cannot launch rdma if still waiting for fma-msg completion */
4473 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4474 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4475 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4477 spin_unlock(&tx->tx_conn->gnc_list_lock);
4480 rc = kgnilnd_send_mapped_tx(tx, 0);
4482 kgnilnd_tx_done(tx, rc);
4485 case GNILND_MSG_PUT_DONE:
4486 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4487 msg->gnm_u.completion.gncm_cookie);
4491 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4492 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4493 "bad tx buftype %d", tx->tx_buftype);
4495 kgnilnd_finalize_rx_done(tx, msg);
4497 case GNILND_MSG_PUT_REQ_REV:
4498 case GNILND_MSG_GET_REQ:
4499 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4500 msg->gnm_srcnid, rx, 1);
4504 case GNILND_MSG_GET_NAK:
4505 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4506 msg->gnm_u.completion.gncm_cookie);
4510 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4511 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4512 "bad tx buftype %d", tx->tx_buftype);
4514 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4517 case GNILND_MSG_GET_DONE:
4518 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4519 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 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4528 msg->gnm_u.completion.gncm_retval);
4530 kgnilnd_finalize_rx_done(tx, msg);
4532 case GNILND_MSG_GET_DONE_REV:
4533 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4534 msg->gnm_u.completion.gncm_cookie);
4538 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4539 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4540 "bad tx buftype %d", tx->tx_buftype);
4542 kgnilnd_finalize_rx_done(tx, msg);
4545 case GNILND_MSG_PUT_DONE_REV:
4546 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4547 msg->gnm_u.completion.gncm_cookie);
4552 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4553 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4554 "bad tx buftype %d", tx->tx_buftype);
4556 kgnilnd_finalize_rx_done(tx, msg);
4558 case GNILND_MSG_PUT_NAK_REV:
4559 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4560 msg->gnm_u.completion.gncm_cookie);
4565 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4566 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4567 "bad tx buftype %d", tx->tx_buftype);
4569 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4574 if (rc < 0) /* protocol/comms error */
4575 kgnilnd_close_conn(conn, rc);
4577 if (repost && rx != NULL) {
4578 kgnilnd_consume_rx(rx);
4581 /* we got an event so assume more there and call for reschedule */
4583 kgnilnd_schedule_conn(conn);
4587 /* Do the failure injections that we need to affect conn processing in the following function.
4588 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4589 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4592 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4596 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4600 /* short circuit out when not set */
4601 if (likely(!cfs_fail_loc)) {
4605 /* failure injection to test for stack reset clean ups */
4606 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4607 /* we can't rely on busy loops being nice enough to get the
4608 * stack reset triggered - it'd just spin on this conn */
4609 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4612 GOTO(did_fail_loc, rc);
4615 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4616 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4618 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4619 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4622 GOTO(did_fail_loc, rc);
4626 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4627 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4628 * will be closed by kgnilnd_complete_closed_conn.
4630 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4631 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4634 GOTO(did_fail_loc, rc);
4637 /* this one is a bit gross - we can't hold the mutex from process_conns
4638 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4639 * and moving onto finish_connect
4640 * so, we'll just set the rc - kgnilnd_process_conns will clear
4641 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4642 * and effectively getting this thread to sleep */
4643 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4646 GOTO(did_fail_loc, rc);
4654 kgnilnd_send_conn_close(kgn_conn_t *conn)
4658 /* we are closing the conn - we will try to send the CLOSE msg
4659 * but will not wait for anything else to flush */
4661 /* send the close if not already done so or received one */
4662 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4663 /* set close_sent regardless of the success of the
4664 * CLOSE message. We are going to try once and then
4665 * kick him out of the sandbox */
4666 conn->gnc_close_sent = 1;
4669 /* EP might be null already if remote side initiated a new connection.
4670 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4671 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4673 if (conn->gnc_ephandle != NULL) {
4676 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4678 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4679 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4680 tx->tx_qtime = jiffies;
4682 if (tx->tx_id.txe_idx == 0) {
4683 rc = kgnilnd_set_tx_id(tx, conn);
4685 kgnilnd_tx_done(tx, rc);
4689 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4692 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4693 kgnilnd_tx_done(tx, -EAGAIN);
4695 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4697 /* It wasnt sent and we dont care. */
4698 kgnilnd_tx_done(tx, rc);
4706 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4707 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4708 conn->gnc_state = GNILND_CONN_CLOSED;
4709 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4710 /* mark this conn as CLOSED now that we processed it
4711 * do after TX, so we can use CLOSING in asserts */
4715 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4716 /* simulate a RX CLOSE after the timeout but before
4717 * the scheduler thread gets it */
4718 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4719 conn->gnc_peer_error = -ETIMEDOUT;
4721 /* schedule to allow potential CLOSE and get the complete phase run */
4722 kgnilnd_schedule_conn(conn);
4726 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4731 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4732 int log_retrans, log_retrans_level;
4733 static int last_map_version;
4736 spin_lock(&dev->gnd_lock);
4737 if (list_empty(&dev->gnd_map_tx)) {
4738 /* if the list is empty make sure we dont have a timer running */
4739 del_singleshot_timer_sync(&dev->gnd_map_timer);
4740 spin_unlock(&dev->gnd_lock);
4744 dev->gnd_sched_alive = jiffies;
4746 /* we'll retry as fast as possible up to 25% of the limit, then we start
4747 * backing off until our map version changes - indicating we unmapped
4749 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4750 if (likely(dev->gnd_map_attempt == 0) ||
4751 time_after_eq(jiffies, dev->gnd_next_map) ||
4752 last_map_version != dev->gnd_map_version) {
4754 /* if this is our first attempt at mapping set last mapped to current
4755 * jiffies so we can timeout our attempt correctly.
4757 if (dev->gnd_map_attempt == 0)
4758 dev->gnd_last_map = jiffies;
4760 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4761 spin_unlock(&dev->gnd_lock);
4765 /* delete the previous timer if it exists */
4766 del_singleshot_timer_sync(&dev->gnd_map_timer);
4767 /* stash the last map version to let us know when a good one was seen */
4768 last_map_version = dev->gnd_map_version;
4770 /* we need to to take the lock and continually refresh the head of the list as
4771 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4772 * allowing them to squeeze in */
4774 while (!list_empty(&dev->gnd_map_tx)) {
4775 /* make sure we break out early on quiesce */
4776 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4777 /* always break with lock held - we unlock outside loop */
4781 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4783 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4786 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4787 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4788 /* if conn is dying, mark tx in tx_ref_table for
4789 * kgnilnd_complete_closed_conn to finish up */
4790 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4793 /* tx was moved to DYING, get next */
4797 spin_unlock(&dev->gnd_lock);
4798 rc = kgnilnd_send_mapped_tx(tx, 1);
4800 /* We made it! skip error handling.. */
4802 /* OK to continue on +ve errors as it won't get seen until
4803 * this function is called again - we operate on a copy of the original
4804 * list and not the live list */
4805 spin_lock(&dev->gnd_lock);
4806 /* reset map attempts back to zero we successfully
4807 * mapped so we can reset our timers */
4808 dev->gnd_map_attempt = 0;
4810 } else if (rc == -EAGAIN) {
4811 spin_lock(&dev->gnd_lock);
4812 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4813 spin_unlock(&dev->gnd_lock);
4814 GOTO(get_out_mapped, rc);
4815 } else if (rc != -ENOMEM) {
4816 /* carp, failure we can't handle */
4817 kgnilnd_tx_done(tx, rc);
4818 spin_lock(&dev->gnd_lock);
4819 /* reset map attempts back to zero we dont know what happened but it
4820 * wasnt a failed mapping
4822 dev->gnd_map_attempt = 0;
4826 /* time to handle the retry cases.. lock so we dont have 2 threads
4827 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4829 spin_lock(&dev->gnd_lock);
4830 dev->gnd_map_attempt++;
4831 if (dev->gnd_map_attempt < fast_remaps) {
4832 /* do nothing we just want it to go as fast as possible.
4833 * just set gnd_next_map to current jiffies so it will process
4834 * as fast as possible.
4836 dev->gnd_next_map = jiffies;
4838 /* Retry based on GNILND_MAP_RETRY_RATE */
4839 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4842 /* only log occasionally once we've retried fast_remaps */
4843 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4844 ((dev->gnd_map_attempt % fast_remaps) == 0);
4845 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4847 /* make sure we are not off in the weeds with this tx */
4848 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4849 GNIDBG_TX(D_NETERROR, tx,
4850 "giving up on TX, too many retries", NULL);
4851 spin_unlock(&dev->gnd_lock);
4852 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4853 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4854 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4856 tx->tx_putinfo.gnpam_dst_cookie,
4857 tx->tx_msg.gnm_srcnid);
4859 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4861 tx->tx_getinfo.gngm_cookie,
4862 tx->tx_msg.gnm_srcnid);
4864 kgnilnd_tx_done(tx, -ENOMEM);
4865 GOTO(get_out_mapped, rc);
4867 GNIDBG_TX(log_retrans_level, tx,
4868 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4870 "nq_map %d mdd# %d/%d GART %ld",
4871 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4872 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4873 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4874 atomic_read(&dev->gnd_nq_map),
4875 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4876 atomic64_read(&dev->gnd_nbytes_map));
4879 /* we need to stop processing the rest of the list, so add it back in */
4880 /* set timer to wake device when we need to schedule this tx */
4881 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4882 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4883 spin_unlock(&dev->gnd_lock);
4884 GOTO(get_out_mapped, rc);
4886 spin_unlock(&dev->gnd_lock);
4892 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4897 int error_inject = 0;
4901 spin_lock(&dev->gnd_lock);
4902 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4903 dev->gnd_sched_alive = jiffies;
4907 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4908 /* break with lock held */
4912 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4913 list_del_init(&conn->gnc_schedlist);
4915 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4918 if (!list_empty(&conn->gnc_delaylist))
4919 list_del_init(&conn->gnc_delaylist);
4920 spin_unlock(&dev->gnd_lock);
4922 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4924 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4925 conn_sched != GNILND_CONN_PROCESS,
4926 "conn %p on ready list but in bad state: %d\n",
4929 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4930 conn, kgnilnd_conn_state2str(conn));
4933 set_mb(conn->gnc_last_sched_do, jiffies);
4935 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4937 /* based on intent see if we should run again. */
4938 rc = kgnilnd_schedule_process_conn(conn, intent);
4940 /* drop ref from gnd_ready_conns */
4941 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4942 down_write(&dev->gnd_conn_sem);
4943 kgnilnd_conn_decref(conn);
4944 up_write(&dev->gnd_conn_sem);
4945 } else if (rc != 1) {
4946 kgnilnd_conn_decref(conn);
4948 /* clear this so that scheduler thread doesn't spin */
4950 /* break with lock held... */
4951 spin_lock(&dev->gnd_lock);
4955 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4956 down_write(&dev->gnd_conn_sem);
4958 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4959 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4960 /* If there are tx's currently in use in another
4961 * thread we dont want to complete the close
4962 * yet. Cycle this conn back through
4964 kgnilnd_schedule_conn(conn);
4966 kgnilnd_complete_closed_conn(conn);
4968 up_write(&dev->gnd_conn_sem);
4969 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4970 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4971 /* serialize SMSG CQs with ep_bind and smsg_release */
4972 down_write(&dev->gnd_conn_sem);
4973 kgnilnd_destroy_conn_ep(conn);
4974 up_write(&dev->gnd_conn_sem);
4975 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4976 /* if we need to do some CLOSE sending, etc done here do it */
4977 down_write(&dev->gnd_conn_sem);
4978 kgnilnd_send_conn_close(conn);
4979 kgnilnd_check_fma_rx(conn);
4980 up_write(&dev->gnd_conn_sem);
4981 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4982 /* start moving traffic if the old conns are cleared out */
4983 down_read(&dev->gnd_conn_sem);
4984 kgnilnd_check_fma_rx(conn);
4985 kgnilnd_process_fmaq(conn);
4986 up_read(&dev->gnd_conn_sem);
4989 rc = kgnilnd_schedule_process_conn(conn, 0);
4991 /* drop ref from gnd_ready_conns */
4992 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4993 down_write(&dev->gnd_conn_sem);
4994 kgnilnd_conn_decref(conn);
4995 up_write(&dev->gnd_conn_sem);
4996 } else if (rc != 1) {
4997 kgnilnd_conn_decref(conn);
5000 /* check list again with lock held */
5001 spin_lock(&dev->gnd_lock);
5004 /* If we are short circuiting due to timing we want to be scheduled
5005 * as soon as possible.
5007 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
5010 spin_unlock(&dev->gnd_lock);
5016 kgnilnd_scheduler(void *arg)
5018 int threadno = (long)arg;
5021 unsigned long deadline = 0;
5024 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
5026 /* all gnilnd threads need to run fairly urgently */
5027 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
5028 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5029 while (!kgnilnd_data.kgn_shutdown) {
5031 /* Safe: kgn_shutdown only set when quiescent */
5033 /* to quiesce or to not quiesce, that is the question */
5035 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5036 KGNILND_SPIN_QUIESCE;
5039 /* tracking for when thread goes AWOL */
5040 dev->gnd_sched_alive = jiffies;
5042 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5043 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5044 /* let folks know we are up and kicking
5045 * - they can use this for latency savings, etc
5046 * - only change if IRQ, if IDLE leave alone as that
5047 * schedule_device calls to put us back to IRQ */
5048 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5050 down_read(&dev->gnd_conn_sem);
5051 /* always check these - they are super low cost */
5052 found_work += kgnilnd_check_fma_send_cq(dev);
5053 found_work += kgnilnd_check_fma_rcv_cq(dev);
5055 /* rdma CQ doesn't care about eps */
5056 found_work += kgnilnd_check_rdma_cq(dev);
5058 /* move some RDMA ? */
5059 found_work += kgnilnd_process_rdmaq(dev);
5061 /* map some pending RDMA requests ? */
5062 found_work += kgnilnd_process_mapped_tx(dev);
5064 /* the EP for a conn is not destroyed until all the references
5065 * to it are gone, so these checks should be safe
5066 * even if run in parallel with the CQ checking functions
5067 * _AND_ a thread that processes the CLOSED->DONE
5071 up_read(&dev->gnd_conn_sem);
5073 /* process all conns ready now */
5074 found_work += kgnilnd_process_conns(dev, deadline);
5076 /* do an eager check to avoid the IRQ disabling in
5077 * prepare_to_wait and friends */
5080 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5081 time_before(jiffies, deadline)) {
5083 if ((busy_loops % 10) == 0) {
5084 /* tickle heartbeat and watchdog to ensure our
5085 * piggishness doesn't turn into heartbeat failure */
5086 touch_nmi_watchdog();
5092 /* if we got here, found_work was zero or busy_loops means we
5093 * need to take a break. We'll clear gnd_ready but we'll check
5094 * one last time if there is an IRQ that needs processing */
5096 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5098 /* the first time this will go LOOP -> IDLE and let us do one final check
5099 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5100 * - this might allow other threads to block us for a bit if they
5101 * try to get the mutex, but that is good as we'd need to wake
5102 * up soon to handle the CQ or other processing anyways */
5104 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5106 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5107 time_after_eq(jiffies, deadline)) {
5109 "yeilding: found_work %d busy_loops %d\n",
5110 found_work, busy_loops);
5112 /* use yield if we are bailing due to busy_loops
5113 * - this will ensure we wake up soonish. This closes
5114 * a race with kgnilnd_device_callback - where it'd
5115 * not call wake_up() because gnd_ready == 1, but then
5116 * we come down and schedule() because of busy_loops.
5117 * We'd not be woken up until something poked our waitq
5118 * again. yield() ensures we wake up without another
5119 * waitq poke in that case */
5120 atomic_inc(&dev->gnd_n_yield);
5121 kgnilnd_data.kgn_last_condresched = jiffies;
5123 CDEBUG(D_INFO, "awake after yeild\n");
5124 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5125 } else if (found_work == GNILND_DEV_IDLE) {
5126 /* busy_loops is low and there is nothing to do,
5127 * go to sleep and wait for a waitq poke */
5129 "scheduling: found_work %d busy_loops %d\n",
5130 found_work, busy_loops);
5131 atomic_inc(&dev->gnd_n_schedule);
5132 kgnilnd_data.kgn_last_scheduled = jiffies;
5134 CDEBUG(D_INFO, "awake after schedule\n");
5135 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5137 finish_wait(&dev->gnd_waitq, &wait);
5140 kgnilnd_thread_fini();