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, lnet_kiov_t *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->kiov_len) {
343 offset -= kiov->kiov_len;
349 /* ignore nob here, if nob < (kiov_len - offset), kiov == 1 */
350 odd = (unsigned long) (kiov[0].kiov_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].kiov_len, offset, nob);
361 for (i = 0; i < nkiov; i++) {
362 pages[i] = kiov[i].kiov_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].kiov_offset + offset, nob);
379 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
380 (void *)addr + kiov[0].kiov_offset + offset, nob);
382 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
383 cksum, tmpck, addr, kiov[0].kiov_offset, nob, offset);
387 fraglen = min(kiov->kiov_len - offset, nob);
389 /* make dang sure we don't send a bogus checksum if somehow we get
390 * an odd length fragment on anything but the last entry in a kiov -
391 * we know from kgnilnd_setup_rdma_buffer that we can't have non
392 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
393 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
394 "odd fraglen %u on nkiov %d, nob %u kiov_len %u offset %u kiov 0x%p\n",
395 fraglen, nkiov, nob, kiov->kiov_len, offset, kiov);
397 addr = (void *)kmap(kiov->kiov_page) + kiov->kiov_offset + offset;
398 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
401 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
402 cksum, tmpck, kiov->kiov_page, kiov->kiov_offset, addr,
408 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
410 kunmap(kiov->kiov_page);
417 /* iov must not run out before end of data */
418 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
423 retsum = csum_fold(cksum);
425 /* don't use magic 'no checksum' value */
429 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
435 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
437 msg->gnm_magic = GNILND_MSG_MAGIC;
438 msg->gnm_version = GNILND_MSG_VERSION;
439 msg->gnm_type = type;
440 msg->gnm_payload_len = 0;
441 msg->gnm_srcnid = source;
442 /* gnm_connstamp gets set when FMA is sent */
443 /* gnm_srcnid is set on creation via function argument
444 * The right interface/net and nid is passed in when the message
450 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
452 kgn_tx_t *tx = kgnilnd_alloc_tx();
455 kgnilnd_init_msg(&tx->tx_msg, type, source);
457 CERROR("couldn't allocate new tx type %s!\n",
458 kgnilnd_msgtype2str(type));
465 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
471 case GNILND_MSG_GET_REQ:
472 case GNILND_MSG_GET_DONE:
473 nak_type = GNILND_MSG_GET_NAK;
475 case GNILND_MSG_PUT_REQ:
476 case GNILND_MSG_PUT_ACK:
477 case GNILND_MSG_PUT_DONE:
478 nak_type = GNILND_MSG_PUT_NAK;
480 case GNILND_MSG_PUT_REQ_REV:
481 case GNILND_MSG_PUT_DONE_REV:
482 nak_type = GNILND_MSG_PUT_NAK_REV;
484 case GNILND_MSG_GET_REQ_REV:
485 case GNILND_MSG_GET_ACK_REV:
486 case GNILND_MSG_GET_DONE_REV:
487 nak_type = GNILND_MSG_GET_NAK_REV;
490 CERROR("invalid msg type %s (%d)\n",
491 kgnilnd_msgtype2str(rx_type), rx_type);
494 /* only allow NAK on error and truncate to zero */
495 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
496 error, conn, cookie);
498 tx = kgnilnd_new_tx_msg(nak_type, source);
500 CNETERR("can't get TX to NAK RDMA to %s\n",
501 libcfs_nid2str(conn->gnc_peer->gnp_nid));
505 tx->tx_msg.gnm_u.completion.gncm_retval = error;
506 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
507 kgnilnd_queue_tx(conn, tx);
511 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
512 struct kvec *iov, lnet_kiov_t *kiov,
513 unsigned int offset, unsigned int nob)
515 kgn_msg_t *msg = &tx->tx_msg;
518 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
519 * gni_smsg_send to send that as the payload */
521 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
524 tx->tx_buffer = NULL;
525 } else if (kiov != NULL) {
527 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
528 niov = ((nob + offset + kiov->kiov_offset + PAGE_SIZE - 1) /
532 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
533 "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n",
534 niov, msg, kiov, iov, offset, nob);
536 while (offset >= kiov->kiov_len) {
537 offset -= kiov->kiov_len;
542 for (i = 0; i < niov; i++) {
543 /* We can't have a kiov_offset on anything but the first entry,
544 * otherwise we'll have a hole at the end of the mapping as we only map
546 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
547 * than kiov_len, we will also have a whole at the end of that page
548 * which isn't allowed */
549 if ((kiov[i].kiov_offset != 0 && i > 0) ||
550 (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_SIZE && i < niov - 1)) {
551 CNETERR("Can't make payload contiguous in I/O VM: page %d, offset %u, nob %u, kiov_offset %u, kiov_len %u\n",
552 i, offset, nob, kiov->kiov_offset, kiov->kiov_len);
555 tx->tx_imm_pages[i] = kiov[i].kiov_page;
558 /* hijack tx_phys for the later unmap */
560 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
562 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) + kiov[0].kiov_offset + offset;
563 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
564 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
565 nob, kiov, tx->tx_buffer);
567 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
568 if (tx->tx_phys == NULL) {
569 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
573 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
574 /* make sure we take into account the kiov offset as the start of the buffer */
575 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].kiov_offset + offset;
576 GNIDBG_TX(D_NET, tx, "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
577 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
579 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
583 /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
584 * could "flatten" the payload into a single contiguous buffer ready
585 * for sending direct over an FMA if we ever needed to. */
589 while (offset >= iov->iov_len) {
590 offset -= iov->iov_len;
596 if (nob > iov->iov_len - offset) {
597 CERROR("Can't handle multiple vaddr fragments\n");
601 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
603 tx->tx_buftype = GNILND_BUF_IMMEDIATE;
607 /* checksum payload early - it shouldn't be changing after lnd_send */
608 if (*kgnilnd_tunables.kgn_checksum >= 2) {
609 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
610 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
611 msg->gnm_payload_cksum += 0xe00e;
613 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
614 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
618 msg->gnm_payload_cksum = 0;
625 kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
626 unsigned int niov, struct kvec *iov,
627 unsigned int offset, unsigned int nob)
632 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
634 while (offset >= iov->iov_len) {
635 offset -= iov->iov_len;
641 if (nob > iov->iov_len - offset) {
642 CERROR("Can't handle multiple vaddr fragments\n");
646 tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
648 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
653 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
654 unsigned int offset, unsigned int nob)
656 gni_mem_segment_t *phys;
658 unsigned int fraglen;
660 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
664 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
666 /* only allocate this if we are going to use it */
667 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
669 if (tx->tx_phys == NULL) {
670 CERROR("failed to allocate tx_phys\n");
675 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
676 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
678 /* if loops changes, please change kgnilnd_cksum_kiov
679 * and kgnilnd_setup_immediate_buffer */
681 while (offset >= kiov->kiov_len) {
682 offset -= kiov->kiov_len;
688 /* at this point, kiov points to the first page that we'll actually map
689 * now that we've seeked into the koiv for offset and dropped any
690 * leading pages that fall entirely within the offset */
691 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
694 /* kiov_offset is start of 'valid' buffer, so index offset past that */
695 tx->tx_buffer = (void *)((unsigned long)(kiov->kiov_offset + offset));
698 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
699 tx, tx->tx_buffer, kiov, kiov->kiov_offset, nkiov, offset);
702 fraglen = min(kiov->kiov_len - offset, nob);
704 /* We can't have a kiov_offset on anything but the first entry,
705 * otherwise we'll have a hole at the end of the mapping as we only map
706 * whole pages. Only the first page is allowed to have an offset -
707 * we'll add that into tx->tx_buffer and that will get used when we
708 * map in the segments (see kgnilnd_map_buffer).
709 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
710 * than kiov_len, we will also have a whole at the end of that page
711 * which isn't allowed */
712 if ((phys != tx->tx_phys) &&
713 ((kiov->kiov_offset != 0) ||
714 ((kiov->kiov_len < PAGE_SIZE) && (nob > kiov->kiov_len)))) {
715 CERROR("Can't make payload contiguous in I/O VM: page %d, offset %u, nob %u, kiov_offset %u, kiov_len %u\n",
716 (int)(phys - tx->tx_phys),
717 offset, nob, kiov->kiov_offset, kiov->kiov_len);
722 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
723 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
728 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
733 CDEBUG(D_BUFFS, "page 0x%p kiov_offset %u kiov_len %u nob %u "
734 "nkiov %u offset %u\n",
735 kiov->kiov_page, kiov->kiov_offset, kiov->kiov_len, nob, nkiov, offset);
737 phys->address = page_to_phys(kiov->kiov_page);
744 /* iov must not run out before end of data */
745 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
749 tx->tx_phys_npages = phys - tx->tx_phys;
754 if (tx->tx_phys != NULL) {
755 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
756 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
757 sizeof(*tx->tx_phys), tx->tx_phys);
764 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
765 struct kvec *iov, lnet_kiov_t *kiov,
766 unsigned int offset, unsigned int nob)
770 LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
771 " offset %d, nob %d, niov %d\n"
772 , iov, kiov, tx, offset, nob, niov);
775 rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
777 rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
782 /* kgnilnd_parse_lnet_rdma()
783 * lntmsg - message passed in from lnet.
784 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
785 * nob - actual number of bytes to in this message.
786 * put_len - It is possible for PUTs to have a different length than the
787 * length stored in lntmsg->msg_len since LNET can adjust this
788 * length based on it's buffer size and offset.
789 * lnet_try_match_md() sets the mlength that we use to do the RDMA
793 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
794 unsigned int *offset, unsigned int *nob,
795 lnet_kiov_t **kiov, int put_len)
797 /* GETs are weird, see kgnilnd_send */
798 if (lntmsg->msg_type == LNET_MSG_GET) {
799 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
802 *kiov = lntmsg->msg_md->md_iov.kiov;
804 *niov = lntmsg->msg_md->md_niov;
805 *nob = lntmsg->msg_md->md_length;
808 *kiov = lntmsg->msg_kiov;
809 *niov = lntmsg->msg_niov;
811 *offset = lntmsg->msg_offset;
816 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
818 unsigned int niov, offset, nob;
820 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
821 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
823 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
824 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
825 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
826 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
827 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
828 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
829 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
831 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
832 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
833 tx->tx_msg.gnm_payload_cksum = 0;
836 if (*kgnilnd_tunables.kgn_checksum < 3) {
837 tx->tx_msg.gnm_payload_cksum = 0;
841 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
843 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
847 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
849 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
851 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
855 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
856 tx->tx_msg.gnm_payload_cksum += 0xd00d;
860 /* kgnilnd_verify_rdma_cksum()
861 * tx - PUT_DONE/GET_DONE matched tx.
862 * rx_cksum - received checksum to compare against.
863 * put_len - see kgnilnd_parse_lnet_rdma comments.
866 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
870 unsigned int niov, offset, nob;
872 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
873 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
875 /* we can only match certain requests */
876 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
877 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
878 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
879 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
880 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
881 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
882 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
884 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
885 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
890 if (*kgnilnd_tunables.kgn_checksum >= 3) {
891 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
892 "no RDMA payload checksum when enabled");
897 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
899 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
902 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
904 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
907 if (cksum != rx_cksum) {
908 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
909 "Bad RDMA payload checksum (%x expected %x); "
910 "kiov 0x%p niov %d nob %u offset %u",
911 cksum, rx_cksum, kiov, niov, nob, offset);
912 switch (dump_on_err) {
915 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
917 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
920 /* fall through to dump log */
922 libcfs_debug_dumplog();
928 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
934 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
938 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
939 "already mapped!", NULL);
941 spin_lock(&dev->gnd_map_lock);
942 switch (tx->tx_buftype) {
944 GNIDBG_TX(D_EMERG, tx,
945 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
946 spin_unlock(&dev->gnd_map_lock);
950 case GNILND_BUF_PHYS_MAPPED:
951 bytes = tx->tx_phys_npages * PAGE_SIZE;
952 dev->gnd_map_nphys++;
953 dev->gnd_map_physnop += tx->tx_phys_npages;
956 case GNILND_BUF_VIRT_MAPPED:
958 dev->gnd_map_nvirt++;
959 dev->gnd_map_virtnob += tx->tx_nob;
963 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
964 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
965 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
966 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
967 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
970 atomic_inc(&dev->gnd_n_mdd);
971 atomic64_add(bytes, &dev->gnd_nbytes_map);
973 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
976 /* we only get here in the valid cases */
977 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
978 dev->gnd_map_version++;
979 spin_unlock(&dev->gnd_map_lock);
983 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
987 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
988 "not mapped!", NULL);
989 spin_lock(&dev->gnd_map_lock);
991 switch (tx->tx_buftype) {
993 GNIDBG_TX(D_EMERG, tx,
994 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
995 spin_unlock(&dev->gnd_map_lock);
999 case GNILND_BUF_PHYS_UNMAPPED:
1000 bytes = tx->tx_phys_npages * PAGE_SIZE;
1001 dev->gnd_map_nphys--;
1002 dev->gnd_map_physnop -= tx->tx_phys_npages;
1005 case GNILND_BUF_VIRT_UNMAPPED:
1007 dev->gnd_map_nvirt--;
1008 dev->gnd_map_virtnob -= tx->tx_nob;
1012 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1013 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1014 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
1015 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
1016 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
1017 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
1018 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1021 atomic_dec(&dev->gnd_n_mdd);
1022 atomic64_sub(bytes, &dev->gnd_nbytes_map);
1024 /* we only get here in the valid cases */
1025 list_del_init(&tx->tx_map_list);
1026 dev->gnd_map_version++;
1027 spin_unlock(&dev->gnd_map_lock);
1031 kgnilnd_map_buffer(kgn_tx_t *tx)
1033 kgn_conn_t *conn = tx->tx_conn;
1034 kgn_device_t *dev = conn->gnc_device;
1035 __u32 flags = GNI_MEM_READWRITE;
1038 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
1039 * be called concurrently as there are internal locks that protect
1040 * any data structures or HW resources. We just need to ensure
1041 * that our concurrency doesn't result in the kgn_device_t
1042 * getting nuked while we are in here */
1044 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
1045 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
1047 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1050 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1051 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1054 switch (tx->tx_buftype) {
1058 case GNILND_BUF_NONE:
1059 case GNILND_BUF_IMMEDIATE:
1060 case GNILND_BUF_IMMEDIATE_KIOV:
1061 case GNILND_BUF_PHYS_MAPPED:
1062 case GNILND_BUF_VIRT_MAPPED:
1065 case GNILND_BUF_PHYS_UNMAPPED:
1066 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1067 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1068 tx->tx_phys, tx->tx_phys_npages, NULL,
1069 GNI_MEM_PHYS_SEGMENTS | flags,
1071 /* could race with other uses of the map counts, but this is ok
1072 * - this needs to turn into a non-fatal error soon to allow
1073 * GART resource, etc starvation handling */
1074 if (rrc != GNI_RC_SUCCESS) {
1075 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1076 "phys %u pp %u, virt %u nob %llu",
1077 tx->tx_phys_npages, dev->gnd_id,
1078 dev->gnd_map_nphys, dev->gnd_map_physnop,
1079 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1080 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1083 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1084 kgnilnd_mem_add_map_list(dev, tx);
1087 case GNILND_BUF_VIRT_UNMAPPED:
1088 rrc = kgnilnd_mem_register(dev->gnd_handle,
1089 (__u64)tx->tx_buffer, tx->tx_nob,
1090 NULL, flags, &tx->tx_map_key);
1091 if (rrc != GNI_RC_SUCCESS) {
1092 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1093 "phys %u pp %u, virt %u nob %llu",
1094 tx->tx_nob, dev->gnd_id,
1095 dev->gnd_map_nphys, dev->gnd_map_physnop,
1096 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1097 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1100 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1101 kgnilnd_mem_add_map_list(dev, tx);
1102 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1103 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1104 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1105 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1106 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1114 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1116 kgn_conn_t *conn = tx->tx_conn;
1117 kgn_mdd_purgatory_t *gmp;
1119 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1120 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1121 " asserting to avoid data corruption\n");
1122 if (tx->tx_buffer_copy)
1123 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1125 gmp->gmp_map_key = tx->tx_map_key;
1127 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1129 /* ensure that we don't have a blank purgatory - indicating the
1130 * conn is not already on purgatory lists - we'd never recover these
1131 * MDD if that were the case */
1132 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1133 "conn 0x%p->%s with NULL purgatory",
1134 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1136 /* link 'er up! - only place we really need to lock for
1137 * concurrent access */
1138 spin_lock(&conn->gnc_list_lock);
1139 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1140 spin_unlock(&conn->gnc_list_lock);
1144 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1148 int hold_timeout = 0;
1150 /* code below relies on +1 relationship ... */
1151 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1152 (GNILND_BUF_PHYS_UNMAPPED + 1));
1153 BUILD_BUG_ON(GNILND_BUF_VIRT_MAPPED !=
1154 (GNILND_BUF_VIRT_UNMAPPED + 1));
1156 switch (tx->tx_buftype) {
1160 case GNILND_BUF_NONE:
1161 case GNILND_BUF_IMMEDIATE:
1162 case GNILND_BUF_PHYS_UNMAPPED:
1163 case GNILND_BUF_VIRT_UNMAPPED:
1165 case GNILND_BUF_IMMEDIATE_KIOV:
1166 if (tx->tx_phys != NULL) {
1167 vunmap(tx->tx_phys);
1168 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1169 kunmap(tx->tx_imm_pages[0]);
1171 /* clear to prevent kgnilnd_free_tx from thinking
1172 * this is a RDMA descriptor */
1176 case GNILND_BUF_PHYS_MAPPED:
1177 case GNILND_BUF_VIRT_MAPPED:
1178 LASSERT(tx->tx_conn != NULL);
1180 dev = tx->tx_conn->gnc_device;
1182 /* only want to hold if we are closing conn without
1183 * verified peer notification - the theory is that
1184 * a TX error can be communicated in all other cases */
1185 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1186 error != -GNILND_NOPURG &&
1187 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1188 kgnilnd_add_purgatory_tx(tx);
1190 /* The timeout we give to kgni is a deadman stop only.
1191 * we are setting high to ensure we don't have the kgni timer
1192 * fire before ours fires _and_ is handled */
1193 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1195 GNIDBG_TX(D_NET, tx,
1196 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1197 tx->tx_conn->gnc_device, hold_timeout,
1198 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1200 if (tx->tx_buffer_copy != NULL) {
1201 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1202 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1203 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1204 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1206 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1207 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1211 kgnilnd_mem_del_map_list(dev, tx);
1217 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1219 struct lnet_msg *lntmsg0, *lntmsg1;
1220 int status0, status1;
1221 struct lnet_ni *ni = NULL;
1222 kgn_conn_t *conn = tx->tx_conn;
1224 LASSERT(!in_interrupt());
1226 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1227 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1230 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1231 !kgnilnd_conn_clean_errno(completion)) {
1232 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1233 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1234 completion, tx, conn ?
1235 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1236 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1237 kgnilnd_tx_state2str(tx->tx_list_state),
1238 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1241 /* The error codes determine if we hold onto the MDD */
1242 kgnilnd_unmap_buffer(tx, completion);
1244 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1245 * we play nice with the error codes to avoid delivering a failed
1246 * REQUEST and then a REPLY event as well */
1248 /* return -EIO to lnet - it is the magic value for failed sends */
1249 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1251 status1 = completion;
1253 status0 = status1 = completion;
1256 tx->tx_buftype = GNILND_BUF_NONE;
1257 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1259 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1260 * set NULL when we are a tx without a conn */
1262 ni = conn->gnc_peer->gnp_net->gnn_ni;
1264 spin_lock(&conn->gnc_tx_lock);
1266 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1267 (volatile unsigned long *)&conn->gnc_tx_bits),
1268 "conn %p tx %p bit %d already cleared\n",
1269 conn, tx, tx->tx_id.txe_idx);
1271 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1272 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1274 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1275 spin_unlock(&conn->gnc_tx_lock);
1278 kgnilnd_free_tx(tx);
1280 /* finalize AFTER freeing lnet msgs */
1282 /* warning - we should hold no locks here - calling lnet_finalize
1283 * could free up lnet credits, resulting in a call chain back into
1284 * the LND via kgnilnd_send and friends */
1286 lnet_finalize(lntmsg0, status0);
1288 if (lntmsg1 != NULL) {
1289 lnet_finalize(lntmsg1, status1);
1294 kgnilnd_txlist_done(struct list_head *txlist, int error)
1297 int err_printed = 0;
1299 if (list_empty(txlist))
1302 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1303 /* only print the first error */
1305 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1306 list_del_init(&tx->tx_list);
1307 kgnilnd_tx_done(tx, error);
1312 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1316 spin_lock(&conn->gnc_tx_lock);
1318 /* ID zero is NOT ALLOWED!!! */
1321 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1322 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1323 if (id == GNILND_MAX_MSG_ID) {
1324 if (conn->gnc_next_tx != 1) {
1325 /* we only searched from next_tx to end and didn't find
1326 * one, so search again from start */
1327 conn->gnc_next_tx = 1;
1330 /* couldn't find one! */
1331 spin_unlock(&conn->gnc_tx_lock);
1335 /* bump next_tx to prevent immediate reuse */
1336 conn->gnc_next_tx = id + 1;
1338 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1339 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1340 "tx 0x%p already at id %d\n",
1341 conn->gnc_tx_ref_table[id], id);
1343 /* delay these until we have a valid ID - prevents bad clear of the bit
1344 * in kgnilnd_tx_done */
1346 tx->tx_id.txe_cqid = conn->gnc_cqid;
1348 tx->tx_id.txe_idx = id;
1349 conn->gnc_tx_ref_table[id] = tx;
1351 /* Using jiffies to help differentiate against TX reuse - with
1352 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1353 * if we are sending to the same node faster than 256000/sec.
1354 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1356 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1358 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1360 spin_unlock(&conn->gnc_tx_lock);
1365 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1369 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1371 /* we don't care about TX timeout - it could be that the network is slower
1372 * or throttled. We'll keep retranmitting - so if the network is so slow
1373 * that we fill up our mailbox, we'll keep trying to resend that msg
1374 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1375 * that he hasn't send us any traffic in return */
1377 /* some reasonable throttling of the debug message */
1379 unsigned long now = jiffies;
1380 /* XXX Nic: Mystical TX debug here... */
1381 /* We expect retransmissions so only log when D_NET is enabled */
1382 GNIDBG_SMSG_CREDS(D_NET, conn);
1383 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1384 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1385 " last_msg %uus/%uus last_cq %uus/%uus",
1386 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1387 tx->tx_id, tx->tx_retrans,
1388 jiffies_to_usecs(now - tx->tx_cred_wait),
1389 jiffies_to_usecs(now - conn->gnc_last_tx),
1390 jiffies_to_usecs(now - conn->gnc_last_rx),
1391 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1392 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1396 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1397 * to avoid bad ordering with state_lock */
1400 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1401 spinlock_t *state_lock, kgn_tx_list_state_t state)
1403 kgn_conn_t *conn = tx->tx_conn;
1404 kgn_msg_t *msg = &tx->tx_msg;
1406 unsigned long newest_last_rx, timeout;
1409 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1410 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1412 "msg 0x%p type %d wrong payload size %d\n",
1413 msg, msg->gnm_type, immediatenob);
1415 /* make sure we catch all the cases where we'd send on a dirty old mbox
1416 * but allow case for sending CLOSE. Since this check is within the CQ
1417 * mutex barrier and the close message is only sent through
1418 * kgnilnd_send_conn_close the last message out the door will be the
1421 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1422 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1423 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1424 /* Return -ETIME, we are closing the connection already so we dont want to
1425 * have this tx hit the wire. The tx will be killed by the calling function.
1426 * Once the EP is marked dirty the close message will be the last
1427 * thing to hit the wire */
1432 timeout = cfs_time_seconds(conn->gnc_timeout);
1434 newest_last_rx = GNILND_LASTRX(conn);
1436 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1437 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1440 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1441 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1442 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1443 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1444 cfs_duration_sec(now - newest_last_rx),
1445 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1446 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1447 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1451 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1452 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1453 * creating the message this allows the message to contain the correct LNET NID/NET needed
1454 * instead of the one that the peer/conn uses for sending the data.
1456 msg->gnm_connstamp = conn->gnc_my_connstamp;
1457 msg->gnm_payload_len = immediatenob;
1458 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1460 /* always init here - kgn_checksum is a /sys module tunable
1461 * and can be flipped at any point, even between msg init and sending */
1463 if (*kgnilnd_tunables.kgn_checksum) {
1464 /* We must set here and not in kgnilnd_init_msg,
1465 * we could resend this msg many times
1466 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1467 * through init_msg again */
1468 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1469 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1470 msg->gnm_cksum += 0xf00f;
1474 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1475 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1476 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1477 tx->tx_id.txe_idx, immediate, immediatenob);
1479 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1480 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1482 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1483 msg, sizeof(*msg), immediate,
1485 tx->tx_id.txe_smsg_id);
1489 case GNI_RC_SUCCESS:
1490 atomic_inc(&conn->gnc_tx_seq);
1491 conn->gnc_last_tx = jiffies;
1492 /* no locking here as LIVE isn't a list */
1493 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1495 /* this needs to be checked under lock as it might be freed from a completion
1498 if (msg->gnm_type == GNILND_MSG_NOOP) {
1499 set_mb(conn->gnc_last_noop_sent, jiffies);
1502 /* serialize with seeing CQ events for completion on this, as well as
1504 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1505 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1507 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1508 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1509 kgnilnd_peer_alive(conn->gnc_peer);
1510 GNIDBG_SMSG_CREDS(D_NET, conn);
1513 case GNI_RC_NOT_DONE:
1514 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1515 * 1: out of mbox credits
1516 * 2: out of mbox payload credits
1517 * 3: On Aries out of dla credits
1519 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1520 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1521 /* We'll handle this error inline - makes the calling logic much more
1524 /* If no lock, caller doesn't want us to retry */
1525 if (state_lock == NULL) {
1529 /* I need kgni credits to send this. Replace tx at the head of the
1530 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1531 * and bump retrans counts since we are requeueing the tx.
1535 conn->gnc_tx_retrans++;
1537 kgnilnd_tx_log_retrans(conn, tx);
1538 /* add to head of list for the state and retries */
1539 spin_lock(state_lock);
1540 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1541 spin_unlock(state_lock);
1543 /* We only reschedule for a certain number of retries, then
1544 * we will wait for the CQ events indicating a release of SMSG
1546 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1547 kgnilnd_schedule_conn(conn);
1550 /* CQ event coming in signifies either TX completed or
1551 * RX receive. Either of these *could* free up credits
1552 * in the SMSG mbox and we should try sending again */
1553 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1554 tx->tx_conn->gnc_cqid);
1555 kgnilnd_schedule_delay_conn(conn);
1556 /* use +ve return code to let upper layers know they
1557 * should stop looping on sends */
1561 /* handle bad retcode gracefully */
1562 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1563 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1568 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1570 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1571 spinlock_t *state_lock, kgn_tx_list_state_t state)
1573 kgn_device_t *dev = tx->tx_conn->gnc_device;
1574 unsigned long timestamp;
1577 timestamp = jiffies;
1578 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1579 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1580 /* delay in jiffies - we are really concerned only with things that
1581 * result in a schedule() or really holding this off for long times .
1582 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1583 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1585 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1591 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1593 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1594 spinlock_t *state_lock, kgn_tx_list_state_t state)
1596 kgn_conn_t *conn = tx->tx_conn;
1597 kgn_device_t *dev = conn->gnc_device;
1598 unsigned long timestamp;
1601 timestamp = jiffies;
1603 /* technically we are doing bad things with the read_lock on the peer_conn
1604 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1605 * for the mutex. I bet lockdep is gonna flag this one though... */
1607 /* there are a few cases where we don't want the immediate send - like
1608 * when we are in the scheduler thread and it'd harm the latency of
1609 * getting messages up to LNet */
1611 /* rmb for gnd_ready */
1613 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1615 atomic_inc(&conn->gnc_device->gnd_fast_block);
1616 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1617 /* dont hit HW during quiesce */
1619 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1620 /* dont hit HW if stale EPs and conns left to close */
1623 atomic_inc(&conn->gnc_device->gnd_fast_try);
1624 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1625 &conn->gnc_smsg_mutex);
1630 /* we got the mutex and weren't blocked */
1632 /* delay in jiffies - we are really concerned only with things that
1633 * result in a schedule() or really holding this off for long times .
1634 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1635 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1637 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1638 tx->tx_qtime = jiffies;
1639 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1640 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1641 /* _nolock unlocks the mutex for us */
1647 /* lets us know if we can push this RDMA through now */
1649 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1653 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1655 if (bytes_left < 0) {
1656 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1657 atomic_inc(&dev->gnd_rdmaq_nstalls);
1660 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1661 dev->gnd_rdmaq_deadline);
1662 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1663 /* we never del this timer - at worst it schedules us.. */
1670 /* this adds a TX to the queue pending throttling authorization before
1671 * we allow our remote peer to launch a PUT at us */
1673 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1677 /* we cannot go into send_mapped_tx from here as we are holding locks
1678 * and mem registration might end up allocating memory in kgni.
1679 * That said, we'll push this as far as we can into the queue process */
1680 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1683 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1684 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1685 /* lets us know how delayed RDMA is */
1686 tx->tx_qtime = jiffies;
1687 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1689 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1690 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1691 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1692 /* lets us know how delayed mapping is */
1693 tx->tx_qtime = jiffies;
1694 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1697 /* make sure we wake up sched to run this */
1698 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1701 /* push TX through state machine */
1703 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1708 /* set the tx_id here, we delay it until we have an actual conn
1710 * in some cases, the tx_id is already set to provide for things
1711 * like RDMA completion cookies, etc */
1712 if (tx->tx_id.txe_idx == 0) {
1713 rc = kgnilnd_set_tx_id(tx, conn);
1715 kgnilnd_tx_done(tx, rc);
1720 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1721 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1723 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1725 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1726 kgnilnd_tx_done(tx, rc);
1730 switch (tx->tx_msg.gnm_type) {
1731 case GNILND_MSG_PUT_ACK:
1732 case GNILND_MSG_GET_REQ:
1733 case GNILND_MSG_PUT_REQ_REV:
1734 case GNILND_MSG_GET_ACK_REV:
1735 /* hijacking time! If this messages will authorize our peer to
1736 * send his dirty little bytes in an RDMA, we need to get permission */
1737 kgnilnd_queue_rdma(conn, tx);
1739 case GNILND_MSG_IMMEDIATE:
1740 /* try to send right now, can help reduce latency */
1741 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1744 /* it was sent, break out of switch to avoid default case of queueing */
1747 /* needs to queue to try again, so fall through to default case */
1748 case GNILND_MSG_NOOP:
1749 /* Just make sure this goes out first for this conn */
1751 /* fall through... */
1753 spin_lock(&conn->gnc_list_lock);
1754 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1755 tx->tx_qtime = jiffies;
1756 spin_unlock(&conn->gnc_list_lock);
1757 kgnilnd_schedule_conn(conn);
1762 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1765 kgn_peer_t *new_peer = NULL;
1766 kgn_conn_t *conn = NULL;
1772 /* If I get here, I've committed to send, so I complete the tx with
1773 * failure on any problems */
1775 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1776 "tx already has connection %p", tx->tx_conn);
1778 /* do all of the peer & conn searching in one swoop - this avoids
1779 * nastiness when dropping locks and needing to maintain a sane state
1780 * in the face of stack reset or something else nuking peers & conns */
1782 /* I expect to find him, so only take a read lock */
1783 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1785 peer = kgnilnd_find_peer_locked(target->nid);
1787 conn = kgnilnd_find_conn_locked(peer);
1788 /* this could be NULL during quiesce */
1790 /* Connection exists; queue message on it */
1791 kgnilnd_queue_tx(conn, tx);
1792 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1796 /* don't create a connection if the peer is marked down */
1797 if (peer->gnp_state != GNILND_PEER_UP) {
1798 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1804 /* creating peer or conn; I'll need a write lock... */
1805 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1807 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1809 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1811 /* NB - this will not block during normal operations -
1812 * the only writer of this is in the startup/shutdown path. */
1813 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1819 /* ignore previous peer entirely - we cycled the lock, so we
1820 * will create new peer and at worst drop it if peer is still
1822 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1824 up_read(&kgnilnd_data.kgn_net_rw_sem);
1828 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1829 up_read(&kgnilnd_data.kgn_net_rw_sem);
1831 /* search for peer again now that we have the lock
1832 * if we don't find it, add our new one to the list */
1833 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1835 /* don't create a connection if the peer is not up */
1836 if (peer->gnp_state != GNILND_PEER_UP) {
1837 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1842 conn = kgnilnd_find_or_create_conn_locked(peer);
1844 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1845 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1850 /* oh hey, found a conn now... magical */
1851 kgnilnd_queue_tx(conn, tx);
1853 /* no conn, must be trying to connect - so we queue for now */
1854 tx->tx_qtime = jiffies;
1855 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1857 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1860 kgnilnd_tx_done(tx, rc);
1865 kgnilnd_rdma(kgn_tx_t *tx, int type,
1866 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1868 kgn_conn_t *conn = tx->tx_conn;
1869 unsigned long timestamp;
1870 gni_post_type_t post_type;
1873 unsigned int desc_nob = nob;
1874 void *desc_buffer = tx->tx_buffer;
1875 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1876 LASSERTF(kgnilnd_tx_mapped(tx),
1877 "unmapped tx %p\n", tx);
1878 LASSERTF(conn != NULL,
1879 "NULL conn on tx %p, naughty, naughty\n", tx);
1880 LASSERTF(nob <= sink->gnrd_nob,
1881 "nob %u > sink->gnrd_nob %d (%p)\n",
1882 nob, sink->gnrd_nob, sink);
1883 LASSERTF(nob <= tx->tx_nob,
1884 "nob %d > tx(%p)->tx_nob %d\n",
1885 nob, tx, tx->tx_nob);
1888 case GNILND_MSG_GET_DONE:
1889 case GNILND_MSG_PUT_DONE:
1890 post_type = GNI_POST_RDMA_PUT;
1892 case GNILND_MSG_GET_DONE_REV:
1893 case GNILND_MSG_PUT_DONE_REV:
1894 post_type = GNI_POST_RDMA_GET;
1897 CERROR("invalid msg type %s (%d)\n",
1898 kgnilnd_msgtype2str(type), type);
1901 if (post_type == GNI_POST_RDMA_GET) {
1902 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1903 * aligned. If the local buffer is not aligned correctly using the copy buffer
1904 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1905 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1906 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1907 * making sure the length and addr are aligned and copying the data into the correct location
1908 * once the transfer has completed.
1910 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1911 (sink->gnrd_addr & 3) ||
1914 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1916 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1918 if ((nob + tx->tx_offset) & 3) {
1919 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1920 atomic_inc(&kgnilnd_data.kgn_rev_length);
1922 desc_nob = (nob + tx->tx_offset);
1925 if (tx->tx_buffer_copy == NULL) {
1926 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1927 * and require at most we allocate a few extra bytes. */
1928 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1930 if (!tx->tx_buffer_copy) {
1931 /* allocation of buffer failed nak the rdma */
1932 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1933 kgnilnd_tx_done(tx, -EFAULT);
1936 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1937 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);
1938 if (rc != GNI_RC_SUCCESS) {
1939 /* Registration Failed nak rdma and kill the tx. */
1940 kgnilnd_vfree(tx->tx_buffer_copy,
1942 tx->tx_buffer_copy = NULL;
1943 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1944 kgnilnd_tx_done(tx, -EFAULT);
1948 desc_map_key = tx->tx_buffer_copy_map_key;
1949 desc_buffer = tx->tx_buffer_copy;
1953 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1954 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1955 tx->tx_rdma_desc.type = post_type;
1956 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1957 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1958 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1959 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1960 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1961 tx->tx_rdma_desc.length = desc_nob;
1962 tx->tx_nob_rdma = nob;
1963 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1964 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1965 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1966 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1967 /* prep final completion message */
1968 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1969 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1970 /* send actual size RDMA'd in retval */
1971 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1973 kgnilnd_compute_rdma_cksum(tx, nob);
1976 kgnilnd_queue_tx(conn, tx);
1980 /* Don't lie (CLOSE == RDMA idle) */
1981 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1982 tx, conn, conn->gnc_close_sent);
1984 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1985 "0x%x cookie:%#llx",
1986 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1988 /* set CQ dedicated for RDMA */
1989 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1991 timestamp = jiffies;
1992 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1993 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1994 /* delay in jiffies - we are really concerned only with things that
1995 * result in a schedule() or really holding this off for long times .
1996 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1997 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1999 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2001 if (rrc == GNI_RC_ERROR_RESOURCE) {
2002 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2003 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2004 kgnilnd_unmap_buffer(tx, 0);
2006 if (tx->tx_buffer_copy != NULL) {
2007 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
2008 tx->tx_buffer_copy = NULL;
2011 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2012 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2014 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2015 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2019 spin_lock(&conn->gnc_list_lock);
2020 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2021 tx->tx_qtime = jiffies;
2022 spin_unlock(&conn->gnc_list_lock);
2023 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2024 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2026 /* XXX Nic: is this a place we should handle more errors for
2027 * robustness sake */
2028 LASSERT(rrc == GNI_RC_SUCCESS);
2033 kgnilnd_alloc_rx(void)
2037 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2039 CERROR("failed to allocate rx\n");
2042 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2045 /* no memset to zero, we'll always fill all members */
2049 /* release is to just free connection resources
2050 * we use this for the eager path after copying */
2052 kgnilnd_release_msg(kgn_conn_t *conn)
2055 unsigned long timestamp;
2057 CDEBUG(D_NET, "consuming %p\n", conn);
2059 timestamp = jiffies;
2060 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2061 /* delay in jiffies - we are really concerned only with things that
2062 * result in a schedule() or really holding this off for long times .
2063 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2064 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2066 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2067 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2069 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2070 GNIDBG_SMSG_CREDS(D_NET, conn);
2072 kgnilnd_schedule_conn(conn);
2076 kgnilnd_consume_rx(kgn_rx_t *rx)
2078 kgn_conn_t *conn = rx->grx_conn;
2079 kgn_msg_t *rxmsg = rx->grx_msg;
2081 /* if we are eager, free the cache alloc'd msg */
2082 if (unlikely(rx->grx_eager)) {
2083 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2084 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2086 /* release ref from eager_recv */
2087 kgnilnd_conn_decref(conn);
2089 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2090 kgnilnd_release_msg(conn);
2093 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2094 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2099 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2101 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2102 int type = lntmsg->msg_type;
2103 struct lnet_process_id target = lntmsg->msg_target;
2104 int target_is_router = lntmsg->msg_target_is_router;
2105 int routing = lntmsg->msg_routing;
2106 unsigned int niov = lntmsg->msg_niov;
2107 struct kvec *iov = lntmsg->msg_iov;
2108 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2109 unsigned int offset = lntmsg->msg_offset;
2110 unsigned int nob = lntmsg->msg_len;
2111 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2112 kgn_net_t *net = ni->ni_data;
2116 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2118 /* NB 'private' is different depending on what we're sending.... */
2119 LASSERT(!in_interrupt());
2121 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2122 type, nob, niov, libcfs_id2str(target));
2124 LASSERTF(nob == 0 || niov > 0,
2125 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2126 LASSERTF(niov <= LNET_MAX_IOV,
2127 "lntmsg %p niov %d\n", lntmsg, niov);
2129 /* payload is either all vaddrs or all pages */
2130 LASSERTF(!(kiov != NULL && iov != NULL),
2131 "lntmsg %p kiov %p iov %p\n", lntmsg, kiov, iov);
2134 mpflag = cfs_memory_pressure_get_and_set();
2138 CERROR("lntmsg %p with unexpected type %d\n",
2143 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2151 if (routing || target_is_router)
2152 break; /* send IMMEDIATE */
2154 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2155 * check the eventual sink buffer here - if small enough, remote
2156 * end is perfectly capable of returning data in short message -
2157 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2158 * for IMMEDIATE messages which will have it send a real reply instead
2159 * of doing kgnilnd_recv to have the RDMA continued */
2160 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2163 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2164 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2166 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2172 /* slightly different options as we might actually have a GET with a
2173 * MD_KIOV set but a non-NULL md_iov.iov */
2174 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
2175 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2176 lntmsg->msg_md->md_iov.iov, NULL,
2177 0, lntmsg->msg_md->md_length);
2179 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2180 NULL, lntmsg->msg_md->md_iov.kiov,
2181 0, lntmsg->msg_md->md_length);
2183 CERROR("unable to setup buffer: %d\n", rc);
2184 kgnilnd_tx_done(tx, rc);
2189 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2190 if (tx->tx_lntmsg[1] == NULL) {
2191 CERROR("Can't create reply for GET to %s\n",
2192 libcfs_nid2str(target.nid));
2193 kgnilnd_tx_done(tx, rc);
2198 tx->tx_lntmsg[0] = lntmsg;
2199 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2200 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2202 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2204 /* rest of tx_msg is setup just before it is sent */
2205 kgnilnd_launch_tx(tx, net, &target);
2207 case LNET_MSG_REPLY:
2209 /* to save on MDDs, we'll handle short kiov by vmap'ing
2210 * and sending via SMSG */
2211 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2214 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2215 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2217 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2224 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2226 kgnilnd_tx_done(tx, rc);
2231 tx->tx_lntmsg[0] = lntmsg;
2232 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2233 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2235 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2237 /* rest of tx_msg is setup just before it is sent */
2238 kgnilnd_launch_tx(tx, net, &target);
2242 /* send IMMEDIATE */
2244 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2245 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2247 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2253 rc = kgnilnd_setup_immediate_buffer(tx, niov, iov, kiov, offset, nob);
2255 kgnilnd_tx_done(tx, rc);
2259 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2260 tx->tx_lntmsg[0] = lntmsg;
2261 kgnilnd_launch_tx(tx, net, &target);
2264 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2266 cfs_memory_pressure_restore(mpflag);
2271 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2273 kgn_conn_t *conn = rx->grx_conn;
2274 kgn_msg_t *rxmsg = rx->grx_msg;
2275 unsigned int niov = lntmsg->msg_niov;
2276 struct kvec *iov = lntmsg->msg_iov;
2277 lnet_kiov_t *kiov = lntmsg->msg_kiov;
2278 unsigned int offset = lntmsg->msg_offset;
2279 unsigned int nob = lntmsg->msg_len;
2284 switch (rxmsg->gnm_type) {
2285 case GNILND_MSG_PUT_REQ_REV:
2286 done_type = GNILND_MSG_PUT_DONE_REV;
2289 case GNILND_MSG_GET_REQ:
2290 done_type = GNILND_MSG_GET_DONE;
2293 CERROR("invalid msg type %s (%d)\n",
2294 kgnilnd_msgtype2str(rxmsg->gnm_type),
2299 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2303 rc = kgnilnd_set_tx_id(tx, conn);
2307 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
2311 tx->tx_lntmsg[0] = lntmsg;
2312 tx->tx_getinfo = rxmsg->gnm_u.get;
2314 /* we only queue from kgnilnd_recv - we might get called from other contexts
2315 * and we don't want to block the mutex in those cases */
2317 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2318 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2319 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2320 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2325 kgnilnd_tx_done(tx, rc);
2326 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2328 lnet_finalize(lntmsg, rc);
2332 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2335 kgn_rx_t *rx = private;
2336 kgn_conn_t *conn = rx->grx_conn;
2337 kgn_msg_t *rxmsg = rx->grx_msg;
2338 kgn_msg_t *eagermsg = NULL;
2339 kgn_peer_t *peer = NULL;
2340 kgn_conn_t *found_conn = NULL;
2342 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2343 conn, rxmsg, lntmsg);
2345 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2346 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2347 rxmsg->gnm_payload_len);
2350 /* Grab a read lock so the connection doesnt disappear on us
2351 * while we look it up
2353 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2355 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2357 found_conn = kgnilnd_find_conn_locked(peer);
2360 /* Verify the connection found is the same one that the message
2361 * is supposed to be using, if it is not output an error message
2364 if (!peer || !found_conn
2365 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2366 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2367 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2368 peer, conn, found_conn);
2370 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2371 " from %s", rxmsg->gnm_connstamp,
2372 found_conn->gnc_peer_connstamp,
2373 libcfs_nid2str(peer->gnp_nid));
2378 /* add conn ref to ensure it doesn't go away until all eager
2379 * messages processed */
2380 kgnilnd_conn_addref(conn);
2382 /* Now that we have verified the connection is valid and added a
2383 * reference we can remove the read_lock on the peer_conn_lock */
2384 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2386 /* we have no credits or buffers for this message, so copy it
2387 * somewhere for a later kgnilnd_recv */
2388 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2389 *kgnilnd_tunables.kgn_eager_credits) {
2390 CERROR("Out of eager credits to %s\n",
2391 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2395 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2397 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2398 if (eagermsg == NULL) {
2399 kgnilnd_conn_decref(conn);
2400 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2401 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2405 /* copy msg and payload */
2406 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2407 rx->grx_msg = eagermsg;
2410 /* stash this for lnet_finalize on cancel-on-conn-close */
2411 rx->grx_lntmsg = lntmsg;
2413 /* keep the same rx_t, it just has a new grx_msg now */
2414 *new_private = private;
2416 /* release SMSG buffer */
2417 kgnilnd_release_msg(conn);
2423 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2424 int delayed, unsigned int niov,
2425 struct kvec *iov, lnet_kiov_t *kiov,
2426 unsigned int offset, unsigned int mlen, unsigned int rlen)
2428 kgn_rx_t *rx = private;
2429 kgn_conn_t *conn = rx->grx_conn;
2430 kgn_msg_t *rxmsg = rx->grx_msg;
2436 LASSERT(!in_interrupt());
2437 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2438 /* Either all pages or all vaddrs */
2439 LASSERTF(!(kiov != NULL && iov != NULL), "kiov %p iov %p\n",
2442 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2443 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2444 conn, rxmsg, lntmsg,
2445 niov, kiov, iov, offset, mlen, rlen);
2447 /* we need to lock here as recv can be called from any context */
2448 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2449 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2450 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2452 /* someone closed the conn after we copied this out, nuke it */
2453 kgnilnd_consume_rx(rx);
2454 lnet_finalize(lntmsg, conn->gnc_error);
2457 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2459 switch (rxmsg->gnm_type) {
2461 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2462 " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
2463 conn, rx, rxmsg, lntmsg, niov, kiov, iov, offset, mlen, rlen);
2466 case GNILND_MSG_IMMEDIATE:
2467 if (mlen > rxmsg->gnm_payload_len) {
2468 GNIDBG_MSG(D_ERROR, rxmsg,
2469 "Immediate message from %s too big: %d > %d",
2470 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2471 rxmsg->gnm_payload_len);
2473 kgnilnd_consume_rx(rx);
2477 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2478 * right after the kgn_msg_t header - so just 'cute' way of saying
2479 * rxmsg + sizeof(kgn_msg_t) */
2481 /* check payload checksum if sent */
2483 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2484 !rxmsg->gnm_payload_cksum &&
2485 rxmsg->gnm_payload_len != 0)
2486 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2488 if (rxmsg->gnm_payload_cksum != 0) {
2489 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2490 * which is what is used to calculate the cksum on the TX side */
2491 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2493 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2494 GNIDBG_MSG(D_NETERROR, rxmsg,
2495 "Bad payload checksum (%x expected %x)",
2496 pload_cksum, rxmsg->gnm_payload_cksum);
2497 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2499 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2500 &rxmsg[1], rxmsg->gnm_payload_len);
2501 /* fall through to dump */
2503 libcfs_debug_dumplog();
2509 /* checksum problems are fatal, kill the conn */
2510 kgnilnd_consume_rx(rx);
2511 kgnilnd_close_conn(conn, rc);
2517 lnet_copy_flat2kiov(
2519 *kgnilnd_tunables.kgn_max_immediate,
2520 &rxmsg[1], 0, mlen);
2524 *kgnilnd_tunables.kgn_max_immediate,
2525 &rxmsg[1], 0, mlen);
2527 kgnilnd_consume_rx(rx);
2528 lnet_finalize(lntmsg, 0);
2531 case GNILND_MSG_PUT_REQ:
2532 /* LNET wants to truncate or drop transaction, sending NAK */
2534 kgnilnd_consume_rx(rx);
2535 lnet_finalize(lntmsg, 0);
2537 /* only error if lntmsg == NULL, otherwise we are just
2538 * short circuiting the rdma process of 0 bytes */
2539 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2540 lntmsg == NULL ? -ENOENT : 0,
2541 rxmsg->gnm_u.get.gngm_cookie,
2545 /* sending ACK with sink buff. info */
2546 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2548 kgnilnd_consume_rx(rx);
2552 rc = kgnilnd_set_tx_id(tx, conn);
2554 GOTO(nak_put_req, rc);
2557 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2559 GOTO(nak_put_req, rc);
2562 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2563 rxmsg->gnm_u.putreq.gnprm_cookie;
2564 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2565 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2566 (__u64)((unsigned long)tx->tx_buffer);
2567 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2569 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2570 tx->tx_qtime = jiffies;
2571 /* we only queue from kgnilnd_recv - we might get called from other contexts
2572 * and we don't want to block the mutex in those cases */
2574 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2575 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2576 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2577 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2579 kgnilnd_consume_rx(rx);
2583 /* make sure we send an error back when the PUT fails */
2584 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2585 kgnilnd_tx_done(tx, rc);
2586 kgnilnd_consume_rx(rx);
2588 /* return magic LNet network error */
2590 case GNILND_MSG_GET_REQ_REV:
2591 /* LNET wants to truncate or drop transaction, sending NAK */
2593 kgnilnd_consume_rx(rx);
2594 lnet_finalize(lntmsg, 0);
2596 /* only error if lntmsg == NULL, otherwise we are just
2597 * short circuiting the rdma process of 0 bytes */
2598 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2599 lntmsg == NULL ? -ENOENT : 0,
2600 rxmsg->gnm_u.get.gngm_cookie,
2604 /* lntmsg can be null when parsing a LNET_GET */
2605 if (lntmsg != NULL) {
2606 /* sending ACK with sink buff. info */
2607 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2609 kgnilnd_consume_rx(rx);
2613 rc = kgnilnd_set_tx_id(tx, conn);
2615 GOTO(nak_get_req_rev, rc);
2618 rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
2620 GOTO(nak_get_req_rev, rc);
2623 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2624 rxmsg->gnm_u.putreq.gnprm_cookie;
2625 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2626 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2627 (__u64)((unsigned long)tx->tx_buffer);
2628 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2630 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2632 /* we only queue from kgnilnd_recv - we might get called from other contexts
2633 * and we don't want to block the mutex in those cases */
2635 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2636 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2637 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2638 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2641 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2643 rxmsg->gnm_u.get.gngm_cookie,
2647 kgnilnd_consume_rx(rx);
2651 /* make sure we send an error back when the GET fails */
2652 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2653 kgnilnd_tx_done(tx, rc);
2654 kgnilnd_consume_rx(rx);
2656 /* return magic LNet network error */
2660 case GNILND_MSG_PUT_REQ_REV:
2661 /* LNET wants to truncate or drop transaction, sending NAK */
2663 kgnilnd_consume_rx(rx);
2664 lnet_finalize(lntmsg, 0);
2666 /* only error if lntmsg == NULL, otherwise we are just
2667 * short circuiting the rdma process of 0 bytes */
2668 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2669 lntmsg == NULL ? -ENOENT : 0,
2670 rxmsg->gnm_u.get.gngm_cookie,
2675 if (lntmsg != NULL) {
2677 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2680 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2682 rxmsg->gnm_u.get.gngm_cookie,
2685 kgnilnd_consume_rx(rx);
2687 case GNILND_MSG_GET_REQ:
2688 if (lntmsg != NULL) {
2690 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2693 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2695 rxmsg->gnm_u.get.gngm_cookie,
2698 kgnilnd_consume_rx(rx);
2704 /* needs write_lock on kgn_peer_conn_lock held */
2706 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2708 unsigned long timeout, keepalive;
2709 unsigned long now = jiffies;
2710 unsigned long newest_last_rx;
2713 /* given that we found this conn hanging off a peer, it better damned
2714 * well be connected */
2715 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2716 "conn 0x%p->%s with bad state%s\n", conn,
2717 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2719 kgnilnd_conn_state2str(conn));
2721 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2722 "rx_diff %lu tx_diff %lu\n",
2723 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2724 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2725 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2726 cfs_duration_sec(now - conn->gnc_last_tx));
2728 timeout = cfs_time_seconds(conn->gnc_timeout);
2729 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2731 /* just in case our lack of RX msg processing is gumming up the works - give the
2732 * remove an extra chance */
2734 newest_last_rx = GNILND_LASTRX(conn);
2736 if (time_after_eq(now, newest_last_rx + timeout)) {
2737 uint32_t level = D_CONSOLE|D_NETERROR;
2739 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2742 GNIDBG_CONN(level, conn,
2743 "No gnilnd traffic received from %s for %lu "
2744 "seconds, terminating connection. Is node down? ",
2745 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2746 cfs_duration_sec(now - newest_last_rx));
2750 /* we don't timeout on last_tx stalls - we are going to trust the
2751 * underlying network to let us know when sends are failing.
2752 * At worst, the peer will timeout our RX stamp and drop the connection
2753 * at that point. We'll then see his CLOSE or at worst his RX
2754 * stamp stop and drop the connection on our end */
2756 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2757 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2758 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2759 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2760 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2762 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2763 conn->gnc_last_noop_cq,
2764 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2765 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2766 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2767 set_mb(conn->gnc_last_noop_want, jiffies);
2768 atomic_inc(&conn->gnc_reaper_noop);
2769 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2772 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2775 kgnilnd_queue_tx(conn, tx);
2781 /* needs write_lock on kgn_peer_conn_lock held */
2783 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2784 struct list_head *souls)
2786 unsigned long timeout;
2787 kgn_conn_t *conn, *connN = NULL;
2793 short releaseconn = 0;
2794 unsigned long first_rx = 0;
2795 int purgatory_conn_cnt = 0;
2797 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2798 peer, libcfs_nid2str(peer->gnp_nid),
2799 peer->gnp_reconnect_interval);
2801 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2802 GNILND_MIN_TIMEOUT));
2804 conn = kgnilnd_find_conn_locked(peer);
2806 /* if there is a valid conn, check the queues for timeouts */
2807 rc = kgnilnd_check_conn_timeouts_locked(conn);
2809 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2810 /* simulate a RX CLOSE after the timeout but before
2811 * the scheduler thread gets it */
2812 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2813 conn->gnc_peer_error = -ETIMEDOUT;
2816 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2818 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2819 CDEBUG(D_WARNING, "%s conn timed out, will "
2820 "reconnect upon request from peer\n",
2821 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2823 /* Once we mark closed, any of the scheduler threads could
2824 * get it and move through before we hit the fail loc code */
2825 kgnilnd_close_conn_locked(conn, rc);
2827 /* first_rx is used to decide when to release a conn from purgatory.
2829 first_rx = conn->gnc_first_rx;
2833 /* now regardless of starting new conn, find tx on peer queue that
2834 * are old and smell bad - do this first so we don't trigger
2835 * reconnect on empty queue if we timeout all */
2836 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2837 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2839 LCONSOLE_INFO("could not send to %s due to connection"
2840 " setup failure after %lu seconds\n",
2841 libcfs_nid2str(peer->gnp_nid),
2842 cfs_duration_sec(jiffies - tx->tx_qtime));
2844 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2846 list_add_tail(&tx->tx_list, todie);
2851 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2852 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2853 count, peer, libcfs_nid2str(peer->gnp_nid));
2854 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2855 if (list_empty(&peer->gnp_tx_queue) ||
2856 peer->gnp_connecting == GNILND_PEER_KILL) {
2857 /* we pass down todie to use a common function - but we know there are
2859 kgnilnd_cancel_peer_connect_locked(peer, todie);
2863 /* Don't reconnect if we are still trying to clear out old conns.
2864 * This prevents us sending traffic on the new mbox before ensuring we are done
2865 * with the old one */
2866 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2867 (atomic_read(&peer->gnp_dirty_eps) == 0);
2869 /* fast reconnect after a timeout */
2870 to_reconn = !conn &&
2871 (peer->gnp_last_errno == -ETIMEDOUT) &&
2872 *kgnilnd_tunables.kgn_fast_reconn;
2874 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2875 * to be sent, we'll check the reconnect interval and fire up a new
2876 * connection request */
2879 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2880 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2881 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2883 CDEBUG(D_NET, "starting connect to %s\n",
2884 libcfs_nid2str(peer->gnp_nid));
2885 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2886 "Peer was idle and we have a write_lock, state issue %d\n",
2887 peer->gnp_connecting);
2889 peer->gnp_connecting = GNILND_PEER_CONNECT;
2890 kgnilnd_peer_addref(peer); /* extra ref for connd */
2892 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2893 list_add_tail(&peer->gnp_connd_list,
2894 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2895 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2897 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2900 /* fail_loc to allow us to delay release of purgatory */
2901 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2904 /* This check allows us to verify that the new conn is actually being used. This allows us to
2905 * pull the old conns out of purgatory if they have actually seen traffic.
2906 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2909 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2910 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2911 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2915 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2916 /* check for purgatory timeouts */
2917 if (conn->gnc_in_purgatory) {
2918 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2919 * that way the next time we check it we can detach it from purgatory
2922 if (conn->gnc_state != GNILND_CONN_DONE) {
2923 /* Skip over conns that are currently not DONE. If they arent already scheduled
2924 * for completion something in the state machine is broken.
2929 /* We only detach a conn that is in purgatory if we have received a close message,
2930 * we have a new valid connection that has successfully received data, or an admin
2931 * command tells us we need to detach.
2934 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2935 unsigned long waiting;
2937 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2939 /* C.E: The remote peer is expected to close the
2940 * connection (see kgnilnd_check_conn_timeouts)
2941 * via the reaper thread and nuke out the MDD and
2942 * FMA resources after conn->gnc_timeout has expired
2943 * without an FMA RX */
2944 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2945 " held resources\n",
2946 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2947 cfs_duration_sec(waiting));
2949 kgnilnd_detach_purgatory_locked(conn, souls);
2951 purgatory_conn_cnt++;
2956 /* If we have too many connections in purgatory we could run out of
2957 * resources. Limit the number of connections to a tunable number,
2958 * clean up to the minimum all in one fell swoop... there are
2959 * situations where dvs will retry tx's and we can eat up several
2960 * hundread connection requests at once.
2962 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2963 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2965 if (conn->gnc_in_purgatory &&
2966 conn->gnc_state == GNILND_CONN_DONE) {
2967 CDEBUG(D_NET, "Dropping Held resource due to"
2968 " resource limits being hit\n");
2969 kgnilnd_detach_purgatory_locked(conn, souls);
2971 if (purgatory_conn_cnt-- <
2972 *kgnilnd_tunables.kgn_max_purgatory)
2980 kgnilnd_reaper_check(int idx)
2982 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2983 struct list_head *ctmp, *ctmpN;
2984 LIST_HEAD(geriatrics);
2987 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2989 list_for_each_safe(ctmp, ctmpN, peers) {
2990 kgn_peer_t *peer = NULL;
2992 /* don't timeout stuff if the network is mucked or shutting down */
2993 if (kgnilnd_check_hw_quiesce()) {
2996 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2998 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
3001 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3003 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3004 kgnilnd_release_purgatory_list(&souls);
3008 kgnilnd_update_reaper_timeout(long timeout)
3010 LASSERT(timeout > 0);
3012 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3014 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3015 kgnilnd_data.kgn_new_min_timeout = timeout;
3017 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3021 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3023 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3027 kgnilnd_reaper(void *arg)
3032 unsigned long next_check_time = jiffies;
3033 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3034 struct timer_list timer;
3037 /* all gnilnd threads need to run fairly urgently */
3038 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3039 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3041 while (!kgnilnd_data.kgn_shutdown) {
3042 /* I wake up every 'p' seconds to check for timeouts on some
3043 * more peers. I try to check every connection 'n' times
3044 * within the global minimum of all keepalive and timeout
3045 * intervals, to ensure I attend to every connection within
3046 * (n+1)/n times its timeout intervals. */
3047 const int p = GNILND_REAPER_THREAD_WAKE;
3048 const int n = GNILND_REAPER_NCHECKS;
3050 /* to quiesce or to not quiesce, that is the question */
3051 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3052 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3053 KGNILND_SPIN_QUIESCE;
3054 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3057 /* careful with the jiffy wrap... */
3058 timeout = (long)(next_check_time - jiffies);
3061 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3062 TASK_INTERRUPTIBLE);
3063 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3064 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3066 mod_timer(&timer, (long) jiffies + timeout);
3068 /* check flag variables before committing */
3069 if (!kgnilnd_data.kgn_shutdown &&
3070 !kgnilnd_data.kgn_quiesce_trigger) {
3071 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3072 timeout, cfs_duration_sec(timeout));
3074 CDEBUG(D_INFO, "awake after schedule\n");
3077 del_singleshot_timer_sync(&timer);
3078 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3079 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3083 /* new_min_timeout is set from the conn timeouts and keepalive
3084 * this should end up with a min timeout of
3085 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3086 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3087 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3088 CDEBUG(D_NET, "Set new min timeout %ld\n",
3089 current_min_timeout);
3092 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3094 /* Compute how many table entries to check now so I get round
3095 * the whole table fast enough given that I do this at fixed
3096 * intervals of 'p' seconds) */
3097 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3098 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3099 chunk = (chunk * n * p) /
3100 kgnilnd_data.kgn_new_min_timeout;
3103 for (i = 0; i < chunk; i++) {
3104 kgnilnd_reaper_check(hash_index);
3105 hash_index = (hash_index + 1) %
3106 *kgnilnd_tunables.kgn_peer_hash_size;
3108 next_check_time = (long) jiffies + cfs_time_seconds(p);
3109 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3111 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3114 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3116 kgnilnd_thread_fini();
3121 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3122 unsigned niov, offset, nob;
3124 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3125 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3128 lnet_copy_flat2kiov(
3131 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3133 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3140 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3143 gni_post_descriptor_t *desc;
3145 kgn_tx_ev_id_t ev_id;
3147 int should_retry, rc;
3148 long num_processed = 0;
3149 kgn_conn_t *conn = NULL;
3150 kgn_tx_t *tx = NULL;
3151 kgn_rdma_desc_t *rdesc;
3156 /* make sure we don't keep looping if we need to reset */
3157 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3158 return num_processed;
3160 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3162 /* we didn't get the mutex, so return that there is still work
3166 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3167 /* a bit gross - but we need a good way to test for
3168 * delayed RDMA completions and the easiest way to do
3169 * that is to delay the RDMA CQ events */
3170 rrc = GNI_RC_NOT_DONE;
3172 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3175 if (rrc == GNI_RC_NOT_DONE) {
3176 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3177 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3178 dev->gnd_id, num_processed);
3179 return num_processed;
3181 dev->gnd_sched_alive = jiffies;
3184 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3185 "this is bad, somehow our credits didn't protect us"
3186 " from CQ overrun\n");
3187 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3188 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3189 event_data, GNI_CQ_GET_TYPE(event_data));
3191 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3193 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3195 /* XXX Nic: Need better error handling here... */
3196 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3197 (rrc == GNI_RC_TRANSACTION_ERROR),
3200 ev_id.txe_cookie = desc->post_id;
3202 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3204 if (conn == NULL || tx == NULL) {
3205 /* either conn or tx was already nuked and this is a "late"
3206 * completion, so drop it */
3210 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3211 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3212 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3213 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3214 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3216 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3218 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3219 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3220 tx->tx_msg.gnm_u.completion.gncm_retval);
3224 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3225 if (tx->tx_buffer_copy != NULL)
3226 kgnilnd_recv_bte_get(tx);
3227 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3230 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3231 if (tx->tx_buffer_copy != NULL)
3232 kgnilnd_recv_bte_get(tx);
3233 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3236 /* remove from rdmaq */
3237 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3238 spin_lock(&conn->gnc_list_lock);
3239 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3240 spin_unlock(&conn->gnc_list_lock);
3241 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3243 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3244 event_data = 1LL << 48;
3248 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3249 atomic_inc(&dev->gnd_rdma_ntx);
3250 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3251 /* transaction succeeded, add into fmaq */
3252 kgnilnd_queue_tx(conn, tx);
3253 kgnilnd_peer_alive(conn->gnc_peer);
3255 /* drop ref from kgnilnd_validate_tx_ev_id */
3256 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3257 kgnilnd_conn_decref(conn);
3262 /* fall through to the TRANSACTION_ERROR case */
3265 /* get stringified version for log messages */
3266 kgnilnd_cq_error_str(event_data, &err_str, 256);
3267 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3269 /* make sure we are not off in the weeds with this tx */
3270 if (tx->tx_retrans >
3271 *kgnilnd_tunables.kgn_max_retransmits) {
3272 GNIDBG_TX(D_NETERROR, tx,
3273 "giving up on TX, too many retries", NULL);
3277 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3278 should_retry ? "transient" : "unrecoverable", err_str);
3280 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3281 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3282 rdesc = &tx->tx_putinfo.gnpam_desc;
3283 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3284 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3286 rdesc = &tx->tx_getinfo.gngm_desc;
3287 rnob = tx->tx_lntmsg[0]->msg_len;
3288 rcookie = tx->tx_getinfo.gngm_cookie;
3293 tx->tx_msg.gnm_type,
3297 kgnilnd_nak_rdma(conn,
3298 tx->tx_msg.gnm_type,
3301 tx->tx_msg.gnm_srcnid);
3302 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3303 kgnilnd_close_conn(conn, -ECOMM);
3306 /* drop ref from kgnilnd_validate_tx_ev_id */
3307 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3308 kgnilnd_conn_decref(conn);
3313 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3317 kgn_tx_ev_id_t ev_id;
3318 kgn_tx_t *tx = NULL;
3319 kgn_conn_t *conn = NULL;
3320 int queued_fma, saw_reply, rc;
3321 long num_processed = 0;
3322 struct list_head *ctmp, *ctmpN;
3325 /* make sure we don't keep looping if we need to reset */
3326 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3327 return num_processed;
3330 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3332 /* we didn't get the mutex, so return that there is still work
3337 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3338 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3340 if (rrc == GNI_RC_NOT_DONE) {
3342 "SMSG send CQ %d not ready (data %#llx) "
3343 "processed %ld\n", dev->gnd_id, event_data,
3346 if (num_processed > 0) {
3347 spin_lock(&dev->gnd_lock);
3348 if (!list_empty(&dev->gnd_delay_conns)) {
3349 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3350 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3351 list_del_init(&conn->gnc_delaylist);
3352 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3353 kgnilnd_schedule_conn_nolock(conn);
3355 spin_unlock(&dev->gnd_lock);
3356 kgnilnd_schedule_device(dev);
3358 spin_unlock(&dev->gnd_lock);
3361 return num_processed;
3364 dev->gnd_sched_alive = jiffies;
3367 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3368 "this is bad, somehow our credits didn't "
3369 "protect us from CQ overrun\n");
3370 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3371 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3372 event_data, GNI_CQ_GET_TYPE(event_data));
3374 /* if SMSG couldn't handle an error, time for conn to die */
3375 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3378 /* need to take the write_lock to ensure atomicity
3379 * on the conn state if we need to close it */
3380 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3381 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3383 /* Conn was destroyed? */
3385 "SMSG CQID lookup %#llx failed\n",
3386 GNI_CQ_GET_INST_ID(event_data));
3387 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3391 kgnilnd_cq_error_str(event_data, &err_str, 256);
3392 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3393 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3395 kgnilnd_close_conn_locked(conn, -ECOMM);
3397 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3399 /* no need to process rest of this tx -
3400 * it is getting canceled */
3404 /* fall through to GNI_RC_SUCCESS case */
3405 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3407 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3408 if (conn == NULL || tx == NULL) {
3409 /* either conn or tx was already nuked and this is a "late"
3410 * completion, so drop it */
3414 tx->tx_conn->gnc_last_tx_cq = jiffies;
3415 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3416 set_mb(conn->gnc_last_noop_cq, jiffies);
3419 /* lock tx_list_state and tx_state */
3420 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3421 spin_lock(&tx->tx_conn->gnc_list_lock);
3423 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3424 "state not GNILND_TX_LIVE_FMAQ", NULL);
3425 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3426 "not waiting for completion", NULL);
3428 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3431 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3433 /* This will trigger other FMA sends that were
3434 * pending this completion */
3435 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3437 /* we either did not expect reply or we already got it */
3438 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3440 spin_unlock(&tx->tx_conn->gnc_list_lock);
3441 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3444 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3446 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3447 kgnilnd_schedule_conn(conn);
3450 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3451 * If saw_reply is true we know that the tx is safe to use as the other thread
3452 * is already finished with it.
3456 /* no longer need to track on the live_fmaq */
3457 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3459 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3460 /* we already got reply & were waiting for
3461 * completion of initial send */
3462 /* to initiate RDMA transaction */
3463 GNIDBG_TX(D_NET, tx,
3464 "Pending RDMA 0x%p type 0x%02x",
3465 tx->tx_msg.gnm_type);
3466 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3467 rc = kgnilnd_send_mapped_tx(tx, 0);
3468 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3470 /* we are done with this tx */
3471 GNIDBG_TX(D_NET, tx,
3472 "Done with tx type 0x%02x",
3473 tx->tx_msg.gnm_type);
3474 kgnilnd_tx_done(tx, tx->tx_rc);
3478 /* drop ref from kgnilnd_validate_tx_ev_id */
3479 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3480 kgnilnd_conn_decref(conn);
3482 /* if we are waiting for a REPLY, we'll handle the tx then */
3483 } /* end for loop */
3487 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3492 long num_processed = 0;
3493 struct list_head *conns;
3494 struct list_head *tmp;
3498 /* make sure we don't keep looping if we need to reset */
3499 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3500 return num_processed;
3503 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3505 /* we didn't get the mutex, so return that there is still work
3509 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3510 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3512 if (rrc == GNI_RC_NOT_DONE) {
3513 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3515 dev->gnd_id, event_data, num_processed);
3516 return num_processed;
3518 dev->gnd_sched_alive = jiffies;
3521 /* this is the only CQ that can really handle transient
3523 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3524 rrc = cfs_fail_val ? cfs_fail_val
3525 : GNI_RC_ERROR_RESOURCE;
3526 if (rrc == GNI_RC_ERROR_RESOURCE) {
3527 /* set overrun too */
3528 event_data |= (1UL << 63);
3529 LASSERTF(GNI_CQ_OVERRUN(event_data),
3530 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3533 /* sender should get error event too and take care
3534 of failed transaction by re-transmitting */
3535 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3536 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3540 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3541 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3542 conn = kgnilnd_cqid2conn_locked(
3543 GNI_CQ_GET_INST_ID(event_data));
3545 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3546 "failed, dropping event %#llx\n",
3547 GNI_CQ_GET_INST_ID(event_data),
3550 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3552 GNI_CQ_GET_INST_ID(event_data),
3553 conn, conn->gnc_peer ?
3554 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3557 conn->gnc_last_rx_cq = jiffies;
3559 /* stash first rx so we can clear out purgatory.
3561 if (conn->gnc_first_rx == 0) {
3562 conn->gnc_first_rx = jiffies;
3564 kgnilnd_peer_alive(conn->gnc_peer);
3565 kgnilnd_schedule_conn(conn);
3567 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3571 /* FMA CQ has overflowed: check ALL conns */
3572 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3573 "conns on device %d\n", dev->gnd_id);
3575 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3577 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3578 conns = &kgnilnd_data.kgn_conns[rc];
3580 list_for_each(tmp, conns) {
3581 conn = list_entry(tmp, kgn_conn_t,
3584 if (conn->gnc_device == dev) {
3585 kgnilnd_schedule_conn(conn);
3586 conn->gnc_last_rx_cq = jiffies;
3590 /* don't block write lockers for too long... */
3591 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3596 /* try_map_if_full should only be used when processing TX from list of
3597 * backlog TX waiting on mappings to free up
3600 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3601 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3604 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3606 /* slight bit of race if multiple people calling, but at worst we'll have
3607 * order altered just a bit... which would not be determenistic anyways */
3608 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3610 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3612 /* We know that we have a GART reservation that should guarantee forward progress.
3613 * This means we don't need to take any extraordinary efforts if we are failing
3614 * mappings here - even if we are holding a very small number of these. */
3616 if (try_map_if_full || (rc == 0)) {
3617 rc = kgnilnd_map_buffer(tx);
3620 /* rc should be 0 if we mapped successfully here, if non-zero
3621 * we are queueing */
3623 /* if try_map_if_full set, they handle requeuing */
3624 if (unlikely(try_map_if_full)) {
3627 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3628 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3629 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3630 /* make sure we wake up sched to run this */
3631 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3632 /* return 0 as this is now queued for later sending */
3637 switch (tx->tx_msg.gnm_type) {
3641 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3642 * remote node where the RDMA will be started
3643 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3644 * be satisified. The rest of the errors are "hard" errors that require
3645 * upper layers to handle themselves.
3646 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3647 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3648 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3649 * GNILND_MSG_GET_DONE_REV.
3651 case GNILND_MSG_GET_REQ:
3652 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3653 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3654 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3655 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3656 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3657 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3658 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3660 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3661 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3663 case GNILND_MSG_PUT_ACK:
3664 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3665 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3666 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3667 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3669 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3670 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3673 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3674 case GNILND_MSG_PUT_DONE:
3675 case GNILND_MSG_PUT_REQ:
3676 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3677 &tx->tx_putinfo.gnpam_desc,
3678 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3679 tx->tx_putinfo.gnpam_dst_cookie);
3680 RETURN(try_map_if_full ? rc : 0);
3682 case GNILND_MSG_GET_DONE:
3683 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3684 &tx->tx_getinfo.gngm_desc,
3685 tx->tx_lntmsg[0]->msg_len,
3686 tx->tx_getinfo.gngm_cookie);
3687 RETURN(try_map_if_full ? rc : 0);
3689 case GNILND_MSG_PUT_REQ_REV:
3690 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3691 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3692 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3693 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3694 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3695 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3696 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3698 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3700 case GNILND_MSG_PUT_DONE_REV:
3701 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3702 &tx->tx_getinfo.gngm_desc,
3704 tx->tx_getinfo.gngm_cookie);
3705 RETURN(try_map_if_full ? rc : 0);
3707 case GNILND_MSG_GET_ACK_REV:
3708 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3709 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3710 /* LNET_GETS are a special case for parse */
3711 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3712 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3714 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3715 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3717 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3718 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3720 case GNILND_MSG_GET_DONE_REV:
3721 case GNILND_MSG_GET_REQ_REV:
3722 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3723 &tx->tx_putinfo.gnpam_desc,
3724 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3725 tx->tx_putinfo.gnpam_dst_cookie);
3726 RETURN(try_map_if_full ? rc : 0);
3734 kgnilnd_process_fmaq(kgn_conn_t *conn)
3737 kgn_tx_t *tx = NULL;
3738 void *buffer = NULL;
3739 unsigned int nob = 0;
3742 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3743 * However I will be rescheduled by an FMA completion event
3744 * when I eventually get some.
3745 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3746 * But it doesn't matter if I try to send a "real" message just
3747 * as I start closing because I'll get scheduled to send the
3750 /* Short circuit if the ep_handle is null we cant send anyway. */
3751 if (conn->gnc_ephandle == NULL)
3754 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3756 spin_lock(&conn->gnc_list_lock);
3758 if (list_empty(&conn->gnc_fmaq)) {
3759 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3761 spin_unlock(&conn->gnc_list_lock);
3763 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3764 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3765 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3766 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3767 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3769 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3770 conn->gnc_last_noop_cq,
3771 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3772 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3773 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3774 atomic_inc(&conn->gnc_sched_noop);
3775 set_mb(conn->gnc_last_noop_want, jiffies);
3777 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3780 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3784 rc = kgnilnd_set_tx_id(tx, conn);
3786 kgnilnd_tx_done(tx, rc);
3792 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3793 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3794 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3795 more_to_do = !list_empty(&conn->gnc_fmaq);
3796 spin_unlock(&conn->gnc_list_lock);
3799 /* if there is no real TX or no NOOP to send, bail */
3804 if (!tx->tx_retrans)
3805 tx->tx_cred_wait = jiffies;
3807 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3808 "tx with zero id", NULL);
3810 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3811 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3812 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3816 switch (tx->tx_msg.gnm_type) {
3820 case GNILND_MSG_NOOP:
3821 case GNILND_MSG_CLOSE:
3822 case GNILND_MSG_IMMEDIATE:
3823 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3824 buffer = tx->tx_buffer;
3828 case GNILND_MSG_GET_DONE:
3829 case GNILND_MSG_PUT_DONE:
3830 case GNILND_MSG_PUT_DONE_REV:
3831 case GNILND_MSG_GET_DONE_REV:
3832 case GNILND_MSG_PUT_NAK:
3833 case GNILND_MSG_GET_NAK:
3834 case GNILND_MSG_GET_NAK_REV:
3835 case GNILND_MSG_PUT_NAK_REV:
3836 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3839 case GNILND_MSG_PUT_REQ:
3840 case GNILND_MSG_GET_REQ_REV:
3841 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3843 case GNILND_MSG_PUT_ACK:
3844 case GNILND_MSG_PUT_REQ_REV:
3845 case GNILND_MSG_GET_ACK_REV:
3846 case GNILND_MSG_GET_REQ:
3847 /* This is really only to handle the retransmit of SMSG once these
3848 * two messages are setup in send_mapped_tx */
3849 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3853 if (likely(rc == 0)) {
3854 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3858 /* don't explicitly reschedule here - we are short credits and will rely on
3859 * kgnilnd_sendmsg to resched the conn if need be */
3861 } else if (rc < 0) {
3862 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3863 * almost certainly a software bug, but lets play nice with the other kids */
3864 kgnilnd_tx_done(tx, rc);
3865 /* just for fun, kick peer in arse - resetting conn might help to correct
3866 * this almost certainly buggy software caused return code */
3867 kgnilnd_close_conn(conn, rc);
3871 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3872 kgnilnd_schedule_conn(conn);
3877 kgnilnd_process_rdmaq(kgn_device_t *dev)
3882 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3886 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3887 unsigned long dead_bump;
3890 /* if we think we need to adjust, take lock to serialize and recheck */
3891 spin_lock(&dev->gnd_rdmaq_lock);
3892 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3893 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3895 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3897 /* roll the bucket forward */
3898 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3900 if (kgnilnd_data.kgn_rdmaq_override &&
3901 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3902 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3907 /* roll current outstanding forward to make sure we carry outstanding
3908 * committment forward
3909 * new_ok starts out as the whole interval value
3910 * - first subtract bytes_out from last interval, as that would push us over
3911 * strict limits for this interval
3912 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3914 * there is a small race here if someone is actively processing mappings and
3915 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3916 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3917 * - if this gives us problems in the future, we could use a read/write lock
3918 * to protect the resetting of these values */
3919 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3920 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3922 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3923 "current out %ld\n",
3924 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3925 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3927 spin_unlock(&dev->gnd_rdmaq_lock);
3930 spin_lock(&dev->gnd_rdmaq_lock);
3931 while (!list_empty(&dev->gnd_rdmaq)) {
3934 /* make sure we break out early on quiesce */
3935 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3936 /* always break with lock held - we unlock outside loop */
3940 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3941 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3944 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3945 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3946 /* if conn is dying, mark tx in tx_ref_table for
3947 * kgnilnd_complete_closed_conn to finish up */
3948 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3950 /* tx was moved to DYING, get next */
3953 spin_unlock(&dev->gnd_rdmaq_lock);
3955 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3956 spin_lock(&dev->gnd_rdmaq_lock);
3959 /* no ticket! add back to head */
3960 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3961 /* clear found_work so scheduler threads wait for timer */
3965 /* TX is GO for launch */
3966 tx->tx_qtime = jiffies;
3967 kgnilnd_send_mapped_tx(tx, 0);
3971 spin_unlock(&dev->gnd_rdmaq_lock);
3977 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3979 __swab64s(&d->gnrd_key.qword1);
3980 __swab64s(&d->gnrd_key.qword2);
3981 __swab64s(&d->gnrd_addr);
3982 __swab32s(&d->gnrd_nob);
3985 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3986 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3989 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3991 kgn_tx_ev_id_t ev_id;
3994 /* we use the cookie from the original TX, so we can find the match
3995 * by parsing that and using the txe_idx */
3996 ev_id.txe_cookie = cookie;
3998 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
4001 /* check tx to make sure kgni didn't eat it */
4002 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
4003 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
4005 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
4006 (tx->tx_id.txe_cookie = cookie)),
4007 "conn 0x%p->%s tx_ref_table hosed: wanted "
4008 "txe_cookie %#llx txe_idx %d "
4009 "found tx %p cookie %#llx txe_idx %d\n",
4010 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4011 cookie, ev_id.txe_idx,
4012 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
4014 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4015 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4016 "Unexpected TX type (%x, %x or %x) "
4017 "or state (%x, expected +%x) "
4018 "matched reply from %s\n",
4019 tx->tx_msg.gnm_type, type1, type2,
4020 tx->tx_state, GNILND_TX_WAITING_REPLY,
4021 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4023 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
4024 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4030 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4033 kgn_conn_t *conn = tx->tx_conn;
4034 __u64 nob = tx->tx_nob;
4035 __u32 physnop = tx->tx_phys_npages;
4036 int id = tx->tx_id.txe_smsg_id;
4037 int buftype = tx->tx_buftype;
4038 gni_mem_handle_t hndl;
4039 hndl.qword1 = tx->tx_map_key.qword1;
4040 hndl.qword2 = tx->tx_map_key.qword2;
4042 spin_lock(&conn->gnc_list_lock);
4044 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4045 "not waiting for reply", NULL);
4048 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4050 if (rc == -EFAULT) {
4051 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
4052 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4054 if(*kgnilnd_tunables.kgn_efault_lbug) {
4055 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4056 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4058 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4059 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4060 kgnilnd_tx_state2str(tx->tx_list_state),
4061 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4066 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4067 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4068 /* sample under lock as follow on steps require gnc_list_lock
4069 * - or call kgnilnd_tx_done which requires no locks held over
4070 * call to lnet_finalize */
4073 spin_unlock(&conn->gnc_list_lock);
4076 kgnilnd_tx_done(tx, tx->tx_rc);
4081 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4084 kgn_conn_t *conn = tx->tx_conn;
4086 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4087 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4089 /* the gncm_retval is passed in for PUTs */
4090 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4091 msg->gnm_u.completion.gncm_retval);
4093 kgnilnd_complete_tx(tx, rc);
4097 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4105 kgn_peer_t *peer = conn->gnc_peer;
4108 __u16 tmp_cksum = 0, msg_cksum = 0;
4109 int repost = 1, saw_complete;
4110 unsigned long timestamp, newest_last_rx, timeout;
4114 /* Short circuit if the ep_handle is null.
4115 * It's likely that its about to be closed as stale.
4117 if (conn->gnc_ephandle == NULL)
4120 timestamp = jiffies;
4121 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4122 /* delay in jiffies - we are really concerned only with things that
4123 * result in a schedule() or really holding this off for long times .
4124 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4125 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4127 /* Resample current time as we have no idea how long it took to get the mutex */
4128 timestamp = jiffies;
4130 /* We check here when the last time we received an rx, we do this before
4131 * we call getnext in case the thread has been blocked for a while. If we
4132 * havent received an rx since our timeout value we close the connection
4133 * as we should assume the other side has closed the connection. This will
4134 * stop us from sending replies to a mailbox that is already in purgatory.
4137 timeout = cfs_time_seconds(conn->gnc_timeout);
4138 newest_last_rx = GNILND_LASTRX(conn);
4140 /* Error injection to validate that timestamp checking works and closing the conn */
4141 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4142 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4145 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4146 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4147 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4148 cfs_duration_sec(timestamp - newest_last_rx),
4149 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4150 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4152 kgnilnd_close_conn(conn, rc);
4156 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4158 if (rrc == GNI_RC_NOT_DONE) {
4159 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4160 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4164 /* Instead of asserting when we get mailbox corruption lets attempt to
4165 * close the conn and recover. We can put the conn/mailbox into
4166 * purgatory and let purgatory deal with the problem. If we see
4167 * this NETTERROR reported on production systems in large amounts
4168 * we will need to revisit the state machine to see if we can tighten
4169 * it up further to improve data protection.
4172 if (rrc == GNI_RC_INVALID_STATE) {
4173 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4174 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4175 "detected closing conn %p from peer %s\n", conn,
4176 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4178 kgnilnd_close_conn(conn, rc);
4182 LASSERTF(rrc == GNI_RC_SUCCESS,
4183 "bad rc %d on conn %p from peer %s\n",
4184 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4186 msg = (kgn_msg_t *)prefix;
4188 rx = kgnilnd_alloc_rx();
4190 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4191 kgnilnd_release_msg(conn);
4192 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4193 conn, libcfs_nid2str(peer->gnp_nid));
4197 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4199 timestamp = conn->gnc_last_rx;
4200 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4201 atomic_inc(&conn->gnc_rx_seq);
4203 conn->gnc_last_rx = jiffies;
4204 /* stash first rx so we can clear out purgatory
4206 if (conn->gnc_first_rx == 0)
4207 conn->gnc_first_rx = jiffies;
4209 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4210 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4211 kgnilnd_peer_alive(conn->gnc_peer);
4214 rx->grx_conn = conn;
4216 rx->grx_received = current_kernel_time();
4218 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4221 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4227 kgnilnd_net_decref(net);
4230 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4231 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4233 /* XXX Nic: Do we need to swab cksum */
4234 if (msg->gnm_cksum != 0) {
4235 msg_cksum = msg->gnm_cksum;
4237 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4239 if (tmp_cksum != msg_cksum) {
4240 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4241 tmp_cksum, msg_cksum);
4242 kgnilnd_dump_msg(D_BUFFS, msg);
4247 /* restore checksum for future debug messages */
4248 msg->gnm_cksum = tmp_cksum;
4250 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4251 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4252 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4253 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4258 __swab32s(&msg->gnm_magic);
4259 __swab16s(&msg->gnm_version);
4260 __swab16s(&msg->gnm_type);
4261 __swab64s(&msg->gnm_srcnid);
4262 __swab64s(&msg->gnm_connstamp);
4263 __swab32s(&msg->gnm_seq);
4265 /* NB message type checked below; NOT here... */
4266 switch (msg->gnm_type) {
4267 case GNILND_MSG_GET_ACK_REV:
4268 case GNILND_MSG_PUT_ACK:
4269 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4272 case GNILND_MSG_PUT_REQ_REV:
4273 case GNILND_MSG_GET_REQ:
4274 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4282 if (msg->gnm_version != GNILND_MSG_VERSION) {
4283 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4284 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4289 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4290 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4291 libcfs_nid2str(msg->gnm_srcnid),
4292 libcfs_nid2str(peer->gnp_nid));
4297 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4298 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4299 " expected) from %s",
4300 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4301 libcfs_nid2str(peer->gnp_nid));
4306 if (msg->gnm_seq != seq) {
4307 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4308 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4313 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4315 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4316 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4317 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4318 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4319 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4320 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4321 /* double check state with lock held */
4322 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4323 /* only error if we are not already closing */
4324 if (conn->gnc_peer_error == -ETIMEDOUT) {
4325 unsigned long now = jiffies;
4326 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4328 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4329 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4330 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4332 cfs_duration_sec(now - timestamp),
4333 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4334 atomic_read(&conn->gnc_tx_seq),
4335 cfs_duration_sec(now - conn->gnc_last_tx),
4336 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4337 cfs_duration_sec(now - conn->gnc_last_noop_want),
4338 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4339 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4340 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4341 cfs_duration_sec(now - conn->gnc_last_sched_do),
4342 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4344 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4346 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4350 if (conn->gnc_close_recvd) {
4351 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4352 kgnilnd_msgtype2str(msg->gnm_type),
4353 msg->gnm_type, conn->gnc_close_recvd,
4354 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4359 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4360 /* XXX Nic: log message received on bad connection state */
4364 switch (msg->gnm_type) {
4365 case GNILND_MSG_NOOP:
4366 /* Nothing to do; just a keepalive */
4369 case GNILND_MSG_IMMEDIATE:
4370 /* only get SMSG payload for IMMEDIATE */
4371 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4372 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4373 msg->gnm_srcnid, rx, 0);
4376 case GNILND_MSG_GET_REQ_REV:
4377 case GNILND_MSG_PUT_REQ:
4378 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4379 msg->gnm_srcnid, rx, 1);
4382 case GNILND_MSG_GET_NAK_REV:
4383 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4384 msg->gnm_u.completion.gncm_cookie);
4388 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4390 case GNILND_MSG_PUT_NAK:
4391 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4392 msg->gnm_u.completion.gncm_cookie);
4396 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4398 case GNILND_MSG_PUT_ACK:
4399 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4400 msg->gnm_u.putack.gnpam_src_cookie);
4404 /* store putack data for later: deferred rdma or re-try */
4405 tx->tx_putinfo = msg->gnm_u.putack;
4408 spin_lock(&tx->tx_conn->gnc_list_lock);
4410 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4411 "not waiting for reply", NULL);
4413 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4415 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4416 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4417 /* sample under lock as follow on steps require gnc_list_lock
4418 * - or call kgnilnd_tx_done which requires no locks held over
4419 * call to lnet_finalize */
4422 /* cannot launch rdma if still waiting for fma-msg completion */
4423 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4424 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4425 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4427 spin_unlock(&tx->tx_conn->gnc_list_lock);
4430 rc = kgnilnd_send_mapped_tx(tx, 0);
4432 kgnilnd_tx_done(tx, rc);
4435 case GNILND_MSG_GET_ACK_REV:
4436 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4437 msg->gnm_u.putack.gnpam_src_cookie);
4441 /* store putack data for later: deferred rdma or re-try */
4442 tx->tx_putinfo = msg->gnm_u.putack;
4444 spin_lock(&tx->tx_conn->gnc_list_lock);
4446 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4447 "not waiting for reply", NULL);
4449 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4451 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4452 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4453 /* sample under lock as follow on steps require gnc_list_lock
4454 * - or call kgnilnd_tx_done which requires no locks held over
4455 * call to lnet_finalize */
4458 /* cannot launch rdma if still waiting for fma-msg completion */
4459 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4460 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4461 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4463 spin_unlock(&tx->tx_conn->gnc_list_lock);
4466 rc = kgnilnd_send_mapped_tx(tx, 0);
4468 kgnilnd_tx_done(tx, rc);
4471 case GNILND_MSG_PUT_DONE:
4472 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4473 msg->gnm_u.completion.gncm_cookie);
4477 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4478 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4479 "bad tx buftype %d", tx->tx_buftype);
4481 kgnilnd_finalize_rx_done(tx, msg);
4483 case GNILND_MSG_PUT_REQ_REV:
4484 case GNILND_MSG_GET_REQ:
4485 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4486 msg->gnm_srcnid, rx, 1);
4490 case GNILND_MSG_GET_NAK:
4491 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4492 msg->gnm_u.completion.gncm_cookie);
4496 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4497 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4498 "bad tx buftype %d", tx->tx_buftype);
4500 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4503 case GNILND_MSG_GET_DONE:
4504 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4505 msg->gnm_u.completion.gncm_cookie);
4509 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4510 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4511 "bad tx buftype %d", tx->tx_buftype);
4513 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4514 msg->gnm_u.completion.gncm_retval);
4516 kgnilnd_finalize_rx_done(tx, msg);
4518 case GNILND_MSG_GET_DONE_REV:
4519 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4520 msg->gnm_u.completion.gncm_cookie);
4524 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4525 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4526 "bad tx buftype %d", tx->tx_buftype);
4528 kgnilnd_finalize_rx_done(tx, msg);
4531 case GNILND_MSG_PUT_DONE_REV:
4532 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4533 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);
4544 case GNILND_MSG_PUT_NAK_REV:
4545 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4546 msg->gnm_u.completion.gncm_cookie);
4551 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4552 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4553 "bad tx buftype %d", tx->tx_buftype);
4555 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4560 if (rc < 0) /* protocol/comms error */
4561 kgnilnd_close_conn(conn, rc);
4563 if (repost && rx != NULL) {
4564 kgnilnd_consume_rx(rx);
4567 /* we got an event so assume more there and call for reschedule */
4569 kgnilnd_schedule_conn(conn);
4573 /* Do the failure injections that we need to affect conn processing in the following function.
4574 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4575 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4578 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4582 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4586 /* short circuit out when not set */
4587 if (likely(!cfs_fail_loc)) {
4591 /* failure injection to test for stack reset clean ups */
4592 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4593 /* we can't rely on busy loops being nice enough to get the
4594 * stack reset triggered - it'd just spin on this conn */
4595 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4598 GOTO(did_fail_loc, rc);
4601 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4602 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4604 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4605 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4608 GOTO(did_fail_loc, rc);
4612 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4613 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4614 * will be closed by kgnilnd_complete_closed_conn.
4616 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4617 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4620 GOTO(did_fail_loc, rc);
4623 /* this one is a bit gross - we can't hold the mutex from process_conns
4624 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4625 * and moving onto finish_connect
4626 * so, we'll just set the rc - kgnilnd_process_conns will clear
4627 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4628 * and effectively getting this thread to sleep */
4629 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4632 GOTO(did_fail_loc, rc);
4640 kgnilnd_send_conn_close(kgn_conn_t *conn)
4644 /* we are closing the conn - we will try to send the CLOSE msg
4645 * but will not wait for anything else to flush */
4647 /* send the close if not already done so or received one */
4648 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4649 /* set close_sent regardless of the success of the
4650 * CLOSE message. We are going to try once and then
4651 * kick him out of the sandbox */
4652 conn->gnc_close_sent = 1;
4655 /* EP might be null already if remote side initiated a new connection.
4656 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4657 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4659 if (conn->gnc_ephandle != NULL) {
4662 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4664 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4665 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4666 tx->tx_qtime = jiffies;
4668 if (tx->tx_id.txe_idx == 0) {
4669 rc = kgnilnd_set_tx_id(tx, conn);
4671 kgnilnd_tx_done(tx, rc);
4675 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4678 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4679 kgnilnd_tx_done(tx, -EAGAIN);
4681 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4683 /* It wasnt sent and we dont care. */
4684 kgnilnd_tx_done(tx, rc);
4692 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4693 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4694 conn->gnc_state = GNILND_CONN_CLOSED;
4695 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4696 /* mark this conn as CLOSED now that we processed it
4697 * do after TX, so we can use CLOSING in asserts */
4701 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4702 /* simulate a RX CLOSE after the timeout but before
4703 * the scheduler thread gets it */
4704 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4705 conn->gnc_peer_error = -ETIMEDOUT;
4707 /* schedule to allow potential CLOSE and get the complete phase run */
4708 kgnilnd_schedule_conn(conn);
4712 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4717 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4718 int log_retrans, log_retrans_level;
4719 static int last_map_version;
4722 spin_lock(&dev->gnd_lock);
4723 if (list_empty(&dev->gnd_map_tx)) {
4724 /* if the list is empty make sure we dont have a timer running */
4725 del_singleshot_timer_sync(&dev->gnd_map_timer);
4726 spin_unlock(&dev->gnd_lock);
4730 dev->gnd_sched_alive = jiffies;
4732 /* we'll retry as fast as possible up to 25% of the limit, then we start
4733 * backing off until our map version changes - indicating we unmapped
4735 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4736 if (likely(dev->gnd_map_attempt == 0) ||
4737 time_after_eq(jiffies, dev->gnd_next_map) ||
4738 last_map_version != dev->gnd_map_version) {
4740 /* if this is our first attempt at mapping set last mapped to current
4741 * jiffies so we can timeout our attempt correctly.
4743 if (dev->gnd_map_attempt == 0)
4744 dev->gnd_last_map = jiffies;
4746 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4747 spin_unlock(&dev->gnd_lock);
4751 /* delete the previous timer if it exists */
4752 del_singleshot_timer_sync(&dev->gnd_map_timer);
4753 /* stash the last map version to let us know when a good one was seen */
4754 last_map_version = dev->gnd_map_version;
4756 /* we need to to take the lock and continually refresh the head of the list as
4757 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4758 * allowing them to squeeze in */
4760 while (!list_empty(&dev->gnd_map_tx)) {
4761 /* make sure we break out early on quiesce */
4762 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4763 /* always break with lock held - we unlock outside loop */
4767 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4769 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4772 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4773 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4774 /* if conn is dying, mark tx in tx_ref_table for
4775 * kgnilnd_complete_closed_conn to finish up */
4776 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4779 /* tx was moved to DYING, get next */
4783 spin_unlock(&dev->gnd_lock);
4784 rc = kgnilnd_send_mapped_tx(tx, 1);
4786 /* We made it! skip error handling.. */
4788 /* OK to continue on +ve errors as it won't get seen until
4789 * this function is called again - we operate on a copy of the original
4790 * list and not the live list */
4791 spin_lock(&dev->gnd_lock);
4792 /* reset map attempts back to zero we successfully
4793 * mapped so we can reset our timers */
4794 dev->gnd_map_attempt = 0;
4796 } else if (rc == -EAGAIN) {
4797 spin_lock(&dev->gnd_lock);
4798 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4799 spin_unlock(&dev->gnd_lock);
4800 GOTO(get_out_mapped, rc);
4801 } else if (rc != -ENOMEM) {
4802 /* carp, failure we can't handle */
4803 kgnilnd_tx_done(tx, rc);
4804 spin_lock(&dev->gnd_lock);
4805 /* reset map attempts back to zero we dont know what happened but it
4806 * wasnt a failed mapping
4808 dev->gnd_map_attempt = 0;
4812 /* time to handle the retry cases.. lock so we dont have 2 threads
4813 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4815 spin_lock(&dev->gnd_lock);
4816 dev->gnd_map_attempt++;
4817 if (dev->gnd_map_attempt < fast_remaps) {
4818 /* do nothing we just want it to go as fast as possible.
4819 * just set gnd_next_map to current jiffies so it will process
4820 * as fast as possible.
4822 dev->gnd_next_map = jiffies;
4824 /* Retry based on GNILND_MAP_RETRY_RATE */
4825 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4828 /* only log occasionally once we've retried fast_remaps */
4829 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4830 ((dev->gnd_map_attempt % fast_remaps) == 0);
4831 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4833 /* make sure we are not off in the weeds with this tx */
4834 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4835 GNIDBG_TX(D_NETERROR, tx,
4836 "giving up on TX, too many retries", NULL);
4837 spin_unlock(&dev->gnd_lock);
4838 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4839 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4840 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4842 tx->tx_putinfo.gnpam_dst_cookie,
4843 tx->tx_msg.gnm_srcnid);
4845 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4847 tx->tx_getinfo.gngm_cookie,
4848 tx->tx_msg.gnm_srcnid);
4850 kgnilnd_tx_done(tx, -ENOMEM);
4851 GOTO(get_out_mapped, rc);
4853 GNIDBG_TX(log_retrans_level, tx,
4854 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4856 "nq_map %d mdd# %d/%d GART %ld",
4857 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4858 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4859 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4860 atomic_read(&dev->gnd_nq_map),
4861 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4862 atomic64_read(&dev->gnd_nbytes_map));
4865 /* we need to stop processing the rest of the list, so add it back in */
4866 /* set timer to wake device when we need to schedule this tx */
4867 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4868 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4869 spin_unlock(&dev->gnd_lock);
4870 GOTO(get_out_mapped, rc);
4872 spin_unlock(&dev->gnd_lock);
4878 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4883 int error_inject = 0;
4887 spin_lock(&dev->gnd_lock);
4888 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4889 dev->gnd_sched_alive = jiffies;
4893 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4894 /* break with lock held */
4898 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4899 list_del_init(&conn->gnc_schedlist);
4901 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4904 if (!list_empty(&conn->gnc_delaylist))
4905 list_del_init(&conn->gnc_delaylist);
4906 spin_unlock(&dev->gnd_lock);
4908 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4910 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4911 conn_sched != GNILND_CONN_PROCESS,
4912 "conn %p on ready list but in bad state: %d\n",
4915 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4916 conn, kgnilnd_conn_state2str(conn));
4919 set_mb(conn->gnc_last_sched_do, jiffies);
4921 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4923 /* based on intent see if we should run again. */
4924 rc = kgnilnd_schedule_process_conn(conn, intent);
4926 /* drop ref from gnd_ready_conns */
4927 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4928 down_write(&dev->gnd_conn_sem);
4929 kgnilnd_conn_decref(conn);
4930 up_write(&dev->gnd_conn_sem);
4931 } else if (rc != 1) {
4932 kgnilnd_conn_decref(conn);
4934 /* clear this so that scheduler thread doesn't spin */
4936 /* break with lock held... */
4937 spin_lock(&dev->gnd_lock);
4941 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4942 down_write(&dev->gnd_conn_sem);
4944 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4945 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4946 /* If there are tx's currently in use in another
4947 * thread we dont want to complete the close
4948 * yet. Cycle this conn back through
4950 kgnilnd_schedule_conn(conn);
4952 kgnilnd_complete_closed_conn(conn);
4954 up_write(&dev->gnd_conn_sem);
4955 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4956 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4957 /* serialize SMSG CQs with ep_bind and smsg_release */
4958 down_write(&dev->gnd_conn_sem);
4959 kgnilnd_destroy_conn_ep(conn);
4960 up_write(&dev->gnd_conn_sem);
4961 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4962 /* if we need to do some CLOSE sending, etc done here do it */
4963 down_write(&dev->gnd_conn_sem);
4964 kgnilnd_send_conn_close(conn);
4965 kgnilnd_check_fma_rx(conn);
4966 up_write(&dev->gnd_conn_sem);
4967 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4968 /* start moving traffic if the old conns are cleared out */
4969 down_read(&dev->gnd_conn_sem);
4970 kgnilnd_check_fma_rx(conn);
4971 kgnilnd_process_fmaq(conn);
4972 up_read(&dev->gnd_conn_sem);
4975 rc = kgnilnd_schedule_process_conn(conn, 0);
4977 /* drop ref from gnd_ready_conns */
4978 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4979 down_write(&dev->gnd_conn_sem);
4980 kgnilnd_conn_decref(conn);
4981 up_write(&dev->gnd_conn_sem);
4982 } else if (rc != 1) {
4983 kgnilnd_conn_decref(conn);
4986 /* check list again with lock held */
4987 spin_lock(&dev->gnd_lock);
4990 /* If we are short circuiting due to timing we want to be scheduled
4991 * as soon as possible.
4993 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4996 spin_unlock(&dev->gnd_lock);
5002 kgnilnd_scheduler(void *arg)
5004 int threadno = (long)arg;
5007 unsigned long deadline = 0;
5010 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
5012 /* all gnilnd threads need to run fairly urgently */
5013 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
5014 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5015 while (!kgnilnd_data.kgn_shutdown) {
5017 /* Safe: kgn_shutdown only set when quiescent */
5019 /* to quiesce or to not quiesce, that is the question */
5021 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5022 KGNILND_SPIN_QUIESCE;
5025 /* tracking for when thread goes AWOL */
5026 dev->gnd_sched_alive = jiffies;
5028 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5029 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5030 /* let folks know we are up and kicking
5031 * - they can use this for latency savings, etc
5032 * - only change if IRQ, if IDLE leave alone as that
5033 * schedule_device calls to put us back to IRQ */
5034 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5036 down_read(&dev->gnd_conn_sem);
5037 /* always check these - they are super low cost */
5038 found_work += kgnilnd_check_fma_send_cq(dev);
5039 found_work += kgnilnd_check_fma_rcv_cq(dev);
5041 /* rdma CQ doesn't care about eps */
5042 found_work += kgnilnd_check_rdma_cq(dev);
5044 /* move some RDMA ? */
5045 found_work += kgnilnd_process_rdmaq(dev);
5047 /* map some pending RDMA requests ? */
5048 found_work += kgnilnd_process_mapped_tx(dev);
5050 /* the EP for a conn is not destroyed until all the references
5051 * to it are gone, so these checks should be safe
5052 * even if run in parallel with the CQ checking functions
5053 * _AND_ a thread that processes the CLOSED->DONE
5057 up_read(&dev->gnd_conn_sem);
5059 /* process all conns ready now */
5060 found_work += kgnilnd_process_conns(dev, deadline);
5062 /* do an eager check to avoid the IRQ disabling in
5063 * prepare_to_wait and friends */
5066 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5067 time_before(jiffies, deadline)) {
5069 if ((busy_loops % 10) == 0) {
5070 /* tickle heartbeat and watchdog to ensure our
5071 * piggishness doesn't turn into heartbeat failure */
5072 touch_nmi_watchdog();
5078 /* if we got here, found_work was zero or busy_loops means we
5079 * need to take a break. We'll clear gnd_ready but we'll check
5080 * one last time if there is an IRQ that needs processing */
5082 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5084 /* the first time this will go LOOP -> IDLE and let us do one final check
5085 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5086 * - this might allow other threads to block us for a bit if they
5087 * try to get the mutex, but that is good as we'd need to wake
5088 * up soon to handle the CQ or other processing anyways */
5090 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5092 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5093 time_after_eq(jiffies, deadline)) {
5095 "yeilding: found_work %d busy_loops %d\n",
5096 found_work, busy_loops);
5098 /* use yield if we are bailing due to busy_loops
5099 * - this will ensure we wake up soonish. This closes
5100 * a race with kgnilnd_device_callback - where it'd
5101 * not call wake_up() because gnd_ready == 1, but then
5102 * we come down and schedule() because of busy_loops.
5103 * We'd not be woken up until something poked our waitq
5104 * again. yield() ensures we wake up without another
5105 * waitq poke in that case */
5106 atomic_inc(&dev->gnd_n_yield);
5107 kgnilnd_data.kgn_last_condresched = jiffies;
5109 CDEBUG(D_INFO, "awake after yeild\n");
5110 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5111 } else if (found_work == GNILND_DEV_IDLE) {
5112 /* busy_loops is low and there is nothing to do,
5113 * go to sleep and wait for a waitq poke */
5115 "scheduling: found_work %d busy_loops %d\n",
5116 found_work, busy_loops);
5117 atomic_inc(&dev->gnd_n_schedule);
5118 kgnilnd_data.kgn_last_scheduled = jiffies;
5120 CDEBUG(D_INFO, "awake after schedule\n");
5121 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5123 finish_wait(&dev->gnd_waitq, &wait);
5126 kgnilnd_thread_fini();