2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by Eric Barton <eric@bartonsoftware.com>
7 * Author: James Shimek <jshimek@cray.com>
8 * Author: Nic Henke <nic@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #include <linux/nmi.h>
29 #include <linux/pagemap.h>
31 #include <libcfs/linux/linux-mem.h>
35 /* this is useful when needed to debug wire corruption. */
37 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
45 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
46 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
49 } else if (len >= 16) {
51 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
52 prefix, ptr, *(ptr), *(ptr + 1));
56 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
65 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
67 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
68 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
69 msg->gnm_magic, msg->gnm_version,
70 msg->gnm_type, msg->gnm_srcnid,
71 msg->gnm_connstamp, msg->gnm_seq,
72 msg->gnm_cksum, msg->gnm_payload_cksum,
73 msg->gnm_payload_len);
77 kgnilnd_schedule_device(kgn_device_t *dev)
79 short already_live = 0;
81 /* we'll only want to wake if the scheduler thread
82 * has come around and set ready to zero */
83 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
86 wake_up_all(&dev->gnd_waitq);
90 void kgnilnd_schedule_device_timer(unsigned long arg)
92 kgn_device_t *dev = (kgn_device_t *) arg;
94 kgnilnd_schedule_device(dev);
98 kgnilnd_device_callback(__u32 devid, __u64 arg)
101 int index = (int) arg;
103 if (index >= kgnilnd_data.kgn_ndevs) {
104 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
105 * interrupt context. */
106 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
111 dev = &kgnilnd_data.kgn_devices[index];
112 /* just basic sanity */
113 if (dev->gnd_id == devid) {
114 kgnilnd_schedule_device(dev);
116 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
121 /* sched_intent values:
122 * < 0 : do not reschedule under any circumstances
123 * == 0: reschedule if someone marked him WANTS_SCHED
124 * > 0 : force a reschedule */
125 /* Return code 0 means it did not schedule the conn, 1
126 * means it successfully scheduled the conn.
130 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
134 /* move back to IDLE but save previous state.
135 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
136 * let the xchg there handle any racing callers to get it
137 * onto gnd_ready_conns */
139 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
140 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
141 conn_sched == GNILND_CONN_PROCESS,
142 "conn %p after process in bad state: %d\n",
145 if (sched_intent >= 0) {
146 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
147 return kgnilnd_schedule_conn_refheld(conn, 1);
153 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
157 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
159 kgn_device_t *dev = conn->gnc_device;
163 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
164 /* we only care about the last person who marked want_sched since they
165 * are most likely the culprit
167 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
168 conn->gnc_sched_line = line;
169 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
170 * the chance to put it onto gnd_ready_conns.
171 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
172 * the conn in process_conns will take care of moving it back to
173 * SCHED when it is done processing */
175 if (sched == GNILND_CONN_IDLE) {
176 /* if the conn is already scheduled, we've already requested
177 * the scheduler thread wakeup */
179 /* Add a reference to the conn if we are not holding a reference
180 * already from the exisiting scheduler. We now use the same
181 * reference if we need to reschedule a conn while in a scheduler
184 kgnilnd_conn_addref(conn);
186 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
189 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
191 spin_lock(&dev->gnd_lock);
192 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
194 spin_unlock(&dev->gnd_lock);
195 set_mb(conn->gnc_last_sched_ask, jiffies);
198 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
202 /* make sure thread(s) going to process conns - but let it make
203 * separate decision from conn schedule */
205 kgnilnd_schedule_device(dev);
210 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
212 kgn_device_t *dev = conn->gnc_device;
214 spin_lock(&dev->gnd_lock);
215 if (list_empty(&conn->gnc_delaylist)) {
216 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
219 spin_unlock(&dev->gnd_lock);
221 kgnilnd_schedule_device(dev);
226 kgnilnd_schedule_dgram(kgn_device_t *dev)
230 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
231 if (wake != GNILND_DGRAM_SCHED) {
232 wake_up(&dev->gnd_dgram_waitq);
234 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
239 kgnilnd_free_tx(kgn_tx_t *tx)
241 /* taken from kgnilnd_tx_add_state_locked */
243 LASSERTF((tx->tx_list_p == NULL &&
244 tx->tx_list_state == GNILND_TX_ALLOCD) &&
245 list_empty(&tx->tx_list),
246 "tx %p with bad state %s (list_p %p) tx_list %s\n",
247 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
248 list_empty(&tx->tx_list) ? "empty" : "not empty");
250 atomic_dec(&kgnilnd_data.kgn_ntx);
252 /* we only allocate this if we need to */
253 if (tx->tx_phys != NULL) {
254 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
255 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
256 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
259 /* Only free the buffer if we used it */
260 if (tx->tx_buffer_copy != NULL) {
261 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
262 tx->tx_buffer_copy = NULL;
263 CDEBUG(D_MALLOC, "vfreed buffer2\n");
266 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
268 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
269 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
273 kgnilnd_alloc_tx (void)
277 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
280 tx = kmem_cache_zalloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
282 CERROR("failed to allocate tx\n");
285 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
288 /* setup everything here to minimize time under the lock */
289 tx->tx_buftype = GNILND_BUF_NONE;
290 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
291 INIT_LIST_HEAD(&tx->tx_list);
292 INIT_LIST_HEAD(&tx->tx_map_list);
293 tx->tx_list_state = GNILND_TX_ALLOCD;
295 atomic_inc(&kgnilnd_data.kgn_ntx);
300 /* csum_fold needs to be run on the return value before shipping over the wire */
301 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
303 /* we don't use offset as every one is passing a buffer reference that already
304 * includes the offset into the base address -
305 * see kgnilnd_setup_virt_buffer and kgnilnd_setup_immediate_buffer */
307 kgnilnd_cksum(void *ptr, size_t nob)
311 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
313 /* don't use magic 'no checksum' value */
317 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
324 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
325 unsigned int offset, unsigned int nob, int dump_blob)
331 unsigned int fraglen;
337 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
338 kiov, nkiov, offset, nob, dump_blob);
340 /* if loops changes, please change kgnilnd_setup_phys_buffer */
342 while (offset >= kiov->bv_len) {
343 offset -= kiov->bv_len;
349 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
350 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
352 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
353 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
355 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
356 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
358 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
359 odd, kiov[0].bv_len, offset, nob);
361 for (i = 0; i < nkiov; i++) {
362 pages[i] = kiov[i].bv_page;
365 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
367 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
369 /* return zero to avoid killing tx - we'll just get warning on console
370 * when remote end sees zero checksum */
373 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
375 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
380 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
381 (void *)addr + kiov[0].bv_offset +
384 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
385 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
389 fraglen = min(kiov->bv_len - offset, nob);
391 /* make dang sure we don't send a bogus checksum if somehow we get
392 * an odd length fragment on anything but the last entry in a kiov -
393 * we know from kgnilnd_setup_rdma_buffer that we can't have non
394 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
395 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
396 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
397 fraglen, nkiov, nob, kiov->bv_len,
400 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
402 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
405 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
406 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
407 addr, fraglen, offset);
412 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
414 kunmap(kiov->bv_page);
421 /* iov must not run out before end of data */
422 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
427 retsum = csum_fold(cksum);
429 /* don't use magic 'no checksum' value */
433 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
439 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
441 msg->gnm_magic = GNILND_MSG_MAGIC;
442 msg->gnm_version = GNILND_MSG_VERSION;
443 msg->gnm_type = type;
444 msg->gnm_payload_len = 0;
445 msg->gnm_srcnid = source;
446 /* gnm_connstamp gets set when FMA is sent */
447 /* gnm_srcnid is set on creation via function argument
448 * The right interface/net and nid is passed in when the message
454 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
456 kgn_tx_t *tx = kgnilnd_alloc_tx();
459 kgnilnd_init_msg(&tx->tx_msg, type, source);
461 CERROR("couldn't allocate new tx type %s!\n",
462 kgnilnd_msgtype2str(type));
469 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
475 case GNILND_MSG_GET_REQ:
476 case GNILND_MSG_GET_DONE:
477 nak_type = GNILND_MSG_GET_NAK;
479 case GNILND_MSG_PUT_REQ:
480 case GNILND_MSG_PUT_ACK:
481 case GNILND_MSG_PUT_DONE:
482 nak_type = GNILND_MSG_PUT_NAK;
484 case GNILND_MSG_PUT_REQ_REV:
485 case GNILND_MSG_PUT_DONE_REV:
486 nak_type = GNILND_MSG_PUT_NAK_REV;
488 case GNILND_MSG_GET_REQ_REV:
489 case GNILND_MSG_GET_ACK_REV:
490 case GNILND_MSG_GET_DONE_REV:
491 nak_type = GNILND_MSG_GET_NAK_REV;
494 CERROR("invalid msg type %s (%d)\n",
495 kgnilnd_msgtype2str(rx_type), rx_type);
498 /* only allow NAK on error and truncate to zero */
499 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
500 error, conn, cookie);
502 tx = kgnilnd_new_tx_msg(nak_type, source);
504 CNETERR("can't get TX to NAK RDMA to %s\n",
505 libcfs_nid2str(conn->gnc_peer->gnp_nid));
509 tx->tx_msg.gnm_u.completion.gncm_retval = error;
510 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
511 kgnilnd_queue_tx(conn, tx);
515 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
516 struct kvec *iov, struct bio_vec *kiov,
517 unsigned int offset, unsigned int nob)
519 kgn_msg_t *msg = &tx->tx_msg;
522 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
523 * gni_smsg_send to send that as the payload */
525 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
528 tx->tx_buffer = NULL;
529 } else if (kiov != NULL) {
531 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
532 niov = round_up(nob + offset + kiov->bv_offset,
536 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
537 "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n",
538 niov, msg, kiov, iov, offset, nob);
540 while (offset >= kiov->bv_len) {
541 offset -= kiov->bv_len;
546 for (i = 0; i < niov; i++) {
547 /* We can't have a bv_offset on anything but the first
548 * entry, otherwise we'll have a hole at the end of the
549 * mapping as we only map whole pages.
550 * Also, if we have a bv_len < PAGE_SIZE but we need to
551 * map more than bv_len, we will also have a whole at
552 * the end of that page which isn't allowed
554 if ((kiov[i].bv_offset != 0 && i > 0) ||
555 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
557 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
558 i, offset, nob, kiov->bv_offset,
562 tx->tx_imm_pages[i] = kiov[i].bv_page;
565 /* hijack tx_phys for the later unmap */
567 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
569 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
570 kiov[0].bv_offset + offset;
571 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
572 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
573 nob, kiov, tx->tx_buffer);
575 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
576 if (tx->tx_phys == NULL) {
577 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
581 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
582 /* make sure we take into account the kiov offset as the
583 * start of the buffer
585 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
588 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
589 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
591 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
595 /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
596 * could "flatten" the payload into a single contiguous buffer ready
597 * for sending direct over an FMA if we ever needed to. */
601 while (offset >= iov->iov_len) {
602 offset -= iov->iov_len;
608 if (nob > iov->iov_len - offset) {
609 CERROR("Can't handle multiple vaddr fragments\n");
613 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
615 tx->tx_buftype = GNILND_BUF_IMMEDIATE;
619 /* checksum payload early - it shouldn't be changing after lnd_send */
620 if (*kgnilnd_tunables.kgn_checksum >= 2) {
621 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
622 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
623 msg->gnm_payload_cksum += 0xe00e;
625 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
626 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
630 msg->gnm_payload_cksum = 0;
637 kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
638 unsigned int niov, struct kvec *iov,
639 unsigned int offset, unsigned int nob)
644 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
646 while (offset >= iov->iov_len) {
647 offset -= iov->iov_len;
653 if (nob > iov->iov_len - offset) {
654 CERROR("Can't handle multiple vaddr fragments\n");
658 tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
660 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
665 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
666 unsigned int offset, unsigned int nob)
668 gni_mem_segment_t *phys;
670 unsigned int fraglen;
672 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
676 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
678 /* only allocate this if we are going to use it */
679 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
681 if (tx->tx_phys == NULL) {
682 CERROR("failed to allocate tx_phys\n");
687 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
688 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
690 /* if loops changes, please change kgnilnd_cksum_kiov
691 * and kgnilnd_setup_immediate_buffer */
693 while (offset >= kiov->bv_len) {
694 offset -= kiov->bv_len;
700 /* at this point, kiov points to the first page that we'll actually map
701 * now that we've seeked into the koiv for offset and dropped any
702 * leading pages that fall entirely within the offset */
703 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
706 /* bv_offset is start of 'valid' buffer, so index offset past that */
707 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
710 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
711 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
714 fraglen = min(kiov->bv_len - offset, nob);
716 /* We can't have a bv_offset on anything but the first entry,
717 * otherwise we'll have a hole at the end of the mapping as we
718 * only map whole pages. Only the first page is allowed to
719 * have an offset - we'll add that into tx->tx_buffer and that
720 * will get used when we map in the segments (see
721 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
722 * but we need to map more than bv_len, we will also have a
723 * whole at the end of that page which isn't allowed
725 if ((phys != tx->tx_phys) &&
726 ((kiov->bv_offset != 0) ||
727 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
728 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
729 (int)(phys - tx->tx_phys),
730 offset, nob, kiov->bv_offset, kiov->bv_len);
735 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
736 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
741 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
747 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
748 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
751 phys->address = page_to_phys(kiov->bv_page);
758 /* iov must not run out before end of data */
759 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
763 tx->tx_phys_npages = phys - tx->tx_phys;
768 if (tx->tx_phys != NULL) {
769 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
770 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
771 sizeof(*tx->tx_phys), tx->tx_phys);
778 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
779 struct kvec *iov, struct bio_vec *kiov,
780 unsigned int offset, unsigned int nob)
784 LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
785 " offset %d, nob %d, niov %d\n"
786 , iov, kiov, tx, offset, nob, niov);
789 rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
791 rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
796 /* kgnilnd_parse_lnet_rdma()
797 * lntmsg - message passed in from lnet.
798 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
799 * nob - actual number of bytes to in this message.
800 * put_len - It is possible for PUTs to have a different length than the
801 * length stored in lntmsg->msg_len since LNET can adjust this
802 * length based on it's buffer size and offset.
803 * lnet_try_match_md() sets the mlength that we use to do the RDMA
807 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
808 unsigned int *offset, unsigned int *nob,
809 struct bio_vec **kiov, int put_len)
811 /* GETs are weird, see kgnilnd_send */
812 if (lntmsg->msg_type == LNET_MSG_GET) {
813 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
816 *kiov = lntmsg->msg_md->md_kiov;
818 *niov = lntmsg->msg_md->md_niov;
819 *nob = lntmsg->msg_md->md_length;
822 *kiov = lntmsg->msg_kiov;
823 *niov = lntmsg->msg_niov;
825 *offset = lntmsg->msg_offset;
830 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
832 unsigned int niov, offset, nob;
833 struct bio_vec *kiov;
834 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
835 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
837 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
838 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
839 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
840 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
841 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
842 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
843 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
845 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
846 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
847 tx->tx_msg.gnm_payload_cksum = 0;
850 if (*kgnilnd_tunables.kgn_checksum < 3) {
851 tx->tx_msg.gnm_payload_cksum = 0;
855 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
857 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
861 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
863 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
865 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
869 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
870 tx->tx_msg.gnm_payload_cksum += 0xd00d;
874 /* kgnilnd_verify_rdma_cksum()
875 * tx - PUT_DONE/GET_DONE matched tx.
876 * rx_cksum - received checksum to compare against.
877 * put_len - see kgnilnd_parse_lnet_rdma comments.
880 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
884 unsigned int niov, offset, nob;
885 struct bio_vec *kiov;
886 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
887 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
889 /* we can only match certain requests */
890 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
891 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
892 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
893 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
894 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
895 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
896 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
898 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
899 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
904 if (*kgnilnd_tunables.kgn_checksum >= 3) {
905 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
906 "no RDMA payload checksum when enabled");
911 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
913 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
916 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
918 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
921 if (cksum != rx_cksum) {
922 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
923 "Bad RDMA payload checksum (%x expected %x); "
924 "kiov 0x%p niov %d nob %u offset %u",
925 cksum, rx_cksum, kiov, niov, nob, offset);
926 switch (dump_on_err) {
929 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
931 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
934 /* fall through to dump log */
936 libcfs_debug_dumplog();
942 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
948 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
952 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
953 "already mapped!", NULL);
955 spin_lock(&dev->gnd_map_lock);
956 switch (tx->tx_buftype) {
958 GNIDBG_TX(D_EMERG, tx,
959 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
960 spin_unlock(&dev->gnd_map_lock);
964 case GNILND_BUF_PHYS_MAPPED:
965 bytes = tx->tx_phys_npages * PAGE_SIZE;
966 dev->gnd_map_nphys++;
967 dev->gnd_map_physnop += tx->tx_phys_npages;
970 case GNILND_BUF_VIRT_MAPPED:
972 dev->gnd_map_nvirt++;
973 dev->gnd_map_virtnob += tx->tx_nob;
977 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
978 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
979 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
980 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
981 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
984 atomic_inc(&dev->gnd_n_mdd);
985 atomic64_add(bytes, &dev->gnd_nbytes_map);
987 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
990 /* we only get here in the valid cases */
991 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
992 dev->gnd_map_version++;
993 spin_unlock(&dev->gnd_map_lock);
997 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
1001 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
1002 "not mapped!", NULL);
1003 spin_lock(&dev->gnd_map_lock);
1005 switch (tx->tx_buftype) {
1007 GNIDBG_TX(D_EMERG, tx,
1008 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
1009 spin_unlock(&dev->gnd_map_lock);
1013 case GNILND_BUF_PHYS_UNMAPPED:
1014 bytes = tx->tx_phys_npages * PAGE_SIZE;
1015 dev->gnd_map_nphys--;
1016 dev->gnd_map_physnop -= tx->tx_phys_npages;
1019 case GNILND_BUF_VIRT_UNMAPPED:
1021 dev->gnd_map_nvirt--;
1022 dev->gnd_map_virtnob -= tx->tx_nob;
1026 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1027 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1028 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
1029 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
1030 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
1031 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
1032 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1035 atomic_dec(&dev->gnd_n_mdd);
1036 atomic64_sub(bytes, &dev->gnd_nbytes_map);
1038 /* we only get here in the valid cases */
1039 list_del_init(&tx->tx_map_list);
1040 dev->gnd_map_version++;
1041 spin_unlock(&dev->gnd_map_lock);
1045 kgnilnd_map_buffer(kgn_tx_t *tx)
1047 kgn_conn_t *conn = tx->tx_conn;
1048 kgn_device_t *dev = conn->gnc_device;
1049 __u32 flags = GNI_MEM_READWRITE;
1052 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
1053 * be called concurrently as there are internal locks that protect
1054 * any data structures or HW resources. We just need to ensure
1055 * that our concurrency doesn't result in the kgn_device_t
1056 * getting nuked while we are in here */
1058 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
1059 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
1061 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1064 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1065 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1068 switch (tx->tx_buftype) {
1072 case GNILND_BUF_NONE:
1073 case GNILND_BUF_IMMEDIATE:
1074 case GNILND_BUF_IMMEDIATE_KIOV:
1075 case GNILND_BUF_PHYS_MAPPED:
1076 case GNILND_BUF_VIRT_MAPPED:
1079 case GNILND_BUF_PHYS_UNMAPPED:
1080 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1081 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1082 tx->tx_phys, tx->tx_phys_npages, NULL,
1083 GNI_MEM_PHYS_SEGMENTS | flags,
1085 /* could race with other uses of the map counts, but this is ok
1086 * - this needs to turn into a non-fatal error soon to allow
1087 * GART resource, etc starvation handling */
1088 if (rrc != GNI_RC_SUCCESS) {
1089 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1090 "phys %u pp %u, virt %u nob %llu",
1091 tx->tx_phys_npages, dev->gnd_id,
1092 dev->gnd_map_nphys, dev->gnd_map_physnop,
1093 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1094 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1097 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1098 kgnilnd_mem_add_map_list(dev, tx);
1101 case GNILND_BUF_VIRT_UNMAPPED:
1102 rrc = kgnilnd_mem_register(dev->gnd_handle,
1103 (__u64)tx->tx_buffer, tx->tx_nob,
1104 NULL, flags, &tx->tx_map_key);
1105 if (rrc != GNI_RC_SUCCESS) {
1106 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1107 "phys %u pp %u, virt %u nob %llu",
1108 tx->tx_nob, dev->gnd_id,
1109 dev->gnd_map_nphys, dev->gnd_map_physnop,
1110 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1111 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1114 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1115 kgnilnd_mem_add_map_list(dev, tx);
1116 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1117 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1118 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1119 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1120 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1128 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1130 kgn_conn_t *conn = tx->tx_conn;
1131 kgn_mdd_purgatory_t *gmp;
1133 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1134 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1135 " asserting to avoid data corruption\n");
1136 if (tx->tx_buffer_copy)
1137 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1139 gmp->gmp_map_key = tx->tx_map_key;
1141 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1143 /* ensure that we don't have a blank purgatory - indicating the
1144 * conn is not already on purgatory lists - we'd never recover these
1145 * MDD if that were the case */
1146 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1147 "conn 0x%p->%s with NULL purgatory",
1148 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1150 /* link 'er up! - only place we really need to lock for
1151 * concurrent access */
1152 spin_lock(&conn->gnc_list_lock);
1153 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1154 spin_unlock(&conn->gnc_list_lock);
1158 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1162 int hold_timeout = 0;
1164 /* code below relies on +1 relationship ... */
1165 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1166 (GNILND_BUF_PHYS_UNMAPPED + 1));
1167 BUILD_BUG_ON(GNILND_BUF_VIRT_MAPPED !=
1168 (GNILND_BUF_VIRT_UNMAPPED + 1));
1170 switch (tx->tx_buftype) {
1174 case GNILND_BUF_NONE:
1175 case GNILND_BUF_IMMEDIATE:
1176 case GNILND_BUF_PHYS_UNMAPPED:
1177 case GNILND_BUF_VIRT_UNMAPPED:
1179 case GNILND_BUF_IMMEDIATE_KIOV:
1180 if (tx->tx_phys != NULL) {
1181 vunmap(tx->tx_phys);
1182 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1183 kunmap(tx->tx_imm_pages[0]);
1185 /* clear to prevent kgnilnd_free_tx from thinking
1186 * this is a RDMA descriptor */
1190 case GNILND_BUF_PHYS_MAPPED:
1191 case GNILND_BUF_VIRT_MAPPED:
1192 LASSERT(tx->tx_conn != NULL);
1194 dev = tx->tx_conn->gnc_device;
1196 /* only want to hold if we are closing conn without
1197 * verified peer notification - the theory is that
1198 * a TX error can be communicated in all other cases */
1199 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1200 error != -GNILND_NOPURG &&
1201 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1202 kgnilnd_add_purgatory_tx(tx);
1204 /* The timeout we give to kgni is a deadman stop only.
1205 * we are setting high to ensure we don't have the kgni timer
1206 * fire before ours fires _and_ is handled */
1207 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1209 GNIDBG_TX(D_NET, tx,
1210 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1211 tx->tx_conn->gnc_device, hold_timeout,
1212 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1214 if (tx->tx_buffer_copy != NULL) {
1215 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1216 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1217 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1218 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1220 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1221 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1225 kgnilnd_mem_del_map_list(dev, tx);
1231 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1233 struct lnet_msg *lntmsg0, *lntmsg1;
1234 int status0, status1;
1235 struct lnet_ni *ni = NULL;
1236 kgn_conn_t *conn = tx->tx_conn;
1238 LASSERT(!in_interrupt());
1240 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1241 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1244 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1245 !kgnilnd_conn_clean_errno(completion)) {
1246 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1247 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1248 completion, tx, conn ?
1249 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1250 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1251 kgnilnd_tx_state2str(tx->tx_list_state),
1252 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1255 /* The error codes determine if we hold onto the MDD */
1256 kgnilnd_unmap_buffer(tx, completion);
1258 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1259 * we play nice with the error codes to avoid delivering a failed
1260 * REQUEST and then a REPLY event as well */
1262 /* return -EIO to lnet - it is the magic value for failed sends */
1263 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1265 status1 = completion;
1267 status0 = status1 = completion;
1270 tx->tx_buftype = GNILND_BUF_NONE;
1271 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1273 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1274 * set NULL when we are a tx without a conn */
1276 ni = conn->gnc_peer->gnp_net->gnn_ni;
1278 spin_lock(&conn->gnc_tx_lock);
1280 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1281 (volatile unsigned long *)&conn->gnc_tx_bits),
1282 "conn %p tx %p bit %d already cleared\n",
1283 conn, tx, tx->tx_id.txe_idx);
1285 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1286 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1288 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1289 spin_unlock(&conn->gnc_tx_lock);
1292 kgnilnd_free_tx(tx);
1294 /* finalize AFTER freeing lnet msgs */
1296 /* warning - we should hold no locks here - calling lnet_finalize
1297 * could free up lnet credits, resulting in a call chain back into
1298 * the LND via kgnilnd_send and friends */
1300 lnet_finalize(lntmsg0, status0);
1302 if (lntmsg1 != NULL) {
1303 lnet_finalize(lntmsg1, status1);
1308 kgnilnd_txlist_done(struct list_head *txlist, int error)
1311 int err_printed = 0;
1313 if (list_empty(txlist))
1316 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1317 /* only print the first error */
1319 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1320 list_del_init(&tx->tx_list);
1321 kgnilnd_tx_done(tx, error);
1326 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1330 spin_lock(&conn->gnc_tx_lock);
1332 /* ID zero is NOT ALLOWED!!! */
1335 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1336 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1337 if (id == GNILND_MAX_MSG_ID) {
1338 if (conn->gnc_next_tx != 1) {
1339 /* we only searched from next_tx to end and didn't find
1340 * one, so search again from start */
1341 conn->gnc_next_tx = 1;
1344 /* couldn't find one! */
1345 spin_unlock(&conn->gnc_tx_lock);
1349 /* bump next_tx to prevent immediate reuse */
1350 conn->gnc_next_tx = id + 1;
1352 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1353 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1354 "tx 0x%p already at id %d\n",
1355 conn->gnc_tx_ref_table[id], id);
1357 /* delay these until we have a valid ID - prevents bad clear of the bit
1358 * in kgnilnd_tx_done */
1360 tx->tx_id.txe_cqid = conn->gnc_cqid;
1362 tx->tx_id.txe_idx = id;
1363 conn->gnc_tx_ref_table[id] = tx;
1365 /* Using jiffies to help differentiate against TX reuse - with
1366 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1367 * if we are sending to the same node faster than 256000/sec.
1368 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1370 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1372 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1374 spin_unlock(&conn->gnc_tx_lock);
1379 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1383 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1385 /* we don't care about TX timeout - it could be that the network is slower
1386 * or throttled. We'll keep retranmitting - so if the network is so slow
1387 * that we fill up our mailbox, we'll keep trying to resend that msg
1388 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1389 * that he hasn't send us any traffic in return */
1391 /* some reasonable throttling of the debug message */
1393 unsigned long now = jiffies;
1394 /* XXX Nic: Mystical TX debug here... */
1395 /* We expect retransmissions so only log when D_NET is enabled */
1396 GNIDBG_SMSG_CREDS(D_NET, conn);
1397 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1398 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1399 " last_msg %uus/%uus last_cq %uus/%uus",
1400 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1401 tx->tx_id, tx->tx_retrans,
1402 jiffies_to_usecs(now - tx->tx_cred_wait),
1403 jiffies_to_usecs(now - conn->gnc_last_tx),
1404 jiffies_to_usecs(now - conn->gnc_last_rx),
1405 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1406 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1410 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1411 * to avoid bad ordering with state_lock */
1414 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1415 spinlock_t *state_lock, kgn_tx_list_state_t state)
1417 kgn_conn_t *conn = tx->tx_conn;
1418 kgn_msg_t *msg = &tx->tx_msg;
1420 unsigned long newest_last_rx, timeout;
1423 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1424 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1426 "msg 0x%p type %d wrong payload size %d\n",
1427 msg, msg->gnm_type, immediatenob);
1429 /* make sure we catch all the cases where we'd send on a dirty old mbox
1430 * but allow case for sending CLOSE. Since this check is within the CQ
1431 * mutex barrier and the close message is only sent through
1432 * kgnilnd_send_conn_close the last message out the door will be the
1435 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1436 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1437 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1438 /* Return -ETIME, we are closing the connection already so we dont want to
1439 * have this tx hit the wire. The tx will be killed by the calling function.
1440 * Once the EP is marked dirty the close message will be the last
1441 * thing to hit the wire */
1446 timeout = cfs_time_seconds(conn->gnc_timeout);
1448 newest_last_rx = GNILND_LASTRX(conn);
1450 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1451 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1454 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1455 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1456 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1457 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1458 cfs_duration_sec(now - newest_last_rx),
1459 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1460 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1461 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1465 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1466 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1467 * creating the message this allows the message to contain the correct LNET NID/NET needed
1468 * instead of the one that the peer/conn uses for sending the data.
1470 msg->gnm_connstamp = conn->gnc_my_connstamp;
1471 msg->gnm_payload_len = immediatenob;
1472 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1474 /* always init here - kgn_checksum is a /sys module tunable
1475 * and can be flipped at any point, even between msg init and sending */
1477 if (*kgnilnd_tunables.kgn_checksum) {
1478 /* We must set here and not in kgnilnd_init_msg,
1479 * we could resend this msg many times
1480 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1481 * through init_msg again */
1482 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1483 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1484 msg->gnm_cksum += 0xf00f;
1488 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1489 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1490 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1491 tx->tx_id.txe_idx, immediate, immediatenob);
1493 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1494 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1496 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1497 msg, sizeof(*msg), immediate,
1499 tx->tx_id.txe_smsg_id);
1503 case GNI_RC_SUCCESS:
1504 atomic_inc(&conn->gnc_tx_seq);
1505 conn->gnc_last_tx = jiffies;
1506 /* no locking here as LIVE isn't a list */
1507 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1509 /* this needs to be checked under lock as it might be freed from a completion
1512 if (msg->gnm_type == GNILND_MSG_NOOP) {
1513 set_mb(conn->gnc_last_noop_sent, jiffies);
1516 /* serialize with seeing CQ events for completion on this, as well as
1518 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1519 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1521 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1522 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1523 kgnilnd_peer_alive(conn->gnc_peer);
1524 GNIDBG_SMSG_CREDS(D_NET, conn);
1527 case GNI_RC_NOT_DONE:
1528 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1529 * 1: out of mbox credits
1530 * 2: out of mbox payload credits
1531 * 3: On Aries out of dla credits
1533 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1534 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1535 /* We'll handle this error inline - makes the calling logic much more
1538 /* If no lock, caller doesn't want us to retry */
1539 if (state_lock == NULL) {
1543 /* I need kgni credits to send this. Replace tx at the head of the
1544 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1545 * and bump retrans counts since we are requeueing the tx.
1549 conn->gnc_tx_retrans++;
1551 kgnilnd_tx_log_retrans(conn, tx);
1552 /* add to head of list for the state and retries */
1553 spin_lock(state_lock);
1554 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1555 spin_unlock(state_lock);
1557 /* We only reschedule for a certain number of retries, then
1558 * we will wait for the CQ events indicating a release of SMSG
1560 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1561 kgnilnd_schedule_conn(conn);
1564 /* CQ event coming in signifies either TX completed or
1565 * RX receive. Either of these *could* free up credits
1566 * in the SMSG mbox and we should try sending again */
1567 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1568 tx->tx_conn->gnc_cqid);
1569 kgnilnd_schedule_delay_conn(conn);
1570 /* use +ve return code to let upper layers know they
1571 * should stop looping on sends */
1575 /* handle bad retcode gracefully */
1576 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1577 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1582 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1584 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1585 spinlock_t *state_lock, kgn_tx_list_state_t state)
1587 kgn_device_t *dev = tx->tx_conn->gnc_device;
1588 unsigned long timestamp;
1591 timestamp = jiffies;
1592 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1593 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1594 /* delay in jiffies - we are really concerned only with things that
1595 * result in a schedule() or really holding this off for long times .
1596 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1597 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1599 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1605 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1607 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1608 spinlock_t *state_lock, kgn_tx_list_state_t state)
1610 kgn_conn_t *conn = tx->tx_conn;
1611 kgn_device_t *dev = conn->gnc_device;
1612 unsigned long timestamp;
1615 timestamp = jiffies;
1617 /* technically we are doing bad things with the read_lock on the peer_conn
1618 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1619 * for the mutex. I bet lockdep is gonna flag this one though... */
1621 /* there are a few cases where we don't want the immediate send - like
1622 * when we are in the scheduler thread and it'd harm the latency of
1623 * getting messages up to LNet */
1625 /* rmb for gnd_ready */
1627 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1629 atomic_inc(&conn->gnc_device->gnd_fast_block);
1630 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1631 /* dont hit HW during quiesce */
1633 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1634 /* dont hit HW if stale EPs and conns left to close */
1637 atomic_inc(&conn->gnc_device->gnd_fast_try);
1638 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1639 &conn->gnc_smsg_mutex);
1644 /* we got the mutex and weren't blocked */
1646 /* delay in jiffies - we are really concerned only with things that
1647 * result in a schedule() or really holding this off for long times .
1648 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1649 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1651 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1652 tx->tx_qtime = jiffies;
1653 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1654 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1655 /* _nolock unlocks the mutex for us */
1661 /* lets us know if we can push this RDMA through now */
1663 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1667 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1669 if (bytes_left < 0) {
1670 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1671 atomic_inc(&dev->gnd_rdmaq_nstalls);
1674 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1675 dev->gnd_rdmaq_deadline);
1676 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1677 /* we never del this timer - at worst it schedules us.. */
1684 /* this adds a TX to the queue pending throttling authorization before
1685 * we allow our remote peer to launch a PUT at us */
1687 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1691 /* we cannot go into send_mapped_tx from here as we are holding locks
1692 * and mem registration might end up allocating memory in kgni.
1693 * That said, we'll push this as far as we can into the queue process */
1694 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1697 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1698 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1699 /* lets us know how delayed RDMA is */
1700 tx->tx_qtime = jiffies;
1701 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1703 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1704 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1705 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1706 /* lets us know how delayed mapping is */
1707 tx->tx_qtime = jiffies;
1708 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1711 /* make sure we wake up sched to run this */
1712 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1715 /* push TX through state machine */
1717 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1722 /* set the tx_id here, we delay it until we have an actual conn
1724 * in some cases, the tx_id is already set to provide for things
1725 * like RDMA completion cookies, etc */
1726 if (tx->tx_id.txe_idx == 0) {
1727 rc = kgnilnd_set_tx_id(tx, conn);
1729 kgnilnd_tx_done(tx, rc);
1734 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1735 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1737 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1739 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1740 kgnilnd_tx_done(tx, rc);
1744 switch (tx->tx_msg.gnm_type) {
1745 case GNILND_MSG_PUT_ACK:
1746 case GNILND_MSG_GET_REQ:
1747 case GNILND_MSG_PUT_REQ_REV:
1748 case GNILND_MSG_GET_ACK_REV:
1749 /* hijacking time! If this messages will authorize our peer to
1750 * send his dirty little bytes in an RDMA, we need to get permission */
1751 kgnilnd_queue_rdma(conn, tx);
1753 case GNILND_MSG_IMMEDIATE:
1754 /* try to send right now, can help reduce latency */
1755 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1758 /* it was sent, break out of switch to avoid default case of queueing */
1761 /* needs to queue to try again, so fall through to default case */
1762 case GNILND_MSG_NOOP:
1763 /* Just make sure this goes out first for this conn */
1765 /* fall through... */
1767 spin_lock(&conn->gnc_list_lock);
1768 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1769 tx->tx_qtime = jiffies;
1770 spin_unlock(&conn->gnc_list_lock);
1771 kgnilnd_schedule_conn(conn);
1776 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1779 kgn_peer_t *new_peer = NULL;
1780 kgn_conn_t *conn = NULL;
1786 /* If I get here, I've committed to send, so I complete the tx with
1787 * failure on any problems */
1789 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1790 "tx already has connection %p", tx->tx_conn);
1792 /* do all of the peer & conn searching in one swoop - this avoids
1793 * nastiness when dropping locks and needing to maintain a sane state
1794 * in the face of stack reset or something else nuking peers & conns */
1796 /* I expect to find him, so only take a read lock */
1797 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1799 peer = kgnilnd_find_peer_locked(target->nid);
1801 conn = kgnilnd_find_conn_locked(peer);
1802 /* this could be NULL during quiesce */
1804 /* Connection exists; queue message on it */
1805 kgnilnd_queue_tx(conn, tx);
1806 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1810 /* don't create a connection if the peer is marked down */
1811 if (peer->gnp_state != GNILND_PEER_UP) {
1812 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1818 /* creating peer or conn; I'll need a write lock... */
1819 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1821 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1823 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1825 /* NB - this will not block during normal operations -
1826 * the only writer of this is in the startup/shutdown path. */
1827 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1833 /* ignore previous peer entirely - we cycled the lock, so we
1834 * will create new peer and at worst drop it if peer is still
1836 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1838 up_read(&kgnilnd_data.kgn_net_rw_sem);
1842 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1843 up_read(&kgnilnd_data.kgn_net_rw_sem);
1845 /* search for peer again now that we have the lock
1846 * if we don't find it, add our new one to the list */
1847 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1849 /* don't create a connection if the peer is not up */
1850 if (peer->gnp_state != GNILND_PEER_UP) {
1851 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1856 conn = kgnilnd_find_or_create_conn_locked(peer);
1858 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1859 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1864 /* oh hey, found a conn now... magical */
1865 kgnilnd_queue_tx(conn, tx);
1867 /* no conn, must be trying to connect - so we queue for now */
1868 tx->tx_qtime = jiffies;
1869 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1871 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1874 kgnilnd_tx_done(tx, rc);
1879 kgnilnd_rdma(kgn_tx_t *tx, int type,
1880 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1882 kgn_conn_t *conn = tx->tx_conn;
1883 unsigned long timestamp;
1884 gni_post_type_t post_type;
1887 unsigned int desc_nob = nob;
1888 void *desc_buffer = tx->tx_buffer;
1889 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1890 LASSERTF(kgnilnd_tx_mapped(tx),
1891 "unmapped tx %p\n", tx);
1892 LASSERTF(conn != NULL,
1893 "NULL conn on tx %p, naughty, naughty\n", tx);
1894 LASSERTF(nob <= sink->gnrd_nob,
1895 "nob %u > sink->gnrd_nob %d (%p)\n",
1896 nob, sink->gnrd_nob, sink);
1897 LASSERTF(nob <= tx->tx_nob,
1898 "nob %d > tx(%p)->tx_nob %d\n",
1899 nob, tx, tx->tx_nob);
1902 case GNILND_MSG_GET_DONE:
1903 case GNILND_MSG_PUT_DONE:
1904 post_type = GNI_POST_RDMA_PUT;
1906 case GNILND_MSG_GET_DONE_REV:
1907 case GNILND_MSG_PUT_DONE_REV:
1908 post_type = GNI_POST_RDMA_GET;
1911 CERROR("invalid msg type %s (%d)\n",
1912 kgnilnd_msgtype2str(type), type);
1915 if (post_type == GNI_POST_RDMA_GET) {
1916 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1917 * aligned. If the local buffer is not aligned correctly using the copy buffer
1918 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1919 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1920 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1921 * making sure the length and addr are aligned and copying the data into the correct location
1922 * once the transfer has completed.
1924 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1925 (sink->gnrd_addr & 3) ||
1928 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1930 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1932 if ((nob + tx->tx_offset) & 3) {
1933 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1934 atomic_inc(&kgnilnd_data.kgn_rev_length);
1936 desc_nob = (nob + tx->tx_offset);
1939 if (tx->tx_buffer_copy == NULL) {
1940 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1941 * and require at most we allocate a few extra bytes. */
1942 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1944 if (!tx->tx_buffer_copy) {
1945 /* allocation of buffer failed nak the rdma */
1946 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1947 kgnilnd_tx_done(tx, -EFAULT);
1950 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1951 rc = kgnilnd_mem_register(conn->gnc_device->gnd_handle, (__u64)tx->tx_buffer_copy, desc_nob, NULL, GNI_MEM_READWRITE, &tx->tx_buffer_copy_map_key);
1952 if (rc != GNI_RC_SUCCESS) {
1953 /* Registration Failed nak rdma and kill the tx. */
1954 kgnilnd_vfree(tx->tx_buffer_copy,
1956 tx->tx_buffer_copy = NULL;
1957 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1958 kgnilnd_tx_done(tx, -EFAULT);
1962 desc_map_key = tx->tx_buffer_copy_map_key;
1963 desc_buffer = tx->tx_buffer_copy;
1967 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1968 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1969 tx->tx_rdma_desc.type = post_type;
1970 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1971 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1972 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1973 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1974 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1975 tx->tx_rdma_desc.length = desc_nob;
1976 tx->tx_nob_rdma = nob;
1977 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1978 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1979 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1980 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1981 /* prep final completion message */
1982 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1983 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1984 /* send actual size RDMA'd in retval */
1985 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1987 kgnilnd_compute_rdma_cksum(tx, nob);
1990 kgnilnd_queue_tx(conn, tx);
1994 /* Don't lie (CLOSE == RDMA idle) */
1995 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1996 tx, conn, conn->gnc_close_sent);
1998 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1999 "0x%x cookie:%#llx",
2000 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
2002 /* set CQ dedicated for RDMA */
2003 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
2005 timestamp = jiffies;
2006 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
2007 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2008 /* delay in jiffies - we are really concerned only with things that
2009 * result in a schedule() or really holding this off for long times .
2010 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2011 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2013 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2015 if (rrc == GNI_RC_ERROR_RESOURCE) {
2016 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2017 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2018 kgnilnd_unmap_buffer(tx, 0);
2020 if (tx->tx_buffer_copy != NULL) {
2021 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
2022 tx->tx_buffer_copy = NULL;
2025 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2026 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2028 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2029 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2033 spin_lock(&conn->gnc_list_lock);
2034 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2035 tx->tx_qtime = jiffies;
2036 spin_unlock(&conn->gnc_list_lock);
2037 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2038 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2040 /* XXX Nic: is this a place we should handle more errors for
2041 * robustness sake */
2042 LASSERT(rrc == GNI_RC_SUCCESS);
2047 kgnilnd_alloc_rx(void)
2051 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2053 CERROR("failed to allocate rx\n");
2056 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2059 /* no memset to zero, we'll always fill all members */
2063 /* release is to just free connection resources
2064 * we use this for the eager path after copying */
2066 kgnilnd_release_msg(kgn_conn_t *conn)
2069 unsigned long timestamp;
2071 CDEBUG(D_NET, "consuming %p\n", conn);
2073 timestamp = jiffies;
2074 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2075 /* delay in jiffies - we are really concerned only with things that
2076 * result in a schedule() or really holding this off for long times .
2077 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2078 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2080 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2081 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2083 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2084 GNIDBG_SMSG_CREDS(D_NET, conn);
2086 kgnilnd_schedule_conn(conn);
2090 kgnilnd_consume_rx(kgn_rx_t *rx)
2092 kgn_conn_t *conn = rx->grx_conn;
2093 kgn_msg_t *rxmsg = rx->grx_msg;
2095 /* if we are eager, free the cache alloc'd msg */
2096 if (unlikely(rx->grx_eager)) {
2097 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2098 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2100 /* release ref from eager_recv */
2101 kgnilnd_conn_decref(conn);
2103 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2104 kgnilnd_release_msg(conn);
2107 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2108 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2113 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2115 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2116 int type = lntmsg->msg_type;
2117 struct lnet_process_id target = lntmsg->msg_target;
2118 int target_is_router = lntmsg->msg_target_is_router;
2119 int routing = lntmsg->msg_routing;
2120 unsigned int niov = lntmsg->msg_niov;
2121 struct bio_vec *kiov = lntmsg->msg_kiov;
2122 unsigned int offset = lntmsg->msg_offset;
2123 unsigned int nob = lntmsg->msg_len;
2124 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2125 kgn_net_t *net = ni->ni_data;
2128 /* '1' for consistency with code that checks !mpflag to restore */
2129 unsigned int mpflag = 1;
2130 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2132 /* NB 'private' is different depending on what we're sending.... */
2133 LASSERT(!in_interrupt());
2135 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2136 type, nob, niov, libcfs_id2str(target));
2138 LASSERTF(nob == 0 || niov > 0,
2139 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2140 LASSERTF(niov <= LNET_MAX_IOV,
2141 "lntmsg %p niov %d\n", lntmsg, niov);
2144 mpflag = memalloc_noreclaim_save();
2148 CERROR("lntmsg %p with unexpected type %d\n",
2153 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2161 if (routing || target_is_router)
2162 break; /* send IMMEDIATE */
2164 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2165 * check the eventual sink buffer here - if small enough, remote
2166 * end is perfectly capable of returning data in short message -
2167 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2168 * for IMMEDIATE messages which will have it send a real reply instead
2169 * of doing kgnilnd_recv to have the RDMA continued */
2170 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2173 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2174 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2176 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2182 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2183 NULL, lntmsg->msg_md->md_kiov,
2184 0, lntmsg->msg_md->md_length);
2186 CERROR("unable to setup buffer: %d\n", rc);
2187 kgnilnd_tx_done(tx, rc);
2192 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2193 if (tx->tx_lntmsg[1] == NULL) {
2194 CERROR("Can't create reply for GET to %s\n",
2195 libcfs_nid2str(target.nid));
2196 kgnilnd_tx_done(tx, rc);
2201 tx->tx_lntmsg[0] = lntmsg;
2202 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2203 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2205 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2207 /* rest of tx_msg is setup just before it is sent */
2208 kgnilnd_launch_tx(tx, net, &target);
2210 case LNET_MSG_REPLY:
2212 /* to save on MDDs, we'll handle short kiov by vmap'ing
2213 * and sending via SMSG */
2214 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2217 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2218 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2220 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2227 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL,
2230 kgnilnd_tx_done(tx, rc);
2235 tx->tx_lntmsg[0] = lntmsg;
2236 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2237 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2239 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2241 /* rest of tx_msg is setup just before it is sent */
2242 kgnilnd_launch_tx(tx, net, &target);
2246 /* send IMMEDIATE */
2248 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2249 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2251 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2257 rc = kgnilnd_setup_immediate_buffer(tx, niov, NULL, kiov, offset, nob);
2259 kgnilnd_tx_done(tx, rc);
2263 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2264 tx->tx_lntmsg[0] = lntmsg;
2265 kgnilnd_launch_tx(tx, net, &target);
2268 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2270 memalloc_noreclaim_restore(mpflag);
2275 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2277 kgn_conn_t *conn = rx->grx_conn;
2278 kgn_msg_t *rxmsg = rx->grx_msg;
2279 unsigned int niov = lntmsg->msg_niov;
2280 struct bio_vec *kiov = lntmsg->msg_kiov;
2281 unsigned int offset = lntmsg->msg_offset;
2282 unsigned int nob = lntmsg->msg_len;
2287 switch (rxmsg->gnm_type) {
2288 case GNILND_MSG_PUT_REQ_REV:
2289 done_type = GNILND_MSG_PUT_DONE_REV;
2292 case GNILND_MSG_GET_REQ:
2293 done_type = GNILND_MSG_GET_DONE;
2296 CERROR("invalid msg type %s (%d)\n",
2297 kgnilnd_msgtype2str(rxmsg->gnm_type),
2302 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2306 rc = kgnilnd_set_tx_id(tx, conn);
2310 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL, kiov, offset, nob);
2314 tx->tx_lntmsg[0] = lntmsg;
2315 tx->tx_getinfo = rxmsg->gnm_u.get;
2317 /* we only queue from kgnilnd_recv - we might get called from other contexts
2318 * and we don't want to block the mutex in those cases */
2320 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2321 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2322 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2323 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2328 kgnilnd_tx_done(tx, rc);
2329 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2331 lnet_finalize(lntmsg, rc);
2335 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2338 kgn_rx_t *rx = private;
2339 kgn_conn_t *conn = rx->grx_conn;
2340 kgn_msg_t *rxmsg = rx->grx_msg;
2341 kgn_msg_t *eagermsg = NULL;
2342 kgn_peer_t *peer = NULL;
2343 kgn_conn_t *found_conn = NULL;
2345 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2346 conn, rxmsg, lntmsg);
2348 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2349 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2350 rxmsg->gnm_payload_len);
2353 /* Grab a read lock so the connection doesnt disappear on us
2354 * while we look it up
2356 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2358 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2360 found_conn = kgnilnd_find_conn_locked(peer);
2363 /* Verify the connection found is the same one that the message
2364 * is supposed to be using, if it is not output an error message
2367 if (!peer || !found_conn
2368 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2369 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2370 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2371 peer, conn, found_conn);
2373 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2374 " from %s", rxmsg->gnm_connstamp,
2375 found_conn->gnc_peer_connstamp,
2376 libcfs_nid2str(peer->gnp_nid));
2381 /* add conn ref to ensure it doesn't go away until all eager
2382 * messages processed */
2383 kgnilnd_conn_addref(conn);
2385 /* Now that we have verified the connection is valid and added a
2386 * reference we can remove the read_lock on the peer_conn_lock */
2387 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2389 /* we have no credits or buffers for this message, so copy it
2390 * somewhere for a later kgnilnd_recv */
2391 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2392 *kgnilnd_tunables.kgn_eager_credits) {
2393 CERROR("Out of eager credits to %s\n",
2394 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2398 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2400 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2401 if (eagermsg == NULL) {
2402 kgnilnd_conn_decref(conn);
2403 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2404 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2408 /* copy msg and payload */
2409 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2410 rx->grx_msg = eagermsg;
2413 /* stash this for lnet_finalize on cancel-on-conn-close */
2414 rx->grx_lntmsg = lntmsg;
2416 /* keep the same rx_t, it just has a new grx_msg now */
2417 *new_private = private;
2419 /* release SMSG buffer */
2420 kgnilnd_release_msg(conn);
2426 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2427 int delayed, unsigned int niov,
2428 struct bio_vec *kiov,
2429 unsigned int offset, unsigned int mlen, unsigned int rlen)
2431 kgn_rx_t *rx = private;
2432 kgn_conn_t *conn = rx->grx_conn;
2433 kgn_msg_t *rxmsg = rx->grx_msg;
2439 LASSERT(!in_interrupt());
2440 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2442 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2443 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2444 conn, rxmsg, lntmsg,
2445 niov, kiov, 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 offset=%d mlen=%d rlen=%d",
2463 conn, rx, rxmsg, lntmsg, niov, kiov, 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);
2516 lnet_copy_flat2kiov(
2518 *kgnilnd_tunables.kgn_max_immediate,
2519 &rxmsg[1], 0, mlen);
2521 kgnilnd_consume_rx(rx);
2522 lnet_finalize(lntmsg, 0);
2525 case GNILND_MSG_PUT_REQ:
2526 /* LNET wants to truncate or drop transaction, sending NAK */
2528 kgnilnd_consume_rx(rx);
2529 lnet_finalize(lntmsg, 0);
2531 /* only error if lntmsg == NULL, otherwise we are just
2532 * short circuiting the rdma process of 0 bytes */
2533 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2534 lntmsg == NULL ? -ENOENT : 0,
2535 rxmsg->gnm_u.get.gngm_cookie,
2539 /* sending ACK with sink buff. info */
2540 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2542 kgnilnd_consume_rx(rx);
2546 rc = kgnilnd_set_tx_id(tx, conn);
2548 GOTO(nak_put_req, rc);
2551 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL,
2552 kiov, offset, mlen);
2554 GOTO(nak_put_req, rc);
2557 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2558 rxmsg->gnm_u.putreq.gnprm_cookie;
2559 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2560 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2561 (__u64)((unsigned long)tx->tx_buffer);
2562 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2564 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2565 tx->tx_qtime = jiffies;
2566 /* we only queue from kgnilnd_recv - we might get called from other contexts
2567 * and we don't want to block the mutex in those cases */
2569 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2570 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2571 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2572 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2574 kgnilnd_consume_rx(rx);
2578 /* make sure we send an error back when the PUT fails */
2579 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2580 kgnilnd_tx_done(tx, rc);
2581 kgnilnd_consume_rx(rx);
2583 /* return magic LNet network error */
2585 case GNILND_MSG_GET_REQ_REV:
2586 /* LNET wants to truncate or drop transaction, sending NAK */
2588 kgnilnd_consume_rx(rx);
2589 lnet_finalize(lntmsg, 0);
2591 /* only error if lntmsg == NULL, otherwise we are just
2592 * short circuiting the rdma process of 0 bytes */
2593 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2594 lntmsg == NULL ? -ENOENT : 0,
2595 rxmsg->gnm_u.get.gngm_cookie,
2599 /* lntmsg can be null when parsing a LNET_GET */
2600 if (lntmsg != NULL) {
2601 /* sending ACK with sink buff. info */
2602 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2604 kgnilnd_consume_rx(rx);
2608 rc = kgnilnd_set_tx_id(tx, conn);
2610 GOTO(nak_get_req_rev, rc);
2613 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL,
2614 kiov, offset, mlen);
2616 GOTO(nak_get_req_rev, rc);
2619 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2620 rxmsg->gnm_u.putreq.gnprm_cookie;
2621 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2622 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2623 (__u64)((unsigned long)tx->tx_buffer);
2624 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2626 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2628 /* we only queue from kgnilnd_recv - we might get called from other contexts
2629 * and we don't want to block the mutex in those cases */
2631 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2632 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2633 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2634 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2637 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2639 rxmsg->gnm_u.get.gngm_cookie,
2643 kgnilnd_consume_rx(rx);
2647 /* make sure we send an error back when the GET fails */
2648 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2649 kgnilnd_tx_done(tx, rc);
2650 kgnilnd_consume_rx(rx);
2652 /* return magic LNet network error */
2656 case GNILND_MSG_PUT_REQ_REV:
2657 /* LNET wants to truncate or drop transaction, sending NAK */
2659 kgnilnd_consume_rx(rx);
2660 lnet_finalize(lntmsg, 0);
2662 /* only error if lntmsg == NULL, otherwise we are just
2663 * short circuiting the rdma process of 0 bytes */
2664 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2665 lntmsg == NULL ? -ENOENT : 0,
2666 rxmsg->gnm_u.get.gngm_cookie,
2671 if (lntmsg != NULL) {
2673 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2676 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2678 rxmsg->gnm_u.get.gngm_cookie,
2681 kgnilnd_consume_rx(rx);
2683 case GNILND_MSG_GET_REQ:
2684 if (lntmsg != NULL) {
2686 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2689 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2691 rxmsg->gnm_u.get.gngm_cookie,
2694 kgnilnd_consume_rx(rx);
2700 /* needs write_lock on kgn_peer_conn_lock held */
2702 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2704 unsigned long timeout, keepalive;
2705 unsigned long now = jiffies;
2706 unsigned long newest_last_rx;
2709 /* given that we found this conn hanging off a peer, it better damned
2710 * well be connected */
2711 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2712 "conn 0x%p->%s with bad state%s\n", conn,
2713 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2715 kgnilnd_conn_state2str(conn));
2717 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2718 "rx_diff %lu tx_diff %lu\n",
2719 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2720 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2721 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2722 cfs_duration_sec(now - conn->gnc_last_tx));
2724 timeout = cfs_time_seconds(conn->gnc_timeout);
2725 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2727 /* just in case our lack of RX msg processing is gumming up the works - give the
2728 * remove an extra chance */
2730 newest_last_rx = GNILND_LASTRX(conn);
2732 if (time_after_eq(now, newest_last_rx + timeout)) {
2733 uint32_t level = D_CONSOLE|D_NETERROR;
2735 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2738 GNIDBG_CONN(level, conn,
2739 "No gnilnd traffic received from %s for %lu "
2740 "seconds, terminating connection. Is node down? ",
2741 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2742 cfs_duration_sec(now - newest_last_rx));
2746 /* we don't timeout on last_tx stalls - we are going to trust the
2747 * underlying network to let us know when sends are failing.
2748 * At worst, the peer will timeout our RX stamp and drop the connection
2749 * at that point. We'll then see his CLOSE or at worst his RX
2750 * stamp stop and drop the connection on our end */
2752 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2753 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2754 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2755 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2756 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2758 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2759 conn->gnc_last_noop_cq,
2760 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2761 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2762 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2763 set_mb(conn->gnc_last_noop_want, jiffies);
2764 atomic_inc(&conn->gnc_reaper_noop);
2765 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2768 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2771 kgnilnd_queue_tx(conn, tx);
2777 /* needs write_lock on kgn_peer_conn_lock held */
2779 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2780 struct list_head *souls)
2782 unsigned long timeout;
2783 kgn_conn_t *conn, *connN = NULL;
2789 short releaseconn = 0;
2790 unsigned long first_rx = 0;
2791 int purgatory_conn_cnt = 0;
2793 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2794 peer, libcfs_nid2str(peer->gnp_nid),
2795 peer->gnp_reconnect_interval);
2797 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2798 GNILND_MIN_TIMEOUT));
2800 conn = kgnilnd_find_conn_locked(peer);
2802 /* if there is a valid conn, check the queues for timeouts */
2803 rc = kgnilnd_check_conn_timeouts_locked(conn);
2805 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2806 /* simulate a RX CLOSE after the timeout but before
2807 * the scheduler thread gets it */
2808 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2809 conn->gnc_peer_error = -ETIMEDOUT;
2812 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2814 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2815 CDEBUG(D_WARNING, "%s conn timed out, will "
2816 "reconnect upon request from peer\n",
2817 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2819 /* Once we mark closed, any of the scheduler threads could
2820 * get it and move through before we hit the fail loc code */
2821 kgnilnd_close_conn_locked(conn, rc);
2823 /* first_rx is used to decide when to release a conn from purgatory.
2825 first_rx = conn->gnc_first_rx;
2829 /* now regardless of starting new conn, find tx on peer queue that
2830 * are old and smell bad - do this first so we don't trigger
2831 * reconnect on empty queue if we timeout all */
2832 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2833 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2835 LCONSOLE_INFO("could not send to %s due to connection"
2836 " setup failure after %lu seconds\n",
2837 libcfs_nid2str(peer->gnp_nid),
2838 cfs_duration_sec(jiffies - tx->tx_qtime));
2840 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2842 list_add_tail(&tx->tx_list, todie);
2847 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2848 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2849 count, peer, libcfs_nid2str(peer->gnp_nid));
2850 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2851 if (list_empty(&peer->gnp_tx_queue) ||
2852 peer->gnp_connecting == GNILND_PEER_KILL) {
2853 /* we pass down todie to use a common function - but we know there are
2855 kgnilnd_cancel_peer_connect_locked(peer, todie);
2859 /* Don't reconnect if we are still trying to clear out old conns.
2860 * This prevents us sending traffic on the new mbox before ensuring we are done
2861 * with the old one */
2862 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2863 (atomic_read(&peer->gnp_dirty_eps) == 0);
2865 /* fast reconnect after a timeout */
2866 to_reconn = !conn &&
2867 (peer->gnp_last_errno == -ETIMEDOUT) &&
2868 *kgnilnd_tunables.kgn_fast_reconn;
2870 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2871 * to be sent, we'll check the reconnect interval and fire up a new
2872 * connection request */
2875 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2876 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2877 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2879 CDEBUG(D_NET, "starting connect to %s\n",
2880 libcfs_nid2str(peer->gnp_nid));
2881 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2882 "Peer was idle and we have a write_lock, state issue %d\n",
2883 peer->gnp_connecting);
2885 peer->gnp_connecting = GNILND_PEER_CONNECT;
2886 kgnilnd_peer_addref(peer); /* extra ref for connd */
2888 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2889 list_add_tail(&peer->gnp_connd_list,
2890 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2891 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2893 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2896 /* fail_loc to allow us to delay release of purgatory */
2897 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2900 /* This check allows us to verify that the new conn is actually being used. This allows us to
2901 * pull the old conns out of purgatory if they have actually seen traffic.
2902 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2905 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2906 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2907 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2911 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2912 /* check for purgatory timeouts */
2913 if (conn->gnc_in_purgatory) {
2914 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2915 * that way the next time we check it we can detach it from purgatory
2918 if (conn->gnc_state != GNILND_CONN_DONE) {
2919 /* Skip over conns that are currently not DONE. If they arent already scheduled
2920 * for completion something in the state machine is broken.
2925 /* We only detach a conn that is in purgatory if we have received a close message,
2926 * we have a new valid connection that has successfully received data, or an admin
2927 * command tells us we need to detach.
2930 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2931 unsigned long waiting;
2933 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2935 /* C.E: The remote peer is expected to close the
2936 * connection (see kgnilnd_check_conn_timeouts)
2937 * via the reaper thread and nuke out the MDD and
2938 * FMA resources after conn->gnc_timeout has expired
2939 * without an FMA RX */
2940 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2941 " held resources\n",
2942 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2943 cfs_duration_sec(waiting));
2945 kgnilnd_detach_purgatory_locked(conn, souls);
2947 purgatory_conn_cnt++;
2952 /* If we have too many connections in purgatory we could run out of
2953 * resources. Limit the number of connections to a tunable number,
2954 * clean up to the minimum all in one fell swoop... there are
2955 * situations where dvs will retry tx's and we can eat up several
2956 * hundread connection requests at once.
2958 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2959 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2961 if (conn->gnc_in_purgatory &&
2962 conn->gnc_state == GNILND_CONN_DONE) {
2963 CDEBUG(D_NET, "Dropping Held resource due to"
2964 " resource limits being hit\n");
2965 kgnilnd_detach_purgatory_locked(conn, souls);
2967 if (purgatory_conn_cnt-- <
2968 *kgnilnd_tunables.kgn_max_purgatory)
2976 kgnilnd_reaper_check(int idx)
2978 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2979 struct list_head *ctmp, *ctmpN;
2980 LIST_HEAD(geriatrics);
2983 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2985 list_for_each_safe(ctmp, ctmpN, peers) {
2986 kgn_peer_t *peer = NULL;
2988 /* don't timeout stuff if the network is mucked or shutting down */
2989 if (kgnilnd_check_hw_quiesce()) {
2992 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2994 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
2997 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2999 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3000 kgnilnd_release_purgatory_list(&souls);
3004 kgnilnd_update_reaper_timeout(long timeout)
3006 LASSERT(timeout > 0);
3008 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3010 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3011 kgnilnd_data.kgn_new_min_timeout = timeout;
3013 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3017 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3019 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3023 kgnilnd_reaper(void *arg)
3028 unsigned long next_check_time = jiffies;
3029 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3030 struct timer_list timer;
3033 /* all gnilnd threads need to run fairly urgently */
3034 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3035 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3037 while (!kgnilnd_data.kgn_shutdown) {
3038 /* I wake up every 'p' seconds to check for timeouts on some
3039 * more peers. I try to check every connection 'n' times
3040 * within the global minimum of all keepalive and timeout
3041 * intervals, to ensure I attend to every connection within
3042 * (n+1)/n times its timeout intervals. */
3043 const int p = GNILND_REAPER_THREAD_WAKE;
3044 const int n = GNILND_REAPER_NCHECKS;
3046 /* to quiesce or to not quiesce, that is the question */
3047 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3048 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3049 KGNILND_SPIN_QUIESCE;
3050 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3053 /* careful with the jiffy wrap... */
3054 timeout = (long)(next_check_time - jiffies);
3057 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3058 TASK_INTERRUPTIBLE);
3059 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3060 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3062 mod_timer(&timer, (long) jiffies + timeout);
3064 /* check flag variables before committing */
3065 if (!kgnilnd_data.kgn_shutdown &&
3066 !kgnilnd_data.kgn_quiesce_trigger) {
3067 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3068 timeout, cfs_duration_sec(timeout));
3070 CDEBUG(D_INFO, "awake after schedule\n");
3073 del_singleshot_timer_sync(&timer);
3074 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3075 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3079 /* new_min_timeout is set from the conn timeouts and keepalive
3080 * this should end up with a min timeout of
3081 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3082 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3083 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3084 CDEBUG(D_NET, "Set new min timeout %ld\n",
3085 current_min_timeout);
3088 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3090 /* Compute how many table entries to check now so I get round
3091 * the whole table fast enough given that I do this at fixed
3092 * intervals of 'p' seconds) */
3093 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3094 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3095 chunk = (chunk * n * p) /
3096 kgnilnd_data.kgn_new_min_timeout;
3099 for (i = 0; i < chunk; i++) {
3100 kgnilnd_reaper_check(hash_index);
3101 hash_index = (hash_index + 1) %
3102 *kgnilnd_tunables.kgn_peer_hash_size;
3104 next_check_time = (long) jiffies + cfs_time_seconds(p);
3105 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3107 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3110 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3112 kgnilnd_thread_fini();
3117 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3118 unsigned niov, offset, nob;
3119 struct bio_vec *kiov;
3120 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3121 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3124 lnet_copy_flat2kiov(
3127 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3129 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3136 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3139 gni_post_descriptor_t *desc;
3141 kgn_tx_ev_id_t ev_id;
3143 int should_retry, rc;
3144 long num_processed = 0;
3145 kgn_conn_t *conn = NULL;
3146 kgn_tx_t *tx = NULL;
3147 kgn_rdma_desc_t *rdesc;
3152 /* make sure we don't keep looping if we need to reset */
3153 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3154 return num_processed;
3156 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3158 /* we didn't get the mutex, so return that there is still work
3162 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3163 /* a bit gross - but we need a good way to test for
3164 * delayed RDMA completions and the easiest way to do
3165 * that is to delay the RDMA CQ events */
3166 rrc = GNI_RC_NOT_DONE;
3168 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3171 if (rrc == GNI_RC_NOT_DONE) {
3172 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3173 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3174 dev->gnd_id, num_processed);
3175 return num_processed;
3177 dev->gnd_sched_alive = jiffies;
3180 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3181 "this is bad, somehow our credits didn't protect us"
3182 " from CQ overrun\n");
3183 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3184 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3185 event_data, GNI_CQ_GET_TYPE(event_data));
3187 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3189 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3191 /* XXX Nic: Need better error handling here... */
3192 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3193 (rrc == GNI_RC_TRANSACTION_ERROR),
3196 ev_id.txe_cookie = desc->post_id;
3198 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3200 if (conn == NULL || tx == NULL) {
3201 /* either conn or tx was already nuked and this is a "late"
3202 * completion, so drop it */
3206 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3207 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3208 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3209 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3210 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3212 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3214 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3215 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3216 tx->tx_msg.gnm_u.completion.gncm_retval);
3220 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3221 if (tx->tx_buffer_copy != NULL)
3222 kgnilnd_recv_bte_get(tx);
3223 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3226 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3227 if (tx->tx_buffer_copy != NULL)
3228 kgnilnd_recv_bte_get(tx);
3229 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3232 /* remove from rdmaq */
3233 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3234 spin_lock(&conn->gnc_list_lock);
3235 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3236 spin_unlock(&conn->gnc_list_lock);
3237 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3239 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3240 event_data = 1LL << 48;
3244 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3245 atomic_inc(&dev->gnd_rdma_ntx);
3246 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3247 /* transaction succeeded, add into fmaq */
3248 kgnilnd_queue_tx(conn, tx);
3249 kgnilnd_peer_alive(conn->gnc_peer);
3251 /* drop ref from kgnilnd_validate_tx_ev_id */
3252 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3253 kgnilnd_conn_decref(conn);
3258 /* fall through to the TRANSACTION_ERROR case */
3261 /* get stringified version for log messages */
3262 kgnilnd_cq_error_str(event_data, &err_str, 256);
3263 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3265 /* make sure we are not off in the weeds with this tx */
3266 if (tx->tx_retrans >
3267 *kgnilnd_tunables.kgn_max_retransmits) {
3268 GNIDBG_TX(D_NETERROR, tx,
3269 "giving up on TX, too many retries", NULL);
3273 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3274 should_retry ? "transient" : "unrecoverable", err_str);
3276 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3277 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3278 rdesc = &tx->tx_putinfo.gnpam_desc;
3279 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3280 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3282 rdesc = &tx->tx_getinfo.gngm_desc;
3283 rnob = tx->tx_lntmsg[0]->msg_len;
3284 rcookie = tx->tx_getinfo.gngm_cookie;
3289 tx->tx_msg.gnm_type,
3293 kgnilnd_nak_rdma(conn,
3294 tx->tx_msg.gnm_type,
3297 tx->tx_msg.gnm_srcnid);
3298 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3299 kgnilnd_close_conn(conn, -ECOMM);
3302 /* drop ref from kgnilnd_validate_tx_ev_id */
3303 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3304 kgnilnd_conn_decref(conn);
3309 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3313 kgn_tx_ev_id_t ev_id;
3314 kgn_tx_t *tx = NULL;
3315 kgn_conn_t *conn = NULL;
3316 int queued_fma, saw_reply, rc;
3317 long num_processed = 0;
3318 struct list_head *ctmp, *ctmpN;
3321 /* make sure we don't keep looping if we need to reset */
3322 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3323 return num_processed;
3326 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3328 /* we didn't get the mutex, so return that there is still work
3333 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3334 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3336 if (rrc == GNI_RC_NOT_DONE) {
3338 "SMSG send CQ %d not ready (data %#llx) "
3339 "processed %ld\n", dev->gnd_id, event_data,
3342 if (num_processed > 0) {
3343 spin_lock(&dev->gnd_lock);
3344 if (!list_empty(&dev->gnd_delay_conns)) {
3345 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3346 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3347 list_del_init(&conn->gnc_delaylist);
3348 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3349 kgnilnd_schedule_conn_nolock(conn);
3351 spin_unlock(&dev->gnd_lock);
3352 kgnilnd_schedule_device(dev);
3354 spin_unlock(&dev->gnd_lock);
3357 return num_processed;
3360 dev->gnd_sched_alive = jiffies;
3363 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3364 "this is bad, somehow our credits didn't "
3365 "protect us from CQ overrun\n");
3366 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3367 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3368 event_data, GNI_CQ_GET_TYPE(event_data));
3370 /* if SMSG couldn't handle an error, time for conn to die */
3371 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3374 /* need to take the write_lock to ensure atomicity
3375 * on the conn state if we need to close it */
3376 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3377 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3379 /* Conn was destroyed? */
3381 "SMSG CQID lookup %#llx failed\n",
3382 GNI_CQ_GET_INST_ID(event_data));
3383 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3387 kgnilnd_cq_error_str(event_data, &err_str, 256);
3388 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3389 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3391 kgnilnd_close_conn_locked(conn, -ECOMM);
3393 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3395 /* no need to process rest of this tx -
3396 * it is getting canceled */
3400 /* fall through to GNI_RC_SUCCESS case */
3401 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3403 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3404 if (conn == NULL || tx == NULL) {
3405 /* either conn or tx was already nuked and this is a "late"
3406 * completion, so drop it */
3410 tx->tx_conn->gnc_last_tx_cq = jiffies;
3411 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3412 set_mb(conn->gnc_last_noop_cq, jiffies);
3415 /* lock tx_list_state and tx_state */
3416 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3417 spin_lock(&tx->tx_conn->gnc_list_lock);
3419 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3420 "state not GNILND_TX_LIVE_FMAQ", NULL);
3421 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3422 "not waiting for completion", NULL);
3424 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3427 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3429 /* This will trigger other FMA sends that were
3430 * pending this completion */
3431 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3433 /* we either did not expect reply or we already got it */
3434 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3436 spin_unlock(&tx->tx_conn->gnc_list_lock);
3437 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3440 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3442 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3443 kgnilnd_schedule_conn(conn);
3446 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3447 * If saw_reply is true we know that the tx is safe to use as the other thread
3448 * is already finished with it.
3452 /* no longer need to track on the live_fmaq */
3453 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3455 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3456 /* we already got reply & were waiting for
3457 * completion of initial send */
3458 /* to initiate RDMA transaction */
3459 GNIDBG_TX(D_NET, tx,
3460 "Pending RDMA 0x%p type 0x%02x",
3461 tx->tx_msg.gnm_type);
3462 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3463 rc = kgnilnd_send_mapped_tx(tx, 0);
3464 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3466 /* we are done with this tx */
3467 GNIDBG_TX(D_NET, tx,
3468 "Done with tx type 0x%02x",
3469 tx->tx_msg.gnm_type);
3470 kgnilnd_tx_done(tx, tx->tx_rc);
3474 /* drop ref from kgnilnd_validate_tx_ev_id */
3475 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3476 kgnilnd_conn_decref(conn);
3478 /* if we are waiting for a REPLY, we'll handle the tx then */
3479 } /* end for loop */
3483 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3488 long num_processed = 0;
3489 struct list_head *conns;
3490 struct list_head *tmp;
3494 /* make sure we don't keep looping if we need to reset */
3495 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3496 return num_processed;
3499 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3501 /* we didn't get the mutex, so return that there is still work
3505 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3506 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3508 if (rrc == GNI_RC_NOT_DONE) {
3509 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3511 dev->gnd_id, event_data, num_processed);
3512 return num_processed;
3514 dev->gnd_sched_alive = jiffies;
3517 /* this is the only CQ that can really handle transient
3519 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3520 rrc = cfs_fail_val ? cfs_fail_val
3521 : GNI_RC_ERROR_RESOURCE;
3522 if (rrc == GNI_RC_ERROR_RESOURCE) {
3523 /* set overrun too */
3524 event_data |= (1UL << 63);
3525 LASSERTF(GNI_CQ_OVERRUN(event_data),
3526 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3529 /* sender should get error event too and take care
3530 of failed transaction by re-transmitting */
3531 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3532 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3536 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3537 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3538 conn = kgnilnd_cqid2conn_locked(
3539 GNI_CQ_GET_INST_ID(event_data));
3541 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3542 "failed, dropping event %#llx\n",
3543 GNI_CQ_GET_INST_ID(event_data),
3546 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3548 GNI_CQ_GET_INST_ID(event_data),
3549 conn, conn->gnc_peer ?
3550 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3553 conn->gnc_last_rx_cq = jiffies;
3555 /* stash first rx so we can clear out purgatory.
3557 if (conn->gnc_first_rx == 0) {
3558 conn->gnc_first_rx = jiffies;
3560 kgnilnd_peer_alive(conn->gnc_peer);
3561 kgnilnd_schedule_conn(conn);
3563 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3567 /* FMA CQ has overflowed: check ALL conns */
3568 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3569 "conns on device %d\n", dev->gnd_id);
3571 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3573 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3574 conns = &kgnilnd_data.kgn_conns[rc];
3576 list_for_each(tmp, conns) {
3577 conn = list_entry(tmp, kgn_conn_t,
3580 if (conn->gnc_device == dev) {
3581 kgnilnd_schedule_conn(conn);
3582 conn->gnc_last_rx_cq = jiffies;
3586 /* don't block write lockers for too long... */
3587 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3592 /* try_map_if_full should only be used when processing TX from list of
3593 * backlog TX waiting on mappings to free up
3596 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3597 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3600 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3602 /* slight bit of race if multiple people calling, but at worst we'll have
3603 * order altered just a bit... which would not be determenistic anyways */
3604 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3606 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3608 /* We know that we have a GART reservation that should guarantee forward progress.
3609 * This means we don't need to take any extraordinary efforts if we are failing
3610 * mappings here - even if we are holding a very small number of these. */
3612 if (try_map_if_full || (rc == 0)) {
3613 rc = kgnilnd_map_buffer(tx);
3616 /* rc should be 0 if we mapped successfully here, if non-zero
3617 * we are queueing */
3619 /* if try_map_if_full set, they handle requeuing */
3620 if (unlikely(try_map_if_full)) {
3623 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3624 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3625 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3626 /* make sure we wake up sched to run this */
3627 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3628 /* return 0 as this is now queued for later sending */
3633 switch (tx->tx_msg.gnm_type) {
3637 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3638 * remote node where the RDMA will be started
3639 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3640 * be satisified. The rest of the errors are "hard" errors that require
3641 * upper layers to handle themselves.
3642 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3643 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3644 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3645 * GNILND_MSG_GET_DONE_REV.
3647 case GNILND_MSG_GET_REQ:
3648 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3649 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3650 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3651 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3652 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3653 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3654 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3656 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3657 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3659 case GNILND_MSG_PUT_ACK:
3660 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3661 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3662 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3663 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3665 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3666 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3669 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3670 case GNILND_MSG_PUT_DONE:
3671 case GNILND_MSG_PUT_REQ:
3672 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3673 &tx->tx_putinfo.gnpam_desc,
3674 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3675 tx->tx_putinfo.gnpam_dst_cookie);
3676 RETURN(try_map_if_full ? rc : 0);
3678 case GNILND_MSG_GET_DONE:
3679 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3680 &tx->tx_getinfo.gngm_desc,
3681 tx->tx_lntmsg[0]->msg_len,
3682 tx->tx_getinfo.gngm_cookie);
3683 RETURN(try_map_if_full ? rc : 0);
3685 case GNILND_MSG_PUT_REQ_REV:
3686 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3687 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3688 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3689 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3690 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3691 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3692 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3694 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3696 case GNILND_MSG_PUT_DONE_REV:
3697 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3698 &tx->tx_getinfo.gngm_desc,
3700 tx->tx_getinfo.gngm_cookie);
3701 RETURN(try_map_if_full ? rc : 0);
3703 case GNILND_MSG_GET_ACK_REV:
3704 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3705 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3706 /* LNET_GETS are a special case for parse */
3707 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3708 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3710 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3711 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3713 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3714 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3716 case GNILND_MSG_GET_DONE_REV:
3717 case GNILND_MSG_GET_REQ_REV:
3718 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3719 &tx->tx_putinfo.gnpam_desc,
3720 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3721 tx->tx_putinfo.gnpam_dst_cookie);
3722 RETURN(try_map_if_full ? rc : 0);
3730 kgnilnd_process_fmaq(kgn_conn_t *conn)
3733 kgn_tx_t *tx = NULL;
3734 void *buffer = NULL;
3735 unsigned int nob = 0;
3738 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3739 * However I will be rescheduled by an FMA completion event
3740 * when I eventually get some.
3741 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3742 * But it doesn't matter if I try to send a "real" message just
3743 * as I start closing because I'll get scheduled to send the
3746 /* Short circuit if the ep_handle is null we cant send anyway. */
3747 if (conn->gnc_ephandle == NULL)
3750 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3752 spin_lock(&conn->gnc_list_lock);
3754 if (list_empty(&conn->gnc_fmaq)) {
3755 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3757 spin_unlock(&conn->gnc_list_lock);
3759 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3760 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3761 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3762 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3763 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3765 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3766 conn->gnc_last_noop_cq,
3767 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3768 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3769 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3770 atomic_inc(&conn->gnc_sched_noop);
3771 set_mb(conn->gnc_last_noop_want, jiffies);
3773 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3776 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3780 rc = kgnilnd_set_tx_id(tx, conn);
3782 kgnilnd_tx_done(tx, rc);
3788 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3789 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3790 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3791 more_to_do = !list_empty(&conn->gnc_fmaq);
3792 spin_unlock(&conn->gnc_list_lock);
3795 /* if there is no real TX or no NOOP to send, bail */
3800 if (!tx->tx_retrans)
3801 tx->tx_cred_wait = jiffies;
3803 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3804 "tx with zero id", NULL);
3806 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3807 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3808 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3812 switch (tx->tx_msg.gnm_type) {
3816 case GNILND_MSG_NOOP:
3817 case GNILND_MSG_CLOSE:
3818 case GNILND_MSG_IMMEDIATE:
3819 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3820 buffer = tx->tx_buffer;
3824 case GNILND_MSG_GET_DONE:
3825 case GNILND_MSG_PUT_DONE:
3826 case GNILND_MSG_PUT_DONE_REV:
3827 case GNILND_MSG_GET_DONE_REV:
3828 case GNILND_MSG_PUT_NAK:
3829 case GNILND_MSG_GET_NAK:
3830 case GNILND_MSG_GET_NAK_REV:
3831 case GNILND_MSG_PUT_NAK_REV:
3832 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3835 case GNILND_MSG_PUT_REQ:
3836 case GNILND_MSG_GET_REQ_REV:
3837 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3839 case GNILND_MSG_PUT_ACK:
3840 case GNILND_MSG_PUT_REQ_REV:
3841 case GNILND_MSG_GET_ACK_REV:
3842 case GNILND_MSG_GET_REQ:
3843 /* This is really only to handle the retransmit of SMSG once these
3844 * two messages are setup in send_mapped_tx */
3845 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3849 if (likely(rc == 0)) {
3850 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3854 /* don't explicitly reschedule here - we are short credits and will rely on
3855 * kgnilnd_sendmsg to resched the conn if need be */
3857 } else if (rc < 0) {
3858 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3859 * almost certainly a software bug, but lets play nice with the other kids */
3860 kgnilnd_tx_done(tx, rc);
3861 /* just for fun, kick peer in arse - resetting conn might help to correct
3862 * this almost certainly buggy software caused return code */
3863 kgnilnd_close_conn(conn, rc);
3867 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3868 kgnilnd_schedule_conn(conn);
3873 kgnilnd_process_rdmaq(kgn_device_t *dev)
3878 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3882 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3883 unsigned long dead_bump;
3886 /* if we think we need to adjust, take lock to serialize and recheck */
3887 spin_lock(&dev->gnd_rdmaq_lock);
3888 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3889 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3891 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3893 /* roll the bucket forward */
3894 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3896 if (kgnilnd_data.kgn_rdmaq_override &&
3897 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3898 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3903 /* roll current outstanding forward to make sure we carry outstanding
3904 * committment forward
3905 * new_ok starts out as the whole interval value
3906 * - first subtract bytes_out from last interval, as that would push us over
3907 * strict limits for this interval
3908 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3910 * there is a small race here if someone is actively processing mappings and
3911 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3912 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3913 * - if this gives us problems in the future, we could use a read/write lock
3914 * to protect the resetting of these values */
3915 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3916 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3918 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3919 "current out %ld\n",
3920 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3921 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3923 spin_unlock(&dev->gnd_rdmaq_lock);
3926 spin_lock(&dev->gnd_rdmaq_lock);
3927 while (!list_empty(&dev->gnd_rdmaq)) {
3930 /* make sure we break out early on quiesce */
3931 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3932 /* always break with lock held - we unlock outside loop */
3936 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3937 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3940 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3941 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3942 /* if conn is dying, mark tx in tx_ref_table for
3943 * kgnilnd_complete_closed_conn to finish up */
3944 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3946 /* tx was moved to DYING, get next */
3949 spin_unlock(&dev->gnd_rdmaq_lock);
3951 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3952 spin_lock(&dev->gnd_rdmaq_lock);
3955 /* no ticket! add back to head */
3956 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3957 /* clear found_work so scheduler threads wait for timer */
3961 /* TX is GO for launch */
3962 tx->tx_qtime = jiffies;
3963 kgnilnd_send_mapped_tx(tx, 0);
3967 spin_unlock(&dev->gnd_rdmaq_lock);
3973 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3975 __swab64s(&d->gnrd_key.qword1);
3976 __swab64s(&d->gnrd_key.qword2);
3977 __swab64s(&d->gnrd_addr);
3978 __swab32s(&d->gnrd_nob);
3981 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3982 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3985 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3987 kgn_tx_ev_id_t ev_id;
3990 /* we use the cookie from the original TX, so we can find the match
3991 * by parsing that and using the txe_idx */
3992 ev_id.txe_cookie = cookie;
3994 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3997 /* check tx to make sure kgni didn't eat it */
3998 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3999 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
4001 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
4002 (tx->tx_id.txe_cookie = cookie)),
4003 "conn 0x%p->%s tx_ref_table hosed: wanted "
4004 "txe_cookie %#llx txe_idx %d "
4005 "found tx %p cookie %#llx txe_idx %d\n",
4006 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4007 cookie, ev_id.txe_idx,
4008 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
4010 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4011 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4012 "Unexpected TX type (%x, %x or %x) "
4013 "or state (%x, expected +%x) "
4014 "matched reply from %s\n",
4015 tx->tx_msg.gnm_type, type1, type2,
4016 tx->tx_state, GNILND_TX_WAITING_REPLY,
4017 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4019 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
4020 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4026 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4029 kgn_conn_t *conn = tx->tx_conn;
4030 __u64 nob = tx->tx_nob;
4031 __u32 physnop = tx->tx_phys_npages;
4032 int id = tx->tx_id.txe_smsg_id;
4033 int buftype = tx->tx_buftype;
4034 gni_mem_handle_t hndl;
4035 hndl.qword1 = tx->tx_map_key.qword1;
4036 hndl.qword2 = tx->tx_map_key.qword2;
4038 spin_lock(&conn->gnc_list_lock);
4040 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4041 "not waiting for reply", NULL);
4044 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4046 if (rc == -EFAULT) {
4047 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
4048 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4050 if(*kgnilnd_tunables.kgn_efault_lbug) {
4051 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4052 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4054 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4055 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4056 kgnilnd_tx_state2str(tx->tx_list_state),
4057 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4062 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4063 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4064 /* sample under lock as follow on steps require gnc_list_lock
4065 * - or call kgnilnd_tx_done which requires no locks held over
4066 * call to lnet_finalize */
4069 spin_unlock(&conn->gnc_list_lock);
4072 kgnilnd_tx_done(tx, tx->tx_rc);
4077 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4080 kgn_conn_t *conn = tx->tx_conn;
4082 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4083 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4085 /* the gncm_retval is passed in for PUTs */
4086 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4087 msg->gnm_u.completion.gncm_retval);
4089 kgnilnd_complete_tx(tx, rc);
4093 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4101 kgn_peer_t *peer = conn->gnc_peer;
4104 __u16 tmp_cksum = 0, msg_cksum = 0;
4105 int repost = 1, saw_complete;
4106 unsigned long timestamp, newest_last_rx, timeout;
4110 /* Short circuit if the ep_handle is null.
4111 * It's likely that its about to be closed as stale.
4113 if (conn->gnc_ephandle == NULL)
4116 timestamp = jiffies;
4117 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4118 /* delay in jiffies - we are really concerned only with things that
4119 * result in a schedule() or really holding this off for long times .
4120 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4121 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4123 /* Resample current time as we have no idea how long it took to get the mutex */
4124 timestamp = jiffies;
4126 /* We check here when the last time we received an rx, we do this before
4127 * we call getnext in case the thread has been blocked for a while. If we
4128 * havent received an rx since our timeout value we close the connection
4129 * as we should assume the other side has closed the connection. This will
4130 * stop us from sending replies to a mailbox that is already in purgatory.
4133 timeout = cfs_time_seconds(conn->gnc_timeout);
4134 newest_last_rx = GNILND_LASTRX(conn);
4136 /* Error injection to validate that timestamp checking works and closing the conn */
4137 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4138 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4141 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4142 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4143 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4144 cfs_duration_sec(timestamp - newest_last_rx),
4145 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4146 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4148 kgnilnd_close_conn(conn, rc);
4152 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4154 if (rrc == GNI_RC_NOT_DONE) {
4155 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4156 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4160 /* Instead of asserting when we get mailbox corruption lets attempt to
4161 * close the conn and recover. We can put the conn/mailbox into
4162 * purgatory and let purgatory deal with the problem. If we see
4163 * this NETTERROR reported on production systems in large amounts
4164 * we will need to revisit the state machine to see if we can tighten
4165 * it up further to improve data protection.
4168 if (rrc == GNI_RC_INVALID_STATE) {
4169 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4170 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4171 "detected closing conn %p from peer %s\n", conn,
4172 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4174 kgnilnd_close_conn(conn, rc);
4178 LASSERTF(rrc == GNI_RC_SUCCESS,
4179 "bad rc %d on conn %p from peer %s\n",
4180 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4182 msg = (kgn_msg_t *)prefix;
4184 rx = kgnilnd_alloc_rx();
4186 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4187 kgnilnd_release_msg(conn);
4188 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4189 conn, libcfs_nid2str(peer->gnp_nid));
4193 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4195 timestamp = conn->gnc_last_rx;
4196 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4197 atomic_inc(&conn->gnc_rx_seq);
4199 conn->gnc_last_rx = jiffies;
4200 /* stash first rx so we can clear out purgatory
4202 if (conn->gnc_first_rx == 0)
4203 conn->gnc_first_rx = jiffies;
4205 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4206 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4207 kgnilnd_peer_alive(conn->gnc_peer);
4210 rx->grx_conn = conn;
4212 rx->grx_received = current_kernel_time();
4214 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4217 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4223 kgnilnd_net_decref(net);
4226 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4227 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4229 /* XXX Nic: Do we need to swab cksum */
4230 if (msg->gnm_cksum != 0) {
4231 msg_cksum = msg->gnm_cksum;
4233 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4235 if (tmp_cksum != msg_cksum) {
4236 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4237 tmp_cksum, msg_cksum);
4238 kgnilnd_dump_msg(D_BUFFS, msg);
4243 /* restore checksum for future debug messages */
4244 msg->gnm_cksum = tmp_cksum;
4246 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4247 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4248 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4249 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4254 __swab32s(&msg->gnm_magic);
4255 __swab16s(&msg->gnm_version);
4256 __swab16s(&msg->gnm_type);
4257 __swab64s(&msg->gnm_srcnid);
4258 __swab64s(&msg->gnm_connstamp);
4259 __swab32s(&msg->gnm_seq);
4261 /* NB message type checked below; NOT here... */
4262 switch (msg->gnm_type) {
4263 case GNILND_MSG_GET_ACK_REV:
4264 case GNILND_MSG_PUT_ACK:
4265 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4268 case GNILND_MSG_PUT_REQ_REV:
4269 case GNILND_MSG_GET_REQ:
4270 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4278 if (msg->gnm_version != GNILND_MSG_VERSION) {
4279 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4280 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4285 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4286 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4287 libcfs_nid2str(msg->gnm_srcnid),
4288 libcfs_nid2str(peer->gnp_nid));
4293 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4294 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4295 " expected) from %s",
4296 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4297 libcfs_nid2str(peer->gnp_nid));
4302 if (msg->gnm_seq != seq) {
4303 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4304 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4309 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4311 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4312 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4313 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4314 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4315 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4316 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4317 /* double check state with lock held */
4318 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4319 /* only error if we are not already closing */
4320 if (conn->gnc_peer_error == -ETIMEDOUT) {
4321 unsigned long now = jiffies;
4322 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4324 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4325 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4326 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4328 cfs_duration_sec(now - timestamp),
4329 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4330 atomic_read(&conn->gnc_tx_seq),
4331 cfs_duration_sec(now - conn->gnc_last_tx),
4332 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4333 cfs_duration_sec(now - conn->gnc_last_noop_want),
4334 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4335 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4336 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4337 cfs_duration_sec(now - conn->gnc_last_sched_do),
4338 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4340 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4342 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4346 if (conn->gnc_close_recvd) {
4347 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4348 kgnilnd_msgtype2str(msg->gnm_type),
4349 msg->gnm_type, conn->gnc_close_recvd,
4350 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4355 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4356 /* XXX Nic: log message received on bad connection state */
4360 switch (msg->gnm_type) {
4361 case GNILND_MSG_NOOP:
4362 /* Nothing to do; just a keepalive */
4365 case GNILND_MSG_IMMEDIATE:
4366 /* only get SMSG payload for IMMEDIATE */
4367 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4368 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4369 msg->gnm_srcnid, rx, 0);
4372 case GNILND_MSG_GET_REQ_REV:
4373 case GNILND_MSG_PUT_REQ:
4374 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4375 msg->gnm_srcnid, rx, 1);
4378 case GNILND_MSG_GET_NAK_REV:
4379 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4380 msg->gnm_u.completion.gncm_cookie);
4384 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4386 case GNILND_MSG_PUT_NAK:
4387 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4388 msg->gnm_u.completion.gncm_cookie);
4392 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4394 case GNILND_MSG_PUT_ACK:
4395 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4396 msg->gnm_u.putack.gnpam_src_cookie);
4400 /* store putack data for later: deferred rdma or re-try */
4401 tx->tx_putinfo = msg->gnm_u.putack;
4404 spin_lock(&tx->tx_conn->gnc_list_lock);
4406 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4407 "not waiting for reply", NULL);
4409 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4411 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4412 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4413 /* sample under lock as follow on steps require gnc_list_lock
4414 * - or call kgnilnd_tx_done which requires no locks held over
4415 * call to lnet_finalize */
4418 /* cannot launch rdma if still waiting for fma-msg completion */
4419 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4420 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4421 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4423 spin_unlock(&tx->tx_conn->gnc_list_lock);
4426 rc = kgnilnd_send_mapped_tx(tx, 0);
4428 kgnilnd_tx_done(tx, rc);
4431 case GNILND_MSG_GET_ACK_REV:
4432 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4433 msg->gnm_u.putack.gnpam_src_cookie);
4437 /* store putack data for later: deferred rdma or re-try */
4438 tx->tx_putinfo = msg->gnm_u.putack;
4440 spin_lock(&tx->tx_conn->gnc_list_lock);
4442 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4443 "not waiting for reply", NULL);
4445 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4447 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4448 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4449 /* sample under lock as follow on steps require gnc_list_lock
4450 * - or call kgnilnd_tx_done which requires no locks held over
4451 * call to lnet_finalize */
4454 /* cannot launch rdma if still waiting for fma-msg completion */
4455 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4456 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4457 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4459 spin_unlock(&tx->tx_conn->gnc_list_lock);
4462 rc = kgnilnd_send_mapped_tx(tx, 0);
4464 kgnilnd_tx_done(tx, rc);
4467 case GNILND_MSG_PUT_DONE:
4468 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4469 msg->gnm_u.completion.gncm_cookie);
4473 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4474 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4475 "bad tx buftype %d", tx->tx_buftype);
4477 kgnilnd_finalize_rx_done(tx, msg);
4479 case GNILND_MSG_PUT_REQ_REV:
4480 case GNILND_MSG_GET_REQ:
4481 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4482 msg->gnm_srcnid, rx, 1);
4486 case GNILND_MSG_GET_NAK:
4487 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4488 msg->gnm_u.completion.gncm_cookie);
4492 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4493 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4494 "bad tx buftype %d", tx->tx_buftype);
4496 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4499 case GNILND_MSG_GET_DONE:
4500 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4501 msg->gnm_u.completion.gncm_cookie);
4505 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4506 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4507 "bad tx buftype %d", tx->tx_buftype);
4509 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4510 msg->gnm_u.completion.gncm_retval);
4512 kgnilnd_finalize_rx_done(tx, msg);
4514 case GNILND_MSG_GET_DONE_REV:
4515 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4516 msg->gnm_u.completion.gncm_cookie);
4520 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4521 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4522 "bad tx buftype %d", tx->tx_buftype);
4524 kgnilnd_finalize_rx_done(tx, msg);
4527 case GNILND_MSG_PUT_DONE_REV:
4528 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4529 msg->gnm_u.completion.gncm_cookie);
4534 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4535 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4536 "bad tx buftype %d", tx->tx_buftype);
4538 kgnilnd_finalize_rx_done(tx, msg);
4540 case GNILND_MSG_PUT_NAK_REV:
4541 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4542 msg->gnm_u.completion.gncm_cookie);
4547 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4548 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4549 "bad tx buftype %d", tx->tx_buftype);
4551 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4556 if (rc < 0) /* protocol/comms error */
4557 kgnilnd_close_conn(conn, rc);
4559 if (repost && rx != NULL) {
4560 kgnilnd_consume_rx(rx);
4563 /* we got an event so assume more there and call for reschedule */
4565 kgnilnd_schedule_conn(conn);
4569 /* Do the failure injections that we need to affect conn processing in the following function.
4570 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4571 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4574 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4578 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4582 /* short circuit out when not set */
4583 if (likely(!cfs_fail_loc)) {
4587 /* failure injection to test for stack reset clean ups */
4588 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4589 /* we can't rely on busy loops being nice enough to get the
4590 * stack reset triggered - it'd just spin on this conn */
4591 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4594 GOTO(did_fail_loc, rc);
4597 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4598 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4600 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4601 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4604 GOTO(did_fail_loc, rc);
4608 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4609 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4610 * will be closed by kgnilnd_complete_closed_conn.
4612 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4613 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4616 GOTO(did_fail_loc, rc);
4619 /* this one is a bit gross - we can't hold the mutex from process_conns
4620 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4621 * and moving onto finish_connect
4622 * so, we'll just set the rc - kgnilnd_process_conns will clear
4623 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4624 * and effectively getting this thread to sleep */
4625 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4628 GOTO(did_fail_loc, rc);
4636 kgnilnd_send_conn_close(kgn_conn_t *conn)
4640 /* we are closing the conn - we will try to send the CLOSE msg
4641 * but will not wait for anything else to flush */
4643 /* send the close if not already done so or received one */
4644 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4645 /* set close_sent regardless of the success of the
4646 * CLOSE message. We are going to try once and then
4647 * kick him out of the sandbox */
4648 conn->gnc_close_sent = 1;
4651 /* EP might be null already if remote side initiated a new connection.
4652 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4653 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4655 if (conn->gnc_ephandle != NULL) {
4658 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4660 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4661 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4662 tx->tx_qtime = jiffies;
4664 if (tx->tx_id.txe_idx == 0) {
4665 rc = kgnilnd_set_tx_id(tx, conn);
4667 kgnilnd_tx_done(tx, rc);
4671 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4674 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4675 kgnilnd_tx_done(tx, -EAGAIN);
4677 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4679 /* It wasnt sent and we dont care. */
4680 kgnilnd_tx_done(tx, rc);
4688 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4689 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4690 conn->gnc_state = GNILND_CONN_CLOSED;
4691 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4692 /* mark this conn as CLOSED now that we processed it
4693 * do after TX, so we can use CLOSING in asserts */
4697 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4698 /* simulate a RX CLOSE after the timeout but before
4699 * the scheduler thread gets it */
4700 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4701 conn->gnc_peer_error = -ETIMEDOUT;
4703 /* schedule to allow potential CLOSE and get the complete phase run */
4704 kgnilnd_schedule_conn(conn);
4708 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4713 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4714 int log_retrans, log_retrans_level;
4715 static int last_map_version;
4718 spin_lock(&dev->gnd_lock);
4719 if (list_empty(&dev->gnd_map_tx)) {
4720 /* if the list is empty make sure we dont have a timer running */
4721 del_singleshot_timer_sync(&dev->gnd_map_timer);
4722 spin_unlock(&dev->gnd_lock);
4726 dev->gnd_sched_alive = jiffies;
4728 /* we'll retry as fast as possible up to 25% of the limit, then we start
4729 * backing off until our map version changes - indicating we unmapped
4731 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4732 if (likely(dev->gnd_map_attempt == 0) ||
4733 time_after_eq(jiffies, dev->gnd_next_map) ||
4734 last_map_version != dev->gnd_map_version) {
4736 /* if this is our first attempt at mapping set last mapped to current
4737 * jiffies so we can timeout our attempt correctly.
4739 if (dev->gnd_map_attempt == 0)
4740 dev->gnd_last_map = jiffies;
4742 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4743 spin_unlock(&dev->gnd_lock);
4747 /* delete the previous timer if it exists */
4748 del_singleshot_timer_sync(&dev->gnd_map_timer);
4749 /* stash the last map version to let us know when a good one was seen */
4750 last_map_version = dev->gnd_map_version;
4752 /* we need to to take the lock and continually refresh the head of the list as
4753 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4754 * allowing them to squeeze in */
4756 while (!list_empty(&dev->gnd_map_tx)) {
4757 /* make sure we break out early on quiesce */
4758 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4759 /* always break with lock held - we unlock outside loop */
4763 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4765 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4768 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4769 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4770 /* if conn is dying, mark tx in tx_ref_table for
4771 * kgnilnd_complete_closed_conn to finish up */
4772 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4775 /* tx was moved to DYING, get next */
4779 spin_unlock(&dev->gnd_lock);
4780 rc = kgnilnd_send_mapped_tx(tx, 1);
4782 /* We made it! skip error handling.. */
4784 /* OK to continue on +ve errors as it won't get seen until
4785 * this function is called again - we operate on a copy of the original
4786 * list and not the live list */
4787 spin_lock(&dev->gnd_lock);
4788 /* reset map attempts back to zero we successfully
4789 * mapped so we can reset our timers */
4790 dev->gnd_map_attempt = 0;
4792 } else if (rc == -EAGAIN) {
4793 spin_lock(&dev->gnd_lock);
4794 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4795 spin_unlock(&dev->gnd_lock);
4796 GOTO(get_out_mapped, rc);
4797 } else if (rc != -ENOMEM) {
4798 /* carp, failure we can't handle */
4799 kgnilnd_tx_done(tx, rc);
4800 spin_lock(&dev->gnd_lock);
4801 /* reset map attempts back to zero we dont know what happened but it
4802 * wasnt a failed mapping
4804 dev->gnd_map_attempt = 0;
4808 /* time to handle the retry cases.. lock so we dont have 2 threads
4809 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4811 spin_lock(&dev->gnd_lock);
4812 dev->gnd_map_attempt++;
4813 if (dev->gnd_map_attempt < fast_remaps) {
4814 /* do nothing we just want it to go as fast as possible.
4815 * just set gnd_next_map to current jiffies so it will process
4816 * as fast as possible.
4818 dev->gnd_next_map = jiffies;
4820 /* Retry based on GNILND_MAP_RETRY_RATE */
4821 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4824 /* only log occasionally once we've retried fast_remaps */
4825 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4826 ((dev->gnd_map_attempt % fast_remaps) == 0);
4827 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4829 /* make sure we are not off in the weeds with this tx */
4830 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4831 GNIDBG_TX(D_NETERROR, tx,
4832 "giving up on TX, too many retries", NULL);
4833 spin_unlock(&dev->gnd_lock);
4834 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4835 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4836 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4838 tx->tx_putinfo.gnpam_dst_cookie,
4839 tx->tx_msg.gnm_srcnid);
4841 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4843 tx->tx_getinfo.gngm_cookie,
4844 tx->tx_msg.gnm_srcnid);
4846 kgnilnd_tx_done(tx, -ENOMEM);
4847 GOTO(get_out_mapped, rc);
4849 GNIDBG_TX(log_retrans_level, tx,
4850 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4852 "nq_map %d mdd# %d/%d GART %ld",
4853 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4854 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4855 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4856 atomic_read(&dev->gnd_nq_map),
4857 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4858 atomic64_read(&dev->gnd_nbytes_map));
4861 /* we need to stop processing the rest of the list, so add it back in */
4862 /* set timer to wake device when we need to schedule this tx */
4863 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4864 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4865 spin_unlock(&dev->gnd_lock);
4866 GOTO(get_out_mapped, rc);
4868 spin_unlock(&dev->gnd_lock);
4874 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4879 int error_inject = 0;
4883 spin_lock(&dev->gnd_lock);
4884 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4885 dev->gnd_sched_alive = jiffies;
4889 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4890 /* break with lock held */
4894 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4895 list_del_init(&conn->gnc_schedlist);
4897 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4900 if (!list_empty(&conn->gnc_delaylist))
4901 list_del_init(&conn->gnc_delaylist);
4902 spin_unlock(&dev->gnd_lock);
4904 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4906 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4907 conn_sched != GNILND_CONN_PROCESS,
4908 "conn %p on ready list but in bad state: %d\n",
4911 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4912 conn, kgnilnd_conn_state2str(conn));
4915 set_mb(conn->gnc_last_sched_do, jiffies);
4917 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4919 /* based on intent see if we should run again. */
4920 rc = kgnilnd_schedule_process_conn(conn, intent);
4922 /* drop ref from gnd_ready_conns */
4923 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4924 down_write(&dev->gnd_conn_sem);
4925 kgnilnd_conn_decref(conn);
4926 up_write(&dev->gnd_conn_sem);
4927 } else if (rc != 1) {
4928 kgnilnd_conn_decref(conn);
4930 /* clear this so that scheduler thread doesn't spin */
4932 /* break with lock held... */
4933 spin_lock(&dev->gnd_lock);
4937 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4938 down_write(&dev->gnd_conn_sem);
4940 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4941 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4942 /* If there are tx's currently in use in another
4943 * thread we dont want to complete the close
4944 * yet. Cycle this conn back through
4946 kgnilnd_schedule_conn(conn);
4948 kgnilnd_complete_closed_conn(conn);
4950 up_write(&dev->gnd_conn_sem);
4951 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4952 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4953 /* serialize SMSG CQs with ep_bind and smsg_release */
4954 down_write(&dev->gnd_conn_sem);
4955 kgnilnd_destroy_conn_ep(conn);
4956 up_write(&dev->gnd_conn_sem);
4957 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4958 /* if we need to do some CLOSE sending, etc done here do it */
4959 down_write(&dev->gnd_conn_sem);
4960 kgnilnd_send_conn_close(conn);
4961 kgnilnd_check_fma_rx(conn);
4962 up_write(&dev->gnd_conn_sem);
4963 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4964 /* start moving traffic if the old conns are cleared out */
4965 down_read(&dev->gnd_conn_sem);
4966 kgnilnd_check_fma_rx(conn);
4967 kgnilnd_process_fmaq(conn);
4968 up_read(&dev->gnd_conn_sem);
4971 rc = kgnilnd_schedule_process_conn(conn, 0);
4973 /* drop ref from gnd_ready_conns */
4974 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4975 down_write(&dev->gnd_conn_sem);
4976 kgnilnd_conn_decref(conn);
4977 up_write(&dev->gnd_conn_sem);
4978 } else if (rc != 1) {
4979 kgnilnd_conn_decref(conn);
4982 /* check list again with lock held */
4983 spin_lock(&dev->gnd_lock);
4986 /* If we are short circuiting due to timing we want to be scheduled
4987 * as soon as possible.
4989 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4992 spin_unlock(&dev->gnd_lock);
4998 kgnilnd_scheduler(void *arg)
5000 int threadno = (long)arg;
5003 unsigned long deadline = 0;
5006 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
5008 /* all gnilnd threads need to run fairly urgently */
5009 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
5010 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5011 while (!kgnilnd_data.kgn_shutdown) {
5013 /* Safe: kgn_shutdown only set when quiescent */
5015 /* to quiesce or to not quiesce, that is the question */
5017 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5018 KGNILND_SPIN_QUIESCE;
5021 /* tracking for when thread goes AWOL */
5022 dev->gnd_sched_alive = jiffies;
5024 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5025 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5026 /* let folks know we are up and kicking
5027 * - they can use this for latency savings, etc
5028 * - only change if IRQ, if IDLE leave alone as that
5029 * schedule_device calls to put us back to IRQ */
5030 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5032 down_read(&dev->gnd_conn_sem);
5033 /* always check these - they are super low cost */
5034 found_work += kgnilnd_check_fma_send_cq(dev);
5035 found_work += kgnilnd_check_fma_rcv_cq(dev);
5037 /* rdma CQ doesn't care about eps */
5038 found_work += kgnilnd_check_rdma_cq(dev);
5040 /* move some RDMA ? */
5041 found_work += kgnilnd_process_rdmaq(dev);
5043 /* map some pending RDMA requests ? */
5044 found_work += kgnilnd_process_mapped_tx(dev);
5046 /* the EP for a conn is not destroyed until all the references
5047 * to it are gone, so these checks should be safe
5048 * even if run in parallel with the CQ checking functions
5049 * _AND_ a thread that processes the CLOSED->DONE
5053 up_read(&dev->gnd_conn_sem);
5055 /* process all conns ready now */
5056 found_work += kgnilnd_process_conns(dev, deadline);
5058 /* do an eager check to avoid the IRQ disabling in
5059 * prepare_to_wait and friends */
5062 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5063 time_before(jiffies, deadline)) {
5065 if ((busy_loops % 10) == 0) {
5066 /* tickle heartbeat and watchdog to ensure our
5067 * piggishness doesn't turn into heartbeat failure */
5068 touch_nmi_watchdog();
5074 /* if we got here, found_work was zero or busy_loops means we
5075 * need to take a break. We'll clear gnd_ready but we'll check
5076 * one last time if there is an IRQ that needs processing */
5078 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5080 /* the first time this will go LOOP -> IDLE and let us do one final check
5081 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5082 * - this might allow other threads to block us for a bit if they
5083 * try to get the mutex, but that is good as we'd need to wake
5084 * up soon to handle the CQ or other processing anyways */
5086 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5088 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5089 time_after_eq(jiffies, deadline)) {
5091 "yeilding: found_work %d busy_loops %d\n",
5092 found_work, busy_loops);
5094 /* use yield if we are bailing due to busy_loops
5095 * - this will ensure we wake up soonish. This closes
5096 * a race with kgnilnd_device_callback - where it'd
5097 * not call wake_up() because gnd_ready == 1, but then
5098 * we come down and schedule() because of busy_loops.
5099 * We'd not be woken up until something poked our waitq
5100 * again. yield() ensures we wake up without another
5101 * waitq poke in that case */
5102 atomic_inc(&dev->gnd_n_yield);
5103 kgnilnd_data.kgn_last_condresched = jiffies;
5105 CDEBUG(D_INFO, "awake after yeild\n");
5106 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5107 } else if (found_work == GNILND_DEV_IDLE) {
5108 /* busy_loops is low and there is nothing to do,
5109 * go to sleep and wait for a waitq poke */
5111 "scheduling: found_work %d busy_loops %d\n",
5112 found_work, busy_loops);
5113 atomic_inc(&dev->gnd_n_schedule);
5114 kgnilnd_data.kgn_last_scheduled = jiffies;
5116 CDEBUG(D_INFO, "awake after schedule\n");
5117 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5119 finish_wait(&dev->gnd_waitq, &wait);
5122 kgnilnd_thread_fini();