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(&dev->gnd_waitq);
89 void kgnilnd_schedule_device_timer(cfs_timer_cb_arg_t data)
91 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_map_timer);
93 kgnilnd_schedule_device(dev);
96 void kgnilnd_schedule_device_timer_rd(cfs_timer_cb_arg_t data)
98 kgn_device_t *dev = cfs_from_timer(dev, data, gnd_rdmaq_timer);
100 kgnilnd_schedule_device(dev);
104 kgnilnd_device_callback(__u32 devid, __u64 arg)
107 int index = (int) arg;
109 if (index >= kgnilnd_data.kgn_ndevs) {
110 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
111 * interrupt context. */
112 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
117 dev = &kgnilnd_data.kgn_devices[index];
118 /* just basic sanity */
119 if (dev->gnd_id == devid) {
120 kgnilnd_schedule_device(dev);
122 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
127 /* sched_intent values:
128 * < 0 : do not reschedule under any circumstances
129 * == 0: reschedule if someone marked him WANTS_SCHED
130 * > 0 : force a reschedule */
131 /* Return code 0 means it did not schedule the conn, 1
132 * means it successfully scheduled the conn.
136 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
140 /* move back to IDLE but save previous state.
141 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
142 * let the xchg there handle any racing callers to get it
143 * onto gnd_ready_conns */
145 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
146 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
147 conn_sched == GNILND_CONN_PROCESS,
148 "conn %p after process in bad state: %d\n",
151 if (sched_intent >= 0) {
152 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
153 return kgnilnd_schedule_conn_refheld(conn, 1);
159 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
163 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
165 kgn_device_t *dev = conn->gnc_device;
169 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
170 /* we only care about the last person who marked want_sched since they
171 * are most likely the culprit
173 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
174 conn->gnc_sched_line = line;
175 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
176 * the chance to put it onto gnd_ready_conns.
177 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
178 * the conn in process_conns will take care of moving it back to
179 * SCHED when it is done processing */
181 if (sched == GNILND_CONN_IDLE) {
182 /* if the conn is already scheduled, we've already requested
183 * the scheduler thread wakeup */
185 /* Add a reference to the conn if we are not holding a reference
186 * already from the exisiting scheduler. We now use the same
187 * reference if we need to reschedule a conn while in a scheduler
190 kgnilnd_conn_addref(conn);
192 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
195 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
197 spin_lock(&dev->gnd_lock);
198 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
200 spin_unlock(&dev->gnd_lock);
201 set_mb(conn->gnc_last_sched_ask, jiffies);
204 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
208 /* make sure thread(s) going to process conns - but let it make
209 * separate decision from conn schedule */
211 kgnilnd_schedule_device(dev);
216 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
218 kgn_device_t *dev = conn->gnc_device;
220 spin_lock(&dev->gnd_lock);
221 if (list_empty(&conn->gnc_delaylist)) {
222 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
225 spin_unlock(&dev->gnd_lock);
227 kgnilnd_schedule_device(dev);
232 kgnilnd_schedule_dgram(kgn_device_t *dev)
236 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
237 if (wake != GNILND_DGRAM_SCHED) {
238 wake_up(&dev->gnd_dgram_waitq);
240 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
245 kgnilnd_free_tx(kgn_tx_t *tx)
247 /* taken from kgnilnd_tx_add_state_locked */
249 LASSERTF((tx->tx_list_p == NULL &&
250 tx->tx_list_state == GNILND_TX_ALLOCD) &&
251 list_empty(&tx->tx_list),
252 "tx %p with bad state %s (list_p %p) tx_list %s\n",
253 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
254 list_empty(&tx->tx_list) ? "empty" : "not empty");
256 atomic_dec(&kgnilnd_data.kgn_ntx);
258 /* we only allocate this if we need to */
259 if (tx->tx_phys != NULL) {
260 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
261 LIBCFS_MEM_MSG(tx->tx_phys,
262 GNILND_MAX_IOV * sizeof(gni_mem_segment_t),
266 /* Only free the buffer if we used it */
267 if (tx->tx_buffer_copy != NULL) {
268 LIBCFS_MEM_MSG(tx->tx_buffer_copy, tx->tx_rdma_desc.length,
270 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
271 tx->tx_buffer_copy = NULL;
274 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
276 LIBCFS_MEM_MSG(tx, sizeof(*tx), "slab-freed");
277 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
281 kgnilnd_alloc_tx (void)
285 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
288 tx = kmem_cache_zalloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
290 CERROR("failed to allocate tx\n");
293 LIBCFS_MEM_MSG(tx, sizeof(*tx), "slab-alloced");
295 /* setup everything here to minimize time under the lock */
296 tx->tx_buftype = GNILND_BUF_NONE;
297 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
298 INIT_LIST_HEAD(&tx->tx_list);
299 INIT_LIST_HEAD(&tx->tx_map_list);
300 tx->tx_list_state = GNILND_TX_ALLOCD;
302 atomic_inc(&kgnilnd_data.kgn_ntx);
307 /* csum_fold needs to be run on the return value before shipping over the wire */
308 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
310 /* we don't use offset as every one is passing a buffer reference that already
311 * includes the offset into the base address.
314 kgnilnd_cksum(void *ptr, size_t nob)
318 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
320 /* don't use magic 'no checksum' value */
324 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
331 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
332 unsigned int offset, unsigned int nob, int dump_blob)
338 unsigned int fraglen;
344 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
345 kiov, nkiov, offset, nob, dump_blob);
347 /* if loops changes, please change kgnilnd_setup_phys_buffer */
349 while (offset >= kiov->bv_len) {
350 offset -= kiov->bv_len;
356 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
357 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
359 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
360 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
362 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
363 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
365 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
366 odd, kiov[0].bv_len, offset, nob);
368 for (i = 0; i < nkiov; i++) {
369 pages[i] = kiov[i].bv_page;
372 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
374 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
376 /* return zero to avoid killing tx - we'll just get warning on console
377 * when remote end sees zero checksum */
380 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
382 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
387 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
388 (void *)addr + kiov[0].bv_offset +
391 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
392 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
396 fraglen = min(kiov->bv_len - offset, nob);
398 /* make dang sure we don't send a bogus checksum if somehow we get
399 * an odd length fragment on anything but the last entry in a kiov -
400 * we know from kgnilnd_setup_rdma_buffer that we can't have non
401 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
402 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
403 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
404 fraglen, nkiov, nob, kiov->bv_len,
407 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
409 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
412 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
413 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
414 addr, fraglen, offset);
419 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
421 kunmap(kiov->bv_page);
428 /* iov must not run out before end of data */
429 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
434 retsum = csum_fold(cksum);
436 /* don't use magic 'no checksum' value */
440 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
446 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
448 msg->gnm_magic = GNILND_MSG_MAGIC;
449 msg->gnm_version = GNILND_MSG_VERSION;
450 msg->gnm_type = type;
451 msg->gnm_payload_len = 0;
452 msg->gnm_srcnid = source;
453 /* gnm_connstamp gets set when FMA is sent */
454 /* gnm_srcnid is set on creation via function argument
455 * The right interface/net and nid is passed in when the message
461 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
463 kgn_tx_t *tx = kgnilnd_alloc_tx();
466 kgnilnd_init_msg(&tx->tx_msg, type, source);
468 CERROR("couldn't allocate new tx type %s!\n",
469 kgnilnd_msgtype2str(type));
476 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
482 case GNILND_MSG_GET_REQ:
483 case GNILND_MSG_GET_DONE:
484 nak_type = GNILND_MSG_GET_NAK;
486 case GNILND_MSG_PUT_REQ:
487 case GNILND_MSG_PUT_ACK:
488 case GNILND_MSG_PUT_DONE:
489 nak_type = GNILND_MSG_PUT_NAK;
491 case GNILND_MSG_PUT_REQ_REV:
492 case GNILND_MSG_PUT_DONE_REV:
493 nak_type = GNILND_MSG_PUT_NAK_REV;
495 case GNILND_MSG_GET_REQ_REV:
496 case GNILND_MSG_GET_ACK_REV:
497 case GNILND_MSG_GET_DONE_REV:
498 nak_type = GNILND_MSG_GET_NAK_REV;
501 CERROR("invalid msg type %s (%d)\n",
502 kgnilnd_msgtype2str(rx_type), rx_type);
505 /* only allow NAK on error and truncate to zero */
506 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
507 error, conn, cookie);
509 tx = kgnilnd_new_tx_msg(nak_type, source);
511 CNETERR("can't get TX to NAK RDMA to %s\n",
512 libcfs_nid2str(conn->gnc_peer->gnp_nid));
516 tx->tx_msg.gnm_u.completion.gncm_retval = error;
517 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
518 kgnilnd_queue_tx(conn, tx);
522 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
523 struct bio_vec *kiov,
524 unsigned int offset, unsigned int nob)
526 kgn_msg_t *msg = &tx->tx_msg;
529 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
530 * gni_smsg_send to send that as the payload */
532 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
535 tx->tx_buffer = NULL;
538 if (niov && niov > (nob >> PAGE_SHIFT))
539 niov = DIV_ROUND_UP(nob + offset + kiov->bv_offset,
542 while (offset >= kiov->bv_len) {
543 offset -= kiov->bv_len;
549 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
550 "bad niov %d msg %p kiov %p offset %d nob%d\n",
551 niov, msg, kiov, offset, nob);
553 for (i = 0; i < niov; i++) {
554 /* We can't have a bv_offset on anything but the first
555 * entry, otherwise we'll have a hole at the end of the
556 * mapping as we only map whole pages.
557 * Also, if we have a bv_len < PAGE_SIZE but we need to
558 * map more than bv_len, we will also have a whole at
559 * the end of that page which isn't allowed
561 if ((kiov[i].bv_offset != 0 && i > 0) ||
562 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
564 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
565 i, offset, nob, kiov->bv_offset,
569 tx->tx_imm_pages[i] = kiov[i].bv_page;
572 /* hijack tx_phys for the later unmap */
574 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
576 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
577 kiov[0].bv_offset + offset;
578 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
579 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
580 nob, kiov, tx->tx_buffer);
582 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
583 if (tx->tx_phys == NULL) {
584 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
588 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
589 /* make sure we take into account the kiov offset as the
590 * start of the buffer
592 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
595 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
596 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
598 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
603 /* checksum payload early - it shouldn't be changing after lnd_send */
604 if (*kgnilnd_tunables.kgn_checksum >= 2) {
605 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
606 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
607 msg->gnm_payload_cksum += 0xe00e;
609 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
610 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
614 msg->gnm_payload_cksum = 0;
621 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
622 unsigned int offset, unsigned int nob)
624 gni_mem_segment_t *phys;
626 unsigned int fraglen;
628 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
632 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
634 /* only allocate this if we are going to use it */
635 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
637 if (tx->tx_phys == NULL) {
638 CERROR("failed to allocate tx_phys\n");
643 LIBCFS_MEM_MSG(tx->tx_phys,
644 GNILND_MAX_IOV * sizeof(gni_mem_segment_t),
647 /* if loops changes, please change kgnilnd_cksum_kiov
648 * and kgnilnd_setup_immediate_buffer */
650 while (offset >= kiov->bv_len) {
651 offset -= kiov->bv_len;
657 /* at this point, kiov points to the first page that we'll actually map
658 * now that we've seeked into the koiv for offset and dropped any
659 * leading pages that fall entirely within the offset */
660 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
663 /* bv_offset is start of 'valid' buffer, so index offset past that */
664 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
667 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
668 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
671 fraglen = min(kiov->bv_len - offset, nob);
673 /* We can't have a bv_offset on anything but the first entry,
674 * otherwise we'll have a hole at the end of the mapping as we
675 * only map whole pages. Only the first page is allowed to
676 * have an offset - we'll add that into tx->tx_buffer and that
677 * will get used when we map in the segments (see
678 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
679 * but we need to map more than bv_len, we will also have a
680 * whole at the end of that page which isn't allowed
682 if ((phys != tx->tx_phys) &&
683 ((kiov->bv_offset != 0) ||
684 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
685 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
686 (int)(phys - tx->tx_phys),
687 offset, nob, kiov->bv_offset, kiov->bv_len);
692 if ((phys - tx->tx_phys) == GNILND_MAX_IOV) {
693 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
698 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
704 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
705 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
708 phys->address = page_to_phys(kiov->bv_page);
715 /* iov must not run out before end of data */
716 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
720 tx->tx_phys_npages = phys - tx->tx_phys;
725 if (tx->tx_phys != NULL) {
726 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
727 LIBCFS_MEM_MSG(tx->tx_phys, sizeof(*tx->tx_phys), "slab-freed");
734 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
735 struct bio_vec *kiov,
736 unsigned int offset, unsigned int nob)
738 return kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
741 /* kgnilnd_parse_lnet_rdma()
742 * lntmsg - message passed in from lnet.
743 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
744 * nob - actual number of bytes to in this message.
745 * put_len - It is possible for PUTs to have a different length than the
746 * length stored in lntmsg->msg_len since LNET can adjust this
747 * length based on it's buffer size and offset.
748 * lnet_try_match_md() sets the mlength that we use to do the RDMA
752 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
753 unsigned int *offset, unsigned int *nob,
754 struct bio_vec **kiov, int put_len)
756 /* GETs are weird, see kgnilnd_send */
757 if (lntmsg->msg_type == LNET_MSG_GET) {
758 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
761 *kiov = lntmsg->msg_md->md_kiov;
763 *niov = lntmsg->msg_md->md_niov;
764 *nob = lntmsg->msg_md->md_length;
767 *kiov = lntmsg->msg_kiov;
768 *niov = lntmsg->msg_niov;
770 *offset = lntmsg->msg_offset;
775 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
777 unsigned int niov, offset, nob;
778 struct bio_vec *kiov;
779 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
780 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
782 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
783 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
784 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
785 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
786 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
787 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
788 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
790 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
791 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
792 tx->tx_msg.gnm_payload_cksum = 0;
795 if (*kgnilnd_tunables.kgn_checksum < 3) {
796 tx->tx_msg.gnm_payload_cksum = 0;
800 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
802 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
806 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
808 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
810 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
814 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
815 tx->tx_msg.gnm_payload_cksum += 0xd00d;
819 /* kgnilnd_verify_rdma_cksum()
820 * tx - PUT_DONE/GET_DONE matched tx.
821 * rx_cksum - received checksum to compare against.
822 * put_len - see kgnilnd_parse_lnet_rdma comments.
825 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
829 unsigned int niov, offset, nob;
830 struct bio_vec *kiov;
831 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
832 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
834 /* we can only match certain requests */
835 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
836 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
837 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
838 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
839 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
840 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
841 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
843 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
844 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
849 if (*kgnilnd_tunables.kgn_checksum >= 3) {
850 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
851 "no RDMA payload checksum when enabled");
856 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
858 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
861 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
863 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
866 if (cksum != rx_cksum) {
867 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
868 "Bad RDMA payload checksum (%x expected %x); "
869 "kiov 0x%p niov %d nob %u offset %u",
870 cksum, rx_cksum, kiov, niov, nob, offset);
871 switch (dump_on_err) {
874 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
876 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
881 libcfs_debug_dumplog();
887 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
893 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
897 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
898 "already mapped!", NULL);
900 spin_lock(&dev->gnd_map_lock);
901 switch (tx->tx_buftype) {
903 GNIDBG_TX(D_EMERG, tx,
904 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
905 spin_unlock(&dev->gnd_map_lock);
909 case GNILND_BUF_PHYS_MAPPED:
910 bytes = tx->tx_phys_npages * PAGE_SIZE;
911 dev->gnd_map_nphys++;
912 dev->gnd_map_physnop += tx->tx_phys_npages;
916 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
917 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
918 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
919 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
920 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
923 atomic_inc(&dev->gnd_n_mdd);
924 atomic64_add(bytes, &dev->gnd_nbytes_map);
926 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
929 /* we only get here in the valid cases */
930 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
931 dev->gnd_map_version++;
932 spin_unlock(&dev->gnd_map_lock);
936 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
940 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
941 "not mapped!", NULL);
942 spin_lock(&dev->gnd_map_lock);
944 switch (tx->tx_buftype) {
946 GNIDBG_TX(D_EMERG, tx,
947 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
948 spin_unlock(&dev->gnd_map_lock);
952 case GNILND_BUF_PHYS_UNMAPPED:
953 bytes = tx->tx_phys_npages * PAGE_SIZE;
954 dev->gnd_map_nphys--;
955 dev->gnd_map_physnop -= tx->tx_phys_npages;
959 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
960 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
961 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
962 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
963 "bytes_out negative! %lld\n",
964 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
965 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
966 bytes, (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
969 atomic_dec(&dev->gnd_n_mdd);
970 atomic64_sub(bytes, &dev->gnd_nbytes_map);
972 /* we only get here in the valid cases */
973 list_del_init(&tx->tx_map_list);
974 dev->gnd_map_version++;
975 spin_unlock(&dev->gnd_map_lock);
979 kgnilnd_map_buffer(kgn_tx_t *tx)
981 kgn_conn_t *conn = tx->tx_conn;
982 kgn_device_t *dev = conn->gnc_device;
983 __u32 flags = GNI_MEM_READWRITE;
986 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
987 * be called concurrently as there are internal locks that protect
988 * any data structures or HW resources. We just need to ensure
989 * that our concurrency doesn't result in the kgn_device_t
990 * getting nuked while we are in here */
992 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
993 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
995 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
998 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
999 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1002 switch (tx->tx_buftype) {
1006 case GNILND_BUF_NONE:
1007 case GNILND_BUF_IMMEDIATE:
1008 case GNILND_BUF_IMMEDIATE_KIOV:
1009 case GNILND_BUF_PHYS_MAPPED:
1012 case GNILND_BUF_PHYS_UNMAPPED:
1013 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1014 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1015 tx->tx_phys, tx->tx_phys_npages, NULL,
1016 GNI_MEM_PHYS_SEGMENTS | flags,
1018 /* could race with other uses of the map counts, but this is ok
1019 * - this needs to turn into a non-fatal error soon to allow
1020 * GART resource, etc starvation handling */
1021 if (rrc != GNI_RC_SUCCESS) {
1022 GNIDBG_TX(D_NET, tx,
1023 "Can't map %d pages: dev %d phys %u pp %u",
1024 tx->tx_phys_npages, dev->gnd_id,
1025 dev->gnd_map_nphys, dev->gnd_map_physnop);
1026 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1029 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1030 kgnilnd_mem_add_map_list(dev, tx);
1036 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1038 kgn_conn_t *conn = tx->tx_conn;
1039 kgn_mdd_purgatory_t *gmp;
1041 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1042 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1043 " asserting to avoid data corruption\n");
1044 if (tx->tx_buffer_copy)
1045 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1047 gmp->gmp_map_key = tx->tx_map_key;
1049 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1051 /* ensure that we don't have a blank purgatory - indicating the
1052 * conn is not already on purgatory lists - we'd never recover these
1053 * MDD if that were the case */
1054 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1055 "conn 0x%p->%s with NULL purgatory",
1056 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1058 /* link 'er up! - only place we really need to lock for
1059 * concurrent access */
1060 spin_lock(&conn->gnc_list_lock);
1061 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1062 spin_unlock(&conn->gnc_list_lock);
1066 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1070 int hold_timeout = 0;
1072 /* code below relies on +1 relationship ... */
1073 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1074 (GNILND_BUF_PHYS_UNMAPPED + 1));
1076 switch (tx->tx_buftype) {
1080 case GNILND_BUF_NONE:
1081 case GNILND_BUF_IMMEDIATE:
1082 case GNILND_BUF_PHYS_UNMAPPED:
1084 case GNILND_BUF_IMMEDIATE_KIOV:
1085 if (tx->tx_phys != NULL) {
1086 vunmap(tx->tx_phys);
1087 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1088 kunmap(tx->tx_imm_pages[0]);
1090 /* clear to prevent kgnilnd_free_tx from thinking
1091 * this is a RDMA descriptor */
1095 case GNILND_BUF_PHYS_MAPPED:
1096 LASSERT(tx->tx_conn != NULL);
1098 dev = tx->tx_conn->gnc_device;
1100 /* only want to hold if we are closing conn without
1101 * verified peer notification - the theory is that
1102 * a TX error can be communicated in all other cases */
1103 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1104 error != -GNILND_NOPURG &&
1105 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1106 kgnilnd_add_purgatory_tx(tx);
1108 /* The timeout we give to kgni is a deadman stop only.
1109 * we are setting high to ensure we don't have the kgni timer
1110 * fire before ours fires _and_ is handled */
1111 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1113 GNIDBG_TX(D_NET, tx,
1114 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1115 tx->tx_conn->gnc_device, hold_timeout,
1116 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1118 if (tx->tx_buffer_copy != NULL) {
1119 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1120 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1121 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1122 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1124 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1125 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1129 kgnilnd_mem_del_map_list(dev, tx);
1135 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1137 struct lnet_msg *lntmsg0, *lntmsg1;
1138 int status0, status1;
1139 struct lnet_ni *ni = NULL;
1140 kgn_conn_t *conn = tx->tx_conn;
1142 LASSERT(!in_interrupt());
1144 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1145 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1148 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1149 !kgnilnd_conn_clean_errno(completion)) {
1150 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1151 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1152 completion, tx, conn ?
1153 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1154 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1155 kgnilnd_tx_state2str(tx->tx_list_state),
1156 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1159 /* The error codes determine if we hold onto the MDD */
1160 kgnilnd_unmap_buffer(tx, completion);
1162 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1163 * we play nice with the error codes to avoid delivering a failed
1164 * REQUEST and then a REPLY event as well */
1166 /* return -EIO to lnet - it is the magic value for failed sends */
1167 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1169 status1 = completion;
1171 status0 = status1 = completion;
1174 tx->tx_buftype = GNILND_BUF_NONE;
1175 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1177 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1178 * set NULL when we are a tx without a conn */
1180 ni = conn->gnc_peer->gnp_net->gnn_ni;
1182 spin_lock(&conn->gnc_tx_lock);
1184 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1185 (volatile unsigned long *)&conn->gnc_tx_bits),
1186 "conn %p tx %p bit %d already cleared\n",
1187 conn, tx, tx->tx_id.txe_idx);
1189 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1190 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1192 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1193 spin_unlock(&conn->gnc_tx_lock);
1196 kgnilnd_free_tx(tx);
1198 /* finalize AFTER freeing lnet msgs */
1200 /* warning - we should hold no locks here - calling lnet_finalize
1201 * could free up lnet credits, resulting in a call chain back into
1202 * the LND via kgnilnd_send and friends */
1204 lnet_finalize(lntmsg0, status0);
1206 if (lntmsg1 != NULL) {
1207 lnet_finalize(lntmsg1, status1);
1212 kgnilnd_txlist_done(struct list_head *txlist, int error)
1215 int err_printed = 0;
1217 if (list_empty(txlist))
1220 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1221 /* only print the first error */
1223 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1224 list_del_init(&tx->tx_list);
1225 kgnilnd_tx_done(tx, error);
1230 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1234 spin_lock(&conn->gnc_tx_lock);
1236 /* ID zero is NOT ALLOWED!!! */
1239 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1240 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1241 if (id == GNILND_MAX_MSG_ID) {
1242 if (conn->gnc_next_tx != 1) {
1243 /* we only searched from next_tx to end and didn't find
1244 * one, so search again from start */
1245 conn->gnc_next_tx = 1;
1248 /* couldn't find one! */
1249 spin_unlock(&conn->gnc_tx_lock);
1253 /* bump next_tx to prevent immediate reuse */
1254 conn->gnc_next_tx = id + 1;
1256 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1257 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1258 "tx 0x%p already at id %d\n",
1259 conn->gnc_tx_ref_table[id], id);
1261 /* delay these until we have a valid ID - prevents bad clear of the bit
1262 * in kgnilnd_tx_done */
1264 tx->tx_id.txe_cqid = conn->gnc_cqid;
1266 tx->tx_id.txe_idx = id;
1267 conn->gnc_tx_ref_table[id] = tx;
1269 /* Using jiffies to help differentiate against TX reuse - with
1270 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1271 * if we are sending to the same node faster than 256000/sec.
1272 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1274 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1276 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1278 spin_unlock(&conn->gnc_tx_lock);
1283 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1287 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1289 /* we don't care about TX timeout - it could be that the network is slower
1290 * or throttled. We'll keep retranmitting - so if the network is so slow
1291 * that we fill up our mailbox, we'll keep trying to resend that msg
1292 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1293 * that he hasn't send us any traffic in return */
1295 /* some reasonable throttling of the debug message */
1297 unsigned long now = jiffies;
1298 /* XXX Nic: Mystical TX debug here... */
1299 /* We expect retransmissions so only log when D_NET is enabled */
1300 GNIDBG_SMSG_CREDS(D_NET, conn);
1301 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1302 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1303 " last_msg %uus/%uus last_cq %uus/%uus",
1304 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1305 tx->tx_id, tx->tx_retrans,
1306 jiffies_to_usecs(now - tx->tx_cred_wait),
1307 jiffies_to_usecs(now - conn->gnc_last_tx),
1308 jiffies_to_usecs(now - conn->gnc_last_rx),
1309 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1310 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1314 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1315 * to avoid bad ordering with state_lock */
1318 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1319 spinlock_t *state_lock, kgn_tx_list_state_t state)
1321 kgn_conn_t *conn = tx->tx_conn;
1322 kgn_msg_t *msg = &tx->tx_msg;
1324 unsigned long newest_last_rx, timeout;
1327 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1328 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1330 "msg 0x%p type %d wrong payload size %d\n",
1331 msg, msg->gnm_type, immediatenob);
1333 /* make sure we catch all the cases where we'd send on a dirty old mbox
1334 * but allow case for sending CLOSE. Since this check is within the CQ
1335 * mutex barrier and the close message is only sent through
1336 * kgnilnd_send_conn_close the last message out the door will be the
1339 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1340 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1341 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1342 /* Return -ETIME, we are closing the connection already so we dont want to
1343 * have this tx hit the wire. The tx will be killed by the calling function.
1344 * Once the EP is marked dirty the close message will be the last
1345 * thing to hit the wire */
1350 timeout = cfs_time_seconds(conn->gnc_timeout);
1352 newest_last_rx = GNILND_LASTRX(conn);
1354 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1355 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1358 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1359 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1360 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1361 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1362 cfs_duration_sec(now - newest_last_rx),
1363 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1364 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1365 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1369 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1370 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1371 * creating the message this allows the message to contain the correct LNET NID/NET needed
1372 * instead of the one that the peer/conn uses for sending the data.
1374 msg->gnm_connstamp = conn->gnc_my_connstamp;
1375 msg->gnm_payload_len = immediatenob;
1376 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1378 /* always init here - kgn_checksum is a /sys module tunable
1379 * and can be flipped at any point, even between msg init and sending */
1381 if (*kgnilnd_tunables.kgn_checksum) {
1382 /* We must set here and not in kgnilnd_init_msg,
1383 * we could resend this msg many times
1384 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1385 * through init_msg again */
1386 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1387 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1388 msg->gnm_cksum += 0xf00f;
1392 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1393 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1394 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1395 tx->tx_id.txe_idx, immediate, immediatenob);
1397 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1398 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1400 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1401 msg, sizeof(*msg), immediate,
1403 tx->tx_id.txe_smsg_id);
1407 case GNI_RC_SUCCESS:
1408 atomic_inc(&conn->gnc_tx_seq);
1409 conn->gnc_last_tx = jiffies;
1410 /* no locking here as LIVE isn't a list */
1411 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1413 /* this needs to be checked under lock as it might be freed from a completion
1416 if (msg->gnm_type == GNILND_MSG_NOOP) {
1417 set_mb(conn->gnc_last_noop_sent, jiffies);
1420 /* serialize with seeing CQ events for completion on this, as well as
1422 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1423 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1425 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1426 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1427 kgnilnd_peer_alive(conn->gnc_peer);
1428 GNIDBG_SMSG_CREDS(D_NET, conn);
1431 case GNI_RC_NOT_DONE:
1432 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1433 * 1: out of mbox credits
1434 * 2: out of mbox payload credits
1435 * 3: On Aries out of dla credits
1437 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1438 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1439 /* We'll handle this error inline - makes the calling logic much more
1442 /* If no lock, caller doesn't want us to retry */
1443 if (state_lock == NULL) {
1447 /* I need kgni credits to send this. Replace tx at the head of the
1448 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1449 * and bump retrans counts since we are requeueing the tx.
1453 conn->gnc_tx_retrans++;
1455 kgnilnd_tx_log_retrans(conn, tx);
1456 /* add to head of list for the state and retries */
1457 spin_lock(state_lock);
1458 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1459 spin_unlock(state_lock);
1461 /* We only reschedule for a certain number of retries, then
1462 * we will wait for the CQ events indicating a release of SMSG
1464 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1465 kgnilnd_schedule_conn(conn);
1468 /* CQ event coming in signifies either TX completed or
1469 * RX receive. Either of these *could* free up credits
1470 * in the SMSG mbox and we should try sending again */
1471 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1472 tx->tx_conn->gnc_cqid);
1473 kgnilnd_schedule_delay_conn(conn);
1474 /* use +ve return code to let upper layers know they
1475 * should stop looping on sends */
1479 /* handle bad retcode gracefully */
1480 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1481 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1486 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1488 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1489 spinlock_t *state_lock, kgn_tx_list_state_t state)
1491 kgn_device_t *dev = tx->tx_conn->gnc_device;
1492 unsigned long timestamp;
1495 timestamp = jiffies;
1496 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1497 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1498 /* delay in jiffies - we are really concerned only with things that
1499 * result in a schedule() or really holding this off for long times .
1500 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1501 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1503 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1509 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1511 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1512 spinlock_t *state_lock, kgn_tx_list_state_t state)
1514 kgn_conn_t *conn = tx->tx_conn;
1515 kgn_device_t *dev = conn->gnc_device;
1516 unsigned long timestamp;
1519 timestamp = jiffies;
1521 /* technically we are doing bad things with the read_lock on the peer_conn
1522 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1523 * for the mutex. I bet lockdep is gonna flag this one though... */
1525 /* there are a few cases where we don't want the immediate send - like
1526 * when we are in the scheduler thread and it'd harm the latency of
1527 * getting messages up to LNet */
1529 /* rmb for gnd_ready */
1531 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1533 atomic_inc(&conn->gnc_device->gnd_fast_block);
1534 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1535 /* dont hit HW during quiesce */
1537 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1538 /* dont hit HW if stale EPs and conns left to close */
1541 atomic_inc(&conn->gnc_device->gnd_fast_try);
1542 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1543 &conn->gnc_smsg_mutex);
1548 /* we got the mutex and weren't blocked */
1550 /* delay in jiffies - we are really concerned only with things that
1551 * result in a schedule() or really holding this off for long times .
1552 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1553 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1555 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1556 tx->tx_qtime = jiffies;
1557 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1558 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1559 /* _nolock unlocks the mutex for us */
1565 /* lets us know if we can push this RDMA through now */
1567 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1571 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1573 if (bytes_left < 0) {
1574 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1575 atomic_inc(&dev->gnd_rdmaq_nstalls);
1578 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1579 dev->gnd_rdmaq_deadline);
1580 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1581 /* we never del this timer - at worst it schedules us.. */
1588 /* this adds a TX to the queue pending throttling authorization before
1589 * we allow our remote peer to launch a PUT at us */
1591 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1595 /* we cannot go into send_mapped_tx from here as we are holding locks
1596 * and mem registration might end up allocating memory in kgni.
1597 * That said, we'll push this as far as we can into the queue process */
1598 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1601 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1602 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1603 /* lets us know how delayed RDMA is */
1604 tx->tx_qtime = jiffies;
1605 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1607 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1608 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1609 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1610 /* lets us know how delayed mapping is */
1611 tx->tx_qtime = jiffies;
1612 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1615 /* make sure we wake up sched to run this */
1616 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1619 /* push TX through state machine */
1621 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1626 /* set the tx_id here, we delay it until we have an actual conn
1628 * in some cases, the tx_id is already set to provide for things
1629 * like RDMA completion cookies, etc */
1630 if (tx->tx_id.txe_idx == 0) {
1631 rc = kgnilnd_set_tx_id(tx, conn);
1633 kgnilnd_tx_done(tx, rc);
1638 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1639 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1641 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1643 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1644 kgnilnd_tx_done(tx, rc);
1648 switch (tx->tx_msg.gnm_type) {
1649 case GNILND_MSG_PUT_ACK:
1650 case GNILND_MSG_GET_REQ:
1651 case GNILND_MSG_PUT_REQ_REV:
1652 case GNILND_MSG_GET_ACK_REV:
1653 /* hijacking time! If this messages will authorize our peer to
1654 * send his dirty little bytes in an RDMA, we need to get permission */
1655 kgnilnd_queue_rdma(conn, tx);
1657 case GNILND_MSG_IMMEDIATE:
1658 /* try to send right now, can help reduce latency */
1659 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1662 /* it was sent, break out of switch to avoid default case of queueing */
1665 /* needs to queue to try again, so... */
1667 case GNILND_MSG_NOOP:
1668 /* Just make sure this goes out first for this conn */
1672 spin_lock(&conn->gnc_list_lock);
1673 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1674 tx->tx_qtime = jiffies;
1675 spin_unlock(&conn->gnc_list_lock);
1676 kgnilnd_schedule_conn(conn);
1681 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_processid *target)
1684 kgn_peer_t *new_peer = NULL;
1685 kgn_conn_t *conn = NULL;
1691 /* If I get here, I've committed to send, so I complete the tx with
1692 * failure on any problems */
1694 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1695 "tx already has connection %p", tx->tx_conn);
1697 /* do all of the peer & conn searching in one swoop - this avoids
1698 * nastiness when dropping locks and needing to maintain a sane state
1699 * in the face of stack reset or something else nuking peers & conns */
1701 /* I expect to find him, so only take a read lock */
1702 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1704 peer = kgnilnd_find_peer_locked(lnet_nid_to_nid4(&target->nid));
1706 conn = kgnilnd_find_conn_locked(peer);
1707 /* this could be NULL during quiesce */
1709 /* Connection exists; queue message on it */
1710 kgnilnd_queue_tx(conn, tx);
1711 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1715 /* don't create a connection if the peer is marked down */
1716 if (peer->gnp_state != GNILND_PEER_UP) {
1717 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1723 /* creating peer or conn; I'll need a write lock... */
1724 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1726 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1728 node_state = kgnilnd_get_node_state(ntohl(target->nid.nid_addr[0]));
1730 /* NB - this will not block during normal operations -
1731 * the only writer of this is in the startup/shutdown path. */
1732 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1738 /* ignore previous peer entirely - we cycled the lock, so we
1739 * will create new peer and at worst drop it if peer is still
1741 rc = kgnilnd_create_peer_safe(&new_peer, lnet_nid_to_nid4(&target->nid),
1744 up_read(&kgnilnd_data.kgn_net_rw_sem);
1748 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1749 up_read(&kgnilnd_data.kgn_net_rw_sem);
1751 /* search for peer again now that we have the lock
1752 * if we don't find it, add our new one to the list */
1753 kgnilnd_add_peer_locked(lnet_nid_to_nid4(&target->nid), new_peer,
1756 /* don't create a connection if the peer is not up */
1757 if (peer->gnp_state != GNILND_PEER_UP) {
1758 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1763 conn = kgnilnd_find_or_create_conn_locked(peer);
1765 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1766 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1771 /* oh hey, found a conn now... magical */
1772 kgnilnd_queue_tx(conn, tx);
1774 /* no conn, must be trying to connect - so we queue for now */
1775 tx->tx_qtime = jiffies;
1776 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1778 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1781 kgnilnd_tx_done(tx, rc);
1786 kgnilnd_rdma(kgn_tx_t *tx, int type,
1787 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1789 kgn_conn_t *conn = tx->tx_conn;
1790 unsigned long timestamp;
1791 gni_post_type_t post_type;
1794 unsigned int desc_nob = nob;
1795 void *desc_buffer = tx->tx_buffer;
1796 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1797 LASSERTF(kgnilnd_tx_mapped(tx),
1798 "unmapped tx %p\n", tx);
1799 LASSERTF(conn != NULL,
1800 "NULL conn on tx %p, naughty, naughty\n", tx);
1801 LASSERTF(nob <= sink->gnrd_nob,
1802 "nob %u > sink->gnrd_nob %d (%p)\n",
1803 nob, sink->gnrd_nob, sink);
1804 LASSERTF(nob <= tx->tx_nob,
1805 "nob %d > tx(%p)->tx_nob %d\n",
1806 nob, tx, tx->tx_nob);
1809 case GNILND_MSG_GET_DONE:
1810 case GNILND_MSG_PUT_DONE:
1811 post_type = GNI_POST_RDMA_PUT;
1813 case GNILND_MSG_GET_DONE_REV:
1814 case GNILND_MSG_PUT_DONE_REV:
1815 post_type = GNI_POST_RDMA_GET;
1818 CERROR("invalid msg type %s (%d)\n",
1819 kgnilnd_msgtype2str(type), type);
1822 if (post_type == GNI_POST_RDMA_GET) {
1823 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1824 * aligned. If the local buffer is not aligned correctly using the copy buffer
1825 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1826 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1827 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1828 * making sure the length and addr are aligned and copying the data into the correct location
1829 * once the transfer has completed.
1831 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1832 (sink->gnrd_addr & 3) ||
1835 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1837 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1839 if ((nob + tx->tx_offset) & 3) {
1840 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1841 atomic_inc(&kgnilnd_data.kgn_rev_length);
1843 desc_nob = (nob + tx->tx_offset);
1846 if (tx->tx_buffer_copy == NULL) {
1847 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1848 * and require at most we allocate a few extra bytes. */
1849 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1851 if (!tx->tx_buffer_copy) {
1852 /* allocation of buffer failed nak the rdma */
1853 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1854 kgnilnd_tx_done(tx, -EFAULT);
1857 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1858 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);
1859 if (rc != GNI_RC_SUCCESS) {
1860 /* Registration Failed nak rdma and kill the tx. */
1861 kgnilnd_vfree(tx->tx_buffer_copy,
1863 tx->tx_buffer_copy = NULL;
1864 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1865 kgnilnd_tx_done(tx, -EFAULT);
1869 desc_map_key = tx->tx_buffer_copy_map_key;
1870 desc_buffer = tx->tx_buffer_copy;
1874 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1875 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1876 tx->tx_rdma_desc.type = post_type;
1877 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1878 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1879 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1880 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1881 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1882 tx->tx_rdma_desc.length = desc_nob;
1883 tx->tx_nob_rdma = nob;
1884 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1885 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1886 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1887 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1888 /* prep final completion message */
1889 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1890 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1891 /* send actual size RDMA'd in retval */
1892 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1894 kgnilnd_compute_rdma_cksum(tx, nob);
1897 kgnilnd_queue_tx(conn, tx);
1901 /* Don't lie (CLOSE == RDMA idle) */
1902 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1903 tx, conn, conn->gnc_close_sent);
1905 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
1906 "0x%x cookie:%#llx",
1907 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
1909 /* set CQ dedicated for RDMA */
1910 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
1912 timestamp = jiffies;
1913 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
1914 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1915 /* delay in jiffies - we are really concerned only with things that
1916 * result in a schedule() or really holding this off for long times .
1917 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1918 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1920 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
1922 if (rrc == GNI_RC_ERROR_RESOURCE) {
1923 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1924 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1925 kgnilnd_unmap_buffer(tx, 0);
1927 if (tx->tx_buffer_copy != NULL) {
1928 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
1929 tx->tx_buffer_copy = NULL;
1932 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1933 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
1935 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1936 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1940 spin_lock(&conn->gnc_list_lock);
1941 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
1942 tx->tx_qtime = jiffies;
1943 spin_unlock(&conn->gnc_list_lock);
1944 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1945 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
1947 /* XXX Nic: is this a place we should handle more errors for
1948 * robustness sake */
1949 LASSERT(rrc == GNI_RC_SUCCESS);
1954 kgnilnd_alloc_rx(void)
1958 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
1960 CERROR("failed to allocate rx\n");
1963 LIBCFS_MEM_MSG(rx, sizeof(*rx), "slab-alloced");
1965 /* no memset to zero, we'll always fill all members */
1969 /* release is to just free connection resources
1970 * we use this for the eager path after copying */
1972 kgnilnd_release_msg(kgn_conn_t *conn)
1975 unsigned long timestamp;
1977 CDEBUG(D_NET, "consuming %p\n", conn);
1979 timestamp = jiffies;
1980 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
1981 /* delay in jiffies - we are really concerned only with things that
1982 * result in a schedule() or really holding this off for long times .
1983 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1984 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
1986 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
1987 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1989 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
1990 GNIDBG_SMSG_CREDS(D_NET, conn);
1992 kgnilnd_schedule_conn(conn);
1996 kgnilnd_consume_rx(kgn_rx_t *rx)
1998 kgn_conn_t *conn = rx->grx_conn;
1999 kgn_msg_t *rxmsg = rx->grx_msg;
2001 /* if we are eager, free the cache alloc'd msg */
2002 if (unlikely(rx->grx_eager)) {
2003 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2004 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2006 /* release ref from eager_recv */
2007 kgnilnd_conn_decref(conn);
2009 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2010 kgnilnd_release_msg(conn);
2013 LIBCFS_MEM_MSG(rx, sizeof(*rx), "slab-freed");
2014 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2018 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2020 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2021 int type = lntmsg->msg_type;
2022 struct lnet_processid *target = &lntmsg->msg_target;
2023 int target_is_router = lntmsg->msg_target_is_router;
2024 int routing = lntmsg->msg_routing;
2025 unsigned int niov = lntmsg->msg_niov;
2026 struct bio_vec *kiov = lntmsg->msg_kiov;
2027 unsigned int offset = lntmsg->msg_offset;
2028 unsigned int nob = lntmsg->msg_len;
2029 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2030 kgn_net_t *net = ni->ni_data;
2033 /* '1' for consistency with code that checks !mpflag to restore */
2034 unsigned int mpflag = 1;
2035 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2037 /* NB 'private' is different depending on what we're sending.... */
2038 LASSERT(!in_interrupt());
2040 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2041 type, nob, niov, libcfs_idstr(target));
2043 LASSERTF(nob == 0 || niov > 0,
2044 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2047 mpflag = memalloc_noreclaim_save();
2051 CERROR("lntmsg %p with unexpected type %d\n",
2056 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2064 if (routing || target_is_router)
2065 break; /* send IMMEDIATE */
2067 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2068 * check the eventual sink buffer here - if small enough, remote
2069 * end is perfectly capable of returning data in short message -
2070 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2071 * for IMMEDIATE messages which will have it send a real reply instead
2072 * of doing kgnilnd_recv to have the RDMA continued */
2073 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2076 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2077 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ,
2078 lnet_nid_to_nid4(&ni->ni_nid));
2080 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV,
2081 lnet_nid_to_nid4(&ni->ni_nid));
2087 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2088 lntmsg->msg_md->md_kiov,
2089 0, lntmsg->msg_md->md_length);
2091 CERROR("unable to setup buffer: %d\n", rc);
2092 kgnilnd_tx_done(tx, rc);
2097 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2098 if (tx->tx_lntmsg[1] == NULL) {
2099 CERROR("Can't create reply for GET to %s\n",
2100 libcfs_nidstr(&target->nid));
2101 kgnilnd_tx_done(tx, rc);
2106 tx->tx_lntmsg[0] = lntmsg;
2107 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2108 lnet_hdr_to_nid4(hdr, &tx->tx_msg.gnm_u.get.gngm_hdr);
2110 lnet_hdr_to_nid4(hdr,
2111 &tx->tx_msg.gnm_u.putreq.gnprm_hdr);
2113 /* rest of tx_msg is setup just before it is sent */
2114 kgnilnd_launch_tx(tx, net, target);
2116 case LNET_MSG_REPLY:
2118 /* to save on MDDs, we'll handle short kiov by vmap'ing
2119 * and sending via SMSG */
2120 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2123 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2124 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ,
2125 lnet_nid_to_nid4(&ni->ni_nid));
2127 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV,
2128 lnet_nid_to_nid4(&ni->ni_nid));
2135 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2138 kgnilnd_tx_done(tx, rc);
2143 tx->tx_lntmsg[0] = lntmsg;
2144 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2145 lnet_hdr_to_nid4(hdr,
2146 &tx->tx_msg.gnm_u.putreq.gnprm_hdr);
2148 lnet_hdr_to_nid4(hdr, &tx->tx_msg.gnm_u.get.gngm_hdr);
2150 /* rest of tx_msg is setup just before it is sent */
2151 kgnilnd_launch_tx(tx, net, target);
2155 /* send IMMEDIATE */
2157 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2158 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2160 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE,
2161 lnet_nid_to_nid4(&ni->ni_nid));
2167 rc = kgnilnd_setup_immediate_buffer(tx, niov, kiov, offset, nob);
2169 kgnilnd_tx_done(tx, rc);
2173 lnet_hdr_to_nid4(hdr, &tx->tx_msg.gnm_u.immediate.gnim_hdr);
2174 tx->tx_lntmsg[0] = lntmsg;
2175 kgnilnd_launch_tx(tx, net, target);
2178 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2180 memalloc_noreclaim_restore(mpflag);
2185 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2187 kgn_conn_t *conn = rx->grx_conn;
2188 kgn_msg_t *rxmsg = rx->grx_msg;
2189 unsigned int niov = lntmsg->msg_niov;
2190 struct bio_vec *kiov = lntmsg->msg_kiov;
2191 unsigned int offset = lntmsg->msg_offset;
2192 unsigned int nob = lntmsg->msg_len;
2197 switch (rxmsg->gnm_type) {
2198 case GNILND_MSG_PUT_REQ_REV:
2199 done_type = GNILND_MSG_PUT_DONE_REV;
2202 case GNILND_MSG_GET_REQ:
2203 done_type = GNILND_MSG_GET_DONE;
2206 CERROR("invalid msg type %s (%d)\n",
2207 kgnilnd_msgtype2str(rxmsg->gnm_type),
2212 tx = kgnilnd_new_tx_msg(done_type, lnet_nid_to_nid4(&ni->ni_nid));
2216 rc = kgnilnd_set_tx_id(tx, conn);
2220 rc = kgnilnd_setup_rdma_buffer(tx, niov, kiov, offset, nob);
2224 tx->tx_lntmsg[0] = lntmsg;
2225 tx->tx_getinfo = rxmsg->gnm_u.get;
2227 /* we only queue from kgnilnd_recv - we might get called from other contexts
2228 * and we don't want to block the mutex in those cases */
2230 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2231 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2232 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2233 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2238 kgnilnd_tx_done(tx, rc);
2239 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie,
2240 lnet_nid_to_nid4(&ni->ni_nid));
2242 lnet_finalize(lntmsg, rc);
2246 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2249 kgn_rx_t *rx = private;
2250 kgn_conn_t *conn = rx->grx_conn;
2251 kgn_msg_t *rxmsg = rx->grx_msg;
2252 kgn_msg_t *eagermsg = NULL;
2253 kgn_peer_t *peer = NULL;
2254 kgn_conn_t *found_conn = NULL;
2256 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2257 conn, rxmsg, lntmsg);
2259 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2260 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2261 rxmsg->gnm_payload_len);
2264 /* Grab a read lock so the connection doesnt disappear on us
2265 * while we look it up
2267 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2269 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2271 found_conn = kgnilnd_find_conn_locked(peer);
2274 /* Verify the connection found is the same one that the message
2275 * is supposed to be using, if it is not output an error message
2278 if (!peer || !found_conn
2279 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2280 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2281 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2282 peer, conn, found_conn);
2284 CERROR("Unexpected connstamp %#llx(%#llx expected) from %s\n",
2285 rxmsg->gnm_connstamp,
2286 found_conn->gnc_peer_connstamp,
2287 libcfs_nid2str(peer->gnp_nid));
2292 /* add conn ref to ensure it doesn't go away until all eager
2293 * messages processed */
2294 kgnilnd_conn_addref(conn);
2296 /* Now that we have verified the connection is valid and added a
2297 * reference we can remove the read_lock on the peer_conn_lock */
2298 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2300 /* we have no credits or buffers for this message, so copy it
2301 * somewhere for a later kgnilnd_recv */
2302 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2303 *kgnilnd_tunables.kgn_eager_credits) {
2304 CERROR("Out of eager credits to %s\n",
2305 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2309 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2311 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2312 if (eagermsg == NULL) {
2313 kgnilnd_conn_decref(conn);
2314 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2315 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2319 /* copy msg and payload */
2320 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2321 rx->grx_msg = eagermsg;
2324 /* stash this for lnet_finalize on cancel-on-conn-close */
2325 rx->grx_lntmsg = lntmsg;
2327 /* keep the same rx_t, it just has a new grx_msg now */
2328 *new_private = private;
2330 /* release SMSG buffer */
2331 kgnilnd_release_msg(conn);
2337 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2338 int delayed, unsigned int niov,
2339 struct bio_vec *kiov,
2340 unsigned int offset, unsigned int mlen, unsigned int rlen)
2342 kgn_rx_t *rx = private;
2343 kgn_conn_t *conn = rx->grx_conn;
2344 kgn_msg_t *rxmsg = rx->grx_msg;
2350 LASSERT(!in_interrupt());
2351 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2353 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2354 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2355 conn, rxmsg, lntmsg,
2356 niov, kiov, offset, mlen, rlen);
2358 /* we need to lock here as recv can be called from any context */
2359 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2360 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2361 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2363 /* someone closed the conn after we copied this out, nuke it */
2364 kgnilnd_consume_rx(rx);
2365 lnet_finalize(lntmsg, conn->gnc_error);
2368 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2370 switch (rxmsg->gnm_type) {
2372 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2373 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2374 conn, rx, rxmsg, lntmsg, niov, kiov, offset, mlen, rlen);
2377 case GNILND_MSG_IMMEDIATE:
2378 if (mlen > rxmsg->gnm_payload_len) {
2379 GNIDBG_MSG(D_ERROR, rxmsg,
2380 "Immediate message from %s too big: %d > %d",
2381 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2382 rxmsg->gnm_payload_len);
2384 kgnilnd_consume_rx(rx);
2388 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2389 * right after the kgn_msg_t header - so just 'cute' way of saying
2390 * rxmsg + sizeof(kgn_msg_t) */
2392 /* check payload checksum if sent */
2394 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2395 !rxmsg->gnm_payload_cksum &&
2396 rxmsg->gnm_payload_len != 0)
2397 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2399 if (rxmsg->gnm_payload_cksum != 0) {
2400 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2401 * which is what is used to calculate the cksum on the TX side */
2402 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2404 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2405 GNIDBG_MSG(D_NETERROR, rxmsg,
2406 "Bad payload checksum (%x expected %x)",
2407 pload_cksum, rxmsg->gnm_payload_cksum);
2408 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2410 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2411 &rxmsg[1], rxmsg->gnm_payload_len);
2414 libcfs_debug_dumplog();
2420 /* checksum problems are fatal, kill the conn */
2421 kgnilnd_consume_rx(rx);
2422 kgnilnd_close_conn(conn, rc);
2427 lnet_copy_flat2kiov(
2429 *kgnilnd_tunables.kgn_max_immediate,
2430 &rxmsg[1], 0, mlen);
2432 kgnilnd_consume_rx(rx);
2433 lnet_finalize(lntmsg, 0);
2436 case GNILND_MSG_PUT_REQ:
2437 /* LNET wants to truncate or drop transaction, sending NAK */
2439 kgnilnd_consume_rx(rx);
2440 lnet_finalize(lntmsg, 0);
2442 /* only error if lntmsg == NULL, otherwise we are just
2443 * short circuiting the rdma process of 0 bytes */
2444 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2445 lntmsg == NULL ? -ENOENT : 0,
2446 rxmsg->gnm_u.get.gngm_cookie,
2447 lnet_nid_to_nid4(&ni->ni_nid));
2450 /* sending ACK with sink buff. info */
2451 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK,
2452 lnet_nid_to_nid4(&ni->ni_nid));
2454 kgnilnd_consume_rx(rx);
2458 rc = kgnilnd_set_tx_id(tx, conn);
2460 GOTO(nak_put_req, rc);
2463 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2464 kiov, offset, mlen);
2466 GOTO(nak_put_req, rc);
2469 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2470 rxmsg->gnm_u.putreq.gnprm_cookie;
2471 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2472 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2473 (__u64)((unsigned long)tx->tx_buffer);
2474 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2476 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2477 tx->tx_qtime = jiffies;
2478 /* we only queue from kgnilnd_recv - we might get called from other contexts
2479 * and we don't want to block the mutex in those cases */
2481 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2482 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2483 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2484 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2486 kgnilnd_consume_rx(rx);
2490 /* make sure we send an error back when the PUT fails */
2491 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc,
2492 rxmsg->gnm_u.get.gngm_cookie,
2493 lnet_nid_to_nid4(&ni->ni_nid));
2494 kgnilnd_tx_done(tx, rc);
2495 kgnilnd_consume_rx(rx);
2497 /* return magic LNet network error */
2499 case GNILND_MSG_GET_REQ_REV:
2500 /* LNET wants to truncate or drop transaction, sending NAK */
2502 kgnilnd_consume_rx(rx);
2503 lnet_finalize(lntmsg, 0);
2505 /* only error if lntmsg == NULL, otherwise we are just
2506 * short circuiting the rdma process of 0 bytes */
2507 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2508 lntmsg == NULL ? -ENOENT : 0,
2509 rxmsg->gnm_u.get.gngm_cookie,
2510 lnet_nid_to_nid4(&ni->ni_nid));
2513 /* lntmsg can be null when parsing a LNET_GET */
2514 if (lntmsg != NULL) {
2515 /* sending ACK with sink buff. info */
2516 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV,
2517 lnet_nid_to_nid4(&ni->ni_nid));
2519 kgnilnd_consume_rx(rx);
2523 rc = kgnilnd_set_tx_id(tx, conn);
2525 GOTO(nak_get_req_rev, rc);
2527 rc = kgnilnd_setup_rdma_buffer(tx, niov,
2528 kiov, offset, mlen);
2530 GOTO(nak_get_req_rev, rc);
2532 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2533 rxmsg->gnm_u.putreq.gnprm_cookie;
2534 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2535 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2536 (__u64)((unsigned long)tx->tx_buffer);
2537 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2539 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2541 /* we only queue from kgnilnd_recv - we might get called from other contexts
2542 * and we don't want to block the mutex in those cases */
2544 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2545 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2546 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2547 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2550 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2552 rxmsg->gnm_u.get.gngm_cookie,
2553 lnet_nid_to_nid4(&ni->ni_nid));
2556 kgnilnd_consume_rx(rx);
2560 /* make sure we send an error back when the GET fails */
2561 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc,
2562 rxmsg->gnm_u.get.gngm_cookie,
2563 lnet_nid_to_nid4(&ni->ni_nid));
2564 kgnilnd_tx_done(tx, rc);
2565 kgnilnd_consume_rx(rx);
2567 /* return magic LNet network error */
2571 case GNILND_MSG_PUT_REQ_REV:
2572 /* LNET wants to truncate or drop transaction, sending NAK */
2574 kgnilnd_consume_rx(rx);
2575 lnet_finalize(lntmsg, 0);
2577 /* only error if lntmsg == NULL, otherwise we are just
2578 * short circuiting the rdma process of 0 bytes */
2579 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2580 lntmsg == NULL ? -ENOENT : 0,
2581 rxmsg->gnm_u.get.gngm_cookie,
2582 lnet_nid_to_nid4(&ni->ni_nid));
2586 if (lntmsg != NULL) {
2588 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2591 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2593 rxmsg->gnm_u.get.gngm_cookie,
2594 lnet_nid_to_nid4(&ni->ni_nid));
2596 kgnilnd_consume_rx(rx);
2598 case GNILND_MSG_GET_REQ:
2599 if (lntmsg != NULL) {
2601 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2604 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2606 rxmsg->gnm_u.get.gngm_cookie,
2607 lnet_nid_to_nid4(&ni->ni_nid));
2609 kgnilnd_consume_rx(rx);
2615 /* needs write_lock on kgn_peer_conn_lock held */
2617 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2619 unsigned long timeout, keepalive;
2620 unsigned long now = jiffies;
2621 unsigned long newest_last_rx;
2624 /* given that we found this conn hanging off a peer, it better damned
2625 * well be connected */
2626 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2627 "conn 0x%p->%s with bad state%s\n", conn,
2628 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2630 kgnilnd_conn_state2str(conn));
2632 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2633 "rx_diff %lu tx_diff %lu\n",
2634 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2635 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2636 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2637 cfs_duration_sec(now - conn->gnc_last_tx));
2639 timeout = cfs_time_seconds(conn->gnc_timeout);
2640 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2642 /* just in case our lack of RX msg processing is gumming up the works - give the
2643 * remove an extra chance */
2645 newest_last_rx = GNILND_LASTRX(conn);
2647 if (time_after_eq(now, newest_last_rx + timeout)) {
2648 uint32_t level = D_CONSOLE|D_NETERROR;
2650 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2653 GNIDBG_CONN(level, conn,
2654 "No gnilnd traffic received from %s for %lu "
2655 "seconds, terminating connection. Is node down? ",
2656 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2657 cfs_duration_sec(now - newest_last_rx));
2661 /* we don't timeout on last_tx stalls - we are going to trust the
2662 * underlying network to let us know when sends are failing.
2663 * At worst, the peer will timeout our RX stamp and drop the connection
2664 * at that point. We'll then see his CLOSE or at worst his RX
2665 * stamp stop and drop the connection on our end */
2667 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2668 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2669 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2670 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2671 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2673 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2674 conn->gnc_last_noop_cq,
2675 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2676 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2677 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2678 set_mb(conn->gnc_last_noop_want, jiffies);
2679 atomic_inc(&conn->gnc_reaper_noop);
2680 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2683 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP,
2684 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid));
2687 kgnilnd_queue_tx(conn, tx);
2693 /* needs write_lock on kgn_peer_conn_lock held */
2695 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2696 struct list_head *souls)
2698 unsigned long timeout;
2699 kgn_conn_t *conn, *connN = NULL;
2705 short releaseconn = 0;
2706 unsigned long first_rx = 0;
2707 int purgatory_conn_cnt = 0;
2709 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2710 peer, libcfs_nid2str(peer->gnp_nid),
2711 peer->gnp_reconnect_interval);
2713 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2714 GNILND_MIN_TIMEOUT));
2716 conn = kgnilnd_find_conn_locked(peer);
2718 /* if there is a valid conn, check the queues for timeouts */
2719 rc = kgnilnd_check_conn_timeouts_locked(conn);
2721 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2722 /* simulate a RX CLOSE after the timeout but before
2723 * the scheduler thread gets it */
2724 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2725 conn->gnc_peer_error = -ETIMEDOUT;
2728 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2730 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2731 CDEBUG(D_WARNING, "%s conn timed out, will "
2732 "reconnect upon request from peer\n",
2733 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2735 /* Once we mark closed, any of the scheduler threads could
2736 * get it and move through before we hit the fail loc code */
2737 kgnilnd_close_conn_locked(conn, rc);
2739 /* first_rx is used to decide when to release a conn from purgatory.
2741 first_rx = conn->gnc_first_rx;
2745 /* now regardless of starting new conn, find tx on peer queue that
2746 * are old and smell bad - do this first so we don't trigger
2747 * reconnect on empty queue if we timeout all */
2748 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2749 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2751 LCONSOLE_INFO("could not send to %s due to connection"
2752 " setup failure after %lu seconds\n",
2753 libcfs_nid2str(peer->gnp_nid),
2754 cfs_duration_sec(jiffies - tx->tx_qtime));
2756 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2758 list_add_tail(&tx->tx_list, todie);
2763 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2764 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2765 count, peer, libcfs_nid2str(peer->gnp_nid));
2766 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2767 if (list_empty(&peer->gnp_tx_queue) ||
2768 peer->gnp_connecting == GNILND_PEER_KILL) {
2769 /* we pass down todie to use a common function - but we know there are
2771 kgnilnd_cancel_peer_connect_locked(peer, todie);
2775 /* Don't reconnect if we are still trying to clear out old conns.
2776 * This prevents us sending traffic on the new mbox before ensuring we are done
2777 * with the old one */
2778 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2779 (atomic_read(&peer->gnp_dirty_eps) == 0);
2781 /* fast reconnect after a timeout */
2782 to_reconn = !conn &&
2783 (peer->gnp_last_errno == -ETIMEDOUT) &&
2784 *kgnilnd_tunables.kgn_fast_reconn;
2786 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2787 * to be sent, we'll check the reconnect interval and fire up a new
2788 * connection request */
2791 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2792 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2793 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2795 CDEBUG(D_NET, "starting connect to %s\n",
2796 libcfs_nid2str(peer->gnp_nid));
2797 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2798 "Peer was idle and we have a write_lock, state issue %d\n",
2799 peer->gnp_connecting);
2801 peer->gnp_connecting = GNILND_PEER_CONNECT;
2802 kgnilnd_peer_addref(peer); /* extra ref for connd */
2804 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2805 list_add_tail(&peer->gnp_connd_list,
2806 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2807 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2809 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2812 /* fail_loc to allow us to delay release of purgatory */
2813 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2816 /* This check allows us to verify that the new conn is actually being used. This allows us to
2817 * pull the old conns out of purgatory if they have actually seen traffic.
2818 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2821 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2822 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2823 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2827 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2828 /* check for purgatory timeouts */
2829 if (conn->gnc_in_purgatory) {
2830 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2831 * that way the next time we check it we can detach it from purgatory
2834 if (conn->gnc_state != GNILND_CONN_DONE) {
2835 /* Skip over conns that are currently not DONE. If they arent already scheduled
2836 * for completion something in the state machine is broken.
2841 /* We only detach a conn that is in purgatory if we have received a close message,
2842 * we have a new valid connection that has successfully received data, or an admin
2843 * command tells us we need to detach.
2846 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2847 unsigned long waiting;
2849 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2851 /* C.E: The remote peer is expected to close the
2852 * connection (see kgnilnd_check_conn_timeouts)
2853 * via the reaper thread and nuke out the MDD and
2854 * FMA resources after conn->gnc_timeout has expired
2855 * without an FMA RX */
2856 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2857 " held resources\n",
2858 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2859 cfs_duration_sec(waiting));
2861 kgnilnd_detach_purgatory_locked(conn, souls);
2863 purgatory_conn_cnt++;
2868 /* If we have too many connections in purgatory we could run out of
2869 * resources. Limit the number of connections to a tunable number,
2870 * clean up to the minimum all in one fell swoop... there are
2871 * situations where dvs will retry tx's and we can eat up several
2872 * hundread connection requests at once.
2874 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2875 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2877 if (conn->gnc_in_purgatory &&
2878 conn->gnc_state == GNILND_CONN_DONE) {
2879 CDEBUG(D_NET, "Dropping Held resource due to"
2880 " resource limits being hit\n");
2881 kgnilnd_detach_purgatory_locked(conn, souls);
2883 if (purgatory_conn_cnt-- <
2884 *kgnilnd_tunables.kgn_max_purgatory)
2892 kgnilnd_reaper_check(int idx)
2894 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2895 struct list_head *ctmp, *ctmpN;
2896 LIST_HEAD(geriatrics);
2899 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2901 list_for_each_safe(ctmp, ctmpN, peers) {
2902 kgn_peer_t *peer = NULL;
2904 /* don't timeout stuff if the network is mucked or shutting down */
2905 if (kgnilnd_check_hw_quiesce()) {
2908 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
2910 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
2913 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2915 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
2916 kgnilnd_release_purgatory_list(&souls);
2920 kgnilnd_update_reaper_timeout(long timeout)
2922 LASSERT(timeout > 0);
2924 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2926 if (timeout < kgnilnd_data.kgn_new_min_timeout)
2927 kgnilnd_data.kgn_new_min_timeout = timeout;
2929 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2933 kgnilnd_reaper_poke_with_stick(cfs_timer_cb_arg_t arg)
2935 wake_up(&kgnilnd_data.kgn_reaper_waitq);
2939 kgnilnd_reaper(void *arg)
2944 unsigned long next_check_time = jiffies;
2945 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
2946 struct timer_list timer;
2949 /* all gnilnd threads need to run fairly urgently */
2950 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2951 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2953 while (!kgnilnd_data.kgn_shutdown) {
2954 /* I wake up every 'p' seconds to check for timeouts on some
2955 * more peers. I try to check every connection 'n' times
2956 * within the global minimum of all keepalive and timeout
2957 * intervals, to ensure I attend to every connection within
2958 * (n+1)/n times its timeout intervals. */
2959 const int p = GNILND_REAPER_THREAD_WAKE;
2960 const int n = GNILND_REAPER_NCHECKS;
2962 /* to quiesce or to not quiesce, that is the question */
2963 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2964 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2965 KGNILND_SPIN_QUIESCE;
2966 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2969 /* careful with the jiffy wrap... */
2970 timeout = (long)(next_check_time - jiffies);
2973 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
2974 TASK_INTERRUPTIBLE);
2975 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2976 cfs_timer_setup(&timer, kgnilnd_reaper_poke_with_stick,
2977 next_check_time, 0);
2978 mod_timer(&timer, (long) jiffies + timeout);
2980 /* check flag variables before committing */
2981 if (!kgnilnd_data.kgn_shutdown &&
2982 !kgnilnd_data.kgn_quiesce_trigger) {
2983 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2984 timeout, cfs_duration_sec(timeout));
2986 CDEBUG(D_INFO, "awake after schedule\n");
2989 timer_delete_sync(&timer);
2990 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2991 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
2995 /* new_min_timeout is set from the conn timeouts and keepalive
2996 * this should end up with a min timeout of
2997 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
2998 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
2999 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3000 CDEBUG(D_NET, "Set new min timeout %ld\n",
3001 current_min_timeout);
3004 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3006 /* Compute how many table entries to check now so I get round
3007 * the whole table fast enough given that I do this at fixed
3008 * intervals of 'p' seconds) */
3009 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3010 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3011 chunk = (chunk * n * p) /
3012 kgnilnd_data.kgn_new_min_timeout;
3015 for (i = 0; i < chunk; i++) {
3016 kgnilnd_reaper_check(hash_index);
3017 hash_index = (hash_index + 1) %
3018 *kgnilnd_tunables.kgn_peer_hash_size;
3020 next_check_time = (long) jiffies + cfs_time_seconds(p);
3021 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3023 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3026 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3028 kgnilnd_thread_fini();
3033 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3034 unsigned niov, offset, nob;
3035 struct bio_vec *kiov;
3036 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3037 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3040 lnet_copy_flat2kiov(
3043 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3045 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3052 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3055 gni_post_descriptor_t *desc;
3057 kgn_tx_ev_id_t ev_id;
3059 int should_retry, rc;
3060 long num_processed = 0;
3061 kgn_conn_t *conn = NULL;
3062 kgn_tx_t *tx = NULL;
3063 kgn_rdma_desc_t *rdesc;
3068 /* make sure we don't keep looping if we need to reset */
3069 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3070 return num_processed;
3072 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3074 /* we didn't get the mutex, so return that there is still work
3078 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3079 /* a bit gross - but we need a good way to test for
3080 * delayed RDMA completions and the easiest way to do
3081 * that is to delay the RDMA CQ events */
3082 rrc = GNI_RC_NOT_DONE;
3084 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3087 if (rrc == GNI_RC_NOT_DONE) {
3088 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3089 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3090 dev->gnd_id, num_processed);
3091 return num_processed;
3093 dev->gnd_sched_alive = jiffies;
3096 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3097 "this is bad, somehow our credits didn't protect us"
3098 " from CQ overrun\n");
3099 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3100 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3101 event_data, GNI_CQ_GET_TYPE(event_data));
3103 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3105 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3107 /* XXX Nic: Need better error handling here... */
3108 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3109 (rrc == GNI_RC_TRANSACTION_ERROR),
3112 ev_id.txe_cookie = desc->post_id;
3114 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3116 if (conn == NULL || tx == NULL) {
3117 /* either conn or tx was already nuked and this is a "late"
3118 * completion, so drop it */
3122 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3123 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3124 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3125 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3126 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3128 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3130 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3131 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3132 tx->tx_msg.gnm_u.completion.gncm_retval);
3136 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3137 if (tx->tx_buffer_copy != NULL)
3138 kgnilnd_recv_bte_get(tx);
3139 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3142 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3143 if (tx->tx_buffer_copy != NULL)
3144 kgnilnd_recv_bte_get(tx);
3145 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3148 /* remove from rdmaq */
3149 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3150 spin_lock(&conn->gnc_list_lock);
3151 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3152 spin_unlock(&conn->gnc_list_lock);
3153 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3155 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3156 event_data = 1LL << 48;
3160 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3161 atomic_inc(&dev->gnd_rdma_ntx);
3162 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3163 /* transaction succeeded, add into fmaq */
3164 kgnilnd_queue_tx(conn, tx);
3165 kgnilnd_peer_alive(conn->gnc_peer);
3167 /* drop ref from kgnilnd_validate_tx_ev_id */
3168 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3169 kgnilnd_conn_decref(conn);
3174 /* fall through to the TRANSACTION_ERROR case */
3177 /* get stringified version for log messages */
3178 kgnilnd_cq_error_str(event_data, &err_str, 256);
3179 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3181 /* make sure we are not off in the weeds with this tx */
3182 if (tx->tx_retrans >
3183 *kgnilnd_tunables.kgn_max_retransmits) {
3184 GNIDBG_TX(D_NETERROR, tx,
3185 "giving up on TX, too many retries", NULL);
3189 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3190 should_retry ? "transient" : "unrecoverable", err_str);
3192 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3193 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3194 rdesc = &tx->tx_putinfo.gnpam_desc;
3195 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3196 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3198 rdesc = &tx->tx_getinfo.gngm_desc;
3199 rnob = tx->tx_lntmsg[0]->msg_len;
3200 rcookie = tx->tx_getinfo.gngm_cookie;
3205 tx->tx_msg.gnm_type,
3209 kgnilnd_nak_rdma(conn,
3210 tx->tx_msg.gnm_type,
3213 tx->tx_msg.gnm_srcnid);
3214 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3215 kgnilnd_close_conn(conn, -ECOMM);
3218 /* drop ref from kgnilnd_validate_tx_ev_id */
3219 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3220 kgnilnd_conn_decref(conn);
3225 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3229 kgn_tx_ev_id_t ev_id;
3230 kgn_tx_t *tx = NULL;
3231 kgn_conn_t *conn = NULL;
3232 int queued_fma, saw_reply, rc;
3233 long num_processed = 0;
3234 struct list_head *ctmp, *ctmpN;
3237 /* make sure we don't keep looping if we need to reset */
3238 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3239 return num_processed;
3242 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3244 /* we didn't get the mutex, so return that there is still work
3249 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3250 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3252 if (rrc == GNI_RC_NOT_DONE) {
3254 "SMSG send CQ %d not ready (data %#llx) "
3255 "processed %ld\n", dev->gnd_id, event_data,
3258 if (num_processed > 0) {
3259 spin_lock(&dev->gnd_lock);
3260 if (!list_empty(&dev->gnd_delay_conns)) {
3261 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3262 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3263 list_del_init(&conn->gnc_delaylist);
3264 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3265 kgnilnd_schedule_conn_nolock(conn);
3267 spin_unlock(&dev->gnd_lock);
3268 kgnilnd_schedule_device(dev);
3270 spin_unlock(&dev->gnd_lock);
3273 return num_processed;
3276 dev->gnd_sched_alive = jiffies;
3279 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3280 "this is bad, somehow our credits didn't "
3281 "protect us from CQ overrun\n");
3282 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3283 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3284 event_data, GNI_CQ_GET_TYPE(event_data));
3286 /* if SMSG couldn't handle an error, time for conn to die */
3287 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3290 /* need to take the write_lock to ensure atomicity
3291 * on the conn state if we need to close it */
3292 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3293 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3295 /* Conn was destroyed? */
3297 "SMSG CQID lookup %#llx failed\n",
3298 GNI_CQ_GET_INST_ID(event_data));
3299 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3303 kgnilnd_cq_error_str(event_data, &err_str, 256);
3304 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3305 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3307 kgnilnd_close_conn_locked(conn, -ECOMM);
3309 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3311 /* no need to process rest of this tx -
3312 * it is getting canceled */
3316 /* fall through to GNI_RC_SUCCESS case */
3317 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3319 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3320 if (conn == NULL || tx == NULL) {
3321 /* either conn or tx was already nuked and this is a "late"
3322 * completion, so drop it */
3326 tx->tx_conn->gnc_last_tx_cq = jiffies;
3327 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3328 set_mb(conn->gnc_last_noop_cq, jiffies);
3331 /* lock tx_list_state and tx_state */
3332 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3333 spin_lock(&tx->tx_conn->gnc_list_lock);
3335 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3336 "state not GNILND_TX_LIVE_FMAQ", NULL);
3337 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3338 "not waiting for completion", NULL);
3340 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3343 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3345 /* This will trigger other FMA sends that were
3346 * pending this completion */
3347 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3349 /* we either did not expect reply or we already got it */
3350 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3352 spin_unlock(&tx->tx_conn->gnc_list_lock);
3353 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3356 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3358 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3359 kgnilnd_schedule_conn(conn);
3362 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3363 * If saw_reply is true we know that the tx is safe to use as the other thread
3364 * is already finished with it.
3368 /* no longer need to track on the live_fmaq */
3369 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3371 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3372 /* we already got reply & were waiting for
3373 * completion of initial send */
3374 /* to initiate RDMA transaction */
3375 GNIDBG_TX(D_NET, tx,
3376 "Pending RDMA 0x%p type 0x%02x",
3377 tx->tx_msg.gnm_type);
3378 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3379 rc = kgnilnd_send_mapped_tx(tx, 0);
3380 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3382 /* we are done with this tx */
3383 GNIDBG_TX(D_NET, tx,
3384 "Done with tx type 0x%02x",
3385 tx->tx_msg.gnm_type);
3386 kgnilnd_tx_done(tx, tx->tx_rc);
3390 /* drop ref from kgnilnd_validate_tx_ev_id */
3391 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3392 kgnilnd_conn_decref(conn);
3394 /* if we are waiting for a REPLY, we'll handle the tx then */
3395 } /* end for loop */
3399 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3404 long num_processed = 0;
3405 struct list_head *conns;
3406 struct list_head *tmp;
3410 /* make sure we don't keep looping if we need to reset */
3411 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3412 return num_processed;
3415 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3417 /* we didn't get the mutex, so return that there is still work
3421 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3422 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3424 if (rrc == GNI_RC_NOT_DONE) {
3425 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3427 dev->gnd_id, event_data, num_processed);
3428 return num_processed;
3430 dev->gnd_sched_alive = jiffies;
3433 /* this is the only CQ that can really handle transient
3435 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3436 rrc = cfs_fail_val ? cfs_fail_val
3437 : GNI_RC_ERROR_RESOURCE;
3438 if (rrc == GNI_RC_ERROR_RESOURCE) {
3439 /* set overrun too */
3440 event_data |= (1UL << 63);
3441 LASSERTF(GNI_CQ_OVERRUN(event_data),
3442 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3445 /* sender should get error event too and take care
3446 of failed transaction by re-transmitting */
3447 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3448 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3452 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3453 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3454 conn = kgnilnd_cqid2conn_locked(
3455 GNI_CQ_GET_INST_ID(event_data));
3457 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3458 "failed, dropping event %#llx\n",
3459 GNI_CQ_GET_INST_ID(event_data),
3462 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3464 GNI_CQ_GET_INST_ID(event_data),
3465 conn, conn->gnc_peer ?
3466 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3469 conn->gnc_last_rx_cq = jiffies;
3471 /* stash first rx so we can clear out purgatory.
3473 if (conn->gnc_first_rx == 0) {
3474 conn->gnc_first_rx = jiffies;
3476 kgnilnd_peer_alive(conn->gnc_peer);
3477 kgnilnd_schedule_conn(conn);
3479 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3483 /* FMA CQ has overflowed: check ALL conns */
3484 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3485 "conns on device %d\n", dev->gnd_id);
3487 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3489 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3490 conns = &kgnilnd_data.kgn_conns[rc];
3492 list_for_each(tmp, conns) {
3493 conn = list_entry(tmp, kgn_conn_t,
3496 if (conn->gnc_device == dev) {
3497 kgnilnd_schedule_conn(conn);
3498 conn->gnc_last_rx_cq = jiffies;
3502 /* don't block write lockers for too long... */
3503 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3508 /* try_map_if_full should only be used when processing TX from list of
3509 * backlog TX waiting on mappings to free up
3512 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3513 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3516 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3518 /* slight bit of race if multiple people calling, but at worst we'll have
3519 * order altered just a bit... which would not be determenistic anyways */
3520 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3522 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3524 /* We know that we have a GART reservation that should guarantee forward progress.
3525 * This means we don't need to take any extraordinary efforts if we are failing
3526 * mappings here - even if we are holding a very small number of these. */
3528 if (try_map_if_full || (rc == 0)) {
3529 rc = kgnilnd_map_buffer(tx);
3532 /* rc should be 0 if we mapped successfully here, if non-zero
3533 * we are queueing */
3535 /* if try_map_if_full set, they handle requeuing */
3536 if (unlikely(try_map_if_full)) {
3539 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3540 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3541 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3542 /* make sure we wake up sched to run this */
3543 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3544 /* return 0 as this is now queued for later sending */
3549 switch (tx->tx_msg.gnm_type) {
3553 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3554 * remote node where the RDMA will be started
3555 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3556 * be satisified. The rest of the errors are "hard" errors that require
3557 * upper layers to handle themselves.
3558 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3559 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3560 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3561 * GNILND_MSG_GET_DONE_REV.
3563 case GNILND_MSG_GET_REQ:
3564 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3565 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3566 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3567 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3568 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3569 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3570 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3572 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3573 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3575 case GNILND_MSG_PUT_ACK:
3576 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3577 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3578 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3579 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3581 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3582 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3585 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3586 case GNILND_MSG_PUT_DONE:
3587 case GNILND_MSG_PUT_REQ:
3588 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3589 &tx->tx_putinfo.gnpam_desc,
3590 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3591 tx->tx_putinfo.gnpam_dst_cookie);
3592 RETURN(try_map_if_full ? rc : 0);
3594 case GNILND_MSG_GET_DONE:
3595 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3596 &tx->tx_getinfo.gngm_desc,
3597 tx->tx_lntmsg[0]->msg_len,
3598 tx->tx_getinfo.gngm_cookie);
3599 RETURN(try_map_if_full ? rc : 0);
3601 case GNILND_MSG_PUT_REQ_REV:
3602 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3603 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3604 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3605 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3606 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3607 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3608 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3610 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3612 case GNILND_MSG_PUT_DONE_REV:
3613 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3614 &tx->tx_getinfo.gngm_desc,
3616 tx->tx_getinfo.gngm_cookie);
3617 RETURN(try_map_if_full ? rc : 0);
3619 case GNILND_MSG_GET_ACK_REV:
3620 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3621 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3622 /* LNET_GETS are a special case for parse */
3623 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3624 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3626 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3627 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3629 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3630 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3632 case GNILND_MSG_GET_DONE_REV:
3633 case GNILND_MSG_GET_REQ_REV:
3634 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3635 &tx->tx_putinfo.gnpam_desc,
3636 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3637 tx->tx_putinfo.gnpam_dst_cookie);
3638 RETURN(try_map_if_full ? rc : 0);
3646 kgnilnd_process_fmaq(kgn_conn_t *conn)
3649 kgn_tx_t *tx = NULL;
3650 void *buffer = NULL;
3651 unsigned int nob = 0;
3654 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3655 * However I will be rescheduled by an FMA completion event
3656 * when I eventually get some.
3657 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3658 * But it doesn't matter if I try to send a "real" message just
3659 * as I start closing because I'll get scheduled to send the
3662 /* Short circuit if the ep_handle is null we cant send anyway. */
3663 if (conn->gnc_ephandle == NULL)
3666 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3668 spin_lock(&conn->gnc_list_lock);
3670 if (list_empty(&conn->gnc_fmaq)) {
3671 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3673 spin_unlock(&conn->gnc_list_lock);
3675 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3676 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3677 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3678 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3679 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3681 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3682 conn->gnc_last_noop_cq,
3683 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3684 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3685 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3686 atomic_inc(&conn->gnc_sched_noop);
3687 set_mb(conn->gnc_last_noop_want, jiffies);
3689 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3692 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP,
3693 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid));
3697 rc = kgnilnd_set_tx_id(tx, conn);
3699 kgnilnd_tx_done(tx, rc);
3705 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3706 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3707 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3708 more_to_do = !list_empty(&conn->gnc_fmaq);
3709 spin_unlock(&conn->gnc_list_lock);
3712 /* if there is no real TX or no NOOP to send, bail */
3717 if (!tx->tx_retrans)
3718 tx->tx_cred_wait = jiffies;
3720 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3721 "tx with zero id", NULL);
3723 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3724 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3725 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3729 switch (tx->tx_msg.gnm_type) {
3733 case GNILND_MSG_NOOP:
3734 case GNILND_MSG_CLOSE:
3735 case GNILND_MSG_IMMEDIATE:
3736 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3737 buffer = tx->tx_buffer;
3741 case GNILND_MSG_GET_DONE:
3742 case GNILND_MSG_PUT_DONE:
3743 case GNILND_MSG_PUT_DONE_REV:
3744 case GNILND_MSG_GET_DONE_REV:
3745 case GNILND_MSG_PUT_NAK:
3746 case GNILND_MSG_GET_NAK:
3747 case GNILND_MSG_GET_NAK_REV:
3748 case GNILND_MSG_PUT_NAK_REV:
3749 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3752 case GNILND_MSG_PUT_REQ:
3753 case GNILND_MSG_GET_REQ_REV:
3754 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3756 case GNILND_MSG_PUT_ACK:
3757 case GNILND_MSG_PUT_REQ_REV:
3758 case GNILND_MSG_GET_ACK_REV:
3759 case GNILND_MSG_GET_REQ:
3760 /* This is really only to handle the retransmit of SMSG once these
3761 * two messages are setup in send_mapped_tx */
3762 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3766 if (likely(rc == 0)) {
3767 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3771 /* don't explicitly reschedule here - we are short credits and will rely on
3772 * kgnilnd_sendmsg to resched the conn if need be */
3774 } else if (rc < 0) {
3775 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3776 * almost certainly a software bug, but lets play nice with the other kids */
3777 kgnilnd_tx_done(tx, rc);
3778 /* just for fun, kick peer in arse - resetting conn might help to correct
3779 * this almost certainly buggy software caused return code */
3780 kgnilnd_close_conn(conn, rc);
3784 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3785 kgnilnd_schedule_conn(conn);
3790 kgnilnd_process_rdmaq(kgn_device_t *dev)
3795 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3799 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3800 unsigned long dead_bump;
3803 /* if we think we need to adjust, take lock to serialize and recheck */
3804 spin_lock(&dev->gnd_rdmaq_lock);
3805 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3806 timer_delete_sync(&dev->gnd_rdmaq_timer);
3808 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3810 /* roll the bucket forward */
3811 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3813 if (kgnilnd_data.kgn_rdmaq_override &&
3814 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3815 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3820 /* roll current outstanding forward to make sure we carry outstanding
3821 * committment forward
3822 * new_ok starts out as the whole interval value
3823 * - first subtract bytes_out from last interval, as that would push us over
3824 * strict limits for this interval
3825 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3827 * there is a small race here if someone is actively processing mappings and
3828 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3829 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3830 * - if this gives us problems in the future, we could use a read/write lock
3831 * to protect the resetting of these values */
3832 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3833 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3835 CDEBUG(D_NET, "resetting rdmaq bytes to %lld, deadline +%lu -> %lu, current out %lld\n",
3836 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3837 (s64)atomic64_read(&dev->gnd_rdmaq_bytes_out));
3839 spin_unlock(&dev->gnd_rdmaq_lock);
3842 spin_lock(&dev->gnd_rdmaq_lock);
3843 while (!list_empty(&dev->gnd_rdmaq)) {
3846 /* make sure we break out early on quiesce */
3847 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3848 /* always break with lock held - we unlock outside loop */
3852 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3853 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3856 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3857 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3858 /* if conn is dying, mark tx in tx_ref_table for
3859 * kgnilnd_complete_closed_conn to finish up */
3860 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3862 /* tx was moved to DYING, get next */
3865 spin_unlock(&dev->gnd_rdmaq_lock);
3867 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3868 spin_lock(&dev->gnd_rdmaq_lock);
3871 /* no ticket! add back to head */
3872 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3873 /* clear found_work so scheduler threads wait for timer */
3877 /* TX is GO for launch */
3878 tx->tx_qtime = jiffies;
3879 kgnilnd_send_mapped_tx(tx, 0);
3883 spin_unlock(&dev->gnd_rdmaq_lock);
3889 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3891 __swab64s(&d->gnrd_key.qword1);
3892 __swab64s(&d->gnrd_key.qword2);
3893 __swab64s(&d->gnrd_addr);
3894 __swab32s(&d->gnrd_nob);
3897 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3898 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3901 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3903 kgn_tx_ev_id_t ev_id;
3906 /* we use the cookie from the original TX, so we can find the match
3907 * by parsing that and using the txe_idx */
3908 ev_id.txe_cookie = cookie;
3910 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
3913 /* check tx to make sure kgni didn't eat it */
3914 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
3915 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
3917 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
3918 (tx->tx_id.txe_cookie = cookie)),
3919 "conn 0x%p->%s tx_ref_table hosed: wanted "
3920 "txe_cookie %#llx txe_idx %d "
3921 "found tx %p cookie %#llx txe_idx %d\n",
3922 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
3923 cookie, ev_id.txe_idx,
3924 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
3926 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
3927 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
3928 "Unexpected TX type (%x, %x or %x) "
3929 "or state (%x, expected +%x) "
3930 "matched reply from %s\n",
3931 tx->tx_msg.gnm_type, type1, type2,
3932 tx->tx_state, GNILND_TX_WAITING_REPLY,
3933 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3935 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
3936 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
3942 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
3945 kgn_conn_t *conn = tx->tx_conn;
3946 __u64 nob = tx->tx_nob;
3947 __u32 physnop = tx->tx_phys_npages;
3948 int id = tx->tx_id.txe_smsg_id;
3949 int buftype = tx->tx_buftype;
3950 gni_mem_handle_t hndl;
3951 hndl.qword1 = tx->tx_map_key.qword1;
3952 hndl.qword2 = tx->tx_map_key.qword2;
3954 spin_lock(&conn->gnc_list_lock);
3956 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
3957 "not waiting for reply", NULL);
3960 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
3962 if (rc == -EFAULT) {
3963 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
3964 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
3966 if(*kgnilnd_tunables.kgn_efault_lbug) {
3967 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
3968 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
3970 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
3971 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
3972 kgnilnd_tx_state2str(tx->tx_list_state),
3973 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
3978 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
3979 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3980 /* sample under lock as follow on steps require gnc_list_lock
3981 * - or call kgnilnd_tx_done which requires no locks held over
3982 * call to lnet_finalize */
3985 spin_unlock(&conn->gnc_list_lock);
3988 kgnilnd_tx_done(tx, tx->tx_rc);
3993 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
3996 kgn_conn_t *conn = tx->tx_conn;
3998 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
3999 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4001 /* the gncm_retval is passed in for PUTs */
4002 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4003 msg->gnm_u.completion.gncm_retval);
4005 kgnilnd_complete_tx(tx, rc);
4009 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4017 kgn_peer_t *peer = conn->gnc_peer;
4020 __u16 tmp_cksum = 0, msg_cksum = 0;
4021 int repost = 1, saw_complete;
4022 unsigned long timestamp, newest_last_rx, timeout;
4024 struct lnet_hdr hdr;
4025 struct lnet_nid srcnid;
4028 /* Short circuit if the ep_handle is null.
4029 * It's likely that its about to be closed as stale.
4031 if (conn->gnc_ephandle == NULL)
4034 timestamp = jiffies;
4035 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4036 /* delay in jiffies - we are really concerned only with things that
4037 * result in a schedule() or really holding this off for long times .
4038 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4039 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4041 /* Resample current time as we have no idea how long it took to get the mutex */
4042 timestamp = jiffies;
4044 /* We check here when the last time we received an rx, we do this before
4045 * we call getnext in case the thread has been blocked for a while. If we
4046 * havent received an rx since our timeout value we close the connection
4047 * as we should assume the other side has closed the connection. This will
4048 * stop us from sending replies to a mailbox that is already in purgatory.
4051 timeout = cfs_time_seconds(conn->gnc_timeout);
4052 newest_last_rx = GNILND_LASTRX(conn);
4054 /* Error injection to validate that timestamp checking works and closing the conn */
4055 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4056 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4059 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4060 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4061 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4062 cfs_duration_sec(timestamp - newest_last_rx),
4063 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4064 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4066 kgnilnd_close_conn(conn, rc);
4070 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4072 if (rrc == GNI_RC_NOT_DONE) {
4073 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4074 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4078 /* Instead of asserting when we get mailbox corruption lets attempt to
4079 * close the conn and recover. We can put the conn/mailbox into
4080 * purgatory and let purgatory deal with the problem. If we see
4081 * this NETTERROR reported on production systems in large amounts
4082 * we will need to revisit the state machine to see if we can tighten
4083 * it up further to improve data protection.
4086 if (rrc == GNI_RC_INVALID_STATE) {
4087 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4088 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4089 "detected closing conn %p from peer %s\n", conn,
4090 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4092 kgnilnd_close_conn(conn, rc);
4096 LASSERTF(rrc == GNI_RC_SUCCESS,
4097 "bad rc %d on conn %p from peer %s\n",
4098 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4100 msg = (kgn_msg_t *)prefix;
4102 rx = kgnilnd_alloc_rx();
4104 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4105 kgnilnd_release_msg(conn);
4106 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4107 conn, libcfs_nid2str(peer->gnp_nid));
4111 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4113 timestamp = conn->gnc_last_rx;
4114 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4115 atomic_inc(&conn->gnc_rx_seq);
4117 conn->gnc_last_rx = jiffies;
4118 /* stash first rx so we can clear out purgatory
4120 if (conn->gnc_first_rx == 0)
4121 conn->gnc_first_rx = jiffies;
4123 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4124 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4125 kgnilnd_peer_alive(conn->gnc_peer);
4128 rx->grx_conn = conn;
4130 ktime_get_ts64(&rx->grx_received);
4132 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4135 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4141 kgnilnd_net_decref(net);
4144 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4145 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4147 /* XXX Nic: Do we need to swab cksum */
4148 if (msg->gnm_cksum != 0) {
4149 msg_cksum = msg->gnm_cksum;
4151 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4153 if (tmp_cksum != msg_cksum) {
4154 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4155 tmp_cksum, msg_cksum);
4156 kgnilnd_dump_msg(D_BUFFS, msg);
4161 /* restore checksum for future debug messages */
4162 msg->gnm_cksum = tmp_cksum;
4164 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4165 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4166 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4167 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4172 __swab32s(&msg->gnm_magic);
4173 __swab16s(&msg->gnm_version);
4174 __swab16s(&msg->gnm_type);
4175 __swab64s(&msg->gnm_srcnid);
4176 __swab64s(&msg->gnm_connstamp);
4177 __swab32s(&msg->gnm_seq);
4179 /* NB message type checked below; NOT here... */
4180 switch (msg->gnm_type) {
4181 case GNILND_MSG_GET_ACK_REV:
4182 case GNILND_MSG_PUT_ACK:
4183 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4186 case GNILND_MSG_PUT_REQ_REV:
4187 case GNILND_MSG_GET_REQ:
4188 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4196 if (msg->gnm_version != GNILND_MSG_VERSION) {
4197 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4198 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4203 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4204 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4205 libcfs_nid2str(msg->gnm_srcnid),
4206 libcfs_nid2str(peer->gnp_nid));
4211 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4212 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4213 " expected) from %s",
4214 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4215 libcfs_nid2str(peer->gnp_nid));
4220 if (msg->gnm_seq != seq) {
4221 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4222 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4227 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4229 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4230 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4231 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4232 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4233 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4234 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4235 /* double check state with lock held */
4236 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4237 /* only error if we are not already closing */
4238 if (conn->gnc_peer_error == -ETIMEDOUT) {
4239 unsigned long now = jiffies;
4240 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4242 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4243 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4244 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4246 cfs_duration_sec(now - timestamp),
4247 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4248 atomic_read(&conn->gnc_tx_seq),
4249 cfs_duration_sec(now - conn->gnc_last_tx),
4250 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4251 cfs_duration_sec(now - conn->gnc_last_noop_want),
4252 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4253 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4254 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4255 cfs_duration_sec(now - conn->gnc_last_sched_do),
4256 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4258 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4260 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4264 if (conn->gnc_close_recvd) {
4265 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4266 kgnilnd_msgtype2str(msg->gnm_type),
4267 msg->gnm_type, conn->gnc_close_recvd,
4268 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4273 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4274 /* XXX Nic: log message received on bad connection state */
4278 switch (msg->gnm_type) {
4279 case GNILND_MSG_NOOP:
4280 /* Nothing to do; just a keepalive */
4283 case GNILND_MSG_IMMEDIATE:
4284 /* only get SMSG payload for IMMEDIATE */
4285 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4286 lnet_hdr_from_nid4(&hdr, &msg->gnm_u.immediate.gnim_hdr);
4287 lnet_nid4_to_nid(msg->gnm_srcnid, &srcnid);
4288 rc = lnet_parse(net->gnn_ni, &hdr, &srcnid, rx, 0);
4291 case GNILND_MSG_GET_REQ_REV:
4292 case GNILND_MSG_PUT_REQ:
4293 lnet_hdr_from_nid4(&hdr, &msg->gnm_u.putreq.gnprm_hdr);
4294 lnet_nid4_to_nid(msg->gnm_srcnid, &srcnid);
4295 rc = lnet_parse(net->gnn_ni, &hdr, &srcnid, rx, 1);
4298 case GNILND_MSG_GET_NAK_REV:
4299 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4300 msg->gnm_u.completion.gncm_cookie);
4304 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4306 case GNILND_MSG_PUT_NAK:
4307 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4308 msg->gnm_u.completion.gncm_cookie);
4312 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4314 case GNILND_MSG_PUT_ACK:
4315 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4316 msg->gnm_u.putack.gnpam_src_cookie);
4320 /* store putack data for later: deferred rdma or re-try */
4321 tx->tx_putinfo = msg->gnm_u.putack;
4324 spin_lock(&tx->tx_conn->gnc_list_lock);
4326 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4327 "not waiting for reply", NULL);
4329 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4331 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4332 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4333 /* sample under lock as follow on steps require gnc_list_lock
4334 * - or call kgnilnd_tx_done which requires no locks held over
4335 * call to lnet_finalize */
4338 /* cannot launch rdma if still waiting for fma-msg completion */
4339 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4340 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4341 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4343 spin_unlock(&tx->tx_conn->gnc_list_lock);
4346 rc = kgnilnd_send_mapped_tx(tx, 0);
4348 kgnilnd_tx_done(tx, rc);
4351 case GNILND_MSG_GET_ACK_REV:
4352 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4353 msg->gnm_u.putack.gnpam_src_cookie);
4357 /* store putack data for later: deferred rdma or re-try */
4358 tx->tx_putinfo = msg->gnm_u.putack;
4360 spin_lock(&tx->tx_conn->gnc_list_lock);
4362 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4363 "not waiting for reply", NULL);
4365 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4367 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4368 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4369 /* sample under lock as follow on steps require gnc_list_lock
4370 * - or call kgnilnd_tx_done which requires no locks held over
4371 * call to lnet_finalize */
4374 /* cannot launch rdma if still waiting for fma-msg completion */
4375 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4376 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4377 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4379 spin_unlock(&tx->tx_conn->gnc_list_lock);
4382 rc = kgnilnd_send_mapped_tx(tx, 0);
4384 kgnilnd_tx_done(tx, rc);
4387 case GNILND_MSG_PUT_DONE:
4388 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4389 msg->gnm_u.completion.gncm_cookie);
4393 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4394 "bad tx buftype %d", tx->tx_buftype);
4396 kgnilnd_finalize_rx_done(tx, msg);
4398 case GNILND_MSG_PUT_REQ_REV:
4399 case GNILND_MSG_GET_REQ:
4400 lnet_hdr_from_nid4(&hdr, &msg->gnm_u.get.gngm_hdr);
4401 lnet_nid4_to_nid(msg->gnm_srcnid, &srcnid);
4402 rc = lnet_parse(net->gnn_ni, &hdr, &srcnid, rx, 1);
4406 case GNILND_MSG_GET_NAK:
4407 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4408 msg->gnm_u.completion.gncm_cookie);
4412 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4413 "bad tx buftype %d", tx->tx_buftype);
4415 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4418 case GNILND_MSG_GET_DONE:
4419 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4420 msg->gnm_u.completion.gncm_cookie);
4424 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4425 "bad tx buftype %d", tx->tx_buftype);
4427 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4428 msg->gnm_u.completion.gncm_retval);
4430 kgnilnd_finalize_rx_done(tx, msg);
4432 case GNILND_MSG_GET_DONE_REV:
4433 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4434 msg->gnm_u.completion.gncm_cookie);
4438 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4439 "bad tx buftype %d", tx->tx_buftype);
4441 kgnilnd_finalize_rx_done(tx, msg);
4444 case GNILND_MSG_PUT_DONE_REV:
4445 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4446 msg->gnm_u.completion.gncm_cookie);
4451 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4452 "bad tx buftype %d", tx->tx_buftype);
4454 kgnilnd_finalize_rx_done(tx, msg);
4456 case GNILND_MSG_PUT_NAK_REV:
4457 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4458 msg->gnm_u.completion.gncm_cookie);
4463 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED,
4464 "bad tx buftype %d", tx->tx_buftype);
4466 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4471 if (rc < 0) /* protocol/comms error */
4472 kgnilnd_close_conn(conn, rc);
4474 if (repost && rx != NULL) {
4475 kgnilnd_consume_rx(rx);
4478 /* we got an event so assume more there and call for reschedule */
4480 kgnilnd_schedule_conn(conn);
4484 /* Do the failure injections that we need to affect conn processing in the following function.
4485 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4486 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4489 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4493 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4497 /* short circuit out when not set */
4498 if (likely(!cfs_fail_loc)) {
4502 /* failure injection to test for stack reset clean ups */
4503 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4504 /* we can't rely on busy loops being nice enough to get the
4505 * stack reset triggered - it'd just spin on this conn */
4506 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4509 GOTO(did_fail_loc, rc);
4512 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4513 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4515 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4516 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4519 GOTO(did_fail_loc, rc);
4523 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4524 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4525 * will be closed by kgnilnd_complete_closed_conn.
4527 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4528 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4531 GOTO(did_fail_loc, rc);
4534 /* this one is a bit gross - we can't hold the mutex from process_conns
4535 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4536 * and moving onto finish_connect
4537 * so, we'll just set the rc - kgnilnd_process_conns will clear
4538 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4539 * and effectively getting this thread to sleep */
4540 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4543 GOTO(did_fail_loc, rc);
4551 kgnilnd_send_conn_close(kgn_conn_t *conn)
4555 /* we are closing the conn - we will try to send the CLOSE msg
4556 * but will not wait for anything else to flush */
4558 /* send the close if not already done so or received one */
4559 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4560 /* set close_sent regardless of the success of the
4561 * CLOSE message. We are going to try once and then
4562 * kick him out of the sandbox */
4563 conn->gnc_close_sent = 1;
4566 /* EP might be null already if remote side initiated a new connection.
4567 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4568 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4570 if (conn->gnc_ephandle != NULL) {
4573 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE,
4574 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid));
4576 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4577 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4578 tx->tx_qtime = jiffies;
4580 if (tx->tx_id.txe_idx == 0) {
4581 rc = kgnilnd_set_tx_id(tx, conn);
4583 kgnilnd_tx_done(tx, rc);
4587 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4590 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4591 kgnilnd_tx_done(tx, -EAGAIN);
4593 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4595 /* It wasnt sent and we dont care. */
4596 kgnilnd_tx_done(tx, rc);
4604 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4605 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4606 conn->gnc_state = GNILND_CONN_CLOSED;
4607 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4608 /* mark this conn as CLOSED now that we processed it
4609 * do after TX, so we can use CLOSING in asserts */
4613 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4614 /* simulate a RX CLOSE after the timeout but before
4615 * the scheduler thread gets it */
4616 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4617 conn->gnc_peer_error = -ETIMEDOUT;
4619 /* schedule to allow potential CLOSE and get the complete phase run */
4620 kgnilnd_schedule_conn(conn);
4624 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4629 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4630 int log_retrans, log_retrans_level;
4631 static int last_map_version;
4634 spin_lock(&dev->gnd_lock);
4635 if (list_empty(&dev->gnd_map_tx)) {
4636 /* if the list is empty make sure we dont have a timer running */
4637 timer_delete_sync(&dev->gnd_map_timer);
4638 spin_unlock(&dev->gnd_lock);
4642 dev->gnd_sched_alive = jiffies;
4644 /* we'll retry as fast as possible up to 25% of the limit, then we start
4645 * backing off until our map version changes - indicating we unmapped
4647 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4648 if (likely(dev->gnd_map_attempt == 0) ||
4649 time_after_eq(jiffies, dev->gnd_next_map) ||
4650 last_map_version != dev->gnd_map_version) {
4652 /* if this is our first attempt at mapping set last mapped to current
4653 * jiffies so we can timeout our attempt correctly.
4655 if (dev->gnd_map_attempt == 0)
4656 dev->gnd_last_map = jiffies;
4658 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4659 spin_unlock(&dev->gnd_lock);
4663 /* delete the previous timer if it exists */
4664 timer_delete_sync(&dev->gnd_map_timer);
4665 /* stash the last map version to let us know when a good one was seen */
4666 last_map_version = dev->gnd_map_version;
4668 /* we need to to take the lock and continually refresh the head of the list as
4669 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4670 * allowing them to squeeze in */
4672 while (!list_empty(&dev->gnd_map_tx)) {
4673 /* make sure we break out early on quiesce */
4674 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4675 /* always break with lock held - we unlock outside loop */
4679 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4681 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4684 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4685 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4686 /* if conn is dying, mark tx in tx_ref_table for
4687 * kgnilnd_complete_closed_conn to finish up */
4688 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4691 /* tx was moved to DYING, get next */
4695 spin_unlock(&dev->gnd_lock);
4696 rc = kgnilnd_send_mapped_tx(tx, 1);
4698 /* We made it! skip error handling.. */
4700 /* OK to continue on +ve errors as it won't get seen until
4701 * this function is called again - we operate on a copy of the original
4702 * list and not the live list */
4703 spin_lock(&dev->gnd_lock);
4704 /* reset map attempts back to zero we successfully
4705 * mapped so we can reset our timers */
4706 dev->gnd_map_attempt = 0;
4708 } else if (rc == -EAGAIN) {
4709 spin_lock(&dev->gnd_lock);
4710 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4711 spin_unlock(&dev->gnd_lock);
4712 GOTO(get_out_mapped, rc);
4713 } else if (rc != -ENOMEM) {
4714 /* carp, failure we can't handle */
4715 kgnilnd_tx_done(tx, rc);
4716 spin_lock(&dev->gnd_lock);
4717 /* reset map attempts back to zero we dont know what happened but it
4718 * wasnt a failed mapping
4720 dev->gnd_map_attempt = 0;
4724 /* time to handle the retry cases.. lock so we dont have 2 threads
4725 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4727 spin_lock(&dev->gnd_lock);
4728 dev->gnd_map_attempt++;
4729 if (dev->gnd_map_attempt < fast_remaps) {
4730 /* do nothing we just want it to go as fast as possible.
4731 * just set gnd_next_map to current jiffies so it will process
4732 * as fast as possible.
4734 dev->gnd_next_map = jiffies;
4736 /* Retry based on GNILND_MAP_RETRY_RATE */
4737 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4740 /* only log occasionally once we've retried fast_remaps */
4741 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4742 ((dev->gnd_map_attempt % fast_remaps) == 0);
4743 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4745 /* make sure we are not off in the weeds with this tx */
4746 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4747 GNIDBG_TX(D_NETERROR, tx,
4748 "giving up on TX, too many retries", NULL);
4749 spin_unlock(&dev->gnd_lock);
4750 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4751 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4752 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4754 tx->tx_putinfo.gnpam_dst_cookie,
4755 tx->tx_msg.gnm_srcnid);
4757 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4759 tx->tx_getinfo.gngm_cookie,
4760 tx->tx_msg.gnm_srcnid);
4762 kgnilnd_tx_done(tx, -ENOMEM);
4763 GOTO(get_out_mapped, rc);
4765 GNIDBG_TX(log_retrans_level, tx,
4766 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4767 "nq_map %d mdd# %d/%d GART %ld",
4768 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4769 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4770 atomic_read(&dev->gnd_nq_map),
4771 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4772 atomic64_read(&dev->gnd_nbytes_map));
4775 /* we need to stop processing the rest of the list, so add it back in */
4776 /* set timer to wake device when we need to schedule this tx */
4777 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4778 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4779 spin_unlock(&dev->gnd_lock);
4780 GOTO(get_out_mapped, rc);
4782 spin_unlock(&dev->gnd_lock);
4788 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4793 int error_inject = 0;
4797 spin_lock(&dev->gnd_lock);
4798 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4799 dev->gnd_sched_alive = jiffies;
4803 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4804 /* break with lock held */
4808 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4809 list_del_init(&conn->gnc_schedlist);
4811 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4814 if (!list_empty(&conn->gnc_delaylist))
4815 list_del_init(&conn->gnc_delaylist);
4816 spin_unlock(&dev->gnd_lock);
4818 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4820 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4821 conn_sched != GNILND_CONN_PROCESS,
4822 "conn %p on ready list but in bad state: %d\n",
4825 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4826 conn, kgnilnd_conn_state2str(conn));
4829 set_mb(conn->gnc_last_sched_do, jiffies);
4831 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4833 /* based on intent see if we should run again. */
4834 rc = kgnilnd_schedule_process_conn(conn, intent);
4836 /* drop ref from gnd_ready_conns */
4837 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4838 down_write(&dev->gnd_conn_sem);
4839 kgnilnd_conn_decref(conn);
4840 up_write(&dev->gnd_conn_sem);
4841 } else if (rc != 1) {
4842 kgnilnd_conn_decref(conn);
4844 /* clear this so that scheduler thread doesn't spin */
4846 /* break with lock held... */
4847 spin_lock(&dev->gnd_lock);
4851 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4852 down_write(&dev->gnd_conn_sem);
4854 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4855 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4856 /* If there are tx's currently in use in another
4857 * thread we dont want to complete the close
4858 * yet. Cycle this conn back through
4860 kgnilnd_schedule_conn(conn);
4862 kgnilnd_complete_closed_conn(conn);
4864 up_write(&dev->gnd_conn_sem);
4865 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4866 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4867 /* serialize SMSG CQs with ep_bind and smsg_release */
4868 down_write(&dev->gnd_conn_sem);
4869 kgnilnd_destroy_conn_ep(conn);
4870 up_write(&dev->gnd_conn_sem);
4871 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4872 /* if we need to do some CLOSE sending, etc done here do it */
4873 down_write(&dev->gnd_conn_sem);
4874 kgnilnd_send_conn_close(conn);
4875 kgnilnd_check_fma_rx(conn);
4876 up_write(&dev->gnd_conn_sem);
4877 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4878 /* start moving traffic if the old conns are cleared out */
4879 down_read(&dev->gnd_conn_sem);
4880 kgnilnd_check_fma_rx(conn);
4881 kgnilnd_process_fmaq(conn);
4882 up_read(&dev->gnd_conn_sem);
4885 rc = kgnilnd_schedule_process_conn(conn, 0);
4887 /* drop ref from gnd_ready_conns */
4888 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4889 down_write(&dev->gnd_conn_sem);
4890 kgnilnd_conn_decref(conn);
4891 up_write(&dev->gnd_conn_sem);
4892 } else if (rc != 1) {
4893 kgnilnd_conn_decref(conn);
4896 /* check list again with lock held */
4897 spin_lock(&dev->gnd_lock);
4900 /* If we are short circuiting due to timing we want to be scheduled
4901 * as soon as possible.
4903 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
4906 spin_unlock(&dev->gnd_lock);
4912 kgnilnd_scheduler(void *arg)
4914 int threadno = (long)arg;
4917 unsigned long deadline = 0;
4920 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
4922 /* all gnilnd threads need to run fairly urgently */
4923 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
4924 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
4925 while (!kgnilnd_data.kgn_shutdown) {
4927 /* Safe: kgn_shutdown only set when quiescent */
4929 /* to quiesce or to not quiesce, that is the question */
4931 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4932 KGNILND_SPIN_QUIESCE;
4935 /* tracking for when thread goes AWOL */
4936 dev->gnd_sched_alive = jiffies;
4938 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
4939 (*kgnilnd_tunables.kgn_sched_timeout + 1));
4940 /* let folks know we are up and kicking
4941 * - they can use this for latency savings, etc
4942 * - only change if IRQ, if IDLE leave alone as that
4943 * schedule_device calls to put us back to IRQ */
4944 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
4946 down_read(&dev->gnd_conn_sem);
4947 /* always check these - they are super low cost */
4948 found_work += kgnilnd_check_fma_send_cq(dev);
4949 found_work += kgnilnd_check_fma_rcv_cq(dev);
4951 /* rdma CQ doesn't care about eps */
4952 found_work += kgnilnd_check_rdma_cq(dev);
4954 /* move some RDMA ? */
4955 found_work += kgnilnd_process_rdmaq(dev);
4957 /* map some pending RDMA requests ? */
4958 found_work += kgnilnd_process_mapped_tx(dev);
4960 /* the EP for a conn is not destroyed until all the references
4961 * to it are gone, so these checks should be safe
4962 * even if run in parallel with the CQ checking functions
4963 * _AND_ a thread that processes the CLOSED->DONE
4967 up_read(&dev->gnd_conn_sem);
4969 /* process all conns ready now */
4970 found_work += kgnilnd_process_conns(dev, deadline);
4972 /* do an eager check to avoid the IRQ disabling in
4973 * prepare_to_wait and friends */
4976 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
4977 time_before(jiffies, deadline)) {
4979 if ((busy_loops % 10) == 0) {
4980 /* tickle heartbeat and watchdog to ensure our
4981 * piggishness doesn't turn into heartbeat failure */
4982 touch_nmi_watchdog();
4988 /* if we got here, found_work was zero or busy_loops means we
4989 * need to take a break. We'll clear gnd_ready but we'll check
4990 * one last time if there is an IRQ that needs processing */
4992 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
4994 /* the first time this will go LOOP -> IDLE and let us do one final check
4995 * during which we might get an IRQ, then IDLE->IDLE and schedule()
4996 * - this might allow other threads to block us for a bit if they
4997 * try to get the mutex, but that is good as we'd need to wake
4998 * up soon to handle the CQ or other processing anyways */
5000 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5002 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5003 time_after_eq(jiffies, deadline)) {
5005 "yeilding: found_work %d busy_loops %d\n",
5006 found_work, busy_loops);
5008 /* use yield if we are bailing due to busy_loops
5009 * - this will ensure we wake up soonish. This closes
5010 * a race with kgnilnd_device_callback - where it'd
5011 * not call wake_up() because gnd_ready == 1, but then
5012 * we come down and schedule() because of busy_loops.
5013 * We'd not be woken up until something poked our waitq
5014 * again. yield() ensures we wake up without another
5015 * waitq poke in that case */
5016 atomic_inc(&dev->gnd_n_yield);
5017 kgnilnd_data.kgn_last_condresched = jiffies;
5019 CDEBUG(D_INFO, "awake after yeild\n");
5020 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5021 } else if (found_work == GNILND_DEV_IDLE) {
5022 /* busy_loops is low and there is nothing to do,
5023 * go to sleep and wait for a waitq poke */
5025 "scheduling: found_work %d busy_loops %d\n",
5026 found_work, busy_loops);
5027 atomic_inc(&dev->gnd_n_schedule);
5028 kgnilnd_data.kgn_last_scheduled = jiffies;
5030 CDEBUG(D_INFO, "awake after schedule\n");
5031 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5033 finish_wait(&dev->gnd_waitq, &wait);
5036 kgnilnd_thread_fini();