2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by Eric Barton <eric@bartonsoftware.com>
7 * Author: James Shimek <jshimek@cray.com>
8 * Author: Nic Henke <nic@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #include <linux/nmi.h>
29 #include <linux/pagemap.h>
31 #include <libcfs/linux/linux-mem.h>
35 /* this is useful when needed to debug wire corruption. */
37 kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
45 "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
46 prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
49 } else if (len >= 16) {
51 "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
52 prefix, ptr, *(ptr), *(ptr + 1));
56 CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
65 kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
67 CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
68 " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
69 msg->gnm_magic, msg->gnm_version,
70 msg->gnm_type, msg->gnm_srcnid,
71 msg->gnm_connstamp, msg->gnm_seq,
72 msg->gnm_cksum, msg->gnm_payload_cksum,
73 msg->gnm_payload_len);
77 kgnilnd_schedule_device(kgn_device_t *dev)
79 short already_live = 0;
81 /* we'll only want to wake if the scheduler thread
82 * has come around and set ready to zero */
83 already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);
86 wake_up_all(&dev->gnd_waitq);
90 void kgnilnd_schedule_device_timer(unsigned long arg)
92 kgn_device_t *dev = (kgn_device_t *) arg;
94 kgnilnd_schedule_device(dev);
98 kgnilnd_device_callback(__u32 devid, __u64 arg)
101 int index = (int) arg;
103 if (index >= kgnilnd_data.kgn_ndevs) {
104 /* use _EMERG instead of an LBUG to prevent LBUG'ing in
105 * interrupt context. */
106 LCONSOLE_EMERG("callback for unknown device %d->%d\n",
111 dev = &kgnilnd_data.kgn_devices[index];
112 /* just basic sanity */
113 if (dev->gnd_id == devid) {
114 kgnilnd_schedule_device(dev);
116 LCONSOLE_EMERG("callback for bad device %d devid %d\n",
121 /* sched_intent values:
122 * < 0 : do not reschedule under any circumstances
123 * == 0: reschedule if someone marked him WANTS_SCHED
124 * > 0 : force a reschedule */
125 /* Return code 0 means it did not schedule the conn, 1
126 * means it successfully scheduled the conn.
130 kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
134 /* move back to IDLE but save previous state.
135 * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
136 * let the xchg there handle any racing callers to get it
137 * onto gnd_ready_conns */
139 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
140 LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
141 conn_sched == GNILND_CONN_PROCESS,
142 "conn %p after process in bad state: %d\n",
145 if (sched_intent >= 0) {
146 if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
147 return kgnilnd_schedule_conn_refheld(conn, 1);
153 /* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
157 _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held)
159 kgn_device_t *dev = conn->gnc_device;
163 sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
164 /* we only care about the last person who marked want_sched since they
165 * are most likely the culprit
167 memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
168 conn->gnc_sched_line = line;
169 /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
170 * the chance to put it onto gnd_ready_conns.
171 * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
172 * the conn in process_conns will take care of moving it back to
173 * SCHED when it is done processing */
175 if (sched == GNILND_CONN_IDLE) {
176 /* if the conn is already scheduled, we've already requested
177 * the scheduler thread wakeup */
179 /* Add a reference to the conn if we are not holding a reference
180 * already from the exisiting scheduler. We now use the same
181 * reference if we need to reschedule a conn while in a scheduler
184 kgnilnd_conn_addref(conn);
186 LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
189 CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);
191 spin_lock(&dev->gnd_lock);
192 list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
194 spin_unlock(&dev->gnd_lock);
195 set_mb(conn->gnc_last_sched_ask, jiffies);
198 CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
202 /* make sure thread(s) going to process conns - but let it make
203 * separate decision from conn schedule */
205 kgnilnd_schedule_device(dev);
210 _kgnilnd_schedule_delay_conn(kgn_conn_t *conn)
212 kgn_device_t *dev = conn->gnc_device;
214 spin_lock(&dev->gnd_lock);
215 if (list_empty(&conn->gnc_delaylist)) {
216 list_add_tail(&conn->gnc_delaylist, &dev->gnd_delay_conns);
219 spin_unlock(&dev->gnd_lock);
221 kgnilnd_schedule_device(dev);
226 kgnilnd_schedule_dgram(kgn_device_t *dev)
230 wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
231 if (wake != GNILND_DGRAM_SCHED) {
232 wake_up(&dev->gnd_dgram_waitq);
234 CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
239 kgnilnd_free_tx(kgn_tx_t *tx)
241 /* taken from kgnilnd_tx_add_state_locked */
243 LASSERTF((tx->tx_list_p == NULL &&
244 tx->tx_list_state == GNILND_TX_ALLOCD) &&
245 list_empty(&tx->tx_list),
246 "tx %p with bad state %s (list_p %p) tx_list %s\n",
247 tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
248 list_empty(&tx->tx_list) ? "empty" : "not empty");
250 atomic_dec(&kgnilnd_data.kgn_ntx);
252 /* we only allocate this if we need to */
253 if (tx->tx_phys != NULL) {
254 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
255 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
256 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
259 /* Only free the buffer if we used it */
260 if (tx->tx_buffer_copy != NULL) {
261 kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
262 tx->tx_buffer_copy = NULL;
263 CDEBUG(D_MALLOC, "vfreed buffer2\n");
266 KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
268 CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
269 kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
273 kgnilnd_alloc_tx (void)
277 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
280 tx = kmem_cache_alloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
282 CERROR("failed to allocate tx\n");
285 CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
288 /* need this memset, cache alloc'd memory is not cleared */
289 memset(tx, 0, sizeof(*tx));
291 /* setup everything here to minimize time under the lock */
292 tx->tx_buftype = GNILND_BUF_NONE;
293 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
294 INIT_LIST_HEAD(&tx->tx_list);
295 INIT_LIST_HEAD(&tx->tx_map_list);
296 tx->tx_list_state = GNILND_TX_ALLOCD;
298 atomic_inc(&kgnilnd_data.kgn_ntx);
303 /* csum_fold needs to be run on the return value before shipping over the wire */
304 #define _kgnilnd_cksum(seed, ptr, nob) csum_partial(ptr, nob, seed)
306 /* we don't use offset as every one is passing a buffer reference that already
307 * includes the offset into the base address -
308 * see kgnilnd_setup_virt_buffer and kgnilnd_setup_immediate_buffer */
310 kgnilnd_cksum(void *ptr, size_t nob)
314 sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));
316 /* don't use magic 'no checksum' value */
320 CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
327 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
328 unsigned int offset, unsigned int nob, int dump_blob)
334 unsigned int fraglen;
340 CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
341 kiov, nkiov, offset, nob, dump_blob);
343 /* if loops changes, please change kgnilnd_setup_phys_buffer */
345 while (offset >= kiov->bv_len) {
346 offset -= kiov->bv_len;
352 /* ignore nob here, if nob < (bv_len - offset), kiov == 1 */
353 odd = (unsigned long) (kiov[0].bv_len - offset) & 1;
355 if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
356 struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];
358 LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
359 get_cpu(), kgnilnd_data.kgn_cksum_map_pages);
361 CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
362 odd, kiov[0].bv_len, offset, nob);
364 for (i = 0; i < nkiov; i++) {
365 pages[i] = kiov[i].bv_page;
368 addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
370 CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
372 /* return zero to avoid killing tx - we'll just get warning on console
373 * when remote end sees zero checksum */
376 atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);
378 tmpck = _kgnilnd_cksum(0, ((void *) addr + kiov[0].bv_offset +
383 kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
384 (void *)addr + kiov[0].bv_offset +
387 CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
388 cksum, tmpck, addr, kiov[0].bv_offset, nob, offset);
392 fraglen = min(kiov->bv_len - offset, nob);
394 /* make dang sure we don't send a bogus checksum if somehow we get
395 * an odd length fragment on anything but the last entry in a kiov -
396 * we know from kgnilnd_setup_rdma_buffer that we can't have non
397 * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
398 LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
399 "odd fraglen %u on nkiov %d, nob %u bv_len %u offset %u kiov 0x%p\n",
400 fraglen, nkiov, nob, kiov->bv_len,
403 addr = (void *)kmap(kiov->bv_page) + kiov->bv_offset +
405 tmpck = _kgnilnd_cksum(cksum, addr, fraglen);
408 "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
409 cksum, tmpck, kiov->bv_page, kiov->bv_offset,
410 addr, fraglen, offset);
415 kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);
417 kunmap(kiov->bv_page);
424 /* iov must not run out before end of data */
425 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
430 retsum = csum_fold(cksum);
432 /* don't use magic 'no checksum' value */
436 CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);
442 kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
444 msg->gnm_magic = GNILND_MSG_MAGIC;
445 msg->gnm_version = GNILND_MSG_VERSION;
446 msg->gnm_type = type;
447 msg->gnm_payload_len = 0;
448 msg->gnm_srcnid = source;
449 /* gnm_connstamp gets set when FMA is sent */
450 /* gnm_srcnid is set on creation via function argument
451 * The right interface/net and nid is passed in when the message
457 kgnilnd_new_tx_msg(int type, lnet_nid_t source)
459 kgn_tx_t *tx = kgnilnd_alloc_tx();
462 kgnilnd_init_msg(&tx->tx_msg, type, source);
464 CERROR("couldn't allocate new tx type %s!\n",
465 kgnilnd_msgtype2str(type));
472 kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
478 case GNILND_MSG_GET_REQ:
479 case GNILND_MSG_GET_DONE:
480 nak_type = GNILND_MSG_GET_NAK;
482 case GNILND_MSG_PUT_REQ:
483 case GNILND_MSG_PUT_ACK:
484 case GNILND_MSG_PUT_DONE:
485 nak_type = GNILND_MSG_PUT_NAK;
487 case GNILND_MSG_PUT_REQ_REV:
488 case GNILND_MSG_PUT_DONE_REV:
489 nak_type = GNILND_MSG_PUT_NAK_REV;
491 case GNILND_MSG_GET_REQ_REV:
492 case GNILND_MSG_GET_ACK_REV:
493 case GNILND_MSG_GET_DONE_REV:
494 nak_type = GNILND_MSG_GET_NAK_REV;
497 CERROR("invalid msg type %s (%d)\n",
498 kgnilnd_msgtype2str(rx_type), rx_type);
501 /* only allow NAK on error and truncate to zero */
502 LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
503 error, conn, cookie);
505 tx = kgnilnd_new_tx_msg(nak_type, source);
507 CNETERR("can't get TX to NAK RDMA to %s\n",
508 libcfs_nid2str(conn->gnc_peer->gnp_nid));
512 tx->tx_msg.gnm_u.completion.gncm_retval = error;
513 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
514 kgnilnd_queue_tx(conn, tx);
518 kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov,
519 struct kvec *iov, struct bio_vec *kiov,
520 unsigned int offset, unsigned int nob)
522 kgn_msg_t *msg = &tx->tx_msg;
525 /* To help save on MDDs for short messages, we'll vmap a kiov to allow
526 * gni_smsg_send to send that as the payload */
528 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
531 tx->tx_buffer = NULL;
532 } else if (kiov != NULL) {
534 if ((niov > 0) && unlikely(niov > (nob/PAGE_SIZE))) {
535 niov = round_up(nob + offset + kiov->bv_offset,
539 LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
540 "bad niov %d msg %p kiov %p iov %p offset %d nob%d\n",
541 niov, msg, kiov, iov, offset, nob);
543 while (offset >= kiov->bv_len) {
544 offset -= kiov->bv_len;
549 for (i = 0; i < niov; i++) {
550 /* We can't have a bv_offset on anything but the first
551 * entry, otherwise we'll have a hole at the end of the
552 * mapping as we only map whole pages.
553 * Also, if we have a bv_len < PAGE_SIZE but we need to
554 * map more than bv_len, we will also have a whole at
555 * the end of that page which isn't allowed
557 if ((kiov[i].bv_offset != 0 && i > 0) ||
558 (kiov[i].bv_offset + kiov[i].bv_len != PAGE_SIZE &&
560 CNETERR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
561 i, offset, nob, kiov->bv_offset,
565 tx->tx_imm_pages[i] = kiov[i].bv_page;
568 /* hijack tx_phys for the later unmap */
570 /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
572 tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) +
573 kiov[0].bv_offset + offset;
574 atomic_inc(&kgnilnd_data.kgn_nkmap_short);
575 GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
576 nob, kiov, tx->tx_buffer);
578 tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
579 if (tx->tx_phys == NULL) {
580 CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
584 atomic_inc(&kgnilnd_data.kgn_nvmap_short);
585 /* make sure we take into account the kiov offset as the
586 * start of the buffer
588 tx->tx_buffer = (void *)tx->tx_phys + kiov[0].bv_offset
591 "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
592 niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
594 tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
598 /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
599 * could "flatten" the payload into a single contiguous buffer ready
600 * for sending direct over an FMA if we ever needed to. */
604 while (offset >= iov->iov_len) {
605 offset -= iov->iov_len;
611 if (nob > iov->iov_len - offset) {
612 CERROR("Can't handle multiple vaddr fragments\n");
616 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
618 tx->tx_buftype = GNILND_BUF_IMMEDIATE;
622 /* checksum payload early - it shouldn't be changing after lnd_send */
623 if (*kgnilnd_tunables.kgn_checksum >= 2) {
624 msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
625 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
626 msg->gnm_payload_cksum += 0xe00e;
628 if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
629 kgnilnd_dump_blob(D_BUFFS, "payload checksum",
633 msg->gnm_payload_cksum = 0;
640 kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
641 unsigned int niov, struct kvec *iov,
642 unsigned int offset, unsigned int nob)
647 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
649 while (offset >= iov->iov_len) {
650 offset -= iov->iov_len;
656 if (nob > iov->iov_len - offset) {
657 CERROR("Can't handle multiple vaddr fragments\n");
661 tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
663 tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
668 kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, struct bio_vec *kiov,
669 unsigned int offset, unsigned int nob)
671 gni_mem_segment_t *phys;
673 unsigned int fraglen;
675 GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);
679 LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
681 /* only allocate this if we are going to use it */
682 tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
684 if (tx->tx_phys == NULL) {
685 CERROR("failed to allocate tx_phys\n");
690 CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
691 LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
693 /* if loops changes, please change kgnilnd_cksum_kiov
694 * and kgnilnd_setup_immediate_buffer */
696 while (offset >= kiov->bv_len) {
697 offset -= kiov->bv_len;
703 /* at this point, kiov points to the first page that we'll actually map
704 * now that we've seeked into the koiv for offset and dropped any
705 * leading pages that fall entirely within the offset */
706 tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
709 /* bv_offset is start of 'valid' buffer, so index offset past that */
710 tx->tx_buffer = (void *)((unsigned long)(kiov->bv_offset + offset));
713 CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
714 tx, tx->tx_buffer, kiov, kiov->bv_offset, nkiov, offset);
717 fraglen = min(kiov->bv_len - offset, nob);
719 /* We can't have a bv_offset on anything but the first entry,
720 * otherwise we'll have a hole at the end of the mapping as we
721 * only map whole pages. Only the first page is allowed to
722 * have an offset - we'll add that into tx->tx_buffer and that
723 * will get used when we map in the segments (see
724 * kgnilnd_map_buffer). Also, if we have a bv_len < PAGE_SIZE
725 * but we need to map more than bv_len, we will also have a
726 * whole at the end of that page which isn't allowed
728 if ((phys != tx->tx_phys) &&
729 ((kiov->bv_offset != 0) ||
730 ((kiov->bv_len < PAGE_SIZE) && (nob > kiov->bv_len)))) {
731 CERROR("Can't make payload contiguous in I/O VM:page %d, offset %u, nob %u, bv_offset %u bv_len %u\n",
732 (int)(phys - tx->tx_phys),
733 offset, nob, kiov->bv_offset, kiov->bv_len);
738 if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
739 CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
744 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
750 "page 0x%p bv_offset %u bv_len %u nob %u nkiov %u offset %u\n",
751 kiov->bv_page, kiov->bv_offset, kiov->bv_len, nob, nkiov,
754 phys->address = page_to_phys(kiov->bv_page);
761 /* iov must not run out before end of data */
762 LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);
766 tx->tx_phys_npages = phys - tx->tx_phys;
771 if (tx->tx_phys != NULL) {
772 kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
773 CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
774 sizeof(*tx->tx_phys), tx->tx_phys);
781 kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
782 struct kvec *iov, struct bio_vec *kiov,
783 unsigned int offset, unsigned int nob)
787 LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
788 " offset %d, nob %d, niov %d\n"
789 , iov, kiov, tx, offset, nob, niov);
792 rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
794 rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
799 /* kgnilnd_parse_lnet_rdma()
800 * lntmsg - message passed in from lnet.
801 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
802 * nob - actual number of bytes to in this message.
803 * put_len - It is possible for PUTs to have a different length than the
804 * length stored in lntmsg->msg_len since LNET can adjust this
805 * length based on it's buffer size and offset.
806 * lnet_try_match_md() sets the mlength that we use to do the RDMA
810 kgnilnd_parse_lnet_rdma(struct lnet_msg *lntmsg, unsigned int *niov,
811 unsigned int *offset, unsigned int *nob,
812 struct bio_vec **kiov, int put_len)
814 /* GETs are weird, see kgnilnd_send */
815 if (lntmsg->msg_type == LNET_MSG_GET) {
816 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
819 *kiov = lntmsg->msg_md->md_kiov;
821 *niov = lntmsg->msg_md->md_niov;
822 *nob = lntmsg->msg_md->md_length;
825 *kiov = lntmsg->msg_kiov;
826 *niov = lntmsg->msg_niov;
828 *offset = lntmsg->msg_offset;
833 kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
835 unsigned int niov, offset, nob;
836 struct bio_vec *kiov;
837 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
838 int dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);
840 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
841 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
842 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
843 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
844 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
845 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
846 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
848 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
849 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
850 tx->tx_msg.gnm_payload_cksum = 0;
853 if (*kgnilnd_tunables.kgn_checksum < 3) {
854 tx->tx_msg.gnm_payload_cksum = 0;
858 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
860 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
864 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
866 tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
868 kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
872 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
873 tx->tx_msg.gnm_payload_cksum += 0xd00d;
877 /* kgnilnd_verify_rdma_cksum()
878 * tx - PUT_DONE/GET_DONE matched tx.
879 * rx_cksum - received checksum to compare against.
880 * put_len - see kgnilnd_parse_lnet_rdma comments.
883 kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
887 unsigned int niov, offset, nob;
888 struct bio_vec *kiov;
889 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
890 int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;
892 /* we can only match certain requests */
893 GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
894 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
895 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
896 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
897 (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
898 (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
899 "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));
901 if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
902 (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
907 if (*kgnilnd_tunables.kgn_checksum >= 3) {
908 GNIDBG_MSG(D_WARNING, &tx->tx_msg,
909 "no RDMA payload checksum when enabled");
914 GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);
916 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);
919 cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
921 cksum = kgnilnd_cksum(tx->tx_buffer, nob);
924 if (cksum != rx_cksum) {
925 GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
926 "Bad RDMA payload checksum (%x expected %x); "
927 "kiov 0x%p niov %d nob %u offset %u",
928 cksum, rx_cksum, kiov, niov, nob, offset);
929 switch (dump_on_err) {
932 kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
934 kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
937 /* fall through to dump log */
939 libcfs_debug_dumplog();
945 /* kgnilnd_check_fma_rx will close conn, kill tx with error */
951 kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
955 GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
956 "already mapped!", NULL);
958 spin_lock(&dev->gnd_map_lock);
959 switch (tx->tx_buftype) {
961 GNIDBG_TX(D_EMERG, tx,
962 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
963 spin_unlock(&dev->gnd_map_lock);
967 case GNILND_BUF_PHYS_MAPPED:
968 bytes = tx->tx_phys_npages * PAGE_SIZE;
969 dev->gnd_map_nphys++;
970 dev->gnd_map_physnop += tx->tx_phys_npages;
973 case GNILND_BUF_VIRT_MAPPED:
975 dev->gnd_map_nvirt++;
976 dev->gnd_map_virtnob += tx->tx_nob;
980 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
981 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
982 atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
983 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
984 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
987 atomic_inc(&dev->gnd_n_mdd);
988 atomic64_add(bytes, &dev->gnd_nbytes_map);
990 /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
993 /* we only get here in the valid cases */
994 list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
995 dev->gnd_map_version++;
996 spin_unlock(&dev->gnd_map_lock);
1000 kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
1004 GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
1005 "not mapped!", NULL);
1006 spin_lock(&dev->gnd_map_lock);
1008 switch (tx->tx_buftype) {
1010 GNIDBG_TX(D_EMERG, tx,
1011 "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
1012 spin_unlock(&dev->gnd_map_lock);
1016 case GNILND_BUF_PHYS_UNMAPPED:
1017 bytes = tx->tx_phys_npages * PAGE_SIZE;
1018 dev->gnd_map_nphys--;
1019 dev->gnd_map_physnop -= tx->tx_phys_npages;
1022 case GNILND_BUF_VIRT_UNMAPPED:
1024 dev->gnd_map_nvirt--;
1025 dev->gnd_map_virtnob -= tx->tx_nob;
1029 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1030 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1031 atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
1032 LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
1033 "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
1034 GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
1035 bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1038 atomic_dec(&dev->gnd_n_mdd);
1039 atomic64_sub(bytes, &dev->gnd_nbytes_map);
1041 /* we only get here in the valid cases */
1042 list_del_init(&tx->tx_map_list);
1043 dev->gnd_map_version++;
1044 spin_unlock(&dev->gnd_map_lock);
1048 kgnilnd_map_buffer(kgn_tx_t *tx)
1050 kgn_conn_t *conn = tx->tx_conn;
1051 kgn_device_t *dev = conn->gnc_device;
1052 __u32 flags = GNI_MEM_READWRITE;
1055 /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
1056 * be called concurrently as there are internal locks that protect
1057 * any data structures or HW resources. We just need to ensure
1058 * that our concurrency doesn't result in the kgn_device_t
1059 * getting nuked while we are in here */
1061 LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
1062 " to set tx_conn before calling %s\n", tx, __FUNCTION__);
1064 if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
1067 if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
1068 flags |= GNI_MEM_RELAXED_PI_ORDERING;
1071 switch (tx->tx_buftype) {
1075 case GNILND_BUF_NONE:
1076 case GNILND_BUF_IMMEDIATE:
1077 case GNILND_BUF_IMMEDIATE_KIOV:
1078 case GNILND_BUF_PHYS_MAPPED:
1079 case GNILND_BUF_VIRT_MAPPED:
1082 case GNILND_BUF_PHYS_UNMAPPED:
1083 GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
1084 rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
1085 tx->tx_phys, tx->tx_phys_npages, NULL,
1086 GNI_MEM_PHYS_SEGMENTS | flags,
1088 /* could race with other uses of the map counts, but this is ok
1089 * - this needs to turn into a non-fatal error soon to allow
1090 * GART resource, etc starvation handling */
1091 if (rrc != GNI_RC_SUCCESS) {
1092 GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
1093 "phys %u pp %u, virt %u nob %llu",
1094 tx->tx_phys_npages, dev->gnd_id,
1095 dev->gnd_map_nphys, dev->gnd_map_physnop,
1096 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1097 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1100 tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
1101 kgnilnd_mem_add_map_list(dev, tx);
1104 case GNILND_BUF_VIRT_UNMAPPED:
1105 rrc = kgnilnd_mem_register(dev->gnd_handle,
1106 (__u64)tx->tx_buffer, tx->tx_nob,
1107 NULL, flags, &tx->tx_map_key);
1108 if (rrc != GNI_RC_SUCCESS) {
1109 GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
1110 "phys %u pp %u, virt %u nob %llu",
1111 tx->tx_nob, dev->gnd_id,
1112 dev->gnd_map_nphys, dev->gnd_map_physnop,
1113 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
1114 RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
1117 tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
1118 kgnilnd_mem_add_map_list(dev, tx);
1119 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
1120 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1121 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
1122 GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
1123 tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
1131 kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
1133 kgn_conn_t *conn = tx->tx_conn;
1134 kgn_mdd_purgatory_t *gmp;
1136 LIBCFS_ALLOC(gmp, sizeof(*gmp));
1137 LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
1138 " asserting to avoid data corruption\n");
1139 if (tx->tx_buffer_copy)
1140 gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
1142 gmp->gmp_map_key = tx->tx_map_key;
1144 atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
1146 /* ensure that we don't have a blank purgatory - indicating the
1147 * conn is not already on purgatory lists - we'd never recover these
1148 * MDD if that were the case */
1149 GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
1150 "conn 0x%p->%s with NULL purgatory",
1151 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1153 /* link 'er up! - only place we really need to lock for
1154 * concurrent access */
1155 spin_lock(&conn->gnc_list_lock);
1156 list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
1157 spin_unlock(&conn->gnc_list_lock);
1161 kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
1165 int hold_timeout = 0;
1167 /* code below relies on +1 relationship ... */
1168 BUILD_BUG_ON(GNILND_BUF_PHYS_MAPPED !=
1169 (GNILND_BUF_PHYS_UNMAPPED + 1));
1170 BUILD_BUG_ON(GNILND_BUF_VIRT_MAPPED !=
1171 (GNILND_BUF_VIRT_UNMAPPED + 1));
1173 switch (tx->tx_buftype) {
1177 case GNILND_BUF_NONE:
1178 case GNILND_BUF_IMMEDIATE:
1179 case GNILND_BUF_PHYS_UNMAPPED:
1180 case GNILND_BUF_VIRT_UNMAPPED:
1182 case GNILND_BUF_IMMEDIATE_KIOV:
1183 if (tx->tx_phys != NULL) {
1184 vunmap(tx->tx_phys);
1185 } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
1186 kunmap(tx->tx_imm_pages[0]);
1188 /* clear to prevent kgnilnd_free_tx from thinking
1189 * this is a RDMA descriptor */
1193 case GNILND_BUF_PHYS_MAPPED:
1194 case GNILND_BUF_VIRT_MAPPED:
1195 LASSERT(tx->tx_conn != NULL);
1197 dev = tx->tx_conn->gnc_device;
1199 /* only want to hold if we are closing conn without
1200 * verified peer notification - the theory is that
1201 * a TX error can be communicated in all other cases */
1202 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
1203 error != -GNILND_NOPURG &&
1204 kgnilnd_check_purgatory_conn(tx->tx_conn)) {
1205 kgnilnd_add_purgatory_tx(tx);
1207 /* The timeout we give to kgni is a deadman stop only.
1208 * we are setting high to ensure we don't have the kgni timer
1209 * fire before ours fires _and_ is handled */
1210 hold_timeout = GNILND_TIMEOUT2DEADMAN;
1212 GNIDBG_TX(D_NET, tx,
1213 "dev %p delaying MDD release for %dms key %#llx.%#llx",
1214 tx->tx_conn->gnc_device, hold_timeout,
1215 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
1217 if (tx->tx_buffer_copy != NULL) {
1218 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
1219 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1220 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
1221 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1223 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
1224 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
1228 kgnilnd_mem_del_map_list(dev, tx);
1234 kgnilnd_tx_done(kgn_tx_t *tx, int completion)
1236 struct lnet_msg *lntmsg0, *lntmsg1;
1237 int status0, status1;
1238 struct lnet_ni *ni = NULL;
1239 kgn_conn_t *conn = tx->tx_conn;
1241 LASSERT(!in_interrupt());
1243 lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
1244 lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
1247 !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
1248 !kgnilnd_conn_clean_errno(completion)) {
1249 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
1250 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
1251 completion, tx, conn ?
1252 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
1253 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
1254 kgnilnd_tx_state2str(tx->tx_list_state),
1255 cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
1258 /* The error codes determine if we hold onto the MDD */
1259 kgnilnd_unmap_buffer(tx, completion);
1261 /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
1262 * we play nice with the error codes to avoid delivering a failed
1263 * REQUEST and then a REPLY event as well */
1265 /* return -EIO to lnet - it is the magic value for failed sends */
1266 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
1268 status1 = completion;
1270 status0 = status1 = completion;
1273 tx->tx_buftype = GNILND_BUF_NONE;
1274 tx->tx_msg.gnm_type = GNILND_MSG_NONE;
1276 /* lnet_finalize doesn't do anything with the *ni, so ok for us to
1277 * set NULL when we are a tx without a conn */
1279 ni = conn->gnc_peer->gnp_net->gnn_ni;
1281 spin_lock(&conn->gnc_tx_lock);
1283 LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
1284 (volatile unsigned long *)&conn->gnc_tx_bits),
1285 "conn %p tx %p bit %d already cleared\n",
1286 conn, tx, tx->tx_id.txe_idx);
1288 LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
1289 "msg_id %d already NULL\n", tx->tx_id.txe_idx);
1291 conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
1292 spin_unlock(&conn->gnc_tx_lock);
1295 kgnilnd_free_tx(tx);
1297 /* finalize AFTER freeing lnet msgs */
1299 /* warning - we should hold no locks here - calling lnet_finalize
1300 * could free up lnet credits, resulting in a call chain back into
1301 * the LND via kgnilnd_send and friends */
1303 lnet_finalize(lntmsg0, status0);
1305 if (lntmsg1 != NULL) {
1306 lnet_finalize(lntmsg1, status1);
1311 kgnilnd_txlist_done(struct list_head *txlist, int error)
1314 int err_printed = 0;
1316 if (list_empty(txlist))
1319 list_for_each_entry_safe(tx, txn, txlist, tx_list) {
1320 /* only print the first error */
1322 tx->tx_state |= GNILND_TX_QUIET_ERROR;
1323 list_del_init(&tx->tx_list);
1324 kgnilnd_tx_done(tx, error);
1329 kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
1333 spin_lock(&conn->gnc_tx_lock);
1335 /* ID zero is NOT ALLOWED!!! */
1338 id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
1339 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
1340 if (id == GNILND_MAX_MSG_ID) {
1341 if (conn->gnc_next_tx != 1) {
1342 /* we only searched from next_tx to end and didn't find
1343 * one, so search again from start */
1344 conn->gnc_next_tx = 1;
1347 /* couldn't find one! */
1348 spin_unlock(&conn->gnc_tx_lock);
1352 /* bump next_tx to prevent immediate reuse */
1353 conn->gnc_next_tx = id + 1;
1355 set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
1356 LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
1357 "tx 0x%p already at id %d\n",
1358 conn->gnc_tx_ref_table[id], id);
1360 /* delay these until we have a valid ID - prevents bad clear of the bit
1361 * in kgnilnd_tx_done */
1363 tx->tx_id.txe_cqid = conn->gnc_cqid;
1365 tx->tx_id.txe_idx = id;
1366 conn->gnc_tx_ref_table[id] = tx;
1368 /* Using jiffies to help differentiate against TX reuse - with
1369 * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
1370 * if we are sending to the same node faster than 256000/sec.
1371 * To help guard against this, we OR in the tx_seq - that is 32 bits */
1373 tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));
1375 GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);
1377 spin_unlock(&conn->gnc_tx_lock);
1382 kgnilnd_tx_log_retrans(kgn_conn_t *conn, kgn_tx_t *tx)
1386 log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0));
1388 /* we don't care about TX timeout - it could be that the network is slower
1389 * or throttled. We'll keep retranmitting - so if the network is so slow
1390 * that we fill up our mailbox, we'll keep trying to resend that msg
1391 * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
1392 * that he hasn't send us any traffic in return */
1394 /* some reasonable throttling of the debug message */
1396 unsigned long now = jiffies;
1397 /* XXX Nic: Mystical TX debug here... */
1398 /* We expect retransmissions so only log when D_NET is enabled */
1399 GNIDBG_SMSG_CREDS(D_NET, conn);
1400 GNIDBG_TOMSG(D_NET, &tx->tx_msg,
1401 "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
1402 " last_msg %uus/%uus last_cq %uus/%uus",
1403 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1404 tx->tx_id, tx->tx_retrans,
1405 jiffies_to_usecs(now - tx->tx_cred_wait),
1406 jiffies_to_usecs(now - conn->gnc_last_tx),
1407 jiffies_to_usecs(now - conn->gnc_last_rx),
1408 jiffies_to_usecs(now - conn->gnc_last_tx_cq),
1409 jiffies_to_usecs(now - conn->gnc_last_rx_cq));
1413 /* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
1414 * to avoid bad ordering with state_lock */
1417 kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1418 spinlock_t *state_lock, kgn_tx_list_state_t state)
1420 kgn_conn_t *conn = tx->tx_conn;
1421 kgn_msg_t *msg = &tx->tx_msg;
1423 unsigned long newest_last_rx, timeout;
1426 LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
1427 immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
1429 "msg 0x%p type %d wrong payload size %d\n",
1430 msg, msg->gnm_type, immediatenob);
1432 /* make sure we catch all the cases where we'd send on a dirty old mbox
1433 * but allow case for sending CLOSE. Since this check is within the CQ
1434 * mutex barrier and the close message is only sent through
1435 * kgnilnd_send_conn_close the last message out the door will be the
1438 if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
1439 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1440 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1441 /* Return -ETIME, we are closing the connection already so we dont want to
1442 * have this tx hit the wire. The tx will be killed by the calling function.
1443 * Once the EP is marked dirty the close message will be the last
1444 * thing to hit the wire */
1449 timeout = cfs_time_seconds(conn->gnc_timeout);
1451 newest_last_rx = GNILND_LASTRX(conn);
1453 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
1454 now = now + (GNILND_TIMEOUTRX(timeout) * 2);
1457 if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
1458 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
1459 "Cant send to %s after timeout lapse of %lu; TO %lu\n",
1460 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1461 cfs_duration_sec(now - newest_last_rx),
1462 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
1463 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1464 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1468 GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
1469 /* msg->gnm_srcnid is set when the message is initialized by whatever function is
1470 * creating the message this allows the message to contain the correct LNET NID/NET needed
1471 * instead of the one that the peer/conn uses for sending the data.
1473 msg->gnm_connstamp = conn->gnc_my_connstamp;
1474 msg->gnm_payload_len = immediatenob;
1475 msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);
1477 /* always init here - kgn_checksum is a /sys module tunable
1478 * and can be flipped at any point, even between msg init and sending */
1480 if (*kgnilnd_tunables.kgn_checksum) {
1481 /* We must set here and not in kgnilnd_init_msg,
1482 * we could resend this msg many times
1483 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
1484 * through init_msg again */
1485 msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
1486 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
1487 msg->gnm_cksum += 0xf00f;
1491 GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
1492 tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1493 sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
1494 tx->tx_id.txe_idx, immediate, immediatenob);
1496 if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
1497 rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
1499 rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
1500 msg, sizeof(*msg), immediate,
1502 tx->tx_id.txe_smsg_id);
1506 case GNI_RC_SUCCESS:
1507 atomic_inc(&conn->gnc_tx_seq);
1508 conn->gnc_last_tx = jiffies;
1509 /* no locking here as LIVE isn't a list */
1510 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);
1512 /* this needs to be checked under lock as it might be freed from a completion
1515 if (msg->gnm_type == GNILND_MSG_NOOP) {
1516 set_mb(conn->gnc_last_noop_sent, jiffies);
1519 /* serialize with seeing CQ events for completion on this, as well as
1521 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1522 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1524 atomic_inc(&conn->gnc_device->gnd_short_ntx);
1525 atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
1526 kgnilnd_peer_alive(conn->gnc_peer);
1527 GNIDBG_SMSG_CREDS(D_NET, conn);
1530 case GNI_RC_NOT_DONE:
1531 /* Jshimek: We can get GNI_RC_NOT_DONE for 3 reasons currently
1532 * 1: out of mbox credits
1533 * 2: out of mbox payload credits
1534 * 3: On Aries out of dla credits
1536 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1537 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1538 /* We'll handle this error inline - makes the calling logic much more
1541 /* If no lock, caller doesn't want us to retry */
1542 if (state_lock == NULL) {
1546 /* I need kgni credits to send this. Replace tx at the head of the
1547 * fmaq and I'll get rescheduled when credits appear. Reset the tx_state
1548 * and bump retrans counts since we are requeueing the tx.
1552 conn->gnc_tx_retrans++;
1554 kgnilnd_tx_log_retrans(conn, tx);
1555 /* add to head of list for the state and retries */
1556 spin_lock(state_lock);
1557 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
1558 spin_unlock(state_lock);
1560 /* We only reschedule for a certain number of retries, then
1561 * we will wait for the CQ events indicating a release of SMSG
1563 if (tx->tx_retrans < *kgnilnd_tunables.kgn_max_retransmits) {
1564 kgnilnd_schedule_conn(conn);
1567 /* CQ event coming in signifies either TX completed or
1568 * RX receive. Either of these *could* free up credits
1569 * in the SMSG mbox and we should try sending again */
1570 GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
1571 tx->tx_conn->gnc_cqid);
1572 kgnilnd_schedule_delay_conn(conn);
1573 /* use +ve return code to let upper layers know they
1574 * should stop looping on sends */
1578 /* handle bad retcode gracefully */
1579 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
1580 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
1585 /* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
1587 kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1588 spinlock_t *state_lock, kgn_tx_list_state_t state)
1590 kgn_device_t *dev = tx->tx_conn->gnc_device;
1591 unsigned long timestamp;
1594 timestamp = jiffies;
1595 kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
1596 kgnilnd_conn_mutex_lock(&tx->tx_conn->gnc_smsg_mutex);
1597 /* delay in jiffies - we are really concerned only with things that
1598 * result in a schedule() or really holding this off for long times .
1599 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1600 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1602 rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);
1608 /* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
1610 kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
1611 spinlock_t *state_lock, kgn_tx_list_state_t state)
1613 kgn_conn_t *conn = tx->tx_conn;
1614 kgn_device_t *dev = conn->gnc_device;
1615 unsigned long timestamp;
1618 timestamp = jiffies;
1620 /* technically we are doing bad things with the read_lock on the peer_conn
1621 * table, but we shouldn't be sleeping inside here - and we don't sleep/block
1622 * for the mutex. I bet lockdep is gonna flag this one though... */
1624 /* there are a few cases where we don't want the immediate send - like
1625 * when we are in the scheduler thread and it'd harm the latency of
1626 * getting messages up to LNet */
1628 /* rmb for gnd_ready */
1630 if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
1632 atomic_inc(&conn->gnc_device->gnd_fast_block);
1633 } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
1634 /* dont hit HW during quiesce */
1636 } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
1637 /* dont hit HW if stale EPs and conns left to close */
1640 atomic_inc(&conn->gnc_device->gnd_fast_try);
1641 rc = kgnilnd_trylock(&conn->gnc_device->gnd_cq_mutex,
1642 &conn->gnc_smsg_mutex);
1647 /* we got the mutex and weren't blocked */
1649 /* delay in jiffies - we are really concerned only with things that
1650 * result in a schedule() or really holding this off for long times .
1651 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
1652 dev->gnd_mutex_delay += (long) jiffies - timestamp;
1654 atomic_inc(&conn->gnc_device->gnd_fast_ok);
1655 tx->tx_qtime = jiffies;
1656 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
1657 rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1658 /* _nolock unlocks the mutex for us */
1664 /* lets us know if we can push this RDMA through now */
1666 kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
1670 bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1672 if (bytes_left < 0) {
1673 atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
1674 atomic_inc(&dev->gnd_rdmaq_nstalls);
1677 CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
1678 dev->gnd_rdmaq_deadline);
1679 mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
1680 /* we never del this timer - at worst it schedules us.. */
1687 /* this adds a TX to the queue pending throttling authorization before
1688 * we allow our remote peer to launch a PUT at us */
1690 kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
1694 /* we cannot go into send_mapped_tx from here as we are holding locks
1695 * and mem registration might end up allocating memory in kgni.
1696 * That said, we'll push this as far as we can into the queue process */
1697 rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);
1700 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
1701 kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
1702 /* lets us know how delayed RDMA is */
1703 tx->tx_qtime = jiffies;
1704 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
1706 /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
1707 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
1708 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
1709 /* lets us know how delayed mapping is */
1710 tx->tx_qtime = jiffies;
1711 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
1714 /* make sure we wake up sched to run this */
1715 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
1718 /* push TX through state machine */
1720 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
1725 /* set the tx_id here, we delay it until we have an actual conn
1727 * in some cases, the tx_id is already set to provide for things
1728 * like RDMA completion cookies, etc */
1729 if (tx->tx_id.txe_idx == 0) {
1730 rc = kgnilnd_set_tx_id(tx, conn);
1732 kgnilnd_tx_done(tx, rc);
1737 CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
1738 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
1740 /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
1742 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
1743 kgnilnd_tx_done(tx, rc);
1747 switch (tx->tx_msg.gnm_type) {
1748 case GNILND_MSG_PUT_ACK:
1749 case GNILND_MSG_GET_REQ:
1750 case GNILND_MSG_PUT_REQ_REV:
1751 case GNILND_MSG_GET_ACK_REV:
1752 /* hijacking time! If this messages will authorize our peer to
1753 * send his dirty little bytes in an RDMA, we need to get permission */
1754 kgnilnd_queue_rdma(conn, tx);
1756 case GNILND_MSG_IMMEDIATE:
1757 /* try to send right now, can help reduce latency */
1758 rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
1761 /* it was sent, break out of switch to avoid default case of queueing */
1764 /* needs to queue to try again, so fall through to default case */
1765 case GNILND_MSG_NOOP:
1766 /* Just make sure this goes out first for this conn */
1768 /* fall through... */
1770 spin_lock(&conn->gnc_list_lock);
1771 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
1772 tx->tx_qtime = jiffies;
1773 spin_unlock(&conn->gnc_list_lock);
1774 kgnilnd_schedule_conn(conn);
1779 kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, struct lnet_process_id *target)
1782 kgn_peer_t *new_peer = NULL;
1783 kgn_conn_t *conn = NULL;
1789 /* If I get here, I've committed to send, so I complete the tx with
1790 * failure on any problems */
1792 GNITX_ASSERTF(tx, tx->tx_conn == NULL,
1793 "tx already has connection %p", tx->tx_conn);
1795 /* do all of the peer & conn searching in one swoop - this avoids
1796 * nastiness when dropping locks and needing to maintain a sane state
1797 * in the face of stack reset or something else nuking peers & conns */
1799 /* I expect to find him, so only take a read lock */
1800 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1802 peer = kgnilnd_find_peer_locked(target->nid);
1804 conn = kgnilnd_find_conn_locked(peer);
1805 /* this could be NULL during quiesce */
1807 /* Connection exists; queue message on it */
1808 kgnilnd_queue_tx(conn, tx);
1809 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1813 /* don't create a connection if the peer is marked down */
1814 if (peer->gnp_state != GNILND_PEER_UP) {
1815 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1821 /* creating peer or conn; I'll need a write lock... */
1822 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1824 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1826 node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
1828 /* NB - this will not block during normal operations -
1829 * the only writer of this is in the startup/shutdown path. */
1830 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1836 /* ignore previous peer entirely - we cycled the lock, so we
1837 * will create new peer and at worst drop it if peer is still
1839 rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
1841 up_read(&kgnilnd_data.kgn_net_rw_sem);
1845 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1846 up_read(&kgnilnd_data.kgn_net_rw_sem);
1848 /* search for peer again now that we have the lock
1849 * if we don't find it, add our new one to the list */
1850 kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
1852 /* don't create a connection if the peer is not up */
1853 if (peer->gnp_state != GNILND_PEER_UP) {
1854 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1859 conn = kgnilnd_find_or_create_conn_locked(peer);
1861 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
1862 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1867 /* oh hey, found a conn now... magical */
1868 kgnilnd_queue_tx(conn, tx);
1870 /* no conn, must be trying to connect - so we queue for now */
1871 tx->tx_qtime = jiffies;
1872 kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
1874 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1877 kgnilnd_tx_done(tx, rc);
1882 kgnilnd_rdma(kgn_tx_t *tx, int type,
1883 kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
1885 kgn_conn_t *conn = tx->tx_conn;
1886 unsigned long timestamp;
1887 gni_post_type_t post_type;
1890 unsigned int desc_nob = nob;
1891 void *desc_buffer = tx->tx_buffer;
1892 gni_mem_handle_t desc_map_key = tx->tx_map_key;
1893 LASSERTF(kgnilnd_tx_mapped(tx),
1894 "unmapped tx %p\n", tx);
1895 LASSERTF(conn != NULL,
1896 "NULL conn on tx %p, naughty, naughty\n", tx);
1897 LASSERTF(nob <= sink->gnrd_nob,
1898 "nob %u > sink->gnrd_nob %d (%p)\n",
1899 nob, sink->gnrd_nob, sink);
1900 LASSERTF(nob <= tx->tx_nob,
1901 "nob %d > tx(%p)->tx_nob %d\n",
1902 nob, tx, tx->tx_nob);
1905 case GNILND_MSG_GET_DONE:
1906 case GNILND_MSG_PUT_DONE:
1907 post_type = GNI_POST_RDMA_PUT;
1909 case GNILND_MSG_GET_DONE_REV:
1910 case GNILND_MSG_PUT_DONE_REV:
1911 post_type = GNI_POST_RDMA_GET;
1914 CERROR("invalid msg type %s (%d)\n",
1915 kgnilnd_msgtype2str(type), type);
1918 if (post_type == GNI_POST_RDMA_GET) {
1919 /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
1920 * aligned. If the local buffer is not aligned correctly using the copy buffer
1921 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
1922 * up transferring extra bytes into the buffer but only copy the correct nob into the original
1923 * buffer. Remote offset correction is done through a combination of adjusting the offset,
1924 * making sure the length and addr are aligned and copying the data into the correct location
1925 * once the transfer has completed.
1927 if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
1928 (sink->gnrd_addr & 3) ||
1931 tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
1933 atomic_inc(&kgnilnd_data.kgn_rev_offset);
1935 if ((nob + tx->tx_offset) & 3) {
1936 desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
1937 atomic_inc(&kgnilnd_data.kgn_rev_length);
1939 desc_nob = (nob + tx->tx_offset);
1942 if (tx->tx_buffer_copy == NULL) {
1943 /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
1944 * and require at most we allocate a few extra bytes. */
1945 tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
1947 if (!tx->tx_buffer_copy) {
1948 /* allocation of buffer failed nak the rdma */
1949 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1950 kgnilnd_tx_done(tx, -EFAULT);
1953 atomic_inc(&kgnilnd_data.kgn_rev_copy_buff);
1954 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);
1955 if (rc != GNI_RC_SUCCESS) {
1956 /* Registration Failed nak rdma and kill the tx. */
1957 kgnilnd_vfree(tx->tx_buffer_copy,
1959 tx->tx_buffer_copy = NULL;
1960 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
1961 kgnilnd_tx_done(tx, -EFAULT);
1965 desc_map_key = tx->tx_buffer_copy_map_key;
1966 desc_buffer = tx->tx_buffer_copy;
1970 memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
1971 tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
1972 tx->tx_rdma_desc.type = post_type;
1973 tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
1974 tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
1975 tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
1976 tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
1977 tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
1978 tx->tx_rdma_desc.length = desc_nob;
1979 tx->tx_nob_rdma = nob;
1980 if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
1981 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
1982 if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
1983 tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
1984 /* prep final completion message */
1985 kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
1986 tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
1987 /* send actual size RDMA'd in retval */
1988 tx->tx_msg.gnm_u.completion.gncm_retval = nob;
1990 kgnilnd_compute_rdma_cksum(tx, nob);
1993 kgnilnd_queue_tx(conn, tx);
1997 /* Don't lie (CLOSE == RDMA idle) */
1998 LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
1999 tx, conn, conn->gnc_close_sent);
2001 GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
2002 "0x%x cookie:%#llx",
2003 type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
2005 /* set CQ dedicated for RDMA */
2006 tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;
2008 timestamp = jiffies;
2009 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
2010 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2011 /* delay in jiffies - we are really concerned only with things that
2012 * result in a schedule() or really holding this off for long times .
2013 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2014 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2016 rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);
2018 if (rrc == GNI_RC_ERROR_RESOURCE) {
2019 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2020 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2021 kgnilnd_unmap_buffer(tx, 0);
2023 if (tx->tx_buffer_copy != NULL) {
2024 kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
2025 tx->tx_buffer_copy = NULL;
2028 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2029 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
2031 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2032 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2036 spin_lock(&conn->gnc_list_lock);
2037 kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
2038 tx->tx_qtime = jiffies;
2039 spin_unlock(&conn->gnc_list_lock);
2040 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2041 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
2043 /* XXX Nic: is this a place we should handle more errors for
2044 * robustness sake */
2045 LASSERT(rrc == GNI_RC_SUCCESS);
2050 kgnilnd_alloc_rx(void)
2054 rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
2056 CERROR("failed to allocate rx\n");
2059 CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
2062 /* no memset to zero, we'll always fill all members */
2066 /* release is to just free connection resources
2067 * we use this for the eager path after copying */
2069 kgnilnd_release_msg(kgn_conn_t *conn)
2072 unsigned long timestamp;
2074 CDEBUG(D_NET, "consuming %p\n", conn);
2076 timestamp = jiffies;
2077 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
2078 /* delay in jiffies - we are really concerned only with things that
2079 * result in a schedule() or really holding this off for long times .
2080 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
2081 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
2083 rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
2084 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
2086 LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
2087 GNIDBG_SMSG_CREDS(D_NET, conn);
2089 kgnilnd_schedule_conn(conn);
2093 kgnilnd_consume_rx(kgn_rx_t *rx)
2095 kgn_conn_t *conn = rx->grx_conn;
2096 kgn_msg_t *rxmsg = rx->grx_msg;
2098 /* if we are eager, free the cache alloc'd msg */
2099 if (unlikely(rx->grx_eager)) {
2100 LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
2101 atomic_dec(&kgnilnd_data.kgn_neager_allocs);
2103 /* release ref from eager_recv */
2104 kgnilnd_conn_decref(conn);
2106 GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
2107 kgnilnd_release_msg(conn);
2110 kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
2111 CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
2116 kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
2118 struct lnet_hdr *hdr = &lntmsg->msg_hdr;
2119 int type = lntmsg->msg_type;
2120 struct lnet_process_id target = lntmsg->msg_target;
2121 int target_is_router = lntmsg->msg_target_is_router;
2122 int routing = lntmsg->msg_routing;
2123 unsigned int niov = lntmsg->msg_niov;
2124 struct bio_vec *kiov = lntmsg->msg_kiov;
2125 unsigned int offset = lntmsg->msg_offset;
2126 unsigned int nob = lntmsg->msg_len;
2127 unsigned int msg_vmflush = lntmsg->msg_vmflush;
2128 kgn_net_t *net = ni->ni_data;
2131 /* '1' for consistency with code that checks !mpflag to restore */
2132 unsigned int mpflag = 1;
2133 int reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;
2135 /* NB 'private' is different depending on what we're sending.... */
2136 LASSERT(!in_interrupt());
2138 CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
2139 type, nob, niov, libcfs_id2str(target));
2141 LASSERTF(nob == 0 || niov > 0,
2142 "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
2143 LASSERTF(niov <= LNET_MAX_IOV,
2144 "lntmsg %p niov %d\n", lntmsg, niov);
2147 mpflag = memalloc_noreclaim_save();
2151 CERROR("lntmsg %p with unexpected type %d\n",
2156 LASSERTF(nob == 0, "lntmsg %p nob %d\n",
2164 if (routing || target_is_router)
2165 break; /* send IMMEDIATE */
2167 /* it is safe to do direct GET with out mapping buffer for RDMA as we
2168 * check the eventual sink buffer here - if small enough, remote
2169 * end is perfectly capable of returning data in short message -
2170 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
2171 * for IMMEDIATE messages which will have it send a real reply instead
2172 * of doing kgnilnd_recv to have the RDMA continued */
2173 if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
2176 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2177 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
2179 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);
2185 rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
2186 NULL, lntmsg->msg_md->md_kiov,
2187 0, lntmsg->msg_md->md_length);
2189 CERROR("unable to setup buffer: %d\n", rc);
2190 kgnilnd_tx_done(tx, rc);
2195 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
2196 if (tx->tx_lntmsg[1] == NULL) {
2197 CERROR("Can't create reply for GET to %s\n",
2198 libcfs_nid2str(target.nid));
2199 kgnilnd_tx_done(tx, rc);
2204 tx->tx_lntmsg[0] = lntmsg;
2205 if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
2206 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2208 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2210 /* rest of tx_msg is setup just before it is sent */
2211 kgnilnd_launch_tx(tx, net, &target);
2213 case LNET_MSG_REPLY:
2215 /* to save on MDDs, we'll handle short kiov by vmap'ing
2216 * and sending via SMSG */
2217 if (nob <= *kgnilnd_tunables.kgn_max_immediate)
2220 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2221 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
2223 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);
2230 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL,
2233 kgnilnd_tx_done(tx, rc);
2238 tx->tx_lntmsg[0] = lntmsg;
2239 if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
2240 tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
2242 tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
2244 /* rest of tx_msg is setup just before it is sent */
2245 kgnilnd_launch_tx(tx, net, &target);
2249 /* send IMMEDIATE */
2251 LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
2252 "lntmsg 0x%p too large %d\n", lntmsg, nob);
2254 tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
2260 rc = kgnilnd_setup_immediate_buffer(tx, niov, NULL, kiov, offset, nob);
2262 kgnilnd_tx_done(tx, rc);
2266 tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
2267 tx->tx_lntmsg[0] = lntmsg;
2268 kgnilnd_launch_tx(tx, net, &target);
2271 /* use stored value as we could have already finalized lntmsg here from a failed launch */
2273 memalloc_noreclaim_restore(mpflag);
2278 kgnilnd_setup_rdma(struct lnet_ni *ni, kgn_rx_t *rx, struct lnet_msg *lntmsg, int mlen)
2280 kgn_conn_t *conn = rx->grx_conn;
2281 kgn_msg_t *rxmsg = rx->grx_msg;
2282 unsigned int niov = lntmsg->msg_niov;
2283 struct bio_vec *kiov = lntmsg->msg_kiov;
2284 unsigned int offset = lntmsg->msg_offset;
2285 unsigned int nob = lntmsg->msg_len;
2290 switch (rxmsg->gnm_type) {
2291 case GNILND_MSG_PUT_REQ_REV:
2292 done_type = GNILND_MSG_PUT_DONE_REV;
2295 case GNILND_MSG_GET_REQ:
2296 done_type = GNILND_MSG_GET_DONE;
2299 CERROR("invalid msg type %s (%d)\n",
2300 kgnilnd_msgtype2str(rxmsg->gnm_type),
2305 tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
2309 rc = kgnilnd_set_tx_id(tx, conn);
2313 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL, kiov, offset, nob);
2317 tx->tx_lntmsg[0] = lntmsg;
2318 tx->tx_getinfo = rxmsg->gnm_u.get;
2320 /* we only queue from kgnilnd_recv - we might get called from other contexts
2321 * and we don't want to block the mutex in those cases */
2323 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2324 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2325 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2326 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2331 kgnilnd_tx_done(tx, rc);
2332 kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2334 lnet_finalize(lntmsg, rc);
2338 kgnilnd_eager_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2341 kgn_rx_t *rx = private;
2342 kgn_conn_t *conn = rx->grx_conn;
2343 kgn_msg_t *rxmsg = rx->grx_msg;
2344 kgn_msg_t *eagermsg = NULL;
2345 kgn_peer_t *peer = NULL;
2346 kgn_conn_t *found_conn = NULL;
2348 GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
2349 conn, rxmsg, lntmsg);
2351 if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
2352 GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
2353 rxmsg->gnm_payload_len);
2356 /* Grab a read lock so the connection doesnt disappear on us
2357 * while we look it up
2359 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2361 peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
2363 found_conn = kgnilnd_find_conn_locked(peer);
2366 /* Verify the connection found is the same one that the message
2367 * is supposed to be using, if it is not output an error message
2370 if (!peer || !found_conn
2371 || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
2372 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2373 CERROR("Couldnt find matching peer %p or conn %p / %p\n",
2374 peer, conn, found_conn);
2376 CERROR("Unexpected connstamp %#llx(%#llx expected)"
2377 " from %s", rxmsg->gnm_connstamp,
2378 found_conn->gnc_peer_connstamp,
2379 libcfs_nid2str(peer->gnp_nid));
2384 /* add conn ref to ensure it doesn't go away until all eager
2385 * messages processed */
2386 kgnilnd_conn_addref(conn);
2388 /* Now that we have verified the connection is valid and added a
2389 * reference we can remove the read_lock on the peer_conn_lock */
2390 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2392 /* we have no credits or buffers for this message, so copy it
2393 * somewhere for a later kgnilnd_recv */
2394 if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
2395 *kgnilnd_tunables.kgn_eager_credits) {
2396 CERROR("Out of eager credits to %s\n",
2397 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2401 atomic_inc(&kgnilnd_data.kgn_neager_allocs);
2403 LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
2404 if (eagermsg == NULL) {
2405 kgnilnd_conn_decref(conn);
2406 CERROR("couldn't allocate eager rx message for conn %p to %s\n",
2407 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
2411 /* copy msg and payload */
2412 memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
2413 rx->grx_msg = eagermsg;
2416 /* stash this for lnet_finalize on cancel-on-conn-close */
2417 rx->grx_lntmsg = lntmsg;
2419 /* keep the same rx_t, it just has a new grx_msg now */
2420 *new_private = private;
2422 /* release SMSG buffer */
2423 kgnilnd_release_msg(conn);
2429 kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
2430 int delayed, unsigned int niov,
2431 struct bio_vec *kiov,
2432 unsigned int offset, unsigned int mlen, unsigned int rlen)
2434 kgn_rx_t *rx = private;
2435 kgn_conn_t *conn = rx->grx_conn;
2436 kgn_msg_t *rxmsg = rx->grx_msg;
2442 LASSERT(!in_interrupt());
2443 LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
2445 GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
2446 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2447 conn, rxmsg, lntmsg,
2448 niov, kiov, offset, mlen, rlen);
2450 /* we need to lock here as recv can be called from any context */
2451 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
2452 if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
2453 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2455 /* someone closed the conn after we copied this out, nuke it */
2456 kgnilnd_consume_rx(rx);
2457 lnet_finalize(lntmsg, conn->gnc_error);
2460 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2462 switch (rxmsg->gnm_type) {
2464 GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
2465 " niov=%d kiov=%p offset=%d mlen=%d rlen=%d",
2466 conn, rx, rxmsg, lntmsg, niov, kiov, offset, mlen, rlen);
2469 case GNILND_MSG_IMMEDIATE:
2470 if (mlen > rxmsg->gnm_payload_len) {
2471 GNIDBG_MSG(D_ERROR, rxmsg,
2472 "Immediate message from %s too big: %d > %d",
2473 libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
2474 rxmsg->gnm_payload_len);
2476 kgnilnd_consume_rx(rx);
2480 /* rxmsg[1] is a pointer to the payload, sitting in the buffer
2481 * right after the kgn_msg_t header - so just 'cute' way of saying
2482 * rxmsg + sizeof(kgn_msg_t) */
2484 /* check payload checksum if sent */
2486 if (*kgnilnd_tunables.kgn_checksum >= 2 &&
2487 !rxmsg->gnm_payload_cksum &&
2488 rxmsg->gnm_payload_len != 0)
2489 GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");
2491 if (rxmsg->gnm_payload_cksum != 0) {
2492 /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
2493 * which is what is used to calculate the cksum on the TX side */
2494 pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);
2496 if (rxmsg->gnm_payload_cksum != pload_cksum) {
2497 GNIDBG_MSG(D_NETERROR, rxmsg,
2498 "Bad payload checksum (%x expected %x)",
2499 pload_cksum, rxmsg->gnm_payload_cksum);
2500 switch (*kgnilnd_tunables.kgn_checksum_dump) {
2502 kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
2503 &rxmsg[1], rxmsg->gnm_payload_len);
2504 /* fall through to dump */
2506 libcfs_debug_dumplog();
2512 /* checksum problems are fatal, kill the conn */
2513 kgnilnd_consume_rx(rx);
2514 kgnilnd_close_conn(conn, rc);
2520 lnet_copy_flat2kiov(
2522 *kgnilnd_tunables.kgn_max_immediate,
2523 &rxmsg[1], 0, mlen);
2527 *kgnilnd_tunables.kgn_max_immediate,
2528 &rxmsg[1], 0, mlen);
2530 kgnilnd_consume_rx(rx);
2531 lnet_finalize(lntmsg, 0);
2534 case GNILND_MSG_PUT_REQ:
2535 /* LNET wants to truncate or drop transaction, sending NAK */
2537 kgnilnd_consume_rx(rx);
2538 lnet_finalize(lntmsg, 0);
2540 /* only error if lntmsg == NULL, otherwise we are just
2541 * short circuiting the rdma process of 0 bytes */
2542 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2543 lntmsg == NULL ? -ENOENT : 0,
2544 rxmsg->gnm_u.get.gngm_cookie,
2548 /* sending ACK with sink buff. info */
2549 tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
2551 kgnilnd_consume_rx(rx);
2555 rc = kgnilnd_set_tx_id(tx, conn);
2557 GOTO(nak_put_req, rc);
2560 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL,
2561 kiov, offset, mlen);
2563 GOTO(nak_put_req, rc);
2566 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2567 rxmsg->gnm_u.putreq.gnprm_cookie;
2568 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2569 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2570 (__u64)((unsigned long)tx->tx_buffer);
2571 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2573 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2574 tx->tx_qtime = jiffies;
2575 /* we only queue from kgnilnd_recv - we might get called from other contexts
2576 * and we don't want to block the mutex in those cases */
2578 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2579 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2580 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2581 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2583 kgnilnd_consume_rx(rx);
2587 /* make sure we send an error back when the PUT fails */
2588 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2589 kgnilnd_tx_done(tx, rc);
2590 kgnilnd_consume_rx(rx);
2592 /* return magic LNet network error */
2594 case GNILND_MSG_GET_REQ_REV:
2595 /* LNET wants to truncate or drop transaction, sending NAK */
2597 kgnilnd_consume_rx(rx);
2598 lnet_finalize(lntmsg, 0);
2600 /* only error if lntmsg == NULL, otherwise we are just
2601 * short circuiting the rdma process of 0 bytes */
2602 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2603 lntmsg == NULL ? -ENOENT : 0,
2604 rxmsg->gnm_u.get.gngm_cookie,
2608 /* lntmsg can be null when parsing a LNET_GET */
2609 if (lntmsg != NULL) {
2610 /* sending ACK with sink buff. info */
2611 tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
2613 kgnilnd_consume_rx(rx);
2617 rc = kgnilnd_set_tx_id(tx, conn);
2619 GOTO(nak_get_req_rev, rc);
2622 rc = kgnilnd_setup_rdma_buffer(tx, niov, NULL,
2623 kiov, offset, mlen);
2625 GOTO(nak_get_req_rev, rc);
2628 tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
2629 rxmsg->gnm_u.putreq.gnprm_cookie;
2630 tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
2631 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
2632 (__u64)((unsigned long)tx->tx_buffer);
2633 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
2635 tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
2637 /* we only queue from kgnilnd_recv - we might get called from other contexts
2638 * and we don't want to block the mutex in those cases */
2640 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
2641 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
2642 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
2643 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
2646 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2648 rxmsg->gnm_u.get.gngm_cookie,
2652 kgnilnd_consume_rx(rx);
2656 /* make sure we send an error back when the GET fails */
2657 kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
2658 kgnilnd_tx_done(tx, rc);
2659 kgnilnd_consume_rx(rx);
2661 /* return magic LNet network error */
2665 case GNILND_MSG_PUT_REQ_REV:
2666 /* LNET wants to truncate or drop transaction, sending NAK */
2668 kgnilnd_consume_rx(rx);
2669 lnet_finalize(lntmsg, 0);
2671 /* only error if lntmsg == NULL, otherwise we are just
2672 * short circuiting the rdma process of 0 bytes */
2673 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2674 lntmsg == NULL ? -ENOENT : 0,
2675 rxmsg->gnm_u.get.gngm_cookie,
2680 if (lntmsg != NULL) {
2682 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2685 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2687 rxmsg->gnm_u.get.gngm_cookie,
2690 kgnilnd_consume_rx(rx);
2692 case GNILND_MSG_GET_REQ:
2693 if (lntmsg != NULL) {
2695 kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
2698 kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
2700 rxmsg->gnm_u.get.gngm_cookie,
2703 kgnilnd_consume_rx(rx);
2709 /* needs write_lock on kgn_peer_conn_lock held */
2711 kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
2713 unsigned long timeout, keepalive;
2714 unsigned long now = jiffies;
2715 unsigned long newest_last_rx;
2718 /* given that we found this conn hanging off a peer, it better damned
2719 * well be connected */
2720 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
2721 "conn 0x%p->%s with bad state%s\n", conn,
2722 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
2724 kgnilnd_conn_state2str(conn));
2726 CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
2727 "rx_diff %lu tx_diff %lu\n",
2728 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
2729 conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
2730 cfs_duration_sec(now - conn->gnc_last_rx_cq),
2731 cfs_duration_sec(now - conn->gnc_last_tx));
2733 timeout = cfs_time_seconds(conn->gnc_timeout);
2734 keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));
2736 /* just in case our lack of RX msg processing is gumming up the works - give the
2737 * remove an extra chance */
2739 newest_last_rx = GNILND_LASTRX(conn);
2741 if (time_after_eq(now, newest_last_rx + timeout)) {
2742 uint32_t level = D_CONSOLE|D_NETERROR;
2744 if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
2747 GNIDBG_CONN(level, conn,
2748 "No gnilnd traffic received from %s for %lu "
2749 "seconds, terminating connection. Is node down? ",
2750 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2751 cfs_duration_sec(now - newest_last_rx));
2755 /* we don't timeout on last_tx stalls - we are going to trust the
2756 * underlying network to let us know when sends are failing.
2757 * At worst, the peer will timeout our RX stamp and drop the connection
2758 * at that point. We'll then see his CLOSE or at worst his RX
2759 * stamp stop and drop the connection on our end */
2761 if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
2762 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
2763 "last %lu/%lu/%lu %lus/%lus/%lus\n",
2764 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
2765 cfs_duration_sec(jiffies - conn->gnc_last_tx),
2767 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
2768 conn->gnc_last_noop_cq,
2769 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
2770 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
2771 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
2772 set_mb(conn->gnc_last_noop_want, jiffies);
2773 atomic_inc(&conn->gnc_reaper_noop);
2774 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
2777 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
2780 kgnilnd_queue_tx(conn, tx);
2786 /* needs write_lock on kgn_peer_conn_lock held */
2788 kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
2789 struct list_head *souls)
2791 unsigned long timeout;
2792 kgn_conn_t *conn, *connN = NULL;
2798 short releaseconn = 0;
2799 unsigned long first_rx = 0;
2800 int purgatory_conn_cnt = 0;
2802 CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
2803 peer, libcfs_nid2str(peer->gnp_nid),
2804 peer->gnp_reconnect_interval);
2806 timeout = cfs_time_seconds(max(*kgnilnd_tunables.kgn_timeout,
2807 GNILND_MIN_TIMEOUT));
2809 conn = kgnilnd_find_conn_locked(peer);
2811 /* if there is a valid conn, check the queues for timeouts */
2812 rc = kgnilnd_check_conn_timeouts_locked(conn);
2814 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
2815 /* simulate a RX CLOSE after the timeout but before
2816 * the scheduler thread gets it */
2817 conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
2818 conn->gnc_peer_error = -ETIMEDOUT;
2821 if (*kgnilnd_tunables.kgn_to_reconn_disable &&
2823 peer->gnp_state = GNILND_PEER_TIMED_OUT;
2824 CDEBUG(D_WARNING, "%s conn timed out, will "
2825 "reconnect upon request from peer\n",
2826 libcfs_nid2str(conn->gnc_peer->gnp_nid));
2828 /* Once we mark closed, any of the scheduler threads could
2829 * get it and move through before we hit the fail loc code */
2830 kgnilnd_close_conn_locked(conn, rc);
2832 /* first_rx is used to decide when to release a conn from purgatory.
2834 first_rx = conn->gnc_first_rx;
2838 /* now regardless of starting new conn, find tx on peer queue that
2839 * are old and smell bad - do this first so we don't trigger
2840 * reconnect on empty queue if we timeout all */
2841 list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
2842 if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
2844 LCONSOLE_INFO("could not send to %s due to connection"
2845 " setup failure after %lu seconds\n",
2846 libcfs_nid2str(peer->gnp_nid),
2847 cfs_duration_sec(jiffies - tx->tx_qtime));
2849 kgnilnd_tx_del_state_locked(tx, peer, NULL,
2851 list_add_tail(&tx->tx_list, todie);
2856 if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
2857 CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
2858 count, peer, libcfs_nid2str(peer->gnp_nid));
2859 /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
2860 if (list_empty(&peer->gnp_tx_queue) ||
2861 peer->gnp_connecting == GNILND_PEER_KILL) {
2862 /* we pass down todie to use a common function - but we know there are
2864 kgnilnd_cancel_peer_connect_locked(peer, todie);
2868 /* Don't reconnect if we are still trying to clear out old conns.
2869 * This prevents us sending traffic on the new mbox before ensuring we are done
2870 * with the old one */
2871 reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
2872 (atomic_read(&peer->gnp_dirty_eps) == 0);
2874 /* fast reconnect after a timeout */
2875 to_reconn = !conn &&
2876 (peer->gnp_last_errno == -ETIMEDOUT) &&
2877 *kgnilnd_tunables.kgn_fast_reconn;
2879 /* if we are not connected and there are tx on the gnp_tx_queue waiting
2880 * to be sent, we'll check the reconnect interval and fire up a new
2881 * connection request */
2884 (peer->gnp_connecting == GNILND_PEER_IDLE) &&
2885 (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
2886 (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {
2888 CDEBUG(D_NET, "starting connect to %s\n",
2889 libcfs_nid2str(peer->gnp_nid));
2890 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE,
2891 "Peer was idle and we have a write_lock, state issue %d\n",
2892 peer->gnp_connecting);
2894 peer->gnp_connecting = GNILND_PEER_CONNECT;
2895 kgnilnd_peer_addref(peer); /* extra ref for connd */
2897 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2898 list_add_tail(&peer->gnp_connd_list,
2899 &peer->gnp_net->gnn_dev->gnd_connd_peers);
2900 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2902 kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
2905 /* fail_loc to allow us to delay release of purgatory */
2906 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
2909 /* This check allows us to verify that the new conn is actually being used. This allows us to
2910 * pull the old conns out of purgatory if they have actually seen traffic.
2911 * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
2914 time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
2915 CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
2916 libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
2920 list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
2921 /* check for purgatory timeouts */
2922 if (conn->gnc_in_purgatory) {
2923 /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
2924 * that way the next time we check it we can detach it from purgatory
2927 if (conn->gnc_state != GNILND_CONN_DONE) {
2928 /* Skip over conns that are currently not DONE. If they arent already scheduled
2929 * for completion something in the state machine is broken.
2934 /* We only detach a conn that is in purgatory if we have received a close message,
2935 * we have a new valid connection that has successfully received data, or an admin
2936 * command tells us we need to detach.
2939 if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
2940 unsigned long waiting;
2942 waiting = (long) jiffies - conn->gnc_last_rx_cq;
2944 /* C.E: The remote peer is expected to close the
2945 * connection (see kgnilnd_check_conn_timeouts)
2946 * via the reaper thread and nuke out the MDD and
2947 * FMA resources after conn->gnc_timeout has expired
2948 * without an FMA RX */
2949 CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
2950 " held resources\n",
2951 libcfs_nid2str(conn->gnc_peer->gnp_nid),
2952 cfs_duration_sec(waiting));
2954 kgnilnd_detach_purgatory_locked(conn, souls);
2956 purgatory_conn_cnt++;
2961 /* If we have too many connections in purgatory we could run out of
2962 * resources. Limit the number of connections to a tunable number,
2963 * clean up to the minimum all in one fell swoop... there are
2964 * situations where dvs will retry tx's and we can eat up several
2965 * hundread connection requests at once.
2967 if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
2968 list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
2970 if (conn->gnc_in_purgatory &&
2971 conn->gnc_state == GNILND_CONN_DONE) {
2972 CDEBUG(D_NET, "Dropping Held resource due to"
2973 " resource limits being hit\n");
2974 kgnilnd_detach_purgatory_locked(conn, souls);
2976 if (purgatory_conn_cnt-- <
2977 *kgnilnd_tunables.kgn_max_purgatory)
2985 kgnilnd_reaper_check(int idx)
2987 struct list_head *peers = &kgnilnd_data.kgn_peers[idx];
2988 struct list_head *ctmp, *ctmpN;
2989 LIST_HEAD(geriatrics);
2992 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2994 list_for_each_safe(ctmp, ctmpN, peers) {
2995 kgn_peer_t *peer = NULL;
2997 /* don't timeout stuff if the network is mucked or shutting down */
2998 if (kgnilnd_check_hw_quiesce()) {
3001 peer = list_entry(ctmp, kgn_peer_t, gnp_list);
3003 kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
3006 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3008 kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
3009 kgnilnd_release_purgatory_list(&souls);
3013 kgnilnd_update_reaper_timeout(long timeout)
3015 LASSERT(timeout > 0);
3017 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3019 if (timeout < kgnilnd_data.kgn_new_min_timeout)
3020 kgnilnd_data.kgn_new_min_timeout = timeout;
3022 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3026 kgnilnd_reaper_poke_with_stick(unsigned long arg)
3028 wake_up(&kgnilnd_data.kgn_reaper_waitq);
3032 kgnilnd_reaper(void *arg)
3037 unsigned long next_check_time = jiffies;
3038 long current_min_timeout = MAX_SCHEDULE_TIMEOUT;
3039 struct timer_list timer;
3042 /* all gnilnd threads need to run fairly urgently */
3043 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
3044 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3046 while (!kgnilnd_data.kgn_shutdown) {
3047 /* I wake up every 'p' seconds to check for timeouts on some
3048 * more peers. I try to check every connection 'n' times
3049 * within the global minimum of all keepalive and timeout
3050 * intervals, to ensure I attend to every connection within
3051 * (n+1)/n times its timeout intervals. */
3052 const int p = GNILND_REAPER_THREAD_WAKE;
3053 const int n = GNILND_REAPER_NCHECKS;
3055 /* to quiesce or to not quiesce, that is the question */
3056 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3057 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3058 KGNILND_SPIN_QUIESCE;
3059 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3062 /* careful with the jiffy wrap... */
3063 timeout = (long)(next_check_time - jiffies);
3066 prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
3067 TASK_INTERRUPTIBLE);
3068 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3069 setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
3071 mod_timer(&timer, (long) jiffies + timeout);
3073 /* check flag variables before committing */
3074 if (!kgnilnd_data.kgn_shutdown &&
3075 !kgnilnd_data.kgn_quiesce_trigger) {
3076 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
3077 timeout, cfs_duration_sec(timeout));
3079 CDEBUG(D_INFO, "awake after schedule\n");
3082 del_singleshot_timer_sync(&timer);
3083 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3084 finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
3088 /* new_min_timeout is set from the conn timeouts and keepalive
3089 * this should end up with a min timeout of
3090 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
3091 if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
3092 current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
3093 CDEBUG(D_NET, "Set new min timeout %ld\n",
3094 current_min_timeout);
3097 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3099 /* Compute how many table entries to check now so I get round
3100 * the whole table fast enough given that I do this at fixed
3101 * intervals of 'p' seconds) */
3102 chunk = *kgnilnd_tunables.kgn_peer_hash_size;
3103 if (kgnilnd_data.kgn_new_min_timeout > n * p)
3104 chunk = (chunk * n * p) /
3105 kgnilnd_data.kgn_new_min_timeout;
3108 for (i = 0; i < chunk; i++) {
3109 kgnilnd_reaper_check(hash_index);
3110 hash_index = (hash_index + 1) %
3111 *kgnilnd_tunables.kgn_peer_hash_size;
3113 next_check_time = (long) jiffies + cfs_time_seconds(p);
3114 CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);
3116 spin_lock(&kgnilnd_data.kgn_reaper_lock);
3119 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
3121 kgnilnd_thread_fini();
3126 kgnilnd_recv_bte_get(kgn_tx_t *tx) {
3127 unsigned niov, offset, nob;
3128 struct bio_vec *kiov;
3129 struct lnet_msg *lntmsg = tx->tx_lntmsg[0];
3130 kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);
3133 lnet_copy_flat2kiov(
3136 tx->tx_buffer_copy + tx->tx_offset, 0, nob);
3138 memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
3145 kgnilnd_check_rdma_cq(kgn_device_t *dev)
3148 gni_post_descriptor_t *desc;
3150 kgn_tx_ev_id_t ev_id;
3152 int should_retry, rc;
3153 long num_processed = 0;
3154 kgn_conn_t *conn = NULL;
3155 kgn_tx_t *tx = NULL;
3156 kgn_rdma_desc_t *rdesc;
3161 /* make sure we don't keep looping if we need to reset */
3162 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3163 return num_processed;
3165 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3167 /* we didn't get the mutex, so return that there is still work
3171 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
3172 /* a bit gross - but we need a good way to test for
3173 * delayed RDMA completions and the easiest way to do
3174 * that is to delay the RDMA CQ events */
3175 rrc = GNI_RC_NOT_DONE;
3177 rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
3180 if (rrc == GNI_RC_NOT_DONE) {
3181 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3182 CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
3183 dev->gnd_id, num_processed);
3184 return num_processed;
3186 dev->gnd_sched_alive = jiffies;
3189 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3190 "this is bad, somehow our credits didn't protect us"
3191 " from CQ overrun\n");
3192 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
3193 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3194 event_data, GNI_CQ_GET_TYPE(event_data));
3196 rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
3198 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3200 /* XXX Nic: Need better error handling here... */
3201 LASSERTF((rrc == GNI_RC_SUCCESS) ||
3202 (rrc == GNI_RC_TRANSACTION_ERROR),
3205 ev_id.txe_cookie = desc->post_id;
3207 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3209 if (conn == NULL || tx == NULL) {
3210 /* either conn or tx was already nuked and this is a "late"
3211 * completion, so drop it */
3215 GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3216 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
3217 tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
3218 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
3219 "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);
3221 GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);
3223 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3224 lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
3225 tx->tx_msg.gnm_u.completion.gncm_retval);
3229 if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3230 if (tx->tx_buffer_copy != NULL)
3231 kgnilnd_recv_bte_get(tx);
3232 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
3235 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
3236 if (tx->tx_buffer_copy != NULL)
3237 kgnilnd_recv_bte_get(tx);
3238 rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
3241 /* remove from rdmaq */
3242 kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
3243 spin_lock(&conn->gnc_list_lock);
3244 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3245 spin_unlock(&conn->gnc_list_lock);
3246 kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
3248 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
3249 event_data = 1LL << 48;
3253 if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
3254 atomic_inc(&dev->gnd_rdma_ntx);
3255 atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
3256 /* transaction succeeded, add into fmaq */
3257 kgnilnd_queue_tx(conn, tx);
3258 kgnilnd_peer_alive(conn->gnc_peer);
3260 /* drop ref from kgnilnd_validate_tx_ev_id */
3261 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3262 kgnilnd_conn_decref(conn);
3267 /* fall through to the TRANSACTION_ERROR case */
3270 /* get stringified version for log messages */
3271 kgnilnd_cq_error_str(event_data, &err_str, 256);
3272 kgnilnd_cq_error_recoverable(event_data, &should_retry);
3274 /* make sure we are not off in the weeds with this tx */
3275 if (tx->tx_retrans >
3276 *kgnilnd_tunables.kgn_max_retransmits) {
3277 GNIDBG_TX(D_NETERROR, tx,
3278 "giving up on TX, too many retries", NULL);
3282 GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
3283 should_retry ? "transient" : "unrecoverable", err_str);
3285 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
3286 tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
3287 rdesc = &tx->tx_putinfo.gnpam_desc;
3288 rnob = tx->tx_putinfo.gnpam_desc.gnrd_nob;
3289 rcookie = tx->tx_putinfo.gnpam_dst_cookie;
3291 rdesc = &tx->tx_getinfo.gngm_desc;
3292 rnob = tx->tx_lntmsg[0]->msg_len;
3293 rcookie = tx->tx_getinfo.gngm_cookie;
3298 tx->tx_msg.gnm_type,
3302 kgnilnd_nak_rdma(conn,
3303 tx->tx_msg.gnm_type,
3306 tx->tx_msg.gnm_srcnid);
3307 kgnilnd_tx_done(tx, -GNILND_NOPURG);
3308 kgnilnd_close_conn(conn, -ECOMM);
3311 /* drop ref from kgnilnd_validate_tx_ev_id */
3312 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3313 kgnilnd_conn_decref(conn);
3318 kgnilnd_check_fma_send_cq(kgn_device_t *dev)
3322 kgn_tx_ev_id_t ev_id;
3323 kgn_tx_t *tx = NULL;
3324 kgn_conn_t *conn = NULL;
3325 int queued_fma, saw_reply, rc;
3326 long num_processed = 0;
3327 struct list_head *ctmp, *ctmpN;
3330 /* make sure we don't keep looping if we need to reset */
3331 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3332 return num_processed;
3335 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3337 /* we didn't get the mutex, so return that there is still work
3342 rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
3343 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3345 if (rrc == GNI_RC_NOT_DONE) {
3347 "SMSG send CQ %d not ready (data %#llx) "
3348 "processed %ld\n", dev->gnd_id, event_data,
3351 if (num_processed > 0) {
3352 spin_lock(&dev->gnd_lock);
3353 if (!list_empty(&dev->gnd_delay_conns)) {
3354 list_for_each_safe(ctmp, ctmpN, &dev->gnd_delay_conns) {
3355 conn = list_entry(ctmp, kgn_conn_t, gnc_delaylist);
3356 list_del_init(&conn->gnc_delaylist);
3357 CDEBUG(D_NET, "Moving Conn %p from delay queue to ready_queue\n", conn);
3358 kgnilnd_schedule_conn_nolock(conn);
3360 spin_unlock(&dev->gnd_lock);
3361 kgnilnd_schedule_device(dev);
3363 spin_unlock(&dev->gnd_lock);
3366 return num_processed;
3369 dev->gnd_sched_alive = jiffies;
3372 LASSERTF(!GNI_CQ_OVERRUN(event_data),
3373 "this is bad, somehow our credits didn't "
3374 "protect us from CQ overrun\n");
3375 LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
3376 "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
3377 event_data, GNI_CQ_GET_TYPE(event_data));
3379 /* if SMSG couldn't handle an error, time for conn to die */
3380 if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
3383 /* need to take the write_lock to ensure atomicity
3384 * on the conn state if we need to close it */
3385 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
3386 conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
3388 /* Conn was destroyed? */
3390 "SMSG CQID lookup %#llx failed\n",
3391 GNI_CQ_GET_INST_ID(event_data));
3392 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3396 kgnilnd_cq_error_str(event_data, &err_str, 256);
3397 CNETERR("SMSG send error to %s: rc %d (%s)\n",
3398 libcfs_nid2str(conn->gnc_peer->gnp_nid),
3400 kgnilnd_close_conn_locked(conn, -ECOMM);
3402 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3404 /* no need to process rest of this tx -
3405 * it is getting canceled */
3409 /* fall through to GNI_RC_SUCCESS case */
3410 ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);
3412 kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
3413 if (conn == NULL || tx == NULL) {
3414 /* either conn or tx was already nuked and this is a "late"
3415 * completion, so drop it */
3419 tx->tx_conn->gnc_last_tx_cq = jiffies;
3420 if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
3421 set_mb(conn->gnc_last_noop_cq, jiffies);
3424 /* lock tx_list_state and tx_state */
3425 kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
3426 spin_lock(&tx->tx_conn->gnc_list_lock);
3428 GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
3429 "state not GNILND_TX_LIVE_FMAQ", NULL);
3430 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
3431 "not waiting for completion", NULL);
3433 GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
3436 tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;
3438 /* This will trigger other FMA sends that were
3439 * pending this completion */
3440 queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);
3442 /* we either did not expect reply or we already got it */
3443 saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);
3445 spin_unlock(&tx->tx_conn->gnc_list_lock);
3446 kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
3449 CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
3451 libcfs_nid2str(conn->gnc_peer->gnp_nid));
3452 kgnilnd_schedule_conn(conn);
3455 /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
3456 * If saw_reply is true we know that the tx is safe to use as the other thread
3457 * is already finished with it.
3461 /* no longer need to track on the live_fmaq */
3462 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3464 if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
3465 /* we already got reply & were waiting for
3466 * completion of initial send */
3467 /* to initiate RDMA transaction */
3468 GNIDBG_TX(D_NET, tx,
3469 "Pending RDMA 0x%p type 0x%02x",
3470 tx->tx_msg.gnm_type);
3471 tx->tx_state &= ~GNILND_TX_PENDING_RDMA;
3472 rc = kgnilnd_send_mapped_tx(tx, 0);
3473 GNITX_ASSERTF(tx, rc == 0, "RDMA send failed: %d\n", rc);
3475 /* we are done with this tx */
3476 GNIDBG_TX(D_NET, tx,
3477 "Done with tx type 0x%02x",
3478 tx->tx_msg.gnm_type);
3479 kgnilnd_tx_done(tx, tx->tx_rc);
3483 /* drop ref from kgnilnd_validate_tx_ev_id */
3484 kgnilnd_admin_decref(conn->gnc_tx_in_use);
3485 kgnilnd_conn_decref(conn);
3487 /* if we are waiting for a REPLY, we'll handle the tx then */
3488 } /* end for loop */
3492 kgnilnd_check_fma_rcv_cq(kgn_device_t *dev)
3497 long num_processed = 0;
3498 struct list_head *conns;
3499 struct list_head *tmp;
3503 /* make sure we don't keep looping if we need to reset */
3504 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3505 return num_processed;
3508 rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
3510 /* we didn't get the mutex, so return that there is still work
3514 rrc = kgnilnd_cq_get_event(dev->gnd_rcv_fma_cqh, &event_data);
3515 kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
3517 if (rrc == GNI_RC_NOT_DONE) {
3518 CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
3520 dev->gnd_id, event_data, num_processed);
3521 return num_processed;
3523 dev->gnd_sched_alive = jiffies;
3526 /* this is the only CQ that can really handle transient
3528 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CQ_GET_EVENT)) {
3529 rrc = cfs_fail_val ? cfs_fail_val
3530 : GNI_RC_ERROR_RESOURCE;
3531 if (rrc == GNI_RC_ERROR_RESOURCE) {
3532 /* set overrun too */
3533 event_data |= (1UL << 63);
3534 LASSERTF(GNI_CQ_OVERRUN(event_data),
3535 "(1UL << 63) is no longer the bit to set to indicate CQ_OVERRUN\n");
3538 /* sender should get error event too and take care
3539 of failed transaction by re-transmitting */
3540 if (rrc == GNI_RC_TRANSACTION_ERROR) {
3541 CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
3545 if (likely(!GNI_CQ_OVERRUN(event_data))) {
3546 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3547 conn = kgnilnd_cqid2conn_locked(
3548 GNI_CQ_GET_INST_ID(event_data));
3550 CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
3551 "failed, dropping event %#llx\n",
3552 GNI_CQ_GET_INST_ID(event_data),
3555 CDEBUG(D_NET, "SMSG RX: CQID %llu "
3557 GNI_CQ_GET_INST_ID(event_data),
3558 conn, conn->gnc_peer ?
3559 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
3562 conn->gnc_last_rx_cq = jiffies;
3564 /* stash first rx so we can clear out purgatory.
3566 if (conn->gnc_first_rx == 0) {
3567 conn->gnc_first_rx = jiffies;
3569 kgnilnd_peer_alive(conn->gnc_peer);
3570 kgnilnd_schedule_conn(conn);
3572 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3576 /* FMA CQ has overflowed: check ALL conns */
3577 CNETERR("SMSG RX CQ overflow: scheduling ALL "
3578 "conns on device %d\n", dev->gnd_id);
3580 for (rc = 0; rc < *kgnilnd_tunables.kgn_peer_hash_size; rc++) {
3582 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
3583 conns = &kgnilnd_data.kgn_conns[rc];
3585 list_for_each(tmp, conns) {
3586 conn = list_entry(tmp, kgn_conn_t,
3589 if (conn->gnc_device == dev) {
3590 kgnilnd_schedule_conn(conn);
3591 conn->gnc_last_rx_cq = jiffies;
3595 /* don't block write lockers for too long... */
3596 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
3601 /* try_map_if_full should only be used when processing TX from list of
3602 * backlog TX waiting on mappings to free up
3605 * try_map_if_full = 0: 0 (sent or queued), (-|+)errno failure of kgnilnd_sendmsg
3606 * try_map_if_full = 1: 0 (sent), -ENOMEM for caller to requeue, (-|+)errno failure of kgnilnd_sendmsg */
3609 kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
3611 /* slight bit of race if multiple people calling, but at worst we'll have
3612 * order altered just a bit... which would not be determenistic anyways */
3613 int rc = atomic_read(&tx->tx_conn->gnc_device->gnd_nq_map);
3615 GNIDBG_TX(D_NET, tx, "try %d nq_map %d", try_map_if_full, rc);
3617 /* We know that we have a GART reservation that should guarantee forward progress.
3618 * This means we don't need to take any extraordinary efforts if we are failing
3619 * mappings here - even if we are holding a very small number of these. */
3621 if (try_map_if_full || (rc == 0)) {
3622 rc = kgnilnd_map_buffer(tx);
3625 /* rc should be 0 if we mapped successfully here, if non-zero
3626 * we are queueing */
3628 /* if try_map_if_full set, they handle requeuing */
3629 if (unlikely(try_map_if_full)) {
3632 spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
3633 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
3634 spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
3635 /* make sure we wake up sched to run this */
3636 kgnilnd_schedule_device(tx->tx_conn->gnc_device);
3637 /* return 0 as this is now queued for later sending */
3642 switch (tx->tx_msg.gnm_type) {
3646 /* GET_REQ and PUT_ACK are outbound messages sending our mapping key to
3647 * remote node where the RDMA will be started
3648 * Special case -EAGAIN logic - this should just queued as if the mapping couldn't
3649 * be satisified. The rest of the errors are "hard" errors that require
3650 * upper layers to handle themselves.
3651 * If kgnilnd_post_rdma returns a resource error, kgnilnd_rdma will put
3652 * the tx back on the TX_MAPQ. When this tx is pulled back off the MAPQ,
3653 * it's gnm_type will now be GNILND_MSG_PUT_DONE or
3654 * GNILND_MSG_GET_DONE_REV.
3656 case GNILND_MSG_GET_REQ:
3657 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3658 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3659 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3660 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3661 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3662 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GET_REQ_AGAIN)) {
3663 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3665 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3666 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3668 case GNILND_MSG_PUT_ACK:
3669 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3670 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3671 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN)) {
3672 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3674 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3675 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3678 /* PUT_REQ and GET_DONE are where we do the actual RDMA */
3679 case GNILND_MSG_PUT_DONE:
3680 case GNILND_MSG_PUT_REQ:
3681 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE,
3682 &tx->tx_putinfo.gnpam_desc,
3683 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3684 tx->tx_putinfo.gnpam_dst_cookie);
3685 RETURN(try_map_if_full ? rc : 0);
3687 case GNILND_MSG_GET_DONE:
3688 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE,
3689 &tx->tx_getinfo.gngm_desc,
3690 tx->tx_lntmsg[0]->msg_len,
3691 tx->tx_getinfo.gngm_cookie);
3692 RETURN(try_map_if_full ? rc : 0);
3694 case GNILND_MSG_PUT_REQ_REV:
3695 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_key = tx->tx_map_key;
3696 tx->tx_msg.gnm_u.get.gngm_cookie = tx->tx_id.txe_cookie;
3697 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_addr = (__u64)((unsigned long)tx->tx_buffer);
3698 tx->tx_msg.gnm_u.get.gngm_desc.gnrd_nob = tx->tx_nob;
3699 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3700 kgnilnd_compute_rdma_cksum(tx, tx->tx_nob);
3701 tx->tx_msg.gnm_u.get.gngm_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3703 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3705 case GNILND_MSG_PUT_DONE_REV:
3706 rc = kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
3707 &tx->tx_getinfo.gngm_desc,
3709 tx->tx_getinfo.gngm_cookie);
3710 RETURN(try_map_if_full ? rc : 0);
3712 case GNILND_MSG_GET_ACK_REV:
3713 tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_key = tx->tx_map_key;
3714 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3715 /* LNET_GETS are a special case for parse */
3716 kgnilnd_compute_rdma_cksum(tx, tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob);
3717 tx->tx_msg.gnm_u.putack.gnpam_payload_cksum = tx->tx_msg.gnm_payload_cksum;
3719 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PUT_ACK_AGAIN))
3720 tx->tx_state |= GNILND_TX_FAIL_SMSG;
3722 /* redirect to FMAQ on failure, no need to infinite loop here in MAPQ */
3723 rc = kgnilnd_sendmsg(tx, NULL, 0, &tx->tx_conn->gnc_list_lock, GNILND_TX_FMAQ);
3725 case GNILND_MSG_GET_DONE_REV:
3726 case GNILND_MSG_GET_REQ_REV:
3727 rc = kgnilnd_rdma(tx, GNILND_MSG_GET_DONE_REV,
3728 &tx->tx_putinfo.gnpam_desc,
3729 tx->tx_putinfo.gnpam_desc.gnrd_nob,
3730 tx->tx_putinfo.gnpam_dst_cookie);
3731 RETURN(try_map_if_full ? rc : 0);
3739 kgnilnd_process_fmaq(kgn_conn_t *conn)
3742 kgn_tx_t *tx = NULL;
3743 void *buffer = NULL;
3744 unsigned int nob = 0;
3747 /* NB 1. kgnilnd_sendmsg() may fail if I'm out of credits right now.
3748 * However I will be rescheduled by an FMA completion event
3749 * when I eventually get some.
3750 * NB 2. Sampling gnc_state here races with setting it elsewhere.
3751 * But it doesn't matter if I try to send a "real" message just
3752 * as I start closing because I'll get scheduled to send the
3755 /* Short circuit if the ep_handle is null we cant send anyway. */
3756 if (conn->gnc_ephandle == NULL)
3759 LASSERTF(!conn->gnc_close_sent, "Conn %p close was sent\n", conn);
3761 spin_lock(&conn->gnc_list_lock);
3763 if (list_empty(&conn->gnc_fmaq)) {
3764 int keepalive = GNILND_TO2KA(conn->gnc_timeout);
3766 spin_unlock(&conn->gnc_list_lock);
3768 if (time_after_eq(jiffies, conn->gnc_last_tx + cfs_time_seconds(keepalive))) {
3769 CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%d)) "
3770 "last %lu/%lu/%lu %lus/%lus/%lus\n",
3771 libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
3772 cfs_duration_sec(jiffies - conn->gnc_last_tx),
3774 conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
3775 conn->gnc_last_noop_cq,
3776 cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
3777 cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
3778 cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
3779 atomic_inc(&conn->gnc_sched_noop);
3780 set_mb(conn->gnc_last_noop_want, jiffies);
3782 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
3785 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
3789 rc = kgnilnd_set_tx_id(tx, conn);
3791 kgnilnd_tx_done(tx, rc);
3797 tx = list_first_entry(&conn->gnc_fmaq, kgn_tx_t, tx_list);
3798 /* move from fmaq to allocd, kgnilnd_sendmsg will move to live_fmaq */
3799 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
3800 more_to_do = !list_empty(&conn->gnc_fmaq);
3801 spin_unlock(&conn->gnc_list_lock);
3804 /* if there is no real TX or no NOOP to send, bail */
3809 if (!tx->tx_retrans)
3810 tx->tx_cred_wait = jiffies;
3812 GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
3813 "tx with zero id", NULL);
3815 CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
3816 tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
3817 tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
3821 switch (tx->tx_msg.gnm_type) {
3825 case GNILND_MSG_NOOP:
3826 case GNILND_MSG_CLOSE:
3827 case GNILND_MSG_IMMEDIATE:
3828 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3829 buffer = tx->tx_buffer;
3833 case GNILND_MSG_GET_DONE:
3834 case GNILND_MSG_PUT_DONE:
3835 case GNILND_MSG_PUT_DONE_REV:
3836 case GNILND_MSG_GET_DONE_REV:
3837 case GNILND_MSG_PUT_NAK:
3838 case GNILND_MSG_GET_NAK:
3839 case GNILND_MSG_GET_NAK_REV:
3840 case GNILND_MSG_PUT_NAK_REV:
3841 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
3844 case GNILND_MSG_PUT_REQ:
3845 case GNILND_MSG_GET_REQ_REV:
3846 tx->tx_msg.gnm_u.putreq.gnprm_cookie = tx->tx_id.txe_cookie;
3848 case GNILND_MSG_PUT_ACK:
3849 case GNILND_MSG_PUT_REQ_REV:
3850 case GNILND_MSG_GET_ACK_REV:
3851 case GNILND_MSG_GET_REQ:
3852 /* This is really only to handle the retransmit of SMSG once these
3853 * two messages are setup in send_mapped_tx */
3854 tx->tx_state = GNILND_TX_WAITING_COMPLETION | GNILND_TX_WAITING_REPLY;
3858 if (likely(rc == 0)) {
3859 rc = kgnilnd_sendmsg(tx, buffer, nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
3863 /* don't explicitly reschedule here - we are short credits and will rely on
3864 * kgnilnd_sendmsg to resched the conn if need be */
3866 } else if (rc < 0) {
3867 /* bail: it wasn't sent and we didn't get EAGAIN indicating we should retrans
3868 * almost certainly a software bug, but lets play nice with the other kids */
3869 kgnilnd_tx_done(tx, rc);
3870 /* just for fun, kick peer in arse - resetting conn might help to correct
3871 * this almost certainly buggy software caused return code */
3872 kgnilnd_close_conn(conn, rc);
3876 CDEBUG(D_NET, "Rescheduling %p (more to do)\n", conn);
3877 kgnilnd_schedule_conn(conn);
3882 kgnilnd_process_rdmaq(kgn_device_t *dev)
3887 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMAQ)) {
3891 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3892 unsigned long dead_bump;
3895 /* if we think we need to adjust, take lock to serialize and recheck */
3896 spin_lock(&dev->gnd_rdmaq_lock);
3897 if (time_after_eq(jiffies, dev->gnd_rdmaq_deadline)) {
3898 del_singleshot_timer_sync(&dev->gnd_rdmaq_timer);
3900 dead_bump = cfs_time_seconds(1) / *kgnilnd_tunables.kgn_rdmaq_intervals;
3902 /* roll the bucket forward */
3903 dev->gnd_rdmaq_deadline = jiffies + dead_bump;
3905 if (kgnilnd_data.kgn_rdmaq_override &&
3906 (*kgnilnd_tunables.kgn_rdmaq_intervals != 0)) {
3907 new_ok = kgnilnd_data.kgn_rdmaq_override / *kgnilnd_tunables.kgn_rdmaq_intervals;
3912 /* roll current outstanding forward to make sure we carry outstanding
3913 * committment forward
3914 * new_ok starts out as the whole interval value
3915 * - first subtract bytes_out from last interval, as that would push us over
3916 * strict limits for this interval
3917 * - second, set bytes_ok to new_ok to ensure it doesn't exceed the current auth
3919 * there is a small race here if someone is actively processing mappings and
3920 * adding to rdmaq_bytes_out, but it should be small as the mappings are triggered
3921 * quite quickly after kgnilnd_auth_rdma_bytes gives us the go-ahead
3922 * - if this gives us problems in the future, we could use a read/write lock
3923 * to protect the resetting of these values */
3924 new_ok -= atomic64_read(&dev->gnd_rdmaq_bytes_out);
3925 atomic64_set(&dev->gnd_rdmaq_bytes_ok, new_ok);
3927 CDEBUG(D_NET, "resetting rdmaq bytes to %ld, deadline +%lu -> %lu, "
3928 "current out %ld\n",
3929 atomic64_read(&dev->gnd_rdmaq_bytes_ok), dead_bump, dev->gnd_rdmaq_deadline,
3930 atomic64_read(&dev->gnd_rdmaq_bytes_out));
3932 spin_unlock(&dev->gnd_rdmaq_lock);
3935 spin_lock(&dev->gnd_rdmaq_lock);
3936 while (!list_empty(&dev->gnd_rdmaq)) {
3939 /* make sure we break out early on quiesce */
3940 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
3941 /* always break with lock held - we unlock outside loop */
3945 tx = list_first_entry(&dev->gnd_rdmaq, kgn_tx_t, tx_list);
3946 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
3949 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
3950 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
3951 /* if conn is dying, mark tx in tx_ref_table for
3952 * kgnilnd_complete_closed_conn to finish up */
3953 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
3955 /* tx was moved to DYING, get next */
3958 spin_unlock(&dev->gnd_rdmaq_lock);
3960 rc = kgnilnd_auth_rdma_bytes(dev, tx);
3961 spin_lock(&dev->gnd_rdmaq_lock);
3964 /* no ticket! add back to head */
3965 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_RDMAQ, 0);
3966 /* clear found_work so scheduler threads wait for timer */
3970 /* TX is GO for launch */
3971 tx->tx_qtime = jiffies;
3972 kgnilnd_send_mapped_tx(tx, 0);
3976 spin_unlock(&dev->gnd_rdmaq_lock);
3982 kgnilnd_swab_rdma_desc(kgn_rdma_desc_t *d)
3984 __swab64s(&d->gnrd_key.qword1);
3985 __swab64s(&d->gnrd_key.qword2);
3986 __swab64s(&d->gnrd_addr);
3987 __swab32s(&d->gnrd_nob);
3990 #define kgnilnd_match_reply_either(w, x, y, z) _kgnilnd_match_reply(w, x, y, z)
3991 #define kgnilnd_match_reply(x, y, z) _kgnilnd_match_reply(x, y, GNILND_MSG_NONE, z)
3994 _kgnilnd_match_reply(kgn_conn_t *conn, int type1, int type2, __u64 cookie)
3996 kgn_tx_ev_id_t ev_id;
3999 /* we use the cookie from the original TX, so we can find the match
4000 * by parsing that and using the txe_idx */
4001 ev_id.txe_cookie = cookie;
4003 tx = conn->gnc_tx_ref_table[ev_id.txe_idx];
4006 /* check tx to make sure kgni didn't eat it */
4007 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
4008 "came back from kgni with bad magic %x\n", tx->tx_msg.gnm_magic);
4010 GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
4011 (tx->tx_id.txe_cookie = cookie)),
4012 "conn 0x%p->%s tx_ref_table hosed: wanted "
4013 "txe_cookie %#llx txe_idx %d "
4014 "found tx %p cookie %#llx txe_idx %d\n",
4015 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4016 cookie, ev_id.txe_idx,
4017 tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
4019 LASSERTF((((tx->tx_msg.gnm_type == type1) || (tx->tx_msg.gnm_type == type2)) &&
4020 (tx->tx_state & GNILND_TX_WAITING_REPLY)),
4021 "Unexpected TX type (%x, %x or %x) "
4022 "or state (%x, expected +%x) "
4023 "matched reply from %s\n",
4024 tx->tx_msg.gnm_type, type1, type2,
4025 tx->tx_state, GNILND_TX_WAITING_REPLY,
4026 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4028 CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
4029 type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
4035 kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
4038 kgn_conn_t *conn = tx->tx_conn;
4039 __u64 nob = tx->tx_nob;
4040 __u32 physnop = tx->tx_phys_npages;
4041 int id = tx->tx_id.txe_smsg_id;
4042 int buftype = tx->tx_buftype;
4043 gni_mem_handle_t hndl;
4044 hndl.qword1 = tx->tx_map_key.qword1;
4045 hndl.qword2 = tx->tx_map_key.qword2;
4047 spin_lock(&conn->gnc_list_lock);
4049 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4050 "not waiting for reply", NULL);
4053 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4055 if (rc == -EFAULT) {
4056 CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
4057 rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
4059 if(*kgnilnd_tunables.kgn_efault_lbug) {
4060 GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
4061 "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
4063 libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
4064 tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
4065 kgnilnd_tx_state2str(tx->tx_list_state),
4066 cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
4071 if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
4072 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4073 /* sample under lock as follow on steps require gnc_list_lock
4074 * - or call kgnilnd_tx_done which requires no locks held over
4075 * call to lnet_finalize */
4078 spin_unlock(&conn->gnc_list_lock);
4081 kgnilnd_tx_done(tx, tx->tx_rc);
4086 kgnilnd_finalize_rx_done(kgn_tx_t *tx, kgn_msg_t *msg)
4089 kgn_conn_t *conn = tx->tx_conn;
4091 atomic_inc(&conn->gnc_device->gnd_rdma_nrx);
4092 atomic64_add(tx->tx_nob, &conn->gnc_device->gnd_rdma_rxbytes);
4094 /* the gncm_retval is passed in for PUTs */
4095 rc = kgnilnd_verify_rdma_cksum(tx, msg->gnm_payload_cksum,
4096 msg->gnm_u.completion.gncm_retval);
4098 kgnilnd_complete_tx(tx, rc);
4102 kgnilnd_check_fma_rx(kgn_conn_t *conn)
4110 kgn_peer_t *peer = conn->gnc_peer;
4113 __u16 tmp_cksum = 0, msg_cksum = 0;
4114 int repost = 1, saw_complete;
4115 unsigned long timestamp, newest_last_rx, timeout;
4119 /* Short circuit if the ep_handle is null.
4120 * It's likely that its about to be closed as stale.
4122 if (conn->gnc_ephandle == NULL)
4125 timestamp = jiffies;
4126 kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
4127 /* delay in jiffies - we are really concerned only with things that
4128 * result in a schedule() or really holding this off for long times .
4129 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
4130 conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;
4132 /* Resample current time as we have no idea how long it took to get the mutex */
4133 timestamp = jiffies;
4135 /* We check here when the last time we received an rx, we do this before
4136 * we call getnext in case the thread has been blocked for a while. If we
4137 * havent received an rx since our timeout value we close the connection
4138 * as we should assume the other side has closed the connection. This will
4139 * stop us from sending replies to a mailbox that is already in purgatory.
4142 timeout = cfs_time_seconds(conn->gnc_timeout);
4143 newest_last_rx = GNILND_LASTRX(conn);
4145 /* Error injection to validate that timestamp checking works and closing the conn */
4146 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RECV_TIMEOUT)) {
4147 timestamp = timestamp + (GNILND_TIMEOUTRX(timeout) * 2);
4150 if (time_after_eq(timestamp, newest_last_rx + (GNILND_TIMEOUTRX(timeout)))) {
4151 GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn, "Cant receive from %s after timeout lapse of %lu; TO %lu",
4152 libcfs_nid2str(conn->gnc_peer->gnp_nid),
4153 cfs_duration_sec(timestamp - newest_last_rx),
4154 cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
4155 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4157 kgnilnd_close_conn(conn, rc);
4161 rrc = kgnilnd_smsg_getnext(conn->gnc_ephandle, &prefix);
4163 if (rrc == GNI_RC_NOT_DONE) {
4164 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4165 CDEBUG(D_INFO, "SMSG RX empty conn 0x%p\n", conn);
4169 /* Instead of asserting when we get mailbox corruption lets attempt to
4170 * close the conn and recover. We can put the conn/mailbox into
4171 * purgatory and let purgatory deal with the problem. If we see
4172 * this NETTERROR reported on production systems in large amounts
4173 * we will need to revisit the state machine to see if we can tighten
4174 * it up further to improve data protection.
4177 if (rrc == GNI_RC_INVALID_STATE) {
4178 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4179 GNIDBG_CONN(D_NETERROR | D_CONSOLE, conn, "Mailbox corruption "
4180 "detected closing conn %p from peer %s\n", conn,
4181 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4183 kgnilnd_close_conn(conn, rc);
4187 LASSERTF(rrc == GNI_RC_SUCCESS,
4188 "bad rc %d on conn %p from peer %s\n",
4189 rrc, conn, libcfs_nid2str(peer->gnp_nid));
4191 msg = (kgn_msg_t *)prefix;
4193 rx = kgnilnd_alloc_rx();
4195 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4196 kgnilnd_release_msg(conn);
4197 GNIDBG_MSG(D_NETERROR, msg, "Dropping SMSG RX from 0x%p->%s, no RX memory",
4198 conn, libcfs_nid2str(peer->gnp_nid));
4202 GNIDBG_MSG(D_INFO, msg, "SMSG RX on %p", conn);
4204 timestamp = conn->gnc_last_rx;
4205 seq = last_seq = atomic_read(&conn->gnc_rx_seq);
4206 atomic_inc(&conn->gnc_rx_seq);
4208 conn->gnc_last_rx = jiffies;
4209 /* stash first rx so we can clear out purgatory
4211 if (conn->gnc_first_rx == 0)
4212 conn->gnc_first_rx = jiffies;
4214 /* needs to linger to protect gnc_rx_seq like we do with gnc_tx_seq */
4215 kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
4216 kgnilnd_peer_alive(conn->gnc_peer);
4219 rx->grx_conn = conn;
4221 rx->grx_received = current_kernel_time();
4223 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NET_LOOKUP)) {
4226 rc = kgnilnd_find_net(msg->gnm_srcnid, &net);
4232 kgnilnd_net_decref(net);
4235 if (*kgnilnd_tunables.kgn_checksum && !msg->gnm_cksum)
4236 GNIDBG_MSG(D_WARNING, msg, "no msg header checksum when enabled");
4238 /* XXX Nic: Do we need to swab cksum */
4239 if (msg->gnm_cksum != 0) {
4240 msg_cksum = msg->gnm_cksum;
4242 tmp_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
4244 if (tmp_cksum != msg_cksum) {
4245 GNIDBG_MSG(D_NETERROR, msg, "Bad hdr checksum (%x expected %x)",
4246 tmp_cksum, msg_cksum);
4247 kgnilnd_dump_msg(D_BUFFS, msg);
4252 /* restore checksum for future debug messages */
4253 msg->gnm_cksum = tmp_cksum;
4255 if (msg->gnm_magic != GNILND_MSG_MAGIC) {
4256 if (__swab32(msg->gnm_magic) != GNILND_MSG_MAGIC) {
4257 GNIDBG_MSG(D_NETERROR, msg, "Unexpected magic %08x from %s",
4258 msg->gnm_magic, libcfs_nid2str(peer->gnp_nid));
4263 __swab32s(&msg->gnm_magic);
4264 __swab16s(&msg->gnm_version);
4265 __swab16s(&msg->gnm_type);
4266 __swab64s(&msg->gnm_srcnid);
4267 __swab64s(&msg->gnm_connstamp);
4268 __swab32s(&msg->gnm_seq);
4270 /* NB message type checked below; NOT here... */
4271 switch (msg->gnm_type) {
4272 case GNILND_MSG_GET_ACK_REV:
4273 case GNILND_MSG_PUT_ACK:
4274 kgnilnd_swab_rdma_desc(&msg->gnm_u.putack.gnpam_desc);
4277 case GNILND_MSG_PUT_REQ_REV:
4278 case GNILND_MSG_GET_REQ:
4279 kgnilnd_swab_rdma_desc(&msg->gnm_u.get.gngm_desc);
4287 if (msg->gnm_version != GNILND_MSG_VERSION) {
4288 GNIDBG_MSG(D_NETERROR, msg, "Unexpected protocol version %d from %s",
4289 msg->gnm_version, libcfs_nid2str(peer->gnp_nid));
4294 if (LNET_NIDADDR(msg->gnm_srcnid) != LNET_NIDADDR(peer->gnp_nid)) {
4295 GNIDBG_MSG(D_NETERROR, msg, "Unexpected peer %s from %s",
4296 libcfs_nid2str(msg->gnm_srcnid),
4297 libcfs_nid2str(peer->gnp_nid));
4302 if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
4303 GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
4304 " expected) from %s",
4305 msg->gnm_connstamp, conn->gnc_peer_connstamp,
4306 libcfs_nid2str(peer->gnp_nid));
4311 if (msg->gnm_seq != seq) {
4312 GNIDBG_MSG(D_NETERROR, msg, "Unexpected sequence number %d(%d expected) from %s",
4313 msg->gnm_seq, seq, libcfs_nid2str(peer->gnp_nid));
4318 atomic_inc(&conn->gnc_device->gnd_short_nrx);
4320 if (msg->gnm_type == GNILND_MSG_CLOSE) {
4321 CDEBUG(D_NETTRACE, "%s sent us CLOSE msg\n",
4322 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4323 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4324 conn->gnc_close_recvd = GNILND_CLOSE_RX;
4325 conn->gnc_peer_error = msg->gnm_u.completion.gncm_retval;
4326 /* double check state with lock held */
4327 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
4328 /* only error if we are not already closing */
4329 if (conn->gnc_peer_error == -ETIMEDOUT) {
4330 unsigned long now = jiffies;
4331 CNETERR("peer 0x%p->%s closed connection 0x%p due to timeout. "
4333 "RX %d @ %lus/%lus; TX %d @ %lus/%lus; "
4334 "NOOP %lus/%lus/%lus; sched %lus/%lus/%lus ago\n",
4335 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
4337 cfs_duration_sec(now - timestamp),
4338 cfs_duration_sec(now - conn->gnc_last_rx_cq),
4339 atomic_read(&conn->gnc_tx_seq),
4340 cfs_duration_sec(now - conn->gnc_last_tx),
4341 cfs_duration_sec(now - conn->gnc_last_tx_cq),
4342 cfs_duration_sec(now - conn->gnc_last_noop_want),
4343 cfs_duration_sec(now - conn->gnc_last_noop_sent),
4344 cfs_duration_sec(now - conn->gnc_last_noop_cq),
4345 cfs_duration_sec(now - conn->gnc_last_sched_ask),
4346 cfs_duration_sec(now - conn->gnc_last_sched_do),
4347 cfs_duration_sec(now - conn->gnc_device->gnd_sched_alive));
4349 kgnilnd_close_conn_locked(conn, -ECONNRESET);
4351 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4355 if (conn->gnc_close_recvd) {
4356 GNIDBG_MSG(D_NETERROR, msg, "Unexpected message %s(%d/%d) after CLOSE from %s",
4357 kgnilnd_msgtype2str(msg->gnm_type),
4358 msg->gnm_type, conn->gnc_close_recvd,
4359 libcfs_nid2str(conn->gnc_peer->gnp_nid));
4364 if (conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4365 /* XXX Nic: log message received on bad connection state */
4369 switch (msg->gnm_type) {
4370 case GNILND_MSG_NOOP:
4371 /* Nothing to do; just a keepalive */
4374 case GNILND_MSG_IMMEDIATE:
4375 /* only get SMSG payload for IMMEDIATE */
4376 atomic64_add(msg->gnm_payload_len, &conn->gnc_device->gnd_short_rxbytes);
4377 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.immediate.gnim_hdr,
4378 msg->gnm_srcnid, rx, 0);
4381 case GNILND_MSG_GET_REQ_REV:
4382 case GNILND_MSG_PUT_REQ:
4383 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.putreq.gnprm_hdr,
4384 msg->gnm_srcnid, rx, 1);
4387 case GNILND_MSG_GET_NAK_REV:
4388 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_GET_REQ_REV, GNILND_MSG_GET_ACK_REV,
4389 msg->gnm_u.completion.gncm_cookie);
4393 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4395 case GNILND_MSG_PUT_NAK:
4396 tx = kgnilnd_match_reply_either(conn, GNILND_MSG_PUT_REQ, GNILND_MSG_PUT_ACK,
4397 msg->gnm_u.completion.gncm_cookie);
4401 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4403 case GNILND_MSG_PUT_ACK:
4404 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ,
4405 msg->gnm_u.putack.gnpam_src_cookie);
4409 /* store putack data for later: deferred rdma or re-try */
4410 tx->tx_putinfo = msg->gnm_u.putack;
4413 spin_lock(&tx->tx_conn->gnc_list_lock);
4415 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4416 "not waiting for reply", NULL);
4418 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4420 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4421 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4422 /* sample under lock as follow on steps require gnc_list_lock
4423 * - or call kgnilnd_tx_done which requires no locks held over
4424 * call to lnet_finalize */
4427 /* cannot launch rdma if still waiting for fma-msg completion */
4428 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4429 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4430 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4432 spin_unlock(&tx->tx_conn->gnc_list_lock);
4435 rc = kgnilnd_send_mapped_tx(tx, 0);
4437 kgnilnd_tx_done(tx, rc);
4440 case GNILND_MSG_GET_ACK_REV:
4441 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ_REV,
4442 msg->gnm_u.putack.gnpam_src_cookie);
4446 /* store putack data for later: deferred rdma or re-try */
4447 tx->tx_putinfo = msg->gnm_u.putack;
4449 spin_lock(&tx->tx_conn->gnc_list_lock);
4451 GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_REPLY,
4452 "not waiting for reply", NULL);
4454 tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
4456 if (likely(!(tx->tx_state & GNILND_TX_WAITING_COMPLETION))) {
4457 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
4458 /* sample under lock as follow on steps require gnc_list_lock
4459 * - or call kgnilnd_tx_done which requires no locks held over
4460 * call to lnet_finalize */
4463 /* cannot launch rdma if still waiting for fma-msg completion */
4464 CDEBUG(D_NET, "tx 0x%p type 0x%02x will need to "
4465 "wait for SMSG completion\n", tx, tx->tx_msg.gnm_type);
4466 tx->tx_state |= GNILND_TX_PENDING_RDMA;
4468 spin_unlock(&tx->tx_conn->gnc_list_lock);
4471 rc = kgnilnd_send_mapped_tx(tx, 0);
4473 kgnilnd_tx_done(tx, rc);
4476 case GNILND_MSG_PUT_DONE:
4477 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_ACK,
4478 msg->gnm_u.completion.gncm_cookie);
4482 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4483 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4484 "bad tx buftype %d", tx->tx_buftype);
4486 kgnilnd_finalize_rx_done(tx, msg);
4488 case GNILND_MSG_PUT_REQ_REV:
4489 case GNILND_MSG_GET_REQ:
4490 rc = lnet_parse(net->gnn_ni, &msg->gnm_u.get.gngm_hdr,
4491 msg->gnm_srcnid, rx, 1);
4495 case GNILND_MSG_GET_NAK:
4496 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4497 msg->gnm_u.completion.gncm_cookie);
4501 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4502 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4503 "bad tx buftype %d", tx->tx_buftype);
4505 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4508 case GNILND_MSG_GET_DONE:
4509 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_REQ,
4510 msg->gnm_u.completion.gncm_cookie);
4514 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4515 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4516 "bad tx buftype %d", tx->tx_buftype);
4518 lnet_set_reply_msg_len(net->gnn_ni, tx->tx_lntmsg[1],
4519 msg->gnm_u.completion.gncm_retval);
4521 kgnilnd_finalize_rx_done(tx, msg);
4523 case GNILND_MSG_GET_DONE_REV:
4524 tx = kgnilnd_match_reply(conn, GNILND_MSG_GET_ACK_REV,
4525 msg->gnm_u.completion.gncm_cookie);
4529 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4530 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4531 "bad tx buftype %d", tx->tx_buftype);
4533 kgnilnd_finalize_rx_done(tx, msg);
4536 case GNILND_MSG_PUT_DONE_REV:
4537 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4538 msg->gnm_u.completion.gncm_cookie);
4543 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4544 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4545 "bad tx buftype %d", tx->tx_buftype);
4547 kgnilnd_finalize_rx_done(tx, msg);
4549 case GNILND_MSG_PUT_NAK_REV:
4550 tx = kgnilnd_match_reply(conn, GNILND_MSG_PUT_REQ_REV,
4551 msg->gnm_u.completion.gncm_cookie);
4556 GNITX_ASSERTF(tx, tx->tx_buftype == GNILND_BUF_PHYS_MAPPED ||
4557 tx->tx_buftype == GNILND_BUF_VIRT_MAPPED,
4558 "bad tx buftype %d", tx->tx_buftype);
4560 kgnilnd_complete_tx(tx, msg->gnm_u.completion.gncm_retval);
4565 if (rc < 0) /* protocol/comms error */
4566 kgnilnd_close_conn(conn, rc);
4568 if (repost && rx != NULL) {
4569 kgnilnd_consume_rx(rx);
4572 /* we got an event so assume more there and call for reschedule */
4574 kgnilnd_schedule_conn(conn);
4578 /* Do the failure injections that we need to affect conn processing in the following function.
4579 * When writing tests that use this function make sure to use a fail_loc with a fail mask.
4580 * If you dont you can cause the scheduler threads to spin on the conn without it leaving
4583 * intent is used to signal the calling function whether or not the conn needs to be rescheduled.
4587 kgnilnd_check_conn_fail_loc(kgn_device_t *dev, kgn_conn_t *conn, int *intent)
4591 /* short circuit out when not set */
4592 if (likely(!cfs_fail_loc)) {
4596 /* failure injection to test for stack reset clean ups */
4597 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_CLOSING)) {
4598 /* we can't rely on busy loops being nice enough to get the
4599 * stack reset triggered - it'd just spin on this conn */
4600 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
4603 GOTO(did_fail_loc, rc);
4606 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
4607 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4609 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DROP_DESTROY_EP)) {
4610 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
4613 GOTO(did_fail_loc, rc);
4617 /* CFS_FAIL_GNI_FINISH_PURG2 is used to stop a connection from fully closing. This scheduler
4618 * will spin on the CFS_FAIL_TIMEOUT until the fail_loc is cleared at which time the connection
4619 * will be closed by kgnilnd_complete_closed_conn.
4621 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG2)) {
4622 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_FINISH_PURG2, 1)) {};
4625 GOTO(did_fail_loc, rc);
4628 /* this one is a bit gross - we can't hold the mutex from process_conns
4629 * across a CFS_RACE here - it'd block the conn threads from doing an ep_bind
4630 * and moving onto finish_connect
4631 * so, we'll just set the rc - kgnilnd_process_conns will clear
4632 * found_work on a fail_loc, getting the scheduler thread to call schedule()
4633 * and effectively getting this thread to sleep */
4634 if ((conn->gnc_state == GNILND_CONN_CLOSED) && CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
4637 GOTO(did_fail_loc, rc);
4645 kgnilnd_send_conn_close(kgn_conn_t *conn)
4649 /* we are closing the conn - we will try to send the CLOSE msg
4650 * but will not wait for anything else to flush */
4652 /* send the close if not already done so or received one */
4653 if (!conn->gnc_close_sent && !conn->gnc_close_recvd) {
4654 /* set close_sent regardless of the success of the
4655 * CLOSE message. We are going to try once and then
4656 * kick him out of the sandbox */
4657 conn->gnc_close_sent = 1;
4660 /* EP might be null already if remote side initiated a new connection.
4661 * kgnilnd_finish_connect destroys existing ep_handles before wiring up the new connection,
4662 * so this check is here to make sure we dont attempt to send with a null ep_handle.
4664 if (conn->gnc_ephandle != NULL) {
4667 tx = kgnilnd_new_tx_msg(GNILND_MSG_CLOSE, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
4669 tx->tx_msg.gnm_u.completion.gncm_retval = conn->gnc_error;
4670 tx->tx_state = GNILND_TX_WAITING_COMPLETION;
4671 tx->tx_qtime = jiffies;
4673 if (tx->tx_id.txe_idx == 0) {
4674 rc = kgnilnd_set_tx_id(tx, conn);
4676 kgnilnd_tx_done(tx, rc);
4680 CDEBUG(D_NETTRACE, "sending close with errno %d\n",
4683 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CLOSE_SEND)) {
4684 kgnilnd_tx_done(tx, -EAGAIN);
4686 rc = kgnilnd_sendmsg(tx, NULL, 0, NULL, GNILND_TX_FMAQ);
4688 /* It wasnt sent and we dont care. */
4689 kgnilnd_tx_done(tx, rc);
4697 /* When changing gnc_state we need to take the kgn_peer_conn_lock */
4698 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
4699 conn->gnc_state = GNILND_CONN_CLOSED;
4700 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
4701 /* mark this conn as CLOSED now that we processed it
4702 * do after TX, so we can use CLOSING in asserts */
4706 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSED)) {
4707 /* simulate a RX CLOSE after the timeout but before
4708 * the scheduler thread gets it */
4709 conn->gnc_close_recvd = GNILND_CLOSE_INJECT2;
4710 conn->gnc_peer_error = -ETIMEDOUT;
4712 /* schedule to allow potential CLOSE and get the complete phase run */
4713 kgnilnd_schedule_conn(conn);
4717 kgnilnd_process_mapped_tx(kgn_device_t *dev)
4722 int fast_remaps = GNILND_FAST_MAPPING_TRY;
4723 int log_retrans, log_retrans_level;
4724 static int last_map_version;
4727 spin_lock(&dev->gnd_lock);
4728 if (list_empty(&dev->gnd_map_tx)) {
4729 /* if the list is empty make sure we dont have a timer running */
4730 del_singleshot_timer_sync(&dev->gnd_map_timer);
4731 spin_unlock(&dev->gnd_lock);
4735 dev->gnd_sched_alive = jiffies;
4737 /* we'll retry as fast as possible up to 25% of the limit, then we start
4738 * backing off until our map version changes - indicating we unmapped
4740 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4741 if (likely(dev->gnd_map_attempt == 0) ||
4742 time_after_eq(jiffies, dev->gnd_next_map) ||
4743 last_map_version != dev->gnd_map_version) {
4745 /* if this is our first attempt at mapping set last mapped to current
4746 * jiffies so we can timeout our attempt correctly.
4748 if (dev->gnd_map_attempt == 0)
4749 dev->gnd_last_map = jiffies;
4751 GNIDBG_TX(D_NET, tx, "waiting for mapping event event to retry", NULL);
4752 spin_unlock(&dev->gnd_lock);
4756 /* delete the previous timer if it exists */
4757 del_singleshot_timer_sync(&dev->gnd_map_timer);
4758 /* stash the last map version to let us know when a good one was seen */
4759 last_map_version = dev->gnd_map_version;
4761 /* we need to to take the lock and continually refresh the head of the list as
4762 * kgnilnd_complete_closed_conn might be nuking stuff and we are cycling the lock
4763 * allowing them to squeeze in */
4765 while (!list_empty(&dev->gnd_map_tx)) {
4766 /* make sure we break out early on quiesce */
4767 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4768 /* always break with lock held - we unlock outside loop */
4772 tx = list_first_entry(&dev->gnd_map_tx, kgn_tx_t, tx_list);
4774 kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);
4777 /* sample with lock held, serializing with kgnilnd_complete_closed_conn */
4778 if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED) {
4779 /* if conn is dying, mark tx in tx_ref_table for
4780 * kgnilnd_complete_closed_conn to finish up */
4781 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_DYING, 1);
4784 /* tx was moved to DYING, get next */
4788 spin_unlock(&dev->gnd_lock);
4789 rc = kgnilnd_send_mapped_tx(tx, 1);
4791 /* We made it! skip error handling.. */
4793 /* OK to continue on +ve errors as it won't get seen until
4794 * this function is called again - we operate on a copy of the original
4795 * list and not the live list */
4796 spin_lock(&dev->gnd_lock);
4797 /* reset map attempts back to zero we successfully
4798 * mapped so we can reset our timers */
4799 dev->gnd_map_attempt = 0;
4801 } else if (rc == -EAGAIN) {
4802 spin_lock(&dev->gnd_lock);
4803 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4804 spin_unlock(&dev->gnd_lock);
4805 GOTO(get_out_mapped, rc);
4806 } else if (rc != -ENOMEM) {
4807 /* carp, failure we can't handle */
4808 kgnilnd_tx_done(tx, rc);
4809 spin_lock(&dev->gnd_lock);
4810 /* reset map attempts back to zero we dont know what happened but it
4811 * wasnt a failed mapping
4813 dev->gnd_map_attempt = 0;
4817 /* time to handle the retry cases.. lock so we dont have 2 threads
4818 * mucking with gnd_map_attempt, or gnd_next_map at the same time.
4820 spin_lock(&dev->gnd_lock);
4821 dev->gnd_map_attempt++;
4822 if (dev->gnd_map_attempt < fast_remaps) {
4823 /* do nothing we just want it to go as fast as possible.
4824 * just set gnd_next_map to current jiffies so it will process
4825 * as fast as possible.
4827 dev->gnd_next_map = jiffies;
4829 /* Retry based on GNILND_MAP_RETRY_RATE */
4830 dev->gnd_next_map = jiffies + GNILND_MAP_RETRY_RATE;
4833 /* only log occasionally once we've retried fast_remaps */
4834 log_retrans = (dev->gnd_map_attempt >= fast_remaps) &&
4835 ((dev->gnd_map_attempt % fast_remaps) == 0);
4836 log_retrans_level = log_retrans ? D_NETERROR : D_NET;
4838 /* make sure we are not off in the weeds with this tx */
4839 if (time_after(jiffies, dev->gnd_last_map + GNILND_MAP_TIMEOUT)) {
4840 GNIDBG_TX(D_NETERROR, tx,
4841 "giving up on TX, too many retries", NULL);
4842 spin_unlock(&dev->gnd_lock);
4843 if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ ||
4844 tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ_REV) {
4845 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4847 tx->tx_putinfo.gnpam_dst_cookie,
4848 tx->tx_msg.gnm_srcnid);
4850 kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type,
4852 tx->tx_getinfo.gngm_cookie,
4853 tx->tx_msg.gnm_srcnid);
4855 kgnilnd_tx_done(tx, -ENOMEM);
4856 GOTO(get_out_mapped, rc);
4858 GNIDBG_TX(log_retrans_level, tx,
4859 "transient map failure #%d %d pages/%d bytes phys %u@%u "
4861 "nq_map %d mdd# %d/%d GART %ld",
4862 dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
4863 dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,
4864 dev->gnd_map_nvirt, dev->gnd_map_virtnob,
4865 atomic_read(&dev->gnd_nq_map),
4866 atomic_read(&dev->gnd_n_mdd), atomic_read(&dev->gnd_n_mdd_held),
4867 atomic64_read(&dev->gnd_nbytes_map));
4870 /* we need to stop processing the rest of the list, so add it back in */
4871 /* set timer to wake device when we need to schedule this tx */
4872 mod_timer(&dev->gnd_map_timer, dev->gnd_next_map);
4873 kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
4874 spin_unlock(&dev->gnd_lock);
4875 GOTO(get_out_mapped, rc);
4877 spin_unlock(&dev->gnd_lock);
4883 kgnilnd_process_conns(kgn_device_t *dev, unsigned long deadline)
4888 int error_inject = 0;
4892 spin_lock(&dev->gnd_lock);
4893 while (!list_empty(&dev->gnd_ready_conns) && time_before(jiffies, deadline)) {
4894 dev->gnd_sched_alive = jiffies;
4898 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
4899 /* break with lock held */
4903 conn = list_first_entry(&dev->gnd_ready_conns, kgn_conn_t, gnc_schedlist);
4904 list_del_init(&conn->gnc_schedlist);
4906 * Since we are processing conn now, we don't need to be on the delaylist any longer.
4909 if (!list_empty(&conn->gnc_delaylist))
4910 list_del_init(&conn->gnc_delaylist);
4911 spin_unlock(&dev->gnd_lock);
4913 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
4915 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
4916 conn_sched != GNILND_CONN_PROCESS,
4917 "conn %p on ready list but in bad state: %d\n",
4920 CDEBUG(D_INFO, "conn %p@%s for processing\n",
4921 conn, kgnilnd_conn_state2str(conn));
4924 set_mb(conn->gnc_last_sched_do, jiffies);
4926 if (kgnilnd_check_conn_fail_loc(dev, conn, &intent)) {
4928 /* based on intent see if we should run again. */
4929 rc = kgnilnd_schedule_process_conn(conn, intent);
4931 /* drop ref from gnd_ready_conns */
4932 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4933 down_write(&dev->gnd_conn_sem);
4934 kgnilnd_conn_decref(conn);
4935 up_write(&dev->gnd_conn_sem);
4936 } else if (rc != 1) {
4937 kgnilnd_conn_decref(conn);
4939 /* clear this so that scheduler thread doesn't spin */
4941 /* break with lock held... */
4942 spin_lock(&dev->gnd_lock);
4946 if (unlikely(conn->gnc_state == GNILND_CONN_CLOSED)) {
4947 down_write(&dev->gnd_conn_sem);
4949 /* CONN_CLOSED set in procces_fmaq when CLOSE is sent */
4950 if (unlikely(atomic_read(&conn->gnc_tx_in_use))) {
4951 /* If there are tx's currently in use in another
4952 * thread we dont want to complete the close
4953 * yet. Cycle this conn back through
4955 kgnilnd_schedule_conn(conn);
4957 kgnilnd_complete_closed_conn(conn);
4959 up_write(&dev->gnd_conn_sem);
4960 } else if (unlikely(conn->gnc_state == GNILND_CONN_DESTROY_EP)) {
4961 /* DESTROY_EP set in kgnilnd_conn_decref on gnc_refcount = 1 */
4962 /* serialize SMSG CQs with ep_bind and smsg_release */
4963 down_write(&dev->gnd_conn_sem);
4964 kgnilnd_destroy_conn_ep(conn);
4965 up_write(&dev->gnd_conn_sem);
4966 } else if (unlikely(conn->gnc_state == GNILND_CONN_CLOSING)) {
4967 /* if we need to do some CLOSE sending, etc done here do it */
4968 down_write(&dev->gnd_conn_sem);
4969 kgnilnd_send_conn_close(conn);
4970 kgnilnd_check_fma_rx(conn);
4971 up_write(&dev->gnd_conn_sem);
4972 } else if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) == 0) {
4973 /* start moving traffic if the old conns are cleared out */
4974 down_read(&dev->gnd_conn_sem);
4975 kgnilnd_check_fma_rx(conn);
4976 kgnilnd_process_fmaq(conn);
4977 up_read(&dev->gnd_conn_sem);
4980 rc = kgnilnd_schedule_process_conn(conn, 0);
4982 /* drop ref from gnd_ready_conns */
4983 if (atomic_read(&conn->gnc_refcount) == 1 && rc != 1) {
4984 down_write(&dev->gnd_conn_sem);
4985 kgnilnd_conn_decref(conn);
4986 up_write(&dev->gnd_conn_sem);
4987 } else if (rc != 1) {
4988 kgnilnd_conn_decref(conn);
4991 /* check list again with lock held */
4992 spin_lock(&dev->gnd_lock);
4995 /* If we are short circuiting due to timing we want to be scheduled
4996 * as soon as possible.
4998 if (!list_empty(&dev->gnd_ready_conns) && !error_inject)
5001 spin_unlock(&dev->gnd_lock);
5007 kgnilnd_scheduler(void *arg)
5009 int threadno = (long)arg;
5012 unsigned long deadline = 0;
5015 dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
5017 /* all gnilnd threads need to run fairly urgently */
5018 set_user_nice(current, *kgnilnd_tunables.kgn_sched_nice);
5019 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5020 while (!kgnilnd_data.kgn_shutdown) {
5022 /* Safe: kgn_shutdown only set when quiescent */
5024 /* to quiesce or to not quiesce, that is the question */
5026 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
5027 KGNILND_SPIN_QUIESCE;
5030 /* tracking for when thread goes AWOL */
5031 dev->gnd_sched_alive = jiffies;
5033 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_SCHED_DEADLINE,
5034 (*kgnilnd_tunables.kgn_sched_timeout + 1));
5035 /* let folks know we are up and kicking
5036 * - they can use this for latency savings, etc
5037 * - only change if IRQ, if IDLE leave alone as that
5038 * schedule_device calls to put us back to IRQ */
5039 (void)cmpxchg(&dev->gnd_ready, GNILND_DEV_IRQ, GNILND_DEV_LOOP);
5041 down_read(&dev->gnd_conn_sem);
5042 /* always check these - they are super low cost */
5043 found_work += kgnilnd_check_fma_send_cq(dev);
5044 found_work += kgnilnd_check_fma_rcv_cq(dev);
5046 /* rdma CQ doesn't care about eps */
5047 found_work += kgnilnd_check_rdma_cq(dev);
5049 /* move some RDMA ? */
5050 found_work += kgnilnd_process_rdmaq(dev);
5052 /* map some pending RDMA requests ? */
5053 found_work += kgnilnd_process_mapped_tx(dev);
5055 /* the EP for a conn is not destroyed until all the references
5056 * to it are gone, so these checks should be safe
5057 * even if run in parallel with the CQ checking functions
5058 * _AND_ a thread that processes the CLOSED->DONE
5062 up_read(&dev->gnd_conn_sem);
5064 /* process all conns ready now */
5065 found_work += kgnilnd_process_conns(dev, deadline);
5067 /* do an eager check to avoid the IRQ disabling in
5068 * prepare_to_wait and friends */
5071 (busy_loops++ < *kgnilnd_tunables.kgn_loops) &&
5072 time_before(jiffies, deadline)) {
5074 if ((busy_loops % 10) == 0) {
5075 /* tickle heartbeat and watchdog to ensure our
5076 * piggishness doesn't turn into heartbeat failure */
5077 touch_nmi_watchdog();
5083 /* if we got here, found_work was zero or busy_loops means we
5084 * need to take a break. We'll clear gnd_ready but we'll check
5085 * one last time if there is an IRQ that needs processing */
5087 prepare_to_wait(&dev->gnd_waitq, &wait, TASK_INTERRUPTIBLE);
5089 /* the first time this will go LOOP -> IDLE and let us do one final check
5090 * during which we might get an IRQ, then IDLE->IDLE and schedule()
5091 * - this might allow other threads to block us for a bit if they
5092 * try to get the mutex, but that is good as we'd need to wake
5093 * up soon to handle the CQ or other processing anyways */
5095 found_work += xchg(&dev->gnd_ready, GNILND_DEV_IDLE);
5097 if ((busy_loops >= *kgnilnd_tunables.kgn_loops) ||
5098 time_after_eq(jiffies, deadline)) {
5100 "yeilding: found_work %d busy_loops %d\n",
5101 found_work, busy_loops);
5103 /* use yield if we are bailing due to busy_loops
5104 * - this will ensure we wake up soonish. This closes
5105 * a race with kgnilnd_device_callback - where it'd
5106 * not call wake_up() because gnd_ready == 1, but then
5107 * we come down and schedule() because of busy_loops.
5108 * We'd not be woken up until something poked our waitq
5109 * again. yield() ensures we wake up without another
5110 * waitq poke in that case */
5111 atomic_inc(&dev->gnd_n_yield);
5112 kgnilnd_data.kgn_last_condresched = jiffies;
5114 CDEBUG(D_INFO, "awake after yeild\n");
5115 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5116 } else if (found_work == GNILND_DEV_IDLE) {
5117 /* busy_loops is low and there is nothing to do,
5118 * go to sleep and wait for a waitq poke */
5120 "scheduling: found_work %d busy_loops %d\n",
5121 found_work, busy_loops);
5122 atomic_inc(&dev->gnd_n_schedule);
5123 kgnilnd_data.kgn_last_scheduled = jiffies;
5125 CDEBUG(D_INFO, "awake after schedule\n");
5126 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_sched_timeout);
5128 finish_wait(&dev->gnd_waitq, &wait);
5131 kgnilnd_thread_fini();