2 * Copyright (C) 2012 Cray, Inc.
4 * Copyright (c) 2014, Intel Corporation.
6 * Author: Nic Henke <nic@cray.com>
7 * Author: James Shimek <jshimek@cray.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 kgnilnd_setup_smsg_attr(gni_smsg_attr_t *smsg_attr)
31 smsg_attr->mbox_maxcredit = *kgnilnd_tunables.kgn_mbox_credits;
32 smsg_attr->msg_maxsize = GNILND_MAX_MSG_SIZE;
33 smsg_attr->msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
37 kgnilnd_map_fmablk(kgn_device_t *device, kgn_fma_memblock_t *fma_blk)
40 __u32 flags = GNI_MEM_READWRITE;
42 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
43 flags |= GNI_MEM_PHYS_CONT;
46 fma_blk->gnm_hold_timeout = 0;
48 /* make sure we are mapping a clean block */
49 LASSERTF(fma_blk->gnm_hndl.qword1 == 0UL, "fma_blk %p dirty\n", fma_blk);
51 rrc = kgnilnd_mem_register(device->gnd_handle, (__u64)fma_blk->gnm_block,
52 fma_blk->gnm_blk_size, device->gnd_rcv_fma_cqh,
53 flags, &fma_blk->gnm_hndl);
54 if (rrc != GNI_RC_SUCCESS) {
55 /* XXX Nic: need a way to silence this for runtime stuff that is ok to fail
56 * -- like when under MDD or GART pressure on big systems
58 CNETERR("register fmablk failed 0x%p mbox_size %d flags %u\n",
59 fma_blk, fma_blk->gnm_mbox_size, flags);
63 /* PHYS_CONT memory isn't really mapped, at least not in GART -
64 * but all mappings chew up a MDD
66 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
67 atomic64_add(fma_blk->gnm_blk_size, &device->gnd_nbytes_map);
70 atomic_inc(&device->gnd_n_mdd);
71 /* nfmablk is live (mapped) blocks */
72 atomic_inc(&device->gnd_nfmablk);
78 kgnilnd_alloc_fmablk(kgn_device_t *device, int use_phys)
82 kgn_fma_memblock_t *fma_blk;
83 gni_smsg_attr_t smsg_attr;
84 unsigned long fmablk_vers;
86 #if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
87 /* We allocate large blocks of memory here potentially leading
88 * to memory exhaustion during massive reconnects during a network
89 * outage. Limit the amount of fma blocks to use by always keeping
90 * a percent of pages free initially set to 25% of total memory. */
91 if (global_page_state(NR_FREE_PAGES) < kgnilnd_data.free_pages_limit) {
92 LCONSOLE_INFO("Exceeding free page limit of %ld. "
93 "Free pages available %ld\n",
94 kgnilnd_data.free_pages_limit,
95 global_page_state(NR_FREE_PAGES));
99 /* we'll use fmablk_vers and the gnd_fmablk_mutex to gate access
100 * to this allocation code. Everyone will sample the version
101 * before and after getting the mutex. If it has changed,
102 * we'll bail out to check the lists again - this indicates that
103 * some sort of change was made to the lists and it is possible
104 * that there is a mailbox for us to find now. This should prevent
105 * a ton of spinning in the case where there are lots of threads
106 * that need a yet-to-be-allocated mailbox for a connection. */
108 fmablk_vers = atomic_read(&device->gnd_fmablk_vers);
109 mutex_lock(&device->gnd_fmablk_mutex);
111 if (fmablk_vers != atomic_read(&device->gnd_fmablk_vers)) {
112 /* version changed while we were waiting for semaphore,
113 * we'll recheck the lists assuming something nice happened */
114 mutex_unlock(&device->gnd_fmablk_mutex);
118 LIBCFS_ALLOC(fma_blk, sizeof(kgn_fma_memblock_t));
119 if (fma_blk == NULL) {
120 CNETERR("could not allocate fma block descriptor\n");
125 INIT_LIST_HEAD(&fma_blk->gnm_bufflist);
127 kgnilnd_setup_smsg_attr(&smsg_attr);
129 gni_smsg_buff_size_needed(&smsg_attr, &fma_blk->gnm_mbox_size);
131 LASSERTF(fma_blk->gnm_mbox_size, "mbox size %d\n", fma_blk->gnm_mbox_size);
133 /* gni_smsg_buff_size_needed calculates the base mailbox size and since
134 * we want to hold kgn_peer_credits worth of messages in both directions,
135 * we add PAYLOAD to grow the mailbox size
138 fma_blk->gnm_mbox_size += GNILND_MBOX_PAYLOAD;
140 /* we'll only use physical during preallocate at startup -- this keeps it nice and
141 * clean for runtime decisions. We'll keep the PHYS ones around until shutdown
142 * as reallocating them is tough if there is memory fragmentation */
145 fma_blk->gnm_block = kmem_cache_alloc(kgnilnd_data.kgn_mbox_cache, GFP_ATOMIC);
146 if (fma_blk->gnm_block == NULL) {
147 CNETERR("could not allocate physical SMSG mailbox memory\n");
151 fma_blk->gnm_blk_size = KMALLOC_MAX_SIZE;
152 num_mbox = fma_blk->gnm_blk_size / fma_blk->gnm_mbox_size;
154 LASSERTF(num_mbox >= 1,
155 "num_mbox %d blk_size %u mbox_size %d\n",
156 num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size);
158 fma_blk->gnm_state = GNILND_FMABLK_PHYS;
161 num_mbox = *kgnilnd_tunables.kgn_mbox_per_block;
162 fma_blk->gnm_blk_size = num_mbox * fma_blk->gnm_mbox_size;
164 LASSERTF(num_mbox >= 1 && num_mbox >= *kgnilnd_tunables.kgn_mbox_per_block,
165 "num_mbox %d blk_size %u mbox_size %d tunable %d\n",
166 num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size,
167 *kgnilnd_tunables.kgn_mbox_per_block);
169 fma_blk->gnm_block = kgnilnd_vzalloc(fma_blk->gnm_blk_size);
170 if (fma_blk->gnm_block == NULL) {
171 CNETERR("could not allocate virtual SMSG mailbox memory, %d bytes\n", fma_blk->gnm_blk_size);
176 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
179 /* allocate just enough space for the bits to track the mailboxes */
180 LIBCFS_ALLOC(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox) * sizeof(unsigned long));
181 if (fma_blk->gnm_bit_array == NULL) {
182 CNETERR("could not allocate mailbox bitmask, %lu bytes for %d mbox\n",
183 sizeof(unsigned long) * BITS_TO_LONGS(num_mbox), num_mbox);
187 bitmap_zero(fma_blk->gnm_bit_array, num_mbox);
189 /* now that the num_mbox is set based on allocation type, get debug info setup */
190 LIBCFS_ALLOC(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t) * num_mbox);
191 if (fma_blk->gnm_mbox_info == NULL) {
192 CNETERR("could not allocate mailbox debug, %lu bytes for %d mbox\n",
193 sizeof(kgn_mbox_info_t) * num_mbox, num_mbox);
198 rc = kgnilnd_map_fmablk(device, fma_blk);
203 fma_blk->gnm_next_avail_mbox = 0;
204 fma_blk->gnm_avail_mboxs = fma_blk->gnm_num_mboxs = num_mbox;
206 CDEBUG(D_MALLOC, "alloc fmablk 0x%p num %d msg_maxsize %d credits %d "
207 "mbox_size %d MDD "LPX64"."LPX64"\n",
208 fma_blk, num_mbox, smsg_attr.msg_maxsize, smsg_attr.mbox_maxcredit,
209 fma_blk->gnm_mbox_size, fma_blk->gnm_hndl.qword1,
210 fma_blk->gnm_hndl.qword2);
212 /* lock Is protecting data structures, not semaphore */
214 spin_lock(&device->gnd_fmablk_lock);
215 list_add_tail(&fma_blk->gnm_bufflist, &device->gnd_fma_buffs);
217 /* toggle under the lock so once they change the list is also
218 * ready for others to traverse */
219 atomic_inc(&device->gnd_fmablk_vers);
221 spin_unlock(&device->gnd_fmablk_lock);
223 mutex_unlock(&device->gnd_fmablk_mutex);
228 LIBCFS_FREE(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t)*num_mbox);
230 LIBCFS_FREE(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox) * sizeof (unsigned long));
232 if (fma_blk->gnm_state == GNILND_FMABLK_VIRT) {
233 LIBCFS_FREE(fma_blk->gnm_block, fma_blk->gnm_blk_size);
235 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
238 LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
240 mutex_unlock(&device->gnd_fmablk_mutex);
245 kgnilnd_unmap_fmablk(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
249 /* if some held, set hold_timeout from conn timeouts used in this block
250 * but not during shutdown, then just nuke and pave
251 * During a stack reset, we need to deregister with a hold timeout
252 * set so we don't use the same mdd after reset is complete */
253 if ((fma_blk->gnm_held_mboxs && !kgnilnd_data.kgn_shutdown) ||
254 kgnilnd_data.kgn_in_reset) {
255 fma_blk->gnm_hold_timeout = GNILND_TIMEOUT2DEADMAN;
258 /* we are changing the state of a block, tickle version to tell
259 * proc code list is stale now */
260 atomic_inc(&dev->gnd_fmablk_vers);
262 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &fma_blk->gnm_hndl, fma_blk->gnm_hold_timeout);
264 CDEBUG(rrc == GNI_RC_SUCCESS ? D_MALLOC : D_CONSOLE|D_NETERROR,
265 "unmap fmablk 0x%p@%s sz %u total %d avail %d held %d mbox_size %d "
267 fma_blk, kgnilnd_fmablk_state2str(fma_blk->gnm_state),
268 fma_blk->gnm_blk_size, fma_blk->gnm_num_mboxs,
269 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs,
270 fma_blk->gnm_mbox_size, fma_blk->gnm_hold_timeout);
272 LASSERTF(rrc == GNI_RC_SUCCESS,
273 "tried to double unmap or something bad, fma_blk %p (rrc %d)\n",
276 if (fma_blk->gnm_hold_timeout &&
277 !(kgnilnd_data.kgn_in_reset &&
278 fma_blk->gnm_state == GNILND_FMABLK_PHYS)) {
279 atomic_inc(&dev->gnd_n_mdd_held);
281 atomic_dec(&dev->gnd_n_mdd);
284 /* PHYS blocks don't get mapped */
285 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
286 atomic64_sub(fma_blk->gnm_blk_size, &dev->gnd_nbytes_map);
287 fma_blk->gnm_state = GNILND_FMABLK_IDLE;
288 } else if (kgnilnd_data.kgn_in_reset) {
289 /* in stack reset, clear MDD handle for PHYS blocks, as we'll
290 * re-use the fma_blk after reset so we don't have to drop/allocate
291 * all of those physical blocks */
292 fma_blk->gnm_hndl.qword1 = fma_blk->gnm_hndl.qword2 = 0UL;
295 /* Decrement here as this is the # of mapped blocks */
296 atomic_dec(&dev->gnd_nfmablk);
300 /* needs lock on gnd_fmablk_lock to cover gnd_fma_buffs */
302 kgnilnd_free_fmablk_locked(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
304 LASSERTF(fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs,
305 "fma_blk %p@%d free in bad state (%d): blk total %d avail %d held %d\n",
306 fma_blk, fma_blk->gnm_state, fma_blk->gnm_hold_timeout, fma_blk->gnm_num_mboxs,
307 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs);
309 atomic_inc(&dev->gnd_fmablk_vers);
311 if (fma_blk->gnm_hold_timeout) {
312 CDEBUG(D_MALLOC, "mdd release fmablk 0x%p sz %u avail %d held %d "
314 fma_blk, fma_blk->gnm_blk_size, fma_blk->gnm_avail_mboxs,
315 fma_blk->gnm_held_mboxs, fma_blk->gnm_mbox_size);
317 /* We leave MDD dangling over stack reset */
318 if (!kgnilnd_data.kgn_in_reset) {
319 kgnilnd_mem_mdd_release(dev->gnd_handle, &fma_blk->gnm_hndl);
321 /* ignoring the return code - if kgni/ghal can't find it
322 * it must be released already */
323 atomic_dec(&dev->gnd_n_mdd_held);
324 atomic_dec(&dev->gnd_n_mdd);
327 /* we cant' free the gnm_block until all the conns have released their
328 * purgatory holds. While we have purgatory holds, we might check the conn
329 * RX mailbox during the CLOSING process. It is possible that kgni might
330 * try to look into the RX side for credits when sending the CLOSE msg too */
331 CDEBUG(D_MALLOC, "fmablk %p free buffer %p mbox_size %d\n",
332 fma_blk, fma_blk->gnm_block, fma_blk->gnm_mbox_size);
334 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
335 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
337 LIBCFS_FREE(fma_blk->gnm_block, fma_blk->gnm_blk_size);
339 fma_blk->gnm_state = GNILND_FMABLK_FREED;
341 list_del(&fma_blk->gnm_bufflist);
343 LIBCFS_FREE(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t)*fma_blk->gnm_num_mboxs);
344 LIBCFS_FREE(fma_blk->gnm_bit_array, BITS_TO_LONGS(fma_blk->gnm_num_mboxs) * sizeof (unsigned long));
345 LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
349 kgnilnd_find_free_mbox(kgn_conn_t *conn)
351 kgn_device_t *dev = conn->gnc_device;
352 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
353 kgn_fma_memblock_t *fma_blk;
354 kgn_mbox_info_t *mbox = NULL;
357 spin_lock(&dev->gnd_fmablk_lock);
359 list_for_each_entry(fma_blk, &conn->gnc_device->gnd_fma_buffs,
361 if (fma_blk->gnm_avail_mboxs <= 0 ||
362 fma_blk->gnm_state <= GNILND_FMABLK_IDLE) {
365 /* look in bitarray for available mailbox */
367 id = find_next_zero_bit(
368 fma_blk->gnm_bit_array,
369 fma_blk->gnm_num_mboxs,
370 fma_blk->gnm_next_avail_mbox);
371 if (id == fma_blk->gnm_num_mboxs &&
372 fma_blk->gnm_next_avail_mbox != 0) {
374 fma_blk->gnm_next_avail_mbox = 0;
380 LASSERTF(id < fma_blk->gnm_num_mboxs, "id %d max %d\n",
381 id, fma_blk->gnm_num_mboxs);
382 set_bit(id, (volatile unsigned long *)fma_blk->gnm_bit_array);
383 conn->gnc_mbox_id = id;
385 fma_blk->gnm_next_avail_mbox =
386 (id == (fma_blk->gnm_num_mboxs - 1)) ? 0 : (id + 1);
387 fma_blk->gnm_avail_mboxs--;
388 conn->gnc_fma_blk = fma_blk;
390 kgnilnd_setup_smsg_attr(smsg_attr);
392 smsg_attr->msg_buffer = fma_blk->gnm_block;
393 smsg_attr->mbox_offset = fma_blk->gnm_mbox_size * id;
394 smsg_attr->mem_hndl = fma_blk->gnm_hndl;
395 smsg_attr->buff_size = fma_blk->gnm_mbox_size;
397 /* We'll set the hndl to zero for PHYS blocks unmapped during stack
398 * reset and re-use the same fma_blk after stack reset. This ensures we've
399 * properly mapped it before we use it */
400 LASSERTF(fma_blk->gnm_hndl.qword1 != 0UL, "unmapped fma_blk %p, state %d\n",
401 fma_blk, fma_blk->gnm_state);
403 CDEBUG(D_NET, "conn %p smsg %p fmablk %p "
404 "allocating SMSG mbox %d buf %p "
405 "offset %u hndl "LPX64"."LPX64"\n",
406 conn, smsg_attr, fma_blk, id,
407 smsg_attr->msg_buffer, smsg_attr->mbox_offset,
408 fma_blk->gnm_hndl.qword1,
409 fma_blk->gnm_hndl.qword2);
411 mbox = &fma_blk->gnm_mbox_info[id];
412 mbox->mbx_create_conn_memset = jiffies;
414 mbox->mbx_nallocs_total++;
416 /* zero mbox to remove any old data from our last use.
417 * this better be safe, if not our purgatory timers
418 * are too short or a peer really is misbehaving */
419 memset(smsg_attr->msg_buffer + smsg_attr->mbox_offset,
420 0, smsg_attr->buff_size);
424 spin_unlock(&dev->gnd_fmablk_lock);
428 kgnilnd_setup_mbox(kgn_conn_t *conn)
430 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
433 smsg_attr->msg_buffer = NULL;
434 /* Look for available mbox */
436 kgnilnd_find_free_mbox(conn);
438 /* nothing in the existing buffers, make a new one */
439 if (smsg_attr->msg_buffer == NULL) {
440 /* for runtime allocations, we only want vmalloc */
441 err = kgnilnd_alloc_fmablk(conn->gnc_device, 0);
446 } while (smsg_attr->msg_buffer == NULL);
449 CNETERR("couldn't allocate SMSG mbox for conn %p Error: %d\n",
455 kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold)
457 kgn_device_t *dev = conn->gnc_device;
458 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
459 kgn_fma_memblock_t *fma_blk = NULL;
460 kgn_mbox_info_t *mbox = NULL;
464 /* if we failed to setup mbox and now destroying conn */
465 if (smsg_attr->msg_buffer == NULL) {
469 id = conn->gnc_mbox_id;
471 spin_lock(&dev->gnd_fmablk_lock);
472 /* make sure our conn points at a valid fma_blk
473 * We use this instead of a mem block search out of smsg_attr
474 * because we could have freed a block for fma_blk #1 but the fma_blk
475 * is still in the list for a purgatory hold. This would induce a false
476 * match if that same block gets reallocated to fma_blk #2 */
477 list_for_each_entry(fma_blk, &dev->gnd_fma_buffs, gnm_bufflist) {
478 if (fma_blk == conn->gnc_fma_blk) {
483 LASSERTF(found, "unable to find conn 0x%p with gnc_fma_blk %p "
484 "anywhere in the world\n", conn, conn->gnc_fma_blk);
486 LASSERTF(id < fma_blk->gnm_num_mboxs,
487 "bad id %d max %d\n",
488 id, fma_blk->gnm_num_mboxs);
490 /* < 0 - was held, now free it
491 * == 0 - just free it
492 * > 0 - hold it for now */
493 if (purgatory_hold == 0) {
494 CDEBUG(D_NET, "conn %p smsg %p fmablk %p freeing SMSG mbox %d "
495 "hndl "LPX64"."LPX64"\n",
496 conn, smsg_attr, fma_blk, id,
497 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
498 fma_blk->gnm_avail_mboxs++;
500 } else if (purgatory_hold > 0) {
501 CDEBUG(D_NET, "conn %p smsg %p fmablk %p holding SMSG mbox %d "
502 "hndl "LPX64"."LPX64"\n",
503 conn, smsg_attr, fma_blk, id,
504 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
506 fma_blk->gnm_held_mboxs++;
507 fma_blk->gnm_max_timeout = MAX(fma_blk->gnm_max_timeout,
510 CDEBUG(D_NET, "conn %p smsg %p fmablk %p release SMSG mbox %d "
511 "hndl "LPX64"."LPX64"\n",
512 conn, smsg_attr, fma_blk, id,
513 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
515 fma_blk->gnm_held_mboxs--;
516 fma_blk->gnm_avail_mboxs++;
519 if (purgatory_hold <= 0) {
520 /* if kgni is retransmitting, freeing the smsg block before the EP
521 * is destroyed gets messy. Bug 768295. */
522 LASSERTF(conn->gnc_ephandle == NULL,
523 "can't release mbox before EP is nuked. conn 0x%p\n", conn);
525 mbox = &fma_blk->gnm_mbox_info[id];
526 mbox->mbx_release_from_purgatory = jiffies;
528 /* clear conn gnc_fmablk if it is gone - this allows us to
529 * not worry about state so much in kgnilnd_destroy_conn
530 * and makes the guaranteed cleanup of the resources easier */
531 LASSERTF(test_and_clear_bit(id, fma_blk->gnm_bit_array),
532 "conn %p bit %d already cleared in fma_blk %p\n",
534 conn->gnc_fma_blk = NULL;
538 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FMABLK_AVAIL)) {
539 CERROR("LBUGs in your future: forcibly marking fma_blk %p "
540 "as mapped\n", fma_blk);
541 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
544 /* we don't release or unmap PHYS blocks as part of the normal cycle --
545 * those are controlled manually from startup/shutdown */
546 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
547 /* we can unmap once all are unused (held or avail)
548 * but check hold_timeout to make sure we are not trying to double
549 * unmap this buffer. If there was no hold_timeout set due to
550 * held_mboxs, we'll free the mobx here shortly and won't have to
551 * worry about catching a double free for a 'clean' fma_blk */
552 if (((fma_blk->gnm_avail_mboxs + fma_blk->gnm_held_mboxs) == fma_blk->gnm_num_mboxs) &&
553 (!fma_blk->gnm_hold_timeout)) {
554 kgnilnd_unmap_fmablk(dev, fma_blk);
557 /* But we can only free once they are all avail */
558 if (fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs &&
559 fma_blk->gnm_held_mboxs == 0) {
560 /* all mailboxes are released, free fma_blk */
561 kgnilnd_free_fmablk_locked(dev, fma_blk);
565 spin_unlock(&dev->gnd_fmablk_lock);
569 kgnilnd_count_phys_mbox(kgn_device_t *device)
572 kgn_fma_memblock_t *fma_blk;
574 spin_lock(&device->gnd_fmablk_lock);
576 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
577 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
578 i += fma_blk->gnm_num_mboxs;
580 spin_unlock(&device->gnd_fmablk_lock);
586 kgnilnd_allocate_phys_fmablk(kgn_device_t *device)
590 while (kgnilnd_count_phys_mbox(device) < *kgnilnd_tunables.kgn_nphys_mbox) {
592 rc = kgnilnd_alloc_fmablk(device, 1);
594 CERROR("failed phys mbox allocation, stopping at %d, rc %d\n",
595 kgnilnd_count_phys_mbox(device), rc);
603 kgnilnd_map_phys_fmablk(kgn_device_t *device)
607 kgn_fma_memblock_t *fma_blk;
609 /* use mutex to gate access to single thread, just in case */
610 mutex_lock(&device->gnd_fmablk_mutex);
612 spin_lock(&device->gnd_fmablk_lock);
614 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
615 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
616 rc = kgnilnd_map_fmablk(device, fma_blk);
621 spin_unlock(&device->gnd_fmablk_lock);
623 mutex_unlock(&device->gnd_fmablk_mutex);
629 kgnilnd_unmap_fma_blocks(kgn_device_t *device)
632 kgn_fma_memblock_t *fma_blk;
634 /* use mutex to gate access to single thread, just in case */
635 mutex_lock(&device->gnd_fmablk_mutex);
637 spin_lock(&device->gnd_fmablk_lock);
639 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
640 kgnilnd_unmap_fmablk(device, fma_blk);
642 spin_unlock(&device->gnd_fmablk_lock);
644 mutex_unlock(&device->gnd_fmablk_mutex);
648 kgnilnd_free_phys_fmablk(kgn_device_t *device)
651 kgn_fma_memblock_t *fma_blk, *fma_blkN;
653 /* use mutex to gate access to single thread, just in case */
654 mutex_lock(&device->gnd_fmablk_mutex);
656 spin_lock(&device->gnd_fmablk_lock);
658 list_for_each_entry_safe(fma_blk, fma_blkN, &device->gnd_fma_buffs, gnm_bufflist) {
659 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
660 kgnilnd_free_fmablk_locked(device, fma_blk);
662 spin_unlock(&device->gnd_fmablk_lock);
664 mutex_unlock(&device->gnd_fmablk_mutex);
667 /* kgnilnd dgram nid->struct managment */
669 static inline struct list_head *
670 kgnilnd_nid2dgramlist(kgn_device_t *dev, lnet_nid_t nid)
672 unsigned int hash = ((unsigned int)nid) % *kgnilnd_tunables.kgn_peer_hash_size;
674 RETURN(&dev->gnd_dgrams[hash]);
678 /* needs dev->gnd_dgram_lock held */
680 kgnilnd_find_dgram_locked(kgn_device_t *dev, lnet_nid_t dst_nid)
682 struct list_head *dgram_list = kgnilnd_nid2dgramlist(dev, dst_nid);
685 list_for_each_entry(dgram, dgram_list, gndg_list) {
687 /* if state > POSTED, we are already handling cancel/completion */
688 if ((dgram->gndg_conn_out.gncr_dstnid != dst_nid) ||
689 dgram->gndg_state > GNILND_DGRAM_POSTED)
692 CDEBUG(D_NET, "got dgram [%p] -> %s\n",
693 dgram, libcfs_nid2str(dst_nid));
700 kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid)
704 spin_lock(&dev->gnd_dgram_lock);
705 dgram = kgnilnd_find_dgram_locked(dev, dst_nid);
708 kgnilnd_cancel_dgram_locked(dgram);
710 spin_unlock(&dev->gnd_dgram_lock);
712 RETURN(!!(dgram == NULL));
716 kgnilnd_pack_connreq(kgn_connreq_t *connreq, kgn_conn_t *conn,
717 lnet_nid_t srcnid, lnet_nid_t dstnid,
718 kgn_connreq_type_t type)
722 /* ensure we haven't violated max datagram size */
723 CLASSERT(sizeof(kgn_connreq_t) <= GNI_DATAGRAM_MAXSIZE);
725 /* no need to zero out, we do that when allocating dgram */
726 connreq->gncr_magic = GNILND_MSG_MAGIC;
728 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_SRCNID)) {
730 } else if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
734 connreq->gncr_srcnid = srcnid;
735 connreq->gncr_dstnid = dstnid;
737 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
738 connreq->gncr_version = 99;
740 connreq->gncr_version = GNILND_CONNREQ_VERSION;
742 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
743 connreq->gncr_type = 99;
745 connreq->gncr_type = type;
747 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
748 connreq->gncr_peerstamp = 0;
750 connreq->gncr_peerstamp = kgnilnd_data.kgn_peerstamp;
752 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
753 connreq->gncr_connstamp = 0;
755 connreq->gncr_connstamp = conn->gnc_my_connstamp;
757 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
758 connreq->gncr_timeout = 0;
760 connreq->gncr_timeout = conn->gnc_timeout;
763 /* the rest pack the data into the payload in other places */
764 if (type == GNILND_CONNREQ_REQ) {
765 kgn_gniparams_t *req_params = &connreq->gncr_gnparams;
766 req_params->gnpr_host_id = conn->gnc_device->gnd_host_id;
767 req_params->gnpr_cqid = conn->gnc_cqid;
769 /* allocate mailbox for this connection */
770 err = kgnilnd_setup_mbox(conn);
772 CERROR("Failed to setup FMA mailbox (%d)\n", err);
774 req_params->gnpr_smsg_attr = conn->gnpr_smsg_attr;
777 /* XXX Nic: TBD - checksum computation */
783 kgnilnd_unpack_connreq(kgn_dgram_t *dgram)
785 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
789 /* the following fields must be handled in a backwards compatible
790 * manner to ensure we can always send and interpret NAKs */
792 if (connreq->gncr_magic != GNILND_MSG_MAGIC &&
793 connreq->gncr_magic != __swab32(GNILND_MSG_MAGIC)) {
794 /* Unexpected magic! */
795 CERROR("Unexpected magic %08x\n",
796 connreq->gncr_magic);
800 swab = (connreq->gncr_magic == __swab32(GNILND_MSG_MAGIC));
802 __swab32s(&connreq->gncr_magic);
803 __swab32s(&connreq->gncr_cksum);
804 __swab16s(&connreq->gncr_type);
805 __swab16s(&connreq->gncr_version);
806 __swab32s(&connreq->gncr_timeout);
807 __swab64s(&connreq->gncr_srcnid);
808 __swab64s(&connreq->gncr_dstnid);
809 __swab64s(&connreq->gncr_peerstamp);
810 __swab64s(&connreq->gncr_connstamp);
813 /* Do NOT return anything but -EBADF before we munge
814 * connreq->gncr_srcnid - we need that to send the nak */
816 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
817 lnet_nid_t incoming = connreq->gncr_srcnid;
819 /* even if the incoming packet is hosed, we know who we sent
820 * the original and can set the srcnid so that we can properly
821 * look up our peer to close the loop on this connreq. We still use
822 * -EBADF to prevent a NAK - just in case there are issues with
823 * the payload coming from a random spot, etc. */
824 connreq->gncr_srcnid = dgram->gndg_conn_out.gncr_dstnid;
826 if (LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid) !=
827 LNET_NIDADDR(incoming)) {
828 /* we got a datagram match for the wrong nid... */
829 CERROR("matched datagram 0x%p with srcnid %s "
830 "(%x), expecting %s (%x)\n",
832 libcfs_nid2str(incoming),
833 LNET_NIDADDR(incoming),
834 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
835 LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid));
839 /* if we have a wildcard datagram it should match an
840 * incoming "active" datagram that should have a fully formed
841 * srcnid and dstnid. If we couldn't unpack it, we drop as
842 * corrupted packet, otherwise we'll just verify that the dstnid
843 * matches the NID for the NET that the dgram was posted */
845 /* make sure their wildcard didn't match ours, that is unpossible */
846 LASSERTF(connreq->gncr_dstnid != LNET_NID_ANY,
847 "dgram 0x%p from %s, connreq 0x%p; "
848 "wildcard matched wildcard \n", dgram,
849 libcfs_nid2str(connreq->gncr_srcnid), connreq);
851 rc = kgnilnd_find_net(connreq->gncr_dstnid, &net);
853 if (rc == -ESHUTDOWN) {
854 CERROR("Looking up network: device is in shutdown");
856 } else if (rc == -ENONET) {
857 CERROR("Connection data from %s: she sent "
858 "dst_nid %s, but net lookup failed on "
860 libcfs_nid2str(connreq->gncr_srcnid),
861 libcfs_nid2str(connreq->gncr_dstnid),
862 dgram, kgnilnd_dgram_type2str(dgram));
866 if (net->gnn_ni->ni_nid != connreq->gncr_dstnid) {
867 CERROR("Bad connection data from %s: she sent "
868 "dst_nid %s, but I am %s with dgram 0x%p@%s\n",
869 libcfs_nid2str(connreq->gncr_srcnid),
870 libcfs_nid2str(connreq->gncr_dstnid),
871 libcfs_nid2str(net->gnn_ni->ni_nid),
872 dgram, kgnilnd_dgram_type2str(dgram));
873 kgnilnd_net_decref(net);
877 /* kgnilnd_find_net takes a ref on the net it finds, You need to decref it when not needed. */
878 kgnilnd_net_decref(net);
881 if (connreq->gncr_version != GNILND_CONNREQ_VERSION) {
882 CERROR("Unexpected version %d\n", connreq->gncr_version);
886 /* XXX Nic: TBD - checksum validation */
887 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_DROP)) {
891 if (swab && connreq->gncr_type == GNILND_CONNREQ_REQ) {
892 __u64 msg_addr = (__u64) connreq->gncr_gnparams.gnpr_smsg_attr.msg_buffer;
894 __swab32s(&connreq->gncr_gnparams.gnpr_host_id);
895 __swab32s(&connreq->gncr_gnparams.gnpr_cqid);
896 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.buff_size);
897 __swab16s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_maxcredit);
898 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_offset);
899 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword1);
900 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword2);
901 __swab64s(&msg_addr);
902 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_maxsize);
903 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_type);
904 } else if (swab && connreq->gncr_type == GNILND_CONNREQ_NAK) {
905 __swab32s(&connreq->gncr_nakdata.gnnd_errno);
908 /* since we use a unique instance ID for each network, the driver
909 * will take care of dropping datagrams if we don't have that network.
912 /* few more idiot software or configuration checks */
914 switch (connreq->gncr_type) {
915 case GNILND_CONNREQ_REQ:
916 /* wire up EP and SMSG block - this will check the incoming data
917 * and barf a NAK back if need to */
918 rc = kgnilnd_set_conn_params(dgram);
922 case GNILND_CONNREQ_NAK:
923 case GNILND_CONNREQ_CLOSE:
926 CERROR("unknown connreq packet type %d\n", connreq->gncr_type);
930 if (connreq->gncr_peerstamp == 0 || connreq->gncr_connstamp == 0) {
931 CERROR("Recived bad timestamps peer "LPU64" conn "LPU64"\n",
932 connreq->gncr_peerstamp, connreq->gncr_connstamp);
936 if (connreq->gncr_timeout < GNILND_MIN_TIMEOUT) {
937 CERROR("Received timeout %d < MIN %d\n",
938 connreq->gncr_timeout, GNILND_MIN_TIMEOUT);
946 kgnilnd_alloc_dgram(kgn_dgram_t **dgramp, kgn_device_t *dev, kgn_dgram_type_t type)
950 dgram = kmem_cache_alloc(kgnilnd_data.kgn_dgram_cache, GFP_ATOMIC);
954 /* cache alloc'd memory is not zeroed */
955 memset((void *)dgram, 0, sizeof(*dgram)) ;
957 INIT_LIST_HEAD(&dgram->gndg_list);
958 dgram->gndg_state = GNILND_DGRAM_USED;
959 dgram->gndg_type = type;
960 dgram->gndg_magic = GNILND_DGRAM_MAGIC;
962 atomic_inc(&dev->gnd_ndgrams);
964 CDEBUG(D_MALLOC|D_NETTRACE, "slab-alloced 'dgram': %lu at %p %s ndgrams"
966 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
967 atomic_read(&dev->gnd_ndgrams));
973 /* call this on a dgram that came back from kgnilnd_ep_postdata_test_by_id
974 * returns < 0 on dgram to be cleaned up
975 * > 0 on dgram that isn't done yet
976 * == 0 on dgram that is ok and needs connreq processing */
978 kgnilnd_process_dgram(kgn_dgram_t *dgram, gni_post_state_t post_state)
982 switch (post_state) {
983 case GNI_POST_COMPLETED:
984 /* normal state for dgrams that need actual processing */
985 /* GOTO to avoid processing dgram as canceled/done */
986 GOTO(process_out, rc);
988 case GNI_POST_PENDING:
989 /* we should only see this if we are testing a WC dgram after a
990 * cancel - it means that it needs a full cycle of waiting
991 * for kgni_sm_task to finish moving it to TERMINATED */
992 LASSERTF((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
993 (dgram->gndg_state == GNILND_DGRAM_CANCELED),
994 "POST_PENDING dgram 0x%p with bad type %d(%s) or state %d(%s)\n",
995 dgram, dgram->gndg_type, kgnilnd_dgram_type2str(dgram),
996 dgram->gndg_state, kgnilnd_dgram_state2str(dgram));
998 /* positive RC as this dgram isn't done yet */
1001 /* GOTO as this isn't done yet */
1002 GOTO(process_out, rc);
1005 case GNI_POST_TERMINATED:
1006 /* we've called cancel and it is done or remote guy called cancel and
1007 * we've receved it on a WC dgram */
1009 /* we are seeing weird terminations on non WC dgrams when we have not
1012 LASSERTF(dgram->gndg_state == GNILND_DGRAM_CANCELED ||
1013 dgram->gndg_conn_out.gncr_dstnid == LNET_NID_ANY,
1014 "dgram 0x%p with bad state %d(%s) or dst nid %s\n",
1015 dgram, dgram->gndg_state, kgnilnd_dgram_state2str(dgram),
1016 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
1019 CDEBUG(D_NETTRACE, "dgram 0x%p saw %s, cleaning it up\n", dgram,
1020 dgram->gndg_state == GNILND_DGRAM_CANCELED ? "canceled" : "terminated");
1025 case GNI_POST_TIMEOUT:
1026 /* we could have a timeout on a wildcard dgram too - if
1027 * we got the incoming request but the remote node beefed
1028 * before kgni could send the match data back. We'll just error
1029 * on the active case and bail out gracefully */
1030 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1031 CNETERR("hardware timeout for connect to "
1032 "%s after %lu seconds. Is node dead?\n",
1033 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1034 cfs_duration_sec(jiffies - dgram->gndg_post_time));
1041 CERROR("dgram 0x%p with bad post_state %d\n", dgram, post_state);
1045 /* now finish cleaning up a dgram that is canceled/terminated and needs to
1048 /* If this was actively canceled, drop the count now that we are processing */
1049 if (dgram->gndg_state == GNILND_DGRAM_CANCELED) {
1050 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1051 /* caller responsible for gndg_list removal */
1059 /* needs dev->gnd_dgram_lock held */
1061 kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram)
1065 if (dgram->gndg_state != GNILND_DGRAM_POSTED) {
1069 LASSERTF(dgram->gndg_conn != NULL,
1070 "dgram 0x%p with NULL conn\n", dgram);
1072 /* C.E - WC dgrams could be canceled immediately but
1073 * if there was some match pending, we need to call
1074 * test_by_id to clear it out. If that test returns
1075 * POST_PENDING, it is half done and needs to go along
1076 * with the rest of dgrams and go through a kgni_sm_task cycle
1077 * and deliver a GNI_POST_TERMINATED event before they
1078 * are actually canceled */
1080 dgram->gndg_state = GNILND_DGRAM_CANCELED;
1082 if (dgram->gndg_conn->gnc_state >= GNILND_CONN_ESTABLISHED) {
1083 /* we don't need to cancel_by_id if the datagram was good */
1087 /* let folks know there are outstanding cancels */
1088 atomic_inc(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1089 /* leave on nid list until cancel is done for debugging fun */
1090 grc = kgnilnd_ep_postdata_cancel_by_id(dgram->gndg_conn->gnc_ephandle, (__u64) dgram);
1092 /* if we don't get success here, we have hosed up the dgram tracking
1093 * code and need to bail out */
1094 LASSERTF(grc == GNI_RC_SUCCESS,
1095 "postdata_cancel returned %d for conn 0x%p to %s\n",
1096 grc, dgram->gndg_conn,
1097 dgram->gndg_conn->gnc_peer ?
1098 libcfs_nid2str(dgram->gndg_conn->gnc_peer->gnp_nid)
1102 "canceled dgram 0x%p conn 0x%p ephandle 0x%p\n",
1103 dgram, dgram->gndg_conn,
1104 dgram->gndg_conn->gnc_ephandle);
1106 if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1107 gni_post_state_t post_state;
1109 __u32 remote_addr = 0, remote_id = 0;
1111 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1112 (__u64)dgram, &post_state,
1113 &remote_addr, &remote_id);
1115 LASSERTF(grc == GNI_RC_NO_MATCH || grc == GNI_RC_SUCCESS,
1116 "bad grc %d from test_by_id on dgram 0x%p\n",
1119 /* if WC was canceled immediately, we get NO_MATCH, if needs to go
1120 * through full cycle, we get SUCCESS and need to parse post_state */
1122 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1123 "remote_addr %u remote_id %u\n", grc, dgram,
1124 kgnilnd_dgram_type2str(dgram),
1125 post_state, remote_addr, remote_id);
1127 if (grc == GNI_RC_NO_MATCH) {
1128 /* she's gone, reduce count and move along */
1129 dgram->gndg_state = GNILND_DGRAM_DONE;
1130 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1134 rc = kgnilnd_process_dgram(dgram, post_state);
1137 /* if for some weird reason we get a valid dgram back, just mark as done
1138 * so we can drop it and move along.
1139 * C.E - if it was completed, we'll just release the conn/mbox
1140 * back into the pool and it'll get reused. That said, we should only
1141 * be canceling a WC dgram on stack rest or shutdown, so that is moot */
1142 dgram->gndg_state = GNILND_DGRAM_DONE;
1143 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1145 /* caller context responsible for calling kgnilnd_release_dgram() */
1147 /* still pending, let it simmer until golden brown and delicious */
1151 /* for non WC dgrams, they are still on the nid list but marked canceled waiting
1152 * for kgni to return their ID to us via probe - that is when we'll complete their
1153 * cancel processing */
1157 kgnilnd_cleanup_dgram(kgn_dgram_t *dgram)
1159 /* release the dgram ref on conn */
1160 if (dgram->gndg_conn) {
1161 kgnilnd_conn_decref(dgram->gndg_conn);
1162 dgram->gndg_conn = NULL;
1167 kgnilnd_free_dgram(kgn_device_t *dev, kgn_dgram_t *dgram)
1169 LASSERTF(dgram->gndg_state == GNILND_DGRAM_USED ||
1170 dgram->gndg_state == GNILND_DGRAM_DONE,
1171 "dgram 0x%p with bad state %s\n",
1172 dgram, kgnilnd_dgram_state2str(dgram));
1174 /* bit of poisoning to help detect bad driver data */
1175 dgram->gndg_magic = 0x6f5a6b5f;
1176 atomic_dec(&dev->gnd_ndgrams);
1178 kmem_cache_free(kgnilnd_data.kgn_dgram_cache, dgram);
1179 CDEBUG(D_MALLOC|D_NETTRACE, "slab-freed 'dgram': %lu at %p %s"
1181 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
1182 atomic_read(&dev->gnd_ndgrams));
1186 kgnilnd_post_dgram(kgn_device_t *dev, lnet_nid_t dstnid, kgn_connreq_type_t type,
1190 kgn_dgram_t *dgram = NULL;
1191 kgn_dgram_t *tmpdgram;
1192 kgn_dgram_type_t dgtype;
1198 case GNILND_CONNREQ_REQ:
1199 if (dstnid == LNET_NID_ANY)
1200 dgtype = GNILND_DGRAM_WC_REQ;
1202 dgtype = GNILND_DGRAM_REQ;
1204 case GNILND_CONNREQ_NAK:
1205 LASSERTF(dstnid != LNET_NID_ANY, "can't NAK to LNET_NID_ANY\n");
1206 dgtype = GNILND_DGRAM_NAK;
1209 CERROR("unknown connreq type %d\n", type);
1213 rc = kgnilnd_alloc_dgram(&dgram, dev, dgtype);
1216 GOTO(post_failed, rc);
1219 rc = kgnilnd_create_conn(&dgram->gndg_conn, dev);
1221 GOTO(post_failed, rc);
1224 if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1225 /* clear buffer for sanity on reuse of wildcard */
1226 memset(&dgram->gndg_conn_in, 0, sizeof(kgn_connreq_t));
1229 if (dstnid == LNET_NID_ANY) {
1230 /* set here to reset any dgram re-use */
1231 dgram->gndg_conn->gnc_state = GNILND_CONN_LISTEN;
1235 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(dstnid), 1, &host_id);
1238 GOTO(post_failed, rc);
1241 dgram->gndg_conn->gnc_state = GNILND_CONN_CONNECTING;
1243 /* don't need to serialize, there are no CQs for the dgram
1244 * EP on the kgn_net_t */
1245 grc = kgnilnd_ep_bind(dgram->gndg_conn->gnc_ephandle, host_id, dev->gnd_id);
1247 if (grc != GNI_RC_SUCCESS) {
1249 GOTO(post_failed, rc);
1254 /* If we are posting wildcards post using a net of 0, otherwise we'll use the
1255 * net of the destination node.
1258 if (dstnid == LNET_NID_ANY) {
1259 srcnid = LNET_MKNID(LNET_MKNET(GNILND, 0), dev->gnd_nid);
1261 srcnid = LNET_MKNID(LNET_NIDNET(dstnid), dev->gnd_nid);
1264 rc = kgnilnd_pack_connreq(&dgram->gndg_conn_out, dgram->gndg_conn,
1265 srcnid, dstnid, type);
1267 GOTO(post_failed, rc);
1270 if (type == GNILND_CONNREQ_NAK)
1271 dgram->gndg_conn_out.gncr_nakdata.gnnd_errno = data_rc;
1273 dgram->gndg_post_time = jiffies;
1275 /* XXX Nic: here is where we'd add in logical network multiplexing */
1277 CDEBUG(D_NETTRACE, "dgram 0x%p type %s %s->%s cdm %d\n",
1278 dgram, kgnilnd_dgram_type2str(dgram),
1279 libcfs_nid2str(srcnid),
1280 libcfs_nid2str(dstnid), dev->gnd_id);
1282 /* this allocates memory, can't hold locks across */
1283 grc = kgnilnd_ep_postdata_w_id(dgram->gndg_conn->gnc_ephandle,
1284 &dgram->gndg_conn_out, sizeof(kgn_connreq_t),
1285 &dgram->gndg_conn_in, sizeof(kgn_connreq_t),
1288 if (grc != GNI_RC_SUCCESS) {
1289 CNETERR("dropping failed dgram post id 0x%p type %s"
1290 " reqtype %s to %s: rc %d\n",
1291 dgram, kgnilnd_dgram_type2str(dgram),
1292 kgnilnd_connreq_type2str(&dgram->gndg_conn_out),
1293 libcfs_nid2str(dstnid), grc);
1294 rc = (grc == GNI_RC_ERROR_NOMEM) ? -ENOMEM : -EBADR;
1295 GOTO(post_failed, rc);
1298 /* we don't need to add earlier - if someone does del_peer during post,
1299 * that peer will get marked as unlinked and the callers wil take care of it.
1300 * The dgram code is largely kgn_peer_t ignorant, so at worst, we'll just drop
1301 * the completed dgram later when we cant find a peer to stuff it into */
1303 spin_lock(&dev->gnd_dgram_lock);
1305 /* make sure we are not double posting targeted dgrams
1306 * - we can multiple post WC dgrams to help with processing speed */
1307 if (dstnid != LNET_NID_ANY) {
1308 tmpdgram = kgnilnd_find_dgram_locked(dev, dstnid);
1310 LASSERTF(tmpdgram == NULL,
1311 "dgram 0x%p->%s already posted\n",
1312 dgram, libcfs_nid2str(dstnid));
1315 /* unmunge dstnid to help processing code cope... */
1316 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
1317 dgram->gndg_conn_out.gncr_dstnid = dstnid;
1320 list_add_tail(&dgram->gndg_list, kgnilnd_nid2dgramlist(dev, dstnid));
1321 dgram->gndg_state = GNILND_DGRAM_POSTED;
1322 spin_unlock(&dev->gnd_dgram_lock);
1325 if (rc < 0 && dgram != NULL) {
1326 kgnilnd_cleanup_dgram(dgram);
1327 kgnilnd_free_dgram(dev, dgram);
1333 /* The shutdown flag is set from the shutdown and stack reset threads. */
1335 kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown)
1337 /* The conns of canceled active dgrams need to be put in purgatory so
1338 * we don't reuse the mailbox */
1339 if (unlikely(dgram->gndg_state == GNILND_DGRAM_CANCELED)) {
1341 kgn_conn_t *conn = dgram->gndg_conn;
1342 lnet_nid_t nid = dgram->gndg_conn_out.gncr_dstnid;
1344 dgram->gndg_state = GNILND_DGRAM_DONE;
1346 /* During shutdown we've already removed the peer so we don't
1347 * need to add a peer. During stack reset we don't care about
1348 * MDDs since they are all released. */
1350 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1351 peer = kgnilnd_find_peer_locked(nid);
1354 CDEBUG(D_NET, "adding peer's conn with nid %s "
1355 "to purgatory\n", libcfs_nid2str(nid));
1356 kgnilnd_conn_addref(conn);
1357 conn->gnc_peer = peer;
1358 kgnilnd_peer_addref(peer);
1359 kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
1360 conn->gnc_state = GNILND_CONN_CLOSED;
1361 list_add_tail(&conn->gnc_list,
1363 kgnilnd_add_purgatory_locked(conn,
1365 kgnilnd_schedule_conn(conn);
1367 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1371 spin_lock(&dev->gnd_dgram_lock);
1372 kgnilnd_cancel_dgram_locked(dgram);
1373 spin_unlock(&dev->gnd_dgram_lock);
1375 kgnilnd_cleanup_dgram(dgram);
1377 /* if the dgram is 'canceled' it needs to be wait until the event
1378 * comes up from kgni that tells us it is safe to release */
1379 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1380 dgram->gndg_state = GNILND_DGRAM_DONE;
1382 LASSERTF(list_empty(&dgram->gndg_list), "dgram 0x%p on list\n", dgram);
1384 /* if it is a wildcard and we are in an appropriate state, repost
1387 if ((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
1388 (!kgnilnd_data.kgn_wc_kill && !kgnilnd_data.kgn_in_reset)) {
1391 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1393 /* We failed to repost the WC dgram for some reason
1394 * mark it so the repost system attempts to repost */
1395 kgnilnd_admin_addref(dev->gnd_nwcdgrams);
1399 /* always free the old dgram */
1400 kgnilnd_free_dgram(dev, dgram);
1406 kgnilnd_probe_for_dgram(kgn_device_t *dev, kgn_dgram_t **dgramp)
1408 kgn_dgram_t *dgram = NULL;
1409 gni_post_state_t post_state;
1413 __u32 remote_addr = 0, remote_id = 0;
1416 /* Probe with the lock held. That way if we get a dgram we dont have it canceled
1417 * between finding the ready dgram and grabbing the lock to remove it from the
1418 * list. Otherwise we could be left in an inconsistent state. We own the dgram
1419 * once its off the list so we don't need to worry about others changing it at
1421 spin_lock(&dev->gnd_dgram_lock);
1422 grc = kgnilnd_postdata_probe_by_id(dev->gnd_handle, &readyid);
1423 if (grc != GNI_RC_SUCCESS) {
1424 spin_unlock(&dev->gnd_dgram_lock);
1425 /* return 0 to indicate nothing happened */
1429 CDEBUG(D_NET, "ready "LPX64" on device 0x%p\n",
1432 dgram = (kgn_dgram_t *)readyid;
1434 LASSERTF(dgram->gndg_magic == GNILND_DGRAM_MAGIC,
1435 "dgram 0x%p from id "LPX64" with bad magic %x\n",
1436 dgram, readyid, dgram->gndg_magic);
1438 LASSERTF(dgram->gndg_state == GNILND_DGRAM_POSTED ||
1439 dgram->gndg_state == GNILND_DGRAM_CANCELED,
1440 "dgram 0x%p with bad state %s\n",
1441 dgram, kgnilnd_dgram_state2str(dgram));
1443 LASSERTF(!list_empty(&dgram->gndg_list),
1444 "dgram 0x%p with bad list state %s type %s\n",
1445 dgram, kgnilnd_dgram_state2str(dgram),
1446 kgnilnd_dgram_type2str(dgram));
1448 /* now we know that the datagram structure is ok, so pull off list */
1449 list_del_init(&dgram->gndg_list);
1451 /* while we have the gnn_dgram_lock and BEFORE we call test_by_id
1452 * change the state from POSTED to PROCESSING to ensure that
1453 * nobody cancels it after we've pulled it from the wire */
1454 if (dgram->gndg_state == GNILND_DGRAM_POSTED) {
1455 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1458 LASSERTF(dgram->gndg_conn != NULL,
1459 "dgram 0x%p with NULL conn\n", dgram);
1461 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1462 (__u64)dgram, &post_state,
1463 &remote_addr, &remote_id);
1465 /* we now "own" this datagram */
1466 spin_unlock(&dev->gnd_dgram_lock);
1468 LASSERTF(grc != GNI_RC_NO_MATCH, "kgni lied! probe_by_id told us that"
1469 " id "LPU64" was ready\n", readyid);
1471 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1472 "remote_addr %u remote_id %u\n", grc, dgram,
1473 kgnilnd_dgram_type2str(dgram),
1474 post_state, remote_addr, remote_id);
1476 if (unlikely(grc != GNI_RC_SUCCESS)) {
1477 CNETERR("getting data for dgram 0x%p->%s failed rc %d. Dropping it\n",
1478 dgram, libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1481 GOTO(probe_for_out, rc);
1484 rc = kgnilnd_process_dgram(dgram, post_state);
1486 /* we should never get probe finding a dgram for us and then it
1487 * being a WC dgram that is still in the middle of processing */
1488 LASSERTF(rc <= 0, "bad rc %d from process_dgram 0x%p state %d\n",
1489 rc, dgram, post_state);
1492 /* dgram is good enough for the data to be used */
1493 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1494 /* fake rc to mark that we've done something */
1497 /* let kgnilnd_release_dgram take care of canceled dgrams */
1498 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1499 dgram->gndg_state = GNILND_DGRAM_DONE;
1508 kgnilnd_release_dgram(dev, dgram, 0);
1513 kgnilnd_setup_wildcard_dgram(kgn_device_t *dev)
1515 /* if kgn_wildcard is zero, return error */
1516 int rc = -ENOENT, i;
1519 for (i = 0; i < *kgnilnd_tunables.kgn_nwildcard; i++) {
1520 rc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1522 CERROR("error %d: could not post wildcard datagram # %d\n",
1534 kgnilnd_cancel_net_dgrams(kgn_net_t *net)
1536 kgn_dgram_t *dg, *dgN;
1537 struct list_head zombies;
1541 /* we want to cancel any outstanding dgrams - we don't want to rely
1542 * on del_peer_or_conn catching all of them. This helps protect us in cases
1543 * where we don't quite keep the peer->dgram mapping in sync due to some
1544 * race conditions */
1546 LASSERTF(net->gnn_shutdown || kgnilnd_data.kgn_in_reset,
1547 "called with LND invalid state: net shutdown %d "
1548 "in reset %d\n", net->gnn_shutdown,
1549 kgnilnd_data.kgn_in_reset);
1551 INIT_LIST_HEAD(&zombies);
1553 spin_lock(&net->gnn_dev->gnd_dgram_lock);
1555 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1556 list_for_each_entry_safe(dg, dgN, &net->gnn_dev->gnd_dgrams[i], gndg_list) {
1558 /* skip nids not on our net or are wildcards */
1561 if (dg->gndg_type == GNILND_DGRAM_WC_REQ ||
1562 net->gnn_netnum != LNET_NETNUM(LNET_NIDNET(dg->gndg_conn_out.gncr_dstnid)))
1565 kgnilnd_cancel_dgram_locked(dg);
1569 spin_unlock(&net->gnn_dev->gnd_dgram_lock);
1575 kgnilnd_cancel_wc_dgrams(kgn_device_t *dev)
1577 kgn_dgram_t *dg, *dgN;
1578 struct list_head zombies;
1581 /* Time to kill the outstanding WC's
1582 * WC's exist on net 0 only but match on any net...
1585 LASSERTF(kgnilnd_data.kgn_in_reset || kgnilnd_data.kgn_wc_kill,
1586 "called with LND invalid state: WC shutdown %d "
1587 "in reset %d\n", kgnilnd_data.kgn_wc_kill,
1588 kgnilnd_data.kgn_in_reset);
1590 INIT_LIST_HEAD(&zombies);
1591 spin_lock(&dev->gnd_dgram_lock);
1594 dg = kgnilnd_find_dgram_locked(dev, LNET_NID_ANY);
1596 LASSERTF(dg->gndg_type == GNILND_DGRAM_WC_REQ,
1597 "dgram 0x%p->%s with bad type %d (%s)\n",
1598 dg, libcfs_nid2str(dg->gndg_conn_out.gncr_dstnid),
1599 dg->gndg_type, kgnilnd_dgram_type2str(dg));
1601 kgnilnd_cancel_dgram_locked(dg);
1603 /* WC could be DONE already, check and if so add to list to be released */
1604 if (dg->gndg_state == GNILND_DGRAM_DONE) {
1605 list_del_init(&dg->gndg_list);
1606 list_add_tail(&dg->gndg_list, &zombies);
1609 } while (dg != NULL);
1611 spin_unlock(&dev->gnd_dgram_lock);
1613 list_for_each_entry_safe(dg, dgN, &zombies, gndg_list) {
1614 list_del_init(&dg->gndg_list);
1615 kgnilnd_release_dgram(dev, dg, 1);
1622 kgnilnd_cancel_dgrams(kgn_device_t *dev)
1624 kgn_dgram_t *dg, *dgN;
1628 /* Cancel any outstanding non wildcard datagrams regardless
1629 * of which net they are on as we are in base shutdown and
1630 * dont care about connecting anymore.
1633 LASSERTF(kgnilnd_data.kgn_wc_kill == 1,"We didnt get called from base shutdown\n");
1635 spin_lock(&dev->gnd_dgram_lock);
1637 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size -1); i++) {
1638 list_for_each_entry_safe(dg, dgN, &dev->gnd_dgrams[i], gndg_list) {
1639 if (dg->gndg_type != GNILND_DGRAM_WC_REQ)
1640 kgnilnd_cancel_dgram_locked(dg);
1644 spin_unlock(&dev->gnd_dgram_lock);
1651 kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev)
1659 /* use do while to get at least one check run to allow
1660 * regression test for 762072 to hit bug if there */
1662 /* This function races with the dgram mover during shutdown so it is possible for
1663 * a dgram to be seen in kgnilnd_postdata_probe_wait_by_id but be handled in the
1664 * dgram mover thread instead of inside of this function.
1667 /* This should only be called from within shutdown, baseshutdown, or stack reset.
1668 * there are no assertions here to verify since base_shutdown has nothing in it we can check
1669 * the net is gone by then.
1674 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1675 "Waiting for %d canceled datagrams to clear on device %d\n",
1676 atomic_read(&dev->gnd_canceled_dgrams), dev->gnd_id);
1678 /* check once a second */
1679 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
1682 if (grc != GNI_RC_SUCCESS)
1685 CDEBUG(D_NET, "ready "LPX64" on device %d->0x%p\n",
1686 readyid, dev->gnd_id, dev);
1688 rc = kgnilnd_probe_for_dgram(dev, &dgram);
1690 /* if we got a valid dgram or one that is now done, clean up */
1691 kgnilnd_release_dgram(dev, dgram, 1);
1693 } while (atomic_read(&dev->gnd_canceled_dgrams));
1697 kgnilnd_start_connect(kgn_peer_t *peer)
1700 /* sync point for kgnilnd_del_peer_locked - do an early check to
1701 * catch the most common hits where del_peer is done by the
1702 * time we get here */
1703 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING1)) {
1704 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING1, 1)) {};
1707 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1708 if (!kgnilnd_peer_active(peer) || peer->gnp_connecting != GNILND_PEER_CONNECT) {
1709 /* raced with peer getting unlinked */
1710 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1714 peer->gnp_connecting = GNILND_PEER_POSTING;
1715 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1717 set_mb(peer->gnp_last_dgram_time, jiffies);
1718 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING2)) {
1719 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING2, 1)) {};
1722 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING3)) {
1723 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING3, 1)) {};
1724 rc = cfs_fail_val ? cfs_fail_val : -ENOMEM;
1726 rc = kgnilnd_post_dgram(peer->gnp_net->gnn_dev,
1727 peer->gnp_nid, GNILND_CONNREQ_REQ, 0);
1730 set_mb(peer->gnp_last_dgram_errno, rc);
1734 /* while we're posting someone could have decided this peer/dgram needed to
1735 * die a quick death, so we check for state change and process accordingly */
1737 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1738 if (!kgnilnd_peer_active(peer) || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1739 if (peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1740 peer->gnp_connecting = GNILND_PEER_KILL;
1742 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1743 /* positive RC to avoid dgram cleanup - we'll have to
1744 * wait for the kgni GNI_POST_TERMINATED event to
1745 * finish cleaning up */
1747 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev, peer->gnp_nid);
1750 peer->gnp_connecting = GNILND_PEER_POSTED;
1751 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1752 /* reaper thread will take care of any timeouts */
1753 CDEBUG(D_NET, "waiting for connect to finish to %s rc %d\n",
1754 libcfs_nid2str(peer->gnp_nid), rc);
1759 CDEBUG(D_NET, "connect to %s failed: rc %d \n",
1760 libcfs_nid2str(peer->gnp_nid), rc);
1766 kgnilnd_finish_connect(kgn_dgram_t *dgram)
1768 kgn_conn_t *conn = dgram->gndg_conn;
1769 lnet_nid_t her_nid = dgram->gndg_conn_in.gncr_srcnid;
1770 kgn_peer_t *new_peer, *peer = NULL;
1773 kgn_mbox_info_t *mbox;
1777 /* try to find a peer that matches the nid we got in the connreq
1778 * kgnilnd_unpack_connreq makes sure that conn_in.gncr_srcnid is
1779 * HER and conn_out.gncr_srcnid is ME for both active and WC dgrams */
1781 /* assume this is a new peer - it makes locking cleaner when it isn't */
1782 /* no holding kgn_net_rw_sem - already are at the kgnilnd_dgram_mover level */
1784 rc = kgnilnd_create_peer_safe(&new_peer, her_nid, NULL, GNILND_RCA_NODE_UP);
1786 CERROR("Can't create peer for %s\n", libcfs_nid2str(her_nid));
1790 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1792 /* this transfers ref from create_peer to the kgn_peer table */
1793 kgnilnd_add_peer_locked(her_nid, new_peer, &peer);
1795 /* if we found an existing peer, is it really ready for a new conn ? */
1796 if (peer != new_peer) {
1797 /* if this was an active connect attempt but we can't find a peer waiting for it
1798 * we will dump in the trash */
1800 if (peer->gnp_connecting == GNILND_PEER_IDLE && dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1801 CDEBUG(D_NET, "dropping completed connreq for %s peer 0x%p->%s\n",
1802 libcfs_nid2str(her_nid), peer, libcfs_nid2str(peer->gnp_nid));
1803 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1808 /* check to see if we can catch a connecting peer before it is
1809 * removed from the connd_peers list - if not, we need to
1810 * let the connreqs race and be handled by kgnilnd_conn_isdup_locked() */
1811 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
1812 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1813 if (!list_empty(&peer->gnp_connd_list)) {
1814 list_del_init(&peer->gnp_connd_list);
1815 /* drop connd ref */
1816 kgnilnd_peer_decref(peer);
1818 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1819 /* clear rc to make sure we don't have fake error */
1823 /* no matter what, we are no longer waiting to connect this peer now */
1824 peer->gnp_connecting = GNILND_PEER_IDLE;
1826 /* Refuse to duplicate an existing connection (both sides might try to
1827 * connect at once). NB we return success! We _are_ connected so we
1828 * _don't_ have any blocked txs to complete with failure. */
1829 rc = kgnilnd_conn_isdup_locked(peer, conn);
1831 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1832 CDEBUG(D_NET, "Not creating duplicate connection to %s: %d\n",
1833 libcfs_nid2str(her_nid), rc);
1839 if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
1840 CNETERR("Received connection request from down nid %s\n",
1841 libcfs_nid2str(her_nid));
1842 peer->gnp_down = GNILND_RCA_NODE_UP;
1845 nstale = kgnilnd_close_stale_conns_locked(peer, conn);
1847 /* either way with peer (new or existing), we are ok with ref counts here as the
1848 * kgnilnd_add_peer_locked will use our ref on new_peer (from create_peer_safe) as the
1849 * ref for the peer table. */
1851 /* at this point, the connection request is a winner */
1853 /* mark 'DONE' to avoid cancel being called from release */
1854 dgram->gndg_state = GNILND_DGRAM_DONE;
1856 /* initialise timestamps before reaper looks at them */
1857 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
1859 /* last_tx is initialized to jiffies - (keepalive*2) so that if the NOOP fails it will
1860 * immediatly send a NOOP in the reaper thread during the call to
1861 * kgnilnd_check_conn_timeouts_locked
1863 conn->gnc_last_tx = jiffies - (cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout)) * 2);
1864 conn->gnc_state = GNILND_CONN_ESTABLISHED;
1866 /* save the dgram type used to establish this connection */
1867 conn->gnc_dgram_type = dgram->gndg_type;
1869 /* refs are not transferred from dgram to tables, so increment to
1871 kgnilnd_conn_addref(conn);
1872 kgnilnd_peer_addref(peer);
1873 conn->gnc_peer = peer;
1874 list_add_tail(&conn->gnc_list, &peer->gnp_conns);
1876 kgnilnd_conn_addref(conn); /* +1 ref for conn table */
1877 list_add_tail(&conn->gnc_hashlist,
1878 kgnilnd_cqid2connlist(conn->gnc_cqid));
1879 kgnilnd_data.kgn_conn_version++;
1881 /* Dont send NOOP if fail_loc is set
1883 if (!CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP)) {
1884 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, peer->gnp_net->gnn_ni->ni_nid);
1886 CNETERR("can't get TX to initiate NOOP to %s\n",
1887 libcfs_nid2str(peer->gnp_nid));
1889 kgnilnd_queue_tx(conn, tx);
1893 /* Schedule all packets blocking for a connection */
1894 list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1895 /* lock held here is the peer_conn lock */
1896 kgnilnd_tx_del_state_locked(tx, peer, NULL, GNILND_TX_ALLOCD);
1897 kgnilnd_queue_tx(conn, tx);
1900 /* If this is an active connection lets mark its timestamp on the MBoX */
1901 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1902 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1903 /* conn->gnc_last_rx is jiffies it better exist as it was just set */
1904 mbox->mbx_release_purg_active_dgram = conn->gnc_last_rx;
1907 /* Bug 765042: wake up scheduler for a race with finish_connect and
1908 * complete_conn_closed with a conn in purgatory
1909 * since we can't use CFS_RACE due to mutex_holds in kgnilnd_process_conns,
1910 * we just check for set and then clear */
1911 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
1913 /* get scheduler thread moving again */
1914 kgnilnd_schedule_device(conn->gnc_device);
1917 CDEBUG(D_NET, "New conn 0x%p->%s dev %d\n",
1918 conn, libcfs_nid2str(her_nid), conn->gnc_device->gnd_id);
1920 /* make sure we reset peer reconnect interval now that we have a good conn */
1921 kgnilnd_peer_alive(peer);
1922 peer->gnp_reconnect_interval = 0;
1924 /* clear the unlink attribute if we dont clear it kgnilnd_del_conn_or_peer will wait
1925 * on the atomic forever
1927 if (peer->gnp_pending_unlink) {
1928 peer->gnp_pending_unlink = 0;
1929 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1930 CDEBUG(D_NET, "Clearing peer unlink %p\n",peer);
1933 /* add ref to make it hang around until after we drop the lock */
1934 kgnilnd_conn_addref(conn);
1936 /* Once the peer_conn lock is dropped, the conn could actually move into
1937 * CLOSING->CLOSED->DONE in the scheduler thread, so hold the
1938 * lock until we are really done */
1939 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1941 /* Notify LNET that we now have a working connection to this peer.
1942 * This is a Cray extension to the "standard" LND behavior. */
1943 lnet_notify(peer->gnp_net->gnn_ni, peer->gnp_nid,
1944 1, cfs_time_current());
1946 /* drop our 'hold' ref */
1947 kgnilnd_conn_decref(conn);
1954 kgnilnd_send_nak(kgn_device_t *dev, lnet_nid_t dst_nid, int error)
1959 LASSERTF(dst_nid != LNET_NID_ANY, "bad dst_nid %s\n", libcfs_nid2str(dst_nid));
1961 CDEBUG(D_NET, "NAK to %s errno %d\n", libcfs_nid2str(dst_nid), error);
1963 rc = kgnilnd_post_dgram(dev, dst_nid, GNILND_CONNREQ_NAK, error);
1966 CDEBUG(D_NET, "NAK to %s failed: rc %d \n", libcfs_nid2str(dst_nid), rc);
1972 kgnilnd_process_nak(kgn_dgram_t *dgram)
1974 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
1975 lnet_nid_t src_nid = connreq->gncr_srcnid;
1976 int errno = connreq->gncr_nakdata.gnnd_errno;
1980 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1982 peer = kgnilnd_find_peer_locked(src_nid);
1984 /* we likely dropped him from bad data when we processed
1985 * the original REQ */
1986 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1990 /* need to check peerstamp/connstamp against the ones we find
1991 * to make sure we don't close new (and good?) conns that we
1992 * formed after this connreq failed */
1993 if (peer->gnp_connecting == GNILND_PEER_IDLE) {
1996 if (list_empty(&peer->gnp_conns)) {
1997 /* assume already procced datagram and it barfed up
1998 * on this side too */
1999 CDEBUG(D_NET, "dropping NAK from %s; "
2000 "peer %s is already not connected\n",
2001 libcfs_nid2str(connreq->gncr_srcnid),
2002 libcfs_nid2str(connreq->gncr_dstnid));
2003 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2007 /* stub up a connection with the connreq XXX_stamps to allow
2008 * use to use close_stale_conns_locked */
2009 conn.gnc_peerstamp = connreq->gncr_peerstamp;
2010 conn.gnc_my_connstamp = connreq->gncr_connstamp;
2011 conn.gnc_peer_connstamp = connreq->gncr_connstamp;
2012 conn.gnc_device = peer->gnp_net->gnn_dev;
2014 rc = kgnilnd_close_stale_conns_locked(peer, &conn);
2016 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
2017 "closed %d connections\n",
2018 libcfs_nid2str(connreq->gncr_srcnid),
2019 libcfs_nid2str(connreq->gncr_dstnid), errno, rc);
2021 spin_lock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2023 if (list_empty(&peer->gnp_connd_list)) {
2024 /* if peer isn't on waiting list, try to find one to nuke */
2025 rc = kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
2029 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
2030 "canceled pending connect request\n",
2031 libcfs_nid2str(connreq->gncr_srcnid),
2032 libcfs_nid2str(connreq->gncr_dstnid), errno);
2035 /* if we can't find a waiting dgram, we just drop the nak - the conn
2036 * connect must have failed (didn't find conn above and clear connecting
2037 * -- so nothing to do besides drop */
2039 /* peer is on list, meaning it is a new connect attempt from the one
2040 * we started that generated the NAK - so just drop NAK */
2042 /* use negative to prevent error message */
2045 spin_unlock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2048 /* success! we found a peer and at least marked pending_nak */
2049 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2055 kgnilnd_process_connreq(kgn_dgram_t *dgram, int *needs_nak)
2059 rc = kgnilnd_unpack_connreq(dgram);
2062 /* only NAK if we have good srcnid to use */
2068 switch (dgram->gndg_conn_in.gncr_type) {
2069 case GNILND_CONNREQ_REQ:
2070 /* wire up peer & conn, send queued TX */
2071 rc = kgnilnd_finish_connect(dgram);
2073 /* don't nak when the nid is hosed */
2079 case GNILND_CONNREQ_NAK:
2080 rc = kgnilnd_process_nak(dgram);
2081 /* return early to prevent reconnect bump */
2084 CERROR("unexpected connreq type %s (%d) from %s\n",
2085 kgnilnd_connreq_type2str(&dgram->gndg_conn_in),
2086 dgram->gndg_conn_in.gncr_type,
2087 libcfs_nid2str(dgram->gndg_conn_in.gncr_srcnid));
2098 kgnilnd_probe_and_process_dgram(kgn_device_t *dev)
2102 lnet_nid_t nak_dstnid = LNET_NID_ANY;
2103 lnet_nid_t orig_dstnid;
2104 kgn_dgram_t *dgram = NULL;
2108 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PAUSE_DGRAM_COMP)) {
2111 rc = kgnilnd_probe_for_dgram(dev, &dgram);
2116 } else if (rc < 0) {
2117 GOTO(inform_peer, rc);
2119 /* rc > 1 means it did something, reset for this func */
2123 switch (dgram->gndg_type) {
2124 case GNILND_DGRAM_WC_REQ:
2125 case GNILND_DGRAM_REQ:
2126 rc = kgnilnd_process_connreq(dgram, &needs_nak);
2128 case GNILND_DGRAM_NAK:
2129 CDEBUG(D_NETTRACE, "NAK to %s done\n",
2130 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
2133 CERROR("unknown datagram type %s (%d)\n",
2134 kgnilnd_dgram_type2str(dgram), dgram->gndg_type);
2138 /* stash data to use after releasing current datagram */
2139 /* don't stash net - we are operating on a net already,
2140 * so the lock on rw_net_lock is sufficient */
2142 nak_dstnid = dgram->gndg_conn_in.gncr_srcnid;
2145 LASSERTF(dgram != NULL, "dgram 0x%p rc %d needs_nak %d\n", dgram, rc, needs_nak);
2147 orig_dstnid = dgram->gndg_conn_out.gncr_dstnid;
2149 kgnilnd_release_dgram(dev, dgram, 0);
2151 CDEBUG(D_NET, "cleaning up dgram to %s, rc %d\n",
2152 libcfs_nid2str(orig_dstnid), rc);
2154 /* if this was a WC_REQ that matched an existing peer, it'll get marked done
2155 * in kgnilnd_finish_connect - if errors are from before we get to there,
2156 * we just drop as it is a WC_REQ - the peer CAN'T be waiting for it */
2157 if ((orig_dstnid != LNET_NID_ANY) && (rc < 0)) {
2158 /* if we have a negative rc, we want to find a peer to inform about
2159 * the bad connection attempt. Sorry buddy, better luck next time! */
2161 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2162 peer = kgnilnd_find_peer_locked(orig_dstnid);
2165 /* add ref to make sure he stays around past the possible unlink
2166 * so we can tell LNet about him */
2167 kgnilnd_peer_addref(peer);
2169 /* if he still cares about the outstanding connect */
2170 if (peer->gnp_connecting >= GNILND_PEER_CONNECT) {
2171 /* check if he is on the connd list and remove.. */
2172 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2173 if (!list_empty(&peer->gnp_connd_list)) {
2174 list_del_init(&peer->gnp_connd_list);
2175 /* drop connd ref */
2176 kgnilnd_peer_decref(peer);
2178 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2180 /* clear gnp_connecting so we don't have a non-connecting peer
2181 * on gnd_connd_list */
2182 peer->gnp_connecting = GNILND_PEER_IDLE;
2184 set_mb(peer->gnp_last_dgram_errno, rc);
2186 kgnilnd_peer_increase_reconnect_locked(peer);
2189 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2191 /* now that we are outside the lock, tell Mommy */
2193 kgnilnd_peer_notify(peer, rc, 0);
2194 kgnilnd_peer_decref(peer);
2199 kgnilnd_send_nak(dev, nak_dstnid, rc);
2206 kgnilnd_reaper_dgram_check(kgn_device_t *dev)
2208 kgn_dgram_t *dgram, *tmp;
2211 spin_lock(&dev->gnd_dgram_lock);
2213 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
2214 list_for_each_entry_safe(dgram, tmp, &dev->gnd_dgrams[i], gndg_list) {
2215 unsigned long now = jiffies;
2216 unsigned long timeout;
2218 /* don't timeout stuff if the network is mucked or shutting down */
2219 if (kgnilnd_check_hw_quiesce()) {
2223 if ((dgram->gndg_state != GNILND_DGRAM_POSTED) ||
2224 (dgram->gndg_type == GNILND_DGRAM_WC_REQ)) {
2227 CDEBUG(D_NETTRACE, "checking dgram 0x%p type %s "
2228 "state %s conn 0x%p to %s age %lus\n",
2229 dgram, kgnilnd_dgram_type2str(dgram),
2230 kgnilnd_dgram_state2str(dgram), dgram->gndg_conn,
2231 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2232 cfs_duration_sec(now - dgram->gndg_post_time));
2234 timeout = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout);
2236 if (time_before(now, (dgram->gndg_post_time + timeout)))
2239 CNETERR("%s datagram to %s timed out @ %lus dgram "
2240 "0x%p state %s conn 0x%p\n",
2241 kgnilnd_dgram_type2str(dgram),
2242 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2243 cfs_duration_sec(now - dgram->gndg_post_time),
2244 dgram, kgnilnd_dgram_state2str(dgram),
2247 kgnilnd_cancel_dgram_locked(dgram);
2250 spin_unlock(&dev->gnd_dgram_lock);
2254 /* use a thread for the possibly long-blocking wait_by_id to prevent
2255 * stalling the global workqueues */
2257 kgnilnd_dgram_waitq(void *arg)
2259 kgn_device_t *dev = (kgn_device_t *) arg;
2263 DEFINE_WAIT(mover_done);
2265 snprintf(name, sizeof(name), "kgnilnd_dgn_%02d", dev->gnd_id);
2266 cfs_block_allsigs();
2268 /* all gnilnd threads need to run fairly urgently */
2269 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2271 /* we dont shut down until the device shuts down ... */
2272 while (!kgnilnd_data.kgn_shutdown) {
2273 /* to quiesce or to not quiesce, that is the question */
2274 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2275 KGNILND_SPIN_QUIESCE;
2278 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_DGRAM_COMP, 1)) {}
2280 /* check once a second */
2281 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
2284 if (grc == GNI_RC_SUCCESS) {
2285 CDEBUG(D_INFO, "waking up dgram mover thread\n");
2286 kgnilnd_schedule_dgram(dev);
2288 /* wait for dgram thread to ping us before spinning again */
2289 prepare_to_wait(&dev->gnd_dgping_waitq, &mover_done,
2290 TASK_INTERRUPTIBLE);
2292 /* don't sleep if we need to quiesce */
2293 if (likely(!kgnilnd_data.kgn_quiesce_trigger)) {
2296 finish_wait(&dev->gnd_dgping_waitq, &mover_done);
2300 kgnilnd_thread_fini();
2305 kgnilnd_start_outbound_dgrams(kgn_device_t *dev, unsigned long deadline)
2307 int did_something = 0, rc;
2308 kgn_peer_t *peer = NULL;
2310 spin_lock(&dev->gnd_connd_lock);
2312 /* Active connect - we added this in kgnilnd_launch_tx */
2313 while (!list_empty(&dev->gnd_connd_peers) && time_before(jiffies, deadline)) {
2314 peer = list_first_entry(&dev->gnd_connd_peers,
2315 kgn_peer_t, gnp_connd_list);
2317 /* ref for connd removed in if/else below */
2318 list_del_init(&peer->gnp_connd_list);
2320 /* gnp_connecting and membership on gnd_connd_peers should be
2321 * done coherently to avoid double adding, etc */
2322 /* don't need kgnilnd_data.kgn_peer_conn_lock here as that is only needed
2323 * to get the peer to gnp_connecting in the first place. We just need to
2324 * rely on gnd_connd_lock to serialize someone pulling him from the list
2325 * BEFORE clearing gnp_connecting */
2326 LASSERTF(peer->gnp_connecting != GNILND_PEER_IDLE, "peer 0x%p->%s not connecting\n",
2327 peer, libcfs_nid2str(peer->gnp_nid));
2329 spin_unlock(&dev->gnd_connd_lock);
2331 CDEBUG(D_NET, "processing connect to %s\n",
2332 libcfs_nid2str(peer->gnp_nid));
2335 rc = kgnilnd_start_connect(peer);
2337 if (likely(rc >= 0)) {
2338 /* 0 on success, positive on 'just drop peer' errors */
2339 kgnilnd_peer_decref(peer);
2340 } else if (rc == -ENOMEM) {
2341 /* if we are out of wildcards, add back to
2342 * connd_list - then break out and we'll try later
2343 * if other errors, we'll bail & cancel pending tx */
2344 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2345 if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2346 peer->gnp_connecting = GNILND_PEER_CONNECT;
2347 spin_lock(&dev->gnd_connd_lock);
2348 list_add_tail(&peer->gnp_connd_list,
2349 &dev->gnd_connd_peers);
2351 /* connecting changed while we were posting */
2353 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2354 " state 0x%p->%s, connecting %d\n",
2355 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2356 peer->gnp_connecting = GNILND_PEER_KILL;
2357 spin_lock(&dev->gnd_connd_lock);
2358 /* remove the peer ref frrom the cond list */
2359 kgnilnd_peer_decref(peer);
2360 /* let the system handle itself */
2362 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2363 /* the datagrams are a global pool,
2364 * so break out of trying and hope some free
2369 /* something bad happened, you lose */
2370 CNETERR("could not start connecting to %s "
2371 "rc %d: Will retry until TX timeout\n",
2372 libcfs_nid2str(peer->gnp_nid), rc);
2373 /* It didnt post so just set connecting back to zero now.
2374 * The reaper will reattempt the connection if it needs too.
2375 * If the peer needs death set it so the reaper will cleanup.
2377 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2378 if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2379 peer->gnp_connecting = GNILND_PEER_IDLE;
2380 kgnilnd_peer_increase_reconnect_locked(peer);
2382 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2383 " state 0x%p->%s, connecting %d\n",
2384 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2385 peer->gnp_connecting = GNILND_PEER_KILL;
2387 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2389 /* hold onto ref until we are really done - if it was
2390 * unlinked this could result in a destroy */
2391 kgnilnd_peer_decref(peer);
2393 spin_lock(&dev->gnd_connd_lock);
2396 spin_unlock(&dev->gnd_connd_lock);
2397 RETURN(did_something);
2401 kgnilnd_repost_wc_dgrams(kgn_device_t *dev)
2403 int did_something = 0, to_repost, i;
2404 to_repost = atomic_read(&dev->gnd_nwcdgrams);
2407 for (i = 0; i < to_repost; ++i) {
2409 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
2411 kgnilnd_admin_decref(dev->gnd_nwcdgrams);
2414 CDEBUG(D_NETERROR, "error %d: dev %d could not post wildcard datagram\n",
2420 RETURN(did_something);
2424 kgnilnd_dgram_poke_with_stick(unsigned long arg)
2427 kgn_device_t *dev = &kgnilnd_data.kgn_devices[dev_id];
2429 wake_up(&dev->gnd_dgram_waitq);
2432 /* use single thread for dgrams - should be sufficient for performance */
2434 kgnilnd_dgram_mover(void *arg)
2436 kgn_device_t *dev = (kgn_device_t *)arg;
2438 int rc, did_something;
2439 unsigned long next_purge_check = jiffies - 1;
2440 unsigned long timeout;
2441 struct timer_list timer;
2442 unsigned long deadline = 0;
2445 snprintf(name, sizeof(name), "kgnilnd_dg_%02d", dev->gnd_id);
2446 cfs_block_allsigs();
2447 /* all gnilnd threads need to run fairly urgently */
2448 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2450 /* we are ok not locking for these variables as the dgram waitq threads
2451 * will block both due to tying up net (kgn_shutdown) and the completion
2452 * event for the dgram_waitq (kgn_quiesce_trigger) */
2453 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2454 while (!kgnilnd_data.kgn_shutdown) {
2455 /* Safe: kgn_shutdown only set when quiescent */
2457 /* race with stack reset - we want to hold off seeing any new incoming dgrams
2458 * so we can force a dirty WC dgram for Bug 762072 - put right before
2459 * quiesce check so that it'll go right into that and not do any
2461 CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);
2463 /* to quiesce or to not quiesce, that is the question */
2464 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2465 KGNILND_SPIN_QUIESCE;
2469 CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);
2471 /* process any newly completed dgrams */
2472 down_read(&kgnilnd_data.kgn_net_rw_sem);
2474 rc = kgnilnd_probe_and_process_dgram(dev);
2476 did_something += rc;
2479 up_read(&kgnilnd_data.kgn_net_rw_sem);
2481 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_DGRAM_DEADLINE,
2482 (*kgnilnd_tunables.kgn_dgram_timeout + 1));
2483 /* start new outbound dgrams */
2484 did_something += kgnilnd_start_outbound_dgrams(dev, deadline);
2486 /* find dead dgrams */
2487 if (time_after_eq(jiffies, next_purge_check)) {
2488 /* these don't need to be checked that often */
2489 kgnilnd_reaper_dgram_check(dev);
2491 next_purge_check = (long) jiffies +
2492 cfs_time_seconds(kgnilnd_data.kgn_new_min_timeout / 4);
2495 did_something += kgnilnd_repost_wc_dgrams(dev);
2497 /* careful with the jiffy wrap... */
2498 timeout = (long)(next_purge_check - jiffies);
2500 CDEBUG(D_INFO, "did %d timeout %lu next %lu jiffies %lu\n",
2501 did_something, timeout, next_purge_check, jiffies);
2503 if ((did_something || timeout <= 0) && time_before(jiffies, deadline)) {
2508 prepare_to_wait(&dev->gnd_dgram_waitq, &wait, TASK_INTERRUPTIBLE);
2510 setup_timer(&timer, kgnilnd_dgram_poke_with_stick, dev->gnd_id);
2511 mod_timer(&timer, (long) jiffies + timeout);
2513 /* last second chance for others to poke us */
2514 did_something += xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_IDLE);
2516 /* check flag variables before comittingi even if we did something;
2517 * if we are after the deadline call schedule */
2518 if ((!did_something || time_after(jiffies, deadline)) &&
2519 !kgnilnd_data.kgn_shutdown &&
2520 !kgnilnd_data.kgn_quiesce_trigger) {
2521 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2522 timeout, cfs_duration_sec(timeout));
2523 wake_up_all(&dev->gnd_dgping_waitq);
2525 CDEBUG(D_INFO, "awake after schedule\n");
2526 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2529 del_singleshot_timer_sync(&timer);
2530 finish_wait(&dev->gnd_dgram_waitq, &wait);
2533 kgnilnd_thread_fini();