2 * Copyright (C) 2012 Cray, Inc.
4 * Author: Nic Henke <nic@cray.com>
5 * Author: James Shimek <jshimek@cray.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 kgnilnd_setup_smsg_attr(gni_smsg_attr_t *smsg_attr)
29 smsg_attr->mbox_maxcredit = *kgnilnd_tunables.kgn_mbox_credits;
30 smsg_attr->msg_maxsize = GNILND_MAX_MSG_SIZE;
31 smsg_attr->msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
35 kgnilnd_map_fmablk(kgn_device_t *device, kgn_fma_memblock_t *fma_blk)
38 __u32 flags = GNI_MEM_READWRITE;
40 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
41 flags |= GNI_MEM_PHYS_CONT;
44 /* make sure we are mapping a clean block */
45 LASSERTF(fma_blk->gnm_hndl.qword1 == 0UL, "fma_blk %p dirty\n", fma_blk);
47 rrc = kgnilnd_mem_register(device->gnd_handle, (__u64)fma_blk->gnm_block,
48 fma_blk->gnm_blk_size, device->gnd_rcv_fma_cqh,
49 flags, &fma_blk->gnm_hndl);
50 if (rrc != GNI_RC_SUCCESS) {
51 /* XXX Nic: need a way to silence this for runtime stuff that is ok to fail
52 * -- like when under MDD or GART pressure on big systems
54 CNETERR("register fmablk failed 0x%p mbox_size %d flags %u\n",
55 fma_blk, fma_blk->gnm_mbox_size, flags);
59 /* PHYS_CONT memory isn't really mapped, at least not in GART -
60 * but all mappings chew up a MDD
62 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
63 atomic64_add(fma_blk->gnm_blk_size, &device->gnd_nbytes_map);
66 atomic_inc(&device->gnd_n_mdd);
67 /* nfmablk is live (mapped) blocks */
68 atomic_inc(&device->gnd_nfmablk);
74 kgnilnd_alloc_fmablk(kgn_device_t *device, int use_phys)
78 kgn_fma_memblock_t *fma_blk;
79 gni_smsg_attr_t smsg_attr;
80 unsigned long fmablk_vers;
82 /* we'll use fmablk_vers and the gnd_fmablk_mutex to gate access
83 * to this allocation code. Everyone will sample the version
84 * before and after getting the mutex. If it has changed,
85 * we'll bail out to check the lists again - this indicates that
86 * some sort of change was made to the lists and it is possible
87 * that there is a mailbox for us to find now. This should prevent
88 * a ton of spinning in the case where there are lots of threads
89 * that need a yet-to-be-allocated mailbox for a connection. */
91 fmablk_vers = atomic_read(&device->gnd_fmablk_vers);
92 mutex_lock(&device->gnd_fmablk_mutex);
94 if (fmablk_vers != atomic_read(&device->gnd_fmablk_vers)) {
95 /* version changed while we were waiting for semaphore,
96 * we'll recheck the lists assuming something nice happened */
97 mutex_unlock(&device->gnd_fmablk_mutex);
101 LIBCFS_ALLOC(fma_blk, sizeof(kgn_fma_memblock_t));
102 if (fma_blk == NULL) {
103 CNETERR("could not allocate fma block descriptor\n");
108 INIT_LIST_HEAD(&fma_blk->gnm_bufflist);
110 kgnilnd_setup_smsg_attr(&smsg_attr);
112 gni_smsg_buff_size_needed(&smsg_attr, &fma_blk->gnm_mbox_size);
114 LASSERTF(fma_blk->gnm_mbox_size, "mbox size %d\n", fma_blk->gnm_mbox_size);
116 /* gni_smsg_buff_size_needed calculates the base mailbox size and since
117 * we want to hold kgn_peer_credits worth of messages in both directions,
118 * we add PAYLOAD to grow the mailbox size
121 fma_blk->gnm_mbox_size += GNILND_MBOX_PAYLOAD;
123 /* we'll only use physical during preallocate at startup -- this keeps it nice and
124 * clean for runtime decisions. We'll keep the PHYS ones around until shutdown
125 * as reallocating them is tough if there is memory fragmentation */
128 fma_blk->gnm_block = kmem_cache_alloc(kgnilnd_data.kgn_mbox_cache, GFP_ATOMIC);
129 if (fma_blk->gnm_block == NULL) {
130 CNETERR("could not allocate physical SMSG mailbox memory\n");
134 fma_blk->gnm_blk_size = KMALLOC_MAX_SIZE;
135 num_mbox = fma_blk->gnm_blk_size / fma_blk->gnm_mbox_size;
137 LASSERTF(num_mbox >= 1,
138 "num_mbox %d blk_size %u mbox_size %d\n",
139 num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size);
141 fma_blk->gnm_state = GNILND_FMABLK_PHYS;
144 num_mbox = *kgnilnd_tunables.kgn_mbox_per_block;
145 fma_blk->gnm_blk_size = num_mbox * fma_blk->gnm_mbox_size;
147 LASSERTF(num_mbox >= 1 && num_mbox >= *kgnilnd_tunables.kgn_mbox_per_block,
148 "num_mbox %d blk_size %u mbox_size %d tunable %d\n",
149 num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size,
150 *kgnilnd_tunables.kgn_mbox_per_block);
152 LIBCFS_ALLOC(fma_blk->gnm_block, fma_blk->gnm_blk_size);
153 if (fma_blk->gnm_block == NULL) {
154 CNETERR("could not allocate virtual SMSG mailbox memory, %d bytes\n", fma_blk->gnm_blk_size);
159 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
162 /* allocate just enough space for the bits to track the mailboxes */
163 LIBCFS_ALLOC(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox) * sizeof(unsigned long));
164 if (fma_blk->gnm_bit_array == NULL) {
165 CNETERR("could not allocate mailbox bitmask, %lu bytes for %d mbox\n",
166 sizeof(unsigned long) * BITS_TO_LONGS(num_mbox), num_mbox);
170 bitmap_zero(fma_blk->gnm_bit_array, num_mbox);
172 /* now that the num_mbox is set based on allocation type, get debug info setup */
173 LIBCFS_ALLOC(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t) * num_mbox);
174 if (fma_blk->gnm_mbox_info == NULL) {
175 CNETERR("could not allocate mailbox debug, %lu bytes for %d mbox\n",
176 sizeof(kgn_mbox_info_t) * num_mbox, num_mbox);
181 rc = kgnilnd_map_fmablk(device, fma_blk);
186 fma_blk->gnm_next_avail_mbox = 0;
187 fma_blk->gnm_avail_mboxs = fma_blk->gnm_num_mboxs = num_mbox;
189 CDEBUG(D_MALLOC, "alloc fmablk 0x%p num %d msg_maxsize %d credits %d "
190 "mbox_size %d MDD "LPX64"."LPX64"\n",
191 fma_blk, num_mbox, smsg_attr.msg_maxsize, smsg_attr.mbox_maxcredit,
192 fma_blk->gnm_mbox_size, fma_blk->gnm_hndl.qword1,
193 fma_blk->gnm_hndl.qword2);
195 /* lock Is protecting data structures, not semaphore */
197 spin_lock(&device->gnd_fmablk_lock);
198 list_add_tail(&fma_blk->gnm_bufflist, &device->gnd_fma_buffs);
200 /* toggle under the lock so once they change the list is also
201 * ready for others to traverse */
202 atomic_inc(&device->gnd_fmablk_vers);
204 spin_unlock(&device->gnd_fmablk_lock);
206 mutex_unlock(&device->gnd_fmablk_mutex);
211 LIBCFS_FREE(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t)*num_mbox);
213 LIBCFS_FREE(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox) * sizeof (unsigned long));
215 if (fma_blk->gnm_state == GNILND_FMABLK_VIRT) {
216 LIBCFS_FREE(fma_blk->gnm_block, fma_blk->gnm_blk_size);
218 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
221 LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
223 mutex_unlock(&device->gnd_fmablk_mutex);
228 kgnilnd_unmap_fmablk(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
232 /* if some held, set hold_timeout from conn timeouts used in this block
233 * but not during shutdown, then just nuke and pave */
234 if (fma_blk->gnm_held_mboxs && (!kgnilnd_data.kgn_shutdown)) {
235 fma_blk->gnm_hold_timeout = GNILND_TIMEOUT2DEADMAN;
238 /* we are changing the state of a block, tickle version to tell
239 * proc code list is stale now */
240 atomic_inc(&dev->gnd_fmablk_vers);
242 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &fma_blk->gnm_hndl, fma_blk->gnm_hold_timeout);
244 CDEBUG(rrc == GNI_RC_SUCCESS ? D_MALLOC : D_CONSOLE|D_NETERROR,
245 "unmap fmablk 0x%p@%s sz %u total %d avail %d held %d mbox_size %d "
247 fma_blk, kgnilnd_fmablk_state2str(fma_blk->gnm_state),
248 fma_blk->gnm_blk_size, fma_blk->gnm_num_mboxs,
249 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs,
250 fma_blk->gnm_mbox_size, fma_blk->gnm_hold_timeout);
252 LASSERTF(rrc == GNI_RC_SUCCESS,
253 "tried to double unmap or something bad, fma_blk %p (rrc %d)\n",
256 if (fma_blk->gnm_hold_timeout) {
257 atomic_inc(&dev->gnd_n_mdd_held);
259 atomic_dec(&dev->gnd_n_mdd);
262 /* PHYS blocks don't get mapped */
263 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
264 atomic64_sub(fma_blk->gnm_blk_size, &dev->gnd_nbytes_map);
265 fma_blk->gnm_state = GNILND_FMABLK_IDLE;
266 } else if (kgnilnd_data.kgn_in_reset) {
267 /* in stack reset, clear MDD handle for PHYS blocks, as we'll
268 * re-use the fma_blk after reset so we don't have to drop/allocate
269 * all of those physical blocks */
270 fma_blk->gnm_hndl.qword1 = fma_blk->gnm_hndl.qword2 = 0UL;
273 /* Decrement here as this is the # of mapped blocks */
274 atomic_dec(&dev->gnd_nfmablk);
278 /* needs lock on gnd_fmablk_lock to cover gnd_fma_buffs */
280 kgnilnd_free_fmablk_locked(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
282 LASSERTF(fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs,
283 "fma_blk %p@%d free in bad state (%d): blk total %d avail %d held %d\n",
284 fma_blk, fma_blk->gnm_state, fma_blk->gnm_hold_timeout, fma_blk->gnm_num_mboxs,
285 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs);
287 atomic_inc(&dev->gnd_fmablk_vers);
289 if (fma_blk->gnm_hold_timeout) {
290 CDEBUG(D_MALLOC, "mdd release fmablk 0x%p sz %u avail %d held %d "
292 fma_blk, fma_blk->gnm_blk_size, fma_blk->gnm_avail_mboxs,
293 fma_blk->gnm_held_mboxs, fma_blk->gnm_mbox_size);
295 /* We leave MDD dangling over stack reset */
296 if (!kgnilnd_data.kgn_in_reset) {
297 kgnilnd_mem_mdd_release(dev->gnd_handle, &fma_blk->gnm_hndl);
299 /* ignoring the return code - if kgni/ghal can't find it
300 * it must be released already */
301 atomic_dec(&dev->gnd_n_mdd_held);
302 atomic_dec(&dev->gnd_n_mdd);
305 /* we cant' free the gnm_block until all the conns have released their
306 * purgatory holds. While we have purgatory holds, we might check the conn
307 * RX mailbox during the CLOSING process. It is possible that kgni might
308 * try to look into the RX side for credits when sending the CLOSE msg too */
309 CDEBUG(D_MALLOC, "fmablk %p free buffer %p mbox_size %d\n",
310 fma_blk, fma_blk->gnm_block, fma_blk->gnm_mbox_size);
312 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
313 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
315 LIBCFS_FREE(fma_blk->gnm_block, fma_blk->gnm_blk_size);
317 fma_blk->gnm_state = GNILND_FMABLK_FREED;
319 list_del(&fma_blk->gnm_bufflist);
321 LIBCFS_FREE(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t)*fma_blk->gnm_num_mboxs);
322 LIBCFS_FREE(fma_blk->gnm_bit_array, BITS_TO_LONGS(fma_blk->gnm_num_mboxs) * sizeof (unsigned long));
323 LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
327 kgnilnd_find_free_mbox(kgn_conn_t *conn)
329 kgn_device_t *dev = conn->gnc_device;
330 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
331 kgn_fma_memblock_t *fma_blk;
332 kgn_mbox_info_t *mbox = NULL;
335 spin_lock(&dev->gnd_fmablk_lock);
337 list_for_each_entry(fma_blk, &conn->gnc_device->gnd_fma_buffs,
339 if (fma_blk->gnm_avail_mboxs <= 0 ||
340 fma_blk->gnm_state <= GNILND_FMABLK_IDLE) {
343 /* look in bitarray for available mailbox */
345 id = find_next_zero_bit(
346 fma_blk->gnm_bit_array,
347 fma_blk->gnm_num_mboxs,
348 fma_blk->gnm_next_avail_mbox);
349 if (id == fma_blk->gnm_num_mboxs &&
350 fma_blk->gnm_next_avail_mbox != 0) {
352 fma_blk->gnm_next_avail_mbox = 0;
358 LASSERTF(id < fma_blk->gnm_num_mboxs, "id %d max %d\n",
359 id, fma_blk->gnm_num_mboxs);
360 set_bit(id, (volatile unsigned long *)fma_blk->gnm_bit_array);
361 conn->gnc_mbox_id = id;
363 fma_blk->gnm_next_avail_mbox =
364 (id == (fma_blk->gnm_num_mboxs - 1)) ? 0 : (id + 1);
365 fma_blk->gnm_avail_mboxs--;
366 conn->gnc_fma_blk = fma_blk;
368 kgnilnd_setup_smsg_attr(smsg_attr);
370 smsg_attr->msg_buffer = fma_blk->gnm_block;
371 smsg_attr->mbox_offset = fma_blk->gnm_mbox_size * id;
372 smsg_attr->mem_hndl = fma_blk->gnm_hndl;
373 smsg_attr->buff_size = fma_blk->gnm_mbox_size;
375 /* We'll set the hndl to zero for PHYS blocks unmapped during stack
376 * reset and re-use the same fma_blk after stack reset. This ensures we've
377 * properly mapped it before we use it */
378 LASSERTF(fma_blk->gnm_hndl.qword1 != 0UL, "unmapped fma_blk %p, state %d\n",
379 fma_blk, fma_blk->gnm_state);
381 CDEBUG(D_NET, "conn %p smsg %p fmablk %p "
382 "allocating SMSG mbox %d buf %p "
383 "offset %u hndl "LPX64"."LPX64"\n",
384 conn, smsg_attr, fma_blk, id,
385 smsg_attr->msg_buffer, smsg_attr->mbox_offset,
386 fma_blk->gnm_hndl.qword1,
387 fma_blk->gnm_hndl.qword2);
389 mbox = &fma_blk->gnm_mbox_info[id];
390 mbox->mbx_create_conn_memset = jiffies;
392 mbox->mbx_nallocs_total++;
394 /* zero mbox to remove any old data from our last use.
395 * this better be safe, if not our purgatory timers
396 * are too short or a peer really is misbehaving */
397 memset(smsg_attr->msg_buffer + smsg_attr->mbox_offset,
398 0, smsg_attr->buff_size);
402 spin_unlock(&dev->gnd_fmablk_lock);
406 kgnilnd_setup_mbox(kgn_conn_t *conn)
408 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
411 smsg_attr->msg_buffer = NULL;
412 /* Look for available mbox */
414 kgnilnd_find_free_mbox(conn);
416 /* nothing in the existing buffers, make a new one */
417 if (smsg_attr->msg_buffer == NULL) {
418 /* for runtime allocations, we only want vmalloc */
419 err = kgnilnd_alloc_fmablk(conn->gnc_device, 0);
424 } while (smsg_attr->msg_buffer == NULL);
427 CNETERR("couldn't allocate SMSG mbox for conn %p Error: %d\n",
433 kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold)
435 kgn_device_t *dev = conn->gnc_device;
436 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
437 kgn_fma_memblock_t *fma_blk = NULL;
438 kgn_mbox_info_t *mbox = NULL;
442 /* if we failed to setup mbox and now destroying conn */
443 if (smsg_attr->msg_buffer == NULL) {
447 id = conn->gnc_mbox_id;
449 spin_lock(&dev->gnd_fmablk_lock);
450 /* make sure our conn points at a valid fma_blk
451 * We use this instead of a mem block search out of smsg_attr
452 * because we could have freed a block for fma_blk #1 but the fma_blk
453 * is still in the list for a purgatory hold. This would induce a false
454 * match if that same block gets reallocated to fma_blk #2 */
455 list_for_each_entry(fma_blk, &dev->gnd_fma_buffs, gnm_bufflist) {
456 if (fma_blk == conn->gnc_fma_blk) {
461 LASSERTF(found, "unable to find conn 0x%p with gnc_fma_blk %p "
462 "anywhere in the world\n", conn, conn->gnc_fma_blk);
464 LASSERTF(id < fma_blk->gnm_num_mboxs,
465 "bad id %d max %d\n",
466 id, fma_blk->gnm_num_mboxs);
468 /* < 0 - was held, now free it
469 * == 0 - just free it
470 * > 0 - hold it for now */
471 if (purgatory_hold == 0) {
472 CDEBUG(D_NET, "conn %p smsg %p fmablk %p freeing SMSG mbox %d "
473 "hndl "LPX64"."LPX64"\n",
474 conn, smsg_attr, fma_blk, id,
475 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
476 fma_blk->gnm_avail_mboxs++;
478 } else if (purgatory_hold > 0) {
479 CDEBUG(D_NET, "conn %p smsg %p fmablk %p holding SMSG mbox %d "
480 "hndl "LPX64"."LPX64"\n",
481 conn, smsg_attr, fma_blk, id,
482 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
484 fma_blk->gnm_held_mboxs++;
485 fma_blk->gnm_max_timeout = MAX(fma_blk->gnm_max_timeout,
488 CDEBUG(D_NET, "conn %p smsg %p fmablk %p release SMSG mbox %d "
489 "hndl "LPX64"."LPX64"\n",
490 conn, smsg_attr, fma_blk, id,
491 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
493 fma_blk->gnm_held_mboxs--;
494 fma_blk->gnm_avail_mboxs++;
497 if (purgatory_hold <= 0) {
498 /* if kgni is retransmitting, freeing the smsg block before the EP
499 * is destroyed gets messy. Bug 768295. */
500 LASSERTF(conn->gnc_ephandle == NULL,
501 "can't release mbox before EP is nuked. conn 0x%p\n", conn);
503 mbox = &fma_blk->gnm_mbox_info[id];
504 mbox->mbx_release_from_purgatory = jiffies;
506 /* clear conn gnc_fmablk if it is gone - this allows us to
507 * not worry about state so much in kgnilnd_destroy_conn
508 * and makes the guaranteed cleanup of the resources easier */
509 LASSERTF(test_and_clear_bit(id, fma_blk->gnm_bit_array),
510 "conn %p bit %d already cleared in fma_blk %p\n",
512 conn->gnc_fma_blk = NULL;
516 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FMABLK_AVAIL)) {
517 CERROR("LBUGs in your future: forcibly marking fma_blk %p "
518 "as mapped\n", fma_blk);
519 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
522 /* we don't release or unmap PHYS blocks as part of the normal cycle --
523 * those are controlled manually from startup/shutdown */
524 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
525 /* we can unmap once all are unused (held or avail)
526 * but check hold_timeout to make sure we are not trying to double
527 * unmap this buffer. If there was no hold_timeout set due to
528 * held_mboxs, we'll free the mobx here shortly and won't have to
529 * worry about catching a double free for a 'clean' fma_blk */
530 if (((fma_blk->gnm_avail_mboxs + fma_blk->gnm_held_mboxs) == fma_blk->gnm_num_mboxs) &&
531 (!fma_blk->gnm_hold_timeout)) {
532 kgnilnd_unmap_fmablk(dev, fma_blk);
535 /* But we can only free once they are all avail */
536 if (fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs &&
537 fma_blk->gnm_held_mboxs == 0) {
538 /* all mailboxes are released, free fma_blk */
539 kgnilnd_free_fmablk_locked(dev, fma_blk);
543 spin_unlock(&dev->gnd_fmablk_lock);
547 kgnilnd_count_phys_mbox(kgn_device_t *device)
550 kgn_fma_memblock_t *fma_blk;
552 spin_lock(&device->gnd_fmablk_lock);
554 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
555 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
556 i += fma_blk->gnm_num_mboxs;
558 spin_unlock(&device->gnd_fmablk_lock);
564 kgnilnd_allocate_phys_fmablk(kgn_device_t *device)
568 while (kgnilnd_count_phys_mbox(device) < *kgnilnd_tunables.kgn_nphys_mbox) {
570 rc = kgnilnd_alloc_fmablk(device, 1);
572 CERROR("failed phys mbox allocation, stopping at %d, rc %d\n",
573 kgnilnd_count_phys_mbox(device), rc);
581 kgnilnd_map_phys_fmablk(kgn_device_t *device)
585 kgn_fma_memblock_t *fma_blk;
587 /* use mutex to gate access to single thread, just in case */
588 mutex_lock(&device->gnd_fmablk_mutex);
590 spin_lock(&device->gnd_fmablk_lock);
592 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
593 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
594 rc = kgnilnd_map_fmablk(device, fma_blk);
599 spin_unlock(&device->gnd_fmablk_lock);
601 mutex_unlock(&device->gnd_fmablk_mutex);
607 kgnilnd_unmap_fma_blocks(kgn_device_t *device)
610 kgn_fma_memblock_t *fma_blk;
612 /* use mutex to gate access to single thread, just in case */
613 mutex_lock(&device->gnd_fmablk_mutex);
615 spin_lock(&device->gnd_fmablk_lock);
617 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
618 kgnilnd_unmap_fmablk(device, fma_blk);
620 spin_unlock(&device->gnd_fmablk_lock);
622 mutex_unlock(&device->gnd_fmablk_mutex);
626 kgnilnd_free_phys_fmablk(kgn_device_t *device)
629 kgn_fma_memblock_t *fma_blk, *fma_blkN;
631 /* use mutex to gate access to single thread, just in case */
632 mutex_lock(&device->gnd_fmablk_mutex);
634 spin_lock(&device->gnd_fmablk_lock);
636 list_for_each_entry_safe(fma_blk, fma_blkN, &device->gnd_fma_buffs, gnm_bufflist) {
637 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
638 kgnilnd_free_fmablk_locked(device, fma_blk);
640 spin_unlock(&device->gnd_fmablk_lock);
642 mutex_unlock(&device->gnd_fmablk_mutex);
645 /* kgnilnd dgram nid->struct managment */
647 static inline struct list_head *
648 kgnilnd_nid2dgramlist(kgn_device_t *dev, lnet_nid_t nid)
650 unsigned int hash = ((unsigned int)nid) % *kgnilnd_tunables.kgn_peer_hash_size;
652 RETURN(&dev->gnd_dgrams[hash]);
656 /* needs dev->gnd_dgram_lock held */
658 kgnilnd_find_dgram_locked(kgn_device_t *dev, lnet_nid_t dst_nid)
660 struct list_head *dgram_list = kgnilnd_nid2dgramlist(dev, dst_nid);
663 list_for_each_entry(dgram, dgram_list, gndg_list) {
665 /* if state > POSTED, we are already handling cancel/completion */
666 if ((dgram->gndg_conn_out.gncr_dstnid != dst_nid) ||
667 dgram->gndg_state > GNILND_DGRAM_POSTED)
670 CDEBUG(D_NET, "got dgram [%p] -> %s\n",
671 dgram, libcfs_nid2str(dst_nid));
678 kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid)
682 spin_lock(&dev->gnd_dgram_lock);
683 dgram = kgnilnd_find_dgram_locked(dev, dst_nid);
686 kgnilnd_cancel_dgram_locked(dgram);
688 spin_unlock(&dev->gnd_dgram_lock);
690 RETURN(!!(dgram == NULL));
694 kgnilnd_pack_connreq(kgn_connreq_t *connreq, kgn_conn_t *conn,
695 lnet_nid_t srcnid, lnet_nid_t dstnid,
696 kgn_connreq_type_t type)
700 /* ensure we haven't violated max datagram size */
701 CLASSERT(sizeof(kgn_connreq_t) <= GNI_DATAGRAM_MAXSIZE);
703 /* no need to zero out, we do that when allocating dgram */
704 connreq->gncr_magic = GNILND_MSG_MAGIC;
706 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_SRCNID)) {
708 } else if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
712 connreq->gncr_srcnid = srcnid;
713 connreq->gncr_dstnid = dstnid;
715 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
716 connreq->gncr_version = 99;
718 connreq->gncr_version = GNILND_CONNREQ_VERSION;
720 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
721 connreq->gncr_type = 99;
723 connreq->gncr_type = type;
725 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
726 connreq->gncr_peerstamp = 0;
728 connreq->gncr_peerstamp = kgnilnd_data.kgn_peerstamp;
730 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
731 connreq->gncr_connstamp = 0;
733 connreq->gncr_connstamp = conn->gnc_my_connstamp;
735 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
736 connreq->gncr_timeout = 0;
738 connreq->gncr_timeout = conn->gnc_timeout;
741 /* the rest pack the data into the payload in other places */
742 if (type == GNILND_CONNREQ_REQ) {
743 kgn_gniparams_t *req_params = &connreq->gncr_gnparams;
744 req_params->gnpr_host_id = conn->gnc_device->gnd_host_id;
745 req_params->gnpr_cqid = conn->gnc_cqid;
747 /* allocate mailbox for this connection */
748 err = kgnilnd_setup_mbox(conn);
750 CERROR("Failed to setup FMA mailbox (%d)\n", err);
752 req_params->gnpr_smsg_attr = conn->gnpr_smsg_attr;
755 /* XXX Nic: TBD - checksum computation */
761 kgnilnd_unpack_connreq(kgn_dgram_t *dgram)
763 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
767 /* the following fields must be handled in a backwards compatible
768 * manner to ensure we can always send and interpret NAKs */
770 if (connreq->gncr_magic != GNILND_MSG_MAGIC &&
771 connreq->gncr_magic != __swab32(GNILND_MSG_MAGIC)) {
772 /* Unexpected magic! */
773 CERROR("Unexpected magic %08x\n",
774 connreq->gncr_magic);
778 swab = (connreq->gncr_magic == __swab32(GNILND_MSG_MAGIC));
780 __swab32s(&connreq->gncr_magic);
781 __swab32s(&connreq->gncr_cksum);
782 __swab16s(&connreq->gncr_type);
783 __swab16s(&connreq->gncr_version);
784 __swab32s(&connreq->gncr_timeout);
785 __swab64s(&connreq->gncr_srcnid);
786 __swab64s(&connreq->gncr_dstnid);
787 __swab64s(&connreq->gncr_peerstamp);
788 __swab64s(&connreq->gncr_connstamp);
791 /* Do NOT return anything but -EBADF before we munge
792 * connreq->gncr_srcnid - we need that to send the nak */
794 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
795 lnet_nid_t incoming = connreq->gncr_srcnid;
797 /* even if the incoming packet is hosed, we know who we sent
798 * the original and can set the srcnid so that we can properly
799 * look up our peer to close the loop on this connreq. We still use
800 * -EBADF to prevent a NAK - just in case there are issues with
801 * the payload coming from a random spot, etc. */
802 connreq->gncr_srcnid = dgram->gndg_conn_out.gncr_dstnid;
804 if (LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid) !=
805 LNET_NIDADDR(incoming)) {
806 /* we got a datagram match for the wrong nid... */
807 CERROR("matched datagram 0x%p with srcnid %s "
808 "(%x), expecting %s (%x)\n",
810 libcfs_nid2str(incoming),
811 LNET_NIDADDR(incoming),
812 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
813 LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid));
817 /* if we have a wildcard datagram it should match an
818 * incoming "active" datagram that should have a fully formed
819 * srcnid and dstnid. If we couldn't unpack it, we drop as
820 * corrupted packet, otherwise we'll just verify that the dstnid
821 * matches the NID for the NET that the dgram was posted */
823 /* make sure their wildcard didn't match ours, that is unpossible */
824 LASSERTF(connreq->gncr_dstnid != LNET_NID_ANY,
825 "dgram 0x%p from %s, connreq 0x%p; "
826 "wildcard matched wildcard \n", dgram,
827 libcfs_nid2str(connreq->gncr_srcnid), connreq);
829 rc = kgnilnd_find_net(connreq->gncr_dstnid, &net);
831 if (rc == -ESHUTDOWN) {
832 CERROR("Looking up network: device is in shutdown");
834 } else if (rc == -ENONET) {
835 CERROR("Connection data from %s: she sent "
836 "dst_nid %s, but net lookup failed on "
838 libcfs_nid2str(connreq->gncr_srcnid),
839 libcfs_nid2str(connreq->gncr_dstnid),
840 dgram, kgnilnd_dgram_type2str(dgram));
844 if (net->gnn_ni->ni_nid != connreq->gncr_dstnid) {
845 CERROR("Bad connection data from %s: she sent "
846 "dst_nid %s, but I am %s with dgram 0x%p@%s\n",
847 libcfs_nid2str(connreq->gncr_srcnid),
848 libcfs_nid2str(connreq->gncr_dstnid),
849 libcfs_nid2str(net->gnn_ni->ni_nid),
850 dgram, kgnilnd_dgram_type2str(dgram));
851 kgnilnd_net_decref(net);
855 /* kgnilnd_find_net takes a ref on the net it finds, You need to decref it when not needed. */
856 kgnilnd_net_decref(net);
859 if (connreq->gncr_version != GNILND_CONNREQ_VERSION) {
860 CERROR("Unexpected version %d\n", connreq->gncr_version);
864 /* XXX Nic: TBD - checksum validation */
865 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_DROP)) {
869 if (swab && connreq->gncr_type == GNILND_CONNREQ_REQ) {
870 __u64 msg_addr = (__u64) connreq->gncr_gnparams.gnpr_smsg_attr.msg_buffer;
872 __swab32s(&connreq->gncr_gnparams.gnpr_host_id);
873 __swab32s(&connreq->gncr_gnparams.gnpr_cqid);
874 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.buff_size);
875 __swab16s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_maxcredit);
876 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_offset);
877 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword1);
878 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword2);
879 __swab64s(&msg_addr);
880 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_maxsize);
881 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_type);
882 } else if (swab && connreq->gncr_type == GNILND_CONNREQ_NAK) {
883 __swab32s(&connreq->gncr_nakdata.gnnd_errno);
886 /* since we use a unique instance ID for each network, the driver
887 * will take care of dropping datagrams if we don't have that network.
890 /* few more idiot software or configuration checks */
892 switch (connreq->gncr_type) {
893 case GNILND_CONNREQ_REQ:
894 /* wire up EP and SMSG block - this will check the incoming data
895 * and barf a NAK back if need to */
896 rc = kgnilnd_set_conn_params(dgram);
900 case GNILND_CONNREQ_NAK:
901 case GNILND_CONNREQ_CLOSE:
904 CERROR("unknown connreq packet type %d\n", connreq->gncr_type);
908 if (connreq->gncr_peerstamp == 0 || connreq->gncr_connstamp == 0) {
909 CERROR("Recived bad timestamps peer "LPU64" conn "LPU64"\n",
910 connreq->gncr_peerstamp, connreq->gncr_connstamp);
914 if (connreq->gncr_timeout < GNILND_MIN_TIMEOUT) {
915 CERROR("Received timeout %d < MIN %d\n",
916 connreq->gncr_timeout, GNILND_MIN_TIMEOUT);
924 kgnilnd_alloc_dgram(kgn_dgram_t **dgramp, kgn_device_t *dev, kgn_dgram_type_t type)
928 dgram = kmem_cache_alloc(kgnilnd_data.kgn_dgram_cache, GFP_ATOMIC);
932 /* cache alloc'd memory is not zeroed */
933 memset((void *)dgram, 0, sizeof(*dgram)) ;
935 INIT_LIST_HEAD(&dgram->gndg_list);
936 dgram->gndg_state = GNILND_DGRAM_USED;
937 dgram->gndg_type = type;
938 dgram->gndg_magic = GNILND_DGRAM_MAGIC;
940 atomic_inc(&dev->gnd_ndgrams);
942 CDEBUG(D_MALLOC|D_NETTRACE, "slab-alloced 'dgram': %lu at %p %s ndgrams"
944 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
945 atomic_read(&dev->gnd_ndgrams));
951 /* call this on a dgram that came back from kgnilnd_ep_postdata_test_by_id
952 * returns < 0 on dgram to be cleaned up
953 * > 0 on dgram that isn't done yet
954 * == 0 on dgram that is ok and needs connreq processing */
956 kgnilnd_process_dgram(kgn_dgram_t *dgram, gni_post_state_t post_state)
960 switch (post_state) {
961 case GNI_POST_COMPLETED:
962 /* normal state for dgrams that need actual processing */
963 /* GOTO to avoid processing dgram as canceled/done */
964 GOTO(process_out, rc);
966 case GNI_POST_PENDING:
967 /* we should only see this if we are testing a WC dgram after a
968 * cancel - it means that it needs a full cycle of waiting
969 * for kgni_sm_task to finish moving it to TERMINATED */
970 LASSERTF((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
971 (dgram->gndg_state == GNILND_DGRAM_CANCELED),
972 "POST_PENDING dgram 0x%p with bad type %d(%s) or state %d(%s)\n",
973 dgram, dgram->gndg_type, kgnilnd_dgram_type2str(dgram),
974 dgram->gndg_state, kgnilnd_dgram_state2str(dgram));
976 /* positive RC as this dgram isn't done yet */
979 /* GOTO as this isn't done yet */
980 GOTO(process_out, rc);
983 case GNI_POST_TERMINATED:
984 /* we've called cancel and it is done or remote guy called cancel and
985 * we've receved it on a WC dgram */
987 /* we are seeing weird terminations on non WC dgrams when we have not
990 LASSERTF(dgram->gndg_state == GNILND_DGRAM_CANCELED ||
991 dgram->gndg_conn_out.gncr_dstnid == LNET_NID_ANY,
992 "dgram 0x%p with bad state %d(%s) or dst nid %s\n",
993 dgram, dgram->gndg_state, kgnilnd_dgram_state2str(dgram),
994 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
997 CDEBUG(D_NETTRACE, "dgram 0x%p saw %s, cleaning it up\n", dgram,
998 dgram->gndg_state == GNILND_DGRAM_CANCELED ? "canceled" : "terminated");
1003 case GNI_POST_TIMEOUT:
1004 /* we could have a timeout on a wildcard dgram too - if
1005 * we got the incoming request but the remote node beefed
1006 * before kgni could send the match data back. We'll just error
1007 * on the active case and bail out gracefully */
1008 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1009 CNETERR("hardware timeout for connect to "
1010 "%s after %lu seconds. Is node dead?\n",
1011 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1012 cfs_duration_sec(jiffies - dgram->gndg_post_time));
1019 CERROR("dgram 0x%p with bad post_state %d\n", dgram, post_state);
1023 /* now finish cleaning up a dgram that is canceled/terminated and needs to
1026 /* If this was actively canceled, drop the count now that we are processing */
1027 if (dgram->gndg_state == GNILND_DGRAM_CANCELED) {
1028 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1029 /* caller responsible for gndg_list removal */
1037 /* needs dev->gnd_dgram_lock held */
1039 kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram)
1043 if (dgram->gndg_state != GNILND_DGRAM_POSTED) {
1047 LASSERTF(dgram->gndg_conn != NULL,
1048 "dgram 0x%p with NULL conn\n", dgram);
1050 /* C.E - WC dgrams could be canceled immediately but
1051 * if there was some match pending, we need to call
1052 * test_by_id to clear it out. If that test returns
1053 * POST_PENDING, it is half done and needs to go along
1054 * with the rest of dgrams and go through a kgni_sm_task cycle
1055 * and deliver a GNI_POST_TERMINATED event before they
1056 * are actually canceled */
1058 dgram->gndg_state = GNILND_DGRAM_CANCELED;
1060 if (dgram->gndg_conn->gnc_state >= GNILND_CONN_ESTABLISHED) {
1061 /* we don't need to cancel_by_id if the datagram was good */
1065 /* let folks know there are outstanding cancels */
1066 atomic_inc(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1067 /* leave on nid list until cancel is done for debugging fun */
1068 grc = kgnilnd_ep_postdata_cancel_by_id(dgram->gndg_conn->gnc_ephandle, (__u64) dgram);
1070 /* if we don't get success here, we have hosed up the dgram tracking
1071 * code and need to bail out */
1072 LASSERTF(grc == GNI_RC_SUCCESS,
1073 "postdata_cancel returned %d for conn 0x%p to %s\n",
1074 grc, dgram->gndg_conn,
1075 dgram->gndg_conn->gnc_peer ?
1076 libcfs_nid2str(dgram->gndg_conn->gnc_peer->gnp_nid)
1080 "canceled dgram 0x%p conn 0x%p ephandle 0x%p\n",
1081 dgram, dgram->gndg_conn,
1082 dgram->gndg_conn->gnc_ephandle);
1084 if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1085 gni_post_state_t post_state;
1087 __u32 remote_addr = 0, remote_id = 0;
1089 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1090 (__u64)dgram, &post_state,
1091 &remote_addr, &remote_id);
1093 LASSERTF(grc == GNI_RC_NO_MATCH || grc == GNI_RC_SUCCESS,
1094 "bad grc %d from test_by_id on dgram 0x%p\n",
1097 /* if WC was canceled immediately, we get NO_MATCH, if needs to go
1098 * through full cycle, we get SUCCESS and need to parse post_state */
1100 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1101 "remote_addr %u remote_id %u\n", grc, dgram,
1102 kgnilnd_dgram_type2str(dgram),
1103 post_state, remote_addr, remote_id);
1105 if (grc == GNI_RC_NO_MATCH) {
1106 /* she's gone, reduce count and move along */
1107 dgram->gndg_state = GNILND_DGRAM_DONE;
1108 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1112 rc = kgnilnd_process_dgram(dgram, post_state);
1115 /* if for some weird reason we get a valid dgram back, just mark as done
1116 * so we can drop it and move along.
1117 * C.E - if it was completed, we'll just release the conn/mbox
1118 * back into the pool and it'll get reused. That said, we should only
1119 * be canceling a WC dgram on stack rest or shutdown, so that is moot */
1120 dgram->gndg_state = GNILND_DGRAM_DONE;
1121 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1123 /* caller context responsible for calling kgnilnd_release_dgram() */
1125 /* still pending, let it simmer until golden brown and delicious */
1129 /* for non WC dgrams, they are still on the nid list but marked canceled waiting
1130 * for kgni to return their ID to us via probe - that is when we'll complete their
1131 * cancel processing */
1135 kgnilnd_cleanup_dgram(kgn_dgram_t *dgram)
1137 /* release the dgram ref on conn */
1138 if (dgram->gndg_conn) {
1139 kgnilnd_conn_decref(dgram->gndg_conn);
1140 dgram->gndg_conn = NULL;
1145 kgnilnd_free_dgram(kgn_device_t *dev, kgn_dgram_t *dgram)
1147 LASSERTF(dgram->gndg_state == GNILND_DGRAM_USED ||
1148 dgram->gndg_state == GNILND_DGRAM_DONE,
1149 "dgram 0x%p with bad state %s\n",
1150 dgram, kgnilnd_dgram_state2str(dgram));
1152 /* bit of poisoning to help detect bad driver data */
1153 dgram->gndg_magic = 0x6f5a6b5f;
1154 atomic_dec(&dev->gnd_ndgrams);
1156 kmem_cache_free(kgnilnd_data.kgn_dgram_cache, dgram);
1157 CDEBUG(D_MALLOC|D_NETTRACE, "slab-freed 'dgram': %lu at %p %s"
1159 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
1160 atomic_read(&dev->gnd_ndgrams));
1164 kgnilnd_post_dgram(kgn_device_t *dev, lnet_nid_t dstnid, kgn_connreq_type_t type,
1168 kgn_dgram_t *dgram = NULL;
1169 kgn_dgram_t *tmpdgram;
1170 kgn_dgram_type_t dgtype;
1176 case GNILND_CONNREQ_REQ:
1177 if (dstnid == LNET_NID_ANY)
1178 dgtype = GNILND_DGRAM_WC_REQ;
1180 dgtype = GNILND_DGRAM_REQ;
1182 case GNILND_CONNREQ_NAK:
1183 LASSERTF(dstnid != LNET_NID_ANY, "can't NAK to LNET_NID_ANY\n");
1184 dgtype = GNILND_DGRAM_NAK;
1187 CERROR("unknown connreq type %d\n", type);
1191 rc = kgnilnd_alloc_dgram(&dgram, dev, dgtype);
1194 GOTO(post_failed, rc);
1197 rc = kgnilnd_create_conn(&dgram->gndg_conn, dev);
1199 GOTO(post_failed, rc);
1202 if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1203 /* clear buffer for sanity on reuse of wildcard */
1204 memset(&dgram->gndg_conn_in, 0, sizeof(kgn_connreq_t));
1207 if (dstnid == LNET_NID_ANY) {
1208 /* set here to reset any dgram re-use */
1209 dgram->gndg_conn->gnc_state = GNILND_CONN_LISTEN;
1213 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(dstnid), 1, &host_id);
1216 GOTO(post_failed, rc);
1219 dgram->gndg_conn->gnc_state = GNILND_CONN_CONNECTING;
1221 /* don't need to serialize, there are no CQs for the dgram
1222 * EP on the kgn_net_t */
1223 grc = kgnilnd_ep_bind(dgram->gndg_conn->gnc_ephandle, host_id, dev->gnd_id);
1225 if (grc != GNI_RC_SUCCESS) {
1227 GOTO(post_failed, rc);
1232 /* If we are posting wildcards post using a net of 0, otherwise we'll use the
1233 * net of the destination node.
1236 if (dstnid == LNET_NID_ANY) {
1237 srcnid = LNET_MKNID(LNET_MKNET(GNILND, 0), dev->gnd_nid);
1239 srcnid = LNET_MKNID(LNET_NIDNET(dstnid), dev->gnd_nid);
1242 rc = kgnilnd_pack_connreq(&dgram->gndg_conn_out, dgram->gndg_conn,
1243 srcnid, dstnid, type);
1245 GOTO(post_failed, rc);
1248 if (type == GNILND_CONNREQ_NAK)
1249 dgram->gndg_conn_out.gncr_nakdata.gnnd_errno = data_rc;
1251 dgram->gndg_post_time = jiffies;
1253 /* XXX Nic: here is where we'd add in logical network multiplexing */
1255 CDEBUG(D_NETTRACE, "dgram 0x%p type %s %s->%s cdm %d\n",
1256 dgram, kgnilnd_dgram_type2str(dgram),
1257 libcfs_nid2str(srcnid),
1258 libcfs_nid2str(dstnid), dev->gnd_id);
1260 /* this allocates memory, can't hold locks across */
1261 grc = kgnilnd_ep_postdata_w_id(dgram->gndg_conn->gnc_ephandle,
1262 &dgram->gndg_conn_out, sizeof(kgn_connreq_t),
1263 &dgram->gndg_conn_in, sizeof(kgn_connreq_t),
1266 if (grc != GNI_RC_SUCCESS) {
1267 CNETERR("dropping failed dgram post id 0x%p type %s"
1268 " reqtype %s to %s: rc %d\n",
1269 dgram, kgnilnd_dgram_type2str(dgram),
1270 kgnilnd_connreq_type2str(&dgram->gndg_conn_out),
1271 libcfs_nid2str(dstnid), grc);
1272 rc = (grc == GNI_RC_ERROR_NOMEM) ? -ENOMEM : -EBADR;
1273 GOTO(post_failed, rc);
1276 /* we don't need to add earlier - if someone does del_peer during post,
1277 * that peer will get marked as unlinked and the callers wil take care of it.
1278 * The dgram code is largely kgn_peer_t ignorant, so at worst, we'll just drop
1279 * the completed dgram later when we cant find a peer to stuff it into */
1281 spin_lock(&dev->gnd_dgram_lock);
1283 /* make sure we are not double posting targeted dgrams
1284 * - we can multiple post WC dgrams to help with processing speed */
1285 if (dstnid != LNET_NID_ANY) {
1286 tmpdgram = kgnilnd_find_dgram_locked(dev, dstnid);
1288 LASSERTF(tmpdgram == NULL,
1289 "dgram 0x%p->%s already posted\n",
1290 dgram, libcfs_nid2str(dstnid));
1293 /* unmunge dstnid to help processing code cope... */
1294 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
1295 dgram->gndg_conn_out.gncr_dstnid = dstnid;
1298 list_add_tail(&dgram->gndg_list, kgnilnd_nid2dgramlist(dev, dstnid));
1299 dgram->gndg_state = GNILND_DGRAM_POSTED;
1300 spin_unlock(&dev->gnd_dgram_lock);
1303 if (rc < 0 && dgram != NULL) {
1304 kgnilnd_cleanup_dgram(dgram);
1305 kgnilnd_free_dgram(dev, dgram);
1311 /* The shutdown flag is set from the shutdown and stack reset threads. */
1313 kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown)
1315 /* The conns of canceled active dgrams need to be put in purgatory so
1316 * we don't reuse the mailbox */
1317 if (unlikely(dgram->gndg_state == GNILND_DGRAM_CANCELED)) {
1319 kgn_conn_t *conn = dgram->gndg_conn;
1320 lnet_nid_t nid = dgram->gndg_conn_out.gncr_dstnid;
1322 dgram->gndg_state = GNILND_DGRAM_DONE;
1324 /* During shutdown we've already removed the peer so we don't
1325 * need to add a peer. During stack reset we don't care about
1326 * MDDs since they are all released. */
1328 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1329 peer = kgnilnd_find_peer_locked(nid);
1332 CDEBUG(D_NET, "adding peer's conn with nid %s "
1333 "to purgatory\n", libcfs_nid2str(nid));
1334 kgnilnd_conn_addref(conn);
1335 conn->gnc_peer = peer;
1336 kgnilnd_peer_addref(peer);
1337 kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
1338 conn->gnc_state = GNILND_CONN_CLOSED;
1339 list_add_tail(&conn->gnc_list,
1341 kgnilnd_add_purgatory_locked(conn,
1343 kgnilnd_schedule_conn(conn);
1345 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1349 spin_lock(&dev->gnd_dgram_lock);
1350 kgnilnd_cancel_dgram_locked(dgram);
1351 spin_unlock(&dev->gnd_dgram_lock);
1353 kgnilnd_cleanup_dgram(dgram);
1355 /* if the dgram is 'canceled' it needs to be wait until the event
1356 * comes up from kgni that tells us it is safe to release */
1357 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1358 dgram->gndg_state = GNILND_DGRAM_DONE;
1360 LASSERTF(list_empty(&dgram->gndg_list), "dgram 0x%p on list\n", dgram);
1362 /* if it is a wildcard and we are in an appropriate state, repost
1365 if ((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
1366 (!kgnilnd_data.kgn_wc_kill && !kgnilnd_data.kgn_in_reset)) {
1369 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1371 /* We failed to repost the WC dgram for some reason
1372 * mark it so the repost system attempts to repost */
1373 kgnilnd_admin_addref(dev->gnd_nwcdgrams);
1377 /* always free the old dgram */
1378 kgnilnd_free_dgram(dev, dgram);
1384 kgnilnd_probe_for_dgram(kgn_device_t *dev, kgn_dgram_t **dgramp)
1386 kgn_dgram_t *dgram = NULL;
1387 gni_post_state_t post_state;
1391 __u32 remote_addr = 0, remote_id = 0;
1394 /* Probe with the lock held. That way if we get a dgram we dont have it canceled
1395 * between finding the ready dgram and grabbing the lock to remove it from the
1396 * list. Otherwise we could be left in an inconsistent state. We own the dgram
1397 * once its off the list so we don't need to worry about others changing it at
1399 spin_lock(&dev->gnd_dgram_lock);
1400 grc = kgnilnd_postdata_probe_by_id(dev->gnd_handle, &readyid);
1401 if (grc != GNI_RC_SUCCESS) {
1402 spin_unlock(&dev->gnd_dgram_lock);
1403 /* return 0 to indicate nothing happened */
1407 CDEBUG(D_NET, "ready "LPX64" on device 0x%p\n",
1410 dgram = (kgn_dgram_t *)readyid;
1412 LASSERTF(dgram->gndg_magic == GNILND_DGRAM_MAGIC,
1413 "dgram 0x%p from id "LPX64" with bad magic %x\n",
1414 dgram, readyid, dgram->gndg_magic);
1416 LASSERTF(dgram->gndg_state == GNILND_DGRAM_POSTED ||
1417 dgram->gndg_state == GNILND_DGRAM_CANCELED,
1418 "dgram 0x%p with bad state %s\n",
1419 dgram, kgnilnd_dgram_state2str(dgram));
1421 LASSERTF(!list_empty(&dgram->gndg_list),
1422 "dgram 0x%p with bad list state %s type %s\n",
1423 dgram, kgnilnd_dgram_state2str(dgram),
1424 kgnilnd_dgram_type2str(dgram));
1426 /* now we know that the datagram structure is ok, so pull off list */
1427 list_del_init(&dgram->gndg_list);
1429 /* while we have the gnn_dgram_lock and BEFORE we call test_by_id
1430 * change the state from POSTED to PROCESSING to ensure that
1431 * nobody cancels it after we've pulled it from the wire */
1432 if (dgram->gndg_state == GNILND_DGRAM_POSTED) {
1433 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1436 LASSERTF(dgram->gndg_conn != NULL,
1437 "dgram 0x%p with NULL conn\n", dgram);
1439 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1440 (__u64)dgram, &post_state,
1441 &remote_addr, &remote_id);
1443 /* we now "own" this datagram */
1444 spin_unlock(&dev->gnd_dgram_lock);
1446 LASSERTF(grc != GNI_RC_NO_MATCH, "kgni lied! probe_by_id told us that"
1447 " id "LPU64" was ready\n", readyid);
1449 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1450 "remote_addr %u remote_id %u\n", grc, dgram,
1451 kgnilnd_dgram_type2str(dgram),
1452 post_state, remote_addr, remote_id);
1454 if (unlikely(grc != GNI_RC_SUCCESS)) {
1455 CNETERR("getting data for dgram 0x%p->%s failed rc %d. Dropping it\n",
1456 dgram, libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1459 GOTO(probe_for_out, rc);
1462 rc = kgnilnd_process_dgram(dgram, post_state);
1464 /* we should never get probe finding a dgram for us and then it
1465 * being a WC dgram that is still in the middle of processing */
1466 LASSERTF(rc <= 0, "bad rc %d from process_dgram 0x%p state %d\n",
1467 rc, dgram, post_state);
1470 /* dgram is good enough for the data to be used */
1471 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1472 /* fake rc to mark that we've done something */
1475 /* let kgnilnd_release_dgram take care of canceled dgrams */
1476 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1477 dgram->gndg_state = GNILND_DGRAM_DONE;
1486 kgnilnd_release_dgram(dev, dgram, 0);
1491 kgnilnd_setup_wildcard_dgram(kgn_device_t *dev)
1493 /* if kgn_wildcard is zero, return error */
1494 int rc = -ENOENT, i;
1497 for (i = 0; i < *kgnilnd_tunables.kgn_nwildcard; i++) {
1498 rc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1500 CERROR("error %d: could not post wildcard datagram # %d\n",
1512 kgnilnd_cancel_net_dgrams(kgn_net_t *net)
1514 kgn_dgram_t *dg, *dgN;
1515 struct list_head zombies;
1519 /* we want to cancel any outstanding dgrams - we don't want to rely
1520 * on del_peer_or_conn catching all of them. This helps protect us in cases
1521 * where we don't quite keep the peer->dgram mapping in sync due to some
1522 * race conditions */
1524 LASSERTF(net->gnn_shutdown || kgnilnd_data.kgn_in_reset,
1525 "called with LND invalid state: net shutdown %d "
1526 "in reset %d\n", net->gnn_shutdown,
1527 kgnilnd_data.kgn_in_reset);
1529 INIT_LIST_HEAD(&zombies);
1531 spin_lock(&net->gnn_dev->gnd_dgram_lock);
1533 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1534 list_for_each_entry_safe(dg, dgN, &net->gnn_dev->gnd_dgrams[i], gndg_list) {
1536 /* skip nids not on our net or are wildcards */
1539 if (dg->gndg_type == GNILND_DGRAM_WC_REQ ||
1540 net->gnn_netnum != LNET_NETNUM(LNET_NIDNET(dg->gndg_conn_out.gncr_dstnid)))
1543 kgnilnd_cancel_dgram_locked(dg);
1547 spin_unlock(&net->gnn_dev->gnd_dgram_lock);
1553 kgnilnd_cancel_wc_dgrams(kgn_device_t *dev)
1555 kgn_dgram_t *dg, *dgN;
1556 struct list_head zombies;
1559 /* Time to kill the outstanding WC's
1560 * WC's exist on net 0 only but match on any net...
1563 LASSERTF(kgnilnd_data.kgn_in_reset || kgnilnd_data.kgn_wc_kill,
1564 "called with LND invalid state: WC shutdown %d "
1565 "in reset %d\n", kgnilnd_data.kgn_wc_kill,
1566 kgnilnd_data.kgn_in_reset);
1568 INIT_LIST_HEAD(&zombies);
1569 spin_lock(&dev->gnd_dgram_lock);
1572 dg = kgnilnd_find_dgram_locked(dev, LNET_NID_ANY);
1574 LASSERTF(dg->gndg_type == GNILND_DGRAM_WC_REQ,
1575 "dgram 0x%p->%s with bad type %d (%s)\n",
1576 dg, libcfs_nid2str(dg->gndg_conn_out.gncr_dstnid),
1577 dg->gndg_type, kgnilnd_dgram_type2str(dg));
1579 kgnilnd_cancel_dgram_locked(dg);
1581 /* WC could be DONE already, check and if so add to list to be released */
1582 if (dg->gndg_state == GNILND_DGRAM_DONE) {
1583 list_del_init(&dg->gndg_list);
1584 list_add_tail(&dg->gndg_list, &zombies);
1587 } while (dg != NULL);
1589 spin_unlock(&dev->gnd_dgram_lock);
1591 list_for_each_entry_safe(dg, dgN, &zombies, gndg_list) {
1592 list_del_init(&dg->gndg_list);
1593 kgnilnd_release_dgram(dev, dg, 1);
1600 kgnilnd_cancel_dgrams(kgn_device_t *dev)
1602 kgn_dgram_t *dg, *dgN;
1606 /* Cancel any outstanding non wildcard datagrams regardless
1607 * of which net they are on as we are in base shutdown and
1608 * dont care about connecting anymore.
1611 LASSERTF(kgnilnd_data.kgn_wc_kill == 1,"We didnt get called from base shutdown\n");
1613 spin_lock(&dev->gnd_dgram_lock);
1615 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size -1); i++) {
1616 list_for_each_entry_safe(dg, dgN, &dev->gnd_dgrams[i], gndg_list) {
1617 if (dg->gndg_type != GNILND_DGRAM_WC_REQ)
1618 kgnilnd_cancel_dgram_locked(dg);
1622 spin_unlock(&dev->gnd_dgram_lock);
1629 kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev)
1637 /* use do while to get at least one check run to allow
1638 * regression test for 762072 to hit bug if there */
1640 /* This function races with the dgram mover during shutdown so it is possible for
1641 * a dgram to be seen in kgnilnd_postdata_probe_wait_by_id but be handled in the
1642 * dgram mover thread instead of inside of this function.
1645 /* This should only be called from within shutdown, baseshutdown, or stack reset.
1646 * there are no assertions here to verify since base_shutdown has nothing in it we can check
1647 * the net is gone by then.
1652 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1653 "Waiting for %d canceled datagrams to clear on device %d\n",
1654 atomic_read(&dev->gnd_canceled_dgrams), dev->gnd_id);
1656 /* check once a second */
1657 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
1660 if (grc != GNI_RC_SUCCESS)
1663 CDEBUG(D_NET, "ready "LPX64" on device %d->0x%p\n",
1664 readyid, dev->gnd_id, dev);
1666 rc = kgnilnd_probe_for_dgram(dev, &dgram);
1668 /* if we got a valid dgram or one that is now done, clean up */
1669 kgnilnd_release_dgram(dev, dgram, 1);
1671 } while (atomic_read(&dev->gnd_canceled_dgrams));
1675 kgnilnd_start_connect(kgn_peer_t *peer)
1678 /* sync point for kgnilnd_del_peer_locked - do an early check to
1679 * catch the most common hits where del_peer is done by the
1680 * time we get here */
1681 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING1)) {
1682 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING1, 1)) {};
1685 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1686 if (!kgnilnd_peer_active(peer) || peer->gnp_connecting != GNILND_PEER_CONNECT) {
1687 /* raced with peer getting unlinked */
1688 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1692 peer->gnp_connecting = GNILND_PEER_POSTING;
1693 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1695 set_mb(peer->gnp_last_dgram_time, jiffies);
1696 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING2)) {
1697 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING2, 1)) {};
1700 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING3)) {
1701 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING3, 1)) {};
1702 rc = cfs_fail_val ? cfs_fail_val : -ENOMEM;
1704 rc = kgnilnd_post_dgram(peer->gnp_net->gnn_dev,
1705 peer->gnp_nid, GNILND_CONNREQ_REQ, 0);
1708 set_mb(peer->gnp_last_dgram_errno, rc);
1712 /* while we're posting someone could have decided this peer/dgram needed to
1713 * die a quick death, so we check for state change and process accordingly */
1715 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1716 if (!kgnilnd_peer_active(peer) || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1717 if (peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1718 peer->gnp_connecting = GNILND_PEER_KILL;
1720 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1721 /* positive RC to avoid dgram cleanup - we'll have to
1722 * wait for the kgni GNI_POST_TERMINATED event to
1723 * finish cleaning up */
1725 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev, peer->gnp_nid);
1728 peer->gnp_connecting = GNILND_PEER_POSTED;
1729 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1730 /* reaper thread will take care of any timeouts */
1731 CDEBUG(D_NET, "waiting for connect to finish to %s rc %d\n",
1732 libcfs_nid2str(peer->gnp_nid), rc);
1737 CDEBUG(D_NET, "connect to %s failed: rc %d \n",
1738 libcfs_nid2str(peer->gnp_nid), rc);
1744 kgnilnd_finish_connect(kgn_dgram_t *dgram)
1746 kgn_conn_t *conn = dgram->gndg_conn;
1747 lnet_nid_t her_nid = dgram->gndg_conn_in.gncr_srcnid;
1748 kgn_peer_t *new_peer, *peer = NULL;
1751 kgn_mbox_info_t *mbox;
1755 /* try to find a peer that matches the nid we got in the connreq
1756 * kgnilnd_unpack_connreq makes sure that conn_in.gncr_srcnid is
1757 * HER and conn_out.gncr_srcnid is ME for both active and WC dgrams */
1759 /* assume this is a new peer - it makes locking cleaner when it isn't */
1760 /* no holding kgn_net_rw_sem - already are at the kgnilnd_dgram_mover level */
1762 rc = kgnilnd_create_peer_safe(&new_peer, her_nid, NULL, GNILND_RCA_NODE_UP);
1764 CERROR("Can't create peer for %s\n", libcfs_nid2str(her_nid));
1768 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1770 /* this transfers ref from create_peer to the kgn_peer table */
1771 kgnilnd_add_peer_locked(her_nid, new_peer, &peer);
1773 /* if we found an existing peer, is it really ready for a new conn ? */
1774 if (peer != new_peer) {
1775 /* if this was an active connect attempt but we can't find a peer waiting for it
1776 * we will dump in the trash */
1778 if (peer->gnp_connecting == GNILND_PEER_IDLE && dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1779 CDEBUG(D_NET, "dropping completed connreq for %s peer 0x%p->%s\n",
1780 libcfs_nid2str(her_nid), peer, libcfs_nid2str(peer->gnp_nid));
1781 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1786 /* check to see if we can catch a connecting peer before it is
1787 * removed from the connd_peers list - if not, we need to
1788 * let the connreqs race and be handled by kgnilnd_conn_isdup_locked() */
1789 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
1790 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1791 if (!list_empty(&peer->gnp_connd_list)) {
1792 list_del_init(&peer->gnp_connd_list);
1793 /* drop connd ref */
1794 kgnilnd_peer_decref(peer);
1796 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1797 /* clear rc to make sure we don't have fake error */
1801 /* no matter what, we are no longer waiting to connect this peer now */
1802 peer->gnp_connecting = GNILND_PEER_IDLE;
1804 /* Refuse to duplicate an existing connection (both sides might try to
1805 * connect at once). NB we return success! We _are_ connected so we
1806 * _don't_ have any blocked txs to complete with failure. */
1807 rc = kgnilnd_conn_isdup_locked(peer, conn);
1809 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1810 CDEBUG(D_NET, "Not creating duplicate connection to %s: %d\n",
1811 libcfs_nid2str(her_nid), rc);
1817 if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
1818 CNETERR("Received connection request from %s that RCA thinks is"
1819 " down.\n", libcfs_nid2str(her_nid));
1820 peer->gnp_down = GNILND_RCA_NODE_UP;
1823 nstale = kgnilnd_close_stale_conns_locked(peer, conn);
1825 /* either way with peer (new or existing), we are ok with ref counts here as the
1826 * kgnilnd_add_peer_locked will use our ref on new_peer (from create_peer_safe) as the
1827 * ref for the peer table. */
1829 /* at this point, the connection request is a winner */
1831 /* mark 'DONE' to avoid cancel being called from release */
1832 dgram->gndg_state = GNILND_DGRAM_DONE;
1834 /* initialise timestamps before reaper looks at them */
1835 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
1837 /* last_tx is initialized to jiffies - (keepalive*2) so that if the NOOP fails it will
1838 * immediatly send a NOOP in the reaper thread during the call to
1839 * kgnilnd_check_conn_timeouts_locked
1841 conn->gnc_last_tx = jiffies - (cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout)) * 2);
1842 conn->gnc_state = GNILND_CONN_ESTABLISHED;
1844 /* save the dgram type used to establish this connection */
1845 conn->gnc_dgram_type = dgram->gndg_type;
1847 /* refs are not transferred from dgram to tables, so increment to
1849 kgnilnd_conn_addref(conn);
1850 kgnilnd_peer_addref(peer);
1851 conn->gnc_peer = peer;
1852 list_add_tail(&conn->gnc_list, &peer->gnp_conns);
1854 kgnilnd_conn_addref(conn); /* +1 ref for conn table */
1855 list_add_tail(&conn->gnc_hashlist,
1856 kgnilnd_cqid2connlist(conn->gnc_cqid));
1857 kgnilnd_data.kgn_conn_version++;
1859 /* Dont send NOOP if fail_loc is set
1861 if (!CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP)) {
1862 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, peer->gnp_net->gnn_ni->ni_nid);
1864 CNETERR("can't get TX to initiate NOOP to %s\n",
1865 libcfs_nid2str(peer->gnp_nid));
1867 kgnilnd_queue_tx(conn, tx);
1871 /* Schedule all packets blocking for a connection */
1872 list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1873 /* lock held here is the peer_conn lock */
1874 kgnilnd_tx_del_state_locked(tx, peer, NULL, GNILND_TX_ALLOCD);
1875 kgnilnd_queue_tx(conn, tx);
1878 /* If this is an active connection lets mark its timestamp on the MBoX */
1879 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1880 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1881 /* conn->gnc_last_rx is jiffies it better exist as it was just set */
1882 mbox->mbx_release_purg_active_dgram = conn->gnc_last_rx;
1885 /* Bug 765042: wake up scheduler for a race with finish_connect and
1886 * complete_conn_closed with a conn in purgatory
1887 * since we can't use CFS_RACE due to mutex_holds in kgnilnd_process_conns,
1888 * we just check for set and then clear */
1889 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
1891 /* get scheduler thread moving again */
1892 kgnilnd_schedule_device(conn->gnc_device);
1895 CDEBUG(D_NET, "New conn 0x%p->%s dev %d\n",
1896 conn, libcfs_nid2str(her_nid), conn->gnc_device->gnd_id);
1898 /* make sure we reset peer reconnect interval now that we have a good conn */
1899 kgnilnd_peer_alive(peer);
1900 peer->gnp_reconnect_interval = 0;
1902 /* clear the unlink attribute if we dont clear it kgnilnd_del_conn_or_peer will wait
1903 * on the atomic forever
1905 if (peer->gnp_pending_unlink) {
1906 peer->gnp_pending_unlink = 0;
1907 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1908 CDEBUG(D_NET, "Clearing peer unlink %p\n",peer);
1911 /* add ref to make it hang around until after we drop the lock */
1912 kgnilnd_conn_addref(conn);
1914 /* Once the peer_conn lock is dropped, the conn could actually move into
1915 * CLOSING->CLOSED->DONE in the scheduler thread, so hold the
1916 * lock until we are really done */
1917 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1919 /* Notify LNET that we now have a working connection to this peer.
1920 * This is a Cray extension to the "standard" LND behavior. */
1921 lnet_notify(peer->gnp_net->gnn_ni, peer->gnp_nid,
1922 1, cfs_time_current());
1924 /* drop our 'hold' ref */
1925 kgnilnd_conn_decref(conn);
1932 kgnilnd_send_nak(kgn_device_t *dev, lnet_nid_t dst_nid, int error)
1937 LASSERTF(dst_nid != LNET_NID_ANY, "bad dst_nid %s\n", libcfs_nid2str(dst_nid));
1939 CDEBUG(D_NET, "NAK to %s errno %d\n", libcfs_nid2str(dst_nid), error);
1941 rc = kgnilnd_post_dgram(dev, dst_nid, GNILND_CONNREQ_NAK, error);
1944 CDEBUG(D_NET, "NAK to %s failed: rc %d \n", libcfs_nid2str(dst_nid), rc);
1950 kgnilnd_process_nak(kgn_dgram_t *dgram)
1952 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
1953 lnet_nid_t src_nid = connreq->gncr_srcnid;
1954 int errno = connreq->gncr_nakdata.gnnd_errno;
1958 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1960 peer = kgnilnd_find_peer_locked(src_nid);
1962 /* we likely dropped him from bad data when we processed
1963 * the original REQ */
1964 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1968 /* need to check peerstamp/connstamp against the ones we find
1969 * to make sure we don't close new (and good?) conns that we
1970 * formed after this connreq failed */
1971 if (peer->gnp_connecting == GNILND_PEER_IDLE) {
1974 if (list_empty(&peer->gnp_conns)) {
1975 /* assume already procced datagram and it barfed up
1976 * on this side too */
1977 CDEBUG(D_NET, "dropping NAK from %s; "
1978 "peer %s is already not connected\n",
1979 libcfs_nid2str(connreq->gncr_srcnid),
1980 libcfs_nid2str(connreq->gncr_dstnid));
1981 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1985 /* stub up a connection with the connreq XXX_stamps to allow
1986 * use to use close_stale_conns_locked */
1987 conn.gnc_peerstamp = connreq->gncr_peerstamp;
1988 conn.gnc_my_connstamp = connreq->gncr_connstamp;
1989 conn.gnc_peer_connstamp = connreq->gncr_connstamp;
1990 conn.gnc_device = peer->gnp_net->gnn_dev;
1992 rc = kgnilnd_close_stale_conns_locked(peer, &conn);
1994 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
1995 "closed %d connections\n",
1996 libcfs_nid2str(connreq->gncr_srcnid),
1997 libcfs_nid2str(connreq->gncr_dstnid), errno, rc);
1999 spin_lock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2001 if (list_empty(&peer->gnp_connd_list)) {
2002 /* if peer isn't on waiting list, try to find one to nuke */
2003 rc = kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
2007 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
2008 "canceled pending connect request\n",
2009 libcfs_nid2str(connreq->gncr_srcnid),
2010 libcfs_nid2str(connreq->gncr_dstnid), errno);
2013 /* if we can't find a waiting dgram, we just drop the nak - the conn
2014 * connect must have failed (didn't find conn above and clear connecting
2015 * -- so nothing to do besides drop */
2017 /* peer is on list, meaning it is a new connect attempt from the one
2018 * we started that generated the NAK - so just drop NAK */
2020 /* use negative to prevent error message */
2023 spin_unlock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2026 /* success! we found a peer and at least marked pending_nak */
2027 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2033 kgnilnd_process_connreq(kgn_dgram_t *dgram, int *needs_nak)
2037 rc = kgnilnd_unpack_connreq(dgram);
2040 /* only NAK if we have good srcnid to use */
2046 switch (dgram->gndg_conn_in.gncr_type) {
2047 case GNILND_CONNREQ_REQ:
2048 /* wire up peer & conn, send queued TX */
2049 rc = kgnilnd_finish_connect(dgram);
2051 /* don't nak when the nid is hosed */
2057 case GNILND_CONNREQ_NAK:
2058 rc = kgnilnd_process_nak(dgram);
2059 /* return early to prevent reconnect bump */
2062 CERROR("unexpected connreq type %s (%d) from %s\n",
2063 kgnilnd_connreq_type2str(&dgram->gndg_conn_in),
2064 dgram->gndg_conn_in.gncr_type,
2065 libcfs_nid2str(dgram->gndg_conn_in.gncr_srcnid));
2076 kgnilnd_probe_and_process_dgram(kgn_device_t *dev)
2080 lnet_nid_t nak_dstnid = LNET_NID_ANY;
2081 lnet_nid_t orig_dstnid;
2082 kgn_dgram_t *dgram = NULL;
2086 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PAUSE_DGRAM_COMP)) {
2089 rc = kgnilnd_probe_for_dgram(dev, &dgram);
2094 } else if (rc < 0) {
2095 GOTO(inform_peer, rc);
2097 /* rc > 1 means it did something, reset for this func */
2101 switch (dgram->gndg_type) {
2102 case GNILND_DGRAM_WC_REQ:
2103 case GNILND_DGRAM_REQ:
2104 rc = kgnilnd_process_connreq(dgram, &needs_nak);
2106 case GNILND_DGRAM_NAK:
2107 CDEBUG(D_NETTRACE, "NAK to %s done\n",
2108 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
2111 CERROR("unknown datagram type %s (%d)\n",
2112 kgnilnd_dgram_type2str(dgram), dgram->gndg_type);
2116 /* stash data to use after releasing current datagram */
2117 /* don't stash net - we are operating on a net already,
2118 * so the lock on rw_net_lock is sufficient */
2120 nak_dstnid = dgram->gndg_conn_in.gncr_srcnid;
2123 LASSERTF(dgram != NULL, "dgram 0x%p rc %d needs_nak %d\n", dgram, rc, needs_nak);
2125 orig_dstnid = dgram->gndg_conn_out.gncr_dstnid;
2127 kgnilnd_release_dgram(dev, dgram, 0);
2129 CDEBUG(D_NET, "cleaning up dgram to %s, rc %d\n",
2130 libcfs_nid2str(orig_dstnid), rc);
2132 /* if this was a WC_REQ that matched an existing peer, it'll get marked done
2133 * in kgnilnd_finish_connect - if errors are from before we get to there,
2134 * we just drop as it is a WC_REQ - the peer CAN'T be waiting for it */
2135 if ((orig_dstnid != LNET_NID_ANY) && (rc < 0)) {
2136 /* if we have a negative rc, we want to find a peer to inform about
2137 * the bad connection attempt. Sorry buddy, better luck next time! */
2139 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2140 peer = kgnilnd_find_peer_locked(orig_dstnid);
2143 /* add ref to make sure he stays around past the possible unlink
2144 * so we can tell LNet about him */
2145 kgnilnd_peer_addref(peer);
2147 /* if he still cares about the outstanding connect */
2148 if (peer->gnp_connecting >= GNILND_PEER_CONNECT) {
2149 /* check if he is on the connd list and remove.. */
2150 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2151 if (!list_empty(&peer->gnp_connd_list)) {
2152 list_del_init(&peer->gnp_connd_list);
2153 /* drop connd ref */
2154 kgnilnd_peer_decref(peer);
2156 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2158 /* clear gnp_connecting so we don't have a non-connecting peer
2159 * on gnd_connd_list */
2160 peer->gnp_connecting = GNILND_PEER_IDLE;
2162 set_mb(peer->gnp_last_dgram_errno, rc);
2164 kgnilnd_peer_increase_reconnect_locked(peer);
2167 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2169 /* now that we are outside the lock, tell Mommy */
2171 kgnilnd_peer_notify(peer, rc);
2172 kgnilnd_peer_decref(peer);
2177 kgnilnd_send_nak(dev, nak_dstnid, rc);
2184 kgnilnd_reaper_dgram_check(kgn_device_t *dev)
2186 kgn_dgram_t *dgram, *tmp;
2189 spin_lock(&dev->gnd_dgram_lock);
2191 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
2192 list_for_each_entry_safe(dgram, tmp, &dev->gnd_dgrams[i], gndg_list) {
2193 unsigned long now = jiffies;
2194 unsigned long timeout;
2196 /* don't timeout stuff if the network is mucked or shutting down */
2197 if (kgnilnd_check_hw_quiesce()) {
2201 if ((dgram->gndg_state != GNILND_DGRAM_POSTED) ||
2202 (dgram->gndg_type == GNILND_DGRAM_WC_REQ)) {
2205 CDEBUG(D_NETTRACE, "checking dgram 0x%p type %s "
2206 "state %s conn 0x%p to %s age %lus\n",
2207 dgram, kgnilnd_dgram_type2str(dgram),
2208 kgnilnd_dgram_state2str(dgram), dgram->gndg_conn,
2209 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2210 cfs_duration_sec(now - dgram->gndg_post_time));
2212 timeout = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout);
2214 if (time_before(now, (dgram->gndg_post_time + timeout)))
2217 CNETERR("%s datagram to %s timed out @ %lus dgram "
2218 "0x%p state %s conn 0x%p\n",
2219 kgnilnd_dgram_type2str(dgram),
2220 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2221 cfs_duration_sec(now - dgram->gndg_post_time),
2222 dgram, kgnilnd_dgram_state2str(dgram),
2225 kgnilnd_cancel_dgram_locked(dgram);
2228 spin_unlock(&dev->gnd_dgram_lock);
2232 /* use a thread for the possibly long-blocking wait_by_id to prevent
2233 * stalling the global workqueues */
2235 kgnilnd_dgram_waitq(void *arg)
2237 kgn_device_t *dev = (kgn_device_t *) arg;
2241 DEFINE_WAIT(mover_done);
2243 snprintf(name, sizeof(name), "kgnilnd_dgn_%02d", dev->gnd_id);
2244 cfs_block_allsigs();
2246 /* all gnilnd threads need to run fairly urgently */
2247 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2249 /* we dont shut down until the device shuts down ... */
2250 while (!kgnilnd_data.kgn_shutdown) {
2251 /* to quiesce or to not quiesce, that is the question */
2252 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2253 KGNILND_SPIN_QUIESCE;
2256 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_DGRAM_COMP, 1)) {}
2258 /* check once a second */
2259 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
2262 if (grc == GNI_RC_SUCCESS) {
2263 CDEBUG(D_INFO, "waking up dgram mover thread\n");
2264 kgnilnd_schedule_dgram(dev);
2266 /* wait for dgram thread to ping us before spinning again */
2267 prepare_to_wait(&dev->gnd_dgping_waitq, &mover_done,
2268 TASK_INTERRUPTIBLE);
2270 /* don't sleep if we need to quiesce */
2271 if (likely(!kgnilnd_data.kgn_quiesce_trigger)) {
2274 finish_wait(&dev->gnd_dgping_waitq, &mover_done);
2278 kgnilnd_thread_fini();
2283 kgnilnd_start_outbound_dgrams(kgn_device_t *dev, unsigned long deadline)
2285 int did_something = 0, rc;
2286 kgn_peer_t *peer = NULL;
2288 spin_lock(&dev->gnd_connd_lock);
2290 /* Active connect - we added this in kgnilnd_launch_tx */
2291 while (!list_empty(&dev->gnd_connd_peers) && time_before(jiffies, deadline)) {
2292 peer = list_first_entry(&dev->gnd_connd_peers,
2293 kgn_peer_t, gnp_connd_list);
2295 /* ref for connd removed in if/else below */
2296 list_del_init(&peer->gnp_connd_list);
2298 /* gnp_connecting and membership on gnd_connd_peers should be
2299 * done coherently to avoid double adding, etc */
2300 /* don't need kgnilnd_data.kgn_peer_conn_lock here as that is only needed
2301 * to get the peer to gnp_connecting in the first place. We just need to
2302 * rely on gnd_connd_lock to serialize someone pulling him from the list
2303 * BEFORE clearing gnp_connecting */
2304 LASSERTF(peer->gnp_connecting != GNILND_PEER_IDLE, "peer 0x%p->%s not connecting\n",
2305 peer, libcfs_nid2str(peer->gnp_nid));
2307 spin_unlock(&dev->gnd_connd_lock);
2309 CDEBUG(D_NET, "processing connect to %s\n",
2310 libcfs_nid2str(peer->gnp_nid));
2313 rc = kgnilnd_start_connect(peer);
2315 if (likely(rc >= 0)) {
2316 /* 0 on success, positive on 'just drop peer' errors */
2317 kgnilnd_peer_decref(peer);
2318 } else if (rc == -ENOMEM) {
2319 /* if we are out of wildcards, add back to
2320 * connd_list - then break out and we'll try later
2321 * if other errors, we'll bail & cancel pending tx */
2322 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2323 if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2324 peer->gnp_connecting = GNILND_PEER_CONNECT;
2325 spin_lock(&dev->gnd_connd_lock);
2326 list_add_tail(&peer->gnp_connd_list,
2327 &dev->gnd_connd_peers);
2329 /* connecting changed while we were posting */
2331 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2332 " state 0x%p->%s, connecting %d\n",
2333 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2334 peer->gnp_connecting = GNILND_PEER_KILL;
2335 spin_lock(&dev->gnd_connd_lock);
2336 /* remove the peer ref frrom the cond list */
2337 kgnilnd_peer_decref(peer);
2338 /* let the system handle itself */
2340 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2341 /* the datagrams are a global pool,
2342 * so break out of trying and hope some free
2347 /* something bad happened, you lose */
2348 CNETERR("could not start connecting to %s "
2349 "rc %d: Will retry until TX timeout\n",
2350 libcfs_nid2str(peer->gnp_nid), rc);
2351 /* It didnt post so just set connecting back to zero now.
2352 * The reaper will reattempt the connection if it needs too.
2353 * If the peer needs death set it so the reaper will cleanup.
2355 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2356 if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2357 peer->gnp_connecting = GNILND_PEER_IDLE;
2358 kgnilnd_peer_increase_reconnect_locked(peer);
2360 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2361 " state 0x%p->%s, connecting %d\n",
2362 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2363 peer->gnp_connecting = GNILND_PEER_KILL;
2365 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2367 /* hold onto ref until we are really done - if it was
2368 * unlinked this could result in a destroy */
2369 kgnilnd_peer_decref(peer);
2371 spin_lock(&dev->gnd_connd_lock);
2374 spin_unlock(&dev->gnd_connd_lock);
2375 RETURN(did_something);
2379 kgnilnd_repost_wc_dgrams(kgn_device_t *dev)
2381 int did_something = 0, to_repost, i;
2382 to_repost = atomic_read(&dev->gnd_nwcdgrams);
2385 for (i = 0; i < to_repost; ++i) {
2387 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
2389 kgnilnd_admin_decref(dev->gnd_nwcdgrams);
2392 CDEBUG(D_NETERROR, "error %d: dev %d could not post wildcard datagram\n",
2398 RETURN(did_something);
2402 kgnilnd_dgram_poke_with_stick(unsigned long arg)
2405 kgn_device_t *dev = &kgnilnd_data.kgn_devices[dev_id];
2407 wake_up(&dev->gnd_dgram_waitq);
2410 /* use single thread for dgrams - should be sufficient for performance */
2412 kgnilnd_dgram_mover(void *arg)
2414 kgn_device_t *dev = (kgn_device_t *)arg;
2416 int rc, did_something;
2417 unsigned long next_purge_check = jiffies - 1;
2418 unsigned long timeout;
2419 struct timer_list timer;
2420 unsigned long deadline = 0;
2423 snprintf(name, sizeof(name), "kgnilnd_dg_%02d", dev->gnd_id);
2424 cfs_block_allsigs();
2425 /* all gnilnd threads need to run fairly urgently */
2426 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2428 /* we are ok not locking for these variables as the dgram waitq threads
2429 * will block both due to tying up net (kgn_shutdown) and the completion
2430 * event for the dgram_waitq (kgn_quiesce_trigger) */
2431 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2432 while (!kgnilnd_data.kgn_shutdown) {
2433 /* Safe: kgn_shutdown only set when quiescent */
2435 /* race with stack reset - we want to hold off seeing any new incoming dgrams
2436 * so we can force a dirty WC dgram for Bug 762072 - put right before
2437 * quiesce check so that it'll go right into that and not do any
2439 CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);
2441 /* to quiesce or to not quiesce, that is the question */
2442 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2443 KGNILND_SPIN_QUIESCE;
2447 CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);
2449 /* process any newly completed dgrams */
2450 down_read(&kgnilnd_data.kgn_net_rw_sem);
2452 rc = kgnilnd_probe_and_process_dgram(dev);
2454 did_something += rc;
2457 up_read(&kgnilnd_data.kgn_net_rw_sem);
2459 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_DGRAM_DEADLINE,
2460 (*kgnilnd_tunables.kgn_dgram_timeout + 1));
2461 /* start new outbound dgrams */
2462 did_something += kgnilnd_start_outbound_dgrams(dev, deadline);
2464 /* find dead dgrams */
2465 if (time_after_eq(jiffies, next_purge_check)) {
2466 /* these don't need to be checked that often */
2467 kgnilnd_reaper_dgram_check(dev);
2469 next_purge_check = (long) jiffies +
2470 cfs_time_seconds(kgnilnd_data.kgn_new_min_timeout / 4);
2473 did_something += kgnilnd_repost_wc_dgrams(dev);
2475 /* careful with the jiffy wrap... */
2476 timeout = (long)(next_purge_check - jiffies);
2478 CDEBUG(D_INFO, "did %d timeout %lu next %lu jiffies %lu\n",
2479 did_something, timeout, next_purge_check, jiffies);
2481 if ((did_something || timeout <= 0) && time_before(jiffies, deadline)) {
2486 prepare_to_wait(&dev->gnd_dgram_waitq, &wait, TASK_INTERRUPTIBLE);
2488 setup_timer(&timer, kgnilnd_dgram_poke_with_stick, dev->gnd_id);
2489 mod_timer(&timer, (long) jiffies + timeout);
2491 /* last second chance for others to poke us */
2492 did_something += xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_IDLE);
2494 /* check flag variables before comittingi even if we did something;
2495 * if we are after the deadline call schedule */
2496 if ((!did_something || time_after(jiffies, deadline)) &&
2497 !kgnilnd_data.kgn_shutdown &&
2498 !kgnilnd_data.kgn_quiesce_trigger) {
2499 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2500 timeout, cfs_duration_sec(timeout));
2501 wake_up_all(&dev->gnd_dgping_waitq);
2503 CDEBUG(D_INFO, "awake after schedule\n");
2504 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2507 del_singleshot_timer_sync(&timer);
2508 finish_wait(&dev->gnd_dgram_waitq, &wait);
2511 kgnilnd_thread_fini();