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 /* make sure we are mapping a clean block */
47 LASSERTF(fma_blk->gnm_hndl.qword1 == 0UL, "fma_blk %p dirty\n", fma_blk);
49 rrc = kgnilnd_mem_register(device->gnd_handle, (__u64)fma_blk->gnm_block,
50 fma_blk->gnm_blk_size, device->gnd_rcv_fma_cqh,
51 flags, &fma_blk->gnm_hndl);
52 if (rrc != GNI_RC_SUCCESS) {
53 /* XXX Nic: need a way to silence this for runtime stuff that is ok to fail
54 * -- like when under MDD or GART pressure on big systems
56 CNETERR("register fmablk failed 0x%p mbox_size %d flags %u\n",
57 fma_blk, fma_blk->gnm_mbox_size, flags);
61 /* PHYS_CONT memory isn't really mapped, at least not in GART -
62 * but all mappings chew up a MDD
64 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
65 atomic64_add(fma_blk->gnm_blk_size, &device->gnd_nbytes_map);
68 atomic_inc(&device->gnd_n_mdd);
69 /* nfmablk is live (mapped) blocks */
70 atomic_inc(&device->gnd_nfmablk);
76 kgnilnd_alloc_fmablk(kgn_device_t *device, int use_phys)
80 kgn_fma_memblock_t *fma_blk;
81 gni_smsg_attr_t smsg_attr;
82 unsigned long fmablk_vers;
84 /* we'll use fmablk_vers and the gnd_fmablk_mutex to gate access
85 * to this allocation code. Everyone will sample the version
86 * before and after getting the mutex. If it has changed,
87 * we'll bail out to check the lists again - this indicates that
88 * some sort of change was made to the lists and it is possible
89 * that there is a mailbox for us to find now. This should prevent
90 * a ton of spinning in the case where there are lots of threads
91 * that need a yet-to-be-allocated mailbox for a connection. */
93 fmablk_vers = atomic_read(&device->gnd_fmablk_vers);
94 mutex_lock(&device->gnd_fmablk_mutex);
96 if (fmablk_vers != atomic_read(&device->gnd_fmablk_vers)) {
97 /* version changed while we were waiting for semaphore,
98 * we'll recheck the lists assuming something nice happened */
99 mutex_unlock(&device->gnd_fmablk_mutex);
103 LIBCFS_ALLOC(fma_blk, sizeof(kgn_fma_memblock_t));
104 if (fma_blk == NULL) {
105 CNETERR("could not allocate fma block descriptor\n");
110 INIT_LIST_HEAD(&fma_blk->gnm_bufflist);
112 kgnilnd_setup_smsg_attr(&smsg_attr);
114 gni_smsg_buff_size_needed(&smsg_attr, &fma_blk->gnm_mbox_size);
116 LASSERTF(fma_blk->gnm_mbox_size, "mbox size %d\n", fma_blk->gnm_mbox_size);
118 /* gni_smsg_buff_size_needed calculates the base mailbox size and since
119 * we want to hold kgn_peer_credits worth of messages in both directions,
120 * we add PAYLOAD to grow the mailbox size
123 fma_blk->gnm_mbox_size += GNILND_MBOX_PAYLOAD;
125 /* we'll only use physical during preallocate at startup -- this keeps it nice and
126 * clean for runtime decisions. We'll keep the PHYS ones around until shutdown
127 * as reallocating them is tough if there is memory fragmentation */
130 fma_blk->gnm_block = kmem_cache_alloc(kgnilnd_data.kgn_mbox_cache, GFP_ATOMIC);
131 if (fma_blk->gnm_block == NULL) {
132 CNETERR("could not allocate physical SMSG mailbox memory\n");
136 fma_blk->gnm_blk_size = KMALLOC_MAX_SIZE;
137 num_mbox = fma_blk->gnm_blk_size / fma_blk->gnm_mbox_size;
139 LASSERTF(num_mbox >= 1,
140 "num_mbox %d blk_size %u mbox_size %d\n",
141 num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size);
143 fma_blk->gnm_state = GNILND_FMABLK_PHYS;
146 num_mbox = *kgnilnd_tunables.kgn_mbox_per_block;
147 fma_blk->gnm_blk_size = num_mbox * fma_blk->gnm_mbox_size;
149 LASSERTF(num_mbox >= 1 && num_mbox >= *kgnilnd_tunables.kgn_mbox_per_block,
150 "num_mbox %d blk_size %u mbox_size %d tunable %d\n",
151 num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size,
152 *kgnilnd_tunables.kgn_mbox_per_block);
154 LIBCFS_ALLOC(fma_blk->gnm_block, fma_blk->gnm_blk_size);
155 if (fma_blk->gnm_block == NULL) {
156 CNETERR("could not allocate virtual SMSG mailbox memory, %d bytes\n", fma_blk->gnm_blk_size);
161 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
164 /* allocate just enough space for the bits to track the mailboxes */
165 LIBCFS_ALLOC(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox) * sizeof(unsigned long));
166 if (fma_blk->gnm_bit_array == NULL) {
167 CNETERR("could not allocate mailbox bitmask, %lu bytes for %d mbox\n",
168 sizeof(unsigned long) * BITS_TO_LONGS(num_mbox), num_mbox);
172 bitmap_zero(fma_blk->gnm_bit_array, num_mbox);
174 /* now that the num_mbox is set based on allocation type, get debug info setup */
175 LIBCFS_ALLOC(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t) * num_mbox);
176 if (fma_blk->gnm_mbox_info == NULL) {
177 CNETERR("could not allocate mailbox debug, %lu bytes for %d mbox\n",
178 sizeof(kgn_mbox_info_t) * num_mbox, num_mbox);
183 rc = kgnilnd_map_fmablk(device, fma_blk);
188 fma_blk->gnm_next_avail_mbox = 0;
189 fma_blk->gnm_avail_mboxs = fma_blk->gnm_num_mboxs = num_mbox;
191 CDEBUG(D_MALLOC, "alloc fmablk 0x%p num %d msg_maxsize %d credits %d "
192 "mbox_size %d MDD "LPX64"."LPX64"\n",
193 fma_blk, num_mbox, smsg_attr.msg_maxsize, smsg_attr.mbox_maxcredit,
194 fma_blk->gnm_mbox_size, fma_blk->gnm_hndl.qword1,
195 fma_blk->gnm_hndl.qword2);
197 /* lock Is protecting data structures, not semaphore */
199 spin_lock(&device->gnd_fmablk_lock);
200 list_add_tail(&fma_blk->gnm_bufflist, &device->gnd_fma_buffs);
202 /* toggle under the lock so once they change the list is also
203 * ready for others to traverse */
204 atomic_inc(&device->gnd_fmablk_vers);
206 spin_unlock(&device->gnd_fmablk_lock);
208 mutex_unlock(&device->gnd_fmablk_mutex);
213 LIBCFS_FREE(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t)*num_mbox);
215 LIBCFS_FREE(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox) * sizeof (unsigned long));
217 if (fma_blk->gnm_state == GNILND_FMABLK_VIRT) {
218 LIBCFS_FREE(fma_blk->gnm_block, fma_blk->gnm_blk_size);
220 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
223 LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
225 mutex_unlock(&device->gnd_fmablk_mutex);
230 kgnilnd_unmap_fmablk(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
234 /* if some held, set hold_timeout from conn timeouts used in this block
235 * but not during shutdown, then just nuke and pave */
236 if (fma_blk->gnm_held_mboxs && (!kgnilnd_data.kgn_shutdown)) {
237 fma_blk->gnm_hold_timeout = GNILND_TIMEOUT2DEADMAN;
240 /* we are changing the state of a block, tickle version to tell
241 * proc code list is stale now */
242 atomic_inc(&dev->gnd_fmablk_vers);
244 rrc = kgnilnd_mem_deregister(dev->gnd_handle, &fma_blk->gnm_hndl, fma_blk->gnm_hold_timeout);
246 CDEBUG(rrc == GNI_RC_SUCCESS ? D_MALLOC : D_CONSOLE|D_NETERROR,
247 "unmap fmablk 0x%p@%s sz %u total %d avail %d held %d mbox_size %d "
249 fma_blk, kgnilnd_fmablk_state2str(fma_blk->gnm_state),
250 fma_blk->gnm_blk_size, fma_blk->gnm_num_mboxs,
251 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs,
252 fma_blk->gnm_mbox_size, fma_blk->gnm_hold_timeout);
254 LASSERTF(rrc == GNI_RC_SUCCESS,
255 "tried to double unmap or something bad, fma_blk %p (rrc %d)\n",
258 if (fma_blk->gnm_hold_timeout) {
259 atomic_inc(&dev->gnd_n_mdd_held);
261 atomic_dec(&dev->gnd_n_mdd);
264 /* PHYS blocks don't get mapped */
265 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
266 atomic64_sub(fma_blk->gnm_blk_size, &dev->gnd_nbytes_map);
267 fma_blk->gnm_state = GNILND_FMABLK_IDLE;
268 } else if (kgnilnd_data.kgn_in_reset) {
269 /* in stack reset, clear MDD handle for PHYS blocks, as we'll
270 * re-use the fma_blk after reset so we don't have to drop/allocate
271 * all of those physical blocks */
272 fma_blk->gnm_hndl.qword1 = fma_blk->gnm_hndl.qword2 = 0UL;
275 /* Decrement here as this is the # of mapped blocks */
276 atomic_dec(&dev->gnd_nfmablk);
280 /* needs lock on gnd_fmablk_lock to cover gnd_fma_buffs */
282 kgnilnd_free_fmablk_locked(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
284 LASSERTF(fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs,
285 "fma_blk %p@%d free in bad state (%d): blk total %d avail %d held %d\n",
286 fma_blk, fma_blk->gnm_state, fma_blk->gnm_hold_timeout, fma_blk->gnm_num_mboxs,
287 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs);
289 atomic_inc(&dev->gnd_fmablk_vers);
291 if (fma_blk->gnm_hold_timeout) {
292 CDEBUG(D_MALLOC, "mdd release fmablk 0x%p sz %u avail %d held %d "
294 fma_blk, fma_blk->gnm_blk_size, fma_blk->gnm_avail_mboxs,
295 fma_blk->gnm_held_mboxs, fma_blk->gnm_mbox_size);
297 /* We leave MDD dangling over stack reset */
298 if (!kgnilnd_data.kgn_in_reset) {
299 kgnilnd_mem_mdd_release(dev->gnd_handle, &fma_blk->gnm_hndl);
301 /* ignoring the return code - if kgni/ghal can't find it
302 * it must be released already */
303 atomic_dec(&dev->gnd_n_mdd_held);
304 atomic_dec(&dev->gnd_n_mdd);
307 /* we cant' free the gnm_block until all the conns have released their
308 * purgatory holds. While we have purgatory holds, we might check the conn
309 * RX mailbox during the CLOSING process. It is possible that kgni might
310 * try to look into the RX side for credits when sending the CLOSE msg too */
311 CDEBUG(D_MALLOC, "fmablk %p free buffer %p mbox_size %d\n",
312 fma_blk, fma_blk->gnm_block, fma_blk->gnm_mbox_size);
314 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
315 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
317 LIBCFS_FREE(fma_blk->gnm_block, fma_blk->gnm_blk_size);
319 fma_blk->gnm_state = GNILND_FMABLK_FREED;
321 list_del(&fma_blk->gnm_bufflist);
323 LIBCFS_FREE(fma_blk->gnm_mbox_info, sizeof(kgn_mbox_info_t)*fma_blk->gnm_num_mboxs);
324 LIBCFS_FREE(fma_blk->gnm_bit_array, BITS_TO_LONGS(fma_blk->gnm_num_mboxs) * sizeof (unsigned long));
325 LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
329 kgnilnd_find_free_mbox(kgn_conn_t *conn)
331 kgn_device_t *dev = conn->gnc_device;
332 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
333 kgn_fma_memblock_t *fma_blk;
334 kgn_mbox_info_t *mbox = NULL;
337 spin_lock(&dev->gnd_fmablk_lock);
339 list_for_each_entry(fma_blk, &conn->gnc_device->gnd_fma_buffs,
341 if (fma_blk->gnm_avail_mboxs <= 0 ||
342 fma_blk->gnm_state <= GNILND_FMABLK_IDLE) {
345 /* look in bitarray for available mailbox */
347 id = find_next_zero_bit(
348 fma_blk->gnm_bit_array,
349 fma_blk->gnm_num_mboxs,
350 fma_blk->gnm_next_avail_mbox);
351 if (id == fma_blk->gnm_num_mboxs &&
352 fma_blk->gnm_next_avail_mbox != 0) {
354 fma_blk->gnm_next_avail_mbox = 0;
360 LASSERTF(id < fma_blk->gnm_num_mboxs, "id %d max %d\n",
361 id, fma_blk->gnm_num_mboxs);
362 set_bit(id, (volatile unsigned long *)fma_blk->gnm_bit_array);
363 conn->gnc_mbox_id = id;
365 fma_blk->gnm_next_avail_mbox =
366 (id == (fma_blk->gnm_num_mboxs - 1)) ? 0 : (id + 1);
367 fma_blk->gnm_avail_mboxs--;
368 conn->gnc_fma_blk = fma_blk;
370 kgnilnd_setup_smsg_attr(smsg_attr);
372 smsg_attr->msg_buffer = fma_blk->gnm_block;
373 smsg_attr->mbox_offset = fma_blk->gnm_mbox_size * id;
374 smsg_attr->mem_hndl = fma_blk->gnm_hndl;
375 smsg_attr->buff_size = fma_blk->gnm_mbox_size;
377 /* We'll set the hndl to zero for PHYS blocks unmapped during stack
378 * reset and re-use the same fma_blk after stack reset. This ensures we've
379 * properly mapped it before we use it */
380 LASSERTF(fma_blk->gnm_hndl.qword1 != 0UL, "unmapped fma_blk %p, state %d\n",
381 fma_blk, fma_blk->gnm_state);
383 CDEBUG(D_NET, "conn %p smsg %p fmablk %p "
384 "allocating SMSG mbox %d buf %p "
385 "offset %u hndl "LPX64"."LPX64"\n",
386 conn, smsg_attr, fma_blk, id,
387 smsg_attr->msg_buffer, smsg_attr->mbox_offset,
388 fma_blk->gnm_hndl.qword1,
389 fma_blk->gnm_hndl.qword2);
391 mbox = &fma_blk->gnm_mbox_info[id];
392 mbox->mbx_create_conn_memset = jiffies;
394 mbox->mbx_nallocs_total++;
396 /* zero mbox to remove any old data from our last use.
397 * this better be safe, if not our purgatory timers
398 * are too short or a peer really is misbehaving */
399 memset(smsg_attr->msg_buffer + smsg_attr->mbox_offset,
400 0, smsg_attr->buff_size);
404 spin_unlock(&dev->gnd_fmablk_lock);
408 kgnilnd_setup_mbox(kgn_conn_t *conn)
410 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
413 smsg_attr->msg_buffer = NULL;
414 /* Look for available mbox */
416 kgnilnd_find_free_mbox(conn);
418 /* nothing in the existing buffers, make a new one */
419 if (smsg_attr->msg_buffer == NULL) {
420 /* for runtime allocations, we only want vmalloc */
421 err = kgnilnd_alloc_fmablk(conn->gnc_device, 0);
426 } while (smsg_attr->msg_buffer == NULL);
429 CNETERR("couldn't allocate SMSG mbox for conn %p Error: %d\n",
435 kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold)
437 kgn_device_t *dev = conn->gnc_device;
438 gni_smsg_attr_t *smsg_attr = &conn->gnpr_smsg_attr;
439 kgn_fma_memblock_t *fma_blk = NULL;
440 kgn_mbox_info_t *mbox = NULL;
444 /* if we failed to setup mbox and now destroying conn */
445 if (smsg_attr->msg_buffer == NULL) {
449 id = conn->gnc_mbox_id;
451 spin_lock(&dev->gnd_fmablk_lock);
452 /* make sure our conn points at a valid fma_blk
453 * We use this instead of a mem block search out of smsg_attr
454 * because we could have freed a block for fma_blk #1 but the fma_blk
455 * is still in the list for a purgatory hold. This would induce a false
456 * match if that same block gets reallocated to fma_blk #2 */
457 list_for_each_entry(fma_blk, &dev->gnd_fma_buffs, gnm_bufflist) {
458 if (fma_blk == conn->gnc_fma_blk) {
463 LASSERTF(found, "unable to find conn 0x%p with gnc_fma_blk %p "
464 "anywhere in the world\n", conn, conn->gnc_fma_blk);
466 LASSERTF(id < fma_blk->gnm_num_mboxs,
467 "bad id %d max %d\n",
468 id, fma_blk->gnm_num_mboxs);
470 /* < 0 - was held, now free it
471 * == 0 - just free it
472 * > 0 - hold it for now */
473 if (purgatory_hold == 0) {
474 CDEBUG(D_NET, "conn %p smsg %p fmablk %p freeing SMSG mbox %d "
475 "hndl "LPX64"."LPX64"\n",
476 conn, smsg_attr, fma_blk, id,
477 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
478 fma_blk->gnm_avail_mboxs++;
480 } else if (purgatory_hold > 0) {
481 CDEBUG(D_NET, "conn %p smsg %p fmablk %p holding SMSG mbox %d "
482 "hndl "LPX64"."LPX64"\n",
483 conn, smsg_attr, fma_blk, id,
484 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
486 fma_blk->gnm_held_mboxs++;
487 fma_blk->gnm_max_timeout = MAX(fma_blk->gnm_max_timeout,
490 CDEBUG(D_NET, "conn %p smsg %p fmablk %p release SMSG mbox %d "
491 "hndl "LPX64"."LPX64"\n",
492 conn, smsg_attr, fma_blk, id,
493 fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
495 fma_blk->gnm_held_mboxs--;
496 fma_blk->gnm_avail_mboxs++;
499 if (purgatory_hold <= 0) {
500 /* if kgni is retransmitting, freeing the smsg block before the EP
501 * is destroyed gets messy. Bug 768295. */
502 LASSERTF(conn->gnc_ephandle == NULL,
503 "can't release mbox before EP is nuked. conn 0x%p\n", conn);
505 mbox = &fma_blk->gnm_mbox_info[id];
506 mbox->mbx_release_from_purgatory = jiffies;
508 /* clear conn gnc_fmablk if it is gone - this allows us to
509 * not worry about state so much in kgnilnd_destroy_conn
510 * and makes the guaranteed cleanup of the resources easier */
511 LASSERTF(test_and_clear_bit(id, fma_blk->gnm_bit_array),
512 "conn %p bit %d already cleared in fma_blk %p\n",
514 conn->gnc_fma_blk = NULL;
518 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FMABLK_AVAIL)) {
519 CERROR("LBUGs in your future: forcibly marking fma_blk %p "
520 "as mapped\n", fma_blk);
521 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
524 /* we don't release or unmap PHYS blocks as part of the normal cycle --
525 * those are controlled manually from startup/shutdown */
526 if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
527 /* we can unmap once all are unused (held or avail)
528 * but check hold_timeout to make sure we are not trying to double
529 * unmap this buffer. If there was no hold_timeout set due to
530 * held_mboxs, we'll free the mobx here shortly and won't have to
531 * worry about catching a double free for a 'clean' fma_blk */
532 if (((fma_blk->gnm_avail_mboxs + fma_blk->gnm_held_mboxs) == fma_blk->gnm_num_mboxs) &&
533 (!fma_blk->gnm_hold_timeout)) {
534 kgnilnd_unmap_fmablk(dev, fma_blk);
537 /* But we can only free once they are all avail */
538 if (fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs &&
539 fma_blk->gnm_held_mboxs == 0) {
540 /* all mailboxes are released, free fma_blk */
541 kgnilnd_free_fmablk_locked(dev, fma_blk);
545 spin_unlock(&dev->gnd_fmablk_lock);
549 kgnilnd_count_phys_mbox(kgn_device_t *device)
552 kgn_fma_memblock_t *fma_blk;
554 spin_lock(&device->gnd_fmablk_lock);
556 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
557 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
558 i += fma_blk->gnm_num_mboxs;
560 spin_unlock(&device->gnd_fmablk_lock);
566 kgnilnd_allocate_phys_fmablk(kgn_device_t *device)
570 while (kgnilnd_count_phys_mbox(device) < *kgnilnd_tunables.kgn_nphys_mbox) {
572 rc = kgnilnd_alloc_fmablk(device, 1);
574 CERROR("failed phys mbox allocation, stopping at %d, rc %d\n",
575 kgnilnd_count_phys_mbox(device), rc);
583 kgnilnd_map_phys_fmablk(kgn_device_t *device)
587 kgn_fma_memblock_t *fma_blk;
589 /* use mutex to gate access to single thread, just in case */
590 mutex_lock(&device->gnd_fmablk_mutex);
592 spin_lock(&device->gnd_fmablk_lock);
594 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
595 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
596 rc = kgnilnd_map_fmablk(device, fma_blk);
601 spin_unlock(&device->gnd_fmablk_lock);
603 mutex_unlock(&device->gnd_fmablk_mutex);
609 kgnilnd_unmap_fma_blocks(kgn_device_t *device)
612 kgn_fma_memblock_t *fma_blk;
614 /* use mutex to gate access to single thread, just in case */
615 mutex_lock(&device->gnd_fmablk_mutex);
617 spin_lock(&device->gnd_fmablk_lock);
619 list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
620 kgnilnd_unmap_fmablk(device, fma_blk);
622 spin_unlock(&device->gnd_fmablk_lock);
624 mutex_unlock(&device->gnd_fmablk_mutex);
628 kgnilnd_free_phys_fmablk(kgn_device_t *device)
631 kgn_fma_memblock_t *fma_blk, *fma_blkN;
633 /* use mutex to gate access to single thread, just in case */
634 mutex_lock(&device->gnd_fmablk_mutex);
636 spin_lock(&device->gnd_fmablk_lock);
638 list_for_each_entry_safe(fma_blk, fma_blkN, &device->gnd_fma_buffs, gnm_bufflist) {
639 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
640 kgnilnd_free_fmablk_locked(device, fma_blk);
642 spin_unlock(&device->gnd_fmablk_lock);
644 mutex_unlock(&device->gnd_fmablk_mutex);
647 /* kgnilnd dgram nid->struct managment */
649 static inline struct list_head *
650 kgnilnd_nid2dgramlist(kgn_device_t *dev, lnet_nid_t nid)
652 unsigned int hash = ((unsigned int)nid) % *kgnilnd_tunables.kgn_peer_hash_size;
654 RETURN(&dev->gnd_dgrams[hash]);
658 /* needs dev->gnd_dgram_lock held */
660 kgnilnd_find_dgram_locked(kgn_device_t *dev, lnet_nid_t dst_nid)
662 struct list_head *dgram_list = kgnilnd_nid2dgramlist(dev, dst_nid);
665 list_for_each_entry(dgram, dgram_list, gndg_list) {
667 /* if state > POSTED, we are already handling cancel/completion */
668 if ((dgram->gndg_conn_out.gncr_dstnid != dst_nid) ||
669 dgram->gndg_state > GNILND_DGRAM_POSTED)
672 CDEBUG(D_NET, "got dgram [%p] -> %s\n",
673 dgram, libcfs_nid2str(dst_nid));
680 kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid)
684 spin_lock(&dev->gnd_dgram_lock);
685 dgram = kgnilnd_find_dgram_locked(dev, dst_nid);
688 kgnilnd_cancel_dgram_locked(dgram);
690 spin_unlock(&dev->gnd_dgram_lock);
692 RETURN(!!(dgram == NULL));
696 kgnilnd_pack_connreq(kgn_connreq_t *connreq, kgn_conn_t *conn,
697 lnet_nid_t srcnid, lnet_nid_t dstnid,
698 kgn_connreq_type_t type)
702 /* ensure we haven't violated max datagram size */
703 CLASSERT(sizeof(kgn_connreq_t) <= GNI_DATAGRAM_MAXSIZE);
705 /* no need to zero out, we do that when allocating dgram */
706 connreq->gncr_magic = GNILND_MSG_MAGIC;
708 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_SRCNID)) {
710 } else if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
714 connreq->gncr_srcnid = srcnid;
715 connreq->gncr_dstnid = dstnid;
717 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
718 connreq->gncr_version = 99;
720 connreq->gncr_version = GNILND_CONNREQ_VERSION;
722 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
723 connreq->gncr_type = 99;
725 connreq->gncr_type = type;
727 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
728 connreq->gncr_peerstamp = 0;
730 connreq->gncr_peerstamp = kgnilnd_data.kgn_peerstamp;
732 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
733 connreq->gncr_connstamp = 0;
735 connreq->gncr_connstamp = conn->gnc_my_connstamp;
737 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
738 connreq->gncr_timeout = 0;
740 connreq->gncr_timeout = conn->gnc_timeout;
743 /* the rest pack the data into the payload in other places */
744 if (type == GNILND_CONNREQ_REQ) {
745 kgn_gniparams_t *req_params = &connreq->gncr_gnparams;
746 req_params->gnpr_host_id = conn->gnc_device->gnd_host_id;
747 req_params->gnpr_cqid = conn->gnc_cqid;
749 /* allocate mailbox for this connection */
750 err = kgnilnd_setup_mbox(conn);
752 CERROR("Failed to setup FMA mailbox (%d)\n", err);
754 req_params->gnpr_smsg_attr = conn->gnpr_smsg_attr;
757 /* XXX Nic: TBD - checksum computation */
763 kgnilnd_unpack_connreq(kgn_dgram_t *dgram)
765 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
769 /* the following fields must be handled in a backwards compatible
770 * manner to ensure we can always send and interpret NAKs */
772 if (connreq->gncr_magic != GNILND_MSG_MAGIC &&
773 connreq->gncr_magic != __swab32(GNILND_MSG_MAGIC)) {
774 /* Unexpected magic! */
775 CERROR("Unexpected magic %08x\n",
776 connreq->gncr_magic);
780 swab = (connreq->gncr_magic == __swab32(GNILND_MSG_MAGIC));
782 __swab32s(&connreq->gncr_magic);
783 __swab32s(&connreq->gncr_cksum);
784 __swab16s(&connreq->gncr_type);
785 __swab16s(&connreq->gncr_version);
786 __swab32s(&connreq->gncr_timeout);
787 __swab64s(&connreq->gncr_srcnid);
788 __swab64s(&connreq->gncr_dstnid);
789 __swab64s(&connreq->gncr_peerstamp);
790 __swab64s(&connreq->gncr_connstamp);
793 /* Do NOT return anything but -EBADF before we munge
794 * connreq->gncr_srcnid - we need that to send the nak */
796 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
797 lnet_nid_t incoming = connreq->gncr_srcnid;
799 /* even if the incoming packet is hosed, we know who we sent
800 * the original and can set the srcnid so that we can properly
801 * look up our peer to close the loop on this connreq. We still use
802 * -EBADF to prevent a NAK - just in case there are issues with
803 * the payload coming from a random spot, etc. */
804 connreq->gncr_srcnid = dgram->gndg_conn_out.gncr_dstnid;
806 if (LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid) !=
807 LNET_NIDADDR(incoming)) {
808 /* we got a datagram match for the wrong nid... */
809 CERROR("matched datagram 0x%p with srcnid %s "
810 "(%x), expecting %s (%x)\n",
812 libcfs_nid2str(incoming),
813 LNET_NIDADDR(incoming),
814 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
815 LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid));
819 /* if we have a wildcard datagram it should match an
820 * incoming "active" datagram that should have a fully formed
821 * srcnid and dstnid. If we couldn't unpack it, we drop as
822 * corrupted packet, otherwise we'll just verify that the dstnid
823 * matches the NID for the NET that the dgram was posted */
825 /* make sure their wildcard didn't match ours, that is unpossible */
826 LASSERTF(connreq->gncr_dstnid != LNET_NID_ANY,
827 "dgram 0x%p from %s, connreq 0x%p; "
828 "wildcard matched wildcard \n", dgram,
829 libcfs_nid2str(connreq->gncr_srcnid), connreq);
831 rc = kgnilnd_find_net(connreq->gncr_dstnid, &net);
833 if (rc == -ESHUTDOWN) {
834 CERROR("Looking up network: device is in shutdown");
836 } else if (rc == -ENONET) {
837 CERROR("Connection data from %s: she sent "
838 "dst_nid %s, but net lookup failed on "
840 libcfs_nid2str(connreq->gncr_srcnid),
841 libcfs_nid2str(connreq->gncr_dstnid),
842 dgram, kgnilnd_dgram_type2str(dgram));
846 if (net->gnn_ni->ni_nid != connreq->gncr_dstnid) {
847 CERROR("Bad connection data from %s: she sent "
848 "dst_nid %s, but I am %s with dgram 0x%p@%s\n",
849 libcfs_nid2str(connreq->gncr_srcnid),
850 libcfs_nid2str(connreq->gncr_dstnid),
851 libcfs_nid2str(net->gnn_ni->ni_nid),
852 dgram, kgnilnd_dgram_type2str(dgram));
853 kgnilnd_net_decref(net);
857 /* kgnilnd_find_net takes a ref on the net it finds, You need to decref it when not needed. */
858 kgnilnd_net_decref(net);
861 if (connreq->gncr_version != GNILND_CONNREQ_VERSION) {
862 CERROR("Unexpected version %d\n", connreq->gncr_version);
866 /* XXX Nic: TBD - checksum validation */
867 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_DROP)) {
871 if (swab && connreq->gncr_type == GNILND_CONNREQ_REQ) {
872 __u64 msg_addr = (__u64) connreq->gncr_gnparams.gnpr_smsg_attr.msg_buffer;
874 __swab32s(&connreq->gncr_gnparams.gnpr_host_id);
875 __swab32s(&connreq->gncr_gnparams.gnpr_cqid);
876 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.buff_size);
877 __swab16s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_maxcredit);
878 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_offset);
879 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword1);
880 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword2);
881 __swab64s(&msg_addr);
882 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_maxsize);
883 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_type);
884 } else if (swab && connreq->gncr_type == GNILND_CONNREQ_NAK) {
885 __swab32s(&connreq->gncr_nakdata.gnnd_errno);
888 /* since we use a unique instance ID for each network, the driver
889 * will take care of dropping datagrams if we don't have that network.
892 /* few more idiot software or configuration checks */
894 switch (connreq->gncr_type) {
895 case GNILND_CONNREQ_REQ:
896 /* wire up EP and SMSG block - this will check the incoming data
897 * and barf a NAK back if need to */
898 rc = kgnilnd_set_conn_params(dgram);
902 case GNILND_CONNREQ_NAK:
903 case GNILND_CONNREQ_CLOSE:
906 CERROR("unknown connreq packet type %d\n", connreq->gncr_type);
910 if (connreq->gncr_peerstamp == 0 || connreq->gncr_connstamp == 0) {
911 CERROR("Recived bad timestamps peer "LPU64" conn "LPU64"\n",
912 connreq->gncr_peerstamp, connreq->gncr_connstamp);
916 if (connreq->gncr_timeout < GNILND_MIN_TIMEOUT) {
917 CERROR("Received timeout %d < MIN %d\n",
918 connreq->gncr_timeout, GNILND_MIN_TIMEOUT);
926 kgnilnd_alloc_dgram(kgn_dgram_t **dgramp, kgn_device_t *dev, kgn_dgram_type_t type)
930 dgram = kmem_cache_alloc(kgnilnd_data.kgn_dgram_cache, GFP_ATOMIC);
934 /* cache alloc'd memory is not zeroed */
935 memset((void *)dgram, 0, sizeof(*dgram)) ;
937 INIT_LIST_HEAD(&dgram->gndg_list);
938 dgram->gndg_state = GNILND_DGRAM_USED;
939 dgram->gndg_type = type;
940 dgram->gndg_magic = GNILND_DGRAM_MAGIC;
942 atomic_inc(&dev->gnd_ndgrams);
944 CDEBUG(D_MALLOC|D_NETTRACE, "slab-alloced 'dgram': %lu at %p %s ndgrams"
946 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
947 atomic_read(&dev->gnd_ndgrams));
953 /* call this on a dgram that came back from kgnilnd_ep_postdata_test_by_id
954 * returns < 0 on dgram to be cleaned up
955 * > 0 on dgram that isn't done yet
956 * == 0 on dgram that is ok and needs connreq processing */
958 kgnilnd_process_dgram(kgn_dgram_t *dgram, gni_post_state_t post_state)
962 switch (post_state) {
963 case GNI_POST_COMPLETED:
964 /* normal state for dgrams that need actual processing */
965 /* GOTO to avoid processing dgram as canceled/done */
966 GOTO(process_out, rc);
968 case GNI_POST_PENDING:
969 /* we should only see this if we are testing a WC dgram after a
970 * cancel - it means that it needs a full cycle of waiting
971 * for kgni_sm_task to finish moving it to TERMINATED */
972 LASSERTF((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
973 (dgram->gndg_state == GNILND_DGRAM_CANCELED),
974 "POST_PENDING dgram 0x%p with bad type %d(%s) or state %d(%s)\n",
975 dgram, dgram->gndg_type, kgnilnd_dgram_type2str(dgram),
976 dgram->gndg_state, kgnilnd_dgram_state2str(dgram));
978 /* positive RC as this dgram isn't done yet */
981 /* GOTO as this isn't done yet */
982 GOTO(process_out, rc);
985 case GNI_POST_TERMINATED:
986 /* we've called cancel and it is done or remote guy called cancel and
987 * we've receved it on a WC dgram */
989 /* we are seeing weird terminations on non WC dgrams when we have not
992 LASSERTF(dgram->gndg_state == GNILND_DGRAM_CANCELED ||
993 dgram->gndg_conn_out.gncr_dstnid == LNET_NID_ANY,
994 "dgram 0x%p with bad state %d(%s) or dst nid %s\n",
995 dgram, dgram->gndg_state, kgnilnd_dgram_state2str(dgram),
996 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
999 CDEBUG(D_NETTRACE, "dgram 0x%p saw %s, cleaning it up\n", dgram,
1000 dgram->gndg_state == GNILND_DGRAM_CANCELED ? "canceled" : "terminated");
1005 case GNI_POST_TIMEOUT:
1006 /* we could have a timeout on a wildcard dgram too - if
1007 * we got the incoming request but the remote node beefed
1008 * before kgni could send the match data back. We'll just error
1009 * on the active case and bail out gracefully */
1010 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1011 CNETERR("hardware timeout for connect to "
1012 "%s after %lu seconds. Is node dead?\n",
1013 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1014 cfs_duration_sec(jiffies - dgram->gndg_post_time));
1021 CERROR("dgram 0x%p with bad post_state %d\n", dgram, post_state);
1025 /* now finish cleaning up a dgram that is canceled/terminated and needs to
1028 /* If this was actively canceled, drop the count now that we are processing */
1029 if (dgram->gndg_state == GNILND_DGRAM_CANCELED) {
1030 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1031 /* caller responsible for gndg_list removal */
1039 /* needs dev->gnd_dgram_lock held */
1041 kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram)
1045 if (dgram->gndg_state != GNILND_DGRAM_POSTED) {
1049 LASSERTF(dgram->gndg_conn != NULL,
1050 "dgram 0x%p with NULL conn\n", dgram);
1052 /* C.E - WC dgrams could be canceled immediately but
1053 * if there was some match pending, we need to call
1054 * test_by_id to clear it out. If that test returns
1055 * POST_PENDING, it is half done and needs to go along
1056 * with the rest of dgrams and go through a kgni_sm_task cycle
1057 * and deliver a GNI_POST_TERMINATED event before they
1058 * are actually canceled */
1060 dgram->gndg_state = GNILND_DGRAM_CANCELED;
1062 if (dgram->gndg_conn->gnc_state >= GNILND_CONN_ESTABLISHED) {
1063 /* we don't need to cancel_by_id if the datagram was good */
1067 /* let folks know there are outstanding cancels */
1068 atomic_inc(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1069 /* leave on nid list until cancel is done for debugging fun */
1070 grc = kgnilnd_ep_postdata_cancel_by_id(dgram->gndg_conn->gnc_ephandle, (__u64) dgram);
1072 /* if we don't get success here, we have hosed up the dgram tracking
1073 * code and need to bail out */
1074 LASSERTF(grc == GNI_RC_SUCCESS,
1075 "postdata_cancel returned %d for conn 0x%p to %s\n",
1076 grc, dgram->gndg_conn,
1077 dgram->gndg_conn->gnc_peer ?
1078 libcfs_nid2str(dgram->gndg_conn->gnc_peer->gnp_nid)
1082 "canceled dgram 0x%p conn 0x%p ephandle 0x%p\n",
1083 dgram, dgram->gndg_conn,
1084 dgram->gndg_conn->gnc_ephandle);
1086 if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1087 gni_post_state_t post_state;
1089 __u32 remote_addr = 0, remote_id = 0;
1091 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1092 (__u64)dgram, &post_state,
1093 &remote_addr, &remote_id);
1095 LASSERTF(grc == GNI_RC_NO_MATCH || grc == GNI_RC_SUCCESS,
1096 "bad grc %d from test_by_id on dgram 0x%p\n",
1099 /* if WC was canceled immediately, we get NO_MATCH, if needs to go
1100 * through full cycle, we get SUCCESS and need to parse post_state */
1102 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1103 "remote_addr %u remote_id %u\n", grc, dgram,
1104 kgnilnd_dgram_type2str(dgram),
1105 post_state, remote_addr, remote_id);
1107 if (grc == GNI_RC_NO_MATCH) {
1108 /* she's gone, reduce count and move along */
1109 dgram->gndg_state = GNILND_DGRAM_DONE;
1110 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1114 rc = kgnilnd_process_dgram(dgram, post_state);
1117 /* if for some weird reason we get a valid dgram back, just mark as done
1118 * so we can drop it and move along.
1119 * C.E - if it was completed, we'll just release the conn/mbox
1120 * back into the pool and it'll get reused. That said, we should only
1121 * be canceling a WC dgram on stack rest or shutdown, so that is moot */
1122 dgram->gndg_state = GNILND_DGRAM_DONE;
1123 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1125 /* caller context responsible for calling kgnilnd_release_dgram() */
1127 /* still pending, let it simmer until golden brown and delicious */
1131 /* for non WC dgrams, they are still on the nid list but marked canceled waiting
1132 * for kgni to return their ID to us via probe - that is when we'll complete their
1133 * cancel processing */
1137 kgnilnd_cleanup_dgram(kgn_dgram_t *dgram)
1139 /* release the dgram ref on conn */
1140 if (dgram->gndg_conn) {
1141 kgnilnd_conn_decref(dgram->gndg_conn);
1142 dgram->gndg_conn = NULL;
1147 kgnilnd_free_dgram(kgn_device_t *dev, kgn_dgram_t *dgram)
1149 LASSERTF(dgram->gndg_state == GNILND_DGRAM_USED ||
1150 dgram->gndg_state == GNILND_DGRAM_DONE,
1151 "dgram 0x%p with bad state %s\n",
1152 dgram, kgnilnd_dgram_state2str(dgram));
1154 /* bit of poisoning to help detect bad driver data */
1155 dgram->gndg_magic = 0x6f5a6b5f;
1156 atomic_dec(&dev->gnd_ndgrams);
1158 kmem_cache_free(kgnilnd_data.kgn_dgram_cache, dgram);
1159 CDEBUG(D_MALLOC|D_NETTRACE, "slab-freed 'dgram': %lu at %p %s"
1161 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
1162 atomic_read(&dev->gnd_ndgrams));
1166 kgnilnd_post_dgram(kgn_device_t *dev, lnet_nid_t dstnid, kgn_connreq_type_t type,
1170 kgn_dgram_t *dgram = NULL;
1171 kgn_dgram_t *tmpdgram;
1172 kgn_dgram_type_t dgtype;
1178 case GNILND_CONNREQ_REQ:
1179 if (dstnid == LNET_NID_ANY)
1180 dgtype = GNILND_DGRAM_WC_REQ;
1182 dgtype = GNILND_DGRAM_REQ;
1184 case GNILND_CONNREQ_NAK:
1185 LASSERTF(dstnid != LNET_NID_ANY, "can't NAK to LNET_NID_ANY\n");
1186 dgtype = GNILND_DGRAM_NAK;
1189 CERROR("unknown connreq type %d\n", type);
1193 rc = kgnilnd_alloc_dgram(&dgram, dev, dgtype);
1196 GOTO(post_failed, rc);
1199 rc = kgnilnd_create_conn(&dgram->gndg_conn, dev);
1201 GOTO(post_failed, rc);
1204 if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1205 /* clear buffer for sanity on reuse of wildcard */
1206 memset(&dgram->gndg_conn_in, 0, sizeof(kgn_connreq_t));
1209 if (dstnid == LNET_NID_ANY) {
1210 /* set here to reset any dgram re-use */
1211 dgram->gndg_conn->gnc_state = GNILND_CONN_LISTEN;
1215 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(dstnid), 1, &host_id);
1218 GOTO(post_failed, rc);
1221 dgram->gndg_conn->gnc_state = GNILND_CONN_CONNECTING;
1223 /* don't need to serialize, there are no CQs for the dgram
1224 * EP on the kgn_net_t */
1225 grc = kgnilnd_ep_bind(dgram->gndg_conn->gnc_ephandle, host_id, dev->gnd_id);
1227 if (grc != GNI_RC_SUCCESS) {
1229 GOTO(post_failed, rc);
1234 /* If we are posting wildcards post using a net of 0, otherwise we'll use the
1235 * net of the destination node.
1238 if (dstnid == LNET_NID_ANY) {
1239 srcnid = LNET_MKNID(LNET_MKNET(GNILND, 0), dev->gnd_nid);
1241 srcnid = LNET_MKNID(LNET_NIDNET(dstnid), dev->gnd_nid);
1244 rc = kgnilnd_pack_connreq(&dgram->gndg_conn_out, dgram->gndg_conn,
1245 srcnid, dstnid, type);
1247 GOTO(post_failed, rc);
1250 if (type == GNILND_CONNREQ_NAK)
1251 dgram->gndg_conn_out.gncr_nakdata.gnnd_errno = data_rc;
1253 dgram->gndg_post_time = jiffies;
1255 /* XXX Nic: here is where we'd add in logical network multiplexing */
1257 CDEBUG(D_NETTRACE, "dgram 0x%p type %s %s->%s cdm %d\n",
1258 dgram, kgnilnd_dgram_type2str(dgram),
1259 libcfs_nid2str(srcnid),
1260 libcfs_nid2str(dstnid), dev->gnd_id);
1262 /* this allocates memory, can't hold locks across */
1263 grc = kgnilnd_ep_postdata_w_id(dgram->gndg_conn->gnc_ephandle,
1264 &dgram->gndg_conn_out, sizeof(kgn_connreq_t),
1265 &dgram->gndg_conn_in, sizeof(kgn_connreq_t),
1268 if (grc != GNI_RC_SUCCESS) {
1269 CNETERR("dropping failed dgram post id 0x%p type %s"
1270 " reqtype %s to %s: rc %d\n",
1271 dgram, kgnilnd_dgram_type2str(dgram),
1272 kgnilnd_connreq_type2str(&dgram->gndg_conn_out),
1273 libcfs_nid2str(dstnid), grc);
1274 rc = (grc == GNI_RC_ERROR_NOMEM) ? -ENOMEM : -EBADR;
1275 GOTO(post_failed, rc);
1278 /* we don't need to add earlier - if someone does del_peer during post,
1279 * that peer will get marked as unlinked and the callers wil take care of it.
1280 * The dgram code is largely kgn_peer_t ignorant, so at worst, we'll just drop
1281 * the completed dgram later when we cant find a peer to stuff it into */
1283 spin_lock(&dev->gnd_dgram_lock);
1285 /* make sure we are not double posting targeted dgrams
1286 * - we can multiple post WC dgrams to help with processing speed */
1287 if (dstnid != LNET_NID_ANY) {
1288 tmpdgram = kgnilnd_find_dgram_locked(dev, dstnid);
1290 LASSERTF(tmpdgram == NULL,
1291 "dgram 0x%p->%s already posted\n",
1292 dgram, libcfs_nid2str(dstnid));
1295 /* unmunge dstnid to help processing code cope... */
1296 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
1297 dgram->gndg_conn_out.gncr_dstnid = dstnid;
1300 list_add_tail(&dgram->gndg_list, kgnilnd_nid2dgramlist(dev, dstnid));
1301 dgram->gndg_state = GNILND_DGRAM_POSTED;
1302 spin_unlock(&dev->gnd_dgram_lock);
1305 if (rc < 0 && dgram != NULL) {
1306 kgnilnd_cleanup_dgram(dgram);
1307 kgnilnd_free_dgram(dev, dgram);
1313 /* The shutdown flag is set from the shutdown and stack reset threads. */
1315 kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown)
1317 /* The conns of canceled active dgrams need to be put in purgatory so
1318 * we don't reuse the mailbox */
1319 if (unlikely(dgram->gndg_state == GNILND_DGRAM_CANCELED)) {
1321 kgn_conn_t *conn = dgram->gndg_conn;
1322 lnet_nid_t nid = dgram->gndg_conn_out.gncr_dstnid;
1324 dgram->gndg_state = GNILND_DGRAM_DONE;
1326 /* During shutdown we've already removed the peer so we don't
1327 * need to add a peer. During stack reset we don't care about
1328 * MDDs since they are all released. */
1330 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1331 peer = kgnilnd_find_peer_locked(nid);
1334 CDEBUG(D_NET, "adding peer's conn with nid %s "
1335 "to purgatory\n", libcfs_nid2str(nid));
1336 kgnilnd_conn_addref(conn);
1337 conn->gnc_peer = peer;
1338 kgnilnd_peer_addref(peer);
1339 kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
1340 conn->gnc_state = GNILND_CONN_CLOSED;
1341 list_add_tail(&conn->gnc_list,
1343 kgnilnd_add_purgatory_locked(conn,
1345 kgnilnd_schedule_conn(conn);
1347 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1351 spin_lock(&dev->gnd_dgram_lock);
1352 kgnilnd_cancel_dgram_locked(dgram);
1353 spin_unlock(&dev->gnd_dgram_lock);
1355 kgnilnd_cleanup_dgram(dgram);
1357 /* if the dgram is 'canceled' it needs to be wait until the event
1358 * comes up from kgni that tells us it is safe to release */
1359 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1360 dgram->gndg_state = GNILND_DGRAM_DONE;
1362 LASSERTF(list_empty(&dgram->gndg_list), "dgram 0x%p on list\n", dgram);
1364 /* if it is a wildcard and we are in an appropriate state, repost
1367 if ((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
1368 (!kgnilnd_data.kgn_wc_kill && !kgnilnd_data.kgn_in_reset)) {
1371 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1373 /* We failed to repost the WC dgram for some reason
1374 * mark it so the repost system attempts to repost */
1375 kgnilnd_admin_addref(dev->gnd_nwcdgrams);
1379 /* always free the old dgram */
1380 kgnilnd_free_dgram(dev, dgram);
1386 kgnilnd_probe_for_dgram(kgn_device_t *dev, kgn_dgram_t **dgramp)
1388 kgn_dgram_t *dgram = NULL;
1389 gni_post_state_t post_state;
1393 __u32 remote_addr = 0, remote_id = 0;
1396 /* Probe with the lock held. That way if we get a dgram we dont have it canceled
1397 * between finding the ready dgram and grabbing the lock to remove it from the
1398 * list. Otherwise we could be left in an inconsistent state. We own the dgram
1399 * once its off the list so we don't need to worry about others changing it at
1401 spin_lock(&dev->gnd_dgram_lock);
1402 grc = kgnilnd_postdata_probe_by_id(dev->gnd_handle, &readyid);
1403 if (grc != GNI_RC_SUCCESS) {
1404 spin_unlock(&dev->gnd_dgram_lock);
1405 /* return 0 to indicate nothing happened */
1409 CDEBUG(D_NET, "ready "LPX64" on device 0x%p\n",
1412 dgram = (kgn_dgram_t *)readyid;
1414 LASSERTF(dgram->gndg_magic == GNILND_DGRAM_MAGIC,
1415 "dgram 0x%p from id "LPX64" with bad magic %x\n",
1416 dgram, readyid, dgram->gndg_magic);
1418 LASSERTF(dgram->gndg_state == GNILND_DGRAM_POSTED ||
1419 dgram->gndg_state == GNILND_DGRAM_CANCELED,
1420 "dgram 0x%p with bad state %s\n",
1421 dgram, kgnilnd_dgram_state2str(dgram));
1423 LASSERTF(!list_empty(&dgram->gndg_list),
1424 "dgram 0x%p with bad list state %s type %s\n",
1425 dgram, kgnilnd_dgram_state2str(dgram),
1426 kgnilnd_dgram_type2str(dgram));
1428 /* now we know that the datagram structure is ok, so pull off list */
1429 list_del_init(&dgram->gndg_list);
1431 /* while we have the gnn_dgram_lock and BEFORE we call test_by_id
1432 * change the state from POSTED to PROCESSING to ensure that
1433 * nobody cancels it after we've pulled it from the wire */
1434 if (dgram->gndg_state == GNILND_DGRAM_POSTED) {
1435 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1438 LASSERTF(dgram->gndg_conn != NULL,
1439 "dgram 0x%p with NULL conn\n", dgram);
1441 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1442 (__u64)dgram, &post_state,
1443 &remote_addr, &remote_id);
1445 /* we now "own" this datagram */
1446 spin_unlock(&dev->gnd_dgram_lock);
1448 LASSERTF(grc != GNI_RC_NO_MATCH, "kgni lied! probe_by_id told us that"
1449 " id "LPU64" was ready\n", readyid);
1451 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1452 "remote_addr %u remote_id %u\n", grc, dgram,
1453 kgnilnd_dgram_type2str(dgram),
1454 post_state, remote_addr, remote_id);
1456 if (unlikely(grc != GNI_RC_SUCCESS)) {
1457 CNETERR("getting data for dgram 0x%p->%s failed rc %d. Dropping it\n",
1458 dgram, libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1461 GOTO(probe_for_out, rc);
1464 rc = kgnilnd_process_dgram(dgram, post_state);
1466 /* we should never get probe finding a dgram for us and then it
1467 * being a WC dgram that is still in the middle of processing */
1468 LASSERTF(rc <= 0, "bad rc %d from process_dgram 0x%p state %d\n",
1469 rc, dgram, post_state);
1472 /* dgram is good enough for the data to be used */
1473 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1474 /* fake rc to mark that we've done something */
1477 /* let kgnilnd_release_dgram take care of canceled dgrams */
1478 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1479 dgram->gndg_state = GNILND_DGRAM_DONE;
1488 kgnilnd_release_dgram(dev, dgram, 0);
1493 kgnilnd_setup_wildcard_dgram(kgn_device_t *dev)
1495 /* if kgn_wildcard is zero, return error */
1496 int rc = -ENOENT, i;
1499 for (i = 0; i < *kgnilnd_tunables.kgn_nwildcard; i++) {
1500 rc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1502 CERROR("error %d: could not post wildcard datagram # %d\n",
1514 kgnilnd_cancel_net_dgrams(kgn_net_t *net)
1516 kgn_dgram_t *dg, *dgN;
1517 struct list_head zombies;
1521 /* we want to cancel any outstanding dgrams - we don't want to rely
1522 * on del_peer_or_conn catching all of them. This helps protect us in cases
1523 * where we don't quite keep the peer->dgram mapping in sync due to some
1524 * race conditions */
1526 LASSERTF(net->gnn_shutdown || kgnilnd_data.kgn_in_reset,
1527 "called with LND invalid state: net shutdown %d "
1528 "in reset %d\n", net->gnn_shutdown,
1529 kgnilnd_data.kgn_in_reset);
1531 INIT_LIST_HEAD(&zombies);
1533 spin_lock(&net->gnn_dev->gnd_dgram_lock);
1535 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1536 list_for_each_entry_safe(dg, dgN, &net->gnn_dev->gnd_dgrams[i], gndg_list) {
1538 /* skip nids not on our net or are wildcards */
1541 if (dg->gndg_type == GNILND_DGRAM_WC_REQ ||
1542 net->gnn_netnum != LNET_NETNUM(LNET_NIDNET(dg->gndg_conn_out.gncr_dstnid)))
1545 kgnilnd_cancel_dgram_locked(dg);
1549 spin_unlock(&net->gnn_dev->gnd_dgram_lock);
1555 kgnilnd_cancel_wc_dgrams(kgn_device_t *dev)
1557 kgn_dgram_t *dg, *dgN;
1558 struct list_head zombies;
1561 /* Time to kill the outstanding WC's
1562 * WC's exist on net 0 only but match on any net...
1565 LASSERTF(kgnilnd_data.kgn_in_reset || kgnilnd_data.kgn_wc_kill,
1566 "called with LND invalid state: WC shutdown %d "
1567 "in reset %d\n", kgnilnd_data.kgn_wc_kill,
1568 kgnilnd_data.kgn_in_reset);
1570 INIT_LIST_HEAD(&zombies);
1571 spin_lock(&dev->gnd_dgram_lock);
1574 dg = kgnilnd_find_dgram_locked(dev, LNET_NID_ANY);
1576 LASSERTF(dg->gndg_type == GNILND_DGRAM_WC_REQ,
1577 "dgram 0x%p->%s with bad type %d (%s)\n",
1578 dg, libcfs_nid2str(dg->gndg_conn_out.gncr_dstnid),
1579 dg->gndg_type, kgnilnd_dgram_type2str(dg));
1581 kgnilnd_cancel_dgram_locked(dg);
1583 /* WC could be DONE already, check and if so add to list to be released */
1584 if (dg->gndg_state == GNILND_DGRAM_DONE) {
1585 list_del_init(&dg->gndg_list);
1586 list_add_tail(&dg->gndg_list, &zombies);
1589 } while (dg != NULL);
1591 spin_unlock(&dev->gnd_dgram_lock);
1593 list_for_each_entry_safe(dg, dgN, &zombies, gndg_list) {
1594 list_del_init(&dg->gndg_list);
1595 kgnilnd_release_dgram(dev, dg, 1);
1602 kgnilnd_cancel_dgrams(kgn_device_t *dev)
1604 kgn_dgram_t *dg, *dgN;
1608 /* Cancel any outstanding non wildcard datagrams regardless
1609 * of which net they are on as we are in base shutdown and
1610 * dont care about connecting anymore.
1613 LASSERTF(kgnilnd_data.kgn_wc_kill == 1,"We didnt get called from base shutdown\n");
1615 spin_lock(&dev->gnd_dgram_lock);
1617 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size -1); i++) {
1618 list_for_each_entry_safe(dg, dgN, &dev->gnd_dgrams[i], gndg_list) {
1619 if (dg->gndg_type != GNILND_DGRAM_WC_REQ)
1620 kgnilnd_cancel_dgram_locked(dg);
1624 spin_unlock(&dev->gnd_dgram_lock);
1631 kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev)
1639 /* use do while to get at least one check run to allow
1640 * regression test for 762072 to hit bug if there */
1642 /* This function races with the dgram mover during shutdown so it is possible for
1643 * a dgram to be seen in kgnilnd_postdata_probe_wait_by_id but be handled in the
1644 * dgram mover thread instead of inside of this function.
1647 /* This should only be called from within shutdown, baseshutdown, or stack reset.
1648 * there are no assertions here to verify since base_shutdown has nothing in it we can check
1649 * the net is gone by then.
1654 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1655 "Waiting for %d canceled datagrams to clear on device %d\n",
1656 atomic_read(&dev->gnd_canceled_dgrams), dev->gnd_id);
1658 /* check once a second */
1659 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
1662 if (grc != GNI_RC_SUCCESS)
1665 CDEBUG(D_NET, "ready "LPX64" on device %d->0x%p\n",
1666 readyid, dev->gnd_id, dev);
1668 rc = kgnilnd_probe_for_dgram(dev, &dgram);
1670 /* if we got a valid dgram or one that is now done, clean up */
1671 kgnilnd_release_dgram(dev, dgram, 1);
1673 } while (atomic_read(&dev->gnd_canceled_dgrams));
1677 kgnilnd_start_connect(kgn_peer_t *peer)
1680 /* sync point for kgnilnd_del_peer_locked - do an early check to
1681 * catch the most common hits where del_peer is done by the
1682 * time we get here */
1683 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING1)) {
1684 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING1, 1)) {};
1687 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1688 if (!kgnilnd_peer_active(peer) || peer->gnp_connecting != GNILND_PEER_CONNECT) {
1689 /* raced with peer getting unlinked */
1690 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1694 peer->gnp_connecting = GNILND_PEER_POSTING;
1695 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1697 set_mb(peer->gnp_last_dgram_time, jiffies);
1698 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING2)) {
1699 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING2, 1)) {};
1702 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING3)) {
1703 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING3, 1)) {};
1704 rc = cfs_fail_val ? cfs_fail_val : -ENOMEM;
1706 rc = kgnilnd_post_dgram(peer->gnp_net->gnn_dev,
1707 peer->gnp_nid, GNILND_CONNREQ_REQ, 0);
1710 set_mb(peer->gnp_last_dgram_errno, rc);
1714 /* while we're posting someone could have decided this peer/dgram needed to
1715 * die a quick death, so we check for state change and process accordingly */
1717 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1718 if (!kgnilnd_peer_active(peer) || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1719 if (peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1720 peer->gnp_connecting = GNILND_PEER_KILL;
1722 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1723 /* positive RC to avoid dgram cleanup - we'll have to
1724 * wait for the kgni GNI_POST_TERMINATED event to
1725 * finish cleaning up */
1727 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev, peer->gnp_nid);
1730 peer->gnp_connecting = GNILND_PEER_POSTED;
1731 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1732 /* reaper thread will take care of any timeouts */
1733 CDEBUG(D_NET, "waiting for connect to finish to %s rc %d\n",
1734 libcfs_nid2str(peer->gnp_nid), rc);
1739 CDEBUG(D_NET, "connect to %s failed: rc %d \n",
1740 libcfs_nid2str(peer->gnp_nid), rc);
1746 kgnilnd_finish_connect(kgn_dgram_t *dgram)
1748 kgn_conn_t *conn = dgram->gndg_conn;
1749 lnet_nid_t her_nid = dgram->gndg_conn_in.gncr_srcnid;
1750 kgn_peer_t *new_peer, *peer = NULL;
1753 kgn_mbox_info_t *mbox;
1757 /* try to find a peer that matches the nid we got in the connreq
1758 * kgnilnd_unpack_connreq makes sure that conn_in.gncr_srcnid is
1759 * HER and conn_out.gncr_srcnid is ME for both active and WC dgrams */
1761 /* assume this is a new peer - it makes locking cleaner when it isn't */
1762 /* no holding kgn_net_rw_sem - already are at the kgnilnd_dgram_mover level */
1764 rc = kgnilnd_create_peer_safe(&new_peer, her_nid, NULL, GNILND_RCA_NODE_UP);
1766 CERROR("Can't create peer for %s\n", libcfs_nid2str(her_nid));
1770 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1772 /* this transfers ref from create_peer to the kgn_peer table */
1773 kgnilnd_add_peer_locked(her_nid, new_peer, &peer);
1775 /* if we found an existing peer, is it really ready for a new conn ? */
1776 if (peer != new_peer) {
1777 /* if this was an active connect attempt but we can't find a peer waiting for it
1778 * we will dump in the trash */
1780 if (peer->gnp_connecting == GNILND_PEER_IDLE && dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1781 CDEBUG(D_NET, "dropping completed connreq for %s peer 0x%p->%s\n",
1782 libcfs_nid2str(her_nid), peer, libcfs_nid2str(peer->gnp_nid));
1783 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1788 /* check to see if we can catch a connecting peer before it is
1789 * removed from the connd_peers list - if not, we need to
1790 * let the connreqs race and be handled by kgnilnd_conn_isdup_locked() */
1791 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
1792 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1793 if (!list_empty(&peer->gnp_connd_list)) {
1794 list_del_init(&peer->gnp_connd_list);
1795 /* drop connd ref */
1796 kgnilnd_peer_decref(peer);
1798 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1799 /* clear rc to make sure we don't have fake error */
1803 /* no matter what, we are no longer waiting to connect this peer now */
1804 peer->gnp_connecting = GNILND_PEER_IDLE;
1806 /* Refuse to duplicate an existing connection (both sides might try to
1807 * connect at once). NB we return success! We _are_ connected so we
1808 * _don't_ have any blocked txs to complete with failure. */
1809 rc = kgnilnd_conn_isdup_locked(peer, conn);
1811 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1812 CDEBUG(D_NET, "Not creating duplicate connection to %s: %d\n",
1813 libcfs_nid2str(her_nid), rc);
1819 if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
1820 CNETERR("Received connection request from %s that RCA thinks is"
1821 " down.\n", libcfs_nid2str(her_nid));
1822 peer->gnp_down = GNILND_RCA_NODE_UP;
1825 nstale = kgnilnd_close_stale_conns_locked(peer, conn);
1827 /* either way with peer (new or existing), we are ok with ref counts here as the
1828 * kgnilnd_add_peer_locked will use our ref on new_peer (from create_peer_safe) as the
1829 * ref for the peer table. */
1831 /* at this point, the connection request is a winner */
1833 /* mark 'DONE' to avoid cancel being called from release */
1834 dgram->gndg_state = GNILND_DGRAM_DONE;
1836 /* initialise timestamps before reaper looks at them */
1837 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
1839 /* last_tx is initialized to jiffies - (keepalive*2) so that if the NOOP fails it will
1840 * immediatly send a NOOP in the reaper thread during the call to
1841 * kgnilnd_check_conn_timeouts_locked
1843 conn->gnc_last_tx = jiffies - (cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout)) * 2);
1844 conn->gnc_state = GNILND_CONN_ESTABLISHED;
1846 /* save the dgram type used to establish this connection */
1847 conn->gnc_dgram_type = dgram->gndg_type;
1849 /* refs are not transferred from dgram to tables, so increment to
1851 kgnilnd_conn_addref(conn);
1852 kgnilnd_peer_addref(peer);
1853 conn->gnc_peer = peer;
1854 list_add_tail(&conn->gnc_list, &peer->gnp_conns);
1856 kgnilnd_conn_addref(conn); /* +1 ref for conn table */
1857 list_add_tail(&conn->gnc_hashlist,
1858 kgnilnd_cqid2connlist(conn->gnc_cqid));
1859 kgnilnd_data.kgn_conn_version++;
1861 /* Dont send NOOP if fail_loc is set
1863 if (!CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP)) {
1864 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, peer->gnp_net->gnn_ni->ni_nid);
1866 CNETERR("can't get TX to initiate NOOP to %s\n",
1867 libcfs_nid2str(peer->gnp_nid));
1869 kgnilnd_queue_tx(conn, tx);
1873 /* Schedule all packets blocking for a connection */
1874 list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1875 /* lock held here is the peer_conn lock */
1876 kgnilnd_tx_del_state_locked(tx, peer, NULL, GNILND_TX_ALLOCD);
1877 kgnilnd_queue_tx(conn, tx);
1880 /* If this is an active connection lets mark its timestamp on the MBoX */
1881 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1882 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1883 /* conn->gnc_last_rx is jiffies it better exist as it was just set */
1884 mbox->mbx_release_purg_active_dgram = conn->gnc_last_rx;
1887 /* Bug 765042: wake up scheduler for a race with finish_connect and
1888 * complete_conn_closed with a conn in purgatory
1889 * since we can't use CFS_RACE due to mutex_holds in kgnilnd_process_conns,
1890 * we just check for set and then clear */
1891 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
1893 /* get scheduler thread moving again */
1894 kgnilnd_schedule_device(conn->gnc_device);
1897 CDEBUG(D_NET, "New conn 0x%p->%s dev %d\n",
1898 conn, libcfs_nid2str(her_nid), conn->gnc_device->gnd_id);
1900 /* make sure we reset peer reconnect interval now that we have a good conn */
1901 kgnilnd_peer_alive(peer);
1902 peer->gnp_reconnect_interval = 0;
1904 /* clear the unlink attribute if we dont clear it kgnilnd_del_conn_or_peer will wait
1905 * on the atomic forever
1907 if (peer->gnp_pending_unlink) {
1908 peer->gnp_pending_unlink = 0;
1909 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1910 CDEBUG(D_NET, "Clearing peer unlink %p\n",peer);
1913 /* add ref to make it hang around until after we drop the lock */
1914 kgnilnd_conn_addref(conn);
1916 /* Once the peer_conn lock is dropped, the conn could actually move into
1917 * CLOSING->CLOSED->DONE in the scheduler thread, so hold the
1918 * lock until we are really done */
1919 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1921 /* Notify LNET that we now have a working connection to this peer.
1922 * This is a Cray extension to the "standard" LND behavior. */
1923 lnet_notify(peer->gnp_net->gnn_ni, peer->gnp_nid,
1924 1, cfs_time_current());
1926 /* drop our 'hold' ref */
1927 kgnilnd_conn_decref(conn);
1934 kgnilnd_send_nak(kgn_device_t *dev, lnet_nid_t dst_nid, int error)
1939 LASSERTF(dst_nid != LNET_NID_ANY, "bad dst_nid %s\n", libcfs_nid2str(dst_nid));
1941 CDEBUG(D_NET, "NAK to %s errno %d\n", libcfs_nid2str(dst_nid), error);
1943 rc = kgnilnd_post_dgram(dev, dst_nid, GNILND_CONNREQ_NAK, error);
1946 CDEBUG(D_NET, "NAK to %s failed: rc %d \n", libcfs_nid2str(dst_nid), rc);
1952 kgnilnd_process_nak(kgn_dgram_t *dgram)
1954 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
1955 lnet_nid_t src_nid = connreq->gncr_srcnid;
1956 int errno = connreq->gncr_nakdata.gnnd_errno;
1960 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1962 peer = kgnilnd_find_peer_locked(src_nid);
1964 /* we likely dropped him from bad data when we processed
1965 * the original REQ */
1966 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1970 /* need to check peerstamp/connstamp against the ones we find
1971 * to make sure we don't close new (and good?) conns that we
1972 * formed after this connreq failed */
1973 if (peer->gnp_connecting == GNILND_PEER_IDLE) {
1976 if (list_empty(&peer->gnp_conns)) {
1977 /* assume already procced datagram and it barfed up
1978 * on this side too */
1979 CDEBUG(D_NET, "dropping NAK from %s; "
1980 "peer %s is already not connected\n",
1981 libcfs_nid2str(connreq->gncr_srcnid),
1982 libcfs_nid2str(connreq->gncr_dstnid));
1983 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1987 /* stub up a connection with the connreq XXX_stamps to allow
1988 * use to use close_stale_conns_locked */
1989 conn.gnc_peerstamp = connreq->gncr_peerstamp;
1990 conn.gnc_my_connstamp = connreq->gncr_connstamp;
1991 conn.gnc_peer_connstamp = connreq->gncr_connstamp;
1992 conn.gnc_device = peer->gnp_net->gnn_dev;
1994 rc = kgnilnd_close_stale_conns_locked(peer, &conn);
1996 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
1997 "closed %d connections\n",
1998 libcfs_nid2str(connreq->gncr_srcnid),
1999 libcfs_nid2str(connreq->gncr_dstnid), errno, rc);
2001 spin_lock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2003 if (list_empty(&peer->gnp_connd_list)) {
2004 /* if peer isn't on waiting list, try to find one to nuke */
2005 rc = kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
2009 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
2010 "canceled pending connect request\n",
2011 libcfs_nid2str(connreq->gncr_srcnid),
2012 libcfs_nid2str(connreq->gncr_dstnid), errno);
2015 /* if we can't find a waiting dgram, we just drop the nak - the conn
2016 * connect must have failed (didn't find conn above and clear connecting
2017 * -- so nothing to do besides drop */
2019 /* peer is on list, meaning it is a new connect attempt from the one
2020 * we started that generated the NAK - so just drop NAK */
2022 /* use negative to prevent error message */
2025 spin_unlock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2028 /* success! we found a peer and at least marked pending_nak */
2029 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2035 kgnilnd_process_connreq(kgn_dgram_t *dgram, int *needs_nak)
2039 rc = kgnilnd_unpack_connreq(dgram);
2042 /* only NAK if we have good srcnid to use */
2048 switch (dgram->gndg_conn_in.gncr_type) {
2049 case GNILND_CONNREQ_REQ:
2050 /* wire up peer & conn, send queued TX */
2051 rc = kgnilnd_finish_connect(dgram);
2053 /* don't nak when the nid is hosed */
2059 case GNILND_CONNREQ_NAK:
2060 rc = kgnilnd_process_nak(dgram);
2061 /* return early to prevent reconnect bump */
2064 CERROR("unexpected connreq type %s (%d) from %s\n",
2065 kgnilnd_connreq_type2str(&dgram->gndg_conn_in),
2066 dgram->gndg_conn_in.gncr_type,
2067 libcfs_nid2str(dgram->gndg_conn_in.gncr_srcnid));
2078 kgnilnd_probe_and_process_dgram(kgn_device_t *dev)
2082 lnet_nid_t nak_dstnid = LNET_NID_ANY;
2083 lnet_nid_t orig_dstnid;
2084 kgn_dgram_t *dgram = NULL;
2088 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PAUSE_DGRAM_COMP)) {
2091 rc = kgnilnd_probe_for_dgram(dev, &dgram);
2096 } else if (rc < 0) {
2097 GOTO(inform_peer, rc);
2099 /* rc > 1 means it did something, reset for this func */
2103 switch (dgram->gndg_type) {
2104 case GNILND_DGRAM_WC_REQ:
2105 case GNILND_DGRAM_REQ:
2106 rc = kgnilnd_process_connreq(dgram, &needs_nak);
2108 case GNILND_DGRAM_NAK:
2109 CDEBUG(D_NETTRACE, "NAK to %s done\n",
2110 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
2113 CERROR("unknown datagram type %s (%d)\n",
2114 kgnilnd_dgram_type2str(dgram), dgram->gndg_type);
2118 /* stash data to use after releasing current datagram */
2119 /* don't stash net - we are operating on a net already,
2120 * so the lock on rw_net_lock is sufficient */
2122 nak_dstnid = dgram->gndg_conn_in.gncr_srcnid;
2125 LASSERTF(dgram != NULL, "dgram 0x%p rc %d needs_nak %d\n", dgram, rc, needs_nak);
2127 orig_dstnid = dgram->gndg_conn_out.gncr_dstnid;
2129 kgnilnd_release_dgram(dev, dgram, 0);
2131 CDEBUG(D_NET, "cleaning up dgram to %s, rc %d\n",
2132 libcfs_nid2str(orig_dstnid), rc);
2134 /* if this was a WC_REQ that matched an existing peer, it'll get marked done
2135 * in kgnilnd_finish_connect - if errors are from before we get to there,
2136 * we just drop as it is a WC_REQ - the peer CAN'T be waiting for it */
2137 if ((orig_dstnid != LNET_NID_ANY) && (rc < 0)) {
2138 /* if we have a negative rc, we want to find a peer to inform about
2139 * the bad connection attempt. Sorry buddy, better luck next time! */
2141 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2142 peer = kgnilnd_find_peer_locked(orig_dstnid);
2145 /* add ref to make sure he stays around past the possible unlink
2146 * so we can tell LNet about him */
2147 kgnilnd_peer_addref(peer);
2149 /* if he still cares about the outstanding connect */
2150 if (peer->gnp_connecting >= GNILND_PEER_CONNECT) {
2151 /* check if he is on the connd list and remove.. */
2152 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2153 if (!list_empty(&peer->gnp_connd_list)) {
2154 list_del_init(&peer->gnp_connd_list);
2155 /* drop connd ref */
2156 kgnilnd_peer_decref(peer);
2158 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2160 /* clear gnp_connecting so we don't have a non-connecting peer
2161 * on gnd_connd_list */
2162 peer->gnp_connecting = GNILND_PEER_IDLE;
2164 set_mb(peer->gnp_last_dgram_errno, rc);
2166 kgnilnd_peer_increase_reconnect_locked(peer);
2169 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2171 /* now that we are outside the lock, tell Mommy */
2173 kgnilnd_peer_notify(peer, rc);
2174 kgnilnd_peer_decref(peer);
2179 kgnilnd_send_nak(dev, nak_dstnid, rc);
2186 kgnilnd_reaper_dgram_check(kgn_device_t *dev)
2188 kgn_dgram_t *dgram, *tmp;
2191 spin_lock(&dev->gnd_dgram_lock);
2193 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
2194 list_for_each_entry_safe(dgram, tmp, &dev->gnd_dgrams[i], gndg_list) {
2195 unsigned long now = jiffies;
2196 unsigned long timeout;
2198 /* don't timeout stuff if the network is mucked or shutting down */
2199 if (kgnilnd_check_hw_quiesce()) {
2203 if ((dgram->gndg_state != GNILND_DGRAM_POSTED) ||
2204 (dgram->gndg_type == GNILND_DGRAM_WC_REQ)) {
2207 CDEBUG(D_NETTRACE, "checking dgram 0x%p type %s "
2208 "state %s conn 0x%p to %s age %lus\n",
2209 dgram, kgnilnd_dgram_type2str(dgram),
2210 kgnilnd_dgram_state2str(dgram), dgram->gndg_conn,
2211 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2212 cfs_duration_sec(now - dgram->gndg_post_time));
2214 timeout = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout);
2216 if (time_before(now, (dgram->gndg_post_time + timeout)))
2219 CNETERR("%s datagram to %s timed out @ %lus dgram "
2220 "0x%p state %s conn 0x%p\n",
2221 kgnilnd_dgram_type2str(dgram),
2222 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2223 cfs_duration_sec(now - dgram->gndg_post_time),
2224 dgram, kgnilnd_dgram_state2str(dgram),
2227 kgnilnd_cancel_dgram_locked(dgram);
2230 spin_unlock(&dev->gnd_dgram_lock);
2234 /* use a thread for the possibly long-blocking wait_by_id to prevent
2235 * stalling the global workqueues */
2237 kgnilnd_dgram_waitq(void *arg)
2239 kgn_device_t *dev = (kgn_device_t *) arg;
2243 DEFINE_WAIT(mover_done);
2245 snprintf(name, sizeof(name), "kgnilnd_dgn_%02d", dev->gnd_id);
2246 cfs_block_allsigs();
2248 /* all gnilnd threads need to run fairly urgently */
2249 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2251 /* we dont shut down until the device shuts down ... */
2252 while (!kgnilnd_data.kgn_shutdown) {
2253 /* to quiesce or to not quiesce, that is the question */
2254 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2255 KGNILND_SPIN_QUIESCE;
2258 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_DGRAM_COMP, 1)) {}
2260 /* check once a second */
2261 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
2264 if (grc == GNI_RC_SUCCESS) {
2265 CDEBUG(D_INFO, "waking up dgram mover thread\n");
2266 kgnilnd_schedule_dgram(dev);
2268 /* wait for dgram thread to ping us before spinning again */
2269 prepare_to_wait(&dev->gnd_dgping_waitq, &mover_done,
2270 TASK_INTERRUPTIBLE);
2272 /* don't sleep if we need to quiesce */
2273 if (likely(!kgnilnd_data.kgn_quiesce_trigger)) {
2276 finish_wait(&dev->gnd_dgping_waitq, &mover_done);
2280 kgnilnd_thread_fini();
2285 kgnilnd_start_outbound_dgrams(kgn_device_t *dev, unsigned long deadline)
2287 int did_something = 0, rc;
2288 kgn_peer_t *peer = NULL;
2290 spin_lock(&dev->gnd_connd_lock);
2292 /* Active connect - we added this in kgnilnd_launch_tx */
2293 while (!list_empty(&dev->gnd_connd_peers) && time_before(jiffies, deadline)) {
2294 peer = list_first_entry(&dev->gnd_connd_peers,
2295 kgn_peer_t, gnp_connd_list);
2297 /* ref for connd removed in if/else below */
2298 list_del_init(&peer->gnp_connd_list);
2300 /* gnp_connecting and membership on gnd_connd_peers should be
2301 * done coherently to avoid double adding, etc */
2302 /* don't need kgnilnd_data.kgn_peer_conn_lock here as that is only needed
2303 * to get the peer to gnp_connecting in the first place. We just need to
2304 * rely on gnd_connd_lock to serialize someone pulling him from the list
2305 * BEFORE clearing gnp_connecting */
2306 LASSERTF(peer->gnp_connecting != GNILND_PEER_IDLE, "peer 0x%p->%s not connecting\n",
2307 peer, libcfs_nid2str(peer->gnp_nid));
2309 spin_unlock(&dev->gnd_connd_lock);
2311 CDEBUG(D_NET, "processing connect to %s\n",
2312 libcfs_nid2str(peer->gnp_nid));
2315 rc = kgnilnd_start_connect(peer);
2317 if (likely(rc >= 0)) {
2318 /* 0 on success, positive on 'just drop peer' errors */
2319 kgnilnd_peer_decref(peer);
2320 } else if (rc == -ENOMEM) {
2321 /* if we are out of wildcards, add back to
2322 * connd_list - then break out and we'll try later
2323 * if other errors, we'll bail & cancel pending tx */
2324 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2325 if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2326 peer->gnp_connecting = GNILND_PEER_CONNECT;
2327 spin_lock(&dev->gnd_connd_lock);
2328 list_add_tail(&peer->gnp_connd_list,
2329 &dev->gnd_connd_peers);
2331 /* connecting changed while we were posting */
2333 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2334 " state 0x%p->%s, connecting %d\n",
2335 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2336 peer->gnp_connecting = GNILND_PEER_KILL;
2337 spin_lock(&dev->gnd_connd_lock);
2338 /* remove the peer ref frrom the cond list */
2339 kgnilnd_peer_decref(peer);
2340 /* let the system handle itself */
2342 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2343 /* the datagrams are a global pool,
2344 * so break out of trying and hope some free
2349 /* something bad happened, you lose */
2350 CNETERR("could not start connecting to %s "
2351 "rc %d: Will retry until TX timeout\n",
2352 libcfs_nid2str(peer->gnp_nid), rc);
2353 /* It didnt post so just set connecting back to zero now.
2354 * The reaper will reattempt the connection if it needs too.
2355 * If the peer needs death set it so the reaper will cleanup.
2357 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2358 if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2359 peer->gnp_connecting = GNILND_PEER_IDLE;
2360 kgnilnd_peer_increase_reconnect_locked(peer);
2362 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2363 " state 0x%p->%s, connecting %d\n",
2364 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2365 peer->gnp_connecting = GNILND_PEER_KILL;
2367 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2369 /* hold onto ref until we are really done - if it was
2370 * unlinked this could result in a destroy */
2371 kgnilnd_peer_decref(peer);
2373 spin_lock(&dev->gnd_connd_lock);
2376 spin_unlock(&dev->gnd_connd_lock);
2377 RETURN(did_something);
2381 kgnilnd_repost_wc_dgrams(kgn_device_t *dev)
2383 int did_something = 0, to_repost, i;
2384 to_repost = atomic_read(&dev->gnd_nwcdgrams);
2387 for (i = 0; i < to_repost; ++i) {
2389 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
2391 kgnilnd_admin_decref(dev->gnd_nwcdgrams);
2394 CDEBUG(D_NETERROR, "error %d: dev %d could not post wildcard datagram\n",
2400 RETURN(did_something);
2404 kgnilnd_dgram_poke_with_stick(unsigned long arg)
2407 kgn_device_t *dev = &kgnilnd_data.kgn_devices[dev_id];
2409 wake_up(&dev->gnd_dgram_waitq);
2412 /* use single thread for dgrams - should be sufficient for performance */
2414 kgnilnd_dgram_mover(void *arg)
2416 kgn_device_t *dev = (kgn_device_t *)arg;
2418 int rc, did_something;
2419 unsigned long next_purge_check = jiffies - 1;
2420 unsigned long timeout;
2421 struct timer_list timer;
2422 unsigned long deadline = 0;
2425 snprintf(name, sizeof(name), "kgnilnd_dg_%02d", dev->gnd_id);
2426 cfs_block_allsigs();
2427 /* all gnilnd threads need to run fairly urgently */
2428 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2430 /* we are ok not locking for these variables as the dgram waitq threads
2431 * will block both due to tying up net (kgn_shutdown) and the completion
2432 * event for the dgram_waitq (kgn_quiesce_trigger) */
2433 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2434 while (!kgnilnd_data.kgn_shutdown) {
2435 /* Safe: kgn_shutdown only set when quiescent */
2437 /* race with stack reset - we want to hold off seeing any new incoming dgrams
2438 * so we can force a dirty WC dgram for Bug 762072 - put right before
2439 * quiesce check so that it'll go right into that and not do any
2441 CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);
2443 /* to quiesce or to not quiesce, that is the question */
2444 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2445 KGNILND_SPIN_QUIESCE;
2449 CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);
2451 /* process any newly completed dgrams */
2452 down_read(&kgnilnd_data.kgn_net_rw_sem);
2454 rc = kgnilnd_probe_and_process_dgram(dev);
2456 did_something += rc;
2459 up_read(&kgnilnd_data.kgn_net_rw_sem);
2461 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_DGRAM_DEADLINE,
2462 (*kgnilnd_tunables.kgn_dgram_timeout + 1));
2463 /* start new outbound dgrams */
2464 did_something += kgnilnd_start_outbound_dgrams(dev, deadline);
2466 /* find dead dgrams */
2467 if (time_after_eq(jiffies, next_purge_check)) {
2468 /* these don't need to be checked that often */
2469 kgnilnd_reaper_dgram_check(dev);
2471 next_purge_check = (long) jiffies +
2472 cfs_time_seconds(kgnilnd_data.kgn_new_min_timeout / 4);
2475 did_something += kgnilnd_repost_wc_dgrams(dev);
2477 /* careful with the jiffy wrap... */
2478 timeout = (long)(next_purge_check - jiffies);
2480 CDEBUG(D_INFO, "did %d timeout %lu next %lu jiffies %lu\n",
2481 did_something, timeout, next_purge_check, jiffies);
2483 if ((did_something || timeout <= 0) && time_before(jiffies, deadline)) {
2488 prepare_to_wait(&dev->gnd_dgram_waitq, &wait, TASK_INTERRUPTIBLE);
2490 setup_timer(&timer, kgnilnd_dgram_poke_with_stick, dev->gnd_id);
2491 mod_timer(&timer, (long) jiffies + timeout);
2493 /* last second chance for others to poke us */
2494 did_something += xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_IDLE);
2496 /* check flag variables before comittingi even if we did something;
2497 * if we are after the deadline call schedule */
2498 if ((!did_something || time_after(jiffies, deadline)) &&
2499 !kgnilnd_data.kgn_shutdown &&
2500 !kgnilnd_data.kgn_quiesce_trigger) {
2501 CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2502 timeout, cfs_duration_sec(timeout));
2503 wake_up_all(&dev->gnd_dgping_waitq);
2505 CDEBUG(D_INFO, "awake after schedule\n");
2506 deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2509 del_singleshot_timer_sync(&timer);
2510 finish_wait(&dev->gnd_dgram_waitq, &wait);
2513 kgnilnd_thread_fini();