lnet/klnds/gnilnd/gnilnd_conn.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* Copyright (C) 2012 Cray, Inc.
   4  *
   5  * Copyright (c) 2014, Intel Corporation.
   6  */
   7
   8 /* This file is part of Lustre, http://www.lustre.org.
   9  *
  10  * Author: Nic Henke <nic@cray.com>
  11  * Author: James Shimek <jshimek@cray.com>
  12  */
  13
  14 #include "gnilnd.h"
  15 #include <linux/swap.h>
  16
  17 void
  18 kgnilnd_setup_smsg_attr(gni_smsg_attr_t *smsg_attr)
  19 {
  20         smsg_attr->mbox_maxcredit = *kgnilnd_tunables.kgn_mbox_credits;
  21         smsg_attr->msg_maxsize = GNILND_MAX_MSG_SIZE;
  22         smsg_attr->msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
  23 }
  24
  25 int
  26 kgnilnd_map_fmablk(kgn_device_t *device, kgn_fma_memblock_t *fma_blk)
  27 {
  28         gni_return_t            rrc;
  29         __u32                   flags = GNI_MEM_READWRITE;
  30         static unsigned long    reg_to;
  31         int                     rfto = *kgnilnd_tunables.kgn_reg_fail_timeout;
  32
  33         if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
  34                 flags |= GNI_MEM_PHYS_CONT;
  35         }
  36
  37         fma_blk->gnm_hold_timeout = 0;
  38
  39         /* make sure we are mapping a clean block */
  40         LASSERTF(fma_blk->gnm_hndl.qword1 == 0UL,
  41                  "fma_blk %px dirty\n", fma_blk);
  42
  43         rrc = kgnilnd_mem_register(device->gnd_handle, (__u64)fma_blk->gnm_block,
  44                                    fma_blk->gnm_blk_size, device->gnd_rcv_fma_cqh,
  45                                    flags, &fma_blk->gnm_hndl);
  46         if (rrc != GNI_RC_SUCCESS) {
  47                 if (rfto != GNILND_REGFAILTO_DISABLE) {
  48                         if (reg_to == 0)
  49                                 reg_to = jiffies + cfs_time_seconds(rfto);
  50                         else
  51                                 LASSERTF(!time_after(jiffies, reg_to),
  52                                          "FATAL:fmablk registration has failed for %ld seconds.\n",
  53                                          cfs_duration_sec(jiffies - reg_to) + rfto);
  54                 }
  55
  56                 CNETERR("register fmablk failed 0x%p mbox_size %d flags %u\n",
  57                         fma_blk, fma_blk->gnm_mbox_size, flags);
  58                 RETURN(-ENOMEM);
  59         }
  60
  61         reg_to = 0;
  62
  63         /* PHYS_CONT memory isn't really mapped, at least not in GART -
  64          *  but all mappings chew up a MDD
  65          */
  66         if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
  67                 atomic64_add(fma_blk->gnm_blk_size, &device->gnd_nbytes_map);
  68         }
  69
  70         atomic_inc(&device->gnd_n_mdd);
  71         /* nfmablk is live (mapped) blocks */
  72         atomic_inc(&device->gnd_nfmablk);
  73
  74         RETURN(0);
  75 }
  76
  77 int
  78 kgnilnd_alloc_fmablk(kgn_device_t *device, int use_phys)
  79 {
  80         int                     rc = 0;
  81         int                     num_mbox;
  82         kgn_fma_memblock_t     *fma_blk;
  83         gni_smsg_attr_t         smsg_attr;
  84         unsigned long           fmablk_vers;
  85
  86 #if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
  87         /* We allocate large blocks of memory here potentially leading
  88          * to memory exhaustion during massive reconnects during a network
  89          * outage. Limit the amount of fma blocks to use by always keeping
  90          * a percent of pages free initially set to 25% of total memory. */
  91         if (nr_free_pages() < kgnilnd_data.free_pages_limit) {
  92                 LCONSOLE_INFO("Exceeding free page limit of %ld. "
  93                               "Free pages available %ld\n",
  94                               kgnilnd_data.free_pages_limit,
  95                               nr_free_pages());
  96                 return -ENOMEM;
  97         }
  98 #endif
  99         /* we'll use fmablk_vers and the gnd_fmablk_mutex to gate access
 100          * to this allocation code. Everyone will sample the version
 101          * before and after getting the mutex. If it has changed,
 102          * we'll bail out to check the lists again - this indicates that
 103          * some sort of change was made to the lists and it is possible
 104          * that there is a mailbox for us to find now. This should prevent
 105          * a ton of spinning in the case where there are lots of threads
 106          * that need a yet-to-be-allocated mailbox for a connection. */
 107
 108         fmablk_vers = atomic_read(&device->gnd_fmablk_vers);
 109         mutex_lock(&device->gnd_fmablk_mutex);
 110
 111         if (fmablk_vers != atomic_read(&device->gnd_fmablk_vers)) {
 112                 /* version changed while we were waiting for semaphore,
 113                  * we'll recheck the lists assuming something nice happened */
 114                 mutex_unlock(&device->gnd_fmablk_mutex);
 115                 return 0;
 116         }
 117
 118         LIBCFS_ALLOC(fma_blk, sizeof(kgn_fma_memblock_t));
 119         if (fma_blk == NULL) {
 120                 CNETERR("could not allocate fma block descriptor\n");
 121                 rc = -ENOMEM;
 122                 GOTO(out, rc);
 123         }
 124
 125         INIT_LIST_HEAD(&fma_blk->gnm_bufflist);
 126
 127         kgnilnd_setup_smsg_attr(&smsg_attr);
 128
 129         gni_smsg_buff_size_needed(&smsg_attr, &fma_blk->gnm_mbox_size);
 130
 131         LASSERTF(fma_blk->gnm_mbox_size, "mbox size %d\n", fma_blk->gnm_mbox_size);
 132
 133         /* gni_smsg_buff_size_needed calculates the base mailbox size and since
 134          * we want to hold kgn_peer_credits worth of messages in both directions,
 135          * we add PAYLOAD to grow the mailbox size
 136          */
 137
 138         fma_blk->gnm_mbox_size += GNILND_MBOX_PAYLOAD;
 139
 140         /* we'll only use physical during preallocate at startup -- this keeps it nice and
 141          * clean for runtime decisions. We'll keep the PHYS ones around until shutdown
 142          * as reallocating them is tough if there is memory fragmentation */
 143
 144         if (use_phys) {
 145                 fma_blk->gnm_block = kmem_cache_alloc(kgnilnd_data.kgn_mbox_cache, GFP_ATOMIC);
 146                 if (fma_blk->gnm_block == NULL) {
 147                         CNETERR("could not allocate physical SMSG mailbox memory\n");
 148                         rc = -ENOMEM;
 149                         GOTO(free_desc, rc);
 150                 }
 151                 fma_blk->gnm_blk_size = GNILND_MBOX_SIZE;
 152                 num_mbox = fma_blk->gnm_blk_size / fma_blk->gnm_mbox_size;
 153
 154                 LASSERTF(num_mbox >= 1,
 155                          "num_mbox %d blk_size %u mbox_size %d\n",
 156                           num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size);
 157
 158                 fma_blk->gnm_state = GNILND_FMABLK_PHYS;
 159
 160         } else {
 161                 num_mbox = *kgnilnd_tunables.kgn_mbox_per_block;
 162                 fma_blk->gnm_blk_size = num_mbox * fma_blk->gnm_mbox_size;
 163
 164                 LASSERTF(num_mbox >= 1 && num_mbox >= *kgnilnd_tunables.kgn_mbox_per_block,
 165                          "num_mbox %d blk_size %u mbox_size %d tunable %d\n",
 166                          num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size,
 167                          *kgnilnd_tunables.kgn_mbox_per_block);
 168
 169                 fma_blk->gnm_block = kgnilnd_vzalloc(fma_blk->gnm_blk_size);
 170                 if (fma_blk->gnm_block == NULL) {
 171                         CNETERR("could not allocate virtual SMSG mailbox memory, %d bytes\n", fma_blk->gnm_blk_size);
 172                         rc = -ENOMEM;
 173                         GOTO(free_desc, rc);
 174                 }
 175
 176                 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
 177         }
 178
 179         /* allocate just enough space for the bits to track the mailboxes */
 180         CFS_ALLOC_PTR_ARRAY(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox));
 181         if (fma_blk->gnm_bit_array == NULL) {
 182                 CNETERR("could not allocate mailbox bitmask, %lu bytes for %d mbox\n",
 183                        sizeof(unsigned long) * BITS_TO_LONGS(num_mbox), num_mbox);
 184                 rc = -ENOMEM;
 185                 GOTO(free_blk, rc);
 186         }
 187         bitmap_zero(fma_blk->gnm_bit_array, num_mbox);
 188
 189         /* now that the num_mbox is set based on allocation type, get debug
 190          * info setup
 191          * */
 192         CFS_ALLOC_PTR_ARRAY(fma_blk->gnm_mbox_info, num_mbox);
 193         if (fma_blk->gnm_mbox_info == NULL) {
 194                 CNETERR("could not allocate mailbox debug, %lu bytes for %d mbox\n",
 195                        sizeof(kgn_mbox_info_t) * num_mbox, num_mbox);
 196                 rc = -ENOMEM;
 197                 GOTO(free_bit, rc);
 198         }
 199
 200         rc = kgnilnd_map_fmablk(device, fma_blk);
 201         if (rc) {
 202                 GOTO(free_info, rc);
 203         }
 204
 205         fma_blk->gnm_next_avail_mbox = 0;
 206         fma_blk->gnm_avail_mboxs = fma_blk->gnm_num_mboxs = num_mbox;
 207
 208         CDEBUG(D_MALLOC, "alloc fmablk 0x%p num %d msg_maxsize %d credits %d "
 209                 "mbox_size %d MDD %#llx.%#llx\n",
 210                 fma_blk, num_mbox, smsg_attr.msg_maxsize, smsg_attr.mbox_maxcredit,
 211                 fma_blk->gnm_mbox_size, fma_blk->gnm_hndl.qword1,
 212                 fma_blk->gnm_hndl.qword2);
 213
 214         /* lock Is protecting data structures, not semaphore */
 215
 216         spin_lock(&device->gnd_fmablk_lock);
 217         list_add_tail(&fma_blk->gnm_bufflist, &device->gnd_fma_buffs);
 218
 219         /* toggle under the lock so once they change the list is also
 220          * ready for others to traverse */
 221         atomic_inc(&device->gnd_fmablk_vers);
 222
 223         spin_unlock(&device->gnd_fmablk_lock);
 224
 225         mutex_unlock(&device->gnd_fmablk_mutex);
 226
 227         return 0;
 228
 229 free_info:
 230         CFS_FREE_PTR_ARRAY(fma_blk->gnm_mbox_info, num_mbox);
 231 free_bit:
 232         CFS_FREE_PTR_ARRAY(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox));
 233 free_blk:
 234         if (fma_blk->gnm_state == GNILND_FMABLK_VIRT) {
 235                 kgnilnd_vfree(fma_blk->gnm_block, fma_blk->gnm_blk_size);
 236         } else {
 237                 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
 238         }
 239 free_desc:
 240         LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
 241 out:
 242         mutex_unlock(&device->gnd_fmablk_mutex);
 243         return rc;
 244 }
 245
 246 void
 247 kgnilnd_unmap_fmablk(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
 248 {
 249         gni_return_t            rrc;
 250
 251         /* if some held, set hold_timeout from conn timeouts used in this block
 252          * but not during shutdown, then just nuke and pave
 253          * During a stack reset, we need to deregister with a hold timeout
 254          * set so we don't use the same mdd after reset is complete */
 255         if ((fma_blk->gnm_held_mboxs && !kgnilnd_data.kgn_shutdown) ||
 256             kgnilnd_data.kgn_in_reset) {
 257                 fma_blk->gnm_hold_timeout = GNILND_TIMEOUT2DEADMAN;
 258         }
 259
 260         /* we are changing the state of a block, tickle version to tell
 261          * proc code list is stale now */
 262         atomic_inc(&dev->gnd_fmablk_vers);
 263
 264         rrc = kgnilnd_mem_deregister(dev->gnd_handle, &fma_blk->gnm_hndl, fma_blk->gnm_hold_timeout);
 265
 266         CDEBUG(rrc == GNI_RC_SUCCESS ? D_MALLOC : D_CONSOLE|D_NETERROR,
 267                "unmap fmablk 0x%p@%s sz %u total %d avail %d held %d mbox_size %d "
 268                 "hold_timeout %d\n",
 269                fma_blk, kgnilnd_fmablk_state2str(fma_blk->gnm_state),
 270                fma_blk->gnm_blk_size, fma_blk->gnm_num_mboxs,
 271                fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs,
 272                fma_blk->gnm_mbox_size, fma_blk->gnm_hold_timeout);
 273
 274         LASSERTF(rrc == GNI_RC_SUCCESS,
 275                 "tried to double unmap or something bad, fma_blk %px (rrc %d)\n",
 276                 fma_blk, rrc);
 277
 278         if (fma_blk->gnm_hold_timeout &&
 279             !(kgnilnd_data.kgn_in_reset &&
 280               fma_blk->gnm_state == GNILND_FMABLK_PHYS)) {
 281                 atomic_inc(&dev->gnd_n_mdd_held);
 282         } else {
 283                 atomic_dec(&dev->gnd_n_mdd);
 284         }
 285
 286         /* PHYS blocks don't get mapped */
 287         if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
 288                 atomic64_sub(fma_blk->gnm_blk_size, &dev->gnd_nbytes_map);
 289                 fma_blk->gnm_state = GNILND_FMABLK_IDLE;
 290         } else if (kgnilnd_data.kgn_in_reset) {
 291                 /* in stack reset, clear MDD handle for PHYS blocks, as we'll
 292                  * re-use the fma_blk after reset so we don't have to drop/allocate
 293                  * all of those physical blocks */
 294                 fma_blk->gnm_hndl.qword1 = fma_blk->gnm_hndl.qword2 = 0UL;
 295         }
 296
 297         /* Decrement here as this is the # of mapped blocks */
 298         atomic_dec(&dev->gnd_nfmablk);
 299 }
 300
 301
 302 /* needs lock on gnd_fmablk_lock to cover gnd_fma_buffs */
 303 void
 304 kgnilnd_free_fmablk_locked(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
 305 {
 306         LASSERTF(fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs,
 307                  "fma_blk %px@%d free in bad state (%d): blk total %d avail %d held %d\n",
 308                  fma_blk, fma_blk->gnm_state, fma_blk->gnm_hold_timeout, fma_blk->gnm_num_mboxs,
 309                 fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs);
 310
 311         atomic_inc(&dev->gnd_fmablk_vers);
 312
 313         if (fma_blk->gnm_hold_timeout) {
 314                 CDEBUG(D_MALLOC, "mdd release fmablk 0x%p sz %u avail %d held %d "
 315                         "mbox_size %d\n",
 316                         fma_blk, fma_blk->gnm_blk_size, fma_blk->gnm_avail_mboxs,
 317                         fma_blk->gnm_held_mboxs, fma_blk->gnm_mbox_size);
 318
 319                 /* We leave MDD dangling over stack reset */
 320                 if (!kgnilnd_data.kgn_in_reset) {
 321                         kgnilnd_mem_mdd_release(dev->gnd_handle, &fma_blk->gnm_hndl);
 322                 }
 323                 /* ignoring the return code - if kgni/ghal can't find it
 324                  * it must be released already */
 325                 atomic_dec(&dev->gnd_n_mdd_held);
 326                 atomic_dec(&dev->gnd_n_mdd);
 327         }
 328
 329         /* we cant' free the gnm_block until all the conns have released their
 330          * purgatory holds. While we have purgatory holds, we might check the conn
 331          * RX mailbox during the CLOSING process. It is possible that kgni might
 332          * try to look into the RX side for credits when sending the CLOSE msg too */
 333         if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
 334                 LIBCFS_MEM_MSG(fma_blk->gnm_block, fma_blk->gnm_mbox_size, "free");
 335                 kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
 336         } else {
 337                 kgnilnd_vfree(fma_blk->gnm_block, fma_blk->gnm_blk_size);
 338         }
 339         fma_blk->gnm_state = GNILND_FMABLK_FREED;
 340
 341         list_del(&fma_blk->gnm_bufflist);
 342
 343         CFS_FREE_PTR_ARRAY(fma_blk->gnm_mbox_info, fma_blk->gnm_num_mboxs);
 344         CFS_FREE_PTR_ARRAY(fma_blk->gnm_bit_array,
 345                            BITS_TO_LONGS(fma_blk->gnm_num_mboxs));
 346         LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
 347 }
 348
 349 void
 350 kgnilnd_find_free_mbox(kgn_conn_t *conn)
 351 {
 352         kgn_device_t            *dev = conn->gnc_device;
 353         gni_smsg_attr_t         *smsg_attr = &conn->gnpr_smsg_attr;
 354         kgn_fma_memblock_t      *fma_blk;
 355         kgn_mbox_info_t         *mbox = NULL;
 356         int                     id;
 357
 358         spin_lock(&dev->gnd_fmablk_lock);
 359
 360         list_for_each_entry(fma_blk, &conn->gnc_device->gnd_fma_buffs,
 361                             gnm_bufflist) {
 362                 if (fma_blk->gnm_avail_mboxs <= 0 ||
 363                     fma_blk->gnm_state <= GNILND_FMABLK_IDLE) {
 364                         continue;
 365                 }
 366                 /* look in bitarray for available mailbox */
 367                 do {
 368                         id = find_next_zero_bit(
 369                                 fma_blk->gnm_bit_array,
 370                                 fma_blk->gnm_num_mboxs,
 371                                 fma_blk->gnm_next_avail_mbox);
 372                       if (id == fma_blk->gnm_num_mboxs &&
 373                           fma_blk->gnm_next_avail_mbox != 0) {
 374                                 /* wrap around */
 375                                 fma_blk->gnm_next_avail_mbox = 0;
 376                         } else {
 377                                 break;
 378                         }
 379                 } while (1);
 380
 381                 LASSERTF(id < fma_blk->gnm_num_mboxs, "id %d max %d\n",
 382                          id, fma_blk->gnm_num_mboxs);
 383                 set_bit(id, (volatile unsigned long *)fma_blk->gnm_bit_array);
 384                 conn->gnc_mbox_id = id;
 385
 386                 fma_blk->gnm_next_avail_mbox =
 387                         (id == (fma_blk->gnm_num_mboxs - 1)) ? 0 : (id + 1);
 388                 fma_blk->gnm_avail_mboxs--;
 389                 conn->gnc_fma_blk = fma_blk;
 390
 391                 kgnilnd_setup_smsg_attr(smsg_attr);
 392
 393                 smsg_attr->msg_buffer = fma_blk->gnm_block;
 394                 smsg_attr->mbox_offset = fma_blk->gnm_mbox_size * id;
 395                 smsg_attr->mem_hndl = fma_blk->gnm_hndl;
 396                 smsg_attr->buff_size = fma_blk->gnm_mbox_size;
 397
 398                 /* We'll set the hndl to zero for PHYS blocks unmapped during stack
 399                  * reset and re-use the same fma_blk after stack reset. This ensures we've
 400                  * properly mapped it before we use it */
 401                 LASSERTF(fma_blk->gnm_hndl.qword1 != 0UL,
 402                         "unmapped fma_blk %px, state %d\n",
 403                          fma_blk, fma_blk->gnm_state);
 404
 405                 CDEBUG(D_NET, "conn %p smsg %p fmablk %p "
 406                         "allocating SMSG mbox %d buf %p "
 407                         "offset %u hndl %#llx.%#llx\n",
 408                         conn, smsg_attr, fma_blk, id,
 409                         smsg_attr->msg_buffer, smsg_attr->mbox_offset,
 410                         fma_blk->gnm_hndl.qword1,
 411                         fma_blk->gnm_hndl.qword2);
 412
 413                 mbox = &fma_blk->gnm_mbox_info[id];
 414                 mbox->mbx_create_conn_memset = jiffies;
 415                 mbox->mbx_nallocs++;
 416                 mbox->mbx_nallocs_total++;
 417
 418                 /* zero mbox to remove any old data from our last use.
 419                  * this better be safe, if not our purgatory timers
 420                  * are too short or a peer really is misbehaving */
 421                 memset(smsg_attr->msg_buffer + smsg_attr->mbox_offset,
 422                        0, smsg_attr->buff_size);
 423                 break;
 424         }
 425
 426         spin_unlock(&dev->gnd_fmablk_lock);
 427 }
 428
 429 int
 430 kgnilnd_setup_mbox(kgn_conn_t *conn)
 431 {
 432         gni_smsg_attr_t         *smsg_attr = &conn->gnpr_smsg_attr;
 433         int                      err = 0;
 434
 435         smsg_attr->msg_buffer = NULL;
 436         /* Look for available mbox */
 437         do {
 438                 kgnilnd_find_free_mbox(conn);
 439
 440                 /* nothing in the existing buffers, make a new one */
 441                 if (smsg_attr->msg_buffer == NULL) {
 442                         /* for runtime allocations, we only want vmalloc */
 443                         err = kgnilnd_alloc_fmablk(conn->gnc_device, 0);
 444                         if (err) {
 445                                 break;
 446                         }
 447                 }
 448         } while (smsg_attr->msg_buffer == NULL);
 449
 450         if (err)
 451                 CNETERR("couldn't allocate SMSG mbox for conn %p Error: %d\n",
 452                         conn, err);
 453         return err;
 454 }
 455
 456 void
 457 kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold)
 458 {
 459         kgn_device_t           *dev = conn->gnc_device;
 460         gni_smsg_attr_t        *smsg_attr = &conn->gnpr_smsg_attr;
 461         kgn_fma_memblock_t     *fma_blk = NULL;
 462         kgn_mbox_info_t        *mbox = NULL;
 463         int                     found = 0;
 464         int                     id;
 465
 466         /* if we failed to setup mbox and now destroying conn */
 467         if (smsg_attr->msg_buffer == NULL) {
 468                 return;
 469         }
 470
 471         id = conn->gnc_mbox_id;
 472
 473         spin_lock(&dev->gnd_fmablk_lock);
 474         /* make sure our conn points at a valid fma_blk
 475          * We use this instead of a mem block search out of smsg_attr
 476          * because we could have freed a block for fma_blk #1 but the fma_blk
 477          * is still in the list for a purgatory hold. This would induce a false
 478          * match if that same block gets reallocated to fma_blk #2 */
 479         list_for_each_entry(fma_blk, &dev->gnd_fma_buffs, gnm_bufflist) {
 480                 if (fma_blk == conn->gnc_fma_blk) {
 481                         found = 1;
 482                         break;
 483                 }
 484         }
 485         LASSERTF(found,
 486                 "unable to find conn 0x%p with gnc_fma_blk %px anywhere in the world\n",
 487                  conn, conn->gnc_fma_blk);
 488
 489         LASSERTF(id < fma_blk->gnm_num_mboxs,
 490                 "bad id %d max %d\n",
 491                 id, fma_blk->gnm_num_mboxs);
 492
 493         /* < 0 - was held, now free it
 494          * == 0 - just free it
 495          * > 0 - hold it for now */
 496         if (purgatory_hold == 0) {
 497                 CDEBUG(D_NET, "conn %p smsg %p fmablk %p freeing SMSG mbox %d "
 498                         "hndl %#llx.%#llx\n",
 499                         conn, smsg_attr, fma_blk, id,
 500                         fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
 501                 fma_blk->gnm_avail_mboxs++;
 502
 503         } else if (purgatory_hold > 0) {
 504                 CDEBUG(D_NET, "conn %p smsg %p fmablk %p holding SMSG mbox %d "
 505                         "hndl %#llx.%#llx\n",
 506                         conn, smsg_attr, fma_blk, id,
 507                         fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
 508
 509                 fma_blk->gnm_held_mboxs++;
 510                 fma_blk->gnm_max_timeout = max_t(long, fma_blk->gnm_max_timeout,
 511                                                  conn->gnc_timeout);
 512         } else {
 513                 CDEBUG(D_NET, "conn %p smsg %p fmablk %p release SMSG mbox %d "
 514                         "hndl %#llx.%#llx\n",
 515                         conn, smsg_attr, fma_blk, id,
 516                         fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
 517
 518                 fma_blk->gnm_held_mboxs--;
 519                 fma_blk->gnm_avail_mboxs++;
 520         }
 521
 522         if (purgatory_hold <= 0) {
 523                 /* if kgni is retransmitting, freeing the smsg block before the EP
 524                  * is destroyed gets messy. Bug 768295. */
 525                 LASSERTF(conn->gnc_ephandle == NULL,
 526                          "can't release mbox before EP is nuked. conn 0x%p\n", conn);
 527
 528                 mbox = &fma_blk->gnm_mbox_info[id];
 529                 mbox->mbx_release_from_purgatory = jiffies;
 530
 531                 /* clear conn gnc_fmablk if it is gone - this allows us to
 532                  * not worry about state so much in kgnilnd_destroy_conn
 533                  * and makes the guaranteed cleanup of the resources easier */
 534                 LASSERTF(test_and_clear_bit(id, fma_blk->gnm_bit_array),
 535                         "conn %px bit %d already cleared in fma_blk %px\n",
 536                          conn, id, fma_blk);
 537                 conn->gnc_fma_blk = NULL;
 538                 mbox->mbx_nallocs--;
 539         }
 540
 541         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FMABLK_AVAIL)) {
 542                 CERROR("LBUGs in your future: forcibly marking fma_blk %p "
 543                        "as mapped\n", fma_blk);
 544                 fma_blk->gnm_state = GNILND_FMABLK_VIRT;
 545         }
 546
 547         /* we don't release or unmap PHYS blocks as part of the normal cycle --
 548          * those are controlled manually from startup/shutdown */
 549         if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
 550                 /* we can unmap once all are unused (held or avail)
 551                  * but check hold_timeout to make sure we are not trying to double
 552                  * unmap this buffer. If there was no hold_timeout set due to
 553                  * held_mboxs, we'll free the mobx here shortly and won't have to
 554                  * worry about catching a double free for a 'clean' fma_blk */
 555                 if (((fma_blk->gnm_avail_mboxs + fma_blk->gnm_held_mboxs) == fma_blk->gnm_num_mboxs) &&
 556                     (!fma_blk->gnm_hold_timeout)) {
 557                         kgnilnd_unmap_fmablk(dev, fma_blk);
 558                 }
 559
 560                 /* But we can only free once they are all avail */
 561                 if (fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs &&
 562                     fma_blk->gnm_held_mboxs == 0) {
 563                         /* all mailboxes are released, free fma_blk */
 564                         kgnilnd_free_fmablk_locked(dev, fma_blk);
 565                 }
 566         }
 567
 568         spin_unlock(&dev->gnd_fmablk_lock);
 569 }
 570
 571 int
 572 kgnilnd_count_phys_mbox(kgn_device_t *device)
 573 {
 574         int                     i = 0;
 575         kgn_fma_memblock_t     *fma_blk;
 576
 577         spin_lock(&device->gnd_fmablk_lock);
 578
 579         list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
 580                 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
 581                         i += fma_blk->gnm_num_mboxs;
 582         }
 583         spin_unlock(&device->gnd_fmablk_lock);
 584
 585         RETURN(i);
 586 }
 587
 588 int
 589 kgnilnd_allocate_phys_fmablk(kgn_device_t *device)
 590 {
 591         int     rc;
 592
 593         while (kgnilnd_count_phys_mbox(device) < *kgnilnd_tunables.kgn_nphys_mbox) {
 594
 595                 rc = kgnilnd_alloc_fmablk(device, 1);
 596                 if (rc) {
 597                         CERROR("failed phys mbox allocation, stopping at %d, rc %d\n",
 598                                 kgnilnd_count_phys_mbox(device), rc);
 599                         RETURN(rc);
 600                 }
 601         }
 602         RETURN(0);
 603 }
 604
 605 int
 606 kgnilnd_map_phys_fmablk(kgn_device_t *device)
 607 {
 608
 609         int                     rc = 0;
 610         kgn_fma_memblock_t     *fma_blk;
 611
 612         /* use mutex to gate access to single thread, just in case */
 613         mutex_lock(&device->gnd_fmablk_mutex);
 614
 615         spin_lock(&device->gnd_fmablk_lock);
 616
 617         list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
 618                 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
 619                         rc = kgnilnd_map_fmablk(device, fma_blk);
 620                         if (rc)
 621                                 break;
 622                 }
 623         }
 624         spin_unlock(&device->gnd_fmablk_lock);
 625
 626         mutex_unlock(&device->gnd_fmablk_mutex);
 627
 628         RETURN(rc);
 629 }
 630
 631 void
 632 kgnilnd_unmap_fma_blocks(kgn_device_t *device)
 633 {
 634
 635         kgn_fma_memblock_t      *fma_blk;
 636
 637         /* use mutex to gate access to single thread, just in case */
 638         mutex_lock(&device->gnd_fmablk_mutex);
 639
 640         spin_lock(&device->gnd_fmablk_lock);
 641
 642         list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
 643                 kgnilnd_unmap_fmablk(device, fma_blk);
 644         }
 645         spin_unlock(&device->gnd_fmablk_lock);
 646
 647         mutex_unlock(&device->gnd_fmablk_mutex);
 648 }
 649
 650 void
 651 kgnilnd_free_phys_fmablk(kgn_device_t *device)
 652 {
 653
 654         kgn_fma_memblock_t      *fma_blk, *fma_blkN;
 655
 656         /* use mutex to gate access to single thread, just in case */
 657         mutex_lock(&device->gnd_fmablk_mutex);
 658
 659         spin_lock(&device->gnd_fmablk_lock);
 660
 661         list_for_each_entry_safe(fma_blk, fma_blkN, &device->gnd_fma_buffs, gnm_bufflist) {
 662                 if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
 663                         kgnilnd_free_fmablk_locked(device, fma_blk);
 664         }
 665         spin_unlock(&device->gnd_fmablk_lock);
 666
 667         mutex_unlock(&device->gnd_fmablk_mutex);
 668 }
 669
 670 /* kgnilnd dgram nid->struct managment */
 671
 672 static inline struct list_head *
 673 kgnilnd_nid2dgramlist(kgn_device_t *dev, lnet_nid_t nid)
 674 {
 675         unsigned int hash = ((unsigned int)nid) % *kgnilnd_tunables.kgn_peer_hash_size;
 676
 677         RETURN(&dev->gnd_dgrams[hash]);
 678 }
 679
 680
 681 /* needs dev->gnd_dgram_lock held */
 682 kgn_dgram_t *
 683 kgnilnd_find_dgram_locked(kgn_device_t *dev, lnet_nid_t dst_nid)
 684 {
 685         struct list_head *dgram_list = kgnilnd_nid2dgramlist(dev, dst_nid);
 686         kgn_dgram_t      *dgram;
 687
 688         list_for_each_entry(dgram, dgram_list, gndg_list) {
 689
 690                 /* if state > POSTED, we are already handling cancel/completion */
 691                 if ((dgram->gndg_conn_out.gncr_dstnid != dst_nid) ||
 692                      dgram->gndg_state > GNILND_DGRAM_POSTED)
 693                         continue;
 694
 695                 CDEBUG(D_NET, "got dgram [%p] -> %s\n",
 696                        dgram, libcfs_nid2str(dst_nid));
 697                 return dgram;
 698         }
 699         return NULL;
 700 }
 701
 702 int
 703 kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid)
 704 {
 705         kgn_dgram_t     *dgram;
 706
 707         spin_lock(&dev->gnd_dgram_lock);
 708         dgram = kgnilnd_find_dgram_locked(dev, dst_nid);
 709
 710         if (dgram) {
 711                 kgnilnd_cancel_dgram_locked(dgram);
 712         }
 713         spin_unlock(&dev->gnd_dgram_lock);
 714
 715         RETURN(!!(dgram == NULL));
 716 }
 717
 718 int
 719 kgnilnd_pack_connreq(kgn_connreq_t *connreq, kgn_conn_t *conn,
 720                      lnet_nid_t srcnid, lnet_nid_t dstnid,
 721                      kgn_connreq_type_t type)
 722 {
 723         int err = 0;
 724
 725         /* ensure we haven't violated max datagram size */
 726         BUILD_BUG_ON(sizeof(kgn_connreq_t) > GNI_DATAGRAM_MAXSIZE);
 727
 728         /* no need to zero out, we do that when allocating dgram */
 729         connreq->gncr_magic     = GNILND_MSG_MAGIC;
 730
 731         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_SRCNID)) {
 732                 srcnid = 0xABADBABE;
 733         } else if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
 734                 dstnid = 0xDEFEC8ED;
 735         }
 736
 737         connreq->gncr_srcnid    = srcnid;
 738         connreq->gncr_dstnid    = dstnid;
 739
 740         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
 741                 connreq->gncr_version = 99;
 742         } else {
 743                 connreq->gncr_version   = GNILND_CONNREQ_VERSION;
 744         }
 745         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
 746                 connreq->gncr_type = 99;
 747         } else {
 748                 connreq->gncr_type      = type;
 749         }
 750         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
 751                 connreq->gncr_peerstamp = 0;
 752         } else {
 753                 connreq->gncr_peerstamp = kgnilnd_data.kgn_peerstamp;
 754         }
 755         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
 756                 connreq->gncr_connstamp = 0;
 757         } else {
 758                 connreq->gncr_connstamp = conn->gnc_my_connstamp;
 759         }
 760         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
 761                 connreq->gncr_timeout = 0;
 762         } else {
 763                 connreq->gncr_timeout   = conn->gnc_timeout;
 764         }
 765
 766         /* the rest pack the data into the payload in other places */
 767         if (type == GNILND_CONNREQ_REQ) {
 768                 kgn_gniparams_t       *req_params = &connreq->gncr_gnparams;
 769                 req_params->gnpr_host_id = conn->gnc_device->gnd_host_id;
 770                 req_params->gnpr_cqid = conn->gnc_cqid;
 771
 772                 /* allocate mailbox for this connection */
 773                 err = kgnilnd_setup_mbox(conn);
 774                 if (err != 0) {
 775                         CERROR("Failed to setup FMA mailbox (%d)\n", err);
 776                 }
 777                 req_params->gnpr_smsg_attr = conn->gnpr_smsg_attr;
 778         }
 779
 780         /* XXX Nic: TBD - checksum computation */
 781
 782         return err;
 783 }
 784
 785 int
 786 kgnilnd_unpack_connreq(kgn_dgram_t *dgram)
 787 {
 788         kgn_connreq_t           *connreq = &dgram->gndg_conn_in;
 789         int                      swab, rc = 0;
 790         kgn_net_t               *net;
 791
 792         /* the following fields must be handled in a backwards compatible
 793          * manner to ensure we can always send and interpret NAKs */
 794
 795         if (connreq->gncr_magic != GNILND_MSG_MAGIC &&
 796             connreq->gncr_magic != __swab32(GNILND_MSG_MAGIC)) {
 797                 /* Unexpected magic! */
 798                 CERROR("Unexpected magic %08x\n",
 799                        connreq->gncr_magic);
 800                 return -EBADF;
 801         }
 802
 803         swab = (connreq->gncr_magic == __swab32(GNILND_MSG_MAGIC));
 804         if (swab) {
 805                 __swab32s(&connreq->gncr_magic);
 806                 __swab32s(&connreq->gncr_cksum);
 807                 __swab16s(&connreq->gncr_type);
 808                 __swab16s(&connreq->gncr_version);
 809                 __swab32s(&connreq->gncr_timeout);
 810                 __swab64s(&connreq->gncr_srcnid);
 811                 __swab64s(&connreq->gncr_dstnid);
 812                 __swab64s(&connreq->gncr_peerstamp);
 813                 __swab64s(&connreq->gncr_connstamp);
 814         }
 815
 816         /* Do NOT return anything but -EBADF before we munge
 817          * connreq->gncr_srcnid - we need that to send the nak */
 818
 819         if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
 820                 lnet_nid_t      incoming = connreq->gncr_srcnid;
 821
 822                 /* even if the incoming packet is hosed, we know who we sent
 823                  * the original and can set the srcnid so that we can properly
 824                  * look up our peer to close the loop on this connreq. We still use
 825                  * -EBADF to prevent a NAK - just in case there are issues with
 826                  * the payload coming from a random spot, etc. */
 827                 connreq->gncr_srcnid = dgram->gndg_conn_out.gncr_dstnid;
 828
 829                 if (LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid) !=
 830                                 LNET_NIDADDR(incoming)) {
 831                         /* we got a datagram match for the wrong nid... */
 832                         CERROR("matched datagram 0x%p with srcnid %s "
 833                                 "(%x), expecting %s (%x)\n",
 834                                 dgram,
 835                                 libcfs_nid2str(incoming),
 836                                 LNET_NIDADDR(incoming),
 837                                 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
 838                                 LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid));
 839                         return -EBADF;
 840                 }
 841         } else {
 842                 /* if we have a wildcard datagram it should match an
 843                  * incoming "active" datagram that should have a fully formed
 844                  * srcnid and dstnid. If we couldn't unpack it, we drop as
 845                  * corrupted packet, otherwise we'll just verify that the dstnid
 846                  * matches the NID for the NET that the dgram was posted */
 847
 848                 /* make sure their wildcard didn't match ours, that is unpossible */
 849                 LASSERTF(connreq->gncr_dstnid != LNET_NID_ANY,
 850                          "dgram 0x%p from %s, connreq 0x%p; "
 851                          "wildcard matched wildcard \n", dgram,
 852                          libcfs_nid2str(connreq->gncr_srcnid), connreq);
 853
 854                 rc = kgnilnd_find_net(connreq->gncr_dstnid, &net);
 855
 856                 if (rc == -ESHUTDOWN) {
 857                         CERROR("Looking up network: device is in shutdown\n");
 858                         return rc;
 859                 } else if (rc == -ENONET) {
 860                         CERROR("Connection data from %s: she sent "
 861                         "dst_nid %s, but net lookup failed on "
 862                         "dgram 0x%p@%s\n",
 863                         libcfs_nid2str(connreq->gncr_srcnid),
 864                         libcfs_nid2str(connreq->gncr_dstnid),
 865                         dgram, kgnilnd_dgram_type2str(dgram));
 866                         return rc;
 867                 }
 868
 869                 if (lnet_nid_to_nid4(&net->gnn_ni->ni_nid) !=
 870                     connreq->gncr_dstnid) {
 871                         CERROR("Bad connection data from %s: she sent "
 872                                "dst_nid %s, but I am %s with dgram 0x%p@%s\n",
 873                                libcfs_nid2str(connreq->gncr_srcnid),
 874                                libcfs_nid2str(connreq->gncr_dstnid),
 875                                libcfs_nidstr(&net->gnn_ni->ni_nid),
 876                                dgram, kgnilnd_dgram_type2str(dgram));
 877                         kgnilnd_net_decref(net);
 878                         return -EBADSLT;
 879                 }
 880
 881                 /* kgnilnd_find_net takes a ref on the net it finds, You need to decref it when not needed. */
 882                 kgnilnd_net_decref(net);
 883         }
 884
 885         if (connreq->gncr_version != GNILND_CONNREQ_VERSION) {
 886                 CERROR("Unexpected version %d\n", connreq->gncr_version);
 887                 return -EPROTO;
 888         }
 889
 890         /* XXX Nic: TBD - checksum validation */
 891         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_DROP)) {
 892                 return -EBADF;
 893         }
 894
 895         if (swab && connreq->gncr_type == GNILND_CONNREQ_REQ) {
 896                 __u64 msg_addr = (__u64) connreq->gncr_gnparams.gnpr_smsg_attr.msg_buffer;
 897
 898                 __swab32s(&connreq->gncr_gnparams.gnpr_host_id);
 899                 __swab32s(&connreq->gncr_gnparams.gnpr_cqid);
 900                 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.buff_size);
 901                 __swab16s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_maxcredit);
 902                 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_offset);
 903                 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword1);
 904                 __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword2);
 905                 __swab64s(&msg_addr);
 906                 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_maxsize);
 907                 __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_type);
 908         } else if (swab && connreq->gncr_type == GNILND_CONNREQ_NAK) {
 909                 __swab32s(&connreq->gncr_nakdata.gnnd_errno);
 910         }
 911
 912         /* since we use a unique instance ID for each network, the driver
 913          * will take care of dropping datagrams if we don't have that network.
 914          */
 915
 916         /* few more idiot software or configuration checks */
 917
 918         switch (connreq->gncr_type) {
 919         case GNILND_CONNREQ_REQ:
 920                 /* wire up EP and SMSG block - this will check the incoming data
 921                  * and barf a NAK back if need to */
 922                 rc = kgnilnd_set_conn_params(dgram);
 923                 if (rc)
 924                         return rc;
 925                 break;
 926         case GNILND_CONNREQ_NAK:
 927         case GNILND_CONNREQ_CLOSE:
 928                 break;
 929         default:
 930                 CERROR("unknown connreq packet type %d\n", connreq->gncr_type);
 931                 return -EPROTO;
 932         }
 933
 934         if (connreq->gncr_peerstamp == 0 || connreq->gncr_connstamp == 0) {
 935                 CERROR("Recived bad timestamps peer %llu conn %llu\n",
 936                 connreq->gncr_peerstamp, connreq->gncr_connstamp);
 937                 return -EPROTO;
 938         }
 939
 940         if (connreq->gncr_timeout < GNILND_MIN_TIMEOUT) {
 941                 CERROR("Received timeout %d < MIN %d\n",
 942                        connreq->gncr_timeout, GNILND_MIN_TIMEOUT);
 943                 return -EPROTO;
 944         }
 945
 946         return 0;
 947 }
 948
 949 int
 950 kgnilnd_alloc_dgram(kgn_dgram_t **dgramp, kgn_device_t *dev, kgn_dgram_type_t type)
 951 {
 952         kgn_dgram_t         *dgram;
 953
 954         dgram = kmem_cache_zalloc(kgnilnd_data.kgn_dgram_cache, GFP_ATOMIC);
 955         if (dgram == NULL)
 956                 return -ENOMEM;
 957
 958         INIT_LIST_HEAD(&dgram->gndg_list);
 959         dgram->gndg_state = GNILND_DGRAM_USED;
 960         dgram->gndg_type = type;
 961         dgram->gndg_magic = GNILND_DGRAM_MAGIC;
 962
 963         atomic_inc(&dev->gnd_ndgrams);
 964
 965         CDEBUG(D_MALLOC|D_NETTRACE, "slab-alloced 'dgram': %lu at %p %s ndgrams"
 966                 " %d\n",
 967                 sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
 968                 atomic_read(&dev->gnd_ndgrams));
 969
 970         *dgramp = dgram;
 971         return 0;
 972 }
 973
 974 /* call this on a dgram that came back from kgnilnd_ep_postdata_test_by_id
 975  * returns < 0 on dgram to be cleaned up
 976  * > 0 on dgram that isn't done yet
 977  * == 0 on dgram that is ok and needs connreq processing */
 978 int
 979 kgnilnd_process_dgram(kgn_dgram_t *dgram, gni_post_state_t post_state)
 980 {
 981         int rc = 0;
 982
 983         switch (post_state) {
 984         case GNI_POST_COMPLETED:
 985                 /* normal state for dgrams that need actual processing */
 986                 /* GOTO to avoid processing dgram as canceled/done */
 987                 GOTO(process_out, rc);
 988
 989         case GNI_POST_PENDING:
 990                 /* we should only see this if we are testing a WC dgram after a
 991                  * cancel - it means that it needs a full cycle of waiting
 992                  * for kgni_sm_task to finish moving it to TERMINATED */
 993                 LASSERTF((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
 994                           (dgram->gndg_state == GNILND_DGRAM_CANCELED),
 995                          "POST_PENDING dgram 0x%p with bad type %d(%s) or state %d(%s)\n",
 996                          dgram, dgram->gndg_type, kgnilnd_dgram_type2str(dgram),
 997                          dgram->gndg_state, kgnilnd_dgram_state2str(dgram));
 998
 999                 /* positive RC as this dgram isn't done yet */
1000                 rc = EINPROGRESS;
1001
1002                 /* GOTO as this isn't done yet */
1003                 GOTO(process_out, rc);
1004                 break;
1005
1006         case GNI_POST_TERMINATED:
1007                 /* we've called cancel and it is done or remote guy called cancel and
1008                  * we've receved it on a WC dgram */
1009 #if 0
1010                 /* we are seeing weird terminations on non WC dgrams when we have not
1011                  * canceled them */
1012
1013                 LASSERTF(dgram->gndg_state == GNILND_DGRAM_CANCELED ||
1014                          dgram->gndg_conn_out.gncr_dstnid == LNET_NID_ANY,
1015                         "dgram 0x%p with bad state %d(%s) or dst nid %s\n",
1016                         dgram, dgram->gndg_state, kgnilnd_dgram_state2str(dgram),
1017                         libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
1018 #endif
1019
1020                 CDEBUG(D_NETTRACE, "dgram 0x%p saw %s, cleaning it up\n", dgram,
1021                        dgram->gndg_state == GNILND_DGRAM_CANCELED ?  "canceled" : "terminated");
1022
1023                 rc =  -ECANCELED;
1024                 break;
1025
1026         case GNI_POST_TIMEOUT:
1027                 /* we could have a timeout on a wildcard dgram too - if
1028                  * we got the incoming request but the remote node beefed
1029                  * before kgni could send the match data back. We'll just error
1030                  * on the active case and bail out gracefully */
1031                 if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1032                         CNETERR("hardware timeout for connect to "
1033                                "%s after %lu seconds. Is node dead?\n",
1034                                libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1035                                cfs_duration_sec(jiffies - dgram->gndg_post_time));
1036                 }
1037
1038                 rc = -ETIMEDOUT;
1039                 break;
1040
1041         default:
1042                 CERROR("dgram 0x%p with bad post_state %d\n", dgram, post_state);
1043                 LBUG();
1044         }
1045
1046         /* now finish cleaning up a dgram that is canceled/terminated and needs to
1047          * go away */
1048
1049         /* If this was actively canceled, drop the count now that we are processing */
1050         if (dgram->gndg_state == GNILND_DGRAM_CANCELED) {
1051                 atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1052                 /* caller responsible for gndg_list removal */
1053         }
1054
1055 process_out:
1056
1057         RETURN(rc);
1058 }
1059
1060 /* needs dev->gnd_dgram_lock held */
1061 void
1062 kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram)
1063 {
1064         gni_return_t            grc;
1065
1066         if (dgram->gndg_state != GNILND_DGRAM_POSTED) {
1067                 return;
1068         }
1069
1070         LASSERTF(dgram->gndg_conn != NULL,
1071                  "dgram 0x%p with NULL conn\n", dgram);
1072
1073         /* C.E - WC dgrams could be canceled immediately but
1074          * if there was some match pending, we need to call
1075          * test_by_id to clear it out. If that test returns
1076          * POST_PENDING, it is half done and needs to go along
1077          * with the rest of dgrams and go through a kgni_sm_task cycle
1078          * and deliver a GNI_POST_TERMINATED event before they
1079          * are actually canceled */
1080
1081         dgram->gndg_state = GNILND_DGRAM_CANCELED;
1082
1083         if (dgram->gndg_conn->gnc_state >= GNILND_CONN_ESTABLISHED) {
1084                 /* we don't need to cancel_by_id if the datagram was good */
1085                 return;
1086         }
1087
1088         /* let folks know there are outstanding cancels */
1089         atomic_inc(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1090         /* leave on nid list until cancel is done for debugging fun */
1091         grc = kgnilnd_ep_postdata_cancel_by_id(dgram->gndg_conn->gnc_ephandle, (__u64) dgram);
1092
1093         /* if we don't get success here, we have hosed up the dgram tracking
1094          * code and need to bail out */
1095         LASSERTF(grc == GNI_RC_SUCCESS,
1096                  "postdata_cancel returned %d for conn 0x%p to %s\n",
1097                  grc, dgram->gndg_conn,
1098                  dgram->gndg_conn->gnc_peer ?
1099                   libcfs_nid2str(dgram->gndg_conn->gnc_peer->gnp_nid)
1100                   : "<?>");
1101
1102         CDEBUG(D_NETTRACE,
1103                 "canceled dgram 0x%p conn 0x%p ephandle 0x%p\n",
1104                 dgram, dgram->gndg_conn,
1105                 dgram->gndg_conn->gnc_ephandle);
1106
1107         if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1108                 gni_post_state_t         post_state;
1109                 int                      rc = 0;
1110                 __u32                    remote_addr = 0, remote_id = 0;
1111
1112                 grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1113                                                      (__u64)dgram, &post_state,
1114                                                      &remote_addr, &remote_id);
1115
1116                 LASSERTF(grc == GNI_RC_NO_MATCH || grc == GNI_RC_SUCCESS,
1117                          "bad grc %d from test_by_id on dgram 0x%p\n",
1118                         grc, dgram);
1119
1120                 /* if WC was canceled immediately, we get NO_MATCH, if needs to go
1121                  * through full cycle, we get SUCCESS and need to parse post_state */
1122
1123                 CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1124                         "remote_addr %u remote_id %u\n", grc, dgram,
1125                         kgnilnd_dgram_type2str(dgram),
1126                         post_state, remote_addr, remote_id);
1127
1128                 if (grc == GNI_RC_NO_MATCH) {
1129                         /* she's gone, reduce count and move along */
1130                         dgram->gndg_state = GNILND_DGRAM_DONE;
1131                         atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1132                         RETURN_EXIT;
1133                 }
1134
1135                 rc = kgnilnd_process_dgram(dgram, post_state);
1136
1137                 if (rc <= 0) {
1138                         /* if for some weird reason we get a valid dgram back, just mark as done
1139                          * so we can drop it and move along.
1140                          * C.E - if it was completed, we'll just release the conn/mbox
1141                          * back into the pool and it'll get reused. That said, we should only
1142                          * be canceling a WC dgram on stack rest or shutdown, so that is moot */
1143                         dgram->gndg_state = GNILND_DGRAM_DONE;
1144                         atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
1145
1146                         /* caller context responsible for calling kgnilnd_release_dgram() */
1147                 } else {
1148                         /* still pending, let it simmer until golden brown and delicious */
1149                 }
1150         }
1151
1152         /* for non WC dgrams, they are still on the nid list but marked canceled waiting
1153          * for kgni to return their ID to us via probe - that is when we'll complete their
1154          * cancel processing */
1155 }
1156
1157 void
1158 kgnilnd_cleanup_dgram(kgn_dgram_t *dgram)
1159 {
1160         /* release the dgram ref on conn */
1161         if (dgram->gndg_conn) {
1162                 kgnilnd_conn_decref(dgram->gndg_conn);
1163                 dgram->gndg_conn = NULL;
1164         }
1165 }
1166
1167 void
1168 kgnilnd_free_dgram(kgn_device_t *dev, kgn_dgram_t *dgram)
1169 {
1170         LASSERTF(dgram->gndg_state == GNILND_DGRAM_USED ||
1171                  dgram->gndg_state == GNILND_DGRAM_DONE,
1172                  "dgram 0x%p with bad state %s\n",
1173                  dgram, kgnilnd_dgram_state2str(dgram));
1174
1175         /* bit of poisoning to help detect bad driver data */
1176         dgram->gndg_magic = 0x6f5a6b5f;
1177         atomic_dec(&dev->gnd_ndgrams);
1178
1179         kmem_cache_free(kgnilnd_data.kgn_dgram_cache, dgram);
1180         CDEBUG(D_MALLOC|D_NETTRACE, "slab-freed 'dgram': %lu at %p %s"
1181                " ndgrams %d\n",
1182                sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
1183                atomic_read(&dev->gnd_ndgrams));
1184 }
1185
1186 int
1187 kgnilnd_post_dgram(kgn_device_t *dev, lnet_nid_t dstnid, kgn_connreq_type_t type,
1188                    int data_rc)
1189 {
1190         int              rc = 0;
1191         kgn_dgram_t     *dgram = NULL;
1192         kgn_dgram_t     *tmpdgram;
1193         kgn_dgram_type_t dgtype;
1194         gni_return_t     grc;
1195         __u64            srcnid;
1196         ENTRY;
1197
1198         switch (type) {
1199         case GNILND_CONNREQ_REQ:
1200                 if (dstnid == LNET_NID_ANY)
1201                         dgtype = GNILND_DGRAM_WC_REQ;
1202                 else
1203                         dgtype = GNILND_DGRAM_REQ;
1204                 break;
1205         case GNILND_CONNREQ_NAK:
1206                 LASSERTF(dstnid != LNET_NID_ANY, "can't NAK to LNET_NID_ANY\n");
1207                 dgtype = GNILND_DGRAM_NAK;
1208                 break;
1209         default:
1210                 CERROR("unknown connreq type %d\n", type);
1211                 LBUG();
1212         }
1213
1214         rc = kgnilnd_alloc_dgram(&dgram, dev, dgtype);
1215         if (rc < 0) {
1216                 rc = -ENOMEM;
1217                 GOTO(post_failed, rc);
1218         }
1219
1220         rc = kgnilnd_create_conn(&dgram->gndg_conn, dev);
1221         if (rc) {
1222                 GOTO(post_failed, rc);
1223         }
1224
1225         if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
1226                 /* clear buffer for sanity on reuse of wildcard */
1227                 memset(&dgram->gndg_conn_in, 0, sizeof(kgn_connreq_t));
1228         }
1229
1230         if (dstnid == LNET_NID_ANY) {
1231                 /* set here to reset any dgram re-use */
1232                 dgram->gndg_conn->gnc_state = GNILND_CONN_LISTEN;
1233         } else {
1234                 __u32            host_id;
1235
1236                 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(dstnid), 1, &host_id);
1237                 if (rc <= 0) {
1238                         rc = -ESRCH;
1239                         GOTO(post_failed, rc);
1240                 }
1241
1242                 dgram->gndg_conn->gnc_state = GNILND_CONN_CONNECTING;
1243
1244                 /* don't need to serialize, there are no CQs for the dgram
1245                  * EP on the kgn_net_t */
1246                 grc = kgnilnd_ep_bind(dgram->gndg_conn->gnc_ephandle, host_id, dev->gnd_id);
1247
1248                 if (grc != GNI_RC_SUCCESS) {
1249                         rc = -ECONNABORTED;
1250                         GOTO(post_failed, rc);
1251                 }
1252
1253         }
1254
1255         /* If we are posting wildcards post using a net of 0, otherwise we'll use the
1256          * net of the destination node.
1257          */
1258
1259         if (dstnid == LNET_NID_ANY) {
1260                 srcnid = LNET_MKNID(LNET_MKNET(GNILND, 0), dev->gnd_nid);
1261         } else {
1262                 srcnid = LNET_MKNID(LNET_NIDNET(dstnid), dev->gnd_nid);
1263         }
1264
1265         rc = kgnilnd_pack_connreq(&dgram->gndg_conn_out, dgram->gndg_conn,
1266                                   srcnid, dstnid, type);
1267         if (rc) {
1268                 GOTO(post_failed, rc);
1269         }
1270
1271         if (type == GNILND_CONNREQ_NAK)
1272                 dgram->gndg_conn_out.gncr_nakdata.gnnd_errno = data_rc;
1273
1274         dgram->gndg_post_time = jiffies;
1275
1276         /* XXX Nic: here is where we'd add in logical network multiplexing */
1277
1278         CDEBUG(D_NETTRACE, "dgram 0x%p type %s %s->%s cdm %d\n",
1279                dgram, kgnilnd_dgram_type2str(dgram),
1280                libcfs_nid2str(srcnid),
1281                libcfs_nid2str(dstnid), dev->gnd_id);
1282
1283         /* this allocates memory, can't hold locks across */
1284         grc = kgnilnd_ep_postdata_w_id(dgram->gndg_conn->gnc_ephandle,
1285                                    &dgram->gndg_conn_out, sizeof(kgn_connreq_t),
1286                                    &dgram->gndg_conn_in, sizeof(kgn_connreq_t),
1287                                    (__u64)dgram);
1288
1289         if (grc != GNI_RC_SUCCESS) {
1290                 CNETERR("dropping failed dgram post id 0x%p type %s"
1291                         " reqtype %s to %s: rc %d\n",
1292                         dgram, kgnilnd_dgram_type2str(dgram),
1293                         kgnilnd_connreq_type2str(&dgram->gndg_conn_out),
1294                         libcfs_nid2str(dstnid), grc);
1295                 rc = (grc == GNI_RC_ERROR_NOMEM) ? -ENOMEM : -EBADR;
1296                 GOTO(post_failed, rc);
1297         }
1298
1299         /* we don't need to add earlier - if someone does del_peer during post,
1300          * that peer will get marked as unlinked and the callers wil take care of it.
1301          * The dgram code is largely kgn_peer_t ignorant, so at worst, we'll just drop
1302          * the completed dgram later when we cant find a peer to stuff it into */
1303
1304         spin_lock(&dev->gnd_dgram_lock);
1305
1306         /* make sure we are not double posting targeted dgrams
1307          * - we can multiple post WC dgrams to help with processing speed */
1308         if (dstnid != LNET_NID_ANY) {
1309                 tmpdgram = kgnilnd_find_dgram_locked(dev, dstnid);
1310
1311                 LASSERTF(tmpdgram == NULL,
1312                         "dgram 0x%p->%s already posted\n",
1313                          dgram, libcfs_nid2str(dstnid));
1314         }
1315
1316         /* unmunge dstnid to help processing code cope... */
1317         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
1318                 dgram->gndg_conn_out.gncr_dstnid = dstnid;
1319         }
1320
1321         list_add_tail(&dgram->gndg_list, kgnilnd_nid2dgramlist(dev, dstnid));
1322         dgram->gndg_state = GNILND_DGRAM_POSTED;
1323         spin_unlock(&dev->gnd_dgram_lock);
1324
1325 post_failed:
1326         if (rc < 0 && dgram != NULL) {
1327                 kgnilnd_cleanup_dgram(dgram);
1328                 kgnilnd_free_dgram(dev, dgram);
1329         }
1330
1331         RETURN(rc);
1332 }
1333
1334 /* The shutdown flag is set from the shutdown and stack reset threads. */
1335 void
1336 kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown)
1337 {
1338         /* The conns of canceled active dgrams need to be put in purgatory so
1339          * we don't reuse the mailbox */
1340         if (unlikely(dgram->gndg_state == GNILND_DGRAM_CANCELED)) {
1341                 kgn_peer_t *peer;
1342                 kgn_conn_t *conn = dgram->gndg_conn;
1343                 lnet_nid_t nid = dgram->gndg_conn_out.gncr_dstnid;
1344
1345                 dgram->gndg_state = GNILND_DGRAM_DONE;
1346
1347                 /* During shutdown we've already removed the peer so we don't
1348                  * need to add a peer. During stack reset we don't care about
1349                  * MDDs since they are all released. */
1350                 if (!shutdown) {
1351                         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1352                         peer = kgnilnd_find_peer_locked(nid);
1353
1354                         if (peer != NULL) {
1355                                 CDEBUG(D_NET, "adding peer's conn with nid %s "
1356                                         "to purgatory\n", libcfs_nid2str(nid));
1357                                 kgnilnd_conn_addref(conn);
1358                                 conn->gnc_peer = peer;
1359                                 kgnilnd_peer_addref(peer);
1360                                 kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
1361                                 conn->gnc_state = GNILND_CONN_CLOSED;
1362                                 list_add_tail(&conn->gnc_list,
1363                                               &peer->gnp_conns);
1364                                 kgnilnd_add_purgatory_locked(conn,
1365                                                              conn->gnc_peer);
1366                                 kgnilnd_schedule_conn(conn);
1367                         }
1368                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1369                 }
1370         }
1371
1372         spin_lock(&dev->gnd_dgram_lock);
1373         kgnilnd_cancel_dgram_locked(dgram);
1374         spin_unlock(&dev->gnd_dgram_lock);
1375
1376         kgnilnd_cleanup_dgram(dgram);
1377
1378         /* if the dgram is 'canceled' it needs to be wait until the event
1379          * comes up from kgni that tells us it is safe to release */
1380         if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1381                 dgram->gndg_state = GNILND_DGRAM_DONE;
1382
1383                 LASSERTF(list_empty(&dgram->gndg_list), "dgram 0x%p on list\n", dgram);
1384
1385                 /* if it is a wildcard and we are in an appropriate state, repost
1386                  * the wildcard */
1387
1388                 if ((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
1389                     (!kgnilnd_data.kgn_wc_kill && !kgnilnd_data.kgn_in_reset)) {
1390                         int     rerc;
1391
1392                         rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1393                         if (rerc != 0) {
1394                                 /* We failed to repost the WC dgram for some reason
1395                                  * mark it so the repost system attempts to repost */
1396                                 kgnilnd_admin_addref(dev->gnd_nwcdgrams);
1397                         }
1398                 }
1399
1400                 /* always free the old dgram */
1401                 kgnilnd_free_dgram(dev, dgram);
1402         }
1403 }
1404
1405
1406 int
1407 kgnilnd_probe_for_dgram(kgn_device_t *dev, kgn_dgram_t **dgramp)
1408 {
1409         kgn_dgram_t             *dgram = NULL;
1410         gni_post_state_t         post_state;
1411         gni_return_t             grc;
1412         int                      rc = 0;
1413         __u64                    readyid;
1414         __u32                    remote_addr = 0, remote_id = 0;
1415         ENTRY;
1416
1417         /* Probe with the lock held. That way if we get a dgram we dont have it canceled
1418          * between finding the ready dgram and grabbing the lock to remove it from the
1419          * list. Otherwise we could be left in an inconsistent state. We own the dgram
1420          * once its off the list so we don't need to worry about others changing it at
1421          * that point. */
1422         spin_lock(&dev->gnd_dgram_lock);
1423         grc = kgnilnd_postdata_probe_by_id(dev->gnd_handle, &readyid);
1424         if (grc != GNI_RC_SUCCESS) {
1425                 spin_unlock(&dev->gnd_dgram_lock);
1426                 /* return 0 to indicate nothing happened */
1427                 RETURN(0);
1428         }
1429
1430         CDEBUG(D_NET, "ready %#llx on device 0x%p\n",
1431                 readyid, dev);
1432
1433         dgram = (kgn_dgram_t *)readyid;
1434
1435         LASSERTF(dgram->gndg_magic == GNILND_DGRAM_MAGIC,
1436                  "dgram 0x%p from id %#llx with bad magic %x\n",
1437                  dgram, readyid, dgram->gndg_magic);
1438
1439         LASSERTF(dgram->gndg_state == GNILND_DGRAM_POSTED ||
1440                  dgram->gndg_state == GNILND_DGRAM_CANCELED,
1441                  "dgram 0x%p with bad state %s\n",
1442                  dgram, kgnilnd_dgram_state2str(dgram));
1443
1444         LASSERTF(!list_empty(&dgram->gndg_list),
1445                  "dgram 0x%p with bad list state %s type %s\n",
1446                  dgram, kgnilnd_dgram_state2str(dgram),
1447                  kgnilnd_dgram_type2str(dgram));
1448
1449         /* now we know that the datagram structure is ok, so pull off list */
1450         list_del_init(&dgram->gndg_list);
1451
1452         /* while we have the gnn_dgram_lock and BEFORE we call test_by_id
1453          * change the state from POSTED to PROCESSING to ensure that
1454          * nobody cancels it after we've pulled it from the wire */
1455         if (dgram->gndg_state == GNILND_DGRAM_POSTED) {
1456                 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1457         }
1458
1459         LASSERTF(dgram->gndg_conn != NULL,
1460                 "dgram 0x%p with NULL conn\n", dgram);
1461
1462         grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
1463                                              (__u64)dgram, &post_state,
1464                                              &remote_addr, &remote_id);
1465
1466         /* we now "own" this datagram */
1467         spin_unlock(&dev->gnd_dgram_lock);
1468
1469         LASSERTF(grc != GNI_RC_NO_MATCH, "kgni lied! probe_by_id told us that"
1470                  " id %llu was ready\n", readyid);
1471
1472         CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
1473                 "remote_addr %u remote_id %u\n", grc, dgram,
1474                 kgnilnd_dgram_type2str(dgram),
1475                 post_state, remote_addr, remote_id);
1476
1477         if (unlikely(grc != GNI_RC_SUCCESS)) {
1478                 CNETERR("getting data for dgram 0x%p->%s failed rc %d. Dropping it\n",
1479                         dgram, libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
1480                         grc);
1481                 rc = -EINVAL;
1482                 GOTO(probe_for_out, rc);
1483         }
1484
1485         rc = kgnilnd_process_dgram(dgram, post_state);
1486
1487         /* we should never get probe finding a dgram for us and then it
1488          * being a WC dgram that is still in the middle of processing */
1489         LASSERTF(rc <= 0, "bad rc %d from process_dgram 0x%p state %d\n",
1490                  rc, dgram, post_state);
1491
1492         if (rc == 0) {
1493                 /* dgram is good enough for the data to be used */
1494                 dgram->gndg_state = GNILND_DGRAM_PROCESSING;
1495                 /* fake rc to mark that we've done something */
1496                 rc = 1;
1497         } else {
1498                 /* let kgnilnd_release_dgram take care of canceled dgrams */
1499                 if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
1500                         dgram->gndg_state = GNILND_DGRAM_DONE;
1501                 }
1502         }
1503
1504         *dgramp = dgram;
1505         RETURN(rc);
1506
1507 probe_for_out:
1508
1509         kgnilnd_release_dgram(dev, dgram, 0);
1510         RETURN(rc);
1511 }
1512
1513 int
1514 kgnilnd_setup_wildcard_dgram(kgn_device_t *dev)
1515 {
1516         /* if kgn_wildcard is zero, return error */
1517         int     rc = -ENOENT, i;
1518         ENTRY;
1519
1520         for (i = 0; i < *kgnilnd_tunables.kgn_nwildcard; i++) {
1521                 rc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
1522                 if (rc < 0) {
1523                         CERROR("error %d: could not post wildcard datagram # %d\n",
1524                                 rc, i);
1525                         rc = -EINVAL;
1526                         GOTO(failed, rc);
1527                 }
1528         }
1529
1530 failed:
1531         RETURN(rc);
1532 }
1533
1534 int
1535 kgnilnd_cancel_net_dgrams(kgn_net_t *net)
1536 {
1537         kgn_dgram_t *dg, *dgN;
1538         LIST_HEAD(zombies);
1539         int i;
1540         ENTRY;
1541
1542         /* we want to cancel any outstanding dgrams - we don't want to rely
1543          * on del_peer_or_conn catching all of them. This helps protect us in cases
1544          * where we don't quite keep the peer->dgram mapping in sync due to some
1545          * race conditions */
1546
1547         LASSERTF(net->gnn_shutdown || kgnilnd_data.kgn_in_reset,
1548                  "called with LND invalid state: net shutdown %d "
1549                  "in reset %d\n", net->gnn_shutdown,
1550                  kgnilnd_data.kgn_in_reset);
1551
1552         spin_lock(&net->gnn_dev->gnd_dgram_lock);
1553
1554         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1555                 list_for_each_entry_safe(dg, dgN, &net->gnn_dev->gnd_dgrams[i], gndg_list) {
1556
1557                         /* skip nids not on our net or are wildcards */
1558
1559
1560                         if (dg->gndg_type == GNILND_DGRAM_WC_REQ ||
1561                                 net->gnn_netnum != LNET_NETNUM(LNET_NIDNET(dg->gndg_conn_out.gncr_dstnid)))
1562                                 continue;
1563
1564                         kgnilnd_cancel_dgram_locked(dg);
1565                 }
1566         }
1567
1568         spin_unlock(&net->gnn_dev->gnd_dgram_lock);
1569
1570         RETURN(0);
1571 }
1572
1573 int
1574 kgnilnd_cancel_wc_dgrams(kgn_device_t *dev)
1575 {
1576         kgn_dgram_t *dg, *dgN;
1577         LIST_HEAD(zombies);
1578         ENTRY;
1579
1580         /* Time to kill the outstanding WC's
1581          * WC's exist on net 0 only but match on any net...
1582          */
1583
1584         LASSERTF(kgnilnd_data.kgn_in_reset || kgnilnd_data.kgn_wc_kill,
1585                 "called with LND invalid state: WC shutdown %d "
1586                 "in reset %d\n", kgnilnd_data.kgn_wc_kill,
1587                 kgnilnd_data.kgn_in_reset);
1588
1589         spin_lock(&dev->gnd_dgram_lock);
1590
1591         do {
1592                 dg = kgnilnd_find_dgram_locked(dev, LNET_NID_ANY);
1593                 if (dg != NULL) {
1594                         LASSERTF(dg->gndg_type == GNILND_DGRAM_WC_REQ,
1595                                  "dgram 0x%p->%s with bad type %d (%s)\n",
1596                                 dg, libcfs_nid2str(dg->gndg_conn_out.gncr_dstnid),
1597                                 dg->gndg_type, kgnilnd_dgram_type2str(dg));
1598
1599                         kgnilnd_cancel_dgram_locked(dg);
1600
1601                         /* WC could be DONE already, check and if so add to list to be released */
1602                         if (dg->gndg_state == GNILND_DGRAM_DONE)
1603                                 list_move_tail(&dg->gndg_list, &zombies);
1604                 }
1605         } while (dg != NULL);
1606
1607         spin_unlock(&dev->gnd_dgram_lock);
1608
1609         list_for_each_entry_safe(dg, dgN, &zombies, gndg_list) {
1610                 list_del_init(&dg->gndg_list);
1611                 kgnilnd_release_dgram(dev, dg, 1);
1612         }
1613         RETURN(0);
1614
1615 }
1616
1617 int
1618 kgnilnd_cancel_dgrams(kgn_device_t *dev)
1619 {
1620         kgn_dgram_t *dg, *dgN;
1621         int i;
1622         ENTRY;
1623
1624         /* Cancel any outstanding non wildcard datagrams regardless
1625          * of which net they are on as we are in base shutdown and
1626          * dont care about connecting anymore.
1627          */
1628
1629         LASSERTF(kgnilnd_data.kgn_wc_kill == 1,"We didnt get called from base shutdown\n");
1630
1631         spin_lock(&dev->gnd_dgram_lock);
1632
1633         for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size -1); i++) {
1634                 list_for_each_entry_safe(dg, dgN, &dev->gnd_dgrams[i], gndg_list) {
1635                         if (dg->gndg_type != GNILND_DGRAM_WC_REQ)
1636                                 kgnilnd_cancel_dgram_locked(dg);
1637                 }
1638         }
1639
1640         spin_unlock(&dev->gnd_dgram_lock);
1641
1642         RETURN(0);
1643 }
1644
1645
1646 void
1647 kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev)
1648 {
1649         int             i = 4;
1650         int             rc;
1651         gni_return_t    grc;
1652         __u64           readyid;
1653         kgn_dgram_t    *dgram;
1654
1655         /* use do while to get at least one check run to allow
1656          * regression test for 762072 to hit bug if there */
1657
1658         /* This function races with the dgram mover during shutdown so it is possible for
1659          * a dgram to be seen in kgnilnd_postdata_probe_wait_by_id but be handled in the
1660          * dgram mover thread instead of inside of this function.
1661          */
1662
1663         /* This should only be called from within shutdown, baseshutdown, or stack reset.
1664          * there are no assertions here to verify since base_shutdown has nothing in it we can check
1665          * the net is gone by then.
1666          */
1667
1668         do {
1669                 i++;
1670                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1671                         "Waiting for %d canceled datagrams to clear on device %d\n",
1672                         atomic_read(&dev->gnd_canceled_dgrams), dev->gnd_id);
1673
1674                 /* check once a second */
1675                 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
1676                        250, &readyid);
1677
1678                 if (grc != GNI_RC_SUCCESS)
1679                         continue;
1680
1681                 CDEBUG(D_NET, "ready %#llx on device %d->0x%p\n",
1682                         readyid, dev->gnd_id, dev);
1683
1684                 rc = kgnilnd_probe_for_dgram(dev, &dgram);
1685                 if (rc != 0) {
1686                         /* if we got a valid dgram or one that is now done, clean up */
1687                         kgnilnd_release_dgram(dev, dgram, 1);
1688                 }
1689         } while (atomic_read(&dev->gnd_canceled_dgrams));
1690 }
1691
1692 int
1693 kgnilnd_start_connect(kgn_peer_t *peer)
1694 {
1695         int              rc = 0;
1696         /* sync point for kgnilnd_del_peer_locked - do an early check to
1697          * catch the most common hits where del_peer is done by the
1698          * time we get here */
1699         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING1)) {
1700                 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING1, 1)) {};
1701         }
1702
1703         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1704         if (!kgnilnd_peer_active(peer) || peer->gnp_connecting != GNILND_PEER_CONNECT) {
1705                 /* raced with peer getting unlinked */
1706                 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1707                 rc = ESTALE;
1708                 GOTO(out, rc);
1709         }
1710         peer->gnp_connecting = GNILND_PEER_POSTING;
1711         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1712
1713         set_mb(peer->gnp_last_dgram_time, jiffies);
1714         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING2)) {
1715                 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING2, 1)) {};
1716         }
1717
1718         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING3)) {
1719                 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING3, 1)) {};
1720                 rc = cfs_fail_val ? cfs_fail_val : -ENOMEM;
1721         } else {
1722                 rc = kgnilnd_post_dgram(peer->gnp_net->gnn_dev,
1723                                         peer->gnp_nid, GNILND_CONNREQ_REQ, 0);
1724         }
1725         if (rc < 0) {
1726                 set_mb(peer->gnp_last_dgram_errno, rc);
1727                 GOTO(failed, rc);
1728         }
1729
1730         /* while we're posting someone could have decided this peer/dgram needed to
1731          * die a quick death, so we check for state change and process accordingly */
1732
1733         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1734         if (!kgnilnd_peer_active(peer) || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1735                 if (peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1736                         peer->gnp_connecting = GNILND_PEER_KILL;
1737                 }
1738                 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1739                 /* positive RC to avoid dgram cleanup - we'll have to
1740                  * wait for the kgni GNI_POST_TERMINATED event to
1741                  * finish cleaning up */
1742                 rc = ESTALE;
1743                 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev, peer->gnp_nid);
1744                 GOTO(out, rc);
1745         }
1746         peer->gnp_connecting = GNILND_PEER_POSTED;
1747         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1748         /* reaper thread will take care of any timeouts */
1749         CDEBUG(D_NET, "waiting for connect to finish to %s rc %d\n",
1750                libcfs_nid2str(peer->gnp_nid), rc);
1751
1752         RETURN(rc);
1753
1754 failed:
1755         CDEBUG(D_NET, "connect to %s failed: rc %d \n",
1756                libcfs_nid2str(peer->gnp_nid), rc);
1757 out:
1758         RETURN(rc);
1759 }
1760
1761 int
1762 kgnilnd_finish_connect(kgn_dgram_t *dgram)
1763 {
1764         kgn_conn_t        *conn = dgram->gndg_conn;
1765         lnet_nid_t         her_nid = dgram->gndg_conn_in.gncr_srcnid;
1766         struct lnet_nid    peer_nid;
1767         kgn_peer_t        *new_peer, *peer = NULL;
1768         kgn_tx_t          *tx;
1769         kgn_tx_t          *txn;
1770         kgn_mbox_info_t   *mbox;
1771         int                rc;
1772         int                nstale;
1773
1774         /* try to find a peer that matches the nid we got in the connreq
1775          * kgnilnd_unpack_connreq makes sure that conn_in.gncr_srcnid is
1776          * HER and conn_out.gncr_srcnid is ME for both active and WC dgrams */
1777
1778         /* assume this is a new peer  - it makes locking cleaner when it isn't */
1779         /* no holding kgn_net_rw_sem - already are at the kgnilnd_dgram_mover level */
1780
1781         rc = kgnilnd_create_peer_safe(&new_peer, her_nid, NULL, GNILND_PEER_UP);
1782         if (rc != 0) {
1783                 CERROR("Can't create peer for %s\n", libcfs_nid2str(her_nid));
1784                 return rc;
1785         }
1786
1787         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1788
1789         /* this transfers ref from create_peer to the kgn_peer table */
1790         kgnilnd_add_peer_locked(her_nid, new_peer, &peer);
1791
1792         /* if we found an existing peer, is it really ready for a new conn ? */
1793         if (peer != new_peer) {
1794                 /* if this was an active connect attempt but we can't find a peer waiting for it
1795                  * we will dump in the trash */
1796
1797                 if (peer->gnp_connecting == GNILND_PEER_IDLE && dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1798                         CDEBUG(D_NET, "dropping completed connreq for %s peer 0x%p->%s\n",
1799                                libcfs_nid2str(her_nid), peer, libcfs_nid2str(peer->gnp_nid));
1800                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1801                         rc = ECANCELED;
1802                         GOTO(out, rc);
1803                 }
1804
1805                 /* check to see if we can catch a connecting peer before it is
1806                  * removed from the connd_peers list - if not, we need to
1807                  * let the connreqs race and be handled by kgnilnd_conn_isdup_locked() */
1808                 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
1809                         spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1810                         if (!list_empty(&peer->gnp_connd_list)) {
1811                                 list_del_init(&peer->gnp_connd_list);
1812                                 /* drop connd ref */
1813                                 kgnilnd_peer_decref(peer);
1814                         }
1815                         spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1816                         /* clear rc to make sure we don't have fake error */
1817                         rc = 0;
1818                 }
1819
1820                 /* no matter what, we are no longer waiting to connect this peer now */
1821                 peer->gnp_connecting = GNILND_PEER_IDLE;
1822
1823                 /* Refuse to duplicate an existing connection (both sides might try to
1824                  * connect at once).  NB we return success!  We _are_ connected so we
1825                  * _don't_ have any blocked txs to complete with failure. */
1826                 rc = kgnilnd_conn_isdup_locked(peer, conn);
1827                 if (rc != 0) {
1828                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1829                         CDEBUG(D_NET, "Not creating duplicate connection to %s: %d\n",
1830                               libcfs_nid2str(her_nid), rc);
1831                         rc = EALREADY;
1832                         GOTO(out, rc);
1833                 }
1834         }
1835
1836         if (peer->gnp_state == GNILND_PEER_DOWN) {
1837                 CNETERR("Received connection request from down nid %s\n",
1838                         libcfs_nid2str(her_nid));
1839         }
1840
1841         peer->gnp_state = GNILND_PEER_UP;
1842         nstale = kgnilnd_close_stale_conns_locked(peer, conn);
1843
1844         /* either way with peer (new or existing), we are ok with ref counts here as the
1845          * kgnilnd_add_peer_locked will use our ref on new_peer (from create_peer_safe) as the
1846          * ref for the peer table. */
1847
1848         /* at this point, the connection request is a winner */
1849
1850         /* mark 'DONE' to avoid cancel being called from release */
1851         dgram->gndg_state = GNILND_DGRAM_DONE;
1852
1853         /* initialise timestamps before reaper looks at them */
1854         conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
1855
1856         /* last_tx is initialized to jiffies - (keepalive*2) so that if the NOOP fails it will
1857          * immediatly send a NOOP in the reaper thread during the call to
1858          * kgnilnd_check_conn_timeouts_locked
1859          */
1860         conn->gnc_last_tx = jiffies - (cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout)) * 2);
1861         conn->gnc_state = GNILND_CONN_ESTABLISHED;
1862
1863         /* save the dgram type used to establish this connection */
1864         conn->gnc_dgram_type = dgram->gndg_type;
1865
1866         /* refs are not transferred from dgram to tables, so increment to
1867          * take ownership */
1868         kgnilnd_conn_addref(conn);
1869         kgnilnd_peer_addref(peer);
1870         conn->gnc_peer = peer;
1871         list_add_tail(&conn->gnc_list, &peer->gnp_conns);
1872
1873         kgnilnd_conn_addref(conn);               /* +1 ref for conn table */
1874         list_add_tail(&conn->gnc_hashlist,
1875                       kgnilnd_cqid2connlist(conn->gnc_cqid));
1876         kgnilnd_data.kgn_conn_version++;
1877
1878         /* Dont send NOOP if fail_loc is set
1879          */
1880         if (!CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP)) {
1881                 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP,
1882                                         lnet_nid_to_nid4(&peer->gnp_net->gnn_ni->ni_nid));
1883                 if (tx == NULL) {
1884                         CNETERR("can't get TX to initiate NOOP to %s\n",
1885                                 libcfs_nid2str(peer->gnp_nid));
1886                 } else {
1887                         kgnilnd_queue_tx(conn, tx);
1888                 }
1889         }
1890
1891         /* Schedule all packets blocking for a connection */
1892         list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1893                 /* lock held here is the peer_conn lock */
1894                 kgnilnd_tx_del_state_locked(tx, peer, NULL, GNILND_TX_ALLOCD);
1895                 kgnilnd_queue_tx(conn, tx);
1896         }
1897
1898         /* If this is an active connection lets mark its timestamp on the MBoX */
1899         if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
1900                 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1901                 /* conn->gnc_last_rx is jiffies it better exist as it was just set */
1902                 mbox->mbx_release_purg_active_dgram = conn->gnc_last_rx;
1903         }
1904
1905         /* Bug 765042: wake up scheduler for a race with finish_connect and
1906          * complete_conn_closed with a conn in purgatory
1907          * since we can't use CFS_RACE due to mutex_holds in kgnilnd_process_conns,
1908          * we just check for set and then clear */
1909         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
1910                 cfs_fail_loc = 0x0;
1911                 /* get scheduler thread moving again */
1912                 kgnilnd_schedule_device(conn->gnc_device);
1913         }
1914
1915         CDEBUG(D_NET, "New conn 0x%p->%s dev %d\n",
1916                conn, libcfs_nid2str(her_nid), conn->gnc_device->gnd_id);
1917
1918         /* make sure we reset peer reconnect interval now that we have a good conn */
1919         kgnilnd_peer_alive(peer);
1920         peer->gnp_reconnect_interval = 0;
1921
1922         /* clear the unlink attribute if we dont clear it kgnilnd_del_conn_or_peer will wait
1923          * on the atomic forever
1924          */
1925         if (peer->gnp_pending_unlink) {
1926                 peer->gnp_pending_unlink = 0;
1927                 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1928                 CDEBUG(D_NET, "Clearing peer unlink %p\n",peer);
1929         }
1930
1931         /* add ref to make it hang around until after we drop the lock */
1932         kgnilnd_conn_addref(conn);
1933
1934         /* Once the peer_conn lock is dropped, the conn could actually move into
1935          * CLOSING->CLOSED->DONE in the scheduler thread, so hold the
1936          * lock until we are really done */
1937         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1938
1939         /* Notify LNET that we now have a working connection to this peer.
1940          * This is a Cray extension to the "standard" LND behavior.
1941          */
1942         lnet_nid4_to_nid(peer->gnp_nid, &peer_nid);
1943         lnet_notify(peer->gnp_net->gnn_ni, &peer_nid, true, true,
1944                     ktime_get_seconds());
1945
1946         /* drop our 'hold' ref */
1947         kgnilnd_conn_decref(conn);
1948
1949 out:
1950         RETURN(rc);
1951 }
1952
1953 void
1954 kgnilnd_send_nak(kgn_device_t *dev, lnet_nid_t dst_nid, int error)
1955 {
1956         int              rc = 0;
1957         ENTRY;
1958
1959         LASSERTF(dst_nid != LNET_NID_ANY, "bad dst_nid %s\n", libcfs_nid2str(dst_nid));
1960
1961         CDEBUG(D_NET, "NAK to %s errno %d\n", libcfs_nid2str(dst_nid), error);
1962
1963         rc = kgnilnd_post_dgram(dev, dst_nid, GNILND_CONNREQ_NAK, error);
1964
1965         if (rc < 0) {
1966                 CDEBUG(D_NET, "NAK to %s failed: rc %d \n", libcfs_nid2str(dst_nid), rc);
1967         }
1968         EXIT;
1969 }
1970
1971 int
1972 kgnilnd_process_nak(kgn_dgram_t *dgram)
1973 {
1974         kgn_connreq_t     *connreq = &dgram->gndg_conn_in;
1975         lnet_nid_t         src_nid = connreq->gncr_srcnid;
1976         int                errno = connreq->gncr_nakdata.gnnd_errno;
1977         kgn_peer_t        *peer;
1978         int                rc = 0;
1979
1980         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1981
1982         peer = kgnilnd_find_peer_locked(src_nid);
1983         if (peer == NULL) {
1984                 /* we likely dropped him from bad data when we processed
1985                  * the original REQ */
1986                 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1987                 return -EBADSLT;
1988         }
1989
1990         /* need to check peerstamp/connstamp against the ones we find
1991          * to make sure we don't close new (and good?) conns that we
1992          * formed after this connreq failed */
1993         if (peer->gnp_connecting == GNILND_PEER_IDLE) {
1994                 kgn_conn_t        conn;
1995
1996                 if (list_empty(&peer->gnp_conns)) {
1997                         /* assume already procced datagram and it barfed up
1998                          * on this side too */
1999                         CDEBUG(D_NET, "dropping NAK from %s; "
2000                                "peer %s is already not connected\n",
2001                                 libcfs_nid2str(connreq->gncr_srcnid),
2002                                 libcfs_nid2str(connreq->gncr_dstnid));
2003                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2004                         return 0;
2005                 }
2006
2007                 /* stub up a connection with the connreq XXX_stamps to allow
2008                  * use to use close_stale_conns_locked */
2009                 conn.gnc_peerstamp = connreq->gncr_peerstamp;
2010                 conn.gnc_my_connstamp = connreq->gncr_connstamp;
2011                 conn.gnc_peer_connstamp = connreq->gncr_connstamp;
2012                 conn.gnc_device = peer->gnp_net->gnn_dev;
2013
2014                 rc = kgnilnd_close_stale_conns_locked(peer, &conn);
2015
2016                 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
2017                         "closed %d connections\n",
2018                         libcfs_nid2str(connreq->gncr_srcnid),
2019                         libcfs_nid2str(connreq->gncr_dstnid), errno, rc);
2020         } else {
2021                 spin_lock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2022
2023                 if (list_empty(&peer->gnp_connd_list)) {
2024                         /* if peer isn't on waiting list, try to find one to nuke */
2025                         rc = kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
2026                                                            peer->gnp_nid);
2027
2028                         if (rc) {
2029                                 LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
2030                                         "canceled pending connect request\n",
2031                                         libcfs_nid2str(connreq->gncr_srcnid),
2032                                         libcfs_nid2str(connreq->gncr_dstnid), errno);
2033                         }
2034
2035                         /* if we can't find a waiting dgram, we just drop the nak - the conn
2036                          * connect must have failed (didn't find conn above and clear connecting
2037                          * -- so nothing to do besides drop */
2038                 } else {
2039                         /* peer is on list, meaning it is a new connect attempt from the one
2040                          * we started that generated the NAK - so just drop NAK */
2041
2042                         /* use negative to prevent error message */
2043                         rc = -EAGAIN;
2044                 }
2045                 spin_unlock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
2046         }
2047
2048         /* success! we found a peer and at least marked pending_nak */
2049         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2050
2051         return rc;
2052 }
2053
2054 int
2055 kgnilnd_process_connreq(kgn_dgram_t *dgram, int *needs_nak)
2056 {
2057         int                      rc;
2058
2059         rc = kgnilnd_unpack_connreq(dgram);
2060         if (rc < 0) {
2061                 if (rc != -EBADF) {
2062                         /* only NAK if we have good srcnid to use */
2063                         *needs_nak = 1;
2064                 }
2065                 goto connreq_out;
2066         }
2067
2068         switch (dgram->gndg_conn_in.gncr_type) {
2069         case GNILND_CONNREQ_REQ:
2070                 /* wire up peer & conn, send queued TX */
2071                 rc = kgnilnd_finish_connect(dgram);
2072
2073                 /* don't nak when the nid is hosed */
2074                 if ((rc < 0)) {
2075                         *needs_nak = 1;
2076                 }
2077
2078                 break;
2079         case GNILND_CONNREQ_NAK:
2080                 rc = kgnilnd_process_nak(dgram);
2081                 /* return early to prevent reconnect bump */
2082                 return rc;
2083         default:
2084                 CERROR("unexpected connreq type %s (%d) from %s\n",
2085                         kgnilnd_connreq_type2str(&dgram->gndg_conn_in),
2086                         dgram->gndg_conn_in.gncr_type,
2087                         libcfs_nid2str(dgram->gndg_conn_in.gncr_srcnid));
2088                 rc = -EINVAL;
2089                 *needs_nak = 1;
2090                 break;
2091         }
2092
2093 connreq_out:
2094         RETURN(rc);
2095 }
2096
2097 int
2098 kgnilnd_probe_and_process_dgram(kgn_device_t *dev)
2099 {
2100         int                      rc;
2101         int                      needs_nak = 0;
2102         lnet_nid_t               nak_dstnid = LNET_NID_ANY;
2103         lnet_nid_t               orig_dstnid;
2104         kgn_dgram_t             *dgram = NULL;
2105         kgn_peer_t              *peer;
2106         ENTRY;
2107
2108         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PAUSE_DGRAM_COMP)) {
2109                 rc = 0;
2110         } else {
2111                 rc = kgnilnd_probe_for_dgram(dev, &dgram);
2112         }
2113
2114         if (rc == 0) {
2115                 RETURN(0);
2116         } else if (rc < 0) {
2117                 GOTO(inform_peer, rc);
2118         } else {
2119                 /* rc > 1 means it did something, reset for this func  */
2120                 rc = 0;
2121         }
2122
2123         switch (dgram->gndg_type) {
2124         case GNILND_DGRAM_WC_REQ:
2125         case GNILND_DGRAM_REQ:
2126                 rc = kgnilnd_process_connreq(dgram, &needs_nak);
2127                 break;
2128         case GNILND_DGRAM_NAK:
2129                 CDEBUG(D_NETTRACE, "NAK to %s done\n",
2130                         libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
2131                 break;
2132         default:
2133                 CERROR("unknown datagram type %s (%d)\n",
2134                        kgnilnd_dgram_type2str(dgram), dgram->gndg_type);
2135                 break;
2136         }
2137
2138         /* stash data to use after releasing current datagram */
2139         /* don't stash net - we are operating on a net already,
2140          * so the lock on rw_net_lock is sufficient */
2141
2142         nak_dstnid = dgram->gndg_conn_in.gncr_srcnid;
2143
2144 inform_peer:
2145         LASSERTF(dgram != NULL, "dgram 0x%p rc %d needs_nak %d\n", dgram, rc, needs_nak);
2146
2147         orig_dstnid = dgram->gndg_conn_out.gncr_dstnid;
2148
2149         kgnilnd_release_dgram(dev, dgram, 0);
2150
2151         CDEBUG(D_NET, "cleaning up dgram to %s, rc %d\n",
2152                libcfs_nid2str(orig_dstnid), rc);
2153
2154         /* if this was a WC_REQ that matched an existing peer, it'll get marked done
2155          * in kgnilnd_finish_connect - if errors are from before we get to there,
2156          * we just drop as it is a WC_REQ - the peer CAN'T be waiting for it */
2157         if ((orig_dstnid != LNET_NID_ANY) && (rc < 0)) {
2158                 /* if we have a negative rc, we want to find a peer to inform about
2159                  * the bad connection attempt. Sorry buddy, better luck next time! */
2160
2161                 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2162                 peer = kgnilnd_find_peer_locked(orig_dstnid);
2163
2164                 if (peer != NULL) {
2165                         /* add ref to make sure he stays around past the possible unlink
2166                          * so we can tell LNet about him */
2167                         kgnilnd_peer_addref(peer);
2168
2169                         /* if he still cares about the outstanding connect */
2170                         if (peer->gnp_connecting >= GNILND_PEER_CONNECT) {
2171                                 /* check if he is on the connd list and remove.. */
2172                                 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2173                                 if (!list_empty(&peer->gnp_connd_list)) {
2174                                         list_del_init(&peer->gnp_connd_list);
2175                                         /* drop connd ref */
2176                                         kgnilnd_peer_decref(peer);
2177                                 }
2178                                 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
2179
2180                                 /* clear gnp_connecting so we don't have a non-connecting peer
2181                                  * on gnd_connd_list */
2182                                 peer->gnp_connecting = GNILND_PEER_IDLE;
2183
2184                                 set_mb(peer->gnp_last_dgram_errno, rc);
2185
2186                                 kgnilnd_peer_increase_reconnect_locked(peer);
2187                         }
2188                 }
2189                 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2190
2191                 /* now that we are outside the lock, tell Mommy */
2192                 if (peer != NULL) {
2193                         kgnilnd_peer_notify(peer, rc, 0);
2194                         kgnilnd_peer_decref(peer);
2195                 }
2196         }
2197
2198         if (needs_nak) {
2199                 kgnilnd_send_nak(dev, nak_dstnid, rc);
2200         }
2201
2202         RETURN(1);
2203 }
2204
2205 void
2206 kgnilnd_reaper_dgram_check(kgn_device_t *dev)
2207 {
2208         kgn_dgram_t    *dgram, *tmp;
2209         int             i;
2210
2211         spin_lock(&dev->gnd_dgram_lock);
2212
2213         for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
2214                 list_for_each_entry_safe(dgram, tmp, &dev->gnd_dgrams[i], gndg_list) {
2215                         unsigned long            now = jiffies;
2216                         unsigned long            timeout;
2217
2218                         /* don't timeout stuff if the network is mucked or shutting down */
2219                         if (kgnilnd_check_hw_quiesce()) {
2220                                 break;
2221                         }
2222
2223                         if ((dgram->gndg_state != GNILND_DGRAM_POSTED) ||
2224                             (dgram->gndg_type == GNILND_DGRAM_WC_REQ)) {
2225                                 continue;
2226                         }
2227                         CDEBUG(D_NETTRACE, "checking dgram 0x%p type %s "
2228                                 "state %s conn 0x%p to %s age %lus\n",
2229                                 dgram, kgnilnd_dgram_type2str(dgram),
2230                                 kgnilnd_dgram_state2str(dgram), dgram->gndg_conn,
2231                                 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2232                                 cfs_duration_sec(now - dgram->gndg_post_time));
2233
2234                         timeout = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout);
2235
2236                         if (time_before(now, (dgram->gndg_post_time + timeout)))
2237                                 continue;
2238
2239                         CNETERR("%s datagram to %s timed out @ %lus dgram "
2240                                 "0x%p state %s conn 0x%p\n",
2241                                 kgnilnd_dgram_type2str(dgram),
2242                                 libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
2243                                 cfs_duration_sec(now - dgram->gndg_post_time),
2244                                 dgram, kgnilnd_dgram_state2str(dgram),
2245                                 dgram->gndg_conn);
2246
2247                         kgnilnd_cancel_dgram_locked(dgram);
2248                 }
2249         }
2250         spin_unlock(&dev->gnd_dgram_lock);
2251 }
2252
2253
2254 /* use a thread for the possibly long-blocking wait_by_id to prevent
2255  * stalling the global workqueues */
2256 int
2257 kgnilnd_dgram_waitq(void *arg)
2258 {
2259         kgn_device_t     *dev = (kgn_device_t *) arg;
2260         char              name[16];
2261         gni_return_t      grc;
2262         __u64             readyid;
2263         DEFINE_WAIT(mover_done);
2264
2265         snprintf(name, sizeof(name), "kgnilnd_dgn_%02d", dev->gnd_id);
2266
2267         /* all gnilnd threads need to run fairly urgently */
2268         set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2269
2270         /* we dont shut down until the device shuts down ... */
2271         while (!kgnilnd_data.kgn_shutdown) {
2272                 /* to quiesce or to not quiesce, that is the question */
2273                 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2274                         KGNILND_SPIN_QUIESCE;
2275                 }
2276
2277                 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_DGRAM_COMP, 1)) {}
2278
2279                 /* check once a second */
2280                 grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
2281                                                        1000, &readyid);
2282
2283                 if (grc == GNI_RC_SUCCESS) {
2284                         CDEBUG(D_INFO, "waking up dgram mover thread\n");
2285                         kgnilnd_schedule_dgram(dev);
2286
2287                         /* wait for dgram thread to ping us before spinning again */
2288                         prepare_to_wait(&dev->gnd_dgping_waitq, &mover_done,
2289                                         TASK_INTERRUPTIBLE);
2290
2291                         /* don't sleep if we need to quiesce */
2292                         if (likely(!kgnilnd_data.kgn_quiesce_trigger)) {
2293                                 schedule();
2294                         }
2295                         finish_wait(&dev->gnd_dgping_waitq, &mover_done);
2296                 }
2297         }
2298
2299         kgnilnd_thread_fini();
2300         return 0;
2301 }
2302
2303 int
2304 kgnilnd_start_outbound_dgrams(kgn_device_t *dev, unsigned long deadline)
2305 {
2306         int                      did_something = 0, rc;
2307         kgn_peer_t              *peer = NULL;
2308
2309         spin_lock(&dev->gnd_connd_lock);
2310
2311         /* Active connect - we added this in kgnilnd_launch_tx */
2312         while (!list_empty(&dev->gnd_connd_peers) && time_before(jiffies, deadline)) {
2313                 peer = list_first_entry(&dev->gnd_connd_peers,
2314                                         kgn_peer_t, gnp_connd_list);
2315
2316                 /* ref for connd removed in if/else below */
2317                list_del_init(&peer->gnp_connd_list);
2318
2319                 /* gnp_connecting and membership on gnd_connd_peers should be
2320                  * done coherently to avoid double adding, etc */
2321                 /* don't need kgnilnd_data.kgn_peer_conn_lock here as that is only needed
2322                  * to get the peer to gnp_connecting in the first place. We just need to
2323                  * rely on gnd_connd_lock to serialize someone pulling him from the list
2324                  * BEFORE clearing gnp_connecting */
2325                 LASSERTF(peer->gnp_connecting != GNILND_PEER_IDLE, "peer 0x%p->%s not connecting\n",
2326                          peer, libcfs_nid2str(peer->gnp_nid));
2327
2328                 spin_unlock(&dev->gnd_connd_lock);
2329
2330                 CDEBUG(D_NET, "processing connect to %s\n",
2331                        libcfs_nid2str(peer->gnp_nid));
2332
2333                 did_something += 1;
2334                 rc = kgnilnd_start_connect(peer);
2335
2336                 if (likely(rc >= 0)) {
2337                         /* 0 on success, positive on 'just drop peer' errors */
2338                         kgnilnd_peer_decref(peer);
2339                 } else if (rc == -ENOMEM) {
2340                         /* if we are out of wildcards, add back to
2341                          * connd_list - then break out and we'll try later
2342                          * if other errors, we'll bail & cancel pending tx */
2343                         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2344                         if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2345                                 peer->gnp_connecting = GNILND_PEER_CONNECT;
2346                                 spin_lock(&dev->gnd_connd_lock);
2347                                 list_add_tail(&peer->gnp_connd_list,
2348                                               &dev->gnd_connd_peers);
2349                         } else {
2350                                 /* connecting changed while we were posting */
2351
2352                                 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2353                                         " state 0x%p->%s, connecting %d\n",
2354                                         peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2355                                 peer->gnp_connecting = GNILND_PEER_KILL;
2356                                 spin_lock(&dev->gnd_connd_lock);
2357                                 /* remove the peer ref frrom the cond list */
2358                                 kgnilnd_peer_decref(peer);
2359                                 /* let the system handle itself */
2360                         }
2361                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2362                         /* the datagrams are a global pool,
2363                          * so break out of trying and hope some free
2364                          * up soon */
2365                         did_something -= 1;
2366                         break;
2367                 } else {
2368                         /* something bad happened, you lose */
2369                         CNETERR("could not start connecting to %s "
2370                                 "rc %d: Will retry until TX timeout\n",
2371                                libcfs_nid2str(peer->gnp_nid), rc);
2372                         /* It didnt post so just set connecting back to zero now.
2373                          * The reaper will reattempt the connection if it needs too.
2374                          * If the peer needs death set it so the reaper will cleanup.
2375                          */
2376                         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
2377                         if (peer->gnp_connecting == GNILND_PEER_POSTING) {
2378                                 peer->gnp_connecting = GNILND_PEER_IDLE;
2379                                 kgnilnd_peer_increase_reconnect_locked(peer);
2380                         } else {
2381                                 LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
2382                                         " state 0x%p->%s, connecting %d\n",
2383                                         peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
2384                                 peer->gnp_connecting = GNILND_PEER_KILL;
2385                         }
2386                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
2387
2388                         /* hold onto ref until we are really done - if it was
2389                          * unlinked this could result in a destroy */
2390                         kgnilnd_peer_decref(peer);
2391                 }
2392                 spin_lock(&dev->gnd_connd_lock);
2393         }
2394
2395         spin_unlock(&dev->gnd_connd_lock);
2396         RETURN(did_something);
2397 }
2398
2399 int
2400 kgnilnd_repost_wc_dgrams(kgn_device_t *dev)
2401 {
2402         int did_something = 0, to_repost, i;
2403         to_repost = atomic_read(&dev->gnd_nwcdgrams);
2404         ENTRY;
2405
2406         for (i = 0; i < to_repost; ++i) {
2407                 int     rerc;
2408                 rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
2409                 if (rerc == 0) {
2410                         kgnilnd_admin_decref(dev->gnd_nwcdgrams);
2411                         did_something += 1;
2412                 } else {
2413                         CDEBUG(D_NETERROR, "error %d: dev %d could not post wildcard datagram\n",
2414                                 rerc, dev->gnd_id);
2415                         break;
2416                 }
2417         }
2418
2419         RETURN(did_something);
2420 }
2421
2422 struct kgnilnd_dgram_timer {
2423         struct timer_list timer;
2424         kgn_device_t *dev;
2425 };
2426
2427 static void
2428 kgnilnd_dgram_poke_with_stick(cfs_timer_cb_arg_t arg)
2429 {
2430         struct kgnilnd_dgram_timer *t = cfs_from_timer(t, arg, timer);
2431
2432         wake_up(&t->dev->gnd_dgram_waitq);
2433 }
2434
2435 /* use single thread for dgrams - should be sufficient for performance */
2436 int
2437 kgnilnd_dgram_mover(void *arg)
2438 {
2439         kgn_device_t            *dev = (kgn_device_t *)arg;
2440         char                     name[16];
2441         int                      rc, did_something;
2442         unsigned long            next_purge_check = jiffies - 1;
2443         unsigned long            timeout;
2444         struct kgnilnd_dgram_timer timer;
2445         unsigned long deadline = 0;
2446         DEFINE_WAIT(wait);
2447
2448         snprintf(name, sizeof(name), "kgnilnd_dg_%02d", dev->gnd_id);
2449
2450         /* all gnilnd threads need to run fairly urgently */
2451         set_user_nice(current, *kgnilnd_tunables.kgn_nice);
2452
2453         /* we are ok not locking for these variables as the dgram waitq threads
2454          * will block both due to tying up net (kgn_shutdown) and the completion
2455          * event for the dgram_waitq (kgn_quiesce_trigger) */
2456         deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2457         while (!kgnilnd_data.kgn_shutdown) {
2458                 /* Safe: kgn_shutdown only set when quiescent */
2459
2460                 /* race with stack reset - we want to hold off seeing any new incoming dgrams
2461                  * so we can force a dirty WC dgram for Bug 762072 - put right before
2462                  * quiesce check so that it'll go right into that and not do any
2463                  * dgram mucking */
2464                 CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);
2465
2466                 /* to quiesce or to not quiesce, that is the question */
2467                 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
2468                         KGNILND_SPIN_QUIESCE;
2469                 }
2470                 did_something = 0;
2471
2472                 CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);
2473
2474                 /* process any newly completed dgrams */
2475                 down_read(&kgnilnd_data.kgn_net_rw_sem);
2476
2477                 rc = kgnilnd_probe_and_process_dgram(dev);
2478                 if (rc > 0) {
2479                         did_something += rc;
2480                 }
2481
2482                 up_read(&kgnilnd_data.kgn_net_rw_sem);
2483
2484                 CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_DGRAM_DEADLINE,
2485                         (*kgnilnd_tunables.kgn_dgram_timeout + 1));
2486                 /* start new outbound dgrams */
2487                 did_something += kgnilnd_start_outbound_dgrams(dev, deadline);
2488
2489                 /* find dead dgrams */
2490                 if (time_after_eq(jiffies, next_purge_check)) {
2491                         /* these don't need to be checked that often */
2492                         kgnilnd_reaper_dgram_check(dev);
2493
2494                         next_purge_check = (long) jiffies +
2495                                       cfs_time_seconds(kgnilnd_data.kgn_new_min_timeout / 4);
2496                 }
2497
2498                 did_something += kgnilnd_repost_wc_dgrams(dev);
2499
2500                 /* careful with the jiffy wrap... */
2501                 timeout = (long)(next_purge_check - jiffies);
2502
2503                 CDEBUG(D_INFO, "did %d timeout %lu next %lu jiffies %lu\n",
2504                        did_something, timeout, next_purge_check, jiffies);
2505
2506                 if ((did_something || timeout <= 0) && time_before(jiffies, deadline)) {
2507                         did_something = 0;
2508                         continue;
2509                 }
2510
2511                 prepare_to_wait(&dev->gnd_dgram_waitq, &wait, TASK_INTERRUPTIBLE);
2512
2513                 cfs_timer_setup(&timer.timer,
2514                                 kgnilnd_dgram_poke_with_stick,
2515                                 dev, 0);
2516                 timer.dev = dev;
2517                 mod_timer(&timer.timer, (long) jiffies + timeout);
2518
2519                 /* last second chance for others to poke us */
2520                 did_something += xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_IDLE);
2521
2522                 /* check flag variables before committing even if we
2523                  * did something; if we are after the deadline call
2524                  * schedule */
2525                 if ((!did_something || time_after(jiffies, deadline)) &&
2526                     !kgnilnd_data.kgn_shutdown &&
2527                     !kgnilnd_data.kgn_quiesce_trigger) {
2528                         CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
2529                                timeout, cfs_duration_sec(timeout));
2530                         wake_up(&dev->gnd_dgping_waitq);
2531                         schedule();
2532                         CDEBUG(D_INFO, "awake after schedule\n");
2533                         deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
2534                 }
2535
2536                 timer_delete_sync(&timer.timer);
2537                 finish_wait(&dev->gnd_dgram_waitq, &wait);
2538         }
2539
2540         kgnilnd_thread_fini();
2541         return 0;
2542 }