lnet/klnds/kfilnd/kfilnd_tn.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright 2022 Hewlett Packard Enterprise Development LP
  24  */
  25 /*
  26  * This file is part of Lustre, http://www.lustre.org/
  27  */
  28 /*
  29  * kfilnd transaction and state machine processing.
  30  */
  31
  32 #include "kfilnd_tn.h"
  33 #include "kfilnd_ep.h"
  34 #include "kfilnd_dev.h"
  35 #include "kfilnd_dom.h"
  36 #include "kfilnd_peer.h"
  37 #include <asm/checksum.h>
  38
  39 static struct kmem_cache *tn_cache;
  40 static struct kmem_cache *imm_buf_cache;
  41
  42 static __sum16 kfilnd_tn_cksum(void *ptr, int nob)
  43 {
  44         if (cksum)
  45                 return csum_fold(csum_partial(ptr, nob, 0));
  46         return NO_CHECKSUM;
  47 }
  48
  49 static int kfilnd_tn_msgtype2size(enum kfilnd_msg_type type)
  50 {
  51         const int hdr_size = offsetof(struct kfilnd_msg, proto);
  52
  53         switch (type) {
  54         case KFILND_MSG_IMMEDIATE:
  55                 return offsetof(struct kfilnd_msg, proto.immed.payload[0]);
  56
  57         case KFILND_MSG_BULK_PUT_REQ:
  58         case KFILND_MSG_BULK_GET_REQ:
  59                 return hdr_size + sizeof(struct kfilnd_bulk_req_msg);
  60
  61         default:
  62                 return -1;
  63         }
  64 }
  65
  66 static void kfilnd_tn_pack_hello_req(struct kfilnd_transaction *tn)
  67 {
  68         struct kfilnd_msg *msg = tn->tn_tx_msg.msg;
  69
  70         /* Pack the protocol header and payload. */
  71         msg->proto.hello.version = KFILND_MSG_VERSION;
  72         msg->proto.hello.rx_base = kfilnd_peer_target_rx_base(tn->tn_kp);
  73         msg->proto.hello.session_key = tn->tn_kp->kp_local_session_key;
  74
  75         /* TODO: Support multiple RX contexts per peer. */
  76         msg->proto.hello.rx_count = 1;
  77
  78         /* Pack the transport header. */
  79         msg->magic = KFILND_MSG_MAGIC;
  80
  81         /* Mesage version zero is only valid for hello requests. */
  82         msg->version = 0;
  83         msg->type = KFILND_MSG_HELLO_REQ;
  84         msg->nob = sizeof(struct kfilnd_hello_msg) +
  85                 offsetof(struct kfilnd_msg, proto);
  86         msg->cksum = NO_CHECKSUM;
  87         msg->srcnid = lnet_nid_to_nid4(&tn->tn_ep->end_dev->kfd_ni->ni_nid);
  88         msg->dstnid = tn->tn_kp->kp_nid;
  89
  90         /* Checksum entire message. */
  91         msg->cksum = kfilnd_tn_cksum(msg, msg->nob);
  92
  93         tn->tn_tx_msg.length = msg->nob;
  94 }
  95
  96 static void kfilnd_tn_pack_hello_rsp(struct kfilnd_transaction *tn)
  97 {
  98         struct kfilnd_msg *msg = tn->tn_tx_msg.msg;
  99
 100         /* Pack the protocol header and payload. */
 101         msg->proto.hello.version = tn->tn_kp->kp_version;
 102         msg->proto.hello.rx_base = kfilnd_peer_target_rx_base(tn->tn_kp);
 103         msg->proto.hello.session_key = tn->tn_kp->kp_local_session_key;
 104
 105         /* TODO: Support multiple RX contexts per peer. */
 106         msg->proto.hello.rx_count = 1;
 107
 108         /* Pack the transport header. */
 109         msg->magic = KFILND_MSG_MAGIC;
 110
 111         /* Mesage version zero is only valid for hello requests. */
 112         msg->version = 0;
 113         msg->type = KFILND_MSG_HELLO_RSP;
 114         msg->nob = sizeof(struct kfilnd_hello_msg) +
 115                 offsetof(struct kfilnd_msg, proto);
 116         msg->cksum = NO_CHECKSUM;
 117         msg->srcnid = lnet_nid_to_nid4(&tn->tn_ep->end_dev->kfd_ni->ni_nid);
 118         msg->dstnid = tn->tn_kp->kp_nid;
 119
 120         /* Checksum entire message. */
 121         msg->cksum = kfilnd_tn_cksum(msg, msg->nob);
 122
 123         tn->tn_tx_msg.length = msg->nob;
 124 }
 125
 126 static void kfilnd_tn_pack_bulk_req(struct kfilnd_transaction *tn)
 127 {
 128         struct kfilnd_msg *msg = tn->tn_tx_msg.msg;
 129
 130         /* Pack the protocol header and payload. */
 131         lnet_hdr_to_nid4(&tn->tn_lntmsg->msg_hdr, &msg->proto.bulk_req.hdr);
 132         msg->proto.bulk_req.key = tn->tn_mr_key;
 133         msg->proto.bulk_req.response_rx = tn->tn_response_rx;
 134
 135         /* Pack the transport header. */
 136         msg->magic = KFILND_MSG_MAGIC;
 137         msg->version = KFILND_MSG_VERSION;
 138         msg->type = tn->msg_type;
 139         msg->nob = sizeof(struct kfilnd_bulk_req_msg) +
 140                 offsetof(struct kfilnd_msg, proto);
 141         msg->cksum = NO_CHECKSUM;
 142         msg->srcnid = lnet_nid_to_nid4(&tn->tn_ep->end_dev->kfd_ni->ni_nid);
 143         msg->dstnid = tn->tn_kp->kp_nid;
 144
 145         /* Checksum entire message. */
 146         msg->cksum = kfilnd_tn_cksum(msg, msg->nob);
 147
 148         tn->tn_tx_msg.length = msg->nob;
 149 }
 150
 151 static void kfilnd_tn_pack_immed_msg(struct kfilnd_transaction *tn)
 152 {
 153         struct kfilnd_msg *msg = tn->tn_tx_msg.msg;
 154
 155         /* Pack the protocol header and payload. */
 156         lnet_hdr_to_nid4(&tn->tn_lntmsg->msg_hdr, &msg->proto.immed.hdr);
 157
 158         lnet_copy_kiov2flat(KFILND_IMMEDIATE_MSG_SIZE,
 159                             msg,
 160                             offsetof(struct kfilnd_msg,
 161                                      proto.immed.payload),
 162                             tn->tn_num_iovec, tn->tn_kiov, 0,
 163                             tn->tn_nob);
 164
 165         /* Pack the transport header. */
 166         msg->magic = KFILND_MSG_MAGIC;
 167         msg->version = KFILND_MSG_VERSION;
 168         msg->type = tn->msg_type;
 169         msg->nob = offsetof(struct kfilnd_msg, proto.immed.payload[tn->tn_nob]);
 170         msg->cksum = NO_CHECKSUM;
 171         msg->srcnid = lnet_nid_to_nid4(&tn->tn_ep->end_dev->kfd_ni->ni_nid);
 172         msg->dstnid = tn->tn_kp->kp_nid;
 173
 174         /* Checksum entire message. */
 175         msg->cksum = kfilnd_tn_cksum(msg, msg->nob);
 176
 177         tn->tn_tx_msg.length = msg->nob;
 178 }
 179
 180 static int kfilnd_tn_unpack_msg(struct kfilnd_ep *ep, struct kfilnd_msg *msg,
 181                                 unsigned int nob)
 182 {
 183         const unsigned int hdr_size = offsetof(struct kfilnd_msg, proto);
 184
 185         if (nob < hdr_size) {
 186                 KFILND_EP_ERROR(ep, "Short message: %u", nob);
 187                 return -EPROTO;
 188         }
 189
 190         /* TODO: Support byte swapping on mixed endian systems. */
 191         if (msg->magic != KFILND_MSG_MAGIC) {
 192                 KFILND_EP_ERROR(ep, "Bad magic: %#x", msg->magic);
 193                 return -EPROTO;
 194         }
 195
 196         /* TODO: Allow for older versions. */
 197         if (msg->version > KFILND_MSG_VERSION) {
 198                 KFILND_EP_ERROR(ep, "Bad version: %#x", msg->version);
 199                 return -EPROTO;
 200         }
 201
 202         if (msg->nob > nob) {
 203                 KFILND_EP_ERROR(ep, "Short message: got=%u, expected=%u", nob,
 204                                 msg->nob);
 205                 return -EPROTO;
 206         }
 207
 208         /* If kfilnd_tn_cksum() returns a non-zero value, checksum is bad. */
 209         if (msg->cksum != NO_CHECKSUM && kfilnd_tn_cksum(msg, msg->nob)) {
 210                 KFILND_EP_ERROR(ep, "Bad checksum");
 211                 return -EPROTO;
 212         }
 213
 214         if (msg->dstnid != lnet_nid_to_nid4(&ep->end_dev->kfd_ni->ni_nid)) {
 215                 KFILND_EP_ERROR(ep, "Bad destination nid: %s",
 216                                 libcfs_nid2str(msg->dstnid));
 217                 return -EPROTO;
 218         }
 219
 220         if (msg->srcnid == LNET_NID_ANY) {
 221                 KFILND_EP_ERROR(ep, "Bad source nid: %s",
 222                                 libcfs_nid2str(msg->srcnid));
 223                 return -EPROTO;
 224         }
 225
 226         if (msg->nob < kfilnd_tn_msgtype2size(msg->type)) {
 227                 KFILND_EP_ERROR(ep, "Short %s: %d(%d)\n",
 228                                 msg_type_to_str(msg->type),
 229                                 msg->nob, kfilnd_tn_msgtype2size(msg->type));
 230                 return -EPROTO;
 231         }
 232
 233         switch ((enum kfilnd_msg_type)msg->type) {
 234         case KFILND_MSG_IMMEDIATE:
 235         case KFILND_MSG_BULK_PUT_REQ:
 236         case KFILND_MSG_BULK_GET_REQ:
 237                 if (msg->version == 0) {
 238                         KFILND_EP_ERROR(ep,
 239                                         "Bad message type and version: type=%s version=%u",
 240                                         msg_type_to_str(msg->type),
 241                                         msg->version);
 242                         return -EPROTO;
 243                 }
 244                 break;
 245
 246         case KFILND_MSG_HELLO_REQ:
 247         case KFILND_MSG_HELLO_RSP:
 248                 if (msg->version != 0) {
 249                         KFILND_EP_ERROR(ep,
 250                                         "Bad message type and version: type=%s version=%u",
 251                                         msg_type_to_str(msg->type),
 252                                         msg->version);
 253                         return -EPROTO;
 254                 }
 255                 break;
 256
 257         default:
 258                 CERROR("Unknown message type %x\n", msg->type);
 259                 return -EPROTO;
 260         }
 261         return 0;
 262 }
 263
 264 static void kfilnd_tn_record_state_change(struct kfilnd_transaction *tn)
 265 {
 266         unsigned int data_size_bucket =
 267                 kfilnd_msg_len_to_data_size_bucket(tn->lnet_msg_len);
 268         struct kfilnd_tn_duration_stat *stat;
 269
 270         if (tn->is_initiator)
 271                 stat = &tn->tn_ep->end_dev->initiator_state_stats.state[tn->tn_state].data_size[data_size_bucket];
 272         else
 273                 stat = &tn->tn_ep->end_dev->target_state_stats.state[tn->tn_state].data_size[data_size_bucket];
 274
 275         atomic64_add(ktime_to_ns(ktime_sub(ktime_get(), tn->tn_state_ts)),
 276                      &stat->accumulated_duration);
 277         atomic_inc(&stat->accumulated_count);
 278 }
 279
 280 static void kfilnd_tn_state_change(struct kfilnd_transaction *tn,
 281                                    enum tn_states new_state)
 282 {
 283         KFILND_TN_DEBUG(tn, "%s -> %s state change",
 284                         tn_state_to_str(tn->tn_state),
 285                         tn_state_to_str(new_state));
 286
 287         kfilnd_tn_record_state_change(tn);
 288
 289         tn->tn_state = new_state;
 290         tn->tn_state_ts = ktime_get();
 291 }
 292
 293 static void kfilnd_tn_status_update(struct kfilnd_transaction *tn, int status,
 294                                     enum lnet_msg_hstatus hstatus)
 295 {
 296         /* Only the first non-ok status will take. */
 297         if (tn->tn_status == 0) {
 298                 KFILND_TN_DEBUG(tn, "%d -> %d status change", tn->tn_status,
 299                                 status);
 300                 tn->tn_status = status;
 301         }
 302
 303         if (tn->hstatus == LNET_MSG_STATUS_OK) {
 304                 KFILND_TN_DEBUG(tn, "%d -> %d health status change",
 305                                 tn->hstatus, hstatus);
 306                 tn->hstatus = hstatus;
 307         }
 308 }
 309
 310 static bool kfilnd_tn_has_failed(struct kfilnd_transaction *tn)
 311 {
 312         return tn->tn_status != 0;
 313 }
 314
 315 /**
 316  * kfilnd_tn_process_rx_event() - Process an immediate receive event.
 317  *
 318  * For each immediate receive, a transaction structure needs to be allocated to
 319  * process the receive.
 320  */
 321 void kfilnd_tn_process_rx_event(struct kfilnd_immediate_buffer *bufdesc,
 322                                 struct kfilnd_msg *rx_msg, int msg_size)
 323 {
 324         struct kfilnd_transaction *tn;
 325         bool alloc_msg = true;
 326         int rc;
 327         enum tn_events event = TN_EVENT_RX_HELLO;
 328
 329         /* Increment buf ref count for this work */
 330         atomic_inc(&bufdesc->immed_ref);
 331
 332         /* Unpack the message */
 333         rc = kfilnd_tn_unpack_msg(bufdesc->immed_end, rx_msg, msg_size);
 334         if (rc || CFS_FAIL_CHECK(CFS_KFI_FAIL_MSG_UNPACK)) {
 335                 kfilnd_ep_imm_buffer_put(bufdesc);
 336                 KFILND_EP_ERROR(bufdesc->immed_end,
 337                                 "Failed to unpack message %d", rc);
 338                 return;
 339         }
 340
 341         switch ((enum kfilnd_msg_type)rx_msg->type) {
 342         case KFILND_MSG_IMMEDIATE:
 343         case KFILND_MSG_BULK_PUT_REQ:
 344         case KFILND_MSG_BULK_GET_REQ:
 345                 event = TN_EVENT_RX_OK;
 346                 fallthrough;
 347         case KFILND_MSG_HELLO_RSP:
 348                 alloc_msg = false;
 349                 fallthrough;
 350         case KFILND_MSG_HELLO_REQ:
 351                 /* Context points to a received buffer and status is the length.
 352                  * Allocate a Tn structure, set its values, then launch the
 353                  * receive.
 354                  */
 355                 tn = kfilnd_tn_alloc(bufdesc->immed_end->end_dev,
 356                                      bufdesc->immed_end->end_cpt,
 357                                      rx_msg->srcnid, alloc_msg, false,
 358                                      false);
 359                 if (IS_ERR(tn)) {
 360                         kfilnd_ep_imm_buffer_put(bufdesc);
 361                         KFILND_EP_ERROR(bufdesc->immed_end,
 362                                         "Failed to allocate transaction struct: rc=%ld",
 363                                         PTR_ERR(tn));
 364                         return;
 365                 }
 366
 367                 tn->tn_rx_msg.msg = rx_msg;
 368                 tn->tn_rx_msg.length = msg_size;
 369                 tn->tn_posted_buf = bufdesc;
 370
 371                 KFILND_EP_DEBUG(bufdesc->immed_end, "%s transaction ID %u",
 372                                 msg_type_to_str((enum kfilnd_msg_type)rx_msg->type),
 373                                 tn->tn_mr_key);
 374                 break;
 375
 376         default:
 377                 KFILND_EP_ERROR(bufdesc->immed_end,
 378                                 "Unhandled kfilnd message type: %d",
 379                                 (enum kfilnd_msg_type)rx_msg->type);
 380                 LBUG();
 381         };
 382
 383         kfilnd_tn_event_handler(tn, event, 0);
 384 }
 385
 386 static void kfilnd_tn_record_duration(struct kfilnd_transaction *tn)
 387 {
 388         unsigned int data_size_bucket =
 389                 kfilnd_msg_len_to_data_size_bucket(tn->lnet_msg_len);
 390         struct kfilnd_tn_duration_stat *stat;
 391
 392         if (tn->is_initiator)
 393                 stat = &tn->tn_ep->end_dev->initiator_stats.data_size[data_size_bucket];
 394         else
 395                 stat = &tn->tn_ep->end_dev->target_stats.data_size[data_size_bucket];
 396
 397         atomic64_add(ktime_to_ns(ktime_sub(ktime_get(), tn->tn_alloc_ts)),
 398                      &stat->accumulated_duration);
 399         atomic_inc(&stat->accumulated_count);
 400 }
 401
 402 /**
 403  * kfilnd_tn_finalize() - Cleanup resources and finalize LNet operation.
 404  *
 405  * All state machine functions should call kfilnd_tn_finalize() instead of
 406  * kfilnd_tn_free(). Once all expected asynchronous events have been received,
 407  * if the transaction lock has not been released, it will now be released,
 408  * transaction resources cleaned up, and LNet finalized will be called.
 409  */
 410 static void kfilnd_tn_finalize(struct kfilnd_transaction *tn, bool *tn_released)
 411 {
 412         if (!*tn_released) {
 413                 mutex_unlock(&tn->tn_lock);
 414                 *tn_released = true;
 415         }
 416
 417         /* Release the reference on the multi-receive buffer. */
 418         if (tn->tn_posted_buf)
 419                 kfilnd_ep_imm_buffer_put(tn->tn_posted_buf);
 420
 421         /* Finalize LNet operation. */
 422         if (tn->tn_lntmsg) {
 423                 tn->tn_lntmsg->msg_health_status = tn->hstatus;
 424                 lnet_finalize(tn->tn_lntmsg, tn->tn_status);
 425         }
 426
 427         if (tn->tn_getreply) {
 428                 tn->tn_getreply->msg_health_status = tn->hstatus;
 429                 lnet_set_reply_msg_len(tn->tn_ep->end_dev->kfd_ni,
 430                                        tn->tn_getreply,
 431                                        tn->tn_status ? 0 : tn->tn_nob);
 432                 lnet_finalize(tn->tn_getreply, tn->tn_status);
 433         }
 434
 435         if (KFILND_TN_PEER_VALID(tn))
 436                 kfilnd_peer_put(tn->tn_kp);
 437
 438         kfilnd_tn_record_state_change(tn);
 439         kfilnd_tn_record_duration(tn);
 440
 441         kfilnd_tn_free(tn);
 442 }
 443
 444 /**
 445  * kfilnd_tn_cancel_tag_recv() - Attempt to cancel a tagged receive.
 446  * @tn: Transaction to have tagged received cancelled.
 447  *
 448  * Return: 0 on success. Else, negative errno. If an error occurs, resources may
 449  * be leaked.
 450  */
 451 static int kfilnd_tn_cancel_tag_recv(struct kfilnd_transaction *tn)
 452 {
 453         int rc;
 454
 455         /* Issue a cancel. A return code of zero means the operation issued an
 456          * async cancel. A return code of -ENOENT means the tagged receive was
 457          * not found. The assumption here is that a tagged send landed thus
 458          * removing the tagged receive buffer from hardware. For both cases,
 459          * async events should occur.
 460          */
 461         rc = kfilnd_ep_cancel_tagged_recv(tn->tn_ep, tn);
 462         if (rc != 0 && rc != -ENOENT) {
 463                 KFILND_TN_ERROR(tn, "Failed to cancel tag receive. Resources may leak.");
 464                 return rc;
 465         }
 466
 467         return 0;
 468 }
 469
 470 static void kfilnd_tn_timeout_work(struct work_struct *work)
 471 {
 472         struct kfilnd_transaction *tn =
 473                 container_of(work, struct kfilnd_transaction, timeout_work);
 474
 475         KFILND_TN_ERROR(tn, "Bulk operation timeout");
 476         kfilnd_tn_event_handler(tn, TN_EVENT_TIMEOUT, 0);
 477 }
 478
 479 static void kfilnd_tn_timeout(cfs_timer_cb_arg_t data)
 480 {
 481         struct kfilnd_transaction *tn = cfs_from_timer(tn, data, timeout_timer);
 482
 483         queue_work(kfilnd_wq, &tn->timeout_work);
 484 }
 485
 486 static bool kfilnd_tn_timeout_cancel(struct kfilnd_transaction *tn)
 487 {
 488         return del_timer(&tn->timeout_timer);
 489 }
 490
 491 static void kfilnd_tn_timeout_enable(struct kfilnd_transaction *tn)
 492 {
 493         ktime_t remaining_time = max_t(ktime_t, 0,
 494                                        tn->deadline - ktime_get_seconds());
 495         unsigned long expires = remaining_time * HZ + jiffies;
 496
 497         if (CFS_FAIL_CHECK(CFS_KFI_FAIL_BULK_TIMEOUT))
 498                 expires = jiffies;
 499
 500         cfs_timer_setup(&tn->timeout_timer, kfilnd_tn_timeout,
 501                         (unsigned long)tn, 0);
 502         mod_timer(&tn->timeout_timer, expires);
 503 }
 504
 505 /*  The following are the state machine routines for the transactions. */
 506 static int kfilnd_tn_state_send_failed(struct kfilnd_transaction *tn,
 507                                        enum tn_events event, int status,
 508                                        bool *tn_released)
 509 {
 510         int rc;
 511
 512         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 513                         status);
 514
 515         switch (event) {
 516         case TN_EVENT_INIT_BULK:
 517                 /* Need to cancel the tagged receive to prevent resources from
 518                  * being leaked.
 519                  */
 520                 rc = kfilnd_tn_cancel_tag_recv(tn);
 521
 522                 switch (rc) {
 523                 /* Async event will progress transaction. */
 524                 case 0:
 525                         kfilnd_tn_state_change(tn, TN_STATE_FAIL);
 526                         return 0;
 527
 528                 /* Need to replay TN_EVENT_INIT_BULK event while in the
 529                  * TN_STATE_SEND_FAILED state.
 530                  */
 531                 case -EAGAIN:
 532                         KFILND_TN_DEBUG(tn,
 533                                         "Need to replay cancel tagged recv");
 534                         return -EAGAIN;
 535
 536                 default:
 537                         KFILND_TN_ERROR(tn,
 538                                         "Unexpected error during cancel tagged receive: rc=%d",
 539                                         rc);
 540                         LBUG();
 541                 }
 542                 break;
 543
 544         default:
 545                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
 546                 LBUG();
 547         }
 548 }
 549
 550 static int kfilnd_tn_state_tagged_recv_posted(struct kfilnd_transaction *tn,
 551                                               enum tn_events event, int status,
 552                                               bool *tn_released)
 553 {
 554         int rc;
 555
 556         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 557                         status);
 558
 559         switch (event) {
 560         case TN_EVENT_INIT_BULK:
 561                 tn->tn_target_addr = kfilnd_peer_get_kfi_addr(tn->tn_kp);
 562                 KFILND_TN_DEBUG(tn, "Using peer %s(%#llx)",
 563                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 564                                 tn->tn_target_addr);
 565
 566                 kfilnd_tn_pack_bulk_req(tn);
 567
 568                 rc = kfilnd_ep_post_send(tn->tn_ep, tn);
 569                 switch (rc) {
 570                 /* Async event will progress immediate send. */
 571                 case 0:
 572                         kfilnd_tn_state_change(tn, TN_STATE_WAIT_COMP);
 573                         return 0;
 574
 575                 /* Need to replay TN_EVENT_INIT_BULK event while in the
 576                  * TN_STATE_TAGGED_RECV_POSTED state.
 577                  */
 578                 case -EAGAIN:
 579                         KFILND_TN_DEBUG(tn,
 580                                         "Need to replay post send to %s(%#llx)",
 581                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 582                                         tn->tn_target_addr);
 583                         return -EAGAIN;
 584
 585                 /* Need to transition to the TN_STATE_SEND_FAILED to cleanup
 586                  * posted tagged receive buffer.
 587                  */
 588                 default:
 589                         KFILND_TN_ERROR(tn,
 590                                         "Failed to post send to %s(%#llx): rc=%d",
 591                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 592                                         tn->tn_target_addr, rc);
 593                         kfilnd_tn_status_update(tn, rc,
 594                                                 LNET_MSG_STATUS_LOCAL_ERROR);
 595                         kfilnd_tn_state_change(tn, TN_STATE_SEND_FAILED);
 596
 597                         /* Propogate TN_EVENT_INIT_BULK event to
 598                          * TN_STATE_SEND_FAILED handler.
 599                          */
 600                         return kfilnd_tn_state_send_failed(tn, event, rc,
 601                                                            tn_released);
 602                 }
 603
 604         default:
 605                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
 606                 LBUG();
 607         }
 608 }
 609
 610 static int kfilnd_tn_state_idle(struct kfilnd_transaction *tn,
 611                                 enum tn_events event, int status,
 612                                 bool *tn_released)
 613 {
 614         struct kfilnd_msg *msg;
 615         int rc;
 616         bool finalize = false;
 617         ktime_t remaining_time;
 618         struct lnet_hdr hdr;
 619         struct lnet_nid srcnid;
 620
 621         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 622                         status);
 623
 624         /* For new peers, send a hello request message and queue the true LNet
 625          * message for replay.
 626          */
 627         if (kfilnd_peer_is_new_peer(tn->tn_kp) &&
 628             (event == TN_EVENT_INIT_IMMEDIATE || event == TN_EVENT_INIT_BULK)) {
 629                 remaining_time = max_t(ktime_t, 0,
 630                                        tn->deadline - ktime_get_seconds());
 631
 632                 /* If transaction deadline has not be met, return -EAGAIN. This
 633                  * will cause this transaction event to be replayed. During this
 634                  * time, an async message from the peer should occur at which
 635                  * point the kfilnd version should be negotiated.
 636                  */
 637                 if (remaining_time > 0) {
 638                         KFILND_TN_DEBUG(tn, "%s hello response pending",
 639                                         libcfs_nid2str(tn->tn_kp->kp_nid));
 640                         return -EAGAIN;
 641                 }
 642
 643                 rc = 0;
 644                 kfilnd_tn_status_update(tn, -ETIMEDOUT,
 645                                         LNET_MSG_STATUS_NETWORK_TIMEOUT);
 646                 goto out;
 647         }
 648
 649         switch (event) {
 650         case TN_EVENT_INIT_IMMEDIATE:
 651         case TN_EVENT_TX_HELLO:
 652                 tn->tn_target_addr = kfilnd_peer_get_kfi_addr(tn->tn_kp);
 653                 KFILND_TN_DEBUG(tn, "Using peer %s(%#llx)",
 654                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 655                                 tn->tn_target_addr);
 656
 657                 if (event == TN_EVENT_INIT_IMMEDIATE)
 658                         kfilnd_tn_pack_immed_msg(tn);
 659                 else
 660                         kfilnd_tn_pack_hello_req(tn);
 661
 662                 /* Send immediate message. */
 663                 rc = kfilnd_ep_post_send(tn->tn_ep, tn);
 664                 switch (rc) {
 665                 /* Async event will progress immediate send. */
 666                 case 0:
 667                         kfilnd_tn_state_change(tn, TN_STATE_IMM_SEND);
 668                         return 0;
 669
 670                 /* Need to TN_EVENT_INIT_IMMEDIATE event while in TN_STATE_IDLE
 671                  * state.
 672                  */
 673                 case -EAGAIN:
 674                         KFILND_TN_DEBUG(tn, "Need to replay send to %s(%#llx)",
 675                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 676                                         tn->tn_target_addr);
 677                         return -EAGAIN;
 678
 679                 default:
 680                         KFILND_TN_ERROR(tn,
 681                                         "Failed to post send to %s(%#llx): rc=%d",
 682                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 683                                         tn->tn_target_addr, rc);
 684                         if (event == TN_EVENT_TX_HELLO)
 685                                 kfilnd_peer_clear_hello_pending(tn->tn_kp);
 686                         kfilnd_tn_status_update(tn, rc,
 687                                                 LNET_MSG_STATUS_LOCAL_ERROR);
 688                 }
 689                 break;
 690
 691         case TN_EVENT_INIT_BULK:
 692                 /* Post tagged receive buffer used to land bulk response. */
 693                 rc = kfilnd_ep_post_tagged_recv(tn->tn_ep, tn);
 694
 695                 switch (rc) {
 696                 /* Transition to TN_STATE_TAGGED_RECV_POSTED on success. */
 697                 case 0:
 698                         kfilnd_tn_state_change(tn, TN_STATE_TAGGED_RECV_POSTED);
 699
 700                         /* Propogate TN_EVENT_INIT_BULK event to
 701                          * TN_STATE_TAGGED_RECV_POSTED handler.
 702                          */
 703                         return kfilnd_tn_state_tagged_recv_posted(tn, event,
 704                                                                   rc,
 705                                                                   tn_released);
 706
 707                 /* Need to replay TN_EVENT_INIT_BULK event in the TN_STATE_IDLE
 708                  * state.
 709                  */
 710                 case -EAGAIN:
 711                         KFILND_TN_DEBUG(tn, "Need to replay tagged recv");
 712                         return -EAGAIN;
 713
 714                 default:
 715                         KFILND_TN_ERROR(tn, "Failed to post tagged recv %d",
 716                                         rc);
 717                         kfilnd_tn_status_update(tn, rc,
 718                                                 LNET_MSG_STATUS_LOCAL_ERROR);
 719                 }
 720                 break;
 721
 722         case TN_EVENT_RX_OK:
 723                 if (kfilnd_peer_needs_hello(tn->tn_kp)) {
 724                         rc = kfilnd_send_hello_request(tn->tn_ep->end_dev,
 725                                                        tn->tn_ep->end_cpt,
 726                                                        tn->tn_kp);
 727                         if (rc)
 728                                 KFILND_TN_ERROR(tn,
 729                                                 "Failed to send hello request: rc=%d",
 730                                                 rc);
 731                         rc = 0;
 732                 }
 733
 734                 /* If this is a new peer then we cannot progress the transaction
 735                  * and must drop it
 736                  */
 737                 if (kfilnd_peer_is_new_peer(tn->tn_kp)) {
 738                         KFILND_TN_ERROR(tn,
 739                                         "Dropping message from %s due to stale peer",
 740                                         libcfs_nid2str(tn->tn_kp->kp_nid));
 741                         kfilnd_tn_status_update(tn, -EPROTO,
 742                                                 LNET_MSG_STATUS_LOCAL_DROPPED);
 743                         rc = 0;
 744                         goto out;
 745                 }
 746
 747                 LASSERT(kfilnd_peer_is_new_peer(tn->tn_kp) == false);
 748                 msg = tn->tn_rx_msg.msg;
 749
 750                 /* Update the NID address with the new preferred RX context. */
 751                 kfilnd_peer_alive(tn->tn_kp);
 752
 753                 /* Pass message up to LNet
 754                  * The TN will be reused in this call chain so we need to
 755                  * release the lock on the TN before proceeding.
 756                  */
 757                 KFILND_TN_DEBUG(tn, "%s -> TN_STATE_IMM_RECV state change",
 758                                 tn_state_to_str(tn->tn_state));
 759
 760                 /* TODO: Do not manually update this state change. */
 761                 tn->tn_state = TN_STATE_IMM_RECV;
 762                 mutex_unlock(&tn->tn_lock);
 763                 *tn_released = true;
 764                 lnet_nid4_to_nid(msg->srcnid, &srcnid);
 765                 if (msg->type == KFILND_MSG_IMMEDIATE) {
 766                         lnet_hdr_from_nid4(&hdr, &msg->proto.immed.hdr);
 767                         rc = lnet_parse(tn->tn_ep->end_dev->kfd_ni,
 768                                         &hdr, &srcnid, tn, 0);
 769                 } else {
 770                         lnet_hdr_from_nid4(&hdr, &msg->proto.bulk_req.hdr);
 771                         rc = lnet_parse(tn->tn_ep->end_dev->kfd_ni,
 772                                         &hdr, &srcnid, tn, 1);
 773                 }
 774
 775                 /* If successful, transaction has been accepted by LNet and we
 776                  * cannot process the transaction anymore within this context.
 777                  */
 778                 if (!rc)
 779                         return 0;
 780
 781                 KFILND_TN_ERROR(tn, "Failed to parse LNet message: rc=%d", rc);
 782                 kfilnd_tn_status_update(tn, rc, LNET_MSG_STATUS_LOCAL_ERROR);
 783                 break;
 784
 785         case TN_EVENT_RX_HELLO:
 786                 msg = tn->tn_rx_msg.msg;
 787
 788                 switch (msg->type) {
 789                 case KFILND_MSG_HELLO_REQ:
 790                         kfilnd_peer_process_hello(tn->tn_kp, msg);
 791                         tn->tn_target_addr = kfilnd_peer_get_kfi_addr(tn->tn_kp);
 792                         KFILND_TN_DEBUG(tn, "Using peer %s(%#llx)",
 793                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 794                                         tn->tn_target_addr);
 795
 796                         kfilnd_tn_pack_hello_rsp(tn);
 797
 798                         /* Send immediate message. */
 799                         rc = kfilnd_ep_post_send(tn->tn_ep, tn);
 800                         switch (rc) {
 801                         case 0:
 802                                 kfilnd_tn_state_change(tn, TN_STATE_IMM_SEND);
 803                                 return 0;
 804
 805                         case -EAGAIN:
 806                                 KFILND_TN_DEBUG(tn, "Need to replay send to %s(%#llx)",
 807                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 808                                                 tn->tn_target_addr);
 809                                 return -EAGAIN;
 810
 811                         default:
 812                                 KFILND_TN_ERROR(tn,
 813                                                 "Failed to post send to %s(%#llx): rc=%d",
 814                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 815                                                 tn->tn_target_addr, rc);
 816                                 kfilnd_tn_status_update(tn, rc,
 817                                                         LNET_MSG_STATUS_LOCAL_ERROR);
 818                         }
 819                         break;
 820
 821                 case KFILND_MSG_HELLO_RSP:
 822                         rc = 0;
 823                         kfilnd_peer_process_hello(tn->tn_kp, msg);
 824                         finalize = true;
 825                         break;
 826
 827                 default:
 828                         KFILND_TN_ERROR(tn, "Invalid message type: %s",
 829                                         msg_type_to_str(msg->type));
 830                         LBUG();
 831                 }
 832                 break;
 833
 834         default:
 835                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
 836                 LBUG();
 837         }
 838
 839 out:
 840         if (kfilnd_tn_has_failed(tn))
 841                 finalize = true;
 842
 843         if (finalize)
 844                 kfilnd_tn_finalize(tn, tn_released);
 845
 846         return rc;
 847 }
 848
 849 static int kfilnd_tn_state_imm_send(struct kfilnd_transaction *tn,
 850                                     enum tn_events event, int status,
 851                                     bool *tn_released)
 852 {
 853         enum lnet_msg_hstatus hstatus;
 854
 855         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 856                         status);
 857
 858         switch (event) {
 859         case TN_EVENT_TX_FAIL:
 860                 if (status == -ETIMEDOUT || status == -EIO)
 861                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
 862                 else
 863                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
 864
 865                 kfilnd_tn_status_update(tn, status, hstatus);
 866                 kfilnd_peer_down(tn->tn_kp);
 867                 if (tn->msg_type == KFILND_MSG_HELLO_REQ)
 868                         kfilnd_peer_clear_hello_pending(tn->tn_kp);
 869                 break;
 870
 871         case TN_EVENT_TX_OK:
 872                 kfilnd_peer_alive(tn->tn_kp);
 873                 break;
 874
 875         default:
 876                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
 877                 LBUG();
 878         }
 879
 880         kfilnd_tn_finalize(tn, tn_released);
 881
 882         return 0;
 883 }
 884
 885 static int kfilnd_tn_state_imm_recv(struct kfilnd_transaction *tn,
 886                                     enum tn_events event, int status,
 887                                     bool *tn_released)
 888 {
 889         int rc = 0;
 890         bool finalize = false;
 891
 892         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 893                         status);
 894
 895         switch (event) {
 896         case TN_EVENT_INIT_TAG_RMA:
 897         case TN_EVENT_SKIP_TAG_RMA:
 898                 /* Release the buffer we received the request on. All relevant
 899                  * information to perform the RMA operation is stored in the
 900                  * transaction structure. This should be done before the RMA
 901                  * operation to prevent two contexts from potentially processing
 902                  * the same transaction.
 903                  *
 904                  * TODO: Prevent this from returning -EAGAIN.
 905                  */
 906                 if (tn->tn_posted_buf) {
 907                         kfilnd_ep_imm_buffer_put(tn->tn_posted_buf);
 908                         tn->tn_posted_buf = NULL;
 909                 }
 910
 911                 /* Update the KFI address to use the response RX context. */
 912                 tn->tn_target_addr =
 913                         kfi_rx_addr(KFILND_BASE_ADDR(tn->tn_kp->kp_addr),
 914                                     tn->tn_response_rx, KFILND_FAB_RX_CTX_BITS);
 915                 KFILND_TN_DEBUG(tn, "Using peer %s(0x%llx)",
 916                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 917                                 tn->tn_target_addr);
 918
 919                 /* Initiate the RMA operation to push/pull the LNet payload or
 920                  * send a tagged message to finalize the bulk operation if the
 921                  * RMA operation should be skipped.
 922                  */
 923                 if (event == TN_EVENT_INIT_TAG_RMA) {
 924                         if (tn->sink_buffer)
 925                                 rc = kfilnd_ep_post_read(tn->tn_ep, tn);
 926                         else
 927                                 rc = kfilnd_ep_post_write(tn->tn_ep, tn);
 928
 929                         switch (rc) {
 930                         /* Async tagged RMA event will progress transaction. */
 931                         case 0:
 932                                 kfilnd_tn_state_change(tn,
 933                                                        TN_STATE_WAIT_TAG_RMA_COMP);
 934                                 return 0;
 935
 936                         /* Need to replay TN_EVENT_INIT_TAG_RMA event while in
 937                          * the TN_STATE_IMM_RECV state.
 938                          */
 939                         case -EAGAIN:
 940                                 KFILND_TN_DEBUG(tn,
 941                                                 "Need to replay tagged %s to %s(%#llx)",
 942                                                 tn->sink_buffer ? "read" : "write",
 943                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 944                                                 tn->tn_target_addr);
 945                                 return -EAGAIN;
 946
 947                         default:
 948                                 KFILND_TN_ERROR(tn,
 949                                                 "Failed to post tagged %s to %s(%#llx): rc=%d",
 950                                                 tn->sink_buffer ? "read" : "write",
 951                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 952                                                 tn->tn_target_addr, rc);
 953                                 kfilnd_tn_status_update(tn, rc,
 954                                                         LNET_MSG_STATUS_LOCAL_ERROR);
 955                         }
 956                 } else {
 957                         kfilnd_tn_status_update(tn, status,
 958                                                 LNET_MSG_STATUS_OK);
 959
 960                         /* Since the LNet initiator has posted a unique tagged
 961                          * buffer specific for this LNet transaction and the
 962                          * LNet target has decide not to push/pull to/for the
 963                          * LNet initiator tagged buffer, a noop operation is
 964                          * done to this tagged buffer (i/e payload transfer size
 965                          * is zero). But, immediate data, which contains the
 966                          * LNet target status for the transaction, is sent to
 967                          * the LNet initiator. Immediate data only appears in
 968                          * the completion event at the LNet initiator and not in
 969                          * the tagged buffer.
 970                          */
 971                         tn->tagged_data = cpu_to_be64(abs(tn->tn_status));
 972
 973                         rc = kfilnd_ep_post_tagged_send(tn->tn_ep, tn);
 974                         switch (rc) {
 975                         /* Async tagged RMA event will progress transaction. */
 976                         case 0:
 977                                 kfilnd_tn_state_change(tn,
 978                                                        TN_STATE_WAIT_TAG_COMP);
 979                                 return 0;
 980
 981                         /* Need to replay TN_EVENT_SKIP_TAG_RMA event while in
 982                          * the TN_STATE_IMM_RECV state.
 983                          */
 984                         case -EAGAIN:
 985                                 KFILND_TN_DEBUG(tn,
 986                                                 "Need to replay tagged send to %s(%#llx)",
 987                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 988                                                 tn->tn_target_addr);
 989                                 return -EAGAIN;
 990
 991                         default:
 992                                 KFILND_TN_ERROR(tn,
 993                                                 "Failed to post tagged send to %s(%#llx): rc=%d",
 994                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 995                                                 tn->tn_target_addr, rc);
 996                                 kfilnd_tn_status_update(tn, rc,
 997                                                         LNET_MSG_STATUS_LOCAL_ERROR);
 998                         }
 999                 }
1000                 break;
1001
1002         case TN_EVENT_RX_OK:
1003                 finalize = true;
1004                 break;
1005
1006         default:
1007                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1008                 LBUG();
1009         }
1010
1011         if (kfilnd_tn_has_failed(tn))
1012                 finalize = true;
1013
1014         if (finalize)
1015                 kfilnd_tn_finalize(tn, tn_released);
1016
1017         return rc;
1018 }
1019
1020 static int kfilnd_tn_state_wait_comp(struct kfilnd_transaction *tn,
1021                                      enum tn_events event, int status,
1022                                      bool *tn_released)
1023 {
1024         int rc;
1025         enum lnet_msg_hstatus hstatus;
1026
1027         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1028                         status);
1029
1030         switch (event) {
1031         case TN_EVENT_TX_OK:
1032                 kfilnd_peer_alive(tn->tn_kp);
1033                 kfilnd_tn_timeout_enable(tn);
1034                 kfilnd_tn_state_change(tn, TN_STATE_WAIT_TAG_COMP);
1035                 break;
1036
1037         case TN_EVENT_TAG_RX_OK:
1038                 kfilnd_tn_state_change(tn, TN_STATE_WAIT_SEND_COMP);
1039                 break;
1040
1041         case TN_EVENT_TX_FAIL:
1042                 if (status == -ETIMEDOUT)
1043                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
1044                 else
1045                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
1046
1047                 kfilnd_tn_status_update(tn, status, hstatus);
1048                 kfilnd_peer_down(tn->tn_kp);
1049
1050                 /* Need to cancel the tagged receive to prevent resources from
1051                  * being leaked.
1052                  */
1053                 rc = kfilnd_tn_cancel_tag_recv(tn);
1054
1055                 switch (rc) {
1056                 /* Async cancel event will progress transaction. */
1057                 case 0:
1058                         kfilnd_tn_status_update(tn, status,
1059                                                 LNET_MSG_STATUS_LOCAL_ERROR);
1060                         kfilnd_tn_state_change(tn, TN_STATE_FAIL);
1061                         return 0;
1062
1063                 /* Need to replay TN_EVENT_INIT_BULK event while in the
1064                  * TN_STATE_SEND_FAILED state.
1065                  */
1066                 case -EAGAIN:
1067                         KFILND_TN_DEBUG(tn,
1068                                         "Need to replay cancel tagged recv");
1069                         return -EAGAIN;
1070
1071                 default:
1072                         KFILND_TN_ERROR(tn,
1073                                         "Unexpected error during cancel tagged receive: rc=%d",
1074                                         rc);
1075                         LBUG();
1076                 }
1077                 break;
1078
1079         case TN_EVENT_TAG_RX_FAIL:
1080                 kfilnd_tn_status_update(tn, status,
1081                                         LNET_MSG_STATUS_LOCAL_ERROR);
1082                 kfilnd_tn_state_change(tn, TN_STATE_FAIL);
1083                 break;
1084
1085         default:
1086                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1087                 LBUG();
1088         }
1089
1090         return 0;
1091 }
1092
1093 static int kfilnd_tn_state_wait_send_comp(struct kfilnd_transaction *tn,
1094                                           enum tn_events event, int status,
1095                                           bool *tn_released)
1096 {
1097         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1098                         status);
1099
1100         if (event == TN_EVENT_TX_OK) {
1101                 kfilnd_peer_alive(tn->tn_kp);
1102                 kfilnd_tn_finalize(tn, tn_released);
1103         } else {
1104                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1105                 LBUG();
1106         }
1107
1108         return 0;
1109 }
1110
1111 static int kfilnd_tn_state_wait_tag_rma_comp(struct kfilnd_transaction *tn,
1112                                              enum tn_events event, int status,
1113                                              bool *tn_released)
1114 {
1115         enum lnet_msg_hstatus hstatus;
1116
1117         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1118                         status);
1119
1120         switch (event) {
1121         case TN_EVENT_TAG_TX_OK:
1122                 kfilnd_peer_alive(tn->tn_kp);
1123                 break;
1124
1125         case TN_EVENT_TAG_TX_FAIL:
1126                 if (status == -ETIMEDOUT)
1127                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
1128                 else
1129                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
1130
1131                 kfilnd_tn_status_update(tn, status, hstatus);
1132                 kfilnd_peer_down(tn->tn_kp);
1133                 break;
1134
1135         default:
1136                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1137                 LBUG();
1138         }
1139
1140         kfilnd_tn_finalize(tn, tn_released);
1141
1142         return 0;
1143 }
1144
1145 static int kfilnd_tn_state_wait_tag_comp(struct kfilnd_transaction *tn,
1146                                          enum tn_events event, int status,
1147                                          bool *tn_released)
1148 {
1149         int rc;
1150         enum lnet_msg_hstatus hstatus;
1151
1152         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1153                         status);
1154
1155         switch (event) {
1156         case TN_EVENT_TAG_RX_FAIL:
1157         case TN_EVENT_TAG_RX_OK:
1158                 /* Status can be set for both TN_EVENT_TAG_RX_FAIL and
1159                  * TN_EVENT_TAG_RX_OK. For TN_EVENT_TAG_RX_OK, if status is set,
1160                  * LNet target returned -ENODATA.
1161                  */
1162                 if (status) {
1163                         if (event == TN_EVENT_TAG_RX_FAIL)
1164                                 kfilnd_tn_status_update(tn, status,
1165                                                         LNET_MSG_STATUS_LOCAL_ERROR);
1166                         else
1167                                 kfilnd_tn_status_update(tn, status,
1168                                                         LNET_MSG_STATUS_OK);
1169                 }
1170
1171                 if (!kfilnd_tn_timeout_cancel(tn)) {
1172                         kfilnd_tn_state_change(tn, TN_STATE_WAIT_TIMEOUT_COMP);
1173                         return 0;
1174                 }
1175                 break;
1176
1177         case TN_EVENT_TIMEOUT:
1178                 /* Need to cancel the tagged receive to prevent resources from
1179                  * being leaked.
1180                  */
1181                 rc = kfilnd_tn_cancel_tag_recv(tn);
1182
1183                 switch (rc) {
1184                 /* Async cancel event will progress transaction. */
1185                 case 0:
1186                         kfilnd_tn_state_change(tn,
1187                                                TN_STATE_WAIT_TIMEOUT_TAG_COMP);
1188                         return 0;
1189
1190                 /* Need to replay TN_EVENT_INIT_BULK event while in the
1191                  * TN_STATE_WAIT_TAG_COMP state.
1192                  */
1193                 case -EAGAIN:
1194                         KFILND_TN_DEBUG(tn,
1195                                         "Need to replay cancel tagged recv");
1196                         return -EAGAIN;
1197
1198                 default:
1199                         KFILND_TN_ERROR(tn,
1200                                         "Unexpected error during cancel tagged receive: rc=%d",
1201                                         rc);
1202                         LBUG();
1203                 }
1204                 break;
1205
1206         case TN_EVENT_TAG_TX_FAIL:
1207                 if (status == -ETIMEDOUT)
1208                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
1209                 else
1210                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
1211
1212                 kfilnd_tn_status_update(tn, status, hstatus);
1213                 kfilnd_peer_down(tn->tn_kp);
1214                 break;
1215
1216         case TN_EVENT_TAG_TX_OK:
1217                 kfilnd_peer_alive(tn->tn_kp);
1218                 break;
1219
1220         default:
1221                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1222                 LBUG();
1223         }
1224
1225         kfilnd_tn_finalize(tn, tn_released);
1226
1227         return 0;
1228 }
1229
1230 static int kfilnd_tn_state_fail(struct kfilnd_transaction *tn,
1231                                 enum tn_events event, int status,
1232                                 bool *tn_released)
1233 {
1234         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1235                         status);
1236
1237         switch (event) {
1238         case TN_EVENT_TX_FAIL:
1239                 kfilnd_peer_down(tn->tn_kp);
1240                 break;
1241
1242         case TN_EVENT_TX_OK:
1243                 kfilnd_peer_alive(tn->tn_kp);
1244                 break;
1245
1246         case TN_EVENT_TAG_RX_FAIL:
1247         case TN_EVENT_TAG_RX_CANCEL:
1248                 break;
1249
1250         default:
1251                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1252                 LBUG();
1253         }
1254
1255         kfilnd_tn_finalize(tn, tn_released);
1256
1257         return 0;
1258 }
1259
1260 static int kfilnd_tn_state_wait_timeout_tag_comp(struct kfilnd_transaction *tn,
1261                                                  enum tn_events event,
1262                                                  int status, bool *tn_released)
1263 {
1264         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1265                         status);
1266
1267         switch (event) {
1268         case TN_EVENT_TAG_RX_CANCEL:
1269                 kfilnd_tn_status_update(tn, -ETIMEDOUT,
1270                                         LNET_MSG_STATUS_REMOTE_TIMEOUT);
1271                 kfilnd_peer_down(tn->tn_kp);
1272                 break;
1273
1274         case TN_EVENT_TAG_RX_FAIL:
1275                 kfilnd_tn_status_update(tn, status,
1276                                         LNET_MSG_STATUS_LOCAL_ERROR);
1277                 break;
1278
1279         case TN_EVENT_TAG_RX_OK:
1280                 break;
1281
1282         default:
1283                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1284                 LBUG();
1285         }
1286
1287         kfilnd_tn_finalize(tn, tn_released);
1288
1289         return 0;
1290 }
1291
1292 static int kfilnd_tn_state_wait_timeout_comp(struct kfilnd_transaction *tn,
1293                                              enum tn_events event, int status,
1294                                              bool *tn_released)
1295 {
1296         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1297                         status);
1298
1299         if (event == TN_EVENT_TIMEOUT) {
1300                 kfilnd_tn_finalize(tn, tn_released);
1301         } else {
1302                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1303                 LBUG();
1304         }
1305
1306         return 0;
1307 }
1308
1309 static int
1310 (* const kfilnd_tn_state_dispatch_table[TN_STATE_MAX])(struct kfilnd_transaction *tn,
1311                                                        enum tn_events event,
1312                                                        int status,
1313                                                        bool *tn_released) = {
1314         [TN_STATE_IDLE] = kfilnd_tn_state_idle,
1315         [TN_STATE_WAIT_TAG_COMP] = kfilnd_tn_state_wait_tag_comp,
1316         [TN_STATE_IMM_SEND] = kfilnd_tn_state_imm_send,
1317         [TN_STATE_TAGGED_RECV_POSTED] = kfilnd_tn_state_tagged_recv_posted,
1318         [TN_STATE_SEND_FAILED] = kfilnd_tn_state_send_failed,
1319         [TN_STATE_WAIT_COMP] = kfilnd_tn_state_wait_comp,
1320         [TN_STATE_WAIT_TIMEOUT_COMP] = kfilnd_tn_state_wait_timeout_comp,
1321         [TN_STATE_WAIT_SEND_COMP] = kfilnd_tn_state_wait_send_comp,
1322         [TN_STATE_WAIT_TIMEOUT_TAG_COMP] =
1323                 kfilnd_tn_state_wait_timeout_tag_comp,
1324         [TN_STATE_FAIL] = kfilnd_tn_state_fail,
1325         [TN_STATE_IMM_RECV] = kfilnd_tn_state_imm_recv,
1326         [TN_STATE_WAIT_TAG_RMA_COMP] = kfilnd_tn_state_wait_tag_rma_comp,
1327 };
1328
1329 /**
1330  * kfilnd_tn_event_handler() - Update transaction state machine with an event.
1331  * @tn: Transaction to be updated.
1332  * @event: Transaction event.
1333  * @status: Errno status associated with the event.
1334  *
1335  * When the transaction event handler is first called on a new transaction, the
1336  * transaction is now own by the transaction system. This means that will be
1337  * freed by the system as the transaction is progressed through the state
1338  * machine.
1339  */
1340 void kfilnd_tn_event_handler(struct kfilnd_transaction *tn,
1341                              enum tn_events event, int status)
1342 {
1343         bool tn_released = false;
1344         int rc;
1345
1346         if (!tn)
1347                 return;
1348
1349         mutex_lock(&tn->tn_lock);
1350         rc = kfilnd_tn_state_dispatch_table[tn->tn_state](tn, event, status,
1351                                                           &tn_released);
1352         if (rc == -EAGAIN) {
1353                 tn->replay_event = event;
1354                 tn->replay_status = status;
1355                 kfilnd_ep_queue_tn_replay(tn->tn_ep, tn);
1356         }
1357
1358         if (!tn_released)
1359                 mutex_unlock(&tn->tn_lock);
1360 }
1361
1362 /**
1363  * kfilnd_tn_free() - Free a transaction.
1364  */
1365 void kfilnd_tn_free(struct kfilnd_transaction *tn)
1366 {
1367         spin_lock(&tn->tn_ep->tn_list_lock);
1368         list_del(&tn->tn_entry);
1369         spin_unlock(&tn->tn_ep->tn_list_lock);
1370
1371         KFILND_TN_DEBUG(tn, "Transaction freed");
1372
1373         if (tn->tn_mr_key)
1374                 kfilnd_ep_put_key(tn->tn_ep, tn->tn_mr_key);
1375
1376         /* Free send message buffer if needed. */
1377         if (tn->tn_tx_msg.msg)
1378                 kmem_cache_free(imm_buf_cache, tn->tn_tx_msg.msg);
1379
1380         kmem_cache_free(tn_cache, tn);
1381 }
1382
1383 /**
1384  * kfilnd_tn_alloc() - Allocate a new KFI LND transaction.
1385  * @dev: KFI LND device used to look the KFI LND endpoint to associate with the
1386  * transaction.
1387  * @cpt: CPT of the transaction.
1388  * @target_nid: Target NID of the transaction.
1389  * @alloc_msg: Allocate an immediate message for the transaction.
1390  * @is_initiator: Is initiator of LNet transaction.
1391  * @key: Is transaction memory region key need.
1392  *
1393  * During transaction allocation, each transaction is associated with a KFI LND
1394  * endpoint use to post data transfer operations. The CPT argument is used to
1395  * lookup the KFI LND endpoint within the KFI LND device.
1396  *
1397  * Return: On success, valid pointer. Else, negative errno pointer.
1398  */
1399 struct kfilnd_transaction *kfilnd_tn_alloc(struct kfilnd_dev *dev, int cpt,
1400                                            lnet_nid_t target_nid,
1401                                            bool alloc_msg, bool is_initiator,
1402                                            bool key)
1403 {
1404         struct kfilnd_transaction *tn;
1405         struct kfilnd_peer *kp;
1406         int rc;
1407
1408         if (!dev) {
1409                 rc = -EINVAL;
1410                 goto err;
1411         }
1412
1413         kp = kfilnd_peer_get(dev, target_nid);
1414         if (IS_ERR(kp)) {
1415                 rc = PTR_ERR(kp);
1416                 goto err;
1417         }
1418
1419         tn = kfilnd_tn_alloc_for_peer(dev, cpt, kp, alloc_msg, is_initiator,
1420                                       key);
1421         if (IS_ERR(tn)) {
1422                 rc = PTR_ERR(tn);
1423                 kfilnd_peer_put(kp);
1424                 goto err;
1425         }
1426
1427         return tn;
1428
1429 err:
1430         return ERR_PTR(rc);
1431 }
1432
1433 /* See kfilnd_tn_alloc()
1434  * Note: Caller must have a reference on @kp
1435  */
1436 struct kfilnd_transaction *kfilnd_tn_alloc_for_peer(struct kfilnd_dev *dev,
1437                                                     int cpt,
1438                                                     struct kfilnd_peer *kp,
1439                                                     bool alloc_msg,
1440                                                     bool is_initiator,
1441                                                     bool key)
1442 {
1443         struct kfilnd_transaction *tn;
1444         struct kfilnd_ep *ep;
1445         int rc;
1446         ktime_t tn_alloc_ts;
1447
1448         if (!dev) {
1449                 rc = -EINVAL;
1450                 goto err;
1451         }
1452
1453         tn_alloc_ts = ktime_get();
1454
1455         /* If the CPT does not fall into the LNet NI CPT range, force the CPT
1456          * into the LNet NI CPT range. This should never happen.
1457          */
1458         ep = dev->cpt_to_endpoint[cpt];
1459         if (!ep) {
1460                 CWARN("%s used invalid cpt=%d\n",
1461                       libcfs_nidstr(&dev->kfd_ni->ni_nid), cpt);
1462                 ep = dev->kfd_endpoints[0];
1463         }
1464
1465         tn = kmem_cache_zalloc(tn_cache, GFP_KERNEL);
1466         if (!tn) {
1467                 rc = -ENOMEM;
1468                 goto err;
1469         }
1470
1471         if (alloc_msg) {
1472                 tn->tn_tx_msg.msg = kmem_cache_alloc(imm_buf_cache, GFP_KERNEL);
1473                 if (!tn->tn_tx_msg.msg) {
1474                         rc = -ENOMEM;
1475                         goto err_free_tn;
1476                 }
1477         }
1478
1479         if (key) {
1480                 rc = kfilnd_ep_get_key(ep);
1481                 if (rc < 0)
1482                         goto err_free_tn;
1483                 tn->tn_mr_key = rc;
1484         }
1485
1486         tn->tn_kp = kp;
1487         refcount_inc(&kp->kp_cnt);
1488
1489         mutex_init(&tn->tn_lock);
1490         tn->tn_ep = ep;
1491         tn->tn_response_rx = ep->end_context_id;
1492         tn->tn_state = TN_STATE_IDLE;
1493         tn->hstatus = LNET_MSG_STATUS_OK;
1494         tn->deadline = ktime_get_seconds() + lnet_get_lnd_timeout();
1495         tn->is_initiator = is_initiator;
1496         INIT_WORK(&tn->timeout_work, kfilnd_tn_timeout_work);
1497
1498         /* Add the transaction to an endpoint.  This is like
1499          * incrementing a ref counter.
1500          */
1501         spin_lock(&ep->tn_list_lock);
1502         list_add_tail(&tn->tn_entry, &ep->tn_list);
1503         spin_unlock(&ep->tn_list_lock);
1504
1505         tn->tn_alloc_ts = tn_alloc_ts;
1506         tn->tn_state_ts = ktime_get();
1507
1508         KFILND_EP_DEBUG(ep, "Transaction ID %u allocated", tn->tn_mr_key);
1509
1510         return tn;
1511
1512 err_free_tn:
1513         if (tn->tn_tx_msg.msg)
1514                 kmem_cache_free(imm_buf_cache, tn->tn_tx_msg.msg);
1515         kmem_cache_free(tn_cache, tn);
1516 err:
1517         return ERR_PTR(rc);
1518 }
1519
1520 /**
1521  * kfilnd_tn_cleanup() - Cleanup KFI LND transaction system.
1522  *
1523  * This function should only be called when there are no outstanding
1524  * transactions.
1525  */
1526 void kfilnd_tn_cleanup(void)
1527 {
1528         kmem_cache_destroy(imm_buf_cache);
1529         kmem_cache_destroy(tn_cache);
1530 }
1531
1532 /**
1533  * kfilnd_tn_init() - Initialize KFI LND transaction system.
1534  *
1535  * Return: On success, zero. Else, negative errno.
1536  */
1537 int kfilnd_tn_init(void)
1538 {
1539         tn_cache = kmem_cache_create("kfilnd_tn",
1540                                      sizeof(struct kfilnd_transaction), 0,
1541                                      SLAB_HWCACHE_ALIGN, NULL);
1542         if (!tn_cache)
1543                 goto err;
1544
1545         imm_buf_cache = kmem_cache_create("kfilnd_imm_buf",
1546                                           KFILND_IMMEDIATE_MSG_SIZE, 0,
1547                                           SLAB_HWCACHE_ALIGN, NULL);
1548         if (!imm_buf_cache)
1549                 goto err_tn_cache_destroy;
1550
1551         return 0;
1552
1553 err_tn_cache_destroy:
1554         kmem_cache_destroy(tn_cache);
1555 err:
1556         return -ENOMEM;
1557 }
1558
1559 /**
1560  * kfilnd_tn_set_kiov_buf() - Set the buffer used for a transaction.
1561  * @tn: Transaction to have buffer set.
1562  * @kiov: LNet KIOV buffer.
1563  * @num_iov: Number of IOVs.
1564  * @offset: Offset into IOVs where the buffer starts.
1565  * @len: Length of the buffer.
1566  *
1567  * This function takes the user provided IOV, offset, and len, and sets the
1568  * transaction buffer. The user provided IOV is an LNet KIOV. When the
1569  * transaction buffer is configured, the user provided offset is applied
1570  * when the transaction buffer is configured (i.e. the transaction buffer
1571  * offset is zero).
1572  */
1573 int kfilnd_tn_set_kiov_buf(struct kfilnd_transaction *tn,
1574                            struct bio_vec *kiov, size_t num_iov,
1575                            size_t offset, size_t len)
1576 {
1577         size_t i;
1578         size_t cur_len = 0;
1579         size_t cur_offset = offset;
1580         size_t cur_iov = 0;
1581         size_t tmp_len;
1582         size_t tmp_offset;
1583
1584         for (i = 0; (i < num_iov) && (cur_len < len); i++) {
1585                 /* Skip KIOVs until a KIOV with a length less than the current
1586                  * offset is found.
1587                  */
1588                 if (kiov[i].bv_len <= cur_offset) {
1589                         cur_offset -= kiov[i].bv_len;
1590                         continue;
1591                 }
1592
1593                 tmp_len = kiov[i].bv_len - cur_offset;
1594                 tmp_offset = kiov[i].bv_len - tmp_len + kiov[i].bv_offset;
1595
1596                 if (tmp_len + cur_len > len)
1597                         tmp_len = len - cur_len;
1598
1599                 /* tn_kiov is an array of size LNET_MAX_IOV */
1600                 if (cur_iov >= LNET_MAX_IOV)
1601                         return -EINVAL;
1602
1603                 tn->tn_kiov[cur_iov].bv_page = kiov[i].bv_page;
1604                 tn->tn_kiov[cur_iov].bv_len = tmp_len;
1605                 tn->tn_kiov[cur_iov].bv_offset = tmp_offset;
1606
1607                 cur_iov++;
1608                 cur_len += tmp_len;
1609                 cur_offset = 0;
1610         }
1611
1612         tn->tn_num_iovec = cur_iov;
1613         tn->tn_nob = cur_len;
1614
1615         return 0;
1616 }