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         struct lnet_hdr hdr;
 618         struct lnet_nid srcnid;
 619
 620         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 621                         status);
 622
 623         /* For new peers, send a hello request message and queue the true LNet
 624          * message for replay.
 625          */
 626         if (kfilnd_peer_needs_throttle(tn->tn_kp) &&
 627             (event == TN_EVENT_INIT_IMMEDIATE || event == TN_EVENT_INIT_BULK)) {
 628                 if (kfilnd_peer_deleted(tn->tn_kp)) {
 629                         /* We'll assign a NETWORK_TIMEOUT message health status
 630                          * below because we don't know why this peer was marked
 631                          * for removal
 632                          */
 633                         rc = -ESTALE;
 634                         KFILND_TN_DEBUG(tn, "Drop message to deleted peer");
 635                 } else if (kfilnd_peer_needs_hello(tn->tn_kp, false)) {
 636                         /* We're throttling transactions to this peer until
 637                          * a handshake can be completed, but there is no HELLO
 638                          * currently in flight. This implies the HELLO has
 639                          * failed, and we should cancel this TN. Otherwise we
 640                          * are stuck waiting for the TN deadline.
 641                          *
 642                          * We assign NETWORK_TIMEOUT health status below because
 643                          * we do not know why the HELLO failed.
 644                          */
 645                         rc = -ECANCELED;
 646                         KFILND_TN_DEBUG(tn, "Cancel throttled TN");
 647                 } else if (ktime_after(tn->deadline, ktime_get_seconds())) {
 648                         /* If transaction deadline has not been met, return
 649                          * -EAGAIN. This will cause this transaction event to be
 650                          * replayed. During this time, an async message from the
 651                          * peer should occur at which point the kfilnd version
 652                          * should be negotiated.
 653                          */
 654                         KFILND_TN_DEBUG(tn, "hello response pending");
 655                         return -EAGAIN;
 656                 } else {
 657                         rc = -ETIMEDOUT;
 658                 }
 659
 660                 kfilnd_tn_status_update(tn, rc,
 661                                         LNET_MSG_STATUS_NETWORK_TIMEOUT);
 662                 rc = 0;
 663                 goto out;
 664         }
 665
 666         switch (event) {
 667         case TN_EVENT_INIT_IMMEDIATE:
 668         case TN_EVENT_TX_HELLO:
 669                 tn->tn_target_addr = kfilnd_peer_get_kfi_addr(tn->tn_kp);
 670                 KFILND_TN_DEBUG(tn, "Using peer %s(%#llx)",
 671                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 672                                 tn->tn_target_addr);
 673
 674                 if (event == TN_EVENT_INIT_IMMEDIATE)
 675                         kfilnd_tn_pack_immed_msg(tn);
 676                 else
 677                         kfilnd_tn_pack_hello_req(tn);
 678
 679                 /* Send immediate message. */
 680                 rc = kfilnd_ep_post_send(tn->tn_ep, tn);
 681                 switch (rc) {
 682                 /* Async event will progress immediate send. */
 683                 case 0:
 684                         kfilnd_tn_state_change(tn, TN_STATE_IMM_SEND);
 685                         return 0;
 686
 687                 /* Need to TN_EVENT_INIT_IMMEDIATE event while in TN_STATE_IDLE
 688                  * state.
 689                  */
 690                 case -EAGAIN:
 691                         KFILND_TN_DEBUG(tn, "Need to replay send to %s(%#llx)",
 692                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 693                                         tn->tn_target_addr);
 694                         return -EAGAIN;
 695
 696                 default:
 697                         KFILND_TN_ERROR(tn,
 698                                         "Failed to post send to %s(%#llx): rc=%d",
 699                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 700                                         tn->tn_target_addr, rc);
 701                         if (event == TN_EVENT_TX_HELLO)
 702                                 kfilnd_peer_clear_hello_pending(tn->tn_kp);
 703                         kfilnd_tn_status_update(tn, rc,
 704                                                 LNET_MSG_STATUS_LOCAL_ERROR);
 705                 }
 706                 break;
 707
 708         case TN_EVENT_INIT_BULK:
 709                 /* Post tagged receive buffer used to land bulk response. */
 710                 rc = kfilnd_ep_post_tagged_recv(tn->tn_ep, tn);
 711
 712                 switch (rc) {
 713                 /* Transition to TN_STATE_TAGGED_RECV_POSTED on success. */
 714                 case 0:
 715                         kfilnd_tn_state_change(tn, TN_STATE_TAGGED_RECV_POSTED);
 716
 717                         /* Propogate TN_EVENT_INIT_BULK event to
 718                          * TN_STATE_TAGGED_RECV_POSTED handler.
 719                          */
 720                         return kfilnd_tn_state_tagged_recv_posted(tn, event,
 721                                                                   rc,
 722                                                                   tn_released);
 723
 724                 /* Need to replay TN_EVENT_INIT_BULK event in the TN_STATE_IDLE
 725                  * state.
 726                  */
 727                 case -EAGAIN:
 728                         KFILND_TN_DEBUG(tn, "Need to replay tagged recv");
 729                         return -EAGAIN;
 730
 731                 default:
 732                         KFILND_TN_ERROR(tn, "Failed to post tagged recv %d",
 733                                         rc);
 734                         kfilnd_tn_status_update(tn, rc,
 735                                                 LNET_MSG_STATUS_LOCAL_ERROR);
 736                 }
 737                 break;
 738
 739         case TN_EVENT_RX_OK:
 740                 if (kfilnd_peer_needs_hello(tn->tn_kp, false)) {
 741                         rc = kfilnd_send_hello_request(tn->tn_ep->end_dev,
 742                                                        tn->tn_ep->end_cpt,
 743                                                        tn->tn_kp);
 744                         if (rc)
 745                                 KFILND_TN_ERROR(tn,
 746                                                 "Failed to send hello request: rc=%d",
 747                                                 rc);
 748                         rc = 0;
 749                 }
 750
 751                 /* If this is a new peer then we cannot progress the transaction
 752                  * and must drop it
 753                  */
 754                 if (kfilnd_peer_is_new_peer(tn->tn_kp)) {
 755                         KFILND_TN_ERROR(tn,
 756                                         "Dropping message from %s due to stale peer",
 757                                         libcfs_nid2str(tn->tn_kp->kp_nid));
 758                         kfilnd_tn_status_update(tn, -EPROTO,
 759                                                 LNET_MSG_STATUS_LOCAL_DROPPED);
 760                         rc = 0;
 761                         goto out;
 762                 }
 763
 764                 LASSERT(kfilnd_peer_is_new_peer(tn->tn_kp) == false);
 765                 msg = tn->tn_rx_msg.msg;
 766
 767                 /* Update the NID address with the new preferred RX context. */
 768                 kfilnd_peer_alive(tn->tn_kp);
 769
 770                 /* Pass message up to LNet
 771                  * The TN will be reused in this call chain so we need to
 772                  * release the lock on the TN before proceeding.
 773                  */
 774                 KFILND_TN_DEBUG(tn, "%s -> TN_STATE_IMM_RECV state change",
 775                                 tn_state_to_str(tn->tn_state));
 776
 777                 /* TODO: Do not manually update this state change. */
 778                 tn->tn_state = TN_STATE_IMM_RECV;
 779                 mutex_unlock(&tn->tn_lock);
 780                 *tn_released = true;
 781                 lnet_nid4_to_nid(msg->srcnid, &srcnid);
 782                 if (msg->type == KFILND_MSG_IMMEDIATE) {
 783                         lnet_hdr_from_nid4(&hdr, &msg->proto.immed.hdr);
 784                         rc = lnet_parse(tn->tn_ep->end_dev->kfd_ni,
 785                                         &hdr, &srcnid, tn, 0);
 786                 } else {
 787                         lnet_hdr_from_nid4(&hdr, &msg->proto.bulk_req.hdr);
 788                         rc = lnet_parse(tn->tn_ep->end_dev->kfd_ni,
 789                                         &hdr, &srcnid, tn, 1);
 790                 }
 791
 792                 /* If successful, transaction has been accepted by LNet and we
 793                  * cannot process the transaction anymore within this context.
 794                  */
 795                 if (!rc)
 796                         return 0;
 797
 798                 KFILND_TN_ERROR(tn, "Failed to parse LNet message: rc=%d", rc);
 799                 kfilnd_tn_status_update(tn, rc, LNET_MSG_STATUS_LOCAL_ERROR);
 800                 break;
 801
 802         case TN_EVENT_RX_HELLO:
 803                 msg = tn->tn_rx_msg.msg;
 804
 805                 kfilnd_peer_alive(tn->tn_kp);
 806
 807                 switch (msg->type) {
 808                 case KFILND_MSG_HELLO_REQ:
 809                         kfilnd_peer_process_hello(tn->tn_kp, msg);
 810                         tn->tn_target_addr = kfilnd_peer_get_kfi_addr(tn->tn_kp);
 811                         KFILND_TN_DEBUG(tn, "Using peer %s(%#llx)",
 812                                         libcfs_nid2str(tn->tn_kp->kp_nid),
 813                                         tn->tn_target_addr);
 814
 815                         kfilnd_tn_pack_hello_rsp(tn);
 816
 817                         /* Send immediate message. */
 818                         rc = kfilnd_ep_post_send(tn->tn_ep, tn);
 819                         switch (rc) {
 820                         case 0:
 821                                 kfilnd_tn_state_change(tn, TN_STATE_IMM_SEND);
 822                                 return 0;
 823
 824                         case -EAGAIN:
 825                                 KFILND_TN_DEBUG(tn, "Need to replay send to %s(%#llx)",
 826                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 827                                                 tn->tn_target_addr);
 828                                 return -EAGAIN;
 829
 830                         default:
 831                                 KFILND_TN_ERROR(tn,
 832                                                 "Failed to post send to %s(%#llx): rc=%d",
 833                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 834                                                 tn->tn_target_addr, rc);
 835                                 kfilnd_tn_status_update(tn, rc,
 836                                                         LNET_MSG_STATUS_LOCAL_ERROR);
 837                         }
 838                         break;
 839
 840                 case KFILND_MSG_HELLO_RSP:
 841                         rc = 0;
 842                         kfilnd_peer_process_hello(tn->tn_kp, msg);
 843                         finalize = true;
 844                         break;
 845
 846                 default:
 847                         KFILND_TN_ERROR(tn, "Invalid message type: %s",
 848                                         msg_type_to_str(msg->type));
 849                         LBUG();
 850                 }
 851                 break;
 852
 853         default:
 854                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
 855                 LBUG();
 856         }
 857
 858 out:
 859         if (kfilnd_tn_has_failed(tn))
 860                 finalize = true;
 861
 862         if (finalize)
 863                 kfilnd_tn_finalize(tn, tn_released);
 864
 865         return rc;
 866 }
 867
 868 static int kfilnd_tn_state_imm_send(struct kfilnd_transaction *tn,
 869                                     enum tn_events event, int status,
 870                                     bool *tn_released)
 871 {
 872         enum lnet_msg_hstatus hstatus;
 873
 874         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 875                         status);
 876
 877         switch (event) {
 878         case TN_EVENT_TX_FAIL:
 879                 if (status == -ETIMEDOUT || status == -EIO)
 880                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
 881                 else
 882                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
 883
 884                 kfilnd_tn_status_update(tn, status, hstatus);
 885                 kfilnd_peer_tn_failed(tn->tn_kp, status);
 886                 if (tn->msg_type == KFILND_MSG_HELLO_REQ)
 887                         kfilnd_peer_clear_hello_pending(tn->tn_kp);
 888                 break;
 889
 890         case TN_EVENT_TX_OK:
 891                 kfilnd_peer_alive(tn->tn_kp);
 892                 break;
 893
 894         default:
 895                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
 896                 LBUG();
 897         }
 898
 899         kfilnd_tn_finalize(tn, tn_released);
 900
 901         return 0;
 902 }
 903
 904 static int kfilnd_tn_state_imm_recv(struct kfilnd_transaction *tn,
 905                                     enum tn_events event, int status,
 906                                     bool *tn_released)
 907 {
 908         int rc = 0;
 909         bool finalize = false;
 910
 911         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
 912                         status);
 913
 914         switch (event) {
 915         case TN_EVENT_INIT_TAG_RMA:
 916         case TN_EVENT_SKIP_TAG_RMA:
 917                 /* Release the buffer we received the request on. All relevant
 918                  * information to perform the RMA operation is stored in the
 919                  * transaction structure. This should be done before the RMA
 920                  * operation to prevent two contexts from potentially processing
 921                  * the same transaction.
 922                  *
 923                  * TODO: Prevent this from returning -EAGAIN.
 924                  */
 925                 if (tn->tn_posted_buf) {
 926                         kfilnd_ep_imm_buffer_put(tn->tn_posted_buf);
 927                         tn->tn_posted_buf = NULL;
 928                 }
 929
 930                 /* Update the KFI address to use the response RX context. */
 931                 tn->tn_target_addr =
 932                         kfi_rx_addr(KFILND_BASE_ADDR(tn->tn_kp->kp_addr),
 933                                     tn->tn_response_rx, KFILND_FAB_RX_CTX_BITS);
 934                 KFILND_TN_DEBUG(tn, "Using peer %s(0x%llx)",
 935                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 936                                 tn->tn_target_addr);
 937
 938                 /* Initiate the RMA operation to push/pull the LNet payload or
 939                  * send a tagged message to finalize the bulk operation if the
 940                  * RMA operation should be skipped.
 941                  */
 942                 if (event == TN_EVENT_INIT_TAG_RMA) {
 943                         if (tn->sink_buffer)
 944                                 rc = kfilnd_ep_post_read(tn->tn_ep, tn);
 945                         else
 946                                 rc = kfilnd_ep_post_write(tn->tn_ep, tn);
 947
 948                         switch (rc) {
 949                         /* Async tagged RMA event will progress transaction. */
 950                         case 0:
 951                                 kfilnd_tn_state_change(tn,
 952                                                        TN_STATE_WAIT_TAG_RMA_COMP);
 953                                 return 0;
 954
 955                         /* Need to replay TN_EVENT_INIT_TAG_RMA event while in
 956                          * the TN_STATE_IMM_RECV state.
 957                          */
 958                         case -EAGAIN:
 959                                 KFILND_TN_DEBUG(tn,
 960                                                 "Need to replay tagged %s to %s(%#llx)",
 961                                                 tn->sink_buffer ? "read" : "write",
 962                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 963                                                 tn->tn_target_addr);
 964                                 return -EAGAIN;
 965
 966                         default:
 967                                 KFILND_TN_ERROR(tn,
 968                                                 "Failed to post tagged %s to %s(%#llx): rc=%d",
 969                                                 tn->sink_buffer ? "read" : "write",
 970                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
 971                                                 tn->tn_target_addr, rc);
 972                                 kfilnd_tn_status_update(tn, rc,
 973                                                         LNET_MSG_STATUS_LOCAL_ERROR);
 974                         }
 975                 } else {
 976                         kfilnd_tn_status_update(tn, status,
 977                                                 LNET_MSG_STATUS_OK);
 978
 979                         /* Since the LNet initiator has posted a unique tagged
 980                          * buffer specific for this LNet transaction and the
 981                          * LNet target has decide not to push/pull to/for the
 982                          * LNet initiator tagged buffer, a noop operation is
 983                          * done to this tagged buffer (i/e payload transfer size
 984                          * is zero). But, immediate data, which contains the
 985                          * LNet target status for the transaction, is sent to
 986                          * the LNet initiator. Immediate data only appears in
 987                          * the completion event at the LNet initiator and not in
 988                          * the tagged buffer.
 989                          */
 990                         tn->tagged_data = cpu_to_be64(abs(tn->tn_status));
 991
 992                         rc = kfilnd_ep_post_tagged_send(tn->tn_ep, tn);
 993                         switch (rc) {
 994                         /* Async tagged RMA event will progress transaction. */
 995                         case 0:
 996                                 kfilnd_tn_state_change(tn,
 997                                                        TN_STATE_WAIT_TAG_COMP);
 998                                 return 0;
 999
1000                         /* Need to replay TN_EVENT_SKIP_TAG_RMA event while in
1001                          * the TN_STATE_IMM_RECV state.
1002                          */
1003                         case -EAGAIN:
1004                                 KFILND_TN_DEBUG(tn,
1005                                                 "Need to replay tagged send to %s(%#llx)",
1006                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
1007                                                 tn->tn_target_addr);
1008                                 return -EAGAIN;
1009
1010                         default:
1011                                 KFILND_TN_ERROR(tn,
1012                                                 "Failed to post tagged send to %s(%#llx): rc=%d",
1013                                                 libcfs_nid2str(tn->tn_kp->kp_nid),
1014                                                 tn->tn_target_addr, rc);
1015                                 kfilnd_tn_status_update(tn, rc,
1016                                                         LNET_MSG_STATUS_LOCAL_ERROR);
1017                         }
1018                 }
1019                 break;
1020
1021         case TN_EVENT_RX_OK:
1022                 finalize = true;
1023                 break;
1024
1025         default:
1026                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1027                 LBUG();
1028         }
1029
1030         if (kfilnd_tn_has_failed(tn))
1031                 finalize = true;
1032
1033         if (finalize)
1034                 kfilnd_tn_finalize(tn, tn_released);
1035
1036         return rc;
1037 }
1038
1039 static int kfilnd_tn_state_wait_comp(struct kfilnd_transaction *tn,
1040                                      enum tn_events event, int status,
1041                                      bool *tn_released)
1042 {
1043         int rc;
1044         enum lnet_msg_hstatus hstatus;
1045
1046         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1047                         status);
1048
1049         switch (event) {
1050         case TN_EVENT_TX_OK:
1051                 kfilnd_peer_alive(tn->tn_kp);
1052                 kfilnd_tn_timeout_enable(tn);
1053                 kfilnd_tn_state_change(tn, TN_STATE_WAIT_TAG_COMP);
1054                 break;
1055
1056         case TN_EVENT_TAG_RX_OK:
1057                 kfilnd_tn_state_change(tn, TN_STATE_WAIT_SEND_COMP);
1058                 break;
1059
1060         case TN_EVENT_TX_FAIL:
1061                 if (status == -ETIMEDOUT)
1062                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
1063                 else
1064                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
1065
1066                 kfilnd_tn_status_update(tn, status, hstatus);
1067                 kfilnd_peer_tn_failed(tn->tn_kp, status);
1068
1069                 /* Need to cancel the tagged receive to prevent resources from
1070                  * being leaked.
1071                  */
1072                 rc = kfilnd_tn_cancel_tag_recv(tn);
1073
1074                 switch (rc) {
1075                 /* Async cancel event will progress transaction. */
1076                 case 0:
1077                         kfilnd_tn_status_update(tn, status,
1078                                                 LNET_MSG_STATUS_LOCAL_ERROR);
1079                         kfilnd_tn_state_change(tn, TN_STATE_FAIL);
1080                         return 0;
1081
1082                 /* Need to replay TN_EVENT_INIT_BULK event while in the
1083                  * TN_STATE_SEND_FAILED state.
1084                  */
1085                 case -EAGAIN:
1086                         KFILND_TN_DEBUG(tn,
1087                                         "Need to replay cancel tagged recv");
1088                         return -EAGAIN;
1089
1090                 default:
1091                         KFILND_TN_ERROR(tn,
1092                                         "Unexpected error during cancel tagged receive: rc=%d",
1093                                         rc);
1094                         LBUG();
1095                 }
1096                 break;
1097
1098         case TN_EVENT_TAG_RX_FAIL:
1099                 kfilnd_tn_status_update(tn, status,
1100                                         LNET_MSG_STATUS_LOCAL_ERROR);
1101                 kfilnd_tn_state_change(tn, TN_STATE_FAIL);
1102                 break;
1103
1104         default:
1105                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1106                 LBUG();
1107         }
1108
1109         return 0;
1110 }
1111
1112 static int kfilnd_tn_state_wait_send_comp(struct kfilnd_transaction *tn,
1113                                           enum tn_events event, int status,
1114                                           bool *tn_released)
1115 {
1116         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1117                         status);
1118
1119         if (event == TN_EVENT_TX_OK) {
1120                 kfilnd_peer_alive(tn->tn_kp);
1121                 kfilnd_tn_finalize(tn, tn_released);
1122         } else {
1123                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1124                 LBUG();
1125         }
1126
1127         return 0;
1128 }
1129
1130 static int kfilnd_tn_state_wait_tag_rma_comp(struct kfilnd_transaction *tn,
1131                                              enum tn_events event, int status,
1132                                              bool *tn_released)
1133 {
1134         enum lnet_msg_hstatus hstatus;
1135
1136         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1137                         status);
1138
1139         switch (event) {
1140         case TN_EVENT_TAG_TX_OK:
1141                 kfilnd_peer_alive(tn->tn_kp);
1142                 break;
1143
1144         case TN_EVENT_TAG_TX_FAIL:
1145                 if (status == -ETIMEDOUT)
1146                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
1147                 else
1148                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
1149
1150                 kfilnd_tn_status_update(tn, status, hstatus);
1151                 kfilnd_peer_tn_failed(tn->tn_kp, status);
1152                 break;
1153
1154         default:
1155                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1156                 LBUG();
1157         }
1158
1159         kfilnd_tn_finalize(tn, tn_released);
1160
1161         return 0;
1162 }
1163
1164 static int kfilnd_tn_state_wait_tag_comp(struct kfilnd_transaction *tn,
1165                                          enum tn_events event, int status,
1166                                          bool *tn_released)
1167 {
1168         int rc;
1169         enum lnet_msg_hstatus hstatus;
1170
1171         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1172                         status);
1173
1174         switch (event) {
1175         case TN_EVENT_TAG_RX_FAIL:
1176         case TN_EVENT_TAG_RX_OK:
1177                 /* Status can be set for both TN_EVENT_TAG_RX_FAIL and
1178                  * TN_EVENT_TAG_RX_OK. For TN_EVENT_TAG_RX_OK, if status is set,
1179                  * LNet target returned -ENODATA.
1180                  */
1181                 if (status) {
1182                         if (event == TN_EVENT_TAG_RX_FAIL)
1183                                 kfilnd_tn_status_update(tn, status,
1184                                                         LNET_MSG_STATUS_LOCAL_ERROR);
1185                         else
1186                                 kfilnd_tn_status_update(tn, status,
1187                                                         LNET_MSG_STATUS_OK);
1188                 }
1189
1190                 if (!kfilnd_tn_timeout_cancel(tn)) {
1191                         kfilnd_tn_state_change(tn, TN_STATE_WAIT_TIMEOUT_COMP);
1192                         return 0;
1193                 }
1194                 break;
1195
1196         case TN_EVENT_TIMEOUT:
1197                 /* Need to cancel the tagged receive to prevent resources from
1198                  * being leaked.
1199                  */
1200                 rc = kfilnd_tn_cancel_tag_recv(tn);
1201
1202                 switch (rc) {
1203                 /* Async cancel event will progress transaction. */
1204                 case 0:
1205                         kfilnd_tn_state_change(tn,
1206                                                TN_STATE_WAIT_TIMEOUT_TAG_COMP);
1207                         return 0;
1208
1209                 /* Need to replay TN_EVENT_INIT_BULK event while in the
1210                  * TN_STATE_WAIT_TAG_COMP state.
1211                  */
1212                 case -EAGAIN:
1213                         KFILND_TN_DEBUG(tn,
1214                                         "Need to replay cancel tagged recv");
1215                         return -EAGAIN;
1216
1217                 default:
1218                         KFILND_TN_ERROR(tn,
1219                                         "Unexpected error during cancel tagged receive: rc=%d",
1220                                         rc);
1221                         LBUG();
1222                 }
1223                 break;
1224
1225         case TN_EVENT_TAG_TX_FAIL:
1226                 if (status == -ETIMEDOUT)
1227                         hstatus = LNET_MSG_STATUS_NETWORK_TIMEOUT;
1228                 else
1229                         hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
1230
1231                 kfilnd_tn_status_update(tn, status, hstatus);
1232                 kfilnd_peer_tn_failed(tn->tn_kp, status);
1233                 break;
1234
1235         case TN_EVENT_TAG_TX_OK:
1236                 kfilnd_peer_alive(tn->tn_kp);
1237                 break;
1238
1239         default:
1240                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1241                 LBUG();
1242         }
1243
1244         kfilnd_tn_finalize(tn, tn_released);
1245
1246         return 0;
1247 }
1248
1249 static int kfilnd_tn_state_fail(struct kfilnd_transaction *tn,
1250                                 enum tn_events event, int status,
1251                                 bool *tn_released)
1252 {
1253         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1254                         status);
1255
1256         switch (event) {
1257         case TN_EVENT_TX_FAIL:
1258                 kfilnd_peer_tn_failed(tn->tn_kp, status);
1259                 break;
1260
1261         case TN_EVENT_TX_OK:
1262                 kfilnd_peer_alive(tn->tn_kp);
1263                 break;
1264
1265         case TN_EVENT_TAG_RX_FAIL:
1266         case TN_EVENT_TAG_RX_CANCEL:
1267                 break;
1268
1269         default:
1270                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1271                 LBUG();
1272         }
1273
1274         kfilnd_tn_finalize(tn, tn_released);
1275
1276         return 0;
1277 }
1278
1279 static int kfilnd_tn_state_wait_timeout_tag_comp(struct kfilnd_transaction *tn,
1280                                                  enum tn_events event,
1281                                                  int status, bool *tn_released)
1282 {
1283         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1284                         status);
1285
1286         switch (event) {
1287         case TN_EVENT_TAG_RX_CANCEL:
1288                 kfilnd_tn_status_update(tn, -ETIMEDOUT,
1289                                         LNET_MSG_STATUS_REMOTE_TIMEOUT);
1290                 kfilnd_peer_tn_failed(tn->tn_kp, -ETIMEDOUT);
1291                 break;
1292
1293         case TN_EVENT_TAG_RX_FAIL:
1294                 kfilnd_tn_status_update(tn, status,
1295                                         LNET_MSG_STATUS_LOCAL_ERROR);
1296                 break;
1297
1298         case TN_EVENT_TAG_RX_OK:
1299                 break;
1300
1301         default:
1302                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1303                 LBUG();
1304         }
1305
1306         kfilnd_tn_finalize(tn, tn_released);
1307
1308         return 0;
1309 }
1310
1311 static int kfilnd_tn_state_wait_timeout_comp(struct kfilnd_transaction *tn,
1312                                              enum tn_events event, int status,
1313                                              bool *tn_released)
1314 {
1315         KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
1316                         status);
1317
1318         if (event == TN_EVENT_TIMEOUT) {
1319                 kfilnd_tn_finalize(tn, tn_released);
1320         } else {
1321                 KFILND_TN_ERROR(tn, "Invalid %s event", tn_event_to_str(event));
1322                 LBUG();
1323         }
1324
1325         return 0;
1326 }
1327
1328 static int
1329 (* const kfilnd_tn_state_dispatch_table[TN_STATE_MAX])(struct kfilnd_transaction *tn,
1330                                                        enum tn_events event,
1331                                                        int status,
1332                                                        bool *tn_released) = {
1333         [TN_STATE_IDLE] = kfilnd_tn_state_idle,
1334         [TN_STATE_WAIT_TAG_COMP] = kfilnd_tn_state_wait_tag_comp,
1335         [TN_STATE_IMM_SEND] = kfilnd_tn_state_imm_send,
1336         [TN_STATE_TAGGED_RECV_POSTED] = kfilnd_tn_state_tagged_recv_posted,
1337         [TN_STATE_SEND_FAILED] = kfilnd_tn_state_send_failed,
1338         [TN_STATE_WAIT_COMP] = kfilnd_tn_state_wait_comp,
1339         [TN_STATE_WAIT_TIMEOUT_COMP] = kfilnd_tn_state_wait_timeout_comp,
1340         [TN_STATE_WAIT_SEND_COMP] = kfilnd_tn_state_wait_send_comp,
1341         [TN_STATE_WAIT_TIMEOUT_TAG_COMP] =
1342                 kfilnd_tn_state_wait_timeout_tag_comp,
1343         [TN_STATE_FAIL] = kfilnd_tn_state_fail,
1344         [TN_STATE_IMM_RECV] = kfilnd_tn_state_imm_recv,
1345         [TN_STATE_WAIT_TAG_RMA_COMP] = kfilnd_tn_state_wait_tag_rma_comp,
1346 };
1347
1348 /**
1349  * kfilnd_tn_event_handler() - Update transaction state machine with an event.
1350  * @tn: Transaction to be updated.
1351  * @event: Transaction event.
1352  * @status: Errno status associated with the event.
1353  *
1354  * When the transaction event handler is first called on a new transaction, the
1355  * transaction is now own by the transaction system. This means that will be
1356  * freed by the system as the transaction is progressed through the state
1357  * machine.
1358  */
1359 void kfilnd_tn_event_handler(struct kfilnd_transaction *tn,
1360                              enum tn_events event, int status)
1361 {
1362         bool tn_released = false;
1363         int rc;
1364
1365         if (!tn)
1366                 return;
1367
1368         mutex_lock(&tn->tn_lock);
1369         rc = kfilnd_tn_state_dispatch_table[tn->tn_state](tn, event, status,
1370                                                           &tn_released);
1371         if (rc == -EAGAIN) {
1372                 tn->replay_event = event;
1373                 tn->replay_status = status;
1374                 kfilnd_ep_queue_tn_replay(tn->tn_ep, tn);
1375         }
1376
1377         if (!tn_released)
1378                 mutex_unlock(&tn->tn_lock);
1379 }
1380
1381 /**
1382  * kfilnd_tn_free() - Free a transaction.
1383  */
1384 void kfilnd_tn_free(struct kfilnd_transaction *tn)
1385 {
1386         spin_lock(&tn->tn_ep->tn_list_lock);
1387         list_del(&tn->tn_entry);
1388         spin_unlock(&tn->tn_ep->tn_list_lock);
1389
1390         KFILND_TN_DEBUG(tn, "Transaction freed");
1391
1392         if (tn->tn_mr_key)
1393                 kfilnd_ep_put_key(tn->tn_ep, tn->tn_mr_key);
1394
1395         /* Free send message buffer if needed. */
1396         if (tn->tn_tx_msg.msg)
1397                 kmem_cache_free(imm_buf_cache, tn->tn_tx_msg.msg);
1398
1399         kmem_cache_free(tn_cache, tn);
1400 }
1401
1402 /**
1403  * kfilnd_tn_alloc() - Allocate a new KFI LND transaction.
1404  * @dev: KFI LND device used to look the KFI LND endpoint to associate with the
1405  * transaction.
1406  * @cpt: CPT of the transaction.
1407  * @target_nid: Target NID of the transaction.
1408  * @alloc_msg: Allocate an immediate message for the transaction.
1409  * @is_initiator: Is initiator of LNet transaction.
1410  * @key: Is transaction memory region key need.
1411  *
1412  * During transaction allocation, each transaction is associated with a KFI LND
1413  * endpoint use to post data transfer operations. The CPT argument is used to
1414  * lookup the KFI LND endpoint within the KFI LND device.
1415  *
1416  * Return: On success, valid pointer. Else, negative errno pointer.
1417  */
1418 struct kfilnd_transaction *kfilnd_tn_alloc(struct kfilnd_dev *dev, int cpt,
1419                                            lnet_nid_t target_nid,
1420                                            bool alloc_msg, bool is_initiator,
1421                                            bool key)
1422 {
1423         struct kfilnd_transaction *tn;
1424         struct kfilnd_peer *kp;
1425         int rc;
1426
1427         if (!dev) {
1428                 rc = -EINVAL;
1429                 goto err;
1430         }
1431
1432         kp = kfilnd_peer_get(dev, target_nid);
1433         if (IS_ERR(kp)) {
1434                 rc = PTR_ERR(kp);
1435                 goto err;
1436         }
1437
1438         tn = kfilnd_tn_alloc_for_peer(dev, cpt, kp, alloc_msg, is_initiator,
1439                                       key);
1440         if (IS_ERR(tn)) {
1441                 rc = PTR_ERR(tn);
1442                 kfilnd_peer_put(kp);
1443                 goto err;
1444         }
1445
1446         return tn;
1447
1448 err:
1449         return ERR_PTR(rc);
1450 }
1451
1452 /* See kfilnd_tn_alloc()
1453  * Note: Caller must have a reference on @kp
1454  */
1455 struct kfilnd_transaction *kfilnd_tn_alloc_for_peer(struct kfilnd_dev *dev,
1456                                                     int cpt,
1457                                                     struct kfilnd_peer *kp,
1458                                                     bool alloc_msg,
1459                                                     bool is_initiator,
1460                                                     bool key)
1461 {
1462         struct kfilnd_transaction *tn;
1463         struct kfilnd_ep *ep;
1464         int rc;
1465         ktime_t tn_alloc_ts;
1466
1467         if (!dev) {
1468                 rc = -EINVAL;
1469                 goto err;
1470         }
1471
1472         tn_alloc_ts = ktime_get();
1473
1474         /* If the CPT does not fall into the LNet NI CPT range, force the CPT
1475          * into the LNet NI CPT range. This should never happen.
1476          */
1477         ep = dev->cpt_to_endpoint[cpt];
1478         if (!ep) {
1479                 CWARN("%s used invalid cpt=%d\n",
1480                       libcfs_nidstr(&dev->kfd_ni->ni_nid), cpt);
1481                 ep = dev->kfd_endpoints[0];
1482         }
1483
1484         tn = kmem_cache_zalloc(tn_cache, GFP_KERNEL);
1485         if (!tn) {
1486                 rc = -ENOMEM;
1487                 goto err;
1488         }
1489
1490         if (alloc_msg) {
1491                 tn->tn_tx_msg.msg = kmem_cache_alloc(imm_buf_cache, GFP_KERNEL);
1492                 if (!tn->tn_tx_msg.msg) {
1493                         rc = -ENOMEM;
1494                         goto err_free_tn;
1495                 }
1496         }
1497
1498         if (key) {
1499                 rc = kfilnd_ep_get_key(ep);
1500                 if (rc < 0)
1501                         goto err_free_tn;
1502                 tn->tn_mr_key = rc;
1503         }
1504
1505         tn->tn_kp = kp;
1506         refcount_inc(&kp->kp_cnt);
1507
1508         mutex_init(&tn->tn_lock);
1509         tn->tn_ep = ep;
1510         tn->tn_response_rx = ep->end_context_id;
1511         tn->tn_state = TN_STATE_IDLE;
1512         tn->hstatus = LNET_MSG_STATUS_OK;
1513         tn->deadline = ktime_get_seconds() + lnet_get_lnd_timeout();
1514         tn->is_initiator = is_initiator;
1515         INIT_WORK(&tn->timeout_work, kfilnd_tn_timeout_work);
1516
1517         /* Add the transaction to an endpoint.  This is like
1518          * incrementing a ref counter.
1519          */
1520         spin_lock(&ep->tn_list_lock);
1521         list_add_tail(&tn->tn_entry, &ep->tn_list);
1522         spin_unlock(&ep->tn_list_lock);
1523
1524         tn->tn_alloc_ts = tn_alloc_ts;
1525         tn->tn_state_ts = ktime_get();
1526
1527         KFILND_EP_DEBUG(ep, "Transaction ID %u allocated", tn->tn_mr_key);
1528
1529         return tn;
1530
1531 err_free_tn:
1532         if (tn->tn_tx_msg.msg)
1533                 kmem_cache_free(imm_buf_cache, tn->tn_tx_msg.msg);
1534         kmem_cache_free(tn_cache, tn);
1535 err:
1536         return ERR_PTR(rc);
1537 }
1538
1539 /**
1540  * kfilnd_tn_cleanup() - Cleanup KFI LND transaction system.
1541  *
1542  * This function should only be called when there are no outstanding
1543  * transactions.
1544  */
1545 void kfilnd_tn_cleanup(void)
1546 {
1547         kmem_cache_destroy(imm_buf_cache);
1548         kmem_cache_destroy(tn_cache);
1549 }
1550
1551 /**
1552  * kfilnd_tn_init() - Initialize KFI LND transaction system.
1553  *
1554  * Return: On success, zero. Else, negative errno.
1555  */
1556 int kfilnd_tn_init(void)
1557 {
1558         tn_cache = kmem_cache_create("kfilnd_tn",
1559                                      sizeof(struct kfilnd_transaction), 0,
1560                                      SLAB_HWCACHE_ALIGN, NULL);
1561         if (!tn_cache)
1562                 goto err;
1563
1564         imm_buf_cache = kmem_cache_create("kfilnd_imm_buf",
1565                                           KFILND_IMMEDIATE_MSG_SIZE, 0,
1566                                           SLAB_HWCACHE_ALIGN, NULL);
1567         if (!imm_buf_cache)
1568                 goto err_tn_cache_destroy;
1569
1570         return 0;
1571
1572 err_tn_cache_destroy:
1573         kmem_cache_destroy(tn_cache);
1574 err:
1575         return -ENOMEM;
1576 }
1577
1578 /**
1579  * kfilnd_tn_set_kiov_buf() - Set the buffer used for a transaction.
1580  * @tn: Transaction to have buffer set.
1581  * @kiov: LNet KIOV buffer.
1582  * @num_iov: Number of IOVs.
1583  * @offset: Offset into IOVs where the buffer starts.
1584  * @len: Length of the buffer.
1585  *
1586  * This function takes the user provided IOV, offset, and len, and sets the
1587  * transaction buffer. The user provided IOV is an LNet KIOV. When the
1588  * transaction buffer is configured, the user provided offset is applied
1589  * when the transaction buffer is configured (i.e. the transaction buffer
1590  * offset is zero).
1591  */
1592 int kfilnd_tn_set_kiov_buf(struct kfilnd_transaction *tn,
1593                            struct bio_vec *kiov, size_t num_iov,
1594                            size_t offset, size_t len)
1595 {
1596         size_t i;
1597         size_t cur_len = 0;
1598         size_t cur_offset = offset;
1599         size_t cur_iov = 0;
1600         size_t tmp_len;
1601         size_t tmp_offset;
1602
1603         for (i = 0; (i < num_iov) && (cur_len < len); i++) {
1604                 /* Skip KIOVs until a KIOV with a length less than the current
1605                  * offset is found.
1606                  */
1607                 if (kiov[i].bv_len <= cur_offset) {
1608                         cur_offset -= kiov[i].bv_len;
1609                         continue;
1610                 }
1611
1612                 tmp_len = kiov[i].bv_len - cur_offset;
1613                 tmp_offset = kiov[i].bv_len - tmp_len + kiov[i].bv_offset;
1614
1615                 if (tmp_len + cur_len > len)
1616                         tmp_len = len - cur_len;
1617
1618                 /* tn_kiov is an array of size LNET_MAX_IOV */
1619                 if (cur_iov >= LNET_MAX_IOV)
1620                         return -EINVAL;
1621
1622                 tn->tn_kiov[cur_iov].bv_page = kiov[i].bv_page;
1623                 tn->tn_kiov[cur_iov].bv_len = tmp_len;
1624                 tn->tn_kiov[cur_iov].bv_offset = tmp_offset;
1625
1626                 cur_iov++;
1627                 cur_len += tmp_len;
1628                 cur_offset = 0;
1629         }
1630
1631         tn->tn_num_iovec = cur_iov;
1632         tn->tn_nob = cur_len;
1633
1634         return 0;
1635 }