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