fs/cifs/smbdirect.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  *   Copyright (C) 2017, Microsoft Corporation.
   4  *
   5  *   Author(s): Long Li <longli@microsoft.com>
   6  */
   7 #include <linux/module.h>
   8 #include <linux/highmem.h>
   9 #include "smbdirect.h"
  10 #include "cifs_debug.h"
  11 #include "cifsproto.h"
  12 #include "smb2proto.h"
  13
  14 static struct smbd_response *get_empty_queue_buffer(
  15                 struct smbd_connection *info);
  16 static struct smbd_response *get_receive_buffer(
  17                 struct smbd_connection *info);
  18 static void put_receive_buffer(
  19                 struct smbd_connection *info,
  20                 struct smbd_response *response);
  21 static int allocate_receive_buffers(struct smbd_connection *info, int num_buf);
  22 static void destroy_receive_buffers(struct smbd_connection *info);
  23
  24 static void put_empty_packet(
  25                 struct smbd_connection *info, struct smbd_response *response);
  26 static void enqueue_reassembly(
  27                 struct smbd_connection *info,
  28                 struct smbd_response *response, int data_length);
  29 static struct smbd_response *_get_first_reassembly(
  30                 struct smbd_connection *info);
  31
  32 static int smbd_post_recv(
  33                 struct smbd_connection *info,
  34                 struct smbd_response *response);
  35
  36 static int smbd_post_send_empty(struct smbd_connection *info);
  37 static int smbd_post_send_data(
  38                 struct smbd_connection *info,
  39                 struct kvec *iov, int n_vec, int remaining_data_length);
  40 static int smbd_post_send_page(struct smbd_connection *info,
  41                 struct page *page, unsigned long offset,
  42                 size_t size, int remaining_data_length);
  43
  44 static void destroy_mr_list(struct smbd_connection *info);
  45 static int allocate_mr_list(struct smbd_connection *info);
  46
  47 /* SMBD version number */
  48 #define SMBD_V1 0x0100
  49
  50 /* Port numbers for SMBD transport */
  51 #define SMB_PORT        445
  52 #define SMBD_PORT       5445
  53
  54 /* Address lookup and resolve timeout in ms */
  55 #define RDMA_RESOLVE_TIMEOUT    5000
  56
  57 /* SMBD negotiation timeout in seconds */
  58 #define SMBD_NEGOTIATE_TIMEOUT  120
  59
  60 /* SMBD minimum receive size and fragmented sized defined in [MS-SMBD] */
  61 #define SMBD_MIN_RECEIVE_SIZE           128
  62 #define SMBD_MIN_FRAGMENTED_SIZE        131072
  63
  64 /*
  65  * Default maximum number of RDMA read/write outstanding on this connection
  66  * This value is possibly decreased during QP creation on hardware limit
  67  */
  68 #define SMBD_CM_RESPONDER_RESOURCES     32
  69
  70 /* Maximum number of retries on data transfer operations */
  71 #define SMBD_CM_RETRY                   6
  72 /* No need to retry on Receiver Not Ready since SMBD manages credits */
  73 #define SMBD_CM_RNR_RETRY               0
  74
  75 /*
  76  * User configurable initial values per SMBD transport connection
  77  * as defined in [MS-SMBD] 3.1.1.1
  78  * Those may change after a SMBD negotiation
  79  */
  80 /* The local peer's maximum number of credits to grant to the peer */
  81 int smbd_receive_credit_max = 255;
  82
  83 /* The remote peer's credit request of local peer */
  84 int smbd_send_credit_target = 255;
  85
  86 /* The maximum single message size can be sent to remote peer */
  87 int smbd_max_send_size = 1364;
  88
  89 /*  The maximum fragmented upper-layer payload receive size supported */
  90 int smbd_max_fragmented_recv_size = 1024 * 1024;
  91
  92 /*  The maximum single-message size which can be received */
  93 int smbd_max_receive_size = 1364;
  94
  95 /* The timeout to initiate send of a keepalive message on idle */
  96 int smbd_keep_alive_interval = 120;
  97
  98 /*
  99  * User configurable initial values for RDMA transport
 100  * The actual values used may be lower and are limited to hardware capabilities
 101  */
 102 /* Default maximum number of pages in a single RDMA write/read */
 103 int smbd_max_frmr_depth = 2048;
 104
 105 /* If payload is less than this byte, use RDMA send/recv not read/write */
 106 int rdma_readwrite_threshold = 4096;
 107
 108 /* Transport logging functions
 109  * Logging are defined as classes. They can be OR'ed to define the actual
 110  * logging level via module parameter smbd_logging_class
 111  * e.g. cifs.smbd_logging_class=0xa0 will log all log_rdma_recv() and
 112  * log_rdma_event()
 113  */
 114 #define LOG_OUTGOING                    0x1
 115 #define LOG_INCOMING                    0x2
 116 #define LOG_READ                        0x4
 117 #define LOG_WRITE                       0x8
 118 #define LOG_RDMA_SEND                   0x10
 119 #define LOG_RDMA_RECV                   0x20
 120 #define LOG_KEEP_ALIVE                  0x40
 121 #define LOG_RDMA_EVENT                  0x80
 122 #define LOG_RDMA_MR                     0x100
 123 static unsigned int smbd_logging_class;
 124 module_param(smbd_logging_class, uint, 0644);
 125 MODULE_PARM_DESC(smbd_logging_class,
 126         "Logging class for SMBD transport 0x0 to 0x100");
 127
 128 #define ERR             0x0
 129 #define INFO            0x1
 130 static unsigned int smbd_logging_level = ERR;
 131 module_param(smbd_logging_level, uint, 0644);
 132 MODULE_PARM_DESC(smbd_logging_level,
 133         "Logging level for SMBD transport, 0 (default): error, 1: info");
 134
 135 #define log_rdma(level, class, fmt, args...)                            \
 136 do {                                                                    \
 137         if (level <= smbd_logging_level || class & smbd_logging_class)  \
 138                 cifs_dbg(VFS, "%s:%d " fmt, __func__, __LINE__, ##args);\
 139 } while (0)
 140
 141 #define log_outgoing(level, fmt, args...) \
 142                 log_rdma(level, LOG_OUTGOING, fmt, ##args)
 143 #define log_incoming(level, fmt, args...) \
 144                 log_rdma(level, LOG_INCOMING, fmt, ##args)
 145 #define log_read(level, fmt, args...)   log_rdma(level, LOG_READ, fmt, ##args)
 146 #define log_write(level, fmt, args...)  log_rdma(level, LOG_WRITE, fmt, ##args)
 147 #define log_rdma_send(level, fmt, args...) \
 148                 log_rdma(level, LOG_RDMA_SEND, fmt, ##args)
 149 #define log_rdma_recv(level, fmt, args...) \
 150                 log_rdma(level, LOG_RDMA_RECV, fmt, ##args)
 151 #define log_keep_alive(level, fmt, args...) \
 152                 log_rdma(level, LOG_KEEP_ALIVE, fmt, ##args)
 153 #define log_rdma_event(level, fmt, args...) \
 154                 log_rdma(level, LOG_RDMA_EVENT, fmt, ##args)
 155 #define log_rdma_mr(level, fmt, args...) \
 156                 log_rdma(level, LOG_RDMA_MR, fmt, ##args)
 157
 158 static void smbd_disconnect_rdma_work(struct work_struct *work)
 159 {
 160         struct smbd_connection *info =
 161                 container_of(work, struct smbd_connection, disconnect_work);
 162
 163         if (info->transport_status == SMBD_CONNECTED) {
 164                 info->transport_status = SMBD_DISCONNECTING;
 165                 rdma_disconnect(info->id);
 166         }
 167 }
 168
 169 static void smbd_disconnect_rdma_connection(struct smbd_connection *info)
 170 {
 171         queue_work(info->workqueue, &info->disconnect_work);
 172 }
 173
 174 /* Upcall from RDMA CM */
 175 static int smbd_conn_upcall(
 176                 struct rdma_cm_id *id, struct rdma_cm_event *event)
 177 {
 178         struct smbd_connection *info = id->context;
 179
 180         log_rdma_event(INFO, "event=%d status=%d\n",
 181                 event->event, event->status);
 182
 183         switch (event->event) {
 184         case RDMA_CM_EVENT_ADDR_RESOLVED:
 185         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 186                 info->ri_rc = 0;
 187                 complete(&info->ri_done);
 188                 break;
 189
 190         case RDMA_CM_EVENT_ADDR_ERROR:
 191                 info->ri_rc = -EHOSTUNREACH;
 192                 complete(&info->ri_done);
 193                 break;
 194
 195         case RDMA_CM_EVENT_ROUTE_ERROR:
 196                 info->ri_rc = -ENETUNREACH;
 197                 complete(&info->ri_done);
 198                 break;
 199
 200         case RDMA_CM_EVENT_ESTABLISHED:
 201                 log_rdma_event(INFO, "connected event=%d\n", event->event);
 202                 info->transport_status = SMBD_CONNECTED;
 203                 wake_up_interruptible(&info->conn_wait);
 204                 break;
 205
 206         case RDMA_CM_EVENT_CONNECT_ERROR:
 207         case RDMA_CM_EVENT_UNREACHABLE:
 208         case RDMA_CM_EVENT_REJECTED:
 209                 log_rdma_event(INFO, "connecting failed event=%d\n", event->event);
 210                 info->transport_status = SMBD_DISCONNECTED;
 211                 wake_up_interruptible(&info->conn_wait);
 212                 break;
 213
 214         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 215         case RDMA_CM_EVENT_DISCONNECTED:
 216                 /* This happenes when we fail the negotiation */
 217                 if (info->transport_status == SMBD_NEGOTIATE_FAILED) {
 218                         info->transport_status = SMBD_DISCONNECTED;
 219                         wake_up(&info->conn_wait);
 220                         break;
 221                 }
 222
 223                 info->transport_status = SMBD_DISCONNECTED;
 224                 wake_up_interruptible(&info->disconn_wait);
 225                 wake_up_interruptible(&info->wait_reassembly_queue);
 226                 wake_up_interruptible_all(&info->wait_send_queue);
 227                 break;
 228
 229         default:
 230                 break;
 231         }
 232
 233         return 0;
 234 }
 235
 236 /* Upcall from RDMA QP */
 237 static void
 238 smbd_qp_async_error_upcall(struct ib_event *event, void *context)
 239 {
 240         struct smbd_connection *info = context;
 241
 242         log_rdma_event(ERR, "%s on device %s info %p\n",
 243                 ib_event_msg(event->event), event->device->name, info);
 244
 245         switch (event->event) {
 246         case IB_EVENT_CQ_ERR:
 247         case IB_EVENT_QP_FATAL:
 248                 smbd_disconnect_rdma_connection(info);
 249                 break;
 250
 251         default:
 252                 break;
 253         }
 254 }
 255
 256 static inline void *smbd_request_payload(struct smbd_request *request)
 257 {
 258         return (void *)request->packet;
 259 }
 260
 261 static inline void *smbd_response_payload(struct smbd_response *response)
 262 {
 263         return (void *)response->packet;
 264 }
 265
 266 /* Called when a RDMA send is done */
 267 static void send_done(struct ib_cq *cq, struct ib_wc *wc)
 268 {
 269         int i;
 270         struct smbd_request *request =
 271                 container_of(wc->wr_cqe, struct smbd_request, cqe);
 272
 273         log_rdma_send(INFO, "smbd_request 0x%p completed wc->status=%d\n",
 274                 request, wc->status);
 275
 276         if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) {
 277                 log_rdma_send(ERR, "wc->status=%d wc->opcode=%d\n",
 278                         wc->status, wc->opcode);
 279                 smbd_disconnect_rdma_connection(request->info);
 280         }
 281
 282         for (i = 0; i < request->num_sge; i++)
 283                 ib_dma_unmap_single(request->info->id->device,
 284                         request->sge[i].addr,
 285                         request->sge[i].length,
 286                         DMA_TO_DEVICE);
 287
 288         if (atomic_dec_and_test(&request->info->send_pending))
 289                 wake_up(&request->info->wait_send_pending);
 290
 291         wake_up(&request->info->wait_post_send);
 292
 293         mempool_free(request, request->info->request_mempool);
 294 }
 295
 296 static void dump_smbd_negotiate_resp(struct smbd_negotiate_resp *resp)
 297 {
 298         log_rdma_event(INFO, "resp message min_version %u max_version %u negotiated_version %u credits_requested %u credits_granted %u status %u max_readwrite_size %u preferred_send_size %u max_receive_size %u max_fragmented_size %u\n",
 299                        resp->min_version, resp->max_version,
 300                        resp->negotiated_version, resp->credits_requested,
 301                        resp->credits_granted, resp->status,
 302                        resp->max_readwrite_size, resp->preferred_send_size,
 303                        resp->max_receive_size, resp->max_fragmented_size);
 304 }
 305
 306 /*
 307  * Process a negotiation response message, according to [MS-SMBD]3.1.5.7
 308  * response, packet_length: the negotiation response message
 309  * return value: true if negotiation is a success, false if failed
 310  */
 311 static bool process_negotiation_response(
 312                 struct smbd_response *response, int packet_length)
 313 {
 314         struct smbd_connection *info = response->info;
 315         struct smbd_negotiate_resp *packet = smbd_response_payload(response);
 316
 317         if (packet_length < sizeof(struct smbd_negotiate_resp)) {
 318                 log_rdma_event(ERR,
 319                         "error: packet_length=%d\n", packet_length);
 320                 return false;
 321         }
 322
 323         if (le16_to_cpu(packet->negotiated_version) != SMBD_V1) {
 324                 log_rdma_event(ERR, "error: negotiated_version=%x\n",
 325                         le16_to_cpu(packet->negotiated_version));
 326                 return false;
 327         }
 328         info->protocol = le16_to_cpu(packet->negotiated_version);
 329
 330         if (packet->credits_requested == 0) {
 331                 log_rdma_event(ERR, "error: credits_requested==0\n");
 332                 return false;
 333         }
 334         info->receive_credit_target = le16_to_cpu(packet->credits_requested);
 335
 336         if (packet->credits_granted == 0) {
 337                 log_rdma_event(ERR, "error: credits_granted==0\n");
 338                 return false;
 339         }
 340         atomic_set(&info->send_credits, le16_to_cpu(packet->credits_granted));
 341
 342         atomic_set(&info->receive_credits, 0);
 343
 344         if (le32_to_cpu(packet->preferred_send_size) > info->max_receive_size) {
 345                 log_rdma_event(ERR, "error: preferred_send_size=%d\n",
 346                         le32_to_cpu(packet->preferred_send_size));
 347                 return false;
 348         }
 349         info->max_receive_size = le32_to_cpu(packet->preferred_send_size);
 350
 351         if (le32_to_cpu(packet->max_receive_size) < SMBD_MIN_RECEIVE_SIZE) {
 352                 log_rdma_event(ERR, "error: max_receive_size=%d\n",
 353                         le32_to_cpu(packet->max_receive_size));
 354                 return false;
 355         }
 356         info->max_send_size = min_t(int, info->max_send_size,
 357                                         le32_to_cpu(packet->max_receive_size));
 358
 359         if (le32_to_cpu(packet->max_fragmented_size) <
 360                         SMBD_MIN_FRAGMENTED_SIZE) {
 361                 log_rdma_event(ERR, "error: max_fragmented_size=%d\n",
 362                         le32_to_cpu(packet->max_fragmented_size));
 363                 return false;
 364         }
 365         info->max_fragmented_send_size =
 366                 le32_to_cpu(packet->max_fragmented_size);
 367         info->rdma_readwrite_threshold =
 368                 rdma_readwrite_threshold > info->max_fragmented_send_size ?
 369                 info->max_fragmented_send_size :
 370                 rdma_readwrite_threshold;
 371
 372
 373         info->max_readwrite_size = min_t(u32,
 374                         le32_to_cpu(packet->max_readwrite_size),
 375                         info->max_frmr_depth * PAGE_SIZE);
 376         info->max_frmr_depth = info->max_readwrite_size / PAGE_SIZE;
 377
 378         return true;
 379 }
 380
 381 static void smbd_post_send_credits(struct work_struct *work)
 382 {
 383         int ret = 0;
 384         int use_receive_queue = 1;
 385         int rc;
 386         struct smbd_response *response;
 387         struct smbd_connection *info =
 388                 container_of(work, struct smbd_connection,
 389                         post_send_credits_work);
 390
 391         if (info->transport_status != SMBD_CONNECTED) {
 392                 wake_up(&info->wait_receive_queues);
 393                 return;
 394         }
 395
 396         if (info->receive_credit_target >
 397                 atomic_read(&info->receive_credits)) {
 398                 while (true) {
 399                         if (use_receive_queue)
 400                                 response = get_receive_buffer(info);
 401                         else
 402                                 response = get_empty_queue_buffer(info);
 403                         if (!response) {
 404                                 /* now switch to emtpy packet queue */
 405                                 if (use_receive_queue) {
 406                                         use_receive_queue = 0;
 407                                         continue;
 408                                 } else
 409                                         break;
 410                         }
 411
 412                         response->type = SMBD_TRANSFER_DATA;
 413                         response->first_segment = false;
 414                         rc = smbd_post_recv(info, response);
 415                         if (rc) {
 416                                 log_rdma_recv(ERR,
 417                                         "post_recv failed rc=%d\n", rc);
 418                                 put_receive_buffer(info, response);
 419                                 break;
 420                         }
 421
 422                         ret++;
 423                 }
 424         }
 425
 426         spin_lock(&info->lock_new_credits_offered);
 427         info->new_credits_offered += ret;
 428         spin_unlock(&info->lock_new_credits_offered);
 429
 430         /* Promptly send an immediate packet as defined in [MS-SMBD] 3.1.1.1 */
 431         info->send_immediate = true;
 432         if (atomic_read(&info->receive_credits) <
 433                 info->receive_credit_target - 1) {
 434                 if (info->keep_alive_requested == KEEP_ALIVE_PENDING ||
 435                     info->send_immediate) {
 436                         log_keep_alive(INFO, "send an empty message\n");
 437                         smbd_post_send_empty(info);
 438                 }
 439         }
 440 }
 441
 442 /* Called from softirq, when recv is done */
 443 static void recv_done(struct ib_cq *cq, struct ib_wc *wc)
 444 {
 445         struct smbd_data_transfer *data_transfer;
 446         struct smbd_response *response =
 447                 container_of(wc->wr_cqe, struct smbd_response, cqe);
 448         struct smbd_connection *info = response->info;
 449         int data_length = 0;
 450
 451         log_rdma_recv(INFO, "response=0x%p type=%d wc status=%d wc opcode %d byte_len=%d pkey_index=%u\n",
 452                       response, response->type, wc->status, wc->opcode,
 453                       wc->byte_len, wc->pkey_index);
 454
 455         if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) {
 456                 log_rdma_recv(INFO, "wc->status=%d opcode=%d\n",
 457                         wc->status, wc->opcode);
 458                 smbd_disconnect_rdma_connection(info);
 459                 goto error;
 460         }
 461
 462         ib_dma_sync_single_for_cpu(
 463                 wc->qp->device,
 464                 response->sge.addr,
 465                 response->sge.length,
 466                 DMA_FROM_DEVICE);
 467
 468         switch (response->type) {
 469         /* SMBD negotiation response */
 470         case SMBD_NEGOTIATE_RESP:
 471                 dump_smbd_negotiate_resp(smbd_response_payload(response));
 472                 info->full_packet_received = true;
 473                 info->negotiate_done =
 474                         process_negotiation_response(response, wc->byte_len);
 475                 complete(&info->negotiate_completion);
 476                 break;
 477
 478         /* SMBD data transfer packet */
 479         case SMBD_TRANSFER_DATA:
 480                 data_transfer = smbd_response_payload(response);
 481                 data_length = le32_to_cpu(data_transfer->data_length);
 482
 483                 /*
 484                  * If this is a packet with data playload place the data in
 485                  * reassembly queue and wake up the reading thread
 486                  */
 487                 if (data_length) {
 488                         if (info->full_packet_received)
 489                                 response->first_segment = true;
 490
 491                         if (le32_to_cpu(data_transfer->remaining_data_length))
 492                                 info->full_packet_received = false;
 493                         else
 494                                 info->full_packet_received = true;
 495
 496                         enqueue_reassembly(
 497                                 info,
 498                                 response,
 499                                 data_length);
 500                 } else
 501                         put_empty_packet(info, response);
 502
 503                 if (data_length)
 504                         wake_up_interruptible(&info->wait_reassembly_queue);
 505
 506                 atomic_dec(&info->receive_credits);
 507                 info->receive_credit_target =
 508                         le16_to_cpu(data_transfer->credits_requested);
 509                 if (le16_to_cpu(data_transfer->credits_granted)) {
 510                         atomic_add(le16_to_cpu(data_transfer->credits_granted),
 511                                 &info->send_credits);
 512                         /*
 513                          * We have new send credits granted from remote peer
 514                          * If any sender is waiting for credits, unblock it
 515                          */
 516                         wake_up_interruptible(&info->wait_send_queue);
 517                 }
 518
 519                 log_incoming(INFO, "data flags %d data_offset %d data_length %d remaining_data_length %d\n",
 520                              le16_to_cpu(data_transfer->flags),
 521                              le32_to_cpu(data_transfer->data_offset),
 522                              le32_to_cpu(data_transfer->data_length),
 523                              le32_to_cpu(data_transfer->remaining_data_length));
 524
 525                 /* Send a KEEP_ALIVE response right away if requested */
 526                 info->keep_alive_requested = KEEP_ALIVE_NONE;
 527                 if (le16_to_cpu(data_transfer->flags) &
 528                                 SMB_DIRECT_RESPONSE_REQUESTED) {
 529                         info->keep_alive_requested = KEEP_ALIVE_PENDING;
 530                 }
 531
 532                 return;
 533
 534         default:
 535                 log_rdma_recv(ERR,
 536                         "unexpected response type=%d\n", response->type);
 537         }
 538
 539 error:
 540         put_receive_buffer(info, response);
 541 }
 542
 543 static struct rdma_cm_id *smbd_create_id(
 544                 struct smbd_connection *info,
 545                 struct sockaddr *dstaddr, int port)
 546 {
 547         struct rdma_cm_id *id;
 548         int rc;
 549         __be16 *sport;
 550
 551         id = rdma_create_id(&init_net, smbd_conn_upcall, info,
 552                 RDMA_PS_TCP, IB_QPT_RC);
 553         if (IS_ERR(id)) {
 554                 rc = PTR_ERR(id);
 555                 log_rdma_event(ERR, "rdma_create_id() failed %i\n", rc);
 556                 return id;
 557         }
 558
 559         if (dstaddr->sa_family == AF_INET6)
 560                 sport = &((struct sockaddr_in6 *)dstaddr)->sin6_port;
 561         else
 562                 sport = &((struct sockaddr_in *)dstaddr)->sin_port;
 563
 564         *sport = htons(port);
 565
 566         init_completion(&info->ri_done);
 567         info->ri_rc = -ETIMEDOUT;
 568
 569         rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)dstaddr,
 570                 RDMA_RESOLVE_TIMEOUT);
 571         if (rc) {
 572                 log_rdma_event(ERR, "rdma_resolve_addr() failed %i\n", rc);
 573                 goto out;
 574         }
 575         rc = wait_for_completion_interruptible_timeout(
 576                 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
 577         /* e.g. if interrupted returns -ERESTARTSYS */
 578         if (rc < 0) {
 579                 log_rdma_event(ERR, "rdma_resolve_addr timeout rc: %i\n", rc);
 580                 goto out;
 581         }
 582         rc = info->ri_rc;
 583         if (rc) {
 584                 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc);
 585                 goto out;
 586         }
 587
 588         info->ri_rc = -ETIMEDOUT;
 589         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 590         if (rc) {
 591                 log_rdma_event(ERR, "rdma_resolve_route() failed %i\n", rc);
 592                 goto out;
 593         }
 594         rc = wait_for_completion_interruptible_timeout(
 595                 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
 596         /* e.g. if interrupted returns -ERESTARTSYS */
 597         if (rc < 0)  {
 598                 log_rdma_event(ERR, "rdma_resolve_addr timeout rc: %i\n", rc);
 599                 goto out;
 600         }
 601         rc = info->ri_rc;
 602         if (rc) {
 603                 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc);
 604                 goto out;
 605         }
 606
 607         return id;
 608
 609 out:
 610         rdma_destroy_id(id);
 611         return ERR_PTR(rc);
 612 }
 613
 614 /*
 615  * Test if FRWR (Fast Registration Work Requests) is supported on the device
 616  * This implementation requries FRWR on RDMA read/write
 617  * return value: true if it is supported
 618  */
 619 static bool frwr_is_supported(struct ib_device_attr *attrs)
 620 {
 621         if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
 622                 return false;
 623         if (attrs->max_fast_reg_page_list_len == 0)
 624                 return false;
 625         return true;
 626 }
 627
 628 static int smbd_ia_open(
 629                 struct smbd_connection *info,
 630                 struct sockaddr *dstaddr, int port)
 631 {
 632         int rc;
 633
 634         info->id = smbd_create_id(info, dstaddr, port);
 635         if (IS_ERR(info->id)) {
 636                 rc = PTR_ERR(info->id);
 637                 goto out1;
 638         }
 639
 640         if (!frwr_is_supported(&info->id->device->attrs)) {
 641                 log_rdma_event(ERR, "Fast Registration Work Requests (FRWR) is not supported\n");
 642                 log_rdma_event(ERR, "Device capability flags = %llx max_fast_reg_page_list_len = %u\n",
 643                                info->id->device->attrs.device_cap_flags,
 644                                info->id->device->attrs.max_fast_reg_page_list_len);
 645                 rc = -EPROTONOSUPPORT;
 646                 goto out2;
 647         }
 648         info->max_frmr_depth = min_t(int,
 649                 smbd_max_frmr_depth,
 650                 info->id->device->attrs.max_fast_reg_page_list_len);
 651         info->mr_type = IB_MR_TYPE_MEM_REG;
 652         if (info->id->device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
 653                 info->mr_type = IB_MR_TYPE_SG_GAPS;
 654
 655         info->pd = ib_alloc_pd(info->id->device, 0);
 656         if (IS_ERR(info->pd)) {
 657                 rc = PTR_ERR(info->pd);
 658                 log_rdma_event(ERR, "ib_alloc_pd() returned %d\n", rc);
 659                 goto out2;
 660         }
 661
 662         return 0;
 663
 664 out2:
 665         rdma_destroy_id(info->id);
 666         info->id = NULL;
 667
 668 out1:
 669         return rc;
 670 }
 671
 672 /*
 673  * Send a negotiation request message to the peer
 674  * The negotiation procedure is in [MS-SMBD] 3.1.5.2 and 3.1.5.3
 675  * After negotiation, the transport is connected and ready for
 676  * carrying upper layer SMB payload
 677  */
 678 static int smbd_post_send_negotiate_req(struct smbd_connection *info)
 679 {
 680         struct ib_send_wr send_wr;
 681         int rc = -ENOMEM;
 682         struct smbd_request *request;
 683         struct smbd_negotiate_req *packet;
 684
 685         request = mempool_alloc(info->request_mempool, GFP_KERNEL);
 686         if (!request)
 687                 return rc;
 688
 689         request->info = info;
 690
 691         packet = smbd_request_payload(request);
 692         packet->min_version = cpu_to_le16(SMBD_V1);
 693         packet->max_version = cpu_to_le16(SMBD_V1);
 694         packet->reserved = 0;
 695         packet->credits_requested = cpu_to_le16(info->send_credit_target);
 696         packet->preferred_send_size = cpu_to_le32(info->max_send_size);
 697         packet->max_receive_size = cpu_to_le32(info->max_receive_size);
 698         packet->max_fragmented_size =
 699                 cpu_to_le32(info->max_fragmented_recv_size);
 700
 701         request->num_sge = 1;
 702         request->sge[0].addr = ib_dma_map_single(
 703                                 info->id->device, (void *)packet,
 704                                 sizeof(*packet), DMA_TO_DEVICE);
 705         if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
 706                 rc = -EIO;
 707                 goto dma_mapping_failed;
 708         }
 709
 710         request->sge[0].length = sizeof(*packet);
 711         request->sge[0].lkey = info->pd->local_dma_lkey;
 712
 713         ib_dma_sync_single_for_device(
 714                 info->id->device, request->sge[0].addr,
 715                 request->sge[0].length, DMA_TO_DEVICE);
 716
 717         request->cqe.done = send_done;
 718
 719         send_wr.next = NULL;
 720         send_wr.wr_cqe = &request->cqe;
 721         send_wr.sg_list = request->sge;
 722         send_wr.num_sge = request->num_sge;
 723         send_wr.opcode = IB_WR_SEND;
 724         send_wr.send_flags = IB_SEND_SIGNALED;
 725
 726         log_rdma_send(INFO, "sge addr=0x%llx length=%u lkey=0x%x\n",
 727                 request->sge[0].addr,
 728                 request->sge[0].length, request->sge[0].lkey);
 729
 730         atomic_inc(&info->send_pending);
 731         rc = ib_post_send(info->id->qp, &send_wr, NULL);
 732         if (!rc)
 733                 return 0;
 734
 735         /* if we reach here, post send failed */
 736         log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
 737         atomic_dec(&info->send_pending);
 738         ib_dma_unmap_single(info->id->device, request->sge[0].addr,
 739                 request->sge[0].length, DMA_TO_DEVICE);
 740
 741         smbd_disconnect_rdma_connection(info);
 742
 743 dma_mapping_failed:
 744         mempool_free(request, info->request_mempool);
 745         return rc;
 746 }
 747
 748 /*
 749  * Extend the credits to remote peer
 750  * This implements [MS-SMBD] 3.1.5.9
 751  * The idea is that we should extend credits to remote peer as quickly as
 752  * it's allowed, to maintain data flow. We allocate as much receive
 753  * buffer as possible, and extend the receive credits to remote peer
 754  * return value: the new credtis being granted.
 755  */
 756 static int manage_credits_prior_sending(struct smbd_connection *info)
 757 {
 758         int new_credits;
 759
 760         spin_lock(&info->lock_new_credits_offered);
 761         new_credits = info->new_credits_offered;
 762         info->new_credits_offered = 0;
 763         spin_unlock(&info->lock_new_credits_offered);
 764
 765         return new_credits;
 766 }
 767
 768 /*
 769  * Check if we need to send a KEEP_ALIVE message
 770  * The idle connection timer triggers a KEEP_ALIVE message when expires
 771  * SMB_DIRECT_RESPONSE_REQUESTED is set in the message flag to have peer send
 772  * back a response.
 773  * return value:
 774  * 1 if SMB_DIRECT_RESPONSE_REQUESTED needs to be set
 775  * 0: otherwise
 776  */
 777 static int manage_keep_alive_before_sending(struct smbd_connection *info)
 778 {
 779         if (info->keep_alive_requested == KEEP_ALIVE_PENDING) {
 780                 info->keep_alive_requested = KEEP_ALIVE_SENT;
 781                 return 1;
 782         }
 783         return 0;
 784 }
 785
 786 /* Post the send request */
 787 static int smbd_post_send(struct smbd_connection *info,
 788                 struct smbd_request *request)
 789 {
 790         struct ib_send_wr send_wr;
 791         int rc, i;
 792
 793         for (i = 0; i < request->num_sge; i++) {
 794                 log_rdma_send(INFO,
 795                         "rdma_request sge[%d] addr=0x%llx length=%u\n",
 796                         i, request->sge[i].addr, request->sge[i].length);
 797                 ib_dma_sync_single_for_device(
 798                         info->id->device,
 799                         request->sge[i].addr,
 800                         request->sge[i].length,
 801                         DMA_TO_DEVICE);
 802         }
 803
 804         request->cqe.done = send_done;
 805
 806         send_wr.next = NULL;
 807         send_wr.wr_cqe = &request->cqe;
 808         send_wr.sg_list = request->sge;
 809         send_wr.num_sge = request->num_sge;
 810         send_wr.opcode = IB_WR_SEND;
 811         send_wr.send_flags = IB_SEND_SIGNALED;
 812
 813         rc = ib_post_send(info->id->qp, &send_wr, NULL);
 814         if (rc) {
 815                 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
 816                 smbd_disconnect_rdma_connection(info);
 817                 rc = -EAGAIN;
 818         } else
 819                 /* Reset timer for idle connection after packet is sent */
 820                 mod_delayed_work(info->workqueue, &info->idle_timer_work,
 821                         info->keep_alive_interval*HZ);
 822
 823         return rc;
 824 }
 825
 826 static int smbd_post_send_sgl(struct smbd_connection *info,
 827         struct scatterlist *sgl, int data_length, int remaining_data_length)
 828 {
 829         int num_sgs;
 830         int i, rc;
 831         int header_length;
 832         struct smbd_request *request;
 833         struct smbd_data_transfer *packet;
 834         int new_credits;
 835         struct scatterlist *sg;
 836
 837 wait_credit:
 838         /* Wait for send credits. A SMBD packet needs one credit */
 839         rc = wait_event_interruptible(info->wait_send_queue,
 840                 atomic_read(&info->send_credits) > 0 ||
 841                 info->transport_status != SMBD_CONNECTED);
 842         if (rc)
 843                 goto err_wait_credit;
 844
 845         if (info->transport_status != SMBD_CONNECTED) {
 846                 log_outgoing(ERR, "disconnected not sending on wait_credit\n");
 847                 rc = -EAGAIN;
 848                 goto err_wait_credit;
 849         }
 850         if (unlikely(atomic_dec_return(&info->send_credits) < 0)) {
 851                 atomic_inc(&info->send_credits);
 852                 goto wait_credit;
 853         }
 854
 855 wait_send_queue:
 856         wait_event(info->wait_post_send,
 857                 atomic_read(&info->send_pending) < info->send_credit_target ||
 858                 info->transport_status != SMBD_CONNECTED);
 859
 860         if (info->transport_status != SMBD_CONNECTED) {
 861                 log_outgoing(ERR, "disconnected not sending on wait_send_queue\n");
 862                 rc = -EAGAIN;
 863                 goto err_wait_send_queue;
 864         }
 865
 866         if (unlikely(atomic_inc_return(&info->send_pending) >
 867                                 info->send_credit_target)) {
 868                 atomic_dec(&info->send_pending);
 869                 goto wait_send_queue;
 870         }
 871
 872         request = mempool_alloc(info->request_mempool, GFP_KERNEL);
 873         if (!request) {
 874                 rc = -ENOMEM;
 875                 goto err_alloc;
 876         }
 877
 878         request->info = info;
 879
 880         /* Fill in the packet header */
 881         packet = smbd_request_payload(request);
 882         packet->credits_requested = cpu_to_le16(info->send_credit_target);
 883
 884         new_credits = manage_credits_prior_sending(info);
 885         atomic_add(new_credits, &info->receive_credits);
 886         packet->credits_granted = cpu_to_le16(new_credits);
 887
 888         info->send_immediate = false;
 889
 890         packet->flags = 0;
 891         if (manage_keep_alive_before_sending(info))
 892                 packet->flags |= cpu_to_le16(SMB_DIRECT_RESPONSE_REQUESTED);
 893
 894         packet->reserved = 0;
 895         if (!data_length)
 896                 packet->data_offset = 0;
 897         else
 898                 packet->data_offset = cpu_to_le32(24);
 899         packet->data_length = cpu_to_le32(data_length);
 900         packet->remaining_data_length = cpu_to_le32(remaining_data_length);
 901         packet->padding = 0;
 902
 903         log_outgoing(INFO, "credits_requested=%d credits_granted=%d data_offset=%d data_length=%d remaining_data_length=%d\n",
 904                      le16_to_cpu(packet->credits_requested),
 905                      le16_to_cpu(packet->credits_granted),
 906                      le32_to_cpu(packet->data_offset),
 907                      le32_to_cpu(packet->data_length),
 908                      le32_to_cpu(packet->remaining_data_length));
 909
 910         /* Map the packet to DMA */
 911         header_length = sizeof(struct smbd_data_transfer);
 912         /* If this is a packet without payload, don't send padding */
 913         if (!data_length)
 914                 header_length = offsetof(struct smbd_data_transfer, padding);
 915
 916         request->num_sge = 1;
 917         request->sge[0].addr = ib_dma_map_single(info->id->device,
 918                                                  (void *)packet,
 919                                                  header_length,
 920                                                  DMA_TO_DEVICE);
 921         if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
 922                 rc = -EIO;
 923                 request->sge[0].addr = 0;
 924                 goto err_dma;
 925         }
 926
 927         request->sge[0].length = header_length;
 928         request->sge[0].lkey = info->pd->local_dma_lkey;
 929
 930         /* Fill in the packet data payload */
 931         num_sgs = sgl ? sg_nents(sgl) : 0;
 932         for_each_sg(sgl, sg, num_sgs, i) {
 933                 request->sge[i+1].addr =
 934                         ib_dma_map_page(info->id->device, sg_page(sg),
 935                                sg->offset, sg->length, DMA_TO_DEVICE);
 936                 if (ib_dma_mapping_error(
 937                                 info->id->device, request->sge[i+1].addr)) {
 938                         rc = -EIO;
 939                         request->sge[i+1].addr = 0;
 940                         goto err_dma;
 941                 }
 942                 request->sge[i+1].length = sg->length;
 943                 request->sge[i+1].lkey = info->pd->local_dma_lkey;
 944                 request->num_sge++;
 945         }
 946
 947         rc = smbd_post_send(info, request);
 948         if (!rc)
 949                 return 0;
 950
 951 err_dma:
 952         for (i = 0; i < request->num_sge; i++)
 953                 if (request->sge[i].addr)
 954                         ib_dma_unmap_single(info->id->device,
 955                                             request->sge[i].addr,
 956                                             request->sge[i].length,
 957                                             DMA_TO_DEVICE);
 958         mempool_free(request, info->request_mempool);
 959
 960         /* roll back receive credits and credits to be offered */
 961         spin_lock(&info->lock_new_credits_offered);
 962         info->new_credits_offered += new_credits;
 963         spin_unlock(&info->lock_new_credits_offered);
 964         atomic_sub(new_credits, &info->receive_credits);
 965
 966 err_alloc:
 967         if (atomic_dec_and_test(&info->send_pending))
 968                 wake_up(&info->wait_send_pending);
 969
 970 err_wait_send_queue:
 971         /* roll back send credits and pending */
 972         atomic_inc(&info->send_credits);
 973
 974 err_wait_credit:
 975         return rc;
 976 }
 977
 978 /*
 979  * Send a page
 980  * page: the page to send
 981  * offset: offset in the page to send
 982  * size: length in the page to send
 983  * remaining_data_length: remaining data to send in this payload
 984  */
 985 static int smbd_post_send_page(struct smbd_connection *info, struct page *page,
 986                 unsigned long offset, size_t size, int remaining_data_length)
 987 {
 988         struct scatterlist sgl;
 989
 990         sg_init_table(&sgl, 1);
 991         sg_set_page(&sgl, page, size, offset);
 992
 993         return smbd_post_send_sgl(info, &sgl, size, remaining_data_length);
 994 }
 995
 996 /*
 997  * Send an empty message
 998  * Empty message is used to extend credits to peer to for keep live
 999  * while there is no upper layer payload to send at the time
1000  */
1001 static int smbd_post_send_empty(struct smbd_connection *info)
1002 {
1003         info->count_send_empty++;
1004         return smbd_post_send_sgl(info, NULL, 0, 0);
1005 }
1006
1007 /*
1008  * Send a data buffer
1009  * iov: the iov array describing the data buffers
1010  * n_vec: number of iov array
1011  * remaining_data_length: remaining data to send following this packet
1012  * in segmented SMBD packet
1013  */
1014 static int smbd_post_send_data(
1015         struct smbd_connection *info, struct kvec *iov, int n_vec,
1016         int remaining_data_length)
1017 {
1018         int i;
1019         u32 data_length = 0;
1020         struct scatterlist sgl[SMBDIRECT_MAX_SEND_SGE - 1];
1021
1022         if (n_vec > SMBDIRECT_MAX_SEND_SGE - 1) {
1023                 cifs_dbg(VFS, "Can't fit data to SGL, n_vec=%d\n", n_vec);
1024                 return -EINVAL;
1025         }
1026
1027         sg_init_table(sgl, n_vec);
1028         for (i = 0; i < n_vec; i++) {
1029                 data_length += iov[i].iov_len;
1030                 sg_set_buf(&sgl[i], iov[i].iov_base, iov[i].iov_len);
1031         }
1032
1033         return smbd_post_send_sgl(info, sgl, data_length, remaining_data_length);
1034 }
1035
1036 /*
1037  * Post a receive request to the transport
1038  * The remote peer can only send data when a receive request is posted
1039  * The interaction is controlled by send/receive credit system
1040  */
1041 static int smbd_post_recv(
1042                 struct smbd_connection *info, struct smbd_response *response)
1043 {
1044         struct ib_recv_wr recv_wr;
1045         int rc = -EIO;
1046
1047         response->sge.addr = ib_dma_map_single(
1048                                 info->id->device, response->packet,
1049                                 info->max_receive_size, DMA_FROM_DEVICE);
1050         if (ib_dma_mapping_error(info->id->device, response->sge.addr))
1051                 return rc;
1052
1053         response->sge.length = info->max_receive_size;
1054         response->sge.lkey = info->pd->local_dma_lkey;
1055
1056         response->cqe.done = recv_done;
1057
1058         recv_wr.wr_cqe = &response->cqe;
1059         recv_wr.next = NULL;
1060         recv_wr.sg_list = &response->sge;
1061         recv_wr.num_sge = 1;
1062
1063         rc = ib_post_recv(info->id->qp, &recv_wr, NULL);
1064         if (rc) {
1065                 ib_dma_unmap_single(info->id->device, response->sge.addr,
1066                                     response->sge.length, DMA_FROM_DEVICE);
1067                 smbd_disconnect_rdma_connection(info);
1068                 log_rdma_recv(ERR, "ib_post_recv failed rc=%d\n", rc);
1069         }
1070
1071         return rc;
1072 }
1073
1074 /* Perform SMBD negotiate according to [MS-SMBD] 3.1.5.2 */
1075 static int smbd_negotiate(struct smbd_connection *info)
1076 {
1077         int rc;
1078         struct smbd_response *response = get_receive_buffer(info);
1079
1080         response->type = SMBD_NEGOTIATE_RESP;
1081         rc = smbd_post_recv(info, response);
1082         log_rdma_event(INFO, "smbd_post_recv rc=%d iov.addr=0x%llx iov.length=%u iov.lkey=0x%x\n",
1083                        rc, response->sge.addr,
1084                        response->sge.length, response->sge.lkey);
1085         if (rc)
1086                 return rc;
1087
1088         init_completion(&info->negotiate_completion);
1089         info->negotiate_done = false;
1090         rc = smbd_post_send_negotiate_req(info);
1091         if (rc)
1092                 return rc;
1093
1094         rc = wait_for_completion_interruptible_timeout(
1095                 &info->negotiate_completion, SMBD_NEGOTIATE_TIMEOUT * HZ);
1096         log_rdma_event(INFO, "wait_for_completion_timeout rc=%d\n", rc);
1097
1098         if (info->negotiate_done)
1099                 return 0;
1100
1101         if (rc == 0)
1102                 rc = -ETIMEDOUT;
1103         else if (rc == -ERESTARTSYS)
1104                 rc = -EINTR;
1105         else
1106                 rc = -ENOTCONN;
1107
1108         return rc;
1109 }
1110
1111 static void put_empty_packet(
1112                 struct smbd_connection *info, struct smbd_response *response)
1113 {
1114         spin_lock(&info->empty_packet_queue_lock);
1115         list_add_tail(&response->list, &info->empty_packet_queue);
1116         info->count_empty_packet_queue++;
1117         spin_unlock(&info->empty_packet_queue_lock);
1118
1119         queue_work(info->workqueue, &info->post_send_credits_work);
1120 }
1121
1122 /*
1123  * Implement Connection.FragmentReassemblyBuffer defined in [MS-SMBD] 3.1.1.1
1124  * This is a queue for reassembling upper layer payload and present to upper
1125  * layer. All the inncoming payload go to the reassembly queue, regardless of
1126  * if reassembly is required. The uuper layer code reads from the queue for all
1127  * incoming payloads.
1128  * Put a received packet to the reassembly queue
1129  * response: the packet received
1130  * data_length: the size of payload in this packet
1131  */
1132 static void enqueue_reassembly(
1133         struct smbd_connection *info,
1134         struct smbd_response *response,
1135         int data_length)
1136 {
1137         spin_lock(&info->reassembly_queue_lock);
1138         list_add_tail(&response->list, &info->reassembly_queue);
1139         info->reassembly_queue_length++;
1140         /*
1141          * Make sure reassembly_data_length is updated after list and
1142          * reassembly_queue_length are updated. On the dequeue side
1143          * reassembly_data_length is checked without a lock to determine
1144          * if reassembly_queue_length and list is up to date
1145          */
1146         virt_wmb();
1147         info->reassembly_data_length += data_length;
1148         spin_unlock(&info->reassembly_queue_lock);
1149         info->count_reassembly_queue++;
1150         info->count_enqueue_reassembly_queue++;
1151 }
1152
1153 /*
1154  * Get the first entry at the front of reassembly queue
1155  * Caller is responsible for locking
1156  * return value: the first entry if any, NULL if queue is empty
1157  */
1158 static struct smbd_response *_get_first_reassembly(struct smbd_connection *info)
1159 {
1160         struct smbd_response *ret = NULL;
1161
1162         if (!list_empty(&info->reassembly_queue)) {
1163                 ret = list_first_entry(
1164                         &info->reassembly_queue,
1165                         struct smbd_response, list);
1166         }
1167         return ret;
1168 }
1169
1170 static struct smbd_response *get_empty_queue_buffer(
1171                 struct smbd_connection *info)
1172 {
1173         struct smbd_response *ret = NULL;
1174         unsigned long flags;
1175
1176         spin_lock_irqsave(&info->empty_packet_queue_lock, flags);
1177         if (!list_empty(&info->empty_packet_queue)) {
1178                 ret = list_first_entry(
1179                         &info->empty_packet_queue,
1180                         struct smbd_response, list);
1181                 list_del(&ret->list);
1182                 info->count_empty_packet_queue--;
1183         }
1184         spin_unlock_irqrestore(&info->empty_packet_queue_lock, flags);
1185
1186         return ret;
1187 }
1188
1189 /*
1190  * Get a receive buffer
1191  * For each remote send, we need to post a receive. The receive buffers are
1192  * pre-allocated in advance.
1193  * return value: the receive buffer, NULL if none is available
1194  */
1195 static struct smbd_response *get_receive_buffer(struct smbd_connection *info)
1196 {
1197         struct smbd_response *ret = NULL;
1198         unsigned long flags;
1199
1200         spin_lock_irqsave(&info->receive_queue_lock, flags);
1201         if (!list_empty(&info->receive_queue)) {
1202                 ret = list_first_entry(
1203                         &info->receive_queue,
1204                         struct smbd_response, list);
1205                 list_del(&ret->list);
1206                 info->count_receive_queue--;
1207                 info->count_get_receive_buffer++;
1208         }
1209         spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1210
1211         return ret;
1212 }
1213
1214 /*
1215  * Return a receive buffer
1216  * Upon returning of a receive buffer, we can post new receive and extend
1217  * more receive credits to remote peer. This is done immediately after a
1218  * receive buffer is returned.
1219  */
1220 static void put_receive_buffer(
1221         struct smbd_connection *info, struct smbd_response *response)
1222 {
1223         unsigned long flags;
1224
1225         ib_dma_unmap_single(info->id->device, response->sge.addr,
1226                 response->sge.length, DMA_FROM_DEVICE);
1227
1228         spin_lock_irqsave(&info->receive_queue_lock, flags);
1229         list_add_tail(&response->list, &info->receive_queue);
1230         info->count_receive_queue++;
1231         info->count_put_receive_buffer++;
1232         spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1233
1234         queue_work(info->workqueue, &info->post_send_credits_work);
1235 }
1236
1237 /* Preallocate all receive buffer on transport establishment */
1238 static int allocate_receive_buffers(struct smbd_connection *info, int num_buf)
1239 {
1240         int i;
1241         struct smbd_response *response;
1242
1243         INIT_LIST_HEAD(&info->reassembly_queue);
1244         spin_lock_init(&info->reassembly_queue_lock);
1245         info->reassembly_data_length = 0;
1246         info->reassembly_queue_length = 0;
1247
1248         INIT_LIST_HEAD(&info->receive_queue);
1249         spin_lock_init(&info->receive_queue_lock);
1250         info->count_receive_queue = 0;
1251
1252         INIT_LIST_HEAD(&info->empty_packet_queue);
1253         spin_lock_init(&info->empty_packet_queue_lock);
1254         info->count_empty_packet_queue = 0;
1255
1256         init_waitqueue_head(&info->wait_receive_queues);
1257
1258         for (i = 0; i < num_buf; i++) {
1259                 response = mempool_alloc(info->response_mempool, GFP_KERNEL);
1260                 if (!response)
1261                         goto allocate_failed;
1262
1263                 response->info = info;
1264                 list_add_tail(&response->list, &info->receive_queue);
1265                 info->count_receive_queue++;
1266         }
1267
1268         return 0;
1269
1270 allocate_failed:
1271         while (!list_empty(&info->receive_queue)) {
1272                 response = list_first_entry(
1273                                 &info->receive_queue,
1274                                 struct smbd_response, list);
1275                 list_del(&response->list);
1276                 info->count_receive_queue--;
1277
1278                 mempool_free(response, info->response_mempool);
1279         }
1280         return -ENOMEM;
1281 }
1282
1283 static void destroy_receive_buffers(struct smbd_connection *info)
1284 {
1285         struct smbd_response *response;
1286
1287         while ((response = get_receive_buffer(info)))
1288                 mempool_free(response, info->response_mempool);
1289
1290         while ((response = get_empty_queue_buffer(info)))
1291                 mempool_free(response, info->response_mempool);
1292 }
1293
1294 /* Implement idle connection timer [MS-SMBD] 3.1.6.2 */
1295 static void idle_connection_timer(struct work_struct *work)
1296 {
1297         struct smbd_connection *info = container_of(
1298                                         work, struct smbd_connection,
1299                                         idle_timer_work.work);
1300
1301         if (info->keep_alive_requested != KEEP_ALIVE_NONE) {
1302                 log_keep_alive(ERR,
1303                         "error status info->keep_alive_requested=%d\n",
1304                         info->keep_alive_requested);
1305                 smbd_disconnect_rdma_connection(info);
1306                 return;
1307         }
1308
1309         log_keep_alive(INFO, "about to send an empty idle message\n");
1310         smbd_post_send_empty(info);
1311
1312         /* Setup the next idle timeout work */
1313         queue_delayed_work(info->workqueue, &info->idle_timer_work,
1314                         info->keep_alive_interval*HZ);
1315 }
1316
1317 /*
1318  * Destroy the transport and related RDMA and memory resources
1319  * Need to go through all the pending counters and make sure on one is using
1320  * the transport while it is destroyed
1321  */
1322 void smbd_destroy(struct TCP_Server_Info *server)
1323 {
1324         struct smbd_connection *info = server->smbd_conn;
1325         struct smbd_response *response;
1326         unsigned long flags;
1327
1328         if (!info) {
1329                 log_rdma_event(INFO, "rdma session already destroyed\n");
1330                 return;
1331         }
1332
1333         log_rdma_event(INFO, "destroying rdma session\n");
1334         if (info->transport_status != SMBD_DISCONNECTED) {
1335                 rdma_disconnect(server->smbd_conn->id);
1336                 log_rdma_event(INFO, "wait for transport being disconnected\n");
1337                 wait_event_interruptible(
1338                         info->disconn_wait,
1339                         info->transport_status == SMBD_DISCONNECTED);
1340         }
1341
1342         log_rdma_event(INFO, "destroying qp\n");
1343         ib_drain_qp(info->id->qp);
1344         rdma_destroy_qp(info->id);
1345
1346         log_rdma_event(INFO, "cancelling idle timer\n");
1347         cancel_delayed_work_sync(&info->idle_timer_work);
1348
1349         log_rdma_event(INFO, "wait for all send posted to IB to finish\n");
1350         wait_event(info->wait_send_pending,
1351                 atomic_read(&info->send_pending) == 0);
1352
1353         /* It's not possible for upper layer to get to reassembly */
1354         log_rdma_event(INFO, "drain the reassembly queue\n");
1355         do {
1356                 spin_lock_irqsave(&info->reassembly_queue_lock, flags);
1357                 response = _get_first_reassembly(info);
1358                 if (response) {
1359                         list_del(&response->list);
1360                         spin_unlock_irqrestore(
1361                                 &info->reassembly_queue_lock, flags);
1362                         put_receive_buffer(info, response);
1363                 } else
1364                         spin_unlock_irqrestore(
1365                                 &info->reassembly_queue_lock, flags);
1366         } while (response);
1367         info->reassembly_data_length = 0;
1368
1369         log_rdma_event(INFO, "free receive buffers\n");
1370         wait_event(info->wait_receive_queues,
1371                 info->count_receive_queue + info->count_empty_packet_queue
1372                         == info->receive_credit_max);
1373         destroy_receive_buffers(info);
1374
1375         /*
1376          * For performance reasons, memory registration and deregistration
1377          * are not locked by srv_mutex. It is possible some processes are
1378          * blocked on transport srv_mutex while holding memory registration.
1379          * Release the transport srv_mutex to allow them to hit the failure
1380          * path when sending data, and then release memory registartions.
1381          */
1382         log_rdma_event(INFO, "freeing mr list\n");
1383         wake_up_interruptible_all(&info->wait_mr);
1384         while (atomic_read(&info->mr_used_count)) {
1385                 cifs_server_unlock(server);
1386                 msleep(1000);
1387                 cifs_server_lock(server);
1388         }
1389         destroy_mr_list(info);
1390
1391         ib_free_cq(info->send_cq);
1392         ib_free_cq(info->recv_cq);
1393         ib_dealloc_pd(info->pd);
1394         rdma_destroy_id(info->id);
1395
1396         /* free mempools */
1397         mempool_destroy(info->request_mempool);
1398         kmem_cache_destroy(info->request_cache);
1399
1400         mempool_destroy(info->response_mempool);
1401         kmem_cache_destroy(info->response_cache);
1402
1403         info->transport_status = SMBD_DESTROYED;
1404
1405         destroy_workqueue(info->workqueue);
1406         log_rdma_event(INFO,  "rdma session destroyed\n");
1407         kfree(info);
1408         server->smbd_conn = NULL;
1409 }
1410
1411 /*
1412  * Reconnect this SMBD connection, called from upper layer
1413  * return value: 0 on success, or actual error code
1414  */
1415 int smbd_reconnect(struct TCP_Server_Info *server)
1416 {
1417         log_rdma_event(INFO, "reconnecting rdma session\n");
1418
1419         if (!server->smbd_conn) {
1420                 log_rdma_event(INFO, "rdma session already destroyed\n");
1421                 goto create_conn;
1422         }
1423
1424         /*
1425          * This is possible if transport is disconnected and we haven't received
1426          * notification from RDMA, but upper layer has detected timeout
1427          */
1428         if (server->smbd_conn->transport_status == SMBD_CONNECTED) {
1429                 log_rdma_event(INFO, "disconnecting transport\n");
1430                 smbd_destroy(server);
1431         }
1432
1433 create_conn:
1434         log_rdma_event(INFO, "creating rdma session\n");
1435         server->smbd_conn = smbd_get_connection(
1436                 server, (struct sockaddr *) &server->dstaddr);
1437
1438         if (server->smbd_conn)
1439                 cifs_dbg(VFS, "RDMA transport re-established\n");
1440
1441         return server->smbd_conn ? 0 : -ENOENT;
1442 }
1443
1444 static void destroy_caches_and_workqueue(struct smbd_connection *info)
1445 {
1446         destroy_receive_buffers(info);
1447         destroy_workqueue(info->workqueue);
1448         mempool_destroy(info->response_mempool);
1449         kmem_cache_destroy(info->response_cache);
1450         mempool_destroy(info->request_mempool);
1451         kmem_cache_destroy(info->request_cache);
1452 }
1453
1454 #define MAX_NAME_LEN    80
1455 static int allocate_caches_and_workqueue(struct smbd_connection *info)
1456 {
1457         char name[MAX_NAME_LEN];
1458         int rc;
1459
1460         scnprintf(name, MAX_NAME_LEN, "smbd_request_%p", info);
1461         info->request_cache =
1462                 kmem_cache_create(
1463                         name,
1464                         sizeof(struct smbd_request) +
1465                                 sizeof(struct smbd_data_transfer),
1466                         0, SLAB_HWCACHE_ALIGN, NULL);
1467         if (!info->request_cache)
1468                 return -ENOMEM;
1469
1470         info->request_mempool =
1471                 mempool_create(info->send_credit_target, mempool_alloc_slab,
1472                         mempool_free_slab, info->request_cache);
1473         if (!info->request_mempool)
1474                 goto out1;
1475
1476         scnprintf(name, MAX_NAME_LEN, "smbd_response_%p", info);
1477         info->response_cache =
1478                 kmem_cache_create(
1479                         name,
1480                         sizeof(struct smbd_response) +
1481                                 info->max_receive_size,
1482                         0, SLAB_HWCACHE_ALIGN, NULL);
1483         if (!info->response_cache)
1484                 goto out2;
1485
1486         info->response_mempool =
1487                 mempool_create(info->receive_credit_max, mempool_alloc_slab,
1488                        mempool_free_slab, info->response_cache);
1489         if (!info->response_mempool)
1490                 goto out3;
1491
1492         scnprintf(name, MAX_NAME_LEN, "smbd_%p", info);
1493         info->workqueue = create_workqueue(name);
1494         if (!info->workqueue)
1495                 goto out4;
1496
1497         rc = allocate_receive_buffers(info, info->receive_credit_max);
1498         if (rc) {
1499                 log_rdma_event(ERR, "failed to allocate receive buffers\n");
1500                 goto out5;
1501         }
1502
1503         return 0;
1504
1505 out5:
1506         destroy_workqueue(info->workqueue);
1507 out4:
1508         mempool_destroy(info->response_mempool);
1509 out3:
1510         kmem_cache_destroy(info->response_cache);
1511 out2:
1512         mempool_destroy(info->request_mempool);
1513 out1:
1514         kmem_cache_destroy(info->request_cache);
1515         return -ENOMEM;
1516 }
1517
1518 /* Create a SMBD connection, called by upper layer */
1519 static struct smbd_connection *_smbd_get_connection(
1520         struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port)
1521 {
1522         int rc;
1523         struct smbd_connection *info;
1524         struct rdma_conn_param conn_param;
1525         struct ib_qp_init_attr qp_attr;
1526         struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
1527         struct ib_port_immutable port_immutable;
1528         u32 ird_ord_hdr[2];
1529
1530         info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
1531         if (!info)
1532                 return NULL;
1533
1534         info->transport_status = SMBD_CONNECTING;
1535         rc = smbd_ia_open(info, dstaddr, port);
1536         if (rc) {
1537                 log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc);
1538                 goto create_id_failed;
1539         }
1540
1541         if (smbd_send_credit_target > info->id->device->attrs.max_cqe ||
1542             smbd_send_credit_target > info->id->device->attrs.max_qp_wr) {
1543                 log_rdma_event(ERR, "consider lowering send_credit_target = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n",
1544                                smbd_send_credit_target,
1545                                info->id->device->attrs.max_cqe,
1546                                info->id->device->attrs.max_qp_wr);
1547                 goto config_failed;
1548         }
1549
1550         if (smbd_receive_credit_max > info->id->device->attrs.max_cqe ||
1551             smbd_receive_credit_max > info->id->device->attrs.max_qp_wr) {
1552                 log_rdma_event(ERR, "consider lowering receive_credit_max = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n",
1553                                smbd_receive_credit_max,
1554                                info->id->device->attrs.max_cqe,
1555                                info->id->device->attrs.max_qp_wr);
1556                 goto config_failed;
1557         }
1558
1559         info->receive_credit_max = smbd_receive_credit_max;
1560         info->send_credit_target = smbd_send_credit_target;
1561         info->max_send_size = smbd_max_send_size;
1562         info->max_fragmented_recv_size = smbd_max_fragmented_recv_size;
1563         info->max_receive_size = smbd_max_receive_size;
1564         info->keep_alive_interval = smbd_keep_alive_interval;
1565
1566         if (info->id->device->attrs.max_send_sge < SMBDIRECT_MAX_SEND_SGE ||
1567             info->id->device->attrs.max_recv_sge < SMBDIRECT_MAX_RECV_SGE) {
1568                 log_rdma_event(ERR,
1569                         "device %.*s max_send_sge/max_recv_sge = %d/%d too small\n",
1570                         IB_DEVICE_NAME_MAX,
1571                         info->id->device->name,
1572                         info->id->device->attrs.max_send_sge,
1573                         info->id->device->attrs.max_recv_sge);
1574                 goto config_failed;
1575         }
1576
1577         info->send_cq = NULL;
1578         info->recv_cq = NULL;
1579         info->send_cq =
1580                 ib_alloc_cq_any(info->id->device, info,
1581                                 info->send_credit_target, IB_POLL_SOFTIRQ);
1582         if (IS_ERR(info->send_cq)) {
1583                 info->send_cq = NULL;
1584                 goto alloc_cq_failed;
1585         }
1586
1587         info->recv_cq =
1588                 ib_alloc_cq_any(info->id->device, info,
1589                                 info->receive_credit_max, IB_POLL_SOFTIRQ);
1590         if (IS_ERR(info->recv_cq)) {
1591                 info->recv_cq = NULL;
1592                 goto alloc_cq_failed;
1593         }
1594
1595         memset(&qp_attr, 0, sizeof(qp_attr));
1596         qp_attr.event_handler = smbd_qp_async_error_upcall;
1597         qp_attr.qp_context = info;
1598         qp_attr.cap.max_send_wr = info->send_credit_target;
1599         qp_attr.cap.max_recv_wr = info->receive_credit_max;
1600         qp_attr.cap.max_send_sge = SMBDIRECT_MAX_SEND_SGE;
1601         qp_attr.cap.max_recv_sge = SMBDIRECT_MAX_RECV_SGE;
1602         qp_attr.cap.max_inline_data = 0;
1603         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1604         qp_attr.qp_type = IB_QPT_RC;
1605         qp_attr.send_cq = info->send_cq;
1606         qp_attr.recv_cq = info->recv_cq;
1607         qp_attr.port_num = ~0;
1608
1609         rc = rdma_create_qp(info->id, info->pd, &qp_attr);
1610         if (rc) {
1611                 log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc);
1612                 goto create_qp_failed;
1613         }
1614
1615         memset(&conn_param, 0, sizeof(conn_param));
1616         conn_param.initiator_depth = 0;
1617
1618         conn_param.responder_resources =
1619                 info->id->device->attrs.max_qp_rd_atom
1620                         < SMBD_CM_RESPONDER_RESOURCES ?
1621                 info->id->device->attrs.max_qp_rd_atom :
1622                 SMBD_CM_RESPONDER_RESOURCES;
1623         info->responder_resources = conn_param.responder_resources;
1624         log_rdma_mr(INFO, "responder_resources=%d\n",
1625                 info->responder_resources);
1626
1627         /* Need to send IRD/ORD in private data for iWARP */
1628         info->id->device->ops.get_port_immutable(
1629                 info->id->device, info->id->port_num, &port_immutable);
1630         if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
1631                 ird_ord_hdr[0] = info->responder_resources;
1632                 ird_ord_hdr[1] = 1;
1633                 conn_param.private_data = ird_ord_hdr;
1634                 conn_param.private_data_len = sizeof(ird_ord_hdr);
1635         } else {
1636                 conn_param.private_data = NULL;
1637                 conn_param.private_data_len = 0;
1638         }
1639
1640         conn_param.retry_count = SMBD_CM_RETRY;
1641         conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY;
1642         conn_param.flow_control = 0;
1643
1644         log_rdma_event(INFO, "connecting to IP %pI4 port %d\n",
1645                 &addr_in->sin_addr, port);
1646
1647         init_waitqueue_head(&info->conn_wait);
1648         init_waitqueue_head(&info->disconn_wait);
1649         init_waitqueue_head(&info->wait_reassembly_queue);
1650         rc = rdma_connect(info->id, &conn_param);
1651         if (rc) {
1652                 log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc);
1653                 goto rdma_connect_failed;
1654         }
1655
1656         wait_event_interruptible(
1657                 info->conn_wait, info->transport_status != SMBD_CONNECTING);
1658
1659         if (info->transport_status != SMBD_CONNECTED) {
1660                 log_rdma_event(ERR, "rdma_connect failed port=%d\n", port);
1661                 goto rdma_connect_failed;
1662         }
1663
1664         log_rdma_event(INFO, "rdma_connect connected\n");
1665
1666         rc = allocate_caches_and_workqueue(info);
1667         if (rc) {
1668                 log_rdma_event(ERR, "cache allocation failed\n");
1669                 goto allocate_cache_failed;
1670         }
1671
1672         init_waitqueue_head(&info->wait_send_queue);
1673         INIT_DELAYED_WORK(&info->idle_timer_work, idle_connection_timer);
1674         queue_delayed_work(info->workqueue, &info->idle_timer_work,
1675                 info->keep_alive_interval*HZ);
1676
1677         init_waitqueue_head(&info->wait_send_pending);
1678         atomic_set(&info->send_pending, 0);
1679
1680         init_waitqueue_head(&info->wait_post_send);
1681
1682         INIT_WORK(&info->disconnect_work, smbd_disconnect_rdma_work);
1683         INIT_WORK(&info->post_send_credits_work, smbd_post_send_credits);
1684         info->new_credits_offered = 0;
1685         spin_lock_init(&info->lock_new_credits_offered);
1686
1687         rc = smbd_negotiate(info);
1688         if (rc) {
1689                 log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc);
1690                 goto negotiation_failed;
1691         }
1692
1693         rc = allocate_mr_list(info);
1694         if (rc) {
1695                 log_rdma_mr(ERR, "memory registration allocation failed\n");
1696                 goto allocate_mr_failed;
1697         }
1698
1699         return info;
1700
1701 allocate_mr_failed:
1702         /* At this point, need to a full transport shutdown */
1703         server->smbd_conn = info;
1704         smbd_destroy(server);
1705         return NULL;
1706
1707 negotiation_failed:
1708         cancel_delayed_work_sync(&info->idle_timer_work);
1709         destroy_caches_and_workqueue(info);
1710         info->transport_status = SMBD_NEGOTIATE_FAILED;
1711         init_waitqueue_head(&info->conn_wait);
1712         rdma_disconnect(info->id);
1713         wait_event(info->conn_wait,
1714                 info->transport_status == SMBD_DISCONNECTED);
1715
1716 allocate_cache_failed:
1717 rdma_connect_failed:
1718         rdma_destroy_qp(info->id);
1719
1720 create_qp_failed:
1721 alloc_cq_failed:
1722         if (info->send_cq)
1723                 ib_free_cq(info->send_cq);
1724         if (info->recv_cq)
1725                 ib_free_cq(info->recv_cq);
1726
1727 config_failed:
1728         ib_dealloc_pd(info->pd);
1729         rdma_destroy_id(info->id);
1730
1731 create_id_failed:
1732         kfree(info);
1733         return NULL;
1734 }
1735
1736 struct smbd_connection *smbd_get_connection(
1737         struct TCP_Server_Info *server, struct sockaddr *dstaddr)
1738 {
1739         struct smbd_connection *ret;
1740         int port = SMBD_PORT;
1741
1742 try_again:
1743         ret = _smbd_get_connection(server, dstaddr, port);
1744
1745         /* Try SMB_PORT if SMBD_PORT doesn't work */
1746         if (!ret && port == SMBD_PORT) {
1747                 port = SMB_PORT;
1748                 goto try_again;
1749         }
1750         return ret;
1751 }
1752
1753 /*
1754  * Receive data from receive reassembly queue
1755  * All the incoming data packets are placed in reassembly queue
1756  * buf: the buffer to read data into
1757  * size: the length of data to read
1758  * return value: actual data read
1759  * Note: this implementation copies the data from reassebmly queue to receive
1760  * buffers used by upper layer. This is not the optimal code path. A better way
1761  * to do it is to not have upper layer allocate its receive buffers but rather
1762  * borrow the buffer from reassembly queue, and return it after data is
1763  * consumed. But this will require more changes to upper layer code, and also
1764  * need to consider packet boundaries while they still being reassembled.
1765  */
1766 static int smbd_recv_buf(struct smbd_connection *info, char *buf,
1767                 unsigned int size)
1768 {
1769         struct smbd_response *response;
1770         struct smbd_data_transfer *data_transfer;
1771         int to_copy, to_read, data_read, offset;
1772         u32 data_length, remaining_data_length, data_offset;
1773         int rc;
1774
1775 again:
1776         /*
1777          * No need to hold the reassembly queue lock all the time as we are
1778          * the only one reading from the front of the queue. The transport
1779          * may add more entries to the back of the queue at the same time
1780          */
1781         log_read(INFO, "size=%d info->reassembly_data_length=%d\n", size,
1782                 info->reassembly_data_length);
1783         if (info->reassembly_data_length >= size) {
1784                 int queue_length;
1785                 int queue_removed = 0;
1786
1787                 /*
1788                  * Need to make sure reassembly_data_length is read before
1789                  * reading reassembly_queue_length and calling
1790                  * _get_first_reassembly. This call is lock free
1791                  * as we never read at the end of the queue which are being
1792                  * updated in SOFTIRQ as more data is received
1793                  */
1794                 virt_rmb();
1795                 queue_length = info->reassembly_queue_length;
1796                 data_read = 0;
1797                 to_read = size;
1798                 offset = info->first_entry_offset;
1799                 while (data_read < size) {
1800                         response = _get_first_reassembly(info);
1801                         data_transfer = smbd_response_payload(response);
1802                         data_length = le32_to_cpu(data_transfer->data_length);
1803                         remaining_data_length =
1804                                 le32_to_cpu(
1805                                         data_transfer->remaining_data_length);
1806                         data_offset = le32_to_cpu(data_transfer->data_offset);
1807
1808                         /*
1809                          * The upper layer expects RFC1002 length at the
1810                          * beginning of the payload. Return it to indicate
1811                          * the total length of the packet. This minimize the
1812                          * change to upper layer packet processing logic. This
1813                          * will be eventually remove when an intermediate
1814                          * transport layer is added
1815                          */
1816                         if (response->first_segment && size == 4) {
1817                                 unsigned int rfc1002_len =
1818                                         data_length + remaining_data_length;
1819                                 *((__be32 *)buf) = cpu_to_be32(rfc1002_len);
1820                                 data_read = 4;
1821                                 response->first_segment = false;
1822                                 log_read(INFO, "returning rfc1002 length %d\n",
1823                                         rfc1002_len);
1824                                 goto read_rfc1002_done;
1825                         }
1826
1827                         to_copy = min_t(int, data_length - offset, to_read);
1828                         memcpy(
1829                                 buf + data_read,
1830                                 (char *)data_transfer + data_offset + offset,
1831                                 to_copy);
1832
1833                         /* move on to the next buffer? */
1834                         if (to_copy == data_length - offset) {
1835                                 queue_length--;
1836                                 /*
1837                                  * No need to lock if we are not at the
1838                                  * end of the queue
1839                                  */
1840                                 if (queue_length)
1841                                         list_del(&response->list);
1842                                 else {
1843                                         spin_lock_irq(
1844                                                 &info->reassembly_queue_lock);
1845                                         list_del(&response->list);
1846                                         spin_unlock_irq(
1847                                                 &info->reassembly_queue_lock);
1848                                 }
1849                                 queue_removed++;
1850                                 info->count_reassembly_queue--;
1851                                 info->count_dequeue_reassembly_queue++;
1852                                 put_receive_buffer(info, response);
1853                                 offset = 0;
1854                                 log_read(INFO, "put_receive_buffer offset=0\n");
1855                         } else
1856                                 offset += to_copy;
1857
1858                         to_read -= to_copy;
1859                         data_read += to_copy;
1860
1861                         log_read(INFO, "_get_first_reassembly memcpy %d bytes data_transfer_length-offset=%d after that to_read=%d data_read=%d offset=%d\n",
1862                                  to_copy, data_length - offset,
1863                                  to_read, data_read, offset);
1864                 }
1865
1866                 spin_lock_irq(&info->reassembly_queue_lock);
1867                 info->reassembly_data_length -= data_read;
1868                 info->reassembly_queue_length -= queue_removed;
1869                 spin_unlock_irq(&info->reassembly_queue_lock);
1870
1871                 info->first_entry_offset = offset;
1872                 log_read(INFO, "returning to thread data_read=%d reassembly_data_length=%d first_entry_offset=%d\n",
1873                          data_read, info->reassembly_data_length,
1874                          info->first_entry_offset);
1875 read_rfc1002_done:
1876                 return data_read;
1877         }
1878
1879         log_read(INFO, "wait_event on more data\n");
1880         rc = wait_event_interruptible(
1881                 info->wait_reassembly_queue,
1882                 info->reassembly_data_length >= size ||
1883                         info->transport_status != SMBD_CONNECTED);
1884         /* Don't return any data if interrupted */
1885         if (rc)
1886                 return rc;
1887
1888         if (info->transport_status != SMBD_CONNECTED) {
1889                 log_read(ERR, "disconnected\n");
1890                 return -ECONNABORTED;
1891         }
1892
1893         goto again;
1894 }
1895
1896 /*
1897  * Receive a page from receive reassembly queue
1898  * page: the page to read data into
1899  * to_read: the length of data to read
1900  * return value: actual data read
1901  */
1902 static int smbd_recv_page(struct smbd_connection *info,
1903                 struct page *page, unsigned int page_offset,
1904                 unsigned int to_read)
1905 {
1906         int ret;
1907         char *to_address;
1908         void *page_address;
1909
1910         /* make sure we have the page ready for read */
1911         ret = wait_event_interruptible(
1912                 info->wait_reassembly_queue,
1913                 info->reassembly_data_length >= to_read ||
1914                         info->transport_status != SMBD_CONNECTED);
1915         if (ret)
1916                 return ret;
1917
1918         /* now we can read from reassembly queue and not sleep */
1919         page_address = kmap_atomic(page);
1920         to_address = (char *) page_address + page_offset;
1921
1922         log_read(INFO, "reading from page=%p address=%p to_read=%d\n",
1923                 page, to_address, to_read);
1924
1925         ret = smbd_recv_buf(info, to_address, to_read);
1926         kunmap_atomic(page_address);
1927
1928         return ret;
1929 }
1930
1931 /*
1932  * Receive data from transport
1933  * msg: a msghdr point to the buffer, can be ITER_KVEC or ITER_BVEC
1934  * return: total bytes read, or 0. SMB Direct will not do partial read.
1935  */
1936 int smbd_recv(struct smbd_connection *info, struct msghdr *msg)
1937 {
1938         char *buf;
1939         struct page *page;
1940         unsigned int to_read, page_offset;
1941         int rc;
1942
1943         if (iov_iter_rw(&msg->msg_iter) == WRITE) {
1944                 /* It's a bug in upper layer to get there */
1945                 cifs_dbg(VFS, "Invalid msg iter dir %u\n",
1946                          iov_iter_rw(&msg->msg_iter));
1947                 rc = -EINVAL;
1948                 goto out;
1949         }
1950
1951         switch (iov_iter_type(&msg->msg_iter)) {
1952         case ITER_KVEC:
1953                 buf = msg->msg_iter.kvec->iov_base;
1954                 to_read = msg->msg_iter.kvec->iov_len;
1955                 rc = smbd_recv_buf(info, buf, to_read);
1956                 break;
1957
1958         case ITER_BVEC:
1959                 page = msg->msg_iter.bvec->bv_page;
1960                 page_offset = msg->msg_iter.bvec->bv_offset;
1961                 to_read = msg->msg_iter.bvec->bv_len;
1962                 rc = smbd_recv_page(info, page, page_offset, to_read);
1963                 break;
1964
1965         default:
1966                 /* It's a bug in upper layer to get there */
1967                 cifs_dbg(VFS, "Invalid msg type %d\n",
1968                          iov_iter_type(&msg->msg_iter));
1969                 rc = -EINVAL;
1970         }
1971
1972 out:
1973         /* SMBDirect will read it all or nothing */
1974         if (rc > 0)
1975                 msg->msg_iter.count = 0;
1976         return rc;
1977 }
1978
1979 /*
1980  * Send data to transport
1981  * Each rqst is transported as a SMBDirect payload
1982  * rqst: the data to write
1983  * return value: 0 if successfully write, otherwise error code
1984  */
1985 int smbd_send(struct TCP_Server_Info *server,
1986         int num_rqst, struct smb_rqst *rqst_array)
1987 {
1988         struct smbd_connection *info = server->smbd_conn;
1989         struct kvec vecs[SMBDIRECT_MAX_SEND_SGE - 1];
1990         int nvecs;
1991         int size;
1992         unsigned int buflen, remaining_data_length;
1993         unsigned int offset, remaining_vec_data_length;
1994         int start, i, j;
1995         int max_iov_size =
1996                 info->max_send_size - sizeof(struct smbd_data_transfer);
1997         struct kvec *iov;
1998         int rc;
1999         struct smb_rqst *rqst;
2000         int rqst_idx;
2001
2002         if (info->transport_status != SMBD_CONNECTED)
2003                 return -EAGAIN;
2004
2005         /*
2006          * Add in the page array if there is one. The caller needs to set
2007          * rq_tailsz to PAGE_SIZE when the buffer has multiple pages and
2008          * ends at page boundary
2009          */
2010         remaining_data_length = 0;
2011         for (i = 0; i < num_rqst; i++)
2012                 remaining_data_length += smb_rqst_len(server, &rqst_array[i]);
2013
2014         if (unlikely(remaining_data_length > info->max_fragmented_send_size)) {
2015                 /* assertion: payload never exceeds negotiated maximum */
2016                 log_write(ERR, "payload size %d > max size %d\n",
2017                         remaining_data_length, info->max_fragmented_send_size);
2018                 return -EINVAL;
2019         }
2020
2021         log_write(INFO, "num_rqst=%d total length=%u\n",
2022                         num_rqst, remaining_data_length);
2023
2024         rqst_idx = 0;
2025         do {
2026                 rqst = &rqst_array[rqst_idx];
2027                 iov = rqst->rq_iov;
2028
2029                 cifs_dbg(FYI, "Sending smb (RDMA): idx=%d smb_len=%lu\n",
2030                         rqst_idx, smb_rqst_len(server, rqst));
2031                 remaining_vec_data_length = 0;
2032                 for (i = 0; i < rqst->rq_nvec; i++) {
2033                         remaining_vec_data_length += iov[i].iov_len;
2034                         dump_smb(iov[i].iov_base, iov[i].iov_len);
2035                 }
2036
2037                 log_write(INFO, "rqst_idx=%d nvec=%d rqst->rq_npages=%d rq_pagesz=%d rq_tailsz=%d buflen=%lu\n",
2038                           rqst_idx, rqst->rq_nvec,
2039                           rqst->rq_npages, rqst->rq_pagesz,
2040                           rqst->rq_tailsz, smb_rqst_len(server, rqst));
2041
2042                 start = 0;
2043                 offset = 0;
2044                 do {
2045                         buflen = 0;
2046                         i = start;
2047                         j = 0;
2048                         while (i < rqst->rq_nvec &&
2049                                 j < SMBDIRECT_MAX_SEND_SGE - 1 &&
2050                                 buflen < max_iov_size) {
2051
2052                                 vecs[j].iov_base = iov[i].iov_base + offset;
2053                                 if (buflen + iov[i].iov_len > max_iov_size) {
2054                                         vecs[j].iov_len =
2055                                                 max_iov_size - iov[i].iov_len;
2056                                         buflen = max_iov_size;
2057                                         offset = vecs[j].iov_len;
2058                                 } else {
2059                                         vecs[j].iov_len =
2060                                                 iov[i].iov_len - offset;
2061                                         buflen += vecs[j].iov_len;
2062                                         offset = 0;
2063                                         ++i;
2064                                 }
2065                                 ++j;
2066                         }
2067
2068                         remaining_vec_data_length -= buflen;
2069                         remaining_data_length -= buflen;
2070                         log_write(INFO, "sending %s iov[%d] from start=%d nvecs=%d remaining_data_length=%d\n",
2071                                         remaining_vec_data_length > 0 ?
2072                                                 "partial" : "complete",
2073                                         rqst->rq_nvec, start, j,
2074                                         remaining_data_length);
2075
2076                         start = i;
2077                         rc = smbd_post_send_data(info, vecs, j, remaining_data_length);
2078                         if (rc)
2079                                 goto done;
2080                 } while (remaining_vec_data_length > 0);
2081
2082                 /* now sending pages if there are any */
2083                 for (i = 0; i < rqst->rq_npages; i++) {
2084                         rqst_page_get_length(rqst, i, &buflen, &offset);
2085                         nvecs = (buflen + max_iov_size - 1) / max_iov_size;
2086                         log_write(INFO, "sending pages buflen=%d nvecs=%d\n",
2087                                 buflen, nvecs);
2088                         for (j = 0; j < nvecs; j++) {
2089                                 size = min_t(unsigned int, max_iov_size, remaining_data_length);
2090                                 remaining_data_length -= size;
2091                                 log_write(INFO, "sending pages i=%d offset=%d size=%d remaining_data_length=%d\n",
2092                                           i, j * max_iov_size + offset, size,
2093                                           remaining_data_length);
2094                                 rc = smbd_post_send_page(
2095                                         info, rqst->rq_pages[i],
2096                                         j*max_iov_size + offset,
2097                                         size, remaining_data_length);
2098                                 if (rc)
2099                                         goto done;
2100                         }
2101                 }
2102         } while (++rqst_idx < num_rqst);
2103
2104 done:
2105         /*
2106          * As an optimization, we don't wait for individual I/O to finish
2107          * before sending the next one.
2108          * Send them all and wait for pending send count to get to 0
2109          * that means all the I/Os have been out and we are good to return
2110          */
2111
2112         wait_event(info->wait_send_pending,
2113                 atomic_read(&info->send_pending) == 0);
2114
2115         return rc;
2116 }
2117
2118 static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc)
2119 {
2120         struct smbd_mr *mr;
2121         struct ib_cqe *cqe;
2122
2123         if (wc->status) {
2124                 log_rdma_mr(ERR, "status=%d\n", wc->status);
2125                 cqe = wc->wr_cqe;
2126                 mr = container_of(cqe, struct smbd_mr, cqe);
2127                 smbd_disconnect_rdma_connection(mr->conn);
2128         }
2129 }
2130
2131 /*
2132  * The work queue function that recovers MRs
2133  * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used
2134  * again. Both calls are slow, so finish them in a workqueue. This will not
2135  * block I/O path.
2136  * There is one workqueue that recovers MRs, there is no need to lock as the
2137  * I/O requests calling smbd_register_mr will never update the links in the
2138  * mr_list.
2139  */
2140 static void smbd_mr_recovery_work(struct work_struct *work)
2141 {
2142         struct smbd_connection *info =
2143                 container_of(work, struct smbd_connection, mr_recovery_work);
2144         struct smbd_mr *smbdirect_mr;
2145         int rc;
2146
2147         list_for_each_entry(smbdirect_mr, &info->mr_list, list) {
2148                 if (smbdirect_mr->state == MR_ERROR) {
2149
2150                         /* recover this MR entry */
2151                         rc = ib_dereg_mr(smbdirect_mr->mr);
2152                         if (rc) {
2153                                 log_rdma_mr(ERR,
2154                                         "ib_dereg_mr failed rc=%x\n",
2155                                         rc);
2156                                 smbd_disconnect_rdma_connection(info);
2157                                 continue;
2158                         }
2159
2160                         smbdirect_mr->mr = ib_alloc_mr(
2161                                 info->pd, info->mr_type,
2162                                 info->max_frmr_depth);
2163                         if (IS_ERR(smbdirect_mr->mr)) {
2164                                 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n",
2165                                             info->mr_type,
2166                                             info->max_frmr_depth);
2167                                 smbd_disconnect_rdma_connection(info);
2168                                 continue;
2169                         }
2170                 } else
2171                         /* This MR is being used, don't recover it */
2172                         continue;
2173
2174                 smbdirect_mr->state = MR_READY;
2175
2176                 /* smbdirect_mr->state is updated by this function
2177                  * and is read and updated by I/O issuing CPUs trying
2178                  * to get a MR, the call to atomic_inc_return
2179                  * implicates a memory barrier and guarantees this
2180                  * value is updated before waking up any calls to
2181                  * get_mr() from the I/O issuing CPUs
2182                  */
2183                 if (atomic_inc_return(&info->mr_ready_count) == 1)
2184                         wake_up_interruptible(&info->wait_mr);
2185         }
2186 }
2187
2188 static void destroy_mr_list(struct smbd_connection *info)
2189 {
2190         struct smbd_mr *mr, *tmp;
2191
2192         cancel_work_sync(&info->mr_recovery_work);
2193         list_for_each_entry_safe(mr, tmp, &info->mr_list, list) {
2194                 if (mr->state == MR_INVALIDATED)
2195                         ib_dma_unmap_sg(info->id->device, mr->sgl,
2196                                 mr->sgl_count, mr->dir);
2197                 ib_dereg_mr(mr->mr);
2198                 kfree(mr->sgl);
2199                 kfree(mr);
2200         }
2201 }
2202
2203 /*
2204  * Allocate MRs used for RDMA read/write
2205  * The number of MRs will not exceed hardware capability in responder_resources
2206  * All MRs are kept in mr_list. The MR can be recovered after it's used
2207  * Recovery is done in smbd_mr_recovery_work. The content of list entry changes
2208  * as MRs are used and recovered for I/O, but the list links will not change
2209  */
2210 static int allocate_mr_list(struct smbd_connection *info)
2211 {
2212         int i;
2213         struct smbd_mr *smbdirect_mr, *tmp;
2214
2215         INIT_LIST_HEAD(&info->mr_list);
2216         init_waitqueue_head(&info->wait_mr);
2217         spin_lock_init(&info->mr_list_lock);
2218         atomic_set(&info->mr_ready_count, 0);
2219         atomic_set(&info->mr_used_count, 0);
2220         init_waitqueue_head(&info->wait_for_mr_cleanup);
2221         INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
2222         /* Allocate more MRs (2x) than hardware responder_resources */
2223         for (i = 0; i < info->responder_resources * 2; i++) {
2224                 smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
2225                 if (!smbdirect_mr)
2226                         goto out;
2227                 smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type,
2228                                         info->max_frmr_depth);
2229                 if (IS_ERR(smbdirect_mr->mr)) {
2230                         log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n",
2231                                     info->mr_type, info->max_frmr_depth);
2232                         goto out;
2233                 }
2234                 smbdirect_mr->sgl = kcalloc(
2235                                         info->max_frmr_depth,
2236                                         sizeof(struct scatterlist),
2237                                         GFP_KERNEL);
2238                 if (!smbdirect_mr->sgl) {
2239                         log_rdma_mr(ERR, "failed to allocate sgl\n");
2240                         ib_dereg_mr(smbdirect_mr->mr);
2241                         goto out;
2242                 }
2243                 smbdirect_mr->state = MR_READY;
2244                 smbdirect_mr->conn = info;
2245
2246                 list_add_tail(&smbdirect_mr->list, &info->mr_list);
2247                 atomic_inc(&info->mr_ready_count);
2248         }
2249         return 0;
2250
2251 out:
2252         kfree(smbdirect_mr);
2253
2254         list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) {
2255                 list_del(&smbdirect_mr->list);
2256                 ib_dereg_mr(smbdirect_mr->mr);
2257                 kfree(smbdirect_mr->sgl);
2258                 kfree(smbdirect_mr);
2259         }
2260         return -ENOMEM;
2261 }
2262
2263 /*
2264  * Get a MR from mr_list. This function waits until there is at least one
2265  * MR available in the list. It may access the list while the
2266  * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock
2267  * as they never modify the same places. However, there may be several CPUs
2268  * issueing I/O trying to get MR at the same time, mr_list_lock is used to
2269  * protect this situation.
2270  */
2271 static struct smbd_mr *get_mr(struct smbd_connection *info)
2272 {
2273         struct smbd_mr *ret;
2274         int rc;
2275 again:
2276         rc = wait_event_interruptible(info->wait_mr,
2277                 atomic_read(&info->mr_ready_count) ||
2278                 info->transport_status != SMBD_CONNECTED);
2279         if (rc) {
2280                 log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc);
2281                 return NULL;
2282         }
2283
2284         if (info->transport_status != SMBD_CONNECTED) {
2285                 log_rdma_mr(ERR, "info->transport_status=%x\n",
2286                         info->transport_status);
2287                 return NULL;
2288         }
2289
2290         spin_lock(&info->mr_list_lock);
2291         list_for_each_entry(ret, &info->mr_list, list) {
2292                 if (ret->state == MR_READY) {
2293                         ret->state = MR_REGISTERED;
2294                         spin_unlock(&info->mr_list_lock);
2295                         atomic_dec(&info->mr_ready_count);
2296                         atomic_inc(&info->mr_used_count);
2297                         return ret;
2298                 }
2299         }
2300
2301         spin_unlock(&info->mr_list_lock);
2302         /*
2303          * It is possible that we could fail to get MR because other processes may
2304          * try to acquire a MR at the same time. If this is the case, retry it.
2305          */
2306         goto again;
2307 }
2308
2309 /*
2310  * Register memory for RDMA read/write
2311  * pages[]: the list of pages to register memory with
2312  * num_pages: the number of pages to register
2313  * tailsz: if non-zero, the bytes to register in the last page
2314  * writing: true if this is a RDMA write (SMB read), false for RDMA read
2315  * need_invalidate: true if this MR needs to be locally invalidated after I/O
2316  * return value: the MR registered, NULL if failed.
2317  */
2318 struct smbd_mr *smbd_register_mr(
2319         struct smbd_connection *info, struct page *pages[], int num_pages,
2320         int offset, int tailsz, bool writing, bool need_invalidate)
2321 {
2322         struct smbd_mr *smbdirect_mr;
2323         int rc, i;
2324         enum dma_data_direction dir;
2325         struct ib_reg_wr *reg_wr;
2326
2327         if (num_pages > info->max_frmr_depth) {
2328                 log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n",
2329                         num_pages, info->max_frmr_depth);
2330                 return NULL;
2331         }
2332
2333         smbdirect_mr = get_mr(info);
2334         if (!smbdirect_mr) {
2335                 log_rdma_mr(ERR, "get_mr returning NULL\n");
2336                 return NULL;
2337         }
2338         smbdirect_mr->need_invalidate = need_invalidate;
2339         smbdirect_mr->sgl_count = num_pages;
2340         sg_init_table(smbdirect_mr->sgl, num_pages);
2341
2342         log_rdma_mr(INFO, "num_pages=0x%x offset=0x%x tailsz=0x%x\n",
2343                         num_pages, offset, tailsz);
2344
2345         if (num_pages == 1) {
2346                 sg_set_page(&smbdirect_mr->sgl[0], pages[0], tailsz, offset);
2347                 goto skip_multiple_pages;
2348         }
2349
2350         /* We have at least two pages to register */
2351         sg_set_page(
2352                 &smbdirect_mr->sgl[0], pages[0], PAGE_SIZE - offset, offset);
2353         i = 1;
2354         while (i < num_pages - 1) {
2355                 sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0);
2356                 i++;
2357         }
2358         sg_set_page(&smbdirect_mr->sgl[i], pages[i],
2359                 tailsz ? tailsz : PAGE_SIZE, 0);
2360
2361 skip_multiple_pages:
2362         dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2363         smbdirect_mr->dir = dir;
2364         rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir);
2365         if (!rc) {
2366                 log_rdma_mr(ERR, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n",
2367                         num_pages, dir, rc);
2368                 goto dma_map_error;
2369         }
2370
2371         rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages,
2372                 NULL, PAGE_SIZE);
2373         if (rc != num_pages) {
2374                 log_rdma_mr(ERR,
2375                         "ib_map_mr_sg failed rc = %d num_pages = %x\n",
2376                         rc, num_pages);
2377                 goto map_mr_error;
2378         }
2379
2380         ib_update_fast_reg_key(smbdirect_mr->mr,
2381                 ib_inc_rkey(smbdirect_mr->mr->rkey));
2382         reg_wr = &smbdirect_mr->wr;
2383         reg_wr->wr.opcode = IB_WR_REG_MR;
2384         smbdirect_mr->cqe.done = register_mr_done;
2385         reg_wr->wr.wr_cqe = &smbdirect_mr->cqe;
2386         reg_wr->wr.num_sge = 0;
2387         reg_wr->wr.send_flags = IB_SEND_SIGNALED;
2388         reg_wr->mr = smbdirect_mr->mr;
2389         reg_wr->key = smbdirect_mr->mr->rkey;
2390         reg_wr->access = writing ?
2391                         IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2392                         IB_ACCESS_REMOTE_READ;
2393
2394         /*
2395          * There is no need for waiting for complemtion on ib_post_send
2396          * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution
2397          * on the next ib_post_send when we actaully send I/O to remote peer
2398          */
2399         rc = ib_post_send(info->id->qp, &reg_wr->wr, NULL);
2400         if (!rc)
2401                 return smbdirect_mr;
2402
2403         log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n",
2404                 rc, reg_wr->key);
2405
2406         /* If all failed, attempt to recover this MR by setting it MR_ERROR*/
2407 map_mr_error:
2408         ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl,
2409                 smbdirect_mr->sgl_count, smbdirect_mr->dir);
2410
2411 dma_map_error:
2412         smbdirect_mr->state = MR_ERROR;
2413         if (atomic_dec_and_test(&info->mr_used_count))
2414                 wake_up(&info->wait_for_mr_cleanup);
2415
2416         smbd_disconnect_rdma_connection(info);
2417
2418         return NULL;
2419 }
2420
2421 static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc)
2422 {
2423         struct smbd_mr *smbdirect_mr;
2424         struct ib_cqe *cqe;
2425
2426         cqe = wc->wr_cqe;
2427         smbdirect_mr = container_of(cqe, struct smbd_mr, cqe);
2428         smbdirect_mr->state = MR_INVALIDATED;
2429         if (wc->status != IB_WC_SUCCESS) {
2430                 log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status);
2431                 smbdirect_mr->state = MR_ERROR;
2432         }
2433         complete(&smbdirect_mr->invalidate_done);
2434 }
2435
2436 /*
2437  * Deregister a MR after I/O is done
2438  * This function may wait if remote invalidation is not used
2439  * and we have to locally invalidate the buffer to prevent data is being
2440  * modified by remote peer after upper layer consumes it
2441  */
2442 int smbd_deregister_mr(struct smbd_mr *smbdirect_mr)
2443 {
2444         struct ib_send_wr *wr;
2445         struct smbd_connection *info = smbdirect_mr->conn;
2446         int rc = 0;
2447
2448         if (smbdirect_mr->need_invalidate) {
2449                 /* Need to finish local invalidation before returning */
2450                 wr = &smbdirect_mr->inv_wr;
2451                 wr->opcode = IB_WR_LOCAL_INV;
2452                 smbdirect_mr->cqe.done = local_inv_done;
2453                 wr->wr_cqe = &smbdirect_mr->cqe;
2454                 wr->num_sge = 0;
2455                 wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey;
2456                 wr->send_flags = IB_SEND_SIGNALED;
2457
2458                 init_completion(&smbdirect_mr->invalidate_done);
2459                 rc = ib_post_send(info->id->qp, wr, NULL);
2460                 if (rc) {
2461                         log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc);
2462                         smbd_disconnect_rdma_connection(info);
2463                         goto done;
2464                 }
2465                 wait_for_completion(&smbdirect_mr->invalidate_done);
2466                 smbdirect_mr->need_invalidate = false;
2467         } else
2468                 /*
2469                  * For remote invalidation, just set it to MR_INVALIDATED
2470                  * and defer to mr_recovery_work to recover the MR for next use
2471                  */
2472                 smbdirect_mr->state = MR_INVALIDATED;
2473
2474         if (smbdirect_mr->state == MR_INVALIDATED) {
2475                 ib_dma_unmap_sg(
2476                         info->id->device, smbdirect_mr->sgl,
2477                         smbdirect_mr->sgl_count,
2478                         smbdirect_mr->dir);
2479                 smbdirect_mr->state = MR_READY;
2480                 if (atomic_inc_return(&info->mr_ready_count) == 1)
2481                         wake_up_interruptible(&info->wait_mr);
2482         } else
2483                 /*
2484                  * Schedule the work to do MR recovery for future I/Os MR
2485                  * recovery is slow and don't want it to block current I/O
2486                  */
2487                 queue_work(info->workqueue, &info->mr_recovery_work);
2488
2489 done:
2490         if (atomic_dec_and_test(&info->mr_used_count))
2491                 wake_up(&info->wait_for_mr_cleanup);
2492
2493         return rc;
2494 }