root/fs/cifs/smbdirect.c

DEFINITIONS

1. smbd_disconnect_rdma_work
2. smbd_disconnect_rdma_connection
3. smbd_conn_upcall
4. smbd_qp_async_error_upcall
5. smbd_request_payload
6. smbd_response_payload
7. send_done
8. dump_smbd_negotiate_resp
9. process_negotiation_response
10. check_and_send_immediate
11. smbd_post_send_credits
12. smbd_recv_done_work
13. recv_done
14. smbd_create_id
15. frwr_is_supported
16. smbd_ia_open
17. smbd_post_send_negotiate_req
18. manage_credits_prior_sending
19. manage_keep_alive_before_sending
20. smbd_create_header
21. smbd_destroy_header
22. smbd_post_send
23. smbd_post_send_sgl
24. smbd_post_send_page
25. smbd_post_send_empty
26. smbd_post_send_data
27. smbd_post_recv
28. smbd_negotiate
29. put_empty_packet
30. enqueue_reassembly
31. _get_first_reassembly
32. get_empty_queue_buffer
33. get_receive_buffer
34. put_receive_buffer
35. allocate_receive_buffers
36. destroy_receive_buffers
37. send_immediate_work
38. idle_connection_timer
39. smbd_destroy
40. smbd_reconnect
41. destroy_caches_and_workqueue
42. allocate_caches_and_workqueue
43. _smbd_get_connection
44. smbd_get_connection
45. smbd_recv_buf
46. smbd_recv_page
47. smbd_recv
48. smbd_send
49. register_mr_done
50. smbd_mr_recovery_work
51. destroy_mr_list
52. allocate_mr_list
53. get_mr
54. smbd_register_mr
55. local_inv_done
56. smbd_deregister_mr

   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 = 8192;
  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 SGEs in a 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 
 250         default:
 251                 break;
 252         }
 253 }
 254 
 255 static inline void *smbd_request_payload(struct smbd_request *request)
 256 {
 257         return (void *)request->packet;
 258 }
 259 
 260 static inline void *smbd_response_payload(struct smbd_response *response)
 261 {
 262         return (void *)response->packet;
 263 }
 264 
 265 /* Called when a RDMA send is done */
 266 static void send_done(struct ib_cq *cq, struct ib_wc *wc)
 267 {
 268         int i;
 269         struct smbd_request *request =
 270                 container_of(wc->wr_cqe, struct smbd_request, cqe);
 271 
 272         log_rdma_send(INFO, "smbd_request %p completed wc->status=%d\n",
 273                 request, wc->status);
 274 
 275         if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) {
 276                 log_rdma_send(ERR, "wc->status=%d wc->opcode=%d\n",
 277                         wc->status, wc->opcode);
 278                 smbd_disconnect_rdma_connection(request->info);
 279         }
 280 
 281         for (i = 0; i < request->num_sge; i++)
 282                 ib_dma_unmap_single(request->info->id->device,
 283                         request->sge[i].addr,
 284                         request->sge[i].length,
 285                         DMA_TO_DEVICE);
 286 
 287         if (request->has_payload) {
 288                 if (atomic_dec_and_test(&request->info->send_payload_pending))
 289                         wake_up(&request->info->wait_send_payload_pending);
 290         } else {
 291                 if (atomic_dec_and_test(&request->info->send_pending))
 292                         wake_up(&request->info->wait_send_pending);
 293         }
 294 
 295         mempool_free(request, request->info->request_mempool);
 296 }
 297 
 298 static void dump_smbd_negotiate_resp(struct smbd_negotiate_resp *resp)
 299 {
 300         log_rdma_event(INFO, "resp message min_version %u max_version %u "
 301                 "negotiated_version %u credits_requested %u "
 302                 "credits_granted %u status %u max_readwrite_size %u "
 303                 "preferred_send_size %u max_receive_size %u "
 304                 "max_fragmented_size %u\n",
 305                 resp->min_version, resp->max_version, resp->negotiated_version,
 306                 resp->credits_requested, resp->credits_granted, resp->status,
 307                 resp->max_readwrite_size, resp->preferred_send_size,
 308                 resp->max_receive_size, resp->max_fragmented_size);
 309 }
 310 
 311 /*
 312  * Process a negotiation response message, according to [MS-SMBD]3.1.5.7
 313  * response, packet_length: the negotiation response message
 314  * return value: true if negotiation is a success, false if failed
 315  */
 316 static bool process_negotiation_response(
 317                 struct smbd_response *response, int packet_length)
 318 {
 319         struct smbd_connection *info = response->info;
 320         struct smbd_negotiate_resp *packet = smbd_response_payload(response);
 321 
 322         if (packet_length < sizeof(struct smbd_negotiate_resp)) {
 323                 log_rdma_event(ERR,
 324                         "error: packet_length=%d\n", packet_length);
 325                 return false;
 326         }
 327 
 328         if (le16_to_cpu(packet->negotiated_version) != SMBD_V1) {
 329                 log_rdma_event(ERR, "error: negotiated_version=%x\n",
 330                         le16_to_cpu(packet->negotiated_version));
 331                 return false;
 332         }
 333         info->protocol = le16_to_cpu(packet->negotiated_version);
 334 
 335         if (packet->credits_requested == 0) {
 336                 log_rdma_event(ERR, "error: credits_requested==0\n");
 337                 return false;
 338         }
 339         info->receive_credit_target = le16_to_cpu(packet->credits_requested);
 340 
 341         if (packet->credits_granted == 0) {
 342                 log_rdma_event(ERR, "error: credits_granted==0\n");
 343                 return false;
 344         }
 345         atomic_set(&info->send_credits, le16_to_cpu(packet->credits_granted));
 346 
 347         atomic_set(&info->receive_credits, 0);
 348 
 349         if (le32_to_cpu(packet->preferred_send_size) > info->max_receive_size) {
 350                 log_rdma_event(ERR, "error: preferred_send_size=%d\n",
 351                         le32_to_cpu(packet->preferred_send_size));
 352                 return false;
 353         }
 354         info->max_receive_size = le32_to_cpu(packet->preferred_send_size);
 355 
 356         if (le32_to_cpu(packet->max_receive_size) < SMBD_MIN_RECEIVE_SIZE) {
 357                 log_rdma_event(ERR, "error: max_receive_size=%d\n",
 358                         le32_to_cpu(packet->max_receive_size));
 359                 return false;
 360         }
 361         info->max_send_size = min_t(int, info->max_send_size,
 362                                         le32_to_cpu(packet->max_receive_size));
 363 
 364         if (le32_to_cpu(packet->max_fragmented_size) <
 365                         SMBD_MIN_FRAGMENTED_SIZE) {
 366                 log_rdma_event(ERR, "error: max_fragmented_size=%d\n",
 367                         le32_to_cpu(packet->max_fragmented_size));
 368                 return false;
 369         }
 370         info->max_fragmented_send_size =
 371                 le32_to_cpu(packet->max_fragmented_size);
 372         info->rdma_readwrite_threshold =
 373                 rdma_readwrite_threshold > info->max_fragmented_send_size ?
 374                 info->max_fragmented_send_size :
 375                 rdma_readwrite_threshold;
 376 
 377 
 378         info->max_readwrite_size = min_t(u32,
 379                         le32_to_cpu(packet->max_readwrite_size),
 380                         info->max_frmr_depth * PAGE_SIZE);
 381         info->max_frmr_depth = info->max_readwrite_size / PAGE_SIZE;
 382 
 383         return true;
 384 }
 385 
 386 /*
 387  * Check and schedule to send an immediate packet
 388  * This is used to extend credtis to remote peer to keep the transport busy
 389  */
 390 static void check_and_send_immediate(struct smbd_connection *info)
 391 {
 392         if (info->transport_status != SMBD_CONNECTED)
 393                 return;
 394 
 395         info->send_immediate = true;
 396 
 397         /*
 398          * Promptly send a packet if our peer is running low on receive
 399          * credits
 400          */
 401         if (atomic_read(&info->receive_credits) <
 402                 info->receive_credit_target - 1)
 403                 queue_delayed_work(
 404                         info->workqueue, &info->send_immediate_work, 0);
 405 }
 406 
 407 static void smbd_post_send_credits(struct work_struct *work)
 408 {
 409         int ret = 0;
 410         int use_receive_queue = 1;
 411         int rc;
 412         struct smbd_response *response;
 413         struct smbd_connection *info =
 414                 container_of(work, struct smbd_connection,
 415                         post_send_credits_work);
 416 
 417         if (info->transport_status != SMBD_CONNECTED) {
 418                 wake_up(&info->wait_receive_queues);
 419                 return;
 420         }
 421 
 422         if (info->receive_credit_target >
 423                 atomic_read(&info->receive_credits)) {
 424                 while (true) {
 425                         if (use_receive_queue)
 426                                 response = get_receive_buffer(info);
 427                         else
 428                                 response = get_empty_queue_buffer(info);
 429                         if (!response) {
 430                                 /* now switch to emtpy packet queue */
 431                                 if (use_receive_queue) {
 432                                         use_receive_queue = 0;
 433                                         continue;
 434                                 } else
 435                                         break;
 436                         }
 437 
 438                         response->type = SMBD_TRANSFER_DATA;
 439                         response->first_segment = false;
 440                         rc = smbd_post_recv(info, response);
 441                         if (rc) {
 442                                 log_rdma_recv(ERR,
 443                                         "post_recv failed rc=%d\n", rc);
 444                                 put_receive_buffer(info, response);
 445                                 break;
 446                         }
 447 
 448                         ret++;
 449                 }
 450         }
 451 
 452         spin_lock(&info->lock_new_credits_offered);
 453         info->new_credits_offered += ret;
 454         spin_unlock(&info->lock_new_credits_offered);
 455 
 456         atomic_add(ret, &info->receive_credits);
 457 
 458         /* Check if we can post new receive and grant credits to peer */
 459         check_and_send_immediate(info);
 460 }
 461 
 462 static void smbd_recv_done_work(struct work_struct *work)
 463 {
 464         struct smbd_connection *info =
 465                 container_of(work, struct smbd_connection, recv_done_work);
 466 
 467         /*
 468          * We may have new send credits granted from remote peer
 469          * If any sender is blcoked on lack of credets, unblock it
 470          */
 471         if (atomic_read(&info->send_credits))
 472                 wake_up_interruptible(&info->wait_send_queue);
 473 
 474         /*
 475          * Check if we need to send something to remote peer to
 476          * grant more credits or respond to KEEP_ALIVE packet
 477          */
 478         check_and_send_immediate(info);
 479 }
 480 
 481 /* Called from softirq, when recv is done */
 482 static void recv_done(struct ib_cq *cq, struct ib_wc *wc)
 483 {
 484         struct smbd_data_transfer *data_transfer;
 485         struct smbd_response *response =
 486                 container_of(wc->wr_cqe, struct smbd_response, cqe);
 487         struct smbd_connection *info = response->info;
 488         int data_length = 0;
 489 
 490         log_rdma_recv(INFO, "response=%p type=%d wc status=%d wc opcode %d "
 491                       "byte_len=%d pkey_index=%x\n",
 492                 response, response->type, wc->status, wc->opcode,
 493                 wc->byte_len, wc->pkey_index);
 494 
 495         if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) {
 496                 log_rdma_recv(INFO, "wc->status=%d opcode=%d\n",
 497                         wc->status, wc->opcode);
 498                 smbd_disconnect_rdma_connection(info);
 499                 goto error;
 500         }
 501 
 502         ib_dma_sync_single_for_cpu(
 503                 wc->qp->device,
 504                 response->sge.addr,
 505                 response->sge.length,
 506                 DMA_FROM_DEVICE);
 507 
 508         switch (response->type) {
 509         /* SMBD negotiation response */
 510         case SMBD_NEGOTIATE_RESP:
 511                 dump_smbd_negotiate_resp(smbd_response_payload(response));
 512                 info->full_packet_received = true;
 513                 info->negotiate_done =
 514                         process_negotiation_response(response, wc->byte_len);
 515                 complete(&info->negotiate_completion);
 516                 break;
 517 
 518         /* SMBD data transfer packet */
 519         case SMBD_TRANSFER_DATA:
 520                 data_transfer = smbd_response_payload(response);
 521                 data_length = le32_to_cpu(data_transfer->data_length);
 522 
 523                 /*
 524                  * If this is a packet with data playload place the data in
 525                  * reassembly queue and wake up the reading thread
 526                  */
 527                 if (data_length) {
 528                         if (info->full_packet_received)
 529                                 response->first_segment = true;
 530 
 531                         if (le32_to_cpu(data_transfer->remaining_data_length))
 532                                 info->full_packet_received = false;
 533                         else
 534                                 info->full_packet_received = true;
 535 
 536                         enqueue_reassembly(
 537                                 info,
 538                                 response,
 539                                 data_length);
 540                 } else
 541                         put_empty_packet(info, response);
 542 
 543                 if (data_length)
 544                         wake_up_interruptible(&info->wait_reassembly_queue);
 545 
 546                 atomic_dec(&info->receive_credits);
 547                 info->receive_credit_target =
 548                         le16_to_cpu(data_transfer->credits_requested);
 549                 atomic_add(le16_to_cpu(data_transfer->credits_granted),
 550                         &info->send_credits);
 551 
 552                 log_incoming(INFO, "data flags %d data_offset %d "
 553                         "data_length %d remaining_data_length %d\n",
 554                         le16_to_cpu(data_transfer->flags),
 555                         le32_to_cpu(data_transfer->data_offset),
 556                         le32_to_cpu(data_transfer->data_length),
 557                         le32_to_cpu(data_transfer->remaining_data_length));
 558 
 559                 /* Send a KEEP_ALIVE response right away if requested */
 560                 info->keep_alive_requested = KEEP_ALIVE_NONE;
 561                 if (le16_to_cpu(data_transfer->flags) &
 562                                 SMB_DIRECT_RESPONSE_REQUESTED) {
 563                         info->keep_alive_requested = KEEP_ALIVE_PENDING;
 564                 }
 565 
 566                 queue_work(info->workqueue, &info->recv_done_work);
 567                 return;
 568 
 569         default:
 570                 log_rdma_recv(ERR,
 571                         "unexpected response type=%d\n", response->type);
 572         }
 573 
 574 error:
 575         put_receive_buffer(info, response);
 576 }
 577 
 578 static struct rdma_cm_id *smbd_create_id(
 579                 struct smbd_connection *info,
 580                 struct sockaddr *dstaddr, int port)
 581 {
 582         struct rdma_cm_id *id;
 583         int rc;
 584         __be16 *sport;
 585 
 586         id = rdma_create_id(&init_net, smbd_conn_upcall, info,
 587                 RDMA_PS_TCP, IB_QPT_RC);
 588         if (IS_ERR(id)) {
 589                 rc = PTR_ERR(id);
 590                 log_rdma_event(ERR, "rdma_create_id() failed %i\n", rc);
 591                 return id;
 592         }
 593 
 594         if (dstaddr->sa_family == AF_INET6)
 595                 sport = &((struct sockaddr_in6 *)dstaddr)->sin6_port;
 596         else
 597                 sport = &((struct sockaddr_in *)dstaddr)->sin_port;
 598 
 599         *sport = htons(port);
 600 
 601         init_completion(&info->ri_done);
 602         info->ri_rc = -ETIMEDOUT;
 603 
 604         rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)dstaddr,
 605                 RDMA_RESOLVE_TIMEOUT);
 606         if (rc) {
 607                 log_rdma_event(ERR, "rdma_resolve_addr() failed %i\n", rc);
 608                 goto out;
 609         }
 610         wait_for_completion_interruptible_timeout(
 611                 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
 612         rc = info->ri_rc;
 613         if (rc) {
 614                 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc);
 615                 goto out;
 616         }
 617 
 618         info->ri_rc = -ETIMEDOUT;
 619         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 620         if (rc) {
 621                 log_rdma_event(ERR, "rdma_resolve_route() failed %i\n", rc);
 622                 goto out;
 623         }
 624         wait_for_completion_interruptible_timeout(
 625                 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
 626         rc = info->ri_rc;
 627         if (rc) {
 628                 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc);
 629                 goto out;
 630         }
 631 
 632         return id;
 633 
 634 out:
 635         rdma_destroy_id(id);
 636         return ERR_PTR(rc);
 637 }
 638 
 639 /*
 640  * Test if FRWR (Fast Registration Work Requests) is supported on the device
 641  * This implementation requries FRWR on RDMA read/write
 642  * return value: true if it is supported
 643  */
 644 static bool frwr_is_supported(struct ib_device_attr *attrs)
 645 {
 646         if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
 647                 return false;
 648         if (attrs->max_fast_reg_page_list_len == 0)
 649                 return false;
 650         return true;
 651 }
 652 
 653 static int smbd_ia_open(
 654                 struct smbd_connection *info,
 655                 struct sockaddr *dstaddr, int port)
 656 {
 657         int rc;
 658 
 659         info->id = smbd_create_id(info, dstaddr, port);
 660         if (IS_ERR(info->id)) {
 661                 rc = PTR_ERR(info->id);
 662                 goto out1;
 663         }
 664 
 665         if (!frwr_is_supported(&info->id->device->attrs)) {
 666                 log_rdma_event(ERR,
 667                         "Fast Registration Work Requests "
 668                         "(FRWR) is not supported\n");
 669                 log_rdma_event(ERR,
 670                         "Device capability flags = %llx "
 671                         "max_fast_reg_page_list_len = %u\n",
 672                         info->id->device->attrs.device_cap_flags,
 673                         info->id->device->attrs.max_fast_reg_page_list_len);
 674                 rc = -EPROTONOSUPPORT;
 675                 goto out2;
 676         }
 677         info->max_frmr_depth = min_t(int,
 678                 smbd_max_frmr_depth,
 679                 info->id->device->attrs.max_fast_reg_page_list_len);
 680         info->mr_type = IB_MR_TYPE_MEM_REG;
 681         if (info->id->device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
 682                 info->mr_type = IB_MR_TYPE_SG_GAPS;
 683 
 684         info->pd = ib_alloc_pd(info->id->device, 0);
 685         if (IS_ERR(info->pd)) {
 686                 rc = PTR_ERR(info->pd);
 687                 log_rdma_event(ERR, "ib_alloc_pd() returned %d\n", rc);
 688                 goto out2;
 689         }
 690 
 691         return 0;
 692 
 693 out2:
 694         rdma_destroy_id(info->id);
 695         info->id = NULL;
 696 
 697 out1:
 698         return rc;
 699 }
 700 
 701 /*
 702  * Send a negotiation request message to the peer
 703  * The negotiation procedure is in [MS-SMBD] 3.1.5.2 and 3.1.5.3
 704  * After negotiation, the transport is connected and ready for
 705  * carrying upper layer SMB payload
 706  */
 707 static int smbd_post_send_negotiate_req(struct smbd_connection *info)
 708 {
 709         struct ib_send_wr send_wr;
 710         int rc = -ENOMEM;
 711         struct smbd_request *request;
 712         struct smbd_negotiate_req *packet;
 713 
 714         request = mempool_alloc(info->request_mempool, GFP_KERNEL);
 715         if (!request)
 716                 return rc;
 717 
 718         request->info = info;
 719 
 720         packet = smbd_request_payload(request);
 721         packet->min_version = cpu_to_le16(SMBD_V1);
 722         packet->max_version = cpu_to_le16(SMBD_V1);
 723         packet->reserved = 0;
 724         packet->credits_requested = cpu_to_le16(info->send_credit_target);
 725         packet->preferred_send_size = cpu_to_le32(info->max_send_size);
 726         packet->max_receive_size = cpu_to_le32(info->max_receive_size);
 727         packet->max_fragmented_size =
 728                 cpu_to_le32(info->max_fragmented_recv_size);
 729 
 730         request->num_sge = 1;
 731         request->sge[0].addr = ib_dma_map_single(
 732                                 info->id->device, (void *)packet,
 733                                 sizeof(*packet), DMA_TO_DEVICE);
 734         if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
 735                 rc = -EIO;
 736                 goto dma_mapping_failed;
 737         }
 738 
 739         request->sge[0].length = sizeof(*packet);
 740         request->sge[0].lkey = info->pd->local_dma_lkey;
 741 
 742         ib_dma_sync_single_for_device(
 743                 info->id->device, request->sge[0].addr,
 744                 request->sge[0].length, DMA_TO_DEVICE);
 745 
 746         request->cqe.done = send_done;
 747 
 748         send_wr.next = NULL;
 749         send_wr.wr_cqe = &request->cqe;
 750         send_wr.sg_list = request->sge;
 751         send_wr.num_sge = request->num_sge;
 752         send_wr.opcode = IB_WR_SEND;
 753         send_wr.send_flags = IB_SEND_SIGNALED;
 754 
 755         log_rdma_send(INFO, "sge addr=%llx length=%x lkey=%x\n",
 756                 request->sge[0].addr,
 757                 request->sge[0].length, request->sge[0].lkey);
 758 
 759         request->has_payload = false;
 760         atomic_inc(&info->send_pending);
 761         rc = ib_post_send(info->id->qp, &send_wr, NULL);
 762         if (!rc)
 763                 return 0;
 764 
 765         /* if we reach here, post send failed */
 766         log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
 767         atomic_dec(&info->send_pending);
 768         ib_dma_unmap_single(info->id->device, request->sge[0].addr,
 769                 request->sge[0].length, DMA_TO_DEVICE);
 770 
 771         smbd_disconnect_rdma_connection(info);
 772 
 773 dma_mapping_failed:
 774         mempool_free(request, info->request_mempool);
 775         return rc;
 776 }
 777 
 778 /*
 779  * Extend the credits to remote peer
 780  * This implements [MS-SMBD] 3.1.5.9
 781  * The idea is that we should extend credits to remote peer as quickly as
 782  * it's allowed, to maintain data flow. We allocate as much receive
 783  * buffer as possible, and extend the receive credits to remote peer
 784  * return value: the new credtis being granted.
 785  */
 786 static int manage_credits_prior_sending(struct smbd_connection *info)
 787 {
 788         int new_credits;
 789 
 790         spin_lock(&info->lock_new_credits_offered);
 791         new_credits = info->new_credits_offered;
 792         info->new_credits_offered = 0;
 793         spin_unlock(&info->lock_new_credits_offered);
 794 
 795         return new_credits;
 796 }
 797 
 798 /*
 799  * Check if we need to send a KEEP_ALIVE message
 800  * The idle connection timer triggers a KEEP_ALIVE message when expires
 801  * SMB_DIRECT_RESPONSE_REQUESTED is set in the message flag to have peer send
 802  * back a response.
 803  * return value:
 804  * 1 if SMB_DIRECT_RESPONSE_REQUESTED needs to be set
 805  * 0: otherwise
 806  */
 807 static int manage_keep_alive_before_sending(struct smbd_connection *info)
 808 {
 809         if (info->keep_alive_requested == KEEP_ALIVE_PENDING) {
 810                 info->keep_alive_requested = KEEP_ALIVE_SENT;
 811                 return 1;
 812         }
 813         return 0;
 814 }
 815 
 816 /*
 817  * Build and prepare the SMBD packet header
 818  * This function waits for avaialbe send credits and build a SMBD packet
 819  * header. The caller then optional append payload to the packet after
 820  * the header
 821  * intput values
 822  * size: the size of the payload
 823  * remaining_data_length: remaining data to send if this is part of a
 824  * fragmented packet
 825  * output values
 826  * request_out: the request allocated from this function
 827  * return values: 0 on success, otherwise actual error code returned
 828  */
 829 static int smbd_create_header(struct smbd_connection *info,
 830                 int size, int remaining_data_length,
 831                 struct smbd_request **request_out)
 832 {
 833         struct smbd_request *request;
 834         struct smbd_data_transfer *packet;
 835         int header_length;
 836         int rc;
 837 
 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                 return rc;
 844 
 845         if (info->transport_status != SMBD_CONNECTED) {
 846                 log_outgoing(ERR, "disconnected not sending\n");
 847                 return -EAGAIN;
 848         }
 849         atomic_dec(&info->send_credits);
 850 
 851         request = mempool_alloc(info->request_mempool, GFP_KERNEL);
 852         if (!request) {
 853                 rc = -ENOMEM;
 854                 goto err;
 855         }
 856 
 857         request->info = info;
 858 
 859         /* Fill in the packet header */
 860         packet = smbd_request_payload(request);
 861         packet->credits_requested = cpu_to_le16(info->send_credit_target);
 862         packet->credits_granted =
 863                 cpu_to_le16(manage_credits_prior_sending(info));
 864         info->send_immediate = false;
 865 
 866         packet->flags = 0;
 867         if (manage_keep_alive_before_sending(info))
 868                 packet->flags |= cpu_to_le16(SMB_DIRECT_RESPONSE_REQUESTED);
 869 
 870         packet->reserved = 0;
 871         if (!size)
 872                 packet->data_offset = 0;
 873         else
 874                 packet->data_offset = cpu_to_le32(24);
 875         packet->data_length = cpu_to_le32(size);
 876         packet->remaining_data_length = cpu_to_le32(remaining_data_length);
 877         packet->padding = 0;
 878 
 879         log_outgoing(INFO, "credits_requested=%d credits_granted=%d "
 880                 "data_offset=%d data_length=%d remaining_data_length=%d\n",
 881                 le16_to_cpu(packet->credits_requested),
 882                 le16_to_cpu(packet->credits_granted),
 883                 le32_to_cpu(packet->data_offset),
 884                 le32_to_cpu(packet->data_length),
 885                 le32_to_cpu(packet->remaining_data_length));
 886 
 887         /* Map the packet to DMA */
 888         header_length = sizeof(struct smbd_data_transfer);
 889         /* If this is a packet without payload, don't send padding */
 890         if (!size)
 891                 header_length = offsetof(struct smbd_data_transfer, padding);
 892 
 893         request->num_sge = 1;
 894         request->sge[0].addr = ib_dma_map_single(info->id->device,
 895                                                  (void *)packet,
 896                                                  header_length,
 897                                                  DMA_TO_DEVICE);
 898         if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
 899                 mempool_free(request, info->request_mempool);
 900                 rc = -EIO;
 901                 goto err;
 902         }
 903 
 904         request->sge[0].length = header_length;
 905         request->sge[0].lkey = info->pd->local_dma_lkey;
 906 
 907         *request_out = request;
 908         return 0;
 909 
 910 err:
 911         atomic_inc(&info->send_credits);
 912         return rc;
 913 }
 914 
 915 static void smbd_destroy_header(struct smbd_connection *info,
 916                 struct smbd_request *request)
 917 {
 918 
 919         ib_dma_unmap_single(info->id->device,
 920                             request->sge[0].addr,
 921                             request->sge[0].length,
 922                             DMA_TO_DEVICE);
 923         mempool_free(request, info->request_mempool);
 924         atomic_inc(&info->send_credits);
 925 }
 926 
 927 /* Post the send request */
 928 static int smbd_post_send(struct smbd_connection *info,
 929                 struct smbd_request *request, bool has_payload)
 930 {
 931         struct ib_send_wr send_wr;
 932         int rc, i;
 933 
 934         for (i = 0; i < request->num_sge; i++) {
 935                 log_rdma_send(INFO,
 936                         "rdma_request sge[%d] addr=%llu length=%u\n",
 937                         i, request->sge[i].addr, request->sge[i].length);
 938                 ib_dma_sync_single_for_device(
 939                         info->id->device,
 940                         request->sge[i].addr,
 941                         request->sge[i].length,
 942                         DMA_TO_DEVICE);
 943         }
 944 
 945         request->cqe.done = send_done;
 946 
 947         send_wr.next = NULL;
 948         send_wr.wr_cqe = &request->cqe;
 949         send_wr.sg_list = request->sge;
 950         send_wr.num_sge = request->num_sge;
 951         send_wr.opcode = IB_WR_SEND;
 952         send_wr.send_flags = IB_SEND_SIGNALED;
 953 
 954         if (has_payload) {
 955                 request->has_payload = true;
 956                 atomic_inc(&info->send_payload_pending);
 957         } else {
 958                 request->has_payload = false;
 959                 atomic_inc(&info->send_pending);
 960         }
 961 
 962         rc = ib_post_send(info->id->qp, &send_wr, NULL);
 963         if (rc) {
 964                 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
 965                 if (has_payload) {
 966                         if (atomic_dec_and_test(&info->send_payload_pending))
 967                                 wake_up(&info->wait_send_payload_pending);
 968                 } else {
 969                         if (atomic_dec_and_test(&info->send_pending))
 970                                 wake_up(&info->wait_send_pending);
 971                 }
 972                 smbd_disconnect_rdma_connection(info);
 973                 rc = -EAGAIN;
 974         } else
 975                 /* Reset timer for idle connection after packet is sent */
 976                 mod_delayed_work(info->workqueue, &info->idle_timer_work,
 977                         info->keep_alive_interval*HZ);
 978 
 979         return rc;
 980 }
 981 
 982 static int smbd_post_send_sgl(struct smbd_connection *info,
 983         struct scatterlist *sgl, int data_length, int remaining_data_length)
 984 {
 985         int num_sgs;
 986         int i, rc;
 987         struct smbd_request *request;
 988         struct scatterlist *sg;
 989 
 990         rc = smbd_create_header(
 991                 info, data_length, remaining_data_length, &request);
 992         if (rc)
 993                 return rc;
 994 
 995         num_sgs = sgl ? sg_nents(sgl) : 0;
 996         for_each_sg(sgl, sg, num_sgs, i) {
 997                 request->sge[i+1].addr =
 998                         ib_dma_map_page(info->id->device, sg_page(sg),
 999                                sg->offset, sg->length, DMA_TO_DEVICE);
1000                 if (ib_dma_mapping_error(
1001                                 info->id->device, request->sge[i+1].addr)) {
1002                         rc = -EIO;
1003                         request->sge[i+1].addr = 0;
1004                         goto dma_mapping_failure;
1005                 }
1006                 request->sge[i+1].length = sg->length;
1007                 request->sge[i+1].lkey = info->pd->local_dma_lkey;
1008                 request->num_sge++;
1009         }
1010 
1011         rc = smbd_post_send(info, request, data_length);
1012         if (!rc)
1013                 return 0;
1014 
1015 dma_mapping_failure:
1016         for (i = 1; i < request->num_sge; i++)
1017                 if (request->sge[i].addr)
1018                         ib_dma_unmap_single(info->id->device,
1019                                             request->sge[i].addr,
1020                                             request->sge[i].length,
1021                                             DMA_TO_DEVICE);
1022         smbd_destroy_header(info, request);
1023         return rc;
1024 }
1025 
1026 /*
1027  * Send a page
1028  * page: the page to send
1029  * offset: offset in the page to send
1030  * size: length in the page to send
1031  * remaining_data_length: remaining data to send in this payload
1032  */
1033 static int smbd_post_send_page(struct smbd_connection *info, struct page *page,
1034                 unsigned long offset, size_t size, int remaining_data_length)
1035 {
1036         struct scatterlist sgl;
1037 
1038         sg_init_table(&sgl, 1);
1039         sg_set_page(&sgl, page, size, offset);
1040 
1041         return smbd_post_send_sgl(info, &sgl, size, remaining_data_length);
1042 }
1043 
1044 /*
1045  * Send an empty message
1046  * Empty message is used to extend credits to peer to for keep live
1047  * while there is no upper layer payload to send at the time
1048  */
1049 static int smbd_post_send_empty(struct smbd_connection *info)
1050 {
1051         info->count_send_empty++;
1052         return smbd_post_send_sgl(info, NULL, 0, 0);
1053 }
1054 
1055 /*
1056  * Send a data buffer
1057  * iov: the iov array describing the data buffers
1058  * n_vec: number of iov array
1059  * remaining_data_length: remaining data to send following this packet
1060  * in segmented SMBD packet
1061  */
1062 static int smbd_post_send_data(
1063         struct smbd_connection *info, struct kvec *iov, int n_vec,
1064         int remaining_data_length)
1065 {
1066         int i;
1067         u32 data_length = 0;
1068         struct scatterlist sgl[SMBDIRECT_MAX_SGE];
1069 
1070         if (n_vec > SMBDIRECT_MAX_SGE) {
1071                 cifs_dbg(VFS, "Can't fit data to SGL, n_vec=%d\n", n_vec);
1072                 return -EINVAL;
1073         }
1074 
1075         sg_init_table(sgl, n_vec);
1076         for (i = 0; i < n_vec; i++) {
1077                 data_length += iov[i].iov_len;
1078                 sg_set_buf(&sgl[i], iov[i].iov_base, iov[i].iov_len);
1079         }
1080 
1081         return smbd_post_send_sgl(info, sgl, data_length, remaining_data_length);
1082 }
1083 
1084 /*
1085  * Post a receive request to the transport
1086  * The remote peer can only send data when a receive request is posted
1087  * The interaction is controlled by send/receive credit system
1088  */
1089 static int smbd_post_recv(
1090                 struct smbd_connection *info, struct smbd_response *response)
1091 {
1092         struct ib_recv_wr recv_wr;
1093         int rc = -EIO;
1094 
1095         response->sge.addr = ib_dma_map_single(
1096                                 info->id->device, response->packet,
1097                                 info->max_receive_size, DMA_FROM_DEVICE);
1098         if (ib_dma_mapping_error(info->id->device, response->sge.addr))
1099                 return rc;
1100 
1101         response->sge.length = info->max_receive_size;
1102         response->sge.lkey = info->pd->local_dma_lkey;
1103 
1104         response->cqe.done = recv_done;
1105 
1106         recv_wr.wr_cqe = &response->cqe;
1107         recv_wr.next = NULL;
1108         recv_wr.sg_list = &response->sge;
1109         recv_wr.num_sge = 1;
1110 
1111         rc = ib_post_recv(info->id->qp, &recv_wr, NULL);
1112         if (rc) {
1113                 ib_dma_unmap_single(info->id->device, response->sge.addr,
1114                                     response->sge.length, DMA_FROM_DEVICE);
1115                 smbd_disconnect_rdma_connection(info);
1116                 log_rdma_recv(ERR, "ib_post_recv failed rc=%d\n", rc);
1117         }
1118 
1119         return rc;
1120 }
1121 
1122 /* Perform SMBD negotiate according to [MS-SMBD] 3.1.5.2 */
1123 static int smbd_negotiate(struct smbd_connection *info)
1124 {
1125         int rc;
1126         struct smbd_response *response = get_receive_buffer(info);
1127 
1128         response->type = SMBD_NEGOTIATE_RESP;
1129         rc = smbd_post_recv(info, response);
1130         log_rdma_event(INFO,
1131                 "smbd_post_recv rc=%d iov.addr=%llx iov.length=%x "
1132                 "iov.lkey=%x\n",
1133                 rc, response->sge.addr,
1134                 response->sge.length, response->sge.lkey);
1135         if (rc)
1136                 return rc;
1137 
1138         init_completion(&info->negotiate_completion);
1139         info->negotiate_done = false;
1140         rc = smbd_post_send_negotiate_req(info);
1141         if (rc)
1142                 return rc;
1143 
1144         rc = wait_for_completion_interruptible_timeout(
1145                 &info->negotiate_completion, SMBD_NEGOTIATE_TIMEOUT * HZ);
1146         log_rdma_event(INFO, "wait_for_completion_timeout rc=%d\n", rc);
1147 
1148         if (info->negotiate_done)
1149                 return 0;
1150 
1151         if (rc == 0)
1152                 rc = -ETIMEDOUT;
1153         else if (rc == -ERESTARTSYS)
1154                 rc = -EINTR;
1155         else
1156                 rc = -ENOTCONN;
1157 
1158         return rc;
1159 }
1160 
1161 static void put_empty_packet(
1162                 struct smbd_connection *info, struct smbd_response *response)
1163 {
1164         spin_lock(&info->empty_packet_queue_lock);
1165         list_add_tail(&response->list, &info->empty_packet_queue);
1166         info->count_empty_packet_queue++;
1167         spin_unlock(&info->empty_packet_queue_lock);
1168 
1169         queue_work(info->workqueue, &info->post_send_credits_work);
1170 }
1171 
1172 /*
1173  * Implement Connection.FragmentReassemblyBuffer defined in [MS-SMBD] 3.1.1.1
1174  * This is a queue for reassembling upper layer payload and present to upper
1175  * layer. All the inncoming payload go to the reassembly queue, regardless of
1176  * if reassembly is required. The uuper layer code reads from the queue for all
1177  * incoming payloads.
1178  * Put a received packet to the reassembly queue
1179  * response: the packet received
1180  * data_length: the size of payload in this packet
1181  */
1182 static void enqueue_reassembly(
1183         struct smbd_connection *info,
1184         struct smbd_response *response,
1185         int data_length)
1186 {
1187         spin_lock(&info->reassembly_queue_lock);
1188         list_add_tail(&response->list, &info->reassembly_queue);
1189         info->reassembly_queue_length++;
1190         /*
1191          * Make sure reassembly_data_length is updated after list and
1192          * reassembly_queue_length are updated. On the dequeue side
1193          * reassembly_data_length is checked without a lock to determine
1194          * if reassembly_queue_length and list is up to date
1195          */
1196         virt_wmb();
1197         info->reassembly_data_length += data_length;
1198         spin_unlock(&info->reassembly_queue_lock);
1199         info->count_reassembly_queue++;
1200         info->count_enqueue_reassembly_queue++;
1201 }
1202 
1203 /*
1204  * Get the first entry at the front of reassembly queue
1205  * Caller is responsible for locking
1206  * return value: the first entry if any, NULL if queue is empty
1207  */
1208 static struct smbd_response *_get_first_reassembly(struct smbd_connection *info)
1209 {
1210         struct smbd_response *ret = NULL;
1211 
1212         if (!list_empty(&info->reassembly_queue)) {
1213                 ret = list_first_entry(
1214                         &info->reassembly_queue,
1215                         struct smbd_response, list);
1216         }
1217         return ret;
1218 }
1219 
1220 static struct smbd_response *get_empty_queue_buffer(
1221                 struct smbd_connection *info)
1222 {
1223         struct smbd_response *ret = NULL;
1224         unsigned long flags;
1225 
1226         spin_lock_irqsave(&info->empty_packet_queue_lock, flags);
1227         if (!list_empty(&info->empty_packet_queue)) {
1228                 ret = list_first_entry(
1229                         &info->empty_packet_queue,
1230                         struct smbd_response, list);
1231                 list_del(&ret->list);
1232                 info->count_empty_packet_queue--;
1233         }
1234         spin_unlock_irqrestore(&info->empty_packet_queue_lock, flags);
1235 
1236         return ret;
1237 }
1238 
1239 /*
1240  * Get a receive buffer
1241  * For each remote send, we need to post a receive. The receive buffers are
1242  * pre-allocated in advance.
1243  * return value: the receive buffer, NULL if none is available
1244  */
1245 static struct smbd_response *get_receive_buffer(struct smbd_connection *info)
1246 {
1247         struct smbd_response *ret = NULL;
1248         unsigned long flags;
1249 
1250         spin_lock_irqsave(&info->receive_queue_lock, flags);
1251         if (!list_empty(&info->receive_queue)) {
1252                 ret = list_first_entry(
1253                         &info->receive_queue,
1254                         struct smbd_response, list);
1255                 list_del(&ret->list);
1256                 info->count_receive_queue--;
1257                 info->count_get_receive_buffer++;
1258         }
1259         spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1260 
1261         return ret;
1262 }
1263 
1264 /*
1265  * Return a receive buffer
1266  * Upon returning of a receive buffer, we can post new receive and extend
1267  * more receive credits to remote peer. This is done immediately after a
1268  * receive buffer is returned.
1269  */
1270 static void put_receive_buffer(
1271         struct smbd_connection *info, struct smbd_response *response)
1272 {
1273         unsigned long flags;
1274 
1275         ib_dma_unmap_single(info->id->device, response->sge.addr,
1276                 response->sge.length, DMA_FROM_DEVICE);
1277 
1278         spin_lock_irqsave(&info->receive_queue_lock, flags);
1279         list_add_tail(&response->list, &info->receive_queue);
1280         info->count_receive_queue++;
1281         info->count_put_receive_buffer++;
1282         spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1283 
1284         queue_work(info->workqueue, &info->post_send_credits_work);
1285 }
1286 
1287 /* Preallocate all receive buffer on transport establishment */
1288 static int allocate_receive_buffers(struct smbd_connection *info, int num_buf)
1289 {
1290         int i;
1291         struct smbd_response *response;
1292 
1293         INIT_LIST_HEAD(&info->reassembly_queue);
1294         spin_lock_init(&info->reassembly_queue_lock);
1295         info->reassembly_data_length = 0;
1296         info->reassembly_queue_length = 0;
1297 
1298         INIT_LIST_HEAD(&info->receive_queue);
1299         spin_lock_init(&info->receive_queue_lock);
1300         info->count_receive_queue = 0;
1301 
1302         INIT_LIST_HEAD(&info->empty_packet_queue);
1303         spin_lock_init(&info->empty_packet_queue_lock);
1304         info->count_empty_packet_queue = 0;
1305 
1306         init_waitqueue_head(&info->wait_receive_queues);
1307 
1308         for (i = 0; i < num_buf; i++) {
1309                 response = mempool_alloc(info->response_mempool, GFP_KERNEL);
1310                 if (!response)
1311                         goto allocate_failed;
1312 
1313                 response->info = info;
1314                 list_add_tail(&response->list, &info->receive_queue);
1315                 info->count_receive_queue++;
1316         }
1317 
1318         return 0;
1319 
1320 allocate_failed:
1321         while (!list_empty(&info->receive_queue)) {
1322                 response = list_first_entry(
1323                                 &info->receive_queue,
1324                                 struct smbd_response, list);
1325                 list_del(&response->list);
1326                 info->count_receive_queue--;
1327 
1328                 mempool_free(response, info->response_mempool);
1329         }
1330         return -ENOMEM;
1331 }
1332 
1333 static void destroy_receive_buffers(struct smbd_connection *info)
1334 {
1335         struct smbd_response *response;
1336 
1337         while ((response = get_receive_buffer(info)))
1338                 mempool_free(response, info->response_mempool);
1339 
1340         while ((response = get_empty_queue_buffer(info)))
1341                 mempool_free(response, info->response_mempool);
1342 }
1343 
1344 /*
1345  * Check and send an immediate or keep alive packet
1346  * The condition to send those packets are defined in [MS-SMBD] 3.1.1.1
1347  * Connection.KeepaliveRequested and Connection.SendImmediate
1348  * The idea is to extend credits to server as soon as it becomes available
1349  */
1350 static void send_immediate_work(struct work_struct *work)
1351 {
1352         struct smbd_connection *info = container_of(
1353                                         work, struct smbd_connection,
1354                                         send_immediate_work.work);
1355 
1356         if (info->keep_alive_requested == KEEP_ALIVE_PENDING ||
1357             info->send_immediate) {
1358                 log_keep_alive(INFO, "send an empty message\n");
1359                 smbd_post_send_empty(info);
1360         }
1361 }
1362 
1363 /* Implement idle connection timer [MS-SMBD] 3.1.6.2 */
1364 static void idle_connection_timer(struct work_struct *work)
1365 {
1366         struct smbd_connection *info = container_of(
1367                                         work, struct smbd_connection,
1368                                         idle_timer_work.work);
1369 
1370         if (info->keep_alive_requested != KEEP_ALIVE_NONE) {
1371                 log_keep_alive(ERR,
1372                         "error status info->keep_alive_requested=%d\n",
1373                         info->keep_alive_requested);
1374                 smbd_disconnect_rdma_connection(info);
1375                 return;
1376         }
1377 
1378         log_keep_alive(INFO, "about to send an empty idle message\n");
1379         smbd_post_send_empty(info);
1380 
1381         /* Setup the next idle timeout work */
1382         queue_delayed_work(info->workqueue, &info->idle_timer_work,
1383                         info->keep_alive_interval*HZ);
1384 }
1385 
1386 /*
1387  * Destroy the transport and related RDMA and memory resources
1388  * Need to go through all the pending counters and make sure on one is using
1389  * the transport while it is destroyed
1390  */
1391 void smbd_destroy(struct TCP_Server_Info *server)
1392 {
1393         struct smbd_connection *info = server->smbd_conn;
1394         struct smbd_response *response;
1395         unsigned long flags;
1396 
1397         if (!info) {
1398                 log_rdma_event(INFO, "rdma session already destroyed\n");
1399                 return;
1400         }
1401 
1402         log_rdma_event(INFO, "destroying rdma session\n");
1403         if (info->transport_status != SMBD_DISCONNECTED) {
1404                 rdma_disconnect(server->smbd_conn->id);
1405                 log_rdma_event(INFO, "wait for transport being disconnected\n");
1406                 wait_event_interruptible(
1407                         info->disconn_wait,
1408                         info->transport_status == SMBD_DISCONNECTED);
1409         }
1410 
1411         log_rdma_event(INFO, "destroying qp\n");
1412         ib_drain_qp(info->id->qp);
1413         rdma_destroy_qp(info->id);
1414 
1415         log_rdma_event(INFO, "cancelling idle timer\n");
1416         cancel_delayed_work_sync(&info->idle_timer_work);
1417         log_rdma_event(INFO, "cancelling send immediate work\n");
1418         cancel_delayed_work_sync(&info->send_immediate_work);
1419 
1420         log_rdma_event(INFO, "wait for all send posted to IB to finish\n");
1421         wait_event(info->wait_send_pending,
1422                 atomic_read(&info->send_pending) == 0);
1423         wait_event(info->wait_send_payload_pending,
1424                 atomic_read(&info->send_payload_pending) == 0);
1425 
1426         /* It's not posssible for upper layer to get to reassembly */
1427         log_rdma_event(INFO, "drain the reassembly queue\n");
1428         do {
1429                 spin_lock_irqsave(&info->reassembly_queue_lock, flags);
1430                 response = _get_first_reassembly(info);
1431                 if (response) {
1432                         list_del(&response->list);
1433                         spin_unlock_irqrestore(
1434                                 &info->reassembly_queue_lock, flags);
1435                         put_receive_buffer(info, response);
1436                 } else
1437                         spin_unlock_irqrestore(
1438                                 &info->reassembly_queue_lock, flags);
1439         } while (response);
1440         info->reassembly_data_length = 0;
1441 
1442         log_rdma_event(INFO, "free receive buffers\n");
1443         wait_event(info->wait_receive_queues,
1444                 info->count_receive_queue + info->count_empty_packet_queue
1445                         == info->receive_credit_max);
1446         destroy_receive_buffers(info);
1447 
1448         /*
1449          * For performance reasons, memory registration and deregistration
1450          * are not locked by srv_mutex. It is possible some processes are
1451          * blocked on transport srv_mutex while holding memory registration.
1452          * Release the transport srv_mutex to allow them to hit the failure
1453          * path when sending data, and then release memory registartions.
1454          */
1455         log_rdma_event(INFO, "freeing mr list\n");
1456         wake_up_interruptible_all(&info->wait_mr);
1457         while (atomic_read(&info->mr_used_count)) {
1458                 mutex_unlock(&server->srv_mutex);
1459                 msleep(1000);
1460                 mutex_lock(&server->srv_mutex);
1461         }
1462         destroy_mr_list(info);
1463 
1464         ib_free_cq(info->send_cq);
1465         ib_free_cq(info->recv_cq);
1466         ib_dealloc_pd(info->pd);
1467         rdma_destroy_id(info->id);
1468 
1469         /* free mempools */
1470         mempool_destroy(info->request_mempool);
1471         kmem_cache_destroy(info->request_cache);
1472 
1473         mempool_destroy(info->response_mempool);
1474         kmem_cache_destroy(info->response_cache);
1475 
1476         info->transport_status = SMBD_DESTROYED;
1477 
1478         destroy_workqueue(info->workqueue);
1479         log_rdma_event(INFO,  "rdma session destroyed\n");
1480         kfree(info);
1481 }
1482 
1483 /*
1484  * Reconnect this SMBD connection, called from upper layer
1485  * return value: 0 on success, or actual error code
1486  */
1487 int smbd_reconnect(struct TCP_Server_Info *server)
1488 {
1489         log_rdma_event(INFO, "reconnecting rdma session\n");
1490 
1491         if (!server->smbd_conn) {
1492                 log_rdma_event(INFO, "rdma session already destroyed\n");
1493                 goto create_conn;
1494         }
1495 
1496         /*
1497          * This is possible if transport is disconnected and we haven't received
1498          * notification from RDMA, but upper layer has detected timeout
1499          */
1500         if (server->smbd_conn->transport_status == SMBD_CONNECTED) {
1501                 log_rdma_event(INFO, "disconnecting transport\n");
1502                 smbd_destroy(server);
1503         }
1504 
1505 create_conn:
1506         log_rdma_event(INFO, "creating rdma session\n");
1507         server->smbd_conn = smbd_get_connection(
1508                 server, (struct sockaddr *) &server->dstaddr);
1509 
1510         if (server->smbd_conn)
1511                 cifs_dbg(VFS, "RDMA transport re-established\n");
1512 
1513         return server->smbd_conn ? 0 : -ENOENT;
1514 }
1515 
1516 static void destroy_caches_and_workqueue(struct smbd_connection *info)
1517 {
1518         destroy_receive_buffers(info);
1519         destroy_workqueue(info->workqueue);
1520         mempool_destroy(info->response_mempool);
1521         kmem_cache_destroy(info->response_cache);
1522         mempool_destroy(info->request_mempool);
1523         kmem_cache_destroy(info->request_cache);
1524 }
1525 
1526 #define MAX_NAME_LEN    80
1527 static int allocate_caches_and_workqueue(struct smbd_connection *info)
1528 {
1529         char name[MAX_NAME_LEN];
1530         int rc;
1531 
1532         scnprintf(name, MAX_NAME_LEN, "smbd_request_%p", info);
1533         info->request_cache =
1534                 kmem_cache_create(
1535                         name,
1536                         sizeof(struct smbd_request) +
1537                                 sizeof(struct smbd_data_transfer),
1538                         0, SLAB_HWCACHE_ALIGN, NULL);
1539         if (!info->request_cache)
1540                 return -ENOMEM;
1541 
1542         info->request_mempool =
1543                 mempool_create(info->send_credit_target, mempool_alloc_slab,
1544                         mempool_free_slab, info->request_cache);
1545         if (!info->request_mempool)
1546                 goto out1;
1547 
1548         scnprintf(name, MAX_NAME_LEN, "smbd_response_%p", info);
1549         info->response_cache =
1550                 kmem_cache_create(
1551                         name,
1552                         sizeof(struct smbd_response) +
1553                                 info->max_receive_size,
1554                         0, SLAB_HWCACHE_ALIGN, NULL);
1555         if (!info->response_cache)
1556                 goto out2;
1557 
1558         info->response_mempool =
1559                 mempool_create(info->receive_credit_max, mempool_alloc_slab,
1560                        mempool_free_slab, info->response_cache);
1561         if (!info->response_mempool)
1562                 goto out3;
1563 
1564         scnprintf(name, MAX_NAME_LEN, "smbd_%p", info);
1565         info->workqueue = create_workqueue(name);
1566         if (!info->workqueue)
1567                 goto out4;
1568 
1569         rc = allocate_receive_buffers(info, info->receive_credit_max);
1570         if (rc) {
1571                 log_rdma_event(ERR, "failed to allocate receive buffers\n");
1572                 goto out5;
1573         }
1574 
1575         return 0;
1576 
1577 out5:
1578         destroy_workqueue(info->workqueue);
1579 out4:
1580         mempool_destroy(info->response_mempool);
1581 out3:
1582         kmem_cache_destroy(info->response_cache);
1583 out2:
1584         mempool_destroy(info->request_mempool);
1585 out1:
1586         kmem_cache_destroy(info->request_cache);
1587         return -ENOMEM;
1588 }
1589 
1590 /* Create a SMBD connection, called by upper layer */
1591 static struct smbd_connection *_smbd_get_connection(
1592         struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port)
1593 {
1594         int rc;
1595         struct smbd_connection *info;
1596         struct rdma_conn_param conn_param;
1597         struct ib_qp_init_attr qp_attr;
1598         struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
1599         struct ib_port_immutable port_immutable;
1600         u32 ird_ord_hdr[2];
1601 
1602         info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
1603         if (!info)
1604                 return NULL;
1605 
1606         info->transport_status = SMBD_CONNECTING;
1607         rc = smbd_ia_open(info, dstaddr, port);
1608         if (rc) {
1609                 log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc);
1610                 goto create_id_failed;
1611         }
1612 
1613         if (smbd_send_credit_target > info->id->device->attrs.max_cqe ||
1614             smbd_send_credit_target > info->id->device->attrs.max_qp_wr) {
1615                 log_rdma_event(ERR,
1616                         "consider lowering send_credit_target = %d. "
1617                         "Possible CQE overrun, device "
1618                         "reporting max_cpe %d max_qp_wr %d\n",
1619                         smbd_send_credit_target,
1620                         info->id->device->attrs.max_cqe,
1621                         info->id->device->attrs.max_qp_wr);
1622                 goto config_failed;
1623         }
1624 
1625         if (smbd_receive_credit_max > info->id->device->attrs.max_cqe ||
1626             smbd_receive_credit_max > info->id->device->attrs.max_qp_wr) {
1627                 log_rdma_event(ERR,
1628                         "consider lowering receive_credit_max = %d. "
1629                         "Possible CQE overrun, device "
1630                         "reporting max_cpe %d max_qp_wr %d\n",
1631                         smbd_receive_credit_max,
1632                         info->id->device->attrs.max_cqe,
1633                         info->id->device->attrs.max_qp_wr);
1634                 goto config_failed;
1635         }
1636 
1637         info->receive_credit_max = smbd_receive_credit_max;
1638         info->send_credit_target = smbd_send_credit_target;
1639         info->max_send_size = smbd_max_send_size;
1640         info->max_fragmented_recv_size = smbd_max_fragmented_recv_size;
1641         info->max_receive_size = smbd_max_receive_size;
1642         info->keep_alive_interval = smbd_keep_alive_interval;
1643 
1644         if (info->id->device->attrs.max_send_sge < SMBDIRECT_MAX_SGE) {
1645                 log_rdma_event(ERR,
1646                         "warning: device max_send_sge = %d too small\n",
1647                         info->id->device->attrs.max_send_sge);
1648                 log_rdma_event(ERR, "Queue Pair creation may fail\n");
1649         }
1650         if (info->id->device->attrs.max_recv_sge < SMBDIRECT_MAX_SGE) {
1651                 log_rdma_event(ERR,
1652                         "warning: device max_recv_sge = %d too small\n",
1653                         info->id->device->attrs.max_recv_sge);
1654                 log_rdma_event(ERR, "Queue Pair creation may fail\n");
1655         }
1656 
1657         info->send_cq = NULL;
1658         info->recv_cq = NULL;
1659         info->send_cq =
1660                 ib_alloc_cq_any(info->id->device, info,
1661                                 info->send_credit_target, IB_POLL_SOFTIRQ);
1662         if (IS_ERR(info->send_cq)) {
1663                 info->send_cq = NULL;
1664                 goto alloc_cq_failed;
1665         }
1666 
1667         info->recv_cq =
1668                 ib_alloc_cq_any(info->id->device, info,
1669                                 info->receive_credit_max, IB_POLL_SOFTIRQ);
1670         if (IS_ERR(info->recv_cq)) {
1671                 info->recv_cq = NULL;
1672                 goto alloc_cq_failed;
1673         }
1674 
1675         memset(&qp_attr, 0, sizeof(qp_attr));
1676         qp_attr.event_handler = smbd_qp_async_error_upcall;
1677         qp_attr.qp_context = info;
1678         qp_attr.cap.max_send_wr = info->send_credit_target;
1679         qp_attr.cap.max_recv_wr = info->receive_credit_max;
1680         qp_attr.cap.max_send_sge = SMBDIRECT_MAX_SGE;
1681         qp_attr.cap.max_recv_sge = SMBDIRECT_MAX_SGE;
1682         qp_attr.cap.max_inline_data = 0;
1683         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1684         qp_attr.qp_type = IB_QPT_RC;
1685         qp_attr.send_cq = info->send_cq;
1686         qp_attr.recv_cq = info->recv_cq;
1687         qp_attr.port_num = ~0;
1688 
1689         rc = rdma_create_qp(info->id, info->pd, &qp_attr);
1690         if (rc) {
1691                 log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc);
1692                 goto create_qp_failed;
1693         }
1694 
1695         memset(&conn_param, 0, sizeof(conn_param));
1696         conn_param.initiator_depth = 0;
1697 
1698         conn_param.responder_resources =
1699                 info->id->device->attrs.max_qp_rd_atom
1700                         < SMBD_CM_RESPONDER_RESOURCES ?
1701                 info->id->device->attrs.max_qp_rd_atom :
1702                 SMBD_CM_RESPONDER_RESOURCES;
1703         info->responder_resources = conn_param.responder_resources;
1704         log_rdma_mr(INFO, "responder_resources=%d\n",
1705                 info->responder_resources);
1706 
1707         /* Need to send IRD/ORD in private data for iWARP */
1708         info->id->device->ops.get_port_immutable(
1709                 info->id->device, info->id->port_num, &port_immutable);
1710         if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
1711                 ird_ord_hdr[0] = info->responder_resources;
1712                 ird_ord_hdr[1] = 1;
1713                 conn_param.private_data = ird_ord_hdr;
1714                 conn_param.private_data_len = sizeof(ird_ord_hdr);
1715         } else {
1716                 conn_param.private_data = NULL;
1717                 conn_param.private_data_len = 0;
1718         }
1719 
1720         conn_param.retry_count = SMBD_CM_RETRY;
1721         conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY;
1722         conn_param.flow_control = 0;
1723 
1724         log_rdma_event(INFO, "connecting to IP %pI4 port %d\n",
1725                 &addr_in->sin_addr, port);
1726 
1727         init_waitqueue_head(&info->conn_wait);
1728         init_waitqueue_head(&info->disconn_wait);
1729         init_waitqueue_head(&info->wait_reassembly_queue);
1730         rc = rdma_connect(info->id, &conn_param);
1731         if (rc) {
1732                 log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc);
1733                 goto rdma_connect_failed;
1734         }
1735 
1736         wait_event_interruptible(
1737                 info->conn_wait, info->transport_status != SMBD_CONNECTING);
1738 
1739         if (info->transport_status != SMBD_CONNECTED) {
1740                 log_rdma_event(ERR, "rdma_connect failed port=%d\n", port);
1741                 goto rdma_connect_failed;
1742         }
1743 
1744         log_rdma_event(INFO, "rdma_connect connected\n");
1745 
1746         rc = allocate_caches_and_workqueue(info);
1747         if (rc) {
1748                 log_rdma_event(ERR, "cache allocation failed\n");
1749                 goto allocate_cache_failed;
1750         }
1751 
1752         init_waitqueue_head(&info->wait_send_queue);
1753         INIT_DELAYED_WORK(&info->idle_timer_work, idle_connection_timer);
1754         INIT_DELAYED_WORK(&info->send_immediate_work, send_immediate_work);
1755         queue_delayed_work(info->workqueue, &info->idle_timer_work,
1756                 info->keep_alive_interval*HZ);
1757 
1758         init_waitqueue_head(&info->wait_send_pending);
1759         atomic_set(&info->send_pending, 0);
1760 
1761         init_waitqueue_head(&info->wait_send_payload_pending);
1762         atomic_set(&info->send_payload_pending, 0);
1763 
1764         INIT_WORK(&info->disconnect_work, smbd_disconnect_rdma_work);
1765         INIT_WORK(&info->recv_done_work, smbd_recv_done_work);
1766         INIT_WORK(&info->post_send_credits_work, smbd_post_send_credits);
1767         info->new_credits_offered = 0;
1768         spin_lock_init(&info->lock_new_credits_offered);
1769 
1770         rc = smbd_negotiate(info);
1771         if (rc) {
1772                 log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc);
1773                 goto negotiation_failed;
1774         }
1775 
1776         rc = allocate_mr_list(info);
1777         if (rc) {
1778                 log_rdma_mr(ERR, "memory registration allocation failed\n");
1779                 goto allocate_mr_failed;
1780         }
1781 
1782         return info;
1783 
1784 allocate_mr_failed:
1785         /* At this point, need to a full transport shutdown */
1786         smbd_destroy(server);
1787         return NULL;
1788 
1789 negotiation_failed:
1790         cancel_delayed_work_sync(&info->idle_timer_work);
1791         destroy_caches_and_workqueue(info);
1792         info->transport_status = SMBD_NEGOTIATE_FAILED;
1793         init_waitqueue_head(&info->conn_wait);
1794         rdma_disconnect(info->id);
1795         wait_event(info->conn_wait,
1796                 info->transport_status == SMBD_DISCONNECTED);
1797 
1798 allocate_cache_failed:
1799 rdma_connect_failed:
1800         rdma_destroy_qp(info->id);
1801 
1802 create_qp_failed:
1803 alloc_cq_failed:
1804         if (info->send_cq)
1805                 ib_free_cq(info->send_cq);
1806         if (info->recv_cq)
1807                 ib_free_cq(info->recv_cq);
1808 
1809 config_failed:
1810         ib_dealloc_pd(info->pd);
1811         rdma_destroy_id(info->id);
1812 
1813 create_id_failed:
1814         kfree(info);
1815         return NULL;
1816 }
1817 
1818 struct smbd_connection *smbd_get_connection(
1819         struct TCP_Server_Info *server, struct sockaddr *dstaddr)
1820 {
1821         struct smbd_connection *ret;
1822         int port = SMBD_PORT;
1823 
1824 try_again:
1825         ret = _smbd_get_connection(server, dstaddr, port);
1826 
1827         /* Try SMB_PORT if SMBD_PORT doesn't work */
1828         if (!ret && port == SMBD_PORT) {
1829                 port = SMB_PORT;
1830                 goto try_again;
1831         }
1832         return ret;
1833 }
1834 
1835 /*
1836  * Receive data from receive reassembly queue
1837  * All the incoming data packets are placed in reassembly queue
1838  * buf: the buffer to read data into
1839  * size: the length of data to read
1840  * return value: actual data read
1841  * Note: this implementation copies the data from reassebmly queue to receive
1842  * buffers used by upper layer. This is not the optimal code path. A better way
1843  * to do it is to not have upper layer allocate its receive buffers but rather
1844  * borrow the buffer from reassembly queue, and return it after data is
1845  * consumed. But this will require more changes to upper layer code, and also
1846  * need to consider packet boundaries while they still being reassembled.
1847  */
1848 static int smbd_recv_buf(struct smbd_connection *info, char *buf,
1849                 unsigned int size)
1850 {
1851         struct smbd_response *response;
1852         struct smbd_data_transfer *data_transfer;
1853         int to_copy, to_read, data_read, offset;
1854         u32 data_length, remaining_data_length, data_offset;
1855         int rc;
1856 
1857 again:
1858         /*
1859          * No need to hold the reassembly queue lock all the time as we are
1860          * the only one reading from the front of the queue. The transport
1861          * may add more entries to the back of the queue at the same time
1862          */
1863         log_read(INFO, "size=%d info->reassembly_data_length=%d\n", size,
1864                 info->reassembly_data_length);
1865         if (info->reassembly_data_length >= size) {
1866                 int queue_length;
1867                 int queue_removed = 0;
1868 
1869                 /*
1870                  * Need to make sure reassembly_data_length is read before
1871                  * reading reassembly_queue_length and calling
1872                  * _get_first_reassembly. This call is lock free
1873                  * as we never read at the end of the queue which are being
1874                  * updated in SOFTIRQ as more data is received
1875                  */
1876                 virt_rmb();
1877                 queue_length = info->reassembly_queue_length;
1878                 data_read = 0;
1879                 to_read = size;
1880                 offset = info->first_entry_offset;
1881                 while (data_read < size) {
1882                         response = _get_first_reassembly(info);
1883                         data_transfer = smbd_response_payload(response);
1884                         data_length = le32_to_cpu(data_transfer->data_length);
1885                         remaining_data_length =
1886                                 le32_to_cpu(
1887                                         data_transfer->remaining_data_length);
1888                         data_offset = le32_to_cpu(data_transfer->data_offset);
1889 
1890                         /*
1891                          * The upper layer expects RFC1002 length at the
1892                          * beginning of the payload. Return it to indicate
1893                          * the total length of the packet. This minimize the
1894                          * change to upper layer packet processing logic. This
1895                          * will be eventually remove when an intermediate
1896                          * transport layer is added
1897                          */
1898                         if (response->first_segment && size == 4) {
1899                                 unsigned int rfc1002_len =
1900                                         data_length + remaining_data_length;
1901                                 *((__be32 *)buf) = cpu_to_be32(rfc1002_len);
1902                                 data_read = 4;
1903                                 response->first_segment = false;
1904                                 log_read(INFO, "returning rfc1002 length %d\n",
1905                                         rfc1002_len);
1906                                 goto read_rfc1002_done;
1907                         }
1908 
1909                         to_copy = min_t(int, data_length - offset, to_read);
1910                         memcpy(
1911                                 buf + data_read,
1912                                 (char *)data_transfer + data_offset + offset,
1913                                 to_copy);
1914 
1915                         /* move on to the next buffer? */
1916                         if (to_copy == data_length - offset) {
1917                                 queue_length--;
1918                                 /*
1919                                  * No need to lock if we are not at the
1920                                  * end of the queue
1921                                  */
1922                                 if (queue_length)
1923                                         list_del(&response->list);
1924                                 else {
1925                                         spin_lock_irq(
1926                                                 &info->reassembly_queue_lock);
1927                                         list_del(&response->list);
1928                                         spin_unlock_irq(
1929                                                 &info->reassembly_queue_lock);
1930                                 }
1931                                 queue_removed++;
1932                                 info->count_reassembly_queue--;
1933                                 info->count_dequeue_reassembly_queue++;
1934                                 put_receive_buffer(info, response);
1935                                 offset = 0;
1936                                 log_read(INFO, "put_receive_buffer offset=0\n");
1937                         } else
1938                                 offset += to_copy;
1939 
1940                         to_read -= to_copy;
1941                         data_read += to_copy;
1942 
1943                         log_read(INFO, "_get_first_reassembly memcpy %d bytes "
1944                                 "data_transfer_length-offset=%d after that "
1945                                 "to_read=%d data_read=%d offset=%d\n",
1946                                 to_copy, data_length - offset,
1947                                 to_read, data_read, offset);
1948                 }
1949 
1950                 spin_lock_irq(&info->reassembly_queue_lock);
1951                 info->reassembly_data_length -= data_read;
1952                 info->reassembly_queue_length -= queue_removed;
1953                 spin_unlock_irq(&info->reassembly_queue_lock);
1954 
1955                 info->first_entry_offset = offset;
1956                 log_read(INFO, "returning to thread data_read=%d "
1957                         "reassembly_data_length=%d first_entry_offset=%d\n",
1958                         data_read, info->reassembly_data_length,
1959                         info->first_entry_offset);
1960 read_rfc1002_done:
1961                 return data_read;
1962         }
1963 
1964         log_read(INFO, "wait_event on more data\n");
1965         rc = wait_event_interruptible(
1966                 info->wait_reassembly_queue,
1967                 info->reassembly_data_length >= size ||
1968                         info->transport_status != SMBD_CONNECTED);
1969         /* Don't return any data if interrupted */
1970         if (rc)
1971                 return rc;
1972 
1973         if (info->transport_status != SMBD_CONNECTED) {
1974                 log_read(ERR, "disconnected\n");
1975                 return -ECONNABORTED;
1976         }
1977 
1978         goto again;
1979 }
1980 
1981 /*
1982  * Receive a page from receive reassembly queue
1983  * page: the page to read data into
1984  * to_read: the length of data to read
1985  * return value: actual data read
1986  */
1987 static int smbd_recv_page(struct smbd_connection *info,
1988                 struct page *page, unsigned int page_offset,
1989                 unsigned int to_read)
1990 {
1991         int ret;
1992         char *to_address;
1993         void *page_address;
1994 
1995         /* make sure we have the page ready for read */
1996         ret = wait_event_interruptible(
1997                 info->wait_reassembly_queue,
1998                 info->reassembly_data_length >= to_read ||
1999                         info->transport_status != SMBD_CONNECTED);
2000         if (ret)
2001                 return ret;
2002 
2003         /* now we can read from reassembly queue and not sleep */
2004         page_address = kmap_atomic(page);
2005         to_address = (char *) page_address + page_offset;
2006 
2007         log_read(INFO, "reading from page=%p address=%p to_read=%d\n",
2008                 page, to_address, to_read);
2009 
2010         ret = smbd_recv_buf(info, to_address, to_read);
2011         kunmap_atomic(page_address);
2012 
2013         return ret;
2014 }
2015 
2016 /*
2017  * Receive data from transport
2018  * msg: a msghdr point to the buffer, can be ITER_KVEC or ITER_BVEC
2019  * return: total bytes read, or 0. SMB Direct will not do partial read.
2020  */
2021 int smbd_recv(struct smbd_connection *info, struct msghdr *msg)
2022 {
2023         char *buf;
2024         struct page *page;
2025         unsigned int to_read, page_offset;
2026         int rc;
2027 
2028         if (iov_iter_rw(&msg->msg_iter) == WRITE) {
2029                 /* It's a bug in upper layer to get there */
2030                 cifs_dbg(VFS, "CIFS: invalid msg iter dir %u\n",
2031                          iov_iter_rw(&msg->msg_iter));
2032                 rc = -EINVAL;
2033                 goto out;
2034         }
2035 
2036         switch (iov_iter_type(&msg->msg_iter)) {
2037         case ITER_KVEC:
2038                 buf = msg->msg_iter.kvec->iov_base;
2039                 to_read = msg->msg_iter.kvec->iov_len;
2040                 rc = smbd_recv_buf(info, buf, to_read);
2041                 break;
2042 
2043         case ITER_BVEC:
2044                 page = msg->msg_iter.bvec->bv_page;
2045                 page_offset = msg->msg_iter.bvec->bv_offset;
2046                 to_read = msg->msg_iter.bvec->bv_len;
2047                 rc = smbd_recv_page(info, page, page_offset, to_read);
2048                 break;
2049 
2050         default:
2051                 /* It's a bug in upper layer to get there */
2052                 cifs_dbg(VFS, "CIFS: invalid msg type %d\n",
2053                          iov_iter_type(&msg->msg_iter));
2054                 rc = -EINVAL;
2055         }
2056 
2057 out:
2058         /* SMBDirect will read it all or nothing */
2059         if (rc > 0)
2060                 msg->msg_iter.count = 0;
2061         return rc;
2062 }
2063 
2064 /*
2065  * Send data to transport
2066  * Each rqst is transported as a SMBDirect payload
2067  * rqst: the data to write
2068  * return value: 0 if successfully write, otherwise error code
2069  */
2070 int smbd_send(struct TCP_Server_Info *server,
2071         int num_rqst, struct smb_rqst *rqst_array)
2072 {
2073         struct smbd_connection *info = server->smbd_conn;
2074         struct kvec vec;
2075         int nvecs;
2076         int size;
2077         unsigned int buflen, remaining_data_length;
2078         int start, i, j;
2079         int max_iov_size =
2080                 info->max_send_size - sizeof(struct smbd_data_transfer);
2081         struct kvec *iov;
2082         int rc;
2083         struct smb_rqst *rqst;
2084         int rqst_idx;
2085 
2086         if (info->transport_status != SMBD_CONNECTED) {
2087                 rc = -EAGAIN;
2088                 goto done;
2089         }
2090 
2091         /*
2092          * Add in the page array if there is one. The caller needs to set
2093          * rq_tailsz to PAGE_SIZE when the buffer has multiple pages and
2094          * ends at page boundary
2095          */
2096         remaining_data_length = 0;
2097         for (i = 0; i < num_rqst; i++)
2098                 remaining_data_length += smb_rqst_len(server, &rqst_array[i]);
2099 
2100         if (remaining_data_length + sizeof(struct smbd_data_transfer) >
2101                 info->max_fragmented_send_size) {
2102                 log_write(ERR, "payload size %d > max size %d\n",
2103                         remaining_data_length, info->max_fragmented_send_size);
2104                 rc = -EINVAL;
2105                 goto done;
2106         }
2107 
2108         log_write(INFO, "num_rqst=%d total length=%u\n",
2109                         num_rqst, remaining_data_length);
2110 
2111         rqst_idx = 0;
2112 next_rqst:
2113         rqst = &rqst_array[rqst_idx];
2114         iov = rqst->rq_iov;
2115 
2116         cifs_dbg(FYI, "Sending smb (RDMA): idx=%d smb_len=%lu\n",
2117                 rqst_idx, smb_rqst_len(server, rqst));
2118         for (i = 0; i < rqst->rq_nvec; i++)
2119                 dump_smb(iov[i].iov_base, iov[i].iov_len);
2120 
2121 
2122         log_write(INFO, "rqst_idx=%d nvec=%d rqst->rq_npages=%d rq_pagesz=%d "
2123                 "rq_tailsz=%d buflen=%lu\n",
2124                 rqst_idx, rqst->rq_nvec, rqst->rq_npages, rqst->rq_pagesz,
2125                 rqst->rq_tailsz, smb_rqst_len(server, rqst));
2126 
2127         start = i = 0;
2128         buflen = 0;
2129         while (true) {
2130                 buflen += iov[i].iov_len;
2131                 if (buflen > max_iov_size) {
2132                         if (i > start) {
2133                                 remaining_data_length -=
2134                                         (buflen-iov[i].iov_len);
2135                                 log_write(INFO, "sending iov[] from start=%d "
2136                                         "i=%d nvecs=%d "
2137                                         "remaining_data_length=%d\n",
2138                                         start, i, i-start,
2139                                         remaining_data_length);
2140                                 rc = smbd_post_send_data(
2141                                         info, &iov[start], i-start,
2142                                         remaining_data_length);
2143                                 if (rc)
2144                                         goto done;
2145                         } else {
2146                                 /* iov[start] is too big, break it */
2147                                 nvecs = (buflen+max_iov_size-1)/max_iov_size;
2148                                 log_write(INFO, "iov[%d] iov_base=%p buflen=%d"
2149                                         " break to %d vectors\n",
2150                                         start, iov[start].iov_base,
2151                                         buflen, nvecs);
2152                                 for (j = 0; j < nvecs; j++) {
2153                                         vec.iov_base =
2154                                                 (char *)iov[start].iov_base +
2155                                                 j*max_iov_size;
2156                                         vec.iov_len = max_iov_size;
2157                                         if (j == nvecs-1)
2158                                                 vec.iov_len =
2159                                                         buflen -
2160                                                         max_iov_size*(nvecs-1);
2161                                         remaining_data_length -= vec.iov_len;
2162                                         log_write(INFO,
2163                                                 "sending vec j=%d iov_base=%p"
2164                                                 " iov_len=%zu "
2165                                                 "remaining_data_length=%d\n",
2166                                                 j, vec.iov_base, vec.iov_len,
2167                                                 remaining_data_length);
2168                                         rc = smbd_post_send_data(
2169                                                 info, &vec, 1,
2170                                                 remaining_data_length);
2171                                         if (rc)
2172                                                 goto done;
2173                                 }
2174                                 i++;
2175                                 if (i == rqst->rq_nvec)
2176                                         break;
2177                         }
2178                         start = i;
2179                         buflen = 0;
2180                 } else {
2181                         i++;
2182                         if (i == rqst->rq_nvec) {
2183                                 /* send out all remaining vecs */
2184                                 remaining_data_length -= buflen;
2185                                 log_write(INFO,
2186                                         "sending iov[] from start=%d i=%d "
2187                                         "nvecs=%d remaining_data_length=%d\n",
2188                                         start, i, i-start,
2189                                         remaining_data_length);
2190                                 rc = smbd_post_send_data(info, &iov[start],
2191                                         i-start, remaining_data_length);
2192                                 if (rc)
2193                                         goto done;
2194                                 break;
2195                         }
2196                 }
2197                 log_write(INFO, "looping i=%d buflen=%d\n", i, buflen);
2198         }
2199 
2200         /* now sending pages if there are any */
2201         for (i = 0; i < rqst->rq_npages; i++) {
2202                 unsigned int offset;
2203 
2204                 rqst_page_get_length(rqst, i, &buflen, &offset);
2205                 nvecs = (buflen + max_iov_size - 1) / max_iov_size;
2206                 log_write(INFO, "sending pages buflen=%d nvecs=%d\n",
2207                         buflen, nvecs);
2208                 for (j = 0; j < nvecs; j++) {
2209                         size = max_iov_size;
2210                         if (j == nvecs-1)
2211                                 size = buflen - j*max_iov_size;
2212                         remaining_data_length -= size;
2213                         log_write(INFO, "sending pages i=%d offset=%d size=%d"
2214                                 " remaining_data_length=%d\n",
2215                                 i, j*max_iov_size+offset, size,
2216                                 remaining_data_length);
2217                         rc = smbd_post_send_page(
2218                                 info, rqst->rq_pages[i],
2219                                 j*max_iov_size + offset,
2220                                 size, remaining_data_length);
2221                         if (rc)
2222                                 goto done;
2223                 }
2224         }
2225 
2226         rqst_idx++;
2227         if (rqst_idx < num_rqst)
2228                 goto next_rqst;
2229 
2230 done:
2231         /*
2232          * As an optimization, we don't wait for individual I/O to finish
2233          * before sending the next one.
2234          * Send them all and wait for pending send count to get to 0
2235          * that means all the I/Os have been out and we are good to return
2236          */
2237 
2238         wait_event(info->wait_send_payload_pending,
2239                 atomic_read(&info->send_payload_pending) == 0);
2240 
2241         return rc;
2242 }
2243 
2244 static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc)
2245 {
2246         struct smbd_mr *mr;
2247         struct ib_cqe *cqe;
2248 
2249         if (wc->status) {
2250                 log_rdma_mr(ERR, "status=%d\n", wc->status);
2251                 cqe = wc->wr_cqe;
2252                 mr = container_of(cqe, struct smbd_mr, cqe);
2253                 smbd_disconnect_rdma_connection(mr->conn);
2254         }
2255 }
2256 
2257 /*
2258  * The work queue function that recovers MRs
2259  * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used
2260  * again. Both calls are slow, so finish them in a workqueue. This will not
2261  * block I/O path.
2262  * There is one workqueue that recovers MRs, there is no need to lock as the
2263  * I/O requests calling smbd_register_mr will never update the links in the
2264  * mr_list.
2265  */
2266 static void smbd_mr_recovery_work(struct work_struct *work)
2267 {
2268         struct smbd_connection *info =
2269                 container_of(work, struct smbd_connection, mr_recovery_work);
2270         struct smbd_mr *smbdirect_mr;
2271         int rc;
2272 
2273         list_for_each_entry(smbdirect_mr, &info->mr_list, list) {
2274                 if (smbdirect_mr->state == MR_ERROR) {
2275 
2276                         /* recover this MR entry */
2277                         rc = ib_dereg_mr(smbdirect_mr->mr);
2278                         if (rc) {
2279                                 log_rdma_mr(ERR,
2280                                         "ib_dereg_mr failed rc=%x\n",
2281                                         rc);
2282                                 smbd_disconnect_rdma_connection(info);
2283                                 continue;
2284                         }
2285 
2286                         smbdirect_mr->mr = ib_alloc_mr(
2287                                 info->pd, info->mr_type,
2288                                 info->max_frmr_depth);
2289                         if (IS_ERR(smbdirect_mr->mr)) {
2290                                 log_rdma_mr(ERR,
2291                                         "ib_alloc_mr failed mr_type=%x "
2292                                         "max_frmr_depth=%x\n",
2293                                         info->mr_type,
2294                                         info->max_frmr_depth);
2295                                 smbd_disconnect_rdma_connection(info);
2296                                 continue;
2297                         }
2298                 } else
2299                         /* This MR is being used, don't recover it */
2300                         continue;
2301 
2302                 smbdirect_mr->state = MR_READY;
2303 
2304                 /* smbdirect_mr->state is updated by this function
2305                  * and is read and updated by I/O issuing CPUs trying
2306                  * to get a MR, the call to atomic_inc_return
2307                  * implicates a memory barrier and guarantees this
2308                  * value is updated before waking up any calls to
2309                  * get_mr() from the I/O issuing CPUs
2310                  */
2311                 if (atomic_inc_return(&info->mr_ready_count) == 1)
2312                         wake_up_interruptible(&info->wait_mr);
2313         }
2314 }
2315 
2316 static void destroy_mr_list(struct smbd_connection *info)
2317 {
2318         struct smbd_mr *mr, *tmp;
2319 
2320         cancel_work_sync(&info->mr_recovery_work);
2321         list_for_each_entry_safe(mr, tmp, &info->mr_list, list) {
2322                 if (mr->state == MR_INVALIDATED)
2323                         ib_dma_unmap_sg(info->id->device, mr->sgl,
2324                                 mr->sgl_count, mr->dir);
2325                 ib_dereg_mr(mr->mr);
2326                 kfree(mr->sgl);
2327                 kfree(mr);
2328         }
2329 }
2330 
2331 /*
2332  * Allocate MRs used for RDMA read/write
2333  * The number of MRs will not exceed hardware capability in responder_resources
2334  * All MRs are kept in mr_list. The MR can be recovered after it's used
2335  * Recovery is done in smbd_mr_recovery_work. The content of list entry changes
2336  * as MRs are used and recovered for I/O, but the list links will not change
2337  */
2338 static int allocate_mr_list(struct smbd_connection *info)
2339 {
2340         int i;
2341         struct smbd_mr *smbdirect_mr, *tmp;
2342 
2343         INIT_LIST_HEAD(&info->mr_list);
2344         init_waitqueue_head(&info->wait_mr);
2345         spin_lock_init(&info->mr_list_lock);
2346         atomic_set(&info->mr_ready_count, 0);
2347         atomic_set(&info->mr_used_count, 0);
2348         init_waitqueue_head(&info->wait_for_mr_cleanup);
2349         /* Allocate more MRs (2x) than hardware responder_resources */
2350         for (i = 0; i < info->responder_resources * 2; i++) {
2351                 smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
2352                 if (!smbdirect_mr)
2353                         goto out;
2354                 smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type,
2355                                         info->max_frmr_depth);
2356                 if (IS_ERR(smbdirect_mr->mr)) {
2357                         log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x "
2358                                 "max_frmr_depth=%x\n",
2359                                 info->mr_type, info->max_frmr_depth);
2360                         goto out;
2361                 }
2362                 smbdirect_mr->sgl = kcalloc(
2363                                         info->max_frmr_depth,
2364                                         sizeof(struct scatterlist),
2365                                         GFP_KERNEL);
2366                 if (!smbdirect_mr->sgl) {
2367                         log_rdma_mr(ERR, "failed to allocate sgl\n");
2368                         ib_dereg_mr(smbdirect_mr->mr);
2369                         goto out;
2370                 }
2371                 smbdirect_mr->state = MR_READY;
2372                 smbdirect_mr->conn = info;
2373 
2374                 list_add_tail(&smbdirect_mr->list, &info->mr_list);
2375                 atomic_inc(&info->mr_ready_count);
2376         }
2377         INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
2378         return 0;
2379 
2380 out:
2381         kfree(smbdirect_mr);
2382 
2383         list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) {
2384                 ib_dereg_mr(smbdirect_mr->mr);
2385                 kfree(smbdirect_mr->sgl);
2386                 kfree(smbdirect_mr);
2387         }
2388         return -ENOMEM;
2389 }
2390 
2391 /*
2392  * Get a MR from mr_list. This function waits until there is at least one
2393  * MR available in the list. It may access the list while the
2394  * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock
2395  * as they never modify the same places. However, there may be several CPUs
2396  * issueing I/O trying to get MR at the same time, mr_list_lock is used to
2397  * protect this situation.
2398  */
2399 static struct smbd_mr *get_mr(struct smbd_connection *info)
2400 {
2401         struct smbd_mr *ret;
2402         int rc;
2403 again:
2404         rc = wait_event_interruptible(info->wait_mr,
2405                 atomic_read(&info->mr_ready_count) ||
2406                 info->transport_status != SMBD_CONNECTED);
2407         if (rc) {
2408                 log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc);
2409                 return NULL;
2410         }
2411 
2412         if (info->transport_status != SMBD_CONNECTED) {
2413                 log_rdma_mr(ERR, "info->transport_status=%x\n",
2414                         info->transport_status);
2415                 return NULL;
2416         }
2417 
2418         spin_lock(&info->mr_list_lock);
2419         list_for_each_entry(ret, &info->mr_list, list) {
2420                 if (ret->state == MR_READY) {
2421                         ret->state = MR_REGISTERED;
2422                         spin_unlock(&info->mr_list_lock);
2423                         atomic_dec(&info->mr_ready_count);
2424                         atomic_inc(&info->mr_used_count);
2425                         return ret;
2426                 }
2427         }
2428 
2429         spin_unlock(&info->mr_list_lock);
2430         /*
2431          * It is possible that we could fail to get MR because other processes may
2432          * try to acquire a MR at the same time. If this is the case, retry it.
2433          */
2434         goto again;
2435 }
2436 
2437 /*
2438  * Register memory for RDMA read/write
2439  * pages[]: the list of pages to register memory with
2440  * num_pages: the number of pages to register
2441  * tailsz: if non-zero, the bytes to register in the last page
2442  * writing: true if this is a RDMA write (SMB read), false for RDMA read
2443  * need_invalidate: true if this MR needs to be locally invalidated after I/O
2444  * return value: the MR registered, NULL if failed.
2445  */
2446 struct smbd_mr *smbd_register_mr(
2447         struct smbd_connection *info, struct page *pages[], int num_pages,
2448         int offset, int tailsz, bool writing, bool need_invalidate)
2449 {
2450         struct smbd_mr *smbdirect_mr;
2451         int rc, i;
2452         enum dma_data_direction dir;
2453         struct ib_reg_wr *reg_wr;
2454 
2455         if (num_pages > info->max_frmr_depth) {
2456                 log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n",
2457                         num_pages, info->max_frmr_depth);
2458                 return NULL;
2459         }
2460 
2461         smbdirect_mr = get_mr(info);
2462         if (!smbdirect_mr) {
2463                 log_rdma_mr(ERR, "get_mr returning NULL\n");
2464                 return NULL;
2465         }
2466         smbdirect_mr->need_invalidate = need_invalidate;
2467         smbdirect_mr->sgl_count = num_pages;
2468         sg_init_table(smbdirect_mr->sgl, num_pages);
2469 
2470         log_rdma_mr(INFO, "num_pages=0x%x offset=0x%x tailsz=0x%x\n",
2471                         num_pages, offset, tailsz);
2472 
2473         if (num_pages == 1) {
2474                 sg_set_page(&smbdirect_mr->sgl[0], pages[0], tailsz, offset);
2475                 goto skip_multiple_pages;
2476         }
2477 
2478         /* We have at least two pages to register */
2479         sg_set_page(
2480                 &smbdirect_mr->sgl[0], pages[0], PAGE_SIZE - offset, offset);
2481         i = 1;
2482         while (i < num_pages - 1) {
2483                 sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0);
2484                 i++;
2485         }
2486         sg_set_page(&smbdirect_mr->sgl[i], pages[i],
2487                 tailsz ? tailsz : PAGE_SIZE, 0);
2488 
2489 skip_multiple_pages:
2490         dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2491         smbdirect_mr->dir = dir;
2492         rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir);
2493         if (!rc) {
2494                 log_rdma_mr(ERR, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n",
2495                         num_pages, dir, rc);
2496                 goto dma_map_error;
2497         }
2498 
2499         rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages,
2500                 NULL, PAGE_SIZE);
2501         if (rc != num_pages) {
2502                 log_rdma_mr(ERR,
2503                         "ib_map_mr_sg failed rc = %d num_pages = %x\n",
2504                         rc, num_pages);
2505                 goto map_mr_error;
2506         }
2507 
2508         ib_update_fast_reg_key(smbdirect_mr->mr,
2509                 ib_inc_rkey(smbdirect_mr->mr->rkey));
2510         reg_wr = &smbdirect_mr->wr;
2511         reg_wr->wr.opcode = IB_WR_REG_MR;
2512         smbdirect_mr->cqe.done = register_mr_done;
2513         reg_wr->wr.wr_cqe = &smbdirect_mr->cqe;
2514         reg_wr->wr.num_sge = 0;
2515         reg_wr->wr.send_flags = IB_SEND_SIGNALED;
2516         reg_wr->mr = smbdirect_mr->mr;
2517         reg_wr->key = smbdirect_mr->mr->rkey;
2518         reg_wr->access = writing ?
2519                         IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2520                         IB_ACCESS_REMOTE_READ;
2521 
2522         /*
2523          * There is no need for waiting for complemtion on ib_post_send
2524          * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution
2525          * on the next ib_post_send when we actaully send I/O to remote peer
2526          */
2527         rc = ib_post_send(info->id->qp, &reg_wr->wr, NULL);
2528         if (!rc)
2529                 return smbdirect_mr;
2530 
2531         log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n",
2532                 rc, reg_wr->key);
2533 
2534         /* If all failed, attempt to recover this MR by setting it MR_ERROR*/
2535 map_mr_error:
2536         ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl,
2537                 smbdirect_mr->sgl_count, smbdirect_mr->dir);
2538 
2539 dma_map_error:
2540         smbdirect_mr->state = MR_ERROR;
2541         if (atomic_dec_and_test(&info->mr_used_count))
2542                 wake_up(&info->wait_for_mr_cleanup);
2543 
2544         smbd_disconnect_rdma_connection(info);
2545 
2546         return NULL;
2547 }
2548 
2549 static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc)
2550 {
2551         struct smbd_mr *smbdirect_mr;
2552         struct ib_cqe *cqe;
2553 
2554         cqe = wc->wr_cqe;
2555         smbdirect_mr = container_of(cqe, struct smbd_mr, cqe);
2556         smbdirect_mr->state = MR_INVALIDATED;
2557         if (wc->status != IB_WC_SUCCESS) {
2558                 log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status);
2559                 smbdirect_mr->state = MR_ERROR;
2560         }
2561         complete(&smbdirect_mr->invalidate_done);
2562 }
2563 
2564 /*
2565  * Deregister a MR after I/O is done
2566  * This function may wait if remote invalidation is not used
2567  * and we have to locally invalidate the buffer to prevent data is being
2568  * modified by remote peer after upper layer consumes it
2569  */
2570 int smbd_deregister_mr(struct smbd_mr *smbdirect_mr)
2571 {
2572         struct ib_send_wr *wr;
2573         struct smbd_connection *info = smbdirect_mr->conn;
2574         int rc = 0;
2575 
2576         if (smbdirect_mr->need_invalidate) {
2577                 /* Need to finish local invalidation before returning */
2578                 wr = &smbdirect_mr->inv_wr;
2579                 wr->opcode = IB_WR_LOCAL_INV;
2580                 smbdirect_mr->cqe.done = local_inv_done;
2581                 wr->wr_cqe = &smbdirect_mr->cqe;
2582                 wr->num_sge = 0;
2583                 wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey;
2584                 wr->send_flags = IB_SEND_SIGNALED;
2585 
2586                 init_completion(&smbdirect_mr->invalidate_done);
2587                 rc = ib_post_send(info->id->qp, wr, NULL);
2588                 if (rc) {
2589                         log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc);
2590                         smbd_disconnect_rdma_connection(info);
2591                         goto done;
2592                 }
2593                 wait_for_completion(&smbdirect_mr->invalidate_done);
2594                 smbdirect_mr->need_invalidate = false;
2595         } else
2596                 /*
2597                  * For remote invalidation, just set it to MR_INVALIDATED
2598                  * and defer to mr_recovery_work to recover the MR for next use
2599                  */
2600                 smbdirect_mr->state = MR_INVALIDATED;
2601 
2602         if (smbdirect_mr->state == MR_INVALIDATED) {
2603                 ib_dma_unmap_sg(
2604                         info->id->device, smbdirect_mr->sgl,
2605                         smbdirect_mr->sgl_count,
2606                         smbdirect_mr->dir);
2607                 smbdirect_mr->state = MR_READY;
2608                 if (atomic_inc_return(&info->mr_ready_count) == 1)
2609                         wake_up_interruptible(&info->wait_mr);
2610         } else
2611                 /*
2612                  * Schedule the work to do MR recovery for future I/Os MR
2613                  * recovery is slow and don't want it to block current I/O
2614                  */
2615                 queue_work(info->workqueue, &info->mr_recovery_work);
2616 
2617 done:
2618         if (atomic_dec_and_test(&info->mr_used_count))
2619                 wake_up(&info->wait_for_mr_cleanup);
2620 
2621         return rc;
2622 }