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 <linux/folio_queue.h> 10 #include "../common/smbdirect/smbdirect_pdu.h" 11 #include "smbdirect.h" 12 #include "cifs_debug.h" 13 #include "cifsproto.h" 14 #include "smb2proto.h" 15 16 const struct smbdirect_socket_parameters *smbd_get_parameters(struct smbd_connection *conn) 17 { 18 struct smbdirect_socket *sc = &conn->socket; 19 20 return &sc->parameters; 21 } 22 23 static struct smbdirect_recv_io *get_receive_buffer( 24 struct smbdirect_socket *sc); 25 static void put_receive_buffer( 26 struct smbdirect_socket *sc, 27 struct smbdirect_recv_io *response); 28 static int allocate_receive_buffers(struct smbdirect_socket *sc, int num_buf); 29 static void destroy_receive_buffers(struct smbdirect_socket *sc); 30 31 static void enqueue_reassembly( 32 struct smbdirect_socket *sc, 33 struct smbdirect_recv_io *response, int data_length); 34 static struct smbdirect_recv_io *_get_first_reassembly( 35 struct smbdirect_socket *sc); 36 37 static int smbd_post_recv( 38 struct smbdirect_socket *sc, 39 struct smbdirect_recv_io *response); 40 41 static int smbd_post_send_empty(struct smbdirect_socket *sc); 42 43 static void destroy_mr_list(struct smbdirect_socket *sc); 44 static int allocate_mr_list(struct smbdirect_socket *sc); 45 46 struct smb_extract_to_rdma { 47 struct ib_sge *sge; 48 unsigned int nr_sge; 49 unsigned int max_sge; 50 struct ib_device *device; 51 u32 local_dma_lkey; 52 enum dma_data_direction direction; 53 }; 54 static ssize_t smb_extract_iter_to_rdma(struct iov_iter *iter, size_t len, 55 struct smb_extract_to_rdma *rdma); 56 57 /* Port numbers for SMBD transport */ 58 #define SMB_PORT 445 59 #define SMBD_PORT 5445 60 61 /* Address lookup and resolve timeout in ms */ 62 #define RDMA_RESOLVE_TIMEOUT 5000 63 64 /* SMBD negotiation timeout in seconds */ 65 #define SMBD_NEGOTIATE_TIMEOUT 120 66 67 /* The timeout to wait for a keepalive message from peer in seconds */ 68 #define KEEPALIVE_RECV_TIMEOUT 5 69 70 /* SMBD minimum receive size and fragmented sized defined in [MS-SMBD] */ 71 #define SMBD_MIN_RECEIVE_SIZE 128 72 #define SMBD_MIN_FRAGMENTED_SIZE 131072 73 74 /* 75 * Default maximum number of RDMA read/write outstanding on this connection 76 * This value is possibly decreased during QP creation on hardware limit 77 */ 78 #define SMBD_CM_RESPONDER_RESOURCES 32 79 80 /* Maximum number of retries on data transfer operations */ 81 #define SMBD_CM_RETRY 6 82 /* No need to retry on Receiver Not Ready since SMBD manages credits */ 83 #define SMBD_CM_RNR_RETRY 0 84 85 /* 86 * User configurable initial values per SMBD transport connection 87 * as defined in [MS-SMBD] 3.1.1.1 88 * Those may change after a SMBD negotiation 89 */ 90 /* The local peer's maximum number of credits to grant to the peer */ 91 int smbd_receive_credit_max = 255; 92 93 /* The remote peer's credit request of local peer */ 94 int smbd_send_credit_target = 255; 95 96 /* The maximum single message size can be sent to remote peer */ 97 int smbd_max_send_size = 1364; 98 99 /* The maximum fragmented upper-layer payload receive size supported */ 100 int smbd_max_fragmented_recv_size = 1024 * 1024; 101 102 /* The maximum single-message size which can be received */ 103 int smbd_max_receive_size = 1364; 104 105 /* The timeout to initiate send of a keepalive message on idle */ 106 int smbd_keep_alive_interval = 120; 107 108 /* 109 * User configurable initial values for RDMA transport 110 * The actual values used may be lower and are limited to hardware capabilities 111 */ 112 /* Default maximum number of pages in a single RDMA write/read */ 113 int smbd_max_frmr_depth = 2048; 114 115 /* If payload is less than this byte, use RDMA send/recv not read/write */ 116 int rdma_readwrite_threshold = 4096; 117 118 /* Transport logging functions 119 * Logging are defined as classes. They can be OR'ed to define the actual 120 * logging level via module parameter smbd_logging_class 121 * e.g. cifs.smbd_logging_class=0xa0 will log all log_rdma_recv() and 122 * log_rdma_event() 123 */ 124 #define LOG_OUTGOING 0x1 125 #define LOG_INCOMING 0x2 126 #define LOG_READ 0x4 127 #define LOG_WRITE 0x8 128 #define LOG_RDMA_SEND 0x10 129 #define LOG_RDMA_RECV 0x20 130 #define LOG_KEEP_ALIVE 0x40 131 #define LOG_RDMA_EVENT 0x80 132 #define LOG_RDMA_MR 0x100 133 static unsigned int smbd_logging_class; 134 module_param(smbd_logging_class, uint, 0644); 135 MODULE_PARM_DESC(smbd_logging_class, 136 "Logging class for SMBD transport 0x0 to 0x100"); 137 138 #define ERR 0x0 139 #define INFO 0x1 140 static unsigned int smbd_logging_level = ERR; 141 module_param(smbd_logging_level, uint, 0644); 142 MODULE_PARM_DESC(smbd_logging_level, 143 "Logging level for SMBD transport, 0 (default): error, 1: info"); 144 145 #define log_rdma(level, class, fmt, args...) \ 146 do { \ 147 if (level <= smbd_logging_level || class & smbd_logging_class) \ 148 cifs_dbg(VFS, "%s:%d " fmt, __func__, __LINE__, ##args);\ 149 } while (0) 150 151 #define log_outgoing(level, fmt, args...) \ 152 log_rdma(level, LOG_OUTGOING, fmt, ##args) 153 #define log_incoming(level, fmt, args...) \ 154 log_rdma(level, LOG_INCOMING, fmt, ##args) 155 #define log_read(level, fmt, args...) log_rdma(level, LOG_READ, fmt, ##args) 156 #define log_write(level, fmt, args...) log_rdma(level, LOG_WRITE, fmt, ##args) 157 #define log_rdma_send(level, fmt, args...) \ 158 log_rdma(level, LOG_RDMA_SEND, fmt, ##args) 159 #define log_rdma_recv(level, fmt, args...) \ 160 log_rdma(level, LOG_RDMA_RECV, fmt, ##args) 161 #define log_keep_alive(level, fmt, args...) \ 162 log_rdma(level, LOG_KEEP_ALIVE, fmt, ##args) 163 #define log_rdma_event(level, fmt, args...) \ 164 log_rdma(level, LOG_RDMA_EVENT, fmt, ##args) 165 #define log_rdma_mr(level, fmt, args...) \ 166 log_rdma(level, LOG_RDMA_MR, fmt, ##args) 167 168 static void smbd_disconnect_wake_up_all(struct smbdirect_socket *sc) 169 { 170 /* 171 * Wake up all waiters in all wait queues 172 * in order to notice the broken connection. 173 */ 174 wake_up_all(&sc->status_wait); 175 wake_up_all(&sc->send_io.credits.wait_queue); 176 wake_up_all(&sc->send_io.pending.dec_wait_queue); 177 wake_up_all(&sc->send_io.pending.zero_wait_queue); 178 wake_up_all(&sc->recv_io.reassembly.wait_queue); 179 wake_up_all(&sc->mr_io.ready.wait_queue); 180 wake_up_all(&sc->mr_io.cleanup.wait_queue); 181 } 182 183 static void smbd_disconnect_rdma_work(struct work_struct *work) 184 { 185 struct smbdirect_socket *sc = 186 container_of(work, struct smbdirect_socket, disconnect_work); 187 188 /* 189 * make sure this and other work is not queued again 190 * but here we don't block and avoid 191 * disable[_delayed]_work_sync() 192 */ 193 disable_work(&sc->disconnect_work); 194 disable_work(&sc->recv_io.posted.refill_work); 195 disable_work(&sc->mr_io.recovery_work); 196 disable_work(&sc->idle.immediate_work); 197 disable_delayed_work(&sc->idle.timer_work); 198 199 if (sc->first_error == 0) 200 sc->first_error = -ECONNABORTED; 201 202 switch (sc->status) { 203 case SMBDIRECT_SOCKET_NEGOTIATE_NEEDED: 204 case SMBDIRECT_SOCKET_NEGOTIATE_RUNNING: 205 case SMBDIRECT_SOCKET_NEGOTIATE_FAILED: 206 case SMBDIRECT_SOCKET_CONNECTED: 207 case SMBDIRECT_SOCKET_ERROR: 208 sc->status = SMBDIRECT_SOCKET_DISCONNECTING; 209 rdma_disconnect(sc->rdma.cm_id); 210 break; 211 212 case SMBDIRECT_SOCKET_CREATED: 213 case SMBDIRECT_SOCKET_RESOLVE_ADDR_NEEDED: 214 case SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING: 215 case SMBDIRECT_SOCKET_RESOLVE_ADDR_FAILED: 216 case SMBDIRECT_SOCKET_RESOLVE_ROUTE_NEEDED: 217 case SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING: 218 case SMBDIRECT_SOCKET_RESOLVE_ROUTE_FAILED: 219 case SMBDIRECT_SOCKET_RDMA_CONNECT_NEEDED: 220 case SMBDIRECT_SOCKET_RDMA_CONNECT_RUNNING: 221 case SMBDIRECT_SOCKET_RDMA_CONNECT_FAILED: 222 /* 223 * rdma_connect() never reached 224 * RDMA_CM_EVENT_ESTABLISHED 225 */ 226 sc->status = SMBDIRECT_SOCKET_DISCONNECTED; 227 break; 228 229 case SMBDIRECT_SOCKET_DISCONNECTING: 230 case SMBDIRECT_SOCKET_DISCONNECTED: 231 case SMBDIRECT_SOCKET_DESTROYED: 232 break; 233 } 234 235 /* 236 * Wake up all waiters in all wait queues 237 * in order to notice the broken connection. 238 */ 239 smbd_disconnect_wake_up_all(sc); 240 } 241 242 static void smbd_disconnect_rdma_connection(struct smbdirect_socket *sc) 243 { 244 /* 245 * make sure other work (than disconnect_work) is 246 * not queued again but here we don't block and avoid 247 * disable[_delayed]_work_sync() 248 */ 249 disable_work(&sc->recv_io.posted.refill_work); 250 disable_work(&sc->mr_io.recovery_work); 251 disable_work(&sc->idle.immediate_work); 252 disable_delayed_work(&sc->idle.timer_work); 253 254 if (sc->first_error == 0) 255 sc->first_error = -ECONNABORTED; 256 257 switch (sc->status) { 258 case SMBDIRECT_SOCKET_RESOLVE_ADDR_FAILED: 259 case SMBDIRECT_SOCKET_RESOLVE_ROUTE_FAILED: 260 case SMBDIRECT_SOCKET_RDMA_CONNECT_FAILED: 261 case SMBDIRECT_SOCKET_NEGOTIATE_FAILED: 262 case SMBDIRECT_SOCKET_ERROR: 263 case SMBDIRECT_SOCKET_DISCONNECTING: 264 case SMBDIRECT_SOCKET_DISCONNECTED: 265 case SMBDIRECT_SOCKET_DESTROYED: 266 /* 267 * Keep the current error status 268 */ 269 break; 270 271 case SMBDIRECT_SOCKET_RESOLVE_ADDR_NEEDED: 272 case SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING: 273 sc->status = SMBDIRECT_SOCKET_RESOLVE_ADDR_FAILED; 274 break; 275 276 case SMBDIRECT_SOCKET_RESOLVE_ROUTE_NEEDED: 277 case SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING: 278 sc->status = SMBDIRECT_SOCKET_RESOLVE_ROUTE_FAILED; 279 break; 280 281 case SMBDIRECT_SOCKET_RDMA_CONNECT_NEEDED: 282 case SMBDIRECT_SOCKET_RDMA_CONNECT_RUNNING: 283 sc->status = SMBDIRECT_SOCKET_RDMA_CONNECT_FAILED; 284 break; 285 286 case SMBDIRECT_SOCKET_NEGOTIATE_NEEDED: 287 case SMBDIRECT_SOCKET_NEGOTIATE_RUNNING: 288 sc->status = SMBDIRECT_SOCKET_NEGOTIATE_FAILED; 289 break; 290 291 case SMBDIRECT_SOCKET_CREATED: 292 case SMBDIRECT_SOCKET_CONNECTED: 293 sc->status = SMBDIRECT_SOCKET_ERROR; 294 break; 295 } 296 297 /* 298 * Wake up all waiters in all wait queues 299 * in order to notice the broken connection. 300 */ 301 smbd_disconnect_wake_up_all(sc); 302 303 queue_work(sc->workqueue, &sc->disconnect_work); 304 } 305 306 /* Upcall from RDMA CM */ 307 static int smbd_conn_upcall( 308 struct rdma_cm_id *id, struct rdma_cm_event *event) 309 { 310 struct smbdirect_socket *sc = id->context; 311 struct smbdirect_socket_parameters *sp = &sc->parameters; 312 const char *event_name = rdma_event_msg(event->event); 313 u8 peer_initiator_depth; 314 u8 peer_responder_resources; 315 316 log_rdma_event(INFO, "event=%s status=%d\n", 317 event_name, event->status); 318 319 switch (event->event) { 320 case RDMA_CM_EVENT_ADDR_RESOLVED: 321 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING); 322 sc->status = SMBDIRECT_SOCKET_RESOLVE_ROUTE_NEEDED; 323 wake_up(&sc->status_wait); 324 break; 325 326 case RDMA_CM_EVENT_ROUTE_RESOLVED: 327 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING); 328 sc->status = SMBDIRECT_SOCKET_RDMA_CONNECT_NEEDED; 329 wake_up(&sc->status_wait); 330 break; 331 332 case RDMA_CM_EVENT_ADDR_ERROR: 333 log_rdma_event(ERR, "connecting failed event=%s\n", event_name); 334 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING); 335 sc->status = SMBDIRECT_SOCKET_RESOLVE_ADDR_FAILED; 336 smbd_disconnect_rdma_work(&sc->disconnect_work); 337 break; 338 339 case RDMA_CM_EVENT_ROUTE_ERROR: 340 log_rdma_event(ERR, "connecting failed event=%s\n", event_name); 341 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING); 342 sc->status = SMBDIRECT_SOCKET_RESOLVE_ROUTE_FAILED; 343 smbd_disconnect_rdma_work(&sc->disconnect_work); 344 break; 345 346 case RDMA_CM_EVENT_ESTABLISHED: 347 log_rdma_event(INFO, "connected event=%s\n", event_name); 348 349 /* 350 * Here we work around an inconsistency between 351 * iWarp and other devices (at least rxe and irdma using RoCEv2) 352 */ 353 if (rdma_protocol_iwarp(id->device, id->port_num)) { 354 /* 355 * iWarp devices report the peer's values 356 * with the perspective of the peer here. 357 * Tested with siw and irdma (in iwarp mode) 358 * We need to change to our perspective here, 359 * so we need to switch the values. 360 */ 361 peer_initiator_depth = event->param.conn.responder_resources; 362 peer_responder_resources = event->param.conn.initiator_depth; 363 } else { 364 /* 365 * Non iWarp devices report the peer's values 366 * already changed to our perspective here. 367 * Tested with rxe and irdma (in roce mode). 368 */ 369 peer_initiator_depth = event->param.conn.initiator_depth; 370 peer_responder_resources = event->param.conn.responder_resources; 371 } 372 if (rdma_protocol_iwarp(id->device, id->port_num) && 373 event->param.conn.private_data_len == 8) { 374 /* 375 * Legacy clients with only iWarp MPA v1 support 376 * need a private blob in order to negotiate 377 * the IRD/ORD values. 378 */ 379 const __be32 *ird_ord_hdr = event->param.conn.private_data; 380 u32 ird32 = be32_to_cpu(ird_ord_hdr[0]); 381 u32 ord32 = be32_to_cpu(ird_ord_hdr[1]); 382 383 /* 384 * cifs.ko sends the legacy IRD/ORD negotiation 385 * event if iWarp MPA v2 was used. 386 * 387 * Here we check that the values match and only 388 * mark the client as legacy if they don't match. 389 */ 390 if ((u32)event->param.conn.initiator_depth != ird32 || 391 (u32)event->param.conn.responder_resources != ord32) { 392 /* 393 * There are broken clients (old cifs.ko) 394 * using little endian and also 395 * struct rdma_conn_param only uses u8 396 * for initiator_depth and responder_resources, 397 * so we truncate the value to U8_MAX. 398 * 399 * smb_direct_accept_client() will then 400 * do the real negotiation in order to 401 * select the minimum between client and 402 * server. 403 */ 404 ird32 = min_t(u32, ird32, U8_MAX); 405 ord32 = min_t(u32, ord32, U8_MAX); 406 407 sc->rdma.legacy_iwarp = true; 408 peer_initiator_depth = (u8)ird32; 409 peer_responder_resources = (u8)ord32; 410 } 411 } 412 413 /* 414 * negotiate the value by using the minimum 415 * between client and server if the client provided 416 * non 0 values. 417 */ 418 if (peer_initiator_depth != 0) 419 sp->initiator_depth = 420 min_t(u8, sp->initiator_depth, 421 peer_initiator_depth); 422 if (peer_responder_resources != 0) 423 sp->responder_resources = 424 min_t(u8, sp->responder_resources, 425 peer_responder_resources); 426 427 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RDMA_CONNECT_RUNNING); 428 sc->status = SMBDIRECT_SOCKET_NEGOTIATE_NEEDED; 429 wake_up(&sc->status_wait); 430 break; 431 432 case RDMA_CM_EVENT_CONNECT_ERROR: 433 case RDMA_CM_EVENT_UNREACHABLE: 434 case RDMA_CM_EVENT_REJECTED: 435 log_rdma_event(ERR, "connecting failed event=%s\n", event_name); 436 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RDMA_CONNECT_RUNNING); 437 sc->status = SMBDIRECT_SOCKET_RDMA_CONNECT_FAILED; 438 smbd_disconnect_rdma_work(&sc->disconnect_work); 439 break; 440 441 case RDMA_CM_EVENT_DEVICE_REMOVAL: 442 case RDMA_CM_EVENT_DISCONNECTED: 443 /* This happens when we fail the negotiation */ 444 if (sc->status == SMBDIRECT_SOCKET_NEGOTIATE_FAILED) { 445 log_rdma_event(ERR, "event=%s during negotiation\n", event_name); 446 } 447 448 sc->status = SMBDIRECT_SOCKET_DISCONNECTED; 449 smbd_disconnect_rdma_work(&sc->disconnect_work); 450 break; 451 452 default: 453 log_rdma_event(ERR, "unexpected event=%s status=%d\n", 454 event_name, event->status); 455 break; 456 } 457 458 return 0; 459 } 460 461 /* Upcall from RDMA QP */ 462 static void 463 smbd_qp_async_error_upcall(struct ib_event *event, void *context) 464 { 465 struct smbdirect_socket *sc = context; 466 467 log_rdma_event(ERR, "%s on device %s socket %p\n", 468 ib_event_msg(event->event), event->device->name, sc); 469 470 switch (event->event) { 471 case IB_EVENT_CQ_ERR: 472 case IB_EVENT_QP_FATAL: 473 smbd_disconnect_rdma_connection(sc); 474 break; 475 476 default: 477 break; 478 } 479 } 480 481 static inline void *smbdirect_send_io_payload(struct smbdirect_send_io *request) 482 { 483 return (void *)request->packet; 484 } 485 486 static inline void *smbdirect_recv_io_payload(struct smbdirect_recv_io *response) 487 { 488 return (void *)response->packet; 489 } 490 491 /* Called when a RDMA send is done */ 492 static void send_done(struct ib_cq *cq, struct ib_wc *wc) 493 { 494 int i; 495 struct smbdirect_send_io *request = 496 container_of(wc->wr_cqe, struct smbdirect_send_io, cqe); 497 struct smbdirect_socket *sc = request->socket; 498 499 log_rdma_send(INFO, "smbdirect_send_io 0x%p completed wc->status=%s\n", 500 request, ib_wc_status_msg(wc->status)); 501 502 for (i = 0; i < request->num_sge; i++) 503 ib_dma_unmap_single(sc->ib.dev, 504 request->sge[i].addr, 505 request->sge[i].length, 506 DMA_TO_DEVICE); 507 508 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) { 509 if (wc->status != IB_WC_WR_FLUSH_ERR) 510 log_rdma_send(ERR, "wc->status=%s wc->opcode=%d\n", 511 ib_wc_status_msg(wc->status), wc->opcode); 512 mempool_free(request, sc->send_io.mem.pool); 513 smbd_disconnect_rdma_connection(sc); 514 return; 515 } 516 517 if (atomic_dec_and_test(&sc->send_io.pending.count)) 518 wake_up(&sc->send_io.pending.zero_wait_queue); 519 520 wake_up(&sc->send_io.pending.dec_wait_queue); 521 522 mempool_free(request, sc->send_io.mem.pool); 523 } 524 525 static void dump_smbdirect_negotiate_resp(struct smbdirect_negotiate_resp *resp) 526 { 527 log_rdma_event(INFO, "resp message min_version %u max_version %u negotiated_version %u credits_requested %u credits_granted %u status %u max_readwrite_size %u preferred_send_size %u max_receive_size %u max_fragmented_size %u\n", 528 resp->min_version, resp->max_version, 529 resp->negotiated_version, resp->credits_requested, 530 resp->credits_granted, resp->status, 531 resp->max_readwrite_size, resp->preferred_send_size, 532 resp->max_receive_size, resp->max_fragmented_size); 533 } 534 535 /* 536 * Process a negotiation response message, according to [MS-SMBD]3.1.5.7 537 * response, packet_length: the negotiation response message 538 * return value: true if negotiation is a success, false if failed 539 */ 540 static bool process_negotiation_response( 541 struct smbdirect_recv_io *response, int packet_length) 542 { 543 struct smbdirect_socket *sc = response->socket; 544 struct smbdirect_socket_parameters *sp = &sc->parameters; 545 struct smbdirect_negotiate_resp *packet = smbdirect_recv_io_payload(response); 546 547 if (packet_length < sizeof(struct smbdirect_negotiate_resp)) { 548 log_rdma_event(ERR, 549 "error: packet_length=%d\n", packet_length); 550 return false; 551 } 552 553 if (le16_to_cpu(packet->negotiated_version) != SMBDIRECT_V1) { 554 log_rdma_event(ERR, "error: negotiated_version=%x\n", 555 le16_to_cpu(packet->negotiated_version)); 556 return false; 557 } 558 559 if (packet->credits_requested == 0) { 560 log_rdma_event(ERR, "error: credits_requested==0\n"); 561 return false; 562 } 563 sc->recv_io.credits.target = le16_to_cpu(packet->credits_requested); 564 sc->recv_io.credits.target = min_t(u16, sc->recv_io.credits.target, sp->recv_credit_max); 565 566 if (packet->credits_granted == 0) { 567 log_rdma_event(ERR, "error: credits_granted==0\n"); 568 return false; 569 } 570 atomic_set(&sc->send_io.credits.count, le16_to_cpu(packet->credits_granted)); 571 572 if (le32_to_cpu(packet->preferred_send_size) > sp->max_recv_size) { 573 log_rdma_event(ERR, "error: preferred_send_size=%d\n", 574 le32_to_cpu(packet->preferred_send_size)); 575 return false; 576 } 577 sp->max_recv_size = le32_to_cpu(packet->preferred_send_size); 578 579 if (le32_to_cpu(packet->max_receive_size) < SMBD_MIN_RECEIVE_SIZE) { 580 log_rdma_event(ERR, "error: max_receive_size=%d\n", 581 le32_to_cpu(packet->max_receive_size)); 582 return false; 583 } 584 sp->max_send_size = min_t(u32, sp->max_send_size, 585 le32_to_cpu(packet->max_receive_size)); 586 587 if (le32_to_cpu(packet->max_fragmented_size) < 588 SMBD_MIN_FRAGMENTED_SIZE) { 589 log_rdma_event(ERR, "error: max_fragmented_size=%d\n", 590 le32_to_cpu(packet->max_fragmented_size)); 591 return false; 592 } 593 sp->max_fragmented_send_size = 594 le32_to_cpu(packet->max_fragmented_size); 595 596 597 sp->max_read_write_size = min_t(u32, 598 le32_to_cpu(packet->max_readwrite_size), 599 sp->max_frmr_depth * PAGE_SIZE); 600 sp->max_frmr_depth = sp->max_read_write_size / PAGE_SIZE; 601 602 sc->recv_io.expected = SMBDIRECT_EXPECT_DATA_TRANSFER; 603 return true; 604 } 605 606 static void smbd_post_send_credits(struct work_struct *work) 607 { 608 int rc; 609 struct smbdirect_recv_io *response; 610 struct smbdirect_socket *sc = 611 container_of(work, struct smbdirect_socket, recv_io.posted.refill_work); 612 613 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) { 614 return; 615 } 616 617 if (sc->recv_io.credits.target > 618 atomic_read(&sc->recv_io.credits.count)) { 619 while (true) { 620 response = get_receive_buffer(sc); 621 if (!response) 622 break; 623 624 response->first_segment = false; 625 rc = smbd_post_recv(sc, response); 626 if (rc) { 627 log_rdma_recv(ERR, 628 "post_recv failed rc=%d\n", rc); 629 put_receive_buffer(sc, response); 630 break; 631 } 632 633 atomic_inc(&sc->recv_io.posted.count); 634 } 635 } 636 637 /* Promptly send an immediate packet as defined in [MS-SMBD] 3.1.1.1 */ 638 if (atomic_read(&sc->recv_io.credits.count) < 639 sc->recv_io.credits.target - 1) { 640 log_keep_alive(INFO, "schedule send of an empty message\n"); 641 queue_work(sc->workqueue, &sc->idle.immediate_work); 642 } 643 } 644 645 /* Called from softirq, when recv is done */ 646 static void recv_done(struct ib_cq *cq, struct ib_wc *wc) 647 { 648 struct smbdirect_data_transfer *data_transfer; 649 struct smbdirect_recv_io *response = 650 container_of(wc->wr_cqe, struct smbdirect_recv_io, cqe); 651 struct smbdirect_socket *sc = response->socket; 652 struct smbdirect_socket_parameters *sp = &sc->parameters; 653 u16 old_recv_credit_target; 654 u32 data_offset = 0; 655 u32 data_length = 0; 656 u32 remaining_data_length = 0; 657 bool negotiate_done = false; 658 659 log_rdma_recv(INFO, 660 "response=0x%p type=%d wc status=%s wc opcode %d byte_len=%d pkey_index=%u\n", 661 response, sc->recv_io.expected, 662 ib_wc_status_msg(wc->status), wc->opcode, 663 wc->byte_len, wc->pkey_index); 664 665 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) { 666 if (wc->status != IB_WC_WR_FLUSH_ERR) 667 log_rdma_recv(ERR, "wc->status=%s opcode=%d\n", 668 ib_wc_status_msg(wc->status), wc->opcode); 669 goto error; 670 } 671 672 ib_dma_sync_single_for_cpu( 673 wc->qp->device, 674 response->sge.addr, 675 response->sge.length, 676 DMA_FROM_DEVICE); 677 678 /* 679 * Reset timer to the keepalive interval in 680 * order to trigger our next keepalive message. 681 */ 682 sc->idle.keepalive = SMBDIRECT_KEEPALIVE_NONE; 683 mod_delayed_work(sc->workqueue, &sc->idle.timer_work, 684 msecs_to_jiffies(sp->keepalive_interval_msec)); 685 686 switch (sc->recv_io.expected) { 687 /* SMBD negotiation response */ 688 case SMBDIRECT_EXPECT_NEGOTIATE_REP: 689 dump_smbdirect_negotiate_resp(smbdirect_recv_io_payload(response)); 690 sc->recv_io.reassembly.full_packet_received = true; 691 negotiate_done = 692 process_negotiation_response(response, wc->byte_len); 693 put_receive_buffer(sc, response); 694 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_NEGOTIATE_RUNNING); 695 if (!negotiate_done) { 696 sc->status = SMBDIRECT_SOCKET_NEGOTIATE_FAILED; 697 smbd_disconnect_rdma_connection(sc); 698 } else { 699 sc->status = SMBDIRECT_SOCKET_CONNECTED; 700 wake_up(&sc->status_wait); 701 } 702 703 return; 704 705 /* SMBD data transfer packet */ 706 case SMBDIRECT_EXPECT_DATA_TRANSFER: 707 data_transfer = smbdirect_recv_io_payload(response); 708 709 if (wc->byte_len < 710 offsetof(struct smbdirect_data_transfer, padding)) 711 goto error; 712 713 remaining_data_length = le32_to_cpu(data_transfer->remaining_data_length); 714 data_offset = le32_to_cpu(data_transfer->data_offset); 715 data_length = le32_to_cpu(data_transfer->data_length); 716 if (wc->byte_len < data_offset || 717 (u64)wc->byte_len < (u64)data_offset + data_length) 718 goto error; 719 720 if (remaining_data_length > sp->max_fragmented_recv_size || 721 data_length > sp->max_fragmented_recv_size || 722 (u64)remaining_data_length + (u64)data_length > (u64)sp->max_fragmented_recv_size) 723 goto error; 724 725 if (data_length) { 726 if (sc->recv_io.reassembly.full_packet_received) 727 response->first_segment = true; 728 729 if (le32_to_cpu(data_transfer->remaining_data_length)) 730 sc->recv_io.reassembly.full_packet_received = false; 731 else 732 sc->recv_io.reassembly.full_packet_received = true; 733 } 734 735 atomic_dec(&sc->recv_io.posted.count); 736 atomic_dec(&sc->recv_io.credits.count); 737 old_recv_credit_target = sc->recv_io.credits.target; 738 sc->recv_io.credits.target = 739 le16_to_cpu(data_transfer->credits_requested); 740 sc->recv_io.credits.target = 741 min_t(u16, sc->recv_io.credits.target, sp->recv_credit_max); 742 sc->recv_io.credits.target = 743 max_t(u16, sc->recv_io.credits.target, 1); 744 if (le16_to_cpu(data_transfer->credits_granted)) { 745 atomic_add(le16_to_cpu(data_transfer->credits_granted), 746 &sc->send_io.credits.count); 747 /* 748 * We have new send credits granted from remote peer 749 * If any sender is waiting for credits, unblock it 750 */ 751 wake_up(&sc->send_io.credits.wait_queue); 752 } 753 754 log_incoming(INFO, "data flags %d data_offset %d data_length %d remaining_data_length %d\n", 755 le16_to_cpu(data_transfer->flags), 756 le32_to_cpu(data_transfer->data_offset), 757 le32_to_cpu(data_transfer->data_length), 758 le32_to_cpu(data_transfer->remaining_data_length)); 759 760 /* Send an immediate response right away if requested */ 761 if (le16_to_cpu(data_transfer->flags) & 762 SMBDIRECT_FLAG_RESPONSE_REQUESTED) { 763 log_keep_alive(INFO, "schedule send of immediate response\n"); 764 queue_work(sc->workqueue, &sc->idle.immediate_work); 765 } 766 767 /* 768 * If this is a packet with data playload place the data in 769 * reassembly queue and wake up the reading thread 770 */ 771 if (data_length) { 772 if (sc->recv_io.credits.target > old_recv_credit_target) 773 queue_work(sc->workqueue, &sc->recv_io.posted.refill_work); 774 775 enqueue_reassembly(sc, response, data_length); 776 wake_up(&sc->recv_io.reassembly.wait_queue); 777 } else 778 put_receive_buffer(sc, response); 779 780 return; 781 782 case SMBDIRECT_EXPECT_NEGOTIATE_REQ: 783 /* Only server... */ 784 break; 785 } 786 787 /* 788 * This is an internal error! 789 */ 790 log_rdma_recv(ERR, "unexpected response type=%d\n", sc->recv_io.expected); 791 WARN_ON_ONCE(sc->recv_io.expected != SMBDIRECT_EXPECT_DATA_TRANSFER); 792 error: 793 put_receive_buffer(sc, response); 794 smbd_disconnect_rdma_connection(sc); 795 } 796 797 static struct rdma_cm_id *smbd_create_id( 798 struct smbdirect_socket *sc, 799 struct sockaddr *dstaddr, int port) 800 { 801 struct smbdirect_socket_parameters *sp = &sc->parameters; 802 struct rdma_cm_id *id; 803 int rc; 804 __be16 *sport; 805 806 id = rdma_create_id(&init_net, smbd_conn_upcall, sc, 807 RDMA_PS_TCP, IB_QPT_RC); 808 if (IS_ERR(id)) { 809 rc = PTR_ERR(id); 810 log_rdma_event(ERR, "rdma_create_id() failed %i\n", rc); 811 return id; 812 } 813 814 if (dstaddr->sa_family == AF_INET6) 815 sport = &((struct sockaddr_in6 *)dstaddr)->sin6_port; 816 else 817 sport = &((struct sockaddr_in *)dstaddr)->sin_port; 818 819 *sport = htons(port); 820 821 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RESOLVE_ADDR_NEEDED); 822 sc->status = SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING; 823 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)dstaddr, 824 sp->resolve_addr_timeout_msec); 825 if (rc) { 826 log_rdma_event(ERR, "rdma_resolve_addr() failed %i\n", rc); 827 goto out; 828 } 829 rc = wait_event_interruptible_timeout( 830 sc->status_wait, 831 sc->status != SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING, 832 msecs_to_jiffies(sp->resolve_addr_timeout_msec)); 833 /* e.g. if interrupted returns -ERESTARTSYS */ 834 if (rc < 0) { 835 log_rdma_event(ERR, "rdma_resolve_addr timeout rc: %i\n", rc); 836 goto out; 837 } 838 if (sc->status == SMBDIRECT_SOCKET_RESOLVE_ADDR_RUNNING) { 839 rc = -ETIMEDOUT; 840 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc); 841 goto out; 842 } 843 if (sc->status != SMBDIRECT_SOCKET_RESOLVE_ROUTE_NEEDED) { 844 rc = -EHOSTUNREACH; 845 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc); 846 goto out; 847 } 848 849 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RESOLVE_ROUTE_NEEDED); 850 sc->status = SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING; 851 rc = rdma_resolve_route(id, sp->resolve_route_timeout_msec); 852 if (rc) { 853 log_rdma_event(ERR, "rdma_resolve_route() failed %i\n", rc); 854 goto out; 855 } 856 rc = wait_event_interruptible_timeout( 857 sc->status_wait, 858 sc->status != SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING, 859 msecs_to_jiffies(sp->resolve_route_timeout_msec)); 860 /* e.g. if interrupted returns -ERESTARTSYS */ 861 if (rc < 0) { 862 log_rdma_event(ERR, "rdma_resolve_addr timeout rc: %i\n", rc); 863 goto out; 864 } 865 if (sc->status == SMBDIRECT_SOCKET_RESOLVE_ROUTE_RUNNING) { 866 rc = -ETIMEDOUT; 867 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc); 868 goto out; 869 } 870 if (sc->status != SMBDIRECT_SOCKET_RDMA_CONNECT_NEEDED) { 871 rc = -ENETUNREACH; 872 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc); 873 goto out; 874 } 875 876 return id; 877 878 out: 879 rdma_destroy_id(id); 880 return ERR_PTR(rc); 881 } 882 883 /* 884 * Test if FRWR (Fast Registration Work Requests) is supported on the device 885 * This implementation requires FRWR on RDMA read/write 886 * return value: true if it is supported 887 */ 888 static bool frwr_is_supported(struct ib_device_attr *attrs) 889 { 890 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)) 891 return false; 892 if (attrs->max_fast_reg_page_list_len == 0) 893 return false; 894 return true; 895 } 896 897 static int smbd_ia_open( 898 struct smbdirect_socket *sc, 899 struct sockaddr *dstaddr, int port) 900 { 901 struct smbdirect_socket_parameters *sp = &sc->parameters; 902 int rc; 903 904 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_CREATED); 905 sc->status = SMBDIRECT_SOCKET_RESOLVE_ADDR_NEEDED; 906 907 sc->rdma.cm_id = smbd_create_id(sc, dstaddr, port); 908 if (IS_ERR(sc->rdma.cm_id)) { 909 rc = PTR_ERR(sc->rdma.cm_id); 910 goto out1; 911 } 912 sc->ib.dev = sc->rdma.cm_id->device; 913 914 if (!frwr_is_supported(&sc->ib.dev->attrs)) { 915 log_rdma_event(ERR, "Fast Registration Work Requests (FRWR) is not supported\n"); 916 log_rdma_event(ERR, "Device capability flags = %llx max_fast_reg_page_list_len = %u\n", 917 sc->ib.dev->attrs.device_cap_flags, 918 sc->ib.dev->attrs.max_fast_reg_page_list_len); 919 rc = -EPROTONOSUPPORT; 920 goto out2; 921 } 922 sp->max_frmr_depth = min_t(u32, 923 sp->max_frmr_depth, 924 sc->ib.dev->attrs.max_fast_reg_page_list_len); 925 sc->mr_io.type = IB_MR_TYPE_MEM_REG; 926 if (sc->ib.dev->attrs.kernel_cap_flags & IBK_SG_GAPS_REG) 927 sc->mr_io.type = IB_MR_TYPE_SG_GAPS; 928 929 return 0; 930 931 out2: 932 rdma_destroy_id(sc->rdma.cm_id); 933 sc->rdma.cm_id = NULL; 934 935 out1: 936 return rc; 937 } 938 939 /* 940 * Send a negotiation request message to the peer 941 * The negotiation procedure is in [MS-SMBD] 3.1.5.2 and 3.1.5.3 942 * After negotiation, the transport is connected and ready for 943 * carrying upper layer SMB payload 944 */ 945 static int smbd_post_send_negotiate_req(struct smbdirect_socket *sc) 946 { 947 struct smbdirect_socket_parameters *sp = &sc->parameters; 948 struct ib_send_wr send_wr; 949 int rc = -ENOMEM; 950 struct smbdirect_send_io *request; 951 struct smbdirect_negotiate_req *packet; 952 953 request = mempool_alloc(sc->send_io.mem.pool, GFP_KERNEL); 954 if (!request) 955 return rc; 956 957 request->socket = sc; 958 959 packet = smbdirect_send_io_payload(request); 960 packet->min_version = cpu_to_le16(SMBDIRECT_V1); 961 packet->max_version = cpu_to_le16(SMBDIRECT_V1); 962 packet->reserved = 0; 963 packet->credits_requested = cpu_to_le16(sp->send_credit_target); 964 packet->preferred_send_size = cpu_to_le32(sp->max_send_size); 965 packet->max_receive_size = cpu_to_le32(sp->max_recv_size); 966 packet->max_fragmented_size = 967 cpu_to_le32(sp->max_fragmented_recv_size); 968 969 request->num_sge = 1; 970 request->sge[0].addr = ib_dma_map_single( 971 sc->ib.dev, (void *)packet, 972 sizeof(*packet), DMA_TO_DEVICE); 973 if (ib_dma_mapping_error(sc->ib.dev, request->sge[0].addr)) { 974 rc = -EIO; 975 goto dma_mapping_failed; 976 } 977 978 request->sge[0].length = sizeof(*packet); 979 request->sge[0].lkey = sc->ib.pd->local_dma_lkey; 980 981 ib_dma_sync_single_for_device( 982 sc->ib.dev, request->sge[0].addr, 983 request->sge[0].length, DMA_TO_DEVICE); 984 985 request->cqe.done = send_done; 986 987 send_wr.next = NULL; 988 send_wr.wr_cqe = &request->cqe; 989 send_wr.sg_list = request->sge; 990 send_wr.num_sge = request->num_sge; 991 send_wr.opcode = IB_WR_SEND; 992 send_wr.send_flags = IB_SEND_SIGNALED; 993 994 log_rdma_send(INFO, "sge addr=0x%llx length=%u lkey=0x%x\n", 995 request->sge[0].addr, 996 request->sge[0].length, request->sge[0].lkey); 997 998 atomic_inc(&sc->send_io.pending.count); 999 rc = ib_post_send(sc->ib.qp, &send_wr, NULL); 1000 if (!rc) 1001 return 0; 1002 1003 /* if we reach here, post send failed */ 1004 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc); 1005 atomic_dec(&sc->send_io.pending.count); 1006 ib_dma_unmap_single(sc->ib.dev, request->sge[0].addr, 1007 request->sge[0].length, DMA_TO_DEVICE); 1008 1009 smbd_disconnect_rdma_connection(sc); 1010 1011 dma_mapping_failed: 1012 mempool_free(request, sc->send_io.mem.pool); 1013 return rc; 1014 } 1015 1016 /* 1017 * Extend the credits to remote peer 1018 * This implements [MS-SMBD] 3.1.5.9 1019 * The idea is that we should extend credits to remote peer as quickly as 1020 * it's allowed, to maintain data flow. We allocate as much receive 1021 * buffer as possible, and extend the receive credits to remote peer 1022 * return value: the new credtis being granted. 1023 */ 1024 static int manage_credits_prior_sending(struct smbdirect_socket *sc) 1025 { 1026 int new_credits; 1027 1028 if (atomic_read(&sc->recv_io.credits.count) >= sc->recv_io.credits.target) 1029 return 0; 1030 1031 new_credits = atomic_read(&sc->recv_io.posted.count); 1032 if (new_credits == 0) 1033 return 0; 1034 1035 new_credits -= atomic_read(&sc->recv_io.credits.count); 1036 if (new_credits <= 0) 1037 return 0; 1038 1039 return new_credits; 1040 } 1041 1042 /* 1043 * Check if we need to send a KEEP_ALIVE message 1044 * The idle connection timer triggers a KEEP_ALIVE message when expires 1045 * SMBDIRECT_FLAG_RESPONSE_REQUESTED is set in the message flag to have peer send 1046 * back a response. 1047 * return value: 1048 * 1 if SMBDIRECT_FLAG_RESPONSE_REQUESTED needs to be set 1049 * 0: otherwise 1050 */ 1051 static int manage_keep_alive_before_sending(struct smbdirect_socket *sc) 1052 { 1053 struct smbdirect_socket_parameters *sp = &sc->parameters; 1054 1055 if (sc->idle.keepalive == SMBDIRECT_KEEPALIVE_PENDING) { 1056 sc->idle.keepalive = SMBDIRECT_KEEPALIVE_SENT; 1057 /* 1058 * Now use the keepalive timeout (instead of keepalive interval) 1059 * in order to wait for a response 1060 */ 1061 mod_delayed_work(sc->workqueue, &sc->idle.timer_work, 1062 msecs_to_jiffies(sp->keepalive_timeout_msec)); 1063 return 1; 1064 } 1065 return 0; 1066 } 1067 1068 /* Post the send request */ 1069 static int smbd_post_send(struct smbdirect_socket *sc, 1070 struct smbdirect_send_io *request) 1071 { 1072 struct ib_send_wr send_wr; 1073 int rc, i; 1074 1075 for (i = 0; i < request->num_sge; i++) { 1076 log_rdma_send(INFO, 1077 "rdma_request sge[%d] addr=0x%llx length=%u\n", 1078 i, request->sge[i].addr, request->sge[i].length); 1079 ib_dma_sync_single_for_device( 1080 sc->ib.dev, 1081 request->sge[i].addr, 1082 request->sge[i].length, 1083 DMA_TO_DEVICE); 1084 } 1085 1086 request->cqe.done = send_done; 1087 1088 send_wr.next = NULL; 1089 send_wr.wr_cqe = &request->cqe; 1090 send_wr.sg_list = request->sge; 1091 send_wr.num_sge = request->num_sge; 1092 send_wr.opcode = IB_WR_SEND; 1093 send_wr.send_flags = IB_SEND_SIGNALED; 1094 1095 rc = ib_post_send(sc->ib.qp, &send_wr, NULL); 1096 if (rc) { 1097 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc); 1098 smbd_disconnect_rdma_connection(sc); 1099 rc = -EAGAIN; 1100 } 1101 1102 return rc; 1103 } 1104 1105 static int smbd_post_send_iter(struct smbdirect_socket *sc, 1106 struct iov_iter *iter, 1107 int *_remaining_data_length) 1108 { 1109 struct smbdirect_socket_parameters *sp = &sc->parameters; 1110 int i, rc; 1111 int header_length; 1112 int data_length; 1113 struct smbdirect_send_io *request; 1114 struct smbdirect_data_transfer *packet; 1115 int new_credits = 0; 1116 1117 wait_credit: 1118 /* Wait for send credits. A SMBD packet needs one credit */ 1119 rc = wait_event_interruptible(sc->send_io.credits.wait_queue, 1120 atomic_read(&sc->send_io.credits.count) > 0 || 1121 sc->status != SMBDIRECT_SOCKET_CONNECTED); 1122 if (rc) 1123 goto err_wait_credit; 1124 1125 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) { 1126 log_outgoing(ERR, "disconnected not sending on wait_credit\n"); 1127 rc = -EAGAIN; 1128 goto err_wait_credit; 1129 } 1130 if (unlikely(atomic_dec_return(&sc->send_io.credits.count) < 0)) { 1131 atomic_inc(&sc->send_io.credits.count); 1132 goto wait_credit; 1133 } 1134 1135 wait_send_queue: 1136 wait_event(sc->send_io.pending.dec_wait_queue, 1137 atomic_read(&sc->send_io.pending.count) < sp->send_credit_target || 1138 sc->status != SMBDIRECT_SOCKET_CONNECTED); 1139 1140 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) { 1141 log_outgoing(ERR, "disconnected not sending on wait_send_queue\n"); 1142 rc = -EAGAIN; 1143 goto err_wait_send_queue; 1144 } 1145 1146 if (unlikely(atomic_inc_return(&sc->send_io.pending.count) > 1147 sp->send_credit_target)) { 1148 atomic_dec(&sc->send_io.pending.count); 1149 goto wait_send_queue; 1150 } 1151 1152 request = mempool_alloc(sc->send_io.mem.pool, GFP_KERNEL); 1153 if (!request) { 1154 rc = -ENOMEM; 1155 goto err_alloc; 1156 } 1157 1158 request->socket = sc; 1159 memset(request->sge, 0, sizeof(request->sge)); 1160 1161 /* Map the packet to DMA */ 1162 header_length = sizeof(struct smbdirect_data_transfer); 1163 /* If this is a packet without payload, don't send padding */ 1164 if (!iter) 1165 header_length = offsetof(struct smbdirect_data_transfer, padding); 1166 1167 packet = smbdirect_send_io_payload(request); 1168 request->sge[0].addr = ib_dma_map_single(sc->ib.dev, 1169 (void *)packet, 1170 header_length, 1171 DMA_TO_DEVICE); 1172 if (ib_dma_mapping_error(sc->ib.dev, request->sge[0].addr)) { 1173 rc = -EIO; 1174 goto err_dma; 1175 } 1176 1177 request->sge[0].length = header_length; 1178 request->sge[0].lkey = sc->ib.pd->local_dma_lkey; 1179 request->num_sge = 1; 1180 1181 /* Fill in the data payload to find out how much data we can add */ 1182 if (iter) { 1183 struct smb_extract_to_rdma extract = { 1184 .nr_sge = request->num_sge, 1185 .max_sge = SMBDIRECT_SEND_IO_MAX_SGE, 1186 .sge = request->sge, 1187 .device = sc->ib.dev, 1188 .local_dma_lkey = sc->ib.pd->local_dma_lkey, 1189 .direction = DMA_TO_DEVICE, 1190 }; 1191 size_t payload_len = umin(*_remaining_data_length, 1192 sp->max_send_size - sizeof(*packet)); 1193 1194 rc = smb_extract_iter_to_rdma(iter, payload_len, 1195 &extract); 1196 if (rc < 0) 1197 goto err_dma; 1198 data_length = rc; 1199 request->num_sge = extract.nr_sge; 1200 *_remaining_data_length -= data_length; 1201 } else { 1202 data_length = 0; 1203 } 1204 1205 /* Fill in the packet header */ 1206 packet->credits_requested = cpu_to_le16(sp->send_credit_target); 1207 1208 new_credits = manage_credits_prior_sending(sc); 1209 atomic_add(new_credits, &sc->recv_io.credits.count); 1210 packet->credits_granted = cpu_to_le16(new_credits); 1211 1212 packet->flags = 0; 1213 if (manage_keep_alive_before_sending(sc)) 1214 packet->flags |= cpu_to_le16(SMBDIRECT_FLAG_RESPONSE_REQUESTED); 1215 1216 packet->reserved = 0; 1217 if (!data_length) 1218 packet->data_offset = 0; 1219 else 1220 packet->data_offset = cpu_to_le32(24); 1221 packet->data_length = cpu_to_le32(data_length); 1222 packet->remaining_data_length = cpu_to_le32(*_remaining_data_length); 1223 packet->padding = 0; 1224 1225 log_outgoing(INFO, "credits_requested=%d credits_granted=%d data_offset=%d data_length=%d remaining_data_length=%d\n", 1226 le16_to_cpu(packet->credits_requested), 1227 le16_to_cpu(packet->credits_granted), 1228 le32_to_cpu(packet->data_offset), 1229 le32_to_cpu(packet->data_length), 1230 le32_to_cpu(packet->remaining_data_length)); 1231 1232 rc = smbd_post_send(sc, request); 1233 if (!rc) 1234 return 0; 1235 1236 err_dma: 1237 for (i = 0; i < request->num_sge; i++) 1238 if (request->sge[i].addr) 1239 ib_dma_unmap_single(sc->ib.dev, 1240 request->sge[i].addr, 1241 request->sge[i].length, 1242 DMA_TO_DEVICE); 1243 mempool_free(request, sc->send_io.mem.pool); 1244 1245 /* roll back the granted receive credits */ 1246 atomic_sub(new_credits, &sc->recv_io.credits.count); 1247 1248 err_alloc: 1249 if (atomic_dec_and_test(&sc->send_io.pending.count)) 1250 wake_up(&sc->send_io.pending.zero_wait_queue); 1251 1252 err_wait_send_queue: 1253 /* roll back send credits and pending */ 1254 atomic_inc(&sc->send_io.credits.count); 1255 1256 err_wait_credit: 1257 return rc; 1258 } 1259 1260 /* 1261 * Send an empty message 1262 * Empty message is used to extend credits to peer to for keep live 1263 * while there is no upper layer payload to send at the time 1264 */ 1265 static int smbd_post_send_empty(struct smbdirect_socket *sc) 1266 { 1267 int remaining_data_length = 0; 1268 1269 sc->statistics.send_empty++; 1270 return smbd_post_send_iter(sc, NULL, &remaining_data_length); 1271 } 1272 1273 static int smbd_post_send_full_iter(struct smbdirect_socket *sc, 1274 struct iov_iter *iter, 1275 int *_remaining_data_length) 1276 { 1277 int rc = 0; 1278 1279 /* 1280 * smbd_post_send_iter() respects the 1281 * negotiated max_send_size, so we need to 1282 * loop until the full iter is posted 1283 */ 1284 1285 while (iov_iter_count(iter) > 0) { 1286 rc = smbd_post_send_iter(sc, iter, _remaining_data_length); 1287 if (rc < 0) 1288 break; 1289 } 1290 1291 return rc; 1292 } 1293 1294 /* 1295 * Post a receive request to the transport 1296 * The remote peer can only send data when a receive request is posted 1297 * The interaction is controlled by send/receive credit system 1298 */ 1299 static int smbd_post_recv( 1300 struct smbdirect_socket *sc, struct smbdirect_recv_io *response) 1301 { 1302 struct smbdirect_socket_parameters *sp = &sc->parameters; 1303 struct ib_recv_wr recv_wr; 1304 int rc = -EIO; 1305 1306 response->sge.addr = ib_dma_map_single( 1307 sc->ib.dev, response->packet, 1308 sp->max_recv_size, DMA_FROM_DEVICE); 1309 if (ib_dma_mapping_error(sc->ib.dev, response->sge.addr)) 1310 return rc; 1311 1312 response->sge.length = sp->max_recv_size; 1313 response->sge.lkey = sc->ib.pd->local_dma_lkey; 1314 1315 response->cqe.done = recv_done; 1316 1317 recv_wr.wr_cqe = &response->cqe; 1318 recv_wr.next = NULL; 1319 recv_wr.sg_list = &response->sge; 1320 recv_wr.num_sge = 1; 1321 1322 rc = ib_post_recv(sc->ib.qp, &recv_wr, NULL); 1323 if (rc) { 1324 ib_dma_unmap_single(sc->ib.dev, response->sge.addr, 1325 response->sge.length, DMA_FROM_DEVICE); 1326 response->sge.length = 0; 1327 smbd_disconnect_rdma_connection(sc); 1328 log_rdma_recv(ERR, "ib_post_recv failed rc=%d\n", rc); 1329 } 1330 1331 return rc; 1332 } 1333 1334 /* Perform SMBD negotiate according to [MS-SMBD] 3.1.5.2 */ 1335 static int smbd_negotiate(struct smbdirect_socket *sc) 1336 { 1337 struct smbdirect_socket_parameters *sp = &sc->parameters; 1338 int rc; 1339 struct smbdirect_recv_io *response = get_receive_buffer(sc); 1340 1341 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_NEGOTIATE_NEEDED); 1342 sc->status = SMBDIRECT_SOCKET_NEGOTIATE_RUNNING; 1343 1344 sc->recv_io.expected = SMBDIRECT_EXPECT_NEGOTIATE_REP; 1345 rc = smbd_post_recv(sc, response); 1346 log_rdma_event(INFO, "smbd_post_recv rc=%d iov.addr=0x%llx iov.length=%u iov.lkey=0x%x\n", 1347 rc, response->sge.addr, 1348 response->sge.length, response->sge.lkey); 1349 if (rc) { 1350 put_receive_buffer(sc, response); 1351 return rc; 1352 } 1353 1354 rc = smbd_post_send_negotiate_req(sc); 1355 if (rc) 1356 return rc; 1357 1358 rc = wait_event_interruptible_timeout( 1359 sc->status_wait, 1360 sc->status != SMBDIRECT_SOCKET_NEGOTIATE_RUNNING, 1361 msecs_to_jiffies(sp->negotiate_timeout_msec)); 1362 log_rdma_event(INFO, "wait_event_interruptible_timeout rc=%d\n", rc); 1363 1364 if (sc->status == SMBDIRECT_SOCKET_CONNECTED) 1365 return 0; 1366 1367 if (rc == 0) 1368 rc = -ETIMEDOUT; 1369 else if (rc == -ERESTARTSYS) 1370 rc = -EINTR; 1371 else 1372 rc = -ENOTCONN; 1373 1374 return rc; 1375 } 1376 1377 /* 1378 * Implement Connection.FragmentReassemblyBuffer defined in [MS-SMBD] 3.1.1.1 1379 * This is a queue for reassembling upper layer payload and present to upper 1380 * layer. All the inncoming payload go to the reassembly queue, regardless of 1381 * if reassembly is required. The uuper layer code reads from the queue for all 1382 * incoming payloads. 1383 * Put a received packet to the reassembly queue 1384 * response: the packet received 1385 * data_length: the size of payload in this packet 1386 */ 1387 static void enqueue_reassembly( 1388 struct smbdirect_socket *sc, 1389 struct smbdirect_recv_io *response, 1390 int data_length) 1391 { 1392 unsigned long flags; 1393 1394 spin_lock_irqsave(&sc->recv_io.reassembly.lock, flags); 1395 list_add_tail(&response->list, &sc->recv_io.reassembly.list); 1396 sc->recv_io.reassembly.queue_length++; 1397 /* 1398 * Make sure reassembly_data_length is updated after list and 1399 * reassembly_queue_length are updated. On the dequeue side 1400 * reassembly_data_length is checked without a lock to determine 1401 * if reassembly_queue_length and list is up to date 1402 */ 1403 virt_wmb(); 1404 sc->recv_io.reassembly.data_length += data_length; 1405 spin_unlock_irqrestore(&sc->recv_io.reassembly.lock, flags); 1406 sc->statistics.enqueue_reassembly_queue++; 1407 } 1408 1409 /* 1410 * Get the first entry at the front of reassembly queue 1411 * Caller is responsible for locking 1412 * return value: the first entry if any, NULL if queue is empty 1413 */ 1414 static struct smbdirect_recv_io *_get_first_reassembly(struct smbdirect_socket *sc) 1415 { 1416 struct smbdirect_recv_io *ret = NULL; 1417 1418 if (!list_empty(&sc->recv_io.reassembly.list)) { 1419 ret = list_first_entry( 1420 &sc->recv_io.reassembly.list, 1421 struct smbdirect_recv_io, list); 1422 } 1423 return ret; 1424 } 1425 1426 /* 1427 * Get a receive buffer 1428 * For each remote send, we need to post a receive. The receive buffers are 1429 * pre-allocated in advance. 1430 * return value: the receive buffer, NULL if none is available 1431 */ 1432 static struct smbdirect_recv_io *get_receive_buffer(struct smbdirect_socket *sc) 1433 { 1434 struct smbdirect_recv_io *ret = NULL; 1435 unsigned long flags; 1436 1437 spin_lock_irqsave(&sc->recv_io.free.lock, flags); 1438 if (!list_empty(&sc->recv_io.free.list)) { 1439 ret = list_first_entry( 1440 &sc->recv_io.free.list, 1441 struct smbdirect_recv_io, list); 1442 list_del(&ret->list); 1443 sc->statistics.get_receive_buffer++; 1444 } 1445 spin_unlock_irqrestore(&sc->recv_io.free.lock, flags); 1446 1447 return ret; 1448 } 1449 1450 /* 1451 * Return a receive buffer 1452 * Upon returning of a receive buffer, we can post new receive and extend 1453 * more receive credits to remote peer. This is done immediately after a 1454 * receive buffer is returned. 1455 */ 1456 static void put_receive_buffer( 1457 struct smbdirect_socket *sc, struct smbdirect_recv_io *response) 1458 { 1459 unsigned long flags; 1460 1461 if (likely(response->sge.length != 0)) { 1462 ib_dma_unmap_single(sc->ib.dev, 1463 response->sge.addr, 1464 response->sge.length, 1465 DMA_FROM_DEVICE); 1466 response->sge.length = 0; 1467 } 1468 1469 spin_lock_irqsave(&sc->recv_io.free.lock, flags); 1470 list_add_tail(&response->list, &sc->recv_io.free.list); 1471 sc->statistics.put_receive_buffer++; 1472 spin_unlock_irqrestore(&sc->recv_io.free.lock, flags); 1473 1474 queue_work(sc->workqueue, &sc->recv_io.posted.refill_work); 1475 } 1476 1477 /* Preallocate all receive buffer on transport establishment */ 1478 static int allocate_receive_buffers(struct smbdirect_socket *sc, int num_buf) 1479 { 1480 struct smbdirect_recv_io *response; 1481 int i; 1482 1483 for (i = 0; i < num_buf; i++) { 1484 response = mempool_alloc(sc->recv_io.mem.pool, GFP_KERNEL); 1485 if (!response) 1486 goto allocate_failed; 1487 1488 response->socket = sc; 1489 response->sge.length = 0; 1490 list_add_tail(&response->list, &sc->recv_io.free.list); 1491 } 1492 1493 return 0; 1494 1495 allocate_failed: 1496 while (!list_empty(&sc->recv_io.free.list)) { 1497 response = list_first_entry( 1498 &sc->recv_io.free.list, 1499 struct smbdirect_recv_io, list); 1500 list_del(&response->list); 1501 1502 mempool_free(response, sc->recv_io.mem.pool); 1503 } 1504 return -ENOMEM; 1505 } 1506 1507 static void destroy_receive_buffers(struct smbdirect_socket *sc) 1508 { 1509 struct smbdirect_recv_io *response; 1510 1511 while ((response = get_receive_buffer(sc))) 1512 mempool_free(response, sc->recv_io.mem.pool); 1513 } 1514 1515 static void send_immediate_empty_message(struct work_struct *work) 1516 { 1517 struct smbdirect_socket *sc = 1518 container_of(work, struct smbdirect_socket, idle.immediate_work); 1519 1520 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) 1521 return; 1522 1523 log_keep_alive(INFO, "send an empty message\n"); 1524 smbd_post_send_empty(sc); 1525 } 1526 1527 /* Implement idle connection timer [MS-SMBD] 3.1.6.2 */ 1528 static void idle_connection_timer(struct work_struct *work) 1529 { 1530 struct smbdirect_socket *sc = 1531 container_of(work, struct smbdirect_socket, idle.timer_work.work); 1532 struct smbdirect_socket_parameters *sp = &sc->parameters; 1533 1534 if (sc->idle.keepalive != SMBDIRECT_KEEPALIVE_NONE) { 1535 log_keep_alive(ERR, 1536 "error status sc->idle.keepalive=%d\n", 1537 sc->idle.keepalive); 1538 smbd_disconnect_rdma_connection(sc); 1539 return; 1540 } 1541 1542 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) 1543 return; 1544 1545 /* 1546 * Now use the keepalive timeout (instead of keepalive interval) 1547 * in order to wait for a response 1548 */ 1549 sc->idle.keepalive = SMBDIRECT_KEEPALIVE_PENDING; 1550 mod_delayed_work(sc->workqueue, &sc->idle.timer_work, 1551 msecs_to_jiffies(sp->keepalive_timeout_msec)); 1552 log_keep_alive(INFO, "schedule send of empty idle message\n"); 1553 queue_work(sc->workqueue, &sc->idle.immediate_work); 1554 } 1555 1556 /* 1557 * Destroy the transport and related RDMA and memory resources 1558 * Need to go through all the pending counters and make sure on one is using 1559 * the transport while it is destroyed 1560 */ 1561 void smbd_destroy(struct TCP_Server_Info *server) 1562 { 1563 struct smbd_connection *info = server->smbd_conn; 1564 struct smbdirect_socket *sc; 1565 struct smbdirect_recv_io *response; 1566 unsigned long flags; 1567 1568 if (!info) { 1569 log_rdma_event(INFO, "rdma session already destroyed\n"); 1570 return; 1571 } 1572 sc = &info->socket; 1573 1574 log_rdma_event(INFO, "cancelling and disable disconnect_work\n"); 1575 disable_work_sync(&sc->disconnect_work); 1576 1577 log_rdma_event(INFO, "destroying rdma session\n"); 1578 if (sc->status < SMBDIRECT_SOCKET_DISCONNECTING) { 1579 smbd_disconnect_rdma_work(&sc->disconnect_work); 1580 log_rdma_event(INFO, "wait for transport being disconnected\n"); 1581 wait_event_interruptible( 1582 sc->status_wait, 1583 sc->status == SMBDIRECT_SOCKET_DISCONNECTED); 1584 } 1585 1586 /* 1587 * Wake up all waiters in all wait queues 1588 * in order to notice the broken connection. 1589 * 1590 * Most likely this was already called via 1591 * smbd_disconnect_rdma_work(), but call it again... 1592 */ 1593 smbd_disconnect_wake_up_all(sc); 1594 1595 log_rdma_event(INFO, "cancelling recv_io.posted.refill_work\n"); 1596 disable_work_sync(&sc->recv_io.posted.refill_work); 1597 1598 log_rdma_event(INFO, "destroying qp\n"); 1599 ib_drain_qp(sc->ib.qp); 1600 rdma_destroy_qp(sc->rdma.cm_id); 1601 sc->ib.qp = NULL; 1602 1603 log_rdma_event(INFO, "cancelling idle timer\n"); 1604 disable_delayed_work_sync(&sc->idle.timer_work); 1605 log_rdma_event(INFO, "cancelling send immediate work\n"); 1606 disable_work_sync(&sc->idle.immediate_work); 1607 1608 /* It's not possible for upper layer to get to reassembly */ 1609 log_rdma_event(INFO, "drain the reassembly queue\n"); 1610 do { 1611 spin_lock_irqsave(&sc->recv_io.reassembly.lock, flags); 1612 response = _get_first_reassembly(sc); 1613 if (response) { 1614 list_del(&response->list); 1615 spin_unlock_irqrestore( 1616 &sc->recv_io.reassembly.lock, flags); 1617 put_receive_buffer(sc, response); 1618 } else 1619 spin_unlock_irqrestore( 1620 &sc->recv_io.reassembly.lock, flags); 1621 } while (response); 1622 sc->recv_io.reassembly.data_length = 0; 1623 1624 log_rdma_event(INFO, "free receive buffers\n"); 1625 destroy_receive_buffers(sc); 1626 1627 /* 1628 * For performance reasons, memory registration and deregistration 1629 * are not locked by srv_mutex. It is possible some processes are 1630 * blocked on transport srv_mutex while holding memory registration. 1631 * Release the transport srv_mutex to allow them to hit the failure 1632 * path when sending data, and then release memory registrations. 1633 */ 1634 log_rdma_event(INFO, "freeing mr list\n"); 1635 while (atomic_read(&sc->mr_io.used.count)) { 1636 cifs_server_unlock(server); 1637 msleep(1000); 1638 cifs_server_lock(server); 1639 } 1640 destroy_mr_list(sc); 1641 1642 ib_free_cq(sc->ib.send_cq); 1643 ib_free_cq(sc->ib.recv_cq); 1644 ib_dealloc_pd(sc->ib.pd); 1645 rdma_destroy_id(sc->rdma.cm_id); 1646 1647 /* free mempools */ 1648 mempool_destroy(sc->send_io.mem.pool); 1649 kmem_cache_destroy(sc->send_io.mem.cache); 1650 1651 mempool_destroy(sc->recv_io.mem.pool); 1652 kmem_cache_destroy(sc->recv_io.mem.cache); 1653 1654 sc->status = SMBDIRECT_SOCKET_DESTROYED; 1655 1656 destroy_workqueue(sc->workqueue); 1657 log_rdma_event(INFO, "rdma session destroyed\n"); 1658 kfree(info); 1659 server->smbd_conn = NULL; 1660 } 1661 1662 /* 1663 * Reconnect this SMBD connection, called from upper layer 1664 * return value: 0 on success, or actual error code 1665 */ 1666 int smbd_reconnect(struct TCP_Server_Info *server) 1667 { 1668 log_rdma_event(INFO, "reconnecting rdma session\n"); 1669 1670 if (!server->smbd_conn) { 1671 log_rdma_event(INFO, "rdma session already destroyed\n"); 1672 goto create_conn; 1673 } 1674 1675 /* 1676 * This is possible if transport is disconnected and we haven't received 1677 * notification from RDMA, but upper layer has detected timeout 1678 */ 1679 if (server->smbd_conn->socket.status == SMBDIRECT_SOCKET_CONNECTED) { 1680 log_rdma_event(INFO, "disconnecting transport\n"); 1681 smbd_destroy(server); 1682 } 1683 1684 create_conn: 1685 log_rdma_event(INFO, "creating rdma session\n"); 1686 server->smbd_conn = smbd_get_connection( 1687 server, (struct sockaddr *) &server->dstaddr); 1688 1689 if (server->smbd_conn) { 1690 cifs_dbg(VFS, "RDMA transport re-established\n"); 1691 trace_smb3_smbd_connect_done(server->hostname, server->conn_id, &server->dstaddr); 1692 return 0; 1693 } 1694 trace_smb3_smbd_connect_err(server->hostname, server->conn_id, &server->dstaddr); 1695 return -ENOENT; 1696 } 1697 1698 static void destroy_caches(struct smbdirect_socket *sc) 1699 { 1700 destroy_receive_buffers(sc); 1701 mempool_destroy(sc->recv_io.mem.pool); 1702 kmem_cache_destroy(sc->recv_io.mem.cache); 1703 mempool_destroy(sc->send_io.mem.pool); 1704 kmem_cache_destroy(sc->send_io.mem.cache); 1705 } 1706 1707 #define MAX_NAME_LEN 80 1708 static int allocate_caches(struct smbdirect_socket *sc) 1709 { 1710 struct smbdirect_socket_parameters *sp = &sc->parameters; 1711 char name[MAX_NAME_LEN]; 1712 int rc; 1713 1714 if (WARN_ON_ONCE(sp->max_recv_size < sizeof(struct smbdirect_data_transfer))) 1715 return -ENOMEM; 1716 1717 scnprintf(name, MAX_NAME_LEN, "smbdirect_send_io_%p", sc); 1718 sc->send_io.mem.cache = 1719 kmem_cache_create( 1720 name, 1721 sizeof(struct smbdirect_send_io) + 1722 sizeof(struct smbdirect_data_transfer), 1723 0, SLAB_HWCACHE_ALIGN, NULL); 1724 if (!sc->send_io.mem.cache) 1725 return -ENOMEM; 1726 1727 sc->send_io.mem.pool = 1728 mempool_create(sp->send_credit_target, mempool_alloc_slab, 1729 mempool_free_slab, sc->send_io.mem.cache); 1730 if (!sc->send_io.mem.pool) 1731 goto out1; 1732 1733 scnprintf(name, MAX_NAME_LEN, "smbdirect_recv_io_%p", sc); 1734 1735 struct kmem_cache_args response_args = { 1736 .align = __alignof__(struct smbdirect_recv_io), 1737 .useroffset = (offsetof(struct smbdirect_recv_io, packet) + 1738 sizeof(struct smbdirect_data_transfer)), 1739 .usersize = sp->max_recv_size - sizeof(struct smbdirect_data_transfer), 1740 }; 1741 sc->recv_io.mem.cache = 1742 kmem_cache_create(name, 1743 sizeof(struct smbdirect_recv_io) + sp->max_recv_size, 1744 &response_args, SLAB_HWCACHE_ALIGN); 1745 if (!sc->recv_io.mem.cache) 1746 goto out2; 1747 1748 sc->recv_io.mem.pool = 1749 mempool_create(sp->recv_credit_max, mempool_alloc_slab, 1750 mempool_free_slab, sc->recv_io.mem.cache); 1751 if (!sc->recv_io.mem.pool) 1752 goto out3; 1753 1754 rc = allocate_receive_buffers(sc, sp->recv_credit_max); 1755 if (rc) { 1756 log_rdma_event(ERR, "failed to allocate receive buffers\n"); 1757 goto out4; 1758 } 1759 1760 return 0; 1761 1762 out4: 1763 mempool_destroy(sc->recv_io.mem.pool); 1764 out3: 1765 kmem_cache_destroy(sc->recv_io.mem.cache); 1766 out2: 1767 mempool_destroy(sc->send_io.mem.pool); 1768 out1: 1769 kmem_cache_destroy(sc->send_io.mem.cache); 1770 return -ENOMEM; 1771 } 1772 1773 /* Create a SMBD connection, called by upper layer */ 1774 static struct smbd_connection *_smbd_get_connection( 1775 struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port) 1776 { 1777 int rc; 1778 struct smbd_connection *info; 1779 struct smbdirect_socket *sc; 1780 struct smbdirect_socket_parameters *sp; 1781 struct rdma_conn_param conn_param; 1782 struct ib_qp_init_attr qp_attr; 1783 struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr; 1784 struct ib_port_immutable port_immutable; 1785 __be32 ird_ord_hdr[2]; 1786 char wq_name[80]; 1787 struct workqueue_struct *workqueue; 1788 1789 info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL); 1790 if (!info) 1791 return NULL; 1792 sc = &info->socket; 1793 scnprintf(wq_name, ARRAY_SIZE(wq_name), "smbd_%p", sc); 1794 workqueue = create_workqueue(wq_name); 1795 if (!workqueue) 1796 goto create_wq_failed; 1797 smbdirect_socket_init(sc); 1798 sc->workqueue = workqueue; 1799 sp = &sc->parameters; 1800 1801 INIT_WORK(&sc->disconnect_work, smbd_disconnect_rdma_work); 1802 1803 sp->resolve_addr_timeout_msec = RDMA_RESOLVE_TIMEOUT; 1804 sp->resolve_route_timeout_msec = RDMA_RESOLVE_TIMEOUT; 1805 sp->rdma_connect_timeout_msec = RDMA_RESOLVE_TIMEOUT; 1806 sp->negotiate_timeout_msec = SMBD_NEGOTIATE_TIMEOUT * 1000; 1807 sp->initiator_depth = 1; 1808 sp->responder_resources = SMBD_CM_RESPONDER_RESOURCES; 1809 sp->recv_credit_max = smbd_receive_credit_max; 1810 sp->send_credit_target = smbd_send_credit_target; 1811 sp->max_send_size = smbd_max_send_size; 1812 sp->max_fragmented_recv_size = smbd_max_fragmented_recv_size; 1813 sp->max_recv_size = smbd_max_receive_size; 1814 sp->max_frmr_depth = smbd_max_frmr_depth; 1815 sp->keepalive_interval_msec = smbd_keep_alive_interval * 1000; 1816 sp->keepalive_timeout_msec = KEEPALIVE_RECV_TIMEOUT * 1000; 1817 1818 rc = smbd_ia_open(sc, dstaddr, port); 1819 if (rc) { 1820 log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc); 1821 goto create_id_failed; 1822 } 1823 1824 if (sp->send_credit_target > sc->ib.dev->attrs.max_cqe || 1825 sp->send_credit_target > sc->ib.dev->attrs.max_qp_wr) { 1826 log_rdma_event(ERR, "consider lowering send_credit_target = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n", 1827 sp->send_credit_target, 1828 sc->ib.dev->attrs.max_cqe, 1829 sc->ib.dev->attrs.max_qp_wr); 1830 goto config_failed; 1831 } 1832 1833 if (sp->recv_credit_max > sc->ib.dev->attrs.max_cqe || 1834 sp->recv_credit_max > sc->ib.dev->attrs.max_qp_wr) { 1835 log_rdma_event(ERR, "consider lowering receive_credit_max = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n", 1836 sp->recv_credit_max, 1837 sc->ib.dev->attrs.max_cqe, 1838 sc->ib.dev->attrs.max_qp_wr); 1839 goto config_failed; 1840 } 1841 1842 if (sc->ib.dev->attrs.max_send_sge < SMBDIRECT_SEND_IO_MAX_SGE || 1843 sc->ib.dev->attrs.max_recv_sge < SMBDIRECT_RECV_IO_MAX_SGE) { 1844 log_rdma_event(ERR, 1845 "device %.*s max_send_sge/max_recv_sge = %d/%d too small\n", 1846 IB_DEVICE_NAME_MAX, 1847 sc->ib.dev->name, 1848 sc->ib.dev->attrs.max_send_sge, 1849 sc->ib.dev->attrs.max_recv_sge); 1850 goto config_failed; 1851 } 1852 1853 sc->ib.pd = ib_alloc_pd(sc->ib.dev, 0); 1854 if (IS_ERR(sc->ib.pd)) { 1855 rc = PTR_ERR(sc->ib.pd); 1856 sc->ib.pd = NULL; 1857 log_rdma_event(ERR, "ib_alloc_pd() returned %d\n", rc); 1858 goto alloc_pd_failed; 1859 } 1860 1861 sc->ib.send_cq = 1862 ib_alloc_cq_any(sc->ib.dev, sc, 1863 sp->send_credit_target, IB_POLL_SOFTIRQ); 1864 if (IS_ERR(sc->ib.send_cq)) { 1865 sc->ib.send_cq = NULL; 1866 goto alloc_cq_failed; 1867 } 1868 1869 sc->ib.recv_cq = 1870 ib_alloc_cq_any(sc->ib.dev, sc, 1871 sp->recv_credit_max, IB_POLL_SOFTIRQ); 1872 if (IS_ERR(sc->ib.recv_cq)) { 1873 sc->ib.recv_cq = NULL; 1874 goto alloc_cq_failed; 1875 } 1876 1877 memset(&qp_attr, 0, sizeof(qp_attr)); 1878 qp_attr.event_handler = smbd_qp_async_error_upcall; 1879 qp_attr.qp_context = sc; 1880 qp_attr.cap.max_send_wr = sp->send_credit_target; 1881 qp_attr.cap.max_recv_wr = sp->recv_credit_max; 1882 qp_attr.cap.max_send_sge = SMBDIRECT_SEND_IO_MAX_SGE; 1883 qp_attr.cap.max_recv_sge = SMBDIRECT_RECV_IO_MAX_SGE; 1884 qp_attr.cap.max_inline_data = 0; 1885 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 1886 qp_attr.qp_type = IB_QPT_RC; 1887 qp_attr.send_cq = sc->ib.send_cq; 1888 qp_attr.recv_cq = sc->ib.recv_cq; 1889 qp_attr.port_num = ~0; 1890 1891 rc = rdma_create_qp(sc->rdma.cm_id, sc->ib.pd, &qp_attr); 1892 if (rc) { 1893 log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc); 1894 goto create_qp_failed; 1895 } 1896 sc->ib.qp = sc->rdma.cm_id->qp; 1897 1898 sp->responder_resources = 1899 min_t(u8, sp->responder_resources, 1900 sc->ib.dev->attrs.max_qp_rd_atom); 1901 log_rdma_mr(INFO, "responder_resources=%d\n", 1902 sp->responder_resources); 1903 1904 memset(&conn_param, 0, sizeof(conn_param)); 1905 conn_param.initiator_depth = sp->initiator_depth; 1906 conn_param.responder_resources = sp->responder_resources; 1907 1908 /* Need to send IRD/ORD in private data for iWARP */ 1909 sc->ib.dev->ops.get_port_immutable( 1910 sc->ib.dev, sc->rdma.cm_id->port_num, &port_immutable); 1911 if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) { 1912 ird_ord_hdr[0] = cpu_to_be32(conn_param.responder_resources); 1913 ird_ord_hdr[1] = cpu_to_be32(conn_param.initiator_depth); 1914 conn_param.private_data = ird_ord_hdr; 1915 conn_param.private_data_len = sizeof(ird_ord_hdr); 1916 } else { 1917 conn_param.private_data = NULL; 1918 conn_param.private_data_len = 0; 1919 } 1920 1921 conn_param.retry_count = SMBD_CM_RETRY; 1922 conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY; 1923 conn_param.flow_control = 0; 1924 1925 log_rdma_event(INFO, "connecting to IP %pI4 port %d\n", 1926 &addr_in->sin_addr, port); 1927 1928 WARN_ON_ONCE(sc->status != SMBDIRECT_SOCKET_RDMA_CONNECT_NEEDED); 1929 sc->status = SMBDIRECT_SOCKET_RDMA_CONNECT_RUNNING; 1930 rc = rdma_connect(sc->rdma.cm_id, &conn_param); 1931 if (rc) { 1932 log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc); 1933 goto rdma_connect_failed; 1934 } 1935 1936 wait_event_interruptible_timeout( 1937 sc->status_wait, 1938 sc->status != SMBDIRECT_SOCKET_RDMA_CONNECT_RUNNING, 1939 msecs_to_jiffies(sp->rdma_connect_timeout_msec)); 1940 1941 if (sc->status != SMBDIRECT_SOCKET_NEGOTIATE_NEEDED) { 1942 log_rdma_event(ERR, "rdma_connect failed port=%d\n", port); 1943 goto rdma_connect_failed; 1944 } 1945 1946 log_rdma_event(INFO, "rdma_connect connected\n"); 1947 1948 rc = allocate_caches(sc); 1949 if (rc) { 1950 log_rdma_event(ERR, "cache allocation failed\n"); 1951 goto allocate_cache_failed; 1952 } 1953 1954 INIT_WORK(&sc->idle.immediate_work, send_immediate_empty_message); 1955 INIT_DELAYED_WORK(&sc->idle.timer_work, idle_connection_timer); 1956 /* 1957 * start with the negotiate timeout and SMBDIRECT_KEEPALIVE_PENDING 1958 * so that the timer will cause a disconnect. 1959 */ 1960 sc->idle.keepalive = SMBDIRECT_KEEPALIVE_PENDING; 1961 mod_delayed_work(sc->workqueue, &sc->idle.timer_work, 1962 msecs_to_jiffies(sp->negotiate_timeout_msec)); 1963 1964 INIT_WORK(&sc->recv_io.posted.refill_work, smbd_post_send_credits); 1965 1966 rc = smbd_negotiate(sc); 1967 if (rc) { 1968 log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc); 1969 goto negotiation_failed; 1970 } 1971 1972 rc = allocate_mr_list(sc); 1973 if (rc) { 1974 log_rdma_mr(ERR, "memory registration allocation failed\n"); 1975 goto allocate_mr_failed; 1976 } 1977 1978 return info; 1979 1980 allocate_mr_failed: 1981 /* At this point, need to a full transport shutdown */ 1982 server->smbd_conn = info; 1983 smbd_destroy(server); 1984 return NULL; 1985 1986 negotiation_failed: 1987 disable_delayed_work_sync(&sc->idle.timer_work); 1988 destroy_caches(sc); 1989 sc->status = SMBDIRECT_SOCKET_NEGOTIATE_FAILED; 1990 rdma_disconnect(sc->rdma.cm_id); 1991 wait_event(sc->status_wait, 1992 sc->status == SMBDIRECT_SOCKET_DISCONNECTED); 1993 1994 allocate_cache_failed: 1995 rdma_connect_failed: 1996 rdma_destroy_qp(sc->rdma.cm_id); 1997 1998 create_qp_failed: 1999 alloc_cq_failed: 2000 if (sc->ib.send_cq) 2001 ib_free_cq(sc->ib.send_cq); 2002 if (sc->ib.recv_cq) 2003 ib_free_cq(sc->ib.recv_cq); 2004 2005 ib_dealloc_pd(sc->ib.pd); 2006 2007 alloc_pd_failed: 2008 config_failed: 2009 rdma_destroy_id(sc->rdma.cm_id); 2010 2011 create_id_failed: 2012 destroy_workqueue(sc->workqueue); 2013 create_wq_failed: 2014 kfree(info); 2015 return NULL; 2016 } 2017 2018 struct smbd_connection *smbd_get_connection( 2019 struct TCP_Server_Info *server, struct sockaddr *dstaddr) 2020 { 2021 struct smbd_connection *ret; 2022 const struct smbdirect_socket_parameters *sp; 2023 int port = SMBD_PORT; 2024 2025 try_again: 2026 ret = _smbd_get_connection(server, dstaddr, port); 2027 2028 /* Try SMB_PORT if SMBD_PORT doesn't work */ 2029 if (!ret && port == SMBD_PORT) { 2030 port = SMB_PORT; 2031 goto try_again; 2032 } 2033 if (!ret) 2034 return NULL; 2035 2036 sp = &ret->socket.parameters; 2037 2038 server->rdma_readwrite_threshold = 2039 rdma_readwrite_threshold > sp->max_fragmented_send_size ? 2040 sp->max_fragmented_send_size : 2041 rdma_readwrite_threshold; 2042 2043 return ret; 2044 } 2045 2046 /* 2047 * Receive data from the transport's receive reassembly queue 2048 * All the incoming data packets are placed in reassembly queue 2049 * iter: the buffer to read data into 2050 * size: the length of data to read 2051 * return value: actual data read 2052 * 2053 * Note: this implementation copies the data from reassembly queue to receive 2054 * buffers used by upper layer. This is not the optimal code path. A better way 2055 * to do it is to not have upper layer allocate its receive buffers but rather 2056 * borrow the buffer from reassembly queue, and return it after data is 2057 * consumed. But this will require more changes to upper layer code, and also 2058 * need to consider packet boundaries while they still being reassembled. 2059 */ 2060 int smbd_recv(struct smbd_connection *info, struct msghdr *msg) 2061 { 2062 struct smbdirect_socket *sc = &info->socket; 2063 struct smbdirect_recv_io *response; 2064 struct smbdirect_data_transfer *data_transfer; 2065 size_t size = iov_iter_count(&msg->msg_iter); 2066 int to_copy, to_read, data_read, offset; 2067 u32 data_length, remaining_data_length, data_offset; 2068 int rc; 2069 2070 if (WARN_ON_ONCE(iov_iter_rw(&msg->msg_iter) == WRITE)) 2071 return -EINVAL; /* It's a bug in upper layer to get there */ 2072 2073 again: 2074 /* 2075 * No need to hold the reassembly queue lock all the time as we are 2076 * the only one reading from the front of the queue. The transport 2077 * may add more entries to the back of the queue at the same time 2078 */ 2079 log_read(INFO, "size=%zd sc->recv_io.reassembly.data_length=%d\n", size, 2080 sc->recv_io.reassembly.data_length); 2081 if (sc->recv_io.reassembly.data_length >= size) { 2082 int queue_length; 2083 int queue_removed = 0; 2084 unsigned long flags; 2085 2086 /* 2087 * Need to make sure reassembly_data_length is read before 2088 * reading reassembly_queue_length and calling 2089 * _get_first_reassembly. This call is lock free 2090 * as we never read at the end of the queue which are being 2091 * updated in SOFTIRQ as more data is received 2092 */ 2093 virt_rmb(); 2094 queue_length = sc->recv_io.reassembly.queue_length; 2095 data_read = 0; 2096 to_read = size; 2097 offset = sc->recv_io.reassembly.first_entry_offset; 2098 while (data_read < size) { 2099 response = _get_first_reassembly(sc); 2100 data_transfer = smbdirect_recv_io_payload(response); 2101 data_length = le32_to_cpu(data_transfer->data_length); 2102 remaining_data_length = 2103 le32_to_cpu( 2104 data_transfer->remaining_data_length); 2105 data_offset = le32_to_cpu(data_transfer->data_offset); 2106 2107 /* 2108 * The upper layer expects RFC1002 length at the 2109 * beginning of the payload. Return it to indicate 2110 * the total length of the packet. This minimize the 2111 * change to upper layer packet processing logic. This 2112 * will be eventually remove when an intermediate 2113 * transport layer is added 2114 */ 2115 if (response->first_segment && size == 4) { 2116 unsigned int rfc1002_len = 2117 data_length + remaining_data_length; 2118 __be32 rfc1002_hdr = cpu_to_be32(rfc1002_len); 2119 if (copy_to_iter(&rfc1002_hdr, sizeof(rfc1002_hdr), 2120 &msg->msg_iter) != sizeof(rfc1002_hdr)) 2121 return -EFAULT; 2122 data_read = 4; 2123 response->first_segment = false; 2124 log_read(INFO, "returning rfc1002 length %d\n", 2125 rfc1002_len); 2126 goto read_rfc1002_done; 2127 } 2128 2129 to_copy = min_t(int, data_length - offset, to_read); 2130 if (copy_to_iter((char *)data_transfer + data_offset + offset, 2131 to_copy, &msg->msg_iter) != to_copy) 2132 return -EFAULT; 2133 2134 /* move on to the next buffer? */ 2135 if (to_copy == data_length - offset) { 2136 queue_length--; 2137 /* 2138 * No need to lock if we are not at the 2139 * end of the queue 2140 */ 2141 if (queue_length) 2142 list_del(&response->list); 2143 else { 2144 spin_lock_irqsave( 2145 &sc->recv_io.reassembly.lock, flags); 2146 list_del(&response->list); 2147 spin_unlock_irqrestore( 2148 &sc->recv_io.reassembly.lock, flags); 2149 } 2150 queue_removed++; 2151 sc->statistics.dequeue_reassembly_queue++; 2152 put_receive_buffer(sc, response); 2153 offset = 0; 2154 log_read(INFO, "put_receive_buffer offset=0\n"); 2155 } else 2156 offset += to_copy; 2157 2158 to_read -= to_copy; 2159 data_read += to_copy; 2160 2161 log_read(INFO, "_get_first_reassembly memcpy %d bytes data_transfer_length-offset=%d after that to_read=%d data_read=%d offset=%d\n", 2162 to_copy, data_length - offset, 2163 to_read, data_read, offset); 2164 } 2165 2166 spin_lock_irqsave(&sc->recv_io.reassembly.lock, flags); 2167 sc->recv_io.reassembly.data_length -= data_read; 2168 sc->recv_io.reassembly.queue_length -= queue_removed; 2169 spin_unlock_irqrestore(&sc->recv_io.reassembly.lock, flags); 2170 2171 sc->recv_io.reassembly.first_entry_offset = offset; 2172 log_read(INFO, "returning to thread data_read=%d reassembly_data_length=%d first_entry_offset=%d\n", 2173 data_read, sc->recv_io.reassembly.data_length, 2174 sc->recv_io.reassembly.first_entry_offset); 2175 read_rfc1002_done: 2176 return data_read; 2177 } 2178 2179 log_read(INFO, "wait_event on more data\n"); 2180 rc = wait_event_interruptible( 2181 sc->recv_io.reassembly.wait_queue, 2182 sc->recv_io.reassembly.data_length >= size || 2183 sc->status != SMBDIRECT_SOCKET_CONNECTED); 2184 /* Don't return any data if interrupted */ 2185 if (rc) 2186 return rc; 2187 2188 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) { 2189 log_read(ERR, "disconnected\n"); 2190 return -ECONNABORTED; 2191 } 2192 2193 goto again; 2194 } 2195 2196 /* 2197 * Send data to transport 2198 * Each rqst is transported as a SMBDirect payload 2199 * rqst: the data to write 2200 * return value: 0 if successfully write, otherwise error code 2201 */ 2202 int smbd_send(struct TCP_Server_Info *server, 2203 int num_rqst, struct smb_rqst *rqst_array) 2204 { 2205 struct smbd_connection *info = server->smbd_conn; 2206 struct smbdirect_socket *sc = &info->socket; 2207 struct smbdirect_socket_parameters *sp = &sc->parameters; 2208 struct smb_rqst *rqst; 2209 struct iov_iter iter; 2210 unsigned int remaining_data_length, klen; 2211 int rc, i, rqst_idx; 2212 2213 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) 2214 return -EAGAIN; 2215 2216 /* 2217 * Add in the page array if there is one. The caller needs to set 2218 * rq_tailsz to PAGE_SIZE when the buffer has multiple pages and 2219 * ends at page boundary 2220 */ 2221 remaining_data_length = 0; 2222 for (i = 0; i < num_rqst; i++) 2223 remaining_data_length += smb_rqst_len(server, &rqst_array[i]); 2224 2225 if (unlikely(remaining_data_length > sp->max_fragmented_send_size)) { 2226 /* assertion: payload never exceeds negotiated maximum */ 2227 log_write(ERR, "payload size %d > max size %d\n", 2228 remaining_data_length, sp->max_fragmented_send_size); 2229 return -EINVAL; 2230 } 2231 2232 log_write(INFO, "num_rqst=%d total length=%u\n", 2233 num_rqst, remaining_data_length); 2234 2235 rqst_idx = 0; 2236 do { 2237 rqst = &rqst_array[rqst_idx]; 2238 2239 cifs_dbg(FYI, "Sending smb (RDMA): idx=%d smb_len=%lu\n", 2240 rqst_idx, smb_rqst_len(server, rqst)); 2241 for (i = 0; i < rqst->rq_nvec; i++) 2242 dump_smb(rqst->rq_iov[i].iov_base, rqst->rq_iov[i].iov_len); 2243 2244 log_write(INFO, "RDMA-WR[%u] nvec=%d len=%u iter=%zu rqlen=%lu\n", 2245 rqst_idx, rqst->rq_nvec, remaining_data_length, 2246 iov_iter_count(&rqst->rq_iter), smb_rqst_len(server, rqst)); 2247 2248 /* Send the metadata pages. */ 2249 klen = 0; 2250 for (i = 0; i < rqst->rq_nvec; i++) 2251 klen += rqst->rq_iov[i].iov_len; 2252 iov_iter_kvec(&iter, ITER_SOURCE, rqst->rq_iov, rqst->rq_nvec, klen); 2253 2254 rc = smbd_post_send_full_iter(sc, &iter, &remaining_data_length); 2255 if (rc < 0) 2256 break; 2257 2258 if (iov_iter_count(&rqst->rq_iter) > 0) { 2259 /* And then the data pages if there are any */ 2260 rc = smbd_post_send_full_iter(sc, &rqst->rq_iter, 2261 &remaining_data_length); 2262 if (rc < 0) 2263 break; 2264 } 2265 2266 } while (++rqst_idx < num_rqst); 2267 2268 /* 2269 * As an optimization, we don't wait for individual I/O to finish 2270 * before sending the next one. 2271 * Send them all and wait for pending send count to get to 0 2272 * that means all the I/Os have been out and we are good to return 2273 */ 2274 2275 wait_event(sc->send_io.pending.zero_wait_queue, 2276 atomic_read(&sc->send_io.pending.count) == 0 || 2277 sc->status != SMBDIRECT_SOCKET_CONNECTED); 2278 2279 if (sc->status != SMBDIRECT_SOCKET_CONNECTED && rc == 0) 2280 rc = -EAGAIN; 2281 2282 return rc; 2283 } 2284 2285 static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc) 2286 { 2287 struct smbdirect_mr_io *mr = 2288 container_of(wc->wr_cqe, struct smbdirect_mr_io, cqe); 2289 struct smbdirect_socket *sc = mr->socket; 2290 2291 if (wc->status) { 2292 log_rdma_mr(ERR, "status=%d\n", wc->status); 2293 smbd_disconnect_rdma_connection(sc); 2294 } 2295 } 2296 2297 /* 2298 * The work queue function that recovers MRs 2299 * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used 2300 * again. Both calls are slow, so finish them in a workqueue. This will not 2301 * block I/O path. 2302 * There is one workqueue that recovers MRs, there is no need to lock as the 2303 * I/O requests calling smbd_register_mr will never update the links in the 2304 * mr_list. 2305 */ 2306 static void smbd_mr_recovery_work(struct work_struct *work) 2307 { 2308 struct smbdirect_socket *sc = 2309 container_of(work, struct smbdirect_socket, mr_io.recovery_work); 2310 struct smbdirect_socket_parameters *sp = &sc->parameters; 2311 struct smbdirect_mr_io *smbdirect_mr; 2312 int rc; 2313 2314 list_for_each_entry(smbdirect_mr, &sc->mr_io.all.list, list) { 2315 if (smbdirect_mr->state == SMBDIRECT_MR_ERROR) { 2316 2317 /* recover this MR entry */ 2318 rc = ib_dereg_mr(smbdirect_mr->mr); 2319 if (rc) { 2320 log_rdma_mr(ERR, 2321 "ib_dereg_mr failed rc=%x\n", 2322 rc); 2323 smbd_disconnect_rdma_connection(sc); 2324 continue; 2325 } 2326 2327 smbdirect_mr->mr = ib_alloc_mr( 2328 sc->ib.pd, sc->mr_io.type, 2329 sp->max_frmr_depth); 2330 if (IS_ERR(smbdirect_mr->mr)) { 2331 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n", 2332 sc->mr_io.type, 2333 sp->max_frmr_depth); 2334 smbd_disconnect_rdma_connection(sc); 2335 continue; 2336 } 2337 } else 2338 /* This MR is being used, don't recover it */ 2339 continue; 2340 2341 smbdirect_mr->state = SMBDIRECT_MR_READY; 2342 2343 /* smbdirect_mr->state is updated by this function 2344 * and is read and updated by I/O issuing CPUs trying 2345 * to get a MR, the call to atomic_inc_return 2346 * implicates a memory barrier and guarantees this 2347 * value is updated before waking up any calls to 2348 * get_mr() from the I/O issuing CPUs 2349 */ 2350 if (atomic_inc_return(&sc->mr_io.ready.count) == 1) 2351 wake_up(&sc->mr_io.ready.wait_queue); 2352 } 2353 } 2354 2355 static void destroy_mr_list(struct smbdirect_socket *sc) 2356 { 2357 struct smbdirect_mr_io *mr, *tmp; 2358 2359 disable_work_sync(&sc->mr_io.recovery_work); 2360 list_for_each_entry_safe(mr, tmp, &sc->mr_io.all.list, list) { 2361 if (mr->state == SMBDIRECT_MR_INVALIDATED) 2362 ib_dma_unmap_sg(sc->ib.dev, mr->sgt.sgl, 2363 mr->sgt.nents, mr->dir); 2364 ib_dereg_mr(mr->mr); 2365 kfree(mr->sgt.sgl); 2366 kfree(mr); 2367 } 2368 } 2369 2370 /* 2371 * Allocate MRs used for RDMA read/write 2372 * The number of MRs will not exceed hardware capability in responder_resources 2373 * All MRs are kept in mr_list. The MR can be recovered after it's used 2374 * Recovery is done in smbd_mr_recovery_work. The content of list entry changes 2375 * as MRs are used and recovered for I/O, but the list links will not change 2376 */ 2377 static int allocate_mr_list(struct smbdirect_socket *sc) 2378 { 2379 struct smbdirect_socket_parameters *sp = &sc->parameters; 2380 int i; 2381 struct smbdirect_mr_io *smbdirect_mr, *tmp; 2382 2383 INIT_WORK(&sc->mr_io.recovery_work, smbd_mr_recovery_work); 2384 2385 if (sp->responder_resources == 0) { 2386 log_rdma_mr(ERR, "responder_resources negotiated as 0\n"); 2387 return -EINVAL; 2388 } 2389 2390 /* Allocate more MRs (2x) than hardware responder_resources */ 2391 for (i = 0; i < sp->responder_resources * 2; i++) { 2392 smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL); 2393 if (!smbdirect_mr) 2394 goto cleanup_entries; 2395 smbdirect_mr->mr = ib_alloc_mr(sc->ib.pd, sc->mr_io.type, 2396 sp->max_frmr_depth); 2397 if (IS_ERR(smbdirect_mr->mr)) { 2398 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n", 2399 sc->mr_io.type, sp->max_frmr_depth); 2400 goto out; 2401 } 2402 smbdirect_mr->sgt.sgl = kcalloc(sp->max_frmr_depth, 2403 sizeof(struct scatterlist), 2404 GFP_KERNEL); 2405 if (!smbdirect_mr->sgt.sgl) { 2406 log_rdma_mr(ERR, "failed to allocate sgl\n"); 2407 ib_dereg_mr(smbdirect_mr->mr); 2408 goto out; 2409 } 2410 smbdirect_mr->state = SMBDIRECT_MR_READY; 2411 smbdirect_mr->socket = sc; 2412 2413 list_add_tail(&smbdirect_mr->list, &sc->mr_io.all.list); 2414 atomic_inc(&sc->mr_io.ready.count); 2415 } 2416 return 0; 2417 2418 out: 2419 kfree(smbdirect_mr); 2420 cleanup_entries: 2421 list_for_each_entry_safe(smbdirect_mr, tmp, &sc->mr_io.all.list, list) { 2422 list_del(&smbdirect_mr->list); 2423 ib_dereg_mr(smbdirect_mr->mr); 2424 kfree(smbdirect_mr->sgt.sgl); 2425 kfree(smbdirect_mr); 2426 } 2427 return -ENOMEM; 2428 } 2429 2430 /* 2431 * Get a MR from mr_list. This function waits until there is at least one 2432 * MR available in the list. It may access the list while the 2433 * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock 2434 * as they never modify the same places. However, there may be several CPUs 2435 * issuing I/O trying to get MR at the same time, mr_list_lock is used to 2436 * protect this situation. 2437 */ 2438 static struct smbdirect_mr_io *get_mr(struct smbdirect_socket *sc) 2439 { 2440 struct smbdirect_mr_io *ret; 2441 unsigned long flags; 2442 int rc; 2443 again: 2444 rc = wait_event_interruptible(sc->mr_io.ready.wait_queue, 2445 atomic_read(&sc->mr_io.ready.count) || 2446 sc->status != SMBDIRECT_SOCKET_CONNECTED); 2447 if (rc) { 2448 log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc); 2449 return NULL; 2450 } 2451 2452 if (sc->status != SMBDIRECT_SOCKET_CONNECTED) { 2453 log_rdma_mr(ERR, "sc->status=%x\n", sc->status); 2454 return NULL; 2455 } 2456 2457 spin_lock_irqsave(&sc->mr_io.all.lock, flags); 2458 list_for_each_entry(ret, &sc->mr_io.all.list, list) { 2459 if (ret->state == SMBDIRECT_MR_READY) { 2460 ret->state = SMBDIRECT_MR_REGISTERED; 2461 spin_unlock_irqrestore(&sc->mr_io.all.lock, flags); 2462 atomic_dec(&sc->mr_io.ready.count); 2463 atomic_inc(&sc->mr_io.used.count); 2464 return ret; 2465 } 2466 } 2467 2468 spin_unlock_irqrestore(&sc->mr_io.all.lock, flags); 2469 /* 2470 * It is possible that we could fail to get MR because other processes may 2471 * try to acquire a MR at the same time. If this is the case, retry it. 2472 */ 2473 goto again; 2474 } 2475 2476 /* 2477 * Transcribe the pages from an iterator into an MR scatterlist. 2478 */ 2479 static int smbd_iter_to_mr(struct iov_iter *iter, 2480 struct sg_table *sgt, 2481 unsigned int max_sg) 2482 { 2483 int ret; 2484 2485 memset(sgt->sgl, 0, max_sg * sizeof(struct scatterlist)); 2486 2487 ret = extract_iter_to_sg(iter, iov_iter_count(iter), sgt, max_sg, 0); 2488 WARN_ON(ret < 0); 2489 if (sgt->nents > 0) 2490 sg_mark_end(&sgt->sgl[sgt->nents - 1]); 2491 return ret; 2492 } 2493 2494 /* 2495 * Register memory for RDMA read/write 2496 * iter: the buffer to register memory with 2497 * writing: true if this is a RDMA write (SMB read), false for RDMA read 2498 * need_invalidate: true if this MR needs to be locally invalidated after I/O 2499 * return value: the MR registered, NULL if failed. 2500 */ 2501 struct smbdirect_mr_io *smbd_register_mr(struct smbd_connection *info, 2502 struct iov_iter *iter, 2503 bool writing, bool need_invalidate) 2504 { 2505 struct smbdirect_socket *sc = &info->socket; 2506 struct smbdirect_socket_parameters *sp = &sc->parameters; 2507 struct smbdirect_mr_io *smbdirect_mr; 2508 int rc, num_pages; 2509 enum dma_data_direction dir; 2510 struct ib_reg_wr *reg_wr; 2511 2512 num_pages = iov_iter_npages(iter, sp->max_frmr_depth + 1); 2513 if (num_pages > sp->max_frmr_depth) { 2514 log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n", 2515 num_pages, sp->max_frmr_depth); 2516 WARN_ON_ONCE(1); 2517 return NULL; 2518 } 2519 2520 smbdirect_mr = get_mr(sc); 2521 if (!smbdirect_mr) { 2522 log_rdma_mr(ERR, "get_mr returning NULL\n"); 2523 return NULL; 2524 } 2525 2526 dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE; 2527 smbdirect_mr->dir = dir; 2528 smbdirect_mr->need_invalidate = need_invalidate; 2529 smbdirect_mr->sgt.nents = 0; 2530 smbdirect_mr->sgt.orig_nents = 0; 2531 2532 log_rdma_mr(INFO, "num_pages=0x%x count=0x%zx depth=%u\n", 2533 num_pages, iov_iter_count(iter), sp->max_frmr_depth); 2534 smbd_iter_to_mr(iter, &smbdirect_mr->sgt, sp->max_frmr_depth); 2535 2536 rc = ib_dma_map_sg(sc->ib.dev, smbdirect_mr->sgt.sgl, 2537 smbdirect_mr->sgt.nents, dir); 2538 if (!rc) { 2539 log_rdma_mr(ERR, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n", 2540 num_pages, dir, rc); 2541 goto dma_map_error; 2542 } 2543 2544 rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgt.sgl, 2545 smbdirect_mr->sgt.nents, NULL, PAGE_SIZE); 2546 if (rc != smbdirect_mr->sgt.nents) { 2547 log_rdma_mr(ERR, 2548 "ib_map_mr_sg failed rc = %d nents = %x\n", 2549 rc, smbdirect_mr->sgt.nents); 2550 goto map_mr_error; 2551 } 2552 2553 ib_update_fast_reg_key(smbdirect_mr->mr, 2554 ib_inc_rkey(smbdirect_mr->mr->rkey)); 2555 reg_wr = &smbdirect_mr->wr; 2556 reg_wr->wr.opcode = IB_WR_REG_MR; 2557 smbdirect_mr->cqe.done = register_mr_done; 2558 reg_wr->wr.wr_cqe = &smbdirect_mr->cqe; 2559 reg_wr->wr.num_sge = 0; 2560 reg_wr->wr.send_flags = IB_SEND_SIGNALED; 2561 reg_wr->mr = smbdirect_mr->mr; 2562 reg_wr->key = smbdirect_mr->mr->rkey; 2563 reg_wr->access = writing ? 2564 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE : 2565 IB_ACCESS_REMOTE_READ; 2566 2567 /* 2568 * There is no need for waiting for complemtion on ib_post_send 2569 * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution 2570 * on the next ib_post_send when we actually send I/O to remote peer 2571 */ 2572 rc = ib_post_send(sc->ib.qp, ®_wr->wr, NULL); 2573 if (!rc) 2574 return smbdirect_mr; 2575 2576 log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n", 2577 rc, reg_wr->key); 2578 2579 /* If all failed, attempt to recover this MR by setting it SMBDIRECT_MR_ERROR*/ 2580 map_mr_error: 2581 ib_dma_unmap_sg(sc->ib.dev, smbdirect_mr->sgt.sgl, 2582 smbdirect_mr->sgt.nents, smbdirect_mr->dir); 2583 2584 dma_map_error: 2585 smbdirect_mr->state = SMBDIRECT_MR_ERROR; 2586 if (atomic_dec_and_test(&sc->mr_io.used.count)) 2587 wake_up(&sc->mr_io.cleanup.wait_queue); 2588 2589 smbd_disconnect_rdma_connection(sc); 2590 2591 return NULL; 2592 } 2593 2594 static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc) 2595 { 2596 struct smbdirect_mr_io *smbdirect_mr; 2597 struct ib_cqe *cqe; 2598 2599 cqe = wc->wr_cqe; 2600 smbdirect_mr = container_of(cqe, struct smbdirect_mr_io, cqe); 2601 smbdirect_mr->state = SMBDIRECT_MR_INVALIDATED; 2602 if (wc->status != IB_WC_SUCCESS) { 2603 log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status); 2604 smbdirect_mr->state = SMBDIRECT_MR_ERROR; 2605 } 2606 complete(&smbdirect_mr->invalidate_done); 2607 } 2608 2609 /* 2610 * Deregister a MR after I/O is done 2611 * This function may wait if remote invalidation is not used 2612 * and we have to locally invalidate the buffer to prevent data is being 2613 * modified by remote peer after upper layer consumes it 2614 */ 2615 int smbd_deregister_mr(struct smbdirect_mr_io *smbdirect_mr) 2616 { 2617 struct ib_send_wr *wr; 2618 struct smbdirect_socket *sc = smbdirect_mr->socket; 2619 int rc = 0; 2620 2621 if (smbdirect_mr->need_invalidate) { 2622 /* Need to finish local invalidation before returning */ 2623 wr = &smbdirect_mr->inv_wr; 2624 wr->opcode = IB_WR_LOCAL_INV; 2625 smbdirect_mr->cqe.done = local_inv_done; 2626 wr->wr_cqe = &smbdirect_mr->cqe; 2627 wr->num_sge = 0; 2628 wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey; 2629 wr->send_flags = IB_SEND_SIGNALED; 2630 2631 init_completion(&smbdirect_mr->invalidate_done); 2632 rc = ib_post_send(sc->ib.qp, wr, NULL); 2633 if (rc) { 2634 log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc); 2635 smbd_disconnect_rdma_connection(sc); 2636 goto done; 2637 } 2638 wait_for_completion(&smbdirect_mr->invalidate_done); 2639 smbdirect_mr->need_invalidate = false; 2640 } else 2641 /* 2642 * For remote invalidation, just set it to SMBDIRECT_MR_INVALIDATED 2643 * and defer to mr_recovery_work to recover the MR for next use 2644 */ 2645 smbdirect_mr->state = SMBDIRECT_MR_INVALIDATED; 2646 2647 if (smbdirect_mr->state == SMBDIRECT_MR_INVALIDATED) { 2648 ib_dma_unmap_sg( 2649 sc->ib.dev, smbdirect_mr->sgt.sgl, 2650 smbdirect_mr->sgt.nents, 2651 smbdirect_mr->dir); 2652 smbdirect_mr->state = SMBDIRECT_MR_READY; 2653 if (atomic_inc_return(&sc->mr_io.ready.count) == 1) 2654 wake_up(&sc->mr_io.ready.wait_queue); 2655 } else 2656 /* 2657 * Schedule the work to do MR recovery for future I/Os MR 2658 * recovery is slow and don't want it to block current I/O 2659 */ 2660 queue_work(sc->workqueue, &sc->mr_io.recovery_work); 2661 2662 done: 2663 if (atomic_dec_and_test(&sc->mr_io.used.count)) 2664 wake_up(&sc->mr_io.cleanup.wait_queue); 2665 2666 return rc; 2667 } 2668 2669 static bool smb_set_sge(struct smb_extract_to_rdma *rdma, 2670 struct page *lowest_page, size_t off, size_t len) 2671 { 2672 struct ib_sge *sge = &rdma->sge[rdma->nr_sge]; 2673 u64 addr; 2674 2675 addr = ib_dma_map_page(rdma->device, lowest_page, 2676 off, len, rdma->direction); 2677 if (ib_dma_mapping_error(rdma->device, addr)) 2678 return false; 2679 2680 sge->addr = addr; 2681 sge->length = len; 2682 sge->lkey = rdma->local_dma_lkey; 2683 rdma->nr_sge++; 2684 return true; 2685 } 2686 2687 /* 2688 * Extract page fragments from a BVEC-class iterator and add them to an RDMA 2689 * element list. The pages are not pinned. 2690 */ 2691 static ssize_t smb_extract_bvec_to_rdma(struct iov_iter *iter, 2692 struct smb_extract_to_rdma *rdma, 2693 ssize_t maxsize) 2694 { 2695 const struct bio_vec *bv = iter->bvec; 2696 unsigned long start = iter->iov_offset; 2697 unsigned int i; 2698 ssize_t ret = 0; 2699 2700 for (i = 0; i < iter->nr_segs; i++) { 2701 size_t off, len; 2702 2703 len = bv[i].bv_len; 2704 if (start >= len) { 2705 start -= len; 2706 continue; 2707 } 2708 2709 len = min_t(size_t, maxsize, len - start); 2710 off = bv[i].bv_offset + start; 2711 2712 if (!smb_set_sge(rdma, bv[i].bv_page, off, len)) 2713 return -EIO; 2714 2715 ret += len; 2716 maxsize -= len; 2717 if (rdma->nr_sge >= rdma->max_sge || maxsize <= 0) 2718 break; 2719 start = 0; 2720 } 2721 2722 if (ret > 0) 2723 iov_iter_advance(iter, ret); 2724 return ret; 2725 } 2726 2727 /* 2728 * Extract fragments from a KVEC-class iterator and add them to an RDMA list. 2729 * This can deal with vmalloc'd buffers as well as kmalloc'd or static buffers. 2730 * The pages are not pinned. 2731 */ 2732 static ssize_t smb_extract_kvec_to_rdma(struct iov_iter *iter, 2733 struct smb_extract_to_rdma *rdma, 2734 ssize_t maxsize) 2735 { 2736 const struct kvec *kv = iter->kvec; 2737 unsigned long start = iter->iov_offset; 2738 unsigned int i; 2739 ssize_t ret = 0; 2740 2741 for (i = 0; i < iter->nr_segs; i++) { 2742 struct page *page; 2743 unsigned long kaddr; 2744 size_t off, len, seg; 2745 2746 len = kv[i].iov_len; 2747 if (start >= len) { 2748 start -= len; 2749 continue; 2750 } 2751 2752 kaddr = (unsigned long)kv[i].iov_base + start; 2753 off = kaddr & ~PAGE_MASK; 2754 len = min_t(size_t, maxsize, len - start); 2755 kaddr &= PAGE_MASK; 2756 2757 maxsize -= len; 2758 do { 2759 seg = min_t(size_t, len, PAGE_SIZE - off); 2760 2761 if (is_vmalloc_or_module_addr((void *)kaddr)) 2762 page = vmalloc_to_page((void *)kaddr); 2763 else 2764 page = virt_to_page((void *)kaddr); 2765 2766 if (!smb_set_sge(rdma, page, off, seg)) 2767 return -EIO; 2768 2769 ret += seg; 2770 len -= seg; 2771 kaddr += PAGE_SIZE; 2772 off = 0; 2773 } while (len > 0 && rdma->nr_sge < rdma->max_sge); 2774 2775 if (rdma->nr_sge >= rdma->max_sge || maxsize <= 0) 2776 break; 2777 start = 0; 2778 } 2779 2780 if (ret > 0) 2781 iov_iter_advance(iter, ret); 2782 return ret; 2783 } 2784 2785 /* 2786 * Extract folio fragments from a FOLIOQ-class iterator and add them to an RDMA 2787 * list. The folios are not pinned. 2788 */ 2789 static ssize_t smb_extract_folioq_to_rdma(struct iov_iter *iter, 2790 struct smb_extract_to_rdma *rdma, 2791 ssize_t maxsize) 2792 { 2793 const struct folio_queue *folioq = iter->folioq; 2794 unsigned int slot = iter->folioq_slot; 2795 ssize_t ret = 0; 2796 size_t offset = iter->iov_offset; 2797 2798 BUG_ON(!folioq); 2799 2800 if (slot >= folioq_nr_slots(folioq)) { 2801 folioq = folioq->next; 2802 if (WARN_ON_ONCE(!folioq)) 2803 return -EIO; 2804 slot = 0; 2805 } 2806 2807 do { 2808 struct folio *folio = folioq_folio(folioq, slot); 2809 size_t fsize = folioq_folio_size(folioq, slot); 2810 2811 if (offset < fsize) { 2812 size_t part = umin(maxsize, fsize - offset); 2813 2814 if (!smb_set_sge(rdma, folio_page(folio, 0), offset, part)) 2815 return -EIO; 2816 2817 offset += part; 2818 ret += part; 2819 maxsize -= part; 2820 } 2821 2822 if (offset >= fsize) { 2823 offset = 0; 2824 slot++; 2825 if (slot >= folioq_nr_slots(folioq)) { 2826 if (!folioq->next) { 2827 WARN_ON_ONCE(ret < iter->count); 2828 break; 2829 } 2830 folioq = folioq->next; 2831 slot = 0; 2832 } 2833 } 2834 } while (rdma->nr_sge < rdma->max_sge && maxsize > 0); 2835 2836 iter->folioq = folioq; 2837 iter->folioq_slot = slot; 2838 iter->iov_offset = offset; 2839 iter->count -= ret; 2840 return ret; 2841 } 2842 2843 /* 2844 * Extract page fragments from up to the given amount of the source iterator 2845 * and build up an RDMA list that refers to all of those bits. The RDMA list 2846 * is appended to, up to the maximum number of elements set in the parameter 2847 * block. 2848 * 2849 * The extracted page fragments are not pinned or ref'd in any way; if an 2850 * IOVEC/UBUF-type iterator is to be used, it should be converted to a 2851 * BVEC-type iterator and the pages pinned, ref'd or otherwise held in some 2852 * way. 2853 */ 2854 static ssize_t smb_extract_iter_to_rdma(struct iov_iter *iter, size_t len, 2855 struct smb_extract_to_rdma *rdma) 2856 { 2857 ssize_t ret; 2858 int before = rdma->nr_sge; 2859 2860 switch (iov_iter_type(iter)) { 2861 case ITER_BVEC: 2862 ret = smb_extract_bvec_to_rdma(iter, rdma, len); 2863 break; 2864 case ITER_KVEC: 2865 ret = smb_extract_kvec_to_rdma(iter, rdma, len); 2866 break; 2867 case ITER_FOLIOQ: 2868 ret = smb_extract_folioq_to_rdma(iter, rdma, len); 2869 break; 2870 default: 2871 WARN_ON_ONCE(1); 2872 return -EIO; 2873 } 2874 2875 if (ret < 0) { 2876 while (rdma->nr_sge > before) { 2877 struct ib_sge *sge = &rdma->sge[rdma->nr_sge--]; 2878 2879 ib_dma_unmap_single(rdma->device, sge->addr, sge->length, 2880 rdma->direction); 2881 sge->addr = 0; 2882 } 2883 } 2884 2885 return ret; 2886 } 2887