1 /* 2 * Copyright (c) 2006 - 2009 Mellanox Technology Inc. All rights reserved. 3 * Copyright (C) 2008 - 2011 Bart Van Assche <bvanassche@acm.org>. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 * 33 */ 34 35 #include <linux/module.h> 36 #include <linux/hex.h> 37 #include <linux/init.h> 38 #include <linux/slab.h> 39 #include <linux/err.h> 40 #include <linux/ctype.h> 41 #include <linux/kthread.h> 42 #include <linux/string.h> 43 #include <linux/delay.h> 44 #include <linux/atomic.h> 45 #include <linux/inet.h> 46 #include <rdma/ib_cache.h> 47 #include <scsi/scsi_proto.h> 48 #include <scsi/scsi_tcq.h> 49 #include <target/target_core_base.h> 50 #include <target/target_core_fabric.h> 51 #include "ib_srpt.h" 52 53 /* Name of this kernel module. */ 54 #define DRV_NAME "ib_srpt" 55 56 #define SRPT_ID_STRING "Linux SRP target" 57 58 #undef pr_fmt 59 #define pr_fmt(fmt) DRV_NAME " " fmt 60 61 MODULE_AUTHOR("Vu Pham and Bart Van Assche"); 62 MODULE_DESCRIPTION("SCSI RDMA Protocol target driver"); 63 MODULE_LICENSE("Dual BSD/GPL"); 64 65 /* 66 * Global Variables 67 */ 68 69 static u64 srpt_service_guid; 70 static DEFINE_SPINLOCK(srpt_dev_lock); /* Protects srpt_dev_list. */ 71 static LIST_HEAD(srpt_dev_list); /* List of srpt_device structures. */ 72 static DEFINE_MUTEX(srpt_mc_mutex); /* Protects srpt_memory_caches. */ 73 static DEFINE_XARRAY(srpt_memory_caches); /* See also srpt_memory_cache_entry */ 74 75 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE; 76 module_param(srp_max_req_size, int, 0444); 77 MODULE_PARM_DESC(srp_max_req_size, 78 "Maximum size of SRP request messages in bytes."); 79 80 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE; 81 module_param(srpt_srq_size, int, 0444); 82 MODULE_PARM_DESC(srpt_srq_size, 83 "Shared receive queue (SRQ) size."); 84 85 static int srpt_set_u64_x(const char *buffer, const struct kernel_param *kp) 86 { 87 return kstrtou64(buffer, 16, (u64 *)kp->arg); 88 } 89 static int srpt_get_u64_x(char *buffer, const struct kernel_param *kp) 90 { 91 return sprintf(buffer, "0x%016llx\n", *(u64 *)kp->arg); 92 } 93 module_param_call(srpt_service_guid, srpt_set_u64_x, srpt_get_u64_x, 94 &srpt_service_guid, 0444); 95 MODULE_PARM_DESC(srpt_service_guid, 96 "Using this value for ioc_guid, id_ext, and cm_listen_id instead of using the node_guid of the first HCA."); 97 98 static struct ib_client srpt_client; 99 /* Protects both rdma_cm_port and rdma_cm_id. */ 100 static DEFINE_MUTEX(rdma_cm_mutex); 101 /* Port number RDMA/CM will bind to. */ 102 static u16 rdma_cm_port; 103 static struct rdma_cm_id *rdma_cm_id; 104 static void srpt_release_cmd(struct se_cmd *se_cmd); 105 static void srpt_free_ch(struct kref *kref); 106 static int srpt_queue_status(struct se_cmd *cmd); 107 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc); 108 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc); 109 static void srpt_process_wait_list(struct srpt_rdma_ch *ch); 110 111 /* Type of the entries in srpt_memory_caches. */ 112 struct srpt_memory_cache_entry { 113 refcount_t ref; 114 struct kmem_cache *c; 115 }; 116 117 static struct kmem_cache *srpt_cache_get(unsigned int object_size) 118 { 119 struct srpt_memory_cache_entry *e; 120 char name[32]; 121 void *res; 122 123 guard(mutex)(&srpt_mc_mutex); 124 e = xa_load(&srpt_memory_caches, object_size); 125 if (e) { 126 refcount_inc(&e->ref); 127 return e->c; 128 } 129 snprintf(name, sizeof(name), "srpt-%u", object_size); 130 e = kmalloc_obj(*e); 131 if (!e) 132 return NULL; 133 refcount_set(&e->ref, 1); 134 e->c = kmem_cache_create(name, object_size, /*align=*/512, 0, NULL); 135 if (!e->c) 136 goto free_entry; 137 res = xa_store(&srpt_memory_caches, object_size, e, GFP_KERNEL); 138 if (xa_is_err(res)) 139 goto destroy_cache; 140 return e->c; 141 142 destroy_cache: 143 kmem_cache_destroy(e->c); 144 145 free_entry: 146 kfree(e); 147 return NULL; 148 } 149 150 static void srpt_cache_put(struct kmem_cache *c) 151 { 152 struct srpt_memory_cache_entry *e = NULL; 153 unsigned long object_size; 154 155 guard(mutex)(&srpt_mc_mutex); 156 xa_for_each(&srpt_memory_caches, object_size, e) 157 if (e->c == c) 158 break; 159 if (WARN_ON_ONCE(!e)) 160 return; 161 if (!refcount_dec_and_test(&e->ref)) 162 return; 163 WARN_ON_ONCE(xa_erase(&srpt_memory_caches, object_size) != e); 164 kmem_cache_destroy(e->c); 165 kfree(e); 166 } 167 168 /* 169 * The only allowed channel state changes are those that change the channel 170 * state into a state with a higher numerical value. Hence the new > prev test. 171 */ 172 static bool srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new) 173 { 174 unsigned long flags; 175 enum rdma_ch_state prev; 176 bool changed = false; 177 178 spin_lock_irqsave(&ch->spinlock, flags); 179 prev = ch->state; 180 if (new > prev) { 181 ch->state = new; 182 changed = true; 183 } 184 spin_unlock_irqrestore(&ch->spinlock, flags); 185 186 return changed; 187 } 188 189 /** 190 * srpt_event_handler - asynchronous IB event callback function 191 * @handler: IB event handler registered by ib_register_event_handler(). 192 * @event: Description of the event that occurred. 193 * 194 * Callback function called by the InfiniBand core when an asynchronous IB 195 * event occurs. This callback may occur in interrupt context. See also 196 * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand 197 * Architecture Specification. 198 */ 199 static void srpt_event_handler(struct ib_event_handler *handler, 200 struct ib_event *event) 201 { 202 struct srpt_device *sdev = 203 container_of(handler, struct srpt_device, event_handler); 204 struct srpt_port *sport; 205 u8 port_num; 206 207 pr_debug("ASYNC event= %d on device= %s\n", event->event, 208 dev_name(&sdev->device->dev)); 209 210 switch (event->event) { 211 case IB_EVENT_PORT_ERR: 212 port_num = event->element.port_num - 1; 213 if (port_num < sdev->device->phys_port_cnt) { 214 sport = &sdev->port[port_num]; 215 sport->lid = 0; 216 sport->sm_lid = 0; 217 } else { 218 WARN(true, "event %d: port_num %d out of range 1..%d\n", 219 event->event, port_num + 1, 220 sdev->device->phys_port_cnt); 221 } 222 break; 223 case IB_EVENT_PORT_ACTIVE: 224 case IB_EVENT_LID_CHANGE: 225 case IB_EVENT_PKEY_CHANGE: 226 case IB_EVENT_SM_CHANGE: 227 case IB_EVENT_CLIENT_REREGISTER: 228 case IB_EVENT_GID_CHANGE: 229 /* Refresh port data asynchronously. */ 230 port_num = event->element.port_num - 1; 231 if (port_num < sdev->device->phys_port_cnt) { 232 sport = &sdev->port[port_num]; 233 if (!sport->lid && !sport->sm_lid) 234 schedule_work(&sport->work); 235 } else { 236 WARN(true, "event %d: port_num %d out of range 1..%d\n", 237 event->event, port_num + 1, 238 sdev->device->phys_port_cnt); 239 } 240 break; 241 default: 242 pr_err("received unrecognized IB event %d\n", event->event); 243 break; 244 } 245 } 246 247 /** 248 * srpt_srq_event - SRQ event callback function 249 * @event: Description of the event that occurred. 250 * @ctx: Context pointer specified at SRQ creation time. 251 */ 252 static void srpt_srq_event(struct ib_event *event, void *ctx) 253 { 254 pr_debug("SRQ event %d\n", event->event); 255 } 256 257 static const char *get_ch_state_name(enum rdma_ch_state s) 258 { 259 switch (s) { 260 case CH_CONNECTING: 261 return "connecting"; 262 case CH_LIVE: 263 return "live"; 264 case CH_DISCONNECTING: 265 return "disconnecting"; 266 case CH_DRAINING: 267 return "draining"; 268 case CH_DISCONNECTED: 269 return "disconnected"; 270 } 271 return "???"; 272 } 273 274 /** 275 * srpt_qp_event - QP event callback function 276 * @event: Description of the event that occurred. 277 * @ptr: SRPT RDMA channel. 278 */ 279 static void srpt_qp_event(struct ib_event *event, void *ptr) 280 { 281 struct srpt_rdma_ch *ch = ptr; 282 283 pr_debug("QP event %d on ch=%p sess_name=%s-%d state=%s\n", 284 event->event, ch, ch->sess_name, ch->qp->qp_num, 285 get_ch_state_name(ch->state)); 286 287 switch (event->event) { 288 case IB_EVENT_COMM_EST: 289 if (ch->using_rdma_cm) 290 rdma_notify(ch->rdma_cm.cm_id, event->event); 291 else 292 ib_cm_notify(ch->ib_cm.cm_id, event->event); 293 break; 294 case IB_EVENT_QP_LAST_WQE_REACHED: 295 pr_debug("%s-%d, state %s: received Last WQE event.\n", 296 ch->sess_name, ch->qp->qp_num, 297 get_ch_state_name(ch->state)); 298 break; 299 default: 300 pr_err("received unrecognized IB QP event %d\n", event->event); 301 break; 302 } 303 } 304 305 /** 306 * srpt_set_ioc - initialize a IOUnitInfo structure 307 * @c_list: controller list. 308 * @slot: one-based slot number. 309 * @value: four-bit value. 310 * 311 * Copies the lowest four bits of value in element slot of the array of four 312 * bit elements called c_list (controller list). The index slot is one-based. 313 */ 314 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value) 315 { 316 u16 id; 317 u8 tmp; 318 319 id = (slot - 1) / 2; 320 if (slot & 0x1) { 321 tmp = c_list[id] & 0xf; 322 c_list[id] = (value << 4) | tmp; 323 } else { 324 tmp = c_list[id] & 0xf0; 325 c_list[id] = (value & 0xf) | tmp; 326 } 327 } 328 329 /** 330 * srpt_get_class_port_info - copy ClassPortInfo to a management datagram 331 * @mad: Datagram that will be sent as response to DM_ATTR_CLASS_PORT_INFO. 332 * 333 * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture 334 * Specification. 335 */ 336 static void srpt_get_class_port_info(struct ib_dm_mad *mad) 337 { 338 struct ib_class_port_info *cif; 339 340 cif = (struct ib_class_port_info *)mad->data; 341 memset(cif, 0, sizeof(*cif)); 342 cif->base_version = 1; 343 cif->class_version = 1; 344 345 ib_set_cpi_resp_time(cif, 20); 346 mad->mad_hdr.status = 0; 347 } 348 349 /** 350 * srpt_get_iou - write IOUnitInfo to a management datagram 351 * @mad: Datagram that will be sent as response to DM_ATTR_IOU_INFO. 352 * 353 * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture 354 * Specification. See also section B.7, table B.6 in the SRP r16a document. 355 */ 356 static void srpt_get_iou(struct ib_dm_mad *mad) 357 { 358 struct ib_dm_iou_info *ioui; 359 u8 slot; 360 int i; 361 362 ioui = (struct ib_dm_iou_info *)mad->data; 363 ioui->change_id = cpu_to_be16(1); 364 ioui->max_controllers = 16; 365 366 /* set present for slot 1 and empty for the rest */ 367 srpt_set_ioc(ioui->controller_list, 1, 1); 368 for (i = 1, slot = 2; i < 16; i++, slot++) 369 srpt_set_ioc(ioui->controller_list, slot, 0); 370 371 mad->mad_hdr.status = 0; 372 } 373 374 /** 375 * srpt_get_ioc - write IOControllerprofile to a management datagram 376 * @sport: HCA port through which the MAD has been received. 377 * @slot: Slot number specified in DM_ATTR_IOC_PROFILE query. 378 * @mad: Datagram that will be sent as response to DM_ATTR_IOC_PROFILE. 379 * 380 * See also section 16.3.3.4 IOControllerProfile in the InfiniBand 381 * Architecture Specification. See also section B.7, table B.7 in the SRP 382 * r16a document. 383 */ 384 static void srpt_get_ioc(struct srpt_port *sport, u32 slot, 385 struct ib_dm_mad *mad) 386 { 387 struct srpt_device *sdev = sport->sdev; 388 struct ib_dm_ioc_profile *iocp; 389 int send_queue_depth; 390 391 iocp = (struct ib_dm_ioc_profile *)mad->data; 392 393 if (!slot || slot > 16) { 394 mad->mad_hdr.status 395 = cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD); 396 return; 397 } 398 399 if (slot > 2) { 400 mad->mad_hdr.status 401 = cpu_to_be16(DM_MAD_STATUS_NO_IOC); 402 return; 403 } 404 405 if (sdev->use_srq) 406 send_queue_depth = sdev->srq_size; 407 else 408 send_queue_depth = min(MAX_SRPT_RQ_SIZE, 409 sdev->device->attrs.max_qp_wr); 410 411 memset(iocp, 0, sizeof(*iocp)); 412 strcpy(iocp->id_string, SRPT_ID_STRING); 413 iocp->guid = cpu_to_be64(srpt_service_guid); 414 iocp->vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id); 415 iocp->device_id = cpu_to_be32(sdev->device->attrs.vendor_part_id); 416 iocp->device_version = cpu_to_be16(sdev->device->attrs.hw_ver); 417 iocp->subsys_vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id); 418 iocp->subsys_device_id = 0x0; 419 iocp->io_class = cpu_to_be16(SRP_REV16A_IB_IO_CLASS); 420 iocp->io_subclass = cpu_to_be16(SRP_IO_SUBCLASS); 421 iocp->protocol = cpu_to_be16(SRP_PROTOCOL); 422 iocp->protocol_version = cpu_to_be16(SRP_PROTOCOL_VERSION); 423 iocp->send_queue_depth = cpu_to_be16(send_queue_depth); 424 iocp->rdma_read_depth = 4; 425 iocp->send_size = cpu_to_be32(srp_max_req_size); 426 iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size, 427 1U << 24)); 428 iocp->num_svc_entries = 1; 429 iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC | 430 SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC; 431 432 mad->mad_hdr.status = 0; 433 } 434 435 /** 436 * srpt_get_svc_entries - write ServiceEntries to a management datagram 437 * @ioc_guid: I/O controller GUID to use in reply. 438 * @slot: I/O controller number. 439 * @hi: End of the range of service entries to be specified in the reply. 440 * @lo: Start of the range of service entries to be specified in the reply.. 441 * @mad: Datagram that will be sent as response to DM_ATTR_SVC_ENTRIES. 442 * 443 * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture 444 * Specification. See also section B.7, table B.8 in the SRP r16a document. 445 */ 446 static void srpt_get_svc_entries(u64 ioc_guid, 447 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad) 448 { 449 struct ib_dm_svc_entries *svc_entries; 450 451 WARN_ON(!ioc_guid); 452 453 if (!slot || slot > 16) { 454 mad->mad_hdr.status 455 = cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD); 456 return; 457 } 458 459 if (slot > 2 || lo > hi || hi > 1) { 460 mad->mad_hdr.status 461 = cpu_to_be16(DM_MAD_STATUS_NO_IOC); 462 return; 463 } 464 465 svc_entries = (struct ib_dm_svc_entries *)mad->data; 466 memset(svc_entries, 0, sizeof(*svc_entries)); 467 svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid); 468 snprintf(svc_entries->service_entries[0].name, 469 sizeof(svc_entries->service_entries[0].name), 470 "%s%016llx", 471 SRP_SERVICE_NAME_PREFIX, 472 ioc_guid); 473 474 mad->mad_hdr.status = 0; 475 } 476 477 /** 478 * srpt_mgmt_method_get - process a received management datagram 479 * @sp: HCA port through which the MAD has been received. 480 * @rq_mad: received MAD. 481 * @rsp_mad: response MAD. 482 */ 483 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad, 484 struct ib_dm_mad *rsp_mad) 485 { 486 u16 attr_id; 487 u32 slot; 488 u8 hi, lo; 489 490 attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id); 491 switch (attr_id) { 492 case DM_ATTR_CLASS_PORT_INFO: 493 srpt_get_class_port_info(rsp_mad); 494 break; 495 case DM_ATTR_IOU_INFO: 496 srpt_get_iou(rsp_mad); 497 break; 498 case DM_ATTR_IOC_PROFILE: 499 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod); 500 srpt_get_ioc(sp, slot, rsp_mad); 501 break; 502 case DM_ATTR_SVC_ENTRIES: 503 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod); 504 hi = (u8) ((slot >> 8) & 0xff); 505 lo = (u8) (slot & 0xff); 506 slot = (u16) ((slot >> 16) & 0xffff); 507 srpt_get_svc_entries(srpt_service_guid, 508 slot, hi, lo, rsp_mad); 509 break; 510 default: 511 rsp_mad->mad_hdr.status = 512 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR); 513 break; 514 } 515 } 516 517 /** 518 * srpt_mad_send_handler - MAD send completion callback 519 * @mad_agent: Return value of ib_register_mad_agent(). 520 * @mad_wc: Work completion reporting that the MAD has been sent. 521 */ 522 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent, 523 struct ib_mad_send_wc *mad_wc) 524 { 525 rdma_destroy_ah(mad_wc->send_buf->ah, RDMA_DESTROY_AH_SLEEPABLE); 526 ib_free_send_mad(mad_wc->send_buf); 527 } 528 529 /** 530 * srpt_mad_recv_handler - MAD reception callback function 531 * @mad_agent: Return value of ib_register_mad_agent(). 532 * @send_buf: Not used. 533 * @mad_wc: Work completion reporting that a MAD has been received. 534 */ 535 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent, 536 struct ib_mad_send_buf *send_buf, 537 struct ib_mad_recv_wc *mad_wc) 538 { 539 struct srpt_port *sport = (struct srpt_port *)mad_agent->context; 540 struct ib_ah *ah; 541 struct ib_mad_send_buf *rsp; 542 struct ib_dm_mad *dm_mad; 543 544 if (!mad_wc || !mad_wc->recv_buf.mad) 545 return; 546 547 ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc, 548 mad_wc->recv_buf.grh, mad_agent->port_num); 549 if (IS_ERR(ah)) 550 goto err; 551 552 BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR); 553 554 rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp, 555 mad_wc->wc->pkey_index, 0, 556 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA, 557 GFP_KERNEL, 558 IB_MGMT_BASE_VERSION); 559 if (IS_ERR(rsp)) 560 goto err_rsp; 561 562 rsp->ah = ah; 563 564 dm_mad = rsp->mad; 565 memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof(*dm_mad)); 566 dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP; 567 dm_mad->mad_hdr.status = 0; 568 569 switch (mad_wc->recv_buf.mad->mad_hdr.method) { 570 case IB_MGMT_METHOD_GET: 571 srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad); 572 break; 573 case IB_MGMT_METHOD_SET: 574 dm_mad->mad_hdr.status = 575 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR); 576 break; 577 default: 578 dm_mad->mad_hdr.status = 579 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD); 580 break; 581 } 582 583 if (!ib_post_send_mad(rsp, NULL)) { 584 ib_free_recv_mad(mad_wc); 585 /* will destroy_ah & free_send_mad in send completion */ 586 return; 587 } 588 589 ib_free_send_mad(rsp); 590 591 err_rsp: 592 rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE); 593 err: 594 ib_free_recv_mad(mad_wc); 595 } 596 597 static int srpt_format_guid(char *buf, unsigned int size, const __be64 *guid) 598 { 599 const __be16 *g = (const __be16 *)guid; 600 601 return snprintf(buf, size, "%04x:%04x:%04x:%04x", 602 be16_to_cpu(g[0]), be16_to_cpu(g[1]), 603 be16_to_cpu(g[2]), be16_to_cpu(g[3])); 604 } 605 606 /** 607 * srpt_refresh_port - configure a HCA port 608 * @sport: SRPT HCA port. 609 * 610 * Enable InfiniBand management datagram processing, update the cached sm_lid, 611 * lid and gid values, and register a callback function for processing MADs 612 * on the specified port. 613 * 614 * Note: It is safe to call this function more than once for the same port. 615 */ 616 static int srpt_refresh_port(struct srpt_port *sport) 617 { 618 struct ib_mad_agent *mad_agent; 619 struct ib_mad_reg_req reg_req; 620 struct ib_port_modify port_modify; 621 struct ib_port_attr port_attr; 622 int ret; 623 624 ret = ib_query_port(sport->sdev->device, sport->port, &port_attr); 625 if (ret) 626 return ret; 627 628 sport->sm_lid = port_attr.sm_lid; 629 sport->lid = port_attr.lid; 630 631 ret = rdma_query_gid(sport->sdev->device, sport->port, 0, &sport->gid); 632 if (ret) 633 return ret; 634 635 srpt_format_guid(sport->guid_name, ARRAY_SIZE(sport->guid_name), 636 &sport->gid.global.interface_id); 637 snprintf(sport->gid_name, ARRAY_SIZE(sport->gid_name), 638 "0x%016llx%016llx", 639 be64_to_cpu(sport->gid.global.subnet_prefix), 640 be64_to_cpu(sport->gid.global.interface_id)); 641 642 if (rdma_protocol_iwarp(sport->sdev->device, sport->port)) 643 return 0; 644 645 memset(&port_modify, 0, sizeof(port_modify)); 646 port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP; 647 port_modify.clr_port_cap_mask = 0; 648 649 ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify); 650 if (ret) { 651 pr_warn("%s-%d: enabling device management failed (%d). Note: this is expected if SR-IOV is enabled.\n", 652 dev_name(&sport->sdev->device->dev), sport->port, ret); 653 return 0; 654 } 655 656 if (!sport->mad_agent) { 657 memset(®_req, 0, sizeof(reg_req)); 658 reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT; 659 reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION; 660 set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask); 661 set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask); 662 663 mad_agent = ib_register_mad_agent(sport->sdev->device, 664 sport->port, 665 IB_QPT_GSI, 666 ®_req, 0, 667 srpt_mad_send_handler, 668 srpt_mad_recv_handler, 669 sport, 0); 670 if (IS_ERR(mad_agent)) { 671 pr_err("%s-%d: MAD agent registration failed (%pe). Note: this is expected if SR-IOV is enabled.\n", 672 dev_name(&sport->sdev->device->dev), sport->port, 673 mad_agent); 674 sport->mad_agent = NULL; 675 memset(&port_modify, 0, sizeof(port_modify)); 676 port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP; 677 ib_modify_port(sport->sdev->device, sport->port, 0, 678 &port_modify); 679 return 0; 680 } 681 682 sport->mad_agent = mad_agent; 683 } 684 685 return 0; 686 } 687 688 /** 689 * srpt_unregister_mad_agent - unregister MAD callback functions 690 * @sdev: SRPT HCA pointer. 691 * @port_cnt: number of ports with registered MAD 692 * 693 * Note: It is safe to call this function more than once for the same device. 694 */ 695 static void srpt_unregister_mad_agent(struct srpt_device *sdev, int port_cnt) 696 { 697 struct ib_port_modify port_modify = { 698 .clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP, 699 }; 700 struct srpt_port *sport; 701 int i; 702 703 for (i = 1; i <= port_cnt; i++) { 704 sport = &sdev->port[i - 1]; 705 WARN_ON(sport->port != i); 706 if (sport->mad_agent) { 707 ib_modify_port(sdev->device, i, 0, &port_modify); 708 ib_unregister_mad_agent(sport->mad_agent); 709 sport->mad_agent = NULL; 710 } 711 } 712 } 713 714 /** 715 * srpt_alloc_ioctx - allocate a SRPT I/O context structure 716 * @sdev: SRPT HCA pointer. 717 * @ioctx_size: I/O context size. 718 * @buf_cache: I/O buffer cache. 719 * @dir: DMA data direction. 720 */ 721 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev, 722 int ioctx_size, 723 struct kmem_cache *buf_cache, 724 enum dma_data_direction dir) 725 { 726 struct srpt_ioctx *ioctx; 727 728 ioctx = kzalloc(ioctx_size, GFP_KERNEL); 729 if (!ioctx) 730 goto err; 731 732 ioctx->buf = kmem_cache_alloc(buf_cache, GFP_KERNEL); 733 if (!ioctx->buf) 734 goto err_free_ioctx; 735 736 ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, 737 kmem_cache_size(buf_cache), dir); 738 if (ib_dma_mapping_error(sdev->device, ioctx->dma)) 739 goto err_free_buf; 740 741 return ioctx; 742 743 err_free_buf: 744 kmem_cache_free(buf_cache, ioctx->buf); 745 err_free_ioctx: 746 kfree(ioctx); 747 err: 748 return NULL; 749 } 750 751 /** 752 * srpt_free_ioctx - free a SRPT I/O context structure 753 * @sdev: SRPT HCA pointer. 754 * @ioctx: I/O context pointer. 755 * @buf_cache: I/O buffer cache. 756 * @dir: DMA data direction. 757 */ 758 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx, 759 struct kmem_cache *buf_cache, 760 enum dma_data_direction dir) 761 { 762 if (!ioctx) 763 return; 764 765 ib_dma_unmap_single(sdev->device, ioctx->dma, 766 kmem_cache_size(buf_cache), dir); 767 kmem_cache_free(buf_cache, ioctx->buf); 768 kfree(ioctx); 769 } 770 771 /** 772 * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures 773 * @sdev: Device to allocate the I/O context ring for. 774 * @ring_size: Number of elements in the I/O context ring. 775 * @ioctx_size: I/O context size. 776 * @buf_cache: I/O buffer cache. 777 * @alignment_offset: Offset in each ring buffer at which the SRP information 778 * unit starts. 779 * @dir: DMA data direction. 780 */ 781 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev, 782 int ring_size, int ioctx_size, 783 struct kmem_cache *buf_cache, 784 int alignment_offset, 785 enum dma_data_direction dir) 786 { 787 struct srpt_ioctx **ring; 788 int i; 789 790 WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) && 791 ioctx_size != sizeof(struct srpt_send_ioctx)); 792 793 ring = kvmalloc_objs(ring[0], ring_size); 794 if (!ring) 795 goto out; 796 for (i = 0; i < ring_size; ++i) { 797 ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, buf_cache, dir); 798 if (!ring[i]) 799 goto err; 800 ring[i]->index = i; 801 ring[i]->offset = alignment_offset; 802 } 803 goto out; 804 805 err: 806 while (--i >= 0) 807 srpt_free_ioctx(sdev, ring[i], buf_cache, dir); 808 kvfree(ring); 809 ring = NULL; 810 out: 811 return ring; 812 } 813 814 /** 815 * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures 816 * @ioctx_ring: I/O context ring to be freed. 817 * @sdev: SRPT HCA pointer. 818 * @ring_size: Number of ring elements. 819 * @buf_cache: I/O buffer cache. 820 * @dir: DMA data direction. 821 */ 822 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring, 823 struct srpt_device *sdev, int ring_size, 824 struct kmem_cache *buf_cache, 825 enum dma_data_direction dir) 826 { 827 int i; 828 829 if (!ioctx_ring) 830 return; 831 832 for (i = 0; i < ring_size; ++i) 833 srpt_free_ioctx(sdev, ioctx_ring[i], buf_cache, dir); 834 kvfree(ioctx_ring); 835 } 836 837 /** 838 * srpt_set_cmd_state - set the state of a SCSI command 839 * @ioctx: Send I/O context. 840 * @new: New I/O context state. 841 * 842 * Does not modify the state of aborted commands. Returns the previous command 843 * state. 844 */ 845 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx, 846 enum srpt_command_state new) 847 { 848 enum srpt_command_state previous; 849 850 previous = ioctx->state; 851 if (previous != SRPT_STATE_DONE) 852 ioctx->state = new; 853 854 return previous; 855 } 856 857 /** 858 * srpt_test_and_set_cmd_state - test and set the state of a command 859 * @ioctx: Send I/O context. 860 * @old: Current I/O context state. 861 * @new: New I/O context state. 862 * 863 * Returns true if and only if the previous command state was equal to 'old'. 864 */ 865 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx, 866 enum srpt_command_state old, 867 enum srpt_command_state new) 868 { 869 enum srpt_command_state previous; 870 871 WARN_ON(!ioctx); 872 WARN_ON(old == SRPT_STATE_DONE); 873 WARN_ON(new == SRPT_STATE_NEW); 874 875 previous = ioctx->state; 876 if (previous == old) 877 ioctx->state = new; 878 879 return previous == old; 880 } 881 882 /** 883 * srpt_post_recv - post an IB receive request 884 * @sdev: SRPT HCA pointer. 885 * @ch: SRPT RDMA channel. 886 * @ioctx: Receive I/O context pointer. 887 */ 888 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch, 889 struct srpt_recv_ioctx *ioctx) 890 { 891 struct ib_sge list; 892 struct ib_recv_wr wr; 893 894 BUG_ON(!sdev); 895 list.addr = ioctx->ioctx.dma + ioctx->ioctx.offset; 896 list.length = srp_max_req_size; 897 list.lkey = sdev->lkey; 898 899 ioctx->ioctx.cqe.done = srpt_recv_done; 900 wr.wr_cqe = &ioctx->ioctx.cqe; 901 wr.next = NULL; 902 wr.sg_list = &list; 903 wr.num_sge = 1; 904 905 if (sdev->use_srq) 906 return ib_post_srq_recv(sdev->srq, &wr, NULL); 907 else 908 return ib_post_recv(ch->qp, &wr, NULL); 909 } 910 911 /** 912 * srpt_zerolength_write - perform a zero-length RDMA write 913 * @ch: SRPT RDMA channel. 914 * 915 * A quote from the InfiniBand specification: C9-88: For an HCA responder 916 * using Reliable Connection service, for each zero-length RDMA READ or WRITE 917 * request, the R_Key shall not be validated, even if the request includes 918 * Immediate data. 919 */ 920 static int srpt_zerolength_write(struct srpt_rdma_ch *ch) 921 { 922 struct ib_rdma_wr wr = { 923 .wr = { 924 .next = NULL, 925 { .wr_cqe = &ch->zw_cqe, }, 926 .opcode = IB_WR_RDMA_WRITE, 927 .send_flags = IB_SEND_SIGNALED, 928 } 929 }; 930 931 pr_debug("%s-%d: queued zerolength write\n", ch->sess_name, 932 ch->qp->qp_num); 933 934 return ib_post_send(ch->qp, &wr.wr, NULL); 935 } 936 937 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc) 938 { 939 struct srpt_rdma_ch *ch = wc->qp->qp_context; 940 941 pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num, 942 wc->status); 943 944 if (wc->status == IB_WC_SUCCESS) { 945 srpt_process_wait_list(ch); 946 } else { 947 if (srpt_set_ch_state(ch, CH_DISCONNECTED)) 948 schedule_work(&ch->release_work); 949 else 950 pr_debug("%s-%d: already disconnected.\n", 951 ch->sess_name, ch->qp->qp_num); 952 } 953 } 954 955 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx, 956 struct srp_direct_buf *db, int nbufs, struct scatterlist **sg, 957 unsigned *sg_cnt) 958 { 959 enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd); 960 struct srpt_rdma_ch *ch = ioctx->ch; 961 struct scatterlist *prev = NULL; 962 unsigned prev_nents; 963 int ret, i; 964 965 if (nbufs == 1) { 966 ioctx->rw_ctxs = &ioctx->s_rw_ctx; 967 } else { 968 ioctx->rw_ctxs = kmalloc_objs(*ioctx->rw_ctxs, nbufs); 969 if (!ioctx->rw_ctxs) 970 return -ENOMEM; 971 } 972 973 for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) { 974 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 975 u64 remote_addr = be64_to_cpu(db->va); 976 u32 size = be32_to_cpu(db->len); 977 u32 rkey = be32_to_cpu(db->key); 978 979 ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false, 980 i < nbufs - 1); 981 if (ret) 982 goto unwind; 983 984 ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port, 985 ctx->sg, ctx->nents, 0, remote_addr, rkey, dir); 986 if (ret < 0) { 987 target_free_sgl(ctx->sg, ctx->nents); 988 goto unwind; 989 } 990 991 ioctx->n_rdma += ret; 992 ioctx->n_rw_ctx++; 993 994 if (prev) { 995 sg_unmark_end(&prev[prev_nents - 1]); 996 sg_chain(prev, prev_nents + 1, ctx->sg); 997 } else { 998 *sg = ctx->sg; 999 } 1000 1001 prev = ctx->sg; 1002 prev_nents = ctx->nents; 1003 1004 *sg_cnt += ctx->nents; 1005 } 1006 1007 return 0; 1008 1009 unwind: 1010 while (--i >= 0) { 1011 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 1012 1013 rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port, 1014 ctx->sg, ctx->nents, dir); 1015 target_free_sgl(ctx->sg, ctx->nents); 1016 } 1017 if (ioctx->rw_ctxs != &ioctx->s_rw_ctx) 1018 kfree(ioctx->rw_ctxs); 1019 return ret; 1020 } 1021 1022 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch, 1023 struct srpt_send_ioctx *ioctx) 1024 { 1025 enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd); 1026 int i; 1027 1028 for (i = 0; i < ioctx->n_rw_ctx; i++) { 1029 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 1030 1031 rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port, 1032 ctx->sg, ctx->nents, dir); 1033 target_free_sgl(ctx->sg, ctx->nents); 1034 } 1035 1036 if (ioctx->rw_ctxs != &ioctx->s_rw_ctx) 1037 kfree(ioctx->rw_ctxs); 1038 } 1039 1040 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd) 1041 { 1042 /* 1043 * The pointer computations below will only be compiled correctly 1044 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check 1045 * whether srp_cmd::add_data has been declared as a byte pointer. 1046 */ 1047 BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) && 1048 !__same_type(srp_cmd->add_data[0], (u8)0)); 1049 1050 /* 1051 * According to the SRP spec, the lower two bits of the 'ADDITIONAL 1052 * CDB LENGTH' field are reserved and the size in bytes of this field 1053 * is four times the value specified in bits 3..7. Hence the "& ~3". 1054 */ 1055 return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3); 1056 } 1057 1058 /** 1059 * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request 1060 * @recv_ioctx: I/O context associated with the received command @srp_cmd. 1061 * @ioctx: I/O context that will be used for responding to the initiator. 1062 * @srp_cmd: Pointer to the SRP_CMD request data. 1063 * @dir: Pointer to the variable to which the transfer direction will be 1064 * written. 1065 * @sg: [out] scatterlist for the parsed SRP_CMD. 1066 * @sg_cnt: [out] length of @sg. 1067 * @data_len: Pointer to the variable to which the total data length of all 1068 * descriptors in the SRP_CMD request will be written. 1069 * @imm_data_offset: [in] Offset in SRP_CMD requests at which immediate data 1070 * starts. 1071 * 1072 * This function initializes ioctx->nrbuf and ioctx->r_bufs. 1073 * 1074 * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors; 1075 * -ENOMEM when memory allocation fails and zero upon success. 1076 */ 1077 static int srpt_get_desc_tbl(struct srpt_recv_ioctx *recv_ioctx, 1078 struct srpt_send_ioctx *ioctx, 1079 struct srp_cmd *srp_cmd, enum dma_data_direction *dir, 1080 struct scatterlist **sg, unsigned int *sg_cnt, u64 *data_len, 1081 u16 imm_data_offset) 1082 { 1083 BUG_ON(!dir); 1084 BUG_ON(!data_len); 1085 1086 /* 1087 * The lower four bits of the buffer format field contain the DATA-IN 1088 * buffer descriptor format, and the highest four bits contain the 1089 * DATA-OUT buffer descriptor format. 1090 */ 1091 if (srp_cmd->buf_fmt & 0xf) 1092 /* DATA-IN: transfer data from target to initiator (read). */ 1093 *dir = DMA_FROM_DEVICE; 1094 else if (srp_cmd->buf_fmt >> 4) 1095 /* DATA-OUT: transfer data from initiator to target (write). */ 1096 *dir = DMA_TO_DEVICE; 1097 else 1098 *dir = DMA_NONE; 1099 1100 /* initialize data_direction early as srpt_alloc_rw_ctxs needs it */ 1101 ioctx->cmd.data_direction = *dir; 1102 1103 if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) || 1104 ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) { 1105 struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd); 1106 1107 *data_len = be32_to_cpu(db->len); 1108 return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt); 1109 } else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) || 1110 ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) { 1111 struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd); 1112 int nbufs = be32_to_cpu(idb->table_desc.len) / 1113 sizeof(struct srp_direct_buf); 1114 1115 if (nbufs > 1116 (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) { 1117 pr_err("received unsupported SRP_CMD request type (%u out + %u in != %u / %zu)\n", 1118 srp_cmd->data_out_desc_cnt, 1119 srp_cmd->data_in_desc_cnt, 1120 be32_to_cpu(idb->table_desc.len), 1121 sizeof(struct srp_direct_buf)); 1122 return -EINVAL; 1123 } 1124 1125 *data_len = be32_to_cpu(idb->len); 1126 return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs, 1127 sg, sg_cnt); 1128 } else if ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_IMM) { 1129 struct srp_imm_buf *imm_buf = srpt_get_desc_buf(srp_cmd); 1130 void *data = (void *)srp_cmd + imm_data_offset; 1131 uint32_t len = be32_to_cpu(imm_buf->len); 1132 uint32_t req_size = imm_data_offset + len; 1133 1134 if (req_size > srp_max_req_size) { 1135 pr_err("Immediate data (length %d + %d) exceeds request size %d\n", 1136 imm_data_offset, len, srp_max_req_size); 1137 return -EINVAL; 1138 } 1139 if (recv_ioctx->byte_len < req_size) { 1140 pr_err("Received too few data - %d < %d\n", 1141 recv_ioctx->byte_len, req_size); 1142 return -EIO; 1143 } 1144 /* 1145 * The immediate data buffer descriptor must occur before the 1146 * immediate data itself. 1147 */ 1148 if ((void *)(imm_buf + 1) > (void *)data) { 1149 pr_err("Received invalid write request\n"); 1150 return -EINVAL; 1151 } 1152 *data_len = len; 1153 ioctx->recv_ioctx = recv_ioctx; 1154 if ((uintptr_t)data & 511) { 1155 pr_warn_once("Internal error - the receive buffers are not aligned properly.\n"); 1156 return -EINVAL; 1157 } 1158 sg_init_one(&ioctx->imm_sg, data, len); 1159 *sg = &ioctx->imm_sg; 1160 *sg_cnt = 1; 1161 return 0; 1162 } else { 1163 *data_len = 0; 1164 return 0; 1165 } 1166 } 1167 1168 /** 1169 * srpt_init_ch_qp - initialize queue pair attributes 1170 * @ch: SRPT RDMA channel. 1171 * @qp: Queue pair pointer. 1172 * 1173 * Initialized the attributes of queue pair 'qp' by allowing local write, 1174 * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT. 1175 */ 1176 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1177 { 1178 struct ib_qp_attr *attr; 1179 int ret; 1180 1181 WARN_ON_ONCE(ch->using_rdma_cm); 1182 1183 attr = kzalloc_obj(*attr); 1184 if (!attr) 1185 return -ENOMEM; 1186 1187 attr->qp_state = IB_QPS_INIT; 1188 attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE; 1189 attr->port_num = ch->sport->port; 1190 1191 ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port, 1192 ch->pkey, &attr->pkey_index); 1193 if (ret < 0) 1194 pr_err("Translating pkey %#x failed (%d) - using index 0\n", 1195 ch->pkey, ret); 1196 1197 ret = ib_modify_qp(qp, attr, 1198 IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT | 1199 IB_QP_PKEY_INDEX); 1200 1201 kfree(attr); 1202 return ret; 1203 } 1204 1205 /** 1206 * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR) 1207 * @ch: channel of the queue pair. 1208 * @qp: queue pair to change the state of. 1209 * 1210 * Returns zero upon success and a negative value upon failure. 1211 * 1212 * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system. 1213 * If this structure ever becomes larger, it might be necessary to allocate 1214 * it dynamically instead of on the stack. 1215 */ 1216 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1217 { 1218 struct ib_qp_attr qp_attr; 1219 int attr_mask; 1220 int ret; 1221 1222 WARN_ON_ONCE(ch->using_rdma_cm); 1223 1224 qp_attr.qp_state = IB_QPS_RTR; 1225 ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask); 1226 if (ret) 1227 goto out; 1228 1229 qp_attr.max_dest_rd_atomic = 4; 1230 1231 ret = ib_modify_qp(qp, &qp_attr, attr_mask); 1232 1233 out: 1234 return ret; 1235 } 1236 1237 /** 1238 * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS) 1239 * @ch: channel of the queue pair. 1240 * @qp: queue pair to change the state of. 1241 * 1242 * Returns zero upon success and a negative value upon failure. 1243 * 1244 * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system. 1245 * If this structure ever becomes larger, it might be necessary to allocate 1246 * it dynamically instead of on the stack. 1247 */ 1248 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1249 { 1250 struct ib_qp_attr qp_attr; 1251 int attr_mask; 1252 int ret; 1253 1254 qp_attr.qp_state = IB_QPS_RTS; 1255 ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask); 1256 if (ret) 1257 goto out; 1258 1259 qp_attr.max_rd_atomic = 4; 1260 1261 ret = ib_modify_qp(qp, &qp_attr, attr_mask); 1262 1263 out: 1264 return ret; 1265 } 1266 1267 /** 1268 * srpt_ch_qp_err - set the channel queue pair state to 'error' 1269 * @ch: SRPT RDMA channel. 1270 */ 1271 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch) 1272 { 1273 struct ib_qp_attr qp_attr; 1274 1275 qp_attr.qp_state = IB_QPS_ERR; 1276 return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE); 1277 } 1278 1279 /** 1280 * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator 1281 * @ch: SRPT RDMA channel. 1282 */ 1283 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch) 1284 { 1285 struct srpt_send_ioctx *ioctx; 1286 int tag, cpu; 1287 1288 BUG_ON(!ch); 1289 1290 tag = sbitmap_queue_get(&ch->sess->sess_tag_pool, &cpu); 1291 if (tag < 0) 1292 return NULL; 1293 1294 ioctx = ch->ioctx_ring[tag]; 1295 BUG_ON(ioctx->ch != ch); 1296 ioctx->state = SRPT_STATE_NEW; 1297 WARN_ON_ONCE(ioctx->recv_ioctx); 1298 ioctx->n_rdma = 0; 1299 ioctx->n_rw_ctx = 0; 1300 ioctx->queue_status_only = false; 1301 /* 1302 * transport_init_se_cmd() does not initialize all fields, so do it 1303 * here. 1304 */ 1305 memset(&ioctx->cmd, 0, sizeof(ioctx->cmd)); 1306 memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data)); 1307 ioctx->cmd.map_tag = tag; 1308 ioctx->cmd.map_cpu = cpu; 1309 1310 return ioctx; 1311 } 1312 1313 /** 1314 * srpt_abort_cmd - abort a SCSI command 1315 * @ioctx: I/O context associated with the SCSI command. 1316 */ 1317 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx) 1318 { 1319 enum srpt_command_state state; 1320 1321 BUG_ON(!ioctx); 1322 1323 /* 1324 * If the command is in a state where the target core is waiting for 1325 * the ib_srpt driver, change the state to the next state. 1326 */ 1327 1328 state = ioctx->state; 1329 switch (state) { 1330 case SRPT_STATE_NEED_DATA: 1331 ioctx->state = SRPT_STATE_DATA_IN; 1332 break; 1333 case SRPT_STATE_CMD_RSP_SENT: 1334 case SRPT_STATE_MGMT_RSP_SENT: 1335 ioctx->state = SRPT_STATE_DONE; 1336 break; 1337 default: 1338 WARN_ONCE(true, "%s: unexpected I/O context state %d\n", 1339 __func__, state); 1340 break; 1341 } 1342 1343 pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state, 1344 ioctx->state, ioctx->cmd.tag); 1345 1346 switch (state) { 1347 case SRPT_STATE_NEW: 1348 case SRPT_STATE_DATA_IN: 1349 case SRPT_STATE_MGMT: 1350 case SRPT_STATE_DONE: 1351 /* 1352 * Do nothing - defer abort processing until 1353 * srpt_queue_response() is invoked. 1354 */ 1355 break; 1356 case SRPT_STATE_NEED_DATA: 1357 pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag); 1358 transport_generic_request_failure(&ioctx->cmd, 1359 TCM_CHECK_CONDITION_ABORT_CMD); 1360 break; 1361 case SRPT_STATE_CMD_RSP_SENT: 1362 /* 1363 * SRP_RSP sending failed or the SRP_RSP send completion has 1364 * not been received in time. 1365 */ 1366 transport_generic_free_cmd(&ioctx->cmd, 0); 1367 break; 1368 case SRPT_STATE_MGMT_RSP_SENT: 1369 transport_generic_free_cmd(&ioctx->cmd, 0); 1370 break; 1371 default: 1372 WARN(1, "Unexpected command state (%d)", state); 1373 break; 1374 } 1375 1376 return state; 1377 } 1378 1379 /** 1380 * srpt_rdma_read_done - RDMA read completion callback 1381 * @cq: Completion queue. 1382 * @wc: Work completion. 1383 * 1384 * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping 1385 * the data that has been transferred via IB RDMA had to be postponed until the 1386 * check_stop_free() callback. None of this is necessary anymore and needs to 1387 * be cleaned up. 1388 */ 1389 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc) 1390 { 1391 struct srpt_rdma_ch *ch = wc->qp->qp_context; 1392 struct srpt_send_ioctx *ioctx = 1393 container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe); 1394 1395 WARN_ON(ioctx->n_rdma <= 0); 1396 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail); 1397 ioctx->n_rdma = 0; 1398 1399 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1400 pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n", 1401 ioctx, wc->status); 1402 srpt_abort_cmd(ioctx); 1403 return; 1404 } 1405 1406 if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA, 1407 SRPT_STATE_DATA_IN)) 1408 target_execute_cmd(&ioctx->cmd); 1409 else 1410 pr_err("%s[%d]: wrong state = %d\n", __func__, 1411 __LINE__, ioctx->state); 1412 } 1413 1414 /** 1415 * srpt_build_cmd_rsp - build a SRP_RSP response 1416 * @ch: RDMA channel through which the request has been received. 1417 * @ioctx: I/O context associated with the SRP_CMD request. The response will 1418 * be built in the buffer ioctx->buf points at and hence this function will 1419 * overwrite the request data. 1420 * @tag: tag of the request for which this response is being generated. 1421 * @status: value for the STATUS field of the SRP_RSP information unit. 1422 * 1423 * Returns the size in bytes of the SRP_RSP response. 1424 * 1425 * An SRP_RSP response contains a SCSI status or service response. See also 1426 * section 6.9 in the SRP r16a document for the format of an SRP_RSP 1427 * response. See also SPC-2 for more information about sense data. 1428 */ 1429 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch, 1430 struct srpt_send_ioctx *ioctx, u64 tag, 1431 int status) 1432 { 1433 struct se_cmd *cmd = &ioctx->cmd; 1434 struct srp_rsp *srp_rsp; 1435 const u8 *sense_data; 1436 int sense_data_len, max_sense_len; 1437 u32 resid = cmd->residual_count; 1438 1439 /* 1440 * The lowest bit of all SAM-3 status codes is zero (see also 1441 * paragraph 5.3 in SAM-3). 1442 */ 1443 WARN_ON(status & 1); 1444 1445 srp_rsp = ioctx->ioctx.buf; 1446 BUG_ON(!srp_rsp); 1447 1448 sense_data = ioctx->sense_data; 1449 sense_data_len = ioctx->cmd.scsi_sense_length; 1450 WARN_ON(sense_data_len > sizeof(ioctx->sense_data)); 1451 1452 memset(srp_rsp, 0, sizeof(*srp_rsp)); 1453 srp_rsp->opcode = SRP_RSP; 1454 srp_rsp->req_lim_delta = 1455 cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0)); 1456 srp_rsp->tag = tag; 1457 srp_rsp->status = status; 1458 1459 if (cmd->se_cmd_flags & SCF_UNDERFLOW_BIT) { 1460 if (cmd->data_direction == DMA_TO_DEVICE) { 1461 /* residual data from an underflow write */ 1462 srp_rsp->flags = SRP_RSP_FLAG_DOUNDER; 1463 srp_rsp->data_out_res_cnt = cpu_to_be32(resid); 1464 } else if (cmd->data_direction == DMA_FROM_DEVICE) { 1465 /* residual data from an underflow read */ 1466 srp_rsp->flags = SRP_RSP_FLAG_DIUNDER; 1467 srp_rsp->data_in_res_cnt = cpu_to_be32(resid); 1468 } 1469 } else if (cmd->se_cmd_flags & SCF_OVERFLOW_BIT) { 1470 if (cmd->data_direction == DMA_TO_DEVICE) { 1471 /* residual data from an overflow write */ 1472 srp_rsp->flags = SRP_RSP_FLAG_DOOVER; 1473 srp_rsp->data_out_res_cnt = cpu_to_be32(resid); 1474 } else if (cmd->data_direction == DMA_FROM_DEVICE) { 1475 /* residual data from an overflow read */ 1476 srp_rsp->flags = SRP_RSP_FLAG_DIOVER; 1477 srp_rsp->data_in_res_cnt = cpu_to_be32(resid); 1478 } 1479 } 1480 1481 if (sense_data_len) { 1482 BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp)); 1483 max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp); 1484 if (sense_data_len > max_sense_len) { 1485 pr_warn("truncated sense data from %d to %d bytes\n", 1486 sense_data_len, max_sense_len); 1487 sense_data_len = max_sense_len; 1488 } 1489 1490 srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID; 1491 srp_rsp->sense_data_len = cpu_to_be32(sense_data_len); 1492 memcpy(srp_rsp->data, sense_data, sense_data_len); 1493 } 1494 1495 return sizeof(*srp_rsp) + sense_data_len; 1496 } 1497 1498 /** 1499 * srpt_build_tskmgmt_rsp - build a task management response 1500 * @ch: RDMA channel through which the request has been received. 1501 * @ioctx: I/O context in which the SRP_RSP response will be built. 1502 * @rsp_code: RSP_CODE that will be stored in the response. 1503 * @tag: Tag of the request for which this response is being generated. 1504 * 1505 * Returns the size in bytes of the SRP_RSP response. 1506 * 1507 * An SRP_RSP response contains a SCSI status or service response. See also 1508 * section 6.9 in the SRP r16a document for the format of an SRP_RSP 1509 * response. 1510 */ 1511 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch, 1512 struct srpt_send_ioctx *ioctx, 1513 u8 rsp_code, u64 tag) 1514 { 1515 struct srp_rsp *srp_rsp; 1516 int resp_data_len; 1517 int resp_len; 1518 1519 resp_data_len = 4; 1520 resp_len = sizeof(*srp_rsp) + resp_data_len; 1521 1522 srp_rsp = ioctx->ioctx.buf; 1523 BUG_ON(!srp_rsp); 1524 memset(srp_rsp, 0, sizeof(*srp_rsp)); 1525 1526 srp_rsp->opcode = SRP_RSP; 1527 srp_rsp->req_lim_delta = 1528 cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0)); 1529 srp_rsp->tag = tag; 1530 1531 srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID; 1532 srp_rsp->resp_data_len = cpu_to_be32(resp_data_len); 1533 srp_rsp->data[3] = rsp_code; 1534 1535 return resp_len; 1536 } 1537 1538 static int srpt_check_stop_free(struct se_cmd *cmd) 1539 { 1540 struct srpt_send_ioctx *ioctx = container_of(cmd, 1541 struct srpt_send_ioctx, cmd); 1542 1543 return target_put_sess_cmd(&ioctx->cmd); 1544 } 1545 1546 /** 1547 * srpt_handle_cmd - process a SRP_CMD information unit 1548 * @ch: SRPT RDMA channel. 1549 * @recv_ioctx: Receive I/O context. 1550 * @send_ioctx: Send I/O context. 1551 */ 1552 static void srpt_handle_cmd(struct srpt_rdma_ch *ch, 1553 struct srpt_recv_ioctx *recv_ioctx, 1554 struct srpt_send_ioctx *send_ioctx) 1555 { 1556 struct se_cmd *cmd; 1557 struct srp_cmd *srp_cmd; 1558 struct scatterlist *sg = NULL; 1559 unsigned sg_cnt = 0; 1560 u64 data_len; 1561 enum dma_data_direction dir; 1562 int rc; 1563 1564 BUG_ON(!send_ioctx); 1565 1566 srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset; 1567 cmd = &send_ioctx->cmd; 1568 cmd->tag = srp_cmd->tag; 1569 1570 switch (srp_cmd->task_attr) { 1571 case SRP_CMD_SIMPLE_Q: 1572 cmd->sam_task_attr = TCM_SIMPLE_TAG; 1573 break; 1574 case SRP_CMD_ORDERED_Q: 1575 default: 1576 cmd->sam_task_attr = TCM_ORDERED_TAG; 1577 break; 1578 case SRP_CMD_HEAD_OF_Q: 1579 cmd->sam_task_attr = TCM_HEAD_TAG; 1580 break; 1581 case SRP_CMD_ACA: 1582 cmd->sam_task_attr = TCM_ACA_TAG; 1583 break; 1584 } 1585 1586 rc = srpt_get_desc_tbl(recv_ioctx, send_ioctx, srp_cmd, &dir, 1587 &sg, &sg_cnt, &data_len, ch->imm_data_offset); 1588 if (rc) { 1589 if (rc != -EAGAIN) { 1590 pr_err("0x%llx: parsing SRP descriptor table failed.\n", 1591 srp_cmd->tag); 1592 } 1593 goto busy; 1594 } 1595 1596 rc = target_init_cmd(cmd, ch->sess, &send_ioctx->sense_data[0], 1597 scsilun_to_int(&srp_cmd->lun), data_len, 1598 TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF); 1599 if (rc != 0) { 1600 pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc, 1601 srp_cmd->tag); 1602 goto busy; 1603 } 1604 1605 if (target_submit_prep(cmd, srp_cmd->cdb, sg, sg_cnt, NULL, 0, NULL, 0, 1606 GFP_KERNEL)) 1607 return; 1608 1609 target_submit(cmd); 1610 return; 1611 1612 busy: 1613 target_send_busy(cmd); 1614 } 1615 1616 static int srp_tmr_to_tcm(int fn) 1617 { 1618 switch (fn) { 1619 case SRP_TSK_ABORT_TASK: 1620 return TMR_ABORT_TASK; 1621 case SRP_TSK_ABORT_TASK_SET: 1622 return TMR_ABORT_TASK_SET; 1623 case SRP_TSK_CLEAR_TASK_SET: 1624 return TMR_CLEAR_TASK_SET; 1625 case SRP_TSK_LUN_RESET: 1626 return TMR_LUN_RESET; 1627 case SRP_TSK_CLEAR_ACA: 1628 return TMR_CLEAR_ACA; 1629 default: 1630 return -1; 1631 } 1632 } 1633 1634 /** 1635 * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit 1636 * @ch: SRPT RDMA channel. 1637 * @recv_ioctx: Receive I/O context. 1638 * @send_ioctx: Send I/O context. 1639 * 1640 * Returns 0 if and only if the request will be processed by the target core. 1641 * 1642 * For more information about SRP_TSK_MGMT information units, see also section 1643 * 6.7 in the SRP r16a document. 1644 */ 1645 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch, 1646 struct srpt_recv_ioctx *recv_ioctx, 1647 struct srpt_send_ioctx *send_ioctx) 1648 { 1649 struct srp_tsk_mgmt *srp_tsk; 1650 struct se_cmd *cmd; 1651 struct se_session *sess = ch->sess; 1652 int tcm_tmr; 1653 int rc; 1654 1655 BUG_ON(!send_ioctx); 1656 1657 srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset; 1658 cmd = &send_ioctx->cmd; 1659 1660 pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n", 1661 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch, 1662 ch->sess); 1663 1664 srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT); 1665 send_ioctx->cmd.tag = srp_tsk->tag; 1666 tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func); 1667 rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, 1668 scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr, 1669 GFP_KERNEL, srp_tsk->task_tag, 1670 TARGET_SCF_ACK_KREF); 1671 if (rc != 0) { 1672 send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED; 1673 cmd->se_tfo->queue_tm_rsp(cmd); 1674 } 1675 return; 1676 } 1677 1678 /** 1679 * srpt_handle_new_iu - process a newly received information unit 1680 * @ch: RDMA channel through which the information unit has been received. 1681 * @recv_ioctx: Receive I/O context associated with the information unit. 1682 */ 1683 static bool 1684 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx) 1685 { 1686 struct srpt_send_ioctx *send_ioctx = NULL; 1687 struct srp_cmd *srp_cmd; 1688 bool res = false; 1689 u8 opcode; 1690 1691 BUG_ON(!ch); 1692 BUG_ON(!recv_ioctx); 1693 1694 if (unlikely(ch->state == CH_CONNECTING)) 1695 goto push; 1696 1697 ib_dma_sync_single_for_cpu(ch->sport->sdev->device, 1698 recv_ioctx->ioctx.dma, 1699 recv_ioctx->ioctx.offset + srp_max_req_size, 1700 DMA_FROM_DEVICE); 1701 1702 srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset; 1703 opcode = srp_cmd->opcode; 1704 if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) { 1705 send_ioctx = srpt_get_send_ioctx(ch); 1706 if (unlikely(!send_ioctx)) 1707 goto push; 1708 } 1709 1710 if (!list_empty(&recv_ioctx->wait_list)) { 1711 WARN_ON_ONCE(!ch->processing_wait_list); 1712 list_del_init(&recv_ioctx->wait_list); 1713 } 1714 1715 switch (opcode) { 1716 case SRP_CMD: 1717 srpt_handle_cmd(ch, recv_ioctx, send_ioctx); 1718 break; 1719 case SRP_TSK_MGMT: 1720 srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx); 1721 break; 1722 case SRP_I_LOGOUT: 1723 pr_err("Not yet implemented: SRP_I_LOGOUT\n"); 1724 break; 1725 case SRP_CRED_RSP: 1726 pr_debug("received SRP_CRED_RSP\n"); 1727 break; 1728 case SRP_AER_RSP: 1729 pr_debug("received SRP_AER_RSP\n"); 1730 break; 1731 case SRP_RSP: 1732 pr_err("Received SRP_RSP\n"); 1733 break; 1734 default: 1735 pr_err("received IU with unknown opcode 0x%x\n", opcode); 1736 break; 1737 } 1738 1739 if (!send_ioctx || !send_ioctx->recv_ioctx) 1740 srpt_post_recv(ch->sport->sdev, ch, recv_ioctx); 1741 res = true; 1742 1743 out: 1744 return res; 1745 1746 push: 1747 if (list_empty(&recv_ioctx->wait_list)) { 1748 WARN_ON_ONCE(ch->processing_wait_list); 1749 list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list); 1750 } 1751 goto out; 1752 } 1753 1754 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc) 1755 { 1756 struct srpt_rdma_ch *ch = wc->qp->qp_context; 1757 struct srpt_recv_ioctx *ioctx = 1758 container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe); 1759 1760 if (wc->status == IB_WC_SUCCESS) { 1761 int req_lim; 1762 1763 req_lim = atomic_dec_return(&ch->req_lim); 1764 if (unlikely(req_lim < 0)) 1765 pr_err("req_lim = %d < 0\n", req_lim); 1766 ioctx->byte_len = wc->byte_len; 1767 srpt_handle_new_iu(ch, ioctx); 1768 } else { 1769 pr_info_ratelimited("receiving failed for ioctx %p with status %d\n", 1770 ioctx, wc->status); 1771 } 1772 } 1773 1774 /* 1775 * This function must be called from the context in which RDMA completions are 1776 * processed because it accesses the wait list without protection against 1777 * access from other threads. 1778 */ 1779 static void srpt_process_wait_list(struct srpt_rdma_ch *ch) 1780 { 1781 struct srpt_recv_ioctx *recv_ioctx, *tmp; 1782 1783 WARN_ON_ONCE(ch->state == CH_CONNECTING); 1784 1785 if (list_empty(&ch->cmd_wait_list)) 1786 return; 1787 1788 WARN_ON_ONCE(ch->processing_wait_list); 1789 ch->processing_wait_list = true; 1790 list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list, 1791 wait_list) { 1792 if (!srpt_handle_new_iu(ch, recv_ioctx)) 1793 break; 1794 } 1795 ch->processing_wait_list = false; 1796 } 1797 1798 /** 1799 * srpt_send_done - send completion callback 1800 * @cq: Completion queue. 1801 * @wc: Work completion. 1802 * 1803 * Note: Although this has not yet been observed during tests, at least in 1804 * theory it is possible that the srpt_get_send_ioctx() call invoked by 1805 * srpt_handle_new_iu() fails. This is possible because the req_lim_delta 1806 * value in each response is set to one, and it is possible that this response 1807 * makes the initiator send a new request before the send completion for that 1808 * response has been processed. This could e.g. happen if the call to 1809 * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or 1810 * if IB retransmission causes generation of the send completion to be 1811 * delayed. Incoming information units for which srpt_get_send_ioctx() fails 1812 * are queued on cmd_wait_list. The code below processes these delayed 1813 * requests one at a time. 1814 */ 1815 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc) 1816 { 1817 struct srpt_rdma_ch *ch = wc->qp->qp_context; 1818 struct srpt_send_ioctx *ioctx = 1819 container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe); 1820 enum srpt_command_state state; 1821 1822 state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE); 1823 1824 WARN_ON(state != SRPT_STATE_CMD_RSP_SENT && 1825 state != SRPT_STATE_MGMT_RSP_SENT); 1826 1827 atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail); 1828 1829 if (wc->status != IB_WC_SUCCESS) 1830 pr_info("sending response for ioctx 0x%p failed with status %d\n", 1831 ioctx, wc->status); 1832 1833 if (state != SRPT_STATE_DONE) { 1834 transport_generic_free_cmd(&ioctx->cmd, 0); 1835 } else { 1836 pr_err("IB completion has been received too late for wr_id = %u.\n", 1837 ioctx->ioctx.index); 1838 } 1839 1840 srpt_process_wait_list(ch); 1841 } 1842 1843 /** 1844 * srpt_create_ch_ib - create receive and send completion queues 1845 * @ch: SRPT RDMA channel. 1846 */ 1847 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch) 1848 { 1849 struct ib_qp_init_attr *qp_init; 1850 struct srpt_port *sport = ch->sport; 1851 struct srpt_device *sdev = sport->sdev; 1852 const struct ib_device_attr *attrs = &sdev->device->attrs; 1853 int sq_size = sport->port_attrib.srp_sq_size; 1854 int i, ret; 1855 1856 WARN_ON(ch->rq_size < 1); 1857 1858 ret = -ENOMEM; 1859 qp_init = kzalloc_obj(*qp_init); 1860 if (!qp_init) 1861 goto out; 1862 1863 retry: 1864 ch->cq = ib_cq_pool_get(sdev->device, ch->rq_size + sq_size, -1, 1865 IB_POLL_WORKQUEUE); 1866 if (IS_ERR(ch->cq)) { 1867 ret = PTR_ERR(ch->cq); 1868 pr_err("failed to create CQ cqe= %d ret= %pe\n", 1869 ch->rq_size + sq_size, ch->cq); 1870 goto out; 1871 } 1872 ch->cq_size = ch->rq_size + sq_size; 1873 1874 qp_init->qp_context = (void *)ch; 1875 qp_init->event_handler = srpt_qp_event; 1876 qp_init->send_cq = ch->cq; 1877 qp_init->recv_cq = ch->cq; 1878 qp_init->sq_sig_type = IB_SIGNAL_REQ_WR; 1879 qp_init->qp_type = IB_QPT_RC; 1880 /* 1881 * We divide up our send queue size into half SEND WRs to send the 1882 * completions, and half R/W contexts to actually do the RDMA 1883 * READ/WRITE transfers. Note that we need to allocate CQ slots for 1884 * both both, as RDMA contexts will also post completions for the 1885 * RDMA READ case. 1886 */ 1887 qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr); 1888 qp_init->cap.max_rdma_ctxs = sq_size / 2; 1889 qp_init->cap.max_send_sge = attrs->max_send_sge; 1890 qp_init->cap.max_recv_sge = 1; 1891 qp_init->port_num = ch->sport->port; 1892 if (sdev->use_srq) 1893 qp_init->srq = sdev->srq; 1894 else 1895 qp_init->cap.max_recv_wr = ch->rq_size; 1896 1897 if (ch->using_rdma_cm) { 1898 ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init); 1899 ch->qp = ch->rdma_cm.cm_id->qp; 1900 } else { 1901 ch->qp = ib_create_qp(sdev->pd, qp_init); 1902 if (!IS_ERR(ch->qp)) { 1903 ret = srpt_init_ch_qp(ch, ch->qp); 1904 if (ret) 1905 ib_destroy_qp(ch->qp); 1906 } else { 1907 ret = PTR_ERR(ch->qp); 1908 } 1909 } 1910 if (ret) { 1911 bool retry = sq_size > MIN_SRPT_SQ_SIZE; 1912 1913 if (retry) { 1914 pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n", 1915 sq_size, ret); 1916 ib_cq_pool_put(ch->cq, ch->cq_size); 1917 sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE); 1918 goto retry; 1919 } else { 1920 pr_err("failed to create queue pair with sq_size = %d (%d)\n", 1921 sq_size, ret); 1922 goto err_destroy_cq; 1923 } 1924 } 1925 1926 atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr); 1927 1928 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n", 1929 __func__, ch->cq->cqe, qp_init->cap.max_send_sge, 1930 qp_init->cap.max_send_wr, ch); 1931 1932 if (!sdev->use_srq) 1933 for (i = 0; i < ch->rq_size; i++) 1934 srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]); 1935 1936 out: 1937 kfree(qp_init); 1938 return ret; 1939 1940 err_destroy_cq: 1941 ch->qp = NULL; 1942 ib_cq_pool_put(ch->cq, ch->cq_size); 1943 goto out; 1944 } 1945 1946 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch) 1947 { 1948 ib_destroy_qp(ch->qp); 1949 ib_cq_pool_put(ch->cq, ch->cq_size); 1950 } 1951 1952 /** 1953 * srpt_close_ch - close a RDMA channel 1954 * @ch: SRPT RDMA channel. 1955 * 1956 * Make sure all resources associated with the channel will be deallocated at 1957 * an appropriate time. 1958 * 1959 * Returns true if and only if the channel state has been modified into 1960 * CH_DRAINING. 1961 */ 1962 static bool srpt_close_ch(struct srpt_rdma_ch *ch) 1963 { 1964 int ret; 1965 1966 if (!srpt_set_ch_state(ch, CH_DRAINING)) { 1967 pr_debug("%s: already closed\n", ch->sess_name); 1968 return false; 1969 } 1970 1971 kref_get(&ch->kref); 1972 1973 ret = srpt_ch_qp_err(ch); 1974 if (ret < 0) 1975 pr_err("%s-%d: changing queue pair into error state failed: %d\n", 1976 ch->sess_name, ch->qp->qp_num, ret); 1977 1978 ret = srpt_zerolength_write(ch); 1979 if (ret < 0) { 1980 pr_err("%s-%d: queuing zero-length write failed: %d\n", 1981 ch->sess_name, ch->qp->qp_num, ret); 1982 if (srpt_set_ch_state(ch, CH_DISCONNECTED)) 1983 schedule_work(&ch->release_work); 1984 else 1985 WARN_ON_ONCE(true); 1986 } 1987 1988 kref_put(&ch->kref, srpt_free_ch); 1989 1990 return true; 1991 } 1992 1993 /* 1994 * Change the channel state into CH_DISCONNECTING. If a channel has not yet 1995 * reached the connected state, close it. If a channel is in the connected 1996 * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is 1997 * the responsibility of the caller to ensure that this function is not 1998 * invoked concurrently with the code that accepts a connection. This means 1999 * that this function must either be invoked from inside a CM callback 2000 * function or that it must be invoked with the srpt_port.mutex held. 2001 */ 2002 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch) 2003 { 2004 int ret; 2005 2006 if (!srpt_set_ch_state(ch, CH_DISCONNECTING)) 2007 return -ENOTCONN; 2008 2009 if (ch->using_rdma_cm) { 2010 ret = rdma_disconnect(ch->rdma_cm.cm_id); 2011 } else { 2012 ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0); 2013 if (ret < 0) 2014 ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0); 2015 } 2016 2017 if (ret < 0 && srpt_close_ch(ch)) 2018 ret = 0; 2019 2020 return ret; 2021 } 2022 2023 /* Send DREQ and wait for DREP. */ 2024 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch) 2025 { 2026 DECLARE_COMPLETION_ONSTACK(closed); 2027 struct srpt_port *sport = ch->sport; 2028 2029 pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num, 2030 ch->state); 2031 2032 ch->closed = &closed; 2033 2034 mutex_lock(&sport->mutex); 2035 srpt_disconnect_ch(ch); 2036 mutex_unlock(&sport->mutex); 2037 2038 while (wait_for_completion_timeout(&closed, 5 * HZ) == 0) 2039 pr_info("%s(%s-%d state %d): still waiting ...\n", __func__, 2040 ch->sess_name, ch->qp->qp_num, ch->state); 2041 2042 } 2043 2044 static void __srpt_close_all_ch(struct srpt_port *sport) 2045 { 2046 struct srpt_nexus *nexus; 2047 struct srpt_rdma_ch *ch; 2048 2049 lockdep_assert_held(&sport->mutex); 2050 2051 list_for_each_entry(nexus, &sport->nexus_list, entry) { 2052 list_for_each_entry(ch, &nexus->ch_list, list) { 2053 if (srpt_disconnect_ch(ch) >= 0) 2054 pr_info("Closing channel %s-%d because target %s_%d has been disabled\n", 2055 ch->sess_name, ch->qp->qp_num, 2056 dev_name(&sport->sdev->device->dev), 2057 sport->port); 2058 srpt_close_ch(ch); 2059 } 2060 } 2061 } 2062 2063 /* 2064 * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if 2065 * it does not yet exist. 2066 */ 2067 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport, 2068 const u8 i_port_id[16], 2069 const u8 t_port_id[16]) 2070 { 2071 struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n; 2072 2073 for (;;) { 2074 mutex_lock(&sport->mutex); 2075 list_for_each_entry(n, &sport->nexus_list, entry) { 2076 if (memcmp(n->i_port_id, i_port_id, 16) == 0 && 2077 memcmp(n->t_port_id, t_port_id, 16) == 0) { 2078 nexus = n; 2079 break; 2080 } 2081 } 2082 if (!nexus && tmp_nexus) { 2083 list_add_tail_rcu(&tmp_nexus->entry, 2084 &sport->nexus_list); 2085 swap(nexus, tmp_nexus); 2086 } 2087 mutex_unlock(&sport->mutex); 2088 2089 if (nexus) 2090 break; 2091 tmp_nexus = kzalloc_obj(*nexus); 2092 if (!tmp_nexus) { 2093 nexus = ERR_PTR(-ENOMEM); 2094 break; 2095 } 2096 INIT_LIST_HEAD(&tmp_nexus->ch_list); 2097 memcpy(tmp_nexus->i_port_id, i_port_id, 16); 2098 memcpy(tmp_nexus->t_port_id, t_port_id, 16); 2099 } 2100 2101 kfree(tmp_nexus); 2102 2103 return nexus; 2104 } 2105 2106 static void srpt_set_enabled(struct srpt_port *sport, bool enabled) 2107 __must_hold(&sport->mutex) 2108 { 2109 lockdep_assert_held(&sport->mutex); 2110 2111 if (sport->enabled == enabled) 2112 return; 2113 sport->enabled = enabled; 2114 if (!enabled) 2115 __srpt_close_all_ch(sport); 2116 } 2117 2118 static void srpt_drop_sport_ref(struct srpt_port *sport) 2119 { 2120 if (atomic_dec_return(&sport->refcount) == 0 && sport->freed_channels) 2121 complete(sport->freed_channels); 2122 } 2123 2124 static void srpt_free_ch(struct kref *kref) 2125 { 2126 struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref); 2127 2128 srpt_drop_sport_ref(ch->sport); 2129 kfree_rcu(ch, rcu); 2130 } 2131 2132 /* 2133 * Shut down the SCSI target session, tell the connection manager to 2134 * disconnect the associated RDMA channel, transition the QP to the error 2135 * state and remove the channel from the channel list. This function is 2136 * typically called from inside srpt_zerolength_write_done(). Concurrent 2137 * srpt_zerolength_write() calls from inside srpt_close_ch() are possible 2138 * as long as the channel is on sport->nexus_list. 2139 */ 2140 static void srpt_release_channel_work(struct work_struct *w) 2141 { 2142 struct srpt_rdma_ch *ch; 2143 struct srpt_device *sdev; 2144 struct srpt_port *sport; 2145 struct se_session *se_sess; 2146 2147 ch = container_of(w, struct srpt_rdma_ch, release_work); 2148 pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num); 2149 2150 sdev = ch->sport->sdev; 2151 BUG_ON(!sdev); 2152 2153 se_sess = ch->sess; 2154 BUG_ON(!se_sess); 2155 2156 target_stop_session(se_sess); 2157 target_wait_for_sess_cmds(se_sess); 2158 2159 target_remove_session(se_sess); 2160 ch->sess = NULL; 2161 2162 if (ch->using_rdma_cm) 2163 rdma_destroy_id(ch->rdma_cm.cm_id); 2164 else 2165 ib_destroy_cm_id(ch->ib_cm.cm_id); 2166 2167 sport = ch->sport; 2168 mutex_lock(&sport->mutex); 2169 list_del_rcu(&ch->list); 2170 mutex_unlock(&sport->mutex); 2171 2172 if (ch->closed) 2173 complete(ch->closed); 2174 2175 srpt_destroy_ch_ib(ch); 2176 2177 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring, 2178 ch->sport->sdev, ch->rq_size, 2179 ch->rsp_buf_cache, DMA_TO_DEVICE); 2180 2181 srpt_cache_put(ch->rsp_buf_cache); 2182 2183 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring, 2184 sdev, ch->rq_size, 2185 ch->req_buf_cache, DMA_FROM_DEVICE); 2186 2187 srpt_cache_put(ch->req_buf_cache); 2188 2189 kref_put(&ch->kref, srpt_free_ch); 2190 } 2191 2192 /** 2193 * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED 2194 * @sdev: HCA through which the login request was received. 2195 * @ib_cm_id: IB/CM connection identifier in case of IB/CM. 2196 * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM. 2197 * @port_num: Port through which the REQ message was received. 2198 * @pkey: P_Key of the incoming connection. 2199 * @req: SRP login request. 2200 * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted 2201 * the login request. 2202 * 2203 * Ownership of the cm_id is transferred to the target session if this 2204 * function returns zero. Otherwise the caller remains the owner of cm_id. 2205 */ 2206 static int srpt_cm_req_recv(struct srpt_device *const sdev, 2207 struct ib_cm_id *ib_cm_id, 2208 struct rdma_cm_id *rdma_cm_id, 2209 u8 port_num, __be16 pkey, 2210 const struct srp_login_req *req, 2211 const char *src_addr) 2212 { 2213 struct srpt_port *sport = &sdev->port[port_num - 1]; 2214 struct srpt_nexus *nexus; 2215 struct srp_login_rsp *rsp = NULL; 2216 struct srp_login_rej *rej = NULL; 2217 union { 2218 struct rdma_conn_param rdma_cm; 2219 struct ib_cm_rep_param ib_cm; 2220 } *rep_param = NULL; 2221 struct srpt_rdma_ch *ch = NULL; 2222 char i_port_id[36]; 2223 u32 it_iu_len; 2224 int i, tag_num, tag_size, ret; 2225 struct srpt_tpg *stpg; 2226 2227 WARN_ON_ONCE(irqs_disabled()); 2228 2229 it_iu_len = be32_to_cpu(req->req_it_iu_len); 2230 2231 pr_info("Received SRP_LOGIN_REQ with i_port_id %pI6, t_port_id %pI6 and it_iu_len %d on port %d (guid=%pI6); pkey %#04x\n", 2232 req->initiator_port_id, req->target_port_id, it_iu_len, 2233 port_num, &sport->gid, be16_to_cpu(pkey)); 2234 2235 nexus = srpt_get_nexus(sport, req->initiator_port_id, 2236 req->target_port_id); 2237 if (IS_ERR(nexus)) { 2238 ret = PTR_ERR(nexus); 2239 goto out; 2240 } 2241 2242 ret = -ENOMEM; 2243 rsp = kzalloc_obj(*rsp); 2244 rej = kzalloc_obj(*rej); 2245 rep_param = kzalloc_obj(*rep_param); 2246 if (!rsp || !rej || !rep_param) 2247 goto out; 2248 2249 ret = -EINVAL; 2250 if (it_iu_len > srp_max_req_size || it_iu_len < 64) { 2251 rej->reason = cpu_to_be32( 2252 SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE); 2253 pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n", 2254 it_iu_len, 64, srp_max_req_size); 2255 goto reject; 2256 } 2257 2258 if (!sport->enabled) { 2259 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2260 pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n", 2261 dev_name(&sport->sdev->device->dev), port_num); 2262 goto reject; 2263 } 2264 2265 if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid) 2266 || *(__be64 *)(req->target_port_id + 8) != 2267 cpu_to_be64(srpt_service_guid)) { 2268 rej->reason = cpu_to_be32( 2269 SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL); 2270 pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n"); 2271 goto reject; 2272 } 2273 2274 ret = -ENOMEM; 2275 ch = kzalloc_obj(*ch); 2276 if (!ch) { 2277 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2278 pr_err("rejected SRP_LOGIN_REQ because out of memory.\n"); 2279 goto reject; 2280 } 2281 2282 kref_init(&ch->kref); 2283 ch->pkey = be16_to_cpu(pkey); 2284 ch->nexus = nexus; 2285 ch->zw_cqe.done = srpt_zerolength_write_done; 2286 INIT_WORK(&ch->release_work, srpt_release_channel_work); 2287 ch->sport = sport; 2288 if (rdma_cm_id) { 2289 ch->using_rdma_cm = true; 2290 ch->rdma_cm.cm_id = rdma_cm_id; 2291 rdma_cm_id->context = ch; 2292 } else { 2293 ch->ib_cm.cm_id = ib_cm_id; 2294 ib_cm_id->context = ch; 2295 } 2296 /* 2297 * ch->rq_size should be at least as large as the initiator queue 2298 * depth to avoid that the initiator driver has to report QUEUE_FULL 2299 * to the SCSI mid-layer. 2300 */ 2301 ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr); 2302 spin_lock_init(&ch->spinlock); 2303 ch->state = CH_CONNECTING; 2304 INIT_LIST_HEAD(&ch->cmd_wait_list); 2305 ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size; 2306 2307 ch->rsp_buf_cache = srpt_cache_get(ch->max_rsp_size); 2308 if (!ch->rsp_buf_cache) 2309 goto free_ch; 2310 2311 ch->ioctx_ring = (struct srpt_send_ioctx **) 2312 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size, 2313 sizeof(*ch->ioctx_ring[0]), 2314 ch->rsp_buf_cache, 0, DMA_TO_DEVICE); 2315 if (!ch->ioctx_ring) { 2316 pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n"); 2317 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2318 goto free_rsp_cache; 2319 } 2320 2321 for (i = 0; i < ch->rq_size; i++) 2322 ch->ioctx_ring[i]->ch = ch; 2323 if (!sdev->use_srq) { 2324 u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ? 2325 be16_to_cpu(req->imm_data_offset) : 0; 2326 u16 alignment_offset; 2327 u32 req_sz; 2328 2329 if (req->req_flags & SRP_IMMED_REQUESTED) 2330 pr_debug("imm_data_offset = %d\n", 2331 be16_to_cpu(req->imm_data_offset)); 2332 if (imm_data_offset >= sizeof(struct srp_cmd)) { 2333 ch->imm_data_offset = imm_data_offset; 2334 rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP; 2335 } else { 2336 ch->imm_data_offset = 0; 2337 } 2338 alignment_offset = round_up(imm_data_offset, 512) - 2339 imm_data_offset; 2340 req_sz = alignment_offset + imm_data_offset + srp_max_req_size; 2341 ch->req_buf_cache = srpt_cache_get(req_sz); 2342 if (!ch->req_buf_cache) 2343 goto free_rsp_ring; 2344 2345 ch->ioctx_recv_ring = (struct srpt_recv_ioctx **) 2346 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size, 2347 sizeof(*ch->ioctx_recv_ring[0]), 2348 ch->req_buf_cache, 2349 alignment_offset, 2350 DMA_FROM_DEVICE); 2351 if (!ch->ioctx_recv_ring) { 2352 pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n"); 2353 rej->reason = 2354 cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2355 goto free_recv_cache; 2356 } 2357 for (i = 0; i < ch->rq_size; i++) 2358 INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list); 2359 } 2360 2361 ret = srpt_create_ch_ib(ch); 2362 if (ret) { 2363 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2364 pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n"); 2365 goto free_recv_ring; 2366 } 2367 2368 strscpy(ch->sess_name, src_addr, sizeof(ch->sess_name)); 2369 snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx", 2370 be64_to_cpu(*(__be64 *)nexus->i_port_id), 2371 be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8))); 2372 2373 pr_debug("registering src addr %s or i_port_id %s\n", ch->sess_name, 2374 i_port_id); 2375 2376 tag_num = ch->rq_size; 2377 tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */ 2378 2379 if (sport->guid_id) { 2380 mutex_lock(&sport->guid_id->mutex); 2381 list_for_each_entry(stpg, &sport->guid_id->tpg_list, entry) { 2382 if (!IS_ERR_OR_NULL(ch->sess)) 2383 break; 2384 ch->sess = target_setup_session(&stpg->tpg, tag_num, 2385 tag_size, TARGET_PROT_NORMAL, 2386 ch->sess_name, ch, NULL); 2387 } 2388 mutex_unlock(&sport->guid_id->mutex); 2389 } 2390 2391 if (sport->gid_id) { 2392 mutex_lock(&sport->gid_id->mutex); 2393 list_for_each_entry(stpg, &sport->gid_id->tpg_list, entry) { 2394 if (!IS_ERR_OR_NULL(ch->sess)) 2395 break; 2396 ch->sess = target_setup_session(&stpg->tpg, tag_num, 2397 tag_size, TARGET_PROT_NORMAL, i_port_id, 2398 ch, NULL); 2399 if (!IS_ERR_OR_NULL(ch->sess)) 2400 break; 2401 /* Retry without leading "0x" */ 2402 ch->sess = target_setup_session(&stpg->tpg, tag_num, 2403 tag_size, TARGET_PROT_NORMAL, 2404 i_port_id + 2, ch, NULL); 2405 } 2406 mutex_unlock(&sport->gid_id->mutex); 2407 } 2408 2409 if (IS_ERR_OR_NULL(ch->sess)) { 2410 WARN_ON_ONCE(ch->sess == NULL); 2411 ret = PTR_ERR(ch->sess); 2412 ch->sess = NULL; 2413 pr_info("Rejected login for initiator %s: ret = %d.\n", 2414 ch->sess_name, ret); 2415 rej->reason = cpu_to_be32(ret == -ENOMEM ? 2416 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES : 2417 SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED); 2418 goto destroy_ib; 2419 } 2420 2421 /* 2422 * Once a session has been created destruction of srpt_rdma_ch objects 2423 * will decrement sport->refcount. Hence increment sport->refcount now. 2424 */ 2425 atomic_inc(&sport->refcount); 2426 2427 mutex_lock(&sport->mutex); 2428 2429 if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) { 2430 struct srpt_rdma_ch *ch2; 2431 2432 list_for_each_entry(ch2, &nexus->ch_list, list) { 2433 if (srpt_disconnect_ch(ch2) < 0) 2434 continue; 2435 pr_info("Relogin - closed existing channel %s\n", 2436 ch2->sess_name); 2437 rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED; 2438 } 2439 } else { 2440 rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED; 2441 } 2442 2443 list_add_tail_rcu(&ch->list, &nexus->ch_list); 2444 2445 if (!sport->enabled) { 2446 rej->reason = cpu_to_be32( 2447 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2448 pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n", 2449 dev_name(&sdev->device->dev), port_num); 2450 mutex_unlock(&sport->mutex); 2451 ret = -EINVAL; 2452 goto reject; 2453 } 2454 2455 mutex_unlock(&sport->mutex); 2456 2457 ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp); 2458 if (ret) { 2459 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2460 pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n", 2461 ret); 2462 goto reject; 2463 } 2464 2465 pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess, 2466 ch->sess_name, ch); 2467 2468 /* create srp_login_response */ 2469 rsp->opcode = SRP_LOGIN_RSP; 2470 rsp->tag = req->tag; 2471 rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size); 2472 rsp->max_ti_iu_len = req->req_it_iu_len; 2473 ch->max_ti_iu_len = it_iu_len; 2474 rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT | 2475 SRP_BUF_FORMAT_INDIRECT); 2476 rsp->req_lim_delta = cpu_to_be32(ch->rq_size); 2477 atomic_set(&ch->req_lim, ch->rq_size); 2478 atomic_set(&ch->req_lim_delta, 0); 2479 2480 /* create cm reply */ 2481 if (ch->using_rdma_cm) { 2482 rep_param->rdma_cm.private_data = (void *)rsp; 2483 rep_param->rdma_cm.private_data_len = sizeof(*rsp); 2484 rep_param->rdma_cm.rnr_retry_count = 7; 2485 rep_param->rdma_cm.flow_control = 1; 2486 rep_param->rdma_cm.responder_resources = 4; 2487 rep_param->rdma_cm.initiator_depth = 4; 2488 } else { 2489 rep_param->ib_cm.qp_num = ch->qp->qp_num; 2490 rep_param->ib_cm.private_data = (void *)rsp; 2491 rep_param->ib_cm.private_data_len = sizeof(*rsp); 2492 rep_param->ib_cm.rnr_retry_count = 7; 2493 rep_param->ib_cm.flow_control = 1; 2494 rep_param->ib_cm.failover_accepted = 0; 2495 rep_param->ib_cm.srq = 1; 2496 rep_param->ib_cm.responder_resources = 4; 2497 rep_param->ib_cm.initiator_depth = 4; 2498 } 2499 2500 /* 2501 * Hold the sport mutex while accepting a connection to avoid that 2502 * srpt_disconnect_ch() is invoked concurrently with this code. 2503 */ 2504 mutex_lock(&sport->mutex); 2505 if (sport->enabled && ch->state == CH_CONNECTING) { 2506 if (ch->using_rdma_cm) 2507 ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm); 2508 else 2509 ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm); 2510 } else { 2511 ret = -EINVAL; 2512 } 2513 mutex_unlock(&sport->mutex); 2514 2515 switch (ret) { 2516 case 0: 2517 break; 2518 case -EINVAL: 2519 goto reject; 2520 default: 2521 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2522 pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n", 2523 ret); 2524 goto reject; 2525 } 2526 2527 goto out; 2528 2529 destroy_ib: 2530 srpt_destroy_ch_ib(ch); 2531 2532 free_recv_ring: 2533 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring, 2534 ch->sport->sdev, ch->rq_size, 2535 ch->req_buf_cache, DMA_FROM_DEVICE); 2536 2537 free_recv_cache: 2538 srpt_cache_put(ch->req_buf_cache); 2539 2540 free_rsp_ring: 2541 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring, 2542 ch->sport->sdev, ch->rq_size, 2543 ch->rsp_buf_cache, DMA_TO_DEVICE); 2544 2545 free_rsp_cache: 2546 srpt_cache_put(ch->rsp_buf_cache); 2547 2548 free_ch: 2549 if (rdma_cm_id) 2550 rdma_cm_id->context = NULL; 2551 else 2552 ib_cm_id->context = NULL; 2553 kfree(ch); 2554 ch = NULL; 2555 2556 WARN_ON_ONCE(ret == 0); 2557 2558 reject: 2559 pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason)); 2560 rej->opcode = SRP_LOGIN_REJ; 2561 rej->tag = req->tag; 2562 rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT | 2563 SRP_BUF_FORMAT_INDIRECT); 2564 2565 if (rdma_cm_id) 2566 rdma_reject(rdma_cm_id, rej, sizeof(*rej), 2567 IB_CM_REJ_CONSUMER_DEFINED); 2568 else 2569 ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0, 2570 rej, sizeof(*rej)); 2571 2572 if (ch && ch->sess) { 2573 srpt_close_ch(ch); 2574 /* 2575 * Tell the caller not to free cm_id since 2576 * srpt_release_channel_work() will do that. 2577 */ 2578 ret = 0; 2579 } 2580 2581 out: 2582 kfree(rep_param); 2583 kfree(rsp); 2584 kfree(rej); 2585 2586 return ret; 2587 } 2588 2589 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id, 2590 const struct ib_cm_req_event_param *param, 2591 void *private_data) 2592 { 2593 char sguid[40]; 2594 2595 srpt_format_guid(sguid, sizeof(sguid), 2596 ¶m->primary_path->dgid.global.interface_id); 2597 2598 return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port, 2599 param->primary_path->pkey, 2600 private_data, sguid); 2601 } 2602 2603 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id, 2604 struct rdma_cm_event *event) 2605 { 2606 struct srpt_device *sdev; 2607 struct srp_login_req req; 2608 const struct srp_login_req_rdma *req_rdma; 2609 struct sa_path_rec *path_rec = cm_id->route.path_rec; 2610 char src_addr[40]; 2611 2612 sdev = ib_get_client_data(cm_id->device, &srpt_client); 2613 if (!sdev) 2614 return -ECONNREFUSED; 2615 2616 if (event->param.conn.private_data_len < sizeof(*req_rdma)) 2617 return -EINVAL; 2618 2619 /* Transform srp_login_req_rdma into srp_login_req. */ 2620 req_rdma = event->param.conn.private_data; 2621 memset(&req, 0, sizeof(req)); 2622 req.opcode = req_rdma->opcode; 2623 req.tag = req_rdma->tag; 2624 req.req_it_iu_len = req_rdma->req_it_iu_len; 2625 req.req_buf_fmt = req_rdma->req_buf_fmt; 2626 req.req_flags = req_rdma->req_flags; 2627 memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16); 2628 memcpy(req.target_port_id, req_rdma->target_port_id, 16); 2629 req.imm_data_offset = req_rdma->imm_data_offset; 2630 2631 snprintf(src_addr, sizeof(src_addr), "%pIS", 2632 &cm_id->route.addr.src_addr); 2633 2634 return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num, 2635 path_rec ? path_rec->pkey : 0, &req, src_addr); 2636 } 2637 2638 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch, 2639 enum ib_cm_rej_reason reason, 2640 const u8 *private_data, 2641 u8 private_data_len) 2642 { 2643 char *priv = NULL; 2644 int i; 2645 2646 if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1, 2647 GFP_KERNEL))) { 2648 for (i = 0; i < private_data_len; i++) 2649 sprintf(priv + 3 * i, " %02x", private_data[i]); 2650 } 2651 pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n", 2652 ch->sess_name, ch->qp->qp_num, reason, private_data_len ? 2653 "; private data" : "", priv ? priv : " (?)"); 2654 kfree(priv); 2655 } 2656 2657 /** 2658 * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event 2659 * @ch: SRPT RDMA channel. 2660 * 2661 * An RTU (ready to use) message indicates that the connection has been 2662 * established and that the recipient may begin transmitting. 2663 */ 2664 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch) 2665 { 2666 int ret; 2667 2668 ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp); 2669 if (ret < 0) { 2670 pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name, 2671 ch->qp->qp_num); 2672 srpt_close_ch(ch); 2673 return; 2674 } 2675 2676 /* 2677 * Note: calling srpt_close_ch() if the transition to the LIVE state 2678 * fails is not necessary since that means that that function has 2679 * already been invoked from another thread. 2680 */ 2681 if (!srpt_set_ch_state(ch, CH_LIVE)) { 2682 pr_err("%s-%d: channel transition to LIVE state failed\n", 2683 ch->sess_name, ch->qp->qp_num); 2684 return; 2685 } 2686 2687 /* Trigger wait list processing. */ 2688 ret = srpt_zerolength_write(ch); 2689 WARN_ONCE(ret < 0, "%d\n", ret); 2690 } 2691 2692 /** 2693 * srpt_cm_handler - IB connection manager callback function 2694 * @cm_id: IB/CM connection identifier. 2695 * @event: IB/CM event. 2696 * 2697 * A non-zero return value will cause the caller destroy the CM ID. 2698 * 2699 * Note: srpt_cm_handler() must only return a non-zero value when transferring 2700 * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning 2701 * a non-zero value in any other case will trigger a race with the 2702 * ib_destroy_cm_id() call in srpt_release_channel(). 2703 */ 2704 static int srpt_cm_handler(struct ib_cm_id *cm_id, 2705 const struct ib_cm_event *event) 2706 { 2707 struct srpt_rdma_ch *ch = cm_id->context; 2708 int ret; 2709 2710 ret = 0; 2711 switch (event->event) { 2712 case IB_CM_REQ_RECEIVED: 2713 ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd, 2714 event->private_data); 2715 break; 2716 case IB_CM_REJ_RECEIVED: 2717 srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason, 2718 event->private_data, 2719 IB_CM_REJ_PRIVATE_DATA_SIZE); 2720 break; 2721 case IB_CM_RTU_RECEIVED: 2722 case IB_CM_USER_ESTABLISHED: 2723 srpt_cm_rtu_recv(ch); 2724 break; 2725 case IB_CM_DREQ_RECEIVED: 2726 srpt_disconnect_ch(ch); 2727 break; 2728 case IB_CM_DREP_RECEIVED: 2729 pr_info("Received CM DREP message for ch %s-%d.\n", 2730 ch->sess_name, ch->qp->qp_num); 2731 srpt_close_ch(ch); 2732 break; 2733 case IB_CM_TIMEWAIT_EXIT: 2734 pr_info("Received CM TimeWait exit for ch %s-%d.\n", 2735 ch->sess_name, ch->qp->qp_num); 2736 srpt_close_ch(ch); 2737 break; 2738 case IB_CM_REP_ERROR: 2739 pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name, 2740 ch->qp->qp_num); 2741 break; 2742 case IB_CM_DREQ_ERROR: 2743 pr_info("Received CM DREQ ERROR event.\n"); 2744 break; 2745 case IB_CM_MRA_RECEIVED: 2746 pr_info("Received CM MRA event\n"); 2747 break; 2748 default: 2749 pr_err("received unrecognized CM event %d\n", event->event); 2750 break; 2751 } 2752 2753 return ret; 2754 } 2755 2756 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id, 2757 struct rdma_cm_event *event) 2758 { 2759 struct srpt_rdma_ch *ch = cm_id->context; 2760 int ret = 0; 2761 2762 switch (event->event) { 2763 case RDMA_CM_EVENT_CONNECT_REQUEST: 2764 ret = srpt_rdma_cm_req_recv(cm_id, event); 2765 break; 2766 case RDMA_CM_EVENT_REJECTED: 2767 srpt_cm_rej_recv(ch, event->status, 2768 event->param.conn.private_data, 2769 event->param.conn.private_data_len); 2770 break; 2771 case RDMA_CM_EVENT_ESTABLISHED: 2772 srpt_cm_rtu_recv(ch); 2773 break; 2774 case RDMA_CM_EVENT_DISCONNECTED: 2775 if (ch->state < CH_DISCONNECTING) 2776 srpt_disconnect_ch(ch); 2777 else 2778 srpt_close_ch(ch); 2779 break; 2780 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2781 srpt_close_ch(ch); 2782 break; 2783 case RDMA_CM_EVENT_UNREACHABLE: 2784 pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name, 2785 ch->qp->qp_num); 2786 break; 2787 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2788 case RDMA_CM_EVENT_ADDR_CHANGE: 2789 break; 2790 default: 2791 pr_err("received unrecognized RDMA CM event %d\n", 2792 event->event); 2793 break; 2794 } 2795 2796 return ret; 2797 } 2798 2799 /* 2800 * srpt_write_pending - Start data transfer from initiator to target (write). 2801 */ 2802 static int srpt_write_pending(struct se_cmd *se_cmd) 2803 { 2804 struct srpt_send_ioctx *ioctx = 2805 container_of(se_cmd, struct srpt_send_ioctx, cmd); 2806 struct srpt_rdma_ch *ch = ioctx->ch; 2807 struct ib_send_wr *first_wr = NULL; 2808 struct ib_cqe *cqe = &ioctx->rdma_cqe; 2809 enum srpt_command_state new_state; 2810 int ret, i; 2811 2812 if (ioctx->recv_ioctx) { 2813 srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN); 2814 target_execute_cmd(&ioctx->cmd); 2815 return 0; 2816 } 2817 2818 new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA); 2819 WARN_ON(new_state == SRPT_STATE_DONE); 2820 2821 if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) { 2822 pr_warn("%s: IB send queue full (needed %d)\n", 2823 __func__, ioctx->n_rdma); 2824 ret = -ENOMEM; 2825 goto out_undo; 2826 } 2827 2828 cqe->done = srpt_rdma_read_done; 2829 for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) { 2830 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 2831 2832 first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port, 2833 cqe, first_wr); 2834 cqe = NULL; 2835 } 2836 2837 ret = ib_post_send(ch->qp, first_wr, NULL); 2838 if (ret) { 2839 pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n", 2840 __func__, ret, ioctx->n_rdma, 2841 atomic_read(&ch->sq_wr_avail)); 2842 goto out_undo; 2843 } 2844 2845 return 0; 2846 out_undo: 2847 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail); 2848 return ret; 2849 } 2850 2851 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status) 2852 { 2853 switch (tcm_mgmt_status) { 2854 case TMR_FUNCTION_COMPLETE: 2855 return SRP_TSK_MGMT_SUCCESS; 2856 case TMR_FUNCTION_REJECTED: 2857 return SRP_TSK_MGMT_FUNC_NOT_SUPP; 2858 } 2859 return SRP_TSK_MGMT_FAILED; 2860 } 2861 2862 /** 2863 * srpt_queue_response - transmit the response to a SCSI command 2864 * @cmd: SCSI target command. 2865 * 2866 * Callback function called by the TCM core. Must not block since it can be 2867 * invoked on the context of the IB completion handler. 2868 */ 2869 static void srpt_queue_response(struct se_cmd *cmd) 2870 { 2871 struct srpt_send_ioctx *ioctx = 2872 container_of(cmd, struct srpt_send_ioctx, cmd); 2873 struct srpt_rdma_ch *ch = ioctx->ch; 2874 struct srpt_device *sdev = ch->sport->sdev; 2875 struct ib_send_wr send_wr, *first_wr = &send_wr; 2876 struct ib_sge sge; 2877 enum srpt_command_state state; 2878 int resp_len, ret, i; 2879 u8 srp_tm_status; 2880 2881 state = ioctx->state; 2882 switch (state) { 2883 case SRPT_STATE_NEW: 2884 case SRPT_STATE_DATA_IN: 2885 ioctx->state = SRPT_STATE_CMD_RSP_SENT; 2886 break; 2887 case SRPT_STATE_MGMT: 2888 ioctx->state = SRPT_STATE_MGMT_RSP_SENT; 2889 break; 2890 default: 2891 WARN(true, "ch %p; cmd %d: unexpected command state %d\n", 2892 ch, ioctx->ioctx.index, ioctx->state); 2893 break; 2894 } 2895 2896 if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT)) 2897 return; 2898 2899 /* For read commands, transfer the data to the initiator. */ 2900 if (ioctx->cmd.data_direction == DMA_FROM_DEVICE && 2901 ioctx->cmd.data_length && 2902 !ioctx->queue_status_only) { 2903 for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) { 2904 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 2905 2906 first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, 2907 ch->sport->port, NULL, first_wr); 2908 } 2909 } 2910 2911 if (state != SRPT_STATE_MGMT) 2912 resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag, 2913 cmd->scsi_status); 2914 else { 2915 srp_tm_status 2916 = tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response); 2917 resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status, 2918 ioctx->cmd.tag); 2919 } 2920 2921 atomic_inc(&ch->req_lim); 2922 2923 if (unlikely(atomic_sub_return(1 + ioctx->n_rdma, 2924 &ch->sq_wr_avail) < 0)) { 2925 pr_warn("%s: IB send queue full (needed %d)\n", 2926 __func__, ioctx->n_rdma); 2927 goto out; 2928 } 2929 2930 ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len, 2931 DMA_TO_DEVICE); 2932 2933 sge.addr = ioctx->ioctx.dma; 2934 sge.length = resp_len; 2935 sge.lkey = sdev->lkey; 2936 2937 ioctx->ioctx.cqe.done = srpt_send_done; 2938 send_wr.next = NULL; 2939 send_wr.wr_cqe = &ioctx->ioctx.cqe; 2940 send_wr.sg_list = &sge; 2941 send_wr.num_sge = 1; 2942 send_wr.opcode = IB_WR_SEND; 2943 send_wr.send_flags = IB_SEND_SIGNALED; 2944 2945 ret = ib_post_send(ch->qp, first_wr, NULL); 2946 if (ret < 0) { 2947 pr_err("%s: sending cmd response failed for tag %llu (%d)\n", 2948 __func__, ioctx->cmd.tag, ret); 2949 goto out; 2950 } 2951 2952 return; 2953 2954 out: 2955 atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail); 2956 atomic_dec(&ch->req_lim); 2957 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE); 2958 target_put_sess_cmd(&ioctx->cmd); 2959 } 2960 2961 static int srpt_queue_data_in(struct se_cmd *cmd) 2962 { 2963 srpt_queue_response(cmd); 2964 return 0; 2965 } 2966 2967 static void srpt_queue_tm_rsp(struct se_cmd *cmd) 2968 { 2969 srpt_queue_response(cmd); 2970 } 2971 2972 /* 2973 * This function is called for aborted commands if no response is sent to the 2974 * initiator. Make sure that the credits freed by aborting a command are 2975 * returned to the initiator the next time a response is sent by incrementing 2976 * ch->req_lim_delta. 2977 */ 2978 static void srpt_aborted_task(struct se_cmd *cmd) 2979 { 2980 struct srpt_send_ioctx *ioctx = container_of(cmd, 2981 struct srpt_send_ioctx, cmd); 2982 struct srpt_rdma_ch *ch = ioctx->ch; 2983 2984 atomic_inc(&ch->req_lim_delta); 2985 } 2986 2987 static int srpt_queue_status(struct se_cmd *cmd) 2988 { 2989 struct srpt_send_ioctx *ioctx; 2990 2991 ioctx = container_of(cmd, struct srpt_send_ioctx, cmd); 2992 BUG_ON(ioctx->sense_data != cmd->sense_buffer); 2993 if (cmd->se_cmd_flags & 2994 (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE)) 2995 WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION); 2996 ioctx->queue_status_only = true; 2997 srpt_queue_response(cmd); 2998 return 0; 2999 } 3000 3001 static void srpt_refresh_port_work(struct work_struct *work) 3002 { 3003 struct srpt_port *sport = container_of(work, struct srpt_port, work); 3004 3005 srpt_refresh_port(sport); 3006 } 3007 3008 /** 3009 * srpt_release_sport - disable login and wait for associated channels 3010 * @sport: SRPT HCA port. 3011 */ 3012 static int srpt_release_sport(struct srpt_port *sport) 3013 { 3014 DECLARE_COMPLETION_ONSTACK(c); 3015 struct srpt_nexus *nexus, *next_n; 3016 struct srpt_rdma_ch *ch; 3017 3018 WARN_ON_ONCE(irqs_disabled()); 3019 3020 sport->freed_channels = &c; 3021 3022 mutex_lock(&sport->mutex); 3023 srpt_set_enabled(sport, false); 3024 mutex_unlock(&sport->mutex); 3025 3026 while (atomic_read(&sport->refcount) > 0 && 3027 wait_for_completion_timeout(&c, 5 * HZ) <= 0) { 3028 pr_info("%s_%d: waiting for unregistration of %d sessions ...\n", 3029 dev_name(&sport->sdev->device->dev), sport->port, 3030 atomic_read(&sport->refcount)); 3031 rcu_read_lock(); 3032 list_for_each_entry(nexus, &sport->nexus_list, entry) { 3033 list_for_each_entry(ch, &nexus->ch_list, list) { 3034 pr_info("%s-%d: state %s\n", 3035 ch->sess_name, ch->qp->qp_num, 3036 get_ch_state_name(ch->state)); 3037 } 3038 } 3039 rcu_read_unlock(); 3040 } 3041 3042 mutex_lock(&sport->mutex); 3043 list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) { 3044 list_del(&nexus->entry); 3045 kfree_rcu(nexus, rcu); 3046 } 3047 mutex_unlock(&sport->mutex); 3048 3049 return 0; 3050 } 3051 3052 struct port_and_port_id { 3053 struct srpt_port *sport; 3054 struct srpt_port_id **port_id; 3055 }; 3056 3057 static struct port_and_port_id __srpt_lookup_port(const char *name) 3058 { 3059 struct ib_device *dev; 3060 struct srpt_device *sdev; 3061 struct srpt_port *sport; 3062 int i; 3063 3064 list_for_each_entry(sdev, &srpt_dev_list, list) { 3065 dev = sdev->device; 3066 if (!dev) 3067 continue; 3068 3069 for (i = 0; i < dev->phys_port_cnt; i++) { 3070 sport = &sdev->port[i]; 3071 3072 if (strcmp(sport->guid_name, name) == 0) { 3073 kref_get(&sdev->refcnt); 3074 return (struct port_and_port_id){ 3075 sport, &sport->guid_id}; 3076 } 3077 if (strcmp(sport->gid_name, name) == 0) { 3078 kref_get(&sdev->refcnt); 3079 return (struct port_and_port_id){ 3080 sport, &sport->gid_id}; 3081 } 3082 } 3083 } 3084 3085 return (struct port_and_port_id){}; 3086 } 3087 3088 /** 3089 * srpt_lookup_port() - Look up an RDMA port by name 3090 * @name: ASCII port name 3091 * 3092 * Increments the RDMA port reference count if an RDMA port pointer is returned. 3093 * The caller must drop that reference count by calling srpt_port_put_ref(). 3094 */ 3095 static struct port_and_port_id srpt_lookup_port(const char *name) 3096 { 3097 struct port_and_port_id papi; 3098 3099 spin_lock(&srpt_dev_lock); 3100 papi = __srpt_lookup_port(name); 3101 spin_unlock(&srpt_dev_lock); 3102 3103 return papi; 3104 } 3105 3106 static void srpt_free_srq(struct srpt_device *sdev) 3107 { 3108 if (!sdev->srq) 3109 return; 3110 3111 ib_destroy_srq(sdev->srq); 3112 srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev, 3113 sdev->srq_size, sdev->req_buf_cache, 3114 DMA_FROM_DEVICE); 3115 srpt_cache_put(sdev->req_buf_cache); 3116 sdev->srq = NULL; 3117 } 3118 3119 static int srpt_alloc_srq(struct srpt_device *sdev) 3120 { 3121 struct ib_srq_init_attr srq_attr = { 3122 .event_handler = srpt_srq_event, 3123 .srq_context = (void *)sdev, 3124 .attr.max_wr = sdev->srq_size, 3125 .attr.max_sge = 1, 3126 .srq_type = IB_SRQT_BASIC, 3127 }; 3128 struct ib_device *device = sdev->device; 3129 struct ib_srq *srq; 3130 int i; 3131 3132 WARN_ON_ONCE(sdev->srq); 3133 srq = ib_create_srq(sdev->pd, &srq_attr); 3134 if (IS_ERR(srq)) { 3135 pr_debug("ib_create_srq() failed: %pe\n", srq); 3136 return PTR_ERR(srq); 3137 } 3138 3139 pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size, 3140 sdev->device->attrs.max_srq_wr, dev_name(&device->dev)); 3141 3142 sdev->req_buf_cache = srpt_cache_get(srp_max_req_size); 3143 if (!sdev->req_buf_cache) 3144 goto free_srq; 3145 3146 sdev->ioctx_ring = (struct srpt_recv_ioctx **) 3147 srpt_alloc_ioctx_ring(sdev, sdev->srq_size, 3148 sizeof(*sdev->ioctx_ring[0]), 3149 sdev->req_buf_cache, 0, DMA_FROM_DEVICE); 3150 if (!sdev->ioctx_ring) 3151 goto free_cache; 3152 3153 sdev->use_srq = true; 3154 sdev->srq = srq; 3155 3156 for (i = 0; i < sdev->srq_size; ++i) { 3157 INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list); 3158 srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]); 3159 } 3160 3161 return 0; 3162 3163 free_cache: 3164 srpt_cache_put(sdev->req_buf_cache); 3165 3166 free_srq: 3167 ib_destroy_srq(srq); 3168 return -ENOMEM; 3169 } 3170 3171 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq) 3172 { 3173 struct ib_device *device = sdev->device; 3174 int ret = 0; 3175 3176 if (!use_srq) { 3177 srpt_free_srq(sdev); 3178 sdev->use_srq = false; 3179 } else if (use_srq && !sdev->srq) { 3180 ret = srpt_alloc_srq(sdev); 3181 } 3182 pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__, 3183 dev_name(&device->dev), sdev->use_srq, ret); 3184 return ret; 3185 } 3186 3187 static void srpt_free_sdev(struct kref *refcnt) 3188 { 3189 struct srpt_device *sdev = container_of(refcnt, typeof(*sdev), refcnt); 3190 3191 kfree(sdev); 3192 } 3193 3194 static void srpt_sdev_put(struct srpt_device *sdev) 3195 { 3196 kref_put(&sdev->refcnt, srpt_free_sdev); 3197 } 3198 3199 /** 3200 * srpt_add_one - InfiniBand device addition callback function 3201 * @device: Describes a HCA. 3202 */ 3203 static int srpt_add_one(struct ib_device *device) 3204 { 3205 struct srpt_device *sdev; 3206 struct srpt_port *sport; 3207 int ret; 3208 u32 i; 3209 3210 pr_debug("device = %p\n", device); 3211 3212 sdev = kzalloc_flex(*sdev, port, device->phys_port_cnt); 3213 if (!sdev) 3214 return -ENOMEM; 3215 3216 kref_init(&sdev->refcnt); 3217 sdev->device = device; 3218 mutex_init(&sdev->sdev_mutex); 3219 3220 sdev->pd = ib_alloc_pd(device, 0); 3221 if (IS_ERR(sdev->pd)) { 3222 ret = PTR_ERR(sdev->pd); 3223 goto free_dev; 3224 } 3225 3226 sdev->lkey = sdev->pd->local_dma_lkey; 3227 3228 sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr); 3229 3230 srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq); 3231 3232 if (!srpt_service_guid) 3233 srpt_service_guid = be64_to_cpu(device->node_guid); 3234 3235 if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND) 3236 sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev); 3237 if (IS_ERR(sdev->cm_id)) { 3238 pr_info("ib_create_cm_id() failed: %pe\n", sdev->cm_id); 3239 ret = PTR_ERR(sdev->cm_id); 3240 sdev->cm_id = NULL; 3241 if (!rdma_cm_id) 3242 goto err_ring; 3243 } 3244 3245 /* print out target login information */ 3246 pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n", 3247 srpt_service_guid, srpt_service_guid, srpt_service_guid); 3248 3249 /* 3250 * We do not have a consistent service_id (ie. also id_ext of target_id) 3251 * to identify this target. We currently use the guid of the first HCA 3252 * in the system as service_id; therefore, the target_id will change 3253 * if this HCA is gone bad and replaced by different HCA 3254 */ 3255 ret = sdev->cm_id ? 3256 ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid)) : 3257 0; 3258 if (ret < 0) { 3259 pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret, 3260 sdev->cm_id->state); 3261 goto err_cm; 3262 } 3263 3264 INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device, 3265 srpt_event_handler); 3266 3267 for (i = 1; i <= sdev->device->phys_port_cnt; i++) { 3268 sport = &sdev->port[i - 1]; 3269 INIT_LIST_HEAD(&sport->nexus_list); 3270 mutex_init(&sport->mutex); 3271 sport->sdev = sdev; 3272 sport->port = i; 3273 sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE; 3274 sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE; 3275 sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE; 3276 sport->port_attrib.use_srq = false; 3277 INIT_WORK(&sport->work, srpt_refresh_port_work); 3278 3279 ret = srpt_refresh_port(sport); 3280 if (ret) { 3281 pr_err("MAD registration failed for %s-%d.\n", 3282 dev_name(&sdev->device->dev), i); 3283 i--; 3284 goto err_port; 3285 } 3286 } 3287 3288 ib_register_event_handler(&sdev->event_handler); 3289 spin_lock(&srpt_dev_lock); 3290 list_add_tail(&sdev->list, &srpt_dev_list); 3291 spin_unlock(&srpt_dev_lock); 3292 3293 ib_set_client_data(device, &srpt_client, sdev); 3294 pr_debug("added %s.\n", dev_name(&device->dev)); 3295 return 0; 3296 3297 err_port: 3298 srpt_unregister_mad_agent(sdev, i); 3299 err_cm: 3300 if (sdev->cm_id) 3301 ib_destroy_cm_id(sdev->cm_id); 3302 err_ring: 3303 srpt_free_srq(sdev); 3304 ib_dealloc_pd(sdev->pd); 3305 free_dev: 3306 srpt_sdev_put(sdev); 3307 pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev)); 3308 return ret; 3309 } 3310 3311 /** 3312 * srpt_remove_one - InfiniBand device removal callback function 3313 * @device: Describes a HCA. 3314 * @client_data: The value passed as the third argument to ib_set_client_data(). 3315 */ 3316 static void srpt_remove_one(struct ib_device *device, void *client_data) 3317 { 3318 struct srpt_device *sdev = client_data; 3319 int i; 3320 3321 srpt_unregister_mad_agent(sdev, sdev->device->phys_port_cnt); 3322 3323 ib_unregister_event_handler(&sdev->event_handler); 3324 3325 /* Cancel any work queued by the just unregistered IB event handler. */ 3326 for (i = 0; i < sdev->device->phys_port_cnt; i++) 3327 cancel_work_sync(&sdev->port[i].work); 3328 3329 if (sdev->cm_id) 3330 ib_destroy_cm_id(sdev->cm_id); 3331 3332 ib_set_client_data(device, &srpt_client, NULL); 3333 3334 /* 3335 * Unregistering a target must happen after destroying sdev->cm_id 3336 * such that no new SRP_LOGIN_REQ information units can arrive while 3337 * destroying the target. 3338 */ 3339 spin_lock(&srpt_dev_lock); 3340 list_del(&sdev->list); 3341 spin_unlock(&srpt_dev_lock); 3342 3343 for (i = 0; i < sdev->device->phys_port_cnt; i++) 3344 srpt_release_sport(&sdev->port[i]); 3345 3346 srpt_free_srq(sdev); 3347 3348 ib_dealloc_pd(sdev->pd); 3349 3350 srpt_sdev_put(sdev); 3351 } 3352 3353 static struct ib_client srpt_client = { 3354 .name = DRV_NAME, 3355 .add = srpt_add_one, 3356 .remove = srpt_remove_one 3357 }; 3358 3359 static int srpt_check_true(struct se_portal_group *se_tpg) 3360 { 3361 return 1; 3362 } 3363 3364 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg) 3365 { 3366 return tpg->se_tpg_wwn->priv; 3367 } 3368 3369 static struct srpt_port_id *srpt_wwn_to_sport_id(struct se_wwn *wwn) 3370 { 3371 struct srpt_port *sport = wwn->priv; 3372 3373 if (sport->guid_id && &sport->guid_id->wwn == wwn) 3374 return sport->guid_id; 3375 if (sport->gid_id && &sport->gid_id->wwn == wwn) 3376 return sport->gid_id; 3377 WARN_ON_ONCE(true); 3378 return NULL; 3379 } 3380 3381 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg) 3382 { 3383 struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg); 3384 3385 return stpg->sport_id->name; 3386 } 3387 3388 static u16 srpt_get_tag(struct se_portal_group *tpg) 3389 { 3390 return 1; 3391 } 3392 3393 static void srpt_release_cmd(struct se_cmd *se_cmd) 3394 { 3395 struct srpt_send_ioctx *ioctx = container_of(se_cmd, 3396 struct srpt_send_ioctx, cmd); 3397 struct srpt_rdma_ch *ch = ioctx->ch; 3398 struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx; 3399 3400 WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE && 3401 !(ioctx->cmd.transport_state & CMD_T_ABORTED)); 3402 3403 if (recv_ioctx) { 3404 WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list)); 3405 ioctx->recv_ioctx = NULL; 3406 srpt_post_recv(ch->sport->sdev, ch, recv_ioctx); 3407 } 3408 3409 if (ioctx->n_rw_ctx) { 3410 srpt_free_rw_ctxs(ch, ioctx); 3411 ioctx->n_rw_ctx = 0; 3412 } 3413 3414 target_free_tag(se_cmd->se_sess, se_cmd); 3415 } 3416 3417 /** 3418 * srpt_close_session - forcibly close a session 3419 * @se_sess: SCSI target session. 3420 * 3421 * Callback function invoked by the TCM core to clean up sessions associated 3422 * with a node ACL when the user invokes 3423 * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id 3424 */ 3425 static void srpt_close_session(struct se_session *se_sess) 3426 { 3427 struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr; 3428 3429 srpt_disconnect_ch_sync(ch); 3430 } 3431 3432 /* Note: only used from inside debug printk's by the TCM core. */ 3433 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd) 3434 { 3435 struct srpt_send_ioctx *ioctx; 3436 3437 ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd); 3438 return ioctx->state; 3439 } 3440 3441 static int srpt_parse_guid(u64 *guid, const char *name) 3442 { 3443 u16 w[4]; 3444 int ret = -EINVAL; 3445 3446 if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4) 3447 goto out; 3448 *guid = get_unaligned_be64(w); 3449 ret = 0; 3450 out: 3451 return ret; 3452 } 3453 3454 /** 3455 * srpt_parse_i_port_id - parse an initiator port ID 3456 * @name: ASCII representation of a 128-bit initiator port ID. 3457 * @i_port_id: Binary 128-bit port ID. 3458 */ 3459 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name) 3460 { 3461 const char *p; 3462 unsigned len, count, leading_zero_bytes; 3463 int ret; 3464 3465 p = name; 3466 if (strncasecmp(p, "0x", 2) == 0) 3467 p += 2; 3468 ret = -EINVAL; 3469 len = strlen(p); 3470 if (len % 2) 3471 goto out; 3472 count = min(len / 2, 16U); 3473 leading_zero_bytes = 16 - count; 3474 memset(i_port_id, 0, leading_zero_bytes); 3475 ret = hex2bin(i_port_id + leading_zero_bytes, p, count); 3476 3477 out: 3478 return ret; 3479 } 3480 3481 /* 3482 * configfs callback function invoked for mkdir 3483 * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id 3484 * 3485 * i_port_id must be an initiator port GUID, GID or IP address. See also the 3486 * target_alloc_session() calls in this driver. Examples of valid initiator 3487 * port IDs: 3488 * 0x0000000000000000505400fffe4a0b7b 3489 * 0000000000000000505400fffe4a0b7b 3490 * 5054:00ff:fe4a:0b7b 3491 * 192.168.122.76 3492 */ 3493 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name) 3494 { 3495 struct sockaddr_storage sa; 3496 u64 guid; 3497 u8 i_port_id[16]; 3498 int ret; 3499 3500 ret = srpt_parse_guid(&guid, name); 3501 if (ret < 0) 3502 ret = srpt_parse_i_port_id(i_port_id, name); 3503 if (ret < 0) 3504 ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL, 3505 &sa); 3506 if (ret < 0) 3507 pr_err("invalid initiator port ID %s\n", name); 3508 return ret; 3509 } 3510 3511 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item, 3512 char *page) 3513 { 3514 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3515 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3516 3517 return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rdma_size); 3518 } 3519 3520 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item, 3521 const char *page, size_t count) 3522 { 3523 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3524 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3525 unsigned long val; 3526 int ret; 3527 3528 ret = kstrtoul(page, 0, &val); 3529 if (ret < 0) { 3530 pr_err("kstrtoul() failed with ret: %d\n", ret); 3531 return -EINVAL; 3532 } 3533 if (val > MAX_SRPT_RDMA_SIZE) { 3534 pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val, 3535 MAX_SRPT_RDMA_SIZE); 3536 return -EINVAL; 3537 } 3538 if (val < DEFAULT_MAX_RDMA_SIZE) { 3539 pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n", 3540 val, DEFAULT_MAX_RDMA_SIZE); 3541 return -EINVAL; 3542 } 3543 sport->port_attrib.srp_max_rdma_size = val; 3544 3545 return count; 3546 } 3547 3548 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item, 3549 char *page) 3550 { 3551 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3552 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3553 3554 return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rsp_size); 3555 } 3556 3557 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item, 3558 const char *page, size_t count) 3559 { 3560 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3561 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3562 unsigned long val; 3563 int ret; 3564 3565 ret = kstrtoul(page, 0, &val); 3566 if (ret < 0) { 3567 pr_err("kstrtoul() failed with ret: %d\n", ret); 3568 return -EINVAL; 3569 } 3570 if (val > MAX_SRPT_RSP_SIZE) { 3571 pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val, 3572 MAX_SRPT_RSP_SIZE); 3573 return -EINVAL; 3574 } 3575 if (val < MIN_MAX_RSP_SIZE) { 3576 pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val, 3577 MIN_MAX_RSP_SIZE); 3578 return -EINVAL; 3579 } 3580 sport->port_attrib.srp_max_rsp_size = val; 3581 3582 return count; 3583 } 3584 3585 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item, 3586 char *page) 3587 { 3588 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3589 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3590 3591 return sysfs_emit(page, "%u\n", sport->port_attrib.srp_sq_size); 3592 } 3593 3594 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item, 3595 const char *page, size_t count) 3596 { 3597 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3598 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3599 unsigned long val; 3600 int ret; 3601 3602 ret = kstrtoul(page, 0, &val); 3603 if (ret < 0) { 3604 pr_err("kstrtoul() failed with ret: %d\n", ret); 3605 return -EINVAL; 3606 } 3607 if (val > MAX_SRPT_SRQ_SIZE) { 3608 pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val, 3609 MAX_SRPT_SRQ_SIZE); 3610 return -EINVAL; 3611 } 3612 if (val < MIN_SRPT_SRQ_SIZE) { 3613 pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val, 3614 MIN_SRPT_SRQ_SIZE); 3615 return -EINVAL; 3616 } 3617 sport->port_attrib.srp_sq_size = val; 3618 3619 return count; 3620 } 3621 3622 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item, 3623 char *page) 3624 { 3625 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3626 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3627 3628 return sysfs_emit(page, "%d\n", sport->port_attrib.use_srq); 3629 } 3630 3631 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item, 3632 const char *page, size_t count) 3633 { 3634 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3635 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3636 struct srpt_device *sdev = sport->sdev; 3637 unsigned long val; 3638 bool enabled; 3639 int ret; 3640 3641 ret = kstrtoul(page, 0, &val); 3642 if (ret < 0) 3643 return ret; 3644 if (val != !!val) 3645 return -EINVAL; 3646 3647 ret = mutex_lock_interruptible(&sdev->sdev_mutex); 3648 if (ret < 0) 3649 return ret; 3650 ret = mutex_lock_interruptible(&sport->mutex); 3651 if (ret < 0) 3652 goto unlock_sdev; 3653 enabled = sport->enabled; 3654 /* Log out all initiator systems before changing 'use_srq'. */ 3655 srpt_set_enabled(sport, false); 3656 sport->port_attrib.use_srq = val; 3657 srpt_use_srq(sdev, sport->port_attrib.use_srq); 3658 srpt_set_enabled(sport, enabled); 3659 ret = count; 3660 mutex_unlock(&sport->mutex); 3661 unlock_sdev: 3662 mutex_unlock(&sdev->sdev_mutex); 3663 3664 return ret; 3665 } 3666 3667 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_max_rdma_size); 3668 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_max_rsp_size); 3669 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_sq_size); 3670 CONFIGFS_ATTR(srpt_tpg_attrib_, use_srq); 3671 3672 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = { 3673 &srpt_tpg_attrib_attr_srp_max_rdma_size, 3674 &srpt_tpg_attrib_attr_srp_max_rsp_size, 3675 &srpt_tpg_attrib_attr_srp_sq_size, 3676 &srpt_tpg_attrib_attr_use_srq, 3677 NULL, 3678 }; 3679 3680 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr) 3681 { 3682 struct rdma_cm_id *rdma_cm_id; 3683 int ret; 3684 3685 rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler, 3686 NULL, RDMA_PS_TCP, IB_QPT_RC); 3687 if (IS_ERR(rdma_cm_id)) { 3688 pr_err("RDMA/CM ID creation failed: %pe\n", rdma_cm_id); 3689 goto out; 3690 } 3691 3692 ret = rdma_bind_addr(rdma_cm_id, listen_addr); 3693 if (ret) { 3694 char addr_str[64]; 3695 3696 snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr); 3697 pr_err("Binding RDMA/CM ID to address %s failed: %d\n", 3698 addr_str, ret); 3699 rdma_destroy_id(rdma_cm_id); 3700 rdma_cm_id = ERR_PTR(ret); 3701 goto out; 3702 } 3703 3704 ret = rdma_listen(rdma_cm_id, 128); 3705 if (ret) { 3706 pr_err("rdma_listen() failed: %d\n", ret); 3707 rdma_destroy_id(rdma_cm_id); 3708 rdma_cm_id = ERR_PTR(ret); 3709 } 3710 3711 out: 3712 return rdma_cm_id; 3713 } 3714 3715 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page) 3716 { 3717 return sysfs_emit(page, "%d\n", rdma_cm_port); 3718 } 3719 3720 static ssize_t srpt_rdma_cm_port_store(struct config_item *item, 3721 const char *page, size_t count) 3722 { 3723 struct sockaddr_in addr4 = { .sin_family = AF_INET }; 3724 struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 }; 3725 struct rdma_cm_id *new_id = NULL; 3726 u16 val; 3727 int ret; 3728 3729 ret = kstrtou16(page, 0, &val); 3730 if (ret < 0) 3731 return ret; 3732 ret = count; 3733 if (rdma_cm_port == val) 3734 goto out; 3735 3736 if (val) { 3737 addr6.sin6_port = cpu_to_be16(val); 3738 new_id = srpt_create_rdma_id((struct sockaddr *)&addr6); 3739 if (IS_ERR(new_id)) { 3740 addr4.sin_port = cpu_to_be16(val); 3741 new_id = srpt_create_rdma_id((struct sockaddr *)&addr4); 3742 if (IS_ERR(new_id)) { 3743 ret = PTR_ERR(new_id); 3744 goto out; 3745 } 3746 } 3747 } 3748 3749 mutex_lock(&rdma_cm_mutex); 3750 rdma_cm_port = val; 3751 swap(rdma_cm_id, new_id); 3752 mutex_unlock(&rdma_cm_mutex); 3753 3754 if (new_id) 3755 rdma_destroy_id(new_id); 3756 ret = count; 3757 out: 3758 return ret; 3759 } 3760 3761 CONFIGFS_ATTR(srpt_, rdma_cm_port); 3762 3763 static struct configfs_attribute *srpt_da_attrs[] = { 3764 &srpt_attr_rdma_cm_port, 3765 NULL, 3766 }; 3767 3768 static int srpt_enable_tpg(struct se_portal_group *se_tpg, bool enable) 3769 { 3770 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3771 3772 mutex_lock(&sport->mutex); 3773 srpt_set_enabled(sport, enable); 3774 mutex_unlock(&sport->mutex); 3775 3776 return 0; 3777 } 3778 3779 /** 3780 * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg 3781 * @wwn: Corresponds to $driver/$port. 3782 * @name: $tpg. 3783 */ 3784 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn, 3785 const char *name) 3786 { 3787 struct srpt_port_id *sport_id = srpt_wwn_to_sport_id(wwn); 3788 struct srpt_tpg *stpg; 3789 int res = -ENOMEM; 3790 3791 stpg = kzalloc_obj(*stpg); 3792 if (!stpg) 3793 return ERR_PTR(res); 3794 stpg->sport_id = sport_id; 3795 res = core_tpg_register(wwn, &stpg->tpg, SCSI_PROTOCOL_SRP); 3796 if (res) { 3797 kfree(stpg); 3798 return ERR_PTR(res); 3799 } 3800 3801 mutex_lock(&sport_id->mutex); 3802 list_add_tail(&stpg->entry, &sport_id->tpg_list); 3803 mutex_unlock(&sport_id->mutex); 3804 3805 return &stpg->tpg; 3806 } 3807 3808 /** 3809 * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg 3810 * @tpg: Target portal group to deregister. 3811 */ 3812 static void srpt_drop_tpg(struct se_portal_group *tpg) 3813 { 3814 struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg); 3815 struct srpt_port_id *sport_id = stpg->sport_id; 3816 struct srpt_port *sport = srpt_tpg_to_sport(tpg); 3817 3818 mutex_lock(&sport_id->mutex); 3819 list_del(&stpg->entry); 3820 mutex_unlock(&sport_id->mutex); 3821 3822 sport->enabled = false; 3823 core_tpg_deregister(tpg); 3824 kfree(stpg); 3825 } 3826 3827 /** 3828 * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port 3829 * @tf: Not used. 3830 * @group: Not used. 3831 * @name: $port. 3832 */ 3833 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf, 3834 struct config_group *group, 3835 const char *name) 3836 { 3837 struct port_and_port_id papi = srpt_lookup_port(name); 3838 struct srpt_port *sport = papi.sport; 3839 struct srpt_port_id *port_id; 3840 3841 if (!papi.port_id) 3842 return ERR_PTR(-EINVAL); 3843 if (*papi.port_id) { 3844 /* Attempt to create a directory that already exists. */ 3845 WARN_ON_ONCE(true); 3846 return &(*papi.port_id)->wwn; 3847 } 3848 port_id = kzalloc_obj(*port_id); 3849 if (!port_id) { 3850 srpt_sdev_put(sport->sdev); 3851 return ERR_PTR(-ENOMEM); 3852 } 3853 mutex_init(&port_id->mutex); 3854 INIT_LIST_HEAD(&port_id->tpg_list); 3855 port_id->wwn.priv = sport; 3856 memcpy(port_id->name, port_id == sport->guid_id ? sport->guid_name : 3857 sport->gid_name, ARRAY_SIZE(port_id->name)); 3858 3859 *papi.port_id = port_id; 3860 3861 return &port_id->wwn; 3862 } 3863 3864 /** 3865 * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port 3866 * @wwn: $port. 3867 */ 3868 static void srpt_drop_tport(struct se_wwn *wwn) 3869 { 3870 struct srpt_port_id *port_id = container_of(wwn, typeof(*port_id), wwn); 3871 struct srpt_port *sport = wwn->priv; 3872 3873 if (sport->guid_id == port_id) 3874 sport->guid_id = NULL; 3875 else if (sport->gid_id == port_id) 3876 sport->gid_id = NULL; 3877 else 3878 WARN_ON_ONCE(true); 3879 3880 srpt_sdev_put(sport->sdev); 3881 kfree(port_id); 3882 } 3883 3884 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf) 3885 { 3886 return sysfs_emit(buf, "\n"); 3887 } 3888 3889 CONFIGFS_ATTR_RO(srpt_wwn_, version); 3890 3891 static struct configfs_attribute *srpt_wwn_attrs[] = { 3892 &srpt_wwn_attr_version, 3893 NULL, 3894 }; 3895 3896 static const struct target_core_fabric_ops srpt_template = { 3897 .module = THIS_MODULE, 3898 .fabric_name = "srpt", 3899 .tpg_get_wwn = srpt_get_fabric_wwn, 3900 .tpg_get_tag = srpt_get_tag, 3901 .tpg_check_demo_mode_cache = srpt_check_true, 3902 .tpg_check_demo_mode_write_protect = srpt_check_true, 3903 .release_cmd = srpt_release_cmd, 3904 .check_stop_free = srpt_check_stop_free, 3905 .close_session = srpt_close_session, 3906 .sess_get_initiator_sid = NULL, 3907 .write_pending = srpt_write_pending, 3908 .get_cmd_state = srpt_get_tcm_cmd_state, 3909 .queue_data_in = srpt_queue_data_in, 3910 .queue_status = srpt_queue_status, 3911 .queue_tm_rsp = srpt_queue_tm_rsp, 3912 .aborted_task = srpt_aborted_task, 3913 /* 3914 * Setup function pointers for generic logic in 3915 * target_core_fabric_configfs.c 3916 */ 3917 .fabric_make_wwn = srpt_make_tport, 3918 .fabric_drop_wwn = srpt_drop_tport, 3919 .fabric_make_tpg = srpt_make_tpg, 3920 .fabric_enable_tpg = srpt_enable_tpg, 3921 .fabric_drop_tpg = srpt_drop_tpg, 3922 .fabric_init_nodeacl = srpt_init_nodeacl, 3923 3924 .tfc_discovery_attrs = srpt_da_attrs, 3925 .tfc_wwn_attrs = srpt_wwn_attrs, 3926 .tfc_tpg_attrib_attrs = srpt_tpg_attrib_attrs, 3927 3928 .default_submit_type = TARGET_DIRECT_SUBMIT, 3929 .direct_submit_supp = 1, 3930 }; 3931 3932 /** 3933 * srpt_init_module - kernel module initialization 3934 * 3935 * Note: Since ib_register_client() registers callback functions, and since at 3936 * least one of these callback functions (srpt_add_one()) calls target core 3937 * functions, this driver must be registered with the target core before 3938 * ib_register_client() is called. 3939 */ 3940 static int __init srpt_init_module(void) 3941 { 3942 int ret; 3943 3944 ret = -EINVAL; 3945 if (srp_max_req_size < MIN_MAX_REQ_SIZE) { 3946 pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n", 3947 srp_max_req_size, MIN_MAX_REQ_SIZE); 3948 goto out; 3949 } 3950 3951 if (srpt_srq_size < MIN_SRPT_SRQ_SIZE 3952 || srpt_srq_size > MAX_SRPT_SRQ_SIZE) { 3953 pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n", 3954 srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE); 3955 goto out; 3956 } 3957 3958 ret = target_register_template(&srpt_template); 3959 if (ret) 3960 goto out; 3961 3962 ret = ib_register_client(&srpt_client); 3963 if (ret) { 3964 pr_err("couldn't register IB client\n"); 3965 goto out_unregister_target; 3966 } 3967 3968 return 0; 3969 3970 out_unregister_target: 3971 target_unregister_template(&srpt_template); 3972 out: 3973 return ret; 3974 } 3975 3976 static void __exit srpt_cleanup_module(void) 3977 { 3978 if (rdma_cm_id) 3979 rdma_destroy_id(rdma_cm_id); 3980 ib_unregister_client(&srpt_client); 3981 target_unregister_template(&srpt_template); 3982 } 3983 3984 module_init(srpt_init_module); 3985 module_exit(srpt_cleanup_module); 3986