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/init.h> 37 #include <linux/slab.h> 38 #include <linux/err.h> 39 #include <linux/ctype.h> 40 #include <linux/kthread.h> 41 #include <linux/string.h> 42 #include <linux/delay.h> 43 #include <linux/atomic.h> 44 #include <linux/inet.h> 45 #include <rdma/ib_cache.h> 46 #include <scsi/scsi_proto.h> 47 #include <scsi/scsi_tcq.h> 48 #include <target/target_core_base.h> 49 #include <target/target_core_fabric.h> 50 #include "ib_srpt.h" 51 52 /* Name of this kernel module. */ 53 #define DRV_NAME "ib_srpt" 54 55 #define SRPT_ID_STRING "Linux SRP target" 56 57 #undef pr_fmt 58 #define pr_fmt(fmt) DRV_NAME " " fmt 59 60 MODULE_AUTHOR("Vu Pham and Bart Van Assche"); 61 MODULE_DESCRIPTION("SCSI RDMA Protocol target driver"); 62 MODULE_LICENSE("Dual BSD/GPL"); 63 64 /* 65 * Global Variables 66 */ 67 68 static u64 srpt_service_guid; 69 static DEFINE_SPINLOCK(srpt_dev_lock); /* Protects srpt_dev_list. */ 70 static LIST_HEAD(srpt_dev_list); /* List of srpt_device structures. */ 71 72 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE; 73 module_param(srp_max_req_size, int, 0444); 74 MODULE_PARM_DESC(srp_max_req_size, 75 "Maximum size of SRP request messages in bytes."); 76 77 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE; 78 module_param(srpt_srq_size, int, 0444); 79 MODULE_PARM_DESC(srpt_srq_size, 80 "Shared receive queue (SRQ) size."); 81 82 static int srpt_get_u64_x(char *buffer, const struct kernel_param *kp) 83 { 84 return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg); 85 } 86 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid, 87 0444); 88 MODULE_PARM_DESC(srpt_service_guid, 89 "Using this value for ioc_guid, id_ext, and cm_listen_id instead of using the node_guid of the first HCA."); 90 91 static struct ib_client srpt_client; 92 /* Protects both rdma_cm_port and rdma_cm_id. */ 93 static DEFINE_MUTEX(rdma_cm_mutex); 94 /* Port number RDMA/CM will bind to. */ 95 static u16 rdma_cm_port; 96 static struct rdma_cm_id *rdma_cm_id; 97 static void srpt_release_cmd(struct se_cmd *se_cmd); 98 static void srpt_free_ch(struct kref *kref); 99 static int srpt_queue_status(struct se_cmd *cmd); 100 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc); 101 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc); 102 static void srpt_process_wait_list(struct srpt_rdma_ch *ch); 103 104 /* 105 * The only allowed channel state changes are those that change the channel 106 * state into a state with a higher numerical value. Hence the new > prev test. 107 */ 108 static bool srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new) 109 { 110 unsigned long flags; 111 enum rdma_ch_state prev; 112 bool changed = false; 113 114 spin_lock_irqsave(&ch->spinlock, flags); 115 prev = ch->state; 116 if (new > prev) { 117 ch->state = new; 118 changed = true; 119 } 120 spin_unlock_irqrestore(&ch->spinlock, flags); 121 122 return changed; 123 } 124 125 /** 126 * srpt_event_handler - asynchronous IB event callback function 127 * @handler: IB event handler registered by ib_register_event_handler(). 128 * @event: Description of the event that occurred. 129 * 130 * Callback function called by the InfiniBand core when an asynchronous IB 131 * event occurs. This callback may occur in interrupt context. See also 132 * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand 133 * Architecture Specification. 134 */ 135 static void srpt_event_handler(struct ib_event_handler *handler, 136 struct ib_event *event) 137 { 138 struct srpt_device *sdev; 139 struct srpt_port *sport; 140 u8 port_num; 141 142 sdev = ib_get_client_data(event->device, &srpt_client); 143 if (!sdev || sdev->device != event->device) 144 return; 145 146 pr_debug("ASYNC event= %d on device= %s\n", event->event, 147 dev_name(&sdev->device->dev)); 148 149 switch (event->event) { 150 case IB_EVENT_PORT_ERR: 151 port_num = event->element.port_num - 1; 152 if (port_num < sdev->device->phys_port_cnt) { 153 sport = &sdev->port[port_num]; 154 sport->lid = 0; 155 sport->sm_lid = 0; 156 } else { 157 WARN(true, "event %d: port_num %d out of range 1..%d\n", 158 event->event, port_num + 1, 159 sdev->device->phys_port_cnt); 160 } 161 break; 162 case IB_EVENT_PORT_ACTIVE: 163 case IB_EVENT_LID_CHANGE: 164 case IB_EVENT_PKEY_CHANGE: 165 case IB_EVENT_SM_CHANGE: 166 case IB_EVENT_CLIENT_REREGISTER: 167 case IB_EVENT_GID_CHANGE: 168 /* Refresh port data asynchronously. */ 169 port_num = event->element.port_num - 1; 170 if (port_num < sdev->device->phys_port_cnt) { 171 sport = &sdev->port[port_num]; 172 if (!sport->lid && !sport->sm_lid) 173 schedule_work(&sport->work); 174 } else { 175 WARN(true, "event %d: port_num %d out of range 1..%d\n", 176 event->event, port_num + 1, 177 sdev->device->phys_port_cnt); 178 } 179 break; 180 default: 181 pr_err("received unrecognized IB event %d\n", event->event); 182 break; 183 } 184 } 185 186 /** 187 * srpt_srq_event - SRQ event callback function 188 * @event: Description of the event that occurred. 189 * @ctx: Context pointer specified at SRQ creation time. 190 */ 191 static void srpt_srq_event(struct ib_event *event, void *ctx) 192 { 193 pr_debug("SRQ event %d\n", event->event); 194 } 195 196 static const char *get_ch_state_name(enum rdma_ch_state s) 197 { 198 switch (s) { 199 case CH_CONNECTING: 200 return "connecting"; 201 case CH_LIVE: 202 return "live"; 203 case CH_DISCONNECTING: 204 return "disconnecting"; 205 case CH_DRAINING: 206 return "draining"; 207 case CH_DISCONNECTED: 208 return "disconnected"; 209 } 210 return "???"; 211 } 212 213 /** 214 * srpt_qp_event - QP event callback function 215 * @event: Description of the event that occurred. 216 * @ch: SRPT RDMA channel. 217 */ 218 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch) 219 { 220 pr_debug("QP event %d on ch=%p sess_name=%s state=%d\n", 221 event->event, ch, ch->sess_name, ch->state); 222 223 switch (event->event) { 224 case IB_EVENT_COMM_EST: 225 if (ch->using_rdma_cm) 226 rdma_notify(ch->rdma_cm.cm_id, event->event); 227 else 228 ib_cm_notify(ch->ib_cm.cm_id, event->event); 229 break; 230 case IB_EVENT_QP_LAST_WQE_REACHED: 231 pr_debug("%s-%d, state %s: received Last WQE event.\n", 232 ch->sess_name, ch->qp->qp_num, 233 get_ch_state_name(ch->state)); 234 break; 235 default: 236 pr_err("received unrecognized IB QP event %d\n", event->event); 237 break; 238 } 239 } 240 241 /** 242 * srpt_set_ioc - initialize a IOUnitInfo structure 243 * @c_list: controller list. 244 * @slot: one-based slot number. 245 * @value: four-bit value. 246 * 247 * Copies the lowest four bits of value in element slot of the array of four 248 * bit elements called c_list (controller list). The index slot is one-based. 249 */ 250 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value) 251 { 252 u16 id; 253 u8 tmp; 254 255 id = (slot - 1) / 2; 256 if (slot & 0x1) { 257 tmp = c_list[id] & 0xf; 258 c_list[id] = (value << 4) | tmp; 259 } else { 260 tmp = c_list[id] & 0xf0; 261 c_list[id] = (value & 0xf) | tmp; 262 } 263 } 264 265 /** 266 * srpt_get_class_port_info - copy ClassPortInfo to a management datagram 267 * @mad: Datagram that will be sent as response to DM_ATTR_CLASS_PORT_INFO. 268 * 269 * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture 270 * Specification. 271 */ 272 static void srpt_get_class_port_info(struct ib_dm_mad *mad) 273 { 274 struct ib_class_port_info *cif; 275 276 cif = (struct ib_class_port_info *)mad->data; 277 memset(cif, 0, sizeof(*cif)); 278 cif->base_version = 1; 279 cif->class_version = 1; 280 281 ib_set_cpi_resp_time(cif, 20); 282 mad->mad_hdr.status = 0; 283 } 284 285 /** 286 * srpt_get_iou - write IOUnitInfo to a management datagram 287 * @mad: Datagram that will be sent as response to DM_ATTR_IOU_INFO. 288 * 289 * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture 290 * Specification. See also section B.7, table B.6 in the SRP r16a document. 291 */ 292 static void srpt_get_iou(struct ib_dm_mad *mad) 293 { 294 struct ib_dm_iou_info *ioui; 295 u8 slot; 296 int i; 297 298 ioui = (struct ib_dm_iou_info *)mad->data; 299 ioui->change_id = cpu_to_be16(1); 300 ioui->max_controllers = 16; 301 302 /* set present for slot 1 and empty for the rest */ 303 srpt_set_ioc(ioui->controller_list, 1, 1); 304 for (i = 1, slot = 2; i < 16; i++, slot++) 305 srpt_set_ioc(ioui->controller_list, slot, 0); 306 307 mad->mad_hdr.status = 0; 308 } 309 310 /** 311 * srpt_get_ioc - write IOControllerprofile to a management datagram 312 * @sport: HCA port through which the MAD has been received. 313 * @slot: Slot number specified in DM_ATTR_IOC_PROFILE query. 314 * @mad: Datagram that will be sent as response to DM_ATTR_IOC_PROFILE. 315 * 316 * See also section 16.3.3.4 IOControllerProfile in the InfiniBand 317 * Architecture Specification. See also section B.7, table B.7 in the SRP 318 * r16a document. 319 */ 320 static void srpt_get_ioc(struct srpt_port *sport, u32 slot, 321 struct ib_dm_mad *mad) 322 { 323 struct srpt_device *sdev = sport->sdev; 324 struct ib_dm_ioc_profile *iocp; 325 int send_queue_depth; 326 327 iocp = (struct ib_dm_ioc_profile *)mad->data; 328 329 if (!slot || slot > 16) { 330 mad->mad_hdr.status 331 = cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD); 332 return; 333 } 334 335 if (slot > 2) { 336 mad->mad_hdr.status 337 = cpu_to_be16(DM_MAD_STATUS_NO_IOC); 338 return; 339 } 340 341 if (sdev->use_srq) 342 send_queue_depth = sdev->srq_size; 343 else 344 send_queue_depth = min(MAX_SRPT_RQ_SIZE, 345 sdev->device->attrs.max_qp_wr); 346 347 memset(iocp, 0, sizeof(*iocp)); 348 strcpy(iocp->id_string, SRPT_ID_STRING); 349 iocp->guid = cpu_to_be64(srpt_service_guid); 350 iocp->vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id); 351 iocp->device_id = cpu_to_be32(sdev->device->attrs.vendor_part_id); 352 iocp->device_version = cpu_to_be16(sdev->device->attrs.hw_ver); 353 iocp->subsys_vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id); 354 iocp->subsys_device_id = 0x0; 355 iocp->io_class = cpu_to_be16(SRP_REV16A_IB_IO_CLASS); 356 iocp->io_subclass = cpu_to_be16(SRP_IO_SUBCLASS); 357 iocp->protocol = cpu_to_be16(SRP_PROTOCOL); 358 iocp->protocol_version = cpu_to_be16(SRP_PROTOCOL_VERSION); 359 iocp->send_queue_depth = cpu_to_be16(send_queue_depth); 360 iocp->rdma_read_depth = 4; 361 iocp->send_size = cpu_to_be32(srp_max_req_size); 362 iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size, 363 1U << 24)); 364 iocp->num_svc_entries = 1; 365 iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC | 366 SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC; 367 368 mad->mad_hdr.status = 0; 369 } 370 371 /** 372 * srpt_get_svc_entries - write ServiceEntries to a management datagram 373 * @ioc_guid: I/O controller GUID to use in reply. 374 * @slot: I/O controller number. 375 * @hi: End of the range of service entries to be specified in the reply. 376 * @lo: Start of the range of service entries to be specified in the reply.. 377 * @mad: Datagram that will be sent as response to DM_ATTR_SVC_ENTRIES. 378 * 379 * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture 380 * Specification. See also section B.7, table B.8 in the SRP r16a document. 381 */ 382 static void srpt_get_svc_entries(u64 ioc_guid, 383 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad) 384 { 385 struct ib_dm_svc_entries *svc_entries; 386 387 WARN_ON(!ioc_guid); 388 389 if (!slot || slot > 16) { 390 mad->mad_hdr.status 391 = cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD); 392 return; 393 } 394 395 if (slot > 2 || lo > hi || hi > 1) { 396 mad->mad_hdr.status 397 = cpu_to_be16(DM_MAD_STATUS_NO_IOC); 398 return; 399 } 400 401 svc_entries = (struct ib_dm_svc_entries *)mad->data; 402 memset(svc_entries, 0, sizeof(*svc_entries)); 403 svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid); 404 snprintf(svc_entries->service_entries[0].name, 405 sizeof(svc_entries->service_entries[0].name), 406 "%s%016llx", 407 SRP_SERVICE_NAME_PREFIX, 408 ioc_guid); 409 410 mad->mad_hdr.status = 0; 411 } 412 413 /** 414 * srpt_mgmt_method_get - process a received management datagram 415 * @sp: HCA port through which the MAD has been received. 416 * @rq_mad: received MAD. 417 * @rsp_mad: response MAD. 418 */ 419 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad, 420 struct ib_dm_mad *rsp_mad) 421 { 422 u16 attr_id; 423 u32 slot; 424 u8 hi, lo; 425 426 attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id); 427 switch (attr_id) { 428 case DM_ATTR_CLASS_PORT_INFO: 429 srpt_get_class_port_info(rsp_mad); 430 break; 431 case DM_ATTR_IOU_INFO: 432 srpt_get_iou(rsp_mad); 433 break; 434 case DM_ATTR_IOC_PROFILE: 435 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod); 436 srpt_get_ioc(sp, slot, rsp_mad); 437 break; 438 case DM_ATTR_SVC_ENTRIES: 439 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod); 440 hi = (u8) ((slot >> 8) & 0xff); 441 lo = (u8) (slot & 0xff); 442 slot = (u16) ((slot >> 16) & 0xffff); 443 srpt_get_svc_entries(srpt_service_guid, 444 slot, hi, lo, rsp_mad); 445 break; 446 default: 447 rsp_mad->mad_hdr.status = 448 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR); 449 break; 450 } 451 } 452 453 /** 454 * srpt_mad_send_handler - MAD send completion callback 455 * @mad_agent: Return value of ib_register_mad_agent(). 456 * @mad_wc: Work completion reporting that the MAD has been sent. 457 */ 458 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent, 459 struct ib_mad_send_wc *mad_wc) 460 { 461 rdma_destroy_ah(mad_wc->send_buf->ah, RDMA_DESTROY_AH_SLEEPABLE); 462 ib_free_send_mad(mad_wc->send_buf); 463 } 464 465 /** 466 * srpt_mad_recv_handler - MAD reception callback function 467 * @mad_agent: Return value of ib_register_mad_agent(). 468 * @send_buf: Not used. 469 * @mad_wc: Work completion reporting that a MAD has been received. 470 */ 471 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent, 472 struct ib_mad_send_buf *send_buf, 473 struct ib_mad_recv_wc *mad_wc) 474 { 475 struct srpt_port *sport = (struct srpt_port *)mad_agent->context; 476 struct ib_ah *ah; 477 struct ib_mad_send_buf *rsp; 478 struct ib_dm_mad *dm_mad; 479 480 if (!mad_wc || !mad_wc->recv_buf.mad) 481 return; 482 483 ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc, 484 mad_wc->recv_buf.grh, mad_agent->port_num); 485 if (IS_ERR(ah)) 486 goto err; 487 488 BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR); 489 490 rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp, 491 mad_wc->wc->pkey_index, 0, 492 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA, 493 GFP_KERNEL, 494 IB_MGMT_BASE_VERSION); 495 if (IS_ERR(rsp)) 496 goto err_rsp; 497 498 rsp->ah = ah; 499 500 dm_mad = rsp->mad; 501 memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof(*dm_mad)); 502 dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP; 503 dm_mad->mad_hdr.status = 0; 504 505 switch (mad_wc->recv_buf.mad->mad_hdr.method) { 506 case IB_MGMT_METHOD_GET: 507 srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad); 508 break; 509 case IB_MGMT_METHOD_SET: 510 dm_mad->mad_hdr.status = 511 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR); 512 break; 513 default: 514 dm_mad->mad_hdr.status = 515 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD); 516 break; 517 } 518 519 if (!ib_post_send_mad(rsp, NULL)) { 520 ib_free_recv_mad(mad_wc); 521 /* will destroy_ah & free_send_mad in send completion */ 522 return; 523 } 524 525 ib_free_send_mad(rsp); 526 527 err_rsp: 528 rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE); 529 err: 530 ib_free_recv_mad(mad_wc); 531 } 532 533 static int srpt_format_guid(char *buf, unsigned int size, const __be64 *guid) 534 { 535 const __be16 *g = (const __be16 *)guid; 536 537 return snprintf(buf, size, "%04x:%04x:%04x:%04x", 538 be16_to_cpu(g[0]), be16_to_cpu(g[1]), 539 be16_to_cpu(g[2]), be16_to_cpu(g[3])); 540 } 541 542 /** 543 * srpt_refresh_port - configure a HCA port 544 * @sport: SRPT HCA port. 545 * 546 * Enable InfiniBand management datagram processing, update the cached sm_lid, 547 * lid and gid values, and register a callback function for processing MADs 548 * on the specified port. 549 * 550 * Note: It is safe to call this function more than once for the same port. 551 */ 552 static int srpt_refresh_port(struct srpt_port *sport) 553 { 554 struct ib_mad_reg_req reg_req; 555 struct ib_port_modify port_modify; 556 struct ib_port_attr port_attr; 557 int ret; 558 559 memset(&port_modify, 0, sizeof(port_modify)); 560 port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP; 561 port_modify.clr_port_cap_mask = 0; 562 563 ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify); 564 if (ret) 565 goto err_mod_port; 566 567 ret = ib_query_port(sport->sdev->device, sport->port, &port_attr); 568 if (ret) 569 goto err_query_port; 570 571 sport->sm_lid = port_attr.sm_lid; 572 sport->lid = port_attr.lid; 573 574 ret = rdma_query_gid(sport->sdev->device, sport->port, 0, &sport->gid); 575 if (ret) 576 goto err_query_port; 577 578 sport->port_guid_wwn.priv = sport; 579 srpt_format_guid(sport->port_guid, sizeof(sport->port_guid), 580 &sport->gid.global.interface_id); 581 sport->port_gid_wwn.priv = sport; 582 snprintf(sport->port_gid, sizeof(sport->port_gid), 583 "0x%016llx%016llx", 584 be64_to_cpu(sport->gid.global.subnet_prefix), 585 be64_to_cpu(sport->gid.global.interface_id)); 586 587 if (!sport->mad_agent) { 588 memset(®_req, 0, sizeof(reg_req)); 589 reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT; 590 reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION; 591 set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask); 592 set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask); 593 594 sport->mad_agent = ib_register_mad_agent(sport->sdev->device, 595 sport->port, 596 IB_QPT_GSI, 597 ®_req, 0, 598 srpt_mad_send_handler, 599 srpt_mad_recv_handler, 600 sport, 0); 601 if (IS_ERR(sport->mad_agent)) { 602 ret = PTR_ERR(sport->mad_agent); 603 sport->mad_agent = NULL; 604 goto err_query_port; 605 } 606 } 607 608 return 0; 609 610 err_query_port: 611 612 port_modify.set_port_cap_mask = 0; 613 port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP; 614 ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify); 615 616 err_mod_port: 617 618 return ret; 619 } 620 621 /** 622 * srpt_unregister_mad_agent - unregister MAD callback functions 623 * @sdev: SRPT HCA pointer. 624 * 625 * Note: It is safe to call this function more than once for the same device. 626 */ 627 static void srpt_unregister_mad_agent(struct srpt_device *sdev) 628 { 629 struct ib_port_modify port_modify = { 630 .clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP, 631 }; 632 struct srpt_port *sport; 633 int i; 634 635 for (i = 1; i <= sdev->device->phys_port_cnt; i++) { 636 sport = &sdev->port[i - 1]; 637 WARN_ON(sport->port != i); 638 if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0) 639 pr_err("disabling MAD processing failed.\n"); 640 if (sport->mad_agent) { 641 ib_unregister_mad_agent(sport->mad_agent); 642 sport->mad_agent = NULL; 643 } 644 } 645 } 646 647 /** 648 * srpt_alloc_ioctx - allocate a SRPT I/O context structure 649 * @sdev: SRPT HCA pointer. 650 * @ioctx_size: I/O context size. 651 * @buf_cache: I/O buffer cache. 652 * @dir: DMA data direction. 653 */ 654 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev, 655 int ioctx_size, 656 struct kmem_cache *buf_cache, 657 enum dma_data_direction dir) 658 { 659 struct srpt_ioctx *ioctx; 660 661 ioctx = kzalloc(ioctx_size, GFP_KERNEL); 662 if (!ioctx) 663 goto err; 664 665 ioctx->buf = kmem_cache_alloc(buf_cache, GFP_KERNEL); 666 if (!ioctx->buf) 667 goto err_free_ioctx; 668 669 ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, 670 kmem_cache_size(buf_cache), dir); 671 if (ib_dma_mapping_error(sdev->device, ioctx->dma)) 672 goto err_free_buf; 673 674 return ioctx; 675 676 err_free_buf: 677 kmem_cache_free(buf_cache, ioctx->buf); 678 err_free_ioctx: 679 kfree(ioctx); 680 err: 681 return NULL; 682 } 683 684 /** 685 * srpt_free_ioctx - free a SRPT I/O context structure 686 * @sdev: SRPT HCA pointer. 687 * @ioctx: I/O context pointer. 688 * @buf_cache: I/O buffer cache. 689 * @dir: DMA data direction. 690 */ 691 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx, 692 struct kmem_cache *buf_cache, 693 enum dma_data_direction dir) 694 { 695 if (!ioctx) 696 return; 697 698 ib_dma_unmap_single(sdev->device, ioctx->dma, 699 kmem_cache_size(buf_cache), dir); 700 kmem_cache_free(buf_cache, ioctx->buf); 701 kfree(ioctx); 702 } 703 704 /** 705 * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures 706 * @sdev: Device to allocate the I/O context ring for. 707 * @ring_size: Number of elements in the I/O context ring. 708 * @ioctx_size: I/O context size. 709 * @buf_cache: I/O buffer cache. 710 * @alignment_offset: Offset in each ring buffer at which the SRP information 711 * unit starts. 712 * @dir: DMA data direction. 713 */ 714 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev, 715 int ring_size, int ioctx_size, 716 struct kmem_cache *buf_cache, 717 int alignment_offset, 718 enum dma_data_direction dir) 719 { 720 struct srpt_ioctx **ring; 721 int i; 722 723 WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) && 724 ioctx_size != sizeof(struct srpt_send_ioctx)); 725 726 ring = kvmalloc_array(ring_size, sizeof(ring[0]), GFP_KERNEL); 727 if (!ring) 728 goto out; 729 for (i = 0; i < ring_size; ++i) { 730 ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, buf_cache, dir); 731 if (!ring[i]) 732 goto err; 733 ring[i]->index = i; 734 ring[i]->offset = alignment_offset; 735 } 736 goto out; 737 738 err: 739 while (--i >= 0) 740 srpt_free_ioctx(sdev, ring[i], buf_cache, dir); 741 kvfree(ring); 742 ring = NULL; 743 out: 744 return ring; 745 } 746 747 /** 748 * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures 749 * @ioctx_ring: I/O context ring to be freed. 750 * @sdev: SRPT HCA pointer. 751 * @ring_size: Number of ring elements. 752 * @buf_cache: I/O buffer cache. 753 * @dir: DMA data direction. 754 */ 755 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring, 756 struct srpt_device *sdev, int ring_size, 757 struct kmem_cache *buf_cache, 758 enum dma_data_direction dir) 759 { 760 int i; 761 762 if (!ioctx_ring) 763 return; 764 765 for (i = 0; i < ring_size; ++i) 766 srpt_free_ioctx(sdev, ioctx_ring[i], buf_cache, dir); 767 kvfree(ioctx_ring); 768 } 769 770 /** 771 * srpt_set_cmd_state - set the state of a SCSI command 772 * @ioctx: Send I/O context. 773 * @new: New I/O context state. 774 * 775 * Does not modify the state of aborted commands. Returns the previous command 776 * state. 777 */ 778 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx, 779 enum srpt_command_state new) 780 { 781 enum srpt_command_state previous; 782 783 previous = ioctx->state; 784 if (previous != SRPT_STATE_DONE) 785 ioctx->state = new; 786 787 return previous; 788 } 789 790 /** 791 * srpt_test_and_set_cmd_state - test and set the state of a command 792 * @ioctx: Send I/O context. 793 * @old: Current I/O context state. 794 * @new: New I/O context state. 795 * 796 * Returns true if and only if the previous command state was equal to 'old'. 797 */ 798 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx, 799 enum srpt_command_state old, 800 enum srpt_command_state new) 801 { 802 enum srpt_command_state previous; 803 804 WARN_ON(!ioctx); 805 WARN_ON(old == SRPT_STATE_DONE); 806 WARN_ON(new == SRPT_STATE_NEW); 807 808 previous = ioctx->state; 809 if (previous == old) 810 ioctx->state = new; 811 812 return previous == old; 813 } 814 815 /** 816 * srpt_post_recv - post an IB receive request 817 * @sdev: SRPT HCA pointer. 818 * @ch: SRPT RDMA channel. 819 * @ioctx: Receive I/O context pointer. 820 */ 821 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch, 822 struct srpt_recv_ioctx *ioctx) 823 { 824 struct ib_sge list; 825 struct ib_recv_wr wr; 826 827 BUG_ON(!sdev); 828 list.addr = ioctx->ioctx.dma + ioctx->ioctx.offset; 829 list.length = srp_max_req_size; 830 list.lkey = sdev->lkey; 831 832 ioctx->ioctx.cqe.done = srpt_recv_done; 833 wr.wr_cqe = &ioctx->ioctx.cqe; 834 wr.next = NULL; 835 wr.sg_list = &list; 836 wr.num_sge = 1; 837 838 if (sdev->use_srq) 839 return ib_post_srq_recv(sdev->srq, &wr, NULL); 840 else 841 return ib_post_recv(ch->qp, &wr, NULL); 842 } 843 844 /** 845 * srpt_zerolength_write - perform a zero-length RDMA write 846 * @ch: SRPT RDMA channel. 847 * 848 * A quote from the InfiniBand specification: C9-88: For an HCA responder 849 * using Reliable Connection service, for each zero-length RDMA READ or WRITE 850 * request, the R_Key shall not be validated, even if the request includes 851 * Immediate data. 852 */ 853 static int srpt_zerolength_write(struct srpt_rdma_ch *ch) 854 { 855 struct ib_rdma_wr wr = { 856 .wr = { 857 .next = NULL, 858 { .wr_cqe = &ch->zw_cqe, }, 859 .opcode = IB_WR_RDMA_WRITE, 860 .send_flags = IB_SEND_SIGNALED, 861 } 862 }; 863 864 pr_debug("%s-%d: queued zerolength write\n", ch->sess_name, 865 ch->qp->qp_num); 866 867 return ib_post_send(ch->qp, &wr.wr, NULL); 868 } 869 870 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc) 871 { 872 struct srpt_rdma_ch *ch = cq->cq_context; 873 874 pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num, 875 wc->status); 876 877 if (wc->status == IB_WC_SUCCESS) { 878 srpt_process_wait_list(ch); 879 } else { 880 if (srpt_set_ch_state(ch, CH_DISCONNECTED)) 881 schedule_work(&ch->release_work); 882 else 883 pr_debug("%s-%d: already disconnected.\n", 884 ch->sess_name, ch->qp->qp_num); 885 } 886 } 887 888 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx, 889 struct srp_direct_buf *db, int nbufs, struct scatterlist **sg, 890 unsigned *sg_cnt) 891 { 892 enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd); 893 struct srpt_rdma_ch *ch = ioctx->ch; 894 struct scatterlist *prev = NULL; 895 unsigned prev_nents; 896 int ret, i; 897 898 if (nbufs == 1) { 899 ioctx->rw_ctxs = &ioctx->s_rw_ctx; 900 } else { 901 ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs), 902 GFP_KERNEL); 903 if (!ioctx->rw_ctxs) 904 return -ENOMEM; 905 } 906 907 for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) { 908 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 909 u64 remote_addr = be64_to_cpu(db->va); 910 u32 size = be32_to_cpu(db->len); 911 u32 rkey = be32_to_cpu(db->key); 912 913 ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false, 914 i < nbufs - 1); 915 if (ret) 916 goto unwind; 917 918 ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port, 919 ctx->sg, ctx->nents, 0, remote_addr, rkey, dir); 920 if (ret < 0) { 921 target_free_sgl(ctx->sg, ctx->nents); 922 goto unwind; 923 } 924 925 ioctx->n_rdma += ret; 926 ioctx->n_rw_ctx++; 927 928 if (prev) { 929 sg_unmark_end(&prev[prev_nents - 1]); 930 sg_chain(prev, prev_nents + 1, ctx->sg); 931 } else { 932 *sg = ctx->sg; 933 } 934 935 prev = ctx->sg; 936 prev_nents = ctx->nents; 937 938 *sg_cnt += ctx->nents; 939 } 940 941 return 0; 942 943 unwind: 944 while (--i >= 0) { 945 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 946 947 rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port, 948 ctx->sg, ctx->nents, dir); 949 target_free_sgl(ctx->sg, ctx->nents); 950 } 951 if (ioctx->rw_ctxs != &ioctx->s_rw_ctx) 952 kfree(ioctx->rw_ctxs); 953 return ret; 954 } 955 956 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch, 957 struct srpt_send_ioctx *ioctx) 958 { 959 enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd); 960 int i; 961 962 for (i = 0; i < ioctx->n_rw_ctx; i++) { 963 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 964 965 rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port, 966 ctx->sg, ctx->nents, dir); 967 target_free_sgl(ctx->sg, ctx->nents); 968 } 969 970 if (ioctx->rw_ctxs != &ioctx->s_rw_ctx) 971 kfree(ioctx->rw_ctxs); 972 } 973 974 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd) 975 { 976 /* 977 * The pointer computations below will only be compiled correctly 978 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check 979 * whether srp_cmd::add_data has been declared as a byte pointer. 980 */ 981 BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) && 982 !__same_type(srp_cmd->add_data[0], (u8)0)); 983 984 /* 985 * According to the SRP spec, the lower two bits of the 'ADDITIONAL 986 * CDB LENGTH' field are reserved and the size in bytes of this field 987 * is four times the value specified in bits 3..7. Hence the "& ~3". 988 */ 989 return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3); 990 } 991 992 /** 993 * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request 994 * @recv_ioctx: I/O context associated with the received command @srp_cmd. 995 * @ioctx: I/O context that will be used for responding to the initiator. 996 * @srp_cmd: Pointer to the SRP_CMD request data. 997 * @dir: Pointer to the variable to which the transfer direction will be 998 * written. 999 * @sg: [out] scatterlist for the parsed SRP_CMD. 1000 * @sg_cnt: [out] length of @sg. 1001 * @data_len: Pointer to the variable to which the total data length of all 1002 * descriptors in the SRP_CMD request will be written. 1003 * @imm_data_offset: [in] Offset in SRP_CMD requests at which immediate data 1004 * starts. 1005 * 1006 * This function initializes ioctx->nrbuf and ioctx->r_bufs. 1007 * 1008 * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors; 1009 * -ENOMEM when memory allocation fails and zero upon success. 1010 */ 1011 static int srpt_get_desc_tbl(struct srpt_recv_ioctx *recv_ioctx, 1012 struct srpt_send_ioctx *ioctx, 1013 struct srp_cmd *srp_cmd, enum dma_data_direction *dir, 1014 struct scatterlist **sg, unsigned int *sg_cnt, u64 *data_len, 1015 u16 imm_data_offset) 1016 { 1017 BUG_ON(!dir); 1018 BUG_ON(!data_len); 1019 1020 /* 1021 * The lower four bits of the buffer format field contain the DATA-IN 1022 * buffer descriptor format, and the highest four bits contain the 1023 * DATA-OUT buffer descriptor format. 1024 */ 1025 if (srp_cmd->buf_fmt & 0xf) 1026 /* DATA-IN: transfer data from target to initiator (read). */ 1027 *dir = DMA_FROM_DEVICE; 1028 else if (srp_cmd->buf_fmt >> 4) 1029 /* DATA-OUT: transfer data from initiator to target (write). */ 1030 *dir = DMA_TO_DEVICE; 1031 else 1032 *dir = DMA_NONE; 1033 1034 /* initialize data_direction early as srpt_alloc_rw_ctxs needs it */ 1035 ioctx->cmd.data_direction = *dir; 1036 1037 if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) || 1038 ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) { 1039 struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd); 1040 1041 *data_len = be32_to_cpu(db->len); 1042 return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt); 1043 } else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) || 1044 ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) { 1045 struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd); 1046 int nbufs = be32_to_cpu(idb->table_desc.len) / 1047 sizeof(struct srp_direct_buf); 1048 1049 if (nbufs > 1050 (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) { 1051 pr_err("received unsupported SRP_CMD request type (%u out + %u in != %u / %zu)\n", 1052 srp_cmd->data_out_desc_cnt, 1053 srp_cmd->data_in_desc_cnt, 1054 be32_to_cpu(idb->table_desc.len), 1055 sizeof(struct srp_direct_buf)); 1056 return -EINVAL; 1057 } 1058 1059 *data_len = be32_to_cpu(idb->len); 1060 return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs, 1061 sg, sg_cnt); 1062 } else if ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_IMM) { 1063 struct srp_imm_buf *imm_buf = srpt_get_desc_buf(srp_cmd); 1064 void *data = (void *)srp_cmd + imm_data_offset; 1065 uint32_t len = be32_to_cpu(imm_buf->len); 1066 uint32_t req_size = imm_data_offset + len; 1067 1068 if (req_size > srp_max_req_size) { 1069 pr_err("Immediate data (length %d + %d) exceeds request size %d\n", 1070 imm_data_offset, len, srp_max_req_size); 1071 return -EINVAL; 1072 } 1073 if (recv_ioctx->byte_len < req_size) { 1074 pr_err("Received too few data - %d < %d\n", 1075 recv_ioctx->byte_len, req_size); 1076 return -EIO; 1077 } 1078 /* 1079 * The immediate data buffer descriptor must occur before the 1080 * immediate data itself. 1081 */ 1082 if ((void *)(imm_buf + 1) > (void *)data) { 1083 pr_err("Received invalid write request\n"); 1084 return -EINVAL; 1085 } 1086 *data_len = len; 1087 ioctx->recv_ioctx = recv_ioctx; 1088 if ((uintptr_t)data & 511) { 1089 pr_warn_once("Internal error - the receive buffers are not aligned properly.\n"); 1090 return -EINVAL; 1091 } 1092 sg_init_one(&ioctx->imm_sg, data, len); 1093 *sg = &ioctx->imm_sg; 1094 *sg_cnt = 1; 1095 return 0; 1096 } else { 1097 *data_len = 0; 1098 return 0; 1099 } 1100 } 1101 1102 /** 1103 * srpt_init_ch_qp - initialize queue pair attributes 1104 * @ch: SRPT RDMA channel. 1105 * @qp: Queue pair pointer. 1106 * 1107 * Initialized the attributes of queue pair 'qp' by allowing local write, 1108 * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT. 1109 */ 1110 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1111 { 1112 struct ib_qp_attr *attr; 1113 int ret; 1114 1115 WARN_ON_ONCE(ch->using_rdma_cm); 1116 1117 attr = kzalloc(sizeof(*attr), GFP_KERNEL); 1118 if (!attr) 1119 return -ENOMEM; 1120 1121 attr->qp_state = IB_QPS_INIT; 1122 attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE; 1123 attr->port_num = ch->sport->port; 1124 1125 ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port, 1126 ch->pkey, &attr->pkey_index); 1127 if (ret < 0) 1128 pr_err("Translating pkey %#x failed (%d) - using index 0\n", 1129 ch->pkey, ret); 1130 1131 ret = ib_modify_qp(qp, attr, 1132 IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT | 1133 IB_QP_PKEY_INDEX); 1134 1135 kfree(attr); 1136 return ret; 1137 } 1138 1139 /** 1140 * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR) 1141 * @ch: channel of the queue pair. 1142 * @qp: queue pair to change the state of. 1143 * 1144 * Returns zero upon success and a negative value upon failure. 1145 * 1146 * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system. 1147 * If this structure ever becomes larger, it might be necessary to allocate 1148 * it dynamically instead of on the stack. 1149 */ 1150 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1151 { 1152 struct ib_qp_attr qp_attr; 1153 int attr_mask; 1154 int ret; 1155 1156 WARN_ON_ONCE(ch->using_rdma_cm); 1157 1158 qp_attr.qp_state = IB_QPS_RTR; 1159 ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask); 1160 if (ret) 1161 goto out; 1162 1163 qp_attr.max_dest_rd_atomic = 4; 1164 1165 ret = ib_modify_qp(qp, &qp_attr, attr_mask); 1166 1167 out: 1168 return ret; 1169 } 1170 1171 /** 1172 * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS) 1173 * @ch: channel of the queue pair. 1174 * @qp: queue pair to change the state of. 1175 * 1176 * Returns zero upon success and a negative value upon failure. 1177 * 1178 * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system. 1179 * If this structure ever becomes larger, it might be necessary to allocate 1180 * it dynamically instead of on the stack. 1181 */ 1182 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1183 { 1184 struct ib_qp_attr qp_attr; 1185 int attr_mask; 1186 int ret; 1187 1188 qp_attr.qp_state = IB_QPS_RTS; 1189 ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask); 1190 if (ret) 1191 goto out; 1192 1193 qp_attr.max_rd_atomic = 4; 1194 1195 ret = ib_modify_qp(qp, &qp_attr, attr_mask); 1196 1197 out: 1198 return ret; 1199 } 1200 1201 /** 1202 * srpt_ch_qp_err - set the channel queue pair state to 'error' 1203 * @ch: SRPT RDMA channel. 1204 */ 1205 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch) 1206 { 1207 struct ib_qp_attr qp_attr; 1208 1209 qp_attr.qp_state = IB_QPS_ERR; 1210 return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE); 1211 } 1212 1213 /** 1214 * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator 1215 * @ch: SRPT RDMA channel. 1216 */ 1217 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch) 1218 { 1219 struct srpt_send_ioctx *ioctx; 1220 int tag, cpu; 1221 1222 BUG_ON(!ch); 1223 1224 tag = sbitmap_queue_get(&ch->sess->sess_tag_pool, &cpu); 1225 if (tag < 0) 1226 return NULL; 1227 1228 ioctx = ch->ioctx_ring[tag]; 1229 BUG_ON(ioctx->ch != ch); 1230 ioctx->state = SRPT_STATE_NEW; 1231 WARN_ON_ONCE(ioctx->recv_ioctx); 1232 ioctx->n_rdma = 0; 1233 ioctx->n_rw_ctx = 0; 1234 ioctx->queue_status_only = false; 1235 /* 1236 * transport_init_se_cmd() does not initialize all fields, so do it 1237 * here. 1238 */ 1239 memset(&ioctx->cmd, 0, sizeof(ioctx->cmd)); 1240 memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data)); 1241 ioctx->cmd.map_tag = tag; 1242 ioctx->cmd.map_cpu = cpu; 1243 1244 return ioctx; 1245 } 1246 1247 /** 1248 * srpt_abort_cmd - abort a SCSI command 1249 * @ioctx: I/O context associated with the SCSI command. 1250 */ 1251 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx) 1252 { 1253 enum srpt_command_state state; 1254 1255 BUG_ON(!ioctx); 1256 1257 /* 1258 * If the command is in a state where the target core is waiting for 1259 * the ib_srpt driver, change the state to the next state. 1260 */ 1261 1262 state = ioctx->state; 1263 switch (state) { 1264 case SRPT_STATE_NEED_DATA: 1265 ioctx->state = SRPT_STATE_DATA_IN; 1266 break; 1267 case SRPT_STATE_CMD_RSP_SENT: 1268 case SRPT_STATE_MGMT_RSP_SENT: 1269 ioctx->state = SRPT_STATE_DONE; 1270 break; 1271 default: 1272 WARN_ONCE(true, "%s: unexpected I/O context state %d\n", 1273 __func__, state); 1274 break; 1275 } 1276 1277 pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state, 1278 ioctx->state, ioctx->cmd.tag); 1279 1280 switch (state) { 1281 case SRPT_STATE_NEW: 1282 case SRPT_STATE_DATA_IN: 1283 case SRPT_STATE_MGMT: 1284 case SRPT_STATE_DONE: 1285 /* 1286 * Do nothing - defer abort processing until 1287 * srpt_queue_response() is invoked. 1288 */ 1289 break; 1290 case SRPT_STATE_NEED_DATA: 1291 pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag); 1292 transport_generic_request_failure(&ioctx->cmd, 1293 TCM_CHECK_CONDITION_ABORT_CMD); 1294 break; 1295 case SRPT_STATE_CMD_RSP_SENT: 1296 /* 1297 * SRP_RSP sending failed or the SRP_RSP send completion has 1298 * not been received in time. 1299 */ 1300 transport_generic_free_cmd(&ioctx->cmd, 0); 1301 break; 1302 case SRPT_STATE_MGMT_RSP_SENT: 1303 transport_generic_free_cmd(&ioctx->cmd, 0); 1304 break; 1305 default: 1306 WARN(1, "Unexpected command state (%d)", state); 1307 break; 1308 } 1309 1310 return state; 1311 } 1312 1313 /** 1314 * srpt_rdma_read_done - RDMA read completion callback 1315 * @cq: Completion queue. 1316 * @wc: Work completion. 1317 * 1318 * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping 1319 * the data that has been transferred via IB RDMA had to be postponed until the 1320 * check_stop_free() callback. None of this is necessary anymore and needs to 1321 * be cleaned up. 1322 */ 1323 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc) 1324 { 1325 struct srpt_rdma_ch *ch = cq->cq_context; 1326 struct srpt_send_ioctx *ioctx = 1327 container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe); 1328 1329 WARN_ON(ioctx->n_rdma <= 0); 1330 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail); 1331 ioctx->n_rdma = 0; 1332 1333 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1334 pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n", 1335 ioctx, wc->status); 1336 srpt_abort_cmd(ioctx); 1337 return; 1338 } 1339 1340 if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA, 1341 SRPT_STATE_DATA_IN)) 1342 target_execute_cmd(&ioctx->cmd); 1343 else 1344 pr_err("%s[%d]: wrong state = %d\n", __func__, 1345 __LINE__, ioctx->state); 1346 } 1347 1348 /** 1349 * srpt_build_cmd_rsp - build a SRP_RSP response 1350 * @ch: RDMA channel through which the request has been received. 1351 * @ioctx: I/O context associated with the SRP_CMD request. The response will 1352 * be built in the buffer ioctx->buf points at and hence this function will 1353 * overwrite the request data. 1354 * @tag: tag of the request for which this response is being generated. 1355 * @status: value for the STATUS field of the SRP_RSP information unit. 1356 * 1357 * Returns the size in bytes of the SRP_RSP response. 1358 * 1359 * An SRP_RSP response contains a SCSI status or service response. See also 1360 * section 6.9 in the SRP r16a document for the format of an SRP_RSP 1361 * response. See also SPC-2 for more information about sense data. 1362 */ 1363 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch, 1364 struct srpt_send_ioctx *ioctx, u64 tag, 1365 int status) 1366 { 1367 struct srp_rsp *srp_rsp; 1368 const u8 *sense_data; 1369 int sense_data_len, max_sense_len; 1370 1371 /* 1372 * The lowest bit of all SAM-3 status codes is zero (see also 1373 * paragraph 5.3 in SAM-3). 1374 */ 1375 WARN_ON(status & 1); 1376 1377 srp_rsp = ioctx->ioctx.buf; 1378 BUG_ON(!srp_rsp); 1379 1380 sense_data = ioctx->sense_data; 1381 sense_data_len = ioctx->cmd.scsi_sense_length; 1382 WARN_ON(sense_data_len > sizeof(ioctx->sense_data)); 1383 1384 memset(srp_rsp, 0, sizeof(*srp_rsp)); 1385 srp_rsp->opcode = SRP_RSP; 1386 srp_rsp->req_lim_delta = 1387 cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0)); 1388 srp_rsp->tag = tag; 1389 srp_rsp->status = status; 1390 1391 if (sense_data_len) { 1392 BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp)); 1393 max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp); 1394 if (sense_data_len > max_sense_len) { 1395 pr_warn("truncated sense data from %d to %d bytes\n", 1396 sense_data_len, max_sense_len); 1397 sense_data_len = max_sense_len; 1398 } 1399 1400 srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID; 1401 srp_rsp->sense_data_len = cpu_to_be32(sense_data_len); 1402 memcpy(srp_rsp + 1, sense_data, sense_data_len); 1403 } 1404 1405 return sizeof(*srp_rsp) + sense_data_len; 1406 } 1407 1408 /** 1409 * srpt_build_tskmgmt_rsp - build a task management response 1410 * @ch: RDMA channel through which the request has been received. 1411 * @ioctx: I/O context in which the SRP_RSP response will be built. 1412 * @rsp_code: RSP_CODE that will be stored in the response. 1413 * @tag: Tag of the request for which this response is being generated. 1414 * 1415 * Returns the size in bytes of the SRP_RSP response. 1416 * 1417 * An SRP_RSP response contains a SCSI status or service response. See also 1418 * section 6.9 in the SRP r16a document for the format of an SRP_RSP 1419 * response. 1420 */ 1421 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch, 1422 struct srpt_send_ioctx *ioctx, 1423 u8 rsp_code, u64 tag) 1424 { 1425 struct srp_rsp *srp_rsp; 1426 int resp_data_len; 1427 int resp_len; 1428 1429 resp_data_len = 4; 1430 resp_len = sizeof(*srp_rsp) + resp_data_len; 1431 1432 srp_rsp = ioctx->ioctx.buf; 1433 BUG_ON(!srp_rsp); 1434 memset(srp_rsp, 0, sizeof(*srp_rsp)); 1435 1436 srp_rsp->opcode = SRP_RSP; 1437 srp_rsp->req_lim_delta = 1438 cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0)); 1439 srp_rsp->tag = tag; 1440 1441 srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID; 1442 srp_rsp->resp_data_len = cpu_to_be32(resp_data_len); 1443 srp_rsp->data[3] = rsp_code; 1444 1445 return resp_len; 1446 } 1447 1448 static int srpt_check_stop_free(struct se_cmd *cmd) 1449 { 1450 struct srpt_send_ioctx *ioctx = container_of(cmd, 1451 struct srpt_send_ioctx, cmd); 1452 1453 return target_put_sess_cmd(&ioctx->cmd); 1454 } 1455 1456 /** 1457 * srpt_handle_cmd - process a SRP_CMD information unit 1458 * @ch: SRPT RDMA channel. 1459 * @recv_ioctx: Receive I/O context. 1460 * @send_ioctx: Send I/O context. 1461 */ 1462 static void srpt_handle_cmd(struct srpt_rdma_ch *ch, 1463 struct srpt_recv_ioctx *recv_ioctx, 1464 struct srpt_send_ioctx *send_ioctx) 1465 { 1466 struct se_cmd *cmd; 1467 struct srp_cmd *srp_cmd; 1468 struct scatterlist *sg = NULL; 1469 unsigned sg_cnt = 0; 1470 u64 data_len; 1471 enum dma_data_direction dir; 1472 int rc; 1473 1474 BUG_ON(!send_ioctx); 1475 1476 srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset; 1477 cmd = &send_ioctx->cmd; 1478 cmd->tag = srp_cmd->tag; 1479 1480 switch (srp_cmd->task_attr) { 1481 case SRP_CMD_SIMPLE_Q: 1482 cmd->sam_task_attr = TCM_SIMPLE_TAG; 1483 break; 1484 case SRP_CMD_ORDERED_Q: 1485 default: 1486 cmd->sam_task_attr = TCM_ORDERED_TAG; 1487 break; 1488 case SRP_CMD_HEAD_OF_Q: 1489 cmd->sam_task_attr = TCM_HEAD_TAG; 1490 break; 1491 case SRP_CMD_ACA: 1492 cmd->sam_task_attr = TCM_ACA_TAG; 1493 break; 1494 } 1495 1496 rc = srpt_get_desc_tbl(recv_ioctx, send_ioctx, srp_cmd, &dir, 1497 &sg, &sg_cnt, &data_len, ch->imm_data_offset); 1498 if (rc) { 1499 if (rc != -EAGAIN) { 1500 pr_err("0x%llx: parsing SRP descriptor table failed.\n", 1501 srp_cmd->tag); 1502 } 1503 goto busy; 1504 } 1505 1506 rc = target_submit_cmd_map_sgls(cmd, ch->sess, srp_cmd->cdb, 1507 &send_ioctx->sense_data[0], 1508 scsilun_to_int(&srp_cmd->lun), data_len, 1509 TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF, 1510 sg, sg_cnt, NULL, 0, NULL, 0); 1511 if (rc != 0) { 1512 pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc, 1513 srp_cmd->tag); 1514 goto busy; 1515 } 1516 return; 1517 1518 busy: 1519 target_send_busy(cmd); 1520 } 1521 1522 static int srp_tmr_to_tcm(int fn) 1523 { 1524 switch (fn) { 1525 case SRP_TSK_ABORT_TASK: 1526 return TMR_ABORT_TASK; 1527 case SRP_TSK_ABORT_TASK_SET: 1528 return TMR_ABORT_TASK_SET; 1529 case SRP_TSK_CLEAR_TASK_SET: 1530 return TMR_CLEAR_TASK_SET; 1531 case SRP_TSK_LUN_RESET: 1532 return TMR_LUN_RESET; 1533 case SRP_TSK_CLEAR_ACA: 1534 return TMR_CLEAR_ACA; 1535 default: 1536 return -1; 1537 } 1538 } 1539 1540 /** 1541 * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit 1542 * @ch: SRPT RDMA channel. 1543 * @recv_ioctx: Receive I/O context. 1544 * @send_ioctx: Send I/O context. 1545 * 1546 * Returns 0 if and only if the request will be processed by the target core. 1547 * 1548 * For more information about SRP_TSK_MGMT information units, see also section 1549 * 6.7 in the SRP r16a document. 1550 */ 1551 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch, 1552 struct srpt_recv_ioctx *recv_ioctx, 1553 struct srpt_send_ioctx *send_ioctx) 1554 { 1555 struct srp_tsk_mgmt *srp_tsk; 1556 struct se_cmd *cmd; 1557 struct se_session *sess = ch->sess; 1558 int tcm_tmr; 1559 int rc; 1560 1561 BUG_ON(!send_ioctx); 1562 1563 srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset; 1564 cmd = &send_ioctx->cmd; 1565 1566 pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n", 1567 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch, 1568 ch->sess); 1569 1570 srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT); 1571 send_ioctx->cmd.tag = srp_tsk->tag; 1572 tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func); 1573 rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, 1574 scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr, 1575 GFP_KERNEL, srp_tsk->task_tag, 1576 TARGET_SCF_ACK_KREF); 1577 if (rc != 0) { 1578 send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED; 1579 cmd->se_tfo->queue_tm_rsp(cmd); 1580 } 1581 return; 1582 } 1583 1584 /** 1585 * srpt_handle_new_iu - process a newly received information unit 1586 * @ch: RDMA channel through which the information unit has been received. 1587 * @recv_ioctx: Receive I/O context associated with the information unit. 1588 */ 1589 static bool 1590 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx) 1591 { 1592 struct srpt_send_ioctx *send_ioctx = NULL; 1593 struct srp_cmd *srp_cmd; 1594 bool res = false; 1595 u8 opcode; 1596 1597 BUG_ON(!ch); 1598 BUG_ON(!recv_ioctx); 1599 1600 if (unlikely(ch->state == CH_CONNECTING)) 1601 goto push; 1602 1603 ib_dma_sync_single_for_cpu(ch->sport->sdev->device, 1604 recv_ioctx->ioctx.dma, 1605 recv_ioctx->ioctx.offset + srp_max_req_size, 1606 DMA_FROM_DEVICE); 1607 1608 srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset; 1609 opcode = srp_cmd->opcode; 1610 if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) { 1611 send_ioctx = srpt_get_send_ioctx(ch); 1612 if (unlikely(!send_ioctx)) 1613 goto push; 1614 } 1615 1616 if (!list_empty(&recv_ioctx->wait_list)) { 1617 WARN_ON_ONCE(!ch->processing_wait_list); 1618 list_del_init(&recv_ioctx->wait_list); 1619 } 1620 1621 switch (opcode) { 1622 case SRP_CMD: 1623 srpt_handle_cmd(ch, recv_ioctx, send_ioctx); 1624 break; 1625 case SRP_TSK_MGMT: 1626 srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx); 1627 break; 1628 case SRP_I_LOGOUT: 1629 pr_err("Not yet implemented: SRP_I_LOGOUT\n"); 1630 break; 1631 case SRP_CRED_RSP: 1632 pr_debug("received SRP_CRED_RSP\n"); 1633 break; 1634 case SRP_AER_RSP: 1635 pr_debug("received SRP_AER_RSP\n"); 1636 break; 1637 case SRP_RSP: 1638 pr_err("Received SRP_RSP\n"); 1639 break; 1640 default: 1641 pr_err("received IU with unknown opcode 0x%x\n", opcode); 1642 break; 1643 } 1644 1645 if (!send_ioctx || !send_ioctx->recv_ioctx) 1646 srpt_post_recv(ch->sport->sdev, ch, recv_ioctx); 1647 res = true; 1648 1649 out: 1650 return res; 1651 1652 push: 1653 if (list_empty(&recv_ioctx->wait_list)) { 1654 WARN_ON_ONCE(ch->processing_wait_list); 1655 list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list); 1656 } 1657 goto out; 1658 } 1659 1660 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc) 1661 { 1662 struct srpt_rdma_ch *ch = cq->cq_context; 1663 struct srpt_recv_ioctx *ioctx = 1664 container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe); 1665 1666 if (wc->status == IB_WC_SUCCESS) { 1667 int req_lim; 1668 1669 req_lim = atomic_dec_return(&ch->req_lim); 1670 if (unlikely(req_lim < 0)) 1671 pr_err("req_lim = %d < 0\n", req_lim); 1672 ioctx->byte_len = wc->byte_len; 1673 srpt_handle_new_iu(ch, ioctx); 1674 } else { 1675 pr_info_ratelimited("receiving failed for ioctx %p with status %d\n", 1676 ioctx, wc->status); 1677 } 1678 } 1679 1680 /* 1681 * This function must be called from the context in which RDMA completions are 1682 * processed because it accesses the wait list without protection against 1683 * access from other threads. 1684 */ 1685 static void srpt_process_wait_list(struct srpt_rdma_ch *ch) 1686 { 1687 struct srpt_recv_ioctx *recv_ioctx, *tmp; 1688 1689 WARN_ON_ONCE(ch->state == CH_CONNECTING); 1690 1691 if (list_empty(&ch->cmd_wait_list)) 1692 return; 1693 1694 WARN_ON_ONCE(ch->processing_wait_list); 1695 ch->processing_wait_list = true; 1696 list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list, 1697 wait_list) { 1698 if (!srpt_handle_new_iu(ch, recv_ioctx)) 1699 break; 1700 } 1701 ch->processing_wait_list = false; 1702 } 1703 1704 /** 1705 * srpt_send_done - send completion callback 1706 * @cq: Completion queue. 1707 * @wc: Work completion. 1708 * 1709 * Note: Although this has not yet been observed during tests, at least in 1710 * theory it is possible that the srpt_get_send_ioctx() call invoked by 1711 * srpt_handle_new_iu() fails. This is possible because the req_lim_delta 1712 * value in each response is set to one, and it is possible that this response 1713 * makes the initiator send a new request before the send completion for that 1714 * response has been processed. This could e.g. happen if the call to 1715 * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or 1716 * if IB retransmission causes generation of the send completion to be 1717 * delayed. Incoming information units for which srpt_get_send_ioctx() fails 1718 * are queued on cmd_wait_list. The code below processes these delayed 1719 * requests one at a time. 1720 */ 1721 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc) 1722 { 1723 struct srpt_rdma_ch *ch = cq->cq_context; 1724 struct srpt_send_ioctx *ioctx = 1725 container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe); 1726 enum srpt_command_state state; 1727 1728 state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE); 1729 1730 WARN_ON(state != SRPT_STATE_CMD_RSP_SENT && 1731 state != SRPT_STATE_MGMT_RSP_SENT); 1732 1733 atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail); 1734 1735 if (wc->status != IB_WC_SUCCESS) 1736 pr_info("sending response for ioctx 0x%p failed with status %d\n", 1737 ioctx, wc->status); 1738 1739 if (state != SRPT_STATE_DONE) { 1740 transport_generic_free_cmd(&ioctx->cmd, 0); 1741 } else { 1742 pr_err("IB completion has been received too late for wr_id = %u.\n", 1743 ioctx->ioctx.index); 1744 } 1745 1746 srpt_process_wait_list(ch); 1747 } 1748 1749 /** 1750 * srpt_create_ch_ib - create receive and send completion queues 1751 * @ch: SRPT RDMA channel. 1752 */ 1753 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch) 1754 { 1755 struct ib_qp_init_attr *qp_init; 1756 struct srpt_port *sport = ch->sport; 1757 struct srpt_device *sdev = sport->sdev; 1758 const struct ib_device_attr *attrs = &sdev->device->attrs; 1759 int sq_size = sport->port_attrib.srp_sq_size; 1760 int i, ret; 1761 1762 WARN_ON(ch->rq_size < 1); 1763 1764 ret = -ENOMEM; 1765 qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL); 1766 if (!qp_init) 1767 goto out; 1768 1769 retry: 1770 ch->cq = ib_alloc_cq_any(sdev->device, ch, ch->rq_size + sq_size, 1771 IB_POLL_WORKQUEUE); 1772 if (IS_ERR(ch->cq)) { 1773 ret = PTR_ERR(ch->cq); 1774 pr_err("failed to create CQ cqe= %d ret= %d\n", 1775 ch->rq_size + sq_size, ret); 1776 goto out; 1777 } 1778 1779 qp_init->qp_context = (void *)ch; 1780 qp_init->event_handler 1781 = (void(*)(struct ib_event *, void*))srpt_qp_event; 1782 qp_init->send_cq = ch->cq; 1783 qp_init->recv_cq = ch->cq; 1784 qp_init->sq_sig_type = IB_SIGNAL_REQ_WR; 1785 qp_init->qp_type = IB_QPT_RC; 1786 /* 1787 * We divide up our send queue size into half SEND WRs to send the 1788 * completions, and half R/W contexts to actually do the RDMA 1789 * READ/WRITE transfers. Note that we need to allocate CQ slots for 1790 * both both, as RDMA contexts will also post completions for the 1791 * RDMA READ case. 1792 */ 1793 qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr); 1794 qp_init->cap.max_rdma_ctxs = sq_size / 2; 1795 qp_init->cap.max_send_sge = min(attrs->max_send_sge, 1796 SRPT_MAX_SG_PER_WQE); 1797 qp_init->cap.max_recv_sge = min(attrs->max_recv_sge, 1798 SRPT_MAX_SG_PER_WQE); 1799 qp_init->port_num = ch->sport->port; 1800 if (sdev->use_srq) { 1801 qp_init->srq = sdev->srq; 1802 } else { 1803 qp_init->cap.max_recv_wr = ch->rq_size; 1804 qp_init->cap.max_recv_sge = min(attrs->max_recv_sge, 1805 SRPT_MAX_SG_PER_WQE); 1806 } 1807 1808 if (ch->using_rdma_cm) { 1809 ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init); 1810 ch->qp = ch->rdma_cm.cm_id->qp; 1811 } else { 1812 ch->qp = ib_create_qp(sdev->pd, qp_init); 1813 if (!IS_ERR(ch->qp)) { 1814 ret = srpt_init_ch_qp(ch, ch->qp); 1815 if (ret) 1816 ib_destroy_qp(ch->qp); 1817 } else { 1818 ret = PTR_ERR(ch->qp); 1819 } 1820 } 1821 if (ret) { 1822 bool retry = sq_size > MIN_SRPT_SQ_SIZE; 1823 1824 if (retry) { 1825 pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n", 1826 sq_size, ret); 1827 ib_free_cq(ch->cq); 1828 sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE); 1829 goto retry; 1830 } else { 1831 pr_err("failed to create queue pair with sq_size = %d (%d)\n", 1832 sq_size, ret); 1833 goto err_destroy_cq; 1834 } 1835 } 1836 1837 atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr); 1838 1839 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n", 1840 __func__, ch->cq->cqe, qp_init->cap.max_send_sge, 1841 qp_init->cap.max_send_wr, ch); 1842 1843 if (!sdev->use_srq) 1844 for (i = 0; i < ch->rq_size; i++) 1845 srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]); 1846 1847 out: 1848 kfree(qp_init); 1849 return ret; 1850 1851 err_destroy_cq: 1852 ch->qp = NULL; 1853 ib_free_cq(ch->cq); 1854 goto out; 1855 } 1856 1857 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch) 1858 { 1859 ib_destroy_qp(ch->qp); 1860 ib_free_cq(ch->cq); 1861 } 1862 1863 /** 1864 * srpt_close_ch - close a RDMA channel 1865 * @ch: SRPT RDMA channel. 1866 * 1867 * Make sure all resources associated with the channel will be deallocated at 1868 * an appropriate time. 1869 * 1870 * Returns true if and only if the channel state has been modified into 1871 * CH_DRAINING. 1872 */ 1873 static bool srpt_close_ch(struct srpt_rdma_ch *ch) 1874 { 1875 int ret; 1876 1877 if (!srpt_set_ch_state(ch, CH_DRAINING)) { 1878 pr_debug("%s: already closed\n", ch->sess_name); 1879 return false; 1880 } 1881 1882 kref_get(&ch->kref); 1883 1884 ret = srpt_ch_qp_err(ch); 1885 if (ret < 0) 1886 pr_err("%s-%d: changing queue pair into error state failed: %d\n", 1887 ch->sess_name, ch->qp->qp_num, ret); 1888 1889 ret = srpt_zerolength_write(ch); 1890 if (ret < 0) { 1891 pr_err("%s-%d: queuing zero-length write failed: %d\n", 1892 ch->sess_name, ch->qp->qp_num, ret); 1893 if (srpt_set_ch_state(ch, CH_DISCONNECTED)) 1894 schedule_work(&ch->release_work); 1895 else 1896 WARN_ON_ONCE(true); 1897 } 1898 1899 kref_put(&ch->kref, srpt_free_ch); 1900 1901 return true; 1902 } 1903 1904 /* 1905 * Change the channel state into CH_DISCONNECTING. If a channel has not yet 1906 * reached the connected state, close it. If a channel is in the connected 1907 * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is 1908 * the responsibility of the caller to ensure that this function is not 1909 * invoked concurrently with the code that accepts a connection. This means 1910 * that this function must either be invoked from inside a CM callback 1911 * function or that it must be invoked with the srpt_port.mutex held. 1912 */ 1913 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch) 1914 { 1915 int ret; 1916 1917 if (!srpt_set_ch_state(ch, CH_DISCONNECTING)) 1918 return -ENOTCONN; 1919 1920 if (ch->using_rdma_cm) { 1921 ret = rdma_disconnect(ch->rdma_cm.cm_id); 1922 } else { 1923 ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0); 1924 if (ret < 0) 1925 ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0); 1926 } 1927 1928 if (ret < 0 && srpt_close_ch(ch)) 1929 ret = 0; 1930 1931 return ret; 1932 } 1933 1934 static bool srpt_ch_closed(struct srpt_port *sport, struct srpt_rdma_ch *ch) 1935 { 1936 struct srpt_nexus *nexus; 1937 struct srpt_rdma_ch *ch2; 1938 bool res = true; 1939 1940 rcu_read_lock(); 1941 list_for_each_entry(nexus, &sport->nexus_list, entry) { 1942 list_for_each_entry(ch2, &nexus->ch_list, list) { 1943 if (ch2 == ch) { 1944 res = false; 1945 goto done; 1946 } 1947 } 1948 } 1949 done: 1950 rcu_read_unlock(); 1951 1952 return res; 1953 } 1954 1955 /* Send DREQ and wait for DREP. */ 1956 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch) 1957 { 1958 struct srpt_port *sport = ch->sport; 1959 1960 pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num, 1961 ch->state); 1962 1963 mutex_lock(&sport->mutex); 1964 srpt_disconnect_ch(ch); 1965 mutex_unlock(&sport->mutex); 1966 1967 while (wait_event_timeout(sport->ch_releaseQ, srpt_ch_closed(sport, ch), 1968 5 * HZ) == 0) 1969 pr_info("%s(%s-%d state %d): still waiting ...\n", __func__, 1970 ch->sess_name, ch->qp->qp_num, ch->state); 1971 1972 } 1973 1974 static void __srpt_close_all_ch(struct srpt_port *sport) 1975 { 1976 struct srpt_nexus *nexus; 1977 struct srpt_rdma_ch *ch; 1978 1979 lockdep_assert_held(&sport->mutex); 1980 1981 list_for_each_entry(nexus, &sport->nexus_list, entry) { 1982 list_for_each_entry(ch, &nexus->ch_list, list) { 1983 if (srpt_disconnect_ch(ch) >= 0) 1984 pr_info("Closing channel %s because target %s_%d has been disabled\n", 1985 ch->sess_name, 1986 dev_name(&sport->sdev->device->dev), 1987 sport->port); 1988 srpt_close_ch(ch); 1989 } 1990 } 1991 } 1992 1993 /* 1994 * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if 1995 * it does not yet exist. 1996 */ 1997 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport, 1998 const u8 i_port_id[16], 1999 const u8 t_port_id[16]) 2000 { 2001 struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n; 2002 2003 for (;;) { 2004 mutex_lock(&sport->mutex); 2005 list_for_each_entry(n, &sport->nexus_list, entry) { 2006 if (memcmp(n->i_port_id, i_port_id, 16) == 0 && 2007 memcmp(n->t_port_id, t_port_id, 16) == 0) { 2008 nexus = n; 2009 break; 2010 } 2011 } 2012 if (!nexus && tmp_nexus) { 2013 list_add_tail_rcu(&tmp_nexus->entry, 2014 &sport->nexus_list); 2015 swap(nexus, tmp_nexus); 2016 } 2017 mutex_unlock(&sport->mutex); 2018 2019 if (nexus) 2020 break; 2021 tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL); 2022 if (!tmp_nexus) { 2023 nexus = ERR_PTR(-ENOMEM); 2024 break; 2025 } 2026 INIT_LIST_HEAD(&tmp_nexus->ch_list); 2027 memcpy(tmp_nexus->i_port_id, i_port_id, 16); 2028 memcpy(tmp_nexus->t_port_id, t_port_id, 16); 2029 } 2030 2031 kfree(tmp_nexus); 2032 2033 return nexus; 2034 } 2035 2036 static void srpt_set_enabled(struct srpt_port *sport, bool enabled) 2037 __must_hold(&sport->mutex) 2038 { 2039 lockdep_assert_held(&sport->mutex); 2040 2041 if (sport->enabled == enabled) 2042 return; 2043 sport->enabled = enabled; 2044 if (!enabled) 2045 __srpt_close_all_ch(sport); 2046 } 2047 2048 static void srpt_free_ch(struct kref *kref) 2049 { 2050 struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref); 2051 2052 kfree_rcu(ch, rcu); 2053 } 2054 2055 /* 2056 * Shut down the SCSI target session, tell the connection manager to 2057 * disconnect the associated RDMA channel, transition the QP to the error 2058 * state and remove the channel from the channel list. This function is 2059 * typically called from inside srpt_zerolength_write_done(). Concurrent 2060 * srpt_zerolength_write() calls from inside srpt_close_ch() are possible 2061 * as long as the channel is on sport->nexus_list. 2062 */ 2063 static void srpt_release_channel_work(struct work_struct *w) 2064 { 2065 struct srpt_rdma_ch *ch; 2066 struct srpt_device *sdev; 2067 struct srpt_port *sport; 2068 struct se_session *se_sess; 2069 2070 ch = container_of(w, struct srpt_rdma_ch, release_work); 2071 pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num); 2072 2073 sdev = ch->sport->sdev; 2074 BUG_ON(!sdev); 2075 2076 se_sess = ch->sess; 2077 BUG_ON(!se_sess); 2078 2079 target_sess_cmd_list_set_waiting(se_sess); 2080 target_wait_for_sess_cmds(se_sess); 2081 2082 target_remove_session(se_sess); 2083 ch->sess = NULL; 2084 2085 if (ch->using_rdma_cm) 2086 rdma_destroy_id(ch->rdma_cm.cm_id); 2087 else 2088 ib_destroy_cm_id(ch->ib_cm.cm_id); 2089 2090 sport = ch->sport; 2091 mutex_lock(&sport->mutex); 2092 list_del_rcu(&ch->list); 2093 mutex_unlock(&sport->mutex); 2094 2095 srpt_destroy_ch_ib(ch); 2096 2097 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring, 2098 ch->sport->sdev, ch->rq_size, 2099 ch->rsp_buf_cache, DMA_TO_DEVICE); 2100 2101 kmem_cache_destroy(ch->rsp_buf_cache); 2102 2103 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring, 2104 sdev, ch->rq_size, 2105 ch->req_buf_cache, DMA_FROM_DEVICE); 2106 2107 kmem_cache_destroy(ch->req_buf_cache); 2108 2109 wake_up(&sport->ch_releaseQ); 2110 2111 kref_put(&ch->kref, srpt_free_ch); 2112 } 2113 2114 /** 2115 * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED 2116 * @sdev: HCA through which the login request was received. 2117 * @ib_cm_id: IB/CM connection identifier in case of IB/CM. 2118 * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM. 2119 * @port_num: Port through which the REQ message was received. 2120 * @pkey: P_Key of the incoming connection. 2121 * @req: SRP login request. 2122 * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted 2123 * the login request. 2124 * 2125 * Ownership of the cm_id is transferred to the target session if this 2126 * function returns zero. Otherwise the caller remains the owner of cm_id. 2127 */ 2128 static int srpt_cm_req_recv(struct srpt_device *const sdev, 2129 struct ib_cm_id *ib_cm_id, 2130 struct rdma_cm_id *rdma_cm_id, 2131 u8 port_num, __be16 pkey, 2132 const struct srp_login_req *req, 2133 const char *src_addr) 2134 { 2135 struct srpt_port *sport = &sdev->port[port_num - 1]; 2136 struct srpt_nexus *nexus; 2137 struct srp_login_rsp *rsp = NULL; 2138 struct srp_login_rej *rej = NULL; 2139 union { 2140 struct rdma_conn_param rdma_cm; 2141 struct ib_cm_rep_param ib_cm; 2142 } *rep_param = NULL; 2143 struct srpt_rdma_ch *ch = NULL; 2144 char i_port_id[36]; 2145 u32 it_iu_len; 2146 int i, tag_num, tag_size, ret; 2147 2148 WARN_ON_ONCE(irqs_disabled()); 2149 2150 if (WARN_ON(!sdev || !req)) 2151 return -EINVAL; 2152 2153 it_iu_len = be32_to_cpu(req->req_it_iu_len); 2154 2155 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", 2156 req->initiator_port_id, req->target_port_id, it_iu_len, 2157 port_num, &sport->gid, be16_to_cpu(pkey)); 2158 2159 nexus = srpt_get_nexus(sport, req->initiator_port_id, 2160 req->target_port_id); 2161 if (IS_ERR(nexus)) { 2162 ret = PTR_ERR(nexus); 2163 goto out; 2164 } 2165 2166 ret = -ENOMEM; 2167 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL); 2168 rej = kzalloc(sizeof(*rej), GFP_KERNEL); 2169 rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL); 2170 if (!rsp || !rej || !rep_param) 2171 goto out; 2172 2173 ret = -EINVAL; 2174 if (it_iu_len > srp_max_req_size || it_iu_len < 64) { 2175 rej->reason = cpu_to_be32( 2176 SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE); 2177 pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n", 2178 it_iu_len, 64, srp_max_req_size); 2179 goto reject; 2180 } 2181 2182 if (!sport->enabled) { 2183 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2184 pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n", 2185 dev_name(&sport->sdev->device->dev), port_num); 2186 goto reject; 2187 } 2188 2189 if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid) 2190 || *(__be64 *)(req->target_port_id + 8) != 2191 cpu_to_be64(srpt_service_guid)) { 2192 rej->reason = cpu_to_be32( 2193 SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL); 2194 pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n"); 2195 goto reject; 2196 } 2197 2198 ret = -ENOMEM; 2199 ch = kzalloc(sizeof(*ch), GFP_KERNEL); 2200 if (!ch) { 2201 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2202 pr_err("rejected SRP_LOGIN_REQ because out of memory.\n"); 2203 goto reject; 2204 } 2205 2206 kref_init(&ch->kref); 2207 ch->pkey = be16_to_cpu(pkey); 2208 ch->nexus = nexus; 2209 ch->zw_cqe.done = srpt_zerolength_write_done; 2210 INIT_WORK(&ch->release_work, srpt_release_channel_work); 2211 ch->sport = sport; 2212 if (ib_cm_id) { 2213 ch->ib_cm.cm_id = ib_cm_id; 2214 ib_cm_id->context = ch; 2215 } else { 2216 ch->using_rdma_cm = true; 2217 ch->rdma_cm.cm_id = rdma_cm_id; 2218 rdma_cm_id->context = ch; 2219 } 2220 /* 2221 * ch->rq_size should be at least as large as the initiator queue 2222 * depth to avoid that the initiator driver has to report QUEUE_FULL 2223 * to the SCSI mid-layer. 2224 */ 2225 ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr); 2226 spin_lock_init(&ch->spinlock); 2227 ch->state = CH_CONNECTING; 2228 INIT_LIST_HEAD(&ch->cmd_wait_list); 2229 ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size; 2230 2231 ch->rsp_buf_cache = kmem_cache_create("srpt-rsp-buf", ch->max_rsp_size, 2232 512, 0, NULL); 2233 if (!ch->rsp_buf_cache) 2234 goto free_ch; 2235 2236 ch->ioctx_ring = (struct srpt_send_ioctx **) 2237 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size, 2238 sizeof(*ch->ioctx_ring[0]), 2239 ch->rsp_buf_cache, 0, DMA_TO_DEVICE); 2240 if (!ch->ioctx_ring) { 2241 pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n"); 2242 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2243 goto free_rsp_cache; 2244 } 2245 2246 for (i = 0; i < ch->rq_size; i++) 2247 ch->ioctx_ring[i]->ch = ch; 2248 if (!sdev->use_srq) { 2249 u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ? 2250 be16_to_cpu(req->imm_data_offset) : 0; 2251 u16 alignment_offset; 2252 u32 req_sz; 2253 2254 if (req->req_flags & SRP_IMMED_REQUESTED) 2255 pr_debug("imm_data_offset = %d\n", 2256 be16_to_cpu(req->imm_data_offset)); 2257 if (imm_data_offset >= sizeof(struct srp_cmd)) { 2258 ch->imm_data_offset = imm_data_offset; 2259 rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP; 2260 } else { 2261 ch->imm_data_offset = 0; 2262 } 2263 alignment_offset = round_up(imm_data_offset, 512) - 2264 imm_data_offset; 2265 req_sz = alignment_offset + imm_data_offset + srp_max_req_size; 2266 ch->req_buf_cache = kmem_cache_create("srpt-req-buf", req_sz, 2267 512, 0, NULL); 2268 if (!ch->req_buf_cache) 2269 goto free_rsp_ring; 2270 2271 ch->ioctx_recv_ring = (struct srpt_recv_ioctx **) 2272 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size, 2273 sizeof(*ch->ioctx_recv_ring[0]), 2274 ch->req_buf_cache, 2275 alignment_offset, 2276 DMA_FROM_DEVICE); 2277 if (!ch->ioctx_recv_ring) { 2278 pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n"); 2279 rej->reason = 2280 cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2281 goto free_recv_cache; 2282 } 2283 for (i = 0; i < ch->rq_size; i++) 2284 INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list); 2285 } 2286 2287 ret = srpt_create_ch_ib(ch); 2288 if (ret) { 2289 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2290 pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n"); 2291 goto free_recv_ring; 2292 } 2293 2294 strlcpy(ch->sess_name, src_addr, sizeof(ch->sess_name)); 2295 snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx", 2296 be64_to_cpu(*(__be64 *)nexus->i_port_id), 2297 be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8))); 2298 2299 pr_debug("registering session %s\n", ch->sess_name); 2300 2301 tag_num = ch->rq_size; 2302 tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */ 2303 if (sport->port_guid_tpg.se_tpg_wwn) 2304 ch->sess = target_setup_session(&sport->port_guid_tpg, tag_num, 2305 tag_size, TARGET_PROT_NORMAL, 2306 ch->sess_name, ch, NULL); 2307 if (sport->port_gid_tpg.se_tpg_wwn && IS_ERR_OR_NULL(ch->sess)) 2308 ch->sess = target_setup_session(&sport->port_gid_tpg, tag_num, 2309 tag_size, TARGET_PROT_NORMAL, i_port_id, 2310 ch, NULL); 2311 /* Retry without leading "0x" */ 2312 if (sport->port_gid_tpg.se_tpg_wwn && IS_ERR_OR_NULL(ch->sess)) 2313 ch->sess = target_setup_session(&sport->port_gid_tpg, tag_num, 2314 tag_size, TARGET_PROT_NORMAL, 2315 i_port_id + 2, ch, NULL); 2316 if (IS_ERR_OR_NULL(ch->sess)) { 2317 WARN_ON_ONCE(ch->sess == NULL); 2318 ret = PTR_ERR(ch->sess); 2319 ch->sess = NULL; 2320 pr_info("Rejected login for initiator %s: ret = %d.\n", 2321 ch->sess_name, ret); 2322 rej->reason = cpu_to_be32(ret == -ENOMEM ? 2323 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES : 2324 SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED); 2325 goto destroy_ib; 2326 } 2327 2328 mutex_lock(&sport->mutex); 2329 2330 if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) { 2331 struct srpt_rdma_ch *ch2; 2332 2333 list_for_each_entry(ch2, &nexus->ch_list, list) { 2334 if (srpt_disconnect_ch(ch2) < 0) 2335 continue; 2336 pr_info("Relogin - closed existing channel %s\n", 2337 ch2->sess_name); 2338 rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED; 2339 } 2340 } else { 2341 rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED; 2342 } 2343 2344 list_add_tail_rcu(&ch->list, &nexus->ch_list); 2345 2346 if (!sport->enabled) { 2347 rej->reason = cpu_to_be32( 2348 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2349 pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n", 2350 dev_name(&sdev->device->dev), port_num); 2351 mutex_unlock(&sport->mutex); 2352 goto reject; 2353 } 2354 2355 mutex_unlock(&sport->mutex); 2356 2357 ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp); 2358 if (ret) { 2359 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2360 pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n", 2361 ret); 2362 goto reject; 2363 } 2364 2365 pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess, 2366 ch->sess_name, ch); 2367 2368 /* create srp_login_response */ 2369 rsp->opcode = SRP_LOGIN_RSP; 2370 rsp->tag = req->tag; 2371 rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size); 2372 rsp->max_ti_iu_len = req->req_it_iu_len; 2373 ch->max_ti_iu_len = it_iu_len; 2374 rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT | 2375 SRP_BUF_FORMAT_INDIRECT); 2376 rsp->req_lim_delta = cpu_to_be32(ch->rq_size); 2377 atomic_set(&ch->req_lim, ch->rq_size); 2378 atomic_set(&ch->req_lim_delta, 0); 2379 2380 /* create cm reply */ 2381 if (ch->using_rdma_cm) { 2382 rep_param->rdma_cm.private_data = (void *)rsp; 2383 rep_param->rdma_cm.private_data_len = sizeof(*rsp); 2384 rep_param->rdma_cm.rnr_retry_count = 7; 2385 rep_param->rdma_cm.flow_control = 1; 2386 rep_param->rdma_cm.responder_resources = 4; 2387 rep_param->rdma_cm.initiator_depth = 4; 2388 } else { 2389 rep_param->ib_cm.qp_num = ch->qp->qp_num; 2390 rep_param->ib_cm.private_data = (void *)rsp; 2391 rep_param->ib_cm.private_data_len = sizeof(*rsp); 2392 rep_param->ib_cm.rnr_retry_count = 7; 2393 rep_param->ib_cm.flow_control = 1; 2394 rep_param->ib_cm.failover_accepted = 0; 2395 rep_param->ib_cm.srq = 1; 2396 rep_param->ib_cm.responder_resources = 4; 2397 rep_param->ib_cm.initiator_depth = 4; 2398 } 2399 2400 /* 2401 * Hold the sport mutex while accepting a connection to avoid that 2402 * srpt_disconnect_ch() is invoked concurrently with this code. 2403 */ 2404 mutex_lock(&sport->mutex); 2405 if (sport->enabled && ch->state == CH_CONNECTING) { 2406 if (ch->using_rdma_cm) 2407 ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm); 2408 else 2409 ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm); 2410 } else { 2411 ret = -EINVAL; 2412 } 2413 mutex_unlock(&sport->mutex); 2414 2415 switch (ret) { 2416 case 0: 2417 break; 2418 case -EINVAL: 2419 goto reject; 2420 default: 2421 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2422 pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n", 2423 ret); 2424 goto reject; 2425 } 2426 2427 goto out; 2428 2429 destroy_ib: 2430 srpt_destroy_ch_ib(ch); 2431 2432 free_recv_ring: 2433 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring, 2434 ch->sport->sdev, ch->rq_size, 2435 ch->req_buf_cache, DMA_FROM_DEVICE); 2436 2437 free_recv_cache: 2438 kmem_cache_destroy(ch->req_buf_cache); 2439 2440 free_rsp_ring: 2441 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring, 2442 ch->sport->sdev, ch->rq_size, 2443 ch->rsp_buf_cache, DMA_TO_DEVICE); 2444 2445 free_rsp_cache: 2446 kmem_cache_destroy(ch->rsp_buf_cache); 2447 2448 free_ch: 2449 if (rdma_cm_id) 2450 rdma_cm_id->context = NULL; 2451 else 2452 ib_cm_id->context = NULL; 2453 kfree(ch); 2454 ch = NULL; 2455 2456 WARN_ON_ONCE(ret == 0); 2457 2458 reject: 2459 pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason)); 2460 rej->opcode = SRP_LOGIN_REJ; 2461 rej->tag = req->tag; 2462 rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT | 2463 SRP_BUF_FORMAT_INDIRECT); 2464 2465 if (rdma_cm_id) 2466 rdma_reject(rdma_cm_id, rej, sizeof(*rej)); 2467 else 2468 ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0, 2469 rej, sizeof(*rej)); 2470 2471 if (ch && ch->sess) { 2472 srpt_close_ch(ch); 2473 /* 2474 * Tell the caller not to free cm_id since 2475 * srpt_release_channel_work() will do that. 2476 */ 2477 ret = 0; 2478 } 2479 2480 out: 2481 kfree(rep_param); 2482 kfree(rsp); 2483 kfree(rej); 2484 2485 return ret; 2486 } 2487 2488 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id, 2489 const struct ib_cm_req_event_param *param, 2490 void *private_data) 2491 { 2492 char sguid[40]; 2493 2494 srpt_format_guid(sguid, sizeof(sguid), 2495 ¶m->primary_path->dgid.global.interface_id); 2496 2497 return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port, 2498 param->primary_path->pkey, 2499 private_data, sguid); 2500 } 2501 2502 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id, 2503 struct rdma_cm_event *event) 2504 { 2505 struct srpt_device *sdev; 2506 struct srp_login_req req; 2507 const struct srp_login_req_rdma *req_rdma; 2508 char src_addr[40]; 2509 2510 sdev = ib_get_client_data(cm_id->device, &srpt_client); 2511 if (!sdev) 2512 return -ECONNREFUSED; 2513 2514 if (event->param.conn.private_data_len < sizeof(*req_rdma)) 2515 return -EINVAL; 2516 2517 /* Transform srp_login_req_rdma into srp_login_req. */ 2518 req_rdma = event->param.conn.private_data; 2519 memset(&req, 0, sizeof(req)); 2520 req.opcode = req_rdma->opcode; 2521 req.tag = req_rdma->tag; 2522 req.req_it_iu_len = req_rdma->req_it_iu_len; 2523 req.req_buf_fmt = req_rdma->req_buf_fmt; 2524 req.req_flags = req_rdma->req_flags; 2525 memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16); 2526 memcpy(req.target_port_id, req_rdma->target_port_id, 16); 2527 req.imm_data_offset = req_rdma->imm_data_offset; 2528 2529 snprintf(src_addr, sizeof(src_addr), "%pIS", 2530 &cm_id->route.addr.src_addr); 2531 2532 return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num, 2533 cm_id->route.path_rec->pkey, &req, src_addr); 2534 } 2535 2536 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch, 2537 enum ib_cm_rej_reason reason, 2538 const u8 *private_data, 2539 u8 private_data_len) 2540 { 2541 char *priv = NULL; 2542 int i; 2543 2544 if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1, 2545 GFP_KERNEL))) { 2546 for (i = 0; i < private_data_len; i++) 2547 sprintf(priv + 3 * i, " %02x", private_data[i]); 2548 } 2549 pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n", 2550 ch->sess_name, ch->qp->qp_num, reason, private_data_len ? 2551 "; private data" : "", priv ? priv : " (?)"); 2552 kfree(priv); 2553 } 2554 2555 /** 2556 * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event 2557 * @ch: SRPT RDMA channel. 2558 * 2559 * An RTU (ready to use) message indicates that the connection has been 2560 * established and that the recipient may begin transmitting. 2561 */ 2562 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch) 2563 { 2564 int ret; 2565 2566 ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp); 2567 if (ret < 0) { 2568 pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name, 2569 ch->qp->qp_num); 2570 srpt_close_ch(ch); 2571 return; 2572 } 2573 2574 /* 2575 * Note: calling srpt_close_ch() if the transition to the LIVE state 2576 * fails is not necessary since that means that that function has 2577 * already been invoked from another thread. 2578 */ 2579 if (!srpt_set_ch_state(ch, CH_LIVE)) { 2580 pr_err("%s-%d: channel transition to LIVE state failed\n", 2581 ch->sess_name, ch->qp->qp_num); 2582 return; 2583 } 2584 2585 /* Trigger wait list processing. */ 2586 ret = srpt_zerolength_write(ch); 2587 WARN_ONCE(ret < 0, "%d\n", ret); 2588 } 2589 2590 /** 2591 * srpt_cm_handler - IB connection manager callback function 2592 * @cm_id: IB/CM connection identifier. 2593 * @event: IB/CM event. 2594 * 2595 * A non-zero return value will cause the caller destroy the CM ID. 2596 * 2597 * Note: srpt_cm_handler() must only return a non-zero value when transferring 2598 * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning 2599 * a non-zero value in any other case will trigger a race with the 2600 * ib_destroy_cm_id() call in srpt_release_channel(). 2601 */ 2602 static int srpt_cm_handler(struct ib_cm_id *cm_id, 2603 const struct ib_cm_event *event) 2604 { 2605 struct srpt_rdma_ch *ch = cm_id->context; 2606 int ret; 2607 2608 ret = 0; 2609 switch (event->event) { 2610 case IB_CM_REQ_RECEIVED: 2611 ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd, 2612 event->private_data); 2613 break; 2614 case IB_CM_REJ_RECEIVED: 2615 srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason, 2616 event->private_data, 2617 IB_CM_REJ_PRIVATE_DATA_SIZE); 2618 break; 2619 case IB_CM_RTU_RECEIVED: 2620 case IB_CM_USER_ESTABLISHED: 2621 srpt_cm_rtu_recv(ch); 2622 break; 2623 case IB_CM_DREQ_RECEIVED: 2624 srpt_disconnect_ch(ch); 2625 break; 2626 case IB_CM_DREP_RECEIVED: 2627 pr_info("Received CM DREP message for ch %s-%d.\n", 2628 ch->sess_name, ch->qp->qp_num); 2629 srpt_close_ch(ch); 2630 break; 2631 case IB_CM_TIMEWAIT_EXIT: 2632 pr_info("Received CM TimeWait exit for ch %s-%d.\n", 2633 ch->sess_name, ch->qp->qp_num); 2634 srpt_close_ch(ch); 2635 break; 2636 case IB_CM_REP_ERROR: 2637 pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name, 2638 ch->qp->qp_num); 2639 break; 2640 case IB_CM_DREQ_ERROR: 2641 pr_info("Received CM DREQ ERROR event.\n"); 2642 break; 2643 case IB_CM_MRA_RECEIVED: 2644 pr_info("Received CM MRA event\n"); 2645 break; 2646 default: 2647 pr_err("received unrecognized CM event %d\n", event->event); 2648 break; 2649 } 2650 2651 return ret; 2652 } 2653 2654 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id, 2655 struct rdma_cm_event *event) 2656 { 2657 struct srpt_rdma_ch *ch = cm_id->context; 2658 int ret = 0; 2659 2660 switch (event->event) { 2661 case RDMA_CM_EVENT_CONNECT_REQUEST: 2662 ret = srpt_rdma_cm_req_recv(cm_id, event); 2663 break; 2664 case RDMA_CM_EVENT_REJECTED: 2665 srpt_cm_rej_recv(ch, event->status, 2666 event->param.conn.private_data, 2667 event->param.conn.private_data_len); 2668 break; 2669 case RDMA_CM_EVENT_ESTABLISHED: 2670 srpt_cm_rtu_recv(ch); 2671 break; 2672 case RDMA_CM_EVENT_DISCONNECTED: 2673 if (ch->state < CH_DISCONNECTING) 2674 srpt_disconnect_ch(ch); 2675 else 2676 srpt_close_ch(ch); 2677 break; 2678 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2679 srpt_close_ch(ch); 2680 break; 2681 case RDMA_CM_EVENT_UNREACHABLE: 2682 pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name, 2683 ch->qp->qp_num); 2684 break; 2685 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2686 case RDMA_CM_EVENT_ADDR_CHANGE: 2687 break; 2688 default: 2689 pr_err("received unrecognized RDMA CM event %d\n", 2690 event->event); 2691 break; 2692 } 2693 2694 return ret; 2695 } 2696 2697 /* 2698 * srpt_write_pending - Start data transfer from initiator to target (write). 2699 */ 2700 static int srpt_write_pending(struct se_cmd *se_cmd) 2701 { 2702 struct srpt_send_ioctx *ioctx = 2703 container_of(se_cmd, struct srpt_send_ioctx, cmd); 2704 struct srpt_rdma_ch *ch = ioctx->ch; 2705 struct ib_send_wr *first_wr = NULL; 2706 struct ib_cqe *cqe = &ioctx->rdma_cqe; 2707 enum srpt_command_state new_state; 2708 int ret, i; 2709 2710 if (ioctx->recv_ioctx) { 2711 srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN); 2712 target_execute_cmd(&ioctx->cmd); 2713 return 0; 2714 } 2715 2716 new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA); 2717 WARN_ON(new_state == SRPT_STATE_DONE); 2718 2719 if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) { 2720 pr_warn("%s: IB send queue full (needed %d)\n", 2721 __func__, ioctx->n_rdma); 2722 ret = -ENOMEM; 2723 goto out_undo; 2724 } 2725 2726 cqe->done = srpt_rdma_read_done; 2727 for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) { 2728 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 2729 2730 first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port, 2731 cqe, first_wr); 2732 cqe = NULL; 2733 } 2734 2735 ret = ib_post_send(ch->qp, first_wr, NULL); 2736 if (ret) { 2737 pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n", 2738 __func__, ret, ioctx->n_rdma, 2739 atomic_read(&ch->sq_wr_avail)); 2740 goto out_undo; 2741 } 2742 2743 return 0; 2744 out_undo: 2745 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail); 2746 return ret; 2747 } 2748 2749 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status) 2750 { 2751 switch (tcm_mgmt_status) { 2752 case TMR_FUNCTION_COMPLETE: 2753 return SRP_TSK_MGMT_SUCCESS; 2754 case TMR_FUNCTION_REJECTED: 2755 return SRP_TSK_MGMT_FUNC_NOT_SUPP; 2756 } 2757 return SRP_TSK_MGMT_FAILED; 2758 } 2759 2760 /** 2761 * srpt_queue_response - transmit the response to a SCSI command 2762 * @cmd: SCSI target command. 2763 * 2764 * Callback function called by the TCM core. Must not block since it can be 2765 * invoked on the context of the IB completion handler. 2766 */ 2767 static void srpt_queue_response(struct se_cmd *cmd) 2768 { 2769 struct srpt_send_ioctx *ioctx = 2770 container_of(cmd, struct srpt_send_ioctx, cmd); 2771 struct srpt_rdma_ch *ch = ioctx->ch; 2772 struct srpt_device *sdev = ch->sport->sdev; 2773 struct ib_send_wr send_wr, *first_wr = &send_wr; 2774 struct ib_sge sge; 2775 enum srpt_command_state state; 2776 int resp_len, ret, i; 2777 u8 srp_tm_status; 2778 2779 BUG_ON(!ch); 2780 2781 state = ioctx->state; 2782 switch (state) { 2783 case SRPT_STATE_NEW: 2784 case SRPT_STATE_DATA_IN: 2785 ioctx->state = SRPT_STATE_CMD_RSP_SENT; 2786 break; 2787 case SRPT_STATE_MGMT: 2788 ioctx->state = SRPT_STATE_MGMT_RSP_SENT; 2789 break; 2790 default: 2791 WARN(true, "ch %p; cmd %d: unexpected command state %d\n", 2792 ch, ioctx->ioctx.index, ioctx->state); 2793 break; 2794 } 2795 2796 if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT)) 2797 return; 2798 2799 /* For read commands, transfer the data to the initiator. */ 2800 if (ioctx->cmd.data_direction == DMA_FROM_DEVICE && 2801 ioctx->cmd.data_length && 2802 !ioctx->queue_status_only) { 2803 for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) { 2804 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 2805 2806 first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, 2807 ch->sport->port, NULL, first_wr); 2808 } 2809 } 2810 2811 if (state != SRPT_STATE_MGMT) 2812 resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag, 2813 cmd->scsi_status); 2814 else { 2815 srp_tm_status 2816 = tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response); 2817 resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status, 2818 ioctx->cmd.tag); 2819 } 2820 2821 atomic_inc(&ch->req_lim); 2822 2823 if (unlikely(atomic_sub_return(1 + ioctx->n_rdma, 2824 &ch->sq_wr_avail) < 0)) { 2825 pr_warn("%s: IB send queue full (needed %d)\n", 2826 __func__, ioctx->n_rdma); 2827 ret = -ENOMEM; 2828 goto out; 2829 } 2830 2831 ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len, 2832 DMA_TO_DEVICE); 2833 2834 sge.addr = ioctx->ioctx.dma; 2835 sge.length = resp_len; 2836 sge.lkey = sdev->lkey; 2837 2838 ioctx->ioctx.cqe.done = srpt_send_done; 2839 send_wr.next = NULL; 2840 send_wr.wr_cqe = &ioctx->ioctx.cqe; 2841 send_wr.sg_list = &sge; 2842 send_wr.num_sge = 1; 2843 send_wr.opcode = IB_WR_SEND; 2844 send_wr.send_flags = IB_SEND_SIGNALED; 2845 2846 ret = ib_post_send(ch->qp, first_wr, NULL); 2847 if (ret < 0) { 2848 pr_err("%s: sending cmd response failed for tag %llu (%d)\n", 2849 __func__, ioctx->cmd.tag, ret); 2850 goto out; 2851 } 2852 2853 return; 2854 2855 out: 2856 atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail); 2857 atomic_dec(&ch->req_lim); 2858 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE); 2859 target_put_sess_cmd(&ioctx->cmd); 2860 } 2861 2862 static int srpt_queue_data_in(struct se_cmd *cmd) 2863 { 2864 srpt_queue_response(cmd); 2865 return 0; 2866 } 2867 2868 static void srpt_queue_tm_rsp(struct se_cmd *cmd) 2869 { 2870 srpt_queue_response(cmd); 2871 } 2872 2873 /* 2874 * This function is called for aborted commands if no response is sent to the 2875 * initiator. Make sure that the credits freed by aborting a command are 2876 * returned to the initiator the next time a response is sent by incrementing 2877 * ch->req_lim_delta. 2878 */ 2879 static void srpt_aborted_task(struct se_cmd *cmd) 2880 { 2881 struct srpt_send_ioctx *ioctx = container_of(cmd, 2882 struct srpt_send_ioctx, cmd); 2883 struct srpt_rdma_ch *ch = ioctx->ch; 2884 2885 atomic_inc(&ch->req_lim_delta); 2886 } 2887 2888 static int srpt_queue_status(struct se_cmd *cmd) 2889 { 2890 struct srpt_send_ioctx *ioctx; 2891 2892 ioctx = container_of(cmd, struct srpt_send_ioctx, cmd); 2893 BUG_ON(ioctx->sense_data != cmd->sense_buffer); 2894 if (cmd->se_cmd_flags & 2895 (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE)) 2896 WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION); 2897 ioctx->queue_status_only = true; 2898 srpt_queue_response(cmd); 2899 return 0; 2900 } 2901 2902 static void srpt_refresh_port_work(struct work_struct *work) 2903 { 2904 struct srpt_port *sport = container_of(work, struct srpt_port, work); 2905 2906 srpt_refresh_port(sport); 2907 } 2908 2909 static bool srpt_ch_list_empty(struct srpt_port *sport) 2910 { 2911 struct srpt_nexus *nexus; 2912 bool res = true; 2913 2914 rcu_read_lock(); 2915 list_for_each_entry(nexus, &sport->nexus_list, entry) 2916 if (!list_empty(&nexus->ch_list)) 2917 res = false; 2918 rcu_read_unlock(); 2919 2920 return res; 2921 } 2922 2923 /** 2924 * srpt_release_sport - disable login and wait for associated channels 2925 * @sport: SRPT HCA port. 2926 */ 2927 static int srpt_release_sport(struct srpt_port *sport) 2928 { 2929 struct srpt_nexus *nexus, *next_n; 2930 struct srpt_rdma_ch *ch; 2931 2932 WARN_ON_ONCE(irqs_disabled()); 2933 2934 mutex_lock(&sport->mutex); 2935 srpt_set_enabled(sport, false); 2936 mutex_unlock(&sport->mutex); 2937 2938 while (wait_event_timeout(sport->ch_releaseQ, 2939 srpt_ch_list_empty(sport), 5 * HZ) <= 0) { 2940 pr_info("%s_%d: waiting for session unregistration ...\n", 2941 dev_name(&sport->sdev->device->dev), sport->port); 2942 rcu_read_lock(); 2943 list_for_each_entry(nexus, &sport->nexus_list, entry) { 2944 list_for_each_entry(ch, &nexus->ch_list, list) { 2945 pr_info("%s-%d: state %s\n", 2946 ch->sess_name, ch->qp->qp_num, 2947 get_ch_state_name(ch->state)); 2948 } 2949 } 2950 rcu_read_unlock(); 2951 } 2952 2953 mutex_lock(&sport->mutex); 2954 list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) { 2955 list_del(&nexus->entry); 2956 kfree_rcu(nexus, rcu); 2957 } 2958 mutex_unlock(&sport->mutex); 2959 2960 return 0; 2961 } 2962 2963 static struct se_wwn *__srpt_lookup_wwn(const char *name) 2964 { 2965 struct ib_device *dev; 2966 struct srpt_device *sdev; 2967 struct srpt_port *sport; 2968 int i; 2969 2970 list_for_each_entry(sdev, &srpt_dev_list, list) { 2971 dev = sdev->device; 2972 if (!dev) 2973 continue; 2974 2975 for (i = 0; i < dev->phys_port_cnt; i++) { 2976 sport = &sdev->port[i]; 2977 2978 if (strcmp(sport->port_guid, name) == 0) 2979 return &sport->port_guid_wwn; 2980 if (strcmp(sport->port_gid, name) == 0) 2981 return &sport->port_gid_wwn; 2982 } 2983 } 2984 2985 return NULL; 2986 } 2987 2988 static struct se_wwn *srpt_lookup_wwn(const char *name) 2989 { 2990 struct se_wwn *wwn; 2991 2992 spin_lock(&srpt_dev_lock); 2993 wwn = __srpt_lookup_wwn(name); 2994 spin_unlock(&srpt_dev_lock); 2995 2996 return wwn; 2997 } 2998 2999 static void srpt_free_srq(struct srpt_device *sdev) 3000 { 3001 if (!sdev->srq) 3002 return; 3003 3004 ib_destroy_srq(sdev->srq); 3005 srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev, 3006 sdev->srq_size, sdev->req_buf_cache, 3007 DMA_FROM_DEVICE); 3008 kmem_cache_destroy(sdev->req_buf_cache); 3009 sdev->srq = NULL; 3010 } 3011 3012 static int srpt_alloc_srq(struct srpt_device *sdev) 3013 { 3014 struct ib_srq_init_attr srq_attr = { 3015 .event_handler = srpt_srq_event, 3016 .srq_context = (void *)sdev, 3017 .attr.max_wr = sdev->srq_size, 3018 .attr.max_sge = 1, 3019 .srq_type = IB_SRQT_BASIC, 3020 }; 3021 struct ib_device *device = sdev->device; 3022 struct ib_srq *srq; 3023 int i; 3024 3025 WARN_ON_ONCE(sdev->srq); 3026 srq = ib_create_srq(sdev->pd, &srq_attr); 3027 if (IS_ERR(srq)) { 3028 pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq)); 3029 return PTR_ERR(srq); 3030 } 3031 3032 pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size, 3033 sdev->device->attrs.max_srq_wr, dev_name(&device->dev)); 3034 3035 sdev->req_buf_cache = kmem_cache_create("srpt-srq-req-buf", 3036 srp_max_req_size, 0, 0, NULL); 3037 if (!sdev->req_buf_cache) 3038 goto free_srq; 3039 3040 sdev->ioctx_ring = (struct srpt_recv_ioctx **) 3041 srpt_alloc_ioctx_ring(sdev, sdev->srq_size, 3042 sizeof(*sdev->ioctx_ring[0]), 3043 sdev->req_buf_cache, 0, DMA_FROM_DEVICE); 3044 if (!sdev->ioctx_ring) 3045 goto free_cache; 3046 3047 sdev->use_srq = true; 3048 sdev->srq = srq; 3049 3050 for (i = 0; i < sdev->srq_size; ++i) { 3051 INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list); 3052 srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]); 3053 } 3054 3055 return 0; 3056 3057 free_cache: 3058 kmem_cache_destroy(sdev->req_buf_cache); 3059 3060 free_srq: 3061 ib_destroy_srq(srq); 3062 return -ENOMEM; 3063 } 3064 3065 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq) 3066 { 3067 struct ib_device *device = sdev->device; 3068 int ret = 0; 3069 3070 if (!use_srq) { 3071 srpt_free_srq(sdev); 3072 sdev->use_srq = false; 3073 } else if (use_srq && !sdev->srq) { 3074 ret = srpt_alloc_srq(sdev); 3075 } 3076 pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__, 3077 dev_name(&device->dev), sdev->use_srq, ret); 3078 return ret; 3079 } 3080 3081 /** 3082 * srpt_add_one - InfiniBand device addition callback function 3083 * @device: Describes a HCA. 3084 */ 3085 static void srpt_add_one(struct ib_device *device) 3086 { 3087 struct srpt_device *sdev; 3088 struct srpt_port *sport; 3089 int i, ret; 3090 3091 pr_debug("device = %p\n", device); 3092 3093 sdev = kzalloc(struct_size(sdev, port, device->phys_port_cnt), 3094 GFP_KERNEL); 3095 if (!sdev) 3096 goto err; 3097 3098 sdev->device = device; 3099 mutex_init(&sdev->sdev_mutex); 3100 3101 sdev->pd = ib_alloc_pd(device, 0); 3102 if (IS_ERR(sdev->pd)) 3103 goto free_dev; 3104 3105 sdev->lkey = sdev->pd->local_dma_lkey; 3106 3107 sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr); 3108 3109 srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq); 3110 3111 if (!srpt_service_guid) 3112 srpt_service_guid = be64_to_cpu(device->node_guid); 3113 3114 if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND) 3115 sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev); 3116 if (IS_ERR(sdev->cm_id)) { 3117 pr_info("ib_create_cm_id() failed: %ld\n", 3118 PTR_ERR(sdev->cm_id)); 3119 sdev->cm_id = NULL; 3120 if (!rdma_cm_id) 3121 goto err_ring; 3122 } 3123 3124 /* print out target login information */ 3125 pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n", 3126 srpt_service_guid, srpt_service_guid, srpt_service_guid); 3127 3128 /* 3129 * We do not have a consistent service_id (ie. also id_ext of target_id) 3130 * to identify this target. We currently use the guid of the first HCA 3131 * in the system as service_id; therefore, the target_id will change 3132 * if this HCA is gone bad and replaced by different HCA 3133 */ 3134 ret = sdev->cm_id ? 3135 ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0) : 3136 0; 3137 if (ret < 0) { 3138 pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret, 3139 sdev->cm_id->state); 3140 goto err_cm; 3141 } 3142 3143 INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device, 3144 srpt_event_handler); 3145 ib_register_event_handler(&sdev->event_handler); 3146 3147 for (i = 1; i <= sdev->device->phys_port_cnt; i++) { 3148 sport = &sdev->port[i - 1]; 3149 INIT_LIST_HEAD(&sport->nexus_list); 3150 init_waitqueue_head(&sport->ch_releaseQ); 3151 mutex_init(&sport->mutex); 3152 sport->sdev = sdev; 3153 sport->port = i; 3154 sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE; 3155 sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE; 3156 sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE; 3157 sport->port_attrib.use_srq = false; 3158 INIT_WORK(&sport->work, srpt_refresh_port_work); 3159 3160 if (srpt_refresh_port(sport)) { 3161 pr_err("MAD registration failed for %s-%d.\n", 3162 dev_name(&sdev->device->dev), i); 3163 goto err_event; 3164 } 3165 } 3166 3167 spin_lock(&srpt_dev_lock); 3168 list_add_tail(&sdev->list, &srpt_dev_list); 3169 spin_unlock(&srpt_dev_lock); 3170 3171 out: 3172 ib_set_client_data(device, &srpt_client, sdev); 3173 pr_debug("added %s.\n", dev_name(&device->dev)); 3174 return; 3175 3176 err_event: 3177 ib_unregister_event_handler(&sdev->event_handler); 3178 err_cm: 3179 if (sdev->cm_id) 3180 ib_destroy_cm_id(sdev->cm_id); 3181 err_ring: 3182 srpt_free_srq(sdev); 3183 ib_dealloc_pd(sdev->pd); 3184 free_dev: 3185 kfree(sdev); 3186 err: 3187 sdev = NULL; 3188 pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev)); 3189 goto out; 3190 } 3191 3192 /** 3193 * srpt_remove_one - InfiniBand device removal callback function 3194 * @device: Describes a HCA. 3195 * @client_data: The value passed as the third argument to ib_set_client_data(). 3196 */ 3197 static void srpt_remove_one(struct ib_device *device, void *client_data) 3198 { 3199 struct srpt_device *sdev = client_data; 3200 int i; 3201 3202 if (!sdev) { 3203 pr_info("%s(%s): nothing to do.\n", __func__, 3204 dev_name(&device->dev)); 3205 return; 3206 } 3207 3208 srpt_unregister_mad_agent(sdev); 3209 3210 ib_unregister_event_handler(&sdev->event_handler); 3211 3212 /* Cancel any work queued by the just unregistered IB event handler. */ 3213 for (i = 0; i < sdev->device->phys_port_cnt; i++) 3214 cancel_work_sync(&sdev->port[i].work); 3215 3216 if (sdev->cm_id) 3217 ib_destroy_cm_id(sdev->cm_id); 3218 3219 ib_set_client_data(device, &srpt_client, NULL); 3220 3221 /* 3222 * Unregistering a target must happen after destroying sdev->cm_id 3223 * such that no new SRP_LOGIN_REQ information units can arrive while 3224 * destroying the target. 3225 */ 3226 spin_lock(&srpt_dev_lock); 3227 list_del(&sdev->list); 3228 spin_unlock(&srpt_dev_lock); 3229 3230 for (i = 0; i < sdev->device->phys_port_cnt; i++) 3231 srpt_release_sport(&sdev->port[i]); 3232 3233 srpt_free_srq(sdev); 3234 3235 ib_dealloc_pd(sdev->pd); 3236 3237 kfree(sdev); 3238 } 3239 3240 static struct ib_client srpt_client = { 3241 .name = DRV_NAME, 3242 .add = srpt_add_one, 3243 .remove = srpt_remove_one 3244 }; 3245 3246 static int srpt_check_true(struct se_portal_group *se_tpg) 3247 { 3248 return 1; 3249 } 3250 3251 static int srpt_check_false(struct se_portal_group *se_tpg) 3252 { 3253 return 0; 3254 } 3255 3256 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg) 3257 { 3258 return tpg->se_tpg_wwn->priv; 3259 } 3260 3261 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg) 3262 { 3263 struct srpt_port *sport = srpt_tpg_to_sport(tpg); 3264 3265 WARN_ON_ONCE(tpg != &sport->port_guid_tpg && 3266 tpg != &sport->port_gid_tpg); 3267 return tpg == &sport->port_guid_tpg ? sport->port_guid : 3268 sport->port_gid; 3269 } 3270 3271 static u16 srpt_get_tag(struct se_portal_group *tpg) 3272 { 3273 return 1; 3274 } 3275 3276 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg) 3277 { 3278 return 1; 3279 } 3280 3281 static void srpt_release_cmd(struct se_cmd *se_cmd) 3282 { 3283 struct srpt_send_ioctx *ioctx = container_of(se_cmd, 3284 struct srpt_send_ioctx, cmd); 3285 struct srpt_rdma_ch *ch = ioctx->ch; 3286 struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx; 3287 3288 WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE && 3289 !(ioctx->cmd.transport_state & CMD_T_ABORTED)); 3290 3291 if (recv_ioctx) { 3292 WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list)); 3293 ioctx->recv_ioctx = NULL; 3294 srpt_post_recv(ch->sport->sdev, ch, recv_ioctx); 3295 } 3296 3297 if (ioctx->n_rw_ctx) { 3298 srpt_free_rw_ctxs(ch, ioctx); 3299 ioctx->n_rw_ctx = 0; 3300 } 3301 3302 target_free_tag(se_cmd->se_sess, se_cmd); 3303 } 3304 3305 /** 3306 * srpt_close_session - forcibly close a session 3307 * @se_sess: SCSI target session. 3308 * 3309 * Callback function invoked by the TCM core to clean up sessions associated 3310 * with a node ACL when the user invokes 3311 * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id 3312 */ 3313 static void srpt_close_session(struct se_session *se_sess) 3314 { 3315 struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr; 3316 3317 srpt_disconnect_ch_sync(ch); 3318 } 3319 3320 /** 3321 * srpt_sess_get_index - return the value of scsiAttIntrPortIndex (SCSI-MIB) 3322 * @se_sess: SCSI target session. 3323 * 3324 * A quote from RFC 4455 (SCSI-MIB) about this MIB object: 3325 * This object represents an arbitrary integer used to uniquely identify a 3326 * particular attached remote initiator port to a particular SCSI target port 3327 * within a particular SCSI target device within a particular SCSI instance. 3328 */ 3329 static u32 srpt_sess_get_index(struct se_session *se_sess) 3330 { 3331 return 0; 3332 } 3333 3334 static void srpt_set_default_node_attrs(struct se_node_acl *nacl) 3335 { 3336 } 3337 3338 /* Note: only used from inside debug printk's by the TCM core. */ 3339 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd) 3340 { 3341 struct srpt_send_ioctx *ioctx; 3342 3343 ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd); 3344 return ioctx->state; 3345 } 3346 3347 static int srpt_parse_guid(u64 *guid, const char *name) 3348 { 3349 u16 w[4]; 3350 int ret = -EINVAL; 3351 3352 if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4) 3353 goto out; 3354 *guid = get_unaligned_be64(w); 3355 ret = 0; 3356 out: 3357 return ret; 3358 } 3359 3360 /** 3361 * srpt_parse_i_port_id - parse an initiator port ID 3362 * @name: ASCII representation of a 128-bit initiator port ID. 3363 * @i_port_id: Binary 128-bit port ID. 3364 */ 3365 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name) 3366 { 3367 const char *p; 3368 unsigned len, count, leading_zero_bytes; 3369 int ret; 3370 3371 p = name; 3372 if (strncasecmp(p, "0x", 2) == 0) 3373 p += 2; 3374 ret = -EINVAL; 3375 len = strlen(p); 3376 if (len % 2) 3377 goto out; 3378 count = min(len / 2, 16U); 3379 leading_zero_bytes = 16 - count; 3380 memset(i_port_id, 0, leading_zero_bytes); 3381 ret = hex2bin(i_port_id + leading_zero_bytes, p, count); 3382 3383 out: 3384 return ret; 3385 } 3386 3387 /* 3388 * configfs callback function invoked for mkdir 3389 * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id 3390 * 3391 * i_port_id must be an initiator port GUID, GID or IP address. See also the 3392 * target_alloc_session() calls in this driver. Examples of valid initiator 3393 * port IDs: 3394 * 0x0000000000000000505400fffe4a0b7b 3395 * 0000000000000000505400fffe4a0b7b 3396 * 5054:00ff:fe4a:0b7b 3397 * 192.168.122.76 3398 */ 3399 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name) 3400 { 3401 struct sockaddr_storage sa; 3402 u64 guid; 3403 u8 i_port_id[16]; 3404 int ret; 3405 3406 ret = srpt_parse_guid(&guid, name); 3407 if (ret < 0) 3408 ret = srpt_parse_i_port_id(i_port_id, name); 3409 if (ret < 0) 3410 ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL, 3411 &sa); 3412 if (ret < 0) 3413 pr_err("invalid initiator port ID %s\n", name); 3414 return ret; 3415 } 3416 3417 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item, 3418 char *page) 3419 { 3420 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3421 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3422 3423 return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size); 3424 } 3425 3426 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item, 3427 const char *page, size_t count) 3428 { 3429 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3430 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3431 unsigned long val; 3432 int ret; 3433 3434 ret = kstrtoul(page, 0, &val); 3435 if (ret < 0) { 3436 pr_err("kstrtoul() failed with ret: %d\n", ret); 3437 return -EINVAL; 3438 } 3439 if (val > MAX_SRPT_RDMA_SIZE) { 3440 pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val, 3441 MAX_SRPT_RDMA_SIZE); 3442 return -EINVAL; 3443 } 3444 if (val < DEFAULT_MAX_RDMA_SIZE) { 3445 pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n", 3446 val, DEFAULT_MAX_RDMA_SIZE); 3447 return -EINVAL; 3448 } 3449 sport->port_attrib.srp_max_rdma_size = val; 3450 3451 return count; 3452 } 3453 3454 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item, 3455 char *page) 3456 { 3457 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3458 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3459 3460 return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size); 3461 } 3462 3463 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item, 3464 const char *page, size_t count) 3465 { 3466 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3467 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3468 unsigned long val; 3469 int ret; 3470 3471 ret = kstrtoul(page, 0, &val); 3472 if (ret < 0) { 3473 pr_err("kstrtoul() failed with ret: %d\n", ret); 3474 return -EINVAL; 3475 } 3476 if (val > MAX_SRPT_RSP_SIZE) { 3477 pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val, 3478 MAX_SRPT_RSP_SIZE); 3479 return -EINVAL; 3480 } 3481 if (val < MIN_MAX_RSP_SIZE) { 3482 pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val, 3483 MIN_MAX_RSP_SIZE); 3484 return -EINVAL; 3485 } 3486 sport->port_attrib.srp_max_rsp_size = val; 3487 3488 return count; 3489 } 3490 3491 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item, 3492 char *page) 3493 { 3494 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3495 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3496 3497 return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size); 3498 } 3499 3500 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item, 3501 const char *page, size_t count) 3502 { 3503 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3504 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3505 unsigned long val; 3506 int ret; 3507 3508 ret = kstrtoul(page, 0, &val); 3509 if (ret < 0) { 3510 pr_err("kstrtoul() failed with ret: %d\n", ret); 3511 return -EINVAL; 3512 } 3513 if (val > MAX_SRPT_SRQ_SIZE) { 3514 pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val, 3515 MAX_SRPT_SRQ_SIZE); 3516 return -EINVAL; 3517 } 3518 if (val < MIN_SRPT_SRQ_SIZE) { 3519 pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val, 3520 MIN_SRPT_SRQ_SIZE); 3521 return -EINVAL; 3522 } 3523 sport->port_attrib.srp_sq_size = val; 3524 3525 return count; 3526 } 3527 3528 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item, 3529 char *page) 3530 { 3531 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3532 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3533 3534 return sprintf(page, "%d\n", sport->port_attrib.use_srq); 3535 } 3536 3537 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item, 3538 const char *page, size_t count) 3539 { 3540 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3541 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3542 struct srpt_device *sdev = sport->sdev; 3543 unsigned long val; 3544 bool enabled; 3545 int ret; 3546 3547 ret = kstrtoul(page, 0, &val); 3548 if (ret < 0) 3549 return ret; 3550 if (val != !!val) 3551 return -EINVAL; 3552 3553 ret = mutex_lock_interruptible(&sdev->sdev_mutex); 3554 if (ret < 0) 3555 return ret; 3556 ret = mutex_lock_interruptible(&sport->mutex); 3557 if (ret < 0) 3558 goto unlock_sdev; 3559 enabled = sport->enabled; 3560 /* Log out all initiator systems before changing 'use_srq'. */ 3561 srpt_set_enabled(sport, false); 3562 sport->port_attrib.use_srq = val; 3563 srpt_use_srq(sdev, sport->port_attrib.use_srq); 3564 srpt_set_enabled(sport, enabled); 3565 ret = count; 3566 mutex_unlock(&sport->mutex); 3567 unlock_sdev: 3568 mutex_unlock(&sdev->sdev_mutex); 3569 3570 return ret; 3571 } 3572 3573 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_max_rdma_size); 3574 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_max_rsp_size); 3575 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_sq_size); 3576 CONFIGFS_ATTR(srpt_tpg_attrib_, use_srq); 3577 3578 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = { 3579 &srpt_tpg_attrib_attr_srp_max_rdma_size, 3580 &srpt_tpg_attrib_attr_srp_max_rsp_size, 3581 &srpt_tpg_attrib_attr_srp_sq_size, 3582 &srpt_tpg_attrib_attr_use_srq, 3583 NULL, 3584 }; 3585 3586 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr) 3587 { 3588 struct rdma_cm_id *rdma_cm_id; 3589 int ret; 3590 3591 rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler, 3592 NULL, RDMA_PS_TCP, IB_QPT_RC); 3593 if (IS_ERR(rdma_cm_id)) { 3594 pr_err("RDMA/CM ID creation failed: %ld\n", 3595 PTR_ERR(rdma_cm_id)); 3596 goto out; 3597 } 3598 3599 ret = rdma_bind_addr(rdma_cm_id, listen_addr); 3600 if (ret) { 3601 char addr_str[64]; 3602 3603 snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr); 3604 pr_err("Binding RDMA/CM ID to address %s failed: %d\n", 3605 addr_str, ret); 3606 rdma_destroy_id(rdma_cm_id); 3607 rdma_cm_id = ERR_PTR(ret); 3608 goto out; 3609 } 3610 3611 ret = rdma_listen(rdma_cm_id, 128); 3612 if (ret) { 3613 pr_err("rdma_listen() failed: %d\n", ret); 3614 rdma_destroy_id(rdma_cm_id); 3615 rdma_cm_id = ERR_PTR(ret); 3616 } 3617 3618 out: 3619 return rdma_cm_id; 3620 } 3621 3622 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page) 3623 { 3624 return sprintf(page, "%d\n", rdma_cm_port); 3625 } 3626 3627 static ssize_t srpt_rdma_cm_port_store(struct config_item *item, 3628 const char *page, size_t count) 3629 { 3630 struct sockaddr_in addr4 = { .sin_family = AF_INET }; 3631 struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 }; 3632 struct rdma_cm_id *new_id = NULL; 3633 u16 val; 3634 int ret; 3635 3636 ret = kstrtou16(page, 0, &val); 3637 if (ret < 0) 3638 return ret; 3639 ret = count; 3640 if (rdma_cm_port == val) 3641 goto out; 3642 3643 if (val) { 3644 addr6.sin6_port = cpu_to_be16(val); 3645 new_id = srpt_create_rdma_id((struct sockaddr *)&addr6); 3646 if (IS_ERR(new_id)) { 3647 addr4.sin_port = cpu_to_be16(val); 3648 new_id = srpt_create_rdma_id((struct sockaddr *)&addr4); 3649 if (IS_ERR(new_id)) { 3650 ret = PTR_ERR(new_id); 3651 goto out; 3652 } 3653 } 3654 } 3655 3656 mutex_lock(&rdma_cm_mutex); 3657 rdma_cm_port = val; 3658 swap(rdma_cm_id, new_id); 3659 mutex_unlock(&rdma_cm_mutex); 3660 3661 if (new_id) 3662 rdma_destroy_id(new_id); 3663 ret = count; 3664 out: 3665 return ret; 3666 } 3667 3668 CONFIGFS_ATTR(srpt_, rdma_cm_port); 3669 3670 static struct configfs_attribute *srpt_da_attrs[] = { 3671 &srpt_attr_rdma_cm_port, 3672 NULL, 3673 }; 3674 3675 static ssize_t srpt_tpg_enable_show(struct config_item *item, char *page) 3676 { 3677 struct se_portal_group *se_tpg = to_tpg(item); 3678 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3679 3680 return snprintf(page, PAGE_SIZE, "%d\n", sport->enabled); 3681 } 3682 3683 static ssize_t srpt_tpg_enable_store(struct config_item *item, 3684 const char *page, size_t count) 3685 { 3686 struct se_portal_group *se_tpg = to_tpg(item); 3687 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3688 unsigned long tmp; 3689 int ret; 3690 3691 ret = kstrtoul(page, 0, &tmp); 3692 if (ret < 0) { 3693 pr_err("Unable to extract srpt_tpg_store_enable\n"); 3694 return -EINVAL; 3695 } 3696 3697 if ((tmp != 0) && (tmp != 1)) { 3698 pr_err("Illegal value for srpt_tpg_store_enable: %lu\n", tmp); 3699 return -EINVAL; 3700 } 3701 3702 mutex_lock(&sport->mutex); 3703 srpt_set_enabled(sport, tmp); 3704 mutex_unlock(&sport->mutex); 3705 3706 return count; 3707 } 3708 3709 CONFIGFS_ATTR(srpt_tpg_, enable); 3710 3711 static struct configfs_attribute *srpt_tpg_attrs[] = { 3712 &srpt_tpg_attr_enable, 3713 NULL, 3714 }; 3715 3716 /** 3717 * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg 3718 * @wwn: Corresponds to $driver/$port. 3719 * @name: $tpg. 3720 */ 3721 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn, 3722 const char *name) 3723 { 3724 struct srpt_port *sport = wwn->priv; 3725 struct se_portal_group *tpg; 3726 int res; 3727 3728 WARN_ON_ONCE(wwn != &sport->port_guid_wwn && 3729 wwn != &sport->port_gid_wwn); 3730 tpg = wwn == &sport->port_guid_wwn ? &sport->port_guid_tpg : 3731 &sport->port_gid_tpg; 3732 res = core_tpg_register(wwn, tpg, SCSI_PROTOCOL_SRP); 3733 if (res) 3734 return ERR_PTR(res); 3735 3736 return tpg; 3737 } 3738 3739 /** 3740 * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg 3741 * @tpg: Target portal group to deregister. 3742 */ 3743 static void srpt_drop_tpg(struct se_portal_group *tpg) 3744 { 3745 struct srpt_port *sport = srpt_tpg_to_sport(tpg); 3746 3747 sport->enabled = false; 3748 core_tpg_deregister(tpg); 3749 } 3750 3751 /** 3752 * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port 3753 * @tf: Not used. 3754 * @group: Not used. 3755 * @name: $port. 3756 */ 3757 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf, 3758 struct config_group *group, 3759 const char *name) 3760 { 3761 return srpt_lookup_wwn(name) ? : ERR_PTR(-EINVAL); 3762 } 3763 3764 /** 3765 * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port 3766 * @wwn: $port. 3767 */ 3768 static void srpt_drop_tport(struct se_wwn *wwn) 3769 { 3770 } 3771 3772 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf) 3773 { 3774 return scnprintf(buf, PAGE_SIZE, "\n"); 3775 } 3776 3777 CONFIGFS_ATTR_RO(srpt_wwn_, version); 3778 3779 static struct configfs_attribute *srpt_wwn_attrs[] = { 3780 &srpt_wwn_attr_version, 3781 NULL, 3782 }; 3783 3784 static const struct target_core_fabric_ops srpt_template = { 3785 .module = THIS_MODULE, 3786 .fabric_name = "srpt", 3787 .tpg_get_wwn = srpt_get_fabric_wwn, 3788 .tpg_get_tag = srpt_get_tag, 3789 .tpg_check_demo_mode = srpt_check_false, 3790 .tpg_check_demo_mode_cache = srpt_check_true, 3791 .tpg_check_demo_mode_write_protect = srpt_check_true, 3792 .tpg_check_prod_mode_write_protect = srpt_check_false, 3793 .tpg_get_inst_index = srpt_tpg_get_inst_index, 3794 .release_cmd = srpt_release_cmd, 3795 .check_stop_free = srpt_check_stop_free, 3796 .close_session = srpt_close_session, 3797 .sess_get_index = srpt_sess_get_index, 3798 .sess_get_initiator_sid = NULL, 3799 .write_pending = srpt_write_pending, 3800 .set_default_node_attributes = srpt_set_default_node_attrs, 3801 .get_cmd_state = srpt_get_tcm_cmd_state, 3802 .queue_data_in = srpt_queue_data_in, 3803 .queue_status = srpt_queue_status, 3804 .queue_tm_rsp = srpt_queue_tm_rsp, 3805 .aborted_task = srpt_aborted_task, 3806 /* 3807 * Setup function pointers for generic logic in 3808 * target_core_fabric_configfs.c 3809 */ 3810 .fabric_make_wwn = srpt_make_tport, 3811 .fabric_drop_wwn = srpt_drop_tport, 3812 .fabric_make_tpg = srpt_make_tpg, 3813 .fabric_drop_tpg = srpt_drop_tpg, 3814 .fabric_init_nodeacl = srpt_init_nodeacl, 3815 3816 .tfc_discovery_attrs = srpt_da_attrs, 3817 .tfc_wwn_attrs = srpt_wwn_attrs, 3818 .tfc_tpg_base_attrs = srpt_tpg_attrs, 3819 .tfc_tpg_attrib_attrs = srpt_tpg_attrib_attrs, 3820 }; 3821 3822 /** 3823 * srpt_init_module - kernel module initialization 3824 * 3825 * Note: Since ib_register_client() registers callback functions, and since at 3826 * least one of these callback functions (srpt_add_one()) calls target core 3827 * functions, this driver must be registered with the target core before 3828 * ib_register_client() is called. 3829 */ 3830 static int __init srpt_init_module(void) 3831 { 3832 int ret; 3833 3834 ret = -EINVAL; 3835 if (srp_max_req_size < MIN_MAX_REQ_SIZE) { 3836 pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n", 3837 srp_max_req_size, MIN_MAX_REQ_SIZE); 3838 goto out; 3839 } 3840 3841 if (srpt_srq_size < MIN_SRPT_SRQ_SIZE 3842 || srpt_srq_size > MAX_SRPT_SRQ_SIZE) { 3843 pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n", 3844 srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE); 3845 goto out; 3846 } 3847 3848 ret = target_register_template(&srpt_template); 3849 if (ret) 3850 goto out; 3851 3852 ret = ib_register_client(&srpt_client); 3853 if (ret) { 3854 pr_err("couldn't register IB client\n"); 3855 goto out_unregister_target; 3856 } 3857 3858 return 0; 3859 3860 out_unregister_target: 3861 target_unregister_template(&srpt_template); 3862 out: 3863 return ret; 3864 } 3865 3866 static void __exit srpt_cleanup_module(void) 3867 { 3868 if (rdma_cm_id) 3869 rdma_destroy_id(rdma_cm_id); 3870 ib_unregister_client(&srpt_client); 3871 target_unregister_template(&srpt_template); 3872 } 3873 3874 module_init(srpt_init_module); 3875 module_exit(srpt_cleanup_module); 3876