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 <rdma/ib_cache.h> 45 #include <scsi/scsi_proto.h> 46 #include <scsi/scsi_tcq.h> 47 #include <target/target_core_base.h> 48 #include <target/target_core_fabric.h> 49 #include "ib_srpt.h" 50 51 /* Name of this kernel module. */ 52 #define DRV_NAME "ib_srpt" 53 #define DRV_VERSION "2.0.0" 54 #define DRV_RELDATE "2011-02-14" 55 56 #define SRPT_ID_STRING "Linux SRP target" 57 58 #undef pr_fmt 59 #define pr_fmt(fmt) DRV_NAME " " fmt 60 61 MODULE_AUTHOR("Vu Pham and Bart Van Assche"); 62 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol target " 63 "v" DRV_VERSION " (" DRV_RELDATE ")"); 64 MODULE_LICENSE("Dual BSD/GPL"); 65 66 /* 67 * Global Variables 68 */ 69 70 static u64 srpt_service_guid; 71 static DEFINE_SPINLOCK(srpt_dev_lock); /* Protects srpt_dev_list. */ 72 static LIST_HEAD(srpt_dev_list); /* List of srpt_device structures. */ 73 74 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE; 75 module_param(srp_max_req_size, int, 0444); 76 MODULE_PARM_DESC(srp_max_req_size, 77 "Maximum size of SRP request messages in bytes."); 78 79 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE; 80 module_param(srpt_srq_size, int, 0444); 81 MODULE_PARM_DESC(srpt_srq_size, 82 "Shared receive queue (SRQ) size."); 83 84 static int srpt_get_u64_x(char *buffer, const struct kernel_param *kp) 85 { 86 return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg); 87 } 88 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid, 89 0444); 90 MODULE_PARM_DESC(srpt_service_guid, 91 "Using this value for ioc_guid, id_ext, and cm_listen_id" 92 " instead of using the node_guid of the first HCA."); 93 94 static struct ib_client srpt_client; 95 static void srpt_release_cmd(struct se_cmd *se_cmd); 96 static void srpt_free_ch(struct kref *kref); 97 static int srpt_queue_status(struct se_cmd *cmd); 98 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc); 99 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc); 100 static void srpt_process_wait_list(struct srpt_rdma_ch *ch); 101 102 /* 103 * The only allowed channel state changes are those that change the channel 104 * state into a state with a higher numerical value. Hence the new > prev test. 105 */ 106 static bool srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new) 107 { 108 unsigned long flags; 109 enum rdma_ch_state prev; 110 bool changed = false; 111 112 spin_lock_irqsave(&ch->spinlock, flags); 113 prev = ch->state; 114 if (new > prev) { 115 ch->state = new; 116 changed = true; 117 } 118 spin_unlock_irqrestore(&ch->spinlock, flags); 119 120 return changed; 121 } 122 123 /** 124 * srpt_event_handler - asynchronous IB event callback function 125 * @handler: IB event handler registered by ib_register_event_handler(). 126 * @event: Description of the event that occurred. 127 * 128 * Callback function called by the InfiniBand core when an asynchronous IB 129 * event occurs. This callback may occur in interrupt context. See also 130 * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand 131 * Architecture Specification. 132 */ 133 static void srpt_event_handler(struct ib_event_handler *handler, 134 struct ib_event *event) 135 { 136 struct srpt_device *sdev; 137 struct srpt_port *sport; 138 u8 port_num; 139 140 sdev = ib_get_client_data(event->device, &srpt_client); 141 if (!sdev || sdev->device != event->device) 142 return; 143 144 pr_debug("ASYNC event= %d on device= %s\n", event->event, 145 sdev->device->name); 146 147 switch (event->event) { 148 case IB_EVENT_PORT_ERR: 149 port_num = event->element.port_num - 1; 150 if (port_num < sdev->device->phys_port_cnt) { 151 sport = &sdev->port[port_num]; 152 sport->lid = 0; 153 sport->sm_lid = 0; 154 } else { 155 WARN(true, "event %d: port_num %d out of range 1..%d\n", 156 event->event, port_num + 1, 157 sdev->device->phys_port_cnt); 158 } 159 break; 160 case IB_EVENT_PORT_ACTIVE: 161 case IB_EVENT_LID_CHANGE: 162 case IB_EVENT_PKEY_CHANGE: 163 case IB_EVENT_SM_CHANGE: 164 case IB_EVENT_CLIENT_REREGISTER: 165 case IB_EVENT_GID_CHANGE: 166 /* Refresh port data asynchronously. */ 167 port_num = event->element.port_num - 1; 168 if (port_num < sdev->device->phys_port_cnt) { 169 sport = &sdev->port[port_num]; 170 if (!sport->lid && !sport->sm_lid) 171 schedule_work(&sport->work); 172 } else { 173 WARN(true, "event %d: port_num %d out of range 1..%d\n", 174 event->event, port_num + 1, 175 sdev->device->phys_port_cnt); 176 } 177 break; 178 default: 179 pr_err("received unrecognized IB event %d\n", event->event); 180 break; 181 } 182 } 183 184 /** 185 * srpt_srq_event - SRQ event callback function 186 * @event: Description of the event that occurred. 187 * @ctx: Context pointer specified at SRQ creation time. 188 */ 189 static void srpt_srq_event(struct ib_event *event, void *ctx) 190 { 191 pr_debug("SRQ event %d\n", event->event); 192 } 193 194 static const char *get_ch_state_name(enum rdma_ch_state s) 195 { 196 switch (s) { 197 case CH_CONNECTING: 198 return "connecting"; 199 case CH_LIVE: 200 return "live"; 201 case CH_DISCONNECTING: 202 return "disconnecting"; 203 case CH_DRAINING: 204 return "draining"; 205 case CH_DISCONNECTED: 206 return "disconnected"; 207 } 208 return "???"; 209 } 210 211 /** 212 * srpt_qp_event - QP event callback function 213 * @event: Description of the event that occurred. 214 * @ch: SRPT RDMA channel. 215 */ 216 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch) 217 { 218 pr_debug("QP event %d on ch=%p sess_name=%s state=%d\n", 219 event->event, ch, ch->sess_name, ch->state); 220 221 switch (event->event) { 222 case IB_EVENT_COMM_EST: 223 ib_cm_notify(ch->ib_cm.cm_id, event->event); 224 break; 225 case IB_EVENT_QP_LAST_WQE_REACHED: 226 pr_debug("%s-%d, state %s: received Last WQE event.\n", 227 ch->sess_name, ch->qp->qp_num, 228 get_ch_state_name(ch->state)); 229 break; 230 default: 231 pr_err("received unrecognized IB QP event %d\n", event->event); 232 break; 233 } 234 } 235 236 /** 237 * srpt_set_ioc - initialize a IOUnitInfo structure 238 * @c_list: controller list. 239 * @slot: one-based slot number. 240 * @value: four-bit value. 241 * 242 * Copies the lowest four bits of value in element slot of the array of four 243 * bit elements called c_list (controller list). The index slot is one-based. 244 */ 245 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value) 246 { 247 u16 id; 248 u8 tmp; 249 250 id = (slot - 1) / 2; 251 if (slot & 0x1) { 252 tmp = c_list[id] & 0xf; 253 c_list[id] = (value << 4) | tmp; 254 } else { 255 tmp = c_list[id] & 0xf0; 256 c_list[id] = (value & 0xf) | tmp; 257 } 258 } 259 260 /** 261 * srpt_get_class_port_info - copy ClassPortInfo to a management datagram 262 * @mad: Datagram that will be sent as response to DM_ATTR_CLASS_PORT_INFO. 263 * 264 * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture 265 * Specification. 266 */ 267 static void srpt_get_class_port_info(struct ib_dm_mad *mad) 268 { 269 struct ib_class_port_info *cif; 270 271 cif = (struct ib_class_port_info *)mad->data; 272 memset(cif, 0, sizeof(*cif)); 273 cif->base_version = 1; 274 cif->class_version = 1; 275 276 ib_set_cpi_resp_time(cif, 20); 277 mad->mad_hdr.status = 0; 278 } 279 280 /** 281 * srpt_get_iou - write IOUnitInfo to a management datagram 282 * @mad: Datagram that will be sent as response to DM_ATTR_IOU_INFO. 283 * 284 * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture 285 * Specification. See also section B.7, table B.6 in the SRP r16a document. 286 */ 287 static void srpt_get_iou(struct ib_dm_mad *mad) 288 { 289 struct ib_dm_iou_info *ioui; 290 u8 slot; 291 int i; 292 293 ioui = (struct ib_dm_iou_info *)mad->data; 294 ioui->change_id = cpu_to_be16(1); 295 ioui->max_controllers = 16; 296 297 /* set present for slot 1 and empty for the rest */ 298 srpt_set_ioc(ioui->controller_list, 1, 1); 299 for (i = 1, slot = 2; i < 16; i++, slot++) 300 srpt_set_ioc(ioui->controller_list, slot, 0); 301 302 mad->mad_hdr.status = 0; 303 } 304 305 /** 306 * srpt_get_ioc - write IOControllerprofile to a management datagram 307 * @sport: HCA port through which the MAD has been received. 308 * @slot: Slot number specified in DM_ATTR_IOC_PROFILE query. 309 * @mad: Datagram that will be sent as response to DM_ATTR_IOC_PROFILE. 310 * 311 * See also section 16.3.3.4 IOControllerProfile in the InfiniBand 312 * Architecture Specification. See also section B.7, table B.7 in the SRP 313 * r16a document. 314 */ 315 static void srpt_get_ioc(struct srpt_port *sport, u32 slot, 316 struct ib_dm_mad *mad) 317 { 318 struct srpt_device *sdev = sport->sdev; 319 struct ib_dm_ioc_profile *iocp; 320 int send_queue_depth; 321 322 iocp = (struct ib_dm_ioc_profile *)mad->data; 323 324 if (!slot || slot > 16) { 325 mad->mad_hdr.status 326 = cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD); 327 return; 328 } 329 330 if (slot > 2) { 331 mad->mad_hdr.status 332 = cpu_to_be16(DM_MAD_STATUS_NO_IOC); 333 return; 334 } 335 336 if (sdev->use_srq) 337 send_queue_depth = sdev->srq_size; 338 else 339 send_queue_depth = min(MAX_SRPT_RQ_SIZE, 340 sdev->device->attrs.max_qp_wr); 341 342 memset(iocp, 0, sizeof(*iocp)); 343 strcpy(iocp->id_string, SRPT_ID_STRING); 344 iocp->guid = cpu_to_be64(srpt_service_guid); 345 iocp->vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id); 346 iocp->device_id = cpu_to_be32(sdev->device->attrs.vendor_part_id); 347 iocp->device_version = cpu_to_be16(sdev->device->attrs.hw_ver); 348 iocp->subsys_vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id); 349 iocp->subsys_device_id = 0x0; 350 iocp->io_class = cpu_to_be16(SRP_REV16A_IB_IO_CLASS); 351 iocp->io_subclass = cpu_to_be16(SRP_IO_SUBCLASS); 352 iocp->protocol = cpu_to_be16(SRP_PROTOCOL); 353 iocp->protocol_version = cpu_to_be16(SRP_PROTOCOL_VERSION); 354 iocp->send_queue_depth = cpu_to_be16(send_queue_depth); 355 iocp->rdma_read_depth = 4; 356 iocp->send_size = cpu_to_be32(srp_max_req_size); 357 iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size, 358 1U << 24)); 359 iocp->num_svc_entries = 1; 360 iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC | 361 SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC; 362 363 mad->mad_hdr.status = 0; 364 } 365 366 /** 367 * srpt_get_svc_entries - write ServiceEntries to a management datagram 368 * @ioc_guid: I/O controller GUID to use in reply. 369 * @slot: I/O controller number. 370 * @hi: End of the range of service entries to be specified in the reply. 371 * @lo: Start of the range of service entries to be specified in the reply.. 372 * @mad: Datagram that will be sent as response to DM_ATTR_SVC_ENTRIES. 373 * 374 * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture 375 * Specification. See also section B.7, table B.8 in the SRP r16a document. 376 */ 377 static void srpt_get_svc_entries(u64 ioc_guid, 378 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad) 379 { 380 struct ib_dm_svc_entries *svc_entries; 381 382 WARN_ON(!ioc_guid); 383 384 if (!slot || slot > 16) { 385 mad->mad_hdr.status 386 = cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD); 387 return; 388 } 389 390 if (slot > 2 || lo > hi || hi > 1) { 391 mad->mad_hdr.status 392 = cpu_to_be16(DM_MAD_STATUS_NO_IOC); 393 return; 394 } 395 396 svc_entries = (struct ib_dm_svc_entries *)mad->data; 397 memset(svc_entries, 0, sizeof(*svc_entries)); 398 svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid); 399 snprintf(svc_entries->service_entries[0].name, 400 sizeof(svc_entries->service_entries[0].name), 401 "%s%016llx", 402 SRP_SERVICE_NAME_PREFIX, 403 ioc_guid); 404 405 mad->mad_hdr.status = 0; 406 } 407 408 /** 409 * srpt_mgmt_method_get - process a received management datagram 410 * @sp: HCA port through which the MAD has been received. 411 * @rq_mad: received MAD. 412 * @rsp_mad: response MAD. 413 */ 414 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad, 415 struct ib_dm_mad *rsp_mad) 416 { 417 u16 attr_id; 418 u32 slot; 419 u8 hi, lo; 420 421 attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id); 422 switch (attr_id) { 423 case DM_ATTR_CLASS_PORT_INFO: 424 srpt_get_class_port_info(rsp_mad); 425 break; 426 case DM_ATTR_IOU_INFO: 427 srpt_get_iou(rsp_mad); 428 break; 429 case DM_ATTR_IOC_PROFILE: 430 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod); 431 srpt_get_ioc(sp, slot, rsp_mad); 432 break; 433 case DM_ATTR_SVC_ENTRIES: 434 slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod); 435 hi = (u8) ((slot >> 8) & 0xff); 436 lo = (u8) (slot & 0xff); 437 slot = (u16) ((slot >> 16) & 0xffff); 438 srpt_get_svc_entries(srpt_service_guid, 439 slot, hi, lo, rsp_mad); 440 break; 441 default: 442 rsp_mad->mad_hdr.status = 443 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR); 444 break; 445 } 446 } 447 448 /** 449 * srpt_mad_send_handler - MAD send completion callback 450 * @mad_agent: Return value of ib_register_mad_agent(). 451 * @mad_wc: Work completion reporting that the MAD has been sent. 452 */ 453 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent, 454 struct ib_mad_send_wc *mad_wc) 455 { 456 rdma_destroy_ah(mad_wc->send_buf->ah); 457 ib_free_send_mad(mad_wc->send_buf); 458 } 459 460 /** 461 * srpt_mad_recv_handler - MAD reception callback function 462 * @mad_agent: Return value of ib_register_mad_agent(). 463 * @send_buf: Not used. 464 * @mad_wc: Work completion reporting that a MAD has been received. 465 */ 466 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent, 467 struct ib_mad_send_buf *send_buf, 468 struct ib_mad_recv_wc *mad_wc) 469 { 470 struct srpt_port *sport = (struct srpt_port *)mad_agent->context; 471 struct ib_ah *ah; 472 struct ib_mad_send_buf *rsp; 473 struct ib_dm_mad *dm_mad; 474 475 if (!mad_wc || !mad_wc->recv_buf.mad) 476 return; 477 478 ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc, 479 mad_wc->recv_buf.grh, mad_agent->port_num); 480 if (IS_ERR(ah)) 481 goto err; 482 483 BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR); 484 485 rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp, 486 mad_wc->wc->pkey_index, 0, 487 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA, 488 GFP_KERNEL, 489 IB_MGMT_BASE_VERSION); 490 if (IS_ERR(rsp)) 491 goto err_rsp; 492 493 rsp->ah = ah; 494 495 dm_mad = rsp->mad; 496 memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof(*dm_mad)); 497 dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP; 498 dm_mad->mad_hdr.status = 0; 499 500 switch (mad_wc->recv_buf.mad->mad_hdr.method) { 501 case IB_MGMT_METHOD_GET: 502 srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad); 503 break; 504 case IB_MGMT_METHOD_SET: 505 dm_mad->mad_hdr.status = 506 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR); 507 break; 508 default: 509 dm_mad->mad_hdr.status = 510 cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD); 511 break; 512 } 513 514 if (!ib_post_send_mad(rsp, NULL)) { 515 ib_free_recv_mad(mad_wc); 516 /* will destroy_ah & free_send_mad in send completion */ 517 return; 518 } 519 520 ib_free_send_mad(rsp); 521 522 err_rsp: 523 rdma_destroy_ah(ah); 524 err: 525 ib_free_recv_mad(mad_wc); 526 } 527 528 static int srpt_format_guid(char *buf, unsigned int size, const __be64 *guid) 529 { 530 const __be16 *g = (const __be16 *)guid; 531 532 return snprintf(buf, size, "%04x:%04x:%04x:%04x", 533 be16_to_cpu(g[0]), be16_to_cpu(g[1]), 534 be16_to_cpu(g[2]), be16_to_cpu(g[3])); 535 } 536 537 /** 538 * srpt_refresh_port - configure a HCA port 539 * @sport: SRPT HCA port. 540 * 541 * Enable InfiniBand management datagram processing, update the cached sm_lid, 542 * lid and gid values, and register a callback function for processing MADs 543 * on the specified port. 544 * 545 * Note: It is safe to call this function more than once for the same port. 546 */ 547 static int srpt_refresh_port(struct srpt_port *sport) 548 { 549 struct ib_mad_reg_req reg_req; 550 struct ib_port_modify port_modify; 551 struct ib_port_attr port_attr; 552 int ret; 553 554 memset(&port_modify, 0, sizeof(port_modify)); 555 port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP; 556 port_modify.clr_port_cap_mask = 0; 557 558 ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify); 559 if (ret) 560 goto err_mod_port; 561 562 ret = ib_query_port(sport->sdev->device, sport->port, &port_attr); 563 if (ret) 564 goto err_query_port; 565 566 sport->sm_lid = port_attr.sm_lid; 567 sport->lid = port_attr.lid; 568 569 ret = ib_query_gid(sport->sdev->device, sport->port, 0, &sport->gid, 570 NULL); 571 if (ret) 572 goto err_query_port; 573 574 sport->port_guid_wwn.priv = sport; 575 srpt_format_guid(sport->port_guid, sizeof(sport->port_guid), 576 &sport->gid.global.interface_id); 577 sport->port_gid_wwn.priv = sport; 578 snprintf(sport->port_gid, sizeof(sport->port_gid), 579 "0x%016llx%016llx", 580 be64_to_cpu(sport->gid.global.subnet_prefix), 581 be64_to_cpu(sport->gid.global.interface_id)); 582 583 if (!sport->mad_agent) { 584 memset(®_req, 0, sizeof(reg_req)); 585 reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT; 586 reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION; 587 set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask); 588 set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask); 589 590 sport->mad_agent = ib_register_mad_agent(sport->sdev->device, 591 sport->port, 592 IB_QPT_GSI, 593 ®_req, 0, 594 srpt_mad_send_handler, 595 srpt_mad_recv_handler, 596 sport, 0); 597 if (IS_ERR(sport->mad_agent)) { 598 ret = PTR_ERR(sport->mad_agent); 599 sport->mad_agent = NULL; 600 goto err_query_port; 601 } 602 } 603 604 return 0; 605 606 err_query_port: 607 608 port_modify.set_port_cap_mask = 0; 609 port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP; 610 ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify); 611 612 err_mod_port: 613 614 return ret; 615 } 616 617 /** 618 * srpt_unregister_mad_agent - unregister MAD callback functions 619 * @sdev: SRPT HCA pointer. 620 * 621 * Note: It is safe to call this function more than once for the same device. 622 */ 623 static void srpt_unregister_mad_agent(struct srpt_device *sdev) 624 { 625 struct ib_port_modify port_modify = { 626 .clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP, 627 }; 628 struct srpt_port *sport; 629 int i; 630 631 for (i = 1; i <= sdev->device->phys_port_cnt; i++) { 632 sport = &sdev->port[i - 1]; 633 WARN_ON(sport->port != i); 634 if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0) 635 pr_err("disabling MAD processing failed.\n"); 636 if (sport->mad_agent) { 637 ib_unregister_mad_agent(sport->mad_agent); 638 sport->mad_agent = NULL; 639 } 640 } 641 } 642 643 /** 644 * srpt_alloc_ioctx - allocate a SRPT I/O context structure 645 * @sdev: SRPT HCA pointer. 646 * @ioctx_size: I/O context size. 647 * @dma_size: Size of I/O context DMA buffer. 648 * @dir: DMA data direction. 649 */ 650 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev, 651 int ioctx_size, int dma_size, 652 enum dma_data_direction dir) 653 { 654 struct srpt_ioctx *ioctx; 655 656 ioctx = kmalloc(ioctx_size, GFP_KERNEL); 657 if (!ioctx) 658 goto err; 659 660 ioctx->buf = kmalloc(dma_size, GFP_KERNEL); 661 if (!ioctx->buf) 662 goto err_free_ioctx; 663 664 ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, dma_size, dir); 665 if (ib_dma_mapping_error(sdev->device, ioctx->dma)) 666 goto err_free_buf; 667 668 return ioctx; 669 670 err_free_buf: 671 kfree(ioctx->buf); 672 err_free_ioctx: 673 kfree(ioctx); 674 err: 675 return NULL; 676 } 677 678 /** 679 * srpt_free_ioctx - free a SRPT I/O context structure 680 * @sdev: SRPT HCA pointer. 681 * @ioctx: I/O context pointer. 682 * @dma_size: Size of I/O context DMA buffer. 683 * @dir: DMA data direction. 684 */ 685 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx, 686 int dma_size, enum dma_data_direction dir) 687 { 688 if (!ioctx) 689 return; 690 691 ib_dma_unmap_single(sdev->device, ioctx->dma, dma_size, dir); 692 kfree(ioctx->buf); 693 kfree(ioctx); 694 } 695 696 /** 697 * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures 698 * @sdev: Device to allocate the I/O context ring for. 699 * @ring_size: Number of elements in the I/O context ring. 700 * @ioctx_size: I/O context size. 701 * @dma_size: DMA buffer size. 702 * @dir: DMA data direction. 703 */ 704 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev, 705 int ring_size, int ioctx_size, 706 int dma_size, enum dma_data_direction dir) 707 { 708 struct srpt_ioctx **ring; 709 int i; 710 711 WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) 712 && ioctx_size != sizeof(struct srpt_send_ioctx)); 713 714 ring = kmalloc(ring_size * sizeof(ring[0]), GFP_KERNEL); 715 if (!ring) 716 goto out; 717 for (i = 0; i < ring_size; ++i) { 718 ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, dma_size, dir); 719 if (!ring[i]) 720 goto err; 721 ring[i]->index = i; 722 } 723 goto out; 724 725 err: 726 while (--i >= 0) 727 srpt_free_ioctx(sdev, ring[i], dma_size, dir); 728 kfree(ring); 729 ring = NULL; 730 out: 731 return ring; 732 } 733 734 /** 735 * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures 736 * @ioctx_ring: I/O context ring to be freed. 737 * @sdev: SRPT HCA pointer. 738 * @ring_size: Number of ring elements. 739 * @dma_size: Size of I/O context DMA buffer. 740 * @dir: DMA data direction. 741 */ 742 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring, 743 struct srpt_device *sdev, int ring_size, 744 int dma_size, enum dma_data_direction dir) 745 { 746 int i; 747 748 if (!ioctx_ring) 749 return; 750 751 for (i = 0; i < ring_size; ++i) 752 srpt_free_ioctx(sdev, ioctx_ring[i], dma_size, dir); 753 kfree(ioctx_ring); 754 } 755 756 /** 757 * srpt_set_cmd_state - set the state of a SCSI command 758 * @ioctx: Send I/O context. 759 * @new: New I/O context state. 760 * 761 * Does not modify the state of aborted commands. Returns the previous command 762 * state. 763 */ 764 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx, 765 enum srpt_command_state new) 766 { 767 enum srpt_command_state previous; 768 769 previous = ioctx->state; 770 if (previous != SRPT_STATE_DONE) 771 ioctx->state = new; 772 773 return previous; 774 } 775 776 /** 777 * srpt_test_and_set_cmd_state - test and set the state of a command 778 * @ioctx: Send I/O context. 779 * @old: Current I/O context state. 780 * @new: New I/O context state. 781 * 782 * Returns true if and only if the previous command state was equal to 'old'. 783 */ 784 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx, 785 enum srpt_command_state old, 786 enum srpt_command_state new) 787 { 788 enum srpt_command_state previous; 789 790 WARN_ON(!ioctx); 791 WARN_ON(old == SRPT_STATE_DONE); 792 WARN_ON(new == SRPT_STATE_NEW); 793 794 previous = ioctx->state; 795 if (previous == old) 796 ioctx->state = new; 797 798 return previous == old; 799 } 800 801 /** 802 * srpt_post_recv - post an IB receive request 803 * @sdev: SRPT HCA pointer. 804 * @ch: SRPT RDMA channel. 805 * @ioctx: Receive I/O context pointer. 806 */ 807 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch, 808 struct srpt_recv_ioctx *ioctx) 809 { 810 struct ib_sge list; 811 struct ib_recv_wr wr, *bad_wr; 812 813 BUG_ON(!sdev); 814 list.addr = ioctx->ioctx.dma; 815 list.length = srp_max_req_size; 816 list.lkey = sdev->lkey; 817 818 ioctx->ioctx.cqe.done = srpt_recv_done; 819 wr.wr_cqe = &ioctx->ioctx.cqe; 820 wr.next = NULL; 821 wr.sg_list = &list; 822 wr.num_sge = 1; 823 824 if (sdev->use_srq) 825 return ib_post_srq_recv(sdev->srq, &wr, &bad_wr); 826 else 827 return ib_post_recv(ch->qp, &wr, &bad_wr); 828 } 829 830 /** 831 * srpt_zerolength_write - perform a zero-length RDMA write 832 * @ch: SRPT RDMA channel. 833 * 834 * A quote from the InfiniBand specification: C9-88: For an HCA responder 835 * using Reliable Connection service, for each zero-length RDMA READ or WRITE 836 * request, the R_Key shall not be validated, even if the request includes 837 * Immediate data. 838 */ 839 static int srpt_zerolength_write(struct srpt_rdma_ch *ch) 840 { 841 struct ib_send_wr wr, *bad_wr; 842 843 pr_debug("%s-%d: queued zerolength write\n", ch->sess_name, 844 ch->qp->qp_num); 845 846 memset(&wr, 0, sizeof(wr)); 847 wr.opcode = IB_WR_RDMA_WRITE; 848 wr.wr_cqe = &ch->zw_cqe; 849 wr.send_flags = IB_SEND_SIGNALED; 850 return ib_post_send(ch->qp, &wr, &bad_wr); 851 } 852 853 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc) 854 { 855 struct srpt_rdma_ch *ch = cq->cq_context; 856 857 pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num, 858 wc->status); 859 860 if (wc->status == IB_WC_SUCCESS) { 861 srpt_process_wait_list(ch); 862 } else { 863 if (srpt_set_ch_state(ch, CH_DISCONNECTED)) 864 schedule_work(&ch->release_work); 865 else 866 pr_debug("%s-%d: already disconnected.\n", 867 ch->sess_name, ch->qp->qp_num); 868 } 869 } 870 871 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx, 872 struct srp_direct_buf *db, int nbufs, struct scatterlist **sg, 873 unsigned *sg_cnt) 874 { 875 enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd); 876 struct srpt_rdma_ch *ch = ioctx->ch; 877 struct scatterlist *prev = NULL; 878 unsigned prev_nents; 879 int ret, i; 880 881 if (nbufs == 1) { 882 ioctx->rw_ctxs = &ioctx->s_rw_ctx; 883 } else { 884 ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs), 885 GFP_KERNEL); 886 if (!ioctx->rw_ctxs) 887 return -ENOMEM; 888 } 889 890 for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) { 891 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 892 u64 remote_addr = be64_to_cpu(db->va); 893 u32 size = be32_to_cpu(db->len); 894 u32 rkey = be32_to_cpu(db->key); 895 896 ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false, 897 i < nbufs - 1); 898 if (ret) 899 goto unwind; 900 901 ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port, 902 ctx->sg, ctx->nents, 0, remote_addr, rkey, dir); 903 if (ret < 0) { 904 target_free_sgl(ctx->sg, ctx->nents); 905 goto unwind; 906 } 907 908 ioctx->n_rdma += ret; 909 ioctx->n_rw_ctx++; 910 911 if (prev) { 912 sg_unmark_end(&prev[prev_nents - 1]); 913 sg_chain(prev, prev_nents + 1, ctx->sg); 914 } else { 915 *sg = ctx->sg; 916 } 917 918 prev = ctx->sg; 919 prev_nents = ctx->nents; 920 921 *sg_cnt += ctx->nents; 922 } 923 924 return 0; 925 926 unwind: 927 while (--i >= 0) { 928 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 929 930 rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port, 931 ctx->sg, ctx->nents, dir); 932 target_free_sgl(ctx->sg, ctx->nents); 933 } 934 if (ioctx->rw_ctxs != &ioctx->s_rw_ctx) 935 kfree(ioctx->rw_ctxs); 936 return ret; 937 } 938 939 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch, 940 struct srpt_send_ioctx *ioctx) 941 { 942 enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd); 943 int i; 944 945 for (i = 0; i < ioctx->n_rw_ctx; i++) { 946 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 947 948 rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port, 949 ctx->sg, ctx->nents, dir); 950 target_free_sgl(ctx->sg, ctx->nents); 951 } 952 953 if (ioctx->rw_ctxs != &ioctx->s_rw_ctx) 954 kfree(ioctx->rw_ctxs); 955 } 956 957 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd) 958 { 959 /* 960 * The pointer computations below will only be compiled correctly 961 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check 962 * whether srp_cmd::add_data has been declared as a byte pointer. 963 */ 964 BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) && 965 !__same_type(srp_cmd->add_data[0], (u8)0)); 966 967 /* 968 * According to the SRP spec, the lower two bits of the 'ADDITIONAL 969 * CDB LENGTH' field are reserved and the size in bytes of this field 970 * is four times the value specified in bits 3..7. Hence the "& ~3". 971 */ 972 return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3); 973 } 974 975 /** 976 * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request 977 * @ioctx: Pointer to the I/O context associated with the request. 978 * @srp_cmd: Pointer to the SRP_CMD request data. 979 * @dir: Pointer to the variable to which the transfer direction will be 980 * written. 981 * @sg: [out] scatterlist allocated for the parsed SRP_CMD. 982 * @sg_cnt: [out] length of @sg. 983 * @data_len: Pointer to the variable to which the total data length of all 984 * descriptors in the SRP_CMD request will be written. 985 * 986 * This function initializes ioctx->nrbuf and ioctx->r_bufs. 987 * 988 * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors; 989 * -ENOMEM when memory allocation fails and zero upon success. 990 */ 991 static int srpt_get_desc_tbl(struct srpt_send_ioctx *ioctx, 992 struct srp_cmd *srp_cmd, enum dma_data_direction *dir, 993 struct scatterlist **sg, unsigned *sg_cnt, u64 *data_len) 994 { 995 BUG_ON(!dir); 996 BUG_ON(!data_len); 997 998 /* 999 * The lower four bits of the buffer format field contain the DATA-IN 1000 * buffer descriptor format, and the highest four bits contain the 1001 * DATA-OUT buffer descriptor format. 1002 */ 1003 if (srp_cmd->buf_fmt & 0xf) 1004 /* DATA-IN: transfer data from target to initiator (read). */ 1005 *dir = DMA_FROM_DEVICE; 1006 else if (srp_cmd->buf_fmt >> 4) 1007 /* DATA-OUT: transfer data from initiator to target (write). */ 1008 *dir = DMA_TO_DEVICE; 1009 else 1010 *dir = DMA_NONE; 1011 1012 /* initialize data_direction early as srpt_alloc_rw_ctxs needs it */ 1013 ioctx->cmd.data_direction = *dir; 1014 1015 if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) || 1016 ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) { 1017 struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd); 1018 1019 *data_len = be32_to_cpu(db->len); 1020 return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt); 1021 } else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) || 1022 ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) { 1023 struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd); 1024 int nbufs = be32_to_cpu(idb->table_desc.len) / 1025 sizeof(struct srp_direct_buf); 1026 1027 if (nbufs > 1028 (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) { 1029 pr_err("received unsupported SRP_CMD request" 1030 " type (%u out + %u in != %u / %zu)\n", 1031 srp_cmd->data_out_desc_cnt, 1032 srp_cmd->data_in_desc_cnt, 1033 be32_to_cpu(idb->table_desc.len), 1034 sizeof(struct srp_direct_buf)); 1035 return -EINVAL; 1036 } 1037 1038 *data_len = be32_to_cpu(idb->len); 1039 return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs, 1040 sg, sg_cnt); 1041 } else { 1042 *data_len = 0; 1043 return 0; 1044 } 1045 } 1046 1047 /** 1048 * srpt_init_ch_qp - initialize queue pair attributes 1049 * @ch: SRPT RDMA channel. 1050 * @qp: Queue pair pointer. 1051 * 1052 * Initialized the attributes of queue pair 'qp' by allowing local write, 1053 * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT. 1054 */ 1055 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1056 { 1057 struct ib_qp_attr *attr; 1058 int ret; 1059 1060 attr = kzalloc(sizeof(*attr), GFP_KERNEL); 1061 if (!attr) 1062 return -ENOMEM; 1063 1064 attr->qp_state = IB_QPS_INIT; 1065 attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE; 1066 attr->port_num = ch->sport->port; 1067 1068 ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port, 1069 ch->pkey, &attr->pkey_index); 1070 if (ret < 0) 1071 pr_err("Translating pkey %#x failed (%d) - using index 0\n", 1072 ch->pkey, ret); 1073 1074 ret = ib_modify_qp(qp, attr, 1075 IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT | 1076 IB_QP_PKEY_INDEX); 1077 1078 kfree(attr); 1079 return ret; 1080 } 1081 1082 /** 1083 * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR) 1084 * @ch: channel of the queue pair. 1085 * @qp: queue pair to change the state of. 1086 * 1087 * Returns zero upon success and a negative value upon failure. 1088 * 1089 * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system. 1090 * If this structure ever becomes larger, it might be necessary to allocate 1091 * it dynamically instead of on the stack. 1092 */ 1093 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1094 { 1095 struct ib_qp_attr qp_attr; 1096 int attr_mask; 1097 int ret; 1098 1099 qp_attr.qp_state = IB_QPS_RTR; 1100 ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask); 1101 if (ret) 1102 goto out; 1103 1104 qp_attr.max_dest_rd_atomic = 4; 1105 1106 ret = ib_modify_qp(qp, &qp_attr, attr_mask); 1107 1108 out: 1109 return ret; 1110 } 1111 1112 /** 1113 * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS) 1114 * @ch: channel of the queue pair. 1115 * @qp: queue pair to change the state of. 1116 * 1117 * Returns zero upon success and a negative value upon failure. 1118 * 1119 * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system. 1120 * If this structure ever becomes larger, it might be necessary to allocate 1121 * it dynamically instead of on the stack. 1122 */ 1123 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp) 1124 { 1125 struct ib_qp_attr qp_attr; 1126 int attr_mask; 1127 int ret; 1128 1129 qp_attr.qp_state = IB_QPS_RTS; 1130 ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask); 1131 if (ret) 1132 goto out; 1133 1134 qp_attr.max_rd_atomic = 4; 1135 1136 ret = ib_modify_qp(qp, &qp_attr, attr_mask); 1137 1138 out: 1139 return ret; 1140 } 1141 1142 /** 1143 * srpt_ch_qp_err - set the channel queue pair state to 'error' 1144 * @ch: SRPT RDMA channel. 1145 */ 1146 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch) 1147 { 1148 struct ib_qp_attr qp_attr; 1149 1150 qp_attr.qp_state = IB_QPS_ERR; 1151 return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE); 1152 } 1153 1154 /** 1155 * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator 1156 * @ch: SRPT RDMA channel. 1157 */ 1158 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch) 1159 { 1160 struct srpt_send_ioctx *ioctx; 1161 unsigned long flags; 1162 1163 BUG_ON(!ch); 1164 1165 ioctx = NULL; 1166 spin_lock_irqsave(&ch->spinlock, flags); 1167 if (!list_empty(&ch->free_list)) { 1168 ioctx = list_first_entry(&ch->free_list, 1169 struct srpt_send_ioctx, free_list); 1170 list_del(&ioctx->free_list); 1171 } 1172 spin_unlock_irqrestore(&ch->spinlock, flags); 1173 1174 if (!ioctx) 1175 return ioctx; 1176 1177 BUG_ON(ioctx->ch != ch); 1178 ioctx->state = SRPT_STATE_NEW; 1179 ioctx->n_rdma = 0; 1180 ioctx->n_rw_ctx = 0; 1181 ioctx->queue_status_only = false; 1182 /* 1183 * transport_init_se_cmd() does not initialize all fields, so do it 1184 * here. 1185 */ 1186 memset(&ioctx->cmd, 0, sizeof(ioctx->cmd)); 1187 memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data)); 1188 1189 return ioctx; 1190 } 1191 1192 /** 1193 * srpt_abort_cmd - abort a SCSI command 1194 * @ioctx: I/O context associated with the SCSI command. 1195 */ 1196 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx) 1197 { 1198 enum srpt_command_state state; 1199 1200 BUG_ON(!ioctx); 1201 1202 /* 1203 * If the command is in a state where the target core is waiting for 1204 * the ib_srpt driver, change the state to the next state. 1205 */ 1206 1207 state = ioctx->state; 1208 switch (state) { 1209 case SRPT_STATE_NEED_DATA: 1210 ioctx->state = SRPT_STATE_DATA_IN; 1211 break; 1212 case SRPT_STATE_CMD_RSP_SENT: 1213 case SRPT_STATE_MGMT_RSP_SENT: 1214 ioctx->state = SRPT_STATE_DONE; 1215 break; 1216 default: 1217 WARN_ONCE(true, "%s: unexpected I/O context state %d\n", 1218 __func__, state); 1219 break; 1220 } 1221 1222 pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state, 1223 ioctx->state, ioctx->cmd.tag); 1224 1225 switch (state) { 1226 case SRPT_STATE_NEW: 1227 case SRPT_STATE_DATA_IN: 1228 case SRPT_STATE_MGMT: 1229 case SRPT_STATE_DONE: 1230 /* 1231 * Do nothing - defer abort processing until 1232 * srpt_queue_response() is invoked. 1233 */ 1234 break; 1235 case SRPT_STATE_NEED_DATA: 1236 pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag); 1237 transport_generic_request_failure(&ioctx->cmd, 1238 TCM_CHECK_CONDITION_ABORT_CMD); 1239 break; 1240 case SRPT_STATE_CMD_RSP_SENT: 1241 /* 1242 * SRP_RSP sending failed or the SRP_RSP send completion has 1243 * not been received in time. 1244 */ 1245 transport_generic_free_cmd(&ioctx->cmd, 0); 1246 break; 1247 case SRPT_STATE_MGMT_RSP_SENT: 1248 transport_generic_free_cmd(&ioctx->cmd, 0); 1249 break; 1250 default: 1251 WARN(1, "Unexpected command state (%d)", state); 1252 break; 1253 } 1254 1255 return state; 1256 } 1257 1258 /** 1259 * srpt_rdma_read_done - RDMA read completion callback 1260 * @cq: Completion queue. 1261 * @wc: Work completion. 1262 * 1263 * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping 1264 * the data that has been transferred via IB RDMA had to be postponed until the 1265 * check_stop_free() callback. None of this is necessary anymore and needs to 1266 * be cleaned up. 1267 */ 1268 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc) 1269 { 1270 struct srpt_rdma_ch *ch = cq->cq_context; 1271 struct srpt_send_ioctx *ioctx = 1272 container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe); 1273 1274 WARN_ON(ioctx->n_rdma <= 0); 1275 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail); 1276 ioctx->n_rdma = 0; 1277 1278 if (unlikely(wc->status != IB_WC_SUCCESS)) { 1279 pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n", 1280 ioctx, wc->status); 1281 srpt_abort_cmd(ioctx); 1282 return; 1283 } 1284 1285 if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA, 1286 SRPT_STATE_DATA_IN)) 1287 target_execute_cmd(&ioctx->cmd); 1288 else 1289 pr_err("%s[%d]: wrong state = %d\n", __func__, 1290 __LINE__, ioctx->state); 1291 } 1292 1293 /** 1294 * srpt_build_cmd_rsp - build a SRP_RSP response 1295 * @ch: RDMA channel through which the request has been received. 1296 * @ioctx: I/O context associated with the SRP_CMD request. The response will 1297 * be built in the buffer ioctx->buf points at and hence this function will 1298 * overwrite the request data. 1299 * @tag: tag of the request for which this response is being generated. 1300 * @status: value for the STATUS field of the SRP_RSP information unit. 1301 * 1302 * Returns the size in bytes of the SRP_RSP response. 1303 * 1304 * An SRP_RSP response contains a SCSI status or service response. See also 1305 * section 6.9 in the SRP r16a document for the format of an SRP_RSP 1306 * response. See also SPC-2 for more information about sense data. 1307 */ 1308 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch, 1309 struct srpt_send_ioctx *ioctx, u64 tag, 1310 int status) 1311 { 1312 struct srp_rsp *srp_rsp; 1313 const u8 *sense_data; 1314 int sense_data_len, max_sense_len; 1315 1316 /* 1317 * The lowest bit of all SAM-3 status codes is zero (see also 1318 * paragraph 5.3 in SAM-3). 1319 */ 1320 WARN_ON(status & 1); 1321 1322 srp_rsp = ioctx->ioctx.buf; 1323 BUG_ON(!srp_rsp); 1324 1325 sense_data = ioctx->sense_data; 1326 sense_data_len = ioctx->cmd.scsi_sense_length; 1327 WARN_ON(sense_data_len > sizeof(ioctx->sense_data)); 1328 1329 memset(srp_rsp, 0, sizeof(*srp_rsp)); 1330 srp_rsp->opcode = SRP_RSP; 1331 srp_rsp->req_lim_delta = 1332 cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0)); 1333 srp_rsp->tag = tag; 1334 srp_rsp->status = status; 1335 1336 if (sense_data_len) { 1337 BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp)); 1338 max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp); 1339 if (sense_data_len > max_sense_len) { 1340 pr_warn("truncated sense data from %d to %d" 1341 " bytes\n", sense_data_len, max_sense_len); 1342 sense_data_len = max_sense_len; 1343 } 1344 1345 srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID; 1346 srp_rsp->sense_data_len = cpu_to_be32(sense_data_len); 1347 memcpy(srp_rsp + 1, sense_data, sense_data_len); 1348 } 1349 1350 return sizeof(*srp_rsp) + sense_data_len; 1351 } 1352 1353 /** 1354 * srpt_build_tskmgmt_rsp - build a task management response 1355 * @ch: RDMA channel through which the request has been received. 1356 * @ioctx: I/O context in which the SRP_RSP response will be built. 1357 * @rsp_code: RSP_CODE that will be stored in the response. 1358 * @tag: Tag of the request for which this response is being generated. 1359 * 1360 * Returns the size in bytes of the SRP_RSP response. 1361 * 1362 * An SRP_RSP response contains a SCSI status or service response. See also 1363 * section 6.9 in the SRP r16a document for the format of an SRP_RSP 1364 * response. 1365 */ 1366 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch, 1367 struct srpt_send_ioctx *ioctx, 1368 u8 rsp_code, u64 tag) 1369 { 1370 struct srp_rsp *srp_rsp; 1371 int resp_data_len; 1372 int resp_len; 1373 1374 resp_data_len = 4; 1375 resp_len = sizeof(*srp_rsp) + resp_data_len; 1376 1377 srp_rsp = ioctx->ioctx.buf; 1378 BUG_ON(!srp_rsp); 1379 memset(srp_rsp, 0, sizeof(*srp_rsp)); 1380 1381 srp_rsp->opcode = SRP_RSP; 1382 srp_rsp->req_lim_delta = 1383 cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0)); 1384 srp_rsp->tag = tag; 1385 1386 srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID; 1387 srp_rsp->resp_data_len = cpu_to_be32(resp_data_len); 1388 srp_rsp->data[3] = rsp_code; 1389 1390 return resp_len; 1391 } 1392 1393 static int srpt_check_stop_free(struct se_cmd *cmd) 1394 { 1395 struct srpt_send_ioctx *ioctx = container_of(cmd, 1396 struct srpt_send_ioctx, cmd); 1397 1398 return target_put_sess_cmd(&ioctx->cmd); 1399 } 1400 1401 /** 1402 * srpt_handle_cmd - process a SRP_CMD information unit 1403 * @ch: SRPT RDMA channel. 1404 * @recv_ioctx: Receive I/O context. 1405 * @send_ioctx: Send I/O context. 1406 */ 1407 static void srpt_handle_cmd(struct srpt_rdma_ch *ch, 1408 struct srpt_recv_ioctx *recv_ioctx, 1409 struct srpt_send_ioctx *send_ioctx) 1410 { 1411 struct se_cmd *cmd; 1412 struct srp_cmd *srp_cmd; 1413 struct scatterlist *sg = NULL; 1414 unsigned sg_cnt = 0; 1415 u64 data_len; 1416 enum dma_data_direction dir; 1417 int rc; 1418 1419 BUG_ON(!send_ioctx); 1420 1421 srp_cmd = recv_ioctx->ioctx.buf; 1422 cmd = &send_ioctx->cmd; 1423 cmd->tag = srp_cmd->tag; 1424 1425 switch (srp_cmd->task_attr) { 1426 case SRP_CMD_SIMPLE_Q: 1427 cmd->sam_task_attr = TCM_SIMPLE_TAG; 1428 break; 1429 case SRP_CMD_ORDERED_Q: 1430 default: 1431 cmd->sam_task_attr = TCM_ORDERED_TAG; 1432 break; 1433 case SRP_CMD_HEAD_OF_Q: 1434 cmd->sam_task_attr = TCM_HEAD_TAG; 1435 break; 1436 case SRP_CMD_ACA: 1437 cmd->sam_task_attr = TCM_ACA_TAG; 1438 break; 1439 } 1440 1441 rc = srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &sg, &sg_cnt, 1442 &data_len); 1443 if (rc) { 1444 if (rc != -EAGAIN) { 1445 pr_err("0x%llx: parsing SRP descriptor table failed.\n", 1446 srp_cmd->tag); 1447 } 1448 goto release_ioctx; 1449 } 1450 1451 rc = target_submit_cmd_map_sgls(cmd, ch->sess, srp_cmd->cdb, 1452 &send_ioctx->sense_data[0], 1453 scsilun_to_int(&srp_cmd->lun), data_len, 1454 TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF, 1455 sg, sg_cnt, NULL, 0, NULL, 0); 1456 if (rc != 0) { 1457 pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc, 1458 srp_cmd->tag); 1459 goto release_ioctx; 1460 } 1461 return; 1462 1463 release_ioctx: 1464 send_ioctx->state = SRPT_STATE_DONE; 1465 srpt_release_cmd(cmd); 1466 } 1467 1468 static int srp_tmr_to_tcm(int fn) 1469 { 1470 switch (fn) { 1471 case SRP_TSK_ABORT_TASK: 1472 return TMR_ABORT_TASK; 1473 case SRP_TSK_ABORT_TASK_SET: 1474 return TMR_ABORT_TASK_SET; 1475 case SRP_TSK_CLEAR_TASK_SET: 1476 return TMR_CLEAR_TASK_SET; 1477 case SRP_TSK_LUN_RESET: 1478 return TMR_LUN_RESET; 1479 case SRP_TSK_CLEAR_ACA: 1480 return TMR_CLEAR_ACA; 1481 default: 1482 return -1; 1483 } 1484 } 1485 1486 /** 1487 * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit 1488 * @ch: SRPT RDMA channel. 1489 * @recv_ioctx: Receive I/O context. 1490 * @send_ioctx: Send I/O context. 1491 * 1492 * Returns 0 if and only if the request will be processed by the target core. 1493 * 1494 * For more information about SRP_TSK_MGMT information units, see also section 1495 * 6.7 in the SRP r16a document. 1496 */ 1497 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch, 1498 struct srpt_recv_ioctx *recv_ioctx, 1499 struct srpt_send_ioctx *send_ioctx) 1500 { 1501 struct srp_tsk_mgmt *srp_tsk; 1502 struct se_cmd *cmd; 1503 struct se_session *sess = ch->sess; 1504 int tcm_tmr; 1505 int rc; 1506 1507 BUG_ON(!send_ioctx); 1508 1509 srp_tsk = recv_ioctx->ioctx.buf; 1510 cmd = &send_ioctx->cmd; 1511 1512 pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n", 1513 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch, 1514 ch->sess); 1515 1516 srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT); 1517 send_ioctx->cmd.tag = srp_tsk->tag; 1518 tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func); 1519 rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, 1520 scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr, 1521 GFP_KERNEL, srp_tsk->task_tag, 1522 TARGET_SCF_ACK_KREF); 1523 if (rc != 0) { 1524 send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED; 1525 goto fail; 1526 } 1527 return; 1528 fail: 1529 transport_send_check_condition_and_sense(cmd, 0, 0); // XXX: 1530 } 1531 1532 /** 1533 * srpt_handle_new_iu - process a newly received information unit 1534 * @ch: RDMA channel through which the information unit has been received. 1535 * @recv_ioctx: Receive I/O context associated with the information unit. 1536 */ 1537 static bool 1538 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx) 1539 { 1540 struct srpt_send_ioctx *send_ioctx = NULL; 1541 struct srp_cmd *srp_cmd; 1542 bool res = false; 1543 u8 opcode; 1544 1545 BUG_ON(!ch); 1546 BUG_ON(!recv_ioctx); 1547 1548 if (unlikely(ch->state == CH_CONNECTING)) 1549 goto push; 1550 1551 ib_dma_sync_single_for_cpu(ch->sport->sdev->device, 1552 recv_ioctx->ioctx.dma, srp_max_req_size, 1553 DMA_FROM_DEVICE); 1554 1555 srp_cmd = recv_ioctx->ioctx.buf; 1556 opcode = srp_cmd->opcode; 1557 if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) { 1558 send_ioctx = srpt_get_send_ioctx(ch); 1559 if (unlikely(!send_ioctx)) 1560 goto push; 1561 } 1562 1563 if (!list_empty(&recv_ioctx->wait_list)) { 1564 WARN_ON_ONCE(!ch->processing_wait_list); 1565 list_del_init(&recv_ioctx->wait_list); 1566 } 1567 1568 switch (opcode) { 1569 case SRP_CMD: 1570 srpt_handle_cmd(ch, recv_ioctx, send_ioctx); 1571 break; 1572 case SRP_TSK_MGMT: 1573 srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx); 1574 break; 1575 case SRP_I_LOGOUT: 1576 pr_err("Not yet implemented: SRP_I_LOGOUT\n"); 1577 break; 1578 case SRP_CRED_RSP: 1579 pr_debug("received SRP_CRED_RSP\n"); 1580 break; 1581 case SRP_AER_RSP: 1582 pr_debug("received SRP_AER_RSP\n"); 1583 break; 1584 case SRP_RSP: 1585 pr_err("Received SRP_RSP\n"); 1586 break; 1587 default: 1588 pr_err("received IU with unknown opcode 0x%x\n", opcode); 1589 break; 1590 } 1591 1592 srpt_post_recv(ch->sport->sdev, ch, recv_ioctx); 1593 res = true; 1594 1595 out: 1596 return res; 1597 1598 push: 1599 if (list_empty(&recv_ioctx->wait_list)) { 1600 WARN_ON_ONCE(ch->processing_wait_list); 1601 list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list); 1602 } 1603 goto out; 1604 } 1605 1606 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc) 1607 { 1608 struct srpt_rdma_ch *ch = cq->cq_context; 1609 struct srpt_recv_ioctx *ioctx = 1610 container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe); 1611 1612 if (wc->status == IB_WC_SUCCESS) { 1613 int req_lim; 1614 1615 req_lim = atomic_dec_return(&ch->req_lim); 1616 if (unlikely(req_lim < 0)) 1617 pr_err("req_lim = %d < 0\n", req_lim); 1618 srpt_handle_new_iu(ch, ioctx); 1619 } else { 1620 pr_info_ratelimited("receiving failed for ioctx %p with status %d\n", 1621 ioctx, wc->status); 1622 } 1623 } 1624 1625 /* 1626 * This function must be called from the context in which RDMA completions are 1627 * processed because it accesses the wait list without protection against 1628 * access from other threads. 1629 */ 1630 static void srpt_process_wait_list(struct srpt_rdma_ch *ch) 1631 { 1632 struct srpt_recv_ioctx *recv_ioctx, *tmp; 1633 1634 WARN_ON_ONCE(ch->state == CH_CONNECTING); 1635 1636 if (list_empty(&ch->cmd_wait_list)) 1637 return; 1638 1639 WARN_ON_ONCE(ch->processing_wait_list); 1640 ch->processing_wait_list = true; 1641 list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list, 1642 wait_list) { 1643 if (!srpt_handle_new_iu(ch, recv_ioctx)) 1644 break; 1645 } 1646 ch->processing_wait_list = false; 1647 } 1648 1649 /** 1650 * srpt_send_done - send completion callback 1651 * @cq: Completion queue. 1652 * @wc: Work completion. 1653 * 1654 * Note: Although this has not yet been observed during tests, at least in 1655 * theory it is possible that the srpt_get_send_ioctx() call invoked by 1656 * srpt_handle_new_iu() fails. This is possible because the req_lim_delta 1657 * value in each response is set to one, and it is possible that this response 1658 * makes the initiator send a new request before the send completion for that 1659 * response has been processed. This could e.g. happen if the call to 1660 * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or 1661 * if IB retransmission causes generation of the send completion to be 1662 * delayed. Incoming information units for which srpt_get_send_ioctx() fails 1663 * are queued on cmd_wait_list. The code below processes these delayed 1664 * requests one at a time. 1665 */ 1666 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc) 1667 { 1668 struct srpt_rdma_ch *ch = cq->cq_context; 1669 struct srpt_send_ioctx *ioctx = 1670 container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe); 1671 enum srpt_command_state state; 1672 1673 state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE); 1674 1675 WARN_ON(state != SRPT_STATE_CMD_RSP_SENT && 1676 state != SRPT_STATE_MGMT_RSP_SENT); 1677 1678 atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail); 1679 1680 if (wc->status != IB_WC_SUCCESS) 1681 pr_info("sending response for ioctx 0x%p failed" 1682 " with status %d\n", ioctx, wc->status); 1683 1684 if (state != SRPT_STATE_DONE) { 1685 transport_generic_free_cmd(&ioctx->cmd, 0); 1686 } else { 1687 pr_err("IB completion has been received too late for" 1688 " wr_id = %u.\n", ioctx->ioctx.index); 1689 } 1690 1691 srpt_process_wait_list(ch); 1692 } 1693 1694 /** 1695 * srpt_create_ch_ib - create receive and send completion queues 1696 * @ch: SRPT RDMA channel. 1697 */ 1698 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch) 1699 { 1700 struct ib_qp_init_attr *qp_init; 1701 struct srpt_port *sport = ch->sport; 1702 struct srpt_device *sdev = sport->sdev; 1703 const struct ib_device_attr *attrs = &sdev->device->attrs; 1704 int sq_size = sport->port_attrib.srp_sq_size; 1705 int i, ret; 1706 1707 WARN_ON(ch->rq_size < 1); 1708 1709 ret = -ENOMEM; 1710 qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL); 1711 if (!qp_init) 1712 goto out; 1713 1714 retry: 1715 ch->cq = ib_alloc_cq(sdev->device, ch, ch->rq_size + sq_size, 1716 0 /* XXX: spread CQs */, IB_POLL_WORKQUEUE); 1717 if (IS_ERR(ch->cq)) { 1718 ret = PTR_ERR(ch->cq); 1719 pr_err("failed to create CQ cqe= %d ret= %d\n", 1720 ch->rq_size + sq_size, ret); 1721 goto out; 1722 } 1723 1724 qp_init->qp_context = (void *)ch; 1725 qp_init->event_handler 1726 = (void(*)(struct ib_event *, void*))srpt_qp_event; 1727 qp_init->send_cq = ch->cq; 1728 qp_init->recv_cq = ch->cq; 1729 qp_init->sq_sig_type = IB_SIGNAL_REQ_WR; 1730 qp_init->qp_type = IB_QPT_RC; 1731 /* 1732 * We divide up our send queue size into half SEND WRs to send the 1733 * completions, and half R/W contexts to actually do the RDMA 1734 * READ/WRITE transfers. Note that we need to allocate CQ slots for 1735 * both both, as RDMA contexts will also post completions for the 1736 * RDMA READ case. 1737 */ 1738 qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr); 1739 qp_init->cap.max_rdma_ctxs = sq_size / 2; 1740 qp_init->cap.max_send_sge = min(attrs->max_sge, SRPT_MAX_SG_PER_WQE); 1741 qp_init->port_num = ch->sport->port; 1742 if (sdev->use_srq) { 1743 qp_init->srq = sdev->srq; 1744 } else { 1745 qp_init->cap.max_recv_wr = ch->rq_size; 1746 qp_init->cap.max_recv_sge = qp_init->cap.max_send_sge; 1747 } 1748 1749 ch->qp = ib_create_qp(sdev->pd, qp_init); 1750 if (IS_ERR(ch->qp)) { 1751 ret = PTR_ERR(ch->qp); 1752 if (ret == -ENOMEM) { 1753 sq_size /= 2; 1754 if (sq_size >= MIN_SRPT_SQ_SIZE) { 1755 ib_destroy_cq(ch->cq); 1756 goto retry; 1757 } 1758 } 1759 pr_err("failed to create_qp ret= %d\n", ret); 1760 goto err_destroy_cq; 1761 } 1762 1763 atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr); 1764 1765 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n", 1766 __func__, ch->cq->cqe, qp_init->cap.max_send_sge, 1767 qp_init->cap.max_send_wr, ch); 1768 1769 ret = srpt_init_ch_qp(ch, ch->qp); 1770 if (ret) 1771 goto err_destroy_qp; 1772 1773 if (!sdev->use_srq) 1774 for (i = 0; i < ch->rq_size; i++) 1775 srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]); 1776 1777 out: 1778 kfree(qp_init); 1779 return ret; 1780 1781 err_destroy_qp: 1782 ib_destroy_qp(ch->qp); 1783 err_destroy_cq: 1784 ib_free_cq(ch->cq); 1785 goto out; 1786 } 1787 1788 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch) 1789 { 1790 ib_destroy_qp(ch->qp); 1791 ib_free_cq(ch->cq); 1792 } 1793 1794 /** 1795 * srpt_close_ch - close a RDMA channel 1796 * @ch: SRPT RDMA channel. 1797 * 1798 * Make sure all resources associated with the channel will be deallocated at 1799 * an appropriate time. 1800 * 1801 * Returns true if and only if the channel state has been modified into 1802 * CH_DRAINING. 1803 */ 1804 static bool srpt_close_ch(struct srpt_rdma_ch *ch) 1805 { 1806 int ret; 1807 1808 if (!srpt_set_ch_state(ch, CH_DRAINING)) { 1809 pr_debug("%s-%d: already closed\n", ch->sess_name, 1810 ch->qp->qp_num); 1811 return false; 1812 } 1813 1814 kref_get(&ch->kref); 1815 1816 ret = srpt_ch_qp_err(ch); 1817 if (ret < 0) 1818 pr_err("%s-%d: changing queue pair into error state failed: %d\n", 1819 ch->sess_name, ch->qp->qp_num, ret); 1820 1821 ret = srpt_zerolength_write(ch); 1822 if (ret < 0) { 1823 pr_err("%s-%d: queuing zero-length write failed: %d\n", 1824 ch->sess_name, ch->qp->qp_num, ret); 1825 if (srpt_set_ch_state(ch, CH_DISCONNECTED)) 1826 schedule_work(&ch->release_work); 1827 else 1828 WARN_ON_ONCE(true); 1829 } 1830 1831 kref_put(&ch->kref, srpt_free_ch); 1832 1833 return true; 1834 } 1835 1836 /* 1837 * Change the channel state into CH_DISCONNECTING. If a channel has not yet 1838 * reached the connected state, close it. If a channel is in the connected 1839 * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is 1840 * the responsibility of the caller to ensure that this function is not 1841 * invoked concurrently with the code that accepts a connection. This means 1842 * that this function must either be invoked from inside a CM callback 1843 * function or that it must be invoked with the srpt_port.mutex held. 1844 */ 1845 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch) 1846 { 1847 int ret; 1848 1849 if (!srpt_set_ch_state(ch, CH_DISCONNECTING)) 1850 return -ENOTCONN; 1851 1852 ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0); 1853 if (ret < 0) 1854 ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0); 1855 1856 if (ret < 0 && srpt_close_ch(ch)) 1857 ret = 0; 1858 1859 return ret; 1860 } 1861 1862 static bool srpt_ch_closed(struct srpt_port *sport, struct srpt_rdma_ch *ch) 1863 { 1864 struct srpt_nexus *nexus; 1865 struct srpt_rdma_ch *ch2; 1866 bool res = true; 1867 1868 rcu_read_lock(); 1869 list_for_each_entry(nexus, &sport->nexus_list, entry) { 1870 list_for_each_entry(ch2, &nexus->ch_list, list) { 1871 if (ch2 == ch) { 1872 res = false; 1873 goto done; 1874 } 1875 } 1876 } 1877 done: 1878 rcu_read_unlock(); 1879 1880 return res; 1881 } 1882 1883 /* Send DREQ and wait for DREP. */ 1884 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch) 1885 { 1886 struct srpt_port *sport = ch->sport; 1887 1888 pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num, 1889 ch->state); 1890 1891 mutex_lock(&sport->mutex); 1892 srpt_disconnect_ch(ch); 1893 mutex_unlock(&sport->mutex); 1894 1895 while (wait_event_timeout(sport->ch_releaseQ, srpt_ch_closed(sport, ch), 1896 5 * HZ) == 0) 1897 pr_info("%s(%s-%d state %d): still waiting ...\n", __func__, 1898 ch->sess_name, ch->qp->qp_num, ch->state); 1899 1900 } 1901 1902 static void __srpt_close_all_ch(struct srpt_port *sport) 1903 { 1904 struct srpt_nexus *nexus; 1905 struct srpt_rdma_ch *ch; 1906 1907 lockdep_assert_held(&sport->mutex); 1908 1909 list_for_each_entry(nexus, &sport->nexus_list, entry) { 1910 list_for_each_entry(ch, &nexus->ch_list, list) { 1911 if (srpt_disconnect_ch(ch) >= 0) 1912 pr_info("Closing channel %s-%d because target %s_%d has been disabled\n", 1913 ch->sess_name, ch->qp->qp_num, 1914 sport->sdev->device->name, sport->port); 1915 srpt_close_ch(ch); 1916 } 1917 } 1918 } 1919 1920 /* 1921 * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if 1922 * it does not yet exist. 1923 */ 1924 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport, 1925 const u8 i_port_id[16], 1926 const u8 t_port_id[16]) 1927 { 1928 struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n; 1929 1930 for (;;) { 1931 mutex_lock(&sport->mutex); 1932 list_for_each_entry(n, &sport->nexus_list, entry) { 1933 if (memcmp(n->i_port_id, i_port_id, 16) == 0 && 1934 memcmp(n->t_port_id, t_port_id, 16) == 0) { 1935 nexus = n; 1936 break; 1937 } 1938 } 1939 if (!nexus && tmp_nexus) { 1940 list_add_tail_rcu(&tmp_nexus->entry, 1941 &sport->nexus_list); 1942 swap(nexus, tmp_nexus); 1943 } 1944 mutex_unlock(&sport->mutex); 1945 1946 if (nexus) 1947 break; 1948 tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL); 1949 if (!tmp_nexus) { 1950 nexus = ERR_PTR(-ENOMEM); 1951 break; 1952 } 1953 INIT_LIST_HEAD(&tmp_nexus->ch_list); 1954 memcpy(tmp_nexus->i_port_id, i_port_id, 16); 1955 memcpy(tmp_nexus->t_port_id, t_port_id, 16); 1956 } 1957 1958 kfree(tmp_nexus); 1959 1960 return nexus; 1961 } 1962 1963 static void srpt_set_enabled(struct srpt_port *sport, bool enabled) 1964 __must_hold(&sport->mutex) 1965 { 1966 lockdep_assert_held(&sport->mutex); 1967 1968 if (sport->enabled == enabled) 1969 return; 1970 sport->enabled = enabled; 1971 if (!enabled) 1972 __srpt_close_all_ch(sport); 1973 } 1974 1975 static void srpt_free_ch(struct kref *kref) 1976 { 1977 struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref); 1978 1979 kfree_rcu(ch, rcu); 1980 } 1981 1982 static void srpt_release_channel_work(struct work_struct *w) 1983 { 1984 struct srpt_rdma_ch *ch; 1985 struct srpt_device *sdev; 1986 struct srpt_port *sport; 1987 struct se_session *se_sess; 1988 1989 ch = container_of(w, struct srpt_rdma_ch, release_work); 1990 pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num); 1991 1992 sdev = ch->sport->sdev; 1993 BUG_ON(!sdev); 1994 1995 se_sess = ch->sess; 1996 BUG_ON(!se_sess); 1997 1998 target_sess_cmd_list_set_waiting(se_sess); 1999 target_wait_for_sess_cmds(se_sess); 2000 2001 transport_deregister_session_configfs(se_sess); 2002 transport_deregister_session(se_sess); 2003 ch->sess = NULL; 2004 2005 ib_destroy_cm_id(ch->ib_cm.cm_id); 2006 2007 srpt_destroy_ch_ib(ch); 2008 2009 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring, 2010 ch->sport->sdev, ch->rq_size, 2011 ch->max_rsp_size, DMA_TO_DEVICE); 2012 2013 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring, 2014 sdev, ch->rq_size, 2015 srp_max_req_size, DMA_FROM_DEVICE); 2016 2017 sport = ch->sport; 2018 mutex_lock(&sport->mutex); 2019 list_del_rcu(&ch->list); 2020 mutex_unlock(&sport->mutex); 2021 2022 wake_up(&sport->ch_releaseQ); 2023 2024 kref_put(&ch->kref, srpt_free_ch); 2025 } 2026 2027 /** 2028 * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED 2029 * @cm_id: IB/CM connection identifier. 2030 * @port_num: Port through which the IB/CM REQ message was received. 2031 * @pkey: P_Key of the incoming connection. 2032 * @req: SRP login request. 2033 * @src_addr: GID of the port that submitted the login request. 2034 * 2035 * Ownership of the cm_id is transferred to the target session if this 2036 * functions returns zero. Otherwise the caller remains the owner of cm_id. 2037 */ 2038 static int srpt_cm_req_recv(struct ib_cm_id *cm_id, 2039 u8 port_num, __be16 pkey, 2040 const struct srp_login_req *req, 2041 const char *src_addr) 2042 { 2043 struct srpt_device *sdev = cm_id->context; 2044 struct srpt_port *sport = &sdev->port[port_num - 1]; 2045 struct srpt_nexus *nexus; 2046 struct srp_login_rsp *rsp = NULL; 2047 struct srp_login_rej *rej = NULL; 2048 struct ib_cm_rep_param *rep_param = NULL; 2049 struct srpt_rdma_ch *ch; 2050 char i_port_id[36]; 2051 u32 it_iu_len; 2052 int i, ret; 2053 2054 WARN_ON_ONCE(irqs_disabled()); 2055 2056 if (WARN_ON(!sdev || !req)) 2057 return -EINVAL; 2058 2059 it_iu_len = be32_to_cpu(req->req_it_iu_len); 2060 2061 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", 2062 req->initiator_port_id, req->target_port_id, it_iu_len, 2063 port_num, &sport->gid, be16_to_cpu(pkey)); 2064 2065 nexus = srpt_get_nexus(sport, req->initiator_port_id, 2066 req->target_port_id); 2067 if (IS_ERR(nexus)) { 2068 ret = PTR_ERR(nexus); 2069 goto out; 2070 } 2071 2072 ret = -ENOMEM; 2073 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL); 2074 rej = kzalloc(sizeof(*rej), GFP_KERNEL); 2075 rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL); 2076 if (!rsp || !rej || !rep_param) 2077 goto out; 2078 2079 ret = -EINVAL; 2080 if (it_iu_len > srp_max_req_size || it_iu_len < 64) { 2081 rej->reason = cpu_to_be32( 2082 SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE); 2083 pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n", 2084 it_iu_len, 64, srp_max_req_size); 2085 goto reject; 2086 } 2087 2088 if (!sport->enabled) { 2089 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2090 pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n", 2091 sport->sdev->device->name, port_num); 2092 goto reject; 2093 } 2094 2095 if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid) 2096 || *(__be64 *)(req->target_port_id + 8) != 2097 cpu_to_be64(srpt_service_guid)) { 2098 rej->reason = cpu_to_be32( 2099 SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL); 2100 pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n"); 2101 goto reject; 2102 } 2103 2104 ret = -ENOMEM; 2105 ch = kzalloc(sizeof(*ch), GFP_KERNEL); 2106 if (!ch) { 2107 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2108 pr_err("rejected SRP_LOGIN_REQ because out of memory.\n"); 2109 goto reject; 2110 } 2111 2112 kref_init(&ch->kref); 2113 ch->pkey = be16_to_cpu(pkey); 2114 ch->nexus = nexus; 2115 ch->zw_cqe.done = srpt_zerolength_write_done; 2116 INIT_WORK(&ch->release_work, srpt_release_channel_work); 2117 ch->sport = sport; 2118 ch->ib_cm.cm_id = cm_id; 2119 cm_id->context = ch; 2120 /* 2121 * ch->rq_size should be at least as large as the initiator queue 2122 * depth to avoid that the initiator driver has to report QUEUE_FULL 2123 * to the SCSI mid-layer. 2124 */ 2125 ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr); 2126 spin_lock_init(&ch->spinlock); 2127 ch->state = CH_CONNECTING; 2128 INIT_LIST_HEAD(&ch->cmd_wait_list); 2129 ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size; 2130 2131 ch->ioctx_ring = (struct srpt_send_ioctx **) 2132 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size, 2133 sizeof(*ch->ioctx_ring[0]), 2134 ch->max_rsp_size, DMA_TO_DEVICE); 2135 if (!ch->ioctx_ring) { 2136 pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n"); 2137 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2138 goto free_ch; 2139 } 2140 2141 INIT_LIST_HEAD(&ch->free_list); 2142 for (i = 0; i < ch->rq_size; i++) { 2143 ch->ioctx_ring[i]->ch = ch; 2144 list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list); 2145 } 2146 if (!sdev->use_srq) { 2147 ch->ioctx_recv_ring = (struct srpt_recv_ioctx **) 2148 srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size, 2149 sizeof(*ch->ioctx_recv_ring[0]), 2150 srp_max_req_size, 2151 DMA_FROM_DEVICE); 2152 if (!ch->ioctx_recv_ring) { 2153 pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n"); 2154 rej->reason = 2155 cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2156 goto free_ring; 2157 } 2158 for (i = 0; i < ch->rq_size; i++) 2159 INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list); 2160 } 2161 2162 ret = srpt_create_ch_ib(ch); 2163 if (ret) { 2164 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2165 pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n"); 2166 goto free_recv_ring; 2167 } 2168 2169 strlcpy(ch->sess_name, src_addr, sizeof(ch->sess_name)); 2170 snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx", 2171 be64_to_cpu(*(__be64 *)nexus->i_port_id), 2172 be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8))); 2173 2174 pr_debug("registering session %s\n", ch->sess_name); 2175 2176 if (sport->port_guid_tpg.se_tpg_wwn) 2177 ch->sess = target_alloc_session(&sport->port_guid_tpg, 0, 0, 2178 TARGET_PROT_NORMAL, 2179 ch->sess_name, ch, NULL); 2180 if (sport->port_gid_tpg.se_tpg_wwn && IS_ERR_OR_NULL(ch->sess)) 2181 ch->sess = target_alloc_session(&sport->port_gid_tpg, 0, 0, 2182 TARGET_PROT_NORMAL, i_port_id, ch, 2183 NULL); 2184 /* Retry without leading "0x" */ 2185 if (sport->port_gid_tpg.se_tpg_wwn && IS_ERR_OR_NULL(ch->sess)) 2186 ch->sess = target_alloc_session(&sport->port_gid_tpg, 0, 0, 2187 TARGET_PROT_NORMAL, 2188 i_port_id + 2, ch, NULL); 2189 if (IS_ERR_OR_NULL(ch->sess)) { 2190 ret = PTR_ERR(ch->sess); 2191 pr_info("Rejected login for initiator %s: ret = %d.\n", 2192 ch->sess_name, ret); 2193 rej->reason = cpu_to_be32(ret == -ENOMEM ? 2194 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES : 2195 SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED); 2196 goto reject; 2197 } 2198 2199 mutex_lock(&sport->mutex); 2200 2201 if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) { 2202 struct srpt_rdma_ch *ch2; 2203 2204 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN; 2205 2206 list_for_each_entry(ch2, &nexus->ch_list, list) { 2207 if (srpt_disconnect_ch(ch2) < 0) 2208 continue; 2209 pr_info("Relogin - closed existing channel %s\n", 2210 ch2->sess_name); 2211 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_TERMINATED; 2212 } 2213 } else { 2214 rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED; 2215 } 2216 2217 list_add_tail_rcu(&ch->list, &nexus->ch_list); 2218 2219 if (!sport->enabled) { 2220 rej->reason = cpu_to_be32( 2221 SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2222 pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n", 2223 sdev->device->name, port_num); 2224 mutex_unlock(&sport->mutex); 2225 goto reject; 2226 } 2227 2228 mutex_unlock(&sport->mutex); 2229 2230 ret = srpt_ch_qp_rtr(ch, ch->qp); 2231 if (ret) { 2232 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2233 pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n", 2234 ret); 2235 goto destroy_ib; 2236 } 2237 2238 pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess, 2239 ch->sess_name, ch); 2240 2241 /* create srp_login_response */ 2242 rsp->opcode = SRP_LOGIN_RSP; 2243 rsp->tag = req->tag; 2244 rsp->max_it_iu_len = req->req_it_iu_len; 2245 rsp->max_ti_iu_len = req->req_it_iu_len; 2246 ch->max_ti_iu_len = it_iu_len; 2247 rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT | 2248 SRP_BUF_FORMAT_INDIRECT); 2249 rsp->req_lim_delta = cpu_to_be32(ch->rq_size); 2250 atomic_set(&ch->req_lim, ch->rq_size); 2251 atomic_set(&ch->req_lim_delta, 0); 2252 2253 /* create cm reply */ 2254 rep_param->qp_num = ch->qp->qp_num; 2255 rep_param->private_data = (void *)rsp; 2256 rep_param->private_data_len = sizeof(*rsp); 2257 rep_param->rnr_retry_count = 7; 2258 rep_param->flow_control = 1; 2259 rep_param->failover_accepted = 0; 2260 rep_param->srq = 1; 2261 rep_param->responder_resources = 4; 2262 rep_param->initiator_depth = 4; 2263 2264 /* 2265 * Hold the sport mutex while accepting a connection to avoid that 2266 * srpt_disconnect_ch() is invoked concurrently with this code. 2267 */ 2268 mutex_lock(&sport->mutex); 2269 if (sport->enabled && ch->state == CH_CONNECTING) 2270 ret = ib_send_cm_rep(cm_id, rep_param); 2271 else 2272 ret = -EINVAL; 2273 mutex_unlock(&sport->mutex); 2274 2275 switch (ret) { 2276 case 0: 2277 break; 2278 case -EINVAL: 2279 goto reject; 2280 default: 2281 rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES); 2282 pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n", 2283 ret); 2284 goto reject; 2285 } 2286 2287 goto out; 2288 2289 destroy_ib: 2290 srpt_destroy_ch_ib(ch); 2291 2292 free_recv_ring: 2293 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring, 2294 ch->sport->sdev, ch->rq_size, 2295 srp_max_req_size, DMA_FROM_DEVICE); 2296 2297 free_ring: 2298 srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring, 2299 ch->sport->sdev, ch->rq_size, 2300 ch->max_rsp_size, DMA_TO_DEVICE); 2301 free_ch: 2302 cm_id->context = NULL; 2303 kfree(ch); 2304 ch = NULL; 2305 2306 WARN_ON_ONCE(ret == 0); 2307 2308 reject: 2309 pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason)); 2310 rej->opcode = SRP_LOGIN_REJ; 2311 rej->tag = req->tag; 2312 rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT | 2313 SRP_BUF_FORMAT_INDIRECT); 2314 2315 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0, 2316 (void *)rej, sizeof(*rej)); 2317 2318 out: 2319 kfree(rep_param); 2320 kfree(rsp); 2321 kfree(rej); 2322 2323 return ret; 2324 } 2325 2326 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id, 2327 struct ib_cm_req_event_param *param, 2328 void *private_data) 2329 { 2330 char sguid[40]; 2331 2332 srpt_format_guid(sguid, sizeof(sguid), 2333 ¶m->primary_path->dgid.global.interface_id); 2334 2335 return srpt_cm_req_recv(cm_id, param->port, param->primary_path->pkey, 2336 private_data, sguid); 2337 } 2338 2339 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch, 2340 enum ib_cm_rej_reason reason, 2341 const u8 *private_data, 2342 u8 private_data_len) 2343 { 2344 char *priv = NULL; 2345 int i; 2346 2347 if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1, 2348 GFP_KERNEL))) { 2349 for (i = 0; i < private_data_len; i++) 2350 sprintf(priv + 3 * i, " %02x", private_data[i]); 2351 } 2352 pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n", 2353 ch->sess_name, ch->qp->qp_num, reason, private_data_len ? 2354 "; private data" : "", priv ? priv : " (?)"); 2355 kfree(priv); 2356 } 2357 2358 /** 2359 * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event 2360 * @ch: SRPT RDMA channel. 2361 * 2362 * An IB_CM_RTU_RECEIVED message indicates that the connection is established 2363 * and that the recipient may begin transmitting (RTU = ready to use). 2364 */ 2365 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch) 2366 { 2367 int ret; 2368 2369 ret = srpt_ch_qp_rts(ch, ch->qp); 2370 if (ret < 0) { 2371 pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name, 2372 ch->qp->qp_num); 2373 srpt_close_ch(ch); 2374 return; 2375 } 2376 2377 /* 2378 * Note: calling srpt_close_ch() if the transition to the LIVE state 2379 * fails is not necessary since that means that that function has 2380 * already been invoked from another thread. 2381 */ 2382 if (!srpt_set_ch_state(ch, CH_LIVE)) { 2383 pr_err("%s-%d: channel transition to LIVE state failed\n", 2384 ch->sess_name, ch->qp->qp_num); 2385 return; 2386 } 2387 2388 /* Trigger wait list processing. */ 2389 ret = srpt_zerolength_write(ch); 2390 WARN_ONCE(ret < 0, "%d\n", ret); 2391 } 2392 2393 /** 2394 * srpt_cm_handler - IB connection manager callback function 2395 * @cm_id: IB/CM connection identifier. 2396 * @event: IB/CM event. 2397 * 2398 * A non-zero return value will cause the caller destroy the CM ID. 2399 * 2400 * Note: srpt_cm_handler() must only return a non-zero value when transferring 2401 * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning 2402 * a non-zero value in any other case will trigger a race with the 2403 * ib_destroy_cm_id() call in srpt_release_channel(). 2404 */ 2405 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 2406 { 2407 struct srpt_rdma_ch *ch = cm_id->context; 2408 int ret; 2409 2410 ret = 0; 2411 switch (event->event) { 2412 case IB_CM_REQ_RECEIVED: 2413 ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd, 2414 event->private_data); 2415 break; 2416 case IB_CM_REJ_RECEIVED: 2417 srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason, 2418 event->private_data, 2419 IB_CM_REJ_PRIVATE_DATA_SIZE); 2420 break; 2421 case IB_CM_RTU_RECEIVED: 2422 case IB_CM_USER_ESTABLISHED: 2423 srpt_cm_rtu_recv(ch); 2424 break; 2425 case IB_CM_DREQ_RECEIVED: 2426 srpt_disconnect_ch(ch); 2427 break; 2428 case IB_CM_DREP_RECEIVED: 2429 pr_info("Received CM DREP message for ch %s-%d.\n", 2430 ch->sess_name, ch->qp->qp_num); 2431 srpt_close_ch(ch); 2432 break; 2433 case IB_CM_TIMEWAIT_EXIT: 2434 pr_info("Received CM TimeWait exit for ch %s-%d.\n", 2435 ch->sess_name, ch->qp->qp_num); 2436 srpt_close_ch(ch); 2437 break; 2438 case IB_CM_REP_ERROR: 2439 pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name, 2440 ch->qp->qp_num); 2441 break; 2442 case IB_CM_DREQ_ERROR: 2443 pr_info("Received CM DREQ ERROR event.\n"); 2444 break; 2445 case IB_CM_MRA_RECEIVED: 2446 pr_info("Received CM MRA event\n"); 2447 break; 2448 default: 2449 pr_err("received unrecognized CM event %d\n", event->event); 2450 break; 2451 } 2452 2453 return ret; 2454 } 2455 2456 static int srpt_write_pending_status(struct se_cmd *se_cmd) 2457 { 2458 struct srpt_send_ioctx *ioctx; 2459 2460 ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd); 2461 return ioctx->state == SRPT_STATE_NEED_DATA; 2462 } 2463 2464 /* 2465 * srpt_write_pending - Start data transfer from initiator to target (write). 2466 */ 2467 static int srpt_write_pending(struct se_cmd *se_cmd) 2468 { 2469 struct srpt_send_ioctx *ioctx = 2470 container_of(se_cmd, struct srpt_send_ioctx, cmd); 2471 struct srpt_rdma_ch *ch = ioctx->ch; 2472 struct ib_send_wr *first_wr = NULL, *bad_wr; 2473 struct ib_cqe *cqe = &ioctx->rdma_cqe; 2474 enum srpt_command_state new_state; 2475 int ret, i; 2476 2477 new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA); 2478 WARN_ON(new_state == SRPT_STATE_DONE); 2479 2480 if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) { 2481 pr_warn("%s: IB send queue full (needed %d)\n", 2482 __func__, ioctx->n_rdma); 2483 ret = -ENOMEM; 2484 goto out_undo; 2485 } 2486 2487 cqe->done = srpt_rdma_read_done; 2488 for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) { 2489 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 2490 2491 first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port, 2492 cqe, first_wr); 2493 cqe = NULL; 2494 } 2495 2496 ret = ib_post_send(ch->qp, first_wr, &bad_wr); 2497 if (ret) { 2498 pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n", 2499 __func__, ret, ioctx->n_rdma, 2500 atomic_read(&ch->sq_wr_avail)); 2501 goto out_undo; 2502 } 2503 2504 return 0; 2505 out_undo: 2506 atomic_add(ioctx->n_rdma, &ch->sq_wr_avail); 2507 return ret; 2508 } 2509 2510 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status) 2511 { 2512 switch (tcm_mgmt_status) { 2513 case TMR_FUNCTION_COMPLETE: 2514 return SRP_TSK_MGMT_SUCCESS; 2515 case TMR_FUNCTION_REJECTED: 2516 return SRP_TSK_MGMT_FUNC_NOT_SUPP; 2517 } 2518 return SRP_TSK_MGMT_FAILED; 2519 } 2520 2521 /** 2522 * srpt_queue_response - transmit the response to a SCSI command 2523 * @cmd: SCSI target command. 2524 * 2525 * Callback function called by the TCM core. Must not block since it can be 2526 * invoked on the context of the IB completion handler. 2527 */ 2528 static void srpt_queue_response(struct se_cmd *cmd) 2529 { 2530 struct srpt_send_ioctx *ioctx = 2531 container_of(cmd, struct srpt_send_ioctx, cmd); 2532 struct srpt_rdma_ch *ch = ioctx->ch; 2533 struct srpt_device *sdev = ch->sport->sdev; 2534 struct ib_send_wr send_wr, *first_wr = &send_wr, *bad_wr; 2535 struct ib_sge sge; 2536 enum srpt_command_state state; 2537 int resp_len, ret, i; 2538 u8 srp_tm_status; 2539 2540 BUG_ON(!ch); 2541 2542 state = ioctx->state; 2543 switch (state) { 2544 case SRPT_STATE_NEW: 2545 case SRPT_STATE_DATA_IN: 2546 ioctx->state = SRPT_STATE_CMD_RSP_SENT; 2547 break; 2548 case SRPT_STATE_MGMT: 2549 ioctx->state = SRPT_STATE_MGMT_RSP_SENT; 2550 break; 2551 default: 2552 WARN(true, "ch %p; cmd %d: unexpected command state %d\n", 2553 ch, ioctx->ioctx.index, ioctx->state); 2554 break; 2555 } 2556 2557 if (unlikely(WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) 2558 return; 2559 2560 /* For read commands, transfer the data to the initiator. */ 2561 if (ioctx->cmd.data_direction == DMA_FROM_DEVICE && 2562 ioctx->cmd.data_length && 2563 !ioctx->queue_status_only) { 2564 for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) { 2565 struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i]; 2566 2567 first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, 2568 ch->sport->port, NULL, first_wr); 2569 } 2570 } 2571 2572 if (state != SRPT_STATE_MGMT) 2573 resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag, 2574 cmd->scsi_status); 2575 else { 2576 srp_tm_status 2577 = tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response); 2578 resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status, 2579 ioctx->cmd.tag); 2580 } 2581 2582 atomic_inc(&ch->req_lim); 2583 2584 if (unlikely(atomic_sub_return(1 + ioctx->n_rdma, 2585 &ch->sq_wr_avail) < 0)) { 2586 pr_warn("%s: IB send queue full (needed %d)\n", 2587 __func__, ioctx->n_rdma); 2588 ret = -ENOMEM; 2589 goto out; 2590 } 2591 2592 ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len, 2593 DMA_TO_DEVICE); 2594 2595 sge.addr = ioctx->ioctx.dma; 2596 sge.length = resp_len; 2597 sge.lkey = sdev->lkey; 2598 2599 ioctx->ioctx.cqe.done = srpt_send_done; 2600 send_wr.next = NULL; 2601 send_wr.wr_cqe = &ioctx->ioctx.cqe; 2602 send_wr.sg_list = &sge; 2603 send_wr.num_sge = 1; 2604 send_wr.opcode = IB_WR_SEND; 2605 send_wr.send_flags = IB_SEND_SIGNALED; 2606 2607 ret = ib_post_send(ch->qp, first_wr, &bad_wr); 2608 if (ret < 0) { 2609 pr_err("%s: sending cmd response failed for tag %llu (%d)\n", 2610 __func__, ioctx->cmd.tag, ret); 2611 goto out; 2612 } 2613 2614 return; 2615 2616 out: 2617 atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail); 2618 atomic_dec(&ch->req_lim); 2619 srpt_set_cmd_state(ioctx, SRPT_STATE_DONE); 2620 target_put_sess_cmd(&ioctx->cmd); 2621 } 2622 2623 static int srpt_queue_data_in(struct se_cmd *cmd) 2624 { 2625 srpt_queue_response(cmd); 2626 return 0; 2627 } 2628 2629 static void srpt_queue_tm_rsp(struct se_cmd *cmd) 2630 { 2631 srpt_queue_response(cmd); 2632 } 2633 2634 static void srpt_aborted_task(struct se_cmd *cmd) 2635 { 2636 } 2637 2638 static int srpt_queue_status(struct se_cmd *cmd) 2639 { 2640 struct srpt_send_ioctx *ioctx; 2641 2642 ioctx = container_of(cmd, struct srpt_send_ioctx, cmd); 2643 BUG_ON(ioctx->sense_data != cmd->sense_buffer); 2644 if (cmd->se_cmd_flags & 2645 (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE)) 2646 WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION); 2647 ioctx->queue_status_only = true; 2648 srpt_queue_response(cmd); 2649 return 0; 2650 } 2651 2652 static void srpt_refresh_port_work(struct work_struct *work) 2653 { 2654 struct srpt_port *sport = container_of(work, struct srpt_port, work); 2655 2656 srpt_refresh_port(sport); 2657 } 2658 2659 static bool srpt_ch_list_empty(struct srpt_port *sport) 2660 { 2661 struct srpt_nexus *nexus; 2662 bool res = true; 2663 2664 rcu_read_lock(); 2665 list_for_each_entry(nexus, &sport->nexus_list, entry) 2666 if (!list_empty(&nexus->ch_list)) 2667 res = false; 2668 rcu_read_unlock(); 2669 2670 return res; 2671 } 2672 2673 /** 2674 * srpt_release_sport - disable login and wait for associated channels 2675 * @sport: SRPT HCA port. 2676 */ 2677 static int srpt_release_sport(struct srpt_port *sport) 2678 { 2679 struct srpt_nexus *nexus, *next_n; 2680 struct srpt_rdma_ch *ch; 2681 2682 WARN_ON_ONCE(irqs_disabled()); 2683 2684 mutex_lock(&sport->mutex); 2685 srpt_set_enabled(sport, false); 2686 mutex_unlock(&sport->mutex); 2687 2688 while (wait_event_timeout(sport->ch_releaseQ, 2689 srpt_ch_list_empty(sport), 5 * HZ) <= 0) { 2690 pr_info("%s_%d: waiting for session unregistration ...\n", 2691 sport->sdev->device->name, sport->port); 2692 rcu_read_lock(); 2693 list_for_each_entry(nexus, &sport->nexus_list, entry) { 2694 list_for_each_entry(ch, &nexus->ch_list, list) { 2695 pr_info("%s-%d: state %s\n", 2696 ch->sess_name, ch->qp->qp_num, 2697 get_ch_state_name(ch->state)); 2698 } 2699 } 2700 rcu_read_unlock(); 2701 } 2702 2703 mutex_lock(&sport->mutex); 2704 list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) { 2705 list_del(&nexus->entry); 2706 kfree_rcu(nexus, rcu); 2707 } 2708 mutex_unlock(&sport->mutex); 2709 2710 return 0; 2711 } 2712 2713 static struct se_wwn *__srpt_lookup_wwn(const char *name) 2714 { 2715 struct ib_device *dev; 2716 struct srpt_device *sdev; 2717 struct srpt_port *sport; 2718 int i; 2719 2720 list_for_each_entry(sdev, &srpt_dev_list, list) { 2721 dev = sdev->device; 2722 if (!dev) 2723 continue; 2724 2725 for (i = 0; i < dev->phys_port_cnt; i++) { 2726 sport = &sdev->port[i]; 2727 2728 if (strcmp(sport->port_guid, name) == 0) 2729 return &sport->port_guid_wwn; 2730 if (strcmp(sport->port_gid, name) == 0) 2731 return &sport->port_gid_wwn; 2732 } 2733 } 2734 2735 return NULL; 2736 } 2737 2738 static struct se_wwn *srpt_lookup_wwn(const char *name) 2739 { 2740 struct se_wwn *wwn; 2741 2742 spin_lock(&srpt_dev_lock); 2743 wwn = __srpt_lookup_wwn(name); 2744 spin_unlock(&srpt_dev_lock); 2745 2746 return wwn; 2747 } 2748 2749 static void srpt_free_srq(struct srpt_device *sdev) 2750 { 2751 if (!sdev->srq) 2752 return; 2753 2754 ib_destroy_srq(sdev->srq); 2755 srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev, 2756 sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE); 2757 sdev->srq = NULL; 2758 } 2759 2760 static int srpt_alloc_srq(struct srpt_device *sdev) 2761 { 2762 struct ib_srq_init_attr srq_attr = { 2763 .event_handler = srpt_srq_event, 2764 .srq_context = (void *)sdev, 2765 .attr.max_wr = sdev->srq_size, 2766 .attr.max_sge = 1, 2767 .srq_type = IB_SRQT_BASIC, 2768 }; 2769 struct ib_device *device = sdev->device; 2770 struct ib_srq *srq; 2771 int i; 2772 2773 WARN_ON_ONCE(sdev->srq); 2774 srq = ib_create_srq(sdev->pd, &srq_attr); 2775 if (IS_ERR(srq)) { 2776 pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq)); 2777 return PTR_ERR(srq); 2778 } 2779 2780 pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size, 2781 sdev->device->attrs.max_srq_wr, device->name); 2782 2783 sdev->ioctx_ring = (struct srpt_recv_ioctx **) 2784 srpt_alloc_ioctx_ring(sdev, sdev->srq_size, 2785 sizeof(*sdev->ioctx_ring[0]), 2786 srp_max_req_size, DMA_FROM_DEVICE); 2787 if (!sdev->ioctx_ring) { 2788 ib_destroy_srq(srq); 2789 return -ENOMEM; 2790 } 2791 2792 sdev->use_srq = true; 2793 sdev->srq = srq; 2794 2795 for (i = 0; i < sdev->srq_size; ++i) { 2796 INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list); 2797 srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]); 2798 } 2799 2800 return 0; 2801 } 2802 2803 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq) 2804 { 2805 struct ib_device *device = sdev->device; 2806 int ret = 0; 2807 2808 if (!use_srq) { 2809 srpt_free_srq(sdev); 2810 sdev->use_srq = false; 2811 } else if (use_srq && !sdev->srq) { 2812 ret = srpt_alloc_srq(sdev); 2813 } 2814 pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__, device->name, 2815 sdev->use_srq, ret); 2816 return ret; 2817 } 2818 2819 /** 2820 * srpt_add_one - InfiniBand device addition callback function 2821 * @device: Describes a HCA. 2822 */ 2823 static void srpt_add_one(struct ib_device *device) 2824 { 2825 struct srpt_device *sdev; 2826 struct srpt_port *sport; 2827 int i; 2828 2829 pr_debug("device = %p\n", device); 2830 2831 sdev = kzalloc(sizeof(*sdev), GFP_KERNEL); 2832 if (!sdev) 2833 goto err; 2834 2835 sdev->device = device; 2836 mutex_init(&sdev->sdev_mutex); 2837 2838 sdev->pd = ib_alloc_pd(device, 0); 2839 if (IS_ERR(sdev->pd)) 2840 goto free_dev; 2841 2842 sdev->lkey = sdev->pd->local_dma_lkey; 2843 2844 sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr); 2845 2846 srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq); 2847 2848 if (!srpt_service_guid) 2849 srpt_service_guid = be64_to_cpu(device->node_guid); 2850 2851 sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev); 2852 if (IS_ERR(sdev->cm_id)) 2853 goto err_ring; 2854 2855 /* print out target login information */ 2856 pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx," 2857 "pkey=ffff,service_id=%016llx\n", srpt_service_guid, 2858 srpt_service_guid, srpt_service_guid); 2859 2860 /* 2861 * We do not have a consistent service_id (ie. also id_ext of target_id) 2862 * to identify this target. We currently use the guid of the first HCA 2863 * in the system as service_id; therefore, the target_id will change 2864 * if this HCA is gone bad and replaced by different HCA 2865 */ 2866 if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0)) 2867 goto err_cm; 2868 2869 INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device, 2870 srpt_event_handler); 2871 ib_register_event_handler(&sdev->event_handler); 2872 2873 WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port)); 2874 2875 for (i = 1; i <= sdev->device->phys_port_cnt; i++) { 2876 sport = &sdev->port[i - 1]; 2877 INIT_LIST_HEAD(&sport->nexus_list); 2878 init_waitqueue_head(&sport->ch_releaseQ); 2879 mutex_init(&sport->mutex); 2880 sport->sdev = sdev; 2881 sport->port = i; 2882 sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE; 2883 sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE; 2884 sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE; 2885 sport->port_attrib.use_srq = false; 2886 INIT_WORK(&sport->work, srpt_refresh_port_work); 2887 2888 if (srpt_refresh_port(sport)) { 2889 pr_err("MAD registration failed for %s-%d.\n", 2890 sdev->device->name, i); 2891 goto err_event; 2892 } 2893 } 2894 2895 spin_lock(&srpt_dev_lock); 2896 list_add_tail(&sdev->list, &srpt_dev_list); 2897 spin_unlock(&srpt_dev_lock); 2898 2899 out: 2900 ib_set_client_data(device, &srpt_client, sdev); 2901 pr_debug("added %s.\n", device->name); 2902 return; 2903 2904 err_event: 2905 ib_unregister_event_handler(&sdev->event_handler); 2906 err_cm: 2907 ib_destroy_cm_id(sdev->cm_id); 2908 err_ring: 2909 srpt_free_srq(sdev); 2910 ib_dealloc_pd(sdev->pd); 2911 free_dev: 2912 kfree(sdev); 2913 err: 2914 sdev = NULL; 2915 pr_info("%s(%s) failed.\n", __func__, device->name); 2916 goto out; 2917 } 2918 2919 /** 2920 * srpt_remove_one - InfiniBand device removal callback function 2921 * @device: Describes a HCA. 2922 * @client_data: The value passed as the third argument to ib_set_client_data(). 2923 */ 2924 static void srpt_remove_one(struct ib_device *device, void *client_data) 2925 { 2926 struct srpt_device *sdev = client_data; 2927 int i; 2928 2929 if (!sdev) { 2930 pr_info("%s(%s): nothing to do.\n", __func__, device->name); 2931 return; 2932 } 2933 2934 srpt_unregister_mad_agent(sdev); 2935 2936 ib_unregister_event_handler(&sdev->event_handler); 2937 2938 /* Cancel any work queued by the just unregistered IB event handler. */ 2939 for (i = 0; i < sdev->device->phys_port_cnt; i++) 2940 cancel_work_sync(&sdev->port[i].work); 2941 2942 ib_destroy_cm_id(sdev->cm_id); 2943 2944 /* 2945 * Unregistering a target must happen after destroying sdev->cm_id 2946 * such that no new SRP_LOGIN_REQ information units can arrive while 2947 * destroying the target. 2948 */ 2949 spin_lock(&srpt_dev_lock); 2950 list_del(&sdev->list); 2951 spin_unlock(&srpt_dev_lock); 2952 2953 for (i = 0; i < sdev->device->phys_port_cnt; i++) 2954 srpt_release_sport(&sdev->port[i]); 2955 2956 srpt_free_srq(sdev); 2957 2958 ib_dealloc_pd(sdev->pd); 2959 2960 kfree(sdev); 2961 } 2962 2963 static struct ib_client srpt_client = { 2964 .name = DRV_NAME, 2965 .add = srpt_add_one, 2966 .remove = srpt_remove_one 2967 }; 2968 2969 static int srpt_check_true(struct se_portal_group *se_tpg) 2970 { 2971 return 1; 2972 } 2973 2974 static int srpt_check_false(struct se_portal_group *se_tpg) 2975 { 2976 return 0; 2977 } 2978 2979 static char *srpt_get_fabric_name(void) 2980 { 2981 return "srpt"; 2982 } 2983 2984 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg) 2985 { 2986 return tpg->se_tpg_wwn->priv; 2987 } 2988 2989 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg) 2990 { 2991 struct srpt_port *sport = srpt_tpg_to_sport(tpg); 2992 2993 WARN_ON_ONCE(tpg != &sport->port_guid_tpg && 2994 tpg != &sport->port_gid_tpg); 2995 return tpg == &sport->port_guid_tpg ? sport->port_guid : 2996 sport->port_gid; 2997 } 2998 2999 static u16 srpt_get_tag(struct se_portal_group *tpg) 3000 { 3001 return 1; 3002 } 3003 3004 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg) 3005 { 3006 return 1; 3007 } 3008 3009 static void srpt_release_cmd(struct se_cmd *se_cmd) 3010 { 3011 struct srpt_send_ioctx *ioctx = container_of(se_cmd, 3012 struct srpt_send_ioctx, cmd); 3013 struct srpt_rdma_ch *ch = ioctx->ch; 3014 unsigned long flags; 3015 3016 WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE && 3017 !(ioctx->cmd.transport_state & CMD_T_ABORTED)); 3018 3019 if (ioctx->n_rw_ctx) { 3020 srpt_free_rw_ctxs(ch, ioctx); 3021 ioctx->n_rw_ctx = 0; 3022 } 3023 3024 spin_lock_irqsave(&ch->spinlock, flags); 3025 list_add(&ioctx->free_list, &ch->free_list); 3026 spin_unlock_irqrestore(&ch->spinlock, flags); 3027 } 3028 3029 /** 3030 * srpt_close_session - forcibly close a session 3031 * @se_sess: SCSI target session. 3032 * 3033 * Callback function invoked by the TCM core to clean up sessions associated 3034 * with a node ACL when the user invokes 3035 * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id 3036 */ 3037 static void srpt_close_session(struct se_session *se_sess) 3038 { 3039 struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr; 3040 3041 srpt_disconnect_ch_sync(ch); 3042 } 3043 3044 /** 3045 * srpt_sess_get_index - return the value of scsiAttIntrPortIndex (SCSI-MIB) 3046 * @se_sess: SCSI target session. 3047 * 3048 * A quote from RFC 4455 (SCSI-MIB) about this MIB object: 3049 * This object represents an arbitrary integer used to uniquely identify a 3050 * particular attached remote initiator port to a particular SCSI target port 3051 * within a particular SCSI target device within a particular SCSI instance. 3052 */ 3053 static u32 srpt_sess_get_index(struct se_session *se_sess) 3054 { 3055 return 0; 3056 } 3057 3058 static void srpt_set_default_node_attrs(struct se_node_acl *nacl) 3059 { 3060 } 3061 3062 /* Note: only used from inside debug printk's by the TCM core. */ 3063 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd) 3064 { 3065 struct srpt_send_ioctx *ioctx; 3066 3067 ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd); 3068 return ioctx->state; 3069 } 3070 3071 static int srpt_parse_guid(u64 *guid, const char *name) 3072 { 3073 u16 w[4]; 3074 int ret = -EINVAL; 3075 3076 if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4) 3077 goto out; 3078 *guid = get_unaligned_be64(w); 3079 ret = 0; 3080 out: 3081 return ret; 3082 } 3083 3084 /** 3085 * srpt_parse_i_port_id - parse an initiator port ID 3086 * @name: ASCII representation of a 128-bit initiator port ID. 3087 * @i_port_id: Binary 128-bit port ID. 3088 */ 3089 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name) 3090 { 3091 const char *p; 3092 unsigned len, count, leading_zero_bytes; 3093 int ret; 3094 3095 p = name; 3096 if (strncasecmp(p, "0x", 2) == 0) 3097 p += 2; 3098 ret = -EINVAL; 3099 len = strlen(p); 3100 if (len % 2) 3101 goto out; 3102 count = min(len / 2, 16U); 3103 leading_zero_bytes = 16 - count; 3104 memset(i_port_id, 0, leading_zero_bytes); 3105 ret = hex2bin(i_port_id + leading_zero_bytes, p, count); 3106 if (ret < 0) 3107 pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", ret); 3108 out: 3109 return ret; 3110 } 3111 3112 /* 3113 * configfs callback function invoked for 3114 * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id 3115 */ 3116 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name) 3117 { 3118 u64 guid; 3119 u8 i_port_id[16]; 3120 int ret; 3121 3122 ret = srpt_parse_guid(&guid, name); 3123 if (ret < 0) 3124 ret = srpt_parse_i_port_id(i_port_id, name); 3125 if (ret < 0) 3126 pr_err("invalid initiator port ID %s\n", name); 3127 return ret; 3128 } 3129 3130 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item, 3131 char *page) 3132 { 3133 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3134 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3135 3136 return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size); 3137 } 3138 3139 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item, 3140 const char *page, size_t count) 3141 { 3142 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3143 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3144 unsigned long val; 3145 int ret; 3146 3147 ret = kstrtoul(page, 0, &val); 3148 if (ret < 0) { 3149 pr_err("kstrtoul() failed with ret: %d\n", ret); 3150 return -EINVAL; 3151 } 3152 if (val > MAX_SRPT_RDMA_SIZE) { 3153 pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val, 3154 MAX_SRPT_RDMA_SIZE); 3155 return -EINVAL; 3156 } 3157 if (val < DEFAULT_MAX_RDMA_SIZE) { 3158 pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n", 3159 val, DEFAULT_MAX_RDMA_SIZE); 3160 return -EINVAL; 3161 } 3162 sport->port_attrib.srp_max_rdma_size = val; 3163 3164 return count; 3165 } 3166 3167 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item, 3168 char *page) 3169 { 3170 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3171 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3172 3173 return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size); 3174 } 3175 3176 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item, 3177 const char *page, size_t count) 3178 { 3179 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3180 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3181 unsigned long val; 3182 int ret; 3183 3184 ret = kstrtoul(page, 0, &val); 3185 if (ret < 0) { 3186 pr_err("kstrtoul() failed with ret: %d\n", ret); 3187 return -EINVAL; 3188 } 3189 if (val > MAX_SRPT_RSP_SIZE) { 3190 pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val, 3191 MAX_SRPT_RSP_SIZE); 3192 return -EINVAL; 3193 } 3194 if (val < MIN_MAX_RSP_SIZE) { 3195 pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val, 3196 MIN_MAX_RSP_SIZE); 3197 return -EINVAL; 3198 } 3199 sport->port_attrib.srp_max_rsp_size = val; 3200 3201 return count; 3202 } 3203 3204 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item, 3205 char *page) 3206 { 3207 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3208 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3209 3210 return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size); 3211 } 3212 3213 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item, 3214 const char *page, size_t count) 3215 { 3216 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3217 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3218 unsigned long val; 3219 int ret; 3220 3221 ret = kstrtoul(page, 0, &val); 3222 if (ret < 0) { 3223 pr_err("kstrtoul() failed with ret: %d\n", ret); 3224 return -EINVAL; 3225 } 3226 if (val > MAX_SRPT_SRQ_SIZE) { 3227 pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val, 3228 MAX_SRPT_SRQ_SIZE); 3229 return -EINVAL; 3230 } 3231 if (val < MIN_SRPT_SRQ_SIZE) { 3232 pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val, 3233 MIN_SRPT_SRQ_SIZE); 3234 return -EINVAL; 3235 } 3236 sport->port_attrib.srp_sq_size = val; 3237 3238 return count; 3239 } 3240 3241 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item, 3242 char *page) 3243 { 3244 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3245 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3246 3247 return sprintf(page, "%d\n", sport->port_attrib.use_srq); 3248 } 3249 3250 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item, 3251 const char *page, size_t count) 3252 { 3253 struct se_portal_group *se_tpg = attrib_to_tpg(item); 3254 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3255 struct srpt_device *sdev = sport->sdev; 3256 unsigned long val; 3257 bool enabled; 3258 int ret; 3259 3260 ret = kstrtoul(page, 0, &val); 3261 if (ret < 0) 3262 return ret; 3263 if (val != !!val) 3264 return -EINVAL; 3265 3266 ret = mutex_lock_interruptible(&sdev->sdev_mutex); 3267 if (ret < 0) 3268 return ret; 3269 ret = mutex_lock_interruptible(&sport->mutex); 3270 if (ret < 0) 3271 goto unlock_sdev; 3272 enabled = sport->enabled; 3273 /* Log out all initiator systems before changing 'use_srq'. */ 3274 srpt_set_enabled(sport, false); 3275 sport->port_attrib.use_srq = val; 3276 srpt_use_srq(sdev, sport->port_attrib.use_srq); 3277 srpt_set_enabled(sport, enabled); 3278 ret = count; 3279 mutex_unlock(&sport->mutex); 3280 unlock_sdev: 3281 mutex_unlock(&sdev->sdev_mutex); 3282 3283 return ret; 3284 } 3285 3286 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_max_rdma_size); 3287 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_max_rsp_size); 3288 CONFIGFS_ATTR(srpt_tpg_attrib_, srp_sq_size); 3289 CONFIGFS_ATTR(srpt_tpg_attrib_, use_srq); 3290 3291 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = { 3292 &srpt_tpg_attrib_attr_srp_max_rdma_size, 3293 &srpt_tpg_attrib_attr_srp_max_rsp_size, 3294 &srpt_tpg_attrib_attr_srp_sq_size, 3295 &srpt_tpg_attrib_attr_use_srq, 3296 NULL, 3297 }; 3298 3299 static ssize_t srpt_tpg_enable_show(struct config_item *item, char *page) 3300 { 3301 struct se_portal_group *se_tpg = to_tpg(item); 3302 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3303 3304 return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0); 3305 } 3306 3307 static ssize_t srpt_tpg_enable_store(struct config_item *item, 3308 const char *page, size_t count) 3309 { 3310 struct se_portal_group *se_tpg = to_tpg(item); 3311 struct srpt_port *sport = srpt_tpg_to_sport(se_tpg); 3312 unsigned long tmp; 3313 int ret; 3314 3315 ret = kstrtoul(page, 0, &tmp); 3316 if (ret < 0) { 3317 pr_err("Unable to extract srpt_tpg_store_enable\n"); 3318 return -EINVAL; 3319 } 3320 3321 if ((tmp != 0) && (tmp != 1)) { 3322 pr_err("Illegal value for srpt_tpg_store_enable: %lu\n", tmp); 3323 return -EINVAL; 3324 } 3325 3326 mutex_lock(&sport->mutex); 3327 srpt_set_enabled(sport, tmp); 3328 mutex_unlock(&sport->mutex); 3329 3330 return count; 3331 } 3332 3333 CONFIGFS_ATTR(srpt_tpg_, enable); 3334 3335 static struct configfs_attribute *srpt_tpg_attrs[] = { 3336 &srpt_tpg_attr_enable, 3337 NULL, 3338 }; 3339 3340 /** 3341 * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg 3342 * @wwn: Corresponds to $driver/$port. 3343 * @group: Not used. 3344 * @name: $tpg. 3345 */ 3346 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn, 3347 struct config_group *group, 3348 const char *name) 3349 { 3350 struct srpt_port *sport = wwn->priv; 3351 static struct se_portal_group *tpg; 3352 int res; 3353 3354 WARN_ON_ONCE(wwn != &sport->port_guid_wwn && 3355 wwn != &sport->port_gid_wwn); 3356 tpg = wwn == &sport->port_guid_wwn ? &sport->port_guid_tpg : 3357 &sport->port_gid_tpg; 3358 res = core_tpg_register(wwn, tpg, SCSI_PROTOCOL_SRP); 3359 if (res) 3360 return ERR_PTR(res); 3361 3362 return tpg; 3363 } 3364 3365 /** 3366 * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg 3367 * @tpg: Target portal group to deregister. 3368 */ 3369 static void srpt_drop_tpg(struct se_portal_group *tpg) 3370 { 3371 struct srpt_port *sport = srpt_tpg_to_sport(tpg); 3372 3373 sport->enabled = false; 3374 core_tpg_deregister(tpg); 3375 } 3376 3377 /** 3378 * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port 3379 * @tf: Not used. 3380 * @group: Not used. 3381 * @name: $port. 3382 */ 3383 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf, 3384 struct config_group *group, 3385 const char *name) 3386 { 3387 return srpt_lookup_wwn(name) ? : ERR_PTR(-EINVAL); 3388 } 3389 3390 /** 3391 * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port 3392 * @wwn: $port. 3393 */ 3394 static void srpt_drop_tport(struct se_wwn *wwn) 3395 { 3396 } 3397 3398 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf) 3399 { 3400 return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION); 3401 } 3402 3403 CONFIGFS_ATTR_RO(srpt_wwn_, version); 3404 3405 static struct configfs_attribute *srpt_wwn_attrs[] = { 3406 &srpt_wwn_attr_version, 3407 NULL, 3408 }; 3409 3410 static const struct target_core_fabric_ops srpt_template = { 3411 .module = THIS_MODULE, 3412 .name = "srpt", 3413 .get_fabric_name = srpt_get_fabric_name, 3414 .tpg_get_wwn = srpt_get_fabric_wwn, 3415 .tpg_get_tag = srpt_get_tag, 3416 .tpg_check_demo_mode = srpt_check_false, 3417 .tpg_check_demo_mode_cache = srpt_check_true, 3418 .tpg_check_demo_mode_write_protect = srpt_check_true, 3419 .tpg_check_prod_mode_write_protect = srpt_check_false, 3420 .tpg_get_inst_index = srpt_tpg_get_inst_index, 3421 .release_cmd = srpt_release_cmd, 3422 .check_stop_free = srpt_check_stop_free, 3423 .close_session = srpt_close_session, 3424 .sess_get_index = srpt_sess_get_index, 3425 .sess_get_initiator_sid = NULL, 3426 .write_pending = srpt_write_pending, 3427 .write_pending_status = srpt_write_pending_status, 3428 .set_default_node_attributes = srpt_set_default_node_attrs, 3429 .get_cmd_state = srpt_get_tcm_cmd_state, 3430 .queue_data_in = srpt_queue_data_in, 3431 .queue_status = srpt_queue_status, 3432 .queue_tm_rsp = srpt_queue_tm_rsp, 3433 .aborted_task = srpt_aborted_task, 3434 /* 3435 * Setup function pointers for generic logic in 3436 * target_core_fabric_configfs.c 3437 */ 3438 .fabric_make_wwn = srpt_make_tport, 3439 .fabric_drop_wwn = srpt_drop_tport, 3440 .fabric_make_tpg = srpt_make_tpg, 3441 .fabric_drop_tpg = srpt_drop_tpg, 3442 .fabric_init_nodeacl = srpt_init_nodeacl, 3443 3444 .tfc_wwn_attrs = srpt_wwn_attrs, 3445 .tfc_tpg_base_attrs = srpt_tpg_attrs, 3446 .tfc_tpg_attrib_attrs = srpt_tpg_attrib_attrs, 3447 }; 3448 3449 /** 3450 * srpt_init_module - kernel module initialization 3451 * 3452 * Note: Since ib_register_client() registers callback functions, and since at 3453 * least one of these callback functions (srpt_add_one()) calls target core 3454 * functions, this driver must be registered with the target core before 3455 * ib_register_client() is called. 3456 */ 3457 static int __init srpt_init_module(void) 3458 { 3459 int ret; 3460 3461 ret = -EINVAL; 3462 if (srp_max_req_size < MIN_MAX_REQ_SIZE) { 3463 pr_err("invalid value %d for kernel module parameter" 3464 " srp_max_req_size -- must be at least %d.\n", 3465 srp_max_req_size, MIN_MAX_REQ_SIZE); 3466 goto out; 3467 } 3468 3469 if (srpt_srq_size < MIN_SRPT_SRQ_SIZE 3470 || srpt_srq_size > MAX_SRPT_SRQ_SIZE) { 3471 pr_err("invalid value %d for kernel module parameter" 3472 " srpt_srq_size -- must be in the range [%d..%d].\n", 3473 srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE); 3474 goto out; 3475 } 3476 3477 ret = target_register_template(&srpt_template); 3478 if (ret) 3479 goto out; 3480 3481 ret = ib_register_client(&srpt_client); 3482 if (ret) { 3483 pr_err("couldn't register IB client\n"); 3484 goto out_unregister_target; 3485 } 3486 3487 return 0; 3488 3489 out_unregister_target: 3490 target_unregister_template(&srpt_template); 3491 out: 3492 return ret; 3493 } 3494 3495 static void __exit srpt_cleanup_module(void) 3496 { 3497 ib_unregister_client(&srpt_client); 3498 target_unregister_template(&srpt_template); 3499 } 3500 3501 module_init(srpt_init_module); 3502 module_exit(srpt_cleanup_module); 3503