1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright 2010 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 /* 28 * RDMA channel interface for Solaris SCSI RDMA Protocol Target (SRP) 29 * transport port provider module for the COMSTAR framework. 30 */ 31 32 #include <sys/cpuvar.h> 33 #include <sys/types.h> 34 #include <sys/conf.h> 35 #include <sys/stat.h> 36 #include <sys/file.h> 37 #include <sys/ddi.h> 38 #include <sys/sunddi.h> 39 #include <sys/modctl.h> 40 #include <sys/sysmacros.h> 41 #include <sys/sdt.h> 42 #include <sys/taskq.h> 43 #include <sys/scsi/scsi.h> 44 #include <sys/ib/ibtl/ibti.h> 45 46 #include <stmf.h> 47 #include <stmf_ioctl.h> 48 #include <portif.h> 49 50 #include "srp.h" 51 #include "srpt_impl.h" 52 #include "srpt_ioc.h" 53 #include "srpt_stp.h" 54 #include "srpt_ch.h" 55 56 extern srpt_ctxt_t *srpt_ctxt; 57 extern uint16_t srpt_send_msg_depth; 58 59 /* 60 * Prototypes. 61 */ 62 static void srpt_ch_scq_hdlr(ibt_cq_hdl_t cq_dhl, void *arg); 63 static void srpt_ch_rcq_hdlr(ibt_cq_hdl_t cq_dhl, void *arg); 64 static void srpt_ch_process_iu(srpt_channel_t *ch, srpt_iu_t *iu); 65 66 /* 67 * srpt_ch_alloc() 68 */ 69 srpt_channel_t * 70 srpt_ch_alloc(srpt_target_port_t *tgt, uint8_t port) 71 { 72 ibt_status_t status; 73 srpt_channel_t *ch; 74 ibt_cq_attr_t cq_attr; 75 ibt_rc_chan_alloc_args_t ch_args; 76 uint32_t cq_real_size; 77 srpt_ioc_t *ioc; 78 79 ASSERT(tgt != NULL); 80 ioc = tgt->tp_ioc; 81 ASSERT(ioc != NULL); 82 83 ch = kmem_zalloc(sizeof (*ch), KM_SLEEP); 84 rw_init(&ch->ch_rwlock, NULL, RW_DRIVER, NULL); 85 mutex_init(&ch->ch_reflock, NULL, MUTEX_DRIVER, NULL); 86 cv_init(&ch->ch_cv_complete, NULL, CV_DRIVER, NULL); 87 ch->ch_refcnt = 1; 88 ch->ch_cv_waiters = 0; 89 90 ch->ch_state = SRPT_CHANNEL_CONNECTING; 91 ch->ch_tgt = tgt; 92 ch->ch_req_lim_delta = 0; 93 ch->ch_ti_iu_len = 0; 94 95 cq_attr.cq_size = srpt_send_msg_depth * 2; 96 cq_attr.cq_sched = 0; 97 cq_attr.cq_flags = IBT_CQ_NO_FLAGS; 98 99 status = ibt_alloc_cq(ioc->ioc_ibt_hdl, &cq_attr, &ch->ch_scq_hdl, 100 &cq_real_size); 101 if (status != IBT_SUCCESS) { 102 SRPT_DPRINTF_L1("ch_alloc, send CQ alloc error (%d)", 103 status); 104 goto scq_alloc_err; 105 } 106 107 cq_attr.cq_size = srpt_send_msg_depth + 1; 108 cq_attr.cq_sched = 0; 109 cq_attr.cq_flags = IBT_CQ_NO_FLAGS; 110 111 status = ibt_alloc_cq(ioc->ioc_ibt_hdl, &cq_attr, &ch->ch_rcq_hdl, 112 &cq_real_size); 113 if (status != IBT_SUCCESS) { 114 SRPT_DPRINTF_L2("ch_alloc, receive CQ alloc error (%d)", 115 status); 116 goto rcq_alloc_err; 117 } 118 119 ibt_set_cq_handler(ch->ch_scq_hdl, srpt_ch_scq_hdlr, ch); 120 ibt_set_cq_handler(ch->ch_rcq_hdl, srpt_ch_rcq_hdlr, ch); 121 (void) ibt_enable_cq_notify(ch->ch_scq_hdl, IBT_NEXT_COMPLETION); 122 (void) ibt_enable_cq_notify(ch->ch_rcq_hdl, IBT_NEXT_COMPLETION); 123 124 ch_args.rc_flags = IBT_WR_SIGNALED; 125 126 /* Maker certain initiator can not read/write our memory */ 127 ch_args.rc_control = 0; 128 129 ch_args.rc_hca_port_num = port; 130 131 /* 132 * Any SRP IU can result in a number of STMF data buffer transfers 133 * and those transfers themselves could span multiple initiator 134 * buffers. Therefore, the number of send WQE's actually required 135 * can vary. Here we assume that on average an I/O will require 136 * no more than SRPT_MAX_OUT_IO_PER_CMD send WQE's. In practice 137 * this will prevent send work queue overrun, but we will also 138 * inform STMF to throttle I/O should the work queue become full. 139 * 140 * If the HCA tells us the max outstanding WRs for a channel is 141 * lower than our default, use the HCA value. 142 */ 143 ch_args.rc_sizes.cs_sq = min(ioc->ioc_attr.hca_max_chan_sz, 144 (srpt_send_msg_depth * SRPT_MAX_OUT_IO_PER_CMD)); 145 ch_args.rc_sizes.cs_rq = 0; 146 ch_args.rc_sizes.cs_sq_sgl = 2; 147 ch_args.rc_sizes.cs_rq_sgl = 0; 148 149 ch_args.rc_scq = ch->ch_scq_hdl; 150 ch_args.rc_rcq = ch->ch_rcq_hdl; 151 ch_args.rc_pd = ioc->ioc_pd_hdl; 152 ch_args.rc_clone_chan = NULL; 153 ch_args.rc_srq = ioc->ioc_srq_hdl; 154 155 status = ibt_alloc_rc_channel(ioc->ioc_ibt_hdl, IBT_ACHAN_USES_SRQ, 156 &ch_args, &ch->ch_chan_hdl, &ch->ch_sizes); 157 if (status != IBT_SUCCESS) { 158 SRPT_DPRINTF_L2("ch_alloc, IBT channel alloc error (%d)", 159 status); 160 goto qp_alloc_err; 161 } 162 163 /* 164 * Create pool of send WQE entries to map send wqe work IDs 165 * to various types (specifically in error cases where OP 166 * is not known). 167 */ 168 ch->ch_num_swqe = ch->ch_sizes.cs_sq; 169 SRPT_DPRINTF_L2("ch_alloc, number of SWQEs = %u", ch->ch_num_swqe); 170 ch->ch_swqe = kmem_zalloc(sizeof (srpt_swqe_t) * ch->ch_num_swqe, 171 KM_SLEEP); 172 if (ch->ch_swqe == NULL) { 173 SRPT_DPRINTF_L2("ch_alloc, SWQE alloc error"); 174 (void) ibt_free_channel(ch->ch_chan_hdl); 175 goto qp_alloc_err; 176 } 177 mutex_init(&ch->ch_swqe_lock, NULL, MUTEX_DRIVER, NULL); 178 ch->ch_head = 1; 179 for (ch->ch_tail = 1; ch->ch_tail < ch->ch_num_swqe -1; ch->ch_tail++) { 180 ch->ch_swqe[ch->ch_tail].sw_next = ch->ch_tail + 1; 181 } 182 ch->ch_swqe[ch->ch_tail].sw_next = 0; 183 184 ibt_set_chan_private(ch->ch_chan_hdl, ch); 185 return (ch); 186 187 qp_alloc_err: 188 (void) ibt_free_cq(ch->ch_rcq_hdl); 189 190 rcq_alloc_err: 191 (void) ibt_free_cq(ch->ch_scq_hdl); 192 193 scq_alloc_err: 194 cv_destroy(&ch->ch_cv_complete); 195 mutex_destroy(&ch->ch_reflock); 196 rw_destroy(&ch->ch_rwlock); 197 kmem_free(ch, sizeof (*ch)); 198 199 return (NULL); 200 } 201 202 /* 203 * srpt_ch_add_ref() 204 */ 205 void 206 srpt_ch_add_ref(srpt_channel_t *ch) 207 { 208 mutex_enter(&ch->ch_reflock); 209 ch->ch_refcnt++; 210 SRPT_DPRINTF_L4("ch_add_ref, ch (%p), refcnt (%d)", 211 (void *)ch, ch->ch_refcnt); 212 ASSERT(ch->ch_refcnt != 0); 213 mutex_exit(&ch->ch_reflock); 214 } 215 216 /* 217 * srpt_ch_release_ref() 218 * 219 * A non-zero value for wait causes thread to block until all references 220 * to channel are released. 221 */ 222 void 223 srpt_ch_release_ref(srpt_channel_t *ch, uint_t wait) 224 { 225 mutex_enter(&ch->ch_reflock); 226 227 SRPT_DPRINTF_L4("ch_release_ref, ch (%p), refcnt (%d), wait (%d)", 228 (void *)ch, ch->ch_refcnt, wait); 229 230 ASSERT(ch->ch_refcnt != 0); 231 232 ch->ch_refcnt--; 233 234 if (ch->ch_refcnt != 0) { 235 if (wait) { 236 ch->ch_cv_waiters++; 237 while (ch->ch_refcnt != 0) { 238 cv_wait(&ch->ch_cv_complete, &ch->ch_reflock); 239 } 240 ch->ch_cv_waiters--; 241 } else { 242 mutex_exit(&ch->ch_reflock); 243 return; 244 } 245 } 246 247 /* 248 * Last thread out frees the IB resources, locks/conditions and memory 249 */ 250 if (ch->ch_cv_waiters > 0) { 251 /* we're not last, wake someone else up */ 252 cv_signal(&ch->ch_cv_complete); 253 mutex_exit(&ch->ch_reflock); 254 return; 255 } 256 257 SRPT_DPRINTF_L3("ch_release_ref - release resources"); 258 if (ch->ch_chan_hdl) { 259 SRPT_DPRINTF_L3("ch_release_ref - free channel"); 260 (void) ibt_free_channel(ch->ch_chan_hdl); 261 } 262 263 if (ch->ch_scq_hdl) { 264 (void) ibt_free_cq(ch->ch_scq_hdl); 265 } 266 267 if (ch->ch_rcq_hdl) { 268 (void) ibt_free_cq(ch->ch_rcq_hdl); 269 } 270 271 /* 272 * There should be no IU's associated with this 273 * channel on the SCSI session. 274 */ 275 if (ch->ch_session != NULL) { 276 ASSERT(list_is_empty(&ch->ch_session->ss_task_list)); 277 278 /* 279 * Currently only have one channel per session, we will 280 * need to release a reference when support is added 281 * for multi-channel target login. 282 */ 283 srpt_stp_free_session(ch->ch_session); 284 ch->ch_session = NULL; 285 } 286 287 kmem_free(ch->ch_swqe, sizeof (srpt_swqe_t) * ch->ch_num_swqe); 288 mutex_destroy(&ch->ch_swqe_lock); 289 mutex_exit(&ch->ch_reflock); 290 mutex_destroy(&ch->ch_reflock); 291 rw_destroy(&ch->ch_rwlock); 292 kmem_free(ch, sizeof (srpt_channel_t)); 293 } 294 295 /* 296 * srpt_ch_disconnect() 297 */ 298 void 299 srpt_ch_disconnect(srpt_channel_t *ch) 300 { 301 ibt_status_t status; 302 303 SRPT_DPRINTF_L3("ch_disconnect, invoked for ch (%p)", 304 (void *)ch); 305 306 rw_enter(&ch->ch_rwlock, RW_WRITER); 307 308 /* 309 * If we are already in the process of disconnecting then 310 * nothing need be done, CM will call-back into us when done. 311 */ 312 if (ch->ch_state == SRPT_CHANNEL_DISCONNECTING) { 313 SRPT_DPRINTF_L2("ch_disconnect, called when" 314 " disconnect in progress"); 315 rw_exit(&ch->ch_rwlock); 316 return; 317 } 318 ch->ch_state = SRPT_CHANNEL_DISCONNECTING; 319 rw_exit(&ch->ch_rwlock); 320 321 /* 322 * Initiate the sending of the CM DREQ message, the private data 323 * should be the SRP Target logout IU. We don't really care about 324 * the remote CM DREP message returned. We issue this in an 325 * asynchronous manner and will cleanup when called back by CM. 326 */ 327 status = ibt_close_rc_channel(ch->ch_chan_hdl, IBT_NONBLOCKING, 328 NULL, 0, NULL, NULL, 0); 329 330 if (status != IBT_SUCCESS) { 331 SRPT_DPRINTF_L2("ch_disconnect, close RC channel" 332 " err(%d)", status); 333 } 334 } 335 336 /* 337 * srpt_ch_cleanup() 338 */ 339 void 340 srpt_ch_cleanup(srpt_channel_t *ch) 341 { 342 srpt_iu_t *iu; 343 srpt_iu_t *next; 344 ibt_wc_t wc; 345 srpt_target_port_t *tgt; 346 srpt_channel_t *tgt_ch; 347 scsi_task_t *iutask; 348 349 SRPT_DPRINTF_L3("ch_cleanup, invoked for ch(%p), state(%d)", 350 (void *)ch, ch->ch_state); 351 352 /* add a ref for the channel until we're done */ 353 srpt_ch_add_ref(ch); 354 355 tgt = ch->ch_tgt; 356 ASSERT(tgt != NULL); 357 358 /* 359 * Make certain the channel is in the target ports list of 360 * known channels and remove it (releasing the target 361 * ports reference to the channel). 362 */ 363 mutex_enter(&tgt->tp_ch_list_lock); 364 tgt_ch = list_head(&tgt->tp_ch_list); 365 while (tgt_ch != NULL) { 366 if (tgt_ch == ch) { 367 list_remove(&tgt->tp_ch_list, tgt_ch); 368 srpt_ch_release_ref(tgt_ch, 0); 369 break; 370 } 371 tgt_ch = list_next(&tgt->tp_ch_list, tgt_ch); 372 } 373 mutex_exit(&tgt->tp_ch_list_lock); 374 375 if (tgt_ch == NULL) { 376 SRPT_DPRINTF_L2("ch_cleanup, target channel no" 377 "longer known to target"); 378 srpt_ch_release_ref(ch, 0); 379 return; 380 } 381 382 rw_enter(&ch->ch_rwlock, RW_WRITER); 383 ch->ch_state = SRPT_CHANNEL_DISCONNECTING; 384 rw_exit(&ch->ch_rwlock); 385 386 387 /* 388 * Generally the IB CQ's will have been drained prior to 389 * getting to this call; but we check here to make certain. 390 */ 391 if (ch->ch_scq_hdl) { 392 SRPT_DPRINTF_L4("ch_cleanup, start drain (%d)", 393 ch->ch_swqe_posted); 394 while ((int)ch->ch_swqe_posted > 0) { 395 delay(drv_usectohz(1000)); 396 } 397 ibt_set_cq_handler(ch->ch_scq_hdl, NULL, NULL); 398 } 399 400 if (ch->ch_rcq_hdl) { 401 ibt_set_cq_handler(ch->ch_rcq_hdl, NULL, NULL); 402 403 while (ibt_poll_cq(ch->ch_rcq_hdl, &wc, 1, NULL) == 404 IBT_SUCCESS) { 405 iu = (srpt_iu_t *)(uintptr_t)wc.wc_id; 406 SRPT_DPRINTF_L4("ch_cleanup, recovering" 407 " outstanding RX iu(%p)", (void *)iu); 408 mutex_enter(&iu->iu_lock); 409 srpt_ioc_repost_recv_iu(iu->iu_ioc, iu); 410 /* 411 * Channel reference has not yet been added for this 412 * IU, so do not decrement. 413 */ 414 mutex_exit(&iu->iu_lock); 415 } 416 } 417 418 /* 419 * Go through the list of outstanding IU for the channel's SCSI 420 * session and for each either abort or complete an abort. 421 */ 422 rw_enter(&ch->ch_rwlock, RW_READER); 423 if (ch->ch_session != NULL) { 424 rw_enter(&ch->ch_session->ss_rwlock, RW_READER); 425 iu = list_head(&ch->ch_session->ss_task_list); 426 while (iu != NULL) { 427 next = list_next(&ch->ch_session->ss_task_list, iu); 428 429 mutex_enter(&iu->iu_lock); 430 if (ch == iu->iu_ch) { 431 if (iu->iu_stmf_task == NULL) { 432 cmn_err(CE_NOTE, 433 "ch_cleanup, NULL stmf task"); 434 ASSERT(0); 435 } 436 iutask = iu->iu_stmf_task; 437 } else { 438 iutask = NULL; 439 } 440 mutex_exit(&iu->iu_lock); 441 442 if (iutask != NULL) { 443 SRPT_DPRINTF_L4("ch_cleanup, aborting " 444 "task(%p)", (void *)iutask); 445 stmf_abort(STMF_QUEUE_TASK_ABORT, iutask, 446 STMF_ABORTED, NULL); 447 } 448 iu = next; 449 } 450 rw_exit(&ch->ch_session->ss_rwlock); 451 } 452 rw_exit(&ch->ch_rwlock); 453 454 srpt_ch_release_ref(ch, 0); 455 } 456 457 /* 458 * srpt_ch_rsp_comp() 459 * 460 * Process a completion for an IB SEND message. A SEND completion 461 * is for a SRP response packet sent back to the initiator. It 462 * will not have a STMF SCSI task associated with it if it was 463 * sent for a rejected IU, or was a task management abort response. 464 */ 465 static void 466 srpt_ch_rsp_comp(srpt_channel_t *ch, srpt_iu_t *iu, 467 ibt_wc_status_t wc_status) 468 { 469 stmf_status_t st = STMF_SUCCESS; 470 471 ASSERT(iu->iu_ch == ch); 472 473 /* 474 * Process the completion regardless whether it's a failure or 475 * success. At this point, we've processed as far as we can and 476 * just need to complete the associated task. 477 */ 478 479 if (wc_status != IBT_SUCCESS) { 480 SRPT_DPRINTF_L2("ch_rsp_comp, WC status err(%d)", 481 wc_status); 482 483 st = STMF_FAILURE; 484 485 if (wc_status != IBT_WC_WR_FLUSHED_ERR) { 486 srpt_ch_disconnect(ch); 487 } 488 } 489 490 /* 491 * If the IU response completion is not associated with 492 * with a SCSI task, release the IU to return the resource 493 * and the reference to the channel it holds. 494 */ 495 mutex_enter(&iu->iu_lock); 496 atomic_dec_32(&iu->iu_sq_posted_cnt); 497 498 if (iu->iu_stmf_task == NULL) { 499 srpt_ioc_repost_recv_iu(iu->iu_ioc, iu); 500 mutex_exit(&iu->iu_lock); 501 srpt_ch_release_ref(ch, 0); 502 return; 503 } 504 505 /* 506 * We should not get a SEND completion where the task has already 507 * completed aborting and STMF has been informed. 508 */ 509 ASSERT((iu->iu_flags & SRPT_IU_ABORTED) == 0); 510 511 /* 512 * Let STMF know we are done. 513 */ 514 mutex_exit(&iu->iu_lock); 515 516 stmf_send_status_done(iu->iu_stmf_task, st, STMF_IOF_LPORT_DONE); 517 } 518 519 /* 520 * srpt_ch_data_comp() 521 * 522 * Process an IB completion for a RDMA operation. This completion 523 * should be associated with the last RDMA operation for any 524 * data buffer transfer. 525 */ 526 static void 527 srpt_ch_data_comp(srpt_channel_t *ch, stmf_data_buf_t *stmf_dbuf, 528 ibt_wc_status_t wc_status) 529 { 530 srpt_ds_dbuf_t *dbuf; 531 srpt_iu_t *iu; 532 stmf_status_t status; 533 534 ASSERT(stmf_dbuf != NULL); 535 536 dbuf = (srpt_ds_dbuf_t *)stmf_dbuf->db_port_private; 537 538 ASSERT(dbuf != NULL); 539 540 iu = dbuf->db_iu; 541 542 ASSERT(iu != NULL); 543 ASSERT(iu->iu_ch == ch); 544 545 /* 546 * If work completion indicates non-flush failure, then 547 * start a channel disconnect (asynchronous) and release 548 * the reference to the IU. The task will be cleaned 549 * up with STMF during channel shutdown processing. 550 */ 551 if (wc_status != IBT_SUCCESS) { 552 SRPT_DPRINTF_L2("ch_data_comp, WC status err(%d)", 553 wc_status); 554 if (wc_status != IBT_WC_WR_FLUSHED_ERR) { 555 srpt_ch_disconnect(ch); 556 } 557 atomic_dec_32(&iu->iu_sq_posted_cnt); 558 return; 559 } 560 561 /* 562 * If STMF has requested this task be aborted, then if this is the 563 * last I/O operation outstanding, notify STMF the task has been 564 * aborted and ignore the completion. 565 */ 566 mutex_enter(&iu->iu_lock); 567 atomic_dec_32(&iu->iu_sq_posted_cnt); 568 569 if ((iu->iu_flags & SRPT_IU_STMF_ABORTING) != 0) { 570 scsi_task_t *abort_task = iu->iu_stmf_task; 571 572 mutex_exit(&iu->iu_lock); 573 stmf_abort(STMF_REQUEUE_TASK_ABORT_LPORT, abort_task, 574 STMF_ABORTED, NULL); 575 return; 576 } 577 578 /* 579 * We should not get an RDMA completion where the task has already 580 * completed aborting and STMF has been informed. 581 */ 582 ASSERT((iu->iu_flags & SRPT_IU_ABORTED) == 0); 583 584 /* 585 * Good completion for last RDMA op associated with a data buffer 586 * I/O, if specified initiate status otherwise let STMF know we are 587 * done. 588 */ 589 stmf_dbuf->db_xfer_status = STMF_SUCCESS; 590 mutex_exit(&iu->iu_lock); 591 592 DTRACE_SRP_8(xfer__done, srpt_channel_t, ch, 593 ibt_wr_ds_t, &(dbuf->db_sge), srpt_iu_t, iu, 594 ibt_send_wr_t, 0, uint32_t, stmf_dbuf->db_data_size, 595 uint32_t, 0, uint32_t, 0, 596 uint32_t, (stmf_dbuf->db_flags & DB_DIRECTION_TO_RPORT) ? 1 : 0); 597 598 if ((stmf_dbuf->db_flags & DB_SEND_STATUS_GOOD) != 0) { 599 status = srpt_stp_send_status(dbuf->db_iu->iu_stmf_task, 0); 600 if (status == STMF_SUCCESS) { 601 return; 602 } 603 stmf_dbuf->db_xfer_status = STMF_FAILURE; 604 } 605 stmf_data_xfer_done(dbuf->db_iu->iu_stmf_task, stmf_dbuf, 0); 606 } 607 608 /* 609 * srpt_ch_scq_hdlr() 610 */ 611 static void 612 srpt_ch_scq_hdlr(ibt_cq_hdl_t cq_hdl, void *arg) 613 { 614 ibt_status_t status; 615 srpt_channel_t *ch = arg; 616 ibt_wc_t wc[SRPT_SEND_WC_POLL_SIZE]; 617 ibt_wc_t *wcp; 618 int i; 619 uint32_t cq_rearmed = 0; 620 uint32_t entries; 621 srpt_swqe_t *swqe; 622 623 ASSERT(ch != NULL); 624 625 /* Reference channel for the duration of this call */ 626 srpt_ch_add_ref(ch); 627 628 for (;;) { 629 status = ibt_poll_cq(cq_hdl, &wc[0], SRPT_SEND_WC_POLL_SIZE, 630 &entries); 631 if (status == IBT_CQ_EMPTY) { 632 /* 633 * CQ drained, if we have not rearmed the CQ 634 * do so and poll to eliminate race; otherwise 635 * we are done. 636 */ 637 if (cq_rearmed == 0) { 638 (void) ibt_enable_cq_notify(ch->ch_scq_hdl, 639 IBT_NEXT_COMPLETION); 640 cq_rearmed = 1; 641 continue; 642 } else { 643 break; 644 } 645 } else if (status != IBT_SUCCESS) { 646 /* 647 * This error should not happen, it indicates something 648 * abnormal has gone wrong and represents either a 649 * hardware or programming logic coding error. 650 */ 651 SRPT_DPRINTF_L2("ch_scq_hdlr, unexpected CQ err(%d)", 652 status); 653 srpt_ch_disconnect(ch); 654 break; 655 } 656 657 for (wcp = wc, i = 0; i < entries; i++, wcp++) { 658 659 /* 660 * A zero work ID indicates this CQE is associated 661 * with an intermediate post of a RDMA data transfer 662 * operation. Since intermediate data requests are 663 * unsignaled, we should only get these if there was 664 * an error. No action is required. 665 */ 666 if (wcp->wc_id == 0) { 667 continue; 668 } 669 swqe = ch->ch_swqe + wcp->wc_id; 670 671 switch (swqe->sw_type) { 672 case SRPT_SWQE_TYPE_RESP: 673 srpt_ch_rsp_comp(ch, (srpt_iu_t *) 674 swqe->sw_addr, wcp->wc_status); 675 break; 676 677 case SRPT_SWQE_TYPE_DATA: 678 srpt_ch_data_comp(ch, (stmf_data_buf_t *) 679 swqe->sw_addr, wcp->wc_status); 680 break; 681 682 default: 683 SRPT_DPRINTF_L2("ch_scq_hdlr, bad type(%d)", 684 swqe->sw_type); 685 ASSERT(0); 686 } 687 688 srpt_ch_free_swqe_wrid(ch, wcp->wc_id); 689 } 690 } 691 692 srpt_ch_release_ref(ch, 0); 693 } 694 695 /* 696 * srpt_ch_rcq_hdlr() 697 */ 698 static void 699 srpt_ch_rcq_hdlr(ibt_cq_hdl_t cq_hdl, void *arg) 700 { 701 ibt_status_t status; 702 srpt_channel_t *ch = arg; 703 ibt_wc_t wc[SRPT_RECV_WC_POLL_SIZE]; 704 ibt_wc_t *wcp; 705 int i; 706 uint32_t entries; 707 srpt_iu_t *iu; 708 uint_t cq_rearmed = 0; 709 710 /* 711 * The channel object will exists while the CQ handler call-back 712 * is installed. 713 */ 714 ASSERT(ch != NULL); 715 srpt_ch_add_ref(ch); 716 717 /* 718 * If we know a channel disconnect has started do nothing 719 * and let channel cleanup code recover resources from the CQ. 720 * We are not concerned about races with the state transition 721 * since the code will do the correct thing either way. This 722 * is simply to circumvent rearming the CQ, and it will 723 * catch the state next time. 724 */ 725 rw_enter(&ch->ch_rwlock, RW_READER); 726 if (ch->ch_state == SRPT_CHANNEL_DISCONNECTING) { 727 SRPT_DPRINTF_L2("ch_rcq_hdlr, channel disconnecting"); 728 rw_exit(&ch->ch_rwlock); 729 srpt_ch_release_ref(ch, 0); 730 return; 731 } 732 rw_exit(&ch->ch_rwlock); 733 734 for (;;) { 735 status = ibt_poll_cq(cq_hdl, &wc[0], SRPT_RECV_WC_POLL_SIZE, 736 &entries); 737 if (status == IBT_CQ_EMPTY) { 738 /* 739 * OK, empty, if we have not rearmed the CQ 740 * do so, and poll to eliminate race; otherwise 741 * we are done. 742 */ 743 if (cq_rearmed == 0) { 744 (void) ibt_enable_cq_notify(ch->ch_rcq_hdl, 745 IBT_NEXT_COMPLETION); 746 cq_rearmed = 1; 747 continue; 748 } else { 749 break; 750 } 751 } else if (status != IBT_SUCCESS) { 752 /* 753 * This error should not happen, it indicates something 754 * abnormal has gone wrong and represents either a 755 * hardware or programming logic coding error. 756 */ 757 SRPT_DPRINTF_L2("ch_rcq_hdlr, unexpected CQ err(%d)", 758 status); 759 srpt_ch_disconnect(ch); 760 break; 761 } 762 763 for (wcp = wc, i = 0; i < entries; i++, wcp++) { 764 765 /* 766 * Check wc_status before proceeding. If the 767 * status indicates a channel problem, stop processing. 768 */ 769 if (wcp->wc_status != IBT_WC_SUCCESS) { 770 if (wcp->wc_status == IBT_WC_WR_FLUSHED_ERR) { 771 SRPT_DPRINTF_L2( 772 "ch_rcq, unexpected" 773 " wc_status err(%d)", 774 wcp->wc_status); 775 srpt_ch_disconnect(ch); 776 /* XXX - verify not leaking IUs */ 777 goto done; 778 } else { 779 /* skip IUs with errors */ 780 SRPT_DPRINTF_L2( 781 "ch_rcq, ERROR comp(%d)", 782 wcp->wc_status); 783 /* XXX - verify not leaking IUs */ 784 continue; 785 } 786 } 787 788 iu = (srpt_iu_t *)(uintptr_t)wcp->wc_id; 789 ASSERT(iu != NULL); 790 791 /* 792 * Process the IU. 793 */ 794 ASSERT(wcp->wc_type == IBT_WRC_RECV); 795 srpt_ch_process_iu(ch, iu); 796 } 797 } 798 799 done: 800 srpt_ch_release_ref(ch, 0); 801 } 802 803 /* 804 * srpt_ch_srp_cmd() 805 */ 806 static int 807 srpt_ch_srp_cmd(srpt_channel_t *ch, srpt_iu_t *iu) 808 { 809 srp_cmd_req_t *cmd = (srp_cmd_req_t *)iu->iu_buf; 810 srp_indirect_desc_t *i_desc; 811 uint_t i_di_cnt; 812 uint_t i_do_cnt; 813 uint8_t do_fmt; 814 uint8_t di_fmt; 815 uint32_t *cur_desc_off; 816 int i; 817 ibt_status_t status; 818 uint8_t addlen; 819 820 821 DTRACE_SRP_2(task__command, srpt_channel_t, ch, srp_cmd_req_t, cmd); 822 iu->iu_ch = ch; 823 iu->iu_tag = cmd->cr_tag; 824 825 /* 826 * The SRP specification and SAM require support for bi-directional 827 * data transfer, so we create a single buffer descriptor list that 828 * in the IU buffer that covers the data-in and data-out buffers. 829 * In practice we will just see unidirectional transfers with either 830 * data-in or data out descriptors. If we were to take that as fact, 831 * we could reduce overhead slightly. 832 */ 833 834 /* 835 * additional length is a 6-bit number in 4-byte words, so multiply by 4 836 * to get bytes. 837 */ 838 addlen = cmd->cr_add_cdb_len & 0x3f; /* mask off 6 bits */ 839 840 cur_desc_off = (uint32_t *)(void *)&cmd->cr_add_data; 841 cur_desc_off += addlen; /* 32-bit arithmetic */ 842 iu->iu_num_rdescs = 0; 843 iu->iu_rdescs = (srp_direct_desc_t *)(void *)cur_desc_off; 844 845 /* 846 * Examine buffer description for Data In (i.e. data flows 847 * to the initiator). 848 */ 849 i_do_cnt = i_di_cnt = 0; 850 di_fmt = cmd->cr_buf_fmt >> 4; 851 if (di_fmt == SRP_DATA_DESC_DIRECT) { 852 iu->iu_num_rdescs = 1; 853 cur_desc_off = (uint32_t *)(void *)&iu->iu_rdescs[1]; 854 } else if (di_fmt == SRP_DATA_DESC_INDIRECT) { 855 i_desc = (srp_indirect_desc_t *)iu->iu_rdescs; 856 i_di_cnt = b2h32(i_desc->id_table.dd_len) / 857 sizeof (srp_direct_desc_t); 858 859 /* 860 * Some initiators like OFED occasionally use the wrong counts, 861 * so check total to allow for this. NOTE: we do not support 862 * reading of the descriptor table from the initiator, so if 863 * not all descriptors are in the IU we drop the task. 864 */ 865 if (i_di_cnt > (cmd->cr_dicnt + cmd->cr_docnt)) { 866 SRPT_DPRINTF_L2("ch_srp_cmd, remote RDMA of" 867 " descriptors not supported"); 868 SRPT_DPRINTF_L2("ch_srp_cmd, sizeof entry (%d)," 869 " i_di_cnt(%d), cr_dicnt(%d)", 870 (uint_t)sizeof (srp_direct_desc_t), 871 i_di_cnt, cmd->cr_dicnt); 872 iu->iu_rdescs = NULL; 873 return (1); 874 } 875 bcopy(&i_desc->id_desc[0], iu->iu_rdescs, 876 sizeof (srp_direct_desc_t) * i_di_cnt); 877 iu->iu_num_rdescs += i_di_cnt; 878 cur_desc_off = (uint32_t *)(void *)&i_desc->id_desc[i_di_cnt]; 879 } 880 881 /* 882 * Examine buffer description for Data Out (i.e. data flows 883 * from the initiator). 884 */ 885 do_fmt = cmd->cr_buf_fmt & 0x0F; 886 if (do_fmt == SRP_DATA_DESC_DIRECT) { 887 if (di_fmt == SRP_DATA_DESC_DIRECT) { 888 bcopy(cur_desc_off, &iu->iu_rdescs[iu->iu_num_rdescs], 889 sizeof (srp_direct_desc_t)); 890 } 891 iu->iu_num_rdescs++; 892 } else if (do_fmt == SRP_DATA_DESC_INDIRECT) { 893 i_desc = (srp_indirect_desc_t *)cur_desc_off; 894 i_do_cnt = b2h32(i_desc->id_table.dd_len) / 895 sizeof (srp_direct_desc_t); 896 897 /* 898 * Some initiators like OFED occasionally use the wrong counts, 899 * so check total to allow for this. NOTE: we do not support 900 * reading of the descriptor table from the initiator, so if 901 * not all descriptors are in the IU we drop the task. 902 */ 903 if ((i_di_cnt + i_do_cnt) > (cmd->cr_dicnt + cmd->cr_docnt)) { 904 SRPT_DPRINTF_L2("ch_srp_cmd, remote RDMA of" 905 " descriptors not supported"); 906 SRPT_DPRINTF_L2("ch_srp_cmd, sizeof entry (%d)," 907 " i_do_cnt(%d), cr_docnt(%d)", 908 (uint_t)sizeof (srp_direct_desc_t), 909 i_do_cnt, cmd->cr_docnt); 910 iu->iu_rdescs = 0; 911 return (1); 912 } 913 bcopy(&i_desc->id_desc[0], &iu->iu_rdescs[iu->iu_num_rdescs], 914 sizeof (srp_direct_desc_t) * i_do_cnt); 915 iu->iu_num_rdescs += i_do_cnt; 916 } 917 918 iu->iu_tot_xfer_len = 0; 919 for (i = 0; i < iu->iu_num_rdescs; i++) { 920 iu->iu_rdescs[i].dd_vaddr = b2h64(iu->iu_rdescs[i].dd_vaddr); 921 iu->iu_rdescs[i].dd_hdl = b2h32(iu->iu_rdescs[i].dd_hdl); 922 iu->iu_rdescs[i].dd_len = b2h32(iu->iu_rdescs[i].dd_len); 923 iu->iu_tot_xfer_len += iu->iu_rdescs[i].dd_len; 924 } 925 926 #ifdef DEBUG 927 if (srpt_errlevel >= SRPT_LOG_L4) { 928 SRPT_DPRINTF_L4("ch_srp_cmd, iu->iu_tot_xfer_len (%d)", 929 iu->iu_tot_xfer_len); 930 for (i = 0; i < iu->iu_num_rdescs; i++) { 931 SRPT_DPRINTF_L4("ch_srp_cmd, rdescs[%d].dd_vaddr" 932 " (0x%08llx)", 933 i, (u_longlong_t)iu->iu_rdescs[i].dd_vaddr); 934 SRPT_DPRINTF_L4("ch_srp_cmd, rdescs[%d].dd_hdl" 935 " (0x%08x)", i, iu->iu_rdescs[i].dd_hdl); 936 SRPT_DPRINTF_L4("ch_srp_cmd, rdescs[%d].dd_len (%d)", 937 i, iu->iu_rdescs[i].dd_len); 938 } 939 SRPT_DPRINTF_L4("ch_srp_cmd, LUN (0x%08lx)", 940 (unsigned long int) *((uint64_t *)(void *) cmd->cr_lun)); 941 } 942 #endif 943 rw_enter(&ch->ch_rwlock, RW_READER); 944 945 if (ch->ch_state == SRPT_CHANNEL_DISCONNECTING) { 946 /* 947 * The channel has begun disconnecting, so ignore the 948 * the command returning the IU resources. 949 */ 950 rw_exit(&ch->ch_rwlock); 951 return (1); 952 } 953 954 /* 955 * Once a SCSI task is allocated and assigned to the IU, it 956 * owns those IU resources, which will be held until STMF 957 * is notified the task is done (from a lport perspective). 958 */ 959 iu->iu_stmf_task = stmf_task_alloc(ch->ch_tgt->tp_lport, 960 ch->ch_session->ss_ss, cmd->cr_lun, 961 SRP_CDB_SIZE + (addlen * 4), 0); 962 if (iu->iu_stmf_task == NULL) { 963 /* 964 * Could not allocate, return status to the initiator 965 * indicating that we are temporarily unable to process 966 * commands. If unable to send, immediately return IU 967 * resource. 968 */ 969 SRPT_DPRINTF_L2("ch_srp_cmd, SCSI task allocation failure"); 970 rw_exit(&ch->ch_rwlock); 971 mutex_enter(&iu->iu_lock); 972 status = srpt_stp_send_response(iu, STATUS_BUSY, 0, 0, 0, 973 NULL, SRPT_NO_FENCE_SEND); 974 mutex_exit(&iu->iu_lock); 975 if (status != IBT_SUCCESS) { 976 SRPT_DPRINTF_L2("ch_srp_cmd, error(%d) posting error" 977 " response", status); 978 return (1); 979 } else { 980 return (0); 981 } 982 } 983 984 iu->iu_stmf_task->task_port_private = iu; 985 iu->iu_stmf_task->task_flags = 0; 986 987 if (di_fmt != 0) { 988 iu->iu_stmf_task->task_flags |= TF_WRITE_DATA; 989 } 990 if (do_fmt != 0) { 991 iu->iu_stmf_task->task_flags |= TF_READ_DATA; 992 } 993 994 switch (cmd->cr_task_attr) { 995 case SRP_TSK_ATTR_QTYPE_SIMPLE: 996 iu->iu_stmf_task->task_flags |= TF_ATTR_SIMPLE_QUEUE; 997 break; 998 999 case SRP_TSK_ATTR_QTYPE_HEAD_OF_Q: 1000 iu->iu_stmf_task->task_flags |= TF_ATTR_HEAD_OF_QUEUE; 1001 break; 1002 1003 case SRP_TSK_ATTR_QTYPE_ORDERED: 1004 iu->iu_stmf_task->task_flags |= TF_ATTR_ORDERED_QUEUE; 1005 break; 1006 1007 case SRP_TSK_ATTR_QTYPE_ACA_Q_TAG: 1008 iu->iu_stmf_task->task_flags |= TF_ATTR_ACA; 1009 break; 1010 1011 default: 1012 SRPT_DPRINTF_L2("ch_srp_cmd, reserved task attr (%d)", 1013 cmd->cr_task_attr); 1014 iu->iu_stmf_task->task_flags |= TF_ATTR_ORDERED_QUEUE; 1015 break; 1016 } 1017 iu->iu_stmf_task->task_additional_flags = 0; 1018 iu->iu_stmf_task->task_priority = 0; 1019 iu->iu_stmf_task->task_mgmt_function = TM_NONE; 1020 iu->iu_stmf_task->task_max_nbufs = STMF_BUFS_MAX; 1021 iu->iu_stmf_task->task_expected_xfer_length = iu->iu_tot_xfer_len; 1022 iu->iu_stmf_task->task_csn_size = 0; 1023 1024 bcopy(cmd->cr_cdb, iu->iu_stmf_task->task_cdb, 1025 SRP_CDB_SIZE); 1026 if (addlen != 0) { 1027 bcopy(&cmd->cr_add_data, 1028 iu->iu_stmf_task->task_cdb + SRP_CDB_SIZE, 1029 addlen * 4); 1030 } 1031 1032 /* 1033 * Add the IU/task to the session and post to STMF. The task will 1034 * remain in the session's list until STMF is informed by SRP that 1035 * it is done with the task. 1036 */ 1037 DTRACE_SRP_3(scsi__command, srpt_channel_t, iu->iu_ch, 1038 scsi_task_t, iu->iu_stmf_task, srp_cmd_req_t, cmd); 1039 srpt_stp_add_task(ch->ch_session, iu); 1040 1041 SRPT_DPRINTF_L3("ch_srp_cmd, new task (%p) posted", 1042 (void *)iu->iu_stmf_task); 1043 stmf_post_task(iu->iu_stmf_task, NULL); 1044 rw_exit(&ch->ch_rwlock); 1045 1046 return (0); 1047 } 1048 1049 /* 1050 * srpt_ch_task_mgmt_abort() 1051 * 1052 * Returns 0 on success, indicating we've sent a management response. 1053 * Returns !0 to indicate failure; the IU should be reposted. 1054 */ 1055 static ibt_status_t 1056 srpt_ch_task_mgmt_abort(srpt_channel_t *ch, srpt_iu_t *iu, 1057 uint64_t tag_to_abort) 1058 { 1059 srpt_session_t *session = ch->ch_session; 1060 srpt_iu_t *ss_iu; 1061 ibt_status_t status; 1062 1063 /* 1064 * Locate the associated task (tag_to_abort) in the 1065 * session's active task list. 1066 */ 1067 rw_enter(&session->ss_rwlock, RW_READER); 1068 ss_iu = list_head(&session->ss_task_list); 1069 while (ss_iu != NULL) { 1070 mutex_enter(&ss_iu->iu_lock); 1071 if ((tag_to_abort == ss_iu->iu_tag)) { 1072 mutex_exit(&ss_iu->iu_lock); 1073 break; 1074 } 1075 mutex_exit(&ss_iu->iu_lock); 1076 ss_iu = list_next(&session->ss_task_list, ss_iu); 1077 } 1078 rw_exit(&session->ss_rwlock); 1079 1080 /* 1081 * Take appropriate action based on state of task 1082 * to be aborted: 1083 * 1) No longer exists - do nothing. 1084 * 2) Previously aborted or status queued - do nothing. 1085 * 3) Otherwise - initiate abort. 1086 */ 1087 if (ss_iu == NULL) { 1088 goto send_mgmt_resp; 1089 } 1090 1091 mutex_enter(&ss_iu->iu_lock); 1092 if ((ss_iu->iu_flags & (SRPT_IU_STMF_ABORTING | 1093 SRPT_IU_ABORTED | SRPT_IU_RESP_SENT)) != 0) { 1094 mutex_exit(&ss_iu->iu_lock); 1095 goto send_mgmt_resp; 1096 } 1097 1098 /* 1099 * Set aborting flag and notify STMF of abort request. No 1100 * additional I/O will be queued for this IU. 1101 */ 1102 SRPT_DPRINTF_L3("ch_task_mgmt_abort, task found"); 1103 ss_iu->iu_flags |= SRPT_IU_SRP_ABORTING; 1104 mutex_exit(&ss_iu->iu_lock); 1105 stmf_abort(STMF_QUEUE_TASK_ABORT, 1106 ss_iu->iu_stmf_task, STMF_ABORTED, NULL); 1107 1108 send_mgmt_resp: 1109 mutex_enter(&iu->iu_lock); 1110 status = srpt_stp_send_mgmt_response(iu, SRP_TM_SUCCESS, 1111 SRPT_FENCE_SEND); 1112 mutex_exit(&iu->iu_lock); 1113 1114 if (status != IBT_SUCCESS) { 1115 SRPT_DPRINTF_L2("ch_task_mgmt_abort, err(%d)" 1116 " posting abort response", status); 1117 } 1118 1119 return (status); 1120 } 1121 1122 /* 1123 * srpt_ch_srp_task_mgmt() 1124 */ 1125 static int 1126 srpt_ch_srp_task_mgmt(srpt_channel_t *ch, srpt_iu_t *iu) 1127 { 1128 srp_tsk_mgmt_t *tsk = (srp_tsk_mgmt_t *)iu->iu_buf; 1129 uint8_t tm_fn; 1130 ibt_status_t status; 1131 1132 SRPT_DPRINTF_L3("ch_srp_task_mgmt, SRP TASK MGMT func(%d)", 1133 tsk->tm_function); 1134 1135 /* 1136 * Both tag and lun fileds have the same corresponding offsets 1137 * in both srp_tsk_mgmt_t and srp_cmd_req_t structures. The 1138 * casting will allow us to use the same dtrace translator. 1139 */ 1140 DTRACE_SRP_2(task__command, srpt_channel_t, ch, 1141 srp_cmd_req_t, (srp_cmd_req_t *)tsk); 1142 1143 iu->iu_ch = ch; 1144 iu->iu_tag = tsk->tm_tag; 1145 1146 /* 1147 * Task management aborts are processed directly by the SRP driver; 1148 * all other task management requests are handed off to STMF. 1149 */ 1150 switch (tsk->tm_function) { 1151 case SRP_TSK_MGMT_ABORT_TASK: 1152 /* 1153 * Initiate SCSI transport protocol specific task abort 1154 * logic. 1155 */ 1156 status = srpt_ch_task_mgmt_abort(ch, iu, tsk->tm_task_tag); 1157 if (status != IBT_SUCCESS) { 1158 /* repost this IU */ 1159 return (1); 1160 } else { 1161 return (0); 1162 } 1163 1164 case SRP_TSK_MGMT_ABORT_TASK_SET: 1165 tm_fn = TM_ABORT_TASK_SET; 1166 break; 1167 1168 case SRP_TSK_MGMT_CLEAR_TASK_SET: 1169 tm_fn = TM_CLEAR_TASK_SET; 1170 break; 1171 1172 case SRP_TSK_MGMT_LUN_RESET: 1173 tm_fn = TM_LUN_RESET; 1174 break; 1175 1176 case SRP_TSK_MGMT_CLEAR_ACA: 1177 tm_fn = TM_CLEAR_ACA; 1178 break; 1179 1180 default: 1181 /* 1182 * SRP does not support the requested task management 1183 * function; return a not supported status in the response. 1184 */ 1185 SRPT_DPRINTF_L2("ch_srp_task_mgmt, SRP task mgmt fn(%d)" 1186 " not supported", tsk->tm_function); 1187 mutex_enter(&iu->iu_lock); 1188 status = srpt_stp_send_mgmt_response(iu, 1189 SRP_TM_NOT_SUPPORTED, SRPT_NO_FENCE_SEND); 1190 mutex_exit(&iu->iu_lock); 1191 if (status != IBT_SUCCESS) { 1192 SRPT_DPRINTF_L2("ch_srp_task_mgmt, err(%d) posting" 1193 " response", status); 1194 return (1); 1195 } 1196 return (0); 1197 } 1198 1199 rw_enter(&ch->ch_rwlock, RW_READER); 1200 if (ch->ch_state == SRPT_CHANNEL_DISCONNECTING) { 1201 /* 1202 * The channel has begun disconnecting, so ignore the 1203 * the command returning the IU resources. 1204 */ 1205 rw_exit(&ch->ch_rwlock); 1206 return (1); 1207 } 1208 1209 /* 1210 * Once a SCSI mgmt task is allocated and assigned to the IU, it 1211 * owns those IU resources, which will be held until we inform 1212 * STMF that we are done with the task (from an lports perspective). 1213 */ 1214 iu->iu_stmf_task = stmf_task_alloc(ch->ch_tgt->tp_lport, 1215 ch->ch_session->ss_ss, tsk->tm_lun, 0, STMF_TASK_EXT_NONE); 1216 if (iu->iu_stmf_task == NULL) { 1217 /* 1218 * Could not allocate, return status to the initiator 1219 * indicating that we are temporarily unable to process 1220 * commands. If unable to send, immediately return IU 1221 * resource. 1222 */ 1223 SRPT_DPRINTF_L2("ch_srp_task_mgmt, SCSI task allocation" 1224 " failure"); 1225 rw_exit(&ch->ch_rwlock); 1226 mutex_enter(&iu->iu_lock); 1227 status = srpt_stp_send_response(iu, STATUS_BUSY, 0, 0, 0, 1228 NULL, SRPT_NO_FENCE_SEND); 1229 mutex_exit(&iu->iu_lock); 1230 if (status != IBT_SUCCESS) { 1231 SRPT_DPRINTF_L2("ch_srp_task_mgmt, err(%d) posting" 1232 "busy response", status); 1233 /* repost the IU */ 1234 return (1); 1235 } 1236 return (0); 1237 } 1238 1239 iu->iu_stmf_task->task_port_private = iu; 1240 iu->iu_stmf_task->task_flags = 0; 1241 iu->iu_stmf_task->task_additional_flags = 1242 TASK_AF_NO_EXPECTED_XFER_LENGTH; 1243 iu->iu_stmf_task->task_priority = 0; 1244 iu->iu_stmf_task->task_mgmt_function = tm_fn; 1245 iu->iu_stmf_task->task_max_nbufs = STMF_BUFS_MAX; 1246 iu->iu_stmf_task->task_expected_xfer_length = 0; 1247 iu->iu_stmf_task->task_csn_size = 0; 1248 1249 /* 1250 * Add the IU/task to the session and post to STMF. The task will 1251 * remain in the session's list until STMF is informed by SRP that 1252 * it is done with the task. 1253 */ 1254 srpt_stp_add_task(ch->ch_session, iu); 1255 1256 SRPT_DPRINTF_L3("ch_srp_task_mgmt, new mgmt task(%p) posted", 1257 (void *)iu->iu_stmf_task); 1258 stmf_post_task(iu->iu_stmf_task, NULL); 1259 rw_exit(&ch->ch_rwlock); 1260 1261 return (0); 1262 } 1263 1264 /* 1265 * srpt_ch_process_iu() 1266 */ 1267 static void 1268 srpt_ch_process_iu(srpt_channel_t *ch, srpt_iu_t *iu) 1269 { 1270 srpt_iu_data_t *iud; 1271 int status = 1; 1272 1273 /* 1274 * IU adds reference to channel which will represent a 1275 * a reference by STMF. If for whatever reason the IU 1276 * is not handed off to STMF, then this reference will be 1277 * released. Otherwise, the reference will be released when 1278 * SRP informs STMF that the associated SCSI task is done. 1279 */ 1280 srpt_ch_add_ref(ch); 1281 1282 /* 1283 * Validate login RC channel state. Normally active, if 1284 * not active then we need to handle a possible race between the 1285 * receipt of a implied RTU and CM calling back to notify of the 1286 * state transition. 1287 */ 1288 rw_enter(&ch->ch_rwlock, RW_READER); 1289 if (ch->ch_state == SRPT_CHANNEL_DISCONNECTING) { 1290 rw_exit(&ch->ch_rwlock); 1291 goto repost_iu; 1292 } 1293 rw_exit(&ch->ch_rwlock); 1294 1295 iud = iu->iu_buf; 1296 1297 switch (iud->rx_iu.srp_op) { 1298 case SRP_IU_CMD: 1299 status = srpt_ch_srp_cmd(ch, iu); 1300 break; 1301 1302 case SRP_IU_TASK_MGMT: 1303 status = srpt_ch_srp_task_mgmt(ch, iu); 1304 return; 1305 1306 case SRP_IU_I_LOGOUT: 1307 SRPT_DPRINTF_L3("ch_process_iu, SRP INITIATOR LOGOUT"); 1308 /* 1309 * Initiators should logout by issuing a CM disconnect 1310 * request (DREQ) with the logout IU in the private data; 1311 * however some initiators have been known to send the 1312 * IU in-band, if this happens just initiate the logout. 1313 * Note that we do not return a response as per the 1314 * specification. 1315 */ 1316 srpt_stp_logout(ch); 1317 break; 1318 1319 case SRP_IU_AER_RSP: 1320 case SRP_IU_CRED_RSP: 1321 default: 1322 /* 1323 * We don't send asynchronous events or ask for credit 1324 * adjustments, so nothing need be done. Log we got an 1325 * unexpected IU but then just repost the IU to the SRQ. 1326 */ 1327 SRPT_DPRINTF_L2("ch_process_iu, invalid IU from initiator," 1328 " IU opcode(%d)", iud->rx_iu.srp_op); 1329 break; 1330 } 1331 1332 if (status == 0) { 1333 return; 1334 } 1335 1336 repost_iu: 1337 SRPT_DPRINTF_L4("process_iu: reposting iu %p", (void *)iu); 1338 mutex_enter(&iu->iu_lock); 1339 srpt_ioc_repost_recv_iu(iu->iu_ioc, iu); 1340 mutex_exit(&iu->iu_lock); 1341 srpt_ch_release_ref(ch, 0); 1342 } 1343 1344 /* 1345 * srpt_ch_post_send 1346 */ 1347 ibt_status_t 1348 srpt_ch_post_send(srpt_channel_t *ch, srpt_iu_t *iu, uint32_t len, 1349 uint_t fence) 1350 { 1351 ibt_status_t status; 1352 ibt_send_wr_t wr; 1353 ibt_wr_ds_t ds; 1354 uint_t posted; 1355 1356 ASSERT(ch != NULL); 1357 ASSERT(iu != NULL); 1358 ASSERT(mutex_owned(&iu->iu_lock)); 1359 1360 rw_enter(&ch->ch_rwlock, RW_READER); 1361 if (ch->ch_state == SRPT_CHANNEL_DISCONNECTING) { 1362 rw_exit(&ch->ch_rwlock); 1363 SRPT_DPRINTF_L2("ch_post_send, bad ch state (%d)", 1364 ch->ch_state); 1365 return (IBT_FAILURE); 1366 } 1367 rw_exit(&ch->ch_rwlock); 1368 1369 wr.wr_id = srpt_ch_alloc_swqe_wrid(ch, SRPT_SWQE_TYPE_RESP, 1370 (void *)iu); 1371 if (wr.wr_id == 0) { 1372 SRPT_DPRINTF_L2("ch_post_send, queue full"); 1373 return (IBT_FAILURE); 1374 } 1375 1376 atomic_inc_32(&iu->iu_sq_posted_cnt); 1377 1378 wr.wr_flags = IBT_WR_SEND_SIGNAL; 1379 if (fence == SRPT_FENCE_SEND) { 1380 wr.wr_flags |= IBT_WR_SEND_FENCE; 1381 } 1382 wr.wr_opcode = IBT_WRC_SEND; 1383 wr.wr_trans = IBT_RC_SRV; 1384 wr.wr_nds = 1; 1385 wr.wr_sgl = &ds; 1386 1387 ds.ds_va = iu->iu_sge.ds_va; 1388 ds.ds_key = iu->iu_sge.ds_key; 1389 ds.ds_len = len; 1390 1391 SRPT_DPRINTF_L4("ch_post_send, posting SRP response to channel" 1392 " ds.ds_va (0x%16llx), ds.ds_key (0x%08x), " 1393 " ds.ds_len (%d)", 1394 (u_longlong_t)ds.ds_va, ds.ds_key, ds.ds_len); 1395 1396 status = ibt_post_send(ch->ch_chan_hdl, &wr, 1, &posted); 1397 if (status != IBT_SUCCESS) { 1398 SRPT_DPRINTF_L2("ch_post_send, post_send failed (%d)", 1399 status); 1400 atomic_dec_32(&iu->iu_sq_posted_cnt); 1401 srpt_ch_free_swqe_wrid(ch, wr.wr_id); 1402 return (status); 1403 } 1404 1405 return (IBT_SUCCESS); 1406 } 1407 1408 /* 1409 * srpt_ch_alloc_swqe_wrid() 1410 */ 1411 ibt_wrid_t 1412 srpt_ch_alloc_swqe_wrid(srpt_channel_t *ch, 1413 srpt_swqe_type_t wqe_type, void *addr) 1414 { 1415 ibt_wrid_t wrid; 1416 1417 mutex_enter(&ch->ch_swqe_lock); 1418 if (ch->ch_head == ch->ch_tail) { 1419 mutex_exit(&ch->ch_swqe_lock); 1420 return ((ibt_wrid_t)0); 1421 } 1422 wrid = (ibt_wrid_t)ch->ch_head; 1423 ch->ch_swqe[ch->ch_head].sw_type = wqe_type; 1424 ch->ch_swqe[ch->ch_head].sw_addr = addr; 1425 ch->ch_head = ch->ch_swqe[ch->ch_head].sw_next; 1426 ch->ch_swqe_posted++; 1427 mutex_exit(&ch->ch_swqe_lock); 1428 return (wrid); 1429 } 1430 1431 /* 1432 * srpt_ch_free_swqe_wrid() 1433 */ 1434 void 1435 srpt_ch_free_swqe_wrid(srpt_channel_t *ch, ibt_wrid_t id) 1436 { 1437 mutex_enter(&ch->ch_swqe_lock); 1438 ch->ch_swqe[ch->ch_tail].sw_next = id; 1439 ch->ch_tail = (uint32_t)id; 1440 ch->ch_swqe_posted--; 1441 mutex_exit(&ch->ch_swqe_lock); 1442 } 1443