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 (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. 24 */ 25 26 /* 27 * hermon_event.c 28 * Hermon Interrupt and Event Processing Routines 29 * 30 * Implements all the routines necessary for allocating, freeing, and 31 * handling all of the various event types that the Hermon hardware can 32 * generate. 33 * These routines include the main Hermon interrupt service routine 34 * (hermon_isr()) as well as all the code necessary to setup and handle 35 * events from each of the many event queues used by the Hermon device. 36 */ 37 38 #include <sys/types.h> 39 #include <sys/conf.h> 40 #include <sys/ddi.h> 41 #include <sys/sunddi.h> 42 #include <sys/modctl.h> 43 44 #include <sys/ib/adapters/hermon/hermon.h> 45 46 static void hermon_eq_poll(hermon_state_t *state, hermon_eqhdl_t eq); 47 static void hermon_eq_catastrophic(hermon_state_t *state); 48 static int hermon_eq_alloc(hermon_state_t *state, uint32_t log_eq_size, 49 uint_t intr, hermon_eqhdl_t *eqhdl); 50 static int hermon_eq_free(hermon_state_t *state, hermon_eqhdl_t *eqhdl); 51 static int hermon_eq_handler_init(hermon_state_t *state, hermon_eqhdl_t eq, 52 uint_t evt_type_mask, int (*eqfunc)(hermon_state_t *state, 53 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe)); 54 static int hermon_eq_handler_fini(hermon_state_t *state, hermon_eqhdl_t eq); 55 static int hermon_port_state_change_handler(hermon_state_t *state, 56 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 57 static int hermon_comm_estbl_handler(hermon_state_t *state, 58 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 59 static int hermon_local_wq_cat_err_handler(hermon_state_t *state, 60 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 61 static int hermon_invreq_local_wq_err_handler(hermon_state_t *state, 62 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 63 static int hermon_local_acc_vio_wq_err_handler(hermon_state_t *state, 64 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 65 static int hermon_sendq_drained_handler(hermon_state_t *state, 66 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 67 static int hermon_path_mig_handler(hermon_state_t *state, 68 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 69 static int hermon_path_mig_err_handler(hermon_state_t *state, 70 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 71 static int hermon_catastrophic_handler(hermon_state_t *state, 72 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 73 static int hermon_srq_last_wqe_reached_handler(hermon_state_t *state, 74 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 75 static int hermon_fexch_error_handler(hermon_state_t *state, 76 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe); 77 static int hermon_no_eqhandler(hermon_state_t *state, hermon_eqhdl_t eq, 78 hermon_hw_eqe_t *eqe); 79 static int hermon_eq_demux(hermon_state_t *state, hermon_eqhdl_t eq, 80 hermon_hw_eqe_t *eqe); 81 82 /* 83 * hermon_eq_init_all 84 * Context: Only called from attach() path context 85 */ 86 int 87 hermon_eq_init_all(hermon_state_t *state) 88 { 89 uint_t log_eq_size, intr_num; 90 uint_t num_eq, num_eq_init, num_eq_unmap, num_eq_rsvd; 91 uint32_t event_mask; /* used for multiple event types */ 92 int status, i, num_extra; 93 struct hermon_sw_eq_s **eq; 94 ddi_acc_handle_t uarhdl = hermon_get_uarhdl(state); 95 96 /* initialize the FMA retry loop */ 97 hermon_pio_init(fm_loop_cnt, fm_status, fm_test); 98 99 /* 100 * For now, all Event Queues default to the same size (pulled from 101 * the current configuration profile) and are all assigned to the 102 * same interrupt or MSI. In the future we may support assigning 103 * EQs to specific interrupts or MSIs XXX 104 */ 105 log_eq_size = state->hs_cfg_profile->cp_log_eq_sz; 106 107 /* 108 * Total number of supported EQs is fixed. Hermon hardware 109 * supports up to 512 EQs, though in theory they will one day be 110 * alloc'd to virtual HCA's. We are currently using only 47 of them 111 * - that is, in Arbel and Tavor, before HERMON, where 112 * we had set aside the first 32 for use with Completion Queues (CQ) 113 * and reserved a few of the other 32 for each specific class of event 114 * 115 * However, with the coming of vitualization, we'll have only 4 per 116 * potential guest - so, we'll try alloc'ing them differntly 117 * (see below for more details). 118 */ 119 num_eq = HERMON_NUM_EQ_USED; 120 num_eq_rsvd = state->hs_rsvd_eqs; 121 eq = &state->hs_eqhdl[num_eq_rsvd]; 122 123 /* 124 * If MSI is to be used, then set intr_num to the MSI number. 125 * Otherwise, for fixed (i.e. 'legacy') interrupts, 126 * it is what the card tells us in 'inta_pin'. 127 */ 128 if (state->hs_intr_type_chosen == DDI_INTR_TYPE_FIXED) { 129 intr_num = state->hs_adapter.inta_pin; 130 num_extra = 0; 131 } else { 132 /* If we have more than one MSI-X vector, init them. */ 133 for (i = 0; i + 1 < state->hs_intrmsi_allocd; i++) { 134 status = hermon_eq_alloc(state, log_eq_size, i, &eq[i]); 135 if (status != DDI_SUCCESS) { 136 while (--i >= 0) { 137 (void) hermon_eq_handler_fini(state, 138 eq[i]); 139 (void) hermon_eq_free(state, &eq[i]); 140 } 141 return (DDI_FAILURE); 142 } 143 144 (void) hermon_eq_handler_init(state, eq[i], 145 HERMON_EVT_NO_MASK, hermon_cq_handler); 146 } 147 intr_num = i; 148 num_extra = i; 149 } 150 151 /* 152 * Allocate and initialize the rest of the Event Queues to be used. 153 * If any of these EQ allocations fail then jump to the end, cleanup 154 * what had been successfully initialized, and return an error. 155 */ 156 for (i = 0; i < num_eq; i++) { 157 status = hermon_eq_alloc(state, log_eq_size, intr_num, 158 &eq[num_extra + i]); 159 if (status != DDI_SUCCESS) { 160 num_eq_init = i; 161 goto all_eq_init_fail; 162 } 163 } 164 num_eq_init = num_eq; 165 /* 166 * The "num_eq_unmap" variable is used in any possible failure 167 * cleanup (below) to indicate which events queues might require 168 * possible event class unmapping. 169 */ 170 num_eq_unmap = 0; 171 172 /* 173 * Setup EQ0 (first avail) for use with Completion Queues. Note: We can 174 * cast the return value to void here because, when we use the 175 * HERMON_EVT_NO_MASK flag, it is not possible for 176 * hermon_eq_handler_init() to return an error. 177 */ 178 (void) hermon_eq_handler_init(state, eq[num_eq_unmap + num_extra], 179 HERMON_EVT_NO_MASK, hermon_cq_handler); 180 181 num_eq_unmap++; 182 183 /* 184 * Setup EQ1 for handling Completion Queue Error Events. 185 * 186 * These events include things like CQ overflow or CQ access 187 * violation errors. If this setup fails for any reason (which, in 188 * general, it really never should), then jump to the end, cleanup 189 * everything that has been successfully initialized, and return an 190 * error. 191 */ 192 status = hermon_eq_handler_init(state, eq[num_eq_unmap + num_extra], 193 HERMON_EVT_MSK_CQ_ERRORS, hermon_cq_err_handler); 194 if (status != DDI_SUCCESS) { 195 goto all_eq_init_fail; 196 } 197 state->hs_cq_erreqnum = num_eq_unmap + num_extra + num_eq_rsvd; 198 num_eq_unmap++; 199 200 /* 201 * Setup EQ2 for handling most other things including: 202 * 203 * Port State Change Events 204 * These events include things like Port Up and Port Down events. 205 * 206 * Communication Established Events 207 * These events correspond to the IB affiliated asynchronous events 208 * that are used for connection management 209 * 210 * Path Migration Succeeded Events 211 * These evens corresponid to the IB affiliated asynchronous events 212 * that are used to indicate successful completion of a 213 * Path Migration. 214 * 215 * Command Completion Events 216 * These events correspond to the Arbel generated events that are used 217 * to indicate Arbel firmware command completion. 218 * 219 * Local WQ Catastrophic Error Events 220 * Invalid Req Local WQ Error Events 221 * Local Access Violation WQ Error Events 222 * SRQ Catastrophic Error Events 223 * SRQ Last WQE Reached Events 224 * ECC error detection events 225 * These events also correspond to the similarly-named IB affiliated 226 * asynchronous error type. 227 * 228 * Send Queue Drained Events 229 * These events correspond to the IB affiliated asynchronous events 230 * that are used to indicate completion of a Send Queue Drained QP 231 * state transition. 232 * 233 * Path Migration Failed Events 234 * These events correspond to the IB affiliated asynchronous events 235 * that are used to indicate that path migration was not successful. 236 * 237 * Fibre Channel Error Event 238 * This event is affiliated with an Fexch QP. 239 * 240 * NOTE: When an event fires on this EQ, it will demux the type and 241 * send it to the right specific handler routine 242 * 243 */ 244 event_mask = 245 HERMON_EVT_MSK_PORT_STATE_CHANGE | 246 HERMON_EVT_MSK_COMM_ESTABLISHED | 247 HERMON_EVT_MSK_COMMAND_INTF_COMP | 248 HERMON_EVT_MSK_LOCAL_WQ_CAT_ERROR | 249 HERMON_EVT_MSK_INV_REQ_LOCAL_WQ_ERROR | 250 HERMON_EVT_MSK_LOCAL_ACC_VIO_WQ_ERROR | 251 HERMON_EVT_MSK_SEND_QUEUE_DRAINED | 252 HERMON_EVT_MSK_PATH_MIGRATED | 253 HERMON_EVT_MSK_PATH_MIGRATE_FAILED | 254 HERMON_EVT_MSK_SRQ_CATASTROPHIC_ERROR | 255 HERMON_EVT_MSK_SRQ_LAST_WQE_REACHED | 256 HERMON_EVT_MSK_FEXCH_ERROR; 257 258 status = hermon_eq_handler_init(state, eq[num_eq_unmap + num_extra], 259 event_mask, hermon_eq_demux); 260 if (status != DDI_SUCCESS) { 261 goto all_eq_init_fail; 262 } 263 num_eq_unmap++; 264 265 /* 266 * Setup EQ3 to catch all other types of events. Specifically, we 267 * do not catch the "Local EEC Catastrophic Error Event" because we 268 * should have no EEC (the Arbel driver does not support RD). We also 269 * choose not to handle any of the address translation page fault 270 * event types. Since we are not doing any page fault handling (and 271 * since the Arbel firmware does not currently support any such 272 * handling), we allow these events to go to the catch-all handler. 273 */ 274 status = hermon_eq_handler_init(state, eq[num_eq_unmap + num_extra], 275 HERMON_EVT_CATCHALL_MASK, hermon_no_eqhandler); 276 if (status != DDI_SUCCESS) { 277 goto all_eq_init_fail; 278 } 279 num_eq_unmap++; 280 281 /* the FMA retry loop starts. */ 282 hermon_pio_start(state, uarhdl, all_eq_init_fail, fm_loop_cnt, 283 fm_status, fm_test); 284 285 /* 286 * Run through and initialize the Consumer Index for each EQC. 287 */ 288 for (i = 0; i < num_eq + num_extra; i++) { 289 ddi_put32(uarhdl, eq[i]->eq_doorbell, 0x0); 290 } 291 292 /* the FMA retry loop ends. */ 293 hermon_pio_end(state, uarhdl, all_eq_init_fail, fm_loop_cnt, 294 fm_status, fm_test); 295 296 return (DDI_SUCCESS); 297 298 all_eq_init_fail: 299 300 /* Unmap any of the partially mapped EQs from above */ 301 for (i = 0; i < num_eq_unmap + num_extra; i++) { 302 (void) hermon_eq_handler_fini(state, eq[i]); 303 } 304 305 /* Free up any of the partially allocated EQs from above */ 306 for (i = 0; i < num_eq_init + num_extra; i++) { 307 (void) hermon_eq_free(state, &eq[i]); 308 } 309 310 /* If a HW error happen during ddi_pio, return DDI_FAILURE */ 311 if (fm_status == HCA_PIO_PERSISTENT) { 312 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_NON_FATAL); 313 status = DDI_FAILURE; 314 } 315 316 return (status); 317 } 318 319 320 /* 321 * hermon_eq_fini_all 322 * Context: Only called from attach() and/or detach() path contexts 323 */ 324 int 325 hermon_eq_fini_all(hermon_state_t *state) 326 { 327 uint_t num_eq, num_eq_rsvd; 328 int status, i; 329 struct hermon_sw_eq_s **eq; 330 331 /* 332 * Grab the total number of supported EQs again. This is the same 333 * hardcoded value that was used above (during the event queue 334 * initialization.) 335 */ 336 num_eq = HERMON_NUM_EQ_USED + state->hs_intrmsi_allocd - 1; 337 num_eq_rsvd = state->hs_rsvd_eqs; 338 eq = &state->hs_eqhdl[num_eq_rsvd]; 339 340 /* 341 * For each of the event queues that we initialized and mapped 342 * earlier, attempt to unmap the events from the EQ. 343 */ 344 for (i = 0; i < num_eq; i++) { 345 status = hermon_eq_handler_fini(state, eq[i]); 346 if (status != DDI_SUCCESS) { 347 return (DDI_FAILURE); 348 } 349 } 350 351 /* 352 * Teardown and free up all the Event Queues that were allocated 353 * earlier. 354 */ 355 for (i = 0; i < num_eq; i++) { 356 status = hermon_eq_free(state, &eq[i]); 357 if (status != DDI_SUCCESS) { 358 return (DDI_FAILURE); 359 } 360 } 361 362 return (DDI_SUCCESS); 363 } 364 365 366 /* 367 * hermon_eq_reset_uar_baseaddr 368 * Context: Only called from attach() 369 */ 370 void 371 hermon_eq_reset_uar_baseaddr(hermon_state_t *state) 372 { 373 int i, num_eq; 374 hermon_eqhdl_t eq, *eqh; 375 376 num_eq = HERMON_NUM_EQ_USED + state->hs_intrmsi_allocd - 1; 377 eqh = &state->hs_eqhdl[state->hs_rsvd_eqs]; 378 for (i = 0; i < num_eq; i++) { 379 eq = eqh[i]; 380 eq->eq_doorbell = (uint32_t *) 381 ((uintptr_t)state->hs_reg_uar_baseaddr + 382 (uint32_t)ARM_EQ_INDEX(eq->eq_eqnum)); 383 } 384 } 385 386 387 /* 388 * hermon_eq_arm_all 389 * Context: Only called from attach() and/or detach() path contexts 390 */ 391 int 392 hermon_eq_arm_all(hermon_state_t *state) 393 { 394 uint_t num_eq, num_eq_rsvd; 395 uint64_t offset; 396 hermon_eqhdl_t eq; 397 uint32_t eq_ci; 398 int i; 399 ddi_acc_handle_t uarhdl = hermon_get_uarhdl(state); 400 401 /* initialize the FMA retry loop */ 402 hermon_pio_init(fm_loop_cnt, fm_status, fm_test); 403 404 num_eq = HERMON_NUM_EQ_USED + state->hs_intrmsi_allocd - 1; 405 num_eq_rsvd = state->hs_rsvd_eqs; 406 407 /* the FMA retry loop starts. */ 408 hermon_pio_start(state, uarhdl, pio_error, fm_loop_cnt, fm_status, 409 fm_test); 410 411 for (i = 0; i < num_eq; i++) { 412 offset = ARM_EQ_INDEX(i + num_eq_rsvd); 413 eq = state->hs_eqhdl[i + num_eq_rsvd]; 414 eq_ci = (eq->eq_consindx & HERMON_EQ_CI_MASK) | EQ_ARM_BIT; 415 ddi_put32(uarhdl, 416 (uint32_t *)((uintptr_t)state->hs_reg_uar_baseaddr + 417 (uint32_t)offset), eq_ci); 418 } 419 420 /* the FMA retry loop ends. */ 421 hermon_pio_end(state, uarhdl, pio_error, fm_loop_cnt, fm_status, 422 fm_test); 423 424 return (DDI_SUCCESS); 425 426 pio_error: 427 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_NON_FATAL); 428 return (DDI_FAILURE); 429 } 430 431 432 /* 433 * hermon_isr() 434 * Context: Only called from interrupt context (and during panic) 435 */ 436 uint_t 437 hermon_isr(caddr_t arg1, caddr_t arg2) 438 { 439 hermon_state_t *state; 440 int i, r; 441 int intr; 442 443 /* 444 * Grab the Hermon softstate pointer from the input parameter 445 */ 446 state = (hermon_state_t *)(void *)arg1; 447 448 /* Get the interrupt number */ 449 intr = (int)(uintptr_t)arg2; 450 451 /* 452 * Clear the interrupt. Note: This is only needed for 453 * fixed interrupts as the framework does what is needed for 454 * MSI-X interrupts. 455 */ 456 if (state->hs_intr_type_chosen == DDI_INTR_TYPE_FIXED) { 457 ddi_acc_handle_t cmdhdl = hermon_get_cmdhdl(state); 458 459 /* initialize the FMA retry loop */ 460 hermon_pio_init(fm_loop_cnt, fm_status, fm_test); 461 462 /* the FMA retry loop starts. */ 463 hermon_pio_start(state, cmdhdl, pio_error, fm_loop_cnt, 464 fm_status, fm_test); 465 466 ddi_put64(cmdhdl, state->hs_cmd_regs.clr_intr, 467 (uint64_t)1 << state->hs_adapter.inta_pin); 468 469 /* the FMA retry loop ends. */ 470 hermon_pio_end(state, cmdhdl, pio_error, fm_loop_cnt, fm_status, 471 fm_test); 472 } 473 474 /* 475 * Loop through all the EQs looking for ones that have "fired". 476 * To determine if an EQ is fired, the ownership will be the SW 477 * (the HW will set the owner appropriately). Update the Consumer Index 478 * of the Event Queue Entry (EQE) and pass it to HW by writing it 479 * to the respective Set CI DB Register. 480 * 481 * The "else" case handles the extra EQs used only for completion 482 * events, whereas the "if" case deals with the required interrupt 483 * vector that is used for all classes of events. 484 */ 485 r = state->hs_rsvd_eqs; 486 487 if (intr + 1 == state->hs_intrmsi_allocd) { /* last intr */ 488 r += state->hs_intrmsi_allocd - 1; 489 for (i = 0; i < HERMON_NUM_EQ_USED; i++) { 490 hermon_eq_poll(state, state->hs_eqhdl[i + r]); 491 } 492 } else { /* only poll the one EQ */ 493 hermon_eq_poll(state, state->hs_eqhdl[intr + r]); 494 } 495 496 return (DDI_INTR_CLAIMED); 497 498 pio_error: 499 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_FATAL); 500 return (DDI_INTR_UNCLAIMED); 501 } 502 503 504 /* 505 * hermon_eq_poll 506 * Context: Only called from interrupt context (and during panic) 507 */ 508 static void 509 hermon_eq_poll(hermon_state_t *state, hermon_eqhdl_t eq) 510 { 511 hermon_hw_eqe_t *eqe; 512 int polled_some; 513 uint32_t cons_indx, wrap_around_mask, shift; 514 int (*eqfunction)(hermon_state_t *state, hermon_eqhdl_t eq, 515 hermon_hw_eqe_t *eqe); 516 ddi_acc_handle_t uarhdl = hermon_get_uarhdl(state); 517 518 /* initialize the FMA retry loop */ 519 hermon_pio_init(fm_loop_cnt, fm_status, fm_test); 520 521 _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*eq)) 522 523 /* Get the consumer pointer index */ 524 cons_indx = eq->eq_consindx; 525 shift = eq->eq_log_eqsz - HERMON_EQE_OWNER_SHIFT; 526 527 /* 528 * Calculate the wrap around mask. Note: This operation only works 529 * because all Hermon event queues have power-of-2 sizes 530 */ 531 wrap_around_mask = (eq->eq_bufsz - 1); 532 533 /* Calculate the pointer to the first EQ entry */ 534 eqe = &eq->eq_buf[(cons_indx & wrap_around_mask)]; 535 536 537 _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*eqe)) 538 539 /* 540 * Pull the handler function for this EQ from the Hermon Event Queue 541 * handle 542 */ 543 eqfunction = eq->eq_func; 544 545 for (;;) { 546 polled_some = 0; 547 while (HERMON_EQE_OWNER_IS_SW(eq, eqe, cons_indx, shift)) { 548 549 /* 550 * Call the EQ handler function. But only call if we 551 * are not in polled I/O mode (i.e. not processing 552 * because of a system panic). Note: We don't call 553 * the EQ handling functions from a system panic 554 * because we are primarily concerned only with 555 * ensuring that the event queues do not overflow (or, 556 * more specifically, the event queue associated with 557 * the CQ that is being used in the sync/dump process). 558 * Also, we don't want to make any upcalls (to the 559 * IBTF) because we can't guarantee when/if those 560 * calls would ever return. And, if we're in panic, 561 * then we reached here through a PollCQ() call (from 562 * hermon_cq_poll()), and we need to ensure that we 563 * successfully return any work completions to the 564 * caller. 565 */ 566 if (ddi_in_panic() == 0) { 567 eqfunction(state, eq, eqe); 568 } 569 570 /* Reset to hardware ownership is implicit */ 571 572 /* Increment the consumer index */ 573 cons_indx++; 574 575 /* Update the pointer to the next EQ entry */ 576 eqe = &eq->eq_buf[(cons_indx & wrap_around_mask)]; 577 578 polled_some = 1; 579 } 580 581 /* 582 * write consumer index via EQ set CI Doorbell, to keep overflow 583 * from occuring during poll 584 */ 585 586 eq->eq_consindx = cons_indx; 587 588 /* the FMA retry loop starts. */ 589 hermon_pio_start(state, uarhdl, pio_error, fm_loop_cnt, 590 fm_status, fm_test); 591 592 ddi_put32(uarhdl, eq->eq_doorbell, 593 (cons_indx & HERMON_EQ_CI_MASK) | EQ_ARM_BIT); 594 595 /* the FMA retry loop starts. */ 596 hermon_pio_end(state, uarhdl, pio_error, fm_loop_cnt, 597 fm_status, fm_test); 598 599 if (polled_some == 0) 600 break; 601 }; 602 return; 603 604 pio_error: 605 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_FATAL); 606 } 607 608 609 /* 610 * hermon_eq_catastrophic 611 * Context: Only called from interrupt context (and during panic) 612 */ 613 static void 614 hermon_eq_catastrophic(hermon_state_t *state) 615 { 616 ddi_acc_handle_t cmdhdl = hermon_get_cmdhdl(state); 617 ibt_async_code_t type; 618 ibc_async_event_t event; 619 uint32_t *base_addr; 620 uint32_t buf_size; 621 uint32_t word; 622 uint8_t err_type; 623 uint32_t err_buf; 624 int i; 625 626 /* initialize the FMA retry loop */ 627 hermon_pio_init(fm_loop_cnt, fm_status, fm_test); 628 629 bzero(&event, sizeof (ibc_async_event_t)); 630 base_addr = state->hs_cmd_regs.fw_err_buf; 631 632 buf_size = state->hs_fw.error_buf_sz; /* in #dwords */ 633 634 /* the FMA retry loop starts. */ 635 hermon_pio_start(state, cmdhdl, pio_error, fm_loop_cnt, fm_status, 636 fm_test); 637 638 word = ddi_get32(cmdhdl, base_addr); 639 640 /* the FMA retry loop ends. */ 641 hermon_pio_end(state, cmdhdl, pio_error, fm_loop_cnt, fm_status, 642 fm_test); 643 644 err_type = (word & 0xFF000000) >> 24; 645 type = IBT_ERROR_LOCAL_CATASTROPHIC; 646 647 switch (err_type) { 648 case HERMON_CATASTROPHIC_INTERNAL_ERROR: 649 cmn_err(CE_WARN, "Catastrophic Internal Error: 0x%02x", 650 err_type); 651 652 break; 653 654 case HERMON_CATASTROPHIC_UPLINK_BUS_ERROR: 655 cmn_err(CE_WARN, "Catastrophic Uplink Bus Error: 0x%02x", 656 err_type); 657 658 break; 659 660 case HERMON_CATASTROPHIC_DDR_DATA_ERROR: 661 cmn_err(CE_WARN, "Catastrophic DDR Data Error: 0x%02x", 662 err_type); 663 664 break; 665 666 case HERMON_CATASTROPHIC_INTERNAL_PARITY_ERROR: 667 cmn_err(CE_WARN, "Catastrophic Internal Parity Error: 0x%02x", 668 err_type); 669 670 break; 671 672 default: 673 /* Unknown type of Catastrophic error */ 674 cmn_err(CE_WARN, "Catastrophic Unknown Error: 0x%02x", 675 err_type); 676 677 break; 678 } 679 680 /* the FMA retry loop starts. */ 681 hermon_pio_start(state, cmdhdl, pio_error, fm_loop_cnt, fm_status, 682 fm_test); 683 684 /* 685 * Read in the catastrophic error buffer from the hardware. 686 */ 687 for (i = 0; i < buf_size; i++) { 688 base_addr = 689 (state->hs_cmd_regs.fw_err_buf + i); 690 err_buf = ddi_get32(cmdhdl, base_addr); 691 cmn_err(CE_NOTE, "hermon%d: catastrophic_error[%02x]: %08X", 692 state->hs_instance, i, err_buf); 693 } 694 695 /* the FMA retry loop ends. */ 696 hermon_pio_end(state, cmdhdl, pio_error, fm_loop_cnt, fm_status, 697 fm_test); 698 699 /* 700 * We also call the IBTF here to inform it of the catastrophic error. 701 * Note: Since no event information (i.e. QP handles, CQ handles, 702 * etc.) is necessary, we pass a NULL pointer instead of a pointer to 703 * an empty ibc_async_event_t struct. 704 * 705 * But we also check if "hs_ibtfpriv" is NULL. If it is then it 706 * means that we've have either received this event before we 707 * finished attaching to the IBTF or we've received it while we 708 * are in the process of detaching. 709 */ 710 if (state->hs_ibtfpriv != NULL) { 711 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 712 } 713 714 pio_error: 715 /* ignore these errors but log them because they're harmless. */ 716 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_NON_FATAL); 717 } 718 719 720 /* 721 * hermon_eq_alloc() 722 * Context: Only called from attach() path context 723 */ 724 static int 725 hermon_eq_alloc(hermon_state_t *state, uint32_t log_eq_size, uint_t intr, 726 hermon_eqhdl_t *eqhdl) 727 { 728 hermon_rsrc_t *eqc, *rsrc; 729 hermon_hw_eqc_t eqc_entry; 730 hermon_eqhdl_t eq; 731 ibt_mr_attr_t mr_attr; 732 hermon_mr_options_t op; 733 hermon_pdhdl_t pd; 734 hermon_mrhdl_t mr; 735 hermon_hw_eqe_t *buf; 736 int status; 737 738 /* Use the internal protection domain (PD) for setting up EQs */ 739 pd = state->hs_pdhdl_internal; 740 741 /* Increment the reference count on the protection domain (PD) */ 742 hermon_pd_refcnt_inc(pd); 743 744 /* 745 * Allocate an EQ context entry. This will be filled in with all 746 * the necessary parameters to define the Event Queue. And then 747 * ownership will be passed to the hardware in the final step 748 * below. If we fail here, we must undo the protection domain 749 * reference count. 750 */ 751 status = hermon_rsrc_alloc(state, HERMON_EQC, 1, HERMON_SLEEP, &eqc); 752 if (status != DDI_SUCCESS) { 753 status = DDI_FAILURE; 754 goto eqalloc_fail1; 755 } 756 757 /* 758 * Allocate the software structure for tracking the event queue (i.e. 759 * the Hermon Event Queue handle). If we fail here, we must undo the 760 * protection domain reference count and the previous resource 761 * allocation. 762 */ 763 status = hermon_rsrc_alloc(state, HERMON_EQHDL, 1, HERMON_SLEEP, &rsrc); 764 if (status != DDI_SUCCESS) { 765 status = DDI_FAILURE; 766 goto eqalloc_fail2; 767 } 768 769 eq = (hermon_eqhdl_t)rsrc->hr_addr; 770 771 /* 772 * Allocate the memory for Event Queue. 773 */ 774 eq->eq_eqinfo.qa_size = (1 << log_eq_size) * sizeof (hermon_hw_eqe_t); 775 eq->eq_eqinfo.qa_alloc_align = eq->eq_eqinfo.qa_bind_align = PAGESIZE; 776 777 eq->eq_eqinfo.qa_location = HERMON_QUEUE_LOCATION_NORMAL; 778 status = hermon_queue_alloc(state, &eq->eq_eqinfo, HERMON_SLEEP); 779 if (status != DDI_SUCCESS) { 780 status = DDI_FAILURE; 781 goto eqalloc_fail3; 782 } 783 784 buf = (hermon_hw_eqe_t *)eq->eq_eqinfo.qa_buf_aligned; 785 /* 786 * Initializing each of the Event Queue Entries (EQE) by setting their 787 * ownership to hardware ("owner" bit set to HW) is now done by HW 788 * when the transfer of ownership (below) of the 789 * EQ context itself is done. 790 */ 791 792 /* 793 * Register the memory for the EQ. 794 * 795 * Because we are in the attach path we use NOSLEEP here so that we 796 * SPIN in the HCR since the event queues are not setup yet, and we 797 * cannot NOSPIN at this point in time. 798 */ 799 800 mr_attr.mr_vaddr = (uint64_t)(uintptr_t)buf; 801 mr_attr.mr_len = eq->eq_eqinfo.qa_size; 802 mr_attr.mr_as = NULL; 803 mr_attr.mr_flags = IBT_MR_NOSLEEP | IBT_MR_ENABLE_LOCAL_WRITE; 804 op.mro_bind_type = state->hs_cfg_profile->cp_iommu_bypass; 805 op.mro_bind_dmahdl = eq->eq_eqinfo.qa_dmahdl; 806 op.mro_bind_override_addr = 0; 807 status = hermon_mr_register(state, pd, &mr_attr, &mr, &op, 808 HERMON_EQ_CMPT); 809 if (status != DDI_SUCCESS) { 810 status = DDI_FAILURE; 811 goto eqalloc_fail4; 812 } 813 _NOTE(NOW_INVISIBLE_TO_OTHER_THREADS(*mr)) 814 815 /* 816 * Fill in the EQC entry. This is the final step before passing 817 * ownership of the EQC entry to the Hermon hardware. We use all of 818 * the information collected/calculated above to fill in the 819 * requisite portions of the EQC. Note: We create all EQs in the 820 * "fired" state. We will arm them later (after our interrupt 821 * routine had been registered.) 822 */ 823 bzero(&eqc_entry, sizeof (hermon_hw_eqc_t)); 824 eqc_entry.state = HERMON_EQ_ARMED; 825 eqc_entry.log_eq_sz = log_eq_size; 826 eqc_entry.intr = intr; 827 eqc_entry.log2_pgsz = mr->mr_log2_pgsz; 828 eqc_entry.pg_offs = eq->eq_eqinfo.qa_pgoffs >> 5; 829 eqc_entry.mtt_base_addrh = (uint32_t)((mr->mr_mttaddr >> 32) & 0xFF); 830 eqc_entry.mtt_base_addrl = mr->mr_mttaddr >> 3; 831 eqc_entry.cons_indx = 0x0; 832 eqc_entry.prod_indx = 0x0; 833 834 /* 835 * Write the EQC entry to hardware. Lastly, we pass ownership of 836 * the entry to the hardware (using the Hermon SW2HW_EQ firmware 837 * command). Note: in general, this operation shouldn't fail. But 838 * if it does, we have to undo everything we've done above before 839 * returning error. 840 */ 841 status = hermon_cmn_ownership_cmd_post(state, SW2HW_EQ, &eqc_entry, 842 sizeof (hermon_hw_eqc_t), eqc->hr_indx, HERMON_CMD_NOSLEEP_SPIN); 843 if (status != HERMON_CMD_SUCCESS) { 844 cmn_err(CE_NOTE, "hermon%d: SW2HW_EQ command failed: %08x\n", 845 state->hs_instance, status); 846 if (status == HERMON_CMD_INVALID_STATUS) { 847 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_SRV_LOST); 848 } 849 status = ibc_get_ci_failure(0); 850 goto eqalloc_fail5; 851 } 852 853 /* 854 * Fill in the rest of the Hermon Event Queue handle. Having 855 * successfully transferred ownership of the EQC, we can update the 856 * following fields for use in further operations on the EQ. 857 */ 858 eq->eq_eqcrsrcp = eqc; 859 eq->eq_rsrcp = rsrc; 860 eq->eq_consindx = 0; 861 eq->eq_eqnum = eqc->hr_indx; 862 eq->eq_buf = buf; 863 eq->eq_bufsz = (1 << log_eq_size); 864 eq->eq_log_eqsz = log_eq_size; 865 eq->eq_mrhdl = mr; 866 eq->eq_doorbell = (uint32_t *)((uintptr_t)state->hs_reg_uar_baseaddr + 867 (uint32_t)ARM_EQ_INDEX(eq->eq_eqnum)); 868 *eqhdl = eq; 869 870 return (DDI_SUCCESS); 871 872 /* 873 * The following is cleanup for all possible failure cases in this routine 874 */ 875 eqalloc_fail5: 876 if (hermon_mr_deregister(state, &mr, HERMON_MR_DEREG_ALL, 877 HERMON_NOSLEEP) != DDI_SUCCESS) { 878 HERMON_WARNING(state, "failed to deregister EQ memory"); 879 } 880 eqalloc_fail4: 881 hermon_queue_free(&eq->eq_eqinfo); 882 eqalloc_fail3: 883 hermon_rsrc_free(state, &rsrc); 884 eqalloc_fail2: 885 hermon_rsrc_free(state, &eqc); 886 eqalloc_fail1: 887 hermon_pd_refcnt_dec(pd); 888 eqalloc_fail: 889 return (status); 890 } 891 892 893 /* 894 * hermon_eq_free() 895 * Context: Only called from attach() and/or detach() path contexts 896 */ 897 static int 898 hermon_eq_free(hermon_state_t *state, hermon_eqhdl_t *eqhdl) 899 { 900 hermon_rsrc_t *eqc, *rsrc; 901 hermon_hw_eqc_t eqc_entry; 902 hermon_pdhdl_t pd; 903 hermon_mrhdl_t mr; 904 hermon_eqhdl_t eq; 905 uint32_t eqnum; 906 int status; 907 908 /* 909 * Pull all the necessary information from the Hermon Event Queue 910 * handle. This is necessary here because the resource for the 911 * EQ handle is going to be freed up as part of this operation. 912 */ 913 eq = *eqhdl; 914 eqc = eq->eq_eqcrsrcp; 915 rsrc = eq->eq_rsrcp; 916 pd = state->hs_pdhdl_internal; 917 mr = eq->eq_mrhdl; 918 eqnum = eq->eq_eqnum; 919 920 /* 921 * Reclaim EQC entry from hardware (using the Hermon HW2SW_EQ 922 * firmware command). If the ownership transfer fails for any reason, 923 * then it is an indication that something (either in HW or SW) has 924 * gone seriously wrong. 925 */ 926 status = hermon_cmn_ownership_cmd_post(state, HW2SW_EQ, &eqc_entry, 927 sizeof (hermon_hw_eqc_t), eqnum, HERMON_CMD_NOSLEEP_SPIN); 928 if (status != HERMON_CMD_SUCCESS) { 929 HERMON_WARNING(state, "failed to reclaim EQC ownership"); 930 cmn_err(CE_CONT, "Hermon: HW2SW_EQ command failed: %08x\n", 931 status); 932 return (DDI_FAILURE); 933 } 934 935 /* 936 * Deregister the memory for the Event Queue. If this fails 937 * for any reason, then it is an indication that something (either 938 * in HW or SW) has gone seriously wrong. So we print a warning 939 * message and continue. 940 */ 941 status = hermon_mr_deregister(state, &mr, HERMON_MR_DEREG_ALL, 942 HERMON_NOSLEEP); 943 if (status != DDI_SUCCESS) { 944 HERMON_WARNING(state, "failed to deregister EQ memory"); 945 } 946 947 /* Free the memory for the EQ */ 948 hermon_queue_free(&eq->eq_eqinfo); 949 950 /* Free the Hermon Event Queue handle */ 951 hermon_rsrc_free(state, &rsrc); 952 953 /* Free up the EQC entry resource */ 954 hermon_rsrc_free(state, &eqc); 955 956 /* Decrement the reference count on the protection domain (PD) */ 957 hermon_pd_refcnt_dec(pd); 958 959 /* Set the eqhdl pointer to NULL and return success */ 960 *eqhdl = NULL; 961 962 return (DDI_SUCCESS); 963 } 964 965 966 /* 967 * hermon_eq_handler_init 968 * Context: Only called from attach() path context 969 */ 970 static int 971 hermon_eq_handler_init(hermon_state_t *state, hermon_eqhdl_t eq, 972 uint_t evt_type_mask, int (*eq_func)(hermon_state_t *state, 973 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe)) 974 { 975 int status; 976 977 /* 978 * Save away the EQ handler function and the event type mask. These 979 * will be used later during interrupt and event queue processing. 980 */ 981 eq->eq_func = eq_func; 982 eq->eq_evttypemask = evt_type_mask; 983 984 /* 985 * Map the EQ to a specific class of event (or events) depending 986 * on the mask value passed in. The HERMON_EVT_NO_MASK means not 987 * to attempt associating the EQ with any specific class of event. 988 * This is particularly useful when initializing the events queues 989 * used for CQ events. The mapping is done using the Hermon MAP_EQ 990 * firmware command. Note: This command should not, in general, fail. 991 * If it does, then something (probably HW related) has gone seriously 992 * wrong. 993 */ 994 if (evt_type_mask != HERMON_EVT_NO_MASK) { 995 status = hermon_map_eq_cmd_post(state, 996 HERMON_CMD_MAP_EQ_EVT_MAP, eq->eq_eqnum, evt_type_mask, 997 HERMON_CMD_NOSLEEP_SPIN); 998 if (status != HERMON_CMD_SUCCESS) { 999 cmn_err(CE_NOTE, "hermon%d: MAP_EQ command failed: " 1000 "%08x\n", state->hs_instance, status); 1001 return (DDI_FAILURE); 1002 } 1003 } 1004 1005 return (DDI_SUCCESS); 1006 } 1007 1008 1009 /* 1010 * hermon_eq_handler_fini 1011 * Context: Only called from attach() and/or detach() path contexts 1012 */ 1013 static int 1014 hermon_eq_handler_fini(hermon_state_t *state, hermon_eqhdl_t eq) 1015 { 1016 int status; 1017 1018 /* 1019 * Unmap the EQ from the event class to which it had been previously 1020 * mapped. The unmapping is done using the Hermon MAP_EQ (in much 1021 * the same way that the initial mapping was done). The difference, 1022 * however, is in the HERMON_EQ_EVT_UNMAP flag that is passed to the 1023 * MAP_EQ firmware command. The HERMON_EVT_NO_MASK (which may have 1024 * been passed in at init time) still means that no association has 1025 * been made between the EQ and any specific class of event (and, 1026 * hence, no unmapping is necessary). Note: This command should not, 1027 * in general, fail. If it does, then something (probably HW related) 1028 * has gone seriously wrong. 1029 */ 1030 if (eq->eq_evttypemask != HERMON_EVT_NO_MASK) { 1031 status = hermon_map_eq_cmd_post(state, 1032 HERMON_CMD_MAP_EQ_EVT_UNMAP, eq->eq_eqnum, 1033 eq->eq_evttypemask, HERMON_CMD_NOSLEEP_SPIN); 1034 if (status != HERMON_CMD_SUCCESS) { 1035 cmn_err(CE_NOTE, "hermon%d: MAP_EQ command failed: " 1036 "%08x\n", state->hs_instance, status); 1037 return (DDI_FAILURE); 1038 } 1039 } 1040 1041 return (DDI_SUCCESS); 1042 } 1043 1044 1045 /* 1046 * hermon_eq_demux() 1047 * Context: Called only from interrupt context 1048 * Usage: to demux the various type reported on one EQ 1049 */ 1050 static int 1051 hermon_eq_demux(hermon_state_t *state, hermon_eqhdl_t eq, 1052 hermon_hw_eqe_t *eqe) 1053 { 1054 uint_t eqe_evttype; 1055 int status = DDI_FAILURE; 1056 1057 eqe_evttype = HERMON_EQE_EVTTYPE_GET(eq, eqe); 1058 1059 switch (eqe_evttype) { 1060 1061 case HERMON_EVT_PORT_STATE_CHANGE: 1062 status = hermon_port_state_change_handler(state, eq, eqe); 1063 break; 1064 1065 case HERMON_EVT_COMM_ESTABLISHED: 1066 status = hermon_comm_estbl_handler(state, eq, eqe); 1067 break; 1068 1069 case HERMON_EVT_COMMAND_INTF_COMP: 1070 status = hermon_cmd_complete_handler(state, eq, eqe); 1071 break; 1072 1073 case HERMON_EVT_LOCAL_WQ_CAT_ERROR: 1074 HERMON_WARNING(state, HERMON_FMA_LOCCAT); 1075 status = hermon_local_wq_cat_err_handler(state, eq, eqe); 1076 break; 1077 1078 case HERMON_EVT_INV_REQ_LOCAL_WQ_ERROR: 1079 HERMON_WARNING(state, HERMON_FMA_LOCINV); 1080 status = hermon_invreq_local_wq_err_handler(state, eq, eqe); 1081 break; 1082 1083 case HERMON_EVT_LOCAL_ACC_VIO_WQ_ERROR: 1084 HERMON_WARNING(state, HERMON_FMA_LOCACEQ); 1085 IBTF_DPRINTF_L2("async", HERMON_FMA_LOCACEQ); 1086 status = hermon_local_acc_vio_wq_err_handler(state, eq, eqe); 1087 break; 1088 case HERMON_EVT_SEND_QUEUE_DRAINED: 1089 status = hermon_sendq_drained_handler(state, eq, eqe); 1090 break; 1091 1092 case HERMON_EVT_PATH_MIGRATED: 1093 status = hermon_path_mig_handler(state, eq, eqe); 1094 break; 1095 1096 case HERMON_EVT_PATH_MIGRATE_FAILED: 1097 HERMON_WARNING(state, HERMON_FMA_PATHMIG); 1098 status = hermon_path_mig_err_handler(state, eq, eqe); 1099 break; 1100 1101 case HERMON_EVT_SRQ_CATASTROPHIC_ERROR: 1102 HERMON_WARNING(state, HERMON_FMA_SRQCAT); 1103 status = hermon_catastrophic_handler(state, eq, eqe); 1104 break; 1105 1106 case HERMON_EVT_SRQ_LAST_WQE_REACHED: 1107 status = hermon_srq_last_wqe_reached_handler(state, eq, eqe); 1108 break; 1109 1110 case HERMON_EVT_FEXCH_ERROR: 1111 status = hermon_fexch_error_handler(state, eq, eqe); 1112 break; 1113 1114 default: 1115 break; 1116 } 1117 return (status); 1118 } 1119 1120 /* 1121 * hermon_port_state_change_handler() 1122 * Context: Only called from interrupt context 1123 */ 1124 /* ARGSUSED */ 1125 static int 1126 hermon_port_state_change_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1127 hermon_hw_eqe_t *eqe) 1128 { 1129 ibc_async_event_t event; 1130 ibt_async_code_t type; 1131 uint_t subtype; 1132 uint8_t port; 1133 char link_msg[24]; 1134 1135 /* 1136 * Depending on the type of Port State Change event, pass the 1137 * appropriate asynch event to the IBTF. 1138 */ 1139 port = (uint8_t)HERMON_EQE_PORTNUM_GET(eq, eqe); 1140 1141 /* Check for valid port number in event */ 1142 if ((port == 0) || (port > state->hs_cfg_profile->cp_num_ports)) { 1143 HERMON_WARNING(state, "Unexpected port number in port state " 1144 "change event"); 1145 cmn_err(CE_CONT, " Port number: %02x\n", port); 1146 return (DDI_FAILURE); 1147 } 1148 1149 subtype = HERMON_EQE_EVTSUBTYPE_GET(eq, eqe); 1150 if (subtype == HERMON_PORT_LINK_ACTIVE) { 1151 event.ev_port = port; 1152 type = IBT_EVENT_PORT_UP; 1153 1154 (void) snprintf(link_msg, 23, "port %d up", port); 1155 ddi_dev_report_fault(state->hs_dip, DDI_SERVICE_RESTORED, 1156 DDI_EXTERNAL_FAULT, link_msg); 1157 } else if (subtype == HERMON_PORT_LINK_DOWN) { 1158 event.ev_port = port; 1159 type = IBT_ERROR_PORT_DOWN; 1160 1161 (void) snprintf(link_msg, 23, "port %d down", port); 1162 ddi_dev_report_fault(state->hs_dip, DDI_SERVICE_LOST, 1163 DDI_EXTERNAL_FAULT, link_msg); 1164 } else { 1165 HERMON_WARNING(state, "Unexpected subtype in port state change " 1166 "event"); 1167 cmn_err(CE_CONT, " Event type: %02x, subtype: %02x\n", 1168 HERMON_EQE_EVTTYPE_GET(eq, eqe), subtype); 1169 return (DDI_FAILURE); 1170 } 1171 1172 /* 1173 * Deliver the event to the IBTF. Note: If "hs_ibtfpriv" is NULL, 1174 * then we have either received this event before we finished 1175 * attaching to the IBTF or we've received it while we are in the 1176 * process of detaching. 1177 */ 1178 if (state->hs_ibtfpriv != NULL) { 1179 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1180 } 1181 1182 return (DDI_SUCCESS); 1183 } 1184 1185 1186 /* 1187 * hermon_comm_estbl_handler() 1188 * Context: Only called from interrupt context 1189 */ 1190 /* ARGSUSED */ 1191 static int 1192 hermon_comm_estbl_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1193 hermon_hw_eqe_t *eqe) 1194 { 1195 hermon_qphdl_t qp; 1196 uint_t qpnum; 1197 ibc_async_event_t event; 1198 ibt_async_code_t type; 1199 1200 /* Get the QP handle from QP number in event descriptor */ 1201 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1202 qp = hermon_qphdl_from_qpnum(state, qpnum); 1203 1204 /* 1205 * If the QP handle is NULL, this is probably an indication 1206 * that the QP has been freed already. In which case, we 1207 * should not deliver this event. 1208 * 1209 * We also check that the QP number in the handle is the 1210 * same as the QP number in the event queue entry. This 1211 * extra check allows us to handle the case where a QP was 1212 * freed and then allocated again in the time it took to 1213 * handle the event queue processing. By constantly incrementing 1214 * the non-constrained portion of the QP number every time 1215 * a new QP is allocated, we mitigate (somewhat) the chance 1216 * that a stale event could be passed to the client's QP 1217 * handler. 1218 * 1219 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1220 * means that we've have either received this event before we 1221 * finished attaching to the IBTF or we've received it while we 1222 * are in the process of detaching. 1223 */ 1224 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1225 (state->hs_ibtfpriv != NULL)) { 1226 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1227 type = IBT_EVENT_COM_EST_QP; 1228 1229 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1230 } 1231 1232 return (DDI_SUCCESS); 1233 } 1234 1235 1236 /* 1237 * hermon_local_wq_cat_err_handler() 1238 * Context: Only called from interrupt context 1239 */ 1240 /* ARGSUSED */ 1241 static int 1242 hermon_local_wq_cat_err_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1243 hermon_hw_eqe_t *eqe) 1244 { 1245 hermon_qphdl_t qp; 1246 uint_t qpnum; 1247 ibc_async_event_t event; 1248 ibt_async_code_t type; 1249 1250 /* Get the QP handle from QP number in event descriptor */ 1251 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1252 qp = hermon_qphdl_from_qpnum(state, qpnum); 1253 1254 /* 1255 * If the QP handle is NULL, this is probably an indication 1256 * that the QP has been freed already. In which case, we 1257 * should not deliver this event. 1258 * 1259 * We also check that the QP number in the handle is the 1260 * same as the QP number in the event queue entry. This 1261 * extra check allows us to handle the case where a QP was 1262 * freed and then allocated again in the time it took to 1263 * handle the event queue processing. By constantly incrementing 1264 * the non-constrained portion of the QP number every time 1265 * a new QP is allocated, we mitigate (somewhat) the chance 1266 * that a stale event could be passed to the client's QP 1267 * handler. 1268 * 1269 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1270 * means that we've have either received this event before we 1271 * finished attaching to the IBTF or we've received it while we 1272 * are in the process of detaching. 1273 */ 1274 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1275 (state->hs_ibtfpriv != NULL)) { 1276 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1277 type = IBT_ERROR_CATASTROPHIC_QP; 1278 1279 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1280 } 1281 1282 return (DDI_SUCCESS); 1283 } 1284 1285 1286 /* 1287 * hermon_invreq_local_wq_err_handler() 1288 * Context: Only called from interrupt context 1289 */ 1290 /* ARGSUSED */ 1291 static int 1292 hermon_invreq_local_wq_err_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1293 hermon_hw_eqe_t *eqe) 1294 { 1295 hermon_qphdl_t qp; 1296 uint_t qpnum; 1297 ibc_async_event_t event; 1298 ibt_async_code_t type; 1299 1300 /* Get the QP handle from QP number in event descriptor */ 1301 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1302 qp = hermon_qphdl_from_qpnum(state, qpnum); 1303 1304 /* 1305 * If the QP handle is NULL, this is probably an indication 1306 * that the QP has been freed already. In which case, we 1307 * should not deliver this event. 1308 * 1309 * We also check that the QP number in the handle is the 1310 * same as the QP number in the event queue entry. This 1311 * extra check allows us to handle the case where a QP was 1312 * freed and then allocated again in the time it took to 1313 * handle the event queue processing. By constantly incrementing 1314 * the non-constrained portion of the QP number every time 1315 * a new QP is allocated, we mitigate (somewhat) the chance 1316 * that a stale event could be passed to the client's QP 1317 * handler. 1318 * 1319 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1320 * means that we've have either received this event before we 1321 * finished attaching to the IBTF or we've received it while we 1322 * are in the process of detaching. 1323 */ 1324 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1325 (state->hs_ibtfpriv != NULL)) { 1326 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1327 type = IBT_ERROR_INVALID_REQUEST_QP; 1328 1329 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1330 } 1331 1332 return (DDI_SUCCESS); 1333 } 1334 1335 1336 /* 1337 * hermon_local_acc_vio_wq_err_handler() 1338 * Context: Only called from interrupt context 1339 */ 1340 /* ARGSUSED */ 1341 static int 1342 hermon_local_acc_vio_wq_err_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1343 hermon_hw_eqe_t *eqe) 1344 { 1345 hermon_qphdl_t qp; 1346 uint_t qpnum; 1347 ibc_async_event_t event; 1348 ibt_async_code_t type; 1349 1350 /* Get the QP handle from QP number in event descriptor */ 1351 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1352 qp = hermon_qphdl_from_qpnum(state, qpnum); 1353 1354 /* 1355 * If the QP handle is NULL, this is probably an indication 1356 * that the QP has been freed already. In which case, we 1357 * should not deliver this event. 1358 * 1359 * We also check that the QP number in the handle is the 1360 * same as the QP number in the event queue entry. This 1361 * extra check allows us to handle the case where a QP was 1362 * freed and then allocated again in the time it took to 1363 * handle the event queue processing. By constantly incrementing 1364 * the non-constrained portion of the QP number every time 1365 * a new QP is allocated, we mitigate (somewhat) the chance 1366 * that a stale event could be passed to the client's QP 1367 * handler. 1368 * 1369 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1370 * means that we've have either received this event before we 1371 * finished attaching to the IBTF or we've received it while we 1372 * are in the process of detaching. 1373 */ 1374 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1375 (state->hs_ibtfpriv != NULL)) { 1376 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1377 type = IBT_ERROR_ACCESS_VIOLATION_QP; 1378 1379 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1380 } 1381 1382 return (DDI_SUCCESS); 1383 } 1384 1385 1386 /* 1387 * hermon_sendq_drained_handler() 1388 * Context: Only called from interrupt context 1389 */ 1390 /* ARGSUSED */ 1391 static int 1392 hermon_sendq_drained_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1393 hermon_hw_eqe_t *eqe) 1394 { 1395 hermon_qphdl_t qp; 1396 uint_t qpnum; 1397 ibc_async_event_t event; 1398 uint_t forward_sqd_event; 1399 ibt_async_code_t type; 1400 1401 /* Get the QP handle from QP number in event descriptor */ 1402 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1403 qp = hermon_qphdl_from_qpnum(state, qpnum); 1404 1405 /* 1406 * If the QP handle is NULL, this is probably an indication 1407 * that the QP has been freed already. In which case, we 1408 * should not deliver this event. 1409 * 1410 * We also check that the QP number in the handle is the 1411 * same as the QP number in the event queue entry. This 1412 * extra check allows us to handle the case where a QP was 1413 * freed and then allocated again in the time it took to 1414 * handle the event queue processing. By constantly incrementing 1415 * the non-constrained portion of the QP number every time 1416 * a new QP is allocated, we mitigate (somewhat) the chance 1417 * that a stale event could be passed to the client's QP 1418 * handler. 1419 * 1420 * And then we check if "hs_ibtfpriv" is NULL. If it is then it 1421 * means that we've have either received this event before we 1422 * finished attaching to the IBTF or we've received it while we 1423 * are in the process of detaching. 1424 */ 1425 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1426 (state->hs_ibtfpriv != NULL)) { 1427 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1428 type = IBT_EVENT_SQD; 1429 1430 /* 1431 * Grab the QP lock and update the QP state to reflect that 1432 * the Send Queue Drained event has arrived. Also determine 1433 * whether the event is intended to be forwarded on to the 1434 * consumer or not. This information is used below in 1435 * determining whether or not to call the IBTF. 1436 */ 1437 mutex_enter(&qp->qp_lock); 1438 forward_sqd_event = qp->qp_forward_sqd_event; 1439 qp->qp_forward_sqd_event = 0; 1440 qp->qp_sqd_still_draining = 0; 1441 mutex_exit(&qp->qp_lock); 1442 1443 if (forward_sqd_event != 0) { 1444 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1445 } 1446 } 1447 1448 return (DDI_SUCCESS); 1449 } 1450 1451 1452 /* 1453 * hermon_path_mig_handler() 1454 * Context: Only called from interrupt context 1455 */ 1456 /* ARGSUSED */ 1457 static int 1458 hermon_path_mig_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1459 hermon_hw_eqe_t *eqe) 1460 { 1461 hermon_qphdl_t qp; 1462 uint_t qpnum; 1463 ibc_async_event_t event; 1464 ibt_async_code_t type; 1465 1466 /* Get the QP handle from QP number in event descriptor */ 1467 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1468 qp = hermon_qphdl_from_qpnum(state, qpnum); 1469 1470 /* 1471 * If the QP handle is NULL, this is probably an indication 1472 * that the QP has been freed already. In which case, we 1473 * should not deliver this event. 1474 * 1475 * We also check that the QP number in the handle is the 1476 * same as the QP number in the event queue entry. This 1477 * extra check allows us to handle the case where a QP was 1478 * freed and then allocated again in the time it took to 1479 * handle the event queue processing. By constantly incrementing 1480 * the non-constrained portion of the QP number every time 1481 * a new QP is allocated, we mitigate (somewhat) the chance 1482 * that a stale event could be passed to the client's QP 1483 * handler. 1484 * 1485 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1486 * means that we've have either received this event before we 1487 * finished attaching to the IBTF or we've received it while we 1488 * are in the process of detaching. 1489 */ 1490 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1491 (state->hs_ibtfpriv != NULL)) { 1492 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1493 type = IBT_EVENT_PATH_MIGRATED_QP; 1494 1495 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1496 } 1497 1498 return (DDI_SUCCESS); 1499 } 1500 1501 1502 /* 1503 * hermon_path_mig_err_handler() 1504 * Context: Only called from interrupt context 1505 */ 1506 /* ARGSUSED */ 1507 static int 1508 hermon_path_mig_err_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1509 hermon_hw_eqe_t *eqe) 1510 { 1511 hermon_qphdl_t qp; 1512 uint_t qpnum; 1513 ibc_async_event_t event; 1514 ibt_async_code_t type; 1515 1516 /* Get the QP handle from QP number in event descriptor */ 1517 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1518 qp = hermon_qphdl_from_qpnum(state, qpnum); 1519 1520 /* 1521 * If the QP handle is NULL, this is probably an indication 1522 * that the QP has been freed already. In which case, we 1523 * should not deliver this event. 1524 * 1525 * We also check that the QP number in the handle is the 1526 * same as the QP number in the event queue entry. This 1527 * extra check allows us to handle the case where a QP was 1528 * freed and then allocated again in the time it took to 1529 * handle the event queue processing. By constantly incrementing 1530 * the non-constrained portion of the QP number every time 1531 * a new QP is allocated, we mitigate (somewhat) the chance 1532 * that a stale event could be passed to the client's QP 1533 * handler. 1534 * 1535 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1536 * means that we've have either received this event before we 1537 * finished attaching to the IBTF or we've received it while we 1538 * are in the process of detaching. 1539 */ 1540 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1541 (state->hs_ibtfpriv != NULL)) { 1542 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1543 type = IBT_ERROR_PATH_MIGRATE_REQ_QP; 1544 1545 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1546 } 1547 1548 return (DDI_SUCCESS); 1549 } 1550 1551 1552 /* 1553 * hermon_catastrophic_handler() 1554 * Context: Only called from interrupt context 1555 */ 1556 /* ARGSUSED */ 1557 static int 1558 hermon_catastrophic_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1559 hermon_hw_eqe_t *eqe) 1560 { 1561 hermon_qphdl_t qp; 1562 uint_t qpnum; 1563 ibc_async_event_t event; 1564 ibt_async_code_t type; 1565 1566 if (eq->eq_evttypemask == HERMON_EVT_MSK_LOCAL_CAT_ERROR) { 1567 HERMON_FMANOTE(state, HERMON_FMA_INTERNAL); 1568 hermon_eq_catastrophic(state); 1569 return (DDI_SUCCESS); 1570 } 1571 1572 /* Get the QP handle from QP number in event descriptor */ 1573 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1574 qp = hermon_qphdl_from_qpnum(state, qpnum); 1575 1576 /* 1577 * If the QP handle is NULL, this is probably an indication 1578 * that the QP has been freed already. In which case, we 1579 * should not deliver this event. 1580 * 1581 * We also check that the QP number in the handle is the 1582 * same as the QP number in the event queue entry. This 1583 * extra check allows us to handle the case where a QP was 1584 * freed and then allocated again in the time it took to 1585 * handle the event queue processing. By constantly incrementing 1586 * the non-constrained portion of the QP number every time 1587 * a new QP is allocated, we mitigate (somewhat) the chance 1588 * that a stale event could be passed to the client's QP 1589 * handler. 1590 * 1591 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1592 * means that we've have either received this event before we 1593 * finished attaching to the IBTF or we've received it while we 1594 * are in the process of detaching. 1595 */ 1596 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1597 (state->hs_ibtfpriv != NULL)) { 1598 event.ev_srq_hdl = (ibt_srq_hdl_t)qp->qp_srqhdl->srq_hdlrarg; 1599 type = IBT_ERROR_CATASTROPHIC_SRQ; 1600 1601 mutex_enter(&qp->qp_srqhdl->srq_lock); 1602 qp->qp_srqhdl->srq_state = HERMON_SRQ_STATE_ERROR; 1603 mutex_exit(&qp->qp_srqhdl->srq_lock); 1604 1605 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1606 } 1607 1608 return (DDI_SUCCESS); 1609 } 1610 1611 1612 /* 1613 * hermon_srq_last_wqe_reached_handler() 1614 * Context: Only called from interrupt context 1615 */ 1616 /* ARGSUSED */ 1617 static int 1618 hermon_srq_last_wqe_reached_handler(hermon_state_t *state, hermon_eqhdl_t eq, 1619 hermon_hw_eqe_t *eqe) 1620 { 1621 hermon_qphdl_t qp; 1622 uint_t qpnum; 1623 ibc_async_event_t event; 1624 ibt_async_code_t type; 1625 1626 /* Get the QP handle from QP number in event descriptor */ 1627 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe); 1628 qp = hermon_qphdl_from_qpnum(state, qpnum); 1629 1630 /* 1631 * If the QP handle is NULL, this is probably an indication 1632 * that the QP has been freed already. In which case, we 1633 * should not deliver this event. 1634 * 1635 * We also check that the QP number in the handle is the 1636 * same as the QP number in the event queue entry. This 1637 * extra check allows us to handle the case where a QP was 1638 * freed and then allocated again in the time it took to 1639 * handle the event queue processing. By constantly incrementing 1640 * the non-constrained portion of the QP number every time 1641 * a new QP is allocated, we mitigate (somewhat) the chance 1642 * that a stale event could be passed to the client's QP 1643 * handler. 1644 * 1645 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1646 * means that we've have either received this event before we 1647 * finished attaching to the IBTF or we've received it while we 1648 * are in the process of detaching. 1649 */ 1650 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1651 (state->hs_ibtfpriv != NULL)) { 1652 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1653 type = IBT_EVENT_EMPTY_CHAN; 1654 1655 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1656 } 1657 1658 return (DDI_SUCCESS); 1659 } 1660 1661 1662 /* ARGSUSED */ 1663 static int hermon_fexch_error_handler(hermon_state_t *state, 1664 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe) 1665 { 1666 hermon_qphdl_t qp; 1667 uint_t qpnum; 1668 ibc_async_event_t event; 1669 ibt_async_code_t type; 1670 1671 /* Get the QP handle from QP number in event descriptor */ 1672 event.ev_port = HERMON_EQE_FEXCH_PORTNUM_GET(eq, eqe); 1673 qpnum = hermon_fcoib_qpnum_from_fexch(state, 1674 event.ev_port, HERMON_EQE_FEXCH_FEXCH_GET(eq, eqe)); 1675 qp = hermon_qphdl_from_qpnum(state, qpnum); 1676 1677 event.ev_fc = HERMON_EQE_FEXCH_SYNDROME_GET(eq, eqe); 1678 1679 /* 1680 * If the QP handle is NULL, this is probably an indication 1681 * that the QP has been freed already. In which case, we 1682 * should not deliver this event. 1683 * 1684 * We also check that the QP number in the handle is the 1685 * same as the QP number in the event queue entry. This 1686 * extra check allows us to handle the case where a QP was 1687 * freed and then allocated again in the time it took to 1688 * handle the event queue processing. By constantly incrementing 1689 * the non-constrained portion of the QP number every time 1690 * a new QP is allocated, we mitigate (somewhat) the chance 1691 * that a stale event could be passed to the client's QP 1692 * handler. 1693 * 1694 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it 1695 * means that we've have either received this event before we 1696 * finished attaching to the IBTF or we've received it while we 1697 * are in the process of detaching. 1698 */ 1699 if ((qp != NULL) && (qp->qp_qpnum == qpnum) && 1700 (state->hs_ibtfpriv != NULL)) { 1701 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg; 1702 type = IBT_FEXCH_ERROR; 1703 1704 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event); 1705 } 1706 1707 return (DDI_SUCCESS); 1708 } 1709 1710 1711 /* 1712 * hermon_no_eqhandler 1713 * Context: Only called from interrupt context 1714 */ 1715 /* ARGSUSED */ 1716 static int 1717 hermon_no_eqhandler(hermon_state_t *state, hermon_eqhdl_t eq, 1718 hermon_hw_eqe_t *eqe) 1719 { 1720 uint_t data; 1721 int i; 1722 1723 /* 1724 * This "unexpected event" handler (or "catch-all" handler) will 1725 * receive all events for which no other handler has been registered. 1726 * If we end up here, then something has probably gone seriously wrong 1727 * with the Hermon hardware (or, perhaps, with the software... though 1728 * it's unlikely in this case). The EQE provides all the information 1729 * about the event. So we print a warning message here along with 1730 * the contents of the EQE. 1731 */ 1732 HERMON_WARNING(state, "Unexpected Event handler"); 1733 cmn_err(CE_CONT, " Event type: %02x, subtype: %02x\n", 1734 HERMON_EQE_EVTTYPE_GET(eq, eqe), 1735 HERMON_EQE_EVTSUBTYPE_GET(eq, eqe)); 1736 for (i = 0; i < sizeof (hermon_hw_eqe_t) >> 2; i++) { 1737 data = ((uint_t *)eqe)[i]; 1738 cmn_err(CE_CONT, " EQE[%02x]: %08x\n", i, data); 1739 } 1740 1741 return (DDI_SUCCESS); 1742 } 1743