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 /* Get the consumer pointer index */
522 cons_indx = eq->eq_consindx;
523 shift = eq->eq_log_eqsz - HERMON_EQE_OWNER_SHIFT;
524
525 /*
526 * Calculate the wrap around mask. Note: This operation only works
527 * because all Hermon event queues have power-of-2 sizes
528 */
529 wrap_around_mask = (eq->eq_bufsz - 1);
530
531 /* Calculate the pointer to the first EQ entry */
532 eqe = &eq->eq_buf[(cons_indx & wrap_around_mask)];
533
534
535 /*
536 * Pull the handler function for this EQ from the Hermon Event Queue
537 * handle
538 */
539 eqfunction = eq->eq_func;
540
541 for (;;) {
542 polled_some = 0;
543 while (HERMON_EQE_OWNER_IS_SW(eq, eqe, cons_indx, shift)) {
544
545 /*
546 * Call the EQ handler function. But only call if we
547 * are not in polled I/O mode (i.e. not processing
548 * because of a system panic). Note: We don't call
549 * the EQ handling functions from a system panic
550 * because we are primarily concerned only with
551 * ensuring that the event queues do not overflow (or,
552 * more specifically, the event queue associated with
553 * the CQ that is being used in the sync/dump process).
554 * Also, we don't want to make any upcalls (to the
555 * IBTF) because we can't guarantee when/if those
556 * calls would ever return. And, if we're in panic,
557 * then we reached here through a PollCQ() call (from
558 * hermon_cq_poll()), and we need to ensure that we
559 * successfully return any work completions to the
560 * caller.
561 */
562 if (ddi_in_panic() == 0) {
563 eqfunction(state, eq, eqe);
564 }
565
566 /* Reset to hardware ownership is implicit */
567
568 /* Increment the consumer index */
569 cons_indx++;
570
571 /* Update the pointer to the next EQ entry */
572 eqe = &eq->eq_buf[(cons_indx & wrap_around_mask)];
573
574 polled_some = 1;
575 }
576
577 /*
578 * write consumer index via EQ set CI Doorbell, to keep overflow
579 * from occuring during poll
580 */
581
582 eq->eq_consindx = cons_indx;
583
584 /* the FMA retry loop starts. */
585 hermon_pio_start(state, uarhdl, pio_error, fm_loop_cnt,
586 fm_status, fm_test);
587
588 ddi_put32(uarhdl, eq->eq_doorbell,
589 (cons_indx & HERMON_EQ_CI_MASK) | EQ_ARM_BIT);
590
591 /* the FMA retry loop starts. */
592 hermon_pio_end(state, uarhdl, pio_error, fm_loop_cnt,
593 fm_status, fm_test);
594
595 if (polled_some == 0)
596 break;
597 };
598 return;
599
600 pio_error:
601 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_FATAL);
602 }
603
604
605 /*
606 * hermon_eq_catastrophic
607 * Context: Only called from interrupt context (and during panic)
608 */
609 static void
610 hermon_eq_catastrophic(hermon_state_t *state)
611 {
612 ddi_acc_handle_t cmdhdl = hermon_get_cmdhdl(state);
613 ibt_async_code_t type;
614 ibc_async_event_t event;
615 uint32_t *base_addr;
616 uint32_t buf_size;
617 uint32_t word;
618 uint8_t err_type;
619 uint32_t err_buf;
620 int i;
621
622 /* initialize the FMA retry loop */
623 hermon_pio_init(fm_loop_cnt, fm_status, fm_test);
624
625 bzero(&event, sizeof (ibc_async_event_t));
626 base_addr = state->hs_cmd_regs.fw_err_buf;
627
628 buf_size = state->hs_fw.error_buf_sz; /* in #dwords */
629
630 /* the FMA retry loop starts. */
631 hermon_pio_start(state, cmdhdl, pio_error, fm_loop_cnt, fm_status,
632 fm_test);
633
634 word = ddi_get32(cmdhdl, base_addr);
635
636 /* the FMA retry loop ends. */
637 hermon_pio_end(state, cmdhdl, pio_error, fm_loop_cnt, fm_status,
638 fm_test);
639
640 err_type = (word & 0xFF000000) >> 24;
641 type = IBT_ERROR_LOCAL_CATASTROPHIC;
642
643 switch (err_type) {
644 case HERMON_CATASTROPHIC_INTERNAL_ERROR:
645 cmn_err(CE_WARN, "Catastrophic Internal Error: 0x%02x",
646 err_type);
647
648 break;
649
650 case HERMON_CATASTROPHIC_UPLINK_BUS_ERROR:
651 cmn_err(CE_WARN, "Catastrophic Uplink Bus Error: 0x%02x",
652 err_type);
653
654 break;
655
656 case HERMON_CATASTROPHIC_DDR_DATA_ERROR:
657 cmn_err(CE_WARN, "Catastrophic DDR Data Error: 0x%02x",
658 err_type);
659
660 break;
661
662 case HERMON_CATASTROPHIC_INTERNAL_PARITY_ERROR:
663 cmn_err(CE_WARN, "Catastrophic Internal Parity Error: 0x%02x",
664 err_type);
665
666 break;
667
668 default:
669 /* Unknown type of Catastrophic error */
670 cmn_err(CE_WARN, "Catastrophic Unknown Error: 0x%02x",
671 err_type);
672
673 break;
674 }
675
676 /* the FMA retry loop starts. */
677 hermon_pio_start(state, cmdhdl, pio_error, fm_loop_cnt, fm_status,
678 fm_test);
679
680 /*
681 * Read in the catastrophic error buffer from the hardware.
682 */
683 for (i = 0; i < buf_size; i++) {
684 base_addr =
685 (state->hs_cmd_regs.fw_err_buf + i);
686 err_buf = ddi_get32(cmdhdl, base_addr);
687 cmn_err(CE_NOTE, "hermon%d: catastrophic_error[%02x]: %08X",
688 state->hs_instance, i, err_buf);
689 }
690
691 /* the FMA retry loop ends. */
692 hermon_pio_end(state, cmdhdl, pio_error, fm_loop_cnt, fm_status,
693 fm_test);
694
695 /*
696 * We also call the IBTF here to inform it of the catastrophic error.
697 * Note: Since no event information (i.e. QP handles, CQ handles,
698 * etc.) is necessary, we pass a NULL pointer instead of a pointer to
699 * an empty ibc_async_event_t struct.
700 *
701 * But we also check if "hs_ibtfpriv" is NULL. If it is then it
702 * means that we've have either received this event before we
703 * finished attaching to the IBTF or we've received it while we
704 * are in the process of detaching.
705 */
706 if (state->hs_ibtfpriv != NULL) {
707 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
708 }
709
710 pio_error:
711 /* ignore these errors but log them because they're harmless. */
712 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_NON_FATAL);
713 }
714
715
716 /*
717 * hermon_eq_alloc()
718 * Context: Only called from attach() path context
719 */
720 static int
721 hermon_eq_alloc(hermon_state_t *state, uint32_t log_eq_size, uint_t intr,
722 hermon_eqhdl_t *eqhdl)
723 {
724 hermon_rsrc_t *eqc, *rsrc;
725 hermon_hw_eqc_t eqc_entry;
726 hermon_eqhdl_t eq;
727 ibt_mr_attr_t mr_attr;
728 hermon_mr_options_t op;
729 hermon_pdhdl_t pd;
730 hermon_mrhdl_t mr;
731 hermon_hw_eqe_t *buf;
732 int status;
733
734 /* Use the internal protection domain (PD) for setting up EQs */
735 pd = state->hs_pdhdl_internal;
736
737 /* Increment the reference count on the protection domain (PD) */
738 hermon_pd_refcnt_inc(pd);
739
740 /*
741 * Allocate an EQ context entry. This will be filled in with all
742 * the necessary parameters to define the Event Queue. And then
743 * ownership will be passed to the hardware in the final step
744 * below. If we fail here, we must undo the protection domain
745 * reference count.
746 */
747 status = hermon_rsrc_alloc(state, HERMON_EQC, 1, HERMON_SLEEP, &eqc);
748 if (status != DDI_SUCCESS) {
749 status = DDI_FAILURE;
750 goto eqalloc_fail1;
751 }
752
753 /*
754 * Allocate the software structure for tracking the event queue (i.e.
755 * the Hermon Event Queue handle). If we fail here, we must undo the
756 * protection domain reference count and the previous resource
757 * allocation.
758 */
759 status = hermon_rsrc_alloc(state, HERMON_EQHDL, 1, HERMON_SLEEP, &rsrc);
760 if (status != DDI_SUCCESS) {
761 status = DDI_FAILURE;
762 goto eqalloc_fail2;
763 }
764
765 eq = (hermon_eqhdl_t)rsrc->hr_addr;
766
767 /*
768 * Allocate the memory for Event Queue.
769 */
770 eq->eq_eqinfo.qa_size = (1 << log_eq_size) * sizeof (hermon_hw_eqe_t);
771 eq->eq_eqinfo.qa_alloc_align = eq->eq_eqinfo.qa_bind_align = PAGESIZE;
772
773 eq->eq_eqinfo.qa_location = HERMON_QUEUE_LOCATION_NORMAL;
774 status = hermon_queue_alloc(state, &eq->eq_eqinfo, HERMON_SLEEP);
775 if (status != DDI_SUCCESS) {
776 status = DDI_FAILURE;
777 goto eqalloc_fail3;
778 }
779
780 buf = (hermon_hw_eqe_t *)eq->eq_eqinfo.qa_buf_aligned;
781 /*
782 * Initializing each of the Event Queue Entries (EQE) by setting their
783 * ownership to hardware ("owner" bit set to HW) is now done by HW
784 * when the transfer of ownership (below) of the
785 * EQ context itself is done.
786 */
787
788 /*
789 * Register the memory for the EQ.
790 *
791 * Because we are in the attach path we use NOSLEEP here so that we
792 * SPIN in the HCR since the event queues are not setup yet, and we
793 * cannot NOSPIN at this point in time.
794 */
795
796 mr_attr.mr_vaddr = (uint64_t)(uintptr_t)buf;
797 mr_attr.mr_len = eq->eq_eqinfo.qa_size;
798 mr_attr.mr_as = NULL;
799 mr_attr.mr_flags = IBT_MR_NOSLEEP | IBT_MR_ENABLE_LOCAL_WRITE;
800 op.mro_bind_type = state->hs_cfg_profile->cp_iommu_bypass;
801 op.mro_bind_dmahdl = eq->eq_eqinfo.qa_dmahdl;
802 op.mro_bind_override_addr = 0;
803 status = hermon_mr_register(state, pd, &mr_attr, &mr, &op,
804 HERMON_EQ_CMPT);
805 if (status != DDI_SUCCESS) {
806 status = DDI_FAILURE;
807 goto eqalloc_fail4;
808 }
809
810 /*
811 * Fill in the EQC entry. This is the final step before passing
812 * ownership of the EQC entry to the Hermon hardware. We use all of
813 * the information collected/calculated above to fill in the
814 * requisite portions of the EQC. Note: We create all EQs in the
815 * "fired" state. We will arm them later (after our interrupt
816 * routine had been registered.)
817 */
818 bzero(&eqc_entry, sizeof (hermon_hw_eqc_t));
819 eqc_entry.state = HERMON_EQ_ARMED;
820 eqc_entry.log_eq_sz = log_eq_size;
821 eqc_entry.intr = intr;
822 eqc_entry.log2_pgsz = mr->mr_log2_pgsz;
823 eqc_entry.pg_offs = eq->eq_eqinfo.qa_pgoffs >> 5;
824 eqc_entry.mtt_base_addrh = (uint32_t)((mr->mr_mttaddr >> 32) & 0xFF);
825 eqc_entry.mtt_base_addrl = mr->mr_mttaddr >> 3;
826 eqc_entry.cons_indx = 0x0;
827 eqc_entry.prod_indx = 0x0;
828
829 /*
830 * Write the EQC entry to hardware. Lastly, we pass ownership of
831 * the entry to the hardware (using the Hermon SW2HW_EQ firmware
832 * command). Note: in general, this operation shouldn't fail. But
833 * if it does, we have to undo everything we've done above before
834 * returning error.
835 */
836 status = hermon_cmn_ownership_cmd_post(state, SW2HW_EQ, &eqc_entry,
837 sizeof (hermon_hw_eqc_t), eqc->hr_indx, HERMON_CMD_NOSLEEP_SPIN);
838 if (status != HERMON_CMD_SUCCESS) {
839 cmn_err(CE_NOTE, "hermon%d: SW2HW_EQ command failed: %08x\n",
840 state->hs_instance, status);
841 if (status == HERMON_CMD_INVALID_STATUS) {
842 hermon_fm_ereport(state, HCA_SYS_ERR, HCA_ERR_SRV_LOST);
843 }
844 status = ibc_get_ci_failure(0);
845 goto eqalloc_fail5;
846 }
847
848 /*
849 * Fill in the rest of the Hermon Event Queue handle. Having
850 * successfully transferred ownership of the EQC, we can update the
851 * following fields for use in further operations on the EQ.
852 */
853 eq->eq_eqcrsrcp = eqc;
854 eq->eq_rsrcp = rsrc;
855 eq->eq_consindx = 0;
856 eq->eq_eqnum = eqc->hr_indx;
857 eq->eq_buf = buf;
858 eq->eq_bufsz = (1 << log_eq_size);
859 eq->eq_log_eqsz = log_eq_size;
860 eq->eq_mrhdl = mr;
861 eq->eq_doorbell = (uint32_t *)((uintptr_t)state->hs_reg_uar_baseaddr +
862 (uint32_t)ARM_EQ_INDEX(eq->eq_eqnum));
863 *eqhdl = eq;
864
865 return (DDI_SUCCESS);
866
867 /*
868 * The following is cleanup for all possible failure cases in this routine
869 */
870 eqalloc_fail5:
871 if (hermon_mr_deregister(state, &mr, HERMON_MR_DEREG_ALL,
872 HERMON_NOSLEEP) != DDI_SUCCESS) {
873 HERMON_WARNING(state, "failed to deregister EQ memory");
874 }
875 eqalloc_fail4:
876 hermon_queue_free(&eq->eq_eqinfo);
877 eqalloc_fail3:
878 hermon_rsrc_free(state, &rsrc);
879 eqalloc_fail2:
880 hermon_rsrc_free(state, &eqc);
881 eqalloc_fail1:
882 hermon_pd_refcnt_dec(pd);
883 eqalloc_fail:
884 return (status);
885 }
886
887
888 /*
889 * hermon_eq_free()
890 * Context: Only called from attach() and/or detach() path contexts
891 */
892 static int
893 hermon_eq_free(hermon_state_t *state, hermon_eqhdl_t *eqhdl)
894 {
895 hermon_rsrc_t *eqc, *rsrc;
896 hermon_hw_eqc_t eqc_entry;
897 hermon_pdhdl_t pd;
898 hermon_mrhdl_t mr;
899 hermon_eqhdl_t eq;
900 uint32_t eqnum;
901 int status;
902
903 /*
904 * Pull all the necessary information from the Hermon Event Queue
905 * handle. This is necessary here because the resource for the
906 * EQ handle is going to be freed up as part of this operation.
907 */
908 eq = *eqhdl;
909 eqc = eq->eq_eqcrsrcp;
910 rsrc = eq->eq_rsrcp;
911 pd = state->hs_pdhdl_internal;
912 mr = eq->eq_mrhdl;
913 eqnum = eq->eq_eqnum;
914
915 /*
916 * Reclaim EQC entry from hardware (using the Hermon HW2SW_EQ
917 * firmware command). If the ownership transfer fails for any reason,
918 * then it is an indication that something (either in HW or SW) has
919 * gone seriously wrong.
920 */
921 status = hermon_cmn_ownership_cmd_post(state, HW2SW_EQ, &eqc_entry,
922 sizeof (hermon_hw_eqc_t), eqnum, HERMON_CMD_NOSLEEP_SPIN);
923 if (status != HERMON_CMD_SUCCESS) {
924 HERMON_WARNING(state, "failed to reclaim EQC ownership");
925 cmn_err(CE_CONT, "Hermon: HW2SW_EQ command failed: %08x\n",
926 status);
927 return (DDI_FAILURE);
928 }
929
930 /*
931 * Deregister the memory for the Event Queue. If this fails
932 * for any reason, then it is an indication that something (either
933 * in HW or SW) has gone seriously wrong. So we print a warning
934 * message and continue.
935 */
936 status = hermon_mr_deregister(state, &mr, HERMON_MR_DEREG_ALL,
937 HERMON_NOSLEEP);
938 if (status != DDI_SUCCESS) {
939 HERMON_WARNING(state, "failed to deregister EQ memory");
940 }
941
942 /* Free the memory for the EQ */
943 hermon_queue_free(&eq->eq_eqinfo);
944
945 /* Free the Hermon Event Queue handle */
946 hermon_rsrc_free(state, &rsrc);
947
948 /* Free up the EQC entry resource */
949 hermon_rsrc_free(state, &eqc);
950
951 /* Decrement the reference count on the protection domain (PD) */
952 hermon_pd_refcnt_dec(pd);
953
954 /* Set the eqhdl pointer to NULL and return success */
955 *eqhdl = NULL;
956
957 return (DDI_SUCCESS);
958 }
959
960
961 /*
962 * hermon_eq_handler_init
963 * Context: Only called from attach() path context
964 */
965 static int
966 hermon_eq_handler_init(hermon_state_t *state, hermon_eqhdl_t eq,
967 uint_t evt_type_mask, int (*eq_func)(hermon_state_t *state,
968 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe))
969 {
970 int status;
971
972 /*
973 * Save away the EQ handler function and the event type mask. These
974 * will be used later during interrupt and event queue processing.
975 */
976 eq->eq_func = eq_func;
977 eq->eq_evttypemask = evt_type_mask;
978
979 /*
980 * Map the EQ to a specific class of event (or events) depending
981 * on the mask value passed in. The HERMON_EVT_NO_MASK means not
982 * to attempt associating the EQ with any specific class of event.
983 * This is particularly useful when initializing the events queues
984 * used for CQ events. The mapping is done using the Hermon MAP_EQ
985 * firmware command. Note: This command should not, in general, fail.
986 * If it does, then something (probably HW related) has gone seriously
987 * wrong.
988 */
989 if (evt_type_mask != HERMON_EVT_NO_MASK) {
990 status = hermon_map_eq_cmd_post(state,
991 HERMON_CMD_MAP_EQ_EVT_MAP, eq->eq_eqnum, evt_type_mask,
992 HERMON_CMD_NOSLEEP_SPIN);
993 if (status != HERMON_CMD_SUCCESS) {
994 cmn_err(CE_NOTE, "hermon%d: MAP_EQ command failed: "
995 "%08x\n", state->hs_instance, status);
996 return (DDI_FAILURE);
997 }
998 }
999
1000 return (DDI_SUCCESS);
1001 }
1002
1003
1004 /*
1005 * hermon_eq_handler_fini
1006 * Context: Only called from attach() and/or detach() path contexts
1007 */
1008 static int
1009 hermon_eq_handler_fini(hermon_state_t *state, hermon_eqhdl_t eq)
1010 {
1011 int status;
1012
1013 /*
1014 * Unmap the EQ from the event class to which it had been previously
1015 * mapped. The unmapping is done using the Hermon MAP_EQ (in much
1016 * the same way that the initial mapping was done). The difference,
1017 * however, is in the HERMON_EQ_EVT_UNMAP flag that is passed to the
1018 * MAP_EQ firmware command. The HERMON_EVT_NO_MASK (which may have
1019 * been passed in at init time) still means that no association has
1020 * been made between the EQ and any specific class of event (and,
1021 * hence, no unmapping is necessary). Note: This command should not,
1022 * in general, fail. If it does, then something (probably HW related)
1023 * has gone seriously wrong.
1024 */
1025 if (eq->eq_evttypemask != HERMON_EVT_NO_MASK) {
1026 status = hermon_map_eq_cmd_post(state,
1027 HERMON_CMD_MAP_EQ_EVT_UNMAP, eq->eq_eqnum,
1028 eq->eq_evttypemask, HERMON_CMD_NOSLEEP_SPIN);
1029 if (status != HERMON_CMD_SUCCESS) {
1030 cmn_err(CE_NOTE, "hermon%d: MAP_EQ command failed: "
1031 "%08x\n", state->hs_instance, status);
1032 return (DDI_FAILURE);
1033 }
1034 }
1035
1036 return (DDI_SUCCESS);
1037 }
1038
1039
1040 /*
1041 * hermon_eq_demux()
1042 * Context: Called only from interrupt context
1043 * Usage: to demux the various type reported on one EQ
1044 */
1045 static int
1046 hermon_eq_demux(hermon_state_t *state, hermon_eqhdl_t eq,
1047 hermon_hw_eqe_t *eqe)
1048 {
1049 uint_t eqe_evttype;
1050 int status = DDI_FAILURE;
1051
1052 eqe_evttype = HERMON_EQE_EVTTYPE_GET(eq, eqe);
1053
1054 switch (eqe_evttype) {
1055
1056 case HERMON_EVT_PORT_STATE_CHANGE:
1057 status = hermon_port_state_change_handler(state, eq, eqe);
1058 break;
1059
1060 case HERMON_EVT_COMM_ESTABLISHED:
1061 status = hermon_comm_estbl_handler(state, eq, eqe);
1062 break;
1063
1064 case HERMON_EVT_COMMAND_INTF_COMP:
1065 status = hermon_cmd_complete_handler(state, eq, eqe);
1066 break;
1067
1068 case HERMON_EVT_LOCAL_WQ_CAT_ERROR:
1069 HERMON_WARNING(state, HERMON_FMA_LOCCAT);
1070 status = hermon_local_wq_cat_err_handler(state, eq, eqe);
1071 break;
1072
1073 case HERMON_EVT_INV_REQ_LOCAL_WQ_ERROR:
1074 HERMON_WARNING(state, HERMON_FMA_LOCINV);
1075 status = hermon_invreq_local_wq_err_handler(state, eq, eqe);
1076 break;
1077
1078 case HERMON_EVT_LOCAL_ACC_VIO_WQ_ERROR:
1079 HERMON_WARNING(state, HERMON_FMA_LOCACEQ);
1080 IBTF_DPRINTF_L2("async", HERMON_FMA_LOCACEQ);
1081 status = hermon_local_acc_vio_wq_err_handler(state, eq, eqe);
1082 break;
1083 case HERMON_EVT_SEND_QUEUE_DRAINED:
1084 status = hermon_sendq_drained_handler(state, eq, eqe);
1085 break;
1086
1087 case HERMON_EVT_PATH_MIGRATED:
1088 status = hermon_path_mig_handler(state, eq, eqe);
1089 break;
1090
1091 case HERMON_EVT_PATH_MIGRATE_FAILED:
1092 HERMON_WARNING(state, HERMON_FMA_PATHMIG);
1093 status = hermon_path_mig_err_handler(state, eq, eqe);
1094 break;
1095
1096 case HERMON_EVT_SRQ_CATASTROPHIC_ERROR:
1097 HERMON_WARNING(state, HERMON_FMA_SRQCAT);
1098 status = hermon_catastrophic_handler(state, eq, eqe);
1099 break;
1100
1101 case HERMON_EVT_SRQ_LAST_WQE_REACHED:
1102 status = hermon_srq_last_wqe_reached_handler(state, eq, eqe);
1103 break;
1104
1105 case HERMON_EVT_FEXCH_ERROR:
1106 status = hermon_fexch_error_handler(state, eq, eqe);
1107 break;
1108
1109 default:
1110 break;
1111 }
1112 return (status);
1113 }
1114
1115 /*
1116 * hermon_port_state_change_handler()
1117 * Context: Only called from interrupt context
1118 */
1119 /* ARGSUSED */
1120 static int
1121 hermon_port_state_change_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1122 hermon_hw_eqe_t *eqe)
1123 {
1124 ibc_async_event_t event;
1125 ibt_async_code_t type;
1126 uint_t subtype;
1127 uint8_t port;
1128 char link_msg[24];
1129
1130 /*
1131 * Depending on the type of Port State Change event, pass the
1132 * appropriate asynch event to the IBTF.
1133 */
1134 port = (uint8_t)HERMON_EQE_PORTNUM_GET(eq, eqe);
1135
1136 /* Check for valid port number in event */
1137 if ((port == 0) || (port > state->hs_cfg_profile->cp_num_ports)) {
1138 HERMON_WARNING(state, "Unexpected port number in port state "
1139 "change event");
1140 cmn_err(CE_CONT, " Port number: %02x\n", port);
1141 return (DDI_FAILURE);
1142 }
1143
1144 subtype = HERMON_EQE_EVTSUBTYPE_GET(eq, eqe);
1145 if (subtype == HERMON_PORT_LINK_ACTIVE) {
1146 event.ev_port = port;
1147 type = IBT_EVENT_PORT_UP;
1148
1149 (void) snprintf(link_msg, 23, "port %d up", port);
1150 ddi_dev_report_fault(state->hs_dip, DDI_SERVICE_RESTORED,
1151 DDI_EXTERNAL_FAULT, link_msg);
1152 } else if (subtype == HERMON_PORT_LINK_DOWN) {
1153 event.ev_port = port;
1154 type = IBT_ERROR_PORT_DOWN;
1155
1156 (void) snprintf(link_msg, 23, "port %d down", port);
1157 ddi_dev_report_fault(state->hs_dip, DDI_SERVICE_LOST,
1158 DDI_EXTERNAL_FAULT, link_msg);
1159 } else {
1160 HERMON_WARNING(state, "Unexpected subtype in port state change "
1161 "event");
1162 cmn_err(CE_CONT, " Event type: %02x, subtype: %02x\n",
1163 HERMON_EQE_EVTTYPE_GET(eq, eqe), subtype);
1164 return (DDI_FAILURE);
1165 }
1166
1167 /*
1168 * Deliver the event to the IBTF. Note: If "hs_ibtfpriv" is NULL,
1169 * then we have either received this event before we finished
1170 * attaching to the IBTF or we've received it while we are in the
1171 * process of detaching.
1172 */
1173 if (state->hs_ibtfpriv != NULL) {
1174 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1175 }
1176
1177 return (DDI_SUCCESS);
1178 }
1179
1180
1181 /*
1182 * hermon_comm_estbl_handler()
1183 * Context: Only called from interrupt context
1184 */
1185 /* ARGSUSED */
1186 static int
1187 hermon_comm_estbl_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1188 hermon_hw_eqe_t *eqe)
1189 {
1190 hermon_qphdl_t qp;
1191 uint_t qpnum;
1192 ibc_async_event_t event;
1193 ibt_async_code_t type;
1194
1195 /* Get the QP handle from QP number in event descriptor */
1196 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1197 qp = hermon_qphdl_from_qpnum(state, qpnum);
1198
1199 /*
1200 * If the QP handle is NULL, this is probably an indication
1201 * that the QP has been freed already. In which case, we
1202 * should not deliver this event.
1203 *
1204 * We also check that the QP number in the handle is the
1205 * same as the QP number in the event queue entry. This
1206 * extra check allows us to handle the case where a QP was
1207 * freed and then allocated again in the time it took to
1208 * handle the event queue processing. By constantly incrementing
1209 * the non-constrained portion of the QP number every time
1210 * a new QP is allocated, we mitigate (somewhat) the chance
1211 * that a stale event could be passed to the client's QP
1212 * handler.
1213 *
1214 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1215 * means that we've have either received this event before we
1216 * finished attaching to the IBTF or we've received it while we
1217 * are in the process of detaching.
1218 */
1219 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1220 (state->hs_ibtfpriv != NULL)) {
1221 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1222 type = IBT_EVENT_COM_EST_QP;
1223
1224 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1225 }
1226
1227 return (DDI_SUCCESS);
1228 }
1229
1230
1231 /*
1232 * hermon_local_wq_cat_err_handler()
1233 * Context: Only called from interrupt context
1234 */
1235 /* ARGSUSED */
1236 static int
1237 hermon_local_wq_cat_err_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1238 hermon_hw_eqe_t *eqe)
1239 {
1240 hermon_qphdl_t qp;
1241 uint_t qpnum;
1242 ibc_async_event_t event;
1243 ibt_async_code_t type;
1244
1245 /* Get the QP handle from QP number in event descriptor */
1246 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1247 qp = hermon_qphdl_from_qpnum(state, qpnum);
1248
1249 /*
1250 * If the QP handle is NULL, this is probably an indication
1251 * that the QP has been freed already. In which case, we
1252 * should not deliver this event.
1253 *
1254 * We also check that the QP number in the handle is the
1255 * same as the QP number in the event queue entry. This
1256 * extra check allows us to handle the case where a QP was
1257 * freed and then allocated again in the time it took to
1258 * handle the event queue processing. By constantly incrementing
1259 * the non-constrained portion of the QP number every time
1260 * a new QP is allocated, we mitigate (somewhat) the chance
1261 * that a stale event could be passed to the client's QP
1262 * handler.
1263 *
1264 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1265 * means that we've have either received this event before we
1266 * finished attaching to the IBTF or we've received it while we
1267 * are in the process of detaching.
1268 */
1269 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1270 (state->hs_ibtfpriv != NULL)) {
1271 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1272 type = IBT_ERROR_CATASTROPHIC_QP;
1273
1274 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1275 }
1276
1277 return (DDI_SUCCESS);
1278 }
1279
1280
1281 /*
1282 * hermon_invreq_local_wq_err_handler()
1283 * Context: Only called from interrupt context
1284 */
1285 /* ARGSUSED */
1286 static int
1287 hermon_invreq_local_wq_err_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1288 hermon_hw_eqe_t *eqe)
1289 {
1290 hermon_qphdl_t qp;
1291 uint_t qpnum;
1292 ibc_async_event_t event;
1293 ibt_async_code_t type;
1294
1295 /* Get the QP handle from QP number in event descriptor */
1296 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1297 qp = hermon_qphdl_from_qpnum(state, qpnum);
1298
1299 /*
1300 * If the QP handle is NULL, this is probably an indication
1301 * that the QP has been freed already. In which case, we
1302 * should not deliver this event.
1303 *
1304 * We also check that the QP number in the handle is the
1305 * same as the QP number in the event queue entry. This
1306 * extra check allows us to handle the case where a QP was
1307 * freed and then allocated again in the time it took to
1308 * handle the event queue processing. By constantly incrementing
1309 * the non-constrained portion of the QP number every time
1310 * a new QP is allocated, we mitigate (somewhat) the chance
1311 * that a stale event could be passed to the client's QP
1312 * handler.
1313 *
1314 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1315 * means that we've have either received this event before we
1316 * finished attaching to the IBTF or we've received it while we
1317 * are in the process of detaching.
1318 */
1319 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1320 (state->hs_ibtfpriv != NULL)) {
1321 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1322 type = IBT_ERROR_INVALID_REQUEST_QP;
1323
1324 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1325 }
1326
1327 return (DDI_SUCCESS);
1328 }
1329
1330
1331 /*
1332 * hermon_local_acc_vio_wq_err_handler()
1333 * Context: Only called from interrupt context
1334 */
1335 /* ARGSUSED */
1336 static int
1337 hermon_local_acc_vio_wq_err_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1338 hermon_hw_eqe_t *eqe)
1339 {
1340 hermon_qphdl_t qp;
1341 uint_t qpnum;
1342 ibc_async_event_t event;
1343 ibt_async_code_t type;
1344
1345 /* Get the QP handle from QP number in event descriptor */
1346 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1347 qp = hermon_qphdl_from_qpnum(state, qpnum);
1348
1349 /*
1350 * If the QP handle is NULL, this is probably an indication
1351 * that the QP has been freed already. In which case, we
1352 * should not deliver this event.
1353 *
1354 * We also check that the QP number in the handle is the
1355 * same as the QP number in the event queue entry. This
1356 * extra check allows us to handle the case where a QP was
1357 * freed and then allocated again in the time it took to
1358 * handle the event queue processing. By constantly incrementing
1359 * the non-constrained portion of the QP number every time
1360 * a new QP is allocated, we mitigate (somewhat) the chance
1361 * that a stale event could be passed to the client's QP
1362 * handler.
1363 *
1364 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1365 * means that we've have either received this event before we
1366 * finished attaching to the IBTF or we've received it while we
1367 * are in the process of detaching.
1368 */
1369 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1370 (state->hs_ibtfpriv != NULL)) {
1371 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1372 type = IBT_ERROR_ACCESS_VIOLATION_QP;
1373
1374 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1375 }
1376
1377 return (DDI_SUCCESS);
1378 }
1379
1380
1381 /*
1382 * hermon_sendq_drained_handler()
1383 * Context: Only called from interrupt context
1384 */
1385 /* ARGSUSED */
1386 static int
1387 hermon_sendq_drained_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1388 hermon_hw_eqe_t *eqe)
1389 {
1390 hermon_qphdl_t qp;
1391 uint_t qpnum;
1392 ibc_async_event_t event;
1393 uint_t forward_sqd_event;
1394 ibt_async_code_t type;
1395
1396 /* Get the QP handle from QP number in event descriptor */
1397 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1398 qp = hermon_qphdl_from_qpnum(state, qpnum);
1399
1400 /*
1401 * If the QP handle is NULL, this is probably an indication
1402 * that the QP has been freed already. In which case, we
1403 * should not deliver this event.
1404 *
1405 * We also check that the QP number in the handle is the
1406 * same as the QP number in the event queue entry. This
1407 * extra check allows us to handle the case where a QP was
1408 * freed and then allocated again in the time it took to
1409 * handle the event queue processing. By constantly incrementing
1410 * the non-constrained portion of the QP number every time
1411 * a new QP is allocated, we mitigate (somewhat) the chance
1412 * that a stale event could be passed to the client's QP
1413 * handler.
1414 *
1415 * And then we check if "hs_ibtfpriv" is NULL. If it is then it
1416 * means that we've have either received this event before we
1417 * finished attaching to the IBTF or we've received it while we
1418 * are in the process of detaching.
1419 */
1420 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1421 (state->hs_ibtfpriv != NULL)) {
1422 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1423 type = IBT_EVENT_SQD;
1424
1425 /*
1426 * Grab the QP lock and update the QP state to reflect that
1427 * the Send Queue Drained event has arrived. Also determine
1428 * whether the event is intended to be forwarded on to the
1429 * consumer or not. This information is used below in
1430 * determining whether or not to call the IBTF.
1431 */
1432 mutex_enter(&qp->qp_lock);
1433 forward_sqd_event = qp->qp_forward_sqd_event;
1434 qp->qp_forward_sqd_event = 0;
1435 qp->qp_sqd_still_draining = 0;
1436 mutex_exit(&qp->qp_lock);
1437
1438 if (forward_sqd_event != 0) {
1439 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1440 }
1441 }
1442
1443 return (DDI_SUCCESS);
1444 }
1445
1446
1447 /*
1448 * hermon_path_mig_handler()
1449 * Context: Only called from interrupt context
1450 */
1451 /* ARGSUSED */
1452 static int
1453 hermon_path_mig_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1454 hermon_hw_eqe_t *eqe)
1455 {
1456 hermon_qphdl_t qp;
1457 uint_t qpnum;
1458 ibc_async_event_t event;
1459 ibt_async_code_t type;
1460
1461 /* Get the QP handle from QP number in event descriptor */
1462 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1463 qp = hermon_qphdl_from_qpnum(state, qpnum);
1464
1465 /*
1466 * If the QP handle is NULL, this is probably an indication
1467 * that the QP has been freed already. In which case, we
1468 * should not deliver this event.
1469 *
1470 * We also check that the QP number in the handle is the
1471 * same as the QP number in the event queue entry. This
1472 * extra check allows us to handle the case where a QP was
1473 * freed and then allocated again in the time it took to
1474 * handle the event queue processing. By constantly incrementing
1475 * the non-constrained portion of the QP number every time
1476 * a new QP is allocated, we mitigate (somewhat) the chance
1477 * that a stale event could be passed to the client's QP
1478 * handler.
1479 *
1480 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1481 * means that we've have either received this event before we
1482 * finished attaching to the IBTF or we've received it while we
1483 * are in the process of detaching.
1484 */
1485 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1486 (state->hs_ibtfpriv != NULL)) {
1487 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1488 type = IBT_EVENT_PATH_MIGRATED_QP;
1489
1490 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1491 }
1492
1493 return (DDI_SUCCESS);
1494 }
1495
1496
1497 /*
1498 * hermon_path_mig_err_handler()
1499 * Context: Only called from interrupt context
1500 */
1501 /* ARGSUSED */
1502 static int
1503 hermon_path_mig_err_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1504 hermon_hw_eqe_t *eqe)
1505 {
1506 hermon_qphdl_t qp;
1507 uint_t qpnum;
1508 ibc_async_event_t event;
1509 ibt_async_code_t type;
1510
1511 /* Get the QP handle from QP number in event descriptor */
1512 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1513 qp = hermon_qphdl_from_qpnum(state, qpnum);
1514
1515 /*
1516 * If the QP handle is NULL, this is probably an indication
1517 * that the QP has been freed already. In which case, we
1518 * should not deliver this event.
1519 *
1520 * We also check that the QP number in the handle is the
1521 * same as the QP number in the event queue entry. This
1522 * extra check allows us to handle the case where a QP was
1523 * freed and then allocated again in the time it took to
1524 * handle the event queue processing. By constantly incrementing
1525 * the non-constrained portion of the QP number every time
1526 * a new QP is allocated, we mitigate (somewhat) the chance
1527 * that a stale event could be passed to the client's QP
1528 * handler.
1529 *
1530 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1531 * means that we've have either received this event before we
1532 * finished attaching to the IBTF or we've received it while we
1533 * are in the process of detaching.
1534 */
1535 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1536 (state->hs_ibtfpriv != NULL)) {
1537 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1538 type = IBT_ERROR_PATH_MIGRATE_REQ_QP;
1539
1540 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1541 }
1542
1543 return (DDI_SUCCESS);
1544 }
1545
1546
1547 /*
1548 * hermon_catastrophic_handler()
1549 * Context: Only called from interrupt context
1550 */
1551 /* ARGSUSED */
1552 static int
1553 hermon_catastrophic_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1554 hermon_hw_eqe_t *eqe)
1555 {
1556 hermon_qphdl_t qp;
1557 uint_t qpnum;
1558 ibc_async_event_t event;
1559 ibt_async_code_t type;
1560
1561 if (eq->eq_evttypemask == HERMON_EVT_MSK_LOCAL_CAT_ERROR) {
1562 HERMON_FMANOTE(state, HERMON_FMA_INTERNAL);
1563 hermon_eq_catastrophic(state);
1564 return (DDI_SUCCESS);
1565 }
1566
1567 /* Get the QP handle from QP number in event descriptor */
1568 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1569 qp = hermon_qphdl_from_qpnum(state, qpnum);
1570
1571 /*
1572 * If the QP handle is NULL, this is probably an indication
1573 * that the QP has been freed already. In which case, we
1574 * should not deliver this event.
1575 *
1576 * We also check that the QP number in the handle is the
1577 * same as the QP number in the event queue entry. This
1578 * extra check allows us to handle the case where a QP was
1579 * freed and then allocated again in the time it took to
1580 * handle the event queue processing. By constantly incrementing
1581 * the non-constrained portion of the QP number every time
1582 * a new QP is allocated, we mitigate (somewhat) the chance
1583 * that a stale event could be passed to the client's QP
1584 * handler.
1585 *
1586 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1587 * means that we've have either received this event before we
1588 * finished attaching to the IBTF or we've received it while we
1589 * are in the process of detaching.
1590 */
1591 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1592 (state->hs_ibtfpriv != NULL)) {
1593 event.ev_srq_hdl = (ibt_srq_hdl_t)qp->qp_srqhdl->srq_hdlrarg;
1594 type = IBT_ERROR_CATASTROPHIC_SRQ;
1595
1596 mutex_enter(&qp->qp_srqhdl->srq_lock);
1597 qp->qp_srqhdl->srq_state = HERMON_SRQ_STATE_ERROR;
1598 mutex_exit(&qp->qp_srqhdl->srq_lock);
1599
1600 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1601 }
1602
1603 return (DDI_SUCCESS);
1604 }
1605
1606
1607 /*
1608 * hermon_srq_last_wqe_reached_handler()
1609 * Context: Only called from interrupt context
1610 */
1611 /* ARGSUSED */
1612 static int
1613 hermon_srq_last_wqe_reached_handler(hermon_state_t *state, hermon_eqhdl_t eq,
1614 hermon_hw_eqe_t *eqe)
1615 {
1616 hermon_qphdl_t qp;
1617 uint_t qpnum;
1618 ibc_async_event_t event;
1619 ibt_async_code_t type;
1620
1621 /* Get the QP handle from QP number in event descriptor */
1622 qpnum = HERMON_EQE_QPNUM_GET(eq, eqe);
1623 qp = hermon_qphdl_from_qpnum(state, qpnum);
1624
1625 /*
1626 * If the QP handle is NULL, this is probably an indication
1627 * that the QP has been freed already. In which case, we
1628 * should not deliver this event.
1629 *
1630 * We also check that the QP number in the handle is the
1631 * same as the QP number in the event queue entry. This
1632 * extra check allows us to handle the case where a QP was
1633 * freed and then allocated again in the time it took to
1634 * handle the event queue processing. By constantly incrementing
1635 * the non-constrained portion of the QP number every time
1636 * a new QP is allocated, we mitigate (somewhat) the chance
1637 * that a stale event could be passed to the client's QP
1638 * handler.
1639 *
1640 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1641 * means that we've have either received this event before we
1642 * finished attaching to the IBTF or we've received it while we
1643 * are in the process of detaching.
1644 */
1645 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1646 (state->hs_ibtfpriv != NULL)) {
1647 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1648 type = IBT_EVENT_EMPTY_CHAN;
1649
1650 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1651 }
1652
1653 return (DDI_SUCCESS);
1654 }
1655
1656
1657 /* ARGSUSED */
1658 static int hermon_fexch_error_handler(hermon_state_t *state,
1659 hermon_eqhdl_t eq, hermon_hw_eqe_t *eqe)
1660 {
1661 hermon_qphdl_t qp;
1662 uint_t qpnum;
1663 ibc_async_event_t event;
1664 ibt_async_code_t type;
1665
1666 /* Get the QP handle from QP number in event descriptor */
1667 event.ev_port = HERMON_EQE_FEXCH_PORTNUM_GET(eq, eqe);
1668 qpnum = hermon_fcoib_qpnum_from_fexch(state,
1669 event.ev_port, HERMON_EQE_FEXCH_FEXCH_GET(eq, eqe));
1670 qp = hermon_qphdl_from_qpnum(state, qpnum);
1671
1672 event.ev_fc = HERMON_EQE_FEXCH_SYNDROME_GET(eq, eqe);
1673
1674 /*
1675 * If the QP handle is NULL, this is probably an indication
1676 * that the QP has been freed already. In which case, we
1677 * should not deliver this event.
1678 *
1679 * We also check that the QP number in the handle is the
1680 * same as the QP number in the event queue entry. This
1681 * extra check allows us to handle the case where a QP was
1682 * freed and then allocated again in the time it took to
1683 * handle the event queue processing. By constantly incrementing
1684 * the non-constrained portion of the QP number every time
1685 * a new QP is allocated, we mitigate (somewhat) the chance
1686 * that a stale event could be passed to the client's QP
1687 * handler.
1688 *
1689 * Lastly, we check if "hs_ibtfpriv" is NULL. If it is then it
1690 * means that we've have either received this event before we
1691 * finished attaching to the IBTF or we've received it while we
1692 * are in the process of detaching.
1693 */
1694 if ((qp != NULL) && (qp->qp_qpnum == qpnum) &&
1695 (state->hs_ibtfpriv != NULL)) {
1696 event.ev_qp_hdl = (ibtl_qp_hdl_t)qp->qp_hdlrarg;
1697 type = IBT_FEXCH_ERROR;
1698
1699 HERMON_DO_IBTF_ASYNC_CALLB(state, type, &event);
1700 }
1701
1702 return (DDI_SUCCESS);
1703 }
1704
1705
1706 /*
1707 * hermon_no_eqhandler
1708 * Context: Only called from interrupt context
1709 */
1710 /* ARGSUSED */
1711 static int
1712 hermon_no_eqhandler(hermon_state_t *state, hermon_eqhdl_t eq,
1713 hermon_hw_eqe_t *eqe)
1714 {
1715 uint_t data;
1716 int i;
1717
1718 /*
1719 * This "unexpected event" handler (or "catch-all" handler) will
1720 * receive all events for which no other handler has been registered.
1721 * If we end up here, then something has probably gone seriously wrong
1722 * with the Hermon hardware (or, perhaps, with the software... though
1723 * it's unlikely in this case). The EQE provides all the information
1724 * about the event. So we print a warning message here along with
1725 * the contents of the EQE.
1726 */
1727 HERMON_WARNING(state, "Unexpected Event handler");
1728 cmn_err(CE_CONT, " Event type: %02x, subtype: %02x\n",
1729 HERMON_EQE_EVTTYPE_GET(eq, eqe),
1730 HERMON_EQE_EVTSUBTYPE_GET(eq, eqe));
1731 for (i = 0; i < sizeof (hermon_hw_eqe_t) >> 2; i++) {
1732 data = ((uint_t *)eqe)[i];
1733 cmn_err(CE_CONT, " EQE[%02x]: %08x\n", i, data);
1734 }
1735
1736 return (DDI_SUCCESS);
1737 }