xref: /freebsd/sys/dev/mpr/mpr.c (revision 907b59d76938e654f0d040a888e8dfca3de1e222)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2016 Avago Technologies
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
29  *
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 /* Communications core for Avago Technologies (LSI) MPT3 */
36 
37 /* TODO Move headers to mprvar */
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/selinfo.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/bus.h>
47 #include <sys/conf.h>
48 #include <sys/bio.h>
49 #include <sys/malloc.h>
50 #include <sys/uio.h>
51 #include <sys/sysctl.h>
52 #include <sys/queue.h>
53 #include <sys/kthread.h>
54 #include <sys/taskqueue.h>
55 #include <sys/endian.h>
56 #include <sys/eventhandler.h>
57 
58 #include <machine/bus.h>
59 #include <machine/resource.h>
60 #include <sys/rman.h>
61 #include <sys/proc.h>
62 
63 #include <dev/pci/pcivar.h>
64 
65 #include <cam/cam.h>
66 #include <cam/scsi/scsi_all.h>
67 
68 #include <dev/mpr/mpi/mpi2_type.h>
69 #include <dev/mpr/mpi/mpi2.h>
70 #include <dev/mpr/mpi/mpi2_ioc.h>
71 #include <dev/mpr/mpi/mpi2_sas.h>
72 #include <dev/mpr/mpi/mpi2_cnfg.h>
73 #include <dev/mpr/mpi/mpi2_init.h>
74 #include <dev/mpr/mpi/mpi2_tool.h>
75 #include <dev/mpr/mpr_ioctl.h>
76 #include <dev/mpr/mprvar.h>
77 #include <dev/mpr/mpr_table.h>
78 
79 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
80 static int mpr_init_queues(struct mpr_softc *sc);
81 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
82 static int mpr_transition_operational(struct mpr_softc *sc);
83 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
84 static void mpr_iocfacts_free(struct mpr_softc *sc);
85 static void mpr_startup(void *arg);
86 static int mpr_send_iocinit(struct mpr_softc *sc);
87 static int mpr_alloc_queues(struct mpr_softc *sc);
88 static int mpr_alloc_replies(struct mpr_softc *sc);
89 static int mpr_alloc_requests(struct mpr_softc *sc);
90 static int mpr_attach_log(struct mpr_softc *sc);
91 static __inline void mpr_complete_command(struct mpr_softc *sc,
92     struct mpr_command *cm);
93 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
94     MPI2_EVENT_NOTIFICATION_REPLY *reply);
95 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
96 static void mpr_periodic(void *);
97 static int mpr_reregister_events(struct mpr_softc *sc);
98 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
99 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
100 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
101 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD, 0, "MPR Driver Parameters");
102 
103 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory");
104 
105 /*
106  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
107  * any state and back to its initialization state machine.
108  */
109 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
110 
111 /*
112  * Added this union to smoothly convert le64toh cm->cm_desc.Words.
113  * Compiler only supports unint64_t to be passed as an argument.
114  * Otherwise it will through this error:
115  * "aggregate value used where an integer was expected"
116  */
117 typedef union _reply_descriptor {
118         u64 word;
119         struct {
120                 u32 low;
121                 u32 high;
122         } u;
123 }reply_descriptor,address_descriptor;
124 
125 /* Rate limit chain-fail messages to 1 per minute */
126 static struct timeval mpr_chainfail_interval = { 60, 0 };
127 
128 /*
129  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
130  * If this function is called from process context, it can sleep
131  * and there is no harm to sleep, in case if this fuction is called
132  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
133  * based on sleep flags driver will call either msleep, pause or DELAY.
134  * msleep and pause are of same variant, but pause is used when mpr_mtx
135  * is not hold by driver.
136  */
137 static int
138 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag)
139 {
140 	uint32_t reg;
141 	int i, error, tries = 0;
142 	uint8_t first_wait_done = FALSE;
143 
144 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
145 
146 	/* Clear any pending interrupts */
147 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
148 
149 	/*
150 	 * Force NO_SLEEP for threads prohibited to sleep
151  	 * e.a Thread from interrupt handler are prohibited to sleep.
152  	 */
153 #if __FreeBSD_version >= 1000029
154 	if (curthread->td_no_sleeping)
155 #else //__FreeBSD_version < 1000029
156 	if (curthread->td_pflags & TDP_NOSLEEPING)
157 #endif //__FreeBSD_version >= 1000029
158 		sleep_flag = NO_SLEEP;
159 
160 	/* Push the magic sequence */
161 	error = ETIMEDOUT;
162 	while (tries++ < 20) {
163 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
164 			mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
165 			    mpt2_reset_magic[i]);
166 
167 		/* wait 100 msec */
168 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
169 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
170 			    "mprdiag", hz/10);
171 		else if (sleep_flag == CAN_SLEEP)
172 			pause("mprdiag", hz/10);
173 		else
174 			DELAY(100 * 1000);
175 
176 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
177 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
178 			error = 0;
179 			break;
180 		}
181 	}
182 	if (error)
183 		return (error);
184 
185 	/* Send the actual reset.  XXX need to refresh the reg? */
186 	mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
187 	    reg | MPI2_DIAG_RESET_ADAPTER);
188 
189 	/* Wait up to 300 seconds in 50ms intervals */
190 	error = ETIMEDOUT;
191 	for (i = 0; i < 6000; i++) {
192 		/*
193 		 * Wait 50 msec. If this is the first time through, wait 256
194 		 * msec to satisfy Diag Reset timing requirements.
195 		 */
196 		if (first_wait_done) {
197 			if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
198 				msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
199 				    "mprdiag", hz/20);
200 			else if (sleep_flag == CAN_SLEEP)
201 				pause("mprdiag", hz/20);
202 			else
203 				DELAY(50 * 1000);
204 		} else {
205 			DELAY(256 * 1000);
206 			first_wait_done = TRUE;
207 		}
208 		/*
209 		 * Check for the RESET_ADAPTER bit to be cleared first, then
210 		 * wait for the RESET state to be cleared, which takes a little
211 		 * longer.
212 		 */
213 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
214 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
215 			continue;
216 		}
217 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
218 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
219 			error = 0;
220 			break;
221 		}
222 	}
223 	if (error)
224 		return (error);
225 
226 	mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
227 
228 	return (0);
229 }
230 
231 static int
232 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag)
233 {
234 
235 	MPR_FUNCTRACE(sc);
236 
237 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
238 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
239 	    MPI2_DOORBELL_FUNCTION_SHIFT);
240 
241 	if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) {
242 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed : <%s>\n",
243 				__func__);
244 		return (ETIMEDOUT);
245 	}
246 
247 	return (0);
248 }
249 
250 static int
251 mpr_transition_ready(struct mpr_softc *sc)
252 {
253 	uint32_t reg, state;
254 	int error, tries = 0;
255 	int sleep_flags;
256 
257 	MPR_FUNCTRACE(sc);
258 	/* If we are in attach call, do not sleep */
259 	sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE)
260 	    ? CAN_SLEEP : NO_SLEEP;
261 
262 	error = 0;
263 	while (tries++ < 1200) {
264 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
265 		mpr_dprint(sc, MPR_INIT, "Doorbell= 0x%x\n", reg);
266 
267 		/*
268 		 * Ensure the IOC is ready to talk.  If it's not, try
269 		 * resetting it.
270 		 */
271 		if (reg & MPI2_DOORBELL_USED) {
272 			mpr_diag_reset(sc, sleep_flags);
273 			DELAY(50000);
274 			continue;
275 		}
276 
277 		/* Is the adapter owned by another peer? */
278 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
279 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
280 			device_printf(sc->mpr_dev, "IOC is under the control "
281 			    "of another peer host, aborting initialization.\n");
282 			return (ENXIO);
283 		}
284 
285 		state = reg & MPI2_IOC_STATE_MASK;
286 		if (state == MPI2_IOC_STATE_READY) {
287 			/* Ready to go! */
288 			error = 0;
289 			break;
290 		} else if (state == MPI2_IOC_STATE_FAULT) {
291 			mpr_dprint(sc, MPR_FAULT, "IOC in fault state 0x%x\n",
292 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
293 			mpr_diag_reset(sc, sleep_flags);
294 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
295 			/* Need to take ownership */
296 			mpr_message_unit_reset(sc, sleep_flags);
297 		} else if (state == MPI2_IOC_STATE_RESET) {
298 			/* Wait a bit, IOC might be in transition */
299 			mpr_dprint(sc, MPR_FAULT,
300 			    "IOC in unexpected reset state\n");
301 		} else {
302 			mpr_dprint(sc, MPR_FAULT,
303 			    "IOC in unknown state 0x%x\n", state);
304 			error = EINVAL;
305 			break;
306 		}
307 
308 		/* Wait 50ms for things to settle down. */
309 		DELAY(50000);
310 	}
311 
312 	if (error)
313 		device_printf(sc->mpr_dev, "Cannot transition IOC to ready\n");
314 
315 	return (error);
316 }
317 
318 static int
319 mpr_transition_operational(struct mpr_softc *sc)
320 {
321 	uint32_t reg, state;
322 	int error;
323 
324 	MPR_FUNCTRACE(sc);
325 
326 	error = 0;
327 	reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
328 	mpr_dprint(sc, MPR_INIT, "Doorbell= 0x%x\n", reg);
329 
330 	state = reg & MPI2_IOC_STATE_MASK;
331 	if (state != MPI2_IOC_STATE_READY) {
332 		if ((error = mpr_transition_ready(sc)) != 0) {
333 			mpr_dprint(sc, MPR_FAULT,
334 			    "%s failed to transition ready\n", __func__);
335 			return (error);
336 		}
337 	}
338 
339 	error = mpr_send_iocinit(sc);
340 	return (error);
341 }
342 
343 /*
344  * This is called during attach and when re-initializing due to a Diag Reset.
345  * IOC Facts is used to allocate many of the structures needed by the driver.
346  * If called from attach, de-allocation is not required because the driver has
347  * not allocated any structures yet, but if called from a Diag Reset, previously
348  * allocated structures based on IOC Facts will need to be freed and re-
349  * allocated bases on the latest IOC Facts.
350  */
351 static int
352 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
353 {
354 	int error;
355 	Mpi2IOCFactsReply_t saved_facts;
356 	uint8_t saved_mode, reallocating;
357 
358 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
359 
360 	/* Save old IOC Facts and then only reallocate if Facts have changed */
361 	if (!attaching) {
362 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
363 	}
364 
365 	/*
366 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
367 	 * a re-initialization and only return the error if attaching so the OS
368 	 * can handle it.
369 	 */
370 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
371 		if (attaching) {
372 			mpr_dprint(sc, MPR_FAULT, "%s failed to get IOC Facts "
373 			    "with error %d\n", __func__, error);
374 			return (error);
375 		} else {
376 			panic("%s failed to get IOC Facts with error %d\n",
377 			    __func__, error);
378 		}
379 	}
380 
381 	mpr_print_iocfacts(sc, sc->facts);
382 
383 	snprintf(sc->fw_version, sizeof(sc->fw_version),
384 	    "%02d.%02d.%02d.%02d",
385 	    sc->facts->FWVersion.Struct.Major,
386 	    sc->facts->FWVersion.Struct.Minor,
387 	    sc->facts->FWVersion.Struct.Unit,
388 	    sc->facts->FWVersion.Struct.Dev);
389 
390 	mpr_printf(sc, "Firmware: %s, Driver: %s\n", sc->fw_version,
391 	    MPR_DRIVER_VERSION);
392 	mpr_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
393 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
394 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
395 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
396 
397 	/*
398 	 * If the chip doesn't support event replay then a hard reset will be
399 	 * required to trigger a full discovery.  Do the reset here then
400 	 * retransition to Ready.  A hard reset might have already been done,
401 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
402 	 * for a Diag Reset.
403 	 */
404 	if (attaching) {
405 		if ((sc->facts->IOCCapabilities &
406 		    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
407 			mpr_diag_reset(sc, NO_SLEEP);
408 			if ((error = mpr_transition_ready(sc)) != 0) {
409 				mpr_dprint(sc, MPR_FAULT, "%s failed to "
410 				    "transition to ready with error %d\n",
411 				    __func__, error);
412 				return (error);
413 			}
414 		}
415 	}
416 
417 	/*
418 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
419 	 * changed from the previous IOC Facts, log a warning, but only if
420 	 * checking this after a Diag Reset and not during attach.
421 	 */
422 	saved_mode = sc->ir_firmware;
423 	if (sc->facts->IOCCapabilities &
424 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
425 		sc->ir_firmware = 1;
426 	if (!attaching) {
427 		if (sc->ir_firmware != saved_mode) {
428 			mpr_dprint(sc, MPR_FAULT, "%s new IR/IT mode in IOC "
429 			    "Facts does not match previous mode\n", __func__);
430 		}
431 	}
432 
433 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
434 	reallocating = FALSE;
435 	if ((!attaching) &&
436 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
437 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
438 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
439 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
440 	    (saved_facts.ProductID != sc->facts->ProductID) ||
441 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
442 	    (saved_facts.IOCRequestFrameSize !=
443 	    sc->facts->IOCRequestFrameSize) ||
444 	    (saved_facts.IOCMaxChainSegmentSize !=
445 	    sc->facts->IOCMaxChainSegmentSize) ||
446 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
447 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
448 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
449 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
450 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
451 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
452 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
453 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
454 	    (saved_facts.MaxPersistentEntries !=
455 	    sc->facts->MaxPersistentEntries))) {
456 		reallocating = TRUE;
457 	}
458 
459 	/*
460 	 * Some things should be done if attaching or re-allocating after a Diag
461 	 * Reset, but are not needed after a Diag Reset if the FW has not
462 	 * changed.
463 	 */
464 	if (attaching || reallocating) {
465 		/*
466 		 * Check if controller supports FW diag buffers and set flag to
467 		 * enable each type.
468 		 */
469 		if (sc->facts->IOCCapabilities &
470 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
471 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
472 			    enabled = TRUE;
473 		if (sc->facts->IOCCapabilities &
474 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
475 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
476 			    enabled = TRUE;
477 		if (sc->facts->IOCCapabilities &
478 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
479 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
480 			    enabled = TRUE;
481 
482 		/*
483 		 * Set flag if EEDP is supported and if TLR is supported.
484 		 */
485 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
486 			sc->eedp_enabled = TRUE;
487 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
488 			sc->control_TLR = TRUE;
489 
490 		/*
491 		 * Size the queues. Since the reply queues always need one free
492 		 * entry, we'll just deduct one reply message here.
493 		 */
494 		sc->num_reqs = MIN(MPR_REQ_FRAMES, sc->facts->RequestCredit);
495 		sc->num_replies = MIN(MPR_REPLY_FRAMES + MPR_EVT_REPLY_FRAMES,
496 		    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
497 
498 		/*
499 		 * Initialize all Tail Queues
500 		 */
501 		TAILQ_INIT(&sc->req_list);
502 		TAILQ_INIT(&sc->high_priority_req_list);
503 		TAILQ_INIT(&sc->chain_list);
504 		TAILQ_INIT(&sc->tm_list);
505 	}
506 
507 	/*
508 	 * If doing a Diag Reset and the FW is significantly different
509 	 * (reallocating will be set above in IOC Facts comparison), then all
510 	 * buffers based on the IOC Facts will need to be freed before they are
511 	 * reallocated.
512 	 */
513 	if (reallocating) {
514 		mpr_iocfacts_free(sc);
515 		mprsas_realloc_targets(sc, saved_facts.MaxTargets);
516 	}
517 
518 	/*
519 	 * Any deallocation has been completed.  Now start reallocating
520 	 * if needed.  Will only need to reallocate if attaching or if the new
521 	 * IOC Facts are different from the previous IOC Facts after a Diag
522 	 * Reset. Targets have already been allocated above if needed.
523 	 */
524 	if (attaching || reallocating) {
525 		if (((error = mpr_alloc_queues(sc)) != 0) ||
526 		    ((error = mpr_alloc_replies(sc)) != 0) ||
527 		    ((error = mpr_alloc_requests(sc)) != 0)) {
528 			if (attaching ) {
529 				mpr_dprint(sc, MPR_FAULT, "%s failed to alloc "
530 				    "queues with error %d\n", __func__, error);
531 				mpr_free(sc);
532 				return (error);
533 			} else {
534 				panic("%s failed to alloc queues with error "
535 				    "%d\n", __func__, error);
536 			}
537 		}
538 	}
539 
540 	/* Always initialize the queues */
541 	bzero(sc->free_queue, sc->fqdepth * 4);
542 	mpr_init_queues(sc);
543 
544 	/*
545 	 * Always get the chip out of the reset state, but only panic if not
546 	 * attaching.  If attaching and there is an error, that is handled by
547 	 * the OS.
548 	 */
549 	error = mpr_transition_operational(sc);
550 	if (error != 0) {
551 		if (attaching) {
552 			mpr_printf(sc, "%s failed to transition to operational "
553 			    "with error %d\n", __func__, error);
554 			mpr_free(sc);
555 			return (error);
556 		} else {
557 			panic("%s failed to transition to operational with "
558 			    "error %d\n", __func__, error);
559 		}
560 	}
561 
562 	/*
563 	 * Finish the queue initialization.
564 	 * These are set here instead of in mpr_init_queues() because the
565 	 * IOC resets these values during the state transition in
566 	 * mpr_transition_operational().  The free index is set to 1
567 	 * because the corresponding index in the IOC is set to 0, and the
568 	 * IOC treats the queues as full if both are set to the same value.
569 	 * Hence the reason that the queue can't hold all of the possible
570 	 * replies.
571 	 */
572 	sc->replypostindex = 0;
573 	mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
574 	mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
575 
576 	/*
577 	 * Attach the subsystems so they can prepare their event masks.
578 	 */
579 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
580 	if (attaching) {
581 		if (((error = mpr_attach_log(sc)) != 0) ||
582 		    ((error = mpr_attach_sas(sc)) != 0) ||
583 		    ((error = mpr_attach_user(sc)) != 0)) {
584 			mpr_printf(sc, "%s failed to attach all subsystems: "
585 			    "error %d\n", __func__, error);
586 			mpr_free(sc);
587 			return (error);
588 		}
589 
590 		if ((error = mpr_pci_setup_interrupts(sc)) != 0) {
591 			mpr_printf(sc, "%s failed to setup interrupts\n",
592 			    __func__);
593 			mpr_free(sc);
594 			return (error);
595 		}
596 	}
597 
598 	return (error);
599 }
600 
601 /*
602  * This is called if memory is being free (during detach for example) and when
603  * buffers need to be reallocated due to a Diag Reset.
604  */
605 static void
606 mpr_iocfacts_free(struct mpr_softc *sc)
607 {
608 	struct mpr_command *cm;
609 	int i;
610 
611 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
612 
613 	if (sc->free_busaddr != 0)
614 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
615 	if (sc->free_queue != NULL)
616 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
617 		    sc->queues_map);
618 	if (sc->queues_dmat != NULL)
619 		bus_dma_tag_destroy(sc->queues_dmat);
620 
621 	if (sc->chain_busaddr != 0)
622 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
623 	if (sc->chain_frames != NULL)
624 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
625 		    sc->chain_map);
626 	if (sc->chain_dmat != NULL)
627 		bus_dma_tag_destroy(sc->chain_dmat);
628 
629 	if (sc->sense_busaddr != 0)
630 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
631 	if (sc->sense_frames != NULL)
632 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
633 		    sc->sense_map);
634 	if (sc->sense_dmat != NULL)
635 		bus_dma_tag_destroy(sc->sense_dmat);
636 
637 	if (sc->reply_busaddr != 0)
638 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
639 	if (sc->reply_frames != NULL)
640 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
641 		    sc->reply_map);
642 	if (sc->reply_dmat != NULL)
643 		bus_dma_tag_destroy(sc->reply_dmat);
644 
645 	if (sc->req_busaddr != 0)
646 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
647 	if (sc->req_frames != NULL)
648 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
649 	if (sc->req_dmat != NULL)
650 		bus_dma_tag_destroy(sc->req_dmat);
651 
652 	if (sc->chains != NULL)
653 		free(sc->chains, M_MPR);
654 	if (sc->commands != NULL) {
655 		for (i = 1; i < sc->num_reqs; i++) {
656 			cm = &sc->commands[i];
657 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
658 		}
659 		free(sc->commands, M_MPR);
660 	}
661 	if (sc->buffer_dmat != NULL)
662 		bus_dma_tag_destroy(sc->buffer_dmat);
663 }
664 
665 /*
666  * The terms diag reset and hard reset are used interchangeably in the MPI
667  * docs to mean resetting the controller chip.  In this code diag reset
668  * cleans everything up, and the hard reset function just sends the reset
669  * sequence to the chip.  This should probably be refactored so that every
670  * subsystem gets a reset notification of some sort, and can clean up
671  * appropriately.
672  */
673 int
674 mpr_reinit(struct mpr_softc *sc)
675 {
676 	int error;
677 	struct mprsas_softc *sassc;
678 
679 	sassc = sc->sassc;
680 
681 	MPR_FUNCTRACE(sc);
682 
683 	mtx_assert(&sc->mpr_mtx, MA_OWNED);
684 
685 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) {
686 		mpr_dprint(sc, MPR_INIT, "%s reset already in progress\n",
687 		    __func__);
688 		return 0;
689 	}
690 
691 	mpr_dprint(sc, MPR_INFO, "Reinitializing controller,\n");
692 	/* make sure the completion callbacks can recognize they're getting
693 	 * a NULL cm_reply due to a reset.
694 	 */
695 	sc->mpr_flags |= MPR_FLAGS_DIAGRESET;
696 
697 	/*
698 	 * Mask interrupts here.
699 	 */
700 	mpr_dprint(sc, MPR_INIT, "%s mask interrupts\n", __func__);
701 	mpr_mask_intr(sc);
702 
703 	error = mpr_diag_reset(sc, CAN_SLEEP);
704 	if (error != 0) {
705 		panic("%s hard reset failed with error %d\n", __func__, error);
706 	}
707 
708 	/* Restore the PCI state, including the MSI-X registers */
709 	mpr_pci_restore(sc);
710 
711 	/* Give the I/O subsystem special priority to get itself prepared */
712 	mprsas_handle_reinit(sc);
713 
714 	/*
715 	 * Get IOC Facts and allocate all structures based on this information.
716 	 * The attach function will also call mpr_iocfacts_allocate at startup.
717 	 * If relevant values have changed in IOC Facts, this function will free
718 	 * all of the memory based on IOC Facts and reallocate that memory.
719 	 */
720 	if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) {
721 		panic("%s IOC Facts based allocation failed with error %d\n",
722 		    __func__, error);
723 	}
724 
725 	/*
726 	 * Mapping structures will be re-allocated after getting IOC Page8, so
727 	 * free these structures here.
728 	 */
729 	mpr_mapping_exit(sc);
730 
731 	/*
732 	 * The static page function currently read is IOC Page8.  Others can be
733 	 * added in future.  It's possible that the values in IOC Page8 have
734 	 * changed after a Diag Reset due to user modification, so always read
735 	 * these.  Interrupts are masked, so unmask them before getting config
736 	 * pages.
737 	 */
738 	mpr_unmask_intr(sc);
739 	sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET;
740 	mpr_base_static_config_pages(sc);
741 
742 	/*
743 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
744 	 * mapping tables.
745 	 */
746 	mpr_mapping_initialize(sc);
747 
748 	/*
749 	 * Restart will reload the event masks clobbered by the reset, and
750 	 * then enable the port.
751 	 */
752 	mpr_reregister_events(sc);
753 
754 	/* the end of discovery will release the simq, so we're done. */
755 	mpr_dprint(sc, MPR_INFO, "%s finished sc %p post %u free %u\n",
756 	    __func__, sc, sc->replypostindex, sc->replyfreeindex);
757 	mprsas_release_simq_reinit(sassc);
758 
759 	return 0;
760 }
761 
762 /* Wait for the chip to ACK a word that we've put into its FIFO
763  * Wait for <timeout> seconds. In single loop wait for busy loop
764  * for 500 microseconds.
765  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
766  * */
767 static int
768 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag)
769 {
770 	u32 cntdn, count;
771 	u32 int_status;
772 	u32 doorbell;
773 
774 	count = 0;
775 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
776 	do {
777 		int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
778 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
779 			mpr_dprint(sc, MPR_INIT, "%s: successful count(%d), "
780 			    "timeout(%d)\n", __func__, count, timeout);
781 			return 0;
782 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
783 			doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
784 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
785 			    MPI2_IOC_STATE_FAULT) {
786 				mpr_dprint(sc, MPR_FAULT,
787 				    "fault_state(0x%04x)!\n", doorbell);
788 				return (EFAULT);
789 			}
790 		} else if (int_status == 0xFFFFFFFF)
791 			goto out;
792 
793 		/*
794 		 * If it can sleep, sleep for 1 milisecond, else busy loop for
795  		 * 0.5 milisecond
796 		 */
797 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
798 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba",
799 			    hz/1000);
800 		else if (sleep_flag == CAN_SLEEP)
801 			pause("mprdba", hz/1000);
802 		else
803 			DELAY(500);
804 		count++;
805 	} while (--cntdn);
806 
807 	out:
808 	mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), "
809 		"int_status(%x)!\n", __func__, count, int_status);
810 	return (ETIMEDOUT);
811 }
812 
813 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
814 static int
815 mpr_wait_db_int(struct mpr_softc *sc)
816 {
817 	int retry;
818 
819 	for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) {
820 		if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
821 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
822 			return (0);
823 		DELAY(2000);
824 	}
825 	return (ETIMEDOUT);
826 }
827 
828 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
829 static int
830 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
831     int req_sz, int reply_sz, int timeout)
832 {
833 	uint32_t *data32;
834 	uint16_t *data16;
835 	int i, count, ioc_sz, residual;
836 	int sleep_flags = CAN_SLEEP;
837 
838 #if __FreeBSD_version >= 1000029
839 	if (curthread->td_no_sleeping)
840 #else //__FreeBSD_version < 1000029
841 	if (curthread->td_pflags & TDP_NOSLEEPING)
842 #endif //__FreeBSD_version >= 1000029
843 		sleep_flags = NO_SLEEP;
844 
845 	/* Step 1 */
846 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
847 
848 	/* Step 2 */
849 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
850 		return (EBUSY);
851 
852 	/* Step 3
853 	 * Announce that a message is coming through the doorbell.  Messages
854 	 * are pushed at 32bit words, so round up if needed.
855 	 */
856 	count = (req_sz + 3) / 4;
857 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
858 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
859 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
860 
861 	/* Step 4 */
862 	if (mpr_wait_db_int(sc) ||
863 	    (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
864 		mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n");
865 		return (ENXIO);
866 	}
867 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
868 	if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
869 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n");
870 		return (ENXIO);
871 	}
872 
873 	/* Step 5 */
874 	/* Clock out the message data synchronously in 32-bit dwords*/
875 	data32 = (uint32_t *)req;
876 	for (i = 0; i < count; i++) {
877 		mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
878 		if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
879 			mpr_dprint(sc, MPR_FAULT,
880 			    "Timeout while writing doorbell\n");
881 			return (ENXIO);
882 		}
883 	}
884 
885 	/* Step 6 */
886 	/* Clock in the reply in 16-bit words.  The total length of the
887 	 * message is always in the 4th byte, so clock out the first 2 words
888 	 * manually, then loop the rest.
889 	 */
890 	data16 = (uint16_t *)reply;
891 	if (mpr_wait_db_int(sc) != 0) {
892 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n");
893 		return (ENXIO);
894 	}
895 	data16[0] =
896 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
897 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
898 	if (mpr_wait_db_int(sc) != 0) {
899 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n");
900 		return (ENXIO);
901 	}
902 	data16[1] =
903 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
904 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
905 
906 	/* Number of 32bit words in the message */
907 	ioc_sz = reply->MsgLength;
908 
909 	/*
910 	 * Figure out how many 16bit words to clock in without overrunning.
911 	 * The precision loss with dividing reply_sz can safely be
912 	 * ignored because the messages can only be multiples of 32bits.
913 	 */
914 	residual = 0;
915 	count = MIN((reply_sz / 4), ioc_sz) * 2;
916 	if (count < ioc_sz * 2) {
917 		residual = ioc_sz * 2 - count;
918 		mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d "
919 		    "residual message words\n", residual);
920 	}
921 
922 	for (i = 2; i < count; i++) {
923 		if (mpr_wait_db_int(sc) != 0) {
924 			mpr_dprint(sc, MPR_FAULT,
925 			    "Timeout reading doorbell %d\n", i);
926 			return (ENXIO);
927 		}
928 		data16[i] = mpr_regread(sc, MPI2_DOORBELL_OFFSET) &
929 		    MPI2_DOORBELL_DATA_MASK;
930 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
931 	}
932 
933 	/*
934 	 * Pull out residual words that won't fit into the provided buffer.
935 	 * This keeps the chip from hanging due to a driver programming
936 	 * error.
937 	 */
938 	while (residual--) {
939 		if (mpr_wait_db_int(sc) != 0) {
940 			mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n");
941 			return (ENXIO);
942 		}
943 		(void)mpr_regread(sc, MPI2_DOORBELL_OFFSET);
944 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
945 	}
946 
947 	/* Step 7 */
948 	if (mpr_wait_db_int(sc) != 0) {
949 		mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n");
950 		return (ENXIO);
951 	}
952 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
953 		mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n");
954 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
955 
956 	return (0);
957 }
958 
959 static void
960 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm)
961 {
962 	reply_descriptor rd;
963 
964 	MPR_FUNCTRACE(sc);
965 	mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n",
966 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
967 
968 	if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags &
969 	    MPR_FLAGS_SHUTDOWN))
970 		mtx_assert(&sc->mpr_mtx, MA_OWNED);
971 
972 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
973 		sc->io_cmds_highwater++;
974 
975 	rd.u.low = cm->cm_desc.Words.Low;
976 	rd.u.high = cm->cm_desc.Words.High;
977 	rd.word = htole64(rd.word);
978 	/* TODO-We may need to make below regwrite atomic */
979 	mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
980 	    rd.u.low);
981 	mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
982 	    rd.u.high);
983 }
984 
985 /*
986  * Just the FACTS, ma'am.
987  */
988 static int
989 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
990 {
991 	MPI2_DEFAULT_REPLY *reply;
992 	MPI2_IOC_FACTS_REQUEST request;
993 	int error, req_sz, reply_sz;
994 
995 	MPR_FUNCTRACE(sc);
996 
997 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
998 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
999 	reply = (MPI2_DEFAULT_REPLY *)facts;
1000 
1001 	bzero(&request, req_sz);
1002 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1003 	error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1004 
1005 	return (error);
1006 }
1007 
1008 static int
1009 mpr_send_iocinit(struct mpr_softc *sc)
1010 {
1011 	MPI2_IOC_INIT_REQUEST	init;
1012 	MPI2_DEFAULT_REPLY	reply;
1013 	int req_sz, reply_sz, error;
1014 	struct timeval now;
1015 	uint64_t time_in_msec;
1016 
1017 	MPR_FUNCTRACE(sc);
1018 
1019 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1020 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1021 	bzero(&init, req_sz);
1022 	bzero(&reply, reply_sz);
1023 
1024 	/*
1025 	 * Fill in the init block.  Note that most addresses are
1026 	 * deliberately in the lower 32bits of memory.  This is a micro-
1027 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1028 	 */
1029 	init.Function = MPI2_FUNCTION_IOC_INIT;
1030 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1031 	init.MsgVersion = htole16(MPI2_VERSION);
1032 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1033 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1034 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1035 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1036 	init.SenseBufferAddressHigh = 0;
1037 	init.SystemReplyAddressHigh = 0;
1038 	init.SystemRequestFrameBaseAddress.High = 0;
1039 	init.SystemRequestFrameBaseAddress.Low =
1040 	    htole32((uint32_t)sc->req_busaddr);
1041 	init.ReplyDescriptorPostQueueAddress.High = 0;
1042 	init.ReplyDescriptorPostQueueAddress.Low =
1043 	    htole32((uint32_t)sc->post_busaddr);
1044 	init.ReplyFreeQueueAddress.High = 0;
1045 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1046 	getmicrotime(&now);
1047 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1048 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1049 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1050 
1051 	error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1052 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1053 		error = ENXIO;
1054 
1055 	mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1056 	return (error);
1057 }
1058 
1059 void
1060 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1061 {
1062 	bus_addr_t *addr;
1063 
1064 	addr = arg;
1065 	*addr = segs[0].ds_addr;
1066 }
1067 
1068 static int
1069 mpr_alloc_queues(struct mpr_softc *sc)
1070 {
1071 	bus_addr_t queues_busaddr;
1072 	uint8_t *queues;
1073 	int qsize, fqsize, pqsize;
1074 
1075 	/*
1076 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1077 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1078 	 * This queue supplies fresh reply frames for the firmware to use.
1079 	 *
1080 	 * The reply descriptor post queue contains 8 byte entries in
1081 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1082 	 * contains filled-in reply frames sent from the firmware to the host.
1083 	 *
1084 	 * These two queues are allocated together for simplicity.
1085 	 */
1086 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1087 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1088 	fqsize= sc->fqdepth * 4;
1089 	pqsize = sc->pqdepth * 8;
1090 	qsize = fqsize + pqsize;
1091 
1092         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1093 				16, 0,			/* algnmnt, boundary */
1094 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1095 				BUS_SPACE_MAXADDR,	/* highaddr */
1096 				NULL, NULL,		/* filter, filterarg */
1097                                 qsize,			/* maxsize */
1098                                 1,			/* nsegments */
1099                                 qsize,			/* maxsegsize */
1100                                 0,			/* flags */
1101                                 NULL, NULL,		/* lockfunc, lockarg */
1102                                 &sc->queues_dmat)) {
1103 		device_printf(sc->mpr_dev, "Cannot allocate queues DMA tag\n");
1104 		return (ENOMEM);
1105         }
1106         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1107 	    &sc->queues_map)) {
1108 		device_printf(sc->mpr_dev, "Cannot allocate queues memory\n");
1109 		return (ENOMEM);
1110         }
1111         bzero(queues, qsize);
1112         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1113 	    mpr_memaddr_cb, &queues_busaddr, 0);
1114 
1115 	sc->free_queue = (uint32_t *)queues;
1116 	sc->free_busaddr = queues_busaddr;
1117 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1118 	sc->post_busaddr = queues_busaddr + fqsize;
1119 
1120 	return (0);
1121 }
1122 
1123 static int
1124 mpr_alloc_replies(struct mpr_softc *sc)
1125 {
1126 	int rsize, num_replies;
1127 
1128 	/*
1129 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1130 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1131 	 * replies can be used at once.
1132 	 */
1133 	num_replies = max(sc->fqdepth, sc->num_replies);
1134 
1135 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1136         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1137 				4, 0,			/* algnmnt, boundary */
1138 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1139 				BUS_SPACE_MAXADDR,	/* highaddr */
1140 				NULL, NULL,		/* filter, filterarg */
1141                                 rsize,			/* maxsize */
1142                                 1,			/* nsegments */
1143                                 rsize,			/* maxsegsize */
1144                                 0,			/* flags */
1145                                 NULL, NULL,		/* lockfunc, lockarg */
1146                                 &sc->reply_dmat)) {
1147 		device_printf(sc->mpr_dev, "Cannot allocate replies DMA tag\n");
1148 		return (ENOMEM);
1149         }
1150         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1151 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1152 		device_printf(sc->mpr_dev, "Cannot allocate replies memory\n");
1153 		return (ENOMEM);
1154         }
1155         bzero(sc->reply_frames, rsize);
1156         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1157 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1158 
1159 	return (0);
1160 }
1161 
1162 static int
1163 mpr_alloc_requests(struct mpr_softc *sc)
1164 {
1165 	struct mpr_command *cm;
1166 	struct mpr_chain *chain;
1167 	int i, rsize, nsegs;
1168 
1169 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1170         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1171 				16, 0,			/* algnmnt, boundary */
1172 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1173 				BUS_SPACE_MAXADDR,	/* highaddr */
1174 				NULL, NULL,		/* filter, filterarg */
1175                                 rsize,			/* maxsize */
1176                                 1,			/* nsegments */
1177                                 rsize,			/* maxsegsize */
1178                                 0,			/* flags */
1179                                 NULL, NULL,		/* lockfunc, lockarg */
1180                                 &sc->req_dmat)) {
1181 		device_printf(sc->mpr_dev, "Cannot allocate request DMA tag\n");
1182 		return (ENOMEM);
1183         }
1184         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1185 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1186 		device_printf(sc->mpr_dev, "Cannot allocate request memory\n");
1187 		return (ENOMEM);
1188         }
1189         bzero(sc->req_frames, rsize);
1190         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1191 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1192 
1193 	/*
1194 	 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to
1195 	 * get the size of a Chain Frame.  Previous versions use the size as a
1196 	 * Request Frame for the Chain Frame size.  If IOCMaxChainSegmentSize
1197 	 * is 0, use the default value.  The IOCMaxChainSegmentSize is the
1198 	 * number of 16-byte elelements that can fit in a Chain Frame, which is
1199 	 * the size of an IEEE Simple SGE.
1200 	 */
1201 	if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) {
1202 		sc->chain_seg_size =
1203 		    htole16(sc->facts->IOCMaxChainSegmentSize);
1204 		if (sc->chain_seg_size == 0) {
1205 			sc->chain_frame_size = MPR_DEFAULT_CHAIN_SEG_SIZE *
1206 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
1207 		} else {
1208 			sc->chain_frame_size = sc->chain_seg_size *
1209 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
1210 		}
1211 	} else {
1212 		sc->chain_frame_size = sc->facts->IOCRequestFrameSize * 4;
1213 	}
1214 	rsize = sc->chain_frame_size * sc->max_chains;
1215         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1216 				16, 0,			/* algnmnt, boundary */
1217 				BUS_SPACE_MAXADDR,	/* lowaddr */
1218 				BUS_SPACE_MAXADDR,	/* highaddr */
1219 				NULL, NULL,		/* filter, filterarg */
1220                                 rsize,			/* maxsize */
1221                                 1,			/* nsegments */
1222                                 rsize,			/* maxsegsize */
1223                                 0,			/* flags */
1224                                 NULL, NULL,		/* lockfunc, lockarg */
1225                                 &sc->chain_dmat)) {
1226 		device_printf(sc->mpr_dev, "Cannot allocate chain DMA tag\n");
1227 		return (ENOMEM);
1228         }
1229         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1230 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1231 		device_printf(sc->mpr_dev, "Cannot allocate chain memory\n");
1232 		return (ENOMEM);
1233         }
1234         bzero(sc->chain_frames, rsize);
1235         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1236 	    mpr_memaddr_cb, &sc->chain_busaddr, 0);
1237 
1238 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1239 	if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1240 				1, 0,			/* algnmnt, boundary */
1241 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1242 				BUS_SPACE_MAXADDR,	/* highaddr */
1243 				NULL, NULL,		/* filter, filterarg */
1244                                 rsize,			/* maxsize */
1245                                 1,			/* nsegments */
1246                                 rsize,			/* maxsegsize */
1247                                 0,			/* flags */
1248                                 NULL, NULL,		/* lockfunc, lockarg */
1249                                 &sc->sense_dmat)) {
1250 		device_printf(sc->mpr_dev, "Cannot allocate sense DMA tag\n");
1251 		return (ENOMEM);
1252         }
1253         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1254 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1255 		device_printf(sc->mpr_dev, "Cannot allocate sense memory\n");
1256 		return (ENOMEM);
1257         }
1258         bzero(sc->sense_frames, rsize);
1259         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1260 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1261 
1262 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->max_chains, M_MPR,
1263 	    M_WAITOK | M_ZERO);
1264 	if (!sc->chains) {
1265 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
1266 		    __func__, __LINE__);
1267 		return (ENOMEM);
1268 	}
1269 	for (i = 0; i < sc->max_chains; i++) {
1270 		chain = &sc->chains[i];
1271 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1272 		    i * sc->chain_frame_size);
1273 		chain->chain_busaddr = sc->chain_busaddr +
1274 		    i * sc->chain_frame_size;
1275 		mpr_free_chain(sc, chain);
1276 		sc->chain_free_lowwater++;
1277 	}
1278 
1279 	/* XXX Need to pick a more precise value */
1280 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1281         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1282 				1, 0,			/* algnmnt, boundary */
1283 				BUS_SPACE_MAXADDR,	/* lowaddr */
1284 				BUS_SPACE_MAXADDR,	/* highaddr */
1285 				NULL, NULL,		/* filter, filterarg */
1286                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1287                                 nsegs,			/* nsegments */
1288                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
1289                                 BUS_DMA_ALLOCNOW,	/* flags */
1290                                 busdma_lock_mutex,	/* lockfunc */
1291 				&sc->mpr_mtx,		/* lockarg */
1292                                 &sc->buffer_dmat)) {
1293 		device_printf(sc->mpr_dev, "Cannot allocate buffer DMA tag\n");
1294 		return (ENOMEM);
1295         }
1296 
1297 	/*
1298 	 * SMID 0 cannot be used as a free command per the firmware spec.
1299 	 * Just drop that command instead of risking accounting bugs.
1300 	 */
1301 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1302 	    M_MPR, M_WAITOK | M_ZERO);
1303 	if (!sc->commands) {
1304 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
1305 		    __func__, __LINE__);
1306 		return (ENOMEM);
1307 	}
1308 	for (i = 1; i < sc->num_reqs; i++) {
1309 		cm = &sc->commands[i];
1310 		cm->cm_req = sc->req_frames +
1311 		    i * sc->facts->IOCRequestFrameSize * 4;
1312 		cm->cm_req_busaddr = sc->req_busaddr +
1313 		    i * sc->facts->IOCRequestFrameSize * 4;
1314 		cm->cm_sense = &sc->sense_frames[i];
1315 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1316 		cm->cm_desc.Default.SMID = i;
1317 		cm->cm_sc = sc;
1318 		TAILQ_INIT(&cm->cm_chain_list);
1319 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1320 
1321 		/* XXX Is a failure here a critical problem? */
1322 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1323 			if (i <= sc->facts->HighPriorityCredit)
1324 				mpr_free_high_priority_command(sc, cm);
1325 			else
1326 				mpr_free_command(sc, cm);
1327 		else {
1328 			panic("failed to allocate command %d\n", i);
1329 			sc->num_reqs = i;
1330 			break;
1331 		}
1332 	}
1333 
1334 	return (0);
1335 }
1336 
1337 static int
1338 mpr_init_queues(struct mpr_softc *sc)
1339 {
1340 	int i;
1341 
1342 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1343 
1344 	/*
1345 	 * According to the spec, we need to use one less reply than we
1346 	 * have space for on the queue.  So sc->num_replies (the number we
1347 	 * use) should be less than sc->fqdepth (allocated size).
1348 	 */
1349 	if (sc->num_replies >= sc->fqdepth)
1350 		return (EINVAL);
1351 
1352 	/*
1353 	 * Initialize all of the free queue entries.
1354 	 */
1355 	for (i = 0; i < sc->fqdepth; i++)
1356 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1357 	sc->replyfreeindex = sc->num_replies;
1358 
1359 	return (0);
1360 }
1361 
1362 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1363  * Next are the global settings, if they exist.  Highest are the per-unit
1364  * settings, if they exist.
1365  */
1366 static void
1367 mpr_get_tunables(struct mpr_softc *sc)
1368 {
1369 	char tmpstr[80];
1370 
1371 	/* XXX default to some debugging for now */
1372 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1373 	sc->disable_msix = 0;
1374 	sc->disable_msi = 0;
1375 	sc->max_chains = MPR_CHAIN_FRAMES;
1376 	sc->max_io_pages = MPR_MAXIO_PAGES;
1377 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1378 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1379 
1380 	/*
1381 	 * Grab the global variables.
1382 	 */
1383 	TUNABLE_INT_FETCH("hw.mpr.debug_level", &sc->mpr_debug);
1384 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1385 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1386 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1387 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1388 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1389 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1390 
1391 	/* Grab the unit-instance variables */
1392 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1393 	    device_get_unit(sc->mpr_dev));
1394 	TUNABLE_INT_FETCH(tmpstr, &sc->mpr_debug);
1395 
1396 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1397 	    device_get_unit(sc->mpr_dev));
1398 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1399 
1400 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1401 	    device_get_unit(sc->mpr_dev));
1402 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1403 
1404 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1405 	    device_get_unit(sc->mpr_dev));
1406 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1407 
1408 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1409 	    device_get_unit(sc->mpr_dev));
1410 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1411 
1412 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1413 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1414 	    device_get_unit(sc->mpr_dev));
1415 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1416 
1417 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1418 	    device_get_unit(sc->mpr_dev));
1419 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1420 
1421 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1422 	    device_get_unit(sc->mpr_dev));
1423 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1424 }
1425 
1426 static void
1427 mpr_setup_sysctl(struct mpr_softc *sc)
1428 {
1429 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1430 	struct sysctl_oid	*sysctl_tree = NULL;
1431 	char tmpstr[80], tmpstr2[80];
1432 
1433 	/*
1434 	 * Setup the sysctl variable so the user can change the debug level
1435 	 * on the fly.
1436 	 */
1437 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1438 	    device_get_unit(sc->mpr_dev));
1439 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1440 
1441 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1442 	if (sysctl_ctx != NULL)
1443 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1444 
1445 	if (sysctl_tree == NULL) {
1446 		sysctl_ctx_init(&sc->sysctl_ctx);
1447 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1448 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1449 		    CTLFLAG_RD, 0, tmpstr);
1450 		if (sc->sysctl_tree == NULL)
1451 			return;
1452 		sysctl_ctx = &sc->sysctl_ctx;
1453 		sysctl_tree = sc->sysctl_tree;
1454 	}
1455 
1456 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1457 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mpr_debug, 0,
1458 	    "mpr debug level");
1459 
1460 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1461 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1462 	    "Disable the use of MSI-X interrupts");
1463 
1464 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1465 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1466 	    "Disable the use of MSI interrupts");
1467 
1468 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1469 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1470 	    strlen(sc->fw_version), "firmware version");
1471 
1472 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1473 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPR_DRIVER_VERSION,
1474 	    strlen(MPR_DRIVER_VERSION), "driver version");
1475 
1476 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1477 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1478 	    &sc->io_cmds_active, 0, "number of currently active commands");
1479 
1480 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1481 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1482 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1483 
1484 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1485 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1486 	    &sc->chain_free, 0, "number of free chain elements");
1487 
1488 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1489 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1490 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1491 
1492 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1493 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1494 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1495 
1496 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1497 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1498 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1499 	    "IOCFacts)");
1500 
1501 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1502 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1503 	    "enable SSU to SATA SSD/HDD at shutdown");
1504 
1505 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1506 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1507 	    &sc->chain_alloc_fail, "chain allocation failures");
1508 
1509 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1510 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1511 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1512 	    "spinup after SATA ID error");
1513 }
1514 
1515 int
1516 mpr_attach(struct mpr_softc *sc)
1517 {
1518 	int error;
1519 
1520 	mpr_get_tunables(sc);
1521 
1522 	MPR_FUNCTRACE(sc);
1523 
1524 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
1525 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
1526 	TAILQ_INIT(&sc->event_list);
1527 	timevalclear(&sc->lastfail);
1528 
1529 	if ((error = mpr_transition_ready(sc)) != 0) {
1530 		mpr_printf(sc, "%s failed to transition ready\n", __func__);
1531 		return (error);
1532 	}
1533 
1534 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
1535 	    M_ZERO|M_NOWAIT);
1536 	if (!sc->facts) {
1537 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
1538 		    __func__, __LINE__);
1539 		return (ENOMEM);
1540 	}
1541 
1542 	/*
1543 	 * Get IOC Facts and allocate all structures based on this information.
1544 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
1545 	 * Facts. If relevant values have changed in IOC Facts, this function
1546 	 * will free all of the memory based on IOC Facts and reallocate that
1547 	 * memory.  If this fails, any allocated memory should already be freed.
1548 	 */
1549 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
1550 		mpr_dprint(sc, MPR_FAULT, "%s IOC Facts based allocation "
1551 		    "failed with error %d\n", __func__, error);
1552 		return (error);
1553 	}
1554 
1555 	/* Start the periodic watchdog check on the IOC Doorbell */
1556 	mpr_periodic(sc);
1557 
1558 	/*
1559 	 * The portenable will kick off discovery events that will drive the
1560 	 * rest of the initialization process.  The CAM/SAS module will
1561 	 * hold up the boot sequence until discovery is complete.
1562 	 */
1563 	sc->mpr_ich.ich_func = mpr_startup;
1564 	sc->mpr_ich.ich_arg = sc;
1565 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
1566 		mpr_dprint(sc, MPR_ERROR, "Cannot establish MPR config hook\n");
1567 		error = EINVAL;
1568 	}
1569 
1570 	/*
1571 	 * Allow IR to shutdown gracefully when shutdown occurs.
1572 	 */
1573 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1574 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1575 
1576 	if (sc->shutdown_eh == NULL)
1577 		mpr_dprint(sc, MPR_ERROR, "shutdown event registration "
1578 		    "failed\n");
1579 
1580 	mpr_setup_sysctl(sc);
1581 
1582 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
1583 
1584 	return (error);
1585 }
1586 
1587 /* Run through any late-start handlers. */
1588 static void
1589 mpr_startup(void *arg)
1590 {
1591 	struct mpr_softc *sc;
1592 
1593 	sc = (struct mpr_softc *)arg;
1594 
1595 	mpr_lock(sc);
1596 	mpr_unmask_intr(sc);
1597 
1598 	/* initialize device mapping tables */
1599 	mpr_base_static_config_pages(sc);
1600 	mpr_mapping_initialize(sc);
1601 	mprsas_startup(sc);
1602 	mpr_unlock(sc);
1603 }
1604 
1605 /* Periodic watchdog.  Is called with the driver lock already held. */
1606 static void
1607 mpr_periodic(void *arg)
1608 {
1609 	struct mpr_softc *sc;
1610 	uint32_t db;
1611 
1612 	sc = (struct mpr_softc *)arg;
1613 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
1614 		return;
1615 
1616 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
1617 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1618 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
1619 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
1620 			panic("TEMPERATURE FAULT: STOPPING.");
1621 		}
1622 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1623 		mpr_reinit(sc);
1624 	}
1625 
1626 	callout_reset(&sc->periodic, MPR_PERIODIC_DELAY * hz, mpr_periodic, sc);
1627 }
1628 
1629 static void
1630 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
1631     MPI2_EVENT_NOTIFICATION_REPLY *event)
1632 {
1633 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1634 
1635 	mpr_print_event(sc, event);
1636 
1637 	switch (event->Event) {
1638 	case MPI2_EVENT_LOG_DATA:
1639 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1640 		if (sc->mpr_debug & MPR_EVENT)
1641 			hexdump(event->EventData, event->EventDataLength, NULL,
1642 			    0);
1643 		break;
1644 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1645 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1646 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1647 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1648 		     entry->LogSequence);
1649 		break;
1650 	default:
1651 		break;
1652 	}
1653 	return;
1654 }
1655 
1656 static int
1657 mpr_attach_log(struct mpr_softc *sc)
1658 {
1659 	uint8_t events[16];
1660 
1661 	bzero(events, 16);
1662 	setbit(events, MPI2_EVENT_LOG_DATA);
1663 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1664 
1665 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
1666 	    &sc->mpr_log_eh);
1667 
1668 	return (0);
1669 }
1670 
1671 static int
1672 mpr_detach_log(struct mpr_softc *sc)
1673 {
1674 
1675 	if (sc->mpr_log_eh != NULL)
1676 		mpr_deregister_events(sc, sc->mpr_log_eh);
1677 	return (0);
1678 }
1679 
1680 /*
1681  * Free all of the driver resources and detach submodules.  Should be called
1682  * without the lock held.
1683  */
1684 int
1685 mpr_free(struct mpr_softc *sc)
1686 {
1687 	int error;
1688 
1689 	/* Turn off the watchdog */
1690 	mpr_lock(sc);
1691 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
1692 	mpr_unlock(sc);
1693 	/* Lock must not be held for this */
1694 	callout_drain(&sc->periodic);
1695 
1696 	if (((error = mpr_detach_log(sc)) != 0) ||
1697 	    ((error = mpr_detach_sas(sc)) != 0))
1698 		return (error);
1699 
1700 	mpr_detach_user(sc);
1701 
1702 	/* Put the IOC back in the READY state. */
1703 	mpr_lock(sc);
1704 	if ((error = mpr_transition_ready(sc)) != 0) {
1705 		mpr_unlock(sc);
1706 		return (error);
1707 	}
1708 	mpr_unlock(sc);
1709 
1710 	if (sc->facts != NULL)
1711 		free(sc->facts, M_MPR);
1712 
1713 	/*
1714 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
1715 	 * to free these buffers too.
1716 	 */
1717 	mpr_iocfacts_free(sc);
1718 
1719 	if (sc->sysctl_tree != NULL)
1720 		sysctl_ctx_free(&sc->sysctl_ctx);
1721 
1722 	/* Deregister the shutdown function */
1723 	if (sc->shutdown_eh != NULL)
1724 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1725 
1726 	mtx_destroy(&sc->mpr_mtx);
1727 
1728 	return (0);
1729 }
1730 
1731 static __inline void
1732 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
1733 {
1734 	MPR_FUNCTRACE(sc);
1735 
1736 	if (cm == NULL) {
1737 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
1738 		return;
1739 	}
1740 
1741 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
1742 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
1743 
1744 	if (cm->cm_complete != NULL) {
1745 		mpr_dprint(sc, MPR_TRACE,
1746 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
1747 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
1748 		    cm->cm_reply);
1749 		cm->cm_complete(sc, cm);
1750 	}
1751 
1752 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
1753 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
1754 		wakeup(cm);
1755 	}
1756 
1757 	if (sc->io_cmds_active != 0) {
1758 		sc->io_cmds_active--;
1759 	} else {
1760 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
1761 		    "out of sync - resynching to 0\n");
1762 	}
1763 }
1764 
1765 static void
1766 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
1767 {
1768 	union loginfo_type {
1769 		u32	loginfo;
1770 		struct {
1771 			u32	subcode:16;
1772 			u32	code:8;
1773 			u32	originator:4;
1774 			u32	bus_type:4;
1775 		} dw;
1776 	};
1777 	union loginfo_type sas_loginfo;
1778 	char *originator_str = NULL;
1779 
1780 	sas_loginfo.loginfo = log_info;
1781 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1782 		return;
1783 
1784 	/* each nexus loss loginfo */
1785 	if (log_info == 0x31170000)
1786 		return;
1787 
1788 	/* eat the loginfos associated with task aborts */
1789 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
1790 	    (log_info == 0x31130000))
1791 		return;
1792 
1793 	switch (sas_loginfo.dw.originator) {
1794 	case 0:
1795 		originator_str = "IOP";
1796 		break;
1797 	case 1:
1798 		originator_str = "PL";
1799 		break;
1800 	case 2:
1801 		originator_str = "IR";
1802 		break;
1803 	}
1804 
1805 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
1806 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
1807 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
1808 }
1809 
1810 static void
1811 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
1812 {
1813 	MPI2DefaultReply_t *mpi_reply;
1814 	u16 sc_status;
1815 
1816 	mpi_reply = (MPI2DefaultReply_t*)reply;
1817 	sc_status = le16toh(mpi_reply->IOCStatus);
1818 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1819 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1820 }
1821 
1822 void
1823 mpr_intr(void *data)
1824 {
1825 	struct mpr_softc *sc;
1826 	uint32_t status;
1827 
1828 	sc = (struct mpr_softc *)data;
1829 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
1830 
1831 	/*
1832 	 * Check interrupt status register to flush the bus.  This is
1833 	 * needed for both INTx interrupts and driver-driven polling
1834 	 */
1835 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1836 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1837 		return;
1838 
1839 	mpr_lock(sc);
1840 	mpr_intr_locked(data);
1841 	mpr_unlock(sc);
1842 	return;
1843 }
1844 
1845 /*
1846  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1847  * chip.  Hopefully this theory is correct.
1848  */
1849 void
1850 mpr_intr_msi(void *data)
1851 {
1852 	struct mpr_softc *sc;
1853 
1854 	sc = (struct mpr_softc *)data;
1855 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
1856 	mpr_lock(sc);
1857 	mpr_intr_locked(data);
1858 	mpr_unlock(sc);
1859 	return;
1860 }
1861 
1862 /*
1863  * The locking is overly broad and simplistic, but easy to deal with for now.
1864  */
1865 void
1866 mpr_intr_locked(void *data)
1867 {
1868 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1869 	struct mpr_softc *sc;
1870 	struct mpr_command *cm = NULL;
1871 	uint8_t flags;
1872 	u_int pq;
1873 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
1874 	mpr_fw_diagnostic_buffer_t *pBuffer;
1875 
1876 	sc = (struct mpr_softc *)data;
1877 
1878 	pq = sc->replypostindex;
1879 	mpr_dprint(sc, MPR_TRACE,
1880 	    "%s sc %p starting with replypostindex %u\n",
1881 	    __func__, sc, sc->replypostindex);
1882 
1883 	for ( ;; ) {
1884 		cm = NULL;
1885 		desc = &sc->post_queue[sc->replypostindex];
1886 		flags = desc->Default.ReplyFlags &
1887 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1888 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
1889 		    (le32toh(desc->Words.High) == 0xffffffff))
1890 			break;
1891 
1892 		/* increment the replypostindex now, so that event handlers
1893 		 * and cm completion handlers which decide to do a diag
1894 		 * reset can zero it without it getting incremented again
1895 		 * afterwards, and we break out of this loop on the next
1896 		 * iteration since the reply post queue has been cleared to
1897 		 * 0xFF and all descriptors look unused (which they are).
1898 		 */
1899 		if (++sc->replypostindex >= sc->pqdepth)
1900 			sc->replypostindex = 0;
1901 
1902 		switch (flags) {
1903 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1904 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
1905 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
1906 			cm->cm_reply = NULL;
1907 			break;
1908 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1909 		{
1910 			uint32_t baddr;
1911 			uint8_t *reply;
1912 
1913 			/*
1914 			 * Re-compose the reply address from the address
1915 			 * sent back from the chip.  The ReplyFrameAddress
1916 			 * is the lower 32 bits of the physical address of
1917 			 * particular reply frame.  Convert that address to
1918 			 * host format, and then use that to provide the
1919 			 * offset against the virtual address base
1920 			 * (sc->reply_frames).
1921 			 */
1922 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1923 			reply = sc->reply_frames +
1924 				(baddr - ((uint32_t)sc->reply_busaddr));
1925 			/*
1926 			 * Make sure the reply we got back is in a valid
1927 			 * range.  If not, go ahead and panic here, since
1928 			 * we'll probably panic as soon as we deference the
1929 			 * reply pointer anyway.
1930 			 */
1931 			if ((reply < sc->reply_frames)
1932 			 || (reply > (sc->reply_frames +
1933 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1934 				printf("%s: WARNING: reply %p out of range!\n",
1935 				       __func__, reply);
1936 				printf("%s: reply_frames %p, fqdepth %d, "
1937 				       "frame size %d\n", __func__,
1938 				       sc->reply_frames, sc->fqdepth,
1939 				       sc->facts->ReplyFrameSize * 4);
1940 				printf("%s: baddr %#x,\n", __func__, baddr);
1941 				/* LSI-TODO. See Linux Code for Graceful exit */
1942 				panic("Reply address out of range");
1943 			}
1944 			if (le16toh(desc->AddressReply.SMID) == 0) {
1945 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
1946 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
1947 					/*
1948 					 * If SMID is 0 for Diag Buffer Post,
1949 					 * this implies that the reply is due to
1950 					 * a release function with a status that
1951 					 * the buffer has been released.  Set
1952 					 * the buffer flags accordingly.
1953 					 */
1954 					rel_rep =
1955 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
1956 					if ((le16toh(rel_rep->IOCStatus) &
1957 					    MPI2_IOCSTATUS_MASK) ==
1958 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
1959 					{
1960 						pBuffer =
1961 						    &sc->fw_diag_buffer_list[
1962 						    rel_rep->BufferType];
1963 						pBuffer->valid_data = TRUE;
1964 						pBuffer->owned_by_firmware =
1965 						    FALSE;
1966 						pBuffer->immediate = FALSE;
1967 					}
1968 				} else
1969 					mpr_dispatch_event(sc, baddr,
1970 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
1971 					    reply);
1972 			} else {
1973 				cm = &sc->commands[
1974 				    le16toh(desc->AddressReply.SMID)];
1975 				cm->cm_reply = reply;
1976 				cm->cm_reply_data =
1977 				    le32toh(desc->AddressReply.
1978 				    ReplyFrameAddress);
1979 			}
1980 			break;
1981 		}
1982 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
1983 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
1984 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
1985 		default:
1986 			/* Unhandled */
1987 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
1988 			    desc->Default.ReplyFlags);
1989 			cm = NULL;
1990 			break;
1991 		}
1992 
1993 		if (cm != NULL) {
1994 			// Print Error reply frame
1995 			if (cm->cm_reply)
1996 				mpr_display_reply_info(sc,cm->cm_reply);
1997 			mpr_complete_command(sc, cm);
1998 		}
1999 
2000 		desc->Words.Low = 0xffffffff;
2001 		desc->Words.High = 0xffffffff;
2002 	}
2003 
2004 	if (pq != sc->replypostindex) {
2005 		mpr_dprint(sc, MPR_TRACE,
2006 		    "%s sc %p writing postindex %d\n",
2007 		    __func__, sc, sc->replypostindex);
2008 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2009 		    sc->replypostindex);
2010 	}
2011 
2012 	return;
2013 }
2014 
2015 static void
2016 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2017     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2018 {
2019 	struct mpr_event_handle *eh;
2020 	int event, handled = 0;
2021 
2022 	event = le16toh(reply->Event);
2023 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2024 		if (isset(eh->mask, event)) {
2025 			eh->callback(sc, data, reply);
2026 			handled++;
2027 		}
2028 	}
2029 
2030 	if (handled == 0)
2031 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2032 		    le16toh(event));
2033 
2034 	/*
2035 	 * This is the only place that the event/reply should be freed.
2036 	 * Anything wanting to hold onto the event data should have
2037 	 * already copied it into their own storage.
2038 	 */
2039 	mpr_free_reply(sc, data);
2040 }
2041 
2042 static void
2043 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2044 {
2045 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2046 
2047 	if (cm->cm_reply)
2048 		mpr_print_event(sc,
2049 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2050 
2051 	mpr_free_command(sc, cm);
2052 
2053 	/* next, send a port enable */
2054 	mprsas_startup(sc);
2055 }
2056 
2057 /*
2058  * For both register_events and update_events, the caller supplies a bitmap
2059  * of events that it _wants_.  These functions then turn that into a bitmask
2060  * suitable for the controller.
2061  */
2062 int
2063 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2064     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2065 {
2066 	struct mpr_event_handle *eh;
2067 	int error = 0;
2068 
2069 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2070 	if (!eh) {
2071 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
2072 		    __func__, __LINE__);
2073 		return (ENOMEM);
2074 	}
2075 	eh->callback = cb;
2076 	eh->data = data;
2077 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2078 	if (mask != NULL)
2079 		error = mpr_update_events(sc, eh, mask);
2080 	*handle = eh;
2081 
2082 	return (error);
2083 }
2084 
2085 int
2086 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2087     uint8_t *mask)
2088 {
2089 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2090 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
2091 	struct mpr_command *cm;
2092 	struct mpr_event_handle *eh;
2093 	int error, i;
2094 
2095 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2096 
2097 	if ((mask != NULL) && (handle != NULL))
2098 		bcopy(mask, &handle->mask[0], 16);
2099 	memset(sc->event_mask, 0xff, 16);
2100 
2101 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2102 		for (i = 0; i < 16; i++)
2103 			sc->event_mask[i] &= ~eh->mask[i];
2104 	}
2105 
2106 	if ((cm = mpr_alloc_command(sc)) == NULL)
2107 		return (EBUSY);
2108 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2109 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2110 	evtreq->MsgFlags = 0;
2111 	evtreq->SASBroadcastPrimitiveMasks = 0;
2112 #ifdef MPR_DEBUG_ALL_EVENTS
2113 	{
2114 		u_char fullmask[16];
2115 		memset(fullmask, 0x00, 16);
2116 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2117 	}
2118 #else
2119 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2120 #endif
2121 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2122 	cm->cm_data = NULL;
2123 
2124 	error = mpr_request_polled(sc, cm);
2125 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2126 	if ((reply == NULL) ||
2127 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2128 		error = ENXIO;
2129 
2130 	if (reply)
2131 		mpr_print_event(sc, reply);
2132 
2133 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2134 
2135 	mpr_free_command(sc, cm);
2136 	return (error);
2137 }
2138 
2139 static int
2140 mpr_reregister_events(struct mpr_softc *sc)
2141 {
2142 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2143 	struct mpr_command *cm;
2144 	struct mpr_event_handle *eh;
2145 	int error, i;
2146 
2147 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2148 
2149 	/* first, reregister events */
2150 
2151 	memset(sc->event_mask, 0xff, 16);
2152 
2153 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2154 		for (i = 0; i < 16; i++)
2155 			sc->event_mask[i] &= ~eh->mask[i];
2156 	}
2157 
2158 	if ((cm = mpr_alloc_command(sc)) == NULL)
2159 		return (EBUSY);
2160 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2161 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2162 	evtreq->MsgFlags = 0;
2163 	evtreq->SASBroadcastPrimitiveMasks = 0;
2164 #ifdef MPR_DEBUG_ALL_EVENTS
2165 	{
2166 		u_char fullmask[16];
2167 		memset(fullmask, 0x00, 16);
2168 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2169 	}
2170 #else
2171 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2172 #endif
2173 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2174 	cm->cm_data = NULL;
2175 	cm->cm_complete = mpr_reregister_events_complete;
2176 
2177 	error = mpr_map_command(sc, cm);
2178 
2179 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2180 	    error);
2181 	return (error);
2182 }
2183 
2184 int
2185 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2186 {
2187 
2188 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2189 	free(handle, M_MPR);
2190 	return (mpr_update_events(sc, NULL, NULL));
2191 }
2192 
2193 /*
2194  * Add a chain element as the next SGE for the specified command.
2195  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
2196  * only required for IEEE commands.  Therefore there is no code for commands
2197  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
2198  * shouldn't be requesting chains).
2199  */
2200 static int
2201 mpr_add_chain(struct mpr_command *cm, int segsleft)
2202 {
2203 	struct mpr_softc *sc = cm->cm_sc;
2204 	MPI2_REQUEST_HEADER *req;
2205 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
2206 	struct mpr_chain *chain;
2207 	int sgc_size, current_segs, rem_segs, segs_per_frame;
2208 	uint8_t next_chain_offset = 0;
2209 
2210 	/*
2211 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
2212 	 * only IEEE commands should be requesting chains.  Return some error
2213 	 * code other than 0.
2214 	 */
2215 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
2216 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
2217 		    "an MPI SGL.\n");
2218 		return(ENOBUFS);
2219 	}
2220 
2221 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
2222 	if (cm->cm_sglsize < sgc_size)
2223 		panic("MPR: Need SGE Error Code\n");
2224 
2225 	chain = mpr_alloc_chain(cm->cm_sc);
2226 	if (chain == NULL)
2227 		return (ENOBUFS);
2228 
2229 	/*
2230 	 * Note: a double-linked list is used to make it easier to walk for
2231 	 * debugging.
2232 	 */
2233 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2234 
2235 	/*
2236 	 * Need to know if the number of frames left is more than 1 or not.  If
2237 	 * more than 1 frame is required, NextChainOffset will need to be set,
2238 	 * which will just be the last segment of the frame.
2239 	 */
2240 	rem_segs = 0;
2241 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
2242 		/*
2243 		 * rem_segs is the number of segements remaining after the
2244 		 * segments that will go into the current frame.  Since it is
2245 		 * known that at least one more frame is required, account for
2246 		 * the chain element.  To know if more than one more frame is
2247 		 * required, just check if there will be a remainder after using
2248 		 * the current frame (with this chain) and the next frame.  If
2249 		 * so the NextChainOffset must be the last element of the next
2250 		 * frame.
2251 		 */
2252 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
2253 		rem_segs = segsleft - current_segs;
2254 		segs_per_frame = sc->chain_frame_size / sgc_size;
2255 		if (rem_segs > segs_per_frame) {
2256 			next_chain_offset = segs_per_frame - 1;
2257 		}
2258 	}
2259 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
2260 	ieee_sgc->Length = next_chain_offset ?
2261 	    htole32((uint32_t)sc->chain_frame_size) :
2262 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
2263 	ieee_sgc->NextChainOffset = next_chain_offset;
2264 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
2265 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
2266 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
2267 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
2268 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
2269 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
2270 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
2271 
2272 	cm->cm_sglsize = sc->chain_frame_size;
2273 	return (0);
2274 }
2275 
2276 /*
2277  * Add one scatter-gather element to the scatter-gather list for a command.
2278  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
2279  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
2280  * chain, so don't consider any chain additions.
2281  */
2282 int
2283 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
2284     int segsleft)
2285 {
2286 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2287 	u32 sge_flags;
2288 
2289 	/*
2290 	 * case 1: >=1 more segment, no room for anything (error)
2291 	 * case 2: 1 more segment and enough room for it
2292          */
2293 
2294 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
2295 		mpr_dprint(cm->cm_sc, MPR_ERROR,
2296 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
2297 		    __func__);
2298 		return(ENOBUFS);
2299 	}
2300 
2301 	KASSERT(segsleft == 1,
2302 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
2303 	    segsleft));
2304 
2305 	/*
2306 	 * There is one more segment left to add for the MPI SGL and there is
2307 	 * enough room in the frame to add it.  This is the normal case because
2308 	 * MPI SGL's don't have chains, otherwise something is wrong.
2309 	 *
2310 	 * If this is a bi-directional request, need to account for that
2311 	 * here.  Save the pre-filled sge values.  These will be used
2312 	 * either for the 2nd SGL or for a single direction SGL.  If
2313 	 * cm_out_len is non-zero, this is a bi-directional request, so
2314 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
2315 	 * fill in the IN SGL.  Note that at this time, when filling in
2316 	 * 2 SGL's for a bi-directional request, they both use the same
2317 	 * DMA buffer (same cm command).
2318 	 */
2319 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
2320 	saved_address_low = sge->Address.Low;
2321 	saved_address_high = sge->Address.High;
2322 	if (cm->cm_out_len) {
2323 		sge->FlagsLength = cm->cm_out_len |
2324 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2325 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2326 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
2327 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2328 		    MPI2_SGE_FLAGS_SHIFT);
2329 		cm->cm_sglsize -= len;
2330 		/* Endian Safe code */
2331 		sge_flags = sge->FlagsLength;
2332 		sge->FlagsLength = htole32(sge_flags);
2333 		sge->Address.High = htole32(sge->Address.High);
2334 		sge->Address.Low = htole32(sge->Address.Low);
2335 		bcopy(sge, cm->cm_sge, len);
2336 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2337 	}
2338 	sge->FlagsLength = saved_buf_len |
2339 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2340 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
2341 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
2342 	    MPI2_SGE_FLAGS_END_OF_LIST |
2343 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2344 	    MPI2_SGE_FLAGS_SHIFT);
2345 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
2346 		sge->FlagsLength |=
2347 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2348 		    MPI2_SGE_FLAGS_SHIFT);
2349 	} else {
2350 		sge->FlagsLength |=
2351 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2352 		    MPI2_SGE_FLAGS_SHIFT);
2353 	}
2354 	sge->Address.Low = saved_address_low;
2355 	sge->Address.High = saved_address_high;
2356 
2357 	cm->cm_sglsize -= len;
2358 	/* Endian Safe code */
2359 	sge_flags = sge->FlagsLength;
2360 	sge->FlagsLength = htole32(sge_flags);
2361 	sge->Address.High = htole32(sge->Address.High);
2362 	sge->Address.Low = htole32(sge->Address.Low);
2363 	bcopy(sge, cm->cm_sge, len);
2364 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2365 	return (0);
2366 }
2367 
2368 /*
2369  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
2370  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
2371  * remaining size and pointer to the next SGE to fill in, respectively.
2372  */
2373 int
2374 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
2375 {
2376 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
2377 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
2378 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2379 	uint32_t sge_length;
2380 
2381 	/*
2382 	 * case 1: No room for chain or segment (error).
2383 	 * case 2: Two or more segments left but only room for chain.
2384 	 * case 3: Last segment and room for it, so set flags.
2385 	 */
2386 
2387 	/*
2388 	 * There should be room for at least one element, or there is a big
2389 	 * problem.
2390 	 */
2391 	if (cm->cm_sglsize < ieee_sge_size)
2392 		panic("MPR: Need SGE Error Code\n");
2393 
2394 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
2395 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
2396 			return (error);
2397 	}
2398 
2399 	if (segsleft == 1) {
2400 		/*
2401 		 * If this is a bi-directional request, need to account for that
2402 		 * here.  Save the pre-filled sge values.  These will be used
2403 		 * either for the 2nd SGL or for a single direction SGL.  If
2404 		 * cm_out_len is non-zero, this is a bi-directional request, so
2405 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2406 		 * fill in the IN SGL.  Note that at this time, when filling in
2407 		 * 2 SGL's for a bi-directional request, they both use the same
2408 		 * DMA buffer (same cm command).
2409 		 */
2410 		saved_buf_len = sge->Length;
2411 		saved_address_low = sge->Address.Low;
2412 		saved_address_high = sge->Address.High;
2413 		if (cm->cm_out_len) {
2414 			sge->Length = cm->cm_out_len;
2415 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
2416 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
2417 			cm->cm_sglsize -= ieee_sge_size;
2418 			/* Endian Safe code */
2419 			sge_length = sge->Length;
2420 			sge->Length = htole32(sge_length);
2421 			sge->Address.High = htole32(sge->Address.High);
2422 			sge->Address.Low = htole32(sge->Address.Low);
2423 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
2424 			cm->cm_sge =
2425 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
2426 			    ieee_sge_size);
2427 		}
2428 		sge->Length = saved_buf_len;
2429 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
2430 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
2431 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
2432 		sge->Address.Low = saved_address_low;
2433 		sge->Address.High = saved_address_high;
2434 	}
2435 
2436 	cm->cm_sglsize -= ieee_sge_size;
2437 	/* Endian Safe code */
2438 	sge_length = sge->Length;
2439 	sge->Length = htole32(sge_length);
2440 	sge->Address.High = htole32(sge->Address.High);
2441 	sge->Address.Low = htole32(sge->Address.Low);
2442 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
2443 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
2444 	    ieee_sge_size);
2445 	return (0);
2446 }
2447 
2448 /*
2449  * Add one dma segment to the scatter-gather list for a command.
2450  */
2451 int
2452 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2453     int segsleft)
2454 {
2455 	MPI2_SGE_SIMPLE64 sge;
2456 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
2457 
2458 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
2459 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
2460 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
2461 		ieee_sge.Length = len;
2462 		mpr_from_u64(pa, &ieee_sge.Address);
2463 
2464 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
2465 	} else {
2466 		/*
2467 		 * This driver always uses 64-bit address elements for
2468 		 * simplicity.
2469 		 */
2470 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2471 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2472 		/* Set Endian safe macro in mpr_push_sge */
2473 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
2474 		mpr_from_u64(pa, &sge.Address);
2475 
2476 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
2477 	}
2478 }
2479 
2480 static void
2481 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2482 {
2483 	struct mpr_softc *sc;
2484 	struct mpr_command *cm;
2485 	u_int i, dir, sflags;
2486 
2487 	cm = (struct mpr_command *)arg;
2488 	sc = cm->cm_sc;
2489 
2490 	/*
2491 	 * In this case, just print out a warning and let the chip tell the
2492 	 * user they did the wrong thing.
2493 	 */
2494 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2495 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
2496 		    "segments, more than the %d allowed\n", __func__, nsegs,
2497 		    cm->cm_max_segs);
2498 	}
2499 
2500 	/*
2501 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2502 	 * here.  In that case, both direction flags will be set.
2503 	 */
2504 	sflags = 0;
2505 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
2506 		/*
2507 		 * We have to add a special case for SMP passthrough, there
2508 		 * is no easy way to generically handle it.  The first
2509 		 * S/G element is used for the command (therefore the
2510 		 * direction bit needs to be set).  The second one is used
2511 		 * for the reply.  We'll leave it to the caller to make
2512 		 * sure we only have two buffers.
2513 		 */
2514 		/*
2515 		 * Even though the busdma man page says it doesn't make
2516 		 * sense to have both direction flags, it does in this case.
2517 		 * We have one s/g element being accessed in each direction.
2518 		 */
2519 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2520 
2521 		/*
2522 		 * Set the direction flag on the first buffer in the SMP
2523 		 * passthrough request.  We'll clear it for the second one.
2524 		 */
2525 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2526 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2527 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
2528 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2529 		dir = BUS_DMASYNC_PREWRITE;
2530 	} else
2531 		dir = BUS_DMASYNC_PREREAD;
2532 
2533 	for (i = 0; i < nsegs; i++) {
2534 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
2535 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2536 		}
2537 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2538 		    sflags, nsegs - i);
2539 		if (error != 0) {
2540 			/* Resource shortage, roll back! */
2541 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
2542 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
2543 				    "consider increasing hw.mpr.max_chains.\n");
2544 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
2545 			mpr_complete_command(sc, cm);
2546 			return;
2547 		}
2548 	}
2549 
2550 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2551 	mpr_enqueue_request(sc, cm);
2552 
2553 	return;
2554 }
2555 
2556 static void
2557 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2558 	     int error)
2559 {
2560 	mpr_data_cb(arg, segs, nsegs, error);
2561 }
2562 
2563 /*
2564  * This is the routine to enqueue commands ansynchronously.
2565  * Note that the only error path here is from bus_dmamap_load(), which can
2566  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2567  * assumed that if you have a command in-hand, then you have enough credits
2568  * to use it.
2569  */
2570 int
2571 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
2572 {
2573 	int error = 0;
2574 
2575 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
2576 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2577 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
2578 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
2579 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2580 		    cm->cm_data, mpr_data_cb, cm, 0);
2581 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2582 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2583 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
2584 	} else {
2585 		/* Add a zero-length element as needed */
2586 		if (cm->cm_sge != NULL)
2587 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
2588 		mpr_enqueue_request(sc, cm);
2589 	}
2590 
2591 	return (error);
2592 }
2593 
2594 /*
2595  * This is the routine to enqueue commands synchronously.  An error of
2596  * EINPROGRESS from mpr_map_command() is ignored since the command will
2597  * be executed and enqueued automatically.  Other errors come from msleep().
2598  */
2599 int
2600 mpr_wait_command(struct mpr_softc *sc, struct mpr_command *cm, int timeout,
2601     int sleep_flag)
2602 {
2603 	int error, rc;
2604 	struct timeval cur_time, start_time;
2605 
2606 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
2607 		return  EBUSY;
2608 
2609 	cm->cm_complete = NULL;
2610 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
2611 	error = mpr_map_command(sc, cm);
2612 	if ((error != 0) && (error != EINPROGRESS))
2613 		return (error);
2614 
2615 	// Check for context and wait for 50 mSec at a time until time has
2616 	// expired or the command has finished.  If msleep can't be used, need
2617 	// to poll.
2618 #if __FreeBSD_version >= 1000029
2619 	if (curthread->td_no_sleeping)
2620 #else //__FreeBSD_version < 1000029
2621 	if (curthread->td_pflags & TDP_NOSLEEPING)
2622 #endif //__FreeBSD_version >= 1000029
2623 		sleep_flag = NO_SLEEP;
2624 	getmicrotime(&start_time);
2625 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
2626 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
2627 	} else {
2628 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
2629 			mpr_intr_locked(sc);
2630 			if (sleep_flag == CAN_SLEEP)
2631 				pause("mprwait", hz/20);
2632 			else
2633 				DELAY(50000);
2634 
2635 			getmicrotime(&cur_time);
2636 			if ((cur_time.tv_sec - start_time.tv_sec) > timeout) {
2637 				error = EWOULDBLOCK;
2638 				break;
2639 			}
2640 		}
2641 	}
2642 
2643 	if (error == EWOULDBLOCK) {
2644 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
2645 		rc = mpr_reinit(sc);
2646 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2647 		    "failed");
2648 		error = ETIMEDOUT;
2649 	}
2650 	return (error);
2651 }
2652 
2653 /*
2654  * This is the routine to enqueue a command synchonously and poll for
2655  * completion.  Its use should be rare.
2656  */
2657 int
2658 mpr_request_polled(struct mpr_softc *sc, struct mpr_command *cm)
2659 {
2660 	int error, timeout = 0, rc;
2661 	struct timeval cur_time, start_time;
2662 
2663 	error = 0;
2664 
2665 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
2666 	cm->cm_complete = NULL;
2667 	mpr_map_command(sc, cm);
2668 
2669 	getmicrotime(&start_time);
2670 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
2671 		mpr_intr_locked(sc);
2672 
2673 		if (mtx_owned(&sc->mpr_mtx))
2674 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
2675 			    "mprpoll", hz/20);
2676 		else
2677 			pause("mprpoll", hz/20);
2678 
2679 		/*
2680 		 * Check for real-time timeout and fail if more than 60 seconds.
2681 		 */
2682 		getmicrotime(&cur_time);
2683 		timeout = cur_time.tv_sec - start_time.tv_sec;
2684 		if (timeout > 60) {
2685 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
2686 			error = ETIMEDOUT;
2687 			break;
2688 		}
2689 	}
2690 
2691 	if (error) {
2692 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
2693 		rc = mpr_reinit(sc);
2694 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2695 		    "failed");
2696 	}
2697 	return (error);
2698 }
2699 
2700 /*
2701  * The MPT driver had a verbose interface for config pages.  In this driver,
2702  * reduce it to much simpler terms, similar to the Linux driver.
2703  */
2704 int
2705 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
2706 {
2707 	MPI2_CONFIG_REQUEST *req;
2708 	struct mpr_command *cm;
2709 	int error;
2710 
2711 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
2712 		return (EBUSY);
2713 	}
2714 
2715 	cm = mpr_alloc_command(sc);
2716 	if (cm == NULL) {
2717 		return (EBUSY);
2718 	}
2719 
2720 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2721 	req->Function = MPI2_FUNCTION_CONFIG;
2722 	req->Action = params->action;
2723 	req->SGLFlags = 0;
2724 	req->ChainOffset = 0;
2725 	req->PageAddress = params->page_address;
2726 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2727 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2728 
2729 		hdr = &params->hdr.Ext;
2730 		req->ExtPageType = hdr->ExtPageType;
2731 		req->ExtPageLength = hdr->ExtPageLength;
2732 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2733 		req->Header.PageLength = 0; /* Must be set to zero */
2734 		req->Header.PageNumber = hdr->PageNumber;
2735 		req->Header.PageVersion = hdr->PageVersion;
2736 	} else {
2737 		MPI2_CONFIG_PAGE_HEADER *hdr;
2738 
2739 		hdr = &params->hdr.Struct;
2740 		req->Header.PageType = hdr->PageType;
2741 		req->Header.PageNumber = hdr->PageNumber;
2742 		req->Header.PageLength = hdr->PageLength;
2743 		req->Header.PageVersion = hdr->PageVersion;
2744 	}
2745 
2746 	cm->cm_data = params->buffer;
2747 	cm->cm_length = params->length;
2748 	if (cm->cm_data != NULL) {
2749 		cm->cm_sge = &req->PageBufferSGE;
2750 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2751 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
2752 	} else
2753 		cm->cm_sge = NULL;
2754 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2755 
2756 	cm->cm_complete_data = params;
2757 	if (params->callback != NULL) {
2758 		cm->cm_complete = mpr_config_complete;
2759 		return (mpr_map_command(sc, cm));
2760 	} else {
2761 		error = mpr_wait_command(sc, cm, 0, CAN_SLEEP);
2762 		if (error) {
2763 			mpr_dprint(sc, MPR_FAULT,
2764 			    "Error %d reading config page\n", error);
2765 			mpr_free_command(sc, cm);
2766 			return (error);
2767 		}
2768 		mpr_config_complete(sc, cm);
2769 	}
2770 
2771 	return (0);
2772 }
2773 
2774 int
2775 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
2776 {
2777 	return (EINVAL);
2778 }
2779 
2780 static void
2781 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
2782 {
2783 	MPI2_CONFIG_REPLY *reply;
2784 	struct mpr_config_params *params;
2785 
2786 	MPR_FUNCTRACE(sc);
2787 	params = cm->cm_complete_data;
2788 
2789 	if (cm->cm_data != NULL) {
2790 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2791 		    BUS_DMASYNC_POSTREAD);
2792 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2793 	}
2794 
2795 	/*
2796 	 * XXX KDM need to do more error recovery?  This results in the
2797 	 * device in question not getting probed.
2798 	 */
2799 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
2800 		params->status = MPI2_IOCSTATUS_BUSY;
2801 		goto done;
2802 	}
2803 
2804 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2805 	if (reply == NULL) {
2806 		params->status = MPI2_IOCSTATUS_BUSY;
2807 		goto done;
2808 	}
2809 	params->status = reply->IOCStatus;
2810 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2811 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2812 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2813 		params->hdr.Ext.PageType = reply->Header.PageType;
2814 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
2815 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
2816 	} else {
2817 		params->hdr.Struct.PageType = reply->Header.PageType;
2818 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2819 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2820 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2821 	}
2822 
2823 done:
2824 	mpr_free_command(sc, cm);
2825 	if (params->callback != NULL)
2826 		params->callback(sc, params);
2827 
2828 	return;
2829 }
2830