xref: /freebsd/sys/dev/mps/mps.c (revision d37eb51047221dc3322b34db1038ff3aa533883f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2009 Yahoo! Inc.
5  * Copyright (c) 2011-2015 LSI Corp.
6  * Copyright (c) 2013-2015 Avago Technologies
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
31  */
32 
33 #include <sys/cdefs.h>
34 /* Communications core for Avago Technologies (LSI) MPT2 */
35 
36 /* TODO Move headers to mpsvar */
37 #include <sys/types.h>
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/selinfo.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/module.h>
45 #include <sys/bus.h>
46 #include <sys/conf.h>
47 #include <sys/bio.h>
48 #include <sys/malloc.h>
49 #include <sys/uio.h>
50 #include <sys/sysctl.h>
51 #include <sys/smp.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 #include <sys/sbuf.h>
58 #include <sys/priv.h>
59 
60 #include <machine/bus.h>
61 #include <machine/resource.h>
62 #include <sys/rman.h>
63 #include <sys/proc.h>
64 
65 #include <dev/pci/pcivar.h>
66 
67 #include <cam/cam.h>
68 #include <cam/scsi/scsi_all.h>
69 
70 #include <dev/mps/mpi/mpi2_type.h>
71 #include <dev/mps/mpi/mpi2.h>
72 #include <dev/mps/mpi/mpi2_ioc.h>
73 #include <dev/mps/mpi/mpi2_sas.h>
74 #include <dev/mps/mpi/mpi2_cnfg.h>
75 #include <dev/mps/mpi/mpi2_init.h>
76 #include <dev/mps/mpi/mpi2_tool.h>
77 #include <dev/mps/mps_ioctl.h>
78 #include <dev/mps/mpsvar.h>
79 #include <dev/mps/mps_table.h>
80 
81 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
82 static int mps_init_queues(struct mps_softc *sc);
83 static void mps_resize_queues(struct mps_softc *sc);
84 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
85 static int mps_transition_operational(struct mps_softc *sc);
86 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
87 static void mps_iocfacts_free(struct mps_softc *sc);
88 static void mps_startup(void *arg);
89 static int mps_send_iocinit(struct mps_softc *sc);
90 static int mps_alloc_queues(struct mps_softc *sc);
91 static int mps_alloc_hw_queues(struct mps_softc *sc);
92 static int mps_alloc_replies(struct mps_softc *sc);
93 static int mps_alloc_requests(struct mps_softc *sc);
94 static int mps_attach_log(struct mps_softc *sc);
95 static __inline void mps_complete_command(struct mps_softc *sc,
96     struct mps_command *cm);
97 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
98     MPI2_EVENT_NOTIFICATION_REPLY *reply);
99 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
100 static void mps_periodic(void *);
101 static int mps_reregister_events(struct mps_softc *sc);
102 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
103 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
104 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
105 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
106 static int mps_dump_reqs(SYSCTL_HANDLER_ARGS);
107 static void mps_parse_debug(struct mps_softc *sc, char *list);
108 
109 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
110     "MPS Driver Parameters");
111 
112 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
113 MALLOC_DECLARE(M_MPSUSER);
114 
115 /*
116  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
117  * any state and back to its initialization state machine.
118  */
119 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
120 
121 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
122  * Compiler only support unint64_t to be passed as argument.
123  * Otherwise it will throw below error
124  * "aggregate value used where an integer was expected"
125  */
126 
127 typedef union {
128         u64 word;
129         struct {
130                 u32 low;
131                 u32 high;
132         } u;
133 } request_descriptor_t;
134 
135 /* Rate limit chain-fail messages to 1 per minute */
136 static struct timeval mps_chainfail_interval = { 60, 0 };
137 
138 /*
139  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
140  * If this function is called from process context, it can sleep
141  * and there is no harm to sleep, in case if this fuction is called
142  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
143  * based on sleep flags driver will call either msleep, pause or DELAY.
144  * msleep and pause are of same variant, but pause is used when mps_mtx
145  * is not hold by driver.
146  *
147  */
148 static int
149 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
150 {
151 	uint32_t reg;
152 	int i, error, tries = 0;
153 	uint8_t first_wait_done = FALSE;
154 
155 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
156 
157 	/* Clear any pending interrupts */
158 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
159 
160 	/*
161 	 * Force NO_SLEEP for threads prohibited to sleep
162  	 * e.a Thread from interrupt handler are prohibited to sleep.
163  	 */
164 	if (curthread->td_no_sleeping != 0)
165 		sleep_flag = NO_SLEEP;
166 
167 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
168 
169 	/* Push the magic sequence */
170 	error = ETIMEDOUT;
171 	while (tries++ < 20) {
172 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
173 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
174 			    mpt2_reset_magic[i]);
175 		/* wait 100 msec */
176 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
177 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
178 			    "mpsdiag", hz/10);
179 		else if (sleep_flag == CAN_SLEEP)
180 			pause("mpsdiag", hz/10);
181 		else
182 			DELAY(100 * 1000);
183 
184 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
185 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
186 			error = 0;
187 			break;
188 		}
189 	}
190 	if (error) {
191 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
192 		    error);
193 		return (error);
194 	}
195 
196 	/* Send the actual reset.  XXX need to refresh the reg? */
197 	reg |= MPI2_DIAG_RESET_ADAPTER;
198 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
199 		reg);
200 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
201 
202 	/* Wait up to 300 seconds in 50ms intervals */
203 	error = ETIMEDOUT;
204 	for (i = 0; i < 6000; i++) {
205 		/*
206 		 * Wait 50 msec. If this is the first time through, wait 256
207 		 * msec to satisfy Diag Reset timing requirements.
208 		 */
209 		if (first_wait_done) {
210 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
211 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
212 				    "mpsdiag", hz/20);
213 			else if (sleep_flag == CAN_SLEEP)
214 				pause("mpsdiag", hz/20);
215 			else
216 				DELAY(50 * 1000);
217 		} else {
218 			DELAY(256 * 1000);
219 			first_wait_done = TRUE;
220 		}
221 		/*
222 		 * Check for the RESET_ADAPTER bit to be cleared first, then
223 		 * wait for the RESET state to be cleared, which takes a little
224 		 * longer.
225 		 */
226 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
227 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
228 			continue;
229 		}
230 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
231 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
232 			error = 0;
233 			break;
234 		}
235 	}
236 	if (error) {
237 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
238 		    error);
239 		return (error);
240 	}
241 
242 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
243 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
244 
245 	return (0);
246 }
247 
248 static int
249 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
250 {
251 	int error;
252 
253 	MPS_FUNCTRACE(sc);
254 
255 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
256 
257 	error = 0;
258 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
259 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
260 	    MPI2_DOORBELL_FUNCTION_SHIFT);
261 
262 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
263 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
264 		    "Doorbell handshake failed\n");
265 		error = ETIMEDOUT;
266 	}
267 
268 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
269 	return (error);
270 }
271 
272 static int
273 mps_transition_ready(struct mps_softc *sc)
274 {
275 	uint32_t reg, state;
276 	int error, tries = 0;
277 	int sleep_flags;
278 
279 	MPS_FUNCTRACE(sc);
280 	/* If we are in attach call, do not sleep */
281 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
282 					? CAN_SLEEP:NO_SLEEP;
283 	error = 0;
284 
285 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
286 	   __func__, sleep_flags);
287 
288 	while (tries++ < 1200) {
289 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
290 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
291 
292 		/*
293 		 * Ensure the IOC is ready to talk.  If it's not, try
294 		 * resetting it.
295 		 */
296 		if (reg & MPI2_DOORBELL_USED) {
297 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
298 			    "reset\n");
299 			mps_diag_reset(sc, sleep_flags);
300 			DELAY(50000);
301 			continue;
302 		}
303 
304 		/* Is the adapter owned by another peer? */
305 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
306 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
307 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
308 			    "control of another peer host, aborting "
309 			    "initialization.\n");
310 			error = ENXIO;
311 			break;
312 		}
313 
314 		state = reg & MPI2_IOC_STATE_MASK;
315 		if (state == MPI2_IOC_STATE_READY) {
316 			/* Ready to go! */
317 			error = 0;
318 			break;
319 		} else if (state == MPI2_IOC_STATE_FAULT) {
320 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
321 			    "state 0x%x, resetting\n",
322 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
323 			mps_diag_reset(sc, sleep_flags);
324 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
325 			/* Need to take ownership */
326 			mps_message_unit_reset(sc, sleep_flags);
327 		} else if (state == MPI2_IOC_STATE_RESET) {
328 			/* Wait a bit, IOC might be in transition */
329 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
330 			    "IOC in unexpected reset state\n");
331 		} else {
332 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
333 			    "IOC in unknown state 0x%x\n", state);
334 			error = EINVAL;
335 			break;
336 		}
337 
338 		/* Wait 50ms for things to settle down. */
339 		DELAY(50000);
340 	}
341 
342 	if (error)
343 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
344 		    "Cannot transition IOC to ready\n");
345 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
346 
347 	return (error);
348 }
349 
350 static int
351 mps_transition_operational(struct mps_softc *sc)
352 {
353 	uint32_t reg, state;
354 	int error;
355 
356 	MPS_FUNCTRACE(sc);
357 
358 	error = 0;
359 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
360 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
361 
362 	state = reg & MPI2_IOC_STATE_MASK;
363 	if (state != MPI2_IOC_STATE_READY) {
364 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
365 		if ((error = mps_transition_ready(sc)) != 0) {
366 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
367 			    "failed to transition ready, exit\n");
368 			return (error);
369 		}
370 	}
371 
372 	error = mps_send_iocinit(sc);
373 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
374 
375 	return (error);
376 }
377 
378 static void
379 mps_resize_queues(struct mps_softc *sc)
380 {
381 	u_int reqcr, prireqcr, maxio, sges_per_frame;
382 
383 	/*
384 	 * Size the queues. Since the reply queues always need one free
385 	 * entry, we'll deduct one reply message here.  The LSI documents
386 	 * suggest instead to add a count to the request queue, but I think
387 	 * that it's better to deduct from reply queue.
388 	 */
389 	prireqcr = MAX(1, sc->max_prireqframes);
390 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
391 
392 	reqcr = MAX(2, sc->max_reqframes);
393 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
394 
395 	sc->num_reqs = prireqcr + reqcr;
396 	sc->num_prireqs = prireqcr;
397 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
398 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
399 
400 	/* Store the request frame size in bytes rather than as 32bit words */
401 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
402 
403 	/*
404 	 * Max IO Size is Page Size * the following:
405 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
406 	 * + 1 for no chain needed in last frame
407 	 *
408 	 * If user suggests a Max IO size to use, use the smaller of the
409 	 * user's value and the calculated value as long as the user's
410 	 * value is larger than 0. The user's value is in pages.
411 	 */
412 	sges_per_frame = sc->reqframesz / sizeof(MPI2_SGE_SIMPLE64) - 1;
413 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
414 
415 	/*
416 	 * If I/O size limitation requested, then use it and pass up to CAM.
417 	 * If not, use maxphys as an optimization hint, but report HW limit.
418 	 */
419 	if (sc->max_io_pages > 0) {
420 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
421 		sc->maxio = maxio;
422 	} else {
423 		sc->maxio = maxio;
424 		maxio = min(maxio, maxphys);
425 	}
426 
427 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
428 	    sges_per_frame * reqcr;
429 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
430 		sc->num_chains = sc->max_chains;
431 
432 	/*
433 	 * Figure out the number of MSIx-based queues.  If the firmware or
434 	 * user has done something crazy and not allowed enough credit for
435 	 * the queues to be useful then don't enable multi-queue.
436 	 */
437 	if (sc->facts->MaxMSIxVectors < 2)
438 		sc->msi_msgs = 1;
439 
440 	if (sc->msi_msgs > 1) {
441 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
442 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
443 		if (sc->num_reqs / sc->msi_msgs < 2)
444 			sc->msi_msgs = 1;
445 	}
446 
447 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
448 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
449 }
450 
451 /*
452  * This is called during attach and when re-initializing due to a Diag Reset.
453  * IOC Facts is used to allocate many of the structures needed by the driver.
454  * If called from attach, de-allocation is not required because the driver has
455  * not allocated any structures yet, but if called from a Diag Reset, previously
456  * allocated structures based on IOC Facts will need to be freed and re-
457  * allocated bases on the latest IOC Facts.
458  */
459 static int
460 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
461 {
462 	int error;
463 	Mpi2IOCFactsReply_t saved_facts;
464 	uint8_t saved_mode, reallocating;
465 
466 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
467 
468 	/* Save old IOC Facts and then only reallocate if Facts have changed */
469 	if (!attaching) {
470 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
471 	}
472 
473 	/*
474 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
475 	 * a re-initialization and only return the error if attaching so the OS
476 	 * can handle it.
477 	 */
478 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
479 		if (attaching) {
480 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
481 			    "IOC Facts with error %d, exit\n", error);
482 			return (error);
483 		} else {
484 			panic("%s failed to get IOC Facts with error %d\n",
485 			    __func__, error);
486 		}
487 	}
488 
489 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
490 
491 	snprintf(sc->fw_version, sizeof(sc->fw_version),
492 	    "%02d.%02d.%02d.%02d",
493 	    sc->facts->FWVersion.Struct.Major,
494 	    sc->facts->FWVersion.Struct.Minor,
495 	    sc->facts->FWVersion.Struct.Unit,
496 	    sc->facts->FWVersion.Struct.Dev);
497 
498 	snprintf(sc->msg_version, sizeof(sc->msg_version), "%d.%d",
499 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MAJOR_MASK) >>
500 	    MPI2_IOCFACTS_MSGVERSION_MAJOR_SHIFT,
501 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MINOR_MASK) >>
502 	    MPI2_IOCFACTS_MSGVERSION_MINOR_SHIFT);
503 
504 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
505 	    MPS_DRIVER_VERSION);
506 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
507 	     sc->facts->IOCCapabilities,
508 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
509 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
510 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
511 
512 	/*
513 	 * If the chip doesn't support event replay then a hard reset will be
514 	 * required to trigger a full discovery.  Do the reset here then
515 	 * retransition to Ready.  A hard reset might have already been done,
516 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
517 	 * for a Diag Reset.
518 	 */
519 	if (attaching && ((sc->facts->IOCCapabilities &
520 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
521 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
522 		mps_diag_reset(sc, NO_SLEEP);
523 		if ((error = mps_transition_ready(sc)) != 0) {
524 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
525 			    "transition to ready with error %d, exit\n",
526 			    error);
527 			return (error);
528 		}
529 	}
530 
531 	/*
532 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
533 	 * changed from the previous IOC Facts, log a warning, but only if
534 	 * checking this after a Diag Reset and not during attach.
535 	 */
536 	saved_mode = sc->ir_firmware;
537 	if (sc->facts->IOCCapabilities &
538 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
539 		sc->ir_firmware = 1;
540 	if (!attaching) {
541 		if (sc->ir_firmware != saved_mode) {
542 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
543 			    "in IOC Facts does not match previous mode\n");
544 		}
545 	}
546 
547 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
548 	reallocating = FALSE;
549 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
550 
551 	if ((!attaching) &&
552 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
553 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
554 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
555 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
556 	    (saved_facts.ProductID != sc->facts->ProductID) ||
557 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
558 	    (saved_facts.IOCRequestFrameSize !=
559 	    sc->facts->IOCRequestFrameSize) ||
560 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
561 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
562 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
563 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
564 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
565 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
566 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
567 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
568 	    (saved_facts.MaxPersistentEntries !=
569 	    sc->facts->MaxPersistentEntries))) {
570 		reallocating = TRUE;
571 
572 		/* Record that we reallocated everything */
573 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
574 	}
575 
576 	/*
577 	 * Some things should be done if attaching or re-allocating after a Diag
578 	 * Reset, but are not needed after a Diag Reset if the FW has not
579 	 * changed.
580 	 */
581 	if (attaching || reallocating) {
582 		/*
583 		 * Check if controller supports FW diag buffers and set flag to
584 		 * enable each type.
585 		 */
586 		if (sc->facts->IOCCapabilities &
587 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
588 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
589 			    enabled = TRUE;
590 		if (sc->facts->IOCCapabilities &
591 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
592 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
593 			    enabled = TRUE;
594 		if (sc->facts->IOCCapabilities &
595 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
596 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
597 			    enabled = TRUE;
598 
599 		/*
600 		 * Set flag if EEDP is supported and if TLR is supported.
601 		 */
602 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
603 			sc->eedp_enabled = TRUE;
604 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
605 			sc->control_TLR = TRUE;
606 
607 		mps_resize_queues(sc);
608 
609 		/*
610 		 * Initialize all Tail Queues
611 		 */
612 		TAILQ_INIT(&sc->req_list);
613 		TAILQ_INIT(&sc->high_priority_req_list);
614 		TAILQ_INIT(&sc->chain_list);
615 		TAILQ_INIT(&sc->tm_list);
616 	}
617 
618 	/*
619 	 * If doing a Diag Reset and the FW is significantly different
620 	 * (reallocating will be set above in IOC Facts comparison), then all
621 	 * buffers based on the IOC Facts will need to be freed before they are
622 	 * reallocated.
623 	 */
624 	if (reallocating) {
625 		mps_iocfacts_free(sc);
626 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
627 		    saved_facts.MaxVolumes);
628 	}
629 
630 	/*
631 	 * Any deallocation has been completed.  Now start reallocating
632 	 * if needed.  Will only need to reallocate if attaching or if the new
633 	 * IOC Facts are different from the previous IOC Facts after a Diag
634 	 * Reset. Targets have already been allocated above if needed.
635 	 */
636 	error = 0;
637 	while (attaching || reallocating) {
638 		if ((error = mps_alloc_hw_queues(sc)) != 0)
639 			break;
640 		if ((error = mps_alloc_replies(sc)) != 0)
641 			break;
642 		if ((error = mps_alloc_requests(sc)) != 0)
643 			break;
644 		if ((error = mps_alloc_queues(sc)) != 0)
645 			break;
646 
647 		break;
648 	}
649 	if (error) {
650 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
651 		    "Failed to alloc queues with error %d\n", error);
652 		mps_free(sc);
653 		return (error);
654 	}
655 
656 	/* Always initialize the queues */
657 	bzero(sc->free_queue, sc->fqdepth * 4);
658 	mps_init_queues(sc);
659 
660 	/*
661 	 * Always get the chip out of the reset state, but only panic if not
662 	 * attaching.  If attaching and there is an error, that is handled by
663 	 * the OS.
664 	 */
665 	error = mps_transition_operational(sc);
666 	if (error != 0) {
667 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
668 		    "transition to operational with error %d\n", error);
669 		mps_free(sc);
670 		return (error);
671 	}
672 
673 	/*
674 	 * Finish the queue initialization.
675 	 * These are set here instead of in mps_init_queues() because the
676 	 * IOC resets these values during the state transition in
677 	 * mps_transition_operational().  The free index is set to 1
678 	 * because the corresponding index in the IOC is set to 0, and the
679 	 * IOC treats the queues as full if both are set to the same value.
680 	 * Hence the reason that the queue can't hold all of the possible
681 	 * replies.
682 	 */
683 	sc->replypostindex = 0;
684 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
685 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
686 
687 	/*
688 	 * Attach the subsystems so they can prepare their event masks.
689 	 * XXX Should be dynamic so that IM/IR and user modules can attach
690 	 */
691 	error = 0;
692 	while (attaching) {
693 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
694 		if ((error = mps_attach_log(sc)) != 0)
695 			break;
696 		if ((error = mps_attach_sas(sc)) != 0)
697 			break;
698 		if ((error = mps_attach_user(sc)) != 0)
699 			break;
700 		break;
701 	}
702 	if (error) {
703 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
704 		    "subsystems: error %d\n", error);
705 		mps_free(sc);
706 		return (error);
707 	}
708 
709 	/*
710 	 * XXX If the number of MSI-X vectors changes during re-init, this
711 	 * won't see it and adjust.
712 	 */
713 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
714 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
715 		    "interrupts\n");
716 		mps_free(sc);
717 		return (error);
718 	}
719 
720 	/*
721 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
722 	 * reset it after a Diag Reset, just in case.
723 	 */
724 	sc->WD_available = FALSE;
725 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
726 		sc->WD_available = TRUE;
727 
728 	return (error);
729 }
730 
731 /*
732  * This is called if memory is being free (during detach for example) and when
733  * buffers need to be reallocated due to a Diag Reset.
734  */
735 static void
736 mps_iocfacts_free(struct mps_softc *sc)
737 {
738 	struct mps_command *cm;
739 	int i;
740 
741 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
742 
743 	if (sc->free_busaddr != 0)
744 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
745 	if (sc->free_queue != NULL)
746 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
747 		    sc->queues_map);
748 	if (sc->queues_dmat != NULL)
749 		bus_dma_tag_destroy(sc->queues_dmat);
750 
751 	if (sc->chain_frames != NULL) {
752 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
753 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
754 		    sc->chain_map);
755 	}
756 	if (sc->chain_dmat != NULL)
757 		bus_dma_tag_destroy(sc->chain_dmat);
758 
759 	if (sc->sense_busaddr != 0)
760 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
761 	if (sc->sense_frames != NULL)
762 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
763 		    sc->sense_map);
764 	if (sc->sense_dmat != NULL)
765 		bus_dma_tag_destroy(sc->sense_dmat);
766 
767 	if (sc->reply_busaddr != 0)
768 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
769 	if (sc->reply_frames != NULL)
770 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
771 		    sc->reply_map);
772 	if (sc->reply_dmat != NULL)
773 		bus_dma_tag_destroy(sc->reply_dmat);
774 
775 	if (sc->req_busaddr != 0)
776 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
777 	if (sc->req_frames != NULL)
778 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
779 	if (sc->req_dmat != NULL)
780 		bus_dma_tag_destroy(sc->req_dmat);
781 
782 	if (sc->chains != NULL)
783 		free(sc->chains, M_MPT2);
784 	if (sc->commands != NULL) {
785 		for (i = 1; i < sc->num_reqs; i++) {
786 			cm = &sc->commands[i];
787 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
788 		}
789 		free(sc->commands, M_MPT2);
790 	}
791 	if (sc->buffer_dmat != NULL)
792 		bus_dma_tag_destroy(sc->buffer_dmat);
793 
794 	mps_pci_free_interrupts(sc);
795 	free(sc->queues, M_MPT2);
796 	sc->queues = NULL;
797 }
798 
799 /*
800  * The terms diag reset and hard reset are used interchangeably in the MPI
801  * docs to mean resetting the controller chip.  In this code diag reset
802  * cleans everything up, and the hard reset function just sends the reset
803  * sequence to the chip.  This should probably be refactored so that every
804  * subsystem gets a reset notification of some sort, and can clean up
805  * appropriately.
806  */
807 int
808 mps_reinit(struct mps_softc *sc)
809 {
810 	int error;
811 	struct mpssas_softc *sassc;
812 
813 	sassc = sc->sassc;
814 
815 	MPS_FUNCTRACE(sc);
816 
817 	mtx_assert(&sc->mps_mtx, MA_OWNED);
818 
819 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
820 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
821 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
822 		return 0;
823 	}
824 
825 	/* make sure the completion callbacks can recognize they're getting
826 	 * a NULL cm_reply due to a reset.
827 	 */
828 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
829 
830 	/*
831 	 * Mask interrupts here.
832 	 */
833 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
834 	mps_mask_intr(sc);
835 
836 	error = mps_diag_reset(sc, CAN_SLEEP);
837 	if (error != 0) {
838 		/* XXXSL No need to panic here */
839 		panic("%s hard reset failed with error %d\n",
840 		    __func__, error);
841 	}
842 
843 	/* Restore the PCI state, including the MSI-X registers */
844 	mps_pci_restore(sc);
845 
846 	/* Give the I/O subsystem special priority to get itself prepared */
847 	mpssas_handle_reinit(sc);
848 
849 	/*
850 	 * Get IOC Facts and allocate all structures based on this information.
851 	 * The attach function will also call mps_iocfacts_allocate at startup.
852 	 * If relevant values have changed in IOC Facts, this function will free
853 	 * all of the memory based on IOC Facts and reallocate that memory.
854 	 */
855 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
856 		panic("%s IOC Facts based allocation failed with error %d\n",
857 		    __func__, error);
858 	}
859 
860 	/*
861 	 * Mapping structures will be re-allocated after getting IOC Page8, so
862 	 * free these structures here.
863 	 */
864 	mps_mapping_exit(sc);
865 
866 	/*
867 	 * The static page function currently read is IOC Page8.  Others can be
868 	 * added in future.  It's possible that the values in IOC Page8 have
869 	 * changed after a Diag Reset due to user modification, so always read
870 	 * these.  Interrupts are masked, so unmask them before getting config
871 	 * pages.
872 	 */
873 	mps_unmask_intr(sc);
874 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
875 	mps_base_static_config_pages(sc);
876 
877 	/*
878 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
879 	 * mapping tables.
880 	 */
881 	mps_mapping_initialize(sc);
882 
883 	/*
884 	 * Restart will reload the event masks clobbered by the reset, and
885 	 * then enable the port.
886 	 */
887 	mps_reregister_events(sc);
888 
889 	/* the end of discovery will release the simq, so we're done. */
890 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
891 	    sc, sc->replypostindex, sc->replyfreeindex);
892 
893 	mpssas_release_simq_reinit(sassc);
894 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
895 
896 	return 0;
897 }
898 
899 /* Wait for the chip to ACK a word that we've put into its FIFO
900  * Wait for <timeout> seconds. In single loop wait for busy loop
901  * for 500 microseconds.
902  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
903  * */
904 static int
905 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
906 {
907 
908 	u32 cntdn, count;
909 	u32 int_status;
910 	u32 doorbell;
911 
912 	count = 0;
913 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
914 	do {
915 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
916 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
917 			mps_dprint(sc, MPS_TRACE,
918 			"%s: successful count(%d), timeout(%d)\n",
919 			__func__, count, timeout);
920 		return 0;
921 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
922 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
923 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
924 				MPI2_IOC_STATE_FAULT) {
925 				mps_dprint(sc, MPS_FAULT,
926 					"fault_state(0x%04x)!\n", doorbell);
927 				return (EFAULT);
928 			}
929 		} else if (int_status == 0xFFFFFFFF)
930 			goto out;
931 
932 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
933 		* 0.5 milisecond */
934 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
935 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
936 			"mpsdba", hz/1000);
937 		else if (sleep_flag == CAN_SLEEP)
938 			pause("mpsdba", hz/1000);
939 		else
940 			DELAY(500);
941 		count++;
942 	} while (--cntdn);
943 
944 	out:
945 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
946 		"int_status(%x)!\n", __func__, count, int_status);
947 	return (ETIMEDOUT);
948 
949 }
950 
951 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
952 static int
953 mps_wait_db_int(struct mps_softc *sc)
954 {
955 	int retry;
956 
957 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
958 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
959 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
960 			return (0);
961 		DELAY(2000);
962 	}
963 	return (ETIMEDOUT);
964 }
965 
966 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
967 static int
968 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
969     int req_sz, int reply_sz, int timeout)
970 {
971 	uint32_t *data32;
972 	uint16_t *data16;
973 	int i, count, ioc_sz, residual;
974 	int sleep_flags = CAN_SLEEP;
975 
976 	if (curthread->td_no_sleeping != 0)
977 		sleep_flags = NO_SLEEP;
978 
979 	/* Step 1 */
980 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
981 
982 	/* Step 2 */
983 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
984 		return (EBUSY);
985 
986 	/* Step 3
987 	 * Announce that a message is coming through the doorbell.  Messages
988 	 * are pushed at 32bit words, so round up if needed.
989 	 */
990 	count = (req_sz + 3) / 4;
991 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
992 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
993 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
994 
995 	/* Step 4 */
996 	if (mps_wait_db_int(sc) ||
997 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
998 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
999 		return (ENXIO);
1000 	}
1001 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1002 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
1003 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
1004 		return (ENXIO);
1005 	}
1006 
1007 	/* Step 5 */
1008 	/* Clock out the message data synchronously in 32-bit dwords*/
1009 	data32 = (uint32_t *)req;
1010 	for (i = 0; i < count; i++) {
1011 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1012 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
1013 			mps_dprint(sc, MPS_FAULT,
1014 			    "Timeout while writing doorbell\n");
1015 			return (ENXIO);
1016 		}
1017 	}
1018 
1019 	/* Step 6 */
1020 	/* Clock in the reply in 16-bit words.  The total length of the
1021 	 * message is always in the 4th byte, so clock out the first 2 words
1022 	 * manually, then loop the rest.
1023 	 */
1024 	data16 = (uint16_t *)reply;
1025 	if (mps_wait_db_int(sc) != 0) {
1026 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
1027 		return (ENXIO);
1028 	}
1029 	data16[0] =
1030 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1031 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1032 	if (mps_wait_db_int(sc) != 0) {
1033 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
1034 		return (ENXIO);
1035 	}
1036 	data16[1] =
1037 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1038 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1039 
1040 	/* Number of 32bit words in the message */
1041 	ioc_sz = reply->MsgLength;
1042 
1043 	/*
1044 	 * Figure out how many 16bit words to clock in without overrunning.
1045 	 * The precision loss with dividing reply_sz can safely be
1046 	 * ignored because the messages can only be multiples of 32bits.
1047 	 */
1048 	residual = 0;
1049 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1050 	if (count < ioc_sz * 2) {
1051 		residual = ioc_sz * 2 - count;
1052 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1053 		    "residual message words\n", residual);
1054 	}
1055 
1056 	for (i = 2; i < count; i++) {
1057 		if (mps_wait_db_int(sc) != 0) {
1058 			mps_dprint(sc, MPS_FAULT,
1059 			    "Timeout reading doorbell %d\n", i);
1060 			return (ENXIO);
1061 		}
1062 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1063 		    MPI2_DOORBELL_DATA_MASK;
1064 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1065 	}
1066 
1067 	/*
1068 	 * Pull out residual words that won't fit into the provided buffer.
1069 	 * This keeps the chip from hanging due to a driver programming
1070 	 * error.
1071 	 */
1072 	while (residual--) {
1073 		if (mps_wait_db_int(sc) != 0) {
1074 			mps_dprint(sc, MPS_FAULT,
1075 			    "Timeout reading doorbell\n");
1076 			return (ENXIO);
1077 		}
1078 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1079 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1080 	}
1081 
1082 	/* Step 7 */
1083 	if (mps_wait_db_int(sc) != 0) {
1084 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1085 		return (ENXIO);
1086 	}
1087 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1088 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1089 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1090 
1091 	return (0);
1092 }
1093 
1094 static void
1095 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1096 {
1097 	request_descriptor_t rd;
1098 	MPS_FUNCTRACE(sc);
1099 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1100 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1101 
1102 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1103 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1104 
1105 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1106 		sc->io_cmds_highwater++;
1107 	rd.u.low = cm->cm_desc.Words.Low;
1108 	rd.u.high = cm->cm_desc.Words.High;
1109 	rd.word = htole64(rd.word);
1110 
1111 	KASSERT(cm->cm_state == MPS_CM_STATE_BUSY,
1112 	    ("command not busy, state = %u\n", cm->cm_state));
1113 	cm->cm_state = MPS_CM_STATE_INQUEUE;
1114 
1115 	/* TODO-We may need to make below regwrite atomic */
1116 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1117 	    rd.u.low);
1118 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1119 	    rd.u.high);
1120 }
1121 
1122 /*
1123  * Just the FACTS, ma'am.
1124  */
1125 static int
1126 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1127 {
1128 	MPI2_DEFAULT_REPLY *reply;
1129 	MPI2_IOC_FACTS_REQUEST request;
1130 	int error, req_sz, reply_sz;
1131 
1132 	MPS_FUNCTRACE(sc);
1133 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1134 
1135 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1136 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1137 	reply = (MPI2_DEFAULT_REPLY *)facts;
1138 
1139 	bzero(&request, req_sz);
1140 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1141 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1142 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1143 
1144 	return (error);
1145 }
1146 
1147 static int
1148 mps_send_iocinit(struct mps_softc *sc)
1149 {
1150 	MPI2_IOC_INIT_REQUEST	init;
1151 	MPI2_DEFAULT_REPLY	reply;
1152 	int req_sz, reply_sz, error;
1153 	struct timeval now;
1154 	uint64_t time_in_msec;
1155 
1156 	MPS_FUNCTRACE(sc);
1157 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1158 
1159 	/* Do a quick sanity check on proper initialization */
1160 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1161 	    || (sc->replyframesz == 0)) {
1162 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1163 		    "Driver not fully initialized for IOCInit\n");
1164 		return (EINVAL);
1165 	}
1166 
1167 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1168 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1169 	bzero(&init, req_sz);
1170 	bzero(&reply, reply_sz);
1171 
1172 	/*
1173 	 * Fill in the init block.  Note that most addresses are
1174 	 * deliberately in the lower 32bits of memory.  This is a micro-
1175 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1176 	 */
1177 	init.Function = MPI2_FUNCTION_IOC_INIT;
1178 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1179 	init.MsgVersion = htole16(MPI2_VERSION);
1180 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1181 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1182 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1183 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1184 	init.SenseBufferAddressHigh = 0;
1185 	init.SystemReplyAddressHigh = 0;
1186 	init.SystemRequestFrameBaseAddress.High = 0;
1187 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1188 	init.ReplyDescriptorPostQueueAddress.High = 0;
1189 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1190 	init.ReplyFreeQueueAddress.High = 0;
1191 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1192 	getmicrotime(&now);
1193 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1194 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1195 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1196 
1197 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1198 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1199 		error = ENXIO;
1200 
1201 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1202 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1203 	return (error);
1204 }
1205 
1206 void
1207 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1208 {
1209 	bus_addr_t *addr;
1210 
1211 	addr = arg;
1212 	*addr = segs[0].ds_addr;
1213 }
1214 
1215 void
1216 mps_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1217 {
1218 	struct mps_busdma_context *ctx;
1219 	int need_unload, need_free;
1220 
1221 	ctx = (struct mps_busdma_context *)arg;
1222 	need_unload = 0;
1223 	need_free = 0;
1224 
1225 	mps_lock(ctx->softc);
1226 	ctx->error = error;
1227 	ctx->completed = 1;
1228 	if ((error == 0) && (ctx->abandoned == 0)) {
1229 		*ctx->addr = segs[0].ds_addr;
1230 	} else {
1231 		if (nsegs != 0)
1232 			need_unload = 1;
1233 		if (ctx->abandoned != 0)
1234 			need_free = 1;
1235 	}
1236 	if (need_free == 0)
1237 		wakeup(ctx);
1238 
1239 	mps_unlock(ctx->softc);
1240 
1241 	if (need_unload != 0) {
1242 		bus_dmamap_unload(ctx->buffer_dmat,
1243 				  ctx->buffer_dmamap);
1244 		*ctx->addr = 0;
1245 	}
1246 
1247 	if (need_free != 0)
1248 		free(ctx, M_MPSUSER);
1249 }
1250 
1251 static int
1252 mps_alloc_queues(struct mps_softc *sc)
1253 {
1254 	struct mps_queue *q;
1255 	u_int nq, i;
1256 
1257 	nq = sc->msi_msgs;
1258 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1259 
1260 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1261 	    M_NOWAIT|M_ZERO);
1262 	if (sc->queues == NULL)
1263 		return (ENOMEM);
1264 
1265 	for (i = 0; i < nq; i++) {
1266 		q = &sc->queues[i];
1267 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1268 		q->sc = sc;
1269 		q->qnum = i;
1270 	}
1271 
1272 	return (0);
1273 }
1274 
1275 static int
1276 mps_alloc_hw_queues(struct mps_softc *sc)
1277 {
1278 	bus_dma_template_t t;
1279 	bus_addr_t queues_busaddr;
1280 	uint8_t *queues;
1281 	int qsize, fqsize, pqsize;
1282 
1283 	/*
1284 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1285 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1286 	 * This queue supplies fresh reply frames for the firmware to use.
1287 	 *
1288 	 * The reply descriptor post queue contains 8 byte entries in
1289 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1290 	 * contains filled-in reply frames sent from the firmware to the host.
1291 	 *
1292 	 * These two queues are allocated together for simplicity.
1293 	 */
1294 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1295 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1296 	fqsize= sc->fqdepth * 4;
1297 	pqsize = sc->pqdepth * 8;
1298 	qsize = fqsize + pqsize;
1299 
1300 	bus_dma_template_init(&t, sc->mps_parent_dmat);
1301 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(qsize),
1302 	    BD_MAXSEGSIZE(qsize), BD_NSEGMENTS(1),
1303 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1304 	if (bus_dma_template_tag(&t, &sc->queues_dmat)) {
1305 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1306 		return (ENOMEM);
1307         }
1308         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1309 	    &sc->queues_map)) {
1310 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1311 		return (ENOMEM);
1312         }
1313         bzero(queues, qsize);
1314         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1315 	    mps_memaddr_cb, &queues_busaddr, 0);
1316 
1317 	sc->free_queue = (uint32_t *)queues;
1318 	sc->free_busaddr = queues_busaddr;
1319 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1320 	sc->post_busaddr = queues_busaddr + fqsize;
1321 	mps_dprint(sc, MPS_INIT, "free queue busaddr= %#016jx size= %d\n",
1322 	    (uintmax_t)sc->free_busaddr, fqsize);
1323 	mps_dprint(sc, MPS_INIT, "reply queue busaddr= %#016jx size= %d\n",
1324 	    (uintmax_t)sc->post_busaddr, pqsize);
1325 
1326 	return (0);
1327 }
1328 
1329 static int
1330 mps_alloc_replies(struct mps_softc *sc)
1331 {
1332 	bus_dma_template_t t;
1333 	int rsize, num_replies;
1334 
1335 	/* Store the reply frame size in bytes rather than as 32bit words */
1336 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1337 
1338 	/*
1339 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1340 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1341 	 * replies can be used at once.
1342 	 */
1343 	num_replies = max(sc->fqdepth, sc->num_replies);
1344 
1345 	rsize = sc->replyframesz * num_replies;
1346 	bus_dma_template_init(&t, sc->mps_parent_dmat);
1347 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize),
1348 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1349 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1350 	if (bus_dma_template_tag(&t, &sc->reply_dmat)) {
1351 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1352 		return (ENOMEM);
1353         }
1354         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1355 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1356 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1357 		return (ENOMEM);
1358         }
1359         bzero(sc->reply_frames, rsize);
1360         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1361 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1362 
1363 	mps_dprint(sc, MPS_INIT, "reply frames busaddr= %#016jx size= %d\n",
1364 	    (uintmax_t)sc->reply_busaddr, rsize);
1365 
1366 	return (0);
1367 }
1368 
1369 static void
1370 mps_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1371 {
1372 	struct mps_softc *sc = arg;
1373 	struct mps_chain *chain;
1374 	bus_size_t bo;
1375 	int i, o, s;
1376 
1377 	if (error != 0)
1378 		return;
1379 
1380 	for (i = 0, o = 0, s = 0; s < nsegs; s++) {
1381 		KASSERT(segs[s].ds_addr + segs[s].ds_len - 1 <= BUS_SPACE_MAXADDR_32BIT,
1382 		    ("mps: Bad segment address %#jx len %#jx\n", (uintmax_t)segs[s].ds_addr,
1383 			(uintmax_t)segs[s].ds_len));
1384 		for (bo = 0; bo + sc->reqframesz <= segs[s].ds_len;
1385 		    bo += sc->reqframesz) {
1386 			chain = &sc->chains[i++];
1387 			chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o);
1388 			chain->chain_busaddr = segs[s].ds_addr + bo;
1389 			o += sc->reqframesz;
1390 			mps_free_chain(sc, chain);
1391 		}
1392 		if (bo != segs[s].ds_len)
1393 			o += segs[s].ds_len - bo;
1394 	}
1395 	sc->chain_free_lowwater = i;
1396 }
1397 
1398 static int
1399 mps_alloc_requests(struct mps_softc *sc)
1400 {
1401 	bus_dma_template_t t;
1402 	struct mps_command *cm;
1403 	int i, rsize, nsegs;
1404 
1405 	rsize = sc->reqframesz * sc->num_reqs;
1406 	bus_dma_template_init(&t, sc->mps_parent_dmat);
1407 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize),
1408 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1409 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1410         if (bus_dma_template_tag(&t, &sc->req_dmat)) {
1411 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1412 		return (ENOMEM);
1413         }
1414         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1415 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1416 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1417 		return (ENOMEM);
1418         }
1419         bzero(sc->req_frames, rsize);
1420         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1421 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1422 	mps_dprint(sc, MPS_INIT, "request frames busaddr= %#016jx size= %d\n",
1423 	    (uintmax_t)sc->req_busaddr, rsize);
1424 
1425 	sc->chains = malloc(sizeof(struct mps_chain) * sc->num_chains, M_MPT2,
1426 	    M_NOWAIT | M_ZERO);
1427 	if (!sc->chains) {
1428 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1429 		return (ENOMEM);
1430 	}
1431 	rsize = sc->reqframesz * sc->num_chains;
1432 	bus_dma_template_clone(&t, sc->req_dmat);
1433 	BUS_DMA_TEMPLATE_FILL(&t, BD_MAXSIZE(rsize), BD_MAXSEGSIZE(rsize),
1434 	    BD_NSEGMENTS(howmany(rsize, PAGE_SIZE)),
1435 	    BD_BOUNDARY(BUS_SPACE_MAXSIZE_32BIT+1));
1436 	if (bus_dma_template_tag(&t, &sc->chain_dmat)) {
1437 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1438 		return (ENOMEM);
1439 	}
1440 	if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1441 	    BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) {
1442 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1443 		return (ENOMEM);
1444 	}
1445 	if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames,
1446 	    rsize, mps_load_chains_cb, sc, BUS_DMA_NOWAIT)) {
1447 		mps_dprint(sc, MPS_ERROR, "Cannot load chain memory\n");
1448 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
1449 		    sc->chain_map);
1450 		return (ENOMEM);
1451 	}
1452 
1453 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1454 	bus_dma_template_clone(&t, sc->req_dmat);
1455 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(1), BD_MAXSIZE(rsize),
1456 	    BD_MAXSEGSIZE(rsize));
1457         if (bus_dma_template_tag(&t, &sc->sense_dmat)) {
1458 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1459 		return (ENOMEM);
1460         }
1461         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1462 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1463 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1464 		return (ENOMEM);
1465         }
1466         bzero(sc->sense_frames, rsize);
1467         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1468 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1469 	mps_dprint(sc, MPS_INIT, "sense frames busaddr= %#016jx size= %d\n",
1470 	    (uintmax_t)sc->sense_busaddr, rsize);
1471 
1472 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1473 	bus_dma_template_init(&t, sc->mps_parent_dmat);
1474 	BUS_DMA_TEMPLATE_FILL(&t, BD_MAXSIZE(BUS_SPACE_MAXSIZE_32BIT),
1475 	    BD_NSEGMENTS(nsegs), BD_MAXSEGSIZE(BUS_SPACE_MAXSIZE_24BIT),
1476 	    BD_FLAGS(BUS_DMA_ALLOCNOW), BD_LOCKFUNC(busdma_lock_mutex),
1477 	    BD_LOCKFUNCARG(&sc->mps_mtx),
1478 	    BD_BOUNDARY(BUS_SPACE_MAXSIZE_32BIT+1));
1479         if (bus_dma_template_tag(&t, &sc->buffer_dmat)) {
1480 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1481 		return (ENOMEM);
1482         }
1483 
1484 	/*
1485 	 * SMID 0 cannot be used as a free command per the firmware spec.
1486 	 * Just drop that command instead of risking accounting bugs.
1487 	 */
1488 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1489 	    M_MPT2, M_WAITOK | M_ZERO);
1490 	for (i = 1; i < sc->num_reqs; i++) {
1491 		cm = &sc->commands[i];
1492 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1493 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1494 		cm->cm_sense = &sc->sense_frames[i];
1495 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1496 		cm->cm_desc.Default.SMID = i;
1497 		cm->cm_sc = sc;
1498 		cm->cm_state = MPS_CM_STATE_BUSY;
1499 		TAILQ_INIT(&cm->cm_chain_list);
1500 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1501 
1502 		/* XXX Is a failure here a critical problem? */
1503 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1504 			if (i <= sc->num_prireqs)
1505 				mps_free_high_priority_command(sc, cm);
1506 			else
1507 				mps_free_command(sc, cm);
1508 		else {
1509 			panic("failed to allocate command %d\n", i);
1510 			sc->num_reqs = i;
1511 			break;
1512 		}
1513 	}
1514 
1515 	return (0);
1516 }
1517 
1518 static int
1519 mps_init_queues(struct mps_softc *sc)
1520 {
1521 	int i;
1522 
1523 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1524 
1525 	/*
1526 	 * According to the spec, we need to use one less reply than we
1527 	 * have space for on the queue.  So sc->num_replies (the number we
1528 	 * use) should be less than sc->fqdepth (allocated size).
1529 	 */
1530 	if (sc->num_replies >= sc->fqdepth)
1531 		return (EINVAL);
1532 
1533 	/*
1534 	 * Initialize all of the free queue entries.
1535 	 */
1536 	for (i = 0; i < sc->fqdepth; i++)
1537 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1538 	sc->replyfreeindex = sc->num_replies;
1539 
1540 	return (0);
1541 }
1542 
1543 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1544  * Next are the global settings, if they exist.  Highest are the per-unit
1545  * settings, if they exist.
1546  */
1547 void
1548 mps_get_tunables(struct mps_softc *sc)
1549 {
1550 	char tmpstr[80], mps_debug[80];
1551 
1552 	/* XXX default to some debugging for now */
1553 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1554 	sc->disable_msix = 0;
1555 	sc->disable_msi = 0;
1556 	sc->max_msix = MPS_MSIX_MAX;
1557 	sc->max_chains = MPS_CHAIN_FRAMES;
1558 	sc->max_io_pages = MPS_MAXIO_PAGES;
1559 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1560 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1561 	sc->use_phynum = 1;
1562 	sc->max_reqframes = MPS_REQ_FRAMES;
1563 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1564 	sc->max_replyframes = MPS_REPLY_FRAMES;
1565 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1566 
1567 	/*
1568 	 * Grab the global variables.
1569 	 */
1570 	bzero(mps_debug, 80);
1571 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1572 		mps_parse_debug(sc, mps_debug);
1573 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1574 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1575 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1576 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1577 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1578 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1579 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1580 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1581 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1582 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1583 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1584 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1585 
1586 	/* Grab the unit-instance variables */
1587 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1588 	    device_get_unit(sc->mps_dev));
1589 	bzero(mps_debug, 80);
1590 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1591 		mps_parse_debug(sc, mps_debug);
1592 
1593 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1594 	    device_get_unit(sc->mps_dev));
1595 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1596 
1597 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1598 	    device_get_unit(sc->mps_dev));
1599 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1600 
1601 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1602 	    device_get_unit(sc->mps_dev));
1603 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1604 
1605 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1606 	    device_get_unit(sc->mps_dev));
1607 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1608 
1609 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1610 	    device_get_unit(sc->mps_dev));
1611 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1612 
1613 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1614 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1615 	    device_get_unit(sc->mps_dev));
1616 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1617 
1618 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1619 	    device_get_unit(sc->mps_dev));
1620 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1621 
1622 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1623 	    device_get_unit(sc->mps_dev));
1624 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1625 
1626 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1627 	    device_get_unit(sc->mps_dev));
1628 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1629 
1630 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1631 	    device_get_unit(sc->mps_dev));
1632 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1633 
1634 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1635 	    device_get_unit(sc->mps_dev));
1636 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1637 
1638 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1639 	    device_get_unit(sc->mps_dev));
1640 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1641 
1642 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1643 	    device_get_unit(sc->mps_dev));
1644 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1645 
1646 }
1647 
1648 static void
1649 mps_setup_sysctl(struct mps_softc *sc)
1650 {
1651 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1652 	struct sysctl_oid	*sysctl_tree = NULL;
1653 	char tmpstr[80], tmpstr2[80];
1654 
1655 	/*
1656 	 * Setup the sysctl variable so the user can change the debug level
1657 	 * on the fly.
1658 	 */
1659 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1660 	    device_get_unit(sc->mps_dev));
1661 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1662 
1663 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1664 	if (sysctl_ctx != NULL)
1665 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1666 
1667 	if (sysctl_tree == NULL) {
1668 		sysctl_ctx_init(&sc->sysctl_ctx);
1669 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1670 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1671 		    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr);
1672 		if (sc->sysctl_tree == NULL)
1673 			return;
1674 		sysctl_ctx = &sc->sysctl_ctx;
1675 		sysctl_tree = sc->sysctl_tree;
1676 	}
1677 
1678 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1679 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1680 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1681 
1682 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1683 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1684 	    "Disable the use of MSI-X interrupts");
1685 
1686 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1687 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1688 	    "Disable the use of MSI interrupts");
1689 
1690 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1691 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1692 	    "User-defined maximum number of MSIX queues");
1693 
1694 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1695 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1696 	    "Negotiated number of MSIX queues");
1697 
1698 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1699 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1700 	    "Total number of allocated request frames");
1701 
1702 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1703 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1704 	    "Total number of allocated high priority request frames");
1705 
1706 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1707 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1708 	    "Total number of allocated reply frames");
1709 
1710 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1711 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1712 	    "Total number of event frames allocated");
1713 
1714 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1715 	    OID_AUTO, "firmware_version", CTLFLAG_RD, sc->fw_version,
1716 	    strlen(sc->fw_version), "firmware version");
1717 
1718 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1719 	    OID_AUTO, "driver_version", CTLFLAG_RD, MPS_DRIVER_VERSION,
1720 	    strlen(MPS_DRIVER_VERSION), "driver version");
1721 
1722 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1723 	    OID_AUTO, "msg_version", CTLFLAG_RD, sc->msg_version,
1724 	    strlen(sc->msg_version), "message interface version (deprecated)");
1725 
1726 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1727 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1728 	    &sc->io_cmds_active, 0, "number of currently active commands");
1729 
1730 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1731 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1732 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1733 
1734 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1735 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1736 	    &sc->chain_free, 0, "number of free chain elements");
1737 
1738 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1739 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1740 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1741 
1742 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1743 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1744 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1745 
1746 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1747 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1748 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1749 	    "IOCFacts)");
1750 
1751 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1752 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1753 	    "enable SSU to SATA SSD/HDD at shutdown");
1754 
1755 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1756 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1757 	    &sc->chain_alloc_fail, "chain allocation failures");
1758 
1759 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1760 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1761 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1762 	    "spinup after SATA ID error");
1763 
1764 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1765 	    OID_AUTO, "mapping_table_dump",
1766 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
1767 	    mps_mapping_dump, "A", "Mapping Table Dump");
1768 
1769 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1770 	    OID_AUTO, "encl_table_dump",
1771 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
1772 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1773 
1774 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1775 	    OID_AUTO, "dump_reqs",
1776 	    CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE,
1777 	    sc, 0, mps_dump_reqs, "I", "Dump Active Requests");
1778 
1779 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1780 	    OID_AUTO, "dump_reqs_alltypes", CTLFLAG_RW,
1781 	    &sc->dump_reqs_alltypes, 0,
1782 	    "dump all request types not just inqueue");
1783 
1784 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1785 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1786 	    "Use the phy number for enumeration");
1787 }
1788 
1789 static struct mps_debug_string {
1790 	char	*name;
1791 	int	flag;
1792 } mps_debug_strings[] = {
1793 	{"info", MPS_INFO},
1794 	{"fault", MPS_FAULT},
1795 	{"event", MPS_EVENT},
1796 	{"log", MPS_LOG},
1797 	{"recovery", MPS_RECOVERY},
1798 	{"error", MPS_ERROR},
1799 	{"init", MPS_INIT},
1800 	{"xinfo", MPS_XINFO},
1801 	{"user", MPS_USER},
1802 	{"mapping", MPS_MAPPING},
1803 	{"trace", MPS_TRACE}
1804 };
1805 
1806 enum mps_debug_level_combiner {
1807 	COMB_NONE,
1808 	COMB_ADD,
1809 	COMB_SUB
1810 };
1811 
1812 static int
1813 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1814 {
1815 	struct mps_softc *sc;
1816 	struct mps_debug_string *string;
1817 	struct sbuf *sbuf;
1818 	char *buffer;
1819 	size_t sz;
1820 	int i, len, debug, error;
1821 
1822 	sc = (struct mps_softc *)arg1;
1823 
1824 	error = sysctl_wire_old_buffer(req, 0);
1825 	if (error != 0)
1826 		return (error);
1827 
1828 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1829 	debug = sc->mps_debug;
1830 
1831 	sbuf_printf(sbuf, "%#x", debug);
1832 
1833 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1834 	for (i = 0; i < sz; i++) {
1835 		string = &mps_debug_strings[i];
1836 		if (debug & string->flag)
1837 			sbuf_printf(sbuf, ",%s", string->name);
1838 	}
1839 
1840 	error = sbuf_finish(sbuf);
1841 	sbuf_delete(sbuf);
1842 
1843 	if (error || req->newptr == NULL)
1844 		return (error);
1845 
1846 	len = req->newlen - req->newidx;
1847 	if (len == 0)
1848 		return (0);
1849 
1850 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1851 	error = SYSCTL_IN(req, buffer, len);
1852 
1853 	mps_parse_debug(sc, buffer);
1854 
1855 	free(buffer, M_MPT2);
1856 	return (error);
1857 }
1858 
1859 static void
1860 mps_parse_debug(struct mps_softc *sc, char *list)
1861 {
1862 	struct mps_debug_string *string;
1863 	enum mps_debug_level_combiner op;
1864 	char *token, *endtoken;
1865 	size_t sz;
1866 	int flags, i;
1867 
1868 	if (list == NULL || *list == '\0')
1869 		return;
1870 
1871 	if (*list == '+') {
1872 		op = COMB_ADD;
1873 		list++;
1874 	} else if (*list == '-') {
1875 		op = COMB_SUB;
1876 		list++;
1877 	} else
1878 		op = COMB_NONE;
1879 	if (*list == '\0')
1880 		return;
1881 
1882 	flags = 0;
1883 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1884 	while ((token = strsep(&list, ":,")) != NULL) {
1885 		/* Handle integer flags */
1886 		flags |= strtol(token, &endtoken, 0);
1887 		if (token != endtoken)
1888 			continue;
1889 
1890 		/* Handle text flags */
1891 		for (i = 0; i < sz; i++) {
1892 			string = &mps_debug_strings[i];
1893 			if (strcasecmp(token, string->name) == 0) {
1894 				flags |= string->flag;
1895 				break;
1896 			}
1897 		}
1898 	}
1899 
1900 	switch (op) {
1901 	case COMB_NONE:
1902 		sc->mps_debug = flags;
1903 		break;
1904 	case COMB_ADD:
1905 		sc->mps_debug |= flags;
1906 		break;
1907 	case COMB_SUB:
1908 		sc->mps_debug &= (~flags);
1909 		break;
1910 	}
1911 
1912 	return;
1913 }
1914 
1915 struct mps_dumpreq_hdr {
1916 	uint32_t	smid;
1917 	uint32_t	state;
1918 	uint32_t	numframes;
1919 	uint32_t	deschi;
1920 	uint32_t	desclo;
1921 };
1922 
1923 static int
1924 mps_dump_reqs(SYSCTL_HANDLER_ARGS)
1925 {
1926 	struct mps_softc *sc;
1927 	struct mps_chain *chain, *chain1;
1928 	struct mps_command *cm;
1929 	struct mps_dumpreq_hdr hdr;
1930 	struct sbuf *sb;
1931 	uint32_t smid, state;
1932 	int i, numreqs, error = 0;
1933 
1934 	sc = (struct mps_softc *)arg1;
1935 
1936 	if ((error = priv_check(curthread, PRIV_DRIVER)) != 0) {
1937 		printf("priv check error %d\n", error);
1938 		return (error);
1939 	}
1940 
1941 	state = MPS_CM_STATE_INQUEUE;
1942 	smid = 1;
1943 	numreqs = sc->num_reqs;
1944 
1945 	if (req->newptr != NULL)
1946 		return (EINVAL);
1947 
1948 	if (smid == 0 || smid > sc->num_reqs)
1949 		return (EINVAL);
1950 	if (numreqs <= 0 || (numreqs + smid > sc->num_reqs))
1951 		numreqs = sc->num_reqs;
1952 	sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req);
1953 
1954 	/* Best effort, no locking */
1955 	for (i = smid; i < numreqs; i++) {
1956 		cm = &sc->commands[i];
1957 		if ((sc->dump_reqs_alltypes == 0) && (cm->cm_state != state))
1958 			continue;
1959 		hdr.smid = i;
1960 		hdr.state = cm->cm_state;
1961 		hdr.numframes = 1;
1962 		hdr.deschi = cm->cm_desc.Words.High;
1963 		hdr.desclo = cm->cm_desc.Words.Low;
1964 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
1965 		   chain1)
1966 			hdr.numframes++;
1967 		sbuf_bcat(sb, &hdr, sizeof(hdr));
1968 		sbuf_bcat(sb, cm->cm_req, 128);
1969 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
1970 		    chain1)
1971 			sbuf_bcat(sb, chain->chain, 128);
1972 	}
1973 
1974 	error = sbuf_finish(sb);
1975 	sbuf_delete(sb);
1976 	return (error);
1977 }
1978 
1979 int
1980 mps_attach(struct mps_softc *sc)
1981 {
1982 	int error;
1983 
1984 	MPS_FUNCTRACE(sc);
1985 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1986 
1987 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1988 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1989 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1990 	TAILQ_INIT(&sc->event_list);
1991 	timevalclear(&sc->lastfail);
1992 
1993 	if ((error = mps_transition_ready(sc)) != 0) {
1994 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1995 		    "ready\n");
1996 		return (error);
1997 	}
1998 
1999 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
2000 	    M_ZERO|M_NOWAIT);
2001 	if(!sc->facts) {
2002 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
2003 		    "exit\n");
2004 		return (ENOMEM);
2005 	}
2006 
2007 	/*
2008 	 * Get IOC Facts and allocate all structures based on this information.
2009 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
2010 	 * Facts. If relevant values have changed in IOC Facts, this function
2011 	 * will free all of the memory based on IOC Facts and reallocate that
2012 	 * memory.  If this fails, any allocated memory should already be freed.
2013 	 */
2014 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
2015 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
2016 		    "failed with error %d, exit\n", error);
2017 		return (error);
2018 	}
2019 
2020 	/* Start the periodic watchdog check on the IOC Doorbell */
2021 	mps_periodic(sc);
2022 
2023 	/*
2024 	 * The portenable will kick off discovery events that will drive the
2025 	 * rest of the initialization process.  The CAM/SAS module will
2026 	 * hold up the boot sequence until discovery is complete.
2027 	 */
2028 	sc->mps_ich.ich_func = mps_startup;
2029 	sc->mps_ich.ich_arg = sc;
2030 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
2031 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
2032 		    "Cannot establish MPS config hook\n");
2033 		error = EINVAL;
2034 	}
2035 
2036 	/*
2037 	 * Allow IR to shutdown gracefully when shutdown occurs.
2038 	 */
2039 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
2040 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
2041 
2042 	if (sc->shutdown_eh == NULL)
2043 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
2044 		    "shutdown event registration failed\n");
2045 
2046 	mps_setup_sysctl(sc);
2047 
2048 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
2049 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
2050 
2051 	return (error);
2052 }
2053 
2054 /* Run through any late-start handlers. */
2055 static void
2056 mps_startup(void *arg)
2057 {
2058 	struct mps_softc *sc;
2059 
2060 	sc = (struct mps_softc *)arg;
2061 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2062 
2063 	mps_lock(sc);
2064 	mps_unmask_intr(sc);
2065 
2066 	/* initialize device mapping tables */
2067 	mps_base_static_config_pages(sc);
2068 	mps_mapping_initialize(sc);
2069 	mpssas_startup(sc);
2070 	mps_unlock(sc);
2071 
2072 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
2073 	config_intrhook_disestablish(&sc->mps_ich);
2074 	sc->mps_ich.ich_arg = NULL;
2075 
2076 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2077 }
2078 
2079 /* Periodic watchdog.  Is called with the driver lock already held. */
2080 static void
2081 mps_periodic(void *arg)
2082 {
2083 	struct mps_softc *sc;
2084 	uint32_t db;
2085 
2086 	sc = (struct mps_softc *)arg;
2087 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
2088 		return;
2089 
2090 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
2091 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2092 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2093 		mps_reinit(sc);
2094 	}
2095 
2096 	callout_reset_sbt(&sc->periodic, MPS_PERIODIC_DELAY * SBT_1S, 0,
2097 	    mps_periodic, sc, C_PREL(1));
2098 }
2099 
2100 static void
2101 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
2102     MPI2_EVENT_NOTIFICATION_REPLY *event)
2103 {
2104 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2105 
2106 	MPS_DPRINT_EVENT(sc, generic, event);
2107 
2108 	switch (event->Event) {
2109 	case MPI2_EVENT_LOG_DATA:
2110 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2111 		if (sc->mps_debug & MPS_EVENT)
2112 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
2113 		break;
2114 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2115 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2116 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2117 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2118 		     entry->LogSequence);
2119 		break;
2120 	default:
2121 		break;
2122 	}
2123 	return;
2124 }
2125 
2126 static int
2127 mps_attach_log(struct mps_softc *sc)
2128 {
2129 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
2130 
2131 	bzero(events, 16);
2132 	setbit(events, MPI2_EVENT_LOG_DATA);
2133 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2134 
2135 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
2136 	    &sc->mps_log_eh);
2137 
2138 	return (0);
2139 }
2140 
2141 static int
2142 mps_detach_log(struct mps_softc *sc)
2143 {
2144 
2145 	if (sc->mps_log_eh != NULL)
2146 		mps_deregister_events(sc, sc->mps_log_eh);
2147 	return (0);
2148 }
2149 
2150 /*
2151  * Free all of the driver resources and detach submodules.  Should be called
2152  * without the lock held.
2153  */
2154 int
2155 mps_free(struct mps_softc *sc)
2156 {
2157 	int error;
2158 
2159 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2160 	/* Turn off the watchdog */
2161 	mps_lock(sc);
2162 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
2163 	mps_unlock(sc);
2164 	/* Lock must not be held for this */
2165 	callout_drain(&sc->periodic);
2166 	callout_drain(&sc->device_check_callout);
2167 
2168 	if (((error = mps_detach_log(sc)) != 0) ||
2169 	    ((error = mps_detach_sas(sc)) != 0)) {
2170 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2171 		    "subsystems, exit\n");
2172 		return (error);
2173 	}
2174 
2175 	mps_detach_user(sc);
2176 
2177 	/* Put the IOC back in the READY state. */
2178 	mps_lock(sc);
2179 	if ((error = mps_transition_ready(sc)) != 0) {
2180 		mps_unlock(sc);
2181 		return (error);
2182 	}
2183 	mps_unlock(sc);
2184 
2185 	if (sc->facts != NULL)
2186 		free(sc->facts, M_MPT2);
2187 
2188 	/*
2189 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2190 	 * to free these buffers too.
2191 	 */
2192 	mps_iocfacts_free(sc);
2193 
2194 	if (sc->sysctl_tree != NULL)
2195 		sysctl_ctx_free(&sc->sysctl_ctx);
2196 
2197 	/* Deregister the shutdown function */
2198 	if (sc->shutdown_eh != NULL)
2199 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2200 
2201 	mtx_destroy(&sc->mps_mtx);
2202 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2203 
2204 	return (0);
2205 }
2206 
2207 static __inline void
2208 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2209 {
2210 	MPS_FUNCTRACE(sc);
2211 
2212 	if (cm == NULL) {
2213 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2214 		return;
2215 	}
2216 
2217 	KASSERT(cm->cm_state == MPS_CM_STATE_INQUEUE,
2218 	    ("command not inqueue, state = %u\n", cm->cm_state));
2219 	cm->cm_state = MPS_CM_STATE_BUSY;
2220 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2221 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2222 
2223 	if (cm->cm_complete != NULL) {
2224 		mps_dprint(sc, MPS_TRACE,
2225 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2226 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2227 			   cm->cm_reply);
2228 		cm->cm_complete(sc, cm);
2229 	}
2230 
2231 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2232 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2233 		wakeup(cm);
2234 	}
2235 
2236 	if (cm->cm_sc->io_cmds_active != 0) {
2237 		cm->cm_sc->io_cmds_active--;
2238 	} else {
2239 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2240 		    "out of sync - resynching to 0\n");
2241 	}
2242 }
2243 
2244 static void
2245 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2246 {
2247 	union loginfo_type {
2248 		u32     loginfo;
2249 		struct {
2250 			u32     subcode:16;
2251 			u32     code:8;
2252 			u32     originator:4;
2253 			u32     bus_type:4;
2254 		} dw;
2255 	};
2256 	union loginfo_type sas_loginfo;
2257 	char *originator_str = NULL;
2258 
2259 	sas_loginfo.loginfo = log_info;
2260 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2261 		return;
2262 
2263 	/* each nexus loss loginfo */
2264 	if (log_info == 0x31170000)
2265 		return;
2266 
2267 	/* eat the loginfos associated with task aborts */
2268 	if ((log_info == 30050000 || log_info ==
2269 	    0x31140000 || log_info == 0x31130000))
2270 		return;
2271 
2272 	switch (sas_loginfo.dw.originator) {
2273 	case 0:
2274 		originator_str = "IOP";
2275 		break;
2276 	case 1:
2277 		originator_str = "PL";
2278 		break;
2279 	case 2:
2280 		originator_str = "IR";
2281 		break;
2282 }
2283 
2284 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2285 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2286 	originator_str, sas_loginfo.dw.code,
2287 	sas_loginfo.dw.subcode);
2288 }
2289 
2290 static void
2291 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2292 {
2293 	MPI2DefaultReply_t *mpi_reply;
2294 	u16 sc_status;
2295 
2296 	mpi_reply = (MPI2DefaultReply_t*)reply;
2297 	sc_status = le16toh(mpi_reply->IOCStatus);
2298 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2299 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2300 }
2301 void
2302 mps_intr(void *data)
2303 {
2304 	struct mps_softc *sc;
2305 	uint32_t status;
2306 
2307 	sc = (struct mps_softc *)data;
2308 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2309 
2310 	/*
2311 	 * Check interrupt status register to flush the bus.  This is
2312 	 * needed for both INTx interrupts and driver-driven polling
2313 	 */
2314 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2315 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2316 		return;
2317 
2318 	mps_lock(sc);
2319 	mps_intr_locked(data);
2320 	mps_unlock(sc);
2321 	return;
2322 }
2323 
2324 /*
2325  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2326  * chip.  Hopefully this theory is correct.
2327  */
2328 void
2329 mps_intr_msi(void *data)
2330 {
2331 	struct mps_softc *sc;
2332 
2333 	sc = (struct mps_softc *)data;
2334 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2335 	mps_lock(sc);
2336 	mps_intr_locked(data);
2337 	mps_unlock(sc);
2338 	return;
2339 }
2340 
2341 /*
2342  * The locking is overly broad and simplistic, but easy to deal with for now.
2343  */
2344 void
2345 mps_intr_locked(void *data)
2346 {
2347 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2348 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2349 	mps_fw_diagnostic_buffer_t *pBuffer;
2350 	struct mps_softc *sc;
2351 	struct mps_command *cm = NULL;
2352 	uint64_t tdesc;
2353 	uint8_t flags;
2354 	u_int pq;
2355 
2356 	sc = (struct mps_softc *)data;
2357 
2358 	pq = sc->replypostindex;
2359 	mps_dprint(sc, MPS_TRACE,
2360 	    "%s sc %p starting with replypostindex %u\n",
2361 	    __func__, sc, sc->replypostindex);
2362 
2363 	for ( ;; ) {
2364 		cm = NULL;
2365 		desc = &sc->post_queue[sc->replypostindex];
2366 
2367 		/*
2368 		 * Copy and clear out the descriptor so that any reentry will
2369 		 * immediately know that this descriptor has already been
2370 		 * looked at.  There is unfortunate casting magic because the
2371 		 * MPI API doesn't have a cardinal 64bit type.
2372 		 */
2373 		tdesc = 0xffffffffffffffff;
2374 		tdesc = atomic_swap_64((uint64_t *)desc, tdesc);
2375 		desc = (MPI2_REPLY_DESCRIPTORS_UNION *)&tdesc;
2376 
2377 		flags = desc->Default.ReplyFlags &
2378 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2379 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2380 		 || (le32toh(desc->Words.High) == 0xffffffff))
2381 			break;
2382 
2383 		/* increment the replypostindex now, so that event handlers
2384 		 * and cm completion handlers which decide to do a diag
2385 		 * reset can zero it without it getting incremented again
2386 		 * afterwards, and we break out of this loop on the next
2387 		 * iteration since the reply post queue has been cleared to
2388 		 * 0xFF and all descriptors look unused (which they are).
2389 		 */
2390 		if (++sc->replypostindex >= sc->pqdepth)
2391 			sc->replypostindex = 0;
2392 
2393 		switch (flags) {
2394 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2395 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2396 			cm->cm_reply = NULL;
2397 			break;
2398 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2399 		{
2400 			uint32_t baddr;
2401 			uint8_t *reply;
2402 
2403 			/*
2404 			 * Re-compose the reply address from the address
2405 			 * sent back from the chip.  The ReplyFrameAddress
2406 			 * is the lower 32 bits of the physical address of
2407 			 * particular reply frame.  Convert that address to
2408 			 * host format, and then use that to provide the
2409 			 * offset against the virtual address base
2410 			 * (sc->reply_frames).
2411 			 */
2412 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2413 			reply = sc->reply_frames +
2414 				(baddr - ((uint32_t)sc->reply_busaddr));
2415 			/*
2416 			 * Make sure the reply we got back is in a valid
2417 			 * range.  If not, go ahead and panic here, since
2418 			 * we'll probably panic as soon as we deference the
2419 			 * reply pointer anyway.
2420 			 */
2421 			if ((reply < sc->reply_frames)
2422 			 || (reply > (sc->reply_frames +
2423 			     (sc->fqdepth * sc->replyframesz)))) {
2424 				printf("%s: WARNING: reply %p out of range!\n",
2425 				       __func__, reply);
2426 				printf("%s: reply_frames %p, fqdepth %d, "
2427 				       "frame size %d\n", __func__,
2428 				       sc->reply_frames, sc->fqdepth,
2429 				       sc->replyframesz);
2430 				printf("%s: baddr %#x,\n", __func__, baddr);
2431 				/* LSI-TODO. See Linux Code for Graceful exit */
2432 				panic("Reply address out of range");
2433 			}
2434 			if (le16toh(desc->AddressReply.SMID) == 0) {
2435 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2436 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2437 					/*
2438 					 * If SMID is 0 for Diag Buffer Post,
2439 					 * this implies that the reply is due to
2440 					 * a release function with a status that
2441 					 * the buffer has been released.  Set
2442 					 * the buffer flags accordingly.
2443 					 */
2444 					rel_rep =
2445 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2446 					if ((le16toh(rel_rep->IOCStatus) &
2447 					    MPI2_IOCSTATUS_MASK) ==
2448 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2449 					{
2450 						pBuffer =
2451 						    &sc->fw_diag_buffer_list[
2452 						    rel_rep->BufferType];
2453 						pBuffer->valid_data = TRUE;
2454 						pBuffer->owned_by_firmware =
2455 						    FALSE;
2456 						pBuffer->immediate = FALSE;
2457 					}
2458 				} else
2459 					mps_dispatch_event(sc, baddr,
2460 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2461 					    reply);
2462 			} else {
2463 				/*
2464 				 * Ignore commands not in INQUEUE state
2465 				 * since they've already been completed
2466 				 * via another path.
2467 				 */
2468 				cm = &sc->commands[
2469 				    le16toh(desc->AddressReply.SMID)];
2470 				if (cm->cm_state == MPS_CM_STATE_INQUEUE) {
2471 					cm->cm_reply = reply;
2472 					cm->cm_reply_data = le32toh(
2473 					    desc->AddressReply.ReplyFrameAddress);
2474 				} else {
2475 					mps_dprint(sc, MPS_RECOVERY,
2476 					    "Bad state for ADDRESS_REPLY status,"
2477 					    " ignoring state %d cm %p\n",
2478 					    cm->cm_state, cm);
2479 				}
2480 			}
2481 			break;
2482 		}
2483 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2484 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2485 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2486 		default:
2487 			/* Unhandled */
2488 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2489 			    desc->Default.ReplyFlags);
2490 			cm = NULL;
2491 			break;
2492 		}
2493 
2494 
2495 		if (cm != NULL) {
2496 			// Print Error reply frame
2497 			if (cm->cm_reply)
2498 				mps_display_reply_info(sc,cm->cm_reply);
2499 			mps_complete_command(sc, cm);
2500 		}
2501 	}
2502 
2503 	if (pq != sc->replypostindex) {
2504 		mps_dprint(sc, MPS_TRACE, "%s sc %p writing postindex %d\n",
2505 		    __func__, sc, sc->replypostindex);
2506 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2507 		    sc->replypostindex);
2508 	}
2509 
2510 	return;
2511 }
2512 
2513 static void
2514 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2515     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2516 {
2517 	struct mps_event_handle *eh;
2518 	int event, handled = 0;
2519 
2520 	event = le16toh(reply->Event);
2521 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2522 		if (isset(eh->mask, event)) {
2523 			eh->callback(sc, data, reply);
2524 			handled++;
2525 		}
2526 	}
2527 
2528 	if (handled == 0)
2529 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2530 
2531 	/*
2532 	 * This is the only place that the event/reply should be freed.
2533 	 * Anything wanting to hold onto the event data should have
2534 	 * already copied it into their own storage.
2535 	 */
2536 	mps_free_reply(sc, data);
2537 }
2538 
2539 static void
2540 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2541 {
2542 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2543 
2544 	if (cm->cm_reply)
2545 		MPS_DPRINT_EVENT(sc, generic,
2546 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2547 
2548 	mps_free_command(sc, cm);
2549 
2550 	/* next, send a port enable */
2551 	mpssas_startup(sc);
2552 }
2553 
2554 /*
2555  * For both register_events and update_events, the caller supplies a bitmap
2556  * of events that it _wants_.  These functions then turn that into a bitmask
2557  * suitable for the controller.
2558  */
2559 int
2560 mps_register_events(struct mps_softc *sc, u32 *mask,
2561     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2562 {
2563 	struct mps_event_handle *eh;
2564 	int error = 0;
2565 
2566 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2567 	eh->callback = cb;
2568 	eh->data = data;
2569 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2570 	if (mask != NULL)
2571 		error = mps_update_events(sc, eh, mask);
2572 	*handle = eh;
2573 
2574 	return (error);
2575 }
2576 
2577 int
2578 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2579     u32 *mask)
2580 {
2581 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2582 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2583 	struct mps_command *cm;
2584 	int error, i;
2585 
2586 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2587 
2588 	if ((mask != NULL) && (handle != NULL))
2589 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2590 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2591 
2592 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2593 		sc->event_mask[i] = -1;
2594 
2595 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2596 		sc->event_mask[i] &= ~handle->mask[i];
2597 
2598 	if ((cm = mps_alloc_command(sc)) == NULL)
2599 		return (EBUSY);
2600 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2601 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2602 	evtreq->MsgFlags = 0;
2603 	evtreq->SASBroadcastPrimitiveMasks = 0;
2604 #ifdef MPS_DEBUG_ALL_EVENTS
2605 	{
2606 		u_char fullmask[16];
2607 		memset(fullmask, 0x00, 16);
2608 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2609 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2610 	}
2611 #else
2612         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2613                 evtreq->EventMasks[i] =
2614                     htole32(sc->event_mask[i]);
2615 #endif
2616 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2617 	cm->cm_data = NULL;
2618 
2619 	error = mps_wait_command(sc, &cm, 60, 0);
2620 	if (cm != NULL)
2621 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2622 	if ((reply == NULL) ||
2623 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2624 		error = ENXIO;
2625 
2626 	if (reply)
2627 		MPS_DPRINT_EVENT(sc, generic, reply);
2628 
2629 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2630 
2631 	if (cm != NULL)
2632 		mps_free_command(sc, cm);
2633 	return (error);
2634 }
2635 
2636 static int
2637 mps_reregister_events(struct mps_softc *sc)
2638 {
2639 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2640 	struct mps_command *cm;
2641 	struct mps_event_handle *eh;
2642 	int error, i;
2643 
2644 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2645 
2646 	/* first, reregister events */
2647 
2648 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2649 		sc->event_mask[i] = -1;
2650 
2651 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2652 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2653 			sc->event_mask[i] &= ~eh->mask[i];
2654 	}
2655 
2656 	if ((cm = mps_alloc_command(sc)) == NULL)
2657 		return (EBUSY);
2658 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2659 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2660 	evtreq->MsgFlags = 0;
2661 	evtreq->SASBroadcastPrimitiveMasks = 0;
2662 #ifdef MPS_DEBUG_ALL_EVENTS
2663 	{
2664 		u_char fullmask[16];
2665 		memset(fullmask, 0x00, 16);
2666 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2667 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2668 	}
2669 #else
2670         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2671                 evtreq->EventMasks[i] =
2672                     htole32(sc->event_mask[i]);
2673 #endif
2674 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2675 	cm->cm_data = NULL;
2676 	cm->cm_complete = mps_reregister_events_complete;
2677 
2678 	error = mps_map_command(sc, cm);
2679 
2680 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2681 	    error);
2682 	return (error);
2683 }
2684 
2685 void
2686 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2687 {
2688 
2689 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2690 	free(handle, M_MPT2);
2691 }
2692 
2693 /*
2694  * Add a chain element as the next SGE for the specified command.
2695  * Reset cm_sge and cm_sgesize to indicate all the available space.
2696  */
2697 static int
2698 mps_add_chain(struct mps_command *cm)
2699 {
2700 	MPI2_SGE_CHAIN64 *sgc;
2701 	struct mps_chain *chain;
2702 	u_int space;
2703 
2704 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2705 		panic("MPS: Need SGE Error Code\n");
2706 
2707 	chain = mps_alloc_chain(cm->cm_sc);
2708 	if (chain == NULL)
2709 		return (ENOBUFS);
2710 
2711 	space = cm->cm_sc->reqframesz;
2712 
2713 	/*
2714 	 * Note: a double-linked list is used to make it easier to
2715 	 * walk for debugging.
2716 	 */
2717 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2718 
2719 	sgc = (MPI2_SGE_CHAIN64 *)&cm->cm_sge->MpiChain;
2720 	sgc->Length = htole16(space);
2721 	sgc->NextChainOffset = 0;
2722 	/* TODO Looks like bug in Setting sgc->Flags.
2723 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2724 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2725 	 *	This is fine.. because we are not using simple element. In case of
2726 	 *	MPI2_SGE_CHAIN64, we have separate Length and Flags field.
2727  	 */
2728 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2729 	sgc->Address.High = htole32(chain->chain_busaddr >> 32);
2730 	sgc->Address.Low = htole32(chain->chain_busaddr);
2731 
2732 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2733 	cm->cm_sglsize = space;
2734 	return (0);
2735 }
2736 
2737 /*
2738  * Add one scatter-gather element (chain, simple, transaction context)
2739  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2740  * cm_sge as the remaining size and pointer to the next SGE to fill
2741  * in, respectively.
2742  */
2743 int
2744 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2745 {
2746 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2747 	MPI2_SGE_SIMPLE64 *sge = sgep;
2748 	int error, type;
2749 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2750 
2751 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2752 
2753 #ifdef INVARIANTS
2754 	switch (type) {
2755 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2756 		if (len != tc->DetailsLength + 4)
2757 			panic("TC %p length %u or %zu?", tc,
2758 			    tc->DetailsLength + 4, len);
2759 		}
2760 		break;
2761 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2762 		/* Driver only uses 64-bit chain elements */
2763 		if (len != MPS_SGC_SIZE)
2764 			panic("CHAIN %p length %u or %zu?", sgep,
2765 			    MPS_SGC_SIZE, len);
2766 		break;
2767 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2768 		/* Driver only uses 64-bit SGE simple elements */
2769 		if (len != MPS_SGE64_SIZE)
2770 			panic("SGE simple %p length %u or %zu?", sge,
2771 			    MPS_SGE64_SIZE, len);
2772 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2773 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2774 			panic("SGE simple %p not marked 64-bit?", sge);
2775 
2776 		break;
2777 	default:
2778 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2779 	}
2780 #endif
2781 
2782 	/*
2783 	 * case 1: 1 more segment, enough room for it
2784 	 * case 2: 2 more segments, enough room for both
2785 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2786 	 * case 4: >=1 more segment, enough room for only a chain
2787 	 * case 5: >=1 more segment, no room for anything (error)
2788          */
2789 
2790 	/*
2791 	 * There should be room for at least a chain element, or this
2792 	 * code is buggy.  Case (5).
2793 	 */
2794 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2795 		panic("MPS: Need SGE Error Code\n");
2796 
2797 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2798 		/*
2799 		 * 1 or more segment, enough room for only a chain.
2800 		 * Hope the previous element wasn't a Simple entry
2801 		 * that needed to be marked with
2802 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2803 		 */
2804 		if ((error = mps_add_chain(cm)) != 0)
2805 			return (error);
2806 	}
2807 
2808 	if (segsleft >= 2 &&
2809 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2810 		/*
2811 		 * There are 2 or more segments left to add, and only
2812 		 * enough room for 1 and a chain.  Case (3).
2813 		 *
2814 		 * Mark as last element in this chain if necessary.
2815 		 */
2816 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2817 			sge->FlagsLength |= htole32(
2818 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2819 		}
2820 
2821 		/*
2822 		 * Add the item then a chain.  Do the chain now,
2823 		 * rather than on the next iteration, to simplify
2824 		 * understanding the code.
2825 		 */
2826 		cm->cm_sglsize -= len;
2827 		bcopy(sgep, cm->cm_sge, len);
2828 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2829 		return (mps_add_chain(cm));
2830 	}
2831 
2832 #ifdef INVARIANTS
2833 	/* Case 1: 1 more segment, enough room for it. */
2834 	if (segsleft == 1 && cm->cm_sglsize < len)
2835 		panic("1 seg left and no room? %u versus %zu",
2836 		    cm->cm_sglsize, len);
2837 
2838 	/* Case 2: 2 more segments, enough room for both */
2839 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2840 		panic("2 segs left and no room? %u versus %zu",
2841 		    cm->cm_sglsize, len);
2842 #endif
2843 
2844 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2845 		/*
2846 		 * If this is a bi-directional request, need to account for that
2847 		 * here.  Save the pre-filled sge values.  These will be used
2848 		 * either for the 2nd SGL or for a single direction SGL.  If
2849 		 * cm_out_len is non-zero, this is a bi-directional request, so
2850 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2851 		 * fill in the IN SGL.  Note that at this time, when filling in
2852 		 * 2 SGL's for a bi-directional request, they both use the same
2853 		 * DMA buffer (same cm command).
2854 		 */
2855 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2856 		saved_address_low = sge->Address.Low;
2857 		saved_address_high = sge->Address.High;
2858 		if (cm->cm_out_len) {
2859 			sge->FlagsLength = htole32(cm->cm_out_len |
2860 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2861 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2862 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2863 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2864 			    MPI2_SGE_FLAGS_SHIFT));
2865 			cm->cm_sglsize -= len;
2866 			bcopy(sgep, cm->cm_sge, len);
2867 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2868 			    + len);
2869 		}
2870 		saved_buf_len |=
2871 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2872 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2873 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2874 		    MPI2_SGE_FLAGS_END_OF_LIST |
2875 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2876 		    MPI2_SGE_FLAGS_SHIFT);
2877 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2878 			saved_buf_len |=
2879 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2880 			    MPI2_SGE_FLAGS_SHIFT);
2881 		} else {
2882 			saved_buf_len |=
2883 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2884 			    MPI2_SGE_FLAGS_SHIFT);
2885 		}
2886 		sge->FlagsLength = htole32(saved_buf_len);
2887 		sge->Address.Low = saved_address_low;
2888 		sge->Address.High = saved_address_high;
2889 	}
2890 
2891 	cm->cm_sglsize -= len;
2892 	bcopy(sgep, cm->cm_sge, len);
2893 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2894 	return (0);
2895 }
2896 
2897 /*
2898  * Add one dma segment to the scatter-gather list for a command.
2899  */
2900 int
2901 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2902     int segsleft)
2903 {
2904 	MPI2_SGE_SIMPLE64 sge;
2905 
2906 	/*
2907 	 * This driver always uses 64-bit address elements for simplicity.
2908 	 */
2909 	bzero(&sge, sizeof(sge));
2910 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2911 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2912 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2913 	mps_from_u64(pa, &sge.Address);
2914 
2915 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2916 }
2917 
2918 static void
2919 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2920 {
2921 	struct mps_softc *sc;
2922 	struct mps_command *cm;
2923 	u_int i, dir, sflags;
2924 
2925 	cm = (struct mps_command *)arg;
2926 	sc = cm->cm_sc;
2927 
2928 	/*
2929 	 * In this case, just print out a warning and let the chip tell the
2930 	 * user they did the wrong thing.
2931 	 */
2932 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2933 		mps_dprint(sc, MPS_ERROR,
2934 			   "%s: warning: busdma returned %d segments, "
2935 			   "more than the %d allowed\n", __func__, nsegs,
2936 			   cm->cm_max_segs);
2937 	}
2938 
2939 	/*
2940 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2941 	 * here.  In that case, both direction flags will be set.
2942 	 */
2943 	sflags = 0;
2944 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2945 		/*
2946 		 * We have to add a special case for SMP passthrough, there
2947 		 * is no easy way to generically handle it.  The first
2948 		 * S/G element is used for the command (therefore the
2949 		 * direction bit needs to be set).  The second one is used
2950 		 * for the reply.  We'll leave it to the caller to make
2951 		 * sure we only have two buffers.
2952 		 */
2953 		/*
2954 		 * Even though the busdma man page says it doesn't make
2955 		 * sense to have both direction flags, it does in this case.
2956 		 * We have one s/g element being accessed in each direction.
2957 		 */
2958 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2959 
2960 		/*
2961 		 * Set the direction flag on the first buffer in the SMP
2962 		 * passthrough request.  We'll clear it for the second one.
2963 		 */
2964 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2965 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2966 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2967 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2968 		dir = BUS_DMASYNC_PREWRITE;
2969 	} else
2970 		dir = BUS_DMASYNC_PREREAD;
2971 
2972 	for (i = 0; i < nsegs; i++) {
2973 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2974 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2975 		}
2976 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2977 		    sflags, nsegs - i);
2978 		if (error != 0) {
2979 			/* Resource shortage, roll back! */
2980 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2981 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2982 				    "consider increasing hw.mps.max_chains.\n");
2983 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2984 			/*
2985 			 * mpr_complete_command can only be called on commands
2986 			 * that are in the queue. Since this is an error path
2987 			 * which gets called before we enqueue, update the state
2988 			 * to meet this requirement before we complete it.
2989 			 */
2990 			cm->cm_state = MPS_CM_STATE_INQUEUE;
2991 			mps_complete_command(sc, cm);
2992 			return;
2993 		}
2994 	}
2995 
2996 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2997 	mps_enqueue_request(sc, cm);
2998 
2999 	return;
3000 }
3001 
3002 static void
3003 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3004 	     int error)
3005 {
3006 	mps_data_cb(arg, segs, nsegs, error);
3007 }
3008 
3009 /*
3010  * This is the routine to enqueue commands ansynchronously.
3011  * Note that the only error path here is from bus_dmamap_load(), which can
3012  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3013  * assumed that if you have a command in-hand, then you have enough credits
3014  * to use it.
3015  */
3016 int
3017 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
3018 {
3019 	int error = 0;
3020 
3021 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
3022 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3023 		    &cm->cm_uio, mps_data_cb2, cm, 0);
3024 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
3025 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3026 		    cm->cm_data, mps_data_cb, cm, 0);
3027 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3028 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3029 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
3030 	} else {
3031 		/* Add a zero-length element as needed */
3032 		if (cm->cm_sge != NULL)
3033 			mps_add_dmaseg(cm, 0, 0, 0, 1);
3034 		mps_enqueue_request(sc, cm);
3035 	}
3036 
3037 	return (error);
3038 }
3039 
3040 /*
3041  * This is the routine to enqueue commands synchronously.  An error of
3042  * EINPROGRESS from mps_map_command() is ignored since the command will
3043  * be executed and enqueued automatically.  Other errors come from msleep().
3044  */
3045 int
3046 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
3047     int sleep_flag)
3048 {
3049 	int error, rc;
3050 	struct timeval cur_time, start_time;
3051 	struct mps_command *cm = *cmp;
3052 
3053 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
3054 		return  EBUSY;
3055 
3056 	cm->cm_complete = NULL;
3057 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
3058 	error = mps_map_command(sc, cm);
3059 	if ((error != 0) && (error != EINPROGRESS))
3060 		return (error);
3061 
3062 	/*
3063 	 * Check for context and wait for 50 mSec at a time until time has
3064 	 * expired or the command has finished.  If msleep can't be used, need
3065 	 * to poll.
3066 	 */
3067 	if (curthread->td_no_sleeping != 0)
3068 		sleep_flag = NO_SLEEP;
3069 	getmicrouptime(&start_time);
3070 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
3071 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
3072 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
3073 		if (error == EWOULDBLOCK) {
3074 			/*
3075 			 * Record the actual elapsed time in the case of a
3076 			 * timeout for the message below.
3077 			 */
3078 			getmicrouptime(&cur_time);
3079 			timevalsub(&cur_time, &start_time);
3080 		}
3081 	} else {
3082 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
3083 			mps_intr_locked(sc);
3084 			if (sleep_flag == CAN_SLEEP)
3085 				pause("mpswait", hz/20);
3086 			else
3087 				DELAY(50000);
3088 
3089 			getmicrouptime(&cur_time);
3090 			timevalsub(&cur_time, &start_time);
3091 			if (cur_time.tv_sec > timeout) {
3092 				error = EWOULDBLOCK;
3093 				break;
3094 			}
3095 		}
3096 	}
3097 
3098 	if (error == EWOULDBLOCK) {
3099 		if (cm->cm_timeout_handler == NULL) {
3100 			mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
3101 			    " elapsed=%jd\n", __func__, timeout,
3102 			    (intmax_t)cur_time.tv_sec);
3103 			rc = mps_reinit(sc);
3104 			mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3105 			    "failed");
3106 		} else
3107 			cm->cm_timeout_handler(sc, cm);
3108 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
3109 			/*
3110 			 * Tell the caller that we freed the command in a
3111 			 * reinit.
3112 			 */
3113 			*cmp = NULL;
3114 		}
3115 		error = ETIMEDOUT;
3116 	}
3117 	return (error);
3118 }
3119 
3120 /*
3121  * The MPT driver had a verbose interface for config pages.  In this driver,
3122  * reduce it to much simpler terms, similar to the Linux driver.
3123  */
3124 int
3125 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
3126 {
3127 	MPI2_CONFIG_REQUEST *req;
3128 	struct mps_command *cm;
3129 	int error;
3130 
3131 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
3132 		return (EBUSY);
3133 	}
3134 
3135 	cm = mps_alloc_command(sc);
3136 	if (cm == NULL) {
3137 		return (EBUSY);
3138 	}
3139 
3140 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3141 	req->Function = MPI2_FUNCTION_CONFIG;
3142 	req->Action = params->action;
3143 	req->SGLFlags = 0;
3144 	req->ChainOffset = 0;
3145 	req->PageAddress = params->page_address;
3146 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3147 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3148 
3149 		hdr = &params->hdr.Ext;
3150 		req->ExtPageType = hdr->ExtPageType;
3151 		req->ExtPageLength = hdr->ExtPageLength;
3152 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3153 		req->Header.PageLength = 0; /* Must be set to zero */
3154 		req->Header.PageNumber = hdr->PageNumber;
3155 		req->Header.PageVersion = hdr->PageVersion;
3156 	} else {
3157 		MPI2_CONFIG_PAGE_HEADER *hdr;
3158 
3159 		hdr = &params->hdr.Struct;
3160 		req->Header.PageType = hdr->PageType;
3161 		req->Header.PageNumber = hdr->PageNumber;
3162 		req->Header.PageLength = hdr->PageLength;
3163 		req->Header.PageVersion = hdr->PageVersion;
3164 	}
3165 
3166 	cm->cm_data = params->buffer;
3167 	cm->cm_length = params->length;
3168 	if (cm->cm_data != NULL) {
3169 		cm->cm_sge = &req->PageBufferSGE;
3170 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3171 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
3172 	} else
3173 		cm->cm_sge = NULL;
3174 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3175 
3176 	cm->cm_complete_data = params;
3177 	if (params->callback != NULL) {
3178 		cm->cm_complete = mps_config_complete;
3179 		return (mps_map_command(sc, cm));
3180 	} else {
3181 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
3182 		if (error) {
3183 			mps_dprint(sc, MPS_FAULT,
3184 			    "Error %d reading config page\n", error);
3185 			if (cm != NULL)
3186 				mps_free_command(sc, cm);
3187 			return (error);
3188 		}
3189 		mps_config_complete(sc, cm);
3190 	}
3191 
3192 	return (0);
3193 }
3194 
3195 int
3196 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3197 {
3198 	return (EINVAL);
3199 }
3200 
3201 static void
3202 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3203 {
3204 	MPI2_CONFIG_REPLY *reply;
3205 	struct mps_config_params *params;
3206 
3207 	MPS_FUNCTRACE(sc);
3208 	params = cm->cm_complete_data;
3209 
3210 	if (cm->cm_data != NULL) {
3211 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3212 		    BUS_DMASYNC_POSTREAD);
3213 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3214 	}
3215 
3216 	/*
3217 	 * XXX KDM need to do more error recovery?  This results in the
3218 	 * device in question not getting probed.
3219 	 */
3220 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3221 		params->status = MPI2_IOCSTATUS_BUSY;
3222 		goto done;
3223 	}
3224 
3225 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3226 	if (reply == NULL) {
3227 		params->status = MPI2_IOCSTATUS_BUSY;
3228 		goto done;
3229 	}
3230 	params->status = reply->IOCStatus;
3231 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3232 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3233 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3234 		params->hdr.Ext.PageType = reply->Header.PageType;
3235 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3236 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3237 	} else {
3238 		params->hdr.Struct.PageType = reply->Header.PageType;
3239 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3240 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3241 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3242 	}
3243 
3244 done:
3245 	mps_free_command(sc, cm);
3246 	if (params->callback != NULL)
3247 		params->callback(sc, params);
3248 
3249 	return;
3250 }
3251