xref: /freebsd/sys/dev/mps/mps.c (revision 3c134670993bf525fcd6c4dfef84a3dfc3d4ed1b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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  * $FreeBSD$
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 /* Communications core for Avago Technologies (LSI) MPT2 */
39 
40 /* TODO Move headers to mpsvar */
41 #include <sys/types.h>
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/selinfo.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/module.h>
49 #include <sys/bus.h>
50 #include <sys/conf.h>
51 #include <sys/bio.h>
52 #include <sys/malloc.h>
53 #include <sys/uio.h>
54 #include <sys/sysctl.h>
55 #include <sys/smp.h>
56 #include <sys/queue.h>
57 #include <sys/kthread.h>
58 #include <sys/taskqueue.h>
59 #include <sys/endian.h>
60 #include <sys/eventhandler.h>
61 #include <sys/sbuf.h>
62 #include <sys/priv.h>
63 
64 #include <machine/bus.h>
65 #include <machine/resource.h>
66 #include <sys/rman.h>
67 #include <sys/proc.h>
68 
69 #include <dev/pci/pcivar.h>
70 
71 #include <cam/cam.h>
72 #include <cam/scsi/scsi_all.h>
73 
74 #include <dev/mps/mpi/mpi2_type.h>
75 #include <dev/mps/mpi/mpi2.h>
76 #include <dev/mps/mpi/mpi2_ioc.h>
77 #include <dev/mps/mpi/mpi2_sas.h>
78 #include <dev/mps/mpi/mpi2_cnfg.h>
79 #include <dev/mps/mpi/mpi2_init.h>
80 #include <dev/mps/mpi/mpi2_tool.h>
81 #include <dev/mps/mps_ioctl.h>
82 #include <dev/mps/mpsvar.h>
83 #include <dev/mps/mps_table.h>
84 
85 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
86 static int mps_init_queues(struct mps_softc *sc);
87 static void mps_resize_queues(struct mps_softc *sc);
88 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
89 static int mps_transition_operational(struct mps_softc *sc);
90 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
91 static void mps_iocfacts_free(struct mps_softc *sc);
92 static void mps_startup(void *arg);
93 static int mps_send_iocinit(struct mps_softc *sc);
94 static int mps_alloc_queues(struct mps_softc *sc);
95 static int mps_alloc_hw_queues(struct mps_softc *sc);
96 static int mps_alloc_replies(struct mps_softc *sc);
97 static int mps_alloc_requests(struct mps_softc *sc);
98 static int mps_attach_log(struct mps_softc *sc);
99 static __inline void mps_complete_command(struct mps_softc *sc,
100     struct mps_command *cm);
101 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
102     MPI2_EVENT_NOTIFICATION_REPLY *reply);
103 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
104 static void mps_periodic(void *);
105 static int mps_reregister_events(struct mps_softc *sc);
106 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
107 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
108 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
109 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
110 static int mps_dump_reqs(SYSCTL_HANDLER_ARGS);
111 static void mps_parse_debug(struct mps_softc *sc, char *list);
112 
113 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
114     "MPS Driver Parameters");
115 
116 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
117 MALLOC_DECLARE(M_MPSUSER);
118 
119 /*
120  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
121  * any state and back to its initialization state machine.
122  */
123 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
124 
125 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
126  * Compiler only support unint64_t to be passed as argument.
127  * Otherwise it will throw below error
128  * "aggregate value used where an integer was expected"
129  */
130 
131 typedef union _reply_descriptor {
132         u64 word;
133         struct {
134                 u32 low;
135                 u32 high;
136         } u;
137 }reply_descriptor,address_descriptor;
138 
139 /* Rate limit chain-fail messages to 1 per minute */
140 static struct timeval mps_chainfail_interval = { 60, 0 };
141 
142 /*
143  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
144  * If this function is called from process context, it can sleep
145  * and there is no harm to sleep, in case if this fuction is called
146  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
147  * based on sleep flags driver will call either msleep, pause or DELAY.
148  * msleep and pause are of same variant, but pause is used when mps_mtx
149  * is not hold by driver.
150  *
151  */
152 static int
153 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
154 {
155 	uint32_t reg;
156 	int i, error, tries = 0;
157 	uint8_t first_wait_done = FALSE;
158 
159 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
160 
161 	/* Clear any pending interrupts */
162 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
163 
164 	/*
165 	 * Force NO_SLEEP for threads prohibited to sleep
166  	 * e.a Thread from interrupt handler are prohibited to sleep.
167  	 */
168 	if (curthread->td_no_sleeping != 0)
169 		sleep_flag = NO_SLEEP;
170 
171 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
172 
173 	/* Push the magic sequence */
174 	error = ETIMEDOUT;
175 	while (tries++ < 20) {
176 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
177 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
178 			    mpt2_reset_magic[i]);
179 		/* wait 100 msec */
180 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
181 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
182 			    "mpsdiag", hz/10);
183 		else if (sleep_flag == CAN_SLEEP)
184 			pause("mpsdiag", hz/10);
185 		else
186 			DELAY(100 * 1000);
187 
188 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
189 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
190 			error = 0;
191 			break;
192 		}
193 	}
194 	if (error) {
195 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
196 		    error);
197 		return (error);
198 	}
199 
200 	/* Send the actual reset.  XXX need to refresh the reg? */
201 	reg |= MPI2_DIAG_RESET_ADAPTER;
202 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
203 		reg);
204 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
205 
206 	/* Wait up to 300 seconds in 50ms intervals */
207 	error = ETIMEDOUT;
208 	for (i = 0; i < 6000; i++) {
209 		/*
210 		 * Wait 50 msec. If this is the first time through, wait 256
211 		 * msec to satisfy Diag Reset timing requirements.
212 		 */
213 		if (first_wait_done) {
214 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
215 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
216 				    "mpsdiag", hz/20);
217 			else if (sleep_flag == CAN_SLEEP)
218 				pause("mpsdiag", hz/20);
219 			else
220 				DELAY(50 * 1000);
221 		} else {
222 			DELAY(256 * 1000);
223 			first_wait_done = TRUE;
224 		}
225 		/*
226 		 * Check for the RESET_ADAPTER bit to be cleared first, then
227 		 * wait for the RESET state to be cleared, which takes a little
228 		 * longer.
229 		 */
230 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
231 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
232 			continue;
233 		}
234 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
235 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
236 			error = 0;
237 			break;
238 		}
239 	}
240 	if (error) {
241 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
242 		    error);
243 		return (error);
244 	}
245 
246 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
247 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
248 
249 	return (0);
250 }
251 
252 static int
253 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
254 {
255 	int error;
256 
257 	MPS_FUNCTRACE(sc);
258 
259 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
260 
261 	error = 0;
262 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
263 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
264 	    MPI2_DOORBELL_FUNCTION_SHIFT);
265 
266 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
267 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
268 		    "Doorbell handshake failed\n");
269 		error = ETIMEDOUT;
270 	}
271 
272 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
273 	return (error);
274 }
275 
276 static int
277 mps_transition_ready(struct mps_softc *sc)
278 {
279 	uint32_t reg, state;
280 	int error, tries = 0;
281 	int sleep_flags;
282 
283 	MPS_FUNCTRACE(sc);
284 	/* If we are in attach call, do not sleep */
285 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
286 					? CAN_SLEEP:NO_SLEEP;
287 	error = 0;
288 
289 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
290 	   __func__, sleep_flags);
291 
292 	while (tries++ < 1200) {
293 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
294 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
295 
296 		/*
297 		 * Ensure the IOC is ready to talk.  If it's not, try
298 		 * resetting it.
299 		 */
300 		if (reg & MPI2_DOORBELL_USED) {
301 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
302 			    "reset\n");
303 			mps_diag_reset(sc, sleep_flags);
304 			DELAY(50000);
305 			continue;
306 		}
307 
308 		/* Is the adapter owned by another peer? */
309 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
310 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
311 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
312 			    "control of another peer host, aborting "
313 			    "initialization.\n");
314 			error = ENXIO;
315 			break;
316 		}
317 
318 		state = reg & MPI2_IOC_STATE_MASK;
319 		if (state == MPI2_IOC_STATE_READY) {
320 			/* Ready to go! */
321 			error = 0;
322 			break;
323 		} else if (state == MPI2_IOC_STATE_FAULT) {
324 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
325 			    "state 0x%x, resetting\n",
326 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
327 			mps_diag_reset(sc, sleep_flags);
328 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
329 			/* Need to take ownership */
330 			mps_message_unit_reset(sc, sleep_flags);
331 		} else if (state == MPI2_IOC_STATE_RESET) {
332 			/* Wait a bit, IOC might be in transition */
333 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
334 			    "IOC in unexpected reset state\n");
335 		} else {
336 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
337 			    "IOC in unknown state 0x%x\n", state);
338 			error = EINVAL;
339 			break;
340 		}
341 
342 		/* Wait 50ms for things to settle down. */
343 		DELAY(50000);
344 	}
345 
346 	if (error)
347 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
348 		    "Cannot transition IOC to ready\n");
349 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
350 
351 	return (error);
352 }
353 
354 static int
355 mps_transition_operational(struct mps_softc *sc)
356 {
357 	uint32_t reg, state;
358 	int error;
359 
360 	MPS_FUNCTRACE(sc);
361 
362 	error = 0;
363 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
364 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
365 
366 	state = reg & MPI2_IOC_STATE_MASK;
367 	if (state != MPI2_IOC_STATE_READY) {
368 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
369 		if ((error = mps_transition_ready(sc)) != 0) {
370 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
371 			    "failed to transition ready, exit\n");
372 			return (error);
373 		}
374 	}
375 
376 	error = mps_send_iocinit(sc);
377 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
378 
379 	return (error);
380 }
381 
382 static void
383 mps_resize_queues(struct mps_softc *sc)
384 {
385 	u_int reqcr, prireqcr, maxio, sges_per_frame;
386 
387 	/*
388 	 * Size the queues. Since the reply queues always need one free
389 	 * entry, we'll deduct one reply message here.  The LSI documents
390 	 * suggest instead to add a count to the request queue, but I think
391 	 * that it's better to deduct from reply queue.
392 	 */
393 	prireqcr = MAX(1, sc->max_prireqframes);
394 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
395 
396 	reqcr = MAX(2, sc->max_reqframes);
397 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
398 
399 	sc->num_reqs = prireqcr + reqcr;
400 	sc->num_prireqs = prireqcr;
401 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
402 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
403 
404 	/* Store the request frame size in bytes rather than as 32bit words */
405 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
406 
407 	/*
408 	 * Max IO Size is Page Size * the following:
409 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
410 	 * + 1 for no chain needed in last frame
411 	 *
412 	 * If user suggests a Max IO size to use, use the smaller of the
413 	 * user's value and the calculated value as long as the user's
414 	 * value is larger than 0. The user's value is in pages.
415 	 */
416 	sges_per_frame = sc->reqframesz / sizeof(MPI2_SGE_SIMPLE64) - 1;
417 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
418 
419 	/*
420 	 * If I/O size limitation requested, then use it and pass up to CAM.
421 	 * If not, use MAXPHYS as an optimization hint, but report HW limit.
422 	 */
423 	if (sc->max_io_pages > 0) {
424 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
425 		sc->maxio = maxio;
426 	} else {
427 		sc->maxio = maxio;
428 		maxio = min(maxio, MAXPHYS);
429 	}
430 
431 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
432 	    sges_per_frame * reqcr;
433 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
434 		sc->num_chains = sc->max_chains;
435 
436 	/*
437 	 * Figure out the number of MSIx-based queues.  If the firmware or
438 	 * user has done something crazy and not allowed enough credit for
439 	 * the queues to be useful then don't enable multi-queue.
440 	 */
441 	if (sc->facts->MaxMSIxVectors < 2)
442 		sc->msi_msgs = 1;
443 
444 	if (sc->msi_msgs > 1) {
445 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
446 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
447 		if (sc->num_reqs / sc->msi_msgs < 2)
448 			sc->msi_msgs = 1;
449 	}
450 
451 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
452 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
453 }
454 
455 /*
456  * This is called during attach and when re-initializing due to a Diag Reset.
457  * IOC Facts is used to allocate many of the structures needed by the driver.
458  * If called from attach, de-allocation is not required because the driver has
459  * not allocated any structures yet, but if called from a Diag Reset, previously
460  * allocated structures based on IOC Facts will need to be freed and re-
461  * allocated bases on the latest IOC Facts.
462  */
463 static int
464 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
465 {
466 	int error;
467 	Mpi2IOCFactsReply_t saved_facts;
468 	uint8_t saved_mode, reallocating;
469 
470 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
471 
472 	/* Save old IOC Facts and then only reallocate if Facts have changed */
473 	if (!attaching) {
474 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
475 	}
476 
477 	/*
478 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
479 	 * a re-initialization and only return the error if attaching so the OS
480 	 * can handle it.
481 	 */
482 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
483 		if (attaching) {
484 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
485 			    "IOC Facts with error %d, exit\n", error);
486 			return (error);
487 		} else {
488 			panic("%s failed to get IOC Facts with error %d\n",
489 			    __func__, error);
490 		}
491 	}
492 
493 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
494 
495 	snprintf(sc->fw_version, sizeof(sc->fw_version),
496 	    "%02d.%02d.%02d.%02d",
497 	    sc->facts->FWVersion.Struct.Major,
498 	    sc->facts->FWVersion.Struct.Minor,
499 	    sc->facts->FWVersion.Struct.Unit,
500 	    sc->facts->FWVersion.Struct.Dev);
501 
502 	snprintf(sc->msg_version, sizeof(sc->msg_version), "%d.%d",
503 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MAJOR_MASK) >>
504 	    MPI2_IOCFACTS_MSGVERSION_MAJOR_SHIFT,
505 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MINOR_MASK) >>
506 	    MPI2_IOCFACTS_MSGVERSION_MINOR_SHIFT);
507 
508 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
509 	    MPS_DRIVER_VERSION);
510 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
511 	     sc->facts->IOCCapabilities,
512 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
513 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
514 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
515 
516 	/*
517 	 * If the chip doesn't support event replay then a hard reset will be
518 	 * required to trigger a full discovery.  Do the reset here then
519 	 * retransition to Ready.  A hard reset might have already been done,
520 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
521 	 * for a Diag Reset.
522 	 */
523 	if (attaching && ((sc->facts->IOCCapabilities &
524 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
525 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
526 		mps_diag_reset(sc, NO_SLEEP);
527 		if ((error = mps_transition_ready(sc)) != 0) {
528 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
529 			    "transition to ready with error %d, exit\n",
530 			    error);
531 			return (error);
532 		}
533 	}
534 
535 	/*
536 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
537 	 * changed from the previous IOC Facts, log a warning, but only if
538 	 * checking this after a Diag Reset and not during attach.
539 	 */
540 	saved_mode = sc->ir_firmware;
541 	if (sc->facts->IOCCapabilities &
542 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
543 		sc->ir_firmware = 1;
544 	if (!attaching) {
545 		if (sc->ir_firmware != saved_mode) {
546 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
547 			    "in IOC Facts does not match previous mode\n");
548 		}
549 	}
550 
551 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
552 	reallocating = FALSE;
553 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
554 
555 	if ((!attaching) &&
556 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
557 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
558 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
559 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
560 	    (saved_facts.ProductID != sc->facts->ProductID) ||
561 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
562 	    (saved_facts.IOCRequestFrameSize !=
563 	    sc->facts->IOCRequestFrameSize) ||
564 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
565 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
566 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
567 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
568 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
569 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
570 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
571 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
572 	    (saved_facts.MaxPersistentEntries !=
573 	    sc->facts->MaxPersistentEntries))) {
574 		reallocating = TRUE;
575 
576 		/* Record that we reallocated everything */
577 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
578 	}
579 
580 	/*
581 	 * Some things should be done if attaching or re-allocating after a Diag
582 	 * Reset, but are not needed after a Diag Reset if the FW has not
583 	 * changed.
584 	 */
585 	if (attaching || reallocating) {
586 		/*
587 		 * Check if controller supports FW diag buffers and set flag to
588 		 * enable each type.
589 		 */
590 		if (sc->facts->IOCCapabilities &
591 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
592 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
593 			    enabled = TRUE;
594 		if (sc->facts->IOCCapabilities &
595 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
596 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
597 			    enabled = TRUE;
598 		if (sc->facts->IOCCapabilities &
599 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
600 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
601 			    enabled = TRUE;
602 
603 		/*
604 		 * Set flag if EEDP is supported and if TLR is supported.
605 		 */
606 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
607 			sc->eedp_enabled = TRUE;
608 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
609 			sc->control_TLR = TRUE;
610 
611 		mps_resize_queues(sc);
612 
613 		/*
614 		 * Initialize all Tail Queues
615 		 */
616 		TAILQ_INIT(&sc->req_list);
617 		TAILQ_INIT(&sc->high_priority_req_list);
618 		TAILQ_INIT(&sc->chain_list);
619 		TAILQ_INIT(&sc->tm_list);
620 	}
621 
622 	/*
623 	 * If doing a Diag Reset and the FW is significantly different
624 	 * (reallocating will be set above in IOC Facts comparison), then all
625 	 * buffers based on the IOC Facts will need to be freed before they are
626 	 * reallocated.
627 	 */
628 	if (reallocating) {
629 		mps_iocfacts_free(sc);
630 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
631 		    saved_facts.MaxVolumes);
632 	}
633 
634 	/*
635 	 * Any deallocation has been completed.  Now start reallocating
636 	 * if needed.  Will only need to reallocate if attaching or if the new
637 	 * IOC Facts are different from the previous IOC Facts after a Diag
638 	 * Reset. Targets have already been allocated above if needed.
639 	 */
640 	error = 0;
641 	while (attaching || reallocating) {
642 		if ((error = mps_alloc_hw_queues(sc)) != 0)
643 			break;
644 		if ((error = mps_alloc_replies(sc)) != 0)
645 			break;
646 		if ((error = mps_alloc_requests(sc)) != 0)
647 			break;
648 		if ((error = mps_alloc_queues(sc)) != 0)
649 			break;
650 
651 		break;
652 	}
653 	if (error) {
654 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
655 		    "Failed to alloc queues with error %d\n", error);
656 		mps_free(sc);
657 		return (error);
658 	}
659 
660 	/* Always initialize the queues */
661 	bzero(sc->free_queue, sc->fqdepth * 4);
662 	mps_init_queues(sc);
663 
664 	/*
665 	 * Always get the chip out of the reset state, but only panic if not
666 	 * attaching.  If attaching and there is an error, that is handled by
667 	 * the OS.
668 	 */
669 	error = mps_transition_operational(sc);
670 	if (error != 0) {
671 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
672 		    "transition to operational with error %d\n", error);
673 		mps_free(sc);
674 		return (error);
675 	}
676 
677 	/*
678 	 * Finish the queue initialization.
679 	 * These are set here instead of in mps_init_queues() because the
680 	 * IOC resets these values during the state transition in
681 	 * mps_transition_operational().  The free index is set to 1
682 	 * because the corresponding index in the IOC is set to 0, and the
683 	 * IOC treats the queues as full if both are set to the same value.
684 	 * Hence the reason that the queue can't hold all of the possible
685 	 * replies.
686 	 */
687 	sc->replypostindex = 0;
688 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
689 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
690 
691 	/*
692 	 * Attach the subsystems so they can prepare their event masks.
693 	 * XXX Should be dynamic so that IM/IR and user modules can attach
694 	 */
695 	error = 0;
696 	while (attaching) {
697 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
698 		if ((error = mps_attach_log(sc)) != 0)
699 			break;
700 		if ((error = mps_attach_sas(sc)) != 0)
701 			break;
702 		if ((error = mps_attach_user(sc)) != 0)
703 			break;
704 		break;
705 	}
706 	if (error) {
707 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
708 		    "subsystems: error %d\n", error);
709 		mps_free(sc);
710 		return (error);
711 	}
712 
713 	/*
714 	 * XXX If the number of MSI-X vectors changes during re-init, this
715 	 * won't see it and adjust.
716 	 */
717 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
718 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
719 		    "interrupts\n");
720 		mps_free(sc);
721 		return (error);
722 	}
723 
724 	/*
725 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
726 	 * reset it after a Diag Reset, just in case.
727 	 */
728 	sc->WD_available = FALSE;
729 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
730 		sc->WD_available = TRUE;
731 
732 	return (error);
733 }
734 
735 /*
736  * This is called if memory is being free (during detach for example) and when
737  * buffers need to be reallocated due to a Diag Reset.
738  */
739 static void
740 mps_iocfacts_free(struct mps_softc *sc)
741 {
742 	struct mps_command *cm;
743 	int i;
744 
745 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
746 
747 	if (sc->free_busaddr != 0)
748 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
749 	if (sc->free_queue != NULL)
750 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
751 		    sc->queues_map);
752 	if (sc->queues_dmat != NULL)
753 		bus_dma_tag_destroy(sc->queues_dmat);
754 
755 	if (sc->chain_frames != NULL) {
756 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
757 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
758 		    sc->chain_map);
759 	}
760 	if (sc->chain_dmat != NULL)
761 		bus_dma_tag_destroy(sc->chain_dmat);
762 
763 	if (sc->sense_busaddr != 0)
764 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
765 	if (sc->sense_frames != NULL)
766 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
767 		    sc->sense_map);
768 	if (sc->sense_dmat != NULL)
769 		bus_dma_tag_destroy(sc->sense_dmat);
770 
771 	if (sc->reply_busaddr != 0)
772 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
773 	if (sc->reply_frames != NULL)
774 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
775 		    sc->reply_map);
776 	if (sc->reply_dmat != NULL)
777 		bus_dma_tag_destroy(sc->reply_dmat);
778 
779 	if (sc->req_busaddr != 0)
780 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
781 	if (sc->req_frames != NULL)
782 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
783 	if (sc->req_dmat != NULL)
784 		bus_dma_tag_destroy(sc->req_dmat);
785 
786 	if (sc->chains != NULL)
787 		free(sc->chains, M_MPT2);
788 	if (sc->commands != NULL) {
789 		for (i = 1; i < sc->num_reqs; i++) {
790 			cm = &sc->commands[i];
791 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
792 		}
793 		free(sc->commands, M_MPT2);
794 	}
795 	if (sc->buffer_dmat != NULL)
796 		bus_dma_tag_destroy(sc->buffer_dmat);
797 
798 	mps_pci_free_interrupts(sc);
799 	free(sc->queues, M_MPT2);
800 	sc->queues = NULL;
801 }
802 
803 /*
804  * The terms diag reset and hard reset are used interchangeably in the MPI
805  * docs to mean resetting the controller chip.  In this code diag reset
806  * cleans everything up, and the hard reset function just sends the reset
807  * sequence to the chip.  This should probably be refactored so that every
808  * subsystem gets a reset notification of some sort, and can clean up
809  * appropriately.
810  */
811 int
812 mps_reinit(struct mps_softc *sc)
813 {
814 	int error;
815 	struct mpssas_softc *sassc;
816 
817 	sassc = sc->sassc;
818 
819 	MPS_FUNCTRACE(sc);
820 
821 	mtx_assert(&sc->mps_mtx, MA_OWNED);
822 
823 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
824 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
825 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
826 		return 0;
827 	}
828 
829 	/* make sure the completion callbacks can recognize they're getting
830 	 * a NULL cm_reply due to a reset.
831 	 */
832 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
833 
834 	/*
835 	 * Mask interrupts here.
836 	 */
837 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
838 	mps_mask_intr(sc);
839 
840 	error = mps_diag_reset(sc, CAN_SLEEP);
841 	if (error != 0) {
842 		/* XXXSL No need to panic here */
843 		panic("%s hard reset failed with error %d\n",
844 		    __func__, error);
845 	}
846 
847 	/* Restore the PCI state, including the MSI-X registers */
848 	mps_pci_restore(sc);
849 
850 	/* Give the I/O subsystem special priority to get itself prepared */
851 	mpssas_handle_reinit(sc);
852 
853 	/*
854 	 * Get IOC Facts and allocate all structures based on this information.
855 	 * The attach function will also call mps_iocfacts_allocate at startup.
856 	 * If relevant values have changed in IOC Facts, this function will free
857 	 * all of the memory based on IOC Facts and reallocate that memory.
858 	 */
859 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
860 		panic("%s IOC Facts based allocation failed with error %d\n",
861 		    __func__, error);
862 	}
863 
864 	/*
865 	 * Mapping structures will be re-allocated after getting IOC Page8, so
866 	 * free these structures here.
867 	 */
868 	mps_mapping_exit(sc);
869 
870 	/*
871 	 * The static page function currently read is IOC Page8.  Others can be
872 	 * added in future.  It's possible that the values in IOC Page8 have
873 	 * changed after a Diag Reset due to user modification, so always read
874 	 * these.  Interrupts are masked, so unmask them before getting config
875 	 * pages.
876 	 */
877 	mps_unmask_intr(sc);
878 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
879 	mps_base_static_config_pages(sc);
880 
881 	/*
882 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
883 	 * mapping tables.
884 	 */
885 	mps_mapping_initialize(sc);
886 
887 	/*
888 	 * Restart will reload the event masks clobbered by the reset, and
889 	 * then enable the port.
890 	 */
891 	mps_reregister_events(sc);
892 
893 	/* the end of discovery will release the simq, so we're done. */
894 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
895 	    sc, sc->replypostindex, sc->replyfreeindex);
896 
897 	mpssas_release_simq_reinit(sassc);
898 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
899 
900 	return 0;
901 }
902 
903 /* Wait for the chip to ACK a word that we've put into its FIFO
904  * Wait for <timeout> seconds. In single loop wait for busy loop
905  * for 500 microseconds.
906  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
907  * */
908 static int
909 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
910 {
911 
912 	u32 cntdn, count;
913 	u32 int_status;
914 	u32 doorbell;
915 
916 	count = 0;
917 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
918 	do {
919 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
920 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
921 			mps_dprint(sc, MPS_TRACE,
922 			"%s: successful count(%d), timeout(%d)\n",
923 			__func__, count, timeout);
924 		return 0;
925 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
926 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
927 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
928 				MPI2_IOC_STATE_FAULT) {
929 				mps_dprint(sc, MPS_FAULT,
930 					"fault_state(0x%04x)!\n", doorbell);
931 				return (EFAULT);
932 			}
933 		} else if (int_status == 0xFFFFFFFF)
934 			goto out;
935 
936 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
937 		* 0.5 milisecond */
938 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
939 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
940 			"mpsdba", hz/1000);
941 		else if (sleep_flag == CAN_SLEEP)
942 			pause("mpsdba", hz/1000);
943 		else
944 			DELAY(500);
945 		count++;
946 	} while (--cntdn);
947 
948 	out:
949 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
950 		"int_status(%x)!\n", __func__, count, int_status);
951 	return (ETIMEDOUT);
952 
953 }
954 
955 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
956 static int
957 mps_wait_db_int(struct mps_softc *sc)
958 {
959 	int retry;
960 
961 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
962 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
963 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
964 			return (0);
965 		DELAY(2000);
966 	}
967 	return (ETIMEDOUT);
968 }
969 
970 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
971 static int
972 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
973     int req_sz, int reply_sz, int timeout)
974 {
975 	uint32_t *data32;
976 	uint16_t *data16;
977 	int i, count, ioc_sz, residual;
978 	int sleep_flags = CAN_SLEEP;
979 
980 	if (curthread->td_no_sleeping != 0)
981 		sleep_flags = NO_SLEEP;
982 
983 	/* Step 1 */
984 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
985 
986 	/* Step 2 */
987 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
988 		return (EBUSY);
989 
990 	/* Step 3
991 	 * Announce that a message is coming through the doorbell.  Messages
992 	 * are pushed at 32bit words, so round up if needed.
993 	 */
994 	count = (req_sz + 3) / 4;
995 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
996 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
997 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
998 
999 	/* Step 4 */
1000 	if (mps_wait_db_int(sc) ||
1001 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
1002 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
1003 		return (ENXIO);
1004 	}
1005 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1006 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
1007 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
1008 		return (ENXIO);
1009 	}
1010 
1011 	/* Step 5 */
1012 	/* Clock out the message data synchronously in 32-bit dwords*/
1013 	data32 = (uint32_t *)req;
1014 	for (i = 0; i < count; i++) {
1015 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1016 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
1017 			mps_dprint(sc, MPS_FAULT,
1018 			    "Timeout while writing doorbell\n");
1019 			return (ENXIO);
1020 		}
1021 	}
1022 
1023 	/* Step 6 */
1024 	/* Clock in the reply in 16-bit words.  The total length of the
1025 	 * message is always in the 4th byte, so clock out the first 2 words
1026 	 * manually, then loop the rest.
1027 	 */
1028 	data16 = (uint16_t *)reply;
1029 	if (mps_wait_db_int(sc) != 0) {
1030 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
1031 		return (ENXIO);
1032 	}
1033 	data16[0] =
1034 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1035 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1036 	if (mps_wait_db_int(sc) != 0) {
1037 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
1038 		return (ENXIO);
1039 	}
1040 	data16[1] =
1041 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1042 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1043 
1044 	/* Number of 32bit words in the message */
1045 	ioc_sz = reply->MsgLength;
1046 
1047 	/*
1048 	 * Figure out how many 16bit words to clock in without overrunning.
1049 	 * The precision loss with dividing reply_sz can safely be
1050 	 * ignored because the messages can only be multiples of 32bits.
1051 	 */
1052 	residual = 0;
1053 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1054 	if (count < ioc_sz * 2) {
1055 		residual = ioc_sz * 2 - count;
1056 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1057 		    "residual message words\n", residual);
1058 	}
1059 
1060 	for (i = 2; i < count; i++) {
1061 		if (mps_wait_db_int(sc) != 0) {
1062 			mps_dprint(sc, MPS_FAULT,
1063 			    "Timeout reading doorbell %d\n", i);
1064 			return (ENXIO);
1065 		}
1066 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1067 		    MPI2_DOORBELL_DATA_MASK;
1068 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1069 	}
1070 
1071 	/*
1072 	 * Pull out residual words that won't fit into the provided buffer.
1073 	 * This keeps the chip from hanging due to a driver programming
1074 	 * error.
1075 	 */
1076 	while (residual--) {
1077 		if (mps_wait_db_int(sc) != 0) {
1078 			mps_dprint(sc, MPS_FAULT,
1079 			    "Timeout reading doorbell\n");
1080 			return (ENXIO);
1081 		}
1082 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1083 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1084 	}
1085 
1086 	/* Step 7 */
1087 	if (mps_wait_db_int(sc) != 0) {
1088 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1089 		return (ENXIO);
1090 	}
1091 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1092 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1093 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1094 
1095 	return (0);
1096 }
1097 
1098 static void
1099 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1100 {
1101 	reply_descriptor rd;
1102 	MPS_FUNCTRACE(sc);
1103 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1104 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1105 
1106 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1107 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1108 
1109 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1110 		sc->io_cmds_highwater++;
1111 	rd.u.low = cm->cm_desc.Words.Low;
1112 	rd.u.high = cm->cm_desc.Words.High;
1113 	rd.word = htole64(rd.word);
1114 
1115 	KASSERT(cm->cm_state == MPS_CM_STATE_BUSY, ("command not busy\n"));
1116 	cm->cm_state = MPS_CM_STATE_INQUEUE;
1117 
1118 	/* TODO-We may need to make below regwrite atomic */
1119 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1120 	    rd.u.low);
1121 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1122 	    rd.u.high);
1123 }
1124 
1125 /*
1126  * Just the FACTS, ma'am.
1127  */
1128 static int
1129 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1130 {
1131 	MPI2_DEFAULT_REPLY *reply;
1132 	MPI2_IOC_FACTS_REQUEST request;
1133 	int error, req_sz, reply_sz;
1134 
1135 	MPS_FUNCTRACE(sc);
1136 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1137 
1138 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1139 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1140 	reply = (MPI2_DEFAULT_REPLY *)facts;
1141 
1142 	bzero(&request, req_sz);
1143 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1144 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1145 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1146 
1147 	return (error);
1148 }
1149 
1150 static int
1151 mps_send_iocinit(struct mps_softc *sc)
1152 {
1153 	MPI2_IOC_INIT_REQUEST	init;
1154 	MPI2_DEFAULT_REPLY	reply;
1155 	int req_sz, reply_sz, error;
1156 	struct timeval now;
1157 	uint64_t time_in_msec;
1158 
1159 	MPS_FUNCTRACE(sc);
1160 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1161 
1162 	/* Do a quick sanity check on proper initialization */
1163 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1164 	    || (sc->replyframesz == 0)) {
1165 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1166 		    "Driver not fully initialized for IOCInit\n");
1167 		return (EINVAL);
1168 	}
1169 
1170 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1171 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1172 	bzero(&init, req_sz);
1173 	bzero(&reply, reply_sz);
1174 
1175 	/*
1176 	 * Fill in the init block.  Note that most addresses are
1177 	 * deliberately in the lower 32bits of memory.  This is a micro-
1178 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1179 	 */
1180 	init.Function = MPI2_FUNCTION_IOC_INIT;
1181 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1182 	init.MsgVersion = htole16(MPI2_VERSION);
1183 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1184 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1185 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1186 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1187 	init.SenseBufferAddressHigh = 0;
1188 	init.SystemReplyAddressHigh = 0;
1189 	init.SystemRequestFrameBaseAddress.High = 0;
1190 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1191 	init.ReplyDescriptorPostQueueAddress.High = 0;
1192 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1193 	init.ReplyFreeQueueAddress.High = 0;
1194 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1195 	getmicrotime(&now);
1196 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1197 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1198 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1199 
1200 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1201 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1202 		error = ENXIO;
1203 
1204 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1205 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1206 	return (error);
1207 }
1208 
1209 void
1210 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1211 {
1212 	bus_addr_t *addr;
1213 
1214 	addr = arg;
1215 	*addr = segs[0].ds_addr;
1216 }
1217 
1218 void
1219 mps_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1220 {
1221 	struct mps_busdma_context *ctx;
1222 	int need_unload, need_free;
1223 
1224 	ctx = (struct mps_busdma_context *)arg;
1225 	need_unload = 0;
1226 	need_free = 0;
1227 
1228 	mps_lock(ctx->softc);
1229 	ctx->error = error;
1230 	ctx->completed = 1;
1231 	if ((error == 0) && (ctx->abandoned == 0)) {
1232 		*ctx->addr = segs[0].ds_addr;
1233 	} else {
1234 		if (nsegs != 0)
1235 			need_unload = 1;
1236 		if (ctx->abandoned != 0)
1237 			need_free = 1;
1238 	}
1239 	if (need_free == 0)
1240 		wakeup(ctx);
1241 
1242 	mps_unlock(ctx->softc);
1243 
1244 	if (need_unload != 0) {
1245 		bus_dmamap_unload(ctx->buffer_dmat,
1246 				  ctx->buffer_dmamap);
1247 		*ctx->addr = 0;
1248 	}
1249 
1250 	if (need_free != 0)
1251 		free(ctx, M_MPSUSER);
1252 }
1253 
1254 static int
1255 mps_alloc_queues(struct mps_softc *sc)
1256 {
1257 	struct mps_queue *q;
1258 	u_int nq, i;
1259 
1260 	nq = sc->msi_msgs;
1261 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1262 
1263 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1264 	    M_NOWAIT|M_ZERO);
1265 	if (sc->queues == NULL)
1266 		return (ENOMEM);
1267 
1268 	for (i = 0; i < nq; i++) {
1269 		q = &sc->queues[i];
1270 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1271 		q->sc = sc;
1272 		q->qnum = i;
1273 	}
1274 
1275 	return (0);
1276 }
1277 
1278 static int
1279 mps_alloc_hw_queues(struct mps_softc *sc)
1280 {
1281 	bus_addr_t queues_busaddr;
1282 	uint8_t *queues;
1283 	int qsize, fqsize, pqsize;
1284 
1285 	/*
1286 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1287 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1288 	 * This queue supplies fresh reply frames for the firmware to use.
1289 	 *
1290 	 * The reply descriptor post queue contains 8 byte entries in
1291 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1292 	 * contains filled-in reply frames sent from the firmware to the host.
1293 	 *
1294 	 * These two queues are allocated together for simplicity.
1295 	 */
1296 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1297 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1298 	fqsize= sc->fqdepth * 4;
1299 	pqsize = sc->pqdepth * 8;
1300 	qsize = fqsize + pqsize;
1301 
1302         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1303 				16, 0,			/* algnmnt, boundary */
1304 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1305 				BUS_SPACE_MAXADDR,	/* highaddr */
1306 				NULL, NULL,		/* filter, filterarg */
1307                                 qsize,			/* maxsize */
1308                                 1,			/* nsegments */
1309                                 qsize,			/* maxsegsize */
1310                                 0,			/* flags */
1311                                 NULL, NULL,		/* lockfunc, lockarg */
1312                                 &sc->queues_dmat)) {
1313 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1314 		return (ENOMEM);
1315         }
1316         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1317 	    &sc->queues_map)) {
1318 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1319 		return (ENOMEM);
1320         }
1321         bzero(queues, qsize);
1322         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1323 	    mps_memaddr_cb, &queues_busaddr, 0);
1324 
1325 	sc->free_queue = (uint32_t *)queues;
1326 	sc->free_busaddr = queues_busaddr;
1327 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1328 	sc->post_busaddr = queues_busaddr + fqsize;
1329 	mps_dprint(sc, MPS_INIT, "free queue busaddr= %#016jx size= %d\n",
1330 	    (uintmax_t)sc->free_busaddr, fqsize);
1331 	mps_dprint(sc, MPS_INIT, "reply queue busaddr= %#016jx size= %d\n",
1332 	    (uintmax_t)sc->post_busaddr, pqsize);
1333 
1334 	return (0);
1335 }
1336 
1337 static int
1338 mps_alloc_replies(struct mps_softc *sc)
1339 {
1340 	int rsize, num_replies;
1341 
1342 	/* Store the reply frame size in bytes rather than as 32bit words */
1343 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1344 
1345 	/*
1346 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1347 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1348 	 * replies can be used at once.
1349 	 */
1350 	num_replies = max(sc->fqdepth, sc->num_replies);
1351 
1352 	rsize = sc->replyframesz * num_replies;
1353         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1354 				4, 0,			/* algnmnt, boundary */
1355 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1356 				BUS_SPACE_MAXADDR,	/* highaddr */
1357 				NULL, NULL,		/* filter, filterarg */
1358                                 rsize,			/* maxsize */
1359                                 1,			/* nsegments */
1360                                 rsize,			/* maxsegsize */
1361                                 0,			/* flags */
1362                                 NULL, NULL,		/* lockfunc, lockarg */
1363                                 &sc->reply_dmat)) {
1364 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1365 		return (ENOMEM);
1366         }
1367         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1368 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1369 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1370 		return (ENOMEM);
1371         }
1372         bzero(sc->reply_frames, rsize);
1373         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1374 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1375 
1376 	mps_dprint(sc, MPS_INIT, "reply frames busaddr= %#016jx size= %d\n",
1377 	    (uintmax_t)sc->reply_busaddr, rsize);
1378 
1379 	return (0);
1380 }
1381 
1382 static void
1383 mps_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1384 {
1385 	struct mps_softc *sc = arg;
1386 	struct mps_chain *chain;
1387 	bus_size_t bo;
1388 	int i, o, s;
1389 
1390 	if (error != 0)
1391 		return;
1392 
1393 	for (i = 0, o = 0, s = 0; s < nsegs; s++) {
1394 		for (bo = 0; bo + sc->reqframesz <= segs[s].ds_len;
1395 		    bo += sc->reqframesz) {
1396 			chain = &sc->chains[i++];
1397 			chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o);
1398 			chain->chain_busaddr = segs[s].ds_addr + bo;
1399 			o += sc->reqframesz;
1400 			mps_free_chain(sc, chain);
1401 		}
1402 		if (bo != segs[s].ds_len)
1403 			o += segs[s].ds_len - bo;
1404 	}
1405 	sc->chain_free_lowwater = i;
1406 }
1407 
1408 static int
1409 mps_alloc_requests(struct mps_softc *sc)
1410 {
1411 	struct mps_command *cm;
1412 	int i, rsize, nsegs;
1413 
1414 	rsize = sc->reqframesz * sc->num_reqs;
1415         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1416 				16, 0,			/* algnmnt, boundary */
1417 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1418 				BUS_SPACE_MAXADDR,	/* highaddr */
1419 				NULL, NULL,		/* filter, filterarg */
1420                                 rsize,			/* maxsize */
1421                                 1,			/* nsegments */
1422                                 rsize,			/* maxsegsize */
1423                                 0,			/* flags */
1424                                 NULL, NULL,		/* lockfunc, lockarg */
1425                                 &sc->req_dmat)) {
1426 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1427 		return (ENOMEM);
1428         }
1429         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1430 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1431 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1432 		return (ENOMEM);
1433         }
1434         bzero(sc->req_frames, rsize);
1435         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1436 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1437 	mps_dprint(sc, MPS_INIT, "request frames busaddr= %#016jx size= %d\n",
1438 	    (uintmax_t)sc->req_busaddr, rsize);
1439 
1440 	sc->chains = malloc(sizeof(struct mps_chain) * sc->num_chains, M_MPT2,
1441 	    M_NOWAIT | M_ZERO);
1442 	if (!sc->chains) {
1443 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1444 		return (ENOMEM);
1445 	}
1446 	rsize = sc->reqframesz * sc->num_chains;
1447 	if (bus_dma_tag_create( sc->mps_parent_dmat,	/* parent */
1448 				16, 0,			/* algnmnt, boundary */
1449 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1450 				BUS_SPACE_MAXADDR,	/* highaddr */
1451 				NULL, NULL,		/* filter, filterarg */
1452 				rsize,			/* maxsize */
1453 				howmany(rsize, PAGE_SIZE), /* nsegments */
1454 				rsize,			/* maxsegsize */
1455 				0,			/* flags */
1456 				NULL, NULL,		/* lockfunc, lockarg */
1457 				&sc->chain_dmat)) {
1458 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1459 		return (ENOMEM);
1460 	}
1461 	if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1462 	    BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) {
1463 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1464 		return (ENOMEM);
1465 	}
1466 	if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames,
1467 	    rsize, mps_load_chains_cb, sc, BUS_DMA_NOWAIT)) {
1468 		mps_dprint(sc, MPS_ERROR, "Cannot load chain memory\n");
1469 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
1470 		    sc->chain_map);
1471 		return (ENOMEM);
1472 	}
1473 
1474 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1475         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1476 				1, 0,			/* algnmnt, boundary */
1477 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1478 				BUS_SPACE_MAXADDR,	/* highaddr */
1479 				NULL, NULL,		/* filter, filterarg */
1480                                 rsize,			/* maxsize */
1481                                 1,			/* nsegments */
1482                                 rsize,			/* maxsegsize */
1483                                 0,			/* flags */
1484                                 NULL, NULL,		/* lockfunc, lockarg */
1485                                 &sc->sense_dmat)) {
1486 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1487 		return (ENOMEM);
1488         }
1489         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1490 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1491 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1492 		return (ENOMEM);
1493         }
1494         bzero(sc->sense_frames, rsize);
1495         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1496 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1497 	mps_dprint(sc, MPS_INIT, "sense frames busaddr= %#016jx size= %d\n",
1498 	    (uintmax_t)sc->sense_busaddr, rsize);
1499 
1500 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1501         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1502 				1, 0,			/* algnmnt, boundary */
1503 				BUS_SPACE_MAXADDR,	/* lowaddr */
1504 				BUS_SPACE_MAXADDR,	/* highaddr */
1505 				NULL, NULL,		/* filter, filterarg */
1506                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1507                                 nsegs,			/* nsegments */
1508                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1509                                 BUS_DMA_ALLOCNOW,	/* flags */
1510                                 busdma_lock_mutex,	/* lockfunc */
1511 				&sc->mps_mtx,		/* lockarg */
1512                                 &sc->buffer_dmat)) {
1513 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1514 		return (ENOMEM);
1515         }
1516 
1517 	/*
1518 	 * SMID 0 cannot be used as a free command per the firmware spec.
1519 	 * Just drop that command instead of risking accounting bugs.
1520 	 */
1521 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1522 	    M_MPT2, M_WAITOK | M_ZERO);
1523 	for (i = 1; i < sc->num_reqs; i++) {
1524 		cm = &sc->commands[i];
1525 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1526 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1527 		cm->cm_sense = &sc->sense_frames[i];
1528 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1529 		cm->cm_desc.Default.SMID = i;
1530 		cm->cm_sc = sc;
1531 		cm->cm_state = MPS_CM_STATE_BUSY;
1532 		TAILQ_INIT(&cm->cm_chain_list);
1533 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1534 
1535 		/* XXX Is a failure here a critical problem? */
1536 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1537 			if (i <= sc->num_prireqs)
1538 				mps_free_high_priority_command(sc, cm);
1539 			else
1540 				mps_free_command(sc, cm);
1541 		else {
1542 			panic("failed to allocate command %d\n", i);
1543 			sc->num_reqs = i;
1544 			break;
1545 		}
1546 	}
1547 
1548 	return (0);
1549 }
1550 
1551 static int
1552 mps_init_queues(struct mps_softc *sc)
1553 {
1554 	int i;
1555 
1556 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1557 
1558 	/*
1559 	 * According to the spec, we need to use one less reply than we
1560 	 * have space for on the queue.  So sc->num_replies (the number we
1561 	 * use) should be less than sc->fqdepth (allocated size).
1562 	 */
1563 	if (sc->num_replies >= sc->fqdepth)
1564 		return (EINVAL);
1565 
1566 	/*
1567 	 * Initialize all of the free queue entries.
1568 	 */
1569 	for (i = 0; i < sc->fqdepth; i++)
1570 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1571 	sc->replyfreeindex = sc->num_replies;
1572 
1573 	return (0);
1574 }
1575 
1576 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1577  * Next are the global settings, if they exist.  Highest are the per-unit
1578  * settings, if they exist.
1579  */
1580 void
1581 mps_get_tunables(struct mps_softc *sc)
1582 {
1583 	char tmpstr[80], mps_debug[80];
1584 
1585 	/* XXX default to some debugging for now */
1586 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1587 	sc->disable_msix = 0;
1588 	sc->disable_msi = 0;
1589 	sc->max_msix = MPS_MSIX_MAX;
1590 	sc->max_chains = MPS_CHAIN_FRAMES;
1591 	sc->max_io_pages = MPS_MAXIO_PAGES;
1592 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1593 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1594 	sc->use_phynum = 1;
1595 	sc->max_reqframes = MPS_REQ_FRAMES;
1596 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1597 	sc->max_replyframes = MPS_REPLY_FRAMES;
1598 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1599 
1600 	/*
1601 	 * Grab the global variables.
1602 	 */
1603 	bzero(mps_debug, 80);
1604 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1605 		mps_parse_debug(sc, mps_debug);
1606 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1607 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1608 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1609 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1610 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1611 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1612 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1613 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1614 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1615 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1616 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1617 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1618 
1619 	/* Grab the unit-instance variables */
1620 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1621 	    device_get_unit(sc->mps_dev));
1622 	bzero(mps_debug, 80);
1623 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1624 		mps_parse_debug(sc, mps_debug);
1625 
1626 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1627 	    device_get_unit(sc->mps_dev));
1628 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1629 
1630 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1631 	    device_get_unit(sc->mps_dev));
1632 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1633 
1634 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1635 	    device_get_unit(sc->mps_dev));
1636 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1637 
1638 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1639 	    device_get_unit(sc->mps_dev));
1640 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1641 
1642 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1643 	    device_get_unit(sc->mps_dev));
1644 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1645 
1646 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1647 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1648 	    device_get_unit(sc->mps_dev));
1649 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1650 
1651 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1652 	    device_get_unit(sc->mps_dev));
1653 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1654 
1655 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1656 	    device_get_unit(sc->mps_dev));
1657 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1658 
1659 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1660 	    device_get_unit(sc->mps_dev));
1661 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1662 
1663 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1664 	    device_get_unit(sc->mps_dev));
1665 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1666 
1667 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1668 	    device_get_unit(sc->mps_dev));
1669 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1670 
1671 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1672 	    device_get_unit(sc->mps_dev));
1673 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1674 
1675 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1676 	    device_get_unit(sc->mps_dev));
1677 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1678 
1679 }
1680 
1681 static void
1682 mps_setup_sysctl(struct mps_softc *sc)
1683 {
1684 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1685 	struct sysctl_oid	*sysctl_tree = NULL;
1686 	char tmpstr[80], tmpstr2[80];
1687 
1688 	/*
1689 	 * Setup the sysctl variable so the user can change the debug level
1690 	 * on the fly.
1691 	 */
1692 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1693 	    device_get_unit(sc->mps_dev));
1694 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1695 
1696 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1697 	if (sysctl_ctx != NULL)
1698 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1699 
1700 	if (sysctl_tree == NULL) {
1701 		sysctl_ctx_init(&sc->sysctl_ctx);
1702 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1703 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1704 		    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr);
1705 		if (sc->sysctl_tree == NULL)
1706 			return;
1707 		sysctl_ctx = &sc->sysctl_ctx;
1708 		sysctl_tree = sc->sysctl_tree;
1709 	}
1710 
1711 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1712 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1713 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1714 
1715 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1716 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1717 	    "Disable the use of MSI-X interrupts");
1718 
1719 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1720 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1721 	    "Disable the use of MSI interrupts");
1722 
1723 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1724 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1725 	    "User-defined maximum number of MSIX queues");
1726 
1727 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1728 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1729 	    "Negotiated number of MSIX queues");
1730 
1731 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1732 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1733 	    "Total number of allocated request frames");
1734 
1735 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1736 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1737 	    "Total number of allocated high priority request frames");
1738 
1739 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1740 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1741 	    "Total number of allocated reply frames");
1742 
1743 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1744 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1745 	    "Total number of event frames allocated");
1746 
1747 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1748 	    OID_AUTO, "firmware_version", CTLFLAG_RD, sc->fw_version,
1749 	    strlen(sc->fw_version), "firmware version");
1750 
1751 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1752 	    OID_AUTO, "driver_version", CTLFLAG_RD, MPS_DRIVER_VERSION,
1753 	    strlen(MPS_DRIVER_VERSION), "driver version");
1754 
1755 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1756 	    OID_AUTO, "msg_version", CTLFLAG_RD, sc->msg_version,
1757 	    strlen(sc->msg_version), "message interface version");
1758 
1759 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1760 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1761 	    &sc->io_cmds_active, 0, "number of currently active commands");
1762 
1763 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1764 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1765 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1766 
1767 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1768 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1769 	    &sc->chain_free, 0, "number of free chain elements");
1770 
1771 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1772 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1773 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1774 
1775 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1776 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1777 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1778 
1779 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1780 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1781 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1782 	    "IOCFacts)");
1783 
1784 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1785 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1786 	    "enable SSU to SATA SSD/HDD at shutdown");
1787 
1788 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1789 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1790 	    &sc->chain_alloc_fail, "chain allocation failures");
1791 
1792 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1793 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1794 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1795 	    "spinup after SATA ID error");
1796 
1797 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1798 	    OID_AUTO, "mapping_table_dump",
1799 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, sc, 0,
1800 	    mps_mapping_dump, "A", "Mapping Table Dump");
1801 
1802 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1803 	    OID_AUTO, "encl_table_dump",
1804 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, sc, 0,
1805 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1806 
1807 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1808 	    OID_AUTO, "dump_reqs",
1809 	    CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT,
1810 	    sc, 0, mps_dump_reqs, "I", "Dump Active Requests");
1811 
1812 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1813 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1814 	    "Use the phy number for enumeration");
1815 }
1816 
1817 static struct mps_debug_string {
1818 	char	*name;
1819 	int	flag;
1820 } mps_debug_strings[] = {
1821 	{"info", MPS_INFO},
1822 	{"fault", MPS_FAULT},
1823 	{"event", MPS_EVENT},
1824 	{"log", MPS_LOG},
1825 	{"recovery", MPS_RECOVERY},
1826 	{"error", MPS_ERROR},
1827 	{"init", MPS_INIT},
1828 	{"xinfo", MPS_XINFO},
1829 	{"user", MPS_USER},
1830 	{"mapping", MPS_MAPPING},
1831 	{"trace", MPS_TRACE}
1832 };
1833 
1834 enum mps_debug_level_combiner {
1835 	COMB_NONE,
1836 	COMB_ADD,
1837 	COMB_SUB
1838 };
1839 
1840 static int
1841 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1842 {
1843 	struct mps_softc *sc;
1844 	struct mps_debug_string *string;
1845 	struct sbuf *sbuf;
1846 	char *buffer;
1847 	size_t sz;
1848 	int i, len, debug, error;
1849 
1850 	sc = (struct mps_softc *)arg1;
1851 
1852 	error = sysctl_wire_old_buffer(req, 0);
1853 	if (error != 0)
1854 		return (error);
1855 
1856 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1857 	debug = sc->mps_debug;
1858 
1859 	sbuf_printf(sbuf, "%#x", debug);
1860 
1861 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1862 	for (i = 0; i < sz; i++) {
1863 		string = &mps_debug_strings[i];
1864 		if (debug & string->flag)
1865 			sbuf_printf(sbuf, ",%s", string->name);
1866 	}
1867 
1868 	error = sbuf_finish(sbuf);
1869 	sbuf_delete(sbuf);
1870 
1871 	if (error || req->newptr == NULL)
1872 		return (error);
1873 
1874 	len = req->newlen - req->newidx;
1875 	if (len == 0)
1876 		return (0);
1877 
1878 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1879 	error = SYSCTL_IN(req, buffer, len);
1880 
1881 	mps_parse_debug(sc, buffer);
1882 
1883 	free(buffer, M_MPT2);
1884 	return (error);
1885 }
1886 
1887 static void
1888 mps_parse_debug(struct mps_softc *sc, char *list)
1889 {
1890 	struct mps_debug_string *string;
1891 	enum mps_debug_level_combiner op;
1892 	char *token, *endtoken;
1893 	size_t sz;
1894 	int flags, i;
1895 
1896 	if (list == NULL || *list == '\0')
1897 		return;
1898 
1899 	if (*list == '+') {
1900 		op = COMB_ADD;
1901 		list++;
1902 	} else if (*list == '-') {
1903 		op = COMB_SUB;
1904 		list++;
1905 	} else
1906 		op = COMB_NONE;
1907 	if (*list == '\0')
1908 		return;
1909 
1910 	flags = 0;
1911 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1912 	while ((token = strsep(&list, ":,")) != NULL) {
1913 		/* Handle integer flags */
1914 		flags |= strtol(token, &endtoken, 0);
1915 		if (token != endtoken)
1916 			continue;
1917 
1918 		/* Handle text flags */
1919 		for (i = 0; i < sz; i++) {
1920 			string = &mps_debug_strings[i];
1921 			if (strcasecmp(token, string->name) == 0) {
1922 				flags |= string->flag;
1923 				break;
1924 			}
1925 		}
1926 	}
1927 
1928 	switch (op) {
1929 	case COMB_NONE:
1930 		sc->mps_debug = flags;
1931 		break;
1932 	case COMB_ADD:
1933 		sc->mps_debug |= flags;
1934 		break;
1935 	case COMB_SUB:
1936 		sc->mps_debug &= (~flags);
1937 		break;
1938 	}
1939 
1940 	return;
1941 }
1942 
1943 struct mps_dumpreq_hdr {
1944 	uint32_t	smid;
1945 	uint32_t	state;
1946 	uint32_t	numframes;
1947 	uint32_t	deschi;
1948 	uint32_t	desclo;
1949 };
1950 
1951 static int
1952 mps_dump_reqs(SYSCTL_HANDLER_ARGS)
1953 {
1954 	struct mps_softc *sc;
1955 	struct mps_chain *chain, *chain1;
1956 	struct mps_command *cm;
1957 	struct mps_dumpreq_hdr hdr;
1958 	struct sbuf *sb;
1959 	uint32_t smid, state;
1960 	int i, numreqs, error = 0;
1961 
1962 	sc = (struct mps_softc *)arg1;
1963 
1964 	if ((error = priv_check(curthread, PRIV_DRIVER)) != 0) {
1965 		printf("priv check error %d\n", error);
1966 		return (error);
1967 	}
1968 
1969 	state = MPS_CM_STATE_INQUEUE;
1970 	smid = 1;
1971 	numreqs = sc->num_reqs;
1972 
1973 	if (req->newptr != NULL)
1974 		return (EINVAL);
1975 
1976 	if (smid == 0 || smid > sc->num_reqs)
1977 		return (EINVAL);
1978 	if (numreqs <= 0 || (numreqs + smid > sc->num_reqs))
1979 		numreqs = sc->num_reqs;
1980 	sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req);
1981 
1982 	/* Best effort, no locking */
1983 	for (i = smid; i < numreqs; i++) {
1984 		cm = &sc->commands[i];
1985 		if (cm->cm_state != state)
1986 			continue;
1987 		hdr.smid = i;
1988 		hdr.state = cm->cm_state;
1989 		hdr.numframes = 1;
1990 		hdr.deschi = cm->cm_desc.Words.High;
1991 		hdr.desclo = cm->cm_desc.Words.Low;
1992 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
1993 		   chain1)
1994 			hdr.numframes++;
1995 		sbuf_bcat(sb, &hdr, sizeof(hdr));
1996 		sbuf_bcat(sb, cm->cm_req, 128);
1997 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
1998 		    chain1)
1999 			sbuf_bcat(sb, chain->chain, 128);
2000 	}
2001 
2002 	error = sbuf_finish(sb);
2003 	sbuf_delete(sb);
2004 	return (error);
2005 }
2006 
2007 int
2008 mps_attach(struct mps_softc *sc)
2009 {
2010 	int error;
2011 
2012 	MPS_FUNCTRACE(sc);
2013 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2014 
2015 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
2016 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
2017 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
2018 	TAILQ_INIT(&sc->event_list);
2019 	timevalclear(&sc->lastfail);
2020 
2021 	if ((error = mps_transition_ready(sc)) != 0) {
2022 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
2023 		    "ready\n");
2024 		return (error);
2025 	}
2026 
2027 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
2028 	    M_ZERO|M_NOWAIT);
2029 	if(!sc->facts) {
2030 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
2031 		    "exit\n");
2032 		return (ENOMEM);
2033 	}
2034 
2035 	/*
2036 	 * Get IOC Facts and allocate all structures based on this information.
2037 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
2038 	 * Facts. If relevant values have changed in IOC Facts, this function
2039 	 * will free all of the memory based on IOC Facts and reallocate that
2040 	 * memory.  If this fails, any allocated memory should already be freed.
2041 	 */
2042 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
2043 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
2044 		    "failed with error %d, exit\n", error);
2045 		return (error);
2046 	}
2047 
2048 	/* Start the periodic watchdog check on the IOC Doorbell */
2049 	mps_periodic(sc);
2050 
2051 	/*
2052 	 * The portenable will kick off discovery events that will drive the
2053 	 * rest of the initialization process.  The CAM/SAS module will
2054 	 * hold up the boot sequence until discovery is complete.
2055 	 */
2056 	sc->mps_ich.ich_func = mps_startup;
2057 	sc->mps_ich.ich_arg = sc;
2058 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
2059 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
2060 		    "Cannot establish MPS config hook\n");
2061 		error = EINVAL;
2062 	}
2063 
2064 	/*
2065 	 * Allow IR to shutdown gracefully when shutdown occurs.
2066 	 */
2067 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
2068 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
2069 
2070 	if (sc->shutdown_eh == NULL)
2071 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
2072 		    "shutdown event registration failed\n");
2073 
2074 	mps_setup_sysctl(sc);
2075 
2076 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
2077 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
2078 
2079 	return (error);
2080 }
2081 
2082 /* Run through any late-start handlers. */
2083 static void
2084 mps_startup(void *arg)
2085 {
2086 	struct mps_softc *sc;
2087 
2088 	sc = (struct mps_softc *)arg;
2089 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2090 
2091 	mps_lock(sc);
2092 	mps_unmask_intr(sc);
2093 
2094 	/* initialize device mapping tables */
2095 	mps_base_static_config_pages(sc);
2096 	mps_mapping_initialize(sc);
2097 	mpssas_startup(sc);
2098 	mps_unlock(sc);
2099 
2100 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
2101 	config_intrhook_disestablish(&sc->mps_ich);
2102 	sc->mps_ich.ich_arg = NULL;
2103 
2104 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2105 }
2106 
2107 /* Periodic watchdog.  Is called with the driver lock already held. */
2108 static void
2109 mps_periodic(void *arg)
2110 {
2111 	struct mps_softc *sc;
2112 	uint32_t db;
2113 
2114 	sc = (struct mps_softc *)arg;
2115 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
2116 		return;
2117 
2118 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
2119 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2120 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2121 		mps_reinit(sc);
2122 	}
2123 
2124 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
2125 }
2126 
2127 static void
2128 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
2129     MPI2_EVENT_NOTIFICATION_REPLY *event)
2130 {
2131 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2132 
2133 	MPS_DPRINT_EVENT(sc, generic, event);
2134 
2135 	switch (event->Event) {
2136 	case MPI2_EVENT_LOG_DATA:
2137 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2138 		if (sc->mps_debug & MPS_EVENT)
2139 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
2140 		break;
2141 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2142 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2143 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2144 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2145 		     entry->LogSequence);
2146 		break;
2147 	default:
2148 		break;
2149 	}
2150 	return;
2151 }
2152 
2153 static int
2154 mps_attach_log(struct mps_softc *sc)
2155 {
2156 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
2157 
2158 	bzero(events, 16);
2159 	setbit(events, MPI2_EVENT_LOG_DATA);
2160 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2161 
2162 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
2163 	    &sc->mps_log_eh);
2164 
2165 	return (0);
2166 }
2167 
2168 static int
2169 mps_detach_log(struct mps_softc *sc)
2170 {
2171 
2172 	if (sc->mps_log_eh != NULL)
2173 		mps_deregister_events(sc, sc->mps_log_eh);
2174 	return (0);
2175 }
2176 
2177 /*
2178  * Free all of the driver resources and detach submodules.  Should be called
2179  * without the lock held.
2180  */
2181 int
2182 mps_free(struct mps_softc *sc)
2183 {
2184 	int error;
2185 
2186 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2187 	/* Turn off the watchdog */
2188 	mps_lock(sc);
2189 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
2190 	mps_unlock(sc);
2191 	/* Lock must not be held for this */
2192 	callout_drain(&sc->periodic);
2193 	callout_drain(&sc->device_check_callout);
2194 
2195 	if (((error = mps_detach_log(sc)) != 0) ||
2196 	    ((error = mps_detach_sas(sc)) != 0)) {
2197 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2198 		    "subsystems, exit\n");
2199 		return (error);
2200 	}
2201 
2202 	mps_detach_user(sc);
2203 
2204 	/* Put the IOC back in the READY state. */
2205 	mps_lock(sc);
2206 	if ((error = mps_transition_ready(sc)) != 0) {
2207 		mps_unlock(sc);
2208 		return (error);
2209 	}
2210 	mps_unlock(sc);
2211 
2212 	if (sc->facts != NULL)
2213 		free(sc->facts, M_MPT2);
2214 
2215 	/*
2216 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2217 	 * to free these buffers too.
2218 	 */
2219 	mps_iocfacts_free(sc);
2220 
2221 	if (sc->sysctl_tree != NULL)
2222 		sysctl_ctx_free(&sc->sysctl_ctx);
2223 
2224 	/* Deregister the shutdown function */
2225 	if (sc->shutdown_eh != NULL)
2226 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2227 
2228 	mtx_destroy(&sc->mps_mtx);
2229 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2230 
2231 	return (0);
2232 }
2233 
2234 static __inline void
2235 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2236 {
2237 	MPS_FUNCTRACE(sc);
2238 
2239 	if (cm == NULL) {
2240 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2241 		return;
2242 	}
2243 
2244 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2245 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2246 
2247 	if (cm->cm_complete != NULL) {
2248 		mps_dprint(sc, MPS_TRACE,
2249 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2250 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2251 			   cm->cm_reply);
2252 		cm->cm_complete(sc, cm);
2253 	}
2254 
2255 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2256 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2257 		wakeup(cm);
2258 	}
2259 
2260 	if (cm->cm_sc->io_cmds_active != 0) {
2261 		cm->cm_sc->io_cmds_active--;
2262 	} else {
2263 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2264 		    "out of sync - resynching to 0\n");
2265 	}
2266 }
2267 
2268 static void
2269 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2270 {
2271 	union loginfo_type {
2272 		u32     loginfo;
2273 		struct {
2274 			u32     subcode:16;
2275 			u32     code:8;
2276 			u32     originator:4;
2277 			u32     bus_type:4;
2278 		} dw;
2279 	};
2280 	union loginfo_type sas_loginfo;
2281 	char *originator_str = NULL;
2282 
2283 	sas_loginfo.loginfo = log_info;
2284 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2285 		return;
2286 
2287 	/* each nexus loss loginfo */
2288 	if (log_info == 0x31170000)
2289 		return;
2290 
2291 	/* eat the loginfos associated with task aborts */
2292 	if ((log_info == 30050000 || log_info ==
2293 	    0x31140000 || log_info == 0x31130000))
2294 		return;
2295 
2296 	switch (sas_loginfo.dw.originator) {
2297 	case 0:
2298 		originator_str = "IOP";
2299 		break;
2300 	case 1:
2301 		originator_str = "PL";
2302 		break;
2303 	case 2:
2304 		originator_str = "IR";
2305 		break;
2306 }
2307 
2308 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2309 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2310 	originator_str, sas_loginfo.dw.code,
2311 	sas_loginfo.dw.subcode);
2312 }
2313 
2314 static void
2315 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2316 {
2317 	MPI2DefaultReply_t *mpi_reply;
2318 	u16 sc_status;
2319 
2320 	mpi_reply = (MPI2DefaultReply_t*)reply;
2321 	sc_status = le16toh(mpi_reply->IOCStatus);
2322 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2323 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2324 }
2325 void
2326 mps_intr(void *data)
2327 {
2328 	struct mps_softc *sc;
2329 	uint32_t status;
2330 
2331 	sc = (struct mps_softc *)data;
2332 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2333 
2334 	/*
2335 	 * Check interrupt status register to flush the bus.  This is
2336 	 * needed for both INTx interrupts and driver-driven polling
2337 	 */
2338 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2339 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2340 		return;
2341 
2342 	mps_lock(sc);
2343 	mps_intr_locked(data);
2344 	mps_unlock(sc);
2345 	return;
2346 }
2347 
2348 /*
2349  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2350  * chip.  Hopefully this theory is correct.
2351  */
2352 void
2353 mps_intr_msi(void *data)
2354 {
2355 	struct mps_softc *sc;
2356 
2357 	sc = (struct mps_softc *)data;
2358 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2359 	mps_lock(sc);
2360 	mps_intr_locked(data);
2361 	mps_unlock(sc);
2362 	return;
2363 }
2364 
2365 /*
2366  * The locking is overly broad and simplistic, but easy to deal with for now.
2367  */
2368 void
2369 mps_intr_locked(void *data)
2370 {
2371 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2372 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2373 	mps_fw_diagnostic_buffer_t *pBuffer;
2374 	struct mps_softc *sc;
2375 	struct mps_command *cm = NULL;
2376 	uint64_t tdesc;
2377 	uint8_t flags;
2378 	u_int pq;
2379 
2380 	sc = (struct mps_softc *)data;
2381 
2382 	pq = sc->replypostindex;
2383 	mps_dprint(sc, MPS_TRACE,
2384 	    "%s sc %p starting with replypostindex %u\n",
2385 	    __func__, sc, sc->replypostindex);
2386 
2387 	for ( ;; ) {
2388 		cm = NULL;
2389 		desc = &sc->post_queue[sc->replypostindex];
2390 
2391 		/*
2392 		 * Copy and clear out the descriptor so that any reentry will
2393 		 * immediately know that this descriptor has already been
2394 		 * looked at.  There is unfortunate casting magic because the
2395 		 * MPI API doesn't have a cardinal 64bit type.
2396 		 */
2397 		tdesc = 0xffffffffffffffff;
2398 		tdesc = atomic_swap_64((uint64_t *)desc, tdesc);
2399 		desc = (MPI2_REPLY_DESCRIPTORS_UNION *)&tdesc;
2400 
2401 		flags = desc->Default.ReplyFlags &
2402 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2403 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2404 		 || (le32toh(desc->Words.High) == 0xffffffff))
2405 			break;
2406 
2407 		/* increment the replypostindex now, so that event handlers
2408 		 * and cm completion handlers which decide to do a diag
2409 		 * reset can zero it without it getting incremented again
2410 		 * afterwards, and we break out of this loop on the next
2411 		 * iteration since the reply post queue has been cleared to
2412 		 * 0xFF and all descriptors look unused (which they are).
2413 		 */
2414 		if (++sc->replypostindex >= sc->pqdepth)
2415 			sc->replypostindex = 0;
2416 
2417 		switch (flags) {
2418 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2419 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2420 			KASSERT(cm->cm_state == MPS_CM_STATE_INQUEUE,
2421 			    ("command not inqueue\n"));
2422 			cm->cm_state = MPS_CM_STATE_BUSY;
2423 			cm->cm_reply = NULL;
2424 			break;
2425 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2426 		{
2427 			uint32_t baddr;
2428 			uint8_t *reply;
2429 
2430 			/*
2431 			 * Re-compose the reply address from the address
2432 			 * sent back from the chip.  The ReplyFrameAddress
2433 			 * is the lower 32 bits of the physical address of
2434 			 * particular reply frame.  Convert that address to
2435 			 * host format, and then use that to provide the
2436 			 * offset against the virtual address base
2437 			 * (sc->reply_frames).
2438 			 */
2439 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2440 			reply = sc->reply_frames +
2441 				(baddr - ((uint32_t)sc->reply_busaddr));
2442 			/*
2443 			 * Make sure the reply we got back is in a valid
2444 			 * range.  If not, go ahead and panic here, since
2445 			 * we'll probably panic as soon as we deference the
2446 			 * reply pointer anyway.
2447 			 */
2448 			if ((reply < sc->reply_frames)
2449 			 || (reply > (sc->reply_frames +
2450 			     (sc->fqdepth * sc->replyframesz)))) {
2451 				printf("%s: WARNING: reply %p out of range!\n",
2452 				       __func__, reply);
2453 				printf("%s: reply_frames %p, fqdepth %d, "
2454 				       "frame size %d\n", __func__,
2455 				       sc->reply_frames, sc->fqdepth,
2456 				       sc->replyframesz);
2457 				printf("%s: baddr %#x,\n", __func__, baddr);
2458 				/* LSI-TODO. See Linux Code for Graceful exit */
2459 				panic("Reply address out of range");
2460 			}
2461 			if (le16toh(desc->AddressReply.SMID) == 0) {
2462 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2463 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2464 					/*
2465 					 * If SMID is 0 for Diag Buffer Post,
2466 					 * this implies that the reply is due to
2467 					 * a release function with a status that
2468 					 * the buffer has been released.  Set
2469 					 * the buffer flags accordingly.
2470 					 */
2471 					rel_rep =
2472 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2473 					if ((le16toh(rel_rep->IOCStatus) &
2474 					    MPI2_IOCSTATUS_MASK) ==
2475 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2476 					{
2477 						pBuffer =
2478 						    &sc->fw_diag_buffer_list[
2479 						    rel_rep->BufferType];
2480 						pBuffer->valid_data = TRUE;
2481 						pBuffer->owned_by_firmware =
2482 						    FALSE;
2483 						pBuffer->immediate = FALSE;
2484 					}
2485 				} else
2486 					mps_dispatch_event(sc, baddr,
2487 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2488 					    reply);
2489 			} else {
2490 				/*
2491 				 * Ignore commands not in INQUEUE state
2492 				 * since they've already been completed
2493 				 * via another path.
2494 				 */
2495 				cm = &sc->commands[
2496 				    le16toh(desc->AddressReply.SMID)];
2497 				if (cm->cm_state == MPS_CM_STATE_INQUEUE) {
2498 					cm->cm_state = MPS_CM_STATE_BUSY;
2499 					cm->cm_reply = reply;
2500 					cm->cm_reply_data = le32toh(
2501 					    desc->AddressReply.ReplyFrameAddress);
2502 				} else {
2503 					mps_dprint(sc, MPS_RECOVERY,
2504 					    "Bad state for ADDRESS_REPLY status,"
2505 					    " ignoring state %d cm %p\n",
2506 					    cm->cm_state, cm);
2507 				}
2508 			}
2509 			break;
2510 		}
2511 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2512 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2513 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2514 		default:
2515 			/* Unhandled */
2516 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2517 			    desc->Default.ReplyFlags);
2518 			cm = NULL;
2519 			break;
2520 		}
2521 
2522 
2523 		if (cm != NULL) {
2524 			// Print Error reply frame
2525 			if (cm->cm_reply)
2526 				mps_display_reply_info(sc,cm->cm_reply);
2527 			mps_complete_command(sc, cm);
2528 		}
2529 	}
2530 
2531 	if (pq != sc->replypostindex) {
2532 		mps_dprint(sc, MPS_TRACE, "%s sc %p writing postindex %d\n",
2533 		    __func__, sc, sc->replypostindex);
2534 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2535 		    sc->replypostindex);
2536 	}
2537 
2538 	return;
2539 }
2540 
2541 static void
2542 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2543     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2544 {
2545 	struct mps_event_handle *eh;
2546 	int event, handled = 0;
2547 
2548 	event = le16toh(reply->Event);
2549 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2550 		if (isset(eh->mask, event)) {
2551 			eh->callback(sc, data, reply);
2552 			handled++;
2553 		}
2554 	}
2555 
2556 	if (handled == 0)
2557 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2558 
2559 	/*
2560 	 * This is the only place that the event/reply should be freed.
2561 	 * Anything wanting to hold onto the event data should have
2562 	 * already copied it into their own storage.
2563 	 */
2564 	mps_free_reply(sc, data);
2565 }
2566 
2567 static void
2568 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2569 {
2570 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2571 
2572 	if (cm->cm_reply)
2573 		MPS_DPRINT_EVENT(sc, generic,
2574 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2575 
2576 	mps_free_command(sc, cm);
2577 
2578 	/* next, send a port enable */
2579 	mpssas_startup(sc);
2580 }
2581 
2582 /*
2583  * For both register_events and update_events, the caller supplies a bitmap
2584  * of events that it _wants_.  These functions then turn that into a bitmask
2585  * suitable for the controller.
2586  */
2587 int
2588 mps_register_events(struct mps_softc *sc, u32 *mask,
2589     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2590 {
2591 	struct mps_event_handle *eh;
2592 	int error = 0;
2593 
2594 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2595 	eh->callback = cb;
2596 	eh->data = data;
2597 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2598 	if (mask != NULL)
2599 		error = mps_update_events(sc, eh, mask);
2600 	*handle = eh;
2601 
2602 	return (error);
2603 }
2604 
2605 int
2606 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2607     u32 *mask)
2608 {
2609 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2610 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2611 	struct mps_command *cm;
2612 	int error, i;
2613 
2614 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2615 
2616 	if ((mask != NULL) && (handle != NULL))
2617 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2618 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2619 
2620 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2621 		sc->event_mask[i] = -1;
2622 
2623 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2624 		sc->event_mask[i] &= ~handle->mask[i];
2625 
2626 	if ((cm = mps_alloc_command(sc)) == NULL)
2627 		return (EBUSY);
2628 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2629 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2630 	evtreq->MsgFlags = 0;
2631 	evtreq->SASBroadcastPrimitiveMasks = 0;
2632 #ifdef MPS_DEBUG_ALL_EVENTS
2633 	{
2634 		u_char fullmask[16];
2635 		memset(fullmask, 0x00, 16);
2636 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2637 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2638 	}
2639 #else
2640         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2641                 evtreq->EventMasks[i] =
2642                     htole32(sc->event_mask[i]);
2643 #endif
2644 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2645 	cm->cm_data = NULL;
2646 
2647 	error = mps_wait_command(sc, &cm, 60, 0);
2648 	if (cm != NULL)
2649 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2650 	if ((reply == NULL) ||
2651 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2652 		error = ENXIO;
2653 
2654 	if (reply)
2655 		MPS_DPRINT_EVENT(sc, generic, reply);
2656 
2657 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2658 
2659 	if (cm != NULL)
2660 		mps_free_command(sc, cm);
2661 	return (error);
2662 }
2663 
2664 static int
2665 mps_reregister_events(struct mps_softc *sc)
2666 {
2667 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2668 	struct mps_command *cm;
2669 	struct mps_event_handle *eh;
2670 	int error, i;
2671 
2672 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2673 
2674 	/* first, reregister events */
2675 
2676 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2677 		sc->event_mask[i] = -1;
2678 
2679 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2680 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2681 			sc->event_mask[i] &= ~eh->mask[i];
2682 	}
2683 
2684 	if ((cm = mps_alloc_command(sc)) == NULL)
2685 		return (EBUSY);
2686 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2687 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2688 	evtreq->MsgFlags = 0;
2689 	evtreq->SASBroadcastPrimitiveMasks = 0;
2690 #ifdef MPS_DEBUG_ALL_EVENTS
2691 	{
2692 		u_char fullmask[16];
2693 		memset(fullmask, 0x00, 16);
2694 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2695 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2696 	}
2697 #else
2698         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2699                 evtreq->EventMasks[i] =
2700                     htole32(sc->event_mask[i]);
2701 #endif
2702 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2703 	cm->cm_data = NULL;
2704 	cm->cm_complete = mps_reregister_events_complete;
2705 
2706 	error = mps_map_command(sc, cm);
2707 
2708 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2709 	    error);
2710 	return (error);
2711 }
2712 
2713 void
2714 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2715 {
2716 
2717 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2718 	free(handle, M_MPT2);
2719 }
2720 
2721 /*
2722  * Add a chain element as the next SGE for the specified command.
2723  * Reset cm_sge and cm_sgesize to indicate all the available space.
2724  */
2725 static int
2726 mps_add_chain(struct mps_command *cm)
2727 {
2728 	MPI2_SGE_CHAIN32 *sgc;
2729 	struct mps_chain *chain;
2730 	u_int space;
2731 
2732 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2733 		panic("MPS: Need SGE Error Code\n");
2734 
2735 	chain = mps_alloc_chain(cm->cm_sc);
2736 	if (chain == NULL)
2737 		return (ENOBUFS);
2738 
2739 	space = cm->cm_sc->reqframesz;
2740 
2741 	/*
2742 	 * Note: a double-linked list is used to make it easier to
2743 	 * walk for debugging.
2744 	 */
2745 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2746 
2747 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2748 	sgc->Length = htole16(space);
2749 	sgc->NextChainOffset = 0;
2750 	/* TODO Looks like bug in Setting sgc->Flags.
2751 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2752 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2753 	 *	This is fine.. because we are not using simple element. In case of
2754 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2755  	 */
2756 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2757 	sgc->Address = htole32(chain->chain_busaddr);
2758 
2759 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2760 	cm->cm_sglsize = space;
2761 	return (0);
2762 }
2763 
2764 /*
2765  * Add one scatter-gather element (chain, simple, transaction context)
2766  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2767  * cm_sge as the remaining size and pointer to the next SGE to fill
2768  * in, respectively.
2769  */
2770 int
2771 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2772 {
2773 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2774 	MPI2_SGE_SIMPLE64 *sge = sgep;
2775 	int error, type;
2776 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2777 
2778 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2779 
2780 #ifdef INVARIANTS
2781 	switch (type) {
2782 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2783 		if (len != tc->DetailsLength + 4)
2784 			panic("TC %p length %u or %zu?", tc,
2785 			    tc->DetailsLength + 4, len);
2786 		}
2787 		break;
2788 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2789 		/* Driver only uses 32-bit chain elements */
2790 		if (len != MPS_SGC_SIZE)
2791 			panic("CHAIN %p length %u or %zu?", sgep,
2792 			    MPS_SGC_SIZE, len);
2793 		break;
2794 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2795 		/* Driver only uses 64-bit SGE simple elements */
2796 		if (len != MPS_SGE64_SIZE)
2797 			panic("SGE simple %p length %u or %zu?", sge,
2798 			    MPS_SGE64_SIZE, len);
2799 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2800 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2801 			panic("SGE simple %p not marked 64-bit?", sge);
2802 
2803 		break;
2804 	default:
2805 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2806 	}
2807 #endif
2808 
2809 	/*
2810 	 * case 1: 1 more segment, enough room for it
2811 	 * case 2: 2 more segments, enough room for both
2812 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2813 	 * case 4: >=1 more segment, enough room for only a chain
2814 	 * case 5: >=1 more segment, no room for anything (error)
2815          */
2816 
2817 	/*
2818 	 * There should be room for at least a chain element, or this
2819 	 * code is buggy.  Case (5).
2820 	 */
2821 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2822 		panic("MPS: Need SGE Error Code\n");
2823 
2824 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2825 		/*
2826 		 * 1 or more segment, enough room for only a chain.
2827 		 * Hope the previous element wasn't a Simple entry
2828 		 * that needed to be marked with
2829 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2830 		 */
2831 		if ((error = mps_add_chain(cm)) != 0)
2832 			return (error);
2833 	}
2834 
2835 	if (segsleft >= 2 &&
2836 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2837 		/*
2838 		 * There are 2 or more segments left to add, and only
2839 		 * enough room for 1 and a chain.  Case (3).
2840 		 *
2841 		 * Mark as last element in this chain if necessary.
2842 		 */
2843 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2844 			sge->FlagsLength |= htole32(
2845 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2846 		}
2847 
2848 		/*
2849 		 * Add the item then a chain.  Do the chain now,
2850 		 * rather than on the next iteration, to simplify
2851 		 * understanding the code.
2852 		 */
2853 		cm->cm_sglsize -= len;
2854 		bcopy(sgep, cm->cm_sge, len);
2855 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2856 		return (mps_add_chain(cm));
2857 	}
2858 
2859 #ifdef INVARIANTS
2860 	/* Case 1: 1 more segment, enough room for it. */
2861 	if (segsleft == 1 && cm->cm_sglsize < len)
2862 		panic("1 seg left and no room? %u versus %zu",
2863 		    cm->cm_sglsize, len);
2864 
2865 	/* Case 2: 2 more segments, enough room for both */
2866 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2867 		panic("2 segs left and no room? %u versus %zu",
2868 		    cm->cm_sglsize, len);
2869 #endif
2870 
2871 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2872 		/*
2873 		 * If this is a bi-directional request, need to account for that
2874 		 * here.  Save the pre-filled sge values.  These will be used
2875 		 * either for the 2nd SGL or for a single direction SGL.  If
2876 		 * cm_out_len is non-zero, this is a bi-directional request, so
2877 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2878 		 * fill in the IN SGL.  Note that at this time, when filling in
2879 		 * 2 SGL's for a bi-directional request, they both use the same
2880 		 * DMA buffer (same cm command).
2881 		 */
2882 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2883 		saved_address_low = sge->Address.Low;
2884 		saved_address_high = sge->Address.High;
2885 		if (cm->cm_out_len) {
2886 			sge->FlagsLength = htole32(cm->cm_out_len |
2887 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2888 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2889 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2890 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2891 			    MPI2_SGE_FLAGS_SHIFT));
2892 			cm->cm_sglsize -= len;
2893 			bcopy(sgep, cm->cm_sge, len);
2894 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2895 			    + len);
2896 		}
2897 		saved_buf_len |=
2898 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2899 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2900 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2901 		    MPI2_SGE_FLAGS_END_OF_LIST |
2902 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2903 		    MPI2_SGE_FLAGS_SHIFT);
2904 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2905 			saved_buf_len |=
2906 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2907 			    MPI2_SGE_FLAGS_SHIFT);
2908 		} else {
2909 			saved_buf_len |=
2910 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2911 			    MPI2_SGE_FLAGS_SHIFT);
2912 		}
2913 		sge->FlagsLength = htole32(saved_buf_len);
2914 		sge->Address.Low = saved_address_low;
2915 		sge->Address.High = saved_address_high;
2916 	}
2917 
2918 	cm->cm_sglsize -= len;
2919 	bcopy(sgep, cm->cm_sge, len);
2920 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2921 	return (0);
2922 }
2923 
2924 /*
2925  * Add one dma segment to the scatter-gather list for a command.
2926  */
2927 int
2928 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2929     int segsleft)
2930 {
2931 	MPI2_SGE_SIMPLE64 sge;
2932 
2933 	/*
2934 	 * This driver always uses 64-bit address elements for simplicity.
2935 	 */
2936 	bzero(&sge, sizeof(sge));
2937 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2938 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2939 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2940 	mps_from_u64(pa, &sge.Address);
2941 
2942 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2943 }
2944 
2945 static void
2946 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2947 {
2948 	struct mps_softc *sc;
2949 	struct mps_command *cm;
2950 	u_int i, dir, sflags;
2951 
2952 	cm = (struct mps_command *)arg;
2953 	sc = cm->cm_sc;
2954 
2955 	/*
2956 	 * In this case, just print out a warning and let the chip tell the
2957 	 * user they did the wrong thing.
2958 	 */
2959 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2960 		mps_dprint(sc, MPS_ERROR,
2961 			   "%s: warning: busdma returned %d segments, "
2962 			   "more than the %d allowed\n", __func__, nsegs,
2963 			   cm->cm_max_segs);
2964 	}
2965 
2966 	/*
2967 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2968 	 * here.  In that case, both direction flags will be set.
2969 	 */
2970 	sflags = 0;
2971 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2972 		/*
2973 		 * We have to add a special case for SMP passthrough, there
2974 		 * is no easy way to generically handle it.  The first
2975 		 * S/G element is used for the command (therefore the
2976 		 * direction bit needs to be set).  The second one is used
2977 		 * for the reply.  We'll leave it to the caller to make
2978 		 * sure we only have two buffers.
2979 		 */
2980 		/*
2981 		 * Even though the busdma man page says it doesn't make
2982 		 * sense to have both direction flags, it does in this case.
2983 		 * We have one s/g element being accessed in each direction.
2984 		 */
2985 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2986 
2987 		/*
2988 		 * Set the direction flag on the first buffer in the SMP
2989 		 * passthrough request.  We'll clear it for the second one.
2990 		 */
2991 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2992 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2993 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2994 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2995 		dir = BUS_DMASYNC_PREWRITE;
2996 	} else
2997 		dir = BUS_DMASYNC_PREREAD;
2998 
2999 	for (i = 0; i < nsegs; i++) {
3000 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
3001 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3002 		}
3003 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3004 		    sflags, nsegs - i);
3005 		if (error != 0) {
3006 			/* Resource shortage, roll back! */
3007 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
3008 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
3009 				    "consider increasing hw.mps.max_chains.\n");
3010 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
3011 			mps_complete_command(sc, cm);
3012 			return;
3013 		}
3014 	}
3015 
3016 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3017 	mps_enqueue_request(sc, cm);
3018 
3019 	return;
3020 }
3021 
3022 static void
3023 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3024 	     int error)
3025 {
3026 	mps_data_cb(arg, segs, nsegs, error);
3027 }
3028 
3029 /*
3030  * This is the routine to enqueue commands ansynchronously.
3031  * Note that the only error path here is from bus_dmamap_load(), which can
3032  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3033  * assumed that if you have a command in-hand, then you have enough credits
3034  * to use it.
3035  */
3036 int
3037 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
3038 {
3039 	int error = 0;
3040 
3041 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
3042 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3043 		    &cm->cm_uio, mps_data_cb2, cm, 0);
3044 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
3045 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3046 		    cm->cm_data, mps_data_cb, cm, 0);
3047 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3048 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3049 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
3050 	} else {
3051 		/* Add a zero-length element as needed */
3052 		if (cm->cm_sge != NULL)
3053 			mps_add_dmaseg(cm, 0, 0, 0, 1);
3054 		mps_enqueue_request(sc, cm);
3055 	}
3056 
3057 	return (error);
3058 }
3059 
3060 /*
3061  * This is the routine to enqueue commands synchronously.  An error of
3062  * EINPROGRESS from mps_map_command() is ignored since the command will
3063  * be executed and enqueued automatically.  Other errors come from msleep().
3064  */
3065 int
3066 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
3067     int sleep_flag)
3068 {
3069 	int error, rc;
3070 	struct timeval cur_time, start_time;
3071 	struct mps_command *cm = *cmp;
3072 
3073 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
3074 		return  EBUSY;
3075 
3076 	cm->cm_complete = NULL;
3077 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
3078 	error = mps_map_command(sc, cm);
3079 	if ((error != 0) && (error != EINPROGRESS))
3080 		return (error);
3081 
3082 	/*
3083 	 * Check for context and wait for 50 mSec at a time until time has
3084 	 * expired or the command has finished.  If msleep can't be used, need
3085 	 * to poll.
3086 	 */
3087 	if (curthread->td_no_sleeping != 0)
3088 		sleep_flag = NO_SLEEP;
3089 	getmicrouptime(&start_time);
3090 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
3091 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
3092 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
3093 		if (error == EWOULDBLOCK) {
3094 			/*
3095 			 * Record the actual elapsed time in the case of a
3096 			 * timeout for the message below.
3097 			 */
3098 			getmicrouptime(&cur_time);
3099 			timevalsub(&cur_time, &start_time);
3100 		}
3101 	} else {
3102 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
3103 			mps_intr_locked(sc);
3104 			if (sleep_flag == CAN_SLEEP)
3105 				pause("mpswait", hz/20);
3106 			else
3107 				DELAY(50000);
3108 
3109 			getmicrouptime(&cur_time);
3110 			timevalsub(&cur_time, &start_time);
3111 			if (cur_time.tv_sec > timeout) {
3112 				error = EWOULDBLOCK;
3113 				break;
3114 			}
3115 		}
3116 	}
3117 
3118 	if (error == EWOULDBLOCK) {
3119 		if (cm->cm_timeout_handler == NULL) {
3120 			mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
3121 			    " elapsed=%jd\n", __func__, timeout,
3122 			    (intmax_t)cur_time.tv_sec);
3123 			rc = mps_reinit(sc);
3124 			mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3125 			    "failed");
3126 		} else
3127 			cm->cm_timeout_handler(sc, cm);
3128 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
3129 			/*
3130 			 * Tell the caller that we freed the command in a
3131 			 * reinit.
3132 			 */
3133 			*cmp = NULL;
3134 		}
3135 		error = ETIMEDOUT;
3136 	}
3137 	return (error);
3138 }
3139 
3140 /*
3141  * The MPT driver had a verbose interface for config pages.  In this driver,
3142  * reduce it to much simpler terms, similar to the Linux driver.
3143  */
3144 int
3145 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
3146 {
3147 	MPI2_CONFIG_REQUEST *req;
3148 	struct mps_command *cm;
3149 	int error;
3150 
3151 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
3152 		return (EBUSY);
3153 	}
3154 
3155 	cm = mps_alloc_command(sc);
3156 	if (cm == NULL) {
3157 		return (EBUSY);
3158 	}
3159 
3160 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3161 	req->Function = MPI2_FUNCTION_CONFIG;
3162 	req->Action = params->action;
3163 	req->SGLFlags = 0;
3164 	req->ChainOffset = 0;
3165 	req->PageAddress = params->page_address;
3166 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3167 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3168 
3169 		hdr = &params->hdr.Ext;
3170 		req->ExtPageType = hdr->ExtPageType;
3171 		req->ExtPageLength = hdr->ExtPageLength;
3172 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3173 		req->Header.PageLength = 0; /* Must be set to zero */
3174 		req->Header.PageNumber = hdr->PageNumber;
3175 		req->Header.PageVersion = hdr->PageVersion;
3176 	} else {
3177 		MPI2_CONFIG_PAGE_HEADER *hdr;
3178 
3179 		hdr = &params->hdr.Struct;
3180 		req->Header.PageType = hdr->PageType;
3181 		req->Header.PageNumber = hdr->PageNumber;
3182 		req->Header.PageLength = hdr->PageLength;
3183 		req->Header.PageVersion = hdr->PageVersion;
3184 	}
3185 
3186 	cm->cm_data = params->buffer;
3187 	cm->cm_length = params->length;
3188 	if (cm->cm_data != NULL) {
3189 		cm->cm_sge = &req->PageBufferSGE;
3190 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3191 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
3192 	} else
3193 		cm->cm_sge = NULL;
3194 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3195 
3196 	cm->cm_complete_data = params;
3197 	if (params->callback != NULL) {
3198 		cm->cm_complete = mps_config_complete;
3199 		return (mps_map_command(sc, cm));
3200 	} else {
3201 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
3202 		if (error) {
3203 			mps_dprint(sc, MPS_FAULT,
3204 			    "Error %d reading config page\n", error);
3205 			if (cm != NULL)
3206 				mps_free_command(sc, cm);
3207 			return (error);
3208 		}
3209 		mps_config_complete(sc, cm);
3210 	}
3211 
3212 	return (0);
3213 }
3214 
3215 int
3216 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3217 {
3218 	return (EINVAL);
3219 }
3220 
3221 static void
3222 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3223 {
3224 	MPI2_CONFIG_REPLY *reply;
3225 	struct mps_config_params *params;
3226 
3227 	MPS_FUNCTRACE(sc);
3228 	params = cm->cm_complete_data;
3229 
3230 	if (cm->cm_data != NULL) {
3231 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3232 		    BUS_DMASYNC_POSTREAD);
3233 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3234 	}
3235 
3236 	/*
3237 	 * XXX KDM need to do more error recovery?  This results in the
3238 	 * device in question not getting probed.
3239 	 */
3240 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3241 		params->status = MPI2_IOCSTATUS_BUSY;
3242 		goto done;
3243 	}
3244 
3245 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3246 	if (reply == NULL) {
3247 		params->status = MPI2_IOCSTATUS_BUSY;
3248 		goto done;
3249 	}
3250 	params->status = reply->IOCStatus;
3251 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3252 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3253 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3254 		params->hdr.Ext.PageType = reply->Header.PageType;
3255 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3256 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3257 	} else {
3258 		params->hdr.Struct.PageType = reply->Header.PageType;
3259 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3260 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3261 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3262 	}
3263 
3264 done:
3265 	mps_free_command(sc, cm);
3266 	if (params->callback != NULL)
3267 		params->callback(sc, params);
3268 
3269 	return;
3270 }
3271