xref: /freebsd/sys/dev/mps/mps.c (revision 05890ca018513401980415f735eca8c94aaa644a)
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 
63 #include <machine/bus.h>
64 #include <machine/resource.h>
65 #include <sys/rman.h>
66 #include <sys/proc.h>
67 
68 #include <dev/pci/pcivar.h>
69 
70 #include <cam/cam.h>
71 #include <cam/scsi/scsi_all.h>
72 
73 #include <dev/mps/mpi/mpi2_type.h>
74 #include <dev/mps/mpi/mpi2.h>
75 #include <dev/mps/mpi/mpi2_ioc.h>
76 #include <dev/mps/mpi/mpi2_sas.h>
77 #include <dev/mps/mpi/mpi2_cnfg.h>
78 #include <dev/mps/mpi/mpi2_init.h>
79 #include <dev/mps/mpi/mpi2_tool.h>
80 #include <dev/mps/mps_ioctl.h>
81 #include <dev/mps/mpsvar.h>
82 #include <dev/mps/mps_table.h>
83 
84 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
85 static int mps_init_queues(struct mps_softc *sc);
86 static void mps_resize_queues(struct mps_softc *sc);
87 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
88 static int mps_transition_operational(struct mps_softc *sc);
89 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
90 static void mps_iocfacts_free(struct mps_softc *sc);
91 static void mps_startup(void *arg);
92 static int mps_send_iocinit(struct mps_softc *sc);
93 static int mps_alloc_queues(struct mps_softc *sc);
94 static int mps_alloc_hw_queues(struct mps_softc *sc);
95 static int mps_alloc_replies(struct mps_softc *sc);
96 static int mps_alloc_requests(struct mps_softc *sc);
97 static int mps_attach_log(struct mps_softc *sc);
98 static __inline void mps_complete_command(struct mps_softc *sc,
99     struct mps_command *cm);
100 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
101     MPI2_EVENT_NOTIFICATION_REPLY *reply);
102 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
103 static void mps_periodic(void *);
104 static int mps_reregister_events(struct mps_softc *sc);
105 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
106 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
107 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
108 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
109 static void mps_parse_debug(struct mps_softc *sc, char *list);
110 
111 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
112 
113 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
114 MALLOC_DECLARE(M_MPSUSER);
115 
116 /*
117  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
118  * any state and back to its initialization state machine.
119  */
120 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
121 
122 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
123  * Compiler only support unint64_t to be passed as argument.
124  * Otherwise it will throw below error
125  * "aggregate value used where an integer was expected"
126  */
127 
128 typedef union _reply_descriptor {
129         u64 word;
130         struct {
131                 u32 low;
132                 u32 high;
133         } u;
134 }reply_descriptor,address_descriptor;
135 
136 /* Rate limit chain-fail messages to 1 per minute */
137 static struct timeval mps_chainfail_interval = { 60, 0 };
138 
139 /*
140  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
141  * If this function is called from process context, it can sleep
142  * and there is no harm to sleep, in case if this fuction is called
143  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
144  * based on sleep flags driver will call either msleep, pause or DELAY.
145  * msleep and pause are of same variant, but pause is used when mps_mtx
146  * is not hold by driver.
147  *
148  */
149 static int
150 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
151 {
152 	uint32_t reg;
153 	int i, error, tries = 0;
154 	uint8_t first_wait_done = FALSE;
155 
156 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
157 
158 	/* Clear any pending interrupts */
159 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
160 
161 	/*
162 	 * Force NO_SLEEP for threads prohibited to sleep
163  	 * e.a Thread from interrupt handler are prohibited to sleep.
164  	 */
165 	if (curthread->td_no_sleeping != 0)
166 		sleep_flag = NO_SLEEP;
167 
168 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
169 
170 	/* Push the magic sequence */
171 	error = ETIMEDOUT;
172 	while (tries++ < 20) {
173 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
174 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
175 			    mpt2_reset_magic[i]);
176 		/* wait 100 msec */
177 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
178 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
179 			    "mpsdiag", hz/10);
180 		else if (sleep_flag == CAN_SLEEP)
181 			pause("mpsdiag", hz/10);
182 		else
183 			DELAY(100 * 1000);
184 
185 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
186 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
187 			error = 0;
188 			break;
189 		}
190 	}
191 	if (error) {
192 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
193 		    error);
194 		return (error);
195 	}
196 
197 	/* Send the actual reset.  XXX need to refresh the reg? */
198 	reg |= MPI2_DIAG_RESET_ADAPTER;
199 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
200 		reg);
201 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
202 
203 	/* Wait up to 300 seconds in 50ms intervals */
204 	error = ETIMEDOUT;
205 	for (i = 0; i < 6000; i++) {
206 		/*
207 		 * Wait 50 msec. If this is the first time through, wait 256
208 		 * msec to satisfy Diag Reset timing requirements.
209 		 */
210 		if (first_wait_done) {
211 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
212 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
213 				    "mpsdiag", hz/20);
214 			else if (sleep_flag == CAN_SLEEP)
215 				pause("mpsdiag", hz/20);
216 			else
217 				DELAY(50 * 1000);
218 		} else {
219 			DELAY(256 * 1000);
220 			first_wait_done = TRUE;
221 		}
222 		/*
223 		 * Check for the RESET_ADAPTER bit to be cleared first, then
224 		 * wait for the RESET state to be cleared, which takes a little
225 		 * longer.
226 		 */
227 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
228 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
229 			continue;
230 		}
231 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
232 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
233 			error = 0;
234 			break;
235 		}
236 	}
237 	if (error) {
238 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
239 		    error);
240 		return (error);
241 	}
242 
243 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
244 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
245 
246 	return (0);
247 }
248 
249 static int
250 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
251 {
252 	int error;
253 
254 	MPS_FUNCTRACE(sc);
255 
256 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
257 
258 	error = 0;
259 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
260 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
261 	    MPI2_DOORBELL_FUNCTION_SHIFT);
262 
263 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
264 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
265 		    "Doorbell handshake failed\n");
266 		error = ETIMEDOUT;
267 	}
268 
269 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
270 	return (error);
271 }
272 
273 static int
274 mps_transition_ready(struct mps_softc *sc)
275 {
276 	uint32_t reg, state;
277 	int error, tries = 0;
278 	int sleep_flags;
279 
280 	MPS_FUNCTRACE(sc);
281 	/* If we are in attach call, do not sleep */
282 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
283 					? CAN_SLEEP:NO_SLEEP;
284 	error = 0;
285 
286 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
287 	   __func__, sleep_flags);
288 
289 	while (tries++ < 1200) {
290 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
291 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
292 
293 		/*
294 		 * Ensure the IOC is ready to talk.  If it's not, try
295 		 * resetting it.
296 		 */
297 		if (reg & MPI2_DOORBELL_USED) {
298 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
299 			    "reset\n");
300 			mps_diag_reset(sc, sleep_flags);
301 			DELAY(50000);
302 			continue;
303 		}
304 
305 		/* Is the adapter owned by another peer? */
306 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
307 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
308 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
309 			    "control of another peer host, aborting "
310 			    "initialization.\n");
311 			error = ENXIO;
312 			break;
313 		}
314 
315 		state = reg & MPI2_IOC_STATE_MASK;
316 		if (state == MPI2_IOC_STATE_READY) {
317 			/* Ready to go! */
318 			error = 0;
319 			break;
320 		} else if (state == MPI2_IOC_STATE_FAULT) {
321 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
322 			    "state 0x%x, resetting\n",
323 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
324 			mps_diag_reset(sc, sleep_flags);
325 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
326 			/* Need to take ownership */
327 			mps_message_unit_reset(sc, sleep_flags);
328 		} else if (state == MPI2_IOC_STATE_RESET) {
329 			/* Wait a bit, IOC might be in transition */
330 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
331 			    "IOC in unexpected reset state\n");
332 		} else {
333 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
334 			    "IOC in unknown state 0x%x\n", state);
335 			error = EINVAL;
336 			break;
337 		}
338 
339 		/* Wait 50ms for things to settle down. */
340 		DELAY(50000);
341 	}
342 
343 	if (error)
344 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
345 		    "Cannot transition IOC to ready\n");
346 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
347 
348 	return (error);
349 }
350 
351 static int
352 mps_transition_operational(struct mps_softc *sc)
353 {
354 	uint32_t reg, state;
355 	int error;
356 
357 	MPS_FUNCTRACE(sc);
358 
359 	error = 0;
360 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
361 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
362 
363 	state = reg & MPI2_IOC_STATE_MASK;
364 	if (state != MPI2_IOC_STATE_READY) {
365 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
366 		if ((error = mps_transition_ready(sc)) != 0) {
367 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
368 			    "failed to transition ready, exit\n");
369 			return (error);
370 		}
371 	}
372 
373 	error = mps_send_iocinit(sc);
374 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
375 
376 	return (error);
377 }
378 
379 static void
380 mps_resize_queues(struct mps_softc *sc)
381 {
382 	int reqcr, prireqcr;
383 
384 	/*
385 	 * Size the queues. Since the reply queues always need one free
386 	 * entry, we'll deduct one reply message here.  The LSI documents
387 	 * suggest instead to add a count to the request queue, but I think
388 	 * that it's better to deduct from reply queue.
389 	 */
390 	prireqcr = MAX(1, sc->max_prireqframes);
391 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
392 
393 	reqcr = MAX(2, sc->max_reqframes);
394 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
395 
396 	sc->num_reqs = prireqcr + reqcr;
397 	sc->num_prireqs = prireqcr;
398 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
399 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
400 
401 	/*
402 	 * Figure out the number of MSIx-based queues.  If the firmware or
403 	 * user has done something crazy and not allowed enough credit for
404 	 * the queues to be useful then don't enable multi-queue.
405 	 */
406 	if (sc->facts->MaxMSIxVectors < 2)
407 		sc->msi_msgs = 1;
408 
409 	if (sc->msi_msgs > 1) {
410 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
411 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
412 		if (sc->num_reqs / sc->msi_msgs < 2)
413 			sc->msi_msgs = 1;
414 	}
415 
416 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
417 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
418 }
419 
420 /*
421  * This is called during attach and when re-initializing due to a Diag Reset.
422  * IOC Facts is used to allocate many of the structures needed by the driver.
423  * If called from attach, de-allocation is not required because the driver has
424  * not allocated any structures yet, but if called from a Diag Reset, previously
425  * allocated structures based on IOC Facts will need to be freed and re-
426  * allocated bases on the latest IOC Facts.
427  */
428 static int
429 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
430 {
431 	int error;
432 	Mpi2IOCFactsReply_t saved_facts;
433 	uint8_t saved_mode, reallocating;
434 
435 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
436 
437 	/* Save old IOC Facts and then only reallocate if Facts have changed */
438 	if (!attaching) {
439 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
440 	}
441 
442 	/*
443 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
444 	 * a re-initialization and only return the error if attaching so the OS
445 	 * can handle it.
446 	 */
447 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
448 		if (attaching) {
449 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
450 			    "IOC Facts with error %d, exit\n", error);
451 			return (error);
452 		} else {
453 			panic("%s failed to get IOC Facts with error %d\n",
454 			    __func__, error);
455 		}
456 	}
457 
458 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
459 
460 	snprintf(sc->fw_version, sizeof(sc->fw_version),
461 	    "%02d.%02d.%02d.%02d",
462 	    sc->facts->FWVersion.Struct.Major,
463 	    sc->facts->FWVersion.Struct.Minor,
464 	    sc->facts->FWVersion.Struct.Unit,
465 	    sc->facts->FWVersion.Struct.Dev);
466 
467 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
468 	    MPS_DRIVER_VERSION);
469 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
470 	     sc->facts->IOCCapabilities,
471 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
472 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
473 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
474 
475 	/*
476 	 * If the chip doesn't support event replay then a hard reset will be
477 	 * required to trigger a full discovery.  Do the reset here then
478 	 * retransition to Ready.  A hard reset might have already been done,
479 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
480 	 * for a Diag Reset.
481 	 */
482 	if (attaching && ((sc->facts->IOCCapabilities &
483 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
484 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
485 		mps_diag_reset(sc, NO_SLEEP);
486 		if ((error = mps_transition_ready(sc)) != 0) {
487 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
488 			    "transition to ready with error %d, exit\n",
489 			    error);
490 			return (error);
491 		}
492 	}
493 
494 	/*
495 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
496 	 * changed from the previous IOC Facts, log a warning, but only if
497 	 * checking this after a Diag Reset and not during attach.
498 	 */
499 	saved_mode = sc->ir_firmware;
500 	if (sc->facts->IOCCapabilities &
501 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
502 		sc->ir_firmware = 1;
503 	if (!attaching) {
504 		if (sc->ir_firmware != saved_mode) {
505 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
506 			    "in IOC Facts does not match previous mode\n");
507 		}
508 	}
509 
510 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
511 	reallocating = FALSE;
512 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
513 
514 	if ((!attaching) &&
515 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
516 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
517 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
518 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
519 	    (saved_facts.ProductID != sc->facts->ProductID) ||
520 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
521 	    (saved_facts.IOCRequestFrameSize !=
522 	    sc->facts->IOCRequestFrameSize) ||
523 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
524 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
525 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
526 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
527 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
528 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
529 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
530 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
531 	    (saved_facts.MaxPersistentEntries !=
532 	    sc->facts->MaxPersistentEntries))) {
533 		reallocating = TRUE;
534 
535 		/* Record that we reallocated everything */
536 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
537 	}
538 
539 	/*
540 	 * Some things should be done if attaching or re-allocating after a Diag
541 	 * Reset, but are not needed after a Diag Reset if the FW has not
542 	 * changed.
543 	 */
544 	if (attaching || reallocating) {
545 		/*
546 		 * Check if controller supports FW diag buffers and set flag to
547 		 * enable each type.
548 		 */
549 		if (sc->facts->IOCCapabilities &
550 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
551 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
552 			    enabled = TRUE;
553 		if (sc->facts->IOCCapabilities &
554 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
555 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
556 			    enabled = TRUE;
557 		if (sc->facts->IOCCapabilities &
558 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
559 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
560 			    enabled = TRUE;
561 
562 		/*
563 		 * Set flag if EEDP is supported and if TLR is supported.
564 		 */
565 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
566 			sc->eedp_enabled = TRUE;
567 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
568 			sc->control_TLR = TRUE;
569 
570 		mps_resize_queues(sc);
571 
572 		/*
573 		 * Initialize all Tail Queues
574 		 */
575 		TAILQ_INIT(&sc->req_list);
576 		TAILQ_INIT(&sc->high_priority_req_list);
577 		TAILQ_INIT(&sc->chain_list);
578 		TAILQ_INIT(&sc->tm_list);
579 	}
580 
581 	/*
582 	 * If doing a Diag Reset and the FW is significantly different
583 	 * (reallocating will be set above in IOC Facts comparison), then all
584 	 * buffers based on the IOC Facts will need to be freed before they are
585 	 * reallocated.
586 	 */
587 	if (reallocating) {
588 		mps_iocfacts_free(sc);
589 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
590 		    saved_facts.MaxVolumes);
591 	}
592 
593 	/*
594 	 * Any deallocation has been completed.  Now start reallocating
595 	 * if needed.  Will only need to reallocate if attaching or if the new
596 	 * IOC Facts are different from the previous IOC Facts after a Diag
597 	 * Reset. Targets have already been allocated above if needed.
598 	 */
599 	error = 0;
600 	while (attaching || reallocating) {
601 		if ((error = mps_alloc_hw_queues(sc)) != 0)
602 			break;
603 		if ((error = mps_alloc_replies(sc)) != 0)
604 			break;
605 		if ((error = mps_alloc_requests(sc)) != 0)
606 			break;
607 		if ((error = mps_alloc_queues(sc)) != 0)
608 			break;
609 
610 		break;
611 	}
612 	if (error) {
613 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
614 		    "Failed to alloc queues with error %d\n", error);
615 		mps_free(sc);
616 		return (error);
617 	}
618 
619 	/* Always initialize the queues */
620 	bzero(sc->free_queue, sc->fqdepth * 4);
621 	mps_init_queues(sc);
622 
623 	/*
624 	 * Always get the chip out of the reset state, but only panic if not
625 	 * attaching.  If attaching and there is an error, that is handled by
626 	 * the OS.
627 	 */
628 	error = mps_transition_operational(sc);
629 	if (error != 0) {
630 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
631 		    "transition to operational with error %d\n", error);
632 		mps_free(sc);
633 		return (error);
634 	}
635 
636 	/*
637 	 * Finish the queue initialization.
638 	 * These are set here instead of in mps_init_queues() because the
639 	 * IOC resets these values during the state transition in
640 	 * mps_transition_operational().  The free index is set to 1
641 	 * because the corresponding index in the IOC is set to 0, and the
642 	 * IOC treats the queues as full if both are set to the same value.
643 	 * Hence the reason that the queue can't hold all of the possible
644 	 * replies.
645 	 */
646 	sc->replypostindex = 0;
647 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
648 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
649 
650 	/*
651 	 * Attach the subsystems so they can prepare their event masks.
652 	 * XXX Should be dynamic so that IM/IR and user modules can attach
653 	 */
654 	error = 0;
655 	while (attaching) {
656 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
657 		if ((error = mps_attach_log(sc)) != 0)
658 			break;
659 		if ((error = mps_attach_sas(sc)) != 0)
660 			break;
661 		if ((error = mps_attach_user(sc)) != 0)
662 			break;
663 		break;
664 	}
665 	if (error) {
666 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
667 		    "subsystems: error %d\n", error);
668 		mps_free(sc);
669 		return (error);
670 	}
671 
672 	/*
673 	 * XXX If the number of MSI-X vectors changes during re-init, this
674 	 * won't see it and adjust.
675 	 */
676 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
677 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
678 		    "interrupts\n");
679 		mps_free(sc);
680 		return (error);
681 	}
682 
683 	/*
684 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
685 	 * reset it after a Diag Reset, just in case.
686 	 */
687 	sc->WD_available = FALSE;
688 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
689 		sc->WD_available = TRUE;
690 
691 	return (error);
692 }
693 
694 /*
695  * This is called if memory is being free (during detach for example) and when
696  * buffers need to be reallocated due to a Diag Reset.
697  */
698 static void
699 mps_iocfacts_free(struct mps_softc *sc)
700 {
701 	struct mps_command *cm;
702 	int i;
703 
704 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
705 
706 	if (sc->free_busaddr != 0)
707 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
708 	if (sc->free_queue != NULL)
709 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
710 		    sc->queues_map);
711 	if (sc->queues_dmat != NULL)
712 		bus_dma_tag_destroy(sc->queues_dmat);
713 
714 	if (sc->chain_busaddr != 0)
715 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
716 	if (sc->chain_frames != NULL)
717 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
718 		    sc->chain_map);
719 	if (sc->chain_dmat != NULL)
720 		bus_dma_tag_destroy(sc->chain_dmat);
721 
722 	if (sc->sense_busaddr != 0)
723 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
724 	if (sc->sense_frames != NULL)
725 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
726 		    sc->sense_map);
727 	if (sc->sense_dmat != NULL)
728 		bus_dma_tag_destroy(sc->sense_dmat);
729 
730 	if (sc->reply_busaddr != 0)
731 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
732 	if (sc->reply_frames != NULL)
733 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
734 		    sc->reply_map);
735 	if (sc->reply_dmat != NULL)
736 		bus_dma_tag_destroy(sc->reply_dmat);
737 
738 	if (sc->req_busaddr != 0)
739 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
740 	if (sc->req_frames != NULL)
741 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
742 	if (sc->req_dmat != NULL)
743 		bus_dma_tag_destroy(sc->req_dmat);
744 
745 	if (sc->chains != NULL)
746 		free(sc->chains, M_MPT2);
747 	if (sc->commands != NULL) {
748 		for (i = 1; i < sc->num_reqs; i++) {
749 			cm = &sc->commands[i];
750 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
751 		}
752 		free(sc->commands, M_MPT2);
753 	}
754 	if (sc->buffer_dmat != NULL)
755 		bus_dma_tag_destroy(sc->buffer_dmat);
756 
757 	mps_pci_free_interrupts(sc);
758 	free(sc->queues, M_MPT2);
759 	sc->queues = NULL;
760 }
761 
762 /*
763  * The terms diag reset and hard reset are used interchangeably in the MPI
764  * docs to mean resetting the controller chip.  In this code diag reset
765  * cleans everything up, and the hard reset function just sends the reset
766  * sequence to the chip.  This should probably be refactored so that every
767  * subsystem gets a reset notification of some sort, and can clean up
768  * appropriately.
769  */
770 int
771 mps_reinit(struct mps_softc *sc)
772 {
773 	int error;
774 	struct mpssas_softc *sassc;
775 
776 	sassc = sc->sassc;
777 
778 	MPS_FUNCTRACE(sc);
779 
780 	mtx_assert(&sc->mps_mtx, MA_OWNED);
781 
782 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
783 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
784 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
785 		return 0;
786 	}
787 
788 	/* make sure the completion callbacks can recognize they're getting
789 	 * a NULL cm_reply due to a reset.
790 	 */
791 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
792 
793 	/*
794 	 * Mask interrupts here.
795 	 */
796 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
797 	mps_mask_intr(sc);
798 
799 	error = mps_diag_reset(sc, CAN_SLEEP);
800 	if (error != 0) {
801 		/* XXXSL No need to panic here */
802 		panic("%s hard reset failed with error %d\n",
803 		    __func__, error);
804 	}
805 
806 	/* Restore the PCI state, including the MSI-X registers */
807 	mps_pci_restore(sc);
808 
809 	/* Give the I/O subsystem special priority to get itself prepared */
810 	mpssas_handle_reinit(sc);
811 
812 	/*
813 	 * Get IOC Facts and allocate all structures based on this information.
814 	 * The attach function will also call mps_iocfacts_allocate at startup.
815 	 * If relevant values have changed in IOC Facts, this function will free
816 	 * all of the memory based on IOC Facts and reallocate that memory.
817 	 */
818 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
819 		panic("%s IOC Facts based allocation failed with error %d\n",
820 		    __func__, error);
821 	}
822 
823 	/*
824 	 * Mapping structures will be re-allocated after getting IOC Page8, so
825 	 * free these structures here.
826 	 */
827 	mps_mapping_exit(sc);
828 
829 	/*
830 	 * The static page function currently read is IOC Page8.  Others can be
831 	 * added in future.  It's possible that the values in IOC Page8 have
832 	 * changed after a Diag Reset due to user modification, so always read
833 	 * these.  Interrupts are masked, so unmask them before getting config
834 	 * pages.
835 	 */
836 	mps_unmask_intr(sc);
837 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
838 	mps_base_static_config_pages(sc);
839 
840 	/*
841 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
842 	 * mapping tables.
843 	 */
844 	mps_mapping_initialize(sc);
845 
846 	/*
847 	 * Restart will reload the event masks clobbered by the reset, and
848 	 * then enable the port.
849 	 */
850 	mps_reregister_events(sc);
851 
852 	/* the end of discovery will release the simq, so we're done. */
853 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
854 	    sc, sc->replypostindex, sc->replyfreeindex);
855 
856 	mpssas_release_simq_reinit(sassc);
857 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
858 
859 	return 0;
860 }
861 
862 /* Wait for the chip to ACK a word that we've put into its FIFO
863  * Wait for <timeout> seconds. In single loop wait for busy loop
864  * for 500 microseconds.
865  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
866  * */
867 static int
868 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
869 {
870 
871 	u32 cntdn, count;
872 	u32 int_status;
873 	u32 doorbell;
874 
875 	count = 0;
876 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
877 	do {
878 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
879 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
880 			mps_dprint(sc, MPS_TRACE,
881 			"%s: successful count(%d), timeout(%d)\n",
882 			__func__, count, timeout);
883 		return 0;
884 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
885 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
886 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
887 				MPI2_IOC_STATE_FAULT) {
888 				mps_dprint(sc, MPS_FAULT,
889 					"fault_state(0x%04x)!\n", doorbell);
890 				return (EFAULT);
891 			}
892 		} else if (int_status == 0xFFFFFFFF)
893 			goto out;
894 
895 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
896 		* 0.5 milisecond */
897 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
898 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
899 			"mpsdba", hz/1000);
900 		else if (sleep_flag == CAN_SLEEP)
901 			pause("mpsdba", hz/1000);
902 		else
903 			DELAY(500);
904 		count++;
905 	} while (--cntdn);
906 
907 	out:
908 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
909 		"int_status(%x)!\n", __func__, count, int_status);
910 	return (ETIMEDOUT);
911 
912 }
913 
914 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
915 static int
916 mps_wait_db_int(struct mps_softc *sc)
917 {
918 	int retry;
919 
920 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
921 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
922 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
923 			return (0);
924 		DELAY(2000);
925 	}
926 	return (ETIMEDOUT);
927 }
928 
929 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
930 static int
931 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
932     int req_sz, int reply_sz, int timeout)
933 {
934 	uint32_t *data32;
935 	uint16_t *data16;
936 	int i, count, ioc_sz, residual;
937 	int sleep_flags = CAN_SLEEP;
938 
939 	if (curthread->td_no_sleeping != 0)
940 		sleep_flags = NO_SLEEP;
941 
942 	/* Step 1 */
943 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
944 
945 	/* Step 2 */
946 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
947 		return (EBUSY);
948 
949 	/* Step 3
950 	 * Announce that a message is coming through the doorbell.  Messages
951 	 * are pushed at 32bit words, so round up if needed.
952 	 */
953 	count = (req_sz + 3) / 4;
954 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
955 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
956 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
957 
958 	/* Step 4 */
959 	if (mps_wait_db_int(sc) ||
960 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
961 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
962 		return (ENXIO);
963 	}
964 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
965 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
966 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
967 		return (ENXIO);
968 	}
969 
970 	/* Step 5 */
971 	/* Clock out the message data synchronously in 32-bit dwords*/
972 	data32 = (uint32_t *)req;
973 	for (i = 0; i < count; i++) {
974 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
975 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
976 			mps_dprint(sc, MPS_FAULT,
977 			    "Timeout while writing doorbell\n");
978 			return (ENXIO);
979 		}
980 	}
981 
982 	/* Step 6 */
983 	/* Clock in the reply in 16-bit words.  The total length of the
984 	 * message is always in the 4th byte, so clock out the first 2 words
985 	 * manually, then loop the rest.
986 	 */
987 	data16 = (uint16_t *)reply;
988 	if (mps_wait_db_int(sc) != 0) {
989 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
990 		return (ENXIO);
991 	}
992 	data16[0] =
993 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
994 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
995 	if (mps_wait_db_int(sc) != 0) {
996 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
997 		return (ENXIO);
998 	}
999 	data16[1] =
1000 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1001 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1002 
1003 	/* Number of 32bit words in the message */
1004 	ioc_sz = reply->MsgLength;
1005 
1006 	/*
1007 	 * Figure out how many 16bit words to clock in without overrunning.
1008 	 * The precision loss with dividing reply_sz can safely be
1009 	 * ignored because the messages can only be multiples of 32bits.
1010 	 */
1011 	residual = 0;
1012 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1013 	if (count < ioc_sz * 2) {
1014 		residual = ioc_sz * 2 - count;
1015 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1016 		    "residual message words\n", residual);
1017 	}
1018 
1019 	for (i = 2; i < count; i++) {
1020 		if (mps_wait_db_int(sc) != 0) {
1021 			mps_dprint(sc, MPS_FAULT,
1022 			    "Timeout reading doorbell %d\n", i);
1023 			return (ENXIO);
1024 		}
1025 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1026 		    MPI2_DOORBELL_DATA_MASK;
1027 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1028 	}
1029 
1030 	/*
1031 	 * Pull out residual words that won't fit into the provided buffer.
1032 	 * This keeps the chip from hanging due to a driver programming
1033 	 * error.
1034 	 */
1035 	while (residual--) {
1036 		if (mps_wait_db_int(sc) != 0) {
1037 			mps_dprint(sc, MPS_FAULT,
1038 			    "Timeout reading doorbell\n");
1039 			return (ENXIO);
1040 		}
1041 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1042 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1043 	}
1044 
1045 	/* Step 7 */
1046 	if (mps_wait_db_int(sc) != 0) {
1047 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1048 		return (ENXIO);
1049 	}
1050 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1051 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1052 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1053 
1054 	return (0);
1055 }
1056 
1057 static void
1058 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1059 {
1060 	reply_descriptor rd;
1061 	MPS_FUNCTRACE(sc);
1062 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1063 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1064 
1065 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1066 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1067 
1068 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1069 		sc->io_cmds_highwater++;
1070 	rd.u.low = cm->cm_desc.Words.Low;
1071 	rd.u.high = cm->cm_desc.Words.High;
1072 	rd.word = htole64(rd.word);
1073 	/* TODO-We may need to make below regwrite atomic */
1074 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1075 	    rd.u.low);
1076 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1077 	    rd.u.high);
1078 }
1079 
1080 /*
1081  * Just the FACTS, ma'am.
1082  */
1083 static int
1084 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1085 {
1086 	MPI2_DEFAULT_REPLY *reply;
1087 	MPI2_IOC_FACTS_REQUEST request;
1088 	int error, req_sz, reply_sz;
1089 
1090 	MPS_FUNCTRACE(sc);
1091 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1092 
1093 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1094 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1095 	reply = (MPI2_DEFAULT_REPLY *)facts;
1096 
1097 	bzero(&request, req_sz);
1098 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1099 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1100 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1101 
1102 	return (error);
1103 }
1104 
1105 static int
1106 mps_send_iocinit(struct mps_softc *sc)
1107 {
1108 	MPI2_IOC_INIT_REQUEST	init;
1109 	MPI2_DEFAULT_REPLY	reply;
1110 	int req_sz, reply_sz, error;
1111 	struct timeval now;
1112 	uint64_t time_in_msec;
1113 
1114 	MPS_FUNCTRACE(sc);
1115 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1116 
1117 	/* Do a quick sanity check on proper initialization */
1118 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1119 	    || (sc->replyframesz == 0)) {
1120 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1121 		    "Driver not fully initialized for IOCInit\n");
1122 		return (EINVAL);
1123 	}
1124 
1125 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1126 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1127 	bzero(&init, req_sz);
1128 	bzero(&reply, reply_sz);
1129 
1130 	/*
1131 	 * Fill in the init block.  Note that most addresses are
1132 	 * deliberately in the lower 32bits of memory.  This is a micro-
1133 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1134 	 */
1135 	init.Function = MPI2_FUNCTION_IOC_INIT;
1136 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1137 	init.MsgVersion = htole16(MPI2_VERSION);
1138 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1139 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1140 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1141 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1142 	init.SenseBufferAddressHigh = 0;
1143 	init.SystemReplyAddressHigh = 0;
1144 	init.SystemRequestFrameBaseAddress.High = 0;
1145 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1146 	init.ReplyDescriptorPostQueueAddress.High = 0;
1147 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1148 	init.ReplyFreeQueueAddress.High = 0;
1149 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1150 	getmicrotime(&now);
1151 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1152 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1153 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1154 
1155 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1156 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1157 		error = ENXIO;
1158 
1159 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1160 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1161 	return (error);
1162 }
1163 
1164 void
1165 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1166 {
1167 	bus_addr_t *addr;
1168 
1169 	addr = arg;
1170 	*addr = segs[0].ds_addr;
1171 }
1172 
1173 void
1174 mps_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1175 {
1176 	struct mps_busdma_context *ctx;
1177 	int need_unload, need_free;
1178 
1179 	ctx = (struct mps_busdma_context *)arg;
1180 	need_unload = 0;
1181 	need_free = 0;
1182 
1183 	mps_lock(ctx->softc);
1184 	ctx->error = error;
1185 	ctx->completed = 1;
1186 	if ((error == 0) && (ctx->abandoned == 0)) {
1187 		*ctx->addr = segs[0].ds_addr;
1188 	} else {
1189 		if (nsegs != 0)
1190 			need_unload = 1;
1191 		if (ctx->abandoned != 0)
1192 			need_free = 1;
1193 	}
1194 	if (need_free == 0)
1195 		wakeup(ctx);
1196 
1197 	mps_unlock(ctx->softc);
1198 
1199 	if (need_unload != 0) {
1200 		bus_dmamap_unload(ctx->buffer_dmat,
1201 				  ctx->buffer_dmamap);
1202 		*ctx->addr = 0;
1203 	}
1204 
1205 	if (need_free != 0)
1206 		free(ctx, M_MPSUSER);
1207 }
1208 
1209 static int
1210 mps_alloc_queues(struct mps_softc *sc)
1211 {
1212 	struct mps_queue *q;
1213 	u_int nq, i;
1214 
1215 	nq = sc->msi_msgs;
1216 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1217 
1218 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1219 	    M_NOWAIT|M_ZERO);
1220 	if (sc->queues == NULL)
1221 		return (ENOMEM);
1222 
1223 	for (i = 0; i < nq; i++) {
1224 		q = &sc->queues[i];
1225 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1226 		q->sc = sc;
1227 		q->qnum = i;
1228 	}
1229 
1230 	return (0);
1231 }
1232 
1233 static int
1234 mps_alloc_hw_queues(struct mps_softc *sc)
1235 {
1236 	bus_addr_t queues_busaddr;
1237 	uint8_t *queues;
1238 	int qsize, fqsize, pqsize;
1239 
1240 	/*
1241 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1242 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1243 	 * This queue supplies fresh reply frames for the firmware to use.
1244 	 *
1245 	 * The reply descriptor post queue contains 8 byte entries in
1246 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1247 	 * contains filled-in reply frames sent from the firmware to the host.
1248 	 *
1249 	 * These two queues are allocated together for simplicity.
1250 	 */
1251 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1252 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1253 	fqsize= sc->fqdepth * 4;
1254 	pqsize = sc->pqdepth * 8;
1255 	qsize = fqsize + pqsize;
1256 
1257         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1258 				16, 0,			/* algnmnt, boundary */
1259 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1260 				BUS_SPACE_MAXADDR,	/* highaddr */
1261 				NULL, NULL,		/* filter, filterarg */
1262                                 qsize,			/* maxsize */
1263                                 1,			/* nsegments */
1264                                 qsize,			/* maxsegsize */
1265                                 0,			/* flags */
1266                                 NULL, NULL,		/* lockfunc, lockarg */
1267                                 &sc->queues_dmat)) {
1268 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1269 		return (ENOMEM);
1270         }
1271         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1272 	    &sc->queues_map)) {
1273 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1274 		return (ENOMEM);
1275         }
1276         bzero(queues, qsize);
1277         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1278 	    mps_memaddr_cb, &queues_busaddr, 0);
1279 
1280 	sc->free_queue = (uint32_t *)queues;
1281 	sc->free_busaddr = queues_busaddr;
1282 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1283 	sc->post_busaddr = queues_busaddr + fqsize;
1284 
1285 	return (0);
1286 }
1287 
1288 static int
1289 mps_alloc_replies(struct mps_softc *sc)
1290 {
1291 	int rsize, num_replies;
1292 
1293 	/* Store the reply frame size in bytes rather than as 32bit words */
1294 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1295 
1296 	/*
1297 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1298 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1299 	 * replies can be used at once.
1300 	 */
1301 	num_replies = max(sc->fqdepth, sc->num_replies);
1302 
1303 	rsize = sc->replyframesz * num_replies;
1304         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1305 				4, 0,			/* algnmnt, boundary */
1306 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1307 				BUS_SPACE_MAXADDR,	/* highaddr */
1308 				NULL, NULL,		/* filter, filterarg */
1309                                 rsize,			/* maxsize */
1310                                 1,			/* nsegments */
1311                                 rsize,			/* maxsegsize */
1312                                 0,			/* flags */
1313                                 NULL, NULL,		/* lockfunc, lockarg */
1314                                 &sc->reply_dmat)) {
1315 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1316 		return (ENOMEM);
1317         }
1318         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1319 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1320 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1321 		return (ENOMEM);
1322         }
1323         bzero(sc->reply_frames, rsize);
1324         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1325 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1326 
1327 	return (0);
1328 }
1329 
1330 static int
1331 mps_alloc_requests(struct mps_softc *sc)
1332 {
1333 	struct mps_command *cm;
1334 	struct mps_chain *chain;
1335 	int i, rsize, nsegs;
1336 
1337 	/* Store the request frame size in bytes rather than as 32bit words */
1338 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
1339 
1340 	rsize = sc->reqframesz * sc->num_reqs;
1341         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1342 				16, 0,			/* algnmnt, boundary */
1343 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1344 				BUS_SPACE_MAXADDR,	/* highaddr */
1345 				NULL, NULL,		/* filter, filterarg */
1346                                 rsize,			/* maxsize */
1347                                 1,			/* nsegments */
1348                                 rsize,			/* maxsegsize */
1349                                 0,			/* flags */
1350                                 NULL, NULL,		/* lockfunc, lockarg */
1351                                 &sc->req_dmat)) {
1352 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1353 		return (ENOMEM);
1354         }
1355         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1356 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1357 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1358 		return (ENOMEM);
1359         }
1360         bzero(sc->req_frames, rsize);
1361         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1362 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1363 
1364 	rsize = sc->reqframesz * sc->max_chains;
1365         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1366 				16, 0,			/* algnmnt, boundary */
1367 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1368 				BUS_SPACE_MAXADDR,	/* highaddr */
1369 				NULL, NULL,		/* filter, filterarg */
1370                                 rsize,			/* maxsize */
1371                                 1,			/* nsegments */
1372                                 rsize,			/* maxsegsize */
1373                                 0,			/* flags */
1374                                 NULL, NULL,		/* lockfunc, lockarg */
1375                                 &sc->chain_dmat)) {
1376 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1377 		return (ENOMEM);
1378         }
1379         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1380 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1381 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1382 		return (ENOMEM);
1383         }
1384         bzero(sc->chain_frames, rsize);
1385         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1386 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1387 
1388 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1389         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1390 				1, 0,			/* algnmnt, boundary */
1391 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1392 				BUS_SPACE_MAXADDR,	/* highaddr */
1393 				NULL, NULL,		/* filter, filterarg */
1394                                 rsize,			/* maxsize */
1395                                 1,			/* nsegments */
1396                                 rsize,			/* maxsegsize */
1397                                 0,			/* flags */
1398                                 NULL, NULL,		/* lockfunc, lockarg */
1399                                 &sc->sense_dmat)) {
1400 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1401 		return (ENOMEM);
1402         }
1403         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1404 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1405 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1406 		return (ENOMEM);
1407         }
1408         bzero(sc->sense_frames, rsize);
1409         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1410 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1411 
1412 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1413 	    M_WAITOK | M_ZERO);
1414 	if(!sc->chains) {
1415 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1416 		return (ENOMEM);
1417 	}
1418 	for (i = 0; i < sc->max_chains; i++) {
1419 		chain = &sc->chains[i];
1420 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1421 		    i * sc->reqframesz);
1422 		chain->chain_busaddr = sc->chain_busaddr +
1423 		    i * sc->reqframesz;
1424 		mps_free_chain(sc, chain);
1425 		sc->chain_free_lowwater++;
1426 	}
1427 
1428 	/* XXX Need to pick a more precise value */
1429 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1430         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1431 				1, 0,			/* algnmnt, boundary */
1432 				BUS_SPACE_MAXADDR,	/* lowaddr */
1433 				BUS_SPACE_MAXADDR,	/* highaddr */
1434 				NULL, NULL,		/* filter, filterarg */
1435                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1436                                 nsegs,			/* nsegments */
1437                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1438                                 BUS_DMA_ALLOCNOW,	/* flags */
1439                                 busdma_lock_mutex,	/* lockfunc */
1440 				&sc->mps_mtx,		/* lockarg */
1441                                 &sc->buffer_dmat)) {
1442 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1443 		return (ENOMEM);
1444         }
1445 
1446 	/*
1447 	 * SMID 0 cannot be used as a free command per the firmware spec.
1448 	 * Just drop that command instead of risking accounting bugs.
1449 	 */
1450 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1451 	    M_MPT2, M_WAITOK | M_ZERO);
1452 	if(!sc->commands) {
1453 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1454 		return (ENOMEM);
1455 	}
1456 	for (i = 1; i < sc->num_reqs; i++) {
1457 		cm = &sc->commands[i];
1458 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1459 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1460 		cm->cm_sense = &sc->sense_frames[i];
1461 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1462 		cm->cm_desc.Default.SMID = i;
1463 		cm->cm_sc = sc;
1464 		TAILQ_INIT(&cm->cm_chain_list);
1465 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1466 
1467 		/* XXX Is a failure here a critical problem? */
1468 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1469 			if (i <= sc->num_prireqs)
1470 				mps_free_high_priority_command(sc, cm);
1471 			else
1472 				mps_free_command(sc, cm);
1473 		else {
1474 			panic("failed to allocate command %d\n", i);
1475 			sc->num_reqs = i;
1476 			break;
1477 		}
1478 	}
1479 
1480 	return (0);
1481 }
1482 
1483 static int
1484 mps_init_queues(struct mps_softc *sc)
1485 {
1486 	int i;
1487 
1488 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1489 
1490 	/*
1491 	 * According to the spec, we need to use one less reply than we
1492 	 * have space for on the queue.  So sc->num_replies (the number we
1493 	 * use) should be less than sc->fqdepth (allocated size).
1494 	 */
1495 	if (sc->num_replies >= sc->fqdepth)
1496 		return (EINVAL);
1497 
1498 	/*
1499 	 * Initialize all of the free queue entries.
1500 	 */
1501 	for (i = 0; i < sc->fqdepth; i++)
1502 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1503 	sc->replyfreeindex = sc->num_replies;
1504 
1505 	return (0);
1506 }
1507 
1508 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1509  * Next are the global settings, if they exist.  Highest are the per-unit
1510  * settings, if they exist.
1511  */
1512 void
1513 mps_get_tunables(struct mps_softc *sc)
1514 {
1515 	char tmpstr[80], mps_debug[80];
1516 
1517 	/* XXX default to some debugging for now */
1518 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1519 	sc->disable_msix = 0;
1520 	sc->disable_msi = 0;
1521 	sc->max_msix = MPS_MSIX_MAX;
1522 	sc->max_chains = MPS_CHAIN_FRAMES;
1523 	sc->max_io_pages = MPS_MAXIO_PAGES;
1524 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1525 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1526 	sc->use_phynum = 1;
1527 	sc->max_reqframes = MPS_REQ_FRAMES;
1528 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1529 	sc->max_replyframes = MPS_REPLY_FRAMES;
1530 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1531 
1532 	/*
1533 	 * Grab the global variables.
1534 	 */
1535 	bzero(mps_debug, 80);
1536 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1537 		mps_parse_debug(sc, mps_debug);
1538 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1539 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1540 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1541 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1542 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1543 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1544 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1545 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1546 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1547 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1548 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1549 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1550 
1551 	/* Grab the unit-instance variables */
1552 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1553 	    device_get_unit(sc->mps_dev));
1554 	bzero(mps_debug, 80);
1555 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1556 		mps_parse_debug(sc, mps_debug);
1557 
1558 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1559 	    device_get_unit(sc->mps_dev));
1560 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1561 
1562 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1563 	    device_get_unit(sc->mps_dev));
1564 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1565 
1566 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1567 	    device_get_unit(sc->mps_dev));
1568 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1569 
1570 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1571 	    device_get_unit(sc->mps_dev));
1572 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1573 
1574 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1575 	    device_get_unit(sc->mps_dev));
1576 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1577 
1578 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1579 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1580 	    device_get_unit(sc->mps_dev));
1581 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1582 
1583 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1584 	    device_get_unit(sc->mps_dev));
1585 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1586 
1587 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1588 	    device_get_unit(sc->mps_dev));
1589 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1590 
1591 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1592 	    device_get_unit(sc->mps_dev));
1593 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1594 
1595 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1596 	    device_get_unit(sc->mps_dev));
1597 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1598 
1599 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1600 	    device_get_unit(sc->mps_dev));
1601 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1602 
1603 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1604 	    device_get_unit(sc->mps_dev));
1605 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1606 
1607 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1608 	    device_get_unit(sc->mps_dev));
1609 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1610 
1611 }
1612 
1613 static void
1614 mps_setup_sysctl(struct mps_softc *sc)
1615 {
1616 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1617 	struct sysctl_oid	*sysctl_tree = NULL;
1618 	char tmpstr[80], tmpstr2[80];
1619 
1620 	/*
1621 	 * Setup the sysctl variable so the user can change the debug level
1622 	 * on the fly.
1623 	 */
1624 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1625 	    device_get_unit(sc->mps_dev));
1626 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1627 
1628 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1629 	if (sysctl_ctx != NULL)
1630 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1631 
1632 	if (sysctl_tree == NULL) {
1633 		sysctl_ctx_init(&sc->sysctl_ctx);
1634 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1635 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1636 		    CTLFLAG_RD, 0, tmpstr);
1637 		if (sc->sysctl_tree == NULL)
1638 			return;
1639 		sysctl_ctx = &sc->sysctl_ctx;
1640 		sysctl_tree = sc->sysctl_tree;
1641 	}
1642 
1643 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1644 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1645 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1646 
1647 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1648 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1649 	    "Disable the use of MSI-X interrupts");
1650 
1651 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1652 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1653 	    "Disable the use of MSI interrupts");
1654 
1655 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1656 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1657 	    "User-defined maximum number of MSIX queues");
1658 
1659 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1660 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1661 	    "Negotiated number of MSIX queues");
1662 
1663 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1664 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1665 	    "Total number of allocated request frames");
1666 
1667 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1668 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1669 	    "Total number of allocated high priority request frames");
1670 
1671 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1672 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1673 	    "Total number of allocated reply frames");
1674 
1675 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1676 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1677 	    "Total number of event frames allocated");
1678 
1679 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1680 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1681 	    strlen(sc->fw_version), "firmware version");
1682 
1683 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1684 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1685 	    strlen(MPS_DRIVER_VERSION), "driver version");
1686 
1687 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1688 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1689 	    &sc->io_cmds_active, 0, "number of currently active commands");
1690 
1691 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1692 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1693 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1694 
1695 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1696 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1697 	    &sc->chain_free, 0, "number of free chain elements");
1698 
1699 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1700 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1701 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1702 
1703 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1704 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1705 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1706 
1707 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1708 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1709 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1710 	    "IOCFacts)");
1711 
1712 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1713 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1714 	    "enable SSU to SATA SSD/HDD at shutdown");
1715 
1716 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1717 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1718 	    &sc->chain_alloc_fail, "chain allocation failures");
1719 
1720 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1721 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1722 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1723 	    "spinup after SATA ID error");
1724 
1725 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1726 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1727 	    mps_mapping_dump, "A", "Mapping Table Dump");
1728 
1729 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1730 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1731 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1732 
1733 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1734 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1735 	    "Use the phy number for enumeration");
1736 }
1737 
1738 static struct mps_debug_string {
1739 	char	*name;
1740 	int	flag;
1741 } mps_debug_strings[] = {
1742 	{"info", MPS_INFO},
1743 	{"fault", MPS_FAULT},
1744 	{"event", MPS_EVENT},
1745 	{"log", MPS_LOG},
1746 	{"recovery", MPS_RECOVERY},
1747 	{"error", MPS_ERROR},
1748 	{"init", MPS_INIT},
1749 	{"xinfo", MPS_XINFO},
1750 	{"user", MPS_USER},
1751 	{"mapping", MPS_MAPPING},
1752 	{"trace", MPS_TRACE}
1753 };
1754 
1755 enum mps_debug_level_combiner {
1756 	COMB_NONE,
1757 	COMB_ADD,
1758 	COMB_SUB
1759 };
1760 
1761 static int
1762 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1763 {
1764 	struct mps_softc *sc;
1765 	struct mps_debug_string *string;
1766 	struct sbuf *sbuf;
1767 	char *buffer;
1768 	size_t sz;
1769 	int i, len, debug, error;
1770 
1771 	sc = (struct mps_softc *)arg1;
1772 
1773 	error = sysctl_wire_old_buffer(req, 0);
1774 	if (error != 0)
1775 		return (error);
1776 
1777 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1778 	debug = sc->mps_debug;
1779 
1780 	sbuf_printf(sbuf, "%#x", debug);
1781 
1782 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1783 	for (i = 0; i < sz; i++) {
1784 		string = &mps_debug_strings[i];
1785 		if (debug & string->flag)
1786 			sbuf_printf(sbuf, ",%s", string->name);
1787 	}
1788 
1789 	error = sbuf_finish(sbuf);
1790 	sbuf_delete(sbuf);
1791 
1792 	if (error || req->newptr == NULL)
1793 		return (error);
1794 
1795 	len = req->newlen - req->newidx;
1796 	if (len == 0)
1797 		return (0);
1798 
1799 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1800 	error = SYSCTL_IN(req, buffer, len);
1801 
1802 	mps_parse_debug(sc, buffer);
1803 
1804 	free(buffer, M_MPT2);
1805 	return (error);
1806 }
1807 
1808 static void
1809 mps_parse_debug(struct mps_softc *sc, char *list)
1810 {
1811 	struct mps_debug_string *string;
1812 	enum mps_debug_level_combiner op;
1813 	char *token, *endtoken;
1814 	size_t sz;
1815 	int flags, i;
1816 
1817 	if (list == NULL || *list == '\0')
1818 		return;
1819 
1820 	if (*list == '+') {
1821 		op = COMB_ADD;
1822 		list++;
1823 	} else if (*list == '-') {
1824 		op = COMB_SUB;
1825 		list++;
1826 	} else
1827 		op = COMB_NONE;
1828 	if (*list == '\0')
1829 		return;
1830 
1831 	flags = 0;
1832 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1833 	while ((token = strsep(&list, ":,")) != NULL) {
1834 
1835 		/* Handle integer flags */
1836 		flags |= strtol(token, &endtoken, 0);
1837 		if (token != endtoken)
1838 			continue;
1839 
1840 		/* Handle text flags */
1841 		for (i = 0; i < sz; i++) {
1842 			string = &mps_debug_strings[i];
1843 			if (strcasecmp(token, string->name) == 0) {
1844 				flags |= string->flag;
1845 				break;
1846 			}
1847 		}
1848 	}
1849 
1850 	switch (op) {
1851 	case COMB_NONE:
1852 		sc->mps_debug = flags;
1853 		break;
1854 	case COMB_ADD:
1855 		sc->mps_debug |= flags;
1856 		break;
1857 	case COMB_SUB:
1858 		sc->mps_debug &= (~flags);
1859 		break;
1860 	}
1861 
1862 	return;
1863 }
1864 
1865 int
1866 mps_attach(struct mps_softc *sc)
1867 {
1868 	int error;
1869 
1870 	MPS_FUNCTRACE(sc);
1871 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1872 
1873 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1874 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1875 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1876 	TAILQ_INIT(&sc->event_list);
1877 	timevalclear(&sc->lastfail);
1878 
1879 	if ((error = mps_transition_ready(sc)) != 0) {
1880 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1881 		    "ready\n");
1882 		return (error);
1883 	}
1884 
1885 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1886 	    M_ZERO|M_NOWAIT);
1887 	if(!sc->facts) {
1888 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1889 		    "exit\n");
1890 		return (ENOMEM);
1891 	}
1892 
1893 	/*
1894 	 * Get IOC Facts and allocate all structures based on this information.
1895 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1896 	 * Facts. If relevant values have changed in IOC Facts, this function
1897 	 * will free all of the memory based on IOC Facts and reallocate that
1898 	 * memory.  If this fails, any allocated memory should already be freed.
1899 	 */
1900 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1901 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1902 		    "failed with error %d, exit\n", error);
1903 		return (error);
1904 	}
1905 
1906 	/* Start the periodic watchdog check on the IOC Doorbell */
1907 	mps_periodic(sc);
1908 
1909 	/*
1910 	 * The portenable will kick off discovery events that will drive the
1911 	 * rest of the initialization process.  The CAM/SAS module will
1912 	 * hold up the boot sequence until discovery is complete.
1913 	 */
1914 	sc->mps_ich.ich_func = mps_startup;
1915 	sc->mps_ich.ich_arg = sc;
1916 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1917 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1918 		    "Cannot establish MPS config hook\n");
1919 		error = EINVAL;
1920 	}
1921 
1922 	/*
1923 	 * Allow IR to shutdown gracefully when shutdown occurs.
1924 	 */
1925 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1926 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1927 
1928 	if (sc->shutdown_eh == NULL)
1929 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1930 		    "shutdown event registration failed\n");
1931 
1932 	mps_setup_sysctl(sc);
1933 
1934 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1935 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1936 
1937 	return (error);
1938 }
1939 
1940 /* Run through any late-start handlers. */
1941 static void
1942 mps_startup(void *arg)
1943 {
1944 	struct mps_softc *sc;
1945 
1946 	sc = (struct mps_softc *)arg;
1947 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1948 
1949 	mps_lock(sc);
1950 	mps_unmask_intr(sc);
1951 
1952 	/* initialize device mapping tables */
1953 	mps_base_static_config_pages(sc);
1954 	mps_mapping_initialize(sc);
1955 	mpssas_startup(sc);
1956 	mps_unlock(sc);
1957 
1958 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
1959 	config_intrhook_disestablish(&sc->mps_ich);
1960 	sc->mps_ich.ich_arg = NULL;
1961 
1962 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1963 }
1964 
1965 /* Periodic watchdog.  Is called with the driver lock already held. */
1966 static void
1967 mps_periodic(void *arg)
1968 {
1969 	struct mps_softc *sc;
1970 	uint32_t db;
1971 
1972 	sc = (struct mps_softc *)arg;
1973 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1974 		return;
1975 
1976 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1977 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1978 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1979 		mps_reinit(sc);
1980 	}
1981 
1982 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1983 }
1984 
1985 static void
1986 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1987     MPI2_EVENT_NOTIFICATION_REPLY *event)
1988 {
1989 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1990 
1991 	MPS_DPRINT_EVENT(sc, generic, event);
1992 
1993 	switch (event->Event) {
1994 	case MPI2_EVENT_LOG_DATA:
1995 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1996 		if (sc->mps_debug & MPS_EVENT)
1997 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1998 		break;
1999 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2000 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2001 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2002 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2003 		     entry->LogSequence);
2004 		break;
2005 	default:
2006 		break;
2007 	}
2008 	return;
2009 }
2010 
2011 static int
2012 mps_attach_log(struct mps_softc *sc)
2013 {
2014 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
2015 
2016 	bzero(events, 16);
2017 	setbit(events, MPI2_EVENT_LOG_DATA);
2018 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2019 
2020 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
2021 	    &sc->mps_log_eh);
2022 
2023 	return (0);
2024 }
2025 
2026 static int
2027 mps_detach_log(struct mps_softc *sc)
2028 {
2029 
2030 	if (sc->mps_log_eh != NULL)
2031 		mps_deregister_events(sc, sc->mps_log_eh);
2032 	return (0);
2033 }
2034 
2035 /*
2036  * Free all of the driver resources and detach submodules.  Should be called
2037  * without the lock held.
2038  */
2039 int
2040 mps_free(struct mps_softc *sc)
2041 {
2042 	int error;
2043 
2044 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2045 	/* Turn off the watchdog */
2046 	mps_lock(sc);
2047 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
2048 	mps_unlock(sc);
2049 	/* Lock must not be held for this */
2050 	callout_drain(&sc->periodic);
2051 	callout_drain(&sc->device_check_callout);
2052 
2053 	if (((error = mps_detach_log(sc)) != 0) ||
2054 	    ((error = mps_detach_sas(sc)) != 0)) {
2055 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2056 		    "subsystems, exit\n");
2057 		return (error);
2058 	}
2059 
2060 	mps_detach_user(sc);
2061 
2062 	/* Put the IOC back in the READY state. */
2063 	mps_lock(sc);
2064 	if ((error = mps_transition_ready(sc)) != 0) {
2065 		mps_unlock(sc);
2066 		return (error);
2067 	}
2068 	mps_unlock(sc);
2069 
2070 	if (sc->facts != NULL)
2071 		free(sc->facts, M_MPT2);
2072 
2073 	/*
2074 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2075 	 * to free these buffers too.
2076 	 */
2077 	mps_iocfacts_free(sc);
2078 
2079 	if (sc->sysctl_tree != NULL)
2080 		sysctl_ctx_free(&sc->sysctl_ctx);
2081 
2082 	/* Deregister the shutdown function */
2083 	if (sc->shutdown_eh != NULL)
2084 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2085 
2086 	mtx_destroy(&sc->mps_mtx);
2087 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2088 
2089 	return (0);
2090 }
2091 
2092 static __inline void
2093 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2094 {
2095 	MPS_FUNCTRACE(sc);
2096 
2097 	if (cm == NULL) {
2098 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2099 		return;
2100 	}
2101 
2102 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2103 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2104 
2105 	if (cm->cm_complete != NULL) {
2106 		mps_dprint(sc, MPS_TRACE,
2107 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2108 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2109 			   cm->cm_reply);
2110 		cm->cm_complete(sc, cm);
2111 	}
2112 
2113 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2114 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2115 		wakeup(cm);
2116 	}
2117 
2118 	if (cm->cm_sc->io_cmds_active != 0) {
2119 		cm->cm_sc->io_cmds_active--;
2120 	} else {
2121 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2122 		    "out of sync - resynching to 0\n");
2123 	}
2124 }
2125 
2126 
2127 static void
2128 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2129 {
2130 	union loginfo_type {
2131 		u32     loginfo;
2132 		struct {
2133 			u32     subcode:16;
2134 			u32     code:8;
2135 			u32     originator:4;
2136 			u32     bus_type:4;
2137 		} dw;
2138 	};
2139 	union loginfo_type sas_loginfo;
2140 	char *originator_str = NULL;
2141 
2142 	sas_loginfo.loginfo = log_info;
2143 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2144 		return;
2145 
2146 	/* each nexus loss loginfo */
2147 	if (log_info == 0x31170000)
2148 		return;
2149 
2150 	/* eat the loginfos associated with task aborts */
2151 	if ((log_info == 30050000 || log_info ==
2152 	    0x31140000 || log_info == 0x31130000))
2153 		return;
2154 
2155 	switch (sas_loginfo.dw.originator) {
2156 	case 0:
2157 		originator_str = "IOP";
2158 		break;
2159 	case 1:
2160 		originator_str = "PL";
2161 		break;
2162 	case 2:
2163 		originator_str = "IR";
2164 		break;
2165 }
2166 
2167 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2168 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2169 	originator_str, sas_loginfo.dw.code,
2170 	sas_loginfo.dw.subcode);
2171 }
2172 
2173 static void
2174 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2175 {
2176 	MPI2DefaultReply_t *mpi_reply;
2177 	u16 sc_status;
2178 
2179 	mpi_reply = (MPI2DefaultReply_t*)reply;
2180 	sc_status = le16toh(mpi_reply->IOCStatus);
2181 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2182 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2183 }
2184 void
2185 mps_intr(void *data)
2186 {
2187 	struct mps_softc *sc;
2188 	uint32_t status;
2189 
2190 	sc = (struct mps_softc *)data;
2191 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2192 
2193 	/*
2194 	 * Check interrupt status register to flush the bus.  This is
2195 	 * needed for both INTx interrupts and driver-driven polling
2196 	 */
2197 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2198 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2199 		return;
2200 
2201 	mps_lock(sc);
2202 	mps_intr_locked(data);
2203 	mps_unlock(sc);
2204 	return;
2205 }
2206 
2207 /*
2208  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2209  * chip.  Hopefully this theory is correct.
2210  */
2211 void
2212 mps_intr_msi(void *data)
2213 {
2214 	struct mps_softc *sc;
2215 
2216 	sc = (struct mps_softc *)data;
2217 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2218 	mps_lock(sc);
2219 	mps_intr_locked(data);
2220 	mps_unlock(sc);
2221 	return;
2222 }
2223 
2224 /*
2225  * The locking is overly broad and simplistic, but easy to deal with for now.
2226  */
2227 void
2228 mps_intr_locked(void *data)
2229 {
2230 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2231 	struct mps_softc *sc;
2232 	struct mps_command *cm = NULL;
2233 	uint8_t flags;
2234 	u_int pq;
2235 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2236 	mps_fw_diagnostic_buffer_t *pBuffer;
2237 
2238 	sc = (struct mps_softc *)data;
2239 
2240 	pq = sc->replypostindex;
2241 	mps_dprint(sc, MPS_TRACE,
2242 	    "%s sc %p starting with replypostindex %u\n",
2243 	    __func__, sc, sc->replypostindex);
2244 
2245 	for ( ;; ) {
2246 		cm = NULL;
2247 		desc = &sc->post_queue[sc->replypostindex];
2248 		flags = desc->Default.ReplyFlags &
2249 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2250 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2251 		 || (le32toh(desc->Words.High) == 0xffffffff))
2252 			break;
2253 
2254 		/* increment the replypostindex now, so that event handlers
2255 		 * and cm completion handlers which decide to do a diag
2256 		 * reset can zero it without it getting incremented again
2257 		 * afterwards, and we break out of this loop on the next
2258 		 * iteration since the reply post queue has been cleared to
2259 		 * 0xFF and all descriptors look unused (which they are).
2260 		 */
2261 		if (++sc->replypostindex >= sc->pqdepth)
2262 			sc->replypostindex = 0;
2263 
2264 		switch (flags) {
2265 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2266 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2267 			cm->cm_reply = NULL;
2268 			break;
2269 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2270 		{
2271 			uint32_t baddr;
2272 			uint8_t *reply;
2273 
2274 			/*
2275 			 * Re-compose the reply address from the address
2276 			 * sent back from the chip.  The ReplyFrameAddress
2277 			 * is the lower 32 bits of the physical address of
2278 			 * particular reply frame.  Convert that address to
2279 			 * host format, and then use that to provide the
2280 			 * offset against the virtual address base
2281 			 * (sc->reply_frames).
2282 			 */
2283 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2284 			reply = sc->reply_frames +
2285 				(baddr - ((uint32_t)sc->reply_busaddr));
2286 			/*
2287 			 * Make sure the reply we got back is in a valid
2288 			 * range.  If not, go ahead and panic here, since
2289 			 * we'll probably panic as soon as we deference the
2290 			 * reply pointer anyway.
2291 			 */
2292 			if ((reply < sc->reply_frames)
2293 			 || (reply > (sc->reply_frames +
2294 			     (sc->fqdepth * sc->replyframesz)))) {
2295 				printf("%s: WARNING: reply %p out of range!\n",
2296 				       __func__, reply);
2297 				printf("%s: reply_frames %p, fqdepth %d, "
2298 				       "frame size %d\n", __func__,
2299 				       sc->reply_frames, sc->fqdepth,
2300 				       sc->replyframesz);
2301 				printf("%s: baddr %#x,\n", __func__, baddr);
2302 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
2303 				panic("Reply address out of range");
2304 			}
2305 			if (le16toh(desc->AddressReply.SMID) == 0) {
2306 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2307 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2308 					/*
2309 					 * If SMID is 0 for Diag Buffer Post,
2310 					 * this implies that the reply is due to
2311 					 * a release function with a status that
2312 					 * the buffer has been released.  Set
2313 					 * the buffer flags accordingly.
2314 					 */
2315 					rel_rep =
2316 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2317 					if ((le16toh(rel_rep->IOCStatus) &
2318 					    MPI2_IOCSTATUS_MASK) ==
2319 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2320 					{
2321 						pBuffer =
2322 						    &sc->fw_diag_buffer_list[
2323 						    rel_rep->BufferType];
2324 						pBuffer->valid_data = TRUE;
2325 						pBuffer->owned_by_firmware =
2326 						    FALSE;
2327 						pBuffer->immediate = FALSE;
2328 					}
2329 				} else
2330 					mps_dispatch_event(sc, baddr,
2331 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2332 					    reply);
2333 			} else {
2334 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2335 				cm->cm_reply = reply;
2336 				cm->cm_reply_data =
2337 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2338 			}
2339 			break;
2340 		}
2341 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2342 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2343 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2344 		default:
2345 			/* Unhandled */
2346 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2347 			    desc->Default.ReplyFlags);
2348 			cm = NULL;
2349 			break;
2350 		}
2351 
2352 
2353 		if (cm != NULL) {
2354 			// Print Error reply frame
2355 			if (cm->cm_reply)
2356 				mps_display_reply_info(sc,cm->cm_reply);
2357 			mps_complete_command(sc, cm);
2358 		}
2359 
2360 		desc->Words.Low = 0xffffffff;
2361 		desc->Words.High = 0xffffffff;
2362 	}
2363 
2364 	if (pq != sc->replypostindex) {
2365 		mps_dprint(sc, MPS_TRACE,
2366 		    "%s sc %p writing postindex %d\n",
2367 		    __func__, sc, sc->replypostindex);
2368 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2369 	}
2370 
2371 	return;
2372 }
2373 
2374 static void
2375 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2376     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2377 {
2378 	struct mps_event_handle *eh;
2379 	int event, handled = 0;
2380 
2381 	event = le16toh(reply->Event);
2382 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2383 		if (isset(eh->mask, event)) {
2384 			eh->callback(sc, data, reply);
2385 			handled++;
2386 		}
2387 	}
2388 
2389 	if (handled == 0)
2390 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2391 
2392 	/*
2393 	 * This is the only place that the event/reply should be freed.
2394 	 * Anything wanting to hold onto the event data should have
2395 	 * already copied it into their own storage.
2396 	 */
2397 	mps_free_reply(sc, data);
2398 }
2399 
2400 static void
2401 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2402 {
2403 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2404 
2405 	if (cm->cm_reply)
2406 		MPS_DPRINT_EVENT(sc, generic,
2407 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2408 
2409 	mps_free_command(sc, cm);
2410 
2411 	/* next, send a port enable */
2412 	mpssas_startup(sc);
2413 }
2414 
2415 /*
2416  * For both register_events and update_events, the caller supplies a bitmap
2417  * of events that it _wants_.  These functions then turn that into a bitmask
2418  * suitable for the controller.
2419  */
2420 int
2421 mps_register_events(struct mps_softc *sc, u32 *mask,
2422     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2423 {
2424 	struct mps_event_handle *eh;
2425 	int error = 0;
2426 
2427 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2428 	if(!eh) {
2429 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2430 		return (ENOMEM);
2431 	}
2432 	eh->callback = cb;
2433 	eh->data = data;
2434 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2435 	if (mask != NULL)
2436 		error = mps_update_events(sc, eh, mask);
2437 	*handle = eh;
2438 
2439 	return (error);
2440 }
2441 
2442 int
2443 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2444     u32 *mask)
2445 {
2446 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2447 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2448 	struct mps_command *cm;
2449 	int error, i;
2450 
2451 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2452 
2453 	if ((mask != NULL) && (handle != NULL))
2454 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2455 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2456 
2457 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2458 		sc->event_mask[i] = -1;
2459 
2460 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2461 		sc->event_mask[i] &= ~handle->mask[i];
2462 
2463 
2464 	if ((cm = mps_alloc_command(sc)) == NULL)
2465 		return (EBUSY);
2466 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2467 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2468 	evtreq->MsgFlags = 0;
2469 	evtreq->SASBroadcastPrimitiveMasks = 0;
2470 #ifdef MPS_DEBUG_ALL_EVENTS
2471 	{
2472 		u_char fullmask[16];
2473 		memset(fullmask, 0x00, 16);
2474 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2475 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2476 	}
2477 #else
2478         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2479                 evtreq->EventMasks[i] =
2480                     htole32(sc->event_mask[i]);
2481 #endif
2482 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2483 	cm->cm_data = NULL;
2484 
2485 	error = mps_wait_command(sc, &cm, 60, 0);
2486 	if (cm != NULL)
2487 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2488 	if ((reply == NULL) ||
2489 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2490 		error = ENXIO;
2491 
2492 	if (reply)
2493 		MPS_DPRINT_EVENT(sc, generic, reply);
2494 
2495 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2496 
2497 	if (cm != NULL)
2498 		mps_free_command(sc, cm);
2499 	return (error);
2500 }
2501 
2502 static int
2503 mps_reregister_events(struct mps_softc *sc)
2504 {
2505 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2506 	struct mps_command *cm;
2507 	struct mps_event_handle *eh;
2508 	int error, i;
2509 
2510 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2511 
2512 	/* first, reregister events */
2513 
2514 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2515 		sc->event_mask[i] = -1;
2516 
2517 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2518 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2519 			sc->event_mask[i] &= ~eh->mask[i];
2520 	}
2521 
2522 	if ((cm = mps_alloc_command(sc)) == NULL)
2523 		return (EBUSY);
2524 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2525 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2526 	evtreq->MsgFlags = 0;
2527 	evtreq->SASBroadcastPrimitiveMasks = 0;
2528 #ifdef MPS_DEBUG_ALL_EVENTS
2529 	{
2530 		u_char fullmask[16];
2531 		memset(fullmask, 0x00, 16);
2532 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2533 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2534 	}
2535 #else
2536         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2537                 evtreq->EventMasks[i] =
2538                     htole32(sc->event_mask[i]);
2539 #endif
2540 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2541 	cm->cm_data = NULL;
2542 	cm->cm_complete = mps_reregister_events_complete;
2543 
2544 	error = mps_map_command(sc, cm);
2545 
2546 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2547 	    error);
2548 	return (error);
2549 }
2550 
2551 void
2552 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2553 {
2554 
2555 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2556 	free(handle, M_MPT2);
2557 }
2558 
2559 /*
2560  * Add a chain element as the next SGE for the specified command.
2561  * Reset cm_sge and cm_sgesize to indicate all the available space.
2562  */
2563 static int
2564 mps_add_chain(struct mps_command *cm)
2565 {
2566 	MPI2_SGE_CHAIN32 *sgc;
2567 	struct mps_chain *chain;
2568 	u_int space;
2569 
2570 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2571 		panic("MPS: Need SGE Error Code\n");
2572 
2573 	chain = mps_alloc_chain(cm->cm_sc);
2574 	if (chain == NULL)
2575 		return (ENOBUFS);
2576 
2577 	space = cm->cm_sc->reqframesz;
2578 
2579 	/*
2580 	 * Note: a double-linked list is used to make it easier to
2581 	 * walk for debugging.
2582 	 */
2583 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2584 
2585 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2586 	sgc->Length = htole16(space);
2587 	sgc->NextChainOffset = 0;
2588 	/* TODO Looks like bug in Setting sgc->Flags.
2589 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2590 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2591 	 *	This is fine.. because we are not using simple element. In case of
2592 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2593  	 */
2594 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2595 	sgc->Address = htole32(chain->chain_busaddr);
2596 
2597 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2598 	cm->cm_sglsize = space;
2599 	return (0);
2600 }
2601 
2602 /*
2603  * Add one scatter-gather element (chain, simple, transaction context)
2604  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2605  * cm_sge as the remaining size and pointer to the next SGE to fill
2606  * in, respectively.
2607  */
2608 int
2609 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2610 {
2611 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2612 	MPI2_SGE_SIMPLE64 *sge = sgep;
2613 	int error, type;
2614 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2615 
2616 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2617 
2618 #ifdef INVARIANTS
2619 	switch (type) {
2620 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2621 		if (len != tc->DetailsLength + 4)
2622 			panic("TC %p length %u or %zu?", tc,
2623 			    tc->DetailsLength + 4, len);
2624 		}
2625 		break;
2626 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2627 		/* Driver only uses 32-bit chain elements */
2628 		if (len != MPS_SGC_SIZE)
2629 			panic("CHAIN %p length %u or %zu?", sgep,
2630 			    MPS_SGC_SIZE, len);
2631 		break;
2632 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2633 		/* Driver only uses 64-bit SGE simple elements */
2634 		if (len != MPS_SGE64_SIZE)
2635 			panic("SGE simple %p length %u or %zu?", sge,
2636 			    MPS_SGE64_SIZE, len);
2637 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2638 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2639 			panic("SGE simple %p not marked 64-bit?", sge);
2640 
2641 		break;
2642 	default:
2643 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2644 	}
2645 #endif
2646 
2647 	/*
2648 	 * case 1: 1 more segment, enough room for it
2649 	 * case 2: 2 more segments, enough room for both
2650 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2651 	 * case 4: >=1 more segment, enough room for only a chain
2652 	 * case 5: >=1 more segment, no room for anything (error)
2653          */
2654 
2655 	/*
2656 	 * There should be room for at least a chain element, or this
2657 	 * code is buggy.  Case (5).
2658 	 */
2659 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2660 		panic("MPS: Need SGE Error Code\n");
2661 
2662 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2663 		/*
2664 		 * 1 or more segment, enough room for only a chain.
2665 		 * Hope the previous element wasn't a Simple entry
2666 		 * that needed to be marked with
2667 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2668 		 */
2669 		if ((error = mps_add_chain(cm)) != 0)
2670 			return (error);
2671 	}
2672 
2673 	if (segsleft >= 2 &&
2674 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2675 		/*
2676 		 * There are 2 or more segments left to add, and only
2677 		 * enough room for 1 and a chain.  Case (3).
2678 		 *
2679 		 * Mark as last element in this chain if necessary.
2680 		 */
2681 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2682 			sge->FlagsLength |= htole32(
2683 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2684 		}
2685 
2686 		/*
2687 		 * Add the item then a chain.  Do the chain now,
2688 		 * rather than on the next iteration, to simplify
2689 		 * understanding the code.
2690 		 */
2691 		cm->cm_sglsize -= len;
2692 		bcopy(sgep, cm->cm_sge, len);
2693 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2694 		return (mps_add_chain(cm));
2695 	}
2696 
2697 #ifdef INVARIANTS
2698 	/* Case 1: 1 more segment, enough room for it. */
2699 	if (segsleft == 1 && cm->cm_sglsize < len)
2700 		panic("1 seg left and no room? %u versus %zu",
2701 		    cm->cm_sglsize, len);
2702 
2703 	/* Case 2: 2 more segments, enough room for both */
2704 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2705 		panic("2 segs left and no room? %u versus %zu",
2706 		    cm->cm_sglsize, len);
2707 #endif
2708 
2709 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2710 		/*
2711 		 * If this is a bi-directional request, need to account for that
2712 		 * here.  Save the pre-filled sge values.  These will be used
2713 		 * either for the 2nd SGL or for a single direction SGL.  If
2714 		 * cm_out_len is non-zero, this is a bi-directional request, so
2715 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2716 		 * fill in the IN SGL.  Note that at this time, when filling in
2717 		 * 2 SGL's for a bi-directional request, they both use the same
2718 		 * DMA buffer (same cm command).
2719 		 */
2720 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2721 		saved_address_low = sge->Address.Low;
2722 		saved_address_high = sge->Address.High;
2723 		if (cm->cm_out_len) {
2724 			sge->FlagsLength = htole32(cm->cm_out_len |
2725 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2726 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2727 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2728 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2729 			    MPI2_SGE_FLAGS_SHIFT));
2730 			cm->cm_sglsize -= len;
2731 			bcopy(sgep, cm->cm_sge, len);
2732 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2733 			    + len);
2734 		}
2735 		saved_buf_len |=
2736 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2737 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2738 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2739 		    MPI2_SGE_FLAGS_END_OF_LIST |
2740 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2741 		    MPI2_SGE_FLAGS_SHIFT);
2742 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2743 			saved_buf_len |=
2744 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2745 			    MPI2_SGE_FLAGS_SHIFT);
2746 		} else {
2747 			saved_buf_len |=
2748 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2749 			    MPI2_SGE_FLAGS_SHIFT);
2750 		}
2751 		sge->FlagsLength = htole32(saved_buf_len);
2752 		sge->Address.Low = saved_address_low;
2753 		sge->Address.High = saved_address_high;
2754 	}
2755 
2756 	cm->cm_sglsize -= len;
2757 	bcopy(sgep, cm->cm_sge, len);
2758 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2759 	return (0);
2760 }
2761 
2762 /*
2763  * Add one dma segment to the scatter-gather list for a command.
2764  */
2765 int
2766 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2767     int segsleft)
2768 {
2769 	MPI2_SGE_SIMPLE64 sge;
2770 
2771 	/*
2772 	 * This driver always uses 64-bit address elements for simplicity.
2773 	 */
2774 	bzero(&sge, sizeof(sge));
2775 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2776 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2777 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2778 	mps_from_u64(pa, &sge.Address);
2779 
2780 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2781 }
2782 
2783 static void
2784 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2785 {
2786 	struct mps_softc *sc;
2787 	struct mps_command *cm;
2788 	u_int i, dir, sflags;
2789 
2790 	cm = (struct mps_command *)arg;
2791 	sc = cm->cm_sc;
2792 
2793 	/*
2794 	 * In this case, just print out a warning and let the chip tell the
2795 	 * user they did the wrong thing.
2796 	 */
2797 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2798 		mps_dprint(sc, MPS_ERROR,
2799 			   "%s: warning: busdma returned %d segments, "
2800 			   "more than the %d allowed\n", __func__, nsegs,
2801 			   cm->cm_max_segs);
2802 	}
2803 
2804 	/*
2805 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2806 	 * here.  In that case, both direction flags will be set.
2807 	 */
2808 	sflags = 0;
2809 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2810 		/*
2811 		 * We have to add a special case for SMP passthrough, there
2812 		 * is no easy way to generically handle it.  The first
2813 		 * S/G element is used for the command (therefore the
2814 		 * direction bit needs to be set).  The second one is used
2815 		 * for the reply.  We'll leave it to the caller to make
2816 		 * sure we only have two buffers.
2817 		 */
2818 		/*
2819 		 * Even though the busdma man page says it doesn't make
2820 		 * sense to have both direction flags, it does in this case.
2821 		 * We have one s/g element being accessed in each direction.
2822 		 */
2823 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2824 
2825 		/*
2826 		 * Set the direction flag on the first buffer in the SMP
2827 		 * passthrough request.  We'll clear it for the second one.
2828 		 */
2829 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2830 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2831 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2832 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2833 		dir = BUS_DMASYNC_PREWRITE;
2834 	} else
2835 		dir = BUS_DMASYNC_PREREAD;
2836 
2837 	for (i = 0; i < nsegs; i++) {
2838 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2839 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2840 		}
2841 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2842 		    sflags, nsegs - i);
2843 		if (error != 0) {
2844 			/* Resource shortage, roll back! */
2845 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2846 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2847 				    "consider increasing hw.mps.max_chains.\n");
2848 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2849 			mps_complete_command(sc, cm);
2850 			return;
2851 		}
2852 	}
2853 
2854 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2855 	mps_enqueue_request(sc, cm);
2856 
2857 	return;
2858 }
2859 
2860 static void
2861 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2862 	     int error)
2863 {
2864 	mps_data_cb(arg, segs, nsegs, error);
2865 }
2866 
2867 /*
2868  * This is the routine to enqueue commands ansynchronously.
2869  * Note that the only error path here is from bus_dmamap_load(), which can
2870  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2871  * assumed that if you have a command in-hand, then you have enough credits
2872  * to use it.
2873  */
2874 int
2875 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2876 {
2877 	int error = 0;
2878 
2879 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2880 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2881 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2882 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2883 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2884 		    cm->cm_data, mps_data_cb, cm, 0);
2885 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2886 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2887 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2888 	} else {
2889 		/* Add a zero-length element as needed */
2890 		if (cm->cm_sge != NULL)
2891 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2892 		mps_enqueue_request(sc, cm);
2893 	}
2894 
2895 	return (error);
2896 }
2897 
2898 /*
2899  * This is the routine to enqueue commands synchronously.  An error of
2900  * EINPROGRESS from mps_map_command() is ignored since the command will
2901  * be executed and enqueued automatically.  Other errors come from msleep().
2902  */
2903 int
2904 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2905     int sleep_flag)
2906 {
2907 	int error, rc;
2908 	struct timeval cur_time, start_time;
2909 	struct mps_command *cm = *cmp;
2910 
2911 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2912 		return  EBUSY;
2913 
2914 	cm->cm_complete = NULL;
2915 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2916 	error = mps_map_command(sc, cm);
2917 	if ((error != 0) && (error != EINPROGRESS))
2918 		return (error);
2919 
2920 	/*
2921 	 * Check for context and wait for 50 mSec at a time until time has
2922 	 * expired or the command has finished.  If msleep can't be used, need
2923 	 * to poll.
2924 	 */
2925 	if (curthread->td_no_sleeping != 0)
2926 		sleep_flag = NO_SLEEP;
2927 	getmicrouptime(&start_time);
2928 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2929 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2930 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2931 		if (error == EWOULDBLOCK) {
2932 			/*
2933 			 * Record the actual elapsed time in the case of a
2934 			 * timeout for the message below.
2935 			 */
2936 			getmicrouptime(&cur_time);
2937 			timevalsub(&cur_time, &start_time);
2938 		}
2939 	} else {
2940 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2941 			mps_intr_locked(sc);
2942 			if (sleep_flag == CAN_SLEEP)
2943 				pause("mpswait", hz/20);
2944 			else
2945 				DELAY(50000);
2946 
2947 			getmicrouptime(&cur_time);
2948 			timevalsub(&cur_time, &start_time);
2949 			if (cur_time.tv_sec > timeout) {
2950 				error = EWOULDBLOCK;
2951 				break;
2952 			}
2953 		}
2954 	}
2955 
2956 	if (error == EWOULDBLOCK) {
2957 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2958 		    " elapsed=%jd\n", __func__, timeout,
2959 		    (intmax_t)cur_time.tv_sec);
2960 		rc = mps_reinit(sc);
2961 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2962 		    "failed");
2963 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
2964 			/*
2965 			 * Tell the caller that we freed the command in a
2966 			 * reinit.
2967 			 */
2968 			*cmp = NULL;
2969 		}
2970 		error = ETIMEDOUT;
2971 	}
2972 	return (error);
2973 }
2974 
2975 /*
2976  * The MPT driver had a verbose interface for config pages.  In this driver,
2977  * reduce it to much simpler terms, similar to the Linux driver.
2978  */
2979 int
2980 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2981 {
2982 	MPI2_CONFIG_REQUEST *req;
2983 	struct mps_command *cm;
2984 	int error;
2985 
2986 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2987 		return (EBUSY);
2988 	}
2989 
2990 	cm = mps_alloc_command(sc);
2991 	if (cm == NULL) {
2992 		return (EBUSY);
2993 	}
2994 
2995 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2996 	req->Function = MPI2_FUNCTION_CONFIG;
2997 	req->Action = params->action;
2998 	req->SGLFlags = 0;
2999 	req->ChainOffset = 0;
3000 	req->PageAddress = params->page_address;
3001 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3002 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3003 
3004 		hdr = &params->hdr.Ext;
3005 		req->ExtPageType = hdr->ExtPageType;
3006 		req->ExtPageLength = hdr->ExtPageLength;
3007 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3008 		req->Header.PageLength = 0; /* Must be set to zero */
3009 		req->Header.PageNumber = hdr->PageNumber;
3010 		req->Header.PageVersion = hdr->PageVersion;
3011 	} else {
3012 		MPI2_CONFIG_PAGE_HEADER *hdr;
3013 
3014 		hdr = &params->hdr.Struct;
3015 		req->Header.PageType = hdr->PageType;
3016 		req->Header.PageNumber = hdr->PageNumber;
3017 		req->Header.PageLength = hdr->PageLength;
3018 		req->Header.PageVersion = hdr->PageVersion;
3019 	}
3020 
3021 	cm->cm_data = params->buffer;
3022 	cm->cm_length = params->length;
3023 	if (cm->cm_data != NULL) {
3024 		cm->cm_sge = &req->PageBufferSGE;
3025 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3026 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
3027 	} else
3028 		cm->cm_sge = NULL;
3029 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3030 
3031 	cm->cm_complete_data = params;
3032 	if (params->callback != NULL) {
3033 		cm->cm_complete = mps_config_complete;
3034 		return (mps_map_command(sc, cm));
3035 	} else {
3036 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
3037 		if (error) {
3038 			mps_dprint(sc, MPS_FAULT,
3039 			    "Error %d reading config page\n", error);
3040 			if (cm != NULL)
3041 				mps_free_command(sc, cm);
3042 			return (error);
3043 		}
3044 		mps_config_complete(sc, cm);
3045 	}
3046 
3047 	return (0);
3048 }
3049 
3050 int
3051 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3052 {
3053 	return (EINVAL);
3054 }
3055 
3056 static void
3057 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3058 {
3059 	MPI2_CONFIG_REPLY *reply;
3060 	struct mps_config_params *params;
3061 
3062 	MPS_FUNCTRACE(sc);
3063 	params = cm->cm_complete_data;
3064 
3065 	if (cm->cm_data != NULL) {
3066 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3067 		    BUS_DMASYNC_POSTREAD);
3068 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3069 	}
3070 
3071 	/*
3072 	 * XXX KDM need to do more error recovery?  This results in the
3073 	 * device in question not getting probed.
3074 	 */
3075 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3076 		params->status = MPI2_IOCSTATUS_BUSY;
3077 		goto done;
3078 	}
3079 
3080 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3081 	if (reply == NULL) {
3082 		params->status = MPI2_IOCSTATUS_BUSY;
3083 		goto done;
3084 	}
3085 	params->status = reply->IOCStatus;
3086 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3087 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3088 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3089 		params->hdr.Ext.PageType = reply->Header.PageType;
3090 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3091 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3092 	} else {
3093 		params->hdr.Struct.PageType = reply->Header.PageType;
3094 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3095 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3096 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3097 	}
3098 
3099 done:
3100 	mps_free_command(sc, cm);
3101 	if (params->callback != NULL)
3102 		params->callback(sc, params);
3103 
3104 	return;
3105 }
3106