xref: /freebsd/sys/dev/mps/mps.c (revision d8a0fe102c0cfdfcd5b818f850eff09d8536c9bc)
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 
115 /*
116  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
117  * any state and back to its initialization state machine.
118  */
119 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
120 
121 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
122  * Compiler only support unint64_t to be passed as argument.
123  * Otherwise it will throw below error
124  * "aggregate value used where an integer was expected"
125  */
126 
127 typedef union _reply_descriptor {
128         u64 word;
129         struct {
130                 u32 low;
131                 u32 high;
132         } u;
133 }reply_descriptor,address_descriptor;
134 
135 /* Rate limit chain-fail messages to 1 per minute */
136 static struct timeval mps_chainfail_interval = { 60, 0 };
137 
138 /*
139  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
140  * If this function is called from process context, it can sleep
141  * and there is no harm to sleep, in case if this fuction is called
142  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
143  * based on sleep flags driver will call either msleep, pause or DELAY.
144  * msleep and pause are of same variant, but pause is used when mps_mtx
145  * is not hold by driver.
146  *
147  */
148 static int
149 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
150 {
151 	uint32_t reg;
152 	int i, error, tries = 0;
153 	uint8_t first_wait_done = FALSE;
154 
155 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
156 
157 	/* Clear any pending interrupts */
158 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
159 
160 	/*
161 	 * Force NO_SLEEP for threads prohibited to sleep
162  	 * e.a Thread from interrupt handler are prohibited to sleep.
163  	 */
164 	if (curthread->td_no_sleeping != 0)
165 		sleep_flag = NO_SLEEP;
166 
167 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
168 
169 	/* Push the magic sequence */
170 	error = ETIMEDOUT;
171 	while (tries++ < 20) {
172 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
173 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
174 			    mpt2_reset_magic[i]);
175 		/* wait 100 msec */
176 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
177 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
178 			    "mpsdiag", hz/10);
179 		else if (sleep_flag == CAN_SLEEP)
180 			pause("mpsdiag", hz/10);
181 		else
182 			DELAY(100 * 1000);
183 
184 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
185 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
186 			error = 0;
187 			break;
188 		}
189 	}
190 	if (error) {
191 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
192 		    error);
193 		return (error);
194 	}
195 
196 	/* Send the actual reset.  XXX need to refresh the reg? */
197 	reg |= MPI2_DIAG_RESET_ADAPTER;
198 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
199 		reg);
200 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
201 
202 	/* Wait up to 300 seconds in 50ms intervals */
203 	error = ETIMEDOUT;
204 	for (i = 0; i < 6000; i++) {
205 		/*
206 		 * Wait 50 msec. If this is the first time through, wait 256
207 		 * msec to satisfy Diag Reset timing requirements.
208 		 */
209 		if (first_wait_done) {
210 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
211 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
212 				    "mpsdiag", hz/20);
213 			else if (sleep_flag == CAN_SLEEP)
214 				pause("mpsdiag", hz/20);
215 			else
216 				DELAY(50 * 1000);
217 		} else {
218 			DELAY(256 * 1000);
219 			first_wait_done = TRUE;
220 		}
221 		/*
222 		 * Check for the RESET_ADAPTER bit to be cleared first, then
223 		 * wait for the RESET state to be cleared, which takes a little
224 		 * longer.
225 		 */
226 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
227 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
228 			continue;
229 		}
230 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
231 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
232 			error = 0;
233 			break;
234 		}
235 	}
236 	if (error) {
237 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
238 		    error);
239 		return (error);
240 	}
241 
242 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
243 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
244 
245 	return (0);
246 }
247 
248 static int
249 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
250 {
251 	int error;
252 
253 	MPS_FUNCTRACE(sc);
254 
255 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
256 
257 	error = 0;
258 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
259 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
260 	    MPI2_DOORBELL_FUNCTION_SHIFT);
261 
262 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
263 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
264 		    "Doorbell handshake failed\n");
265 		error = ETIMEDOUT;
266 	}
267 
268 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
269 	return (error);
270 }
271 
272 static int
273 mps_transition_ready(struct mps_softc *sc)
274 {
275 	uint32_t reg, state;
276 	int error, tries = 0;
277 	int sleep_flags;
278 
279 	MPS_FUNCTRACE(sc);
280 	/* If we are in attach call, do not sleep */
281 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
282 					? CAN_SLEEP:NO_SLEEP;
283 	error = 0;
284 
285 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
286 	   __func__, sleep_flags);
287 
288 	while (tries++ < 1200) {
289 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
290 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
291 
292 		/*
293 		 * Ensure the IOC is ready to talk.  If it's not, try
294 		 * resetting it.
295 		 */
296 		if (reg & MPI2_DOORBELL_USED) {
297 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
298 			    "reset\n");
299 			mps_diag_reset(sc, sleep_flags);
300 			DELAY(50000);
301 			continue;
302 		}
303 
304 		/* Is the adapter owned by another peer? */
305 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
306 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
307 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
308 			    "control of another peer host, aborting "
309 			    "initialization.\n");
310 			error = ENXIO;
311 			break;
312 		}
313 
314 		state = reg & MPI2_IOC_STATE_MASK;
315 		if (state == MPI2_IOC_STATE_READY) {
316 			/* Ready to go! */
317 			error = 0;
318 			break;
319 		} else if (state == MPI2_IOC_STATE_FAULT) {
320 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
321 			    "state 0x%x, resetting\n",
322 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
323 			mps_diag_reset(sc, sleep_flags);
324 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
325 			/* Need to take ownership */
326 			mps_message_unit_reset(sc, sleep_flags);
327 		} else if (state == MPI2_IOC_STATE_RESET) {
328 			/* Wait a bit, IOC might be in transition */
329 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
330 			    "IOC in unexpected reset state\n");
331 		} else {
332 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
333 			    "IOC in unknown state 0x%x\n", state);
334 			error = EINVAL;
335 			break;
336 		}
337 
338 		/* Wait 50ms for things to settle down. */
339 		DELAY(50000);
340 	}
341 
342 	if (error)
343 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
344 		    "Cannot transition IOC to ready\n");
345 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
346 
347 	return (error);
348 }
349 
350 static int
351 mps_transition_operational(struct mps_softc *sc)
352 {
353 	uint32_t reg, state;
354 	int error;
355 
356 	MPS_FUNCTRACE(sc);
357 
358 	error = 0;
359 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
360 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
361 
362 	state = reg & MPI2_IOC_STATE_MASK;
363 	if (state != MPI2_IOC_STATE_READY) {
364 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
365 		if ((error = mps_transition_ready(sc)) != 0) {
366 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
367 			    "failed to transition ready, exit\n");
368 			return (error);
369 		}
370 	}
371 
372 	error = mps_send_iocinit(sc);
373 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
374 
375 	return (error);
376 }
377 
378 static void
379 mps_resize_queues(struct mps_softc *sc)
380 {
381 	int reqcr, prireqcr;
382 
383 	/*
384 	 * Size the queues. Since the reply queues always need one free
385 	 * entry, we'll deduct one reply message here.  The LSI documents
386 	 * suggest instead to add a count to the request queue, but I think
387 	 * that it's better to deduct from reply queue.
388 	 */
389 	prireqcr = MAX(1, sc->max_prireqframes);
390 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
391 
392 	reqcr = MAX(2, sc->max_reqframes);
393 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
394 
395 	sc->num_reqs = prireqcr + reqcr;
396 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
397 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
398 
399 	/*
400 	 * Figure out the number of MSIx-based queues.  If the firmware or
401 	 * user has done something crazy and not allowed enough credit for
402 	 * the queues to be useful then don't enable multi-queue.
403 	 */
404 	if (sc->facts->MaxMSIxVectors < 2)
405 		sc->msi_msgs = 1;
406 
407 	if (sc->msi_msgs > 1) {
408 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
409 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
410 		if (sc->num_reqs / sc->msi_msgs < 2)
411 			sc->msi_msgs = 1;
412 	}
413 
414 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
415 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
416 }
417 
418 /*
419  * This is called during attach and when re-initializing due to a Diag Reset.
420  * IOC Facts is used to allocate many of the structures needed by the driver.
421  * If called from attach, de-allocation is not required because the driver has
422  * not allocated any structures yet, but if called from a Diag Reset, previously
423  * allocated structures based on IOC Facts will need to be freed and re-
424  * allocated bases on the latest IOC Facts.
425  */
426 static int
427 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
428 {
429 	int error;
430 	Mpi2IOCFactsReply_t saved_facts;
431 	uint8_t saved_mode, reallocating;
432 
433 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
434 
435 	/* Save old IOC Facts and then only reallocate if Facts have changed */
436 	if (!attaching) {
437 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
438 	}
439 
440 	/*
441 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
442 	 * a re-initialization and only return the error if attaching so the OS
443 	 * can handle it.
444 	 */
445 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
446 		if (attaching) {
447 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
448 			    "IOC Facts with error %d, exit\n", error);
449 			return (error);
450 		} else {
451 			panic("%s failed to get IOC Facts with error %d\n",
452 			    __func__, error);
453 		}
454 	}
455 
456 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
457 
458 	snprintf(sc->fw_version, sizeof(sc->fw_version),
459 	    "%02d.%02d.%02d.%02d",
460 	    sc->facts->FWVersion.Struct.Major,
461 	    sc->facts->FWVersion.Struct.Minor,
462 	    sc->facts->FWVersion.Struct.Unit,
463 	    sc->facts->FWVersion.Struct.Dev);
464 
465 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
466 	    MPS_DRIVER_VERSION);
467 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
468 	     sc->facts->IOCCapabilities,
469 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
470 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
471 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
472 
473 	/*
474 	 * If the chip doesn't support event replay then a hard reset will be
475 	 * required to trigger a full discovery.  Do the reset here then
476 	 * retransition to Ready.  A hard reset might have already been done,
477 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
478 	 * for a Diag Reset.
479 	 */
480 	if (attaching && ((sc->facts->IOCCapabilities &
481 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
482 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
483 		mps_diag_reset(sc, NO_SLEEP);
484 		if ((error = mps_transition_ready(sc)) != 0) {
485 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
486 			    "transition to ready with error %d, exit\n",
487 			    error);
488 			return (error);
489 		}
490 	}
491 
492 	/*
493 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
494 	 * changed from the previous IOC Facts, log a warning, but only if
495 	 * checking this after a Diag Reset and not during attach.
496 	 */
497 	saved_mode = sc->ir_firmware;
498 	if (sc->facts->IOCCapabilities &
499 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
500 		sc->ir_firmware = 1;
501 	if (!attaching) {
502 		if (sc->ir_firmware != saved_mode) {
503 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
504 			    "in IOC Facts does not match previous mode\n");
505 		}
506 	}
507 
508 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
509 	reallocating = FALSE;
510 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
511 
512 	if ((!attaching) &&
513 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
514 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
515 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
516 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
517 	    (saved_facts.ProductID != sc->facts->ProductID) ||
518 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
519 	    (saved_facts.IOCRequestFrameSize !=
520 	    sc->facts->IOCRequestFrameSize) ||
521 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
522 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
523 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
524 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
525 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
526 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
527 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
528 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
529 	    (saved_facts.MaxPersistentEntries !=
530 	    sc->facts->MaxPersistentEntries))) {
531 		reallocating = TRUE;
532 
533 		/* Record that we reallocated everything */
534 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
535 	}
536 
537 	/*
538 	 * Some things should be done if attaching or re-allocating after a Diag
539 	 * Reset, but are not needed after a Diag Reset if the FW has not
540 	 * changed.
541 	 */
542 	if (attaching || reallocating) {
543 		/*
544 		 * Check if controller supports FW diag buffers and set flag to
545 		 * enable each type.
546 		 */
547 		if (sc->facts->IOCCapabilities &
548 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
549 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
550 			    enabled = TRUE;
551 		if (sc->facts->IOCCapabilities &
552 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
553 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
554 			    enabled = TRUE;
555 		if (sc->facts->IOCCapabilities &
556 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
557 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
558 			    enabled = TRUE;
559 
560 		/*
561 		 * Set flag if EEDP is supported and if TLR is supported.
562 		 */
563 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
564 			sc->eedp_enabled = TRUE;
565 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
566 			sc->control_TLR = TRUE;
567 
568 		mps_resize_queues(sc);
569 
570 		/*
571 		 * Initialize all Tail Queues
572 		 */
573 		TAILQ_INIT(&sc->req_list);
574 		TAILQ_INIT(&sc->high_priority_req_list);
575 		TAILQ_INIT(&sc->chain_list);
576 		TAILQ_INIT(&sc->tm_list);
577 	}
578 
579 	/*
580 	 * If doing a Diag Reset and the FW is significantly different
581 	 * (reallocating will be set above in IOC Facts comparison), then all
582 	 * buffers based on the IOC Facts will need to be freed before they are
583 	 * reallocated.
584 	 */
585 	if (reallocating) {
586 		mps_iocfacts_free(sc);
587 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
588 		    saved_facts.MaxVolumes);
589 	}
590 
591 	/*
592 	 * Any deallocation has been completed.  Now start reallocating
593 	 * if needed.  Will only need to reallocate if attaching or if the new
594 	 * IOC Facts are different from the previous IOC Facts after a Diag
595 	 * Reset. Targets have already been allocated above if needed.
596 	 */
597 	error = 0;
598 	while (attaching || reallocating) {
599 		if ((error = mps_alloc_hw_queues(sc)) != 0)
600 			break;
601 		if ((error = mps_alloc_replies(sc)) != 0)
602 			break;
603 		if ((error = mps_alloc_requests(sc)) != 0)
604 			break;
605 		if ((error = mps_alloc_queues(sc)) != 0)
606 			break;
607 
608 		break;
609 	}
610 	if (error) {
611 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
612 		    "Failed to alloc queues with error %d\n", error);
613 		mps_free(sc);
614 		return (error);
615 	}
616 
617 	/* Always initialize the queues */
618 	bzero(sc->free_queue, sc->fqdepth * 4);
619 	mps_init_queues(sc);
620 
621 	/*
622 	 * Always get the chip out of the reset state, but only panic if not
623 	 * attaching.  If attaching and there is an error, that is handled by
624 	 * the OS.
625 	 */
626 	error = mps_transition_operational(sc);
627 	if (error != 0) {
628 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
629 		    "transition to operational with error %d\n", error);
630 		mps_free(sc);
631 		return (error);
632 	}
633 
634 	/*
635 	 * Finish the queue initialization.
636 	 * These are set here instead of in mps_init_queues() because the
637 	 * IOC resets these values during the state transition in
638 	 * mps_transition_operational().  The free index is set to 1
639 	 * because the corresponding index in the IOC is set to 0, and the
640 	 * IOC treats the queues as full if both are set to the same value.
641 	 * Hence the reason that the queue can't hold all of the possible
642 	 * replies.
643 	 */
644 	sc->replypostindex = 0;
645 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
646 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
647 
648 	/*
649 	 * Attach the subsystems so they can prepare their event masks.
650 	 * XXX Should be dynamic so that IM/IR and user modules can attach
651 	 */
652 	error = 0;
653 	while (attaching) {
654 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
655 		if ((error = mps_attach_log(sc)) != 0)
656 			break;
657 		if ((error = mps_attach_sas(sc)) != 0)
658 			break;
659 		if ((error = mps_attach_user(sc)) != 0)
660 			break;
661 		break;
662 	}
663 	if (error) {
664 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
665 		    "subsystems: error %d\n", error);
666 		mps_free(sc);
667 		return (error);
668 	}
669 
670 	/*
671 	 * XXX If the number of MSI-X vectors changes during re-init, this
672 	 * won't see it and adjust.
673 	 */
674 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
675 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
676 		    "interrupts\n");
677 		mps_free(sc);
678 		return (error);
679 	}
680 
681 	/*
682 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
683 	 * reset it after a Diag Reset, just in case.
684 	 */
685 	sc->WD_available = FALSE;
686 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
687 		sc->WD_available = TRUE;
688 
689 	return (error);
690 }
691 
692 /*
693  * This is called if memory is being free (during detach for example) and when
694  * buffers need to be reallocated due to a Diag Reset.
695  */
696 static void
697 mps_iocfacts_free(struct mps_softc *sc)
698 {
699 	struct mps_command *cm;
700 	int i;
701 
702 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
703 
704 	if (sc->free_busaddr != 0)
705 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
706 	if (sc->free_queue != NULL)
707 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
708 		    sc->queues_map);
709 	if (sc->queues_dmat != NULL)
710 		bus_dma_tag_destroy(sc->queues_dmat);
711 
712 	if (sc->chain_busaddr != 0)
713 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
714 	if (sc->chain_frames != NULL)
715 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
716 		    sc->chain_map);
717 	if (sc->chain_dmat != NULL)
718 		bus_dma_tag_destroy(sc->chain_dmat);
719 
720 	if (sc->sense_busaddr != 0)
721 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
722 	if (sc->sense_frames != NULL)
723 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
724 		    sc->sense_map);
725 	if (sc->sense_dmat != NULL)
726 		bus_dma_tag_destroy(sc->sense_dmat);
727 
728 	if (sc->reply_busaddr != 0)
729 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
730 	if (sc->reply_frames != NULL)
731 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
732 		    sc->reply_map);
733 	if (sc->reply_dmat != NULL)
734 		bus_dma_tag_destroy(sc->reply_dmat);
735 
736 	if (sc->req_busaddr != 0)
737 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
738 	if (sc->req_frames != NULL)
739 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
740 	if (sc->req_dmat != NULL)
741 		bus_dma_tag_destroy(sc->req_dmat);
742 
743 	if (sc->chains != NULL)
744 		free(sc->chains, M_MPT2);
745 	if (sc->commands != NULL) {
746 		for (i = 1; i < sc->num_reqs; i++) {
747 			cm = &sc->commands[i];
748 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
749 		}
750 		free(sc->commands, M_MPT2);
751 	}
752 	if (sc->buffer_dmat != NULL)
753 		bus_dma_tag_destroy(sc->buffer_dmat);
754 
755 	mps_pci_free_interrupts(sc);
756 	free(sc->queues, M_MPT2);
757 	sc->queues = NULL;
758 }
759 
760 /*
761  * The terms diag reset and hard reset are used interchangeably in the MPI
762  * docs to mean resetting the controller chip.  In this code diag reset
763  * cleans everything up, and the hard reset function just sends the reset
764  * sequence to the chip.  This should probably be refactored so that every
765  * subsystem gets a reset notification of some sort, and can clean up
766  * appropriately.
767  */
768 int
769 mps_reinit(struct mps_softc *sc)
770 {
771 	int error;
772 	struct mpssas_softc *sassc;
773 
774 	sassc = sc->sassc;
775 
776 	MPS_FUNCTRACE(sc);
777 
778 	mtx_assert(&sc->mps_mtx, MA_OWNED);
779 
780 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
781 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
782 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
783 		return 0;
784 	}
785 
786 	/* make sure the completion callbacks can recognize they're getting
787 	 * a NULL cm_reply due to a reset.
788 	 */
789 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
790 
791 	/*
792 	 * Mask interrupts here.
793 	 */
794 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
795 	mps_mask_intr(sc);
796 
797 	error = mps_diag_reset(sc, CAN_SLEEP);
798 	if (error != 0) {
799 		/* XXXSL No need to panic here */
800 		panic("%s hard reset failed with error %d\n",
801 		    __func__, error);
802 	}
803 
804 	/* Restore the PCI state, including the MSI-X registers */
805 	mps_pci_restore(sc);
806 
807 	/* Give the I/O subsystem special priority to get itself prepared */
808 	mpssas_handle_reinit(sc);
809 
810 	/*
811 	 * Get IOC Facts and allocate all structures based on this information.
812 	 * The attach function will also call mps_iocfacts_allocate at startup.
813 	 * If relevant values have changed in IOC Facts, this function will free
814 	 * all of the memory based on IOC Facts and reallocate that memory.
815 	 */
816 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
817 		panic("%s IOC Facts based allocation failed with error %d\n",
818 		    __func__, error);
819 	}
820 
821 	/*
822 	 * Mapping structures will be re-allocated after getting IOC Page8, so
823 	 * free these structures here.
824 	 */
825 	mps_mapping_exit(sc);
826 
827 	/*
828 	 * The static page function currently read is IOC Page8.  Others can be
829 	 * added in future.  It's possible that the values in IOC Page8 have
830 	 * changed after a Diag Reset due to user modification, so always read
831 	 * these.  Interrupts are masked, so unmask them before getting config
832 	 * pages.
833 	 */
834 	mps_unmask_intr(sc);
835 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
836 	mps_base_static_config_pages(sc);
837 
838 	/*
839 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
840 	 * mapping tables.
841 	 */
842 	mps_mapping_initialize(sc);
843 
844 	/*
845 	 * Restart will reload the event masks clobbered by the reset, and
846 	 * then enable the port.
847 	 */
848 	mps_reregister_events(sc);
849 
850 	/* the end of discovery will release the simq, so we're done. */
851 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
852 	    sc, sc->replypostindex, sc->replyfreeindex);
853 
854 	mpssas_release_simq_reinit(sassc);
855 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
856 
857 	return 0;
858 }
859 
860 /* Wait for the chip to ACK a word that we've put into its FIFO
861  * Wait for <timeout> seconds. In single loop wait for busy loop
862  * for 500 microseconds.
863  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
864  * */
865 static int
866 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
867 {
868 
869 	u32 cntdn, count;
870 	u32 int_status;
871 	u32 doorbell;
872 
873 	count = 0;
874 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
875 	do {
876 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
877 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
878 			mps_dprint(sc, MPS_TRACE,
879 			"%s: successful count(%d), timeout(%d)\n",
880 			__func__, count, timeout);
881 		return 0;
882 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
883 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
884 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
885 				MPI2_IOC_STATE_FAULT) {
886 				mps_dprint(sc, MPS_FAULT,
887 					"fault_state(0x%04x)!\n", doorbell);
888 				return (EFAULT);
889 			}
890 		} else if (int_status == 0xFFFFFFFF)
891 			goto out;
892 
893 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
894 		* 0.5 milisecond */
895 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
896 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
897 			"mpsdba", hz/1000);
898 		else if (sleep_flag == CAN_SLEEP)
899 			pause("mpsdba", hz/1000);
900 		else
901 			DELAY(500);
902 		count++;
903 	} while (--cntdn);
904 
905 	out:
906 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
907 		"int_status(%x)!\n", __func__, count, int_status);
908 	return (ETIMEDOUT);
909 
910 }
911 
912 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
913 static int
914 mps_wait_db_int(struct mps_softc *sc)
915 {
916 	int retry;
917 
918 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
919 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
920 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
921 			return (0);
922 		DELAY(2000);
923 	}
924 	return (ETIMEDOUT);
925 }
926 
927 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
928 static int
929 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
930     int req_sz, int reply_sz, int timeout)
931 {
932 	uint32_t *data32;
933 	uint16_t *data16;
934 	int i, count, ioc_sz, residual;
935 	int sleep_flags = CAN_SLEEP;
936 
937 	if (curthread->td_no_sleeping != 0)
938 		sleep_flags = NO_SLEEP;
939 
940 	/* Step 1 */
941 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
942 
943 	/* Step 2 */
944 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
945 		return (EBUSY);
946 
947 	/* Step 3
948 	 * Announce that a message is coming through the doorbell.  Messages
949 	 * are pushed at 32bit words, so round up if needed.
950 	 */
951 	count = (req_sz + 3) / 4;
952 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
953 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
954 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
955 
956 	/* Step 4 */
957 	if (mps_wait_db_int(sc) ||
958 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
959 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
960 		return (ENXIO);
961 	}
962 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
963 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
964 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
965 		return (ENXIO);
966 	}
967 
968 	/* Step 5 */
969 	/* Clock out the message data synchronously in 32-bit dwords*/
970 	data32 = (uint32_t *)req;
971 	for (i = 0; i < count; i++) {
972 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
973 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
974 			mps_dprint(sc, MPS_FAULT,
975 			    "Timeout while writing doorbell\n");
976 			return (ENXIO);
977 		}
978 	}
979 
980 	/* Step 6 */
981 	/* Clock in the reply in 16-bit words.  The total length of the
982 	 * message is always in the 4th byte, so clock out the first 2 words
983 	 * manually, then loop the rest.
984 	 */
985 	data16 = (uint16_t *)reply;
986 	if (mps_wait_db_int(sc) != 0) {
987 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
988 		return (ENXIO);
989 	}
990 	data16[0] =
991 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
992 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
993 	if (mps_wait_db_int(sc) != 0) {
994 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
995 		return (ENXIO);
996 	}
997 	data16[1] =
998 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
999 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1000 
1001 	/* Number of 32bit words in the message */
1002 	ioc_sz = reply->MsgLength;
1003 
1004 	/*
1005 	 * Figure out how many 16bit words to clock in without overrunning.
1006 	 * The precision loss with dividing reply_sz can safely be
1007 	 * ignored because the messages can only be multiples of 32bits.
1008 	 */
1009 	residual = 0;
1010 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1011 	if (count < ioc_sz * 2) {
1012 		residual = ioc_sz * 2 - count;
1013 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1014 		    "residual message words\n", residual);
1015 	}
1016 
1017 	for (i = 2; i < count; i++) {
1018 		if (mps_wait_db_int(sc) != 0) {
1019 			mps_dprint(sc, MPS_FAULT,
1020 			    "Timeout reading doorbell %d\n", i);
1021 			return (ENXIO);
1022 		}
1023 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1024 		    MPI2_DOORBELL_DATA_MASK;
1025 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1026 	}
1027 
1028 	/*
1029 	 * Pull out residual words that won't fit into the provided buffer.
1030 	 * This keeps the chip from hanging due to a driver programming
1031 	 * error.
1032 	 */
1033 	while (residual--) {
1034 		if (mps_wait_db_int(sc) != 0) {
1035 			mps_dprint(sc, MPS_FAULT,
1036 			    "Timeout reading doorbell\n");
1037 			return (ENXIO);
1038 		}
1039 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1040 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1041 	}
1042 
1043 	/* Step 7 */
1044 	if (mps_wait_db_int(sc) != 0) {
1045 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1046 		return (ENXIO);
1047 	}
1048 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1049 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1050 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1051 
1052 	return (0);
1053 }
1054 
1055 static void
1056 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1057 {
1058 	reply_descriptor rd;
1059 	MPS_FUNCTRACE(sc);
1060 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1061 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1062 
1063 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1064 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1065 
1066 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1067 		sc->io_cmds_highwater++;
1068 	rd.u.low = cm->cm_desc.Words.Low;
1069 	rd.u.high = cm->cm_desc.Words.High;
1070 	rd.word = htole64(rd.word);
1071 	/* TODO-We may need to make below regwrite atomic */
1072 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1073 	    rd.u.low);
1074 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1075 	    rd.u.high);
1076 }
1077 
1078 /*
1079  * Just the FACTS, ma'am.
1080  */
1081 static int
1082 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1083 {
1084 	MPI2_DEFAULT_REPLY *reply;
1085 	MPI2_IOC_FACTS_REQUEST request;
1086 	int error, req_sz, reply_sz;
1087 
1088 	MPS_FUNCTRACE(sc);
1089 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1090 
1091 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1092 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1093 	reply = (MPI2_DEFAULT_REPLY *)facts;
1094 
1095 	bzero(&request, req_sz);
1096 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1097 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1098 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1099 
1100 	return (error);
1101 }
1102 
1103 static int
1104 mps_send_iocinit(struct mps_softc *sc)
1105 {
1106 	MPI2_IOC_INIT_REQUEST	init;
1107 	MPI2_DEFAULT_REPLY	reply;
1108 	int req_sz, reply_sz, error;
1109 	struct timeval now;
1110 	uint64_t time_in_msec;
1111 
1112 	MPS_FUNCTRACE(sc);
1113 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1114 
1115 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1116 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1117 	bzero(&init, req_sz);
1118 	bzero(&reply, reply_sz);
1119 
1120 	/*
1121 	 * Fill in the init block.  Note that most addresses are
1122 	 * deliberately in the lower 32bits of memory.  This is a micro-
1123 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1124 	 */
1125 	init.Function = MPI2_FUNCTION_IOC_INIT;
1126 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1127 	init.MsgVersion = htole16(MPI2_VERSION);
1128 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1129 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1130 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1131 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1132 	init.SenseBufferAddressHigh = 0;
1133 	init.SystemReplyAddressHigh = 0;
1134 	init.SystemRequestFrameBaseAddress.High = 0;
1135 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1136 	init.ReplyDescriptorPostQueueAddress.High = 0;
1137 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1138 	init.ReplyFreeQueueAddress.High = 0;
1139 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1140 	getmicrotime(&now);
1141 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1142 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1143 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1144 
1145 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1146 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1147 		error = ENXIO;
1148 
1149 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1150 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1151 	return (error);
1152 }
1153 
1154 void
1155 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1156 {
1157 	bus_addr_t *addr;
1158 
1159 	addr = arg;
1160 	*addr = segs[0].ds_addr;
1161 }
1162 
1163 static int
1164 mps_alloc_queues(struct mps_softc *sc)
1165 {
1166 	struct mps_queue *q;
1167 	int nq, i;
1168 
1169 	nq = sc->msi_msgs;
1170 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1171 
1172 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1173 	    M_NOWAIT|M_ZERO);
1174 	if (sc->queues == NULL)
1175 		return (ENOMEM);
1176 
1177 	for (i = 0; i < nq; i++) {
1178 		q = &sc->queues[i];
1179 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1180 		q->sc = sc;
1181 		q->qnum = i;
1182 	}
1183 
1184 	return (0);
1185 }
1186 
1187 static int
1188 mps_alloc_hw_queues(struct mps_softc *sc)
1189 {
1190 	bus_addr_t queues_busaddr;
1191 	uint8_t *queues;
1192 	int qsize, fqsize, pqsize;
1193 
1194 	/*
1195 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1196 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1197 	 * This queue supplies fresh reply frames for the firmware to use.
1198 	 *
1199 	 * The reply descriptor post queue contains 8 byte entries in
1200 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1201 	 * contains filled-in reply frames sent from the firmware to the host.
1202 	 *
1203 	 * These two queues are allocated together for simplicity.
1204 	 */
1205 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1206 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1207 	fqsize= sc->fqdepth * 4;
1208 	pqsize = sc->pqdepth * 8;
1209 	qsize = fqsize + pqsize;
1210 
1211         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1212 				16, 0,			/* algnmnt, boundary */
1213 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1214 				BUS_SPACE_MAXADDR,	/* highaddr */
1215 				NULL, NULL,		/* filter, filterarg */
1216                                 qsize,			/* maxsize */
1217                                 1,			/* nsegments */
1218                                 qsize,			/* maxsegsize */
1219                                 0,			/* flags */
1220                                 NULL, NULL,		/* lockfunc, lockarg */
1221                                 &sc->queues_dmat)) {
1222 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1223 		return (ENOMEM);
1224         }
1225         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1226 	    &sc->queues_map)) {
1227 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1228 		return (ENOMEM);
1229         }
1230         bzero(queues, qsize);
1231         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1232 	    mps_memaddr_cb, &queues_busaddr, 0);
1233 
1234 	sc->free_queue = (uint32_t *)queues;
1235 	sc->free_busaddr = queues_busaddr;
1236 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1237 	sc->post_busaddr = queues_busaddr + fqsize;
1238 
1239 	return (0);
1240 }
1241 
1242 static int
1243 mps_alloc_replies(struct mps_softc *sc)
1244 {
1245 	int rsize, num_replies;
1246 
1247 	/*
1248 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1249 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1250 	 * replies can be used at once.
1251 	 */
1252 	num_replies = max(sc->fqdepth, sc->num_replies);
1253 
1254 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1255         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1256 				4, 0,			/* algnmnt, boundary */
1257 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1258 				BUS_SPACE_MAXADDR,	/* highaddr */
1259 				NULL, NULL,		/* filter, filterarg */
1260                                 rsize,			/* maxsize */
1261                                 1,			/* nsegments */
1262                                 rsize,			/* maxsegsize */
1263                                 0,			/* flags */
1264                                 NULL, NULL,		/* lockfunc, lockarg */
1265                                 &sc->reply_dmat)) {
1266 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1267 		return (ENOMEM);
1268         }
1269         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1270 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1271 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1272 		return (ENOMEM);
1273         }
1274         bzero(sc->reply_frames, rsize);
1275         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1276 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1277 
1278 	return (0);
1279 }
1280 
1281 static int
1282 mps_alloc_requests(struct mps_softc *sc)
1283 {
1284 	struct mps_command *cm;
1285 	struct mps_chain *chain;
1286 	int i, rsize, nsegs;
1287 
1288 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1289         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1290 				16, 0,			/* algnmnt, boundary */
1291 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1292 				BUS_SPACE_MAXADDR,	/* highaddr */
1293 				NULL, NULL,		/* filter, filterarg */
1294                                 rsize,			/* maxsize */
1295                                 1,			/* nsegments */
1296                                 rsize,			/* maxsegsize */
1297                                 0,			/* flags */
1298                                 NULL, NULL,		/* lockfunc, lockarg */
1299                                 &sc->req_dmat)) {
1300 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1301 		return (ENOMEM);
1302         }
1303         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1304 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1305 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1306 		return (ENOMEM);
1307         }
1308         bzero(sc->req_frames, rsize);
1309         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1310 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1311 
1312 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1313         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1314 				16, 0,			/* algnmnt, boundary */
1315 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1316 				BUS_SPACE_MAXADDR,	/* highaddr */
1317 				NULL, NULL,		/* filter, filterarg */
1318                                 rsize,			/* maxsize */
1319                                 1,			/* nsegments */
1320                                 rsize,			/* maxsegsize */
1321                                 0,			/* flags */
1322                                 NULL, NULL,		/* lockfunc, lockarg */
1323                                 &sc->chain_dmat)) {
1324 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1325 		return (ENOMEM);
1326         }
1327         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1328 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1329 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1330 		return (ENOMEM);
1331         }
1332         bzero(sc->chain_frames, rsize);
1333         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1334 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1335 
1336 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1337         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1338 				1, 0,			/* algnmnt, boundary */
1339 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1340 				BUS_SPACE_MAXADDR,	/* highaddr */
1341 				NULL, NULL,		/* filter, filterarg */
1342                                 rsize,			/* maxsize */
1343                                 1,			/* nsegments */
1344                                 rsize,			/* maxsegsize */
1345                                 0,			/* flags */
1346                                 NULL, NULL,		/* lockfunc, lockarg */
1347                                 &sc->sense_dmat)) {
1348 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1349 		return (ENOMEM);
1350         }
1351         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1352 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1353 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1354 		return (ENOMEM);
1355         }
1356         bzero(sc->sense_frames, rsize);
1357         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1358 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1359 
1360 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1361 	    M_WAITOK | M_ZERO);
1362 	if(!sc->chains) {
1363 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1364 		return (ENOMEM);
1365 	}
1366 	for (i = 0; i < sc->max_chains; i++) {
1367 		chain = &sc->chains[i];
1368 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1369 		    i * sc->facts->IOCRequestFrameSize * 4);
1370 		chain->chain_busaddr = sc->chain_busaddr +
1371 		    i * sc->facts->IOCRequestFrameSize * 4;
1372 		mps_free_chain(sc, chain);
1373 		sc->chain_free_lowwater++;
1374 	}
1375 
1376 	/* XXX Need to pick a more precise value */
1377 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1378         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1379 				1, 0,			/* algnmnt, boundary */
1380 				BUS_SPACE_MAXADDR,	/* lowaddr */
1381 				BUS_SPACE_MAXADDR,	/* highaddr */
1382 				NULL, NULL,		/* filter, filterarg */
1383                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1384                                 nsegs,			/* nsegments */
1385                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1386                                 BUS_DMA_ALLOCNOW,	/* flags */
1387                                 busdma_lock_mutex,	/* lockfunc */
1388 				&sc->mps_mtx,		/* lockarg */
1389                                 &sc->buffer_dmat)) {
1390 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1391 		return (ENOMEM);
1392         }
1393 
1394 	/*
1395 	 * SMID 0 cannot be used as a free command per the firmware spec.
1396 	 * Just drop that command instead of risking accounting bugs.
1397 	 */
1398 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1399 	    M_MPT2, M_WAITOK | M_ZERO);
1400 	if(!sc->commands) {
1401 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1402 		return (ENOMEM);
1403 	}
1404 	for (i = 1; i < sc->num_reqs; i++) {
1405 		cm = &sc->commands[i];
1406 		cm->cm_req = sc->req_frames +
1407 		    i * sc->facts->IOCRequestFrameSize * 4;
1408 		cm->cm_req_busaddr = sc->req_busaddr +
1409 		    i * sc->facts->IOCRequestFrameSize * 4;
1410 		cm->cm_sense = &sc->sense_frames[i];
1411 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1412 		cm->cm_desc.Default.SMID = i;
1413 		cm->cm_sc = sc;
1414 		TAILQ_INIT(&cm->cm_chain_list);
1415 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1416 
1417 		/* XXX Is a failure here a critical problem? */
1418 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1419 			if (i <= sc->facts->HighPriorityCredit)
1420 				mps_free_high_priority_command(sc, cm);
1421 			else
1422 				mps_free_command(sc, cm);
1423 		else {
1424 			panic("failed to allocate command %d\n", i);
1425 			sc->num_reqs = i;
1426 			break;
1427 		}
1428 	}
1429 
1430 	return (0);
1431 }
1432 
1433 static int
1434 mps_init_queues(struct mps_softc *sc)
1435 {
1436 	int i;
1437 
1438 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1439 
1440 	/*
1441 	 * According to the spec, we need to use one less reply than we
1442 	 * have space for on the queue.  So sc->num_replies (the number we
1443 	 * use) should be less than sc->fqdepth (allocated size).
1444 	 */
1445 	if (sc->num_replies >= sc->fqdepth)
1446 		return (EINVAL);
1447 
1448 	/*
1449 	 * Initialize all of the free queue entries.
1450 	 */
1451 	for (i = 0; i < sc->fqdepth; i++)
1452 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1453 	sc->replyfreeindex = sc->num_replies;
1454 
1455 	return (0);
1456 }
1457 
1458 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1459  * Next are the global settings, if they exist.  Highest are the per-unit
1460  * settings, if they exist.
1461  */
1462 void
1463 mps_get_tunables(struct mps_softc *sc)
1464 {
1465 	char tmpstr[80], mps_debug[80];
1466 
1467 	/* XXX default to some debugging for now */
1468 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1469 	sc->disable_msix = 0;
1470 	sc->disable_msi = 0;
1471 	sc->max_msix = MPS_MSIX_MAX;
1472 	sc->max_chains = MPS_CHAIN_FRAMES;
1473 	sc->max_io_pages = MPS_MAXIO_PAGES;
1474 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1475 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1476 	sc->use_phynum = 1;
1477 	sc->max_reqframes = MPS_REQ_FRAMES;
1478 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1479 	sc->max_replyframes = MPS_REPLY_FRAMES;
1480 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1481 
1482 	/*
1483 	 * Grab the global variables.
1484 	 */
1485 	bzero(mps_debug, 80);
1486 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1487 		mps_parse_debug(sc, mps_debug);
1488 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1489 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1490 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1491 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1492 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1493 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1494 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1495 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1496 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1497 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1498 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1499 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1500 
1501 	/* Grab the unit-instance variables */
1502 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1503 	    device_get_unit(sc->mps_dev));
1504 	bzero(mps_debug, 80);
1505 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1506 		mps_parse_debug(sc, mps_debug);
1507 
1508 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1509 	    device_get_unit(sc->mps_dev));
1510 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1511 
1512 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1513 	    device_get_unit(sc->mps_dev));
1514 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1515 
1516 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1517 	    device_get_unit(sc->mps_dev));
1518 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1519 
1520 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1521 	    device_get_unit(sc->mps_dev));
1522 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1523 
1524 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1525 	    device_get_unit(sc->mps_dev));
1526 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1527 
1528 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1529 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1530 	    device_get_unit(sc->mps_dev));
1531 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1532 
1533 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1534 	    device_get_unit(sc->mps_dev));
1535 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1536 
1537 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1538 	    device_get_unit(sc->mps_dev));
1539 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1540 
1541 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1542 	    device_get_unit(sc->mps_dev));
1543 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1544 
1545 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1546 	    device_get_unit(sc->mps_dev));
1547 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1548 
1549 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1550 	    device_get_unit(sc->mps_dev));
1551 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1552 
1553 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1554 	    device_get_unit(sc->mps_dev));
1555 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1556 
1557 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1558 	    device_get_unit(sc->mps_dev));
1559 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1560 
1561 }
1562 
1563 static void
1564 mps_setup_sysctl(struct mps_softc *sc)
1565 {
1566 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1567 	struct sysctl_oid	*sysctl_tree = NULL;
1568 	char tmpstr[80], tmpstr2[80];
1569 
1570 	/*
1571 	 * Setup the sysctl variable so the user can change the debug level
1572 	 * on the fly.
1573 	 */
1574 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1575 	    device_get_unit(sc->mps_dev));
1576 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1577 
1578 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1579 	if (sysctl_ctx != NULL)
1580 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1581 
1582 	if (sysctl_tree == NULL) {
1583 		sysctl_ctx_init(&sc->sysctl_ctx);
1584 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1585 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1586 		    CTLFLAG_RD, 0, tmpstr);
1587 		if (sc->sysctl_tree == NULL)
1588 			return;
1589 		sysctl_ctx = &sc->sysctl_ctx;
1590 		sysctl_tree = sc->sysctl_tree;
1591 	}
1592 
1593 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1594 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1595 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1596 
1597 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1598 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1599 	    "Disable the use of MSI-X interrupts");
1600 
1601 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1602 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1603 	    "Disable the use of MSI interrupts");
1604 
1605 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1606 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1607 	    "User-defined maximum number of MSIX queues");
1608 
1609 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1610 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1611 	    "Negotiated number of MSIX queues");
1612 
1613 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1614 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1615 	    "Total number of allocated request frames");
1616 
1617 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1618 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1619 	    "Total number of allocated high priority request frames");
1620 
1621 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1622 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1623 	    "Total number of allocated reply frames");
1624 
1625 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1626 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1627 	    "Total number of event frames allocated");
1628 
1629 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1630 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1631 	    strlen(sc->fw_version), "firmware version");
1632 
1633 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1634 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1635 	    strlen(MPS_DRIVER_VERSION), "driver version");
1636 
1637 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1638 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1639 	    &sc->io_cmds_active, 0, "number of currently active commands");
1640 
1641 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1642 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1643 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1644 
1645 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1646 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1647 	    &sc->chain_free, 0, "number of free chain elements");
1648 
1649 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1650 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1651 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1652 
1653 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1654 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1655 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1656 
1657 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1658 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1659 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1660 	    "IOCFacts)");
1661 
1662 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1663 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1664 	    "enable SSU to SATA SSD/HDD at shutdown");
1665 
1666 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1667 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1668 	    &sc->chain_alloc_fail, "chain allocation failures");
1669 
1670 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1671 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1672 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1673 	    "spinup after SATA ID error");
1674 
1675 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1676 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1677 	    mps_mapping_dump, "A", "Mapping Table Dump");
1678 
1679 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1680 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1681 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1682 
1683 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1684 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1685 	    "Use the phy number for enumeration");
1686 }
1687 
1688 static struct mps_debug_string {
1689 	char	*name;
1690 	int	flag;
1691 } mps_debug_strings[] = {
1692 	{"info", MPS_INFO},
1693 	{"fault", MPS_FAULT},
1694 	{"event", MPS_EVENT},
1695 	{"log", MPS_LOG},
1696 	{"recovery", MPS_RECOVERY},
1697 	{"error", MPS_ERROR},
1698 	{"init", MPS_INIT},
1699 	{"xinfo", MPS_XINFO},
1700 	{"user", MPS_USER},
1701 	{"mapping", MPS_MAPPING},
1702 	{"trace", MPS_TRACE}
1703 };
1704 
1705 enum mps_debug_level_combiner {
1706 	COMB_NONE,
1707 	COMB_ADD,
1708 	COMB_SUB
1709 };
1710 
1711 static int
1712 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1713 {
1714 	struct mps_softc *sc;
1715 	struct mps_debug_string *string;
1716 	struct sbuf *sbuf;
1717 	char *buffer;
1718 	size_t sz;
1719 	int i, len, debug, error;
1720 
1721 	sc = (struct mps_softc *)arg1;
1722 
1723 	error = sysctl_wire_old_buffer(req, 0);
1724 	if (error != 0)
1725 		return (error);
1726 
1727 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1728 	debug = sc->mps_debug;
1729 
1730 	sbuf_printf(sbuf, "%#x", debug);
1731 
1732 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1733 	for (i = 0; i < sz; i++) {
1734 		string = &mps_debug_strings[i];
1735 		if (debug & string->flag)
1736 			sbuf_printf(sbuf, ",%s", string->name);
1737 	}
1738 
1739 	error = sbuf_finish(sbuf);
1740 	sbuf_delete(sbuf);
1741 
1742 	if (error || req->newptr == NULL)
1743 		return (error);
1744 
1745 	len = req->newlen - req->newidx;
1746 	if (len == 0)
1747 		return (0);
1748 
1749 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1750 	error = SYSCTL_IN(req, buffer, len);
1751 
1752 	mps_parse_debug(sc, buffer);
1753 
1754 	free(buffer, M_MPT2);
1755 	return (error);
1756 }
1757 
1758 static void
1759 mps_parse_debug(struct mps_softc *sc, char *list)
1760 {
1761 	struct mps_debug_string *string;
1762 	enum mps_debug_level_combiner op;
1763 	char *token, *endtoken;
1764 	size_t sz;
1765 	int flags, i;
1766 
1767 	if (list == NULL || *list == '\0')
1768 		return;
1769 
1770 	if (*list == '+') {
1771 		op = COMB_ADD;
1772 		list++;
1773 	} else if (*list == '-') {
1774 		op = COMB_SUB;
1775 		list++;
1776 	} else
1777 		op = COMB_NONE;
1778 	if (*list == '\0')
1779 		return;
1780 
1781 	flags = 0;
1782 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1783 	while ((token = strsep(&list, ":,")) != NULL) {
1784 
1785 		/* Handle integer flags */
1786 		flags |= strtol(token, &endtoken, 0);
1787 		if (token != endtoken)
1788 			continue;
1789 
1790 		/* Handle text flags */
1791 		for (i = 0; i < sz; i++) {
1792 			string = &mps_debug_strings[i];
1793 			if (strcasecmp(token, string->name) == 0) {
1794 				flags |= string->flag;
1795 				break;
1796 			}
1797 		}
1798 	}
1799 
1800 	switch (op) {
1801 	case COMB_NONE:
1802 		sc->mps_debug = flags;
1803 		break;
1804 	case COMB_ADD:
1805 		sc->mps_debug |= flags;
1806 		break;
1807 	case COMB_SUB:
1808 		sc->mps_debug &= (~flags);
1809 		break;
1810 	}
1811 
1812 	return;
1813 }
1814 
1815 int
1816 mps_attach(struct mps_softc *sc)
1817 {
1818 	int error;
1819 
1820 	MPS_FUNCTRACE(sc);
1821 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1822 
1823 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1824 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1825 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1826 	TAILQ_INIT(&sc->event_list);
1827 	timevalclear(&sc->lastfail);
1828 
1829 	if ((error = mps_transition_ready(sc)) != 0) {
1830 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1831 		    "ready\n");
1832 		return (error);
1833 	}
1834 
1835 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1836 	    M_ZERO|M_NOWAIT);
1837 	if(!sc->facts) {
1838 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1839 		    "exit\n");
1840 		return (ENOMEM);
1841 	}
1842 
1843 	/*
1844 	 * Get IOC Facts and allocate all structures based on this information.
1845 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1846 	 * Facts. If relevant values have changed in IOC Facts, this function
1847 	 * will free all of the memory based on IOC Facts and reallocate that
1848 	 * memory.  If this fails, any allocated memory should already be freed.
1849 	 */
1850 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1851 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1852 		    "failed with error %d, exit\n", error);
1853 		return (error);
1854 	}
1855 
1856 	/* Start the periodic watchdog check on the IOC Doorbell */
1857 	mps_periodic(sc);
1858 
1859 	/*
1860 	 * The portenable will kick off discovery events that will drive the
1861 	 * rest of the initialization process.  The CAM/SAS module will
1862 	 * hold up the boot sequence until discovery is complete.
1863 	 */
1864 	sc->mps_ich.ich_func = mps_startup;
1865 	sc->mps_ich.ich_arg = sc;
1866 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1867 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1868 		    "Cannot establish MPS config hook\n");
1869 		error = EINVAL;
1870 	}
1871 
1872 	/*
1873 	 * Allow IR to shutdown gracefully when shutdown occurs.
1874 	 */
1875 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1876 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1877 
1878 	if (sc->shutdown_eh == NULL)
1879 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1880 		    "shutdown event registration failed\n");
1881 
1882 	mps_setup_sysctl(sc);
1883 
1884 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1885 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1886 
1887 	return (error);
1888 }
1889 
1890 /* Run through any late-start handlers. */
1891 static void
1892 mps_startup(void *arg)
1893 {
1894 	struct mps_softc *sc;
1895 
1896 	sc = (struct mps_softc *)arg;
1897 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1898 
1899 	mps_lock(sc);
1900 	mps_unmask_intr(sc);
1901 
1902 	/* initialize device mapping tables */
1903 	mps_base_static_config_pages(sc);
1904 	mps_mapping_initialize(sc);
1905 	mpssas_startup(sc);
1906 	mps_unlock(sc);
1907 
1908 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
1909 	config_intrhook_disestablish(&sc->mps_ich);
1910 	sc->mps_ich.ich_arg = NULL;
1911 
1912 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1913 }
1914 
1915 /* Periodic watchdog.  Is called with the driver lock already held. */
1916 static void
1917 mps_periodic(void *arg)
1918 {
1919 	struct mps_softc *sc;
1920 	uint32_t db;
1921 
1922 	sc = (struct mps_softc *)arg;
1923 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1924 		return;
1925 
1926 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1927 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1928 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1929 		mps_reinit(sc);
1930 	}
1931 
1932 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1933 }
1934 
1935 static void
1936 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1937     MPI2_EVENT_NOTIFICATION_REPLY *event)
1938 {
1939 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1940 
1941 	MPS_DPRINT_EVENT(sc, generic, event);
1942 
1943 	switch (event->Event) {
1944 	case MPI2_EVENT_LOG_DATA:
1945 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1946 		if (sc->mps_debug & MPS_EVENT)
1947 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1948 		break;
1949 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1950 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1951 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1952 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1953 		     entry->LogSequence);
1954 		break;
1955 	default:
1956 		break;
1957 	}
1958 	return;
1959 }
1960 
1961 static int
1962 mps_attach_log(struct mps_softc *sc)
1963 {
1964 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1965 
1966 	bzero(events, 16);
1967 	setbit(events, MPI2_EVENT_LOG_DATA);
1968 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1969 
1970 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1971 	    &sc->mps_log_eh);
1972 
1973 	return (0);
1974 }
1975 
1976 static int
1977 mps_detach_log(struct mps_softc *sc)
1978 {
1979 
1980 	if (sc->mps_log_eh != NULL)
1981 		mps_deregister_events(sc, sc->mps_log_eh);
1982 	return (0);
1983 }
1984 
1985 /*
1986  * Free all of the driver resources and detach submodules.  Should be called
1987  * without the lock held.
1988  */
1989 int
1990 mps_free(struct mps_softc *sc)
1991 {
1992 	int error;
1993 
1994 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1995 	/* Turn off the watchdog */
1996 	mps_lock(sc);
1997 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1998 	mps_unlock(sc);
1999 	/* Lock must not be held for this */
2000 	callout_drain(&sc->periodic);
2001 	callout_drain(&sc->device_check_callout);
2002 
2003 	if (((error = mps_detach_log(sc)) != 0) ||
2004 	    ((error = mps_detach_sas(sc)) != 0)) {
2005 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2006 		    "subsystems, exit\n");
2007 		return (error);
2008 	}
2009 
2010 	mps_detach_user(sc);
2011 
2012 	/* Put the IOC back in the READY state. */
2013 	mps_lock(sc);
2014 	if ((error = mps_transition_ready(sc)) != 0) {
2015 		mps_unlock(sc);
2016 		return (error);
2017 	}
2018 	mps_unlock(sc);
2019 
2020 	if (sc->facts != NULL)
2021 		free(sc->facts, M_MPT2);
2022 
2023 	/*
2024 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2025 	 * to free these buffers too.
2026 	 */
2027 	mps_iocfacts_free(sc);
2028 
2029 	if (sc->sysctl_tree != NULL)
2030 		sysctl_ctx_free(&sc->sysctl_ctx);
2031 
2032 	/* Deregister the shutdown function */
2033 	if (sc->shutdown_eh != NULL)
2034 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2035 
2036 	mtx_destroy(&sc->mps_mtx);
2037 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2038 
2039 	return (0);
2040 }
2041 
2042 static __inline void
2043 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2044 {
2045 	MPS_FUNCTRACE(sc);
2046 
2047 	if (cm == NULL) {
2048 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2049 		return;
2050 	}
2051 
2052 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2053 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2054 
2055 	if (cm->cm_complete != NULL) {
2056 		mps_dprint(sc, MPS_TRACE,
2057 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2058 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2059 			   cm->cm_reply);
2060 		cm->cm_complete(sc, cm);
2061 	}
2062 
2063 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2064 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2065 		wakeup(cm);
2066 	}
2067 
2068 	if (cm->cm_sc->io_cmds_active != 0) {
2069 		cm->cm_sc->io_cmds_active--;
2070 	} else {
2071 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2072 		    "out of sync - resynching to 0\n");
2073 	}
2074 }
2075 
2076 
2077 static void
2078 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2079 {
2080 	union loginfo_type {
2081 		u32     loginfo;
2082 		struct {
2083 			u32     subcode:16;
2084 			u32     code:8;
2085 			u32     originator:4;
2086 			u32     bus_type:4;
2087 		} dw;
2088 	};
2089 	union loginfo_type sas_loginfo;
2090 	char *originator_str = NULL;
2091 
2092 	sas_loginfo.loginfo = log_info;
2093 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2094 		return;
2095 
2096 	/* each nexus loss loginfo */
2097 	if (log_info == 0x31170000)
2098 		return;
2099 
2100 	/* eat the loginfos associated with task aborts */
2101 	if ((log_info == 30050000 || log_info ==
2102 	    0x31140000 || log_info == 0x31130000))
2103 		return;
2104 
2105 	switch (sas_loginfo.dw.originator) {
2106 	case 0:
2107 		originator_str = "IOP";
2108 		break;
2109 	case 1:
2110 		originator_str = "PL";
2111 		break;
2112 	case 2:
2113 		originator_str = "IR";
2114 		break;
2115 }
2116 
2117 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2118 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2119 	originator_str, sas_loginfo.dw.code,
2120 	sas_loginfo.dw.subcode);
2121 }
2122 
2123 static void
2124 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2125 {
2126 	MPI2DefaultReply_t *mpi_reply;
2127 	u16 sc_status;
2128 
2129 	mpi_reply = (MPI2DefaultReply_t*)reply;
2130 	sc_status = le16toh(mpi_reply->IOCStatus);
2131 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2132 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2133 }
2134 void
2135 mps_intr(void *data)
2136 {
2137 	struct mps_softc *sc;
2138 	uint32_t status;
2139 
2140 	sc = (struct mps_softc *)data;
2141 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2142 
2143 	/*
2144 	 * Check interrupt status register to flush the bus.  This is
2145 	 * needed for both INTx interrupts and driver-driven polling
2146 	 */
2147 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2148 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2149 		return;
2150 
2151 	mps_lock(sc);
2152 	mps_intr_locked(data);
2153 	mps_unlock(sc);
2154 	return;
2155 }
2156 
2157 /*
2158  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2159  * chip.  Hopefully this theory is correct.
2160  */
2161 void
2162 mps_intr_msi(void *data)
2163 {
2164 	struct mps_softc *sc;
2165 
2166 	sc = (struct mps_softc *)data;
2167 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2168 	mps_lock(sc);
2169 	mps_intr_locked(data);
2170 	mps_unlock(sc);
2171 	return;
2172 }
2173 
2174 /*
2175  * The locking is overly broad and simplistic, but easy to deal with for now.
2176  */
2177 void
2178 mps_intr_locked(void *data)
2179 {
2180 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2181 	struct mps_softc *sc;
2182 	struct mps_command *cm = NULL;
2183 	uint8_t flags;
2184 	u_int pq;
2185 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2186 	mps_fw_diagnostic_buffer_t *pBuffer;
2187 
2188 	sc = (struct mps_softc *)data;
2189 
2190 	pq = sc->replypostindex;
2191 	mps_dprint(sc, MPS_TRACE,
2192 	    "%s sc %p starting with replypostindex %u\n",
2193 	    __func__, sc, sc->replypostindex);
2194 
2195 	for ( ;; ) {
2196 		cm = NULL;
2197 		desc = &sc->post_queue[sc->replypostindex];
2198 		flags = desc->Default.ReplyFlags &
2199 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2200 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2201 		 || (le32toh(desc->Words.High) == 0xffffffff))
2202 			break;
2203 
2204 		/* increment the replypostindex now, so that event handlers
2205 		 * and cm completion handlers which decide to do a diag
2206 		 * reset can zero it without it getting incremented again
2207 		 * afterwards, and we break out of this loop on the next
2208 		 * iteration since the reply post queue has been cleared to
2209 		 * 0xFF and all descriptors look unused (which they are).
2210 		 */
2211 		if (++sc->replypostindex >= sc->pqdepth)
2212 			sc->replypostindex = 0;
2213 
2214 		switch (flags) {
2215 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2216 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2217 			cm->cm_reply = NULL;
2218 			break;
2219 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2220 		{
2221 			uint32_t baddr;
2222 			uint8_t *reply;
2223 
2224 			/*
2225 			 * Re-compose the reply address from the address
2226 			 * sent back from the chip.  The ReplyFrameAddress
2227 			 * is the lower 32 bits of the physical address of
2228 			 * particular reply frame.  Convert that address to
2229 			 * host format, and then use that to provide the
2230 			 * offset against the virtual address base
2231 			 * (sc->reply_frames).
2232 			 */
2233 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2234 			reply = sc->reply_frames +
2235 				(baddr - ((uint32_t)sc->reply_busaddr));
2236 			/*
2237 			 * Make sure the reply we got back is in a valid
2238 			 * range.  If not, go ahead and panic here, since
2239 			 * we'll probably panic as soon as we deference the
2240 			 * reply pointer anyway.
2241 			 */
2242 			if ((reply < sc->reply_frames)
2243 			 || (reply > (sc->reply_frames +
2244 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
2245 				printf("%s: WARNING: reply %p out of range!\n",
2246 				       __func__, reply);
2247 				printf("%s: reply_frames %p, fqdepth %d, "
2248 				       "frame size %d\n", __func__,
2249 				       sc->reply_frames, sc->fqdepth,
2250 				       sc->facts->ReplyFrameSize * 4);
2251 				printf("%s: baddr %#x,\n", __func__, baddr);
2252 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
2253 				panic("Reply address out of range");
2254 			}
2255 			if (le16toh(desc->AddressReply.SMID) == 0) {
2256 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2257 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2258 					/*
2259 					 * If SMID is 0 for Diag Buffer Post,
2260 					 * this implies that the reply is due to
2261 					 * a release function with a status that
2262 					 * the buffer has been released.  Set
2263 					 * the buffer flags accordingly.
2264 					 */
2265 					rel_rep =
2266 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2267 					if ((le16toh(rel_rep->IOCStatus) &
2268 					    MPI2_IOCSTATUS_MASK) ==
2269 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2270 					{
2271 						pBuffer =
2272 						    &sc->fw_diag_buffer_list[
2273 						    rel_rep->BufferType];
2274 						pBuffer->valid_data = TRUE;
2275 						pBuffer->owned_by_firmware =
2276 						    FALSE;
2277 						pBuffer->immediate = FALSE;
2278 					}
2279 				} else
2280 					mps_dispatch_event(sc, baddr,
2281 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2282 					    reply);
2283 			} else {
2284 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2285 				cm->cm_reply = reply;
2286 				cm->cm_reply_data =
2287 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2288 			}
2289 			break;
2290 		}
2291 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2292 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2293 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2294 		default:
2295 			/* Unhandled */
2296 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2297 			    desc->Default.ReplyFlags);
2298 			cm = NULL;
2299 			break;
2300 		}
2301 
2302 
2303 		if (cm != NULL) {
2304 			// Print Error reply frame
2305 			if (cm->cm_reply)
2306 				mps_display_reply_info(sc,cm->cm_reply);
2307 			mps_complete_command(sc, cm);
2308 		}
2309 
2310 		desc->Words.Low = 0xffffffff;
2311 		desc->Words.High = 0xffffffff;
2312 	}
2313 
2314 	if (pq != sc->replypostindex) {
2315 		mps_dprint(sc, MPS_TRACE,
2316 		    "%s sc %p writing postindex %d\n",
2317 		    __func__, sc, sc->replypostindex);
2318 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2319 	}
2320 
2321 	return;
2322 }
2323 
2324 static void
2325 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2326     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2327 {
2328 	struct mps_event_handle *eh;
2329 	int event, handled = 0;
2330 
2331 	event = le16toh(reply->Event);
2332 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2333 		if (isset(eh->mask, event)) {
2334 			eh->callback(sc, data, reply);
2335 			handled++;
2336 		}
2337 	}
2338 
2339 	if (handled == 0)
2340 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2341 
2342 	/*
2343 	 * This is the only place that the event/reply should be freed.
2344 	 * Anything wanting to hold onto the event data should have
2345 	 * already copied it into their own storage.
2346 	 */
2347 	mps_free_reply(sc, data);
2348 }
2349 
2350 static void
2351 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2352 {
2353 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2354 
2355 	if (cm->cm_reply)
2356 		MPS_DPRINT_EVENT(sc, generic,
2357 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2358 
2359 	mps_free_command(sc, cm);
2360 
2361 	/* next, send a port enable */
2362 	mpssas_startup(sc);
2363 }
2364 
2365 /*
2366  * For both register_events and update_events, the caller supplies a bitmap
2367  * of events that it _wants_.  These functions then turn that into a bitmask
2368  * suitable for the controller.
2369  */
2370 int
2371 mps_register_events(struct mps_softc *sc, u32 *mask,
2372     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2373 {
2374 	struct mps_event_handle *eh;
2375 	int error = 0;
2376 
2377 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2378 	if(!eh) {
2379 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2380 		return (ENOMEM);
2381 	}
2382 	eh->callback = cb;
2383 	eh->data = data;
2384 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2385 	if (mask != NULL)
2386 		error = mps_update_events(sc, eh, mask);
2387 	*handle = eh;
2388 
2389 	return (error);
2390 }
2391 
2392 int
2393 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2394     u32 *mask)
2395 {
2396 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2397 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2398 	struct mps_command *cm;
2399 	int error, i;
2400 
2401 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2402 
2403 	if ((mask != NULL) && (handle != NULL))
2404 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2405 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2406 
2407 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2408 		sc->event_mask[i] = -1;
2409 
2410 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2411 		sc->event_mask[i] &= ~handle->mask[i];
2412 
2413 
2414 	if ((cm = mps_alloc_command(sc)) == NULL)
2415 		return (EBUSY);
2416 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2417 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2418 	evtreq->MsgFlags = 0;
2419 	evtreq->SASBroadcastPrimitiveMasks = 0;
2420 #ifdef MPS_DEBUG_ALL_EVENTS
2421 	{
2422 		u_char fullmask[16];
2423 		memset(fullmask, 0x00, 16);
2424 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2425 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2426 	}
2427 #else
2428         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2429                 evtreq->EventMasks[i] =
2430                     htole32(sc->event_mask[i]);
2431 #endif
2432 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2433 	cm->cm_data = NULL;
2434 
2435 	error = mps_wait_command(sc, &cm, 60, 0);
2436 	if (cm != NULL)
2437 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2438 	if ((reply == NULL) ||
2439 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2440 		error = ENXIO;
2441 
2442 	if (reply)
2443 		MPS_DPRINT_EVENT(sc, generic, reply);
2444 
2445 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2446 
2447 	if (cm != NULL)
2448 		mps_free_command(sc, cm);
2449 	return (error);
2450 }
2451 
2452 static int
2453 mps_reregister_events(struct mps_softc *sc)
2454 {
2455 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2456 	struct mps_command *cm;
2457 	struct mps_event_handle *eh;
2458 	int error, i;
2459 
2460 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2461 
2462 	/* first, reregister events */
2463 
2464 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2465 		sc->event_mask[i] = -1;
2466 
2467 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2468 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2469 			sc->event_mask[i] &= ~eh->mask[i];
2470 	}
2471 
2472 	if ((cm = mps_alloc_command(sc)) == NULL)
2473 		return (EBUSY);
2474 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2475 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2476 	evtreq->MsgFlags = 0;
2477 	evtreq->SASBroadcastPrimitiveMasks = 0;
2478 #ifdef MPS_DEBUG_ALL_EVENTS
2479 	{
2480 		u_char fullmask[16];
2481 		memset(fullmask, 0x00, 16);
2482 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2483 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2484 	}
2485 #else
2486         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2487                 evtreq->EventMasks[i] =
2488                     htole32(sc->event_mask[i]);
2489 #endif
2490 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2491 	cm->cm_data = NULL;
2492 	cm->cm_complete = mps_reregister_events_complete;
2493 
2494 	error = mps_map_command(sc, cm);
2495 
2496 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2497 	    error);
2498 	return (error);
2499 }
2500 
2501 void
2502 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2503 {
2504 
2505 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2506 	free(handle, M_MPT2);
2507 }
2508 
2509 /*
2510  * Add a chain element as the next SGE for the specified command.
2511  * Reset cm_sge and cm_sgesize to indicate all the available space.
2512  */
2513 static int
2514 mps_add_chain(struct mps_command *cm)
2515 {
2516 	MPI2_SGE_CHAIN32 *sgc;
2517 	struct mps_chain *chain;
2518 	int space;
2519 
2520 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2521 		panic("MPS: Need SGE Error Code\n");
2522 
2523 	chain = mps_alloc_chain(cm->cm_sc);
2524 	if (chain == NULL)
2525 		return (ENOBUFS);
2526 
2527 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2528 
2529 	/*
2530 	 * Note: a double-linked list is used to make it easier to
2531 	 * walk for debugging.
2532 	 */
2533 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2534 
2535 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2536 	sgc->Length = htole16(space);
2537 	sgc->NextChainOffset = 0;
2538 	/* TODO Looks like bug in Setting sgc->Flags.
2539 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2540 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2541 	 *	This is fine.. because we are not using simple element. In case of
2542 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2543  	 */
2544 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2545 	sgc->Address = htole32(chain->chain_busaddr);
2546 
2547 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2548 	cm->cm_sglsize = space;
2549 	return (0);
2550 }
2551 
2552 /*
2553  * Add one scatter-gather element (chain, simple, transaction context)
2554  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2555  * cm_sge as the remaining size and pointer to the next SGE to fill
2556  * in, respectively.
2557  */
2558 int
2559 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2560 {
2561 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2562 	MPI2_SGE_SIMPLE64 *sge = sgep;
2563 	int error, type;
2564 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2565 
2566 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2567 
2568 #ifdef INVARIANTS
2569 	switch (type) {
2570 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2571 		if (len != tc->DetailsLength + 4)
2572 			panic("TC %p length %u or %zu?", tc,
2573 			    tc->DetailsLength + 4, len);
2574 		}
2575 		break;
2576 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2577 		/* Driver only uses 32-bit chain elements */
2578 		if (len != MPS_SGC_SIZE)
2579 			panic("CHAIN %p length %u or %zu?", sgep,
2580 			    MPS_SGC_SIZE, len);
2581 		break;
2582 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2583 		/* Driver only uses 64-bit SGE simple elements */
2584 		if (len != MPS_SGE64_SIZE)
2585 			panic("SGE simple %p length %u or %zu?", sge,
2586 			    MPS_SGE64_SIZE, len);
2587 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2588 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2589 			panic("SGE simple %p not marked 64-bit?", sge);
2590 
2591 		break;
2592 	default:
2593 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2594 	}
2595 #endif
2596 
2597 	/*
2598 	 * case 1: 1 more segment, enough room for it
2599 	 * case 2: 2 more segments, enough room for both
2600 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2601 	 * case 4: >=1 more segment, enough room for only a chain
2602 	 * case 5: >=1 more segment, no room for anything (error)
2603          */
2604 
2605 	/*
2606 	 * There should be room for at least a chain element, or this
2607 	 * code is buggy.  Case (5).
2608 	 */
2609 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2610 		panic("MPS: Need SGE Error Code\n");
2611 
2612 	if (segsleft >= 2 &&
2613 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2614 		/*
2615 		 * There are 2 or more segments left to add, and only
2616 		 * enough room for 1 and a chain.  Case (3).
2617 		 *
2618 		 * Mark as last element in this chain if necessary.
2619 		 */
2620 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2621 			sge->FlagsLength |= htole32(
2622 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2623 		}
2624 
2625 		/*
2626 		 * Add the item then a chain.  Do the chain now,
2627 		 * rather than on the next iteration, to simplify
2628 		 * understanding the code.
2629 		 */
2630 		cm->cm_sglsize -= len;
2631 		bcopy(sgep, cm->cm_sge, len);
2632 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2633 		return (mps_add_chain(cm));
2634 	}
2635 
2636 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2637 		/*
2638 		 * 1 or more segment, enough room for only a chain.
2639 		 * Hope the previous element wasn't a Simple entry
2640 		 * that needed to be marked with
2641 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2642 		 */
2643 		if ((error = mps_add_chain(cm)) != 0)
2644 			return (error);
2645 	}
2646 
2647 #ifdef INVARIANTS
2648 	/* Case 1: 1 more segment, enough room for it. */
2649 	if (segsleft == 1 && cm->cm_sglsize < len)
2650 		panic("1 seg left and no room? %u versus %zu",
2651 		    cm->cm_sglsize, len);
2652 
2653 	/* Case 2: 2 more segments, enough room for both */
2654 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2655 		panic("2 segs left and no room? %u versus %zu",
2656 		    cm->cm_sglsize, len);
2657 #endif
2658 
2659 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2660 		/*
2661 		 * If this is a bi-directional request, need to account for that
2662 		 * here.  Save the pre-filled sge values.  These will be used
2663 		 * either for the 2nd SGL or for a single direction SGL.  If
2664 		 * cm_out_len is non-zero, this is a bi-directional request, so
2665 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2666 		 * fill in the IN SGL.  Note that at this time, when filling in
2667 		 * 2 SGL's for a bi-directional request, they both use the same
2668 		 * DMA buffer (same cm command).
2669 		 */
2670 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2671 		saved_address_low = sge->Address.Low;
2672 		saved_address_high = sge->Address.High;
2673 		if (cm->cm_out_len) {
2674 			sge->FlagsLength = htole32(cm->cm_out_len |
2675 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2676 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2677 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2678 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2679 			    MPI2_SGE_FLAGS_SHIFT));
2680 			cm->cm_sglsize -= len;
2681 			bcopy(sgep, cm->cm_sge, len);
2682 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2683 			    + len);
2684 		}
2685 		saved_buf_len |=
2686 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2687 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2688 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2689 		    MPI2_SGE_FLAGS_END_OF_LIST |
2690 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2691 		    MPI2_SGE_FLAGS_SHIFT);
2692 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2693 			saved_buf_len |=
2694 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2695 			    MPI2_SGE_FLAGS_SHIFT);
2696 		} else {
2697 			saved_buf_len |=
2698 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2699 			    MPI2_SGE_FLAGS_SHIFT);
2700 		}
2701 		sge->FlagsLength = htole32(saved_buf_len);
2702 		sge->Address.Low = saved_address_low;
2703 		sge->Address.High = saved_address_high;
2704 	}
2705 
2706 	cm->cm_sglsize -= len;
2707 	bcopy(sgep, cm->cm_sge, len);
2708 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2709 	return (0);
2710 }
2711 
2712 /*
2713  * Add one dma segment to the scatter-gather list for a command.
2714  */
2715 int
2716 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2717     int segsleft)
2718 {
2719 	MPI2_SGE_SIMPLE64 sge;
2720 
2721 	/*
2722 	 * This driver always uses 64-bit address elements for simplicity.
2723 	 */
2724 	bzero(&sge, sizeof(sge));
2725 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2726 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2727 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2728 	mps_from_u64(pa, &sge.Address);
2729 
2730 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2731 }
2732 
2733 static void
2734 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2735 {
2736 	struct mps_softc *sc;
2737 	struct mps_command *cm;
2738 	u_int i, dir, sflags;
2739 
2740 	cm = (struct mps_command *)arg;
2741 	sc = cm->cm_sc;
2742 
2743 	/*
2744 	 * In this case, just print out a warning and let the chip tell the
2745 	 * user they did the wrong thing.
2746 	 */
2747 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2748 		mps_dprint(sc, MPS_ERROR,
2749 			   "%s: warning: busdma returned %d segments, "
2750 			   "more than the %d allowed\n", __func__, nsegs,
2751 			   cm->cm_max_segs);
2752 	}
2753 
2754 	/*
2755 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2756 	 * here.  In that case, both direction flags will be set.
2757 	 */
2758 	sflags = 0;
2759 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2760 		/*
2761 		 * We have to add a special case for SMP passthrough, there
2762 		 * is no easy way to generically handle it.  The first
2763 		 * S/G element is used for the command (therefore the
2764 		 * direction bit needs to be set).  The second one is used
2765 		 * for the reply.  We'll leave it to the caller to make
2766 		 * sure we only have two buffers.
2767 		 */
2768 		/*
2769 		 * Even though the busdma man page says it doesn't make
2770 		 * sense to have both direction flags, it does in this case.
2771 		 * We have one s/g element being accessed in each direction.
2772 		 */
2773 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2774 
2775 		/*
2776 		 * Set the direction flag on the first buffer in the SMP
2777 		 * passthrough request.  We'll clear it for the second one.
2778 		 */
2779 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2780 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2781 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2782 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2783 		dir = BUS_DMASYNC_PREWRITE;
2784 	} else
2785 		dir = BUS_DMASYNC_PREREAD;
2786 
2787 	for (i = 0; i < nsegs; i++) {
2788 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2789 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2790 		}
2791 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2792 		    sflags, nsegs - i);
2793 		if (error != 0) {
2794 			/* Resource shortage, roll back! */
2795 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2796 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2797 				    "consider increasing hw.mps.max_chains.\n");
2798 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2799 			mps_complete_command(sc, cm);
2800 			return;
2801 		}
2802 	}
2803 
2804 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2805 	mps_enqueue_request(sc, cm);
2806 
2807 	return;
2808 }
2809 
2810 static void
2811 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2812 	     int error)
2813 {
2814 	mps_data_cb(arg, segs, nsegs, error);
2815 }
2816 
2817 /*
2818  * This is the routine to enqueue commands ansynchronously.
2819  * Note that the only error path here is from bus_dmamap_load(), which can
2820  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2821  * assumed that if you have a command in-hand, then you have enough credits
2822  * to use it.
2823  */
2824 int
2825 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2826 {
2827 	int error = 0;
2828 
2829 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2830 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2831 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2832 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2833 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2834 		    cm->cm_data, mps_data_cb, cm, 0);
2835 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2836 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2837 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2838 	} else {
2839 		/* Add a zero-length element as needed */
2840 		if (cm->cm_sge != NULL)
2841 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2842 		mps_enqueue_request(sc, cm);
2843 	}
2844 
2845 	return (error);
2846 }
2847 
2848 /*
2849  * This is the routine to enqueue commands synchronously.  An error of
2850  * EINPROGRESS from mps_map_command() is ignored since the command will
2851  * be executed and enqueued automatically.  Other errors come from msleep().
2852  */
2853 int
2854 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2855     int sleep_flag)
2856 {
2857 	int error, rc;
2858 	struct timeval cur_time, start_time;
2859 	struct mps_command *cm = *cmp;
2860 
2861 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2862 		return  EBUSY;
2863 
2864 	cm->cm_complete = NULL;
2865 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2866 	error = mps_map_command(sc, cm);
2867 	if ((error != 0) && (error != EINPROGRESS))
2868 		return (error);
2869 
2870 	/*
2871 	 * Check for context and wait for 50 mSec at a time until time has
2872 	 * expired or the command has finished.  If msleep can't be used, need
2873 	 * to poll.
2874 	 */
2875 	if (curthread->td_no_sleeping != 0)
2876 		sleep_flag = NO_SLEEP;
2877 	getmicrouptime(&start_time);
2878 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2879 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2880 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2881 		if (error == EWOULDBLOCK) {
2882 			/*
2883 			 * Record the actual elapsed time in the case of a
2884 			 * timeout for the message below.
2885 			 */
2886 			getmicrouptime(&cur_time);
2887 			timevalsub(&cur_time, &start_time);
2888 		}
2889 	} else {
2890 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2891 			mps_intr_locked(sc);
2892 			if (sleep_flag == CAN_SLEEP)
2893 				pause("mpswait", hz/20);
2894 			else
2895 				DELAY(50000);
2896 
2897 			getmicrouptime(&cur_time);
2898 			timevalsub(&cur_time, &start_time);
2899 			if (cur_time.tv_sec > timeout) {
2900 				error = EWOULDBLOCK;
2901 				break;
2902 			}
2903 		}
2904 	}
2905 
2906 	if (error == EWOULDBLOCK) {
2907 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2908 		    " elapsed=%jd\n", __func__, timeout,
2909 		    (intmax_t)cur_time.tv_sec);
2910 		rc = mps_reinit(sc);
2911 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2912 		    "failed");
2913 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
2914 			/*
2915 			 * Tell the caller that we freed the command in a
2916 			 * reinit.
2917 			 */
2918 			*cmp = NULL;
2919 		}
2920 		error = ETIMEDOUT;
2921 	}
2922 	return (error);
2923 }
2924 
2925 /*
2926  * The MPT driver had a verbose interface for config pages.  In this driver,
2927  * reduce it to much simpler terms, similar to the Linux driver.
2928  */
2929 int
2930 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2931 {
2932 	MPI2_CONFIG_REQUEST *req;
2933 	struct mps_command *cm;
2934 	int error;
2935 
2936 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2937 		return (EBUSY);
2938 	}
2939 
2940 	cm = mps_alloc_command(sc);
2941 	if (cm == NULL) {
2942 		return (EBUSY);
2943 	}
2944 
2945 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2946 	req->Function = MPI2_FUNCTION_CONFIG;
2947 	req->Action = params->action;
2948 	req->SGLFlags = 0;
2949 	req->ChainOffset = 0;
2950 	req->PageAddress = params->page_address;
2951 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2952 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2953 
2954 		hdr = &params->hdr.Ext;
2955 		req->ExtPageType = hdr->ExtPageType;
2956 		req->ExtPageLength = hdr->ExtPageLength;
2957 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2958 		req->Header.PageLength = 0; /* Must be set to zero */
2959 		req->Header.PageNumber = hdr->PageNumber;
2960 		req->Header.PageVersion = hdr->PageVersion;
2961 	} else {
2962 		MPI2_CONFIG_PAGE_HEADER *hdr;
2963 
2964 		hdr = &params->hdr.Struct;
2965 		req->Header.PageType = hdr->PageType;
2966 		req->Header.PageNumber = hdr->PageNumber;
2967 		req->Header.PageLength = hdr->PageLength;
2968 		req->Header.PageVersion = hdr->PageVersion;
2969 	}
2970 
2971 	cm->cm_data = params->buffer;
2972 	cm->cm_length = params->length;
2973 	if (cm->cm_data != NULL) {
2974 		cm->cm_sge = &req->PageBufferSGE;
2975 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2976 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2977 	} else
2978 		cm->cm_sge = NULL;
2979 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2980 
2981 	cm->cm_complete_data = params;
2982 	if (params->callback != NULL) {
2983 		cm->cm_complete = mps_config_complete;
2984 		return (mps_map_command(sc, cm));
2985 	} else {
2986 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
2987 		if (error) {
2988 			mps_dprint(sc, MPS_FAULT,
2989 			    "Error %d reading config page\n", error);
2990 			if (cm != NULL)
2991 				mps_free_command(sc, cm);
2992 			return (error);
2993 		}
2994 		mps_config_complete(sc, cm);
2995 	}
2996 
2997 	return (0);
2998 }
2999 
3000 int
3001 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3002 {
3003 	return (EINVAL);
3004 }
3005 
3006 static void
3007 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3008 {
3009 	MPI2_CONFIG_REPLY *reply;
3010 	struct mps_config_params *params;
3011 
3012 	MPS_FUNCTRACE(sc);
3013 	params = cm->cm_complete_data;
3014 
3015 	if (cm->cm_data != NULL) {
3016 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3017 		    BUS_DMASYNC_POSTREAD);
3018 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3019 	}
3020 
3021 	/*
3022 	 * XXX KDM need to do more error recovery?  This results in the
3023 	 * device in question not getting probed.
3024 	 */
3025 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3026 		params->status = MPI2_IOCSTATUS_BUSY;
3027 		goto done;
3028 	}
3029 
3030 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3031 	if (reply == NULL) {
3032 		params->status = MPI2_IOCSTATUS_BUSY;
3033 		goto done;
3034 	}
3035 	params->status = reply->IOCStatus;
3036 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3037 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3038 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3039 		params->hdr.Ext.PageType = reply->Header.PageType;
3040 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3041 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3042 	} else {
3043 		params->hdr.Struct.PageType = reply->Header.PageType;
3044 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3045 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3046 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3047 	}
3048 
3049 done:
3050 	mps_free_command(sc, cm);
3051 	if (params->callback != NULL)
3052 		params->callback(sc, params);
3053 
3054 	return;
3055 }
3056