xref: /freebsd/sys/dev/mps/mps.c (revision 8a0a413e)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2015 Avago Technologies
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
29  *
30  * $FreeBSD$
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 /* Communications core for Avago Technologies (LSI) MPT2 */
37 
38 /* TODO Move headers to mpsvar */
39 #include <sys/types.h>
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/selinfo.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/module.h>
47 #include <sys/bus.h>
48 #include <sys/conf.h>
49 #include <sys/bio.h>
50 #include <sys/malloc.h>
51 #include <sys/uio.h>
52 #include <sys/sysctl.h>
53 #include <sys/smp.h>
54 #include <sys/queue.h>
55 #include <sys/kthread.h>
56 #include <sys/taskqueue.h>
57 #include <sys/endian.h>
58 #include <sys/eventhandler.h>
59 #include <sys/sbuf.h>
60 
61 #include <machine/bus.h>
62 #include <machine/resource.h>
63 #include <sys/rman.h>
64 #include <sys/proc.h>
65 
66 #include <dev/pci/pcivar.h>
67 
68 #include <cam/cam.h>
69 #include <cam/scsi/scsi_all.h>
70 
71 #include <dev/mps/mpi/mpi2_type.h>
72 #include <dev/mps/mpi/mpi2.h>
73 #include <dev/mps/mpi/mpi2_ioc.h>
74 #include <dev/mps/mpi/mpi2_sas.h>
75 #include <dev/mps/mpi/mpi2_cnfg.h>
76 #include <dev/mps/mpi/mpi2_init.h>
77 #include <dev/mps/mpi/mpi2_tool.h>
78 #include <dev/mps/mps_ioctl.h>
79 #include <dev/mps/mpsvar.h>
80 #include <dev/mps/mps_table.h>
81 
82 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
83 static int mps_init_queues(struct mps_softc *sc);
84 static void mps_resize_queues(struct mps_softc *sc);
85 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
86 static int mps_transition_operational(struct mps_softc *sc);
87 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
88 static void mps_iocfacts_free(struct mps_softc *sc);
89 static void mps_startup(void *arg);
90 static int mps_send_iocinit(struct mps_softc *sc);
91 static int mps_alloc_queues(struct mps_softc *sc);
92 static int mps_alloc_hw_queues(struct mps_softc *sc);
93 static int mps_alloc_replies(struct mps_softc *sc);
94 static int mps_alloc_requests(struct mps_softc *sc);
95 static int mps_attach_log(struct mps_softc *sc);
96 static __inline void mps_complete_command(struct mps_softc *sc,
97     struct mps_command *cm);
98 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
99     MPI2_EVENT_NOTIFICATION_REPLY *reply);
100 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
101 static void mps_periodic(void *);
102 static int mps_reregister_events(struct mps_softc *sc);
103 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
104 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
105 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
106 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
107 static void mps_parse_debug(struct mps_softc *sc, char *list);
108 
109 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
110 
111 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
112 
113 /*
114  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
115  * any state and back to its initialization state machine.
116  */
117 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
118 
119 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
120  * Compiler only support unint64_t to be passed as argument.
121  * Otherwise it will throw below error
122  * "aggregate value used where an integer was expected"
123  */
124 
125 typedef union _reply_descriptor {
126         u64 word;
127         struct {
128                 u32 low;
129                 u32 high;
130         } u;
131 }reply_descriptor,address_descriptor;
132 
133 /* Rate limit chain-fail messages to 1 per minute */
134 static struct timeval mps_chainfail_interval = { 60, 0 };
135 
136 /*
137  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
138  * If this function is called from process context, it can sleep
139  * and there is no harm to sleep, in case if this fuction is called
140  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
141  * based on sleep flags driver will call either msleep, pause or DELAY.
142  * msleep and pause are of same variant, but pause is used when mps_mtx
143  * is not hold by driver.
144  *
145  */
146 static int
147 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
148 {
149 	uint32_t reg;
150 	int i, error, tries = 0;
151 	uint8_t first_wait_done = FALSE;
152 
153 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
154 
155 	/* Clear any pending interrupts */
156 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
157 
158 	/*
159 	 * Force NO_SLEEP for threads prohibited to sleep
160  	 * e.a Thread from interrupt handler are prohibited to sleep.
161  	 */
162 	if (curthread->td_no_sleeping != 0)
163 		sleep_flag = NO_SLEEP;
164 
165 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
166 
167 	/* Push the magic sequence */
168 	error = ETIMEDOUT;
169 	while (tries++ < 20) {
170 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
171 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
172 			    mpt2_reset_magic[i]);
173 		/* wait 100 msec */
174 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
175 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
176 			    "mpsdiag", hz/10);
177 		else if (sleep_flag == CAN_SLEEP)
178 			pause("mpsdiag", hz/10);
179 		else
180 			DELAY(100 * 1000);
181 
182 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
183 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
184 			error = 0;
185 			break;
186 		}
187 	}
188 	if (error) {
189 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
190 		    error);
191 		return (error);
192 	}
193 
194 	/* Send the actual reset.  XXX need to refresh the reg? */
195 	reg |= MPI2_DIAG_RESET_ADAPTER;
196 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
197 		reg);
198 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
199 
200 	/* Wait up to 300 seconds in 50ms intervals */
201 	error = ETIMEDOUT;
202 	for (i = 0; i < 6000; i++) {
203 		/*
204 		 * Wait 50 msec. If this is the first time through, wait 256
205 		 * msec to satisfy Diag Reset timing requirements.
206 		 */
207 		if (first_wait_done) {
208 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
209 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
210 				    "mpsdiag", hz/20);
211 			else if (sleep_flag == CAN_SLEEP)
212 				pause("mpsdiag", hz/20);
213 			else
214 				DELAY(50 * 1000);
215 		} else {
216 			DELAY(256 * 1000);
217 			first_wait_done = TRUE;
218 		}
219 		/*
220 		 * Check for the RESET_ADAPTER bit to be cleared first, then
221 		 * wait for the RESET state to be cleared, which takes a little
222 		 * longer.
223 		 */
224 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
225 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
226 			continue;
227 		}
228 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
229 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
230 			error = 0;
231 			break;
232 		}
233 	}
234 	if (error) {
235 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
236 		    error);
237 		return (error);
238 	}
239 
240 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
241 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
242 
243 	return (0);
244 }
245 
246 static int
247 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
248 {
249 	int error;
250 
251 	MPS_FUNCTRACE(sc);
252 
253 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
254 
255 	error = 0;
256 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
257 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
258 	    MPI2_DOORBELL_FUNCTION_SHIFT);
259 
260 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
261 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
262 		    "Doorbell handshake failed\n");
263 		error = ETIMEDOUT;
264 	}
265 
266 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
267 	return (error);
268 }
269 
270 static int
271 mps_transition_ready(struct mps_softc *sc)
272 {
273 	uint32_t reg, state;
274 	int error, tries = 0;
275 	int sleep_flags;
276 
277 	MPS_FUNCTRACE(sc);
278 	/* If we are in attach call, do not sleep */
279 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
280 					? CAN_SLEEP:NO_SLEEP;
281 	error = 0;
282 
283 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
284 	   __func__, sleep_flags);
285 
286 	while (tries++ < 1200) {
287 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
288 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
289 
290 		/*
291 		 * Ensure the IOC is ready to talk.  If it's not, try
292 		 * resetting it.
293 		 */
294 		if (reg & MPI2_DOORBELL_USED) {
295 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
296 			    "reset\n");
297 			mps_diag_reset(sc, sleep_flags);
298 			DELAY(50000);
299 			continue;
300 		}
301 
302 		/* Is the adapter owned by another peer? */
303 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
304 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
305 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
306 			    "control of another peer host, aborting "
307 			    "initialization.\n");
308 			error = ENXIO;
309 			break;
310 		}
311 
312 		state = reg & MPI2_IOC_STATE_MASK;
313 		if (state == MPI2_IOC_STATE_READY) {
314 			/* Ready to go! */
315 			error = 0;
316 			break;
317 		} else if (state == MPI2_IOC_STATE_FAULT) {
318 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
319 			    "state 0x%x, resetting\n",
320 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
321 			mps_diag_reset(sc, sleep_flags);
322 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
323 			/* Need to take ownership */
324 			mps_message_unit_reset(sc, sleep_flags);
325 		} else if (state == MPI2_IOC_STATE_RESET) {
326 			/* Wait a bit, IOC might be in transition */
327 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
328 			    "IOC in unexpected reset state\n");
329 		} else {
330 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
331 			    "IOC in unknown state 0x%x\n", state);
332 			error = EINVAL;
333 			break;
334 		}
335 
336 		/* Wait 50ms for things to settle down. */
337 		DELAY(50000);
338 	}
339 
340 	if (error)
341 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
342 		    "Cannot transition IOC to ready\n");
343 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
344 
345 	return (error);
346 }
347 
348 static int
349 mps_transition_operational(struct mps_softc *sc)
350 {
351 	uint32_t reg, state;
352 	int error;
353 
354 	MPS_FUNCTRACE(sc);
355 
356 	error = 0;
357 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
358 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
359 
360 	state = reg & MPI2_IOC_STATE_MASK;
361 	if (state != MPI2_IOC_STATE_READY) {
362 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
363 		if ((error = mps_transition_ready(sc)) != 0) {
364 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
365 			    "failed to transition ready, exit\n");
366 			return (error);
367 		}
368 	}
369 
370 	error = mps_send_iocinit(sc);
371 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
372 
373 	return (error);
374 }
375 
376 static void
377 mps_resize_queues(struct mps_softc *sc)
378 {
379 	int reqcr, prireqcr;
380 
381 	/*
382 	 * Size the queues. Since the reply queues always need one free
383 	 * entry, we'll deduct one reply message here.  The LSI documents
384 	 * suggest instead to add a count to the request queue, but I think
385 	 * that it's better to deduct from reply queue.
386 	 */
387 	prireqcr = MAX(1, sc->max_prireqframes);
388 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
389 
390 	reqcr = MAX(2, sc->max_reqframes);
391 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
392 
393 	sc->num_reqs = prireqcr + reqcr;
394 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
395 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
396 
397 	/*
398 	 * Figure out the number of MSIx-based queues.  If the firmware or
399 	 * user has done something crazy and not allowed enough credit for
400 	 * the queues to be useful then don't enable multi-queue.
401 	 */
402 	if (sc->facts->MaxMSIxVectors < 2)
403 		sc->msi_msgs = 1;
404 
405 	if (sc->msi_msgs > 1) {
406 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
407 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
408 		if (sc->num_reqs / sc->msi_msgs < 2)
409 			sc->msi_msgs = 1;
410 	}
411 
412 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
413 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
414 }
415 
416 /*
417  * This is called during attach and when re-initializing due to a Diag Reset.
418  * IOC Facts is used to allocate many of the structures needed by the driver.
419  * If called from attach, de-allocation is not required because the driver has
420  * not allocated any structures yet, but if called from a Diag Reset, previously
421  * allocated structures based on IOC Facts will need to be freed and re-
422  * allocated bases on the latest IOC Facts.
423  */
424 static int
425 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
426 {
427 	int error;
428 	Mpi2IOCFactsReply_t saved_facts;
429 	uint8_t saved_mode, reallocating;
430 
431 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
432 
433 	/* Save old IOC Facts and then only reallocate if Facts have changed */
434 	if (!attaching) {
435 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
436 	}
437 
438 	/*
439 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
440 	 * a re-initialization and only return the error if attaching so the OS
441 	 * can handle it.
442 	 */
443 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
444 		if (attaching) {
445 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
446 			    "IOC Facts with error %d, exit\n", error);
447 			return (error);
448 		} else {
449 			panic("%s failed to get IOC Facts with error %d\n",
450 			    __func__, error);
451 		}
452 	}
453 
454 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
455 
456 	snprintf(sc->fw_version, sizeof(sc->fw_version),
457 	    "%02d.%02d.%02d.%02d",
458 	    sc->facts->FWVersion.Struct.Major,
459 	    sc->facts->FWVersion.Struct.Minor,
460 	    sc->facts->FWVersion.Struct.Unit,
461 	    sc->facts->FWVersion.Struct.Dev);
462 
463 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
464 	    MPS_DRIVER_VERSION);
465 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
466 	     sc->facts->IOCCapabilities,
467 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
468 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
469 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
470 
471 	/*
472 	 * If the chip doesn't support event replay then a hard reset will be
473 	 * required to trigger a full discovery.  Do the reset here then
474 	 * retransition to Ready.  A hard reset might have already been done,
475 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
476 	 * for a Diag Reset.
477 	 */
478 	if (attaching && ((sc->facts->IOCCapabilities &
479 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
480 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
481 		mps_diag_reset(sc, NO_SLEEP);
482 		if ((error = mps_transition_ready(sc)) != 0) {
483 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
484 			    "transition to ready with error %d, exit\n",
485 			    error);
486 			return (error);
487 		}
488 	}
489 
490 	/*
491 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
492 	 * changed from the previous IOC Facts, log a warning, but only if
493 	 * checking this after a Diag Reset and not during attach.
494 	 */
495 	saved_mode = sc->ir_firmware;
496 	if (sc->facts->IOCCapabilities &
497 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
498 		sc->ir_firmware = 1;
499 	if (!attaching) {
500 		if (sc->ir_firmware != saved_mode) {
501 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
502 			    "in IOC Facts does not match previous mode\n");
503 		}
504 	}
505 
506 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
507 	reallocating = FALSE;
508 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
509 
510 	if ((!attaching) &&
511 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
512 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
513 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
514 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
515 	    (saved_facts.ProductID != sc->facts->ProductID) ||
516 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
517 	    (saved_facts.IOCRequestFrameSize !=
518 	    sc->facts->IOCRequestFrameSize) ||
519 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
520 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
521 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
522 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
523 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
524 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
525 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
526 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
527 	    (saved_facts.MaxPersistentEntries !=
528 	    sc->facts->MaxPersistentEntries))) {
529 		reallocating = TRUE;
530 
531 		/* Record that we reallocated everything */
532 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
533 	}
534 
535 	/*
536 	 * Some things should be done if attaching or re-allocating after a Diag
537 	 * Reset, but are not needed after a Diag Reset if the FW has not
538 	 * changed.
539 	 */
540 	if (attaching || reallocating) {
541 		/*
542 		 * Check if controller supports FW diag buffers and set flag to
543 		 * enable each type.
544 		 */
545 		if (sc->facts->IOCCapabilities &
546 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
547 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
548 			    enabled = TRUE;
549 		if (sc->facts->IOCCapabilities &
550 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
551 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
552 			    enabled = TRUE;
553 		if (sc->facts->IOCCapabilities &
554 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
555 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
556 			    enabled = TRUE;
557 
558 		/*
559 		 * Set flag if EEDP is supported and if TLR is supported.
560 		 */
561 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
562 			sc->eedp_enabled = TRUE;
563 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
564 			sc->control_TLR = TRUE;
565 
566 		mps_resize_queues(sc);
567 
568 		/*
569 		 * Initialize all Tail Queues
570 		 */
571 		TAILQ_INIT(&sc->req_list);
572 		TAILQ_INIT(&sc->high_priority_req_list);
573 		TAILQ_INIT(&sc->chain_list);
574 		TAILQ_INIT(&sc->tm_list);
575 	}
576 
577 	/*
578 	 * If doing a Diag Reset and the FW is significantly different
579 	 * (reallocating will be set above in IOC Facts comparison), then all
580 	 * buffers based on the IOC Facts will need to be freed before they are
581 	 * reallocated.
582 	 */
583 	if (reallocating) {
584 		mps_iocfacts_free(sc);
585 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
586 		    saved_facts.MaxVolumes);
587 	}
588 
589 	/*
590 	 * Any deallocation has been completed.  Now start reallocating
591 	 * if needed.  Will only need to reallocate if attaching or if the new
592 	 * IOC Facts are different from the previous IOC Facts after a Diag
593 	 * Reset. Targets have already been allocated above if needed.
594 	 */
595 	error = 0;
596 	while (attaching || reallocating) {
597 		if ((error = mps_alloc_hw_queues(sc)) != 0)
598 			break;
599 		if ((error = mps_alloc_replies(sc)) != 0)
600 			break;
601 		if ((error = mps_alloc_requests(sc)) != 0)
602 			break;
603 		if ((error = mps_alloc_queues(sc)) != 0)
604 			break;
605 
606 		break;
607 	}
608 	if (error) {
609 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
610 		    "Failed to alloc queues with error %d\n", error);
611 		mps_free(sc);
612 		return (error);
613 	}
614 
615 	/* Always initialize the queues */
616 	bzero(sc->free_queue, sc->fqdepth * 4);
617 	mps_init_queues(sc);
618 
619 	/*
620 	 * Always get the chip out of the reset state, but only panic if not
621 	 * attaching.  If attaching and there is an error, that is handled by
622 	 * the OS.
623 	 */
624 	error = mps_transition_operational(sc);
625 	if (error != 0) {
626 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
627 		    "transition to operational with error %d\n", error);
628 		mps_free(sc);
629 		return (error);
630 	}
631 
632 	/*
633 	 * Finish the queue initialization.
634 	 * These are set here instead of in mps_init_queues() because the
635 	 * IOC resets these values during the state transition in
636 	 * mps_transition_operational().  The free index is set to 1
637 	 * because the corresponding index in the IOC is set to 0, and the
638 	 * IOC treats the queues as full if both are set to the same value.
639 	 * Hence the reason that the queue can't hold all of the possible
640 	 * replies.
641 	 */
642 	sc->replypostindex = 0;
643 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
644 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
645 
646 	/*
647 	 * Attach the subsystems so they can prepare their event masks.
648 	 * XXX Should be dynamic so that IM/IR and user modules can attach
649 	 */
650 	error = 0;
651 	while (attaching) {
652 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
653 		if ((error = mps_attach_log(sc)) != 0)
654 			break;
655 		if ((error = mps_attach_sas(sc)) != 0)
656 			break;
657 		if ((error = mps_attach_user(sc)) != 0)
658 			break;
659 		break;
660 	}
661 	if (error) {
662 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
663 		    "subsystems: error %d\n", error);
664 		mps_free(sc);
665 		return (error);
666 	}
667 
668 	/*
669 	 * XXX If the number of MSI-X vectors changes during re-init, this
670 	 * won't see it and adjust.
671 	 */
672 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
673 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
674 		    "interrupts\n");
675 		mps_free(sc);
676 		return (error);
677 	}
678 
679 	/*
680 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
681 	 * reset it after a Diag Reset, just in case.
682 	 */
683 	sc->WD_available = FALSE;
684 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
685 		sc->WD_available = TRUE;
686 
687 	return (error);
688 }
689 
690 /*
691  * This is called if memory is being free (during detach for example) and when
692  * buffers need to be reallocated due to a Diag Reset.
693  */
694 static void
695 mps_iocfacts_free(struct mps_softc *sc)
696 {
697 	struct mps_command *cm;
698 	int i;
699 
700 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
701 
702 	if (sc->free_busaddr != 0)
703 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
704 	if (sc->free_queue != NULL)
705 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
706 		    sc->queues_map);
707 	if (sc->queues_dmat != NULL)
708 		bus_dma_tag_destroy(sc->queues_dmat);
709 
710 	if (sc->chain_busaddr != 0)
711 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
712 	if (sc->chain_frames != NULL)
713 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
714 		    sc->chain_map);
715 	if (sc->chain_dmat != NULL)
716 		bus_dma_tag_destroy(sc->chain_dmat);
717 
718 	if (sc->sense_busaddr != 0)
719 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
720 	if (sc->sense_frames != NULL)
721 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
722 		    sc->sense_map);
723 	if (sc->sense_dmat != NULL)
724 		bus_dma_tag_destroy(sc->sense_dmat);
725 
726 	if (sc->reply_busaddr != 0)
727 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
728 	if (sc->reply_frames != NULL)
729 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
730 		    sc->reply_map);
731 	if (sc->reply_dmat != NULL)
732 		bus_dma_tag_destroy(sc->reply_dmat);
733 
734 	if (sc->req_busaddr != 0)
735 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
736 	if (sc->req_frames != NULL)
737 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
738 	if (sc->req_dmat != NULL)
739 		bus_dma_tag_destroy(sc->req_dmat);
740 
741 	if (sc->chains != NULL)
742 		free(sc->chains, M_MPT2);
743 	if (sc->commands != NULL) {
744 		for (i = 1; i < sc->num_reqs; i++) {
745 			cm = &sc->commands[i];
746 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
747 		}
748 		free(sc->commands, M_MPT2);
749 	}
750 	if (sc->buffer_dmat != NULL)
751 		bus_dma_tag_destroy(sc->buffer_dmat);
752 
753 	mps_pci_free_interrupts(sc);
754 	free(sc->queues, M_MPT2);
755 	sc->queues = NULL;
756 }
757 
758 /*
759  * The terms diag reset and hard reset are used interchangeably in the MPI
760  * docs to mean resetting the controller chip.  In this code diag reset
761  * cleans everything up, and the hard reset function just sends the reset
762  * sequence to the chip.  This should probably be refactored so that every
763  * subsystem gets a reset notification of some sort, and can clean up
764  * appropriately.
765  */
766 int
767 mps_reinit(struct mps_softc *sc)
768 {
769 	int error;
770 	struct mpssas_softc *sassc;
771 
772 	sassc = sc->sassc;
773 
774 	MPS_FUNCTRACE(sc);
775 
776 	mtx_assert(&sc->mps_mtx, MA_OWNED);
777 
778 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
779 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
780 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
781 		return 0;
782 	}
783 
784 	/* make sure the completion callbacks can recognize they're getting
785 	 * a NULL cm_reply due to a reset.
786 	 */
787 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
788 
789 	/*
790 	 * Mask interrupts here.
791 	 */
792 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
793 	mps_mask_intr(sc);
794 
795 	error = mps_diag_reset(sc, CAN_SLEEP);
796 	if (error != 0) {
797 		/* XXXSL No need to panic here */
798 		panic("%s hard reset failed with error %d\n",
799 		    __func__, error);
800 	}
801 
802 	/* Restore the PCI state, including the MSI-X registers */
803 	mps_pci_restore(sc);
804 
805 	/* Give the I/O subsystem special priority to get itself prepared */
806 	mpssas_handle_reinit(sc);
807 
808 	/*
809 	 * Get IOC Facts and allocate all structures based on this information.
810 	 * The attach function will also call mps_iocfacts_allocate at startup.
811 	 * If relevant values have changed in IOC Facts, this function will free
812 	 * all of the memory based on IOC Facts and reallocate that memory.
813 	 */
814 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
815 		panic("%s IOC Facts based allocation failed with error %d\n",
816 		    __func__, error);
817 	}
818 
819 	/*
820 	 * Mapping structures will be re-allocated after getting IOC Page8, so
821 	 * free these structures here.
822 	 */
823 	mps_mapping_exit(sc);
824 
825 	/*
826 	 * The static page function currently read is IOC Page8.  Others can be
827 	 * added in future.  It's possible that the values in IOC Page8 have
828 	 * changed after a Diag Reset due to user modification, so always read
829 	 * these.  Interrupts are masked, so unmask them before getting config
830 	 * pages.
831 	 */
832 	mps_unmask_intr(sc);
833 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
834 	mps_base_static_config_pages(sc);
835 
836 	/*
837 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
838 	 * mapping tables.
839 	 */
840 	mps_mapping_initialize(sc);
841 
842 	/*
843 	 * Restart will reload the event masks clobbered by the reset, and
844 	 * then enable the port.
845 	 */
846 	mps_reregister_events(sc);
847 
848 	/* the end of discovery will release the simq, so we're done. */
849 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
850 	    sc, sc->replypostindex, sc->replyfreeindex);
851 
852 	mpssas_release_simq_reinit(sassc);
853 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
854 
855 	return 0;
856 }
857 
858 /* Wait for the chip to ACK a word that we've put into its FIFO
859  * Wait for <timeout> seconds. In single loop wait for busy loop
860  * for 500 microseconds.
861  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
862  * */
863 static int
864 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
865 {
866 
867 	u32 cntdn, count;
868 	u32 int_status;
869 	u32 doorbell;
870 
871 	count = 0;
872 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
873 	do {
874 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
875 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
876 			mps_dprint(sc, MPS_TRACE,
877 			"%s: successful count(%d), timeout(%d)\n",
878 			__func__, count, timeout);
879 		return 0;
880 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
881 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
882 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
883 				MPI2_IOC_STATE_FAULT) {
884 				mps_dprint(sc, MPS_FAULT,
885 					"fault_state(0x%04x)!\n", doorbell);
886 				return (EFAULT);
887 			}
888 		} else if (int_status == 0xFFFFFFFF)
889 			goto out;
890 
891 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
892 		* 0.5 milisecond */
893 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
894 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
895 			"mpsdba", hz/1000);
896 		else if (sleep_flag == CAN_SLEEP)
897 			pause("mpsdba", hz/1000);
898 		else
899 			DELAY(500);
900 		count++;
901 	} while (--cntdn);
902 
903 	out:
904 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
905 		"int_status(%x)!\n", __func__, count, int_status);
906 	return (ETIMEDOUT);
907 
908 }
909 
910 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
911 static int
912 mps_wait_db_int(struct mps_softc *sc)
913 {
914 	int retry;
915 
916 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
917 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
918 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
919 			return (0);
920 		DELAY(2000);
921 	}
922 	return (ETIMEDOUT);
923 }
924 
925 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
926 static int
927 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
928     int req_sz, int reply_sz, int timeout)
929 {
930 	uint32_t *data32;
931 	uint16_t *data16;
932 	int i, count, ioc_sz, residual;
933 	int sleep_flags = CAN_SLEEP;
934 
935 	if (curthread->td_no_sleeping != 0)
936 		sleep_flags = NO_SLEEP;
937 
938 	/* Step 1 */
939 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
940 
941 	/* Step 2 */
942 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
943 		return (EBUSY);
944 
945 	/* Step 3
946 	 * Announce that a message is coming through the doorbell.  Messages
947 	 * are pushed at 32bit words, so round up if needed.
948 	 */
949 	count = (req_sz + 3) / 4;
950 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
951 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
952 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
953 
954 	/* Step 4 */
955 	if (mps_wait_db_int(sc) ||
956 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
957 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
958 		return (ENXIO);
959 	}
960 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
961 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
962 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
963 		return (ENXIO);
964 	}
965 
966 	/* Step 5 */
967 	/* Clock out the message data synchronously in 32-bit dwords*/
968 	data32 = (uint32_t *)req;
969 	for (i = 0; i < count; i++) {
970 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
971 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
972 			mps_dprint(sc, MPS_FAULT,
973 			    "Timeout while writing doorbell\n");
974 			return (ENXIO);
975 		}
976 	}
977 
978 	/* Step 6 */
979 	/* Clock in the reply in 16-bit words.  The total length of the
980 	 * message is always in the 4th byte, so clock out the first 2 words
981 	 * manually, then loop the rest.
982 	 */
983 	data16 = (uint16_t *)reply;
984 	if (mps_wait_db_int(sc) != 0) {
985 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
986 		return (ENXIO);
987 	}
988 	data16[0] =
989 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
990 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
991 	if (mps_wait_db_int(sc) != 0) {
992 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
993 		return (ENXIO);
994 	}
995 	data16[1] =
996 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
997 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
998 
999 	/* Number of 32bit words in the message */
1000 	ioc_sz = reply->MsgLength;
1001 
1002 	/*
1003 	 * Figure out how many 16bit words to clock in without overrunning.
1004 	 * The precision loss with dividing reply_sz can safely be
1005 	 * ignored because the messages can only be multiples of 32bits.
1006 	 */
1007 	residual = 0;
1008 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1009 	if (count < ioc_sz * 2) {
1010 		residual = ioc_sz * 2 - count;
1011 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1012 		    "residual message words\n", residual);
1013 	}
1014 
1015 	for (i = 2; i < count; i++) {
1016 		if (mps_wait_db_int(sc) != 0) {
1017 			mps_dprint(sc, MPS_FAULT,
1018 			    "Timeout reading doorbell %d\n", i);
1019 			return (ENXIO);
1020 		}
1021 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1022 		    MPI2_DOORBELL_DATA_MASK;
1023 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1024 	}
1025 
1026 	/*
1027 	 * Pull out residual words that won't fit into the provided buffer.
1028 	 * This keeps the chip from hanging due to a driver programming
1029 	 * error.
1030 	 */
1031 	while (residual--) {
1032 		if (mps_wait_db_int(sc) != 0) {
1033 			mps_dprint(sc, MPS_FAULT,
1034 			    "Timeout reading doorbell\n");
1035 			return (ENXIO);
1036 		}
1037 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1038 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1039 	}
1040 
1041 	/* Step 7 */
1042 	if (mps_wait_db_int(sc) != 0) {
1043 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1044 		return (ENXIO);
1045 	}
1046 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1047 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1048 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1049 
1050 	return (0);
1051 }
1052 
1053 static void
1054 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1055 {
1056 	reply_descriptor rd;
1057 	MPS_FUNCTRACE(sc);
1058 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1059 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1060 
1061 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1062 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1063 
1064 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1065 		sc->io_cmds_highwater++;
1066 	rd.u.low = cm->cm_desc.Words.Low;
1067 	rd.u.high = cm->cm_desc.Words.High;
1068 	rd.word = htole64(rd.word);
1069 	/* TODO-We may need to make below regwrite atomic */
1070 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1071 	    rd.u.low);
1072 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1073 	    rd.u.high);
1074 }
1075 
1076 /*
1077  * Just the FACTS, ma'am.
1078  */
1079 static int
1080 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1081 {
1082 	MPI2_DEFAULT_REPLY *reply;
1083 	MPI2_IOC_FACTS_REQUEST request;
1084 	int error, req_sz, reply_sz;
1085 
1086 	MPS_FUNCTRACE(sc);
1087 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1088 
1089 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1090 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1091 	reply = (MPI2_DEFAULT_REPLY *)facts;
1092 
1093 	bzero(&request, req_sz);
1094 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1095 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1096 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1097 
1098 	return (error);
1099 }
1100 
1101 static int
1102 mps_send_iocinit(struct mps_softc *sc)
1103 {
1104 	MPI2_IOC_INIT_REQUEST	init;
1105 	MPI2_DEFAULT_REPLY	reply;
1106 	int req_sz, reply_sz, error;
1107 	struct timeval now;
1108 	uint64_t time_in_msec;
1109 
1110 	MPS_FUNCTRACE(sc);
1111 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1112 
1113 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1114 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1115 	bzero(&init, req_sz);
1116 	bzero(&reply, reply_sz);
1117 
1118 	/*
1119 	 * Fill in the init block.  Note that most addresses are
1120 	 * deliberately in the lower 32bits of memory.  This is a micro-
1121 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1122 	 */
1123 	init.Function = MPI2_FUNCTION_IOC_INIT;
1124 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1125 	init.MsgVersion = htole16(MPI2_VERSION);
1126 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1127 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1128 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1129 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1130 	init.SenseBufferAddressHigh = 0;
1131 	init.SystemReplyAddressHigh = 0;
1132 	init.SystemRequestFrameBaseAddress.High = 0;
1133 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1134 	init.ReplyDescriptorPostQueueAddress.High = 0;
1135 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1136 	init.ReplyFreeQueueAddress.High = 0;
1137 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1138 	getmicrotime(&now);
1139 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1140 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1141 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1142 
1143 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1144 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1145 		error = ENXIO;
1146 
1147 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1148 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1149 	return (error);
1150 }
1151 
1152 void
1153 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1154 {
1155 	bus_addr_t *addr;
1156 
1157 	addr = arg;
1158 	*addr = segs[0].ds_addr;
1159 }
1160 
1161 static int
1162 mps_alloc_queues(struct mps_softc *sc)
1163 {
1164 	struct mps_queue *q;
1165 	int nq, i;
1166 
1167 	nq = sc->msi_msgs;
1168 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1169 
1170 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1171 	    M_NOWAIT|M_ZERO);
1172 	if (sc->queues == NULL)
1173 		return (ENOMEM);
1174 
1175 	for (i = 0; i < nq; i++) {
1176 		q = &sc->queues[i];
1177 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1178 		q->sc = sc;
1179 		q->qnum = i;
1180 	}
1181 
1182 	return (0);
1183 }
1184 
1185 static int
1186 mps_alloc_hw_queues(struct mps_softc *sc)
1187 {
1188 	bus_addr_t queues_busaddr;
1189 	uint8_t *queues;
1190 	int qsize, fqsize, pqsize;
1191 
1192 	/*
1193 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1194 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1195 	 * This queue supplies fresh reply frames for the firmware to use.
1196 	 *
1197 	 * The reply descriptor post queue contains 8 byte entries in
1198 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1199 	 * contains filled-in reply frames sent from the firmware to the host.
1200 	 *
1201 	 * These two queues are allocated together for simplicity.
1202 	 */
1203 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1204 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1205 	fqsize= sc->fqdepth * 4;
1206 	pqsize = sc->pqdepth * 8;
1207 	qsize = fqsize + pqsize;
1208 
1209         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1210 				16, 0,			/* algnmnt, boundary */
1211 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1212 				BUS_SPACE_MAXADDR,	/* highaddr */
1213 				NULL, NULL,		/* filter, filterarg */
1214                                 qsize,			/* maxsize */
1215                                 1,			/* nsegments */
1216                                 qsize,			/* maxsegsize */
1217                                 0,			/* flags */
1218                                 NULL, NULL,		/* lockfunc, lockarg */
1219                                 &sc->queues_dmat)) {
1220 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1221 		return (ENOMEM);
1222         }
1223         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1224 	    &sc->queues_map)) {
1225 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1226 		return (ENOMEM);
1227         }
1228         bzero(queues, qsize);
1229         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1230 	    mps_memaddr_cb, &queues_busaddr, 0);
1231 
1232 	sc->free_queue = (uint32_t *)queues;
1233 	sc->free_busaddr = queues_busaddr;
1234 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1235 	sc->post_busaddr = queues_busaddr + fqsize;
1236 
1237 	return (0);
1238 }
1239 
1240 static int
1241 mps_alloc_replies(struct mps_softc *sc)
1242 {
1243 	int rsize, num_replies;
1244 
1245 	/*
1246 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1247 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1248 	 * replies can be used at once.
1249 	 */
1250 	num_replies = max(sc->fqdepth, sc->num_replies);
1251 
1252 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1253         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1254 				4, 0,			/* algnmnt, boundary */
1255 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1256 				BUS_SPACE_MAXADDR,	/* highaddr */
1257 				NULL, NULL,		/* filter, filterarg */
1258                                 rsize,			/* maxsize */
1259                                 1,			/* nsegments */
1260                                 rsize,			/* maxsegsize */
1261                                 0,			/* flags */
1262                                 NULL, NULL,		/* lockfunc, lockarg */
1263                                 &sc->reply_dmat)) {
1264 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1265 		return (ENOMEM);
1266         }
1267         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1268 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1269 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1270 		return (ENOMEM);
1271         }
1272         bzero(sc->reply_frames, rsize);
1273         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1274 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1275 
1276 	return (0);
1277 }
1278 
1279 static int
1280 mps_alloc_requests(struct mps_softc *sc)
1281 {
1282 	struct mps_command *cm;
1283 	struct mps_chain *chain;
1284 	int i, rsize, nsegs;
1285 
1286 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1287         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1288 				16, 0,			/* algnmnt, boundary */
1289 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1290 				BUS_SPACE_MAXADDR,	/* highaddr */
1291 				NULL, NULL,		/* filter, filterarg */
1292                                 rsize,			/* maxsize */
1293                                 1,			/* nsegments */
1294                                 rsize,			/* maxsegsize */
1295                                 0,			/* flags */
1296                                 NULL, NULL,		/* lockfunc, lockarg */
1297                                 &sc->req_dmat)) {
1298 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1299 		return (ENOMEM);
1300         }
1301         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1302 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1303 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1304 		return (ENOMEM);
1305         }
1306         bzero(sc->req_frames, rsize);
1307         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1308 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1309 
1310 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1311         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1312 				16, 0,			/* algnmnt, boundary */
1313 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1314 				BUS_SPACE_MAXADDR,	/* highaddr */
1315 				NULL, NULL,		/* filter, filterarg */
1316                                 rsize,			/* maxsize */
1317                                 1,			/* nsegments */
1318                                 rsize,			/* maxsegsize */
1319                                 0,			/* flags */
1320                                 NULL, NULL,		/* lockfunc, lockarg */
1321                                 &sc->chain_dmat)) {
1322 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1323 		return (ENOMEM);
1324         }
1325         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1326 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1327 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1328 		return (ENOMEM);
1329         }
1330         bzero(sc->chain_frames, rsize);
1331         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1332 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1333 
1334 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1335         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1336 				1, 0,			/* algnmnt, boundary */
1337 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1338 				BUS_SPACE_MAXADDR,	/* highaddr */
1339 				NULL, NULL,		/* filter, filterarg */
1340                                 rsize,			/* maxsize */
1341                                 1,			/* nsegments */
1342                                 rsize,			/* maxsegsize */
1343                                 0,			/* flags */
1344                                 NULL, NULL,		/* lockfunc, lockarg */
1345                                 &sc->sense_dmat)) {
1346 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1347 		return (ENOMEM);
1348         }
1349         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1350 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1351 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1352 		return (ENOMEM);
1353         }
1354         bzero(sc->sense_frames, rsize);
1355         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1356 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1357 
1358 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1359 	    M_WAITOK | M_ZERO);
1360 	if(!sc->chains) {
1361 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1362 		return (ENOMEM);
1363 	}
1364 	for (i = 0; i < sc->max_chains; i++) {
1365 		chain = &sc->chains[i];
1366 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1367 		    i * sc->facts->IOCRequestFrameSize * 4);
1368 		chain->chain_busaddr = sc->chain_busaddr +
1369 		    i * sc->facts->IOCRequestFrameSize * 4;
1370 		mps_free_chain(sc, chain);
1371 		sc->chain_free_lowwater++;
1372 	}
1373 
1374 	/* XXX Need to pick a more precise value */
1375 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1376         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1377 				1, 0,			/* algnmnt, boundary */
1378 				BUS_SPACE_MAXADDR,	/* lowaddr */
1379 				BUS_SPACE_MAXADDR,	/* highaddr */
1380 				NULL, NULL,		/* filter, filterarg */
1381                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1382                                 nsegs,			/* nsegments */
1383                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1384                                 BUS_DMA_ALLOCNOW,	/* flags */
1385                                 busdma_lock_mutex,	/* lockfunc */
1386 				&sc->mps_mtx,		/* lockarg */
1387                                 &sc->buffer_dmat)) {
1388 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1389 		return (ENOMEM);
1390         }
1391 
1392 	/*
1393 	 * SMID 0 cannot be used as a free command per the firmware spec.
1394 	 * Just drop that command instead of risking accounting bugs.
1395 	 */
1396 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1397 	    M_MPT2, M_WAITOK | M_ZERO);
1398 	if(!sc->commands) {
1399 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1400 		return (ENOMEM);
1401 	}
1402 	for (i = 1; i < sc->num_reqs; i++) {
1403 		cm = &sc->commands[i];
1404 		cm->cm_req = sc->req_frames +
1405 		    i * sc->facts->IOCRequestFrameSize * 4;
1406 		cm->cm_req_busaddr = sc->req_busaddr +
1407 		    i * sc->facts->IOCRequestFrameSize * 4;
1408 		cm->cm_sense = &sc->sense_frames[i];
1409 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1410 		cm->cm_desc.Default.SMID = i;
1411 		cm->cm_sc = sc;
1412 		TAILQ_INIT(&cm->cm_chain_list);
1413 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1414 
1415 		/* XXX Is a failure here a critical problem? */
1416 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1417 			if (i <= sc->facts->HighPriorityCredit)
1418 				mps_free_high_priority_command(sc, cm);
1419 			else
1420 				mps_free_command(sc, cm);
1421 		else {
1422 			panic("failed to allocate command %d\n", i);
1423 			sc->num_reqs = i;
1424 			break;
1425 		}
1426 	}
1427 
1428 	return (0);
1429 }
1430 
1431 static int
1432 mps_init_queues(struct mps_softc *sc)
1433 {
1434 	int i;
1435 
1436 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1437 
1438 	/*
1439 	 * According to the spec, we need to use one less reply than we
1440 	 * have space for on the queue.  So sc->num_replies (the number we
1441 	 * use) should be less than sc->fqdepth (allocated size).
1442 	 */
1443 	if (sc->num_replies >= sc->fqdepth)
1444 		return (EINVAL);
1445 
1446 	/*
1447 	 * Initialize all of the free queue entries.
1448 	 */
1449 	for (i = 0; i < sc->fqdepth; i++)
1450 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1451 	sc->replyfreeindex = sc->num_replies;
1452 
1453 	return (0);
1454 }
1455 
1456 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1457  * Next are the global settings, if they exist.  Highest are the per-unit
1458  * settings, if they exist.
1459  */
1460 void
1461 mps_get_tunables(struct mps_softc *sc)
1462 {
1463 	char tmpstr[80], mps_debug[80];
1464 
1465 	/* XXX default to some debugging for now */
1466 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1467 	sc->disable_msix = 0;
1468 	sc->disable_msi = 0;
1469 	sc->max_msix = MPS_MSIX_MAX;
1470 	sc->max_chains = MPS_CHAIN_FRAMES;
1471 	sc->max_io_pages = MPS_MAXIO_PAGES;
1472 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1473 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1474 	sc->use_phynum = 1;
1475 	sc->max_reqframes = MPS_REQ_FRAMES;
1476 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1477 	sc->max_replyframes = MPS_REPLY_FRAMES;
1478 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1479 
1480 	/*
1481 	 * Grab the global variables.
1482 	 */
1483 	bzero(mps_debug, 80);
1484 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1485 		mps_parse_debug(sc, mps_debug);
1486 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1487 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1488 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1489 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1490 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1491 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1492 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1493 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1494 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1495 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1496 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1497 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1498 
1499 	/* Grab the unit-instance variables */
1500 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1501 	    device_get_unit(sc->mps_dev));
1502 	bzero(mps_debug, 80);
1503 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1504 		mps_parse_debug(sc, mps_debug);
1505 
1506 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1507 	    device_get_unit(sc->mps_dev));
1508 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1509 
1510 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1511 	    device_get_unit(sc->mps_dev));
1512 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1513 
1514 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1515 	    device_get_unit(sc->mps_dev));
1516 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1517 
1518 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1519 	    device_get_unit(sc->mps_dev));
1520 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1521 
1522 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1523 	    device_get_unit(sc->mps_dev));
1524 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1525 
1526 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1527 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1528 	    device_get_unit(sc->mps_dev));
1529 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1530 
1531 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1532 	    device_get_unit(sc->mps_dev));
1533 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1534 
1535 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1536 	    device_get_unit(sc->mps_dev));
1537 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1538 
1539 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1540 	    device_get_unit(sc->mps_dev));
1541 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1542 
1543 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1544 	    device_get_unit(sc->mps_dev));
1545 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1546 
1547 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1548 	    device_get_unit(sc->mps_dev));
1549 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1550 
1551 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1552 	    device_get_unit(sc->mps_dev));
1553 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1554 
1555 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1556 	    device_get_unit(sc->mps_dev));
1557 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1558 
1559 }
1560 
1561 static void
1562 mps_setup_sysctl(struct mps_softc *sc)
1563 {
1564 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1565 	struct sysctl_oid	*sysctl_tree = NULL;
1566 	char tmpstr[80], tmpstr2[80];
1567 
1568 	/*
1569 	 * Setup the sysctl variable so the user can change the debug level
1570 	 * on the fly.
1571 	 */
1572 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1573 	    device_get_unit(sc->mps_dev));
1574 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1575 
1576 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1577 	if (sysctl_ctx != NULL)
1578 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1579 
1580 	if (sysctl_tree == NULL) {
1581 		sysctl_ctx_init(&sc->sysctl_ctx);
1582 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1583 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1584 		    CTLFLAG_RD, 0, tmpstr);
1585 		if (sc->sysctl_tree == NULL)
1586 			return;
1587 		sysctl_ctx = &sc->sysctl_ctx;
1588 		sysctl_tree = sc->sysctl_tree;
1589 	}
1590 
1591 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1592 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1593 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1594 
1595 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1596 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1597 	    "Disable the use of MSI-X interrupts");
1598 
1599 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1600 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1601 	    "Disable the use of MSI interrupts");
1602 
1603 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1604 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1605 	    "User-defined maximum number of MSIX queues");
1606 
1607 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1608 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1609 	    "Negotiated number of MSIX queues");
1610 
1611 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1612 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1613 	    "Total number of allocated request frames");
1614 
1615 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1616 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1617 	    "Total number of allocated high priority request frames");
1618 
1619 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1620 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1621 	    "Total number of allocated reply frames");
1622 
1623 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1624 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1625 	    "Total number of event frames allocated");
1626 
1627 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1628 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1629 	    strlen(sc->fw_version), "firmware version");
1630 
1631 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1632 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1633 	    strlen(MPS_DRIVER_VERSION), "driver version");
1634 
1635 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1636 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1637 	    &sc->io_cmds_active, 0, "number of currently active commands");
1638 
1639 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1640 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1641 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1642 
1643 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1644 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1645 	    &sc->chain_free, 0, "number of free chain elements");
1646 
1647 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1648 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1649 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1650 
1651 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1652 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1653 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1654 
1655 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1656 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1657 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1658 	    "IOCFacts)");
1659 
1660 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1661 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1662 	    "enable SSU to SATA SSD/HDD at shutdown");
1663 
1664 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1665 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1666 	    &sc->chain_alloc_fail, "chain allocation failures");
1667 
1668 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1669 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1670 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1671 	    "spinup after SATA ID error");
1672 
1673 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1674 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1675 	    mps_mapping_dump, "A", "Mapping Table Dump");
1676 
1677 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1678 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1679 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1680 
1681 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1682 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1683 	    "Use the phy number for enumeration");
1684 }
1685 
1686 static struct mps_debug_string {
1687 	char	*name;
1688 	int	flag;
1689 } mps_debug_strings[] = {
1690 	{"info", MPS_INFO},
1691 	{"fault", MPS_FAULT},
1692 	{"event", MPS_EVENT},
1693 	{"log", MPS_LOG},
1694 	{"recovery", MPS_RECOVERY},
1695 	{"error", MPS_ERROR},
1696 	{"init", MPS_INIT},
1697 	{"xinfo", MPS_XINFO},
1698 	{"user", MPS_USER},
1699 	{"mapping", MPS_MAPPING},
1700 	{"trace", MPS_TRACE}
1701 };
1702 
1703 enum mps_debug_level_combiner {
1704 	COMB_NONE,
1705 	COMB_ADD,
1706 	COMB_SUB
1707 };
1708 
1709 static int
1710 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1711 {
1712 	struct mps_softc *sc;
1713 	struct mps_debug_string *string;
1714 	struct sbuf *sbuf;
1715 	char *buffer;
1716 	size_t sz;
1717 	int i, len, debug, error;
1718 
1719 	sc = (struct mps_softc *)arg1;
1720 
1721 	error = sysctl_wire_old_buffer(req, 0);
1722 	if (error != 0)
1723 		return (error);
1724 
1725 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1726 	debug = sc->mps_debug;
1727 
1728 	sbuf_printf(sbuf, "%#x", debug);
1729 
1730 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1731 	for (i = 0; i < sz; i++) {
1732 		string = &mps_debug_strings[i];
1733 		if (debug & string->flag)
1734 			sbuf_printf(sbuf, ",%s", string->name);
1735 	}
1736 
1737 	error = sbuf_finish(sbuf);
1738 	sbuf_delete(sbuf);
1739 
1740 	if (error || req->newptr == NULL)
1741 		return (error);
1742 
1743 	len = req->newlen - req->newidx;
1744 	if (len == 0)
1745 		return (0);
1746 
1747 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1748 	error = SYSCTL_IN(req, buffer, len);
1749 
1750 	mps_parse_debug(sc, buffer);
1751 
1752 	free(buffer, M_MPT2);
1753 	return (error);
1754 }
1755 
1756 static void
1757 mps_parse_debug(struct mps_softc *sc, char *list)
1758 {
1759 	struct mps_debug_string *string;
1760 	enum mps_debug_level_combiner op;
1761 	char *token, *endtoken;
1762 	size_t sz;
1763 	int flags, i;
1764 
1765 	if (list == NULL || *list == '\0')
1766 		return;
1767 
1768 	if (*list == '+') {
1769 		op = COMB_ADD;
1770 		list++;
1771 	} else if (*list == '-') {
1772 		op = COMB_SUB;
1773 		list++;
1774 	} else
1775 		op = COMB_NONE;
1776 	if (*list == '\0')
1777 		return;
1778 
1779 	flags = 0;
1780 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1781 	while ((token = strsep(&list, ":,")) != NULL) {
1782 
1783 		/* Handle integer flags */
1784 		flags |= strtol(token, &endtoken, 0);
1785 		if (token != endtoken)
1786 			continue;
1787 
1788 		/* Handle text flags */
1789 		for (i = 0; i < sz; i++) {
1790 			string = &mps_debug_strings[i];
1791 			if (strcasecmp(token, string->name) == 0) {
1792 				flags |= string->flag;
1793 				break;
1794 			}
1795 		}
1796 	}
1797 
1798 	switch (op) {
1799 	case COMB_NONE:
1800 		sc->mps_debug = flags;
1801 		break;
1802 	case COMB_ADD:
1803 		sc->mps_debug |= flags;
1804 		break;
1805 	case COMB_SUB:
1806 		sc->mps_debug &= (~flags);
1807 		break;
1808 	}
1809 
1810 	return;
1811 }
1812 
1813 int
1814 mps_attach(struct mps_softc *sc)
1815 {
1816 	int error;
1817 
1818 	MPS_FUNCTRACE(sc);
1819 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1820 
1821 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1822 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1823 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1824 	TAILQ_INIT(&sc->event_list);
1825 	timevalclear(&sc->lastfail);
1826 
1827 	if ((error = mps_transition_ready(sc)) != 0) {
1828 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1829 		    "ready\n");
1830 		return (error);
1831 	}
1832 
1833 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1834 	    M_ZERO|M_NOWAIT);
1835 	if(!sc->facts) {
1836 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1837 		    "exit\n");
1838 		return (ENOMEM);
1839 	}
1840 
1841 	/*
1842 	 * Get IOC Facts and allocate all structures based on this information.
1843 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1844 	 * Facts. If relevant values have changed in IOC Facts, this function
1845 	 * will free all of the memory based on IOC Facts and reallocate that
1846 	 * memory.  If this fails, any allocated memory should already be freed.
1847 	 */
1848 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1849 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1850 		    "failed with error %d, exit\n", error);
1851 		return (error);
1852 	}
1853 
1854 	/* Start the periodic watchdog check on the IOC Doorbell */
1855 	mps_periodic(sc);
1856 
1857 	/*
1858 	 * The portenable will kick off discovery events that will drive the
1859 	 * rest of the initialization process.  The CAM/SAS module will
1860 	 * hold up the boot sequence until discovery is complete.
1861 	 */
1862 	sc->mps_ich.ich_func = mps_startup;
1863 	sc->mps_ich.ich_arg = sc;
1864 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1865 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1866 		    "Cannot establish MPS config hook\n");
1867 		error = EINVAL;
1868 	}
1869 
1870 	/*
1871 	 * Allow IR to shutdown gracefully when shutdown occurs.
1872 	 */
1873 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1874 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1875 
1876 	if (sc->shutdown_eh == NULL)
1877 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1878 		    "shutdown event registration failed\n");
1879 
1880 	mps_setup_sysctl(sc);
1881 
1882 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1883 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1884 
1885 	return (error);
1886 }
1887 
1888 /* Run through any late-start handlers. */
1889 static void
1890 mps_startup(void *arg)
1891 {
1892 	struct mps_softc *sc;
1893 
1894 	sc = (struct mps_softc *)arg;
1895 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1896 
1897 	mps_lock(sc);
1898 	mps_unmask_intr(sc);
1899 
1900 	/* initialize device mapping tables */
1901 	mps_base_static_config_pages(sc);
1902 	mps_mapping_initialize(sc);
1903 	mpssas_startup(sc);
1904 	mps_unlock(sc);
1905 
1906 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
1907 	config_intrhook_disestablish(&sc->mps_ich);
1908 	sc->mps_ich.ich_arg = NULL;
1909 
1910 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1911 }
1912 
1913 /* Periodic watchdog.  Is called with the driver lock already held. */
1914 static void
1915 mps_periodic(void *arg)
1916 {
1917 	struct mps_softc *sc;
1918 	uint32_t db;
1919 
1920 	sc = (struct mps_softc *)arg;
1921 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1922 		return;
1923 
1924 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1925 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1926 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1927 		mps_reinit(sc);
1928 	}
1929 
1930 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1931 }
1932 
1933 static void
1934 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1935     MPI2_EVENT_NOTIFICATION_REPLY *event)
1936 {
1937 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1938 
1939 	MPS_DPRINT_EVENT(sc, generic, event);
1940 
1941 	switch (event->Event) {
1942 	case MPI2_EVENT_LOG_DATA:
1943 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1944 		if (sc->mps_debug & MPS_EVENT)
1945 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1946 		break;
1947 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1948 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1949 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1950 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1951 		     entry->LogSequence);
1952 		break;
1953 	default:
1954 		break;
1955 	}
1956 	return;
1957 }
1958 
1959 static int
1960 mps_attach_log(struct mps_softc *sc)
1961 {
1962 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1963 
1964 	bzero(events, 16);
1965 	setbit(events, MPI2_EVENT_LOG_DATA);
1966 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1967 
1968 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1969 	    &sc->mps_log_eh);
1970 
1971 	return (0);
1972 }
1973 
1974 static int
1975 mps_detach_log(struct mps_softc *sc)
1976 {
1977 
1978 	if (sc->mps_log_eh != NULL)
1979 		mps_deregister_events(sc, sc->mps_log_eh);
1980 	return (0);
1981 }
1982 
1983 /*
1984  * Free all of the driver resources and detach submodules.  Should be called
1985  * without the lock held.
1986  */
1987 int
1988 mps_free(struct mps_softc *sc)
1989 {
1990 	int error;
1991 
1992 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1993 	/* Turn off the watchdog */
1994 	mps_lock(sc);
1995 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1996 	mps_unlock(sc);
1997 	/* Lock must not be held for this */
1998 	callout_drain(&sc->periodic);
1999 	callout_drain(&sc->device_check_callout);
2000 
2001 	if (((error = mps_detach_log(sc)) != 0) ||
2002 	    ((error = mps_detach_sas(sc)) != 0)) {
2003 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2004 		    "subsystems, exit\n");
2005 		return (error);
2006 	}
2007 
2008 	mps_detach_user(sc);
2009 
2010 	/* Put the IOC back in the READY state. */
2011 	mps_lock(sc);
2012 	if ((error = mps_transition_ready(sc)) != 0) {
2013 		mps_unlock(sc);
2014 		return (error);
2015 	}
2016 	mps_unlock(sc);
2017 
2018 	if (sc->facts != NULL)
2019 		free(sc->facts, M_MPT2);
2020 
2021 	/*
2022 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2023 	 * to free these buffers too.
2024 	 */
2025 	mps_iocfacts_free(sc);
2026 
2027 	if (sc->sysctl_tree != NULL)
2028 		sysctl_ctx_free(&sc->sysctl_ctx);
2029 
2030 	/* Deregister the shutdown function */
2031 	if (sc->shutdown_eh != NULL)
2032 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2033 
2034 	mtx_destroy(&sc->mps_mtx);
2035 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2036 
2037 	return (0);
2038 }
2039 
2040 static __inline void
2041 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2042 {
2043 	MPS_FUNCTRACE(sc);
2044 
2045 	if (cm == NULL) {
2046 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2047 		return;
2048 	}
2049 
2050 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2051 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2052 
2053 	if (cm->cm_complete != NULL) {
2054 		mps_dprint(sc, MPS_TRACE,
2055 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2056 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2057 			   cm->cm_reply);
2058 		cm->cm_complete(sc, cm);
2059 	}
2060 
2061 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2062 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2063 		wakeup(cm);
2064 	}
2065 
2066 	if (cm->cm_sc->io_cmds_active != 0) {
2067 		cm->cm_sc->io_cmds_active--;
2068 	} else {
2069 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2070 		    "out of sync - resynching to 0\n");
2071 	}
2072 }
2073 
2074 
2075 static void
2076 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2077 {
2078 	union loginfo_type {
2079 		u32     loginfo;
2080 		struct {
2081 			u32     subcode:16;
2082 			u32     code:8;
2083 			u32     originator:4;
2084 			u32     bus_type:4;
2085 		} dw;
2086 	};
2087 	union loginfo_type sas_loginfo;
2088 	char *originator_str = NULL;
2089 
2090 	sas_loginfo.loginfo = log_info;
2091 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2092 		return;
2093 
2094 	/* each nexus loss loginfo */
2095 	if (log_info == 0x31170000)
2096 		return;
2097 
2098 	/* eat the loginfos associated with task aborts */
2099 	if ((log_info == 30050000 || log_info ==
2100 	    0x31140000 || log_info == 0x31130000))
2101 		return;
2102 
2103 	switch (sas_loginfo.dw.originator) {
2104 	case 0:
2105 		originator_str = "IOP";
2106 		break;
2107 	case 1:
2108 		originator_str = "PL";
2109 		break;
2110 	case 2:
2111 		originator_str = "IR";
2112 		break;
2113 }
2114 
2115 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2116 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2117 	originator_str, sas_loginfo.dw.code,
2118 	sas_loginfo.dw.subcode);
2119 }
2120 
2121 static void
2122 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2123 {
2124 	MPI2DefaultReply_t *mpi_reply;
2125 	u16 sc_status;
2126 
2127 	mpi_reply = (MPI2DefaultReply_t*)reply;
2128 	sc_status = le16toh(mpi_reply->IOCStatus);
2129 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2130 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2131 }
2132 void
2133 mps_intr(void *data)
2134 {
2135 	struct mps_softc *sc;
2136 	uint32_t status;
2137 
2138 	sc = (struct mps_softc *)data;
2139 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2140 
2141 	/*
2142 	 * Check interrupt status register to flush the bus.  This is
2143 	 * needed for both INTx interrupts and driver-driven polling
2144 	 */
2145 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2146 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2147 		return;
2148 
2149 	mps_lock(sc);
2150 	mps_intr_locked(data);
2151 	mps_unlock(sc);
2152 	return;
2153 }
2154 
2155 /*
2156  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2157  * chip.  Hopefully this theory is correct.
2158  */
2159 void
2160 mps_intr_msi(void *data)
2161 {
2162 	struct mps_softc *sc;
2163 
2164 	sc = (struct mps_softc *)data;
2165 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2166 	mps_lock(sc);
2167 	mps_intr_locked(data);
2168 	mps_unlock(sc);
2169 	return;
2170 }
2171 
2172 /*
2173  * The locking is overly broad and simplistic, but easy to deal with for now.
2174  */
2175 void
2176 mps_intr_locked(void *data)
2177 {
2178 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2179 	struct mps_softc *sc;
2180 	struct mps_command *cm = NULL;
2181 	uint8_t flags;
2182 	u_int pq;
2183 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2184 	mps_fw_diagnostic_buffer_t *pBuffer;
2185 
2186 	sc = (struct mps_softc *)data;
2187 
2188 	pq = sc->replypostindex;
2189 	mps_dprint(sc, MPS_TRACE,
2190 	    "%s sc %p starting with replypostindex %u\n",
2191 	    __func__, sc, sc->replypostindex);
2192 
2193 	for ( ;; ) {
2194 		cm = NULL;
2195 		desc = &sc->post_queue[sc->replypostindex];
2196 		flags = desc->Default.ReplyFlags &
2197 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2198 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2199 		 || (le32toh(desc->Words.High) == 0xffffffff))
2200 			break;
2201 
2202 		/* increment the replypostindex now, so that event handlers
2203 		 * and cm completion handlers which decide to do a diag
2204 		 * reset can zero it without it getting incremented again
2205 		 * afterwards, and we break out of this loop on the next
2206 		 * iteration since the reply post queue has been cleared to
2207 		 * 0xFF and all descriptors look unused (which they are).
2208 		 */
2209 		if (++sc->replypostindex >= sc->pqdepth)
2210 			sc->replypostindex = 0;
2211 
2212 		switch (flags) {
2213 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2214 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2215 			cm->cm_reply = NULL;
2216 			break;
2217 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2218 		{
2219 			uint32_t baddr;
2220 			uint8_t *reply;
2221 
2222 			/*
2223 			 * Re-compose the reply address from the address
2224 			 * sent back from the chip.  The ReplyFrameAddress
2225 			 * is the lower 32 bits of the physical address of
2226 			 * particular reply frame.  Convert that address to
2227 			 * host format, and then use that to provide the
2228 			 * offset against the virtual address base
2229 			 * (sc->reply_frames).
2230 			 */
2231 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2232 			reply = sc->reply_frames +
2233 				(baddr - ((uint32_t)sc->reply_busaddr));
2234 			/*
2235 			 * Make sure the reply we got back is in a valid
2236 			 * range.  If not, go ahead and panic here, since
2237 			 * we'll probably panic as soon as we deference the
2238 			 * reply pointer anyway.
2239 			 */
2240 			if ((reply < sc->reply_frames)
2241 			 || (reply > (sc->reply_frames +
2242 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
2243 				printf("%s: WARNING: reply %p out of range!\n",
2244 				       __func__, reply);
2245 				printf("%s: reply_frames %p, fqdepth %d, "
2246 				       "frame size %d\n", __func__,
2247 				       sc->reply_frames, sc->fqdepth,
2248 				       sc->facts->ReplyFrameSize * 4);
2249 				printf("%s: baddr %#x,\n", __func__, baddr);
2250 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
2251 				panic("Reply address out of range");
2252 			}
2253 			if (le16toh(desc->AddressReply.SMID) == 0) {
2254 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2255 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2256 					/*
2257 					 * If SMID is 0 for Diag Buffer Post,
2258 					 * this implies that the reply is due to
2259 					 * a release function with a status that
2260 					 * the buffer has been released.  Set
2261 					 * the buffer flags accordingly.
2262 					 */
2263 					rel_rep =
2264 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2265 					if ((le16toh(rel_rep->IOCStatus) &
2266 					    MPI2_IOCSTATUS_MASK) ==
2267 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2268 					{
2269 						pBuffer =
2270 						    &sc->fw_diag_buffer_list[
2271 						    rel_rep->BufferType];
2272 						pBuffer->valid_data = TRUE;
2273 						pBuffer->owned_by_firmware =
2274 						    FALSE;
2275 						pBuffer->immediate = FALSE;
2276 					}
2277 				} else
2278 					mps_dispatch_event(sc, baddr,
2279 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2280 					    reply);
2281 			} else {
2282 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2283 				cm->cm_reply = reply;
2284 				cm->cm_reply_data =
2285 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2286 			}
2287 			break;
2288 		}
2289 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2290 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2291 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2292 		default:
2293 			/* Unhandled */
2294 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2295 			    desc->Default.ReplyFlags);
2296 			cm = NULL;
2297 			break;
2298 		}
2299 
2300 
2301 		if (cm != NULL) {
2302 			// Print Error reply frame
2303 			if (cm->cm_reply)
2304 				mps_display_reply_info(sc,cm->cm_reply);
2305 			mps_complete_command(sc, cm);
2306 		}
2307 
2308 		desc->Words.Low = 0xffffffff;
2309 		desc->Words.High = 0xffffffff;
2310 	}
2311 
2312 	if (pq != sc->replypostindex) {
2313 		mps_dprint(sc, MPS_TRACE,
2314 		    "%s sc %p writing postindex %d\n",
2315 		    __func__, sc, sc->replypostindex);
2316 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2317 	}
2318 
2319 	return;
2320 }
2321 
2322 static void
2323 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2324     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2325 {
2326 	struct mps_event_handle *eh;
2327 	int event, handled = 0;
2328 
2329 	event = le16toh(reply->Event);
2330 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2331 		if (isset(eh->mask, event)) {
2332 			eh->callback(sc, data, reply);
2333 			handled++;
2334 		}
2335 	}
2336 
2337 	if (handled == 0)
2338 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2339 
2340 	/*
2341 	 * This is the only place that the event/reply should be freed.
2342 	 * Anything wanting to hold onto the event data should have
2343 	 * already copied it into their own storage.
2344 	 */
2345 	mps_free_reply(sc, data);
2346 }
2347 
2348 static void
2349 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2350 {
2351 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2352 
2353 	if (cm->cm_reply)
2354 		MPS_DPRINT_EVENT(sc, generic,
2355 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2356 
2357 	mps_free_command(sc, cm);
2358 
2359 	/* next, send a port enable */
2360 	mpssas_startup(sc);
2361 }
2362 
2363 /*
2364  * For both register_events and update_events, the caller supplies a bitmap
2365  * of events that it _wants_.  These functions then turn that into a bitmask
2366  * suitable for the controller.
2367  */
2368 int
2369 mps_register_events(struct mps_softc *sc, u32 *mask,
2370     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2371 {
2372 	struct mps_event_handle *eh;
2373 	int error = 0;
2374 
2375 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2376 	if(!eh) {
2377 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2378 		return (ENOMEM);
2379 	}
2380 	eh->callback = cb;
2381 	eh->data = data;
2382 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2383 	if (mask != NULL)
2384 		error = mps_update_events(sc, eh, mask);
2385 	*handle = eh;
2386 
2387 	return (error);
2388 }
2389 
2390 int
2391 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2392     u32 *mask)
2393 {
2394 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2395 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2396 	struct mps_command *cm;
2397 	int error, i;
2398 
2399 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2400 
2401 	if ((mask != NULL) && (handle != NULL))
2402 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2403 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2404 
2405 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2406 		sc->event_mask[i] = -1;
2407 
2408 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2409 		sc->event_mask[i] &= ~handle->mask[i];
2410 
2411 
2412 	if ((cm = mps_alloc_command(sc)) == NULL)
2413 		return (EBUSY);
2414 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2415 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2416 	evtreq->MsgFlags = 0;
2417 	evtreq->SASBroadcastPrimitiveMasks = 0;
2418 #ifdef MPS_DEBUG_ALL_EVENTS
2419 	{
2420 		u_char fullmask[16];
2421 		memset(fullmask, 0x00, 16);
2422 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2423 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2424 	}
2425 #else
2426         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2427                 evtreq->EventMasks[i] =
2428                     htole32(sc->event_mask[i]);
2429 #endif
2430 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2431 	cm->cm_data = NULL;
2432 
2433 	error = mps_wait_command(sc, &cm, 60, 0);
2434 	if (cm != NULL)
2435 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2436 	if ((reply == NULL) ||
2437 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2438 		error = ENXIO;
2439 
2440 	if (reply)
2441 		MPS_DPRINT_EVENT(sc, generic, reply);
2442 
2443 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2444 
2445 	if (cm != NULL)
2446 		mps_free_command(sc, cm);
2447 	return (error);
2448 }
2449 
2450 static int
2451 mps_reregister_events(struct mps_softc *sc)
2452 {
2453 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2454 	struct mps_command *cm;
2455 	struct mps_event_handle *eh;
2456 	int error, i;
2457 
2458 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2459 
2460 	/* first, reregister events */
2461 
2462 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2463 		sc->event_mask[i] = -1;
2464 
2465 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2466 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2467 			sc->event_mask[i] &= ~eh->mask[i];
2468 	}
2469 
2470 	if ((cm = mps_alloc_command(sc)) == NULL)
2471 		return (EBUSY);
2472 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2473 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2474 	evtreq->MsgFlags = 0;
2475 	evtreq->SASBroadcastPrimitiveMasks = 0;
2476 #ifdef MPS_DEBUG_ALL_EVENTS
2477 	{
2478 		u_char fullmask[16];
2479 		memset(fullmask, 0x00, 16);
2480 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2481 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2482 	}
2483 #else
2484         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2485                 evtreq->EventMasks[i] =
2486                     htole32(sc->event_mask[i]);
2487 #endif
2488 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2489 	cm->cm_data = NULL;
2490 	cm->cm_complete = mps_reregister_events_complete;
2491 
2492 	error = mps_map_command(sc, cm);
2493 
2494 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2495 	    error);
2496 	return (error);
2497 }
2498 
2499 void
2500 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2501 {
2502 
2503 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2504 	free(handle, M_MPT2);
2505 }
2506 
2507 /*
2508  * Add a chain element as the next SGE for the specified command.
2509  * Reset cm_sge and cm_sgesize to indicate all the available space.
2510  */
2511 static int
2512 mps_add_chain(struct mps_command *cm)
2513 {
2514 	MPI2_SGE_CHAIN32 *sgc;
2515 	struct mps_chain *chain;
2516 	int space;
2517 
2518 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2519 		panic("MPS: Need SGE Error Code\n");
2520 
2521 	chain = mps_alloc_chain(cm->cm_sc);
2522 	if (chain == NULL)
2523 		return (ENOBUFS);
2524 
2525 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2526 
2527 	/*
2528 	 * Note: a double-linked list is used to make it easier to
2529 	 * walk for debugging.
2530 	 */
2531 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2532 
2533 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2534 	sgc->Length = htole16(space);
2535 	sgc->NextChainOffset = 0;
2536 	/* TODO Looks like bug in Setting sgc->Flags.
2537 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2538 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2539 	 *	This is fine.. because we are not using simple element. In case of
2540 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2541  	 */
2542 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2543 	sgc->Address = htole32(chain->chain_busaddr);
2544 
2545 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2546 	cm->cm_sglsize = space;
2547 	return (0);
2548 }
2549 
2550 /*
2551  * Add one scatter-gather element (chain, simple, transaction context)
2552  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2553  * cm_sge as the remaining size and pointer to the next SGE to fill
2554  * in, respectively.
2555  */
2556 int
2557 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2558 {
2559 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2560 	MPI2_SGE_SIMPLE64 *sge = sgep;
2561 	int error, type;
2562 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2563 
2564 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2565 
2566 #ifdef INVARIANTS
2567 	switch (type) {
2568 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2569 		if (len != tc->DetailsLength + 4)
2570 			panic("TC %p length %u or %zu?", tc,
2571 			    tc->DetailsLength + 4, len);
2572 		}
2573 		break;
2574 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2575 		/* Driver only uses 32-bit chain elements */
2576 		if (len != MPS_SGC_SIZE)
2577 			panic("CHAIN %p length %u or %zu?", sgep,
2578 			    MPS_SGC_SIZE, len);
2579 		break;
2580 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2581 		/* Driver only uses 64-bit SGE simple elements */
2582 		if (len != MPS_SGE64_SIZE)
2583 			panic("SGE simple %p length %u or %zu?", sge,
2584 			    MPS_SGE64_SIZE, len);
2585 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2586 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2587 			panic("SGE simple %p not marked 64-bit?", sge);
2588 
2589 		break;
2590 	default:
2591 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2592 	}
2593 #endif
2594 
2595 	/*
2596 	 * case 1: 1 more segment, enough room for it
2597 	 * case 2: 2 more segments, enough room for both
2598 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2599 	 * case 4: >=1 more segment, enough room for only a chain
2600 	 * case 5: >=1 more segment, no room for anything (error)
2601          */
2602 
2603 	/*
2604 	 * There should be room for at least a chain element, or this
2605 	 * code is buggy.  Case (5).
2606 	 */
2607 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2608 		panic("MPS: Need SGE Error Code\n");
2609 
2610 	if (segsleft >= 2 &&
2611 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2612 		/*
2613 		 * There are 2 or more segments left to add, and only
2614 		 * enough room for 1 and a chain.  Case (3).
2615 		 *
2616 		 * Mark as last element in this chain if necessary.
2617 		 */
2618 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2619 			sge->FlagsLength |= htole32(
2620 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2621 		}
2622 
2623 		/*
2624 		 * Add the item then a chain.  Do the chain now,
2625 		 * rather than on the next iteration, to simplify
2626 		 * understanding the code.
2627 		 */
2628 		cm->cm_sglsize -= len;
2629 		bcopy(sgep, cm->cm_sge, len);
2630 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2631 		return (mps_add_chain(cm));
2632 	}
2633 
2634 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2635 		/*
2636 		 * 1 or more segment, enough room for only a chain.
2637 		 * Hope the previous element wasn't a Simple entry
2638 		 * that needed to be marked with
2639 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2640 		 */
2641 		if ((error = mps_add_chain(cm)) != 0)
2642 			return (error);
2643 	}
2644 
2645 #ifdef INVARIANTS
2646 	/* Case 1: 1 more segment, enough room for it. */
2647 	if (segsleft == 1 && cm->cm_sglsize < len)
2648 		panic("1 seg left and no room? %u versus %zu",
2649 		    cm->cm_sglsize, len);
2650 
2651 	/* Case 2: 2 more segments, enough room for both */
2652 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2653 		panic("2 segs left and no room? %u versus %zu",
2654 		    cm->cm_sglsize, len);
2655 #endif
2656 
2657 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2658 		/*
2659 		 * If this is a bi-directional request, need to account for that
2660 		 * here.  Save the pre-filled sge values.  These will be used
2661 		 * either for the 2nd SGL or for a single direction SGL.  If
2662 		 * cm_out_len is non-zero, this is a bi-directional request, so
2663 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2664 		 * fill in the IN SGL.  Note that at this time, when filling in
2665 		 * 2 SGL's for a bi-directional request, they both use the same
2666 		 * DMA buffer (same cm command).
2667 		 */
2668 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2669 		saved_address_low = sge->Address.Low;
2670 		saved_address_high = sge->Address.High;
2671 		if (cm->cm_out_len) {
2672 			sge->FlagsLength = htole32(cm->cm_out_len |
2673 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2674 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2675 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2676 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2677 			    MPI2_SGE_FLAGS_SHIFT));
2678 			cm->cm_sglsize -= len;
2679 			bcopy(sgep, cm->cm_sge, len);
2680 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2681 			    + len);
2682 		}
2683 		saved_buf_len |=
2684 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2685 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2686 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2687 		    MPI2_SGE_FLAGS_END_OF_LIST |
2688 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2689 		    MPI2_SGE_FLAGS_SHIFT);
2690 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2691 			saved_buf_len |=
2692 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2693 			    MPI2_SGE_FLAGS_SHIFT);
2694 		} else {
2695 			saved_buf_len |=
2696 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2697 			    MPI2_SGE_FLAGS_SHIFT);
2698 		}
2699 		sge->FlagsLength = htole32(saved_buf_len);
2700 		sge->Address.Low = saved_address_low;
2701 		sge->Address.High = saved_address_high;
2702 	}
2703 
2704 	cm->cm_sglsize -= len;
2705 	bcopy(sgep, cm->cm_sge, len);
2706 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2707 	return (0);
2708 }
2709 
2710 /*
2711  * Add one dma segment to the scatter-gather list for a command.
2712  */
2713 int
2714 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2715     int segsleft)
2716 {
2717 	MPI2_SGE_SIMPLE64 sge;
2718 
2719 	/*
2720 	 * This driver always uses 64-bit address elements for simplicity.
2721 	 */
2722 	bzero(&sge, sizeof(sge));
2723 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2724 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2725 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2726 	mps_from_u64(pa, &sge.Address);
2727 
2728 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2729 }
2730 
2731 static void
2732 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2733 {
2734 	struct mps_softc *sc;
2735 	struct mps_command *cm;
2736 	u_int i, dir, sflags;
2737 
2738 	cm = (struct mps_command *)arg;
2739 	sc = cm->cm_sc;
2740 
2741 	/*
2742 	 * In this case, just print out a warning and let the chip tell the
2743 	 * user they did the wrong thing.
2744 	 */
2745 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2746 		mps_dprint(sc, MPS_ERROR,
2747 			   "%s: warning: busdma returned %d segments, "
2748 			   "more than the %d allowed\n", __func__, nsegs,
2749 			   cm->cm_max_segs);
2750 	}
2751 
2752 	/*
2753 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2754 	 * here.  In that case, both direction flags will be set.
2755 	 */
2756 	sflags = 0;
2757 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2758 		/*
2759 		 * We have to add a special case for SMP passthrough, there
2760 		 * is no easy way to generically handle it.  The first
2761 		 * S/G element is used for the command (therefore the
2762 		 * direction bit needs to be set).  The second one is used
2763 		 * for the reply.  We'll leave it to the caller to make
2764 		 * sure we only have two buffers.
2765 		 */
2766 		/*
2767 		 * Even though the busdma man page says it doesn't make
2768 		 * sense to have both direction flags, it does in this case.
2769 		 * We have one s/g element being accessed in each direction.
2770 		 */
2771 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2772 
2773 		/*
2774 		 * Set the direction flag on the first buffer in the SMP
2775 		 * passthrough request.  We'll clear it for the second one.
2776 		 */
2777 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2778 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2779 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2780 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2781 		dir = BUS_DMASYNC_PREWRITE;
2782 	} else
2783 		dir = BUS_DMASYNC_PREREAD;
2784 
2785 	for (i = 0; i < nsegs; i++) {
2786 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2787 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2788 		}
2789 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2790 		    sflags, nsegs - i);
2791 		if (error != 0) {
2792 			/* Resource shortage, roll back! */
2793 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2794 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2795 				    "consider increasing hw.mps.max_chains.\n");
2796 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2797 			mps_complete_command(sc, cm);
2798 			return;
2799 		}
2800 	}
2801 
2802 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2803 	mps_enqueue_request(sc, cm);
2804 
2805 	return;
2806 }
2807 
2808 static void
2809 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2810 	     int error)
2811 {
2812 	mps_data_cb(arg, segs, nsegs, error);
2813 }
2814 
2815 /*
2816  * This is the routine to enqueue commands ansynchronously.
2817  * Note that the only error path here is from bus_dmamap_load(), which can
2818  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2819  * assumed that if you have a command in-hand, then you have enough credits
2820  * to use it.
2821  */
2822 int
2823 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2824 {
2825 	int error = 0;
2826 
2827 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2828 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2829 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2830 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2831 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2832 		    cm->cm_data, mps_data_cb, cm, 0);
2833 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2834 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2835 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2836 	} else {
2837 		/* Add a zero-length element as needed */
2838 		if (cm->cm_sge != NULL)
2839 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2840 		mps_enqueue_request(sc, cm);
2841 	}
2842 
2843 	return (error);
2844 }
2845 
2846 /*
2847  * This is the routine to enqueue commands synchronously.  An error of
2848  * EINPROGRESS from mps_map_command() is ignored since the command will
2849  * be executed and enqueued automatically.  Other errors come from msleep().
2850  */
2851 int
2852 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2853     int sleep_flag)
2854 {
2855 	int error, rc;
2856 	struct timeval cur_time, start_time;
2857 	struct mps_command *cm = *cmp;
2858 
2859 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2860 		return  EBUSY;
2861 
2862 	cm->cm_complete = NULL;
2863 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2864 	error = mps_map_command(sc, cm);
2865 	if ((error != 0) && (error != EINPROGRESS))
2866 		return (error);
2867 
2868 	/*
2869 	 * Check for context and wait for 50 mSec at a time until time has
2870 	 * expired or the command has finished.  If msleep can't be used, need
2871 	 * to poll.
2872 	 */
2873 	if (curthread->td_no_sleeping != 0)
2874 		sleep_flag = NO_SLEEP;
2875 	getmicrouptime(&start_time);
2876 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2877 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2878 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2879 		if (error == EWOULDBLOCK) {
2880 			/*
2881 			 * Record the actual elapsed time in the case of a
2882 			 * timeout for the message below.
2883 			 */
2884 			getmicrouptime(&cur_time);
2885 			timevalsub(&cur_time, &start_time);
2886 		}
2887 	} else {
2888 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2889 			mps_intr_locked(sc);
2890 			if (sleep_flag == CAN_SLEEP)
2891 				pause("mpswait", hz/20);
2892 			else
2893 				DELAY(50000);
2894 
2895 			getmicrouptime(&cur_time);
2896 			timevalsub(&cur_time, &start_time);
2897 			if (cur_time.tv_sec > timeout) {
2898 				error = EWOULDBLOCK;
2899 				break;
2900 			}
2901 		}
2902 	}
2903 
2904 	if (error == EWOULDBLOCK) {
2905 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2906 		    " elapsed=%jd\n", __func__, timeout,
2907 		    (intmax_t)cur_time.tv_sec);
2908 		rc = mps_reinit(sc);
2909 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2910 		    "failed");
2911 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
2912 			/*
2913 			 * Tell the caller that we freed the command in a
2914 			 * reinit.
2915 			 */
2916 			*cmp = NULL;
2917 		}
2918 		error = ETIMEDOUT;
2919 	}
2920 	return (error);
2921 }
2922 
2923 /*
2924  * The MPT driver had a verbose interface for config pages.  In this driver,
2925  * reduce it to much simpler terms, similar to the Linux driver.
2926  */
2927 int
2928 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2929 {
2930 	MPI2_CONFIG_REQUEST *req;
2931 	struct mps_command *cm;
2932 	int error;
2933 
2934 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2935 		return (EBUSY);
2936 	}
2937 
2938 	cm = mps_alloc_command(sc);
2939 	if (cm == NULL) {
2940 		return (EBUSY);
2941 	}
2942 
2943 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2944 	req->Function = MPI2_FUNCTION_CONFIG;
2945 	req->Action = params->action;
2946 	req->SGLFlags = 0;
2947 	req->ChainOffset = 0;
2948 	req->PageAddress = params->page_address;
2949 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2950 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2951 
2952 		hdr = &params->hdr.Ext;
2953 		req->ExtPageType = hdr->ExtPageType;
2954 		req->ExtPageLength = hdr->ExtPageLength;
2955 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2956 		req->Header.PageLength = 0; /* Must be set to zero */
2957 		req->Header.PageNumber = hdr->PageNumber;
2958 		req->Header.PageVersion = hdr->PageVersion;
2959 	} else {
2960 		MPI2_CONFIG_PAGE_HEADER *hdr;
2961 
2962 		hdr = &params->hdr.Struct;
2963 		req->Header.PageType = hdr->PageType;
2964 		req->Header.PageNumber = hdr->PageNumber;
2965 		req->Header.PageLength = hdr->PageLength;
2966 		req->Header.PageVersion = hdr->PageVersion;
2967 	}
2968 
2969 	cm->cm_data = params->buffer;
2970 	cm->cm_length = params->length;
2971 	if (cm->cm_data != NULL) {
2972 		cm->cm_sge = &req->PageBufferSGE;
2973 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2974 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2975 	} else
2976 		cm->cm_sge = NULL;
2977 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2978 
2979 	cm->cm_complete_data = params;
2980 	if (params->callback != NULL) {
2981 		cm->cm_complete = mps_config_complete;
2982 		return (mps_map_command(sc, cm));
2983 	} else {
2984 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
2985 		if (error) {
2986 			mps_dprint(sc, MPS_FAULT,
2987 			    "Error %d reading config page\n", error);
2988 			if (cm != NULL)
2989 				mps_free_command(sc, cm);
2990 			return (error);
2991 		}
2992 		mps_config_complete(sc, cm);
2993 	}
2994 
2995 	return (0);
2996 }
2997 
2998 int
2999 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3000 {
3001 	return (EINVAL);
3002 }
3003 
3004 static void
3005 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3006 {
3007 	MPI2_CONFIG_REPLY *reply;
3008 	struct mps_config_params *params;
3009 
3010 	MPS_FUNCTRACE(sc);
3011 	params = cm->cm_complete_data;
3012 
3013 	if (cm->cm_data != NULL) {
3014 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3015 		    BUS_DMASYNC_POSTREAD);
3016 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3017 	}
3018 
3019 	/*
3020 	 * XXX KDM need to do more error recovery?  This results in the
3021 	 * device in question not getting probed.
3022 	 */
3023 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3024 		params->status = MPI2_IOCSTATUS_BUSY;
3025 		goto done;
3026 	}
3027 
3028 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3029 	if (reply == NULL) {
3030 		params->status = MPI2_IOCSTATUS_BUSY;
3031 		goto done;
3032 	}
3033 	params->status = reply->IOCStatus;
3034 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3035 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3036 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3037 		params->hdr.Ext.PageType = reply->Header.PageType;
3038 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3039 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3040 	} else {
3041 		params->hdr.Struct.PageType = reply->Header.PageType;
3042 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3043 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3044 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3045 	}
3046 
3047 done:
3048 	mps_free_command(sc, cm);
3049 	if (params->callback != NULL)
3050 		params->callback(sc, params);
3051 
3052 	return;
3053 }
3054