xref: /freebsd/sys/dev/mps/mps.c (revision 40427cca7a9ae77b095936fb1954417c290cfb17)
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 
60 #include <machine/bus.h>
61 #include <machine/resource.h>
62 #include <sys/rman.h>
63 #include <sys/proc.h>
64 
65 #include <dev/pci/pcivar.h>
66 
67 #include <cam/cam.h>
68 #include <cam/scsi/scsi_all.h>
69 
70 #include <dev/mps/mpi/mpi2_type.h>
71 #include <dev/mps/mpi/mpi2.h>
72 #include <dev/mps/mpi/mpi2_ioc.h>
73 #include <dev/mps/mpi/mpi2_sas.h>
74 #include <dev/mps/mpi/mpi2_cnfg.h>
75 #include <dev/mps/mpi/mpi2_init.h>
76 #include <dev/mps/mpi/mpi2_tool.h>
77 #include <dev/mps/mps_ioctl.h>
78 #include <dev/mps/mpsvar.h>
79 #include <dev/mps/mps_table.h>
80 
81 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
82 static int mps_init_queues(struct mps_softc *sc);
83 static void mps_resize_queues(struct mps_softc *sc);
84 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
85 static int mps_transition_operational(struct mps_softc *sc);
86 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
87 static void mps_iocfacts_free(struct mps_softc *sc);
88 static void mps_startup(void *arg);
89 static int mps_send_iocinit(struct mps_softc *sc);
90 static int mps_alloc_queues(struct mps_softc *sc);
91 static int mps_alloc_hw_queues(struct mps_softc *sc);
92 static int mps_alloc_replies(struct mps_softc *sc);
93 static int mps_alloc_requests(struct mps_softc *sc);
94 static int mps_attach_log(struct mps_softc *sc);
95 static __inline void mps_complete_command(struct mps_softc *sc,
96     struct mps_command *cm);
97 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
98     MPI2_EVENT_NOTIFICATION_REPLY *reply);
99 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
100 static void mps_periodic(void *);
101 static int mps_reregister_events(struct mps_softc *sc);
102 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
103 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
104 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
105 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
106 
107 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
108 
109 /*
110  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
111  * any state and back to its initialization state machine.
112  */
113 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
114 
115 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
116  * Compiler only support unint64_t to be passed as argument.
117  * Otherwise it will throw below error
118  * "aggregate value used where an integer was expected"
119  */
120 
121 typedef union _reply_descriptor {
122         u64 word;
123         struct {
124                 u32 low;
125                 u32 high;
126         } u;
127 }reply_descriptor,address_descriptor;
128 
129 /* Rate limit chain-fail messages to 1 per minute */
130 static struct timeval mps_chainfail_interval = { 60, 0 };
131 
132 /*
133  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
134  * If this function is called from process context, it can sleep
135  * and there is no harm to sleep, in case if this fuction is called
136  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
137  * based on sleep flags driver will call either msleep, pause or DELAY.
138  * msleep and pause are of same variant, but pause is used when mps_mtx
139  * is not hold by driver.
140  *
141  */
142 static int
143 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
144 {
145 	uint32_t reg;
146 	int i, error, tries = 0;
147 	uint8_t first_wait_done = FALSE;
148 
149 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
150 
151 	/* Clear any pending interrupts */
152 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
153 
154 	/*
155 	 * Force NO_SLEEP for threads prohibited to sleep
156  	 * e.a Thread from interrupt handler are prohibited to sleep.
157  	 */
158 	if (curthread->td_no_sleeping != 0)
159 		sleep_flag = NO_SLEEP;
160 
161 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
162 
163 	/* Push the magic sequence */
164 	error = ETIMEDOUT;
165 	while (tries++ < 20) {
166 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
167 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
168 			    mpt2_reset_magic[i]);
169 		/* wait 100 msec */
170 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
171 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
172 			    "mpsdiag", hz/10);
173 		else if (sleep_flag == CAN_SLEEP)
174 			pause("mpsdiag", hz/10);
175 		else
176 			DELAY(100 * 1000);
177 
178 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
179 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
180 			error = 0;
181 			break;
182 		}
183 	}
184 	if (error) {
185 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
186 		    error);
187 		return (error);
188 	}
189 
190 	/* Send the actual reset.  XXX need to refresh the reg? */
191 	reg |= MPI2_DIAG_RESET_ADAPTER;
192 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
193 		reg);
194 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
195 
196 	/* Wait up to 300 seconds in 50ms intervals */
197 	error = ETIMEDOUT;
198 	for (i = 0; i < 6000; i++) {
199 		/*
200 		 * Wait 50 msec. If this is the first time through, wait 256
201 		 * msec to satisfy Diag Reset timing requirements.
202 		 */
203 		if (first_wait_done) {
204 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
205 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
206 				    "mpsdiag", hz/20);
207 			else if (sleep_flag == CAN_SLEEP)
208 				pause("mpsdiag", hz/20);
209 			else
210 				DELAY(50 * 1000);
211 		} else {
212 			DELAY(256 * 1000);
213 			first_wait_done = TRUE;
214 		}
215 		/*
216 		 * Check for the RESET_ADAPTER bit to be cleared first, then
217 		 * wait for the RESET state to be cleared, which takes a little
218 		 * longer.
219 		 */
220 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
221 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
222 			continue;
223 		}
224 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
225 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
226 			error = 0;
227 			break;
228 		}
229 	}
230 	if (error) {
231 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
232 		    error);
233 		return (error);
234 	}
235 
236 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
237 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
238 
239 	return (0);
240 }
241 
242 static int
243 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
244 {
245 	int error;
246 
247 	MPS_FUNCTRACE(sc);
248 
249 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
250 
251 	error = 0;
252 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
253 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
254 	    MPI2_DOORBELL_FUNCTION_SHIFT);
255 
256 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
257 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
258 		    "Doorbell handshake failed\n");
259 		error = ETIMEDOUT;
260 	}
261 
262 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
263 	return (error);
264 }
265 
266 static int
267 mps_transition_ready(struct mps_softc *sc)
268 {
269 	uint32_t reg, state;
270 	int error, tries = 0;
271 	int sleep_flags;
272 
273 	MPS_FUNCTRACE(sc);
274 	/* If we are in attach call, do not sleep */
275 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
276 					? CAN_SLEEP:NO_SLEEP;
277 	error = 0;
278 
279 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
280 	   __func__, sleep_flags);
281 
282 	while (tries++ < 1200) {
283 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
284 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
285 
286 		/*
287 		 * Ensure the IOC is ready to talk.  If it's not, try
288 		 * resetting it.
289 		 */
290 		if (reg & MPI2_DOORBELL_USED) {
291 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
292 			    "reset\n");
293 			mps_diag_reset(sc, sleep_flags);
294 			DELAY(50000);
295 			continue;
296 		}
297 
298 		/* Is the adapter owned by another peer? */
299 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
300 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
301 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
302 			    "control of another peer host, aborting "
303 			    "initialization.\n");
304 			error = ENXIO;
305 			break;
306 		}
307 
308 		state = reg & MPI2_IOC_STATE_MASK;
309 		if (state == MPI2_IOC_STATE_READY) {
310 			/* Ready to go! */
311 			error = 0;
312 			break;
313 		} else if (state == MPI2_IOC_STATE_FAULT) {
314 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
315 			    "state 0x%x, resetting\n",
316 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
317 			mps_diag_reset(sc, sleep_flags);
318 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
319 			/* Need to take ownership */
320 			mps_message_unit_reset(sc, sleep_flags);
321 		} else if (state == MPI2_IOC_STATE_RESET) {
322 			/* Wait a bit, IOC might be in transition */
323 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
324 			    "IOC in unexpected reset state\n");
325 		} else {
326 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
327 			    "IOC in unknown state 0x%x\n", state);
328 			error = EINVAL;
329 			break;
330 		}
331 
332 		/* Wait 50ms for things to settle down. */
333 		DELAY(50000);
334 	}
335 
336 	if (error)
337 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
338 		    "Cannot transition IOC to ready\n");
339 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
340 
341 	return (error);
342 }
343 
344 static int
345 mps_transition_operational(struct mps_softc *sc)
346 {
347 	uint32_t reg, state;
348 	int error;
349 
350 	MPS_FUNCTRACE(sc);
351 
352 	error = 0;
353 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
354 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
355 
356 	state = reg & MPI2_IOC_STATE_MASK;
357 	if (state != MPI2_IOC_STATE_READY) {
358 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
359 		if ((error = mps_transition_ready(sc)) != 0) {
360 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
361 			    "failed to transition ready, exit\n");
362 			return (error);
363 		}
364 	}
365 
366 	error = mps_send_iocinit(sc);
367 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
368 
369 	return (error);
370 }
371 
372 static void
373 mps_resize_queues(struct mps_softc *sc)
374 {
375 	int reqcr, prireqcr;
376 
377 	/*
378 	 * Size the queues. Since the reply queues always need one free
379 	 * entry, we'll deduct one reply message here.  The LSI documents
380 	 * suggest instead to add a count to the request queue, but I think
381 	 * that it's better to deduct from reply queue.
382 	 */
383 	prireqcr = MAX(1, sc->max_prireqframes);
384 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
385 
386 	reqcr = MAX(2, sc->max_reqframes);
387 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
388 
389 	sc->num_reqs = prireqcr + reqcr;
390 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
391 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
392 
393 	/*
394 	 * Figure out the number of MSIx-based queues.  If the firmware or
395 	 * user has done something crazy and not allowed enough credit for
396 	 * the queues to be useful then don't enable multi-queue.
397 	 */
398 	if (sc->facts->MaxMSIxVectors < 2)
399 		sc->msi_msgs = 1;
400 
401 	if (sc->msi_msgs > 1) {
402 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
403 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
404 		if (sc->num_reqs / sc->msi_msgs < 2)
405 			sc->msi_msgs = 1;
406 	}
407 
408 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
409 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
410 }
411 
412 /*
413  * This is called during attach and when re-initializing due to a Diag Reset.
414  * IOC Facts is used to allocate many of the structures needed by the driver.
415  * If called from attach, de-allocation is not required because the driver has
416  * not allocated any structures yet, but if called from a Diag Reset, previously
417  * allocated structures based on IOC Facts will need to be freed and re-
418  * allocated bases on the latest IOC Facts.
419  */
420 static int
421 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
422 {
423 	int error;
424 	Mpi2IOCFactsReply_t saved_facts;
425 	uint8_t saved_mode, reallocating;
426 
427 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
428 
429 	/* Save old IOC Facts and then only reallocate if Facts have changed */
430 	if (!attaching) {
431 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
432 	}
433 
434 	/*
435 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
436 	 * a re-initialization and only return the error if attaching so the OS
437 	 * can handle it.
438 	 */
439 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
440 		if (attaching) {
441 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
442 			    "IOC Facts with error %d, exit\n", error);
443 			return (error);
444 		} else {
445 			panic("%s failed to get IOC Facts with error %d\n",
446 			    __func__, error);
447 		}
448 	}
449 
450 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
451 
452 	snprintf(sc->fw_version, sizeof(sc->fw_version),
453 	    "%02d.%02d.%02d.%02d",
454 	    sc->facts->FWVersion.Struct.Major,
455 	    sc->facts->FWVersion.Struct.Minor,
456 	    sc->facts->FWVersion.Struct.Unit,
457 	    sc->facts->FWVersion.Struct.Dev);
458 
459 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
460 	    MPS_DRIVER_VERSION);
461 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
462 	     sc->facts->IOCCapabilities,
463 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
464 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
465 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
466 
467 	/*
468 	 * If the chip doesn't support event replay then a hard reset will be
469 	 * required to trigger a full discovery.  Do the reset here then
470 	 * retransition to Ready.  A hard reset might have already been done,
471 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
472 	 * for a Diag Reset.
473 	 */
474 	if (attaching && ((sc->facts->IOCCapabilities &
475 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
476 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
477 		mps_diag_reset(sc, NO_SLEEP);
478 		if ((error = mps_transition_ready(sc)) != 0) {
479 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
480 			    "transition to ready with error %d, exit\n",
481 			    error);
482 			return (error);
483 		}
484 	}
485 
486 	/*
487 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
488 	 * changed from the previous IOC Facts, log a warning, but only if
489 	 * checking this after a Diag Reset and not during attach.
490 	 */
491 	saved_mode = sc->ir_firmware;
492 	if (sc->facts->IOCCapabilities &
493 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
494 		sc->ir_firmware = 1;
495 	if (!attaching) {
496 		if (sc->ir_firmware != saved_mode) {
497 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
498 			    "in IOC Facts does not match previous mode\n");
499 		}
500 	}
501 
502 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
503 	reallocating = FALSE;
504 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
505 
506 	if ((!attaching) &&
507 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
508 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
509 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
510 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
511 	    (saved_facts.ProductID != sc->facts->ProductID) ||
512 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
513 	    (saved_facts.IOCRequestFrameSize !=
514 	    sc->facts->IOCRequestFrameSize) ||
515 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
516 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
517 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
518 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
519 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
520 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
521 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
522 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
523 	    (saved_facts.MaxPersistentEntries !=
524 	    sc->facts->MaxPersistentEntries))) {
525 		reallocating = TRUE;
526 
527 		/* Record that we reallocated everything */
528 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
529 	}
530 
531 	/*
532 	 * Some things should be done if attaching or re-allocating after a Diag
533 	 * Reset, but are not needed after a Diag Reset if the FW has not
534 	 * changed.
535 	 */
536 	if (attaching || reallocating) {
537 		/*
538 		 * Check if controller supports FW diag buffers and set flag to
539 		 * enable each type.
540 		 */
541 		if (sc->facts->IOCCapabilities &
542 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
543 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
544 			    enabled = TRUE;
545 		if (sc->facts->IOCCapabilities &
546 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
547 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
548 			    enabled = TRUE;
549 		if (sc->facts->IOCCapabilities &
550 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
551 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
552 			    enabled = TRUE;
553 
554 		/*
555 		 * Set flag if EEDP is supported and if TLR is supported.
556 		 */
557 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
558 			sc->eedp_enabled = TRUE;
559 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
560 			sc->control_TLR = TRUE;
561 
562 		mps_resize_queues(sc);
563 
564 		/*
565 		 * Initialize all Tail Queues
566 		 */
567 		TAILQ_INIT(&sc->req_list);
568 		TAILQ_INIT(&sc->high_priority_req_list);
569 		TAILQ_INIT(&sc->chain_list);
570 		TAILQ_INIT(&sc->tm_list);
571 	}
572 
573 	/*
574 	 * If doing a Diag Reset and the FW is significantly different
575 	 * (reallocating will be set above in IOC Facts comparison), then all
576 	 * buffers based on the IOC Facts will need to be freed before they are
577 	 * reallocated.
578 	 */
579 	if (reallocating) {
580 		mps_iocfacts_free(sc);
581 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
582 		    saved_facts.MaxVolumes);
583 	}
584 
585 	/*
586 	 * Any deallocation has been completed.  Now start reallocating
587 	 * if needed.  Will only need to reallocate if attaching or if the new
588 	 * IOC Facts are different from the previous IOC Facts after a Diag
589 	 * Reset. Targets have already been allocated above if needed.
590 	 */
591 	error = 0;
592 	while (attaching || reallocating) {
593 		if ((error = mps_alloc_hw_queues(sc)) != 0)
594 			break;
595 		if ((error = mps_alloc_replies(sc)) != 0)
596 			break;
597 		if ((error = mps_alloc_requests(sc)) != 0)
598 			break;
599 		if ((error = mps_alloc_queues(sc)) != 0)
600 			break;
601 
602 		break;
603 	}
604 	if (error) {
605 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
606 		    "Failed to alloc queues with error %d\n", error);
607 		mps_free(sc);
608 		return (error);
609 	}
610 
611 	/* Always initialize the queues */
612 	bzero(sc->free_queue, sc->fqdepth * 4);
613 	mps_init_queues(sc);
614 
615 	/*
616 	 * Always get the chip out of the reset state, but only panic if not
617 	 * attaching.  If attaching and there is an error, that is handled by
618 	 * the OS.
619 	 */
620 	error = mps_transition_operational(sc);
621 	if (error != 0) {
622 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
623 		    "transition to operational with error %d\n", error);
624 		mps_free(sc);
625 		return (error);
626 	}
627 
628 	/*
629 	 * Finish the queue initialization.
630 	 * These are set here instead of in mps_init_queues() because the
631 	 * IOC resets these values during the state transition in
632 	 * mps_transition_operational().  The free index is set to 1
633 	 * because the corresponding index in the IOC is set to 0, and the
634 	 * IOC treats the queues as full if both are set to the same value.
635 	 * Hence the reason that the queue can't hold all of the possible
636 	 * replies.
637 	 */
638 	sc->replypostindex = 0;
639 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
640 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
641 
642 	/*
643 	 * Attach the subsystems so they can prepare their event masks.
644 	 * XXX Should be dynamic so that IM/IR and user modules can attach
645 	 */
646 	error = 0;
647 	while (attaching) {
648 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
649 		if ((error = mps_attach_log(sc)) != 0)
650 			break;
651 		if ((error = mps_attach_sas(sc)) != 0)
652 			break;
653 		if ((error = mps_attach_user(sc)) != 0)
654 			break;
655 		break;
656 	}
657 	if (error) {
658 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
659 		    "subsystems: error %d\n", error);
660 		mps_free(sc);
661 		return (error);
662 	}
663 
664 	if ((error = mps_pci_setup_interrupts(sc)) != 0) {
665 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
666 		    "interrupts\n");
667 		mps_free(sc);
668 		return (error);
669 	}
670 
671 	/*
672 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
673 	 * reset it after a Diag Reset, just in case.
674 	 */
675 	sc->WD_available = FALSE;
676 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
677 		sc->WD_available = TRUE;
678 
679 	return (error);
680 }
681 
682 /*
683  * This is called if memory is being free (during detach for example) and when
684  * buffers need to be reallocated due to a Diag Reset.
685  */
686 static void
687 mps_iocfacts_free(struct mps_softc *sc)
688 {
689 	struct mps_command *cm;
690 	int i;
691 
692 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
693 
694 	if (sc->free_busaddr != 0)
695 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
696 	if (sc->free_queue != NULL)
697 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
698 		    sc->queues_map);
699 	if (sc->queues_dmat != NULL)
700 		bus_dma_tag_destroy(sc->queues_dmat);
701 
702 	if (sc->chain_busaddr != 0)
703 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
704 	if (sc->chain_frames != NULL)
705 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
706 		    sc->chain_map);
707 	if (sc->chain_dmat != NULL)
708 		bus_dma_tag_destroy(sc->chain_dmat);
709 
710 	if (sc->sense_busaddr != 0)
711 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
712 	if (sc->sense_frames != NULL)
713 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
714 		    sc->sense_map);
715 	if (sc->sense_dmat != NULL)
716 		bus_dma_tag_destroy(sc->sense_dmat);
717 
718 	if (sc->reply_busaddr != 0)
719 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
720 	if (sc->reply_frames != NULL)
721 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
722 		    sc->reply_map);
723 	if (sc->reply_dmat != NULL)
724 		bus_dma_tag_destroy(sc->reply_dmat);
725 
726 	if (sc->req_busaddr != 0)
727 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
728 	if (sc->req_frames != NULL)
729 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
730 	if (sc->req_dmat != NULL)
731 		bus_dma_tag_destroy(sc->req_dmat);
732 
733 	if (sc->chains != NULL)
734 		free(sc->chains, M_MPT2);
735 	if (sc->commands != NULL) {
736 		for (i = 1; i < sc->num_reqs; i++) {
737 			cm = &sc->commands[i];
738 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
739 		}
740 		free(sc->commands, M_MPT2);
741 	}
742 	if (sc->buffer_dmat != NULL)
743 		bus_dma_tag_destroy(sc->buffer_dmat);
744 
745 	mps_pci_free_interrupts(sc);
746 	free(sc->queues, M_MPT2);
747 	sc->queues = NULL;
748 }
749 
750 /*
751  * The terms diag reset and hard reset are used interchangeably in the MPI
752  * docs to mean resetting the controller chip.  In this code diag reset
753  * cleans everything up, and the hard reset function just sends the reset
754  * sequence to the chip.  This should probably be refactored so that every
755  * subsystem gets a reset notification of some sort, and can clean up
756  * appropriately.
757  */
758 int
759 mps_reinit(struct mps_softc *sc)
760 {
761 	int error;
762 	struct mpssas_softc *sassc;
763 
764 	sassc = sc->sassc;
765 
766 	MPS_FUNCTRACE(sc);
767 
768 	mtx_assert(&sc->mps_mtx, MA_OWNED);
769 
770 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
771 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
772 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
773 		return 0;
774 	}
775 
776 	/* make sure the completion callbacks can recognize they're getting
777 	 * a NULL cm_reply due to a reset.
778 	 */
779 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
780 
781 	/*
782 	 * Mask interrupts here.
783 	 */
784 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
785 	mps_mask_intr(sc);
786 
787 	error = mps_diag_reset(sc, CAN_SLEEP);
788 	if (error != 0) {
789 		/* XXXSL No need to panic here */
790 		panic("%s hard reset failed with error %d\n",
791 		    __func__, error);
792 	}
793 
794 	/* Restore the PCI state, including the MSI-X registers */
795 	mps_pci_restore(sc);
796 
797 	/* Give the I/O subsystem special priority to get itself prepared */
798 	mpssas_handle_reinit(sc);
799 
800 	/*
801 	 * Get IOC Facts and allocate all structures based on this information.
802 	 * The attach function will also call mps_iocfacts_allocate at startup.
803 	 * If relevant values have changed in IOC Facts, this function will free
804 	 * all of the memory based on IOC Facts and reallocate that memory.
805 	 */
806 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
807 		panic("%s IOC Facts based allocation failed with error %d\n",
808 		    __func__, error);
809 	}
810 
811 	/*
812 	 * Mapping structures will be re-allocated after getting IOC Page8, so
813 	 * free these structures here.
814 	 */
815 	mps_mapping_exit(sc);
816 
817 	/*
818 	 * The static page function currently read is IOC Page8.  Others can be
819 	 * added in future.  It's possible that the values in IOC Page8 have
820 	 * changed after a Diag Reset due to user modification, so always read
821 	 * these.  Interrupts are masked, so unmask them before getting config
822 	 * pages.
823 	 */
824 	mps_unmask_intr(sc);
825 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
826 	mps_base_static_config_pages(sc);
827 
828 	/*
829 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
830 	 * mapping tables.
831 	 */
832 	mps_mapping_initialize(sc);
833 
834 	/*
835 	 * Restart will reload the event masks clobbered by the reset, and
836 	 * then enable the port.
837 	 */
838 	mps_reregister_events(sc);
839 
840 	/* the end of discovery will release the simq, so we're done. */
841 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
842 	    sc, sc->replypostindex, sc->replyfreeindex);
843 
844 	mpssas_release_simq_reinit(sassc);
845 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
846 
847 	return 0;
848 }
849 
850 /* Wait for the chip to ACK a word that we've put into its FIFO
851  * Wait for <timeout> seconds. In single loop wait for busy loop
852  * for 500 microseconds.
853  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
854  * */
855 static int
856 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
857 {
858 
859 	u32 cntdn, count;
860 	u32 int_status;
861 	u32 doorbell;
862 
863 	count = 0;
864 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
865 	do {
866 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
867 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
868 			mps_dprint(sc, MPS_TRACE,
869 			"%s: successful count(%d), timeout(%d)\n",
870 			__func__, count, timeout);
871 		return 0;
872 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
873 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
874 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
875 				MPI2_IOC_STATE_FAULT) {
876 				mps_dprint(sc, MPS_FAULT,
877 					"fault_state(0x%04x)!\n", doorbell);
878 				return (EFAULT);
879 			}
880 		} else if (int_status == 0xFFFFFFFF)
881 			goto out;
882 
883 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
884 		* 0.5 milisecond */
885 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
886 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
887 			"mpsdba", hz/1000);
888 		else if (sleep_flag == CAN_SLEEP)
889 			pause("mpsdba", hz/1000);
890 		else
891 			DELAY(500);
892 		count++;
893 	} while (--cntdn);
894 
895 	out:
896 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
897 		"int_status(%x)!\n", __func__, count, int_status);
898 	return (ETIMEDOUT);
899 
900 }
901 
902 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
903 static int
904 mps_wait_db_int(struct mps_softc *sc)
905 {
906 	int retry;
907 
908 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
909 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
910 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
911 			return (0);
912 		DELAY(2000);
913 	}
914 	return (ETIMEDOUT);
915 }
916 
917 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
918 static int
919 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
920     int req_sz, int reply_sz, int timeout)
921 {
922 	uint32_t *data32;
923 	uint16_t *data16;
924 	int i, count, ioc_sz, residual;
925 	int sleep_flags = CAN_SLEEP;
926 
927 	if (curthread->td_no_sleeping != 0)
928 		sleep_flags = NO_SLEEP;
929 
930 	/* Step 1 */
931 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
932 
933 	/* Step 2 */
934 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
935 		return (EBUSY);
936 
937 	/* Step 3
938 	 * Announce that a message is coming through the doorbell.  Messages
939 	 * are pushed at 32bit words, so round up if needed.
940 	 */
941 	count = (req_sz + 3) / 4;
942 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
943 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
944 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
945 
946 	/* Step 4 */
947 	if (mps_wait_db_int(sc) ||
948 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
949 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
950 		return (ENXIO);
951 	}
952 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
953 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
954 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
955 		return (ENXIO);
956 	}
957 
958 	/* Step 5 */
959 	/* Clock out the message data synchronously in 32-bit dwords*/
960 	data32 = (uint32_t *)req;
961 	for (i = 0; i < count; i++) {
962 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
963 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
964 			mps_dprint(sc, MPS_FAULT,
965 			    "Timeout while writing doorbell\n");
966 			return (ENXIO);
967 		}
968 	}
969 
970 	/* Step 6 */
971 	/* Clock in the reply in 16-bit words.  The total length of the
972 	 * message is always in the 4th byte, so clock out the first 2 words
973 	 * manually, then loop the rest.
974 	 */
975 	data16 = (uint16_t *)reply;
976 	if (mps_wait_db_int(sc) != 0) {
977 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
978 		return (ENXIO);
979 	}
980 	data16[0] =
981 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
982 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
983 	if (mps_wait_db_int(sc) != 0) {
984 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
985 		return (ENXIO);
986 	}
987 	data16[1] =
988 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
989 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
990 
991 	/* Number of 32bit words in the message */
992 	ioc_sz = reply->MsgLength;
993 
994 	/*
995 	 * Figure out how many 16bit words to clock in without overrunning.
996 	 * The precision loss with dividing reply_sz can safely be
997 	 * ignored because the messages can only be multiples of 32bits.
998 	 */
999 	residual = 0;
1000 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1001 	if (count < ioc_sz * 2) {
1002 		residual = ioc_sz * 2 - count;
1003 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1004 		    "residual message words\n", residual);
1005 	}
1006 
1007 	for (i = 2; i < count; i++) {
1008 		if (mps_wait_db_int(sc) != 0) {
1009 			mps_dprint(sc, MPS_FAULT,
1010 			    "Timeout reading doorbell %d\n", i);
1011 			return (ENXIO);
1012 		}
1013 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1014 		    MPI2_DOORBELL_DATA_MASK;
1015 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1016 	}
1017 
1018 	/*
1019 	 * Pull out residual words that won't fit into the provided buffer.
1020 	 * This keeps the chip from hanging due to a driver programming
1021 	 * error.
1022 	 */
1023 	while (residual--) {
1024 		if (mps_wait_db_int(sc) != 0) {
1025 			mps_dprint(sc, MPS_FAULT,
1026 			    "Timeout reading doorbell\n");
1027 			return (ENXIO);
1028 		}
1029 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1030 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1031 	}
1032 
1033 	/* Step 7 */
1034 	if (mps_wait_db_int(sc) != 0) {
1035 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1036 		return (ENXIO);
1037 	}
1038 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1039 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1040 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1041 
1042 	return (0);
1043 }
1044 
1045 static void
1046 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1047 {
1048 	reply_descriptor rd;
1049 	MPS_FUNCTRACE(sc);
1050 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1051 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1052 
1053 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1054 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1055 
1056 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1057 		sc->io_cmds_highwater++;
1058 	rd.u.low = cm->cm_desc.Words.Low;
1059 	rd.u.high = cm->cm_desc.Words.High;
1060 	rd.word = htole64(rd.word);
1061 	/* TODO-We may need to make below regwrite atomic */
1062 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1063 	    rd.u.low);
1064 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1065 	    rd.u.high);
1066 }
1067 
1068 /*
1069  * Just the FACTS, ma'am.
1070  */
1071 static int
1072 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1073 {
1074 	MPI2_DEFAULT_REPLY *reply;
1075 	MPI2_IOC_FACTS_REQUEST request;
1076 	int error, req_sz, reply_sz;
1077 
1078 	MPS_FUNCTRACE(sc);
1079 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1080 
1081 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1082 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1083 	reply = (MPI2_DEFAULT_REPLY *)facts;
1084 
1085 	bzero(&request, req_sz);
1086 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1087 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1088 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1089 
1090 	return (error);
1091 }
1092 
1093 static int
1094 mps_send_iocinit(struct mps_softc *sc)
1095 {
1096 	MPI2_IOC_INIT_REQUEST	init;
1097 	MPI2_DEFAULT_REPLY	reply;
1098 	int req_sz, reply_sz, error;
1099 	struct timeval now;
1100 	uint64_t time_in_msec;
1101 
1102 	MPS_FUNCTRACE(sc);
1103 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1104 
1105 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1106 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1107 	bzero(&init, req_sz);
1108 	bzero(&reply, reply_sz);
1109 
1110 	/*
1111 	 * Fill in the init block.  Note that most addresses are
1112 	 * deliberately in the lower 32bits of memory.  This is a micro-
1113 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1114 	 */
1115 	init.Function = MPI2_FUNCTION_IOC_INIT;
1116 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1117 	init.MsgVersion = htole16(MPI2_VERSION);
1118 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1119 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1120 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1121 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1122 	init.SenseBufferAddressHigh = 0;
1123 	init.SystemReplyAddressHigh = 0;
1124 	init.SystemRequestFrameBaseAddress.High = 0;
1125 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1126 	init.ReplyDescriptorPostQueueAddress.High = 0;
1127 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1128 	init.ReplyFreeQueueAddress.High = 0;
1129 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1130 	getmicrotime(&now);
1131 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1132 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1133 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1134 
1135 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1136 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1137 		error = ENXIO;
1138 
1139 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1140 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1141 	return (error);
1142 }
1143 
1144 void
1145 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1146 {
1147 	bus_addr_t *addr;
1148 
1149 	addr = arg;
1150 	*addr = segs[0].ds_addr;
1151 }
1152 
1153 static int
1154 mps_alloc_queues(struct mps_softc *sc)
1155 {
1156 	struct mps_queue *q;
1157 	int nq, i;
1158 
1159 	nq = sc->msi_msgs;
1160 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1161 
1162 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1163 	    M_NOWAIT|M_ZERO);
1164 	if (sc->queues == NULL)
1165 		return (ENOMEM);
1166 
1167 	for (i = 0; i < nq; i++) {
1168 		q = &sc->queues[i];
1169 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1170 		q->sc = sc;
1171 		q->qnum = i;
1172 	}
1173 
1174 	return (0);
1175 }
1176 
1177 static int
1178 mps_alloc_hw_queues(struct mps_softc *sc)
1179 {
1180 	bus_addr_t queues_busaddr;
1181 	uint8_t *queues;
1182 	int qsize, fqsize, pqsize;
1183 
1184 	/*
1185 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1186 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1187 	 * This queue supplies fresh reply frames for the firmware to use.
1188 	 *
1189 	 * The reply descriptor post queue contains 8 byte entries in
1190 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1191 	 * contains filled-in reply frames sent from the firmware to the host.
1192 	 *
1193 	 * These two queues are allocated together for simplicity.
1194 	 */
1195 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1196 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1197 	fqsize= sc->fqdepth * 4;
1198 	pqsize = sc->pqdepth * 8;
1199 	qsize = fqsize + pqsize;
1200 
1201         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1202 				16, 0,			/* algnmnt, boundary */
1203 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1204 				BUS_SPACE_MAXADDR,	/* highaddr */
1205 				NULL, NULL,		/* filter, filterarg */
1206                                 qsize,			/* maxsize */
1207                                 1,			/* nsegments */
1208                                 qsize,			/* maxsegsize */
1209                                 0,			/* flags */
1210                                 NULL, NULL,		/* lockfunc, lockarg */
1211                                 &sc->queues_dmat)) {
1212 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1213 		return (ENOMEM);
1214         }
1215         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1216 	    &sc->queues_map)) {
1217 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1218 		return (ENOMEM);
1219         }
1220         bzero(queues, qsize);
1221         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1222 	    mps_memaddr_cb, &queues_busaddr, 0);
1223 
1224 	sc->free_queue = (uint32_t *)queues;
1225 	sc->free_busaddr = queues_busaddr;
1226 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1227 	sc->post_busaddr = queues_busaddr + fqsize;
1228 
1229 	return (0);
1230 }
1231 
1232 static int
1233 mps_alloc_replies(struct mps_softc *sc)
1234 {
1235 	int rsize, num_replies;
1236 
1237 	/*
1238 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1239 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1240 	 * replies can be used at once.
1241 	 */
1242 	num_replies = max(sc->fqdepth, sc->num_replies);
1243 
1244 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1245         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1246 				4, 0,			/* algnmnt, boundary */
1247 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1248 				BUS_SPACE_MAXADDR,	/* highaddr */
1249 				NULL, NULL,		/* filter, filterarg */
1250                                 rsize,			/* maxsize */
1251                                 1,			/* nsegments */
1252                                 rsize,			/* maxsegsize */
1253                                 0,			/* flags */
1254                                 NULL, NULL,		/* lockfunc, lockarg */
1255                                 &sc->reply_dmat)) {
1256 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1257 		return (ENOMEM);
1258         }
1259         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1260 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1261 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1262 		return (ENOMEM);
1263         }
1264         bzero(sc->reply_frames, rsize);
1265         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1266 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1267 
1268 	return (0);
1269 }
1270 
1271 static int
1272 mps_alloc_requests(struct mps_softc *sc)
1273 {
1274 	struct mps_command *cm;
1275 	struct mps_chain *chain;
1276 	int i, rsize, nsegs;
1277 
1278 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1279         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1280 				16, 0,			/* algnmnt, boundary */
1281 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1282 				BUS_SPACE_MAXADDR,	/* highaddr */
1283 				NULL, NULL,		/* filter, filterarg */
1284                                 rsize,			/* maxsize */
1285                                 1,			/* nsegments */
1286                                 rsize,			/* maxsegsize */
1287                                 0,			/* flags */
1288                                 NULL, NULL,		/* lockfunc, lockarg */
1289                                 &sc->req_dmat)) {
1290 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1291 		return (ENOMEM);
1292         }
1293         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1294 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1295 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1296 		return (ENOMEM);
1297         }
1298         bzero(sc->req_frames, rsize);
1299         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1300 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1301 
1302 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1303         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1304 				16, 0,			/* algnmnt, boundary */
1305 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1306 				BUS_SPACE_MAXADDR,	/* highaddr */
1307 				NULL, NULL,		/* filter, filterarg */
1308                                 rsize,			/* maxsize */
1309                                 1,			/* nsegments */
1310                                 rsize,			/* maxsegsize */
1311                                 0,			/* flags */
1312                                 NULL, NULL,		/* lockfunc, lockarg */
1313                                 &sc->chain_dmat)) {
1314 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1315 		return (ENOMEM);
1316         }
1317         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1318 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1319 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1320 		return (ENOMEM);
1321         }
1322         bzero(sc->chain_frames, rsize);
1323         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1324 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1325 
1326 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1327         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1328 				1, 0,			/* algnmnt, boundary */
1329 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1330 				BUS_SPACE_MAXADDR,	/* highaddr */
1331 				NULL, NULL,		/* filter, filterarg */
1332                                 rsize,			/* maxsize */
1333                                 1,			/* nsegments */
1334                                 rsize,			/* maxsegsize */
1335                                 0,			/* flags */
1336                                 NULL, NULL,		/* lockfunc, lockarg */
1337                                 &sc->sense_dmat)) {
1338 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1339 		return (ENOMEM);
1340         }
1341         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1342 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1343 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1344 		return (ENOMEM);
1345         }
1346         bzero(sc->sense_frames, rsize);
1347         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1348 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1349 
1350 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1351 	    M_WAITOK | M_ZERO);
1352 	if(!sc->chains) {
1353 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1354 		return (ENOMEM);
1355 	}
1356 	for (i = 0; i < sc->max_chains; i++) {
1357 		chain = &sc->chains[i];
1358 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1359 		    i * sc->facts->IOCRequestFrameSize * 4);
1360 		chain->chain_busaddr = sc->chain_busaddr +
1361 		    i * sc->facts->IOCRequestFrameSize * 4;
1362 		mps_free_chain(sc, chain);
1363 		sc->chain_free_lowwater++;
1364 	}
1365 
1366 	/* XXX Need to pick a more precise value */
1367 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1368         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1369 				1, 0,			/* algnmnt, boundary */
1370 				BUS_SPACE_MAXADDR,	/* lowaddr */
1371 				BUS_SPACE_MAXADDR,	/* highaddr */
1372 				NULL, NULL,		/* filter, filterarg */
1373                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1374                                 nsegs,			/* nsegments */
1375                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1376                                 BUS_DMA_ALLOCNOW,	/* flags */
1377                                 busdma_lock_mutex,	/* lockfunc */
1378 				&sc->mps_mtx,		/* lockarg */
1379                                 &sc->buffer_dmat)) {
1380 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1381 		return (ENOMEM);
1382         }
1383 
1384 	/*
1385 	 * SMID 0 cannot be used as a free command per the firmware spec.
1386 	 * Just drop that command instead of risking accounting bugs.
1387 	 */
1388 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1389 	    M_MPT2, M_WAITOK | M_ZERO);
1390 	if(!sc->commands) {
1391 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1392 		return (ENOMEM);
1393 	}
1394 	for (i = 1; i < sc->num_reqs; i++) {
1395 		cm = &sc->commands[i];
1396 		cm->cm_req = sc->req_frames +
1397 		    i * sc->facts->IOCRequestFrameSize * 4;
1398 		cm->cm_req_busaddr = sc->req_busaddr +
1399 		    i * sc->facts->IOCRequestFrameSize * 4;
1400 		cm->cm_sense = &sc->sense_frames[i];
1401 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1402 		cm->cm_desc.Default.SMID = i;
1403 		cm->cm_sc = sc;
1404 		TAILQ_INIT(&cm->cm_chain_list);
1405 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1406 
1407 		/* XXX Is a failure here a critical problem? */
1408 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1409 			if (i <= sc->facts->HighPriorityCredit)
1410 				mps_free_high_priority_command(sc, cm);
1411 			else
1412 				mps_free_command(sc, cm);
1413 		else {
1414 			panic("failed to allocate command %d\n", i);
1415 			sc->num_reqs = i;
1416 			break;
1417 		}
1418 	}
1419 
1420 	return (0);
1421 }
1422 
1423 static int
1424 mps_init_queues(struct mps_softc *sc)
1425 {
1426 	int i;
1427 
1428 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1429 
1430 	/*
1431 	 * According to the spec, we need to use one less reply than we
1432 	 * have space for on the queue.  So sc->num_replies (the number we
1433 	 * use) should be less than sc->fqdepth (allocated size).
1434 	 */
1435 	if (sc->num_replies >= sc->fqdepth)
1436 		return (EINVAL);
1437 
1438 	/*
1439 	 * Initialize all of the free queue entries.
1440 	 */
1441 	for (i = 0; i < sc->fqdepth; i++)
1442 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1443 	sc->replyfreeindex = sc->num_replies;
1444 
1445 	return (0);
1446 }
1447 
1448 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1449  * Next are the global settings, if they exist.  Highest are the per-unit
1450  * settings, if they exist.
1451  */
1452 void
1453 mps_get_tunables(struct mps_softc *sc)
1454 {
1455 	char tmpstr[80];
1456 
1457 	/* XXX default to some debugging for now */
1458 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1459 	sc->disable_msix = 0;
1460 	sc->disable_msi = 0;
1461 	sc->max_msix = MPS_MSIX_MAX;
1462 	sc->max_chains = MPS_CHAIN_FRAMES;
1463 	sc->max_io_pages = MPS_MAXIO_PAGES;
1464 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1465 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1466 	sc->use_phynum = 1;
1467 	sc->max_reqframes = MPS_REQ_FRAMES;
1468 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1469 	sc->max_replyframes = MPS_REPLY_FRAMES;
1470 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1471 
1472 	/*
1473 	 * Grab the global variables.
1474 	 */
1475 	TUNABLE_INT_FETCH("hw.mps.debug_level", &sc->mps_debug);
1476 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1477 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1478 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1479 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1480 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1481 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1482 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1483 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1484 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1485 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1486 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1487 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1488 
1489 	/* Grab the unit-instance variables */
1490 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1491 	    device_get_unit(sc->mps_dev));
1492 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
1493 
1494 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1495 	    device_get_unit(sc->mps_dev));
1496 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1497 
1498 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1499 	    device_get_unit(sc->mps_dev));
1500 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1501 
1502 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1503 	    device_get_unit(sc->mps_dev));
1504 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1505 
1506 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1507 	    device_get_unit(sc->mps_dev));
1508 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1509 
1510 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1511 	    device_get_unit(sc->mps_dev));
1512 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1513 
1514 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1515 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1516 	    device_get_unit(sc->mps_dev));
1517 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1518 
1519 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1520 	    device_get_unit(sc->mps_dev));
1521 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1522 
1523 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1524 	    device_get_unit(sc->mps_dev));
1525 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1526 
1527 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1528 	    device_get_unit(sc->mps_dev));
1529 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1530 
1531 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1532 	    device_get_unit(sc->mps_dev));
1533 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1534 
1535 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1536 	    device_get_unit(sc->mps_dev));
1537 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1538 
1539 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1540 	    device_get_unit(sc->mps_dev));
1541 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1542 
1543 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1544 	    device_get_unit(sc->mps_dev));
1545 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1546 
1547 }
1548 
1549 static void
1550 mps_setup_sysctl(struct mps_softc *sc)
1551 {
1552 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1553 	struct sysctl_oid	*sysctl_tree = NULL;
1554 	char tmpstr[80], tmpstr2[80];
1555 
1556 	/*
1557 	 * Setup the sysctl variable so the user can change the debug level
1558 	 * on the fly.
1559 	 */
1560 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1561 	    device_get_unit(sc->mps_dev));
1562 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1563 
1564 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1565 	if (sysctl_ctx != NULL)
1566 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1567 
1568 	if (sysctl_tree == NULL) {
1569 		sysctl_ctx_init(&sc->sysctl_ctx);
1570 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1571 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1572 		    CTLFLAG_RD, 0, tmpstr);
1573 		if (sc->sysctl_tree == NULL)
1574 			return;
1575 		sysctl_ctx = &sc->sysctl_ctx;
1576 		sysctl_tree = sc->sysctl_tree;
1577 	}
1578 
1579 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1580 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
1581 	    "mps debug level");
1582 
1583 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1584 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1585 	    "Disable the use of MSI-X interrupts");
1586 
1587 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1588 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1589 	    "Disable the use of MSI interrupts");
1590 
1591 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1592 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1593 	    "User-defined maximum number of MSIX queues");
1594 
1595 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1596 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1597 	    "Negotiated number of MSIX queues");
1598 
1599 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1600 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1601 	    "Total number of allocated request frames");
1602 
1603 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1604 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1605 	    "Total number of allocated high priority request frames");
1606 
1607 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1608 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1609 	    "Total number of allocated reply frames");
1610 
1611 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1612 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1613 	    "Total number of event frames allocated");
1614 
1615 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1616 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1617 	    strlen(sc->fw_version), "firmware version");
1618 
1619 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1620 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1621 	    strlen(MPS_DRIVER_VERSION), "driver version");
1622 
1623 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1624 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1625 	    &sc->io_cmds_active, 0, "number of currently active commands");
1626 
1627 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1628 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1629 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1630 
1631 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1632 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1633 	    &sc->chain_free, 0, "number of free chain elements");
1634 
1635 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1636 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1637 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1638 
1639 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1640 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1641 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1642 
1643 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1644 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1645 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1646 	    "IOCFacts)");
1647 
1648 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1649 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1650 	    "enable SSU to SATA SSD/HDD at shutdown");
1651 
1652 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1653 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1654 	    &sc->chain_alloc_fail, "chain allocation failures");
1655 
1656 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1657 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1658 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1659 	    "spinup after SATA ID error");
1660 
1661 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1662 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1663 	    mps_mapping_dump, "A", "Mapping Table Dump");
1664 
1665 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1666 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1667 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1668 
1669 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1670 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1671 	    "Use the phy number for enumeration");
1672 }
1673 
1674 int
1675 mps_attach(struct mps_softc *sc)
1676 {
1677 	int error;
1678 
1679 	MPS_FUNCTRACE(sc);
1680 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1681 
1682 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1683 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1684 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1685 	TAILQ_INIT(&sc->event_list);
1686 	timevalclear(&sc->lastfail);
1687 
1688 	if ((error = mps_transition_ready(sc)) != 0) {
1689 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1690 		    "ready\n");
1691 		return (error);
1692 	}
1693 
1694 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1695 	    M_ZERO|M_NOWAIT);
1696 	if(!sc->facts) {
1697 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1698 		    "exit\n");
1699 		return (ENOMEM);
1700 	}
1701 
1702 	/*
1703 	 * Get IOC Facts and allocate all structures based on this information.
1704 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1705 	 * Facts. If relevant values have changed in IOC Facts, this function
1706 	 * will free all of the memory based on IOC Facts and reallocate that
1707 	 * memory.  If this fails, any allocated memory should already be freed.
1708 	 */
1709 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1710 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1711 		    "failed with error %d, exit\n", error);
1712 		return (error);
1713 	}
1714 
1715 	/* Start the periodic watchdog check on the IOC Doorbell */
1716 	mps_periodic(sc);
1717 
1718 	/*
1719 	 * The portenable will kick off discovery events that will drive the
1720 	 * rest of the initialization process.  The CAM/SAS module will
1721 	 * hold up the boot sequence until discovery is complete.
1722 	 */
1723 	sc->mps_ich.ich_func = mps_startup;
1724 	sc->mps_ich.ich_arg = sc;
1725 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1726 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1727 		    "Cannot establish MPS config hook\n");
1728 		error = EINVAL;
1729 	}
1730 
1731 	/*
1732 	 * Allow IR to shutdown gracefully when shutdown occurs.
1733 	 */
1734 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1735 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1736 
1737 	if (sc->shutdown_eh == NULL)
1738 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1739 		    "shutdown event registration failed\n");
1740 
1741 	mps_setup_sysctl(sc);
1742 
1743 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1744 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1745 
1746 	return (error);
1747 }
1748 
1749 /* Run through any late-start handlers. */
1750 static void
1751 mps_startup(void *arg)
1752 {
1753 	struct mps_softc *sc;
1754 
1755 	sc = (struct mps_softc *)arg;
1756 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1757 
1758 	mps_lock(sc);
1759 	mps_unmask_intr(sc);
1760 
1761 	/* initialize device mapping tables */
1762 	mps_base_static_config_pages(sc);
1763 	mps_mapping_initialize(sc);
1764 	mpssas_startup(sc);
1765 	mps_unlock(sc);
1766 
1767 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
1768 	config_intrhook_disestablish(&sc->mps_ich);
1769 	sc->mps_ich.ich_arg = NULL;
1770 
1771 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1772 }
1773 
1774 /* Periodic watchdog.  Is called with the driver lock already held. */
1775 static void
1776 mps_periodic(void *arg)
1777 {
1778 	struct mps_softc *sc;
1779 	uint32_t db;
1780 
1781 	sc = (struct mps_softc *)arg;
1782 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1783 		return;
1784 
1785 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1786 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1787 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1788 		mps_reinit(sc);
1789 	}
1790 
1791 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1792 }
1793 
1794 static void
1795 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1796     MPI2_EVENT_NOTIFICATION_REPLY *event)
1797 {
1798 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1799 
1800 	MPS_DPRINT_EVENT(sc, generic, event);
1801 
1802 	switch (event->Event) {
1803 	case MPI2_EVENT_LOG_DATA:
1804 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1805 		if (sc->mps_debug & MPS_EVENT)
1806 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1807 		break;
1808 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1809 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1810 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1811 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1812 		     entry->LogSequence);
1813 		break;
1814 	default:
1815 		break;
1816 	}
1817 	return;
1818 }
1819 
1820 static int
1821 mps_attach_log(struct mps_softc *sc)
1822 {
1823 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1824 
1825 	bzero(events, 16);
1826 	setbit(events, MPI2_EVENT_LOG_DATA);
1827 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1828 
1829 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1830 	    &sc->mps_log_eh);
1831 
1832 	return (0);
1833 }
1834 
1835 static int
1836 mps_detach_log(struct mps_softc *sc)
1837 {
1838 
1839 	if (sc->mps_log_eh != NULL)
1840 		mps_deregister_events(sc, sc->mps_log_eh);
1841 	return (0);
1842 }
1843 
1844 /*
1845  * Free all of the driver resources and detach submodules.  Should be called
1846  * without the lock held.
1847  */
1848 int
1849 mps_free(struct mps_softc *sc)
1850 {
1851 	int error;
1852 
1853 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1854 	/* Turn off the watchdog */
1855 	mps_lock(sc);
1856 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1857 	mps_unlock(sc);
1858 	/* Lock must not be held for this */
1859 	callout_drain(&sc->periodic);
1860 	callout_drain(&sc->device_check_callout);
1861 
1862 	if (((error = mps_detach_log(sc)) != 0) ||
1863 	    ((error = mps_detach_sas(sc)) != 0)) {
1864 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
1865 		    "subsystems, exit\n");
1866 		return (error);
1867 	}
1868 
1869 	mps_detach_user(sc);
1870 
1871 	/* Put the IOC back in the READY state. */
1872 	mps_lock(sc);
1873 	if ((error = mps_transition_ready(sc)) != 0) {
1874 		mps_unlock(sc);
1875 		return (error);
1876 	}
1877 	mps_unlock(sc);
1878 
1879 	if (sc->facts != NULL)
1880 		free(sc->facts, M_MPT2);
1881 
1882 	/*
1883 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
1884 	 * to free these buffers too.
1885 	 */
1886 	mps_iocfacts_free(sc);
1887 
1888 	if (sc->sysctl_tree != NULL)
1889 		sysctl_ctx_free(&sc->sysctl_ctx);
1890 
1891 	/* Deregister the shutdown function */
1892 	if (sc->shutdown_eh != NULL)
1893 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1894 
1895 	mtx_destroy(&sc->mps_mtx);
1896 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1897 
1898 	return (0);
1899 }
1900 
1901 static __inline void
1902 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
1903 {
1904 	MPS_FUNCTRACE(sc);
1905 
1906 	if (cm == NULL) {
1907 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
1908 		return;
1909 	}
1910 
1911 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1912 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1913 
1914 	if (cm->cm_complete != NULL) {
1915 		mps_dprint(sc, MPS_TRACE,
1916 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
1917 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
1918 			   cm->cm_reply);
1919 		cm->cm_complete(sc, cm);
1920 	}
1921 
1922 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1923 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
1924 		wakeup(cm);
1925 	}
1926 
1927 	if (cm->cm_sc->io_cmds_active != 0) {
1928 		cm->cm_sc->io_cmds_active--;
1929 	} else {
1930 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
1931 		    "out of sync - resynching to 0\n");
1932 	}
1933 }
1934 
1935 
1936 static void
1937 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
1938 {
1939 	union loginfo_type {
1940 		u32     loginfo;
1941 		struct {
1942 			u32     subcode:16;
1943 			u32     code:8;
1944 			u32     originator:4;
1945 			u32     bus_type:4;
1946 		} dw;
1947 	};
1948 	union loginfo_type sas_loginfo;
1949 	char *originator_str = NULL;
1950 
1951 	sas_loginfo.loginfo = log_info;
1952 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1953 		return;
1954 
1955 	/* each nexus loss loginfo */
1956 	if (log_info == 0x31170000)
1957 		return;
1958 
1959 	/* eat the loginfos associated with task aborts */
1960 	if ((log_info == 30050000 || log_info ==
1961 	    0x31140000 || log_info == 0x31130000))
1962 		return;
1963 
1964 	switch (sas_loginfo.dw.originator) {
1965 	case 0:
1966 		originator_str = "IOP";
1967 		break;
1968 	case 1:
1969 		originator_str = "PL";
1970 		break;
1971 	case 2:
1972 		originator_str = "IR";
1973 		break;
1974 }
1975 
1976 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
1977 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
1978 	originator_str, sas_loginfo.dw.code,
1979 	sas_loginfo.dw.subcode);
1980 }
1981 
1982 static void
1983 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
1984 {
1985 	MPI2DefaultReply_t *mpi_reply;
1986 	u16 sc_status;
1987 
1988 	mpi_reply = (MPI2DefaultReply_t*)reply;
1989 	sc_status = le16toh(mpi_reply->IOCStatus);
1990 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1991 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1992 }
1993 void
1994 mps_intr(void *data)
1995 {
1996 	struct mps_softc *sc;
1997 	uint32_t status;
1998 
1999 	sc = (struct mps_softc *)data;
2000 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2001 
2002 	/*
2003 	 * Check interrupt status register to flush the bus.  This is
2004 	 * needed for both INTx interrupts and driver-driven polling
2005 	 */
2006 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2007 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2008 		return;
2009 
2010 	mps_lock(sc);
2011 	mps_intr_locked(data);
2012 	mps_unlock(sc);
2013 	return;
2014 }
2015 
2016 /*
2017  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2018  * chip.  Hopefully this theory is correct.
2019  */
2020 void
2021 mps_intr_msi(void *data)
2022 {
2023 	struct mps_softc *sc;
2024 
2025 	sc = (struct mps_softc *)data;
2026 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2027 	mps_lock(sc);
2028 	mps_intr_locked(data);
2029 	mps_unlock(sc);
2030 	return;
2031 }
2032 
2033 /*
2034  * The locking is overly broad and simplistic, but easy to deal with for now.
2035  */
2036 void
2037 mps_intr_locked(void *data)
2038 {
2039 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2040 	struct mps_softc *sc;
2041 	struct mps_command *cm = NULL;
2042 	uint8_t flags;
2043 	u_int pq;
2044 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2045 	mps_fw_diagnostic_buffer_t *pBuffer;
2046 
2047 	sc = (struct mps_softc *)data;
2048 
2049 	pq = sc->replypostindex;
2050 	mps_dprint(sc, MPS_TRACE,
2051 	    "%s sc %p starting with replypostindex %u\n",
2052 	    __func__, sc, sc->replypostindex);
2053 
2054 	for ( ;; ) {
2055 		cm = NULL;
2056 		desc = &sc->post_queue[sc->replypostindex];
2057 		flags = desc->Default.ReplyFlags &
2058 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2059 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2060 		 || (le32toh(desc->Words.High) == 0xffffffff))
2061 			break;
2062 
2063 		/* increment the replypostindex now, so that event handlers
2064 		 * and cm completion handlers which decide to do a diag
2065 		 * reset can zero it without it getting incremented again
2066 		 * afterwards, and we break out of this loop on the next
2067 		 * iteration since the reply post queue has been cleared to
2068 		 * 0xFF and all descriptors look unused (which they are).
2069 		 */
2070 		if (++sc->replypostindex >= sc->pqdepth)
2071 			sc->replypostindex = 0;
2072 
2073 		switch (flags) {
2074 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2075 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2076 			cm->cm_reply = NULL;
2077 			break;
2078 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2079 		{
2080 			uint32_t baddr;
2081 			uint8_t *reply;
2082 
2083 			/*
2084 			 * Re-compose the reply address from the address
2085 			 * sent back from the chip.  The ReplyFrameAddress
2086 			 * is the lower 32 bits of the physical address of
2087 			 * particular reply frame.  Convert that address to
2088 			 * host format, and then use that to provide the
2089 			 * offset against the virtual address base
2090 			 * (sc->reply_frames).
2091 			 */
2092 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2093 			reply = sc->reply_frames +
2094 				(baddr - ((uint32_t)sc->reply_busaddr));
2095 			/*
2096 			 * Make sure the reply we got back is in a valid
2097 			 * range.  If not, go ahead and panic here, since
2098 			 * we'll probably panic as soon as we deference the
2099 			 * reply pointer anyway.
2100 			 */
2101 			if ((reply < sc->reply_frames)
2102 			 || (reply > (sc->reply_frames +
2103 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
2104 				printf("%s: WARNING: reply %p out of range!\n",
2105 				       __func__, reply);
2106 				printf("%s: reply_frames %p, fqdepth %d, "
2107 				       "frame size %d\n", __func__,
2108 				       sc->reply_frames, sc->fqdepth,
2109 				       sc->facts->ReplyFrameSize * 4);
2110 				printf("%s: baddr %#x,\n", __func__, baddr);
2111 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
2112 				panic("Reply address out of range");
2113 			}
2114 			if (le16toh(desc->AddressReply.SMID) == 0) {
2115 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2116 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2117 					/*
2118 					 * If SMID is 0 for Diag Buffer Post,
2119 					 * this implies that the reply is due to
2120 					 * a release function with a status that
2121 					 * the buffer has been released.  Set
2122 					 * the buffer flags accordingly.
2123 					 */
2124 					rel_rep =
2125 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2126 					if ((le16toh(rel_rep->IOCStatus) &
2127 					    MPI2_IOCSTATUS_MASK) ==
2128 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2129 					{
2130 						pBuffer =
2131 						    &sc->fw_diag_buffer_list[
2132 						    rel_rep->BufferType];
2133 						pBuffer->valid_data = TRUE;
2134 						pBuffer->owned_by_firmware =
2135 						    FALSE;
2136 						pBuffer->immediate = FALSE;
2137 					}
2138 				} else
2139 					mps_dispatch_event(sc, baddr,
2140 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2141 					    reply);
2142 			} else {
2143 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2144 				cm->cm_reply = reply;
2145 				cm->cm_reply_data =
2146 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2147 			}
2148 			break;
2149 		}
2150 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2151 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2152 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2153 		default:
2154 			/* Unhandled */
2155 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2156 			    desc->Default.ReplyFlags);
2157 			cm = NULL;
2158 			break;
2159 		}
2160 
2161 
2162 		if (cm != NULL) {
2163 			// Print Error reply frame
2164 			if (cm->cm_reply)
2165 				mps_display_reply_info(sc,cm->cm_reply);
2166 			mps_complete_command(sc, cm);
2167 		}
2168 
2169 		desc->Words.Low = 0xffffffff;
2170 		desc->Words.High = 0xffffffff;
2171 	}
2172 
2173 	if (pq != sc->replypostindex) {
2174 		mps_dprint(sc, MPS_TRACE,
2175 		    "%s sc %p writing postindex %d\n",
2176 		    __func__, sc, sc->replypostindex);
2177 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2178 	}
2179 
2180 	return;
2181 }
2182 
2183 static void
2184 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2185     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2186 {
2187 	struct mps_event_handle *eh;
2188 	int event, handled = 0;
2189 
2190 	event = le16toh(reply->Event);
2191 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2192 		if (isset(eh->mask, event)) {
2193 			eh->callback(sc, data, reply);
2194 			handled++;
2195 		}
2196 	}
2197 
2198 	if (handled == 0)
2199 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2200 
2201 	/*
2202 	 * This is the only place that the event/reply should be freed.
2203 	 * Anything wanting to hold onto the event data should have
2204 	 * already copied it into their own storage.
2205 	 */
2206 	mps_free_reply(sc, data);
2207 }
2208 
2209 static void
2210 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2211 {
2212 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2213 
2214 	if (cm->cm_reply)
2215 		MPS_DPRINT_EVENT(sc, generic,
2216 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2217 
2218 	mps_free_command(sc, cm);
2219 
2220 	/* next, send a port enable */
2221 	mpssas_startup(sc);
2222 }
2223 
2224 /*
2225  * For both register_events and update_events, the caller supplies a bitmap
2226  * of events that it _wants_.  These functions then turn that into a bitmask
2227  * suitable for the controller.
2228  */
2229 int
2230 mps_register_events(struct mps_softc *sc, u32 *mask,
2231     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2232 {
2233 	struct mps_event_handle *eh;
2234 	int error = 0;
2235 
2236 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2237 	if(!eh) {
2238 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2239 		return (ENOMEM);
2240 	}
2241 	eh->callback = cb;
2242 	eh->data = data;
2243 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2244 	if (mask != NULL)
2245 		error = mps_update_events(sc, eh, mask);
2246 	*handle = eh;
2247 
2248 	return (error);
2249 }
2250 
2251 int
2252 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2253     u32 *mask)
2254 {
2255 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2256 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2257 	struct mps_command *cm;
2258 	int error, i;
2259 
2260 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2261 
2262 	if ((mask != NULL) && (handle != NULL))
2263 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2264 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2265 
2266 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2267 		sc->event_mask[i] = -1;
2268 
2269 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2270 		sc->event_mask[i] &= ~handle->mask[i];
2271 
2272 
2273 	if ((cm = mps_alloc_command(sc)) == NULL)
2274 		return (EBUSY);
2275 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2276 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2277 	evtreq->MsgFlags = 0;
2278 	evtreq->SASBroadcastPrimitiveMasks = 0;
2279 #ifdef MPS_DEBUG_ALL_EVENTS
2280 	{
2281 		u_char fullmask[16];
2282 		memset(fullmask, 0x00, 16);
2283 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2284 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2285 	}
2286 #else
2287         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2288                 evtreq->EventMasks[i] =
2289                     htole32(sc->event_mask[i]);
2290 #endif
2291 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2292 	cm->cm_data = NULL;
2293 
2294 	error = mps_wait_command(sc, &cm, 60, 0);
2295 	if (cm != NULL)
2296 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2297 	if ((reply == NULL) ||
2298 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2299 		error = ENXIO;
2300 
2301 	if (reply)
2302 		MPS_DPRINT_EVENT(sc, generic, reply);
2303 
2304 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2305 
2306 	if (cm != NULL)
2307 		mps_free_command(sc, cm);
2308 	return (error);
2309 }
2310 
2311 static int
2312 mps_reregister_events(struct mps_softc *sc)
2313 {
2314 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2315 	struct mps_command *cm;
2316 	struct mps_event_handle *eh;
2317 	int error, i;
2318 
2319 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2320 
2321 	/* first, reregister events */
2322 
2323 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2324 		sc->event_mask[i] = -1;
2325 
2326 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2327 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2328 			sc->event_mask[i] &= ~eh->mask[i];
2329 	}
2330 
2331 	if ((cm = mps_alloc_command(sc)) == NULL)
2332 		return (EBUSY);
2333 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2334 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2335 	evtreq->MsgFlags = 0;
2336 	evtreq->SASBroadcastPrimitiveMasks = 0;
2337 #ifdef MPS_DEBUG_ALL_EVENTS
2338 	{
2339 		u_char fullmask[16];
2340 		memset(fullmask, 0x00, 16);
2341 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2342 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2343 	}
2344 #else
2345         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2346                 evtreq->EventMasks[i] =
2347                     htole32(sc->event_mask[i]);
2348 #endif
2349 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2350 	cm->cm_data = NULL;
2351 	cm->cm_complete = mps_reregister_events_complete;
2352 
2353 	error = mps_map_command(sc, cm);
2354 
2355 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2356 	    error);
2357 	return (error);
2358 }
2359 
2360 void
2361 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2362 {
2363 
2364 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2365 	free(handle, M_MPT2);
2366 }
2367 
2368 /*
2369  * Add a chain element as the next SGE for the specified command.
2370  * Reset cm_sge and cm_sgesize to indicate all the available space.
2371  */
2372 static int
2373 mps_add_chain(struct mps_command *cm)
2374 {
2375 	MPI2_SGE_CHAIN32 *sgc;
2376 	struct mps_chain *chain;
2377 	int space;
2378 
2379 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2380 		panic("MPS: Need SGE Error Code\n");
2381 
2382 	chain = mps_alloc_chain(cm->cm_sc);
2383 	if (chain == NULL)
2384 		return (ENOBUFS);
2385 
2386 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2387 
2388 	/*
2389 	 * Note: a double-linked list is used to make it easier to
2390 	 * walk for debugging.
2391 	 */
2392 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2393 
2394 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2395 	sgc->Length = htole16(space);
2396 	sgc->NextChainOffset = 0;
2397 	/* TODO Looks like bug in Setting sgc->Flags.
2398 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2399 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2400 	 *	This is fine.. because we are not using simple element. In case of
2401 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2402  	 */
2403 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2404 	sgc->Address = htole32(chain->chain_busaddr);
2405 
2406 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2407 	cm->cm_sglsize = space;
2408 	return (0);
2409 }
2410 
2411 /*
2412  * Add one scatter-gather element (chain, simple, transaction context)
2413  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2414  * cm_sge as the remaining size and pointer to the next SGE to fill
2415  * in, respectively.
2416  */
2417 int
2418 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2419 {
2420 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2421 	MPI2_SGE_SIMPLE64 *sge = sgep;
2422 	int error, type;
2423 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2424 
2425 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2426 
2427 #ifdef INVARIANTS
2428 	switch (type) {
2429 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2430 		if (len != tc->DetailsLength + 4)
2431 			panic("TC %p length %u or %zu?", tc,
2432 			    tc->DetailsLength + 4, len);
2433 		}
2434 		break;
2435 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2436 		/* Driver only uses 32-bit chain elements */
2437 		if (len != MPS_SGC_SIZE)
2438 			panic("CHAIN %p length %u or %zu?", sgep,
2439 			    MPS_SGC_SIZE, len);
2440 		break;
2441 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2442 		/* Driver only uses 64-bit SGE simple elements */
2443 		if (len != MPS_SGE64_SIZE)
2444 			panic("SGE simple %p length %u or %zu?", sge,
2445 			    MPS_SGE64_SIZE, len);
2446 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2447 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2448 			panic("SGE simple %p not marked 64-bit?", sge);
2449 
2450 		break;
2451 	default:
2452 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2453 	}
2454 #endif
2455 
2456 	/*
2457 	 * case 1: 1 more segment, enough room for it
2458 	 * case 2: 2 more segments, enough room for both
2459 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2460 	 * case 4: >=1 more segment, enough room for only a chain
2461 	 * case 5: >=1 more segment, no room for anything (error)
2462          */
2463 
2464 	/*
2465 	 * There should be room for at least a chain element, or this
2466 	 * code is buggy.  Case (5).
2467 	 */
2468 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2469 		panic("MPS: Need SGE Error Code\n");
2470 
2471 	if (segsleft >= 2 &&
2472 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2473 		/*
2474 		 * There are 2 or more segments left to add, and only
2475 		 * enough room for 1 and a chain.  Case (3).
2476 		 *
2477 		 * Mark as last element in this chain if necessary.
2478 		 */
2479 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2480 			sge->FlagsLength |= htole32(
2481 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2482 		}
2483 
2484 		/*
2485 		 * Add the item then a chain.  Do the chain now,
2486 		 * rather than on the next iteration, to simplify
2487 		 * understanding the code.
2488 		 */
2489 		cm->cm_sglsize -= len;
2490 		bcopy(sgep, cm->cm_sge, len);
2491 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2492 		return (mps_add_chain(cm));
2493 	}
2494 
2495 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2496 		/*
2497 		 * 1 or more segment, enough room for only a chain.
2498 		 * Hope the previous element wasn't a Simple entry
2499 		 * that needed to be marked with
2500 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2501 		 */
2502 		if ((error = mps_add_chain(cm)) != 0)
2503 			return (error);
2504 	}
2505 
2506 #ifdef INVARIANTS
2507 	/* Case 1: 1 more segment, enough room for it. */
2508 	if (segsleft == 1 && cm->cm_sglsize < len)
2509 		panic("1 seg left and no room? %u versus %zu",
2510 		    cm->cm_sglsize, len);
2511 
2512 	/* Case 2: 2 more segments, enough room for both */
2513 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2514 		panic("2 segs left and no room? %u versus %zu",
2515 		    cm->cm_sglsize, len);
2516 #endif
2517 
2518 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2519 		/*
2520 		 * If this is a bi-directional request, need to account for that
2521 		 * here.  Save the pre-filled sge values.  These will be used
2522 		 * either for the 2nd SGL or for a single direction SGL.  If
2523 		 * cm_out_len is non-zero, this is a bi-directional request, so
2524 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2525 		 * fill in the IN SGL.  Note that at this time, when filling in
2526 		 * 2 SGL's for a bi-directional request, they both use the same
2527 		 * DMA buffer (same cm command).
2528 		 */
2529 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2530 		saved_address_low = sge->Address.Low;
2531 		saved_address_high = sge->Address.High;
2532 		if (cm->cm_out_len) {
2533 			sge->FlagsLength = htole32(cm->cm_out_len |
2534 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2535 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2536 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2537 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2538 			    MPI2_SGE_FLAGS_SHIFT));
2539 			cm->cm_sglsize -= len;
2540 			bcopy(sgep, cm->cm_sge, len);
2541 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2542 			    + len);
2543 		}
2544 		saved_buf_len |=
2545 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2546 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2547 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2548 		    MPI2_SGE_FLAGS_END_OF_LIST |
2549 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2550 		    MPI2_SGE_FLAGS_SHIFT);
2551 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2552 			saved_buf_len |=
2553 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2554 			    MPI2_SGE_FLAGS_SHIFT);
2555 		} else {
2556 			saved_buf_len |=
2557 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2558 			    MPI2_SGE_FLAGS_SHIFT);
2559 		}
2560 		sge->FlagsLength = htole32(saved_buf_len);
2561 		sge->Address.Low = saved_address_low;
2562 		sge->Address.High = saved_address_high;
2563 	}
2564 
2565 	cm->cm_sglsize -= len;
2566 	bcopy(sgep, cm->cm_sge, len);
2567 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2568 	return (0);
2569 }
2570 
2571 /*
2572  * Add one dma segment to the scatter-gather list for a command.
2573  */
2574 int
2575 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2576     int segsleft)
2577 {
2578 	MPI2_SGE_SIMPLE64 sge;
2579 
2580 	/*
2581 	 * This driver always uses 64-bit address elements for simplicity.
2582 	 */
2583 	bzero(&sge, sizeof(sge));
2584 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2585 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2586 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2587 	mps_from_u64(pa, &sge.Address);
2588 
2589 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2590 }
2591 
2592 static void
2593 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2594 {
2595 	struct mps_softc *sc;
2596 	struct mps_command *cm;
2597 	u_int i, dir, sflags;
2598 
2599 	cm = (struct mps_command *)arg;
2600 	sc = cm->cm_sc;
2601 
2602 	/*
2603 	 * In this case, just print out a warning and let the chip tell the
2604 	 * user they did the wrong thing.
2605 	 */
2606 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2607 		mps_dprint(sc, MPS_ERROR,
2608 			   "%s: warning: busdma returned %d segments, "
2609 			   "more than the %d allowed\n", __func__, nsegs,
2610 			   cm->cm_max_segs);
2611 	}
2612 
2613 	/*
2614 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2615 	 * here.  In that case, both direction flags will be set.
2616 	 */
2617 	sflags = 0;
2618 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2619 		/*
2620 		 * We have to add a special case for SMP passthrough, there
2621 		 * is no easy way to generically handle it.  The first
2622 		 * S/G element is used for the command (therefore the
2623 		 * direction bit needs to be set).  The second one is used
2624 		 * for the reply.  We'll leave it to the caller to make
2625 		 * sure we only have two buffers.
2626 		 */
2627 		/*
2628 		 * Even though the busdma man page says it doesn't make
2629 		 * sense to have both direction flags, it does in this case.
2630 		 * We have one s/g element being accessed in each direction.
2631 		 */
2632 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2633 
2634 		/*
2635 		 * Set the direction flag on the first buffer in the SMP
2636 		 * passthrough request.  We'll clear it for the second one.
2637 		 */
2638 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2639 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2640 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2641 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2642 		dir = BUS_DMASYNC_PREWRITE;
2643 	} else
2644 		dir = BUS_DMASYNC_PREREAD;
2645 
2646 	for (i = 0; i < nsegs; i++) {
2647 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2648 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2649 		}
2650 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2651 		    sflags, nsegs - i);
2652 		if (error != 0) {
2653 			/* Resource shortage, roll back! */
2654 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2655 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2656 				    "consider increasing hw.mps.max_chains.\n");
2657 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2658 			mps_complete_command(sc, cm);
2659 			return;
2660 		}
2661 	}
2662 
2663 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2664 	mps_enqueue_request(sc, cm);
2665 
2666 	return;
2667 }
2668 
2669 static void
2670 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2671 	     int error)
2672 {
2673 	mps_data_cb(arg, segs, nsegs, error);
2674 }
2675 
2676 /*
2677  * This is the routine to enqueue commands ansynchronously.
2678  * Note that the only error path here is from bus_dmamap_load(), which can
2679  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2680  * assumed that if you have a command in-hand, then you have enough credits
2681  * to use it.
2682  */
2683 int
2684 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2685 {
2686 	int error = 0;
2687 
2688 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2689 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2690 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2691 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2692 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2693 		    cm->cm_data, mps_data_cb, cm, 0);
2694 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2695 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2696 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2697 	} else {
2698 		/* Add a zero-length element as needed */
2699 		if (cm->cm_sge != NULL)
2700 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2701 		mps_enqueue_request(sc, cm);
2702 	}
2703 
2704 	return (error);
2705 }
2706 
2707 /*
2708  * This is the routine to enqueue commands synchronously.  An error of
2709  * EINPROGRESS from mps_map_command() is ignored since the command will
2710  * be executed and enqueued automatically.  Other errors come from msleep().
2711  */
2712 int
2713 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2714     int sleep_flag)
2715 {
2716 	int error, rc;
2717 	struct timeval cur_time, start_time;
2718 	struct mps_command *cm = *cmp;
2719 
2720 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2721 		return  EBUSY;
2722 
2723 	cm->cm_complete = NULL;
2724 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2725 	error = mps_map_command(sc, cm);
2726 	if ((error != 0) && (error != EINPROGRESS))
2727 		return (error);
2728 
2729 	/*
2730 	 * Check for context and wait for 50 mSec at a time until time has
2731 	 * expired or the command has finished.  If msleep can't be used, need
2732 	 * to poll.
2733 	 */
2734 	if (curthread->td_no_sleeping != 0)
2735 		sleep_flag = NO_SLEEP;
2736 	getmicrouptime(&start_time);
2737 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2738 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2739 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2740 		if (error == EWOULDBLOCK) {
2741 			/*
2742 			 * Record the actual elapsed time in the case of a
2743 			 * timeout for the message below.
2744 			 */
2745 			getmicrouptime(&cur_time);
2746 			timevalsub(&cur_time, &start_time);
2747 		}
2748 	} else {
2749 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2750 			mps_intr_locked(sc);
2751 			if (sleep_flag == CAN_SLEEP)
2752 				pause("mpswait", hz/20);
2753 			else
2754 				DELAY(50000);
2755 
2756 			getmicrouptime(&cur_time);
2757 			timevalsub(&cur_time, &start_time);
2758 			if (cur_time.tv_sec > timeout) {
2759 				error = EWOULDBLOCK;
2760 				break;
2761 			}
2762 		}
2763 	}
2764 
2765 	if (error == EWOULDBLOCK) {
2766 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2767 		    " elapsed=%jd\n", __func__, timeout,
2768 		    (intmax_t)cur_time.tv_sec);
2769 		rc = mps_reinit(sc);
2770 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2771 		    "failed");
2772 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
2773 			/*
2774 			 * Tell the caller that we freed the command in a
2775 			 * reinit.
2776 			 */
2777 			*cmp = NULL;
2778 		}
2779 		error = ETIMEDOUT;
2780 	}
2781 	return (error);
2782 }
2783 
2784 /*
2785  * The MPT driver had a verbose interface for config pages.  In this driver,
2786  * reduce it to much simpler terms, similar to the Linux driver.
2787  */
2788 int
2789 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2790 {
2791 	MPI2_CONFIG_REQUEST *req;
2792 	struct mps_command *cm;
2793 	int error;
2794 
2795 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2796 		return (EBUSY);
2797 	}
2798 
2799 	cm = mps_alloc_command(sc);
2800 	if (cm == NULL) {
2801 		return (EBUSY);
2802 	}
2803 
2804 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2805 	req->Function = MPI2_FUNCTION_CONFIG;
2806 	req->Action = params->action;
2807 	req->SGLFlags = 0;
2808 	req->ChainOffset = 0;
2809 	req->PageAddress = params->page_address;
2810 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2811 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2812 
2813 		hdr = &params->hdr.Ext;
2814 		req->ExtPageType = hdr->ExtPageType;
2815 		req->ExtPageLength = hdr->ExtPageLength;
2816 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2817 		req->Header.PageLength = 0; /* Must be set to zero */
2818 		req->Header.PageNumber = hdr->PageNumber;
2819 		req->Header.PageVersion = hdr->PageVersion;
2820 	} else {
2821 		MPI2_CONFIG_PAGE_HEADER *hdr;
2822 
2823 		hdr = &params->hdr.Struct;
2824 		req->Header.PageType = hdr->PageType;
2825 		req->Header.PageNumber = hdr->PageNumber;
2826 		req->Header.PageLength = hdr->PageLength;
2827 		req->Header.PageVersion = hdr->PageVersion;
2828 	}
2829 
2830 	cm->cm_data = params->buffer;
2831 	cm->cm_length = params->length;
2832 	if (cm->cm_data != NULL) {
2833 		cm->cm_sge = &req->PageBufferSGE;
2834 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2835 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2836 	} else
2837 		cm->cm_sge = NULL;
2838 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2839 
2840 	cm->cm_complete_data = params;
2841 	if (params->callback != NULL) {
2842 		cm->cm_complete = mps_config_complete;
2843 		return (mps_map_command(sc, cm));
2844 	} else {
2845 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
2846 		if (error) {
2847 			mps_dprint(sc, MPS_FAULT,
2848 			    "Error %d reading config page\n", error);
2849 			if (cm != NULL)
2850 				mps_free_command(sc, cm);
2851 			return (error);
2852 		}
2853 		mps_config_complete(sc, cm);
2854 	}
2855 
2856 	return (0);
2857 }
2858 
2859 int
2860 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2861 {
2862 	return (EINVAL);
2863 }
2864 
2865 static void
2866 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
2867 {
2868 	MPI2_CONFIG_REPLY *reply;
2869 	struct mps_config_params *params;
2870 
2871 	MPS_FUNCTRACE(sc);
2872 	params = cm->cm_complete_data;
2873 
2874 	if (cm->cm_data != NULL) {
2875 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2876 		    BUS_DMASYNC_POSTREAD);
2877 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2878 	}
2879 
2880 	/*
2881 	 * XXX KDM need to do more error recovery?  This results in the
2882 	 * device in question not getting probed.
2883 	 */
2884 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
2885 		params->status = MPI2_IOCSTATUS_BUSY;
2886 		goto done;
2887 	}
2888 
2889 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2890 	if (reply == NULL) {
2891 		params->status = MPI2_IOCSTATUS_BUSY;
2892 		goto done;
2893 	}
2894 	params->status = reply->IOCStatus;
2895 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2896 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2897 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2898 		params->hdr.Ext.PageType = reply->Header.PageType;
2899 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
2900 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
2901 	} else {
2902 		params->hdr.Struct.PageType = reply->Header.PageType;
2903 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2904 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2905 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2906 	}
2907 
2908 done:
2909 	mps_free_command(sc, cm);
2910 	if (params->callback != NULL)
2911 		params->callback(sc, params);
2912 
2913 	return;
2914 }
2915