xref: /freebsd/sys/dev/mps/mps.c (revision 38069501)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2015 Avago Technologies
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
29  *
30  * $FreeBSD$
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 /* Communications core for Avago Technologies (LSI) MPT2 */
37 
38 /* TODO Move headers to mpsvar */
39 #include <sys/types.h>
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/selinfo.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/module.h>
47 #include <sys/bus.h>
48 #include <sys/conf.h>
49 #include <sys/bio.h>
50 #include <sys/malloc.h>
51 #include <sys/uio.h>
52 #include <sys/sysctl.h>
53 #include <sys/smp.h>
54 #include <sys/queue.h>
55 #include <sys/kthread.h>
56 #include <sys/taskqueue.h>
57 #include <sys/endian.h>
58 #include <sys/eventhandler.h>
59 #include <sys/sbuf.h>
60 
61 #include <machine/bus.h>
62 #include <machine/resource.h>
63 #include <sys/rman.h>
64 #include <sys/proc.h>
65 
66 #include <dev/pci/pcivar.h>
67 
68 #include <cam/cam.h>
69 #include <cam/scsi/scsi_all.h>
70 
71 #include <dev/mps/mpi/mpi2_type.h>
72 #include <dev/mps/mpi/mpi2.h>
73 #include <dev/mps/mpi/mpi2_ioc.h>
74 #include <dev/mps/mpi/mpi2_sas.h>
75 #include <dev/mps/mpi/mpi2_cnfg.h>
76 #include <dev/mps/mpi/mpi2_init.h>
77 #include <dev/mps/mpi/mpi2_tool.h>
78 #include <dev/mps/mps_ioctl.h>
79 #include <dev/mps/mpsvar.h>
80 #include <dev/mps/mps_table.h>
81 
82 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
83 static int mps_init_queues(struct mps_softc *sc);
84 static void mps_resize_queues(struct mps_softc *sc);
85 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
86 static int mps_transition_operational(struct mps_softc *sc);
87 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
88 static void mps_iocfacts_free(struct mps_softc *sc);
89 static void mps_startup(void *arg);
90 static int mps_send_iocinit(struct mps_softc *sc);
91 static int mps_alloc_queues(struct mps_softc *sc);
92 static int mps_alloc_hw_queues(struct mps_softc *sc);
93 static int mps_alloc_replies(struct mps_softc *sc);
94 static int mps_alloc_requests(struct mps_softc *sc);
95 static int mps_attach_log(struct mps_softc *sc);
96 static __inline void mps_complete_command(struct mps_softc *sc,
97     struct mps_command *cm);
98 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
99     MPI2_EVENT_NOTIFICATION_REPLY *reply);
100 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
101 static void mps_periodic(void *);
102 static int mps_reregister_events(struct mps_softc *sc);
103 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
104 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
105 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
106 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
107 static void mps_parse_debug(struct mps_softc *sc, char *list);
108 
109 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
110 
111 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
112 
113 /*
114  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
115  * any state and back to its initialization state machine.
116  */
117 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
118 
119 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
120  * Compiler only support unint64_t to be passed as argument.
121  * Otherwise it will throw below error
122  * "aggregate value used where an integer was expected"
123  */
124 
125 typedef union _reply_descriptor {
126         u64 word;
127         struct {
128                 u32 low;
129                 u32 high;
130         } u;
131 }reply_descriptor,address_descriptor;
132 
133 /* Rate limit chain-fail messages to 1 per minute */
134 static struct timeval mps_chainfail_interval = { 60, 0 };
135 
136 /*
137  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
138  * If this function is called from process context, it can sleep
139  * and there is no harm to sleep, in case if this fuction is called
140  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
141  * based on sleep flags driver will call either msleep, pause or DELAY.
142  * msleep and pause are of same variant, but pause is used when mps_mtx
143  * is not hold by driver.
144  *
145  */
146 static int
147 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
148 {
149 	uint32_t reg;
150 	int i, error, tries = 0;
151 	uint8_t first_wait_done = FALSE;
152 
153 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
154 
155 	/* Clear any pending interrupts */
156 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
157 
158 	/*
159 	 * Force NO_SLEEP for threads prohibited to sleep
160  	 * e.a Thread from interrupt handler are prohibited to sleep.
161  	 */
162 	if (curthread->td_no_sleeping != 0)
163 		sleep_flag = NO_SLEEP;
164 
165 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
166 
167 	/* Push the magic sequence */
168 	error = ETIMEDOUT;
169 	while (tries++ < 20) {
170 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
171 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
172 			    mpt2_reset_magic[i]);
173 		/* wait 100 msec */
174 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
175 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
176 			    "mpsdiag", hz/10);
177 		else if (sleep_flag == CAN_SLEEP)
178 			pause("mpsdiag", hz/10);
179 		else
180 			DELAY(100 * 1000);
181 
182 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
183 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
184 			error = 0;
185 			break;
186 		}
187 	}
188 	if (error) {
189 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
190 		    error);
191 		return (error);
192 	}
193 
194 	/* Send the actual reset.  XXX need to refresh the reg? */
195 	reg |= MPI2_DIAG_RESET_ADAPTER;
196 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
197 		reg);
198 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
199 
200 	/* Wait up to 300 seconds in 50ms intervals */
201 	error = ETIMEDOUT;
202 	for (i = 0; i < 6000; i++) {
203 		/*
204 		 * Wait 50 msec. If this is the first time through, wait 256
205 		 * msec to satisfy Diag Reset timing requirements.
206 		 */
207 		if (first_wait_done) {
208 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
209 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
210 				    "mpsdiag", hz/20);
211 			else if (sleep_flag == CAN_SLEEP)
212 				pause("mpsdiag", hz/20);
213 			else
214 				DELAY(50 * 1000);
215 		} else {
216 			DELAY(256 * 1000);
217 			first_wait_done = TRUE;
218 		}
219 		/*
220 		 * Check for the RESET_ADAPTER bit to be cleared first, then
221 		 * wait for the RESET state to be cleared, which takes a little
222 		 * longer.
223 		 */
224 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
225 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
226 			continue;
227 		}
228 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
229 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
230 			error = 0;
231 			break;
232 		}
233 	}
234 	if (error) {
235 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
236 		    error);
237 		return (error);
238 	}
239 
240 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
241 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
242 
243 	return (0);
244 }
245 
246 static int
247 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
248 {
249 	int error;
250 
251 	MPS_FUNCTRACE(sc);
252 
253 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
254 
255 	error = 0;
256 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
257 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
258 	    MPI2_DOORBELL_FUNCTION_SHIFT);
259 
260 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
261 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
262 		    "Doorbell handshake failed\n");
263 		error = ETIMEDOUT;
264 	}
265 
266 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
267 	return (error);
268 }
269 
270 static int
271 mps_transition_ready(struct mps_softc *sc)
272 {
273 	uint32_t reg, state;
274 	int error, tries = 0;
275 	int sleep_flags;
276 
277 	MPS_FUNCTRACE(sc);
278 	/* If we are in attach call, do not sleep */
279 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
280 					? CAN_SLEEP:NO_SLEEP;
281 	error = 0;
282 
283 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
284 	   __func__, sleep_flags);
285 
286 	while (tries++ < 1200) {
287 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
288 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
289 
290 		/*
291 		 * Ensure the IOC is ready to talk.  If it's not, try
292 		 * resetting it.
293 		 */
294 		if (reg & MPI2_DOORBELL_USED) {
295 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
296 			    "reset\n");
297 			mps_diag_reset(sc, sleep_flags);
298 			DELAY(50000);
299 			continue;
300 		}
301 
302 		/* Is the adapter owned by another peer? */
303 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
304 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
305 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
306 			    "control of another peer host, aborting "
307 			    "initialization.\n");
308 			error = ENXIO;
309 			break;
310 		}
311 
312 		state = reg & MPI2_IOC_STATE_MASK;
313 		if (state == MPI2_IOC_STATE_READY) {
314 			/* Ready to go! */
315 			error = 0;
316 			break;
317 		} else if (state == MPI2_IOC_STATE_FAULT) {
318 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
319 			    "state 0x%x, resetting\n",
320 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
321 			mps_diag_reset(sc, sleep_flags);
322 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
323 			/* Need to take ownership */
324 			mps_message_unit_reset(sc, sleep_flags);
325 		} else if (state == MPI2_IOC_STATE_RESET) {
326 			/* Wait a bit, IOC might be in transition */
327 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
328 			    "IOC in unexpected reset state\n");
329 		} else {
330 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
331 			    "IOC in unknown state 0x%x\n", state);
332 			error = EINVAL;
333 			break;
334 		}
335 
336 		/* Wait 50ms for things to settle down. */
337 		DELAY(50000);
338 	}
339 
340 	if (error)
341 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
342 		    "Cannot transition IOC to ready\n");
343 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
344 
345 	return (error);
346 }
347 
348 static int
349 mps_transition_operational(struct mps_softc *sc)
350 {
351 	uint32_t reg, state;
352 	int error;
353 
354 	MPS_FUNCTRACE(sc);
355 
356 	error = 0;
357 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
358 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
359 
360 	state = reg & MPI2_IOC_STATE_MASK;
361 	if (state != MPI2_IOC_STATE_READY) {
362 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
363 		if ((error = mps_transition_ready(sc)) != 0) {
364 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
365 			    "failed to transition ready, exit\n");
366 			return (error);
367 		}
368 	}
369 
370 	error = mps_send_iocinit(sc);
371 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
372 
373 	return (error);
374 }
375 
376 static void
377 mps_resize_queues(struct mps_softc *sc)
378 {
379 	int reqcr, prireqcr;
380 
381 	/*
382 	 * Size the queues. Since the reply queues always need one free
383 	 * entry, we'll deduct one reply message here.  The LSI documents
384 	 * suggest instead to add a count to the request queue, but I think
385 	 * that it's better to deduct from reply queue.
386 	 */
387 	prireqcr = MAX(1, sc->max_prireqframes);
388 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
389 
390 	reqcr = MAX(2, sc->max_reqframes);
391 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
392 
393 	sc->num_reqs = prireqcr + reqcr;
394 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
395 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
396 
397 	/*
398 	 * Figure out the number of MSIx-based queues.  If the firmware or
399 	 * user has done something crazy and not allowed enough credit for
400 	 * the queues to be useful then don't enable multi-queue.
401 	 */
402 	if (sc->facts->MaxMSIxVectors < 2)
403 		sc->msi_msgs = 1;
404 
405 	if (sc->msi_msgs > 1) {
406 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
407 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
408 		if (sc->num_reqs / sc->msi_msgs < 2)
409 			sc->msi_msgs = 1;
410 	}
411 
412 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
413 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
414 }
415 
416 /*
417  * This is called during attach and when re-initializing due to a Diag Reset.
418  * IOC Facts is used to allocate many of the structures needed by the driver.
419  * If called from attach, de-allocation is not required because the driver has
420  * not allocated any structures yet, but if called from a Diag Reset, previously
421  * allocated structures based on IOC Facts will need to be freed and re-
422  * allocated bases on the latest IOC Facts.
423  */
424 static int
425 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
426 {
427 	int error;
428 	Mpi2IOCFactsReply_t saved_facts;
429 	uint8_t saved_mode, reallocating;
430 
431 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
432 
433 	/* Save old IOC Facts and then only reallocate if Facts have changed */
434 	if (!attaching) {
435 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
436 	}
437 
438 	/*
439 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
440 	 * a re-initialization and only return the error if attaching so the OS
441 	 * can handle it.
442 	 */
443 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
444 		if (attaching) {
445 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
446 			    "IOC Facts with error %d, exit\n", error);
447 			return (error);
448 		} else {
449 			panic("%s failed to get IOC Facts with error %d\n",
450 			    __func__, error);
451 		}
452 	}
453 
454 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
455 
456 	snprintf(sc->fw_version, sizeof(sc->fw_version),
457 	    "%02d.%02d.%02d.%02d",
458 	    sc->facts->FWVersion.Struct.Major,
459 	    sc->facts->FWVersion.Struct.Minor,
460 	    sc->facts->FWVersion.Struct.Unit,
461 	    sc->facts->FWVersion.Struct.Dev);
462 
463 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
464 	    MPS_DRIVER_VERSION);
465 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
466 	     sc->facts->IOCCapabilities,
467 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
468 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
469 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
470 
471 	/*
472 	 * If the chip doesn't support event replay then a hard reset will be
473 	 * required to trigger a full discovery.  Do the reset here then
474 	 * retransition to Ready.  A hard reset might have already been done,
475 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
476 	 * for a Diag Reset.
477 	 */
478 	if (attaching && ((sc->facts->IOCCapabilities &
479 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
480 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
481 		mps_diag_reset(sc, NO_SLEEP);
482 		if ((error = mps_transition_ready(sc)) != 0) {
483 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
484 			    "transition to ready with error %d, exit\n",
485 			    error);
486 			return (error);
487 		}
488 	}
489 
490 	/*
491 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
492 	 * changed from the previous IOC Facts, log a warning, but only if
493 	 * checking this after a Diag Reset and not during attach.
494 	 */
495 	saved_mode = sc->ir_firmware;
496 	if (sc->facts->IOCCapabilities &
497 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
498 		sc->ir_firmware = 1;
499 	if (!attaching) {
500 		if (sc->ir_firmware != saved_mode) {
501 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
502 			    "in IOC Facts does not match previous mode\n");
503 		}
504 	}
505 
506 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
507 	reallocating = FALSE;
508 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
509 
510 	if ((!attaching) &&
511 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
512 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
513 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
514 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
515 	    (saved_facts.ProductID != sc->facts->ProductID) ||
516 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
517 	    (saved_facts.IOCRequestFrameSize !=
518 	    sc->facts->IOCRequestFrameSize) ||
519 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
520 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
521 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
522 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
523 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
524 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
525 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
526 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
527 	    (saved_facts.MaxPersistentEntries !=
528 	    sc->facts->MaxPersistentEntries))) {
529 		reallocating = TRUE;
530 
531 		/* Record that we reallocated everything */
532 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
533 	}
534 
535 	/*
536 	 * Some things should be done if attaching or re-allocating after a Diag
537 	 * Reset, but are not needed after a Diag Reset if the FW has not
538 	 * changed.
539 	 */
540 	if (attaching || reallocating) {
541 		/*
542 		 * Check if controller supports FW diag buffers and set flag to
543 		 * enable each type.
544 		 */
545 		if (sc->facts->IOCCapabilities &
546 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
547 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
548 			    enabled = TRUE;
549 		if (sc->facts->IOCCapabilities &
550 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
551 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
552 			    enabled = TRUE;
553 		if (sc->facts->IOCCapabilities &
554 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
555 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
556 			    enabled = TRUE;
557 
558 		/*
559 		 * Set flag if EEDP is supported and if TLR is supported.
560 		 */
561 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
562 			sc->eedp_enabled = TRUE;
563 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
564 			sc->control_TLR = TRUE;
565 
566 		mps_resize_queues(sc);
567 
568 		/*
569 		 * Initialize all Tail Queues
570 		 */
571 		TAILQ_INIT(&sc->req_list);
572 		TAILQ_INIT(&sc->high_priority_req_list);
573 		TAILQ_INIT(&sc->chain_list);
574 		TAILQ_INIT(&sc->tm_list);
575 	}
576 
577 	/*
578 	 * If doing a Diag Reset and the FW is significantly different
579 	 * (reallocating will be set above in IOC Facts comparison), then all
580 	 * buffers based on the IOC Facts will need to be freed before they are
581 	 * reallocated.
582 	 */
583 	if (reallocating) {
584 		mps_iocfacts_free(sc);
585 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
586 		    saved_facts.MaxVolumes);
587 	}
588 
589 	/*
590 	 * Any deallocation has been completed.  Now start reallocating
591 	 * if needed.  Will only need to reallocate if attaching or if the new
592 	 * IOC Facts are different from the previous IOC Facts after a Diag
593 	 * Reset. Targets have already been allocated above if needed.
594 	 */
595 	error = 0;
596 	while (attaching || reallocating) {
597 		if ((error = mps_alloc_hw_queues(sc)) != 0)
598 			break;
599 		if ((error = mps_alloc_replies(sc)) != 0)
600 			break;
601 		if ((error = mps_alloc_requests(sc)) != 0)
602 			break;
603 		if ((error = mps_alloc_queues(sc)) != 0)
604 			break;
605 
606 		break;
607 	}
608 	if (error) {
609 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
610 		    "Failed to alloc queues with error %d\n", error);
611 		mps_free(sc);
612 		return (error);
613 	}
614 
615 	/* Always initialize the queues */
616 	bzero(sc->free_queue, sc->fqdepth * 4);
617 	mps_init_queues(sc);
618 
619 	/*
620 	 * Always get the chip out of the reset state, but only panic if not
621 	 * attaching.  If attaching and there is an error, that is handled by
622 	 * the OS.
623 	 */
624 	error = mps_transition_operational(sc);
625 	if (error != 0) {
626 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
627 		    "transition to operational with error %d\n", error);
628 		mps_free(sc);
629 		return (error);
630 	}
631 
632 	/*
633 	 * Finish the queue initialization.
634 	 * These are set here instead of in mps_init_queues() because the
635 	 * IOC resets these values during the state transition in
636 	 * mps_transition_operational().  The free index is set to 1
637 	 * because the corresponding index in the IOC is set to 0, and the
638 	 * IOC treats the queues as full if both are set to the same value.
639 	 * Hence the reason that the queue can't hold all of the possible
640 	 * replies.
641 	 */
642 	sc->replypostindex = 0;
643 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
644 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
645 
646 	/*
647 	 * Attach the subsystems so they can prepare their event masks.
648 	 * XXX Should be dynamic so that IM/IR and user modules can attach
649 	 */
650 	error = 0;
651 	while (attaching) {
652 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
653 		if ((error = mps_attach_log(sc)) != 0)
654 			break;
655 		if ((error = mps_attach_sas(sc)) != 0)
656 			break;
657 		if ((error = mps_attach_user(sc)) != 0)
658 			break;
659 		break;
660 	}
661 	if (error) {
662 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
663 		    "subsystems: error %d\n", error);
664 		mps_free(sc);
665 		return (error);
666 	}
667 
668 	if ((error = mps_pci_setup_interrupts(sc)) != 0) {
669 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
670 		    "interrupts\n");
671 		mps_free(sc);
672 		return (error);
673 	}
674 
675 	/*
676 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
677 	 * reset it after a Diag Reset, just in case.
678 	 */
679 	sc->WD_available = FALSE;
680 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
681 		sc->WD_available = TRUE;
682 
683 	return (error);
684 }
685 
686 /*
687  * This is called if memory is being free (during detach for example) and when
688  * buffers need to be reallocated due to a Diag Reset.
689  */
690 static void
691 mps_iocfacts_free(struct mps_softc *sc)
692 {
693 	struct mps_command *cm;
694 	int i;
695 
696 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
697 
698 	if (sc->free_busaddr != 0)
699 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
700 	if (sc->free_queue != NULL)
701 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
702 		    sc->queues_map);
703 	if (sc->queues_dmat != NULL)
704 		bus_dma_tag_destroy(sc->queues_dmat);
705 
706 	if (sc->chain_busaddr != 0)
707 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
708 	if (sc->chain_frames != NULL)
709 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
710 		    sc->chain_map);
711 	if (sc->chain_dmat != NULL)
712 		bus_dma_tag_destroy(sc->chain_dmat);
713 
714 	if (sc->sense_busaddr != 0)
715 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
716 	if (sc->sense_frames != NULL)
717 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
718 		    sc->sense_map);
719 	if (sc->sense_dmat != NULL)
720 		bus_dma_tag_destroy(sc->sense_dmat);
721 
722 	if (sc->reply_busaddr != 0)
723 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
724 	if (sc->reply_frames != NULL)
725 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
726 		    sc->reply_map);
727 	if (sc->reply_dmat != NULL)
728 		bus_dma_tag_destroy(sc->reply_dmat);
729 
730 	if (sc->req_busaddr != 0)
731 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
732 	if (sc->req_frames != NULL)
733 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
734 	if (sc->req_dmat != NULL)
735 		bus_dma_tag_destroy(sc->req_dmat);
736 
737 	if (sc->chains != NULL)
738 		free(sc->chains, M_MPT2);
739 	if (sc->commands != NULL) {
740 		for (i = 1; i < sc->num_reqs; i++) {
741 			cm = &sc->commands[i];
742 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
743 		}
744 		free(sc->commands, M_MPT2);
745 	}
746 	if (sc->buffer_dmat != NULL)
747 		bus_dma_tag_destroy(sc->buffer_dmat);
748 
749 	mps_pci_free_interrupts(sc);
750 	free(sc->queues, M_MPT2);
751 	sc->queues = NULL;
752 }
753 
754 /*
755  * The terms diag reset and hard reset are used interchangeably in the MPI
756  * docs to mean resetting the controller chip.  In this code diag reset
757  * cleans everything up, and the hard reset function just sends the reset
758  * sequence to the chip.  This should probably be refactored so that every
759  * subsystem gets a reset notification of some sort, and can clean up
760  * appropriately.
761  */
762 int
763 mps_reinit(struct mps_softc *sc)
764 {
765 	int error;
766 	struct mpssas_softc *sassc;
767 
768 	sassc = sc->sassc;
769 
770 	MPS_FUNCTRACE(sc);
771 
772 	mtx_assert(&sc->mps_mtx, MA_OWNED);
773 
774 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
775 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
776 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
777 		return 0;
778 	}
779 
780 	/* make sure the completion callbacks can recognize they're getting
781 	 * a NULL cm_reply due to a reset.
782 	 */
783 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
784 
785 	/*
786 	 * Mask interrupts here.
787 	 */
788 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
789 	mps_mask_intr(sc);
790 
791 	error = mps_diag_reset(sc, CAN_SLEEP);
792 	if (error != 0) {
793 		/* XXXSL No need to panic here */
794 		panic("%s hard reset failed with error %d\n",
795 		    __func__, error);
796 	}
797 
798 	/* Restore the PCI state, including the MSI-X registers */
799 	mps_pci_restore(sc);
800 
801 	/* Give the I/O subsystem special priority to get itself prepared */
802 	mpssas_handle_reinit(sc);
803 
804 	/*
805 	 * Get IOC Facts and allocate all structures based on this information.
806 	 * The attach function will also call mps_iocfacts_allocate at startup.
807 	 * If relevant values have changed in IOC Facts, this function will free
808 	 * all of the memory based on IOC Facts and reallocate that memory.
809 	 */
810 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
811 		panic("%s IOC Facts based allocation failed with error %d\n",
812 		    __func__, error);
813 	}
814 
815 	/*
816 	 * Mapping structures will be re-allocated after getting IOC Page8, so
817 	 * free these structures here.
818 	 */
819 	mps_mapping_exit(sc);
820 
821 	/*
822 	 * The static page function currently read is IOC Page8.  Others can be
823 	 * added in future.  It's possible that the values in IOC Page8 have
824 	 * changed after a Diag Reset due to user modification, so always read
825 	 * these.  Interrupts are masked, so unmask them before getting config
826 	 * pages.
827 	 */
828 	mps_unmask_intr(sc);
829 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
830 	mps_base_static_config_pages(sc);
831 
832 	/*
833 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
834 	 * mapping tables.
835 	 */
836 	mps_mapping_initialize(sc);
837 
838 	/*
839 	 * Restart will reload the event masks clobbered by the reset, and
840 	 * then enable the port.
841 	 */
842 	mps_reregister_events(sc);
843 
844 	/* the end of discovery will release the simq, so we're done. */
845 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
846 	    sc, sc->replypostindex, sc->replyfreeindex);
847 
848 	mpssas_release_simq_reinit(sassc);
849 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
850 
851 	return 0;
852 }
853 
854 /* Wait for the chip to ACK a word that we've put into its FIFO
855  * Wait for <timeout> seconds. In single loop wait for busy loop
856  * for 500 microseconds.
857  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
858  * */
859 static int
860 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
861 {
862 
863 	u32 cntdn, count;
864 	u32 int_status;
865 	u32 doorbell;
866 
867 	count = 0;
868 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
869 	do {
870 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
871 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
872 			mps_dprint(sc, MPS_TRACE,
873 			"%s: successful count(%d), timeout(%d)\n",
874 			__func__, count, timeout);
875 		return 0;
876 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
877 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
878 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
879 				MPI2_IOC_STATE_FAULT) {
880 				mps_dprint(sc, MPS_FAULT,
881 					"fault_state(0x%04x)!\n", doorbell);
882 				return (EFAULT);
883 			}
884 		} else if (int_status == 0xFFFFFFFF)
885 			goto out;
886 
887 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
888 		* 0.5 milisecond */
889 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
890 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
891 			"mpsdba", hz/1000);
892 		else if (sleep_flag == CAN_SLEEP)
893 			pause("mpsdba", hz/1000);
894 		else
895 			DELAY(500);
896 		count++;
897 	} while (--cntdn);
898 
899 	out:
900 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
901 		"int_status(%x)!\n", __func__, count, int_status);
902 	return (ETIMEDOUT);
903 
904 }
905 
906 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
907 static int
908 mps_wait_db_int(struct mps_softc *sc)
909 {
910 	int retry;
911 
912 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
913 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
914 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
915 			return (0);
916 		DELAY(2000);
917 	}
918 	return (ETIMEDOUT);
919 }
920 
921 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
922 static int
923 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
924     int req_sz, int reply_sz, int timeout)
925 {
926 	uint32_t *data32;
927 	uint16_t *data16;
928 	int i, count, ioc_sz, residual;
929 	int sleep_flags = CAN_SLEEP;
930 
931 	if (curthread->td_no_sleeping != 0)
932 		sleep_flags = NO_SLEEP;
933 
934 	/* Step 1 */
935 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
936 
937 	/* Step 2 */
938 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
939 		return (EBUSY);
940 
941 	/* Step 3
942 	 * Announce that a message is coming through the doorbell.  Messages
943 	 * are pushed at 32bit words, so round up if needed.
944 	 */
945 	count = (req_sz + 3) / 4;
946 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
947 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
948 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
949 
950 	/* Step 4 */
951 	if (mps_wait_db_int(sc) ||
952 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
953 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
954 		return (ENXIO);
955 	}
956 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
957 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
958 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
959 		return (ENXIO);
960 	}
961 
962 	/* Step 5 */
963 	/* Clock out the message data synchronously in 32-bit dwords*/
964 	data32 = (uint32_t *)req;
965 	for (i = 0; i < count; i++) {
966 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
967 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
968 			mps_dprint(sc, MPS_FAULT,
969 			    "Timeout while writing doorbell\n");
970 			return (ENXIO);
971 		}
972 	}
973 
974 	/* Step 6 */
975 	/* Clock in the reply in 16-bit words.  The total length of the
976 	 * message is always in the 4th byte, so clock out the first 2 words
977 	 * manually, then loop the rest.
978 	 */
979 	data16 = (uint16_t *)reply;
980 	if (mps_wait_db_int(sc) != 0) {
981 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
982 		return (ENXIO);
983 	}
984 	data16[0] =
985 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
986 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
987 	if (mps_wait_db_int(sc) != 0) {
988 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
989 		return (ENXIO);
990 	}
991 	data16[1] =
992 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
993 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
994 
995 	/* Number of 32bit words in the message */
996 	ioc_sz = reply->MsgLength;
997 
998 	/*
999 	 * Figure out how many 16bit words to clock in without overrunning.
1000 	 * The precision loss with dividing reply_sz can safely be
1001 	 * ignored because the messages can only be multiples of 32bits.
1002 	 */
1003 	residual = 0;
1004 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1005 	if (count < ioc_sz * 2) {
1006 		residual = ioc_sz * 2 - count;
1007 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1008 		    "residual message words\n", residual);
1009 	}
1010 
1011 	for (i = 2; i < count; i++) {
1012 		if (mps_wait_db_int(sc) != 0) {
1013 			mps_dprint(sc, MPS_FAULT,
1014 			    "Timeout reading doorbell %d\n", i);
1015 			return (ENXIO);
1016 		}
1017 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1018 		    MPI2_DOORBELL_DATA_MASK;
1019 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1020 	}
1021 
1022 	/*
1023 	 * Pull out residual words that won't fit into the provided buffer.
1024 	 * This keeps the chip from hanging due to a driver programming
1025 	 * error.
1026 	 */
1027 	while (residual--) {
1028 		if (mps_wait_db_int(sc) != 0) {
1029 			mps_dprint(sc, MPS_FAULT,
1030 			    "Timeout reading doorbell\n");
1031 			return (ENXIO);
1032 		}
1033 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1034 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1035 	}
1036 
1037 	/* Step 7 */
1038 	if (mps_wait_db_int(sc) != 0) {
1039 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1040 		return (ENXIO);
1041 	}
1042 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1043 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1044 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1045 
1046 	return (0);
1047 }
1048 
1049 static void
1050 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1051 {
1052 	reply_descriptor rd;
1053 	MPS_FUNCTRACE(sc);
1054 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1055 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1056 
1057 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1058 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1059 
1060 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1061 		sc->io_cmds_highwater++;
1062 	rd.u.low = cm->cm_desc.Words.Low;
1063 	rd.u.high = cm->cm_desc.Words.High;
1064 	rd.word = htole64(rd.word);
1065 	/* TODO-We may need to make below regwrite atomic */
1066 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1067 	    rd.u.low);
1068 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1069 	    rd.u.high);
1070 }
1071 
1072 /*
1073  * Just the FACTS, ma'am.
1074  */
1075 static int
1076 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1077 {
1078 	MPI2_DEFAULT_REPLY *reply;
1079 	MPI2_IOC_FACTS_REQUEST request;
1080 	int error, req_sz, reply_sz;
1081 
1082 	MPS_FUNCTRACE(sc);
1083 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1084 
1085 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1086 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1087 	reply = (MPI2_DEFAULT_REPLY *)facts;
1088 
1089 	bzero(&request, req_sz);
1090 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1091 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1092 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1093 
1094 	return (error);
1095 }
1096 
1097 static int
1098 mps_send_iocinit(struct mps_softc *sc)
1099 {
1100 	MPI2_IOC_INIT_REQUEST	init;
1101 	MPI2_DEFAULT_REPLY	reply;
1102 	int req_sz, reply_sz, error;
1103 	struct timeval now;
1104 	uint64_t time_in_msec;
1105 
1106 	MPS_FUNCTRACE(sc);
1107 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1108 
1109 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1110 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1111 	bzero(&init, req_sz);
1112 	bzero(&reply, reply_sz);
1113 
1114 	/*
1115 	 * Fill in the init block.  Note that most addresses are
1116 	 * deliberately in the lower 32bits of memory.  This is a micro-
1117 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1118 	 */
1119 	init.Function = MPI2_FUNCTION_IOC_INIT;
1120 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1121 	init.MsgVersion = htole16(MPI2_VERSION);
1122 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1123 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1124 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1125 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1126 	init.SenseBufferAddressHigh = 0;
1127 	init.SystemReplyAddressHigh = 0;
1128 	init.SystemRequestFrameBaseAddress.High = 0;
1129 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1130 	init.ReplyDescriptorPostQueueAddress.High = 0;
1131 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1132 	init.ReplyFreeQueueAddress.High = 0;
1133 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1134 	getmicrotime(&now);
1135 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1136 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1137 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1138 
1139 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1140 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1141 		error = ENXIO;
1142 
1143 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1144 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1145 	return (error);
1146 }
1147 
1148 void
1149 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1150 {
1151 	bus_addr_t *addr;
1152 
1153 	addr = arg;
1154 	*addr = segs[0].ds_addr;
1155 }
1156 
1157 static int
1158 mps_alloc_queues(struct mps_softc *sc)
1159 {
1160 	struct mps_queue *q;
1161 	int nq, i;
1162 
1163 	nq = sc->msi_msgs;
1164 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1165 
1166 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1167 	    M_NOWAIT|M_ZERO);
1168 	if (sc->queues == NULL)
1169 		return (ENOMEM);
1170 
1171 	for (i = 0; i < nq; i++) {
1172 		q = &sc->queues[i];
1173 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1174 		q->sc = sc;
1175 		q->qnum = i;
1176 	}
1177 
1178 	return (0);
1179 }
1180 
1181 static int
1182 mps_alloc_hw_queues(struct mps_softc *sc)
1183 {
1184 	bus_addr_t queues_busaddr;
1185 	uint8_t *queues;
1186 	int qsize, fqsize, pqsize;
1187 
1188 	/*
1189 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1190 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1191 	 * This queue supplies fresh reply frames for the firmware to use.
1192 	 *
1193 	 * The reply descriptor post queue contains 8 byte entries in
1194 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1195 	 * contains filled-in reply frames sent from the firmware to the host.
1196 	 *
1197 	 * These two queues are allocated together for simplicity.
1198 	 */
1199 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1200 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1201 	fqsize= sc->fqdepth * 4;
1202 	pqsize = sc->pqdepth * 8;
1203 	qsize = fqsize + pqsize;
1204 
1205         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1206 				16, 0,			/* algnmnt, boundary */
1207 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1208 				BUS_SPACE_MAXADDR,	/* highaddr */
1209 				NULL, NULL,		/* filter, filterarg */
1210                                 qsize,			/* maxsize */
1211                                 1,			/* nsegments */
1212                                 qsize,			/* maxsegsize */
1213                                 0,			/* flags */
1214                                 NULL, NULL,		/* lockfunc, lockarg */
1215                                 &sc->queues_dmat)) {
1216 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1217 		return (ENOMEM);
1218         }
1219         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1220 	    &sc->queues_map)) {
1221 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1222 		return (ENOMEM);
1223         }
1224         bzero(queues, qsize);
1225         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1226 	    mps_memaddr_cb, &queues_busaddr, 0);
1227 
1228 	sc->free_queue = (uint32_t *)queues;
1229 	sc->free_busaddr = queues_busaddr;
1230 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1231 	sc->post_busaddr = queues_busaddr + fqsize;
1232 
1233 	return (0);
1234 }
1235 
1236 static int
1237 mps_alloc_replies(struct mps_softc *sc)
1238 {
1239 	int rsize, num_replies;
1240 
1241 	/*
1242 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1243 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1244 	 * replies can be used at once.
1245 	 */
1246 	num_replies = max(sc->fqdepth, sc->num_replies);
1247 
1248 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1249         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1250 				4, 0,			/* algnmnt, boundary */
1251 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1252 				BUS_SPACE_MAXADDR,	/* highaddr */
1253 				NULL, NULL,		/* filter, filterarg */
1254                                 rsize,			/* maxsize */
1255                                 1,			/* nsegments */
1256                                 rsize,			/* maxsegsize */
1257                                 0,			/* flags */
1258                                 NULL, NULL,		/* lockfunc, lockarg */
1259                                 &sc->reply_dmat)) {
1260 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1261 		return (ENOMEM);
1262         }
1263         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1264 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1265 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1266 		return (ENOMEM);
1267         }
1268         bzero(sc->reply_frames, rsize);
1269         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1270 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1271 
1272 	return (0);
1273 }
1274 
1275 static int
1276 mps_alloc_requests(struct mps_softc *sc)
1277 {
1278 	struct mps_command *cm;
1279 	struct mps_chain *chain;
1280 	int i, rsize, nsegs;
1281 
1282 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1283         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1284 				16, 0,			/* algnmnt, boundary */
1285 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1286 				BUS_SPACE_MAXADDR,	/* highaddr */
1287 				NULL, NULL,		/* filter, filterarg */
1288                                 rsize,			/* maxsize */
1289                                 1,			/* nsegments */
1290                                 rsize,			/* maxsegsize */
1291                                 0,			/* flags */
1292                                 NULL, NULL,		/* lockfunc, lockarg */
1293                                 &sc->req_dmat)) {
1294 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1295 		return (ENOMEM);
1296         }
1297         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1298 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1299 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1300 		return (ENOMEM);
1301         }
1302         bzero(sc->req_frames, rsize);
1303         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1304 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1305 
1306 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1307         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1308 				16, 0,			/* algnmnt, boundary */
1309 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1310 				BUS_SPACE_MAXADDR,	/* highaddr */
1311 				NULL, NULL,		/* filter, filterarg */
1312                                 rsize,			/* maxsize */
1313                                 1,			/* nsegments */
1314                                 rsize,			/* maxsegsize */
1315                                 0,			/* flags */
1316                                 NULL, NULL,		/* lockfunc, lockarg */
1317                                 &sc->chain_dmat)) {
1318 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1319 		return (ENOMEM);
1320         }
1321         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1322 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1323 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1324 		return (ENOMEM);
1325         }
1326         bzero(sc->chain_frames, rsize);
1327         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1328 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1329 
1330 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1331         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1332 				1, 0,			/* algnmnt, boundary */
1333 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1334 				BUS_SPACE_MAXADDR,	/* highaddr */
1335 				NULL, NULL,		/* filter, filterarg */
1336                                 rsize,			/* maxsize */
1337                                 1,			/* nsegments */
1338                                 rsize,			/* maxsegsize */
1339                                 0,			/* flags */
1340                                 NULL, NULL,		/* lockfunc, lockarg */
1341                                 &sc->sense_dmat)) {
1342 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1343 		return (ENOMEM);
1344         }
1345         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1346 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1347 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1348 		return (ENOMEM);
1349         }
1350         bzero(sc->sense_frames, rsize);
1351         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1352 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1353 
1354 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1355 	    M_WAITOK | M_ZERO);
1356 	if(!sc->chains) {
1357 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1358 		return (ENOMEM);
1359 	}
1360 	for (i = 0; i < sc->max_chains; i++) {
1361 		chain = &sc->chains[i];
1362 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1363 		    i * sc->facts->IOCRequestFrameSize * 4);
1364 		chain->chain_busaddr = sc->chain_busaddr +
1365 		    i * sc->facts->IOCRequestFrameSize * 4;
1366 		mps_free_chain(sc, chain);
1367 		sc->chain_free_lowwater++;
1368 	}
1369 
1370 	/* XXX Need to pick a more precise value */
1371 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1372         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1373 				1, 0,			/* algnmnt, boundary */
1374 				BUS_SPACE_MAXADDR,	/* lowaddr */
1375 				BUS_SPACE_MAXADDR,	/* highaddr */
1376 				NULL, NULL,		/* filter, filterarg */
1377                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1378                                 nsegs,			/* nsegments */
1379                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1380                                 BUS_DMA_ALLOCNOW,	/* flags */
1381                                 busdma_lock_mutex,	/* lockfunc */
1382 				&sc->mps_mtx,		/* lockarg */
1383                                 &sc->buffer_dmat)) {
1384 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1385 		return (ENOMEM);
1386         }
1387 
1388 	/*
1389 	 * SMID 0 cannot be used as a free command per the firmware spec.
1390 	 * Just drop that command instead of risking accounting bugs.
1391 	 */
1392 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1393 	    M_MPT2, M_WAITOK | M_ZERO);
1394 	if(!sc->commands) {
1395 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1396 		return (ENOMEM);
1397 	}
1398 	for (i = 1; i < sc->num_reqs; i++) {
1399 		cm = &sc->commands[i];
1400 		cm->cm_req = sc->req_frames +
1401 		    i * sc->facts->IOCRequestFrameSize * 4;
1402 		cm->cm_req_busaddr = sc->req_busaddr +
1403 		    i * sc->facts->IOCRequestFrameSize * 4;
1404 		cm->cm_sense = &sc->sense_frames[i];
1405 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1406 		cm->cm_desc.Default.SMID = i;
1407 		cm->cm_sc = sc;
1408 		TAILQ_INIT(&cm->cm_chain_list);
1409 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1410 
1411 		/* XXX Is a failure here a critical problem? */
1412 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1413 			if (i <= sc->facts->HighPriorityCredit)
1414 				mps_free_high_priority_command(sc, cm);
1415 			else
1416 				mps_free_command(sc, cm);
1417 		else {
1418 			panic("failed to allocate command %d\n", i);
1419 			sc->num_reqs = i;
1420 			break;
1421 		}
1422 	}
1423 
1424 	return (0);
1425 }
1426 
1427 static int
1428 mps_init_queues(struct mps_softc *sc)
1429 {
1430 	int i;
1431 
1432 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1433 
1434 	/*
1435 	 * According to the spec, we need to use one less reply than we
1436 	 * have space for on the queue.  So sc->num_replies (the number we
1437 	 * use) should be less than sc->fqdepth (allocated size).
1438 	 */
1439 	if (sc->num_replies >= sc->fqdepth)
1440 		return (EINVAL);
1441 
1442 	/*
1443 	 * Initialize all of the free queue entries.
1444 	 */
1445 	for (i = 0; i < sc->fqdepth; i++)
1446 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1447 	sc->replyfreeindex = sc->num_replies;
1448 
1449 	return (0);
1450 }
1451 
1452 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1453  * Next are the global settings, if they exist.  Highest are the per-unit
1454  * settings, if they exist.
1455  */
1456 void
1457 mps_get_tunables(struct mps_softc *sc)
1458 {
1459 	char tmpstr[80], mps_debug[80];
1460 
1461 	/* XXX default to some debugging for now */
1462 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1463 	sc->disable_msix = 0;
1464 	sc->disable_msi = 0;
1465 	sc->max_msix = MPS_MSIX_MAX;
1466 	sc->max_chains = MPS_CHAIN_FRAMES;
1467 	sc->max_io_pages = MPS_MAXIO_PAGES;
1468 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1469 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1470 	sc->use_phynum = 1;
1471 	sc->max_reqframes = MPS_REQ_FRAMES;
1472 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1473 	sc->max_replyframes = MPS_REPLY_FRAMES;
1474 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1475 
1476 	/*
1477 	 * Grab the global variables.
1478 	 */
1479 	bzero(mps_debug, 80);
1480 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1481 		mps_parse_debug(sc, mps_debug);
1482 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1483 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1484 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1485 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1486 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1487 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1488 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1489 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1490 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1491 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1492 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1493 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1494 
1495 	/* Grab the unit-instance variables */
1496 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1497 	    device_get_unit(sc->mps_dev));
1498 	bzero(mps_debug, 80);
1499 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1500 		mps_parse_debug(sc, mps_debug);
1501 
1502 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1503 	    device_get_unit(sc->mps_dev));
1504 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1505 
1506 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1507 	    device_get_unit(sc->mps_dev));
1508 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1509 
1510 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1511 	    device_get_unit(sc->mps_dev));
1512 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1513 
1514 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1515 	    device_get_unit(sc->mps_dev));
1516 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1517 
1518 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1519 	    device_get_unit(sc->mps_dev));
1520 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1521 
1522 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1523 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1524 	    device_get_unit(sc->mps_dev));
1525 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1526 
1527 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1528 	    device_get_unit(sc->mps_dev));
1529 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1530 
1531 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1532 	    device_get_unit(sc->mps_dev));
1533 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1534 
1535 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1536 	    device_get_unit(sc->mps_dev));
1537 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1538 
1539 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1540 	    device_get_unit(sc->mps_dev));
1541 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1542 
1543 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1544 	    device_get_unit(sc->mps_dev));
1545 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1546 
1547 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1548 	    device_get_unit(sc->mps_dev));
1549 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1550 
1551 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1552 	    device_get_unit(sc->mps_dev));
1553 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1554 
1555 }
1556 
1557 static void
1558 mps_setup_sysctl(struct mps_softc *sc)
1559 {
1560 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1561 	struct sysctl_oid	*sysctl_tree = NULL;
1562 	char tmpstr[80], tmpstr2[80];
1563 
1564 	/*
1565 	 * Setup the sysctl variable so the user can change the debug level
1566 	 * on the fly.
1567 	 */
1568 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1569 	    device_get_unit(sc->mps_dev));
1570 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1571 
1572 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1573 	if (sysctl_ctx != NULL)
1574 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1575 
1576 	if (sysctl_tree == NULL) {
1577 		sysctl_ctx_init(&sc->sysctl_ctx);
1578 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1579 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1580 		    CTLFLAG_RD, 0, tmpstr);
1581 		if (sc->sysctl_tree == NULL)
1582 			return;
1583 		sysctl_ctx = &sc->sysctl_ctx;
1584 		sysctl_tree = sc->sysctl_tree;
1585 	}
1586 
1587 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1588 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1589 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1590 
1591 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1592 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1593 	    "Disable the use of MSI-X interrupts");
1594 
1595 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1596 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1597 	    "Disable the use of MSI interrupts");
1598 
1599 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1600 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1601 	    "User-defined maximum number of MSIX queues");
1602 
1603 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1604 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1605 	    "Negotiated number of MSIX queues");
1606 
1607 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1608 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1609 	    "Total number of allocated request frames");
1610 
1611 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1612 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1613 	    "Total number of allocated high priority request frames");
1614 
1615 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1616 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1617 	    "Total number of allocated reply frames");
1618 
1619 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1620 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1621 	    "Total number of event frames allocated");
1622 
1623 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1624 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1625 	    strlen(sc->fw_version), "firmware version");
1626 
1627 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1628 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1629 	    strlen(MPS_DRIVER_VERSION), "driver version");
1630 
1631 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1632 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1633 	    &sc->io_cmds_active, 0, "number of currently active commands");
1634 
1635 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1636 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1637 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1638 
1639 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1640 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1641 	    &sc->chain_free, 0, "number of free chain elements");
1642 
1643 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1644 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1645 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1646 
1647 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1648 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1649 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1650 
1651 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1652 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1653 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1654 	    "IOCFacts)");
1655 
1656 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1657 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1658 	    "enable SSU to SATA SSD/HDD at shutdown");
1659 
1660 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1661 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1662 	    &sc->chain_alloc_fail, "chain allocation failures");
1663 
1664 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1665 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1666 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1667 	    "spinup after SATA ID error");
1668 
1669 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1670 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1671 	    mps_mapping_dump, "A", "Mapping Table Dump");
1672 
1673 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1674 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1675 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1676 
1677 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1678 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1679 	    "Use the phy number for enumeration");
1680 }
1681 
1682 static struct mps_debug_string {
1683 	char	*name;
1684 	int	flag;
1685 } mps_debug_strings[] = {
1686 	{"info", MPS_INFO},
1687 	{"fault", MPS_FAULT},
1688 	{"event", MPS_EVENT},
1689 	{"log", MPS_LOG},
1690 	{"recovery", MPS_RECOVERY},
1691 	{"error", MPS_ERROR},
1692 	{"init", MPS_INIT},
1693 	{"xinfo", MPS_XINFO},
1694 	{"user", MPS_USER},
1695 	{"mapping", MPS_MAPPING},
1696 	{"trace", MPS_TRACE}
1697 };
1698 
1699 enum mps_debug_level_combiner {
1700 	COMB_NONE,
1701 	COMB_ADD,
1702 	COMB_SUB
1703 };
1704 
1705 static int
1706 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1707 {
1708 	struct mps_softc *sc;
1709 	struct mps_debug_string *string;
1710 	struct sbuf *sbuf;
1711 	char *buffer;
1712 	size_t sz;
1713 	int i, len, debug, error;
1714 
1715 	sc = (struct mps_softc *)arg1;
1716 
1717 	error = sysctl_wire_old_buffer(req, 0);
1718 	if (error != 0)
1719 		return (error);
1720 
1721 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1722 	debug = sc->mps_debug;
1723 
1724 	sbuf_printf(sbuf, "%#x", debug);
1725 
1726 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1727 	for (i = 0; i < sz; i++) {
1728 		string = &mps_debug_strings[i];
1729 		if (debug & string->flag)
1730 			sbuf_printf(sbuf, ",%s", string->name);
1731 	}
1732 
1733 	error = sbuf_finish(sbuf);
1734 	sbuf_delete(sbuf);
1735 
1736 	if (error || req->newptr == NULL)
1737 		return (error);
1738 
1739 	len = req->newlen - req->newidx;
1740 	if (len == 0)
1741 		return (0);
1742 
1743 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1744 	error = SYSCTL_IN(req, buffer, len);
1745 
1746 	mps_parse_debug(sc, buffer);
1747 
1748 	free(buffer, M_MPT2);
1749 	return (error);
1750 }
1751 
1752 static void
1753 mps_parse_debug(struct mps_softc *sc, char *list)
1754 {
1755 	struct mps_debug_string *string;
1756 	enum mps_debug_level_combiner op;
1757 	char *token, *endtoken;
1758 	size_t sz;
1759 	int flags, i;
1760 
1761 	if (list == NULL || *list == '\0')
1762 		return;
1763 
1764 	if (*list == '+') {
1765 		op = COMB_ADD;
1766 		list++;
1767 	} else if (*list == '-') {
1768 		op = COMB_SUB;
1769 		list++;
1770 	} else
1771 		op = COMB_NONE;
1772 	if (*list == '\0')
1773 		return;
1774 
1775 	flags = 0;
1776 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1777 	while ((token = strsep(&list, ":,")) != NULL) {
1778 
1779 		/* Handle integer flags */
1780 		flags |= strtol(token, &endtoken, 0);
1781 		if (token != endtoken)
1782 			continue;
1783 
1784 		/* Handle text flags */
1785 		for (i = 0; i < sz; i++) {
1786 			string = &mps_debug_strings[i];
1787 			if (strcasecmp(token, string->name) == 0) {
1788 				flags |= string->flag;
1789 				break;
1790 			}
1791 		}
1792 	}
1793 
1794 	switch (op) {
1795 	case COMB_NONE:
1796 		sc->mps_debug = flags;
1797 		break;
1798 	case COMB_ADD:
1799 		sc->mps_debug |= flags;
1800 		break;
1801 	case COMB_SUB:
1802 		sc->mps_debug &= (~flags);
1803 		break;
1804 	}
1805 
1806 	return;
1807 }
1808 
1809 int
1810 mps_attach(struct mps_softc *sc)
1811 {
1812 	int error;
1813 
1814 	MPS_FUNCTRACE(sc);
1815 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1816 
1817 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1818 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1819 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1820 	TAILQ_INIT(&sc->event_list);
1821 	timevalclear(&sc->lastfail);
1822 
1823 	if ((error = mps_transition_ready(sc)) != 0) {
1824 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1825 		    "ready\n");
1826 		return (error);
1827 	}
1828 
1829 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1830 	    M_ZERO|M_NOWAIT);
1831 	if(!sc->facts) {
1832 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1833 		    "exit\n");
1834 		return (ENOMEM);
1835 	}
1836 
1837 	/*
1838 	 * Get IOC Facts and allocate all structures based on this information.
1839 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1840 	 * Facts. If relevant values have changed in IOC Facts, this function
1841 	 * will free all of the memory based on IOC Facts and reallocate that
1842 	 * memory.  If this fails, any allocated memory should already be freed.
1843 	 */
1844 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1845 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1846 		    "failed with error %d, exit\n", error);
1847 		return (error);
1848 	}
1849 
1850 	/* Start the periodic watchdog check on the IOC Doorbell */
1851 	mps_periodic(sc);
1852 
1853 	/*
1854 	 * The portenable will kick off discovery events that will drive the
1855 	 * rest of the initialization process.  The CAM/SAS module will
1856 	 * hold up the boot sequence until discovery is complete.
1857 	 */
1858 	sc->mps_ich.ich_func = mps_startup;
1859 	sc->mps_ich.ich_arg = sc;
1860 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1861 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1862 		    "Cannot establish MPS config hook\n");
1863 		error = EINVAL;
1864 	}
1865 
1866 	/*
1867 	 * Allow IR to shutdown gracefully when shutdown occurs.
1868 	 */
1869 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1870 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1871 
1872 	if (sc->shutdown_eh == NULL)
1873 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1874 		    "shutdown event registration failed\n");
1875 
1876 	mps_setup_sysctl(sc);
1877 
1878 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1879 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1880 
1881 	return (error);
1882 }
1883 
1884 /* Run through any late-start handlers. */
1885 static void
1886 mps_startup(void *arg)
1887 {
1888 	struct mps_softc *sc;
1889 
1890 	sc = (struct mps_softc *)arg;
1891 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1892 
1893 	mps_lock(sc);
1894 	mps_unmask_intr(sc);
1895 
1896 	/* initialize device mapping tables */
1897 	mps_base_static_config_pages(sc);
1898 	mps_mapping_initialize(sc);
1899 	mpssas_startup(sc);
1900 	mps_unlock(sc);
1901 
1902 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
1903 	config_intrhook_disestablish(&sc->mps_ich);
1904 	sc->mps_ich.ich_arg = NULL;
1905 
1906 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1907 }
1908 
1909 /* Periodic watchdog.  Is called with the driver lock already held. */
1910 static void
1911 mps_periodic(void *arg)
1912 {
1913 	struct mps_softc *sc;
1914 	uint32_t db;
1915 
1916 	sc = (struct mps_softc *)arg;
1917 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1918 		return;
1919 
1920 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1921 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1922 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1923 		mps_reinit(sc);
1924 	}
1925 
1926 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1927 }
1928 
1929 static void
1930 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1931     MPI2_EVENT_NOTIFICATION_REPLY *event)
1932 {
1933 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1934 
1935 	MPS_DPRINT_EVENT(sc, generic, event);
1936 
1937 	switch (event->Event) {
1938 	case MPI2_EVENT_LOG_DATA:
1939 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1940 		if (sc->mps_debug & MPS_EVENT)
1941 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1942 		break;
1943 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1944 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1945 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1946 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1947 		     entry->LogSequence);
1948 		break;
1949 	default:
1950 		break;
1951 	}
1952 	return;
1953 }
1954 
1955 static int
1956 mps_attach_log(struct mps_softc *sc)
1957 {
1958 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1959 
1960 	bzero(events, 16);
1961 	setbit(events, MPI2_EVENT_LOG_DATA);
1962 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1963 
1964 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1965 	    &sc->mps_log_eh);
1966 
1967 	return (0);
1968 }
1969 
1970 static int
1971 mps_detach_log(struct mps_softc *sc)
1972 {
1973 
1974 	if (sc->mps_log_eh != NULL)
1975 		mps_deregister_events(sc, sc->mps_log_eh);
1976 	return (0);
1977 }
1978 
1979 /*
1980  * Free all of the driver resources and detach submodules.  Should be called
1981  * without the lock held.
1982  */
1983 int
1984 mps_free(struct mps_softc *sc)
1985 {
1986 	int error;
1987 
1988 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1989 	/* Turn off the watchdog */
1990 	mps_lock(sc);
1991 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1992 	mps_unlock(sc);
1993 	/* Lock must not be held for this */
1994 	callout_drain(&sc->periodic);
1995 	callout_drain(&sc->device_check_callout);
1996 
1997 	if (((error = mps_detach_log(sc)) != 0) ||
1998 	    ((error = mps_detach_sas(sc)) != 0)) {
1999 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2000 		    "subsystems, exit\n");
2001 		return (error);
2002 	}
2003 
2004 	mps_detach_user(sc);
2005 
2006 	/* Put the IOC back in the READY state. */
2007 	mps_lock(sc);
2008 	if ((error = mps_transition_ready(sc)) != 0) {
2009 		mps_unlock(sc);
2010 		return (error);
2011 	}
2012 	mps_unlock(sc);
2013 
2014 	if (sc->facts != NULL)
2015 		free(sc->facts, M_MPT2);
2016 
2017 	/*
2018 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2019 	 * to free these buffers too.
2020 	 */
2021 	mps_iocfacts_free(sc);
2022 
2023 	if (sc->sysctl_tree != NULL)
2024 		sysctl_ctx_free(&sc->sysctl_ctx);
2025 
2026 	/* Deregister the shutdown function */
2027 	if (sc->shutdown_eh != NULL)
2028 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2029 
2030 	mtx_destroy(&sc->mps_mtx);
2031 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2032 
2033 	return (0);
2034 }
2035 
2036 static __inline void
2037 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2038 {
2039 	MPS_FUNCTRACE(sc);
2040 
2041 	if (cm == NULL) {
2042 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2043 		return;
2044 	}
2045 
2046 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2047 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2048 
2049 	if (cm->cm_complete != NULL) {
2050 		mps_dprint(sc, MPS_TRACE,
2051 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2052 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2053 			   cm->cm_reply);
2054 		cm->cm_complete(sc, cm);
2055 	}
2056 
2057 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2058 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2059 		wakeup(cm);
2060 	}
2061 
2062 	if (cm->cm_sc->io_cmds_active != 0) {
2063 		cm->cm_sc->io_cmds_active--;
2064 	} else {
2065 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2066 		    "out of sync - resynching to 0\n");
2067 	}
2068 }
2069 
2070 
2071 static void
2072 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2073 {
2074 	union loginfo_type {
2075 		u32     loginfo;
2076 		struct {
2077 			u32     subcode:16;
2078 			u32     code:8;
2079 			u32     originator:4;
2080 			u32     bus_type:4;
2081 		} dw;
2082 	};
2083 	union loginfo_type sas_loginfo;
2084 	char *originator_str = NULL;
2085 
2086 	sas_loginfo.loginfo = log_info;
2087 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2088 		return;
2089 
2090 	/* each nexus loss loginfo */
2091 	if (log_info == 0x31170000)
2092 		return;
2093 
2094 	/* eat the loginfos associated with task aborts */
2095 	if ((log_info == 30050000 || log_info ==
2096 	    0x31140000 || log_info == 0x31130000))
2097 		return;
2098 
2099 	switch (sas_loginfo.dw.originator) {
2100 	case 0:
2101 		originator_str = "IOP";
2102 		break;
2103 	case 1:
2104 		originator_str = "PL";
2105 		break;
2106 	case 2:
2107 		originator_str = "IR";
2108 		break;
2109 }
2110 
2111 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2112 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2113 	originator_str, sas_loginfo.dw.code,
2114 	sas_loginfo.dw.subcode);
2115 }
2116 
2117 static void
2118 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2119 {
2120 	MPI2DefaultReply_t *mpi_reply;
2121 	u16 sc_status;
2122 
2123 	mpi_reply = (MPI2DefaultReply_t*)reply;
2124 	sc_status = le16toh(mpi_reply->IOCStatus);
2125 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2126 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2127 }
2128 void
2129 mps_intr(void *data)
2130 {
2131 	struct mps_softc *sc;
2132 	uint32_t status;
2133 
2134 	sc = (struct mps_softc *)data;
2135 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2136 
2137 	/*
2138 	 * Check interrupt status register to flush the bus.  This is
2139 	 * needed for both INTx interrupts and driver-driven polling
2140 	 */
2141 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2142 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2143 		return;
2144 
2145 	mps_lock(sc);
2146 	mps_intr_locked(data);
2147 	mps_unlock(sc);
2148 	return;
2149 }
2150 
2151 /*
2152  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2153  * chip.  Hopefully this theory is correct.
2154  */
2155 void
2156 mps_intr_msi(void *data)
2157 {
2158 	struct mps_softc *sc;
2159 
2160 	sc = (struct mps_softc *)data;
2161 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2162 	mps_lock(sc);
2163 	mps_intr_locked(data);
2164 	mps_unlock(sc);
2165 	return;
2166 }
2167 
2168 /*
2169  * The locking is overly broad and simplistic, but easy to deal with for now.
2170  */
2171 void
2172 mps_intr_locked(void *data)
2173 {
2174 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2175 	struct mps_softc *sc;
2176 	struct mps_command *cm = NULL;
2177 	uint8_t flags;
2178 	u_int pq;
2179 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2180 	mps_fw_diagnostic_buffer_t *pBuffer;
2181 
2182 	sc = (struct mps_softc *)data;
2183 
2184 	pq = sc->replypostindex;
2185 	mps_dprint(sc, MPS_TRACE,
2186 	    "%s sc %p starting with replypostindex %u\n",
2187 	    __func__, sc, sc->replypostindex);
2188 
2189 	for ( ;; ) {
2190 		cm = NULL;
2191 		desc = &sc->post_queue[sc->replypostindex];
2192 		flags = desc->Default.ReplyFlags &
2193 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2194 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2195 		 || (le32toh(desc->Words.High) == 0xffffffff))
2196 			break;
2197 
2198 		/* increment the replypostindex now, so that event handlers
2199 		 * and cm completion handlers which decide to do a diag
2200 		 * reset can zero it without it getting incremented again
2201 		 * afterwards, and we break out of this loop on the next
2202 		 * iteration since the reply post queue has been cleared to
2203 		 * 0xFF and all descriptors look unused (which they are).
2204 		 */
2205 		if (++sc->replypostindex >= sc->pqdepth)
2206 			sc->replypostindex = 0;
2207 
2208 		switch (flags) {
2209 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2210 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2211 			cm->cm_reply = NULL;
2212 			break;
2213 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2214 		{
2215 			uint32_t baddr;
2216 			uint8_t *reply;
2217 
2218 			/*
2219 			 * Re-compose the reply address from the address
2220 			 * sent back from the chip.  The ReplyFrameAddress
2221 			 * is the lower 32 bits of the physical address of
2222 			 * particular reply frame.  Convert that address to
2223 			 * host format, and then use that to provide the
2224 			 * offset against the virtual address base
2225 			 * (sc->reply_frames).
2226 			 */
2227 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2228 			reply = sc->reply_frames +
2229 				(baddr - ((uint32_t)sc->reply_busaddr));
2230 			/*
2231 			 * Make sure the reply we got back is in a valid
2232 			 * range.  If not, go ahead and panic here, since
2233 			 * we'll probably panic as soon as we deference the
2234 			 * reply pointer anyway.
2235 			 */
2236 			if ((reply < sc->reply_frames)
2237 			 || (reply > (sc->reply_frames +
2238 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
2239 				printf("%s: WARNING: reply %p out of range!\n",
2240 				       __func__, reply);
2241 				printf("%s: reply_frames %p, fqdepth %d, "
2242 				       "frame size %d\n", __func__,
2243 				       sc->reply_frames, sc->fqdepth,
2244 				       sc->facts->ReplyFrameSize * 4);
2245 				printf("%s: baddr %#x,\n", __func__, baddr);
2246 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
2247 				panic("Reply address out of range");
2248 			}
2249 			if (le16toh(desc->AddressReply.SMID) == 0) {
2250 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2251 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2252 					/*
2253 					 * If SMID is 0 for Diag Buffer Post,
2254 					 * this implies that the reply is due to
2255 					 * a release function with a status that
2256 					 * the buffer has been released.  Set
2257 					 * the buffer flags accordingly.
2258 					 */
2259 					rel_rep =
2260 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2261 					if ((le16toh(rel_rep->IOCStatus) &
2262 					    MPI2_IOCSTATUS_MASK) ==
2263 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2264 					{
2265 						pBuffer =
2266 						    &sc->fw_diag_buffer_list[
2267 						    rel_rep->BufferType];
2268 						pBuffer->valid_data = TRUE;
2269 						pBuffer->owned_by_firmware =
2270 						    FALSE;
2271 						pBuffer->immediate = FALSE;
2272 					}
2273 				} else
2274 					mps_dispatch_event(sc, baddr,
2275 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2276 					    reply);
2277 			} else {
2278 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2279 				cm->cm_reply = reply;
2280 				cm->cm_reply_data =
2281 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2282 			}
2283 			break;
2284 		}
2285 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2286 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2287 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2288 		default:
2289 			/* Unhandled */
2290 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2291 			    desc->Default.ReplyFlags);
2292 			cm = NULL;
2293 			break;
2294 		}
2295 
2296 
2297 		if (cm != NULL) {
2298 			// Print Error reply frame
2299 			if (cm->cm_reply)
2300 				mps_display_reply_info(sc,cm->cm_reply);
2301 			mps_complete_command(sc, cm);
2302 		}
2303 
2304 		desc->Words.Low = 0xffffffff;
2305 		desc->Words.High = 0xffffffff;
2306 	}
2307 
2308 	if (pq != sc->replypostindex) {
2309 		mps_dprint(sc, MPS_TRACE,
2310 		    "%s sc %p writing postindex %d\n",
2311 		    __func__, sc, sc->replypostindex);
2312 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2313 	}
2314 
2315 	return;
2316 }
2317 
2318 static void
2319 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2320     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2321 {
2322 	struct mps_event_handle *eh;
2323 	int event, handled = 0;
2324 
2325 	event = le16toh(reply->Event);
2326 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2327 		if (isset(eh->mask, event)) {
2328 			eh->callback(sc, data, reply);
2329 			handled++;
2330 		}
2331 	}
2332 
2333 	if (handled == 0)
2334 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2335 
2336 	/*
2337 	 * This is the only place that the event/reply should be freed.
2338 	 * Anything wanting to hold onto the event data should have
2339 	 * already copied it into their own storage.
2340 	 */
2341 	mps_free_reply(sc, data);
2342 }
2343 
2344 static void
2345 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2346 {
2347 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2348 
2349 	if (cm->cm_reply)
2350 		MPS_DPRINT_EVENT(sc, generic,
2351 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2352 
2353 	mps_free_command(sc, cm);
2354 
2355 	/* next, send a port enable */
2356 	mpssas_startup(sc);
2357 }
2358 
2359 /*
2360  * For both register_events and update_events, the caller supplies a bitmap
2361  * of events that it _wants_.  These functions then turn that into a bitmask
2362  * suitable for the controller.
2363  */
2364 int
2365 mps_register_events(struct mps_softc *sc, u32 *mask,
2366     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2367 {
2368 	struct mps_event_handle *eh;
2369 	int error = 0;
2370 
2371 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2372 	if(!eh) {
2373 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2374 		return (ENOMEM);
2375 	}
2376 	eh->callback = cb;
2377 	eh->data = data;
2378 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2379 	if (mask != NULL)
2380 		error = mps_update_events(sc, eh, mask);
2381 	*handle = eh;
2382 
2383 	return (error);
2384 }
2385 
2386 int
2387 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2388     u32 *mask)
2389 {
2390 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2391 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2392 	struct mps_command *cm;
2393 	int error, i;
2394 
2395 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2396 
2397 	if ((mask != NULL) && (handle != NULL))
2398 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2399 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2400 
2401 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2402 		sc->event_mask[i] = -1;
2403 
2404 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2405 		sc->event_mask[i] &= ~handle->mask[i];
2406 
2407 
2408 	if ((cm = mps_alloc_command(sc)) == NULL)
2409 		return (EBUSY);
2410 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2411 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2412 	evtreq->MsgFlags = 0;
2413 	evtreq->SASBroadcastPrimitiveMasks = 0;
2414 #ifdef MPS_DEBUG_ALL_EVENTS
2415 	{
2416 		u_char fullmask[16];
2417 		memset(fullmask, 0x00, 16);
2418 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2419 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2420 	}
2421 #else
2422         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2423                 evtreq->EventMasks[i] =
2424                     htole32(sc->event_mask[i]);
2425 #endif
2426 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2427 	cm->cm_data = NULL;
2428 
2429 	error = mps_wait_command(sc, &cm, 60, 0);
2430 	if (cm != NULL)
2431 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2432 	if ((reply == NULL) ||
2433 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2434 		error = ENXIO;
2435 
2436 	if (reply)
2437 		MPS_DPRINT_EVENT(sc, generic, reply);
2438 
2439 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2440 
2441 	if (cm != NULL)
2442 		mps_free_command(sc, cm);
2443 	return (error);
2444 }
2445 
2446 static int
2447 mps_reregister_events(struct mps_softc *sc)
2448 {
2449 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2450 	struct mps_command *cm;
2451 	struct mps_event_handle *eh;
2452 	int error, i;
2453 
2454 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2455 
2456 	/* first, reregister events */
2457 
2458 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2459 		sc->event_mask[i] = -1;
2460 
2461 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2462 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2463 			sc->event_mask[i] &= ~eh->mask[i];
2464 	}
2465 
2466 	if ((cm = mps_alloc_command(sc)) == NULL)
2467 		return (EBUSY);
2468 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2469 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2470 	evtreq->MsgFlags = 0;
2471 	evtreq->SASBroadcastPrimitiveMasks = 0;
2472 #ifdef MPS_DEBUG_ALL_EVENTS
2473 	{
2474 		u_char fullmask[16];
2475 		memset(fullmask, 0x00, 16);
2476 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2477 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2478 	}
2479 #else
2480         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2481                 evtreq->EventMasks[i] =
2482                     htole32(sc->event_mask[i]);
2483 #endif
2484 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2485 	cm->cm_data = NULL;
2486 	cm->cm_complete = mps_reregister_events_complete;
2487 
2488 	error = mps_map_command(sc, cm);
2489 
2490 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2491 	    error);
2492 	return (error);
2493 }
2494 
2495 void
2496 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2497 {
2498 
2499 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2500 	free(handle, M_MPT2);
2501 }
2502 
2503 /*
2504  * Add a chain element as the next SGE for the specified command.
2505  * Reset cm_sge and cm_sgesize to indicate all the available space.
2506  */
2507 static int
2508 mps_add_chain(struct mps_command *cm)
2509 {
2510 	MPI2_SGE_CHAIN32 *sgc;
2511 	struct mps_chain *chain;
2512 	int space;
2513 
2514 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2515 		panic("MPS: Need SGE Error Code\n");
2516 
2517 	chain = mps_alloc_chain(cm->cm_sc);
2518 	if (chain == NULL)
2519 		return (ENOBUFS);
2520 
2521 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2522 
2523 	/*
2524 	 * Note: a double-linked list is used to make it easier to
2525 	 * walk for debugging.
2526 	 */
2527 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2528 
2529 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2530 	sgc->Length = htole16(space);
2531 	sgc->NextChainOffset = 0;
2532 	/* TODO Looks like bug in Setting sgc->Flags.
2533 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2534 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2535 	 *	This is fine.. because we are not using simple element. In case of
2536 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2537  	 */
2538 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2539 	sgc->Address = htole32(chain->chain_busaddr);
2540 
2541 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2542 	cm->cm_sglsize = space;
2543 	return (0);
2544 }
2545 
2546 /*
2547  * Add one scatter-gather element (chain, simple, transaction context)
2548  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2549  * cm_sge as the remaining size and pointer to the next SGE to fill
2550  * in, respectively.
2551  */
2552 int
2553 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2554 {
2555 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2556 	MPI2_SGE_SIMPLE64 *sge = sgep;
2557 	int error, type;
2558 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2559 
2560 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2561 
2562 #ifdef INVARIANTS
2563 	switch (type) {
2564 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2565 		if (len != tc->DetailsLength + 4)
2566 			panic("TC %p length %u or %zu?", tc,
2567 			    tc->DetailsLength + 4, len);
2568 		}
2569 		break;
2570 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2571 		/* Driver only uses 32-bit chain elements */
2572 		if (len != MPS_SGC_SIZE)
2573 			panic("CHAIN %p length %u or %zu?", sgep,
2574 			    MPS_SGC_SIZE, len);
2575 		break;
2576 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2577 		/* Driver only uses 64-bit SGE simple elements */
2578 		if (len != MPS_SGE64_SIZE)
2579 			panic("SGE simple %p length %u or %zu?", sge,
2580 			    MPS_SGE64_SIZE, len);
2581 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2582 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2583 			panic("SGE simple %p not marked 64-bit?", sge);
2584 
2585 		break;
2586 	default:
2587 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2588 	}
2589 #endif
2590 
2591 	/*
2592 	 * case 1: 1 more segment, enough room for it
2593 	 * case 2: 2 more segments, enough room for both
2594 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2595 	 * case 4: >=1 more segment, enough room for only a chain
2596 	 * case 5: >=1 more segment, no room for anything (error)
2597          */
2598 
2599 	/*
2600 	 * There should be room for at least a chain element, or this
2601 	 * code is buggy.  Case (5).
2602 	 */
2603 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2604 		panic("MPS: Need SGE Error Code\n");
2605 
2606 	if (segsleft >= 2 &&
2607 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2608 		/*
2609 		 * There are 2 or more segments left to add, and only
2610 		 * enough room for 1 and a chain.  Case (3).
2611 		 *
2612 		 * Mark as last element in this chain if necessary.
2613 		 */
2614 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2615 			sge->FlagsLength |= htole32(
2616 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2617 		}
2618 
2619 		/*
2620 		 * Add the item then a chain.  Do the chain now,
2621 		 * rather than on the next iteration, to simplify
2622 		 * understanding the code.
2623 		 */
2624 		cm->cm_sglsize -= len;
2625 		bcopy(sgep, cm->cm_sge, len);
2626 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2627 		return (mps_add_chain(cm));
2628 	}
2629 
2630 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2631 		/*
2632 		 * 1 or more segment, enough room for only a chain.
2633 		 * Hope the previous element wasn't a Simple entry
2634 		 * that needed to be marked with
2635 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2636 		 */
2637 		if ((error = mps_add_chain(cm)) != 0)
2638 			return (error);
2639 	}
2640 
2641 #ifdef INVARIANTS
2642 	/* Case 1: 1 more segment, enough room for it. */
2643 	if (segsleft == 1 && cm->cm_sglsize < len)
2644 		panic("1 seg left and no room? %u versus %zu",
2645 		    cm->cm_sglsize, len);
2646 
2647 	/* Case 2: 2 more segments, enough room for both */
2648 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2649 		panic("2 segs left and no room? %u versus %zu",
2650 		    cm->cm_sglsize, len);
2651 #endif
2652 
2653 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2654 		/*
2655 		 * If this is a bi-directional request, need to account for that
2656 		 * here.  Save the pre-filled sge values.  These will be used
2657 		 * either for the 2nd SGL or for a single direction SGL.  If
2658 		 * cm_out_len is non-zero, this is a bi-directional request, so
2659 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2660 		 * fill in the IN SGL.  Note that at this time, when filling in
2661 		 * 2 SGL's for a bi-directional request, they both use the same
2662 		 * DMA buffer (same cm command).
2663 		 */
2664 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2665 		saved_address_low = sge->Address.Low;
2666 		saved_address_high = sge->Address.High;
2667 		if (cm->cm_out_len) {
2668 			sge->FlagsLength = htole32(cm->cm_out_len |
2669 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2670 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2671 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2672 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2673 			    MPI2_SGE_FLAGS_SHIFT));
2674 			cm->cm_sglsize -= len;
2675 			bcopy(sgep, cm->cm_sge, len);
2676 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2677 			    + len);
2678 		}
2679 		saved_buf_len |=
2680 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2681 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2682 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2683 		    MPI2_SGE_FLAGS_END_OF_LIST |
2684 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2685 		    MPI2_SGE_FLAGS_SHIFT);
2686 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2687 			saved_buf_len |=
2688 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2689 			    MPI2_SGE_FLAGS_SHIFT);
2690 		} else {
2691 			saved_buf_len |=
2692 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2693 			    MPI2_SGE_FLAGS_SHIFT);
2694 		}
2695 		sge->FlagsLength = htole32(saved_buf_len);
2696 		sge->Address.Low = saved_address_low;
2697 		sge->Address.High = saved_address_high;
2698 	}
2699 
2700 	cm->cm_sglsize -= len;
2701 	bcopy(sgep, cm->cm_sge, len);
2702 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2703 	return (0);
2704 }
2705 
2706 /*
2707  * Add one dma segment to the scatter-gather list for a command.
2708  */
2709 int
2710 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2711     int segsleft)
2712 {
2713 	MPI2_SGE_SIMPLE64 sge;
2714 
2715 	/*
2716 	 * This driver always uses 64-bit address elements for simplicity.
2717 	 */
2718 	bzero(&sge, sizeof(sge));
2719 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2720 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2721 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2722 	mps_from_u64(pa, &sge.Address);
2723 
2724 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2725 }
2726 
2727 static void
2728 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2729 {
2730 	struct mps_softc *sc;
2731 	struct mps_command *cm;
2732 	u_int i, dir, sflags;
2733 
2734 	cm = (struct mps_command *)arg;
2735 	sc = cm->cm_sc;
2736 
2737 	/*
2738 	 * In this case, just print out a warning and let the chip tell the
2739 	 * user they did the wrong thing.
2740 	 */
2741 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2742 		mps_dprint(sc, MPS_ERROR,
2743 			   "%s: warning: busdma returned %d segments, "
2744 			   "more than the %d allowed\n", __func__, nsegs,
2745 			   cm->cm_max_segs);
2746 	}
2747 
2748 	/*
2749 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2750 	 * here.  In that case, both direction flags will be set.
2751 	 */
2752 	sflags = 0;
2753 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2754 		/*
2755 		 * We have to add a special case for SMP passthrough, there
2756 		 * is no easy way to generically handle it.  The first
2757 		 * S/G element is used for the command (therefore the
2758 		 * direction bit needs to be set).  The second one is used
2759 		 * for the reply.  We'll leave it to the caller to make
2760 		 * sure we only have two buffers.
2761 		 */
2762 		/*
2763 		 * Even though the busdma man page says it doesn't make
2764 		 * sense to have both direction flags, it does in this case.
2765 		 * We have one s/g element being accessed in each direction.
2766 		 */
2767 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2768 
2769 		/*
2770 		 * Set the direction flag on the first buffer in the SMP
2771 		 * passthrough request.  We'll clear it for the second one.
2772 		 */
2773 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2774 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2775 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2776 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2777 		dir = BUS_DMASYNC_PREWRITE;
2778 	} else
2779 		dir = BUS_DMASYNC_PREREAD;
2780 
2781 	for (i = 0; i < nsegs; i++) {
2782 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2783 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2784 		}
2785 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2786 		    sflags, nsegs - i);
2787 		if (error != 0) {
2788 			/* Resource shortage, roll back! */
2789 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2790 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2791 				    "consider increasing hw.mps.max_chains.\n");
2792 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2793 			mps_complete_command(sc, cm);
2794 			return;
2795 		}
2796 	}
2797 
2798 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2799 	mps_enqueue_request(sc, cm);
2800 
2801 	return;
2802 }
2803 
2804 static void
2805 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2806 	     int error)
2807 {
2808 	mps_data_cb(arg, segs, nsegs, error);
2809 }
2810 
2811 /*
2812  * This is the routine to enqueue commands ansynchronously.
2813  * Note that the only error path here is from bus_dmamap_load(), which can
2814  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2815  * assumed that if you have a command in-hand, then you have enough credits
2816  * to use it.
2817  */
2818 int
2819 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2820 {
2821 	int error = 0;
2822 
2823 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2824 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2825 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2826 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2827 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2828 		    cm->cm_data, mps_data_cb, cm, 0);
2829 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2830 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2831 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2832 	} else {
2833 		/* Add a zero-length element as needed */
2834 		if (cm->cm_sge != NULL)
2835 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2836 		mps_enqueue_request(sc, cm);
2837 	}
2838 
2839 	return (error);
2840 }
2841 
2842 /*
2843  * This is the routine to enqueue commands synchronously.  An error of
2844  * EINPROGRESS from mps_map_command() is ignored since the command will
2845  * be executed and enqueued automatically.  Other errors come from msleep().
2846  */
2847 int
2848 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2849     int sleep_flag)
2850 {
2851 	int error, rc;
2852 	struct timeval cur_time, start_time;
2853 	struct mps_command *cm = *cmp;
2854 
2855 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2856 		return  EBUSY;
2857 
2858 	cm->cm_complete = NULL;
2859 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2860 	error = mps_map_command(sc, cm);
2861 	if ((error != 0) && (error != EINPROGRESS))
2862 		return (error);
2863 
2864 	/*
2865 	 * Check for context and wait for 50 mSec at a time until time has
2866 	 * expired or the command has finished.  If msleep can't be used, need
2867 	 * to poll.
2868 	 */
2869 	if (curthread->td_no_sleeping != 0)
2870 		sleep_flag = NO_SLEEP;
2871 	getmicrouptime(&start_time);
2872 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2873 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2874 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2875 		if (error == EWOULDBLOCK) {
2876 			/*
2877 			 * Record the actual elapsed time in the case of a
2878 			 * timeout for the message below.
2879 			 */
2880 			getmicrouptime(&cur_time);
2881 			timevalsub(&cur_time, &start_time);
2882 		}
2883 	} else {
2884 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2885 			mps_intr_locked(sc);
2886 			if (sleep_flag == CAN_SLEEP)
2887 				pause("mpswait", hz/20);
2888 			else
2889 				DELAY(50000);
2890 
2891 			getmicrouptime(&cur_time);
2892 			timevalsub(&cur_time, &start_time);
2893 			if (cur_time.tv_sec > timeout) {
2894 				error = EWOULDBLOCK;
2895 				break;
2896 			}
2897 		}
2898 	}
2899 
2900 	if (error == EWOULDBLOCK) {
2901 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2902 		    " elapsed=%jd\n", __func__, timeout,
2903 		    (intmax_t)cur_time.tv_sec);
2904 		rc = mps_reinit(sc);
2905 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2906 		    "failed");
2907 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
2908 			/*
2909 			 * Tell the caller that we freed the command in a
2910 			 * reinit.
2911 			 */
2912 			*cmp = NULL;
2913 		}
2914 		error = ETIMEDOUT;
2915 	}
2916 	return (error);
2917 }
2918 
2919 /*
2920  * The MPT driver had a verbose interface for config pages.  In this driver,
2921  * reduce it to much simpler terms, similar to the Linux driver.
2922  */
2923 int
2924 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2925 {
2926 	MPI2_CONFIG_REQUEST *req;
2927 	struct mps_command *cm;
2928 	int error;
2929 
2930 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2931 		return (EBUSY);
2932 	}
2933 
2934 	cm = mps_alloc_command(sc);
2935 	if (cm == NULL) {
2936 		return (EBUSY);
2937 	}
2938 
2939 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2940 	req->Function = MPI2_FUNCTION_CONFIG;
2941 	req->Action = params->action;
2942 	req->SGLFlags = 0;
2943 	req->ChainOffset = 0;
2944 	req->PageAddress = params->page_address;
2945 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2946 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2947 
2948 		hdr = &params->hdr.Ext;
2949 		req->ExtPageType = hdr->ExtPageType;
2950 		req->ExtPageLength = hdr->ExtPageLength;
2951 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2952 		req->Header.PageLength = 0; /* Must be set to zero */
2953 		req->Header.PageNumber = hdr->PageNumber;
2954 		req->Header.PageVersion = hdr->PageVersion;
2955 	} else {
2956 		MPI2_CONFIG_PAGE_HEADER *hdr;
2957 
2958 		hdr = &params->hdr.Struct;
2959 		req->Header.PageType = hdr->PageType;
2960 		req->Header.PageNumber = hdr->PageNumber;
2961 		req->Header.PageLength = hdr->PageLength;
2962 		req->Header.PageVersion = hdr->PageVersion;
2963 	}
2964 
2965 	cm->cm_data = params->buffer;
2966 	cm->cm_length = params->length;
2967 	if (cm->cm_data != NULL) {
2968 		cm->cm_sge = &req->PageBufferSGE;
2969 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2970 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2971 	} else
2972 		cm->cm_sge = NULL;
2973 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2974 
2975 	cm->cm_complete_data = params;
2976 	if (params->callback != NULL) {
2977 		cm->cm_complete = mps_config_complete;
2978 		return (mps_map_command(sc, cm));
2979 	} else {
2980 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
2981 		if (error) {
2982 			mps_dprint(sc, MPS_FAULT,
2983 			    "Error %d reading config page\n", error);
2984 			if (cm != NULL)
2985 				mps_free_command(sc, cm);
2986 			return (error);
2987 		}
2988 		mps_config_complete(sc, cm);
2989 	}
2990 
2991 	return (0);
2992 }
2993 
2994 int
2995 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2996 {
2997 	return (EINVAL);
2998 }
2999 
3000 static void
3001 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3002 {
3003 	MPI2_CONFIG_REPLY *reply;
3004 	struct mps_config_params *params;
3005 
3006 	MPS_FUNCTRACE(sc);
3007 	params = cm->cm_complete_data;
3008 
3009 	if (cm->cm_data != NULL) {
3010 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3011 		    BUS_DMASYNC_POSTREAD);
3012 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3013 	}
3014 
3015 	/*
3016 	 * XXX KDM need to do more error recovery?  This results in the
3017 	 * device in question not getting probed.
3018 	 */
3019 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3020 		params->status = MPI2_IOCSTATUS_BUSY;
3021 		goto done;
3022 	}
3023 
3024 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3025 	if (reply == NULL) {
3026 		params->status = MPI2_IOCSTATUS_BUSY;
3027 		goto done;
3028 	}
3029 	params->status = reply->IOCStatus;
3030 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3031 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3032 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3033 		params->hdr.Ext.PageType = reply->Header.PageType;
3034 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3035 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3036 	} else {
3037 		params->hdr.Struct.PageType = reply->Header.PageType;
3038 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3039 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3040 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3041 	}
3042 
3043 done:
3044 	mps_free_command(sc, cm);
3045 	if (params->callback != NULL)
3046 		params->callback(sc, params);
3047 
3048 	return;
3049 }
3050