xref: /freebsd/sys/dev/mps/mps.c (revision 834063202a16592e1ef5c3a9fbd04ca5f1df3ed0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2009 Yahoo! Inc.
5  * Copyright (c) 2011-2015 LSI Corp.
6  * Copyright (c) 2013-2015 Avago Technologies
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
31  *
32  * $FreeBSD$
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 /* Communications core for Avago Technologies (LSI) MPT2 */
39 
40 /* TODO Move headers to mpsvar */
41 #include <sys/types.h>
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/selinfo.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/module.h>
49 #include <sys/bus.h>
50 #include <sys/conf.h>
51 #include <sys/bio.h>
52 #include <sys/malloc.h>
53 #include <sys/uio.h>
54 #include <sys/sysctl.h>
55 #include <sys/smp.h>
56 #include <sys/queue.h>
57 #include <sys/kthread.h>
58 #include <sys/taskqueue.h>
59 #include <sys/endian.h>
60 #include <sys/eventhandler.h>
61 #include <sys/sbuf.h>
62 
63 #include <machine/bus.h>
64 #include <machine/resource.h>
65 #include <sys/rman.h>
66 #include <sys/proc.h>
67 
68 #include <dev/pci/pcivar.h>
69 
70 #include <cam/cam.h>
71 #include <cam/scsi/scsi_all.h>
72 
73 #include <dev/mps/mpi/mpi2_type.h>
74 #include <dev/mps/mpi/mpi2.h>
75 #include <dev/mps/mpi/mpi2_ioc.h>
76 #include <dev/mps/mpi/mpi2_sas.h>
77 #include <dev/mps/mpi/mpi2_cnfg.h>
78 #include <dev/mps/mpi/mpi2_init.h>
79 #include <dev/mps/mpi/mpi2_tool.h>
80 #include <dev/mps/mps_ioctl.h>
81 #include <dev/mps/mpsvar.h>
82 #include <dev/mps/mps_table.h>
83 
84 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
85 static int mps_init_queues(struct mps_softc *sc);
86 static void mps_resize_queues(struct mps_softc *sc);
87 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
88 static int mps_transition_operational(struct mps_softc *sc);
89 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
90 static void mps_iocfacts_free(struct mps_softc *sc);
91 static void mps_startup(void *arg);
92 static int mps_send_iocinit(struct mps_softc *sc);
93 static int mps_alloc_queues(struct mps_softc *sc);
94 static int mps_alloc_hw_queues(struct mps_softc *sc);
95 static int mps_alloc_replies(struct mps_softc *sc);
96 static int mps_alloc_requests(struct mps_softc *sc);
97 static int mps_attach_log(struct mps_softc *sc);
98 static __inline void mps_complete_command(struct mps_softc *sc,
99     struct mps_command *cm);
100 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
101     MPI2_EVENT_NOTIFICATION_REPLY *reply);
102 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
103 static void mps_periodic(void *);
104 static int mps_reregister_events(struct mps_softc *sc);
105 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
106 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
107 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
108 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
109 static void mps_parse_debug(struct mps_softc *sc, char *list);
110 
111 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
112 
113 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
114 MALLOC_DECLARE(M_MPSUSER);
115 
116 /*
117  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
118  * any state and back to its initialization state machine.
119  */
120 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
121 
122 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
123  * Compiler only support unint64_t to be passed as argument.
124  * Otherwise it will throw below error
125  * "aggregate value used where an integer was expected"
126  */
127 
128 typedef union _reply_descriptor {
129         u64 word;
130         struct {
131                 u32 low;
132                 u32 high;
133         } u;
134 }reply_descriptor,address_descriptor;
135 
136 /* Rate limit chain-fail messages to 1 per minute */
137 static struct timeval mps_chainfail_interval = { 60, 0 };
138 
139 /*
140  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
141  * If this function is called from process context, it can sleep
142  * and there is no harm to sleep, in case if this fuction is called
143  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
144  * based on sleep flags driver will call either msleep, pause or DELAY.
145  * msleep and pause are of same variant, but pause is used when mps_mtx
146  * is not hold by driver.
147  *
148  */
149 static int
150 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
151 {
152 	uint32_t reg;
153 	int i, error, tries = 0;
154 	uint8_t first_wait_done = FALSE;
155 
156 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
157 
158 	/* Clear any pending interrupts */
159 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
160 
161 	/*
162 	 * Force NO_SLEEP for threads prohibited to sleep
163  	 * e.a Thread from interrupt handler are prohibited to sleep.
164  	 */
165 	if (curthread->td_no_sleeping != 0)
166 		sleep_flag = NO_SLEEP;
167 
168 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
169 
170 	/* Push the magic sequence */
171 	error = ETIMEDOUT;
172 	while (tries++ < 20) {
173 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
174 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
175 			    mpt2_reset_magic[i]);
176 		/* wait 100 msec */
177 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
178 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
179 			    "mpsdiag", hz/10);
180 		else if (sleep_flag == CAN_SLEEP)
181 			pause("mpsdiag", hz/10);
182 		else
183 			DELAY(100 * 1000);
184 
185 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
186 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
187 			error = 0;
188 			break;
189 		}
190 	}
191 	if (error) {
192 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
193 		    error);
194 		return (error);
195 	}
196 
197 	/* Send the actual reset.  XXX need to refresh the reg? */
198 	reg |= MPI2_DIAG_RESET_ADAPTER;
199 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
200 		reg);
201 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
202 
203 	/* Wait up to 300 seconds in 50ms intervals */
204 	error = ETIMEDOUT;
205 	for (i = 0; i < 6000; i++) {
206 		/*
207 		 * Wait 50 msec. If this is the first time through, wait 256
208 		 * msec to satisfy Diag Reset timing requirements.
209 		 */
210 		if (first_wait_done) {
211 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
212 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
213 				    "mpsdiag", hz/20);
214 			else if (sleep_flag == CAN_SLEEP)
215 				pause("mpsdiag", hz/20);
216 			else
217 				DELAY(50 * 1000);
218 		} else {
219 			DELAY(256 * 1000);
220 			first_wait_done = TRUE;
221 		}
222 		/*
223 		 * Check for the RESET_ADAPTER bit to be cleared first, then
224 		 * wait for the RESET state to be cleared, which takes a little
225 		 * longer.
226 		 */
227 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
228 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
229 			continue;
230 		}
231 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
232 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
233 			error = 0;
234 			break;
235 		}
236 	}
237 	if (error) {
238 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
239 		    error);
240 		return (error);
241 	}
242 
243 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
244 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
245 
246 	return (0);
247 }
248 
249 static int
250 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
251 {
252 	int error;
253 
254 	MPS_FUNCTRACE(sc);
255 
256 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
257 
258 	error = 0;
259 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
260 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
261 	    MPI2_DOORBELL_FUNCTION_SHIFT);
262 
263 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
264 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
265 		    "Doorbell handshake failed\n");
266 		error = ETIMEDOUT;
267 	}
268 
269 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
270 	return (error);
271 }
272 
273 static int
274 mps_transition_ready(struct mps_softc *sc)
275 {
276 	uint32_t reg, state;
277 	int error, tries = 0;
278 	int sleep_flags;
279 
280 	MPS_FUNCTRACE(sc);
281 	/* If we are in attach call, do not sleep */
282 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
283 					? CAN_SLEEP:NO_SLEEP;
284 	error = 0;
285 
286 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
287 	   __func__, sleep_flags);
288 
289 	while (tries++ < 1200) {
290 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
291 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
292 
293 		/*
294 		 * Ensure the IOC is ready to talk.  If it's not, try
295 		 * resetting it.
296 		 */
297 		if (reg & MPI2_DOORBELL_USED) {
298 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
299 			    "reset\n");
300 			mps_diag_reset(sc, sleep_flags);
301 			DELAY(50000);
302 			continue;
303 		}
304 
305 		/* Is the adapter owned by another peer? */
306 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
307 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
308 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
309 			    "control of another peer host, aborting "
310 			    "initialization.\n");
311 			error = ENXIO;
312 			break;
313 		}
314 
315 		state = reg & MPI2_IOC_STATE_MASK;
316 		if (state == MPI2_IOC_STATE_READY) {
317 			/* Ready to go! */
318 			error = 0;
319 			break;
320 		} else if (state == MPI2_IOC_STATE_FAULT) {
321 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
322 			    "state 0x%x, resetting\n",
323 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
324 			mps_diag_reset(sc, sleep_flags);
325 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
326 			/* Need to take ownership */
327 			mps_message_unit_reset(sc, sleep_flags);
328 		} else if (state == MPI2_IOC_STATE_RESET) {
329 			/* Wait a bit, IOC might be in transition */
330 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
331 			    "IOC in unexpected reset state\n");
332 		} else {
333 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
334 			    "IOC in unknown state 0x%x\n", state);
335 			error = EINVAL;
336 			break;
337 		}
338 
339 		/* Wait 50ms for things to settle down. */
340 		DELAY(50000);
341 	}
342 
343 	if (error)
344 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
345 		    "Cannot transition IOC to ready\n");
346 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
347 
348 	return (error);
349 }
350 
351 static int
352 mps_transition_operational(struct mps_softc *sc)
353 {
354 	uint32_t reg, state;
355 	int error;
356 
357 	MPS_FUNCTRACE(sc);
358 
359 	error = 0;
360 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
361 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
362 
363 	state = reg & MPI2_IOC_STATE_MASK;
364 	if (state != MPI2_IOC_STATE_READY) {
365 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
366 		if ((error = mps_transition_ready(sc)) != 0) {
367 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
368 			    "failed to transition ready, exit\n");
369 			return (error);
370 		}
371 	}
372 
373 	error = mps_send_iocinit(sc);
374 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
375 
376 	return (error);
377 }
378 
379 static void
380 mps_resize_queues(struct mps_softc *sc)
381 {
382 	u_int reqcr, prireqcr, maxio, sges_per_frame;
383 
384 	/*
385 	 * Size the queues. Since the reply queues always need one free
386 	 * entry, we'll deduct one reply message here.  The LSI documents
387 	 * suggest instead to add a count to the request queue, but I think
388 	 * that it's better to deduct from reply queue.
389 	 */
390 	prireqcr = MAX(1, sc->max_prireqframes);
391 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
392 
393 	reqcr = MAX(2, sc->max_reqframes);
394 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
395 
396 	sc->num_reqs = prireqcr + reqcr;
397 	sc->num_prireqs = prireqcr;
398 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
399 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
400 
401 	/* Store the request frame size in bytes rather than as 32bit words */
402 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
403 
404 	/*
405 	 * Max IO Size is Page Size * the following:
406 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
407 	 * + 1 for no chain needed in last frame
408 	 *
409 	 * If user suggests a Max IO size to use, use the smaller of the
410 	 * user's value and the calculated value as long as the user's
411 	 * value is larger than 0. The user's value is in pages.
412 	 */
413 	sges_per_frame = sc->reqframesz / sizeof(MPI2_SGE_SIMPLE64) - 1;
414 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
415 
416 	/*
417 	 * If I/O size limitation requested, then use it and pass up to CAM.
418 	 * If not, use MAXPHYS as an optimization hint, but report HW limit.
419 	 */
420 	if (sc->max_io_pages > 0) {
421 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
422 		sc->maxio = maxio;
423 	} else {
424 		sc->maxio = maxio;
425 		maxio = min(maxio, MAXPHYS);
426 	}
427 
428 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
429 	    sges_per_frame * reqcr;
430 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
431 		sc->num_chains = sc->max_chains;
432 
433 	/*
434 	 * Figure out the number of MSIx-based queues.  If the firmware or
435 	 * user has done something crazy and not allowed enough credit for
436 	 * the queues to be useful then don't enable multi-queue.
437 	 */
438 	if (sc->facts->MaxMSIxVectors < 2)
439 		sc->msi_msgs = 1;
440 
441 	if (sc->msi_msgs > 1) {
442 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
443 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
444 		if (sc->num_reqs / sc->msi_msgs < 2)
445 			sc->msi_msgs = 1;
446 	}
447 
448 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
449 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
450 }
451 
452 /*
453  * This is called during attach and when re-initializing due to a Diag Reset.
454  * IOC Facts is used to allocate many of the structures needed by the driver.
455  * If called from attach, de-allocation is not required because the driver has
456  * not allocated any structures yet, but if called from a Diag Reset, previously
457  * allocated structures based on IOC Facts will need to be freed and re-
458  * allocated bases on the latest IOC Facts.
459  */
460 static int
461 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
462 {
463 	int error;
464 	Mpi2IOCFactsReply_t saved_facts;
465 	uint8_t saved_mode, reallocating;
466 
467 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
468 
469 	/* Save old IOC Facts and then only reallocate if Facts have changed */
470 	if (!attaching) {
471 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
472 	}
473 
474 	/*
475 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
476 	 * a re-initialization and only return the error if attaching so the OS
477 	 * can handle it.
478 	 */
479 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
480 		if (attaching) {
481 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
482 			    "IOC Facts with error %d, exit\n", error);
483 			return (error);
484 		} else {
485 			panic("%s failed to get IOC Facts with error %d\n",
486 			    __func__, error);
487 		}
488 	}
489 
490 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
491 
492 	snprintf(sc->fw_version, sizeof(sc->fw_version),
493 	    "%02d.%02d.%02d.%02d",
494 	    sc->facts->FWVersion.Struct.Major,
495 	    sc->facts->FWVersion.Struct.Minor,
496 	    sc->facts->FWVersion.Struct.Unit,
497 	    sc->facts->FWVersion.Struct.Dev);
498 
499 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
500 	    MPS_DRIVER_VERSION);
501 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
502 	     sc->facts->IOCCapabilities,
503 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
504 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
505 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
506 
507 	/*
508 	 * If the chip doesn't support event replay then a hard reset will be
509 	 * required to trigger a full discovery.  Do the reset here then
510 	 * retransition to Ready.  A hard reset might have already been done,
511 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
512 	 * for a Diag Reset.
513 	 */
514 	if (attaching && ((sc->facts->IOCCapabilities &
515 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
516 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
517 		mps_diag_reset(sc, NO_SLEEP);
518 		if ((error = mps_transition_ready(sc)) != 0) {
519 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
520 			    "transition to ready with error %d, exit\n",
521 			    error);
522 			return (error);
523 		}
524 	}
525 
526 	/*
527 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
528 	 * changed from the previous IOC Facts, log a warning, but only if
529 	 * checking this after a Diag Reset and not during attach.
530 	 */
531 	saved_mode = sc->ir_firmware;
532 	if (sc->facts->IOCCapabilities &
533 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
534 		sc->ir_firmware = 1;
535 	if (!attaching) {
536 		if (sc->ir_firmware != saved_mode) {
537 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
538 			    "in IOC Facts does not match previous mode\n");
539 		}
540 	}
541 
542 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
543 	reallocating = FALSE;
544 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
545 
546 	if ((!attaching) &&
547 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
548 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
549 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
550 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
551 	    (saved_facts.ProductID != sc->facts->ProductID) ||
552 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
553 	    (saved_facts.IOCRequestFrameSize !=
554 	    sc->facts->IOCRequestFrameSize) ||
555 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
556 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
557 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
558 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
559 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
560 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
561 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
562 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
563 	    (saved_facts.MaxPersistentEntries !=
564 	    sc->facts->MaxPersistentEntries))) {
565 		reallocating = TRUE;
566 
567 		/* Record that we reallocated everything */
568 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
569 	}
570 
571 	/*
572 	 * Some things should be done if attaching or re-allocating after a Diag
573 	 * Reset, but are not needed after a Diag Reset if the FW has not
574 	 * changed.
575 	 */
576 	if (attaching || reallocating) {
577 		/*
578 		 * Check if controller supports FW diag buffers and set flag to
579 		 * enable each type.
580 		 */
581 		if (sc->facts->IOCCapabilities &
582 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
583 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
584 			    enabled = TRUE;
585 		if (sc->facts->IOCCapabilities &
586 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
587 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
588 			    enabled = TRUE;
589 		if (sc->facts->IOCCapabilities &
590 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
591 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
592 			    enabled = TRUE;
593 
594 		/*
595 		 * Set flag if EEDP is supported and if TLR is supported.
596 		 */
597 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
598 			sc->eedp_enabled = TRUE;
599 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
600 			sc->control_TLR = TRUE;
601 
602 		mps_resize_queues(sc);
603 
604 		/*
605 		 * Initialize all Tail Queues
606 		 */
607 		TAILQ_INIT(&sc->req_list);
608 		TAILQ_INIT(&sc->high_priority_req_list);
609 		TAILQ_INIT(&sc->chain_list);
610 		TAILQ_INIT(&sc->tm_list);
611 	}
612 
613 	/*
614 	 * If doing a Diag Reset and the FW is significantly different
615 	 * (reallocating will be set above in IOC Facts comparison), then all
616 	 * buffers based on the IOC Facts will need to be freed before they are
617 	 * reallocated.
618 	 */
619 	if (reallocating) {
620 		mps_iocfacts_free(sc);
621 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
622 		    saved_facts.MaxVolumes);
623 	}
624 
625 	/*
626 	 * Any deallocation has been completed.  Now start reallocating
627 	 * if needed.  Will only need to reallocate if attaching or if the new
628 	 * IOC Facts are different from the previous IOC Facts after a Diag
629 	 * Reset. Targets have already been allocated above if needed.
630 	 */
631 	error = 0;
632 	while (attaching || reallocating) {
633 		if ((error = mps_alloc_hw_queues(sc)) != 0)
634 			break;
635 		if ((error = mps_alloc_replies(sc)) != 0)
636 			break;
637 		if ((error = mps_alloc_requests(sc)) != 0)
638 			break;
639 		if ((error = mps_alloc_queues(sc)) != 0)
640 			break;
641 
642 		break;
643 	}
644 	if (error) {
645 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
646 		    "Failed to alloc queues with error %d\n", error);
647 		mps_free(sc);
648 		return (error);
649 	}
650 
651 	/* Always initialize the queues */
652 	bzero(sc->free_queue, sc->fqdepth * 4);
653 	mps_init_queues(sc);
654 
655 	/*
656 	 * Always get the chip out of the reset state, but only panic if not
657 	 * attaching.  If attaching and there is an error, that is handled by
658 	 * the OS.
659 	 */
660 	error = mps_transition_operational(sc);
661 	if (error != 0) {
662 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
663 		    "transition to operational with error %d\n", error);
664 		mps_free(sc);
665 		return (error);
666 	}
667 
668 	/*
669 	 * Finish the queue initialization.
670 	 * These are set here instead of in mps_init_queues() because the
671 	 * IOC resets these values during the state transition in
672 	 * mps_transition_operational().  The free index is set to 1
673 	 * because the corresponding index in the IOC is set to 0, and the
674 	 * IOC treats the queues as full if both are set to the same value.
675 	 * Hence the reason that the queue can't hold all of the possible
676 	 * replies.
677 	 */
678 	sc->replypostindex = 0;
679 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
680 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
681 
682 	/*
683 	 * Attach the subsystems so they can prepare their event masks.
684 	 * XXX Should be dynamic so that IM/IR and user modules can attach
685 	 */
686 	error = 0;
687 	while (attaching) {
688 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
689 		if ((error = mps_attach_log(sc)) != 0)
690 			break;
691 		if ((error = mps_attach_sas(sc)) != 0)
692 			break;
693 		if ((error = mps_attach_user(sc)) != 0)
694 			break;
695 		break;
696 	}
697 	if (error) {
698 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
699 		    "subsystems: error %d\n", error);
700 		mps_free(sc);
701 		return (error);
702 	}
703 
704 	/*
705 	 * XXX If the number of MSI-X vectors changes during re-init, this
706 	 * won't see it and adjust.
707 	 */
708 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
709 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
710 		    "interrupts\n");
711 		mps_free(sc);
712 		return (error);
713 	}
714 
715 	/*
716 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
717 	 * reset it after a Diag Reset, just in case.
718 	 */
719 	sc->WD_available = FALSE;
720 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
721 		sc->WD_available = TRUE;
722 
723 	return (error);
724 }
725 
726 /*
727  * This is called if memory is being free (during detach for example) and when
728  * buffers need to be reallocated due to a Diag Reset.
729  */
730 static void
731 mps_iocfacts_free(struct mps_softc *sc)
732 {
733 	struct mps_command *cm;
734 	int i;
735 
736 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
737 
738 	if (sc->free_busaddr != 0)
739 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
740 	if (sc->free_queue != NULL)
741 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
742 		    sc->queues_map);
743 	if (sc->queues_dmat != NULL)
744 		bus_dma_tag_destroy(sc->queues_dmat);
745 
746 	if (sc->chain_busaddr != 0)
747 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
748 	if (sc->chain_frames != NULL)
749 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
750 		    sc->chain_map);
751 	if (sc->chain_dmat != NULL)
752 		bus_dma_tag_destroy(sc->chain_dmat);
753 
754 	if (sc->sense_busaddr != 0)
755 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
756 	if (sc->sense_frames != NULL)
757 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
758 		    sc->sense_map);
759 	if (sc->sense_dmat != NULL)
760 		bus_dma_tag_destroy(sc->sense_dmat);
761 
762 	if (sc->reply_busaddr != 0)
763 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
764 	if (sc->reply_frames != NULL)
765 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
766 		    sc->reply_map);
767 	if (sc->reply_dmat != NULL)
768 		bus_dma_tag_destroy(sc->reply_dmat);
769 
770 	if (sc->req_busaddr != 0)
771 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
772 	if (sc->req_frames != NULL)
773 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
774 	if (sc->req_dmat != NULL)
775 		bus_dma_tag_destroy(sc->req_dmat);
776 
777 	if (sc->chains != NULL)
778 		free(sc->chains, M_MPT2);
779 	if (sc->commands != NULL) {
780 		for (i = 1; i < sc->num_reqs; i++) {
781 			cm = &sc->commands[i];
782 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
783 		}
784 		free(sc->commands, M_MPT2);
785 	}
786 	if (sc->buffer_dmat != NULL)
787 		bus_dma_tag_destroy(sc->buffer_dmat);
788 
789 	mps_pci_free_interrupts(sc);
790 	free(sc->queues, M_MPT2);
791 	sc->queues = NULL;
792 }
793 
794 /*
795  * The terms diag reset and hard reset are used interchangeably in the MPI
796  * docs to mean resetting the controller chip.  In this code diag reset
797  * cleans everything up, and the hard reset function just sends the reset
798  * sequence to the chip.  This should probably be refactored so that every
799  * subsystem gets a reset notification of some sort, and can clean up
800  * appropriately.
801  */
802 int
803 mps_reinit(struct mps_softc *sc)
804 {
805 	int error;
806 	struct mpssas_softc *sassc;
807 
808 	sassc = sc->sassc;
809 
810 	MPS_FUNCTRACE(sc);
811 
812 	mtx_assert(&sc->mps_mtx, MA_OWNED);
813 
814 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
815 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
816 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
817 		return 0;
818 	}
819 
820 	/* make sure the completion callbacks can recognize they're getting
821 	 * a NULL cm_reply due to a reset.
822 	 */
823 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
824 
825 	/*
826 	 * Mask interrupts here.
827 	 */
828 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
829 	mps_mask_intr(sc);
830 
831 	error = mps_diag_reset(sc, CAN_SLEEP);
832 	if (error != 0) {
833 		/* XXXSL No need to panic here */
834 		panic("%s hard reset failed with error %d\n",
835 		    __func__, error);
836 	}
837 
838 	/* Restore the PCI state, including the MSI-X registers */
839 	mps_pci_restore(sc);
840 
841 	/* Give the I/O subsystem special priority to get itself prepared */
842 	mpssas_handle_reinit(sc);
843 
844 	/*
845 	 * Get IOC Facts and allocate all structures based on this information.
846 	 * The attach function will also call mps_iocfacts_allocate at startup.
847 	 * If relevant values have changed in IOC Facts, this function will free
848 	 * all of the memory based on IOC Facts and reallocate that memory.
849 	 */
850 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
851 		panic("%s IOC Facts based allocation failed with error %d\n",
852 		    __func__, error);
853 	}
854 
855 	/*
856 	 * Mapping structures will be re-allocated after getting IOC Page8, so
857 	 * free these structures here.
858 	 */
859 	mps_mapping_exit(sc);
860 
861 	/*
862 	 * The static page function currently read is IOC Page8.  Others can be
863 	 * added in future.  It's possible that the values in IOC Page8 have
864 	 * changed after a Diag Reset due to user modification, so always read
865 	 * these.  Interrupts are masked, so unmask them before getting config
866 	 * pages.
867 	 */
868 	mps_unmask_intr(sc);
869 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
870 	mps_base_static_config_pages(sc);
871 
872 	/*
873 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
874 	 * mapping tables.
875 	 */
876 	mps_mapping_initialize(sc);
877 
878 	/*
879 	 * Restart will reload the event masks clobbered by the reset, and
880 	 * then enable the port.
881 	 */
882 	mps_reregister_events(sc);
883 
884 	/* the end of discovery will release the simq, so we're done. */
885 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
886 	    sc, sc->replypostindex, sc->replyfreeindex);
887 
888 	mpssas_release_simq_reinit(sassc);
889 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
890 
891 	return 0;
892 }
893 
894 /* Wait for the chip to ACK a word that we've put into its FIFO
895  * Wait for <timeout> seconds. In single loop wait for busy loop
896  * for 500 microseconds.
897  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
898  * */
899 static int
900 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
901 {
902 
903 	u32 cntdn, count;
904 	u32 int_status;
905 	u32 doorbell;
906 
907 	count = 0;
908 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
909 	do {
910 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
911 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
912 			mps_dprint(sc, MPS_TRACE,
913 			"%s: successful count(%d), timeout(%d)\n",
914 			__func__, count, timeout);
915 		return 0;
916 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
917 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
918 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
919 				MPI2_IOC_STATE_FAULT) {
920 				mps_dprint(sc, MPS_FAULT,
921 					"fault_state(0x%04x)!\n", doorbell);
922 				return (EFAULT);
923 			}
924 		} else if (int_status == 0xFFFFFFFF)
925 			goto out;
926 
927 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
928 		* 0.5 milisecond */
929 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
930 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
931 			"mpsdba", hz/1000);
932 		else if (sleep_flag == CAN_SLEEP)
933 			pause("mpsdba", hz/1000);
934 		else
935 			DELAY(500);
936 		count++;
937 	} while (--cntdn);
938 
939 	out:
940 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
941 		"int_status(%x)!\n", __func__, count, int_status);
942 	return (ETIMEDOUT);
943 
944 }
945 
946 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
947 static int
948 mps_wait_db_int(struct mps_softc *sc)
949 {
950 	int retry;
951 
952 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
953 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
954 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
955 			return (0);
956 		DELAY(2000);
957 	}
958 	return (ETIMEDOUT);
959 }
960 
961 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
962 static int
963 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
964     int req_sz, int reply_sz, int timeout)
965 {
966 	uint32_t *data32;
967 	uint16_t *data16;
968 	int i, count, ioc_sz, residual;
969 	int sleep_flags = CAN_SLEEP;
970 
971 	if (curthread->td_no_sleeping != 0)
972 		sleep_flags = NO_SLEEP;
973 
974 	/* Step 1 */
975 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
976 
977 	/* Step 2 */
978 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
979 		return (EBUSY);
980 
981 	/* Step 3
982 	 * Announce that a message is coming through the doorbell.  Messages
983 	 * are pushed at 32bit words, so round up if needed.
984 	 */
985 	count = (req_sz + 3) / 4;
986 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
987 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
988 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
989 
990 	/* Step 4 */
991 	if (mps_wait_db_int(sc) ||
992 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
993 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
994 		return (ENXIO);
995 	}
996 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
997 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
998 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
999 		return (ENXIO);
1000 	}
1001 
1002 	/* Step 5 */
1003 	/* Clock out the message data synchronously in 32-bit dwords*/
1004 	data32 = (uint32_t *)req;
1005 	for (i = 0; i < count; i++) {
1006 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1007 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
1008 			mps_dprint(sc, MPS_FAULT,
1009 			    "Timeout while writing doorbell\n");
1010 			return (ENXIO);
1011 		}
1012 	}
1013 
1014 	/* Step 6 */
1015 	/* Clock in the reply in 16-bit words.  The total length of the
1016 	 * message is always in the 4th byte, so clock out the first 2 words
1017 	 * manually, then loop the rest.
1018 	 */
1019 	data16 = (uint16_t *)reply;
1020 	if (mps_wait_db_int(sc) != 0) {
1021 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
1022 		return (ENXIO);
1023 	}
1024 	data16[0] =
1025 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1026 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1027 	if (mps_wait_db_int(sc) != 0) {
1028 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
1029 		return (ENXIO);
1030 	}
1031 	data16[1] =
1032 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1033 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1034 
1035 	/* Number of 32bit words in the message */
1036 	ioc_sz = reply->MsgLength;
1037 
1038 	/*
1039 	 * Figure out how many 16bit words to clock in without overrunning.
1040 	 * The precision loss with dividing reply_sz can safely be
1041 	 * ignored because the messages can only be multiples of 32bits.
1042 	 */
1043 	residual = 0;
1044 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1045 	if (count < ioc_sz * 2) {
1046 		residual = ioc_sz * 2 - count;
1047 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1048 		    "residual message words\n", residual);
1049 	}
1050 
1051 	for (i = 2; i < count; i++) {
1052 		if (mps_wait_db_int(sc) != 0) {
1053 			mps_dprint(sc, MPS_FAULT,
1054 			    "Timeout reading doorbell %d\n", i);
1055 			return (ENXIO);
1056 		}
1057 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1058 		    MPI2_DOORBELL_DATA_MASK;
1059 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1060 	}
1061 
1062 	/*
1063 	 * Pull out residual words that won't fit into the provided buffer.
1064 	 * This keeps the chip from hanging due to a driver programming
1065 	 * error.
1066 	 */
1067 	while (residual--) {
1068 		if (mps_wait_db_int(sc) != 0) {
1069 			mps_dprint(sc, MPS_FAULT,
1070 			    "Timeout reading doorbell\n");
1071 			return (ENXIO);
1072 		}
1073 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1074 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1075 	}
1076 
1077 	/* Step 7 */
1078 	if (mps_wait_db_int(sc) != 0) {
1079 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1080 		return (ENXIO);
1081 	}
1082 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1083 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1084 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1085 
1086 	return (0);
1087 }
1088 
1089 static void
1090 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1091 {
1092 	reply_descriptor rd;
1093 	MPS_FUNCTRACE(sc);
1094 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1095 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1096 
1097 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1098 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1099 
1100 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1101 		sc->io_cmds_highwater++;
1102 	rd.u.low = cm->cm_desc.Words.Low;
1103 	rd.u.high = cm->cm_desc.Words.High;
1104 	rd.word = htole64(rd.word);
1105 	/* TODO-We may need to make below regwrite atomic */
1106 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1107 	    rd.u.low);
1108 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1109 	    rd.u.high);
1110 }
1111 
1112 /*
1113  * Just the FACTS, ma'am.
1114  */
1115 static int
1116 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1117 {
1118 	MPI2_DEFAULT_REPLY *reply;
1119 	MPI2_IOC_FACTS_REQUEST request;
1120 	int error, req_sz, reply_sz;
1121 
1122 	MPS_FUNCTRACE(sc);
1123 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1124 
1125 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1126 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1127 	reply = (MPI2_DEFAULT_REPLY *)facts;
1128 
1129 	bzero(&request, req_sz);
1130 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1131 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1132 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1133 
1134 	return (error);
1135 }
1136 
1137 static int
1138 mps_send_iocinit(struct mps_softc *sc)
1139 {
1140 	MPI2_IOC_INIT_REQUEST	init;
1141 	MPI2_DEFAULT_REPLY	reply;
1142 	int req_sz, reply_sz, error;
1143 	struct timeval now;
1144 	uint64_t time_in_msec;
1145 
1146 	MPS_FUNCTRACE(sc);
1147 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1148 
1149 	/* Do a quick sanity check on proper initialization */
1150 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1151 	    || (sc->replyframesz == 0)) {
1152 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1153 		    "Driver not fully initialized for IOCInit\n");
1154 		return (EINVAL);
1155 	}
1156 
1157 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1158 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1159 	bzero(&init, req_sz);
1160 	bzero(&reply, reply_sz);
1161 
1162 	/*
1163 	 * Fill in the init block.  Note that most addresses are
1164 	 * deliberately in the lower 32bits of memory.  This is a micro-
1165 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1166 	 */
1167 	init.Function = MPI2_FUNCTION_IOC_INIT;
1168 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1169 	init.MsgVersion = htole16(MPI2_VERSION);
1170 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1171 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1172 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1173 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1174 	init.SenseBufferAddressHigh = 0;
1175 	init.SystemReplyAddressHigh = 0;
1176 	init.SystemRequestFrameBaseAddress.High = 0;
1177 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1178 	init.ReplyDescriptorPostQueueAddress.High = 0;
1179 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1180 	init.ReplyFreeQueueAddress.High = 0;
1181 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1182 	getmicrotime(&now);
1183 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1184 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1185 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1186 
1187 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1188 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1189 		error = ENXIO;
1190 
1191 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1192 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1193 	return (error);
1194 }
1195 
1196 void
1197 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1198 {
1199 	bus_addr_t *addr;
1200 
1201 	addr = arg;
1202 	*addr = segs[0].ds_addr;
1203 }
1204 
1205 void
1206 mps_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1207 {
1208 	struct mps_busdma_context *ctx;
1209 	int need_unload, need_free;
1210 
1211 	ctx = (struct mps_busdma_context *)arg;
1212 	need_unload = 0;
1213 	need_free = 0;
1214 
1215 	mps_lock(ctx->softc);
1216 	ctx->error = error;
1217 	ctx->completed = 1;
1218 	if ((error == 0) && (ctx->abandoned == 0)) {
1219 		*ctx->addr = segs[0].ds_addr;
1220 	} else {
1221 		if (nsegs != 0)
1222 			need_unload = 1;
1223 		if (ctx->abandoned != 0)
1224 			need_free = 1;
1225 	}
1226 	if (need_free == 0)
1227 		wakeup(ctx);
1228 
1229 	mps_unlock(ctx->softc);
1230 
1231 	if (need_unload != 0) {
1232 		bus_dmamap_unload(ctx->buffer_dmat,
1233 				  ctx->buffer_dmamap);
1234 		*ctx->addr = 0;
1235 	}
1236 
1237 	if (need_free != 0)
1238 		free(ctx, M_MPSUSER);
1239 }
1240 
1241 static int
1242 mps_alloc_queues(struct mps_softc *sc)
1243 {
1244 	struct mps_queue *q;
1245 	u_int nq, i;
1246 
1247 	nq = sc->msi_msgs;
1248 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1249 
1250 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1251 	    M_NOWAIT|M_ZERO);
1252 	if (sc->queues == NULL)
1253 		return (ENOMEM);
1254 
1255 	for (i = 0; i < nq; i++) {
1256 		q = &sc->queues[i];
1257 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1258 		q->sc = sc;
1259 		q->qnum = i;
1260 	}
1261 
1262 	return (0);
1263 }
1264 
1265 static int
1266 mps_alloc_hw_queues(struct mps_softc *sc)
1267 {
1268 	bus_addr_t queues_busaddr;
1269 	uint8_t *queues;
1270 	int qsize, fqsize, pqsize;
1271 
1272 	/*
1273 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1274 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1275 	 * This queue supplies fresh reply frames for the firmware to use.
1276 	 *
1277 	 * The reply descriptor post queue contains 8 byte entries in
1278 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1279 	 * contains filled-in reply frames sent from the firmware to the host.
1280 	 *
1281 	 * These two queues are allocated together for simplicity.
1282 	 */
1283 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1284 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1285 	fqsize= sc->fqdepth * 4;
1286 	pqsize = sc->pqdepth * 8;
1287 	qsize = fqsize + pqsize;
1288 
1289         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1290 				16, 0,			/* algnmnt, boundary */
1291 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1292 				BUS_SPACE_MAXADDR,	/* highaddr */
1293 				NULL, NULL,		/* filter, filterarg */
1294                                 qsize,			/* maxsize */
1295                                 1,			/* nsegments */
1296                                 qsize,			/* maxsegsize */
1297                                 0,			/* flags */
1298                                 NULL, NULL,		/* lockfunc, lockarg */
1299                                 &sc->queues_dmat)) {
1300 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1301 		return (ENOMEM);
1302         }
1303         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1304 	    &sc->queues_map)) {
1305 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1306 		return (ENOMEM);
1307         }
1308         bzero(queues, qsize);
1309         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1310 	    mps_memaddr_cb, &queues_busaddr, 0);
1311 
1312 	sc->free_queue = (uint32_t *)queues;
1313 	sc->free_busaddr = queues_busaddr;
1314 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1315 	sc->post_busaddr = queues_busaddr + fqsize;
1316 	mps_dprint(sc, MPS_INIT, "free queue busaddr= %#016jx size= %d\n",
1317 	    (uintmax_t)sc->free_busaddr, fqsize);
1318 	mps_dprint(sc, MPS_INIT, "reply queue busaddr= %#016jx size= %d\n",
1319 	    (uintmax_t)sc->post_busaddr, pqsize);
1320 
1321 	return (0);
1322 }
1323 
1324 static int
1325 mps_alloc_replies(struct mps_softc *sc)
1326 {
1327 	int rsize, num_replies;
1328 
1329 	/* Store the reply frame size in bytes rather than as 32bit words */
1330 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1331 
1332 	/*
1333 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1334 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1335 	 * replies can be used at once.
1336 	 */
1337 	num_replies = max(sc->fqdepth, sc->num_replies);
1338 
1339 	rsize = sc->replyframesz * num_replies;
1340         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1341 				4, 0,			/* algnmnt, boundary */
1342 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1343 				BUS_SPACE_MAXADDR,	/* highaddr */
1344 				NULL, NULL,		/* filter, filterarg */
1345                                 rsize,			/* maxsize */
1346                                 1,			/* nsegments */
1347                                 rsize,			/* maxsegsize */
1348                                 0,			/* flags */
1349                                 NULL, NULL,		/* lockfunc, lockarg */
1350                                 &sc->reply_dmat)) {
1351 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1352 		return (ENOMEM);
1353         }
1354         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1355 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1356 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1357 		return (ENOMEM);
1358         }
1359         bzero(sc->reply_frames, rsize);
1360         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1361 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1362 
1363 	mps_dprint(sc, MPS_INIT, "reply frames busaddr= %#016jx size= %d\n",
1364 	    (uintmax_t)sc->reply_busaddr, rsize);
1365 
1366 	return (0);
1367 }
1368 
1369 static int
1370 mps_alloc_requests(struct mps_softc *sc)
1371 {
1372 	struct mps_command *cm;
1373 	struct mps_chain *chain;
1374 	int i, rsize, nsegs;
1375 
1376 	rsize = sc->reqframesz * sc->num_reqs;
1377         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1378 				16, 0,			/* algnmnt, boundary */
1379 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1380 				BUS_SPACE_MAXADDR,	/* highaddr */
1381 				NULL, NULL,		/* filter, filterarg */
1382                                 rsize,			/* maxsize */
1383                                 1,			/* nsegments */
1384                                 rsize,			/* maxsegsize */
1385                                 0,			/* flags */
1386                                 NULL, NULL,		/* lockfunc, lockarg */
1387                                 &sc->req_dmat)) {
1388 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1389 		return (ENOMEM);
1390         }
1391         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1392 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1393 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1394 		return (ENOMEM);
1395         }
1396         bzero(sc->req_frames, rsize);
1397         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1398 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1399 	mps_dprint(sc, MPS_INIT, "request frames busaddr= %#016jx size= %d\n",
1400 	    (uintmax_t)sc->req_busaddr, rsize);
1401 
1402 	rsize = sc->reqframesz * sc->num_chains;
1403         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1404 				16, 0,			/* algnmnt, boundary */
1405 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1406 				BUS_SPACE_MAXADDR,	/* highaddr */
1407 				NULL, NULL,		/* filter, filterarg */
1408                                 rsize,			/* maxsize */
1409                                 1,			/* nsegments */
1410                                 rsize,			/* maxsegsize */
1411                                 0,			/* flags */
1412                                 NULL, NULL,		/* lockfunc, lockarg */
1413                                 &sc->chain_dmat)) {
1414 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1415 		return (ENOMEM);
1416         }
1417         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1418 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1419 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1420 		return (ENOMEM);
1421         }
1422         bzero(sc->chain_frames, rsize);
1423         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1424 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1425 	mps_dprint(sc, MPS_INIT, "chain frames busaddr= %#016jx size= %d\n",
1426 	    (uintmax_t)sc->chain_busaddr, rsize);
1427 
1428 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1429         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1430 				1, 0,			/* algnmnt, boundary */
1431 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1432 				BUS_SPACE_MAXADDR,	/* highaddr */
1433 				NULL, NULL,		/* filter, filterarg */
1434                                 rsize,			/* maxsize */
1435                                 1,			/* nsegments */
1436                                 rsize,			/* maxsegsize */
1437                                 0,			/* flags */
1438                                 NULL, NULL,		/* lockfunc, lockarg */
1439                                 &sc->sense_dmat)) {
1440 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1441 		return (ENOMEM);
1442         }
1443         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1444 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1445 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1446 		return (ENOMEM);
1447         }
1448         bzero(sc->sense_frames, rsize);
1449         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1450 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1451 	mps_dprint(sc, MPS_INIT, "sense frames busaddr= %#016jx size= %d\n",
1452 	    (uintmax_t)sc->sense_busaddr, rsize);
1453 
1454 	sc->chains = malloc(sizeof(struct mps_chain) * sc->num_chains, M_MPT2,
1455 	    M_WAITOK | M_ZERO);
1456 	if(!sc->chains) {
1457 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1458 		return (ENOMEM);
1459 	}
1460 	for (i = 0; i < sc->num_chains; i++) {
1461 		chain = &sc->chains[i];
1462 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1463 		    i * sc->reqframesz);
1464 		chain->chain_busaddr = sc->chain_busaddr +
1465 		    i * sc->reqframesz;
1466 		mps_free_chain(sc, chain);
1467 		sc->chain_free_lowwater++;
1468 	}
1469 
1470 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1471         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1472 				1, 0,			/* algnmnt, boundary */
1473 				BUS_SPACE_MAXADDR,	/* lowaddr */
1474 				BUS_SPACE_MAXADDR,	/* highaddr */
1475 				NULL, NULL,		/* filter, filterarg */
1476                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1477                                 nsegs,			/* nsegments */
1478                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1479                                 BUS_DMA_ALLOCNOW,	/* flags */
1480                                 busdma_lock_mutex,	/* lockfunc */
1481 				&sc->mps_mtx,		/* lockarg */
1482                                 &sc->buffer_dmat)) {
1483 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1484 		return (ENOMEM);
1485         }
1486 
1487 	/*
1488 	 * SMID 0 cannot be used as a free command per the firmware spec.
1489 	 * Just drop that command instead of risking accounting bugs.
1490 	 */
1491 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1492 	    M_MPT2, M_WAITOK | M_ZERO);
1493 	if(!sc->commands) {
1494 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1495 		return (ENOMEM);
1496 	}
1497 	for (i = 1; i < sc->num_reqs; i++) {
1498 		cm = &sc->commands[i];
1499 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1500 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1501 		cm->cm_sense = &sc->sense_frames[i];
1502 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1503 		cm->cm_desc.Default.SMID = i;
1504 		cm->cm_sc = sc;
1505 		TAILQ_INIT(&cm->cm_chain_list);
1506 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1507 
1508 		/* XXX Is a failure here a critical problem? */
1509 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1510 			if (i <= sc->num_prireqs)
1511 				mps_free_high_priority_command(sc, cm);
1512 			else
1513 				mps_free_command(sc, cm);
1514 		else {
1515 			panic("failed to allocate command %d\n", i);
1516 			sc->num_reqs = i;
1517 			break;
1518 		}
1519 	}
1520 
1521 	return (0);
1522 }
1523 
1524 static int
1525 mps_init_queues(struct mps_softc *sc)
1526 {
1527 	int i;
1528 
1529 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1530 
1531 	/*
1532 	 * According to the spec, we need to use one less reply than we
1533 	 * have space for on the queue.  So sc->num_replies (the number we
1534 	 * use) should be less than sc->fqdepth (allocated size).
1535 	 */
1536 	if (sc->num_replies >= sc->fqdepth)
1537 		return (EINVAL);
1538 
1539 	/*
1540 	 * Initialize all of the free queue entries.
1541 	 */
1542 	for (i = 0; i < sc->fqdepth; i++)
1543 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1544 	sc->replyfreeindex = sc->num_replies;
1545 
1546 	return (0);
1547 }
1548 
1549 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1550  * Next are the global settings, if they exist.  Highest are the per-unit
1551  * settings, if they exist.
1552  */
1553 void
1554 mps_get_tunables(struct mps_softc *sc)
1555 {
1556 	char tmpstr[80], mps_debug[80];
1557 
1558 	/* XXX default to some debugging for now */
1559 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1560 	sc->disable_msix = 0;
1561 	sc->disable_msi = 0;
1562 	sc->max_msix = MPS_MSIX_MAX;
1563 	sc->max_chains = MPS_CHAIN_FRAMES;
1564 	sc->max_io_pages = MPS_MAXIO_PAGES;
1565 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1566 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1567 	sc->use_phynum = 1;
1568 	sc->max_reqframes = MPS_REQ_FRAMES;
1569 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1570 	sc->max_replyframes = MPS_REPLY_FRAMES;
1571 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1572 
1573 	/*
1574 	 * Grab the global variables.
1575 	 */
1576 	bzero(mps_debug, 80);
1577 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1578 		mps_parse_debug(sc, mps_debug);
1579 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1580 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1581 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1582 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1583 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1584 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1585 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1586 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1587 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1588 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1589 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1590 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1591 
1592 	/* Grab the unit-instance variables */
1593 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1594 	    device_get_unit(sc->mps_dev));
1595 	bzero(mps_debug, 80);
1596 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1597 		mps_parse_debug(sc, mps_debug);
1598 
1599 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1600 	    device_get_unit(sc->mps_dev));
1601 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1602 
1603 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1604 	    device_get_unit(sc->mps_dev));
1605 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1606 
1607 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1608 	    device_get_unit(sc->mps_dev));
1609 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1610 
1611 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1612 	    device_get_unit(sc->mps_dev));
1613 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1614 
1615 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1616 	    device_get_unit(sc->mps_dev));
1617 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1618 
1619 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1620 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1621 	    device_get_unit(sc->mps_dev));
1622 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1623 
1624 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1625 	    device_get_unit(sc->mps_dev));
1626 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1627 
1628 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1629 	    device_get_unit(sc->mps_dev));
1630 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1631 
1632 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1633 	    device_get_unit(sc->mps_dev));
1634 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1635 
1636 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1637 	    device_get_unit(sc->mps_dev));
1638 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1639 
1640 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1641 	    device_get_unit(sc->mps_dev));
1642 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1643 
1644 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1645 	    device_get_unit(sc->mps_dev));
1646 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1647 
1648 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1649 	    device_get_unit(sc->mps_dev));
1650 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1651 
1652 }
1653 
1654 static void
1655 mps_setup_sysctl(struct mps_softc *sc)
1656 {
1657 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1658 	struct sysctl_oid	*sysctl_tree = NULL;
1659 	char tmpstr[80], tmpstr2[80];
1660 
1661 	/*
1662 	 * Setup the sysctl variable so the user can change the debug level
1663 	 * on the fly.
1664 	 */
1665 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1666 	    device_get_unit(sc->mps_dev));
1667 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1668 
1669 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1670 	if (sysctl_ctx != NULL)
1671 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1672 
1673 	if (sysctl_tree == NULL) {
1674 		sysctl_ctx_init(&sc->sysctl_ctx);
1675 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1676 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1677 		    CTLFLAG_RD, 0, tmpstr);
1678 		if (sc->sysctl_tree == NULL)
1679 			return;
1680 		sysctl_ctx = &sc->sysctl_ctx;
1681 		sysctl_tree = sc->sysctl_tree;
1682 	}
1683 
1684 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1685 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1686 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1687 
1688 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1689 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1690 	    "Disable the use of MSI-X interrupts");
1691 
1692 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1693 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1694 	    "Disable the use of MSI interrupts");
1695 
1696 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1697 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1698 	    "User-defined maximum number of MSIX queues");
1699 
1700 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1701 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1702 	    "Negotiated number of MSIX queues");
1703 
1704 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1705 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1706 	    "Total number of allocated request frames");
1707 
1708 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1709 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1710 	    "Total number of allocated high priority request frames");
1711 
1712 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1713 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1714 	    "Total number of allocated reply frames");
1715 
1716 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1717 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1718 	    "Total number of event frames allocated");
1719 
1720 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1721 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1722 	    strlen(sc->fw_version), "firmware version");
1723 
1724 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1725 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1726 	    strlen(MPS_DRIVER_VERSION), "driver version");
1727 
1728 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1729 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1730 	    &sc->io_cmds_active, 0, "number of currently active commands");
1731 
1732 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1733 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1734 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1735 
1736 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1737 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1738 	    &sc->chain_free, 0, "number of free chain elements");
1739 
1740 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1741 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1742 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1743 
1744 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1745 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1746 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1747 
1748 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1749 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1750 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1751 	    "IOCFacts)");
1752 
1753 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1754 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1755 	    "enable SSU to SATA SSD/HDD at shutdown");
1756 
1757 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1758 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1759 	    &sc->chain_alloc_fail, "chain allocation failures");
1760 
1761 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1762 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1763 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1764 	    "spinup after SATA ID error");
1765 
1766 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1767 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1768 	    mps_mapping_dump, "A", "Mapping Table Dump");
1769 
1770 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1771 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1772 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1773 
1774 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1775 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1776 	    "Use the phy number for enumeration");
1777 }
1778 
1779 static struct mps_debug_string {
1780 	char	*name;
1781 	int	flag;
1782 } mps_debug_strings[] = {
1783 	{"info", MPS_INFO},
1784 	{"fault", MPS_FAULT},
1785 	{"event", MPS_EVENT},
1786 	{"log", MPS_LOG},
1787 	{"recovery", MPS_RECOVERY},
1788 	{"error", MPS_ERROR},
1789 	{"init", MPS_INIT},
1790 	{"xinfo", MPS_XINFO},
1791 	{"user", MPS_USER},
1792 	{"mapping", MPS_MAPPING},
1793 	{"trace", MPS_TRACE}
1794 };
1795 
1796 enum mps_debug_level_combiner {
1797 	COMB_NONE,
1798 	COMB_ADD,
1799 	COMB_SUB
1800 };
1801 
1802 static int
1803 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1804 {
1805 	struct mps_softc *sc;
1806 	struct mps_debug_string *string;
1807 	struct sbuf *sbuf;
1808 	char *buffer;
1809 	size_t sz;
1810 	int i, len, debug, error;
1811 
1812 	sc = (struct mps_softc *)arg1;
1813 
1814 	error = sysctl_wire_old_buffer(req, 0);
1815 	if (error != 0)
1816 		return (error);
1817 
1818 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1819 	debug = sc->mps_debug;
1820 
1821 	sbuf_printf(sbuf, "%#x", debug);
1822 
1823 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1824 	for (i = 0; i < sz; i++) {
1825 		string = &mps_debug_strings[i];
1826 		if (debug & string->flag)
1827 			sbuf_printf(sbuf, ",%s", string->name);
1828 	}
1829 
1830 	error = sbuf_finish(sbuf);
1831 	sbuf_delete(sbuf);
1832 
1833 	if (error || req->newptr == NULL)
1834 		return (error);
1835 
1836 	len = req->newlen - req->newidx;
1837 	if (len == 0)
1838 		return (0);
1839 
1840 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1841 	error = SYSCTL_IN(req, buffer, len);
1842 
1843 	mps_parse_debug(sc, buffer);
1844 
1845 	free(buffer, M_MPT2);
1846 	return (error);
1847 }
1848 
1849 static void
1850 mps_parse_debug(struct mps_softc *sc, char *list)
1851 {
1852 	struct mps_debug_string *string;
1853 	enum mps_debug_level_combiner op;
1854 	char *token, *endtoken;
1855 	size_t sz;
1856 	int flags, i;
1857 
1858 	if (list == NULL || *list == '\0')
1859 		return;
1860 
1861 	if (*list == '+') {
1862 		op = COMB_ADD;
1863 		list++;
1864 	} else if (*list == '-') {
1865 		op = COMB_SUB;
1866 		list++;
1867 	} else
1868 		op = COMB_NONE;
1869 	if (*list == '\0')
1870 		return;
1871 
1872 	flags = 0;
1873 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1874 	while ((token = strsep(&list, ":,")) != NULL) {
1875 
1876 		/* Handle integer flags */
1877 		flags |= strtol(token, &endtoken, 0);
1878 		if (token != endtoken)
1879 			continue;
1880 
1881 		/* Handle text flags */
1882 		for (i = 0; i < sz; i++) {
1883 			string = &mps_debug_strings[i];
1884 			if (strcasecmp(token, string->name) == 0) {
1885 				flags |= string->flag;
1886 				break;
1887 			}
1888 		}
1889 	}
1890 
1891 	switch (op) {
1892 	case COMB_NONE:
1893 		sc->mps_debug = flags;
1894 		break;
1895 	case COMB_ADD:
1896 		sc->mps_debug |= flags;
1897 		break;
1898 	case COMB_SUB:
1899 		sc->mps_debug &= (~flags);
1900 		break;
1901 	}
1902 
1903 	return;
1904 }
1905 
1906 int
1907 mps_attach(struct mps_softc *sc)
1908 {
1909 	int error;
1910 
1911 	MPS_FUNCTRACE(sc);
1912 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1913 
1914 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1915 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1916 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1917 	TAILQ_INIT(&sc->event_list);
1918 	timevalclear(&sc->lastfail);
1919 
1920 	if ((error = mps_transition_ready(sc)) != 0) {
1921 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1922 		    "ready\n");
1923 		return (error);
1924 	}
1925 
1926 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1927 	    M_ZERO|M_NOWAIT);
1928 	if(!sc->facts) {
1929 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1930 		    "exit\n");
1931 		return (ENOMEM);
1932 	}
1933 
1934 	/*
1935 	 * Get IOC Facts and allocate all structures based on this information.
1936 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1937 	 * Facts. If relevant values have changed in IOC Facts, this function
1938 	 * will free all of the memory based on IOC Facts and reallocate that
1939 	 * memory.  If this fails, any allocated memory should already be freed.
1940 	 */
1941 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1942 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1943 		    "failed with error %d, exit\n", error);
1944 		return (error);
1945 	}
1946 
1947 	/* Start the periodic watchdog check on the IOC Doorbell */
1948 	mps_periodic(sc);
1949 
1950 	/*
1951 	 * The portenable will kick off discovery events that will drive the
1952 	 * rest of the initialization process.  The CAM/SAS module will
1953 	 * hold up the boot sequence until discovery is complete.
1954 	 */
1955 	sc->mps_ich.ich_func = mps_startup;
1956 	sc->mps_ich.ich_arg = sc;
1957 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1958 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1959 		    "Cannot establish MPS config hook\n");
1960 		error = EINVAL;
1961 	}
1962 
1963 	/*
1964 	 * Allow IR to shutdown gracefully when shutdown occurs.
1965 	 */
1966 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1967 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1968 
1969 	if (sc->shutdown_eh == NULL)
1970 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1971 		    "shutdown event registration failed\n");
1972 
1973 	mps_setup_sysctl(sc);
1974 
1975 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1976 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1977 
1978 	return (error);
1979 }
1980 
1981 /* Run through any late-start handlers. */
1982 static void
1983 mps_startup(void *arg)
1984 {
1985 	struct mps_softc *sc;
1986 
1987 	sc = (struct mps_softc *)arg;
1988 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1989 
1990 	mps_lock(sc);
1991 	mps_unmask_intr(sc);
1992 
1993 	/* initialize device mapping tables */
1994 	mps_base_static_config_pages(sc);
1995 	mps_mapping_initialize(sc);
1996 	mpssas_startup(sc);
1997 	mps_unlock(sc);
1998 
1999 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
2000 	config_intrhook_disestablish(&sc->mps_ich);
2001 	sc->mps_ich.ich_arg = NULL;
2002 
2003 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2004 }
2005 
2006 /* Periodic watchdog.  Is called with the driver lock already held. */
2007 static void
2008 mps_periodic(void *arg)
2009 {
2010 	struct mps_softc *sc;
2011 	uint32_t db;
2012 
2013 	sc = (struct mps_softc *)arg;
2014 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
2015 		return;
2016 
2017 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
2018 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2019 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2020 		mps_reinit(sc);
2021 	}
2022 
2023 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
2024 }
2025 
2026 static void
2027 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
2028     MPI2_EVENT_NOTIFICATION_REPLY *event)
2029 {
2030 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2031 
2032 	MPS_DPRINT_EVENT(sc, generic, event);
2033 
2034 	switch (event->Event) {
2035 	case MPI2_EVENT_LOG_DATA:
2036 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2037 		if (sc->mps_debug & MPS_EVENT)
2038 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
2039 		break;
2040 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2041 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2042 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2043 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2044 		     entry->LogSequence);
2045 		break;
2046 	default:
2047 		break;
2048 	}
2049 	return;
2050 }
2051 
2052 static int
2053 mps_attach_log(struct mps_softc *sc)
2054 {
2055 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
2056 
2057 	bzero(events, 16);
2058 	setbit(events, MPI2_EVENT_LOG_DATA);
2059 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2060 
2061 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
2062 	    &sc->mps_log_eh);
2063 
2064 	return (0);
2065 }
2066 
2067 static int
2068 mps_detach_log(struct mps_softc *sc)
2069 {
2070 
2071 	if (sc->mps_log_eh != NULL)
2072 		mps_deregister_events(sc, sc->mps_log_eh);
2073 	return (0);
2074 }
2075 
2076 /*
2077  * Free all of the driver resources and detach submodules.  Should be called
2078  * without the lock held.
2079  */
2080 int
2081 mps_free(struct mps_softc *sc)
2082 {
2083 	int error;
2084 
2085 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2086 	/* Turn off the watchdog */
2087 	mps_lock(sc);
2088 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
2089 	mps_unlock(sc);
2090 	/* Lock must not be held for this */
2091 	callout_drain(&sc->periodic);
2092 	callout_drain(&sc->device_check_callout);
2093 
2094 	if (((error = mps_detach_log(sc)) != 0) ||
2095 	    ((error = mps_detach_sas(sc)) != 0)) {
2096 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2097 		    "subsystems, exit\n");
2098 		return (error);
2099 	}
2100 
2101 	mps_detach_user(sc);
2102 
2103 	/* Put the IOC back in the READY state. */
2104 	mps_lock(sc);
2105 	if ((error = mps_transition_ready(sc)) != 0) {
2106 		mps_unlock(sc);
2107 		return (error);
2108 	}
2109 	mps_unlock(sc);
2110 
2111 	if (sc->facts != NULL)
2112 		free(sc->facts, M_MPT2);
2113 
2114 	/*
2115 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2116 	 * to free these buffers too.
2117 	 */
2118 	mps_iocfacts_free(sc);
2119 
2120 	if (sc->sysctl_tree != NULL)
2121 		sysctl_ctx_free(&sc->sysctl_ctx);
2122 
2123 	/* Deregister the shutdown function */
2124 	if (sc->shutdown_eh != NULL)
2125 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2126 
2127 	mtx_destroy(&sc->mps_mtx);
2128 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2129 
2130 	return (0);
2131 }
2132 
2133 static __inline void
2134 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2135 {
2136 	MPS_FUNCTRACE(sc);
2137 
2138 	if (cm == NULL) {
2139 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2140 		return;
2141 	}
2142 
2143 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2144 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2145 
2146 	if (cm->cm_complete != NULL) {
2147 		mps_dprint(sc, MPS_TRACE,
2148 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2149 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2150 			   cm->cm_reply);
2151 		cm->cm_complete(sc, cm);
2152 	}
2153 
2154 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2155 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2156 		wakeup(cm);
2157 	}
2158 
2159 	if (cm->cm_sc->io_cmds_active != 0) {
2160 		cm->cm_sc->io_cmds_active--;
2161 	} else {
2162 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2163 		    "out of sync - resynching to 0\n");
2164 	}
2165 }
2166 
2167 
2168 static void
2169 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2170 {
2171 	union loginfo_type {
2172 		u32     loginfo;
2173 		struct {
2174 			u32     subcode:16;
2175 			u32     code:8;
2176 			u32     originator:4;
2177 			u32     bus_type:4;
2178 		} dw;
2179 	};
2180 	union loginfo_type sas_loginfo;
2181 	char *originator_str = NULL;
2182 
2183 	sas_loginfo.loginfo = log_info;
2184 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2185 		return;
2186 
2187 	/* each nexus loss loginfo */
2188 	if (log_info == 0x31170000)
2189 		return;
2190 
2191 	/* eat the loginfos associated with task aborts */
2192 	if ((log_info == 30050000 || log_info ==
2193 	    0x31140000 || log_info == 0x31130000))
2194 		return;
2195 
2196 	switch (sas_loginfo.dw.originator) {
2197 	case 0:
2198 		originator_str = "IOP";
2199 		break;
2200 	case 1:
2201 		originator_str = "PL";
2202 		break;
2203 	case 2:
2204 		originator_str = "IR";
2205 		break;
2206 }
2207 
2208 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2209 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2210 	originator_str, sas_loginfo.dw.code,
2211 	sas_loginfo.dw.subcode);
2212 }
2213 
2214 static void
2215 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2216 {
2217 	MPI2DefaultReply_t *mpi_reply;
2218 	u16 sc_status;
2219 
2220 	mpi_reply = (MPI2DefaultReply_t*)reply;
2221 	sc_status = le16toh(mpi_reply->IOCStatus);
2222 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2223 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2224 }
2225 void
2226 mps_intr(void *data)
2227 {
2228 	struct mps_softc *sc;
2229 	uint32_t status;
2230 
2231 	sc = (struct mps_softc *)data;
2232 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2233 
2234 	/*
2235 	 * Check interrupt status register to flush the bus.  This is
2236 	 * needed for both INTx interrupts and driver-driven polling
2237 	 */
2238 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2239 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2240 		return;
2241 
2242 	mps_lock(sc);
2243 	mps_intr_locked(data);
2244 	mps_unlock(sc);
2245 	return;
2246 }
2247 
2248 /*
2249  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2250  * chip.  Hopefully this theory is correct.
2251  */
2252 void
2253 mps_intr_msi(void *data)
2254 {
2255 	struct mps_softc *sc;
2256 
2257 	sc = (struct mps_softc *)data;
2258 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2259 	mps_lock(sc);
2260 	mps_intr_locked(data);
2261 	mps_unlock(sc);
2262 	return;
2263 }
2264 
2265 /*
2266  * The locking is overly broad and simplistic, but easy to deal with for now.
2267  */
2268 void
2269 mps_intr_locked(void *data)
2270 {
2271 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2272 	struct mps_softc *sc;
2273 	struct mps_command *cm = NULL;
2274 	uint8_t flags;
2275 	u_int pq;
2276 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2277 	mps_fw_diagnostic_buffer_t *pBuffer;
2278 
2279 	sc = (struct mps_softc *)data;
2280 
2281 	pq = sc->replypostindex;
2282 	mps_dprint(sc, MPS_TRACE,
2283 	    "%s sc %p starting with replypostindex %u\n",
2284 	    __func__, sc, sc->replypostindex);
2285 
2286 	for ( ;; ) {
2287 		cm = NULL;
2288 		desc = &sc->post_queue[sc->replypostindex];
2289 		flags = desc->Default.ReplyFlags &
2290 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2291 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2292 		 || (le32toh(desc->Words.High) == 0xffffffff))
2293 			break;
2294 
2295 		/* increment the replypostindex now, so that event handlers
2296 		 * and cm completion handlers which decide to do a diag
2297 		 * reset can zero it without it getting incremented again
2298 		 * afterwards, and we break out of this loop on the next
2299 		 * iteration since the reply post queue has been cleared to
2300 		 * 0xFF and all descriptors look unused (which they are).
2301 		 */
2302 		if (++sc->replypostindex >= sc->pqdepth)
2303 			sc->replypostindex = 0;
2304 
2305 		switch (flags) {
2306 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2307 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2308 			cm->cm_reply = NULL;
2309 			break;
2310 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2311 		{
2312 			uint32_t baddr;
2313 			uint8_t *reply;
2314 
2315 			/*
2316 			 * Re-compose the reply address from the address
2317 			 * sent back from the chip.  The ReplyFrameAddress
2318 			 * is the lower 32 bits of the physical address of
2319 			 * particular reply frame.  Convert that address to
2320 			 * host format, and then use that to provide the
2321 			 * offset against the virtual address base
2322 			 * (sc->reply_frames).
2323 			 */
2324 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2325 			reply = sc->reply_frames +
2326 				(baddr - ((uint32_t)sc->reply_busaddr));
2327 			/*
2328 			 * Make sure the reply we got back is in a valid
2329 			 * range.  If not, go ahead and panic here, since
2330 			 * we'll probably panic as soon as we deference the
2331 			 * reply pointer anyway.
2332 			 */
2333 			if ((reply < sc->reply_frames)
2334 			 || (reply > (sc->reply_frames +
2335 			     (sc->fqdepth * sc->replyframesz)))) {
2336 				printf("%s: WARNING: reply %p out of range!\n",
2337 				       __func__, reply);
2338 				printf("%s: reply_frames %p, fqdepth %d, "
2339 				       "frame size %d\n", __func__,
2340 				       sc->reply_frames, sc->fqdepth,
2341 				       sc->replyframesz);
2342 				printf("%s: baddr %#x,\n", __func__, baddr);
2343 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
2344 				panic("Reply address out of range");
2345 			}
2346 			if (le16toh(desc->AddressReply.SMID) == 0) {
2347 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2348 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2349 					/*
2350 					 * If SMID is 0 for Diag Buffer Post,
2351 					 * this implies that the reply is due to
2352 					 * a release function with a status that
2353 					 * the buffer has been released.  Set
2354 					 * the buffer flags accordingly.
2355 					 */
2356 					rel_rep =
2357 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2358 					if ((le16toh(rel_rep->IOCStatus) &
2359 					    MPI2_IOCSTATUS_MASK) ==
2360 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2361 					{
2362 						pBuffer =
2363 						    &sc->fw_diag_buffer_list[
2364 						    rel_rep->BufferType];
2365 						pBuffer->valid_data = TRUE;
2366 						pBuffer->owned_by_firmware =
2367 						    FALSE;
2368 						pBuffer->immediate = FALSE;
2369 					}
2370 				} else
2371 					mps_dispatch_event(sc, baddr,
2372 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2373 					    reply);
2374 			} else {
2375 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2376 				cm->cm_reply = reply;
2377 				cm->cm_reply_data =
2378 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2379 			}
2380 			break;
2381 		}
2382 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2383 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2384 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2385 		default:
2386 			/* Unhandled */
2387 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2388 			    desc->Default.ReplyFlags);
2389 			cm = NULL;
2390 			break;
2391 		}
2392 
2393 
2394 		if (cm != NULL) {
2395 			// Print Error reply frame
2396 			if (cm->cm_reply)
2397 				mps_display_reply_info(sc,cm->cm_reply);
2398 			mps_complete_command(sc, cm);
2399 		}
2400 
2401 		desc->Words.Low = 0xffffffff;
2402 		desc->Words.High = 0xffffffff;
2403 	}
2404 
2405 	if (pq != sc->replypostindex) {
2406 		mps_dprint(sc, MPS_TRACE,
2407 		    "%s sc %p writing postindex %d\n",
2408 		    __func__, sc, sc->replypostindex);
2409 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2410 	}
2411 
2412 	return;
2413 }
2414 
2415 static void
2416 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2417     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2418 {
2419 	struct mps_event_handle *eh;
2420 	int event, handled = 0;
2421 
2422 	event = le16toh(reply->Event);
2423 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2424 		if (isset(eh->mask, event)) {
2425 			eh->callback(sc, data, reply);
2426 			handled++;
2427 		}
2428 	}
2429 
2430 	if (handled == 0)
2431 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2432 
2433 	/*
2434 	 * This is the only place that the event/reply should be freed.
2435 	 * Anything wanting to hold onto the event data should have
2436 	 * already copied it into their own storage.
2437 	 */
2438 	mps_free_reply(sc, data);
2439 }
2440 
2441 static void
2442 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2443 {
2444 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2445 
2446 	if (cm->cm_reply)
2447 		MPS_DPRINT_EVENT(sc, generic,
2448 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2449 
2450 	mps_free_command(sc, cm);
2451 
2452 	/* next, send a port enable */
2453 	mpssas_startup(sc);
2454 }
2455 
2456 /*
2457  * For both register_events and update_events, the caller supplies a bitmap
2458  * of events that it _wants_.  These functions then turn that into a bitmask
2459  * suitable for the controller.
2460  */
2461 int
2462 mps_register_events(struct mps_softc *sc, u32 *mask,
2463     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2464 {
2465 	struct mps_event_handle *eh;
2466 	int error = 0;
2467 
2468 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2469 	if(!eh) {
2470 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2471 		return (ENOMEM);
2472 	}
2473 	eh->callback = cb;
2474 	eh->data = data;
2475 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2476 	if (mask != NULL)
2477 		error = mps_update_events(sc, eh, mask);
2478 	*handle = eh;
2479 
2480 	return (error);
2481 }
2482 
2483 int
2484 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2485     u32 *mask)
2486 {
2487 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2488 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2489 	struct mps_command *cm;
2490 	int error, i;
2491 
2492 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2493 
2494 	if ((mask != NULL) && (handle != NULL))
2495 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2496 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2497 
2498 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2499 		sc->event_mask[i] = -1;
2500 
2501 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2502 		sc->event_mask[i] &= ~handle->mask[i];
2503 
2504 
2505 	if ((cm = mps_alloc_command(sc)) == NULL)
2506 		return (EBUSY);
2507 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2508 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2509 	evtreq->MsgFlags = 0;
2510 	evtreq->SASBroadcastPrimitiveMasks = 0;
2511 #ifdef MPS_DEBUG_ALL_EVENTS
2512 	{
2513 		u_char fullmask[16];
2514 		memset(fullmask, 0x00, 16);
2515 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2516 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2517 	}
2518 #else
2519         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2520                 evtreq->EventMasks[i] =
2521                     htole32(sc->event_mask[i]);
2522 #endif
2523 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2524 	cm->cm_data = NULL;
2525 
2526 	error = mps_wait_command(sc, &cm, 60, 0);
2527 	if (cm != NULL)
2528 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2529 	if ((reply == NULL) ||
2530 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2531 		error = ENXIO;
2532 
2533 	if (reply)
2534 		MPS_DPRINT_EVENT(sc, generic, reply);
2535 
2536 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2537 
2538 	if (cm != NULL)
2539 		mps_free_command(sc, cm);
2540 	return (error);
2541 }
2542 
2543 static int
2544 mps_reregister_events(struct mps_softc *sc)
2545 {
2546 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2547 	struct mps_command *cm;
2548 	struct mps_event_handle *eh;
2549 	int error, i;
2550 
2551 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2552 
2553 	/* first, reregister events */
2554 
2555 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2556 		sc->event_mask[i] = -1;
2557 
2558 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2559 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2560 			sc->event_mask[i] &= ~eh->mask[i];
2561 	}
2562 
2563 	if ((cm = mps_alloc_command(sc)) == NULL)
2564 		return (EBUSY);
2565 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2566 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2567 	evtreq->MsgFlags = 0;
2568 	evtreq->SASBroadcastPrimitiveMasks = 0;
2569 #ifdef MPS_DEBUG_ALL_EVENTS
2570 	{
2571 		u_char fullmask[16];
2572 		memset(fullmask, 0x00, 16);
2573 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2574 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2575 	}
2576 #else
2577         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2578                 evtreq->EventMasks[i] =
2579                     htole32(sc->event_mask[i]);
2580 #endif
2581 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2582 	cm->cm_data = NULL;
2583 	cm->cm_complete = mps_reregister_events_complete;
2584 
2585 	error = mps_map_command(sc, cm);
2586 
2587 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2588 	    error);
2589 	return (error);
2590 }
2591 
2592 void
2593 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2594 {
2595 
2596 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2597 	free(handle, M_MPT2);
2598 }
2599 
2600 /*
2601  * Add a chain element as the next SGE for the specified command.
2602  * Reset cm_sge and cm_sgesize to indicate all the available space.
2603  */
2604 static int
2605 mps_add_chain(struct mps_command *cm)
2606 {
2607 	MPI2_SGE_CHAIN32 *sgc;
2608 	struct mps_chain *chain;
2609 	u_int space;
2610 
2611 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2612 		panic("MPS: Need SGE Error Code\n");
2613 
2614 	chain = mps_alloc_chain(cm->cm_sc);
2615 	if (chain == NULL)
2616 		return (ENOBUFS);
2617 
2618 	space = cm->cm_sc->reqframesz;
2619 
2620 	/*
2621 	 * Note: a double-linked list is used to make it easier to
2622 	 * walk for debugging.
2623 	 */
2624 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2625 
2626 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2627 	sgc->Length = htole16(space);
2628 	sgc->NextChainOffset = 0;
2629 	/* TODO Looks like bug in Setting sgc->Flags.
2630 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2631 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2632 	 *	This is fine.. because we are not using simple element. In case of
2633 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2634  	 */
2635 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2636 	sgc->Address = htole32(chain->chain_busaddr);
2637 
2638 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2639 	cm->cm_sglsize = space;
2640 	return (0);
2641 }
2642 
2643 /*
2644  * Add one scatter-gather element (chain, simple, transaction context)
2645  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2646  * cm_sge as the remaining size and pointer to the next SGE to fill
2647  * in, respectively.
2648  */
2649 int
2650 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2651 {
2652 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2653 	MPI2_SGE_SIMPLE64 *sge = sgep;
2654 	int error, type;
2655 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2656 
2657 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2658 
2659 #ifdef INVARIANTS
2660 	switch (type) {
2661 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2662 		if (len != tc->DetailsLength + 4)
2663 			panic("TC %p length %u or %zu?", tc,
2664 			    tc->DetailsLength + 4, len);
2665 		}
2666 		break;
2667 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2668 		/* Driver only uses 32-bit chain elements */
2669 		if (len != MPS_SGC_SIZE)
2670 			panic("CHAIN %p length %u or %zu?", sgep,
2671 			    MPS_SGC_SIZE, len);
2672 		break;
2673 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2674 		/* Driver only uses 64-bit SGE simple elements */
2675 		if (len != MPS_SGE64_SIZE)
2676 			panic("SGE simple %p length %u or %zu?", sge,
2677 			    MPS_SGE64_SIZE, len);
2678 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2679 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2680 			panic("SGE simple %p not marked 64-bit?", sge);
2681 
2682 		break;
2683 	default:
2684 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2685 	}
2686 #endif
2687 
2688 	/*
2689 	 * case 1: 1 more segment, enough room for it
2690 	 * case 2: 2 more segments, enough room for both
2691 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2692 	 * case 4: >=1 more segment, enough room for only a chain
2693 	 * case 5: >=1 more segment, no room for anything (error)
2694          */
2695 
2696 	/*
2697 	 * There should be room for at least a chain element, or this
2698 	 * code is buggy.  Case (5).
2699 	 */
2700 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2701 		panic("MPS: Need SGE Error Code\n");
2702 
2703 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2704 		/*
2705 		 * 1 or more segment, enough room for only a chain.
2706 		 * Hope the previous element wasn't a Simple entry
2707 		 * that needed to be marked with
2708 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2709 		 */
2710 		if ((error = mps_add_chain(cm)) != 0)
2711 			return (error);
2712 	}
2713 
2714 	if (segsleft >= 2 &&
2715 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2716 		/*
2717 		 * There are 2 or more segments left to add, and only
2718 		 * enough room for 1 and a chain.  Case (3).
2719 		 *
2720 		 * Mark as last element in this chain if necessary.
2721 		 */
2722 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2723 			sge->FlagsLength |= htole32(
2724 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2725 		}
2726 
2727 		/*
2728 		 * Add the item then a chain.  Do the chain now,
2729 		 * rather than on the next iteration, to simplify
2730 		 * understanding the code.
2731 		 */
2732 		cm->cm_sglsize -= len;
2733 		bcopy(sgep, cm->cm_sge, len);
2734 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2735 		return (mps_add_chain(cm));
2736 	}
2737 
2738 #ifdef INVARIANTS
2739 	/* Case 1: 1 more segment, enough room for it. */
2740 	if (segsleft == 1 && cm->cm_sglsize < len)
2741 		panic("1 seg left and no room? %u versus %zu",
2742 		    cm->cm_sglsize, len);
2743 
2744 	/* Case 2: 2 more segments, enough room for both */
2745 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2746 		panic("2 segs left and no room? %u versus %zu",
2747 		    cm->cm_sglsize, len);
2748 #endif
2749 
2750 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2751 		/*
2752 		 * If this is a bi-directional request, need to account for that
2753 		 * here.  Save the pre-filled sge values.  These will be used
2754 		 * either for the 2nd SGL or for a single direction SGL.  If
2755 		 * cm_out_len is non-zero, this is a bi-directional request, so
2756 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2757 		 * fill in the IN SGL.  Note that at this time, when filling in
2758 		 * 2 SGL's for a bi-directional request, they both use the same
2759 		 * DMA buffer (same cm command).
2760 		 */
2761 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2762 		saved_address_low = sge->Address.Low;
2763 		saved_address_high = sge->Address.High;
2764 		if (cm->cm_out_len) {
2765 			sge->FlagsLength = htole32(cm->cm_out_len |
2766 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2767 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2768 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2769 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2770 			    MPI2_SGE_FLAGS_SHIFT));
2771 			cm->cm_sglsize -= len;
2772 			bcopy(sgep, cm->cm_sge, len);
2773 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2774 			    + len);
2775 		}
2776 		saved_buf_len |=
2777 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2778 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2779 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2780 		    MPI2_SGE_FLAGS_END_OF_LIST |
2781 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2782 		    MPI2_SGE_FLAGS_SHIFT);
2783 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2784 			saved_buf_len |=
2785 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2786 			    MPI2_SGE_FLAGS_SHIFT);
2787 		} else {
2788 			saved_buf_len |=
2789 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2790 			    MPI2_SGE_FLAGS_SHIFT);
2791 		}
2792 		sge->FlagsLength = htole32(saved_buf_len);
2793 		sge->Address.Low = saved_address_low;
2794 		sge->Address.High = saved_address_high;
2795 	}
2796 
2797 	cm->cm_sglsize -= len;
2798 	bcopy(sgep, cm->cm_sge, len);
2799 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2800 	return (0);
2801 }
2802 
2803 /*
2804  * Add one dma segment to the scatter-gather list for a command.
2805  */
2806 int
2807 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2808     int segsleft)
2809 {
2810 	MPI2_SGE_SIMPLE64 sge;
2811 
2812 	/*
2813 	 * This driver always uses 64-bit address elements for simplicity.
2814 	 */
2815 	bzero(&sge, sizeof(sge));
2816 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2817 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2818 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2819 	mps_from_u64(pa, &sge.Address);
2820 
2821 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2822 }
2823 
2824 static void
2825 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2826 {
2827 	struct mps_softc *sc;
2828 	struct mps_command *cm;
2829 	u_int i, dir, sflags;
2830 
2831 	cm = (struct mps_command *)arg;
2832 	sc = cm->cm_sc;
2833 
2834 	/*
2835 	 * In this case, just print out a warning and let the chip tell the
2836 	 * user they did the wrong thing.
2837 	 */
2838 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2839 		mps_dprint(sc, MPS_ERROR,
2840 			   "%s: warning: busdma returned %d segments, "
2841 			   "more than the %d allowed\n", __func__, nsegs,
2842 			   cm->cm_max_segs);
2843 	}
2844 
2845 	/*
2846 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2847 	 * here.  In that case, both direction flags will be set.
2848 	 */
2849 	sflags = 0;
2850 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2851 		/*
2852 		 * We have to add a special case for SMP passthrough, there
2853 		 * is no easy way to generically handle it.  The first
2854 		 * S/G element is used for the command (therefore the
2855 		 * direction bit needs to be set).  The second one is used
2856 		 * for the reply.  We'll leave it to the caller to make
2857 		 * sure we only have two buffers.
2858 		 */
2859 		/*
2860 		 * Even though the busdma man page says it doesn't make
2861 		 * sense to have both direction flags, it does in this case.
2862 		 * We have one s/g element being accessed in each direction.
2863 		 */
2864 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2865 
2866 		/*
2867 		 * Set the direction flag on the first buffer in the SMP
2868 		 * passthrough request.  We'll clear it for the second one.
2869 		 */
2870 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2871 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2872 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2873 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2874 		dir = BUS_DMASYNC_PREWRITE;
2875 	} else
2876 		dir = BUS_DMASYNC_PREREAD;
2877 
2878 	for (i = 0; i < nsegs; i++) {
2879 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2880 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2881 		}
2882 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2883 		    sflags, nsegs - i);
2884 		if (error != 0) {
2885 			/* Resource shortage, roll back! */
2886 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2887 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2888 				    "consider increasing hw.mps.max_chains.\n");
2889 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2890 			mps_complete_command(sc, cm);
2891 			return;
2892 		}
2893 	}
2894 
2895 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2896 	mps_enqueue_request(sc, cm);
2897 
2898 	return;
2899 }
2900 
2901 static void
2902 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2903 	     int error)
2904 {
2905 	mps_data_cb(arg, segs, nsegs, error);
2906 }
2907 
2908 /*
2909  * This is the routine to enqueue commands ansynchronously.
2910  * Note that the only error path here is from bus_dmamap_load(), which can
2911  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2912  * assumed that if you have a command in-hand, then you have enough credits
2913  * to use it.
2914  */
2915 int
2916 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2917 {
2918 	int error = 0;
2919 
2920 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2921 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2922 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2923 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2924 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2925 		    cm->cm_data, mps_data_cb, cm, 0);
2926 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2927 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2928 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2929 	} else {
2930 		/* Add a zero-length element as needed */
2931 		if (cm->cm_sge != NULL)
2932 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2933 		mps_enqueue_request(sc, cm);
2934 	}
2935 
2936 	return (error);
2937 }
2938 
2939 /*
2940  * This is the routine to enqueue commands synchronously.  An error of
2941  * EINPROGRESS from mps_map_command() is ignored since the command will
2942  * be executed and enqueued automatically.  Other errors come from msleep().
2943  */
2944 int
2945 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2946     int sleep_flag)
2947 {
2948 	int error, rc;
2949 	struct timeval cur_time, start_time;
2950 	struct mps_command *cm = *cmp;
2951 
2952 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2953 		return  EBUSY;
2954 
2955 	cm->cm_complete = NULL;
2956 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2957 	error = mps_map_command(sc, cm);
2958 	if ((error != 0) && (error != EINPROGRESS))
2959 		return (error);
2960 
2961 	/*
2962 	 * Check for context and wait for 50 mSec at a time until time has
2963 	 * expired or the command has finished.  If msleep can't be used, need
2964 	 * to poll.
2965 	 */
2966 	if (curthread->td_no_sleeping != 0)
2967 		sleep_flag = NO_SLEEP;
2968 	getmicrouptime(&start_time);
2969 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2970 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2971 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2972 		if (error == EWOULDBLOCK) {
2973 			/*
2974 			 * Record the actual elapsed time in the case of a
2975 			 * timeout for the message below.
2976 			 */
2977 			getmicrouptime(&cur_time);
2978 			timevalsub(&cur_time, &start_time);
2979 		}
2980 	} else {
2981 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2982 			mps_intr_locked(sc);
2983 			if (sleep_flag == CAN_SLEEP)
2984 				pause("mpswait", hz/20);
2985 			else
2986 				DELAY(50000);
2987 
2988 			getmicrouptime(&cur_time);
2989 			timevalsub(&cur_time, &start_time);
2990 			if (cur_time.tv_sec > timeout) {
2991 				error = EWOULDBLOCK;
2992 				break;
2993 			}
2994 		}
2995 	}
2996 
2997 	if (error == EWOULDBLOCK) {
2998 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2999 		    " elapsed=%jd\n", __func__, timeout,
3000 		    (intmax_t)cur_time.tv_sec);
3001 		rc = mps_reinit(sc);
3002 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3003 		    "failed");
3004 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
3005 			/*
3006 			 * Tell the caller that we freed the command in a
3007 			 * reinit.
3008 			 */
3009 			*cmp = NULL;
3010 		}
3011 		error = ETIMEDOUT;
3012 	}
3013 	return (error);
3014 }
3015 
3016 /*
3017  * The MPT driver had a verbose interface for config pages.  In this driver,
3018  * reduce it to much simpler terms, similar to the Linux driver.
3019  */
3020 int
3021 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
3022 {
3023 	MPI2_CONFIG_REQUEST *req;
3024 	struct mps_command *cm;
3025 	int error;
3026 
3027 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
3028 		return (EBUSY);
3029 	}
3030 
3031 	cm = mps_alloc_command(sc);
3032 	if (cm == NULL) {
3033 		return (EBUSY);
3034 	}
3035 
3036 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3037 	req->Function = MPI2_FUNCTION_CONFIG;
3038 	req->Action = params->action;
3039 	req->SGLFlags = 0;
3040 	req->ChainOffset = 0;
3041 	req->PageAddress = params->page_address;
3042 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3043 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3044 
3045 		hdr = &params->hdr.Ext;
3046 		req->ExtPageType = hdr->ExtPageType;
3047 		req->ExtPageLength = hdr->ExtPageLength;
3048 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3049 		req->Header.PageLength = 0; /* Must be set to zero */
3050 		req->Header.PageNumber = hdr->PageNumber;
3051 		req->Header.PageVersion = hdr->PageVersion;
3052 	} else {
3053 		MPI2_CONFIG_PAGE_HEADER *hdr;
3054 
3055 		hdr = &params->hdr.Struct;
3056 		req->Header.PageType = hdr->PageType;
3057 		req->Header.PageNumber = hdr->PageNumber;
3058 		req->Header.PageLength = hdr->PageLength;
3059 		req->Header.PageVersion = hdr->PageVersion;
3060 	}
3061 
3062 	cm->cm_data = params->buffer;
3063 	cm->cm_length = params->length;
3064 	if (cm->cm_data != NULL) {
3065 		cm->cm_sge = &req->PageBufferSGE;
3066 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3067 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
3068 	} else
3069 		cm->cm_sge = NULL;
3070 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3071 
3072 	cm->cm_complete_data = params;
3073 	if (params->callback != NULL) {
3074 		cm->cm_complete = mps_config_complete;
3075 		return (mps_map_command(sc, cm));
3076 	} else {
3077 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
3078 		if (error) {
3079 			mps_dprint(sc, MPS_FAULT,
3080 			    "Error %d reading config page\n", error);
3081 			if (cm != NULL)
3082 				mps_free_command(sc, cm);
3083 			return (error);
3084 		}
3085 		mps_config_complete(sc, cm);
3086 	}
3087 
3088 	return (0);
3089 }
3090 
3091 int
3092 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3093 {
3094 	return (EINVAL);
3095 }
3096 
3097 static void
3098 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3099 {
3100 	MPI2_CONFIG_REPLY *reply;
3101 	struct mps_config_params *params;
3102 
3103 	MPS_FUNCTRACE(sc);
3104 	params = cm->cm_complete_data;
3105 
3106 	if (cm->cm_data != NULL) {
3107 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3108 		    BUS_DMASYNC_POSTREAD);
3109 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3110 	}
3111 
3112 	/*
3113 	 * XXX KDM need to do more error recovery?  This results in the
3114 	 * device in question not getting probed.
3115 	 */
3116 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3117 		params->status = MPI2_IOCSTATUS_BUSY;
3118 		goto done;
3119 	}
3120 
3121 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3122 	if (reply == NULL) {
3123 		params->status = MPI2_IOCSTATUS_BUSY;
3124 		goto done;
3125 	}
3126 	params->status = reply->IOCStatus;
3127 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3128 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3129 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3130 		params->hdr.Ext.PageType = reply->Header.PageType;
3131 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3132 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3133 	} else {
3134 		params->hdr.Struct.PageType = reply->Header.PageType;
3135 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3136 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3137 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3138 	}
3139 
3140 done:
3141 	mps_free_command(sc, cm);
3142 	if (params->callback != NULL)
3143 		params->callback(sc, params);
3144 
3145 	return;
3146 }
3147