xref: /freebsd/sys/dev/mps/mps.c (revision 3ea909cc7605dfabe34e3b01b07c038cda0afb86)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2015 Avago Technologies
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
29  *
30  * $FreeBSD$
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 /* Communications core for Avago Technologies (LSI) MPT2 */
37 
38 /* TODO Move headers to mpsvar */
39 #include <sys/types.h>
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/selinfo.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/module.h>
47 #include <sys/bus.h>
48 #include <sys/conf.h>
49 #include <sys/bio.h>
50 #include <sys/malloc.h>
51 #include <sys/uio.h>
52 #include <sys/sysctl.h>
53 #include <sys/queue.h>
54 #include <sys/kthread.h>
55 #include <sys/taskqueue.h>
56 #include <sys/endian.h>
57 #include <sys/eventhandler.h>
58 
59 #include <machine/bus.h>
60 #include <machine/resource.h>
61 #include <sys/rman.h>
62 #include <sys/proc.h>
63 
64 #include <dev/pci/pcivar.h>
65 
66 #include <cam/cam.h>
67 #include <cam/scsi/scsi_all.h>
68 
69 #include <dev/mps/mpi/mpi2_type.h>
70 #include <dev/mps/mpi/mpi2.h>
71 #include <dev/mps/mpi/mpi2_ioc.h>
72 #include <dev/mps/mpi/mpi2_sas.h>
73 #include <dev/mps/mpi/mpi2_cnfg.h>
74 #include <dev/mps/mpi/mpi2_init.h>
75 #include <dev/mps/mpi/mpi2_tool.h>
76 #include <dev/mps/mps_ioctl.h>
77 #include <dev/mps/mpsvar.h>
78 #include <dev/mps/mps_table.h>
79 
80 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
81 static int mps_init_queues(struct mps_softc *sc);
82 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
83 static int mps_transition_operational(struct mps_softc *sc);
84 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
85 static void mps_iocfacts_free(struct mps_softc *sc);
86 static void mps_startup(void *arg);
87 static int mps_send_iocinit(struct mps_softc *sc);
88 static int mps_alloc_queues(struct mps_softc *sc);
89 static int mps_alloc_replies(struct mps_softc *sc);
90 static int mps_alloc_requests(struct mps_softc *sc);
91 static int mps_attach_log(struct mps_softc *sc);
92 static __inline void mps_complete_command(struct mps_softc *sc,
93     struct mps_command *cm);
94 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
95     MPI2_EVENT_NOTIFICATION_REPLY *reply);
96 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
97 static void mps_periodic(void *);
98 static int mps_reregister_events(struct mps_softc *sc);
99 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
100 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
101 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
102 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
103 
104 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
105 
106 /*
107  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
108  * any state and back to its initialization state machine.
109  */
110 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
111 
112 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
113  * Compiler only support unint64_t to be passed as argument.
114  * Otherwise it will throw below error
115  * "aggregate value used where an integer was expected"
116  */
117 
118 typedef union _reply_descriptor {
119         u64 word;
120         struct {
121                 u32 low;
122                 u32 high;
123         } u;
124 }reply_descriptor,address_descriptor;
125 
126 /* Rate limit chain-fail messages to 1 per minute */
127 static struct timeval mps_chainfail_interval = { 60, 0 };
128 
129 /*
130  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
131  * If this function is called from process context, it can sleep
132  * and there is no harm to sleep, in case if this fuction is called
133  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
134  * based on sleep flags driver will call either msleep, pause or DELAY.
135  * msleep and pause are of same variant, but pause is used when mps_mtx
136  * is not hold by driver.
137  *
138  */
139 static int
140 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
141 {
142 	uint32_t reg;
143 	int i, error, tries = 0;
144 	uint8_t first_wait_done = FALSE;
145 
146 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
147 
148 	/* Clear any pending interrupts */
149 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
150 
151 	/*
152 	 * Force NO_SLEEP for threads prohibited to sleep
153  	 * e.a Thread from interrupt handler are prohibited to sleep.
154  	 */
155 	if (curthread->td_no_sleeping != 0)
156 		sleep_flag = NO_SLEEP;
157 
158 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
159 
160 	/* Push the magic sequence */
161 	error = ETIMEDOUT;
162 	while (tries++ < 20) {
163 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
164 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
165 			    mpt2_reset_magic[i]);
166 		/* wait 100 msec */
167 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
168 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
169 			    "mpsdiag", hz/10);
170 		else if (sleep_flag == CAN_SLEEP)
171 			pause("mpsdiag", hz/10);
172 		else
173 			DELAY(100 * 1000);
174 
175 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
176 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
177 			error = 0;
178 			break;
179 		}
180 	}
181 	if (error) {
182 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
183 		    error);
184 		return (error);
185 	}
186 
187 	/* Send the actual reset.  XXX need to refresh the reg? */
188 	reg |= MPI2_DIAG_RESET_ADAPTER;
189 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
190 		reg);
191 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
192 
193 	/* Wait up to 300 seconds in 50ms intervals */
194 	error = ETIMEDOUT;
195 	for (i = 0; i < 6000; i++) {
196 		/*
197 		 * Wait 50 msec. If this is the first time through, wait 256
198 		 * msec to satisfy Diag Reset timing requirements.
199 		 */
200 		if (first_wait_done) {
201 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
202 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
203 				    "mpsdiag", hz/20);
204 			else if (sleep_flag == CAN_SLEEP)
205 				pause("mpsdiag", hz/20);
206 			else
207 				DELAY(50 * 1000);
208 		} else {
209 			DELAY(256 * 1000);
210 			first_wait_done = TRUE;
211 		}
212 		/*
213 		 * Check for the RESET_ADAPTER bit to be cleared first, then
214 		 * wait for the RESET state to be cleared, which takes a little
215 		 * longer.
216 		 */
217 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
218 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
219 			continue;
220 		}
221 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
222 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
223 			error = 0;
224 			break;
225 		}
226 	}
227 	if (error) {
228 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
229 		    error);
230 		return (error);
231 	}
232 
233 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
234 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
235 
236 	return (0);
237 }
238 
239 static int
240 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
241 {
242 	int error;
243 
244 	MPS_FUNCTRACE(sc);
245 
246 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
247 
248 	error = 0;
249 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
250 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
251 	    MPI2_DOORBELL_FUNCTION_SHIFT);
252 
253 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
254 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
255 		    "Doorbell handshake failed\n");
256 		error = ETIMEDOUT;
257 	}
258 
259 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
260 	return (error);
261 }
262 
263 static int
264 mps_transition_ready(struct mps_softc *sc)
265 {
266 	uint32_t reg, state;
267 	int error, tries = 0;
268 	int sleep_flags;
269 
270 	MPS_FUNCTRACE(sc);
271 	/* If we are in attach call, do not sleep */
272 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
273 					? CAN_SLEEP:NO_SLEEP;
274 	error = 0;
275 
276 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
277 	   __func__, sleep_flags);
278 
279 	while (tries++ < 1200) {
280 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
281 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
282 
283 		/*
284 		 * Ensure the IOC is ready to talk.  If it's not, try
285 		 * resetting it.
286 		 */
287 		if (reg & MPI2_DOORBELL_USED) {
288 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
289 			    "reset\n");
290 			mps_diag_reset(sc, sleep_flags);
291 			DELAY(50000);
292 			continue;
293 		}
294 
295 		/* Is the adapter owned by another peer? */
296 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
297 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
298 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
299 			    "control of another peer host, aborting "
300 			    "initialization.\n");
301 			error = ENXIO;
302 			break;
303 		}
304 
305 		state = reg & MPI2_IOC_STATE_MASK;
306 		if (state == MPI2_IOC_STATE_READY) {
307 			/* Ready to go! */
308 			error = 0;
309 			break;
310 		} else if (state == MPI2_IOC_STATE_FAULT) {
311 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
312 			    "state 0x%x, resetting\n",
313 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
314 			mps_diag_reset(sc, sleep_flags);
315 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
316 			/* Need to take ownership */
317 			mps_message_unit_reset(sc, sleep_flags);
318 		} else if (state == MPI2_IOC_STATE_RESET) {
319 			/* Wait a bit, IOC might be in transition */
320 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
321 			    "IOC in unexpected reset state\n");
322 		} else {
323 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
324 			    "IOC in unknown state 0x%x\n", state);
325 			error = EINVAL;
326 			break;
327 		}
328 
329 		/* Wait 50ms for things to settle down. */
330 		DELAY(50000);
331 	}
332 
333 	if (error)
334 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
335 		    "Cannot transition IOC to ready\n");
336 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
337 
338 	return (error);
339 }
340 
341 static int
342 mps_transition_operational(struct mps_softc *sc)
343 {
344 	uint32_t reg, state;
345 	int error;
346 
347 	MPS_FUNCTRACE(sc);
348 
349 	error = 0;
350 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
351 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
352 
353 	state = reg & MPI2_IOC_STATE_MASK;
354 	if (state != MPI2_IOC_STATE_READY) {
355 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
356 		if ((error = mps_transition_ready(sc)) != 0) {
357 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
358 			    "failed to transition ready, exit\n");
359 			return (error);
360 		}
361 	}
362 
363 	error = mps_send_iocinit(sc);
364 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
365 
366 	return (error);
367 }
368 
369 /*
370  * This is called during attach and when re-initializing due to a Diag Reset.
371  * IOC Facts is used to allocate many of the structures needed by the driver.
372  * If called from attach, de-allocation is not required because the driver has
373  * not allocated any structures yet, but if called from a Diag Reset, previously
374  * allocated structures based on IOC Facts will need to be freed and re-
375  * allocated bases on the latest IOC Facts.
376  */
377 static int
378 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
379 {
380 	int error;
381 	Mpi2IOCFactsReply_t saved_facts;
382 	uint8_t saved_mode, reallocating;
383 
384 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
385 
386 	/* Save old IOC Facts and then only reallocate if Facts have changed */
387 	if (!attaching) {
388 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
389 	}
390 
391 	/*
392 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
393 	 * a re-initialization and only return the error if attaching so the OS
394 	 * can handle it.
395 	 */
396 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
397 		if (attaching) {
398 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
399 			    "IOC Facts with error %d, exit\n", error);
400 			return (error);
401 		} else {
402 			panic("%s failed to get IOC Facts with error %d\n",
403 			    __func__, error);
404 		}
405 	}
406 
407 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
408 
409 	snprintf(sc->fw_version, sizeof(sc->fw_version),
410 	    "%02d.%02d.%02d.%02d",
411 	    sc->facts->FWVersion.Struct.Major,
412 	    sc->facts->FWVersion.Struct.Minor,
413 	    sc->facts->FWVersion.Struct.Unit,
414 	    sc->facts->FWVersion.Struct.Dev);
415 
416 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
417 	    MPS_DRIVER_VERSION);
418 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
419 	     sc->facts->IOCCapabilities,
420 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
421 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
422 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
423 
424 	/*
425 	 * If the chip doesn't support event replay then a hard reset will be
426 	 * required to trigger a full discovery.  Do the reset here then
427 	 * retransition to Ready.  A hard reset might have already been done,
428 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
429 	 * for a Diag Reset.
430 	 */
431 	if (attaching && ((sc->facts->IOCCapabilities &
432 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
433 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
434 		mps_diag_reset(sc, NO_SLEEP);
435 		if ((error = mps_transition_ready(sc)) != 0) {
436 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
437 			    "transition to ready with error %d, exit\n",
438 			    error);
439 			return (error);
440 		}
441 	}
442 
443 	/*
444 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
445 	 * changed from the previous IOC Facts, log a warning, but only if
446 	 * checking this after a Diag Reset and not during attach.
447 	 */
448 	saved_mode = sc->ir_firmware;
449 	if (sc->facts->IOCCapabilities &
450 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
451 		sc->ir_firmware = 1;
452 	if (!attaching) {
453 		if (sc->ir_firmware != saved_mode) {
454 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
455 			    "in IOC Facts does not match previous mode\n");
456 		}
457 	}
458 
459 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
460 	reallocating = FALSE;
461 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
462 
463 	if ((!attaching) &&
464 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
465 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
466 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
467 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
468 	    (saved_facts.ProductID != sc->facts->ProductID) ||
469 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
470 	    (saved_facts.IOCRequestFrameSize !=
471 	    sc->facts->IOCRequestFrameSize) ||
472 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
473 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
474 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
475 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
476 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
477 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
478 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
479 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
480 	    (saved_facts.MaxPersistentEntries !=
481 	    sc->facts->MaxPersistentEntries))) {
482 		reallocating = TRUE;
483 
484 		/* Record that we reallocated everything */
485 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
486 	}
487 
488 	/*
489 	 * Some things should be done if attaching or re-allocating after a Diag
490 	 * Reset, but are not needed after a Diag Reset if the FW has not
491 	 * changed.
492 	 */
493 	if (attaching || reallocating) {
494 		/*
495 		 * Check if controller supports FW diag buffers and set flag to
496 		 * enable each type.
497 		 */
498 		if (sc->facts->IOCCapabilities &
499 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
500 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
501 			    enabled = TRUE;
502 		if (sc->facts->IOCCapabilities &
503 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
504 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
505 			    enabled = TRUE;
506 		if (sc->facts->IOCCapabilities &
507 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
508 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
509 			    enabled = TRUE;
510 
511 		/*
512 		 * Set flag if EEDP is supported and if TLR is supported.
513 		 */
514 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
515 			sc->eedp_enabled = TRUE;
516 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
517 			sc->control_TLR = TRUE;
518 
519 		/*
520 		 * Size the queues. Since the reply queues always need one free
521 		 * entry, we'll just deduct one reply message here.
522 		 */
523 		sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
524 		sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
525 		    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
526 
527 		/*
528 		 * Initialize all Tail Queues
529 		 */
530 		TAILQ_INIT(&sc->req_list);
531 		TAILQ_INIT(&sc->high_priority_req_list);
532 		TAILQ_INIT(&sc->chain_list);
533 		TAILQ_INIT(&sc->tm_list);
534 	}
535 
536 	/*
537 	 * If doing a Diag Reset and the FW is significantly different
538 	 * (reallocating will be set above in IOC Facts comparison), then all
539 	 * buffers based on the IOC Facts will need to be freed before they are
540 	 * reallocated.
541 	 */
542 	if (reallocating) {
543 		mps_iocfacts_free(sc);
544 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
545 		    saved_facts.MaxVolumes);
546 	}
547 
548 	/*
549 	 * Any deallocation has been completed.  Now start reallocating
550 	 * if needed.  Will only need to reallocate if attaching or if the new
551 	 * IOC Facts are different from the previous IOC Facts after a Diag
552 	 * Reset. Targets have already been allocated above if needed.
553 	 */
554 	if (attaching || reallocating) {
555 		if (((error = mps_alloc_queues(sc)) != 0) ||
556 		    ((error = mps_alloc_replies(sc)) != 0) ||
557 		    ((error = mps_alloc_requests(sc)) != 0)) {
558 			if (attaching ) {
559 				mps_dprint(sc, MPS_INIT|MPS_FAULT,
560 				    "Failed to alloc queues with error %d\n",
561 				    error);
562 				mps_free(sc);
563 				return (error);
564 			} else {
565 				panic("%s failed to alloc queues with error "
566 				    "%d\n", __func__, error);
567 			}
568 		}
569 	}
570 
571 	/* Always initialize the queues */
572 	bzero(sc->free_queue, sc->fqdepth * 4);
573 	mps_init_queues(sc);
574 
575 	/*
576 	 * Always get the chip out of the reset state, but only panic if not
577 	 * attaching.  If attaching and there is an error, that is handled by
578 	 * the OS.
579 	 */
580 	error = mps_transition_operational(sc);
581 	if (error != 0) {
582 		if (attaching) {
583 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
584 			    "transition to operational with error %d\n", error);
585 			mps_free(sc);
586 			return (error);
587 		} else {
588 			panic("%s failed to transition to operational with "
589 			    "error %d\n", __func__, error);
590 		}
591 	}
592 
593 	/*
594 	 * Finish the queue initialization.
595 	 * These are set here instead of in mps_init_queues() because the
596 	 * IOC resets these values during the state transition in
597 	 * mps_transition_operational().  The free index is set to 1
598 	 * because the corresponding index in the IOC is set to 0, and the
599 	 * IOC treats the queues as full if both are set to the same value.
600 	 * Hence the reason that the queue can't hold all of the possible
601 	 * replies.
602 	 */
603 	sc->replypostindex = 0;
604 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
605 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
606 
607 	/*
608 	 * Attach the subsystems so they can prepare their event masks.
609 	 */
610 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
611 	if (attaching) {
612 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
613 		if (((error = mps_attach_log(sc)) != 0) ||
614 		    ((error = mps_attach_sas(sc)) != 0) ||
615 		    ((error = mps_attach_user(sc)) != 0)) {
616 			mps_dprint(sc, MPS_INIT|MPS_FAULT,"Failed to attach "
617 			    "all subsystems: error %d\n", error);
618 			mps_free(sc);
619 			return (error);
620 		}
621 
622 		if ((error = mps_pci_setup_interrupts(sc)) != 0) {
623 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
624 			    "interrupts\n");
625 			mps_free(sc);
626 			return (error);
627 		}
628 	}
629 
630 	/*
631 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
632 	 * reset it after a Diag Reset, just in case.
633 	 */
634 	sc->WD_available = FALSE;
635 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
636 		sc->WD_available = TRUE;
637 
638 	return (error);
639 }
640 
641 /*
642  * This is called if memory is being free (during detach for example) and when
643  * buffers need to be reallocated due to a Diag Reset.
644  */
645 static void
646 mps_iocfacts_free(struct mps_softc *sc)
647 {
648 	struct mps_command *cm;
649 	int i;
650 
651 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
652 
653 	if (sc->free_busaddr != 0)
654 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
655 	if (sc->free_queue != NULL)
656 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
657 		    sc->queues_map);
658 	if (sc->queues_dmat != NULL)
659 		bus_dma_tag_destroy(sc->queues_dmat);
660 
661 	if (sc->chain_busaddr != 0)
662 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
663 	if (sc->chain_frames != NULL)
664 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
665 		    sc->chain_map);
666 	if (sc->chain_dmat != NULL)
667 		bus_dma_tag_destroy(sc->chain_dmat);
668 
669 	if (sc->sense_busaddr != 0)
670 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
671 	if (sc->sense_frames != NULL)
672 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
673 		    sc->sense_map);
674 	if (sc->sense_dmat != NULL)
675 		bus_dma_tag_destroy(sc->sense_dmat);
676 
677 	if (sc->reply_busaddr != 0)
678 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
679 	if (sc->reply_frames != NULL)
680 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
681 		    sc->reply_map);
682 	if (sc->reply_dmat != NULL)
683 		bus_dma_tag_destroy(sc->reply_dmat);
684 
685 	if (sc->req_busaddr != 0)
686 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
687 	if (sc->req_frames != NULL)
688 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
689 	if (sc->req_dmat != NULL)
690 		bus_dma_tag_destroy(sc->req_dmat);
691 
692 	if (sc->chains != NULL)
693 		free(sc->chains, M_MPT2);
694 	if (sc->commands != NULL) {
695 		for (i = 1; i < sc->num_reqs; i++) {
696 			cm = &sc->commands[i];
697 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
698 		}
699 		free(sc->commands, M_MPT2);
700 	}
701 	if (sc->buffer_dmat != NULL)
702 		bus_dma_tag_destroy(sc->buffer_dmat);
703 }
704 
705 /*
706  * The terms diag reset and hard reset are used interchangeably in the MPI
707  * docs to mean resetting the controller chip.  In this code diag reset
708  * cleans everything up, and the hard reset function just sends the reset
709  * sequence to the chip.  This should probably be refactored so that every
710  * subsystem gets a reset notification of some sort, and can clean up
711  * appropriately.
712  */
713 int
714 mps_reinit(struct mps_softc *sc)
715 {
716 	int error;
717 	struct mpssas_softc *sassc;
718 
719 	sassc = sc->sassc;
720 
721 	MPS_FUNCTRACE(sc);
722 
723 	mtx_assert(&sc->mps_mtx, MA_OWNED);
724 
725 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
726 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
727 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
728 		return 0;
729 	}
730 
731 	/* make sure the completion callbacks can recognize they're getting
732 	 * a NULL cm_reply due to a reset.
733 	 */
734 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
735 
736 	/*
737 	 * Mask interrupts here.
738 	 */
739 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
740 	mps_mask_intr(sc);
741 
742 	error = mps_diag_reset(sc, CAN_SLEEP);
743 	if (error != 0) {
744 		/* XXXSL No need to panic here */
745 		panic("%s hard reset failed with error %d\n",
746 		    __func__, error);
747 	}
748 
749 	/* Restore the PCI state, including the MSI-X registers */
750 	mps_pci_restore(sc);
751 
752 	/* Give the I/O subsystem special priority to get itself prepared */
753 	mpssas_handle_reinit(sc);
754 
755 	/*
756 	 * Get IOC Facts and allocate all structures based on this information.
757 	 * The attach function will also call mps_iocfacts_allocate at startup.
758 	 * If relevant values have changed in IOC Facts, this function will free
759 	 * all of the memory based on IOC Facts and reallocate that memory.
760 	 */
761 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
762 		panic("%s IOC Facts based allocation failed with error %d\n",
763 		    __func__, error);
764 	}
765 
766 	/*
767 	 * Mapping structures will be re-allocated after getting IOC Page8, so
768 	 * free these structures here.
769 	 */
770 	mps_mapping_exit(sc);
771 
772 	/*
773 	 * The static page function currently read is IOC Page8.  Others can be
774 	 * added in future.  It's possible that the values in IOC Page8 have
775 	 * changed after a Diag Reset due to user modification, so always read
776 	 * these.  Interrupts are masked, so unmask them before getting config
777 	 * pages.
778 	 */
779 	mps_unmask_intr(sc);
780 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
781 	mps_base_static_config_pages(sc);
782 
783 	/*
784 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
785 	 * mapping tables.
786 	 */
787 	mps_mapping_initialize(sc);
788 
789 	/*
790 	 * Restart will reload the event masks clobbered by the reset, and
791 	 * then enable the port.
792 	 */
793 	mps_reregister_events(sc);
794 
795 	/* the end of discovery will release the simq, so we're done. */
796 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
797 	    sc, sc->replypostindex, sc->replyfreeindex);
798 
799 	mpssas_release_simq_reinit(sassc);
800 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
801 
802 	return 0;
803 }
804 
805 /* Wait for the chip to ACK a word that we've put into its FIFO
806  * Wait for <timeout> seconds. In single loop wait for busy loop
807  * for 500 microseconds.
808  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
809  * */
810 static int
811 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
812 {
813 
814 	u32 cntdn, count;
815 	u32 int_status;
816 	u32 doorbell;
817 
818 	count = 0;
819 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
820 	do {
821 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
822 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
823 			mps_dprint(sc, MPS_TRACE,
824 			"%s: successful count(%d), timeout(%d)\n",
825 			__func__, count, timeout);
826 		return 0;
827 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
828 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
829 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
830 				MPI2_IOC_STATE_FAULT) {
831 				mps_dprint(sc, MPS_FAULT,
832 					"fault_state(0x%04x)!\n", doorbell);
833 				return (EFAULT);
834 			}
835 		} else if (int_status == 0xFFFFFFFF)
836 			goto out;
837 
838 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
839 		* 0.5 milisecond */
840 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
841 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
842 			"mpsdba", hz/1000);
843 		else if (sleep_flag == CAN_SLEEP)
844 			pause("mpsdba", hz/1000);
845 		else
846 			DELAY(500);
847 		count++;
848 	} while (--cntdn);
849 
850 	out:
851 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
852 		"int_status(%x)!\n", __func__, count, int_status);
853 	return (ETIMEDOUT);
854 
855 }
856 
857 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
858 static int
859 mps_wait_db_int(struct mps_softc *sc)
860 {
861 	int retry;
862 
863 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
864 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
865 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
866 			return (0);
867 		DELAY(2000);
868 	}
869 	return (ETIMEDOUT);
870 }
871 
872 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
873 static int
874 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
875     int req_sz, int reply_sz, int timeout)
876 {
877 	uint32_t *data32;
878 	uint16_t *data16;
879 	int i, count, ioc_sz, residual;
880 	int sleep_flags = CAN_SLEEP;
881 
882 	if (curthread->td_no_sleeping != 0)
883 		sleep_flags = NO_SLEEP;
884 
885 	/* Step 1 */
886 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
887 
888 	/* Step 2 */
889 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
890 		return (EBUSY);
891 
892 	/* Step 3
893 	 * Announce that a message is coming through the doorbell.  Messages
894 	 * are pushed at 32bit words, so round up if needed.
895 	 */
896 	count = (req_sz + 3) / 4;
897 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
898 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
899 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
900 
901 	/* Step 4 */
902 	if (mps_wait_db_int(sc) ||
903 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
904 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
905 		return (ENXIO);
906 	}
907 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
908 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
909 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
910 		return (ENXIO);
911 	}
912 
913 	/* Step 5 */
914 	/* Clock out the message data synchronously in 32-bit dwords*/
915 	data32 = (uint32_t *)req;
916 	for (i = 0; i < count; i++) {
917 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
918 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
919 			mps_dprint(sc, MPS_FAULT,
920 			    "Timeout while writing doorbell\n");
921 			return (ENXIO);
922 		}
923 	}
924 
925 	/* Step 6 */
926 	/* Clock in the reply in 16-bit words.  The total length of the
927 	 * message is always in the 4th byte, so clock out the first 2 words
928 	 * manually, then loop the rest.
929 	 */
930 	data16 = (uint16_t *)reply;
931 	if (mps_wait_db_int(sc) != 0) {
932 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
933 		return (ENXIO);
934 	}
935 	data16[0] =
936 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
937 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
938 	if (mps_wait_db_int(sc) != 0) {
939 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
940 		return (ENXIO);
941 	}
942 	data16[1] =
943 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
944 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
945 
946 	/* Number of 32bit words in the message */
947 	ioc_sz = reply->MsgLength;
948 
949 	/*
950 	 * Figure out how many 16bit words to clock in without overrunning.
951 	 * The precision loss with dividing reply_sz can safely be
952 	 * ignored because the messages can only be multiples of 32bits.
953 	 */
954 	residual = 0;
955 	count = MIN((reply_sz / 4), ioc_sz) * 2;
956 	if (count < ioc_sz * 2) {
957 		residual = ioc_sz * 2 - count;
958 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
959 		    "residual message words\n", residual);
960 	}
961 
962 	for (i = 2; i < count; i++) {
963 		if (mps_wait_db_int(sc) != 0) {
964 			mps_dprint(sc, MPS_FAULT,
965 			    "Timeout reading doorbell %d\n", i);
966 			return (ENXIO);
967 		}
968 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
969 		    MPI2_DOORBELL_DATA_MASK;
970 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
971 	}
972 
973 	/*
974 	 * Pull out residual words that won't fit into the provided buffer.
975 	 * This keeps the chip from hanging due to a driver programming
976 	 * error.
977 	 */
978 	while (residual--) {
979 		if (mps_wait_db_int(sc) != 0) {
980 			mps_dprint(sc, MPS_FAULT,
981 			    "Timeout reading doorbell\n");
982 			return (ENXIO);
983 		}
984 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
985 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
986 	}
987 
988 	/* Step 7 */
989 	if (mps_wait_db_int(sc) != 0) {
990 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
991 		return (ENXIO);
992 	}
993 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
994 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
995 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
996 
997 	return (0);
998 }
999 
1000 static void
1001 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1002 {
1003 	reply_descriptor rd;
1004 	MPS_FUNCTRACE(sc);
1005 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1006 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1007 
1008 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1009 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1010 
1011 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1012 		sc->io_cmds_highwater++;
1013 	rd.u.low = cm->cm_desc.Words.Low;
1014 	rd.u.high = cm->cm_desc.Words.High;
1015 	rd.word = htole64(rd.word);
1016 	/* TODO-We may need to make below regwrite atomic */
1017 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1018 	    rd.u.low);
1019 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1020 	    rd.u.high);
1021 }
1022 
1023 /*
1024  * Just the FACTS, ma'am.
1025  */
1026 static int
1027 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1028 {
1029 	MPI2_DEFAULT_REPLY *reply;
1030 	MPI2_IOC_FACTS_REQUEST request;
1031 	int error, req_sz, reply_sz;
1032 
1033 	MPS_FUNCTRACE(sc);
1034 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1035 
1036 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1037 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1038 	reply = (MPI2_DEFAULT_REPLY *)facts;
1039 
1040 	bzero(&request, req_sz);
1041 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1042 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1043 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1044 
1045 	return (error);
1046 }
1047 
1048 static int
1049 mps_send_iocinit(struct mps_softc *sc)
1050 {
1051 	MPI2_IOC_INIT_REQUEST	init;
1052 	MPI2_DEFAULT_REPLY	reply;
1053 	int req_sz, reply_sz, error;
1054 	struct timeval now;
1055 	uint64_t time_in_msec;
1056 
1057 	MPS_FUNCTRACE(sc);
1058 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1059 
1060 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1061 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1062 	bzero(&init, req_sz);
1063 	bzero(&reply, reply_sz);
1064 
1065 	/*
1066 	 * Fill in the init block.  Note that most addresses are
1067 	 * deliberately in the lower 32bits of memory.  This is a micro-
1068 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1069 	 */
1070 	init.Function = MPI2_FUNCTION_IOC_INIT;
1071 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1072 	init.MsgVersion = htole16(MPI2_VERSION);
1073 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1074 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1075 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1076 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1077 	init.SenseBufferAddressHigh = 0;
1078 	init.SystemReplyAddressHigh = 0;
1079 	init.SystemRequestFrameBaseAddress.High = 0;
1080 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1081 	init.ReplyDescriptorPostQueueAddress.High = 0;
1082 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1083 	init.ReplyFreeQueueAddress.High = 0;
1084 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1085 	getmicrotime(&now);
1086 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1087 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1088 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1089 
1090 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1091 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1092 		error = ENXIO;
1093 
1094 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1095 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1096 	return (error);
1097 }
1098 
1099 void
1100 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1101 {
1102 	bus_addr_t *addr;
1103 
1104 	addr = arg;
1105 	*addr = segs[0].ds_addr;
1106 }
1107 
1108 static int
1109 mps_alloc_queues(struct mps_softc *sc)
1110 {
1111 	bus_addr_t queues_busaddr;
1112 	uint8_t *queues;
1113 	int qsize, fqsize, pqsize;
1114 
1115 	/*
1116 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1117 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1118 	 * This queue supplies fresh reply frames for the firmware to use.
1119 	 *
1120 	 * The reply descriptor post queue contains 8 byte entries in
1121 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1122 	 * contains filled-in reply frames sent from the firmware to the host.
1123 	 *
1124 	 * These two queues are allocated together for simplicity.
1125 	 */
1126 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1127 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1128 	fqsize= sc->fqdepth * 4;
1129 	pqsize = sc->pqdepth * 8;
1130 	qsize = fqsize + pqsize;
1131 
1132         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1133 				16, 0,			/* algnmnt, boundary */
1134 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1135 				BUS_SPACE_MAXADDR,	/* highaddr */
1136 				NULL, NULL,		/* filter, filterarg */
1137                                 qsize,			/* maxsize */
1138                                 1,			/* nsegments */
1139                                 qsize,			/* maxsegsize */
1140                                 0,			/* flags */
1141                                 NULL, NULL,		/* lockfunc, lockarg */
1142                                 &sc->queues_dmat)) {
1143 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1144 		return (ENOMEM);
1145         }
1146         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1147 	    &sc->queues_map)) {
1148 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1149 		return (ENOMEM);
1150         }
1151         bzero(queues, qsize);
1152         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1153 	    mps_memaddr_cb, &queues_busaddr, 0);
1154 
1155 	sc->free_queue = (uint32_t *)queues;
1156 	sc->free_busaddr = queues_busaddr;
1157 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1158 	sc->post_busaddr = queues_busaddr + fqsize;
1159 
1160 	return (0);
1161 }
1162 
1163 static int
1164 mps_alloc_replies(struct mps_softc *sc)
1165 {
1166 	int rsize, num_replies;
1167 
1168 	/*
1169 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1170 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1171 	 * replies can be used at once.
1172 	 */
1173 	num_replies = max(sc->fqdepth, sc->num_replies);
1174 
1175 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1176         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1177 				4, 0,			/* algnmnt, boundary */
1178 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1179 				BUS_SPACE_MAXADDR,	/* highaddr */
1180 				NULL, NULL,		/* filter, filterarg */
1181                                 rsize,			/* maxsize */
1182                                 1,			/* nsegments */
1183                                 rsize,			/* maxsegsize */
1184                                 0,			/* flags */
1185                                 NULL, NULL,		/* lockfunc, lockarg */
1186                                 &sc->reply_dmat)) {
1187 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1188 		return (ENOMEM);
1189         }
1190         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1191 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1192 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1193 		return (ENOMEM);
1194         }
1195         bzero(sc->reply_frames, rsize);
1196         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1197 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1198 
1199 	return (0);
1200 }
1201 
1202 static int
1203 mps_alloc_requests(struct mps_softc *sc)
1204 {
1205 	struct mps_command *cm;
1206 	struct mps_chain *chain;
1207 	int i, rsize, nsegs;
1208 
1209 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1210         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1211 				16, 0,			/* algnmnt, boundary */
1212 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1213 				BUS_SPACE_MAXADDR,	/* highaddr */
1214 				NULL, NULL,		/* filter, filterarg */
1215                                 rsize,			/* maxsize */
1216                                 1,			/* nsegments */
1217                                 rsize,			/* maxsegsize */
1218                                 0,			/* flags */
1219                                 NULL, NULL,		/* lockfunc, lockarg */
1220                                 &sc->req_dmat)) {
1221 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1222 		return (ENOMEM);
1223         }
1224         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1225 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1226 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1227 		return (ENOMEM);
1228         }
1229         bzero(sc->req_frames, rsize);
1230         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1231 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1232 
1233 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1234         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1235 				16, 0,			/* algnmnt, boundary */
1236 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1237 				BUS_SPACE_MAXADDR,	/* highaddr */
1238 				NULL, NULL,		/* filter, filterarg */
1239                                 rsize,			/* maxsize */
1240                                 1,			/* nsegments */
1241                                 rsize,			/* maxsegsize */
1242                                 0,			/* flags */
1243                                 NULL, NULL,		/* lockfunc, lockarg */
1244                                 &sc->chain_dmat)) {
1245 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1246 		return (ENOMEM);
1247         }
1248         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1249 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1250 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1251 		return (ENOMEM);
1252         }
1253         bzero(sc->chain_frames, rsize);
1254         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1255 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1256 
1257 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1258         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1259 				1, 0,			/* algnmnt, boundary */
1260 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1261 				BUS_SPACE_MAXADDR,	/* highaddr */
1262 				NULL, NULL,		/* filter, filterarg */
1263                                 rsize,			/* maxsize */
1264                                 1,			/* nsegments */
1265                                 rsize,			/* maxsegsize */
1266                                 0,			/* flags */
1267                                 NULL, NULL,		/* lockfunc, lockarg */
1268                                 &sc->sense_dmat)) {
1269 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1270 		return (ENOMEM);
1271         }
1272         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1273 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1274 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1275 		return (ENOMEM);
1276         }
1277         bzero(sc->sense_frames, rsize);
1278         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1279 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1280 
1281 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1282 	    M_WAITOK | M_ZERO);
1283 	if(!sc->chains) {
1284 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chains memory\n");
1285 		return (ENOMEM);
1286 	}
1287 	for (i = 0; i < sc->max_chains; i++) {
1288 		chain = &sc->chains[i];
1289 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1290 		    i * sc->facts->IOCRequestFrameSize * 4);
1291 		chain->chain_busaddr = sc->chain_busaddr +
1292 		    i * sc->facts->IOCRequestFrameSize * 4;
1293 		mps_free_chain(sc, chain);
1294 		sc->chain_free_lowwater++;
1295 	}
1296 
1297 	/* XXX Need to pick a more precise value */
1298 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1299         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1300 				1, 0,			/* algnmnt, boundary */
1301 				BUS_SPACE_MAXADDR,	/* lowaddr */
1302 				BUS_SPACE_MAXADDR,	/* highaddr */
1303 				NULL, NULL,		/* filter, filterarg */
1304                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1305                                 nsegs,			/* nsegments */
1306                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1307                                 BUS_DMA_ALLOCNOW,	/* flags */
1308                                 busdma_lock_mutex,	/* lockfunc */
1309 				&sc->mps_mtx,		/* lockarg */
1310                                 &sc->buffer_dmat)) {
1311 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1312 		return (ENOMEM);
1313         }
1314 
1315 	/*
1316 	 * SMID 0 cannot be used as a free command per the firmware spec.
1317 	 * Just drop that command instead of risking accounting bugs.
1318 	 */
1319 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1320 	    M_MPT2, M_WAITOK | M_ZERO);
1321 	if(!sc->commands) {
1322 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1323 		return (ENOMEM);
1324 	}
1325 	for (i = 1; i < sc->num_reqs; i++) {
1326 		cm = &sc->commands[i];
1327 		cm->cm_req = sc->req_frames +
1328 		    i * sc->facts->IOCRequestFrameSize * 4;
1329 		cm->cm_req_busaddr = sc->req_busaddr +
1330 		    i * sc->facts->IOCRequestFrameSize * 4;
1331 		cm->cm_sense = &sc->sense_frames[i];
1332 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1333 		cm->cm_desc.Default.SMID = i;
1334 		cm->cm_sc = sc;
1335 		TAILQ_INIT(&cm->cm_chain_list);
1336 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1337 
1338 		/* XXX Is a failure here a critical problem? */
1339 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1340 			if (i <= sc->facts->HighPriorityCredit)
1341 				mps_free_high_priority_command(sc, cm);
1342 			else
1343 				mps_free_command(sc, cm);
1344 		else {
1345 			panic("failed to allocate command %d\n", i);
1346 			sc->num_reqs = i;
1347 			break;
1348 		}
1349 	}
1350 
1351 	return (0);
1352 }
1353 
1354 static int
1355 mps_init_queues(struct mps_softc *sc)
1356 {
1357 	int i;
1358 
1359 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1360 
1361 	/*
1362 	 * According to the spec, we need to use one less reply than we
1363 	 * have space for on the queue.  So sc->num_replies (the number we
1364 	 * use) should be less than sc->fqdepth (allocated size).
1365 	 */
1366 	if (sc->num_replies >= sc->fqdepth)
1367 		return (EINVAL);
1368 
1369 	/*
1370 	 * Initialize all of the free queue entries.
1371 	 */
1372 	for (i = 0; i < sc->fqdepth; i++)
1373 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1374 	sc->replyfreeindex = sc->num_replies;
1375 
1376 	return (0);
1377 }
1378 
1379 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1380  * Next are the global settings, if they exist.  Highest are the per-unit
1381  * settings, if they exist.
1382  */
1383 void
1384 mps_get_tunables(struct mps_softc *sc)
1385 {
1386 	char tmpstr[80];
1387 
1388 	/* XXX default to some debugging for now */
1389 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1390 	sc->disable_msix = 0;
1391 	sc->disable_msi = 0;
1392 	sc->max_chains = MPS_CHAIN_FRAMES;
1393 	sc->max_io_pages = MPS_MAXIO_PAGES;
1394 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1395 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1396 	sc->use_phynum = 1;
1397 
1398 	/*
1399 	 * Grab the global variables.
1400 	 */
1401 	TUNABLE_INT_FETCH("hw.mps.debug_level", &sc->mps_debug);
1402 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1403 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1404 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1405 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1406 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1407 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1408 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1409 
1410 	/* Grab the unit-instance variables */
1411 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1412 	    device_get_unit(sc->mps_dev));
1413 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
1414 
1415 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1416 	    device_get_unit(sc->mps_dev));
1417 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1418 
1419 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1420 	    device_get_unit(sc->mps_dev));
1421 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1422 
1423 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1424 	    device_get_unit(sc->mps_dev));
1425 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1426 
1427 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1428 	    device_get_unit(sc->mps_dev));
1429 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1430 
1431 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1432 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1433 	    device_get_unit(sc->mps_dev));
1434 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1435 
1436 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1437 	    device_get_unit(sc->mps_dev));
1438 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1439 
1440 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1441 	    device_get_unit(sc->mps_dev));
1442 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1443 
1444 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1445 	    device_get_unit(sc->mps_dev));
1446 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1447 }
1448 
1449 static void
1450 mps_setup_sysctl(struct mps_softc *sc)
1451 {
1452 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1453 	struct sysctl_oid	*sysctl_tree = NULL;
1454 	char tmpstr[80], tmpstr2[80];
1455 
1456 	/*
1457 	 * Setup the sysctl variable so the user can change the debug level
1458 	 * on the fly.
1459 	 */
1460 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1461 	    device_get_unit(sc->mps_dev));
1462 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1463 
1464 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1465 	if (sysctl_ctx != NULL)
1466 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1467 
1468 	if (sysctl_tree == NULL) {
1469 		sysctl_ctx_init(&sc->sysctl_ctx);
1470 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1471 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1472 		    CTLFLAG_RD, 0, tmpstr);
1473 		if (sc->sysctl_tree == NULL)
1474 			return;
1475 		sysctl_ctx = &sc->sysctl_ctx;
1476 		sysctl_tree = sc->sysctl_tree;
1477 	}
1478 
1479 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1480 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
1481 	    "mps debug level");
1482 
1483 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1484 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1485 	    "Disable the use of MSI-X interrupts");
1486 
1487 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1488 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1489 	    "Disable the use of MSI interrupts");
1490 
1491 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1492 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1493 	    strlen(sc->fw_version), "firmware version");
1494 
1495 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1496 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1497 	    strlen(MPS_DRIVER_VERSION), "driver version");
1498 
1499 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1500 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1501 	    &sc->io_cmds_active, 0, "number of currently active commands");
1502 
1503 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1504 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1505 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1506 
1507 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1508 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1509 	    &sc->chain_free, 0, "number of free chain elements");
1510 
1511 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1512 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1513 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1514 
1515 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1516 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1517 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1518 
1519 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1520 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1521 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1522 	    "IOCFacts)");
1523 
1524 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1525 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1526 	    "enable SSU to SATA SSD/HDD at shutdown");
1527 
1528 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1529 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1530 	    &sc->chain_alloc_fail, "chain allocation failures");
1531 
1532 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1533 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1534 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1535 	    "spinup after SATA ID error");
1536 
1537 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1538 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1539 	    mps_mapping_dump, "A", "Mapping Table Dump");
1540 
1541 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1542 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1543 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1544 
1545 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1546 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1547 	    "Use the phy number for enumeration");
1548 }
1549 
1550 int
1551 mps_attach(struct mps_softc *sc)
1552 {
1553 	int error;
1554 
1555 	MPS_FUNCTRACE(sc);
1556 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1557 
1558 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1559 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1560 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1561 	TAILQ_INIT(&sc->event_list);
1562 	timevalclear(&sc->lastfail);
1563 
1564 	if ((error = mps_transition_ready(sc)) != 0) {
1565 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1566 		    "ready\n");
1567 		return (error);
1568 	}
1569 
1570 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1571 	    M_ZERO|M_NOWAIT);
1572 	if(!sc->facts) {
1573 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1574 		    "exit\n");
1575 		return (ENOMEM);
1576 	}
1577 
1578 	/*
1579 	 * Get IOC Facts and allocate all structures based on this information.
1580 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1581 	 * Facts. If relevant values have changed in IOC Facts, this function
1582 	 * will free all of the memory based on IOC Facts and reallocate that
1583 	 * memory.  If this fails, any allocated memory should already be freed.
1584 	 */
1585 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1586 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1587 		    "failed with error %d, exit\n", error);
1588 		return (error);
1589 	}
1590 
1591 	/* Start the periodic watchdog check on the IOC Doorbell */
1592 	mps_periodic(sc);
1593 
1594 	/*
1595 	 * The portenable will kick off discovery events that will drive the
1596 	 * rest of the initialization process.  The CAM/SAS module will
1597 	 * hold up the boot sequence until discovery is complete.
1598 	 */
1599 	sc->mps_ich.ich_func = mps_startup;
1600 	sc->mps_ich.ich_arg = sc;
1601 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1602 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1603 		    "Cannot establish MPS config hook\n");
1604 		error = EINVAL;
1605 	}
1606 
1607 	/*
1608 	 * Allow IR to shutdown gracefully when shutdown occurs.
1609 	 */
1610 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1611 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1612 
1613 	if (sc->shutdown_eh == NULL)
1614 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1615 		    "shutdown event registration failed\n");
1616 
1617 	mps_setup_sysctl(sc);
1618 
1619 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1620 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1621 
1622 	return (error);
1623 }
1624 
1625 /* Run through any late-start handlers. */
1626 static void
1627 mps_startup(void *arg)
1628 {
1629 	struct mps_softc *sc;
1630 
1631 	sc = (struct mps_softc *)arg;
1632 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1633 
1634 	mps_lock(sc);
1635 	mps_unmask_intr(sc);
1636 
1637 	/* initialize device mapping tables */
1638 	mps_base_static_config_pages(sc);
1639 	mps_mapping_initialize(sc);
1640 	mpssas_startup(sc);
1641 	mps_unlock(sc);
1642 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1643 }
1644 
1645 /* Periodic watchdog.  Is called with the driver lock already held. */
1646 static void
1647 mps_periodic(void *arg)
1648 {
1649 	struct mps_softc *sc;
1650 	uint32_t db;
1651 
1652 	sc = (struct mps_softc *)arg;
1653 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1654 		return;
1655 
1656 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1657 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1658 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1659 		mps_reinit(sc);
1660 	}
1661 
1662 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1663 }
1664 
1665 static void
1666 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1667     MPI2_EVENT_NOTIFICATION_REPLY *event)
1668 {
1669 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1670 
1671 	MPS_DPRINT_EVENT(sc, generic, event);
1672 
1673 	switch (event->Event) {
1674 	case MPI2_EVENT_LOG_DATA:
1675 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1676 		if (sc->mps_debug & MPS_EVENT)
1677 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1678 		break;
1679 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1680 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1681 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1682 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1683 		     entry->LogSequence);
1684 		break;
1685 	default:
1686 		break;
1687 	}
1688 	return;
1689 }
1690 
1691 static int
1692 mps_attach_log(struct mps_softc *sc)
1693 {
1694 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1695 
1696 	bzero(events, 16);
1697 	setbit(events, MPI2_EVENT_LOG_DATA);
1698 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1699 
1700 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1701 	    &sc->mps_log_eh);
1702 
1703 	return (0);
1704 }
1705 
1706 static int
1707 mps_detach_log(struct mps_softc *sc)
1708 {
1709 
1710 	if (sc->mps_log_eh != NULL)
1711 		mps_deregister_events(sc, sc->mps_log_eh);
1712 	return (0);
1713 }
1714 
1715 /*
1716  * Free all of the driver resources and detach submodules.  Should be called
1717  * without the lock held.
1718  */
1719 int
1720 mps_free(struct mps_softc *sc)
1721 {
1722 	int error;
1723 
1724 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1725 	/* Turn off the watchdog */
1726 	mps_lock(sc);
1727 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1728 	mps_unlock(sc);
1729 	/* Lock must not be held for this */
1730 	callout_drain(&sc->periodic);
1731 	callout_drain(&sc->device_check_callout);
1732 
1733 	if (((error = mps_detach_log(sc)) != 0) ||
1734 	    ((error = mps_detach_sas(sc)) != 0)) {
1735 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
1736 		    "subsystems, exit\n");
1737 		return (error);
1738 	}
1739 
1740 	mps_detach_user(sc);
1741 
1742 	/* Put the IOC back in the READY state. */
1743 	mps_lock(sc);
1744 	if ((error = mps_transition_ready(sc)) != 0) {
1745 		mps_unlock(sc);
1746 		return (error);
1747 	}
1748 	mps_unlock(sc);
1749 
1750 	if (sc->facts != NULL)
1751 		free(sc->facts, M_MPT2);
1752 
1753 	/*
1754 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
1755 	 * to free these buffers too.
1756 	 */
1757 	mps_iocfacts_free(sc);
1758 
1759 	if (sc->sysctl_tree != NULL)
1760 		sysctl_ctx_free(&sc->sysctl_ctx);
1761 
1762 	/* Deregister the shutdown function */
1763 	if (sc->shutdown_eh != NULL)
1764 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1765 
1766 	mtx_destroy(&sc->mps_mtx);
1767 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1768 
1769 	return (0);
1770 }
1771 
1772 static __inline void
1773 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
1774 {
1775 	MPS_FUNCTRACE(sc);
1776 
1777 	if (cm == NULL) {
1778 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
1779 		return;
1780 	}
1781 
1782 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1783 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1784 
1785 	if (cm->cm_complete != NULL) {
1786 		mps_dprint(sc, MPS_TRACE,
1787 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
1788 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
1789 			   cm->cm_reply);
1790 		cm->cm_complete(sc, cm);
1791 	}
1792 
1793 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1794 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
1795 		wakeup(cm);
1796 	}
1797 
1798 	if (cm->cm_sc->io_cmds_active != 0) {
1799 		cm->cm_sc->io_cmds_active--;
1800 	} else {
1801 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
1802 		    "out of sync - resynching to 0\n");
1803 	}
1804 }
1805 
1806 
1807 static void
1808 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
1809 {
1810 	union loginfo_type {
1811 		u32     loginfo;
1812 		struct {
1813 			u32     subcode:16;
1814 			u32     code:8;
1815 			u32     originator:4;
1816 			u32     bus_type:4;
1817 		} dw;
1818 	};
1819 	union loginfo_type sas_loginfo;
1820 	char *originator_str = NULL;
1821 
1822 	sas_loginfo.loginfo = log_info;
1823 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1824 		return;
1825 
1826 	/* each nexus loss loginfo */
1827 	if (log_info == 0x31170000)
1828 		return;
1829 
1830 	/* eat the loginfos associated with task aborts */
1831 	if ((log_info == 30050000 || log_info ==
1832 	    0x31140000 || log_info == 0x31130000))
1833 		return;
1834 
1835 	switch (sas_loginfo.dw.originator) {
1836 	case 0:
1837 		originator_str = "IOP";
1838 		break;
1839 	case 1:
1840 		originator_str = "PL";
1841 		break;
1842 	case 2:
1843 		originator_str = "IR";
1844 		break;
1845 }
1846 
1847 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
1848 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
1849 	originator_str, sas_loginfo.dw.code,
1850 	sas_loginfo.dw.subcode);
1851 }
1852 
1853 static void
1854 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
1855 {
1856 	MPI2DefaultReply_t *mpi_reply;
1857 	u16 sc_status;
1858 
1859 	mpi_reply = (MPI2DefaultReply_t*)reply;
1860 	sc_status = le16toh(mpi_reply->IOCStatus);
1861 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1862 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1863 }
1864 void
1865 mps_intr(void *data)
1866 {
1867 	struct mps_softc *sc;
1868 	uint32_t status;
1869 
1870 	sc = (struct mps_softc *)data;
1871 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1872 
1873 	/*
1874 	 * Check interrupt status register to flush the bus.  This is
1875 	 * needed for both INTx interrupts and driver-driven polling
1876 	 */
1877 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1878 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1879 		return;
1880 
1881 	mps_lock(sc);
1882 	mps_intr_locked(data);
1883 	mps_unlock(sc);
1884 	return;
1885 }
1886 
1887 /*
1888  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1889  * chip.  Hopefully this theory is correct.
1890  */
1891 void
1892 mps_intr_msi(void *data)
1893 {
1894 	struct mps_softc *sc;
1895 
1896 	sc = (struct mps_softc *)data;
1897 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1898 	mps_lock(sc);
1899 	mps_intr_locked(data);
1900 	mps_unlock(sc);
1901 	return;
1902 }
1903 
1904 /*
1905  * The locking is overly broad and simplistic, but easy to deal with for now.
1906  */
1907 void
1908 mps_intr_locked(void *data)
1909 {
1910 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1911 	struct mps_softc *sc;
1912 	struct mps_command *cm = NULL;
1913 	uint8_t flags;
1914 	u_int pq;
1915 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
1916 	mps_fw_diagnostic_buffer_t *pBuffer;
1917 
1918 	sc = (struct mps_softc *)data;
1919 
1920 	pq = sc->replypostindex;
1921 	mps_dprint(sc, MPS_TRACE,
1922 	    "%s sc %p starting with replypostindex %u\n",
1923 	    __func__, sc, sc->replypostindex);
1924 
1925 	for ( ;; ) {
1926 		cm = NULL;
1927 		desc = &sc->post_queue[sc->replypostindex];
1928 		flags = desc->Default.ReplyFlags &
1929 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1930 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
1931 		 || (le32toh(desc->Words.High) == 0xffffffff))
1932 			break;
1933 
1934 		/* increment the replypostindex now, so that event handlers
1935 		 * and cm completion handlers which decide to do a diag
1936 		 * reset can zero it without it getting incremented again
1937 		 * afterwards, and we break out of this loop on the next
1938 		 * iteration since the reply post queue has been cleared to
1939 		 * 0xFF and all descriptors look unused (which they are).
1940 		 */
1941 		if (++sc->replypostindex >= sc->pqdepth)
1942 			sc->replypostindex = 0;
1943 
1944 		switch (flags) {
1945 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1946 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
1947 			cm->cm_reply = NULL;
1948 			break;
1949 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1950 		{
1951 			uint32_t baddr;
1952 			uint8_t *reply;
1953 
1954 			/*
1955 			 * Re-compose the reply address from the address
1956 			 * sent back from the chip.  The ReplyFrameAddress
1957 			 * is the lower 32 bits of the physical address of
1958 			 * particular reply frame.  Convert that address to
1959 			 * host format, and then use that to provide the
1960 			 * offset against the virtual address base
1961 			 * (sc->reply_frames).
1962 			 */
1963 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1964 			reply = sc->reply_frames +
1965 				(baddr - ((uint32_t)sc->reply_busaddr));
1966 			/*
1967 			 * Make sure the reply we got back is in a valid
1968 			 * range.  If not, go ahead and panic here, since
1969 			 * we'll probably panic as soon as we deference the
1970 			 * reply pointer anyway.
1971 			 */
1972 			if ((reply < sc->reply_frames)
1973 			 || (reply > (sc->reply_frames +
1974 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1975 				printf("%s: WARNING: reply %p out of range!\n",
1976 				       __func__, reply);
1977 				printf("%s: reply_frames %p, fqdepth %d, "
1978 				       "frame size %d\n", __func__,
1979 				       sc->reply_frames, sc->fqdepth,
1980 				       sc->facts->ReplyFrameSize * 4);
1981 				printf("%s: baddr %#x,\n", __func__, baddr);
1982 				/* LSI-TODO. See Linux Code. Need Graceful exit*/
1983 				panic("Reply address out of range");
1984 			}
1985 			if (le16toh(desc->AddressReply.SMID) == 0) {
1986 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
1987 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
1988 					/*
1989 					 * If SMID is 0 for Diag Buffer Post,
1990 					 * this implies that the reply is due to
1991 					 * a release function with a status that
1992 					 * the buffer has been released.  Set
1993 					 * the buffer flags accordingly.
1994 					 */
1995 					rel_rep =
1996 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
1997 					if ((le16toh(rel_rep->IOCStatus) &
1998 					    MPI2_IOCSTATUS_MASK) ==
1999 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2000 					{
2001 						pBuffer =
2002 						    &sc->fw_diag_buffer_list[
2003 						    rel_rep->BufferType];
2004 						pBuffer->valid_data = TRUE;
2005 						pBuffer->owned_by_firmware =
2006 						    FALSE;
2007 						pBuffer->immediate = FALSE;
2008 					}
2009 				} else
2010 					mps_dispatch_event(sc, baddr,
2011 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2012 					    reply);
2013 			} else {
2014 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
2015 				cm->cm_reply = reply;
2016 				cm->cm_reply_data =
2017 				    le32toh(desc->AddressReply.ReplyFrameAddress);
2018 			}
2019 			break;
2020 		}
2021 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2022 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2023 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2024 		default:
2025 			/* Unhandled */
2026 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2027 			    desc->Default.ReplyFlags);
2028 			cm = NULL;
2029 			break;
2030 		}
2031 
2032 
2033 		if (cm != NULL) {
2034 			// Print Error reply frame
2035 			if (cm->cm_reply)
2036 				mps_display_reply_info(sc,cm->cm_reply);
2037 			mps_complete_command(sc, cm);
2038 		}
2039 
2040 		desc->Words.Low = 0xffffffff;
2041 		desc->Words.High = 0xffffffff;
2042 	}
2043 
2044 	if (pq != sc->replypostindex) {
2045 		mps_dprint(sc, MPS_TRACE,
2046 		    "%s sc %p writing postindex %d\n",
2047 		    __func__, sc, sc->replypostindex);
2048 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
2049 	}
2050 
2051 	return;
2052 }
2053 
2054 static void
2055 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2056     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2057 {
2058 	struct mps_event_handle *eh;
2059 	int event, handled = 0;
2060 
2061 	event = le16toh(reply->Event);
2062 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2063 		if (isset(eh->mask, event)) {
2064 			eh->callback(sc, data, reply);
2065 			handled++;
2066 		}
2067 	}
2068 
2069 	if (handled == 0)
2070 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2071 
2072 	/*
2073 	 * This is the only place that the event/reply should be freed.
2074 	 * Anything wanting to hold onto the event data should have
2075 	 * already copied it into their own storage.
2076 	 */
2077 	mps_free_reply(sc, data);
2078 }
2079 
2080 static void
2081 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2082 {
2083 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2084 
2085 	if (cm->cm_reply)
2086 		MPS_DPRINT_EVENT(sc, generic,
2087 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2088 
2089 	mps_free_command(sc, cm);
2090 
2091 	/* next, send a port enable */
2092 	mpssas_startup(sc);
2093 }
2094 
2095 /*
2096  * For both register_events and update_events, the caller supplies a bitmap
2097  * of events that it _wants_.  These functions then turn that into a bitmask
2098  * suitable for the controller.
2099  */
2100 int
2101 mps_register_events(struct mps_softc *sc, u32 *mask,
2102     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2103 {
2104 	struct mps_event_handle *eh;
2105 	int error = 0;
2106 
2107 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2108 	if(!eh) {
2109 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2110 		return (ENOMEM);
2111 	}
2112 	eh->callback = cb;
2113 	eh->data = data;
2114 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2115 	if (mask != NULL)
2116 		error = mps_update_events(sc, eh, mask);
2117 	*handle = eh;
2118 
2119 	return (error);
2120 }
2121 
2122 int
2123 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2124     u32 *mask)
2125 {
2126 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2127 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2128 	struct mps_command *cm;
2129 	int error, i;
2130 
2131 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2132 
2133 	if ((mask != NULL) && (handle != NULL))
2134 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2135 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2136 
2137 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2138 		sc->event_mask[i] = -1;
2139 
2140 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2141 		sc->event_mask[i] &= ~handle->mask[i];
2142 
2143 
2144 	if ((cm = mps_alloc_command(sc)) == NULL)
2145 		return (EBUSY);
2146 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2147 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2148 	evtreq->MsgFlags = 0;
2149 	evtreq->SASBroadcastPrimitiveMasks = 0;
2150 #ifdef MPS_DEBUG_ALL_EVENTS
2151 	{
2152 		u_char fullmask[16];
2153 		memset(fullmask, 0x00, 16);
2154 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2155 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2156 	}
2157 #else
2158         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2159                 evtreq->EventMasks[i] =
2160                     htole32(sc->event_mask[i]);
2161 #endif
2162 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2163 	cm->cm_data = NULL;
2164 
2165 	error = mps_wait_command(sc, &cm, 60, 0);
2166 	if (cm != NULL)
2167 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2168 	if ((reply == NULL) ||
2169 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2170 		error = ENXIO;
2171 
2172 	if (reply)
2173 		MPS_DPRINT_EVENT(sc, generic, reply);
2174 
2175 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2176 
2177 	if (cm != NULL)
2178 		mps_free_command(sc, cm);
2179 	return (error);
2180 }
2181 
2182 static int
2183 mps_reregister_events(struct mps_softc *sc)
2184 {
2185 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2186 	struct mps_command *cm;
2187 	struct mps_event_handle *eh;
2188 	int error, i;
2189 
2190 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2191 
2192 	/* first, reregister events */
2193 
2194 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2195 		sc->event_mask[i] = -1;
2196 
2197 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2198 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2199 			sc->event_mask[i] &= ~eh->mask[i];
2200 	}
2201 
2202 	if ((cm = mps_alloc_command(sc)) == NULL)
2203 		return (EBUSY);
2204 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2205 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2206 	evtreq->MsgFlags = 0;
2207 	evtreq->SASBroadcastPrimitiveMasks = 0;
2208 #ifdef MPS_DEBUG_ALL_EVENTS
2209 	{
2210 		u_char fullmask[16];
2211 		memset(fullmask, 0x00, 16);
2212 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2213 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2214 	}
2215 #else
2216         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2217                 evtreq->EventMasks[i] =
2218                     htole32(sc->event_mask[i]);
2219 #endif
2220 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2221 	cm->cm_data = NULL;
2222 	cm->cm_complete = mps_reregister_events_complete;
2223 
2224 	error = mps_map_command(sc, cm);
2225 
2226 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2227 	    error);
2228 	return (error);
2229 }
2230 
2231 void
2232 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2233 {
2234 
2235 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2236 	free(handle, M_MPT2);
2237 }
2238 
2239 /*
2240  * Add a chain element as the next SGE for the specified command.
2241  * Reset cm_sge and cm_sgesize to indicate all the available space.
2242  */
2243 static int
2244 mps_add_chain(struct mps_command *cm)
2245 {
2246 	MPI2_SGE_CHAIN32 *sgc;
2247 	struct mps_chain *chain;
2248 	int space;
2249 
2250 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2251 		panic("MPS: Need SGE Error Code\n");
2252 
2253 	chain = mps_alloc_chain(cm->cm_sc);
2254 	if (chain == NULL)
2255 		return (ENOBUFS);
2256 
2257 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2258 
2259 	/*
2260 	 * Note: a double-linked list is used to make it easier to
2261 	 * walk for debugging.
2262 	 */
2263 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2264 
2265 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2266 	sgc->Length = htole16(space);
2267 	sgc->NextChainOffset = 0;
2268 	/* TODO Looks like bug in Setting sgc->Flags.
2269 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2270 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2271 	 *	This is fine.. because we are not using simple element. In case of
2272 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2273  	 */
2274 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2275 	sgc->Address = htole32(chain->chain_busaddr);
2276 
2277 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2278 	cm->cm_sglsize = space;
2279 	return (0);
2280 }
2281 
2282 /*
2283  * Add one scatter-gather element (chain, simple, transaction context)
2284  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2285  * cm_sge as the remaining size and pointer to the next SGE to fill
2286  * in, respectively.
2287  */
2288 int
2289 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2290 {
2291 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2292 	MPI2_SGE_SIMPLE64 *sge = sgep;
2293 	int error, type;
2294 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2295 
2296 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2297 
2298 #ifdef INVARIANTS
2299 	switch (type) {
2300 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2301 		if (len != tc->DetailsLength + 4)
2302 			panic("TC %p length %u or %zu?", tc,
2303 			    tc->DetailsLength + 4, len);
2304 		}
2305 		break;
2306 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2307 		/* Driver only uses 32-bit chain elements */
2308 		if (len != MPS_SGC_SIZE)
2309 			panic("CHAIN %p length %u or %zu?", sgep,
2310 			    MPS_SGC_SIZE, len);
2311 		break;
2312 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2313 		/* Driver only uses 64-bit SGE simple elements */
2314 		if (len != MPS_SGE64_SIZE)
2315 			panic("SGE simple %p length %u or %zu?", sge,
2316 			    MPS_SGE64_SIZE, len);
2317 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2318 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2319 			panic("SGE simple %p not marked 64-bit?", sge);
2320 
2321 		break;
2322 	default:
2323 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2324 	}
2325 #endif
2326 
2327 	/*
2328 	 * case 1: 1 more segment, enough room for it
2329 	 * case 2: 2 more segments, enough room for both
2330 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2331 	 * case 4: >=1 more segment, enough room for only a chain
2332 	 * case 5: >=1 more segment, no room for anything (error)
2333          */
2334 
2335 	/*
2336 	 * There should be room for at least a chain element, or this
2337 	 * code is buggy.  Case (5).
2338 	 */
2339 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2340 		panic("MPS: Need SGE Error Code\n");
2341 
2342 	if (segsleft >= 2 &&
2343 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2344 		/*
2345 		 * There are 2 or more segments left to add, and only
2346 		 * enough room for 1 and a chain.  Case (3).
2347 		 *
2348 		 * Mark as last element in this chain if necessary.
2349 		 */
2350 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2351 			sge->FlagsLength |= htole32(
2352 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2353 		}
2354 
2355 		/*
2356 		 * Add the item then a chain.  Do the chain now,
2357 		 * rather than on the next iteration, to simplify
2358 		 * understanding the code.
2359 		 */
2360 		cm->cm_sglsize -= len;
2361 		bcopy(sgep, cm->cm_sge, len);
2362 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2363 		return (mps_add_chain(cm));
2364 	}
2365 
2366 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2367 		/*
2368 		 * 1 or more segment, enough room for only a chain.
2369 		 * Hope the previous element wasn't a Simple entry
2370 		 * that needed to be marked with
2371 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2372 		 */
2373 		if ((error = mps_add_chain(cm)) != 0)
2374 			return (error);
2375 	}
2376 
2377 #ifdef INVARIANTS
2378 	/* Case 1: 1 more segment, enough room for it. */
2379 	if (segsleft == 1 && cm->cm_sglsize < len)
2380 		panic("1 seg left and no room? %u versus %zu",
2381 		    cm->cm_sglsize, len);
2382 
2383 	/* Case 2: 2 more segments, enough room for both */
2384 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2385 		panic("2 segs left and no room? %u versus %zu",
2386 		    cm->cm_sglsize, len);
2387 #endif
2388 
2389 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2390 		/*
2391 		 * If this is a bi-directional request, need to account for that
2392 		 * here.  Save the pre-filled sge values.  These will be used
2393 		 * either for the 2nd SGL or for a single direction SGL.  If
2394 		 * cm_out_len is non-zero, this is a bi-directional request, so
2395 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2396 		 * fill in the IN SGL.  Note that at this time, when filling in
2397 		 * 2 SGL's for a bi-directional request, they both use the same
2398 		 * DMA buffer (same cm command).
2399 		 */
2400 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2401 		saved_address_low = sge->Address.Low;
2402 		saved_address_high = sge->Address.High;
2403 		if (cm->cm_out_len) {
2404 			sge->FlagsLength = htole32(cm->cm_out_len |
2405 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2406 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2407 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2408 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2409 			    MPI2_SGE_FLAGS_SHIFT));
2410 			cm->cm_sglsize -= len;
2411 			bcopy(sgep, cm->cm_sge, len);
2412 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2413 			    + len);
2414 		}
2415 		saved_buf_len |=
2416 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2417 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2418 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2419 		    MPI2_SGE_FLAGS_END_OF_LIST |
2420 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2421 		    MPI2_SGE_FLAGS_SHIFT);
2422 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2423 			saved_buf_len |=
2424 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2425 			    MPI2_SGE_FLAGS_SHIFT);
2426 		} else {
2427 			saved_buf_len |=
2428 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2429 			    MPI2_SGE_FLAGS_SHIFT);
2430 		}
2431 		sge->FlagsLength = htole32(saved_buf_len);
2432 		sge->Address.Low = saved_address_low;
2433 		sge->Address.High = saved_address_high;
2434 	}
2435 
2436 	cm->cm_sglsize -= len;
2437 	bcopy(sgep, cm->cm_sge, len);
2438 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2439 	return (0);
2440 }
2441 
2442 /*
2443  * Add one dma segment to the scatter-gather list for a command.
2444  */
2445 int
2446 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2447     int segsleft)
2448 {
2449 	MPI2_SGE_SIMPLE64 sge;
2450 
2451 	/*
2452 	 * This driver always uses 64-bit address elements for simplicity.
2453 	 */
2454 	bzero(&sge, sizeof(sge));
2455 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2456 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2457 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2458 	mps_from_u64(pa, &sge.Address);
2459 
2460 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2461 }
2462 
2463 static void
2464 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2465 {
2466 	struct mps_softc *sc;
2467 	struct mps_command *cm;
2468 	u_int i, dir, sflags;
2469 
2470 	cm = (struct mps_command *)arg;
2471 	sc = cm->cm_sc;
2472 
2473 	/*
2474 	 * In this case, just print out a warning and let the chip tell the
2475 	 * user they did the wrong thing.
2476 	 */
2477 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2478 		mps_dprint(sc, MPS_ERROR,
2479 			   "%s: warning: busdma returned %d segments, "
2480 			   "more than the %d allowed\n", __func__, nsegs,
2481 			   cm->cm_max_segs);
2482 	}
2483 
2484 	/*
2485 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2486 	 * here.  In that case, both direction flags will be set.
2487 	 */
2488 	sflags = 0;
2489 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2490 		/*
2491 		 * We have to add a special case for SMP passthrough, there
2492 		 * is no easy way to generically handle it.  The first
2493 		 * S/G element is used for the command (therefore the
2494 		 * direction bit needs to be set).  The second one is used
2495 		 * for the reply.  We'll leave it to the caller to make
2496 		 * sure we only have two buffers.
2497 		 */
2498 		/*
2499 		 * Even though the busdma man page says it doesn't make
2500 		 * sense to have both direction flags, it does in this case.
2501 		 * We have one s/g element being accessed in each direction.
2502 		 */
2503 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2504 
2505 		/*
2506 		 * Set the direction flag on the first buffer in the SMP
2507 		 * passthrough request.  We'll clear it for the second one.
2508 		 */
2509 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2510 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2511 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2512 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2513 		dir = BUS_DMASYNC_PREWRITE;
2514 	} else
2515 		dir = BUS_DMASYNC_PREREAD;
2516 
2517 	for (i = 0; i < nsegs; i++) {
2518 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2519 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2520 		}
2521 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2522 		    sflags, nsegs - i);
2523 		if (error != 0) {
2524 			/* Resource shortage, roll back! */
2525 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2526 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2527 				    "consider increasing hw.mps.max_chains.\n");
2528 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2529 			mps_complete_command(sc, cm);
2530 			return;
2531 		}
2532 	}
2533 
2534 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2535 	mps_enqueue_request(sc, cm);
2536 
2537 	return;
2538 }
2539 
2540 static void
2541 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2542 	     int error)
2543 {
2544 	mps_data_cb(arg, segs, nsegs, error);
2545 }
2546 
2547 /*
2548  * This is the routine to enqueue commands ansynchronously.
2549  * Note that the only error path here is from bus_dmamap_load(), which can
2550  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2551  * assumed that if you have a command in-hand, then you have enough credits
2552  * to use it.
2553  */
2554 int
2555 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2556 {
2557 	int error = 0;
2558 
2559 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2560 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2561 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2562 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2563 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2564 		    cm->cm_data, mps_data_cb, cm, 0);
2565 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2566 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2567 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2568 	} else {
2569 		/* Add a zero-length element as needed */
2570 		if (cm->cm_sge != NULL)
2571 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2572 		mps_enqueue_request(sc, cm);
2573 	}
2574 
2575 	return (error);
2576 }
2577 
2578 /*
2579  * This is the routine to enqueue commands synchronously.  An error of
2580  * EINPROGRESS from mps_map_command() is ignored since the command will
2581  * be executed and enqueued automatically.  Other errors come from msleep().
2582  */
2583 int
2584 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2585     int sleep_flag)
2586 {
2587 	int error, rc;
2588 	struct timeval cur_time, start_time;
2589 	struct mps_command *cm = *cmp;
2590 
2591 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2592 		return  EBUSY;
2593 
2594 	cm->cm_complete = NULL;
2595 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2596 	error = mps_map_command(sc, cm);
2597 	if ((error != 0) && (error != EINPROGRESS))
2598 		return (error);
2599 
2600 	/*
2601 	 * Check for context and wait for 50 mSec at a time until time has
2602 	 * expired or the command has finished.  If msleep can't be used, need
2603 	 * to poll.
2604 	 */
2605 	if (curthread->td_no_sleeping != 0)
2606 		sleep_flag = NO_SLEEP;
2607 	getmicrouptime(&start_time);
2608 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2609 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2610 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2611 		if (error == EWOULDBLOCK) {
2612 			/*
2613 			 * Record the actual elapsed time in the case of a
2614 			 * timeout for the message below.
2615 			 */
2616 			getmicrouptime(&cur_time);
2617 			timevalsub(&cur_time, &start_time);
2618 		}
2619 	} else {
2620 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2621 			mps_intr_locked(sc);
2622 			if (sleep_flag == CAN_SLEEP)
2623 				pause("mpswait", hz/20);
2624 			else
2625 				DELAY(50000);
2626 
2627 			getmicrouptime(&cur_time);
2628 			timevalsub(&cur_time, &start_time);
2629 			if (cur_time.tv_sec > timeout) {
2630 				error = EWOULDBLOCK;
2631 				break;
2632 			}
2633 		}
2634 	}
2635 
2636 	if (error == EWOULDBLOCK) {
2637 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
2638 		    " elapsed=%jd\n", __func__, timeout,
2639 		    (intmax_t)cur_time.tv_sec);
2640 		rc = mps_reinit(sc);
2641 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2642 		    "failed");
2643 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
2644 			/*
2645 			 * Tell the caller that we freed the command in a
2646 			 * reinit.
2647 			 */
2648 			*cmp = NULL;
2649 		}
2650 		error = ETIMEDOUT;
2651 	}
2652 	return (error);
2653 }
2654 
2655 /*
2656  * The MPT driver had a verbose interface for config pages.  In this driver,
2657  * reduce it to much simpler terms, similar to the Linux driver.
2658  */
2659 int
2660 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2661 {
2662 	MPI2_CONFIG_REQUEST *req;
2663 	struct mps_command *cm;
2664 	int error;
2665 
2666 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2667 		return (EBUSY);
2668 	}
2669 
2670 	cm = mps_alloc_command(sc);
2671 	if (cm == NULL) {
2672 		return (EBUSY);
2673 	}
2674 
2675 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2676 	req->Function = MPI2_FUNCTION_CONFIG;
2677 	req->Action = params->action;
2678 	req->SGLFlags = 0;
2679 	req->ChainOffset = 0;
2680 	req->PageAddress = params->page_address;
2681 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2682 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2683 
2684 		hdr = &params->hdr.Ext;
2685 		req->ExtPageType = hdr->ExtPageType;
2686 		req->ExtPageLength = hdr->ExtPageLength;
2687 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2688 		req->Header.PageLength = 0; /* Must be set to zero */
2689 		req->Header.PageNumber = hdr->PageNumber;
2690 		req->Header.PageVersion = hdr->PageVersion;
2691 	} else {
2692 		MPI2_CONFIG_PAGE_HEADER *hdr;
2693 
2694 		hdr = &params->hdr.Struct;
2695 		req->Header.PageType = hdr->PageType;
2696 		req->Header.PageNumber = hdr->PageNumber;
2697 		req->Header.PageLength = hdr->PageLength;
2698 		req->Header.PageVersion = hdr->PageVersion;
2699 	}
2700 
2701 	cm->cm_data = params->buffer;
2702 	cm->cm_length = params->length;
2703 	if (cm->cm_data != NULL) {
2704 		cm->cm_sge = &req->PageBufferSGE;
2705 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2706 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2707 	} else
2708 		cm->cm_sge = NULL;
2709 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2710 
2711 	cm->cm_complete_data = params;
2712 	if (params->callback != NULL) {
2713 		cm->cm_complete = mps_config_complete;
2714 		return (mps_map_command(sc, cm));
2715 	} else {
2716 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
2717 		if (error) {
2718 			mps_dprint(sc, MPS_FAULT,
2719 			    "Error %d reading config page\n", error);
2720 			if (cm != NULL)
2721 				mps_free_command(sc, cm);
2722 			return (error);
2723 		}
2724 		mps_config_complete(sc, cm);
2725 	}
2726 
2727 	return (0);
2728 }
2729 
2730 int
2731 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2732 {
2733 	return (EINVAL);
2734 }
2735 
2736 static void
2737 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
2738 {
2739 	MPI2_CONFIG_REPLY *reply;
2740 	struct mps_config_params *params;
2741 
2742 	MPS_FUNCTRACE(sc);
2743 	params = cm->cm_complete_data;
2744 
2745 	if (cm->cm_data != NULL) {
2746 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2747 		    BUS_DMASYNC_POSTREAD);
2748 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2749 	}
2750 
2751 	/*
2752 	 * XXX KDM need to do more error recovery?  This results in the
2753 	 * device in question not getting probed.
2754 	 */
2755 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
2756 		params->status = MPI2_IOCSTATUS_BUSY;
2757 		goto done;
2758 	}
2759 
2760 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2761 	if (reply == NULL) {
2762 		params->status = MPI2_IOCSTATUS_BUSY;
2763 		goto done;
2764 	}
2765 	params->status = reply->IOCStatus;
2766 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2767 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2768 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2769 		params->hdr.Ext.PageType = reply->Header.PageType;
2770 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
2771 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
2772 	} else {
2773 		params->hdr.Struct.PageType = reply->Header.PageType;
2774 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2775 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2776 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2777 	}
2778 
2779 done:
2780 	mps_free_command(sc, cm);
2781 	if (params->callback != NULL)
2782 		params->callback(sc, params);
2783 
2784 	return;
2785 }
2786