xref: /dragonfly/sys/dev/raid/mps/mps.c (revision 52f9f0d9)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  */
27 /*-
28  * Copyright (c) 2011 LSI Corp.
29  * All rights reserved.
30  *
31  * Redistribution and use in source and binary forms, with or without
32  * modification, are permitted provided that the following conditions
33  * are met:
34  * 1. Redistributions of source code must retain the above copyright
35  *    notice, this list of conditions and the following disclaimer.
36  * 2. Redistributions in binary form must reproduce the above copyright
37  *    notice, this list of conditions and the following disclaimer in the
38  *    documentation and/or other materials provided with the distribution.
39  *
40  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
41  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
42  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
43  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
44  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
45  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
46  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
47  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
48  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
49  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
50  * SUCH DAMAGE.
51  *
52  * LSI MPT-Fusion Host Adapter FreeBSD
53  *
54  * $FreeBSD: src/sys/dev/mps/mps.c,v 1.14 2012/01/26 18:17:21 ken Exp $
55  */
56 
57 /* Communications core for LSI MPT2 */
58 
59 /* TODO Move headers to mpsvar */
60 #include <sys/types.h>
61 #include <sys/param.h>
62 #include <sys/systm.h>
63 #include <sys/kernel.h>
64 #include <sys/lock.h>
65 #include <sys/globaldata.h>
66 #include <sys/module.h>
67 #include <sys/bus.h>
68 #include <sys/conf.h>
69 #include <sys/bio.h>
70 #include <sys/malloc.h>
71 #include <sys/uio.h>
72 #include <sys/sysctl.h>
73 #include <sys/queue.h>
74 #include <sys/kthread.h>
75 #include <sys/endian.h>
76 #include <sys/eventhandler.h>
77 
78 #include <sys/rman.h>
79 
80 #include <bus/pci/pcivar.h>
81 
82 #include <bus/cam/scsi/scsi_all.h>
83 
84 #include <dev/raid/mps/mpi/mpi2_type.h>
85 #include <dev/raid/mps/mpi/mpi2.h>
86 #include <dev/raid/mps/mpi/mpi2_ioc.h>
87 #include <dev/raid/mps/mpi/mpi2_sas.h>
88 #include <dev/raid/mps/mpi/mpi2_cnfg.h>
89 #include <dev/raid/mps/mpi/mpi2_init.h>
90 #include <dev/raid/mps/mpi/mpi2_tool.h>
91 #include <dev/raid/mps/mps_ioctl.h>
92 #include <dev/raid/mps/mpsvar.h>
93 #include <dev/raid/mps/mps_table.h>
94 
95 static int mps_diag_reset(struct mps_softc *sc);
96 static int mps_init_queues(struct mps_softc *sc);
97 static int mps_message_unit_reset(struct mps_softc *sc);
98 static int mps_transition_operational(struct mps_softc *sc);
99 static void mps_startup(void *arg);
100 static int mps_send_iocinit(struct mps_softc *sc);
101 static int mps_attach_log(struct mps_softc *sc);
102 static __inline void mps_complete_command(struct mps_command *cm);
103 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
104     MPI2_EVENT_NOTIFICATION_REPLY *reply);
105 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
106 static void mps_periodic(void *);
107 static int mps_reregister_events(struct mps_softc *sc);
108 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
109 
110 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
111 
112 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
113 
114 /*
115  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
116  * any state and back to its initialization state machine.
117  */
118 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
119 
120 static int
121 mps_diag_reset(struct mps_softc *sc)
122 {
123 	uint32_t reg;
124 	int i, error, tries = 0;
125 
126 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
127 
128 	/* Clear any pending interrupts */
129 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
130 
131 	/* Push the magic sequence */
132 	error = ETIMEDOUT;
133 	while (tries++ < 20) {
134 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
135 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
136 			    mpt2_reset_magic[i]);
137 
138 		DELAY(100 * 1000);
139 
140 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
141 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
142 			error = 0;
143 			break;
144 		}
145 	}
146 	if (error)
147 		return (error);
148 
149 	/* Send the actual reset.  XXX need to refresh the reg? */
150 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
151 	    reg | MPI2_DIAG_RESET_ADAPTER);
152 
153 	/* Wait up to 300 seconds in 50ms intervals */
154 	error = ETIMEDOUT;
155 	for (i = 0; i < 60000; i++) {
156 		DELAY(50000);
157 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
158 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
159 			error = 0;
160 			break;
161 		}
162 	}
163 	if (error)
164 		return (error);
165 
166 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
167 
168 	return (0);
169 }
170 
171 static int
172 mps_message_unit_reset(struct mps_softc *sc)
173 {
174 
175 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
176 
177 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
178 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
179 	    MPI2_DOORBELL_FUNCTION_SHIFT);
180 	DELAY(50000);
181 
182 	return (0);
183 }
184 
185 static int
186 mps_transition_ready(struct mps_softc *sc)
187 {
188 	uint32_t reg, state;
189 	int error, tries = 0;
190 
191 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
192 
193 	error = 0;
194 	while (tries++ < 5) {
195 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
196 		mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
197 
198 		/*
199 		 * Ensure the IOC is ready to talk.  If it's not, try
200 		 * resetting it.
201 		 */
202 		if (reg & MPI2_DOORBELL_USED) {
203 			mps_diag_reset(sc);
204 			DELAY(50000);
205 			continue;
206 		}
207 
208 		/* Is the adapter owned by another peer? */
209 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
210 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
211 			device_printf(sc->mps_dev, "IOC is under the control "
212 			    "of another peer host, aborting initialization.\n");
213 			return (ENXIO);
214 		}
215 
216 		state = reg & MPI2_IOC_STATE_MASK;
217 		if (state == MPI2_IOC_STATE_READY) {
218 			/* Ready to go! */
219 			error = 0;
220 			break;
221 		} else if (state == MPI2_IOC_STATE_FAULT) {
222 			mps_dprint(sc, MPS_INFO, "IOC in fault state 0x%x\n",
223 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
224 			mps_diag_reset(sc);
225 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
226 			/* Need to take ownership */
227 			mps_message_unit_reset(sc);
228 		} else if (state == MPI2_IOC_STATE_RESET) {
229 			/* Wait a bit, IOC might be in transition */
230 			mps_dprint(sc, MPS_FAULT,
231 			    "IOC in unexpected reset state\n");
232 		} else {
233 			mps_dprint(sc, MPS_FAULT,
234 			    "IOC in unknown state 0x%x\n", state);
235 			error = EINVAL;
236 			break;
237 		}
238 
239 		/* Wait 50ms for things to settle down. */
240 		DELAY(50000);
241 	}
242 
243 	if (error)
244 		device_printf(sc->mps_dev, "Cannot transition IOC to ready\n");
245 
246 	return (error);
247 }
248 
249 static int
250 mps_transition_operational(struct mps_softc *sc)
251 {
252 	uint32_t reg, state;
253 	int error;
254 
255 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
256 
257 	error = 0;
258 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
259 	mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
260 
261 	state = reg & MPI2_IOC_STATE_MASK;
262 	if (state != MPI2_IOC_STATE_READY) {
263 		if ((error = mps_transition_ready(sc)) != 0) {
264 			mps_dprint(sc, MPS_FAULT,
265 			    "%s failed to transition ready\n", __func__);
266 			return (error);
267 		}
268 	}
269 
270 	error = mps_send_iocinit(sc);
271 	return (error);
272 }
273 
274 /*
275  * XXX Some of this should probably move to mps.c
276  *
277  * The terms diag reset and hard reset are used interchangeably in the MPI
278  * docs to mean resetting the controller chip.  In this code diag reset
279  * cleans everything up, and the hard reset function just sends the reset
280  * sequence to the chip.  This should probably be refactored so that every
281  * subsystem gets a reset notification of some sort, and can clean up
282  * appropriately.
283  */
284 int
285 mps_reinit(struct mps_softc *sc)
286 {
287 	int error;
288 	uint32_t db;
289 
290 	mps_printf(sc, "%s sc %p\n", __func__, sc);
291 
292 	KKASSERT(lockstatus(&sc->mps_lock, curthread) != 0);
293 
294 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
295 		mps_printf(sc, "%s reset already in progress\n", __func__);
296 		return 0;
297 	}
298 
299 	/* make sure the completion callbacks can recognize they're getting
300 	 * a NULL cm_reply due to a reset.
301 	 */
302 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
303 
304 	mps_printf(sc, "%s mask interrupts\n", __func__);
305 	mps_mask_intr(sc);
306 
307 	error = mps_diag_reset(sc);
308 	if (error != 0) {
309 		panic("%s hard reset failed with error %d",
310 		    __func__, error);
311 	}
312 
313 	/* Restore the PCI state, including the MSI-X registers */
314 	mps_pci_restore(sc);
315 
316 	/* Give the I/O subsystem special priority to get itself prepared */
317 	mpssas_handle_reinit(sc);
318 
319 	/* reinitialize queues after the reset */
320 	bzero(sc->free_queue, sc->fqdepth * 4);
321 	mps_init_queues(sc);
322 
323 	/* get the chip out of the reset state */
324 	error = mps_transition_operational(sc);
325 	if (error != 0)
326 		panic("%s transition operational failed with error %d",
327 		    __func__, error);
328 
329 	/* Reinitialize the reply queue. This is delicate because this
330 	 * function is typically invoked by task mgmt completion callbacks,
331 	 * which are called by the interrupt thread.  We need to make sure
332 	 * the interrupt handler loop will exit when we return to it, and
333 	 * that it will recognize the indexes we've changed.
334 	 */
335 	sc->replypostindex = 0;
336 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
337 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
338 
339 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
340 	mps_printf(sc, "%s doorbell 0x%08x\n", __func__, db);
341 
342 	mps_printf(sc, "%s unmask interrupts post %u free %u\n", __func__,
343 	    sc->replypostindex, sc->replyfreeindex);
344 
345 	mps_unmask_intr(sc);
346 
347 	mps_printf(sc, "%s restarting post %u free %u\n", __func__,
348 	    sc->replypostindex, sc->replyfreeindex);
349 
350 	/* restart will reload the event masks clobbered by the reset, and
351 	 * then enable the port.
352 	 */
353 	mps_reregister_events(sc);
354 
355 	/* the end of discovery will release the simq, so we're done. */
356 	mps_printf(sc, "%s finished sc %p post %u free %u\n",
357 	    __func__, sc,
358 	    sc->replypostindex, sc->replyfreeindex);
359 
360 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
361 
362 	return 0;
363 }
364 
365 /* Wait for the chip to ACK a word that we've put into its FIFO */
366 static int
367 mps_wait_db_ack(struct mps_softc *sc)
368 {
369 	int retry;
370 
371 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
372 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
373 		    MPI2_HIS_SYS2IOC_DB_STATUS) == 0)
374 			return (0);
375 		DELAY(2000);
376 	}
377 	return (ETIMEDOUT);
378 }
379 
380 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
381 static int
382 mps_wait_db_int(struct mps_softc *sc)
383 {
384 	int retry;
385 
386 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
387 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
388 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
389 			return (0);
390 		DELAY(2000);
391 	}
392 	return (ETIMEDOUT);
393 }
394 
395 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
396 static int
397 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
398     int req_sz, int reply_sz, int timeout)
399 {
400 	uint32_t *data32;
401 	uint16_t *data16;
402 	int i, count, ioc_sz, residual;
403 
404 	/* Step 1 */
405 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
406 
407 	/* Step 2 */
408 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
409 		return (EBUSY);
410 
411 	/* Step 3
412 	 * Announce that a message is coming through the doorbell.  Messages
413 	 * are pushed at 32bit words, so round up if needed.
414 	 */
415 	count = (req_sz + 3) / 4;
416 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
417 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
418 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
419 
420 	/* Step 4 */
421 	if (mps_wait_db_int(sc) ||
422 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
423 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
424 		return (ENXIO);
425 	}
426 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
427 	if (mps_wait_db_ack(sc) != 0) {
428 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
429 		return (ENXIO);
430 	}
431 
432 	/* Step 5 */
433 	/* Clock out the message data synchronously in 32-bit dwords*/
434 	data32 = (uint32_t *)req;
435 	for (i = 0; i < count; i++) {
436 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, data32[i]);
437 		if (mps_wait_db_ack(sc) != 0) {
438 			mps_dprint(sc, MPS_FAULT,
439 			    "Timeout while writing doorbell\n");
440 			return (ENXIO);
441 		}
442 	}
443 
444 	/* Step 6 */
445 	/* Clock in the reply in 16-bit words.  The total length of the
446 	 * message is always in the 4th byte, so clock out the first 2 words
447 	 * manually, then loop the rest.
448 	 */
449 	data16 = (uint16_t *)reply;
450 	if (mps_wait_db_int(sc) != 0) {
451 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
452 		return (ENXIO);
453 	}
454 	data16[0] =
455 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
456 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
457 	if (mps_wait_db_int(sc) != 0) {
458 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
459 		return (ENXIO);
460 	}
461 	data16[1] =
462 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
463 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
464 
465 	/* Number of 32bit words in the message */
466 	ioc_sz = reply->MsgLength;
467 
468 	/*
469 	 * Figure out how many 16bit words to clock in without overrunning.
470 	 * The precision loss with dividing reply_sz can safely be
471 	 * ignored because the messages can only be multiples of 32bits.
472 	 */
473 	residual = 0;
474 	count = MIN((reply_sz / 4), ioc_sz) * 2;
475 	if (count < ioc_sz * 2) {
476 		residual = ioc_sz * 2 - count;
477 		mps_dprint(sc, MPS_FAULT, "Driver error, throwing away %d "
478 		    "residual message words\n", residual);
479 	}
480 
481 	for (i = 2; i < count; i++) {
482 		if (mps_wait_db_int(sc) != 0) {
483 			mps_dprint(sc, MPS_FAULT,
484 			    "Timeout reading doorbell %d\n", i);
485 			return (ENXIO);
486 		}
487 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
488 		    MPI2_DOORBELL_DATA_MASK;
489 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
490 	}
491 
492 	/*
493 	 * Pull out residual words that won't fit into the provided buffer.
494 	 * This keeps the chip from hanging due to a driver programming
495 	 * error.
496 	 */
497 	while (residual--) {
498 		if (mps_wait_db_int(sc) != 0) {
499 			mps_dprint(sc, MPS_FAULT,
500 			    "Timeout reading doorbell\n");
501 			return (ENXIO);
502 		}
503 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
504 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
505 	}
506 
507 	/* Step 7 */
508 	if (mps_wait_db_int(sc) != 0) {
509 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
510 		return (ENXIO);
511 	}
512 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
513 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
514 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
515 
516 	return (0);
517 }
518 
519 static void
520 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
521 {
522 
523 	mps_dprint(sc, MPS_TRACE, "%s SMID %u cm %p ccb %p\n", __func__,
524 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
525 
526 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
527 		KKASSERT(lockstatus(&sc->mps_lock, curthread) != 0);
528 
529 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
530 		sc->io_cmds_highwater++;
531 
532 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
533 	    cm->cm_desc.Words.Low);
534 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
535 	    cm->cm_desc.Words.High);
536 }
537 
538 /*
539  * Just the FACTS, ma'am.
540  */
541 static int
542 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
543 {
544 	MPI2_DEFAULT_REPLY *reply;
545 	MPI2_IOC_FACTS_REQUEST request;
546 	int error, req_sz, reply_sz;
547 
548 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
549 
550 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
551 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
552 	reply = (MPI2_DEFAULT_REPLY *)facts;
553 
554 	bzero(&request, req_sz);
555 	request.Function = MPI2_FUNCTION_IOC_FACTS;
556 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
557 
558 	return (error);
559 }
560 
561 static int
562 mps_get_portfacts(struct mps_softc *sc, MPI2_PORT_FACTS_REPLY *facts, int port)
563 {
564 	MPI2_PORT_FACTS_REQUEST *request;
565 	MPI2_PORT_FACTS_REPLY *reply;
566 	struct mps_command *cm;
567 	int error;
568 
569 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
570 
571 	if ((cm = mps_alloc_command(sc)) == NULL)
572 		return (EBUSY);
573 	request = (MPI2_PORT_FACTS_REQUEST *)cm->cm_req;
574 	request->Function = MPI2_FUNCTION_PORT_FACTS;
575 	request->PortNumber = port;
576 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
577 	cm->cm_data = NULL;
578 	error = mps_request_polled(sc, cm);
579 	reply = (MPI2_PORT_FACTS_REPLY *)cm->cm_reply;
580 	if (reply == NULL) {
581 		mps_printf(sc, "%s NULL reply\n", __func__);
582 		goto done;
583 	}
584 	if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS) {
585 		mps_printf(sc,
586 		    "%s error %d iocstatus 0x%x iocloginfo 0x%x type 0x%x\n",
587 		    __func__, error, reply->IOCStatus, reply->IOCLogInfo,
588 		    reply->PortType);
589 		error = ENXIO;
590 	}
591 	bcopy(reply, facts, sizeof(MPI2_PORT_FACTS_REPLY));
592 done:
593 	mps_free_command(sc, cm);
594 
595 	return (error);
596 }
597 
598 static int
599 mps_send_iocinit(struct mps_softc *sc)
600 {
601 	MPI2_IOC_INIT_REQUEST	init;
602 	MPI2_DEFAULT_REPLY	reply;
603 	int req_sz, reply_sz, error;
604 
605 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
606 
607 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
608 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
609 	bzero(&init, req_sz);
610 	bzero(&reply, reply_sz);
611 
612 	/*
613 	 * Fill in the init block.  Note that most addresses are
614 	 * deliberately in the lower 32bits of memory.  This is a micro-
615 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
616 	 */
617 	init.Function = MPI2_FUNCTION_IOC_INIT;
618 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
619 	init.MsgVersion = MPI2_VERSION;
620 	init.HeaderVersion = MPI2_HEADER_VERSION;
621 	init.SystemRequestFrameSize = sc->facts->IOCRequestFrameSize;
622 	init.ReplyDescriptorPostQueueDepth = sc->pqdepth;
623 	init.ReplyFreeQueueDepth = sc->fqdepth;
624 	init.SenseBufferAddressHigh = 0;
625 	init.SystemReplyAddressHigh = 0;
626 	init.SystemRequestFrameBaseAddress.High = 0;
627 	init.SystemRequestFrameBaseAddress.Low = (uint32_t)sc->req_busaddr;
628 	init.ReplyDescriptorPostQueueAddress.High = 0;
629 	init.ReplyDescriptorPostQueueAddress.Low = (uint32_t)sc->post_busaddr;
630 	init.ReplyFreeQueueAddress.High = 0;
631 	init.ReplyFreeQueueAddress.Low = (uint32_t)sc->free_busaddr;
632 	init.TimeStamp.High = 0;
633 	init.TimeStamp.Low = (uint32_t)time_second;
634 
635 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
636 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
637 		error = ENXIO;
638 
639 	mps_dprint(sc, MPS_INFO, "IOCInit status= 0x%x\n", reply.IOCStatus);
640 	return (error);
641 }
642 
643 void
644 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
645 {
646 	bus_addr_t *addr;
647 
648 	addr = arg;
649 	*addr = segs[0].ds_addr;
650 }
651 
652 static int
653 mps_alloc_queues(struct mps_softc *sc)
654 {
655 	bus_addr_t queues_busaddr;
656 	uint8_t *queues;
657 	int qsize, fqsize, pqsize;
658 
659 	/*
660 	 * The reply free queue contains 4 byte entries in multiples of 16 and
661 	 * aligned on a 16 byte boundary. There must always be an unused entry.
662 	 * This queue supplies fresh reply frames for the firmware to use.
663 	 *
664 	 * The reply descriptor post queue contains 8 byte entries in
665 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
666 	 * contains filled-in reply frames sent from the firmware to the host.
667 	 *
668 	 * These two queues are allocated together for simplicity.
669 	 */
670 	sc->fqdepth = roundup2((sc->num_replies + 1), 16);
671 	sc->pqdepth = roundup2((sc->num_replies + 1), 16);
672 	fqsize= sc->fqdepth * 4;
673 	pqsize = sc->pqdepth * 8;
674 	qsize = fqsize + pqsize;
675 
676         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
677 				16, 0,			/* algnmnt, boundary */
678 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
679 				BUS_SPACE_MAXADDR,	/* highaddr */
680 				NULL, NULL,		/* filter, filterarg */
681                                 qsize,			/* maxsize */
682                                 1,			/* nsegments */
683                                 qsize,			/* maxsegsize */
684                                 0,			/* flags */
685                                 &sc->queues_dmat)) {
686 		device_printf(sc->mps_dev, "Cannot allocate queues DMA tag\n");
687 		return (ENOMEM);
688         }
689         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
690 	    &sc->queues_map)) {
691 		device_printf(sc->mps_dev, "Cannot allocate queues memory\n");
692 		return (ENOMEM);
693         }
694         bzero(queues, qsize);
695         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
696 	    mps_memaddr_cb, &queues_busaddr, 0);
697 
698 	sc->free_queue = (uint32_t *)queues;
699 	sc->free_busaddr = queues_busaddr;
700 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
701 	sc->post_busaddr = queues_busaddr + fqsize;
702 
703 	return (0);
704 }
705 
706 static int
707 mps_alloc_replies(struct mps_softc *sc)
708 {
709 	int rsize, num_replies;
710 
711 	/*
712 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
713 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
714 	 * replies can be used at once.
715 	 */
716 	num_replies = max(sc->fqdepth, sc->num_replies);
717 
718 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
719         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
720 				4, 0,			/* algnmnt, boundary */
721 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
722 				BUS_SPACE_MAXADDR,	/* highaddr */
723 				NULL, NULL,		/* filter, filterarg */
724                                 rsize,			/* maxsize */
725                                 1,			/* nsegments */
726                                 rsize,			/* maxsegsize */
727                                 0,			/* flags */
728                                 &sc->reply_dmat)) {
729 		device_printf(sc->mps_dev, "Cannot allocate replies DMA tag\n");
730 		return (ENOMEM);
731         }
732         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
733 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
734 		device_printf(sc->mps_dev, "Cannot allocate replies memory\n");
735 		return (ENOMEM);
736         }
737         bzero(sc->reply_frames, rsize);
738         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
739 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
740 
741 	return (0);
742 }
743 
744 static int
745 mps_alloc_requests(struct mps_softc *sc)
746 {
747 	struct mps_command *cm;
748 	struct mps_chain *chain;
749 	int i, rsize, nsegs;
750 
751 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
752         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
753 				16, 0,			/* algnmnt, boundary */
754 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
755 				BUS_SPACE_MAXADDR,	/* highaddr */
756 				NULL, NULL,		/* filter, filterarg */
757                                 rsize,			/* maxsize */
758                                 1,			/* nsegments */
759                                 rsize,			/* maxsegsize */
760                                 0,			/* flags */
761                                 &sc->req_dmat)) {
762 		device_printf(sc->mps_dev, "Cannot allocate request DMA tag\n");
763 		return (ENOMEM);
764         }
765         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
766 	    BUS_DMA_NOWAIT, &sc->req_map)) {
767 		device_printf(sc->mps_dev, "Cannot allocate request memory\n");
768 		return (ENOMEM);
769         }
770         bzero(sc->req_frames, rsize);
771         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
772 	    mps_memaddr_cb, &sc->req_busaddr, 0);
773 
774 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
775         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
776 				16, 0,			/* algnmnt, boundary */
777 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
778 				BUS_SPACE_MAXADDR,	/* highaddr */
779 				NULL, NULL,		/* filter, filterarg */
780                                 rsize,			/* maxsize */
781                                 1,			/* nsegments */
782                                 rsize,			/* maxsegsize */
783                                 0,			/* flags */
784                                 &sc->chain_dmat)) {
785 		device_printf(sc->mps_dev, "Cannot allocate chain DMA tag\n");
786 		return (ENOMEM);
787         }
788         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
789 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
790 		device_printf(sc->mps_dev, "Cannot allocate chain memory\n");
791 		return (ENOMEM);
792         }
793         bzero(sc->chain_frames, rsize);
794         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
795 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
796 
797 	rsize = MPS_SENSE_LEN * sc->num_reqs;
798         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
799 				1, 0,			/* algnmnt, boundary */
800 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
801 				BUS_SPACE_MAXADDR,	/* highaddr */
802 				NULL, NULL,		/* filter, filterarg */
803                                 rsize,			/* maxsize */
804                                 1,			/* nsegments */
805                                 rsize,			/* maxsegsize */
806                                 0,			/* flags */
807                                 &sc->sense_dmat)) {
808 		device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
809 		return (ENOMEM);
810         }
811         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
812 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
813 		device_printf(sc->mps_dev, "Cannot allocate sense memory\n");
814 		return (ENOMEM);
815         }
816         bzero(sc->sense_frames, rsize);
817         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
818 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
819 
820 	sc->chains = kmalloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
821 	    M_WAITOK | M_ZERO);
822 	for (i = 0; i < sc->max_chains; i++) {
823 		chain = &sc->chains[i];
824 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
825 		    i * sc->facts->IOCRequestFrameSize * 4);
826 		chain->chain_busaddr = sc->chain_busaddr +
827 		    i * sc->facts->IOCRequestFrameSize * 4;
828 		mps_free_chain(sc, chain);
829 		sc->chain_free_lowwater++;
830 	}
831 
832 	/* XXX Need to pick a more precise value */
833 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
834         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
835 				1, 0,			/* algnmnt, boundary */
836 				BUS_SPACE_MAXADDR,	/* lowaddr */
837 				BUS_SPACE_MAXADDR,	/* highaddr */
838 				NULL, NULL,		/* filter, filterarg */
839                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
840                                 nsegs,			/* nsegments */
841                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
842                                 BUS_DMA_ALLOCNOW,	/* flags */
843                                 &sc->buffer_dmat)) {
844 		device_printf(sc->mps_dev, "Cannot allocate buffer DMA tag\n");
845 		return (ENOMEM);
846         }
847 
848 	/*
849 	 * SMID 0 cannot be used as a free command per the firmware spec.
850 	 * Just drop that command instead of risking accounting bugs.
851 	 */
852 	sc->commands = kmalloc(sizeof(struct mps_command) * sc->num_reqs,
853 	    M_MPT2, M_WAITOK | M_ZERO);
854 	for (i = 1; i < sc->num_reqs; i++) {
855 		cm = &sc->commands[i];
856 		cm->cm_req = sc->req_frames +
857 		    i * sc->facts->IOCRequestFrameSize * 4;
858 		cm->cm_req_busaddr = sc->req_busaddr +
859 		    i * sc->facts->IOCRequestFrameSize * 4;
860 		cm->cm_sense = &sc->sense_frames[i];
861 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
862 		cm->cm_desc.Default.SMID = i;
863 		cm->cm_sc = sc;
864 		TAILQ_INIT(&cm->cm_chain_list);
865 		callout_init(&cm->cm_callout);
866 
867 		/* XXX Is a failure here a critical problem? */
868 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
869 			if (i <= sc->facts->HighPriorityCredit)
870 				mps_free_high_priority_command(sc, cm);
871 			else
872 				mps_free_command(sc, cm);
873 		else {
874 			panic("failed to allocate command %d", i);
875 			sc->num_reqs = i;
876 			break;
877 		}
878 	}
879 
880 	return (0);
881 }
882 
883 static int
884 mps_init_queues(struct mps_softc *sc)
885 {
886 	int i;
887 
888 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
889 
890 	/*
891 	 * According to the spec, we need to use one less reply than we
892 	 * have space for on the queue.  So sc->num_replies (the number we
893 	 * use) should be less than sc->fqdepth (allocated size).
894 	 */
895 	if (sc->num_replies >= sc->fqdepth)
896 		return (EINVAL);
897 
898 	/*
899 	 * Initialize all of the free queue entries.
900 	 */
901 	for (i = 0; i < sc->fqdepth; i++)
902 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
903 	sc->replyfreeindex = sc->num_replies;
904 
905 	return (0);
906 }
907 
908 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
909  * Next are the global settings, if they exist.  Highest are the per-unit
910  * settings, if they exist.
911  */
912 static void
913 mps_get_tunables(struct mps_softc *sc)
914 {
915 	char tmpstr[80];
916 
917 	/* XXX default to some debugging for now */
918 	sc->mps_debug = MPS_FAULT;
919 #if 0 /* XXX swildner */
920 	sc->disable_msix = 0;
921 #endif
922 	sc->enable_msi = 1;
923 	sc->max_chains = MPS_CHAIN_FRAMES;
924 
925 	/*
926 	 * Grab the global variables.
927 	 */
928 	TUNABLE_INT_FETCH("hw.mps.debug_level", &sc->mps_debug);
929 #if 0 /* XXX swildner */
930 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
931 #endif
932 	TUNABLE_INT_FETCH("hw.mps.msi.enable", &sc->enable_msi);
933 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
934 
935 	/* Grab the unit-instance variables */
936 	ksnprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
937 	    device_get_unit(sc->mps_dev));
938 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
939 
940 #if 0 /* XXX swildner */
941 	ksnprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
942 	    device_get_unit(sc->mps_dev));
943 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
944 #endif
945 
946 	ksnprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_msi",
947 	    device_get_unit(sc->mps_dev));
948 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_msi);
949 
950 	ksnprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
951 	    device_get_unit(sc->mps_dev));
952 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
953 }
954 
955 static void
956 mps_setup_sysctl(struct mps_softc *sc)
957 {
958 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
959 	struct sysctl_oid	*sysctl_tree = NULL;
960 	char tmpstr[80], tmpstr2[80];
961 
962 	/*
963 	 * Setup the sysctl variable so the user can change the debug level
964 	 * on the fly.
965 	 */
966 	ksnprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
967 	    device_get_unit(sc->mps_dev));
968 	ksnprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
969 
970 	sysctl_ctx_init(&sc->sysctl_ctx);
971 	sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
972 	    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
973 	    CTLFLAG_RD, 0, tmpstr);
974 	if (sc->sysctl_tree == NULL)
975 		return;
976 	sysctl_ctx = &sc->sysctl_ctx;
977 	sysctl_tree = sc->sysctl_tree;
978 
979 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
980 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
981 	    "mps debug level");
982 
983 #if 0 /* XXX swildner */
984 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
985 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
986 	    "Disable the use of MSI-X interrupts");
987 #endif
988 
989 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
990 	    OID_AUTO, "enable_msi", CTLFLAG_RD, &sc->enable_msi, 0,
991 	    "Enable the use of MSI interrupts");
992 
993 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
994 	    OID_AUTO, "firmware_version", CTLFLAG_RW, &sc->fw_version,
995 	    strlen(sc->fw_version), "firmware version");
996 
997 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
998 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
999 	    strlen(MPS_DRIVER_VERSION), "driver version");
1000 
1001 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1002 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1003 	    &sc->io_cmds_active, 0, "number of currently active commands");
1004 
1005 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1006 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1007 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1008 
1009 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1010 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1011 	    &sc->chain_free, 0, "number of free chain elements");
1012 
1013 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1014 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1015 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1016 
1017 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1018 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1019 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1020 
1021 #if __FreeBSD_version >= 900030
1022 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1023 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1024 	    &sc->chain_alloc_fail, "chain allocation failures");
1025 #endif //FreeBSD_version >= 900030
1026 }
1027 
1028 int
1029 mps_attach(struct mps_softc *sc)
1030 {
1031 	int i, error;
1032 
1033 	mps_get_tunables(sc);
1034 
1035 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1036 
1037 	lockinit(&sc->mps_lock, "MPT2SAS lock", 0, LK_CANRECURSE);
1038 	callout_init(&sc->periodic);
1039 	TAILQ_INIT(&sc->event_list);
1040 
1041 	if ((error = mps_transition_ready(sc)) != 0) {
1042 		mps_printf(sc, "%s failed to transition ready\n", __func__);
1043 		return (error);
1044 	}
1045 
1046 	sc->facts = kmalloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1047 	    M_ZERO|M_WAITOK);
1048 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0)
1049 		return (error);
1050 
1051 	mps_print_iocfacts(sc, sc->facts);
1052 
1053 	ksnprintf(sc->fw_version, sizeof(sc->fw_version),
1054 	    "%02d.%02d.%02d.%02d",
1055 	    sc->facts->FWVersion.Struct.Major,
1056 	    sc->facts->FWVersion.Struct.Minor,
1057 	    sc->facts->FWVersion.Struct.Unit,
1058 	    sc->facts->FWVersion.Struct.Dev);
1059 
1060 	mps_printf(sc, "Firmware: %s, Driver: %s\n", sc->fw_version,
1061 	    MPS_DRIVER_VERSION);
1062 	mps_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
1063 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
1064 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
1065 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
1066 
1067 	/*
1068 	 * If the chip doesn't support event replay then a hard reset will be
1069 	 * required to trigger a full discovery.  Do the reset here then
1070 	 * retransition to Ready.  A hard reset might have already been done,
1071 	 * but it doesn't hurt to do it again.
1072 	 */
1073 	if ((sc->facts->IOCCapabilities &
1074 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
1075 		mps_diag_reset(sc);
1076 		if ((error = mps_transition_ready(sc)) != 0)
1077 			return (error);
1078 	}
1079 
1080 	/*
1081 	 * Set flag if IR Firmware is loaded.
1082 	 */
1083 	if (sc->facts->IOCCapabilities &
1084 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
1085 		sc->ir_firmware = 1;
1086 
1087 	/*
1088 	 * Check if controller supports FW diag buffers and set flag to enable
1089 	 * each type.
1090 	 */
1091 	if (sc->facts->IOCCapabilities &
1092 	    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
1093 		sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].enabled =
1094 		    TRUE;
1095 	if (sc->facts->IOCCapabilities &
1096 	    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
1097 		sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].enabled =
1098 		    TRUE;
1099 	if (sc->facts->IOCCapabilities &
1100 	    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
1101 		sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].enabled =
1102 		    TRUE;
1103 
1104 	/*
1105 	 * Set flag if EEDP is supported and if TLR is supported.
1106 	 */
1107 	if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
1108 		sc->eedp_enabled = TRUE;
1109 	if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
1110 		sc->control_TLR = TRUE;
1111 
1112 	/*
1113 	 * Size the queues. Since the reply queues always need one free entry,
1114 	 * we'll just deduct one reply message here.
1115 	 */
1116 	sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
1117 	sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
1118 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
1119 	TAILQ_INIT(&sc->req_list);
1120 	TAILQ_INIT(&sc->high_priority_req_list);
1121 	TAILQ_INIT(&sc->chain_list);
1122 	TAILQ_INIT(&sc->tm_list);
1123 
1124 	if (((error = mps_alloc_queues(sc)) != 0) ||
1125 	    ((error = mps_alloc_replies(sc)) != 0) ||
1126 	    ((error = mps_alloc_requests(sc)) != 0)) {
1127 		mps_printf(sc, "%s failed to alloc\n", __func__);
1128 		mps_free(sc);
1129 		return (error);
1130 	}
1131 
1132 	if (((error = mps_init_queues(sc)) != 0) ||
1133 	    ((error = mps_transition_operational(sc)) != 0)) {
1134 		mps_printf(sc, "%s failed to transition operational\n", __func__);
1135 		mps_free(sc);
1136 		return (error);
1137 	}
1138 
1139 	/*
1140 	 * Finish the queue initialization.
1141 	 * These are set here instead of in mps_init_queues() because the
1142 	 * IOC resets these values during the state transition in
1143 	 * mps_transition_operational().  The free index is set to 1
1144 	 * because the corresponding index in the IOC is set to 0, and the
1145 	 * IOC treats the queues as full if both are set to the same value.
1146 	 * Hence the reason that the queue can't hold all of the possible
1147 	 * replies.
1148 	 */
1149 	sc->replypostindex = 0;
1150 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
1151 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
1152 
1153 	sc->pfacts = kmalloc(sizeof(MPI2_PORT_FACTS_REPLY) *
1154 	    sc->facts->NumberOfPorts, M_MPT2, M_ZERO|M_WAITOK);
1155 	for (i = 0; i < sc->facts->NumberOfPorts; i++) {
1156 		if ((error = mps_get_portfacts(sc, &sc->pfacts[i], i)) != 0) {
1157 			mps_printf(sc, "%s failed to get portfacts for port %d\n",
1158 			    __func__, i);
1159 			mps_free(sc);
1160 			return (error);
1161 		}
1162 		mps_print_portfacts(sc, &sc->pfacts[i]);
1163 	}
1164 
1165 	/* Attach the subsystems so they can prepare their event masks. */
1166 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
1167 	if (((error = mps_attach_log(sc)) != 0) ||
1168 	    ((error = mps_attach_sas(sc)) != 0) ||
1169 	    ((error = mps_attach_user(sc)) != 0)) {
1170 		mps_printf(sc, "%s failed to attach all subsystems: error %d\n",
1171 		    __func__, error);
1172 		mps_free(sc);
1173 		return (error);
1174 	}
1175 
1176 	if ((error = mps_pci_setup_interrupts(sc)) != 0) {
1177 		mps_printf(sc, "%s failed to setup interrupts\n", __func__);
1178 		mps_free(sc);
1179 		return (error);
1180 	}
1181 
1182 	/*
1183 	 * The static page function currently read is ioc page8.  Others can be
1184 	 * added in future.
1185 	 */
1186 	mps_base_static_config_pages(sc);
1187 
1188 	/* Start the periodic watchdog check on the IOC Doorbell */
1189 	mps_periodic(sc);
1190 
1191 	/*
1192 	 * The portenable will kick off discovery events that will drive the
1193 	 * rest of the initialization process.  The CAM/SAS module will
1194 	 * hold up the boot sequence until discovery is complete.
1195 	 */
1196 	sc->mps_ich.ich_func = mps_startup;
1197 	sc->mps_ich.ich_arg = sc;
1198 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1199 		mps_dprint(sc, MPS_FAULT, "Cannot establish MPS config hook\n");
1200 		error = EINVAL;
1201 	}
1202 
1203 	/*
1204 	 * Allow IR to shutdown gracefully when shutdown occurs.
1205 	 */
1206 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1207 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1208 
1209 	if (sc->shutdown_eh == NULL)
1210 		mps_dprint(sc, MPS_FAULT, "shutdown event registration "
1211 		    "failed\n");
1212 
1213 	mps_setup_sysctl(sc);
1214 
1215 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1216 
1217 	return (error);
1218 }
1219 
1220 /* Run through any late-start handlers. */
1221 static void
1222 mps_startup(void *arg)
1223 {
1224 	struct mps_softc *sc;
1225 
1226 	sc = (struct mps_softc *)arg;
1227 
1228 	mps_lock(sc);
1229 	mps_unmask_intr(sc);
1230 	/* initialize device mapping tables */
1231 	mps_mapping_initialize(sc);
1232 	mpssas_startup(sc);
1233 	mps_unlock(sc);
1234 }
1235 
1236 /* Periodic watchdog.  Is called with the driver lock already held. */
1237 static void
1238 mps_periodic(void *arg)
1239 {
1240 	struct mps_softc *sc;
1241 	uint32_t db;
1242 
1243 	sc = (struct mps_softc *)arg;
1244 	mps_lock(sc);
1245 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN) {
1246 		mps_unlock(sc);
1247 		return;
1248 	}
1249 
1250 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1251 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1252 		device_printf(sc->mps_dev, "IOC Fault 0x%08x, Resetting\n", db);
1253 
1254 		mps_reinit(sc);
1255 	}
1256 
1257 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1258 	mps_unlock(sc);
1259 }
1260 
1261 static void
1262 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1263     MPI2_EVENT_NOTIFICATION_REPLY *event)
1264 {
1265 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1266 
1267 	mps_print_event(sc, event);
1268 
1269 	switch (event->Event) {
1270 	case MPI2_EVENT_LOG_DATA:
1271 		device_printf(sc->mps_dev, "MPI2_EVENT_LOG_DATA:\n");
1272 		hexdump(event->EventData, event->EventDataLength, NULL, 0);
1273 		break;
1274 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1275 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1276 		mps_dprint(sc, MPS_INFO, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1277 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1278 		     entry->LogSequence);
1279 		break;
1280 	default:
1281 		break;
1282 	}
1283 	return;
1284 }
1285 
1286 static int
1287 mps_attach_log(struct mps_softc *sc)
1288 {
1289 	uint8_t events[16];
1290 
1291 	bzero(events, 16);
1292 	setbit(events, MPI2_EVENT_LOG_DATA);
1293 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1294 
1295 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1296 	    &sc->mps_log_eh);
1297 
1298 	return (0);
1299 }
1300 
1301 static int
1302 mps_detach_log(struct mps_softc *sc)
1303 {
1304 
1305 	if (sc->mps_log_eh != NULL)
1306 		mps_deregister_events(sc, sc->mps_log_eh);
1307 	return (0);
1308 }
1309 
1310 /*
1311  * Free all of the driver resources and detach submodules.  Should be called
1312  * without the lock held.
1313  */
1314 int
1315 mps_free(struct mps_softc *sc)
1316 {
1317 	struct mps_command *cm;
1318 	int i, error;
1319 
1320 	/* Turn off the watchdog */
1321 	mps_lock(sc);
1322 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1323 	mps_unlock(sc);
1324 #if 0 /* XXX swildner */
1325 	/* Lock must not be held for this */
1326 	callout_drain(&sc->periodic);
1327 #else
1328 	callout_stop(&sc->periodic);
1329 #endif
1330 
1331 	if (((error = mps_detach_log(sc)) != 0) ||
1332 	    ((error = mps_detach_sas(sc)) != 0))
1333 		return (error);
1334 
1335 	/* Put the IOC back in the READY state. */
1336 	mps_lock(sc);
1337 	if ((error = mps_transition_ready(sc)) != 0) {
1338 		mps_unlock(sc);
1339 		return (error);
1340 	}
1341 	mps_unlock(sc);
1342 
1343 	if (sc->facts != NULL)
1344 		kfree(sc->facts, M_MPT2);
1345 
1346 	if (sc->pfacts != NULL)
1347 		kfree(sc->pfacts, M_MPT2);
1348 
1349 	if (sc->post_busaddr != 0)
1350 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
1351 	if (sc->post_queue != NULL)
1352 		bus_dmamem_free(sc->queues_dmat, sc->post_queue,
1353 		    sc->queues_map);
1354 	if (sc->queues_dmat != NULL)
1355 		bus_dma_tag_destroy(sc->queues_dmat);
1356 
1357 	if (sc->chain_busaddr != 0)
1358 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
1359 	if (sc->chain_frames != NULL)
1360 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,sc->chain_map);
1361 	if (sc->chain_dmat != NULL)
1362 		bus_dma_tag_destroy(sc->chain_dmat);
1363 
1364 	if (sc->sense_busaddr != 0)
1365 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
1366 	if (sc->sense_frames != NULL)
1367 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,sc->sense_map);
1368 	if (sc->sense_dmat != NULL)
1369 		bus_dma_tag_destroy(sc->sense_dmat);
1370 
1371 	if (sc->reply_busaddr != 0)
1372 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
1373 	if (sc->reply_frames != NULL)
1374 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,sc->reply_map);
1375 	if (sc->reply_dmat != NULL)
1376 		bus_dma_tag_destroy(sc->reply_dmat);
1377 
1378 	if (sc->req_busaddr != 0)
1379 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
1380 	if (sc->req_frames != NULL)
1381 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
1382 	if (sc->req_dmat != NULL)
1383 		bus_dma_tag_destroy(sc->req_dmat);
1384 
1385 	if (sc->chains != NULL)
1386 		kfree(sc->chains, M_MPT2);
1387 	if (sc->commands != NULL) {
1388 		for (i = 1; i < sc->num_reqs; i++) {
1389 			cm = &sc->commands[i];
1390 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
1391 		}
1392 		kfree(sc->commands, M_MPT2);
1393 	}
1394 	if (sc->buffer_dmat != NULL)
1395 		bus_dma_tag_destroy(sc->buffer_dmat);
1396 
1397 	if (sc->sysctl_tree != NULL)
1398 		sysctl_ctx_free(&sc->sysctl_ctx);
1399 
1400 	mps_mapping_free_memory(sc);
1401 
1402 	/* Deregister the shutdown function */
1403 	if (sc->shutdown_eh != NULL)
1404 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1405 
1406 	lockuninit(&sc->mps_lock);
1407 
1408 	return (0);
1409 }
1410 
1411 static __inline void
1412 mps_complete_command(struct mps_command *cm)
1413 {
1414 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1415 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1416 
1417 	if (cm->cm_complete != NULL) {
1418 		mps_dprint(cm->cm_sc, MPS_TRACE,
1419 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
1420 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
1421 			   cm->cm_reply);
1422 		cm->cm_complete(cm->cm_sc, cm);
1423 	}
1424 
1425 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1426 		mps_dprint(cm->cm_sc, MPS_TRACE, "%s: waking up %p\n",
1427 			   __func__, cm);
1428 		wakeup(cm);
1429 	}
1430 
1431 	if (cm->cm_sc->io_cmds_active != 0) {
1432 		cm->cm_sc->io_cmds_active--;
1433 	} else {
1434 		mps_dprint(cm->cm_sc, MPS_INFO, "Warning: io_cmds_active is "
1435 		    "out of sync - resynching to 0\n");
1436 	}
1437 }
1438 
1439 void
1440 mps_intr(void *data)
1441 {
1442 	struct mps_softc *sc;
1443 	uint32_t status;
1444 
1445 	sc = (struct mps_softc *)data;
1446 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1447 
1448 	/*
1449 	 * Check interrupt status register to flush the bus.  This is
1450 	 * needed for both INTx interrupts and driver-driven polling
1451 	 */
1452 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1453 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1454 		return;
1455 
1456 	mps_lock(sc);
1457 	mps_intr_locked(data);
1458 	mps_unlock(sc);
1459 	return;
1460 }
1461 
1462 /*
1463  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1464  * chip.  Hopefully this theory is correct.
1465  */
1466 void
1467 mps_intr_msi(void *data)
1468 {
1469 	struct mps_softc *sc;
1470 
1471 	sc = (struct mps_softc *)data;
1472 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1473 	mps_lock(sc);
1474 	mps_intr_locked(data);
1475 	mps_unlock(sc);
1476 	return;
1477 }
1478 
1479 /*
1480  * The locking is overly broad and simplistic, but easy to deal with for now.
1481  */
1482 void
1483 mps_intr_locked(void *data)
1484 {
1485 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1486 	struct mps_softc *sc;
1487 	struct mps_command *cm = NULL;
1488 	uint8_t flags;
1489 	u_int pq;
1490 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
1491 	mps_fw_diagnostic_buffer_t *pBuffer;
1492 
1493 	sc = (struct mps_softc *)data;
1494 
1495 	pq = sc->replypostindex;
1496 	mps_dprint(sc, MPS_TRACE,
1497 	    "%s sc %p starting with replypostindex %u\n",
1498 	    __func__, sc, sc->replypostindex);
1499 
1500 	for ( ;; ) {
1501 		cm = NULL;
1502 		desc = &sc->post_queue[sc->replypostindex];
1503 		flags = desc->Default.ReplyFlags &
1504 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1505 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
1506 		 || (desc->Words.High == 0xffffffff))
1507 			break;
1508 
1509 		/* increment the replypostindex now, so that event handlers
1510 		 * and cm completion handlers which decide to do a diag
1511 		 * reset can zero it without it getting incremented again
1512 		 * afterwards, and we break out of this loop on the next
1513 		 * iteration since the reply post queue has been cleared to
1514 		 * 0xFF and all descriptors look unused (which they are).
1515 		 */
1516 		if (++sc->replypostindex >= sc->pqdepth)
1517 			sc->replypostindex = 0;
1518 
1519 		switch (flags) {
1520 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1521 			cm = &sc->commands[desc->SCSIIOSuccess.SMID];
1522 			cm->cm_reply = NULL;
1523 			break;
1524 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1525 		{
1526 			uint32_t baddr;
1527 			uint8_t *reply;
1528 
1529 			/*
1530 			 * Re-compose the reply address from the address
1531 			 * sent back from the chip.  The ReplyFrameAddress
1532 			 * is the lower 32 bits of the physical address of
1533 			 * particular reply frame.  Convert that address to
1534 			 * host format, and then use that to provide the
1535 			 * offset against the virtual address base
1536 			 * (sc->reply_frames).
1537 			 */
1538 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1539 			reply = sc->reply_frames +
1540 				(baddr - ((uint32_t)sc->reply_busaddr));
1541 			/*
1542 			 * Make sure the reply we got back is in a valid
1543 			 * range.  If not, go ahead and panic here, since
1544 			 * we'll probably panic as soon as we deference the
1545 			 * reply pointer anyway.
1546 			 */
1547 			if ((reply < sc->reply_frames)
1548 			 || (reply > (sc->reply_frames +
1549 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1550 				kprintf("%s: WARNING: reply %p out of range!\n",
1551 				       __func__, reply);
1552 				kprintf("%s: reply_frames %p, fqdepth %d, "
1553 				       "frame size %d\n", __func__,
1554 				       sc->reply_frames, sc->fqdepth,
1555 				       sc->facts->ReplyFrameSize * 4);
1556 				kprintf("%s: baddr %#x,\n", __func__, baddr);
1557 				panic("Reply address out of range");
1558 			}
1559 			if (desc->AddressReply.SMID == 0) {
1560 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
1561 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
1562 					/*
1563 					 * If SMID is 0 for Diag Buffer Post,
1564 					 * this implies that the reply is due to
1565 					 * a release function with a status that
1566 					 * the buffer has been released.  Set
1567 					 * the buffer flags accordingly.
1568 					 */
1569 					rel_rep =
1570 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
1571 					if (rel_rep->IOCStatus ==
1572 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
1573 					    {
1574 						pBuffer =
1575 						    &sc->fw_diag_buffer_list[
1576 						    rel_rep->BufferType];
1577 						pBuffer->valid_data = TRUE;
1578 						pBuffer->owned_by_firmware =
1579 						    FALSE;
1580 						pBuffer->immediate = FALSE;
1581 					}
1582 				} else
1583 					mps_dispatch_event(sc, baddr,
1584 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
1585 					    reply);
1586 			} else {
1587 				cm = &sc->commands[desc->AddressReply.SMID];
1588 				cm->cm_reply = reply;
1589 				cm->cm_reply_data =
1590 				    desc->AddressReply.ReplyFrameAddress;
1591 			}
1592 			break;
1593 		}
1594 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
1595 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
1596 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
1597 		default:
1598 			/* Unhandled */
1599 			device_printf(sc->mps_dev, "Unhandled reply 0x%x\n",
1600 			    desc->Default.ReplyFlags);
1601 			cm = NULL;
1602 			break;
1603 		}
1604 
1605 		if (cm != NULL)
1606 			mps_complete_command(cm);
1607 
1608 		desc->Words.Low = 0xffffffff;
1609 		desc->Words.High = 0xffffffff;
1610 	}
1611 
1612 	if (pq != sc->replypostindex) {
1613 		mps_dprint(sc, MPS_TRACE,
1614 		    "%s sc %p writing postindex %d\n",
1615 		    __func__, sc, sc->replypostindex);
1616 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
1617 	}
1618 
1619 	return;
1620 }
1621 
1622 static void
1623 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
1624     MPI2_EVENT_NOTIFICATION_REPLY *reply)
1625 {
1626 	struct mps_event_handle *eh;
1627 	int event, handled = 0;
1628 
1629 	event = reply->Event;
1630 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1631 		if (isset(eh->mask, event)) {
1632 			eh->callback(sc, data, reply);
1633 			handled++;
1634 		}
1635 	}
1636 
1637 	if (handled == 0)
1638 		device_printf(sc->mps_dev, "Unhandled event 0x%x\n", event);
1639 
1640 	/*
1641 	 * This is the only place that the event/reply should be freed.
1642 	 * Anything wanting to hold onto the event data should have
1643 	 * already copied it into their own storage.
1644 	 */
1645 	mps_free_reply(sc, data);
1646 }
1647 
1648 static void
1649 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
1650 {
1651 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1652 
1653 	if (cm->cm_reply)
1654 		mps_print_event(sc,
1655 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
1656 
1657 	mps_free_command(sc, cm);
1658 
1659 	/* next, send a port enable */
1660 	mpssas_startup(sc);
1661 }
1662 
1663 /*
1664  * For both register_events and update_events, the caller supplies a bitmap
1665  * of events that it _wants_.  These functions then turn that into a bitmask
1666  * suitable for the controller.
1667  */
1668 int
1669 mps_register_events(struct mps_softc *sc, uint8_t *mask,
1670     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
1671 {
1672 	struct mps_event_handle *eh;
1673 	int error = 0;
1674 
1675 	eh = kmalloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
1676 	eh->callback = cb;
1677 	eh->data = data;
1678 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
1679 	if (mask != NULL)
1680 		error = mps_update_events(sc, eh, mask);
1681 	*handle = eh;
1682 
1683 	return (error);
1684 }
1685 
1686 int
1687 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
1688     uint8_t *mask)
1689 {
1690 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
1691 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
1692 	struct mps_command *cm;
1693 	struct mps_event_handle *eh;
1694 	int error, i;
1695 
1696 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1697 
1698 	if ((mask != NULL) && (handle != NULL))
1699 		bcopy(mask, &handle->mask[0], 16);
1700 	memset(sc->event_mask, 0xff, 16);
1701 
1702 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1703 		for (i = 0; i < 16; i++)
1704 			sc->event_mask[i] &= ~eh->mask[i];
1705 	}
1706 
1707 	if ((cm = mps_alloc_command(sc)) == NULL)
1708 		return (EBUSY);
1709 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
1710 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
1711 	evtreq->MsgFlags = 0;
1712 	evtreq->SASBroadcastPrimitiveMasks = 0;
1713 #ifdef MPS_DEBUG_ALL_EVENTS
1714 	{
1715 		u_char fullmask[16];
1716 		memset(fullmask, 0x00, 16);
1717 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
1718 	}
1719 #else
1720 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
1721 #endif
1722 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
1723 	cm->cm_data = NULL;
1724 
1725 	error = mps_request_polled(sc, cm);
1726 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
1727 	if ((reply == NULL) ||
1728 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1729 		error = ENXIO;
1730 	mps_print_event(sc, reply);
1731 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
1732 
1733 	mps_free_command(sc, cm);
1734 	return (error);
1735 }
1736 
1737 static int
1738 mps_reregister_events(struct mps_softc *sc)
1739 {
1740 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
1741 	struct mps_command *cm;
1742 	struct mps_event_handle *eh;
1743 	int error, i;
1744 
1745 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1746 
1747 	/* first, reregister events */
1748 
1749 	memset(sc->event_mask, 0xff, 16);
1750 
1751 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1752 		for (i = 0; i < 16; i++)
1753 			sc->event_mask[i] &= ~eh->mask[i];
1754 	}
1755 
1756 	if ((cm = mps_alloc_command(sc)) == NULL)
1757 		return (EBUSY);
1758 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
1759 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
1760 	evtreq->MsgFlags = 0;
1761 	evtreq->SASBroadcastPrimitiveMasks = 0;
1762 #ifdef MPS_DEBUG_ALL_EVENTS
1763 	{
1764 		u_char fullmask[16];
1765 		memset(fullmask, 0x00, 16);
1766 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
1767 	}
1768 #else
1769 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
1770 #endif
1771 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
1772 	cm->cm_data = NULL;
1773 	cm->cm_complete = mps_reregister_events_complete;
1774 
1775 	error = mps_map_command(sc, cm);
1776 
1777 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__, error);
1778 	return (error);
1779 }
1780 
1781 int
1782 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
1783 {
1784 
1785 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
1786 	kfree(handle, M_MPT2);
1787 	return (mps_update_events(sc, NULL, NULL));
1788 }
1789 
1790 /*
1791  * Add a chain element as the next SGE for the specified command.
1792  * Reset cm_sge and cm_sgesize to indicate all the available space.
1793  */
1794 static int
1795 mps_add_chain(struct mps_command *cm)
1796 {
1797 	MPI2_SGE_CHAIN32 *sgc;
1798 	struct mps_chain *chain;
1799 	int space;
1800 
1801 	if (cm->cm_sglsize < MPS_SGC_SIZE)
1802 		panic("MPS: Need SGE Error Code");
1803 
1804 	chain = mps_alloc_chain(cm->cm_sc);
1805 	if (chain == NULL)
1806 		return (ENOBUFS);
1807 
1808 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
1809 
1810 	/*
1811 	 * Note: a double-linked list is used to make it easier to
1812 	 * walk for debugging.
1813 	 */
1814 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
1815 
1816 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
1817 	sgc->Length = space;
1818 	sgc->NextChainOffset = 0;
1819 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
1820 	sgc->Address = chain->chain_busaddr;
1821 
1822 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
1823 	cm->cm_sglsize = space;
1824 	return (0);
1825 }
1826 
1827 /*
1828  * Add one scatter-gather element (chain, simple, transaction context)
1829  * to the scatter-gather list for a command.  Maintain cm_sglsize and
1830  * cm_sge as the remaining size and pointer to the next SGE to fill
1831  * in, respectively.
1832  */
1833 int
1834 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
1835 {
1836 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
1837 	MPI2_SGE_SIMPLE64 *sge = sgep;
1838 	int error, type;
1839 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
1840 
1841 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
1842 
1843 #ifdef INVARIANTS
1844 	switch (type) {
1845 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
1846 		if (len != tc->DetailsLength + 4)
1847 			panic("TC %p length %u or %zu?", tc,
1848 			    tc->DetailsLength + 4, len);
1849 		}
1850 		break;
1851 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
1852 		/* Driver only uses 32-bit chain elements */
1853 		if (len != MPS_SGC_SIZE)
1854 			panic("CHAIN %p length %u or %zu?", sgep,
1855 			    MPS_SGC_SIZE, len);
1856 		break;
1857 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
1858 		/* Driver only uses 64-bit SGE simple elements */
1859 		sge = sgep;
1860 		if (len != MPS_SGE64_SIZE)
1861 			panic("SGE simple %p length %u or %zu?", sge,
1862 			    MPS_SGE64_SIZE, len);
1863 		if (((sge->FlagsLength >> MPI2_SGE_FLAGS_SHIFT) &
1864 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
1865 			panic("SGE simple %p flags %02x not marked 64-bit?",
1866 			    sge, sge->FlagsLength >> MPI2_SGE_FLAGS_SHIFT);
1867 
1868 		break;
1869 	default:
1870 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
1871 	}
1872 #endif
1873 
1874 	/*
1875 	 * case 1: 1 more segment, enough room for it
1876 	 * case 2: 2 more segments, enough room for both
1877 	 * case 3: >=2 more segments, only enough room for 1 and a chain
1878 	 * case 4: >=1 more segment, enough room for only a chain
1879 	 * case 5: >=1 more segment, no room for anything (error)
1880          */
1881 
1882 	/*
1883 	 * There should be room for at least a chain element, or this
1884 	 * code is buggy.  Case (5).
1885 	 */
1886 	if (cm->cm_sglsize < MPS_SGC_SIZE)
1887 		panic("MPS: Need SGE Error Code");
1888 
1889 	if (segsleft >= 2 &&
1890 	    cm->cm_sglsize >= len + MPS_SGC_SIZE &&
1891 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
1892 		/*
1893 		 * There are 2 or more segments left to add, and only
1894 		 * enough room for 1 and a chain.  Case (3).
1895 		 *
1896 		 * Mark as last element in this chain if necessary.
1897 		 */
1898 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
1899 			sge->FlagsLength |=
1900 				(MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
1901 		}
1902 
1903 		/*
1904 		 * Add the item then a chain.  Do the chain now,
1905 		 * rather than on the next iteration, to simplify
1906 		 * understanding the code.
1907 		 */
1908 		cm->cm_sglsize -= len;
1909 		bcopy(sgep, cm->cm_sge, len);
1910 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
1911 		return (mps_add_chain(cm));
1912 	}
1913 
1914 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
1915 		/*
1916 		 * 1 or more segment, enough room for only a chain.
1917 		 * Hope the previous element wasn't a Simple entry
1918 		 * that needed to be marked with
1919 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
1920 		 */
1921 		if ((error = mps_add_chain(cm)) != 0)
1922 			return (error);
1923 	}
1924 
1925 #ifdef INVARIANTS
1926 	/* Case 1: 1 more segment, enough room for it. */
1927 	if (segsleft == 1 && cm->cm_sglsize < len)
1928 		panic("1 seg left and no room? %u versus %zu",
1929 		    cm->cm_sglsize, len);
1930 
1931 	/* Case 2: 2 more segments, enough room for both */
1932 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
1933 		panic("2 segs left and no room? %u versus %zu",
1934 		    cm->cm_sglsize, len);
1935 #endif
1936 
1937 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
1938 		/*
1939 		 * If this is a bi-directional request, need to account for that
1940 		 * here.  Save the pre-filled sge values.  These will be used
1941 		 * either for the 2nd SGL or for a single direction SGL.  If
1942 		 * cm_out_len is non-zero, this is a bi-directional request, so
1943 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
1944 		 * fill in the IN SGL.  Note that at this time, when filling in
1945 		 * 2 SGL's for a bi-directional request, they both use the same
1946 		 * DMA buffer (same cm command).
1947 		 */
1948 		saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
1949 		saved_address_low = sge->Address.Low;
1950 		saved_address_high = sge->Address.High;
1951 		if (cm->cm_out_len) {
1952 			sge->FlagsLength = cm->cm_out_len |
1953 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
1954 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
1955 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
1956 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
1957 			    MPI2_SGE_FLAGS_SHIFT);
1958 			cm->cm_sglsize -= len;
1959 			bcopy(sgep, cm->cm_sge, len);
1960 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
1961 			    + len);
1962 		}
1963 		sge->FlagsLength = saved_buf_len |
1964 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
1965 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
1966 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
1967 		    MPI2_SGE_FLAGS_END_OF_LIST |
1968 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
1969 		    MPI2_SGE_FLAGS_SHIFT);
1970 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
1971 			sge->FlagsLength |=
1972 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
1973 			    MPI2_SGE_FLAGS_SHIFT);
1974 		} else {
1975 			sge->FlagsLength |=
1976 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
1977 			    MPI2_SGE_FLAGS_SHIFT);
1978 		}
1979 		sge->Address.Low = saved_address_low;
1980 		sge->Address.High = saved_address_high;
1981 	}
1982 
1983 	cm->cm_sglsize -= len;
1984 	bcopy(sgep, cm->cm_sge, len);
1985 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
1986 	return (0);
1987 }
1988 
1989 /*
1990  * Add one dma segment to the scatter-gather list for a command.
1991  */
1992 int
1993 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
1994     int segsleft)
1995 {
1996 	MPI2_SGE_SIMPLE64 sge;
1997 
1998 	/*
1999 	 * This driver always uses 64-bit address elements for simplicity.
2000 	 */
2001 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2002 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2003 	sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
2004 	mps_from_u64(pa, &sge.Address);
2005 
2006 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2007 }
2008 
2009 static void
2010 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2011 {
2012 	struct mps_softc *sc;
2013 	struct mps_command *cm;
2014 	u_int i, dir, sflags;
2015 
2016 	cm = (struct mps_command *)arg;
2017 	sc = cm->cm_sc;
2018 
2019 	/*
2020 	 * In this case, just print out a warning and let the chip tell the
2021 	 * user they did the wrong thing.
2022 	 */
2023 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2024 		mps_printf(sc, "%s: warning: busdma returned %d segments, "
2025 			   "more than the %d allowed\n", __func__, nsegs,
2026 			   cm->cm_max_segs);
2027 	}
2028 
2029 	/*
2030 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2031 	 * here.  In that case, both direction flags will be set.
2032 	 */
2033 	sflags = 0;
2034 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2035 		/*
2036 		 * We have to add a special case for SMP passthrough, there
2037 		 * is no easy way to generically handle it.  The first
2038 		 * S/G element is used for the command (therefore the
2039 		 * direction bit needs to be set).  The second one is used
2040 		 * for the reply.  We'll leave it to the caller to make
2041 		 * sure we only have two buffers.
2042 		 */
2043 		/*
2044 		 * Even though the busdma man page says it doesn't make
2045 		 * sense to have both direction flags, it does in this case.
2046 		 * We have one s/g element being accessed in each direction.
2047 		 */
2048 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2049 
2050 		/*
2051 		 * Set the direction flag on the first buffer in the SMP
2052 		 * passthrough request.  We'll clear it for the second one.
2053 		 */
2054 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2055 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2056 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2057 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2058 		dir = BUS_DMASYNC_PREWRITE;
2059 	} else
2060 		dir = BUS_DMASYNC_PREREAD;
2061 
2062 	for (i = 0; i < nsegs; i++) {
2063 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2064 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2065 		}
2066 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2067 		    sflags, nsegs - i);
2068 		if (error != 0) {
2069 			/* Resource shortage, roll back! */
2070 			mps_dprint(sc, MPS_INFO, "out of chain frames\n");
2071 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2072 			mps_complete_command(cm);
2073 			return;
2074 		}
2075 	}
2076 
2077 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2078 	mps_enqueue_request(sc, cm);
2079 
2080 	return;
2081 }
2082 
2083 static void
2084 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2085 	     int error)
2086 {
2087 	mps_data_cb(arg, segs, nsegs, error);
2088 }
2089 
2090 /*
2091  * This is the routine to enqueue commands ansynchronously.
2092  * Note that the only error path here is from bus_dmamap_load(), which can
2093  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2094  * assumed that if you have a command in-hand, then you have enough credits
2095  * to use it.
2096  */
2097 int
2098 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2099 {
2100 	MPI2_SGE_SIMPLE32 *sge;
2101 	int error = 0;
2102 
2103 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2104 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2105 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2106 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2107 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2108 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2109 	} else {
2110 		/* Add a zero-length element as needed */
2111 		if (cm->cm_sge != NULL) {
2112 			sge = (MPI2_SGE_SIMPLE32 *)cm->cm_sge;
2113 			sge->FlagsLength = (MPI2_SGE_FLAGS_LAST_ELEMENT |
2114 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2115 			    MPI2_SGE_FLAGS_END_OF_LIST |
2116 			    MPI2_SGE_FLAGS_SIMPLE_ELEMENT) <<
2117 			    MPI2_SGE_FLAGS_SHIFT;
2118 			sge->Address = 0;
2119 		}
2120 		mps_enqueue_request(sc, cm);
2121 	}
2122 
2123 	return (error);
2124 }
2125 
2126 /*
2127  * This is the routine to enqueue commands synchronously.  An error of
2128  * EINPROGRESS from mps_map_command() is ignored since the command will
2129  * be executed and enqueued automatically.  Other errors come from msleep().
2130  */
2131 int
2132 mps_wait_command(struct mps_softc *sc, struct mps_command *cm, int timeout)
2133 {
2134 	int error;
2135 
2136 	KKASSERT(lockstatus(&sc->mps_lock, curthread) != 0);
2137 
2138 	cm->cm_complete = NULL;
2139 	cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2140 	error = mps_map_command(sc, cm);
2141 	if ((error != 0) && (error != EINPROGRESS))
2142 		return (error);
2143 	error = lksleep(cm, &sc->mps_lock, 0, "mpswait", timeout);
2144 	if (error == EWOULDBLOCK)
2145 		error = ETIMEDOUT;
2146 	return (error);
2147 }
2148 
2149 /*
2150  * This is the routine to enqueue a command synchonously and poll for
2151  * completion.  Its use should be rare.
2152  */
2153 int
2154 mps_request_polled(struct mps_softc *sc, struct mps_command *cm)
2155 {
2156 	int error, timeout = 0;
2157 
2158 	error = 0;
2159 
2160 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2161 	cm->cm_complete = NULL;
2162 	mps_map_command(sc, cm);
2163 
2164 	while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2165 		mps_intr_locked(sc);
2166 		DELAY(50 * 1000);
2167 		if (timeout++ > 1000) {
2168 			mps_dprint(sc, MPS_FAULT, "polling failed\n");
2169 			error = ETIMEDOUT;
2170 			break;
2171 		}
2172 	}
2173 
2174 	return (error);
2175 }
2176 
2177 /*
2178  * The MPT driver had a verbose interface for config pages.  In this driver,
2179  * reduce it to much simplier terms, similar to the Linux driver.
2180  */
2181 int
2182 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2183 {
2184 	MPI2_CONFIG_REQUEST *req;
2185 	struct mps_command *cm;
2186 	int error;
2187 
2188 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2189 		return (EBUSY);
2190 	}
2191 
2192 	cm = mps_alloc_command(sc);
2193 	if (cm == NULL) {
2194 		return (EBUSY);
2195 	}
2196 
2197 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2198 	req->Function = MPI2_FUNCTION_CONFIG;
2199 	req->Action = params->action;
2200 	req->SGLFlags = 0;
2201 	req->ChainOffset = 0;
2202 	req->PageAddress = params->page_address;
2203 	if (params->hdr.Ext.ExtPageType != 0) {
2204 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2205 
2206 		hdr = &params->hdr.Ext;
2207 		req->ExtPageType = hdr->ExtPageType;
2208 		req->ExtPageLength = hdr->ExtPageLength;
2209 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2210 		req->Header.PageLength = 0; /* Must be set to zero */
2211 		req->Header.PageNumber = hdr->PageNumber;
2212 		req->Header.PageVersion = hdr->PageVersion;
2213 	} else {
2214 		MPI2_CONFIG_PAGE_HEADER *hdr;
2215 
2216 		hdr = &params->hdr.Struct;
2217 		req->Header.PageType = hdr->PageType;
2218 		req->Header.PageNumber = hdr->PageNumber;
2219 		req->Header.PageLength = hdr->PageLength;
2220 		req->Header.PageVersion = hdr->PageVersion;
2221 	}
2222 
2223 	cm->cm_data = params->buffer;
2224 	cm->cm_length = params->length;
2225 	cm->cm_sge = &req->PageBufferSGE;
2226 	cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2227 	cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2228 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2229 
2230 	cm->cm_complete_data = params;
2231 	if (params->callback != NULL) {
2232 		cm->cm_complete = mps_config_complete;
2233 		return (mps_map_command(sc, cm));
2234 	} else {
2235 		error = mps_wait_command(sc, cm, 0);
2236 		if (error) {
2237 			mps_dprint(sc, MPS_FAULT,
2238 			    "Error %d reading config page\n", error);
2239 			mps_free_command(sc, cm);
2240 			return (error);
2241 		}
2242 		mps_config_complete(sc, cm);
2243 	}
2244 
2245 	return (0);
2246 }
2247 
2248 int
2249 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2250 {
2251 	return (EINVAL);
2252 }
2253 
2254 static void
2255 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
2256 {
2257 	MPI2_CONFIG_REPLY *reply;
2258 	struct mps_config_params *params;
2259 
2260 	params = cm->cm_complete_data;
2261 
2262 	if (cm->cm_data != NULL) {
2263 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2264 		    BUS_DMASYNC_POSTREAD);
2265 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2266 	}
2267 
2268 	/*
2269 	 * XXX KDM need to do more error recovery?  This results in the
2270 	 * device in question not getting probed.
2271 	 */
2272 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
2273 		params->status = MPI2_IOCSTATUS_BUSY;
2274 		goto done;
2275 	}
2276 
2277 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2278 	if (reply == NULL) {
2279 		params->status = MPI2_IOCSTATUS_BUSY;
2280 		goto done;
2281 	}
2282 	params->status = reply->IOCStatus;
2283 	if (params->hdr.Ext.ExtPageType != 0) {
2284 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2285 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2286 	} else {
2287 		params->hdr.Struct.PageType = reply->Header.PageType;
2288 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2289 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2290 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2291 	}
2292 
2293 done:
2294 	mps_free_command(sc, cm);
2295 	if (params->callback != NULL)
2296 		params->callback(sc, params);
2297 
2298 	return;
2299 }
2300