xref: /freebsd/sys/dev/mps/mps.c (revision 884a2a699669ec61e2366e3e358342dbc94be24a)
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 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 /* Communications core for LSI MPT2 */
31 
32 #include <sys/types.h>
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/selinfo.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/module.h>
40 #include <sys/bus.h>
41 #include <sys/conf.h>
42 #include <sys/bio.h>
43 #include <sys/malloc.h>
44 #include <sys/uio.h>
45 #include <sys/sysctl.h>
46 #include <sys/endian.h>
47 
48 #include <machine/bus.h>
49 #include <machine/resource.h>
50 #include <sys/rman.h>
51 
52 #include <cam/scsi/scsi_all.h>
53 
54 #include <dev/mps/mpi/mpi2_type.h>
55 #include <dev/mps/mpi/mpi2.h>
56 #include <dev/mps/mpi/mpi2_ioc.h>
57 #include <dev/mps/mpi/mpi2_cnfg.h>
58 #include <dev/mps/mpsvar.h>
59 #include <dev/mps/mps_table.h>
60 
61 static void mps_startup(void *arg);
62 static void mps_startup_complete(struct mps_softc *sc, struct mps_command *cm);
63 static int mps_send_iocinit(struct mps_softc *sc);
64 static int mps_attach_log(struct mps_softc *sc);
65 static __inline void mps_complete_command(struct mps_command *cm);
66 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data, MPI2_EVENT_NOTIFICATION_REPLY *reply);
67 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
68 static void mps_periodic(void *);
69 
70 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
71 
72 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
73 
74 /*
75  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
76  * any state and back to its initialization state machine.
77  */
78 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
79 
80 static int
81 mps_hard_reset(struct mps_softc *sc)
82 {
83 	uint32_t reg;
84 	int i, error, tries = 0;
85 
86 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
87 
88 	/* Clear any pending interrupts */
89 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
90 
91 	/* Push the magic sequence */
92 	error = ETIMEDOUT;
93 	while (tries++ < 20) {
94 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
95 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
96 			    mpt2_reset_magic[i]);
97 
98 		DELAY(100 * 1000);
99 
100 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
101 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
102 			error = 0;
103 			break;
104 		}
105 	}
106 	if (error)
107 		return (error);
108 
109 	/* Send the actual reset.  XXX need to refresh the reg? */
110 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
111 	    reg | MPI2_DIAG_RESET_ADAPTER);
112 
113 	/* Wait up to 300 seconds in 50ms intervals */
114 	error = ETIMEDOUT;
115 	for (i = 0; i < 60000; i++) {
116 		DELAY(50000);
117 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
118 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
119 			error = 0;
120 			break;
121 		}
122 	}
123 	if (error)
124 		return (error);
125 
126 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
127 
128 	return (0);
129 }
130 
131 static int
132 mps_soft_reset(struct mps_softc *sc)
133 {
134 
135 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
136 
137 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
138 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
139 	    MPI2_DOORBELL_FUNCTION_SHIFT);
140 	DELAY(50000);
141 
142 	return (0);
143 }
144 
145 static int
146 mps_transition_ready(struct mps_softc *sc)
147 {
148 	uint32_t reg, state;
149 	int error, tries = 0;
150 
151 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
152 
153 	error = 0;
154 	while (tries++ < 5) {
155 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
156 		mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
157 
158 		/*
159 		 * Ensure the IOC is ready to talk.  If it's not, try
160 		 * resetting it.
161 		 */
162 		if (reg & MPI2_DOORBELL_USED) {
163 			mps_hard_reset(sc);
164 			DELAY(50000);
165 			continue;
166 		}
167 
168 		/* Is the adapter owned by another peer? */
169 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
170 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
171 			device_printf(sc->mps_dev, "IOC is under the control "
172 			    "of another peer host, aborting initialization.\n");
173 			return (ENXIO);
174 		}
175 
176 		state = reg & MPI2_IOC_STATE_MASK;
177 		if (state == MPI2_IOC_STATE_READY) {
178 			/* Ready to go! */
179 			error = 0;
180 			break;
181 		} else if (state == MPI2_IOC_STATE_FAULT) {
182 			mps_dprint(sc, MPS_INFO, "IOC in fault state 0x%x\n",
183 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
184 			mps_hard_reset(sc);
185 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
186 			/* Need to take ownership */
187 			mps_soft_reset(sc);
188 		} else if (state == MPI2_IOC_STATE_RESET) {
189 			/* Wait a bit, IOC might be in transition */
190 			mps_dprint(sc, MPS_FAULT,
191 			    "IOC in unexpected reset state\n");
192 		} else {
193 			mps_dprint(sc, MPS_FAULT,
194 			    "IOC in unknown state 0x%x\n", state);
195 			error = EINVAL;
196 			break;
197 		}
198 
199 		/* Wait 50ms for things to settle down. */
200 		DELAY(50000);
201 	}
202 
203 	if (error)
204 		device_printf(sc->mps_dev, "Cannot transition IOC to ready\n");
205 
206 	return (error);
207 }
208 
209 static int
210 mps_transition_operational(struct mps_softc *sc)
211 {
212 	uint32_t reg, state;
213 	int error;
214 
215 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
216 
217 	error = 0;
218 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
219 	mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
220 
221 	state = reg & MPI2_IOC_STATE_MASK;
222 	if (state != MPI2_IOC_STATE_READY) {
223 		if ((error = mps_transition_ready(sc)) != 0)
224 			return (error);
225 	}
226 
227 	error = mps_send_iocinit(sc);
228 	return (error);
229 }
230 
231 /* Wait for the chip to ACK a word that we've put into its FIFO */
232 static int
233 mps_wait_db_ack(struct mps_softc *sc)
234 {
235 	int retry;
236 
237 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
238 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
239 		    MPI2_HIS_SYS2IOC_DB_STATUS) == 0)
240 			return (0);
241 		DELAY(2000);
242 	}
243 	return (ETIMEDOUT);
244 }
245 
246 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
247 static int
248 mps_wait_db_int(struct mps_softc *sc)
249 {
250 	int retry;
251 
252 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
253 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
254 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
255 			return (0);
256 		DELAY(2000);
257 	}
258 	return (ETIMEDOUT);
259 }
260 
261 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
262 static int
263 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
264     int req_sz, int reply_sz, int timeout)
265 {
266 	uint32_t *data32;
267 	uint16_t *data16;
268 	int i, count, ioc_sz, residual;
269 
270 	/* Step 1 */
271 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
272 
273 	/* Step 2 */
274 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
275 		return (EBUSY);
276 
277 	/* Step 3
278 	 * Announce that a message is coming through the doorbell.  Messages
279 	 * are pushed at 32bit words, so round up if needed.
280 	 */
281 	count = (req_sz + 3) / 4;
282 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
283 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
284 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
285 
286 	/* Step 4 */
287 	if (mps_wait_db_int(sc) ||
288 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
289 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
290 		return (ENXIO);
291 	}
292 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
293 	if (mps_wait_db_ack(sc) != 0) {
294 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
295 		return (ENXIO);
296 	}
297 
298 	/* Step 5 */
299 	/* Clock out the message data synchronously in 32-bit dwords*/
300 	data32 = (uint32_t *)req;
301 	for (i = 0; i < count; i++) {
302 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, data32[i]);
303 		if (mps_wait_db_ack(sc) != 0) {
304 			mps_dprint(sc, MPS_FAULT,
305 			    "Timeout while writing doorbell\n");
306 			return (ENXIO);
307 		}
308 	}
309 
310 	/* Step 6 */
311 	/* Clock in the reply in 16-bit words.  The total length of the
312 	 * message is always in the 4th byte, so clock out the first 2 words
313 	 * manually, then loop the rest.
314 	 */
315 	data16 = (uint16_t *)reply;
316 	if (mps_wait_db_int(sc) != 0) {
317 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
318 		return (ENXIO);
319 	}
320 	data16[0] =
321 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
322 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
323 	if (mps_wait_db_int(sc) != 0) {
324 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
325 		return (ENXIO);
326 	}
327 	data16[1] =
328 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
329 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
330 
331 	/* Number of 32bit words in the message */
332 	ioc_sz = reply->MsgLength;
333 
334 	/*
335 	 * Figure out how many 16bit words to clock in without overrunning.
336 	 * The precision loss with dividing reply_sz can safely be
337 	 * ignored because the messages can only be multiples of 32bits.
338 	 */
339 	residual = 0;
340 	count = MIN((reply_sz / 4), ioc_sz) * 2;
341 	if (count < ioc_sz * 2) {
342 		residual = ioc_sz * 2 - count;
343 		mps_dprint(sc, MPS_FAULT, "Driver error, throwing away %d "
344 		    "residual message words\n", residual);
345 	}
346 
347 	for (i = 2; i < count; i++) {
348 		if (mps_wait_db_int(sc) != 0) {
349 			mps_dprint(sc, MPS_FAULT,
350 			    "Timeout reading doorbell %d\n", i);
351 			return (ENXIO);
352 		}
353 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
354 		    MPI2_DOORBELL_DATA_MASK;
355 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
356 	}
357 
358 	/*
359 	 * Pull out residual words that won't fit into the provided buffer.
360 	 * This keeps the chip from hanging due to a driver programming
361 	 * error.
362 	 */
363 	while (residual--) {
364 		if (mps_wait_db_int(sc) != 0) {
365 			mps_dprint(sc, MPS_FAULT,
366 			    "Timeout reading doorbell\n");
367 			return (ENXIO);
368 		}
369 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
370 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
371 	}
372 
373 	/* Step 7 */
374 	if (mps_wait_db_int(sc) != 0) {
375 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
376 		return (ENXIO);
377 	}
378 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
379 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
380 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
381 
382 	return (0);
383 }
384 
385 void
386 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
387 {
388 
389 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
390 
391 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
392 		mtx_assert(&sc->mps_mtx, MA_OWNED);
393 
394 	if ((cm->cm_desc.Default.SMID < 1)
395 	 || (cm->cm_desc.Default.SMID >= sc->num_reqs)) {
396 		mps_printf(sc, "%s: invalid SMID %d, desc %#x %#x\n",
397 			   __func__, cm->cm_desc.Default.SMID,
398 			   cm->cm_desc.Words.High, cm->cm_desc.Words.Low);
399 	}
400 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
401 	    cm->cm_desc.Words.Low);
402 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
403 	    cm->cm_desc.Words.High);
404 }
405 
406 int
407 mps_request_polled(struct mps_softc *sc, struct mps_command *cm)
408 {
409 	int error, timeout = 0;
410 
411 	error = 0;
412 
413 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
414 	cm->cm_complete = NULL;
415 	mps_map_command(sc, cm);
416 
417 	while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
418 		mps_intr(sc);
419 		DELAY(50 * 1000);
420 		if (timeout++ > 1000) {
421 			mps_dprint(sc, MPS_FAULT, "polling failed\n");
422 			error = ETIMEDOUT;
423 			break;
424 		}
425 	}
426 
427 	return (error);
428 }
429 
430 /*
431  * Just the FACTS, ma'am.
432  */
433 static int
434 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
435 {
436 	MPI2_DEFAULT_REPLY *reply;
437 	MPI2_IOC_FACTS_REQUEST request;
438 	int error, req_sz, reply_sz;
439 
440 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
441 
442 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
443 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
444 	reply = (MPI2_DEFAULT_REPLY *)facts;
445 
446 	bzero(&request, req_sz);
447 	request.Function = MPI2_FUNCTION_IOC_FACTS;
448 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
449 
450 	return (error);
451 }
452 
453 static int
454 mps_get_portfacts(struct mps_softc *sc, MPI2_PORT_FACTS_REPLY *facts, int port)
455 {
456 	MPI2_PORT_FACTS_REQUEST *request;
457 	MPI2_PORT_FACTS_REPLY *reply;
458 	struct mps_command *cm;
459 	int error;
460 
461 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
462 
463 	if ((cm = mps_alloc_command(sc)) == NULL)
464 		return (EBUSY);
465 	request = (MPI2_PORT_FACTS_REQUEST *)cm->cm_req;
466 	request->Function = MPI2_FUNCTION_PORT_FACTS;
467 	request->PortNumber = port;
468 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
469 	cm->cm_data = NULL;
470 	error = mps_request_polled(sc, cm);
471 	reply = (MPI2_PORT_FACTS_REPLY *)cm->cm_reply;
472 	if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
473 		error = ENXIO;
474 	bcopy(reply, facts, sizeof(MPI2_PORT_FACTS_REPLY));
475 	mps_free_command(sc, cm);
476 
477 	return (error);
478 }
479 
480 static int
481 mps_send_iocinit(struct mps_softc *sc)
482 {
483 	MPI2_IOC_INIT_REQUEST	init;
484 	MPI2_DEFAULT_REPLY	reply;
485 	int req_sz, reply_sz, error;
486 
487 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
488 
489 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
490 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
491 	bzero(&init, req_sz);
492 	bzero(&reply, reply_sz);
493 
494 	/*
495 	 * Fill in the init block.  Note that most addresses are
496 	 * deliberately in the lower 32bits of memory.  This is a micro-
497 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
498 	 */
499 	init.Function = MPI2_FUNCTION_IOC_INIT;
500 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
501 	init.MsgVersion = MPI2_VERSION;
502 	init.HeaderVersion = MPI2_HEADER_VERSION;
503 	init.SystemRequestFrameSize = sc->facts->IOCRequestFrameSize;
504 	init.ReplyDescriptorPostQueueDepth = sc->pqdepth;
505 	init.ReplyFreeQueueDepth = sc->fqdepth;
506 	init.SenseBufferAddressHigh = 0;
507 	init.SystemReplyAddressHigh = 0;
508 	init.SystemRequestFrameBaseAddress.High = 0;
509 	init.SystemRequestFrameBaseAddress.Low = (uint32_t)sc->req_busaddr;
510 	init.ReplyDescriptorPostQueueAddress.High = 0;
511 	init.ReplyDescriptorPostQueueAddress.Low = (uint32_t)sc->post_busaddr;
512 	init.ReplyFreeQueueAddress.High = 0;
513 	init.ReplyFreeQueueAddress.Low = (uint32_t)sc->free_busaddr;
514 	init.TimeStamp.High = 0;
515 	init.TimeStamp.Low = (uint32_t)time_uptime;
516 
517 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
518 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
519 		error = ENXIO;
520 
521 	mps_dprint(sc, MPS_INFO, "IOCInit status= 0x%x\n", reply.IOCStatus);
522 	return (error);
523 }
524 
525 static int
526 mps_send_portenable(struct mps_softc *sc)
527 {
528 	MPI2_PORT_ENABLE_REQUEST *request;
529 	struct mps_command *cm;
530 
531 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
532 
533 	if ((cm = mps_alloc_command(sc)) == NULL)
534 		return (EBUSY);
535 	request = (MPI2_PORT_ENABLE_REQUEST *)cm->cm_req;
536 	request->Function = MPI2_FUNCTION_PORT_ENABLE;
537 	request->MsgFlags = 0;
538 	request->VP_ID = 0;
539 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
540 	cm->cm_complete = mps_startup_complete;
541 
542 	mps_enqueue_request(sc, cm);
543 	return (0);
544 }
545 
546 static int
547 mps_send_mur(struct mps_softc *sc)
548 {
549 
550 	/* Placeholder */
551 	return (0);
552 }
553 
554 void
555 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
556 {
557 	bus_addr_t *addr;
558 
559 	addr = arg;
560 	*addr = segs[0].ds_addr;
561 }
562 
563 static int
564 mps_alloc_queues(struct mps_softc *sc)
565 {
566 	bus_addr_t queues_busaddr;
567 	uint8_t *queues;
568 	int qsize, fqsize, pqsize;
569 
570 	/*
571 	 * The reply free queue contains 4 byte entries in multiples of 16 and
572 	 * aligned on a 16 byte boundary. There must always be an unused entry.
573 	 * This queue supplies fresh reply frames for the firmware to use.
574 	 *
575 	 * The reply descriptor post queue contains 8 byte entries in
576 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
577 	 * contains filled-in reply frames sent from the firmware to the host.
578 	 *
579 	 * These two queues are allocated together for simplicity.
580 	 */
581 	sc->fqdepth = roundup2((sc->num_replies + 1), 16);
582 	sc->pqdepth = roundup2((sc->num_replies + 1), 16);
583 	fqsize= sc->fqdepth * 4;
584 	pqsize = sc->pqdepth * 8;
585 	qsize = fqsize + pqsize;
586 
587         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
588 				16, 0,			/* algnmnt, boundary */
589 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
590 				BUS_SPACE_MAXADDR,	/* highaddr */
591 				NULL, NULL,		/* filter, filterarg */
592                                 qsize,			/* maxsize */
593                                 1,			/* nsegments */
594                                 qsize,			/* maxsegsize */
595                                 0,			/* flags */
596                                 NULL, NULL,		/* lockfunc, lockarg */
597                                 &sc->queues_dmat)) {
598 		device_printf(sc->mps_dev, "Cannot allocate queues DMA tag\n");
599 		return (ENOMEM);
600         }
601         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
602 	    &sc->queues_map)) {
603 		device_printf(sc->mps_dev, "Cannot allocate queues memory\n");
604 		return (ENOMEM);
605         }
606         bzero(queues, qsize);
607         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
608 	    mps_memaddr_cb, &queues_busaddr, 0);
609 
610 	sc->free_queue = (uint32_t *)queues;
611 	sc->free_busaddr = queues_busaddr;
612 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
613 	sc->post_busaddr = queues_busaddr + fqsize;
614 
615 	return (0);
616 }
617 
618 static int
619 mps_alloc_replies(struct mps_softc *sc)
620 {
621 	int rsize, num_replies;
622 
623 	/*
624 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
625 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
626 	 * replies can be used at once.
627 	 */
628 	num_replies = max(sc->fqdepth, sc->num_replies);
629 
630 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
631         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
632 				4, 0,			/* algnmnt, boundary */
633 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
634 				BUS_SPACE_MAXADDR,	/* highaddr */
635 				NULL, NULL,		/* filter, filterarg */
636                                 rsize,			/* maxsize */
637                                 1,			/* nsegments */
638                                 rsize,			/* maxsegsize */
639                                 0,			/* flags */
640                                 NULL, NULL,		/* lockfunc, lockarg */
641                                 &sc->reply_dmat)) {
642 		device_printf(sc->mps_dev, "Cannot allocate replies DMA tag\n");
643 		return (ENOMEM);
644         }
645         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
646 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
647 		device_printf(sc->mps_dev, "Cannot allocate replies memory\n");
648 		return (ENOMEM);
649         }
650         bzero(sc->reply_frames, rsize);
651         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
652 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
653 
654 	return (0);
655 }
656 
657 static int
658 mps_alloc_requests(struct mps_softc *sc)
659 {
660 	struct mps_command *cm;
661 	struct mps_chain *chain;
662 	int i, rsize, nsegs;
663 
664 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
665         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
666 				16, 0,			/* algnmnt, boundary */
667 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
668 				BUS_SPACE_MAXADDR,	/* highaddr */
669 				NULL, NULL,		/* filter, filterarg */
670                                 rsize,			/* maxsize */
671                                 1,			/* nsegments */
672                                 rsize,			/* maxsegsize */
673                                 0,			/* flags */
674                                 NULL, NULL,		/* lockfunc, lockarg */
675                                 &sc->req_dmat)) {
676 		device_printf(sc->mps_dev, "Cannot allocate request DMA tag\n");
677 		return (ENOMEM);
678         }
679         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
680 	    BUS_DMA_NOWAIT, &sc->req_map)) {
681 		device_printf(sc->mps_dev, "Cannot allocate request memory\n");
682 		return (ENOMEM);
683         }
684         bzero(sc->req_frames, rsize);
685         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
686 	    mps_memaddr_cb, &sc->req_busaddr, 0);
687 
688 	rsize = sc->facts->IOCRequestFrameSize * MPS_CHAIN_FRAMES * 4;
689         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
690 				16, 0,			/* algnmnt, boundary */
691 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
692 				BUS_SPACE_MAXADDR,	/* highaddr */
693 				NULL, NULL,		/* filter, filterarg */
694                                 rsize,			/* maxsize */
695                                 1,			/* nsegments */
696                                 rsize,			/* maxsegsize */
697                                 0,			/* flags */
698                                 NULL, NULL,		/* lockfunc, lockarg */
699                                 &sc->chain_dmat)) {
700 		device_printf(sc->mps_dev, "Cannot allocate chain DMA tag\n");
701 		return (ENOMEM);
702         }
703         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
704 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
705 		device_printf(sc->mps_dev, "Cannot allocate chain memory\n");
706 		return (ENOMEM);
707         }
708         bzero(sc->chain_frames, rsize);
709         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
710 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
711 
712 	rsize = MPS_SENSE_LEN * sc->num_reqs;
713         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
714 				1, 0,			/* algnmnt, boundary */
715 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
716 				BUS_SPACE_MAXADDR,	/* highaddr */
717 				NULL, NULL,		/* filter, filterarg */
718                                 rsize,			/* maxsize */
719                                 1,			/* nsegments */
720                                 rsize,			/* maxsegsize */
721                                 0,			/* flags */
722                                 NULL, NULL,		/* lockfunc, lockarg */
723                                 &sc->sense_dmat)) {
724 		device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
725 		return (ENOMEM);
726         }
727         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
728 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
729 		device_printf(sc->mps_dev, "Cannot allocate sense memory\n");
730 		return (ENOMEM);
731         }
732         bzero(sc->sense_frames, rsize);
733         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
734 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
735 
736 	sc->chains = malloc(sizeof(struct mps_chain) * MPS_CHAIN_FRAMES,
737 	    M_MPT2, M_WAITOK | M_ZERO);
738 	for (i = 0; i < MPS_CHAIN_FRAMES; i++) {
739 		chain = &sc->chains[i];
740 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
741 		    i * sc->facts->IOCRequestFrameSize * 4);
742 		chain->chain_busaddr = sc->chain_busaddr +
743 		    i * sc->facts->IOCRequestFrameSize * 4;
744 		mps_free_chain(sc, chain);
745 		sc->chain_free_lowwater++;
746 	}
747 
748 	/* XXX Need to pick a more precise value */
749 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
750         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
751 				1, 0,			/* algnmnt, boundary */
752 				BUS_SPACE_MAXADDR,	/* lowaddr */
753 				BUS_SPACE_MAXADDR,	/* highaddr */
754 				NULL, NULL,		/* filter, filterarg */
755                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
756                                 nsegs,			/* nsegments */
757                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
758                                 BUS_DMA_ALLOCNOW,	/* flags */
759                                 busdma_lock_mutex,	/* lockfunc */
760 				&sc->mps_mtx,		/* lockarg */
761                                 &sc->buffer_dmat)) {
762 		device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
763 		return (ENOMEM);
764         }
765 
766 	/*
767 	 * SMID 0 cannot be used as a free command per the firmware spec.
768 	 * Just drop that command instead of risking accounting bugs.
769 	 */
770 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
771 	    M_MPT2, M_WAITOK | M_ZERO);
772 	for (i = 1; i < sc->num_reqs; i++) {
773 		cm = &sc->commands[i];
774 		cm->cm_req = sc->req_frames +
775 		    i * sc->facts->IOCRequestFrameSize * 4;
776 		cm->cm_req_busaddr = sc->req_busaddr +
777 		    i * sc->facts->IOCRequestFrameSize * 4;
778 		cm->cm_sense = &sc->sense_frames[i];
779 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
780 		cm->cm_desc.Default.SMID = i;
781 		cm->cm_sc = sc;
782 		TAILQ_INIT(&cm->cm_chain_list);
783 		callout_init(&cm->cm_callout, 1 /*MPSAFE*/);
784 
785 		/* XXX Is a failure here a critical problem? */
786 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
787 			mps_free_command(sc, cm);
788 		else {
789 			sc->num_reqs = i;
790 			break;
791 		}
792 	}
793 
794 	return (0);
795 }
796 
797 static int
798 mps_init_queues(struct mps_softc *sc)
799 {
800 	int i;
801 
802 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
803 
804 	/*
805 	 * According to the spec, we need to use one less reply than we
806 	 * have space for on the queue.  So sc->num_replies (the number we
807 	 * use) should be less than sc->fqdepth (allocated size).
808 	 */
809 	if (sc->num_replies >= sc->fqdepth)
810 		return (EINVAL);
811 
812 	/*
813 	 * Initialize all of the free queue entries.
814 	 */
815 	for (i = 0; i < sc->fqdepth; i++)
816 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
817 	sc->replyfreeindex = sc->num_replies;
818 
819 	return (0);
820 }
821 
822 int
823 mps_attach(struct mps_softc *sc)
824 {
825 	int i, error;
826 	char tmpstr[80], tmpstr2[80];
827 
828 	/*
829 	 * Grab any tunable-set debug level so that tracing works as early
830 	 * as possible.
831 	 */
832 	snprintf(tmpstr, sizeof(tmpstr), "hw.mps.%d.debug_level",
833 	    device_get_unit(sc->mps_dev));
834 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
835 	snprintf(tmpstr, sizeof(tmpstr), "hw.mps.%d.allow_multiple_tm_cmds",
836 	    device_get_unit(sc->mps_dev));
837 	TUNABLE_INT_FETCH(tmpstr, &sc->allow_multiple_tm_cmds);
838 
839 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
840 
841 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
842 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
843 	TAILQ_INIT(&sc->event_list);
844 
845 	/*
846 	 * Setup the sysctl variable so the user can change the debug level
847 	 * on the fly.
848 	 */
849 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
850 	    device_get_unit(sc->mps_dev));
851 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
852 
853 	sysctl_ctx_init(&sc->sysctl_ctx);
854 	sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
855 	    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2, CTLFLAG_RD,
856 	    0, tmpstr);
857 	if (sc->sysctl_tree == NULL)
858 		return (ENOMEM);
859 
860 	SYSCTL_ADD_UINT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
861 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
862 	    "mps debug level");
863 
864 	SYSCTL_ADD_UINT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
865 	    OID_AUTO, "allow_multiple_tm_cmds", CTLFLAG_RW,
866 	    &sc->allow_multiple_tm_cmds, 0,
867 	    "allow multiple simultaneous task management cmds");
868 
869 	SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
870 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
871 	    &sc->io_cmds_active, 0, "number of currently active commands");
872 
873 	SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
874 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
875 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
876 
877 	SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
878 	    OID_AUTO, "chain_free", CTLFLAG_RD,
879 	    &sc->chain_free, 0, "number of free chain elements");
880 
881 	SYSCTL_ADD_INT(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
882 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
883 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
884 
885 	SYSCTL_ADD_UQUAD(&sc->sysctl_ctx, SYSCTL_CHILDREN(sc->sysctl_tree),
886 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
887 	    &sc->chain_alloc_fail, "chain allocation failures");
888 
889 	if ((error = mps_transition_ready(sc)) != 0)
890 		return (error);
891 
892 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
893 	    M_ZERO|M_NOWAIT);
894 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0)
895 		return (error);
896 
897 	mps_print_iocfacts(sc, sc->facts);
898 
899 	mps_printf(sc, "Firmware: %02d.%02d.%02d.%02d\n",
900 	    sc->facts->FWVersion.Struct.Major,
901 	    sc->facts->FWVersion.Struct.Minor,
902 	    sc->facts->FWVersion.Struct.Unit,
903 	    sc->facts->FWVersion.Struct.Dev);
904 	mps_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
905 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
906 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
907 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
908 
909 	/*
910 	 * If the chip doesn't support event replay then a hard reset will be
911 	 * required to trigger a full discovery.  Do the reset here then
912 	 * retransition to Ready.  A hard reset might have already been done,
913 	 * but it doesn't hurt to do it again.
914 	 */
915 	if ((sc->facts->IOCCapabilities &
916 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
917 		mps_hard_reset(sc);
918 		if ((error = mps_transition_ready(sc)) != 0)
919 			return (error);
920 	}
921 
922 	/*
923 	 * Size the queues. Since the reply queues always need one free entry,
924 	 * we'll just deduct one reply message here.
925 	 */
926 	sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
927 	sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
928 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
929 	mps_dprint(sc, MPS_INFO, "num_reqs %d, num_replies %d\n", sc->num_reqs,
930 		   sc->num_replies);
931 	TAILQ_INIT(&sc->req_list);
932 	TAILQ_INIT(&sc->chain_list);
933 	TAILQ_INIT(&sc->tm_list);
934 	TAILQ_INIT(&sc->io_list);
935 
936 	if (((error = mps_alloc_queues(sc)) != 0) ||
937 	    ((error = mps_alloc_replies(sc)) != 0) ||
938 	    ((error = mps_alloc_requests(sc)) != 0)) {
939 		mps_free(sc);
940 		return (error);
941 	}
942 
943 	if (((error = mps_init_queues(sc)) != 0) ||
944 	    ((error = mps_transition_operational(sc)) != 0)) {
945 		mps_free(sc);
946 		return (error);
947 	}
948 
949 	/*
950 	 * Finish the queue initialization.
951 	 * These are set here instead of in mps_init_queues() because the
952 	 * IOC resets these values during the state transition in
953 	 * mps_transition_operational().  The free index is set to 1
954 	 * because the corresponding index in the IOC is set to 0, and the
955 	 * IOC treats the queues as full if both are set to the same value.
956 	 * Hence the reason that the queue can't hold all of the possible
957 	 * replies.
958 	 */
959 	sc->replypostindex = 0;
960 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
961 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
962 
963 	sc->pfacts = malloc(sizeof(MPI2_PORT_FACTS_REPLY) *
964 	    sc->facts->NumberOfPorts, M_MPT2, M_ZERO|M_WAITOK);
965 	for (i = 0; i < sc->facts->NumberOfPorts; i++) {
966 		if ((error = mps_get_portfacts(sc, &sc->pfacts[i], i)) != 0) {
967 			mps_free(sc);
968 			return (error);
969 		}
970 		mps_print_portfacts(sc, &sc->pfacts[i]);
971 	}
972 
973 	/* Attach the subsystems so they can prepare their event masks. */
974 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
975 	if (((error = mps_attach_log(sc)) != 0) ||
976 	    ((error = mps_attach_sas(sc)) != 0) ||
977 	    ((error = mps_attach_user(sc)) != 0)) {
978 		mps_printf(sc, "%s failed to attach all subsystems: error %d\n",
979 		    __func__, error);
980 		mps_free(sc);
981 		return (error);
982 	}
983 
984 	if ((error = mps_pci_setup_interrupts(sc)) != 0) {
985 		mps_free(sc);
986 		return (error);
987 	}
988 
989 	/* Start the periodic watchdog check on the IOC Doorbell */
990 	mps_periodic(sc);
991 
992 	/*
993 	 * The portenable will kick off discovery events that will drive the
994 	 * rest of the initialization process.  The CAM/SAS module will
995 	 * hold up the boot sequence until discovery is complete.
996 	 */
997 	sc->mps_ich.ich_func = mps_startup;
998 	sc->mps_ich.ich_arg = sc;
999 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1000 		mps_dprint(sc, MPS_FAULT, "Cannot establish MPS config hook\n");
1001 		error = EINVAL;
1002 	}
1003 
1004 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1005 
1006 	return (error);
1007 }
1008 
1009 static void
1010 mps_startup(void *arg)
1011 {
1012 	struct mps_softc *sc;
1013 
1014 	sc = (struct mps_softc *)arg;
1015 
1016 	mps_lock(sc);
1017 	mps_unmask_intr(sc);
1018 	mps_send_portenable(sc);
1019 	mps_unlock(sc);
1020 }
1021 
1022 /* Periodic watchdog.  Is called with the driver lock already held. */
1023 static void
1024 mps_periodic(void *arg)
1025 {
1026 	struct mps_softc *sc;
1027 	uint32_t db;
1028 
1029 	sc = (struct mps_softc *)arg;
1030 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1031 		return;
1032 
1033 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1034 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1035 		device_printf(sc->mps_dev, "IOC Fault 0x%08x, Resetting\n", db);
1036 		/* XXX Need to broaden this to re-initialize the chip */
1037 		mps_hard_reset(sc);
1038 		db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1039 		if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1040 			device_printf(sc->mps_dev, "Second IOC Fault 0x%08x, "
1041 			    "Giving up!\n", db);
1042 			return;
1043 		}
1044 	}
1045 
1046 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1047 }
1048 
1049 static void
1050 mps_startup_complete(struct mps_softc *sc, struct mps_command *cm)
1051 {
1052 	MPI2_PORT_ENABLE_REPLY *reply;
1053 
1054 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1055 
1056 	reply = (MPI2_PORT_ENABLE_REPLY *)cm->cm_reply;
1057 	if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1058 		mps_dprint(sc, MPS_FAULT, "Portenable failed\n");
1059 
1060 	mps_free_command(sc, cm);
1061 	config_intrhook_disestablish(&sc->mps_ich);
1062 
1063 }
1064 
1065 static void
1066 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1067     MPI2_EVENT_NOTIFICATION_REPLY *event)
1068 {
1069 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1070 
1071 	mps_print_event(sc, event);
1072 
1073 	switch (event->Event) {
1074 	case MPI2_EVENT_LOG_DATA:
1075 		device_printf(sc->mps_dev, "MPI2_EVENT_LOG_DATA:\n");
1076 		hexdump(event->EventData, event->EventDataLength, NULL, 0);
1077 		break;
1078 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1079 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1080 		mps_dprint(sc, MPS_INFO, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1081 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1082 		     entry->LogSequence);
1083 		break;
1084 	default:
1085 		break;
1086 	}
1087 	return;
1088 }
1089 
1090 static int
1091 mps_attach_log(struct mps_softc *sc)
1092 {
1093 	uint8_t events[16];
1094 
1095 	bzero(events, 16);
1096 	setbit(events, MPI2_EVENT_LOG_DATA);
1097 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1098 
1099 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1100 	    &sc->mps_log_eh);
1101 
1102 	return (0);
1103 }
1104 
1105 static int
1106 mps_detach_log(struct mps_softc *sc)
1107 {
1108 
1109 	if (sc->mps_log_eh != NULL)
1110 		mps_deregister_events(sc, sc->mps_log_eh);
1111 	return (0);
1112 }
1113 
1114 /*
1115  * Free all of the driver resources and detach submodules.  Should be called
1116  * without the lock held.
1117  */
1118 int
1119 mps_free(struct mps_softc *sc)
1120 {
1121 	struct mps_command *cm;
1122 	int i, error;
1123 
1124 	/* Turn off the watchdog */
1125 	mps_lock(sc);
1126 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1127 	mps_unlock(sc);
1128 	/* Lock must not be held for this */
1129 	callout_drain(&sc->periodic);
1130 
1131 	if (((error = mps_detach_log(sc)) != 0) ||
1132 	    ((error = mps_detach_sas(sc)) != 0))
1133 		return (error);
1134 
1135 	/* Put the IOC back in the READY state. */
1136 	mps_lock(sc);
1137 	if ((error = mps_send_mur(sc)) != 0) {
1138 		mps_unlock(sc);
1139 		return (error);
1140 	}
1141 	mps_unlock(sc);
1142 
1143 	if (sc->facts != NULL)
1144 		free(sc->facts, M_MPT2);
1145 
1146 	if (sc->pfacts != NULL)
1147 		free(sc->pfacts, M_MPT2);
1148 
1149 	if (sc->post_busaddr != 0)
1150 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
1151 	if (sc->post_queue != NULL)
1152 		bus_dmamem_free(sc->queues_dmat, sc->post_queue,
1153 		    sc->queues_map);
1154 	if (sc->queues_dmat != NULL)
1155 		bus_dma_tag_destroy(sc->queues_dmat);
1156 
1157 	if (sc->chain_busaddr != 0)
1158 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
1159 	if (sc->chain_frames != NULL)
1160 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,sc->chain_map);
1161 	if (sc->chain_dmat != NULL)
1162 		bus_dma_tag_destroy(sc->chain_dmat);
1163 
1164 	if (sc->sense_busaddr != 0)
1165 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
1166 	if (sc->sense_frames != NULL)
1167 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,sc->sense_map);
1168 	if (sc->sense_dmat != NULL)
1169 		bus_dma_tag_destroy(sc->sense_dmat);
1170 
1171 	if (sc->reply_busaddr != 0)
1172 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
1173 	if (sc->reply_frames != NULL)
1174 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,sc->reply_map);
1175 	if (sc->reply_dmat != NULL)
1176 		bus_dma_tag_destroy(sc->reply_dmat);
1177 
1178 	if (sc->req_busaddr != 0)
1179 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
1180 	if (sc->req_frames != NULL)
1181 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
1182 	if (sc->req_dmat != NULL)
1183 		bus_dma_tag_destroy(sc->req_dmat);
1184 
1185 	if (sc->chains != NULL)
1186 		free(sc->chains, M_MPT2);
1187 	if (sc->commands != NULL) {
1188 		for (i = 1; i < sc->num_reqs; i++) {
1189 			cm = &sc->commands[i];
1190 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
1191 		}
1192 		free(sc->commands, M_MPT2);
1193 	}
1194 	if (sc->buffer_dmat != NULL)
1195 		bus_dma_tag_destroy(sc->buffer_dmat);
1196 
1197 	if (sc->sysctl_tree != NULL)
1198 		sysctl_ctx_free(&sc->sysctl_ctx);
1199 
1200 	mtx_destroy(&sc->mps_mtx);
1201 
1202 	return (0);
1203 }
1204 
1205 static __inline void
1206 mps_complete_command(struct mps_command *cm)
1207 {
1208 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1209 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1210 
1211 	if (cm->cm_complete != NULL)
1212 		cm->cm_complete(cm->cm_sc, cm);
1213 
1214 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1215 		mps_dprint(cm->cm_sc, MPS_TRACE, "%s: waking up %p\n",
1216 			   __func__, cm);
1217 		wakeup(cm);
1218 	}
1219 }
1220 
1221 void
1222 mps_intr(void *data)
1223 {
1224 	struct mps_softc *sc;
1225 	uint32_t status;
1226 
1227 	sc = (struct mps_softc *)data;
1228 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1229 
1230 	/*
1231 	 * Check interrupt status register to flush the bus.  This is
1232 	 * needed for both INTx interrupts and driver-driven polling
1233 	 */
1234 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1235 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1236 		return;
1237 
1238 	mps_lock(sc);
1239 	mps_intr_locked(data);
1240 	mps_unlock(sc);
1241 	return;
1242 }
1243 
1244 /*
1245  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1246  * chip.  Hopefully this theory is correct.
1247  */
1248 void
1249 mps_intr_msi(void *data)
1250 {
1251 	struct mps_softc *sc;
1252 
1253 	sc = (struct mps_softc *)data;
1254 	mps_lock(sc);
1255 	mps_intr_locked(data);
1256 	mps_unlock(sc);
1257 	return;
1258 }
1259 
1260 /*
1261  * The locking is overly broad and simplistic, but easy to deal with for now.
1262  */
1263 void
1264 mps_intr_locked(void *data)
1265 {
1266 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1267 	struct mps_softc *sc;
1268 	struct mps_command *cm = NULL;
1269 	uint8_t flags;
1270 	u_int pq;
1271 
1272 	sc = (struct mps_softc *)data;
1273 
1274 	pq = sc->replypostindex;
1275 
1276 	for ( ;; ) {
1277 		cm = NULL;
1278 		desc = &sc->post_queue[pq];
1279 		flags = desc->Default.ReplyFlags &
1280 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1281 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
1282 		 || (desc->Words.High == 0xffffffff))
1283 			break;
1284 
1285 		switch (flags) {
1286 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1287 			cm = &sc->commands[desc->SCSIIOSuccess.SMID];
1288 			cm->cm_reply = NULL;
1289 			break;
1290 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1291 		{
1292 			uint32_t baddr;
1293 			uint8_t *reply;
1294 
1295 			/*
1296 			 * Re-compose the reply address from the address
1297 			 * sent back from the chip.  The ReplyFrameAddress
1298 			 * is the lower 32 bits of the physical address of
1299 			 * particular reply frame.  Convert that address to
1300 			 * host format, and then use that to provide the
1301 			 * offset against the virtual address base
1302 			 * (sc->reply_frames).
1303 			 */
1304 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1305 			reply = sc->reply_frames +
1306 				(baddr - ((uint32_t)sc->reply_busaddr));
1307 			/*
1308 			 * Make sure the reply we got back is in a valid
1309 			 * range.  If not, go ahead and panic here, since
1310 			 * we'll probably panic as soon as we deference the
1311 			 * reply pointer anyway.
1312 			 */
1313 			if ((reply < sc->reply_frames)
1314 			 || (reply > (sc->reply_frames +
1315 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1316 				printf("%s: WARNING: reply %p out of range!\n",
1317 				       __func__, reply);
1318 				printf("%s: reply_frames %p, fqdepth %d, "
1319 				       "frame size %d\n", __func__,
1320 				       sc->reply_frames, sc->fqdepth,
1321 				       sc->facts->ReplyFrameSize * 4);
1322 				printf("%s: baddr %#x,\n", __func__, baddr);
1323 				panic("Reply address out of range");
1324 			}
1325 			if (desc->AddressReply.SMID == 0) {
1326 				mps_dispatch_event(sc, baddr,
1327 				   (MPI2_EVENT_NOTIFICATION_REPLY *) reply);
1328 			} else {
1329 				cm = &sc->commands[desc->AddressReply.SMID];
1330 				cm->cm_reply = reply;
1331 				cm->cm_reply_data =
1332 				    desc->AddressReply.ReplyFrameAddress;
1333 			}
1334 			break;
1335 		}
1336 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
1337 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
1338 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
1339 		default:
1340 			/* Unhandled */
1341 			device_printf(sc->mps_dev, "Unhandled reply 0x%x\n",
1342 			    desc->Default.ReplyFlags);
1343 			cm = NULL;
1344 			break;
1345 		}
1346 
1347 		if (cm != NULL)
1348 			mps_complete_command(cm);
1349 
1350 		desc->Words.Low = 0xffffffff;
1351 		desc->Words.High = 0xffffffff;
1352 		if (++pq >= sc->pqdepth)
1353 			pq = 0;
1354 	}
1355 
1356 	if (pq != sc->replypostindex) {
1357 		mps_dprint(sc, MPS_INFO, "writing postindex %d\n", pq);
1358 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, pq);
1359 		sc->replypostindex = pq;
1360 	}
1361 
1362 	return;
1363 }
1364 
1365 static void
1366 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
1367     MPI2_EVENT_NOTIFICATION_REPLY *reply)
1368 {
1369 	struct mps_event_handle *eh;
1370 	int event, handled = 0;
1371 
1372 	event = reply->Event;
1373 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1374 		if (isset(eh->mask, event)) {
1375 			eh->callback(sc, data, reply);
1376 			handled++;
1377 		}
1378 	}
1379 
1380 	if (handled == 0)
1381 		device_printf(sc->mps_dev, "Unhandled event 0x%x\n", event);
1382 }
1383 
1384 /*
1385  * For both register_events and update_events, the caller supplies a bitmap
1386  * of events that it _wants_.  These functions then turn that into a bitmask
1387  * suitable for the controller.
1388  */
1389 int
1390 mps_register_events(struct mps_softc *sc, uint8_t *mask,
1391     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
1392 {
1393 	struct mps_event_handle *eh;
1394 	int error = 0;
1395 
1396 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
1397 	eh->callback = cb;
1398 	eh->data = data;
1399 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
1400 	if (mask != NULL)
1401 		error = mps_update_events(sc, eh, mask);
1402 	*handle = eh;
1403 
1404 	return (error);
1405 }
1406 
1407 int
1408 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
1409     uint8_t *mask)
1410 {
1411 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
1412 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
1413 	struct mps_command *cm;
1414 	struct mps_event_handle *eh;
1415 	int error, i;
1416 
1417 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1418 
1419 	if ((mask != NULL) && (handle != NULL))
1420 		bcopy(mask, &handle->mask[0], 16);
1421 	memset(sc->event_mask, 0xff, 16);
1422 
1423 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1424 		for (i = 0; i < 16; i++)
1425 			sc->event_mask[i] &= ~eh->mask[i];
1426 	}
1427 
1428 	if ((cm = mps_alloc_command(sc)) == NULL)
1429 		return (EBUSY);
1430 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
1431 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
1432 	evtreq->MsgFlags = 0;
1433 	evtreq->SASBroadcastPrimitiveMasks = 0;
1434 #ifdef MPS_DEBUG_ALL_EVENTS
1435 	{
1436 		u_char fullmask[16];
1437 		memset(fullmask, 0x00, 16);
1438 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
1439 	}
1440 #else
1441 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
1442 #endif
1443 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
1444 	cm->cm_data = NULL;
1445 
1446 	error = mps_request_polled(sc, cm);
1447 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
1448 	if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1449 		error = ENXIO;
1450 	mps_print_event(sc, reply);
1451 
1452 	mps_free_command(sc, cm);
1453 	return (error);
1454 }
1455 
1456 int
1457 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
1458 {
1459 
1460 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
1461 	free(handle, M_MPT2);
1462 	return (mps_update_events(sc, NULL, NULL));
1463 }
1464 
1465 /*
1466  * Add a chain element as the next SGE for the specified command.
1467  * Reset cm_sge and cm_sgesize to indicate all the available space.
1468  */
1469 static int
1470 mps_add_chain(struct mps_command *cm)
1471 {
1472 	MPI2_SGE_CHAIN32 *sgc;
1473 	struct mps_chain *chain;
1474 	int space;
1475 
1476 	if (cm->cm_sglsize < MPS_SGC_SIZE)
1477 		panic("MPS: Need SGE Error Code\n");
1478 
1479 	chain = mps_alloc_chain(cm->cm_sc);
1480 	if (chain == NULL)
1481 		return (ENOBUFS);
1482 
1483 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
1484 
1485 	/*
1486 	 * Note: a double-linked list is used to make it easier to
1487 	 * walk for debugging.
1488 	 */
1489 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
1490 
1491 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
1492 	sgc->Length = space;
1493 	sgc->NextChainOffset = 0;
1494 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
1495 	sgc->Address = chain->chain_busaddr;
1496 
1497 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
1498 	cm->cm_sglsize = space;
1499 	return (0);
1500 }
1501 
1502 /*
1503  * Add one scatter-gather element (chain, simple, transaction context)
1504  * to the scatter-gather list for a command.  Maintain cm_sglsize and
1505  * cm_sge as the remaining size and pointer to the next SGE to fill
1506  * in, respectively.
1507  */
1508 int
1509 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
1510 {
1511 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
1512 	MPI2_SGE_SIMPLE64 *sge = sgep;
1513 	int error, type;
1514 
1515 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
1516 
1517 #ifdef INVARIANTS
1518 	switch (type) {
1519 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
1520 		if (len != tc->DetailsLength + 4)
1521 			panic("TC %p length %u or %zu?", tc,
1522 			    tc->DetailsLength + 4, len);
1523 		}
1524 		break;
1525 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
1526 		/* Driver only uses 32-bit chain elements */
1527 		if (len != MPS_SGC_SIZE)
1528 			panic("CHAIN %p length %u or %zu?", sgep,
1529 			    MPS_SGC_SIZE, len);
1530 		break;
1531 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
1532 		/* Driver only uses 64-bit SGE simple elements */
1533 		sge = sgep;
1534 		if (len != MPS_SGE64_SIZE)
1535 			panic("SGE simple %p length %u or %zu?", sge,
1536 			    MPS_SGE64_SIZE, len);
1537 		if (((sge->FlagsLength >> MPI2_SGE_FLAGS_SHIFT) &
1538 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
1539 			panic("SGE simple %p flags %02x not marked 64-bit?",
1540 			    sge, sge->FlagsLength >> MPI2_SGE_FLAGS_SHIFT);
1541 
1542 		break;
1543 	default:
1544 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
1545 	}
1546 #endif
1547 
1548 	/*
1549 	 * case 1: 1 more segment, enough room for it
1550 	 * case 2: 2 more segments, enough room for both
1551 	 * case 3: >=2 more segments, only enough room for 1 and a chain
1552 	 * case 4: >=1 more segment, enough room for only a chain
1553 	 * case 5: >=1 more segment, no room for anything (error)
1554          */
1555 
1556 	/*
1557 	 * There should be room for at least a chain element, or this
1558 	 * code is buggy.  Case (5).
1559 	 */
1560 	if (cm->cm_sglsize < MPS_SGC_SIZE)
1561 		panic("MPS: Need SGE Error Code\n");
1562 
1563 	if (segsleft >= 2 &&
1564 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
1565 		/*
1566 		 * There are 2 or more segments left to add, and only
1567 		 * enough room for 1 and a chain.  Case (3).
1568 		 *
1569 		 * Mark as last element in this chain if necessary.
1570 		 */
1571 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
1572 			sge->FlagsLength |=
1573 				(MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
1574 		}
1575 
1576 		/*
1577 		 * Add the item then a chain.  Do the chain now,
1578 		 * rather than on the next iteration, to simplify
1579 		 * understanding the code.
1580 		 */
1581 		cm->cm_sglsize -= len;
1582 		bcopy(sgep, cm->cm_sge, len);
1583 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
1584 		return (mps_add_chain(cm));
1585 	}
1586 
1587 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
1588 		/*
1589 		 * 1 or more segment, enough room for only a chain.
1590 		 * Hope the previous element wasn't a Simple entry
1591 		 * that needed to be marked with
1592 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
1593 		 */
1594 		if ((error = mps_add_chain(cm)) != 0)
1595 			return (error);
1596 	}
1597 
1598 #ifdef INVARIANTS
1599 	/* Case 1: 1 more segment, enough room for it. */
1600 	if (segsleft == 1 && cm->cm_sglsize < len)
1601 		panic("1 seg left and no room? %u versus %zu",
1602 		    cm->cm_sglsize, len);
1603 
1604 	/* Case 2: 2 more segments, enough room for both */
1605 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
1606 		panic("2 segs left and no room? %u versus %zu",
1607 		    cm->cm_sglsize, len);
1608 #endif
1609 
1610 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
1611 		/*
1612 		 * Last element of the last segment of the entire
1613 		 * buffer.
1614 		 */
1615 		sge->FlagsLength |= ((MPI2_SGE_FLAGS_LAST_ELEMENT |
1616 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
1617 		    MPI2_SGE_FLAGS_END_OF_LIST) << MPI2_SGE_FLAGS_SHIFT);
1618 	}
1619 
1620 	cm->cm_sglsize -= len;
1621 	bcopy(sgep, cm->cm_sge, len);
1622 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
1623 	return (0);
1624 }
1625 
1626 /*
1627  * Add one dma segment to the scatter-gather list for a command.
1628  */
1629 int
1630 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
1631     int segsleft)
1632 {
1633 	MPI2_SGE_SIMPLE64 sge;
1634 
1635 	/*
1636 	 * This driver always uses 64-bit address elements for
1637 	 * simplicity.
1638 	 */
1639 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT | MPI2_SGE_FLAGS_ADDRESS_SIZE;
1640 	sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
1641 	mps_from_u64(pa, &sge.Address);
1642 
1643 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
1644 }
1645 
1646 static void
1647 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1648 {
1649 	struct mps_softc *sc;
1650 	struct mps_command *cm;
1651 	u_int i, dir, sflags;
1652 
1653 	cm = (struct mps_command *)arg;
1654 	sc = cm->cm_sc;
1655 
1656 	/*
1657 	 * In this case, just print out a warning and let the chip tell the
1658 	 * user they did the wrong thing.
1659 	 */
1660 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
1661 		mps_printf(sc, "%s: warning: busdma returned %d segments, "
1662 			   "more than the %d allowed\n", __func__, nsegs,
1663 			   cm->cm_max_segs);
1664 	}
1665 
1666 	/*
1667 	 * Set up DMA direction flags.  Note that we don't support
1668 	 * bi-directional transfers, with the exception of SMP passthrough.
1669 	 */
1670 	sflags = 0;
1671 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
1672 		/*
1673 		 * We have to add a special case for SMP passthrough, there
1674 		 * is no easy way to generically handle it.  The first
1675 		 * S/G element is used for the command (therefore the
1676 		 * direction bit needs to be set).  The second one is used
1677 		 * for the reply.  We'll leave it to the caller to make
1678 		 * sure we only have two buffers.
1679 		 */
1680 		/*
1681 		 * Even though the busdma man page says it doesn't make
1682 		 * sense to have both direction flags, it does in this case.
1683 		 * We have one s/g element being accessed in each direction.
1684 		 */
1685 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
1686 
1687 		/*
1688 		 * Set the direction flag on the first buffer in the SMP
1689 		 * passthrough request.  We'll clear it for the second one.
1690 		 */
1691 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
1692 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
1693 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
1694 		sflags |= MPI2_SGE_FLAGS_DIRECTION;
1695 		dir = BUS_DMASYNC_PREWRITE;
1696 	} else
1697 		dir = BUS_DMASYNC_PREREAD;
1698 
1699 	for (i = 0; i < nsegs; i++) {
1700 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS)
1701 		 && (i != 0)) {
1702 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
1703 		}
1704 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
1705 		    sflags, nsegs - i);
1706 		if (error != 0) {
1707 			/* Resource shortage, roll back! */
1708 			mps_dprint(sc, MPS_INFO, "out of chain frames\n");
1709 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
1710 			mps_complete_command(cm);
1711 			return;
1712 		}
1713 	}
1714 
1715 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
1716 	mps_enqueue_request(sc, cm);
1717 
1718 	return;
1719 }
1720 
1721 static void
1722 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
1723 	     int error)
1724 {
1725 	mps_data_cb(arg, segs, nsegs, error);
1726 }
1727 
1728 /*
1729  * Note that the only error path here is from bus_dmamap_load(), which can
1730  * return EINPROGRESS if it is waiting for resources.
1731  */
1732 int
1733 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
1734 {
1735 	MPI2_SGE_SIMPLE32 *sge;
1736 	int error = 0;
1737 
1738 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
1739 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
1740 		    &cm->cm_uio, mps_data_cb2, cm, 0);
1741 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
1742 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
1743 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
1744 	} else {
1745 		/* Add a zero-length element as needed */
1746 		if (cm->cm_sge != NULL) {
1747 			sge = (MPI2_SGE_SIMPLE32 *)cm->cm_sge;
1748 			sge->FlagsLength = (MPI2_SGE_FLAGS_LAST_ELEMENT |
1749 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
1750 			    MPI2_SGE_FLAGS_END_OF_LIST |
1751 			    MPI2_SGE_FLAGS_SIMPLE_ELEMENT) <<
1752 			    MPI2_SGE_FLAGS_SHIFT;
1753 			sge->Address = 0;
1754 		}
1755 		mps_enqueue_request(sc, cm);
1756 	}
1757 
1758 	return (error);
1759 }
1760 
1761 /*
1762  * The MPT driver had a verbose interface for config pages.  In this driver,
1763  * reduce it to much simplier terms, similar to the Linux driver.
1764  */
1765 int
1766 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
1767 {
1768 	MPI2_CONFIG_REQUEST *req;
1769 	struct mps_command *cm;
1770 	int error;
1771 
1772 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
1773 		return (EBUSY);
1774 	}
1775 
1776 	cm = mps_alloc_command(sc);
1777 	if (cm == NULL) {
1778 		return (EBUSY);
1779 	}
1780 
1781 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
1782 	req->Function = MPI2_FUNCTION_CONFIG;
1783 	req->Action = params->action;
1784 	req->SGLFlags = 0;
1785 	req->ChainOffset = 0;
1786 	req->PageAddress = params->page_address;
1787 	if (params->hdr.Ext.ExtPageType != 0) {
1788 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
1789 
1790 		hdr = &params->hdr.Ext;
1791 		req->ExtPageType = hdr->ExtPageType;
1792 		req->ExtPageLength = hdr->ExtPageLength;
1793 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
1794 		req->Header.PageLength = 0; /* Must be set to zero */
1795 		req->Header.PageNumber = hdr->PageNumber;
1796 		req->Header.PageVersion = hdr->PageVersion;
1797 	} else {
1798 		MPI2_CONFIG_PAGE_HEADER *hdr;
1799 
1800 		hdr = &params->hdr.Struct;
1801 		req->Header.PageType = hdr->PageType;
1802 		req->Header.PageNumber = hdr->PageNumber;
1803 		req->Header.PageLength = hdr->PageLength;
1804 		req->Header.PageVersion = hdr->PageVersion;
1805 	}
1806 
1807 	cm->cm_data = params->buffer;
1808 	cm->cm_length = params->length;
1809 	cm->cm_sge = &req->PageBufferSGE;
1810 	cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
1811 	cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
1812 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
1813 
1814 	cm->cm_complete_data = params;
1815 	if (params->callback != NULL) {
1816 		cm->cm_complete = mps_config_complete;
1817 		return (mps_map_command(sc, cm));
1818 	} else {
1819 		cm->cm_complete = NULL;
1820 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
1821 		if ((error = mps_map_command(sc, cm)) != 0)
1822 			return (error);
1823 		msleep(cm, &sc->mps_mtx, 0, "mpswait", 0);
1824 		mps_config_complete(sc, cm);
1825 	}
1826 
1827 	return (0);
1828 }
1829 
1830 int
1831 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
1832 {
1833 	return (EINVAL);
1834 }
1835 
1836 static void
1837 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
1838 {
1839 	MPI2_CONFIG_REPLY *reply;
1840 	struct mps_config_params *params;
1841 
1842 	params = cm->cm_complete_data;
1843 
1844 	if (cm->cm_data != NULL) {
1845 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
1846 		    BUS_DMASYNC_POSTREAD);
1847 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
1848 	}
1849 
1850 	/*
1851 	 * XXX KDM need to do more error recovery?  This results in the
1852 	 * device in question not getting probed.
1853 	 */
1854 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
1855 		params->status = MPI2_IOCSTATUS_BUSY;
1856 		goto bailout;
1857 	}
1858 
1859 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
1860 	params->status = reply->IOCStatus;
1861 	if (params->hdr.Ext.ExtPageType != 0) {
1862 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
1863 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
1864 	} else {
1865 		params->hdr.Struct.PageType = reply->Header.PageType;
1866 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
1867 		params->hdr.Struct.PageLength = reply->Header.PageLength;
1868 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
1869 	}
1870 
1871 bailout:
1872 
1873 	mps_free_command(sc, cm);
1874 	if (params->callback != NULL)
1875 		params->callback(sc, params);
1876 
1877 	return;
1878 }
1879