xref: /freebsd/sys/dev/mps/mps.c (revision 95d45410b5100e07f6f98450bcd841a8945d4726)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2012 LSI Corp.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * LSI MPT-Fusion Host Adapter FreeBSD
28  *
29  * $FreeBSD$
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 /* Communications core for LSI MPT2 */
36 
37 /* TODO Move headers to mpsvar */
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/selinfo.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/bus.h>
47 #include <sys/conf.h>
48 #include <sys/bio.h>
49 #include <sys/malloc.h>
50 #include <sys/uio.h>
51 #include <sys/sysctl.h>
52 #include <sys/queue.h>
53 #include <sys/kthread.h>
54 #include <sys/taskqueue.h>
55 #include <sys/endian.h>
56 #include <sys/eventhandler.h>
57 
58 #include <machine/bus.h>
59 #include <machine/resource.h>
60 #include <sys/rman.h>
61 #include <sys/proc.h>
62 
63 #include <dev/pci/pcivar.h>
64 
65 #include <cam/cam.h>
66 #include <cam/scsi/scsi_all.h>
67 
68 #include <dev/mps/mpi/mpi2_type.h>
69 #include <dev/mps/mpi/mpi2.h>
70 #include <dev/mps/mpi/mpi2_ioc.h>
71 #include <dev/mps/mpi/mpi2_sas.h>
72 #include <dev/mps/mpi/mpi2_cnfg.h>
73 #include <dev/mps/mpi/mpi2_init.h>
74 #include <dev/mps/mpi/mpi2_tool.h>
75 #include <dev/mps/mps_ioctl.h>
76 #include <dev/mps/mpsvar.h>
77 #include <dev/mps/mps_table.h>
78 
79 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
80 static int mps_init_queues(struct mps_softc *sc);
81 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
82 static int mps_transition_operational(struct mps_softc *sc);
83 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
84 static void mps_iocfacts_free(struct mps_softc *sc);
85 static void mps_startup(void *arg);
86 static int mps_send_iocinit(struct mps_softc *sc);
87 static int mps_alloc_queues(struct mps_softc *sc);
88 static int mps_alloc_replies(struct mps_softc *sc);
89 static int mps_alloc_requests(struct mps_softc *sc);
90 static int mps_attach_log(struct mps_softc *sc);
91 static __inline void mps_complete_command(struct mps_softc *sc,
92     struct mps_command *cm);
93 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
94     MPI2_EVENT_NOTIFICATION_REPLY *reply);
95 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
96 static void mps_periodic(void *);
97 static int mps_reregister_events(struct mps_softc *sc);
98 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
99 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
100 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
101 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
102 
103 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
104 
105 /*
106  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
107  * any state and back to its initialization state machine.
108  */
109 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
110 
111 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
112  * Compiler only support unint64_t to be passed as argument.
113  * Otherwise it will through below error
114  * "aggregate value used where an integer was expected"
115  */
116 
117 typedef union _reply_descriptor {
118         u64 word;
119         struct {
120                 u32 low;
121                 u32 high;
122         } u;
123 }reply_descriptor,address_descriptor;
124 
125 /* Rate limit chain-fail messages to 1 per minute */
126 static struct timeval mps_chainfail_interval = { 60, 0 };
127 
128 /*
129  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
130  * If this function is called from process context, it can sleep
131  * and there is no harm to sleep, in case if this fuction is called
132  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
133  * based on sleep flags driver will call either msleep, pause or DELAY.
134  * msleep and pause are of same variant, but pause is used when mps_mtx
135  * is not hold by driver.
136  *
137  */
138 static int
139 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
140 {
141 	uint32_t reg;
142 	int i, error, tries = 0;
143 
144 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
145 
146 	/* Clear any pending interrupts */
147 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
148 
149 	/*Force NO_SLEEP for threads prohibited to sleep
150  	* e.a Thread from interrupt handler are prohibited to sleep.
151  	*/
152 	if (curthread->td_no_sleeping != 0)
153 		sleep_flag = NO_SLEEP;
154 
155 	/* Push the magic sequence */
156 	error = ETIMEDOUT;
157 	while (tries++ < 20) {
158 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
159 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
160 			    mpt2_reset_magic[i]);
161 		/* wait 100 msec */
162 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
163 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
164 			    "mpsdiag", hz/10);
165 		else if (sleep_flag == CAN_SLEEP)
166 			pause("mpsdiag", hz/10);
167 		else
168 			DELAY(100 * 1000);
169 
170 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
171 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
172 			error = 0;
173 			break;
174 		}
175 	}
176 	if (error)
177 		return (error);
178 
179 	/* Send the actual reset.  XXX need to refresh the reg? */
180 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
181 	    reg | MPI2_DIAG_RESET_ADAPTER);
182 
183 	/* Wait up to 300 seconds in 50ms intervals */
184 	error = ETIMEDOUT;
185 	for (i = 0; i < 60000; i++) {
186 		/* wait 50 msec */
187 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
188 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
189 			    "mpsdiag", hz/20);
190 		else if (sleep_flag == CAN_SLEEP)
191 			pause("mpsdiag", hz/20);
192 		else
193 			DELAY(50 * 1000);
194 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
195 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
196 			error = 0;
197 			break;
198 		}
199 	}
200 	if (error)
201 		return (error);
202 
203 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
204 
205 	return (0);
206 }
207 
208 static int
209 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
210 {
211 
212 	MPS_FUNCTRACE(sc);
213 
214 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
215 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
216 	    MPI2_DOORBELL_FUNCTION_SHIFT);
217 
218 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
219 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed : <%s>\n",
220 				__func__);
221 		return (ETIMEDOUT);
222 	}
223 
224 	return (0);
225 }
226 
227 static int
228 mps_transition_ready(struct mps_softc *sc)
229 {
230 	uint32_t reg, state;
231 	int error, tries = 0;
232 	int sleep_flags;
233 
234 	MPS_FUNCTRACE(sc);
235 	/* If we are in attach call, do not sleep */
236 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
237 					? CAN_SLEEP:NO_SLEEP;
238 	error = 0;
239 	while (tries++ < 5) {
240 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
241 		mps_dprint(sc, MPS_INIT, "Doorbell= 0x%x\n", reg);
242 
243 		/*
244 		 * Ensure the IOC is ready to talk.  If it's not, try
245 		 * resetting it.
246 		 */
247 		if (reg & MPI2_DOORBELL_USED) {
248 			mps_diag_reset(sc, sleep_flags);
249 			DELAY(50000);
250 			continue;
251 		}
252 
253 		/* Is the adapter owned by another peer? */
254 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
255 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
256 			device_printf(sc->mps_dev, "IOC is under the control "
257 			    "of another peer host, aborting initialization.\n");
258 			return (ENXIO);
259 		}
260 
261 		state = reg & MPI2_IOC_STATE_MASK;
262 		if (state == MPI2_IOC_STATE_READY) {
263 			/* Ready to go! */
264 			error = 0;
265 			break;
266 		} else if (state == MPI2_IOC_STATE_FAULT) {
267 			mps_dprint(sc, MPS_FAULT, "IOC in fault state 0x%x, resetting\n",
268 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
269 			mps_diag_reset(sc, sleep_flags);
270 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
271 			/* Need to take ownership */
272 			mps_message_unit_reset(sc, sleep_flags);
273 		} else if (state == MPI2_IOC_STATE_RESET) {
274 			/* Wait a bit, IOC might be in transition */
275 			mps_dprint(sc, MPS_FAULT,
276 			    "IOC in unexpected reset state\n");
277 		} else {
278 			mps_dprint(sc, MPS_FAULT,
279 			    "IOC in unknown state 0x%x\n", state);
280 			error = EINVAL;
281 			break;
282 		}
283 
284 		/* Wait 50ms for things to settle down. */
285 		DELAY(50000);
286 	}
287 
288 	if (error)
289 		device_printf(sc->mps_dev, "Cannot transition IOC to ready\n");
290 
291 	return (error);
292 }
293 
294 static int
295 mps_transition_operational(struct mps_softc *sc)
296 {
297 	uint32_t reg, state;
298 	int error;
299 
300 	MPS_FUNCTRACE(sc);
301 
302 	error = 0;
303 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
304 	mps_dprint(sc, MPS_INIT, "Doorbell= 0x%x\n", reg);
305 
306 	state = reg & MPI2_IOC_STATE_MASK;
307 	if (state != MPI2_IOC_STATE_READY) {
308 		if ((error = mps_transition_ready(sc)) != 0) {
309 			mps_dprint(sc, MPS_FAULT,
310 			    "%s failed to transition ready\n", __func__);
311 			return (error);
312 		}
313 	}
314 
315 	error = mps_send_iocinit(sc);
316 	return (error);
317 }
318 
319 /*
320  * This is called during attach and when re-initializing due to a Diag Reset.
321  * IOC Facts is used to allocate many of the structures needed by the driver.
322  * If called from attach, de-allocation is not required because the driver has
323  * not allocated any structures yet, but if called from a Diag Reset, previously
324  * allocated structures based on IOC Facts will need to be freed and re-
325  * allocated bases on the latest IOC Facts.
326  */
327 static int
328 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
329 {
330 	int error;
331 	Mpi2IOCFactsReply_t saved_facts;
332 	uint8_t saved_mode, reallocating;
333 
334 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
335 
336 	/* Save old IOC Facts and then only reallocate if Facts have changed */
337 	if (!attaching) {
338 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
339 	}
340 
341 	/*
342 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
343 	 * a re-initialization and only return the error if attaching so the OS
344 	 * can handle it.
345 	 */
346 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
347 		if (attaching) {
348 			mps_dprint(sc, MPS_FAULT, "%s failed to get IOC Facts "
349 			    "with error %d\n", __func__, error);
350 			return (error);
351 		} else {
352 			panic("%s failed to get IOC Facts with error %d\n",
353 			    __func__, error);
354 		}
355 	}
356 
357 	mps_print_iocfacts(sc, sc->facts);
358 
359 	snprintf(sc->fw_version, sizeof(sc->fw_version),
360 	    "%02d.%02d.%02d.%02d",
361 	    sc->facts->FWVersion.Struct.Major,
362 	    sc->facts->FWVersion.Struct.Minor,
363 	    sc->facts->FWVersion.Struct.Unit,
364 	    sc->facts->FWVersion.Struct.Dev);
365 
366 	mps_printf(sc, "Firmware: %s, Driver: %s\n", sc->fw_version,
367 	    MPS_DRIVER_VERSION);
368 	mps_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
369 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
370 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
371 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
372 
373 	/*
374 	 * If the chip doesn't support event replay then a hard reset will be
375 	 * required to trigger a full discovery.  Do the reset here then
376 	 * retransition to Ready.  A hard reset might have already been done,
377 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
378 	 * for a Diag Reset.
379 	 */
380 	if (attaching) {
381 		if ((sc->facts->IOCCapabilities &
382 		    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
383 			mps_diag_reset(sc, NO_SLEEP);
384 			if ((error = mps_transition_ready(sc)) != 0) {
385 				mps_dprint(sc, MPS_FAULT, "%s failed to "
386 				    "transition to ready with error %d\n",
387 				    __func__, error);
388 				return (error);
389 			}
390 		}
391 	}
392 
393 	/*
394 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
395 	 * changed from the previous IOC Facts, log a warning, but only if
396 	 * checking this after a Diag Reset and not during attach.
397 	 */
398 	saved_mode = sc->ir_firmware;
399 	if (sc->facts->IOCCapabilities &
400 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
401 		sc->ir_firmware = 1;
402 	if (!attaching) {
403 		if (sc->ir_firmware != saved_mode) {
404 			mps_dprint(sc, MPS_FAULT, "%s new IR/IT mode in IOC "
405 			    "Facts does not match previous mode\n", __func__);
406 		}
407 	}
408 
409 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
410 	reallocating = FALSE;
411 	if ((!attaching) &&
412 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
413 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
414 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
415 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
416 	    (saved_facts.ProductID != sc->facts->ProductID) ||
417 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
418 	    (saved_facts.IOCRequestFrameSize !=
419 	    sc->facts->IOCRequestFrameSize) ||
420 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
421 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
422 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
423 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
424 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
425 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
426 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
427 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
428 	    (saved_facts.MaxPersistentEntries !=
429 	    sc->facts->MaxPersistentEntries))) {
430 		reallocating = TRUE;
431 	}
432 
433 	/*
434 	 * Some things should be done if attaching or re-allocating after a Diag
435 	 * Reset, but are not needed after a Diag Reset if the FW has not
436 	 * changed.
437 	 */
438 	if (attaching || reallocating) {
439 		/*
440 		 * Check if controller supports FW diag buffers and set flag to
441 		 * enable each type.
442 		 */
443 		if (sc->facts->IOCCapabilities &
444 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
445 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
446 			    enabled = TRUE;
447 		if (sc->facts->IOCCapabilities &
448 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
449 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
450 			    enabled = TRUE;
451 		if (sc->facts->IOCCapabilities &
452 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
453 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
454 			    enabled = TRUE;
455 
456 		/*
457 		 * Set flag if EEDP is supported and if TLR is supported.
458 		 */
459 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
460 			sc->eedp_enabled = TRUE;
461 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
462 			sc->control_TLR = TRUE;
463 
464 		/*
465 		 * Size the queues. Since the reply queues always need one free
466 		 * entry, we'll just deduct one reply message here.
467 		 */
468 		sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
469 		sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
470 		    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
471 
472 		/*
473 		 * Initialize all Tail Queues
474 		 */
475 		TAILQ_INIT(&sc->req_list);
476 		TAILQ_INIT(&sc->high_priority_req_list);
477 		TAILQ_INIT(&sc->chain_list);
478 		TAILQ_INIT(&sc->tm_list);
479 	}
480 
481 	/*
482 	 * If doing a Diag Reset and the FW is significantly different
483 	 * (reallocating will be set above in IOC Facts comparison), then all
484 	 * buffers based on the IOC Facts will need to be freed before they are
485 	 * reallocated.
486 	 */
487 	if (reallocating) {
488 		mps_iocfacts_free(sc);
489 		mpssas_realloc_targets(sc, saved_facts.MaxTargets);
490 	}
491 
492 	/*
493 	 * Any deallocation has been completed.  Now start reallocating
494 	 * if needed.  Will only need to reallocate if attaching or if the new
495 	 * IOC Facts are different from the previous IOC Facts after a Diag
496 	 * Reset. Targets have already been allocated above if needed.
497 	 */
498 	if (attaching || reallocating) {
499 		if (((error = mps_alloc_queues(sc)) != 0) ||
500 		    ((error = mps_alloc_replies(sc)) != 0) ||
501 		    ((error = mps_alloc_requests(sc)) != 0)) {
502 			if (attaching ) {
503 				mps_dprint(sc, MPS_FAULT, "%s failed to alloc "
504 				    "queues with error %d\n", __func__, error);
505 				mps_free(sc);
506 				return (error);
507 			} else {
508 				panic("%s failed to alloc queues with error "
509 				    "%d\n", __func__, error);
510 			}
511 		}
512 	}
513 
514 	/* Always initialize the queues */
515 	bzero(sc->free_queue, sc->fqdepth * 4);
516 	mps_init_queues(sc);
517 
518 	/*
519 	 * Always get the chip out of the reset state, but only panic if not
520 	 * attaching.  If attaching and there is an error, that is handled by
521 	 * the OS.
522 	 */
523 	error = mps_transition_operational(sc);
524 	if (error != 0) {
525 		if (attaching) {
526 			mps_printf(sc, "%s failed to transition to operational "
527 			    "with error %d\n", __func__, error);
528 			mps_free(sc);
529 			return (error);
530 		} else {
531 			panic("%s failed to transition to operational with "
532 			    "error %d\n", __func__, error);
533 		}
534 	}
535 
536 	/*
537 	 * Finish the queue initialization.
538 	 * These are set here instead of in mps_init_queues() because the
539 	 * IOC resets these values during the state transition in
540 	 * mps_transition_operational().  The free index is set to 1
541 	 * because the corresponding index in the IOC is set to 0, and the
542 	 * IOC treats the queues as full if both are set to the same value.
543 	 * Hence the reason that the queue can't hold all of the possible
544 	 * replies.
545 	 */
546 	sc->replypostindex = 0;
547 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
548 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
549 
550 	/*
551 	 * Attach the subsystems so they can prepare their event masks.
552 	 */
553 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
554 	if (attaching) {
555 		if (((error = mps_attach_log(sc)) != 0) ||
556 		    ((error = mps_attach_sas(sc)) != 0) ||
557 		    ((error = mps_attach_user(sc)) != 0)) {
558 			mps_printf(sc, "%s failed to attach all subsystems: "
559 			    "error %d\n", __func__, error);
560 			mps_free(sc);
561 			return (error);
562 		}
563 
564 		if ((error = mps_pci_setup_interrupts(sc)) != 0) {
565 			mps_printf(sc, "%s failed to setup interrupts\n",
566 			    __func__);
567 			mps_free(sc);
568 			return (error);
569 		}
570 	}
571 
572 	/*
573 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
574 	 * reset it after a Diag Reset, just in case.
575 	 */
576 	sc->WD_available = FALSE;
577 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
578 		sc->WD_available = TRUE;
579 
580 	return (error);
581 }
582 
583 /*
584  * This is called if memory is being free (during detach for example) and when
585  * buffers need to be reallocated due to a Diag Reset.
586  */
587 static void
588 mps_iocfacts_free(struct mps_softc *sc)
589 {
590 	struct mps_command *cm;
591 	int i;
592 
593 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
594 
595 	if (sc->post_busaddr != 0)
596 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
597 	if (sc->free_queue != NULL)
598 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
599 		    sc->queues_map);
600 	if (sc->queues_dmat != NULL)
601 		bus_dma_tag_destroy(sc->queues_dmat);
602 
603 	if (sc->chain_busaddr != 0)
604 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
605 	if (sc->chain_frames != NULL)
606 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
607 		    sc->chain_map);
608 	if (sc->chain_dmat != NULL)
609 		bus_dma_tag_destroy(sc->chain_dmat);
610 
611 	if (sc->sense_busaddr != 0)
612 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
613 	if (sc->sense_frames != NULL)
614 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
615 		    sc->sense_map);
616 	if (sc->sense_dmat != NULL)
617 		bus_dma_tag_destroy(sc->sense_dmat);
618 
619 	if (sc->reply_busaddr != 0)
620 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
621 	if (sc->reply_frames != NULL)
622 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
623 		    sc->reply_map);
624 	if (sc->reply_dmat != NULL)
625 		bus_dma_tag_destroy(sc->reply_dmat);
626 
627 	if (sc->req_busaddr != 0)
628 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
629 	if (sc->req_frames != NULL)
630 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
631 	if (sc->req_dmat != NULL)
632 		bus_dma_tag_destroy(sc->req_dmat);
633 
634 	if (sc->chains != NULL)
635 		free(sc->chains, M_MPT2);
636 	if (sc->commands != NULL) {
637 		for (i = 1; i < sc->num_reqs; i++) {
638 			cm = &sc->commands[i];
639 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
640 		}
641 		free(sc->commands, M_MPT2);
642 	}
643 	if (sc->buffer_dmat != NULL)
644 		bus_dma_tag_destroy(sc->buffer_dmat);
645 }
646 
647 /*
648  * The terms diag reset and hard reset are used interchangeably in the MPI
649  * docs to mean resetting the controller chip.  In this code diag reset
650  * cleans everything up, and the hard reset function just sends the reset
651  * sequence to the chip.  This should probably be refactored so that every
652  * subsystem gets a reset notification of some sort, and can clean up
653  * appropriately.
654  */
655 int
656 mps_reinit(struct mps_softc *sc)
657 {
658 	int error;
659 
660 	MPS_FUNCTRACE(sc);
661 
662 	mtx_assert(&sc->mps_mtx, MA_OWNED);
663 
664 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
665 		mps_dprint(sc, MPS_INIT, "%s reset already in progress\n",
666 			   __func__);
667 		return 0;
668 	}
669 
670 	mps_dprint(sc, MPS_INFO, "Reinitializing controller,\n");
671 	/* make sure the completion callbacks can recognize they're getting
672 	 * a NULL cm_reply due to a reset.
673 	 */
674 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
675 
676 	/*
677 	 * Mask interrupts here.
678 	 */
679 	mps_dprint(sc, MPS_INIT, "%s mask interrupts\n", __func__);
680 	mps_mask_intr(sc);
681 
682 	error = mps_diag_reset(sc, CAN_SLEEP);
683 	if (error != 0) {
684 		/* XXXSL No need to panic here */
685 		panic("%s hard reset failed with error %d\n",
686 		    __func__, error);
687 	}
688 
689 	/* Restore the PCI state, including the MSI-X registers */
690 	mps_pci_restore(sc);
691 
692 	/* Give the I/O subsystem special priority to get itself prepared */
693 	mpssas_handle_reinit(sc);
694 
695 	/*
696 	 * Get IOC Facts and allocate all structures based on this information.
697 	 * The attach function will also call mps_iocfacts_allocate at startup.
698 	 * If relevant values have changed in IOC Facts, this function will free
699 	 * all of the memory based on IOC Facts and reallocate that memory.
700 	 */
701 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
702 		panic("%s IOC Facts based allocation failed with error %d\n",
703 		    __func__, error);
704 	}
705 
706 	/*
707 	 * Mapping structures will be re-allocated after getting IOC Page8, so
708 	 * free these structures here.
709 	 */
710 	mps_mapping_exit(sc);
711 
712 	/*
713 	 * The static page function currently read is IOC Page8.  Others can be
714 	 * added in future.  It's possible that the values in IOC Page8 have
715 	 * changed after a Diag Reset due to user modification, so always read
716 	 * these.  Interrupts are masked, so unmask them before getting config
717 	 * pages.
718 	 */
719 	mps_unmask_intr(sc);
720 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
721 	mps_base_static_config_pages(sc);
722 
723 	/*
724 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
725 	 * mapping tables.
726 	 */
727 	mps_mapping_initialize(sc);
728 
729 	/*
730 	 * Restart will reload the event masks clobbered by the reset, and
731 	 * then enable the port.
732 	 */
733 	mps_reregister_events(sc);
734 
735 	/* the end of discovery will release the simq, so we're done. */
736 	mps_dprint(sc, MPS_INFO, "%s finished sc %p post %u free %u\n",
737 	    __func__, sc, sc->replypostindex, sc->replyfreeindex);
738 
739 	return 0;
740 }
741 
742 /* Wait for the chip to ACK a word that we've put into its FIFO
743  * Wait for <timeout> seconds. In single loop wait for busy loop
744  * for 500 microseconds.
745  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
746  * */
747 static int
748 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
749 {
750 
751 	u32 cntdn, count;
752 	u32 int_status;
753 	u32 doorbell;
754 
755 	count = 0;
756 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
757 	do {
758 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
759 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
760 			mps_dprint(sc, MPS_INIT,
761 			"%s: successfull count(%d), timeout(%d)\n",
762 			__func__, count, timeout);
763 		return 0;
764 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
765 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
766 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
767 				MPI2_IOC_STATE_FAULT) {
768 				mps_dprint(sc, MPS_FAULT,
769 					"fault_state(0x%04x)!\n", doorbell);
770 				return (EFAULT);
771 			}
772 		} else if (int_status == 0xFFFFFFFF)
773 			goto out;
774 
775 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
776 		* 0.5 milisecond */
777 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
778 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
779 			"mpsdba", hz/1000);
780 		else if (sleep_flag == CAN_SLEEP)
781 			pause("mpsdba", hz/1000);
782 		else
783 			DELAY(500);
784 		count++;
785 	} while (--cntdn);
786 
787 	out:
788 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
789 		"int_status(%x)!\n", __func__, count, int_status);
790 	return (ETIMEDOUT);
791 
792 }
793 
794 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
795 static int
796 mps_wait_db_int(struct mps_softc *sc)
797 {
798 	int retry;
799 
800 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
801 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
802 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
803 			return (0);
804 		DELAY(2000);
805 	}
806 	return (ETIMEDOUT);
807 }
808 
809 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
810 static int
811 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
812     int req_sz, int reply_sz, int timeout)
813 {
814 	uint32_t *data32;
815 	uint16_t *data16;
816 	int i, count, ioc_sz, residual;
817 	int sleep_flags = CAN_SLEEP;
818 
819 	if (curthread->td_no_sleeping != 0)
820 		sleep_flags = NO_SLEEP;
821 
822 	/* Step 1 */
823 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
824 
825 	/* Step 2 */
826 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
827 		return (EBUSY);
828 
829 	/* Step 3
830 	 * Announce that a message is coming through the doorbell.  Messages
831 	 * are pushed at 32bit words, so round up if needed.
832 	 */
833 	count = (req_sz + 3) / 4;
834 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
835 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
836 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
837 
838 	/* Step 4 */
839 	if (mps_wait_db_int(sc) ||
840 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
841 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
842 		return (ENXIO);
843 	}
844 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
845 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
846 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
847 		return (ENXIO);
848 	}
849 
850 	/* Step 5 */
851 	/* Clock out the message data synchronously in 32-bit dwords*/
852 	data32 = (uint32_t *)req;
853 	for (i = 0; i < count; i++) {
854 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
855 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
856 			mps_dprint(sc, MPS_FAULT,
857 			    "Timeout while writing doorbell\n");
858 			return (ENXIO);
859 		}
860 	}
861 
862 	/* Step 6 */
863 	/* Clock in the reply in 16-bit words.  The total length of the
864 	 * message is always in the 4th byte, so clock out the first 2 words
865 	 * manually, then loop the rest.
866 	 */
867 	data16 = (uint16_t *)reply;
868 	if (mps_wait_db_int(sc) != 0) {
869 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
870 		return (ENXIO);
871 	}
872 	data16[0] =
873 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
874 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
875 	if (mps_wait_db_int(sc) != 0) {
876 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
877 		return (ENXIO);
878 	}
879 	data16[1] =
880 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
881 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
882 
883 	/* Number of 32bit words in the message */
884 	ioc_sz = reply->MsgLength;
885 
886 	/*
887 	 * Figure out how many 16bit words to clock in without overrunning.
888 	 * The precision loss with dividing reply_sz can safely be
889 	 * ignored because the messages can only be multiples of 32bits.
890 	 */
891 	residual = 0;
892 	count = MIN((reply_sz / 4), ioc_sz) * 2;
893 	if (count < ioc_sz * 2) {
894 		residual = ioc_sz * 2 - count;
895 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
896 		    "residual message words\n", residual);
897 	}
898 
899 	for (i = 2; i < count; i++) {
900 		if (mps_wait_db_int(sc) != 0) {
901 			mps_dprint(sc, MPS_FAULT,
902 			    "Timeout reading doorbell %d\n", i);
903 			return (ENXIO);
904 		}
905 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
906 		    MPI2_DOORBELL_DATA_MASK;
907 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
908 	}
909 
910 	/*
911 	 * Pull out residual words that won't fit into the provided buffer.
912 	 * This keeps the chip from hanging due to a driver programming
913 	 * error.
914 	 */
915 	while (residual--) {
916 		if (mps_wait_db_int(sc) != 0) {
917 			mps_dprint(sc, MPS_FAULT,
918 			    "Timeout reading doorbell\n");
919 			return (ENXIO);
920 		}
921 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
922 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
923 	}
924 
925 	/* Step 7 */
926 	if (mps_wait_db_int(sc) != 0) {
927 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
928 		return (ENXIO);
929 	}
930 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
931 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
932 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
933 
934 	return (0);
935 }
936 
937 static void
938 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
939 {
940 	reply_descriptor rd;
941 	MPS_FUNCTRACE(sc);
942 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
943 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
944 
945 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
946 		mtx_assert(&sc->mps_mtx, MA_OWNED);
947 
948 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
949 		sc->io_cmds_highwater++;
950 	rd.u.low = cm->cm_desc.Words.Low;
951 	rd.u.high = cm->cm_desc.Words.High;
952 	rd.word = htole64(rd.word);
953 	/* TODO-We may need to make below regwrite atomic */
954 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
955 	    rd.u.low);
956 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
957 	    rd.u.high);
958 }
959 
960 /*
961  * Just the FACTS, ma'am.
962  */
963 static int
964 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
965 {
966 	MPI2_DEFAULT_REPLY *reply;
967 	MPI2_IOC_FACTS_REQUEST request;
968 	int error, req_sz, reply_sz;
969 
970 	MPS_FUNCTRACE(sc);
971 
972 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
973 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
974 	reply = (MPI2_DEFAULT_REPLY *)facts;
975 
976 	bzero(&request, req_sz);
977 	request.Function = MPI2_FUNCTION_IOC_FACTS;
978 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
979 
980 	return (error);
981 }
982 
983 static int
984 mps_send_iocinit(struct mps_softc *sc)
985 {
986 	MPI2_IOC_INIT_REQUEST	init;
987 	MPI2_DEFAULT_REPLY	reply;
988 	int req_sz, reply_sz, error;
989 	struct timeval now;
990 	uint64_t time_in_msec;
991 
992 	MPS_FUNCTRACE(sc);
993 
994 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
995 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
996 	bzero(&init, req_sz);
997 	bzero(&reply, reply_sz);
998 
999 	/*
1000 	 * Fill in the init block.  Note that most addresses are
1001 	 * deliberately in the lower 32bits of memory.  This is a micro-
1002 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1003 	 */
1004 	init.Function = MPI2_FUNCTION_IOC_INIT;
1005 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1006 	init.MsgVersion = htole16(MPI2_VERSION);
1007 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1008 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1009 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1010 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1011 	init.SenseBufferAddressHigh = 0;
1012 	init.SystemReplyAddressHigh = 0;
1013 	init.SystemRequestFrameBaseAddress.High = 0;
1014 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1015 	init.ReplyDescriptorPostQueueAddress.High = 0;
1016 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1017 	init.ReplyFreeQueueAddress.High = 0;
1018 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1019 	getmicrotime(&now);
1020 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1021 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1022 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1023 
1024 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1025 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1026 		error = ENXIO;
1027 
1028 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1029 	return (error);
1030 }
1031 
1032 void
1033 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1034 {
1035 	bus_addr_t *addr;
1036 
1037 	addr = arg;
1038 	*addr = segs[0].ds_addr;
1039 }
1040 
1041 static int
1042 mps_alloc_queues(struct mps_softc *sc)
1043 {
1044 	bus_addr_t queues_busaddr;
1045 	uint8_t *queues;
1046 	int qsize, fqsize, pqsize;
1047 
1048 	/*
1049 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1050 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1051 	 * This queue supplies fresh reply frames for the firmware to use.
1052 	 *
1053 	 * The reply descriptor post queue contains 8 byte entries in
1054 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1055 	 * contains filled-in reply frames sent from the firmware to the host.
1056 	 *
1057 	 * These two queues are allocated together for simplicity.
1058 	 */
1059 	sc->fqdepth = roundup2((sc->num_replies + 1), 16);
1060 	sc->pqdepth = roundup2((sc->num_replies + 1), 16);
1061 	fqsize= sc->fqdepth * 4;
1062 	pqsize = sc->pqdepth * 8;
1063 	qsize = fqsize + pqsize;
1064 
1065         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1066 				16, 0,			/* algnmnt, boundary */
1067 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1068 				BUS_SPACE_MAXADDR,	/* highaddr */
1069 				NULL, NULL,		/* filter, filterarg */
1070                                 qsize,			/* maxsize */
1071                                 1,			/* nsegments */
1072                                 qsize,			/* maxsegsize */
1073                                 0,			/* flags */
1074                                 NULL, NULL,		/* lockfunc, lockarg */
1075                                 &sc->queues_dmat)) {
1076 		device_printf(sc->mps_dev, "Cannot allocate queues DMA tag\n");
1077 		return (ENOMEM);
1078         }
1079         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1080 	    &sc->queues_map)) {
1081 		device_printf(sc->mps_dev, "Cannot allocate queues memory\n");
1082 		return (ENOMEM);
1083         }
1084         bzero(queues, qsize);
1085         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1086 	    mps_memaddr_cb, &queues_busaddr, 0);
1087 
1088 	sc->free_queue = (uint32_t *)queues;
1089 	sc->free_busaddr = queues_busaddr;
1090 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1091 	sc->post_busaddr = queues_busaddr + fqsize;
1092 
1093 	return (0);
1094 }
1095 
1096 static int
1097 mps_alloc_replies(struct mps_softc *sc)
1098 {
1099 	int rsize, num_replies;
1100 
1101 	/*
1102 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1103 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1104 	 * replies can be used at once.
1105 	 */
1106 	num_replies = max(sc->fqdepth, sc->num_replies);
1107 
1108 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1109         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1110 				4, 0,			/* algnmnt, boundary */
1111 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1112 				BUS_SPACE_MAXADDR,	/* highaddr */
1113 				NULL, NULL,		/* filter, filterarg */
1114                                 rsize,			/* maxsize */
1115                                 1,			/* nsegments */
1116                                 rsize,			/* maxsegsize */
1117                                 0,			/* flags */
1118                                 NULL, NULL,		/* lockfunc, lockarg */
1119                                 &sc->reply_dmat)) {
1120 		device_printf(sc->mps_dev, "Cannot allocate replies DMA tag\n");
1121 		return (ENOMEM);
1122         }
1123         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1124 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1125 		device_printf(sc->mps_dev, "Cannot allocate replies memory\n");
1126 		return (ENOMEM);
1127         }
1128         bzero(sc->reply_frames, rsize);
1129         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1130 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1131 
1132 	return (0);
1133 }
1134 
1135 static int
1136 mps_alloc_requests(struct mps_softc *sc)
1137 {
1138 	struct mps_command *cm;
1139 	struct mps_chain *chain;
1140 	int i, rsize, nsegs;
1141 
1142 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1143         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1144 				16, 0,			/* algnmnt, boundary */
1145 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1146 				BUS_SPACE_MAXADDR,	/* highaddr */
1147 				NULL, NULL,		/* filter, filterarg */
1148                                 rsize,			/* maxsize */
1149                                 1,			/* nsegments */
1150                                 rsize,			/* maxsegsize */
1151                                 0,			/* flags */
1152                                 NULL, NULL,		/* lockfunc, lockarg */
1153                                 &sc->req_dmat)) {
1154 		device_printf(sc->mps_dev, "Cannot allocate request DMA tag\n");
1155 		return (ENOMEM);
1156         }
1157         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1158 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1159 		device_printf(sc->mps_dev, "Cannot allocate request memory\n");
1160 		return (ENOMEM);
1161         }
1162         bzero(sc->req_frames, rsize);
1163         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1164 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1165 
1166 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1167         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1168 				16, 0,			/* algnmnt, boundary */
1169 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1170 				BUS_SPACE_MAXADDR,	/* highaddr */
1171 				NULL, NULL,		/* filter, filterarg */
1172                                 rsize,			/* maxsize */
1173                                 1,			/* nsegments */
1174                                 rsize,			/* maxsegsize */
1175                                 0,			/* flags */
1176                                 NULL, NULL,		/* lockfunc, lockarg */
1177                                 &sc->chain_dmat)) {
1178 		device_printf(sc->mps_dev, "Cannot allocate chain DMA tag\n");
1179 		return (ENOMEM);
1180         }
1181         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1182 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1183 		device_printf(sc->mps_dev, "Cannot allocate chain memory\n");
1184 		return (ENOMEM);
1185         }
1186         bzero(sc->chain_frames, rsize);
1187         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1188 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1189 
1190 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1191         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1192 				1, 0,			/* algnmnt, boundary */
1193 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1194 				BUS_SPACE_MAXADDR,	/* highaddr */
1195 				NULL, NULL,		/* filter, filterarg */
1196                                 rsize,			/* maxsize */
1197                                 1,			/* nsegments */
1198                                 rsize,			/* maxsegsize */
1199                                 0,			/* flags */
1200                                 NULL, NULL,		/* lockfunc, lockarg */
1201                                 &sc->sense_dmat)) {
1202 		device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
1203 		return (ENOMEM);
1204         }
1205         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1206 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1207 		device_printf(sc->mps_dev, "Cannot allocate sense memory\n");
1208 		return (ENOMEM);
1209         }
1210         bzero(sc->sense_frames, rsize);
1211         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1212 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1213 
1214 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1215 	    M_WAITOK | M_ZERO);
1216 	if(!sc->chains) {
1217 		device_printf(sc->mps_dev,
1218 		"Cannot allocate chains memory %s %d\n",
1219 		 __func__, __LINE__);
1220 		return (ENOMEM);
1221 	}
1222 	for (i = 0; i < sc->max_chains; i++) {
1223 		chain = &sc->chains[i];
1224 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1225 		    i * sc->facts->IOCRequestFrameSize * 4);
1226 		chain->chain_busaddr = sc->chain_busaddr +
1227 		    i * sc->facts->IOCRequestFrameSize * 4;
1228 		mps_free_chain(sc, chain);
1229 		sc->chain_free_lowwater++;
1230 	}
1231 
1232 	/* XXX Need to pick a more precise value */
1233 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1234         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1235 				1, 0,			/* algnmnt, boundary */
1236 				BUS_SPACE_MAXADDR,	/* lowaddr */
1237 				BUS_SPACE_MAXADDR,	/* highaddr */
1238 				NULL, NULL,		/* filter, filterarg */
1239                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1240                                 nsegs,			/* nsegments */
1241                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1242                                 BUS_DMA_ALLOCNOW,	/* flags */
1243                                 busdma_lock_mutex,	/* lockfunc */
1244 				&sc->mps_mtx,		/* lockarg */
1245                                 &sc->buffer_dmat)) {
1246 		device_printf(sc->mps_dev, "Cannot allocate buffer DMA tag\n");
1247 		return (ENOMEM);
1248         }
1249 
1250 	/*
1251 	 * SMID 0 cannot be used as a free command per the firmware spec.
1252 	 * Just drop that command instead of risking accounting bugs.
1253 	 */
1254 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1255 	    M_MPT2, M_WAITOK | M_ZERO);
1256 	if(!sc->commands) {
1257 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1258 		 __func__, __LINE__);
1259 		return (ENOMEM);
1260 	}
1261 	for (i = 1; i < sc->num_reqs; i++) {
1262 		cm = &sc->commands[i];
1263 		cm->cm_req = sc->req_frames +
1264 		    i * sc->facts->IOCRequestFrameSize * 4;
1265 		cm->cm_req_busaddr = sc->req_busaddr +
1266 		    i * sc->facts->IOCRequestFrameSize * 4;
1267 		cm->cm_sense = &sc->sense_frames[i];
1268 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1269 		cm->cm_desc.Default.SMID = i;
1270 		cm->cm_sc = sc;
1271 		TAILQ_INIT(&cm->cm_chain_list);
1272 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1273 
1274 		/* XXX Is a failure here a critical problem? */
1275 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1276 			if (i <= sc->facts->HighPriorityCredit)
1277 				mps_free_high_priority_command(sc, cm);
1278 			else
1279 				mps_free_command(sc, cm);
1280 		else {
1281 			panic("failed to allocate command %d\n", i);
1282 			sc->num_reqs = i;
1283 			break;
1284 		}
1285 	}
1286 
1287 	return (0);
1288 }
1289 
1290 static int
1291 mps_init_queues(struct mps_softc *sc)
1292 {
1293 	int i;
1294 
1295 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1296 
1297 	/*
1298 	 * According to the spec, we need to use one less reply than we
1299 	 * have space for on the queue.  So sc->num_replies (the number we
1300 	 * use) should be less than sc->fqdepth (allocated size).
1301 	 */
1302 	if (sc->num_replies >= sc->fqdepth)
1303 		return (EINVAL);
1304 
1305 	/*
1306 	 * Initialize all of the free queue entries.
1307 	 */
1308 	for (i = 0; i < sc->fqdepth; i++)
1309 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1310 	sc->replyfreeindex = sc->num_replies;
1311 
1312 	return (0);
1313 }
1314 
1315 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1316  * Next are the global settings, if they exist.  Highest are the per-unit
1317  * settings, if they exist.
1318  */
1319 static void
1320 mps_get_tunables(struct mps_softc *sc)
1321 {
1322 	char tmpstr[80];
1323 
1324 	/* XXX default to some debugging for now */
1325 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1326 	sc->disable_msix = 0;
1327 	sc->disable_msi = 0;
1328 	sc->max_chains = MPS_CHAIN_FRAMES;
1329 
1330 	/*
1331 	 * Grab the global variables.
1332 	 */
1333 	TUNABLE_INT_FETCH("hw.mps.debug_level", &sc->mps_debug);
1334 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1335 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1336 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1337 
1338 	/* Grab the unit-instance variables */
1339 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1340 	    device_get_unit(sc->mps_dev));
1341 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
1342 
1343 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1344 	    device_get_unit(sc->mps_dev));
1345 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1346 
1347 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1348 	    device_get_unit(sc->mps_dev));
1349 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1350 
1351 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1352 	    device_get_unit(sc->mps_dev));
1353 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1354 
1355 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1356 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1357 	    device_get_unit(sc->mps_dev));
1358 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1359 }
1360 
1361 static void
1362 mps_setup_sysctl(struct mps_softc *sc)
1363 {
1364 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1365 	struct sysctl_oid	*sysctl_tree = NULL;
1366 	char tmpstr[80], tmpstr2[80];
1367 
1368 	/*
1369 	 * Setup the sysctl variable so the user can change the debug level
1370 	 * on the fly.
1371 	 */
1372 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1373 	    device_get_unit(sc->mps_dev));
1374 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1375 
1376 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1377 	if (sysctl_ctx != NULL)
1378 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1379 
1380 	if (sysctl_tree == NULL) {
1381 		sysctl_ctx_init(&sc->sysctl_ctx);
1382 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1383 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1384 		    CTLFLAG_RD, 0, tmpstr);
1385 		if (sc->sysctl_tree == NULL)
1386 			return;
1387 		sysctl_ctx = &sc->sysctl_ctx;
1388 		sysctl_tree = sc->sysctl_tree;
1389 	}
1390 
1391 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1392 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
1393 	    "mps debug level");
1394 
1395 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1396 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1397 	    "Disable the use of MSI-X interrupts");
1398 
1399 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1400 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1401 	    "Disable the use of MSI interrupts");
1402 
1403 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1404 	    OID_AUTO, "firmware_version", CTLFLAG_RW, &sc->fw_version,
1405 	    strlen(sc->fw_version), "firmware version");
1406 
1407 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1408 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1409 	    strlen(MPS_DRIVER_VERSION), "driver version");
1410 
1411 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1412 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1413 	    &sc->io_cmds_active, 0, "number of currently active commands");
1414 
1415 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1416 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1417 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1418 
1419 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1420 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1421 	    &sc->chain_free, 0, "number of free chain elements");
1422 
1423 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1424 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1425 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1426 
1427 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1428 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1429 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1430 
1431 #if __FreeBSD_version >= 900030
1432 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1433 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1434 	    &sc->chain_alloc_fail, "chain allocation failures");
1435 #endif //FreeBSD_version >= 900030
1436 }
1437 
1438 int
1439 mps_attach(struct mps_softc *sc)
1440 {
1441 	int error;
1442 
1443 	mps_get_tunables(sc);
1444 
1445 	MPS_FUNCTRACE(sc);
1446 
1447 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1448 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1449 	TAILQ_INIT(&sc->event_list);
1450 	timevalclear(&sc->lastfail);
1451 
1452 	if ((error = mps_transition_ready(sc)) != 0) {
1453 		mps_printf(sc, "%s failed to transition ready\n", __func__);
1454 		return (error);
1455 	}
1456 
1457 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1458 	    M_ZERO|M_NOWAIT);
1459 	if(!sc->facts) {
1460 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1461 		 __func__, __LINE__);
1462 		return (ENOMEM);
1463 	}
1464 
1465 	/*
1466 	 * Get IOC Facts and allocate all structures based on this information.
1467 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1468 	 * Facts. If relevant values have changed in IOC Facts, this function
1469 	 * will free all of the memory based on IOC Facts and reallocate that
1470 	 * memory.  If this fails, any allocated memory should already be freed.
1471 	 */
1472 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1473 		mps_dprint(sc, MPS_FAULT, "%s IOC Facts based allocation "
1474 		    "failed with error %d\n", __func__, error);
1475 		return (error);
1476 	}
1477 
1478 	/* Start the periodic watchdog check on the IOC Doorbell */
1479 	mps_periodic(sc);
1480 
1481 	/*
1482 	 * The portenable will kick off discovery events that will drive the
1483 	 * rest of the initialization process.  The CAM/SAS module will
1484 	 * hold up the boot sequence until discovery is complete.
1485 	 */
1486 	sc->mps_ich.ich_func = mps_startup;
1487 	sc->mps_ich.ich_arg = sc;
1488 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1489 		mps_dprint(sc, MPS_ERROR, "Cannot establish MPS config hook\n");
1490 		error = EINVAL;
1491 	}
1492 
1493 	/*
1494 	 * Allow IR to shutdown gracefully when shutdown occurs.
1495 	 */
1496 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1497 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1498 
1499 	if (sc->shutdown_eh == NULL)
1500 		mps_dprint(sc, MPS_ERROR, "shutdown event registration "
1501 		    "failed\n");
1502 
1503 	mps_setup_sysctl(sc);
1504 
1505 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1506 
1507 	return (error);
1508 }
1509 
1510 /* Run through any late-start handlers. */
1511 static void
1512 mps_startup(void *arg)
1513 {
1514 	struct mps_softc *sc;
1515 
1516 	sc = (struct mps_softc *)arg;
1517 
1518 	mps_lock(sc);
1519 	mps_unmask_intr(sc);
1520 
1521 	/* initialize device mapping tables */
1522 	mps_base_static_config_pages(sc);
1523 	mps_mapping_initialize(sc);
1524 	mpssas_startup(sc);
1525 	mps_unlock(sc);
1526 }
1527 
1528 /* Periodic watchdog.  Is called with the driver lock already held. */
1529 static void
1530 mps_periodic(void *arg)
1531 {
1532 	struct mps_softc *sc;
1533 	uint32_t db;
1534 
1535 	sc = (struct mps_softc *)arg;
1536 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1537 		return;
1538 
1539 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1540 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1541 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1542 		mps_reinit(sc);
1543 	}
1544 
1545 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1546 }
1547 
1548 static void
1549 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1550     MPI2_EVENT_NOTIFICATION_REPLY *event)
1551 {
1552 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1553 
1554 	mps_print_event(sc, event);
1555 
1556 	switch (event->Event) {
1557 	case MPI2_EVENT_LOG_DATA:
1558 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1559 		if (sc->mps_debug & MPS_EVENT)
1560 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1561 		break;
1562 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1563 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1564 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1565 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1566 		     entry->LogSequence);
1567 		break;
1568 	default:
1569 		break;
1570 	}
1571 	return;
1572 }
1573 
1574 static int
1575 mps_attach_log(struct mps_softc *sc)
1576 {
1577 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1578 
1579 	bzero(events, 16);
1580 	setbit(events, MPI2_EVENT_LOG_DATA);
1581 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1582 
1583 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1584 	    &sc->mps_log_eh);
1585 
1586 	return (0);
1587 }
1588 
1589 static int
1590 mps_detach_log(struct mps_softc *sc)
1591 {
1592 
1593 	if (sc->mps_log_eh != NULL)
1594 		mps_deregister_events(sc, sc->mps_log_eh);
1595 	return (0);
1596 }
1597 
1598 /*
1599  * Free all of the driver resources and detach submodules.  Should be called
1600  * without the lock held.
1601  */
1602 int
1603 mps_free(struct mps_softc *sc)
1604 {
1605 	int error;
1606 
1607 	/* Turn off the watchdog */
1608 	mps_lock(sc);
1609 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1610 	mps_unlock(sc);
1611 	/* Lock must not be held for this */
1612 	callout_drain(&sc->periodic);
1613 
1614 	if (((error = mps_detach_log(sc)) != 0) ||
1615 	    ((error = mps_detach_sas(sc)) != 0))
1616 		return (error);
1617 
1618 	mps_detach_user(sc);
1619 
1620 	/* Put the IOC back in the READY state. */
1621 	mps_lock(sc);
1622 	if ((error = mps_transition_ready(sc)) != 0) {
1623 		mps_unlock(sc);
1624 		return (error);
1625 	}
1626 	mps_unlock(sc);
1627 
1628 	if (sc->facts != NULL)
1629 		free(sc->facts, M_MPT2);
1630 
1631 	/*
1632 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
1633 	 * to free these buffers too.
1634 	 */
1635 	mps_iocfacts_free(sc);
1636 
1637 	if (sc->sysctl_tree != NULL)
1638 		sysctl_ctx_free(&sc->sysctl_ctx);
1639 
1640 	/* Deregister the shutdown function */
1641 	if (sc->shutdown_eh != NULL)
1642 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1643 
1644 	mtx_destroy(&sc->mps_mtx);
1645 
1646 	return (0);
1647 }
1648 
1649 static __inline void
1650 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
1651 {
1652 	MPS_FUNCTRACE(sc);
1653 
1654 	if (cm == NULL) {
1655 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
1656 		return;
1657 	}
1658 
1659 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1660 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1661 
1662 	if (cm->cm_complete != NULL) {
1663 		mps_dprint(sc, MPS_TRACE,
1664 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
1665 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
1666 			   cm->cm_reply);
1667 		cm->cm_complete(sc, cm);
1668 	}
1669 
1670 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1671 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
1672 		wakeup(cm);
1673 	}
1674 
1675 	if (cm->cm_sc->io_cmds_active != 0) {
1676 		cm->cm_sc->io_cmds_active--;
1677 	} else {
1678 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
1679 		    "out of sync - resynching to 0\n");
1680 	}
1681 }
1682 
1683 
1684 static void
1685 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
1686 {
1687 	union loginfo_type {
1688 		u32     loginfo;
1689 		struct {
1690 			u32     subcode:16;
1691 			u32     code:8;
1692 			u32     originator:4;
1693 			u32     bus_type:4;
1694 		} dw;
1695 	};
1696 	union loginfo_type sas_loginfo;
1697 	char *originator_str = NULL;
1698 
1699 	sas_loginfo.loginfo = log_info;
1700 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1701 		return;
1702 
1703 	/* each nexus loss loginfo */
1704 	if (log_info == 0x31170000)
1705 		return;
1706 
1707 	/* eat the loginfos associated with task aborts */
1708 	if ((log_info == 30050000 || log_info ==
1709 	    0x31140000 || log_info == 0x31130000))
1710 		return;
1711 
1712 	switch (sas_loginfo.dw.originator) {
1713 	case 0:
1714 		originator_str = "IOP";
1715 		break;
1716 	case 1:
1717 		originator_str = "PL";
1718 		break;
1719 	case 2:
1720 		originator_str = "IR";
1721 		break;
1722 }
1723 
1724 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
1725 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
1726 	originator_str, sas_loginfo.dw.code,
1727 	sas_loginfo.dw.subcode);
1728 }
1729 
1730 static void
1731 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
1732 {
1733 	MPI2DefaultReply_t *mpi_reply;
1734 	u16 sc_status;
1735 
1736 	mpi_reply = (MPI2DefaultReply_t*)reply;
1737 	sc_status = le16toh(mpi_reply->IOCStatus);
1738 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1739 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1740 }
1741 void
1742 mps_intr(void *data)
1743 {
1744 	struct mps_softc *sc;
1745 	uint32_t status;
1746 
1747 	sc = (struct mps_softc *)data;
1748 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1749 
1750 	/*
1751 	 * Check interrupt status register to flush the bus.  This is
1752 	 * needed for both INTx interrupts and driver-driven polling
1753 	 */
1754 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1755 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1756 		return;
1757 
1758 	mps_lock(sc);
1759 	mps_intr_locked(data);
1760 	mps_unlock(sc);
1761 	return;
1762 }
1763 
1764 /*
1765  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1766  * chip.  Hopefully this theory is correct.
1767  */
1768 void
1769 mps_intr_msi(void *data)
1770 {
1771 	struct mps_softc *sc;
1772 
1773 	sc = (struct mps_softc *)data;
1774 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1775 	mps_lock(sc);
1776 	mps_intr_locked(data);
1777 	mps_unlock(sc);
1778 	return;
1779 }
1780 
1781 /*
1782  * The locking is overly broad and simplistic, but easy to deal with for now.
1783  */
1784 void
1785 mps_intr_locked(void *data)
1786 {
1787 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1788 	struct mps_softc *sc;
1789 	struct mps_command *cm = NULL;
1790 	uint8_t flags;
1791 	u_int pq;
1792 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
1793 	mps_fw_diagnostic_buffer_t *pBuffer;
1794 
1795 	sc = (struct mps_softc *)data;
1796 
1797 	pq = sc->replypostindex;
1798 	mps_dprint(sc, MPS_TRACE,
1799 	    "%s sc %p starting with replypostindex %u\n",
1800 	    __func__, sc, sc->replypostindex);
1801 
1802 	for ( ;; ) {
1803 		cm = NULL;
1804 		desc = &sc->post_queue[sc->replypostindex];
1805 		flags = desc->Default.ReplyFlags &
1806 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1807 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
1808 		 || (le32toh(desc->Words.High) == 0xffffffff))
1809 			break;
1810 
1811 		/* increment the replypostindex now, so that event handlers
1812 		 * and cm completion handlers which decide to do a diag
1813 		 * reset can zero it without it getting incremented again
1814 		 * afterwards, and we break out of this loop on the next
1815 		 * iteration since the reply post queue has been cleared to
1816 		 * 0xFF and all descriptors look unused (which they are).
1817 		 */
1818 		if (++sc->replypostindex >= sc->pqdepth)
1819 			sc->replypostindex = 0;
1820 
1821 		switch (flags) {
1822 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1823 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
1824 			cm->cm_reply = NULL;
1825 			break;
1826 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1827 		{
1828 			uint32_t baddr;
1829 			uint8_t *reply;
1830 
1831 			/*
1832 			 * Re-compose the reply address from the address
1833 			 * sent back from the chip.  The ReplyFrameAddress
1834 			 * is the lower 32 bits of the physical address of
1835 			 * particular reply frame.  Convert that address to
1836 			 * host format, and then use that to provide the
1837 			 * offset against the virtual address base
1838 			 * (sc->reply_frames).
1839 			 */
1840 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1841 			reply = sc->reply_frames +
1842 				(baddr - ((uint32_t)sc->reply_busaddr));
1843 			/*
1844 			 * Make sure the reply we got back is in a valid
1845 			 * range.  If not, go ahead and panic here, since
1846 			 * we'll probably panic as soon as we deference the
1847 			 * reply pointer anyway.
1848 			 */
1849 			if ((reply < sc->reply_frames)
1850 			 || (reply > (sc->reply_frames +
1851 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1852 				printf("%s: WARNING: reply %p out of range!\n",
1853 				       __func__, reply);
1854 				printf("%s: reply_frames %p, fqdepth %d, "
1855 				       "frame size %d\n", __func__,
1856 				       sc->reply_frames, sc->fqdepth,
1857 				       sc->facts->ReplyFrameSize * 4);
1858 				printf("%s: baddr %#x,\n", __func__, baddr);
1859 				/* LSI-TODO. See Linux Code. Need Gracefull exit*/
1860 				panic("Reply address out of range");
1861 			}
1862 			if (le16toh(desc->AddressReply.SMID) == 0) {
1863 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
1864 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
1865 					/*
1866 					 * If SMID is 0 for Diag Buffer Post,
1867 					 * this implies that the reply is due to
1868 					 * a release function with a status that
1869 					 * the buffer has been released.  Set
1870 					 * the buffer flags accordingly.
1871 					 */
1872 					rel_rep =
1873 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
1874 					if (le16toh(rel_rep->IOCStatus) ==
1875 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
1876 					    {
1877 						pBuffer =
1878 						    &sc->fw_diag_buffer_list[
1879 						    rel_rep->BufferType];
1880 						pBuffer->valid_data = TRUE;
1881 						pBuffer->owned_by_firmware =
1882 						    FALSE;
1883 						pBuffer->immediate = FALSE;
1884 					}
1885 				} else
1886 					mps_dispatch_event(sc, baddr,
1887 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
1888 					    reply);
1889 			} else {
1890 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
1891 				cm->cm_reply = reply;
1892 				cm->cm_reply_data =
1893 				    le32toh(desc->AddressReply.ReplyFrameAddress);
1894 			}
1895 			break;
1896 		}
1897 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
1898 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
1899 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
1900 		default:
1901 			/* Unhandled */
1902 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
1903 			    desc->Default.ReplyFlags);
1904 			cm = NULL;
1905 			break;
1906 		}
1907 
1908 
1909 		if (cm != NULL) {
1910 			// Print Error reply frame
1911 			if (cm->cm_reply)
1912 				mps_display_reply_info(sc,cm->cm_reply);
1913 			mps_complete_command(sc, cm);
1914 		}
1915 
1916 		desc->Words.Low = 0xffffffff;
1917 		desc->Words.High = 0xffffffff;
1918 	}
1919 
1920 	if (pq != sc->replypostindex) {
1921 		mps_dprint(sc, MPS_TRACE,
1922 		    "%s sc %p writing postindex %d\n",
1923 		    __func__, sc, sc->replypostindex);
1924 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
1925 	}
1926 
1927 	return;
1928 }
1929 
1930 static void
1931 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
1932     MPI2_EVENT_NOTIFICATION_REPLY *reply)
1933 {
1934 	struct mps_event_handle *eh;
1935 	int event, handled = 0;
1936 
1937 	event = le16toh(reply->Event);
1938 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1939 		if (isset(eh->mask, event)) {
1940 			eh->callback(sc, data, reply);
1941 			handled++;
1942 		}
1943 	}
1944 
1945 	if (handled == 0)
1946 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
1947 
1948 	/*
1949 	 * This is the only place that the event/reply should be freed.
1950 	 * Anything wanting to hold onto the event data should have
1951 	 * already copied it into their own storage.
1952 	 */
1953 	mps_free_reply(sc, data);
1954 }
1955 
1956 static void
1957 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
1958 {
1959 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1960 
1961 	if (cm->cm_reply)
1962 		mps_print_event(sc,
1963 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
1964 
1965 	mps_free_command(sc, cm);
1966 
1967 	/* next, send a port enable */
1968 	mpssas_startup(sc);
1969 }
1970 
1971 /*
1972  * For both register_events and update_events, the caller supplies a bitmap
1973  * of events that it _wants_.  These functions then turn that into a bitmask
1974  * suitable for the controller.
1975  */
1976 int
1977 mps_register_events(struct mps_softc *sc, u32 *mask,
1978     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
1979 {
1980 	struct mps_event_handle *eh;
1981 	int error = 0;
1982 
1983 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
1984 	if(!eh) {
1985 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1986 		 __func__, __LINE__);
1987 		return (ENOMEM);
1988 	}
1989 	eh->callback = cb;
1990 	eh->data = data;
1991 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
1992 	if (mask != NULL)
1993 		error = mps_update_events(sc, eh, mask);
1994 	*handle = eh;
1995 
1996 	return (error);
1997 }
1998 
1999 int
2000 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2001     u32 *mask)
2002 {
2003 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2004 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
2005 	struct mps_command *cm;
2006 	int error, i;
2007 
2008 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2009 
2010 	if ((mask != NULL) && (handle != NULL))
2011 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2012 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2013 
2014 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2015 		sc->event_mask[i] = -1;
2016 
2017 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2018 		sc->event_mask[i] &= ~handle->mask[i];
2019 
2020 
2021 	if ((cm = mps_alloc_command(sc)) == NULL)
2022 		return (EBUSY);
2023 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2024 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2025 	evtreq->MsgFlags = 0;
2026 	evtreq->SASBroadcastPrimitiveMasks = 0;
2027 #ifdef MPS_DEBUG_ALL_EVENTS
2028 	{
2029 		u_char fullmask[16];
2030 		memset(fullmask, 0x00, 16);
2031 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2032 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2033 	}
2034 #else
2035         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2036                 evtreq->EventMasks[i] =
2037                     htole32(sc->event_mask[i]);
2038 #endif
2039 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2040 	cm->cm_data = NULL;
2041 
2042 	error = mps_request_polled(sc, cm);
2043 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2044 	if ((reply == NULL) ||
2045 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2046 		error = ENXIO;
2047 	mps_print_event(sc, reply);
2048 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2049 
2050 	mps_free_command(sc, cm);
2051 	return (error);
2052 }
2053 
2054 static int
2055 mps_reregister_events(struct mps_softc *sc)
2056 {
2057 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2058 	struct mps_command *cm;
2059 	struct mps_event_handle *eh;
2060 	int error, i;
2061 
2062 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2063 
2064 	/* first, reregister events */
2065 
2066 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2067 		sc->event_mask[i] = -1;
2068 
2069 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2070 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2071 			sc->event_mask[i] &= ~eh->mask[i];
2072 	}
2073 
2074 	if ((cm = mps_alloc_command(sc)) == NULL)
2075 		return (EBUSY);
2076 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2077 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2078 	evtreq->MsgFlags = 0;
2079 	evtreq->SASBroadcastPrimitiveMasks = 0;
2080 #ifdef MPS_DEBUG_ALL_EVENTS
2081 	{
2082 		u_char fullmask[16];
2083 		memset(fullmask, 0x00, 16);
2084 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2085 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2086 	}
2087 #else
2088         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2089                 evtreq->EventMasks[i] =
2090                     htole32(sc->event_mask[i]);
2091 #endif
2092 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2093 	cm->cm_data = NULL;
2094 	cm->cm_complete = mps_reregister_events_complete;
2095 
2096 	error = mps_map_command(sc, cm);
2097 
2098 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2099 	    error);
2100 	return (error);
2101 }
2102 
2103 void
2104 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2105 {
2106 
2107 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2108 	free(handle, M_MPT2);
2109 }
2110 
2111 /*
2112  * Add a chain element as the next SGE for the specified command.
2113  * Reset cm_sge and cm_sgesize to indicate all the available space.
2114  */
2115 static int
2116 mps_add_chain(struct mps_command *cm)
2117 {
2118 	MPI2_SGE_CHAIN32 *sgc;
2119 	struct mps_chain *chain;
2120 	int space;
2121 
2122 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2123 		panic("MPS: Need SGE Error Code\n");
2124 
2125 	chain = mps_alloc_chain(cm->cm_sc);
2126 	if (chain == NULL)
2127 		return (ENOBUFS);
2128 
2129 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2130 
2131 	/*
2132 	 * Note: a double-linked list is used to make it easier to
2133 	 * walk for debugging.
2134 	 */
2135 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2136 
2137 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2138 	sgc->Length = htole16(space);
2139 	sgc->NextChainOffset = 0;
2140 	/* TODO Looks like bug in Setting sgc->Flags.
2141 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2142 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2143 	 *	This is fine.. because we are not using simple element. In case of
2144 	 *	MPI2_SGE_CHAIN32, we have seperate Length and Flags feild.
2145  	 */
2146 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2147 	sgc->Address = htole32(chain->chain_busaddr);
2148 
2149 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2150 	cm->cm_sglsize = space;
2151 	return (0);
2152 }
2153 
2154 /*
2155  * Add one scatter-gather element (chain, simple, transaction context)
2156  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2157  * cm_sge as the remaining size and pointer to the next SGE to fill
2158  * in, respectively.
2159  */
2160 int
2161 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2162 {
2163 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2164 	MPI2_SGE_SIMPLE64 *sge = sgep;
2165 	int error, type;
2166 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2167 
2168 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2169 
2170 #ifdef INVARIANTS
2171 	switch (type) {
2172 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2173 		if (len != tc->DetailsLength + 4)
2174 			panic("TC %p length %u or %zu?", tc,
2175 			    tc->DetailsLength + 4, len);
2176 		}
2177 		break;
2178 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2179 		/* Driver only uses 32-bit chain elements */
2180 		if (len != MPS_SGC_SIZE)
2181 			panic("CHAIN %p length %u or %zu?", sgep,
2182 			    MPS_SGC_SIZE, len);
2183 		break;
2184 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2185 		/* Driver only uses 64-bit SGE simple elements */
2186 		if (len != MPS_SGE64_SIZE)
2187 			panic("SGE simple %p length %u or %zu?", sge,
2188 			    MPS_SGE64_SIZE, len);
2189 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2190 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2191 			panic("SGE simple %p not marked 64-bit?", sge);
2192 
2193 		break;
2194 	default:
2195 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2196 	}
2197 #endif
2198 
2199 	/*
2200 	 * case 1: 1 more segment, enough room for it
2201 	 * case 2: 2 more segments, enough room for both
2202 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2203 	 * case 4: >=1 more segment, enough room for only a chain
2204 	 * case 5: >=1 more segment, no room for anything (error)
2205          */
2206 
2207 	/*
2208 	 * There should be room for at least a chain element, or this
2209 	 * code is buggy.  Case (5).
2210 	 */
2211 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2212 		panic("MPS: Need SGE Error Code\n");
2213 
2214 	if (segsleft >= 2 &&
2215 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2216 		/*
2217 		 * There are 2 or more segments left to add, and only
2218 		 * enough room for 1 and a chain.  Case (3).
2219 		 *
2220 		 * Mark as last element in this chain if necessary.
2221 		 */
2222 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2223 			sge->FlagsLength |= htole32(
2224 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2225 		}
2226 
2227 		/*
2228 		 * Add the item then a chain.  Do the chain now,
2229 		 * rather than on the next iteration, to simplify
2230 		 * understanding the code.
2231 		 */
2232 		cm->cm_sglsize -= len;
2233 		bcopy(sgep, cm->cm_sge, len);
2234 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2235 		return (mps_add_chain(cm));
2236 	}
2237 
2238 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2239 		/*
2240 		 * 1 or more segment, enough room for only a chain.
2241 		 * Hope the previous element wasn't a Simple entry
2242 		 * that needed to be marked with
2243 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2244 		 */
2245 		if ((error = mps_add_chain(cm)) != 0)
2246 			return (error);
2247 	}
2248 
2249 #ifdef INVARIANTS
2250 	/* Case 1: 1 more segment, enough room for it. */
2251 	if (segsleft == 1 && cm->cm_sglsize < len)
2252 		panic("1 seg left and no room? %u versus %zu",
2253 		    cm->cm_sglsize, len);
2254 
2255 	/* Case 2: 2 more segments, enough room for both */
2256 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2257 		panic("2 segs left and no room? %u versus %zu",
2258 		    cm->cm_sglsize, len);
2259 #endif
2260 
2261 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2262 		/*
2263 		 * If this is a bi-directional request, need to account for that
2264 		 * here.  Save the pre-filled sge values.  These will be used
2265 		 * either for the 2nd SGL or for a single direction SGL.  If
2266 		 * cm_out_len is non-zero, this is a bi-directional request, so
2267 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2268 		 * fill in the IN SGL.  Note that at this time, when filling in
2269 		 * 2 SGL's for a bi-directional request, they both use the same
2270 		 * DMA buffer (same cm command).
2271 		 */
2272 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2273 		saved_address_low = sge->Address.Low;
2274 		saved_address_high = sge->Address.High;
2275 		if (cm->cm_out_len) {
2276 			sge->FlagsLength = htole32(cm->cm_out_len |
2277 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2278 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2279 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2280 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2281 			    MPI2_SGE_FLAGS_SHIFT));
2282 			cm->cm_sglsize -= len;
2283 			bcopy(sgep, cm->cm_sge, len);
2284 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2285 			    + len);
2286 		}
2287 		saved_buf_len |=
2288 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2289 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2290 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2291 		    MPI2_SGE_FLAGS_END_OF_LIST |
2292 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2293 		    MPI2_SGE_FLAGS_SHIFT);
2294 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2295 			saved_buf_len |=
2296 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2297 			    MPI2_SGE_FLAGS_SHIFT);
2298 		} else {
2299 			saved_buf_len |=
2300 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2301 			    MPI2_SGE_FLAGS_SHIFT);
2302 		}
2303 		sge->FlagsLength = htole32(saved_buf_len);
2304 		sge->Address.Low = saved_address_low;
2305 		sge->Address.High = saved_address_high;
2306 	}
2307 
2308 	cm->cm_sglsize -= len;
2309 	bcopy(sgep, cm->cm_sge, len);
2310 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2311 	return (0);
2312 }
2313 
2314 /*
2315  * Add one dma segment to the scatter-gather list for a command.
2316  */
2317 int
2318 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2319     int segsleft)
2320 {
2321 	MPI2_SGE_SIMPLE64 sge;
2322 
2323 	/*
2324 	 * This driver always uses 64-bit address elements for simplicity.
2325 	 */
2326 	bzero(&sge, sizeof(sge));
2327 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2328 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2329 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2330 	mps_from_u64(pa, &sge.Address);
2331 
2332 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2333 }
2334 
2335 static void
2336 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2337 {
2338 	struct mps_softc *sc;
2339 	struct mps_command *cm;
2340 	u_int i, dir, sflags;
2341 
2342 	cm = (struct mps_command *)arg;
2343 	sc = cm->cm_sc;
2344 
2345 	/*
2346 	 * In this case, just print out a warning and let the chip tell the
2347 	 * user they did the wrong thing.
2348 	 */
2349 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2350 		mps_dprint(sc, MPS_ERROR,
2351 			   "%s: warning: busdma returned %d segments, "
2352 			   "more than the %d allowed\n", __func__, nsegs,
2353 			   cm->cm_max_segs);
2354 	}
2355 
2356 	/*
2357 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2358 	 * here.  In that case, both direction flags will be set.
2359 	 */
2360 	sflags = 0;
2361 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2362 		/*
2363 		 * We have to add a special case for SMP passthrough, there
2364 		 * is no easy way to generically handle it.  The first
2365 		 * S/G element is used for the command (therefore the
2366 		 * direction bit needs to be set).  The second one is used
2367 		 * for the reply.  We'll leave it to the caller to make
2368 		 * sure we only have two buffers.
2369 		 */
2370 		/*
2371 		 * Even though the busdma man page says it doesn't make
2372 		 * sense to have both direction flags, it does in this case.
2373 		 * We have one s/g element being accessed in each direction.
2374 		 */
2375 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2376 
2377 		/*
2378 		 * Set the direction flag on the first buffer in the SMP
2379 		 * passthrough request.  We'll clear it for the second one.
2380 		 */
2381 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2382 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2383 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2384 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2385 		dir = BUS_DMASYNC_PREWRITE;
2386 	} else
2387 		dir = BUS_DMASYNC_PREREAD;
2388 
2389 	for (i = 0; i < nsegs; i++) {
2390 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2391 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2392 		}
2393 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2394 		    sflags, nsegs - i);
2395 		if (error != 0) {
2396 			/* Resource shortage, roll back! */
2397 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2398 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2399 				    "consider increasing hw.mps.max_chains.\n");
2400 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2401 			mps_complete_command(sc, cm);
2402 			return;
2403 		}
2404 	}
2405 
2406 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2407 	mps_enqueue_request(sc, cm);
2408 
2409 	return;
2410 }
2411 
2412 static void
2413 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2414 	     int error)
2415 {
2416 	mps_data_cb(arg, segs, nsegs, error);
2417 }
2418 
2419 /*
2420  * This is the routine to enqueue commands ansynchronously.
2421  * Note that the only error path here is from bus_dmamap_load(), which can
2422  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2423  * assumed that if you have a command in-hand, then you have enough credits
2424  * to use it.
2425  */
2426 int
2427 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2428 {
2429 	int error = 0;
2430 
2431 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2432 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2433 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2434 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2435 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2436 		    cm->cm_data, mps_data_cb, cm, 0);
2437 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2438 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2439 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2440 	} else {
2441 		/* Add a zero-length element as needed */
2442 		if (cm->cm_sge != NULL)
2443 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2444 		mps_enqueue_request(sc, cm);
2445 	}
2446 
2447 	return (error);
2448 }
2449 
2450 /*
2451  * This is the routine to enqueue commands synchronously.  An error of
2452  * EINPROGRESS from mps_map_command() is ignored since the command will
2453  * be executed and enqueued automatically.  Other errors come from msleep().
2454  */
2455 int
2456 mps_wait_command(struct mps_softc *sc, struct mps_command *cm, int timeout,
2457     int sleep_flag)
2458 {
2459 	int error, rc;
2460 	struct timeval cur_time, start_time;
2461 
2462 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2463 		return  EBUSY;
2464 
2465 	cm->cm_complete = NULL;
2466 	cm->cm_flags |= (MPS_CM_FLAGS_WAKEUP + MPS_CM_FLAGS_POLLED);
2467 	error = mps_map_command(sc, cm);
2468 	if ((error != 0) && (error != EINPROGRESS))
2469 		return (error);
2470 
2471 	// Check for context and wait for 50 mSec at a time until time has
2472 	// expired or the command has finished.  If msleep can't be used, need
2473 	// to poll.
2474 	if (curthread->td_no_sleeping != 0)
2475 		sleep_flag = NO_SLEEP;
2476 	getmicrotime(&start_time);
2477 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2478 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2479 	} else {
2480 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2481 			mps_intr_locked(sc);
2482 			if (sleep_flag == CAN_SLEEP)
2483 				pause("mpswait", hz/20);
2484 			else
2485 				DELAY(50000);
2486 
2487 			getmicrotime(&cur_time);
2488 			if ((cur_time.tv_sec - start_time.tv_sec) > timeout) {
2489 				error = EWOULDBLOCK;
2490 				break;
2491 			}
2492 		}
2493 	}
2494 
2495 	if (error == EWOULDBLOCK) {
2496 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s\n", __func__);
2497 		rc = mps_reinit(sc);
2498 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2499 		    "failed");
2500 		error = ETIMEDOUT;
2501 	}
2502 	return (error);
2503 }
2504 
2505 /*
2506  * This is the routine to enqueue a command synchonously and poll for
2507  * completion.  Its use should be rare.
2508  */
2509 int
2510 mps_request_polled(struct mps_softc *sc, struct mps_command *cm)
2511 {
2512 	int error, timeout = 0, rc;
2513 
2514 	error = 0;
2515 
2516 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2517 	cm->cm_complete = NULL;
2518 	mps_map_command(sc, cm);
2519 
2520 	while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2521 		mps_intr_locked(sc);
2522 
2523 		DELAY(50 * 1000);
2524 		if (timeout++ > 1000) {
2525 			mps_dprint(sc, MPS_FAULT, "polling failed\n");
2526 			error = ETIMEDOUT;
2527 			break;
2528 		}
2529 	}
2530 
2531 	if (error) {
2532 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s\n", __func__);
2533 		rc = mps_reinit(sc);
2534 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n",
2535 				(rc == 0) ? "success" : "failed");
2536 	}
2537 
2538 	return (error);
2539 }
2540 
2541 /*
2542  * The MPT driver had a verbose interface for config pages.  In this driver,
2543  * reduce it to much simplier terms, similar to the Linux driver.
2544  */
2545 int
2546 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2547 {
2548 	MPI2_CONFIG_REQUEST *req;
2549 	struct mps_command *cm;
2550 	int error;
2551 
2552 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2553 		return (EBUSY);
2554 	}
2555 
2556 	cm = mps_alloc_command(sc);
2557 	if (cm == NULL) {
2558 		return (EBUSY);
2559 	}
2560 
2561 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2562 	req->Function = MPI2_FUNCTION_CONFIG;
2563 	req->Action = params->action;
2564 	req->SGLFlags = 0;
2565 	req->ChainOffset = 0;
2566 	req->PageAddress = params->page_address;
2567 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2568 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2569 
2570 		hdr = &params->hdr.Ext;
2571 		req->ExtPageType = hdr->ExtPageType;
2572 		req->ExtPageLength = hdr->ExtPageLength;
2573 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2574 		req->Header.PageLength = 0; /* Must be set to zero */
2575 		req->Header.PageNumber = hdr->PageNumber;
2576 		req->Header.PageVersion = hdr->PageVersion;
2577 	} else {
2578 		MPI2_CONFIG_PAGE_HEADER *hdr;
2579 
2580 		hdr = &params->hdr.Struct;
2581 		req->Header.PageType = hdr->PageType;
2582 		req->Header.PageNumber = hdr->PageNumber;
2583 		req->Header.PageLength = hdr->PageLength;
2584 		req->Header.PageVersion = hdr->PageVersion;
2585 	}
2586 
2587 	cm->cm_data = params->buffer;
2588 	cm->cm_length = params->length;
2589 	cm->cm_sge = &req->PageBufferSGE;
2590 	cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2591 	cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2592 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2593 
2594 	cm->cm_complete_data = params;
2595 	if (params->callback != NULL) {
2596 		cm->cm_complete = mps_config_complete;
2597 		return (mps_map_command(sc, cm));
2598 	} else {
2599 		error = mps_wait_command(sc, cm, 0, CAN_SLEEP);
2600 		if (error) {
2601 			mps_dprint(sc, MPS_FAULT,
2602 			    "Error %d reading config page\n", error);
2603 			mps_free_command(sc, cm);
2604 			return (error);
2605 		}
2606 		mps_config_complete(sc, cm);
2607 	}
2608 
2609 	return (0);
2610 }
2611 
2612 int
2613 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2614 {
2615 	return (EINVAL);
2616 }
2617 
2618 static void
2619 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
2620 {
2621 	MPI2_CONFIG_REPLY *reply;
2622 	struct mps_config_params *params;
2623 
2624 	MPS_FUNCTRACE(sc);
2625 	params = cm->cm_complete_data;
2626 
2627 	if (cm->cm_data != NULL) {
2628 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2629 		    BUS_DMASYNC_POSTREAD);
2630 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2631 	}
2632 
2633 	/*
2634 	 * XXX KDM need to do more error recovery?  This results in the
2635 	 * device in question not getting probed.
2636 	 */
2637 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
2638 		params->status = MPI2_IOCSTATUS_BUSY;
2639 		goto done;
2640 	}
2641 
2642 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2643 	if (reply == NULL) {
2644 		params->status = MPI2_IOCSTATUS_BUSY;
2645 		goto done;
2646 	}
2647 	params->status = reply->IOCStatus;
2648 	if (params->hdr.Ext.ExtPageType != 0) {
2649 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2650 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2651 	} else {
2652 		params->hdr.Struct.PageType = reply->Header.PageType;
2653 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2654 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2655 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2656 	}
2657 
2658 done:
2659 	mps_free_command(sc, cm);
2660 	if (params->callback != NULL)
2661 		params->callback(sc, params);
2662 
2663 	return;
2664 }
2665