xref: /freebsd/sys/dev/mps/mps.c (revision 313376588638950ba1e93c403dd8c97bc52fd3a2)
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 	uint8_t first_wait_done = FALSE;
144 
145 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
146 
147 	/* Clear any pending interrupts */
148 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
149 
150 	/*Force NO_SLEEP for threads prohibited to sleep
151  	* e.a Thread from interrupt handler are prohibited to sleep.
152  	*/
153 	if (curthread->td_no_sleeping != 0)
154 		sleep_flag = NO_SLEEP;
155 
156 	/* Push the magic sequence */
157 	error = ETIMEDOUT;
158 	while (tries++ < 20) {
159 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
160 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
161 			    mpt2_reset_magic[i]);
162 		/* wait 100 msec */
163 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
164 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
165 			    "mpsdiag", hz/10);
166 		else if (sleep_flag == CAN_SLEEP)
167 			pause("mpsdiag", hz/10);
168 		else
169 			DELAY(100 * 1000);
170 
171 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
172 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
173 			error = 0;
174 			break;
175 		}
176 	}
177 	if (error)
178 		return (error);
179 
180 	/* Send the actual reset.  XXX need to refresh the reg? */
181 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
182 	    reg | MPI2_DIAG_RESET_ADAPTER);
183 
184 	/* Wait up to 300 seconds in 50ms intervals */
185 	error = ETIMEDOUT;
186 	for (i = 0; i < 6000; i++) {
187 		/*
188 		 * Wait 50 msec. If this is the first time through, wait 256
189 		 * msec to satisfy Diag Reset timing requirements.
190 		 */
191 		if (first_wait_done) {
192 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
193 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
194 				    "mpsdiag", hz/20);
195 			else if (sleep_flag == CAN_SLEEP)
196 				pause("mpsdiag", hz/20);
197 			else
198 				DELAY(50 * 1000);
199 		} else {
200 			DELAY(256 * 1000);
201 			first_wait_done = TRUE;
202 		}
203 		/*
204 		 * Check for the RESET_ADAPTER bit to be cleared first, then
205 		 * wait for the RESET state to be cleared, which takes a little
206 		 * longer.
207 		 */
208 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
209 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
210 			continue;
211 		}
212 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
213 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
214 			error = 0;
215 			break;
216 		}
217 	}
218 	if (error)
219 		return (error);
220 
221 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
222 
223 	return (0);
224 }
225 
226 static int
227 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
228 {
229 
230 	MPS_FUNCTRACE(sc);
231 
232 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
233 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
234 	    MPI2_DOORBELL_FUNCTION_SHIFT);
235 
236 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
237 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed : <%s>\n",
238 				__func__);
239 		return (ETIMEDOUT);
240 	}
241 
242 	return (0);
243 }
244 
245 static int
246 mps_transition_ready(struct mps_softc *sc)
247 {
248 	uint32_t reg, state;
249 	int error, tries = 0;
250 	int sleep_flags;
251 
252 	MPS_FUNCTRACE(sc);
253 	/* If we are in attach call, do not sleep */
254 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
255 					? CAN_SLEEP:NO_SLEEP;
256 	error = 0;
257 	while (tries++ < 1200) {
258 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
259 		mps_dprint(sc, MPS_INIT, "Doorbell= 0x%x\n", reg);
260 
261 		/*
262 		 * Ensure the IOC is ready to talk.  If it's not, try
263 		 * resetting it.
264 		 */
265 		if (reg & MPI2_DOORBELL_USED) {
266 			mps_diag_reset(sc, sleep_flags);
267 			DELAY(50000);
268 			continue;
269 		}
270 
271 		/* Is the adapter owned by another peer? */
272 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
273 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
274 			device_printf(sc->mps_dev, "IOC is under the control "
275 			    "of another peer host, aborting initialization.\n");
276 			return (ENXIO);
277 		}
278 
279 		state = reg & MPI2_IOC_STATE_MASK;
280 		if (state == MPI2_IOC_STATE_READY) {
281 			/* Ready to go! */
282 			error = 0;
283 			break;
284 		} else if (state == MPI2_IOC_STATE_FAULT) {
285 			mps_dprint(sc, MPS_FAULT, "IOC in fault state 0x%x, resetting\n",
286 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
287 			mps_diag_reset(sc, sleep_flags);
288 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
289 			/* Need to take ownership */
290 			mps_message_unit_reset(sc, sleep_flags);
291 		} else if (state == MPI2_IOC_STATE_RESET) {
292 			/* Wait a bit, IOC might be in transition */
293 			mps_dprint(sc, MPS_FAULT,
294 			    "IOC in unexpected reset state\n");
295 		} else {
296 			mps_dprint(sc, MPS_FAULT,
297 			    "IOC in unknown state 0x%x\n", state);
298 			error = EINVAL;
299 			break;
300 		}
301 
302 		/* Wait 50ms for things to settle down. */
303 		DELAY(50000);
304 	}
305 
306 	if (error)
307 		device_printf(sc->mps_dev, "Cannot transition IOC to ready\n");
308 
309 	return (error);
310 }
311 
312 static int
313 mps_transition_operational(struct mps_softc *sc)
314 {
315 	uint32_t reg, state;
316 	int error;
317 
318 	MPS_FUNCTRACE(sc);
319 
320 	error = 0;
321 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
322 	mps_dprint(sc, MPS_INIT, "Doorbell= 0x%x\n", reg);
323 
324 	state = reg & MPI2_IOC_STATE_MASK;
325 	if (state != MPI2_IOC_STATE_READY) {
326 		if ((error = mps_transition_ready(sc)) != 0) {
327 			mps_dprint(sc, MPS_FAULT,
328 			    "%s failed to transition ready\n", __func__);
329 			return (error);
330 		}
331 	}
332 
333 	error = mps_send_iocinit(sc);
334 	return (error);
335 }
336 
337 /*
338  * This is called during attach and when re-initializing due to a Diag Reset.
339  * IOC Facts is used to allocate many of the structures needed by the driver.
340  * If called from attach, de-allocation is not required because the driver has
341  * not allocated any structures yet, but if called from a Diag Reset, previously
342  * allocated structures based on IOC Facts will need to be freed and re-
343  * allocated bases on the latest IOC Facts.
344  */
345 static int
346 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
347 {
348 	int error;
349 	Mpi2IOCFactsReply_t saved_facts;
350 	uint8_t saved_mode, reallocating;
351 
352 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
353 
354 	/* Save old IOC Facts and then only reallocate if Facts have changed */
355 	if (!attaching) {
356 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
357 	}
358 
359 	/*
360 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
361 	 * a re-initialization and only return the error if attaching so the OS
362 	 * can handle it.
363 	 */
364 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
365 		if (attaching) {
366 			mps_dprint(sc, MPS_FAULT, "%s failed to get IOC Facts "
367 			    "with error %d\n", __func__, error);
368 			return (error);
369 		} else {
370 			panic("%s failed to get IOC Facts with error %d\n",
371 			    __func__, error);
372 		}
373 	}
374 
375 	mps_print_iocfacts(sc, sc->facts);
376 
377 	snprintf(sc->fw_version, sizeof(sc->fw_version),
378 	    "%02d.%02d.%02d.%02d",
379 	    sc->facts->FWVersion.Struct.Major,
380 	    sc->facts->FWVersion.Struct.Minor,
381 	    sc->facts->FWVersion.Struct.Unit,
382 	    sc->facts->FWVersion.Struct.Dev);
383 
384 	mps_printf(sc, "Firmware: %s, Driver: %s\n", sc->fw_version,
385 	    MPS_DRIVER_VERSION);
386 	mps_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
387 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
388 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
389 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
390 
391 	/*
392 	 * If the chip doesn't support event replay then a hard reset will be
393 	 * required to trigger a full discovery.  Do the reset here then
394 	 * retransition to Ready.  A hard reset might have already been done,
395 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
396 	 * for a Diag Reset.
397 	 */
398 	if (attaching) {
399 		if ((sc->facts->IOCCapabilities &
400 		    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
401 			mps_diag_reset(sc, NO_SLEEP);
402 			if ((error = mps_transition_ready(sc)) != 0) {
403 				mps_dprint(sc, MPS_FAULT, "%s failed to "
404 				    "transition to ready with error %d\n",
405 				    __func__, error);
406 				return (error);
407 			}
408 		}
409 	}
410 
411 	/*
412 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
413 	 * changed from the previous IOC Facts, log a warning, but only if
414 	 * checking this after a Diag Reset and not during attach.
415 	 */
416 	saved_mode = sc->ir_firmware;
417 	if (sc->facts->IOCCapabilities &
418 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
419 		sc->ir_firmware = 1;
420 	if (!attaching) {
421 		if (sc->ir_firmware != saved_mode) {
422 			mps_dprint(sc, MPS_FAULT, "%s new IR/IT mode in IOC "
423 			    "Facts does not match previous mode\n", __func__);
424 		}
425 	}
426 
427 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
428 	reallocating = FALSE;
429 	if ((!attaching) &&
430 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
431 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
432 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
433 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
434 	    (saved_facts.ProductID != sc->facts->ProductID) ||
435 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
436 	    (saved_facts.IOCRequestFrameSize !=
437 	    sc->facts->IOCRequestFrameSize) ||
438 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
439 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
440 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
441 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
442 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
443 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
444 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
445 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
446 	    (saved_facts.MaxPersistentEntries !=
447 	    sc->facts->MaxPersistentEntries))) {
448 		reallocating = TRUE;
449 	}
450 
451 	/*
452 	 * Some things should be done if attaching or re-allocating after a Diag
453 	 * Reset, but are not needed after a Diag Reset if the FW has not
454 	 * changed.
455 	 */
456 	if (attaching || reallocating) {
457 		/*
458 		 * Check if controller supports FW diag buffers and set flag to
459 		 * enable each type.
460 		 */
461 		if (sc->facts->IOCCapabilities &
462 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
463 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
464 			    enabled = TRUE;
465 		if (sc->facts->IOCCapabilities &
466 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
467 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
468 			    enabled = TRUE;
469 		if (sc->facts->IOCCapabilities &
470 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
471 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
472 			    enabled = TRUE;
473 
474 		/*
475 		 * Set flag if EEDP is supported and if TLR is supported.
476 		 */
477 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
478 			sc->eedp_enabled = TRUE;
479 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
480 			sc->control_TLR = TRUE;
481 
482 		/*
483 		 * Size the queues. Since the reply queues always need one free
484 		 * entry, we'll just deduct one reply message here.
485 		 */
486 		sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
487 		sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
488 		    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
489 
490 		/*
491 		 * Initialize all Tail Queues
492 		 */
493 		TAILQ_INIT(&sc->req_list);
494 		TAILQ_INIT(&sc->high_priority_req_list);
495 		TAILQ_INIT(&sc->chain_list);
496 		TAILQ_INIT(&sc->tm_list);
497 	}
498 
499 	/*
500 	 * If doing a Diag Reset and the FW is significantly different
501 	 * (reallocating will be set above in IOC Facts comparison), then all
502 	 * buffers based on the IOC Facts will need to be freed before they are
503 	 * reallocated.
504 	 */
505 	if (reallocating) {
506 		mps_iocfacts_free(sc);
507 		mpssas_realloc_targets(sc, saved_facts.MaxTargets);
508 	}
509 
510 	/*
511 	 * Any deallocation has been completed.  Now start reallocating
512 	 * if needed.  Will only need to reallocate if attaching or if the new
513 	 * IOC Facts are different from the previous IOC Facts after a Diag
514 	 * Reset. Targets have already been allocated above if needed.
515 	 */
516 	if (attaching || reallocating) {
517 		if (((error = mps_alloc_queues(sc)) != 0) ||
518 		    ((error = mps_alloc_replies(sc)) != 0) ||
519 		    ((error = mps_alloc_requests(sc)) != 0)) {
520 			if (attaching ) {
521 				mps_dprint(sc, MPS_FAULT, "%s failed to alloc "
522 				    "queues with error %d\n", __func__, error);
523 				mps_free(sc);
524 				return (error);
525 			} else {
526 				panic("%s failed to alloc queues with error "
527 				    "%d\n", __func__, error);
528 			}
529 		}
530 	}
531 
532 	/* Always initialize the queues */
533 	bzero(sc->free_queue, sc->fqdepth * 4);
534 	mps_init_queues(sc);
535 
536 	/*
537 	 * Always get the chip out of the reset state, but only panic if not
538 	 * attaching.  If attaching and there is an error, that is handled by
539 	 * the OS.
540 	 */
541 	error = mps_transition_operational(sc);
542 	if (error != 0) {
543 		if (attaching) {
544 			mps_printf(sc, "%s failed to transition to operational "
545 			    "with error %d\n", __func__, error);
546 			mps_free(sc);
547 			return (error);
548 		} else {
549 			panic("%s failed to transition to operational with "
550 			    "error %d\n", __func__, error);
551 		}
552 	}
553 
554 	/*
555 	 * Finish the queue initialization.
556 	 * These are set here instead of in mps_init_queues() because the
557 	 * IOC resets these values during the state transition in
558 	 * mps_transition_operational().  The free index is set to 1
559 	 * because the corresponding index in the IOC is set to 0, and the
560 	 * IOC treats the queues as full if both are set to the same value.
561 	 * Hence the reason that the queue can't hold all of the possible
562 	 * replies.
563 	 */
564 	sc->replypostindex = 0;
565 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
566 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
567 
568 	/*
569 	 * Attach the subsystems so they can prepare their event masks.
570 	 */
571 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
572 	if (attaching) {
573 		if (((error = mps_attach_log(sc)) != 0) ||
574 		    ((error = mps_attach_sas(sc)) != 0) ||
575 		    ((error = mps_attach_user(sc)) != 0)) {
576 			mps_printf(sc, "%s failed to attach all subsystems: "
577 			    "error %d\n", __func__, error);
578 			mps_free(sc);
579 			return (error);
580 		}
581 
582 		if ((error = mps_pci_setup_interrupts(sc)) != 0) {
583 			mps_printf(sc, "%s failed to setup interrupts\n",
584 			    __func__);
585 			mps_free(sc);
586 			return (error);
587 		}
588 	}
589 
590 	/*
591 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
592 	 * reset it after a Diag Reset, just in case.
593 	 */
594 	sc->WD_available = FALSE;
595 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
596 		sc->WD_available = TRUE;
597 
598 	return (error);
599 }
600 
601 /*
602  * This is called if memory is being free (during detach for example) and when
603  * buffers need to be reallocated due to a Diag Reset.
604  */
605 static void
606 mps_iocfacts_free(struct mps_softc *sc)
607 {
608 	struct mps_command *cm;
609 	int i;
610 
611 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
612 
613 	if (sc->free_busaddr != 0)
614 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
615 	if (sc->free_queue != NULL)
616 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
617 		    sc->queues_map);
618 	if (sc->queues_dmat != NULL)
619 		bus_dma_tag_destroy(sc->queues_dmat);
620 
621 	if (sc->chain_busaddr != 0)
622 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
623 	if (sc->chain_frames != NULL)
624 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
625 		    sc->chain_map);
626 	if (sc->chain_dmat != NULL)
627 		bus_dma_tag_destroy(sc->chain_dmat);
628 
629 	if (sc->sense_busaddr != 0)
630 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
631 	if (sc->sense_frames != NULL)
632 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
633 		    sc->sense_map);
634 	if (sc->sense_dmat != NULL)
635 		bus_dma_tag_destroy(sc->sense_dmat);
636 
637 	if (sc->reply_busaddr != 0)
638 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
639 	if (sc->reply_frames != NULL)
640 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
641 		    sc->reply_map);
642 	if (sc->reply_dmat != NULL)
643 		bus_dma_tag_destroy(sc->reply_dmat);
644 
645 	if (sc->req_busaddr != 0)
646 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
647 	if (sc->req_frames != NULL)
648 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
649 	if (sc->req_dmat != NULL)
650 		bus_dma_tag_destroy(sc->req_dmat);
651 
652 	if (sc->chains != NULL)
653 		free(sc->chains, M_MPT2);
654 	if (sc->commands != NULL) {
655 		for (i = 1; i < sc->num_reqs; i++) {
656 			cm = &sc->commands[i];
657 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
658 		}
659 		free(sc->commands, M_MPT2);
660 	}
661 	if (sc->buffer_dmat != NULL)
662 		bus_dma_tag_destroy(sc->buffer_dmat);
663 }
664 
665 /*
666  * The terms diag reset and hard reset are used interchangeably in the MPI
667  * docs to mean resetting the controller chip.  In this code diag reset
668  * cleans everything up, and the hard reset function just sends the reset
669  * sequence to the chip.  This should probably be refactored so that every
670  * subsystem gets a reset notification of some sort, and can clean up
671  * appropriately.
672  */
673 int
674 mps_reinit(struct mps_softc *sc)
675 {
676 	int error;
677 	struct mpssas_softc *sassc;
678 
679 	sassc = sc->sassc;
680 
681 	MPS_FUNCTRACE(sc);
682 
683 	mtx_assert(&sc->mps_mtx, MA_OWNED);
684 
685 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
686 		mps_dprint(sc, MPS_INIT, "%s reset already in progress\n",
687 			   __func__);
688 		return 0;
689 	}
690 
691 	mps_dprint(sc, MPS_INFO, "Reinitializing controller,\n");
692 	/* make sure the completion callbacks can recognize they're getting
693 	 * a NULL cm_reply due to a reset.
694 	 */
695 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
696 
697 	/*
698 	 * Mask interrupts here.
699 	 */
700 	mps_dprint(sc, MPS_INIT, "%s mask interrupts\n", __func__);
701 	mps_mask_intr(sc);
702 
703 	error = mps_diag_reset(sc, CAN_SLEEP);
704 	if (error != 0) {
705 		/* XXXSL No need to panic here */
706 		panic("%s hard reset failed with error %d\n",
707 		    __func__, error);
708 	}
709 
710 	/* Restore the PCI state, including the MSI-X registers */
711 	mps_pci_restore(sc);
712 
713 	/* Give the I/O subsystem special priority to get itself prepared */
714 	mpssas_handle_reinit(sc);
715 
716 	/*
717 	 * Get IOC Facts and allocate all structures based on this information.
718 	 * The attach function will also call mps_iocfacts_allocate at startup.
719 	 * If relevant values have changed in IOC Facts, this function will free
720 	 * all of the memory based on IOC Facts and reallocate that memory.
721 	 */
722 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
723 		panic("%s IOC Facts based allocation failed with error %d\n",
724 		    __func__, error);
725 	}
726 
727 	/*
728 	 * Mapping structures will be re-allocated after getting IOC Page8, so
729 	 * free these structures here.
730 	 */
731 	mps_mapping_exit(sc);
732 
733 	/*
734 	 * The static page function currently read is IOC Page8.  Others can be
735 	 * added in future.  It's possible that the values in IOC Page8 have
736 	 * changed after a Diag Reset due to user modification, so always read
737 	 * these.  Interrupts are masked, so unmask them before getting config
738 	 * pages.
739 	 */
740 	mps_unmask_intr(sc);
741 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
742 	mps_base_static_config_pages(sc);
743 
744 	/*
745 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
746 	 * mapping tables.
747 	 */
748 	mps_mapping_initialize(sc);
749 
750 	/*
751 	 * Restart will reload the event masks clobbered by the reset, and
752 	 * then enable the port.
753 	 */
754 	mps_reregister_events(sc);
755 
756 	/* the end of discovery will release the simq, so we're done. */
757 	mps_dprint(sc, MPS_INFO, "%s finished sc %p post %u free %u\n",
758 	    __func__, sc, sc->replypostindex, sc->replyfreeindex);
759 
760 	mpssas_release_simq_reinit(sassc);
761 
762 	return 0;
763 }
764 
765 /* Wait for the chip to ACK a word that we've put into its FIFO
766  * Wait for <timeout> seconds. In single loop wait for busy loop
767  * for 500 microseconds.
768  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
769  * */
770 static int
771 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
772 {
773 
774 	u32 cntdn, count;
775 	u32 int_status;
776 	u32 doorbell;
777 
778 	count = 0;
779 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
780 	do {
781 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
782 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
783 			mps_dprint(sc, MPS_INIT,
784 			"%s: successfull count(%d), timeout(%d)\n",
785 			__func__, count, timeout);
786 		return 0;
787 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
788 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
789 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
790 				MPI2_IOC_STATE_FAULT) {
791 				mps_dprint(sc, MPS_FAULT,
792 					"fault_state(0x%04x)!\n", doorbell);
793 				return (EFAULT);
794 			}
795 		} else if (int_status == 0xFFFFFFFF)
796 			goto out;
797 
798 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
799 		* 0.5 milisecond */
800 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
801 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
802 			"mpsdba", hz/1000);
803 		else if (sleep_flag == CAN_SLEEP)
804 			pause("mpsdba", hz/1000);
805 		else
806 			DELAY(500);
807 		count++;
808 	} while (--cntdn);
809 
810 	out:
811 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
812 		"int_status(%x)!\n", __func__, count, int_status);
813 	return (ETIMEDOUT);
814 
815 }
816 
817 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
818 static int
819 mps_wait_db_int(struct mps_softc *sc)
820 {
821 	int retry;
822 
823 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
824 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
825 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
826 			return (0);
827 		DELAY(2000);
828 	}
829 	return (ETIMEDOUT);
830 }
831 
832 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
833 static int
834 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
835     int req_sz, int reply_sz, int timeout)
836 {
837 	uint32_t *data32;
838 	uint16_t *data16;
839 	int i, count, ioc_sz, residual;
840 	int sleep_flags = CAN_SLEEP;
841 
842 	if (curthread->td_no_sleeping != 0)
843 		sleep_flags = NO_SLEEP;
844 
845 	/* Step 1 */
846 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
847 
848 	/* Step 2 */
849 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
850 		return (EBUSY);
851 
852 	/* Step 3
853 	 * Announce that a message is coming through the doorbell.  Messages
854 	 * are pushed at 32bit words, so round up if needed.
855 	 */
856 	count = (req_sz + 3) / 4;
857 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
858 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
859 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
860 
861 	/* Step 4 */
862 	if (mps_wait_db_int(sc) ||
863 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
864 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
865 		return (ENXIO);
866 	}
867 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
868 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
869 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
870 		return (ENXIO);
871 	}
872 
873 	/* Step 5 */
874 	/* Clock out the message data synchronously in 32-bit dwords*/
875 	data32 = (uint32_t *)req;
876 	for (i = 0; i < count; i++) {
877 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
878 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
879 			mps_dprint(sc, MPS_FAULT,
880 			    "Timeout while writing doorbell\n");
881 			return (ENXIO);
882 		}
883 	}
884 
885 	/* Step 6 */
886 	/* Clock in the reply in 16-bit words.  The total length of the
887 	 * message is always in the 4th byte, so clock out the first 2 words
888 	 * manually, then loop the rest.
889 	 */
890 	data16 = (uint16_t *)reply;
891 	if (mps_wait_db_int(sc) != 0) {
892 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
893 		return (ENXIO);
894 	}
895 	data16[0] =
896 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
897 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
898 	if (mps_wait_db_int(sc) != 0) {
899 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
900 		return (ENXIO);
901 	}
902 	data16[1] =
903 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
904 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
905 
906 	/* Number of 32bit words in the message */
907 	ioc_sz = reply->MsgLength;
908 
909 	/*
910 	 * Figure out how many 16bit words to clock in without overrunning.
911 	 * The precision loss with dividing reply_sz can safely be
912 	 * ignored because the messages can only be multiples of 32bits.
913 	 */
914 	residual = 0;
915 	count = MIN((reply_sz / 4), ioc_sz) * 2;
916 	if (count < ioc_sz * 2) {
917 		residual = ioc_sz * 2 - count;
918 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
919 		    "residual message words\n", residual);
920 	}
921 
922 	for (i = 2; i < count; i++) {
923 		if (mps_wait_db_int(sc) != 0) {
924 			mps_dprint(sc, MPS_FAULT,
925 			    "Timeout reading doorbell %d\n", i);
926 			return (ENXIO);
927 		}
928 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
929 		    MPI2_DOORBELL_DATA_MASK;
930 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
931 	}
932 
933 	/*
934 	 * Pull out residual words that won't fit into the provided buffer.
935 	 * This keeps the chip from hanging due to a driver programming
936 	 * error.
937 	 */
938 	while (residual--) {
939 		if (mps_wait_db_int(sc) != 0) {
940 			mps_dprint(sc, MPS_FAULT,
941 			    "Timeout reading doorbell\n");
942 			return (ENXIO);
943 		}
944 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
945 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
946 	}
947 
948 	/* Step 7 */
949 	if (mps_wait_db_int(sc) != 0) {
950 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
951 		return (ENXIO);
952 	}
953 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
954 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
955 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
956 
957 	return (0);
958 }
959 
960 static void
961 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
962 {
963 	reply_descriptor rd;
964 	MPS_FUNCTRACE(sc);
965 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
966 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
967 
968 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
969 		mtx_assert(&sc->mps_mtx, MA_OWNED);
970 
971 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
972 		sc->io_cmds_highwater++;
973 	rd.u.low = cm->cm_desc.Words.Low;
974 	rd.u.high = cm->cm_desc.Words.High;
975 	rd.word = htole64(rd.word);
976 	/* TODO-We may need to make below regwrite atomic */
977 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
978 	    rd.u.low);
979 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
980 	    rd.u.high);
981 }
982 
983 /*
984  * Just the FACTS, ma'am.
985  */
986 static int
987 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
988 {
989 	MPI2_DEFAULT_REPLY *reply;
990 	MPI2_IOC_FACTS_REQUEST request;
991 	int error, req_sz, reply_sz;
992 
993 	MPS_FUNCTRACE(sc);
994 
995 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
996 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
997 	reply = (MPI2_DEFAULT_REPLY *)facts;
998 
999 	bzero(&request, req_sz);
1000 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1001 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1002 
1003 	return (error);
1004 }
1005 
1006 static int
1007 mps_send_iocinit(struct mps_softc *sc)
1008 {
1009 	MPI2_IOC_INIT_REQUEST	init;
1010 	MPI2_DEFAULT_REPLY	reply;
1011 	int req_sz, reply_sz, error;
1012 	struct timeval now;
1013 	uint64_t time_in_msec;
1014 
1015 	MPS_FUNCTRACE(sc);
1016 
1017 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1018 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1019 	bzero(&init, req_sz);
1020 	bzero(&reply, reply_sz);
1021 
1022 	/*
1023 	 * Fill in the init block.  Note that most addresses are
1024 	 * deliberately in the lower 32bits of memory.  This is a micro-
1025 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1026 	 */
1027 	init.Function = MPI2_FUNCTION_IOC_INIT;
1028 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1029 	init.MsgVersion = htole16(MPI2_VERSION);
1030 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1031 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1032 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1033 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1034 	init.SenseBufferAddressHigh = 0;
1035 	init.SystemReplyAddressHigh = 0;
1036 	init.SystemRequestFrameBaseAddress.High = 0;
1037 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1038 	init.ReplyDescriptorPostQueueAddress.High = 0;
1039 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1040 	init.ReplyFreeQueueAddress.High = 0;
1041 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1042 	getmicrotime(&now);
1043 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1044 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1045 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1046 
1047 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1048 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1049 		error = ENXIO;
1050 
1051 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1052 	return (error);
1053 }
1054 
1055 void
1056 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1057 {
1058 	bus_addr_t *addr;
1059 
1060 	addr = arg;
1061 	*addr = segs[0].ds_addr;
1062 }
1063 
1064 static int
1065 mps_alloc_queues(struct mps_softc *sc)
1066 {
1067 	bus_addr_t queues_busaddr;
1068 	uint8_t *queues;
1069 	int qsize, fqsize, pqsize;
1070 
1071 	/*
1072 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1073 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1074 	 * This queue supplies fresh reply frames for the firmware to use.
1075 	 *
1076 	 * The reply descriptor post queue contains 8 byte entries in
1077 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1078 	 * contains filled-in reply frames sent from the firmware to the host.
1079 	 *
1080 	 * These two queues are allocated together for simplicity.
1081 	 */
1082 	sc->fqdepth = roundup2((sc->num_replies + 1), 16);
1083 	sc->pqdepth = roundup2((sc->num_replies + 1), 16);
1084 	fqsize= sc->fqdepth * 4;
1085 	pqsize = sc->pqdepth * 8;
1086 	qsize = fqsize + pqsize;
1087 
1088         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1089 				16, 0,			/* algnmnt, boundary */
1090 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1091 				BUS_SPACE_MAXADDR,	/* highaddr */
1092 				NULL, NULL,		/* filter, filterarg */
1093                                 qsize,			/* maxsize */
1094                                 1,			/* nsegments */
1095                                 qsize,			/* maxsegsize */
1096                                 0,			/* flags */
1097                                 NULL, NULL,		/* lockfunc, lockarg */
1098                                 &sc->queues_dmat)) {
1099 		device_printf(sc->mps_dev, "Cannot allocate queues DMA tag\n");
1100 		return (ENOMEM);
1101         }
1102         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1103 	    &sc->queues_map)) {
1104 		device_printf(sc->mps_dev, "Cannot allocate queues memory\n");
1105 		return (ENOMEM);
1106         }
1107         bzero(queues, qsize);
1108         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1109 	    mps_memaddr_cb, &queues_busaddr, 0);
1110 
1111 	sc->free_queue = (uint32_t *)queues;
1112 	sc->free_busaddr = queues_busaddr;
1113 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1114 	sc->post_busaddr = queues_busaddr + fqsize;
1115 
1116 	return (0);
1117 }
1118 
1119 static int
1120 mps_alloc_replies(struct mps_softc *sc)
1121 {
1122 	int rsize, num_replies;
1123 
1124 	/*
1125 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1126 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1127 	 * replies can be used at once.
1128 	 */
1129 	num_replies = max(sc->fqdepth, sc->num_replies);
1130 
1131 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1132         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1133 				4, 0,			/* algnmnt, boundary */
1134 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1135 				BUS_SPACE_MAXADDR,	/* highaddr */
1136 				NULL, NULL,		/* filter, filterarg */
1137                                 rsize,			/* maxsize */
1138                                 1,			/* nsegments */
1139                                 rsize,			/* maxsegsize */
1140                                 0,			/* flags */
1141                                 NULL, NULL,		/* lockfunc, lockarg */
1142                                 &sc->reply_dmat)) {
1143 		device_printf(sc->mps_dev, "Cannot allocate replies DMA tag\n");
1144 		return (ENOMEM);
1145         }
1146         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1147 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1148 		device_printf(sc->mps_dev, "Cannot allocate replies memory\n");
1149 		return (ENOMEM);
1150         }
1151         bzero(sc->reply_frames, rsize);
1152         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1153 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1154 
1155 	return (0);
1156 }
1157 
1158 static int
1159 mps_alloc_requests(struct mps_softc *sc)
1160 {
1161 	struct mps_command *cm;
1162 	struct mps_chain *chain;
1163 	int i, rsize, nsegs;
1164 
1165 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1166         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1167 				16, 0,			/* algnmnt, boundary */
1168 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1169 				BUS_SPACE_MAXADDR,	/* highaddr */
1170 				NULL, NULL,		/* filter, filterarg */
1171                                 rsize,			/* maxsize */
1172                                 1,			/* nsegments */
1173                                 rsize,			/* maxsegsize */
1174                                 0,			/* flags */
1175                                 NULL, NULL,		/* lockfunc, lockarg */
1176                                 &sc->req_dmat)) {
1177 		device_printf(sc->mps_dev, "Cannot allocate request DMA tag\n");
1178 		return (ENOMEM);
1179         }
1180         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1181 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1182 		device_printf(sc->mps_dev, "Cannot allocate request memory\n");
1183 		return (ENOMEM);
1184         }
1185         bzero(sc->req_frames, rsize);
1186         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1187 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1188 
1189 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
1190         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1191 				16, 0,			/* algnmnt, boundary */
1192 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1193 				BUS_SPACE_MAXADDR,	/* highaddr */
1194 				NULL, NULL,		/* filter, filterarg */
1195                                 rsize,			/* maxsize */
1196                                 1,			/* nsegments */
1197                                 rsize,			/* maxsegsize */
1198                                 0,			/* flags */
1199                                 NULL, NULL,		/* lockfunc, lockarg */
1200                                 &sc->chain_dmat)) {
1201 		device_printf(sc->mps_dev, "Cannot allocate chain DMA tag\n");
1202 		return (ENOMEM);
1203         }
1204         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1205 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1206 		device_printf(sc->mps_dev, "Cannot allocate chain memory\n");
1207 		return (ENOMEM);
1208         }
1209         bzero(sc->chain_frames, rsize);
1210         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1211 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
1212 
1213 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1214         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1215 				1, 0,			/* algnmnt, boundary */
1216 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1217 				BUS_SPACE_MAXADDR,	/* highaddr */
1218 				NULL, NULL,		/* filter, filterarg */
1219                                 rsize,			/* maxsize */
1220                                 1,			/* nsegments */
1221                                 rsize,			/* maxsegsize */
1222                                 0,			/* flags */
1223                                 NULL, NULL,		/* lockfunc, lockarg */
1224                                 &sc->sense_dmat)) {
1225 		device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
1226 		return (ENOMEM);
1227         }
1228         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1229 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1230 		device_printf(sc->mps_dev, "Cannot allocate sense memory\n");
1231 		return (ENOMEM);
1232         }
1233         bzero(sc->sense_frames, rsize);
1234         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1235 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1236 
1237 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
1238 	    M_WAITOK | M_ZERO);
1239 	if(!sc->chains) {
1240 		device_printf(sc->mps_dev,
1241 		"Cannot allocate chains memory %s %d\n",
1242 		 __func__, __LINE__);
1243 		return (ENOMEM);
1244 	}
1245 	for (i = 0; i < sc->max_chains; i++) {
1246 		chain = &sc->chains[i];
1247 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1248 		    i * sc->facts->IOCRequestFrameSize * 4);
1249 		chain->chain_busaddr = sc->chain_busaddr +
1250 		    i * sc->facts->IOCRequestFrameSize * 4;
1251 		mps_free_chain(sc, chain);
1252 		sc->chain_free_lowwater++;
1253 	}
1254 
1255 	/* XXX Need to pick a more precise value */
1256 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1257         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1258 				1, 0,			/* algnmnt, boundary */
1259 				BUS_SPACE_MAXADDR,	/* lowaddr */
1260 				BUS_SPACE_MAXADDR,	/* highaddr */
1261 				NULL, NULL,		/* filter, filterarg */
1262                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1263                                 nsegs,			/* nsegments */
1264                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1265                                 BUS_DMA_ALLOCNOW,	/* flags */
1266                                 busdma_lock_mutex,	/* lockfunc */
1267 				&sc->mps_mtx,		/* lockarg */
1268                                 &sc->buffer_dmat)) {
1269 		device_printf(sc->mps_dev, "Cannot allocate buffer DMA tag\n");
1270 		return (ENOMEM);
1271         }
1272 
1273 	/*
1274 	 * SMID 0 cannot be used as a free command per the firmware spec.
1275 	 * Just drop that command instead of risking accounting bugs.
1276 	 */
1277 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1278 	    M_MPT2, M_WAITOK | M_ZERO);
1279 	if(!sc->commands) {
1280 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1281 		 __func__, __LINE__);
1282 		return (ENOMEM);
1283 	}
1284 	for (i = 1; i < sc->num_reqs; i++) {
1285 		cm = &sc->commands[i];
1286 		cm->cm_req = sc->req_frames +
1287 		    i * sc->facts->IOCRequestFrameSize * 4;
1288 		cm->cm_req_busaddr = sc->req_busaddr +
1289 		    i * sc->facts->IOCRequestFrameSize * 4;
1290 		cm->cm_sense = &sc->sense_frames[i];
1291 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1292 		cm->cm_desc.Default.SMID = i;
1293 		cm->cm_sc = sc;
1294 		TAILQ_INIT(&cm->cm_chain_list);
1295 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1296 
1297 		/* XXX Is a failure here a critical problem? */
1298 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1299 			if (i <= sc->facts->HighPriorityCredit)
1300 				mps_free_high_priority_command(sc, cm);
1301 			else
1302 				mps_free_command(sc, cm);
1303 		else {
1304 			panic("failed to allocate command %d\n", i);
1305 			sc->num_reqs = i;
1306 			break;
1307 		}
1308 	}
1309 
1310 	return (0);
1311 }
1312 
1313 static int
1314 mps_init_queues(struct mps_softc *sc)
1315 {
1316 	int i;
1317 
1318 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1319 
1320 	/*
1321 	 * According to the spec, we need to use one less reply than we
1322 	 * have space for on the queue.  So sc->num_replies (the number we
1323 	 * use) should be less than sc->fqdepth (allocated size).
1324 	 */
1325 	if (sc->num_replies >= sc->fqdepth)
1326 		return (EINVAL);
1327 
1328 	/*
1329 	 * Initialize all of the free queue entries.
1330 	 */
1331 	for (i = 0; i < sc->fqdepth; i++)
1332 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
1333 	sc->replyfreeindex = sc->num_replies;
1334 
1335 	return (0);
1336 }
1337 
1338 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1339  * Next are the global settings, if they exist.  Highest are the per-unit
1340  * settings, if they exist.
1341  */
1342 static void
1343 mps_get_tunables(struct mps_softc *sc)
1344 {
1345 	char tmpstr[80];
1346 
1347 	/* XXX default to some debugging for now */
1348 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1349 	sc->disable_msix = 0;
1350 	sc->disable_msi = 0;
1351 	sc->max_chains = MPS_CHAIN_FRAMES;
1352 
1353 	/*
1354 	 * Grab the global variables.
1355 	 */
1356 	TUNABLE_INT_FETCH("hw.mps.debug_level", &sc->mps_debug);
1357 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1358 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1359 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1360 
1361 	/* Grab the unit-instance variables */
1362 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1363 	    device_get_unit(sc->mps_dev));
1364 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
1365 
1366 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1367 	    device_get_unit(sc->mps_dev));
1368 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1369 
1370 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1371 	    device_get_unit(sc->mps_dev));
1372 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1373 
1374 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1375 	    device_get_unit(sc->mps_dev));
1376 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1377 
1378 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1379 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1380 	    device_get_unit(sc->mps_dev));
1381 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1382 }
1383 
1384 static void
1385 mps_setup_sysctl(struct mps_softc *sc)
1386 {
1387 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1388 	struct sysctl_oid	*sysctl_tree = NULL;
1389 	char tmpstr[80], tmpstr2[80];
1390 
1391 	/*
1392 	 * Setup the sysctl variable so the user can change the debug level
1393 	 * on the fly.
1394 	 */
1395 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1396 	    device_get_unit(sc->mps_dev));
1397 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1398 
1399 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1400 	if (sysctl_ctx != NULL)
1401 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1402 
1403 	if (sysctl_tree == NULL) {
1404 		sysctl_ctx_init(&sc->sysctl_ctx);
1405 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1406 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1407 		    CTLFLAG_RD, 0, tmpstr);
1408 		if (sc->sysctl_tree == NULL)
1409 			return;
1410 		sysctl_ctx = &sc->sysctl_ctx;
1411 		sysctl_tree = sc->sysctl_tree;
1412 	}
1413 
1414 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1415 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
1416 	    "mps debug level");
1417 
1418 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1419 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1420 	    "Disable the use of MSI-X interrupts");
1421 
1422 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1423 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1424 	    "Disable the use of MSI interrupts");
1425 
1426 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1427 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1428 	    strlen(sc->fw_version), "firmware version");
1429 
1430 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1431 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1432 	    strlen(MPS_DRIVER_VERSION), "driver version");
1433 
1434 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1435 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1436 	    &sc->io_cmds_active, 0, "number of currently active commands");
1437 
1438 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1439 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1440 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1441 
1442 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1443 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1444 	    &sc->chain_free, 0, "number of free chain elements");
1445 
1446 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1447 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1448 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1449 
1450 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1451 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1452 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1453 
1454 #if __FreeBSD_version >= 900030
1455 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1456 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1457 	    &sc->chain_alloc_fail, "chain allocation failures");
1458 #endif //FreeBSD_version >= 900030
1459 }
1460 
1461 int
1462 mps_attach(struct mps_softc *sc)
1463 {
1464 	int error;
1465 
1466 	mps_get_tunables(sc);
1467 
1468 	MPS_FUNCTRACE(sc);
1469 
1470 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1471 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1472 	TAILQ_INIT(&sc->event_list);
1473 	timevalclear(&sc->lastfail);
1474 
1475 	if ((error = mps_transition_ready(sc)) != 0) {
1476 		mps_printf(sc, "%s failed to transition ready\n", __func__);
1477 		return (error);
1478 	}
1479 
1480 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1481 	    M_ZERO|M_NOWAIT);
1482 	if(!sc->facts) {
1483 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1484 		 __func__, __LINE__);
1485 		return (ENOMEM);
1486 	}
1487 
1488 	/*
1489 	 * Get IOC Facts and allocate all structures based on this information.
1490 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1491 	 * Facts. If relevant values have changed in IOC Facts, this function
1492 	 * will free all of the memory based on IOC Facts and reallocate that
1493 	 * memory.  If this fails, any allocated memory should already be freed.
1494 	 */
1495 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1496 		mps_dprint(sc, MPS_FAULT, "%s IOC Facts based allocation "
1497 		    "failed with error %d\n", __func__, error);
1498 		return (error);
1499 	}
1500 
1501 	/* Start the periodic watchdog check on the IOC Doorbell */
1502 	mps_periodic(sc);
1503 
1504 	/*
1505 	 * The portenable will kick off discovery events that will drive the
1506 	 * rest of the initialization process.  The CAM/SAS module will
1507 	 * hold up the boot sequence until discovery is complete.
1508 	 */
1509 	sc->mps_ich.ich_func = mps_startup;
1510 	sc->mps_ich.ich_arg = sc;
1511 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1512 		mps_dprint(sc, MPS_ERROR, "Cannot establish MPS config hook\n");
1513 		error = EINVAL;
1514 	}
1515 
1516 	/*
1517 	 * Allow IR to shutdown gracefully when shutdown occurs.
1518 	 */
1519 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1520 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1521 
1522 	if (sc->shutdown_eh == NULL)
1523 		mps_dprint(sc, MPS_ERROR, "shutdown event registration "
1524 		    "failed\n");
1525 
1526 	mps_setup_sysctl(sc);
1527 
1528 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1529 
1530 	return (error);
1531 }
1532 
1533 /* Run through any late-start handlers. */
1534 static void
1535 mps_startup(void *arg)
1536 {
1537 	struct mps_softc *sc;
1538 
1539 	sc = (struct mps_softc *)arg;
1540 
1541 	mps_lock(sc);
1542 	mps_unmask_intr(sc);
1543 
1544 	/* initialize device mapping tables */
1545 	mps_base_static_config_pages(sc);
1546 	mps_mapping_initialize(sc);
1547 	mpssas_startup(sc);
1548 	mps_unlock(sc);
1549 }
1550 
1551 /* Periodic watchdog.  Is called with the driver lock already held. */
1552 static void
1553 mps_periodic(void *arg)
1554 {
1555 	struct mps_softc *sc;
1556 	uint32_t db;
1557 
1558 	sc = (struct mps_softc *)arg;
1559 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1560 		return;
1561 
1562 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1563 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1564 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1565 		mps_reinit(sc);
1566 	}
1567 
1568 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1569 }
1570 
1571 static void
1572 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1573     MPI2_EVENT_NOTIFICATION_REPLY *event)
1574 {
1575 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1576 
1577 	mps_print_event(sc, event);
1578 
1579 	switch (event->Event) {
1580 	case MPI2_EVENT_LOG_DATA:
1581 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1582 		if (sc->mps_debug & MPS_EVENT)
1583 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
1584 		break;
1585 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1586 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1587 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1588 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1589 		     entry->LogSequence);
1590 		break;
1591 	default:
1592 		break;
1593 	}
1594 	return;
1595 }
1596 
1597 static int
1598 mps_attach_log(struct mps_softc *sc)
1599 {
1600 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1601 
1602 	bzero(events, 16);
1603 	setbit(events, MPI2_EVENT_LOG_DATA);
1604 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1605 
1606 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1607 	    &sc->mps_log_eh);
1608 
1609 	return (0);
1610 }
1611 
1612 static int
1613 mps_detach_log(struct mps_softc *sc)
1614 {
1615 
1616 	if (sc->mps_log_eh != NULL)
1617 		mps_deregister_events(sc, sc->mps_log_eh);
1618 	return (0);
1619 }
1620 
1621 /*
1622  * Free all of the driver resources and detach submodules.  Should be called
1623  * without the lock held.
1624  */
1625 int
1626 mps_free(struct mps_softc *sc)
1627 {
1628 	int error;
1629 
1630 	/* Turn off the watchdog */
1631 	mps_lock(sc);
1632 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1633 	mps_unlock(sc);
1634 	/* Lock must not be held for this */
1635 	callout_drain(&sc->periodic);
1636 
1637 	if (((error = mps_detach_log(sc)) != 0) ||
1638 	    ((error = mps_detach_sas(sc)) != 0))
1639 		return (error);
1640 
1641 	mps_detach_user(sc);
1642 
1643 	/* Put the IOC back in the READY state. */
1644 	mps_lock(sc);
1645 	if ((error = mps_transition_ready(sc)) != 0) {
1646 		mps_unlock(sc);
1647 		return (error);
1648 	}
1649 	mps_unlock(sc);
1650 
1651 	if (sc->facts != NULL)
1652 		free(sc->facts, M_MPT2);
1653 
1654 	/*
1655 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
1656 	 * to free these buffers too.
1657 	 */
1658 	mps_iocfacts_free(sc);
1659 
1660 	if (sc->sysctl_tree != NULL)
1661 		sysctl_ctx_free(&sc->sysctl_ctx);
1662 
1663 	/* Deregister the shutdown function */
1664 	if (sc->shutdown_eh != NULL)
1665 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1666 
1667 	mtx_destroy(&sc->mps_mtx);
1668 
1669 	return (0);
1670 }
1671 
1672 static __inline void
1673 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
1674 {
1675 	MPS_FUNCTRACE(sc);
1676 
1677 	if (cm == NULL) {
1678 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
1679 		return;
1680 	}
1681 
1682 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1683 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1684 
1685 	if (cm->cm_complete != NULL) {
1686 		mps_dprint(sc, MPS_TRACE,
1687 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
1688 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
1689 			   cm->cm_reply);
1690 		cm->cm_complete(sc, cm);
1691 	}
1692 
1693 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1694 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
1695 		wakeup(cm);
1696 	}
1697 
1698 	if (cm->cm_sc->io_cmds_active != 0) {
1699 		cm->cm_sc->io_cmds_active--;
1700 	} else {
1701 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
1702 		    "out of sync - resynching to 0\n");
1703 	}
1704 }
1705 
1706 
1707 static void
1708 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
1709 {
1710 	union loginfo_type {
1711 		u32     loginfo;
1712 		struct {
1713 			u32     subcode:16;
1714 			u32     code:8;
1715 			u32     originator:4;
1716 			u32     bus_type:4;
1717 		} dw;
1718 	};
1719 	union loginfo_type sas_loginfo;
1720 	char *originator_str = NULL;
1721 
1722 	sas_loginfo.loginfo = log_info;
1723 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1724 		return;
1725 
1726 	/* each nexus loss loginfo */
1727 	if (log_info == 0x31170000)
1728 		return;
1729 
1730 	/* eat the loginfos associated with task aborts */
1731 	if ((log_info == 30050000 || log_info ==
1732 	    0x31140000 || log_info == 0x31130000))
1733 		return;
1734 
1735 	switch (sas_loginfo.dw.originator) {
1736 	case 0:
1737 		originator_str = "IOP";
1738 		break;
1739 	case 1:
1740 		originator_str = "PL";
1741 		break;
1742 	case 2:
1743 		originator_str = "IR";
1744 		break;
1745 }
1746 
1747 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
1748 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
1749 	originator_str, sas_loginfo.dw.code,
1750 	sas_loginfo.dw.subcode);
1751 }
1752 
1753 static void
1754 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
1755 {
1756 	MPI2DefaultReply_t *mpi_reply;
1757 	u16 sc_status;
1758 
1759 	mpi_reply = (MPI2DefaultReply_t*)reply;
1760 	sc_status = le16toh(mpi_reply->IOCStatus);
1761 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1762 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1763 }
1764 void
1765 mps_intr(void *data)
1766 {
1767 	struct mps_softc *sc;
1768 	uint32_t status;
1769 
1770 	sc = (struct mps_softc *)data;
1771 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1772 
1773 	/*
1774 	 * Check interrupt status register to flush the bus.  This is
1775 	 * needed for both INTx interrupts and driver-driven polling
1776 	 */
1777 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1778 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1779 		return;
1780 
1781 	mps_lock(sc);
1782 	mps_intr_locked(data);
1783 	mps_unlock(sc);
1784 	return;
1785 }
1786 
1787 /*
1788  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1789  * chip.  Hopefully this theory is correct.
1790  */
1791 void
1792 mps_intr_msi(void *data)
1793 {
1794 	struct mps_softc *sc;
1795 
1796 	sc = (struct mps_softc *)data;
1797 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1798 	mps_lock(sc);
1799 	mps_intr_locked(data);
1800 	mps_unlock(sc);
1801 	return;
1802 }
1803 
1804 /*
1805  * The locking is overly broad and simplistic, but easy to deal with for now.
1806  */
1807 void
1808 mps_intr_locked(void *data)
1809 {
1810 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1811 	struct mps_softc *sc;
1812 	struct mps_command *cm = NULL;
1813 	uint8_t flags;
1814 	u_int pq;
1815 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
1816 	mps_fw_diagnostic_buffer_t *pBuffer;
1817 
1818 	sc = (struct mps_softc *)data;
1819 
1820 	pq = sc->replypostindex;
1821 	mps_dprint(sc, MPS_TRACE,
1822 	    "%s sc %p starting with replypostindex %u\n",
1823 	    __func__, sc, sc->replypostindex);
1824 
1825 	for ( ;; ) {
1826 		cm = NULL;
1827 		desc = &sc->post_queue[sc->replypostindex];
1828 		flags = desc->Default.ReplyFlags &
1829 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1830 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
1831 		 || (le32toh(desc->Words.High) == 0xffffffff))
1832 			break;
1833 
1834 		/* increment the replypostindex now, so that event handlers
1835 		 * and cm completion handlers which decide to do a diag
1836 		 * reset can zero it without it getting incremented again
1837 		 * afterwards, and we break out of this loop on the next
1838 		 * iteration since the reply post queue has been cleared to
1839 		 * 0xFF and all descriptors look unused (which they are).
1840 		 */
1841 		if (++sc->replypostindex >= sc->pqdepth)
1842 			sc->replypostindex = 0;
1843 
1844 		switch (flags) {
1845 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1846 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
1847 			cm->cm_reply = NULL;
1848 			break;
1849 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1850 		{
1851 			uint32_t baddr;
1852 			uint8_t *reply;
1853 
1854 			/*
1855 			 * Re-compose the reply address from the address
1856 			 * sent back from the chip.  The ReplyFrameAddress
1857 			 * is the lower 32 bits of the physical address of
1858 			 * particular reply frame.  Convert that address to
1859 			 * host format, and then use that to provide the
1860 			 * offset against the virtual address base
1861 			 * (sc->reply_frames).
1862 			 */
1863 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1864 			reply = sc->reply_frames +
1865 				(baddr - ((uint32_t)sc->reply_busaddr));
1866 			/*
1867 			 * Make sure the reply we got back is in a valid
1868 			 * range.  If not, go ahead and panic here, since
1869 			 * we'll probably panic as soon as we deference the
1870 			 * reply pointer anyway.
1871 			 */
1872 			if ((reply < sc->reply_frames)
1873 			 || (reply > (sc->reply_frames +
1874 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1875 				printf("%s: WARNING: reply %p out of range!\n",
1876 				       __func__, reply);
1877 				printf("%s: reply_frames %p, fqdepth %d, "
1878 				       "frame size %d\n", __func__,
1879 				       sc->reply_frames, sc->fqdepth,
1880 				       sc->facts->ReplyFrameSize * 4);
1881 				printf("%s: baddr %#x,\n", __func__, baddr);
1882 				/* LSI-TODO. See Linux Code. Need Gracefull exit*/
1883 				panic("Reply address out of range");
1884 			}
1885 			if (le16toh(desc->AddressReply.SMID) == 0) {
1886 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
1887 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
1888 					/*
1889 					 * If SMID is 0 for Diag Buffer Post,
1890 					 * this implies that the reply is due to
1891 					 * a release function with a status that
1892 					 * the buffer has been released.  Set
1893 					 * the buffer flags accordingly.
1894 					 */
1895 					rel_rep =
1896 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
1897 					if (le16toh(rel_rep->IOCStatus) ==
1898 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
1899 					    {
1900 						pBuffer =
1901 						    &sc->fw_diag_buffer_list[
1902 						    rel_rep->BufferType];
1903 						pBuffer->valid_data = TRUE;
1904 						pBuffer->owned_by_firmware =
1905 						    FALSE;
1906 						pBuffer->immediate = FALSE;
1907 					}
1908 				} else
1909 					mps_dispatch_event(sc, baddr,
1910 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
1911 					    reply);
1912 			} else {
1913 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
1914 				cm->cm_reply = reply;
1915 				cm->cm_reply_data =
1916 				    le32toh(desc->AddressReply.ReplyFrameAddress);
1917 			}
1918 			break;
1919 		}
1920 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
1921 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
1922 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
1923 		default:
1924 			/* Unhandled */
1925 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
1926 			    desc->Default.ReplyFlags);
1927 			cm = NULL;
1928 			break;
1929 		}
1930 
1931 
1932 		if (cm != NULL) {
1933 			// Print Error reply frame
1934 			if (cm->cm_reply)
1935 				mps_display_reply_info(sc,cm->cm_reply);
1936 			mps_complete_command(sc, cm);
1937 		}
1938 
1939 		desc->Words.Low = 0xffffffff;
1940 		desc->Words.High = 0xffffffff;
1941 	}
1942 
1943 	if (pq != sc->replypostindex) {
1944 		mps_dprint(sc, MPS_TRACE,
1945 		    "%s sc %p writing postindex %d\n",
1946 		    __func__, sc, sc->replypostindex);
1947 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
1948 	}
1949 
1950 	return;
1951 }
1952 
1953 static void
1954 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
1955     MPI2_EVENT_NOTIFICATION_REPLY *reply)
1956 {
1957 	struct mps_event_handle *eh;
1958 	int event, handled = 0;
1959 
1960 	event = le16toh(reply->Event);
1961 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1962 		if (isset(eh->mask, event)) {
1963 			eh->callback(sc, data, reply);
1964 			handled++;
1965 		}
1966 	}
1967 
1968 	if (handled == 0)
1969 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
1970 
1971 	/*
1972 	 * This is the only place that the event/reply should be freed.
1973 	 * Anything wanting to hold onto the event data should have
1974 	 * already copied it into their own storage.
1975 	 */
1976 	mps_free_reply(sc, data);
1977 }
1978 
1979 static void
1980 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
1981 {
1982 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1983 
1984 	if (cm->cm_reply)
1985 		mps_print_event(sc,
1986 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
1987 
1988 	mps_free_command(sc, cm);
1989 
1990 	/* next, send a port enable */
1991 	mpssas_startup(sc);
1992 }
1993 
1994 /*
1995  * For both register_events and update_events, the caller supplies a bitmap
1996  * of events that it _wants_.  These functions then turn that into a bitmask
1997  * suitable for the controller.
1998  */
1999 int
2000 mps_register_events(struct mps_softc *sc, u32 *mask,
2001     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2002 {
2003 	struct mps_event_handle *eh;
2004 	int error = 0;
2005 
2006 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2007 	if(!eh) {
2008 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
2009 		 __func__, __LINE__);
2010 		return (ENOMEM);
2011 	}
2012 	eh->callback = cb;
2013 	eh->data = data;
2014 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2015 	if (mask != NULL)
2016 		error = mps_update_events(sc, eh, mask);
2017 	*handle = eh;
2018 
2019 	return (error);
2020 }
2021 
2022 int
2023 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2024     u32 *mask)
2025 {
2026 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2027 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
2028 	struct mps_command *cm;
2029 	int error, i;
2030 
2031 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2032 
2033 	if ((mask != NULL) && (handle != NULL))
2034 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2035 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2036 
2037 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2038 		sc->event_mask[i] = -1;
2039 
2040 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2041 		sc->event_mask[i] &= ~handle->mask[i];
2042 
2043 
2044 	if ((cm = mps_alloc_command(sc)) == NULL)
2045 		return (EBUSY);
2046 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2047 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2048 	evtreq->MsgFlags = 0;
2049 	evtreq->SASBroadcastPrimitiveMasks = 0;
2050 #ifdef MPS_DEBUG_ALL_EVENTS
2051 	{
2052 		u_char fullmask[16];
2053 		memset(fullmask, 0x00, 16);
2054 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2055 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2056 	}
2057 #else
2058         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2059                 evtreq->EventMasks[i] =
2060                     htole32(sc->event_mask[i]);
2061 #endif
2062 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2063 	cm->cm_data = NULL;
2064 
2065 	error = mps_request_polled(sc, cm);
2066 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2067 	if ((reply == NULL) ||
2068 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2069 		error = ENXIO;
2070 	mps_print_event(sc, reply);
2071 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2072 
2073 	mps_free_command(sc, cm);
2074 	return (error);
2075 }
2076 
2077 static int
2078 mps_reregister_events(struct mps_softc *sc)
2079 {
2080 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2081 	struct mps_command *cm;
2082 	struct mps_event_handle *eh;
2083 	int error, i;
2084 
2085 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2086 
2087 	/* first, reregister events */
2088 
2089 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2090 		sc->event_mask[i] = -1;
2091 
2092 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2093 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2094 			sc->event_mask[i] &= ~eh->mask[i];
2095 	}
2096 
2097 	if ((cm = mps_alloc_command(sc)) == NULL)
2098 		return (EBUSY);
2099 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2100 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2101 	evtreq->MsgFlags = 0;
2102 	evtreq->SASBroadcastPrimitiveMasks = 0;
2103 #ifdef MPS_DEBUG_ALL_EVENTS
2104 	{
2105 		u_char fullmask[16];
2106 		memset(fullmask, 0x00, 16);
2107 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2108 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2109 	}
2110 #else
2111         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2112                 evtreq->EventMasks[i] =
2113                     htole32(sc->event_mask[i]);
2114 #endif
2115 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2116 	cm->cm_data = NULL;
2117 	cm->cm_complete = mps_reregister_events_complete;
2118 
2119 	error = mps_map_command(sc, cm);
2120 
2121 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2122 	    error);
2123 	return (error);
2124 }
2125 
2126 void
2127 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2128 {
2129 
2130 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2131 	free(handle, M_MPT2);
2132 }
2133 
2134 /*
2135  * Add a chain element as the next SGE for the specified command.
2136  * Reset cm_sge and cm_sgesize to indicate all the available space.
2137  */
2138 static int
2139 mps_add_chain(struct mps_command *cm)
2140 {
2141 	MPI2_SGE_CHAIN32 *sgc;
2142 	struct mps_chain *chain;
2143 	int space;
2144 
2145 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2146 		panic("MPS: Need SGE Error Code\n");
2147 
2148 	chain = mps_alloc_chain(cm->cm_sc);
2149 	if (chain == NULL)
2150 		return (ENOBUFS);
2151 
2152 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
2153 
2154 	/*
2155 	 * Note: a double-linked list is used to make it easier to
2156 	 * walk for debugging.
2157 	 */
2158 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2159 
2160 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2161 	sgc->Length = htole16(space);
2162 	sgc->NextChainOffset = 0;
2163 	/* TODO Looks like bug in Setting sgc->Flags.
2164 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2165 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2166 	 *	This is fine.. because we are not using simple element. In case of
2167 	 *	MPI2_SGE_CHAIN32, we have seperate Length and Flags feild.
2168  	 */
2169 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2170 	sgc->Address = htole32(chain->chain_busaddr);
2171 
2172 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2173 	cm->cm_sglsize = space;
2174 	return (0);
2175 }
2176 
2177 /*
2178  * Add one scatter-gather element (chain, simple, transaction context)
2179  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2180  * cm_sge as the remaining size and pointer to the next SGE to fill
2181  * in, respectively.
2182  */
2183 int
2184 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2185 {
2186 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2187 	MPI2_SGE_SIMPLE64 *sge = sgep;
2188 	int error, type;
2189 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2190 
2191 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2192 
2193 #ifdef INVARIANTS
2194 	switch (type) {
2195 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2196 		if (len != tc->DetailsLength + 4)
2197 			panic("TC %p length %u or %zu?", tc,
2198 			    tc->DetailsLength + 4, len);
2199 		}
2200 		break;
2201 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2202 		/* Driver only uses 32-bit chain elements */
2203 		if (len != MPS_SGC_SIZE)
2204 			panic("CHAIN %p length %u or %zu?", sgep,
2205 			    MPS_SGC_SIZE, len);
2206 		break;
2207 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2208 		/* Driver only uses 64-bit SGE simple elements */
2209 		if (len != MPS_SGE64_SIZE)
2210 			panic("SGE simple %p length %u or %zu?", sge,
2211 			    MPS_SGE64_SIZE, len);
2212 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2213 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2214 			panic("SGE simple %p not marked 64-bit?", sge);
2215 
2216 		break;
2217 	default:
2218 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2219 	}
2220 #endif
2221 
2222 	/*
2223 	 * case 1: 1 more segment, enough room for it
2224 	 * case 2: 2 more segments, enough room for both
2225 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2226 	 * case 4: >=1 more segment, enough room for only a chain
2227 	 * case 5: >=1 more segment, no room for anything (error)
2228          */
2229 
2230 	/*
2231 	 * There should be room for at least a chain element, or this
2232 	 * code is buggy.  Case (5).
2233 	 */
2234 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2235 		panic("MPS: Need SGE Error Code\n");
2236 
2237 	if (segsleft >= 2 &&
2238 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2239 		/*
2240 		 * There are 2 or more segments left to add, and only
2241 		 * enough room for 1 and a chain.  Case (3).
2242 		 *
2243 		 * Mark as last element in this chain if necessary.
2244 		 */
2245 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2246 			sge->FlagsLength |= htole32(
2247 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2248 		}
2249 
2250 		/*
2251 		 * Add the item then a chain.  Do the chain now,
2252 		 * rather than on the next iteration, to simplify
2253 		 * understanding the code.
2254 		 */
2255 		cm->cm_sglsize -= len;
2256 		bcopy(sgep, cm->cm_sge, len);
2257 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2258 		return (mps_add_chain(cm));
2259 	}
2260 
2261 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2262 		/*
2263 		 * 1 or more segment, enough room for only a chain.
2264 		 * Hope the previous element wasn't a Simple entry
2265 		 * that needed to be marked with
2266 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2267 		 */
2268 		if ((error = mps_add_chain(cm)) != 0)
2269 			return (error);
2270 	}
2271 
2272 #ifdef INVARIANTS
2273 	/* Case 1: 1 more segment, enough room for it. */
2274 	if (segsleft == 1 && cm->cm_sglsize < len)
2275 		panic("1 seg left and no room? %u versus %zu",
2276 		    cm->cm_sglsize, len);
2277 
2278 	/* Case 2: 2 more segments, enough room for both */
2279 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2280 		panic("2 segs left and no room? %u versus %zu",
2281 		    cm->cm_sglsize, len);
2282 #endif
2283 
2284 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2285 		/*
2286 		 * If this is a bi-directional request, need to account for that
2287 		 * here.  Save the pre-filled sge values.  These will be used
2288 		 * either for the 2nd SGL or for a single direction SGL.  If
2289 		 * cm_out_len is non-zero, this is a bi-directional request, so
2290 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2291 		 * fill in the IN SGL.  Note that at this time, when filling in
2292 		 * 2 SGL's for a bi-directional request, they both use the same
2293 		 * DMA buffer (same cm command).
2294 		 */
2295 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2296 		saved_address_low = sge->Address.Low;
2297 		saved_address_high = sge->Address.High;
2298 		if (cm->cm_out_len) {
2299 			sge->FlagsLength = htole32(cm->cm_out_len |
2300 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2301 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2302 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2303 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2304 			    MPI2_SGE_FLAGS_SHIFT));
2305 			cm->cm_sglsize -= len;
2306 			bcopy(sgep, cm->cm_sge, len);
2307 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2308 			    + len);
2309 		}
2310 		saved_buf_len |=
2311 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2312 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2313 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2314 		    MPI2_SGE_FLAGS_END_OF_LIST |
2315 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2316 		    MPI2_SGE_FLAGS_SHIFT);
2317 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2318 			saved_buf_len |=
2319 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2320 			    MPI2_SGE_FLAGS_SHIFT);
2321 		} else {
2322 			saved_buf_len |=
2323 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2324 			    MPI2_SGE_FLAGS_SHIFT);
2325 		}
2326 		sge->FlagsLength = htole32(saved_buf_len);
2327 		sge->Address.Low = saved_address_low;
2328 		sge->Address.High = saved_address_high;
2329 	}
2330 
2331 	cm->cm_sglsize -= len;
2332 	bcopy(sgep, cm->cm_sge, len);
2333 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2334 	return (0);
2335 }
2336 
2337 /*
2338  * Add one dma segment to the scatter-gather list for a command.
2339  */
2340 int
2341 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2342     int segsleft)
2343 {
2344 	MPI2_SGE_SIMPLE64 sge;
2345 
2346 	/*
2347 	 * This driver always uses 64-bit address elements for simplicity.
2348 	 */
2349 	bzero(&sge, sizeof(sge));
2350 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2351 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2352 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2353 	mps_from_u64(pa, &sge.Address);
2354 
2355 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2356 }
2357 
2358 static void
2359 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2360 {
2361 	struct mps_softc *sc;
2362 	struct mps_command *cm;
2363 	u_int i, dir, sflags;
2364 
2365 	cm = (struct mps_command *)arg;
2366 	sc = cm->cm_sc;
2367 
2368 	/*
2369 	 * In this case, just print out a warning and let the chip tell the
2370 	 * user they did the wrong thing.
2371 	 */
2372 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2373 		mps_dprint(sc, MPS_ERROR,
2374 			   "%s: warning: busdma returned %d segments, "
2375 			   "more than the %d allowed\n", __func__, nsegs,
2376 			   cm->cm_max_segs);
2377 	}
2378 
2379 	/*
2380 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2381 	 * here.  In that case, both direction flags will be set.
2382 	 */
2383 	sflags = 0;
2384 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2385 		/*
2386 		 * We have to add a special case for SMP passthrough, there
2387 		 * is no easy way to generically handle it.  The first
2388 		 * S/G element is used for the command (therefore the
2389 		 * direction bit needs to be set).  The second one is used
2390 		 * for the reply.  We'll leave it to the caller to make
2391 		 * sure we only have two buffers.
2392 		 */
2393 		/*
2394 		 * Even though the busdma man page says it doesn't make
2395 		 * sense to have both direction flags, it does in this case.
2396 		 * We have one s/g element being accessed in each direction.
2397 		 */
2398 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2399 
2400 		/*
2401 		 * Set the direction flag on the first buffer in the SMP
2402 		 * passthrough request.  We'll clear it for the second one.
2403 		 */
2404 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2405 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2406 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2407 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2408 		dir = BUS_DMASYNC_PREWRITE;
2409 	} else
2410 		dir = BUS_DMASYNC_PREREAD;
2411 
2412 	for (i = 0; i < nsegs; i++) {
2413 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2414 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2415 		}
2416 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2417 		    sflags, nsegs - i);
2418 		if (error != 0) {
2419 			/* Resource shortage, roll back! */
2420 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2421 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2422 				    "consider increasing hw.mps.max_chains.\n");
2423 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2424 			mps_complete_command(sc, cm);
2425 			return;
2426 		}
2427 	}
2428 
2429 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2430 	mps_enqueue_request(sc, cm);
2431 
2432 	return;
2433 }
2434 
2435 static void
2436 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2437 	     int error)
2438 {
2439 	mps_data_cb(arg, segs, nsegs, error);
2440 }
2441 
2442 /*
2443  * This is the routine to enqueue commands ansynchronously.
2444  * Note that the only error path here is from bus_dmamap_load(), which can
2445  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2446  * assumed that if you have a command in-hand, then you have enough credits
2447  * to use it.
2448  */
2449 int
2450 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2451 {
2452 	int error = 0;
2453 
2454 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2455 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2456 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2457 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2458 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2459 		    cm->cm_data, mps_data_cb, cm, 0);
2460 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2461 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2462 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2463 	} else {
2464 		/* Add a zero-length element as needed */
2465 		if (cm->cm_sge != NULL)
2466 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2467 		mps_enqueue_request(sc, cm);
2468 	}
2469 
2470 	return (error);
2471 }
2472 
2473 /*
2474  * This is the routine to enqueue commands synchronously.  An error of
2475  * EINPROGRESS from mps_map_command() is ignored since the command will
2476  * be executed and enqueued automatically.  Other errors come from msleep().
2477  */
2478 int
2479 mps_wait_command(struct mps_softc *sc, struct mps_command *cm, int timeout,
2480     int sleep_flag)
2481 {
2482 	int error, rc;
2483 	struct timeval cur_time, start_time;
2484 
2485 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2486 		return  EBUSY;
2487 
2488 	cm->cm_complete = NULL;
2489 	cm->cm_flags |= (MPS_CM_FLAGS_WAKEUP + MPS_CM_FLAGS_POLLED);
2490 	error = mps_map_command(sc, cm);
2491 	if ((error != 0) && (error != EINPROGRESS))
2492 		return (error);
2493 
2494 	// Check for context and wait for 50 mSec at a time until time has
2495 	// expired or the command has finished.  If msleep can't be used, need
2496 	// to poll.
2497 	if (curthread->td_no_sleeping != 0)
2498 		sleep_flag = NO_SLEEP;
2499 	getmicrotime(&start_time);
2500 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2501 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2502 	} else {
2503 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2504 			mps_intr_locked(sc);
2505 			if (sleep_flag == CAN_SLEEP)
2506 				pause("mpswait", hz/20);
2507 			else
2508 				DELAY(50000);
2509 
2510 			getmicrotime(&cur_time);
2511 			if ((cur_time.tv_sec - start_time.tv_sec) > timeout) {
2512 				error = EWOULDBLOCK;
2513 				break;
2514 			}
2515 		}
2516 	}
2517 
2518 	if (error == EWOULDBLOCK) {
2519 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s\n", __func__);
2520 		rc = mps_reinit(sc);
2521 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2522 		    "failed");
2523 		error = ETIMEDOUT;
2524 	}
2525 	return (error);
2526 }
2527 
2528 /*
2529  * This is the routine to enqueue a command synchonously and poll for
2530  * completion.  Its use should be rare.
2531  */
2532 int
2533 mps_request_polled(struct mps_softc *sc, struct mps_command *cm)
2534 {
2535 	int error, timeout = 0, rc;
2536 	struct timeval cur_time, start_time;
2537 
2538 	error = 0;
2539 
2540 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2541 	cm->cm_complete = NULL;
2542 	mps_map_command(sc, cm);
2543 
2544 	getmicrotime(&start_time);
2545 	while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2546 		mps_intr_locked(sc);
2547 
2548 		if (mtx_owned(&sc->mps_mtx))
2549 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
2550 			    "mpspoll", hz/20);
2551 		else
2552 			pause("mpsdiag", hz/20);
2553 
2554 		/*
2555 		 * Check for real-time timeout and fail if more than 60 seconds.
2556 		 */
2557 		getmicrotime(&cur_time);
2558 		timeout = cur_time.tv_sec - start_time.tv_sec;
2559 		if (timeout > 60) {
2560 			mps_dprint(sc, MPS_FAULT, "polling failed\n");
2561 			error = ETIMEDOUT;
2562 			break;
2563 		}
2564 	}
2565 
2566 	if (error) {
2567 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s\n", __func__);
2568 		rc = mps_reinit(sc);
2569 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
2570 		    "failed");
2571 	}
2572 
2573 	return (error);
2574 }
2575 
2576 /*
2577  * The MPT driver had a verbose interface for config pages.  In this driver,
2578  * reduce it to much simplier terms, similar to the Linux driver.
2579  */
2580 int
2581 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2582 {
2583 	MPI2_CONFIG_REQUEST *req;
2584 	struct mps_command *cm;
2585 	int error;
2586 
2587 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2588 		return (EBUSY);
2589 	}
2590 
2591 	cm = mps_alloc_command(sc);
2592 	if (cm == NULL) {
2593 		return (EBUSY);
2594 	}
2595 
2596 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2597 	req->Function = MPI2_FUNCTION_CONFIG;
2598 	req->Action = params->action;
2599 	req->SGLFlags = 0;
2600 	req->ChainOffset = 0;
2601 	req->PageAddress = params->page_address;
2602 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
2603 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2604 
2605 		hdr = &params->hdr.Ext;
2606 		req->ExtPageType = hdr->ExtPageType;
2607 		req->ExtPageLength = hdr->ExtPageLength;
2608 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2609 		req->Header.PageLength = 0; /* Must be set to zero */
2610 		req->Header.PageNumber = hdr->PageNumber;
2611 		req->Header.PageVersion = hdr->PageVersion;
2612 	} else {
2613 		MPI2_CONFIG_PAGE_HEADER *hdr;
2614 
2615 		hdr = &params->hdr.Struct;
2616 		req->Header.PageType = hdr->PageType;
2617 		req->Header.PageNumber = hdr->PageNumber;
2618 		req->Header.PageLength = hdr->PageLength;
2619 		req->Header.PageVersion = hdr->PageVersion;
2620 	}
2621 
2622 	cm->cm_data = params->buffer;
2623 	cm->cm_length = params->length;
2624 	cm->cm_sge = &req->PageBufferSGE;
2625 	cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2626 	cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2627 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2628 
2629 	cm->cm_complete_data = params;
2630 	if (params->callback != NULL) {
2631 		cm->cm_complete = mps_config_complete;
2632 		return (mps_map_command(sc, cm));
2633 	} else {
2634 		error = mps_wait_command(sc, cm, 0, CAN_SLEEP);
2635 		if (error) {
2636 			mps_dprint(sc, MPS_FAULT,
2637 			    "Error %d reading config page\n", error);
2638 			mps_free_command(sc, cm);
2639 			return (error);
2640 		}
2641 		mps_config_complete(sc, cm);
2642 	}
2643 
2644 	return (0);
2645 }
2646 
2647 int
2648 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2649 {
2650 	return (EINVAL);
2651 }
2652 
2653 static void
2654 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
2655 {
2656 	MPI2_CONFIG_REPLY *reply;
2657 	struct mps_config_params *params;
2658 
2659 	MPS_FUNCTRACE(sc);
2660 	params = cm->cm_complete_data;
2661 
2662 	if (cm->cm_data != NULL) {
2663 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2664 		    BUS_DMASYNC_POSTREAD);
2665 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2666 	}
2667 
2668 	/*
2669 	 * XXX KDM need to do more error recovery?  This results in the
2670 	 * device in question not getting probed.
2671 	 */
2672 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
2673 		params->status = MPI2_IOCSTATUS_BUSY;
2674 		goto done;
2675 	}
2676 
2677 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2678 	if (reply == NULL) {
2679 		params->status = MPI2_IOCSTATUS_BUSY;
2680 		goto done;
2681 	}
2682 	params->status = reply->IOCStatus;
2683 	if (params->hdr.Ext.ExtPageType != 0) {
2684 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2685 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2686 	} else {
2687 		params->hdr.Struct.PageType = reply->Header.PageType;
2688 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2689 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2690 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2691 	}
2692 
2693 done:
2694 	mps_free_command(sc, cm);
2695 	if (params->callback != NULL)
2696 		params->callback(sc, params);
2697 
2698 	return;
2699 }
2700