xref: /freebsd/sys/dev/mps/mps.c (revision 118063fb70c3f74f16678dc7ddbc8528c483a528)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2009 Yahoo! Inc.
5  * Copyright (c) 2011-2015 LSI Corp.
6  * Copyright (c) 2013-2015 Avago Technologies
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
31  *
32  * $FreeBSD$
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 /* Communications core for Avago Technologies (LSI) MPT2 */
39 
40 /* TODO Move headers to mpsvar */
41 #include <sys/types.h>
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/selinfo.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/module.h>
49 #include <sys/bus.h>
50 #include <sys/conf.h>
51 #include <sys/bio.h>
52 #include <sys/malloc.h>
53 #include <sys/uio.h>
54 #include <sys/sysctl.h>
55 #include <sys/smp.h>
56 #include <sys/queue.h>
57 #include <sys/kthread.h>
58 #include <sys/taskqueue.h>
59 #include <sys/endian.h>
60 #include <sys/eventhandler.h>
61 #include <sys/sbuf.h>
62 
63 #include <machine/bus.h>
64 #include <machine/resource.h>
65 #include <sys/rman.h>
66 #include <sys/proc.h>
67 
68 #include <dev/pci/pcivar.h>
69 
70 #include <cam/cam.h>
71 #include <cam/scsi/scsi_all.h>
72 
73 #include <dev/mps/mpi/mpi2_type.h>
74 #include <dev/mps/mpi/mpi2.h>
75 #include <dev/mps/mpi/mpi2_ioc.h>
76 #include <dev/mps/mpi/mpi2_sas.h>
77 #include <dev/mps/mpi/mpi2_cnfg.h>
78 #include <dev/mps/mpi/mpi2_init.h>
79 #include <dev/mps/mpi/mpi2_tool.h>
80 #include <dev/mps/mps_ioctl.h>
81 #include <dev/mps/mpsvar.h>
82 #include <dev/mps/mps_table.h>
83 
84 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
85 static int mps_init_queues(struct mps_softc *sc);
86 static void mps_resize_queues(struct mps_softc *sc);
87 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
88 static int mps_transition_operational(struct mps_softc *sc);
89 static int mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching);
90 static void mps_iocfacts_free(struct mps_softc *sc);
91 static void mps_startup(void *arg);
92 static int mps_send_iocinit(struct mps_softc *sc);
93 static int mps_alloc_queues(struct mps_softc *sc);
94 static int mps_alloc_hw_queues(struct mps_softc *sc);
95 static int mps_alloc_replies(struct mps_softc *sc);
96 static int mps_alloc_requests(struct mps_softc *sc);
97 static int mps_attach_log(struct mps_softc *sc);
98 static __inline void mps_complete_command(struct mps_softc *sc,
99     struct mps_command *cm);
100 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
101     MPI2_EVENT_NOTIFICATION_REPLY *reply);
102 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
103 static void mps_periodic(void *);
104 static int mps_reregister_events(struct mps_softc *sc);
105 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
106 static int mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
107 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
108 static int mps_debug_sysctl(SYSCTL_HANDLER_ARGS);
109 static void mps_parse_debug(struct mps_softc *sc, char *list);
110 
111 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
112 
113 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
114 MALLOC_DECLARE(M_MPSUSER);
115 
116 /*
117  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
118  * any state and back to its initialization state machine.
119  */
120 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
121 
122 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
123  * Compiler only support unint64_t to be passed as argument.
124  * Otherwise it will throw below error
125  * "aggregate value used where an integer was expected"
126  */
127 
128 typedef union _reply_descriptor {
129         u64 word;
130         struct {
131                 u32 low;
132                 u32 high;
133         } u;
134 }reply_descriptor,address_descriptor;
135 
136 /* Rate limit chain-fail messages to 1 per minute */
137 static struct timeval mps_chainfail_interval = { 60, 0 };
138 
139 /*
140  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
141  * If this function is called from process context, it can sleep
142  * and there is no harm to sleep, in case if this fuction is called
143  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
144  * based on sleep flags driver will call either msleep, pause or DELAY.
145  * msleep and pause are of same variant, but pause is used when mps_mtx
146  * is not hold by driver.
147  *
148  */
149 static int
150 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
151 {
152 	uint32_t reg;
153 	int i, error, tries = 0;
154 	uint8_t first_wait_done = FALSE;
155 
156 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
157 
158 	/* Clear any pending interrupts */
159 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
160 
161 	/*
162 	 * Force NO_SLEEP for threads prohibited to sleep
163  	 * e.a Thread from interrupt handler are prohibited to sleep.
164  	 */
165 	if (curthread->td_no_sleeping != 0)
166 		sleep_flag = NO_SLEEP;
167 
168 	mps_dprint(sc, MPS_INIT, "sequence start, sleep_flag= %d\n", sleep_flag);
169 
170 	/* Push the magic sequence */
171 	error = ETIMEDOUT;
172 	while (tries++ < 20) {
173 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
174 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
175 			    mpt2_reset_magic[i]);
176 		/* wait 100 msec */
177 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
178 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
179 			    "mpsdiag", hz/10);
180 		else if (sleep_flag == CAN_SLEEP)
181 			pause("mpsdiag", hz/10);
182 		else
183 			DELAY(100 * 1000);
184 
185 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
186 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
187 			error = 0;
188 			break;
189 		}
190 	}
191 	if (error) {
192 		mps_dprint(sc, MPS_INIT, "sequence failed, error=%d, exit\n",
193 		    error);
194 		return (error);
195 	}
196 
197 	/* Send the actual reset.  XXX need to refresh the reg? */
198 	reg |= MPI2_DIAG_RESET_ADAPTER;
199 	mps_dprint(sc, MPS_INIT, "sequence success, sending reset, reg= 0x%x\n",
200 		reg);
201 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
202 
203 	/* Wait up to 300 seconds in 50ms intervals */
204 	error = ETIMEDOUT;
205 	for (i = 0; i < 6000; i++) {
206 		/*
207 		 * Wait 50 msec. If this is the first time through, wait 256
208 		 * msec to satisfy Diag Reset timing requirements.
209 		 */
210 		if (first_wait_done) {
211 			if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
212 				msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
213 				    "mpsdiag", hz/20);
214 			else if (sleep_flag == CAN_SLEEP)
215 				pause("mpsdiag", hz/20);
216 			else
217 				DELAY(50 * 1000);
218 		} else {
219 			DELAY(256 * 1000);
220 			first_wait_done = TRUE;
221 		}
222 		/*
223 		 * Check for the RESET_ADAPTER bit to be cleared first, then
224 		 * wait for the RESET state to be cleared, which takes a little
225 		 * longer.
226 		 */
227 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
228 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
229 			continue;
230 		}
231 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
232 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
233 			error = 0;
234 			break;
235 		}
236 	}
237 	if (error) {
238 		mps_dprint(sc, MPS_INIT, "reset failed, error= %d, exit\n",
239 		    error);
240 		return (error);
241 	}
242 
243 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
244 	mps_dprint(sc, MPS_INIT, "diag reset success, exit\n");
245 
246 	return (0);
247 }
248 
249 static int
250 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
251 {
252 	int error;
253 
254 	MPS_FUNCTRACE(sc);
255 
256 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
257 
258 	error = 0;
259 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
260 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
261 	    MPI2_DOORBELL_FUNCTION_SHIFT);
262 
263 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
264 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
265 		    "Doorbell handshake failed\n");
266 		error = ETIMEDOUT;
267 	}
268 
269 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
270 	return (error);
271 }
272 
273 static int
274 mps_transition_ready(struct mps_softc *sc)
275 {
276 	uint32_t reg, state;
277 	int error, tries = 0;
278 	int sleep_flags;
279 
280 	MPS_FUNCTRACE(sc);
281 	/* If we are in attach call, do not sleep */
282 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
283 					? CAN_SLEEP:NO_SLEEP;
284 	error = 0;
285 
286 	mps_dprint(sc, MPS_INIT, "%s entered, sleep_flags= %d\n",
287 	   __func__, sleep_flags);
288 
289 	while (tries++ < 1200) {
290 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
291 		mps_dprint(sc, MPS_INIT, "  Doorbell= 0x%x\n", reg);
292 
293 		/*
294 		 * Ensure the IOC is ready to talk.  If it's not, try
295 		 * resetting it.
296 		 */
297 		if (reg & MPI2_DOORBELL_USED) {
298 			mps_dprint(sc, MPS_INIT, "  Not ready, sending diag "
299 			    "reset\n");
300 			mps_diag_reset(sc, sleep_flags);
301 			DELAY(50000);
302 			continue;
303 		}
304 
305 		/* Is the adapter owned by another peer? */
306 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
307 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
308 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC is under the "
309 			    "control of another peer host, aborting "
310 			    "initialization.\n");
311 			error = ENXIO;
312 			break;
313 		}
314 
315 		state = reg & MPI2_IOC_STATE_MASK;
316 		if (state == MPI2_IOC_STATE_READY) {
317 			/* Ready to go! */
318 			error = 0;
319 			break;
320 		} else if (state == MPI2_IOC_STATE_FAULT) {
321 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC in fault "
322 			    "state 0x%x, resetting\n",
323 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
324 			mps_diag_reset(sc, sleep_flags);
325 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
326 			/* Need to take ownership */
327 			mps_message_unit_reset(sc, sleep_flags);
328 		} else if (state == MPI2_IOC_STATE_RESET) {
329 			/* Wait a bit, IOC might be in transition */
330 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
331 			    "IOC in unexpected reset state\n");
332 		} else {
333 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
334 			    "IOC in unknown state 0x%x\n", state);
335 			error = EINVAL;
336 			break;
337 		}
338 
339 		/* Wait 50ms for things to settle down. */
340 		DELAY(50000);
341 	}
342 
343 	if (error)
344 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
345 		    "Cannot transition IOC to ready\n");
346 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
347 
348 	return (error);
349 }
350 
351 static int
352 mps_transition_operational(struct mps_softc *sc)
353 {
354 	uint32_t reg, state;
355 	int error;
356 
357 	MPS_FUNCTRACE(sc);
358 
359 	error = 0;
360 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
361 	mps_dprint(sc, MPS_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
362 
363 	state = reg & MPI2_IOC_STATE_MASK;
364 	if (state != MPI2_IOC_STATE_READY) {
365 		mps_dprint(sc, MPS_INIT, "IOC not ready\n");
366 		if ((error = mps_transition_ready(sc)) != 0) {
367 			mps_dprint(sc, MPS_INIT|MPS_FAULT,
368 			    "failed to transition ready, exit\n");
369 			return (error);
370 		}
371 	}
372 
373 	error = mps_send_iocinit(sc);
374 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
375 
376 	return (error);
377 }
378 
379 static void
380 mps_resize_queues(struct mps_softc *sc)
381 {
382 	u_int reqcr, prireqcr, maxio, sges_per_frame;
383 
384 	/*
385 	 * Size the queues. Since the reply queues always need one free
386 	 * entry, we'll deduct one reply message here.  The LSI documents
387 	 * suggest instead to add a count to the request queue, but I think
388 	 * that it's better to deduct from reply queue.
389 	 */
390 	prireqcr = MAX(1, sc->max_prireqframes);
391 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
392 
393 	reqcr = MAX(2, sc->max_reqframes);
394 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
395 
396 	sc->num_reqs = prireqcr + reqcr;
397 	sc->num_prireqs = prireqcr;
398 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
399 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
400 
401 	/* Store the request frame size in bytes rather than as 32bit words */
402 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
403 
404 	/*
405 	 * Max IO Size is Page Size * the following:
406 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
407 	 * + 1 for no chain needed in last frame
408 	 *
409 	 * If user suggests a Max IO size to use, use the smaller of the
410 	 * user's value and the calculated value as long as the user's
411 	 * value is larger than 0. The user's value is in pages.
412 	 */
413 	sges_per_frame = sc->reqframesz / sizeof(MPI2_SGE_SIMPLE64) - 1;
414 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
415 
416 	/*
417 	 * If I/O size limitation requested, then use it and pass up to CAM.
418 	 * If not, use MAXPHYS as an optimization hint, but report HW limit.
419 	 */
420 	if (sc->max_io_pages > 0) {
421 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
422 		sc->maxio = maxio;
423 	} else {
424 		sc->maxio = maxio;
425 		maxio = min(maxio, MAXPHYS);
426 	}
427 
428 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
429 	    sges_per_frame * reqcr;
430 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
431 		sc->num_chains = sc->max_chains;
432 
433 	/*
434 	 * Figure out the number of MSIx-based queues.  If the firmware or
435 	 * user has done something crazy and not allowed enough credit for
436 	 * the queues to be useful then don't enable multi-queue.
437 	 */
438 	if (sc->facts->MaxMSIxVectors < 2)
439 		sc->msi_msgs = 1;
440 
441 	if (sc->msi_msgs > 1) {
442 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
443 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
444 		if (sc->num_reqs / sc->msi_msgs < 2)
445 			sc->msi_msgs = 1;
446 	}
447 
448 	mps_dprint(sc, MPS_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
449 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
450 }
451 
452 /*
453  * This is called during attach and when re-initializing due to a Diag Reset.
454  * IOC Facts is used to allocate many of the structures needed by the driver.
455  * If called from attach, de-allocation is not required because the driver has
456  * not allocated any structures yet, but if called from a Diag Reset, previously
457  * allocated structures based on IOC Facts will need to be freed and re-
458  * allocated bases on the latest IOC Facts.
459  */
460 static int
461 mps_iocfacts_allocate(struct mps_softc *sc, uint8_t attaching)
462 {
463 	int error;
464 	Mpi2IOCFactsReply_t saved_facts;
465 	uint8_t saved_mode, reallocating;
466 
467 	mps_dprint(sc, MPS_INIT|MPS_TRACE, "%s entered\n", __func__);
468 
469 	/* Save old IOC Facts and then only reallocate if Facts have changed */
470 	if (!attaching) {
471 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
472 	}
473 
474 	/*
475 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
476 	 * a re-initialization and only return the error if attaching so the OS
477 	 * can handle it.
478 	 */
479 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0) {
480 		if (attaching) {
481 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to get "
482 			    "IOC Facts with error %d, exit\n", error);
483 			return (error);
484 		} else {
485 			panic("%s failed to get IOC Facts with error %d\n",
486 			    __func__, error);
487 		}
488 	}
489 
490 	MPS_DPRINT_PAGE(sc, MPS_XINFO, iocfacts, sc->facts);
491 
492 	snprintf(sc->fw_version, sizeof(sc->fw_version),
493 	    "%02d.%02d.%02d.%02d",
494 	    sc->facts->FWVersion.Struct.Major,
495 	    sc->facts->FWVersion.Struct.Minor,
496 	    sc->facts->FWVersion.Struct.Unit,
497 	    sc->facts->FWVersion.Struct.Dev);
498 
499 	mps_dprint(sc, MPS_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
500 	    MPS_DRIVER_VERSION);
501 	mps_dprint(sc, MPS_INFO, "IOCCapabilities: %b\n",
502 	     sc->facts->IOCCapabilities,
503 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
504 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
505 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
506 
507 	/*
508 	 * If the chip doesn't support event replay then a hard reset will be
509 	 * required to trigger a full discovery.  Do the reset here then
510 	 * retransition to Ready.  A hard reset might have already been done,
511 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
512 	 * for a Diag Reset.
513 	 */
514 	if (attaching && ((sc->facts->IOCCapabilities &
515 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
516 		mps_dprint(sc, MPS_INIT, "No event replay, reseting\n");
517 		mps_diag_reset(sc, NO_SLEEP);
518 		if ((error = mps_transition_ready(sc)) != 0) {
519 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
520 			    "transition to ready with error %d, exit\n",
521 			    error);
522 			return (error);
523 		}
524 	}
525 
526 	/*
527 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
528 	 * changed from the previous IOC Facts, log a warning, but only if
529 	 * checking this after a Diag Reset and not during attach.
530 	 */
531 	saved_mode = sc->ir_firmware;
532 	if (sc->facts->IOCCapabilities &
533 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
534 		sc->ir_firmware = 1;
535 	if (!attaching) {
536 		if (sc->ir_firmware != saved_mode) {
537 			mps_dprint(sc, MPS_INIT|MPS_FAULT, "new IR/IT mode "
538 			    "in IOC Facts does not match previous mode\n");
539 		}
540 	}
541 
542 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
543 	reallocating = FALSE;
544 	sc->mps_flags &= ~MPS_FLAGS_REALLOCATED;
545 
546 	if ((!attaching) &&
547 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
548 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
549 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
550 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
551 	    (saved_facts.ProductID != sc->facts->ProductID) ||
552 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
553 	    (saved_facts.IOCRequestFrameSize !=
554 	    sc->facts->IOCRequestFrameSize) ||
555 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
556 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
557 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
558 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
559 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
560 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
561 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
562 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
563 	    (saved_facts.MaxPersistentEntries !=
564 	    sc->facts->MaxPersistentEntries))) {
565 		reallocating = TRUE;
566 
567 		/* Record that we reallocated everything */
568 		sc->mps_flags |= MPS_FLAGS_REALLOCATED;
569 	}
570 
571 	/*
572 	 * Some things should be done if attaching or re-allocating after a Diag
573 	 * Reset, but are not needed after a Diag Reset if the FW has not
574 	 * changed.
575 	 */
576 	if (attaching || reallocating) {
577 		/*
578 		 * Check if controller supports FW diag buffers and set flag to
579 		 * enable each type.
580 		 */
581 		if (sc->facts->IOCCapabilities &
582 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
583 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
584 			    enabled = TRUE;
585 		if (sc->facts->IOCCapabilities &
586 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
587 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
588 			    enabled = TRUE;
589 		if (sc->facts->IOCCapabilities &
590 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
591 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
592 			    enabled = TRUE;
593 
594 		/*
595 		 * Set flag if EEDP is supported and if TLR is supported.
596 		 */
597 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
598 			sc->eedp_enabled = TRUE;
599 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
600 			sc->control_TLR = TRUE;
601 
602 		mps_resize_queues(sc);
603 
604 		/*
605 		 * Initialize all Tail Queues
606 		 */
607 		TAILQ_INIT(&sc->req_list);
608 		TAILQ_INIT(&sc->high_priority_req_list);
609 		TAILQ_INIT(&sc->chain_list);
610 		TAILQ_INIT(&sc->tm_list);
611 	}
612 
613 	/*
614 	 * If doing a Diag Reset and the FW is significantly different
615 	 * (reallocating will be set above in IOC Facts comparison), then all
616 	 * buffers based on the IOC Facts will need to be freed before they are
617 	 * reallocated.
618 	 */
619 	if (reallocating) {
620 		mps_iocfacts_free(sc);
621 		mpssas_realloc_targets(sc, saved_facts.MaxTargets +
622 		    saved_facts.MaxVolumes);
623 	}
624 
625 	/*
626 	 * Any deallocation has been completed.  Now start reallocating
627 	 * if needed.  Will only need to reallocate if attaching or if the new
628 	 * IOC Facts are different from the previous IOC Facts after a Diag
629 	 * Reset. Targets have already been allocated above if needed.
630 	 */
631 	error = 0;
632 	while (attaching || reallocating) {
633 		if ((error = mps_alloc_hw_queues(sc)) != 0)
634 			break;
635 		if ((error = mps_alloc_replies(sc)) != 0)
636 			break;
637 		if ((error = mps_alloc_requests(sc)) != 0)
638 			break;
639 		if ((error = mps_alloc_queues(sc)) != 0)
640 			break;
641 
642 		break;
643 	}
644 	if (error) {
645 		mps_dprint(sc, MPS_INIT|MPS_FAULT,
646 		    "Failed to alloc queues with error %d\n", error);
647 		mps_free(sc);
648 		return (error);
649 	}
650 
651 	/* Always initialize the queues */
652 	bzero(sc->free_queue, sc->fqdepth * 4);
653 	mps_init_queues(sc);
654 
655 	/*
656 	 * Always get the chip out of the reset state, but only panic if not
657 	 * attaching.  If attaching and there is an error, that is handled by
658 	 * the OS.
659 	 */
660 	error = mps_transition_operational(sc);
661 	if (error != 0) {
662 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to "
663 		    "transition to operational with error %d\n", error);
664 		mps_free(sc);
665 		return (error);
666 	}
667 
668 	/*
669 	 * Finish the queue initialization.
670 	 * These are set here instead of in mps_init_queues() because the
671 	 * IOC resets these values during the state transition in
672 	 * mps_transition_operational().  The free index is set to 1
673 	 * because the corresponding index in the IOC is set to 0, and the
674 	 * IOC treats the queues as full if both are set to the same value.
675 	 * Hence the reason that the queue can't hold all of the possible
676 	 * replies.
677 	 */
678 	sc->replypostindex = 0;
679 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
680 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
681 
682 	/*
683 	 * Attach the subsystems so they can prepare their event masks.
684 	 * XXX Should be dynamic so that IM/IR and user modules can attach
685 	 */
686 	error = 0;
687 	while (attaching) {
688 		mps_dprint(sc, MPS_INIT, "Attaching subsystems\n");
689 		if ((error = mps_attach_log(sc)) != 0)
690 			break;
691 		if ((error = mps_attach_sas(sc)) != 0)
692 			break;
693 		if ((error = mps_attach_user(sc)) != 0)
694 			break;
695 		break;
696 	}
697 	if (error) {
698 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to attach all "
699 		    "subsystems: error %d\n", error);
700 		mps_free(sc);
701 		return (error);
702 	}
703 
704 	/*
705 	 * XXX If the number of MSI-X vectors changes during re-init, this
706 	 * won't see it and adjust.
707 	 */
708 	if (attaching && (error = mps_pci_setup_interrupts(sc)) != 0) {
709 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Failed to setup "
710 		    "interrupts\n");
711 		mps_free(sc);
712 		return (error);
713 	}
714 
715 	/*
716 	 * Set flag if this is a WD controller.  This shouldn't ever change, but
717 	 * reset it after a Diag Reset, just in case.
718 	 */
719 	sc->WD_available = FALSE;
720 	if (pci_get_device(sc->mps_dev) == MPI2_MFGPAGE_DEVID_SSS6200)
721 		sc->WD_available = TRUE;
722 
723 	return (error);
724 }
725 
726 /*
727  * This is called if memory is being free (during detach for example) and when
728  * buffers need to be reallocated due to a Diag Reset.
729  */
730 static void
731 mps_iocfacts_free(struct mps_softc *sc)
732 {
733 	struct mps_command *cm;
734 	int i;
735 
736 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
737 
738 	if (sc->free_busaddr != 0)
739 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
740 	if (sc->free_queue != NULL)
741 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
742 		    sc->queues_map);
743 	if (sc->queues_dmat != NULL)
744 		bus_dma_tag_destroy(sc->queues_dmat);
745 
746 	if (sc->chain_frames != NULL) {
747 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
748 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
749 		    sc->chain_map);
750 	}
751 	if (sc->chain_dmat != NULL)
752 		bus_dma_tag_destroy(sc->chain_dmat);
753 
754 	if (sc->sense_busaddr != 0)
755 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
756 	if (sc->sense_frames != NULL)
757 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
758 		    sc->sense_map);
759 	if (sc->sense_dmat != NULL)
760 		bus_dma_tag_destroy(sc->sense_dmat);
761 
762 	if (sc->reply_busaddr != 0)
763 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
764 	if (sc->reply_frames != NULL)
765 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
766 		    sc->reply_map);
767 	if (sc->reply_dmat != NULL)
768 		bus_dma_tag_destroy(sc->reply_dmat);
769 
770 	if (sc->req_busaddr != 0)
771 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
772 	if (sc->req_frames != NULL)
773 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
774 	if (sc->req_dmat != NULL)
775 		bus_dma_tag_destroy(sc->req_dmat);
776 
777 	if (sc->chains != NULL)
778 		free(sc->chains, M_MPT2);
779 	if (sc->commands != NULL) {
780 		for (i = 1; i < sc->num_reqs; i++) {
781 			cm = &sc->commands[i];
782 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
783 		}
784 		free(sc->commands, M_MPT2);
785 	}
786 	if (sc->buffer_dmat != NULL)
787 		bus_dma_tag_destroy(sc->buffer_dmat);
788 
789 	mps_pci_free_interrupts(sc);
790 	free(sc->queues, M_MPT2);
791 	sc->queues = NULL;
792 }
793 
794 /*
795  * The terms diag reset and hard reset are used interchangeably in the MPI
796  * docs to mean resetting the controller chip.  In this code diag reset
797  * cleans everything up, and the hard reset function just sends the reset
798  * sequence to the chip.  This should probably be refactored so that every
799  * subsystem gets a reset notification of some sort, and can clean up
800  * appropriately.
801  */
802 int
803 mps_reinit(struct mps_softc *sc)
804 {
805 	int error;
806 	struct mpssas_softc *sassc;
807 
808 	sassc = sc->sassc;
809 
810 	MPS_FUNCTRACE(sc);
811 
812 	mtx_assert(&sc->mps_mtx, MA_OWNED);
813 
814 	mps_dprint(sc, MPS_INIT|MPS_INFO, "Reinitializing controller\n");
815 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
816 		mps_dprint(sc, MPS_INIT, "Reset already in progress\n");
817 		return 0;
818 	}
819 
820 	/* make sure the completion callbacks can recognize they're getting
821 	 * a NULL cm_reply due to a reset.
822 	 */
823 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
824 
825 	/*
826 	 * Mask interrupts here.
827 	 */
828 	mps_dprint(sc, MPS_INIT, "masking interrupts and resetting\n");
829 	mps_mask_intr(sc);
830 
831 	error = mps_diag_reset(sc, CAN_SLEEP);
832 	if (error != 0) {
833 		/* XXXSL No need to panic here */
834 		panic("%s hard reset failed with error %d\n",
835 		    __func__, error);
836 	}
837 
838 	/* Restore the PCI state, including the MSI-X registers */
839 	mps_pci_restore(sc);
840 
841 	/* Give the I/O subsystem special priority to get itself prepared */
842 	mpssas_handle_reinit(sc);
843 
844 	/*
845 	 * Get IOC Facts and allocate all structures based on this information.
846 	 * The attach function will also call mps_iocfacts_allocate at startup.
847 	 * If relevant values have changed in IOC Facts, this function will free
848 	 * all of the memory based on IOC Facts and reallocate that memory.
849 	 */
850 	if ((error = mps_iocfacts_allocate(sc, FALSE)) != 0) {
851 		panic("%s IOC Facts based allocation failed with error %d\n",
852 		    __func__, error);
853 	}
854 
855 	/*
856 	 * Mapping structures will be re-allocated after getting IOC Page8, so
857 	 * free these structures here.
858 	 */
859 	mps_mapping_exit(sc);
860 
861 	/*
862 	 * The static page function currently read is IOC Page8.  Others can be
863 	 * added in future.  It's possible that the values in IOC Page8 have
864 	 * changed after a Diag Reset due to user modification, so always read
865 	 * these.  Interrupts are masked, so unmask them before getting config
866 	 * pages.
867 	 */
868 	mps_unmask_intr(sc);
869 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
870 	mps_base_static_config_pages(sc);
871 
872 	/*
873 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
874 	 * mapping tables.
875 	 */
876 	mps_mapping_initialize(sc);
877 
878 	/*
879 	 * Restart will reload the event masks clobbered by the reset, and
880 	 * then enable the port.
881 	 */
882 	mps_reregister_events(sc);
883 
884 	/* the end of discovery will release the simq, so we're done. */
885 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Finished sc %p post %u free %u\n",
886 	    sc, sc->replypostindex, sc->replyfreeindex);
887 
888 	mpssas_release_simq_reinit(sassc);
889 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
890 
891 	return 0;
892 }
893 
894 /* Wait for the chip to ACK a word that we've put into its FIFO
895  * Wait for <timeout> seconds. In single loop wait for busy loop
896  * for 500 microseconds.
897  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
898  * */
899 static int
900 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
901 {
902 
903 	u32 cntdn, count;
904 	u32 int_status;
905 	u32 doorbell;
906 
907 	count = 0;
908 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
909 	do {
910 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
911 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
912 			mps_dprint(sc, MPS_TRACE,
913 			"%s: successful count(%d), timeout(%d)\n",
914 			__func__, count, timeout);
915 		return 0;
916 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
917 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
918 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
919 				MPI2_IOC_STATE_FAULT) {
920 				mps_dprint(sc, MPS_FAULT,
921 					"fault_state(0x%04x)!\n", doorbell);
922 				return (EFAULT);
923 			}
924 		} else if (int_status == 0xFFFFFFFF)
925 			goto out;
926 
927 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
928 		* 0.5 milisecond */
929 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
930 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
931 			"mpsdba", hz/1000);
932 		else if (sleep_flag == CAN_SLEEP)
933 			pause("mpsdba", hz/1000);
934 		else
935 			DELAY(500);
936 		count++;
937 	} while (--cntdn);
938 
939 	out:
940 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
941 		"int_status(%x)!\n", __func__, count, int_status);
942 	return (ETIMEDOUT);
943 
944 }
945 
946 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
947 static int
948 mps_wait_db_int(struct mps_softc *sc)
949 {
950 	int retry;
951 
952 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
953 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
954 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
955 			return (0);
956 		DELAY(2000);
957 	}
958 	return (ETIMEDOUT);
959 }
960 
961 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
962 static int
963 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
964     int req_sz, int reply_sz, int timeout)
965 {
966 	uint32_t *data32;
967 	uint16_t *data16;
968 	int i, count, ioc_sz, residual;
969 	int sleep_flags = CAN_SLEEP;
970 
971 	if (curthread->td_no_sleeping != 0)
972 		sleep_flags = NO_SLEEP;
973 
974 	/* Step 1 */
975 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
976 
977 	/* Step 2 */
978 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
979 		return (EBUSY);
980 
981 	/* Step 3
982 	 * Announce that a message is coming through the doorbell.  Messages
983 	 * are pushed at 32bit words, so round up if needed.
984 	 */
985 	count = (req_sz + 3) / 4;
986 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
987 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
988 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
989 
990 	/* Step 4 */
991 	if (mps_wait_db_int(sc) ||
992 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
993 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
994 		return (ENXIO);
995 	}
996 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
997 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
998 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
999 		return (ENXIO);
1000 	}
1001 
1002 	/* Step 5 */
1003 	/* Clock out the message data synchronously in 32-bit dwords*/
1004 	data32 = (uint32_t *)req;
1005 	for (i = 0; i < count; i++) {
1006 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1007 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
1008 			mps_dprint(sc, MPS_FAULT,
1009 			    "Timeout while writing doorbell\n");
1010 			return (ENXIO);
1011 		}
1012 	}
1013 
1014 	/* Step 6 */
1015 	/* Clock in the reply in 16-bit words.  The total length of the
1016 	 * message is always in the 4th byte, so clock out the first 2 words
1017 	 * manually, then loop the rest.
1018 	 */
1019 	data16 = (uint16_t *)reply;
1020 	if (mps_wait_db_int(sc) != 0) {
1021 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
1022 		return (ENXIO);
1023 	}
1024 	data16[0] =
1025 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1026 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1027 	if (mps_wait_db_int(sc) != 0) {
1028 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
1029 		return (ENXIO);
1030 	}
1031 	data16[1] =
1032 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1033 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1034 
1035 	/* Number of 32bit words in the message */
1036 	ioc_sz = reply->MsgLength;
1037 
1038 	/*
1039 	 * Figure out how many 16bit words to clock in without overrunning.
1040 	 * The precision loss with dividing reply_sz can safely be
1041 	 * ignored because the messages can only be multiples of 32bits.
1042 	 */
1043 	residual = 0;
1044 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1045 	if (count < ioc_sz * 2) {
1046 		residual = ioc_sz * 2 - count;
1047 		mps_dprint(sc, MPS_ERROR, "Driver error, throwing away %d "
1048 		    "residual message words\n", residual);
1049 	}
1050 
1051 	for (i = 2; i < count; i++) {
1052 		if (mps_wait_db_int(sc) != 0) {
1053 			mps_dprint(sc, MPS_FAULT,
1054 			    "Timeout reading doorbell %d\n", i);
1055 			return (ENXIO);
1056 		}
1057 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
1058 		    MPI2_DOORBELL_DATA_MASK;
1059 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1060 	}
1061 
1062 	/*
1063 	 * Pull out residual words that won't fit into the provided buffer.
1064 	 * This keeps the chip from hanging due to a driver programming
1065 	 * error.
1066 	 */
1067 	while (residual--) {
1068 		if (mps_wait_db_int(sc) != 0) {
1069 			mps_dprint(sc, MPS_FAULT,
1070 			    "Timeout reading doorbell\n");
1071 			return (ENXIO);
1072 		}
1073 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
1074 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1075 	}
1076 
1077 	/* Step 7 */
1078 	if (mps_wait_db_int(sc) != 0) {
1079 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
1080 		return (ENXIO);
1081 	}
1082 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1083 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
1084 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1085 
1086 	return (0);
1087 }
1088 
1089 static void
1090 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
1091 {
1092 	reply_descriptor rd;
1093 	MPS_FUNCTRACE(sc);
1094 	mps_dprint(sc, MPS_TRACE, "SMID %u cm %p ccb %p\n",
1095 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1096 
1097 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
1098 		mtx_assert(&sc->mps_mtx, MA_OWNED);
1099 
1100 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1101 		sc->io_cmds_highwater++;
1102 	rd.u.low = cm->cm_desc.Words.Low;
1103 	rd.u.high = cm->cm_desc.Words.High;
1104 	rd.word = htole64(rd.word);
1105 
1106 	KASSERT(cm->cm_state == MPS_CM_STATE_BUSY, ("command not busy\n"));
1107 	cm->cm_state = MPS_CM_STATE_INQUEUE;
1108 
1109 	/* TODO-We may need to make below regwrite atomic */
1110 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1111 	    rd.u.low);
1112 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1113 	    rd.u.high);
1114 }
1115 
1116 /*
1117  * Just the FACTS, ma'am.
1118  */
1119 static int
1120 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1121 {
1122 	MPI2_DEFAULT_REPLY *reply;
1123 	MPI2_IOC_FACTS_REQUEST request;
1124 	int error, req_sz, reply_sz;
1125 
1126 	MPS_FUNCTRACE(sc);
1127 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1128 
1129 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1130 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1131 	reply = (MPI2_DEFAULT_REPLY *)facts;
1132 
1133 	bzero(&request, req_sz);
1134 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1135 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1136 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1137 
1138 	return (error);
1139 }
1140 
1141 static int
1142 mps_send_iocinit(struct mps_softc *sc)
1143 {
1144 	MPI2_IOC_INIT_REQUEST	init;
1145 	MPI2_DEFAULT_REPLY	reply;
1146 	int req_sz, reply_sz, error;
1147 	struct timeval now;
1148 	uint64_t time_in_msec;
1149 
1150 	MPS_FUNCTRACE(sc);
1151 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1152 
1153 	/* Do a quick sanity check on proper initialization */
1154 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1155 	    || (sc->replyframesz == 0)) {
1156 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1157 		    "Driver not fully initialized for IOCInit\n");
1158 		return (EINVAL);
1159 	}
1160 
1161 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1162 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1163 	bzero(&init, req_sz);
1164 	bzero(&reply, reply_sz);
1165 
1166 	/*
1167 	 * Fill in the init block.  Note that most addresses are
1168 	 * deliberately in the lower 32bits of memory.  This is a micro-
1169 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1170 	 */
1171 	init.Function = MPI2_FUNCTION_IOC_INIT;
1172 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1173 	init.MsgVersion = htole16(MPI2_VERSION);
1174 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1175 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1176 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1177 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1178 	init.SenseBufferAddressHigh = 0;
1179 	init.SystemReplyAddressHigh = 0;
1180 	init.SystemRequestFrameBaseAddress.High = 0;
1181 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
1182 	init.ReplyDescriptorPostQueueAddress.High = 0;
1183 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
1184 	init.ReplyFreeQueueAddress.High = 0;
1185 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1186 	getmicrotime(&now);
1187 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1188 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1189 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1190 
1191 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1192 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1193 		error = ENXIO;
1194 
1195 	mps_dprint(sc, MPS_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1196 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
1197 	return (error);
1198 }
1199 
1200 void
1201 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1202 {
1203 	bus_addr_t *addr;
1204 
1205 	addr = arg;
1206 	*addr = segs[0].ds_addr;
1207 }
1208 
1209 void
1210 mps_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1211 {
1212 	struct mps_busdma_context *ctx;
1213 	int need_unload, need_free;
1214 
1215 	ctx = (struct mps_busdma_context *)arg;
1216 	need_unload = 0;
1217 	need_free = 0;
1218 
1219 	mps_lock(ctx->softc);
1220 	ctx->error = error;
1221 	ctx->completed = 1;
1222 	if ((error == 0) && (ctx->abandoned == 0)) {
1223 		*ctx->addr = segs[0].ds_addr;
1224 	} else {
1225 		if (nsegs != 0)
1226 			need_unload = 1;
1227 		if (ctx->abandoned != 0)
1228 			need_free = 1;
1229 	}
1230 	if (need_free == 0)
1231 		wakeup(ctx);
1232 
1233 	mps_unlock(ctx->softc);
1234 
1235 	if (need_unload != 0) {
1236 		bus_dmamap_unload(ctx->buffer_dmat,
1237 				  ctx->buffer_dmamap);
1238 		*ctx->addr = 0;
1239 	}
1240 
1241 	if (need_free != 0)
1242 		free(ctx, M_MPSUSER);
1243 }
1244 
1245 static int
1246 mps_alloc_queues(struct mps_softc *sc)
1247 {
1248 	struct mps_queue *q;
1249 	u_int nq, i;
1250 
1251 	nq = sc->msi_msgs;
1252 	mps_dprint(sc, MPS_INIT|MPS_XINFO, "Allocating %d I/O queues\n", nq);
1253 
1254 	sc->queues = malloc(sizeof(struct mps_queue) * nq, M_MPT2,
1255 	    M_NOWAIT|M_ZERO);
1256 	if (sc->queues == NULL)
1257 		return (ENOMEM);
1258 
1259 	for (i = 0; i < nq; i++) {
1260 		q = &sc->queues[i];
1261 		mps_dprint(sc, MPS_INIT, "Configuring queue %d %p\n", i, q);
1262 		q->sc = sc;
1263 		q->qnum = i;
1264 	}
1265 
1266 	return (0);
1267 }
1268 
1269 static int
1270 mps_alloc_hw_queues(struct mps_softc *sc)
1271 {
1272 	bus_addr_t queues_busaddr;
1273 	uint8_t *queues;
1274 	int qsize, fqsize, pqsize;
1275 
1276 	/*
1277 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1278 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1279 	 * This queue supplies fresh reply frames for the firmware to use.
1280 	 *
1281 	 * The reply descriptor post queue contains 8 byte entries in
1282 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1283 	 * contains filled-in reply frames sent from the firmware to the host.
1284 	 *
1285 	 * These two queues are allocated together for simplicity.
1286 	 */
1287 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1288 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1289 	fqsize= sc->fqdepth * 4;
1290 	pqsize = sc->pqdepth * 8;
1291 	qsize = fqsize + pqsize;
1292 
1293         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1294 				16, 0,			/* algnmnt, boundary */
1295 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1296 				BUS_SPACE_MAXADDR,	/* highaddr */
1297 				NULL, NULL,		/* filter, filterarg */
1298                                 qsize,			/* maxsize */
1299                                 1,			/* nsegments */
1300                                 qsize,			/* maxsegsize */
1301                                 0,			/* flags */
1302                                 NULL, NULL,		/* lockfunc, lockarg */
1303                                 &sc->queues_dmat)) {
1304 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues DMA tag\n");
1305 		return (ENOMEM);
1306         }
1307         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1308 	    &sc->queues_map)) {
1309 		mps_dprint(sc, MPS_ERROR, "Cannot allocate queues memory\n");
1310 		return (ENOMEM);
1311         }
1312         bzero(queues, qsize);
1313         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1314 	    mps_memaddr_cb, &queues_busaddr, 0);
1315 
1316 	sc->free_queue = (uint32_t *)queues;
1317 	sc->free_busaddr = queues_busaddr;
1318 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1319 	sc->post_busaddr = queues_busaddr + fqsize;
1320 	mps_dprint(sc, MPS_INIT, "free queue busaddr= %#016jx size= %d\n",
1321 	    (uintmax_t)sc->free_busaddr, fqsize);
1322 	mps_dprint(sc, MPS_INIT, "reply queue busaddr= %#016jx size= %d\n",
1323 	    (uintmax_t)sc->post_busaddr, pqsize);
1324 
1325 	return (0);
1326 }
1327 
1328 static int
1329 mps_alloc_replies(struct mps_softc *sc)
1330 {
1331 	int rsize, num_replies;
1332 
1333 	/* Store the reply frame size in bytes rather than as 32bit words */
1334 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1335 
1336 	/*
1337 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1338 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1339 	 * replies can be used at once.
1340 	 */
1341 	num_replies = max(sc->fqdepth, sc->num_replies);
1342 
1343 	rsize = sc->replyframesz * num_replies;
1344         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1345 				4, 0,			/* algnmnt, boundary */
1346 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1347 				BUS_SPACE_MAXADDR,	/* highaddr */
1348 				NULL, NULL,		/* filter, filterarg */
1349                                 rsize,			/* maxsize */
1350                                 1,			/* nsegments */
1351                                 rsize,			/* maxsegsize */
1352                                 0,			/* flags */
1353                                 NULL, NULL,		/* lockfunc, lockarg */
1354                                 &sc->reply_dmat)) {
1355 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies DMA tag\n");
1356 		return (ENOMEM);
1357         }
1358         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1359 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1360 		mps_dprint(sc, MPS_ERROR, "Cannot allocate replies memory\n");
1361 		return (ENOMEM);
1362         }
1363         bzero(sc->reply_frames, rsize);
1364         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1365 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
1366 
1367 	mps_dprint(sc, MPS_INIT, "reply frames busaddr= %#016jx size= %d\n",
1368 	    (uintmax_t)sc->reply_busaddr, rsize);
1369 
1370 	return (0);
1371 }
1372 
1373 static void
1374 mps_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1375 {
1376 	struct mps_softc *sc = arg;
1377 	struct mps_chain *chain;
1378 	bus_size_t bo;
1379 	int i, o, s;
1380 
1381 	if (error != 0)
1382 		return;
1383 
1384 	for (i = 0, o = 0, s = 0; s < nsegs; s++) {
1385 		for (bo = 0; bo + sc->reqframesz <= segs[s].ds_len;
1386 		    bo += sc->reqframesz) {
1387 			chain = &sc->chains[i++];
1388 			chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o);
1389 			chain->chain_busaddr = segs[s].ds_addr + bo;
1390 			o += sc->reqframesz;
1391 			mps_free_chain(sc, chain);
1392 		}
1393 		if (bo != segs[s].ds_len)
1394 			o += segs[s].ds_len - bo;
1395 	}
1396 	sc->chain_free_lowwater = i;
1397 }
1398 
1399 static int
1400 mps_alloc_requests(struct mps_softc *sc)
1401 {
1402 	struct mps_command *cm;
1403 	int i, rsize, nsegs;
1404 
1405 	rsize = sc->reqframesz * sc->num_reqs;
1406         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1407 				16, 0,			/* algnmnt, boundary */
1408 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1409 				BUS_SPACE_MAXADDR,	/* highaddr */
1410 				NULL, NULL,		/* filter, filterarg */
1411                                 rsize,			/* maxsize */
1412                                 1,			/* nsegments */
1413                                 rsize,			/* maxsegsize */
1414                                 0,			/* flags */
1415                                 NULL, NULL,		/* lockfunc, lockarg */
1416                                 &sc->req_dmat)) {
1417 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request DMA tag\n");
1418 		return (ENOMEM);
1419         }
1420         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1421 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1422 		mps_dprint(sc, MPS_ERROR, "Cannot allocate request memory\n");
1423 		return (ENOMEM);
1424         }
1425         bzero(sc->req_frames, rsize);
1426         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1427 	    mps_memaddr_cb, &sc->req_busaddr, 0);
1428 	mps_dprint(sc, MPS_INIT, "request frames busaddr= %#016jx size= %d\n",
1429 	    (uintmax_t)sc->req_busaddr, rsize);
1430 
1431 	sc->chains = malloc(sizeof(struct mps_chain) * sc->num_chains, M_MPT2,
1432 	    M_NOWAIT | M_ZERO);
1433 	if (!sc->chains) {
1434 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1435 		return (ENOMEM);
1436 	}
1437 	rsize = sc->reqframesz * sc->num_chains;
1438 	if (bus_dma_tag_create( sc->mps_parent_dmat,	/* parent */
1439 				16, 0,			/* algnmnt, boundary */
1440 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1441 				BUS_SPACE_MAXADDR,	/* highaddr */
1442 				NULL, NULL,		/* filter, filterarg */
1443 				rsize,			/* maxsize */
1444 				howmany(rsize, PAGE_SIZE), /* nsegments */
1445 				rsize,			/* maxsegsize */
1446 				0,			/* flags */
1447 				NULL, NULL,		/* lockfunc, lockarg */
1448 				&sc->chain_dmat)) {
1449 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain DMA tag\n");
1450 		return (ENOMEM);
1451 	}
1452 	if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1453 	    BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) {
1454 		mps_dprint(sc, MPS_ERROR, "Cannot allocate chain memory\n");
1455 		return (ENOMEM);
1456 	}
1457 	if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames,
1458 	    rsize, mps_load_chains_cb, sc, BUS_DMA_NOWAIT)) {
1459 		mps_dprint(sc, MPS_ERROR, "Cannot load chain memory\n");
1460 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
1461 		    sc->chain_map);
1462 		return (ENOMEM);
1463 	}
1464 
1465 	rsize = MPS_SENSE_LEN * sc->num_reqs;
1466         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1467 				1, 0,			/* algnmnt, boundary */
1468 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1469 				BUS_SPACE_MAXADDR,	/* highaddr */
1470 				NULL, NULL,		/* filter, filterarg */
1471                                 rsize,			/* maxsize */
1472                                 1,			/* nsegments */
1473                                 rsize,			/* maxsegsize */
1474                                 0,			/* flags */
1475                                 NULL, NULL,		/* lockfunc, lockarg */
1476                                 &sc->sense_dmat)) {
1477 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense DMA tag\n");
1478 		return (ENOMEM);
1479         }
1480         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1481 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1482 		mps_dprint(sc, MPS_ERROR, "Cannot allocate sense memory\n");
1483 		return (ENOMEM);
1484         }
1485         bzero(sc->sense_frames, rsize);
1486         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1487 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
1488 	mps_dprint(sc, MPS_INIT, "sense frames busaddr= %#016jx size= %d\n",
1489 	    (uintmax_t)sc->sense_busaddr, rsize);
1490 
1491 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1492         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
1493 				1, 0,			/* algnmnt, boundary */
1494 				BUS_SPACE_MAXADDR,	/* lowaddr */
1495 				BUS_SPACE_MAXADDR,	/* highaddr */
1496 				NULL, NULL,		/* filter, filterarg */
1497                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1498                                 nsegs,			/* nsegments */
1499                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
1500                                 BUS_DMA_ALLOCNOW,	/* flags */
1501                                 busdma_lock_mutex,	/* lockfunc */
1502 				&sc->mps_mtx,		/* lockarg */
1503                                 &sc->buffer_dmat)) {
1504 		mps_dprint(sc, MPS_ERROR, "Cannot allocate buffer DMA tag\n");
1505 		return (ENOMEM);
1506         }
1507 
1508 	/*
1509 	 * SMID 0 cannot be used as a free command per the firmware spec.
1510 	 * Just drop that command instead of risking accounting bugs.
1511 	 */
1512 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
1513 	    M_MPT2, M_WAITOK | M_ZERO);
1514 	if(!sc->commands) {
1515 		mps_dprint(sc, MPS_ERROR, "Cannot allocate command memory\n");
1516 		return (ENOMEM);
1517 	}
1518 	for (i = 1; i < sc->num_reqs; i++) {
1519 		cm = &sc->commands[i];
1520 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1521 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1522 		cm->cm_sense = &sc->sense_frames[i];
1523 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
1524 		cm->cm_desc.Default.SMID = i;
1525 		cm->cm_sc = sc;
1526 		cm->cm_state = MPS_CM_STATE_BUSY;
1527 		TAILQ_INIT(&cm->cm_chain_list);
1528 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
1529 
1530 		/* XXX Is a failure here a critical problem? */
1531 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
1532 			if (i <= sc->num_prireqs)
1533 				mps_free_high_priority_command(sc, cm);
1534 			else
1535 				mps_free_command(sc, cm);
1536 		else {
1537 			panic("failed to allocate command %d\n", i);
1538 			sc->num_reqs = i;
1539 			break;
1540 		}
1541 	}
1542 
1543 	return (0);
1544 }
1545 
1546 static int
1547 mps_init_queues(struct mps_softc *sc)
1548 {
1549 	int i;
1550 
1551 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1552 
1553 	/*
1554 	 * According to the spec, we need to use one less reply than we
1555 	 * have space for on the queue.  So sc->num_replies (the number we
1556 	 * use) should be less than sc->fqdepth (allocated size).
1557 	 */
1558 	if (sc->num_replies >= sc->fqdepth)
1559 		return (EINVAL);
1560 
1561 	/*
1562 	 * Initialize all of the free queue entries.
1563 	 */
1564 	for (i = 0; i < sc->fqdepth; i++)
1565 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1566 	sc->replyfreeindex = sc->num_replies;
1567 
1568 	return (0);
1569 }
1570 
1571 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1572  * Next are the global settings, if they exist.  Highest are the per-unit
1573  * settings, if they exist.
1574  */
1575 void
1576 mps_get_tunables(struct mps_softc *sc)
1577 {
1578 	char tmpstr[80], mps_debug[80];
1579 
1580 	/* XXX default to some debugging for now */
1581 	sc->mps_debug = MPS_INFO|MPS_FAULT;
1582 	sc->disable_msix = 0;
1583 	sc->disable_msi = 0;
1584 	sc->max_msix = MPS_MSIX_MAX;
1585 	sc->max_chains = MPS_CHAIN_FRAMES;
1586 	sc->max_io_pages = MPS_MAXIO_PAGES;
1587 	sc->enable_ssu = MPS_SSU_ENABLE_SSD_DISABLE_HDD;
1588 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1589 	sc->use_phynum = 1;
1590 	sc->max_reqframes = MPS_REQ_FRAMES;
1591 	sc->max_prireqframes = MPS_PRI_REQ_FRAMES;
1592 	sc->max_replyframes = MPS_REPLY_FRAMES;
1593 	sc->max_evtframes = MPS_EVT_REPLY_FRAMES;
1594 
1595 	/*
1596 	 * Grab the global variables.
1597 	 */
1598 	bzero(mps_debug, 80);
1599 	if (TUNABLE_STR_FETCH("hw.mps.debug_level", mps_debug, 80) != 0)
1600 		mps_parse_debug(sc, mps_debug);
1601 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1602 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1603 	TUNABLE_INT_FETCH("hw.mps.max_msix", &sc->max_msix);
1604 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1605 	TUNABLE_INT_FETCH("hw.mps.max_io_pages", &sc->max_io_pages);
1606 	TUNABLE_INT_FETCH("hw.mps.enable_ssu", &sc->enable_ssu);
1607 	TUNABLE_INT_FETCH("hw.mps.spinup_wait_time", &sc->spinup_wait_time);
1608 	TUNABLE_INT_FETCH("hw.mps.use_phy_num", &sc->use_phynum);
1609 	TUNABLE_INT_FETCH("hw.mps.max_reqframes", &sc->max_reqframes);
1610 	TUNABLE_INT_FETCH("hw.mps.max_prireqframes", &sc->max_prireqframes);
1611 	TUNABLE_INT_FETCH("hw.mps.max_replyframes", &sc->max_replyframes);
1612 	TUNABLE_INT_FETCH("hw.mps.max_evtframes", &sc->max_evtframes);
1613 
1614 	/* Grab the unit-instance variables */
1615 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1616 	    device_get_unit(sc->mps_dev));
1617 	bzero(mps_debug, 80);
1618 	if (TUNABLE_STR_FETCH(tmpstr, mps_debug, 80) != 0)
1619 		mps_parse_debug(sc, mps_debug);
1620 
1621 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1622 	    device_get_unit(sc->mps_dev));
1623 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1624 
1625 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1626 	    device_get_unit(sc->mps_dev));
1627 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1628 
1629 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_msix",
1630 	    device_get_unit(sc->mps_dev));
1631 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1632 
1633 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1634 	    device_get_unit(sc->mps_dev));
1635 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1636 
1637 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_io_pages",
1638 	    device_get_unit(sc->mps_dev));
1639 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1640 
1641 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1642 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.exclude_ids",
1643 	    device_get_unit(sc->mps_dev));
1644 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1645 
1646 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.enable_ssu",
1647 	    device_get_unit(sc->mps_dev));
1648 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1649 
1650 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.spinup_wait_time",
1651 	    device_get_unit(sc->mps_dev));
1652 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1653 
1654 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.use_phy_num",
1655 	    device_get_unit(sc->mps_dev));
1656 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1657 
1658 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_reqframes",
1659 	    device_get_unit(sc->mps_dev));
1660 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1661 
1662 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_prireqframes",
1663 	    device_get_unit(sc->mps_dev));
1664 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1665 
1666 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_replyframes",
1667 	    device_get_unit(sc->mps_dev));
1668 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1669 
1670 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_evtframes",
1671 	    device_get_unit(sc->mps_dev));
1672 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1673 
1674 }
1675 
1676 static void
1677 mps_setup_sysctl(struct mps_softc *sc)
1678 {
1679 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1680 	struct sysctl_oid	*sysctl_tree = NULL;
1681 	char tmpstr[80], tmpstr2[80];
1682 
1683 	/*
1684 	 * Setup the sysctl variable so the user can change the debug level
1685 	 * on the fly.
1686 	 */
1687 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1688 	    device_get_unit(sc->mps_dev));
1689 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1690 
1691 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1692 	if (sysctl_ctx != NULL)
1693 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1694 
1695 	if (sysctl_tree == NULL) {
1696 		sysctl_ctx_init(&sc->sysctl_ctx);
1697 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1698 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1699 		    CTLFLAG_RD, 0, tmpstr);
1700 		if (sc->sysctl_tree == NULL)
1701 			return;
1702 		sysctl_ctx = &sc->sysctl_ctx;
1703 		sysctl_tree = sc->sysctl_tree;
1704 	}
1705 
1706 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1707 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW |CTLFLAG_MPSAFE,
1708 	    sc, 0, mps_debug_sysctl, "A", "mps debug level");
1709 
1710 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1711 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1712 	    "Disable the use of MSI-X interrupts");
1713 
1714 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1715 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1716 	    "Disable the use of MSI interrupts");
1717 
1718 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1719 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1720 	    "User-defined maximum number of MSIX queues");
1721 
1722 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1723 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1724 	    "Negotiated number of MSIX queues");
1725 
1726 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1727 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1728 	    "Total number of allocated request frames");
1729 
1730 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1731 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1732 	    "Total number of allocated high priority request frames");
1733 
1734 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1735 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1736 	    "Total number of allocated reply frames");
1737 
1738 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1739 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1740 	    "Total number of event frames allocated");
1741 
1742 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1743 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1744 	    strlen(sc->fw_version), "firmware version");
1745 
1746 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1747 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1748 	    strlen(MPS_DRIVER_VERSION), "driver version");
1749 
1750 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1751 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1752 	    &sc->io_cmds_active, 0, "number of currently active commands");
1753 
1754 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1755 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1756 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1757 
1758 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1759 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1760 	    &sc->chain_free, 0, "number of free chain elements");
1761 
1762 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1763 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1764 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1765 
1766 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1767 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1768 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1769 
1770 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1771 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1772 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1773 	    "IOCFacts)");
1774 
1775 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1776 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1777 	    "enable SSU to SATA SSD/HDD at shutdown");
1778 
1779 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1780 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1781 	    &sc->chain_alloc_fail, "chain allocation failures");
1782 
1783 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1784 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1785 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1786 	    "spinup after SATA ID error");
1787 
1788 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1789 	    OID_AUTO, "mapping_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1790 	    mps_mapping_dump, "A", "Mapping Table Dump");
1791 
1792 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1793 	    OID_AUTO, "encl_table_dump", CTLTYPE_STRING | CTLFLAG_RD, sc, 0,
1794 	    mps_mapping_encl_dump, "A", "Enclosure Table Dump");
1795 
1796 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1797 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1798 	    "Use the phy number for enumeration");
1799 }
1800 
1801 static struct mps_debug_string {
1802 	char	*name;
1803 	int	flag;
1804 } mps_debug_strings[] = {
1805 	{"info", MPS_INFO},
1806 	{"fault", MPS_FAULT},
1807 	{"event", MPS_EVENT},
1808 	{"log", MPS_LOG},
1809 	{"recovery", MPS_RECOVERY},
1810 	{"error", MPS_ERROR},
1811 	{"init", MPS_INIT},
1812 	{"xinfo", MPS_XINFO},
1813 	{"user", MPS_USER},
1814 	{"mapping", MPS_MAPPING},
1815 	{"trace", MPS_TRACE}
1816 };
1817 
1818 enum mps_debug_level_combiner {
1819 	COMB_NONE,
1820 	COMB_ADD,
1821 	COMB_SUB
1822 };
1823 
1824 static int
1825 mps_debug_sysctl(SYSCTL_HANDLER_ARGS)
1826 {
1827 	struct mps_softc *sc;
1828 	struct mps_debug_string *string;
1829 	struct sbuf *sbuf;
1830 	char *buffer;
1831 	size_t sz;
1832 	int i, len, debug, error;
1833 
1834 	sc = (struct mps_softc *)arg1;
1835 
1836 	error = sysctl_wire_old_buffer(req, 0);
1837 	if (error != 0)
1838 		return (error);
1839 
1840 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1841 	debug = sc->mps_debug;
1842 
1843 	sbuf_printf(sbuf, "%#x", debug);
1844 
1845 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1846 	for (i = 0; i < sz; i++) {
1847 		string = &mps_debug_strings[i];
1848 		if (debug & string->flag)
1849 			sbuf_printf(sbuf, ",%s", string->name);
1850 	}
1851 
1852 	error = sbuf_finish(sbuf);
1853 	sbuf_delete(sbuf);
1854 
1855 	if (error || req->newptr == NULL)
1856 		return (error);
1857 
1858 	len = req->newlen - req->newidx;
1859 	if (len == 0)
1860 		return (0);
1861 
1862 	buffer = malloc(len, M_MPT2, M_ZERO|M_WAITOK);
1863 	error = SYSCTL_IN(req, buffer, len);
1864 
1865 	mps_parse_debug(sc, buffer);
1866 
1867 	free(buffer, M_MPT2);
1868 	return (error);
1869 }
1870 
1871 static void
1872 mps_parse_debug(struct mps_softc *sc, char *list)
1873 {
1874 	struct mps_debug_string *string;
1875 	enum mps_debug_level_combiner op;
1876 	char *token, *endtoken;
1877 	size_t sz;
1878 	int flags, i;
1879 
1880 	if (list == NULL || *list == '\0')
1881 		return;
1882 
1883 	if (*list == '+') {
1884 		op = COMB_ADD;
1885 		list++;
1886 	} else if (*list == '-') {
1887 		op = COMB_SUB;
1888 		list++;
1889 	} else
1890 		op = COMB_NONE;
1891 	if (*list == '\0')
1892 		return;
1893 
1894 	flags = 0;
1895 	sz = sizeof(mps_debug_strings) / sizeof(mps_debug_strings[0]);
1896 	while ((token = strsep(&list, ":,")) != NULL) {
1897 
1898 		/* Handle integer flags */
1899 		flags |= strtol(token, &endtoken, 0);
1900 		if (token != endtoken)
1901 			continue;
1902 
1903 		/* Handle text flags */
1904 		for (i = 0; i < sz; i++) {
1905 			string = &mps_debug_strings[i];
1906 			if (strcasecmp(token, string->name) == 0) {
1907 				flags |= string->flag;
1908 				break;
1909 			}
1910 		}
1911 	}
1912 
1913 	switch (op) {
1914 	case COMB_NONE:
1915 		sc->mps_debug = flags;
1916 		break;
1917 	case COMB_ADD:
1918 		sc->mps_debug |= flags;
1919 		break;
1920 	case COMB_SUB:
1921 		sc->mps_debug &= (~flags);
1922 		break;
1923 	}
1924 
1925 	return;
1926 }
1927 
1928 int
1929 mps_attach(struct mps_softc *sc)
1930 {
1931 	int error;
1932 
1933 	MPS_FUNCTRACE(sc);
1934 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
1935 
1936 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1937 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1938 	callout_init_mtx(&sc->device_check_callout, &sc->mps_mtx, 0);
1939 	TAILQ_INIT(&sc->event_list);
1940 	timevalclear(&sc->lastfail);
1941 
1942 	if ((error = mps_transition_ready(sc)) != 0) {
1943 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to transition "
1944 		    "ready\n");
1945 		return (error);
1946 	}
1947 
1948 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1949 	    M_ZERO|M_NOWAIT);
1950 	if(!sc->facts) {
1951 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "Cannot allocate memory, "
1952 		    "exit\n");
1953 		return (ENOMEM);
1954 	}
1955 
1956 	/*
1957 	 * Get IOC Facts and allocate all structures based on this information.
1958 	 * A Diag Reset will also call mps_iocfacts_allocate and re-read the IOC
1959 	 * Facts. If relevant values have changed in IOC Facts, this function
1960 	 * will free all of the memory based on IOC Facts and reallocate that
1961 	 * memory.  If this fails, any allocated memory should already be freed.
1962 	 */
1963 	if ((error = mps_iocfacts_allocate(sc, TRUE)) != 0) {
1964 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "IOC Facts based allocation "
1965 		    "failed with error %d, exit\n", error);
1966 		return (error);
1967 	}
1968 
1969 	/* Start the periodic watchdog check on the IOC Doorbell */
1970 	mps_periodic(sc);
1971 
1972 	/*
1973 	 * The portenable will kick off discovery events that will drive the
1974 	 * rest of the initialization process.  The CAM/SAS module will
1975 	 * hold up the boot sequence until discovery is complete.
1976 	 */
1977 	sc->mps_ich.ich_func = mps_startup;
1978 	sc->mps_ich.ich_arg = sc;
1979 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1980 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1981 		    "Cannot establish MPS config hook\n");
1982 		error = EINVAL;
1983 	}
1984 
1985 	/*
1986 	 * Allow IR to shutdown gracefully when shutdown occurs.
1987 	 */
1988 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1989 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1990 
1991 	if (sc->shutdown_eh == NULL)
1992 		mps_dprint(sc, MPS_INIT|MPS_ERROR,
1993 		    "shutdown event registration failed\n");
1994 
1995 	mps_setup_sysctl(sc);
1996 
1997 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1998 	mps_dprint(sc, MPS_INIT, "%s exit error= %d\n", __func__, error);
1999 
2000 	return (error);
2001 }
2002 
2003 /* Run through any late-start handlers. */
2004 static void
2005 mps_startup(void *arg)
2006 {
2007 	struct mps_softc *sc;
2008 
2009 	sc = (struct mps_softc *)arg;
2010 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2011 
2012 	mps_lock(sc);
2013 	mps_unmask_intr(sc);
2014 
2015 	/* initialize device mapping tables */
2016 	mps_base_static_config_pages(sc);
2017 	mps_mapping_initialize(sc);
2018 	mpssas_startup(sc);
2019 	mps_unlock(sc);
2020 
2021 	mps_dprint(sc, MPS_INIT, "disestablish config intrhook\n");
2022 	config_intrhook_disestablish(&sc->mps_ich);
2023 	sc->mps_ich.ich_arg = NULL;
2024 
2025 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2026 }
2027 
2028 /* Periodic watchdog.  Is called with the driver lock already held. */
2029 static void
2030 mps_periodic(void *arg)
2031 {
2032 	struct mps_softc *sc;
2033 	uint32_t db;
2034 
2035 	sc = (struct mps_softc *)arg;
2036 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
2037 		return;
2038 
2039 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
2040 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2041 		mps_dprint(sc, MPS_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2042 		mps_reinit(sc);
2043 	}
2044 
2045 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
2046 }
2047 
2048 static void
2049 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
2050     MPI2_EVENT_NOTIFICATION_REPLY *event)
2051 {
2052 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2053 
2054 	MPS_DPRINT_EVENT(sc, generic, event);
2055 
2056 	switch (event->Event) {
2057 	case MPI2_EVENT_LOG_DATA:
2058 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2059 		if (sc->mps_debug & MPS_EVENT)
2060 			hexdump(event->EventData, event->EventDataLength, NULL, 0);
2061 		break;
2062 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2063 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2064 		mps_dprint(sc, MPS_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2065 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2066 		     entry->LogSequence);
2067 		break;
2068 	default:
2069 		break;
2070 	}
2071 	return;
2072 }
2073 
2074 static int
2075 mps_attach_log(struct mps_softc *sc)
2076 {
2077 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
2078 
2079 	bzero(events, 16);
2080 	setbit(events, MPI2_EVENT_LOG_DATA);
2081 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2082 
2083 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
2084 	    &sc->mps_log_eh);
2085 
2086 	return (0);
2087 }
2088 
2089 static int
2090 mps_detach_log(struct mps_softc *sc)
2091 {
2092 
2093 	if (sc->mps_log_eh != NULL)
2094 		mps_deregister_events(sc, sc->mps_log_eh);
2095 	return (0);
2096 }
2097 
2098 /*
2099  * Free all of the driver resources and detach submodules.  Should be called
2100  * without the lock held.
2101  */
2102 int
2103 mps_free(struct mps_softc *sc)
2104 {
2105 	int error;
2106 
2107 	mps_dprint(sc, MPS_INIT, "%s entered\n", __func__);
2108 	/* Turn off the watchdog */
2109 	mps_lock(sc);
2110 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
2111 	mps_unlock(sc);
2112 	/* Lock must not be held for this */
2113 	callout_drain(&sc->periodic);
2114 	callout_drain(&sc->device_check_callout);
2115 
2116 	if (((error = mps_detach_log(sc)) != 0) ||
2117 	    ((error = mps_detach_sas(sc)) != 0)) {
2118 		mps_dprint(sc, MPS_INIT|MPS_FAULT, "failed to detach "
2119 		    "subsystems, exit\n");
2120 		return (error);
2121 	}
2122 
2123 	mps_detach_user(sc);
2124 
2125 	/* Put the IOC back in the READY state. */
2126 	mps_lock(sc);
2127 	if ((error = mps_transition_ready(sc)) != 0) {
2128 		mps_unlock(sc);
2129 		return (error);
2130 	}
2131 	mps_unlock(sc);
2132 
2133 	if (sc->facts != NULL)
2134 		free(sc->facts, M_MPT2);
2135 
2136 	/*
2137 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2138 	 * to free these buffers too.
2139 	 */
2140 	mps_iocfacts_free(sc);
2141 
2142 	if (sc->sysctl_tree != NULL)
2143 		sysctl_ctx_free(&sc->sysctl_ctx);
2144 
2145 	/* Deregister the shutdown function */
2146 	if (sc->shutdown_eh != NULL)
2147 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2148 
2149 	mtx_destroy(&sc->mps_mtx);
2150 	mps_dprint(sc, MPS_INIT, "%s exit\n", __func__);
2151 
2152 	return (0);
2153 }
2154 
2155 static __inline void
2156 mps_complete_command(struct mps_softc *sc, struct mps_command *cm)
2157 {
2158 	MPS_FUNCTRACE(sc);
2159 
2160 	if (cm == NULL) {
2161 		mps_dprint(sc, MPS_ERROR, "Completing NULL command\n");
2162 		return;
2163 	}
2164 
2165 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
2166 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
2167 
2168 	if (cm->cm_complete != NULL) {
2169 		mps_dprint(sc, MPS_TRACE,
2170 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
2171 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
2172 			   cm->cm_reply);
2173 		cm->cm_complete(sc, cm);
2174 	}
2175 
2176 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
2177 		mps_dprint(sc, MPS_TRACE, "waking up %p\n", cm);
2178 		wakeup(cm);
2179 	}
2180 
2181 	if (cm->cm_sc->io_cmds_active != 0) {
2182 		cm->cm_sc->io_cmds_active--;
2183 	} else {
2184 		mps_dprint(sc, MPS_ERROR, "Warning: io_cmds_active is "
2185 		    "out of sync - resynching to 0\n");
2186 	}
2187 }
2188 
2189 
2190 static void
2191 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
2192 {
2193 	union loginfo_type {
2194 		u32     loginfo;
2195 		struct {
2196 			u32     subcode:16;
2197 			u32     code:8;
2198 			u32     originator:4;
2199 			u32     bus_type:4;
2200 		} dw;
2201 	};
2202 	union loginfo_type sas_loginfo;
2203 	char *originator_str = NULL;
2204 
2205 	sas_loginfo.loginfo = log_info;
2206 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2207 		return;
2208 
2209 	/* each nexus loss loginfo */
2210 	if (log_info == 0x31170000)
2211 		return;
2212 
2213 	/* eat the loginfos associated with task aborts */
2214 	if ((log_info == 30050000 || log_info ==
2215 	    0x31140000 || log_info == 0x31130000))
2216 		return;
2217 
2218 	switch (sas_loginfo.dw.originator) {
2219 	case 0:
2220 		originator_str = "IOP";
2221 		break;
2222 	case 1:
2223 		originator_str = "PL";
2224 		break;
2225 	case 2:
2226 		originator_str = "IR";
2227 		break;
2228 }
2229 
2230 	mps_dprint(sc, MPS_LOG, "log_info(0x%08x): originator(%s), "
2231 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
2232 	originator_str, sas_loginfo.dw.code,
2233 	sas_loginfo.dw.subcode);
2234 }
2235 
2236 static void
2237 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
2238 {
2239 	MPI2DefaultReply_t *mpi_reply;
2240 	u16 sc_status;
2241 
2242 	mpi_reply = (MPI2DefaultReply_t*)reply;
2243 	sc_status = le16toh(mpi_reply->IOCStatus);
2244 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2245 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2246 }
2247 void
2248 mps_intr(void *data)
2249 {
2250 	struct mps_softc *sc;
2251 	uint32_t status;
2252 
2253 	sc = (struct mps_softc *)data;
2254 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2255 
2256 	/*
2257 	 * Check interrupt status register to flush the bus.  This is
2258 	 * needed for both INTx interrupts and driver-driven polling
2259 	 */
2260 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2261 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2262 		return;
2263 
2264 	mps_lock(sc);
2265 	mps_intr_locked(data);
2266 	mps_unlock(sc);
2267 	return;
2268 }
2269 
2270 /*
2271  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2272  * chip.  Hopefully this theory is correct.
2273  */
2274 void
2275 mps_intr_msi(void *data)
2276 {
2277 	struct mps_softc *sc;
2278 
2279 	sc = (struct mps_softc *)data;
2280 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2281 	mps_lock(sc);
2282 	mps_intr_locked(data);
2283 	mps_unlock(sc);
2284 	return;
2285 }
2286 
2287 /*
2288  * The locking is overly broad and simplistic, but easy to deal with for now.
2289  */
2290 void
2291 mps_intr_locked(void *data)
2292 {
2293 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2294 	struct mps_softc *sc;
2295 	struct mps_command *cm = NULL;
2296 	uint8_t flags;
2297 	u_int pq;
2298 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2299 	mps_fw_diagnostic_buffer_t *pBuffer;
2300 
2301 	sc = (struct mps_softc *)data;
2302 
2303 	pq = sc->replypostindex;
2304 	mps_dprint(sc, MPS_TRACE,
2305 	    "%s sc %p starting with replypostindex %u\n",
2306 	    __func__, sc, sc->replypostindex);
2307 
2308 	for ( ;; ) {
2309 		cm = NULL;
2310 		desc = &sc->post_queue[sc->replypostindex];
2311 		flags = desc->Default.ReplyFlags &
2312 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2313 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
2314 		 || (le32toh(desc->Words.High) == 0xffffffff))
2315 			break;
2316 
2317 		/* increment the replypostindex now, so that event handlers
2318 		 * and cm completion handlers which decide to do a diag
2319 		 * reset can zero it without it getting incremented again
2320 		 * afterwards, and we break out of this loop on the next
2321 		 * iteration since the reply post queue has been cleared to
2322 		 * 0xFF and all descriptors look unused (which they are).
2323 		 */
2324 		if (++sc->replypostindex >= sc->pqdepth)
2325 			sc->replypostindex = 0;
2326 
2327 		switch (flags) {
2328 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2329 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2330 			KASSERT(cm->cm_state == MPS_CM_STATE_INQUEUE,
2331 			    ("command not inqueue\n"));
2332 			cm->cm_state = MPS_CM_STATE_BUSY;
2333 			cm->cm_reply = NULL;
2334 			break;
2335 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2336 		{
2337 			uint32_t baddr;
2338 			uint8_t *reply;
2339 
2340 			/*
2341 			 * Re-compose the reply address from the address
2342 			 * sent back from the chip.  The ReplyFrameAddress
2343 			 * is the lower 32 bits of the physical address of
2344 			 * particular reply frame.  Convert that address to
2345 			 * host format, and then use that to provide the
2346 			 * offset against the virtual address base
2347 			 * (sc->reply_frames).
2348 			 */
2349 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2350 			reply = sc->reply_frames +
2351 				(baddr - ((uint32_t)sc->reply_busaddr));
2352 			/*
2353 			 * Make sure the reply we got back is in a valid
2354 			 * range.  If not, go ahead and panic here, since
2355 			 * we'll probably panic as soon as we deference the
2356 			 * reply pointer anyway.
2357 			 */
2358 			if ((reply < sc->reply_frames)
2359 			 || (reply > (sc->reply_frames +
2360 			     (sc->fqdepth * sc->replyframesz)))) {
2361 				printf("%s: WARNING: reply %p out of range!\n",
2362 				       __func__, reply);
2363 				printf("%s: reply_frames %p, fqdepth %d, "
2364 				       "frame size %d\n", __func__,
2365 				       sc->reply_frames, sc->fqdepth,
2366 				       sc->replyframesz);
2367 				printf("%s: baddr %#x,\n", __func__, baddr);
2368 				/* LSI-TODO. See Linux Code for Graceful exit */
2369 				panic("Reply address out of range");
2370 			}
2371 			if (le16toh(desc->AddressReply.SMID) == 0) {
2372 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2373 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2374 					/*
2375 					 * If SMID is 0 for Diag Buffer Post,
2376 					 * this implies that the reply is due to
2377 					 * a release function with a status that
2378 					 * the buffer has been released.  Set
2379 					 * the buffer flags accordingly.
2380 					 */
2381 					rel_rep =
2382 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2383 					if ((le16toh(rel_rep->IOCStatus) &
2384 					    MPI2_IOCSTATUS_MASK) ==
2385 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2386 					{
2387 						pBuffer =
2388 						    &sc->fw_diag_buffer_list[
2389 						    rel_rep->BufferType];
2390 						pBuffer->valid_data = TRUE;
2391 						pBuffer->owned_by_firmware =
2392 						    FALSE;
2393 						pBuffer->immediate = FALSE;
2394 					}
2395 				} else
2396 					mps_dispatch_event(sc, baddr,
2397 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2398 					    reply);
2399 			} else {
2400 				cm = &sc->commands[
2401 				    le16toh(desc->AddressReply.SMID)];
2402 				KASSERT(cm->cm_state == MPS_CM_STATE_INQUEUE,
2403 				    ("command not inqueue\n"));
2404 				cm->cm_state = MPS_CM_STATE_BUSY;
2405 				cm->cm_reply = reply;
2406 				cm->cm_reply_data = le32toh(
2407 				    desc->AddressReply.ReplyFrameAddress);
2408 			}
2409 			break;
2410 		}
2411 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2412 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2413 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2414 		default:
2415 			/* Unhandled */
2416 			mps_dprint(sc, MPS_ERROR, "Unhandled reply 0x%x\n",
2417 			    desc->Default.ReplyFlags);
2418 			cm = NULL;
2419 			break;
2420 		}
2421 
2422 
2423 		if (cm != NULL) {
2424 			// Print Error reply frame
2425 			if (cm->cm_reply)
2426 				mps_display_reply_info(sc,cm->cm_reply);
2427 			mps_complete_command(sc, cm);
2428 		}
2429 
2430 		desc->Words.Low = 0xffffffff;
2431 		desc->Words.High = 0xffffffff;
2432 	}
2433 
2434 	if (pq != sc->replypostindex) {
2435 		mps_dprint(sc, MPS_TRACE, "%s sc %p writing postindex %d\n",
2436 		    __func__, sc, sc->replypostindex);
2437 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2438 		    sc->replypostindex);
2439 	}
2440 
2441 	return;
2442 }
2443 
2444 static void
2445 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
2446     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2447 {
2448 	struct mps_event_handle *eh;
2449 	int event, handled = 0;
2450 
2451 	event = le16toh(reply->Event);
2452 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2453 		if (isset(eh->mask, event)) {
2454 			eh->callback(sc, data, reply);
2455 			handled++;
2456 		}
2457 	}
2458 
2459 	if (handled == 0)
2460 		mps_dprint(sc, MPS_EVENT, "Unhandled event 0x%x\n", le16toh(event));
2461 
2462 	/*
2463 	 * This is the only place that the event/reply should be freed.
2464 	 * Anything wanting to hold onto the event data should have
2465 	 * already copied it into their own storage.
2466 	 */
2467 	mps_free_reply(sc, data);
2468 }
2469 
2470 static void
2471 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
2472 {
2473 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2474 
2475 	if (cm->cm_reply)
2476 		MPS_DPRINT_EVENT(sc, generic,
2477 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2478 
2479 	mps_free_command(sc, cm);
2480 
2481 	/* next, send a port enable */
2482 	mpssas_startup(sc);
2483 }
2484 
2485 /*
2486  * For both register_events and update_events, the caller supplies a bitmap
2487  * of events that it _wants_.  These functions then turn that into a bitmask
2488  * suitable for the controller.
2489  */
2490 int
2491 mps_register_events(struct mps_softc *sc, u32 *mask,
2492     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
2493 {
2494 	struct mps_event_handle *eh;
2495 	int error = 0;
2496 
2497 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
2498 	if(!eh) {
2499 		mps_dprint(sc, MPS_ERROR, "Cannot allocate event memory\n");
2500 		return (ENOMEM);
2501 	}
2502 	eh->callback = cb;
2503 	eh->data = data;
2504 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2505 	if (mask != NULL)
2506 		error = mps_update_events(sc, eh, mask);
2507 	*handle = eh;
2508 
2509 	return (error);
2510 }
2511 
2512 int
2513 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
2514     u32 *mask)
2515 {
2516 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2517 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2518 	struct mps_command *cm;
2519 	int error, i;
2520 
2521 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2522 
2523 	if ((mask != NULL) && (handle != NULL))
2524 		bcopy(mask, &handle->mask[0], sizeof(u32) *
2525 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2526 
2527 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2528 		sc->event_mask[i] = -1;
2529 
2530 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2531 		sc->event_mask[i] &= ~handle->mask[i];
2532 
2533 
2534 	if ((cm = mps_alloc_command(sc)) == NULL)
2535 		return (EBUSY);
2536 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2537 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2538 	evtreq->MsgFlags = 0;
2539 	evtreq->SASBroadcastPrimitiveMasks = 0;
2540 #ifdef MPS_DEBUG_ALL_EVENTS
2541 	{
2542 		u_char fullmask[16];
2543 		memset(fullmask, 0x00, 16);
2544 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2545 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2546 	}
2547 #else
2548         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2549                 evtreq->EventMasks[i] =
2550                     htole32(sc->event_mask[i]);
2551 #endif
2552 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2553 	cm->cm_data = NULL;
2554 
2555 	error = mps_wait_command(sc, &cm, 60, 0);
2556 	if (cm != NULL)
2557 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2558 	if ((reply == NULL) ||
2559 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2560 		error = ENXIO;
2561 
2562 	if (reply)
2563 		MPS_DPRINT_EVENT(sc, generic, reply);
2564 
2565 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
2566 
2567 	if (cm != NULL)
2568 		mps_free_command(sc, cm);
2569 	return (error);
2570 }
2571 
2572 static int
2573 mps_reregister_events(struct mps_softc *sc)
2574 {
2575 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2576 	struct mps_command *cm;
2577 	struct mps_event_handle *eh;
2578 	int error, i;
2579 
2580 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
2581 
2582 	/* first, reregister events */
2583 
2584 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2585 		sc->event_mask[i] = -1;
2586 
2587 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2588 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2589 			sc->event_mask[i] &= ~eh->mask[i];
2590 	}
2591 
2592 	if ((cm = mps_alloc_command(sc)) == NULL)
2593 		return (EBUSY);
2594 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2595 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2596 	evtreq->MsgFlags = 0;
2597 	evtreq->SASBroadcastPrimitiveMasks = 0;
2598 #ifdef MPS_DEBUG_ALL_EVENTS
2599 	{
2600 		u_char fullmask[16];
2601 		memset(fullmask, 0x00, 16);
2602 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
2603 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
2604 	}
2605 #else
2606         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2607                 evtreq->EventMasks[i] =
2608                     htole32(sc->event_mask[i]);
2609 #endif
2610 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2611 	cm->cm_data = NULL;
2612 	cm->cm_complete = mps_reregister_events_complete;
2613 
2614 	error = mps_map_command(sc, cm);
2615 
2616 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__,
2617 	    error);
2618 	return (error);
2619 }
2620 
2621 void
2622 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
2623 {
2624 
2625 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2626 	free(handle, M_MPT2);
2627 }
2628 
2629 /*
2630  * Add a chain element as the next SGE for the specified command.
2631  * Reset cm_sge and cm_sgesize to indicate all the available space.
2632  */
2633 static int
2634 mps_add_chain(struct mps_command *cm)
2635 {
2636 	MPI2_SGE_CHAIN32 *sgc;
2637 	struct mps_chain *chain;
2638 	u_int space;
2639 
2640 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2641 		panic("MPS: Need SGE Error Code\n");
2642 
2643 	chain = mps_alloc_chain(cm->cm_sc);
2644 	if (chain == NULL)
2645 		return (ENOBUFS);
2646 
2647 	space = cm->cm_sc->reqframesz;
2648 
2649 	/*
2650 	 * Note: a double-linked list is used to make it easier to
2651 	 * walk for debugging.
2652 	 */
2653 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2654 
2655 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
2656 	sgc->Length = htole16(space);
2657 	sgc->NextChainOffset = 0;
2658 	/* TODO Looks like bug in Setting sgc->Flags.
2659 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
2660 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
2661 	 *	This is fine.. because we are not using simple element. In case of
2662 	 *	MPI2_SGE_CHAIN32, we have separate Length and Flags feild.
2663  	 */
2664 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
2665 	sgc->Address = htole32(chain->chain_busaddr);
2666 
2667 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2668 	cm->cm_sglsize = space;
2669 	return (0);
2670 }
2671 
2672 /*
2673  * Add one scatter-gather element (chain, simple, transaction context)
2674  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2675  * cm_sge as the remaining size and pointer to the next SGE to fill
2676  * in, respectively.
2677  */
2678 int
2679 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2680 {
2681 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2682 	MPI2_SGE_SIMPLE64 *sge = sgep;
2683 	int error, type;
2684 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2685 
2686 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2687 
2688 #ifdef INVARIANTS
2689 	switch (type) {
2690 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2691 		if (len != tc->DetailsLength + 4)
2692 			panic("TC %p length %u or %zu?", tc,
2693 			    tc->DetailsLength + 4, len);
2694 		}
2695 		break;
2696 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2697 		/* Driver only uses 32-bit chain elements */
2698 		if (len != MPS_SGC_SIZE)
2699 			panic("CHAIN %p length %u or %zu?", sgep,
2700 			    MPS_SGC_SIZE, len);
2701 		break;
2702 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2703 		/* Driver only uses 64-bit SGE simple elements */
2704 		if (len != MPS_SGE64_SIZE)
2705 			panic("SGE simple %p length %u or %zu?", sge,
2706 			    MPS_SGE64_SIZE, len);
2707 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2708 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2709 			panic("SGE simple %p not marked 64-bit?", sge);
2710 
2711 		break;
2712 	default:
2713 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2714 	}
2715 #endif
2716 
2717 	/*
2718 	 * case 1: 1 more segment, enough room for it
2719 	 * case 2: 2 more segments, enough room for both
2720 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2721 	 * case 4: >=1 more segment, enough room for only a chain
2722 	 * case 5: >=1 more segment, no room for anything (error)
2723          */
2724 
2725 	/*
2726 	 * There should be room for at least a chain element, or this
2727 	 * code is buggy.  Case (5).
2728 	 */
2729 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2730 		panic("MPS: Need SGE Error Code\n");
2731 
2732 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2733 		/*
2734 		 * 1 or more segment, enough room for only a chain.
2735 		 * Hope the previous element wasn't a Simple entry
2736 		 * that needed to be marked with
2737 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2738 		 */
2739 		if ((error = mps_add_chain(cm)) != 0)
2740 			return (error);
2741 	}
2742 
2743 	if (segsleft >= 2 &&
2744 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2745 		/*
2746 		 * There are 2 or more segments left to add, and only
2747 		 * enough room for 1 and a chain.  Case (3).
2748 		 *
2749 		 * Mark as last element in this chain if necessary.
2750 		 */
2751 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2752 			sge->FlagsLength |= htole32(
2753 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2754 		}
2755 
2756 		/*
2757 		 * Add the item then a chain.  Do the chain now,
2758 		 * rather than on the next iteration, to simplify
2759 		 * understanding the code.
2760 		 */
2761 		cm->cm_sglsize -= len;
2762 		bcopy(sgep, cm->cm_sge, len);
2763 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2764 		return (mps_add_chain(cm));
2765 	}
2766 
2767 #ifdef INVARIANTS
2768 	/* Case 1: 1 more segment, enough room for it. */
2769 	if (segsleft == 1 && cm->cm_sglsize < len)
2770 		panic("1 seg left and no room? %u versus %zu",
2771 		    cm->cm_sglsize, len);
2772 
2773 	/* Case 2: 2 more segments, enough room for both */
2774 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2775 		panic("2 segs left and no room? %u versus %zu",
2776 		    cm->cm_sglsize, len);
2777 #endif
2778 
2779 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2780 		/*
2781 		 * If this is a bi-directional request, need to account for that
2782 		 * here.  Save the pre-filled sge values.  These will be used
2783 		 * either for the 2nd SGL or for a single direction SGL.  If
2784 		 * cm_out_len is non-zero, this is a bi-directional request, so
2785 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2786 		 * fill in the IN SGL.  Note that at this time, when filling in
2787 		 * 2 SGL's for a bi-directional request, they both use the same
2788 		 * DMA buffer (same cm command).
2789 		 */
2790 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2791 		saved_address_low = sge->Address.Low;
2792 		saved_address_high = sge->Address.High;
2793 		if (cm->cm_out_len) {
2794 			sge->FlagsLength = htole32(cm->cm_out_len |
2795 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2796 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2797 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2798 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2799 			    MPI2_SGE_FLAGS_SHIFT));
2800 			cm->cm_sglsize -= len;
2801 			bcopy(sgep, cm->cm_sge, len);
2802 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2803 			    + len);
2804 		}
2805 		saved_buf_len |=
2806 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2807 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2808 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2809 		    MPI2_SGE_FLAGS_END_OF_LIST |
2810 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2811 		    MPI2_SGE_FLAGS_SHIFT);
2812 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2813 			saved_buf_len |=
2814 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2815 			    MPI2_SGE_FLAGS_SHIFT);
2816 		} else {
2817 			saved_buf_len |=
2818 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2819 			    MPI2_SGE_FLAGS_SHIFT);
2820 		}
2821 		sge->FlagsLength = htole32(saved_buf_len);
2822 		sge->Address.Low = saved_address_low;
2823 		sge->Address.High = saved_address_high;
2824 	}
2825 
2826 	cm->cm_sglsize -= len;
2827 	bcopy(sgep, cm->cm_sge, len);
2828 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2829 	return (0);
2830 }
2831 
2832 /*
2833  * Add one dma segment to the scatter-gather list for a command.
2834  */
2835 int
2836 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2837     int segsleft)
2838 {
2839 	MPI2_SGE_SIMPLE64 sge;
2840 
2841 	/*
2842 	 * This driver always uses 64-bit address elements for simplicity.
2843 	 */
2844 	bzero(&sge, sizeof(sge));
2845 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2846 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2847 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2848 	mps_from_u64(pa, &sge.Address);
2849 
2850 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2851 }
2852 
2853 static void
2854 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2855 {
2856 	struct mps_softc *sc;
2857 	struct mps_command *cm;
2858 	u_int i, dir, sflags;
2859 
2860 	cm = (struct mps_command *)arg;
2861 	sc = cm->cm_sc;
2862 
2863 	/*
2864 	 * In this case, just print out a warning and let the chip tell the
2865 	 * user they did the wrong thing.
2866 	 */
2867 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2868 		mps_dprint(sc, MPS_ERROR,
2869 			   "%s: warning: busdma returned %d segments, "
2870 			   "more than the %d allowed\n", __func__, nsegs,
2871 			   cm->cm_max_segs);
2872 	}
2873 
2874 	/*
2875 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2876 	 * here.  In that case, both direction flags will be set.
2877 	 */
2878 	sflags = 0;
2879 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2880 		/*
2881 		 * We have to add a special case for SMP passthrough, there
2882 		 * is no easy way to generically handle it.  The first
2883 		 * S/G element is used for the command (therefore the
2884 		 * direction bit needs to be set).  The second one is used
2885 		 * for the reply.  We'll leave it to the caller to make
2886 		 * sure we only have two buffers.
2887 		 */
2888 		/*
2889 		 * Even though the busdma man page says it doesn't make
2890 		 * sense to have both direction flags, it does in this case.
2891 		 * We have one s/g element being accessed in each direction.
2892 		 */
2893 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2894 
2895 		/*
2896 		 * Set the direction flag on the first buffer in the SMP
2897 		 * passthrough request.  We'll clear it for the second one.
2898 		 */
2899 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2900 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2901 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2902 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2903 		dir = BUS_DMASYNC_PREWRITE;
2904 	} else
2905 		dir = BUS_DMASYNC_PREREAD;
2906 
2907 	for (i = 0; i < nsegs; i++) {
2908 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2909 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2910 		}
2911 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2912 		    sflags, nsegs - i);
2913 		if (error != 0) {
2914 			/* Resource shortage, roll back! */
2915 			if (ratecheck(&sc->lastfail, &mps_chainfail_interval))
2916 				mps_dprint(sc, MPS_INFO, "Out of chain frames, "
2917 				    "consider increasing hw.mps.max_chains.\n");
2918 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2919 			mps_complete_command(sc, cm);
2920 			return;
2921 		}
2922 	}
2923 
2924 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2925 	mps_enqueue_request(sc, cm);
2926 
2927 	return;
2928 }
2929 
2930 static void
2931 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2932 	     int error)
2933 {
2934 	mps_data_cb(arg, segs, nsegs, error);
2935 }
2936 
2937 /*
2938  * This is the routine to enqueue commands ansynchronously.
2939  * Note that the only error path here is from bus_dmamap_load(), which can
2940  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2941  * assumed that if you have a command in-hand, then you have enough credits
2942  * to use it.
2943  */
2944 int
2945 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2946 {
2947 	int error = 0;
2948 
2949 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2950 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2951 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2952 	} else if (cm->cm_flags & MPS_CM_FLAGS_USE_CCB) {
2953 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
2954 		    cm->cm_data, mps_data_cb, cm, 0);
2955 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2956 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2957 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2958 	} else {
2959 		/* Add a zero-length element as needed */
2960 		if (cm->cm_sge != NULL)
2961 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2962 		mps_enqueue_request(sc, cm);
2963 	}
2964 
2965 	return (error);
2966 }
2967 
2968 /*
2969  * This is the routine to enqueue commands synchronously.  An error of
2970  * EINPROGRESS from mps_map_command() is ignored since the command will
2971  * be executed and enqueued automatically.  Other errors come from msleep().
2972  */
2973 int
2974 mps_wait_command(struct mps_softc *sc, struct mps_command **cmp, int timeout,
2975     int sleep_flag)
2976 {
2977 	int error, rc;
2978 	struct timeval cur_time, start_time;
2979 	struct mps_command *cm = *cmp;
2980 
2981 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET)
2982 		return  EBUSY;
2983 
2984 	cm->cm_complete = NULL;
2985 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2986 	error = mps_map_command(sc, cm);
2987 	if ((error != 0) && (error != EINPROGRESS))
2988 		return (error);
2989 
2990 	/*
2991 	 * Check for context and wait for 50 mSec at a time until time has
2992 	 * expired or the command has finished.  If msleep can't be used, need
2993 	 * to poll.
2994 	 */
2995 	if (curthread->td_no_sleeping != 0)
2996 		sleep_flag = NO_SLEEP;
2997 	getmicrouptime(&start_time);
2998 	if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP) {
2999 		cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
3000 		error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
3001 		if (error == EWOULDBLOCK) {
3002 			/*
3003 			 * Record the actual elapsed time in the case of a
3004 			 * timeout for the message below.
3005 			 */
3006 			getmicrouptime(&cur_time);
3007 			timevalsub(&cur_time, &start_time);
3008 		}
3009 	} else {
3010 		while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
3011 			mps_intr_locked(sc);
3012 			if (sleep_flag == CAN_SLEEP)
3013 				pause("mpswait", hz/20);
3014 			else
3015 				DELAY(50000);
3016 
3017 			getmicrouptime(&cur_time);
3018 			timevalsub(&cur_time, &start_time);
3019 			if (cur_time.tv_sec > timeout) {
3020 				error = EWOULDBLOCK;
3021 				break;
3022 			}
3023 		}
3024 	}
3025 
3026 	if (error == EWOULDBLOCK) {
3027 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s, timeout=%d,"
3028 		    " elapsed=%jd\n", __func__, timeout,
3029 		    (intmax_t)cur_time.tv_sec);
3030 		rc = mps_reinit(sc);
3031 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3032 		    "failed");
3033 		if (sc->mps_flags & MPS_FLAGS_REALLOCATED) {
3034 			/*
3035 			 * Tell the caller that we freed the command in a
3036 			 * reinit.
3037 			 */
3038 			*cmp = NULL;
3039 		}
3040 		error = ETIMEDOUT;
3041 	}
3042 	return (error);
3043 }
3044 
3045 /*
3046  * The MPT driver had a verbose interface for config pages.  In this driver,
3047  * reduce it to much simpler terms, similar to the Linux driver.
3048  */
3049 int
3050 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
3051 {
3052 	MPI2_CONFIG_REQUEST *req;
3053 	struct mps_command *cm;
3054 	int error;
3055 
3056 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
3057 		return (EBUSY);
3058 	}
3059 
3060 	cm = mps_alloc_command(sc);
3061 	if (cm == NULL) {
3062 		return (EBUSY);
3063 	}
3064 
3065 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3066 	req->Function = MPI2_FUNCTION_CONFIG;
3067 	req->Action = params->action;
3068 	req->SGLFlags = 0;
3069 	req->ChainOffset = 0;
3070 	req->PageAddress = params->page_address;
3071 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3072 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3073 
3074 		hdr = &params->hdr.Ext;
3075 		req->ExtPageType = hdr->ExtPageType;
3076 		req->ExtPageLength = hdr->ExtPageLength;
3077 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3078 		req->Header.PageLength = 0; /* Must be set to zero */
3079 		req->Header.PageNumber = hdr->PageNumber;
3080 		req->Header.PageVersion = hdr->PageVersion;
3081 	} else {
3082 		MPI2_CONFIG_PAGE_HEADER *hdr;
3083 
3084 		hdr = &params->hdr.Struct;
3085 		req->Header.PageType = hdr->PageType;
3086 		req->Header.PageNumber = hdr->PageNumber;
3087 		req->Header.PageLength = hdr->PageLength;
3088 		req->Header.PageVersion = hdr->PageVersion;
3089 	}
3090 
3091 	cm->cm_data = params->buffer;
3092 	cm->cm_length = params->length;
3093 	if (cm->cm_data != NULL) {
3094 		cm->cm_sge = &req->PageBufferSGE;
3095 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3096 		cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
3097 	} else
3098 		cm->cm_sge = NULL;
3099 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3100 
3101 	cm->cm_complete_data = params;
3102 	if (params->callback != NULL) {
3103 		cm->cm_complete = mps_config_complete;
3104 		return (mps_map_command(sc, cm));
3105 	} else {
3106 		error = mps_wait_command(sc, &cm, 0, CAN_SLEEP);
3107 		if (error) {
3108 			mps_dprint(sc, MPS_FAULT,
3109 			    "Error %d reading config page\n", error);
3110 			if (cm != NULL)
3111 				mps_free_command(sc, cm);
3112 			return (error);
3113 		}
3114 		mps_config_complete(sc, cm);
3115 	}
3116 
3117 	return (0);
3118 }
3119 
3120 int
3121 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
3122 {
3123 	return (EINVAL);
3124 }
3125 
3126 static void
3127 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
3128 {
3129 	MPI2_CONFIG_REPLY *reply;
3130 	struct mps_config_params *params;
3131 
3132 	MPS_FUNCTRACE(sc);
3133 	params = cm->cm_complete_data;
3134 
3135 	if (cm->cm_data != NULL) {
3136 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3137 		    BUS_DMASYNC_POSTREAD);
3138 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3139 	}
3140 
3141 	/*
3142 	 * XXX KDM need to do more error recovery?  This results in the
3143 	 * device in question not getting probed.
3144 	 */
3145 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
3146 		params->status = MPI2_IOCSTATUS_BUSY;
3147 		goto done;
3148 	}
3149 
3150 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3151 	if (reply == NULL) {
3152 		params->status = MPI2_IOCSTATUS_BUSY;
3153 		goto done;
3154 	}
3155 	params->status = reply->IOCStatus;
3156 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3157 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3158 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3159 		params->hdr.Ext.PageType = reply->Header.PageType;
3160 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3161 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3162 	} else {
3163 		params->hdr.Struct.PageType = reply->Header.PageType;
3164 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3165 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3166 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3167 	}
3168 
3169 done:
3170 	mps_free_command(sc, cm);
3171 	if (params->callback != NULL)
3172 		params->callback(sc, params);
3173 
3174 	return;
3175 }
3176