xref: /freebsd/sys/dev/mpr/mpr.c (revision 68b2efbd3b74f0d45bbbf07cef5408e455eefbd1)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2016 Avago Technologies
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
29  *
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 /* Communications core for Avago Technologies (LSI) MPT3 */
36 
37 /* TODO Move headers to mprvar */
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/selinfo.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/bus.h>
47 #include <sys/conf.h>
48 #include <sys/bio.h>
49 #include <sys/malloc.h>
50 #include <sys/uio.h>
51 #include <sys/sysctl.h>
52 #include <sys/smp.h>
53 #include <sys/queue.h>
54 #include <sys/kthread.h>
55 #include <sys/taskqueue.h>
56 #include <sys/endian.h>
57 #include <sys/eventhandler.h>
58 #include <sys/sbuf.h>
59 
60 #include <machine/bus.h>
61 #include <machine/resource.h>
62 #include <sys/rman.h>
63 #include <sys/proc.h>
64 
65 #include <dev/pci/pcivar.h>
66 
67 #include <cam/cam.h>
68 #include <cam/cam_ccb.h>
69 #include <cam/scsi/scsi_all.h>
70 
71 #include <dev/mpr/mpi/mpi2_type.h>
72 #include <dev/mpr/mpi/mpi2.h>
73 #include <dev/mpr/mpi/mpi2_ioc.h>
74 #include <dev/mpr/mpi/mpi2_sas.h>
75 #include <dev/mpr/mpi/mpi2_pci.h>
76 #include <dev/mpr/mpi/mpi2_cnfg.h>
77 #include <dev/mpr/mpi/mpi2_init.h>
78 #include <dev/mpr/mpi/mpi2_tool.h>
79 #include <dev/mpr/mpr_ioctl.h>
80 #include <dev/mpr/mprvar.h>
81 #include <dev/mpr/mpr_table.h>
82 #include <dev/mpr/mpr_sas.h>
83 
84 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
85 static int mpr_init_queues(struct mpr_softc *sc);
86 static void mpr_resize_queues(struct mpr_softc *sc);
87 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
88 static int mpr_transition_operational(struct mpr_softc *sc);
89 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
90 static void mpr_iocfacts_free(struct mpr_softc *sc);
91 static void mpr_startup(void *arg);
92 static int mpr_send_iocinit(struct mpr_softc *sc);
93 static int mpr_alloc_queues(struct mpr_softc *sc);
94 static int mpr_alloc_hw_queues(struct mpr_softc *sc);
95 static int mpr_alloc_replies(struct mpr_softc *sc);
96 static int mpr_alloc_requests(struct mpr_softc *sc);
97 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc);
98 static int mpr_attach_log(struct mpr_softc *sc);
99 static __inline void mpr_complete_command(struct mpr_softc *sc,
100     struct mpr_command *cm);
101 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
102     MPI2_EVENT_NOTIFICATION_REPLY *reply);
103 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
104 static void mpr_periodic(void *);
105 static int mpr_reregister_events(struct mpr_softc *sc);
106 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
107 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
108 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
109 static int mpr_debug_sysctl(SYSCTL_HANDLER_ARGS);
110 static void mpr_parse_debug(struct mpr_softc *sc, char *list);
111 
112 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD, 0, "MPR Driver Parameters");
113 
114 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory");
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 /*
123  * Added this union to smoothly convert le64toh cm->cm_desc.Words.
124  * Compiler only supports uint64_t to be passed as an argument.
125  * Otherwise it will throw this error:
126  * "aggregate value used where an integer was expected"
127  */
128 typedef union _reply_descriptor {
129         u64 word;
130         struct {
131                 u32 low;
132                 u32 high;
133         } u;
134 } reply_descriptor, request_descriptor;
135 
136 /* Rate limit chain-fail messages to 1 per minute */
137 static struct timeval mpr_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 mpr_mtx
146  * is not hold by driver.
147  */
148 static int
149 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag)
150 {
151 	uint32_t reg;
152 	int i, error, tries = 0;
153 	uint8_t first_wait_done = FALSE;
154 
155 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
156 
157 	/* Clear any pending interrupts */
158 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
159 
160 	/*
161 	 * Force NO_SLEEP for threads prohibited to sleep
162  	 * e.a Thread from interrupt handler are prohibited to sleep.
163  	 */
164 #if __FreeBSD_version >= 1000029
165 	if (curthread->td_no_sleeping)
166 #else //__FreeBSD_version < 1000029
167 	if (curthread->td_pflags & TDP_NOSLEEPING)
168 #endif //__FreeBSD_version >= 1000029
169 		sleep_flag = NO_SLEEP;
170 
171 	mpr_dprint(sc, MPR_INIT, "sequence start, sleep_flag=%d\n", sleep_flag);
172 	/* Push the magic sequence */
173 	error = ETIMEDOUT;
174 	while (tries++ < 20) {
175 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
176 			mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
177 			    mpt2_reset_magic[i]);
178 
179 		/* wait 100 msec */
180 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
181 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
182 			    "mprdiag", hz/10);
183 		else if (sleep_flag == CAN_SLEEP)
184 			pause("mprdiag", hz/10);
185 		else
186 			DELAY(100 * 1000);
187 
188 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
189 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
190 			error = 0;
191 			break;
192 		}
193 	}
194 	if (error) {
195 		mpr_dprint(sc, MPR_INIT, "sequence failed, error=%d, exit\n",
196 		    error);
197 		return (error);
198 	}
199 
200 	/* Send the actual reset.  XXX need to refresh the reg? */
201 	reg |= MPI2_DIAG_RESET_ADAPTER;
202 	mpr_dprint(sc, MPR_INIT, "sequence success, sending reset, reg= 0x%x\n",
203 	    reg);
204 	mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
205 
206 	/* Wait up to 300 seconds in 50ms intervals */
207 	error = ETIMEDOUT;
208 	for (i = 0; i < 6000; i++) {
209 		/*
210 		 * Wait 50 msec. If this is the first time through, wait 256
211 		 * msec to satisfy Diag Reset timing requirements.
212 		 */
213 		if (first_wait_done) {
214 			if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
215 				msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
216 				    "mprdiag", hz/20);
217 			else if (sleep_flag == CAN_SLEEP)
218 				pause("mprdiag", hz/20);
219 			else
220 				DELAY(50 * 1000);
221 		} else {
222 			DELAY(256 * 1000);
223 			first_wait_done = TRUE;
224 		}
225 		/*
226 		 * Check for the RESET_ADAPTER bit to be cleared first, then
227 		 * wait for the RESET state to be cleared, which takes a little
228 		 * longer.
229 		 */
230 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
231 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
232 			continue;
233 		}
234 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
235 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
236 			error = 0;
237 			break;
238 		}
239 	}
240 	if (error) {
241 		mpr_dprint(sc, MPR_INIT, "reset failed, error= %d, exit\n",
242 		    error);
243 		return (error);
244 	}
245 
246 	mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
247 	mpr_dprint(sc, MPR_INIT, "diag reset success, exit\n");
248 
249 	return (0);
250 }
251 
252 static int
253 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag)
254 {
255 	int error;
256 
257 	MPR_FUNCTRACE(sc);
258 
259 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
260 
261 	error = 0;
262 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
263 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
264 	    MPI2_DOORBELL_FUNCTION_SHIFT);
265 
266 	if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) {
267 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
268 		    "Doorbell handshake failed\n");
269 		error = ETIMEDOUT;
270 	}
271 
272 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
273 	return (error);
274 }
275 
276 static int
277 mpr_transition_ready(struct mpr_softc *sc)
278 {
279 	uint32_t reg, state;
280 	int error, tries = 0;
281 	int sleep_flags;
282 
283 	MPR_FUNCTRACE(sc);
284 	/* If we are in attach call, do not sleep */
285 	sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE)
286 	    ? CAN_SLEEP : NO_SLEEP;
287 
288 	error = 0;
289 
290 	mpr_dprint(sc, MPR_INIT, "%s entered, sleep_flags= %d\n",
291 	    __func__, sleep_flags);
292 
293 	while (tries++ < 1200) {
294 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
295 		mpr_dprint(sc, MPR_INIT, "  Doorbell= 0x%x\n", reg);
296 
297 		/*
298 		 * Ensure the IOC is ready to talk.  If it's not, try
299 		 * resetting it.
300 		 */
301 		if (reg & MPI2_DOORBELL_USED) {
302 			mpr_dprint(sc, MPR_INIT, "  Not ready, sending diag "
303 			    "reset\n");
304 			mpr_diag_reset(sc, sleep_flags);
305 			DELAY(50000);
306 			continue;
307 		}
308 
309 		/* Is the adapter owned by another peer? */
310 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
311 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
312 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC is under the "
313 			    "control of another peer host, aborting "
314 			    "initialization.\n");
315 			error = ENXIO;
316 			break;
317 		}
318 
319 		state = reg & MPI2_IOC_STATE_MASK;
320 		if (state == MPI2_IOC_STATE_READY) {
321 			/* Ready to go! */
322 			error = 0;
323 			break;
324 		} else if (state == MPI2_IOC_STATE_FAULT) {
325 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC in fault "
326 			    "state 0x%x, resetting\n",
327 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
328 			mpr_diag_reset(sc, sleep_flags);
329 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
330 			/* Need to take ownership */
331 			mpr_message_unit_reset(sc, sleep_flags);
332 		} else if (state == MPI2_IOC_STATE_RESET) {
333 			/* Wait a bit, IOC might be in transition */
334 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
335 			    "IOC in unexpected reset state\n");
336 		} else {
337 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
338 			    "IOC in unknown state 0x%x\n", state);
339 			error = EINVAL;
340 			break;
341 		}
342 
343 		/* Wait 50ms for things to settle down. */
344 		DELAY(50000);
345 	}
346 
347 	if (error)
348 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
349 		    "Cannot transition IOC to ready\n");
350 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
351 	return (error);
352 }
353 
354 static int
355 mpr_transition_operational(struct mpr_softc *sc)
356 {
357 	uint32_t reg, state;
358 	int error;
359 
360 	MPR_FUNCTRACE(sc);
361 
362 	error = 0;
363 	reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
364 	mpr_dprint(sc, MPR_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
365 
366 	state = reg & MPI2_IOC_STATE_MASK;
367 	if (state != MPI2_IOC_STATE_READY) {
368 		mpr_dprint(sc, MPR_INIT, "IOC not ready\n");
369 		if ((error = mpr_transition_ready(sc)) != 0) {
370 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
371 			    "failed to transition ready, exit\n");
372 			return (error);
373 		}
374 	}
375 
376 	error = mpr_send_iocinit(sc);
377 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
378 
379 	return (error);
380 }
381 
382 static void
383 mpr_resize_queues(struct mpr_softc *sc)
384 {
385 	int reqcr, prireqcr;
386 
387 	/*
388 	 * Size the queues. Since the reply queues always need one free
389 	 * entry, we'll deduct one reply message here.  The LSI documents
390 	 * suggest instead to add a count to the request queue, but I think
391 	 * that it's better to deduct from reply queue.
392 	 */
393 	prireqcr = MAX(1, sc->max_prireqframes);
394 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
395 
396 	reqcr = MAX(2, sc->max_reqframes);
397 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
398 
399 	sc->num_reqs = prireqcr + reqcr;
400 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
401 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
402 
403 	/*
404 	 * Figure out the number of MSIx-based queues.  If the firmware or
405 	 * user has done something crazy and not allowed enough credit for
406 	 * the queues to be useful then don't enable multi-queue.
407 	 */
408 	if (sc->facts->MaxMSIxVectors < 2)
409 		sc->msi_msgs = 1;
410 
411 	if (sc->msi_msgs > 1) {
412 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
413 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
414 		if (sc->num_reqs / sc->msi_msgs < 2)
415 			sc->msi_msgs = 1;
416 	}
417 
418 	mpr_dprint(sc, MPR_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
419 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
420 }
421 
422 /*
423  * This is called during attach and when re-initializing due to a Diag Reset.
424  * IOC Facts is used to allocate many of the structures needed by the driver.
425  * If called from attach, de-allocation is not required because the driver has
426  * not allocated any structures yet, but if called from a Diag Reset, previously
427  * allocated structures based on IOC Facts will need to be freed and re-
428  * allocated bases on the latest IOC Facts.
429  */
430 static int
431 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
432 {
433 	int error;
434 	Mpi2IOCFactsReply_t saved_facts;
435 	uint8_t saved_mode, reallocating;
436 
437 	mpr_dprint(sc, MPR_INIT|MPR_TRACE, "%s entered\n", __func__);
438 
439 	/* Save old IOC Facts and then only reallocate if Facts have changed */
440 	if (!attaching) {
441 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
442 	}
443 
444 	/*
445 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
446 	 * a re-initialization and only return the error if attaching so the OS
447 	 * can handle it.
448 	 */
449 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
450 		if (attaching) {
451 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to get "
452 			    "IOC Facts with error %d, exit\n", error);
453 			return (error);
454 		} else {
455 			panic("%s failed to get IOC Facts with error %d\n",
456 			    __func__, error);
457 		}
458 	}
459 
460 	MPR_DPRINT_PAGE(sc, MPR_XINFO, iocfacts, sc->facts);
461 
462 	snprintf(sc->fw_version, sizeof(sc->fw_version),
463 	    "%02d.%02d.%02d.%02d",
464 	    sc->facts->FWVersion.Struct.Major,
465 	    sc->facts->FWVersion.Struct.Minor,
466 	    sc->facts->FWVersion.Struct.Unit,
467 	    sc->facts->FWVersion.Struct.Dev);
468 
469 	mpr_dprint(sc, MPR_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
470 	    MPR_DRIVER_VERSION);
471 	mpr_dprint(sc, MPR_INFO,
472 	    "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
473 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
474 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
475 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc"
476 	    "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV");
477 
478 	/*
479 	 * If the chip doesn't support event replay then a hard reset will be
480 	 * required to trigger a full discovery.  Do the reset here then
481 	 * retransition to Ready.  A hard reset might have already been done,
482 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
483 	 * for a Diag Reset.
484 	 */
485 	if (attaching && ((sc->facts->IOCCapabilities &
486 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
487 		mpr_dprint(sc, MPR_INIT, "No event replay, resetting\n");
488 		mpr_diag_reset(sc, NO_SLEEP);
489 		if ((error = mpr_transition_ready(sc)) != 0) {
490 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
491 			    "transition to ready with error %d, exit\n",
492 			    error);
493 			return (error);
494 		}
495 	}
496 
497 	/*
498 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
499 	 * changed from the previous IOC Facts, log a warning, but only if
500 	 * checking this after a Diag Reset and not during attach.
501 	 */
502 	saved_mode = sc->ir_firmware;
503 	if (sc->facts->IOCCapabilities &
504 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
505 		sc->ir_firmware = 1;
506 	if (!attaching) {
507 		if (sc->ir_firmware != saved_mode) {
508 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "new IR/IT mode "
509 			    "in IOC Facts does not match previous mode\n");
510 		}
511 	}
512 
513 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
514 	reallocating = FALSE;
515 	sc->mpr_flags &= ~MPR_FLAGS_REALLOCATED;
516 
517 	if ((!attaching) &&
518 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
519 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
520 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
521 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
522 	    (saved_facts.ProductID != sc->facts->ProductID) ||
523 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
524 	    (saved_facts.IOCRequestFrameSize !=
525 	    sc->facts->IOCRequestFrameSize) ||
526 	    (saved_facts.IOCMaxChainSegmentSize !=
527 	    sc->facts->IOCMaxChainSegmentSize) ||
528 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
529 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
530 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
531 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
532 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
533 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
534 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
535 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
536 	    (saved_facts.MaxPersistentEntries !=
537 	    sc->facts->MaxPersistentEntries))) {
538 		reallocating = TRUE;
539 
540 		/* Record that we reallocated everything */
541 		sc->mpr_flags |= MPR_FLAGS_REALLOCATED;
542 	}
543 
544 	/*
545 	 * Some things should be done if attaching or re-allocating after a Diag
546 	 * Reset, but are not needed after a Diag Reset if the FW has not
547 	 * changed.
548 	 */
549 	if (attaching || reallocating) {
550 		/*
551 		 * Check if controller supports FW diag buffers and set flag to
552 		 * enable each type.
553 		 */
554 		if (sc->facts->IOCCapabilities &
555 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
556 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
557 			    enabled = TRUE;
558 		if (sc->facts->IOCCapabilities &
559 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
560 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
561 			    enabled = TRUE;
562 		if (sc->facts->IOCCapabilities &
563 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
564 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
565 			    enabled = TRUE;
566 
567 		/*
568 		 * Set flags for some supported items.
569 		 */
570 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
571 			sc->eedp_enabled = TRUE;
572 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
573 			sc->control_TLR = TRUE;
574 		if (sc->facts->IOCCapabilities &
575 		    MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ)
576 			sc->atomic_desc_capable = TRUE;
577 
578 		mpr_resize_queues(sc);
579 
580 		/*
581 		 * Initialize all Tail Queues
582 		 */
583 		TAILQ_INIT(&sc->req_list);
584 		TAILQ_INIT(&sc->high_priority_req_list);
585 		TAILQ_INIT(&sc->chain_list);
586 		TAILQ_INIT(&sc->prp_page_list);
587 		TAILQ_INIT(&sc->tm_list);
588 	}
589 
590 	/*
591 	 * If doing a Diag Reset and the FW is significantly different
592 	 * (reallocating will be set above in IOC Facts comparison), then all
593 	 * buffers based on the IOC Facts will need to be freed before they are
594 	 * reallocated.
595 	 */
596 	if (reallocating) {
597 		mpr_iocfacts_free(sc);
598 		mprsas_realloc_targets(sc, saved_facts.MaxTargets +
599 		    saved_facts.MaxVolumes);
600 	}
601 
602 	/*
603 	 * Any deallocation has been completed.  Now start reallocating
604 	 * if needed.  Will only need to reallocate if attaching or if the new
605 	 * IOC Facts are different from the previous IOC Facts after a Diag
606 	 * Reset. Targets have already been allocated above if needed.
607 	 */
608 	error = 0;
609 	while (attaching || reallocating) {
610 		if ((error = mpr_alloc_hw_queues(sc)) != 0)
611 			break;
612 		if ((error = mpr_alloc_replies(sc)) != 0)
613 			break;
614 		if ((error = mpr_alloc_requests(sc)) != 0)
615 			break;
616 		if ((error = mpr_alloc_queues(sc)) != 0)
617 			break;
618 		break;
619 	}
620 	if (error) {
621 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
622 		    "Failed to alloc queues with error %d\n", error);
623 		mpr_free(sc);
624 		return (error);
625 	}
626 
627 	/* Always initialize the queues */
628 	bzero(sc->free_queue, sc->fqdepth * 4);
629 	mpr_init_queues(sc);
630 
631 	/*
632 	 * Always get the chip out of the reset state, but only panic if not
633 	 * attaching.  If attaching and there is an error, that is handled by
634 	 * the OS.
635 	 */
636 	error = mpr_transition_operational(sc);
637 	if (error != 0) {
638 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
639 		    "transition to operational with error %d\n", error);
640 		mpr_free(sc);
641 		return (error);
642 	}
643 
644 	/*
645 	 * Finish the queue initialization.
646 	 * These are set here instead of in mpr_init_queues() because the
647 	 * IOC resets these values during the state transition in
648 	 * mpr_transition_operational().  The free index is set to 1
649 	 * because the corresponding index in the IOC is set to 0, and the
650 	 * IOC treats the queues as full if both are set to the same value.
651 	 * Hence the reason that the queue can't hold all of the possible
652 	 * replies.
653 	 */
654 	sc->replypostindex = 0;
655 	mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
656 	mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
657 
658 	/*
659 	 * Attach the subsystems so they can prepare their event masks.
660 	 * XXX Should be dynamic so that IM/IR and user modules can attach
661 	 */
662 	error = 0;
663 	while (attaching) {
664 		mpr_dprint(sc, MPR_INIT, "Attaching subsystems\n");
665 		if ((error = mpr_attach_log(sc)) != 0)
666 			break;
667 		if ((error = mpr_attach_sas(sc)) != 0)
668 			break;
669 		if ((error = mpr_attach_user(sc)) != 0)
670 			break;
671 		break;
672 	}
673 	if (error) {
674 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
675 		    "Failed to attach all subsystems: error %d\n", error);
676 		mpr_free(sc);
677 		return (error);
678 	}
679 
680 	if ((error = mpr_pci_setup_interrupts(sc)) != 0) {
681 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
682 		    "Failed to setup interrupts\n");
683 		mpr_free(sc);
684 		return (error);
685 	}
686 
687 	return (error);
688 }
689 
690 /*
691  * This is called if memory is being free (during detach for example) and when
692  * buffers need to be reallocated due to a Diag Reset.
693  */
694 static void
695 mpr_iocfacts_free(struct mpr_softc *sc)
696 {
697 	struct mpr_command *cm;
698 	int i;
699 
700 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
701 
702 	if (sc->free_busaddr != 0)
703 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
704 	if (sc->free_queue != NULL)
705 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
706 		    sc->queues_map);
707 	if (sc->queues_dmat != NULL)
708 		bus_dma_tag_destroy(sc->queues_dmat);
709 
710 	if (sc->chain_busaddr != 0)
711 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
712 	if (sc->chain_frames != NULL)
713 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
714 		    sc->chain_map);
715 	if (sc->chain_dmat != NULL)
716 		bus_dma_tag_destroy(sc->chain_dmat);
717 
718 	if (sc->sense_busaddr != 0)
719 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
720 	if (sc->sense_frames != NULL)
721 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
722 		    sc->sense_map);
723 	if (sc->sense_dmat != NULL)
724 		bus_dma_tag_destroy(sc->sense_dmat);
725 
726 	if (sc->prp_page_busaddr != 0)
727 		bus_dmamap_unload(sc->prp_page_dmat, sc->prp_page_map);
728 	if (sc->prp_pages != NULL)
729 		bus_dmamem_free(sc->prp_page_dmat, sc->prp_pages,
730 		    sc->prp_page_map);
731 	if (sc->prp_page_dmat != NULL)
732 		bus_dma_tag_destroy(sc->prp_page_dmat);
733 
734 	if (sc->reply_busaddr != 0)
735 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
736 	if (sc->reply_frames != NULL)
737 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
738 		    sc->reply_map);
739 	if (sc->reply_dmat != NULL)
740 		bus_dma_tag_destroy(sc->reply_dmat);
741 
742 	if (sc->req_busaddr != 0)
743 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
744 	if (sc->req_frames != NULL)
745 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
746 	if (sc->req_dmat != NULL)
747 		bus_dma_tag_destroy(sc->req_dmat);
748 
749 	if (sc->chains != NULL)
750 		free(sc->chains, M_MPR);
751 	if (sc->prps != NULL)
752 		free(sc->prps, M_MPR);
753 	if (sc->commands != NULL) {
754 		for (i = 1; i < sc->num_reqs; i++) {
755 			cm = &sc->commands[i];
756 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
757 		}
758 		free(sc->commands, M_MPR);
759 	}
760 	if (sc->buffer_dmat != NULL)
761 		bus_dma_tag_destroy(sc->buffer_dmat);
762 
763 	mpr_pci_free_interrupts(sc);
764 	free(sc->queues, M_MPR);
765 	sc->queues = NULL;
766 }
767 
768 /*
769  * The terms diag reset and hard reset are used interchangeably in the MPI
770  * docs to mean resetting the controller chip.  In this code diag reset
771  * cleans everything up, and the hard reset function just sends the reset
772  * sequence to the chip.  This should probably be refactored so that every
773  * subsystem gets a reset notification of some sort, and can clean up
774  * appropriately.
775  */
776 int
777 mpr_reinit(struct mpr_softc *sc)
778 {
779 	int error;
780 	struct mprsas_softc *sassc;
781 
782 	sassc = sc->sassc;
783 
784 	MPR_FUNCTRACE(sc);
785 
786 	mtx_assert(&sc->mpr_mtx, MA_OWNED);
787 
788 	mpr_dprint(sc, MPR_INIT|MPR_INFO, "Reinitializing controller\n");
789 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) {
790 		mpr_dprint(sc, MPR_INIT, "Reset already in progress\n");
791 		return 0;
792 	}
793 
794 	/*
795 	 * Make sure the completion callbacks can recognize they're getting
796 	 * a NULL cm_reply due to a reset.
797 	 */
798 	sc->mpr_flags |= MPR_FLAGS_DIAGRESET;
799 
800 	/*
801 	 * Mask interrupts here.
802 	 */
803 	mpr_dprint(sc, MPR_INIT, "Masking interrupts and resetting\n");
804 	mpr_mask_intr(sc);
805 
806 	error = mpr_diag_reset(sc, CAN_SLEEP);
807 	if (error != 0) {
808 		panic("%s hard reset failed with error %d\n", __func__, error);
809 	}
810 
811 	/* Restore the PCI state, including the MSI-X registers */
812 	mpr_pci_restore(sc);
813 
814 	/* Give the I/O subsystem special priority to get itself prepared */
815 	mprsas_handle_reinit(sc);
816 
817 	/*
818 	 * Get IOC Facts and allocate all structures based on this information.
819 	 * The attach function will also call mpr_iocfacts_allocate at startup.
820 	 * If relevant values have changed in IOC Facts, this function will free
821 	 * all of the memory based on IOC Facts and reallocate that memory.
822 	 */
823 	if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) {
824 		panic("%s IOC Facts based allocation failed with error %d\n",
825 		    __func__, error);
826 	}
827 
828 	/*
829 	 * Mapping structures will be re-allocated after getting IOC Page8, so
830 	 * free these structures here.
831 	 */
832 	mpr_mapping_exit(sc);
833 
834 	/*
835 	 * The static page function currently read is IOC Page8.  Others can be
836 	 * added in future.  It's possible that the values in IOC Page8 have
837 	 * changed after a Diag Reset due to user modification, so always read
838 	 * these.  Interrupts are masked, so unmask them before getting config
839 	 * pages.
840 	 */
841 	mpr_unmask_intr(sc);
842 	sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET;
843 	mpr_base_static_config_pages(sc);
844 
845 	/*
846 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
847 	 * mapping tables.
848 	 */
849 	mpr_mapping_initialize(sc);
850 
851 	/*
852 	 * Restart will reload the event masks clobbered by the reset, and
853 	 * then enable the port.
854 	 */
855 	mpr_reregister_events(sc);
856 
857 	/* the end of discovery will release the simq, so we're done. */
858 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Finished sc %p post %u free %u\n",
859 	    sc, sc->replypostindex, sc->replyfreeindex);
860 	mprsas_release_simq_reinit(sassc);
861 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
862 
863 	return 0;
864 }
865 
866 /* Wait for the chip to ACK a word that we've put into its FIFO
867  * Wait for <timeout> seconds. In single loop wait for busy loop
868  * for 500 microseconds.
869  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
870  * */
871 static int
872 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag)
873 {
874 	u32 cntdn, count;
875 	u32 int_status;
876 	u32 doorbell;
877 
878 	count = 0;
879 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
880 	do {
881 		int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
882 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
883 			mpr_dprint(sc, MPR_TRACE, "%s: successful count(%d), "
884 			    "timeout(%d)\n", __func__, count, timeout);
885 			return 0;
886 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
887 			doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
888 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
889 			    MPI2_IOC_STATE_FAULT) {
890 				mpr_dprint(sc, MPR_FAULT,
891 				    "fault_state(0x%04x)!\n", doorbell);
892 				return (EFAULT);
893 			}
894 		} else if (int_status == 0xFFFFFFFF)
895 			goto out;
896 
897 		/*
898 		 * If it can sleep, sleep for 1 milisecond, else busy loop for
899  		 * 0.5 milisecond
900 		 */
901 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
902 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba",
903 			    hz/1000);
904 		else if (sleep_flag == CAN_SLEEP)
905 			pause("mprdba", hz/1000);
906 		else
907 			DELAY(500);
908 		count++;
909 	} while (--cntdn);
910 
911 out:
912 	mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), "
913 		"int_status(%x)!\n", __func__, count, int_status);
914 	return (ETIMEDOUT);
915 }
916 
917 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
918 static int
919 mpr_wait_db_int(struct mpr_softc *sc)
920 {
921 	int retry;
922 
923 	for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) {
924 		if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
925 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
926 			return (0);
927 		DELAY(2000);
928 	}
929 	return (ETIMEDOUT);
930 }
931 
932 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
933 static int
934 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
935     int req_sz, int reply_sz, int timeout)
936 {
937 	uint32_t *data32;
938 	uint16_t *data16;
939 	int i, count, ioc_sz, residual;
940 	int sleep_flags = CAN_SLEEP;
941 
942 #if __FreeBSD_version >= 1000029
943 	if (curthread->td_no_sleeping)
944 #else //__FreeBSD_version < 1000029
945 	if (curthread->td_pflags & TDP_NOSLEEPING)
946 #endif //__FreeBSD_version >= 1000029
947 		sleep_flags = NO_SLEEP;
948 
949 	/* Step 1 */
950 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
951 
952 	/* Step 2 */
953 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
954 		return (EBUSY);
955 
956 	/* Step 3
957 	 * Announce that a message is coming through the doorbell.  Messages
958 	 * are pushed at 32bit words, so round up if needed.
959 	 */
960 	count = (req_sz + 3) / 4;
961 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
962 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
963 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
964 
965 	/* Step 4 */
966 	if (mpr_wait_db_int(sc) ||
967 	    (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
968 		mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n");
969 		return (ENXIO);
970 	}
971 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
972 	if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
973 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n");
974 		return (ENXIO);
975 	}
976 
977 	/* Step 5 */
978 	/* Clock out the message data synchronously in 32-bit dwords*/
979 	data32 = (uint32_t *)req;
980 	for (i = 0; i < count; i++) {
981 		mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
982 		if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
983 			mpr_dprint(sc, MPR_FAULT,
984 			    "Timeout while writing doorbell\n");
985 			return (ENXIO);
986 		}
987 	}
988 
989 	/* Step 6 */
990 	/* Clock in the reply in 16-bit words.  The total length of the
991 	 * message is always in the 4th byte, so clock out the first 2 words
992 	 * manually, then loop the rest.
993 	 */
994 	data16 = (uint16_t *)reply;
995 	if (mpr_wait_db_int(sc) != 0) {
996 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n");
997 		return (ENXIO);
998 	}
999 	data16[0] =
1000 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1001 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1002 	if (mpr_wait_db_int(sc) != 0) {
1003 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n");
1004 		return (ENXIO);
1005 	}
1006 	data16[1] =
1007 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1008 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1009 
1010 	/* Number of 32bit words in the message */
1011 	ioc_sz = reply->MsgLength;
1012 
1013 	/*
1014 	 * Figure out how many 16bit words to clock in without overrunning.
1015 	 * The precision loss with dividing reply_sz can safely be
1016 	 * ignored because the messages can only be multiples of 32bits.
1017 	 */
1018 	residual = 0;
1019 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1020 	if (count < ioc_sz * 2) {
1021 		residual = ioc_sz * 2 - count;
1022 		mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d "
1023 		    "residual message words\n", residual);
1024 	}
1025 
1026 	for (i = 2; i < count; i++) {
1027 		if (mpr_wait_db_int(sc) != 0) {
1028 			mpr_dprint(sc, MPR_FAULT,
1029 			    "Timeout reading doorbell %d\n", i);
1030 			return (ENXIO);
1031 		}
1032 		data16[i] = mpr_regread(sc, MPI2_DOORBELL_OFFSET) &
1033 		    MPI2_DOORBELL_DATA_MASK;
1034 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1035 	}
1036 
1037 	/*
1038 	 * Pull out residual words that won't fit into the provided buffer.
1039 	 * This keeps the chip from hanging due to a driver programming
1040 	 * error.
1041 	 */
1042 	while (residual--) {
1043 		if (mpr_wait_db_int(sc) != 0) {
1044 			mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n");
1045 			return (ENXIO);
1046 		}
1047 		(void)mpr_regread(sc, MPI2_DOORBELL_OFFSET);
1048 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1049 	}
1050 
1051 	/* Step 7 */
1052 	if (mpr_wait_db_int(sc) != 0) {
1053 		mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n");
1054 		return (ENXIO);
1055 	}
1056 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1057 		mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n");
1058 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1059 
1060 	return (0);
1061 }
1062 
1063 static void
1064 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm)
1065 {
1066 	request_descriptor rd;
1067 
1068 	MPR_FUNCTRACE(sc);
1069 	mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n",
1070 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1071 
1072 	if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags &
1073 	    MPR_FLAGS_SHUTDOWN))
1074 		mtx_assert(&sc->mpr_mtx, MA_OWNED);
1075 
1076 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1077 		sc->io_cmds_highwater++;
1078 
1079 	if (sc->atomic_desc_capable) {
1080 		rd.u.low = cm->cm_desc.Words.Low;
1081 		mpr_regwrite(sc, MPI26_ATOMIC_REQUEST_DESCRIPTOR_POST_OFFSET,
1082 		    rd.u.low);
1083 	} else {
1084 		rd.u.low = cm->cm_desc.Words.Low;
1085 		rd.u.high = cm->cm_desc.Words.High;
1086 		rd.word = htole64(rd.word);
1087 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1088 		    rd.u.low);
1089 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1090 		    rd.u.high);
1091 	}
1092 }
1093 
1094 /*
1095  * Just the FACTS, ma'am.
1096  */
1097 static int
1098 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1099 {
1100 	MPI2_DEFAULT_REPLY *reply;
1101 	MPI2_IOC_FACTS_REQUEST request;
1102 	int error, req_sz, reply_sz;
1103 
1104 	MPR_FUNCTRACE(sc);
1105 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1106 
1107 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1108 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1109 	reply = (MPI2_DEFAULT_REPLY *)facts;
1110 
1111 	bzero(&request, req_sz);
1112 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1113 	error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1114 
1115 	mpr_dprint(sc, MPR_INIT, "%s exit, error= %d\n", __func__, error);
1116 	return (error);
1117 }
1118 
1119 static int
1120 mpr_send_iocinit(struct mpr_softc *sc)
1121 {
1122 	MPI2_IOC_INIT_REQUEST	init;
1123 	MPI2_DEFAULT_REPLY	reply;
1124 	int req_sz, reply_sz, error;
1125 	struct timeval now;
1126 	uint64_t time_in_msec;
1127 
1128 	MPR_FUNCTRACE(sc);
1129 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1130 
1131 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1132 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1133 	bzero(&init, req_sz);
1134 	bzero(&reply, reply_sz);
1135 
1136 	/*
1137 	 * Fill in the init block.  Note that most addresses are
1138 	 * deliberately in the lower 32bits of memory.  This is a micro-
1139 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1140 	 */
1141 	init.Function = MPI2_FUNCTION_IOC_INIT;
1142 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1143 	init.MsgVersion = htole16(MPI2_VERSION);
1144 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1145 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1146 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1147 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1148 	init.SenseBufferAddressHigh = 0;
1149 	init.SystemReplyAddressHigh = 0;
1150 	init.SystemRequestFrameBaseAddress.High = 0;
1151 	init.SystemRequestFrameBaseAddress.Low =
1152 	    htole32((uint32_t)sc->req_busaddr);
1153 	init.ReplyDescriptorPostQueueAddress.High = 0;
1154 	init.ReplyDescriptorPostQueueAddress.Low =
1155 	    htole32((uint32_t)sc->post_busaddr);
1156 	init.ReplyFreeQueueAddress.High = 0;
1157 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1158 	getmicrotime(&now);
1159 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1160 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1161 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1162 	init.HostPageSize = HOST_PAGE_SIZE_4K;
1163 
1164 	error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1165 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1166 		error = ENXIO;
1167 
1168 	mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1169 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
1170 	return (error);
1171 }
1172 
1173 void
1174 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1175 {
1176 	bus_addr_t *addr;
1177 
1178 	addr = arg;
1179 	*addr = segs[0].ds_addr;
1180 }
1181 
1182 static int
1183 mpr_alloc_queues(struct mpr_softc *sc)
1184 {
1185 	struct mpr_queue *q;
1186 	int nq, i;
1187 
1188 	nq = sc->msi_msgs;
1189 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Allocating %d I/O queues\n", nq);
1190 
1191 	sc->queues = malloc(sizeof(struct mpr_queue) * nq, M_MPR,
1192 	     M_NOWAIT|M_ZERO);
1193 	if (sc->queues == NULL)
1194 		return (ENOMEM);
1195 
1196 	for (i = 0; i < nq; i++) {
1197 		q = &sc->queues[i];
1198 		mpr_dprint(sc, MPR_INIT, "Configuring queue %d %p\n", i, q);
1199 		q->sc = sc;
1200 		q->qnum = i;
1201 	}
1202 	return (0);
1203 }
1204 
1205 static int
1206 mpr_alloc_hw_queues(struct mpr_softc *sc)
1207 {
1208 	bus_addr_t queues_busaddr;
1209 	uint8_t *queues;
1210 	int qsize, fqsize, pqsize;
1211 
1212 	/*
1213 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1214 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1215 	 * This queue supplies fresh reply frames for the firmware to use.
1216 	 *
1217 	 * The reply descriptor post queue contains 8 byte entries in
1218 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1219 	 * contains filled-in reply frames sent from the firmware to the host.
1220 	 *
1221 	 * These two queues are allocated together for simplicity.
1222 	 */
1223 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1224 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1225 	fqsize= sc->fqdepth * 4;
1226 	pqsize = sc->pqdepth * 8;
1227 	qsize = fqsize + pqsize;
1228 
1229         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1230 				16, 0,			/* algnmnt, boundary */
1231 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1232 				BUS_SPACE_MAXADDR,	/* highaddr */
1233 				NULL, NULL,		/* filter, filterarg */
1234                                 qsize,			/* maxsize */
1235                                 1,			/* nsegments */
1236                                 qsize,			/* maxsegsize */
1237                                 0,			/* flags */
1238                                 NULL, NULL,		/* lockfunc, lockarg */
1239                                 &sc->queues_dmat)) {
1240 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues DMA tag\n");
1241 		return (ENOMEM);
1242         }
1243         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1244 	    &sc->queues_map)) {
1245 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues memory\n");
1246 		return (ENOMEM);
1247         }
1248         bzero(queues, qsize);
1249         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1250 	    mpr_memaddr_cb, &queues_busaddr, 0);
1251 
1252 	sc->free_queue = (uint32_t *)queues;
1253 	sc->free_busaddr = queues_busaddr;
1254 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1255 	sc->post_busaddr = queues_busaddr + fqsize;
1256 
1257 	return (0);
1258 }
1259 
1260 static int
1261 mpr_alloc_replies(struct mpr_softc *sc)
1262 {
1263 	int rsize, num_replies;
1264 
1265 	/*
1266 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1267 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1268 	 * replies can be used at once.
1269 	 */
1270 	num_replies = max(sc->fqdepth, sc->num_replies);
1271 
1272 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1273         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1274 				4, 0,			/* algnmnt, boundary */
1275 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1276 				BUS_SPACE_MAXADDR,	/* highaddr */
1277 				NULL, NULL,		/* filter, filterarg */
1278                                 rsize,			/* maxsize */
1279                                 1,			/* nsegments */
1280                                 rsize,			/* maxsegsize */
1281                                 0,			/* flags */
1282                                 NULL, NULL,		/* lockfunc, lockarg */
1283                                 &sc->reply_dmat)) {
1284 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies DMA tag\n");
1285 		return (ENOMEM);
1286         }
1287         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1288 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1289 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies memory\n");
1290 		return (ENOMEM);
1291         }
1292         bzero(sc->reply_frames, rsize);
1293         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1294 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1295 
1296 	return (0);
1297 }
1298 
1299 static int
1300 mpr_alloc_requests(struct mpr_softc *sc)
1301 {
1302 	struct mpr_command *cm;
1303 	struct mpr_chain *chain;
1304 	int i, rsize, nsegs;
1305 
1306 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1307         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1308 				16, 0,			/* algnmnt, boundary */
1309 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1310 				BUS_SPACE_MAXADDR,	/* highaddr */
1311 				NULL, NULL,		/* filter, filterarg */
1312                                 rsize,			/* maxsize */
1313                                 1,			/* nsegments */
1314                                 rsize,			/* maxsegsize */
1315                                 0,			/* flags */
1316                                 NULL, NULL,		/* lockfunc, lockarg */
1317                                 &sc->req_dmat)) {
1318 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request DMA tag\n");
1319 		return (ENOMEM);
1320         }
1321         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1322 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1323 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request memory\n");
1324 		return (ENOMEM);
1325         }
1326         bzero(sc->req_frames, rsize);
1327         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1328 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1329 
1330 	/*
1331 	 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to
1332 	 * get the size of a Chain Frame.  Previous versions use the size as a
1333 	 * Request Frame for the Chain Frame size.  If IOCMaxChainSegmentSize
1334 	 * is 0, use the default value.  The IOCMaxChainSegmentSize is the
1335 	 * number of 16-byte elelements that can fit in a Chain Frame, which is
1336 	 * the size of an IEEE Simple SGE.
1337 	 */
1338 	if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) {
1339 		sc->chain_seg_size =
1340 		    htole16(sc->facts->IOCMaxChainSegmentSize);
1341 		if (sc->chain_seg_size == 0) {
1342 			sc->chain_frame_size = MPR_DEFAULT_CHAIN_SEG_SIZE *
1343 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
1344 		} else {
1345 			sc->chain_frame_size = sc->chain_seg_size *
1346 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
1347 		}
1348 	} else {
1349 		sc->chain_frame_size = sc->facts->IOCRequestFrameSize * 4;
1350 	}
1351 	rsize = sc->chain_frame_size * sc->max_chains;
1352         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1353 				16, 0,			/* algnmnt, boundary */
1354 				BUS_SPACE_MAXADDR,	/* lowaddr */
1355 				BUS_SPACE_MAXADDR,	/* highaddr */
1356 				NULL, NULL,		/* filter, filterarg */
1357                                 rsize,			/* maxsize */
1358                                 1,			/* nsegments */
1359                                 rsize,			/* maxsegsize */
1360                                 0,			/* flags */
1361                                 NULL, NULL,		/* lockfunc, lockarg */
1362                                 &sc->chain_dmat)) {
1363 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain DMA tag\n");
1364 		return (ENOMEM);
1365         }
1366         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1367 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1368 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1369 		return (ENOMEM);
1370         }
1371         bzero(sc->chain_frames, rsize);
1372         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1373 	    mpr_memaddr_cb, &sc->chain_busaddr, 0);
1374 
1375 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1376 	if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1377 				1, 0,			/* algnmnt, boundary */
1378 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1379 				BUS_SPACE_MAXADDR,	/* highaddr */
1380 				NULL, NULL,		/* filter, filterarg */
1381                                 rsize,			/* maxsize */
1382                                 1,			/* nsegments */
1383                                 rsize,			/* maxsegsize */
1384                                 0,			/* flags */
1385                                 NULL, NULL,		/* lockfunc, lockarg */
1386                                 &sc->sense_dmat)) {
1387 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense DMA tag\n");
1388 		return (ENOMEM);
1389         }
1390         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1391 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1392 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense memory\n");
1393 		return (ENOMEM);
1394         }
1395         bzero(sc->sense_frames, rsize);
1396         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1397 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1398 
1399 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->max_chains, M_MPR,
1400 	    M_WAITOK | M_ZERO);
1401 	if (!sc->chains) {
1402 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1403 		return (ENOMEM);
1404 	}
1405 	for (i = 0; i < sc->max_chains; i++) {
1406 		chain = &sc->chains[i];
1407 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1408 		    i * sc->chain_frame_size);
1409 		chain->chain_busaddr = sc->chain_busaddr +
1410 		    i * sc->chain_frame_size;
1411 		mpr_free_chain(sc, chain);
1412 		sc->chain_free_lowwater++;
1413 	}
1414 
1415 	/*
1416 	 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports
1417 	 * these devices.
1418 	 */
1419 	if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) &&
1420 	    (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) {
1421 		if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM)
1422 			return (ENOMEM);
1423 	}
1424 
1425 	/* XXX Need to pick a more precise value */
1426 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1427         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1428 				1, 0,			/* algnmnt, boundary */
1429 				BUS_SPACE_MAXADDR,	/* lowaddr */
1430 				BUS_SPACE_MAXADDR,	/* highaddr */
1431 				NULL, NULL,		/* filter, filterarg */
1432                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1433                                 nsegs,			/* nsegments */
1434                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
1435                                 BUS_DMA_ALLOCNOW,	/* flags */
1436                                 busdma_lock_mutex,	/* lockfunc */
1437 				&sc->mpr_mtx,		/* lockarg */
1438                                 &sc->buffer_dmat)) {
1439 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate buffer DMA tag\n");
1440 		return (ENOMEM);
1441         }
1442 
1443 	/*
1444 	 * SMID 0 cannot be used as a free command per the firmware spec.
1445 	 * Just drop that command instead of risking accounting bugs.
1446 	 */
1447 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1448 	    M_MPR, M_WAITOK | M_ZERO);
1449 	if (!sc->commands) {
1450 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate command memory\n");
1451 		return (ENOMEM);
1452 	}
1453 	for (i = 1; i < sc->num_reqs; i++) {
1454 		cm = &sc->commands[i];
1455 		cm->cm_req = sc->req_frames +
1456 		    i * sc->facts->IOCRequestFrameSize * 4;
1457 		cm->cm_req_busaddr = sc->req_busaddr +
1458 		    i * sc->facts->IOCRequestFrameSize * 4;
1459 		cm->cm_sense = &sc->sense_frames[i];
1460 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1461 		cm->cm_desc.Default.SMID = i;
1462 		cm->cm_sc = sc;
1463 		TAILQ_INIT(&cm->cm_chain_list);
1464 		TAILQ_INIT(&cm->cm_prp_page_list);
1465 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1466 
1467 		/* XXX Is a failure here a critical problem? */
1468 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap)
1469 		    == 0) {
1470 			if (i <= sc->facts->HighPriorityCredit)
1471 				mpr_free_high_priority_command(sc, cm);
1472 			else
1473 				mpr_free_command(sc, cm);
1474 		} else {
1475 			panic("failed to allocate command %d\n", i);
1476 			sc->num_reqs = i;
1477 			break;
1478 		}
1479 	}
1480 
1481 	return (0);
1482 }
1483 
1484 /*
1485  * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs,
1486  * which are scatter/gather lists for NVMe devices.
1487  *
1488  * This buffer must be contiguous due to the nature of how NVMe PRPs are built
1489  * and translated by FW.
1490  *
1491  * returns ENOMEM if memory could not be allocated, otherwise returns 0.
1492  */
1493 static int
1494 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc)
1495 {
1496 	int PRPs_per_page, PRPs_required, pages_required;
1497 	int rsize, i;
1498 	struct mpr_prp_page *prp_page;
1499 
1500 	/*
1501 	 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number
1502 	 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is:
1503 	 * MAX_IO_SIZE / PAGE_SIZE = 256
1504 	 *
1505 	 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs
1506 	 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one
1507 	 * page (4096 / 8 = 512), so only one page is required for each I/O.
1508 	 *
1509 	 * Each of these buffers will need to be contiguous. For simplicity,
1510 	 * only one buffer is allocated here, which has all of the space
1511 	 * required for the NVMe Queue Depth. If there are problems allocating
1512 	 * this one buffer, this function will need to change to allocate
1513 	 * individual, contiguous NVME_QDEPTH buffers.
1514 	 *
1515 	 * The real calculation will use the real max io size. Above is just an
1516 	 * example.
1517 	 *
1518 	 */
1519 	PRPs_required = sc->maxio / PAGE_SIZE;
1520 	PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1;
1521 	pages_required = (PRPs_required / PRPs_per_page) + 1;
1522 
1523 	sc->prp_buffer_size = PAGE_SIZE * pages_required;
1524 	rsize = sc->prp_buffer_size * NVME_QDEPTH;
1525 	if (bus_dma_tag_create( sc->mpr_parent_dmat,	/* parent */
1526 				4, 0,			/* algnmnt, boundary */
1527 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1528 				BUS_SPACE_MAXADDR,	/* highaddr */
1529 				NULL, NULL,		/* filter, filterarg */
1530 				rsize,			/* maxsize */
1531 				1,			/* nsegments */
1532 				rsize,			/* maxsegsize */
1533 				0,			/* flags */
1534 				NULL, NULL,		/* lockfunc, lockarg */
1535 				&sc->prp_page_dmat)) {
1536 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP DMA "
1537 		    "tag\n");
1538 		return (ENOMEM);
1539 	}
1540 	if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages,
1541 	    BUS_DMA_NOWAIT, &sc->prp_page_map)) {
1542 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP memory\n");
1543 		return (ENOMEM);
1544 	}
1545 	bzero(sc->prp_pages, rsize);
1546 	bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages,
1547 	    rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0);
1548 
1549 	sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR,
1550 	    M_WAITOK | M_ZERO);
1551 	for (i = 0; i < NVME_QDEPTH; i++) {
1552 		prp_page = &sc->prps[i];
1553 		prp_page->prp_page = (uint64_t *)(sc->prp_pages +
1554 		    i * sc->prp_buffer_size);
1555 		prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr +
1556 		    i * sc->prp_buffer_size);
1557 		mpr_free_prp_page(sc, prp_page);
1558 		sc->prp_pages_free_lowwater++;
1559 	}
1560 
1561 	return (0);
1562 }
1563 
1564 static int
1565 mpr_init_queues(struct mpr_softc *sc)
1566 {
1567 	int i;
1568 
1569 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1570 
1571 	/*
1572 	 * According to the spec, we need to use one less reply than we
1573 	 * have space for on the queue.  So sc->num_replies (the number we
1574 	 * use) should be less than sc->fqdepth (allocated size).
1575 	 */
1576 	if (sc->num_replies >= sc->fqdepth)
1577 		return (EINVAL);
1578 
1579 	/*
1580 	 * Initialize all of the free queue entries.
1581 	 */
1582 	for (i = 0; i < sc->fqdepth; i++) {
1583 		sc->free_queue[i] = sc->reply_busaddr +
1584 		    (i * sc->facts->ReplyFrameSize * 4);
1585 	}
1586 	sc->replyfreeindex = sc->num_replies;
1587 
1588 	return (0);
1589 }
1590 
1591 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1592  * Next are the global settings, if they exist.  Highest are the per-unit
1593  * settings, if they exist.
1594  */
1595 void
1596 mpr_get_tunables(struct mpr_softc *sc)
1597 {
1598 	char tmpstr[80], mpr_debug[80];
1599 
1600 	/* XXX default to some debugging for now */
1601 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1602 	sc->disable_msix = 0;
1603 	sc->disable_msi = 0;
1604 	sc->max_msix = MPR_MSIX_MAX;
1605 	sc->max_chains = MPR_CHAIN_FRAMES;
1606 	sc->max_io_pages = MPR_MAXIO_PAGES;
1607 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1608 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1609 	sc->use_phynum = 1;
1610 	sc->max_reqframes = MPR_REQ_FRAMES;
1611 	sc->max_prireqframes = MPR_PRI_REQ_FRAMES;
1612 	sc->max_replyframes = MPR_REPLY_FRAMES;
1613 	sc->max_evtframes = MPR_EVT_REPLY_FRAMES;
1614 
1615 	/*
1616 	 * Grab the global variables.
1617 	 */
1618 	bzero(mpr_debug, 80);
1619 	if (TUNABLE_STR_FETCH("hw.mpr.debug_level", mpr_debug, 80) != 0)
1620 		mpr_parse_debug(sc, mpr_debug);
1621 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1622 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1623 	TUNABLE_INT_FETCH("hw.mpr.max_msix", &sc->max_msix);
1624 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1625 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1626 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1627 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1628 	TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum);
1629 	TUNABLE_INT_FETCH("hw.mpr.max_reqframes", &sc->max_reqframes);
1630 	TUNABLE_INT_FETCH("hw.mpr.max_prireqframes", &sc->max_prireqframes);
1631 	TUNABLE_INT_FETCH("hw.mpr.max_replyframes", &sc->max_replyframes);
1632 	TUNABLE_INT_FETCH("hw.mpr.max_evtframes", &sc->max_evtframes);
1633 
1634 	/* Grab the unit-instance variables */
1635 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1636 	    device_get_unit(sc->mpr_dev));
1637 	bzero(mpr_debug, 80);
1638 	if (TUNABLE_STR_FETCH(tmpstr, mpr_debug, 80) != 0)
1639 		mpr_parse_debug(sc, mpr_debug);
1640 
1641 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1642 	    device_get_unit(sc->mpr_dev));
1643 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1644 
1645 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1646 	    device_get_unit(sc->mpr_dev));
1647 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1648 
1649 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_msix",
1650 	    device_get_unit(sc->mpr_dev));
1651 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1652 
1653 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1654 	    device_get_unit(sc->mpr_dev));
1655 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1656 
1657 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1658 	    device_get_unit(sc->mpr_dev));
1659 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1660 
1661 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1662 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1663 	    device_get_unit(sc->mpr_dev));
1664 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1665 
1666 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1667 	    device_get_unit(sc->mpr_dev));
1668 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1669 
1670 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1671 	    device_get_unit(sc->mpr_dev));
1672 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1673 
1674 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num",
1675 	    device_get_unit(sc->mpr_dev));
1676 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1677 
1678 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_reqframes",
1679 	    device_get_unit(sc->mpr_dev));
1680 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1681 
1682 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_prireqframes",
1683 	    device_get_unit(sc->mpr_dev));
1684 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1685 
1686 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_replyframes",
1687 	    device_get_unit(sc->mpr_dev));
1688 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1689 
1690 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_evtframes",
1691 	    device_get_unit(sc->mpr_dev));
1692 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1693 }
1694 
1695 static void
1696 mpr_setup_sysctl(struct mpr_softc *sc)
1697 {
1698 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1699 	struct sysctl_oid	*sysctl_tree = NULL;
1700 	char tmpstr[80], tmpstr2[80];
1701 
1702 	/*
1703 	 * Setup the sysctl variable so the user can change the debug level
1704 	 * on the fly.
1705 	 */
1706 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1707 	    device_get_unit(sc->mpr_dev));
1708 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1709 
1710 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1711 	if (sysctl_ctx != NULL)
1712 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1713 
1714 	if (sysctl_tree == NULL) {
1715 		sysctl_ctx_init(&sc->sysctl_ctx);
1716 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1717 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1718 		    CTLFLAG_RD, 0, tmpstr);
1719 		if (sc->sysctl_tree == NULL)
1720 			return;
1721 		sysctl_ctx = &sc->sysctl_ctx;
1722 		sysctl_tree = sc->sysctl_tree;
1723 	}
1724 
1725 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1726 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1727 	    sc, 0, mpr_debug_sysctl, "A", "mpr debug level");
1728 
1729 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1730 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1731 	    "Disable the use of MSI-X interrupts");
1732 
1733 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1734 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1735 	    "User-defined maximum number of MSIX queues");
1736 
1737 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1738 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1739 	    "Negotiated number of MSIX queues");
1740 
1741 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1742 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1743 	    "Total number of allocated request frames");
1744 
1745 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1746 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1747 	    "Total number of allocated high priority request frames");
1748 
1749 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1750 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1751 	    "Total number of allocated reply frames");
1752 
1753 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1754 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1755 	    "Total number of event frames allocated");
1756 
1757 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1758 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1759 	    strlen(sc->fw_version), "firmware version");
1760 
1761 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1762 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPR_DRIVER_VERSION,
1763 	    strlen(MPR_DRIVER_VERSION), "driver version");
1764 
1765 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1766 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1767 	    &sc->io_cmds_active, 0, "number of currently active commands");
1768 
1769 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1770 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1771 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1772 
1773 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1774 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1775 	    &sc->chain_free, 0, "number of free chain elements");
1776 
1777 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1778 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1779 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1780 
1781 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1782 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1783 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1784 
1785 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1786 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1787 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1788 	    "IOCFacts)");
1789 
1790 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1791 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1792 	    "enable SSU to SATA SSD/HDD at shutdown");
1793 
1794 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1795 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1796 	    &sc->chain_alloc_fail, "chain allocation failures");
1797 
1798 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1799 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1800 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1801 	    "spinup after SATA ID error");
1802 
1803 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1804 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1805 	    "Use the phy number for enumeration");
1806 
1807 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1808 	    OID_AUTO, "prp_pages_free", CTLFLAG_RD,
1809 	    &sc->prp_pages_free, 0, "number of free PRP pages");
1810 
1811 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1812 	    OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD,
1813 	    &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages");
1814 
1815 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1816 	    OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD,
1817 	    &sc->prp_page_alloc_fail, "PRP page allocation failures");
1818 }
1819 
1820 static struct mpr_debug_string {
1821 	char *name;
1822 	int flag;
1823 } mpr_debug_strings[] = {
1824 	{"info", MPR_INFO},
1825 	{"fault", MPR_FAULT},
1826 	{"event", MPR_EVENT},
1827 	{"log", MPR_LOG},
1828 	{"recovery", MPR_RECOVERY},
1829 	{"error", MPR_ERROR},
1830 	{"init", MPR_INIT},
1831 	{"xinfo", MPR_XINFO},
1832 	{"user", MPR_USER},
1833 	{"mapping", MPR_MAPPING},
1834 	{"trace", MPR_TRACE}
1835 };
1836 
1837 static int
1838 mpr_debug_sysctl(SYSCTL_HANDLER_ARGS)
1839 {
1840 	struct mpr_softc *sc;
1841 	struct mpr_debug_string *string;
1842 	struct sbuf *sbuf;
1843 	char *buffer;
1844 	size_t sz;
1845 	int i, len, debug, error;
1846 
1847 	sc = (struct mpr_softc *)arg1;
1848 
1849 	error = sysctl_wire_old_buffer(req, 0);
1850 	if (error != 0)
1851 		return (error);
1852 
1853 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1854 	debug = sc->mpr_debug;
1855 
1856 	sbuf_printf(sbuf, "%#x", debug);
1857 
1858 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1859 	for (i = 0; i < sz; i++) {
1860 		string = &mpr_debug_strings[i];
1861 		if (debug & string->flag)
1862 			sbuf_printf(sbuf, ",%s", string->name);
1863 	}
1864 
1865 	error = sbuf_finish(sbuf);
1866 	sbuf_delete(sbuf);
1867 
1868 	if (error || req->newptr == NULL)
1869 		return (error);
1870 
1871 	len = req->newlen - req->newidx;
1872 	if (len == 0)
1873 		return (0);
1874 
1875 	buffer = malloc(len, M_MPR, M_ZERO|M_WAITOK);
1876 	error = SYSCTL_IN(req, buffer, len);
1877 
1878 	mpr_parse_debug(sc, buffer);
1879 
1880 	free(buffer, M_MPR);
1881 	return (error);
1882 }
1883 
1884 static void
1885 mpr_parse_debug(struct mpr_softc *sc, char *list)
1886 {
1887 	struct mpr_debug_string *string;
1888 	char *token, *endtoken;
1889 	size_t sz;
1890 	int flags, i;
1891 
1892 	if (list == NULL || *list == '\0')
1893 		return;
1894 
1895 	flags = 0;
1896 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1897 	while ((token = strsep(&list, ":,")) != NULL) {
1898 
1899 		/* Handle integer flags */
1900 		flags |= strtol(token, &endtoken, 0);
1901 		if (token != endtoken)
1902 			continue;
1903 
1904 		/* Handle text flags */
1905 		for (i = 0; i < sz; i++) {
1906 			string = &mpr_debug_strings[i];
1907 			if (strcasecmp(token, string->name) == 0) {
1908 				flags |= string->flag;
1909 				break;
1910 			}
1911 		}
1912 	}
1913 
1914 	sc->mpr_debug = flags;
1915 	return;
1916 }
1917 
1918 int
1919 mpr_attach(struct mpr_softc *sc)
1920 {
1921 	int error;
1922 
1923 	MPR_FUNCTRACE(sc);
1924 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1925 
1926 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
1927 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
1928 	callout_init_mtx(&sc->device_check_callout, &sc->mpr_mtx, 0);
1929 	TAILQ_INIT(&sc->event_list);
1930 	timevalclear(&sc->lastfail);
1931 
1932 	if ((error = mpr_transition_ready(sc)) != 0) {
1933 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
1934 		    "Failed to transition ready\n");
1935 		return (error);
1936 	}
1937 
1938 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
1939 	    M_ZERO|M_NOWAIT);
1940 	if (!sc->facts) {
1941 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
1942 		    "Cannot allocate memory, exit\n");
1943 		return (ENOMEM);
1944 	}
1945 
1946 	/*
1947 	 * Get IOC Facts and allocate all structures based on this information.
1948 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
1949 	 * Facts. If relevant values have changed in IOC Facts, this function
1950 	 * will free all of the memory based on IOC Facts and reallocate that
1951 	 * memory.  If this fails, any allocated memory should already be freed.
1952 	 */
1953 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
1954 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC Facts allocation "
1955 		    "failed with error %d\n", error);
1956 		return (error);
1957 	}
1958 
1959 	/* Start the periodic watchdog check on the IOC Doorbell */
1960 	mpr_periodic(sc);
1961 
1962 	/*
1963 	 * The portenable will kick off discovery events that will drive the
1964 	 * rest of the initialization process.  The CAM/SAS module will
1965 	 * hold up the boot sequence until discovery is complete.
1966 	 */
1967 	sc->mpr_ich.ich_func = mpr_startup;
1968 	sc->mpr_ich.ich_arg = sc;
1969 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
1970 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
1971 		    "Cannot establish MPR config hook\n");
1972 		error = EINVAL;
1973 	}
1974 
1975 	/*
1976 	 * Allow IR to shutdown gracefully when shutdown occurs.
1977 	 */
1978 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1979 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1980 
1981 	if (sc->shutdown_eh == NULL)
1982 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
1983 		    "shutdown event registration failed\n");
1984 
1985 	mpr_setup_sysctl(sc);
1986 
1987 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
1988 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
1989 
1990 	return (error);
1991 }
1992 
1993 /* Run through any late-start handlers. */
1994 static void
1995 mpr_startup(void *arg)
1996 {
1997 	struct mpr_softc *sc;
1998 
1999 	sc = (struct mpr_softc *)arg;
2000 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2001 
2002 	mpr_lock(sc);
2003 	mpr_unmask_intr(sc);
2004 
2005 	/* initialize device mapping tables */
2006 	mpr_base_static_config_pages(sc);
2007 	mpr_mapping_initialize(sc);
2008 	mprsas_startup(sc);
2009 	mpr_unlock(sc);
2010 
2011 	mpr_dprint(sc, MPR_INIT, "disestablish config intrhook\n");
2012 	config_intrhook_disestablish(&sc->mpr_ich);
2013 	sc->mpr_ich.ich_arg = NULL;
2014 
2015 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2016 }
2017 
2018 /* Periodic watchdog.  Is called with the driver lock already held. */
2019 static void
2020 mpr_periodic(void *arg)
2021 {
2022 	struct mpr_softc *sc;
2023 	uint32_t db;
2024 
2025 	sc = (struct mpr_softc *)arg;
2026 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
2027 		return;
2028 
2029 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
2030 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2031 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
2032 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
2033 			panic("TEMPERATURE FAULT: STOPPING.");
2034 		}
2035 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2036 		mpr_reinit(sc);
2037 	}
2038 
2039 	callout_reset(&sc->periodic, MPR_PERIODIC_DELAY * hz, mpr_periodic, sc);
2040 }
2041 
2042 static void
2043 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
2044     MPI2_EVENT_NOTIFICATION_REPLY *event)
2045 {
2046 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2047 
2048 	MPR_DPRINT_EVENT(sc, generic, event);
2049 
2050 	switch (event->Event) {
2051 	case MPI2_EVENT_LOG_DATA:
2052 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2053 		if (sc->mpr_debug & MPR_EVENT)
2054 			hexdump(event->EventData, event->EventDataLength, NULL,
2055 			    0);
2056 		break;
2057 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2058 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2059 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2060 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2061 		     entry->LogSequence);
2062 		break;
2063 	default:
2064 		break;
2065 	}
2066 	return;
2067 }
2068 
2069 static int
2070 mpr_attach_log(struct mpr_softc *sc)
2071 {
2072 	uint8_t events[16];
2073 
2074 	bzero(events, 16);
2075 	setbit(events, MPI2_EVENT_LOG_DATA);
2076 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2077 
2078 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
2079 	    &sc->mpr_log_eh);
2080 
2081 	return (0);
2082 }
2083 
2084 static int
2085 mpr_detach_log(struct mpr_softc *sc)
2086 {
2087 
2088 	if (sc->mpr_log_eh != NULL)
2089 		mpr_deregister_events(sc, sc->mpr_log_eh);
2090 	return (0);
2091 }
2092 
2093 /*
2094  * Free all of the driver resources and detach submodules.  Should be called
2095  * without the lock held.
2096  */
2097 int
2098 mpr_free(struct mpr_softc *sc)
2099 {
2100 	int error;
2101 
2102 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2103 	/* Turn off the watchdog */
2104 	mpr_lock(sc);
2105 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
2106 	mpr_unlock(sc);
2107 	/* Lock must not be held for this */
2108 	callout_drain(&sc->periodic);
2109 	callout_drain(&sc->device_check_callout);
2110 
2111 	if (((error = mpr_detach_log(sc)) != 0) ||
2112 	    ((error = mpr_detach_sas(sc)) != 0)) {
2113 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "failed to detach "
2114 		    "subsystems, error= %d, exit\n", error);
2115 		return (error);
2116 	}
2117 
2118 	mpr_detach_user(sc);
2119 
2120 	/* Put the IOC back in the READY state. */
2121 	mpr_lock(sc);
2122 	if ((error = mpr_transition_ready(sc)) != 0) {
2123 		mpr_unlock(sc);
2124 		return (error);
2125 	}
2126 	mpr_unlock(sc);
2127 
2128 	if (sc->facts != NULL)
2129 		free(sc->facts, M_MPR);
2130 
2131 	/*
2132 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2133 	 * to free these buffers too.
2134 	 */
2135 	mpr_iocfacts_free(sc);
2136 
2137 	if (sc->sysctl_tree != NULL)
2138 		sysctl_ctx_free(&sc->sysctl_ctx);
2139 
2140 	/* Deregister the shutdown function */
2141 	if (sc->shutdown_eh != NULL)
2142 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2143 
2144 	mtx_destroy(&sc->mpr_mtx);
2145 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2146 
2147 	return (0);
2148 }
2149 
2150 static __inline void
2151 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
2152 {
2153 	MPR_FUNCTRACE(sc);
2154 
2155 	if (cm == NULL) {
2156 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
2157 		return;
2158 	}
2159 
2160 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
2161 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
2162 
2163 	if (cm->cm_complete != NULL) {
2164 		mpr_dprint(sc, MPR_TRACE,
2165 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
2166 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
2167 		    cm->cm_reply);
2168 		cm->cm_complete(sc, cm);
2169 	}
2170 
2171 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
2172 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
2173 		wakeup(cm);
2174 	}
2175 
2176 	if (sc->io_cmds_active != 0) {
2177 		sc->io_cmds_active--;
2178 	} else {
2179 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
2180 		    "out of sync - resynching to 0\n");
2181 	}
2182 }
2183 
2184 static void
2185 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
2186 {
2187 	union loginfo_type {
2188 		u32	loginfo;
2189 		struct {
2190 			u32	subcode:16;
2191 			u32	code:8;
2192 			u32	originator:4;
2193 			u32	bus_type:4;
2194 		} dw;
2195 	};
2196 	union loginfo_type sas_loginfo;
2197 	char *originator_str = NULL;
2198 
2199 	sas_loginfo.loginfo = log_info;
2200 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2201 		return;
2202 
2203 	/* each nexus loss loginfo */
2204 	if (log_info == 0x31170000)
2205 		return;
2206 
2207 	/* eat the loginfos associated with task aborts */
2208 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
2209 	    (log_info == 0x31130000))
2210 		return;
2211 
2212 	switch (sas_loginfo.dw.originator) {
2213 	case 0:
2214 		originator_str = "IOP";
2215 		break;
2216 	case 1:
2217 		originator_str = "PL";
2218 		break;
2219 	case 2:
2220 		originator_str = "IR";
2221 		break;
2222 	}
2223 
2224 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
2225 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
2226 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
2227 }
2228 
2229 static void
2230 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
2231 {
2232 	MPI2DefaultReply_t *mpi_reply;
2233 	u16 sc_status;
2234 
2235 	mpi_reply = (MPI2DefaultReply_t*)reply;
2236 	sc_status = le16toh(mpi_reply->IOCStatus);
2237 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2238 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2239 }
2240 
2241 void
2242 mpr_intr(void *data)
2243 {
2244 	struct mpr_softc *sc;
2245 	uint32_t status;
2246 
2247 	sc = (struct mpr_softc *)data;
2248 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2249 
2250 	/*
2251 	 * Check interrupt status register to flush the bus.  This is
2252 	 * needed for both INTx interrupts and driver-driven polling
2253 	 */
2254 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2255 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2256 		return;
2257 
2258 	mpr_lock(sc);
2259 	mpr_intr_locked(data);
2260 	mpr_unlock(sc);
2261 	return;
2262 }
2263 
2264 /*
2265  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2266  * chip.  Hopefully this theory is correct.
2267  */
2268 void
2269 mpr_intr_msi(void *data)
2270 {
2271 	struct mpr_softc *sc;
2272 
2273 	sc = (struct mpr_softc *)data;
2274 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2275 	mpr_lock(sc);
2276 	mpr_intr_locked(data);
2277 	mpr_unlock(sc);
2278 	return;
2279 }
2280 
2281 /*
2282  * The locking is overly broad and simplistic, but easy to deal with for now.
2283  */
2284 void
2285 mpr_intr_locked(void *data)
2286 {
2287 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2288 	struct mpr_softc *sc;
2289 	struct mpr_command *cm = NULL;
2290 	uint8_t flags;
2291 	u_int pq;
2292 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2293 	mpr_fw_diagnostic_buffer_t *pBuffer;
2294 
2295 	sc = (struct mpr_softc *)data;
2296 
2297 	pq = sc->replypostindex;
2298 	mpr_dprint(sc, MPR_TRACE,
2299 	    "%s sc %p starting with replypostindex %u\n",
2300 	    __func__, sc, sc->replypostindex);
2301 
2302 	for ( ;; ) {
2303 		cm = NULL;
2304 		desc = &sc->post_queue[sc->replypostindex];
2305 		flags = desc->Default.ReplyFlags &
2306 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2307 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
2308 		    (le32toh(desc->Words.High) == 0xffffffff))
2309 			break;
2310 
2311 		/* increment the replypostindex now, so that event handlers
2312 		 * and cm completion handlers which decide to do a diag
2313 		 * reset can zero it without it getting incremented again
2314 		 * afterwards, and we break out of this loop on the next
2315 		 * iteration since the reply post queue has been cleared to
2316 		 * 0xFF and all descriptors look unused (which they are).
2317 		 */
2318 		if (++sc->replypostindex >= sc->pqdepth)
2319 			sc->replypostindex = 0;
2320 
2321 		switch (flags) {
2322 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2323 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
2324 		case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS:
2325 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2326 			cm->cm_reply = NULL;
2327 			break;
2328 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2329 		{
2330 			uint32_t baddr;
2331 			uint8_t *reply;
2332 
2333 			/*
2334 			 * Re-compose the reply address from the address
2335 			 * sent back from the chip.  The ReplyFrameAddress
2336 			 * is the lower 32 bits of the physical address of
2337 			 * particular reply frame.  Convert that address to
2338 			 * host format, and then use that to provide the
2339 			 * offset against the virtual address base
2340 			 * (sc->reply_frames).
2341 			 */
2342 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2343 			reply = sc->reply_frames +
2344 				(baddr - ((uint32_t)sc->reply_busaddr));
2345 			/*
2346 			 * Make sure the reply we got back is in a valid
2347 			 * range.  If not, go ahead and panic here, since
2348 			 * we'll probably panic as soon as we deference the
2349 			 * reply pointer anyway.
2350 			 */
2351 			if ((reply < sc->reply_frames)
2352 			 || (reply > (sc->reply_frames +
2353 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
2354 				printf("%s: WARNING: reply %p out of range!\n",
2355 				       __func__, reply);
2356 				printf("%s: reply_frames %p, fqdepth %d, "
2357 				       "frame size %d\n", __func__,
2358 				       sc->reply_frames, sc->fqdepth,
2359 				       sc->facts->ReplyFrameSize * 4);
2360 				printf("%s: baddr %#x,\n", __func__, baddr);
2361 				/* LSI-TODO. See Linux Code for Graceful exit */
2362 				panic("Reply address out of range");
2363 			}
2364 			if (le16toh(desc->AddressReply.SMID) == 0) {
2365 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2366 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2367 					/*
2368 					 * If SMID is 0 for Diag Buffer Post,
2369 					 * this implies that the reply is due to
2370 					 * a release function with a status that
2371 					 * the buffer has been released.  Set
2372 					 * the buffer flags accordingly.
2373 					 */
2374 					rel_rep =
2375 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2376 					if ((le16toh(rel_rep->IOCStatus) &
2377 					    MPI2_IOCSTATUS_MASK) ==
2378 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2379 					{
2380 						pBuffer =
2381 						    &sc->fw_diag_buffer_list[
2382 						    rel_rep->BufferType];
2383 						pBuffer->valid_data = TRUE;
2384 						pBuffer->owned_by_firmware =
2385 						    FALSE;
2386 						pBuffer->immediate = FALSE;
2387 					}
2388 				} else
2389 					mpr_dispatch_event(sc, baddr,
2390 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2391 					    reply);
2392 			} else {
2393 				cm = &sc->commands[
2394 				    le16toh(desc->AddressReply.SMID)];
2395 				cm->cm_reply = reply;
2396 				cm->cm_reply_data =
2397 				    le32toh(desc->AddressReply.
2398 				    ReplyFrameAddress);
2399 			}
2400 			break;
2401 		}
2402 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2403 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2404 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2405 		default:
2406 			/* Unhandled */
2407 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
2408 			    desc->Default.ReplyFlags);
2409 			cm = NULL;
2410 			break;
2411 		}
2412 
2413 		if (cm != NULL) {
2414 			// Print Error reply frame
2415 			if (cm->cm_reply)
2416 				mpr_display_reply_info(sc,cm->cm_reply);
2417 			mpr_complete_command(sc, cm);
2418 		}
2419 
2420 		desc->Words.Low = 0xffffffff;
2421 		desc->Words.High = 0xffffffff;
2422 	}
2423 
2424 	if (pq != sc->replypostindex) {
2425 		mpr_dprint(sc, MPR_TRACE,
2426 		    "%s sc %p writing postindex %d\n",
2427 		    __func__, sc, sc->replypostindex);
2428 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2429 		    sc->replypostindex);
2430 	}
2431 
2432 	return;
2433 }
2434 
2435 static void
2436 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2437     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2438 {
2439 	struct mpr_event_handle *eh;
2440 	int event, handled = 0;
2441 
2442 	event = le16toh(reply->Event);
2443 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2444 		if (isset(eh->mask, event)) {
2445 			eh->callback(sc, data, reply);
2446 			handled++;
2447 		}
2448 	}
2449 
2450 	if (handled == 0)
2451 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2452 		    le16toh(event));
2453 
2454 	/*
2455 	 * This is the only place that the event/reply should be freed.
2456 	 * Anything wanting to hold onto the event data should have
2457 	 * already copied it into their own storage.
2458 	 */
2459 	mpr_free_reply(sc, data);
2460 }
2461 
2462 static void
2463 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2464 {
2465 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2466 
2467 	if (cm->cm_reply)
2468 		MPR_DPRINT_EVENT(sc, generic,
2469 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2470 
2471 	mpr_free_command(sc, cm);
2472 
2473 	/* next, send a port enable */
2474 	mprsas_startup(sc);
2475 }
2476 
2477 /*
2478  * For both register_events and update_events, the caller supplies a bitmap
2479  * of events that it _wants_.  These functions then turn that into a bitmask
2480  * suitable for the controller.
2481  */
2482 int
2483 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2484     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2485 {
2486 	struct mpr_event_handle *eh;
2487 	int error = 0;
2488 
2489 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2490 	if (!eh) {
2491 		mpr_dprint(sc, MPR_EVENT|MPR_ERROR,
2492 		    "Cannot allocate event memory\n");
2493 		return (ENOMEM);
2494 	}
2495 	eh->callback = cb;
2496 	eh->data = data;
2497 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2498 	if (mask != NULL)
2499 		error = mpr_update_events(sc, eh, mask);
2500 	*handle = eh;
2501 
2502 	return (error);
2503 }
2504 
2505 int
2506 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2507     uint8_t *mask)
2508 {
2509 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2510 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2511 	struct mpr_command *cm = NULL;
2512 	struct mpr_event_handle *eh;
2513 	int error, i;
2514 
2515 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2516 
2517 	if ((mask != NULL) && (handle != NULL))
2518 		bcopy(mask, &handle->mask[0], 16);
2519 	memset(sc->event_mask, 0xff, 16);
2520 
2521 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2522 		for (i = 0; i < 16; i++)
2523 			sc->event_mask[i] &= ~eh->mask[i];
2524 	}
2525 
2526 	if ((cm = mpr_alloc_command(sc)) == NULL)
2527 		return (EBUSY);
2528 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2529 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2530 	evtreq->MsgFlags = 0;
2531 	evtreq->SASBroadcastPrimitiveMasks = 0;
2532 #ifdef MPR_DEBUG_ALL_EVENTS
2533 	{
2534 		u_char fullmask[16];
2535 		memset(fullmask, 0x00, 16);
2536 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2537 	}
2538 #else
2539 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2540 #endif
2541 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2542 	cm->cm_data = NULL;
2543 
2544 	error = mpr_request_polled(sc, &cm);
2545 	if (cm != NULL)
2546 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2547 	if ((reply == NULL) ||
2548 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2549 		error = ENXIO;
2550 
2551 	if (reply)
2552 		MPR_DPRINT_EVENT(sc, generic, reply);
2553 
2554 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2555 
2556 	if (cm != NULL)
2557 		mpr_free_command(sc, cm);
2558 	return (error);
2559 }
2560 
2561 static int
2562 mpr_reregister_events(struct mpr_softc *sc)
2563 {
2564 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2565 	struct mpr_command *cm;
2566 	struct mpr_event_handle *eh;
2567 	int error, i;
2568 
2569 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2570 
2571 	/* first, reregister events */
2572 
2573 	memset(sc->event_mask, 0xff, 16);
2574 
2575 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2576 		for (i = 0; i < 16; i++)
2577 			sc->event_mask[i] &= ~eh->mask[i];
2578 	}
2579 
2580 	if ((cm = mpr_alloc_command(sc)) == NULL)
2581 		return (EBUSY);
2582 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2583 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2584 	evtreq->MsgFlags = 0;
2585 	evtreq->SASBroadcastPrimitiveMasks = 0;
2586 #ifdef MPR_DEBUG_ALL_EVENTS
2587 	{
2588 		u_char fullmask[16];
2589 		memset(fullmask, 0x00, 16);
2590 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2591 	}
2592 #else
2593 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2594 #endif
2595 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2596 	cm->cm_data = NULL;
2597 	cm->cm_complete = mpr_reregister_events_complete;
2598 
2599 	error = mpr_map_command(sc, cm);
2600 
2601 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2602 	    error);
2603 	return (error);
2604 }
2605 
2606 int
2607 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2608 {
2609 
2610 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2611 	free(handle, M_MPR);
2612 	return (mpr_update_events(sc, NULL, NULL));
2613 }
2614 
2615 /**
2616 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a
2617 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry
2618 * of the NVMe message (PRP1). If the data buffer is small enough to be described
2619 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to
2620 * describe a larger data buffer. If the data buffer is too large to describe
2621 * using the two PRP entriess inside the NVMe message, then PRP1 describes the
2622 * first data memory segment, and PRP2 contains a pointer to a PRP list located
2623 * elsewhere in memory to describe the remaining data memory segments. The PRP
2624 * list will be contiguous.
2625 
2626 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP
2627 * consists of a list of PRP entries to describe a number of noncontigous
2628 * physical memory segments as a single memory buffer, just as a SGL does. Note
2629 * however, that this function is only used by the IOCTL call, so the memory
2630 * given will be guaranteed to be contiguous. There is no need to translate
2631 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous
2632 * space that is one page size each.
2633 *
2634 * Each NVMe message contains two PRP entries. The first (PRP1) either contains
2635 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains
2636 * the second PRP element if the memory being described fits within 2 PRP
2637 * entries, or a PRP list pointer if the PRP spans more than two entries.
2638 *
2639 * A PRP list pointer contains the address of a PRP list, structured as a linear
2640 * array of PRP entries. Each PRP entry in this list describes a segment of
2641 * physical memory.
2642 *
2643 * Each 64-bit PRP entry comprises an address and an offset field. The address
2644 * always points to the beginning of a PAGE_SIZE physical memory page, and the
2645 * offset describes where within that page the memory segment begins. Only the
2646 * first element in a PRP list may contain a non-zero offest, implying that all
2647 * memory segments following the first begin at the start of a PAGE_SIZE page.
2648 *
2649 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory,
2650 * with exceptions for the first and last elements in the list. If the memory
2651 * being described by the list begins at a non-zero offset within the first page,
2652 * then the first PRP element will contain a non-zero offset indicating where the
2653 * region begins within the page. The last memory segment may end before the end
2654 * of the PAGE_SIZE segment, depending upon the overall size of the memory being
2655 * described by the PRP list.
2656 *
2657 * Since PRP entries lack any indication of size, the overall data buffer length
2658 * is used to determine where the end of the data memory buffer is located, and
2659 * how many PRP entries are required to describe it.
2660 *
2661 * Returns nothing.
2662 */
2663 void
2664 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm,
2665     Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data,
2666     uint32_t data_in_sz, uint32_t data_out_sz)
2667 {
2668 	int			prp_size = PRP_ENTRY_SIZE;
2669 	uint64_t		*prp_entry, *prp1_entry, *prp2_entry;
2670 	uint64_t		*prp_entry_phys, *prp_page, *prp_page_phys;
2671 	uint32_t		offset, entry_len, page_mask_result, page_mask;
2672 	bus_addr_t		paddr;
2673 	size_t			length;
2674 	struct mpr_prp_page	*prp_page_info = NULL;
2675 
2676 	/*
2677 	 * Not all commands require a data transfer. If no data, just return
2678 	 * without constructing any PRP.
2679 	 */
2680 	if (!data_in_sz && !data_out_sz)
2681 		return;
2682 
2683 	/*
2684 	 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is
2685 	 * located at a 24 byte offset from the start of the NVMe command. Then
2686 	 * set the current PRP entry pointer to PRP1.
2687 	 */
2688 	prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2689 	    NVME_CMD_PRP1_OFFSET);
2690 	prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2691 	    NVME_CMD_PRP2_OFFSET);
2692 	prp_entry = prp1_entry;
2693 
2694 	/*
2695 	 * For the PRP entries, use the specially allocated buffer of
2696 	 * contiguous memory. PRP Page allocation failures should not happen
2697 	 * because there should be enough PRP page buffers to account for the
2698 	 * possible NVMe QDepth.
2699 	 */
2700 	prp_page_info = mpr_alloc_prp_page(sc);
2701 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2702 	    "used for building a native NVMe SGL.\n", __func__));
2703 	prp_page = (uint64_t *)prp_page_info->prp_page;
2704 	prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2705 
2706 	/*
2707 	 * Insert the allocated PRP page into the command's PRP page list. This
2708 	 * will be freed when the command is freed.
2709 	 */
2710 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2711 
2712 	/*
2713 	 * Check if we are within 1 entry of a page boundary we don't want our
2714 	 * first entry to be a PRP List entry.
2715 	 */
2716 	page_mask = PAGE_SIZE - 1;
2717 	page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) &
2718 	    page_mask;
2719 	if (!page_mask_result)
2720 	{
2721 		/* Bump up to next page boundary. */
2722 		prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size);
2723 		prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys +
2724 		    prp_size);
2725 	}
2726 
2727 	/*
2728 	 * Set PRP physical pointer, which initially points to the current PRP
2729 	 * DMA memory page.
2730 	 */
2731 	prp_entry_phys = prp_page_phys;
2732 
2733 	/* Get physical address and length of the data buffer. */
2734 	paddr = (bus_addr_t)data;
2735 	if (data_in_sz)
2736 		length = data_in_sz;
2737 	else
2738 		length = data_out_sz;
2739 
2740 	/* Loop while the length is not zero. */
2741 	while (length)
2742 	{
2743 		/*
2744 		 * Check if we need to put a list pointer here if we are at page
2745 		 * boundary - prp_size (8 bytes).
2746 		 */
2747 		page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys +
2748 		    prp_size) & page_mask;
2749 		if (!page_mask_result)
2750 		{
2751 			/*
2752 			 * This is the last entry in a PRP List, so we need to
2753 			 * put a PRP list pointer here. What this does is:
2754 			 *   - bump the current memory pointer to the next
2755 			 *     address, which will be the next full page.
2756 			 *   - set the PRP Entry to point to that page. This is
2757 			 *     now the PRP List pointer.
2758 			 *   - bump the PRP Entry pointer the start of the next
2759 			 *     page. Since all of this PRP memory is contiguous,
2760 			 *     no need to get a new page - it's just the next
2761 			 *     address.
2762 			 */
2763 			prp_entry_phys++;
2764 			*prp_entry =
2765 			    htole64((uint64_t)(uintptr_t)prp_entry_phys);
2766 			prp_entry++;
2767 		}
2768 
2769 		/* Need to handle if entry will be part of a page. */
2770 		offset = (uint32_t)paddr & page_mask;
2771 		entry_len = PAGE_SIZE - offset;
2772 
2773 		if (prp_entry == prp1_entry)
2774 		{
2775 			/*
2776 			 * Must fill in the first PRP pointer (PRP1) before
2777 			 * moving on.
2778 			 */
2779 			*prp1_entry = htole64((uint64_t)paddr);
2780 
2781 			/*
2782 			 * Now point to the second PRP entry within the
2783 			 * command (PRP2).
2784 			 */
2785 			prp_entry = prp2_entry;
2786 		}
2787 		else if (prp_entry == prp2_entry)
2788 		{
2789 			/*
2790 			 * Should the PRP2 entry be a PRP List pointer or just a
2791 			 * regular PRP pointer? If there is more than one more
2792 			 * page of data, must use a PRP List pointer.
2793 			 */
2794 			if (length > PAGE_SIZE)
2795 			{
2796 				/*
2797 				 * PRP2 will contain a PRP List pointer because
2798 				 * more PRP's are needed with this command. The
2799 				 * list will start at the beginning of the
2800 				 * contiguous buffer.
2801 				 */
2802 				*prp2_entry =
2803 				    htole64(
2804 				    (uint64_t)(uintptr_t)prp_entry_phys);
2805 
2806 				/*
2807 				 * The next PRP Entry will be the start of the
2808 				 * first PRP List.
2809 				 */
2810 				prp_entry = prp_page;
2811 			}
2812 			else
2813 			{
2814 				/*
2815 				 * After this, the PRP Entries are complete.
2816 				 * This command uses 2 PRP's and no PRP list.
2817 				 */
2818 				*prp2_entry = htole64((uint64_t)paddr);
2819 			}
2820 		}
2821 		else
2822 		{
2823 			/*
2824 			 * Put entry in list and bump the addresses.
2825 			 *
2826 			 * After PRP1 and PRP2 are filled in, this will fill in
2827 			 * all remaining PRP entries in a PRP List, one per each
2828 			 * time through the loop.
2829 			 */
2830 			*prp_entry = htole64((uint64_t)paddr);
2831 			prp_entry++;
2832 			prp_entry_phys++;
2833 		}
2834 
2835 		/*
2836 		 * Bump the phys address of the command's data buffer by the
2837 		 * entry_len.
2838 		 */
2839 		paddr += entry_len;
2840 
2841 		/* Decrement length accounting for last partial page. */
2842 		if (entry_len > length)
2843 			length = 0;
2844 		else
2845 			length -= entry_len;
2846 	}
2847 }
2848 
2849 /*
2850  * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to
2851  * determine if the driver needs to build a native SGL. If so, that native SGL
2852  * is built in the contiguous buffers allocated especially for PCIe SGL
2853  * creation. If the driver will not build a native SGL, return TRUE and a
2854  * normal IEEE SGL will be built. Currently this routine supports NVMe devices
2855  * only.
2856  *
2857  * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built.
2858  */
2859 static int
2860 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm,
2861     bus_dma_segment_t *segs, int segs_left)
2862 {
2863 	uint32_t		i, sge_dwords, length, offset, entry_len;
2864 	uint32_t		num_entries, buff_len = 0, sges_in_segment;
2865 	uint32_t		page_mask, page_mask_result, *curr_buff;
2866 	uint32_t		*ptr_sgl, *ptr_first_sgl, first_page_offset;
2867 	uint32_t		first_page_data_size, end_residual;
2868 	uint64_t		*msg_phys;
2869 	bus_addr_t		paddr;
2870 	int			build_native_sgl = 0, first_prp_entry;
2871 	int			prp_size = PRP_ENTRY_SIZE;
2872 	Mpi25IeeeSgeChain64_t	*main_chain_element = NULL;
2873 	struct mpr_prp_page	*prp_page_info = NULL;
2874 
2875 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2876 
2877 	/*
2878 	 * Add up the sizes of each segment length to get the total transfer
2879 	 * size, which will be checked against the Maximum Data Transfer Size.
2880 	 * If the data transfer length exceeds the MDTS for this device, just
2881 	 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O
2882 	 * up into multiple I/O's. [nvme_mdts = 0 means unlimited]
2883 	 */
2884 	for (i = 0; i < segs_left; i++)
2885 		buff_len += htole32(segs[i].ds_len);
2886 	if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS))
2887 		return 1;
2888 
2889 	/* Create page_mask (to get offset within page) */
2890 	page_mask = PAGE_SIZE - 1;
2891 
2892 	/*
2893 	 * Check if the number of elements exceeds the max number that can be
2894 	 * put in the main message frame (H/W can only translate an SGL that
2895 	 * is contained entirely in the main message frame).
2896 	 */
2897 	sges_in_segment = (sc->facts->IOCRequestFrameSize -
2898 	    offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION);
2899 	if (segs_left > sges_in_segment)
2900 		build_native_sgl = 1;
2901 	else
2902 	{
2903 		/*
2904 		 * NVMe uses one PRP for each physical page (or part of physical
2905 		 * page).
2906 		 *    if 4 pages or less then IEEE is OK
2907 		 *    if > 5 pages then we need to build a native SGL
2908 		 *    if > 4 and <= 5 pages, then check the physical address of
2909 		 *      the first SG entry, then if this first size in the page
2910 		 *      is >= the residual beyond 4 pages then use IEEE,
2911 		 *      otherwise use native SGL
2912 		 */
2913 		if (buff_len > (PAGE_SIZE * 5))
2914 			build_native_sgl = 1;
2915 		else if ((buff_len > (PAGE_SIZE * 4)) &&
2916 		    (buff_len <= (PAGE_SIZE * 5)) )
2917 		{
2918 			msg_phys = (uint64_t *)segs[0].ds_addr;
2919 			first_page_offset =
2920 			    ((uint32_t)(uint64_t)(uintptr_t)msg_phys &
2921 			    page_mask);
2922 			first_page_data_size = PAGE_SIZE - first_page_offset;
2923 			end_residual = buff_len % PAGE_SIZE;
2924 
2925 			/*
2926 			 * If offset into first page pushes the end of the data
2927 			 * beyond end of the 5th page, we need the extra PRP
2928 			 * list.
2929 			 */
2930 			if (first_page_data_size < end_residual)
2931 				build_native_sgl = 1;
2932 
2933 			/*
2934 			 * Check if first SG entry size is < residual beyond 4
2935 			 * pages.
2936 			 */
2937 			if (htole32(segs[0].ds_len) <
2938 			    (buff_len - (PAGE_SIZE * 4)))
2939 				build_native_sgl = 1;
2940 		}
2941 	}
2942 
2943 	/* check if native SGL is needed */
2944 	if (!build_native_sgl)
2945 		return 1;
2946 
2947 	/*
2948 	 * Native SGL is needed.
2949 	 * Put a chain element in main message frame that points to the first
2950 	 * chain buffer.
2951 	 *
2952 	 * NOTE:  The ChainOffset field must be 0 when using a chain pointer to
2953 	 *        a native SGL.
2954 	 */
2955 
2956 	/* Set main message chain element pointer */
2957 	main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge;
2958 
2959 	/*
2960 	 * For NVMe the chain element needs to be the 2nd SGL entry in the main
2961 	 * message.
2962 	 */
2963 	main_chain_element = (Mpi25IeeeSgeChain64_t *)
2964 	    ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64));
2965 
2966 	/*
2967 	 * For the PRP entries, use the specially allocated buffer of
2968 	 * contiguous memory. PRP Page allocation failures should not happen
2969 	 * because there should be enough PRP page buffers to account for the
2970 	 * possible NVMe QDepth.
2971 	 */
2972 	prp_page_info = mpr_alloc_prp_page(sc);
2973 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2974 	    "used for building a native NVMe SGL.\n", __func__));
2975 	curr_buff = (uint32_t *)prp_page_info->prp_page;
2976 	msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2977 
2978 	/*
2979 	 * Insert the allocated PRP page into the command's PRP page list. This
2980 	 * will be freed when the command is freed.
2981 	 */
2982 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2983 
2984 	/*
2985 	 * Check if we are within 1 entry of a page boundary we don't want our
2986 	 * first entry to be a PRP List entry.
2987 	 */
2988 	page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) &
2989 	    page_mask;
2990 	if (!page_mask_result) {
2991 		/* Bump up to next page boundary. */
2992 		curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size);
2993 		msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size);
2994 	}
2995 
2996 	/* Fill in the chain element and make it an NVMe segment type. */
2997 	main_chain_element->Address.High =
2998 	    htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32));
2999 	main_chain_element->Address.Low =
3000 	    htole32((uint32_t)(uintptr_t)msg_phys);
3001 	main_chain_element->NextChainOffset = 0;
3002 	main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3003 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3004 	    MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP;
3005 
3006 	/* Set SGL pointer to start of contiguous PCIe buffer. */
3007 	ptr_sgl = curr_buff;
3008 	sge_dwords = 2;
3009 	num_entries = 0;
3010 
3011 	/*
3012 	 * NVMe has a very convoluted PRP format. One PRP is required for each
3013 	 * page or partial page. We need to split up OS SG entries if they are
3014 	 * longer than one page or cross a page boundary. We also have to insert
3015 	 * a PRP list pointer entry as the last entry in each physical page of
3016 	 * the PRP list.
3017 	 *
3018 	 * NOTE: The first PRP "entry" is actually placed in the first SGL entry
3019 	 * in the main message in IEEE 64 format. The 2nd entry in the main
3020 	 * message is the chain element, and the rest of the PRP entries are
3021 	 * built in the contiguous PCIe buffer.
3022 	 */
3023 	first_prp_entry = 1;
3024 	ptr_first_sgl = (uint32_t *)cm->cm_sge;
3025 
3026 	for (i = 0; i < segs_left; i++) {
3027 		/* Get physical address and length of this SG entry. */
3028 		paddr = segs[i].ds_addr;
3029 		length = segs[i].ds_len;
3030 
3031 		/*
3032 		 * Check whether a given SGE buffer lies on a non-PAGED
3033 		 * boundary if this is not the first page. If so, this is not
3034 		 * expected so have FW build the SGL.
3035 		 */
3036 		if ((i != 0) && (((uint32_t)paddr & page_mask) != 0)) {
3037 			mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while "
3038 			    "building NVMe PRPs, low address is 0x%x\n",
3039 			    (uint32_t)paddr);
3040 			return 1;
3041 		}
3042 
3043 		/* Apart from last SGE, if any other SGE boundary is not page
3044 		 * aligned then it means that hole exists. Existence of hole
3045 		 * leads to data corruption. So fallback to IEEE SGEs.
3046 		 */
3047 		if (i != (segs_left - 1)) {
3048 			if (((uint32_t)paddr + length) & page_mask) {
3049 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE "
3050 				    "boundary while building NVMe PRPs, low "
3051 				    "address: 0x%x and length: %u\n",
3052 				    (uint32_t)paddr, length);
3053 				return 1;
3054 			}
3055 		}
3056 
3057 		/* Loop while the length is not zero. */
3058 		while (length) {
3059 			/*
3060 			 * Check if we need to put a list pointer here if we are
3061 			 * at page boundary - prp_size.
3062 			 */
3063 			page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl +
3064 			    prp_size) & page_mask;
3065 			if (!page_mask_result) {
3066 				/*
3067 				 * Need to put a PRP list pointer here.
3068 				 */
3069 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3070 				    prp_size);
3071 				*ptr_sgl = htole32((uintptr_t)msg_phys);
3072 				*(ptr_sgl+1) = htole32((uint64_t)(uintptr_t)
3073 				    msg_phys >> 32);
3074 				ptr_sgl += sge_dwords;
3075 				num_entries++;
3076 			}
3077 
3078 			/* Need to handle if entry will be part of a page. */
3079 			offset = (uint32_t)paddr & page_mask;
3080 			entry_len = PAGE_SIZE - offset;
3081 			if (first_prp_entry) {
3082 				/*
3083 				 * Put IEEE entry in first SGE in main message.
3084 				 * (Simple element, System addr, not end of
3085 				 * list.)
3086 				 */
3087 				*ptr_first_sgl = htole32((uint32_t)paddr);
3088 				*(ptr_first_sgl + 1) =
3089 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3090 				*(ptr_first_sgl + 2) = htole32(entry_len);
3091 				*(ptr_first_sgl + 3) = 0;
3092 
3093 				/* No longer the first PRP entry. */
3094 				first_prp_entry = 0;
3095 			} else {
3096 				/* Put entry in list. */
3097 				*ptr_sgl = htole32((uint32_t)paddr);
3098 				*(ptr_sgl + 1) =
3099 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3100 
3101 				/* Bump ptr_sgl, msg_phys, and num_entries. */
3102 				ptr_sgl += sge_dwords;
3103 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3104 				    prp_size);
3105 				num_entries++;
3106 			}
3107 
3108 			/* Bump the phys address by the entry_len. */
3109 			paddr += entry_len;
3110 
3111 			/* Decrement length accounting for last partial page. */
3112 			if (entry_len > length)
3113 				length = 0;
3114 			else
3115 				length -= entry_len;
3116 		}
3117 	}
3118 
3119 	/* Set chain element Length. */
3120 	main_chain_element->Length = htole32(num_entries * prp_size);
3121 
3122 	/* Return 0, indicating we built a native SGL. */
3123 	return 0;
3124 }
3125 
3126 /*
3127  * Add a chain element as the next SGE for the specified command.
3128  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
3129  * only required for IEEE commands.  Therefore there is no code for commands
3130  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
3131  * shouldn't be requesting chains).
3132  */
3133 static int
3134 mpr_add_chain(struct mpr_command *cm, int segsleft)
3135 {
3136 	struct mpr_softc *sc = cm->cm_sc;
3137 	MPI2_REQUEST_HEADER *req;
3138 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
3139 	struct mpr_chain *chain;
3140 	int sgc_size, current_segs, rem_segs, segs_per_frame;
3141 	uint8_t next_chain_offset = 0;
3142 
3143 	/*
3144 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
3145 	 * only IEEE commands should be requesting chains.  Return some error
3146 	 * code other than 0.
3147 	 */
3148 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
3149 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
3150 		    "an MPI SGL.\n");
3151 		return(ENOBUFS);
3152 	}
3153 
3154 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
3155 	if (cm->cm_sglsize < sgc_size)
3156 		panic("MPR: Need SGE Error Code\n");
3157 
3158 	chain = mpr_alloc_chain(cm->cm_sc);
3159 	if (chain == NULL)
3160 		return (ENOBUFS);
3161 
3162 	/*
3163 	 * Note: a double-linked list is used to make it easier to walk for
3164 	 * debugging.
3165 	 */
3166 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
3167 
3168 	/*
3169 	 * Need to know if the number of frames left is more than 1 or not.  If
3170 	 * more than 1 frame is required, NextChainOffset will need to be set,
3171 	 * which will just be the last segment of the frame.
3172 	 */
3173 	rem_segs = 0;
3174 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
3175 		/*
3176 		 * rem_segs is the number of segements remaining after the
3177 		 * segments that will go into the current frame.  Since it is
3178 		 * known that at least one more frame is required, account for
3179 		 * the chain element.  To know if more than one more frame is
3180 		 * required, just check if there will be a remainder after using
3181 		 * the current frame (with this chain) and the next frame.  If
3182 		 * so the NextChainOffset must be the last element of the next
3183 		 * frame.
3184 		 */
3185 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
3186 		rem_segs = segsleft - current_segs;
3187 		segs_per_frame = sc->chain_frame_size / sgc_size;
3188 		if (rem_segs > segs_per_frame) {
3189 			next_chain_offset = segs_per_frame - 1;
3190 		}
3191 	}
3192 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
3193 	ieee_sgc->Length = next_chain_offset ?
3194 	    htole32((uint32_t)sc->chain_frame_size) :
3195 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
3196 	ieee_sgc->NextChainOffset = next_chain_offset;
3197 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3198 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3199 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
3200 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
3201 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
3202 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
3203 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
3204 
3205 	cm->cm_sglsize = sc->chain_frame_size;
3206 	return (0);
3207 }
3208 
3209 /*
3210  * Add one scatter-gather element to the scatter-gather list for a command.
3211  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
3212  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
3213  * chain, so don't consider any chain additions.
3214  */
3215 int
3216 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
3217     int segsleft)
3218 {
3219 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3220 	u32 sge_flags;
3221 
3222 	/*
3223 	 * case 1: >=1 more segment, no room for anything (error)
3224 	 * case 2: 1 more segment and enough room for it
3225          */
3226 
3227 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
3228 		mpr_dprint(cm->cm_sc, MPR_ERROR,
3229 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
3230 		    __func__);
3231 		return(ENOBUFS);
3232 	}
3233 
3234 	KASSERT(segsleft == 1,
3235 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
3236 	    segsleft));
3237 
3238 	/*
3239 	 * There is one more segment left to add for the MPI SGL and there is
3240 	 * enough room in the frame to add it.  This is the normal case because
3241 	 * MPI SGL's don't have chains, otherwise something is wrong.
3242 	 *
3243 	 * If this is a bi-directional request, need to account for that
3244 	 * here.  Save the pre-filled sge values.  These will be used
3245 	 * either for the 2nd SGL or for a single direction SGL.  If
3246 	 * cm_out_len is non-zero, this is a bi-directional request, so
3247 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
3248 	 * fill in the IN SGL.  Note that at this time, when filling in
3249 	 * 2 SGL's for a bi-directional request, they both use the same
3250 	 * DMA buffer (same cm command).
3251 	 */
3252 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
3253 	saved_address_low = sge->Address.Low;
3254 	saved_address_high = sge->Address.High;
3255 	if (cm->cm_out_len) {
3256 		sge->FlagsLength = cm->cm_out_len |
3257 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3258 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
3259 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
3260 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3261 		    MPI2_SGE_FLAGS_SHIFT);
3262 		cm->cm_sglsize -= len;
3263 		/* Endian Safe code */
3264 		sge_flags = sge->FlagsLength;
3265 		sge->FlagsLength = htole32(sge_flags);
3266 		sge->Address.High = htole32(sge->Address.High);
3267 		sge->Address.Low = htole32(sge->Address.Low);
3268 		bcopy(sge, cm->cm_sge, len);
3269 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3270 	}
3271 	sge->FlagsLength = saved_buf_len |
3272 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3273 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
3274 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
3275 	    MPI2_SGE_FLAGS_END_OF_LIST |
3276 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3277 	    MPI2_SGE_FLAGS_SHIFT);
3278 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
3279 		sge->FlagsLength |=
3280 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
3281 		    MPI2_SGE_FLAGS_SHIFT);
3282 	} else {
3283 		sge->FlagsLength |=
3284 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
3285 		    MPI2_SGE_FLAGS_SHIFT);
3286 	}
3287 	sge->Address.Low = saved_address_low;
3288 	sge->Address.High = saved_address_high;
3289 
3290 	cm->cm_sglsize -= len;
3291 	/* Endian Safe code */
3292 	sge_flags = sge->FlagsLength;
3293 	sge->FlagsLength = htole32(sge_flags);
3294 	sge->Address.High = htole32(sge->Address.High);
3295 	sge->Address.Low = htole32(sge->Address.Low);
3296 	bcopy(sge, cm->cm_sge, len);
3297 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3298 	return (0);
3299 }
3300 
3301 /*
3302  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
3303  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
3304  * remaining size and pointer to the next SGE to fill in, respectively.
3305  */
3306 int
3307 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
3308 {
3309 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
3310 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
3311 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3312 	uint32_t sge_length;
3313 
3314 	/*
3315 	 * case 1: No room for chain or segment (error).
3316 	 * case 2: Two or more segments left but only room for chain.
3317 	 * case 3: Last segment and room for it, so set flags.
3318 	 */
3319 
3320 	/*
3321 	 * There should be room for at least one element, or there is a big
3322 	 * problem.
3323 	 */
3324 	if (cm->cm_sglsize < ieee_sge_size)
3325 		panic("MPR: Need SGE Error Code\n");
3326 
3327 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
3328 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
3329 			return (error);
3330 	}
3331 
3332 	if (segsleft == 1) {
3333 		/*
3334 		 * If this is a bi-directional request, need to account for that
3335 		 * here.  Save the pre-filled sge values.  These will be used
3336 		 * either for the 2nd SGL or for a single direction SGL.  If
3337 		 * cm_out_len is non-zero, this is a bi-directional request, so
3338 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
3339 		 * fill in the IN SGL.  Note that at this time, when filling in
3340 		 * 2 SGL's for a bi-directional request, they both use the same
3341 		 * DMA buffer (same cm command).
3342 		 */
3343 		saved_buf_len = sge->Length;
3344 		saved_address_low = sge->Address.Low;
3345 		saved_address_high = sge->Address.High;
3346 		if (cm->cm_out_len) {
3347 			sge->Length = cm->cm_out_len;
3348 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3349 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3350 			cm->cm_sglsize -= ieee_sge_size;
3351 			/* Endian Safe code */
3352 			sge_length = sge->Length;
3353 			sge->Length = htole32(sge_length);
3354 			sge->Address.High = htole32(sge->Address.High);
3355 			sge->Address.Low = htole32(sge->Address.Low);
3356 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
3357 			cm->cm_sge =
3358 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3359 			    ieee_sge_size);
3360 		}
3361 		sge->Length = saved_buf_len;
3362 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3363 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3364 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
3365 		sge->Address.Low = saved_address_low;
3366 		sge->Address.High = saved_address_high;
3367 	}
3368 
3369 	cm->cm_sglsize -= ieee_sge_size;
3370 	/* Endian Safe code */
3371 	sge_length = sge->Length;
3372 	sge->Length = htole32(sge_length);
3373 	sge->Address.High = htole32(sge->Address.High);
3374 	sge->Address.Low = htole32(sge->Address.Low);
3375 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
3376 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3377 	    ieee_sge_size);
3378 	return (0);
3379 }
3380 
3381 /*
3382  * Add one dma segment to the scatter-gather list for a command.
3383  */
3384 int
3385 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
3386     int segsleft)
3387 {
3388 	MPI2_SGE_SIMPLE64 sge;
3389 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
3390 
3391 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
3392 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3393 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3394 		ieee_sge.Length = len;
3395 		mpr_from_u64(pa, &ieee_sge.Address);
3396 
3397 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
3398 	} else {
3399 		/*
3400 		 * This driver always uses 64-bit address elements for
3401 		 * simplicity.
3402 		 */
3403 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3404 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
3405 		/* Set Endian safe macro in mpr_push_sge */
3406 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
3407 		mpr_from_u64(pa, &sge.Address);
3408 
3409 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
3410 	}
3411 }
3412 
3413 static void
3414 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
3415 {
3416 	struct mpr_softc *sc;
3417 	struct mpr_command *cm;
3418 	u_int i, dir, sflags;
3419 
3420 	cm = (struct mpr_command *)arg;
3421 	sc = cm->cm_sc;
3422 
3423 	/*
3424 	 * In this case, just print out a warning and let the chip tell the
3425 	 * user they did the wrong thing.
3426 	 */
3427 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
3428 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
3429 		    "segments, more than the %d allowed\n", __func__, nsegs,
3430 		    cm->cm_max_segs);
3431 	}
3432 
3433 	/*
3434 	 * Set up DMA direction flags.  Bi-directional requests are also handled
3435 	 * here.  In that case, both direction flags will be set.
3436 	 */
3437 	sflags = 0;
3438 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
3439 		/*
3440 		 * We have to add a special case for SMP passthrough, there
3441 		 * is no easy way to generically handle it.  The first
3442 		 * S/G element is used for the command (therefore the
3443 		 * direction bit needs to be set).  The second one is used
3444 		 * for the reply.  We'll leave it to the caller to make
3445 		 * sure we only have two buffers.
3446 		 */
3447 		/*
3448 		 * Even though the busdma man page says it doesn't make
3449 		 * sense to have both direction flags, it does in this case.
3450 		 * We have one s/g element being accessed in each direction.
3451 		 */
3452 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
3453 
3454 		/*
3455 		 * Set the direction flag on the first buffer in the SMP
3456 		 * passthrough request.  We'll clear it for the second one.
3457 		 */
3458 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
3459 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
3460 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
3461 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
3462 		dir = BUS_DMASYNC_PREWRITE;
3463 	} else
3464 		dir = BUS_DMASYNC_PREREAD;
3465 
3466 	/* Check if a native SG list is needed for an NVMe PCIe device. */
3467 	if (cm->cm_targ && cm->cm_targ->is_nvme &&
3468 	    mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) {
3469 		/* A native SG list was built, skip to end. */
3470 		goto out;
3471 	}
3472 
3473 	for (i = 0; i < nsegs; i++) {
3474 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
3475 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3476 		}
3477 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3478 		    sflags, nsegs - i);
3479 		if (error != 0) {
3480 			/* Resource shortage, roll back! */
3481 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
3482 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
3483 				    "consider increasing hw.mpr.max_chains.\n");
3484 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
3485 			mpr_complete_command(sc, cm);
3486 			return;
3487 		}
3488 	}
3489 
3490 out:
3491 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3492 	mpr_enqueue_request(sc, cm);
3493 
3494 	return;
3495 }
3496 
3497 static void
3498 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3499 	     int error)
3500 {
3501 	mpr_data_cb(arg, segs, nsegs, error);
3502 }
3503 
3504 /*
3505  * This is the routine to enqueue commands ansynchronously.
3506  * Note that the only error path here is from bus_dmamap_load(), which can
3507  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3508  * assumed that if you have a command in-hand, then you have enough credits
3509  * to use it.
3510  */
3511 int
3512 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
3513 {
3514 	int error = 0;
3515 
3516 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
3517 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3518 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
3519 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
3520 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3521 		    cm->cm_data, mpr_data_cb, cm, 0);
3522 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3523 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3524 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
3525 	} else {
3526 		/* Add a zero-length element as needed */
3527 		if (cm->cm_sge != NULL)
3528 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
3529 		mpr_enqueue_request(sc, cm);
3530 	}
3531 
3532 	return (error);
3533 }
3534 
3535 /*
3536  * This is the routine to enqueue commands synchronously.  An error of
3537  * EINPROGRESS from mpr_map_command() is ignored since the command will
3538  * be executed and enqueued automatically.  Other errors come from msleep().
3539  */
3540 int
3541 mpr_wait_command(struct mpr_softc *sc, struct mpr_command **cmp, int timeout,
3542     int sleep_flag)
3543 {
3544 	int error, rc;
3545 	struct timeval cur_time, start_time;
3546 	struct mpr_command *cm = *cmp;
3547 
3548 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
3549 		return  EBUSY;
3550 
3551 	cm->cm_complete = NULL;
3552 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
3553 	error = mpr_map_command(sc, cm);
3554 	if ((error != 0) && (error != EINPROGRESS))
3555 		return (error);
3556 
3557 	// Check for context and wait for 50 mSec at a time until time has
3558 	// expired or the command has finished.  If msleep can't be used, need
3559 	// to poll.
3560 #if __FreeBSD_version >= 1000029
3561 	if (curthread->td_no_sleeping)
3562 #else //__FreeBSD_version < 1000029
3563 	if (curthread->td_pflags & TDP_NOSLEEPING)
3564 #endif //__FreeBSD_version >= 1000029
3565 		sleep_flag = NO_SLEEP;
3566 	getmicrouptime(&start_time);
3567 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
3568 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
3569 		if (error == EWOULDBLOCK) {
3570 			/*
3571 			 * Record the actual elapsed time in the case of a
3572 			 * timeout for the message below.
3573 			 */
3574 			getmicrouptime(&cur_time);
3575 			timevalsub(&cur_time, &start_time);
3576 		}
3577 	} else {
3578 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3579 			mpr_intr_locked(sc);
3580 			if (sleep_flag == CAN_SLEEP)
3581 				pause("mprwait", hz/20);
3582 			else
3583 				DELAY(50000);
3584 
3585 			getmicrouptime(&cur_time);
3586 			timevalsub(&cur_time, &start_time);
3587 			if (cur_time.tv_sec > timeout) {
3588 				error = EWOULDBLOCK;
3589 				break;
3590 			}
3591 		}
3592 	}
3593 
3594 	if (error == EWOULDBLOCK) {
3595 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s, timeout=%d,"
3596 		    " elapsed=%jd\n", __func__, timeout,
3597 		    (intmax_t)cur_time.tv_sec);
3598 		rc = mpr_reinit(sc);
3599 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3600 		    "failed");
3601 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3602 			/*
3603 			 * Tell the caller that we freed the command in a
3604 			 * reinit.
3605 			 */
3606 			*cmp = NULL;
3607 		}
3608 		error = ETIMEDOUT;
3609 	}
3610 	return (error);
3611 }
3612 
3613 /*
3614  * This is the routine to enqueue a command synchonously and poll for
3615  * completion.  Its use should be rare.
3616  */
3617 int
3618 mpr_request_polled(struct mpr_softc *sc, struct mpr_command **cmp)
3619 {
3620 	int error, rc;
3621 	struct timeval cur_time, start_time;
3622 	struct mpr_command *cm = *cmp;
3623 
3624 	error = 0;
3625 
3626 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
3627 	cm->cm_complete = NULL;
3628 	mpr_map_command(sc, cm);
3629 
3630 	getmicrouptime(&start_time);
3631 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3632 		mpr_intr_locked(sc);
3633 
3634 		if (mtx_owned(&sc->mpr_mtx))
3635 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
3636 			    "mprpoll", hz/20);
3637 		else
3638 			pause("mprpoll", hz/20);
3639 
3640 		/*
3641 		 * Check for real-time timeout and fail if more than 60 seconds.
3642 		 */
3643 		getmicrouptime(&cur_time);
3644 		timevalsub(&cur_time, &start_time);
3645 		if (cur_time.tv_sec > 60) {
3646 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
3647 			error = ETIMEDOUT;
3648 			break;
3649 		}
3650 	}
3651 
3652 	if (error) {
3653 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3654 		rc = mpr_reinit(sc);
3655 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3656 		    "failed");
3657 
3658 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3659 			/*
3660 			 * Tell the caller that we freed the command in a
3661 			 * reinit.
3662 			 */
3663 			*cmp = NULL;
3664 		}
3665 	}
3666 	return (error);
3667 }
3668 
3669 /*
3670  * The MPT driver had a verbose interface for config pages.  In this driver,
3671  * reduce it to much simpler terms, similar to the Linux driver.
3672  */
3673 int
3674 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3675 {
3676 	MPI2_CONFIG_REQUEST *req;
3677 	struct mpr_command *cm;
3678 	int error;
3679 
3680 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
3681 		return (EBUSY);
3682 	}
3683 
3684 	cm = mpr_alloc_command(sc);
3685 	if (cm == NULL) {
3686 		return (EBUSY);
3687 	}
3688 
3689 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3690 	req->Function = MPI2_FUNCTION_CONFIG;
3691 	req->Action = params->action;
3692 	req->SGLFlags = 0;
3693 	req->ChainOffset = 0;
3694 	req->PageAddress = params->page_address;
3695 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3696 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3697 
3698 		hdr = &params->hdr.Ext;
3699 		req->ExtPageType = hdr->ExtPageType;
3700 		req->ExtPageLength = hdr->ExtPageLength;
3701 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3702 		req->Header.PageLength = 0; /* Must be set to zero */
3703 		req->Header.PageNumber = hdr->PageNumber;
3704 		req->Header.PageVersion = hdr->PageVersion;
3705 	} else {
3706 		MPI2_CONFIG_PAGE_HEADER *hdr;
3707 
3708 		hdr = &params->hdr.Struct;
3709 		req->Header.PageType = hdr->PageType;
3710 		req->Header.PageNumber = hdr->PageNumber;
3711 		req->Header.PageLength = hdr->PageLength;
3712 		req->Header.PageVersion = hdr->PageVersion;
3713 	}
3714 
3715 	cm->cm_data = params->buffer;
3716 	cm->cm_length = params->length;
3717 	if (cm->cm_data != NULL) {
3718 		cm->cm_sge = &req->PageBufferSGE;
3719 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3720 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
3721 	} else
3722 		cm->cm_sge = NULL;
3723 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3724 
3725 	cm->cm_complete_data = params;
3726 	if (params->callback != NULL) {
3727 		cm->cm_complete = mpr_config_complete;
3728 		return (mpr_map_command(sc, cm));
3729 	} else {
3730 		error = mpr_wait_command(sc, &cm, 0, CAN_SLEEP);
3731 		if (error) {
3732 			mpr_dprint(sc, MPR_FAULT,
3733 			    "Error %d reading config page\n", error);
3734 			if (cm != NULL)
3735 				mpr_free_command(sc, cm);
3736 			return (error);
3737 		}
3738 		mpr_config_complete(sc, cm);
3739 	}
3740 
3741 	return (0);
3742 }
3743 
3744 int
3745 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3746 {
3747 	return (EINVAL);
3748 }
3749 
3750 static void
3751 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
3752 {
3753 	MPI2_CONFIG_REPLY *reply;
3754 	struct mpr_config_params *params;
3755 
3756 	MPR_FUNCTRACE(sc);
3757 	params = cm->cm_complete_data;
3758 
3759 	if (cm->cm_data != NULL) {
3760 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3761 		    BUS_DMASYNC_POSTREAD);
3762 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3763 	}
3764 
3765 	/*
3766 	 * XXX KDM need to do more error recovery?  This results in the
3767 	 * device in question not getting probed.
3768 	 */
3769 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
3770 		params->status = MPI2_IOCSTATUS_BUSY;
3771 		goto done;
3772 	}
3773 
3774 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3775 	if (reply == NULL) {
3776 		params->status = MPI2_IOCSTATUS_BUSY;
3777 		goto done;
3778 	}
3779 	params->status = reply->IOCStatus;
3780 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3781 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3782 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3783 		params->hdr.Ext.PageType = reply->Header.PageType;
3784 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3785 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3786 	} else {
3787 		params->hdr.Struct.PageType = reply->Header.PageType;
3788 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3789 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3790 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3791 	}
3792 
3793 done:
3794 	mpr_free_command(sc, cm);
3795 	if (params->callback != NULL)
3796 		params->callback(sc, params);
3797 
3798 	return;
3799 }
3800