xref: /freebsd/sys/dev/mpr/mpr.c (revision 99429157e8615dc3b7f11afbe3ed92de7476a5db)
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/queue.h>
53 #include <sys/kthread.h>
54 #include <sys/taskqueue.h>
55 #include <sys/endian.h>
56 #include <sys/eventhandler.h>
57 
58 #include <machine/bus.h>
59 #include <machine/resource.h>
60 #include <sys/rman.h>
61 #include <sys/proc.h>
62 
63 #include <dev/pci/pcivar.h>
64 
65 #include <cam/cam.h>
66 #include <cam/cam_ccb.h>
67 #include <cam/scsi/scsi_all.h>
68 
69 #include <dev/mpr/mpi/mpi2_type.h>
70 #include <dev/mpr/mpi/mpi2.h>
71 #include <dev/mpr/mpi/mpi2_ioc.h>
72 #include <dev/mpr/mpi/mpi2_sas.h>
73 #include <dev/mpr/mpi/mpi2_pci.h>
74 #include <dev/mpr/mpi/mpi2_cnfg.h>
75 #include <dev/mpr/mpi/mpi2_init.h>
76 #include <dev/mpr/mpi/mpi2_tool.h>
77 #include <dev/mpr/mpr_ioctl.h>
78 #include <dev/mpr/mprvar.h>
79 #include <dev/mpr/mpr_table.h>
80 #include <dev/mpr/mpr_sas.h>
81 
82 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
83 static int mpr_init_queues(struct mpr_softc *sc);
84 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
85 static int mpr_transition_operational(struct mpr_softc *sc);
86 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
87 static void mpr_iocfacts_free(struct mpr_softc *sc);
88 static void mpr_startup(void *arg);
89 static int mpr_send_iocinit(struct mpr_softc *sc);
90 static int mpr_alloc_queues(struct mpr_softc *sc);
91 static int mpr_alloc_replies(struct mpr_softc *sc);
92 static int mpr_alloc_requests(struct mpr_softc *sc);
93 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc);
94 static int mpr_attach_log(struct mpr_softc *sc);
95 static __inline void mpr_complete_command(struct mpr_softc *sc,
96     struct mpr_command *cm);
97 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
98     MPI2_EVENT_NOTIFICATION_REPLY *reply);
99 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
100 static void mpr_periodic(void *);
101 static int mpr_reregister_events(struct mpr_softc *sc);
102 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
103 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
104 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
105 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD, 0, "MPR Driver Parameters");
106 
107 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory");
108 
109 /*
110  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
111  * any state and back to its initialization state machine.
112  */
113 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
114 
115 /*
116  * Added this union to smoothly convert le64toh cm->cm_desc.Words.
117  * Compiler only supports uint64_t to be passed as an argument.
118  * Otherwise it will through this error:
119  * "aggregate value used where an integer was expected"
120  */
121 typedef union _reply_descriptor {
122         u64 word;
123         struct {
124                 u32 low;
125                 u32 high;
126         } u;
127 } reply_descriptor, request_descriptor;
128 
129 /* Rate limit chain-fail messages to 1 per minute */
130 static struct timeval mpr_chainfail_interval = { 60, 0 };
131 
132 /*
133  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
134  * If this function is called from process context, it can sleep
135  * and there is no harm to sleep, in case if this fuction is called
136  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
137  * based on sleep flags driver will call either msleep, pause or DELAY.
138  * msleep and pause are of same variant, but pause is used when mpr_mtx
139  * is not hold by driver.
140  */
141 static int
142 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag)
143 {
144 	uint32_t reg;
145 	int i, error, tries = 0;
146 	uint8_t first_wait_done = FALSE;
147 
148 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
149 
150 	/* Clear any pending interrupts */
151 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
152 
153 	/*
154 	 * Force NO_SLEEP for threads prohibited to sleep
155  	 * e.a Thread from interrupt handler are prohibited to sleep.
156  	 */
157 #if __FreeBSD_version >= 1000029
158 	if (curthread->td_no_sleeping)
159 #else //__FreeBSD_version < 1000029
160 	if (curthread->td_pflags & TDP_NOSLEEPING)
161 #endif //__FreeBSD_version >= 1000029
162 		sleep_flag = NO_SLEEP;
163 
164 	/* Push the magic sequence */
165 	error = ETIMEDOUT;
166 	while (tries++ < 20) {
167 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
168 			mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
169 			    mpt2_reset_magic[i]);
170 
171 		/* wait 100 msec */
172 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
173 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
174 			    "mprdiag", hz/10);
175 		else if (sleep_flag == CAN_SLEEP)
176 			pause("mprdiag", hz/10);
177 		else
178 			DELAY(100 * 1000);
179 
180 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
181 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
182 			error = 0;
183 			break;
184 		}
185 	}
186 	if (error)
187 		return (error);
188 
189 	/* Send the actual reset.  XXX need to refresh the reg? */
190 	mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
191 	    reg | MPI2_DIAG_RESET_ADAPTER);
192 
193 	/* Wait up to 300 seconds in 50ms intervals */
194 	error = ETIMEDOUT;
195 	for (i = 0; i < 6000; i++) {
196 		/*
197 		 * Wait 50 msec. If this is the first time through, wait 256
198 		 * msec to satisfy Diag Reset timing requirements.
199 		 */
200 		if (first_wait_done) {
201 			if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
202 				msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
203 				    "mprdiag", hz/20);
204 			else if (sleep_flag == CAN_SLEEP)
205 				pause("mprdiag", hz/20);
206 			else
207 				DELAY(50 * 1000);
208 		} else {
209 			DELAY(256 * 1000);
210 			first_wait_done = TRUE;
211 		}
212 		/*
213 		 * Check for the RESET_ADAPTER bit to be cleared first, then
214 		 * wait for the RESET state to be cleared, which takes a little
215 		 * longer.
216 		 */
217 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
218 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
219 			continue;
220 		}
221 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
222 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
223 			error = 0;
224 			break;
225 		}
226 	}
227 	if (error)
228 		return (error);
229 
230 	mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
231 
232 	return (0);
233 }
234 
235 static int
236 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag)
237 {
238 
239 	MPR_FUNCTRACE(sc);
240 
241 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
242 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
243 	    MPI2_DOORBELL_FUNCTION_SHIFT);
244 
245 	if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) {
246 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed : <%s>\n",
247 				__func__);
248 		return (ETIMEDOUT);
249 	}
250 
251 	return (0);
252 }
253 
254 static int
255 mpr_transition_ready(struct mpr_softc *sc)
256 {
257 	uint32_t reg, state;
258 	int error, tries = 0;
259 	int sleep_flags;
260 
261 	MPR_FUNCTRACE(sc);
262 	/* If we are in attach call, do not sleep */
263 	sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE)
264 	    ? CAN_SLEEP : NO_SLEEP;
265 
266 	error = 0;
267 	while (tries++ < 1200) {
268 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
269 		mpr_dprint(sc, MPR_INIT, "Doorbell= 0x%x\n", reg);
270 
271 		/*
272 		 * Ensure the IOC is ready to talk.  If it's not, try
273 		 * resetting it.
274 		 */
275 		if (reg & MPI2_DOORBELL_USED) {
276 			mpr_diag_reset(sc, sleep_flags);
277 			DELAY(50000);
278 			continue;
279 		}
280 
281 		/* Is the adapter owned by another peer? */
282 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
283 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
284 			device_printf(sc->mpr_dev, "IOC is under the control "
285 			    "of another peer host, aborting initialization.\n");
286 			return (ENXIO);
287 		}
288 
289 		state = reg & MPI2_IOC_STATE_MASK;
290 		if (state == MPI2_IOC_STATE_READY) {
291 			/* Ready to go! */
292 			error = 0;
293 			break;
294 		} else if (state == MPI2_IOC_STATE_FAULT) {
295 			mpr_dprint(sc, MPR_FAULT, "IOC in fault state 0x%x\n",
296 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
297 			mpr_diag_reset(sc, sleep_flags);
298 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
299 			/* Need to take ownership */
300 			mpr_message_unit_reset(sc, sleep_flags);
301 		} else if (state == MPI2_IOC_STATE_RESET) {
302 			/* Wait a bit, IOC might be in transition */
303 			mpr_dprint(sc, MPR_FAULT,
304 			    "IOC in unexpected reset state\n");
305 		} else {
306 			mpr_dprint(sc, MPR_FAULT,
307 			    "IOC in unknown state 0x%x\n", state);
308 			error = EINVAL;
309 			break;
310 		}
311 
312 		/* Wait 50ms for things to settle down. */
313 		DELAY(50000);
314 	}
315 
316 	if (error)
317 		device_printf(sc->mpr_dev, "Cannot transition IOC to ready\n");
318 	return (error);
319 }
320 
321 static int
322 mpr_transition_operational(struct mpr_softc *sc)
323 {
324 	uint32_t reg, state;
325 	int error;
326 
327 	MPR_FUNCTRACE(sc);
328 
329 	error = 0;
330 	reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
331 	mpr_dprint(sc, MPR_INIT, "Doorbell= 0x%x\n", reg);
332 
333 	state = reg & MPI2_IOC_STATE_MASK;
334 	if (state != MPI2_IOC_STATE_READY) {
335 		if ((error = mpr_transition_ready(sc)) != 0) {
336 			mpr_dprint(sc, MPR_FAULT,
337 			    "%s failed to transition ready\n", __func__);
338 			return (error);
339 		}
340 	}
341 
342 	error = mpr_send_iocinit(sc);
343 	return (error);
344 }
345 
346 /*
347  * This is called during attach and when re-initializing due to a Diag Reset.
348  * IOC Facts is used to allocate many of the structures needed by the driver.
349  * If called from attach, de-allocation is not required because the driver has
350  * not allocated any structures yet, but if called from a Diag Reset, previously
351  * allocated structures based on IOC Facts will need to be freed and re-
352  * allocated bases on the latest IOC Facts.
353  */
354 static int
355 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
356 {
357 	int error;
358 	Mpi2IOCFactsReply_t saved_facts;
359 	uint8_t saved_mode, reallocating;
360 
361 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
362 
363 	/* Save old IOC Facts and then only reallocate if Facts have changed */
364 	if (!attaching) {
365 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
366 	}
367 
368 	/*
369 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
370 	 * a re-initialization and only return the error if attaching so the OS
371 	 * can handle it.
372 	 */
373 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
374 		if (attaching) {
375 			mpr_dprint(sc, MPR_FAULT, "%s failed to get IOC Facts "
376 			    "with error %d\n", __func__, error);
377 			return (error);
378 		} else {
379 			panic("%s failed to get IOC Facts with error %d\n",
380 			    __func__, error);
381 		}
382 	}
383 
384 	mpr_print_iocfacts(sc, sc->facts);
385 
386 	snprintf(sc->fw_version, sizeof(sc->fw_version),
387 	    "%02d.%02d.%02d.%02d",
388 	    sc->facts->FWVersion.Struct.Major,
389 	    sc->facts->FWVersion.Struct.Minor,
390 	    sc->facts->FWVersion.Struct.Unit,
391 	    sc->facts->FWVersion.Struct.Dev);
392 
393 	mpr_printf(sc, "Firmware: %s, Driver: %s\n", sc->fw_version,
394 	    MPR_DRIVER_VERSION);
395 	mpr_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
396 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
397 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
398 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc"
399 	    "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV");
400 
401 	/*
402 	 * If the chip doesn't support event replay then a hard reset will be
403 	 * required to trigger a full discovery.  Do the reset here then
404 	 * retransition to Ready.  A hard reset might have already been done,
405 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
406 	 * for a Diag Reset.
407 	 */
408 	if (attaching) {
409 		if ((sc->facts->IOCCapabilities &
410 		    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
411 			mpr_diag_reset(sc, NO_SLEEP);
412 			if ((error = mpr_transition_ready(sc)) != 0) {
413 				mpr_dprint(sc, MPR_FAULT, "%s failed to "
414 				    "transition to ready with error %d\n",
415 				    __func__, error);
416 				return (error);
417 			}
418 		}
419 	}
420 
421 	/*
422 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
423 	 * changed from the previous IOC Facts, log a warning, but only if
424 	 * checking this after a Diag Reset and not during attach.
425 	 */
426 	saved_mode = sc->ir_firmware;
427 	if (sc->facts->IOCCapabilities &
428 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
429 		sc->ir_firmware = 1;
430 	if (!attaching) {
431 		if (sc->ir_firmware != saved_mode) {
432 			mpr_dprint(sc, MPR_FAULT, "%s new IR/IT mode in IOC "
433 			    "Facts does not match previous mode\n", __func__);
434 		}
435 	}
436 
437 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
438 	reallocating = FALSE;
439 	if ((!attaching) &&
440 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
441 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
442 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
443 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
444 	    (saved_facts.ProductID != sc->facts->ProductID) ||
445 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
446 	    (saved_facts.IOCRequestFrameSize !=
447 	    sc->facts->IOCRequestFrameSize) ||
448 	    (saved_facts.IOCMaxChainSegmentSize !=
449 	    sc->facts->IOCMaxChainSegmentSize) ||
450 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
451 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
452 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
453 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
454 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
455 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
456 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
457 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
458 	    (saved_facts.MaxPersistentEntries !=
459 	    sc->facts->MaxPersistentEntries))) {
460 		reallocating = TRUE;
461 	}
462 
463 	/*
464 	 * Some things should be done if attaching or re-allocating after a Diag
465 	 * Reset, but are not needed after a Diag Reset if the FW has not
466 	 * changed.
467 	 */
468 	if (attaching || reallocating) {
469 		/*
470 		 * Check if controller supports FW diag buffers and set flag to
471 		 * enable each type.
472 		 */
473 		if (sc->facts->IOCCapabilities &
474 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
475 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
476 			    enabled = TRUE;
477 		if (sc->facts->IOCCapabilities &
478 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
479 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
480 			    enabled = TRUE;
481 		if (sc->facts->IOCCapabilities &
482 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
483 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
484 			    enabled = TRUE;
485 
486 		/*
487 		 * Set flags for some supported items.
488 		 */
489 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
490 			sc->eedp_enabled = TRUE;
491 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
492 			sc->control_TLR = TRUE;
493 		if (sc->facts->IOCCapabilities &
494 		    MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ)
495 			sc->atomic_desc_capable = TRUE;
496 
497 		/*
498 		 * Size the queues. Since the reply queues always need one free
499 		 * entry, we'll just deduct one reply message here.
500 		 */
501 		sc->num_reqs = MIN(MPR_REQ_FRAMES, sc->facts->RequestCredit);
502 		sc->num_replies = MIN(MPR_REPLY_FRAMES + MPR_EVT_REPLY_FRAMES,
503 		    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
504 
505 		/*
506 		 * Initialize all Tail Queues
507 		 */
508 		TAILQ_INIT(&sc->req_list);
509 		TAILQ_INIT(&sc->high_priority_req_list);
510 		TAILQ_INIT(&sc->chain_list);
511 		TAILQ_INIT(&sc->prp_page_list);
512 		TAILQ_INIT(&sc->tm_list);
513 	}
514 
515 	/*
516 	 * If doing a Diag Reset and the FW is significantly different
517 	 * (reallocating will be set above in IOC Facts comparison), then all
518 	 * buffers based on the IOC Facts will need to be freed before they are
519 	 * reallocated.
520 	 */
521 	if (reallocating) {
522 		mpr_iocfacts_free(sc);
523 		mprsas_realloc_targets(sc, saved_facts.MaxTargets);
524 	}
525 
526 	/*
527 	 * Any deallocation has been completed.  Now start reallocating
528 	 * if needed.  Will only need to reallocate if attaching or if the new
529 	 * IOC Facts are different from the previous IOC Facts after a Diag
530 	 * Reset. Targets have already been allocated above if needed.
531 	 */
532 	if (attaching || reallocating) {
533 		if (((error = mpr_alloc_queues(sc)) != 0) ||
534 		    ((error = mpr_alloc_replies(sc)) != 0) ||
535 		    ((error = mpr_alloc_requests(sc)) != 0)) {
536 			if (attaching ) {
537 				mpr_dprint(sc, MPR_FAULT, "%s failed to alloc "
538 				    "queues with error %d\n", __func__, error);
539 				mpr_free(sc);
540 				return (error);
541 			} else {
542 				panic("%s failed to alloc queues with error "
543 				    "%d\n", __func__, error);
544 			}
545 		}
546 	}
547 
548 	/* Always initialize the queues */
549 	bzero(sc->free_queue, sc->fqdepth * 4);
550 	mpr_init_queues(sc);
551 
552 	/*
553 	 * Always get the chip out of the reset state, but only panic if not
554 	 * attaching.  If attaching and there is an error, that is handled by
555 	 * the OS.
556 	 */
557 	error = mpr_transition_operational(sc);
558 	if (error != 0) {
559 		if (attaching) {
560 			mpr_printf(sc, "%s failed to transition to operational "
561 			    "with error %d\n", __func__, error);
562 			mpr_free(sc);
563 			return (error);
564 		} else {
565 			panic("%s failed to transition to operational with "
566 			    "error %d\n", __func__, error);
567 		}
568 	}
569 
570 	/*
571 	 * Finish the queue initialization.
572 	 * These are set here instead of in mpr_init_queues() because the
573 	 * IOC resets these values during the state transition in
574 	 * mpr_transition_operational().  The free index is set to 1
575 	 * because the corresponding index in the IOC is set to 0, and the
576 	 * IOC treats the queues as full if both are set to the same value.
577 	 * Hence the reason that the queue can't hold all of the possible
578 	 * replies.
579 	 */
580 	sc->replypostindex = 0;
581 	mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
582 	mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
583 
584 	/*
585 	 * Attach the subsystems so they can prepare their event masks.
586 	 */
587 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
588 	if (attaching) {
589 		if (((error = mpr_attach_log(sc)) != 0) ||
590 		    ((error = mpr_attach_sas(sc)) != 0) ||
591 		    ((error = mpr_attach_user(sc)) != 0)) {
592 			mpr_printf(sc, "%s failed to attach all subsystems: "
593 			    "error %d\n", __func__, error);
594 			mpr_free(sc);
595 			return (error);
596 		}
597 
598 		if ((error = mpr_pci_setup_interrupts(sc)) != 0) {
599 			mpr_printf(sc, "%s failed to setup interrupts\n",
600 			    __func__);
601 			mpr_free(sc);
602 			return (error);
603 		}
604 	}
605 
606 	return (error);
607 }
608 
609 /*
610  * This is called if memory is being free (during detach for example) and when
611  * buffers need to be reallocated due to a Diag Reset.
612  */
613 static void
614 mpr_iocfacts_free(struct mpr_softc *sc)
615 {
616 	struct mpr_command *cm;
617 	int i;
618 
619 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
620 
621 	if (sc->free_busaddr != 0)
622 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
623 	if (sc->free_queue != NULL)
624 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
625 		    sc->queues_map);
626 	if (sc->queues_dmat != NULL)
627 		bus_dma_tag_destroy(sc->queues_dmat);
628 
629 	if (sc->chain_busaddr != 0)
630 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
631 	if (sc->chain_frames != NULL)
632 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
633 		    sc->chain_map);
634 	if (sc->chain_dmat != NULL)
635 		bus_dma_tag_destroy(sc->chain_dmat);
636 
637 	if (sc->sense_busaddr != 0)
638 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
639 	if (sc->sense_frames != NULL)
640 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
641 		    sc->sense_map);
642 	if (sc->sense_dmat != NULL)
643 		bus_dma_tag_destroy(sc->sense_dmat);
644 
645 	if (sc->prp_page_busaddr != 0)
646 		bus_dmamap_unload(sc->prp_page_dmat, sc->prp_page_map);
647 	if (sc->prp_pages != NULL)
648 		bus_dmamem_free(sc->prp_page_dmat, sc->prp_pages,
649 		    sc->prp_page_map);
650 	if (sc->prp_page_dmat != NULL)
651 		bus_dma_tag_destroy(sc->prp_page_dmat);
652 
653 	if (sc->reply_busaddr != 0)
654 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
655 	if (sc->reply_frames != NULL)
656 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
657 		    sc->reply_map);
658 	if (sc->reply_dmat != NULL)
659 		bus_dma_tag_destroy(sc->reply_dmat);
660 
661 	if (sc->req_busaddr != 0)
662 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
663 	if (sc->req_frames != NULL)
664 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
665 	if (sc->req_dmat != NULL)
666 		bus_dma_tag_destroy(sc->req_dmat);
667 
668 	if (sc->chains != NULL)
669 		free(sc->chains, M_MPR);
670 	if (sc->prps != NULL)
671 		free(sc->prps, M_MPR);
672 	if (sc->commands != NULL) {
673 		for (i = 1; i < sc->num_reqs; i++) {
674 			cm = &sc->commands[i];
675 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
676 		}
677 		free(sc->commands, M_MPR);
678 	}
679 	if (sc->buffer_dmat != NULL)
680 		bus_dma_tag_destroy(sc->buffer_dmat);
681 }
682 
683 /*
684  * The terms diag reset and hard reset are used interchangeably in the MPI
685  * docs to mean resetting the controller chip.  In this code diag reset
686  * cleans everything up, and the hard reset function just sends the reset
687  * sequence to the chip.  This should probably be refactored so that every
688  * subsystem gets a reset notification of some sort, and can clean up
689  * appropriately.
690  */
691 int
692 mpr_reinit(struct mpr_softc *sc)
693 {
694 	int error;
695 	struct mprsas_softc *sassc;
696 
697 	sassc = sc->sassc;
698 
699 	MPR_FUNCTRACE(sc);
700 
701 	mtx_assert(&sc->mpr_mtx, MA_OWNED);
702 
703 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) {
704 		mpr_dprint(sc, MPR_INIT, "%s reset already in progress\n",
705 		    __func__);
706 		return 0;
707 	}
708 
709 	mpr_dprint(sc, MPR_INFO, "Reinitializing controller,\n");
710 	/* make sure the completion callbacks can recognize they're getting
711 	 * a NULL cm_reply due to a reset.
712 	 */
713 	sc->mpr_flags |= MPR_FLAGS_DIAGRESET;
714 
715 	/*
716 	 * Mask interrupts here.
717 	 */
718 	mpr_dprint(sc, MPR_INIT, "%s mask interrupts\n", __func__);
719 	mpr_mask_intr(sc);
720 
721 	error = mpr_diag_reset(sc, CAN_SLEEP);
722 	if (error != 0) {
723 		panic("%s hard reset failed with error %d\n", __func__, error);
724 	}
725 
726 	/* Restore the PCI state, including the MSI-X registers */
727 	mpr_pci_restore(sc);
728 
729 	/* Give the I/O subsystem special priority to get itself prepared */
730 	mprsas_handle_reinit(sc);
731 
732 	/*
733 	 * Get IOC Facts and allocate all structures based on this information.
734 	 * The attach function will also call mpr_iocfacts_allocate at startup.
735 	 * If relevant values have changed in IOC Facts, this function will free
736 	 * all of the memory based on IOC Facts and reallocate that memory.
737 	 */
738 	if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) {
739 		panic("%s IOC Facts based allocation failed with error %d\n",
740 		    __func__, error);
741 	}
742 
743 	/*
744 	 * Mapping structures will be re-allocated after getting IOC Page8, so
745 	 * free these structures here.
746 	 */
747 	mpr_mapping_exit(sc);
748 
749 	/*
750 	 * The static page function currently read is IOC Page8.  Others can be
751 	 * added in future.  It's possible that the values in IOC Page8 have
752 	 * changed after a Diag Reset due to user modification, so always read
753 	 * these.  Interrupts are masked, so unmask them before getting config
754 	 * pages.
755 	 */
756 	mpr_unmask_intr(sc);
757 	sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET;
758 	mpr_base_static_config_pages(sc);
759 
760 	/*
761 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
762 	 * mapping tables.
763 	 */
764 	mpr_mapping_initialize(sc);
765 
766 	/*
767 	 * Restart will reload the event masks clobbered by the reset, and
768 	 * then enable the port.
769 	 */
770 	mpr_reregister_events(sc);
771 
772 	/* the end of discovery will release the simq, so we're done. */
773 	mpr_dprint(sc, MPR_INFO, "%s finished sc %p post %u free %u\n",
774 	    __func__, sc, sc->replypostindex, sc->replyfreeindex);
775 	mprsas_release_simq_reinit(sassc);
776 
777 	return 0;
778 }
779 
780 /* Wait for the chip to ACK a word that we've put into its FIFO
781  * Wait for <timeout> seconds. In single loop wait for busy loop
782  * for 500 microseconds.
783  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
784  * */
785 static int
786 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag)
787 {
788 	u32 cntdn, count;
789 	u32 int_status;
790 	u32 doorbell;
791 
792 	count = 0;
793 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
794 	do {
795 		int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
796 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
797 			mpr_dprint(sc, MPR_INIT, "%s: successful count(%d), "
798 			    "timeout(%d)\n", __func__, count, timeout);
799 			return 0;
800 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
801 			doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
802 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
803 			    MPI2_IOC_STATE_FAULT) {
804 				mpr_dprint(sc, MPR_FAULT,
805 				    "fault_state(0x%04x)!\n", doorbell);
806 				return (EFAULT);
807 			}
808 		} else if (int_status == 0xFFFFFFFF)
809 			goto out;
810 
811 		/*
812 		 * If it can sleep, sleep for 1 milisecond, else busy loop for
813  		 * 0.5 milisecond
814 		 */
815 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
816 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba",
817 			    hz/1000);
818 		else if (sleep_flag == CAN_SLEEP)
819 			pause("mprdba", hz/1000);
820 		else
821 			DELAY(500);
822 		count++;
823 	} while (--cntdn);
824 
825 out:
826 	mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), "
827 		"int_status(%x)!\n", __func__, count, int_status);
828 	return (ETIMEDOUT);
829 }
830 
831 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
832 static int
833 mpr_wait_db_int(struct mpr_softc *sc)
834 {
835 	int retry;
836 
837 	for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) {
838 		if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
839 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
840 			return (0);
841 		DELAY(2000);
842 	}
843 	return (ETIMEDOUT);
844 }
845 
846 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
847 static int
848 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
849     int req_sz, int reply_sz, int timeout)
850 {
851 	uint32_t *data32;
852 	uint16_t *data16;
853 	int i, count, ioc_sz, residual;
854 	int sleep_flags = CAN_SLEEP;
855 
856 #if __FreeBSD_version >= 1000029
857 	if (curthread->td_no_sleeping)
858 #else //__FreeBSD_version < 1000029
859 	if (curthread->td_pflags & TDP_NOSLEEPING)
860 #endif //__FreeBSD_version >= 1000029
861 		sleep_flags = NO_SLEEP;
862 
863 	/* Step 1 */
864 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
865 
866 	/* Step 2 */
867 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
868 		return (EBUSY);
869 
870 	/* Step 3
871 	 * Announce that a message is coming through the doorbell.  Messages
872 	 * are pushed at 32bit words, so round up if needed.
873 	 */
874 	count = (req_sz + 3) / 4;
875 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
876 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
877 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
878 
879 	/* Step 4 */
880 	if (mpr_wait_db_int(sc) ||
881 	    (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
882 		mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n");
883 		return (ENXIO);
884 	}
885 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
886 	if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
887 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n");
888 		return (ENXIO);
889 	}
890 
891 	/* Step 5 */
892 	/* Clock out the message data synchronously in 32-bit dwords*/
893 	data32 = (uint32_t *)req;
894 	for (i = 0; i < count; i++) {
895 		mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
896 		if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
897 			mpr_dprint(sc, MPR_FAULT,
898 			    "Timeout while writing doorbell\n");
899 			return (ENXIO);
900 		}
901 	}
902 
903 	/* Step 6 */
904 	/* Clock in the reply in 16-bit words.  The total length of the
905 	 * message is always in the 4th byte, so clock out the first 2 words
906 	 * manually, then loop the rest.
907 	 */
908 	data16 = (uint16_t *)reply;
909 	if (mpr_wait_db_int(sc) != 0) {
910 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n");
911 		return (ENXIO);
912 	}
913 	data16[0] =
914 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
915 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
916 	if (mpr_wait_db_int(sc) != 0) {
917 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n");
918 		return (ENXIO);
919 	}
920 	data16[1] =
921 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
922 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
923 
924 	/* Number of 32bit words in the message */
925 	ioc_sz = reply->MsgLength;
926 
927 	/*
928 	 * Figure out how many 16bit words to clock in without overrunning.
929 	 * The precision loss with dividing reply_sz can safely be
930 	 * ignored because the messages can only be multiples of 32bits.
931 	 */
932 	residual = 0;
933 	count = MIN((reply_sz / 4), ioc_sz) * 2;
934 	if (count < ioc_sz * 2) {
935 		residual = ioc_sz * 2 - count;
936 		mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d "
937 		    "residual message words\n", residual);
938 	}
939 
940 	for (i = 2; i < count; i++) {
941 		if (mpr_wait_db_int(sc) != 0) {
942 			mpr_dprint(sc, MPR_FAULT,
943 			    "Timeout reading doorbell %d\n", i);
944 			return (ENXIO);
945 		}
946 		data16[i] = mpr_regread(sc, MPI2_DOORBELL_OFFSET) &
947 		    MPI2_DOORBELL_DATA_MASK;
948 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
949 	}
950 
951 	/*
952 	 * Pull out residual words that won't fit into the provided buffer.
953 	 * This keeps the chip from hanging due to a driver programming
954 	 * error.
955 	 */
956 	while (residual--) {
957 		if (mpr_wait_db_int(sc) != 0) {
958 			mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n");
959 			return (ENXIO);
960 		}
961 		(void)mpr_regread(sc, MPI2_DOORBELL_OFFSET);
962 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
963 	}
964 
965 	/* Step 7 */
966 	if (mpr_wait_db_int(sc) != 0) {
967 		mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n");
968 		return (ENXIO);
969 	}
970 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
971 		mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n");
972 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
973 
974 	return (0);
975 }
976 
977 static void
978 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm)
979 {
980 	request_descriptor rd;
981 
982 	MPR_FUNCTRACE(sc);
983 	mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n",
984 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
985 
986 	if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags &
987 	    MPR_FLAGS_SHUTDOWN))
988 		mtx_assert(&sc->mpr_mtx, MA_OWNED);
989 
990 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
991 		sc->io_cmds_highwater++;
992 
993 	if (sc->atomic_desc_capable) {
994 		rd.u.low = cm->cm_desc.Words.Low;
995 		mpr_regwrite(sc, MPI26_ATOMIC_REQUEST_DESCRIPTOR_POST_OFFSET,
996 		    rd.u.low);
997 	} else {
998 		rd.u.low = cm->cm_desc.Words.Low;
999 		rd.u.high = cm->cm_desc.Words.High;
1000 		rd.word = htole64(rd.word);
1001 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1002 		    rd.u.low);
1003 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1004 		    rd.u.high);
1005 	}
1006 }
1007 
1008 /*
1009  * Just the FACTS, ma'am.
1010  */
1011 static int
1012 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1013 {
1014 	MPI2_DEFAULT_REPLY *reply;
1015 	MPI2_IOC_FACTS_REQUEST request;
1016 	int error, req_sz, reply_sz;
1017 
1018 	MPR_FUNCTRACE(sc);
1019 
1020 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1021 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1022 	reply = (MPI2_DEFAULT_REPLY *)facts;
1023 
1024 	bzero(&request, req_sz);
1025 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1026 	error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1027 
1028 	return (error);
1029 }
1030 
1031 static int
1032 mpr_send_iocinit(struct mpr_softc *sc)
1033 {
1034 	MPI2_IOC_INIT_REQUEST	init;
1035 	MPI2_DEFAULT_REPLY	reply;
1036 	int req_sz, reply_sz, error;
1037 	struct timeval now;
1038 	uint64_t time_in_msec;
1039 
1040 	MPR_FUNCTRACE(sc);
1041 
1042 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1043 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1044 	bzero(&init, req_sz);
1045 	bzero(&reply, reply_sz);
1046 
1047 	/*
1048 	 * Fill in the init block.  Note that most addresses are
1049 	 * deliberately in the lower 32bits of memory.  This is a micro-
1050 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1051 	 */
1052 	init.Function = MPI2_FUNCTION_IOC_INIT;
1053 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1054 	init.MsgVersion = htole16(MPI2_VERSION);
1055 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1056 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
1057 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1058 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1059 	init.SenseBufferAddressHigh = 0;
1060 	init.SystemReplyAddressHigh = 0;
1061 	init.SystemRequestFrameBaseAddress.High = 0;
1062 	init.SystemRequestFrameBaseAddress.Low =
1063 	    htole32((uint32_t)sc->req_busaddr);
1064 	init.ReplyDescriptorPostQueueAddress.High = 0;
1065 	init.ReplyDescriptorPostQueueAddress.Low =
1066 	    htole32((uint32_t)sc->post_busaddr);
1067 	init.ReplyFreeQueueAddress.High = 0;
1068 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1069 	getmicrotime(&now);
1070 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1071 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1072 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1073 	init.HostPageSize = HOST_PAGE_SIZE_4K;
1074 
1075 	error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1076 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1077 		error = ENXIO;
1078 
1079 	mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1080 	return (error);
1081 }
1082 
1083 void
1084 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1085 {
1086 	bus_addr_t *addr;
1087 
1088 	addr = arg;
1089 	*addr = segs[0].ds_addr;
1090 }
1091 
1092 static int
1093 mpr_alloc_queues(struct mpr_softc *sc)
1094 {
1095 	bus_addr_t queues_busaddr;
1096 	uint8_t *queues;
1097 	int qsize, fqsize, pqsize;
1098 
1099 	/*
1100 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1101 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1102 	 * This queue supplies fresh reply frames for the firmware to use.
1103 	 *
1104 	 * The reply descriptor post queue contains 8 byte entries in
1105 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1106 	 * contains filled-in reply frames sent from the firmware to the host.
1107 	 *
1108 	 * These two queues are allocated together for simplicity.
1109 	 */
1110 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1111 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1112 	fqsize= sc->fqdepth * 4;
1113 	pqsize = sc->pqdepth * 8;
1114 	qsize = fqsize + pqsize;
1115 
1116         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1117 				16, 0,			/* algnmnt, boundary */
1118 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1119 				BUS_SPACE_MAXADDR,	/* highaddr */
1120 				NULL, NULL,		/* filter, filterarg */
1121                                 qsize,			/* maxsize */
1122                                 1,			/* nsegments */
1123                                 qsize,			/* maxsegsize */
1124                                 0,			/* flags */
1125                                 NULL, NULL,		/* lockfunc, lockarg */
1126                                 &sc->queues_dmat)) {
1127 		device_printf(sc->mpr_dev, "Cannot allocate queues DMA tag\n");
1128 		return (ENOMEM);
1129         }
1130         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1131 	    &sc->queues_map)) {
1132 		device_printf(sc->mpr_dev, "Cannot allocate queues memory\n");
1133 		return (ENOMEM);
1134         }
1135         bzero(queues, qsize);
1136         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1137 	    mpr_memaddr_cb, &queues_busaddr, 0);
1138 
1139 	sc->free_queue = (uint32_t *)queues;
1140 	sc->free_busaddr = queues_busaddr;
1141 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1142 	sc->post_busaddr = queues_busaddr + fqsize;
1143 
1144 	return (0);
1145 }
1146 
1147 static int
1148 mpr_alloc_replies(struct mpr_softc *sc)
1149 {
1150 	int rsize, num_replies;
1151 
1152 	/*
1153 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1154 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1155 	 * replies can be used at once.
1156 	 */
1157 	num_replies = max(sc->fqdepth, sc->num_replies);
1158 
1159 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
1160         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1161 				4, 0,			/* algnmnt, boundary */
1162 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1163 				BUS_SPACE_MAXADDR,	/* highaddr */
1164 				NULL, NULL,		/* filter, filterarg */
1165                                 rsize,			/* maxsize */
1166                                 1,			/* nsegments */
1167                                 rsize,			/* maxsegsize */
1168                                 0,			/* flags */
1169                                 NULL, NULL,		/* lockfunc, lockarg */
1170                                 &sc->reply_dmat)) {
1171 		device_printf(sc->mpr_dev, "Cannot allocate replies DMA tag\n");
1172 		return (ENOMEM);
1173         }
1174         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1175 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1176 		device_printf(sc->mpr_dev, "Cannot allocate replies memory\n");
1177 		return (ENOMEM);
1178         }
1179         bzero(sc->reply_frames, rsize);
1180         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1181 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1182 
1183 	return (0);
1184 }
1185 
1186 static int
1187 mpr_alloc_requests(struct mpr_softc *sc)
1188 {
1189 	struct mpr_command *cm;
1190 	struct mpr_chain *chain;
1191 	int i, rsize, nsegs;
1192 
1193 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
1194         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1195 				16, 0,			/* algnmnt, boundary */
1196 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1197 				BUS_SPACE_MAXADDR,	/* highaddr */
1198 				NULL, NULL,		/* filter, filterarg */
1199                                 rsize,			/* maxsize */
1200                                 1,			/* nsegments */
1201                                 rsize,			/* maxsegsize */
1202                                 0,			/* flags */
1203                                 NULL, NULL,		/* lockfunc, lockarg */
1204                                 &sc->req_dmat)) {
1205 		device_printf(sc->mpr_dev, "Cannot allocate request DMA tag\n");
1206 		return (ENOMEM);
1207         }
1208         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1209 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1210 		device_printf(sc->mpr_dev, "Cannot allocate request memory\n");
1211 		return (ENOMEM);
1212         }
1213         bzero(sc->req_frames, rsize);
1214         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1215 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1216 
1217 	/*
1218 	 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to
1219 	 * get the size of a Chain Frame.  Previous versions use the size as a
1220 	 * Request Frame for the Chain Frame size.  If IOCMaxChainSegmentSize
1221 	 * is 0, use the default value.  The IOCMaxChainSegmentSize is the
1222 	 * number of 16-byte elelements that can fit in a Chain Frame, which is
1223 	 * the size of an IEEE Simple SGE.
1224 	 */
1225 	if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) {
1226 		sc->chain_seg_size =
1227 		    htole16(sc->facts->IOCMaxChainSegmentSize);
1228 		if (sc->chain_seg_size == 0) {
1229 			sc->chain_frame_size = MPR_DEFAULT_CHAIN_SEG_SIZE *
1230 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
1231 		} else {
1232 			sc->chain_frame_size = sc->chain_seg_size *
1233 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
1234 		}
1235 	} else {
1236 		sc->chain_frame_size = sc->facts->IOCRequestFrameSize * 4;
1237 	}
1238 	rsize = sc->chain_frame_size * sc->max_chains;
1239         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1240 				16, 0,			/* algnmnt, boundary */
1241 				BUS_SPACE_MAXADDR,	/* lowaddr */
1242 				BUS_SPACE_MAXADDR,	/* highaddr */
1243 				NULL, NULL,		/* filter, filterarg */
1244                                 rsize,			/* maxsize */
1245                                 1,			/* nsegments */
1246                                 rsize,			/* maxsegsize */
1247                                 0,			/* flags */
1248                                 NULL, NULL,		/* lockfunc, lockarg */
1249                                 &sc->chain_dmat)) {
1250 		device_printf(sc->mpr_dev, "Cannot allocate chain DMA tag\n");
1251 		return (ENOMEM);
1252         }
1253         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1254 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1255 		device_printf(sc->mpr_dev, "Cannot allocate chain memory\n");
1256 		return (ENOMEM);
1257         }
1258         bzero(sc->chain_frames, rsize);
1259         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1260 	    mpr_memaddr_cb, &sc->chain_busaddr, 0);
1261 
1262 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1263 	if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1264 				1, 0,			/* algnmnt, boundary */
1265 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1266 				BUS_SPACE_MAXADDR,	/* highaddr */
1267 				NULL, NULL,		/* filter, filterarg */
1268                                 rsize,			/* maxsize */
1269                                 1,			/* nsegments */
1270                                 rsize,			/* maxsegsize */
1271                                 0,			/* flags */
1272                                 NULL, NULL,		/* lockfunc, lockarg */
1273                                 &sc->sense_dmat)) {
1274 		device_printf(sc->mpr_dev, "Cannot allocate sense DMA tag\n");
1275 		return (ENOMEM);
1276         }
1277         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1278 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1279 		device_printf(sc->mpr_dev, "Cannot allocate sense memory\n");
1280 		return (ENOMEM);
1281         }
1282         bzero(sc->sense_frames, rsize);
1283         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1284 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1285 
1286 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->max_chains, M_MPR,
1287 	    M_WAITOK | M_ZERO);
1288 	if (!sc->chains) {
1289 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
1290 		    __func__, __LINE__);
1291 		return (ENOMEM);
1292 	}
1293 	for (i = 0; i < sc->max_chains; i++) {
1294 		chain = &sc->chains[i];
1295 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1296 		    i * sc->chain_frame_size);
1297 		chain->chain_busaddr = sc->chain_busaddr +
1298 		    i * sc->chain_frame_size;
1299 		mpr_free_chain(sc, chain);
1300 		sc->chain_free_lowwater++;
1301 	}
1302 
1303 	/*
1304 	 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports
1305 	 * these devices.
1306 	 */
1307 	if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) &&
1308 	    (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) {
1309 		if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM)
1310 			return (ENOMEM);
1311 	}
1312 
1313 	/* XXX Need to pick a more precise value */
1314 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
1315         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1316 				1, 0,			/* algnmnt, boundary */
1317 				BUS_SPACE_MAXADDR,	/* lowaddr */
1318 				BUS_SPACE_MAXADDR,	/* highaddr */
1319 				NULL, NULL,		/* filter, filterarg */
1320                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1321                                 nsegs,			/* nsegments */
1322                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
1323                                 BUS_DMA_ALLOCNOW,	/* flags */
1324                                 busdma_lock_mutex,	/* lockfunc */
1325 				&sc->mpr_mtx,		/* lockarg */
1326                                 &sc->buffer_dmat)) {
1327 		device_printf(sc->mpr_dev, "Cannot allocate buffer DMA tag\n");
1328 		return (ENOMEM);
1329         }
1330 
1331 	/*
1332 	 * SMID 0 cannot be used as a free command per the firmware spec.
1333 	 * Just drop that command instead of risking accounting bugs.
1334 	 */
1335 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1336 	    M_MPR, M_WAITOK | M_ZERO);
1337 	if (!sc->commands) {
1338 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
1339 		    __func__, __LINE__);
1340 		return (ENOMEM);
1341 	}
1342 	for (i = 1; i < sc->num_reqs; i++) {
1343 		cm = &sc->commands[i];
1344 		cm->cm_req = sc->req_frames +
1345 		    i * sc->facts->IOCRequestFrameSize * 4;
1346 		cm->cm_req_busaddr = sc->req_busaddr +
1347 		    i * sc->facts->IOCRequestFrameSize * 4;
1348 		cm->cm_sense = &sc->sense_frames[i];
1349 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1350 		cm->cm_desc.Default.SMID = i;
1351 		cm->cm_sc = sc;
1352 		TAILQ_INIT(&cm->cm_chain_list);
1353 		TAILQ_INIT(&cm->cm_prp_page_list);
1354 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1355 
1356 		/* XXX Is a failure here a critical problem? */
1357 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap)
1358 		    == 0) {
1359 			if (i <= sc->facts->HighPriorityCredit)
1360 				mpr_free_high_priority_command(sc, cm);
1361 			else
1362 				mpr_free_command(sc, cm);
1363 		} else {
1364 			panic("failed to allocate command %d\n", i);
1365 			sc->num_reqs = i;
1366 			break;
1367 		}
1368 	}
1369 
1370 	return (0);
1371 }
1372 
1373 /*
1374  * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs,
1375  * which are scatter/gather lists for NVMe devices.
1376  *
1377  * This buffer must be contiguous due to the nature of how NVMe PRPs are built
1378  * and translated by FW.
1379  *
1380  * returns ENOMEM if memory could not be allocated, otherwise returns 0.
1381  */
1382 static int
1383 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc)
1384 {
1385 	int PRPs_per_page, PRPs_required, pages_required;
1386 	int rsize, i;
1387 	struct mpr_prp_page *prp_page;
1388 
1389 	/*
1390 	 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number
1391 	 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is:
1392 	 * MAX_IO_SIZE / PAGE_SIZE = 256
1393 	 *
1394 	 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs
1395 	 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one
1396 	 * page (4096 / 8 = 512), so only one page is required for each I/O.
1397 	 *
1398 	 * Each of these buffers will need to be contiguous. For simplicity,
1399 	 * only one buffer is allocated here, which has all of the space
1400 	 * required for the NVMe Queue Depth. If there are problems allocating
1401 	 * this one buffer, this function will need to change to allocate
1402 	 * individual, contiguous NVME_QDEPTH buffers.
1403 	 *
1404 	 * The real calculation will use the real max io size. Above is just an
1405 	 * example.
1406 	 *
1407 	 */
1408 	PRPs_required = sc->maxio / PAGE_SIZE;
1409 	PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1;
1410 	pages_required = (PRPs_required / PRPs_per_page) + 1;
1411 
1412 	sc->prp_buffer_size = PAGE_SIZE * pages_required;
1413 	rsize = sc->prp_buffer_size * NVME_QDEPTH;
1414 	if (bus_dma_tag_create( sc->mpr_parent_dmat,	/* parent */
1415 				4, 0,			/* algnmnt, boundary */
1416 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1417 				BUS_SPACE_MAXADDR,	/* highaddr */
1418 				NULL, NULL,		/* filter, filterarg */
1419 				rsize,			/* maxsize */
1420 				1,			/* nsegments */
1421 				rsize,			/* maxsegsize */
1422 				0,			/* flags */
1423 				NULL, NULL,		/* lockfunc, lockarg */
1424 				&sc->prp_page_dmat)) {
1425 		device_printf(sc->mpr_dev, "Cannot allocate NVMe PRP DMA "
1426 		    "tag\n");
1427 		return (ENOMEM);
1428 	}
1429 	if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages,
1430 	    BUS_DMA_NOWAIT, &sc->prp_page_map)) {
1431 		device_printf(sc->mpr_dev, "Cannot allocate NVMe PRP memory\n");
1432 		return (ENOMEM);
1433 	}
1434 	bzero(sc->prp_pages, rsize);
1435 	bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages,
1436 	    rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0);
1437 
1438 	sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR,
1439 	    M_WAITOK | M_ZERO);
1440 	for (i = 0; i < NVME_QDEPTH; i++) {
1441 		prp_page = &sc->prps[i];
1442 		prp_page->prp_page = (uint64_t *)(sc->prp_pages +
1443 		    i * sc->prp_buffer_size);
1444 		prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr +
1445 		    i * sc->prp_buffer_size);
1446 		mpr_free_prp_page(sc, prp_page);
1447 		sc->prp_pages_free_lowwater++;
1448 	}
1449 
1450 	return (0);
1451 }
1452 
1453 static int
1454 mpr_init_queues(struct mpr_softc *sc)
1455 {
1456 	int i;
1457 
1458 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1459 
1460 	/*
1461 	 * According to the spec, we need to use one less reply than we
1462 	 * have space for on the queue.  So sc->num_replies (the number we
1463 	 * use) should be less than sc->fqdepth (allocated size).
1464 	 */
1465 	if (sc->num_replies >= sc->fqdepth)
1466 		return (EINVAL);
1467 
1468 	/*
1469 	 * Initialize all of the free queue entries.
1470 	 */
1471 	for (i = 0; i < sc->fqdepth; i++) {
1472 		sc->free_queue[i] = sc->reply_busaddr +
1473 		    (i * sc->facts->ReplyFrameSize * 4);
1474 	}
1475 	sc->replyfreeindex = sc->num_replies;
1476 
1477 	return (0);
1478 }
1479 
1480 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1481  * Next are the global settings, if they exist.  Highest are the per-unit
1482  * settings, if they exist.
1483  */
1484 static void
1485 mpr_get_tunables(struct mpr_softc *sc)
1486 {
1487 	char tmpstr[80];
1488 
1489 	/* XXX default to some debugging for now */
1490 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1491 	sc->disable_msix = 0;
1492 	sc->disable_msi = 0;
1493 	sc->max_chains = MPR_CHAIN_FRAMES;
1494 	sc->max_io_pages = MPR_MAXIO_PAGES;
1495 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1496 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1497 	sc->use_phynum = 1;
1498 
1499 	/*
1500 	 * Grab the global variables.
1501 	 */
1502 	TUNABLE_INT_FETCH("hw.mpr.debug_level", &sc->mpr_debug);
1503 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1504 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1505 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1506 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1507 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1508 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1509 	TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum);
1510 
1511 	/* Grab the unit-instance variables */
1512 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1513 	    device_get_unit(sc->mpr_dev));
1514 	TUNABLE_INT_FETCH(tmpstr, &sc->mpr_debug);
1515 
1516 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1517 	    device_get_unit(sc->mpr_dev));
1518 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1519 
1520 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1521 	    device_get_unit(sc->mpr_dev));
1522 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1523 
1524 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1525 	    device_get_unit(sc->mpr_dev));
1526 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1527 
1528 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1529 	    device_get_unit(sc->mpr_dev));
1530 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1531 
1532 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1533 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1534 	    device_get_unit(sc->mpr_dev));
1535 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1536 
1537 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1538 	    device_get_unit(sc->mpr_dev));
1539 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1540 
1541 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1542 	    device_get_unit(sc->mpr_dev));
1543 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1544 
1545 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num",
1546 	    device_get_unit(sc->mpr_dev));
1547 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1548 }
1549 
1550 static void
1551 mpr_setup_sysctl(struct mpr_softc *sc)
1552 {
1553 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1554 	struct sysctl_oid	*sysctl_tree = NULL;
1555 	char tmpstr[80], tmpstr2[80];
1556 
1557 	/*
1558 	 * Setup the sysctl variable so the user can change the debug level
1559 	 * on the fly.
1560 	 */
1561 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1562 	    device_get_unit(sc->mpr_dev));
1563 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1564 
1565 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1566 	if (sysctl_ctx != NULL)
1567 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1568 
1569 	if (sysctl_tree == NULL) {
1570 		sysctl_ctx_init(&sc->sysctl_ctx);
1571 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1572 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1573 		    CTLFLAG_RD, 0, tmpstr);
1574 		if (sc->sysctl_tree == NULL)
1575 			return;
1576 		sysctl_ctx = &sc->sysctl_ctx;
1577 		sysctl_tree = sc->sysctl_tree;
1578 	}
1579 
1580 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1581 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mpr_debug, 0,
1582 	    "mpr debug level");
1583 
1584 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1585 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1586 	    "Disable the use of MSI-X interrupts");
1587 
1588 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1589 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1590 	    "Disable the use of MSI interrupts");
1591 
1592 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1593 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1594 	    strlen(sc->fw_version), "firmware version");
1595 
1596 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1597 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPR_DRIVER_VERSION,
1598 	    strlen(MPR_DRIVER_VERSION), "driver version");
1599 
1600 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1601 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1602 	    &sc->io_cmds_active, 0, "number of currently active commands");
1603 
1604 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1605 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1606 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1607 
1608 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1609 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1610 	    &sc->chain_free, 0, "number of free chain elements");
1611 
1612 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1613 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1614 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1615 
1616 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1617 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1618 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1619 
1620 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1621 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1622 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1623 	    "IOCFacts)");
1624 
1625 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1626 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1627 	    "enable SSU to SATA SSD/HDD at shutdown");
1628 
1629 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1630 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1631 	    &sc->chain_alloc_fail, "chain allocation failures");
1632 
1633 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1634 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1635 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1636 	    "spinup after SATA ID error");
1637 
1638 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1639 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1640 	    "Use the phy number for enumeration");
1641 
1642 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1643 	    OID_AUTO, "prp_pages_free", CTLFLAG_RD,
1644 	    &sc->prp_pages_free, 0, "number of free PRP pages");
1645 
1646 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1647 	    OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD,
1648 	    &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages");
1649 
1650 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1651 	    OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD,
1652 	    &sc->prp_page_alloc_fail, "PRP page allocation failures");
1653 }
1654 
1655 int
1656 mpr_attach(struct mpr_softc *sc)
1657 {
1658 	int error;
1659 
1660 	mpr_get_tunables(sc);
1661 
1662 	MPR_FUNCTRACE(sc);
1663 
1664 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
1665 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
1666 	TAILQ_INIT(&sc->event_list);
1667 	timevalclear(&sc->lastfail);
1668 
1669 	if ((error = mpr_transition_ready(sc)) != 0) {
1670 		mpr_printf(sc, "%s failed to transition ready\n", __func__);
1671 		return (error);
1672 	}
1673 
1674 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
1675 	    M_ZERO|M_NOWAIT);
1676 	if (!sc->facts) {
1677 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
1678 		    __func__, __LINE__);
1679 		return (ENOMEM);
1680 	}
1681 
1682 	/*
1683 	 * Get IOC Facts and allocate all structures based on this information.
1684 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
1685 	 * Facts. If relevant values have changed in IOC Facts, this function
1686 	 * will free all of the memory based on IOC Facts and reallocate that
1687 	 * memory.  If this fails, any allocated memory should already be freed.
1688 	 */
1689 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
1690 		mpr_dprint(sc, MPR_FAULT, "%s IOC Facts based allocation "
1691 		    "failed with error %d\n", __func__, error);
1692 		return (error);
1693 	}
1694 
1695 	/* Start the periodic watchdog check on the IOC Doorbell */
1696 	mpr_periodic(sc);
1697 
1698 	/*
1699 	 * The portenable will kick off discovery events that will drive the
1700 	 * rest of the initialization process.  The CAM/SAS module will
1701 	 * hold up the boot sequence until discovery is complete.
1702 	 */
1703 	sc->mpr_ich.ich_func = mpr_startup;
1704 	sc->mpr_ich.ich_arg = sc;
1705 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
1706 		mpr_dprint(sc, MPR_ERROR, "Cannot establish MPR config hook\n");
1707 		error = EINVAL;
1708 	}
1709 
1710 	/*
1711 	 * Allow IR to shutdown gracefully when shutdown occurs.
1712 	 */
1713 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1714 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1715 
1716 	if (sc->shutdown_eh == NULL)
1717 		mpr_dprint(sc, MPR_ERROR, "shutdown event registration "
1718 		    "failed\n");
1719 
1720 	mpr_setup_sysctl(sc);
1721 
1722 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
1723 
1724 	return (error);
1725 }
1726 
1727 /* Run through any late-start handlers. */
1728 static void
1729 mpr_startup(void *arg)
1730 {
1731 	struct mpr_softc *sc;
1732 
1733 	sc = (struct mpr_softc *)arg;
1734 
1735 	mpr_lock(sc);
1736 	mpr_unmask_intr(sc);
1737 
1738 	/* initialize device mapping tables */
1739 	mpr_base_static_config_pages(sc);
1740 	mpr_mapping_initialize(sc);
1741 	mprsas_startup(sc);
1742 	mpr_unlock(sc);
1743 }
1744 
1745 /* Periodic watchdog.  Is called with the driver lock already held. */
1746 static void
1747 mpr_periodic(void *arg)
1748 {
1749 	struct mpr_softc *sc;
1750 	uint32_t db;
1751 
1752 	sc = (struct mpr_softc *)arg;
1753 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
1754 		return;
1755 
1756 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
1757 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1758 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
1759 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
1760 			panic("TEMPERATURE FAULT: STOPPING.");
1761 		}
1762 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
1763 		mpr_reinit(sc);
1764 	}
1765 
1766 	callout_reset(&sc->periodic, MPR_PERIODIC_DELAY * hz, mpr_periodic, sc);
1767 }
1768 
1769 static void
1770 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
1771     MPI2_EVENT_NOTIFICATION_REPLY *event)
1772 {
1773 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1774 
1775 	mpr_print_event(sc, event);
1776 
1777 	switch (event->Event) {
1778 	case MPI2_EVENT_LOG_DATA:
1779 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
1780 		if (sc->mpr_debug & MPR_EVENT)
1781 			hexdump(event->EventData, event->EventDataLength, NULL,
1782 			    0);
1783 		break;
1784 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1785 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1786 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1787 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1788 		     entry->LogSequence);
1789 		break;
1790 	default:
1791 		break;
1792 	}
1793 	return;
1794 }
1795 
1796 static int
1797 mpr_attach_log(struct mpr_softc *sc)
1798 {
1799 	uint8_t events[16];
1800 
1801 	bzero(events, 16);
1802 	setbit(events, MPI2_EVENT_LOG_DATA);
1803 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1804 
1805 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
1806 	    &sc->mpr_log_eh);
1807 
1808 	return (0);
1809 }
1810 
1811 static int
1812 mpr_detach_log(struct mpr_softc *sc)
1813 {
1814 
1815 	if (sc->mpr_log_eh != NULL)
1816 		mpr_deregister_events(sc, sc->mpr_log_eh);
1817 	return (0);
1818 }
1819 
1820 /*
1821  * Free all of the driver resources and detach submodules.  Should be called
1822  * without the lock held.
1823  */
1824 int
1825 mpr_free(struct mpr_softc *sc)
1826 {
1827 	int error;
1828 
1829 	/* Turn off the watchdog */
1830 	mpr_lock(sc);
1831 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
1832 	mpr_unlock(sc);
1833 	/* Lock must not be held for this */
1834 	callout_drain(&sc->periodic);
1835 
1836 	if (((error = mpr_detach_log(sc)) != 0) ||
1837 	    ((error = mpr_detach_sas(sc)) != 0))
1838 		return (error);
1839 
1840 	mpr_detach_user(sc);
1841 
1842 	/* Put the IOC back in the READY state. */
1843 	mpr_lock(sc);
1844 	if ((error = mpr_transition_ready(sc)) != 0) {
1845 		mpr_unlock(sc);
1846 		return (error);
1847 	}
1848 	mpr_unlock(sc);
1849 
1850 	if (sc->facts != NULL)
1851 		free(sc->facts, M_MPR);
1852 
1853 	/*
1854 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
1855 	 * to free these buffers too.
1856 	 */
1857 	mpr_iocfacts_free(sc);
1858 
1859 	if (sc->sysctl_tree != NULL)
1860 		sysctl_ctx_free(&sc->sysctl_ctx);
1861 
1862 	/* Deregister the shutdown function */
1863 	if (sc->shutdown_eh != NULL)
1864 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1865 
1866 	mtx_destroy(&sc->mpr_mtx);
1867 
1868 	return (0);
1869 }
1870 
1871 static __inline void
1872 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
1873 {
1874 	MPR_FUNCTRACE(sc);
1875 
1876 	if (cm == NULL) {
1877 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
1878 		return;
1879 	}
1880 
1881 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
1882 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
1883 
1884 	if (cm->cm_complete != NULL) {
1885 		mpr_dprint(sc, MPR_TRACE,
1886 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
1887 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
1888 		    cm->cm_reply);
1889 		cm->cm_complete(sc, cm);
1890 	}
1891 
1892 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
1893 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
1894 		wakeup(cm);
1895 	}
1896 
1897 	if (sc->io_cmds_active != 0) {
1898 		sc->io_cmds_active--;
1899 	} else {
1900 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
1901 		    "out of sync - resynching to 0\n");
1902 	}
1903 }
1904 
1905 static void
1906 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
1907 {
1908 	union loginfo_type {
1909 		u32	loginfo;
1910 		struct {
1911 			u32	subcode:16;
1912 			u32	code:8;
1913 			u32	originator:4;
1914 			u32	bus_type:4;
1915 		} dw;
1916 	};
1917 	union loginfo_type sas_loginfo;
1918 	char *originator_str = NULL;
1919 
1920 	sas_loginfo.loginfo = log_info;
1921 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1922 		return;
1923 
1924 	/* each nexus loss loginfo */
1925 	if (log_info == 0x31170000)
1926 		return;
1927 
1928 	/* eat the loginfos associated with task aborts */
1929 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
1930 	    (log_info == 0x31130000))
1931 		return;
1932 
1933 	switch (sas_loginfo.dw.originator) {
1934 	case 0:
1935 		originator_str = "IOP";
1936 		break;
1937 	case 1:
1938 		originator_str = "PL";
1939 		break;
1940 	case 2:
1941 		originator_str = "IR";
1942 		break;
1943 	}
1944 
1945 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
1946 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
1947 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
1948 }
1949 
1950 static void
1951 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
1952 {
1953 	MPI2DefaultReply_t *mpi_reply;
1954 	u16 sc_status;
1955 
1956 	mpi_reply = (MPI2DefaultReply_t*)reply;
1957 	sc_status = le16toh(mpi_reply->IOCStatus);
1958 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1959 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1960 }
1961 
1962 void
1963 mpr_intr(void *data)
1964 {
1965 	struct mpr_softc *sc;
1966 	uint32_t status;
1967 
1968 	sc = (struct mpr_softc *)data;
1969 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
1970 
1971 	/*
1972 	 * Check interrupt status register to flush the bus.  This is
1973 	 * needed for both INTx interrupts and driver-driven polling
1974 	 */
1975 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1976 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1977 		return;
1978 
1979 	mpr_lock(sc);
1980 	mpr_intr_locked(data);
1981 	mpr_unlock(sc);
1982 	return;
1983 }
1984 
1985 /*
1986  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1987  * chip.  Hopefully this theory is correct.
1988  */
1989 void
1990 mpr_intr_msi(void *data)
1991 {
1992 	struct mpr_softc *sc;
1993 
1994 	sc = (struct mpr_softc *)data;
1995 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
1996 	mpr_lock(sc);
1997 	mpr_intr_locked(data);
1998 	mpr_unlock(sc);
1999 	return;
2000 }
2001 
2002 /*
2003  * The locking is overly broad and simplistic, but easy to deal with for now.
2004  */
2005 void
2006 mpr_intr_locked(void *data)
2007 {
2008 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2009 	struct mpr_softc *sc;
2010 	struct mpr_command *cm = NULL;
2011 	uint8_t flags;
2012 	u_int pq;
2013 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2014 	mpr_fw_diagnostic_buffer_t *pBuffer;
2015 
2016 	sc = (struct mpr_softc *)data;
2017 
2018 	pq = sc->replypostindex;
2019 	mpr_dprint(sc, MPR_TRACE,
2020 	    "%s sc %p starting with replypostindex %u\n",
2021 	    __func__, sc, sc->replypostindex);
2022 
2023 	for ( ;; ) {
2024 		cm = NULL;
2025 		desc = &sc->post_queue[sc->replypostindex];
2026 		flags = desc->Default.ReplyFlags &
2027 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2028 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
2029 		    (le32toh(desc->Words.High) == 0xffffffff))
2030 			break;
2031 
2032 		/* increment the replypostindex now, so that event handlers
2033 		 * and cm completion handlers which decide to do a diag
2034 		 * reset can zero it without it getting incremented again
2035 		 * afterwards, and we break out of this loop on the next
2036 		 * iteration since the reply post queue has been cleared to
2037 		 * 0xFF and all descriptors look unused (which they are).
2038 		 */
2039 		if (++sc->replypostindex >= sc->pqdepth)
2040 			sc->replypostindex = 0;
2041 
2042 		switch (flags) {
2043 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2044 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
2045 		case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS:
2046 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2047 			cm->cm_reply = NULL;
2048 			break;
2049 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2050 		{
2051 			uint32_t baddr;
2052 			uint8_t *reply;
2053 
2054 			/*
2055 			 * Re-compose the reply address from the address
2056 			 * sent back from the chip.  The ReplyFrameAddress
2057 			 * is the lower 32 bits of the physical address of
2058 			 * particular reply frame.  Convert that address to
2059 			 * host format, and then use that to provide the
2060 			 * offset against the virtual address base
2061 			 * (sc->reply_frames).
2062 			 */
2063 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2064 			reply = sc->reply_frames +
2065 				(baddr - ((uint32_t)sc->reply_busaddr));
2066 			/*
2067 			 * Make sure the reply we got back is in a valid
2068 			 * range.  If not, go ahead and panic here, since
2069 			 * we'll probably panic as soon as we deference the
2070 			 * reply pointer anyway.
2071 			 */
2072 			if ((reply < sc->reply_frames)
2073 			 || (reply > (sc->reply_frames +
2074 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
2075 				printf("%s: WARNING: reply %p out of range!\n",
2076 				       __func__, reply);
2077 				printf("%s: reply_frames %p, fqdepth %d, "
2078 				       "frame size %d\n", __func__,
2079 				       sc->reply_frames, sc->fqdepth,
2080 				       sc->facts->ReplyFrameSize * 4);
2081 				printf("%s: baddr %#x,\n", __func__, baddr);
2082 				/* LSI-TODO. See Linux Code for Graceful exit */
2083 				panic("Reply address out of range");
2084 			}
2085 			if (le16toh(desc->AddressReply.SMID) == 0) {
2086 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2087 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2088 					/*
2089 					 * If SMID is 0 for Diag Buffer Post,
2090 					 * this implies that the reply is due to
2091 					 * a release function with a status that
2092 					 * the buffer has been released.  Set
2093 					 * the buffer flags accordingly.
2094 					 */
2095 					rel_rep =
2096 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2097 					if ((le16toh(rel_rep->IOCStatus) &
2098 					    MPI2_IOCSTATUS_MASK) ==
2099 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2100 					{
2101 						pBuffer =
2102 						    &sc->fw_diag_buffer_list[
2103 						    rel_rep->BufferType];
2104 						pBuffer->valid_data = TRUE;
2105 						pBuffer->owned_by_firmware =
2106 						    FALSE;
2107 						pBuffer->immediate = FALSE;
2108 					}
2109 				} else
2110 					mpr_dispatch_event(sc, baddr,
2111 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2112 					    reply);
2113 			} else {
2114 				cm = &sc->commands[
2115 				    le16toh(desc->AddressReply.SMID)];
2116 				cm->cm_reply = reply;
2117 				cm->cm_reply_data =
2118 				    le32toh(desc->AddressReply.
2119 				    ReplyFrameAddress);
2120 			}
2121 			break;
2122 		}
2123 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2124 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2125 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2126 		default:
2127 			/* Unhandled */
2128 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
2129 			    desc->Default.ReplyFlags);
2130 			cm = NULL;
2131 			break;
2132 		}
2133 
2134 		if (cm != NULL) {
2135 			// Print Error reply frame
2136 			if (cm->cm_reply)
2137 				mpr_display_reply_info(sc,cm->cm_reply);
2138 			mpr_complete_command(sc, cm);
2139 		}
2140 
2141 		desc->Words.Low = 0xffffffff;
2142 		desc->Words.High = 0xffffffff;
2143 	}
2144 
2145 	if (pq != sc->replypostindex) {
2146 		mpr_dprint(sc, MPR_TRACE,
2147 		    "%s sc %p writing postindex %d\n",
2148 		    __func__, sc, sc->replypostindex);
2149 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2150 		    sc->replypostindex);
2151 	}
2152 
2153 	return;
2154 }
2155 
2156 static void
2157 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2158     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2159 {
2160 	struct mpr_event_handle *eh;
2161 	int event, handled = 0;
2162 
2163 	event = le16toh(reply->Event);
2164 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2165 		if (isset(eh->mask, event)) {
2166 			eh->callback(sc, data, reply);
2167 			handled++;
2168 		}
2169 	}
2170 
2171 	if (handled == 0)
2172 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2173 		    le16toh(event));
2174 
2175 	/*
2176 	 * This is the only place that the event/reply should be freed.
2177 	 * Anything wanting to hold onto the event data should have
2178 	 * already copied it into their own storage.
2179 	 */
2180 	mpr_free_reply(sc, data);
2181 }
2182 
2183 static void
2184 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2185 {
2186 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2187 
2188 	if (cm->cm_reply)
2189 		mpr_print_event(sc,
2190 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2191 
2192 	mpr_free_command(sc, cm);
2193 
2194 	/* next, send a port enable */
2195 	mprsas_startup(sc);
2196 }
2197 
2198 /*
2199  * For both register_events and update_events, the caller supplies a bitmap
2200  * of events that it _wants_.  These functions then turn that into a bitmask
2201  * suitable for the controller.
2202  */
2203 int
2204 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2205     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2206 {
2207 	struct mpr_event_handle *eh;
2208 	int error = 0;
2209 
2210 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2211 	if (!eh) {
2212 		device_printf(sc->mpr_dev, "Cannot allocate memory %s %d\n",
2213 		    __func__, __LINE__);
2214 		return (ENOMEM);
2215 	}
2216 	eh->callback = cb;
2217 	eh->data = data;
2218 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2219 	if (mask != NULL)
2220 		error = mpr_update_events(sc, eh, mask);
2221 	*handle = eh;
2222 
2223 	return (error);
2224 }
2225 
2226 int
2227 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2228     uint8_t *mask)
2229 {
2230 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2231 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
2232 	struct mpr_command *cm;
2233 	struct mpr_event_handle *eh;
2234 	int error, i;
2235 
2236 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2237 
2238 	if ((mask != NULL) && (handle != NULL))
2239 		bcopy(mask, &handle->mask[0], 16);
2240 	memset(sc->event_mask, 0xff, 16);
2241 
2242 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2243 		for (i = 0; i < 16; i++)
2244 			sc->event_mask[i] &= ~eh->mask[i];
2245 	}
2246 
2247 	if ((cm = mpr_alloc_command(sc)) == NULL)
2248 		return (EBUSY);
2249 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2250 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2251 	evtreq->MsgFlags = 0;
2252 	evtreq->SASBroadcastPrimitiveMasks = 0;
2253 #ifdef MPR_DEBUG_ALL_EVENTS
2254 	{
2255 		u_char fullmask[16];
2256 		memset(fullmask, 0x00, 16);
2257 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2258 	}
2259 #else
2260 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2261 #endif
2262 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2263 	cm->cm_data = NULL;
2264 
2265 	error = mpr_request_polled(sc, cm);
2266 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2267 	if ((reply == NULL) ||
2268 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2269 		error = ENXIO;
2270 
2271 	if (reply)
2272 		mpr_print_event(sc, reply);
2273 
2274 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2275 
2276 	mpr_free_command(sc, cm);
2277 	return (error);
2278 }
2279 
2280 static int
2281 mpr_reregister_events(struct mpr_softc *sc)
2282 {
2283 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2284 	struct mpr_command *cm;
2285 	struct mpr_event_handle *eh;
2286 	int error, i;
2287 
2288 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2289 
2290 	/* first, reregister events */
2291 
2292 	memset(sc->event_mask, 0xff, 16);
2293 
2294 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2295 		for (i = 0; i < 16; i++)
2296 			sc->event_mask[i] &= ~eh->mask[i];
2297 	}
2298 
2299 	if ((cm = mpr_alloc_command(sc)) == NULL)
2300 		return (EBUSY);
2301 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2302 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2303 	evtreq->MsgFlags = 0;
2304 	evtreq->SASBroadcastPrimitiveMasks = 0;
2305 #ifdef MPR_DEBUG_ALL_EVENTS
2306 	{
2307 		u_char fullmask[16];
2308 		memset(fullmask, 0x00, 16);
2309 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2310 	}
2311 #else
2312 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2313 #endif
2314 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2315 	cm->cm_data = NULL;
2316 	cm->cm_complete = mpr_reregister_events_complete;
2317 
2318 	error = mpr_map_command(sc, cm);
2319 
2320 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2321 	    error);
2322 	return (error);
2323 }
2324 
2325 int
2326 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2327 {
2328 
2329 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2330 	free(handle, M_MPR);
2331 	return (mpr_update_events(sc, NULL, NULL));
2332 }
2333 
2334 /**
2335 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a
2336 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry
2337 * of the NVMe message (PRP1). If the data buffer is small enough to be described
2338 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to
2339 * describe a larger data buffer. If the data buffer is too large to describe
2340 * using the two PRP entriess inside the NVMe message, then PRP1 describes the
2341 * first data memory segment, and PRP2 contains a pointer to a PRP list located
2342 * elsewhere in memory to describe the remaining data memory segments. The PRP
2343 * list will be contiguous.
2344 
2345 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP
2346 * consists of a list of PRP entries to describe a number of noncontigous
2347 * physical memory segments as a single memory buffer, just as a SGL does. Note
2348 * however, that this function is only used by the IOCTL call, so the memory
2349 * given will be guaranteed to be contiguous. There is no need to translate
2350 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous
2351 * space that is one page size each.
2352 *
2353 * Each NVMe message contains two PRP entries. The first (PRP1) either contains
2354 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains
2355 * the second PRP element if the memory being described fits within 2 PRP
2356 * entries, or a PRP list pointer if the PRP spans more than two entries.
2357 *
2358 * A PRP list pointer contains the address of a PRP list, structured as a linear
2359 * array of PRP entries. Each PRP entry in this list describes a segment of
2360 * physical memory.
2361 *
2362 * Each 64-bit PRP entry comprises an address and an offset field. The address
2363 * always points to the beginning of a PAGE_SIZE physical memory page, and the
2364 * offset describes where within that page the memory segment begins. Only the
2365 * first element in a PRP list may contain a non-zero offest, implying that all
2366 * memory segments following the first begin at the start of a PAGE_SIZE page.
2367 *
2368 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory,
2369 * with exceptions for the first and last elements in the list. If the memory
2370 * being described by the list begins at a non-zero offset within the first page,
2371 * then the first PRP element will contain a non-zero offset indicating where the
2372 * region begins within the page. The last memory segment may end before the end
2373 * of the PAGE_SIZE segment, depending upon the overall size of the memory being
2374 * described by the PRP list.
2375 *
2376 * Since PRP entries lack any indication of size, the overall data buffer length
2377 * is used to determine where the end of the data memory buffer is located, and
2378 * how many PRP entries are required to describe it.
2379 *
2380 * Returns nothing.
2381 */
2382 void
2383 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm,
2384     Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data,
2385     uint32_t data_in_sz, uint32_t data_out_sz)
2386 {
2387 	int			prp_size = PRP_ENTRY_SIZE;
2388 	uint64_t		*prp_entry, *prp1_entry, *prp2_entry;
2389 	uint64_t		*prp_entry_phys, *prp_page, *prp_page_phys;
2390 	uint32_t		offset, entry_len, page_mask_result, page_mask;
2391 	bus_addr_t		paddr;
2392 	size_t			length;
2393 	struct mpr_prp_page	*prp_page_info = NULL;
2394 
2395 	/*
2396 	 * Not all commands require a data transfer. If no data, just return
2397 	 * without constructing any PRP.
2398 	 */
2399 	if (!data_in_sz && !data_out_sz)
2400 		return;
2401 
2402 	/*
2403 	 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is
2404 	 * located at a 24 byte offset from the start of the NVMe command. Then
2405 	 * set the current PRP entry pointer to PRP1.
2406 	 */
2407 	prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2408 	    NVME_CMD_PRP1_OFFSET);
2409 	prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2410 	    NVME_CMD_PRP2_OFFSET);
2411 	prp_entry = prp1_entry;
2412 
2413 	/*
2414 	 * For the PRP entries, use the specially allocated buffer of
2415 	 * contiguous memory. PRP Page allocation failures should not happen
2416 	 * because there should be enough PRP page buffers to account for the
2417 	 * possible NVMe QDepth.
2418 	 */
2419 	prp_page_info = mpr_alloc_prp_page(sc);
2420 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2421 	    "used for building a native NVMe SGL.\n", __func__));
2422 	prp_page = (uint64_t *)prp_page_info->prp_page;
2423 	prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2424 
2425 	/*
2426 	 * Insert the allocated PRP page into the command's PRP page list. This
2427 	 * will be freed when the command is freed.
2428 	 */
2429 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2430 
2431 	/*
2432 	 * Check if we are within 1 entry of a page boundary we don't want our
2433 	 * first entry to be a PRP List entry.
2434 	 */
2435 	page_mask = PAGE_SIZE - 1;
2436 	page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) &
2437 	    page_mask;
2438 	if (!page_mask_result)
2439 	{
2440 		/* Bump up to next page boundary. */
2441 		prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size);
2442 		prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys +
2443 		    prp_size);
2444 	}
2445 
2446 	/*
2447 	 * Set PRP physical pointer, which initially points to the current PRP
2448 	 * DMA memory page.
2449 	 */
2450 	prp_entry_phys = prp_page_phys;
2451 
2452 	/* Get physical address and length of the data buffer. */
2453 	paddr = (bus_addr_t)data;
2454 	if (data_in_sz)
2455 		length = data_in_sz;
2456 	else
2457 		length = data_out_sz;
2458 
2459 	/* Loop while the length is not zero. */
2460 	while (length)
2461 	{
2462 		/*
2463 		 * Check if we need to put a list pointer here if we are at page
2464 		 * boundary - prp_size (8 bytes).
2465 		 */
2466 		page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys +
2467 		    prp_size) & page_mask;
2468 		if (!page_mask_result)
2469 		{
2470 			/*
2471 			 * This is the last entry in a PRP List, so we need to
2472 			 * put a PRP list pointer here. What this does is:
2473 			 *   - bump the current memory pointer to the next
2474 			 *     address, which will be the next full page.
2475 			 *   - set the PRP Entry to point to that page. This is
2476 			 *     now the PRP List pointer.
2477 			 *   - bump the PRP Entry pointer the start of the next
2478 			 *     page. Since all of this PRP memory is contiguous,
2479 			 *     no need to get a new page - it's just the next
2480 			 *     address.
2481 			 */
2482 			prp_entry_phys++;
2483 			*prp_entry =
2484 			    htole64((uint64_t)(uintptr_t)prp_entry_phys);
2485 			prp_entry++;
2486 		}
2487 
2488 		/* Need to handle if entry will be part of a page. */
2489 		offset = (uint32_t)paddr & page_mask;
2490 		entry_len = PAGE_SIZE - offset;
2491 
2492 		if (prp_entry == prp1_entry)
2493 		{
2494 			/*
2495 			 * Must fill in the first PRP pointer (PRP1) before
2496 			 * moving on.
2497 			 */
2498 			*prp1_entry = htole64((uint64_t)paddr);
2499 
2500 			/*
2501 			 * Now point to the second PRP entry within the
2502 			 * command (PRP2).
2503 			 */
2504 			prp_entry = prp2_entry;
2505 		}
2506 		else if (prp_entry == prp2_entry)
2507 		{
2508 			/*
2509 			 * Should the PRP2 entry be a PRP List pointer or just a
2510 			 * regular PRP pointer? If there is more than one more
2511 			 * page of data, must use a PRP List pointer.
2512 			 */
2513 			if (length > PAGE_SIZE)
2514 			{
2515 				/*
2516 				 * PRP2 will contain a PRP List pointer because
2517 				 * more PRP's are needed with this command. The
2518 				 * list will start at the beginning of the
2519 				 * contiguous buffer.
2520 				 */
2521 				*prp2_entry =
2522 				    htole64(
2523 				    (uint64_t)(uintptr_t)prp_entry_phys);
2524 
2525 				/*
2526 				 * The next PRP Entry will be the start of the
2527 				 * first PRP List.
2528 				 */
2529 				prp_entry = prp_page;
2530 			}
2531 			else
2532 			{
2533 				/*
2534 				 * After this, the PRP Entries are complete.
2535 				 * This command uses 2 PRP's and no PRP list.
2536 				 */
2537 				*prp2_entry = htole64((uint64_t)paddr);
2538 			}
2539 		}
2540 		else
2541 		{
2542 			/*
2543 			 * Put entry in list and bump the addresses.
2544 			 *
2545 			 * After PRP1 and PRP2 are filled in, this will fill in
2546 			 * all remaining PRP entries in a PRP List, one per each
2547 			 * time through the loop.
2548 			 */
2549 			*prp_entry = htole64((uint64_t)paddr);
2550 			prp_entry++;
2551 			prp_entry_phys++;
2552 		}
2553 
2554 		/*
2555 		 * Bump the phys address of the command's data buffer by the
2556 		 * entry_len.
2557 		 */
2558 		paddr += entry_len;
2559 
2560 		/* Decrement length accounting for last partial page. */
2561 		if (entry_len > length)
2562 			length = 0;
2563 		else
2564 			length -= entry_len;
2565 	}
2566 }
2567 
2568 /*
2569  * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to
2570  * determine if the driver needs to build a native SGL. If so, that native SGL
2571  * is built in the contiguous buffers allocated especially for PCIe SGL
2572  * creation. If the driver will not build a native SGL, return TRUE and a
2573  * normal IEEE SGL will be built. Currently this routine supports NVMe devices
2574  * only.
2575  *
2576  * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built.
2577  */
2578 static int
2579 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm,
2580     bus_dma_segment_t *segs, int segs_left)
2581 {
2582 	uint32_t		i, sge_dwords, length, offset, entry_len;
2583 	uint32_t		num_entries, buff_len = 0, sges_in_segment;
2584 	uint32_t		page_mask, page_mask_result, *curr_buff;
2585 	uint32_t		*ptr_sgl, *ptr_first_sgl, first_page_offset;
2586 	uint32_t		first_page_data_size, end_residual;
2587 	uint64_t		*msg_phys;
2588 	bus_addr_t		paddr;
2589 	int			build_native_sgl = 0, first_prp_entry;
2590 	int			prp_size = PRP_ENTRY_SIZE;
2591 	Mpi25IeeeSgeChain64_t	*main_chain_element = NULL;
2592 	struct mpr_prp_page	*prp_page_info = NULL;
2593 
2594 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2595 
2596 	/*
2597 	 * Add up the sizes of each segment length to get the total transfer
2598 	 * size, which will be checked against the Maximum Data Transfer Size.
2599 	 * If the data transfer length exceeds the MDTS for this device, just
2600 	 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O
2601 	 * up into multiple I/O's. [nvme_mdts = 0 means unlimited]
2602 	 */
2603 	for (i = 0; i < segs_left; i++)
2604 		buff_len += htole32(segs[i].ds_len);
2605 	if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS))
2606 		return 1;
2607 
2608 	/* Create page_mask (to get offset within page) */
2609 	page_mask = PAGE_SIZE - 1;
2610 
2611 	/*
2612 	 * Check if the number of elements exceeds the max number that can be
2613 	 * put in the main message frame (H/W can only translate an SGL that
2614 	 * is contained entirely in the main message frame).
2615 	 */
2616 	sges_in_segment = (sc->facts->IOCRequestFrameSize -
2617 	    offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION);
2618 	if (segs_left > sges_in_segment)
2619 		build_native_sgl = 1;
2620 	else
2621 	{
2622 		/*
2623 		 * NVMe uses one PRP for each physical page (or part of physical
2624 		 * page).
2625 		 *    if 4 pages or less then IEEE is OK
2626 		 *    if > 5 pages then we need to build a native SGL
2627 		 *    if > 4 and <= 5 pages, then check the physical address of
2628 		 *      the first SG entry, then if this first size in the page
2629 		 *      is >= the residual beyond 4 pages then use IEEE,
2630 		 *      otherwise use native SGL
2631 		 */
2632 		if (buff_len > (PAGE_SIZE * 5))
2633 			build_native_sgl = 1;
2634 		else if ((buff_len > (PAGE_SIZE * 4)) &&
2635 		    (buff_len <= (PAGE_SIZE * 5)) )
2636 		{
2637 			msg_phys = (uint64_t *)segs[0].ds_addr;
2638 			first_page_offset =
2639 			    ((uint32_t)(uint64_t)(uintptr_t)msg_phys &
2640 			    page_mask);
2641 			first_page_data_size = PAGE_SIZE - first_page_offset;
2642 			end_residual = buff_len % PAGE_SIZE;
2643 
2644 			/*
2645 			 * If offset into first page pushes the end of the data
2646 			 * beyond end of the 5th page, we need the extra PRP
2647 			 * list.
2648 			 */
2649 			if (first_page_data_size < end_residual)
2650 				build_native_sgl = 1;
2651 
2652 			/*
2653 			 * Check if first SG entry size is < residual beyond 4
2654 			 * pages.
2655 			 */
2656 			if (htole32(segs[0].ds_len) <
2657 			    (buff_len - (PAGE_SIZE * 4)))
2658 				build_native_sgl = 1;
2659 		}
2660 	}
2661 
2662 	/* check if native SGL is needed */
2663 	if (!build_native_sgl)
2664 		return 1;
2665 
2666 	/*
2667 	 * Native SGL is needed.
2668 	 * Put a chain element in main message frame that points to the first
2669 	 * chain buffer.
2670 	 *
2671 	 * NOTE:  The ChainOffset field must be 0 when using a chain pointer to
2672 	 *        a native SGL.
2673 	 */
2674 
2675 	/* Set main message chain element pointer */
2676 	main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge;
2677 
2678 	/*
2679 	 * For NVMe the chain element needs to be the 2nd SGL entry in the main
2680 	 * message.
2681 	 */
2682 	main_chain_element = (Mpi25IeeeSgeChain64_t *)
2683 	    ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64));
2684 
2685 	/*
2686 	 * For the PRP entries, use the specially allocated buffer of
2687 	 * contiguous memory. PRP Page allocation failures should not happen
2688 	 * because there should be enough PRP page buffers to account for the
2689 	 * possible NVMe QDepth.
2690 	 */
2691 	prp_page_info = mpr_alloc_prp_page(sc);
2692 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2693 	    "used for building a native NVMe SGL.\n", __func__));
2694 	curr_buff = (uint32_t *)prp_page_info->prp_page;
2695 	msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2696 
2697 	/*
2698 	 * Insert the allocated PRP page into the command's PRP page list. This
2699 	 * will be freed when the command is freed.
2700 	 */
2701 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2702 
2703 	/*
2704 	 * Check if we are within 1 entry of a page boundary we don't want our
2705 	 * first entry to be a PRP List entry.
2706 	 */
2707 	page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) &
2708 	    page_mask;
2709 	if (!page_mask_result) {
2710 		/* Bump up to next page boundary. */
2711 		curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size);
2712 		msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size);
2713 	}
2714 
2715 	/* Fill in the chain element and make it an NVMe segment type. */
2716 	main_chain_element->Address.High =
2717 	    htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32));
2718 	main_chain_element->Address.Low =
2719 	    htole32((uint32_t)(uintptr_t)msg_phys);
2720 	main_chain_element->NextChainOffset = 0;
2721 	main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
2722 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
2723 	    MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP;
2724 
2725 	/* Set SGL pointer to start of contiguous PCIe buffer. */
2726 	ptr_sgl = curr_buff;
2727 	sge_dwords = 2;
2728 	num_entries = 0;
2729 
2730 	/*
2731 	 * NVMe has a very convoluted PRP format. One PRP is required for each
2732 	 * page or partial page. We need to split up OS SG entries if they are
2733 	 * longer than one page or cross a page boundary. We also have to insert
2734 	 * a PRP list pointer entry as the last entry in each physical page of
2735 	 * the PRP list.
2736 	 *
2737 	 * NOTE: The first PRP "entry" is actually placed in the first SGL entry
2738 	 * in the main message in IEEE 64 format. The 2nd entry in the main
2739 	 * message is the chain element, and the rest of the PRP entries are
2740 	 * built in the contiguous PCIe buffer.
2741 	 */
2742 	first_prp_entry = 1;
2743 	ptr_first_sgl = (uint32_t *)cm->cm_sge;
2744 
2745 	for (i = 0; i < segs_left; i++) {
2746 		/* Get physical address and length of this SG entry. */
2747 		paddr = segs[i].ds_addr;
2748 		length = segs[i].ds_len;
2749 
2750 		/*
2751 		 * Check whether a given SGE buffer lies on a non-PAGED
2752 		 * boundary if this is not the first page. If so, this is not
2753 		 * expected so have FW build the SGL.
2754 		 */
2755 		if (i) {
2756 			if ((uint32_t)paddr & page_mask) {
2757 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while "
2758 				    "building NVMe PRPs, low address is 0x%x\n",
2759 				    (uint32_t)paddr);
2760 				return 1;
2761 			}
2762 		}
2763 
2764 		/* Apart from last SGE, if any other SGE boundary is not page
2765 		 * aligned then it means that hole exists. Existence of hole
2766 		 * leads to data corruption. So fallback to IEEE SGEs.
2767 		 */
2768 		if (i != (segs_left - 1)) {
2769 			if (((uint32_t)paddr + length) & page_mask) {
2770 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE "
2771 				    "boundary while building NVMe PRPs, low "
2772 				    "address: 0x%x and length: %u\n",
2773 				    (uint32_t)paddr, length);
2774 				return 1;
2775 			}
2776 		}
2777 
2778 		/* Loop while the length is not zero. */
2779 		while (length) {
2780 			/*
2781 			 * Check if we need to put a list pointer here if we are
2782 			 * at page boundary - prp_size.
2783 			 */
2784 			page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl +
2785 			    prp_size) & page_mask;
2786 			if (!page_mask_result) {
2787 				/*
2788 				 * Need to put a PRP list pointer here.
2789 				 */
2790 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
2791 				    prp_size);
2792 				*ptr_sgl = htole32((uintptr_t)msg_phys);
2793 				*(ptr_sgl+1) = htole32((uint64_t)(uintptr_t)
2794 				    msg_phys >> 32);
2795 				ptr_sgl += sge_dwords;
2796 				num_entries++;
2797 			}
2798 
2799 			/* Need to handle if entry will be part of a page. */
2800 			offset = (uint32_t)paddr & page_mask;
2801 			entry_len = PAGE_SIZE - offset;
2802 			if (first_prp_entry) {
2803 				/*
2804 				 * Put IEEE entry in first SGE in main message.
2805 				 * (Simple element, System addr, not end of
2806 				 * list.)
2807 				 */
2808 				*ptr_first_sgl = htole32((uint32_t)paddr);
2809 				*(ptr_first_sgl + 1) =
2810 				    htole32((uint32_t)((uint64_t)paddr >> 32));
2811 				*(ptr_first_sgl + 2) = htole32(entry_len);
2812 				*(ptr_first_sgl + 3) = 0;
2813 
2814 				/* No longer the first PRP entry. */
2815 				first_prp_entry = 0;
2816 			} else {
2817 				/* Put entry in list. */
2818 				*ptr_sgl = htole32((uint32_t)paddr);
2819 				*(ptr_sgl + 1) =
2820 				    htole32((uint32_t)((uint64_t)paddr >> 32));
2821 
2822 				/* Bump ptr_sgl, msg_phys, and num_entries. */
2823 				ptr_sgl += sge_dwords;
2824 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
2825 				    prp_size);
2826 				num_entries++;
2827 			}
2828 
2829 			/* Bump the phys address by the entry_len. */
2830 			paddr += entry_len;
2831 
2832 			/* Decrement length accounting for last partial page. */
2833 			if (entry_len > length)
2834 				length = 0;
2835 			else
2836 				length -= entry_len;
2837 		}
2838 	}
2839 
2840 	/* Set chain element Length. */
2841 	main_chain_element->Length = htole32(num_entries * prp_size);
2842 
2843 	/* Return 0, indicating we built a native SGL. */
2844 	return 0;
2845 }
2846 
2847 /*
2848  * Add a chain element as the next SGE for the specified command.
2849  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
2850  * only required for IEEE commands.  Therefore there is no code for commands
2851  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
2852  * shouldn't be requesting chains).
2853  */
2854 static int
2855 mpr_add_chain(struct mpr_command *cm, int segsleft)
2856 {
2857 	struct mpr_softc *sc = cm->cm_sc;
2858 	MPI2_REQUEST_HEADER *req;
2859 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
2860 	struct mpr_chain *chain;
2861 	int sgc_size, current_segs, rem_segs, segs_per_frame;
2862 	uint8_t next_chain_offset = 0;
2863 
2864 	/*
2865 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
2866 	 * only IEEE commands should be requesting chains.  Return some error
2867 	 * code other than 0.
2868 	 */
2869 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
2870 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
2871 		    "an MPI SGL.\n");
2872 		return(ENOBUFS);
2873 	}
2874 
2875 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
2876 	if (cm->cm_sglsize < sgc_size)
2877 		panic("MPR: Need SGE Error Code\n");
2878 
2879 	chain = mpr_alloc_chain(cm->cm_sc);
2880 	if (chain == NULL)
2881 		return (ENOBUFS);
2882 
2883 	/*
2884 	 * Note: a double-linked list is used to make it easier to walk for
2885 	 * debugging.
2886 	 */
2887 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
2888 
2889 	/*
2890 	 * Need to know if the number of frames left is more than 1 or not.  If
2891 	 * more than 1 frame is required, NextChainOffset will need to be set,
2892 	 * which will just be the last segment of the frame.
2893 	 */
2894 	rem_segs = 0;
2895 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
2896 		/*
2897 		 * rem_segs is the number of segements remaining after the
2898 		 * segments that will go into the current frame.  Since it is
2899 		 * known that at least one more frame is required, account for
2900 		 * the chain element.  To know if more than one more frame is
2901 		 * required, just check if there will be a remainder after using
2902 		 * the current frame (with this chain) and the next frame.  If
2903 		 * so the NextChainOffset must be the last element of the next
2904 		 * frame.
2905 		 */
2906 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
2907 		rem_segs = segsleft - current_segs;
2908 		segs_per_frame = sc->chain_frame_size / sgc_size;
2909 		if (rem_segs > segs_per_frame) {
2910 			next_chain_offset = segs_per_frame - 1;
2911 		}
2912 	}
2913 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
2914 	ieee_sgc->Length = next_chain_offset ?
2915 	    htole32((uint32_t)sc->chain_frame_size) :
2916 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
2917 	ieee_sgc->NextChainOffset = next_chain_offset;
2918 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
2919 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
2920 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
2921 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
2922 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
2923 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
2924 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
2925 
2926 	cm->cm_sglsize = sc->chain_frame_size;
2927 	return (0);
2928 }
2929 
2930 /*
2931  * Add one scatter-gather element to the scatter-gather list for a command.
2932  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
2933  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
2934  * chain, so don't consider any chain additions.
2935  */
2936 int
2937 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
2938     int segsleft)
2939 {
2940 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2941 	u32 sge_flags;
2942 
2943 	/*
2944 	 * case 1: >=1 more segment, no room for anything (error)
2945 	 * case 2: 1 more segment and enough room for it
2946          */
2947 
2948 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
2949 		mpr_dprint(cm->cm_sc, MPR_ERROR,
2950 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
2951 		    __func__);
2952 		return(ENOBUFS);
2953 	}
2954 
2955 	KASSERT(segsleft == 1,
2956 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
2957 	    segsleft));
2958 
2959 	/*
2960 	 * There is one more segment left to add for the MPI SGL and there is
2961 	 * enough room in the frame to add it.  This is the normal case because
2962 	 * MPI SGL's don't have chains, otherwise something is wrong.
2963 	 *
2964 	 * If this is a bi-directional request, need to account for that
2965 	 * here.  Save the pre-filled sge values.  These will be used
2966 	 * either for the 2nd SGL or for a single direction SGL.  If
2967 	 * cm_out_len is non-zero, this is a bi-directional request, so
2968 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
2969 	 * fill in the IN SGL.  Note that at this time, when filling in
2970 	 * 2 SGL's for a bi-directional request, they both use the same
2971 	 * DMA buffer (same cm command).
2972 	 */
2973 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
2974 	saved_address_low = sge->Address.Low;
2975 	saved_address_high = sge->Address.High;
2976 	if (cm->cm_out_len) {
2977 		sge->FlagsLength = cm->cm_out_len |
2978 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2979 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2980 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
2981 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2982 		    MPI2_SGE_FLAGS_SHIFT);
2983 		cm->cm_sglsize -= len;
2984 		/* Endian Safe code */
2985 		sge_flags = sge->FlagsLength;
2986 		sge->FlagsLength = htole32(sge_flags);
2987 		sge->Address.High = htole32(sge->Address.High);
2988 		sge->Address.Low = htole32(sge->Address.Low);
2989 		bcopy(sge, cm->cm_sge, len);
2990 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2991 	}
2992 	sge->FlagsLength = saved_buf_len |
2993 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2994 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
2995 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
2996 	    MPI2_SGE_FLAGS_END_OF_LIST |
2997 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2998 	    MPI2_SGE_FLAGS_SHIFT);
2999 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
3000 		sge->FlagsLength |=
3001 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
3002 		    MPI2_SGE_FLAGS_SHIFT);
3003 	} else {
3004 		sge->FlagsLength |=
3005 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
3006 		    MPI2_SGE_FLAGS_SHIFT);
3007 	}
3008 	sge->Address.Low = saved_address_low;
3009 	sge->Address.High = saved_address_high;
3010 
3011 	cm->cm_sglsize -= len;
3012 	/* Endian Safe code */
3013 	sge_flags = sge->FlagsLength;
3014 	sge->FlagsLength = htole32(sge_flags);
3015 	sge->Address.High = htole32(sge->Address.High);
3016 	sge->Address.Low = htole32(sge->Address.Low);
3017 	bcopy(sge, cm->cm_sge, len);
3018 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3019 	return (0);
3020 }
3021 
3022 /*
3023  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
3024  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
3025  * remaining size and pointer to the next SGE to fill in, respectively.
3026  */
3027 int
3028 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
3029 {
3030 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
3031 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
3032 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3033 	uint32_t sge_length;
3034 
3035 	/*
3036 	 * case 1: No room for chain or segment (error).
3037 	 * case 2: Two or more segments left but only room for chain.
3038 	 * case 3: Last segment and room for it, so set flags.
3039 	 */
3040 
3041 	/*
3042 	 * There should be room for at least one element, or there is a big
3043 	 * problem.
3044 	 */
3045 	if (cm->cm_sglsize < ieee_sge_size)
3046 		panic("MPR: Need SGE Error Code\n");
3047 
3048 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
3049 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
3050 			return (error);
3051 	}
3052 
3053 	if (segsleft == 1) {
3054 		/*
3055 		 * If this is a bi-directional request, need to account for that
3056 		 * here.  Save the pre-filled sge values.  These will be used
3057 		 * either for the 2nd SGL or for a single direction SGL.  If
3058 		 * cm_out_len is non-zero, this is a bi-directional request, so
3059 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
3060 		 * fill in the IN SGL.  Note that at this time, when filling in
3061 		 * 2 SGL's for a bi-directional request, they both use the same
3062 		 * DMA buffer (same cm command).
3063 		 */
3064 		saved_buf_len = sge->Length;
3065 		saved_address_low = sge->Address.Low;
3066 		saved_address_high = sge->Address.High;
3067 		if (cm->cm_out_len) {
3068 			sge->Length = cm->cm_out_len;
3069 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3070 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3071 			cm->cm_sglsize -= ieee_sge_size;
3072 			/* Endian Safe code */
3073 			sge_length = sge->Length;
3074 			sge->Length = htole32(sge_length);
3075 			sge->Address.High = htole32(sge->Address.High);
3076 			sge->Address.Low = htole32(sge->Address.Low);
3077 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
3078 			cm->cm_sge =
3079 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3080 			    ieee_sge_size);
3081 		}
3082 		sge->Length = saved_buf_len;
3083 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3084 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3085 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
3086 		sge->Address.Low = saved_address_low;
3087 		sge->Address.High = saved_address_high;
3088 	}
3089 
3090 	cm->cm_sglsize -= ieee_sge_size;
3091 	/* Endian Safe code */
3092 	sge_length = sge->Length;
3093 	sge->Length = htole32(sge_length);
3094 	sge->Address.High = htole32(sge->Address.High);
3095 	sge->Address.Low = htole32(sge->Address.Low);
3096 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
3097 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3098 	    ieee_sge_size);
3099 	return (0);
3100 }
3101 
3102 /*
3103  * Add one dma segment to the scatter-gather list for a command.
3104  */
3105 int
3106 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
3107     int segsleft)
3108 {
3109 	MPI2_SGE_SIMPLE64 sge;
3110 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
3111 
3112 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
3113 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3114 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3115 		ieee_sge.Length = len;
3116 		mpr_from_u64(pa, &ieee_sge.Address);
3117 
3118 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
3119 	} else {
3120 		/*
3121 		 * This driver always uses 64-bit address elements for
3122 		 * simplicity.
3123 		 */
3124 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3125 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
3126 		/* Set Endian safe macro in mpr_push_sge */
3127 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
3128 		mpr_from_u64(pa, &sge.Address);
3129 
3130 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
3131 	}
3132 }
3133 
3134 static void
3135 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
3136 {
3137 	struct mpr_softc *sc;
3138 	struct mpr_command *cm;
3139 	u_int i, dir, sflags;
3140 
3141 	cm = (struct mpr_command *)arg;
3142 	sc = cm->cm_sc;
3143 
3144 	/*
3145 	 * In this case, just print out a warning and let the chip tell the
3146 	 * user they did the wrong thing.
3147 	 */
3148 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
3149 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
3150 		    "segments, more than the %d allowed\n", __func__, nsegs,
3151 		    cm->cm_max_segs);
3152 	}
3153 
3154 	/*
3155 	 * Set up DMA direction flags.  Bi-directional requests are also handled
3156 	 * here.  In that case, both direction flags will be set.
3157 	 */
3158 	sflags = 0;
3159 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
3160 		/*
3161 		 * We have to add a special case for SMP passthrough, there
3162 		 * is no easy way to generically handle it.  The first
3163 		 * S/G element is used for the command (therefore the
3164 		 * direction bit needs to be set).  The second one is used
3165 		 * for the reply.  We'll leave it to the caller to make
3166 		 * sure we only have two buffers.
3167 		 */
3168 		/*
3169 		 * Even though the busdma man page says it doesn't make
3170 		 * sense to have both direction flags, it does in this case.
3171 		 * We have one s/g element being accessed in each direction.
3172 		 */
3173 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
3174 
3175 		/*
3176 		 * Set the direction flag on the first buffer in the SMP
3177 		 * passthrough request.  We'll clear it for the second one.
3178 		 */
3179 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
3180 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
3181 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
3182 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
3183 		dir = BUS_DMASYNC_PREWRITE;
3184 	} else
3185 		dir = BUS_DMASYNC_PREREAD;
3186 
3187 	/* Check if a native SG list is needed for an NVMe PCIe device. */
3188 	if (cm->cm_targ && cm->cm_targ->is_nvme &&
3189 	    mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) {
3190 		/* A native SG list was built, skip to end. */
3191 		goto out;
3192 	}
3193 
3194 	for (i = 0; i < nsegs; i++) {
3195 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
3196 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3197 		}
3198 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3199 		    sflags, nsegs - i);
3200 		if (error != 0) {
3201 			/* Resource shortage, roll back! */
3202 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
3203 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
3204 				    "consider increasing hw.mpr.max_chains.\n");
3205 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
3206 			mpr_complete_command(sc, cm);
3207 			return;
3208 		}
3209 	}
3210 
3211 out:
3212 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3213 	mpr_enqueue_request(sc, cm);
3214 
3215 	return;
3216 }
3217 
3218 static void
3219 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3220 	     int error)
3221 {
3222 	mpr_data_cb(arg, segs, nsegs, error);
3223 }
3224 
3225 /*
3226  * This is the routine to enqueue commands ansynchronously.
3227  * Note that the only error path here is from bus_dmamap_load(), which can
3228  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3229  * assumed that if you have a command in-hand, then you have enough credits
3230  * to use it.
3231  */
3232 int
3233 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
3234 {
3235 	int error = 0;
3236 
3237 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
3238 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3239 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
3240 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
3241 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3242 		    cm->cm_data, mpr_data_cb, cm, 0);
3243 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3244 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3245 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
3246 	} else {
3247 		/* Add a zero-length element as needed */
3248 		if (cm->cm_sge != NULL)
3249 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
3250 		mpr_enqueue_request(sc, cm);
3251 	}
3252 
3253 	return (error);
3254 }
3255 
3256 /*
3257  * This is the routine to enqueue commands synchronously.  An error of
3258  * EINPROGRESS from mpr_map_command() is ignored since the command will
3259  * be executed and enqueued automatically.  Other errors come from msleep().
3260  */
3261 int
3262 mpr_wait_command(struct mpr_softc *sc, struct mpr_command *cm, int timeout,
3263     int sleep_flag)
3264 {
3265 	int error, rc;
3266 	struct timeval cur_time, start_time;
3267 
3268 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
3269 		return  EBUSY;
3270 
3271 	cm->cm_complete = NULL;
3272 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
3273 	error = mpr_map_command(sc, cm);
3274 	if ((error != 0) && (error != EINPROGRESS))
3275 		return (error);
3276 
3277 	// Check for context and wait for 50 mSec at a time until time has
3278 	// expired or the command has finished.  If msleep can't be used, need
3279 	// to poll.
3280 #if __FreeBSD_version >= 1000029
3281 	if (curthread->td_no_sleeping)
3282 #else //__FreeBSD_version < 1000029
3283 	if (curthread->td_pflags & TDP_NOSLEEPING)
3284 #endif //__FreeBSD_version >= 1000029
3285 		sleep_flag = NO_SLEEP;
3286 	getmicrotime(&start_time);
3287 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
3288 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
3289 	} else {
3290 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3291 			mpr_intr_locked(sc);
3292 			if (sleep_flag == CAN_SLEEP)
3293 				pause("mprwait", hz/20);
3294 			else
3295 				DELAY(50000);
3296 
3297 			getmicrotime(&cur_time);
3298 			if ((cur_time.tv_sec - start_time.tv_sec) > timeout) {
3299 				error = EWOULDBLOCK;
3300 				break;
3301 			}
3302 		}
3303 	}
3304 
3305 	if (error == EWOULDBLOCK) {
3306 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3307 		rc = mpr_reinit(sc);
3308 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3309 		    "failed");
3310 		error = ETIMEDOUT;
3311 	}
3312 	return (error);
3313 }
3314 
3315 /*
3316  * This is the routine to enqueue a command synchonously and poll for
3317  * completion.  Its use should be rare.
3318  */
3319 int
3320 mpr_request_polled(struct mpr_softc *sc, struct mpr_command *cm)
3321 {
3322 	int error, timeout = 0, rc;
3323 	struct timeval cur_time, start_time;
3324 
3325 	error = 0;
3326 
3327 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
3328 	cm->cm_complete = NULL;
3329 	mpr_map_command(sc, cm);
3330 
3331 	getmicrotime(&start_time);
3332 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3333 		mpr_intr_locked(sc);
3334 
3335 		if (mtx_owned(&sc->mpr_mtx))
3336 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
3337 			    "mprpoll", hz/20);
3338 		else
3339 			pause("mprpoll", hz/20);
3340 
3341 		/*
3342 		 * Check for real-time timeout and fail if more than 60 seconds.
3343 		 */
3344 		getmicrotime(&cur_time);
3345 		timeout = cur_time.tv_sec - start_time.tv_sec;
3346 		if (timeout > 60) {
3347 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
3348 			error = ETIMEDOUT;
3349 			break;
3350 		}
3351 	}
3352 
3353 	if (error) {
3354 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3355 		rc = mpr_reinit(sc);
3356 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3357 		    "failed");
3358 	}
3359 	return (error);
3360 }
3361 
3362 /*
3363  * The MPT driver had a verbose interface for config pages.  In this driver,
3364  * reduce it to much simpler terms, similar to the Linux driver.
3365  */
3366 int
3367 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3368 {
3369 	MPI2_CONFIG_REQUEST *req;
3370 	struct mpr_command *cm;
3371 	int error;
3372 
3373 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
3374 		return (EBUSY);
3375 	}
3376 
3377 	cm = mpr_alloc_command(sc);
3378 	if (cm == NULL) {
3379 		return (EBUSY);
3380 	}
3381 
3382 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3383 	req->Function = MPI2_FUNCTION_CONFIG;
3384 	req->Action = params->action;
3385 	req->SGLFlags = 0;
3386 	req->ChainOffset = 0;
3387 	req->PageAddress = params->page_address;
3388 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3389 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3390 
3391 		hdr = &params->hdr.Ext;
3392 		req->ExtPageType = hdr->ExtPageType;
3393 		req->ExtPageLength = hdr->ExtPageLength;
3394 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3395 		req->Header.PageLength = 0; /* Must be set to zero */
3396 		req->Header.PageNumber = hdr->PageNumber;
3397 		req->Header.PageVersion = hdr->PageVersion;
3398 	} else {
3399 		MPI2_CONFIG_PAGE_HEADER *hdr;
3400 
3401 		hdr = &params->hdr.Struct;
3402 		req->Header.PageType = hdr->PageType;
3403 		req->Header.PageNumber = hdr->PageNumber;
3404 		req->Header.PageLength = hdr->PageLength;
3405 		req->Header.PageVersion = hdr->PageVersion;
3406 	}
3407 
3408 	cm->cm_data = params->buffer;
3409 	cm->cm_length = params->length;
3410 	if (cm->cm_data != NULL) {
3411 		cm->cm_sge = &req->PageBufferSGE;
3412 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3413 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
3414 	} else
3415 		cm->cm_sge = NULL;
3416 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3417 
3418 	cm->cm_complete_data = params;
3419 	if (params->callback != NULL) {
3420 		cm->cm_complete = mpr_config_complete;
3421 		return (mpr_map_command(sc, cm));
3422 	} else {
3423 		error = mpr_wait_command(sc, cm, 0, CAN_SLEEP);
3424 		if (error) {
3425 			mpr_dprint(sc, MPR_FAULT,
3426 			    "Error %d reading config page\n", error);
3427 			mpr_free_command(sc, cm);
3428 			return (error);
3429 		}
3430 		mpr_config_complete(sc, cm);
3431 	}
3432 
3433 	return (0);
3434 }
3435 
3436 int
3437 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3438 {
3439 	return (EINVAL);
3440 }
3441 
3442 static void
3443 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
3444 {
3445 	MPI2_CONFIG_REPLY *reply;
3446 	struct mpr_config_params *params;
3447 
3448 	MPR_FUNCTRACE(sc);
3449 	params = cm->cm_complete_data;
3450 
3451 	if (cm->cm_data != NULL) {
3452 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3453 		    BUS_DMASYNC_POSTREAD);
3454 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3455 	}
3456 
3457 	/*
3458 	 * XXX KDM need to do more error recovery?  This results in the
3459 	 * device in question not getting probed.
3460 	 */
3461 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
3462 		params->status = MPI2_IOCSTATUS_BUSY;
3463 		goto done;
3464 	}
3465 
3466 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3467 	if (reply == NULL) {
3468 		params->status = MPI2_IOCSTATUS_BUSY;
3469 		goto done;
3470 	}
3471 	params->status = reply->IOCStatus;
3472 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3473 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3474 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3475 		params->hdr.Ext.PageType = reply->Header.PageType;
3476 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3477 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3478 	} else {
3479 		params->hdr.Struct.PageType = reply->Header.PageType;
3480 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3481 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3482 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3483 	}
3484 
3485 done:
3486 	mpr_free_command(sc, cm);
3487 	if (params->callback != NULL)
3488 		params->callback(sc, params);
3489 
3490 	return;
3491 }
3492