xref: /freebsd/sys/dev/mpr/mpr.c (revision e0c4386e7e71d93b0edc0c8fa156263fc4a8b0b6)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2016 Avago Technologies
5  * Copyright 2000-2020 Broadcom Inc.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  * Broadcom Inc. (LSI) MPT-Fusion Host Adapter FreeBSD
30  *
31  */
32 
33 #include <sys/cdefs.h>
34 /* Communications core for Avago Technologies (LSI) MPT3 */
35 
36 /* TODO Move headers to mprvar */
37 #include <sys/types.h>
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/selinfo.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/module.h>
45 #include <sys/bus.h>
46 #include <sys/conf.h>
47 #include <sys/bio.h>
48 #include <sys/malloc.h>
49 #include <sys/uio.h>
50 #include <sys/sysctl.h>
51 #include <sys/smp.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 #include <sys/sbuf.h>
58 #include <sys/priv.h>
59 
60 #include <machine/bus.h>
61 #include <machine/resource.h>
62 #include <sys/rman.h>
63 #include <sys/proc.h>
64 
65 #include <dev/pci/pcivar.h>
66 
67 #include <cam/cam.h>
68 #include <cam/cam_ccb.h>
69 #include <cam/scsi/scsi_all.h>
70 
71 #include <dev/mpr/mpi/mpi2_type.h>
72 #include <dev/mpr/mpi/mpi2.h>
73 #include <dev/mpr/mpi/mpi2_ioc.h>
74 #include <dev/mpr/mpi/mpi2_sas.h>
75 #include <dev/mpr/mpi/mpi2_pci.h>
76 #include <dev/mpr/mpi/mpi2_cnfg.h>
77 #include <dev/mpr/mpi/mpi2_init.h>
78 #include <dev/mpr/mpi/mpi2_tool.h>
79 #include <dev/mpr/mpr_ioctl.h>
80 #include <dev/mpr/mprvar.h>
81 #include <dev/mpr/mpr_table.h>
82 #include <dev/mpr/mpr_sas.h>
83 
84 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
85 static int mpr_init_queues(struct mpr_softc *sc);
86 static void mpr_resize_queues(struct mpr_softc *sc);
87 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
88 static int mpr_transition_operational(struct mpr_softc *sc);
89 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
90 static void mpr_iocfacts_free(struct mpr_softc *sc);
91 static void mpr_startup(void *arg);
92 static int mpr_send_iocinit(struct mpr_softc *sc);
93 static int mpr_alloc_queues(struct mpr_softc *sc);
94 static int mpr_alloc_hw_queues(struct mpr_softc *sc);
95 static int mpr_alloc_replies(struct mpr_softc *sc);
96 static int mpr_alloc_requests(struct mpr_softc *sc);
97 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc);
98 static int mpr_attach_log(struct mpr_softc *sc);
99 static __inline void mpr_complete_command(struct mpr_softc *sc,
100     struct mpr_command *cm);
101 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
102     MPI2_EVENT_NOTIFICATION_REPLY *reply);
103 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
104 static void mpr_periodic(void *);
105 static int mpr_reregister_events(struct mpr_softc *sc);
106 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
107 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
108 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
109 static int mpr_debug_sysctl(SYSCTL_HANDLER_ARGS);
110 static int mpr_dump_reqs(SYSCTL_HANDLER_ARGS);
111 static void mpr_parse_debug(struct mpr_softc *sc, char *list);
112 static void adjust_iocfacts_endianness(MPI2_IOC_FACTS_REPLY *facts);
113 
114 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
115     "MPR Driver Parameters");
116 
117 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory");
118 
119 /*
120  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
121  * any state and back to its initialization state machine.
122  */
123 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
124 
125 /*
126  * Added this union to smoothly convert le64toh cm->cm_desc.Words.
127  * Compiler only supports uint64_t to be passed as an argument.
128  * Otherwise it will throw this error:
129  * "aggregate value used where an integer was expected"
130  */
131 typedef union {
132         u64 word;
133         struct {
134                 u32 low;
135                 u32 high;
136         } u;
137 } request_descriptor_t;
138 
139 /* Rate limit chain-fail messages to 1 per minute */
140 static struct timeval mpr_chainfail_interval = { 60, 0 };
141 
142 /*
143  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
144  * If this function is called from process context, it can sleep
145  * and there is no harm to sleep, in case if this fuction is called
146  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
147  * based on sleep flags driver will call either msleep, pause or DELAY.
148  * msleep and pause are of same variant, but pause is used when mpr_mtx
149  * is not hold by driver.
150  */
151 static int
152 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag)
153 {
154 	uint32_t reg;
155 	int i, error, tries = 0;
156 	uint8_t first_wait_done = FALSE;
157 
158 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
159 
160 	/* Clear any pending interrupts */
161 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
162 
163 	/*
164 	 * Force NO_SLEEP for threads prohibited to sleep
165  	 * e.a Thread from interrupt handler are prohibited to sleep.
166  	 */
167 	if (curthread->td_no_sleeping)
168 		sleep_flag = NO_SLEEP;
169 
170 	mpr_dprint(sc, MPR_INIT, "sequence start, sleep_flag=%d\n", sleep_flag);
171 	/* Push the magic sequence */
172 	error = ETIMEDOUT;
173 	while (tries++ < 20) {
174 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
175 			mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
176 			    mpt2_reset_magic[i]);
177 
178 		/* wait 100 msec */
179 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
180 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
181 			    "mprdiag", hz/10);
182 		else if (sleep_flag == CAN_SLEEP)
183 			pause("mprdiag", hz/10);
184 		else
185 			DELAY(100 * 1000);
186 
187 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
188 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
189 			error = 0;
190 			break;
191 		}
192 	}
193 	if (error) {
194 		mpr_dprint(sc, MPR_INIT, "sequence failed, error=%d, exit\n",
195 		    error);
196 		return (error);
197 	}
198 
199 	/* Send the actual reset.  XXX need to refresh the reg? */
200 	reg |= MPI2_DIAG_RESET_ADAPTER;
201 	mpr_dprint(sc, MPR_INIT, "sequence success, sending reset, reg= 0x%x\n",
202 	    reg);
203 	mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
204 
205 	/* Wait up to 300 seconds in 50ms intervals */
206 	error = ETIMEDOUT;
207 	for (i = 0; i < 6000; i++) {
208 		/*
209 		 * Wait 50 msec. If this is the first time through, wait 256
210 		 * msec to satisfy Diag Reset timing requirements.
211 		 */
212 		if (first_wait_done) {
213 			if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
214 				msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
215 				    "mprdiag", hz/20);
216 			else if (sleep_flag == CAN_SLEEP)
217 				pause("mprdiag", hz/20);
218 			else
219 				DELAY(50 * 1000);
220 		} else {
221 			DELAY(256 * 1000);
222 			first_wait_done = TRUE;
223 		}
224 		/*
225 		 * Check for the RESET_ADAPTER bit to be cleared first, then
226 		 * wait for the RESET state to be cleared, which takes a little
227 		 * longer.
228 		 */
229 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
230 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
231 			continue;
232 		}
233 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
234 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
235 			error = 0;
236 			break;
237 		}
238 	}
239 	if (error) {
240 		mpr_dprint(sc, MPR_INIT, "reset failed, error= %d, exit\n",
241 		    error);
242 		return (error);
243 	}
244 
245 	mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
246 	mpr_dprint(sc, MPR_INIT, "diag reset success, exit\n");
247 
248 	return (0);
249 }
250 
251 static int
252 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag)
253 {
254 	int error;
255 
256 	MPR_FUNCTRACE(sc);
257 
258 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
259 
260 	error = 0;
261 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
262 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
263 	    MPI2_DOORBELL_FUNCTION_SHIFT);
264 
265 	if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) {
266 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
267 		    "Doorbell handshake failed\n");
268 		error = ETIMEDOUT;
269 	}
270 
271 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
272 	return (error);
273 }
274 
275 static int
276 mpr_transition_ready(struct mpr_softc *sc)
277 {
278 	uint32_t reg, state;
279 	int error, tries = 0;
280 	int sleep_flags;
281 
282 	MPR_FUNCTRACE(sc);
283 	/* If we are in attach call, do not sleep */
284 	sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE)
285 	    ? CAN_SLEEP : NO_SLEEP;
286 
287 	error = 0;
288 
289 	mpr_dprint(sc, MPR_INIT, "%s entered, sleep_flags= %d\n",
290 	    __func__, sleep_flags);
291 
292 	while (tries++ < 1200) {
293 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
294 		mpr_dprint(sc, MPR_INIT, "  Doorbell= 0x%x\n", reg);
295 
296 		/*
297 		 * Ensure the IOC is ready to talk.  If it's not, try
298 		 * resetting it.
299 		 */
300 		if (reg & MPI2_DOORBELL_USED) {
301 			mpr_dprint(sc, MPR_INIT, "  Not ready, sending diag "
302 			    "reset\n");
303 			mpr_diag_reset(sc, sleep_flags);
304 			DELAY(50000);
305 			continue;
306 		}
307 
308 		/* Is the adapter owned by another peer? */
309 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
310 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
311 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC is under the "
312 			    "control of another peer host, aborting "
313 			    "initialization.\n");
314 			error = ENXIO;
315 			break;
316 		}
317 
318 		state = reg & MPI2_IOC_STATE_MASK;
319 		if (state == MPI2_IOC_STATE_READY) {
320 			/* Ready to go! */
321 			error = 0;
322 			break;
323 		} else if (state == MPI2_IOC_STATE_FAULT) {
324 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC in fault "
325 			    "state 0x%x, resetting\n",
326 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
327 			mpr_diag_reset(sc, sleep_flags);
328 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
329 			/* Need to take ownership */
330 			mpr_message_unit_reset(sc, sleep_flags);
331 		} else if (state == MPI2_IOC_STATE_RESET) {
332 			/* Wait a bit, IOC might be in transition */
333 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
334 			    "IOC in unexpected reset state\n");
335 		} else {
336 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
337 			    "IOC in unknown state 0x%x\n", state);
338 			error = EINVAL;
339 			break;
340 		}
341 
342 		/* Wait 50ms for things to settle down. */
343 		DELAY(50000);
344 	}
345 
346 	if (error)
347 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
348 		    "Cannot transition IOC to ready\n");
349 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
350 	return (error);
351 }
352 
353 static int
354 mpr_transition_operational(struct mpr_softc *sc)
355 {
356 	uint32_t reg, state;
357 	int error;
358 
359 	MPR_FUNCTRACE(sc);
360 
361 	error = 0;
362 	reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
363 	mpr_dprint(sc, MPR_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
364 
365 	state = reg & MPI2_IOC_STATE_MASK;
366 	if (state != MPI2_IOC_STATE_READY) {
367 		mpr_dprint(sc, MPR_INIT, "IOC not ready\n");
368 		if ((error = mpr_transition_ready(sc)) != 0) {
369 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
370 			    "failed to transition ready, exit\n");
371 			return (error);
372 		}
373 	}
374 
375 	error = mpr_send_iocinit(sc);
376 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
377 
378 	return (error);
379 }
380 
381 static void
382 mpr_resize_queues(struct mpr_softc *sc)
383 {
384 	u_int reqcr, prireqcr, maxio, sges_per_frame, chain_seg_size;
385 
386 	/*
387 	 * Size the queues. Since the reply queues always need one free
388 	 * entry, we'll deduct one reply message here.  The LSI documents
389 	 * suggest instead to add a count to the request queue, but I think
390 	 * that it's better to deduct from reply queue.
391 	 */
392 	prireqcr = MAX(1, sc->max_prireqframes);
393 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
394 
395 	reqcr = MAX(2, sc->max_reqframes);
396 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
397 
398 	sc->num_reqs = prireqcr + reqcr;
399 	sc->num_prireqs = prireqcr;
400 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
401 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
402 
403 	/* Store the request frame size in bytes rather than as 32bit words */
404 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
405 
406 	/*
407 	 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to
408 	 * get the size of a Chain Frame.  Previous versions use the size as a
409 	 * Request Frame for the Chain Frame size.  If IOCMaxChainSegmentSize
410 	 * is 0, use the default value.  The IOCMaxChainSegmentSize is the
411 	 * number of 16-byte elelements that can fit in a Chain Frame, which is
412 	 * the size of an IEEE Simple SGE.
413 	 */
414 	if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) {
415 		chain_seg_size = sc->facts->IOCMaxChainSegmentSize;
416 		if (chain_seg_size == 0)
417 			chain_seg_size = MPR_DEFAULT_CHAIN_SEG_SIZE;
418 		sc->chain_frame_size = chain_seg_size *
419 		    MPR_MAX_CHAIN_ELEMENT_SIZE;
420 	} else {
421 		sc->chain_frame_size = sc->reqframesz;
422 	}
423 
424 	/*
425 	 * Max IO Size is Page Size * the following:
426 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
427 	 * + 1 for no chain needed in last frame
428 	 *
429 	 * If user suggests a Max IO size to use, use the smaller of the
430 	 * user's value and the calculated value as long as the user's
431 	 * value is larger than 0. The user's value is in pages.
432 	 */
433 	sges_per_frame = sc->chain_frame_size/sizeof(MPI2_IEEE_SGE_SIMPLE64)-1;
434 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
435 
436 	/*
437 	 * If I/O size limitation requested then use it and pass up to CAM.
438 	 * If not, use maxphys as an optimization hint, but report HW limit.
439 	 */
440 	if (sc->max_io_pages > 0) {
441 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
442 		sc->maxio = maxio;
443 	} else {
444 		sc->maxio = maxio;
445 		maxio = min(maxio, maxphys);
446 	}
447 
448 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
449 	    sges_per_frame * reqcr;
450 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
451 		sc->num_chains = sc->max_chains;
452 
453 	/*
454 	 * Figure out the number of MSIx-based queues.  If the firmware or
455 	 * user has done something crazy and not allowed enough credit for
456 	 * the queues to be useful then don't enable multi-queue.
457 	 */
458 	if (sc->facts->MaxMSIxVectors < 2)
459 		sc->msi_msgs = 1;
460 
461 	if (sc->msi_msgs > 1) {
462 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
463 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
464 		if (sc->num_reqs / sc->msi_msgs < 2)
465 			sc->msi_msgs = 1;
466 	}
467 
468 	mpr_dprint(sc, MPR_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
469 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
470 }
471 
472 /*
473  * This is called during attach and when re-initializing due to a Diag Reset.
474  * IOC Facts is used to allocate many of the structures needed by the driver.
475  * If called from attach, de-allocation is not required because the driver has
476  * not allocated any structures yet, but if called from a Diag Reset, previously
477  * allocated structures based on IOC Facts will need to be freed and re-
478  * allocated bases on the latest IOC Facts.
479  */
480 static int
481 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
482 {
483 	int error;
484 	Mpi2IOCFactsReply_t saved_facts;
485 	uint8_t saved_mode, reallocating;
486 
487 	mpr_dprint(sc, MPR_INIT|MPR_TRACE, "%s entered\n", __func__);
488 
489 	/* Save old IOC Facts and then only reallocate if Facts have changed */
490 	if (!attaching) {
491 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
492 	}
493 
494 	/*
495 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
496 	 * a re-initialization and only return the error if attaching so the OS
497 	 * can handle it.
498 	 */
499 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
500 		if (attaching) {
501 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to get "
502 			    "IOC Facts with error %d, exit\n", error);
503 			return (error);
504 		} else {
505 			panic("%s failed to get IOC Facts with error %d\n",
506 			    __func__, error);
507 		}
508 	}
509 
510 	MPR_DPRINT_PAGE(sc, MPR_XINFO, iocfacts, sc->facts);
511 
512 	snprintf(sc->fw_version, sizeof(sc->fw_version),
513 	    "%02d.%02d.%02d.%02d",
514 	    sc->facts->FWVersion.Struct.Major,
515 	    sc->facts->FWVersion.Struct.Minor,
516 	    sc->facts->FWVersion.Struct.Unit,
517 	    sc->facts->FWVersion.Struct.Dev);
518 
519 	snprintf(sc->msg_version, sizeof(sc->msg_version), "%d.%d",
520 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MAJOR_MASK) >>
521 	    MPI2_IOCFACTS_MSGVERSION_MAJOR_SHIFT,
522 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MINOR_MASK) >>
523 	    MPI2_IOCFACTS_MSGVERSION_MINOR_SHIFT);
524 
525 	mpr_dprint(sc, MPR_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
526 	    MPR_DRIVER_VERSION);
527 	mpr_dprint(sc, MPR_INFO,
528 	    "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
529 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
530 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
531 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc"
532 	    "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV");
533 
534 	/*
535 	 * If the chip doesn't support event replay then a hard reset will be
536 	 * required to trigger a full discovery.  Do the reset here then
537 	 * retransition to Ready.  A hard reset might have already been done,
538 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
539 	 * for a Diag Reset.
540 	 */
541 	if (attaching && ((sc->facts->IOCCapabilities &
542 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
543 		mpr_dprint(sc, MPR_INIT, "No event replay, resetting\n");
544 		mpr_diag_reset(sc, NO_SLEEP);
545 		if ((error = mpr_transition_ready(sc)) != 0) {
546 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
547 			    "transition to ready with error %d, exit\n",
548 			    error);
549 			return (error);
550 		}
551 	}
552 
553 	/*
554 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
555 	 * changed from the previous IOC Facts, log a warning, but only if
556 	 * checking this after a Diag Reset and not during attach.
557 	 */
558 	saved_mode = sc->ir_firmware;
559 	if (sc->facts->IOCCapabilities &
560 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
561 		sc->ir_firmware = 1;
562 	if (!attaching) {
563 		if (sc->ir_firmware != saved_mode) {
564 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "new IR/IT mode "
565 			    "in IOC Facts does not match previous mode\n");
566 		}
567 	}
568 
569 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
570 	reallocating = FALSE;
571 	sc->mpr_flags &= ~MPR_FLAGS_REALLOCATED;
572 
573 	if ((!attaching) &&
574 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
575 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
576 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
577 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
578 	    (saved_facts.ProductID != sc->facts->ProductID) ||
579 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
580 	    (saved_facts.IOCRequestFrameSize !=
581 	    sc->facts->IOCRequestFrameSize) ||
582 	    (saved_facts.IOCMaxChainSegmentSize !=
583 	    sc->facts->IOCMaxChainSegmentSize) ||
584 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
585 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
586 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
587 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
588 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
589 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
590 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
591 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
592 	    (saved_facts.MaxPersistentEntries !=
593 	    sc->facts->MaxPersistentEntries))) {
594 		reallocating = TRUE;
595 
596 		/* Record that we reallocated everything */
597 		sc->mpr_flags |= MPR_FLAGS_REALLOCATED;
598 	}
599 
600 	/*
601 	 * Some things should be done if attaching or re-allocating after a Diag
602 	 * Reset, but are not needed after a Diag Reset if the FW has not
603 	 * changed.
604 	 */
605 	if (attaching || reallocating) {
606 		/*
607 		 * Check if controller supports FW diag buffers and set flag to
608 		 * enable each type.
609 		 */
610 		if (sc->facts->IOCCapabilities &
611 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
612 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
613 			    enabled = TRUE;
614 		if (sc->facts->IOCCapabilities &
615 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
616 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
617 			    enabled = TRUE;
618 		if (sc->facts->IOCCapabilities &
619 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
620 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
621 			    enabled = TRUE;
622 
623 		/*
624 		 * Set flags for some supported items.
625 		 */
626 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
627 			sc->eedp_enabled = TRUE;
628 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
629 			sc->control_TLR = TRUE;
630 		if ((sc->facts->IOCCapabilities &
631 		    MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ) &&
632 		    (sc->mpr_flags & MPR_FLAGS_SEA_IOC))
633 			sc->atomic_desc_capable = TRUE;
634 
635 		mpr_resize_queues(sc);
636 
637 		/*
638 		 * Initialize all Tail Queues
639 		 */
640 		TAILQ_INIT(&sc->req_list);
641 		TAILQ_INIT(&sc->high_priority_req_list);
642 		TAILQ_INIT(&sc->chain_list);
643 		TAILQ_INIT(&sc->prp_page_list);
644 		TAILQ_INIT(&sc->tm_list);
645 	}
646 
647 	/*
648 	 * If doing a Diag Reset and the FW is significantly different
649 	 * (reallocating will be set above in IOC Facts comparison), then all
650 	 * buffers based on the IOC Facts will need to be freed before they are
651 	 * reallocated.
652 	 */
653 	if (reallocating) {
654 		mpr_iocfacts_free(sc);
655 		mprsas_realloc_targets(sc, saved_facts.MaxTargets +
656 		    saved_facts.MaxVolumes);
657 	}
658 
659 	/*
660 	 * Any deallocation has been completed.  Now start reallocating
661 	 * if needed.  Will only need to reallocate if attaching or if the new
662 	 * IOC Facts are different from the previous IOC Facts after a Diag
663 	 * Reset. Targets have already been allocated above if needed.
664 	 */
665 	error = 0;
666 	while (attaching || reallocating) {
667 		if ((error = mpr_alloc_hw_queues(sc)) != 0)
668 			break;
669 		if ((error = mpr_alloc_replies(sc)) != 0)
670 			break;
671 		if ((error = mpr_alloc_requests(sc)) != 0)
672 			break;
673 		if ((error = mpr_alloc_queues(sc)) != 0)
674 			break;
675 		break;
676 	}
677 	if (error) {
678 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
679 		    "Failed to alloc queues with error %d\n", error);
680 		mpr_free(sc);
681 		return (error);
682 	}
683 
684 	/* Always initialize the queues */
685 	bzero(sc->free_queue, sc->fqdepth * 4);
686 	mpr_init_queues(sc);
687 
688 	/*
689 	 * Always get the chip out of the reset state, but only panic if not
690 	 * attaching.  If attaching and there is an error, that is handled by
691 	 * the OS.
692 	 */
693 	error = mpr_transition_operational(sc);
694 	if (error != 0) {
695 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
696 		    "transition to operational with error %d\n", error);
697 		mpr_free(sc);
698 		return (error);
699 	}
700 
701 	/*
702 	 * Finish the queue initialization.
703 	 * These are set here instead of in mpr_init_queues() because the
704 	 * IOC resets these values during the state transition in
705 	 * mpr_transition_operational().  The free index is set to 1
706 	 * because the corresponding index in the IOC is set to 0, and the
707 	 * IOC treats the queues as full if both are set to the same value.
708 	 * Hence the reason that the queue can't hold all of the possible
709 	 * replies.
710 	 */
711 	sc->replypostindex = 0;
712 	mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
713 	mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
714 
715 	/*
716 	 * Attach the subsystems so they can prepare their event masks.
717 	 * XXX Should be dynamic so that IM/IR and user modules can attach
718 	 */
719 	error = 0;
720 	while (attaching) {
721 		mpr_dprint(sc, MPR_INIT, "Attaching subsystems\n");
722 		if ((error = mpr_attach_log(sc)) != 0)
723 			break;
724 		if ((error = mpr_attach_sas(sc)) != 0)
725 			break;
726 		if ((error = mpr_attach_user(sc)) != 0)
727 			break;
728 		break;
729 	}
730 	if (error) {
731 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
732 		    "Failed to attach all subsystems: error %d\n", error);
733 		mpr_free(sc);
734 		return (error);
735 	}
736 
737 	/*
738 	 * XXX If the number of MSI-X vectors changes during re-init, this
739 	 * won't see it and adjust.
740 	 */
741 	if ((attaching || reallocating) && (error = mpr_pci_setup_interrupts(sc)) != 0) {
742 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
743 		    "Failed to setup interrupts\n");
744 		mpr_free(sc);
745 		return (error);
746 	}
747 
748 	return (error);
749 }
750 
751 /*
752  * This is called if memory is being free (during detach for example) and when
753  * buffers need to be reallocated due to a Diag Reset.
754  */
755 static void
756 mpr_iocfacts_free(struct mpr_softc *sc)
757 {
758 	struct mpr_command *cm;
759 	int i;
760 
761 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
762 
763 	if (sc->free_busaddr != 0)
764 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
765 	if (sc->free_queue != NULL)
766 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
767 		    sc->queues_map);
768 	if (sc->queues_dmat != NULL)
769 		bus_dma_tag_destroy(sc->queues_dmat);
770 
771 	if (sc->chain_frames != NULL) {
772 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
773 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
774 		    sc->chain_map);
775 	}
776 	if (sc->chain_dmat != NULL)
777 		bus_dma_tag_destroy(sc->chain_dmat);
778 
779 	if (sc->sense_busaddr != 0)
780 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
781 	if (sc->sense_frames != NULL)
782 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
783 		    sc->sense_map);
784 	if (sc->sense_dmat != NULL)
785 		bus_dma_tag_destroy(sc->sense_dmat);
786 
787 	if (sc->prp_page_busaddr != 0)
788 		bus_dmamap_unload(sc->prp_page_dmat, sc->prp_page_map);
789 	if (sc->prp_pages != NULL)
790 		bus_dmamem_free(sc->prp_page_dmat, sc->prp_pages,
791 		    sc->prp_page_map);
792 	if (sc->prp_page_dmat != NULL)
793 		bus_dma_tag_destroy(sc->prp_page_dmat);
794 
795 	if (sc->reply_busaddr != 0)
796 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
797 	if (sc->reply_frames != NULL)
798 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
799 		    sc->reply_map);
800 	if (sc->reply_dmat != NULL)
801 		bus_dma_tag_destroy(sc->reply_dmat);
802 
803 	if (sc->req_busaddr != 0)
804 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
805 	if (sc->req_frames != NULL)
806 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
807 	if (sc->req_dmat != NULL)
808 		bus_dma_tag_destroy(sc->req_dmat);
809 
810 	if (sc->chains != NULL)
811 		free(sc->chains, M_MPR);
812 	if (sc->prps != NULL)
813 		free(sc->prps, M_MPR);
814 	if (sc->commands != NULL) {
815 		for (i = 1; i < sc->num_reqs; i++) {
816 			cm = &sc->commands[i];
817 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
818 		}
819 		free(sc->commands, M_MPR);
820 	}
821 	if (sc->buffer_dmat != NULL)
822 		bus_dma_tag_destroy(sc->buffer_dmat);
823 
824 	mpr_pci_free_interrupts(sc);
825 	free(sc->queues, M_MPR);
826 	sc->queues = NULL;
827 }
828 
829 /*
830  * The terms diag reset and hard reset are used interchangeably in the MPI
831  * docs to mean resetting the controller chip.  In this code diag reset
832  * cleans everything up, and the hard reset function just sends the reset
833  * sequence to the chip.  This should probably be refactored so that every
834  * subsystem gets a reset notification of some sort, and can clean up
835  * appropriately.
836  */
837 int
838 mpr_reinit(struct mpr_softc *sc)
839 {
840 	int error;
841 	struct mprsas_softc *sassc;
842 
843 	sassc = sc->sassc;
844 
845 	MPR_FUNCTRACE(sc);
846 
847 	mtx_assert(&sc->mpr_mtx, MA_OWNED);
848 
849 	mpr_dprint(sc, MPR_INIT|MPR_INFO, "Reinitializing controller\n");
850 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) {
851 		mpr_dprint(sc, MPR_INIT, "Reset already in progress\n");
852 		return 0;
853 	}
854 
855 	/*
856 	 * Make sure the completion callbacks can recognize they're getting
857 	 * a NULL cm_reply due to a reset.
858 	 */
859 	sc->mpr_flags |= MPR_FLAGS_DIAGRESET;
860 
861 	/*
862 	 * Mask interrupts here.
863 	 */
864 	mpr_dprint(sc, MPR_INIT, "Masking interrupts and resetting\n");
865 	mpr_mask_intr(sc);
866 
867 	error = mpr_diag_reset(sc, CAN_SLEEP);
868 	if (error != 0) {
869 		panic("%s hard reset failed with error %d\n", __func__, error);
870 	}
871 
872 	/* Restore the PCI state, including the MSI-X registers */
873 	mpr_pci_restore(sc);
874 
875 	/* Give the I/O subsystem special priority to get itself prepared */
876 	mprsas_handle_reinit(sc);
877 
878 	/*
879 	 * Get IOC Facts and allocate all structures based on this information.
880 	 * The attach function will also call mpr_iocfacts_allocate at startup.
881 	 * If relevant values have changed in IOC Facts, this function will free
882 	 * all of the memory based on IOC Facts and reallocate that memory.
883 	 */
884 	if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) {
885 		panic("%s IOC Facts based allocation failed with error %d\n",
886 		    __func__, error);
887 	}
888 
889 	/*
890 	 * Mapping structures will be re-allocated after getting IOC Page8, so
891 	 * free these structures here.
892 	 */
893 	mpr_mapping_exit(sc);
894 
895 	/*
896 	 * The static page function currently read is IOC Page8.  Others can be
897 	 * added in future.  It's possible that the values in IOC Page8 have
898 	 * changed after a Diag Reset due to user modification, so always read
899 	 * these.  Interrupts are masked, so unmask them before getting config
900 	 * pages.
901 	 */
902 	mpr_unmask_intr(sc);
903 	sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET;
904 	mpr_base_static_config_pages(sc);
905 
906 	/*
907 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
908 	 * mapping tables.
909 	 */
910 	mpr_mapping_initialize(sc);
911 
912 	/*
913 	 * Restart will reload the event masks clobbered by the reset, and
914 	 * then enable the port.
915 	 */
916 	mpr_reregister_events(sc);
917 
918 	/* the end of discovery will release the simq, so we're done. */
919 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Finished sc %p post %u free %u\n",
920 	    sc, sc->replypostindex, sc->replyfreeindex);
921 	mprsas_release_simq_reinit(sassc);
922 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
923 
924 	return 0;
925 }
926 
927 /* Wait for the chip to ACK a word that we've put into its FIFO
928  * Wait for <timeout> seconds. In single loop wait for busy loop
929  * for 500 microseconds.
930  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
931  * */
932 static int
933 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag)
934 {
935 	u32 cntdn, count;
936 	u32 int_status;
937 	u32 doorbell;
938 
939 	count = 0;
940 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
941 	do {
942 		int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
943 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
944 			mpr_dprint(sc, MPR_TRACE, "%s: successful count(%d), "
945 			    "timeout(%d)\n", __func__, count, timeout);
946 			return 0;
947 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
948 			doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
949 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
950 			    MPI2_IOC_STATE_FAULT) {
951 				mpr_dprint(sc, MPR_FAULT,
952 				    "fault_state(0x%04x)!\n", doorbell);
953 				return (EFAULT);
954 			}
955 		} else if (int_status == 0xFFFFFFFF)
956 			goto out;
957 
958 		/*
959 		 * If it can sleep, sleep for 1 milisecond, else busy loop for
960  		 * 0.5 milisecond
961 		 */
962 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
963 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba",
964 			    hz/1000);
965 		else if (sleep_flag == CAN_SLEEP)
966 			pause("mprdba", hz/1000);
967 		else
968 			DELAY(500);
969 		count++;
970 	} while (--cntdn);
971 
972 out:
973 	mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), "
974 		"int_status(%x)!\n", __func__, count, int_status);
975 	return (ETIMEDOUT);
976 }
977 
978 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
979 static int
980 mpr_wait_db_int(struct mpr_softc *sc)
981 {
982 	int retry;
983 
984 	for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) {
985 		if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
986 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
987 			return (0);
988 		DELAY(2000);
989 	}
990 	return (ETIMEDOUT);
991 }
992 
993 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
994 static int
995 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
996     int req_sz, int reply_sz, int timeout)
997 {
998 	uint32_t *data32;
999 	uint16_t *data16;
1000 	int i, count, ioc_sz, residual;
1001 	int sleep_flags = CAN_SLEEP;
1002 
1003 	if (curthread->td_no_sleeping)
1004 		sleep_flags = NO_SLEEP;
1005 
1006 	/* Step 1 */
1007 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1008 
1009 	/* Step 2 */
1010 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1011 		return (EBUSY);
1012 
1013 	/* Step 3
1014 	 * Announce that a message is coming through the doorbell.  Messages
1015 	 * are pushed at 32bit words, so round up if needed.
1016 	 */
1017 	count = (req_sz + 3) / 4;
1018 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
1019 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
1020 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
1021 
1022 	/* Step 4 */
1023 	if (mpr_wait_db_int(sc) ||
1024 	    (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
1025 		mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n");
1026 		return (ENXIO);
1027 	}
1028 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1029 	if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
1030 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n");
1031 		return (ENXIO);
1032 	}
1033 
1034 	/* Step 5 */
1035 	/* Clock out the message data synchronously in 32-bit dwords*/
1036 	data32 = (uint32_t *)req;
1037 	for (i = 0; i < count; i++) {
1038 		mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1039 		if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
1040 			mpr_dprint(sc, MPR_FAULT,
1041 			    "Timeout while writing doorbell\n");
1042 			return (ENXIO);
1043 		}
1044 	}
1045 
1046 	/* Step 6 */
1047 	/* Clock in the reply in 16-bit words.  The total length of the
1048 	 * message is always in the 4th byte, so clock out the first 2 words
1049 	 * manually, then loop the rest.
1050 	 */
1051 	data16 = (uint16_t *)reply;
1052 	if (mpr_wait_db_int(sc) != 0) {
1053 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n");
1054 		return (ENXIO);
1055 	}
1056 
1057 	/*
1058 	 * If in a BE platform, swap bytes using le16toh to not
1059 	 * disturb 8 bit field neighbors in destination structure
1060 	 * pointed by data16.
1061 	 */
1062 	data16[0] =
1063 	    le16toh(mpr_regread(sc, MPI2_DOORBELL_OFFSET)) & MPI2_DOORBELL_DATA_MASK;
1064 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1065 	if (mpr_wait_db_int(sc) != 0) {
1066 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n");
1067 		return (ENXIO);
1068 	}
1069 	data16[1] =
1070 	    le16toh(mpr_regread(sc, MPI2_DOORBELL_OFFSET)) & MPI2_DOORBELL_DATA_MASK;
1071 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1072 
1073 	/* Number of 32bit words in the message */
1074 	ioc_sz = reply->MsgLength;
1075 
1076 	/*
1077 	 * Figure out how many 16bit words to clock in without overrunning.
1078 	 * The precision loss with dividing reply_sz can safely be
1079 	 * ignored because the messages can only be multiples of 32bits.
1080 	 */
1081 	residual = 0;
1082 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1083 	if (count < ioc_sz * 2) {
1084 		residual = ioc_sz * 2 - count;
1085 		mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d "
1086 		    "residual message words\n", residual);
1087 	}
1088 
1089 	for (i = 2; i < count; i++) {
1090 		if (mpr_wait_db_int(sc) != 0) {
1091 			mpr_dprint(sc, MPR_FAULT,
1092 			    "Timeout reading doorbell %d\n", i);
1093 			return (ENXIO);
1094 		}
1095 		data16[i] = le16toh(mpr_regread(sc, MPI2_DOORBELL_OFFSET)) &
1096 		    MPI2_DOORBELL_DATA_MASK;
1097 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1098 	}
1099 
1100 	/*
1101 	 * Pull out residual words that won't fit into the provided buffer.
1102 	 * This keeps the chip from hanging due to a driver programming
1103 	 * error.
1104 	 */
1105 	while (residual--) {
1106 		if (mpr_wait_db_int(sc) != 0) {
1107 			mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n");
1108 			return (ENXIO);
1109 		}
1110 		(void)mpr_regread(sc, MPI2_DOORBELL_OFFSET);
1111 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1112 	}
1113 
1114 	/* Step 7 */
1115 	if (mpr_wait_db_int(sc) != 0) {
1116 		mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n");
1117 		return (ENXIO);
1118 	}
1119 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1120 		mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n");
1121 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1122 
1123 	return (0);
1124 }
1125 
1126 static void
1127 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm)
1128 {
1129 	request_descriptor_t rd;
1130 
1131 	MPR_FUNCTRACE(sc);
1132 	mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n",
1133 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1134 
1135 	if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags &
1136 	    MPR_FLAGS_SHUTDOWN))
1137 		mtx_assert(&sc->mpr_mtx, MA_OWNED);
1138 
1139 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1140 		sc->io_cmds_highwater++;
1141 
1142 	KASSERT(cm->cm_state == MPR_CM_STATE_BUSY,
1143 	    ("command not busy, state = %u\n", cm->cm_state));
1144 	cm->cm_state = MPR_CM_STATE_INQUEUE;
1145 
1146 	if (sc->atomic_desc_capable) {
1147 		rd.u.low = cm->cm_desc.Words.Low;
1148 		mpr_regwrite(sc, MPI26_ATOMIC_REQUEST_DESCRIPTOR_POST_OFFSET,
1149 		    rd.u.low);
1150 	} else {
1151 		rd.u.low = htole32(cm->cm_desc.Words.Low);
1152 		rd.u.high = htole32(cm->cm_desc.Words.High);
1153 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1154 		    rd.u.low);
1155 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1156 		    rd.u.high);
1157 	}
1158 }
1159 
1160 /*
1161  * Ioc facts are read in 16 bit words and and stored with le16toh,
1162  * this takes care of proper U8 fields endianness in
1163  * MPI2_IOC_FACTS_REPLY, but we still need to swap back U16 fields.
1164  */
1165 static void
1166 adjust_iocfacts_endianness(MPI2_IOC_FACTS_REPLY *facts)
1167 {
1168 	facts->HeaderVersion = le16toh(facts->HeaderVersion);
1169 	facts->Reserved1 = le16toh(facts->Reserved1);
1170 	facts->IOCExceptions = le16toh(facts->IOCExceptions);
1171 	facts->IOCStatus = le16toh(facts->IOCStatus);
1172 	facts->IOCLogInfo = le32toh(facts->IOCLogInfo);
1173 	facts->RequestCredit = le16toh(facts->RequestCredit);
1174 	facts->ProductID = le16toh(facts->ProductID);
1175 	facts->IOCCapabilities = le32toh(facts->IOCCapabilities);
1176 	facts->IOCRequestFrameSize = le16toh(facts->IOCRequestFrameSize);
1177 	facts->IOCMaxChainSegmentSize = le16toh(facts->IOCMaxChainSegmentSize);
1178 	facts->MaxInitiators = le16toh(facts->MaxInitiators);
1179 	facts->MaxTargets = le16toh(facts->MaxTargets);
1180 	facts->MaxSasExpanders = le16toh(facts->MaxSasExpanders);
1181 	facts->MaxEnclosures = le16toh(facts->MaxEnclosures);
1182 	facts->ProtocolFlags = le16toh(facts->ProtocolFlags);
1183 	facts->HighPriorityCredit = le16toh(facts->HighPriorityCredit);
1184 	facts->MaxReplyDescriptorPostQueueDepth = le16toh(facts->MaxReplyDescriptorPostQueueDepth);
1185 	facts->MaxDevHandle = le16toh(facts->MaxDevHandle);
1186 	facts->MaxPersistentEntries = le16toh(facts->MaxPersistentEntries);
1187 	facts->MinDevHandle = le16toh(facts->MinDevHandle);
1188 }
1189 
1190 /*
1191  * Just the FACTS, ma'am.
1192  */
1193 static int
1194 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1195 {
1196 	MPI2_DEFAULT_REPLY *reply;
1197 	MPI2_IOC_FACTS_REQUEST request;
1198 	int error, req_sz, reply_sz;
1199 
1200 	MPR_FUNCTRACE(sc);
1201 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1202 
1203 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1204 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1205 	reply = (MPI2_DEFAULT_REPLY *)facts;
1206 
1207 	bzero(&request, req_sz);
1208 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1209 	error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1210 
1211 	adjust_iocfacts_endianness(facts);
1212 	mpr_dprint(sc, MPR_TRACE, "facts->IOCCapabilities 0x%x\n", facts->IOCCapabilities);
1213 
1214 	mpr_dprint(sc, MPR_INIT, "%s exit, error= %d\n", __func__, error);
1215 	return (error);
1216 }
1217 
1218 static int
1219 mpr_send_iocinit(struct mpr_softc *sc)
1220 {
1221 	MPI2_IOC_INIT_REQUEST	init;
1222 	MPI2_DEFAULT_REPLY	reply;
1223 	int req_sz, reply_sz, error;
1224 	struct timeval now;
1225 	uint64_t time_in_msec;
1226 
1227 	MPR_FUNCTRACE(sc);
1228 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1229 
1230 	/* Do a quick sanity check on proper initialization */
1231 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1232 	    || (sc->replyframesz == 0)) {
1233 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
1234 		    "Driver not fully initialized for IOCInit\n");
1235 		return (EINVAL);
1236 	}
1237 
1238 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1239 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1240 	bzero(&init, req_sz);
1241 	bzero(&reply, reply_sz);
1242 
1243 	/*
1244 	 * Fill in the init block.  Note that most addresses are
1245 	 * deliberately in the lower 32bits of memory.  This is a micro-
1246 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1247 	 */
1248 	init.Function = MPI2_FUNCTION_IOC_INIT;
1249 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1250 	init.MsgVersion = htole16(MPI2_VERSION);
1251 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1252 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1253 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1254 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1255 	init.SenseBufferAddressHigh = 0;
1256 	init.SystemReplyAddressHigh = 0;
1257 	init.SystemRequestFrameBaseAddress.High = 0;
1258 	init.SystemRequestFrameBaseAddress.Low =
1259 	    htole32((uint32_t)sc->req_busaddr);
1260 	init.ReplyDescriptorPostQueueAddress.High = 0;
1261 	init.ReplyDescriptorPostQueueAddress.Low =
1262 	    htole32((uint32_t)sc->post_busaddr);
1263 	init.ReplyFreeQueueAddress.High = 0;
1264 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1265 	getmicrotime(&now);
1266 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1267 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1268 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1269 	init.HostPageSize = HOST_PAGE_SIZE_4K;
1270 
1271 	error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1272 	if ((le16toh(reply.IOCStatus) & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1273 		error = ENXIO;
1274 
1275 	mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", le16toh(reply.IOCStatus));
1276 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
1277 	return (error);
1278 }
1279 
1280 void
1281 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1282 {
1283 	bus_addr_t *addr;
1284 
1285 	addr = arg;
1286 	*addr = segs[0].ds_addr;
1287 }
1288 
1289 void
1290 mpr_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1291 {
1292 	struct mpr_busdma_context *ctx;
1293 	int need_unload, need_free;
1294 
1295 	ctx = (struct mpr_busdma_context *)arg;
1296 	need_unload = 0;
1297 	need_free = 0;
1298 
1299 	mpr_lock(ctx->softc);
1300 	ctx->error = error;
1301 	ctx->completed = 1;
1302 	if ((error == 0) && (ctx->abandoned == 0)) {
1303 		*ctx->addr = segs[0].ds_addr;
1304 	} else {
1305 		if (nsegs != 0)
1306 			need_unload = 1;
1307 		if (ctx->abandoned != 0)
1308 			need_free = 1;
1309 	}
1310 	if (need_free == 0)
1311 		wakeup(ctx);
1312 
1313 	mpr_unlock(ctx->softc);
1314 
1315 	if (need_unload != 0) {
1316 		bus_dmamap_unload(ctx->buffer_dmat,
1317 				  ctx->buffer_dmamap);
1318 		*ctx->addr = 0;
1319 	}
1320 
1321 	if (need_free != 0)
1322 		free(ctx, M_MPR);
1323 }
1324 
1325 static int
1326 mpr_alloc_queues(struct mpr_softc *sc)
1327 {
1328 	struct mpr_queue *q;
1329 	int nq, i;
1330 
1331 	nq = sc->msi_msgs;
1332 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Allocating %d I/O queues\n", nq);
1333 
1334 	sc->queues = malloc(sizeof(struct mpr_queue) * nq, M_MPR,
1335 	     M_NOWAIT|M_ZERO);
1336 	if (sc->queues == NULL)
1337 		return (ENOMEM);
1338 
1339 	for (i = 0; i < nq; i++) {
1340 		q = &sc->queues[i];
1341 		mpr_dprint(sc, MPR_INIT, "Configuring queue %d %p\n", i, q);
1342 		q->sc = sc;
1343 		q->qnum = i;
1344 	}
1345 	return (0);
1346 }
1347 
1348 static int
1349 mpr_alloc_hw_queues(struct mpr_softc *sc)
1350 {
1351 	bus_dma_template_t t;
1352 	bus_addr_t queues_busaddr;
1353 	uint8_t *queues;
1354 	int qsize, fqsize, pqsize;
1355 
1356 	/*
1357 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1358 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1359 	 * This queue supplies fresh reply frames for the firmware to use.
1360 	 *
1361 	 * The reply descriptor post queue contains 8 byte entries in
1362 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1363 	 * contains filled-in reply frames sent from the firmware to the host.
1364 	 *
1365 	 * These two queues are allocated together for simplicity.
1366 	 */
1367 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1368 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1369 	fqsize= sc->fqdepth * 4;
1370 	pqsize = sc->pqdepth * 8;
1371 	qsize = fqsize + pqsize;
1372 
1373 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1374 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(qsize),
1375 	    BD_MAXSEGSIZE(qsize), BD_NSEGMENTS(1),
1376 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1377 	if (bus_dma_template_tag(&t, &sc->queues_dmat)) {
1378 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues DMA tag\n");
1379 		return (ENOMEM);
1380         }
1381         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1382 	    &sc->queues_map)) {
1383 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues memory\n");
1384 		return (ENOMEM);
1385         }
1386         bzero(queues, qsize);
1387         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1388 	    mpr_memaddr_cb, &queues_busaddr, 0);
1389 
1390 	sc->free_queue = (uint32_t *)queues;
1391 	sc->free_busaddr = queues_busaddr;
1392 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1393 	sc->post_busaddr = queues_busaddr + fqsize;
1394 	mpr_dprint(sc, MPR_INIT, "free queue busaddr= %#016jx size= %d\n",
1395 	    (uintmax_t)sc->free_busaddr, fqsize);
1396 	mpr_dprint(sc, MPR_INIT, "reply queue busaddr= %#016jx size= %d\n",
1397 	    (uintmax_t)sc->post_busaddr, pqsize);
1398 
1399 	return (0);
1400 }
1401 
1402 static int
1403 mpr_alloc_replies(struct mpr_softc *sc)
1404 {
1405 	bus_dma_template_t t;
1406 	int rsize, num_replies;
1407 
1408 	/* Store the reply frame size in bytes rather than as 32bit words */
1409 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1410 
1411 	/*
1412 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1413 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1414 	 * replies can be used at once.
1415 	 */
1416 	num_replies = max(sc->fqdepth, sc->num_replies);
1417 
1418 	rsize = sc->replyframesz * num_replies;
1419 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1420 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize),
1421 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1422 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1423 	if (bus_dma_template_tag(&t, &sc->reply_dmat)) {
1424 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies DMA tag\n");
1425 		return (ENOMEM);
1426         }
1427         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1428 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1429 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies memory\n");
1430 		return (ENOMEM);
1431         }
1432         bzero(sc->reply_frames, rsize);
1433         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1434 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1435 	mpr_dprint(sc, MPR_INIT, "reply frames busaddr= %#016jx size= %d\n",
1436 	    (uintmax_t)sc->reply_busaddr, rsize);
1437 
1438 	return (0);
1439 }
1440 
1441 static void
1442 mpr_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1443 {
1444 	struct mpr_softc *sc = arg;
1445 	struct mpr_chain *chain;
1446 	bus_size_t bo;
1447 	int i, o, s;
1448 
1449 	if (error != 0)
1450 		return;
1451 
1452 	for (i = 0, o = 0, s = 0; s < nsegs; s++) {
1453 		for (bo = 0; bo + sc->chain_frame_size <= segs[s].ds_len;
1454 		    bo += sc->chain_frame_size) {
1455 			chain = &sc->chains[i++];
1456 			chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o);
1457 			chain->chain_busaddr = segs[s].ds_addr + bo;
1458 			o += sc->chain_frame_size;
1459 			mpr_free_chain(sc, chain);
1460 		}
1461 		if (bo != segs[s].ds_len)
1462 			o += segs[s].ds_len - bo;
1463 	}
1464 	sc->chain_free_lowwater = i;
1465 }
1466 
1467 static int
1468 mpr_alloc_requests(struct mpr_softc *sc)
1469 {
1470 	bus_dma_template_t t;
1471 	struct mpr_command *cm;
1472 	int i, rsize, nsegs;
1473 
1474 	rsize = sc->reqframesz * sc->num_reqs;
1475 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1476 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize),
1477 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1478 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1479 	if (bus_dma_template_tag(&t, &sc->req_dmat)) {
1480 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request DMA tag\n");
1481 		return (ENOMEM);
1482         }
1483         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1484 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1485 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request memory\n");
1486 		return (ENOMEM);
1487         }
1488         bzero(sc->req_frames, rsize);
1489         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1490 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1491 	mpr_dprint(sc, MPR_INIT, "request frames busaddr= %#016jx size= %d\n",
1492 	    (uintmax_t)sc->req_busaddr, rsize);
1493 
1494 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->num_chains, M_MPR,
1495 	    M_NOWAIT | M_ZERO);
1496 	if (!sc->chains) {
1497 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1498 		return (ENOMEM);
1499 	}
1500 	rsize = sc->chain_frame_size * sc->num_chains;
1501 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1502 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize),
1503 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS((howmany(rsize, PAGE_SIZE))),
1504 	    BD_BOUNDARY(BUS_SPACE_MAXSIZE_32BIT+1));
1505 	if (bus_dma_template_tag(&t, &sc->chain_dmat)) {
1506 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain DMA tag\n");
1507 		return (ENOMEM);
1508 	}
1509 	if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1510 	    BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) {
1511 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1512 		return (ENOMEM);
1513 	}
1514 	if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames,
1515 	    rsize, mpr_load_chains_cb, sc, BUS_DMA_NOWAIT)) {
1516 		mpr_dprint(sc, MPR_ERROR, "Cannot load chain memory\n");
1517 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
1518 		    sc->chain_map);
1519 		return (ENOMEM);
1520 	}
1521 
1522 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1523 	bus_dma_template_clone(&t, sc->req_dmat);
1524 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(1), BD_MAXSIZE(rsize),
1525 	    BD_MAXSEGSIZE(rsize));
1526 	if (bus_dma_template_tag(&t, &sc->sense_dmat)) {
1527 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense DMA tag\n");
1528 		return (ENOMEM);
1529         }
1530         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1531 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1532 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense memory\n");
1533 		return (ENOMEM);
1534         }
1535         bzero(sc->sense_frames, rsize);
1536         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1537 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1538 	mpr_dprint(sc, MPR_INIT, "sense frames busaddr= %#016jx size= %d\n",
1539 	    (uintmax_t)sc->sense_busaddr, rsize);
1540 
1541 	/*
1542 	 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports
1543 	 * these devices.
1544 	 */
1545 	if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) &&
1546 	    (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) {
1547 		if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM)
1548 			return (ENOMEM);
1549 	}
1550 
1551 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1552 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1553 	BUS_DMA_TEMPLATE_FILL(&t, BD_MAXSIZE(BUS_SPACE_MAXSIZE_32BIT),
1554 	    BD_NSEGMENTS(nsegs), BD_MAXSEGSIZE(BUS_SPACE_MAXSIZE_32BIT),
1555 	    BD_FLAGS(BUS_DMA_ALLOCNOW), BD_LOCKFUNC(busdma_lock_mutex),
1556 	    BD_LOCKFUNCARG(&sc->mpr_mtx),
1557 	    BD_BOUNDARY(BUS_SPACE_MAXSIZE_32BIT+1));
1558 	if (bus_dma_template_tag(&t, &sc->buffer_dmat)) {
1559 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate buffer DMA tag\n");
1560 		return (ENOMEM);
1561         }
1562 
1563 	/*
1564 	 * SMID 0 cannot be used as a free command per the firmware spec.
1565 	 * Just drop that command instead of risking accounting bugs.
1566 	 */
1567 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1568 	    M_MPR, M_WAITOK | M_ZERO);
1569 	for (i = 1; i < sc->num_reqs; i++) {
1570 		cm = &sc->commands[i];
1571 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1572 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1573 		cm->cm_sense = &sc->sense_frames[i];
1574 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1575 		cm->cm_desc.Default.SMID = htole16(i);
1576 		cm->cm_sc = sc;
1577 		cm->cm_state = MPR_CM_STATE_BUSY;
1578 		TAILQ_INIT(&cm->cm_chain_list);
1579 		TAILQ_INIT(&cm->cm_prp_page_list);
1580 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1581 
1582 		/* XXX Is a failure here a critical problem? */
1583 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap)
1584 		    == 0) {
1585 			if (i <= sc->num_prireqs)
1586 				mpr_free_high_priority_command(sc, cm);
1587 			else
1588 				mpr_free_command(sc, cm);
1589 		} else {
1590 			panic("failed to allocate command %d\n", i);
1591 			sc->num_reqs = i;
1592 			break;
1593 		}
1594 	}
1595 
1596 	return (0);
1597 }
1598 
1599 /*
1600  * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs,
1601  * which are scatter/gather lists for NVMe devices.
1602  *
1603  * This buffer must be contiguous due to the nature of how NVMe PRPs are built
1604  * and translated by FW.
1605  *
1606  * returns ENOMEM if memory could not be allocated, otherwise returns 0.
1607  */
1608 static int
1609 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc)
1610 {
1611 	bus_dma_template_t t;
1612 	struct mpr_prp_page *prp_page;
1613 	int PRPs_per_page, PRPs_required, pages_required;
1614 	int rsize, i;
1615 
1616 	/*
1617 	 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number
1618 	 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is:
1619 	 * MAX_IO_SIZE / PAGE_SIZE = 256
1620 	 *
1621 	 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs
1622 	 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one
1623 	 * page (4096 / 8 = 512), so only one page is required for each I/O.
1624 	 *
1625 	 * Each of these buffers will need to be contiguous. For simplicity,
1626 	 * only one buffer is allocated here, which has all of the space
1627 	 * required for the NVMe Queue Depth. If there are problems allocating
1628 	 * this one buffer, this function will need to change to allocate
1629 	 * individual, contiguous NVME_QDEPTH buffers.
1630 	 *
1631 	 * The real calculation will use the real max io size. Above is just an
1632 	 * example.
1633 	 *
1634 	 */
1635 	PRPs_required = sc->maxio / PAGE_SIZE;
1636 	PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1;
1637 	pages_required = (PRPs_required / PRPs_per_page) + 1;
1638 
1639 	sc->prp_buffer_size = PAGE_SIZE * pages_required;
1640 	rsize = sc->prp_buffer_size * NVME_QDEPTH;
1641 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1642 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize),
1643 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1644 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1645 	if (bus_dma_template_tag(&t, &sc->prp_page_dmat)) {
1646 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP DMA "
1647 		    "tag\n");
1648 		return (ENOMEM);
1649 	}
1650 	if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages,
1651 	    BUS_DMA_NOWAIT, &sc->prp_page_map)) {
1652 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP memory\n");
1653 		return (ENOMEM);
1654 	}
1655 	bzero(sc->prp_pages, rsize);
1656 	bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages,
1657 	    rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0);
1658 
1659 	sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR,
1660 	    M_WAITOK | M_ZERO);
1661 	for (i = 0; i < NVME_QDEPTH; i++) {
1662 		prp_page = &sc->prps[i];
1663 		prp_page->prp_page = (uint64_t *)(sc->prp_pages +
1664 		    i * sc->prp_buffer_size);
1665 		prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr +
1666 		    i * sc->prp_buffer_size);
1667 		mpr_free_prp_page(sc, prp_page);
1668 		sc->prp_pages_free_lowwater++;
1669 	}
1670 
1671 	return (0);
1672 }
1673 
1674 static int
1675 mpr_init_queues(struct mpr_softc *sc)
1676 {
1677 	int i;
1678 
1679 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1680 
1681 	/*
1682 	 * According to the spec, we need to use one less reply than we
1683 	 * have space for on the queue.  So sc->num_replies (the number we
1684 	 * use) should be less than sc->fqdepth (allocated size).
1685 	 */
1686 	if (sc->num_replies >= sc->fqdepth)
1687 		return (EINVAL);
1688 
1689 	/*
1690 	 * Initialize all of the free queue entries.
1691 	 */
1692 	for (i = 0; i < sc->fqdepth; i++) {
1693 		sc->free_queue[i] = htole32(sc->reply_busaddr + (i * sc->replyframesz));
1694 	}
1695 	sc->replyfreeindex = sc->num_replies;
1696 
1697 	return (0);
1698 }
1699 
1700 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1701  * Next are the global settings, if they exist.  Highest are the per-unit
1702  * settings, if they exist.
1703  */
1704 void
1705 mpr_get_tunables(struct mpr_softc *sc)
1706 {
1707 	char tmpstr[80], mpr_debug[80];
1708 
1709 	/* XXX default to some debugging for now */
1710 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1711 	sc->disable_msix = 0;
1712 	sc->disable_msi = 0;
1713 	sc->max_msix = MPR_MSIX_MAX;
1714 	sc->max_chains = MPR_CHAIN_FRAMES;
1715 	sc->max_io_pages = MPR_MAXIO_PAGES;
1716 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1717 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1718 	sc->use_phynum = 1;
1719 	sc->max_reqframes = MPR_REQ_FRAMES;
1720 	sc->max_prireqframes = MPR_PRI_REQ_FRAMES;
1721 	sc->max_replyframes = MPR_REPLY_FRAMES;
1722 	sc->max_evtframes = MPR_EVT_REPLY_FRAMES;
1723 
1724 	/*
1725 	 * Grab the global variables.
1726 	 */
1727 	bzero(mpr_debug, 80);
1728 	if (TUNABLE_STR_FETCH("hw.mpr.debug_level", mpr_debug, 80) != 0)
1729 		mpr_parse_debug(sc, mpr_debug);
1730 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1731 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1732 	TUNABLE_INT_FETCH("hw.mpr.max_msix", &sc->max_msix);
1733 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1734 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1735 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1736 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1737 	TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum);
1738 	TUNABLE_INT_FETCH("hw.mpr.max_reqframes", &sc->max_reqframes);
1739 	TUNABLE_INT_FETCH("hw.mpr.max_prireqframes", &sc->max_prireqframes);
1740 	TUNABLE_INT_FETCH("hw.mpr.max_replyframes", &sc->max_replyframes);
1741 	TUNABLE_INT_FETCH("hw.mpr.max_evtframes", &sc->max_evtframes);
1742 
1743 	/* Grab the unit-instance variables */
1744 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1745 	    device_get_unit(sc->mpr_dev));
1746 	bzero(mpr_debug, 80);
1747 	if (TUNABLE_STR_FETCH(tmpstr, mpr_debug, 80) != 0)
1748 		mpr_parse_debug(sc, mpr_debug);
1749 
1750 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1751 	    device_get_unit(sc->mpr_dev));
1752 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1753 
1754 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1755 	    device_get_unit(sc->mpr_dev));
1756 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1757 
1758 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_msix",
1759 	    device_get_unit(sc->mpr_dev));
1760 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1761 
1762 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1763 	    device_get_unit(sc->mpr_dev));
1764 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1765 
1766 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1767 	    device_get_unit(sc->mpr_dev));
1768 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1769 
1770 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1771 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1772 	    device_get_unit(sc->mpr_dev));
1773 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1774 
1775 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1776 	    device_get_unit(sc->mpr_dev));
1777 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1778 
1779 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1780 	    device_get_unit(sc->mpr_dev));
1781 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1782 
1783 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num",
1784 	    device_get_unit(sc->mpr_dev));
1785 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1786 
1787 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_reqframes",
1788 	    device_get_unit(sc->mpr_dev));
1789 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1790 
1791 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_prireqframes",
1792 	    device_get_unit(sc->mpr_dev));
1793 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1794 
1795 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_replyframes",
1796 	    device_get_unit(sc->mpr_dev));
1797 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1798 
1799 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_evtframes",
1800 	    device_get_unit(sc->mpr_dev));
1801 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1802 }
1803 
1804 static void
1805 mpr_setup_sysctl(struct mpr_softc *sc)
1806 {
1807 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1808 	struct sysctl_oid	*sysctl_tree = NULL;
1809 	char tmpstr[80], tmpstr2[80];
1810 
1811 	/*
1812 	 * Setup the sysctl variable so the user can change the debug level
1813 	 * on the fly.
1814 	 */
1815 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1816 	    device_get_unit(sc->mpr_dev));
1817 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1818 
1819 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1820 	if (sysctl_ctx != NULL)
1821 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1822 
1823 	if (sysctl_tree == NULL) {
1824 		sysctl_ctx_init(&sc->sysctl_ctx);
1825 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1826 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1827 		    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr);
1828 		if (sc->sysctl_tree == NULL)
1829 			return;
1830 		sysctl_ctx = &sc->sysctl_ctx;
1831 		sysctl_tree = sc->sysctl_tree;
1832 	}
1833 
1834 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1835 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1836 	    sc, 0, mpr_debug_sysctl, "A", "mpr debug level");
1837 
1838 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1839 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1840 	    "Disable the use of MSI-X interrupts");
1841 
1842 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1843 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1844 	    "User-defined maximum number of MSIX queues");
1845 
1846 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1847 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1848 	    "Negotiated number of MSIX queues");
1849 
1850 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1851 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1852 	    "Total number of allocated request frames");
1853 
1854 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1855 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1856 	    "Total number of allocated high priority request frames");
1857 
1858 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1859 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1860 	    "Total number of allocated reply frames");
1861 
1862 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1863 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1864 	    "Total number of event frames allocated");
1865 
1866 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1867 	    OID_AUTO, "firmware_version", CTLFLAG_RD, sc->fw_version,
1868 	    strlen(sc->fw_version), "firmware version");
1869 
1870 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1871 	    OID_AUTO, "driver_version", CTLFLAG_RD, MPR_DRIVER_VERSION,
1872 	    strlen(MPR_DRIVER_VERSION), "driver version");
1873 
1874 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1875 	    OID_AUTO, "msg_version", CTLFLAG_RD, sc->msg_version,
1876 	    strlen(sc->msg_version), "message interface version (deprecated)");
1877 
1878 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1879 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1880 	    &sc->io_cmds_active, 0, "number of currently active commands");
1881 
1882 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1883 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1884 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1885 
1886 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1887 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1888 	    &sc->chain_free, 0, "number of free chain elements");
1889 
1890 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1891 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1892 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1893 
1894 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1895 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1896 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1897 
1898 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1899 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1900 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1901 	    "IOCFacts)");
1902 
1903 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1904 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1905 	    "enable SSU to SATA SSD/HDD at shutdown");
1906 
1907 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1908 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1909 	    &sc->chain_alloc_fail, "chain allocation failures");
1910 
1911 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1912 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1913 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1914 	    "spinup after SATA ID error");
1915 
1916 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1917 	    OID_AUTO, "dump_reqs",
1918 	    CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE,
1919 	    sc, 0, mpr_dump_reqs, "I", "Dump Active Requests");
1920 
1921 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1922 	    OID_AUTO, "dump_reqs_alltypes", CTLFLAG_RW,
1923 	    &sc->dump_reqs_alltypes, 0,
1924 	    "dump all request types not just inqueue");
1925 
1926 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1927 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1928 	    "Use the phy number for enumeration");
1929 
1930 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1931 	    OID_AUTO, "prp_pages_free", CTLFLAG_RD,
1932 	    &sc->prp_pages_free, 0, "number of free PRP pages");
1933 
1934 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1935 	    OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD,
1936 	    &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages");
1937 
1938 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1939 	    OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD,
1940 	    &sc->prp_page_alloc_fail, "PRP page allocation failures");
1941 }
1942 
1943 static struct mpr_debug_string {
1944 	char *name;
1945 	int flag;
1946 } mpr_debug_strings[] = {
1947 	{"info", MPR_INFO},
1948 	{"fault", MPR_FAULT},
1949 	{"event", MPR_EVENT},
1950 	{"log", MPR_LOG},
1951 	{"recovery", MPR_RECOVERY},
1952 	{"error", MPR_ERROR},
1953 	{"init", MPR_INIT},
1954 	{"xinfo", MPR_XINFO},
1955 	{"user", MPR_USER},
1956 	{"mapping", MPR_MAPPING},
1957 	{"trace", MPR_TRACE}
1958 };
1959 
1960 enum mpr_debug_level_combiner {
1961 	COMB_NONE,
1962 	COMB_ADD,
1963 	COMB_SUB
1964 };
1965 
1966 static int
1967 mpr_debug_sysctl(SYSCTL_HANDLER_ARGS)
1968 {
1969 	struct mpr_softc *sc;
1970 	struct mpr_debug_string *string;
1971 	struct sbuf *sbuf;
1972 	char *buffer;
1973 	size_t sz;
1974 	int i, len, debug, error;
1975 
1976 	sc = (struct mpr_softc *)arg1;
1977 
1978 	error = sysctl_wire_old_buffer(req, 0);
1979 	if (error != 0)
1980 		return (error);
1981 
1982 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1983 	debug = sc->mpr_debug;
1984 
1985 	sbuf_printf(sbuf, "%#x", debug);
1986 
1987 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1988 	for (i = 0; i < sz; i++) {
1989 		string = &mpr_debug_strings[i];
1990 		if (debug & string->flag)
1991 			sbuf_printf(sbuf, ",%s", string->name);
1992 	}
1993 
1994 	error = sbuf_finish(sbuf);
1995 	sbuf_delete(sbuf);
1996 
1997 	if (error || req->newptr == NULL)
1998 		return (error);
1999 
2000 	len = req->newlen - req->newidx;
2001 	if (len == 0)
2002 		return (0);
2003 
2004 	buffer = malloc(len, M_MPR, M_ZERO|M_WAITOK);
2005 	error = SYSCTL_IN(req, buffer, len);
2006 
2007 	mpr_parse_debug(sc, buffer);
2008 
2009 	free(buffer, M_MPR);
2010 	return (error);
2011 }
2012 
2013 static void
2014 mpr_parse_debug(struct mpr_softc *sc, char *list)
2015 {
2016 	struct mpr_debug_string *string;
2017 	enum mpr_debug_level_combiner op;
2018 	char *token, *endtoken;
2019 	size_t sz;
2020 	int flags, i;
2021 
2022 	if (list == NULL || *list == '\0')
2023 		return;
2024 
2025 	if (*list == '+') {
2026 		op = COMB_ADD;
2027 		list++;
2028 	} else if (*list == '-') {
2029 		op = COMB_SUB;
2030 		list++;
2031 	} else
2032 		op = COMB_NONE;
2033 	if (*list == '\0')
2034 		return;
2035 
2036 	flags = 0;
2037 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
2038 	while ((token = strsep(&list, ":,")) != NULL) {
2039 		/* Handle integer flags */
2040 		flags |= strtol(token, &endtoken, 0);
2041 		if (token != endtoken)
2042 			continue;
2043 
2044 		/* Handle text flags */
2045 		for (i = 0; i < sz; i++) {
2046 			string = &mpr_debug_strings[i];
2047 			if (strcasecmp(token, string->name) == 0) {
2048 				flags |= string->flag;
2049 				break;
2050 			}
2051 		}
2052 	}
2053 
2054 	switch (op) {
2055 	case COMB_NONE:
2056 		sc->mpr_debug = flags;
2057 		break;
2058 	case COMB_ADD:
2059 		sc->mpr_debug |= flags;
2060 		break;
2061 	case COMB_SUB:
2062 		sc->mpr_debug &= (~flags);
2063 		break;
2064 	}
2065 	return;
2066 }
2067 
2068 struct mpr_dumpreq_hdr {
2069 	uint32_t	smid;
2070 	uint32_t	state;
2071 	uint32_t	numframes;
2072 	uint32_t	deschi;
2073 	uint32_t	desclo;
2074 };
2075 
2076 static int
2077 mpr_dump_reqs(SYSCTL_HANDLER_ARGS)
2078 {
2079 	struct mpr_softc *sc;
2080 	struct mpr_chain *chain, *chain1;
2081 	struct mpr_command *cm;
2082 	struct mpr_dumpreq_hdr hdr;
2083 	struct sbuf *sb;
2084 	uint32_t smid, state;
2085 	int i, numreqs, error = 0;
2086 
2087 	sc = (struct mpr_softc *)arg1;
2088 
2089 	if ((error = priv_check(curthread, PRIV_DRIVER)) != 0) {
2090 		printf("priv check error %d\n", error);
2091 		return (error);
2092 	}
2093 
2094 	state = MPR_CM_STATE_INQUEUE;
2095 	smid = 1;
2096 	numreqs = sc->num_reqs;
2097 
2098 	if (req->newptr != NULL)
2099 		return (EINVAL);
2100 
2101 	if (smid == 0 || smid > sc->num_reqs)
2102 		return (EINVAL);
2103 	if (numreqs <= 0 || (numreqs + smid > sc->num_reqs))
2104 		numreqs = sc->num_reqs;
2105 	sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req);
2106 
2107 	/* Best effort, no locking */
2108 	for (i = smid; i < numreqs; i++) {
2109 		cm = &sc->commands[i];
2110 		if ((sc->dump_reqs_alltypes == 0) && (cm->cm_state != state))
2111 			continue;
2112 		hdr.smid = i;
2113 		hdr.state = cm->cm_state;
2114 		hdr.numframes = 1;
2115 		hdr.deschi = cm->cm_desc.Words.High;
2116 		hdr.desclo = cm->cm_desc.Words.Low;
2117 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
2118 		   chain1)
2119 			hdr.numframes++;
2120 		sbuf_bcat(sb, &hdr, sizeof(hdr));
2121 		sbuf_bcat(sb, cm->cm_req, 128);
2122 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
2123 		    chain1)
2124 			sbuf_bcat(sb, chain->chain, 128);
2125 	}
2126 
2127 	error = sbuf_finish(sb);
2128 	sbuf_delete(sb);
2129 	return (error);
2130 }
2131 
2132 int
2133 mpr_attach(struct mpr_softc *sc)
2134 {
2135 	int error;
2136 
2137 	MPR_FUNCTRACE(sc);
2138 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2139 
2140 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
2141 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
2142 	callout_init_mtx(&sc->device_check_callout, &sc->mpr_mtx, 0);
2143 	TAILQ_INIT(&sc->event_list);
2144 	timevalclear(&sc->lastfail);
2145 
2146 	if ((error = mpr_transition_ready(sc)) != 0) {
2147 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2148 		    "Failed to transition ready\n");
2149 		return (error);
2150 	}
2151 
2152 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
2153 	    M_ZERO|M_NOWAIT);
2154 	if (!sc->facts) {
2155 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2156 		    "Cannot allocate memory, exit\n");
2157 		return (ENOMEM);
2158 	}
2159 
2160 	/*
2161 	 * Get IOC Facts and allocate all structures based on this information.
2162 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
2163 	 * Facts. If relevant values have changed in IOC Facts, this function
2164 	 * will free all of the memory based on IOC Facts and reallocate that
2165 	 * memory.  If this fails, any allocated memory should already be freed.
2166 	 */
2167 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
2168 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC Facts allocation "
2169 		    "failed with error %d\n", error);
2170 		return (error);
2171 	}
2172 
2173 	/* Start the periodic watchdog check on the IOC Doorbell */
2174 	mpr_periodic(sc);
2175 
2176 	/*
2177 	 * The portenable will kick off discovery events that will drive the
2178 	 * rest of the initialization process.  The CAM/SAS module will
2179 	 * hold up the boot sequence until discovery is complete.
2180 	 */
2181 	sc->mpr_ich.ich_func = mpr_startup;
2182 	sc->mpr_ich.ich_arg = sc;
2183 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
2184 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2185 		    "Cannot establish MPR config hook\n");
2186 		error = EINVAL;
2187 	}
2188 
2189 	/*
2190 	 * Allow IR to shutdown gracefully when shutdown occurs.
2191 	 */
2192 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
2193 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
2194 
2195 	if (sc->shutdown_eh == NULL)
2196 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2197 		    "shutdown event registration failed\n");
2198 
2199 	mpr_setup_sysctl(sc);
2200 
2201 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
2202 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
2203 
2204 	return (error);
2205 }
2206 
2207 /* Run through any late-start handlers. */
2208 static void
2209 mpr_startup(void *arg)
2210 {
2211 	struct mpr_softc *sc;
2212 
2213 	sc = (struct mpr_softc *)arg;
2214 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2215 
2216 	mpr_lock(sc);
2217 	mpr_unmask_intr(sc);
2218 
2219 	/* initialize device mapping tables */
2220 	mpr_base_static_config_pages(sc);
2221 	mpr_mapping_initialize(sc);
2222 	mprsas_startup(sc);
2223 	mpr_unlock(sc);
2224 
2225 	mpr_dprint(sc, MPR_INIT, "disestablish config intrhook\n");
2226 	config_intrhook_disestablish(&sc->mpr_ich);
2227 	sc->mpr_ich.ich_arg = NULL;
2228 
2229 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2230 }
2231 
2232 /* Periodic watchdog.  Is called with the driver lock already held. */
2233 static void
2234 mpr_periodic(void *arg)
2235 {
2236 	struct mpr_softc *sc;
2237 	uint32_t db;
2238 
2239 	sc = (struct mpr_softc *)arg;
2240 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
2241 		return;
2242 
2243 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
2244 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2245 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
2246 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
2247 			panic("TEMPERATURE FAULT: STOPPING.");
2248 		}
2249 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2250 		mpr_reinit(sc);
2251 	}
2252 
2253 	callout_reset_sbt(&sc->periodic, MPR_PERIODIC_DELAY * SBT_1S, 0,
2254 	    mpr_periodic, sc, C_PREL(1));
2255 }
2256 
2257 static void
2258 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
2259     MPI2_EVENT_NOTIFICATION_REPLY *event)
2260 {
2261 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2262 
2263 	MPR_DPRINT_EVENT(sc, generic, event);
2264 
2265 	switch (event->Event) {
2266 	case MPI2_EVENT_LOG_DATA:
2267 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2268 		if (sc->mpr_debug & MPR_EVENT)
2269 			hexdump(event->EventData, event->EventDataLength, NULL,
2270 			    0);
2271 		break;
2272 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2273 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2274 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2275 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2276 		     entry->LogSequence);
2277 		break;
2278 	default:
2279 		break;
2280 	}
2281 	return;
2282 }
2283 
2284 static int
2285 mpr_attach_log(struct mpr_softc *sc)
2286 {
2287 	uint8_t events[16];
2288 
2289 	bzero(events, 16);
2290 	setbit(events, MPI2_EVENT_LOG_DATA);
2291 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2292 
2293 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
2294 	    &sc->mpr_log_eh);
2295 
2296 	return (0);
2297 }
2298 
2299 static int
2300 mpr_detach_log(struct mpr_softc *sc)
2301 {
2302 
2303 	if (sc->mpr_log_eh != NULL)
2304 		mpr_deregister_events(sc, sc->mpr_log_eh);
2305 	return (0);
2306 }
2307 
2308 /*
2309  * Free all of the driver resources and detach submodules.  Should be called
2310  * without the lock held.
2311  */
2312 int
2313 mpr_free(struct mpr_softc *sc)
2314 {
2315 	int error;
2316 
2317 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2318 	/* Turn off the watchdog */
2319 	mpr_lock(sc);
2320 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
2321 	mpr_unlock(sc);
2322 	/* Lock must not be held for this */
2323 	callout_drain(&sc->periodic);
2324 	callout_drain(&sc->device_check_callout);
2325 
2326 	if (((error = mpr_detach_log(sc)) != 0) ||
2327 	    ((error = mpr_detach_sas(sc)) != 0)) {
2328 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "failed to detach "
2329 		    "subsystems, error= %d, exit\n", error);
2330 		return (error);
2331 	}
2332 
2333 	mpr_detach_user(sc);
2334 
2335 	/* Put the IOC back in the READY state. */
2336 	mpr_lock(sc);
2337 	if ((error = mpr_transition_ready(sc)) != 0) {
2338 		mpr_unlock(sc);
2339 		return (error);
2340 	}
2341 	mpr_unlock(sc);
2342 
2343 	if (sc->facts != NULL)
2344 		free(sc->facts, M_MPR);
2345 
2346 	/*
2347 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2348 	 * to free these buffers too.
2349 	 */
2350 	mpr_iocfacts_free(sc);
2351 
2352 	if (sc->sysctl_tree != NULL)
2353 		sysctl_ctx_free(&sc->sysctl_ctx);
2354 
2355 	/* Deregister the shutdown function */
2356 	if (sc->shutdown_eh != NULL)
2357 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2358 
2359 	mtx_destroy(&sc->mpr_mtx);
2360 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2361 
2362 	return (0);
2363 }
2364 
2365 static __inline void
2366 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
2367 {
2368 	MPR_FUNCTRACE(sc);
2369 
2370 	if (cm == NULL) {
2371 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
2372 		return;
2373 	}
2374 
2375 	KASSERT(cm->cm_state == MPR_CM_STATE_INQUEUE,
2376 	    ("command not inqueue, state = %u\n", cm->cm_state));
2377 	cm->cm_state = MPR_CM_STATE_BUSY;
2378 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
2379 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
2380 
2381 	if (cm->cm_complete != NULL) {
2382 		mpr_dprint(sc, MPR_TRACE,
2383 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
2384 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
2385 		    cm->cm_reply);
2386 		cm->cm_complete(sc, cm);
2387 	}
2388 
2389 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
2390 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
2391 		wakeup(cm);
2392 	}
2393 
2394 	if (sc->io_cmds_active != 0) {
2395 		sc->io_cmds_active--;
2396 	} else {
2397 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
2398 		    "out of sync - resynching to 0\n");
2399 	}
2400 }
2401 
2402 static void
2403 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
2404 {
2405 	union loginfo_type {
2406 		u32	loginfo;
2407 		struct {
2408 			u32	subcode:16;
2409 			u32	code:8;
2410 			u32	originator:4;
2411 			u32	bus_type:4;
2412 		} dw;
2413 	};
2414 	union loginfo_type sas_loginfo;
2415 	char *originator_str = NULL;
2416 
2417 	sas_loginfo.loginfo = log_info;
2418 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2419 		return;
2420 
2421 	/* each nexus loss loginfo */
2422 	if (log_info == 0x31170000)
2423 		return;
2424 
2425 	/* eat the loginfos associated with task aborts */
2426 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
2427 	    (log_info == 0x31130000))
2428 		return;
2429 
2430 	switch (sas_loginfo.dw.originator) {
2431 	case 0:
2432 		originator_str = "IOP";
2433 		break;
2434 	case 1:
2435 		originator_str = "PL";
2436 		break;
2437 	case 2:
2438 		originator_str = "IR";
2439 		break;
2440 	}
2441 
2442 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
2443 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
2444 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
2445 }
2446 
2447 static void
2448 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
2449 {
2450 	MPI2DefaultReply_t *mpi_reply;
2451 	u16 sc_status;
2452 
2453 	mpi_reply = (MPI2DefaultReply_t*)reply;
2454 	sc_status = le16toh(mpi_reply->IOCStatus);
2455 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2456 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2457 }
2458 
2459 void
2460 mpr_intr(void *data)
2461 {
2462 	struct mpr_softc *sc;
2463 	uint32_t status;
2464 
2465 	sc = (struct mpr_softc *)data;
2466 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2467 
2468 	/*
2469 	 * Check interrupt status register to flush the bus.  This is
2470 	 * needed for both INTx interrupts and driver-driven polling
2471 	 */
2472 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2473 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2474 		return;
2475 
2476 	mpr_lock(sc);
2477 	mpr_intr_locked(data);
2478 	mpr_unlock(sc);
2479 	return;
2480 }
2481 
2482 /*
2483  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2484  * chip.  Hopefully this theory is correct.
2485  */
2486 void
2487 mpr_intr_msi(void *data)
2488 {
2489 	struct mpr_softc *sc;
2490 
2491 	sc = (struct mpr_softc *)data;
2492 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2493 	mpr_lock(sc);
2494 	mpr_intr_locked(data);
2495 	mpr_unlock(sc);
2496 	return;
2497 }
2498 
2499 /*
2500  * The locking is overly broad and simplistic, but easy to deal with for now.
2501  */
2502 void
2503 mpr_intr_locked(void *data)
2504 {
2505 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2506 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2507 	mpr_fw_diagnostic_buffer_t *pBuffer;
2508 	struct mpr_softc *sc;
2509 	uint64_t tdesc;
2510 	struct mpr_command *cm = NULL;
2511 	uint8_t flags;
2512 	u_int pq;
2513 
2514 	sc = (struct mpr_softc *)data;
2515 
2516 	pq = sc->replypostindex;
2517 	mpr_dprint(sc, MPR_TRACE,
2518 	    "%s sc %p starting with replypostindex %u\n",
2519 	    __func__, sc, sc->replypostindex);
2520 
2521 	for ( ;; ) {
2522 		cm = NULL;
2523 		desc = &sc->post_queue[sc->replypostindex];
2524 
2525 		/*
2526 		 * Copy and clear out the descriptor so that any reentry will
2527 		 * immediately know that this descriptor has already been
2528 		 * looked at.  There is unfortunate casting magic because the
2529 		 * MPI API doesn't have a cardinal 64bit type.
2530 		 */
2531 		tdesc = 0xffffffffffffffff;
2532 		tdesc = atomic_swap_64((uint64_t *)desc, tdesc);
2533 		desc = (MPI2_REPLY_DESCRIPTORS_UNION *)&tdesc;
2534 
2535 		flags = desc->Default.ReplyFlags &
2536 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2537 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
2538 		    (le32toh(desc->Words.High) == 0xffffffff))
2539 			break;
2540 
2541 		/* increment the replypostindex now, so that event handlers
2542 		 * and cm completion handlers which decide to do a diag
2543 		 * reset can zero it without it getting incremented again
2544 		 * afterwards, and we break out of this loop on the next
2545 		 * iteration since the reply post queue has been cleared to
2546 		 * 0xFF and all descriptors look unused (which they are).
2547 		 */
2548 		if (++sc->replypostindex >= sc->pqdepth)
2549 			sc->replypostindex = 0;
2550 
2551 		switch (flags) {
2552 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2553 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
2554 		case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS:
2555 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2556 			cm->cm_reply = NULL;
2557 			break;
2558 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2559 		{
2560 			uint32_t baddr;
2561 			uint8_t *reply;
2562 
2563 			/*
2564 			 * Re-compose the reply address from the address
2565 			 * sent back from the chip.  The ReplyFrameAddress
2566 			 * is the lower 32 bits of the physical address of
2567 			 * particular reply frame.  Convert that address to
2568 			 * host format, and then use that to provide the
2569 			 * offset against the virtual address base
2570 			 * (sc->reply_frames).
2571 			 */
2572 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2573 			reply = sc->reply_frames +
2574 				(baddr - ((uint32_t)sc->reply_busaddr));
2575 			/*
2576 			 * Make sure the reply we got back is in a valid
2577 			 * range.  If not, go ahead and panic here, since
2578 			 * we'll probably panic as soon as we deference the
2579 			 * reply pointer anyway.
2580 			 */
2581 			if ((reply < sc->reply_frames)
2582 			 || (reply > (sc->reply_frames +
2583 			     (sc->fqdepth * sc->replyframesz)))) {
2584 				printf("%s: WARNING: reply %p out of range!\n",
2585 				       __func__, reply);
2586 				printf("%s: reply_frames %p, fqdepth %d, "
2587 				       "frame size %d\n", __func__,
2588 				       sc->reply_frames, sc->fqdepth,
2589 				       sc->replyframesz);
2590 				printf("%s: baddr %#x,\n", __func__, baddr);
2591 				/* LSI-TODO. See Linux Code for Graceful exit */
2592 				panic("Reply address out of range");
2593 			}
2594 			if (le16toh(desc->AddressReply.SMID) == 0) {
2595 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2596 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2597 					/*
2598 					 * If SMID is 0 for Diag Buffer Post,
2599 					 * this implies that the reply is due to
2600 					 * a release function with a status that
2601 					 * the buffer has been released.  Set
2602 					 * the buffer flags accordingly.
2603 					 */
2604 					rel_rep =
2605 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2606 					if ((le16toh(rel_rep->IOCStatus) &
2607 					    MPI2_IOCSTATUS_MASK) ==
2608 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2609 					{
2610 						pBuffer =
2611 						    &sc->fw_diag_buffer_list[
2612 						    rel_rep->BufferType];
2613 						pBuffer->valid_data = TRUE;
2614 						pBuffer->owned_by_firmware =
2615 						    FALSE;
2616 						pBuffer->immediate = FALSE;
2617 					}
2618 				} else
2619 					mpr_dispatch_event(sc, baddr,
2620 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2621 					    reply);
2622 			} else {
2623 				cm = &sc->commands[
2624 				    le16toh(desc->AddressReply.SMID)];
2625 				if (cm->cm_state == MPR_CM_STATE_INQUEUE) {
2626 					cm->cm_reply = reply;
2627 					cm->cm_reply_data =
2628 					    le32toh(desc->AddressReply.
2629 						ReplyFrameAddress);
2630 				} else {
2631 					mpr_dprint(sc, MPR_RECOVERY,
2632 					    "Bad state for ADDRESS_REPLY status,"
2633 					    " ignoring state %d cm %p\n",
2634 					    cm->cm_state, cm);
2635 				}
2636 			}
2637 			break;
2638 		}
2639 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2640 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2641 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2642 		default:
2643 			/* Unhandled */
2644 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
2645 			    desc->Default.ReplyFlags);
2646 			cm = NULL;
2647 			break;
2648 		}
2649 
2650 		if (cm != NULL) {
2651 			// Print Error reply frame
2652 			if (cm->cm_reply)
2653 				mpr_display_reply_info(sc,cm->cm_reply);
2654 			mpr_complete_command(sc, cm);
2655 		}
2656 	}
2657 
2658 	if (pq != sc->replypostindex) {
2659 		mpr_dprint(sc, MPR_TRACE, "%s sc %p writing postindex %d\n",
2660 		    __func__, sc, sc->replypostindex);
2661 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2662 		    sc->replypostindex);
2663 	}
2664 
2665 	return;
2666 }
2667 
2668 static void
2669 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2670     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2671 {
2672 	struct mpr_event_handle *eh;
2673 	int event, handled = 0;
2674 
2675 	event = le16toh(reply->Event);
2676 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2677 		if (isset(eh->mask, event)) {
2678 			eh->callback(sc, data, reply);
2679 			handled++;
2680 		}
2681 	}
2682 
2683 	if (handled == 0)
2684 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2685 		    le16toh(event));
2686 
2687 	/*
2688 	 * This is the only place that the event/reply should be freed.
2689 	 * Anything wanting to hold onto the event data should have
2690 	 * already copied it into their own storage.
2691 	 */
2692 	mpr_free_reply(sc, data);
2693 }
2694 
2695 static void
2696 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2697 {
2698 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2699 
2700 	if (cm->cm_reply)
2701 		MPR_DPRINT_EVENT(sc, generic,
2702 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2703 
2704 	mpr_free_command(sc, cm);
2705 
2706 	/* next, send a port enable */
2707 	mprsas_startup(sc);
2708 }
2709 
2710 /*
2711  * For both register_events and update_events, the caller supplies a bitmap
2712  * of events that it _wants_.  These functions then turn that into a bitmask
2713  * suitable for the controller.
2714  */
2715 int
2716 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2717     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2718 {
2719 	struct mpr_event_handle *eh;
2720 	int error = 0;
2721 
2722 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2723 	eh->callback = cb;
2724 	eh->data = data;
2725 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2726 	if (mask != NULL)
2727 		error = mpr_update_events(sc, eh, mask);
2728 	*handle = eh;
2729 
2730 	return (error);
2731 }
2732 
2733 int
2734 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2735     uint8_t *mask)
2736 {
2737 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2738 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2739 	struct mpr_command *cm = NULL;
2740 	struct mpr_event_handle *eh;
2741 	int error, i;
2742 
2743 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2744 
2745 	if ((mask != NULL) && (handle != NULL))
2746 		bcopy(mask, &handle->mask[0], 16);
2747 	memset(sc->event_mask, 0xff, 16);
2748 
2749 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2750 		for (i = 0; i < 16; i++)
2751 			sc->event_mask[i] &= ~eh->mask[i];
2752 	}
2753 
2754 	if ((cm = mpr_alloc_command(sc)) == NULL)
2755 		return (EBUSY);
2756 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2757 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2758 	evtreq->MsgFlags = 0;
2759 	evtreq->SASBroadcastPrimitiveMasks = 0;
2760 #ifdef MPR_DEBUG_ALL_EVENTS
2761 	{
2762 		u_char fullmask[sizeof(evtreq->EventMasks)];
2763 		memset(fullmask, 0x00, sizeof(fullmask));
2764 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, sizeof(fullmask));
2765 	}
2766 #else
2767 	bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, sizeof(sc->event_mask));
2768 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2769 		evtreq->EventMasks[i] = htole32(evtreq->EventMasks[i]);
2770 #endif
2771 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2772 	cm->cm_data = NULL;
2773 
2774 	error = mpr_request_polled(sc, &cm);
2775 	if (cm != NULL)
2776 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2777 	if ((reply == NULL) ||
2778 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2779 		error = ENXIO;
2780 
2781 	if (reply)
2782 		MPR_DPRINT_EVENT(sc, generic, reply);
2783 
2784 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2785 
2786 	if (cm != NULL)
2787 		mpr_free_command(sc, cm);
2788 	return (error);
2789 }
2790 
2791 static int
2792 mpr_reregister_events(struct mpr_softc *sc)
2793 {
2794 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2795 	struct mpr_command *cm;
2796 	struct mpr_event_handle *eh;
2797 	int error, i;
2798 
2799 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2800 
2801 	/* first, reregister events */
2802 
2803 	memset(sc->event_mask, 0xff, 16);
2804 
2805 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2806 		for (i = 0; i < 16; i++)
2807 			sc->event_mask[i] &= ~eh->mask[i];
2808 	}
2809 
2810 	if ((cm = mpr_alloc_command(sc)) == NULL)
2811 		return (EBUSY);
2812 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2813 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2814 	evtreq->MsgFlags = 0;
2815 	evtreq->SASBroadcastPrimitiveMasks = 0;
2816 #ifdef MPR_DEBUG_ALL_EVENTS
2817 	{
2818 		u_char fullmask[sizeof(evtreq->EventMasks)];
2819 		memset(fullmask, 0x00, sizeof(fullmask));
2820 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, sizeof(fullmask));
2821 	}
2822 #else
2823 	bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, sizeof(sc->event_mask));
2824 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
2825 		evtreq->EventMasks[i] = htole32(evtreq->EventMasks[i]);
2826 #endif
2827 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2828 	cm->cm_data = NULL;
2829 	cm->cm_complete = mpr_reregister_events_complete;
2830 
2831 	error = mpr_map_command(sc, cm);
2832 
2833 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2834 	    error);
2835 	return (error);
2836 }
2837 
2838 int
2839 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2840 {
2841 
2842 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2843 	free(handle, M_MPR);
2844 	return (mpr_update_events(sc, NULL, NULL));
2845 }
2846 
2847 /**
2848 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a
2849 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry
2850 * of the NVMe message (PRP1). If the data buffer is small enough to be described
2851 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to
2852 * describe a larger data buffer. If the data buffer is too large to describe
2853 * using the two PRP entriess inside the NVMe message, then PRP1 describes the
2854 * first data memory segment, and PRP2 contains a pointer to a PRP list located
2855 * elsewhere in memory to describe the remaining data memory segments. The PRP
2856 * list will be contiguous.
2857 
2858 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP
2859 * consists of a list of PRP entries to describe a number of noncontigous
2860 * physical memory segments as a single memory buffer, just as a SGL does. Note
2861 * however, that this function is only used by the IOCTL call, so the memory
2862 * given will be guaranteed to be contiguous. There is no need to translate
2863 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous
2864 * space that is one page size each.
2865 *
2866 * Each NVMe message contains two PRP entries. The first (PRP1) either contains
2867 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains
2868 * the second PRP element if the memory being described fits within 2 PRP
2869 * entries, or a PRP list pointer if the PRP spans more than two entries.
2870 *
2871 * A PRP list pointer contains the address of a PRP list, structured as a linear
2872 * array of PRP entries. Each PRP entry in this list describes a segment of
2873 * physical memory.
2874 *
2875 * Each 64-bit PRP entry comprises an address and an offset field. The address
2876 * always points to the beginning of a PAGE_SIZE physical memory page, and the
2877 * offset describes where within that page the memory segment begins. Only the
2878 * first element in a PRP list may contain a non-zero offest, implying that all
2879 * memory segments following the first begin at the start of a PAGE_SIZE page.
2880 *
2881 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory,
2882 * with exceptions for the first and last elements in the list. If the memory
2883 * being described by the list begins at a non-zero offset within the first page,
2884 * then the first PRP element will contain a non-zero offset indicating where the
2885 * region begins within the page. The last memory segment may end before the end
2886 * of the PAGE_SIZE segment, depending upon the overall size of the memory being
2887 * described by the PRP list.
2888 *
2889 * Since PRP entries lack any indication of size, the overall data buffer length
2890 * is used to determine where the end of the data memory buffer is located, and
2891 * how many PRP entries are required to describe it.
2892 *
2893 * Returns nothing.
2894 */
2895 void
2896 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm,
2897     Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data,
2898     uint32_t data_in_sz, uint32_t data_out_sz)
2899 {
2900 	int			prp_size = PRP_ENTRY_SIZE;
2901 	uint64_t		*prp_entry, *prp1_entry, *prp2_entry;
2902 	uint64_t		*prp_entry_phys, *prp_page, *prp_page_phys;
2903 	uint32_t		offset, entry_len, page_mask_result, page_mask;
2904 	bus_addr_t		paddr;
2905 	size_t			length;
2906 	struct mpr_prp_page	*prp_page_info = NULL;
2907 
2908 	/*
2909 	 * Not all commands require a data transfer. If no data, just return
2910 	 * without constructing any PRP.
2911 	 */
2912 	if (!data_in_sz && !data_out_sz)
2913 		return;
2914 
2915 	/*
2916 	 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is
2917 	 * located at a 24 byte offset from the start of the NVMe command. Then
2918 	 * set the current PRP entry pointer to PRP1.
2919 	 */
2920 	prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2921 	    NVME_CMD_PRP1_OFFSET);
2922 	prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2923 	    NVME_CMD_PRP2_OFFSET);
2924 	prp_entry = prp1_entry;
2925 
2926 	/*
2927 	 * For the PRP entries, use the specially allocated buffer of
2928 	 * contiguous memory. PRP Page allocation failures should not happen
2929 	 * because there should be enough PRP page buffers to account for the
2930 	 * possible NVMe QDepth.
2931 	 */
2932 	prp_page_info = mpr_alloc_prp_page(sc);
2933 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2934 	    "used for building a native NVMe SGL.\n", __func__));
2935 	prp_page = (uint64_t *)prp_page_info->prp_page;
2936 	prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2937 
2938 	/*
2939 	 * Insert the allocated PRP page into the command's PRP page list. This
2940 	 * will be freed when the command is freed.
2941 	 */
2942 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2943 
2944 	/*
2945 	 * Check if we are within 1 entry of a page boundary we don't want our
2946 	 * first entry to be a PRP List entry.
2947 	 */
2948 	page_mask = PAGE_SIZE - 1;
2949 	page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) &
2950 	    page_mask;
2951 	if (!page_mask_result)
2952 	{
2953 		/* Bump up to next page boundary. */
2954 		prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size);
2955 		prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys +
2956 		    prp_size);
2957 	}
2958 
2959 	/*
2960 	 * Set PRP physical pointer, which initially points to the current PRP
2961 	 * DMA memory page.
2962 	 */
2963 	prp_entry_phys = prp_page_phys;
2964 
2965 	/* Get physical address and length of the data buffer. */
2966 	paddr = (bus_addr_t)(uintptr_t)data;
2967 	if (data_in_sz)
2968 		length = data_in_sz;
2969 	else
2970 		length = data_out_sz;
2971 
2972 	/* Loop while the length is not zero. */
2973 	while (length)
2974 	{
2975 		/*
2976 		 * Check if we need to put a list pointer here if we are at page
2977 		 * boundary - prp_size (8 bytes).
2978 		 */
2979 		page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys +
2980 		    prp_size) & page_mask;
2981 		if (!page_mask_result)
2982 		{
2983 			/*
2984 			 * This is the last entry in a PRP List, so we need to
2985 			 * put a PRP list pointer here. What this does is:
2986 			 *   - bump the current memory pointer to the next
2987 			 *     address, which will be the next full page.
2988 			 *   - set the PRP Entry to point to that page. This is
2989 			 *     now the PRP List pointer.
2990 			 *   - bump the PRP Entry pointer the start of the next
2991 			 *     page. Since all of this PRP memory is contiguous,
2992 			 *     no need to get a new page - it's just the next
2993 			 *     address.
2994 			 */
2995 			prp_entry_phys++;
2996 			*prp_entry =
2997 			    htole64((uint64_t)(uintptr_t)prp_entry_phys);
2998 			prp_entry++;
2999 		}
3000 
3001 		/* Need to handle if entry will be part of a page. */
3002 		offset = (uint32_t)paddr & page_mask;
3003 		entry_len = PAGE_SIZE - offset;
3004 
3005 		if (prp_entry == prp1_entry)
3006 		{
3007 			/*
3008 			 * Must fill in the first PRP pointer (PRP1) before
3009 			 * moving on.
3010 			 */
3011 			*prp1_entry = htole64((uint64_t)paddr);
3012 
3013 			/*
3014 			 * Now point to the second PRP entry within the
3015 			 * command (PRP2).
3016 			 */
3017 			prp_entry = prp2_entry;
3018 		}
3019 		else if (prp_entry == prp2_entry)
3020 		{
3021 			/*
3022 			 * Should the PRP2 entry be a PRP List pointer or just a
3023 			 * regular PRP pointer? If there is more than one more
3024 			 * page of data, must use a PRP List pointer.
3025 			 */
3026 			if (length > PAGE_SIZE)
3027 			{
3028 				/*
3029 				 * PRP2 will contain a PRP List pointer because
3030 				 * more PRP's are needed with this command. The
3031 				 * list will start at the beginning of the
3032 				 * contiguous buffer.
3033 				 */
3034 				*prp2_entry =
3035 				    htole64(
3036 				    (uint64_t)(uintptr_t)prp_entry_phys);
3037 
3038 				/*
3039 				 * The next PRP Entry will be the start of the
3040 				 * first PRP List.
3041 				 */
3042 				prp_entry = prp_page;
3043 			}
3044 			else
3045 			{
3046 				/*
3047 				 * After this, the PRP Entries are complete.
3048 				 * This command uses 2 PRP's and no PRP list.
3049 				 */
3050 				*prp2_entry = htole64((uint64_t)paddr);
3051 			}
3052 		}
3053 		else
3054 		{
3055 			/*
3056 			 * Put entry in list and bump the addresses.
3057 			 *
3058 			 * After PRP1 and PRP2 are filled in, this will fill in
3059 			 * all remaining PRP entries in a PRP List, one per each
3060 			 * time through the loop.
3061 			 */
3062 			*prp_entry = htole64((uint64_t)paddr);
3063 			prp_entry++;
3064 			prp_entry_phys++;
3065 		}
3066 
3067 		/*
3068 		 * Bump the phys address of the command's data buffer by the
3069 		 * entry_len.
3070 		 */
3071 		paddr += entry_len;
3072 
3073 		/* Decrement length accounting for last partial page. */
3074 		if (entry_len > length)
3075 			length = 0;
3076 		else
3077 			length -= entry_len;
3078 	}
3079 }
3080 
3081 /*
3082  * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to
3083  * determine if the driver needs to build a native SGL. If so, that native SGL
3084  * is built in the contiguous buffers allocated especially for PCIe SGL
3085  * creation. If the driver will not build a native SGL, return TRUE and a
3086  * normal IEEE SGL will be built. Currently this routine supports NVMe devices
3087  * only.
3088  *
3089  * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built.
3090  */
3091 static int
3092 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm,
3093     bus_dma_segment_t *segs, int segs_left)
3094 {
3095 	uint32_t		i, sge_dwords, length, offset, entry_len;
3096 	uint32_t		num_entries, buff_len = 0, sges_in_segment;
3097 	uint32_t		page_mask, page_mask_result, *curr_buff;
3098 	uint32_t		*ptr_sgl, *ptr_first_sgl, first_page_offset;
3099 	uint32_t		first_page_data_size, end_residual;
3100 	uint64_t		*msg_phys;
3101 	bus_addr_t		paddr;
3102 	int			build_native_sgl = 0, first_prp_entry;
3103 	int			prp_size = PRP_ENTRY_SIZE;
3104 	Mpi25IeeeSgeChain64_t	*main_chain_element = NULL;
3105 	struct mpr_prp_page	*prp_page_info = NULL;
3106 
3107 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
3108 
3109 	/*
3110 	 * Add up the sizes of each segment length to get the total transfer
3111 	 * size, which will be checked against the Maximum Data Transfer Size.
3112 	 * If the data transfer length exceeds the MDTS for this device, just
3113 	 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O
3114 	 * up into multiple I/O's. [nvme_mdts = 0 means unlimited]
3115 	 */
3116 	for (i = 0; i < segs_left; i++)
3117 		buff_len += htole32(segs[i].ds_len);
3118 	if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS))
3119 		return 1;
3120 
3121 	/* Create page_mask (to get offset within page) */
3122 	page_mask = PAGE_SIZE - 1;
3123 
3124 	/*
3125 	 * Check if the number of elements exceeds the max number that can be
3126 	 * put in the main message frame (H/W can only translate an SGL that
3127 	 * is contained entirely in the main message frame).
3128 	 */
3129 	sges_in_segment = (sc->reqframesz -
3130 	    offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION);
3131 	if (segs_left > sges_in_segment)
3132 		build_native_sgl = 1;
3133 	else
3134 	{
3135 		/*
3136 		 * NVMe uses one PRP for each physical page (or part of physical
3137 		 * page).
3138 		 *    if 4 pages or less then IEEE is OK
3139 		 *    if > 5 pages then we need to build a native SGL
3140 		 *    if > 4 and <= 5 pages, then check the physical address of
3141 		 *      the first SG entry, then if this first size in the page
3142 		 *      is >= the residual beyond 4 pages then use IEEE,
3143 		 *      otherwise use native SGL
3144 		 */
3145 		if (buff_len > (PAGE_SIZE * 5))
3146 			build_native_sgl = 1;
3147 		else if ((buff_len > (PAGE_SIZE * 4)) &&
3148 		    (buff_len <= (PAGE_SIZE * 5)) )
3149 		{
3150 			msg_phys = (uint64_t *)(uintptr_t)segs[0].ds_addr;
3151 			first_page_offset =
3152 			    ((uint32_t)(uint64_t)(uintptr_t)msg_phys &
3153 			    page_mask);
3154 			first_page_data_size = PAGE_SIZE - first_page_offset;
3155 			end_residual = buff_len % PAGE_SIZE;
3156 
3157 			/*
3158 			 * If offset into first page pushes the end of the data
3159 			 * beyond end of the 5th page, we need the extra PRP
3160 			 * list.
3161 			 */
3162 			if (first_page_data_size < end_residual)
3163 				build_native_sgl = 1;
3164 
3165 			/*
3166 			 * Check if first SG entry size is < residual beyond 4
3167 			 * pages.
3168 			 */
3169 			if (htole32(segs[0].ds_len) <
3170 			    (buff_len - (PAGE_SIZE * 4)))
3171 				build_native_sgl = 1;
3172 		}
3173 	}
3174 
3175 	/* check if native SGL is needed */
3176 	if (!build_native_sgl)
3177 		return 1;
3178 
3179 	/*
3180 	 * Native SGL is needed.
3181 	 * Put a chain element in main message frame that points to the first
3182 	 * chain buffer.
3183 	 *
3184 	 * NOTE:  The ChainOffset field must be 0 when using a chain pointer to
3185 	 *        a native SGL.
3186 	 */
3187 
3188 	/* Set main message chain element pointer */
3189 	main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge;
3190 
3191 	/*
3192 	 * For NVMe the chain element needs to be the 2nd SGL entry in the main
3193 	 * message.
3194 	 */
3195 	main_chain_element = (Mpi25IeeeSgeChain64_t *)
3196 	    ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64));
3197 
3198 	/*
3199 	 * For the PRP entries, use the specially allocated buffer of
3200 	 * contiguous memory. PRP Page allocation failures should not happen
3201 	 * because there should be enough PRP page buffers to account for the
3202 	 * possible NVMe QDepth.
3203 	 */
3204 	prp_page_info = mpr_alloc_prp_page(sc);
3205 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
3206 	    "used for building a native NVMe SGL.\n", __func__));
3207 	curr_buff = (uint32_t *)prp_page_info->prp_page;
3208 	msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
3209 
3210 	/*
3211 	 * Insert the allocated PRP page into the command's PRP page list. This
3212 	 * will be freed when the command is freed.
3213 	 */
3214 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
3215 
3216 	/*
3217 	 * Check if we are within 1 entry of a page boundary we don't want our
3218 	 * first entry to be a PRP List entry.
3219 	 */
3220 	page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) &
3221 	    page_mask;
3222 	if (!page_mask_result) {
3223 		/* Bump up to next page boundary. */
3224 		curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size);
3225 		msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size);
3226 	}
3227 
3228 	/* Fill in the chain element and make it an NVMe segment type. */
3229 	main_chain_element->Address.High =
3230 	    htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32));
3231 	main_chain_element->Address.Low =
3232 	    htole32((uint32_t)(uintptr_t)msg_phys);
3233 	main_chain_element->NextChainOffset = 0;
3234 	main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3235 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3236 	    MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP;
3237 
3238 	/* Set SGL pointer to start of contiguous PCIe buffer. */
3239 	ptr_sgl = curr_buff;
3240 	sge_dwords = 2;
3241 	num_entries = 0;
3242 
3243 	/*
3244 	 * NVMe has a very convoluted PRP format. One PRP is required for each
3245 	 * page or partial page. We need to split up OS SG entries if they are
3246 	 * longer than one page or cross a page boundary. We also have to insert
3247 	 * a PRP list pointer entry as the last entry in each physical page of
3248 	 * the PRP list.
3249 	 *
3250 	 * NOTE: The first PRP "entry" is actually placed in the first SGL entry
3251 	 * in the main message in IEEE 64 format. The 2nd entry in the main
3252 	 * message is the chain element, and the rest of the PRP entries are
3253 	 * built in the contiguous PCIe buffer.
3254 	 */
3255 	first_prp_entry = 1;
3256 	ptr_first_sgl = (uint32_t *)cm->cm_sge;
3257 
3258 	for (i = 0; i < segs_left; i++) {
3259 		/* Get physical address and length of this SG entry. */
3260 		paddr = segs[i].ds_addr;
3261 		length = segs[i].ds_len;
3262 
3263 		/*
3264 		 * Check whether a given SGE buffer lies on a non-PAGED
3265 		 * boundary if this is not the first page. If so, this is not
3266 		 * expected so have FW build the SGL.
3267 		 */
3268 		if ((i != 0) && (((uint32_t)paddr & page_mask) != 0)) {
3269 			mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while "
3270 			    "building NVMe PRPs, low address is 0x%x\n",
3271 			    (uint32_t)paddr);
3272 			return 1;
3273 		}
3274 
3275 		/* Apart from last SGE, if any other SGE boundary is not page
3276 		 * aligned then it means that hole exists. Existence of hole
3277 		 * leads to data corruption. So fallback to IEEE SGEs.
3278 		 */
3279 		if (i != (segs_left - 1)) {
3280 			if (((uint32_t)paddr + length) & page_mask) {
3281 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE "
3282 				    "boundary while building NVMe PRPs, low "
3283 				    "address: 0x%x and length: %u\n",
3284 				    (uint32_t)paddr, length);
3285 				return 1;
3286 			}
3287 		}
3288 
3289 		/* Loop while the length is not zero. */
3290 		while (length) {
3291 			/*
3292 			 * Check if we need to put a list pointer here if we are
3293 			 * at page boundary - prp_size.
3294 			 */
3295 			page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl +
3296 			    prp_size) & page_mask;
3297 			if (!page_mask_result) {
3298 				/*
3299 				 * Need to put a PRP list pointer here.
3300 				 */
3301 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3302 				    prp_size);
3303 				*ptr_sgl = htole32((uintptr_t)msg_phys);
3304 				*(ptr_sgl+1) = htole32((uint64_t)(uintptr_t)
3305 				    msg_phys >> 32);
3306 				ptr_sgl += sge_dwords;
3307 				num_entries++;
3308 			}
3309 
3310 			/* Need to handle if entry will be part of a page. */
3311 			offset = (uint32_t)paddr & page_mask;
3312 			entry_len = PAGE_SIZE - offset;
3313 			if (first_prp_entry) {
3314 				/*
3315 				 * Put IEEE entry in first SGE in main message.
3316 				 * (Simple element, System addr, not end of
3317 				 * list.)
3318 				 */
3319 				*ptr_first_sgl = htole32((uint32_t)paddr);
3320 				*(ptr_first_sgl + 1) =
3321 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3322 				*(ptr_first_sgl + 2) = htole32(entry_len);
3323 				*(ptr_first_sgl + 3) = 0;
3324 
3325 				/* No longer the first PRP entry. */
3326 				first_prp_entry = 0;
3327 			} else {
3328 				/* Put entry in list. */
3329 				*ptr_sgl = htole32((uint32_t)paddr);
3330 				*(ptr_sgl + 1) =
3331 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3332 
3333 				/* Bump ptr_sgl, msg_phys, and num_entries. */
3334 				ptr_sgl += sge_dwords;
3335 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3336 				    prp_size);
3337 				num_entries++;
3338 			}
3339 
3340 			/* Bump the phys address by the entry_len. */
3341 			paddr += entry_len;
3342 
3343 			/* Decrement length accounting for last partial page. */
3344 			if (entry_len > length)
3345 				length = 0;
3346 			else
3347 				length -= entry_len;
3348 		}
3349 	}
3350 
3351 	/* Set chain element Length. */
3352 	main_chain_element->Length = htole32(num_entries * prp_size);
3353 
3354 	/* Return 0, indicating we built a native SGL. */
3355 	return 0;
3356 }
3357 
3358 /*
3359  * Add a chain element as the next SGE for the specified command.
3360  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
3361  * only required for IEEE commands.  Therefore there is no code for commands
3362  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
3363  * shouldn't be requesting chains).
3364  */
3365 static int
3366 mpr_add_chain(struct mpr_command *cm, int segsleft)
3367 {
3368 	struct mpr_softc *sc = cm->cm_sc;
3369 	MPI2_REQUEST_HEADER *req;
3370 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
3371 	struct mpr_chain *chain;
3372 	int sgc_size, current_segs, rem_segs, segs_per_frame;
3373 	uint8_t next_chain_offset = 0;
3374 
3375 	/*
3376 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
3377 	 * only IEEE commands should be requesting chains.  Return some error
3378 	 * code other than 0.
3379 	 */
3380 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
3381 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
3382 		    "an MPI SGL.\n");
3383 		return(ENOBUFS);
3384 	}
3385 
3386 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
3387 	if (cm->cm_sglsize < sgc_size)
3388 		panic("MPR: Need SGE Error Code\n");
3389 
3390 	chain = mpr_alloc_chain(cm->cm_sc);
3391 	if (chain == NULL)
3392 		return (ENOBUFS);
3393 
3394 	/*
3395 	 * Note: a double-linked list is used to make it easier to walk for
3396 	 * debugging.
3397 	 */
3398 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
3399 
3400 	/*
3401 	 * Need to know if the number of frames left is more than 1 or not.  If
3402 	 * more than 1 frame is required, NextChainOffset will need to be set,
3403 	 * which will just be the last segment of the frame.
3404 	 */
3405 	rem_segs = 0;
3406 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
3407 		/*
3408 		 * rem_segs is the number of segment remaining after the
3409 		 * segments that will go into the current frame.  Since it is
3410 		 * known that at least one more frame is required, account for
3411 		 * the chain element.  To know if more than one more frame is
3412 		 * required, just check if there will be a remainder after using
3413 		 * the current frame (with this chain) and the next frame.  If
3414 		 * so the NextChainOffset must be the last element of the next
3415 		 * frame.
3416 		 */
3417 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
3418 		rem_segs = segsleft - current_segs;
3419 		segs_per_frame = sc->chain_frame_size / sgc_size;
3420 		if (rem_segs > segs_per_frame) {
3421 			next_chain_offset = segs_per_frame - 1;
3422 		}
3423 	}
3424 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
3425 	ieee_sgc->Length = next_chain_offset ?
3426 	    htole32((uint32_t)sc->chain_frame_size) :
3427 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
3428 	ieee_sgc->NextChainOffset = next_chain_offset;
3429 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3430 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3431 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
3432 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
3433 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
3434 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
3435 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
3436 
3437 	cm->cm_sglsize = sc->chain_frame_size;
3438 	return (0);
3439 }
3440 
3441 /*
3442  * Add one scatter-gather element to the scatter-gather list for a command.
3443  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
3444  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
3445  * chain, so don't consider any chain additions.
3446  */
3447 int
3448 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
3449     int segsleft)
3450 {
3451 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3452 	u32 sge_flags;
3453 
3454 	/*
3455 	 * case 1: >=1 more segment, no room for anything (error)
3456 	 * case 2: 1 more segment and enough room for it
3457          */
3458 
3459 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
3460 		mpr_dprint(cm->cm_sc, MPR_ERROR,
3461 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
3462 		    __func__);
3463 		return(ENOBUFS);
3464 	}
3465 
3466 	KASSERT(segsleft == 1,
3467 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
3468 	    segsleft));
3469 
3470 	/*
3471 	 * There is one more segment left to add for the MPI SGL and there is
3472 	 * enough room in the frame to add it.  This is the normal case because
3473 	 * MPI SGL's don't have chains, otherwise something is wrong.
3474 	 *
3475 	 * If this is a bi-directional request, need to account for that
3476 	 * here.  Save the pre-filled sge values.  These will be used
3477 	 * either for the 2nd SGL or for a single direction SGL.  If
3478 	 * cm_out_len is non-zero, this is a bi-directional request, so
3479 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
3480 	 * fill in the IN SGL.  Note that at this time, when filling in
3481 	 * 2 SGL's for a bi-directional request, they both use the same
3482 	 * DMA buffer (same cm command).
3483 	 */
3484 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
3485 	saved_address_low = sge->Address.Low;
3486 	saved_address_high = sge->Address.High;
3487 	if (cm->cm_out_len) {
3488 		sge->FlagsLength = cm->cm_out_len |
3489 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3490 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
3491 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
3492 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3493 		    MPI2_SGE_FLAGS_SHIFT);
3494 		cm->cm_sglsize -= len;
3495 		/* Endian Safe code */
3496 		sge_flags = sge->FlagsLength;
3497 		sge->FlagsLength = htole32(sge_flags);
3498 		bcopy(sge, cm->cm_sge, len);
3499 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3500 	}
3501 	sge->FlagsLength = saved_buf_len |
3502 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3503 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
3504 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
3505 	    MPI2_SGE_FLAGS_END_OF_LIST |
3506 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3507 	    MPI2_SGE_FLAGS_SHIFT);
3508 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
3509 		sge->FlagsLength |=
3510 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
3511 		    MPI2_SGE_FLAGS_SHIFT);
3512 	} else {
3513 		sge->FlagsLength |=
3514 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
3515 		    MPI2_SGE_FLAGS_SHIFT);
3516 	}
3517 	sge->Address.Low = saved_address_low;
3518 	sge->Address.High = saved_address_high;
3519 
3520 	cm->cm_sglsize -= len;
3521 	/* Endian Safe code */
3522 	sge_flags = sge->FlagsLength;
3523 	sge->FlagsLength = htole32(sge_flags);
3524 	bcopy(sge, cm->cm_sge, len);
3525 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3526 	return (0);
3527 }
3528 
3529 /*
3530  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
3531  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
3532  * remaining size and pointer to the next SGE to fill in, respectively.
3533  */
3534 int
3535 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
3536 {
3537 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
3538 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
3539 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3540 	uint32_t sge_length;
3541 
3542 	/*
3543 	 * case 1: No room for chain or segment (error).
3544 	 * case 2: Two or more segments left but only room for chain.
3545 	 * case 3: Last segment and room for it, so set flags.
3546 	 */
3547 
3548 	/*
3549 	 * There should be room for at least one element, or there is a big
3550 	 * problem.
3551 	 */
3552 	if (cm->cm_sglsize < ieee_sge_size)
3553 		panic("MPR: Need SGE Error Code\n");
3554 
3555 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
3556 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
3557 			return (error);
3558 	}
3559 
3560 	if (segsleft == 1) {
3561 		/*
3562 		 * If this is a bi-directional request, need to account for that
3563 		 * here.  Save the pre-filled sge values.  These will be used
3564 		 * either for the 2nd SGL or for a single direction SGL.  If
3565 		 * cm_out_len is non-zero, this is a bi-directional request, so
3566 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
3567 		 * fill in the IN SGL.  Note that at this time, when filling in
3568 		 * 2 SGL's for a bi-directional request, they both use the same
3569 		 * DMA buffer (same cm command).
3570 		 */
3571 		saved_buf_len = sge->Length;
3572 		saved_address_low = sge->Address.Low;
3573 		saved_address_high = sge->Address.High;
3574 		if (cm->cm_out_len) {
3575 			sge->Length = cm->cm_out_len;
3576 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3577 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3578 			cm->cm_sglsize -= ieee_sge_size;
3579 			/* Endian Safe code */
3580 			sge_length = sge->Length;
3581 			sge->Length = htole32(sge_length);
3582 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
3583 			cm->cm_sge =
3584 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3585 			    ieee_sge_size);
3586 		}
3587 		sge->Length = saved_buf_len;
3588 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3589 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3590 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
3591 		sge->Address.Low = saved_address_low;
3592 		sge->Address.High = saved_address_high;
3593 	}
3594 
3595 	cm->cm_sglsize -= ieee_sge_size;
3596 	/* Endian Safe code */
3597 	sge_length = sge->Length;
3598 	sge->Length = htole32(sge_length);
3599 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
3600 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3601 	    ieee_sge_size);
3602 	return (0);
3603 }
3604 
3605 /*
3606  * Add one dma segment to the scatter-gather list for a command.
3607  */
3608 int
3609 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
3610     int segsleft)
3611 {
3612 	MPI2_SGE_SIMPLE64 sge;
3613 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
3614 
3615 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
3616 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3617 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3618 		ieee_sge.Length = len;
3619 		mpr_from_u64(pa, &ieee_sge.Address);
3620 
3621 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
3622 	} else {
3623 		/*
3624 		 * This driver always uses 64-bit address elements for
3625 		 * simplicity.
3626 		 */
3627 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3628 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
3629 		/* Set Endian safe macro in mpr_push_sge */
3630 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
3631 		mpr_from_u64(pa, &sge.Address);
3632 
3633 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
3634 	}
3635 }
3636 
3637 static void
3638 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
3639 {
3640 	struct mpr_softc *sc;
3641 	struct mpr_command *cm;
3642 	u_int i, dir, sflags;
3643 
3644 	cm = (struct mpr_command *)arg;
3645 	sc = cm->cm_sc;
3646 
3647 	/*
3648 	 * In this case, just print out a warning and let the chip tell the
3649 	 * user they did the wrong thing.
3650 	 */
3651 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
3652 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
3653 		    "segments, more than the %d allowed\n", __func__, nsegs,
3654 		    cm->cm_max_segs);
3655 	}
3656 
3657 	/*
3658 	 * Set up DMA direction flags.  Bi-directional requests are also handled
3659 	 * here.  In that case, both direction flags will be set.
3660 	 */
3661 	sflags = 0;
3662 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
3663 		/*
3664 		 * We have to add a special case for SMP passthrough, there
3665 		 * is no easy way to generically handle it.  The first
3666 		 * S/G element is used for the command (therefore the
3667 		 * direction bit needs to be set).  The second one is used
3668 		 * for the reply.  We'll leave it to the caller to make
3669 		 * sure we only have two buffers.
3670 		 */
3671 		/*
3672 		 * Even though the busdma man page says it doesn't make
3673 		 * sense to have both direction flags, it does in this case.
3674 		 * We have one s/g element being accessed in each direction.
3675 		 */
3676 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
3677 
3678 		/*
3679 		 * Set the direction flag on the first buffer in the SMP
3680 		 * passthrough request.  We'll clear it for the second one.
3681 		 */
3682 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
3683 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
3684 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
3685 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
3686 		dir = BUS_DMASYNC_PREWRITE;
3687 	} else
3688 		dir = BUS_DMASYNC_PREREAD;
3689 
3690 	/* Check if a native SG list is needed for an NVMe PCIe device. */
3691 	if (cm->cm_targ && cm->cm_targ->is_nvme &&
3692 	    mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) {
3693 		/* A native SG list was built, skip to end. */
3694 		goto out;
3695 	}
3696 
3697 	for (i = 0; i < nsegs; i++) {
3698 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
3699 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3700 		}
3701 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3702 		    sflags, nsegs - i);
3703 		if (error != 0) {
3704 			/* Resource shortage, roll back! */
3705 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
3706 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
3707 				    "consider increasing hw.mpr.max_chains.\n");
3708 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
3709 			/*
3710 			 * mpr_complete_command can only be called on commands
3711 			 * that are in the queue. Since this is an error path
3712 			 * which gets called before we enqueue, update the state
3713 			 * to meet this requirement before we complete it.
3714 			 */
3715 			cm->cm_state = MPR_CM_STATE_INQUEUE;
3716 			mpr_complete_command(sc, cm);
3717 			return;
3718 		}
3719 	}
3720 
3721 out:
3722 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3723 	mpr_enqueue_request(sc, cm);
3724 
3725 	return;
3726 }
3727 
3728 static void
3729 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3730 	     int error)
3731 {
3732 	mpr_data_cb(arg, segs, nsegs, error);
3733 }
3734 
3735 /*
3736  * This is the routine to enqueue commands ansynchronously.
3737  * Note that the only error path here is from bus_dmamap_load(), which can
3738  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3739  * assumed that if you have a command in-hand, then you have enough credits
3740  * to use it.
3741  */
3742 int
3743 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
3744 {
3745 	int error = 0;
3746 
3747 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
3748 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3749 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
3750 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
3751 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3752 		    cm->cm_data, mpr_data_cb, cm, 0);
3753 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3754 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3755 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
3756 	} else {
3757 		/* Add a zero-length element as needed */
3758 		if (cm->cm_sge != NULL)
3759 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
3760 		mpr_enqueue_request(sc, cm);
3761 	}
3762 
3763 	return (error);
3764 }
3765 
3766 /*
3767  * This is the routine to enqueue commands synchronously.  An error of
3768  * EINPROGRESS from mpr_map_command() is ignored since the command will
3769  * be executed and enqueued automatically.  Other errors come from msleep().
3770  */
3771 int
3772 mpr_wait_command(struct mpr_softc *sc, struct mpr_command **cmp, int timeout,
3773     int sleep_flag)
3774 {
3775 	int error, rc;
3776 	struct timeval cur_time, start_time;
3777 	struct mpr_command *cm = *cmp;
3778 
3779 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
3780 		return  EBUSY;
3781 
3782 	cm->cm_complete = NULL;
3783 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
3784 	error = mpr_map_command(sc, cm);
3785 	if ((error != 0) && (error != EINPROGRESS))
3786 		return (error);
3787 
3788 	// Check for context and wait for 50 mSec at a time until time has
3789 	// expired or the command has finished.  If msleep can't be used, need
3790 	// to poll.
3791 	if (curthread->td_no_sleeping)
3792 		sleep_flag = NO_SLEEP;
3793 	getmicrouptime(&start_time);
3794 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
3795 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
3796 		if (error == EWOULDBLOCK) {
3797 			/*
3798 			 * Record the actual elapsed time in the case of a
3799 			 * timeout for the message below.
3800 			 */
3801 			getmicrouptime(&cur_time);
3802 			timevalsub(&cur_time, &start_time);
3803 		}
3804 	} else {
3805 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3806 			mpr_intr_locked(sc);
3807 			if (sleep_flag == CAN_SLEEP)
3808 				pause("mprwait", hz/20);
3809 			else
3810 				DELAY(50000);
3811 
3812 			getmicrouptime(&cur_time);
3813 			timevalsub(&cur_time, &start_time);
3814 			if (cur_time.tv_sec > timeout) {
3815 				error = EWOULDBLOCK;
3816 				break;
3817 			}
3818 		}
3819 	}
3820 
3821 	if (error == EWOULDBLOCK) {
3822 		if (cm->cm_timeout_handler == NULL) {
3823 			mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s, timeout=%d,"
3824 			    " elapsed=%jd\n", __func__, timeout,
3825 			    (intmax_t)cur_time.tv_sec);
3826 			rc = mpr_reinit(sc);
3827 			mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3828 			    "failed");
3829 		} else
3830 			cm->cm_timeout_handler(sc, cm);
3831 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3832 			/*
3833 			 * Tell the caller that we freed the command in a
3834 			 * reinit.
3835 			 */
3836 			*cmp = NULL;
3837 		}
3838 		error = ETIMEDOUT;
3839 	}
3840 	return (error);
3841 }
3842 
3843 /*
3844  * This is the routine to enqueue a command synchonously and poll for
3845  * completion.  Its use should be rare.
3846  */
3847 int
3848 mpr_request_polled(struct mpr_softc *sc, struct mpr_command **cmp)
3849 {
3850 	int error, rc;
3851 	struct timeval cur_time, start_time;
3852 	struct mpr_command *cm = *cmp;
3853 
3854 	error = 0;
3855 
3856 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
3857 	cm->cm_complete = NULL;
3858 	mpr_map_command(sc, cm);
3859 
3860 	getmicrouptime(&start_time);
3861 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3862 		mpr_intr_locked(sc);
3863 
3864 		if (mtx_owned(&sc->mpr_mtx))
3865 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
3866 			    "mprpoll", hz/20);
3867 		else
3868 			pause("mprpoll", hz/20);
3869 
3870 		/*
3871 		 * Check for real-time timeout and fail if more than 60 seconds.
3872 		 */
3873 		getmicrouptime(&cur_time);
3874 		timevalsub(&cur_time, &start_time);
3875 		if (cur_time.tv_sec > 60) {
3876 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
3877 			error = ETIMEDOUT;
3878 			break;
3879 		}
3880 	}
3881 	cm->cm_state = MPR_CM_STATE_BUSY;
3882 	if (error) {
3883 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3884 		rc = mpr_reinit(sc);
3885 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3886 		    "failed");
3887 
3888 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3889 			/*
3890 			 * Tell the caller that we freed the command in a
3891 			 * reinit.
3892 			 */
3893 			*cmp = NULL;
3894 		}
3895 	}
3896 	return (error);
3897 }
3898 
3899 /*
3900  * The MPT driver had a verbose interface for config pages.  In this driver,
3901  * reduce it to much simpler terms, similar to the Linux driver.
3902  */
3903 int
3904 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3905 {
3906 	MPI2_CONFIG_REQUEST *req;
3907 	struct mpr_command *cm;
3908 	int error;
3909 
3910 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
3911 		return (EBUSY);
3912 	}
3913 
3914 	cm = mpr_alloc_command(sc);
3915 	if (cm == NULL) {
3916 		return (EBUSY);
3917 	}
3918 
3919 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3920 	req->Function = MPI2_FUNCTION_CONFIG;
3921 	req->Action = params->action;
3922 	req->SGLFlags = 0;
3923 	req->ChainOffset = 0;
3924 	req->PageAddress = params->page_address;
3925 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3926 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3927 
3928 		hdr = &params->hdr.Ext;
3929 		req->ExtPageType = hdr->ExtPageType;
3930 		req->ExtPageLength = hdr->ExtPageLength;
3931 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3932 		req->Header.PageLength = 0; /* Must be set to zero */
3933 		req->Header.PageNumber = hdr->PageNumber;
3934 		req->Header.PageVersion = hdr->PageVersion;
3935 	} else {
3936 		MPI2_CONFIG_PAGE_HEADER *hdr;
3937 
3938 		hdr = &params->hdr.Struct;
3939 		req->Header.PageType = hdr->PageType;
3940 		req->Header.PageNumber = hdr->PageNumber;
3941 		req->Header.PageLength = hdr->PageLength;
3942 		req->Header.PageVersion = hdr->PageVersion;
3943 	}
3944 
3945 	cm->cm_data = params->buffer;
3946 	cm->cm_length = params->length;
3947 	if (cm->cm_data != NULL) {
3948 		cm->cm_sge = &req->PageBufferSGE;
3949 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3950 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
3951 	} else
3952 		cm->cm_sge = NULL;
3953 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3954 
3955 	cm->cm_complete_data = params;
3956 	if (params->callback != NULL) {
3957 		cm->cm_complete = mpr_config_complete;
3958 		return (mpr_map_command(sc, cm));
3959 	} else {
3960 		error = mpr_wait_command(sc, &cm, 0, CAN_SLEEP);
3961 		if (error) {
3962 			mpr_dprint(sc, MPR_FAULT,
3963 			    "Error %d reading config page\n", error);
3964 			if (cm != NULL)
3965 				mpr_free_command(sc, cm);
3966 			return (error);
3967 		}
3968 		mpr_config_complete(sc, cm);
3969 	}
3970 
3971 	return (0);
3972 }
3973 
3974 int
3975 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3976 {
3977 	return (EINVAL);
3978 }
3979 
3980 static void
3981 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
3982 {
3983 	MPI2_CONFIG_REPLY *reply;
3984 	struct mpr_config_params *params;
3985 
3986 	MPR_FUNCTRACE(sc);
3987 	params = cm->cm_complete_data;
3988 
3989 	if (cm->cm_data != NULL) {
3990 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3991 		    BUS_DMASYNC_POSTREAD);
3992 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3993 	}
3994 
3995 	/*
3996 	 * XXX KDM need to do more error recovery?  This results in the
3997 	 * device in question not getting probed.
3998 	 */
3999 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
4000 		params->status = MPI2_IOCSTATUS_BUSY;
4001 		goto done;
4002 	}
4003 
4004 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
4005 	if (reply == NULL) {
4006 		params->status = MPI2_IOCSTATUS_BUSY;
4007 		goto done;
4008 	}
4009 	params->status = reply->IOCStatus;
4010 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
4011 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
4012 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
4013 		params->hdr.Ext.PageType = reply->Header.PageType;
4014 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
4015 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
4016 	} else {
4017 		params->hdr.Struct.PageType = reply->Header.PageType;
4018 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
4019 		params->hdr.Struct.PageLength = reply->Header.PageLength;
4020 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
4021 	}
4022 
4023 done:
4024 	mpr_free_command(sc, cm);
4025 	if (params->callback != NULL)
4026 		params->callback(sc, params);
4027 
4028 	return;
4029 }
4030