xref: /freebsd/sys/dev/mpr/mpr.c (revision 8bcb0991864975618c09697b1aca10683346d9f0)
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 __FBSDID("$FreeBSD$");
35 
36 /* Communications core for Avago Technologies (LSI) MPT3 */
37 
38 /* TODO Move headers to mprvar */
39 #include <sys/types.h>
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/selinfo.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/module.h>
47 #include <sys/bus.h>
48 #include <sys/conf.h>
49 #include <sys/bio.h>
50 #include <sys/malloc.h>
51 #include <sys/uio.h>
52 #include <sys/sysctl.h>
53 #include <sys/smp.h>
54 #include <sys/queue.h>
55 #include <sys/kthread.h>
56 #include <sys/taskqueue.h>
57 #include <sys/endian.h>
58 #include <sys/eventhandler.h>
59 #include <sys/sbuf.h>
60 #include <sys/priv.h>
61 
62 #include <machine/bus.h>
63 #include <machine/resource.h>
64 #include <sys/rman.h>
65 #include <sys/proc.h>
66 
67 #include <dev/pci/pcivar.h>
68 
69 #include <cam/cam.h>
70 #include <cam/cam_ccb.h>
71 #include <cam/scsi/scsi_all.h>
72 
73 #include <dev/mpr/mpi/mpi2_type.h>
74 #include <dev/mpr/mpi/mpi2.h>
75 #include <dev/mpr/mpi/mpi2_ioc.h>
76 #include <dev/mpr/mpi/mpi2_sas.h>
77 #include <dev/mpr/mpi/mpi2_pci.h>
78 #include <dev/mpr/mpi/mpi2_cnfg.h>
79 #include <dev/mpr/mpi/mpi2_init.h>
80 #include <dev/mpr/mpi/mpi2_tool.h>
81 #include <dev/mpr/mpr_ioctl.h>
82 #include <dev/mpr/mprvar.h>
83 #include <dev/mpr/mpr_table.h>
84 #include <dev/mpr/mpr_sas.h>
85 
86 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
87 static int mpr_init_queues(struct mpr_softc *sc);
88 static void mpr_resize_queues(struct mpr_softc *sc);
89 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
90 static int mpr_transition_operational(struct mpr_softc *sc);
91 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
92 static void mpr_iocfacts_free(struct mpr_softc *sc);
93 static void mpr_startup(void *arg);
94 static int mpr_send_iocinit(struct mpr_softc *sc);
95 static int mpr_alloc_queues(struct mpr_softc *sc);
96 static int mpr_alloc_hw_queues(struct mpr_softc *sc);
97 static int mpr_alloc_replies(struct mpr_softc *sc);
98 static int mpr_alloc_requests(struct mpr_softc *sc);
99 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc);
100 static int mpr_attach_log(struct mpr_softc *sc);
101 static __inline void mpr_complete_command(struct mpr_softc *sc,
102     struct mpr_command *cm);
103 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
104     MPI2_EVENT_NOTIFICATION_REPLY *reply);
105 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
106 static void mpr_periodic(void *);
107 static int mpr_reregister_events(struct mpr_softc *sc);
108 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
109 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
110 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
111 static int mpr_debug_sysctl(SYSCTL_HANDLER_ARGS);
112 static int mpr_dump_reqs(SYSCTL_HANDLER_ARGS);
113 static void mpr_parse_debug(struct mpr_softc *sc, char *list);
114 
115 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD, 0, "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 _reply_descriptor {
132         u64 word;
133         struct {
134                 u32 low;
135                 u32 high;
136         } u;
137 } reply_descriptor, request_descriptor;
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 __FreeBSD_version >= 1000029
168 	if (curthread->td_no_sleeping)
169 #else //__FreeBSD_version < 1000029
170 	if (curthread->td_pflags & TDP_NOSLEEPING)
171 #endif //__FreeBSD_version >= 1000029
172 		sleep_flag = NO_SLEEP;
173 
174 	mpr_dprint(sc, MPR_INIT, "sequence start, sleep_flag=%d\n", sleep_flag);
175 	/* Push the magic sequence */
176 	error = ETIMEDOUT;
177 	while (tries++ < 20) {
178 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
179 			mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
180 			    mpt2_reset_magic[i]);
181 
182 		/* wait 100 msec */
183 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
184 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
185 			    "mprdiag", hz/10);
186 		else if (sleep_flag == CAN_SLEEP)
187 			pause("mprdiag", hz/10);
188 		else
189 			DELAY(100 * 1000);
190 
191 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
192 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
193 			error = 0;
194 			break;
195 		}
196 	}
197 	if (error) {
198 		mpr_dprint(sc, MPR_INIT, "sequence failed, error=%d, exit\n",
199 		    error);
200 		return (error);
201 	}
202 
203 	/* Send the actual reset.  XXX need to refresh the reg? */
204 	reg |= MPI2_DIAG_RESET_ADAPTER;
205 	mpr_dprint(sc, MPR_INIT, "sequence success, sending reset, reg= 0x%x\n",
206 	    reg);
207 	mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
208 
209 	/* Wait up to 300 seconds in 50ms intervals */
210 	error = ETIMEDOUT;
211 	for (i = 0; i < 6000; i++) {
212 		/*
213 		 * Wait 50 msec. If this is the first time through, wait 256
214 		 * msec to satisfy Diag Reset timing requirements.
215 		 */
216 		if (first_wait_done) {
217 			if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
218 				msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
219 				    "mprdiag", hz/20);
220 			else if (sleep_flag == CAN_SLEEP)
221 				pause("mprdiag", hz/20);
222 			else
223 				DELAY(50 * 1000);
224 		} else {
225 			DELAY(256 * 1000);
226 			first_wait_done = TRUE;
227 		}
228 		/*
229 		 * Check for the RESET_ADAPTER bit to be cleared first, then
230 		 * wait for the RESET state to be cleared, which takes a little
231 		 * longer.
232 		 */
233 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
234 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
235 			continue;
236 		}
237 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
238 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
239 			error = 0;
240 			break;
241 		}
242 	}
243 	if (error) {
244 		mpr_dprint(sc, MPR_INIT, "reset failed, error= %d, exit\n",
245 		    error);
246 		return (error);
247 	}
248 
249 	mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
250 	mpr_dprint(sc, MPR_INIT, "diag reset success, exit\n");
251 
252 	return (0);
253 }
254 
255 static int
256 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag)
257 {
258 	int error;
259 
260 	MPR_FUNCTRACE(sc);
261 
262 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
263 
264 	error = 0;
265 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
266 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
267 	    MPI2_DOORBELL_FUNCTION_SHIFT);
268 
269 	if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) {
270 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
271 		    "Doorbell handshake failed\n");
272 		error = ETIMEDOUT;
273 	}
274 
275 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
276 	return (error);
277 }
278 
279 static int
280 mpr_transition_ready(struct mpr_softc *sc)
281 {
282 	uint32_t reg, state;
283 	int error, tries = 0;
284 	int sleep_flags;
285 
286 	MPR_FUNCTRACE(sc);
287 	/* If we are in attach call, do not sleep */
288 	sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE)
289 	    ? CAN_SLEEP : NO_SLEEP;
290 
291 	error = 0;
292 
293 	mpr_dprint(sc, MPR_INIT, "%s entered, sleep_flags= %d\n",
294 	    __func__, sleep_flags);
295 
296 	while (tries++ < 1200) {
297 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
298 		mpr_dprint(sc, MPR_INIT, "  Doorbell= 0x%x\n", reg);
299 
300 		/*
301 		 * Ensure the IOC is ready to talk.  If it's not, try
302 		 * resetting it.
303 		 */
304 		if (reg & MPI2_DOORBELL_USED) {
305 			mpr_dprint(sc, MPR_INIT, "  Not ready, sending diag "
306 			    "reset\n");
307 			mpr_diag_reset(sc, sleep_flags);
308 			DELAY(50000);
309 			continue;
310 		}
311 
312 		/* Is the adapter owned by another peer? */
313 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
314 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
315 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC is under the "
316 			    "control of another peer host, aborting "
317 			    "initialization.\n");
318 			error = ENXIO;
319 			break;
320 		}
321 
322 		state = reg & MPI2_IOC_STATE_MASK;
323 		if (state == MPI2_IOC_STATE_READY) {
324 			/* Ready to go! */
325 			error = 0;
326 			break;
327 		} else if (state == MPI2_IOC_STATE_FAULT) {
328 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC in fault "
329 			    "state 0x%x, resetting\n",
330 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
331 			mpr_diag_reset(sc, sleep_flags);
332 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
333 			/* Need to take ownership */
334 			mpr_message_unit_reset(sc, sleep_flags);
335 		} else if (state == MPI2_IOC_STATE_RESET) {
336 			/* Wait a bit, IOC might be in transition */
337 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
338 			    "IOC in unexpected reset state\n");
339 		} else {
340 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
341 			    "IOC in unknown state 0x%x\n", state);
342 			error = EINVAL;
343 			break;
344 		}
345 
346 		/* Wait 50ms for things to settle down. */
347 		DELAY(50000);
348 	}
349 
350 	if (error)
351 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
352 		    "Cannot transition IOC to ready\n");
353 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
354 	return (error);
355 }
356 
357 static int
358 mpr_transition_operational(struct mpr_softc *sc)
359 {
360 	uint32_t reg, state;
361 	int error;
362 
363 	MPR_FUNCTRACE(sc);
364 
365 	error = 0;
366 	reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
367 	mpr_dprint(sc, MPR_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
368 
369 	state = reg & MPI2_IOC_STATE_MASK;
370 	if (state != MPI2_IOC_STATE_READY) {
371 		mpr_dprint(sc, MPR_INIT, "IOC not ready\n");
372 		if ((error = mpr_transition_ready(sc)) != 0) {
373 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
374 			    "failed to transition ready, exit\n");
375 			return (error);
376 		}
377 	}
378 
379 	error = mpr_send_iocinit(sc);
380 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
381 
382 	return (error);
383 }
384 
385 static void
386 mpr_resize_queues(struct mpr_softc *sc)
387 {
388 	u_int reqcr, prireqcr, maxio, sges_per_frame, chain_seg_size;
389 
390 	/*
391 	 * Size the queues. Since the reply queues always need one free
392 	 * entry, we'll deduct one reply message here.  The LSI documents
393 	 * suggest instead to add a count to the request queue, but I think
394 	 * that it's better to deduct from reply queue.
395 	 */
396 	prireqcr = MAX(1, sc->max_prireqframes);
397 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
398 
399 	reqcr = MAX(2, sc->max_reqframes);
400 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
401 
402 	sc->num_reqs = prireqcr + reqcr;
403 	sc->num_prireqs = prireqcr;
404 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
405 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
406 
407 	/* Store the request frame size in bytes rather than as 32bit words */
408 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
409 
410 	/*
411 	 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to
412 	 * get the size of a Chain Frame.  Previous versions use the size as a
413 	 * Request Frame for the Chain Frame size.  If IOCMaxChainSegmentSize
414 	 * is 0, use the default value.  The IOCMaxChainSegmentSize is the
415 	 * number of 16-byte elelements that can fit in a Chain Frame, which is
416 	 * the size of an IEEE Simple SGE.
417 	 */
418 	if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) {
419 		chain_seg_size = htole16(sc->facts->IOCMaxChainSegmentSize);
420 		if (chain_seg_size == 0)
421 			chain_seg_size = MPR_DEFAULT_CHAIN_SEG_SIZE;
422 		sc->chain_frame_size = chain_seg_size *
423 		    MPR_MAX_CHAIN_ELEMENT_SIZE;
424 	} else {
425 		sc->chain_frame_size = sc->reqframesz;
426 	}
427 
428 	/*
429 	 * Max IO Size is Page Size * the following:
430 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
431 	 * + 1 for no chain needed in last frame
432 	 *
433 	 * If user suggests a Max IO size to use, use the smaller of the
434 	 * user's value and the calculated value as long as the user's
435 	 * value is larger than 0. The user's value is in pages.
436 	 */
437 	sges_per_frame = sc->chain_frame_size/sizeof(MPI2_IEEE_SGE_SIMPLE64)-1;
438 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
439 
440 	/*
441 	 * If I/O size limitation requested then use it and pass up to CAM.
442 	 * If not, use MAXPHYS as an optimization hint, but report HW limit.
443 	 */
444 	if (sc->max_io_pages > 0) {
445 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
446 		sc->maxio = maxio;
447 	} else {
448 		sc->maxio = maxio;
449 		maxio = min(maxio, MAXPHYS);
450 	}
451 
452 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
453 	    sges_per_frame * reqcr;
454 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
455 		sc->num_chains = sc->max_chains;
456 
457 	/*
458 	 * Figure out the number of MSIx-based queues.  If the firmware or
459 	 * user has done something crazy and not allowed enough credit for
460 	 * the queues to be useful then don't enable multi-queue.
461 	 */
462 	if (sc->facts->MaxMSIxVectors < 2)
463 		sc->msi_msgs = 1;
464 
465 	if (sc->msi_msgs > 1) {
466 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
467 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
468 		if (sc->num_reqs / sc->msi_msgs < 2)
469 			sc->msi_msgs = 1;
470 	}
471 
472 	mpr_dprint(sc, MPR_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
473 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
474 }
475 
476 /*
477  * This is called during attach and when re-initializing due to a Diag Reset.
478  * IOC Facts is used to allocate many of the structures needed by the driver.
479  * If called from attach, de-allocation is not required because the driver has
480  * not allocated any structures yet, but if called from a Diag Reset, previously
481  * allocated structures based on IOC Facts will need to be freed and re-
482  * allocated bases on the latest IOC Facts.
483  */
484 static int
485 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
486 {
487 	int error;
488 	Mpi2IOCFactsReply_t saved_facts;
489 	uint8_t saved_mode, reallocating;
490 
491 	mpr_dprint(sc, MPR_INIT|MPR_TRACE, "%s entered\n", __func__);
492 
493 	/* Save old IOC Facts and then only reallocate if Facts have changed */
494 	if (!attaching) {
495 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
496 	}
497 
498 	/*
499 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
500 	 * a re-initialization and only return the error if attaching so the OS
501 	 * can handle it.
502 	 */
503 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
504 		if (attaching) {
505 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to get "
506 			    "IOC Facts with error %d, exit\n", error);
507 			return (error);
508 		} else {
509 			panic("%s failed to get IOC Facts with error %d\n",
510 			    __func__, error);
511 		}
512 	}
513 
514 	MPR_DPRINT_PAGE(sc, MPR_XINFO, iocfacts, sc->facts);
515 
516 	snprintf(sc->fw_version, sizeof(sc->fw_version),
517 	    "%02d.%02d.%02d.%02d",
518 	    sc->facts->FWVersion.Struct.Major,
519 	    sc->facts->FWVersion.Struct.Minor,
520 	    sc->facts->FWVersion.Struct.Unit,
521 	    sc->facts->FWVersion.Struct.Dev);
522 
523 	mpr_dprint(sc, MPR_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
524 	    MPR_DRIVER_VERSION);
525 	mpr_dprint(sc, MPR_INFO,
526 	    "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
527 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
528 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
529 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc"
530 	    "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV");
531 
532 	/*
533 	 * If the chip doesn't support event replay then a hard reset will be
534 	 * required to trigger a full discovery.  Do the reset here then
535 	 * retransition to Ready.  A hard reset might have already been done,
536 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
537 	 * for a Diag Reset.
538 	 */
539 	if (attaching && ((sc->facts->IOCCapabilities &
540 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
541 		mpr_dprint(sc, MPR_INIT, "No event replay, resetting\n");
542 		mpr_diag_reset(sc, NO_SLEEP);
543 		if ((error = mpr_transition_ready(sc)) != 0) {
544 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
545 			    "transition to ready with error %d, exit\n",
546 			    error);
547 			return (error);
548 		}
549 	}
550 
551 	/*
552 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
553 	 * changed from the previous IOC Facts, log a warning, but only if
554 	 * checking this after a Diag Reset and not during attach.
555 	 */
556 	saved_mode = sc->ir_firmware;
557 	if (sc->facts->IOCCapabilities &
558 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
559 		sc->ir_firmware = 1;
560 	if (!attaching) {
561 		if (sc->ir_firmware != saved_mode) {
562 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "new IR/IT mode "
563 			    "in IOC Facts does not match previous mode\n");
564 		}
565 	}
566 
567 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
568 	reallocating = FALSE;
569 	sc->mpr_flags &= ~MPR_FLAGS_REALLOCATED;
570 
571 	if ((!attaching) &&
572 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
573 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
574 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
575 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
576 	    (saved_facts.ProductID != sc->facts->ProductID) ||
577 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
578 	    (saved_facts.IOCRequestFrameSize !=
579 	    sc->facts->IOCRequestFrameSize) ||
580 	    (saved_facts.IOCMaxChainSegmentSize !=
581 	    sc->facts->IOCMaxChainSegmentSize) ||
582 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
583 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
584 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
585 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
586 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
587 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
588 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
589 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
590 	    (saved_facts.MaxPersistentEntries !=
591 	    sc->facts->MaxPersistentEntries))) {
592 		reallocating = TRUE;
593 
594 		/* Record that we reallocated everything */
595 		sc->mpr_flags |= MPR_FLAGS_REALLOCATED;
596 	}
597 
598 	/*
599 	 * Some things should be done if attaching or re-allocating after a Diag
600 	 * Reset, but are not needed after a Diag Reset if the FW has not
601 	 * changed.
602 	 */
603 	if (attaching || reallocating) {
604 		/*
605 		 * Check if controller supports FW diag buffers and set flag to
606 		 * enable each type.
607 		 */
608 		if (sc->facts->IOCCapabilities &
609 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
610 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
611 			    enabled = TRUE;
612 		if (sc->facts->IOCCapabilities &
613 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
614 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
615 			    enabled = TRUE;
616 		if (sc->facts->IOCCapabilities &
617 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
618 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
619 			    enabled = TRUE;
620 
621 		/*
622 		 * Set flags for some supported items.
623 		 */
624 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
625 			sc->eedp_enabled = TRUE;
626 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
627 			sc->control_TLR = TRUE;
628 		if ((sc->facts->IOCCapabilities &
629 		    MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ) &&
630 		    (sc->mpr_flags & MPR_FLAGS_SEA_IOC))
631 			sc->atomic_desc_capable = TRUE;
632 
633 		mpr_resize_queues(sc);
634 
635 		/*
636 		 * Initialize all Tail Queues
637 		 */
638 		TAILQ_INIT(&sc->req_list);
639 		TAILQ_INIT(&sc->high_priority_req_list);
640 		TAILQ_INIT(&sc->chain_list);
641 		TAILQ_INIT(&sc->prp_page_list);
642 		TAILQ_INIT(&sc->tm_list);
643 	}
644 
645 	/*
646 	 * If doing a Diag Reset and the FW is significantly different
647 	 * (reallocating will be set above in IOC Facts comparison), then all
648 	 * buffers based on the IOC Facts will need to be freed before they are
649 	 * reallocated.
650 	 */
651 	if (reallocating) {
652 		mpr_iocfacts_free(sc);
653 		mprsas_realloc_targets(sc, saved_facts.MaxTargets +
654 		    saved_facts.MaxVolumes);
655 	}
656 
657 	/*
658 	 * Any deallocation has been completed.  Now start reallocating
659 	 * if needed.  Will only need to reallocate if attaching or if the new
660 	 * IOC Facts are different from the previous IOC Facts after a Diag
661 	 * Reset. Targets have already been allocated above if needed.
662 	 */
663 	error = 0;
664 	while (attaching || reallocating) {
665 		if ((error = mpr_alloc_hw_queues(sc)) != 0)
666 			break;
667 		if ((error = mpr_alloc_replies(sc)) != 0)
668 			break;
669 		if ((error = mpr_alloc_requests(sc)) != 0)
670 			break;
671 		if ((error = mpr_alloc_queues(sc)) != 0)
672 			break;
673 		break;
674 	}
675 	if (error) {
676 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
677 		    "Failed to alloc queues with error %d\n", error);
678 		mpr_free(sc);
679 		return (error);
680 	}
681 
682 	/* Always initialize the queues */
683 	bzero(sc->free_queue, sc->fqdepth * 4);
684 	mpr_init_queues(sc);
685 
686 	/*
687 	 * Always get the chip out of the reset state, but only panic if not
688 	 * attaching.  If attaching and there is an error, that is handled by
689 	 * the OS.
690 	 */
691 	error = mpr_transition_operational(sc);
692 	if (error != 0) {
693 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
694 		    "transition to operational with error %d\n", error);
695 		mpr_free(sc);
696 		return (error);
697 	}
698 
699 	/*
700 	 * Finish the queue initialization.
701 	 * These are set here instead of in mpr_init_queues() because the
702 	 * IOC resets these values during the state transition in
703 	 * mpr_transition_operational().  The free index is set to 1
704 	 * because the corresponding index in the IOC is set to 0, and the
705 	 * IOC treats the queues as full if both are set to the same value.
706 	 * Hence the reason that the queue can't hold all of the possible
707 	 * replies.
708 	 */
709 	sc->replypostindex = 0;
710 	mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
711 	mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
712 
713 	/*
714 	 * Attach the subsystems so they can prepare their event masks.
715 	 * XXX Should be dynamic so that IM/IR and user modules can attach
716 	 */
717 	error = 0;
718 	while (attaching) {
719 		mpr_dprint(sc, MPR_INIT, "Attaching subsystems\n");
720 		if ((error = mpr_attach_log(sc)) != 0)
721 			break;
722 		if ((error = mpr_attach_sas(sc)) != 0)
723 			break;
724 		if ((error = mpr_attach_user(sc)) != 0)
725 			break;
726 		break;
727 	}
728 	if (error) {
729 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
730 		    "Failed to attach all subsystems: error %d\n", error);
731 		mpr_free(sc);
732 		return (error);
733 	}
734 
735 	/*
736 	 * XXX If the number of MSI-X vectors changes during re-init, this
737 	 * won't see it and adjust.
738 	 */
739 	if (attaching && (error = mpr_pci_setup_interrupts(sc)) != 0) {
740 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
741 		    "Failed to setup interrupts\n");
742 		mpr_free(sc);
743 		return (error);
744 	}
745 
746 	return (error);
747 }
748 
749 /*
750  * This is called if memory is being free (during detach for example) and when
751  * buffers need to be reallocated due to a Diag Reset.
752  */
753 static void
754 mpr_iocfacts_free(struct mpr_softc *sc)
755 {
756 	struct mpr_command *cm;
757 	int i;
758 
759 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
760 
761 	if (sc->free_busaddr != 0)
762 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
763 	if (sc->free_queue != NULL)
764 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
765 		    sc->queues_map);
766 	if (sc->queues_dmat != NULL)
767 		bus_dma_tag_destroy(sc->queues_dmat);
768 
769 	if (sc->chain_frames != NULL) {
770 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
771 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
772 		    sc->chain_map);
773 	}
774 	if (sc->chain_dmat != NULL)
775 		bus_dma_tag_destroy(sc->chain_dmat);
776 
777 	if (sc->sense_busaddr != 0)
778 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
779 	if (sc->sense_frames != NULL)
780 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
781 		    sc->sense_map);
782 	if (sc->sense_dmat != NULL)
783 		bus_dma_tag_destroy(sc->sense_dmat);
784 
785 	if (sc->prp_page_busaddr != 0)
786 		bus_dmamap_unload(sc->prp_page_dmat, sc->prp_page_map);
787 	if (sc->prp_pages != NULL)
788 		bus_dmamem_free(sc->prp_page_dmat, sc->prp_pages,
789 		    sc->prp_page_map);
790 	if (sc->prp_page_dmat != NULL)
791 		bus_dma_tag_destroy(sc->prp_page_dmat);
792 
793 	if (sc->reply_busaddr != 0)
794 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
795 	if (sc->reply_frames != NULL)
796 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
797 		    sc->reply_map);
798 	if (sc->reply_dmat != NULL)
799 		bus_dma_tag_destroy(sc->reply_dmat);
800 
801 	if (sc->req_busaddr != 0)
802 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
803 	if (sc->req_frames != NULL)
804 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
805 	if (sc->req_dmat != NULL)
806 		bus_dma_tag_destroy(sc->req_dmat);
807 
808 	if (sc->chains != NULL)
809 		free(sc->chains, M_MPR);
810 	if (sc->prps != NULL)
811 		free(sc->prps, M_MPR);
812 	if (sc->commands != NULL) {
813 		for (i = 1; i < sc->num_reqs; i++) {
814 			cm = &sc->commands[i];
815 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
816 		}
817 		free(sc->commands, M_MPR);
818 	}
819 	if (sc->buffer_dmat != NULL)
820 		bus_dma_tag_destroy(sc->buffer_dmat);
821 
822 	mpr_pci_free_interrupts(sc);
823 	free(sc->queues, M_MPR);
824 	sc->queues = NULL;
825 }
826 
827 /*
828  * The terms diag reset and hard reset are used interchangeably in the MPI
829  * docs to mean resetting the controller chip.  In this code diag reset
830  * cleans everything up, and the hard reset function just sends the reset
831  * sequence to the chip.  This should probably be refactored so that every
832  * subsystem gets a reset notification of some sort, and can clean up
833  * appropriately.
834  */
835 int
836 mpr_reinit(struct mpr_softc *sc)
837 {
838 	int error;
839 	struct mprsas_softc *sassc;
840 
841 	sassc = sc->sassc;
842 
843 	MPR_FUNCTRACE(sc);
844 
845 	mtx_assert(&sc->mpr_mtx, MA_OWNED);
846 
847 	mpr_dprint(sc, MPR_INIT|MPR_INFO, "Reinitializing controller\n");
848 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) {
849 		mpr_dprint(sc, MPR_INIT, "Reset already in progress\n");
850 		return 0;
851 	}
852 
853 	/*
854 	 * Make sure the completion callbacks can recognize they're getting
855 	 * a NULL cm_reply due to a reset.
856 	 */
857 	sc->mpr_flags |= MPR_FLAGS_DIAGRESET;
858 
859 	/*
860 	 * Mask interrupts here.
861 	 */
862 	mpr_dprint(sc, MPR_INIT, "Masking interrupts and resetting\n");
863 	mpr_mask_intr(sc);
864 
865 	error = mpr_diag_reset(sc, CAN_SLEEP);
866 	if (error != 0) {
867 		panic("%s hard reset failed with error %d\n", __func__, error);
868 	}
869 
870 	/* Restore the PCI state, including the MSI-X registers */
871 	mpr_pci_restore(sc);
872 
873 	/* Give the I/O subsystem special priority to get itself prepared */
874 	mprsas_handle_reinit(sc);
875 
876 	/*
877 	 * Get IOC Facts and allocate all structures based on this information.
878 	 * The attach function will also call mpr_iocfacts_allocate at startup.
879 	 * If relevant values have changed in IOC Facts, this function will free
880 	 * all of the memory based on IOC Facts and reallocate that memory.
881 	 */
882 	if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) {
883 		panic("%s IOC Facts based allocation failed with error %d\n",
884 		    __func__, error);
885 	}
886 
887 	/*
888 	 * Mapping structures will be re-allocated after getting IOC Page8, so
889 	 * free these structures here.
890 	 */
891 	mpr_mapping_exit(sc);
892 
893 	/*
894 	 * The static page function currently read is IOC Page8.  Others can be
895 	 * added in future.  It's possible that the values in IOC Page8 have
896 	 * changed after a Diag Reset due to user modification, so always read
897 	 * these.  Interrupts are masked, so unmask them before getting config
898 	 * pages.
899 	 */
900 	mpr_unmask_intr(sc);
901 	sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET;
902 	mpr_base_static_config_pages(sc);
903 
904 	/*
905 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
906 	 * mapping tables.
907 	 */
908 	mpr_mapping_initialize(sc);
909 
910 	/*
911 	 * Restart will reload the event masks clobbered by the reset, and
912 	 * then enable the port.
913 	 */
914 	mpr_reregister_events(sc);
915 
916 	/* the end of discovery will release the simq, so we're done. */
917 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Finished sc %p post %u free %u\n",
918 	    sc, sc->replypostindex, sc->replyfreeindex);
919 	mprsas_release_simq_reinit(sassc);
920 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
921 
922 	return 0;
923 }
924 
925 /* Wait for the chip to ACK a word that we've put into its FIFO
926  * Wait for <timeout> seconds. In single loop wait for busy loop
927  * for 500 microseconds.
928  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
929  * */
930 static int
931 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag)
932 {
933 	u32 cntdn, count;
934 	u32 int_status;
935 	u32 doorbell;
936 
937 	count = 0;
938 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
939 	do {
940 		int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
941 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
942 			mpr_dprint(sc, MPR_TRACE, "%s: successful count(%d), "
943 			    "timeout(%d)\n", __func__, count, timeout);
944 			return 0;
945 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
946 			doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
947 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
948 			    MPI2_IOC_STATE_FAULT) {
949 				mpr_dprint(sc, MPR_FAULT,
950 				    "fault_state(0x%04x)!\n", doorbell);
951 				return (EFAULT);
952 			}
953 		} else if (int_status == 0xFFFFFFFF)
954 			goto out;
955 
956 		/*
957 		 * If it can sleep, sleep for 1 milisecond, else busy loop for
958  		 * 0.5 milisecond
959 		 */
960 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
961 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba",
962 			    hz/1000);
963 		else if (sleep_flag == CAN_SLEEP)
964 			pause("mprdba", hz/1000);
965 		else
966 			DELAY(500);
967 		count++;
968 	} while (--cntdn);
969 
970 out:
971 	mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), "
972 		"int_status(%x)!\n", __func__, count, int_status);
973 	return (ETIMEDOUT);
974 }
975 
976 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
977 static int
978 mpr_wait_db_int(struct mpr_softc *sc)
979 {
980 	int retry;
981 
982 	for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) {
983 		if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
984 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
985 			return (0);
986 		DELAY(2000);
987 	}
988 	return (ETIMEDOUT);
989 }
990 
991 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
992 static int
993 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
994     int req_sz, int reply_sz, int timeout)
995 {
996 	uint32_t *data32;
997 	uint16_t *data16;
998 	int i, count, ioc_sz, residual;
999 	int sleep_flags = CAN_SLEEP;
1000 
1001 #if __FreeBSD_version >= 1000029
1002 	if (curthread->td_no_sleeping)
1003 #else //__FreeBSD_version < 1000029
1004 	if (curthread->td_pflags & TDP_NOSLEEPING)
1005 #endif //__FreeBSD_version >= 1000029
1006 		sleep_flags = NO_SLEEP;
1007 
1008 	/* Step 1 */
1009 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1010 
1011 	/* Step 2 */
1012 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1013 		return (EBUSY);
1014 
1015 	/* Step 3
1016 	 * Announce that a message is coming through the doorbell.  Messages
1017 	 * are pushed at 32bit words, so round up if needed.
1018 	 */
1019 	count = (req_sz + 3) / 4;
1020 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
1021 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
1022 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
1023 
1024 	/* Step 4 */
1025 	if (mpr_wait_db_int(sc) ||
1026 	    (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
1027 		mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n");
1028 		return (ENXIO);
1029 	}
1030 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1031 	if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
1032 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n");
1033 		return (ENXIO);
1034 	}
1035 
1036 	/* Step 5 */
1037 	/* Clock out the message data synchronously in 32-bit dwords*/
1038 	data32 = (uint32_t *)req;
1039 	for (i = 0; i < count; i++) {
1040 		mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1041 		if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
1042 			mpr_dprint(sc, MPR_FAULT,
1043 			    "Timeout while writing doorbell\n");
1044 			return (ENXIO);
1045 		}
1046 	}
1047 
1048 	/* Step 6 */
1049 	/* Clock in the reply in 16-bit words.  The total length of the
1050 	 * message is always in the 4th byte, so clock out the first 2 words
1051 	 * manually, then loop the rest.
1052 	 */
1053 	data16 = (uint16_t *)reply;
1054 	if (mpr_wait_db_int(sc) != 0) {
1055 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n");
1056 		return (ENXIO);
1057 	}
1058 	data16[0] =
1059 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1060 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1061 	if (mpr_wait_db_int(sc) != 0) {
1062 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n");
1063 		return (ENXIO);
1064 	}
1065 	data16[1] =
1066 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1067 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1068 
1069 	/* Number of 32bit words in the message */
1070 	ioc_sz = reply->MsgLength;
1071 
1072 	/*
1073 	 * Figure out how many 16bit words to clock in without overrunning.
1074 	 * The precision loss with dividing reply_sz can safely be
1075 	 * ignored because the messages can only be multiples of 32bits.
1076 	 */
1077 	residual = 0;
1078 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1079 	if (count < ioc_sz * 2) {
1080 		residual = ioc_sz * 2 - count;
1081 		mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d "
1082 		    "residual message words\n", residual);
1083 	}
1084 
1085 	for (i = 2; i < count; i++) {
1086 		if (mpr_wait_db_int(sc) != 0) {
1087 			mpr_dprint(sc, MPR_FAULT,
1088 			    "Timeout reading doorbell %d\n", i);
1089 			return (ENXIO);
1090 		}
1091 		data16[i] = mpr_regread(sc, MPI2_DOORBELL_OFFSET) &
1092 		    MPI2_DOORBELL_DATA_MASK;
1093 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1094 	}
1095 
1096 	/*
1097 	 * Pull out residual words that won't fit into the provided buffer.
1098 	 * This keeps the chip from hanging due to a driver programming
1099 	 * error.
1100 	 */
1101 	while (residual--) {
1102 		if (mpr_wait_db_int(sc) != 0) {
1103 			mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n");
1104 			return (ENXIO);
1105 		}
1106 		(void)mpr_regread(sc, MPI2_DOORBELL_OFFSET);
1107 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1108 	}
1109 
1110 	/* Step 7 */
1111 	if (mpr_wait_db_int(sc) != 0) {
1112 		mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n");
1113 		return (ENXIO);
1114 	}
1115 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1116 		mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n");
1117 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1118 
1119 	return (0);
1120 }
1121 
1122 static void
1123 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm)
1124 {
1125 	request_descriptor rd;
1126 
1127 	MPR_FUNCTRACE(sc);
1128 	mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n",
1129 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1130 
1131 	if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags &
1132 	    MPR_FLAGS_SHUTDOWN))
1133 		mtx_assert(&sc->mpr_mtx, MA_OWNED);
1134 
1135 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1136 		sc->io_cmds_highwater++;
1137 
1138 	KASSERT(cm->cm_state == MPR_CM_STATE_BUSY, ("command not busy\n"));
1139 	cm->cm_state = MPR_CM_STATE_INQUEUE;
1140 
1141 	if (sc->atomic_desc_capable) {
1142 		rd.u.low = cm->cm_desc.Words.Low;
1143 		mpr_regwrite(sc, MPI26_ATOMIC_REQUEST_DESCRIPTOR_POST_OFFSET,
1144 		    rd.u.low);
1145 	} else {
1146 		rd.u.low = cm->cm_desc.Words.Low;
1147 		rd.u.high = cm->cm_desc.Words.High;
1148 		rd.word = htole64(rd.word);
1149 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1150 		    rd.u.low);
1151 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1152 		    rd.u.high);
1153 	}
1154 }
1155 
1156 /*
1157  * Just the FACTS, ma'am.
1158  */
1159 static int
1160 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1161 {
1162 	MPI2_DEFAULT_REPLY *reply;
1163 	MPI2_IOC_FACTS_REQUEST request;
1164 	int error, req_sz, reply_sz;
1165 
1166 	MPR_FUNCTRACE(sc);
1167 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1168 
1169 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1170 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1171 	reply = (MPI2_DEFAULT_REPLY *)facts;
1172 
1173 	bzero(&request, req_sz);
1174 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1175 	error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1176 
1177 	mpr_dprint(sc, MPR_INIT, "%s exit, error= %d\n", __func__, error);
1178 	return (error);
1179 }
1180 
1181 static int
1182 mpr_send_iocinit(struct mpr_softc *sc)
1183 {
1184 	MPI2_IOC_INIT_REQUEST	init;
1185 	MPI2_DEFAULT_REPLY	reply;
1186 	int req_sz, reply_sz, error;
1187 	struct timeval now;
1188 	uint64_t time_in_msec;
1189 
1190 	MPR_FUNCTRACE(sc);
1191 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1192 
1193 	/* Do a quick sanity check on proper initialization */
1194 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1195 	    || (sc->replyframesz == 0)) {
1196 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
1197 		    "Driver not fully initialized for IOCInit\n");
1198 		return (EINVAL);
1199 	}
1200 
1201 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1202 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1203 	bzero(&init, req_sz);
1204 	bzero(&reply, reply_sz);
1205 
1206 	/*
1207 	 * Fill in the init block.  Note that most addresses are
1208 	 * deliberately in the lower 32bits of memory.  This is a micro-
1209 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1210 	 */
1211 	init.Function = MPI2_FUNCTION_IOC_INIT;
1212 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1213 	init.MsgVersion = htole16(MPI2_VERSION);
1214 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1215 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1216 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1217 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1218 	init.SenseBufferAddressHigh = 0;
1219 	init.SystemReplyAddressHigh = 0;
1220 	init.SystemRequestFrameBaseAddress.High = 0;
1221 	init.SystemRequestFrameBaseAddress.Low =
1222 	    htole32((uint32_t)sc->req_busaddr);
1223 	init.ReplyDescriptorPostQueueAddress.High = 0;
1224 	init.ReplyDescriptorPostQueueAddress.Low =
1225 	    htole32((uint32_t)sc->post_busaddr);
1226 	init.ReplyFreeQueueAddress.High = 0;
1227 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1228 	getmicrotime(&now);
1229 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1230 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1231 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1232 	init.HostPageSize = HOST_PAGE_SIZE_4K;
1233 
1234 	error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1235 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1236 		error = ENXIO;
1237 
1238 	mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1239 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
1240 	return (error);
1241 }
1242 
1243 void
1244 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1245 {
1246 	bus_addr_t *addr;
1247 
1248 	addr = arg;
1249 	*addr = segs[0].ds_addr;
1250 }
1251 
1252 void
1253 mpr_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1254 {
1255 	struct mpr_busdma_context *ctx;
1256 	int need_unload, need_free;
1257 
1258 	ctx = (struct mpr_busdma_context *)arg;
1259 	need_unload = 0;
1260 	need_free = 0;
1261 
1262 	mpr_lock(ctx->softc);
1263 	ctx->error = error;
1264 	ctx->completed = 1;
1265 	if ((error == 0) && (ctx->abandoned == 0)) {
1266 		*ctx->addr = segs[0].ds_addr;
1267 	} else {
1268 		if (nsegs != 0)
1269 			need_unload = 1;
1270 		if (ctx->abandoned != 0)
1271 			need_free = 1;
1272 	}
1273 	if (need_free == 0)
1274 		wakeup(ctx);
1275 
1276 	mpr_unlock(ctx->softc);
1277 
1278 	if (need_unload != 0) {
1279 		bus_dmamap_unload(ctx->buffer_dmat,
1280 				  ctx->buffer_dmamap);
1281 		*ctx->addr = 0;
1282 	}
1283 
1284 	if (need_free != 0)
1285 		free(ctx, M_MPR);
1286 }
1287 
1288 static int
1289 mpr_alloc_queues(struct mpr_softc *sc)
1290 {
1291 	struct mpr_queue *q;
1292 	int nq, i;
1293 
1294 	nq = sc->msi_msgs;
1295 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Allocating %d I/O queues\n", nq);
1296 
1297 	sc->queues = malloc(sizeof(struct mpr_queue) * nq, M_MPR,
1298 	     M_NOWAIT|M_ZERO);
1299 	if (sc->queues == NULL)
1300 		return (ENOMEM);
1301 
1302 	for (i = 0; i < nq; i++) {
1303 		q = &sc->queues[i];
1304 		mpr_dprint(sc, MPR_INIT, "Configuring queue %d %p\n", i, q);
1305 		q->sc = sc;
1306 		q->qnum = i;
1307 	}
1308 	return (0);
1309 }
1310 
1311 static int
1312 mpr_alloc_hw_queues(struct mpr_softc *sc)
1313 {
1314 	bus_dma_tag_template_t t;
1315 	bus_addr_t queues_busaddr;
1316 	uint8_t *queues;
1317 	int qsize, fqsize, pqsize;
1318 
1319 	/*
1320 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1321 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1322 	 * This queue supplies fresh reply frames for the firmware to use.
1323 	 *
1324 	 * The reply descriptor post queue contains 8 byte entries in
1325 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1326 	 * contains filled-in reply frames sent from the firmware to the host.
1327 	 *
1328 	 * These two queues are allocated together for simplicity.
1329 	 */
1330 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1331 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1332 	fqsize= sc->fqdepth * 4;
1333 	pqsize = sc->pqdepth * 8;
1334 	qsize = fqsize + pqsize;
1335 
1336 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1337 	t.alignment = 16;
1338 	t.lowaddr = BUS_SPACE_MAXADDR_32BIT;
1339 	t.maxsize = t.maxsegsize = qsize;
1340 	t.nsegments = 1;
1341 	if (bus_dma_template_tag(&t, &sc->queues_dmat)) {
1342 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues DMA tag\n");
1343 		return (ENOMEM);
1344         }
1345         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1346 	    &sc->queues_map)) {
1347 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues memory\n");
1348 		return (ENOMEM);
1349         }
1350         bzero(queues, qsize);
1351         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1352 	    mpr_memaddr_cb, &queues_busaddr, 0);
1353 
1354 	sc->free_queue = (uint32_t *)queues;
1355 	sc->free_busaddr = queues_busaddr;
1356 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1357 	sc->post_busaddr = queues_busaddr + fqsize;
1358 	mpr_dprint(sc, MPR_INIT, "free queue busaddr= %#016jx size= %d\n",
1359 	    (uintmax_t)sc->free_busaddr, fqsize);
1360 	mpr_dprint(sc, MPR_INIT, "reply queue busaddr= %#016jx size= %d\n",
1361 	    (uintmax_t)sc->post_busaddr, pqsize);
1362 
1363 	return (0);
1364 }
1365 
1366 static int
1367 mpr_alloc_replies(struct mpr_softc *sc)
1368 {
1369 	bus_dma_tag_template_t t;
1370 	int rsize, num_replies;
1371 
1372 	/* Store the reply frame size in bytes rather than as 32bit words */
1373 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1374 
1375 	/*
1376 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1377 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1378 	 * replies can be used at once.
1379 	 */
1380 	num_replies = max(sc->fqdepth, sc->num_replies);
1381 
1382 	rsize = sc->replyframesz * num_replies;
1383 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1384 	t.alignment = 4;
1385 	t.lowaddr = BUS_SPACE_MAXADDR_32BIT;
1386 	t.maxsize = t.maxsegsize = rsize;
1387 	t.nsegments = 1;
1388 	if (bus_dma_template_tag(&t, &sc->reply_dmat)) {
1389 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies DMA tag\n");
1390 		return (ENOMEM);
1391         }
1392         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1393 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1394 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies memory\n");
1395 		return (ENOMEM);
1396         }
1397         bzero(sc->reply_frames, rsize);
1398         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1399 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1400 	mpr_dprint(sc, MPR_INIT, "reply frames busaddr= %#016jx size= %d\n",
1401 	    (uintmax_t)sc->reply_busaddr, rsize);
1402 
1403 	return (0);
1404 }
1405 
1406 static void
1407 mpr_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1408 {
1409 	struct mpr_softc *sc = arg;
1410 	struct mpr_chain *chain;
1411 	bus_size_t bo;
1412 	int i, o, s;
1413 
1414 	if (error != 0)
1415 		return;
1416 
1417 	for (i = 0, o = 0, s = 0; s < nsegs; s++) {
1418 		for (bo = 0; bo + sc->chain_frame_size <= segs[s].ds_len;
1419 		    bo += sc->chain_frame_size) {
1420 			chain = &sc->chains[i++];
1421 			chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o);
1422 			chain->chain_busaddr = segs[s].ds_addr + bo;
1423 			o += sc->chain_frame_size;
1424 			mpr_free_chain(sc, chain);
1425 		}
1426 		if (bo != segs[s].ds_len)
1427 			o += segs[s].ds_len - bo;
1428 	}
1429 	sc->chain_free_lowwater = i;
1430 }
1431 
1432 static int
1433 mpr_alloc_requests(struct mpr_softc *sc)
1434 {
1435 	bus_dma_tag_template_t t;
1436 	struct mpr_command *cm;
1437 	int i, rsize, nsegs;
1438 
1439 	rsize = sc->reqframesz * sc->num_reqs;
1440 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1441 	t.alignment = 16;
1442 	t.lowaddr = BUS_SPACE_MAXADDR_32BIT;
1443 	t.maxsize = t.maxsegsize = rsize;
1444 	t.nsegments = 1;
1445 	if (bus_dma_template_tag(&t, &sc->req_dmat)) {
1446 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request DMA tag\n");
1447 		return (ENOMEM);
1448         }
1449         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1450 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1451 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request memory\n");
1452 		return (ENOMEM);
1453         }
1454         bzero(sc->req_frames, rsize);
1455         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1456 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1457 	mpr_dprint(sc, MPR_INIT, "request frames busaddr= %#016jx size= %d\n",
1458 	    (uintmax_t)sc->req_busaddr, rsize);
1459 
1460 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->num_chains, M_MPR,
1461 	    M_NOWAIT | M_ZERO);
1462 	if (!sc->chains) {
1463 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1464 		return (ENOMEM);
1465 	}
1466 	rsize = sc->chain_frame_size * sc->num_chains;
1467 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1468 	t.alignment = 16;
1469 	t.maxsize = t.maxsegsize = rsize;
1470 	t.nsegments = howmany(rsize, PAGE_SIZE);
1471 	if (bus_dma_template_tag(&t, &sc->chain_dmat)) {
1472 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain DMA tag\n");
1473 		return (ENOMEM);
1474 	}
1475 	if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1476 	    BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) {
1477 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1478 		return (ENOMEM);
1479 	}
1480 	if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames,
1481 	    rsize, mpr_load_chains_cb, sc, BUS_DMA_NOWAIT)) {
1482 		mpr_dprint(sc, MPR_ERROR, "Cannot load chain memory\n");
1483 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
1484 		    sc->chain_map);
1485 		return (ENOMEM);
1486 	}
1487 
1488 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1489 	bus_dma_template_clone(&t, sc->req_dmat);
1490 	t.maxsize = t.maxsegsize = rsize;
1491 	if (bus_dma_template_tag(&t, &sc->sense_dmat)) {
1492 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense DMA tag\n");
1493 		return (ENOMEM);
1494         }
1495         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1496 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1497 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense memory\n");
1498 		return (ENOMEM);
1499         }
1500         bzero(sc->sense_frames, rsize);
1501         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1502 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1503 	mpr_dprint(sc, MPR_INIT, "sense frames busaddr= %#016jx size= %d\n",
1504 	    (uintmax_t)sc->sense_busaddr, rsize);
1505 
1506 	/*
1507 	 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports
1508 	 * these devices.
1509 	 */
1510 	if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) &&
1511 	    (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) {
1512 		if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM)
1513 			return (ENOMEM);
1514 	}
1515 
1516 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1517 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1518 	t.nsegments = nsegs;
1519 	t.flags = BUS_DMA_ALLOCNOW;
1520 	t.lockfunc = busdma_lock_mutex;
1521 	t.lockfuncarg = &sc->mpr_mtx;
1522 	if (bus_dma_template_tag(&t, &sc->buffer_dmat)) {
1523 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate buffer DMA tag\n");
1524 		return (ENOMEM);
1525         }
1526 
1527 	/*
1528 	 * SMID 0 cannot be used as a free command per the firmware spec.
1529 	 * Just drop that command instead of risking accounting bugs.
1530 	 */
1531 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1532 	    M_MPR, M_WAITOK | M_ZERO);
1533 	if (!sc->commands) {
1534 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate command memory\n");
1535 		return (ENOMEM);
1536 	}
1537 	for (i = 1; i < sc->num_reqs; i++) {
1538 		cm = &sc->commands[i];
1539 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1540 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1541 		cm->cm_sense = &sc->sense_frames[i];
1542 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1543 		cm->cm_desc.Default.SMID = i;
1544 		cm->cm_sc = sc;
1545 		cm->cm_state = MPR_CM_STATE_BUSY;
1546 		TAILQ_INIT(&cm->cm_chain_list);
1547 		TAILQ_INIT(&cm->cm_prp_page_list);
1548 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1549 
1550 		/* XXX Is a failure here a critical problem? */
1551 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap)
1552 		    == 0) {
1553 			if (i <= sc->num_prireqs)
1554 				mpr_free_high_priority_command(sc, cm);
1555 			else
1556 				mpr_free_command(sc, cm);
1557 		} else {
1558 			panic("failed to allocate command %d\n", i);
1559 			sc->num_reqs = i;
1560 			break;
1561 		}
1562 	}
1563 
1564 	return (0);
1565 }
1566 
1567 /*
1568  * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs,
1569  * which are scatter/gather lists for NVMe devices.
1570  *
1571  * This buffer must be contiguous due to the nature of how NVMe PRPs are built
1572  * and translated by FW.
1573  *
1574  * returns ENOMEM if memory could not be allocated, otherwise returns 0.
1575  */
1576 static int
1577 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc)
1578 {
1579 	bus_dma_tag_template_t t;
1580 	struct mpr_prp_page *prp_page;
1581 	int PRPs_per_page, PRPs_required, pages_required;
1582 	int rsize, i;
1583 
1584 	/*
1585 	 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number
1586 	 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is:
1587 	 * MAX_IO_SIZE / PAGE_SIZE = 256
1588 	 *
1589 	 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs
1590 	 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one
1591 	 * page (4096 / 8 = 512), so only one page is required for each I/O.
1592 	 *
1593 	 * Each of these buffers will need to be contiguous. For simplicity,
1594 	 * only one buffer is allocated here, which has all of the space
1595 	 * required for the NVMe Queue Depth. If there are problems allocating
1596 	 * this one buffer, this function will need to change to allocate
1597 	 * individual, contiguous NVME_QDEPTH buffers.
1598 	 *
1599 	 * The real calculation will use the real max io size. Above is just an
1600 	 * example.
1601 	 *
1602 	 */
1603 	PRPs_required = sc->maxio / PAGE_SIZE;
1604 	PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1;
1605 	pages_required = (PRPs_required / PRPs_per_page) + 1;
1606 
1607 	sc->prp_buffer_size = PAGE_SIZE * pages_required;
1608 	rsize = sc->prp_buffer_size * NVME_QDEPTH;
1609 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1610 	t.alignment = 4;
1611 	t.lowaddr = BUS_SPACE_MAXADDR_32BIT;
1612 	t.maxsize = t.maxsegsize = rsize;
1613 	t.nsegments = 1;
1614 	if (bus_dma_template_tag(&t, &sc->prp_page_dmat)) {
1615 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP DMA "
1616 		    "tag\n");
1617 		return (ENOMEM);
1618 	}
1619 	if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages,
1620 	    BUS_DMA_NOWAIT, &sc->prp_page_map)) {
1621 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP memory\n");
1622 		return (ENOMEM);
1623 	}
1624 	bzero(sc->prp_pages, rsize);
1625 	bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages,
1626 	    rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0);
1627 
1628 	sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR,
1629 	    M_WAITOK | M_ZERO);
1630 	for (i = 0; i < NVME_QDEPTH; i++) {
1631 		prp_page = &sc->prps[i];
1632 		prp_page->prp_page = (uint64_t *)(sc->prp_pages +
1633 		    i * sc->prp_buffer_size);
1634 		prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr +
1635 		    i * sc->prp_buffer_size);
1636 		mpr_free_prp_page(sc, prp_page);
1637 		sc->prp_pages_free_lowwater++;
1638 	}
1639 
1640 	return (0);
1641 }
1642 
1643 static int
1644 mpr_init_queues(struct mpr_softc *sc)
1645 {
1646 	int i;
1647 
1648 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1649 
1650 	/*
1651 	 * According to the spec, we need to use one less reply than we
1652 	 * have space for on the queue.  So sc->num_replies (the number we
1653 	 * use) should be less than sc->fqdepth (allocated size).
1654 	 */
1655 	if (sc->num_replies >= sc->fqdepth)
1656 		return (EINVAL);
1657 
1658 	/*
1659 	 * Initialize all of the free queue entries.
1660 	 */
1661 	for (i = 0; i < sc->fqdepth; i++) {
1662 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1663 	}
1664 	sc->replyfreeindex = sc->num_replies;
1665 
1666 	return (0);
1667 }
1668 
1669 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1670  * Next are the global settings, if they exist.  Highest are the per-unit
1671  * settings, if they exist.
1672  */
1673 void
1674 mpr_get_tunables(struct mpr_softc *sc)
1675 {
1676 	char tmpstr[80], mpr_debug[80];
1677 
1678 	/* XXX default to some debugging for now */
1679 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1680 	sc->disable_msix = 0;
1681 	sc->disable_msi = 0;
1682 	sc->max_msix = MPR_MSIX_MAX;
1683 	sc->max_chains = MPR_CHAIN_FRAMES;
1684 	sc->max_io_pages = MPR_MAXIO_PAGES;
1685 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1686 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1687 	sc->use_phynum = 1;
1688 	sc->max_reqframes = MPR_REQ_FRAMES;
1689 	sc->max_prireqframes = MPR_PRI_REQ_FRAMES;
1690 	sc->max_replyframes = MPR_REPLY_FRAMES;
1691 	sc->max_evtframes = MPR_EVT_REPLY_FRAMES;
1692 
1693 	/*
1694 	 * Grab the global variables.
1695 	 */
1696 	bzero(mpr_debug, 80);
1697 	if (TUNABLE_STR_FETCH("hw.mpr.debug_level", mpr_debug, 80) != 0)
1698 		mpr_parse_debug(sc, mpr_debug);
1699 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1700 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1701 	TUNABLE_INT_FETCH("hw.mpr.max_msix", &sc->max_msix);
1702 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1703 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1704 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1705 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1706 	TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum);
1707 	TUNABLE_INT_FETCH("hw.mpr.max_reqframes", &sc->max_reqframes);
1708 	TUNABLE_INT_FETCH("hw.mpr.max_prireqframes", &sc->max_prireqframes);
1709 	TUNABLE_INT_FETCH("hw.mpr.max_replyframes", &sc->max_replyframes);
1710 	TUNABLE_INT_FETCH("hw.mpr.max_evtframes", &sc->max_evtframes);
1711 
1712 	/* Grab the unit-instance variables */
1713 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1714 	    device_get_unit(sc->mpr_dev));
1715 	bzero(mpr_debug, 80);
1716 	if (TUNABLE_STR_FETCH(tmpstr, mpr_debug, 80) != 0)
1717 		mpr_parse_debug(sc, mpr_debug);
1718 
1719 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1720 	    device_get_unit(sc->mpr_dev));
1721 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1722 
1723 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1724 	    device_get_unit(sc->mpr_dev));
1725 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1726 
1727 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_msix",
1728 	    device_get_unit(sc->mpr_dev));
1729 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1730 
1731 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1732 	    device_get_unit(sc->mpr_dev));
1733 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1734 
1735 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1736 	    device_get_unit(sc->mpr_dev));
1737 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1738 
1739 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1740 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1741 	    device_get_unit(sc->mpr_dev));
1742 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1743 
1744 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1745 	    device_get_unit(sc->mpr_dev));
1746 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1747 
1748 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1749 	    device_get_unit(sc->mpr_dev));
1750 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1751 
1752 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num",
1753 	    device_get_unit(sc->mpr_dev));
1754 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1755 
1756 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_reqframes",
1757 	    device_get_unit(sc->mpr_dev));
1758 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1759 
1760 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_prireqframes",
1761 	    device_get_unit(sc->mpr_dev));
1762 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1763 
1764 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_replyframes",
1765 	    device_get_unit(sc->mpr_dev));
1766 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1767 
1768 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_evtframes",
1769 	    device_get_unit(sc->mpr_dev));
1770 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1771 }
1772 
1773 static void
1774 mpr_setup_sysctl(struct mpr_softc *sc)
1775 {
1776 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1777 	struct sysctl_oid	*sysctl_tree = NULL;
1778 	char tmpstr[80], tmpstr2[80];
1779 
1780 	/*
1781 	 * Setup the sysctl variable so the user can change the debug level
1782 	 * on the fly.
1783 	 */
1784 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1785 	    device_get_unit(sc->mpr_dev));
1786 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1787 
1788 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1789 	if (sysctl_ctx != NULL)
1790 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1791 
1792 	if (sysctl_tree == NULL) {
1793 		sysctl_ctx_init(&sc->sysctl_ctx);
1794 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1795 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1796 		    CTLFLAG_RD, 0, tmpstr);
1797 		if (sc->sysctl_tree == NULL)
1798 			return;
1799 		sysctl_ctx = &sc->sysctl_ctx;
1800 		sysctl_tree = sc->sysctl_tree;
1801 	}
1802 
1803 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1804 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1805 	    sc, 0, mpr_debug_sysctl, "A", "mpr debug level");
1806 
1807 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1808 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1809 	    "Disable the use of MSI-X interrupts");
1810 
1811 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1812 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1813 	    "User-defined maximum number of MSIX queues");
1814 
1815 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1816 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1817 	    "Negotiated number of MSIX queues");
1818 
1819 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1820 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1821 	    "Total number of allocated request frames");
1822 
1823 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1824 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1825 	    "Total number of allocated high priority request frames");
1826 
1827 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1828 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1829 	    "Total number of allocated reply frames");
1830 
1831 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1832 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1833 	    "Total number of event frames allocated");
1834 
1835 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1836 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1837 	    strlen(sc->fw_version), "firmware version");
1838 
1839 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1840 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPR_DRIVER_VERSION,
1841 	    strlen(MPR_DRIVER_VERSION), "driver version");
1842 
1843 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1844 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1845 	    &sc->io_cmds_active, 0, "number of currently active commands");
1846 
1847 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1848 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1849 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1850 
1851 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1852 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1853 	    &sc->chain_free, 0, "number of free chain elements");
1854 
1855 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1856 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1857 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1858 
1859 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1860 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1861 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1862 
1863 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1864 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1865 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1866 	    "IOCFacts)");
1867 
1868 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1869 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1870 	    "enable SSU to SATA SSD/HDD at shutdown");
1871 
1872 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1873 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1874 	    &sc->chain_alloc_fail, "chain allocation failures");
1875 
1876 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1877 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1878 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1879 	    "spinup after SATA ID error");
1880 
1881 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1882 	    OID_AUTO, "dump_reqs", CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_SKIP, sc, 0,
1883 	    mpr_dump_reqs, "I", "Dump Active Requests");
1884 
1885 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1886 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1887 	    "Use the phy number for enumeration");
1888 
1889 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1890 	    OID_AUTO, "prp_pages_free", CTLFLAG_RD,
1891 	    &sc->prp_pages_free, 0, "number of free PRP pages");
1892 
1893 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1894 	    OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD,
1895 	    &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages");
1896 
1897 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1898 	    OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD,
1899 	    &sc->prp_page_alloc_fail, "PRP page allocation failures");
1900 }
1901 
1902 static struct mpr_debug_string {
1903 	char *name;
1904 	int flag;
1905 } mpr_debug_strings[] = {
1906 	{"info", MPR_INFO},
1907 	{"fault", MPR_FAULT},
1908 	{"event", MPR_EVENT},
1909 	{"log", MPR_LOG},
1910 	{"recovery", MPR_RECOVERY},
1911 	{"error", MPR_ERROR},
1912 	{"init", MPR_INIT},
1913 	{"xinfo", MPR_XINFO},
1914 	{"user", MPR_USER},
1915 	{"mapping", MPR_MAPPING},
1916 	{"trace", MPR_TRACE}
1917 };
1918 
1919 enum mpr_debug_level_combiner {
1920 	COMB_NONE,
1921 	COMB_ADD,
1922 	COMB_SUB
1923 };
1924 
1925 static int
1926 mpr_debug_sysctl(SYSCTL_HANDLER_ARGS)
1927 {
1928 	struct mpr_softc *sc;
1929 	struct mpr_debug_string *string;
1930 	struct sbuf *sbuf;
1931 	char *buffer;
1932 	size_t sz;
1933 	int i, len, debug, error;
1934 
1935 	sc = (struct mpr_softc *)arg1;
1936 
1937 	error = sysctl_wire_old_buffer(req, 0);
1938 	if (error != 0)
1939 		return (error);
1940 
1941 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1942 	debug = sc->mpr_debug;
1943 
1944 	sbuf_printf(sbuf, "%#x", debug);
1945 
1946 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1947 	for (i = 0; i < sz; i++) {
1948 		string = &mpr_debug_strings[i];
1949 		if (debug & string->flag)
1950 			sbuf_printf(sbuf, ",%s", string->name);
1951 	}
1952 
1953 	error = sbuf_finish(sbuf);
1954 	sbuf_delete(sbuf);
1955 
1956 	if (error || req->newptr == NULL)
1957 		return (error);
1958 
1959 	len = req->newlen - req->newidx;
1960 	if (len == 0)
1961 		return (0);
1962 
1963 	buffer = malloc(len, M_MPR, M_ZERO|M_WAITOK);
1964 	error = SYSCTL_IN(req, buffer, len);
1965 
1966 	mpr_parse_debug(sc, buffer);
1967 
1968 	free(buffer, M_MPR);
1969 	return (error);
1970 }
1971 
1972 static void
1973 mpr_parse_debug(struct mpr_softc *sc, char *list)
1974 {
1975 	struct mpr_debug_string *string;
1976 	enum mpr_debug_level_combiner op;
1977 	char *token, *endtoken;
1978 	size_t sz;
1979 	int flags, i;
1980 
1981 	if (list == NULL || *list == '\0')
1982 		return;
1983 
1984 	if (*list == '+') {
1985 		op = COMB_ADD;
1986 		list++;
1987 	} else if (*list == '-') {
1988 		op = COMB_SUB;
1989 		list++;
1990 	} else
1991 		op = COMB_NONE;
1992 	if (*list == '\0')
1993 		return;
1994 
1995 	flags = 0;
1996 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1997 	while ((token = strsep(&list, ":,")) != NULL) {
1998 
1999 		/* Handle integer flags */
2000 		flags |= strtol(token, &endtoken, 0);
2001 		if (token != endtoken)
2002 			continue;
2003 
2004 		/* Handle text flags */
2005 		for (i = 0; i < sz; i++) {
2006 			string = &mpr_debug_strings[i];
2007 			if (strcasecmp(token, string->name) == 0) {
2008 				flags |= string->flag;
2009 				break;
2010 			}
2011 		}
2012 	}
2013 
2014 	switch (op) {
2015 	case COMB_NONE:
2016 		sc->mpr_debug = flags;
2017 		break;
2018 	case COMB_ADD:
2019 		sc->mpr_debug |= flags;
2020 		break;
2021 	case COMB_SUB:
2022 		sc->mpr_debug &= (~flags);
2023 		break;
2024 	}
2025 	return;
2026 }
2027 
2028 struct mpr_dumpreq_hdr {
2029 	uint32_t	smid;
2030 	uint32_t	state;
2031 	uint32_t	numframes;
2032 	uint32_t	deschi;
2033 	uint32_t	desclo;
2034 };
2035 
2036 static int
2037 mpr_dump_reqs(SYSCTL_HANDLER_ARGS)
2038 {
2039 	struct mpr_softc *sc;
2040 	struct mpr_chain *chain, *chain1;
2041 	struct mpr_command *cm;
2042 	struct mpr_dumpreq_hdr hdr;
2043 	struct sbuf *sb;
2044 	uint32_t smid, state;
2045 	int i, numreqs, error = 0;
2046 
2047 	sc = (struct mpr_softc *)arg1;
2048 
2049 	if ((error = priv_check(curthread, PRIV_DRIVER)) != 0) {
2050 		printf("priv check error %d\n", error);
2051 		return (error);
2052 	}
2053 
2054 	state = MPR_CM_STATE_INQUEUE;
2055 	smid = 1;
2056 	numreqs = sc->num_reqs;
2057 
2058 	if (req->newptr != NULL)
2059 		return (EINVAL);
2060 
2061 	if (smid == 0 || smid > sc->num_reqs)
2062 		return (EINVAL);
2063 	if (numreqs <= 0 || (numreqs + smid > sc->num_reqs))
2064 		numreqs = sc->num_reqs;
2065 	sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req);
2066 
2067 	/* Best effort, no locking */
2068 	for (i = smid; i < numreqs; i++) {
2069 		cm = &sc->commands[i];
2070 		if (cm->cm_state != state)
2071 			continue;
2072 		hdr.smid = i;
2073 		hdr.state = cm->cm_state;
2074 		hdr.numframes = 1;
2075 		hdr.deschi = cm->cm_desc.Words.High;
2076 		hdr.desclo = cm->cm_desc.Words.Low;
2077 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
2078 		   chain1)
2079 			hdr.numframes++;
2080 		sbuf_bcat(sb, &hdr, sizeof(hdr));
2081 		sbuf_bcat(sb, cm->cm_req, 128);
2082 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
2083 		    chain1)
2084 			sbuf_bcat(sb, chain->chain, 128);
2085 	}
2086 
2087 	error = sbuf_finish(sb);
2088 	sbuf_delete(sb);
2089 	return (error);
2090 }
2091 
2092 int
2093 mpr_attach(struct mpr_softc *sc)
2094 {
2095 	int error;
2096 
2097 	MPR_FUNCTRACE(sc);
2098 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2099 
2100 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
2101 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
2102 	callout_init_mtx(&sc->device_check_callout, &sc->mpr_mtx, 0);
2103 	TAILQ_INIT(&sc->event_list);
2104 	timevalclear(&sc->lastfail);
2105 
2106 	if ((error = mpr_transition_ready(sc)) != 0) {
2107 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2108 		    "Failed to transition ready\n");
2109 		return (error);
2110 	}
2111 
2112 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
2113 	    M_ZERO|M_NOWAIT);
2114 	if (!sc->facts) {
2115 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2116 		    "Cannot allocate memory, exit\n");
2117 		return (ENOMEM);
2118 	}
2119 
2120 	/*
2121 	 * Get IOC Facts and allocate all structures based on this information.
2122 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
2123 	 * Facts. If relevant values have changed in IOC Facts, this function
2124 	 * will free all of the memory based on IOC Facts and reallocate that
2125 	 * memory.  If this fails, any allocated memory should already be freed.
2126 	 */
2127 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
2128 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC Facts allocation "
2129 		    "failed with error %d\n", error);
2130 		return (error);
2131 	}
2132 
2133 	/* Start the periodic watchdog check on the IOC Doorbell */
2134 	mpr_periodic(sc);
2135 
2136 	/*
2137 	 * The portenable will kick off discovery events that will drive the
2138 	 * rest of the initialization process.  The CAM/SAS module will
2139 	 * hold up the boot sequence until discovery is complete.
2140 	 */
2141 	sc->mpr_ich.ich_func = mpr_startup;
2142 	sc->mpr_ich.ich_arg = sc;
2143 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
2144 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2145 		    "Cannot establish MPR config hook\n");
2146 		error = EINVAL;
2147 	}
2148 
2149 	/*
2150 	 * Allow IR to shutdown gracefully when shutdown occurs.
2151 	 */
2152 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
2153 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
2154 
2155 	if (sc->shutdown_eh == NULL)
2156 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2157 		    "shutdown event registration failed\n");
2158 
2159 	mpr_setup_sysctl(sc);
2160 
2161 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
2162 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
2163 
2164 	return (error);
2165 }
2166 
2167 /* Run through any late-start handlers. */
2168 static void
2169 mpr_startup(void *arg)
2170 {
2171 	struct mpr_softc *sc;
2172 
2173 	sc = (struct mpr_softc *)arg;
2174 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2175 
2176 	mpr_lock(sc);
2177 	mpr_unmask_intr(sc);
2178 
2179 	/* initialize device mapping tables */
2180 	mpr_base_static_config_pages(sc);
2181 	mpr_mapping_initialize(sc);
2182 	mprsas_startup(sc);
2183 	mpr_unlock(sc);
2184 
2185 	mpr_dprint(sc, MPR_INIT, "disestablish config intrhook\n");
2186 	config_intrhook_disestablish(&sc->mpr_ich);
2187 	sc->mpr_ich.ich_arg = NULL;
2188 
2189 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2190 }
2191 
2192 /* Periodic watchdog.  Is called with the driver lock already held. */
2193 static void
2194 mpr_periodic(void *arg)
2195 {
2196 	struct mpr_softc *sc;
2197 	uint32_t db;
2198 
2199 	sc = (struct mpr_softc *)arg;
2200 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
2201 		return;
2202 
2203 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
2204 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2205 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
2206 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
2207 			panic("TEMPERATURE FAULT: STOPPING.");
2208 		}
2209 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2210 		mpr_reinit(sc);
2211 	}
2212 
2213 	callout_reset(&sc->periodic, MPR_PERIODIC_DELAY * hz, mpr_periodic, sc);
2214 }
2215 
2216 static void
2217 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
2218     MPI2_EVENT_NOTIFICATION_REPLY *event)
2219 {
2220 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2221 
2222 	MPR_DPRINT_EVENT(sc, generic, event);
2223 
2224 	switch (event->Event) {
2225 	case MPI2_EVENT_LOG_DATA:
2226 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2227 		if (sc->mpr_debug & MPR_EVENT)
2228 			hexdump(event->EventData, event->EventDataLength, NULL,
2229 			    0);
2230 		break;
2231 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2232 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2233 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2234 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2235 		     entry->LogSequence);
2236 		break;
2237 	default:
2238 		break;
2239 	}
2240 	return;
2241 }
2242 
2243 static int
2244 mpr_attach_log(struct mpr_softc *sc)
2245 {
2246 	uint8_t events[16];
2247 
2248 	bzero(events, 16);
2249 	setbit(events, MPI2_EVENT_LOG_DATA);
2250 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2251 
2252 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
2253 	    &sc->mpr_log_eh);
2254 
2255 	return (0);
2256 }
2257 
2258 static int
2259 mpr_detach_log(struct mpr_softc *sc)
2260 {
2261 
2262 	if (sc->mpr_log_eh != NULL)
2263 		mpr_deregister_events(sc, sc->mpr_log_eh);
2264 	return (0);
2265 }
2266 
2267 /*
2268  * Free all of the driver resources and detach submodules.  Should be called
2269  * without the lock held.
2270  */
2271 int
2272 mpr_free(struct mpr_softc *sc)
2273 {
2274 	int error;
2275 
2276 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2277 	/* Turn off the watchdog */
2278 	mpr_lock(sc);
2279 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
2280 	mpr_unlock(sc);
2281 	/* Lock must not be held for this */
2282 	callout_drain(&sc->periodic);
2283 	callout_drain(&sc->device_check_callout);
2284 
2285 	if (((error = mpr_detach_log(sc)) != 0) ||
2286 	    ((error = mpr_detach_sas(sc)) != 0)) {
2287 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "failed to detach "
2288 		    "subsystems, error= %d, exit\n", error);
2289 		return (error);
2290 	}
2291 
2292 	mpr_detach_user(sc);
2293 
2294 	/* Put the IOC back in the READY state. */
2295 	mpr_lock(sc);
2296 	if ((error = mpr_transition_ready(sc)) != 0) {
2297 		mpr_unlock(sc);
2298 		return (error);
2299 	}
2300 	mpr_unlock(sc);
2301 
2302 	if (sc->facts != NULL)
2303 		free(sc->facts, M_MPR);
2304 
2305 	/*
2306 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2307 	 * to free these buffers too.
2308 	 */
2309 	mpr_iocfacts_free(sc);
2310 
2311 	if (sc->sysctl_tree != NULL)
2312 		sysctl_ctx_free(&sc->sysctl_ctx);
2313 
2314 	/* Deregister the shutdown function */
2315 	if (sc->shutdown_eh != NULL)
2316 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2317 
2318 	mtx_destroy(&sc->mpr_mtx);
2319 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2320 
2321 	return (0);
2322 }
2323 
2324 static __inline void
2325 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
2326 {
2327 	MPR_FUNCTRACE(sc);
2328 
2329 	if (cm == NULL) {
2330 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
2331 		return;
2332 	}
2333 
2334 	cm->cm_state = MPR_CM_STATE_BUSY;
2335 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
2336 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
2337 
2338 	if (cm->cm_complete != NULL) {
2339 		mpr_dprint(sc, MPR_TRACE,
2340 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
2341 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
2342 		    cm->cm_reply);
2343 		cm->cm_complete(sc, cm);
2344 	}
2345 
2346 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
2347 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
2348 		wakeup(cm);
2349 	}
2350 
2351 	if (sc->io_cmds_active != 0) {
2352 		sc->io_cmds_active--;
2353 	} else {
2354 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
2355 		    "out of sync - resynching to 0\n");
2356 	}
2357 }
2358 
2359 static void
2360 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
2361 {
2362 	union loginfo_type {
2363 		u32	loginfo;
2364 		struct {
2365 			u32	subcode:16;
2366 			u32	code:8;
2367 			u32	originator:4;
2368 			u32	bus_type:4;
2369 		} dw;
2370 	};
2371 	union loginfo_type sas_loginfo;
2372 	char *originator_str = NULL;
2373 
2374 	sas_loginfo.loginfo = log_info;
2375 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2376 		return;
2377 
2378 	/* each nexus loss loginfo */
2379 	if (log_info == 0x31170000)
2380 		return;
2381 
2382 	/* eat the loginfos associated with task aborts */
2383 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
2384 	    (log_info == 0x31130000))
2385 		return;
2386 
2387 	switch (sas_loginfo.dw.originator) {
2388 	case 0:
2389 		originator_str = "IOP";
2390 		break;
2391 	case 1:
2392 		originator_str = "PL";
2393 		break;
2394 	case 2:
2395 		originator_str = "IR";
2396 		break;
2397 	}
2398 
2399 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
2400 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
2401 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
2402 }
2403 
2404 static void
2405 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
2406 {
2407 	MPI2DefaultReply_t *mpi_reply;
2408 	u16 sc_status;
2409 
2410 	mpi_reply = (MPI2DefaultReply_t*)reply;
2411 	sc_status = le16toh(mpi_reply->IOCStatus);
2412 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2413 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2414 }
2415 
2416 void
2417 mpr_intr(void *data)
2418 {
2419 	struct mpr_softc *sc;
2420 	uint32_t status;
2421 
2422 	sc = (struct mpr_softc *)data;
2423 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2424 
2425 	/*
2426 	 * Check interrupt status register to flush the bus.  This is
2427 	 * needed for both INTx interrupts and driver-driven polling
2428 	 */
2429 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2430 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2431 		return;
2432 
2433 	mpr_lock(sc);
2434 	mpr_intr_locked(data);
2435 	mpr_unlock(sc);
2436 	return;
2437 }
2438 
2439 /*
2440  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2441  * chip.  Hopefully this theory is correct.
2442  */
2443 void
2444 mpr_intr_msi(void *data)
2445 {
2446 	struct mpr_softc *sc;
2447 
2448 	sc = (struct mpr_softc *)data;
2449 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2450 	mpr_lock(sc);
2451 	mpr_intr_locked(data);
2452 	mpr_unlock(sc);
2453 	return;
2454 }
2455 
2456 /*
2457  * The locking is overly broad and simplistic, but easy to deal with for now.
2458  */
2459 void
2460 mpr_intr_locked(void *data)
2461 {
2462 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2463 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2464 	mpr_fw_diagnostic_buffer_t *pBuffer;
2465 	struct mpr_softc *sc;
2466 	uint64_t tdesc;
2467 	struct mpr_command *cm = NULL;
2468 	uint8_t flags;
2469 	u_int pq;
2470 
2471 	sc = (struct mpr_softc *)data;
2472 
2473 	pq = sc->replypostindex;
2474 	mpr_dprint(sc, MPR_TRACE,
2475 	    "%s sc %p starting with replypostindex %u\n",
2476 	    __func__, sc, sc->replypostindex);
2477 
2478 	for ( ;; ) {
2479 		cm = NULL;
2480 		desc = &sc->post_queue[sc->replypostindex];
2481 
2482 		/*
2483 		 * Copy and clear out the descriptor so that any reentry will
2484 		 * immediately know that this descriptor has already been
2485 		 * looked at.  There is unfortunate casting magic because the
2486 		 * MPI API doesn't have a cardinal 64bit type.
2487 		 */
2488 		tdesc = 0xffffffffffffffff;
2489 		tdesc = atomic_swap_64((uint64_t *)desc, tdesc);
2490 		desc = (MPI2_REPLY_DESCRIPTORS_UNION *)&tdesc;
2491 
2492 		flags = desc->Default.ReplyFlags &
2493 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2494 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
2495 		    (le32toh(desc->Words.High) == 0xffffffff))
2496 			break;
2497 
2498 		/* increment the replypostindex now, so that event handlers
2499 		 * and cm completion handlers which decide to do a diag
2500 		 * reset can zero it without it getting incremented again
2501 		 * afterwards, and we break out of this loop on the next
2502 		 * iteration since the reply post queue has been cleared to
2503 		 * 0xFF and all descriptors look unused (which they are).
2504 		 */
2505 		if (++sc->replypostindex >= sc->pqdepth)
2506 			sc->replypostindex = 0;
2507 
2508 		switch (flags) {
2509 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2510 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
2511 		case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS:
2512 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2513 			KASSERT(cm->cm_state == MPR_CM_STATE_INQUEUE,
2514 			    ("command not inqueue\n"));
2515 			cm->cm_state = MPR_CM_STATE_BUSY;
2516 			cm->cm_reply = NULL;
2517 			break;
2518 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2519 		{
2520 			uint32_t baddr;
2521 			uint8_t *reply;
2522 
2523 			/*
2524 			 * Re-compose the reply address from the address
2525 			 * sent back from the chip.  The ReplyFrameAddress
2526 			 * is the lower 32 bits of the physical address of
2527 			 * particular reply frame.  Convert that address to
2528 			 * host format, and then use that to provide the
2529 			 * offset against the virtual address base
2530 			 * (sc->reply_frames).
2531 			 */
2532 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2533 			reply = sc->reply_frames +
2534 				(baddr - ((uint32_t)sc->reply_busaddr));
2535 			/*
2536 			 * Make sure the reply we got back is in a valid
2537 			 * range.  If not, go ahead and panic here, since
2538 			 * we'll probably panic as soon as we deference the
2539 			 * reply pointer anyway.
2540 			 */
2541 			if ((reply < sc->reply_frames)
2542 			 || (reply > (sc->reply_frames +
2543 			     (sc->fqdepth * sc->replyframesz)))) {
2544 				printf("%s: WARNING: reply %p out of range!\n",
2545 				       __func__, reply);
2546 				printf("%s: reply_frames %p, fqdepth %d, "
2547 				       "frame size %d\n", __func__,
2548 				       sc->reply_frames, sc->fqdepth,
2549 				       sc->replyframesz);
2550 				printf("%s: baddr %#x,\n", __func__, baddr);
2551 				/* LSI-TODO. See Linux Code for Graceful exit */
2552 				panic("Reply address out of range");
2553 			}
2554 			if (le16toh(desc->AddressReply.SMID) == 0) {
2555 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2556 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2557 					/*
2558 					 * If SMID is 0 for Diag Buffer Post,
2559 					 * this implies that the reply is due to
2560 					 * a release function with a status that
2561 					 * the buffer has been released.  Set
2562 					 * the buffer flags accordingly.
2563 					 */
2564 					rel_rep =
2565 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2566 					if ((le16toh(rel_rep->IOCStatus) &
2567 					    MPI2_IOCSTATUS_MASK) ==
2568 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2569 					{
2570 						pBuffer =
2571 						    &sc->fw_diag_buffer_list[
2572 						    rel_rep->BufferType];
2573 						pBuffer->valid_data = TRUE;
2574 						pBuffer->owned_by_firmware =
2575 						    FALSE;
2576 						pBuffer->immediate = FALSE;
2577 					}
2578 				} else
2579 					mpr_dispatch_event(sc, baddr,
2580 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2581 					    reply);
2582 			} else {
2583 				cm = &sc->commands[
2584 				    le16toh(desc->AddressReply.SMID)];
2585 				if (cm->cm_state == MPR_CM_STATE_INQUEUE) {
2586 					cm->cm_reply = reply;
2587 					cm->cm_reply_data =
2588 					    le32toh(desc->AddressReply.
2589 						ReplyFrameAddress);
2590 				} else {
2591 					mpr_dprint(sc, MPR_RECOVERY,
2592 					    "Bad state for ADDRESS_REPLY status,"
2593 					    " ignoring state %d cm %p\n",
2594 					    cm->cm_state, cm);
2595 				}
2596 			}
2597 			break;
2598 		}
2599 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2600 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2601 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2602 		default:
2603 			/* Unhandled */
2604 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
2605 			    desc->Default.ReplyFlags);
2606 			cm = NULL;
2607 			break;
2608 		}
2609 
2610 		if (cm != NULL) {
2611 			// Print Error reply frame
2612 			if (cm->cm_reply)
2613 				mpr_display_reply_info(sc,cm->cm_reply);
2614 			mpr_complete_command(sc, cm);
2615 		}
2616 	}
2617 
2618 	if (pq != sc->replypostindex) {
2619 		mpr_dprint(sc, MPR_TRACE, "%s sc %p writing postindex %d\n",
2620 		    __func__, sc, sc->replypostindex);
2621 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2622 		    sc->replypostindex);
2623 	}
2624 
2625 	return;
2626 }
2627 
2628 static void
2629 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2630     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2631 {
2632 	struct mpr_event_handle *eh;
2633 	int event, handled = 0;
2634 
2635 	event = le16toh(reply->Event);
2636 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2637 		if (isset(eh->mask, event)) {
2638 			eh->callback(sc, data, reply);
2639 			handled++;
2640 		}
2641 	}
2642 
2643 	if (handled == 0)
2644 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2645 		    le16toh(event));
2646 
2647 	/*
2648 	 * This is the only place that the event/reply should be freed.
2649 	 * Anything wanting to hold onto the event data should have
2650 	 * already copied it into their own storage.
2651 	 */
2652 	mpr_free_reply(sc, data);
2653 }
2654 
2655 static void
2656 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2657 {
2658 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2659 
2660 	if (cm->cm_reply)
2661 		MPR_DPRINT_EVENT(sc, generic,
2662 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2663 
2664 	mpr_free_command(sc, cm);
2665 
2666 	/* next, send a port enable */
2667 	mprsas_startup(sc);
2668 }
2669 
2670 /*
2671  * For both register_events and update_events, the caller supplies a bitmap
2672  * of events that it _wants_.  These functions then turn that into a bitmask
2673  * suitable for the controller.
2674  */
2675 int
2676 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2677     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2678 {
2679 	struct mpr_event_handle *eh;
2680 	int error = 0;
2681 
2682 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2683 	if (!eh) {
2684 		mpr_dprint(sc, MPR_EVENT|MPR_ERROR,
2685 		    "Cannot allocate event memory\n");
2686 		return (ENOMEM);
2687 	}
2688 	eh->callback = cb;
2689 	eh->data = data;
2690 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2691 	if (mask != NULL)
2692 		error = mpr_update_events(sc, eh, mask);
2693 	*handle = eh;
2694 
2695 	return (error);
2696 }
2697 
2698 int
2699 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2700     uint8_t *mask)
2701 {
2702 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2703 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2704 	struct mpr_command *cm = NULL;
2705 	struct mpr_event_handle *eh;
2706 	int error, i;
2707 
2708 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2709 
2710 	if ((mask != NULL) && (handle != NULL))
2711 		bcopy(mask, &handle->mask[0], 16);
2712 	memset(sc->event_mask, 0xff, 16);
2713 
2714 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2715 		for (i = 0; i < 16; i++)
2716 			sc->event_mask[i] &= ~eh->mask[i];
2717 	}
2718 
2719 	if ((cm = mpr_alloc_command(sc)) == NULL)
2720 		return (EBUSY);
2721 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2722 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2723 	evtreq->MsgFlags = 0;
2724 	evtreq->SASBroadcastPrimitiveMasks = 0;
2725 #ifdef MPR_DEBUG_ALL_EVENTS
2726 	{
2727 		u_char fullmask[16];
2728 		memset(fullmask, 0x00, 16);
2729 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2730 	}
2731 #else
2732 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2733 #endif
2734 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2735 	cm->cm_data = NULL;
2736 
2737 	error = mpr_request_polled(sc, &cm);
2738 	if (cm != NULL)
2739 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2740 	if ((reply == NULL) ||
2741 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2742 		error = ENXIO;
2743 
2744 	if (reply)
2745 		MPR_DPRINT_EVENT(sc, generic, reply);
2746 
2747 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2748 
2749 	if (cm != NULL)
2750 		mpr_free_command(sc, cm);
2751 	return (error);
2752 }
2753 
2754 static int
2755 mpr_reregister_events(struct mpr_softc *sc)
2756 {
2757 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2758 	struct mpr_command *cm;
2759 	struct mpr_event_handle *eh;
2760 	int error, i;
2761 
2762 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2763 
2764 	/* first, reregister events */
2765 
2766 	memset(sc->event_mask, 0xff, 16);
2767 
2768 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2769 		for (i = 0; i < 16; i++)
2770 			sc->event_mask[i] &= ~eh->mask[i];
2771 	}
2772 
2773 	if ((cm = mpr_alloc_command(sc)) == NULL)
2774 		return (EBUSY);
2775 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2776 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2777 	evtreq->MsgFlags = 0;
2778 	evtreq->SASBroadcastPrimitiveMasks = 0;
2779 #ifdef MPR_DEBUG_ALL_EVENTS
2780 	{
2781 		u_char fullmask[16];
2782 		memset(fullmask, 0x00, 16);
2783 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2784 	}
2785 #else
2786 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2787 #endif
2788 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2789 	cm->cm_data = NULL;
2790 	cm->cm_complete = mpr_reregister_events_complete;
2791 
2792 	error = mpr_map_command(sc, cm);
2793 
2794 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2795 	    error);
2796 	return (error);
2797 }
2798 
2799 int
2800 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2801 {
2802 
2803 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2804 	free(handle, M_MPR);
2805 	return (mpr_update_events(sc, NULL, NULL));
2806 }
2807 
2808 /**
2809 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a
2810 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry
2811 * of the NVMe message (PRP1). If the data buffer is small enough to be described
2812 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to
2813 * describe a larger data buffer. If the data buffer is too large to describe
2814 * using the two PRP entriess inside the NVMe message, then PRP1 describes the
2815 * first data memory segment, and PRP2 contains a pointer to a PRP list located
2816 * elsewhere in memory to describe the remaining data memory segments. The PRP
2817 * list will be contiguous.
2818 
2819 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP
2820 * consists of a list of PRP entries to describe a number of noncontigous
2821 * physical memory segments as a single memory buffer, just as a SGL does. Note
2822 * however, that this function is only used by the IOCTL call, so the memory
2823 * given will be guaranteed to be contiguous. There is no need to translate
2824 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous
2825 * space that is one page size each.
2826 *
2827 * Each NVMe message contains two PRP entries. The first (PRP1) either contains
2828 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains
2829 * the second PRP element if the memory being described fits within 2 PRP
2830 * entries, or a PRP list pointer if the PRP spans more than two entries.
2831 *
2832 * A PRP list pointer contains the address of a PRP list, structured as a linear
2833 * array of PRP entries. Each PRP entry in this list describes a segment of
2834 * physical memory.
2835 *
2836 * Each 64-bit PRP entry comprises an address and an offset field. The address
2837 * always points to the beginning of a PAGE_SIZE physical memory page, and the
2838 * offset describes where within that page the memory segment begins. Only the
2839 * first element in a PRP list may contain a non-zero offest, implying that all
2840 * memory segments following the first begin at the start of a PAGE_SIZE page.
2841 *
2842 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory,
2843 * with exceptions for the first and last elements in the list. If the memory
2844 * being described by the list begins at a non-zero offset within the first page,
2845 * then the first PRP element will contain a non-zero offset indicating where the
2846 * region begins within the page. The last memory segment may end before the end
2847 * of the PAGE_SIZE segment, depending upon the overall size of the memory being
2848 * described by the PRP list.
2849 *
2850 * Since PRP entries lack any indication of size, the overall data buffer length
2851 * is used to determine where the end of the data memory buffer is located, and
2852 * how many PRP entries are required to describe it.
2853 *
2854 * Returns nothing.
2855 */
2856 void
2857 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm,
2858     Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data,
2859     uint32_t data_in_sz, uint32_t data_out_sz)
2860 {
2861 	int			prp_size = PRP_ENTRY_SIZE;
2862 	uint64_t		*prp_entry, *prp1_entry, *prp2_entry;
2863 	uint64_t		*prp_entry_phys, *prp_page, *prp_page_phys;
2864 	uint32_t		offset, entry_len, page_mask_result, page_mask;
2865 	bus_addr_t		paddr;
2866 	size_t			length;
2867 	struct mpr_prp_page	*prp_page_info = NULL;
2868 
2869 	/*
2870 	 * Not all commands require a data transfer. If no data, just return
2871 	 * without constructing any PRP.
2872 	 */
2873 	if (!data_in_sz && !data_out_sz)
2874 		return;
2875 
2876 	/*
2877 	 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is
2878 	 * located at a 24 byte offset from the start of the NVMe command. Then
2879 	 * set the current PRP entry pointer to PRP1.
2880 	 */
2881 	prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2882 	    NVME_CMD_PRP1_OFFSET);
2883 	prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2884 	    NVME_CMD_PRP2_OFFSET);
2885 	prp_entry = prp1_entry;
2886 
2887 	/*
2888 	 * For the PRP entries, use the specially allocated buffer of
2889 	 * contiguous memory. PRP Page allocation failures should not happen
2890 	 * because there should be enough PRP page buffers to account for the
2891 	 * possible NVMe QDepth.
2892 	 */
2893 	prp_page_info = mpr_alloc_prp_page(sc);
2894 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2895 	    "used for building a native NVMe SGL.\n", __func__));
2896 	prp_page = (uint64_t *)prp_page_info->prp_page;
2897 	prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2898 
2899 	/*
2900 	 * Insert the allocated PRP page into the command's PRP page list. This
2901 	 * will be freed when the command is freed.
2902 	 */
2903 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2904 
2905 	/*
2906 	 * Check if we are within 1 entry of a page boundary we don't want our
2907 	 * first entry to be a PRP List entry.
2908 	 */
2909 	page_mask = PAGE_SIZE - 1;
2910 	page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) &
2911 	    page_mask;
2912 	if (!page_mask_result)
2913 	{
2914 		/* Bump up to next page boundary. */
2915 		prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size);
2916 		prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys +
2917 		    prp_size);
2918 	}
2919 
2920 	/*
2921 	 * Set PRP physical pointer, which initially points to the current PRP
2922 	 * DMA memory page.
2923 	 */
2924 	prp_entry_phys = prp_page_phys;
2925 
2926 	/* Get physical address and length of the data buffer. */
2927 	paddr = (bus_addr_t)(uintptr_t)data;
2928 	if (data_in_sz)
2929 		length = data_in_sz;
2930 	else
2931 		length = data_out_sz;
2932 
2933 	/* Loop while the length is not zero. */
2934 	while (length)
2935 	{
2936 		/*
2937 		 * Check if we need to put a list pointer here if we are at page
2938 		 * boundary - prp_size (8 bytes).
2939 		 */
2940 		page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys +
2941 		    prp_size) & page_mask;
2942 		if (!page_mask_result)
2943 		{
2944 			/*
2945 			 * This is the last entry in a PRP List, so we need to
2946 			 * put a PRP list pointer here. What this does is:
2947 			 *   - bump the current memory pointer to the next
2948 			 *     address, which will be the next full page.
2949 			 *   - set the PRP Entry to point to that page. This is
2950 			 *     now the PRP List pointer.
2951 			 *   - bump the PRP Entry pointer the start of the next
2952 			 *     page. Since all of this PRP memory is contiguous,
2953 			 *     no need to get a new page - it's just the next
2954 			 *     address.
2955 			 */
2956 			prp_entry_phys++;
2957 			*prp_entry =
2958 			    htole64((uint64_t)(uintptr_t)prp_entry_phys);
2959 			prp_entry++;
2960 		}
2961 
2962 		/* Need to handle if entry will be part of a page. */
2963 		offset = (uint32_t)paddr & page_mask;
2964 		entry_len = PAGE_SIZE - offset;
2965 
2966 		if (prp_entry == prp1_entry)
2967 		{
2968 			/*
2969 			 * Must fill in the first PRP pointer (PRP1) before
2970 			 * moving on.
2971 			 */
2972 			*prp1_entry = htole64((uint64_t)paddr);
2973 
2974 			/*
2975 			 * Now point to the second PRP entry within the
2976 			 * command (PRP2).
2977 			 */
2978 			prp_entry = prp2_entry;
2979 		}
2980 		else if (prp_entry == prp2_entry)
2981 		{
2982 			/*
2983 			 * Should the PRP2 entry be a PRP List pointer or just a
2984 			 * regular PRP pointer? If there is more than one more
2985 			 * page of data, must use a PRP List pointer.
2986 			 */
2987 			if (length > PAGE_SIZE)
2988 			{
2989 				/*
2990 				 * PRP2 will contain a PRP List pointer because
2991 				 * more PRP's are needed with this command. The
2992 				 * list will start at the beginning of the
2993 				 * contiguous buffer.
2994 				 */
2995 				*prp2_entry =
2996 				    htole64(
2997 				    (uint64_t)(uintptr_t)prp_entry_phys);
2998 
2999 				/*
3000 				 * The next PRP Entry will be the start of the
3001 				 * first PRP List.
3002 				 */
3003 				prp_entry = prp_page;
3004 			}
3005 			else
3006 			{
3007 				/*
3008 				 * After this, the PRP Entries are complete.
3009 				 * This command uses 2 PRP's and no PRP list.
3010 				 */
3011 				*prp2_entry = htole64((uint64_t)paddr);
3012 			}
3013 		}
3014 		else
3015 		{
3016 			/*
3017 			 * Put entry in list and bump the addresses.
3018 			 *
3019 			 * After PRP1 and PRP2 are filled in, this will fill in
3020 			 * all remaining PRP entries in a PRP List, one per each
3021 			 * time through the loop.
3022 			 */
3023 			*prp_entry = htole64((uint64_t)paddr);
3024 			prp_entry++;
3025 			prp_entry_phys++;
3026 		}
3027 
3028 		/*
3029 		 * Bump the phys address of the command's data buffer by the
3030 		 * entry_len.
3031 		 */
3032 		paddr += entry_len;
3033 
3034 		/* Decrement length accounting for last partial page. */
3035 		if (entry_len > length)
3036 			length = 0;
3037 		else
3038 			length -= entry_len;
3039 	}
3040 }
3041 
3042 /*
3043  * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to
3044  * determine if the driver needs to build a native SGL. If so, that native SGL
3045  * is built in the contiguous buffers allocated especially for PCIe SGL
3046  * creation. If the driver will not build a native SGL, return TRUE and a
3047  * normal IEEE SGL will be built. Currently this routine supports NVMe devices
3048  * only.
3049  *
3050  * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built.
3051  */
3052 static int
3053 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm,
3054     bus_dma_segment_t *segs, int segs_left)
3055 {
3056 	uint32_t		i, sge_dwords, length, offset, entry_len;
3057 	uint32_t		num_entries, buff_len = 0, sges_in_segment;
3058 	uint32_t		page_mask, page_mask_result, *curr_buff;
3059 	uint32_t		*ptr_sgl, *ptr_first_sgl, first_page_offset;
3060 	uint32_t		first_page_data_size, end_residual;
3061 	uint64_t		*msg_phys;
3062 	bus_addr_t		paddr;
3063 	int			build_native_sgl = 0, first_prp_entry;
3064 	int			prp_size = PRP_ENTRY_SIZE;
3065 	Mpi25IeeeSgeChain64_t	*main_chain_element = NULL;
3066 	struct mpr_prp_page	*prp_page_info = NULL;
3067 
3068 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
3069 
3070 	/*
3071 	 * Add up the sizes of each segment length to get the total transfer
3072 	 * size, which will be checked against the Maximum Data Transfer Size.
3073 	 * If the data transfer length exceeds the MDTS for this device, just
3074 	 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O
3075 	 * up into multiple I/O's. [nvme_mdts = 0 means unlimited]
3076 	 */
3077 	for (i = 0; i < segs_left; i++)
3078 		buff_len += htole32(segs[i].ds_len);
3079 	if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS))
3080 		return 1;
3081 
3082 	/* Create page_mask (to get offset within page) */
3083 	page_mask = PAGE_SIZE - 1;
3084 
3085 	/*
3086 	 * Check if the number of elements exceeds the max number that can be
3087 	 * put in the main message frame (H/W can only translate an SGL that
3088 	 * is contained entirely in the main message frame).
3089 	 */
3090 	sges_in_segment = (sc->reqframesz -
3091 	    offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION);
3092 	if (segs_left > sges_in_segment)
3093 		build_native_sgl = 1;
3094 	else
3095 	{
3096 		/*
3097 		 * NVMe uses one PRP for each physical page (or part of physical
3098 		 * page).
3099 		 *    if 4 pages or less then IEEE is OK
3100 		 *    if > 5 pages then we need to build a native SGL
3101 		 *    if > 4 and <= 5 pages, then check the physical address of
3102 		 *      the first SG entry, then if this first size in the page
3103 		 *      is >= the residual beyond 4 pages then use IEEE,
3104 		 *      otherwise use native SGL
3105 		 */
3106 		if (buff_len > (PAGE_SIZE * 5))
3107 			build_native_sgl = 1;
3108 		else if ((buff_len > (PAGE_SIZE * 4)) &&
3109 		    (buff_len <= (PAGE_SIZE * 5)) )
3110 		{
3111 			msg_phys = (uint64_t *)(uintptr_t)segs[0].ds_addr;
3112 			first_page_offset =
3113 			    ((uint32_t)(uint64_t)(uintptr_t)msg_phys &
3114 			    page_mask);
3115 			first_page_data_size = PAGE_SIZE - first_page_offset;
3116 			end_residual = buff_len % PAGE_SIZE;
3117 
3118 			/*
3119 			 * If offset into first page pushes the end of the data
3120 			 * beyond end of the 5th page, we need the extra PRP
3121 			 * list.
3122 			 */
3123 			if (first_page_data_size < end_residual)
3124 				build_native_sgl = 1;
3125 
3126 			/*
3127 			 * Check if first SG entry size is < residual beyond 4
3128 			 * pages.
3129 			 */
3130 			if (htole32(segs[0].ds_len) <
3131 			    (buff_len - (PAGE_SIZE * 4)))
3132 				build_native_sgl = 1;
3133 		}
3134 	}
3135 
3136 	/* check if native SGL is needed */
3137 	if (!build_native_sgl)
3138 		return 1;
3139 
3140 	/*
3141 	 * Native SGL is needed.
3142 	 * Put a chain element in main message frame that points to the first
3143 	 * chain buffer.
3144 	 *
3145 	 * NOTE:  The ChainOffset field must be 0 when using a chain pointer to
3146 	 *        a native SGL.
3147 	 */
3148 
3149 	/* Set main message chain element pointer */
3150 	main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge;
3151 
3152 	/*
3153 	 * For NVMe the chain element needs to be the 2nd SGL entry in the main
3154 	 * message.
3155 	 */
3156 	main_chain_element = (Mpi25IeeeSgeChain64_t *)
3157 	    ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64));
3158 
3159 	/*
3160 	 * For the PRP entries, use the specially allocated buffer of
3161 	 * contiguous memory. PRP Page allocation failures should not happen
3162 	 * because there should be enough PRP page buffers to account for the
3163 	 * possible NVMe QDepth.
3164 	 */
3165 	prp_page_info = mpr_alloc_prp_page(sc);
3166 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
3167 	    "used for building a native NVMe SGL.\n", __func__));
3168 	curr_buff = (uint32_t *)prp_page_info->prp_page;
3169 	msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
3170 
3171 	/*
3172 	 * Insert the allocated PRP page into the command's PRP page list. This
3173 	 * will be freed when the command is freed.
3174 	 */
3175 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
3176 
3177 	/*
3178 	 * Check if we are within 1 entry of a page boundary we don't want our
3179 	 * first entry to be a PRP List entry.
3180 	 */
3181 	page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) &
3182 	    page_mask;
3183 	if (!page_mask_result) {
3184 		/* Bump up to next page boundary. */
3185 		curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size);
3186 		msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size);
3187 	}
3188 
3189 	/* Fill in the chain element and make it an NVMe segment type. */
3190 	main_chain_element->Address.High =
3191 	    htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32));
3192 	main_chain_element->Address.Low =
3193 	    htole32((uint32_t)(uintptr_t)msg_phys);
3194 	main_chain_element->NextChainOffset = 0;
3195 	main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3196 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3197 	    MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP;
3198 
3199 	/* Set SGL pointer to start of contiguous PCIe buffer. */
3200 	ptr_sgl = curr_buff;
3201 	sge_dwords = 2;
3202 	num_entries = 0;
3203 
3204 	/*
3205 	 * NVMe has a very convoluted PRP format. One PRP is required for each
3206 	 * page or partial page. We need to split up OS SG entries if they are
3207 	 * longer than one page or cross a page boundary. We also have to insert
3208 	 * a PRP list pointer entry as the last entry in each physical page of
3209 	 * the PRP list.
3210 	 *
3211 	 * NOTE: The first PRP "entry" is actually placed in the first SGL entry
3212 	 * in the main message in IEEE 64 format. The 2nd entry in the main
3213 	 * message is the chain element, and the rest of the PRP entries are
3214 	 * built in the contiguous PCIe buffer.
3215 	 */
3216 	first_prp_entry = 1;
3217 	ptr_first_sgl = (uint32_t *)cm->cm_sge;
3218 
3219 	for (i = 0; i < segs_left; i++) {
3220 		/* Get physical address and length of this SG entry. */
3221 		paddr = segs[i].ds_addr;
3222 		length = segs[i].ds_len;
3223 
3224 		/*
3225 		 * Check whether a given SGE buffer lies on a non-PAGED
3226 		 * boundary if this is not the first page. If so, this is not
3227 		 * expected so have FW build the SGL.
3228 		 */
3229 		if ((i != 0) && (((uint32_t)paddr & page_mask) != 0)) {
3230 			mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while "
3231 			    "building NVMe PRPs, low address is 0x%x\n",
3232 			    (uint32_t)paddr);
3233 			return 1;
3234 		}
3235 
3236 		/* Apart from last SGE, if any other SGE boundary is not page
3237 		 * aligned then it means that hole exists. Existence of hole
3238 		 * leads to data corruption. So fallback to IEEE SGEs.
3239 		 */
3240 		if (i != (segs_left - 1)) {
3241 			if (((uint32_t)paddr + length) & page_mask) {
3242 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE "
3243 				    "boundary while building NVMe PRPs, low "
3244 				    "address: 0x%x and length: %u\n",
3245 				    (uint32_t)paddr, length);
3246 				return 1;
3247 			}
3248 		}
3249 
3250 		/* Loop while the length is not zero. */
3251 		while (length) {
3252 			/*
3253 			 * Check if we need to put a list pointer here if we are
3254 			 * at page boundary - prp_size.
3255 			 */
3256 			page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl +
3257 			    prp_size) & page_mask;
3258 			if (!page_mask_result) {
3259 				/*
3260 				 * Need to put a PRP list pointer here.
3261 				 */
3262 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3263 				    prp_size);
3264 				*ptr_sgl = htole32((uintptr_t)msg_phys);
3265 				*(ptr_sgl+1) = htole32((uint64_t)(uintptr_t)
3266 				    msg_phys >> 32);
3267 				ptr_sgl += sge_dwords;
3268 				num_entries++;
3269 			}
3270 
3271 			/* Need to handle if entry will be part of a page. */
3272 			offset = (uint32_t)paddr & page_mask;
3273 			entry_len = PAGE_SIZE - offset;
3274 			if (first_prp_entry) {
3275 				/*
3276 				 * Put IEEE entry in first SGE in main message.
3277 				 * (Simple element, System addr, not end of
3278 				 * list.)
3279 				 */
3280 				*ptr_first_sgl = htole32((uint32_t)paddr);
3281 				*(ptr_first_sgl + 1) =
3282 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3283 				*(ptr_first_sgl + 2) = htole32(entry_len);
3284 				*(ptr_first_sgl + 3) = 0;
3285 
3286 				/* No longer the first PRP entry. */
3287 				first_prp_entry = 0;
3288 			} else {
3289 				/* Put entry in list. */
3290 				*ptr_sgl = htole32((uint32_t)paddr);
3291 				*(ptr_sgl + 1) =
3292 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3293 
3294 				/* Bump ptr_sgl, msg_phys, and num_entries. */
3295 				ptr_sgl += sge_dwords;
3296 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3297 				    prp_size);
3298 				num_entries++;
3299 			}
3300 
3301 			/* Bump the phys address by the entry_len. */
3302 			paddr += entry_len;
3303 
3304 			/* Decrement length accounting for last partial page. */
3305 			if (entry_len > length)
3306 				length = 0;
3307 			else
3308 				length -= entry_len;
3309 		}
3310 	}
3311 
3312 	/* Set chain element Length. */
3313 	main_chain_element->Length = htole32(num_entries * prp_size);
3314 
3315 	/* Return 0, indicating we built a native SGL. */
3316 	return 0;
3317 }
3318 
3319 /*
3320  * Add a chain element as the next SGE for the specified command.
3321  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
3322  * only required for IEEE commands.  Therefore there is no code for commands
3323  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
3324  * shouldn't be requesting chains).
3325  */
3326 static int
3327 mpr_add_chain(struct mpr_command *cm, int segsleft)
3328 {
3329 	struct mpr_softc *sc = cm->cm_sc;
3330 	MPI2_REQUEST_HEADER *req;
3331 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
3332 	struct mpr_chain *chain;
3333 	int sgc_size, current_segs, rem_segs, segs_per_frame;
3334 	uint8_t next_chain_offset = 0;
3335 
3336 	/*
3337 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
3338 	 * only IEEE commands should be requesting chains.  Return some error
3339 	 * code other than 0.
3340 	 */
3341 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
3342 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
3343 		    "an MPI SGL.\n");
3344 		return(ENOBUFS);
3345 	}
3346 
3347 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
3348 	if (cm->cm_sglsize < sgc_size)
3349 		panic("MPR: Need SGE Error Code\n");
3350 
3351 	chain = mpr_alloc_chain(cm->cm_sc);
3352 	if (chain == NULL)
3353 		return (ENOBUFS);
3354 
3355 	/*
3356 	 * Note: a double-linked list is used to make it easier to walk for
3357 	 * debugging.
3358 	 */
3359 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
3360 
3361 	/*
3362 	 * Need to know if the number of frames left is more than 1 or not.  If
3363 	 * more than 1 frame is required, NextChainOffset will need to be set,
3364 	 * which will just be the last segment of the frame.
3365 	 */
3366 	rem_segs = 0;
3367 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
3368 		/*
3369 		 * rem_segs is the number of segements remaining after the
3370 		 * segments that will go into the current frame.  Since it is
3371 		 * known that at least one more frame is required, account for
3372 		 * the chain element.  To know if more than one more frame is
3373 		 * required, just check if there will be a remainder after using
3374 		 * the current frame (with this chain) and the next frame.  If
3375 		 * so the NextChainOffset must be the last element of the next
3376 		 * frame.
3377 		 */
3378 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
3379 		rem_segs = segsleft - current_segs;
3380 		segs_per_frame = sc->chain_frame_size / sgc_size;
3381 		if (rem_segs > segs_per_frame) {
3382 			next_chain_offset = segs_per_frame - 1;
3383 		}
3384 	}
3385 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
3386 	ieee_sgc->Length = next_chain_offset ?
3387 	    htole32((uint32_t)sc->chain_frame_size) :
3388 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
3389 	ieee_sgc->NextChainOffset = next_chain_offset;
3390 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3391 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3392 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
3393 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
3394 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
3395 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
3396 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
3397 
3398 	cm->cm_sglsize = sc->chain_frame_size;
3399 	return (0);
3400 }
3401 
3402 /*
3403  * Add one scatter-gather element to the scatter-gather list for a command.
3404  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
3405  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
3406  * chain, so don't consider any chain additions.
3407  */
3408 int
3409 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
3410     int segsleft)
3411 {
3412 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3413 	u32 sge_flags;
3414 
3415 	/*
3416 	 * case 1: >=1 more segment, no room for anything (error)
3417 	 * case 2: 1 more segment and enough room for it
3418          */
3419 
3420 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
3421 		mpr_dprint(cm->cm_sc, MPR_ERROR,
3422 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
3423 		    __func__);
3424 		return(ENOBUFS);
3425 	}
3426 
3427 	KASSERT(segsleft == 1,
3428 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
3429 	    segsleft));
3430 
3431 	/*
3432 	 * There is one more segment left to add for the MPI SGL and there is
3433 	 * enough room in the frame to add it.  This is the normal case because
3434 	 * MPI SGL's don't have chains, otherwise something is wrong.
3435 	 *
3436 	 * If this is a bi-directional request, need to account for that
3437 	 * here.  Save the pre-filled sge values.  These will be used
3438 	 * either for the 2nd SGL or for a single direction SGL.  If
3439 	 * cm_out_len is non-zero, this is a bi-directional request, so
3440 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
3441 	 * fill in the IN SGL.  Note that at this time, when filling in
3442 	 * 2 SGL's for a bi-directional request, they both use the same
3443 	 * DMA buffer (same cm command).
3444 	 */
3445 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
3446 	saved_address_low = sge->Address.Low;
3447 	saved_address_high = sge->Address.High;
3448 	if (cm->cm_out_len) {
3449 		sge->FlagsLength = cm->cm_out_len |
3450 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3451 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
3452 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
3453 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3454 		    MPI2_SGE_FLAGS_SHIFT);
3455 		cm->cm_sglsize -= len;
3456 		/* Endian Safe code */
3457 		sge_flags = sge->FlagsLength;
3458 		sge->FlagsLength = htole32(sge_flags);
3459 		sge->Address.High = htole32(sge->Address.High);
3460 		sge->Address.Low = htole32(sge->Address.Low);
3461 		bcopy(sge, cm->cm_sge, len);
3462 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3463 	}
3464 	sge->FlagsLength = saved_buf_len |
3465 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3466 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
3467 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
3468 	    MPI2_SGE_FLAGS_END_OF_LIST |
3469 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3470 	    MPI2_SGE_FLAGS_SHIFT);
3471 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
3472 		sge->FlagsLength |=
3473 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
3474 		    MPI2_SGE_FLAGS_SHIFT);
3475 	} else {
3476 		sge->FlagsLength |=
3477 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
3478 		    MPI2_SGE_FLAGS_SHIFT);
3479 	}
3480 	sge->Address.Low = saved_address_low;
3481 	sge->Address.High = saved_address_high;
3482 
3483 	cm->cm_sglsize -= len;
3484 	/* Endian Safe code */
3485 	sge_flags = sge->FlagsLength;
3486 	sge->FlagsLength = htole32(sge_flags);
3487 	sge->Address.High = htole32(sge->Address.High);
3488 	sge->Address.Low = htole32(sge->Address.Low);
3489 	bcopy(sge, cm->cm_sge, len);
3490 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3491 	return (0);
3492 }
3493 
3494 /*
3495  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
3496  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
3497  * remaining size and pointer to the next SGE to fill in, respectively.
3498  */
3499 int
3500 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
3501 {
3502 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
3503 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
3504 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3505 	uint32_t sge_length;
3506 
3507 	/*
3508 	 * case 1: No room for chain or segment (error).
3509 	 * case 2: Two or more segments left but only room for chain.
3510 	 * case 3: Last segment and room for it, so set flags.
3511 	 */
3512 
3513 	/*
3514 	 * There should be room for at least one element, or there is a big
3515 	 * problem.
3516 	 */
3517 	if (cm->cm_sglsize < ieee_sge_size)
3518 		panic("MPR: Need SGE Error Code\n");
3519 
3520 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
3521 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
3522 			return (error);
3523 	}
3524 
3525 	if (segsleft == 1) {
3526 		/*
3527 		 * If this is a bi-directional request, need to account for that
3528 		 * here.  Save the pre-filled sge values.  These will be used
3529 		 * either for the 2nd SGL or for a single direction SGL.  If
3530 		 * cm_out_len is non-zero, this is a bi-directional request, so
3531 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
3532 		 * fill in the IN SGL.  Note that at this time, when filling in
3533 		 * 2 SGL's for a bi-directional request, they both use the same
3534 		 * DMA buffer (same cm command).
3535 		 */
3536 		saved_buf_len = sge->Length;
3537 		saved_address_low = sge->Address.Low;
3538 		saved_address_high = sge->Address.High;
3539 		if (cm->cm_out_len) {
3540 			sge->Length = cm->cm_out_len;
3541 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3542 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3543 			cm->cm_sglsize -= ieee_sge_size;
3544 			/* Endian Safe code */
3545 			sge_length = sge->Length;
3546 			sge->Length = htole32(sge_length);
3547 			sge->Address.High = htole32(sge->Address.High);
3548 			sge->Address.Low = htole32(sge->Address.Low);
3549 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
3550 			cm->cm_sge =
3551 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3552 			    ieee_sge_size);
3553 		}
3554 		sge->Length = saved_buf_len;
3555 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3556 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3557 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
3558 		sge->Address.Low = saved_address_low;
3559 		sge->Address.High = saved_address_high;
3560 	}
3561 
3562 	cm->cm_sglsize -= ieee_sge_size;
3563 	/* Endian Safe code */
3564 	sge_length = sge->Length;
3565 	sge->Length = htole32(sge_length);
3566 	sge->Address.High = htole32(sge->Address.High);
3567 	sge->Address.Low = htole32(sge->Address.Low);
3568 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
3569 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3570 	    ieee_sge_size);
3571 	return (0);
3572 }
3573 
3574 /*
3575  * Add one dma segment to the scatter-gather list for a command.
3576  */
3577 int
3578 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
3579     int segsleft)
3580 {
3581 	MPI2_SGE_SIMPLE64 sge;
3582 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
3583 
3584 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
3585 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3586 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3587 		ieee_sge.Length = len;
3588 		mpr_from_u64(pa, &ieee_sge.Address);
3589 
3590 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
3591 	} else {
3592 		/*
3593 		 * This driver always uses 64-bit address elements for
3594 		 * simplicity.
3595 		 */
3596 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3597 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
3598 		/* Set Endian safe macro in mpr_push_sge */
3599 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
3600 		mpr_from_u64(pa, &sge.Address);
3601 
3602 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
3603 	}
3604 }
3605 
3606 static void
3607 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
3608 {
3609 	struct mpr_softc *sc;
3610 	struct mpr_command *cm;
3611 	u_int i, dir, sflags;
3612 
3613 	cm = (struct mpr_command *)arg;
3614 	sc = cm->cm_sc;
3615 
3616 	/*
3617 	 * In this case, just print out a warning and let the chip tell the
3618 	 * user they did the wrong thing.
3619 	 */
3620 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
3621 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
3622 		    "segments, more than the %d allowed\n", __func__, nsegs,
3623 		    cm->cm_max_segs);
3624 	}
3625 
3626 	/*
3627 	 * Set up DMA direction flags.  Bi-directional requests are also handled
3628 	 * here.  In that case, both direction flags will be set.
3629 	 */
3630 	sflags = 0;
3631 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
3632 		/*
3633 		 * We have to add a special case for SMP passthrough, there
3634 		 * is no easy way to generically handle it.  The first
3635 		 * S/G element is used for the command (therefore the
3636 		 * direction bit needs to be set).  The second one is used
3637 		 * for the reply.  We'll leave it to the caller to make
3638 		 * sure we only have two buffers.
3639 		 */
3640 		/*
3641 		 * Even though the busdma man page says it doesn't make
3642 		 * sense to have both direction flags, it does in this case.
3643 		 * We have one s/g element being accessed in each direction.
3644 		 */
3645 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
3646 
3647 		/*
3648 		 * Set the direction flag on the first buffer in the SMP
3649 		 * passthrough request.  We'll clear it for the second one.
3650 		 */
3651 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
3652 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
3653 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
3654 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
3655 		dir = BUS_DMASYNC_PREWRITE;
3656 	} else
3657 		dir = BUS_DMASYNC_PREREAD;
3658 
3659 	/* Check if a native SG list is needed for an NVMe PCIe device. */
3660 	if (cm->cm_targ && cm->cm_targ->is_nvme &&
3661 	    mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) {
3662 		/* A native SG list was built, skip to end. */
3663 		goto out;
3664 	}
3665 
3666 	for (i = 0; i < nsegs; i++) {
3667 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
3668 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3669 		}
3670 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3671 		    sflags, nsegs - i);
3672 		if (error != 0) {
3673 			/* Resource shortage, roll back! */
3674 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
3675 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
3676 				    "consider increasing hw.mpr.max_chains.\n");
3677 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
3678 			mpr_complete_command(sc, cm);
3679 			return;
3680 		}
3681 	}
3682 
3683 out:
3684 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3685 	mpr_enqueue_request(sc, cm);
3686 
3687 	return;
3688 }
3689 
3690 static void
3691 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3692 	     int error)
3693 {
3694 	mpr_data_cb(arg, segs, nsegs, error);
3695 }
3696 
3697 /*
3698  * This is the routine to enqueue commands ansynchronously.
3699  * Note that the only error path here is from bus_dmamap_load(), which can
3700  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3701  * assumed that if you have a command in-hand, then you have enough credits
3702  * to use it.
3703  */
3704 int
3705 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
3706 {
3707 	int error = 0;
3708 
3709 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
3710 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3711 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
3712 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
3713 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3714 		    cm->cm_data, mpr_data_cb, cm, 0);
3715 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3716 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3717 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
3718 	} else {
3719 		/* Add a zero-length element as needed */
3720 		if (cm->cm_sge != NULL)
3721 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
3722 		mpr_enqueue_request(sc, cm);
3723 	}
3724 
3725 	return (error);
3726 }
3727 
3728 /*
3729  * This is the routine to enqueue commands synchronously.  An error of
3730  * EINPROGRESS from mpr_map_command() is ignored since the command will
3731  * be executed and enqueued automatically.  Other errors come from msleep().
3732  */
3733 int
3734 mpr_wait_command(struct mpr_softc *sc, struct mpr_command **cmp, int timeout,
3735     int sleep_flag)
3736 {
3737 	int error, rc;
3738 	struct timeval cur_time, start_time;
3739 	struct mpr_command *cm = *cmp;
3740 
3741 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
3742 		return  EBUSY;
3743 
3744 	cm->cm_complete = NULL;
3745 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
3746 	error = mpr_map_command(sc, cm);
3747 	if ((error != 0) && (error != EINPROGRESS))
3748 		return (error);
3749 
3750 	// Check for context and wait for 50 mSec at a time until time has
3751 	// expired or the command has finished.  If msleep can't be used, need
3752 	// to poll.
3753 #if __FreeBSD_version >= 1000029
3754 	if (curthread->td_no_sleeping)
3755 #else //__FreeBSD_version < 1000029
3756 	if (curthread->td_pflags & TDP_NOSLEEPING)
3757 #endif //__FreeBSD_version >= 1000029
3758 		sleep_flag = NO_SLEEP;
3759 	getmicrouptime(&start_time);
3760 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
3761 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
3762 		if (error == EWOULDBLOCK) {
3763 			/*
3764 			 * Record the actual elapsed time in the case of a
3765 			 * timeout for the message below.
3766 			 */
3767 			getmicrouptime(&cur_time);
3768 			timevalsub(&cur_time, &start_time);
3769 		}
3770 	} else {
3771 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3772 			mpr_intr_locked(sc);
3773 			if (sleep_flag == CAN_SLEEP)
3774 				pause("mprwait", hz/20);
3775 			else
3776 				DELAY(50000);
3777 
3778 			getmicrouptime(&cur_time);
3779 			timevalsub(&cur_time, &start_time);
3780 			if (cur_time.tv_sec > timeout) {
3781 				error = EWOULDBLOCK;
3782 				break;
3783 			}
3784 		}
3785 	}
3786 
3787 	if (error == EWOULDBLOCK) {
3788 		if (cm->cm_timeout_handler == NULL) {
3789 			mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s, timeout=%d,"
3790 			    " elapsed=%jd\n", __func__, timeout,
3791 			    (intmax_t)cur_time.tv_sec);
3792 			rc = mpr_reinit(sc);
3793 			mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3794 			    "failed");
3795 		} else
3796 			cm->cm_timeout_handler(sc, cm);
3797 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3798 			/*
3799 			 * Tell the caller that we freed the command in a
3800 			 * reinit.
3801 			 */
3802 			*cmp = NULL;
3803 		}
3804 		error = ETIMEDOUT;
3805 	}
3806 	return (error);
3807 }
3808 
3809 /*
3810  * This is the routine to enqueue a command synchonously and poll for
3811  * completion.  Its use should be rare.
3812  */
3813 int
3814 mpr_request_polled(struct mpr_softc *sc, struct mpr_command **cmp)
3815 {
3816 	int error, rc;
3817 	struct timeval cur_time, start_time;
3818 	struct mpr_command *cm = *cmp;
3819 
3820 	error = 0;
3821 
3822 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
3823 	cm->cm_complete = NULL;
3824 	mpr_map_command(sc, cm);
3825 
3826 	getmicrouptime(&start_time);
3827 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3828 		mpr_intr_locked(sc);
3829 
3830 		if (mtx_owned(&sc->mpr_mtx))
3831 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
3832 			    "mprpoll", hz/20);
3833 		else
3834 			pause("mprpoll", hz/20);
3835 
3836 		/*
3837 		 * Check for real-time timeout and fail if more than 60 seconds.
3838 		 */
3839 		getmicrouptime(&cur_time);
3840 		timevalsub(&cur_time, &start_time);
3841 		if (cur_time.tv_sec > 60) {
3842 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
3843 			error = ETIMEDOUT;
3844 			break;
3845 		}
3846 	}
3847 	cm->cm_state = MPR_CM_STATE_BUSY;
3848 	if (error) {
3849 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3850 		rc = mpr_reinit(sc);
3851 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3852 		    "failed");
3853 
3854 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3855 			/*
3856 			 * Tell the caller that we freed the command in a
3857 			 * reinit.
3858 			 */
3859 			*cmp = NULL;
3860 		}
3861 	}
3862 	return (error);
3863 }
3864 
3865 /*
3866  * The MPT driver had a verbose interface for config pages.  In this driver,
3867  * reduce it to much simpler terms, similar to the Linux driver.
3868  */
3869 int
3870 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3871 {
3872 	MPI2_CONFIG_REQUEST *req;
3873 	struct mpr_command *cm;
3874 	int error;
3875 
3876 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
3877 		return (EBUSY);
3878 	}
3879 
3880 	cm = mpr_alloc_command(sc);
3881 	if (cm == NULL) {
3882 		return (EBUSY);
3883 	}
3884 
3885 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3886 	req->Function = MPI2_FUNCTION_CONFIG;
3887 	req->Action = params->action;
3888 	req->SGLFlags = 0;
3889 	req->ChainOffset = 0;
3890 	req->PageAddress = params->page_address;
3891 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3892 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3893 
3894 		hdr = &params->hdr.Ext;
3895 		req->ExtPageType = hdr->ExtPageType;
3896 		req->ExtPageLength = hdr->ExtPageLength;
3897 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3898 		req->Header.PageLength = 0; /* Must be set to zero */
3899 		req->Header.PageNumber = hdr->PageNumber;
3900 		req->Header.PageVersion = hdr->PageVersion;
3901 	} else {
3902 		MPI2_CONFIG_PAGE_HEADER *hdr;
3903 
3904 		hdr = &params->hdr.Struct;
3905 		req->Header.PageType = hdr->PageType;
3906 		req->Header.PageNumber = hdr->PageNumber;
3907 		req->Header.PageLength = hdr->PageLength;
3908 		req->Header.PageVersion = hdr->PageVersion;
3909 	}
3910 
3911 	cm->cm_data = params->buffer;
3912 	cm->cm_length = params->length;
3913 	if (cm->cm_data != NULL) {
3914 		cm->cm_sge = &req->PageBufferSGE;
3915 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3916 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
3917 	} else
3918 		cm->cm_sge = NULL;
3919 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3920 
3921 	cm->cm_complete_data = params;
3922 	if (params->callback != NULL) {
3923 		cm->cm_complete = mpr_config_complete;
3924 		return (mpr_map_command(sc, cm));
3925 	} else {
3926 		error = mpr_wait_command(sc, &cm, 0, CAN_SLEEP);
3927 		if (error) {
3928 			mpr_dprint(sc, MPR_FAULT,
3929 			    "Error %d reading config page\n", error);
3930 			if (cm != NULL)
3931 				mpr_free_command(sc, cm);
3932 			return (error);
3933 		}
3934 		mpr_config_complete(sc, cm);
3935 	}
3936 
3937 	return (0);
3938 }
3939 
3940 int
3941 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3942 {
3943 	return (EINVAL);
3944 }
3945 
3946 static void
3947 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
3948 {
3949 	MPI2_CONFIG_REPLY *reply;
3950 	struct mpr_config_params *params;
3951 
3952 	MPR_FUNCTRACE(sc);
3953 	params = cm->cm_complete_data;
3954 
3955 	if (cm->cm_data != NULL) {
3956 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3957 		    BUS_DMASYNC_POSTREAD);
3958 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3959 	}
3960 
3961 	/*
3962 	 * XXX KDM need to do more error recovery?  This results in the
3963 	 * device in question not getting probed.
3964 	 */
3965 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
3966 		params->status = MPI2_IOCSTATUS_BUSY;
3967 		goto done;
3968 	}
3969 
3970 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3971 	if (reply == NULL) {
3972 		params->status = MPI2_IOCSTATUS_BUSY;
3973 		goto done;
3974 	}
3975 	params->status = reply->IOCStatus;
3976 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3977 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3978 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3979 		params->hdr.Ext.PageType = reply->Header.PageType;
3980 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3981 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3982 	} else {
3983 		params->hdr.Struct.PageType = reply->Header.PageType;
3984 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3985 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3986 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3987 	}
3988 
3989 done:
3990 	mpr_free_command(sc, cm);
3991 	if (params->callback != NULL)
3992 		params->callback(sc, params);
3993 
3994 	return;
3995 }
3996