xref: /freebsd/sys/dev/mpr/mpr.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2016 Avago Technologies
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * Avago Technologies (LSI) MPT-Fusion Host Adapter FreeBSD
29  *
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 /* Communications core for Avago Technologies (LSI) MPT3 */
36 
37 /* TODO Move headers to mprvar */
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/selinfo.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/bus.h>
47 #include <sys/conf.h>
48 #include <sys/bio.h>
49 #include <sys/malloc.h>
50 #include <sys/uio.h>
51 #include <sys/sysctl.h>
52 #include <sys/smp.h>
53 #include <sys/queue.h>
54 #include <sys/kthread.h>
55 #include <sys/taskqueue.h>
56 #include <sys/endian.h>
57 #include <sys/eventhandler.h>
58 #include <sys/sbuf.h>
59 
60 #include <machine/bus.h>
61 #include <machine/resource.h>
62 #include <sys/rman.h>
63 #include <sys/proc.h>
64 
65 #include <dev/pci/pcivar.h>
66 
67 #include <cam/cam.h>
68 #include <cam/cam_ccb.h>
69 #include <cam/scsi/scsi_all.h>
70 
71 #include <dev/mpr/mpi/mpi2_type.h>
72 #include <dev/mpr/mpi/mpi2.h>
73 #include <dev/mpr/mpi/mpi2_ioc.h>
74 #include <dev/mpr/mpi/mpi2_sas.h>
75 #include <dev/mpr/mpi/mpi2_pci.h>
76 #include <dev/mpr/mpi/mpi2_cnfg.h>
77 #include <dev/mpr/mpi/mpi2_init.h>
78 #include <dev/mpr/mpi/mpi2_tool.h>
79 #include <dev/mpr/mpr_ioctl.h>
80 #include <dev/mpr/mprvar.h>
81 #include <dev/mpr/mpr_table.h>
82 #include <dev/mpr/mpr_sas.h>
83 
84 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
85 static int mpr_init_queues(struct mpr_softc *sc);
86 static void mpr_resize_queues(struct mpr_softc *sc);
87 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
88 static int mpr_transition_operational(struct mpr_softc *sc);
89 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
90 static void mpr_iocfacts_free(struct mpr_softc *sc);
91 static void mpr_startup(void *arg);
92 static int mpr_send_iocinit(struct mpr_softc *sc);
93 static int mpr_alloc_queues(struct mpr_softc *sc);
94 static int mpr_alloc_hw_queues(struct mpr_softc *sc);
95 static int mpr_alloc_replies(struct mpr_softc *sc);
96 static int mpr_alloc_requests(struct mpr_softc *sc);
97 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc);
98 static int mpr_attach_log(struct mpr_softc *sc);
99 static __inline void mpr_complete_command(struct mpr_softc *sc,
100     struct mpr_command *cm);
101 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
102     MPI2_EVENT_NOTIFICATION_REPLY *reply);
103 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
104 static void mpr_periodic(void *);
105 static int mpr_reregister_events(struct mpr_softc *sc);
106 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
107 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
108 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
109 static int mpr_debug_sysctl(SYSCTL_HANDLER_ARGS);
110 static void mpr_parse_debug(struct mpr_softc *sc, char *list);
111 
112 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD, 0, "MPR Driver Parameters");
113 
114 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory");
115 
116 /*
117  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
118  * any state and back to its initialization state machine.
119  */
120 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
121 
122 /*
123  * Added this union to smoothly convert le64toh cm->cm_desc.Words.
124  * Compiler only supports uint64_t to be passed as an argument.
125  * Otherwise it will throw this error:
126  * "aggregate value used where an integer was expected"
127  */
128 typedef union _reply_descriptor {
129         u64 word;
130         struct {
131                 u32 low;
132                 u32 high;
133         } u;
134 } reply_descriptor, request_descriptor;
135 
136 /* Rate limit chain-fail messages to 1 per minute */
137 static struct timeval mpr_chainfail_interval = { 60, 0 };
138 
139 /*
140  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
141  * If this function is called from process context, it can sleep
142  * and there is no harm to sleep, in case if this fuction is called
143  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
144  * based on sleep flags driver will call either msleep, pause or DELAY.
145  * msleep and pause are of same variant, but pause is used when mpr_mtx
146  * is not hold by driver.
147  */
148 static int
149 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag)
150 {
151 	uint32_t reg;
152 	int i, error, tries = 0;
153 	uint8_t first_wait_done = FALSE;
154 
155 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
156 
157 	/* Clear any pending interrupts */
158 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
159 
160 	/*
161 	 * Force NO_SLEEP for threads prohibited to sleep
162  	 * e.a Thread from interrupt handler are prohibited to sleep.
163  	 */
164 #if __FreeBSD_version >= 1000029
165 	if (curthread->td_no_sleeping)
166 #else //__FreeBSD_version < 1000029
167 	if (curthread->td_pflags & TDP_NOSLEEPING)
168 #endif //__FreeBSD_version >= 1000029
169 		sleep_flag = NO_SLEEP;
170 
171 	mpr_dprint(sc, MPR_INIT, "sequence start, sleep_flag=%d\n", sleep_flag);
172 	/* Push the magic sequence */
173 	error = ETIMEDOUT;
174 	while (tries++ < 20) {
175 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
176 			mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
177 			    mpt2_reset_magic[i]);
178 
179 		/* wait 100 msec */
180 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
181 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
182 			    "mprdiag", hz/10);
183 		else if (sleep_flag == CAN_SLEEP)
184 			pause("mprdiag", hz/10);
185 		else
186 			DELAY(100 * 1000);
187 
188 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
189 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
190 			error = 0;
191 			break;
192 		}
193 	}
194 	if (error) {
195 		mpr_dprint(sc, MPR_INIT, "sequence failed, error=%d, exit\n",
196 		    error);
197 		return (error);
198 	}
199 
200 	/* Send the actual reset.  XXX need to refresh the reg? */
201 	reg |= MPI2_DIAG_RESET_ADAPTER;
202 	mpr_dprint(sc, MPR_INIT, "sequence success, sending reset, reg= 0x%x\n",
203 	    reg);
204 	mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg);
205 
206 	/* Wait up to 300 seconds in 50ms intervals */
207 	error = ETIMEDOUT;
208 	for (i = 0; i < 6000; i++) {
209 		/*
210 		 * Wait 50 msec. If this is the first time through, wait 256
211 		 * msec to satisfy Diag Reset timing requirements.
212 		 */
213 		if (first_wait_done) {
214 			if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
215 				msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
216 				    "mprdiag", hz/20);
217 			else if (sleep_flag == CAN_SLEEP)
218 				pause("mprdiag", hz/20);
219 			else
220 				DELAY(50 * 1000);
221 		} else {
222 			DELAY(256 * 1000);
223 			first_wait_done = TRUE;
224 		}
225 		/*
226 		 * Check for the RESET_ADAPTER bit to be cleared first, then
227 		 * wait for the RESET state to be cleared, which takes a little
228 		 * longer.
229 		 */
230 		reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
231 		if (reg & MPI2_DIAG_RESET_ADAPTER) {
232 			continue;
233 		}
234 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
235 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
236 			error = 0;
237 			break;
238 		}
239 	}
240 	if (error) {
241 		mpr_dprint(sc, MPR_INIT, "reset failed, error= %d, exit\n",
242 		    error);
243 		return (error);
244 	}
245 
246 	mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
247 	mpr_dprint(sc, MPR_INIT, "diag reset success, exit\n");
248 
249 	return (0);
250 }
251 
252 static int
253 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag)
254 {
255 	int error;
256 
257 	MPR_FUNCTRACE(sc);
258 
259 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
260 
261 	error = 0;
262 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
263 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
264 	    MPI2_DOORBELL_FUNCTION_SHIFT);
265 
266 	if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) {
267 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
268 		    "Doorbell handshake failed\n");
269 		error = ETIMEDOUT;
270 	}
271 
272 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
273 	return (error);
274 }
275 
276 static int
277 mpr_transition_ready(struct mpr_softc *sc)
278 {
279 	uint32_t reg, state;
280 	int error, tries = 0;
281 	int sleep_flags;
282 
283 	MPR_FUNCTRACE(sc);
284 	/* If we are in attach call, do not sleep */
285 	sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE)
286 	    ? CAN_SLEEP : NO_SLEEP;
287 
288 	error = 0;
289 
290 	mpr_dprint(sc, MPR_INIT, "%s entered, sleep_flags= %d\n",
291 	    __func__, sleep_flags);
292 
293 	while (tries++ < 1200) {
294 		reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
295 		mpr_dprint(sc, MPR_INIT, "  Doorbell= 0x%x\n", reg);
296 
297 		/*
298 		 * Ensure the IOC is ready to talk.  If it's not, try
299 		 * resetting it.
300 		 */
301 		if (reg & MPI2_DOORBELL_USED) {
302 			mpr_dprint(sc, MPR_INIT, "  Not ready, sending diag "
303 			    "reset\n");
304 			mpr_diag_reset(sc, sleep_flags);
305 			DELAY(50000);
306 			continue;
307 		}
308 
309 		/* Is the adapter owned by another peer? */
310 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
311 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
312 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC is under the "
313 			    "control of another peer host, aborting "
314 			    "initialization.\n");
315 			error = ENXIO;
316 			break;
317 		}
318 
319 		state = reg & MPI2_IOC_STATE_MASK;
320 		if (state == MPI2_IOC_STATE_READY) {
321 			/* Ready to go! */
322 			error = 0;
323 			break;
324 		} else if (state == MPI2_IOC_STATE_FAULT) {
325 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC in fault "
326 			    "state 0x%x, resetting\n",
327 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
328 			mpr_diag_reset(sc, sleep_flags);
329 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
330 			/* Need to take ownership */
331 			mpr_message_unit_reset(sc, sleep_flags);
332 		} else if (state == MPI2_IOC_STATE_RESET) {
333 			/* Wait a bit, IOC might be in transition */
334 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
335 			    "IOC in unexpected reset state\n");
336 		} else {
337 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
338 			    "IOC in unknown state 0x%x\n", state);
339 			error = EINVAL;
340 			break;
341 		}
342 
343 		/* Wait 50ms for things to settle down. */
344 		DELAY(50000);
345 	}
346 
347 	if (error)
348 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
349 		    "Cannot transition IOC to ready\n");
350 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
351 	return (error);
352 }
353 
354 static int
355 mpr_transition_operational(struct mpr_softc *sc)
356 {
357 	uint32_t reg, state;
358 	int error;
359 
360 	MPR_FUNCTRACE(sc);
361 
362 	error = 0;
363 	reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
364 	mpr_dprint(sc, MPR_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg);
365 
366 	state = reg & MPI2_IOC_STATE_MASK;
367 	if (state != MPI2_IOC_STATE_READY) {
368 		mpr_dprint(sc, MPR_INIT, "IOC not ready\n");
369 		if ((error = mpr_transition_ready(sc)) != 0) {
370 			mpr_dprint(sc, MPR_INIT|MPR_FAULT,
371 			    "failed to transition ready, exit\n");
372 			return (error);
373 		}
374 	}
375 
376 	error = mpr_send_iocinit(sc);
377 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
378 
379 	return (error);
380 }
381 
382 static void
383 mpr_resize_queues(struct mpr_softc *sc)
384 {
385 	u_int reqcr, prireqcr, maxio, sges_per_frame;
386 
387 	/*
388 	 * Size the queues. Since the reply queues always need one free
389 	 * entry, we'll deduct one reply message here.  The LSI documents
390 	 * suggest instead to add a count to the request queue, but I think
391 	 * that it's better to deduct from reply queue.
392 	 */
393 	prireqcr = MAX(1, sc->max_prireqframes);
394 	prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit);
395 
396 	reqcr = MAX(2, sc->max_reqframes);
397 	reqcr = MIN(reqcr, sc->facts->RequestCredit);
398 
399 	sc->num_reqs = prireqcr + reqcr;
400 	sc->num_prireqs = prireqcr;
401 	sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes,
402 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
403 
404 	/* Store the request frame size in bytes rather than as 32bit words */
405 	sc->reqframesz = sc->facts->IOCRequestFrameSize * 4;
406 
407 	/*
408 	 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to
409 	 * get the size of a Chain Frame.  Previous versions use the size as a
410 	 * Request Frame for the Chain Frame size.  If IOCMaxChainSegmentSize
411 	 * is 0, use the default value.  The IOCMaxChainSegmentSize is the
412 	 * number of 16-byte elelements that can fit in a Chain Frame, which is
413 	 * the size of an IEEE Simple SGE.
414 	 */
415 	if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) {
416 		sc->chain_seg_size =
417 		    htole16(sc->facts->IOCMaxChainSegmentSize);
418 		if (sc->chain_seg_size == 0) {
419 			sc->chain_frame_size = MPR_DEFAULT_CHAIN_SEG_SIZE *
420 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
421 		} else {
422 			sc->chain_frame_size = sc->chain_seg_size *
423 			    MPR_MAX_CHAIN_ELEMENT_SIZE;
424 		}
425 	} else {
426 		sc->chain_frame_size = sc->reqframesz;
427 	}
428 
429 	/*
430 	 * Max IO Size is Page Size * the following:
431 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
432 	 * + 1 for no chain needed in last frame
433 	 *
434 	 * If user suggests a Max IO size to use, use the smaller of the
435 	 * user's value and the calculated value as long as the user's
436 	 * value is larger than 0. The user's value is in pages.
437 	 */
438 	sges_per_frame = sc->chain_frame_size/sizeof(MPI2_IEEE_SGE_SIMPLE64)-1;
439 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
440 
441 	/*
442 	 * If I/O size limitation requested then use it and pass up to CAM.
443 	 * If not, use MAXPHYS as an optimization hint, but report HW limit.
444 	 */
445 	if (sc->max_io_pages > 0) {
446 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
447 		sc->maxio = maxio;
448 	} else {
449 		sc->maxio = maxio;
450 		maxio = min(maxio, MAXPHYS);
451 	}
452 
453 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
454 	    sges_per_frame * reqcr;
455 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
456 		sc->num_chains = sc->max_chains;
457 
458 	/*
459 	 * Figure out the number of MSIx-based queues.  If the firmware or
460 	 * user has done something crazy and not allowed enough credit for
461 	 * the queues to be useful then don't enable multi-queue.
462 	 */
463 	if (sc->facts->MaxMSIxVectors < 2)
464 		sc->msi_msgs = 1;
465 
466 	if (sc->msi_msgs > 1) {
467 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
468 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
469 		if (sc->num_reqs / sc->msi_msgs < 2)
470 			sc->msi_msgs = 1;
471 	}
472 
473 	mpr_dprint(sc, MPR_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
474 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
475 }
476 
477 /*
478  * This is called during attach and when re-initializing due to a Diag Reset.
479  * IOC Facts is used to allocate many of the structures needed by the driver.
480  * If called from attach, de-allocation is not required because the driver has
481  * not allocated any structures yet, but if called from a Diag Reset, previously
482  * allocated structures based on IOC Facts will need to be freed and re-
483  * allocated bases on the latest IOC Facts.
484  */
485 static int
486 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
487 {
488 	int error;
489 	Mpi2IOCFactsReply_t saved_facts;
490 	uint8_t saved_mode, reallocating;
491 
492 	mpr_dprint(sc, MPR_INIT|MPR_TRACE, "%s entered\n", __func__);
493 
494 	/* Save old IOC Facts and then only reallocate if Facts have changed */
495 	if (!attaching) {
496 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
497 	}
498 
499 	/*
500 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
501 	 * a re-initialization and only return the error if attaching so the OS
502 	 * can handle it.
503 	 */
504 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
505 		if (attaching) {
506 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to get "
507 			    "IOC Facts with error %d, exit\n", error);
508 			return (error);
509 		} else {
510 			panic("%s failed to get IOC Facts with error %d\n",
511 			    __func__, error);
512 		}
513 	}
514 
515 	MPR_DPRINT_PAGE(sc, MPR_XINFO, iocfacts, sc->facts);
516 
517 	snprintf(sc->fw_version, sizeof(sc->fw_version),
518 	    "%02d.%02d.%02d.%02d",
519 	    sc->facts->FWVersion.Struct.Major,
520 	    sc->facts->FWVersion.Struct.Minor,
521 	    sc->facts->FWVersion.Struct.Unit,
522 	    sc->facts->FWVersion.Struct.Dev);
523 
524 	mpr_dprint(sc, MPR_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
525 	    MPR_DRIVER_VERSION);
526 	mpr_dprint(sc, MPR_INFO,
527 	    "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
528 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
529 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
530 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc"
531 	    "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV");
532 
533 	/*
534 	 * If the chip doesn't support event replay then a hard reset will be
535 	 * required to trigger a full discovery.  Do the reset here then
536 	 * retransition to Ready.  A hard reset might have already been done,
537 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
538 	 * for a Diag Reset.
539 	 */
540 	if (attaching && ((sc->facts->IOCCapabilities &
541 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
542 		mpr_dprint(sc, MPR_INIT, "No event replay, resetting\n");
543 		mpr_diag_reset(sc, NO_SLEEP);
544 		if ((error = mpr_transition_ready(sc)) != 0) {
545 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
546 			    "transition to ready with error %d, exit\n",
547 			    error);
548 			return (error);
549 		}
550 	}
551 
552 	/*
553 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
554 	 * changed from the previous IOC Facts, log a warning, but only if
555 	 * checking this after a Diag Reset and not during attach.
556 	 */
557 	saved_mode = sc->ir_firmware;
558 	if (sc->facts->IOCCapabilities &
559 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
560 		sc->ir_firmware = 1;
561 	if (!attaching) {
562 		if (sc->ir_firmware != saved_mode) {
563 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "new IR/IT mode "
564 			    "in IOC Facts does not match previous mode\n");
565 		}
566 	}
567 
568 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
569 	reallocating = FALSE;
570 	sc->mpr_flags &= ~MPR_FLAGS_REALLOCATED;
571 
572 	if ((!attaching) &&
573 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
574 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
575 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
576 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
577 	    (saved_facts.ProductID != sc->facts->ProductID) ||
578 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
579 	    (saved_facts.IOCRequestFrameSize !=
580 	    sc->facts->IOCRequestFrameSize) ||
581 	    (saved_facts.IOCMaxChainSegmentSize !=
582 	    sc->facts->IOCMaxChainSegmentSize) ||
583 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
584 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
585 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
586 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
587 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
588 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
589 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
590 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
591 	    (saved_facts.MaxPersistentEntries !=
592 	    sc->facts->MaxPersistentEntries))) {
593 		reallocating = TRUE;
594 
595 		/* Record that we reallocated everything */
596 		sc->mpr_flags |= MPR_FLAGS_REALLOCATED;
597 	}
598 
599 	/*
600 	 * Some things should be done if attaching or re-allocating after a Diag
601 	 * Reset, but are not needed after a Diag Reset if the FW has not
602 	 * changed.
603 	 */
604 	if (attaching || reallocating) {
605 		/*
606 		 * Check if controller supports FW diag buffers and set flag to
607 		 * enable each type.
608 		 */
609 		if (sc->facts->IOCCapabilities &
610 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
611 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
612 			    enabled = TRUE;
613 		if (sc->facts->IOCCapabilities &
614 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
615 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
616 			    enabled = TRUE;
617 		if (sc->facts->IOCCapabilities &
618 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
619 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
620 			    enabled = TRUE;
621 
622 		/*
623 		 * Set flags for some supported items.
624 		 */
625 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
626 			sc->eedp_enabled = TRUE;
627 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
628 			sc->control_TLR = TRUE;
629 		if (sc->facts->IOCCapabilities &
630 		    MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ)
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_busaddr != 0)
770 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
771 	if (sc->chain_frames != NULL)
772 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
773 		    sc->chain_map);
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_addr_t queues_busaddr;
1315 	uint8_t *queues;
1316 	int qsize, fqsize, pqsize;
1317 
1318 	/*
1319 	 * The reply free queue contains 4 byte entries in multiples of 16 and
1320 	 * aligned on a 16 byte boundary. There must always be an unused entry.
1321 	 * This queue supplies fresh reply frames for the firmware to use.
1322 	 *
1323 	 * The reply descriptor post queue contains 8 byte entries in
1324 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
1325 	 * contains filled-in reply frames sent from the firmware to the host.
1326 	 *
1327 	 * These two queues are allocated together for simplicity.
1328 	 */
1329 	sc->fqdepth = roundup2(sc->num_replies + 1, 16);
1330 	sc->pqdepth = roundup2(sc->num_replies + 1, 16);
1331 	fqsize= sc->fqdepth * 4;
1332 	pqsize = sc->pqdepth * 8;
1333 	qsize = fqsize + pqsize;
1334 
1335         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1336 				16, 0,			/* algnmnt, boundary */
1337 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1338 				BUS_SPACE_MAXADDR,	/* highaddr */
1339 				NULL, NULL,		/* filter, filterarg */
1340                                 qsize,			/* maxsize */
1341                                 1,			/* nsegments */
1342                                 qsize,			/* maxsegsize */
1343                                 0,			/* flags */
1344                                 NULL, NULL,		/* lockfunc, lockarg */
1345                                 &sc->queues_dmat)) {
1346 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues DMA tag\n");
1347 		return (ENOMEM);
1348         }
1349         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1350 	    &sc->queues_map)) {
1351 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues memory\n");
1352 		return (ENOMEM);
1353         }
1354         bzero(queues, qsize);
1355         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1356 	    mpr_memaddr_cb, &queues_busaddr, 0);
1357 
1358 	sc->free_queue = (uint32_t *)queues;
1359 	sc->free_busaddr = queues_busaddr;
1360 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1361 	sc->post_busaddr = queues_busaddr + fqsize;
1362 	mpr_dprint(sc, MPR_INIT, "free queue busaddr= %#016jx size= %d\n",
1363 	    (uintmax_t)sc->free_busaddr, fqsize);
1364 	mpr_dprint(sc, MPR_INIT, "reply queue busaddr= %#016jx size= %d\n",
1365 	    (uintmax_t)sc->post_busaddr, pqsize);
1366 
1367 	return (0);
1368 }
1369 
1370 static int
1371 mpr_alloc_replies(struct mpr_softc *sc)
1372 {
1373 	int rsize, num_replies;
1374 
1375 	/* Store the reply frame size in bytes rather than as 32bit words */
1376 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1377 
1378 	/*
1379 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1380 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1381 	 * replies can be used at once.
1382 	 */
1383 	num_replies = max(sc->fqdepth, sc->num_replies);
1384 
1385 	rsize = sc->replyframesz * num_replies;
1386         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1387 				4, 0,			/* algnmnt, boundary */
1388 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1389 				BUS_SPACE_MAXADDR,	/* highaddr */
1390 				NULL, NULL,		/* filter, filterarg */
1391                                 rsize,			/* maxsize */
1392                                 1,			/* nsegments */
1393                                 rsize,			/* maxsegsize */
1394                                 0,			/* flags */
1395                                 NULL, NULL,		/* lockfunc, lockarg */
1396                                 &sc->reply_dmat)) {
1397 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies DMA tag\n");
1398 		return (ENOMEM);
1399         }
1400         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1401 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1402 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies memory\n");
1403 		return (ENOMEM);
1404         }
1405         bzero(sc->reply_frames, rsize);
1406         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1407 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1408 	mpr_dprint(sc, MPR_INIT, "reply frames busaddr= %#016jx size= %d\n",
1409 	    (uintmax_t)sc->reply_busaddr, rsize);
1410 
1411 	return (0);
1412 }
1413 
1414 static int
1415 mpr_alloc_requests(struct mpr_softc *sc)
1416 {
1417 	struct mpr_command *cm;
1418 	struct mpr_chain *chain;
1419 	int i, rsize, nsegs;
1420 
1421 	rsize = sc->reqframesz * sc->num_reqs;
1422         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1423 				16, 0,			/* algnmnt, boundary */
1424 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1425 				BUS_SPACE_MAXADDR,	/* highaddr */
1426 				NULL, NULL,		/* filter, filterarg */
1427                                 rsize,			/* maxsize */
1428                                 1,			/* nsegments */
1429                                 rsize,			/* maxsegsize */
1430                                 0,			/* flags */
1431                                 NULL, NULL,		/* lockfunc, lockarg */
1432                                 &sc->req_dmat)) {
1433 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request DMA tag\n");
1434 		return (ENOMEM);
1435         }
1436         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1437 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1438 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request memory\n");
1439 		return (ENOMEM);
1440         }
1441         bzero(sc->req_frames, rsize);
1442         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1443 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1444 	mpr_dprint(sc, MPR_INIT, "request frames busaddr= %#016jx size= %d\n",
1445 	    (uintmax_t)sc->req_busaddr, rsize);
1446 
1447 	rsize = sc->chain_frame_size * sc->num_chains;
1448         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1449 				16, 0,			/* algnmnt, boundary */
1450 				BUS_SPACE_MAXADDR,	/* lowaddr */
1451 				BUS_SPACE_MAXADDR,	/* highaddr */
1452 				NULL, NULL,		/* filter, filterarg */
1453                                 rsize,			/* maxsize */
1454                                 1,			/* nsegments */
1455                                 rsize,			/* maxsegsize */
1456                                 0,			/* flags */
1457                                 NULL, NULL,		/* lockfunc, lockarg */
1458                                 &sc->chain_dmat)) {
1459 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain DMA tag\n");
1460 		return (ENOMEM);
1461         }
1462         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1463 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
1464 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1465 		return (ENOMEM);
1466         }
1467         bzero(sc->chain_frames, rsize);
1468         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
1469 	    mpr_memaddr_cb, &sc->chain_busaddr, 0);
1470 	mpr_dprint(sc, MPR_INIT, "chain frames busaddr= %#016jx size= %d\n",
1471 	    (uintmax_t)sc->chain_busaddr, rsize);
1472 
1473 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1474 	if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1475 				1, 0,			/* algnmnt, boundary */
1476 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1477 				BUS_SPACE_MAXADDR,	/* highaddr */
1478 				NULL, NULL,		/* filter, filterarg */
1479                                 rsize,			/* maxsize */
1480                                 1,			/* nsegments */
1481                                 rsize,			/* maxsegsize */
1482                                 0,			/* flags */
1483                                 NULL, NULL,		/* lockfunc, lockarg */
1484                                 &sc->sense_dmat)) {
1485 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense DMA tag\n");
1486 		return (ENOMEM);
1487         }
1488         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1489 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1490 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense memory\n");
1491 		return (ENOMEM);
1492         }
1493         bzero(sc->sense_frames, rsize);
1494         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1495 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1496 	mpr_dprint(sc, MPR_INIT, "sense frames busaddr= %#016jx size= %d\n",
1497 	    (uintmax_t)sc->sense_busaddr, rsize);
1498 
1499 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->num_chains, M_MPR,
1500 	    M_WAITOK | M_ZERO);
1501 	if (!sc->chains) {
1502 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1503 		return (ENOMEM);
1504 	}
1505 	for (i = 0; i < sc->num_chains; i++) {
1506 		chain = &sc->chains[i];
1507 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
1508 		    i * sc->chain_frame_size);
1509 		chain->chain_busaddr = sc->chain_busaddr +
1510 		    i * sc->chain_frame_size;
1511 		mpr_free_chain(sc, chain);
1512 		sc->chain_free_lowwater++;
1513 	}
1514 
1515 	/*
1516 	 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports
1517 	 * these devices.
1518 	 */
1519 	if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) &&
1520 	    (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) {
1521 		if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM)
1522 			return (ENOMEM);
1523 	}
1524 
1525 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1526         if (bus_dma_tag_create( sc->mpr_parent_dmat,    /* parent */
1527 				1, 0,			/* algnmnt, boundary */
1528 				BUS_SPACE_MAXADDR,	/* lowaddr */
1529 				BUS_SPACE_MAXADDR,	/* highaddr */
1530 				NULL, NULL,		/* filter, filterarg */
1531                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
1532                                 nsegs,			/* nsegments */
1533                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
1534                                 BUS_DMA_ALLOCNOW,	/* flags */
1535                                 busdma_lock_mutex,	/* lockfunc */
1536 				&sc->mpr_mtx,		/* lockarg */
1537                                 &sc->buffer_dmat)) {
1538 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate buffer DMA tag\n");
1539 		return (ENOMEM);
1540         }
1541 
1542 	/*
1543 	 * SMID 0 cannot be used as a free command per the firmware spec.
1544 	 * Just drop that command instead of risking accounting bugs.
1545 	 */
1546 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1547 	    M_MPR, M_WAITOK | M_ZERO);
1548 	if (!sc->commands) {
1549 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate command memory\n");
1550 		return (ENOMEM);
1551 	}
1552 	for (i = 1; i < sc->num_reqs; i++) {
1553 		cm = &sc->commands[i];
1554 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1555 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1556 		cm->cm_sense = &sc->sense_frames[i];
1557 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1558 		cm->cm_desc.Default.SMID = i;
1559 		cm->cm_sc = sc;
1560 		cm->cm_state = MPR_CM_STATE_BUSY;
1561 		TAILQ_INIT(&cm->cm_chain_list);
1562 		TAILQ_INIT(&cm->cm_prp_page_list);
1563 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1564 
1565 		/* XXX Is a failure here a critical problem? */
1566 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap)
1567 		    == 0) {
1568 			if (i <= sc->num_prireqs)
1569 				mpr_free_high_priority_command(sc, cm);
1570 			else
1571 				mpr_free_command(sc, cm);
1572 		} else {
1573 			panic("failed to allocate command %d\n", i);
1574 			sc->num_reqs = i;
1575 			break;
1576 		}
1577 	}
1578 
1579 	return (0);
1580 }
1581 
1582 /*
1583  * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs,
1584  * which are scatter/gather lists for NVMe devices.
1585  *
1586  * This buffer must be contiguous due to the nature of how NVMe PRPs are built
1587  * and translated by FW.
1588  *
1589  * returns ENOMEM if memory could not be allocated, otherwise returns 0.
1590  */
1591 static int
1592 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc)
1593 {
1594 	int PRPs_per_page, PRPs_required, pages_required;
1595 	int rsize, i;
1596 	struct mpr_prp_page *prp_page;
1597 
1598 	/*
1599 	 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number
1600 	 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is:
1601 	 * MAX_IO_SIZE / PAGE_SIZE = 256
1602 	 *
1603 	 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs
1604 	 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one
1605 	 * page (4096 / 8 = 512), so only one page is required for each I/O.
1606 	 *
1607 	 * Each of these buffers will need to be contiguous. For simplicity,
1608 	 * only one buffer is allocated here, which has all of the space
1609 	 * required for the NVMe Queue Depth. If there are problems allocating
1610 	 * this one buffer, this function will need to change to allocate
1611 	 * individual, contiguous NVME_QDEPTH buffers.
1612 	 *
1613 	 * The real calculation will use the real max io size. Above is just an
1614 	 * example.
1615 	 *
1616 	 */
1617 	PRPs_required = sc->maxio / PAGE_SIZE;
1618 	PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1;
1619 	pages_required = (PRPs_required / PRPs_per_page) + 1;
1620 
1621 	sc->prp_buffer_size = PAGE_SIZE * pages_required;
1622 	rsize = sc->prp_buffer_size * NVME_QDEPTH;
1623 	if (bus_dma_tag_create( sc->mpr_parent_dmat,	/* parent */
1624 				4, 0,			/* algnmnt, boundary */
1625 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
1626 				BUS_SPACE_MAXADDR,	/* highaddr */
1627 				NULL, NULL,		/* filter, filterarg */
1628 				rsize,			/* maxsize */
1629 				1,			/* nsegments */
1630 				rsize,			/* maxsegsize */
1631 				0,			/* flags */
1632 				NULL, NULL,		/* lockfunc, lockarg */
1633 				&sc->prp_page_dmat)) {
1634 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP DMA "
1635 		    "tag\n");
1636 		return (ENOMEM);
1637 	}
1638 	if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages,
1639 	    BUS_DMA_NOWAIT, &sc->prp_page_map)) {
1640 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP memory\n");
1641 		return (ENOMEM);
1642 	}
1643 	bzero(sc->prp_pages, rsize);
1644 	bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages,
1645 	    rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0);
1646 
1647 	sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR,
1648 	    M_WAITOK | M_ZERO);
1649 	for (i = 0; i < NVME_QDEPTH; i++) {
1650 		prp_page = &sc->prps[i];
1651 		prp_page->prp_page = (uint64_t *)(sc->prp_pages +
1652 		    i * sc->prp_buffer_size);
1653 		prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr +
1654 		    i * sc->prp_buffer_size);
1655 		mpr_free_prp_page(sc, prp_page);
1656 		sc->prp_pages_free_lowwater++;
1657 	}
1658 
1659 	return (0);
1660 }
1661 
1662 static int
1663 mpr_init_queues(struct mpr_softc *sc)
1664 {
1665 	int i;
1666 
1667 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1668 
1669 	/*
1670 	 * According to the spec, we need to use one less reply than we
1671 	 * have space for on the queue.  So sc->num_replies (the number we
1672 	 * use) should be less than sc->fqdepth (allocated size).
1673 	 */
1674 	if (sc->num_replies >= sc->fqdepth)
1675 		return (EINVAL);
1676 
1677 	/*
1678 	 * Initialize all of the free queue entries.
1679 	 */
1680 	for (i = 0; i < sc->fqdepth; i++) {
1681 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1682 	}
1683 	sc->replyfreeindex = sc->num_replies;
1684 
1685 	return (0);
1686 }
1687 
1688 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1689  * Next are the global settings, if they exist.  Highest are the per-unit
1690  * settings, if they exist.
1691  */
1692 void
1693 mpr_get_tunables(struct mpr_softc *sc)
1694 {
1695 	char tmpstr[80], mpr_debug[80];
1696 
1697 	/* XXX default to some debugging for now */
1698 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1699 	sc->disable_msix = 0;
1700 	sc->disable_msi = 0;
1701 	sc->max_msix = MPR_MSIX_MAX;
1702 	sc->max_chains = MPR_CHAIN_FRAMES;
1703 	sc->max_io_pages = MPR_MAXIO_PAGES;
1704 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1705 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1706 	sc->use_phynum = 1;
1707 	sc->max_reqframes = MPR_REQ_FRAMES;
1708 	sc->max_prireqframes = MPR_PRI_REQ_FRAMES;
1709 	sc->max_replyframes = MPR_REPLY_FRAMES;
1710 	sc->max_evtframes = MPR_EVT_REPLY_FRAMES;
1711 
1712 	/*
1713 	 * Grab the global variables.
1714 	 */
1715 	bzero(mpr_debug, 80);
1716 	if (TUNABLE_STR_FETCH("hw.mpr.debug_level", mpr_debug, 80) != 0)
1717 		mpr_parse_debug(sc, mpr_debug);
1718 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1719 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1720 	TUNABLE_INT_FETCH("hw.mpr.max_msix", &sc->max_msix);
1721 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1722 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1723 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1724 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1725 	TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum);
1726 	TUNABLE_INT_FETCH("hw.mpr.max_reqframes", &sc->max_reqframes);
1727 	TUNABLE_INT_FETCH("hw.mpr.max_prireqframes", &sc->max_prireqframes);
1728 	TUNABLE_INT_FETCH("hw.mpr.max_replyframes", &sc->max_replyframes);
1729 	TUNABLE_INT_FETCH("hw.mpr.max_evtframes", &sc->max_evtframes);
1730 
1731 	/* Grab the unit-instance variables */
1732 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1733 	    device_get_unit(sc->mpr_dev));
1734 	bzero(mpr_debug, 80);
1735 	if (TUNABLE_STR_FETCH(tmpstr, mpr_debug, 80) != 0)
1736 		mpr_parse_debug(sc, mpr_debug);
1737 
1738 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1739 	    device_get_unit(sc->mpr_dev));
1740 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1741 
1742 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1743 	    device_get_unit(sc->mpr_dev));
1744 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1745 
1746 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_msix",
1747 	    device_get_unit(sc->mpr_dev));
1748 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1749 
1750 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1751 	    device_get_unit(sc->mpr_dev));
1752 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1753 
1754 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1755 	    device_get_unit(sc->mpr_dev));
1756 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1757 
1758 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1759 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1760 	    device_get_unit(sc->mpr_dev));
1761 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1762 
1763 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1764 	    device_get_unit(sc->mpr_dev));
1765 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1766 
1767 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1768 	    device_get_unit(sc->mpr_dev));
1769 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1770 
1771 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num",
1772 	    device_get_unit(sc->mpr_dev));
1773 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1774 
1775 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_reqframes",
1776 	    device_get_unit(sc->mpr_dev));
1777 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1778 
1779 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_prireqframes",
1780 	    device_get_unit(sc->mpr_dev));
1781 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1782 
1783 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_replyframes",
1784 	    device_get_unit(sc->mpr_dev));
1785 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1786 
1787 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_evtframes",
1788 	    device_get_unit(sc->mpr_dev));
1789 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1790 }
1791 
1792 static void
1793 mpr_setup_sysctl(struct mpr_softc *sc)
1794 {
1795 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1796 	struct sysctl_oid	*sysctl_tree = NULL;
1797 	char tmpstr[80], tmpstr2[80];
1798 
1799 	/*
1800 	 * Setup the sysctl variable so the user can change the debug level
1801 	 * on the fly.
1802 	 */
1803 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1804 	    device_get_unit(sc->mpr_dev));
1805 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1806 
1807 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1808 	if (sysctl_ctx != NULL)
1809 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1810 
1811 	if (sysctl_tree == NULL) {
1812 		sysctl_ctx_init(&sc->sysctl_ctx);
1813 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1814 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1815 		    CTLFLAG_RD, 0, tmpstr);
1816 		if (sc->sysctl_tree == NULL)
1817 			return;
1818 		sysctl_ctx = &sc->sysctl_ctx;
1819 		sysctl_tree = sc->sysctl_tree;
1820 	}
1821 
1822 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1823 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1824 	    sc, 0, mpr_debug_sysctl, "A", "mpr debug level");
1825 
1826 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1827 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1828 	    "Disable the use of MSI-X interrupts");
1829 
1830 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1831 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1832 	    "User-defined maximum number of MSIX queues");
1833 
1834 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1835 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1836 	    "Negotiated number of MSIX queues");
1837 
1838 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1839 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1840 	    "Total number of allocated request frames");
1841 
1842 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1843 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1844 	    "Total number of allocated high priority request frames");
1845 
1846 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1847 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1848 	    "Total number of allocated reply frames");
1849 
1850 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1851 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1852 	    "Total number of event frames allocated");
1853 
1854 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1855 	    OID_AUTO, "firmware_version", CTLFLAG_RW, sc->fw_version,
1856 	    strlen(sc->fw_version), "firmware version");
1857 
1858 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1859 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPR_DRIVER_VERSION,
1860 	    strlen(MPR_DRIVER_VERSION), "driver version");
1861 
1862 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1863 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1864 	    &sc->io_cmds_active, 0, "number of currently active commands");
1865 
1866 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1867 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1868 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1869 
1870 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1871 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1872 	    &sc->chain_free, 0, "number of free chain elements");
1873 
1874 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1875 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1876 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1877 
1878 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1879 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1880 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1881 
1882 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1883 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1884 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1885 	    "IOCFacts)");
1886 
1887 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1888 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1889 	    "enable SSU to SATA SSD/HDD at shutdown");
1890 
1891 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1892 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1893 	    &sc->chain_alloc_fail, "chain allocation failures");
1894 
1895 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1896 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1897 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1898 	    "spinup after SATA ID error");
1899 
1900 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1901 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1902 	    "Use the phy number for enumeration");
1903 
1904 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1905 	    OID_AUTO, "prp_pages_free", CTLFLAG_RD,
1906 	    &sc->prp_pages_free, 0, "number of free PRP pages");
1907 
1908 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1909 	    OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD,
1910 	    &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages");
1911 
1912 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1913 	    OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD,
1914 	    &sc->prp_page_alloc_fail, "PRP page allocation failures");
1915 }
1916 
1917 static struct mpr_debug_string {
1918 	char *name;
1919 	int flag;
1920 } mpr_debug_strings[] = {
1921 	{"info", MPR_INFO},
1922 	{"fault", MPR_FAULT},
1923 	{"event", MPR_EVENT},
1924 	{"log", MPR_LOG},
1925 	{"recovery", MPR_RECOVERY},
1926 	{"error", MPR_ERROR},
1927 	{"init", MPR_INIT},
1928 	{"xinfo", MPR_XINFO},
1929 	{"user", MPR_USER},
1930 	{"mapping", MPR_MAPPING},
1931 	{"trace", MPR_TRACE}
1932 };
1933 
1934 enum mpr_debug_level_combiner {
1935 	COMB_NONE,
1936 	COMB_ADD,
1937 	COMB_SUB
1938 };
1939 
1940 static int
1941 mpr_debug_sysctl(SYSCTL_HANDLER_ARGS)
1942 {
1943 	struct mpr_softc *sc;
1944 	struct mpr_debug_string *string;
1945 	struct sbuf *sbuf;
1946 	char *buffer;
1947 	size_t sz;
1948 	int i, len, debug, error;
1949 
1950 	sc = (struct mpr_softc *)arg1;
1951 
1952 	error = sysctl_wire_old_buffer(req, 0);
1953 	if (error != 0)
1954 		return (error);
1955 
1956 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1957 	debug = sc->mpr_debug;
1958 
1959 	sbuf_printf(sbuf, "%#x", debug);
1960 
1961 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1962 	for (i = 0; i < sz; i++) {
1963 		string = &mpr_debug_strings[i];
1964 		if (debug & string->flag)
1965 			sbuf_printf(sbuf, ",%s", string->name);
1966 	}
1967 
1968 	error = sbuf_finish(sbuf);
1969 	sbuf_delete(sbuf);
1970 
1971 	if (error || req->newptr == NULL)
1972 		return (error);
1973 
1974 	len = req->newlen - req->newidx;
1975 	if (len == 0)
1976 		return (0);
1977 
1978 	buffer = malloc(len, M_MPR, M_ZERO|M_WAITOK);
1979 	error = SYSCTL_IN(req, buffer, len);
1980 
1981 	mpr_parse_debug(sc, buffer);
1982 
1983 	free(buffer, M_MPR);
1984 	return (error);
1985 }
1986 
1987 static void
1988 mpr_parse_debug(struct mpr_softc *sc, char *list)
1989 {
1990 	struct mpr_debug_string *string;
1991 	enum mpr_debug_level_combiner op;
1992 	char *token, *endtoken;
1993 	size_t sz;
1994 	int flags, i;
1995 
1996 	if (list == NULL || *list == '\0')
1997 		return;
1998 
1999 	if (*list == '+') {
2000 		op = COMB_ADD;
2001 		list++;
2002 	} else if (*list == '-') {
2003 		op = COMB_SUB;
2004 		list++;
2005 	} else
2006 		op = COMB_NONE;
2007 	if (*list == '\0')
2008 		return;
2009 
2010 	flags = 0;
2011 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
2012 	while ((token = strsep(&list, ":,")) != NULL) {
2013 
2014 		/* Handle integer flags */
2015 		flags |= strtol(token, &endtoken, 0);
2016 		if (token != endtoken)
2017 			continue;
2018 
2019 		/* Handle text flags */
2020 		for (i = 0; i < sz; i++) {
2021 			string = &mpr_debug_strings[i];
2022 			if (strcasecmp(token, string->name) == 0) {
2023 				flags |= string->flag;
2024 				break;
2025 			}
2026 		}
2027 	}
2028 
2029 	switch (op) {
2030 	case COMB_NONE:
2031 		sc->mpr_debug = flags;
2032 		break;
2033 	case COMB_ADD:
2034 		sc->mpr_debug |= flags;
2035 		break;
2036 	case COMB_SUB:
2037 		sc->mpr_debug &= (~flags);
2038 		break;
2039 	}
2040 	return;
2041 }
2042 
2043 int
2044 mpr_attach(struct mpr_softc *sc)
2045 {
2046 	int error;
2047 
2048 	MPR_FUNCTRACE(sc);
2049 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2050 
2051 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
2052 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
2053 	callout_init_mtx(&sc->device_check_callout, &sc->mpr_mtx, 0);
2054 	TAILQ_INIT(&sc->event_list);
2055 	timevalclear(&sc->lastfail);
2056 
2057 	if ((error = mpr_transition_ready(sc)) != 0) {
2058 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2059 		    "Failed to transition ready\n");
2060 		return (error);
2061 	}
2062 
2063 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
2064 	    M_ZERO|M_NOWAIT);
2065 	if (!sc->facts) {
2066 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2067 		    "Cannot allocate memory, exit\n");
2068 		return (ENOMEM);
2069 	}
2070 
2071 	/*
2072 	 * Get IOC Facts and allocate all structures based on this information.
2073 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
2074 	 * Facts. If relevant values have changed in IOC Facts, this function
2075 	 * will free all of the memory based on IOC Facts and reallocate that
2076 	 * memory.  If this fails, any allocated memory should already be freed.
2077 	 */
2078 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
2079 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC Facts allocation "
2080 		    "failed with error %d\n", error);
2081 		return (error);
2082 	}
2083 
2084 	/* Start the periodic watchdog check on the IOC Doorbell */
2085 	mpr_periodic(sc);
2086 
2087 	/*
2088 	 * The portenable will kick off discovery events that will drive the
2089 	 * rest of the initialization process.  The CAM/SAS module will
2090 	 * hold up the boot sequence until discovery is complete.
2091 	 */
2092 	sc->mpr_ich.ich_func = mpr_startup;
2093 	sc->mpr_ich.ich_arg = sc;
2094 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
2095 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2096 		    "Cannot establish MPR config hook\n");
2097 		error = EINVAL;
2098 	}
2099 
2100 	/*
2101 	 * Allow IR to shutdown gracefully when shutdown occurs.
2102 	 */
2103 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
2104 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
2105 
2106 	if (sc->shutdown_eh == NULL)
2107 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2108 		    "shutdown event registration failed\n");
2109 
2110 	mpr_setup_sysctl(sc);
2111 
2112 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
2113 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
2114 
2115 	return (error);
2116 }
2117 
2118 /* Run through any late-start handlers. */
2119 static void
2120 mpr_startup(void *arg)
2121 {
2122 	struct mpr_softc *sc;
2123 
2124 	sc = (struct mpr_softc *)arg;
2125 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2126 
2127 	mpr_lock(sc);
2128 	mpr_unmask_intr(sc);
2129 
2130 	/* initialize device mapping tables */
2131 	mpr_base_static_config_pages(sc);
2132 	mpr_mapping_initialize(sc);
2133 	mprsas_startup(sc);
2134 	mpr_unlock(sc);
2135 
2136 	mpr_dprint(sc, MPR_INIT, "disestablish config intrhook\n");
2137 	config_intrhook_disestablish(&sc->mpr_ich);
2138 	sc->mpr_ich.ich_arg = NULL;
2139 
2140 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2141 }
2142 
2143 /* Periodic watchdog.  Is called with the driver lock already held. */
2144 static void
2145 mpr_periodic(void *arg)
2146 {
2147 	struct mpr_softc *sc;
2148 	uint32_t db;
2149 
2150 	sc = (struct mpr_softc *)arg;
2151 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
2152 		return;
2153 
2154 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
2155 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2156 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
2157 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
2158 			panic("TEMPERATURE FAULT: STOPPING.");
2159 		}
2160 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2161 		mpr_reinit(sc);
2162 	}
2163 
2164 	callout_reset(&sc->periodic, MPR_PERIODIC_DELAY * hz, mpr_periodic, sc);
2165 }
2166 
2167 static void
2168 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
2169     MPI2_EVENT_NOTIFICATION_REPLY *event)
2170 {
2171 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2172 
2173 	MPR_DPRINT_EVENT(sc, generic, event);
2174 
2175 	switch (event->Event) {
2176 	case MPI2_EVENT_LOG_DATA:
2177 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2178 		if (sc->mpr_debug & MPR_EVENT)
2179 			hexdump(event->EventData, event->EventDataLength, NULL,
2180 			    0);
2181 		break;
2182 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2183 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2184 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2185 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2186 		     entry->LogSequence);
2187 		break;
2188 	default:
2189 		break;
2190 	}
2191 	return;
2192 }
2193 
2194 static int
2195 mpr_attach_log(struct mpr_softc *sc)
2196 {
2197 	uint8_t events[16];
2198 
2199 	bzero(events, 16);
2200 	setbit(events, MPI2_EVENT_LOG_DATA);
2201 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2202 
2203 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
2204 	    &sc->mpr_log_eh);
2205 
2206 	return (0);
2207 }
2208 
2209 static int
2210 mpr_detach_log(struct mpr_softc *sc)
2211 {
2212 
2213 	if (sc->mpr_log_eh != NULL)
2214 		mpr_deregister_events(sc, sc->mpr_log_eh);
2215 	return (0);
2216 }
2217 
2218 /*
2219  * Free all of the driver resources and detach submodules.  Should be called
2220  * without the lock held.
2221  */
2222 int
2223 mpr_free(struct mpr_softc *sc)
2224 {
2225 	int error;
2226 
2227 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2228 	/* Turn off the watchdog */
2229 	mpr_lock(sc);
2230 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
2231 	mpr_unlock(sc);
2232 	/* Lock must not be held for this */
2233 	callout_drain(&sc->periodic);
2234 	callout_drain(&sc->device_check_callout);
2235 
2236 	if (((error = mpr_detach_log(sc)) != 0) ||
2237 	    ((error = mpr_detach_sas(sc)) != 0)) {
2238 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "failed to detach "
2239 		    "subsystems, error= %d, exit\n", error);
2240 		return (error);
2241 	}
2242 
2243 	mpr_detach_user(sc);
2244 
2245 	/* Put the IOC back in the READY state. */
2246 	mpr_lock(sc);
2247 	if ((error = mpr_transition_ready(sc)) != 0) {
2248 		mpr_unlock(sc);
2249 		return (error);
2250 	}
2251 	mpr_unlock(sc);
2252 
2253 	if (sc->facts != NULL)
2254 		free(sc->facts, M_MPR);
2255 
2256 	/*
2257 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2258 	 * to free these buffers too.
2259 	 */
2260 	mpr_iocfacts_free(sc);
2261 
2262 	if (sc->sysctl_tree != NULL)
2263 		sysctl_ctx_free(&sc->sysctl_ctx);
2264 
2265 	/* Deregister the shutdown function */
2266 	if (sc->shutdown_eh != NULL)
2267 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2268 
2269 	mtx_destroy(&sc->mpr_mtx);
2270 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2271 
2272 	return (0);
2273 }
2274 
2275 static __inline void
2276 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
2277 {
2278 	MPR_FUNCTRACE(sc);
2279 
2280 	if (cm == NULL) {
2281 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
2282 		return;
2283 	}
2284 
2285 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
2286 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
2287 
2288 	if (cm->cm_complete != NULL) {
2289 		mpr_dprint(sc, MPR_TRACE,
2290 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
2291 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
2292 		    cm->cm_reply);
2293 		cm->cm_complete(sc, cm);
2294 	}
2295 
2296 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
2297 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
2298 		wakeup(cm);
2299 	}
2300 
2301 	if (sc->io_cmds_active != 0) {
2302 		sc->io_cmds_active--;
2303 	} else {
2304 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
2305 		    "out of sync - resynching to 0\n");
2306 	}
2307 }
2308 
2309 static void
2310 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
2311 {
2312 	union loginfo_type {
2313 		u32	loginfo;
2314 		struct {
2315 			u32	subcode:16;
2316 			u32	code:8;
2317 			u32	originator:4;
2318 			u32	bus_type:4;
2319 		} dw;
2320 	};
2321 	union loginfo_type sas_loginfo;
2322 	char *originator_str = NULL;
2323 
2324 	sas_loginfo.loginfo = log_info;
2325 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2326 		return;
2327 
2328 	/* each nexus loss loginfo */
2329 	if (log_info == 0x31170000)
2330 		return;
2331 
2332 	/* eat the loginfos associated with task aborts */
2333 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
2334 	    (log_info == 0x31130000))
2335 		return;
2336 
2337 	switch (sas_loginfo.dw.originator) {
2338 	case 0:
2339 		originator_str = "IOP";
2340 		break;
2341 	case 1:
2342 		originator_str = "PL";
2343 		break;
2344 	case 2:
2345 		originator_str = "IR";
2346 		break;
2347 	}
2348 
2349 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
2350 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
2351 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
2352 }
2353 
2354 static void
2355 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
2356 {
2357 	MPI2DefaultReply_t *mpi_reply;
2358 	u16 sc_status;
2359 
2360 	mpi_reply = (MPI2DefaultReply_t*)reply;
2361 	sc_status = le16toh(mpi_reply->IOCStatus);
2362 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2363 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2364 }
2365 
2366 void
2367 mpr_intr(void *data)
2368 {
2369 	struct mpr_softc *sc;
2370 	uint32_t status;
2371 
2372 	sc = (struct mpr_softc *)data;
2373 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2374 
2375 	/*
2376 	 * Check interrupt status register to flush the bus.  This is
2377 	 * needed for both INTx interrupts and driver-driven polling
2378 	 */
2379 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2380 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2381 		return;
2382 
2383 	mpr_lock(sc);
2384 	mpr_intr_locked(data);
2385 	mpr_unlock(sc);
2386 	return;
2387 }
2388 
2389 /*
2390  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2391  * chip.  Hopefully this theory is correct.
2392  */
2393 void
2394 mpr_intr_msi(void *data)
2395 {
2396 	struct mpr_softc *sc;
2397 
2398 	sc = (struct mpr_softc *)data;
2399 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2400 	mpr_lock(sc);
2401 	mpr_intr_locked(data);
2402 	mpr_unlock(sc);
2403 	return;
2404 }
2405 
2406 /*
2407  * The locking is overly broad and simplistic, but easy to deal with for now.
2408  */
2409 void
2410 mpr_intr_locked(void *data)
2411 {
2412 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2413 	struct mpr_softc *sc;
2414 	struct mpr_command *cm = NULL;
2415 	uint8_t flags;
2416 	u_int pq;
2417 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2418 	mpr_fw_diagnostic_buffer_t *pBuffer;
2419 
2420 	sc = (struct mpr_softc *)data;
2421 
2422 	pq = sc->replypostindex;
2423 	mpr_dprint(sc, MPR_TRACE,
2424 	    "%s sc %p starting with replypostindex %u\n",
2425 	    __func__, sc, sc->replypostindex);
2426 
2427 	for ( ;; ) {
2428 		cm = NULL;
2429 		desc = &sc->post_queue[sc->replypostindex];
2430 		flags = desc->Default.ReplyFlags &
2431 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2432 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
2433 		    (le32toh(desc->Words.High) == 0xffffffff))
2434 			break;
2435 
2436 		/* increment the replypostindex now, so that event handlers
2437 		 * and cm completion handlers which decide to do a diag
2438 		 * reset can zero it without it getting incremented again
2439 		 * afterwards, and we break out of this loop on the next
2440 		 * iteration since the reply post queue has been cleared to
2441 		 * 0xFF and all descriptors look unused (which they are).
2442 		 */
2443 		if (++sc->replypostindex >= sc->pqdepth)
2444 			sc->replypostindex = 0;
2445 
2446 		switch (flags) {
2447 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2448 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
2449 		case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS:
2450 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2451 			KASSERT(cm->cm_state == MPR_CM_STATE_INQUEUE,
2452 			    ("command not inqueue\n"));
2453 			cm->cm_state = MPR_CM_STATE_BUSY;
2454 			cm->cm_reply = NULL;
2455 			break;
2456 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2457 		{
2458 			uint32_t baddr;
2459 			uint8_t *reply;
2460 
2461 			/*
2462 			 * Re-compose the reply address from the address
2463 			 * sent back from the chip.  The ReplyFrameAddress
2464 			 * is the lower 32 bits of the physical address of
2465 			 * particular reply frame.  Convert that address to
2466 			 * host format, and then use that to provide the
2467 			 * offset against the virtual address base
2468 			 * (sc->reply_frames).
2469 			 */
2470 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2471 			reply = sc->reply_frames +
2472 				(baddr - ((uint32_t)sc->reply_busaddr));
2473 			/*
2474 			 * Make sure the reply we got back is in a valid
2475 			 * range.  If not, go ahead and panic here, since
2476 			 * we'll probably panic as soon as we deference the
2477 			 * reply pointer anyway.
2478 			 */
2479 			if ((reply < sc->reply_frames)
2480 			 || (reply > (sc->reply_frames +
2481 			     (sc->fqdepth * sc->replyframesz)))) {
2482 				printf("%s: WARNING: reply %p out of range!\n",
2483 				       __func__, reply);
2484 				printf("%s: reply_frames %p, fqdepth %d, "
2485 				       "frame size %d\n", __func__,
2486 				       sc->reply_frames, sc->fqdepth,
2487 				       sc->replyframesz);
2488 				printf("%s: baddr %#x,\n", __func__, baddr);
2489 				/* LSI-TODO. See Linux Code for Graceful exit */
2490 				panic("Reply address out of range");
2491 			}
2492 			if (le16toh(desc->AddressReply.SMID) == 0) {
2493 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2494 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2495 					/*
2496 					 * If SMID is 0 for Diag Buffer Post,
2497 					 * this implies that the reply is due to
2498 					 * a release function with a status that
2499 					 * the buffer has been released.  Set
2500 					 * the buffer flags accordingly.
2501 					 */
2502 					rel_rep =
2503 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2504 					if ((le16toh(rel_rep->IOCStatus) &
2505 					    MPI2_IOCSTATUS_MASK) ==
2506 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2507 					{
2508 						pBuffer =
2509 						    &sc->fw_diag_buffer_list[
2510 						    rel_rep->BufferType];
2511 						pBuffer->valid_data = TRUE;
2512 						pBuffer->owned_by_firmware =
2513 						    FALSE;
2514 						pBuffer->immediate = FALSE;
2515 					}
2516 				} else
2517 					mpr_dispatch_event(sc, baddr,
2518 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2519 					    reply);
2520 			} else {
2521 				cm = &sc->commands[
2522 				    le16toh(desc->AddressReply.SMID)];
2523 				KASSERT(cm->cm_state == MPR_CM_STATE_INQUEUE,
2524 				    ("command not inqueue\n"));
2525 				cm->cm_state = MPR_CM_STATE_BUSY;
2526 				cm->cm_reply = reply;
2527 				cm->cm_reply_data =
2528 				    le32toh(desc->AddressReply.
2529 				    ReplyFrameAddress);
2530 			}
2531 			break;
2532 		}
2533 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2534 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2535 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2536 		default:
2537 			/* Unhandled */
2538 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
2539 			    desc->Default.ReplyFlags);
2540 			cm = NULL;
2541 			break;
2542 		}
2543 
2544 		if (cm != NULL) {
2545 			// Print Error reply frame
2546 			if (cm->cm_reply)
2547 				mpr_display_reply_info(sc,cm->cm_reply);
2548 			mpr_complete_command(sc, cm);
2549 		}
2550 
2551 		desc->Words.Low = 0xffffffff;
2552 		desc->Words.High = 0xffffffff;
2553 	}
2554 
2555 	if (pq != sc->replypostindex) {
2556 		mpr_dprint(sc, MPR_TRACE, "%s sc %p writing postindex %d\n",
2557 		    __func__, sc, sc->replypostindex);
2558 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2559 		    sc->replypostindex);
2560 	}
2561 
2562 	return;
2563 }
2564 
2565 static void
2566 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2567     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2568 {
2569 	struct mpr_event_handle *eh;
2570 	int event, handled = 0;
2571 
2572 	event = le16toh(reply->Event);
2573 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2574 		if (isset(eh->mask, event)) {
2575 			eh->callback(sc, data, reply);
2576 			handled++;
2577 		}
2578 	}
2579 
2580 	if (handled == 0)
2581 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2582 		    le16toh(event));
2583 
2584 	/*
2585 	 * This is the only place that the event/reply should be freed.
2586 	 * Anything wanting to hold onto the event data should have
2587 	 * already copied it into their own storage.
2588 	 */
2589 	mpr_free_reply(sc, data);
2590 }
2591 
2592 static void
2593 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2594 {
2595 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2596 
2597 	if (cm->cm_reply)
2598 		MPR_DPRINT_EVENT(sc, generic,
2599 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2600 
2601 	mpr_free_command(sc, cm);
2602 
2603 	/* next, send a port enable */
2604 	mprsas_startup(sc);
2605 }
2606 
2607 /*
2608  * For both register_events and update_events, the caller supplies a bitmap
2609  * of events that it _wants_.  These functions then turn that into a bitmask
2610  * suitable for the controller.
2611  */
2612 int
2613 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2614     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2615 {
2616 	struct mpr_event_handle *eh;
2617 	int error = 0;
2618 
2619 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2620 	if (!eh) {
2621 		mpr_dprint(sc, MPR_EVENT|MPR_ERROR,
2622 		    "Cannot allocate event memory\n");
2623 		return (ENOMEM);
2624 	}
2625 	eh->callback = cb;
2626 	eh->data = data;
2627 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2628 	if (mask != NULL)
2629 		error = mpr_update_events(sc, eh, mask);
2630 	*handle = eh;
2631 
2632 	return (error);
2633 }
2634 
2635 int
2636 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2637     uint8_t *mask)
2638 {
2639 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2640 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2641 	struct mpr_command *cm = NULL;
2642 	struct mpr_event_handle *eh;
2643 	int error, i;
2644 
2645 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2646 
2647 	if ((mask != NULL) && (handle != NULL))
2648 		bcopy(mask, &handle->mask[0], 16);
2649 	memset(sc->event_mask, 0xff, 16);
2650 
2651 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2652 		for (i = 0; i < 16; i++)
2653 			sc->event_mask[i] &= ~eh->mask[i];
2654 	}
2655 
2656 	if ((cm = mpr_alloc_command(sc)) == NULL)
2657 		return (EBUSY);
2658 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2659 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2660 	evtreq->MsgFlags = 0;
2661 	evtreq->SASBroadcastPrimitiveMasks = 0;
2662 #ifdef MPR_DEBUG_ALL_EVENTS
2663 	{
2664 		u_char fullmask[16];
2665 		memset(fullmask, 0x00, 16);
2666 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2667 	}
2668 #else
2669 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2670 #endif
2671 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2672 	cm->cm_data = NULL;
2673 
2674 	error = mpr_request_polled(sc, &cm);
2675 	if (cm != NULL)
2676 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2677 	if ((reply == NULL) ||
2678 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2679 		error = ENXIO;
2680 
2681 	if (reply)
2682 		MPR_DPRINT_EVENT(sc, generic, reply);
2683 
2684 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2685 
2686 	if (cm != NULL)
2687 		mpr_free_command(sc, cm);
2688 	return (error);
2689 }
2690 
2691 static int
2692 mpr_reregister_events(struct mpr_softc *sc)
2693 {
2694 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2695 	struct mpr_command *cm;
2696 	struct mpr_event_handle *eh;
2697 	int error, i;
2698 
2699 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2700 
2701 	/* first, reregister events */
2702 
2703 	memset(sc->event_mask, 0xff, 16);
2704 
2705 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2706 		for (i = 0; i < 16; i++)
2707 			sc->event_mask[i] &= ~eh->mask[i];
2708 	}
2709 
2710 	if ((cm = mpr_alloc_command(sc)) == NULL)
2711 		return (EBUSY);
2712 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2713 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2714 	evtreq->MsgFlags = 0;
2715 	evtreq->SASBroadcastPrimitiveMasks = 0;
2716 #ifdef MPR_DEBUG_ALL_EVENTS
2717 	{
2718 		u_char fullmask[16];
2719 		memset(fullmask, 0x00, 16);
2720 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2721 	}
2722 #else
2723 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2724 #endif
2725 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2726 	cm->cm_data = NULL;
2727 	cm->cm_complete = mpr_reregister_events_complete;
2728 
2729 	error = mpr_map_command(sc, cm);
2730 
2731 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2732 	    error);
2733 	return (error);
2734 }
2735 
2736 int
2737 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2738 {
2739 
2740 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2741 	free(handle, M_MPR);
2742 	return (mpr_update_events(sc, NULL, NULL));
2743 }
2744 
2745 /**
2746 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a
2747 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry
2748 * of the NVMe message (PRP1). If the data buffer is small enough to be described
2749 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to
2750 * describe a larger data buffer. If the data buffer is too large to describe
2751 * using the two PRP entriess inside the NVMe message, then PRP1 describes the
2752 * first data memory segment, and PRP2 contains a pointer to a PRP list located
2753 * elsewhere in memory to describe the remaining data memory segments. The PRP
2754 * list will be contiguous.
2755 
2756 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP
2757 * consists of a list of PRP entries to describe a number of noncontigous
2758 * physical memory segments as a single memory buffer, just as a SGL does. Note
2759 * however, that this function is only used by the IOCTL call, so the memory
2760 * given will be guaranteed to be contiguous. There is no need to translate
2761 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous
2762 * space that is one page size each.
2763 *
2764 * Each NVMe message contains two PRP entries. The first (PRP1) either contains
2765 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains
2766 * the second PRP element if the memory being described fits within 2 PRP
2767 * entries, or a PRP list pointer if the PRP spans more than two entries.
2768 *
2769 * A PRP list pointer contains the address of a PRP list, structured as a linear
2770 * array of PRP entries. Each PRP entry in this list describes a segment of
2771 * physical memory.
2772 *
2773 * Each 64-bit PRP entry comprises an address and an offset field. The address
2774 * always points to the beginning of a PAGE_SIZE physical memory page, and the
2775 * offset describes where within that page the memory segment begins. Only the
2776 * first element in a PRP list may contain a non-zero offest, implying that all
2777 * memory segments following the first begin at the start of a PAGE_SIZE page.
2778 *
2779 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory,
2780 * with exceptions for the first and last elements in the list. If the memory
2781 * being described by the list begins at a non-zero offset within the first page,
2782 * then the first PRP element will contain a non-zero offset indicating where the
2783 * region begins within the page. The last memory segment may end before the end
2784 * of the PAGE_SIZE segment, depending upon the overall size of the memory being
2785 * described by the PRP list.
2786 *
2787 * Since PRP entries lack any indication of size, the overall data buffer length
2788 * is used to determine where the end of the data memory buffer is located, and
2789 * how many PRP entries are required to describe it.
2790 *
2791 * Returns nothing.
2792 */
2793 void
2794 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm,
2795     Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data,
2796     uint32_t data_in_sz, uint32_t data_out_sz)
2797 {
2798 	int			prp_size = PRP_ENTRY_SIZE;
2799 	uint64_t		*prp_entry, *prp1_entry, *prp2_entry;
2800 	uint64_t		*prp_entry_phys, *prp_page, *prp_page_phys;
2801 	uint32_t		offset, entry_len, page_mask_result, page_mask;
2802 	bus_addr_t		paddr;
2803 	size_t			length;
2804 	struct mpr_prp_page	*prp_page_info = NULL;
2805 
2806 	/*
2807 	 * Not all commands require a data transfer. If no data, just return
2808 	 * without constructing any PRP.
2809 	 */
2810 	if (!data_in_sz && !data_out_sz)
2811 		return;
2812 
2813 	/*
2814 	 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is
2815 	 * located at a 24 byte offset from the start of the NVMe command. Then
2816 	 * set the current PRP entry pointer to PRP1.
2817 	 */
2818 	prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2819 	    NVME_CMD_PRP1_OFFSET);
2820 	prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2821 	    NVME_CMD_PRP2_OFFSET);
2822 	prp_entry = prp1_entry;
2823 
2824 	/*
2825 	 * For the PRP entries, use the specially allocated buffer of
2826 	 * contiguous memory. PRP Page allocation failures should not happen
2827 	 * because there should be enough PRP page buffers to account for the
2828 	 * possible NVMe QDepth.
2829 	 */
2830 	prp_page_info = mpr_alloc_prp_page(sc);
2831 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2832 	    "used for building a native NVMe SGL.\n", __func__));
2833 	prp_page = (uint64_t *)prp_page_info->prp_page;
2834 	prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2835 
2836 	/*
2837 	 * Insert the allocated PRP page into the command's PRP page list. This
2838 	 * will be freed when the command is freed.
2839 	 */
2840 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2841 
2842 	/*
2843 	 * Check if we are within 1 entry of a page boundary we don't want our
2844 	 * first entry to be a PRP List entry.
2845 	 */
2846 	page_mask = PAGE_SIZE - 1;
2847 	page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) &
2848 	    page_mask;
2849 	if (!page_mask_result)
2850 	{
2851 		/* Bump up to next page boundary. */
2852 		prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size);
2853 		prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys +
2854 		    prp_size);
2855 	}
2856 
2857 	/*
2858 	 * Set PRP physical pointer, which initially points to the current PRP
2859 	 * DMA memory page.
2860 	 */
2861 	prp_entry_phys = prp_page_phys;
2862 
2863 	/* Get physical address and length of the data buffer. */
2864 	paddr = (bus_addr_t)(uintptr_t)data;
2865 	if (data_in_sz)
2866 		length = data_in_sz;
2867 	else
2868 		length = data_out_sz;
2869 
2870 	/* Loop while the length is not zero. */
2871 	while (length)
2872 	{
2873 		/*
2874 		 * Check if we need to put a list pointer here if we are at page
2875 		 * boundary - prp_size (8 bytes).
2876 		 */
2877 		page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys +
2878 		    prp_size) & page_mask;
2879 		if (!page_mask_result)
2880 		{
2881 			/*
2882 			 * This is the last entry in a PRP List, so we need to
2883 			 * put a PRP list pointer here. What this does is:
2884 			 *   - bump the current memory pointer to the next
2885 			 *     address, which will be the next full page.
2886 			 *   - set the PRP Entry to point to that page. This is
2887 			 *     now the PRP List pointer.
2888 			 *   - bump the PRP Entry pointer the start of the next
2889 			 *     page. Since all of this PRP memory is contiguous,
2890 			 *     no need to get a new page - it's just the next
2891 			 *     address.
2892 			 */
2893 			prp_entry_phys++;
2894 			*prp_entry =
2895 			    htole64((uint64_t)(uintptr_t)prp_entry_phys);
2896 			prp_entry++;
2897 		}
2898 
2899 		/* Need to handle if entry will be part of a page. */
2900 		offset = (uint32_t)paddr & page_mask;
2901 		entry_len = PAGE_SIZE - offset;
2902 
2903 		if (prp_entry == prp1_entry)
2904 		{
2905 			/*
2906 			 * Must fill in the first PRP pointer (PRP1) before
2907 			 * moving on.
2908 			 */
2909 			*prp1_entry = htole64((uint64_t)paddr);
2910 
2911 			/*
2912 			 * Now point to the second PRP entry within the
2913 			 * command (PRP2).
2914 			 */
2915 			prp_entry = prp2_entry;
2916 		}
2917 		else if (prp_entry == prp2_entry)
2918 		{
2919 			/*
2920 			 * Should the PRP2 entry be a PRP List pointer or just a
2921 			 * regular PRP pointer? If there is more than one more
2922 			 * page of data, must use a PRP List pointer.
2923 			 */
2924 			if (length > PAGE_SIZE)
2925 			{
2926 				/*
2927 				 * PRP2 will contain a PRP List pointer because
2928 				 * more PRP's are needed with this command. The
2929 				 * list will start at the beginning of the
2930 				 * contiguous buffer.
2931 				 */
2932 				*prp2_entry =
2933 				    htole64(
2934 				    (uint64_t)(uintptr_t)prp_entry_phys);
2935 
2936 				/*
2937 				 * The next PRP Entry will be the start of the
2938 				 * first PRP List.
2939 				 */
2940 				prp_entry = prp_page;
2941 			}
2942 			else
2943 			{
2944 				/*
2945 				 * After this, the PRP Entries are complete.
2946 				 * This command uses 2 PRP's and no PRP list.
2947 				 */
2948 				*prp2_entry = htole64((uint64_t)paddr);
2949 			}
2950 		}
2951 		else
2952 		{
2953 			/*
2954 			 * Put entry in list and bump the addresses.
2955 			 *
2956 			 * After PRP1 and PRP2 are filled in, this will fill in
2957 			 * all remaining PRP entries in a PRP List, one per each
2958 			 * time through the loop.
2959 			 */
2960 			*prp_entry = htole64((uint64_t)paddr);
2961 			prp_entry++;
2962 			prp_entry_phys++;
2963 		}
2964 
2965 		/*
2966 		 * Bump the phys address of the command's data buffer by the
2967 		 * entry_len.
2968 		 */
2969 		paddr += entry_len;
2970 
2971 		/* Decrement length accounting for last partial page. */
2972 		if (entry_len > length)
2973 			length = 0;
2974 		else
2975 			length -= entry_len;
2976 	}
2977 }
2978 
2979 /*
2980  * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to
2981  * determine if the driver needs to build a native SGL. If so, that native SGL
2982  * is built in the contiguous buffers allocated especially for PCIe SGL
2983  * creation. If the driver will not build a native SGL, return TRUE and a
2984  * normal IEEE SGL will be built. Currently this routine supports NVMe devices
2985  * only.
2986  *
2987  * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built.
2988  */
2989 static int
2990 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm,
2991     bus_dma_segment_t *segs, int segs_left)
2992 {
2993 	uint32_t		i, sge_dwords, length, offset, entry_len;
2994 	uint32_t		num_entries, buff_len = 0, sges_in_segment;
2995 	uint32_t		page_mask, page_mask_result, *curr_buff;
2996 	uint32_t		*ptr_sgl, *ptr_first_sgl, first_page_offset;
2997 	uint32_t		first_page_data_size, end_residual;
2998 	uint64_t		*msg_phys;
2999 	bus_addr_t		paddr;
3000 	int			build_native_sgl = 0, first_prp_entry;
3001 	int			prp_size = PRP_ENTRY_SIZE;
3002 	Mpi25IeeeSgeChain64_t	*main_chain_element = NULL;
3003 	struct mpr_prp_page	*prp_page_info = NULL;
3004 
3005 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
3006 
3007 	/*
3008 	 * Add up the sizes of each segment length to get the total transfer
3009 	 * size, which will be checked against the Maximum Data Transfer Size.
3010 	 * If the data transfer length exceeds the MDTS for this device, just
3011 	 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O
3012 	 * up into multiple I/O's. [nvme_mdts = 0 means unlimited]
3013 	 */
3014 	for (i = 0; i < segs_left; i++)
3015 		buff_len += htole32(segs[i].ds_len);
3016 	if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS))
3017 		return 1;
3018 
3019 	/* Create page_mask (to get offset within page) */
3020 	page_mask = PAGE_SIZE - 1;
3021 
3022 	/*
3023 	 * Check if the number of elements exceeds the max number that can be
3024 	 * put in the main message frame (H/W can only translate an SGL that
3025 	 * is contained entirely in the main message frame).
3026 	 */
3027 	sges_in_segment = (sc->reqframesz -
3028 	    offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION);
3029 	if (segs_left > sges_in_segment)
3030 		build_native_sgl = 1;
3031 	else
3032 	{
3033 		/*
3034 		 * NVMe uses one PRP for each physical page (or part of physical
3035 		 * page).
3036 		 *    if 4 pages or less then IEEE is OK
3037 		 *    if > 5 pages then we need to build a native SGL
3038 		 *    if > 4 and <= 5 pages, then check the physical address of
3039 		 *      the first SG entry, then if this first size in the page
3040 		 *      is >= the residual beyond 4 pages then use IEEE,
3041 		 *      otherwise use native SGL
3042 		 */
3043 		if (buff_len > (PAGE_SIZE * 5))
3044 			build_native_sgl = 1;
3045 		else if ((buff_len > (PAGE_SIZE * 4)) &&
3046 		    (buff_len <= (PAGE_SIZE * 5)) )
3047 		{
3048 			msg_phys = (uint64_t *)(uintptr_t)segs[0].ds_addr;
3049 			first_page_offset =
3050 			    ((uint32_t)(uint64_t)(uintptr_t)msg_phys &
3051 			    page_mask);
3052 			first_page_data_size = PAGE_SIZE - first_page_offset;
3053 			end_residual = buff_len % PAGE_SIZE;
3054 
3055 			/*
3056 			 * If offset into first page pushes the end of the data
3057 			 * beyond end of the 5th page, we need the extra PRP
3058 			 * list.
3059 			 */
3060 			if (first_page_data_size < end_residual)
3061 				build_native_sgl = 1;
3062 
3063 			/*
3064 			 * Check if first SG entry size is < residual beyond 4
3065 			 * pages.
3066 			 */
3067 			if (htole32(segs[0].ds_len) <
3068 			    (buff_len - (PAGE_SIZE * 4)))
3069 				build_native_sgl = 1;
3070 		}
3071 	}
3072 
3073 	/* check if native SGL is needed */
3074 	if (!build_native_sgl)
3075 		return 1;
3076 
3077 	/*
3078 	 * Native SGL is needed.
3079 	 * Put a chain element in main message frame that points to the first
3080 	 * chain buffer.
3081 	 *
3082 	 * NOTE:  The ChainOffset field must be 0 when using a chain pointer to
3083 	 *        a native SGL.
3084 	 */
3085 
3086 	/* Set main message chain element pointer */
3087 	main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge;
3088 
3089 	/*
3090 	 * For NVMe the chain element needs to be the 2nd SGL entry in the main
3091 	 * message.
3092 	 */
3093 	main_chain_element = (Mpi25IeeeSgeChain64_t *)
3094 	    ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64));
3095 
3096 	/*
3097 	 * For the PRP entries, use the specially allocated buffer of
3098 	 * contiguous memory. PRP Page allocation failures should not happen
3099 	 * because there should be enough PRP page buffers to account for the
3100 	 * possible NVMe QDepth.
3101 	 */
3102 	prp_page_info = mpr_alloc_prp_page(sc);
3103 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
3104 	    "used for building a native NVMe SGL.\n", __func__));
3105 	curr_buff = (uint32_t *)prp_page_info->prp_page;
3106 	msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
3107 
3108 	/*
3109 	 * Insert the allocated PRP page into the command's PRP page list. This
3110 	 * will be freed when the command is freed.
3111 	 */
3112 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
3113 
3114 	/*
3115 	 * Check if we are within 1 entry of a page boundary we don't want our
3116 	 * first entry to be a PRP List entry.
3117 	 */
3118 	page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) &
3119 	    page_mask;
3120 	if (!page_mask_result) {
3121 		/* Bump up to next page boundary. */
3122 		curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size);
3123 		msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size);
3124 	}
3125 
3126 	/* Fill in the chain element and make it an NVMe segment type. */
3127 	main_chain_element->Address.High =
3128 	    htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32));
3129 	main_chain_element->Address.Low =
3130 	    htole32((uint32_t)(uintptr_t)msg_phys);
3131 	main_chain_element->NextChainOffset = 0;
3132 	main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3133 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3134 	    MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP;
3135 
3136 	/* Set SGL pointer to start of contiguous PCIe buffer. */
3137 	ptr_sgl = curr_buff;
3138 	sge_dwords = 2;
3139 	num_entries = 0;
3140 
3141 	/*
3142 	 * NVMe has a very convoluted PRP format. One PRP is required for each
3143 	 * page or partial page. We need to split up OS SG entries if they are
3144 	 * longer than one page or cross a page boundary. We also have to insert
3145 	 * a PRP list pointer entry as the last entry in each physical page of
3146 	 * the PRP list.
3147 	 *
3148 	 * NOTE: The first PRP "entry" is actually placed in the first SGL entry
3149 	 * in the main message in IEEE 64 format. The 2nd entry in the main
3150 	 * message is the chain element, and the rest of the PRP entries are
3151 	 * built in the contiguous PCIe buffer.
3152 	 */
3153 	first_prp_entry = 1;
3154 	ptr_first_sgl = (uint32_t *)cm->cm_sge;
3155 
3156 	for (i = 0; i < segs_left; i++) {
3157 		/* Get physical address and length of this SG entry. */
3158 		paddr = segs[i].ds_addr;
3159 		length = segs[i].ds_len;
3160 
3161 		/*
3162 		 * Check whether a given SGE buffer lies on a non-PAGED
3163 		 * boundary if this is not the first page. If so, this is not
3164 		 * expected so have FW build the SGL.
3165 		 */
3166 		if ((i != 0) && (((uint32_t)paddr & page_mask) != 0)) {
3167 			mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while "
3168 			    "building NVMe PRPs, low address is 0x%x\n",
3169 			    (uint32_t)paddr);
3170 			return 1;
3171 		}
3172 
3173 		/* Apart from last SGE, if any other SGE boundary is not page
3174 		 * aligned then it means that hole exists. Existence of hole
3175 		 * leads to data corruption. So fallback to IEEE SGEs.
3176 		 */
3177 		if (i != (segs_left - 1)) {
3178 			if (((uint32_t)paddr + length) & page_mask) {
3179 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE "
3180 				    "boundary while building NVMe PRPs, low "
3181 				    "address: 0x%x and length: %u\n",
3182 				    (uint32_t)paddr, length);
3183 				return 1;
3184 			}
3185 		}
3186 
3187 		/* Loop while the length is not zero. */
3188 		while (length) {
3189 			/*
3190 			 * Check if we need to put a list pointer here if we are
3191 			 * at page boundary - prp_size.
3192 			 */
3193 			page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl +
3194 			    prp_size) & page_mask;
3195 			if (!page_mask_result) {
3196 				/*
3197 				 * Need to put a PRP list pointer here.
3198 				 */
3199 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3200 				    prp_size);
3201 				*ptr_sgl = htole32((uintptr_t)msg_phys);
3202 				*(ptr_sgl+1) = htole32((uint64_t)(uintptr_t)
3203 				    msg_phys >> 32);
3204 				ptr_sgl += sge_dwords;
3205 				num_entries++;
3206 			}
3207 
3208 			/* Need to handle if entry will be part of a page. */
3209 			offset = (uint32_t)paddr & page_mask;
3210 			entry_len = PAGE_SIZE - offset;
3211 			if (first_prp_entry) {
3212 				/*
3213 				 * Put IEEE entry in first SGE in main message.
3214 				 * (Simple element, System addr, not end of
3215 				 * list.)
3216 				 */
3217 				*ptr_first_sgl = htole32((uint32_t)paddr);
3218 				*(ptr_first_sgl + 1) =
3219 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3220 				*(ptr_first_sgl + 2) = htole32(entry_len);
3221 				*(ptr_first_sgl + 3) = 0;
3222 
3223 				/* No longer the first PRP entry. */
3224 				first_prp_entry = 0;
3225 			} else {
3226 				/* Put entry in list. */
3227 				*ptr_sgl = htole32((uint32_t)paddr);
3228 				*(ptr_sgl + 1) =
3229 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3230 
3231 				/* Bump ptr_sgl, msg_phys, and num_entries. */
3232 				ptr_sgl += sge_dwords;
3233 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3234 				    prp_size);
3235 				num_entries++;
3236 			}
3237 
3238 			/* Bump the phys address by the entry_len. */
3239 			paddr += entry_len;
3240 
3241 			/* Decrement length accounting for last partial page. */
3242 			if (entry_len > length)
3243 				length = 0;
3244 			else
3245 				length -= entry_len;
3246 		}
3247 	}
3248 
3249 	/* Set chain element Length. */
3250 	main_chain_element->Length = htole32(num_entries * prp_size);
3251 
3252 	/* Return 0, indicating we built a native SGL. */
3253 	return 0;
3254 }
3255 
3256 /*
3257  * Add a chain element as the next SGE for the specified command.
3258  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
3259  * only required for IEEE commands.  Therefore there is no code for commands
3260  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
3261  * shouldn't be requesting chains).
3262  */
3263 static int
3264 mpr_add_chain(struct mpr_command *cm, int segsleft)
3265 {
3266 	struct mpr_softc *sc = cm->cm_sc;
3267 	MPI2_REQUEST_HEADER *req;
3268 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
3269 	struct mpr_chain *chain;
3270 	int sgc_size, current_segs, rem_segs, segs_per_frame;
3271 	uint8_t next_chain_offset = 0;
3272 
3273 	/*
3274 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
3275 	 * only IEEE commands should be requesting chains.  Return some error
3276 	 * code other than 0.
3277 	 */
3278 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
3279 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
3280 		    "an MPI SGL.\n");
3281 		return(ENOBUFS);
3282 	}
3283 
3284 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
3285 	if (cm->cm_sglsize < sgc_size)
3286 		panic("MPR: Need SGE Error Code\n");
3287 
3288 	chain = mpr_alloc_chain(cm->cm_sc);
3289 	if (chain == NULL)
3290 		return (ENOBUFS);
3291 
3292 	/*
3293 	 * Note: a double-linked list is used to make it easier to walk for
3294 	 * debugging.
3295 	 */
3296 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
3297 
3298 	/*
3299 	 * Need to know if the number of frames left is more than 1 or not.  If
3300 	 * more than 1 frame is required, NextChainOffset will need to be set,
3301 	 * which will just be the last segment of the frame.
3302 	 */
3303 	rem_segs = 0;
3304 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
3305 		/*
3306 		 * rem_segs is the number of segements remaining after the
3307 		 * segments that will go into the current frame.  Since it is
3308 		 * known that at least one more frame is required, account for
3309 		 * the chain element.  To know if more than one more frame is
3310 		 * required, just check if there will be a remainder after using
3311 		 * the current frame (with this chain) and the next frame.  If
3312 		 * so the NextChainOffset must be the last element of the next
3313 		 * frame.
3314 		 */
3315 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
3316 		rem_segs = segsleft - current_segs;
3317 		segs_per_frame = sc->chain_frame_size / sgc_size;
3318 		if (rem_segs > segs_per_frame) {
3319 			next_chain_offset = segs_per_frame - 1;
3320 		}
3321 	}
3322 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
3323 	ieee_sgc->Length = next_chain_offset ?
3324 	    htole32((uint32_t)sc->chain_frame_size) :
3325 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
3326 	ieee_sgc->NextChainOffset = next_chain_offset;
3327 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3328 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3329 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
3330 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
3331 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
3332 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
3333 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
3334 
3335 	cm->cm_sglsize = sc->chain_frame_size;
3336 	return (0);
3337 }
3338 
3339 /*
3340  * Add one scatter-gather element to the scatter-gather list for a command.
3341  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
3342  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
3343  * chain, so don't consider any chain additions.
3344  */
3345 int
3346 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
3347     int segsleft)
3348 {
3349 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3350 	u32 sge_flags;
3351 
3352 	/*
3353 	 * case 1: >=1 more segment, no room for anything (error)
3354 	 * case 2: 1 more segment and enough room for it
3355          */
3356 
3357 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
3358 		mpr_dprint(cm->cm_sc, MPR_ERROR,
3359 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
3360 		    __func__);
3361 		return(ENOBUFS);
3362 	}
3363 
3364 	KASSERT(segsleft == 1,
3365 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
3366 	    segsleft));
3367 
3368 	/*
3369 	 * There is one more segment left to add for the MPI SGL and there is
3370 	 * enough room in the frame to add it.  This is the normal case because
3371 	 * MPI SGL's don't have chains, otherwise something is wrong.
3372 	 *
3373 	 * If this is a bi-directional request, need to account for that
3374 	 * here.  Save the pre-filled sge values.  These will be used
3375 	 * either for the 2nd SGL or for a single direction SGL.  If
3376 	 * cm_out_len is non-zero, this is a bi-directional request, so
3377 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
3378 	 * fill in the IN SGL.  Note that at this time, when filling in
3379 	 * 2 SGL's for a bi-directional request, they both use the same
3380 	 * DMA buffer (same cm command).
3381 	 */
3382 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
3383 	saved_address_low = sge->Address.Low;
3384 	saved_address_high = sge->Address.High;
3385 	if (cm->cm_out_len) {
3386 		sge->FlagsLength = cm->cm_out_len |
3387 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3388 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
3389 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
3390 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3391 		    MPI2_SGE_FLAGS_SHIFT);
3392 		cm->cm_sglsize -= len;
3393 		/* Endian Safe code */
3394 		sge_flags = sge->FlagsLength;
3395 		sge->FlagsLength = htole32(sge_flags);
3396 		sge->Address.High = htole32(sge->Address.High);
3397 		sge->Address.Low = htole32(sge->Address.Low);
3398 		bcopy(sge, cm->cm_sge, len);
3399 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3400 	}
3401 	sge->FlagsLength = saved_buf_len |
3402 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3403 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
3404 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
3405 	    MPI2_SGE_FLAGS_END_OF_LIST |
3406 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3407 	    MPI2_SGE_FLAGS_SHIFT);
3408 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
3409 		sge->FlagsLength |=
3410 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
3411 		    MPI2_SGE_FLAGS_SHIFT);
3412 	} else {
3413 		sge->FlagsLength |=
3414 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
3415 		    MPI2_SGE_FLAGS_SHIFT);
3416 	}
3417 	sge->Address.Low = saved_address_low;
3418 	sge->Address.High = saved_address_high;
3419 
3420 	cm->cm_sglsize -= len;
3421 	/* Endian Safe code */
3422 	sge_flags = sge->FlagsLength;
3423 	sge->FlagsLength = htole32(sge_flags);
3424 	sge->Address.High = htole32(sge->Address.High);
3425 	sge->Address.Low = htole32(sge->Address.Low);
3426 	bcopy(sge, cm->cm_sge, len);
3427 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3428 	return (0);
3429 }
3430 
3431 /*
3432  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
3433  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
3434  * remaining size and pointer to the next SGE to fill in, respectively.
3435  */
3436 int
3437 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
3438 {
3439 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
3440 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
3441 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3442 	uint32_t sge_length;
3443 
3444 	/*
3445 	 * case 1: No room for chain or segment (error).
3446 	 * case 2: Two or more segments left but only room for chain.
3447 	 * case 3: Last segment and room for it, so set flags.
3448 	 */
3449 
3450 	/*
3451 	 * There should be room for at least one element, or there is a big
3452 	 * problem.
3453 	 */
3454 	if (cm->cm_sglsize < ieee_sge_size)
3455 		panic("MPR: Need SGE Error Code\n");
3456 
3457 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
3458 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
3459 			return (error);
3460 	}
3461 
3462 	if (segsleft == 1) {
3463 		/*
3464 		 * If this is a bi-directional request, need to account for that
3465 		 * here.  Save the pre-filled sge values.  These will be used
3466 		 * either for the 2nd SGL or for a single direction SGL.  If
3467 		 * cm_out_len is non-zero, this is a bi-directional request, so
3468 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
3469 		 * fill in the IN SGL.  Note that at this time, when filling in
3470 		 * 2 SGL's for a bi-directional request, they both use the same
3471 		 * DMA buffer (same cm command).
3472 		 */
3473 		saved_buf_len = sge->Length;
3474 		saved_address_low = sge->Address.Low;
3475 		saved_address_high = sge->Address.High;
3476 		if (cm->cm_out_len) {
3477 			sge->Length = cm->cm_out_len;
3478 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3479 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3480 			cm->cm_sglsize -= ieee_sge_size;
3481 			/* Endian Safe code */
3482 			sge_length = sge->Length;
3483 			sge->Length = htole32(sge_length);
3484 			sge->Address.High = htole32(sge->Address.High);
3485 			sge->Address.Low = htole32(sge->Address.Low);
3486 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
3487 			cm->cm_sge =
3488 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3489 			    ieee_sge_size);
3490 		}
3491 		sge->Length = saved_buf_len;
3492 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3493 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3494 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
3495 		sge->Address.Low = saved_address_low;
3496 		sge->Address.High = saved_address_high;
3497 	}
3498 
3499 	cm->cm_sglsize -= ieee_sge_size;
3500 	/* Endian Safe code */
3501 	sge_length = sge->Length;
3502 	sge->Length = htole32(sge_length);
3503 	sge->Address.High = htole32(sge->Address.High);
3504 	sge->Address.Low = htole32(sge->Address.Low);
3505 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
3506 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3507 	    ieee_sge_size);
3508 	return (0);
3509 }
3510 
3511 /*
3512  * Add one dma segment to the scatter-gather list for a command.
3513  */
3514 int
3515 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
3516     int segsleft)
3517 {
3518 	MPI2_SGE_SIMPLE64 sge;
3519 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
3520 
3521 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
3522 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3523 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3524 		ieee_sge.Length = len;
3525 		mpr_from_u64(pa, &ieee_sge.Address);
3526 
3527 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
3528 	} else {
3529 		/*
3530 		 * This driver always uses 64-bit address elements for
3531 		 * simplicity.
3532 		 */
3533 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3534 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
3535 		/* Set Endian safe macro in mpr_push_sge */
3536 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
3537 		mpr_from_u64(pa, &sge.Address);
3538 
3539 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
3540 	}
3541 }
3542 
3543 static void
3544 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
3545 {
3546 	struct mpr_softc *sc;
3547 	struct mpr_command *cm;
3548 	u_int i, dir, sflags;
3549 
3550 	cm = (struct mpr_command *)arg;
3551 	sc = cm->cm_sc;
3552 
3553 	/*
3554 	 * In this case, just print out a warning and let the chip tell the
3555 	 * user they did the wrong thing.
3556 	 */
3557 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
3558 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
3559 		    "segments, more than the %d allowed\n", __func__, nsegs,
3560 		    cm->cm_max_segs);
3561 	}
3562 
3563 	/*
3564 	 * Set up DMA direction flags.  Bi-directional requests are also handled
3565 	 * here.  In that case, both direction flags will be set.
3566 	 */
3567 	sflags = 0;
3568 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
3569 		/*
3570 		 * We have to add a special case for SMP passthrough, there
3571 		 * is no easy way to generically handle it.  The first
3572 		 * S/G element is used for the command (therefore the
3573 		 * direction bit needs to be set).  The second one is used
3574 		 * for the reply.  We'll leave it to the caller to make
3575 		 * sure we only have two buffers.
3576 		 */
3577 		/*
3578 		 * Even though the busdma man page says it doesn't make
3579 		 * sense to have both direction flags, it does in this case.
3580 		 * We have one s/g element being accessed in each direction.
3581 		 */
3582 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
3583 
3584 		/*
3585 		 * Set the direction flag on the first buffer in the SMP
3586 		 * passthrough request.  We'll clear it for the second one.
3587 		 */
3588 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
3589 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
3590 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
3591 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
3592 		dir = BUS_DMASYNC_PREWRITE;
3593 	} else
3594 		dir = BUS_DMASYNC_PREREAD;
3595 
3596 	/* Check if a native SG list is needed for an NVMe PCIe device. */
3597 	if (cm->cm_targ && cm->cm_targ->is_nvme &&
3598 	    mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) {
3599 		/* A native SG list was built, skip to end. */
3600 		goto out;
3601 	}
3602 
3603 	for (i = 0; i < nsegs; i++) {
3604 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
3605 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3606 		}
3607 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3608 		    sflags, nsegs - i);
3609 		if (error != 0) {
3610 			/* Resource shortage, roll back! */
3611 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
3612 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
3613 				    "consider increasing hw.mpr.max_chains.\n");
3614 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
3615 			mpr_complete_command(sc, cm);
3616 			return;
3617 		}
3618 	}
3619 
3620 out:
3621 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3622 	mpr_enqueue_request(sc, cm);
3623 
3624 	return;
3625 }
3626 
3627 static void
3628 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3629 	     int error)
3630 {
3631 	mpr_data_cb(arg, segs, nsegs, error);
3632 }
3633 
3634 /*
3635  * This is the routine to enqueue commands ansynchronously.
3636  * Note that the only error path here is from bus_dmamap_load(), which can
3637  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3638  * assumed that if you have a command in-hand, then you have enough credits
3639  * to use it.
3640  */
3641 int
3642 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
3643 {
3644 	int error = 0;
3645 
3646 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
3647 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3648 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
3649 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
3650 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3651 		    cm->cm_data, mpr_data_cb, cm, 0);
3652 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3653 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3654 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
3655 	} else {
3656 		/* Add a zero-length element as needed */
3657 		if (cm->cm_sge != NULL)
3658 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
3659 		mpr_enqueue_request(sc, cm);
3660 	}
3661 
3662 	return (error);
3663 }
3664 
3665 /*
3666  * This is the routine to enqueue commands synchronously.  An error of
3667  * EINPROGRESS from mpr_map_command() is ignored since the command will
3668  * be executed and enqueued automatically.  Other errors come from msleep().
3669  */
3670 int
3671 mpr_wait_command(struct mpr_softc *sc, struct mpr_command **cmp, int timeout,
3672     int sleep_flag)
3673 {
3674 	int error, rc;
3675 	struct timeval cur_time, start_time;
3676 	struct mpr_command *cm = *cmp;
3677 
3678 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
3679 		return  EBUSY;
3680 
3681 	cm->cm_complete = NULL;
3682 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
3683 	error = mpr_map_command(sc, cm);
3684 	if ((error != 0) && (error != EINPROGRESS))
3685 		return (error);
3686 
3687 	// Check for context and wait for 50 mSec at a time until time has
3688 	// expired or the command has finished.  If msleep can't be used, need
3689 	// to poll.
3690 #if __FreeBSD_version >= 1000029
3691 	if (curthread->td_no_sleeping)
3692 #else //__FreeBSD_version < 1000029
3693 	if (curthread->td_pflags & TDP_NOSLEEPING)
3694 #endif //__FreeBSD_version >= 1000029
3695 		sleep_flag = NO_SLEEP;
3696 	getmicrouptime(&start_time);
3697 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
3698 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
3699 		if (error == EWOULDBLOCK) {
3700 			/*
3701 			 * Record the actual elapsed time in the case of a
3702 			 * timeout for the message below.
3703 			 */
3704 			getmicrouptime(&cur_time);
3705 			timevalsub(&cur_time, &start_time);
3706 		}
3707 	} else {
3708 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3709 			mpr_intr_locked(sc);
3710 			if (sleep_flag == CAN_SLEEP)
3711 				pause("mprwait", hz/20);
3712 			else
3713 				DELAY(50000);
3714 
3715 			getmicrouptime(&cur_time);
3716 			timevalsub(&cur_time, &start_time);
3717 			if (cur_time.tv_sec > timeout) {
3718 				error = EWOULDBLOCK;
3719 				break;
3720 			}
3721 		}
3722 	}
3723 
3724 	if (error == EWOULDBLOCK) {
3725 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s, timeout=%d,"
3726 		    " elapsed=%jd\n", __func__, timeout,
3727 		    (intmax_t)cur_time.tv_sec);
3728 		rc = mpr_reinit(sc);
3729 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3730 		    "failed");
3731 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3732 			/*
3733 			 * Tell the caller that we freed the command in a
3734 			 * reinit.
3735 			 */
3736 			*cmp = NULL;
3737 		}
3738 		error = ETIMEDOUT;
3739 	}
3740 	return (error);
3741 }
3742 
3743 /*
3744  * This is the routine to enqueue a command synchonously and poll for
3745  * completion.  Its use should be rare.
3746  */
3747 int
3748 mpr_request_polled(struct mpr_softc *sc, struct mpr_command **cmp)
3749 {
3750 	int error, rc;
3751 	struct timeval cur_time, start_time;
3752 	struct mpr_command *cm = *cmp;
3753 
3754 	error = 0;
3755 
3756 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
3757 	cm->cm_complete = NULL;
3758 	mpr_map_command(sc, cm);
3759 
3760 	getmicrouptime(&start_time);
3761 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3762 		mpr_intr_locked(sc);
3763 
3764 		if (mtx_owned(&sc->mpr_mtx))
3765 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
3766 			    "mprpoll", hz/20);
3767 		else
3768 			pause("mprpoll", hz/20);
3769 
3770 		/*
3771 		 * Check for real-time timeout and fail if more than 60 seconds.
3772 		 */
3773 		getmicrouptime(&cur_time);
3774 		timevalsub(&cur_time, &start_time);
3775 		if (cur_time.tv_sec > 60) {
3776 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
3777 			error = ETIMEDOUT;
3778 			break;
3779 		}
3780 	}
3781 
3782 	if (error) {
3783 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3784 		rc = mpr_reinit(sc);
3785 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3786 		    "failed");
3787 
3788 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3789 			/*
3790 			 * Tell the caller that we freed the command in a
3791 			 * reinit.
3792 			 */
3793 			*cmp = NULL;
3794 		}
3795 	}
3796 	return (error);
3797 }
3798 
3799 /*
3800  * The MPT driver had a verbose interface for config pages.  In this driver,
3801  * reduce it to much simpler terms, similar to the Linux driver.
3802  */
3803 int
3804 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3805 {
3806 	MPI2_CONFIG_REQUEST *req;
3807 	struct mpr_command *cm;
3808 	int error;
3809 
3810 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
3811 		return (EBUSY);
3812 	}
3813 
3814 	cm = mpr_alloc_command(sc);
3815 	if (cm == NULL) {
3816 		return (EBUSY);
3817 	}
3818 
3819 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3820 	req->Function = MPI2_FUNCTION_CONFIG;
3821 	req->Action = params->action;
3822 	req->SGLFlags = 0;
3823 	req->ChainOffset = 0;
3824 	req->PageAddress = params->page_address;
3825 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3826 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3827 
3828 		hdr = &params->hdr.Ext;
3829 		req->ExtPageType = hdr->ExtPageType;
3830 		req->ExtPageLength = hdr->ExtPageLength;
3831 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3832 		req->Header.PageLength = 0; /* Must be set to zero */
3833 		req->Header.PageNumber = hdr->PageNumber;
3834 		req->Header.PageVersion = hdr->PageVersion;
3835 	} else {
3836 		MPI2_CONFIG_PAGE_HEADER *hdr;
3837 
3838 		hdr = &params->hdr.Struct;
3839 		req->Header.PageType = hdr->PageType;
3840 		req->Header.PageNumber = hdr->PageNumber;
3841 		req->Header.PageLength = hdr->PageLength;
3842 		req->Header.PageVersion = hdr->PageVersion;
3843 	}
3844 
3845 	cm->cm_data = params->buffer;
3846 	cm->cm_length = params->length;
3847 	if (cm->cm_data != NULL) {
3848 		cm->cm_sge = &req->PageBufferSGE;
3849 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3850 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
3851 	} else
3852 		cm->cm_sge = NULL;
3853 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3854 
3855 	cm->cm_complete_data = params;
3856 	if (params->callback != NULL) {
3857 		cm->cm_complete = mpr_config_complete;
3858 		return (mpr_map_command(sc, cm));
3859 	} else {
3860 		error = mpr_wait_command(sc, &cm, 0, CAN_SLEEP);
3861 		if (error) {
3862 			mpr_dprint(sc, MPR_FAULT,
3863 			    "Error %d reading config page\n", error);
3864 			if (cm != NULL)
3865 				mpr_free_command(sc, cm);
3866 			return (error);
3867 		}
3868 		mpr_config_complete(sc, cm);
3869 	}
3870 
3871 	return (0);
3872 }
3873 
3874 int
3875 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3876 {
3877 	return (EINVAL);
3878 }
3879 
3880 static void
3881 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
3882 {
3883 	MPI2_CONFIG_REPLY *reply;
3884 	struct mpr_config_params *params;
3885 
3886 	MPR_FUNCTRACE(sc);
3887 	params = cm->cm_complete_data;
3888 
3889 	if (cm->cm_data != NULL) {
3890 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3891 		    BUS_DMASYNC_POSTREAD);
3892 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3893 	}
3894 
3895 	/*
3896 	 * XXX KDM need to do more error recovery?  This results in the
3897 	 * device in question not getting probed.
3898 	 */
3899 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
3900 		params->status = MPI2_IOCSTATUS_BUSY;
3901 		goto done;
3902 	}
3903 
3904 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3905 	if (reply == NULL) {
3906 		params->status = MPI2_IOCSTATUS_BUSY;
3907 		goto done;
3908 	}
3909 	params->status = reply->IOCStatus;
3910 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3911 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3912 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3913 		params->hdr.Ext.PageType = reply->Header.PageType;
3914 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3915 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3916 	} else {
3917 		params->hdr.Struct.PageType = reply->Header.PageType;
3918 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3919 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3920 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3921 	}
3922 
3923 done:
3924 	mpr_free_command(sc, cm);
3925 	if (params->callback != NULL)
3926 		params->callback(sc, params);
3927 
3928 	return;
3929 }
3930