xref: /freebsd/sys/dev/mpr/mpr.c (revision 89e3d5671ba13dceca272d5b159c9bd805f3f504)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2011-2015 LSI Corp.
4  * Copyright (c) 2013-2016 Avago Technologies
5  * Copyright 2000-2020 Broadcom Inc.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  * Broadcom Inc. (LSI) MPT-Fusion Host Adapter FreeBSD
30  *
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 /* Communications core for Avago Technologies (LSI) MPT3 */
37 
38 /* TODO Move headers to mprvar */
39 #include <sys/types.h>
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/selinfo.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/module.h>
47 #include <sys/bus.h>
48 #include <sys/conf.h>
49 #include <sys/bio.h>
50 #include <sys/malloc.h>
51 #include <sys/uio.h>
52 #include <sys/sysctl.h>
53 #include <sys/smp.h>
54 #include <sys/queue.h>
55 #include <sys/kthread.h>
56 #include <sys/taskqueue.h>
57 #include <sys/endian.h>
58 #include <sys/eventhandler.h>
59 #include <sys/sbuf.h>
60 #include <sys/priv.h>
61 
62 #include <machine/bus.h>
63 #include <machine/resource.h>
64 #include <sys/rman.h>
65 #include <sys/proc.h>
66 
67 #include <dev/pci/pcivar.h>
68 
69 #include <cam/cam.h>
70 #include <cam/cam_ccb.h>
71 #include <cam/scsi/scsi_all.h>
72 
73 #include <dev/mpr/mpi/mpi2_type.h>
74 #include <dev/mpr/mpi/mpi2.h>
75 #include <dev/mpr/mpi/mpi2_ioc.h>
76 #include <dev/mpr/mpi/mpi2_sas.h>
77 #include <dev/mpr/mpi/mpi2_pci.h>
78 #include <dev/mpr/mpi/mpi2_cnfg.h>
79 #include <dev/mpr/mpi/mpi2_init.h>
80 #include <dev/mpr/mpi/mpi2_tool.h>
81 #include <dev/mpr/mpr_ioctl.h>
82 #include <dev/mpr/mprvar.h>
83 #include <dev/mpr/mpr_table.h>
84 #include <dev/mpr/mpr_sas.h>
85 
86 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag);
87 static int mpr_init_queues(struct mpr_softc *sc);
88 static void mpr_resize_queues(struct mpr_softc *sc);
89 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag);
90 static int mpr_transition_operational(struct mpr_softc *sc);
91 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching);
92 static void mpr_iocfacts_free(struct mpr_softc *sc);
93 static void mpr_startup(void *arg);
94 static int mpr_send_iocinit(struct mpr_softc *sc);
95 static int mpr_alloc_queues(struct mpr_softc *sc);
96 static int mpr_alloc_hw_queues(struct mpr_softc *sc);
97 static int mpr_alloc_replies(struct mpr_softc *sc);
98 static int mpr_alloc_requests(struct mpr_softc *sc);
99 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc);
100 static int mpr_attach_log(struct mpr_softc *sc);
101 static __inline void mpr_complete_command(struct mpr_softc *sc,
102     struct mpr_command *cm);
103 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
104     MPI2_EVENT_NOTIFICATION_REPLY *reply);
105 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm);
106 static void mpr_periodic(void *);
107 static int mpr_reregister_events(struct mpr_softc *sc);
108 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm);
109 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts);
110 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag);
111 static int mpr_debug_sysctl(SYSCTL_HANDLER_ARGS);
112 static int mpr_dump_reqs(SYSCTL_HANDLER_ARGS);
113 static void mpr_parse_debug(struct mpr_softc *sc, char *list);
114 
115 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
116     "MPR Driver Parameters");
117 
118 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory");
119 
120 /*
121  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
122  * any state and back to its initialization state machine.
123  */
124 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
125 
126 /*
127  * Added this union to smoothly convert le64toh cm->cm_desc.Words.
128  * Compiler only supports uint64_t to be passed as an argument.
129  * Otherwise it will throw this error:
130  * "aggregate value used where an integer was expected"
131  */
132 typedef union {
133         u64 word;
134         struct {
135                 u32 low;
136                 u32 high;
137         } u;
138 } request_descriptor_t;
139 
140 /* Rate limit chain-fail messages to 1 per minute */
141 static struct timeval mpr_chainfail_interval = { 60, 0 };
142 
143 /*
144  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
145  * If this function is called from process context, it can sleep
146  * and there is no harm to sleep, in case if this fuction is called
147  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
148  * based on sleep flags driver will call either msleep, pause or DELAY.
149  * msleep and pause are of same variant, but pause is used when mpr_mtx
150  * is not hold by driver.
151  */
152 static int
153 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag)
154 {
155 	uint32_t reg;
156 	int i, error, tries = 0;
157 	uint8_t first_wait_done = FALSE;
158 
159 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
160 
161 	/* Clear any pending interrupts */
162 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
163 
164 	/*
165 	 * Force NO_SLEEP for threads prohibited to sleep
166  	 * e.a Thread from interrupt handler are prohibited to sleep.
167  	 */
168 	if (curthread->td_no_sleeping)
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, chain_seg_size;
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 		chain_seg_size = htole16(sc->facts->IOCMaxChainSegmentSize);
417 		if (chain_seg_size == 0)
418 			chain_seg_size = MPR_DEFAULT_CHAIN_SEG_SIZE;
419 		sc->chain_frame_size = chain_seg_size *
420 		    MPR_MAX_CHAIN_ELEMENT_SIZE;
421 	} else {
422 		sc->chain_frame_size = sc->reqframesz;
423 	}
424 
425 	/*
426 	 * Max IO Size is Page Size * the following:
427 	 * ((SGEs per frame - 1 for chain element) * Max Chain Depth)
428 	 * + 1 for no chain needed in last frame
429 	 *
430 	 * If user suggests a Max IO size to use, use the smaller of the
431 	 * user's value and the calculated value as long as the user's
432 	 * value is larger than 0. The user's value is in pages.
433 	 */
434 	sges_per_frame = sc->chain_frame_size/sizeof(MPI2_IEEE_SGE_SIMPLE64)-1;
435 	maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE;
436 
437 	/*
438 	 * If I/O size limitation requested then use it and pass up to CAM.
439 	 * If not, use maxphys as an optimization hint, but report HW limit.
440 	 */
441 	if (sc->max_io_pages > 0) {
442 		maxio = min(maxio, sc->max_io_pages * PAGE_SIZE);
443 		sc->maxio = maxio;
444 	} else {
445 		sc->maxio = maxio;
446 		maxio = min(maxio, maxphys);
447 	}
448 
449 	sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) /
450 	    sges_per_frame * reqcr;
451 	if (sc->max_chains > 0 && sc->max_chains < sc->num_chains)
452 		sc->num_chains = sc->max_chains;
453 
454 	/*
455 	 * Figure out the number of MSIx-based queues.  If the firmware or
456 	 * user has done something crazy and not allowed enough credit for
457 	 * the queues to be useful then don't enable multi-queue.
458 	 */
459 	if (sc->facts->MaxMSIxVectors < 2)
460 		sc->msi_msgs = 1;
461 
462 	if (sc->msi_msgs > 1) {
463 		sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus);
464 		sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors);
465 		if (sc->num_reqs / sc->msi_msgs < 2)
466 			sc->msi_msgs = 1;
467 	}
468 
469 	mpr_dprint(sc, MPR_INIT, "Sized queues to q=%d reqs=%d replies=%d\n",
470 	    sc->msi_msgs, sc->num_reqs, sc->num_replies);
471 }
472 
473 /*
474  * This is called during attach and when re-initializing due to a Diag Reset.
475  * IOC Facts is used to allocate many of the structures needed by the driver.
476  * If called from attach, de-allocation is not required because the driver has
477  * not allocated any structures yet, but if called from a Diag Reset, previously
478  * allocated structures based on IOC Facts will need to be freed and re-
479  * allocated bases on the latest IOC Facts.
480  */
481 static int
482 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching)
483 {
484 	int error;
485 	Mpi2IOCFactsReply_t saved_facts;
486 	uint8_t saved_mode, reallocating;
487 
488 	mpr_dprint(sc, MPR_INIT|MPR_TRACE, "%s entered\n", __func__);
489 
490 	/* Save old IOC Facts and then only reallocate if Facts have changed */
491 	if (!attaching) {
492 		bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY));
493 	}
494 
495 	/*
496 	 * Get IOC Facts.  In all cases throughout this function, panic if doing
497 	 * a re-initialization and only return the error if attaching so the OS
498 	 * can handle it.
499 	 */
500 	if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) {
501 		if (attaching) {
502 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to get "
503 			    "IOC Facts with error %d, exit\n", error);
504 			return (error);
505 		} else {
506 			panic("%s failed to get IOC Facts with error %d\n",
507 			    __func__, error);
508 		}
509 	}
510 
511 	MPR_DPRINT_PAGE(sc, MPR_XINFO, iocfacts, sc->facts);
512 
513 	snprintf(sc->fw_version, sizeof(sc->fw_version),
514 	    "%02d.%02d.%02d.%02d",
515 	    sc->facts->FWVersion.Struct.Major,
516 	    sc->facts->FWVersion.Struct.Minor,
517 	    sc->facts->FWVersion.Struct.Unit,
518 	    sc->facts->FWVersion.Struct.Dev);
519 
520 	snprintf(sc->msg_version, sizeof(sc->msg_version), "%d.%d",
521 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MAJOR_MASK) >>
522 	    MPI2_IOCFACTS_MSGVERSION_MAJOR_SHIFT,
523 	    (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MINOR_MASK) >>
524 	    MPI2_IOCFACTS_MSGVERSION_MINOR_SHIFT);
525 
526 	mpr_dprint(sc, MPR_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version,
527 	    MPR_DRIVER_VERSION);
528 	mpr_dprint(sc, MPR_INFO,
529 	    "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
530 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
531 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
532 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc"
533 	    "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV");
534 
535 	/*
536 	 * If the chip doesn't support event replay then a hard reset will be
537 	 * required to trigger a full discovery.  Do the reset here then
538 	 * retransition to Ready.  A hard reset might have already been done,
539 	 * but it doesn't hurt to do it again.  Only do this if attaching, not
540 	 * for a Diag Reset.
541 	 */
542 	if (attaching && ((sc->facts->IOCCapabilities &
543 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) {
544 		mpr_dprint(sc, MPR_INIT, "No event replay, resetting\n");
545 		mpr_diag_reset(sc, NO_SLEEP);
546 		if ((error = mpr_transition_ready(sc)) != 0) {
547 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
548 			    "transition to ready with error %d, exit\n",
549 			    error);
550 			return (error);
551 		}
552 	}
553 
554 	/*
555 	 * Set flag if IR Firmware is loaded.  If the RAID Capability has
556 	 * changed from the previous IOC Facts, log a warning, but only if
557 	 * checking this after a Diag Reset and not during attach.
558 	 */
559 	saved_mode = sc->ir_firmware;
560 	if (sc->facts->IOCCapabilities &
561 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
562 		sc->ir_firmware = 1;
563 	if (!attaching) {
564 		if (sc->ir_firmware != saved_mode) {
565 			mpr_dprint(sc, MPR_INIT|MPR_FAULT, "new IR/IT mode "
566 			    "in IOC Facts does not match previous mode\n");
567 		}
568 	}
569 
570 	/* Only deallocate and reallocate if relevant IOC Facts have changed */
571 	reallocating = FALSE;
572 	sc->mpr_flags &= ~MPR_FLAGS_REALLOCATED;
573 
574 	if ((!attaching) &&
575 	    ((saved_facts.MsgVersion != sc->facts->MsgVersion) ||
576 	    (saved_facts.HeaderVersion != sc->facts->HeaderVersion) ||
577 	    (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) ||
578 	    (saved_facts.RequestCredit != sc->facts->RequestCredit) ||
579 	    (saved_facts.ProductID != sc->facts->ProductID) ||
580 	    (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) ||
581 	    (saved_facts.IOCRequestFrameSize !=
582 	    sc->facts->IOCRequestFrameSize) ||
583 	    (saved_facts.IOCMaxChainSegmentSize !=
584 	    sc->facts->IOCMaxChainSegmentSize) ||
585 	    (saved_facts.MaxTargets != sc->facts->MaxTargets) ||
586 	    (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) ||
587 	    (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) ||
588 	    (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) ||
589 	    (saved_facts.MaxReplyDescriptorPostQueueDepth !=
590 	    sc->facts->MaxReplyDescriptorPostQueueDepth) ||
591 	    (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) ||
592 	    (saved_facts.MaxVolumes != sc->facts->MaxVolumes) ||
593 	    (saved_facts.MaxPersistentEntries !=
594 	    sc->facts->MaxPersistentEntries))) {
595 		reallocating = TRUE;
596 
597 		/* Record that we reallocated everything */
598 		sc->mpr_flags |= MPR_FLAGS_REALLOCATED;
599 	}
600 
601 	/*
602 	 * Some things should be done if attaching or re-allocating after a Diag
603 	 * Reset, but are not needed after a Diag Reset if the FW has not
604 	 * changed.
605 	 */
606 	if (attaching || reallocating) {
607 		/*
608 		 * Check if controller supports FW diag buffers and set flag to
609 		 * enable each type.
610 		 */
611 		if (sc->facts->IOCCapabilities &
612 		    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
613 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].
614 			    enabled = TRUE;
615 		if (sc->facts->IOCCapabilities &
616 		    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
617 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].
618 			    enabled = TRUE;
619 		if (sc->facts->IOCCapabilities &
620 		    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
621 			sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].
622 			    enabled = TRUE;
623 
624 		/*
625 		 * Set flags for some supported items.
626 		 */
627 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
628 			sc->eedp_enabled = TRUE;
629 		if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
630 			sc->control_TLR = TRUE;
631 		if ((sc->facts->IOCCapabilities &
632 		    MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ) &&
633 		    (sc->mpr_flags & MPR_FLAGS_SEA_IOC))
634 			sc->atomic_desc_capable = TRUE;
635 
636 		mpr_resize_queues(sc);
637 
638 		/*
639 		 * Initialize all Tail Queues
640 		 */
641 		TAILQ_INIT(&sc->req_list);
642 		TAILQ_INIT(&sc->high_priority_req_list);
643 		TAILQ_INIT(&sc->chain_list);
644 		TAILQ_INIT(&sc->prp_page_list);
645 		TAILQ_INIT(&sc->tm_list);
646 	}
647 
648 	/*
649 	 * If doing a Diag Reset and the FW is significantly different
650 	 * (reallocating will be set above in IOC Facts comparison), then all
651 	 * buffers based on the IOC Facts will need to be freed before they are
652 	 * reallocated.
653 	 */
654 	if (reallocating) {
655 		mpr_iocfacts_free(sc);
656 		mprsas_realloc_targets(sc, saved_facts.MaxTargets +
657 		    saved_facts.MaxVolumes);
658 	}
659 
660 	/*
661 	 * Any deallocation has been completed.  Now start reallocating
662 	 * if needed.  Will only need to reallocate if attaching or if the new
663 	 * IOC Facts are different from the previous IOC Facts after a Diag
664 	 * Reset. Targets have already been allocated above if needed.
665 	 */
666 	error = 0;
667 	while (attaching || reallocating) {
668 		if ((error = mpr_alloc_hw_queues(sc)) != 0)
669 			break;
670 		if ((error = mpr_alloc_replies(sc)) != 0)
671 			break;
672 		if ((error = mpr_alloc_requests(sc)) != 0)
673 			break;
674 		if ((error = mpr_alloc_queues(sc)) != 0)
675 			break;
676 		break;
677 	}
678 	if (error) {
679 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
680 		    "Failed to alloc queues with error %d\n", error);
681 		mpr_free(sc);
682 		return (error);
683 	}
684 
685 	/* Always initialize the queues */
686 	bzero(sc->free_queue, sc->fqdepth * 4);
687 	mpr_init_queues(sc);
688 
689 	/*
690 	 * Always get the chip out of the reset state, but only panic if not
691 	 * attaching.  If attaching and there is an error, that is handled by
692 	 * the OS.
693 	 */
694 	error = mpr_transition_operational(sc);
695 	if (error != 0) {
696 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to "
697 		    "transition to operational with error %d\n", error);
698 		mpr_free(sc);
699 		return (error);
700 	}
701 
702 	/*
703 	 * Finish the queue initialization.
704 	 * These are set here instead of in mpr_init_queues() because the
705 	 * IOC resets these values during the state transition in
706 	 * mpr_transition_operational().  The free index is set to 1
707 	 * because the corresponding index in the IOC is set to 0, and the
708 	 * IOC treats the queues as full if both are set to the same value.
709 	 * Hence the reason that the queue can't hold all of the possible
710 	 * replies.
711 	 */
712 	sc->replypostindex = 0;
713 	mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
714 	mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
715 
716 	/*
717 	 * Attach the subsystems so they can prepare their event masks.
718 	 * XXX Should be dynamic so that IM/IR and user modules can attach
719 	 */
720 	error = 0;
721 	while (attaching) {
722 		mpr_dprint(sc, MPR_INIT, "Attaching subsystems\n");
723 		if ((error = mpr_attach_log(sc)) != 0)
724 			break;
725 		if ((error = mpr_attach_sas(sc)) != 0)
726 			break;
727 		if ((error = mpr_attach_user(sc)) != 0)
728 			break;
729 		break;
730 	}
731 	if (error) {
732 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
733 		    "Failed to attach all subsystems: error %d\n", error);
734 		mpr_free(sc);
735 		return (error);
736 	}
737 
738 	/*
739 	 * XXX If the number of MSI-X vectors changes during re-init, this
740 	 * won't see it and adjust.
741 	 */
742 	if (attaching && (error = mpr_pci_setup_interrupts(sc)) != 0) {
743 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
744 		    "Failed to setup interrupts\n");
745 		mpr_free(sc);
746 		return (error);
747 	}
748 
749 	return (error);
750 }
751 
752 /*
753  * This is called if memory is being free (during detach for example) and when
754  * buffers need to be reallocated due to a Diag Reset.
755  */
756 static void
757 mpr_iocfacts_free(struct mpr_softc *sc)
758 {
759 	struct mpr_command *cm;
760 	int i;
761 
762 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
763 
764 	if (sc->free_busaddr != 0)
765 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
766 	if (sc->free_queue != NULL)
767 		bus_dmamem_free(sc->queues_dmat, sc->free_queue,
768 		    sc->queues_map);
769 	if (sc->queues_dmat != NULL)
770 		bus_dma_tag_destroy(sc->queues_dmat);
771 
772 	if (sc->chain_frames != NULL) {
773 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
774 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
775 		    sc->chain_map);
776 	}
777 	if (sc->chain_dmat != NULL)
778 		bus_dma_tag_destroy(sc->chain_dmat);
779 
780 	if (sc->sense_busaddr != 0)
781 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
782 	if (sc->sense_frames != NULL)
783 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,
784 		    sc->sense_map);
785 	if (sc->sense_dmat != NULL)
786 		bus_dma_tag_destroy(sc->sense_dmat);
787 
788 	if (sc->prp_page_busaddr != 0)
789 		bus_dmamap_unload(sc->prp_page_dmat, sc->prp_page_map);
790 	if (sc->prp_pages != NULL)
791 		bus_dmamem_free(sc->prp_page_dmat, sc->prp_pages,
792 		    sc->prp_page_map);
793 	if (sc->prp_page_dmat != NULL)
794 		bus_dma_tag_destroy(sc->prp_page_dmat);
795 
796 	if (sc->reply_busaddr != 0)
797 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
798 	if (sc->reply_frames != NULL)
799 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,
800 		    sc->reply_map);
801 	if (sc->reply_dmat != NULL)
802 		bus_dma_tag_destroy(sc->reply_dmat);
803 
804 	if (sc->req_busaddr != 0)
805 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
806 	if (sc->req_frames != NULL)
807 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
808 	if (sc->req_dmat != NULL)
809 		bus_dma_tag_destroy(sc->req_dmat);
810 
811 	if (sc->chains != NULL)
812 		free(sc->chains, M_MPR);
813 	if (sc->prps != NULL)
814 		free(sc->prps, M_MPR);
815 	if (sc->commands != NULL) {
816 		for (i = 1; i < sc->num_reqs; i++) {
817 			cm = &sc->commands[i];
818 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
819 		}
820 		free(sc->commands, M_MPR);
821 	}
822 	if (sc->buffer_dmat != NULL)
823 		bus_dma_tag_destroy(sc->buffer_dmat);
824 
825 	mpr_pci_free_interrupts(sc);
826 	free(sc->queues, M_MPR);
827 	sc->queues = NULL;
828 }
829 
830 /*
831  * The terms diag reset and hard reset are used interchangeably in the MPI
832  * docs to mean resetting the controller chip.  In this code diag reset
833  * cleans everything up, and the hard reset function just sends the reset
834  * sequence to the chip.  This should probably be refactored so that every
835  * subsystem gets a reset notification of some sort, and can clean up
836  * appropriately.
837  */
838 int
839 mpr_reinit(struct mpr_softc *sc)
840 {
841 	int error;
842 	struct mprsas_softc *sassc;
843 
844 	sassc = sc->sassc;
845 
846 	MPR_FUNCTRACE(sc);
847 
848 	mtx_assert(&sc->mpr_mtx, MA_OWNED);
849 
850 	mpr_dprint(sc, MPR_INIT|MPR_INFO, "Reinitializing controller\n");
851 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) {
852 		mpr_dprint(sc, MPR_INIT, "Reset already in progress\n");
853 		return 0;
854 	}
855 
856 	/*
857 	 * Make sure the completion callbacks can recognize they're getting
858 	 * a NULL cm_reply due to a reset.
859 	 */
860 	sc->mpr_flags |= MPR_FLAGS_DIAGRESET;
861 
862 	/*
863 	 * Mask interrupts here.
864 	 */
865 	mpr_dprint(sc, MPR_INIT, "Masking interrupts and resetting\n");
866 	mpr_mask_intr(sc);
867 
868 	error = mpr_diag_reset(sc, CAN_SLEEP);
869 	if (error != 0) {
870 		panic("%s hard reset failed with error %d\n", __func__, error);
871 	}
872 
873 	/* Restore the PCI state, including the MSI-X registers */
874 	mpr_pci_restore(sc);
875 
876 	/* Give the I/O subsystem special priority to get itself prepared */
877 	mprsas_handle_reinit(sc);
878 
879 	/*
880 	 * Get IOC Facts and allocate all structures based on this information.
881 	 * The attach function will also call mpr_iocfacts_allocate at startup.
882 	 * If relevant values have changed in IOC Facts, this function will free
883 	 * all of the memory based on IOC Facts and reallocate that memory.
884 	 */
885 	if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) {
886 		panic("%s IOC Facts based allocation failed with error %d\n",
887 		    __func__, error);
888 	}
889 
890 	/*
891 	 * Mapping structures will be re-allocated after getting IOC Page8, so
892 	 * free these structures here.
893 	 */
894 	mpr_mapping_exit(sc);
895 
896 	/*
897 	 * The static page function currently read is IOC Page8.  Others can be
898 	 * added in future.  It's possible that the values in IOC Page8 have
899 	 * changed after a Diag Reset due to user modification, so always read
900 	 * these.  Interrupts are masked, so unmask them before getting config
901 	 * pages.
902 	 */
903 	mpr_unmask_intr(sc);
904 	sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET;
905 	mpr_base_static_config_pages(sc);
906 
907 	/*
908 	 * Some mapping info is based in IOC Page8 data, so re-initialize the
909 	 * mapping tables.
910 	 */
911 	mpr_mapping_initialize(sc);
912 
913 	/*
914 	 * Restart will reload the event masks clobbered by the reset, and
915 	 * then enable the port.
916 	 */
917 	mpr_reregister_events(sc);
918 
919 	/* the end of discovery will release the simq, so we're done. */
920 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Finished sc %p post %u free %u\n",
921 	    sc, sc->replypostindex, sc->replyfreeindex);
922 	mprsas_release_simq_reinit(sassc);
923 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
924 
925 	return 0;
926 }
927 
928 /* Wait for the chip to ACK a word that we've put into its FIFO
929  * Wait for <timeout> seconds. In single loop wait for busy loop
930  * for 500 microseconds.
931  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
932  * */
933 static int
934 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag)
935 {
936 	u32 cntdn, count;
937 	u32 int_status;
938 	u32 doorbell;
939 
940 	count = 0;
941 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
942 	do {
943 		int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
944 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
945 			mpr_dprint(sc, MPR_TRACE, "%s: successful count(%d), "
946 			    "timeout(%d)\n", __func__, count, timeout);
947 			return 0;
948 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
949 			doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
950 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
951 			    MPI2_IOC_STATE_FAULT) {
952 				mpr_dprint(sc, MPR_FAULT,
953 				    "fault_state(0x%04x)!\n", doorbell);
954 				return (EFAULT);
955 			}
956 		} else if (int_status == 0xFFFFFFFF)
957 			goto out;
958 
959 		/*
960 		 * If it can sleep, sleep for 1 milisecond, else busy loop for
961  		 * 0.5 milisecond
962 		 */
963 		if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP)
964 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba",
965 			    hz/1000);
966 		else if (sleep_flag == CAN_SLEEP)
967 			pause("mprdba", hz/1000);
968 		else
969 			DELAY(500);
970 		count++;
971 	} while (--cntdn);
972 
973 out:
974 	mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), "
975 		"int_status(%x)!\n", __func__, count, int_status);
976 	return (ETIMEDOUT);
977 }
978 
979 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
980 static int
981 mpr_wait_db_int(struct mpr_softc *sc)
982 {
983 	int retry;
984 
985 	for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) {
986 		if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
987 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
988 			return (0);
989 		DELAY(2000);
990 	}
991 	return (ETIMEDOUT);
992 }
993 
994 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
995 static int
996 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
997     int req_sz, int reply_sz, int timeout)
998 {
999 	uint32_t *data32;
1000 	uint16_t *data16;
1001 	int i, count, ioc_sz, residual;
1002 	int sleep_flags = CAN_SLEEP;
1003 
1004 	if (curthread->td_no_sleeping)
1005 		sleep_flags = NO_SLEEP;
1006 
1007 	/* Step 1 */
1008 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1009 
1010 	/* Step 2 */
1011 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1012 		return (EBUSY);
1013 
1014 	/* Step 3
1015 	 * Announce that a message is coming through the doorbell.  Messages
1016 	 * are pushed at 32bit words, so round up if needed.
1017 	 */
1018 	count = (req_sz + 3) / 4;
1019 	mpr_regwrite(sc, MPI2_DOORBELL_OFFSET,
1020 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
1021 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
1022 
1023 	/* Step 4 */
1024 	if (mpr_wait_db_int(sc) ||
1025 	    (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
1026 		mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n");
1027 		return (ENXIO);
1028 	}
1029 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1030 	if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
1031 		mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n");
1032 		return (ENXIO);
1033 	}
1034 
1035 	/* Step 5 */
1036 	/* Clock out the message data synchronously in 32-bit dwords*/
1037 	data32 = (uint32_t *)req;
1038 	for (i = 0; i < count; i++) {
1039 		mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
1040 		if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) {
1041 			mpr_dprint(sc, MPR_FAULT,
1042 			    "Timeout while writing doorbell\n");
1043 			return (ENXIO);
1044 		}
1045 	}
1046 
1047 	/* Step 6 */
1048 	/* Clock in the reply in 16-bit words.  The total length of the
1049 	 * message is always in the 4th byte, so clock out the first 2 words
1050 	 * manually, then loop the rest.
1051 	 */
1052 	data16 = (uint16_t *)reply;
1053 	if (mpr_wait_db_int(sc) != 0) {
1054 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n");
1055 		return (ENXIO);
1056 	}
1057 	data16[0] =
1058 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1059 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1060 	if (mpr_wait_db_int(sc) != 0) {
1061 		mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n");
1062 		return (ENXIO);
1063 	}
1064 	data16[1] =
1065 	    mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
1066 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1067 
1068 	/* Number of 32bit words in the message */
1069 	ioc_sz = reply->MsgLength;
1070 
1071 	/*
1072 	 * Figure out how many 16bit words to clock in without overrunning.
1073 	 * The precision loss with dividing reply_sz can safely be
1074 	 * ignored because the messages can only be multiples of 32bits.
1075 	 */
1076 	residual = 0;
1077 	count = MIN((reply_sz / 4), ioc_sz) * 2;
1078 	if (count < ioc_sz * 2) {
1079 		residual = ioc_sz * 2 - count;
1080 		mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d "
1081 		    "residual message words\n", residual);
1082 	}
1083 
1084 	for (i = 2; i < count; i++) {
1085 		if (mpr_wait_db_int(sc) != 0) {
1086 			mpr_dprint(sc, MPR_FAULT,
1087 			    "Timeout reading doorbell %d\n", i);
1088 			return (ENXIO);
1089 		}
1090 		data16[i] = mpr_regread(sc, MPI2_DOORBELL_OFFSET) &
1091 		    MPI2_DOORBELL_DATA_MASK;
1092 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1093 	}
1094 
1095 	/*
1096 	 * Pull out residual words that won't fit into the provided buffer.
1097 	 * This keeps the chip from hanging due to a driver programming
1098 	 * error.
1099 	 */
1100 	while (residual--) {
1101 		if (mpr_wait_db_int(sc) != 0) {
1102 			mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n");
1103 			return (ENXIO);
1104 		}
1105 		(void)mpr_regread(sc, MPI2_DOORBELL_OFFSET);
1106 		mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1107 	}
1108 
1109 	/* Step 7 */
1110 	if (mpr_wait_db_int(sc) != 0) {
1111 		mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n");
1112 		return (ENXIO);
1113 	}
1114 	if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
1115 		mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n");
1116 	mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
1117 
1118 	return (0);
1119 }
1120 
1121 static void
1122 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm)
1123 {
1124 	request_descriptor_t rd;
1125 
1126 	MPR_FUNCTRACE(sc);
1127 	mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n",
1128 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
1129 
1130 	if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags &
1131 	    MPR_FLAGS_SHUTDOWN))
1132 		mtx_assert(&sc->mpr_mtx, MA_OWNED);
1133 
1134 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
1135 		sc->io_cmds_highwater++;
1136 
1137 	KASSERT(cm->cm_state == MPR_CM_STATE_BUSY, ("command not busy\n"));
1138 	cm->cm_state = MPR_CM_STATE_INQUEUE;
1139 
1140 	if (sc->atomic_desc_capable) {
1141 		rd.u.low = cm->cm_desc.Words.Low;
1142 		mpr_regwrite(sc, MPI26_ATOMIC_REQUEST_DESCRIPTOR_POST_OFFSET,
1143 		    rd.u.low);
1144 	} else {
1145 		rd.u.low = cm->cm_desc.Words.Low;
1146 		rd.u.high = cm->cm_desc.Words.High;
1147 		rd.word = htole64(rd.word);
1148 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
1149 		    rd.u.low);
1150 		mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
1151 		    rd.u.high);
1152 	}
1153 }
1154 
1155 /*
1156  * Just the FACTS, ma'am.
1157  */
1158 static int
1159 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
1160 {
1161 	MPI2_DEFAULT_REPLY *reply;
1162 	MPI2_IOC_FACTS_REQUEST request;
1163 	int error, req_sz, reply_sz;
1164 
1165 	MPR_FUNCTRACE(sc);
1166 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1167 
1168 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
1169 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
1170 	reply = (MPI2_DEFAULT_REPLY *)facts;
1171 
1172 	bzero(&request, req_sz);
1173 	request.Function = MPI2_FUNCTION_IOC_FACTS;
1174 	error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
1175 
1176 	mpr_dprint(sc, MPR_INIT, "%s exit, error= %d\n", __func__, error);
1177 	return (error);
1178 }
1179 
1180 static int
1181 mpr_send_iocinit(struct mpr_softc *sc)
1182 {
1183 	MPI2_IOC_INIT_REQUEST	init;
1184 	MPI2_DEFAULT_REPLY	reply;
1185 	int req_sz, reply_sz, error;
1186 	struct timeval now;
1187 	uint64_t time_in_msec;
1188 
1189 	MPR_FUNCTRACE(sc);
1190 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
1191 
1192 	/* Do a quick sanity check on proper initialization */
1193 	if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0)
1194 	    || (sc->replyframesz == 0)) {
1195 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
1196 		    "Driver not fully initialized for IOCInit\n");
1197 		return (EINVAL);
1198 	}
1199 
1200 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
1201 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
1202 	bzero(&init, req_sz);
1203 	bzero(&reply, reply_sz);
1204 
1205 	/*
1206 	 * Fill in the init block.  Note that most addresses are
1207 	 * deliberately in the lower 32bits of memory.  This is a micro-
1208 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
1209 	 */
1210 	init.Function = MPI2_FUNCTION_IOC_INIT;
1211 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
1212 	init.MsgVersion = htole16(MPI2_VERSION);
1213 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
1214 	init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4));
1215 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
1216 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
1217 	init.SenseBufferAddressHigh = 0;
1218 	init.SystemReplyAddressHigh = 0;
1219 	init.SystemRequestFrameBaseAddress.High = 0;
1220 	init.SystemRequestFrameBaseAddress.Low =
1221 	    htole32((uint32_t)sc->req_busaddr);
1222 	init.ReplyDescriptorPostQueueAddress.High = 0;
1223 	init.ReplyDescriptorPostQueueAddress.Low =
1224 	    htole32((uint32_t)sc->post_busaddr);
1225 	init.ReplyFreeQueueAddress.High = 0;
1226 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
1227 	getmicrotime(&now);
1228 	time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000);
1229 	init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF);
1230 	init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF);
1231 	init.HostPageSize = HOST_PAGE_SIZE_4K;
1232 
1233 	error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
1234 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1235 		error = ENXIO;
1236 
1237 	mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", reply.IOCStatus);
1238 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
1239 	return (error);
1240 }
1241 
1242 void
1243 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1244 {
1245 	bus_addr_t *addr;
1246 
1247 	addr = arg;
1248 	*addr = segs[0].ds_addr;
1249 }
1250 
1251 void
1252 mpr_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1253 {
1254 	struct mpr_busdma_context *ctx;
1255 	int need_unload, need_free;
1256 
1257 	ctx = (struct mpr_busdma_context *)arg;
1258 	need_unload = 0;
1259 	need_free = 0;
1260 
1261 	mpr_lock(ctx->softc);
1262 	ctx->error = error;
1263 	ctx->completed = 1;
1264 	if ((error == 0) && (ctx->abandoned == 0)) {
1265 		*ctx->addr = segs[0].ds_addr;
1266 	} else {
1267 		if (nsegs != 0)
1268 			need_unload = 1;
1269 		if (ctx->abandoned != 0)
1270 			need_free = 1;
1271 	}
1272 	if (need_free == 0)
1273 		wakeup(ctx);
1274 
1275 	mpr_unlock(ctx->softc);
1276 
1277 	if (need_unload != 0) {
1278 		bus_dmamap_unload(ctx->buffer_dmat,
1279 				  ctx->buffer_dmamap);
1280 		*ctx->addr = 0;
1281 	}
1282 
1283 	if (need_free != 0)
1284 		free(ctx, M_MPR);
1285 }
1286 
1287 static int
1288 mpr_alloc_queues(struct mpr_softc *sc)
1289 {
1290 	struct mpr_queue *q;
1291 	int nq, i;
1292 
1293 	nq = sc->msi_msgs;
1294 	mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Allocating %d I/O queues\n", nq);
1295 
1296 	sc->queues = malloc(sizeof(struct mpr_queue) * nq, M_MPR,
1297 	     M_NOWAIT|M_ZERO);
1298 	if (sc->queues == NULL)
1299 		return (ENOMEM);
1300 
1301 	for (i = 0; i < nq; i++) {
1302 		q = &sc->queues[i];
1303 		mpr_dprint(sc, MPR_INIT, "Configuring queue %d %p\n", i, q);
1304 		q->sc = sc;
1305 		q->qnum = i;
1306 	}
1307 	return (0);
1308 }
1309 
1310 static int
1311 mpr_alloc_hw_queues(struct mpr_softc *sc)
1312 {
1313 	bus_dma_template_t t;
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 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1336 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(qsize),
1337 	    BD_MAXSEGSIZE(qsize), BD_NSEGMENTS(1),
1338 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1339 	if (bus_dma_template_tag(&t, &sc->queues_dmat)) {
1340 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues DMA tag\n");
1341 		return (ENOMEM);
1342         }
1343         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
1344 	    &sc->queues_map)) {
1345 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues memory\n");
1346 		return (ENOMEM);
1347         }
1348         bzero(queues, qsize);
1349         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
1350 	    mpr_memaddr_cb, &queues_busaddr, 0);
1351 
1352 	sc->free_queue = (uint32_t *)queues;
1353 	sc->free_busaddr = queues_busaddr;
1354 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
1355 	sc->post_busaddr = queues_busaddr + fqsize;
1356 	mpr_dprint(sc, MPR_INIT, "free queue busaddr= %#016jx size= %d\n",
1357 	    (uintmax_t)sc->free_busaddr, fqsize);
1358 	mpr_dprint(sc, MPR_INIT, "reply queue busaddr= %#016jx size= %d\n",
1359 	    (uintmax_t)sc->post_busaddr, pqsize);
1360 
1361 	return (0);
1362 }
1363 
1364 static int
1365 mpr_alloc_replies(struct mpr_softc *sc)
1366 {
1367 	bus_dma_template_t t;
1368 	int rsize, num_replies;
1369 
1370 	/* Store the reply frame size in bytes rather than as 32bit words */
1371 	sc->replyframesz = sc->facts->ReplyFrameSize * 4;
1372 
1373 	/*
1374 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
1375 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
1376 	 * replies can be used at once.
1377 	 */
1378 	num_replies = max(sc->fqdepth, sc->num_replies);
1379 
1380 	rsize = sc->replyframesz * num_replies;
1381 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1382 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize),
1383 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1384 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1385 	if (bus_dma_template_tag(&t, &sc->reply_dmat)) {
1386 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies DMA tag\n");
1387 		return (ENOMEM);
1388         }
1389         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
1390 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
1391 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies memory\n");
1392 		return (ENOMEM);
1393         }
1394         bzero(sc->reply_frames, rsize);
1395         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
1396 	    mpr_memaddr_cb, &sc->reply_busaddr, 0);
1397 	mpr_dprint(sc, MPR_INIT, "reply frames busaddr= %#016jx size= %d\n",
1398 	    (uintmax_t)sc->reply_busaddr, rsize);
1399 
1400 	return (0);
1401 }
1402 
1403 static void
1404 mpr_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
1405 {
1406 	struct mpr_softc *sc = arg;
1407 	struct mpr_chain *chain;
1408 	bus_size_t bo;
1409 	int i, o, s;
1410 
1411 	if (error != 0)
1412 		return;
1413 
1414 	for (i = 0, o = 0, s = 0; s < nsegs; s++) {
1415 		for (bo = 0; bo + sc->chain_frame_size <= segs[s].ds_len;
1416 		    bo += sc->chain_frame_size) {
1417 			chain = &sc->chains[i++];
1418 			chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o);
1419 			chain->chain_busaddr = segs[s].ds_addr + bo;
1420 			o += sc->chain_frame_size;
1421 			mpr_free_chain(sc, chain);
1422 		}
1423 		if (bo != segs[s].ds_len)
1424 			o += segs[s].ds_len - bo;
1425 	}
1426 	sc->chain_free_lowwater = i;
1427 }
1428 
1429 static int
1430 mpr_alloc_requests(struct mpr_softc *sc)
1431 {
1432 	bus_dma_template_t t;
1433 	struct mpr_command *cm;
1434 	int i, rsize, nsegs;
1435 
1436 	rsize = sc->reqframesz * sc->num_reqs;
1437 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1438 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize),
1439 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1440 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1441 	if (bus_dma_template_tag(&t, &sc->req_dmat)) {
1442 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request DMA tag\n");
1443 		return (ENOMEM);
1444         }
1445         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
1446 	    BUS_DMA_NOWAIT, &sc->req_map)) {
1447 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate request memory\n");
1448 		return (ENOMEM);
1449         }
1450         bzero(sc->req_frames, rsize);
1451         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
1452 	    mpr_memaddr_cb, &sc->req_busaddr, 0);
1453 	mpr_dprint(sc, MPR_INIT, "request frames busaddr= %#016jx size= %d\n",
1454 	    (uintmax_t)sc->req_busaddr, rsize);
1455 
1456 	sc->chains = malloc(sizeof(struct mpr_chain) * sc->num_chains, M_MPR,
1457 	    M_NOWAIT | M_ZERO);
1458 	if (!sc->chains) {
1459 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1460 		return (ENOMEM);
1461 	}
1462 	rsize = sc->chain_frame_size * sc->num_chains;
1463 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1464 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize),
1465 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS((howmany(rsize, PAGE_SIZE))));
1466 	if (bus_dma_template_tag(&t, &sc->chain_dmat)) {
1467 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain DMA tag\n");
1468 		return (ENOMEM);
1469 	}
1470 	if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
1471 	    BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) {
1472 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n");
1473 		return (ENOMEM);
1474 	}
1475 	if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames,
1476 	    rsize, mpr_load_chains_cb, sc, BUS_DMA_NOWAIT)) {
1477 		mpr_dprint(sc, MPR_ERROR, "Cannot load chain memory\n");
1478 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,
1479 		    sc->chain_map);
1480 		return (ENOMEM);
1481 	}
1482 
1483 	rsize = MPR_SENSE_LEN * sc->num_reqs;
1484 	bus_dma_template_clone(&t, sc->req_dmat);
1485 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(1), BD_MAXSIZE(rsize),
1486 	    BD_MAXSEGSIZE(rsize));
1487 	if (bus_dma_template_tag(&t, &sc->sense_dmat)) {
1488 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense DMA tag\n");
1489 		return (ENOMEM);
1490         }
1491         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
1492 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
1493 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense memory\n");
1494 		return (ENOMEM);
1495         }
1496         bzero(sc->sense_frames, rsize);
1497         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
1498 	    mpr_memaddr_cb, &sc->sense_busaddr, 0);
1499 	mpr_dprint(sc, MPR_INIT, "sense frames busaddr= %#016jx size= %d\n",
1500 	    (uintmax_t)sc->sense_busaddr, rsize);
1501 
1502 	/*
1503 	 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports
1504 	 * these devices.
1505 	 */
1506 	if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) &&
1507 	    (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) {
1508 		if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM)
1509 			return (ENOMEM);
1510 	}
1511 
1512 	nsegs = (sc->maxio / PAGE_SIZE) + 1;
1513 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1514 	BUS_DMA_TEMPLATE_FILL(&t, BD_MAXSIZE(BUS_SPACE_MAXSIZE_32BIT),
1515 	    BD_NSEGMENTS(nsegs), BD_MAXSEGSIZE(BUS_SPACE_MAXSIZE_32BIT),
1516 	    BD_FLAGS(BUS_DMA_ALLOCNOW), BD_LOCKFUNC(busdma_lock_mutex),
1517 	    BD_LOCKFUNCARG(&sc->mpr_mtx));
1518 	if (bus_dma_template_tag(&t, &sc->buffer_dmat)) {
1519 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate buffer DMA tag\n");
1520 		return (ENOMEM);
1521         }
1522 
1523 	/*
1524 	 * SMID 0 cannot be used as a free command per the firmware spec.
1525 	 * Just drop that command instead of risking accounting bugs.
1526 	 */
1527 	sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs,
1528 	    M_MPR, M_WAITOK | M_ZERO);
1529 	for (i = 1; i < sc->num_reqs; i++) {
1530 		cm = &sc->commands[i];
1531 		cm->cm_req = sc->req_frames + i * sc->reqframesz;
1532 		cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz;
1533 		cm->cm_sense = &sc->sense_frames[i];
1534 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN;
1535 		cm->cm_desc.Default.SMID = i;
1536 		cm->cm_sc = sc;
1537 		cm->cm_state = MPR_CM_STATE_BUSY;
1538 		TAILQ_INIT(&cm->cm_chain_list);
1539 		TAILQ_INIT(&cm->cm_prp_page_list);
1540 		callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0);
1541 
1542 		/* XXX Is a failure here a critical problem? */
1543 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap)
1544 		    == 0) {
1545 			if (i <= sc->num_prireqs)
1546 				mpr_free_high_priority_command(sc, cm);
1547 			else
1548 				mpr_free_command(sc, cm);
1549 		} else {
1550 			panic("failed to allocate command %d\n", i);
1551 			sc->num_reqs = i;
1552 			break;
1553 		}
1554 	}
1555 
1556 	return (0);
1557 }
1558 
1559 /*
1560  * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs,
1561  * which are scatter/gather lists for NVMe devices.
1562  *
1563  * This buffer must be contiguous due to the nature of how NVMe PRPs are built
1564  * and translated by FW.
1565  *
1566  * returns ENOMEM if memory could not be allocated, otherwise returns 0.
1567  */
1568 static int
1569 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc)
1570 {
1571 	bus_dma_template_t t;
1572 	struct mpr_prp_page *prp_page;
1573 	int PRPs_per_page, PRPs_required, pages_required;
1574 	int rsize, i;
1575 
1576 	/*
1577 	 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number
1578 	 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is:
1579 	 * MAX_IO_SIZE / PAGE_SIZE = 256
1580 	 *
1581 	 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs
1582 	 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one
1583 	 * page (4096 / 8 = 512), so only one page is required for each I/O.
1584 	 *
1585 	 * Each of these buffers will need to be contiguous. For simplicity,
1586 	 * only one buffer is allocated here, which has all of the space
1587 	 * required for the NVMe Queue Depth. If there are problems allocating
1588 	 * this one buffer, this function will need to change to allocate
1589 	 * individual, contiguous NVME_QDEPTH buffers.
1590 	 *
1591 	 * The real calculation will use the real max io size. Above is just an
1592 	 * example.
1593 	 *
1594 	 */
1595 	PRPs_required = sc->maxio / PAGE_SIZE;
1596 	PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1;
1597 	pages_required = (PRPs_required / PRPs_per_page) + 1;
1598 
1599 	sc->prp_buffer_size = PAGE_SIZE * pages_required;
1600 	rsize = sc->prp_buffer_size * NVME_QDEPTH;
1601 	bus_dma_template_init(&t, sc->mpr_parent_dmat);
1602 	BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize),
1603 	    BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1),
1604 	    BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT));
1605 	if (bus_dma_template_tag(&t, &sc->prp_page_dmat)) {
1606 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP DMA "
1607 		    "tag\n");
1608 		return (ENOMEM);
1609 	}
1610 	if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages,
1611 	    BUS_DMA_NOWAIT, &sc->prp_page_map)) {
1612 		mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP memory\n");
1613 		return (ENOMEM);
1614 	}
1615 	bzero(sc->prp_pages, rsize);
1616 	bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages,
1617 	    rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0);
1618 
1619 	sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR,
1620 	    M_WAITOK | M_ZERO);
1621 	for (i = 0; i < NVME_QDEPTH; i++) {
1622 		prp_page = &sc->prps[i];
1623 		prp_page->prp_page = (uint64_t *)(sc->prp_pages +
1624 		    i * sc->prp_buffer_size);
1625 		prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr +
1626 		    i * sc->prp_buffer_size);
1627 		mpr_free_prp_page(sc, prp_page);
1628 		sc->prp_pages_free_lowwater++;
1629 	}
1630 
1631 	return (0);
1632 }
1633 
1634 static int
1635 mpr_init_queues(struct mpr_softc *sc)
1636 {
1637 	int i;
1638 
1639 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
1640 
1641 	/*
1642 	 * According to the spec, we need to use one less reply than we
1643 	 * have space for on the queue.  So sc->num_replies (the number we
1644 	 * use) should be less than sc->fqdepth (allocated size).
1645 	 */
1646 	if (sc->num_replies >= sc->fqdepth)
1647 		return (EINVAL);
1648 
1649 	/*
1650 	 * Initialize all of the free queue entries.
1651 	 */
1652 	for (i = 0; i < sc->fqdepth; i++) {
1653 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->replyframesz);
1654 	}
1655 	sc->replyfreeindex = sc->num_replies;
1656 
1657 	return (0);
1658 }
1659 
1660 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1661  * Next are the global settings, if they exist.  Highest are the per-unit
1662  * settings, if they exist.
1663  */
1664 void
1665 mpr_get_tunables(struct mpr_softc *sc)
1666 {
1667 	char tmpstr[80], mpr_debug[80];
1668 
1669 	/* XXX default to some debugging for now */
1670 	sc->mpr_debug = MPR_INFO | MPR_FAULT;
1671 	sc->disable_msix = 0;
1672 	sc->disable_msi = 0;
1673 	sc->max_msix = MPR_MSIX_MAX;
1674 	sc->max_chains = MPR_CHAIN_FRAMES;
1675 	sc->max_io_pages = MPR_MAXIO_PAGES;
1676 	sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD;
1677 	sc->spinup_wait_time = DEFAULT_SPINUP_WAIT;
1678 	sc->use_phynum = 1;
1679 	sc->max_reqframes = MPR_REQ_FRAMES;
1680 	sc->max_prireqframes = MPR_PRI_REQ_FRAMES;
1681 	sc->max_replyframes = MPR_REPLY_FRAMES;
1682 	sc->max_evtframes = MPR_EVT_REPLY_FRAMES;
1683 
1684 	/*
1685 	 * Grab the global variables.
1686 	 */
1687 	bzero(mpr_debug, 80);
1688 	if (TUNABLE_STR_FETCH("hw.mpr.debug_level", mpr_debug, 80) != 0)
1689 		mpr_parse_debug(sc, mpr_debug);
1690 	TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix);
1691 	TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi);
1692 	TUNABLE_INT_FETCH("hw.mpr.max_msix", &sc->max_msix);
1693 	TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains);
1694 	TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages);
1695 	TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu);
1696 	TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time);
1697 	TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum);
1698 	TUNABLE_INT_FETCH("hw.mpr.max_reqframes", &sc->max_reqframes);
1699 	TUNABLE_INT_FETCH("hw.mpr.max_prireqframes", &sc->max_prireqframes);
1700 	TUNABLE_INT_FETCH("hw.mpr.max_replyframes", &sc->max_replyframes);
1701 	TUNABLE_INT_FETCH("hw.mpr.max_evtframes", &sc->max_evtframes);
1702 
1703 	/* Grab the unit-instance variables */
1704 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level",
1705 	    device_get_unit(sc->mpr_dev));
1706 	bzero(mpr_debug, 80);
1707 	if (TUNABLE_STR_FETCH(tmpstr, mpr_debug, 80) != 0)
1708 		mpr_parse_debug(sc, mpr_debug);
1709 
1710 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix",
1711 	    device_get_unit(sc->mpr_dev));
1712 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1713 
1714 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi",
1715 	    device_get_unit(sc->mpr_dev));
1716 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1717 
1718 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_msix",
1719 	    device_get_unit(sc->mpr_dev));
1720 	TUNABLE_INT_FETCH(tmpstr, &sc->max_msix);
1721 
1722 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains",
1723 	    device_get_unit(sc->mpr_dev));
1724 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1725 
1726 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages",
1727 	    device_get_unit(sc->mpr_dev));
1728 	TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages);
1729 
1730 	bzero(sc->exclude_ids, sizeof(sc->exclude_ids));
1731 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids",
1732 	    device_get_unit(sc->mpr_dev));
1733 	TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids));
1734 
1735 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu",
1736 	    device_get_unit(sc->mpr_dev));
1737 	TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu);
1738 
1739 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time",
1740 	    device_get_unit(sc->mpr_dev));
1741 	TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time);
1742 
1743 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num",
1744 	    device_get_unit(sc->mpr_dev));
1745 	TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum);
1746 
1747 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_reqframes",
1748 	    device_get_unit(sc->mpr_dev));
1749 	TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes);
1750 
1751 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_prireqframes",
1752 	    device_get_unit(sc->mpr_dev));
1753 	TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes);
1754 
1755 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_replyframes",
1756 	    device_get_unit(sc->mpr_dev));
1757 	TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes);
1758 
1759 	snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_evtframes",
1760 	    device_get_unit(sc->mpr_dev));
1761 	TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes);
1762 }
1763 
1764 static void
1765 mpr_setup_sysctl(struct mpr_softc *sc)
1766 {
1767 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1768 	struct sysctl_oid	*sysctl_tree = NULL;
1769 	char tmpstr[80], tmpstr2[80];
1770 
1771 	/*
1772 	 * Setup the sysctl variable so the user can change the debug level
1773 	 * on the fly.
1774 	 */
1775 	snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d",
1776 	    device_get_unit(sc->mpr_dev));
1777 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev));
1778 
1779 	sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev);
1780 	if (sysctl_ctx != NULL)
1781 		sysctl_tree = device_get_sysctl_tree(sc->mpr_dev);
1782 
1783 	if (sysctl_tree == NULL) {
1784 		sysctl_ctx_init(&sc->sysctl_ctx);
1785 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1786 		    SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2,
1787 		    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr);
1788 		if (sc->sysctl_tree == NULL)
1789 			return;
1790 		sysctl_ctx = &sc->sysctl_ctx;
1791 		sysctl_tree = sc->sysctl_tree;
1792 	}
1793 
1794 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1795 	    OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1796 	    sc, 0, mpr_debug_sysctl, "A", "mpr debug level");
1797 
1798 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1799 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1800 	    "Disable the use of MSI-X interrupts");
1801 
1802 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1803 	    OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0,
1804 	    "User-defined maximum number of MSIX queues");
1805 
1806 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1807 	    OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0,
1808 	    "Negotiated number of MSIX queues");
1809 
1810 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1811 	    OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0,
1812 	    "Total number of allocated request frames");
1813 
1814 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1815 	    OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0,
1816 	    "Total number of allocated high priority request frames");
1817 
1818 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1819 	    OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0,
1820 	    "Total number of allocated reply frames");
1821 
1822 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1823 	    OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0,
1824 	    "Total number of event frames allocated");
1825 
1826 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1827 	    OID_AUTO, "firmware_version", CTLFLAG_RD, sc->fw_version,
1828 	    strlen(sc->fw_version), "firmware version");
1829 
1830 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1831 	    OID_AUTO, "driver_version", CTLFLAG_RD, MPR_DRIVER_VERSION,
1832 	    strlen(MPR_DRIVER_VERSION), "driver version");
1833 
1834 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1835 	    OID_AUTO, "msg_version", CTLFLAG_RD, sc->msg_version,
1836 	    strlen(sc->msg_version), "message interface version");
1837 
1838 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1839 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1840 	    &sc->io_cmds_active, 0, "number of currently active commands");
1841 
1842 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1843 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1844 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1845 
1846 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1847 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1848 	    &sc->chain_free, 0, "number of free chain elements");
1849 
1850 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1851 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1852 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1853 
1854 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1855 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1856 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1857 
1858 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1859 	    OID_AUTO, "max_io_pages", CTLFLAG_RD,
1860 	    &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use "
1861 	    "IOCFacts)");
1862 
1863 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1864 	    OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0,
1865 	    "enable SSU to SATA SSD/HDD at shutdown");
1866 
1867 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1868 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1869 	    &sc->chain_alloc_fail, "chain allocation failures");
1870 
1871 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1872 	    OID_AUTO, "spinup_wait_time", CTLFLAG_RD,
1873 	    &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for "
1874 	    "spinup after SATA ID error");
1875 
1876 	SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1877 	    OID_AUTO, "dump_reqs",
1878 	    CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT,
1879 	    sc, 0, mpr_dump_reqs, "I", "Dump Active Requests");
1880 
1881 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1882 	    OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0,
1883 	    "Use the phy number for enumeration");
1884 
1885 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1886 	    OID_AUTO, "prp_pages_free", CTLFLAG_RD,
1887 	    &sc->prp_pages_free, 0, "number of free PRP pages");
1888 
1889 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1890 	    OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD,
1891 	    &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages");
1892 
1893 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1894 	    OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD,
1895 	    &sc->prp_page_alloc_fail, "PRP page allocation failures");
1896 }
1897 
1898 static struct mpr_debug_string {
1899 	char *name;
1900 	int flag;
1901 } mpr_debug_strings[] = {
1902 	{"info", MPR_INFO},
1903 	{"fault", MPR_FAULT},
1904 	{"event", MPR_EVENT},
1905 	{"log", MPR_LOG},
1906 	{"recovery", MPR_RECOVERY},
1907 	{"error", MPR_ERROR},
1908 	{"init", MPR_INIT},
1909 	{"xinfo", MPR_XINFO},
1910 	{"user", MPR_USER},
1911 	{"mapping", MPR_MAPPING},
1912 	{"trace", MPR_TRACE}
1913 };
1914 
1915 enum mpr_debug_level_combiner {
1916 	COMB_NONE,
1917 	COMB_ADD,
1918 	COMB_SUB
1919 };
1920 
1921 static int
1922 mpr_debug_sysctl(SYSCTL_HANDLER_ARGS)
1923 {
1924 	struct mpr_softc *sc;
1925 	struct mpr_debug_string *string;
1926 	struct sbuf *sbuf;
1927 	char *buffer;
1928 	size_t sz;
1929 	int i, len, debug, error;
1930 
1931 	sc = (struct mpr_softc *)arg1;
1932 
1933 	error = sysctl_wire_old_buffer(req, 0);
1934 	if (error != 0)
1935 		return (error);
1936 
1937 	sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req);
1938 	debug = sc->mpr_debug;
1939 
1940 	sbuf_printf(sbuf, "%#x", debug);
1941 
1942 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1943 	for (i = 0; i < sz; i++) {
1944 		string = &mpr_debug_strings[i];
1945 		if (debug & string->flag)
1946 			sbuf_printf(sbuf, ",%s", string->name);
1947 	}
1948 
1949 	error = sbuf_finish(sbuf);
1950 	sbuf_delete(sbuf);
1951 
1952 	if (error || req->newptr == NULL)
1953 		return (error);
1954 
1955 	len = req->newlen - req->newidx;
1956 	if (len == 0)
1957 		return (0);
1958 
1959 	buffer = malloc(len, M_MPR, M_ZERO|M_WAITOK);
1960 	error = SYSCTL_IN(req, buffer, len);
1961 
1962 	mpr_parse_debug(sc, buffer);
1963 
1964 	free(buffer, M_MPR);
1965 	return (error);
1966 }
1967 
1968 static void
1969 mpr_parse_debug(struct mpr_softc *sc, char *list)
1970 {
1971 	struct mpr_debug_string *string;
1972 	enum mpr_debug_level_combiner op;
1973 	char *token, *endtoken;
1974 	size_t sz;
1975 	int flags, i;
1976 
1977 	if (list == NULL || *list == '\0')
1978 		return;
1979 
1980 	if (*list == '+') {
1981 		op = COMB_ADD;
1982 		list++;
1983 	} else if (*list == '-') {
1984 		op = COMB_SUB;
1985 		list++;
1986 	} else
1987 		op = COMB_NONE;
1988 	if (*list == '\0')
1989 		return;
1990 
1991 	flags = 0;
1992 	sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]);
1993 	while ((token = strsep(&list, ":,")) != NULL) {
1994 		/* Handle integer flags */
1995 		flags |= strtol(token, &endtoken, 0);
1996 		if (token != endtoken)
1997 			continue;
1998 
1999 		/* Handle text flags */
2000 		for (i = 0; i < sz; i++) {
2001 			string = &mpr_debug_strings[i];
2002 			if (strcasecmp(token, string->name) == 0) {
2003 				flags |= string->flag;
2004 				break;
2005 			}
2006 		}
2007 	}
2008 
2009 	switch (op) {
2010 	case COMB_NONE:
2011 		sc->mpr_debug = flags;
2012 		break;
2013 	case COMB_ADD:
2014 		sc->mpr_debug |= flags;
2015 		break;
2016 	case COMB_SUB:
2017 		sc->mpr_debug &= (~flags);
2018 		break;
2019 	}
2020 	return;
2021 }
2022 
2023 struct mpr_dumpreq_hdr {
2024 	uint32_t	smid;
2025 	uint32_t	state;
2026 	uint32_t	numframes;
2027 	uint32_t	deschi;
2028 	uint32_t	desclo;
2029 };
2030 
2031 static int
2032 mpr_dump_reqs(SYSCTL_HANDLER_ARGS)
2033 {
2034 	struct mpr_softc *sc;
2035 	struct mpr_chain *chain, *chain1;
2036 	struct mpr_command *cm;
2037 	struct mpr_dumpreq_hdr hdr;
2038 	struct sbuf *sb;
2039 	uint32_t smid, state;
2040 	int i, numreqs, error = 0;
2041 
2042 	sc = (struct mpr_softc *)arg1;
2043 
2044 	if ((error = priv_check(curthread, PRIV_DRIVER)) != 0) {
2045 		printf("priv check error %d\n", error);
2046 		return (error);
2047 	}
2048 
2049 	state = MPR_CM_STATE_INQUEUE;
2050 	smid = 1;
2051 	numreqs = sc->num_reqs;
2052 
2053 	if (req->newptr != NULL)
2054 		return (EINVAL);
2055 
2056 	if (smid == 0 || smid > sc->num_reqs)
2057 		return (EINVAL);
2058 	if (numreqs <= 0 || (numreqs + smid > sc->num_reqs))
2059 		numreqs = sc->num_reqs;
2060 	sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req);
2061 
2062 	/* Best effort, no locking */
2063 	for (i = smid; i < numreqs; i++) {
2064 		cm = &sc->commands[i];
2065 		if (cm->cm_state != state)
2066 			continue;
2067 		hdr.smid = i;
2068 		hdr.state = cm->cm_state;
2069 		hdr.numframes = 1;
2070 		hdr.deschi = cm->cm_desc.Words.High;
2071 		hdr.desclo = cm->cm_desc.Words.Low;
2072 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
2073 		   chain1)
2074 			hdr.numframes++;
2075 		sbuf_bcat(sb, &hdr, sizeof(hdr));
2076 		sbuf_bcat(sb, cm->cm_req, 128);
2077 		TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link,
2078 		    chain1)
2079 			sbuf_bcat(sb, chain->chain, 128);
2080 	}
2081 
2082 	error = sbuf_finish(sb);
2083 	sbuf_delete(sb);
2084 	return (error);
2085 }
2086 
2087 int
2088 mpr_attach(struct mpr_softc *sc)
2089 {
2090 	int error;
2091 
2092 	MPR_FUNCTRACE(sc);
2093 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2094 
2095 	mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF);
2096 	callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0);
2097 	callout_init_mtx(&sc->device_check_callout, &sc->mpr_mtx, 0);
2098 	TAILQ_INIT(&sc->event_list);
2099 	timevalclear(&sc->lastfail);
2100 
2101 	if ((error = mpr_transition_ready(sc)) != 0) {
2102 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2103 		    "Failed to transition ready\n");
2104 		return (error);
2105 	}
2106 
2107 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR,
2108 	    M_ZERO|M_NOWAIT);
2109 	if (!sc->facts) {
2110 		mpr_dprint(sc, MPR_INIT|MPR_FAULT,
2111 		    "Cannot allocate memory, exit\n");
2112 		return (ENOMEM);
2113 	}
2114 
2115 	/*
2116 	 * Get IOC Facts and allocate all structures based on this information.
2117 	 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC
2118 	 * Facts. If relevant values have changed in IOC Facts, this function
2119 	 * will free all of the memory based on IOC Facts and reallocate that
2120 	 * memory.  If this fails, any allocated memory should already be freed.
2121 	 */
2122 	if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) {
2123 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC Facts allocation "
2124 		    "failed with error %d\n", error);
2125 		return (error);
2126 	}
2127 
2128 	/* Start the periodic watchdog check on the IOC Doorbell */
2129 	mpr_periodic(sc);
2130 
2131 	/*
2132 	 * The portenable will kick off discovery events that will drive the
2133 	 * rest of the initialization process.  The CAM/SAS module will
2134 	 * hold up the boot sequence until discovery is complete.
2135 	 */
2136 	sc->mpr_ich.ich_func = mpr_startup;
2137 	sc->mpr_ich.ich_arg = sc;
2138 	if (config_intrhook_establish(&sc->mpr_ich) != 0) {
2139 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2140 		    "Cannot establish MPR config hook\n");
2141 		error = EINVAL;
2142 	}
2143 
2144 	/*
2145 	 * Allow IR to shutdown gracefully when shutdown occurs.
2146 	 */
2147 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
2148 	    mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
2149 
2150 	if (sc->shutdown_eh == NULL)
2151 		mpr_dprint(sc, MPR_INIT|MPR_ERROR,
2152 		    "shutdown event registration failed\n");
2153 
2154 	mpr_setup_sysctl(sc);
2155 
2156 	sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE;
2157 	mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error);
2158 
2159 	return (error);
2160 }
2161 
2162 /* Run through any late-start handlers. */
2163 static void
2164 mpr_startup(void *arg)
2165 {
2166 	struct mpr_softc *sc;
2167 
2168 	sc = (struct mpr_softc *)arg;
2169 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2170 
2171 	mpr_lock(sc);
2172 	mpr_unmask_intr(sc);
2173 
2174 	/* initialize device mapping tables */
2175 	mpr_base_static_config_pages(sc);
2176 	mpr_mapping_initialize(sc);
2177 	mprsas_startup(sc);
2178 	mpr_unlock(sc);
2179 
2180 	mpr_dprint(sc, MPR_INIT, "disestablish config intrhook\n");
2181 	config_intrhook_disestablish(&sc->mpr_ich);
2182 	sc->mpr_ich.ich_arg = NULL;
2183 
2184 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2185 }
2186 
2187 /* Periodic watchdog.  Is called with the driver lock already held. */
2188 static void
2189 mpr_periodic(void *arg)
2190 {
2191 	struct mpr_softc *sc;
2192 	uint32_t db;
2193 
2194 	sc = (struct mpr_softc *)arg;
2195 	if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN)
2196 		return;
2197 
2198 	db = mpr_regread(sc, MPI2_DOORBELL_OFFSET);
2199 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
2200 		if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) ==
2201 		    IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) {
2202 			panic("TEMPERATURE FAULT: STOPPING.");
2203 		}
2204 		mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db);
2205 		mpr_reinit(sc);
2206 	}
2207 
2208 	callout_reset(&sc->periodic, MPR_PERIODIC_DELAY * hz, mpr_periodic, sc);
2209 }
2210 
2211 static void
2212 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data,
2213     MPI2_EVENT_NOTIFICATION_REPLY *event)
2214 {
2215 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
2216 
2217 	MPR_DPRINT_EVENT(sc, generic, event);
2218 
2219 	switch (event->Event) {
2220 	case MPI2_EVENT_LOG_DATA:
2221 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n");
2222 		if (sc->mpr_debug & MPR_EVENT)
2223 			hexdump(event->EventData, event->EventDataLength, NULL,
2224 			    0);
2225 		break;
2226 	case MPI2_EVENT_LOG_ENTRY_ADDED:
2227 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
2228 		mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event "
2229 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
2230 		     entry->LogSequence);
2231 		break;
2232 	default:
2233 		break;
2234 	}
2235 	return;
2236 }
2237 
2238 static int
2239 mpr_attach_log(struct mpr_softc *sc)
2240 {
2241 	uint8_t events[16];
2242 
2243 	bzero(events, 16);
2244 	setbit(events, MPI2_EVENT_LOG_DATA);
2245 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
2246 
2247 	mpr_register_events(sc, events, mpr_log_evt_handler, NULL,
2248 	    &sc->mpr_log_eh);
2249 
2250 	return (0);
2251 }
2252 
2253 static int
2254 mpr_detach_log(struct mpr_softc *sc)
2255 {
2256 
2257 	if (sc->mpr_log_eh != NULL)
2258 		mpr_deregister_events(sc, sc->mpr_log_eh);
2259 	return (0);
2260 }
2261 
2262 /*
2263  * Free all of the driver resources and detach submodules.  Should be called
2264  * without the lock held.
2265  */
2266 int
2267 mpr_free(struct mpr_softc *sc)
2268 {
2269 	int error;
2270 
2271 	mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__);
2272 	/* Turn off the watchdog */
2273 	mpr_lock(sc);
2274 	sc->mpr_flags |= MPR_FLAGS_SHUTDOWN;
2275 	mpr_unlock(sc);
2276 	/* Lock must not be held for this */
2277 	callout_drain(&sc->periodic);
2278 	callout_drain(&sc->device_check_callout);
2279 
2280 	if (((error = mpr_detach_log(sc)) != 0) ||
2281 	    ((error = mpr_detach_sas(sc)) != 0)) {
2282 		mpr_dprint(sc, MPR_INIT|MPR_FAULT, "failed to detach "
2283 		    "subsystems, error= %d, exit\n", error);
2284 		return (error);
2285 	}
2286 
2287 	mpr_detach_user(sc);
2288 
2289 	/* Put the IOC back in the READY state. */
2290 	mpr_lock(sc);
2291 	if ((error = mpr_transition_ready(sc)) != 0) {
2292 		mpr_unlock(sc);
2293 		return (error);
2294 	}
2295 	mpr_unlock(sc);
2296 
2297 	if (sc->facts != NULL)
2298 		free(sc->facts, M_MPR);
2299 
2300 	/*
2301 	 * Free all buffers that are based on IOC Facts.  A Diag Reset may need
2302 	 * to free these buffers too.
2303 	 */
2304 	mpr_iocfacts_free(sc);
2305 
2306 	if (sc->sysctl_tree != NULL)
2307 		sysctl_ctx_free(&sc->sysctl_ctx);
2308 
2309 	/* Deregister the shutdown function */
2310 	if (sc->shutdown_eh != NULL)
2311 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
2312 
2313 	mtx_destroy(&sc->mpr_mtx);
2314 	mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__);
2315 
2316 	return (0);
2317 }
2318 
2319 static __inline void
2320 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm)
2321 {
2322 	MPR_FUNCTRACE(sc);
2323 
2324 	if (cm == NULL) {
2325 		mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n");
2326 		return;
2327 	}
2328 
2329 	cm->cm_state = MPR_CM_STATE_BUSY;
2330 	if (cm->cm_flags & MPR_CM_FLAGS_POLLED)
2331 		cm->cm_flags |= MPR_CM_FLAGS_COMPLETE;
2332 
2333 	if (cm->cm_complete != NULL) {
2334 		mpr_dprint(sc, MPR_TRACE,
2335 		    "%s cm %p calling cm_complete %p data %p reply %p\n",
2336 		    __func__, cm, cm->cm_complete, cm->cm_complete_data,
2337 		    cm->cm_reply);
2338 		cm->cm_complete(sc, cm);
2339 	}
2340 
2341 	if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) {
2342 		mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm);
2343 		wakeup(cm);
2344 	}
2345 
2346 	if (sc->io_cmds_active != 0) {
2347 		sc->io_cmds_active--;
2348 	} else {
2349 		mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is "
2350 		    "out of sync - resynching to 0\n");
2351 	}
2352 }
2353 
2354 static void
2355 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info)
2356 {
2357 	union loginfo_type {
2358 		u32	loginfo;
2359 		struct {
2360 			u32	subcode:16;
2361 			u32	code:8;
2362 			u32	originator:4;
2363 			u32	bus_type:4;
2364 		} dw;
2365 	};
2366 	union loginfo_type sas_loginfo;
2367 	char *originator_str = NULL;
2368 
2369 	sas_loginfo.loginfo = log_info;
2370 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
2371 		return;
2372 
2373 	/* each nexus loss loginfo */
2374 	if (log_info == 0x31170000)
2375 		return;
2376 
2377 	/* eat the loginfos associated with task aborts */
2378 	if ((log_info == 30050000) || (log_info == 0x31140000) ||
2379 	    (log_info == 0x31130000))
2380 		return;
2381 
2382 	switch (sas_loginfo.dw.originator) {
2383 	case 0:
2384 		originator_str = "IOP";
2385 		break;
2386 	case 1:
2387 		originator_str = "PL";
2388 		break;
2389 	case 2:
2390 		originator_str = "IR";
2391 		break;
2392 	}
2393 
2394 	mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), "
2395 	    "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str,
2396 	    sas_loginfo.dw.code, sas_loginfo.dw.subcode);
2397 }
2398 
2399 static void
2400 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply)
2401 {
2402 	MPI2DefaultReply_t *mpi_reply;
2403 	u16 sc_status;
2404 
2405 	mpi_reply = (MPI2DefaultReply_t*)reply;
2406 	sc_status = le16toh(mpi_reply->IOCStatus);
2407 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
2408 		mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
2409 }
2410 
2411 void
2412 mpr_intr(void *data)
2413 {
2414 	struct mpr_softc *sc;
2415 	uint32_t status;
2416 
2417 	sc = (struct mpr_softc *)data;
2418 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2419 
2420 	/*
2421 	 * Check interrupt status register to flush the bus.  This is
2422 	 * needed for both INTx interrupts and driver-driven polling
2423 	 */
2424 	status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
2425 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
2426 		return;
2427 
2428 	mpr_lock(sc);
2429 	mpr_intr_locked(data);
2430 	mpr_unlock(sc);
2431 	return;
2432 }
2433 
2434 /*
2435  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
2436  * chip.  Hopefully this theory is correct.
2437  */
2438 void
2439 mpr_intr_msi(void *data)
2440 {
2441 	struct mpr_softc *sc;
2442 
2443 	sc = (struct mpr_softc *)data;
2444 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2445 	mpr_lock(sc);
2446 	mpr_intr_locked(data);
2447 	mpr_unlock(sc);
2448 	return;
2449 }
2450 
2451 /*
2452  * The locking is overly broad and simplistic, but easy to deal with for now.
2453  */
2454 void
2455 mpr_intr_locked(void *data)
2456 {
2457 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
2458 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
2459 	mpr_fw_diagnostic_buffer_t *pBuffer;
2460 	struct mpr_softc *sc;
2461 	uint64_t tdesc;
2462 	struct mpr_command *cm = NULL;
2463 	uint8_t flags;
2464 	u_int pq;
2465 
2466 	sc = (struct mpr_softc *)data;
2467 
2468 	pq = sc->replypostindex;
2469 	mpr_dprint(sc, MPR_TRACE,
2470 	    "%s sc %p starting with replypostindex %u\n",
2471 	    __func__, sc, sc->replypostindex);
2472 
2473 	for ( ;; ) {
2474 		cm = NULL;
2475 		desc = &sc->post_queue[sc->replypostindex];
2476 
2477 		/*
2478 		 * Copy and clear out the descriptor so that any reentry will
2479 		 * immediately know that this descriptor has already been
2480 		 * looked at.  There is unfortunate casting magic because the
2481 		 * MPI API doesn't have a cardinal 64bit type.
2482 		 */
2483 		tdesc = 0xffffffffffffffff;
2484 		tdesc = atomic_swap_64((uint64_t *)desc, tdesc);
2485 		desc = (MPI2_REPLY_DESCRIPTORS_UNION *)&tdesc;
2486 
2487 		flags = desc->Default.ReplyFlags &
2488 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
2489 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) ||
2490 		    (le32toh(desc->Words.High) == 0xffffffff))
2491 			break;
2492 
2493 		/* increment the replypostindex now, so that event handlers
2494 		 * and cm completion handlers which decide to do a diag
2495 		 * reset can zero it without it getting incremented again
2496 		 * afterwards, and we break out of this loop on the next
2497 		 * iteration since the reply post queue has been cleared to
2498 		 * 0xFF and all descriptors look unused (which they are).
2499 		 */
2500 		if (++sc->replypostindex >= sc->pqdepth)
2501 			sc->replypostindex = 0;
2502 
2503 		switch (flags) {
2504 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
2505 		case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS:
2506 		case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS:
2507 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
2508 			KASSERT(cm->cm_state == MPR_CM_STATE_INQUEUE,
2509 			    ("command not inqueue\n"));
2510 			cm->cm_state = MPR_CM_STATE_BUSY;
2511 			cm->cm_reply = NULL;
2512 			break;
2513 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
2514 		{
2515 			uint32_t baddr;
2516 			uint8_t *reply;
2517 
2518 			/*
2519 			 * Re-compose the reply address from the address
2520 			 * sent back from the chip.  The ReplyFrameAddress
2521 			 * is the lower 32 bits of the physical address of
2522 			 * particular reply frame.  Convert that address to
2523 			 * host format, and then use that to provide the
2524 			 * offset against the virtual address base
2525 			 * (sc->reply_frames).
2526 			 */
2527 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
2528 			reply = sc->reply_frames +
2529 				(baddr - ((uint32_t)sc->reply_busaddr));
2530 			/*
2531 			 * Make sure the reply we got back is in a valid
2532 			 * range.  If not, go ahead and panic here, since
2533 			 * we'll probably panic as soon as we deference the
2534 			 * reply pointer anyway.
2535 			 */
2536 			if ((reply < sc->reply_frames)
2537 			 || (reply > (sc->reply_frames +
2538 			     (sc->fqdepth * sc->replyframesz)))) {
2539 				printf("%s: WARNING: reply %p out of range!\n",
2540 				       __func__, reply);
2541 				printf("%s: reply_frames %p, fqdepth %d, "
2542 				       "frame size %d\n", __func__,
2543 				       sc->reply_frames, sc->fqdepth,
2544 				       sc->replyframesz);
2545 				printf("%s: baddr %#x,\n", __func__, baddr);
2546 				/* LSI-TODO. See Linux Code for Graceful exit */
2547 				panic("Reply address out of range");
2548 			}
2549 			if (le16toh(desc->AddressReply.SMID) == 0) {
2550 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
2551 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
2552 					/*
2553 					 * If SMID is 0 for Diag Buffer Post,
2554 					 * this implies that the reply is due to
2555 					 * a release function with a status that
2556 					 * the buffer has been released.  Set
2557 					 * the buffer flags accordingly.
2558 					 */
2559 					rel_rep =
2560 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
2561 					if ((le16toh(rel_rep->IOCStatus) &
2562 					    MPI2_IOCSTATUS_MASK) ==
2563 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
2564 					{
2565 						pBuffer =
2566 						    &sc->fw_diag_buffer_list[
2567 						    rel_rep->BufferType];
2568 						pBuffer->valid_data = TRUE;
2569 						pBuffer->owned_by_firmware =
2570 						    FALSE;
2571 						pBuffer->immediate = FALSE;
2572 					}
2573 				} else
2574 					mpr_dispatch_event(sc, baddr,
2575 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
2576 					    reply);
2577 			} else {
2578 				cm = &sc->commands[
2579 				    le16toh(desc->AddressReply.SMID)];
2580 				if (cm->cm_state == MPR_CM_STATE_INQUEUE) {
2581 					cm->cm_reply = reply;
2582 					cm->cm_reply_data =
2583 					    le32toh(desc->AddressReply.
2584 						ReplyFrameAddress);
2585 				} else {
2586 					mpr_dprint(sc, MPR_RECOVERY,
2587 					    "Bad state for ADDRESS_REPLY status,"
2588 					    " ignoring state %d cm %p\n",
2589 					    cm->cm_state, cm);
2590 				}
2591 			}
2592 			break;
2593 		}
2594 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
2595 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
2596 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
2597 		default:
2598 			/* Unhandled */
2599 			mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n",
2600 			    desc->Default.ReplyFlags);
2601 			cm = NULL;
2602 			break;
2603 		}
2604 
2605 		if (cm != NULL) {
2606 			// Print Error reply frame
2607 			if (cm->cm_reply)
2608 				mpr_display_reply_info(sc,cm->cm_reply);
2609 			mpr_complete_command(sc, cm);
2610 		}
2611 	}
2612 
2613 	if (pq != sc->replypostindex) {
2614 		mpr_dprint(sc, MPR_TRACE, "%s sc %p writing postindex %d\n",
2615 		    __func__, sc, sc->replypostindex);
2616 		mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET,
2617 		    sc->replypostindex);
2618 	}
2619 
2620 	return;
2621 }
2622 
2623 static void
2624 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data,
2625     MPI2_EVENT_NOTIFICATION_REPLY *reply)
2626 {
2627 	struct mpr_event_handle *eh;
2628 	int event, handled = 0;
2629 
2630 	event = le16toh(reply->Event);
2631 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2632 		if (isset(eh->mask, event)) {
2633 			eh->callback(sc, data, reply);
2634 			handled++;
2635 		}
2636 	}
2637 
2638 	if (handled == 0)
2639 		mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n",
2640 		    le16toh(event));
2641 
2642 	/*
2643 	 * This is the only place that the event/reply should be freed.
2644 	 * Anything wanting to hold onto the event data should have
2645 	 * already copied it into their own storage.
2646 	 */
2647 	mpr_free_reply(sc, data);
2648 }
2649 
2650 static void
2651 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm)
2652 {
2653 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2654 
2655 	if (cm->cm_reply)
2656 		MPR_DPRINT_EVENT(sc, generic,
2657 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
2658 
2659 	mpr_free_command(sc, cm);
2660 
2661 	/* next, send a port enable */
2662 	mprsas_startup(sc);
2663 }
2664 
2665 /*
2666  * For both register_events and update_events, the caller supplies a bitmap
2667  * of events that it _wants_.  These functions then turn that into a bitmask
2668  * suitable for the controller.
2669  */
2670 int
2671 mpr_register_events(struct mpr_softc *sc, uint8_t *mask,
2672     mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle)
2673 {
2674 	struct mpr_event_handle *eh;
2675 	int error = 0;
2676 
2677 	eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO);
2678 	eh->callback = cb;
2679 	eh->data = data;
2680 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
2681 	if (mask != NULL)
2682 		error = mpr_update_events(sc, eh, mask);
2683 	*handle = eh;
2684 
2685 	return (error);
2686 }
2687 
2688 int
2689 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle,
2690     uint8_t *mask)
2691 {
2692 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2693 	MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL;
2694 	struct mpr_command *cm = NULL;
2695 	struct mpr_event_handle *eh;
2696 	int error, i;
2697 
2698 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2699 
2700 	if ((mask != NULL) && (handle != NULL))
2701 		bcopy(mask, &handle->mask[0], 16);
2702 	memset(sc->event_mask, 0xff, 16);
2703 
2704 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2705 		for (i = 0; i < 16; i++)
2706 			sc->event_mask[i] &= ~eh->mask[i];
2707 	}
2708 
2709 	if ((cm = mpr_alloc_command(sc)) == NULL)
2710 		return (EBUSY);
2711 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2712 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2713 	evtreq->MsgFlags = 0;
2714 	evtreq->SASBroadcastPrimitiveMasks = 0;
2715 #ifdef MPR_DEBUG_ALL_EVENTS
2716 	{
2717 		u_char fullmask[16];
2718 		memset(fullmask, 0x00, 16);
2719 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2720 	}
2721 #else
2722 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2723 #endif
2724 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2725 	cm->cm_data = NULL;
2726 
2727 	error = mpr_request_polled(sc, &cm);
2728 	if (cm != NULL)
2729 		reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
2730 	if ((reply == NULL) ||
2731 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
2732 		error = ENXIO;
2733 
2734 	if (reply)
2735 		MPR_DPRINT_EVENT(sc, generic, reply);
2736 
2737 	mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error);
2738 
2739 	if (cm != NULL)
2740 		mpr_free_command(sc, cm);
2741 	return (error);
2742 }
2743 
2744 static int
2745 mpr_reregister_events(struct mpr_softc *sc)
2746 {
2747 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
2748 	struct mpr_command *cm;
2749 	struct mpr_event_handle *eh;
2750 	int error, i;
2751 
2752 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
2753 
2754 	/* first, reregister events */
2755 
2756 	memset(sc->event_mask, 0xff, 16);
2757 
2758 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
2759 		for (i = 0; i < 16; i++)
2760 			sc->event_mask[i] &= ~eh->mask[i];
2761 	}
2762 
2763 	if ((cm = mpr_alloc_command(sc)) == NULL)
2764 		return (EBUSY);
2765 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
2766 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
2767 	evtreq->MsgFlags = 0;
2768 	evtreq->SASBroadcastPrimitiveMasks = 0;
2769 #ifdef MPR_DEBUG_ALL_EVENTS
2770 	{
2771 		u_char fullmask[16];
2772 		memset(fullmask, 0x00, 16);
2773 		bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, 16);
2774 	}
2775 #else
2776 		bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, 16);
2777 #endif
2778 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2779 	cm->cm_data = NULL;
2780 	cm->cm_complete = mpr_reregister_events_complete;
2781 
2782 	error = mpr_map_command(sc, cm);
2783 
2784 	mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__,
2785 	    error);
2786 	return (error);
2787 }
2788 
2789 int
2790 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle)
2791 {
2792 
2793 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
2794 	free(handle, M_MPR);
2795 	return (mpr_update_events(sc, NULL, NULL));
2796 }
2797 
2798 /**
2799 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a
2800 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry
2801 * of the NVMe message (PRP1). If the data buffer is small enough to be described
2802 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to
2803 * describe a larger data buffer. If the data buffer is too large to describe
2804 * using the two PRP entriess inside the NVMe message, then PRP1 describes the
2805 * first data memory segment, and PRP2 contains a pointer to a PRP list located
2806 * elsewhere in memory to describe the remaining data memory segments. The PRP
2807 * list will be contiguous.
2808 
2809 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP
2810 * consists of a list of PRP entries to describe a number of noncontigous
2811 * physical memory segments as a single memory buffer, just as a SGL does. Note
2812 * however, that this function is only used by the IOCTL call, so the memory
2813 * given will be guaranteed to be contiguous. There is no need to translate
2814 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous
2815 * space that is one page size each.
2816 *
2817 * Each NVMe message contains two PRP entries. The first (PRP1) either contains
2818 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains
2819 * the second PRP element if the memory being described fits within 2 PRP
2820 * entries, or a PRP list pointer if the PRP spans more than two entries.
2821 *
2822 * A PRP list pointer contains the address of a PRP list, structured as a linear
2823 * array of PRP entries. Each PRP entry in this list describes a segment of
2824 * physical memory.
2825 *
2826 * Each 64-bit PRP entry comprises an address and an offset field. The address
2827 * always points to the beginning of a PAGE_SIZE physical memory page, and the
2828 * offset describes where within that page the memory segment begins. Only the
2829 * first element in a PRP list may contain a non-zero offest, implying that all
2830 * memory segments following the first begin at the start of a PAGE_SIZE page.
2831 *
2832 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory,
2833 * with exceptions for the first and last elements in the list. If the memory
2834 * being described by the list begins at a non-zero offset within the first page,
2835 * then the first PRP element will contain a non-zero offset indicating where the
2836 * region begins within the page. The last memory segment may end before the end
2837 * of the PAGE_SIZE segment, depending upon the overall size of the memory being
2838 * described by the PRP list.
2839 *
2840 * Since PRP entries lack any indication of size, the overall data buffer length
2841 * is used to determine where the end of the data memory buffer is located, and
2842 * how many PRP entries are required to describe it.
2843 *
2844 * Returns nothing.
2845 */
2846 void
2847 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm,
2848     Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data,
2849     uint32_t data_in_sz, uint32_t data_out_sz)
2850 {
2851 	int			prp_size = PRP_ENTRY_SIZE;
2852 	uint64_t		*prp_entry, *prp1_entry, *prp2_entry;
2853 	uint64_t		*prp_entry_phys, *prp_page, *prp_page_phys;
2854 	uint32_t		offset, entry_len, page_mask_result, page_mask;
2855 	bus_addr_t		paddr;
2856 	size_t			length;
2857 	struct mpr_prp_page	*prp_page_info = NULL;
2858 
2859 	/*
2860 	 * Not all commands require a data transfer. If no data, just return
2861 	 * without constructing any PRP.
2862 	 */
2863 	if (!data_in_sz && !data_out_sz)
2864 		return;
2865 
2866 	/*
2867 	 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is
2868 	 * located at a 24 byte offset from the start of the NVMe command. Then
2869 	 * set the current PRP entry pointer to PRP1.
2870 	 */
2871 	prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2872 	    NVME_CMD_PRP1_OFFSET);
2873 	prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command +
2874 	    NVME_CMD_PRP2_OFFSET);
2875 	prp_entry = prp1_entry;
2876 
2877 	/*
2878 	 * For the PRP entries, use the specially allocated buffer of
2879 	 * contiguous memory. PRP Page allocation failures should not happen
2880 	 * because there should be enough PRP page buffers to account for the
2881 	 * possible NVMe QDepth.
2882 	 */
2883 	prp_page_info = mpr_alloc_prp_page(sc);
2884 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
2885 	    "used for building a native NVMe SGL.\n", __func__));
2886 	prp_page = (uint64_t *)prp_page_info->prp_page;
2887 	prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
2888 
2889 	/*
2890 	 * Insert the allocated PRP page into the command's PRP page list. This
2891 	 * will be freed when the command is freed.
2892 	 */
2893 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
2894 
2895 	/*
2896 	 * Check if we are within 1 entry of a page boundary we don't want our
2897 	 * first entry to be a PRP List entry.
2898 	 */
2899 	page_mask = PAGE_SIZE - 1;
2900 	page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) &
2901 	    page_mask;
2902 	if (!page_mask_result)
2903 	{
2904 		/* Bump up to next page boundary. */
2905 		prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size);
2906 		prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys +
2907 		    prp_size);
2908 	}
2909 
2910 	/*
2911 	 * Set PRP physical pointer, which initially points to the current PRP
2912 	 * DMA memory page.
2913 	 */
2914 	prp_entry_phys = prp_page_phys;
2915 
2916 	/* Get physical address and length of the data buffer. */
2917 	paddr = (bus_addr_t)(uintptr_t)data;
2918 	if (data_in_sz)
2919 		length = data_in_sz;
2920 	else
2921 		length = data_out_sz;
2922 
2923 	/* Loop while the length is not zero. */
2924 	while (length)
2925 	{
2926 		/*
2927 		 * Check if we need to put a list pointer here if we are at page
2928 		 * boundary - prp_size (8 bytes).
2929 		 */
2930 		page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys +
2931 		    prp_size) & page_mask;
2932 		if (!page_mask_result)
2933 		{
2934 			/*
2935 			 * This is the last entry in a PRP List, so we need to
2936 			 * put a PRP list pointer here. What this does is:
2937 			 *   - bump the current memory pointer to the next
2938 			 *     address, which will be the next full page.
2939 			 *   - set the PRP Entry to point to that page. This is
2940 			 *     now the PRP List pointer.
2941 			 *   - bump the PRP Entry pointer the start of the next
2942 			 *     page. Since all of this PRP memory is contiguous,
2943 			 *     no need to get a new page - it's just the next
2944 			 *     address.
2945 			 */
2946 			prp_entry_phys++;
2947 			*prp_entry =
2948 			    htole64((uint64_t)(uintptr_t)prp_entry_phys);
2949 			prp_entry++;
2950 		}
2951 
2952 		/* Need to handle if entry will be part of a page. */
2953 		offset = (uint32_t)paddr & page_mask;
2954 		entry_len = PAGE_SIZE - offset;
2955 
2956 		if (prp_entry == prp1_entry)
2957 		{
2958 			/*
2959 			 * Must fill in the first PRP pointer (PRP1) before
2960 			 * moving on.
2961 			 */
2962 			*prp1_entry = htole64((uint64_t)paddr);
2963 
2964 			/*
2965 			 * Now point to the second PRP entry within the
2966 			 * command (PRP2).
2967 			 */
2968 			prp_entry = prp2_entry;
2969 		}
2970 		else if (prp_entry == prp2_entry)
2971 		{
2972 			/*
2973 			 * Should the PRP2 entry be a PRP List pointer or just a
2974 			 * regular PRP pointer? If there is more than one more
2975 			 * page of data, must use a PRP List pointer.
2976 			 */
2977 			if (length > PAGE_SIZE)
2978 			{
2979 				/*
2980 				 * PRP2 will contain a PRP List pointer because
2981 				 * more PRP's are needed with this command. The
2982 				 * list will start at the beginning of the
2983 				 * contiguous buffer.
2984 				 */
2985 				*prp2_entry =
2986 				    htole64(
2987 				    (uint64_t)(uintptr_t)prp_entry_phys);
2988 
2989 				/*
2990 				 * The next PRP Entry will be the start of the
2991 				 * first PRP List.
2992 				 */
2993 				prp_entry = prp_page;
2994 			}
2995 			else
2996 			{
2997 				/*
2998 				 * After this, the PRP Entries are complete.
2999 				 * This command uses 2 PRP's and no PRP list.
3000 				 */
3001 				*prp2_entry = htole64((uint64_t)paddr);
3002 			}
3003 		}
3004 		else
3005 		{
3006 			/*
3007 			 * Put entry in list and bump the addresses.
3008 			 *
3009 			 * After PRP1 and PRP2 are filled in, this will fill in
3010 			 * all remaining PRP entries in a PRP List, one per each
3011 			 * time through the loop.
3012 			 */
3013 			*prp_entry = htole64((uint64_t)paddr);
3014 			prp_entry++;
3015 			prp_entry_phys++;
3016 		}
3017 
3018 		/*
3019 		 * Bump the phys address of the command's data buffer by the
3020 		 * entry_len.
3021 		 */
3022 		paddr += entry_len;
3023 
3024 		/* Decrement length accounting for last partial page. */
3025 		if (entry_len > length)
3026 			length = 0;
3027 		else
3028 			length -= entry_len;
3029 	}
3030 }
3031 
3032 /*
3033  * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to
3034  * determine if the driver needs to build a native SGL. If so, that native SGL
3035  * is built in the contiguous buffers allocated especially for PCIe SGL
3036  * creation. If the driver will not build a native SGL, return TRUE and a
3037  * normal IEEE SGL will be built. Currently this routine supports NVMe devices
3038  * only.
3039  *
3040  * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built.
3041  */
3042 static int
3043 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm,
3044     bus_dma_segment_t *segs, int segs_left)
3045 {
3046 	uint32_t		i, sge_dwords, length, offset, entry_len;
3047 	uint32_t		num_entries, buff_len = 0, sges_in_segment;
3048 	uint32_t		page_mask, page_mask_result, *curr_buff;
3049 	uint32_t		*ptr_sgl, *ptr_first_sgl, first_page_offset;
3050 	uint32_t		first_page_data_size, end_residual;
3051 	uint64_t		*msg_phys;
3052 	bus_addr_t		paddr;
3053 	int			build_native_sgl = 0, first_prp_entry;
3054 	int			prp_size = PRP_ENTRY_SIZE;
3055 	Mpi25IeeeSgeChain64_t	*main_chain_element = NULL;
3056 	struct mpr_prp_page	*prp_page_info = NULL;
3057 
3058 	mpr_dprint(sc, MPR_TRACE, "%s\n", __func__);
3059 
3060 	/*
3061 	 * Add up the sizes of each segment length to get the total transfer
3062 	 * size, which will be checked against the Maximum Data Transfer Size.
3063 	 * If the data transfer length exceeds the MDTS for this device, just
3064 	 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O
3065 	 * up into multiple I/O's. [nvme_mdts = 0 means unlimited]
3066 	 */
3067 	for (i = 0; i < segs_left; i++)
3068 		buff_len += htole32(segs[i].ds_len);
3069 	if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS))
3070 		return 1;
3071 
3072 	/* Create page_mask (to get offset within page) */
3073 	page_mask = PAGE_SIZE - 1;
3074 
3075 	/*
3076 	 * Check if the number of elements exceeds the max number that can be
3077 	 * put in the main message frame (H/W can only translate an SGL that
3078 	 * is contained entirely in the main message frame).
3079 	 */
3080 	sges_in_segment = (sc->reqframesz -
3081 	    offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION);
3082 	if (segs_left > sges_in_segment)
3083 		build_native_sgl = 1;
3084 	else
3085 	{
3086 		/*
3087 		 * NVMe uses one PRP for each physical page (or part of physical
3088 		 * page).
3089 		 *    if 4 pages or less then IEEE is OK
3090 		 *    if > 5 pages then we need to build a native SGL
3091 		 *    if > 4 and <= 5 pages, then check the physical address of
3092 		 *      the first SG entry, then if this first size in the page
3093 		 *      is >= the residual beyond 4 pages then use IEEE,
3094 		 *      otherwise use native SGL
3095 		 */
3096 		if (buff_len > (PAGE_SIZE * 5))
3097 			build_native_sgl = 1;
3098 		else if ((buff_len > (PAGE_SIZE * 4)) &&
3099 		    (buff_len <= (PAGE_SIZE * 5)) )
3100 		{
3101 			msg_phys = (uint64_t *)(uintptr_t)segs[0].ds_addr;
3102 			first_page_offset =
3103 			    ((uint32_t)(uint64_t)(uintptr_t)msg_phys &
3104 			    page_mask);
3105 			first_page_data_size = PAGE_SIZE - first_page_offset;
3106 			end_residual = buff_len % PAGE_SIZE;
3107 
3108 			/*
3109 			 * If offset into first page pushes the end of the data
3110 			 * beyond end of the 5th page, we need the extra PRP
3111 			 * list.
3112 			 */
3113 			if (first_page_data_size < end_residual)
3114 				build_native_sgl = 1;
3115 
3116 			/*
3117 			 * Check if first SG entry size is < residual beyond 4
3118 			 * pages.
3119 			 */
3120 			if (htole32(segs[0].ds_len) <
3121 			    (buff_len - (PAGE_SIZE * 4)))
3122 				build_native_sgl = 1;
3123 		}
3124 	}
3125 
3126 	/* check if native SGL is needed */
3127 	if (!build_native_sgl)
3128 		return 1;
3129 
3130 	/*
3131 	 * Native SGL is needed.
3132 	 * Put a chain element in main message frame that points to the first
3133 	 * chain buffer.
3134 	 *
3135 	 * NOTE:  The ChainOffset field must be 0 when using a chain pointer to
3136 	 *        a native SGL.
3137 	 */
3138 
3139 	/* Set main message chain element pointer */
3140 	main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge;
3141 
3142 	/*
3143 	 * For NVMe the chain element needs to be the 2nd SGL entry in the main
3144 	 * message.
3145 	 */
3146 	main_chain_element = (Mpi25IeeeSgeChain64_t *)
3147 	    ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64));
3148 
3149 	/*
3150 	 * For the PRP entries, use the specially allocated buffer of
3151 	 * contiguous memory. PRP Page allocation failures should not happen
3152 	 * because there should be enough PRP page buffers to account for the
3153 	 * possible NVMe QDepth.
3154 	 */
3155 	prp_page_info = mpr_alloc_prp_page(sc);
3156 	KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be "
3157 	    "used for building a native NVMe SGL.\n", __func__));
3158 	curr_buff = (uint32_t *)prp_page_info->prp_page;
3159 	msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr;
3160 
3161 	/*
3162 	 * Insert the allocated PRP page into the command's PRP page list. This
3163 	 * will be freed when the command is freed.
3164 	 */
3165 	TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link);
3166 
3167 	/*
3168 	 * Check if we are within 1 entry of a page boundary we don't want our
3169 	 * first entry to be a PRP List entry.
3170 	 */
3171 	page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) &
3172 	    page_mask;
3173 	if (!page_mask_result) {
3174 		/* Bump up to next page boundary. */
3175 		curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size);
3176 		msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size);
3177 	}
3178 
3179 	/* Fill in the chain element and make it an NVMe segment type. */
3180 	main_chain_element->Address.High =
3181 	    htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32));
3182 	main_chain_element->Address.Low =
3183 	    htole32((uint32_t)(uintptr_t)msg_phys);
3184 	main_chain_element->NextChainOffset = 0;
3185 	main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3186 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3187 	    MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP;
3188 
3189 	/* Set SGL pointer to start of contiguous PCIe buffer. */
3190 	ptr_sgl = curr_buff;
3191 	sge_dwords = 2;
3192 	num_entries = 0;
3193 
3194 	/*
3195 	 * NVMe has a very convoluted PRP format. One PRP is required for each
3196 	 * page or partial page. We need to split up OS SG entries if they are
3197 	 * longer than one page or cross a page boundary. We also have to insert
3198 	 * a PRP list pointer entry as the last entry in each physical page of
3199 	 * the PRP list.
3200 	 *
3201 	 * NOTE: The first PRP "entry" is actually placed in the first SGL entry
3202 	 * in the main message in IEEE 64 format. The 2nd entry in the main
3203 	 * message is the chain element, and the rest of the PRP entries are
3204 	 * built in the contiguous PCIe buffer.
3205 	 */
3206 	first_prp_entry = 1;
3207 	ptr_first_sgl = (uint32_t *)cm->cm_sge;
3208 
3209 	for (i = 0; i < segs_left; i++) {
3210 		/* Get physical address and length of this SG entry. */
3211 		paddr = segs[i].ds_addr;
3212 		length = segs[i].ds_len;
3213 
3214 		/*
3215 		 * Check whether a given SGE buffer lies on a non-PAGED
3216 		 * boundary if this is not the first page. If so, this is not
3217 		 * expected so have FW build the SGL.
3218 		 */
3219 		if ((i != 0) && (((uint32_t)paddr & page_mask) != 0)) {
3220 			mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while "
3221 			    "building NVMe PRPs, low address is 0x%x\n",
3222 			    (uint32_t)paddr);
3223 			return 1;
3224 		}
3225 
3226 		/* Apart from last SGE, if any other SGE boundary is not page
3227 		 * aligned then it means that hole exists. Existence of hole
3228 		 * leads to data corruption. So fallback to IEEE SGEs.
3229 		 */
3230 		if (i != (segs_left - 1)) {
3231 			if (((uint32_t)paddr + length) & page_mask) {
3232 				mpr_dprint(sc, MPR_ERROR, "Unaligned SGE "
3233 				    "boundary while building NVMe PRPs, low "
3234 				    "address: 0x%x and length: %u\n",
3235 				    (uint32_t)paddr, length);
3236 				return 1;
3237 			}
3238 		}
3239 
3240 		/* Loop while the length is not zero. */
3241 		while (length) {
3242 			/*
3243 			 * Check if we need to put a list pointer here if we are
3244 			 * at page boundary - prp_size.
3245 			 */
3246 			page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl +
3247 			    prp_size) & page_mask;
3248 			if (!page_mask_result) {
3249 				/*
3250 				 * Need to put a PRP list pointer here.
3251 				 */
3252 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3253 				    prp_size);
3254 				*ptr_sgl = htole32((uintptr_t)msg_phys);
3255 				*(ptr_sgl+1) = htole32((uint64_t)(uintptr_t)
3256 				    msg_phys >> 32);
3257 				ptr_sgl += sge_dwords;
3258 				num_entries++;
3259 			}
3260 
3261 			/* Need to handle if entry will be part of a page. */
3262 			offset = (uint32_t)paddr & page_mask;
3263 			entry_len = PAGE_SIZE - offset;
3264 			if (first_prp_entry) {
3265 				/*
3266 				 * Put IEEE entry in first SGE in main message.
3267 				 * (Simple element, System addr, not end of
3268 				 * list.)
3269 				 */
3270 				*ptr_first_sgl = htole32((uint32_t)paddr);
3271 				*(ptr_first_sgl + 1) =
3272 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3273 				*(ptr_first_sgl + 2) = htole32(entry_len);
3274 				*(ptr_first_sgl + 3) = 0;
3275 
3276 				/* No longer the first PRP entry. */
3277 				first_prp_entry = 0;
3278 			} else {
3279 				/* Put entry in list. */
3280 				*ptr_sgl = htole32((uint32_t)paddr);
3281 				*(ptr_sgl + 1) =
3282 				    htole32((uint32_t)((uint64_t)paddr >> 32));
3283 
3284 				/* Bump ptr_sgl, msg_phys, and num_entries. */
3285 				ptr_sgl += sge_dwords;
3286 				msg_phys = (uint64_t *)((uint8_t *)msg_phys +
3287 				    prp_size);
3288 				num_entries++;
3289 			}
3290 
3291 			/* Bump the phys address by the entry_len. */
3292 			paddr += entry_len;
3293 
3294 			/* Decrement length accounting for last partial page. */
3295 			if (entry_len > length)
3296 				length = 0;
3297 			else
3298 				length -= entry_len;
3299 		}
3300 	}
3301 
3302 	/* Set chain element Length. */
3303 	main_chain_element->Length = htole32(num_entries * prp_size);
3304 
3305 	/* Return 0, indicating we built a native SGL. */
3306 	return 0;
3307 }
3308 
3309 /*
3310  * Add a chain element as the next SGE for the specified command.
3311  * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are
3312  * only required for IEEE commands.  Therefore there is no code for commands
3313  * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands
3314  * shouldn't be requesting chains).
3315  */
3316 static int
3317 mpr_add_chain(struct mpr_command *cm, int segsleft)
3318 {
3319 	struct mpr_softc *sc = cm->cm_sc;
3320 	MPI2_REQUEST_HEADER *req;
3321 	MPI25_IEEE_SGE_CHAIN64 *ieee_sgc;
3322 	struct mpr_chain *chain;
3323 	int sgc_size, current_segs, rem_segs, segs_per_frame;
3324 	uint8_t next_chain_offset = 0;
3325 
3326 	/*
3327 	 * Fail if a command is requesting a chain for SIMPLE SGE's.  For SAS3
3328 	 * only IEEE commands should be requesting chains.  Return some error
3329 	 * code other than 0.
3330 	 */
3331 	if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) {
3332 		mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to "
3333 		    "an MPI SGL.\n");
3334 		return(ENOBUFS);
3335 	}
3336 
3337 	sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64);
3338 	if (cm->cm_sglsize < sgc_size)
3339 		panic("MPR: Need SGE Error Code\n");
3340 
3341 	chain = mpr_alloc_chain(cm->cm_sc);
3342 	if (chain == NULL)
3343 		return (ENOBUFS);
3344 
3345 	/*
3346 	 * Note: a double-linked list is used to make it easier to walk for
3347 	 * debugging.
3348 	 */
3349 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
3350 
3351 	/*
3352 	 * Need to know if the number of frames left is more than 1 or not.  If
3353 	 * more than 1 frame is required, NextChainOffset will need to be set,
3354 	 * which will just be the last segment of the frame.
3355 	 */
3356 	rem_segs = 0;
3357 	if (cm->cm_sglsize < (sgc_size * segsleft)) {
3358 		/*
3359 		 * rem_segs is the number of segements remaining after the
3360 		 * segments that will go into the current frame.  Since it is
3361 		 * known that at least one more frame is required, account for
3362 		 * the chain element.  To know if more than one more frame is
3363 		 * required, just check if there will be a remainder after using
3364 		 * the current frame (with this chain) and the next frame.  If
3365 		 * so the NextChainOffset must be the last element of the next
3366 		 * frame.
3367 		 */
3368 		current_segs = (cm->cm_sglsize / sgc_size) - 1;
3369 		rem_segs = segsleft - current_segs;
3370 		segs_per_frame = sc->chain_frame_size / sgc_size;
3371 		if (rem_segs > segs_per_frame) {
3372 			next_chain_offset = segs_per_frame - 1;
3373 		}
3374 	}
3375 	ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain;
3376 	ieee_sgc->Length = next_chain_offset ?
3377 	    htole32((uint32_t)sc->chain_frame_size) :
3378 	    htole32((uint32_t)rem_segs * (uint32_t)sgc_size);
3379 	ieee_sgc->NextChainOffset = next_chain_offset;
3380 	ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT |
3381 	    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3382 	ieee_sgc->Address.Low = htole32(chain->chain_busaddr);
3383 	ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32);
3384 	cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple;
3385 	req = (MPI2_REQUEST_HEADER *)cm->cm_req;
3386 	req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4;
3387 
3388 	cm->cm_sglsize = sc->chain_frame_size;
3389 	return (0);
3390 }
3391 
3392 /*
3393  * Add one scatter-gather element to the scatter-gather list for a command.
3394  * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the
3395  * next SGE to fill in, respectively.  In Gen3, the MPI SGL does not have a
3396  * chain, so don't consider any chain additions.
3397  */
3398 int
3399 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len,
3400     int segsleft)
3401 {
3402 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3403 	u32 sge_flags;
3404 
3405 	/*
3406 	 * case 1: >=1 more segment, no room for anything (error)
3407 	 * case 2: 1 more segment and enough room for it
3408          */
3409 
3410 	if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) {
3411 		mpr_dprint(cm->cm_sc, MPR_ERROR,
3412 		    "%s: warning: Not enough room for MPI SGL in frame.\n",
3413 		    __func__);
3414 		return(ENOBUFS);
3415 	}
3416 
3417 	KASSERT(segsleft == 1,
3418 	    ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n",
3419 	    segsleft));
3420 
3421 	/*
3422 	 * There is one more segment left to add for the MPI SGL and there is
3423 	 * enough room in the frame to add it.  This is the normal case because
3424 	 * MPI SGL's don't have chains, otherwise something is wrong.
3425 	 *
3426 	 * If this is a bi-directional request, need to account for that
3427 	 * here.  Save the pre-filled sge values.  These will be used
3428 	 * either for the 2nd SGL or for a single direction SGL.  If
3429 	 * cm_out_len is non-zero, this is a bi-directional request, so
3430 	 * fill in the OUT SGL first, then the IN SGL, otherwise just
3431 	 * fill in the IN SGL.  Note that at this time, when filling in
3432 	 * 2 SGL's for a bi-directional request, they both use the same
3433 	 * DMA buffer (same cm command).
3434 	 */
3435 	saved_buf_len = sge->FlagsLength & 0x00FFFFFF;
3436 	saved_address_low = sge->Address.Low;
3437 	saved_address_high = sge->Address.High;
3438 	if (cm->cm_out_len) {
3439 		sge->FlagsLength = cm->cm_out_len |
3440 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3441 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
3442 		    MPI2_SGE_FLAGS_HOST_TO_IOC |
3443 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3444 		    MPI2_SGE_FLAGS_SHIFT);
3445 		cm->cm_sglsize -= len;
3446 		/* Endian Safe code */
3447 		sge_flags = sge->FlagsLength;
3448 		sge->FlagsLength = htole32(sge_flags);
3449 		sge->Address.High = htole32(sge->Address.High);
3450 		sge->Address.Low = htole32(sge->Address.Low);
3451 		bcopy(sge, cm->cm_sge, len);
3452 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3453 	}
3454 	sge->FlagsLength = saved_buf_len |
3455 	    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3456 	    MPI2_SGE_FLAGS_END_OF_BUFFER |
3457 	    MPI2_SGE_FLAGS_LAST_ELEMENT |
3458 	    MPI2_SGE_FLAGS_END_OF_LIST |
3459 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
3460 	    MPI2_SGE_FLAGS_SHIFT);
3461 	if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) {
3462 		sge->FlagsLength |=
3463 		    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
3464 		    MPI2_SGE_FLAGS_SHIFT);
3465 	} else {
3466 		sge->FlagsLength |=
3467 		    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
3468 		    MPI2_SGE_FLAGS_SHIFT);
3469 	}
3470 	sge->Address.Low = saved_address_low;
3471 	sge->Address.High = saved_address_high;
3472 
3473 	cm->cm_sglsize -= len;
3474 	/* Endian Safe code */
3475 	sge_flags = sge->FlagsLength;
3476 	sge->FlagsLength = htole32(sge_flags);
3477 	sge->Address.High = htole32(sge->Address.High);
3478 	sge->Address.Low = htole32(sge->Address.Low);
3479 	bcopy(sge, cm->cm_sge, len);
3480 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
3481 	return (0);
3482 }
3483 
3484 /*
3485  * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter-
3486  * gather list for a command.  Maintain cm_sglsize and cm_sge as the
3487  * remaining size and pointer to the next SGE to fill in, respectively.
3488  */
3489 int
3490 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft)
3491 {
3492 	MPI2_IEEE_SGE_SIMPLE64 *sge = sgep;
3493 	int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION);
3494 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
3495 	uint32_t sge_length;
3496 
3497 	/*
3498 	 * case 1: No room for chain or segment (error).
3499 	 * case 2: Two or more segments left but only room for chain.
3500 	 * case 3: Last segment and room for it, so set flags.
3501 	 */
3502 
3503 	/*
3504 	 * There should be room for at least one element, or there is a big
3505 	 * problem.
3506 	 */
3507 	if (cm->cm_sglsize < ieee_sge_size)
3508 		panic("MPR: Need SGE Error Code\n");
3509 
3510 	if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) {
3511 		if ((error = mpr_add_chain(cm, segsleft)) != 0)
3512 			return (error);
3513 	}
3514 
3515 	if (segsleft == 1) {
3516 		/*
3517 		 * If this is a bi-directional request, need to account for that
3518 		 * here.  Save the pre-filled sge values.  These will be used
3519 		 * either for the 2nd SGL or for a single direction SGL.  If
3520 		 * cm_out_len is non-zero, this is a bi-directional request, so
3521 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
3522 		 * fill in the IN SGL.  Note that at this time, when filling in
3523 		 * 2 SGL's for a bi-directional request, they both use the same
3524 		 * DMA buffer (same cm command).
3525 		 */
3526 		saved_buf_len = sge->Length;
3527 		saved_address_low = sge->Address.Low;
3528 		saved_address_high = sge->Address.High;
3529 		if (cm->cm_out_len) {
3530 			sge->Length = cm->cm_out_len;
3531 			sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3532 			    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3533 			cm->cm_sglsize -= ieee_sge_size;
3534 			/* Endian Safe code */
3535 			sge_length = sge->Length;
3536 			sge->Length = htole32(sge_length);
3537 			sge->Address.High = htole32(sge->Address.High);
3538 			sge->Address.Low = htole32(sge->Address.Low);
3539 			bcopy(sgep, cm->cm_sge, ieee_sge_size);
3540 			cm->cm_sge =
3541 			    (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3542 			    ieee_sge_size);
3543 		}
3544 		sge->Length = saved_buf_len;
3545 		sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3546 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR |
3547 		    MPI25_IEEE_SGE_FLAGS_END_OF_LIST);
3548 		sge->Address.Low = saved_address_low;
3549 		sge->Address.High = saved_address_high;
3550 	}
3551 
3552 	cm->cm_sglsize -= ieee_sge_size;
3553 	/* Endian Safe code */
3554 	sge_length = sge->Length;
3555 	sge->Length = htole32(sge_length);
3556 	sge->Address.High = htole32(sge->Address.High);
3557 	sge->Address.Low = htole32(sge->Address.Low);
3558 	bcopy(sgep, cm->cm_sge, ieee_sge_size);
3559 	cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge +
3560 	    ieee_sge_size);
3561 	return (0);
3562 }
3563 
3564 /*
3565  * Add one dma segment to the scatter-gather list for a command.
3566  */
3567 int
3568 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags,
3569     int segsleft)
3570 {
3571 	MPI2_SGE_SIMPLE64 sge;
3572 	MPI2_IEEE_SGE_SIMPLE64 ieee_sge;
3573 
3574 	if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) {
3575 		ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT |
3576 		    MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR);
3577 		ieee_sge.Length = len;
3578 		mpr_from_u64(pa, &ieee_sge.Address);
3579 
3580 		return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft));
3581 	} else {
3582 		/*
3583 		 * This driver always uses 64-bit address elements for
3584 		 * simplicity.
3585 		 */
3586 		flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
3587 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
3588 		/* Set Endian safe macro in mpr_push_sge */
3589 		sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT);
3590 		mpr_from_u64(pa, &sge.Address);
3591 
3592 		return (mpr_push_sge(cm, &sge, sizeof sge, segsleft));
3593 	}
3594 }
3595 
3596 static void
3597 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
3598 {
3599 	struct mpr_softc *sc;
3600 	struct mpr_command *cm;
3601 	u_int i, dir, sflags;
3602 
3603 	cm = (struct mpr_command *)arg;
3604 	sc = cm->cm_sc;
3605 
3606 	/*
3607 	 * In this case, just print out a warning and let the chip tell the
3608 	 * user they did the wrong thing.
3609 	 */
3610 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
3611 		mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d "
3612 		    "segments, more than the %d allowed\n", __func__, nsegs,
3613 		    cm->cm_max_segs);
3614 	}
3615 
3616 	/*
3617 	 * Set up DMA direction flags.  Bi-directional requests are also handled
3618 	 * here.  In that case, both direction flags will be set.
3619 	 */
3620 	sflags = 0;
3621 	if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) {
3622 		/*
3623 		 * We have to add a special case for SMP passthrough, there
3624 		 * is no easy way to generically handle it.  The first
3625 		 * S/G element is used for the command (therefore the
3626 		 * direction bit needs to be set).  The second one is used
3627 		 * for the reply.  We'll leave it to the caller to make
3628 		 * sure we only have two buffers.
3629 		 */
3630 		/*
3631 		 * Even though the busdma man page says it doesn't make
3632 		 * sense to have both direction flags, it does in this case.
3633 		 * We have one s/g element being accessed in each direction.
3634 		 */
3635 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
3636 
3637 		/*
3638 		 * Set the direction flag on the first buffer in the SMP
3639 		 * passthrough request.  We'll clear it for the second one.
3640 		 */
3641 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
3642 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
3643 	} else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) {
3644 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
3645 		dir = BUS_DMASYNC_PREWRITE;
3646 	} else
3647 		dir = BUS_DMASYNC_PREREAD;
3648 
3649 	/* Check if a native SG list is needed for an NVMe PCIe device. */
3650 	if (cm->cm_targ && cm->cm_targ->is_nvme &&
3651 	    mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) {
3652 		/* A native SG list was built, skip to end. */
3653 		goto out;
3654 	}
3655 
3656 	for (i = 0; i < nsegs; i++) {
3657 		if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) {
3658 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
3659 		}
3660 		error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
3661 		    sflags, nsegs - i);
3662 		if (error != 0) {
3663 			/* Resource shortage, roll back! */
3664 			if (ratecheck(&sc->lastfail, &mpr_chainfail_interval))
3665 				mpr_dprint(sc, MPR_INFO, "Out of chain frames, "
3666 				    "consider increasing hw.mpr.max_chains.\n");
3667 			cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED;
3668 			mpr_complete_command(sc, cm);
3669 			return;
3670 		}
3671 	}
3672 
3673 out:
3674 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
3675 	mpr_enqueue_request(sc, cm);
3676 
3677 	return;
3678 }
3679 
3680 static void
3681 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
3682 	     int error)
3683 {
3684 	mpr_data_cb(arg, segs, nsegs, error);
3685 }
3686 
3687 /*
3688  * This is the routine to enqueue commands ansynchronously.
3689  * Note that the only error path here is from bus_dmamap_load(), which can
3690  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
3691  * assumed that if you have a command in-hand, then you have enough credits
3692  * to use it.
3693  */
3694 int
3695 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm)
3696 {
3697 	int error = 0;
3698 
3699 	if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) {
3700 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
3701 		    &cm->cm_uio, mpr_data_cb2, cm, 0);
3702 	} else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) {
3703 		error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap,
3704 		    cm->cm_data, mpr_data_cb, cm, 0);
3705 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
3706 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
3707 		    cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0);
3708 	} else {
3709 		/* Add a zero-length element as needed */
3710 		if (cm->cm_sge != NULL)
3711 			mpr_add_dmaseg(cm, 0, 0, 0, 1);
3712 		mpr_enqueue_request(sc, cm);
3713 	}
3714 
3715 	return (error);
3716 }
3717 
3718 /*
3719  * This is the routine to enqueue commands synchronously.  An error of
3720  * EINPROGRESS from mpr_map_command() is ignored since the command will
3721  * be executed and enqueued automatically.  Other errors come from msleep().
3722  */
3723 int
3724 mpr_wait_command(struct mpr_softc *sc, struct mpr_command **cmp, int timeout,
3725     int sleep_flag)
3726 {
3727 	int error, rc;
3728 	struct timeval cur_time, start_time;
3729 	struct mpr_command *cm = *cmp;
3730 
3731 	if (sc->mpr_flags & MPR_FLAGS_DIAGRESET)
3732 		return  EBUSY;
3733 
3734 	cm->cm_complete = NULL;
3735 	cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED);
3736 	error = mpr_map_command(sc, cm);
3737 	if ((error != 0) && (error != EINPROGRESS))
3738 		return (error);
3739 
3740 	// Check for context and wait for 50 mSec at a time until time has
3741 	// expired or the command has finished.  If msleep can't be used, need
3742 	// to poll.
3743 	if (curthread->td_no_sleeping)
3744 		sleep_flag = NO_SLEEP;
3745 	getmicrouptime(&start_time);
3746 	if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) {
3747 		error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz);
3748 		if (error == EWOULDBLOCK) {
3749 			/*
3750 			 * Record the actual elapsed time in the case of a
3751 			 * timeout for the message below.
3752 			 */
3753 			getmicrouptime(&cur_time);
3754 			timevalsub(&cur_time, &start_time);
3755 		}
3756 	} else {
3757 		while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3758 			mpr_intr_locked(sc);
3759 			if (sleep_flag == CAN_SLEEP)
3760 				pause("mprwait", hz/20);
3761 			else
3762 				DELAY(50000);
3763 
3764 			getmicrouptime(&cur_time);
3765 			timevalsub(&cur_time, &start_time);
3766 			if (cur_time.tv_sec > timeout) {
3767 				error = EWOULDBLOCK;
3768 				break;
3769 			}
3770 		}
3771 	}
3772 
3773 	if (error == EWOULDBLOCK) {
3774 		if (cm->cm_timeout_handler == NULL) {
3775 			mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s, timeout=%d,"
3776 			    " elapsed=%jd\n", __func__, timeout,
3777 			    (intmax_t)cur_time.tv_sec);
3778 			rc = mpr_reinit(sc);
3779 			mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3780 			    "failed");
3781 		} else
3782 			cm->cm_timeout_handler(sc, cm);
3783 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3784 			/*
3785 			 * Tell the caller that we freed the command in a
3786 			 * reinit.
3787 			 */
3788 			*cmp = NULL;
3789 		}
3790 		error = ETIMEDOUT;
3791 	}
3792 	return (error);
3793 }
3794 
3795 /*
3796  * This is the routine to enqueue a command synchonously and poll for
3797  * completion.  Its use should be rare.
3798  */
3799 int
3800 mpr_request_polled(struct mpr_softc *sc, struct mpr_command **cmp)
3801 {
3802 	int error, rc;
3803 	struct timeval cur_time, start_time;
3804 	struct mpr_command *cm = *cmp;
3805 
3806 	error = 0;
3807 
3808 	cm->cm_flags |= MPR_CM_FLAGS_POLLED;
3809 	cm->cm_complete = NULL;
3810 	mpr_map_command(sc, cm);
3811 
3812 	getmicrouptime(&start_time);
3813 	while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) {
3814 		mpr_intr_locked(sc);
3815 
3816 		if (mtx_owned(&sc->mpr_mtx))
3817 			msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0,
3818 			    "mprpoll", hz/20);
3819 		else
3820 			pause("mprpoll", hz/20);
3821 
3822 		/*
3823 		 * Check for real-time timeout and fail if more than 60 seconds.
3824 		 */
3825 		getmicrouptime(&cur_time);
3826 		timevalsub(&cur_time, &start_time);
3827 		if (cur_time.tv_sec > 60) {
3828 			mpr_dprint(sc, MPR_FAULT, "polling failed\n");
3829 			error = ETIMEDOUT;
3830 			break;
3831 		}
3832 	}
3833 	cm->cm_state = MPR_CM_STATE_BUSY;
3834 	if (error) {
3835 		mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__);
3836 		rc = mpr_reinit(sc);
3837 		mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" :
3838 		    "failed");
3839 
3840 		if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) {
3841 			/*
3842 			 * Tell the caller that we freed the command in a
3843 			 * reinit.
3844 			 */
3845 			*cmp = NULL;
3846 		}
3847 	}
3848 	return (error);
3849 }
3850 
3851 /*
3852  * The MPT driver had a verbose interface for config pages.  In this driver,
3853  * reduce it to much simpler terms, similar to the Linux driver.
3854  */
3855 int
3856 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3857 {
3858 	MPI2_CONFIG_REQUEST *req;
3859 	struct mpr_command *cm;
3860 	int error;
3861 
3862 	if (sc->mpr_flags & MPR_FLAGS_BUSY) {
3863 		return (EBUSY);
3864 	}
3865 
3866 	cm = mpr_alloc_command(sc);
3867 	if (cm == NULL) {
3868 		return (EBUSY);
3869 	}
3870 
3871 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
3872 	req->Function = MPI2_FUNCTION_CONFIG;
3873 	req->Action = params->action;
3874 	req->SGLFlags = 0;
3875 	req->ChainOffset = 0;
3876 	req->PageAddress = params->page_address;
3877 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3878 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
3879 
3880 		hdr = &params->hdr.Ext;
3881 		req->ExtPageType = hdr->ExtPageType;
3882 		req->ExtPageLength = hdr->ExtPageLength;
3883 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
3884 		req->Header.PageLength = 0; /* Must be set to zero */
3885 		req->Header.PageNumber = hdr->PageNumber;
3886 		req->Header.PageVersion = hdr->PageVersion;
3887 	} else {
3888 		MPI2_CONFIG_PAGE_HEADER *hdr;
3889 
3890 		hdr = &params->hdr.Struct;
3891 		req->Header.PageType = hdr->PageType;
3892 		req->Header.PageNumber = hdr->PageNumber;
3893 		req->Header.PageLength = hdr->PageLength;
3894 		req->Header.PageVersion = hdr->PageVersion;
3895 	}
3896 
3897 	cm->cm_data = params->buffer;
3898 	cm->cm_length = params->length;
3899 	if (cm->cm_data != NULL) {
3900 		cm->cm_sge = &req->PageBufferSGE;
3901 		cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
3902 		cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN;
3903 	} else
3904 		cm->cm_sge = NULL;
3905 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
3906 
3907 	cm->cm_complete_data = params;
3908 	if (params->callback != NULL) {
3909 		cm->cm_complete = mpr_config_complete;
3910 		return (mpr_map_command(sc, cm));
3911 	} else {
3912 		error = mpr_wait_command(sc, &cm, 0, CAN_SLEEP);
3913 		if (error) {
3914 			mpr_dprint(sc, MPR_FAULT,
3915 			    "Error %d reading config page\n", error);
3916 			if (cm != NULL)
3917 				mpr_free_command(sc, cm);
3918 			return (error);
3919 		}
3920 		mpr_config_complete(sc, cm);
3921 	}
3922 
3923 	return (0);
3924 }
3925 
3926 int
3927 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params)
3928 {
3929 	return (EINVAL);
3930 }
3931 
3932 static void
3933 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm)
3934 {
3935 	MPI2_CONFIG_REPLY *reply;
3936 	struct mpr_config_params *params;
3937 
3938 	MPR_FUNCTRACE(sc);
3939 	params = cm->cm_complete_data;
3940 
3941 	if (cm->cm_data != NULL) {
3942 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
3943 		    BUS_DMASYNC_POSTREAD);
3944 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
3945 	}
3946 
3947 	/*
3948 	 * XXX KDM need to do more error recovery?  This results in the
3949 	 * device in question not getting probed.
3950 	 */
3951 	if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) {
3952 		params->status = MPI2_IOCSTATUS_BUSY;
3953 		goto done;
3954 	}
3955 
3956 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
3957 	if (reply == NULL) {
3958 		params->status = MPI2_IOCSTATUS_BUSY;
3959 		goto done;
3960 	}
3961 	params->status = reply->IOCStatus;
3962 	if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) {
3963 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
3964 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
3965 		params->hdr.Ext.PageType = reply->Header.PageType;
3966 		params->hdr.Ext.PageNumber = reply->Header.PageNumber;
3967 		params->hdr.Ext.PageVersion = reply->Header.PageVersion;
3968 	} else {
3969 		params->hdr.Struct.PageType = reply->Header.PageType;
3970 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
3971 		params->hdr.Struct.PageLength = reply->Header.PageLength;
3972 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
3973 	}
3974 
3975 done:
3976 	mpr_free_command(sc, cm);
3977 	if (params->callback != NULL)
3978 		params->callback(sc, params);
3979 
3980 	return;
3981 }
3982