xref: /freebsd/sys/dev/mps/mps.c (revision f5f47d5068fb97df18eb114a66ae8ef51a0b3c8c)
1 /*-
2  * Copyright (c) 2009 Yahoo! Inc.
3  * Copyright (c) 2012 LSI Corp.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * LSI MPT-Fusion Host Adapter FreeBSD
28  *
29  * $FreeBSD$
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 /* Communications core for LSI MPT2 */
36 
37 /* TODO Move headers to mpsvar */
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/selinfo.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/bus.h>
47 #include <sys/conf.h>
48 #include <sys/bio.h>
49 #include <sys/malloc.h>
50 #include <sys/uio.h>
51 #include <sys/sysctl.h>
52 #include <sys/queue.h>
53 #include <sys/kthread.h>
54 #include <sys/endian.h>
55 #include <sys/eventhandler.h>
56 
57 #include <machine/bus.h>
58 #include <machine/resource.h>
59 #include <sys/rman.h>
60 #include <sys/proc.h>
61 
62 #include <dev/pci/pcivar.h>
63 
64 #include <cam/scsi/scsi_all.h>
65 
66 #include <dev/mps/mpi/mpi2_type.h>
67 #include <dev/mps/mpi/mpi2.h>
68 #include <dev/mps/mpi/mpi2_ioc.h>
69 #include <dev/mps/mpi/mpi2_sas.h>
70 #include <dev/mps/mpi/mpi2_cnfg.h>
71 #include <dev/mps/mpi/mpi2_init.h>
72 #include <dev/mps/mpi/mpi2_tool.h>
73 #include <dev/mps/mps_ioctl.h>
74 #include <dev/mps/mpsvar.h>
75 #include <dev/mps/mps_table.h>
76 
77 static int mps_diag_reset(struct mps_softc *sc, int sleep_flag);
78 static int mps_init_queues(struct mps_softc *sc);
79 static int mps_message_unit_reset(struct mps_softc *sc, int sleep_flag);
80 static int mps_transition_operational(struct mps_softc *sc);
81 static void mps_startup(void *arg);
82 static int mps_send_iocinit(struct mps_softc *sc);
83 static int mps_attach_log(struct mps_softc *sc);
84 static __inline void mps_complete_command(struct mps_command *cm);
85 static void mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
86     MPI2_EVENT_NOTIFICATION_REPLY *reply);
87 static void mps_config_complete(struct mps_softc *sc, struct mps_command *cm);
88 static void mps_periodic(void *);
89 static int mps_reregister_events(struct mps_softc *sc);
90 static void mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm);
91 static int mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag);
92 SYSCTL_NODE(_hw, OID_AUTO, mps, CTLFLAG_RD, 0, "MPS Driver Parameters");
93 
94 MALLOC_DEFINE(M_MPT2, "mps", "mpt2 driver memory");
95 
96 /*
97  * Do a "Diagnostic Reset" aka a hard reset.  This should get the chip out of
98  * any state and back to its initialization state machine.
99  */
100 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d };
101 
102 /* Added this union to smoothly convert le64toh cm->cm_desc.Words.
103  * Compiler only support unint64_t to be passed as argument.
104  * Otherwise it will through below error
105  * "aggregate value used where an integer was expected"
106  */
107 
108 typedef union _reply_descriptor {
109         u64 word;
110         struct {
111                 u32 low;
112                 u32 high;
113         } u;
114 }reply_descriptor,address_descriptor;
115 
116 /*
117  * sleep_flag can be either CAN_SLEEP or NO_SLEEP.
118  * If this function is called from process context, it can sleep
119  * and there is no harm to sleep, in case if this fuction is called
120  * from Interrupt handler, we can not sleep and need NO_SLEEP flag set.
121  * based on sleep flags driver will call either msleep, pause or DELAY.
122  * msleep and pause are of same variant, but pause is used when mps_mtx
123  * is not hold by driver.
124  *
125  */
126 static int
127 mps_diag_reset(struct mps_softc *sc,int sleep_flag)
128 {
129 	uint32_t reg;
130 	int i, error, tries = 0;
131 
132 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
133 
134 	/* Clear any pending interrupts */
135 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
136 
137 	/*Force NO_SLEEP for threads prohibited to sleep
138  	* e.a Thread from interrupt handler are prohibited to sleep.
139  	*/
140 	if(curthread->td_pflags & TDP_NOSLEEPING)
141 		sleep_flag = NO_SLEEP;
142 
143 	/* Push the magic sequence */
144 	error = ETIMEDOUT;
145 	while (tries++ < 20) {
146 		for (i = 0; i < sizeof(mpt2_reset_magic); i++)
147 			mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET,
148 			    mpt2_reset_magic[i]);
149 		/* wait 100 msec */
150 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
151 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0, "mpsdiag", hz/10);
152 		else if (sleep_flag == CAN_SLEEP)
153 			pause("mpsdiag", hz/10);
154 		else
155 			DELAY(100 * 1000);
156 
157 		reg = mps_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET);
158 		if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) {
159 			error = 0;
160 			break;
161 		}
162 	}
163 	if (error)
164 		return (error);
165 
166 	/* Send the actual reset.  XXX need to refresh the reg? */
167 	mps_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET,
168 	    reg | MPI2_DIAG_RESET_ADAPTER);
169 
170 	/* Wait up to 300 seconds in 50ms intervals */
171 	error = ETIMEDOUT;
172 	for (i = 0; i < 60000; i++) {
173 		/* wait 50 msec */
174 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
175 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0, "mpsdiag", hz/20);
176 		else if (sleep_flag == CAN_SLEEP)
177 			pause("mpsdiag", hz/20);
178 		else
179 			DELAY(50 * 1000);
180 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
181 		if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) {
182 			error = 0;
183 			break;
184 		}
185 	}
186 	if (error)
187 		return (error);
188 
189 	mps_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0);
190 
191 	return (0);
192 }
193 
194 static int
195 mps_message_unit_reset(struct mps_softc *sc, int sleep_flag)
196 {
197 
198 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
199 
200 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
201 	    MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET <<
202 	    MPI2_DOORBELL_FUNCTION_SHIFT);
203 
204 	if (mps_wait_db_ack(sc, 5, sleep_flag) != 0) {
205 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed : <%s>\n",
206 				__func__);
207 		return (ETIMEDOUT);
208 	}
209 
210 	return (0);
211 }
212 
213 static int
214 mps_transition_ready(struct mps_softc *sc)
215 {
216 	uint32_t reg, state;
217 	int error, tries = 0;
218 	int sleep_flags;
219 
220 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
221 	/* If we are in attach call, do not sleep */
222 	sleep_flags = (sc->mps_flags & MPS_FLAGS_ATTACH_DONE)
223 					? CAN_SLEEP:NO_SLEEP;
224 	error = 0;
225 	while (tries++ < 5) {
226 		reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
227 		mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
228 
229 		/*
230 		 * Ensure the IOC is ready to talk.  If it's not, try
231 		 * resetting it.
232 		 */
233 		if (reg & MPI2_DOORBELL_USED) {
234 			mps_diag_reset(sc, sleep_flags);
235 			DELAY(50000);
236 			continue;
237 		}
238 
239 		/* Is the adapter owned by another peer? */
240 		if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) ==
241 		    (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) {
242 			device_printf(sc->mps_dev, "IOC is under the control "
243 			    "of another peer host, aborting initialization.\n");
244 			return (ENXIO);
245 		}
246 
247 		state = reg & MPI2_IOC_STATE_MASK;
248 		if (state == MPI2_IOC_STATE_READY) {
249 			/* Ready to go! */
250 			error = 0;
251 			break;
252 		} else if (state == MPI2_IOC_STATE_FAULT) {
253 			mps_dprint(sc, MPS_INFO, "IOC in fault state 0x%x\n",
254 			    state & MPI2_DOORBELL_FAULT_CODE_MASK);
255 			mps_diag_reset(sc, sleep_flags);
256 		} else if (state == MPI2_IOC_STATE_OPERATIONAL) {
257 			/* Need to take ownership */
258 			mps_message_unit_reset(sc, sleep_flags);
259 		} else if (state == MPI2_IOC_STATE_RESET) {
260 			/* Wait a bit, IOC might be in transition */
261 			mps_dprint(sc, MPS_FAULT,
262 			    "IOC in unexpected reset state\n");
263 		} else {
264 			mps_dprint(sc, MPS_FAULT,
265 			    "IOC in unknown state 0x%x\n", state);
266 			error = EINVAL;
267 			break;
268 		}
269 
270 		/* Wait 50ms for things to settle down. */
271 		DELAY(50000);
272 	}
273 
274 	if (error)
275 		device_printf(sc->mps_dev, "Cannot transition IOC to ready\n");
276 
277 	return (error);
278 }
279 
280 static int
281 mps_transition_operational(struct mps_softc *sc)
282 {
283 	uint32_t reg, state;
284 	int error;
285 
286 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
287 
288 	error = 0;
289 	reg = mps_regread(sc, MPI2_DOORBELL_OFFSET);
290 	mps_dprint(sc, MPS_INFO, "Doorbell= 0x%x\n", reg);
291 
292 	state = reg & MPI2_IOC_STATE_MASK;
293 	if (state != MPI2_IOC_STATE_READY) {
294 		if ((error = mps_transition_ready(sc)) != 0) {
295 			mps_dprint(sc, MPS_FAULT,
296 			    "%s failed to transition ready\n", __func__);
297 			return (error);
298 		}
299 	}
300 
301 	error = mps_send_iocinit(sc);
302 	return (error);
303 }
304 
305 /*
306  * XXX Some of this should probably move to mps.c
307  *
308  * The terms diag reset and hard reset are used interchangeably in the MPI
309  * docs to mean resetting the controller chip.  In this code diag reset
310  * cleans everything up, and the hard reset function just sends the reset
311  * sequence to the chip.  This should probably be refactored so that every
312  * subsystem gets a reset notification of some sort, and can clean up
313  * appropriately.
314  */
315 int
316 mps_reinit(struct mps_softc *sc)
317 {
318 	int error;
319 	uint32_t db;
320 
321 	mps_printf(sc, "%s sc %p\n", __func__, sc);
322 
323 	mtx_assert(&sc->mps_mtx, MA_OWNED);
324 
325 	if (sc->mps_flags & MPS_FLAGS_DIAGRESET) {
326 		mps_printf(sc, "%s reset already in progress\n", __func__);
327 		return 0;
328 	}
329 
330 	/* make sure the completion callbacks can recognize they're getting
331 	 * a NULL cm_reply due to a reset.
332 	 */
333 	sc->mps_flags |= MPS_FLAGS_DIAGRESET;
334 
335 	mps_printf(sc, "%s mask interrupts\n", __func__);
336 	mps_mask_intr(sc);
337 
338 	error = mps_diag_reset(sc, CAN_SLEEP);
339 	if (error != 0) {
340 		panic("%s hard reset failed with error %d\n",
341 		    __func__, error);
342 	}
343 
344 	/* Restore the PCI state, including the MSI-X registers */
345 	mps_pci_restore(sc);
346 
347 	/* Give the I/O subsystem special priority to get itself prepared */
348 	mpssas_handle_reinit(sc);
349 
350 	/* reinitialize queues after the reset */
351 	bzero(sc->free_queue, sc->fqdepth * 4);
352 	mps_init_queues(sc);
353 
354 	/* get the chip out of the reset state */
355 	error = mps_transition_operational(sc);
356 	if (error != 0)
357 		panic("%s transition operational failed with error %d\n",
358 		    __func__, error);
359 
360 	/* Reinitialize the reply queue. This is delicate because this
361 	 * function is typically invoked by task mgmt completion callbacks,
362 	 * which are called by the interrupt thread.  We need to make sure
363 	 * the interrupt handler loop will exit when we return to it, and
364 	 * that it will recognize the indexes we've changed.
365 	 */
366 	sc->replypostindex = 0;
367 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
368 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
369 
370 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
371 	mps_printf(sc, "%s doorbell 0x%08x\n", __func__, db);
372 
373 	mps_printf(sc, "%s unmask interrupts post %u free %u\n", __func__,
374 	    sc->replypostindex, sc->replyfreeindex);
375 
376 	mps_unmask_intr(sc);
377 
378 	mps_printf(sc, "%s restarting post %u free %u\n", __func__,
379 	    sc->replypostindex, sc->replyfreeindex);
380 
381 	/* restart will reload the event masks clobbered by the reset, and
382 	 * then enable the port.
383 	 */
384 	mps_reregister_events(sc);
385 
386 	/* the end of discovery will release the simq, so we're done. */
387 	mps_printf(sc, "%s finished sc %p post %u free %u\n",
388 	    __func__, sc,
389 	    sc->replypostindex, sc->replyfreeindex);
390 
391 	sc->mps_flags &= ~MPS_FLAGS_DIAGRESET;
392 
393 	return 0;
394 }
395 
396 /* Wait for the chip to ACK a word that we've put into its FIFO
397  * Wait for <timeout> seconds. In single loop wait for busy loop
398  * for 500 microseconds.
399  * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds.
400  * */
401 static int
402 mps_wait_db_ack(struct mps_softc *sc, int timeout, int sleep_flag)
403 {
404 
405 	u32 cntdn, count;
406 	u32 int_status;
407 	u32 doorbell;
408 
409 	count = 0;
410 	cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout;
411 	do {
412 		int_status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
413 		if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) {
414 			mps_dprint(sc, MPS_INFO,
415 			"%s: successfull count(%d), timeout(%d)\n",
416 			__func__, count, timeout);
417 		return 0;
418 		} else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) {
419 			doorbell = mps_regread(sc, MPI2_DOORBELL_OFFSET);
420 			if ((doorbell & MPI2_IOC_STATE_MASK) ==
421 				MPI2_IOC_STATE_FAULT) {
422 				mps_dprint(sc, MPS_FAULT,
423 					"fault_state(0x%04x)!\n", doorbell);
424 				return (EFAULT);
425 			}
426 		} else if (int_status == 0xFFFFFFFF)
427 			goto out;
428 
429 		/* If it can sleep, sleep for 1 milisecond, else busy loop for
430 		* 0.5 milisecond */
431 		if (mtx_owned(&sc->mps_mtx) && sleep_flag == CAN_SLEEP)
432 			msleep(&sc->msleep_fake_chan, &sc->mps_mtx, 0,
433 			"mpsdba", hz/1000);
434 		else if (sleep_flag == CAN_SLEEP)
435 			pause("mpsdba", hz/1000);
436 		else
437 			DELAY(500);
438 		count++;
439 	} while (--cntdn);
440 
441 	out:
442 	mps_dprint(sc, MPS_FAULT, "%s: failed due to timeout count(%d), "
443 		"int_status(%x)!\n", __func__, count, int_status);
444 	return (ETIMEDOUT);
445 
446 }
447 
448 /* Wait for the chip to signal that the next word in its FIFO can be fetched */
449 static int
450 mps_wait_db_int(struct mps_softc *sc)
451 {
452 	int retry;
453 
454 	for (retry = 0; retry < MPS_DB_MAX_WAIT; retry++) {
455 		if ((mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) &
456 		    MPI2_HIS_IOC2SYS_DB_STATUS) != 0)
457 			return (0);
458 		DELAY(2000);
459 	}
460 	return (ETIMEDOUT);
461 }
462 
463 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */
464 static int
465 mps_request_sync(struct mps_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply,
466     int req_sz, int reply_sz, int timeout)
467 {
468 	uint32_t *data32;
469 	uint16_t *data16;
470 	int i, count, ioc_sz, residual;
471 	int sleep_flags = CAN_SLEEP;
472 
473 	if(curthread->td_pflags & TDP_NOSLEEPING)
474 		sleep_flags = NO_SLEEP;
475 
476 	/* Step 1 */
477 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
478 
479 	/* Step 2 */
480 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
481 		return (EBUSY);
482 
483 	/* Step 3
484 	 * Announce that a message is coming through the doorbell.  Messages
485 	 * are pushed at 32bit words, so round up if needed.
486 	 */
487 	count = (req_sz + 3) / 4;
488 	mps_regwrite(sc, MPI2_DOORBELL_OFFSET,
489 	    (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) |
490 	    (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT));
491 
492 	/* Step 4 */
493 	if (mps_wait_db_int(sc) ||
494 	    (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) {
495 		mps_dprint(sc, MPS_FAULT, "Doorbell failed to activate\n");
496 		return (ENXIO);
497 	}
498 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
499 	if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
500 		mps_dprint(sc, MPS_FAULT, "Doorbell handshake failed\n");
501 		return (ENXIO);
502 	}
503 
504 	/* Step 5 */
505 	/* Clock out the message data synchronously in 32-bit dwords*/
506 	data32 = (uint32_t *)req;
507 	for (i = 0; i < count; i++) {
508 		mps_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i]));
509 		if (mps_wait_db_ack(sc, 5, sleep_flags) != 0) {
510 			mps_dprint(sc, MPS_FAULT,
511 			    "Timeout while writing doorbell\n");
512 			return (ENXIO);
513 		}
514 	}
515 
516 	/* Step 6 */
517 	/* Clock in the reply in 16-bit words.  The total length of the
518 	 * message is always in the 4th byte, so clock out the first 2 words
519 	 * manually, then loop the rest.
520 	 */
521 	data16 = (uint16_t *)reply;
522 	if (mps_wait_db_int(sc) != 0) {
523 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 0\n");
524 		return (ENXIO);
525 	}
526 	data16[0] =
527 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
528 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
529 	if (mps_wait_db_int(sc) != 0) {
530 		mps_dprint(sc, MPS_FAULT, "Timeout reading doorbell 1\n");
531 		return (ENXIO);
532 	}
533 	data16[1] =
534 	    mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_DATA_MASK;
535 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
536 
537 	/* Number of 32bit words in the message */
538 	ioc_sz = reply->MsgLength;
539 
540 	/*
541 	 * Figure out how many 16bit words to clock in without overrunning.
542 	 * The precision loss with dividing reply_sz can safely be
543 	 * ignored because the messages can only be multiples of 32bits.
544 	 */
545 	residual = 0;
546 	count = MIN((reply_sz / 4), ioc_sz) * 2;
547 	if (count < ioc_sz * 2) {
548 		residual = ioc_sz * 2 - count;
549 		mps_dprint(sc, MPS_FAULT, "Driver error, throwing away %d "
550 		    "residual message words\n", residual);
551 	}
552 
553 	for (i = 2; i < count; i++) {
554 		if (mps_wait_db_int(sc) != 0) {
555 			mps_dprint(sc, MPS_FAULT,
556 			    "Timeout reading doorbell %d\n", i);
557 			return (ENXIO);
558 		}
559 		data16[i] = mps_regread(sc, MPI2_DOORBELL_OFFSET) &
560 		    MPI2_DOORBELL_DATA_MASK;
561 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
562 	}
563 
564 	/*
565 	 * Pull out residual words that won't fit into the provided buffer.
566 	 * This keeps the chip from hanging due to a driver programming
567 	 * error.
568 	 */
569 	while (residual--) {
570 		if (mps_wait_db_int(sc) != 0) {
571 			mps_dprint(sc, MPS_FAULT,
572 			    "Timeout reading doorbell\n");
573 			return (ENXIO);
574 		}
575 		(void)mps_regread(sc, MPI2_DOORBELL_OFFSET);
576 		mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
577 	}
578 
579 	/* Step 7 */
580 	if (mps_wait_db_int(sc) != 0) {
581 		mps_dprint(sc, MPS_FAULT, "Timeout waiting to exit doorbell\n");
582 		return (ENXIO);
583 	}
584 	if (mps_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED)
585 		mps_dprint(sc, MPS_FAULT, "Warning, doorbell still active\n");
586 	mps_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0);
587 
588 	return (0);
589 }
590 
591 static void
592 mps_enqueue_request(struct mps_softc *sc, struct mps_command *cm)
593 {
594 	reply_descriptor rd;
595 	mps_dprint(sc, MPS_TRACE, "%s SMID %u cm %p ccb %p\n", __func__,
596 	    cm->cm_desc.Default.SMID, cm, cm->cm_ccb);
597 
598 	if (sc->mps_flags & MPS_FLAGS_ATTACH_DONE && !(sc->mps_flags & MPS_FLAGS_SHUTDOWN))
599 		mtx_assert(&sc->mps_mtx, MA_OWNED);
600 
601 	if (++sc->io_cmds_active > sc->io_cmds_highwater)
602 		sc->io_cmds_highwater++;
603 	rd.u.low = cm->cm_desc.Words.Low;
604 	rd.u.high = cm->cm_desc.Words.High;
605 	rd.word = htole64(rd.word);
606 	/* TODO-We may need to make below regwrite atomic */
607 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET,
608 	    rd.u.low);
609 	mps_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET,
610 	    rd.u.high);
611 }
612 
613 /*
614  * Just the FACTS, ma'am.
615  */
616 static int
617 mps_get_iocfacts(struct mps_softc *sc, MPI2_IOC_FACTS_REPLY *facts)
618 {
619 	MPI2_DEFAULT_REPLY *reply;
620 	MPI2_IOC_FACTS_REQUEST request;
621 	int error, req_sz, reply_sz;
622 
623 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
624 
625 	req_sz = sizeof(MPI2_IOC_FACTS_REQUEST);
626 	reply_sz = sizeof(MPI2_IOC_FACTS_REPLY);
627 	reply = (MPI2_DEFAULT_REPLY *)facts;
628 
629 	bzero(&request, req_sz);
630 	request.Function = MPI2_FUNCTION_IOC_FACTS;
631 	error = mps_request_sync(sc, &request, reply, req_sz, reply_sz, 5);
632 
633 	return (error);
634 }
635 
636 static int
637 mps_get_portfacts(struct mps_softc *sc, MPI2_PORT_FACTS_REPLY *facts, int port)
638 {
639 	MPI2_PORT_FACTS_REQUEST *request;
640 	MPI2_PORT_FACTS_REPLY *reply;
641 	struct mps_command *cm;
642 	int error;
643 
644 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
645 
646 	if ((cm = mps_alloc_command(sc)) == NULL)
647 		return (EBUSY);
648 	request = (MPI2_PORT_FACTS_REQUEST *)cm->cm_req;
649 	request->Function = MPI2_FUNCTION_PORT_FACTS;
650 	request->PortNumber = port;
651 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
652 	cm->cm_data = NULL;
653 	error = mps_request_polled(sc, cm);
654 	reply = (MPI2_PORT_FACTS_REPLY *)cm->cm_reply;
655 	if (reply == NULL) {
656 		mps_printf(sc, "%s NULL reply\n", __func__);
657 		goto done;
658 	}
659 	if ((reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS) {
660 		mps_printf(sc,
661 		    "%s error %d iocstatus 0x%x iocloginfo 0x%x type 0x%x\n",
662 		    __func__, error, reply->IOCStatus, reply->IOCLogInfo,
663 		    reply->PortType);
664 		error = ENXIO;
665 	}
666 	bcopy(reply, facts, sizeof(MPI2_PORT_FACTS_REPLY));
667 done:
668 	mps_free_command(sc, cm);
669 
670 	return (error);
671 }
672 
673 static int
674 mps_send_iocinit(struct mps_softc *sc)
675 {
676 	MPI2_IOC_INIT_REQUEST	init;
677 	MPI2_DEFAULT_REPLY	reply;
678 	int req_sz, reply_sz, error;
679 
680 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
681 
682 	req_sz = sizeof(MPI2_IOC_INIT_REQUEST);
683 	reply_sz = sizeof(MPI2_IOC_INIT_REPLY);
684 	bzero(&init, req_sz);
685 	bzero(&reply, reply_sz);
686 
687 	/*
688 	 * Fill in the init block.  Note that most addresses are
689 	 * deliberately in the lower 32bits of memory.  This is a micro-
690 	 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe.
691 	 */
692 	init.Function = MPI2_FUNCTION_IOC_INIT;
693 	init.WhoInit = MPI2_WHOINIT_HOST_DRIVER;
694 	init.MsgVersion = htole16(MPI2_VERSION);
695 	init.HeaderVersion = htole16(MPI2_HEADER_VERSION);
696 	init.SystemRequestFrameSize = htole16(sc->facts->IOCRequestFrameSize);
697 	init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth);
698 	init.ReplyFreeQueueDepth = htole16(sc->fqdepth);
699 	init.SenseBufferAddressHigh = 0;
700 	init.SystemReplyAddressHigh = 0;
701 	init.SystemRequestFrameBaseAddress.High = 0;
702 	init.SystemRequestFrameBaseAddress.Low = htole32((uint32_t)sc->req_busaddr);
703 	init.ReplyDescriptorPostQueueAddress.High = 0;
704 	init.ReplyDescriptorPostQueueAddress.Low = htole32((uint32_t)sc->post_busaddr);
705 	init.ReplyFreeQueueAddress.High = 0;
706 	init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr);
707 	init.TimeStamp.High = 0;
708 	init.TimeStamp.Low = htole32((uint32_t)time_uptime);
709 
710 	error = mps_request_sync(sc, &init, &reply, req_sz, reply_sz, 5);
711 	if ((reply.IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
712 		error = ENXIO;
713 
714 	mps_dprint(sc, MPS_INFO, "IOCInit status= 0x%x\n", reply.IOCStatus);
715 	return (error);
716 }
717 
718 void
719 mps_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
720 {
721 	bus_addr_t *addr;
722 
723 	addr = arg;
724 	*addr = segs[0].ds_addr;
725 }
726 
727 static int
728 mps_alloc_queues(struct mps_softc *sc)
729 {
730 	bus_addr_t queues_busaddr;
731 	uint8_t *queues;
732 	int qsize, fqsize, pqsize;
733 
734 	/*
735 	 * The reply free queue contains 4 byte entries in multiples of 16 and
736 	 * aligned on a 16 byte boundary. There must always be an unused entry.
737 	 * This queue supplies fresh reply frames for the firmware to use.
738 	 *
739 	 * The reply descriptor post queue contains 8 byte entries in
740 	 * multiples of 16 and aligned on a 16 byte boundary.  This queue
741 	 * contains filled-in reply frames sent from the firmware to the host.
742 	 *
743 	 * These two queues are allocated together for simplicity.
744 	 */
745 	sc->fqdepth = roundup2((sc->num_replies + 1), 16);
746 	sc->pqdepth = roundup2((sc->num_replies + 1), 16);
747 	fqsize= sc->fqdepth * 4;
748 	pqsize = sc->pqdepth * 8;
749 	qsize = fqsize + pqsize;
750 
751         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
752 				16, 0,			/* algnmnt, boundary */
753 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
754 				BUS_SPACE_MAXADDR,	/* highaddr */
755 				NULL, NULL,		/* filter, filterarg */
756                                 qsize,			/* maxsize */
757                                 1,			/* nsegments */
758                                 qsize,			/* maxsegsize */
759                                 0,			/* flags */
760                                 NULL, NULL,		/* lockfunc, lockarg */
761                                 &sc->queues_dmat)) {
762 		device_printf(sc->mps_dev, "Cannot allocate queues DMA tag\n");
763 		return (ENOMEM);
764         }
765         if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT,
766 	    &sc->queues_map)) {
767 		device_printf(sc->mps_dev, "Cannot allocate queues memory\n");
768 		return (ENOMEM);
769         }
770         bzero(queues, qsize);
771         bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize,
772 	    mps_memaddr_cb, &queues_busaddr, 0);
773 
774 	sc->free_queue = (uint32_t *)queues;
775 	sc->free_busaddr = queues_busaddr;
776 	sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize);
777 	sc->post_busaddr = queues_busaddr + fqsize;
778 
779 	return (0);
780 }
781 
782 static int
783 mps_alloc_replies(struct mps_softc *sc)
784 {
785 	int rsize, num_replies;
786 
787 	/*
788 	 * sc->num_replies should be one less than sc->fqdepth.  We need to
789 	 * allocate space for sc->fqdepth replies, but only sc->num_replies
790 	 * replies can be used at once.
791 	 */
792 	num_replies = max(sc->fqdepth, sc->num_replies);
793 
794 	rsize = sc->facts->ReplyFrameSize * num_replies * 4;
795         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
796 				4, 0,			/* algnmnt, boundary */
797 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
798 				BUS_SPACE_MAXADDR,	/* highaddr */
799 				NULL, NULL,		/* filter, filterarg */
800                                 rsize,			/* maxsize */
801                                 1,			/* nsegments */
802                                 rsize,			/* maxsegsize */
803                                 0,			/* flags */
804                                 NULL, NULL,		/* lockfunc, lockarg */
805                                 &sc->reply_dmat)) {
806 		device_printf(sc->mps_dev, "Cannot allocate replies DMA tag\n");
807 		return (ENOMEM);
808         }
809         if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames,
810 	    BUS_DMA_NOWAIT, &sc->reply_map)) {
811 		device_printf(sc->mps_dev, "Cannot allocate replies memory\n");
812 		return (ENOMEM);
813         }
814         bzero(sc->reply_frames, rsize);
815         bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize,
816 	    mps_memaddr_cb, &sc->reply_busaddr, 0);
817 
818 	return (0);
819 }
820 
821 static int
822 mps_alloc_requests(struct mps_softc *sc)
823 {
824 	struct mps_command *cm;
825 	struct mps_chain *chain;
826 	int i, rsize, nsegs;
827 
828 	rsize = sc->facts->IOCRequestFrameSize * sc->num_reqs * 4;
829         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
830 				16, 0,			/* algnmnt, boundary */
831 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
832 				BUS_SPACE_MAXADDR,	/* highaddr */
833 				NULL, NULL,		/* filter, filterarg */
834                                 rsize,			/* maxsize */
835                                 1,			/* nsegments */
836                                 rsize,			/* maxsegsize */
837                                 0,			/* flags */
838                                 NULL, NULL,		/* lockfunc, lockarg */
839                                 &sc->req_dmat)) {
840 		device_printf(sc->mps_dev, "Cannot allocate request DMA tag\n");
841 		return (ENOMEM);
842         }
843         if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames,
844 	    BUS_DMA_NOWAIT, &sc->req_map)) {
845 		device_printf(sc->mps_dev, "Cannot allocate request memory\n");
846 		return (ENOMEM);
847         }
848         bzero(sc->req_frames, rsize);
849         bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize,
850 	    mps_memaddr_cb, &sc->req_busaddr, 0);
851 
852 	rsize = sc->facts->IOCRequestFrameSize * sc->max_chains * 4;
853         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
854 				16, 0,			/* algnmnt, boundary */
855 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
856 				BUS_SPACE_MAXADDR,	/* highaddr */
857 				NULL, NULL,		/* filter, filterarg */
858                                 rsize,			/* maxsize */
859                                 1,			/* nsegments */
860                                 rsize,			/* maxsegsize */
861                                 0,			/* flags */
862                                 NULL, NULL,		/* lockfunc, lockarg */
863                                 &sc->chain_dmat)) {
864 		device_printf(sc->mps_dev, "Cannot allocate chain DMA tag\n");
865 		return (ENOMEM);
866         }
867         if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames,
868 	    BUS_DMA_NOWAIT, &sc->chain_map)) {
869 		device_printf(sc->mps_dev, "Cannot allocate chain memory\n");
870 		return (ENOMEM);
871         }
872         bzero(sc->chain_frames, rsize);
873         bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, rsize,
874 	    mps_memaddr_cb, &sc->chain_busaddr, 0);
875 
876 	rsize = MPS_SENSE_LEN * sc->num_reqs;
877         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
878 				1, 0,			/* algnmnt, boundary */
879 				BUS_SPACE_MAXADDR_32BIT,/* lowaddr */
880 				BUS_SPACE_MAXADDR,	/* highaddr */
881 				NULL, NULL,		/* filter, filterarg */
882                                 rsize,			/* maxsize */
883                                 1,			/* nsegments */
884                                 rsize,			/* maxsegsize */
885                                 0,			/* flags */
886                                 NULL, NULL,		/* lockfunc, lockarg */
887                                 &sc->sense_dmat)) {
888 		device_printf(sc->mps_dev, "Cannot allocate sense DMA tag\n");
889 		return (ENOMEM);
890         }
891         if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames,
892 	    BUS_DMA_NOWAIT, &sc->sense_map)) {
893 		device_printf(sc->mps_dev, "Cannot allocate sense memory\n");
894 		return (ENOMEM);
895         }
896         bzero(sc->sense_frames, rsize);
897         bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize,
898 	    mps_memaddr_cb, &sc->sense_busaddr, 0);
899 
900 	sc->chains = malloc(sizeof(struct mps_chain) * sc->max_chains, M_MPT2,
901 	    M_WAITOK | M_ZERO);
902 	if(!sc->chains) {
903 		device_printf(sc->mps_dev,
904 		"Cannot allocate chains memory %s %d\n",
905 		 __func__, __LINE__);
906 		return (ENOMEM);
907 	}
908 	for (i = 0; i < sc->max_chains; i++) {
909 		chain = &sc->chains[i];
910 		chain->chain = (MPI2_SGE_IO_UNION *)(sc->chain_frames +
911 		    i * sc->facts->IOCRequestFrameSize * 4);
912 		chain->chain_busaddr = sc->chain_busaddr +
913 		    i * sc->facts->IOCRequestFrameSize * 4;
914 		mps_free_chain(sc, chain);
915 		sc->chain_free_lowwater++;
916 	}
917 
918 	/* XXX Need to pick a more precise value */
919 	nsegs = (MAXPHYS / PAGE_SIZE) + 1;
920         if (bus_dma_tag_create( sc->mps_parent_dmat,    /* parent */
921 				1, 0,			/* algnmnt, boundary */
922 				BUS_SPACE_MAXADDR,	/* lowaddr */
923 				BUS_SPACE_MAXADDR,	/* highaddr */
924 				NULL, NULL,		/* filter, filterarg */
925                                 BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
926                                 nsegs,			/* nsegments */
927                                 BUS_SPACE_MAXSIZE_24BIT,/* maxsegsize */
928                                 BUS_DMA_ALLOCNOW,	/* flags */
929                                 busdma_lock_mutex,	/* lockfunc */
930 				&sc->mps_mtx,		/* lockarg */
931                                 &sc->buffer_dmat)) {
932 		device_printf(sc->mps_dev, "Cannot allocate buffer DMA tag\n");
933 		return (ENOMEM);
934         }
935 
936 	/*
937 	 * SMID 0 cannot be used as a free command per the firmware spec.
938 	 * Just drop that command instead of risking accounting bugs.
939 	 */
940 	sc->commands = malloc(sizeof(struct mps_command) * sc->num_reqs,
941 	    M_MPT2, M_WAITOK | M_ZERO);
942 	if(!sc->commands) {
943 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
944 		 __func__, __LINE__);
945 		return (ENOMEM);
946 	}
947 	for (i = 1; i < sc->num_reqs; i++) {
948 		cm = &sc->commands[i];
949 		cm->cm_req = sc->req_frames +
950 		    i * sc->facts->IOCRequestFrameSize * 4;
951 		cm->cm_req_busaddr = sc->req_busaddr +
952 		    i * sc->facts->IOCRequestFrameSize * 4;
953 		cm->cm_sense = &sc->sense_frames[i];
954 		cm->cm_sense_busaddr = sc->sense_busaddr + i * MPS_SENSE_LEN;
955 		cm->cm_desc.Default.SMID = i;
956 		cm->cm_sc = sc;
957 		TAILQ_INIT(&cm->cm_chain_list);
958 		callout_init_mtx(&cm->cm_callout, &sc->mps_mtx, 0);
959 
960 		/* XXX Is a failure here a critical problem? */
961 		if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) == 0)
962 			if (i <= sc->facts->HighPriorityCredit)
963 				mps_free_high_priority_command(sc, cm);
964 			else
965 				mps_free_command(sc, cm);
966 		else {
967 			panic("failed to allocate command %d\n", i);
968 			sc->num_reqs = i;
969 			break;
970 		}
971 	}
972 
973 	return (0);
974 }
975 
976 static int
977 mps_init_queues(struct mps_softc *sc)
978 {
979 	int i;
980 
981 	memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8);
982 
983 	/*
984 	 * According to the spec, we need to use one less reply than we
985 	 * have space for on the queue.  So sc->num_replies (the number we
986 	 * use) should be less than sc->fqdepth (allocated size).
987 	 */
988 	if (sc->num_replies >= sc->fqdepth)
989 		return (EINVAL);
990 
991 	/*
992 	 * Initialize all of the free queue entries.
993 	 */
994 	for (i = 0; i < sc->fqdepth; i++)
995 		sc->free_queue[i] = sc->reply_busaddr + (i * sc->facts->ReplyFrameSize * 4);
996 	sc->replyfreeindex = sc->num_replies;
997 
998 	return (0);
999 }
1000 
1001 /* Get the driver parameter tunables.  Lowest priority are the driver defaults.
1002  * Next are the global settings, if they exist.  Highest are the per-unit
1003  * settings, if they exist.
1004  */
1005 static void
1006 mps_get_tunables(struct mps_softc *sc)
1007 {
1008 	char tmpstr[80];
1009 
1010 	/* XXX default to some debugging for now */
1011 	sc->mps_debug = MPS_FAULT;
1012 	sc->disable_msix = 0;
1013 	sc->disable_msi = 0;
1014 	sc->max_chains = MPS_CHAIN_FRAMES;
1015 
1016 	/*
1017 	 * Grab the global variables.
1018 	 */
1019 	TUNABLE_INT_FETCH("hw.mps.debug_level", &sc->mps_debug);
1020 	TUNABLE_INT_FETCH("hw.mps.disable_msix", &sc->disable_msix);
1021 	TUNABLE_INT_FETCH("hw.mps.disable_msi", &sc->disable_msi);
1022 	TUNABLE_INT_FETCH("hw.mps.max_chains", &sc->max_chains);
1023 
1024 	/* Grab the unit-instance variables */
1025 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.debug_level",
1026 	    device_get_unit(sc->mps_dev));
1027 	TUNABLE_INT_FETCH(tmpstr, &sc->mps_debug);
1028 
1029 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msix",
1030 	    device_get_unit(sc->mps_dev));
1031 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix);
1032 
1033 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.disable_msi",
1034 	    device_get_unit(sc->mps_dev));
1035 	TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi);
1036 
1037 	snprintf(tmpstr, sizeof(tmpstr), "dev.mps.%d.max_chains",
1038 	    device_get_unit(sc->mps_dev));
1039 	TUNABLE_INT_FETCH(tmpstr, &sc->max_chains);
1040 }
1041 
1042 static void
1043 mps_setup_sysctl(struct mps_softc *sc)
1044 {
1045 	struct sysctl_ctx_list	*sysctl_ctx = NULL;
1046 	struct sysctl_oid	*sysctl_tree = NULL;
1047 	char tmpstr[80], tmpstr2[80];
1048 
1049 	/*
1050 	 * Setup the sysctl variable so the user can change the debug level
1051 	 * on the fly.
1052 	 */
1053 	snprintf(tmpstr, sizeof(tmpstr), "MPS controller %d",
1054 	    device_get_unit(sc->mps_dev));
1055 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mps_dev));
1056 
1057 	sysctl_ctx = device_get_sysctl_ctx(sc->mps_dev);
1058 	if (sysctl_ctx != NULL)
1059 		sysctl_tree = device_get_sysctl_tree(sc->mps_dev);
1060 
1061 	if (sysctl_tree == NULL) {
1062 		sysctl_ctx_init(&sc->sysctl_ctx);
1063 		sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx,
1064 		    SYSCTL_STATIC_CHILDREN(_hw_mps), OID_AUTO, tmpstr2,
1065 		    CTLFLAG_RD, 0, tmpstr);
1066 		if (sc->sysctl_tree == NULL)
1067 			return;
1068 		sysctl_ctx = &sc->sysctl_ctx;
1069 		sysctl_tree = sc->sysctl_tree;
1070 	}
1071 
1072 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1073 	    OID_AUTO, "debug_level", CTLFLAG_RW, &sc->mps_debug, 0,
1074 	    "mps debug level");
1075 
1076 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1077 	    OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0,
1078 	    "Disable the use of MSI-X interrupts");
1079 
1080 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1081 	    OID_AUTO, "disable_msi", CTLFLAG_RD, &sc->disable_msi, 0,
1082 	    "Disable the use of MSI interrupts");
1083 
1084 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1085 	    OID_AUTO, "firmware_version", CTLFLAG_RW, &sc->fw_version,
1086 	    strlen(sc->fw_version), "firmware version");
1087 
1088 	SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1089 	    OID_AUTO, "driver_version", CTLFLAG_RW, MPS_DRIVER_VERSION,
1090 	    strlen(MPS_DRIVER_VERSION), "driver version");
1091 
1092 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1093 	    OID_AUTO, "io_cmds_active", CTLFLAG_RD,
1094 	    &sc->io_cmds_active, 0, "number of currently active commands");
1095 
1096 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1097 	    OID_AUTO, "io_cmds_highwater", CTLFLAG_RD,
1098 	    &sc->io_cmds_highwater, 0, "maximum active commands seen");
1099 
1100 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1101 	    OID_AUTO, "chain_free", CTLFLAG_RD,
1102 	    &sc->chain_free, 0, "number of free chain elements");
1103 
1104 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1105 	    OID_AUTO, "chain_free_lowwater", CTLFLAG_RD,
1106 	    &sc->chain_free_lowwater, 0,"lowest number of free chain elements");
1107 
1108 	SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1109 	    OID_AUTO, "max_chains", CTLFLAG_RD,
1110 	    &sc->max_chains, 0,"maximum chain frames that will be allocated");
1111 
1112 #if __FreeBSD_version >= 900030
1113 	SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree),
1114 	    OID_AUTO, "chain_alloc_fail", CTLFLAG_RD,
1115 	    &sc->chain_alloc_fail, "chain allocation failures");
1116 #endif //FreeBSD_version >= 900030
1117 }
1118 
1119 int
1120 mps_attach(struct mps_softc *sc)
1121 {
1122 	int i, error;
1123 
1124 	mps_get_tunables(sc);
1125 
1126 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1127 
1128 	mtx_init(&sc->mps_mtx, "MPT2SAS lock", NULL, MTX_DEF);
1129 	callout_init_mtx(&sc->periodic, &sc->mps_mtx, 0);
1130 	TAILQ_INIT(&sc->event_list);
1131 
1132 	if ((error = mps_transition_ready(sc)) != 0) {
1133 		mps_printf(sc, "%s failed to transition ready\n", __func__);
1134 		return (error);
1135 	}
1136 
1137 	sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPT2,
1138 	    M_ZERO|M_NOWAIT);
1139 	if(!sc->facts) {
1140 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1141 		 __func__, __LINE__);
1142 		return (ENOMEM);
1143 	}
1144 	if ((error = mps_get_iocfacts(sc, sc->facts)) != 0)
1145 		return (error);
1146 
1147 	mps_print_iocfacts(sc, sc->facts);
1148 
1149 	snprintf(sc->fw_version, sizeof(sc->fw_version),
1150 	    "%02d.%02d.%02d.%02d",
1151 	    sc->facts->FWVersion.Struct.Major,
1152 	    sc->facts->FWVersion.Struct.Minor,
1153 	    sc->facts->FWVersion.Struct.Unit,
1154 	    sc->facts->FWVersion.Struct.Dev);
1155 
1156 	mps_printf(sc, "Firmware: %s, Driver: %s\n", sc->fw_version,
1157 	    MPS_DRIVER_VERSION);
1158 	mps_printf(sc, "IOCCapabilities: %b\n", sc->facts->IOCCapabilities,
1159 	    "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf"
1160 	    "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR"
1161 	    "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc");
1162 
1163 	/*
1164 	 * If the chip doesn't support event replay then a hard reset will be
1165 	 * required to trigger a full discovery.  Do the reset here then
1166 	 * retransition to Ready.  A hard reset might have already been done,
1167 	 * but it doesn't hurt to do it again.
1168 	 */
1169 	if ((sc->facts->IOCCapabilities &
1170 	    MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0) {
1171 		mps_diag_reset(sc, NO_SLEEP);
1172 		if ((error = mps_transition_ready(sc)) != 0)
1173 			return (error);
1174 	}
1175 
1176 	/*
1177 	 * Set flag if IR Firmware is loaded.
1178 	 */
1179 	if (sc->facts->IOCCapabilities &
1180 	    MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID)
1181 		sc->ir_firmware = 1;
1182 
1183 	/*
1184 	 * Check if controller supports FW diag buffers and set flag to enable
1185 	 * each type.
1186 	 */
1187 	if (sc->facts->IOCCapabilities &
1188 	    MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER)
1189 		sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE].enabled =
1190 		    TRUE;
1191 	if (sc->facts->IOCCapabilities &
1192 	    MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER)
1193 		sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT].enabled =
1194 		    TRUE;
1195 	if (sc->facts->IOCCapabilities &
1196 	    MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER)
1197 		sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED].enabled =
1198 		    TRUE;
1199 
1200 	/*
1201 	 * Set flag if EEDP is supported and if TLR is supported.
1202 	 */
1203 	if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP)
1204 		sc->eedp_enabled = TRUE;
1205 	if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR)
1206 		sc->control_TLR = TRUE;
1207 
1208 	/*
1209 	 * Size the queues. Since the reply queues always need one free entry,
1210 	 * we'll just deduct one reply message here.
1211 	 */
1212 	sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
1213 	sc->num_replies = MIN(MPS_REPLY_FRAMES + MPS_EVT_REPLY_FRAMES,
1214 	    sc->facts->MaxReplyDescriptorPostQueueDepth) - 1;
1215 	TAILQ_INIT(&sc->req_list);
1216 	TAILQ_INIT(&sc->high_priority_req_list);
1217 	TAILQ_INIT(&sc->chain_list);
1218 	TAILQ_INIT(&sc->tm_list);
1219 
1220 	if (((error = mps_alloc_queues(sc)) != 0) ||
1221 	    ((error = mps_alloc_replies(sc)) != 0) ||
1222 	    ((error = mps_alloc_requests(sc)) != 0)) {
1223 		mps_printf(sc, "%s failed to alloc\n", __func__);
1224 		mps_free(sc);
1225 		return (error);
1226 	}
1227 
1228 	if (((error = mps_init_queues(sc)) != 0) ||
1229 	    ((error = mps_transition_operational(sc)) != 0)) {
1230 		mps_printf(sc, "%s failed to transition operational\n", __func__);
1231 		mps_free(sc);
1232 		return (error);
1233 	}
1234 
1235 	/*
1236 	 * Finish the queue initialization.
1237 	 * These are set here instead of in mps_init_queues() because the
1238 	 * IOC resets these values during the state transition in
1239 	 * mps_transition_operational().  The free index is set to 1
1240 	 * because the corresponding index in the IOC is set to 0, and the
1241 	 * IOC treats the queues as full if both are set to the same value.
1242 	 * Hence the reason that the queue can't hold all of the possible
1243 	 * replies.
1244 	 */
1245 	sc->replypostindex = 0;
1246 	mps_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex);
1247 	mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0);
1248 
1249 	sc->pfacts = malloc(sizeof(MPI2_PORT_FACTS_REPLY) *
1250 	    sc->facts->NumberOfPorts, M_MPT2, M_ZERO|M_WAITOK);
1251 	if(!sc->pfacts) {
1252 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1253 		 __func__, __LINE__);
1254 		return (ENOMEM);
1255 	}
1256 	for (i = 0; i < sc->facts->NumberOfPorts; i++) {
1257 		if ((error = mps_get_portfacts(sc, &sc->pfacts[i], i)) != 0) {
1258 			mps_printf(sc, "%s failed to get portfacts for port %d\n",
1259 			    __func__, i);
1260 			mps_free(sc);
1261 			return (error);
1262 		}
1263 		mps_print_portfacts(sc, &sc->pfacts[i]);
1264 	}
1265 
1266 	/* Attach the subsystems so they can prepare their event masks. */
1267 	/* XXX Should be dynamic so that IM/IR and user modules can attach */
1268 	if (((error = mps_attach_log(sc)) != 0) ||
1269 	    ((error = mps_attach_sas(sc)) != 0) ||
1270 	    ((error = mps_attach_user(sc)) != 0)) {
1271 		mps_printf(sc, "%s failed to attach all subsystems: error %d\n",
1272 		    __func__, error);
1273 		mps_free(sc);
1274 		return (error);
1275 	}
1276 
1277 	if ((error = mps_pci_setup_interrupts(sc)) != 0) {
1278 		mps_printf(sc, "%s failed to setup interrupts\n", __func__);
1279 		mps_free(sc);
1280 		return (error);
1281 	}
1282 
1283 	/*
1284 	 * The static page function currently read is ioc page8.  Others can be
1285 	 * added in future.
1286 	 */
1287 	mps_base_static_config_pages(sc);
1288 
1289 	/* Start the periodic watchdog check on the IOC Doorbell */
1290 	mps_periodic(sc);
1291 
1292 	/*
1293 	 * The portenable will kick off discovery events that will drive the
1294 	 * rest of the initialization process.  The CAM/SAS module will
1295 	 * hold up the boot sequence until discovery is complete.
1296 	 */
1297 	sc->mps_ich.ich_func = mps_startup;
1298 	sc->mps_ich.ich_arg = sc;
1299 	if (config_intrhook_establish(&sc->mps_ich) != 0) {
1300 		mps_dprint(sc, MPS_FAULT, "Cannot establish MPS config hook\n");
1301 		error = EINVAL;
1302 	}
1303 
1304 	/*
1305 	 * Allow IR to shutdown gracefully when shutdown occurs.
1306 	 */
1307 	sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final,
1308 	    mpssas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT);
1309 
1310 	if (sc->shutdown_eh == NULL)
1311 		mps_dprint(sc, MPS_FAULT, "shutdown event registration "
1312 		    "failed\n");
1313 
1314 	mps_setup_sysctl(sc);
1315 
1316 	sc->mps_flags |= MPS_FLAGS_ATTACH_DONE;
1317 
1318 	return (error);
1319 }
1320 
1321 /* Run through any late-start handlers. */
1322 static void
1323 mps_startup(void *arg)
1324 {
1325 	struct mps_softc *sc;
1326 
1327 	sc = (struct mps_softc *)arg;
1328 
1329 	mps_lock(sc);
1330 	mps_unmask_intr(sc);
1331 	/* initialize device mapping tables */
1332 	mps_mapping_initialize(sc);
1333 	mpssas_startup(sc);
1334 	mps_unlock(sc);
1335 }
1336 
1337 /* Periodic watchdog.  Is called with the driver lock already held. */
1338 static void
1339 mps_periodic(void *arg)
1340 {
1341 	struct mps_softc *sc;
1342 	uint32_t db;
1343 
1344 	sc = (struct mps_softc *)arg;
1345 	if (sc->mps_flags & MPS_FLAGS_SHUTDOWN)
1346 		return;
1347 
1348 	db = mps_regread(sc, MPI2_DOORBELL_OFFSET);
1349 	if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) {
1350 		device_printf(sc->mps_dev, "IOC Fault 0x%08x, Resetting\n", db);
1351 
1352 		mps_reinit(sc);
1353 	}
1354 
1355 	callout_reset(&sc->periodic, MPS_PERIODIC_DELAY * hz, mps_periodic, sc);
1356 }
1357 
1358 static void
1359 mps_log_evt_handler(struct mps_softc *sc, uintptr_t data,
1360     MPI2_EVENT_NOTIFICATION_REPLY *event)
1361 {
1362 	MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry;
1363 
1364 	mps_print_event(sc, event);
1365 
1366 	switch (event->Event) {
1367 	case MPI2_EVENT_LOG_DATA:
1368 		device_printf(sc->mps_dev, "MPI2_EVENT_LOG_DATA:\n");
1369 		hexdump(event->EventData, event->EventDataLength, NULL, 0);
1370 		break;
1371 	case MPI2_EVENT_LOG_ENTRY_ADDED:
1372 		entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData;
1373 		mps_dprint(sc, MPS_INFO, "MPI2_EVENT_LOG_ENTRY_ADDED event "
1374 		    "0x%x Sequence %d:\n", entry->LogEntryQualifier,
1375 		     entry->LogSequence);
1376 		break;
1377 	default:
1378 		break;
1379 	}
1380 	return;
1381 }
1382 
1383 static int
1384 mps_attach_log(struct mps_softc *sc)
1385 {
1386 	u32 events[MPI2_EVENT_NOTIFY_EVENTMASK_WORDS];
1387 
1388 	bzero(events, 16);
1389 	setbit(events, MPI2_EVENT_LOG_DATA);
1390 	setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED);
1391 
1392 	mps_register_events(sc, events, mps_log_evt_handler, NULL,
1393 	    &sc->mps_log_eh);
1394 
1395 	return (0);
1396 }
1397 
1398 static int
1399 mps_detach_log(struct mps_softc *sc)
1400 {
1401 
1402 	if (sc->mps_log_eh != NULL)
1403 		mps_deregister_events(sc, sc->mps_log_eh);
1404 	return (0);
1405 }
1406 
1407 /*
1408  * Free all of the driver resources and detach submodules.  Should be called
1409  * without the lock held.
1410  */
1411 int
1412 mps_free(struct mps_softc *sc)
1413 {
1414 	struct mps_command *cm;
1415 	int i, error;
1416 
1417 	/* Turn off the watchdog */
1418 	mps_lock(sc);
1419 	sc->mps_flags |= MPS_FLAGS_SHUTDOWN;
1420 	mps_unlock(sc);
1421 	/* Lock must not be held for this */
1422 	callout_drain(&sc->periodic);
1423 
1424 	if (((error = mps_detach_log(sc)) != 0) ||
1425 	    ((error = mps_detach_sas(sc)) != 0))
1426 		return (error);
1427 
1428 	mps_detach_user(sc);
1429 
1430 	/* Put the IOC back in the READY state. */
1431 	mps_lock(sc);
1432 	if ((error = mps_transition_ready(sc)) != 0) {
1433 		mps_unlock(sc);
1434 		return (error);
1435 	}
1436 	mps_unlock(sc);
1437 
1438 	if (sc->facts != NULL)
1439 		free(sc->facts, M_MPT2);
1440 
1441 	if (sc->pfacts != NULL)
1442 		free(sc->pfacts, M_MPT2);
1443 
1444 	if (sc->post_busaddr != 0)
1445 		bus_dmamap_unload(sc->queues_dmat, sc->queues_map);
1446 	if (sc->post_queue != NULL)
1447 		bus_dmamem_free(sc->queues_dmat, sc->post_queue,
1448 		    sc->queues_map);
1449 	if (sc->queues_dmat != NULL)
1450 		bus_dma_tag_destroy(sc->queues_dmat);
1451 
1452 	if (sc->chain_busaddr != 0)
1453 		bus_dmamap_unload(sc->chain_dmat, sc->chain_map);
1454 	if (sc->chain_frames != NULL)
1455 		bus_dmamem_free(sc->chain_dmat, sc->chain_frames,sc->chain_map);
1456 	if (sc->chain_dmat != NULL)
1457 		bus_dma_tag_destroy(sc->chain_dmat);
1458 
1459 	if (sc->sense_busaddr != 0)
1460 		bus_dmamap_unload(sc->sense_dmat, sc->sense_map);
1461 	if (sc->sense_frames != NULL)
1462 		bus_dmamem_free(sc->sense_dmat, sc->sense_frames,sc->sense_map);
1463 	if (sc->sense_dmat != NULL)
1464 		bus_dma_tag_destroy(sc->sense_dmat);
1465 
1466 	if (sc->reply_busaddr != 0)
1467 		bus_dmamap_unload(sc->reply_dmat, sc->reply_map);
1468 	if (sc->reply_frames != NULL)
1469 		bus_dmamem_free(sc->reply_dmat, sc->reply_frames,sc->reply_map);
1470 	if (sc->reply_dmat != NULL)
1471 		bus_dma_tag_destroy(sc->reply_dmat);
1472 
1473 	if (sc->req_busaddr != 0)
1474 		bus_dmamap_unload(sc->req_dmat, sc->req_map);
1475 	if (sc->req_frames != NULL)
1476 		bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map);
1477 	if (sc->req_dmat != NULL)
1478 		bus_dma_tag_destroy(sc->req_dmat);
1479 
1480 	if (sc->chains != NULL)
1481 		free(sc->chains, M_MPT2);
1482 	if (sc->commands != NULL) {
1483 		for (i = 1; i < sc->num_reqs; i++) {
1484 			cm = &sc->commands[i];
1485 			bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap);
1486 		}
1487 		free(sc->commands, M_MPT2);
1488 	}
1489 	if (sc->buffer_dmat != NULL)
1490 		bus_dma_tag_destroy(sc->buffer_dmat);
1491 
1492 	if (sc->sysctl_tree != NULL)
1493 		sysctl_ctx_free(&sc->sysctl_ctx);
1494 
1495 	mps_mapping_free_memory(sc);
1496 
1497 	/* Deregister the shutdown function */
1498 	if (sc->shutdown_eh != NULL)
1499 		EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh);
1500 
1501 	mtx_destroy(&sc->mps_mtx);
1502 
1503 	return (0);
1504 }
1505 
1506 static __inline void
1507 mps_complete_command(struct mps_command *cm)
1508 {
1509 	if (cm->cm_flags & MPS_CM_FLAGS_POLLED)
1510 		cm->cm_flags |= MPS_CM_FLAGS_COMPLETE;
1511 
1512 	if (cm->cm_complete != NULL) {
1513 		mps_dprint(cm->cm_sc, MPS_TRACE,
1514 			   "%s cm %p calling cm_complete %p data %p reply %p\n",
1515 			   __func__, cm, cm->cm_complete, cm->cm_complete_data,
1516 			   cm->cm_reply);
1517 		cm->cm_complete(cm->cm_sc, cm);
1518 	}
1519 
1520 	if (cm->cm_flags & MPS_CM_FLAGS_WAKEUP) {
1521 		mps_dprint(cm->cm_sc, MPS_TRACE, "%s: waking up %p\n",
1522 			   __func__, cm);
1523 		wakeup(cm);
1524 	}
1525 
1526 	if (cm->cm_sc->io_cmds_active != 0) {
1527 		cm->cm_sc->io_cmds_active--;
1528 	} else {
1529 		mps_dprint(cm->cm_sc, MPS_INFO, "Warning: io_cmds_active is "
1530 		    "out of sync - resynching to 0\n");
1531 	}
1532 }
1533 
1534 
1535 static void
1536 mps_sas_log_info(struct mps_softc *sc , u32 log_info)
1537 {
1538 	union loginfo_type {
1539 		u32     loginfo;
1540 		struct {
1541 			u32     subcode:16;
1542 			u32     code:8;
1543 			u32     originator:4;
1544 			u32     bus_type:4;
1545 		} dw;
1546 	};
1547 	union loginfo_type sas_loginfo;
1548 	char *originator_str = NULL;
1549 
1550 	sas_loginfo.loginfo = log_info;
1551 	if (sas_loginfo.dw.bus_type != 3 /*SAS*/)
1552 		return;
1553 
1554 	/* each nexus loss loginfo */
1555 	if (log_info == 0x31170000)
1556 		return;
1557 
1558 	/* eat the loginfos associated with task aborts */
1559 	if ((log_info == 30050000 || log_info ==
1560 	    0x31140000 || log_info == 0x31130000))
1561 		return;
1562 
1563 	switch (sas_loginfo.dw.originator) {
1564 	case 0:
1565 		originator_str = "IOP";
1566 		break;
1567 	case 1:
1568 		originator_str = "PL";
1569 		break;
1570 	case 2:
1571 		originator_str = "IR";
1572 		break;
1573 }
1574 
1575 	mps_dprint(sc, MPS_INFO, "log_info(0x%08x): originator(%s), "
1576 	"code(0x%02x), sub_code(0x%04x)\n", log_info,
1577 	originator_str, sas_loginfo.dw.code,
1578 	sas_loginfo.dw.subcode);
1579 }
1580 
1581 static void
1582 mps_display_reply_info(struct mps_softc *sc, uint8_t *reply)
1583 {
1584 	MPI2DefaultReply_t *mpi_reply;
1585 	u16 sc_status;
1586 
1587 	mpi_reply = (MPI2DefaultReply_t*)reply;
1588 	sc_status = le16toh(mpi_reply->IOCStatus);
1589 	if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE)
1590 		mps_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo));
1591 }
1592 void
1593 mps_intr(void *data)
1594 {
1595 	struct mps_softc *sc;
1596 	uint32_t status;
1597 
1598 	sc = (struct mps_softc *)data;
1599 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1600 
1601 	/*
1602 	 * Check interrupt status register to flush the bus.  This is
1603 	 * needed for both INTx interrupts and driver-driven polling
1604 	 */
1605 	status = mps_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET);
1606 	if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0)
1607 		return;
1608 
1609 	mps_lock(sc);
1610 	mps_intr_locked(data);
1611 	mps_unlock(sc);
1612 	return;
1613 }
1614 
1615 /*
1616  * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the
1617  * chip.  Hopefully this theory is correct.
1618  */
1619 void
1620 mps_intr_msi(void *data)
1621 {
1622 	struct mps_softc *sc;
1623 
1624 	sc = (struct mps_softc *)data;
1625 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1626 	mps_lock(sc);
1627 	mps_intr_locked(data);
1628 	mps_unlock(sc);
1629 	return;
1630 }
1631 
1632 /*
1633  * The locking is overly broad and simplistic, but easy to deal with for now.
1634  */
1635 void
1636 mps_intr_locked(void *data)
1637 {
1638 	MPI2_REPLY_DESCRIPTORS_UNION *desc;
1639 	struct mps_softc *sc;
1640 	struct mps_command *cm = NULL;
1641 	uint8_t flags;
1642 	u_int pq;
1643 	MPI2_DIAG_RELEASE_REPLY *rel_rep;
1644 	mps_fw_diagnostic_buffer_t *pBuffer;
1645 
1646 	sc = (struct mps_softc *)data;
1647 
1648 	pq = sc->replypostindex;
1649 	mps_dprint(sc, MPS_TRACE,
1650 	    "%s sc %p starting with replypostindex %u\n",
1651 	    __func__, sc, sc->replypostindex);
1652 
1653 	for ( ;; ) {
1654 		cm = NULL;
1655 		desc = &sc->post_queue[sc->replypostindex];
1656 		flags = desc->Default.ReplyFlags &
1657 		    MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK;
1658 		if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED)
1659 		 || (le32toh(desc->Words.High) == 0xffffffff))
1660 			break;
1661 
1662 		/* increment the replypostindex now, so that event handlers
1663 		 * and cm completion handlers which decide to do a diag
1664 		 * reset can zero it without it getting incremented again
1665 		 * afterwards, and we break out of this loop on the next
1666 		 * iteration since the reply post queue has been cleared to
1667 		 * 0xFF and all descriptors look unused (which they are).
1668 		 */
1669 		if (++sc->replypostindex >= sc->pqdepth)
1670 			sc->replypostindex = 0;
1671 
1672 		switch (flags) {
1673 		case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS:
1674 			cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)];
1675 			cm->cm_reply = NULL;
1676 			break;
1677 		case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY:
1678 		{
1679 			uint32_t baddr;
1680 			uint8_t *reply;
1681 
1682 			/*
1683 			 * Re-compose the reply address from the address
1684 			 * sent back from the chip.  The ReplyFrameAddress
1685 			 * is the lower 32 bits of the physical address of
1686 			 * particular reply frame.  Convert that address to
1687 			 * host format, and then use that to provide the
1688 			 * offset against the virtual address base
1689 			 * (sc->reply_frames).
1690 			 */
1691 			baddr = le32toh(desc->AddressReply.ReplyFrameAddress);
1692 			reply = sc->reply_frames +
1693 				(baddr - ((uint32_t)sc->reply_busaddr));
1694 			/*
1695 			 * Make sure the reply we got back is in a valid
1696 			 * range.  If not, go ahead and panic here, since
1697 			 * we'll probably panic as soon as we deference the
1698 			 * reply pointer anyway.
1699 			 */
1700 			if ((reply < sc->reply_frames)
1701 			 || (reply > (sc->reply_frames +
1702 			     (sc->fqdepth * sc->facts->ReplyFrameSize * 4)))) {
1703 				printf("%s: WARNING: reply %p out of range!\n",
1704 				       __func__, reply);
1705 				printf("%s: reply_frames %p, fqdepth %d, "
1706 				       "frame size %d\n", __func__,
1707 				       sc->reply_frames, sc->fqdepth,
1708 				       sc->facts->ReplyFrameSize * 4);
1709 				printf("%s: baddr %#x,\n", __func__, baddr);
1710 				/* LSI-TODO. See Linux Code. Need Gracefull exit*/
1711 				panic("Reply address out of range");
1712 			}
1713 			if (le16toh(desc->AddressReply.SMID) == 0) {
1714 				if (((MPI2_DEFAULT_REPLY *)reply)->Function ==
1715 				    MPI2_FUNCTION_DIAG_BUFFER_POST) {
1716 					/*
1717 					 * If SMID is 0 for Diag Buffer Post,
1718 					 * this implies that the reply is due to
1719 					 * a release function with a status that
1720 					 * the buffer has been released.  Set
1721 					 * the buffer flags accordingly.
1722 					 */
1723 					rel_rep =
1724 					    (MPI2_DIAG_RELEASE_REPLY *)reply;
1725 					if (le16toh(rel_rep->IOCStatus) ==
1726 					    MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED)
1727 					    {
1728 						pBuffer =
1729 						    &sc->fw_diag_buffer_list[
1730 						    rel_rep->BufferType];
1731 						pBuffer->valid_data = TRUE;
1732 						pBuffer->owned_by_firmware =
1733 						    FALSE;
1734 						pBuffer->immediate = FALSE;
1735 					}
1736 				} else
1737 					mps_dispatch_event(sc, baddr,
1738 					    (MPI2_EVENT_NOTIFICATION_REPLY *)
1739 					    reply);
1740 			} else {
1741 				cm = &sc->commands[le16toh(desc->AddressReply.SMID)];
1742 				cm->cm_reply = reply;
1743 				cm->cm_reply_data =
1744 				    le32toh(desc->AddressReply.ReplyFrameAddress);
1745 			}
1746 			break;
1747 		}
1748 		case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS:
1749 		case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER:
1750 		case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS:
1751 		default:
1752 			/* Unhandled */
1753 			device_printf(sc->mps_dev, "Unhandled reply 0x%x\n",
1754 			    desc->Default.ReplyFlags);
1755 			cm = NULL;
1756 			break;
1757 		}
1758 
1759 
1760 		if (cm != NULL) {
1761 			// Print Error reply frame
1762 			if (cm->cm_reply)
1763 				mps_display_reply_info(sc,cm->cm_reply);
1764 			mps_complete_command(cm);
1765 		}
1766 
1767 		desc->Words.Low = 0xffffffff;
1768 		desc->Words.High = 0xffffffff;
1769 	}
1770 
1771 	if (pq != sc->replypostindex) {
1772 		mps_dprint(sc, MPS_TRACE,
1773 		    "%s sc %p writing postindex %d\n",
1774 		    __func__, sc, sc->replypostindex);
1775 		mps_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, sc->replypostindex);
1776 	}
1777 
1778 	return;
1779 }
1780 
1781 static void
1782 mps_dispatch_event(struct mps_softc *sc, uintptr_t data,
1783     MPI2_EVENT_NOTIFICATION_REPLY *reply)
1784 {
1785 	struct mps_event_handle *eh;
1786 	int event, handled = 0;
1787 
1788 	event = le16toh(reply->Event);
1789 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1790 		if (isset(eh->mask, event)) {
1791 			eh->callback(sc, data, reply);
1792 			handled++;
1793 		}
1794 	}
1795 
1796 	if (handled == 0)
1797 		device_printf(sc->mps_dev, "Unhandled event 0x%x\n", le16toh(event));
1798 
1799 	/*
1800 	 * This is the only place that the event/reply should be freed.
1801 	 * Anything wanting to hold onto the event data should have
1802 	 * already copied it into their own storage.
1803 	 */
1804 	mps_free_reply(sc, data);
1805 }
1806 
1807 static void
1808 mps_reregister_events_complete(struct mps_softc *sc, struct mps_command *cm)
1809 {
1810 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1811 
1812 	if (cm->cm_reply)
1813 		mps_print_event(sc,
1814 			(MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply);
1815 
1816 	mps_free_command(sc, cm);
1817 
1818 	/* next, send a port enable */
1819 	mpssas_startup(sc);
1820 }
1821 
1822 /*
1823  * For both register_events and update_events, the caller supplies a bitmap
1824  * of events that it _wants_.  These functions then turn that into a bitmask
1825  * suitable for the controller.
1826  */
1827 int
1828 mps_register_events(struct mps_softc *sc, u32 *mask,
1829     mps_evt_callback_t *cb, void *data, struct mps_event_handle **handle)
1830 {
1831 	struct mps_event_handle *eh;
1832 	int error = 0;
1833 
1834 	eh = malloc(sizeof(struct mps_event_handle), M_MPT2, M_WAITOK|M_ZERO);
1835 	if(!eh) {
1836 		device_printf(sc->mps_dev, "Cannot allocate memory %s %d\n",
1837 		 __func__, __LINE__);
1838 		return (ENOMEM);
1839 	}
1840 	eh->callback = cb;
1841 	eh->data = data;
1842 	TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list);
1843 	if (mask != NULL)
1844 		error = mps_update_events(sc, eh, mask);
1845 	*handle = eh;
1846 
1847 	return (error);
1848 }
1849 
1850 int
1851 mps_update_events(struct mps_softc *sc, struct mps_event_handle *handle,
1852     u32 *mask)
1853 {
1854 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
1855 	MPI2_EVENT_NOTIFICATION_REPLY *reply;
1856 	struct mps_command *cm;
1857 	int error, i;
1858 
1859 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1860 
1861 	if ((mask != NULL) && (handle != NULL))
1862 		bcopy(mask, &handle->mask[0], sizeof(u32) *
1863 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
1864 
1865 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
1866 		sc->event_mask[i] = -1;
1867 
1868 	for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
1869 		sc->event_mask[i] &= ~handle->mask[i];
1870 
1871 
1872 	if ((cm = mps_alloc_command(sc)) == NULL)
1873 		return (EBUSY);
1874 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
1875 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
1876 	evtreq->MsgFlags = 0;
1877 	evtreq->SASBroadcastPrimitiveMasks = 0;
1878 #ifdef MPS_DEBUG_ALL_EVENTS
1879 	{
1880 		u_char fullmask[16];
1881 		memset(fullmask, 0x00, 16);
1882 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
1883 				MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
1884 	}
1885 #else
1886         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
1887                 evtreq->EventMasks[i] =
1888                     htole32(sc->event_mask[i]);
1889 #endif
1890 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
1891 	cm->cm_data = NULL;
1892 
1893 	error = mps_request_polled(sc, cm);
1894 	reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply;
1895 	if ((reply == NULL) ||
1896 	    (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS)
1897 		error = ENXIO;
1898 	mps_print_event(sc, reply);
1899 	mps_dprint(sc, MPS_TRACE, "%s finished error %d\n", __func__, error);
1900 
1901 	mps_free_command(sc, cm);
1902 	return (error);
1903 }
1904 
1905 static int
1906 mps_reregister_events(struct mps_softc *sc)
1907 {
1908 	MPI2_EVENT_NOTIFICATION_REQUEST *evtreq;
1909 	struct mps_command *cm;
1910 	struct mps_event_handle *eh;
1911 	int error, i;
1912 
1913 	mps_dprint(sc, MPS_TRACE, "%s\n", __func__);
1914 
1915 	/* first, reregister events */
1916 
1917     for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
1918 		sc->event_mask[i] = -1;
1919 
1920 	TAILQ_FOREACH(eh, &sc->event_list, eh_list) {
1921 		for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
1922 			sc->event_mask[i] &= ~eh->mask[i];
1923 	}
1924 
1925 	if ((cm = mps_alloc_command(sc)) == NULL)
1926 		return (EBUSY);
1927 	evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req;
1928 	evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION;
1929 	evtreq->MsgFlags = 0;
1930 	evtreq->SASBroadcastPrimitiveMasks = 0;
1931 #ifdef MPS_DEBUG_ALL_EVENTS
1932 	{
1933 		u_char fullmask[16];
1934 		memset(fullmask, 0x00, 16);
1935 		bcopy(fullmask, &evtreq->EventMasks[0], sizeof(u32) *
1936 			MPI2_EVENT_NOTIFY_EVENTMASK_WORDS);
1937 	}
1938 #else
1939         for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++)
1940                 evtreq->EventMasks[i] =
1941                     htole32(sc->event_mask[i]);
1942 #endif
1943 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
1944 	cm->cm_data = NULL;
1945 	cm->cm_complete = mps_reregister_events_complete;
1946 
1947 	error = mps_map_command(sc, cm);
1948 
1949 	mps_dprint(sc, MPS_TRACE, "%s finished with error %d\n", __func__, error);
1950 	return (error);
1951 }
1952 
1953 void
1954 mps_deregister_events(struct mps_softc *sc, struct mps_event_handle *handle)
1955 {
1956 
1957 	TAILQ_REMOVE(&sc->event_list, handle, eh_list);
1958 	free(handle, M_MPT2);
1959 }
1960 
1961 /*
1962  * Add a chain element as the next SGE for the specified command.
1963  * Reset cm_sge and cm_sgesize to indicate all the available space.
1964  */
1965 static int
1966 mps_add_chain(struct mps_command *cm)
1967 {
1968 	MPI2_SGE_CHAIN32 *sgc;
1969 	struct mps_chain *chain;
1970 	int space;
1971 
1972 	if (cm->cm_sglsize < MPS_SGC_SIZE)
1973 		panic("MPS: Need SGE Error Code\n");
1974 
1975 	chain = mps_alloc_chain(cm->cm_sc);
1976 	if (chain == NULL)
1977 		return (ENOBUFS);
1978 
1979 	space = (int)cm->cm_sc->facts->IOCRequestFrameSize * 4;
1980 
1981 	/*
1982 	 * Note: a double-linked list is used to make it easier to
1983 	 * walk for debugging.
1984 	 */
1985 	TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link);
1986 
1987 	sgc = (MPI2_SGE_CHAIN32 *)&cm->cm_sge->MpiChain;
1988 	sgc->Length = htole16(space);
1989 	sgc->NextChainOffset = 0;
1990 	/* TODO Looks like bug in Setting sgc->Flags.
1991 	 *	sgc->Flags = ( MPI2_SGE_FLAGS_CHAIN_ELEMENT | MPI2_SGE_FLAGS_64_BIT_ADDRESSING |
1992 	 *	            MPI2_SGE_FLAGS_SYSTEM_ADDRESS) << MPI2_SGE_FLAGS_SHIFT
1993 	 *	This is fine.. because we are not using simple element. In case of
1994 	 *	MPI2_SGE_CHAIN32, we have seperate Length and Flags feild.
1995  	 */
1996 	sgc->Flags = MPI2_SGE_FLAGS_CHAIN_ELEMENT;
1997 	sgc->Address = htole32(chain->chain_busaddr);
1998 
1999 	cm->cm_sge = (MPI2_SGE_IO_UNION *)&chain->chain->MpiSimple;
2000 	cm->cm_sglsize = space;
2001 	return (0);
2002 }
2003 
2004 /*
2005  * Add one scatter-gather element (chain, simple, transaction context)
2006  * to the scatter-gather list for a command.  Maintain cm_sglsize and
2007  * cm_sge as the remaining size and pointer to the next SGE to fill
2008  * in, respectively.
2009  */
2010 int
2011 mps_push_sge(struct mps_command *cm, void *sgep, size_t len, int segsleft)
2012 {
2013 	MPI2_SGE_TRANSACTION_UNION *tc = sgep;
2014 	MPI2_SGE_SIMPLE64 *sge = sgep;
2015 	int error, type;
2016 	uint32_t saved_buf_len, saved_address_low, saved_address_high;
2017 
2018 	type = (tc->Flags & MPI2_SGE_FLAGS_ELEMENT_MASK);
2019 
2020 #ifdef INVARIANTS
2021 	switch (type) {
2022 	case MPI2_SGE_FLAGS_TRANSACTION_ELEMENT: {
2023 		if (len != tc->DetailsLength + 4)
2024 			panic("TC %p length %u or %zu?", tc,
2025 			    tc->DetailsLength + 4, len);
2026 		}
2027 		break;
2028 	case MPI2_SGE_FLAGS_CHAIN_ELEMENT:
2029 		/* Driver only uses 32-bit chain elements */
2030 		if (len != MPS_SGC_SIZE)
2031 			panic("CHAIN %p length %u or %zu?", sgep,
2032 			    MPS_SGC_SIZE, len);
2033 		break;
2034 	case MPI2_SGE_FLAGS_SIMPLE_ELEMENT:
2035 		/* Driver only uses 64-bit SGE simple elements */
2036 		if (len != MPS_SGE64_SIZE)
2037 			panic("SGE simple %p length %u or %zu?", sge,
2038 			    MPS_SGE64_SIZE, len);
2039 		if (((le32toh(sge->FlagsLength) >> MPI2_SGE_FLAGS_SHIFT) &
2040 		    MPI2_SGE_FLAGS_ADDRESS_SIZE) == 0)
2041 			panic("SGE simple %p not marked 64-bit?", sge);
2042 
2043 		break;
2044 	default:
2045 		panic("Unexpected SGE %p, flags %02x", tc, tc->Flags);
2046 	}
2047 #endif
2048 
2049 	/*
2050 	 * case 1: 1 more segment, enough room for it
2051 	 * case 2: 2 more segments, enough room for both
2052 	 * case 3: >=2 more segments, only enough room for 1 and a chain
2053 	 * case 4: >=1 more segment, enough room for only a chain
2054 	 * case 5: >=1 more segment, no room for anything (error)
2055          */
2056 
2057 	/*
2058 	 * There should be room for at least a chain element, or this
2059 	 * code is buggy.  Case (5).
2060 	 */
2061 	if (cm->cm_sglsize < MPS_SGC_SIZE)
2062 		panic("MPS: Need SGE Error Code\n");
2063 
2064 	if (segsleft >= 2 &&
2065 	    cm->cm_sglsize < len + MPS_SGC_SIZE + MPS_SGE64_SIZE) {
2066 		/*
2067 		 * There are 2 or more segments left to add, and only
2068 		 * enough room for 1 and a chain.  Case (3).
2069 		 *
2070 		 * Mark as last element in this chain if necessary.
2071 		 */
2072 		if (type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2073 			sge->FlagsLength |= htole32(
2074 			    MPI2_SGE_FLAGS_LAST_ELEMENT << MPI2_SGE_FLAGS_SHIFT);
2075 		}
2076 
2077 		/*
2078 		 * Add the item then a chain.  Do the chain now,
2079 		 * rather than on the next iteration, to simplify
2080 		 * understanding the code.
2081 		 */
2082 		cm->cm_sglsize -= len;
2083 		bcopy(sgep, cm->cm_sge, len);
2084 		cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2085 		return (mps_add_chain(cm));
2086 	}
2087 
2088 	if (segsleft >= 1 && cm->cm_sglsize < len + MPS_SGC_SIZE) {
2089 		/*
2090 		 * 1 or more segment, enough room for only a chain.
2091 		 * Hope the previous element wasn't a Simple entry
2092 		 * that needed to be marked with
2093 		 * MPI2_SGE_FLAGS_LAST_ELEMENT.  Case (4).
2094 		 */
2095 		if ((error = mps_add_chain(cm)) != 0)
2096 			return (error);
2097 	}
2098 
2099 #ifdef INVARIANTS
2100 	/* Case 1: 1 more segment, enough room for it. */
2101 	if (segsleft == 1 && cm->cm_sglsize < len)
2102 		panic("1 seg left and no room? %u versus %zu",
2103 		    cm->cm_sglsize, len);
2104 
2105 	/* Case 2: 2 more segments, enough room for both */
2106 	if (segsleft == 2 && cm->cm_sglsize < len + MPS_SGE64_SIZE)
2107 		panic("2 segs left and no room? %u versus %zu",
2108 		    cm->cm_sglsize, len);
2109 #endif
2110 
2111 	if (segsleft == 1 && type == MPI2_SGE_FLAGS_SIMPLE_ELEMENT) {
2112 		/*
2113 		 * If this is a bi-directional request, need to account for that
2114 		 * here.  Save the pre-filled sge values.  These will be used
2115 		 * either for the 2nd SGL or for a single direction SGL.  If
2116 		 * cm_out_len is non-zero, this is a bi-directional request, so
2117 		 * fill in the OUT SGL first, then the IN SGL, otherwise just
2118 		 * fill in the IN SGL.  Note that at this time, when filling in
2119 		 * 2 SGL's for a bi-directional request, they both use the same
2120 		 * DMA buffer (same cm command).
2121 		 */
2122 		saved_buf_len = le32toh(sge->FlagsLength) & 0x00FFFFFF;
2123 		saved_address_low = sge->Address.Low;
2124 		saved_address_high = sge->Address.High;
2125 		if (cm->cm_out_len) {
2126 			sge->FlagsLength = htole32(cm->cm_out_len |
2127 			    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2128 			    MPI2_SGE_FLAGS_END_OF_BUFFER |
2129 			    MPI2_SGE_FLAGS_HOST_TO_IOC |
2130 			    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2131 			    MPI2_SGE_FLAGS_SHIFT));
2132 			cm->cm_sglsize -= len;
2133 			bcopy(sgep, cm->cm_sge, len);
2134 			cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge
2135 			    + len);
2136 		}
2137 		saved_buf_len |=
2138 		    ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2139 		    MPI2_SGE_FLAGS_END_OF_BUFFER |
2140 		    MPI2_SGE_FLAGS_LAST_ELEMENT |
2141 		    MPI2_SGE_FLAGS_END_OF_LIST |
2142 		    MPI2_SGE_FLAGS_64_BIT_ADDRESSING) <<
2143 		    MPI2_SGE_FLAGS_SHIFT);
2144 		if (cm->cm_flags & MPS_CM_FLAGS_DATAIN) {
2145 			saved_buf_len |=
2146 			    ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) <<
2147 			    MPI2_SGE_FLAGS_SHIFT);
2148 		} else {
2149 			saved_buf_len |=
2150 			    ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) <<
2151 			    MPI2_SGE_FLAGS_SHIFT);
2152 		}
2153 		sge->FlagsLength = htole32(saved_buf_len);
2154 		sge->Address.Low = saved_address_low;
2155 		sge->Address.High = saved_address_high;
2156 	}
2157 
2158 	cm->cm_sglsize -= len;
2159 	bcopy(sgep, cm->cm_sge, len);
2160 	cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len);
2161 	return (0);
2162 }
2163 
2164 /*
2165  * Add one dma segment to the scatter-gather list for a command.
2166  */
2167 int
2168 mps_add_dmaseg(struct mps_command *cm, vm_paddr_t pa, size_t len, u_int flags,
2169     int segsleft)
2170 {
2171 	MPI2_SGE_SIMPLE64 sge;
2172 
2173 	/*
2174 	 * This driver always uses 64-bit address elements for simplicity.
2175 	 */
2176 	bzero(&sge, sizeof(sge));
2177 	flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT |
2178 	    MPI2_SGE_FLAGS_64_BIT_ADDRESSING;
2179 	sge.FlagsLength = htole32(len | (flags << MPI2_SGE_FLAGS_SHIFT));
2180 	mps_from_u64(pa, &sge.Address);
2181 
2182 	return (mps_push_sge(cm, &sge, sizeof sge, segsleft));
2183 }
2184 
2185 static void
2186 mps_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
2187 {
2188 	struct mps_softc *sc;
2189 	struct mps_command *cm;
2190 	u_int i, dir, sflags;
2191 
2192 	cm = (struct mps_command *)arg;
2193 	sc = cm->cm_sc;
2194 
2195 	/*
2196 	 * In this case, just print out a warning and let the chip tell the
2197 	 * user they did the wrong thing.
2198 	 */
2199 	if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) {
2200 		mps_printf(sc, "%s: warning: busdma returned %d segments, "
2201 			   "more than the %d allowed\n", __func__, nsegs,
2202 			   cm->cm_max_segs);
2203 	}
2204 
2205 	/*
2206 	 * Set up DMA direction flags.  Bi-directional requests are also handled
2207 	 * here.  In that case, both direction flags will be set.
2208 	 */
2209 	sflags = 0;
2210 	if (cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) {
2211 		/*
2212 		 * We have to add a special case for SMP passthrough, there
2213 		 * is no easy way to generically handle it.  The first
2214 		 * S/G element is used for the command (therefore the
2215 		 * direction bit needs to be set).  The second one is used
2216 		 * for the reply.  We'll leave it to the caller to make
2217 		 * sure we only have two buffers.
2218 		 */
2219 		/*
2220 		 * Even though the busdma man page says it doesn't make
2221 		 * sense to have both direction flags, it does in this case.
2222 		 * We have one s/g element being accessed in each direction.
2223 		 */
2224 		dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD;
2225 
2226 		/*
2227 		 * Set the direction flag on the first buffer in the SMP
2228 		 * passthrough request.  We'll clear it for the second one.
2229 		 */
2230 		sflags |= MPI2_SGE_FLAGS_DIRECTION |
2231 			  MPI2_SGE_FLAGS_END_OF_BUFFER;
2232 	} else if (cm->cm_flags & MPS_CM_FLAGS_DATAOUT) {
2233 		sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC;
2234 		dir = BUS_DMASYNC_PREWRITE;
2235 	} else
2236 		dir = BUS_DMASYNC_PREREAD;
2237 
2238 	for (i = 0; i < nsegs; i++) {
2239 		if ((cm->cm_flags & MPS_CM_FLAGS_SMP_PASS) && (i != 0)) {
2240 			sflags &= ~MPI2_SGE_FLAGS_DIRECTION;
2241 		}
2242 		error = mps_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len,
2243 		    sflags, nsegs - i);
2244 		if (error != 0) {
2245 			/* Resource shortage, roll back! */
2246 			mps_dprint(sc, MPS_INFO, "out of chain frames\n");
2247 			cm->cm_flags |= MPS_CM_FLAGS_CHAIN_FAILED;
2248 			mps_complete_command(cm);
2249 			return;
2250 		}
2251 	}
2252 
2253 	bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir);
2254 	mps_enqueue_request(sc, cm);
2255 
2256 	return;
2257 }
2258 
2259 static void
2260 mps_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize,
2261 	     int error)
2262 {
2263 	mps_data_cb(arg, segs, nsegs, error);
2264 }
2265 
2266 /*
2267  * This is the routine to enqueue commands ansynchronously.
2268  * Note that the only error path here is from bus_dmamap_load(), which can
2269  * return EINPROGRESS if it is waiting for resources.  Other than this, it's
2270  * assumed that if you have a command in-hand, then you have enough credits
2271  * to use it.
2272  */
2273 int
2274 mps_map_command(struct mps_softc *sc, struct mps_command *cm)
2275 {
2276 	int error = 0;
2277 
2278 	if (cm->cm_flags & MPS_CM_FLAGS_USE_UIO) {
2279 		error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap,
2280 		    &cm->cm_uio, mps_data_cb2, cm, 0);
2281 	} else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) {
2282 		error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap,
2283 		    cm->cm_data, cm->cm_length, mps_data_cb, cm, 0);
2284 	} else {
2285 		/* Add a zero-length element as needed */
2286 		if (cm->cm_sge != NULL)
2287 			mps_add_dmaseg(cm, 0, 0, 0, 1);
2288 		mps_enqueue_request(sc, cm);
2289 	}
2290 
2291 	return (error);
2292 }
2293 
2294 /*
2295  * This is the routine to enqueue commands synchronously.  An error of
2296  * EINPROGRESS from mps_map_command() is ignored since the command will
2297  * be executed and enqueued automatically.  Other errors come from msleep().
2298  */
2299 int
2300 mps_wait_command(struct mps_softc *sc, struct mps_command *cm, int timeout)
2301 {
2302 	int error, rc;
2303 
2304 	mtx_assert(&sc->mps_mtx, MA_OWNED);
2305 
2306 	if(sc->mps_flags & MPS_FLAGS_DIAGRESET)
2307 		return  EBUSY;
2308 
2309 	cm->cm_complete = NULL;
2310 	cm->cm_flags |= MPS_CM_FLAGS_WAKEUP;
2311 	error = mps_map_command(sc, cm);
2312 	if ((error != 0) && (error != EINPROGRESS))
2313 		return (error);
2314 	error = msleep(cm, &sc->mps_mtx, 0, "mpswait", timeout*hz);
2315 	if (error == EWOULDBLOCK) {
2316 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s\n", __func__);
2317 		rc = mps_reinit(sc);
2318 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n",
2319 				(rc == 0) ? "success" : "failed");
2320 		error = ETIMEDOUT;
2321 	}
2322 	return (error);
2323 }
2324 
2325 /*
2326  * This is the routine to enqueue a command synchonously and poll for
2327  * completion.  Its use should be rare.
2328  */
2329 int
2330 mps_request_polled(struct mps_softc *sc, struct mps_command *cm)
2331 {
2332 	int error, timeout = 0, rc;
2333 
2334 	error = 0;
2335 
2336 	cm->cm_flags |= MPS_CM_FLAGS_POLLED;
2337 	cm->cm_complete = NULL;
2338 	mps_map_command(sc, cm);
2339 
2340 	while ((cm->cm_flags & MPS_CM_FLAGS_COMPLETE) == 0) {
2341 		mps_intr_locked(sc);
2342 
2343 		DELAY(50 * 1000);
2344 		if (timeout++ > 1000) {
2345 			mps_dprint(sc, MPS_FAULT, "polling failed\n");
2346 			error = ETIMEDOUT;
2347 			break;
2348 		}
2349 	}
2350 
2351 	if (error) {
2352 		mps_dprint(sc, MPS_FAULT, "Calling Reinit from %s\n", __func__);
2353 		rc = mps_reinit(sc);
2354 		mps_dprint(sc, MPS_FAULT, "Reinit %s\n",
2355 				(rc == 0) ? "success" : "failed");
2356 	}
2357 
2358 	return (error);
2359 }
2360 
2361 /*
2362  * The MPT driver had a verbose interface for config pages.  In this driver,
2363  * reduce it to much simplier terms, similar to the Linux driver.
2364  */
2365 int
2366 mps_read_config_page(struct mps_softc *sc, struct mps_config_params *params)
2367 {
2368 	MPI2_CONFIG_REQUEST *req;
2369 	struct mps_command *cm;
2370 	int error;
2371 
2372 	if (sc->mps_flags & MPS_FLAGS_BUSY) {
2373 		return (EBUSY);
2374 	}
2375 
2376 	cm = mps_alloc_command(sc);
2377 	if (cm == NULL) {
2378 		return (EBUSY);
2379 	}
2380 
2381 	req = (MPI2_CONFIG_REQUEST *)cm->cm_req;
2382 	req->Function = MPI2_FUNCTION_CONFIG;
2383 	req->Action = params->action;
2384 	req->SGLFlags = 0;
2385 	req->ChainOffset = 0;
2386 	req->PageAddress = params->page_address;
2387 	if (params->hdr.Ext.ExtPageType != 0) {
2388 		MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr;
2389 
2390 		hdr = &params->hdr.Ext;
2391 		req->ExtPageType = hdr->ExtPageType;
2392 		req->ExtPageLength = hdr->ExtPageLength;
2393 		req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED;
2394 		req->Header.PageLength = 0; /* Must be set to zero */
2395 		req->Header.PageNumber = hdr->PageNumber;
2396 		req->Header.PageVersion = hdr->PageVersion;
2397 	} else {
2398 		MPI2_CONFIG_PAGE_HEADER *hdr;
2399 
2400 		hdr = &params->hdr.Struct;
2401 		req->Header.PageType = hdr->PageType;
2402 		req->Header.PageNumber = hdr->PageNumber;
2403 		req->Header.PageLength = hdr->PageLength;
2404 		req->Header.PageVersion = hdr->PageVersion;
2405 	}
2406 
2407 	cm->cm_data = params->buffer;
2408 	cm->cm_length = params->length;
2409 	cm->cm_sge = &req->PageBufferSGE;
2410 	cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION);
2411 	cm->cm_flags = MPS_CM_FLAGS_SGE_SIMPLE | MPS_CM_FLAGS_DATAIN;
2412 	cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE;
2413 
2414 	cm->cm_complete_data = params;
2415 	if (params->callback != NULL) {
2416 		cm->cm_complete = mps_config_complete;
2417 		return (mps_map_command(sc, cm));
2418 	} else {
2419 		error = mps_wait_command(sc, cm, 0);
2420 		if (error) {
2421 			mps_dprint(sc, MPS_FAULT,
2422 			    "Error %d reading config page\n", error);
2423 			mps_free_command(sc, cm);
2424 			return (error);
2425 		}
2426 		mps_config_complete(sc, cm);
2427 	}
2428 
2429 	return (0);
2430 }
2431 
2432 int
2433 mps_write_config_page(struct mps_softc *sc, struct mps_config_params *params)
2434 {
2435 	return (EINVAL);
2436 }
2437 
2438 static void
2439 mps_config_complete(struct mps_softc *sc, struct mps_command *cm)
2440 {
2441 	MPI2_CONFIG_REPLY *reply;
2442 	struct mps_config_params *params;
2443 
2444 	params = cm->cm_complete_data;
2445 
2446 	if (cm->cm_data != NULL) {
2447 		bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap,
2448 		    BUS_DMASYNC_POSTREAD);
2449 		bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap);
2450 	}
2451 
2452 	/*
2453 	 * XXX KDM need to do more error recovery?  This results in the
2454 	 * device in question not getting probed.
2455 	 */
2456 	if ((cm->cm_flags & MPS_CM_FLAGS_ERROR_MASK) != 0) {
2457 		params->status = MPI2_IOCSTATUS_BUSY;
2458 		goto done;
2459 	}
2460 
2461 	reply = (MPI2_CONFIG_REPLY *)cm->cm_reply;
2462 	if (reply == NULL) {
2463 		params->status = MPI2_IOCSTATUS_BUSY;
2464 		goto done;
2465 	}
2466 	params->status = reply->IOCStatus;
2467 	if (params->hdr.Ext.ExtPageType != 0) {
2468 		params->hdr.Ext.ExtPageType = reply->ExtPageType;
2469 		params->hdr.Ext.ExtPageLength = reply->ExtPageLength;
2470 	} else {
2471 		params->hdr.Struct.PageType = reply->Header.PageType;
2472 		params->hdr.Struct.PageNumber = reply->Header.PageNumber;
2473 		params->hdr.Struct.PageLength = reply->Header.PageLength;
2474 		params->hdr.Struct.PageVersion = reply->Header.PageVersion;
2475 	}
2476 
2477 done:
2478 	mps_free_command(sc, cm);
2479 	if (params->callback != NULL)
2480 		params->callback(sc, params);
2481 
2482 	return;
2483 }
2484