xref: /netbsd-src/sys/arch/sparc64/dev/iommu.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: iommu.c,v 1.37 2001/08/06 22:02:58 eeh Exp $	*/
2 
3 /*
4  * Copyright (c) 1999, 2000 Matthew R. Green
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
23  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
25  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
26  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 /*-
32  * Copyright (c) 1998 The NetBSD Foundation, Inc.
33  * All rights reserved.
34  *
35  * This code is derived from software contributed to The NetBSD Foundation
36  * by Paul Kranenburg.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *        This product includes software developed by the NetBSD
49  *        Foundation, Inc. and its contributors.
50  * 4. Neither the name of The NetBSD Foundation nor the names of its
51  *    contributors may be used to endorse or promote products derived
52  *    from this software without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
55  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
56  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
57  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
58  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
59  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
60  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
61  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
62  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
63  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
64  * POSSIBILITY OF SUCH DAMAGE.
65  */
66 
67 /*
68  * Copyright (c) 1992, 1993
69  *	The Regents of the University of California.  All rights reserved.
70  *
71  * This software was developed by the Computer Systems Engineering group
72  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
73  * contributed to Berkeley.
74  *
75  * All advertising materials mentioning features or use of this software
76  * must display the following acknowledgement:
77  *	This product includes software developed by the University of
78  *	California, Lawrence Berkeley Laboratory.
79  *
80  * Redistribution and use in source and binary forms, with or without
81  * modification, are permitted provided that the following conditions
82  * are met:
83  * 1. Redistributions of source code must retain the above copyright
84  *    notice, this list of conditions and the following disclaimer.
85  * 2. Redistributions in binary form must reproduce the above copyright
86  *    notice, this list of conditions and the following disclaimer in the
87  *    documentation and/or other materials provided with the distribution.
88  * 3. All advertising materials mentioning features or use of this software
89  *    must display the following acknowledgement:
90  *	This product includes software developed by the University of
91  *	California, Berkeley and its contributors.
92  * 4. Neither the name of the University nor the names of its contributors
93  *    may be used to endorse or promote products derived from this software
94  *    without specific prior written permission.
95  *
96  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
97  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
98  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
99  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
100  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
101  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
102  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
103  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
104  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
105  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
106  * SUCH DAMAGE.
107  *
108  *	from: NetBSD: sbus.c,v 1.13 1999/05/23 07:24:02 mrg Exp
109  *	from: @(#)sbus.c	8.1 (Berkeley) 6/11/93
110  */
111 
112 /*
113  * UltraSPARC IOMMU support; used by both the sbus and pci code.
114  */
115 #include "opt_ddb.h"
116 
117 #include <sys/param.h>
118 #include <sys/extent.h>
119 #include <sys/malloc.h>
120 #include <sys/systm.h>
121 #include <sys/device.h>
122 
123 #include <uvm/uvm_extern.h>
124 
125 #include <machine/bus.h>
126 #include <sparc64/sparc64/cache.h>
127 #include <sparc64/dev/iommureg.h>
128 #include <sparc64/dev/iommuvar.h>
129 
130 #include <machine/autoconf.h>
131 #include <machine/cpu.h>
132 
133 #ifdef DEBUG
134 #define IDB_BUSDMA	0x1
135 #define IDB_IOMMU	0x2
136 #define IDB_INFO	0x4
137 #define	IDB_SYNC	0x8
138 int iommudebug = 0x0;
139 #define DPRINTF(l, s)   do { if (iommudebug & l) printf s; } while (0)
140 #else
141 #define DPRINTF(l, s)
142 #endif
143 
144 #define iommu_strbuf_flush(i,v) bus_space_write_8((i)->is_bustag, \
145 	(bus_space_handle_t)(u_long)&(i)->is_sb->strbuf_pgflush, 0, (v))
146 static	int iommu_strbuf_flush_done __P((struct iommu_state *));
147 
148 /*
149  * initialise the UltraSPARC IOMMU (SBUS or PCI):
150  *	- allocate and setup the iotsb.
151  *	- enable the IOMMU
152  *	- initialise the streaming buffers (if they exist)
153  *	- create a private DVMA map.
154  */
155 void
156 iommu_init(name, is, tsbsize, iovabase)
157 	char *name;
158 	struct iommu_state *is;
159 	int tsbsize;
160 	u_int32_t iovabase;
161 {
162 	psize_t size;
163 	vaddr_t va;
164 	paddr_t pa;
165 	struct vm_page *m;
166 	struct pglist mlist;
167 
168 	/*
169 	 * Setup the iommu.
170 	 *
171 	 * The sun4u iommu is part of the SBUS or PCI controller so we
172 	 * will deal with it here..
173 	 *
174 	 * The IOMMU address space always ends at 0xffffe000, but the starting
175 	 * address depends on the size of the map.  The map size is 1024 * 2 ^
176 	 * is->is_tsbsize entries, where each entry is 8 bytes.  The start of
177 	 * the map can be calculated by (0xffffe000 << (8 + is->is_tsbsize)).
178 	 */
179 	is->is_cr = (tsbsize << 16) | IOMMUCR_EN;
180 	is->is_tsbsize = tsbsize;
181 	is->is_dvmabase = iovabase;
182 	if (iovabase == -1) is->is_dvmabase = IOTSB_VSTART(is->is_tsbsize);
183 
184 	/*
185 	 * Allocate memory for I/O pagetables.  They need to be physically
186 	 * contiguous.
187 	 */
188 
189 	size = NBPG<<(is->is_tsbsize);
190 	TAILQ_INIT(&mlist);
191 	if (uvm_pglistalloc((psize_t)size, (paddr_t)0, (paddr_t)-1,
192 		(paddr_t)NBPG, (paddr_t)0, &mlist, 1, 0) != 0)
193 		panic("iommu_init: no memory");
194 
195 	va = uvm_km_valloc(kernel_map, size);
196 	if (va == 0)
197 		panic("iommu_init: no memory");
198 	is->is_tsb = (int64_t *)va;
199 
200 	m = TAILQ_FIRST(&mlist);
201 	is->is_ptsb = VM_PAGE_TO_PHYS(m);
202 
203 	/* Map the pages */
204 	for (; m != NULL; m = TAILQ_NEXT(m,pageq)) {
205 		pa = VM_PAGE_TO_PHYS(m);
206 		pmap_enter(pmap_kernel(), va, pa | PMAP_NVC,
207 			VM_PROT_READ|VM_PROT_WRITE,
208 			VM_PROT_READ|VM_PROT_WRITE|PMAP_WIRED);
209 		va += NBPG;
210 	}
211 	pmap_update();
212 	bzero(is->is_tsb, size);
213 
214 #ifdef DEBUG
215 	if (iommudebug & IDB_INFO)
216 	{
217 		/* Probe the iommu */
218 		struct iommureg *regs = is->is_iommu;
219 
220 		printf("iommu regs at: cr=%lx tsb=%lx flush=%lx\n",
221 		    (u_long)&regs->iommu_cr,
222 		    (u_long)&regs->iommu_tsb,
223 		    (u_long)&regs->iommu_flush);
224 		printf("iommu cr=%llx tsb=%llx\n", (unsigned long long)regs->iommu_cr, (unsigned long long)regs->iommu_tsb);
225 		printf("TSB base %p phys %llx\n", (void *)is->is_tsb, (unsigned long long)is->is_ptsb);
226 		delay(1000000); /* 1 s */
227 	}
228 #endif
229 
230 	/*
231 	 * Initialize streaming buffer, if it is there.
232 	 */
233 	if (is->is_sb)
234 		(void)pmap_extract(pmap_kernel(), (vaddr_t)&is->is_flush,
235 		    (paddr_t *)&is->is_flushpa);
236 
237 	/*
238 	 * now actually start up the IOMMU
239 	 */
240 	iommu_reset(is);
241 
242 	/*
243 	 * Now all the hardware's working we need to allocate a dvma map.
244 	 */
245 	printf("DVMA map: %x to %x\n",
246 		(unsigned int)is->is_dvmabase,
247 		(unsigned int)(is->is_dvmabase+(size<<10)));
248 	is->is_dvmamap = extent_create(name,
249 				       is->is_dvmabase, (u_long)IOTSB_VEND,
250 				       M_DEVBUF, 0, 0, EX_NOWAIT);
251 }
252 
253 /*
254  * Streaming buffers don't exist on the UltraSPARC IIi; we should have
255  * detected that already and disabled them.  If not, we will notice that
256  * they aren't there when the STRBUF_EN bit does not remain.
257  */
258 void
259 iommu_reset(is)
260 	struct iommu_state *is;
261 {
262 
263 	/* Need to do 64-bit stores */
264 	bus_space_write_8(is->is_bustag,
265 			  (bus_space_handle_t)(u_long)&is->is_iommu->iommu_tsb,
266 			  0, is->is_ptsb);
267 	/* Enable IOMMU in diagnostic mode */
268 	bus_space_write_8(is->is_bustag,
269 			  (bus_space_handle_t)(u_long)&is->is_iommu->iommu_cr, 0,
270 			  is->is_cr|IOMMUCR_DE);
271 
272 
273 	if (!is->is_sb)
274 		return;
275 
276 	/* Enable diagnostics mode? */
277 	bus_space_write_8(is->is_bustag,
278 			  (bus_space_handle_t)(u_long)&is->is_sb->strbuf_ctl,
279 			  0, STRBUF_EN);
280 
281 	/* No streaming buffers? Disable them */
282 	if (bus_space_read_8(is->is_bustag,
283 			     (bus_space_handle_t)(u_long)&is->is_sb->strbuf_ctl,
284 			     0) == 0)
285 		is->is_sb = 0;
286 }
287 
288 /*
289  * Here are the iommu control routines.
290  */
291 void
292 iommu_enter(is, va, pa, flags)
293 	struct iommu_state *is;
294 	vaddr_t va;
295 	int64_t pa;
296 	int flags;
297 {
298 	int64_t tte;
299 
300 #ifdef DIAGNOSTIC
301 	if (va < is->is_dvmabase)
302 		panic("iommu_enter: va %#lx not in DVMA space", va);
303 #endif
304 
305 	tte = MAKEIOTTE(pa, !(flags&BUS_DMA_NOWRITE), !(flags&BUS_DMA_NOCACHE),
306 			(flags&BUS_DMA_STREAMING));
307 
308 	/* Is the streamcache flush really needed? */
309 	if (is->is_sb) {
310 		iommu_strbuf_flush(is, va);
311 		iommu_strbuf_flush_done(is);
312 	}
313 	DPRINTF(IDB_IOMMU, ("Clearing TSB slot %d for va %p\n",
314 		       (int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va));
315 	is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] = tte;
316 	bus_space_write_8(is->is_bustag, (bus_space_handle_t)(u_long)
317 			  &is->is_iommu->iommu_flush, 0, va);
318 	DPRINTF(IDB_IOMMU, ("iommu_enter: va %lx pa %lx TSB[%lx]@%p=%lx\n",
319 		       va, (long)pa, (u_long)IOTSBSLOT(va,is->is_tsbsize),
320 		       (void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
321 		       (u_long)tte));
322 }
323 
324 /*
325  * iommu_remove: removes mappings created by iommu_enter
326  *
327  * Only demap from IOMMU if flag is set.
328  *
329  * XXX: this function needs better internal error checking.
330  */
331 void
332 iommu_remove(is, va, len)
333 	struct iommu_state *is;
334 	vaddr_t va;
335 	size_t len;
336 {
337 
338 #ifdef DIAGNOSTIC
339 	if (va < is->is_dvmabase)
340 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
341 	if ((long)(va + len) < (long)va)
342 		panic("iommu_remove: va 0x%lx + len 0x%lx wraps",
343 		      (long) va, (long) len);
344 	if (len & ~0xfffffff)
345 		panic("iommu_remove: rediculous len 0x%lx", (u_long)len);
346 #endif
347 
348 	va = trunc_page(va);
349 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx TSB[%lx]@%p\n",
350 	    va, (u_long)IOTSBSLOT(va,is->is_tsbsize),
351 	    &is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)]));
352 	while (len > 0) {
353 		DPRINTF(IDB_IOMMU, ("iommu_remove: clearing TSB slot %d for va %p size %lx\n",
354 		    (int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va, (u_long)len));
355 		if (is->is_sb) {
356 			DPRINTF(IDB_IOMMU, ("iommu_remove: flushing va %p TSB[%lx]@%p=%lx, %lu bytes left\n",
357 			       (void *)(u_long)va, (long)IOTSBSLOT(va,is->is_tsbsize),
358 			       (void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
359 			       (long)(is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)]),
360 			       (u_long)len));
361 			iommu_strbuf_flush(is, va);
362 			if (len <= NBPG)
363 				iommu_strbuf_flush_done(is);
364 			DPRINTF(IDB_IOMMU, ("iommu_remove: flushed va %p TSB[%lx]@%p=%lx, %lu bytes left\n",
365 			       (void *)(u_long)va, (long)IOTSBSLOT(va,is->is_tsbsize),
366 			       (void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
367 			       (long)(is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)]),
368 			       (u_long)len));
369 		}
370 
371 		if (len <= NBPG)
372 			len = 0;
373 		else
374 			len -= NBPG;
375 
376 		is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] = 0;
377 		bus_space_write_8(is->is_bustag, (bus_space_handle_t)(u_long)
378 				  &is->is_iommu->iommu_flush, 0, va);
379 		va += NBPG;
380 	}
381 }
382 
383 static int
384 iommu_strbuf_flush_done(is)
385 	struct iommu_state *is;
386 {
387 	struct timeval cur, flushtimeout;
388 
389 #define BUMPTIME(t, usec) { \
390 	register volatile struct timeval *tp = (t); \
391 	register long us; \
392  \
393 	tp->tv_usec = us = tp->tv_usec + (usec); \
394 	if (us >= 1000000) { \
395 		tp->tv_usec = us - 1000000; \
396 		tp->tv_sec++; \
397 	} \
398 }
399 
400 	if (!is->is_sb)
401 		return (0);
402 
403 	/*
404 	 * Streaming buffer flushes:
405 	 *
406 	 *   1 Tell strbuf to flush by storing va to strbuf_pgflush.  If
407 	 *     we're not on a cache line boundary (64-bits):
408 	 *   2 Store 0 in flag
409 	 *   3 Store pointer to flag in flushsync
410 	 *   4 wait till flushsync becomes 0x1
411 	 *
412 	 * If it takes more than .5 sec, something
413 	 * went wrong.
414 	 */
415 
416 	is->is_flush = 0;
417 	membar_sync();	/* #StoreStore is prolly enuf. */
418 	bus_space_write_8(is->is_bustag, (bus_space_handle_t)(u_long)
419 			  &is->is_sb->strbuf_flushsync, 0, is->is_flushpa);
420 	membar_sync();	/* Prolly not needed at all. */
421 
422 	microtime(&flushtimeout);
423 	cur = flushtimeout;
424 	BUMPTIME(&flushtimeout, 500000); /* 1/2 sec */
425 
426 	DPRINTF(IDB_IOMMU, ("iommu_strbuf_flush_done: flush = %lx at va = %lx pa = %lx now=%lx:%lx until = %lx:%lx\n",
427 		       (long)is->is_flush, (long)&is->is_flush,
428 		       (long)is->is_flushpa, cur.tv_sec, cur.tv_usec,
429 		       flushtimeout.tv_sec, flushtimeout.tv_usec));
430 	/* Bypass non-coherent D$ */
431 	while (!ldxa(is->is_flushpa, ASI_PHYS_CACHED) &&
432 	       ((cur.tv_sec <= flushtimeout.tv_sec) &&
433 		(cur.tv_usec <= flushtimeout.tv_usec)))
434 		microtime(&cur);
435 
436 #ifdef DIAGNOSTIC
437 	if (!ldxa(is->is_flushpa, ASI_PHYS_CACHED)) {
438 		printf("iommu_strbuf_flush_done: flush timeout %p at %p\n",
439 		    (void *)(u_long)is->is_flush,
440 		    (void *)(u_long)is->is_flushpa); /* panic? */
441 #ifdef DDB
442 		Debugger();
443 #endif
444 	}
445 #endif
446 	DPRINTF(IDB_IOMMU, ("iommu_strbuf_flush_done: flushed\n"));
447 	return (is->is_flush);
448 }
449 
450 /*
451  * IOMMU DVMA operations, common to SBUS and PCI.
452  */
453 int
454 iommu_dvmamap_load(t, is, map, buf, buflen, p, flags)
455 	bus_dma_tag_t t;
456 	struct iommu_state *is;
457 	bus_dmamap_t map;
458 	void *buf;
459 	bus_size_t buflen;
460 	struct proc *p;
461 	int flags;
462 {
463 	int s;
464 	int err;
465 	bus_size_t sgsize;
466 	paddr_t curaddr;
467 	u_long dvmaddr;
468 	bus_size_t align, boundary;
469 	vaddr_t vaddr = (vaddr_t)buf;
470 	pmap_t pmap;
471 
472 	if (map->dm_nsegs) {
473 		/* Already in use?? */
474 #ifdef DIAGNOSTIC
475 		printf("iommu_dvmamap_load: map still in use\n");
476 #endif
477 		bus_dmamap_unload(t, map);
478 	}
479 	/*
480 	 * Make sure that on error condition we return "no valid mappings".
481 	 */
482 	map->dm_nsegs = 0;
483 
484 	if (buflen > map->_dm_size) {
485 		DPRINTF(IDB_BUSDMA,
486 		    ("iommu_dvmamap_load(): error %d > %d -- "
487 		     "map size exceeded!\n", (int)buflen, (int)map->_dm_size));
488 		return (EINVAL);
489 	}
490 
491 	sgsize = round_page(buflen + ((int)vaddr & PGOFSET));
492 
493 	/*
494 	 * A boundary presented to bus_dmamem_alloc() takes precedence
495 	 * over boundary in the map.
496 	 */
497 	if ((boundary = (map->dm_segs[0]._ds_boundary)) == 0)
498 		boundary = map->_dm_boundary;
499 	align = max(map->dm_segs[0]._ds_align, NBPG);
500 	s = splhigh();
501 	err = extent_alloc(is->is_dvmamap, sgsize, align,
502 	    boundary, EX_NOWAIT|EX_BOUNDZERO, (u_long *)&dvmaddr);
503 	splx(s);
504 
505 #ifdef DEBUG
506 	if (err || (dvmaddr == (bus_addr_t)-1))
507 	{
508 		printf("iommu_dvmamap_load(): extent_alloc(%d, %x) failed!\n",
509 		    (int)sgsize, flags);
510 		Debugger();
511 	}
512 #endif
513 	if (err != 0)
514 		return (err);
515 
516 	if (dvmaddr == (bus_addr_t)-1)
517 		return (ENOMEM);
518 
519 	/*
520 	 * We always use just one segment.
521 	 */
522 	map->dm_mapsize = buflen;
523 	map->dm_nsegs = 1;
524 	map->dm_segs[0].ds_addr = dvmaddr + (vaddr & PGOFSET);
525 	map->dm_segs[0].ds_len = buflen;
526 
527 	if (p != NULL)
528 		pmap = p->p_vmspace->vm_map.pmap;
529 	else
530 		pmap = pmap_kernel();
531 
532 	dvmaddr = trunc_page(map->dm_segs[0].ds_addr);
533 	for (; buflen > 0; ) {
534 		/*
535 		 * Get the physical address for this page.
536 		 */
537 		if (pmap_extract(pmap, (vaddr_t)vaddr, &curaddr) == FALSE) {
538 			bus_dmamap_unload(t, map);
539 			return (-1);
540 		}
541 
542 		/*
543 		 * Compute the segment size, and adjust counts.
544 		 */
545 		sgsize = NBPG - ((u_long)vaddr & PGOFSET);
546 		if (buflen < sgsize)
547 			sgsize = buflen;
548 
549 		DPRINTF(IDB_BUSDMA,
550 		    ("iommu_dvmamap_load: map %p loading va %p "
551 			    "dva %lx at pa %lx\n",
552 			    map, (void *)vaddr, (long)dvmaddr,
553 			    (long)(curaddr&~(NBPG-1))));
554 		iommu_enter(is, trunc_page(dvmaddr), trunc_page(curaddr),
555 		    flags);
556 
557 		dvmaddr += PAGE_SIZE;
558 		vaddr += sgsize;
559 		buflen -= sgsize;
560 	}
561 	return (0);
562 }
563 
564 
565 void
566 iommu_dvmamap_unload(t, is, map)
567 	bus_dma_tag_t t;
568 	struct iommu_state *is;
569 	bus_dmamap_t map;
570 {
571 	vaddr_t addr, offset;
572 	size_t len;
573 	int error, s, i;
574 	bus_addr_t dvmaddr;
575 	bus_size_t sgsize;
576 	paddr_t pa;
577 
578 	dvmaddr = (map->dm_segs[0].ds_addr & ~PGOFSET);
579 	pa = 0;
580 	sgsize = 0;
581 	for (i = 0; i<map->dm_nsegs; i++) {
582 
583 		addr = trunc_page(map->dm_segs[i].ds_addr);
584 		offset = map->dm_segs[i].ds_addr & PGOFSET;
585 		len = map->dm_segs[i].ds_len;
586 		if (len == 0 || addr == 0)
587 			printf("iommu_dvmamap_unload: map = %p, i = %d, len = %d, addr = %lx\n",
588 				map, (int)i, (int)len, (unsigned long)addr);
589 
590 		DPRINTF(IDB_BUSDMA,
591 			("iommu_dvmamap_unload: map %p removing va %lx size %lx\n",
592 			 map, (long)addr, (long)len));
593 		iommu_remove(is, addr, len);
594 
595 		if (trunc_page(pa) == addr)
596 			sgsize += trunc_page(len + offset);
597 		else
598 			sgsize += round_page(len + offset);
599 		pa = addr + offset + len;
600 
601 	}
602 	/* Flush the caches */
603 	bus_dmamap_unload(t->_parent, map);
604 
605 	/* Mark the mappings as invalid. */
606 	map->dm_mapsize = 0;
607 	map->dm_nsegs = 0;
608 
609 	s = splhigh();
610 	error = extent_free(is->is_dvmamap, dvmaddr, sgsize, EX_NOWAIT);
611 	splx(s);
612 	if (error != 0)
613 		printf("warning: %qd of DVMA space lost\n", (long long)sgsize);
614 }
615 
616 
617 int
618 iommu_dvmamap_load_raw(t, is, map, segs, nsegs, flags, size)
619 	bus_dma_tag_t t;
620 	struct iommu_state *is;
621 	bus_dmamap_t map;
622 	bus_dma_segment_t *segs;
623 	int nsegs;
624 	int flags;
625 	bus_size_t size;
626 {
627 	struct vm_page *m;
628 	int i, s;
629 	int left;
630 	int err;
631 	bus_size_t sgsize;
632 	paddr_t pa;
633 	bus_size_t boundary, align;
634 	u_long dvmaddr;
635 	struct pglist *mlist;
636 	int pagesz = PAGE_SIZE;
637 
638 	if (map->dm_nsegs) {
639 		/* Already in use?? */
640 #ifdef DIAGNOSTIC
641 		printf("iommu_dvmamap_load_raw: map still in use\n");
642 #endif
643 		bus_dmamap_unload(t, map);
644 	}
645 	/*
646 	 * Make sure that on error condition we return "no valid mappings".
647 	 */
648 	map->dm_nsegs = 0;
649 	/* Count up the total number of pages we need */
650 	pa = segs[0].ds_addr;
651 	sgsize = 0;
652 	for (i=0; i<nsegs; i++) {
653 		if (round_page(pa) != round_page(segs[i].ds_addr))
654 			sgsize = round_page(sgsize);
655 		sgsize += segs[i].ds_len;
656 		pa = segs[i].ds_addr + segs[i].ds_len;
657 	}
658 	sgsize = round_page(sgsize);
659 
660 	/*
661 	 * A boundary presented to bus_dmamem_alloc() takes precedence
662 	 * over boundary in the map.
663 	 */
664 	if ((boundary = segs[0]._ds_boundary) == 0)
665 		boundary = map->_dm_boundary;
666 
667 	align = max(segs[0]._ds_align, NBPG);
668 	s = splhigh();
669 	err = extent_alloc(is->is_dvmamap, sgsize, align, boundary,
670 	   ((flags & BUS_DMA_NOWAIT) == 0 ? EX_WAITOK : EX_NOWAIT)|EX_BOUNDZERO,
671 	    (u_long *)&dvmaddr);
672 	splx(s);
673 
674 	if (err != 0)
675 		return (err);
676 
677 #ifdef DEBUG
678 	if (dvmaddr == (bus_addr_t)-1)
679 	{
680 		printf("iommu_dvmamap_load_raw(): extent_alloc(%d, %x) failed!\n",
681 		    (int)sgsize, flags);
682 		Debugger();
683 	}
684 #endif
685 	if (dvmaddr == (bus_addr_t)-1)
686 		return (ENOMEM);
687 
688 	if ((mlist = segs[0]._ds_mlist) == NULL) {
689 		u_long prev_va = NULL;
690 		/*
691 		 * This segs is made up of individual physical pages,
692 		 * probably by _bus_dmamap_load_uio() or
693 		 * _bus_dmamap_load_mbuf().  Ignore the mlist and
694 		 * load each segment individually.
695 		 */
696 
697 		/* We'll never end up with less segments than we got as input.
698 		   this gives us a chance to fail quickly */
699 		if (nsegs > map->_dm_segcnt)
700 			return (E2BIG);
701 
702 		i = 0;
703 		dvmaddr += (segs[i].ds_addr & PGOFSET);
704 		map->dm_segs[i].ds_addr = dvmaddr;
705 		map->dm_segs[i].ds_len = left = segs[i].ds_len;
706 		pa = segs[i].ds_addr;
707 
708 		while (left > 0) {
709 			int incr;
710 
711 			if (sgsize == 0)
712 				panic("iommu_dmamap_load_raw: size botch");
713 			DPRINTF(IDB_BUSDMA,
714 				("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
715 				 map, (long)dvmaddr, (long)(pa)));
716 			/* Enter it if we haven't before. */
717 			if (prev_va != trunc_page(dvmaddr))
718 				iommu_enter(is, prev_va = trunc_page(dvmaddr),
719 					    trunc_page(pa), flags);
720 			incr = min(pagesz, left);
721 			dvmaddr += incr;
722 			pa += incr;
723 			left -= incr;
724 
725 			/* Next segment */
726 			if (left <= 0 && ++i < nsegs) {
727 				u_long offset;
728 
729 				/*
730 				 * If the two segs are on different physical pages
731 				 * move to a new virtual page.
732 				 */
733 				offset = (segs[i].ds_addr & PGOFSET);
734 				if (trunc_page(pa) != trunc_page(segs[i].ds_addr))
735 					dvmaddr += NBPG;
736 
737 				pa = segs[i].ds_addr;
738 				dvmaddr = trunc_page(dvmaddr) + offset;
739 
740 				map->dm_segs[i].ds_len = left = segs[i].ds_len;
741 				map->dm_segs[i].ds_addr = dvmaddr;
742 			}
743 		}
744 		map->dm_nsegs = i;
745 
746 		/* bail out if we created more segments than the dmamap is
747 		   allowed to carry */
748 		if (i > map->_dm_segcnt) {
749 			iommu_dvmamap_unload(t, is, map);
750 			return (E2BIG);
751 		}
752 
753 		return (0);
754 	}
755 	/*
756 	 * This was allocated with bus_dmamem_alloc.  We only
757 	 * have one segment, and the pages are on an `mlist'.
758 	 */
759 	map->dm_mapsize = size;
760 	i = 0;
761 	map->dm_segs[0].ds_addr = dvmaddr;
762 	map->dm_segs[0].ds_len = size;
763 
764 	for (m = TAILQ_FIRST(mlist); m != NULL; m = TAILQ_NEXT(m,pageq)) {
765 		if (sgsize == 0)
766 			panic("iommu_dmamap_load_raw: size botch");
767 		pa = VM_PAGE_TO_PHYS(m);
768 
769 		DPRINTF(IDB_BUSDMA,
770 		    ("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
771 		    map, (long)dvmaddr, (long)(pa)));
772 		iommu_enter(is, dvmaddr, pa, flags);
773 
774 		dvmaddr += pagesz;
775 		sgsize -= pagesz;
776 	}
777 	map->dm_nsegs = i;
778 	return (0);
779 }
780 
781 void
782 iommu_dvmamap_sync(t, is, map, offset, len, ops)
783 	bus_dma_tag_t t;
784 	struct iommu_state *is;
785 	bus_dmamap_t map;
786 	bus_addr_t offset;
787 	bus_size_t len;
788 	int ops;
789 {
790 	vaddr_t va = map->dm_segs[0].ds_addr + offset;
791 
792 	/*
793 	 * We only support one DMA segment; supporting more makes this code
794          * too unweildy.
795 	 */
796 
797 	if (ops & BUS_DMASYNC_PREREAD) {
798 		DPRINTF(IDB_SYNC,
799 		    ("iommu_dvmamap_sync: syncing va %p len %lu "
800 		     "BUS_DMASYNC_PREREAD\n", (void *)(u_long)va, (u_long)len));
801 
802 		/* Nothing to do */;
803 	}
804 	if (ops & BUS_DMASYNC_POSTREAD) {
805 		DPRINTF(IDB_SYNC,
806 		    ("iommu_dvmamap_sync: syncing va %p len %lu "
807 		     "BUS_DMASYNC_POSTREAD\n", (void *)(u_long)va, (u_long)len));
808 		/* if we have a streaming buffer, flush it here first */
809 		if (is->is_sb)
810 			while (len > 0) {
811 				DPRINTF(IDB_BUSDMA,
812 				    ("iommu_dvmamap_sync: flushing va %p, %lu "
813 				     "bytes left\n", (void *)(u_long)va, (u_long)len));
814 				iommu_strbuf_flush(is, va);
815 				if (len <= NBPG) {
816 					iommu_strbuf_flush_done(is);
817 					len = 0;
818 				} else
819 					len -= NBPG;
820 				va += NBPG;
821 			}
822 	}
823 	if (ops & BUS_DMASYNC_PREWRITE) {
824 		DPRINTF(IDB_SYNC,
825 		    ("iommu_dvmamap_sync: syncing va %p len %lu "
826 		     "BUS_DMASYNC_PREWRITE\n", (void *)(u_long)va, (u_long)len));
827 		/* if we have a streaming buffer, flush it here first */
828 		if (is->is_sb)
829 			while (len > 0) {
830 				DPRINTF(IDB_BUSDMA,
831 				    ("iommu_dvmamap_sync: flushing va %p, %lu "
832 				     "bytes left\n", (void *)(u_long)va, (u_long)len));
833 				iommu_strbuf_flush(is, va);
834 				if (len <= NBPG) {
835 					iommu_strbuf_flush_done(is);
836 					len = 0;
837 				} else
838 					len -= NBPG;
839 				va += NBPG;
840 			}
841 	}
842 	if (ops & BUS_DMASYNC_POSTWRITE) {
843 		DPRINTF(IDB_SYNC,
844 		    ("iommu_dvmamap_sync: syncing va %p len %lu "
845 		     "BUS_DMASYNC_POSTWRITE\n", (void *)(u_long)va, (u_long)len));
846 		/* Nothing to do */;
847 	}
848 }
849 
850 int
851 iommu_dvmamem_alloc(t, is, size, alignment, boundary, segs, nsegs, rsegs, flags)
852 	bus_dma_tag_t t;
853 	struct iommu_state *is;
854 	bus_size_t size, alignment, boundary;
855 	bus_dma_segment_t *segs;
856 	int nsegs;
857 	int *rsegs;
858 	int flags;
859 {
860 
861 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_alloc: sz %llx align %llx bound %llx "
862 	   "segp %p flags %d\n", (unsigned long long)size,
863 	   (unsigned long long)alignment, (unsigned long long)boundary,
864 	   segs, flags));
865 	return (bus_dmamem_alloc(t->_parent, size, alignment, boundary,
866 	    segs, nsegs, rsegs, flags|BUS_DMA_DVMA));
867 }
868 
869 void
870 iommu_dvmamem_free(t, is, segs, nsegs)
871 	bus_dma_tag_t t;
872 	struct iommu_state *is;
873 	bus_dma_segment_t *segs;
874 	int nsegs;
875 {
876 
877 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_free: segp %p nsegs %d\n",
878 	    segs, nsegs));
879 	bus_dmamem_free(t->_parent, segs, nsegs);
880 }
881 
882 /*
883  * Map the DVMA mappings into the kernel pmap.
884  * Check the flags to see whether we're streaming or coherent.
885  */
886 int
887 iommu_dvmamem_map(t, is, segs, nsegs, size, kvap, flags)
888 	bus_dma_tag_t t;
889 	struct iommu_state *is;
890 	bus_dma_segment_t *segs;
891 	int nsegs;
892 	size_t size;
893 	caddr_t *kvap;
894 	int flags;
895 {
896 	struct vm_page *m;
897 	vaddr_t va;
898 	bus_addr_t addr;
899 	struct pglist *mlist;
900 	int cbit;
901 
902 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: segp %p nsegs %d size %lx\n",
903 	    segs, nsegs, size));
904 
905 	/*
906 	 * Allocate some space in the kernel map, and then map these pages
907 	 * into this space.
908 	 */
909 	size = round_page(size);
910 	va = uvm_km_valloc(kernel_map, size);
911 	if (va == 0)
912 		return (ENOMEM);
913 
914 	*kvap = (caddr_t)va;
915 
916 	/*
917 	 * digest flags:
918 	 */
919 	cbit = 0;
920 	if (flags & BUS_DMA_COHERENT)	/* Disable vcache */
921 		cbit |= PMAP_NVC;
922 	if (flags & BUS_DMA_NOCACHE)	/* sideffects */
923 		cbit |= PMAP_NC;
924 
925 	/*
926 	 * Now take this and map it into the CPU.
927 	 */
928 	mlist = segs[0]._ds_mlist;
929 	for (m = mlist->tqh_first; m != NULL; m = m->pageq.tqe_next) {
930 #ifdef DIAGNOSTIC
931 		if (size == 0)
932 			panic("iommu_dvmamem_map: size botch");
933 #endif
934 		addr = VM_PAGE_TO_PHYS(m);
935 		DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: "
936 		    "mapping va %lx at %llx\n", va, (unsigned long long)addr | cbit));
937 		pmap_enter(pmap_kernel(), va, addr | cbit,
938 		    VM_PROT_READ | VM_PROT_WRITE,
939 		    VM_PROT_READ | VM_PROT_WRITE | PMAP_WIRED);
940 		va += PAGE_SIZE;
941 		size -= PAGE_SIZE;
942 	}
943 	pmap_update();
944 
945 	return (0);
946 }
947 
948 /*
949  * Unmap DVMA mappings from kernel
950  */
951 void
952 iommu_dvmamem_unmap(t, is, kva, size)
953 	bus_dma_tag_t t;
954 	struct iommu_state *is;
955 	caddr_t kva;
956 	size_t size;
957 {
958 
959 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_unmap: kvm %p size %lx\n",
960 	    kva, size));
961 
962 #ifdef DIAGNOSTIC
963 	if ((u_long)kva & PGOFSET)
964 		panic("iommu_dvmamem_unmap");
965 #endif
966 
967 	size = round_page(size);
968 	pmap_remove(pmap_kernel(), (vaddr_t)kva, size);
969 	pmap_update();
970 #if 0
971 	/*
972 	 * XXX ? is this necessary? i think so and i think other
973 	 * implementations are missing it.
974 	 */
975 	uvm_km_free(kernel_map, (vaddr_t)kva, size);
976 #endif
977 }
978