xref: /netbsd-src/sys/arch/sparc64/dev/iommu.c (revision 001c68bd94f75ce9270b69227c4199fbf34ee396)
1 /*	$NetBSD: iommu.c,v 1.65 2003/05/17 01:49:59 nakayama Exp $	*/
2 
3 /*
4  * Copyright (c) 2001, 2002 Eduardo Horvath
5  * Copyright (c) 1999, 2000 Matthew R. Green
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. The name of the author may not be used to endorse or promote products
17  *    derived from this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
24  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
25  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
26  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
27  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * UltraSPARC IOMMU support; used by both the sbus and pci code.
34  */
35 #include "opt_ddb.h"
36 
37 #include <sys/param.h>
38 #include <sys/extent.h>
39 #include <sys/malloc.h>
40 #include <sys/systm.h>
41 #include <sys/device.h>
42 #include <sys/proc.h>
43 
44 #include <uvm/uvm_extern.h>
45 
46 #include <machine/bus.h>
47 #include <sparc64/sparc64/cache.h>
48 #include <sparc64/dev/iommureg.h>
49 #include <sparc64/dev/iommuvar.h>
50 
51 #include <machine/autoconf.h>
52 #include <machine/cpu.h>
53 
54 #ifdef DEBUG
55 #define IDB_BUSDMA	0x1
56 #define IDB_IOMMU	0x2
57 #define IDB_INFO	0x4
58 #define	IDB_SYNC	0x8
59 int iommudebug = 0x0;
60 #define DPRINTF(l, s)   do { if (iommudebug & l) printf s; } while (0)
61 #else
62 #define DPRINTF(l, s)
63 #endif
64 
65 #define iommu_strbuf_flush(i, v) do {					\
66 	if ((i)->sb_flush)						\
67 		bus_space_write_8((i)->sb_is->is_bustag, (i)->sb_sb,	\
68 			STRBUFREG(strbuf_pgflush), (v));		\
69 	} while (0)
70 
71 static	int iommu_strbuf_flush_done __P((struct strbuf_ctl *));
72 
73 /*
74  * initialise the UltraSPARC IOMMU (SBUS or PCI):
75  *	- allocate and setup the iotsb.
76  *	- enable the IOMMU
77  *	- initialise the streaming buffers (if they exist)
78  *	- create a private DVMA map.
79  */
80 void
81 iommu_init(name, is, tsbsize, iovabase)
82 	char *name;
83 	struct iommu_state *is;
84 	int tsbsize;
85 	u_int32_t iovabase;
86 {
87 	psize_t size;
88 	vaddr_t va;
89 	paddr_t pa;
90 	struct vm_page *pg;
91 	struct pglist pglist;
92 
93 	/*
94 	 * Setup the iommu.
95 	 *
96 	 * The sun4u iommu is part of the SBUS or PCI controller so we will
97 	 * deal with it here..
98 	 *
99 	 * For sysio and psycho/psycho+ the IOMMU address space always ends at
100 	 * 0xffffe000, but the starting address depends on the size of the
101 	 * map.  The map size is 1024 * 2 ^ is->is_tsbsize entries, where each
102 	 * entry is 8 bytes.  The start of the map can be calculated by
103 	 * (0xffffe000 << (8 + is->is_tsbsize)).
104 	 *
105 	 * But sabre and hummingbird use a different scheme that seems to
106 	 * be hard-wired, so we read the start and size from the PROM and
107 	 * just use those values.
108 	 */
109 	is->is_cr = (tsbsize << 16) | IOMMUCR_EN;
110 	is->is_tsbsize = tsbsize;
111 	if (iovabase == -1) {
112 		is->is_dvmabase = IOTSB_VSTART(is->is_tsbsize);
113 		is->is_dvmaend = IOTSB_VEND;
114 	} else {
115 		is->is_dvmabase = iovabase;
116 		is->is_dvmaend = iovabase + IOTSB_VSIZE(tsbsize);
117 	}
118 
119 	/*
120 	 * Allocate memory for I/O pagetables.  They need to be physically
121 	 * contiguous.
122 	 */
123 
124 	size = PAGE_SIZE << is->is_tsbsize;
125 	if (uvm_pglistalloc((psize_t)size, (paddr_t)0, (paddr_t)-1,
126 		(paddr_t)PAGE_SIZE, (paddr_t)0, &pglist, 1, 0) != 0)
127 		panic("iommu_init: no memory");
128 
129 	va = uvm_km_valloc(kernel_map, size);
130 	if (va == 0)
131 		panic("iommu_init: no memory");
132 	is->is_tsb = (int64_t *)va;
133 
134 	is->is_ptsb = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pglist));
135 
136 	/* Map the pages */
137 	TAILQ_FOREACH(pg, &pglist, pageq) {
138 		pa = VM_PAGE_TO_PHYS(pg);
139 		pmap_kenter_pa(va, pa | PMAP_NVC, VM_PROT_READ | VM_PROT_WRITE);
140 		va += PAGE_SIZE;
141 	}
142 	pmap_update(pmap_kernel());
143 	memset(is->is_tsb, 0, size);
144 
145 #ifdef DEBUG
146 	if (iommudebug & IDB_INFO)
147 	{
148 		/* Probe the iommu */
149 
150 		printf("iommu regs at: cr=%lx tsb=%lx flush=%lx\n",
151 			(u_long)bus_space_read_8(is->is_bustag, is->is_iommu,
152 				offsetof (struct iommureg, iommu_cr)),
153 			(u_long)bus_space_read_8(is->is_bustag, is->is_iommu,
154 				offsetof (struct iommureg, iommu_tsb)),
155 			(u_long)bus_space_read_8(is->is_bustag, is->is_iommu,
156 				offsetof (struct iommureg, iommu_flush)));
157 		printf("iommu cr=%llx tsb=%llx\n",
158 			(unsigned long long)bus_space_read_8(is->is_bustag,
159 				is->is_iommu,
160 				offsetof (struct iommureg, iommu_cr)),
161 			(unsigned long long)bus_space_read_8(is->is_bustag,
162 				is->is_iommu,
163 				offsetof (struct iommureg, iommu_tsb)));
164 		printf("TSB base %p phys %llx\n", (void *)is->is_tsb,
165 			(unsigned long long)is->is_ptsb);
166 		delay(1000000); /* 1 s */
167 	}
168 #endif
169 
170 	/*
171 	 * now actually start up the IOMMU
172 	 */
173 	iommu_reset(is);
174 
175 	/*
176 	 * Now all the hardware's working we need to allocate a dvma map.
177 	 */
178 	printf("DVMA map: %x to %x\n",
179 		(unsigned int)is->is_dvmabase,
180 		(unsigned int)is->is_dvmaend);
181 	printf("IOTSB: %llx to %llx\n",
182 		(unsigned long long)is->is_ptsb,
183 		(unsigned long long)(is->is_ptsb + size));
184 	is->is_dvmamap = extent_create(name,
185 	    is->is_dvmabase, is->is_dvmaend - PAGE_SIZE,
186 	    M_DEVBUF, 0, 0, EX_NOWAIT);
187 }
188 
189 /*
190  * Streaming buffers don't exist on the UltraSPARC IIi; we should have
191  * detected that already and disabled them.  If not, we will notice that
192  * they aren't there when the STRBUF_EN bit does not remain.
193  */
194 void
195 iommu_reset(is)
196 	struct iommu_state *is;
197 {
198 	int i;
199 	struct strbuf_ctl *sb;
200 
201 	/* Need to do 64-bit stores */
202 	bus_space_write_8(is->is_bustag, is->is_iommu, IOMMUREG(iommu_tsb),
203 		is->is_ptsb);
204 
205 	/* Enable IOMMU in diagnostic mode */
206 	bus_space_write_8(is->is_bustag, is->is_iommu, IOMMUREG(iommu_cr),
207 		is->is_cr|IOMMUCR_DE);
208 
209 	for (i = 0; i < 2; i++) {
210 		if ((sb = is->is_sb[i])) {
211 
212 			/* Enable diagnostics mode? */
213 			bus_space_write_8(is->is_bustag, is->is_sb[i]->sb_sb,
214 				STRBUFREG(strbuf_ctl), STRBUF_EN);
215 
216 			/* No streaming buffers? Disable them */
217 			if (bus_space_read_8(is->is_bustag,
218 				is->is_sb[i]->sb_sb,
219 				STRBUFREG(strbuf_ctl)) == 0) {
220 				is->is_sb[i]->sb_flush = NULL;
221 			} else {
222 
223 				/*
224 				 * locate the pa of the flush buffer.
225 				 */
226 				(void)pmap_extract(pmap_kernel(),
227 					(vaddr_t)is->is_sb[i]->sb_flush,
228 					&is->is_sb[i]->sb_flushpa);
229 			}
230 		}
231 	}
232 }
233 
234 /*
235  * Here are the iommu control routines.
236  */
237 void
238 iommu_enter(sb, va, pa, flags)
239 	struct strbuf_ctl *sb;
240 	vaddr_t va;
241 	int64_t pa;
242 	int flags;
243 {
244 	struct iommu_state *is = sb->sb_is;
245 	int strbuf = (flags & BUS_DMA_STREAMING);
246 	int64_t tte;
247 
248 #ifdef DIAGNOSTIC
249 	if (va < is->is_dvmabase || va > is->is_dvmaend)
250 		panic("iommu_enter: va %#lx not in DVMA space", va);
251 #endif
252 
253 	/* Is the streamcache flush really needed? */
254 	if (sb->sb_flush) {
255 		iommu_strbuf_flush(sb, va);
256 		iommu_strbuf_flush_done(sb);
257 	} else
258 		/* If we can't flush the strbuf don't enable it. */
259 		strbuf = 0;
260 
261 	tte = MAKEIOTTE(pa, !(flags & BUS_DMA_NOWRITE),
262 		!(flags & BUS_DMA_NOCACHE), (strbuf));
263 #ifdef DEBUG
264 	tte |= (flags & 0xff000LL)<<(4*8);
265 #endif
266 
267 	DPRINTF(IDB_IOMMU, ("Clearing TSB slot %d for va %p\n",
268 		       (int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va));
269 	is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] = tte;
270 	bus_space_write_8(is->is_bustag, is->is_iommu,
271 		IOMMUREG(iommu_flush), va);
272 	DPRINTF(IDB_IOMMU, ("iommu_enter: va %lx pa %lx TSB[%lx]@%p=%lx\n",
273 		va, (long)pa, (u_long)IOTSBSLOT(va,is->is_tsbsize),
274 		(void *)(u_long)&is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)],
275 		(u_long)tte));
276 }
277 
278 /*
279  * Find the value of a DVMA address (debug routine).
280  */
281 paddr_t
282 iommu_extract(is, dva)
283 	struct iommu_state *is;
284 	vaddr_t dva;
285 {
286 	int64_t tte = 0;
287 
288 	if (dva >= is->is_dvmabase && dva < is->is_dvmaend)
289 		tte = is->is_tsb[IOTSBSLOT(dva, is->is_tsbsize)];
290 
291 	if ((tte & IOTTE_V) == 0)
292 		return ((paddr_t)-1L);
293 	return (tte & IOTTE_PAMASK);
294 }
295 
296 /*
297  * iommu_remove: removes mappings created by iommu_enter
298  *
299  * Only demap from IOMMU if flag is set.
300  *
301  * XXX: this function needs better internal error checking.
302  */
303 void
304 iommu_remove(is, va, len)
305 	struct iommu_state *is;
306 	vaddr_t va;
307 	size_t len;
308 {
309 
310 #ifdef DIAGNOSTIC
311 	if (va < is->is_dvmabase || va > is->is_dvmaend)
312 		panic("iommu_remove: va 0x%lx not in DVMA space", (u_long)va);
313 	if ((long)(va + len) < (long)va)
314 		panic("iommu_remove: va 0x%lx + len 0x%lx wraps",
315 		      (long) va, (long) len);
316 	if (len & ~0xfffffff)
317 		panic("iommu_remove: rediculous len 0x%lx", (u_long)len);
318 #endif
319 
320 	va = trunc_page(va);
321 	DPRINTF(IDB_IOMMU, ("iommu_remove: va %lx TSB[%lx]@%p\n",
322 		va, (u_long)IOTSBSLOT(va, is->is_tsbsize),
323 		&is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)]));
324 	while (len > 0) {
325 		DPRINTF(IDB_IOMMU, ("iommu_remove: clearing TSB slot %d "
326 			"for va %p size %lx\n",
327 			(int)IOTSBSLOT(va,is->is_tsbsize), (void *)(u_long)va,
328 			(u_long)len));
329 		if (len <= PAGE_SIZE)
330 			len = 0;
331 		else
332 			len -= PAGE_SIZE;
333 
334 		/* XXX Zero-ing the entry would not require RMW */
335 		is->is_tsb[IOTSBSLOT(va,is->is_tsbsize)] &= ~IOTTE_V;
336 		bus_space_write_8(is->is_bustag, is->is_iommu,
337 			IOMMUREG(iommu_flush), va);
338 		va += PAGE_SIZE;
339 	}
340 }
341 
342 static int
343 iommu_strbuf_flush_done(sb)
344 	struct strbuf_ctl *sb;
345 {
346 	struct iommu_state *is = sb->sb_is;
347 	struct timeval cur, flushtimeout;
348 
349 #define BUMPTIME(t, usec) { \
350 	register volatile struct timeval *tp = (t); \
351 	register long us; \
352  \
353 	tp->tv_usec = us = tp->tv_usec + (usec); \
354 	if (us >= 1000000) { \
355 		tp->tv_usec = us - 1000000; \
356 		tp->tv_sec++; \
357 	} \
358 }
359 
360 	if (!sb->sb_flush)
361 		return (0);
362 
363 	/*
364 	 * Streaming buffer flushes:
365 	 *
366 	 *   1 Tell strbuf to flush by storing va to strbuf_pgflush.  If
367 	 *     we're not on a cache line boundary (64-bits):
368 	 *   2 Store 0 in flag
369 	 *   3 Store pointer to flag in flushsync
370 	 *   4 wait till flushsync becomes 0x1
371 	 *
372 	 * If it takes more than .5 sec, something
373 	 * went wrong.
374 	 */
375 
376 	*sb->sb_flush = 0;
377 	bus_space_write_8(is->is_bustag, sb->sb_sb,
378 		STRBUFREG(strbuf_flushsync), sb->sb_flushpa);
379 
380 	microtime(&flushtimeout);
381 	cur = flushtimeout;
382 	BUMPTIME(&flushtimeout, 500000); /* 1/2 sec */
383 
384 	DPRINTF(IDB_IOMMU, ("iommu_strbuf_flush_done: flush = %lx "
385 		"at va = %lx pa = %lx now=%lx:%lx until = %lx:%lx\n",
386 		(long)*sb->sb_flush, (long)sb->sb_flush, (long)sb->sb_flushpa,
387 		cur.tv_sec, cur.tv_usec,
388 		flushtimeout.tv_sec, flushtimeout.tv_usec));
389 
390 	/* Bypass non-coherent D$ */
391 	while ((!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) &&
392 		timercmp(&cur, &flushtimeout, <=))
393 		microtime(&cur);
394 
395 #ifdef DIAGNOSTIC
396 	if (!ldxa(sb->sb_flushpa, ASI_PHYS_CACHED)) {
397 		printf("iommu_strbuf_flush_done: flush timeout %p, at %p\n",
398 			(void *)(u_long)*sb->sb_flush,
399 			(void *)(u_long)sb->sb_flushpa); /* panic? */
400 #ifdef DDB
401 		Debugger();
402 #endif
403 	}
404 #endif
405 	DPRINTF(IDB_IOMMU, ("iommu_strbuf_flush_done: flushed\n"));
406 	return (*sb->sb_flush);
407 }
408 
409 /*
410  * IOMMU DVMA operations, common to SBUS and PCI.
411  */
412 int
413 iommu_dvmamap_load(t, sb, map, buf, buflen, p, flags)
414 	bus_dma_tag_t t;
415 	struct strbuf_ctl *sb;
416 	bus_dmamap_t map;
417 	void *buf;
418 	bus_size_t buflen;
419 	struct proc *p;
420 	int flags;
421 {
422 	struct iommu_state *is = sb->sb_is;
423 	int s;
424 	int err;
425 	bus_size_t sgsize;
426 	paddr_t curaddr;
427 	u_long dvmaddr, sgstart, sgend;
428 	bus_size_t align, boundary;
429 	vaddr_t vaddr = (vaddr_t)buf;
430 	int seg;
431 	struct pmap *pmap;
432 
433 	if (map->dm_nsegs) {
434 		/* Already in use?? */
435 #ifdef DIAGNOSTIC
436 		printf("iommu_dvmamap_load: map still in use\n");
437 #endif
438 		bus_dmamap_unload(t, map);
439 	}
440 
441 	/*
442 	 * Make sure that on error condition we return "no valid mappings".
443 	 */
444 	map->dm_nsegs = 0;
445 	if (buflen > map->_dm_size) {
446 		DPRINTF(IDB_BUSDMA,
447 		    ("iommu_dvmamap_load(): error %d > %d -- "
448 		     "map size exceeded!\n", (int)buflen, (int)map->_dm_size));
449 		return (EINVAL);
450 	}
451 
452 	sgsize = round_page(buflen + ((int)vaddr & PGOFSET));
453 
454 	/*
455 	 * A boundary presented to bus_dmamem_alloc() takes precedence
456 	 * over boundary in the map.
457 	 */
458 	if ((boundary = (map->dm_segs[0]._ds_boundary)) == 0)
459 		boundary = map->_dm_boundary;
460 	align = max(map->dm_segs[0]._ds_align, PAGE_SIZE);
461 
462 	/*
463 	 * If our segment size is larger than the boundary we need to
464 	 * split the transfer up int little pieces ourselves.
465 	 */
466 	s = splhigh();
467 	err = extent_alloc(is->is_dvmamap, sgsize, align,
468 		(sgsize > boundary) ? 0 : boundary,
469 		EX_NOWAIT|EX_BOUNDZERO, &dvmaddr);
470 	splx(s);
471 
472 #ifdef DEBUG
473 	if (err || (dvmaddr == (u_long)-1))
474 	{
475 		printf("iommu_dvmamap_load(): extent_alloc(%d, %x) failed!\n",
476 		    (int)sgsize, flags);
477 #ifdef DDB
478 		Debugger();
479 #endif
480 	}
481 #endif
482 	if (err != 0)
483 		return (err);
484 
485 	if (dvmaddr == (u_long)-1)
486 		return (ENOMEM);
487 
488 	/* Set the active DVMA map */
489 	map->_dm_dvmastart = dvmaddr;
490 	map->_dm_dvmasize = sgsize;
491 
492 	/*
493 	 * Now split the DVMA range into segments, not crossing
494 	 * the boundary.
495 	 */
496 	seg = 0;
497 	sgstart = dvmaddr + (vaddr & PGOFSET);
498 	sgend = sgstart + buflen - 1;
499 	map->dm_segs[seg].ds_addr = sgstart;
500 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: boundary %lx boundary-1 %lx "
501 		"~(boundary-1) %lx\n", (long)boundary, (long)(boundary-1), (long)~(boundary-1)));
502 	while ((sgstart & ~(boundary - 1)) != (sgend & ~(boundary - 1))) {
503 		/* Oops.  We crossed a boundary.  Split the xfer. */
504 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
505 			"seg %d start %lx size %lx\n", seg,
506 			(long)map->dm_segs[seg].ds_addr,
507 			(long)map->dm_segs[seg].ds_len));
508 		map->dm_segs[seg].ds_len =
509 		    boundary - (sgstart & (boundary - 1));
510 		if (++seg >= map->_dm_segcnt) {
511 			/* Too many segments.  Fail the operation. */
512 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
513 				"too many segments %d\n", seg));
514 			s = splhigh();
515 			/* How can this fail?  And if it does what can we do? */
516 			err = extent_free(is->is_dvmamap,
517 				dvmaddr, sgsize, EX_NOWAIT);
518 			map->_dm_dvmastart = 0;
519 			map->_dm_dvmasize = 0;
520 			splx(s);
521 			return (E2BIG);
522 		}
523 		sgstart = roundup(sgstart, boundary);
524 		map->dm_segs[seg].ds_addr = sgstart;
525 	}
526 	map->dm_segs[seg].ds_len = sgend - sgstart + 1;
527 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load: "
528 		"seg %d start %lx size %lx\n", seg,
529 		(long)map->dm_segs[seg].ds_addr, (long)map->dm_segs[seg].ds_len));
530 	map->dm_nsegs = seg+1;
531 	map->dm_mapsize = buflen;
532 
533 	if (p != NULL)
534 		pmap = p->p_vmspace->vm_map.pmap;
535 	else
536 		pmap = pmap_kernel();
537 
538 	for (; buflen > 0; ) {
539 
540 		/*
541 		 * Get the physical address for this page.
542 		 */
543 		if (pmap_extract(pmap, (vaddr_t)vaddr, &curaddr) == FALSE) {
544 			bus_dmamap_unload(t, map);
545 			return (-1);
546 		}
547 
548 		/*
549 		 * Compute the segment size, and adjust counts.
550 		 */
551 		sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
552 		if (buflen < sgsize)
553 			sgsize = buflen;
554 
555 		DPRINTF(IDB_BUSDMA,
556 		    ("iommu_dvmamap_load: map %p loading va %p "
557 			    "dva %lx at pa %lx\n",
558 			    map, (void *)vaddr, (long)dvmaddr,
559 			    (long)(curaddr & ~(PAGE_SIZE-1))));
560 		iommu_enter(sb, trunc_page(dvmaddr), trunc_page(curaddr),
561 		    flags|0x4000);
562 
563 		dvmaddr += PAGE_SIZE;
564 		vaddr += sgsize;
565 		buflen -= sgsize;
566 	}
567 #ifdef DIAGNOSTIC
568 	for (seg = 0; seg < map->dm_nsegs; seg++) {
569 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
570 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
571 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
572 				seg, (long)map->dm_segs[seg].ds_addr,
573 				is->is_dvmabase, is->is_dvmaend);
574 #ifdef DDB
575 			Debugger();
576 #endif
577 		}
578 	}
579 #endif
580 	return (0);
581 }
582 
583 
584 void
585 iommu_dvmamap_unload(t, sb, map)
586 	bus_dma_tag_t t;
587 	struct strbuf_ctl *sb;
588 	bus_dmamap_t map;
589 {
590 	struct iommu_state *is = sb->sb_is;
591 	int error, s;
592 	bus_size_t sgsize;
593 
594 	/* Flush the iommu */
595 #ifdef DEBUG
596 	if (!map->_dm_dvmastart) {
597 		printf("iommu_dvmamap_unload: No dvmastart is zero\n");
598 #ifdef DDB
599 		Debugger();
600 #endif
601 	}
602 #endif
603 	iommu_remove(is, map->_dm_dvmastart, map->_dm_dvmasize);
604 
605 	/* Flush the caches */
606 	bus_dmamap_unload(t->_parent, map);
607 
608 	/* Mark the mappings as invalid. */
609 	map->dm_mapsize = 0;
610 	map->dm_nsegs = 0;
611 
612 	s = splhigh();
613 	error = extent_free(is->is_dvmamap, map->_dm_dvmastart,
614 		map->_dm_dvmasize, EX_NOWAIT);
615 	map->_dm_dvmastart = 0;
616 	map->_dm_dvmasize = 0;
617 	splx(s);
618 	if (error != 0)
619 		printf("warning: %qd of DVMA space lost\n", (long long)sgsize);
620 
621 	/* Clear the map */
622 }
623 
624 
625 int
626 iommu_dvmamap_load_raw(t, sb, map, segs, nsegs, flags, size)
627 	bus_dma_tag_t t;
628 	struct strbuf_ctl *sb;
629 	bus_dmamap_t map;
630 	bus_dma_segment_t *segs;
631 	int nsegs;
632 	int flags;
633 	bus_size_t size;
634 {
635 	struct iommu_state *is = sb->sb_is;
636 	struct vm_page *pg;
637 	int i, j, s;
638 	int left;
639 	int err;
640 	bus_size_t sgsize;
641 	paddr_t pa;
642 	bus_size_t boundary, align;
643 	u_long dvmaddr, sgstart, sgend;
644 	struct pglist *pglist;
645 	int pagesz = PAGE_SIZE;
646 	int npg = 0; /* DEBUG */
647 
648 	if (map->dm_nsegs) {
649 		/* Already in use?? */
650 #ifdef DIAGNOSTIC
651 		printf("iommu_dvmamap_load_raw: map still in use\n");
652 #endif
653 		bus_dmamap_unload(t, map);
654 	}
655 
656 	/*
657 	 * A boundary presented to bus_dmamem_alloc() takes precedence
658 	 * over boundary in the map.
659 	 */
660 	if ((boundary = segs[0]._ds_boundary) == 0)
661 		boundary = map->_dm_boundary;
662 
663 	align = max(segs[0]._ds_align, pagesz);
664 
665 	/*
666 	 * Make sure that on error condition we return "no valid mappings".
667 	 */
668 	map->dm_nsegs = 0;
669 	/* Count up the total number of pages we need */
670 	pa = segs[0].ds_addr;
671 	sgsize = 0;
672 	left = size;
673 	for (i = 0; left && i < nsegs; i++) {
674 		if (round_page(pa) != round_page(segs[i].ds_addr))
675 			sgsize = round_page(sgsize);
676 		sgsize += min(left, segs[i].ds_len);
677 		left -= segs[i].ds_len;
678 		pa = segs[i].ds_addr + segs[i].ds_len;
679 	}
680 	sgsize = round_page(sgsize);
681 
682 	s = splhigh();
683 	/*
684 	 * If our segment size is larger than the boundary we need to
685 	 * split the transfer up into little pieces ourselves.
686 	 */
687 	err = extent_alloc(is->is_dvmamap, sgsize, align,
688 		(sgsize > boundary) ? 0 : boundary,
689 		((flags & BUS_DMA_NOWAIT) == 0 ? EX_WAITOK : EX_NOWAIT) |
690 		EX_BOUNDZERO, &dvmaddr);
691 	splx(s);
692 
693 	if (err != 0)
694 		return (err);
695 
696 #ifdef DEBUG
697 	if (dvmaddr == (u_long)-1)
698 	{
699 		printf("iommu_dvmamap_load_raw(): extent_alloc(%d, %x) failed!\n",
700 		    (int)sgsize, flags);
701 #ifdef DDB
702 		Debugger();
703 #endif
704 	}
705 #endif
706 	if (dvmaddr == (u_long)-1)
707 		return (ENOMEM);
708 
709 	/* Set the active DVMA map */
710 	map->_dm_dvmastart = dvmaddr;
711 	map->_dm_dvmasize = sgsize;
712 
713 	if ((pglist = segs[0]._ds_mlist) == NULL) {
714 		u_long prev_va = NULL;
715 		paddr_t prev_pa = 0;
716 		int end = 0, offset;
717 
718 		/*
719 		 * This segs is made up of individual physical
720 		 *  segments, probably by _bus_dmamap_load_uio() or
721 		 * _bus_dmamap_load_mbuf().  Ignore the mlist and
722 		 * load each one individually.
723 		 */
724 		map->dm_mapsize = size;
725 
726 		j = 0;
727 		for (i = 0; i < nsegs ; i++) {
728 
729 			pa = segs[i].ds_addr;
730 			offset = (pa & PGOFSET);
731 			pa = trunc_page(pa);
732 			dvmaddr = trunc_page(dvmaddr);
733 			left = min(size, segs[i].ds_len);
734 
735 			DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: converting "
736 				"physseg %d start %lx size %lx\n", i,
737 				(long)segs[i].ds_addr, (long)segs[i].ds_len));
738 
739 			if ((pa == prev_pa) &&
740 				((offset != 0) || (end != offset))) {
741 				/* We can re-use this mapping */
742 				dvmaddr = prev_va;
743 			}
744 
745 			sgstart = dvmaddr + offset;
746 			sgend = sgstart + left - 1;
747 
748 			/* Are the segments virtually adjacent? */
749 			if ((j > 0) && (end == offset) &&
750 				((offset == 0) || (pa == prev_pa))) {
751 				/* Just append to the previous segment. */
752 				map->dm_segs[--j].ds_len += left;
753 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
754 					"appending seg %d start %lx size %lx\n", j,
755 					(long)map->dm_segs[j].ds_addr,
756 					(long)map->dm_segs[j].ds_len));
757 			} else {
758 				if (j >= map->_dm_segcnt) {
759 					iommu_dvmamap_unload(t, sb, map);
760 					return (E2BIG);
761 				}
762 				map->dm_segs[j].ds_addr = sgstart;
763 				map->dm_segs[j].ds_len = left;
764 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
765 					"seg %d start %lx size %lx\n", j,
766 					(long)map->dm_segs[j].ds_addr,
767 					(long)map->dm_segs[j].ds_len));
768 			}
769 			end = (offset + left) & PGOFSET;
770 
771 			/* Check for boundary issues */
772 			while ((sgstart & ~(boundary - 1)) !=
773 				(sgend & ~(boundary - 1))) {
774 				/* Need a new segment. */
775 				map->dm_segs[j].ds_len =
776 					boundary - (sgstart & (boundary - 1));
777 				DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
778 					"seg %d start %lx size %lx\n", j,
779 					(long)map->dm_segs[j].ds_addr,
780 					(long)map->dm_segs[j].ds_len));
781 				if (++j >= map->_dm_segcnt) {
782 					iommu_dvmamap_unload(t, sb, map);
783 					return (E2BIG);
784 				}
785 				sgstart = roundup(sgstart, boundary);
786 				map->dm_segs[j].ds_addr = sgstart;
787 				map->dm_segs[j].ds_len = sgend - sgstart + 1;
788 			}
789 
790 			if (sgsize == 0)
791 				panic("iommu_dmamap_load_raw: size botch");
792 
793 			/* Now map a series of pages. */
794 			while (dvmaddr <= sgend) {
795 				DPRINTF(IDB_BUSDMA,
796 					("iommu_dvmamap_load_raw: map %p "
797 						"loading va %lx at pa %lx\n",
798 						map, (long)dvmaddr,
799 						(long)(pa)));
800 				/* Enter it if we haven't before. */
801 				if (prev_va != dvmaddr)
802 					iommu_enter(sb, prev_va = dvmaddr,
803 						prev_pa = pa,
804 						flags | (++npg << 12));
805 				dvmaddr += pagesz;
806 				pa += pagesz;
807 			}
808 
809 			size -= left;
810 			++j;
811 		}
812 
813 		map->dm_nsegs = j;
814 #ifdef DIAGNOSTIC
815 		{ int seg;
816 	for (seg = 0; seg < map->dm_nsegs; seg++) {
817 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
818 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
819 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
820 				seg, (long)map->dm_segs[seg].ds_addr,
821 				is->is_dvmabase, is->is_dvmaend);
822 #ifdef DDB
823 			Debugger();
824 #endif
825 		}
826 	}
827 		}
828 #endif
829 		return (0);
830 	}
831 
832 	/*
833 	 * This was allocated with bus_dmamem_alloc.
834 	 * The pages are on a `pglist'.
835 	 */
836 	map->dm_mapsize = size;
837 	i = 0;
838 	sgstart = dvmaddr;
839 	sgend = sgstart + size - 1;
840 	map->dm_segs[i].ds_addr = sgstart;
841 	while ((sgstart & ~(boundary - 1)) != (sgend & ~(boundary - 1))) {
842 		/* Oops.  We crossed a boundary.  Split the xfer. */
843 		map->dm_segs[i].ds_len = boundary - (sgstart & (boundary - 1));
844 		DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
845 			"seg %d start %lx size %lx\n", i,
846 			(long)map->dm_segs[i].ds_addr,
847 			(long)map->dm_segs[i].ds_len));
848 		if (++i >= map->_dm_segcnt) {
849 			/* Too many segments.  Fail the operation. */
850 			s = splhigh();
851 			/* How can this fail?  And if it does what can we do? */
852 			err = extent_free(is->is_dvmamap,
853 				dvmaddr, sgsize, EX_NOWAIT);
854 			map->_dm_dvmastart = 0;
855 			map->_dm_dvmasize = 0;
856 			splx(s);
857 			return (E2BIG);
858 		}
859 		sgstart = roundup(sgstart, boundary);
860 		map->dm_segs[i].ds_addr = sgstart;
861 	}
862 	DPRINTF(IDB_INFO, ("iommu_dvmamap_load_raw: "
863 			"seg %d start %lx size %lx\n", i,
864 			(long)map->dm_segs[i].ds_addr, (long)map->dm_segs[i].ds_len));
865 	map->dm_segs[i].ds_len = sgend - sgstart + 1;
866 
867 	TAILQ_FOREACH(pg, pglist, pageq) {
868 		if (sgsize == 0)
869 			panic("iommu_dmamap_load_raw: size botch");
870 		pa = VM_PAGE_TO_PHYS(pg);
871 
872 		DPRINTF(IDB_BUSDMA,
873 		    ("iommu_dvmamap_load_raw: map %p loading va %lx at pa %lx\n",
874 		    map, (long)dvmaddr, (long)(pa)));
875 		iommu_enter(sb, dvmaddr, pa, flags|0x8000);
876 
877 		dvmaddr += pagesz;
878 		sgsize -= pagesz;
879 	}
880 	map->dm_mapsize = size;
881 	map->dm_nsegs = i+1;
882 #ifdef DIAGNOSTIC
883 	{ int seg;
884 	for (seg = 0; seg < map->dm_nsegs; seg++) {
885 		if (map->dm_segs[seg].ds_addr < is->is_dvmabase ||
886 			map->dm_segs[seg].ds_addr > is->is_dvmaend) {
887 			printf("seg %d dvmaddr %lx out of range %x - %x\n",
888 				seg, (long)map->dm_segs[seg].ds_addr,
889 				is->is_dvmabase, is->is_dvmaend);
890 #ifdef DDB
891 			Debugger();
892 #endif
893 		}
894 	}
895 	}
896 #endif
897 	return (0);
898 }
899 
900 void
901 iommu_dvmamap_sync(t, sb, map, offset, len, ops)
902 	bus_dma_tag_t t;
903 	struct strbuf_ctl *sb;
904 	bus_dmamap_t map;
905 	bus_addr_t offset;
906 	bus_size_t len;
907 	int ops;
908 {
909 	struct iommu_state *is = sb->sb_is;
910 	vaddr_t va = map->dm_segs[0].ds_addr + offset;
911 	int64_t tte;
912 	vaddr_t vaend;
913 
914 	if (!sb->sb_flush)
915 		return;
916 
917 	/*
918 	 * We only support one DMA segment; supporting more makes this code
919 	 * too unwieldy.
920 	 */
921 	if (map->dm_nsegs > 1)
922 		panic("iommu_dvmamap_sync: %d segments", map->dm_nsegs);
923 
924 	DPRINTF(IDB_SYNC,
925 	    ("iommu_dvmamap_sync: syncing va %p len %lu "
926 	     "ops 0x%x\n", (void *)(u_long)va, (u_long)len, ops));
927 
928 	if (ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_POSTWRITE)) {
929 		/* Nothing to do */;
930 	}
931 
932 	if (ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_PREWRITE)) {
933 #ifdef DIAGNOSTIC
934 		if (va < is->is_dvmabase || va >= is->is_dvmaend)
935 			panic("iommu_dvmamap_sync: invalid dva %lx", va);
936 #endif
937 		tte = is->is_tsb[IOTSBSLOT(va, is->is_tsbsize)];
938 
939 		/* if we have a streaming buffer, flush it here first */
940 		if ((tte & IOTTE_STREAM) && sb->sb_flush) {
941 			vaend = (va + len + PGOFSET) & ~PGOFSET;
942 
943 			for (va &= ~PGOFSET; va <= vaend; va += PAGE_SIZE) {
944 				DPRINTF(IDB_BUSDMA,
945 					("iommu_dvmamap_sync: flushing va %p\n",
946 					 (void *)(u_long)va));
947 				iommu_strbuf_flush(sb, va);
948 			}
949 
950 			iommu_strbuf_flush_done(sb);
951 		}
952 	}
953 }
954 
955 int
956 iommu_dvmamem_alloc(t, sb, size, alignment, boundary, segs, nsegs, rsegs, flags)
957 	bus_dma_tag_t t;
958 	struct strbuf_ctl *sb;
959 	bus_size_t size, alignment, boundary;
960 	bus_dma_segment_t *segs;
961 	int nsegs;
962 	int *rsegs;
963 	int flags;
964 {
965 
966 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_alloc: sz %llx align %llx bound %llx "
967 	   "segp %p flags %d\n", (unsigned long long)size,
968 	   (unsigned long long)alignment, (unsigned long long)boundary,
969 	   segs, flags));
970 	return (bus_dmamem_alloc(t->_parent, size, alignment, boundary,
971 	    segs, nsegs, rsegs, flags|BUS_DMA_DVMA));
972 }
973 
974 void
975 iommu_dvmamem_free(t, sb, segs, nsegs)
976 	bus_dma_tag_t t;
977 	struct strbuf_ctl *sb;
978 	bus_dma_segment_t *segs;
979 	int nsegs;
980 {
981 
982 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_free: segp %p nsegs %d\n",
983 	    segs, nsegs));
984 	bus_dmamem_free(t->_parent, segs, nsegs);
985 }
986 
987 /*
988  * Map the DVMA mappings into the kernel pmap.
989  * Check the flags to see whether we're streaming or coherent.
990  */
991 int
992 iommu_dvmamem_map(t, sb, segs, nsegs, size, kvap, flags)
993 	bus_dma_tag_t t;
994 	struct strbuf_ctl *sb;
995 	bus_dma_segment_t *segs;
996 	int nsegs;
997 	size_t size;
998 	caddr_t *kvap;
999 	int flags;
1000 {
1001 	struct vm_page *pg;
1002 	vaddr_t va;
1003 	bus_addr_t addr;
1004 	struct pglist *pglist;
1005 	int cbit;
1006 
1007 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: segp %p nsegs %d size %lx\n",
1008 	    segs, nsegs, size));
1009 
1010 	/*
1011 	 * Allocate some space in the kernel map, and then map these pages
1012 	 * into this space.
1013 	 */
1014 	size = round_page(size);
1015 	va = uvm_km_valloc(kernel_map, size);
1016 	if (va == 0)
1017 		return (ENOMEM);
1018 
1019 	*kvap = (caddr_t)va;
1020 
1021 	/*
1022 	 * digest flags:
1023 	 */
1024 	cbit = 0;
1025 	if (flags & BUS_DMA_COHERENT)	/* Disable vcache */
1026 		cbit |= PMAP_NVC;
1027 	if (flags & BUS_DMA_NOCACHE)	/* sideffects */
1028 		cbit |= PMAP_NC;
1029 
1030 	/*
1031 	 * Now take this and map it into the CPU.
1032 	 */
1033 	pglist = segs[0]._ds_mlist;
1034 	TAILQ_FOREACH(pg, pglist, pageq) {
1035 #ifdef DIAGNOSTIC
1036 		if (size == 0)
1037 			panic("iommu_dvmamem_map: size botch");
1038 #endif
1039 		addr = VM_PAGE_TO_PHYS(pg);
1040 		DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_map: "
1041 		    "mapping va %lx at %llx\n", va, (unsigned long long)addr | cbit));
1042 		pmap_kenter_pa(va, addr | cbit, VM_PROT_READ | VM_PROT_WRITE);
1043 		va += PAGE_SIZE;
1044 		size -= PAGE_SIZE;
1045 	}
1046 	pmap_update(pmap_kernel());
1047 	return (0);
1048 }
1049 
1050 /*
1051  * Unmap DVMA mappings from kernel
1052  */
1053 void
1054 iommu_dvmamem_unmap(t, sb, kva, size)
1055 	bus_dma_tag_t t;
1056 	struct strbuf_ctl *sb;
1057 	caddr_t kva;
1058 	size_t size;
1059 {
1060 
1061 	DPRINTF(IDB_BUSDMA, ("iommu_dvmamem_unmap: kvm %p size %lx\n",
1062 	    kva, size));
1063 
1064 #ifdef DIAGNOSTIC
1065 	if ((u_long)kva & PGOFSET)
1066 		panic("iommu_dvmamem_unmap");
1067 #endif
1068 
1069 	size = round_page(size);
1070 	pmap_kremove((vaddr_t)kva, size);
1071 	pmap_update(pmap_kernel());
1072 	uvm_km_free(kernel_map, (vaddr_t)kva, size);
1073 }
1074