bus_dma.c revision 1.62 1 /* $NetBSD: bus_dma.c,v 1.62 2012/10/19 11:57:58 matt Exp $ */
2
3 /*-
4 * Copyright (c) 1996, 1997, 1998 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9 * NASA Ames Research Center.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30 * POSSIBILITY OF SUCH DAMAGE.
31 */
32
33 #define _ARM32_BUS_DMA_PRIVATE
34
35 #include <sys/cdefs.h>
36 __KERNEL_RCSID(0, "$NetBSD: bus_dma.c,v 1.62 2012/10/19 11:57:58 matt Exp $");
37
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/proc.h>
42 #include <sys/buf.h>
43 #include <sys/reboot.h>
44 #include <sys/conf.h>
45 #include <sys/file.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/vnode.h>
49 #include <sys/device.h>
50
51 #include <uvm/uvm.h>
52
53 #include <sys/bus.h>
54 #include <machine/cpu.h>
55
56 #include <arm/cpufunc.h>
57
58 static struct evcnt bus_dma_creates =
59 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "creates");
60 static struct evcnt bus_dma_bounced_creates =
61 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced creates");
62 static struct evcnt bus_dma_loads =
63 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "loads");
64 static struct evcnt bus_dma_bounced_loads =
65 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced loads");
66 static struct evcnt bus_dma_read_bounces =
67 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "read bounces");
68 static struct evcnt bus_dma_write_bounces =
69 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "write bounces");
70 static struct evcnt bus_dma_bounced_unloads =
71 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced unloads");
72 static struct evcnt bus_dma_unloads =
73 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "unloads");
74 static struct evcnt bus_dma_bounced_destroys =
75 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced destroys");
76 static struct evcnt bus_dma_destroys =
77 EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "destroys");
78
79 EVCNT_ATTACH_STATIC(bus_dma_creates);
80 EVCNT_ATTACH_STATIC(bus_dma_bounced_creates);
81 EVCNT_ATTACH_STATIC(bus_dma_loads);
82 EVCNT_ATTACH_STATIC(bus_dma_bounced_loads);
83 EVCNT_ATTACH_STATIC(bus_dma_read_bounces);
84 EVCNT_ATTACH_STATIC(bus_dma_write_bounces);
85 EVCNT_ATTACH_STATIC(bus_dma_unloads);
86 EVCNT_ATTACH_STATIC(bus_dma_bounced_unloads);
87 EVCNT_ATTACH_STATIC(bus_dma_destroys);
88 EVCNT_ATTACH_STATIC(bus_dma_bounced_destroys);
89
90 #define STAT_INCR(x) (bus_dma_ ## x.ev_count++)
91
92 int _bus_dmamap_load_buffer(bus_dma_tag_t, bus_dmamap_t, void *,
93 bus_size_t, struct vmspace *, int);
94 static struct arm32_dma_range *
95 _bus_dma_paddr_inrange(struct arm32_dma_range *, int, paddr_t);
96
97 /*
98 * Check to see if the specified page is in an allowed DMA range.
99 */
100 inline struct arm32_dma_range *
101 _bus_dma_paddr_inrange(struct arm32_dma_range *ranges, int nranges,
102 bus_addr_t curaddr)
103 {
104 struct arm32_dma_range *dr;
105 int i;
106
107 for (i = 0, dr = ranges; i < nranges; i++, dr++) {
108 if (curaddr >= dr->dr_sysbase &&
109 round_page(curaddr) <= (dr->dr_sysbase + dr->dr_len))
110 return (dr);
111 }
112
113 return (NULL);
114 }
115
116 /*
117 * Check to see if the specified busaddr is in an allowed DMA range.
118 */
119 static inline paddr_t
120 _bus_dma_busaddr_to_paddr(bus_dma_tag_t t, bus_addr_t curaddr)
121 {
122 struct arm32_dma_range *dr;
123 u_int i;
124
125 if (t->_nranges == 0)
126 return curaddr;
127
128 for (i = 0, dr = t->_ranges; i < t->_nranges; i++, dr++) {
129 if (dr->dr_busbase <= curaddr
130 && round_page(curaddr) <= dr->dr_busbase + dr->dr_len)
131 return curaddr - dr->dr_busbase + dr->dr_sysbase;
132 }
133 panic("%s: curaddr %#lx not in range", __func__, curaddr);
134 }
135
136 /*
137 * Common function to load the specified physical address into the
138 * DMA map, coalescing segments and boundary checking as necessary.
139 */
140 static int
141 _bus_dmamap_load_paddr(bus_dma_tag_t t, bus_dmamap_t map,
142 bus_addr_t paddr, bus_size_t size, bool coherent)
143 {
144 bus_dma_segment_t * const segs = map->dm_segs;
145 int nseg = map->dm_nsegs;
146 bus_addr_t lastaddr;
147 bus_addr_t bmask = ~(map->_dm_boundary - 1);
148 bus_addr_t curaddr;
149 bus_size_t sgsize;
150 uint32_t _ds_flags = coherent ? _BUS_DMAMAP_COHERENT : 0;
151
152 if (nseg > 0)
153 lastaddr = segs[nseg-1].ds_addr + segs[nseg-1].ds_len;
154 else
155 lastaddr = 0xdead;
156
157 again:
158 sgsize = size;
159
160 /* Make sure we're in an allowed DMA range. */
161 if (t->_ranges != NULL) {
162 /* XXX cache last result? */
163 const struct arm32_dma_range * const dr =
164 _bus_dma_paddr_inrange(t->_ranges, t->_nranges, paddr);
165 if (dr == NULL)
166 return (EINVAL);
167
168 /*
169 * If this region is coherent, mark the segment as coherent.
170 */
171 _ds_flags |= dr->dr_flags & _BUS_DMAMAP_COHERENT;
172 #if 0
173 printf("%p: %#lx: range %#lx/%#lx/%#lx/%#x: %#x\n",
174 t, paddr, dr->dr_sysbase, dr->dr_busbase,
175 dr->dr_len, dr->dr_flags, _ds_flags);
176 #endif
177 /*
178 * In a valid DMA range. Translate the physical
179 * memory address to an address in the DMA window.
180 */
181 curaddr = (paddr - dr->dr_sysbase) + dr->dr_busbase;
182 } else
183 curaddr = paddr;
184
185 /*
186 * Make sure we don't cross any boundaries.
187 */
188 if (map->_dm_boundary > 0) {
189 bus_addr_t baddr; /* next boundary address */
190
191 baddr = (curaddr + map->_dm_boundary) & bmask;
192 if (sgsize > (baddr - curaddr))
193 sgsize = (baddr - curaddr);
194 }
195
196 /*
197 * Insert chunk into a segment, coalescing with the
198 * previous segment if possible.
199 */
200 if (nseg > 0 && curaddr == lastaddr &&
201 segs[nseg-1].ds_len + sgsize <= map->dm_maxsegsz &&
202 ((segs[nseg-1]._ds_flags ^ _ds_flags) & _BUS_DMAMAP_COHERENT) == 0 &&
203 (map->_dm_boundary == 0 ||
204 (segs[nseg-1].ds_addr & bmask) == (curaddr & bmask))) {
205 /* coalesce */
206 segs[nseg-1].ds_len += sgsize;
207 } else if (nseg >= map->_dm_segcnt) {
208 return (EFBIG);
209 } else {
210 /* new segment */
211 segs[nseg].ds_addr = curaddr;
212 segs[nseg].ds_len = sgsize;
213 segs[nseg]._ds_flags = _ds_flags;
214 nseg++;
215 }
216
217 lastaddr = curaddr + sgsize;
218
219 paddr += sgsize;
220 size -= sgsize;
221 if (size > 0)
222 goto again;
223
224 map->_dm_flags &= (_ds_flags & _BUS_DMAMAP_COHERENT);
225 map->dm_nsegs = nseg;
226 return (0);
227 }
228
229 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
230 static int _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
231 bus_size_t size, int flags);
232 static void _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map);
233 static int _bus_dma_uiomove(void *buf, struct uio *uio, size_t n,
234 int direction);
235
236 static int
237 _bus_dma_load_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
238 size_t buflen, int buftype, int flags)
239 {
240 struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
241 struct vmspace * const vm = vmspace_kernel();
242 int error;
243
244 KASSERT(cookie != NULL);
245 KASSERT(cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE);
246
247 /*
248 * Allocate bounce pages, if necessary.
249 */
250 if ((cookie->id_flags & _BUS_DMA_HAS_BOUNCE) == 0) {
251 error = _bus_dma_alloc_bouncebuf(t, map, buflen, flags);
252 if (error)
253 return (error);
254 }
255
256 /*
257 * Cache a pointer to the caller's buffer and load the DMA map
258 * with the bounce buffer.
259 */
260 cookie->id_origbuf = buf;
261 cookie->id_origbuflen = buflen;
262 error = _bus_dmamap_load_buffer(t, map, cookie->id_bouncebuf,
263 buflen, vm, flags);
264 if (error)
265 return (error);
266
267 STAT_INCR(bounced_loads);
268 map->dm_mapsize = buflen;
269 map->_dm_vmspace = vm;
270 map->_dm_buftype = buftype;
271
272 /* ...so _bus_dmamap_sync() knows we're bouncing */
273 cookie->id_flags |= _BUS_DMA_IS_BOUNCING;
274 return 0;
275 }
276 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
277
278 /*
279 * Common function for DMA map creation. May be called by bus-specific
280 * DMA map creation functions.
281 */
282 int
283 _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
284 bus_size_t maxsegsz, bus_size_t boundary, int flags, bus_dmamap_t *dmamp)
285 {
286 struct arm32_bus_dmamap *map;
287 void *mapstore;
288 size_t mapsize;
289
290 #ifdef DEBUG_DMA
291 printf("dmamap_create: t=%p size=%lx nseg=%x msegsz=%lx boundary=%lx flags=%x\n",
292 t, size, nsegments, maxsegsz, boundary, flags);
293 #endif /* DEBUG_DMA */
294
295 /*
296 * Allocate and initialize the DMA map. The end of the map
297 * is a variable-sized array of segments, so we allocate enough
298 * room for them in one shot.
299 *
300 * Note we don't preserve the WAITOK or NOWAIT flags. Preservation
301 * of ALLOCNOW notifies others that we've reserved these resources,
302 * and they are not to be freed.
303 *
304 * The bus_dmamap_t includes one bus_dma_segment_t, hence
305 * the (nsegments - 1).
306 */
307 mapsize = sizeof(struct arm32_bus_dmamap) +
308 (sizeof(bus_dma_segment_t) * (nsegments - 1));
309 const int mallocflags = M_ZERO|(flags & BUS_DMA_NOWAIT) ? M_NOWAIT : M_WAITOK;
310 if ((mapstore = malloc(mapsize, M_DMAMAP, mallocflags)) == NULL)
311 return (ENOMEM);
312
313 map = (struct arm32_bus_dmamap *)mapstore;
314 map->_dm_size = size;
315 map->_dm_segcnt = nsegments;
316 map->_dm_maxmaxsegsz = maxsegsz;
317 map->_dm_boundary = boundary;
318 map->_dm_flags = flags & ~(BUS_DMA_WAITOK|BUS_DMA_NOWAIT);
319 map->_dm_origbuf = NULL;
320 map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
321 map->_dm_vmspace = vmspace_kernel();
322 map->_dm_cookie = NULL;
323 map->dm_maxsegsz = maxsegsz;
324 map->dm_mapsize = 0; /* no valid mappings */
325 map->dm_nsegs = 0;
326
327 *dmamp = map;
328
329 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
330 struct arm32_bus_dma_cookie *cookie;
331 int cookieflags;
332 void *cookiestore;
333 size_t cookiesize;
334 int error;
335
336 cookieflags = 0;
337
338 if (t->_may_bounce != NULL) {
339 error = (*t->_may_bounce)(t, map, flags, &cookieflags);
340 if (error != 0)
341 goto out;
342 }
343
344 if (t->_ranges != NULL)
345 cookieflags |= _BUS_DMA_MIGHT_NEED_BOUNCE;
346
347 if ((cookieflags & _BUS_DMA_MIGHT_NEED_BOUNCE) == 0) {
348 STAT_INCR(creates);
349 return 0;
350 }
351
352 cookiesize = sizeof(struct arm32_bus_dma_cookie) +
353 (sizeof(bus_dma_segment_t) * map->_dm_segcnt);
354
355 /*
356 * Allocate our cookie.
357 */
358 if ((cookiestore = malloc(cookiesize, M_DMAMAP, mallocflags)) == NULL) {
359 error = ENOMEM;
360 goto out;
361 }
362 cookie = (struct arm32_bus_dma_cookie *)cookiestore;
363 cookie->id_flags = cookieflags;
364 map->_dm_cookie = cookie;
365 STAT_INCR(bounced_creates);
366
367 error = _bus_dma_alloc_bouncebuf(t, map, size, flags);
368 out:
369 if (error)
370 _bus_dmamap_destroy(t, map);
371 #else
372 STAT_INCR(creates);
373 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
374
375 #ifdef DEBUG_DMA
376 printf("dmamap_create:map=%p\n", map);
377 #endif /* DEBUG_DMA */
378 return (0);
379 }
380
381 /*
382 * Common function for DMA map destruction. May be called by bus-specific
383 * DMA map destruction functions.
384 */
385 void
386 _bus_dmamap_destroy(bus_dma_tag_t t, bus_dmamap_t map)
387 {
388
389 #ifdef DEBUG_DMA
390 printf("dmamap_destroy: t=%p map=%p\n", t, map);
391 #endif /* DEBUG_DMA */
392 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
393 struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
394
395 /*
396 * Free any bounce pages this map might hold.
397 */
398 if (cookie != NULL) {
399 if (cookie->id_flags & _BUS_DMA_IS_BOUNCING)
400 STAT_INCR(bounced_unloads);
401 map->dm_nsegs = 0;
402 if (cookie->id_flags & _BUS_DMA_HAS_BOUNCE)
403 _bus_dma_free_bouncebuf(t, map);
404 STAT_INCR(bounced_destroys);
405 free(cookie, M_DMAMAP);
406 } else
407 #endif
408 STAT_INCR(destroys);
409
410 if (map->dm_nsegs > 0)
411 STAT_INCR(unloads);
412
413 free(map, M_DMAMAP);
414 }
415
416 /*
417 * Common function for loading a DMA map with a linear buffer. May
418 * be called by bus-specific DMA map load functions.
419 */
420 int
421 _bus_dmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
422 bus_size_t buflen, struct proc *p, int flags)
423 {
424 struct vmspace *vm;
425 int error;
426
427 #ifdef DEBUG_DMA
428 printf("dmamap_load: t=%p map=%p buf=%p len=%lx p=%p f=%d\n",
429 t, map, buf, buflen, p, flags);
430 #endif /* DEBUG_DMA */
431
432 if (map->dm_nsegs > 0) {
433 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
434 struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
435 if (cookie != NULL) {
436 if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
437 STAT_INCR(bounced_unloads);
438 cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
439 }
440 } else
441 #endif
442 STAT_INCR(unloads);
443 }
444
445 /*
446 * Make sure that on error condition we return "no valid mappings".
447 */
448 map->dm_mapsize = 0;
449 map->dm_nsegs = 0;
450 map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
451 KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
452
453 if (buflen > map->_dm_size)
454 return (EINVAL);
455
456 if (p != NULL) {
457 vm = p->p_vmspace;
458 } else {
459 vm = vmspace_kernel();
460 }
461
462 /* _bus_dmamap_load_buffer() clears this if we're not... */
463 map->_dm_flags |= _BUS_DMAMAP_COHERENT;
464
465 error = _bus_dmamap_load_buffer(t, map, buf, buflen, vm, flags);
466 if (error == 0) {
467 map->dm_mapsize = buflen;
468 map->_dm_vmspace = vm;
469 map->_dm_origbuf = buf;
470 map->_dm_buftype = _BUS_DMA_BUFTYPE_LINEAR;
471 return 0;
472 }
473 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
474 struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
475 if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
476 error = _bus_dma_load_bouncebuf(t, map, buf, buflen,
477 _BUS_DMA_BUFTYPE_LINEAR, flags);
478 }
479 #endif
480 return (error);
481 }
482
483 /*
484 * Like _bus_dmamap_load(), but for mbufs.
485 */
486 int
487 _bus_dmamap_load_mbuf(bus_dma_tag_t t, bus_dmamap_t map, struct mbuf *m0,
488 int flags)
489 {
490 int error;
491 struct mbuf *m;
492
493 #ifdef DEBUG_DMA
494 printf("dmamap_load_mbuf: t=%p map=%p m0=%p f=%d\n",
495 t, map, m0, flags);
496 #endif /* DEBUG_DMA */
497
498 if (map->dm_nsegs > 0) {
499 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
500 struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
501 if (cookie != NULL) {
502 if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
503 STAT_INCR(bounced_unloads);
504 cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
505 }
506 } else
507 #endif
508 STAT_INCR(unloads);
509 }
510
511 /*
512 * Make sure that on error condition we return "no valid mappings."
513 */
514 map->dm_mapsize = 0;
515 map->dm_nsegs = 0;
516 map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
517 KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
518
519 #ifdef DIAGNOSTIC
520 if ((m0->m_flags & M_PKTHDR) == 0)
521 panic("_bus_dmamap_load_mbuf: no packet header");
522 #endif /* DIAGNOSTIC */
523
524 if (m0->m_pkthdr.len > map->_dm_size)
525 return (EINVAL);
526
527 /* _bus_dmamap_load_paddr() clears this if we're not... */
528 map->_dm_flags |= _BUS_DMAMAP_COHERENT;
529
530 error = 0;
531 for (m = m0; m != NULL && error == 0; m = m->m_next) {
532 int offset;
533 int remainbytes;
534 const struct vm_page * const *pgs;
535 paddr_t paddr;
536 int size;
537
538 if (m->m_len == 0)
539 continue;
540 /*
541 * Don't allow reads in read-only mbufs.
542 */
543 if (M_ROMAP(m) && (flags & BUS_DMA_READ)) {
544 error = EFAULT;
545 break;
546 }
547 switch (m->m_flags & (M_EXT|M_CLUSTER|M_EXT_PAGES)) {
548 case M_EXT|M_CLUSTER:
549 /* XXX KDASSERT */
550 KASSERT(m->m_ext.ext_paddr != M_PADDR_INVALID);
551 paddr = m->m_ext.ext_paddr +
552 (m->m_data - m->m_ext.ext_buf);
553 size = m->m_len;
554 error = _bus_dmamap_load_paddr(t, map, paddr, size,
555 false);
556 break;
557
558 case M_EXT|M_EXT_PAGES:
559 KASSERT(m->m_ext.ext_buf <= m->m_data);
560 KASSERT(m->m_data <=
561 m->m_ext.ext_buf + m->m_ext.ext_size);
562
563 offset = (vaddr_t)m->m_data -
564 trunc_page((vaddr_t)m->m_ext.ext_buf);
565 remainbytes = m->m_len;
566
567 /* skip uninteresting pages */
568 pgs = (const struct vm_page * const *)
569 m->m_ext.ext_pgs + (offset >> PAGE_SHIFT);
570
571 offset &= PAGE_MASK; /* offset in the first page */
572
573 /* load each page */
574 while (remainbytes > 0) {
575 const struct vm_page *pg;
576
577 size = MIN(remainbytes, PAGE_SIZE - offset);
578
579 pg = *pgs++;
580 KASSERT(pg);
581 paddr = VM_PAGE_TO_PHYS(pg) + offset;
582
583 error = _bus_dmamap_load_paddr(t, map,
584 paddr, size, false);
585 if (error)
586 break;
587 offset = 0;
588 remainbytes -= size;
589 }
590 break;
591
592 case 0:
593 paddr = m->m_paddr + M_BUFOFFSET(m) +
594 (m->m_data - M_BUFADDR(m));
595 size = m->m_len;
596 error = _bus_dmamap_load_paddr(t, map, paddr, size,
597 false);
598 break;
599
600 default:
601 error = _bus_dmamap_load_buffer(t, map, m->m_data,
602 m->m_len, vmspace_kernel(), flags);
603 }
604 }
605 if (error == 0) {
606 map->dm_mapsize = m0->m_pkthdr.len;
607 map->_dm_origbuf = m0;
608 map->_dm_buftype = _BUS_DMA_BUFTYPE_MBUF;
609 map->_dm_vmspace = vmspace_kernel(); /* always kernel */
610 return 0;
611 }
612 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
613 struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
614 if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
615 error = _bus_dma_load_bouncebuf(t, map, m0, m0->m_pkthdr.len,
616 _BUS_DMA_BUFTYPE_MBUF, flags);
617 }
618 #endif
619 return (error);
620 }
621
622 /*
623 * Like _bus_dmamap_load(), but for uios.
624 */
625 int
626 _bus_dmamap_load_uio(bus_dma_tag_t t, bus_dmamap_t map, struct uio *uio,
627 int flags)
628 {
629 int i, error;
630 bus_size_t minlen, resid;
631 struct iovec *iov;
632 void *addr;
633
634 /*
635 * Make sure that on error condition we return "no valid mappings."
636 */
637 map->dm_mapsize = 0;
638 map->dm_nsegs = 0;
639 KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
640
641 resid = uio->uio_resid;
642 iov = uio->uio_iov;
643
644 /* _bus_dmamap_load_buffer() clears this if we're not... */
645 map->_dm_flags |= _BUS_DMAMAP_COHERENT;
646
647 error = 0;
648 for (i = 0; i < uio->uio_iovcnt && resid != 0 && error == 0; i++) {
649 /*
650 * Now at the first iovec to load. Load each iovec
651 * until we have exhausted the residual count.
652 */
653 minlen = resid < iov[i].iov_len ? resid : iov[i].iov_len;
654 addr = (void *)iov[i].iov_base;
655
656 error = _bus_dmamap_load_buffer(t, map, addr, minlen,
657 uio->uio_vmspace, flags);
658
659 resid -= minlen;
660 }
661 if (error == 0) {
662 map->dm_mapsize = uio->uio_resid;
663 map->_dm_origbuf = uio;
664 map->_dm_buftype = _BUS_DMA_BUFTYPE_UIO;
665 map->_dm_vmspace = uio->uio_vmspace;
666 }
667 return (error);
668 }
669
670 /*
671 * Like _bus_dmamap_load(), but for raw memory allocated with
672 * bus_dmamem_alloc().
673 */
674 int
675 _bus_dmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
676 bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
677 {
678
679 panic("_bus_dmamap_load_raw: not implemented");
680 }
681
682 /*
683 * Common function for unloading a DMA map. May be called by
684 * bus-specific DMA map unload functions.
685 */
686 void
687 _bus_dmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
688 {
689
690 #ifdef DEBUG_DMA
691 printf("dmamap_unload: t=%p map=%p\n", t, map);
692 #endif /* DEBUG_DMA */
693
694 /*
695 * No resources to free; just mark the mappings as
696 * invalid.
697 */
698 map->dm_mapsize = 0;
699 map->dm_nsegs = 0;
700 map->_dm_origbuf = NULL;
701 map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
702 map->_dm_vmspace = NULL;
703 }
704
705 static void
706 _bus_dmamap_sync_segment(vaddr_t va, paddr_t pa, vsize_t len, int ops, bool readonly_p)
707 {
708 KASSERT((va & PAGE_MASK) == (pa & PAGE_MASK));
709 #if 0
710 printf("sync_segment: va=%#lx pa=%#lx len=%#lx ops=%#x ro=%d\n",
711 va, pa, len, ops, readonly_p);
712 #endif
713
714 switch (ops) {
715 case BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE:
716 if (!readonly_p) {
717 cpu_dcache_wbinv_range(va, len);
718 cpu_sdcache_wbinv_range(va, pa, len);
719 break;
720 }
721 /* FALLTHROUGH */
722
723 case BUS_DMASYNC_PREREAD: {
724 const size_t line_size = arm_dcache_align;
725 const size_t line_mask = arm_dcache_align_mask;
726 vsize_t misalignment = va & line_mask;
727 if (misalignment) {
728 va -= misalignment;
729 pa -= misalignment;
730 len += misalignment;
731 cpu_dcache_wbinv_range(va, line_size);
732 cpu_sdcache_wbinv_range(va, pa, line_size);
733 if (len <= line_size)
734 break;
735 va += line_size;
736 pa += line_size;
737 len -= line_size;
738 }
739 misalignment = len & line_mask;
740 len -= misalignment;
741 cpu_dcache_inv_range(va, len);
742 cpu_sdcache_inv_range(va, pa, len);
743 if (misalignment) {
744 va += len;
745 pa += len;
746 cpu_dcache_wbinv_range(va, line_size);
747 cpu_sdcache_wbinv_range(va, pa, line_size);
748 }
749 break;
750 }
751
752 case BUS_DMASYNC_PREWRITE:
753 cpu_dcache_wb_range(va, len);
754 cpu_sdcache_wb_range(va, pa, len);
755 break;
756 }
757 }
758
759 static inline void
760 _bus_dmamap_sync_linear(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
761 bus_size_t len, int ops)
762 {
763 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
764 struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
765 bool bouncing = (cookie != NULL && (cookie->id_flags & _BUS_DMA_IS_BOUNCING));
766 #endif
767 bus_dma_segment_t *ds = map->dm_segs;
768 vaddr_t va = (vaddr_t) map->_dm_origbuf;
769 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
770 if (bouncing) {
771 va = (vaddr_t) cookie->id_bouncebuf;
772 }
773 #endif
774
775 while (len > 0) {
776 while (offset >= ds->ds_len) {
777 offset -= ds->ds_len;
778 va += ds->ds_len;
779 ds++;
780 }
781
782 paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + offset);
783 size_t seglen = min(len, ds->ds_len - offset);
784
785 if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
786 _bus_dmamap_sync_segment(va + offset, pa, seglen, ops,
787 false);
788
789 offset += seglen;
790 len -= seglen;
791 }
792 }
793
794 static inline void
795 _bus_dmamap_sync_mbuf(bus_dma_tag_t t, bus_dmamap_t map, bus_size_t offset,
796 bus_size_t len, int ops)
797 {
798 bus_dma_segment_t *ds = map->dm_segs;
799 struct mbuf *m = map->_dm_origbuf;
800 bus_size_t voff = offset;
801 bus_size_t ds_off = offset;
802
803 while (len > 0) {
804 /* Find the current dma segment */
805 while (ds_off >= ds->ds_len) {
806 ds_off -= ds->ds_len;
807 ds++;
808 }
809 /* Find the current mbuf. */
810 while (voff >= m->m_len) {
811 voff -= m->m_len;
812 m = m->m_next;
813 }
814
815 /*
816 * Now at the first mbuf to sync; nail each one until
817 * we have exhausted the length.
818 */
819 vsize_t seglen = min(len, min(m->m_len - voff, ds->ds_len - ds_off));
820 vaddr_t va = mtod(m, vaddr_t) + voff;
821 paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
822
823 /*
824 * We can save a lot of work here if we know the mapping
825 * is read-only at the MMU:
826 *
827 * If a mapping is read-only, no dirty cache blocks will
828 * exist for it. If a writable mapping was made read-only,
829 * we know any dirty cache lines for the range will have
830 * been cleaned for us already. Therefore, if the upper
831 * layer can tell us we have a read-only mapping, we can
832 * skip all cache cleaning.
833 *
834 * NOTE: This only works if we know the pmap cleans pages
835 * before making a read-write -> read-only transition. If
836 * this ever becomes non-true (e.g. Physically Indexed
837 * cache), this will have to be revisited.
838 */
839
840 if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
841 _bus_dmamap_sync_segment(va, pa, seglen, ops,
842 M_ROMAP(m));
843 voff += seglen;
844 ds_off += seglen;
845 len -= seglen;
846 }
847 }
848
849 static inline void
850 _bus_dmamap_sync_uio(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
851 bus_size_t len, int ops)
852 {
853 bus_dma_segment_t *ds = map->dm_segs;
854 struct uio *uio = map->_dm_origbuf;
855 struct iovec *iov = uio->uio_iov;
856 bus_size_t voff = offset;
857 bus_size_t ds_off = offset;
858
859 while (len > 0) {
860 /* Find the current dma segment */
861 while (ds_off >= ds->ds_len) {
862 ds_off -= ds->ds_len;
863 ds++;
864 }
865
866 /* Find the current iovec. */
867 while (voff >= iov->iov_len) {
868 voff -= iov->iov_len;
869 iov++;
870 }
871
872 /*
873 * Now at the first iovec to sync; nail each one until
874 * we have exhausted the length.
875 */
876 vsize_t seglen = min(len, min(iov->iov_len - voff, ds->ds_len - ds_off));
877 vaddr_t va = (vaddr_t) iov->iov_base + voff;
878 paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
879
880 if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
881 _bus_dmamap_sync_segment(va, pa, seglen, ops, false);
882
883 voff += seglen;
884 ds_off += seglen;
885 len -= seglen;
886 }
887 }
888
889 /*
890 * Common function for DMA map synchronization. May be called
891 * by bus-specific DMA map synchronization functions.
892 *
893 * This version works for the Virtually Indexed Virtually Tagged
894 * cache found on 32-bit ARM processors.
895 *
896 * XXX Should have separate versions for write-through vs.
897 * XXX write-back caches. We currently assume write-back
898 * XXX here, which is not as efficient as it could be for
899 * XXX the write-through case.
900 */
901 void
902 _bus_dmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
903 bus_size_t len, int ops)
904 {
905 bool bouncing = false;
906
907 #ifdef DEBUG_DMA
908 printf("dmamap_sync: t=%p map=%p offset=%lx len=%lx ops=%x\n",
909 t, map, offset, len, ops);
910 #endif /* DEBUG_DMA */
911
912 /*
913 * Mixing of PRE and POST operations is not allowed.
914 */
915 if ((ops & (BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE)) != 0 &&
916 (ops & (BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE)) != 0)
917 panic("_bus_dmamap_sync: mix PRE and POST");
918
919 #ifdef DIAGNOSTIC
920 if (offset >= map->dm_mapsize)
921 panic("_bus_dmamap_sync: bad offset %lu (map size is %lu)",
922 offset, map->dm_mapsize);
923 if (len == 0 || (offset + len) > map->dm_mapsize)
924 panic("_bus_dmamap_sync: bad length");
925 #endif
926
927 /*
928 * For a virtually-indexed write-back cache, we need
929 * to do the following things:
930 *
931 * PREREAD -- Invalidate the D-cache. We do this
932 * here in case a write-back is required by the back-end.
933 *
934 * PREWRITE -- Write-back the D-cache. Note that if
935 * we are doing a PREREAD|PREWRITE, we can collapse
936 * the whole thing into a single Wb-Inv.
937 *
938 * POSTREAD -- Nothing.
939 *
940 * POSTWRITE -- Nothing.
941 */
942 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
943 struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
944 bouncing = (cookie != NULL && (cookie->id_flags & _BUS_DMA_IS_BOUNCING));
945 #endif
946
947 const int pre_ops = ops & (BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
948 if (!bouncing && pre_ops == 0) {
949 return;
950 }
951
952 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
953 if (bouncing && (ops & BUS_DMASYNC_PREWRITE)) {
954 STAT_INCR(write_bounces);
955 char * const dataptr = (char *)cookie->id_bouncebuf + offset;
956 /*
957 * Copy the caller's buffer to the bounce buffer.
958 */
959 switch (map->_dm_buftype) {
960 case _BUS_DMA_BUFTYPE_LINEAR:
961 memcpy(dataptr, cookie->id_origlinearbuf + offset, len);
962 break;
963 case _BUS_DMA_BUFTYPE_MBUF:
964 m_copydata(cookie->id_origmbuf, offset, len, dataptr);
965 break;
966 case _BUS_DMA_BUFTYPE_UIO:
967 _bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_WRITE);
968 break;
969 #ifdef DIAGNOSTIC
970 case _BUS_DMA_BUFTYPE_RAW:
971 panic("_bus_dmamap_sync(pre): _BUS_DMA_BUFTYPE_RAW");
972 break;
973
974 case _BUS_DMA_BUFTYPE_INVALID:
975 panic("_bus_dmamap_sync(pre): _BUS_DMA_BUFTYPE_INVALID");
976 break;
977
978 default:
979 panic("_bus_dmamap_sync(pre): map %p: unknown buffer type %d\n",
980 map, map->_dm_buftype);
981 break;
982 #endif /* DIAGNOSTIC */
983 }
984 }
985 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
986
987 /* Skip cache frobbing if mapping was COHERENT. */
988 if (!bouncing && (map->_dm_flags & _BUS_DMAMAP_COHERENT)) {
989 /* Drain the write buffer. */
990 cpu_drain_writebuf();
991 return;
992 }
993
994 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
995 if (bouncing && ((map->_dm_flags & _BUS_DMAMAP_COHERENT) || pre_ops == 0)) {
996 goto bounce_it;
997 }
998 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
999
1000 /*
1001 * If the mapping belongs to a non-kernel vmspace, and the
1002 * vmspace has not been active since the last time a full
1003 * cache flush was performed, we don't need to do anything.
1004 */
1005 if (__predict_false(!VMSPACE_IS_KERNEL_P(map->_dm_vmspace) &&
1006 vm_map_pmap(&map->_dm_vmspace->vm_map)->pm_cstate.cs_cache_d == 0))
1007 return;
1008
1009 int buftype = map->_dm_buftype;
1010 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1011 if (bouncing) {
1012 buftype = _BUS_DMA_BUFTYPE_LINEAR;
1013 }
1014 #endif
1015
1016 switch (buftype) {
1017 case _BUS_DMA_BUFTYPE_LINEAR:
1018 _bus_dmamap_sync_linear(t, map, offset, len, ops);
1019 break;
1020
1021 case _BUS_DMA_BUFTYPE_MBUF:
1022 _bus_dmamap_sync_mbuf(t, map, offset, len, ops);
1023 break;
1024
1025 case _BUS_DMA_BUFTYPE_UIO:
1026 _bus_dmamap_sync_uio(t, map, offset, len, ops);
1027 break;
1028
1029 case _BUS_DMA_BUFTYPE_RAW:
1030 panic("_bus_dmamap_sync: _BUS_DMA_BUFTYPE_RAW");
1031 break;
1032
1033 case _BUS_DMA_BUFTYPE_INVALID:
1034 panic("_bus_dmamap_sync: _BUS_DMA_BUFTYPE_INVALID");
1035 break;
1036
1037 default:
1038 panic("_bus_dmamap_sync: map %p: unknown buffer type %d\n",
1039 map, map->_dm_buftype);
1040 }
1041
1042 /* Drain the write buffer. */
1043 cpu_drain_writebuf();
1044
1045 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1046 bounce_it:
1047 if ((ops & BUS_DMASYNC_POSTREAD) == 0
1048 || cookie == NULL
1049 || (cookie->id_flags & _BUS_DMA_IS_BOUNCING) == 0)
1050 return;
1051
1052 char * const dataptr = (char *)cookie->id_bouncebuf + offset;
1053 STAT_INCR(read_bounces);
1054 /*
1055 * Copy the bounce buffer to the caller's buffer.
1056 */
1057 switch (map->_dm_buftype) {
1058 case _BUS_DMA_BUFTYPE_LINEAR:
1059 memcpy(cookie->id_origlinearbuf + offset, dataptr, len);
1060 break;
1061
1062 case _BUS_DMA_BUFTYPE_MBUF:
1063 m_copyback(cookie->id_origmbuf, offset, len, dataptr);
1064 break;
1065
1066 case _BUS_DMA_BUFTYPE_UIO:
1067 _bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_READ);
1068 break;
1069 #ifdef DIAGNOSTIC
1070 case _BUS_DMA_BUFTYPE_RAW:
1071 panic("_bus_dmamap_sync(post): _BUS_DMA_BUFTYPE_RAW");
1072 break;
1073
1074 case _BUS_DMA_BUFTYPE_INVALID:
1075 panic("_bus_dmamap_sync(post): _BUS_DMA_BUFTYPE_INVALID");
1076 break;
1077
1078 default:
1079 panic("_bus_dmamap_sync(post): map %p: unknown buffer type %d\n",
1080 map, map->_dm_buftype);
1081 break;
1082 #endif
1083 }
1084 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1085 }
1086
1087 /*
1088 * Common function for DMA-safe memory allocation. May be called
1089 * by bus-specific DMA memory allocation functions.
1090 */
1091
1092 extern paddr_t physical_start;
1093 extern paddr_t physical_end;
1094
1095 int
1096 _bus_dmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1097 bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1098 int flags)
1099 {
1100 struct arm32_dma_range *dr;
1101 int error, i;
1102
1103 #ifdef DEBUG_DMA
1104 printf("dmamem_alloc t=%p size=%lx align=%lx boundary=%lx "
1105 "segs=%p nsegs=%x rsegs=%p flags=%x\n", t, size, alignment,
1106 boundary, segs, nsegs, rsegs, flags);
1107 #endif
1108
1109 if ((dr = t->_ranges) != NULL) {
1110 error = ENOMEM;
1111 for (i = 0; i < t->_nranges; i++, dr++) {
1112 if (dr->dr_len == 0)
1113 continue;
1114 error = _bus_dmamem_alloc_range(t, size, alignment,
1115 boundary, segs, nsegs, rsegs, flags,
1116 trunc_page(dr->dr_sysbase),
1117 trunc_page(dr->dr_sysbase + dr->dr_len));
1118 if (error == 0)
1119 break;
1120 }
1121 } else {
1122 error = _bus_dmamem_alloc_range(t, size, alignment, boundary,
1123 segs, nsegs, rsegs, flags, trunc_page(physical_start),
1124 trunc_page(physical_end));
1125 }
1126
1127 #ifdef DEBUG_DMA
1128 printf("dmamem_alloc: =%d\n", error);
1129 #endif
1130
1131 return(error);
1132 }
1133
1134 /*
1135 * Common function for freeing DMA-safe memory. May be called by
1136 * bus-specific DMA memory free functions.
1137 */
1138 void
1139 _bus_dmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
1140 {
1141 struct vm_page *m;
1142 bus_addr_t addr;
1143 struct pglist mlist;
1144 int curseg;
1145
1146 #ifdef DEBUG_DMA
1147 printf("dmamem_free: t=%p segs=%p nsegs=%x\n", t, segs, nsegs);
1148 #endif /* DEBUG_DMA */
1149
1150 /*
1151 * Build a list of pages to free back to the VM system.
1152 */
1153 TAILQ_INIT(&mlist);
1154 for (curseg = 0; curseg < nsegs; curseg++) {
1155 for (addr = segs[curseg].ds_addr;
1156 addr < (segs[curseg].ds_addr + segs[curseg].ds_len);
1157 addr += PAGE_SIZE) {
1158 m = PHYS_TO_VM_PAGE(addr);
1159 TAILQ_INSERT_TAIL(&mlist, m, pageq.queue);
1160 }
1161 }
1162 uvm_pglistfree(&mlist);
1163 }
1164
1165 /*
1166 * Common function for mapping DMA-safe memory. May be called by
1167 * bus-specific DMA memory map functions.
1168 */
1169 int
1170 _bus_dmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1171 size_t size, void **kvap, int flags)
1172 {
1173 vaddr_t va;
1174 paddr_t pa;
1175 int curseg;
1176 pt_entry_t *ptep/*, pte*/;
1177 const uvm_flag_t kmflags =
1178 (flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0;
1179
1180 #ifdef DEBUG_DMA
1181 printf("dmamem_map: t=%p segs=%p nsegs=%x size=%lx flags=%x\n", t,
1182 segs, nsegs, (unsigned long)size, flags);
1183 #endif /* DEBUG_DMA */
1184
1185 #ifdef PMAP_MAP_POOLPAGE
1186 /*
1187 * If all of memory is mapped, and we are mapping a single physically
1188 * contiguous area then this area is already mapped. Let's see if we
1189 * avoid having a separate mapping for it.
1190 */
1191 if (nsegs == 1) {
1192 paddr_t paddr = segs[0].ds_addr;
1193 /*
1194 * If this is a non-COHERENT mapping, then the existing kernel
1195 * mapping is already compatible with it.
1196 */
1197 if ((flags & _BUS_DMAMAP_COHERENT) == 0) {
1198 *kvap = (void *)PMAP_MAP_POOLPAGE(paddr);
1199 return 0;
1200 }
1201 /*
1202 * This is a COHERENT mapping, which unless this address is in
1203 * a COHERENT dma range, will not be compatible.
1204 */
1205 if (t->_ranges != NULL) {
1206 const struct arm32_dma_range * const dr =
1207 _bus_dma_paddr_inrange(t->_ranges, t->_nranges,
1208 paddr);
1209 if (dr != NULL
1210 && (dr->dr_flags & _BUS_DMAMAP_COHERENT) != 0) {
1211 *kvap = (void *)PMAP_MAP_POOLPAGE(paddr);
1212 return 0;
1213 }
1214 }
1215 }
1216 #endif
1217
1218 size = round_page(size);
1219 va = uvm_km_alloc(kernel_map, size, 0, UVM_KMF_VAONLY | kmflags);
1220
1221 if (va == 0)
1222 return (ENOMEM);
1223
1224 *kvap = (void *)va;
1225
1226 for (curseg = 0; curseg < nsegs; curseg++) {
1227 for (pa = segs[curseg].ds_addr;
1228 pa < (segs[curseg].ds_addr + segs[curseg].ds_len);
1229 pa += PAGE_SIZE, va += PAGE_SIZE, size -= PAGE_SIZE) {
1230 #ifdef DEBUG_DMA
1231 printf("wiring p%lx to v%lx", pa, va);
1232 #endif /* DEBUG_DMA */
1233 if (size == 0)
1234 panic("_bus_dmamem_map: size botch");
1235 pmap_enter(pmap_kernel(), va, pa,
1236 VM_PROT_READ | VM_PROT_WRITE,
1237 VM_PROT_READ | VM_PROT_WRITE | PMAP_WIRED);
1238
1239 /*
1240 * If the memory must remain coherent with the
1241 * cache then we must make the memory uncacheable
1242 * in order to maintain virtual cache coherency.
1243 * We must also guarantee the cache does not already
1244 * contain the virtal addresses we are making
1245 * uncacheable.
1246 */
1247
1248 bool uncached = (flags & BUS_DMA_COHERENT);
1249 if (uncached) {
1250 const struct arm32_dma_range * const dr =
1251 _bus_dma_paddr_inrange(t->_ranges,
1252 t->_nranges, pa);
1253 /*
1254 * If this dma region is coherent then there is
1255 * no need for an uncached mapping.
1256 */
1257 if (dr != NULL
1258 && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
1259 uncached = false;
1260 }
1261 }
1262 if (uncached) {
1263 cpu_dcache_wbinv_range(va, PAGE_SIZE);
1264 cpu_sdcache_wbinv_range(va, pa, PAGE_SIZE);
1265 cpu_drain_writebuf();
1266 ptep = vtopte(va);
1267 *ptep &= ~L2_S_CACHE_MASK;
1268 PTE_SYNC(ptep);
1269 tlb_flush();
1270 }
1271 #ifdef DEBUG_DMA
1272 ptep = vtopte(va);
1273 printf(" pte=v%p *pte=%x\n", ptep, *ptep);
1274 #endif /* DEBUG_DMA */
1275 }
1276 }
1277 pmap_update(pmap_kernel());
1278 #ifdef DEBUG_DMA
1279 printf("dmamem_map: =%p\n", *kvap);
1280 #endif /* DEBUG_DMA */
1281 return (0);
1282 }
1283
1284 /*
1285 * Common function for unmapping DMA-safe memory. May be called by
1286 * bus-specific DMA memory unmapping functions.
1287 */
1288 void
1289 _bus_dmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
1290 {
1291
1292 #ifdef DEBUG_DMA
1293 printf("dmamem_unmap: t=%p kva=%p size=%lx\n", t, kva,
1294 (unsigned long)size);
1295 #endif /* DEBUG_DMA */
1296 #ifdef DIAGNOSTIC
1297 if ((u_long)kva & PGOFSET)
1298 panic("_bus_dmamem_unmap");
1299 #endif /* DIAGNOSTIC */
1300
1301 size = round_page(size);
1302 pmap_remove(pmap_kernel(), (vaddr_t)kva, (vaddr_t)kva + size);
1303 pmap_update(pmap_kernel());
1304 uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
1305 }
1306
1307 /*
1308 * Common functin for mmap(2)'ing DMA-safe memory. May be called by
1309 * bus-specific DMA mmap(2)'ing functions.
1310 */
1311 paddr_t
1312 _bus_dmamem_mmap(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1313 off_t off, int prot, int flags)
1314 {
1315 int i;
1316
1317 for (i = 0; i < nsegs; i++) {
1318 #ifdef DIAGNOSTIC
1319 if (off & PGOFSET)
1320 panic("_bus_dmamem_mmap: offset unaligned");
1321 if (segs[i].ds_addr & PGOFSET)
1322 panic("_bus_dmamem_mmap: segment unaligned");
1323 if (segs[i].ds_len & PGOFSET)
1324 panic("_bus_dmamem_mmap: segment size not multiple"
1325 " of page size");
1326 #endif /* DIAGNOSTIC */
1327 if (off >= segs[i].ds_len) {
1328 off -= segs[i].ds_len;
1329 continue;
1330 }
1331
1332 return (arm_btop((u_long)segs[i].ds_addr + off));
1333 }
1334
1335 /* Page not found. */
1336 return (-1);
1337 }
1338
1339 /**********************************************************************
1340 * DMA utility functions
1341 **********************************************************************/
1342
1343 /*
1344 * Utility function to load a linear buffer. lastaddrp holds state
1345 * between invocations (for multiple-buffer loads). segp contains
1346 * the starting segment on entrace, and the ending segment on exit.
1347 * first indicates if this is the first invocation of this function.
1348 */
1349 int
1350 _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
1351 bus_size_t buflen, struct vmspace *vm, int flags)
1352 {
1353 bus_size_t sgsize;
1354 bus_addr_t curaddr;
1355 vaddr_t vaddr = (vaddr_t)buf;
1356 int error;
1357 pmap_t pmap;
1358
1359 #ifdef DEBUG_DMA
1360 printf("_bus_dmamem_load_buffer(buf=%p, len=%lx, flags=%d)\n",
1361 buf, buflen, flags);
1362 #endif /* DEBUG_DMA */
1363
1364 pmap = vm_map_pmap(&vm->vm_map);
1365
1366 while (buflen > 0) {
1367 /*
1368 * Get the physical address for this segment.
1369 *
1370 * XXX Doesn't support checking for coherent mappings
1371 * XXX in user address space.
1372 */
1373 bool coherent;
1374 if (__predict_true(pmap == pmap_kernel())) {
1375 pd_entry_t *pde;
1376 pt_entry_t *ptep;
1377 (void) pmap_get_pde_pte(pmap, vaddr, &pde, &ptep);
1378 if (__predict_false(pmap_pde_section(pde))) {
1379 paddr_t s_frame = L1_S_FRAME;
1380 paddr_t s_offset = L1_S_OFFSET;
1381 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
1382 if (__predict_false(pmap_pde_supersection(pde))) {
1383 s_frame = L1_SS_FRAME;
1384 s_offset = L1_SS_OFFSET;
1385 }
1386 #endif
1387 curaddr = (*pde & s_frame) | (vaddr & s_offset);
1388 coherent = (*pde & L1_S_CACHE_MASK) != 0;
1389 } else {
1390 pt_entry_t pte = *ptep;
1391 KDASSERT((pte & L2_TYPE_MASK) != L2_TYPE_INV);
1392 if (__predict_false((pte & L2_TYPE_MASK)
1393 == L2_TYPE_L)) {
1394 curaddr = (pte & L2_L_FRAME) |
1395 (vaddr & L2_L_OFFSET);
1396 coherent = (pte & L2_L_CACHE_MASK) != 0;
1397 } else {
1398 curaddr = (pte & L2_S_FRAME) |
1399 (vaddr & L2_S_OFFSET);
1400 coherent = (pte & L2_S_CACHE_MASK) != 0;
1401 }
1402 }
1403 } else {
1404 (void) pmap_extract(pmap, vaddr, &curaddr);
1405 coherent = false;
1406 }
1407
1408 /*
1409 * Compute the segment size, and adjust counts.
1410 */
1411 sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
1412 if (buflen < sgsize)
1413 sgsize = buflen;
1414
1415 error = _bus_dmamap_load_paddr(t, map, curaddr, sgsize,
1416 coherent);
1417 if (error)
1418 return (error);
1419
1420 vaddr += sgsize;
1421 buflen -= sgsize;
1422 }
1423
1424 return (0);
1425 }
1426
1427 /*
1428 * Allocate physical memory from the given physical address range.
1429 * Called by DMA-safe memory allocation methods.
1430 */
1431 int
1432 _bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1433 bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1434 int flags, paddr_t low, paddr_t high)
1435 {
1436 paddr_t curaddr, lastaddr;
1437 struct vm_page *m;
1438 struct pglist mlist;
1439 int curseg, error;
1440
1441 #ifdef DEBUG_DMA
1442 printf("alloc_range: t=%p size=%lx align=%lx boundary=%lx segs=%p nsegs=%x rsegs=%p flags=%x lo=%lx hi=%lx\n",
1443 t, size, alignment, boundary, segs, nsegs, rsegs, flags, low, high);
1444 #endif /* DEBUG_DMA */
1445
1446 /* Always round the size. */
1447 size = round_page(size);
1448
1449 /*
1450 * Allocate pages from the VM system.
1451 */
1452 error = uvm_pglistalloc(size, low, high, alignment, boundary,
1453 &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
1454 if (error)
1455 return (error);
1456
1457 /*
1458 * Compute the location, size, and number of segments actually
1459 * returned by the VM code.
1460 */
1461 m = TAILQ_FIRST(&mlist);
1462 curseg = 0;
1463 lastaddr = segs[curseg].ds_addr = VM_PAGE_TO_PHYS(m);
1464 segs[curseg].ds_len = PAGE_SIZE;
1465 #ifdef DEBUG_DMA
1466 printf("alloc: page %lx\n", lastaddr);
1467 #endif /* DEBUG_DMA */
1468 m = TAILQ_NEXT(m, pageq.queue);
1469
1470 for (; m != NULL; m = TAILQ_NEXT(m, pageq.queue)) {
1471 curaddr = VM_PAGE_TO_PHYS(m);
1472 #ifdef DIAGNOSTIC
1473 if (curaddr < low || curaddr >= high) {
1474 printf("uvm_pglistalloc returned non-sensical"
1475 " address 0x%lx\n", curaddr);
1476 panic("_bus_dmamem_alloc_range");
1477 }
1478 #endif /* DIAGNOSTIC */
1479 #ifdef DEBUG_DMA
1480 printf("alloc: page %lx\n", curaddr);
1481 #endif /* DEBUG_DMA */
1482 if (curaddr == (lastaddr + PAGE_SIZE))
1483 segs[curseg].ds_len += PAGE_SIZE;
1484 else {
1485 curseg++;
1486 segs[curseg].ds_addr = curaddr;
1487 segs[curseg].ds_len = PAGE_SIZE;
1488 }
1489 lastaddr = curaddr;
1490 }
1491
1492 *rsegs = curseg + 1;
1493
1494 return (0);
1495 }
1496
1497 /*
1498 * Check if a memory region intersects with a DMA range, and return the
1499 * page-rounded intersection if it does.
1500 */
1501 int
1502 arm32_dma_range_intersect(struct arm32_dma_range *ranges, int nranges,
1503 paddr_t pa, psize_t size, paddr_t *pap, psize_t *sizep)
1504 {
1505 struct arm32_dma_range *dr;
1506 int i;
1507
1508 if (ranges == NULL)
1509 return (0);
1510
1511 for (i = 0, dr = ranges; i < nranges; i++, dr++) {
1512 if (dr->dr_sysbase <= pa &&
1513 pa < (dr->dr_sysbase + dr->dr_len)) {
1514 /*
1515 * Beginning of region intersects with this range.
1516 */
1517 *pap = trunc_page(pa);
1518 *sizep = round_page(min(pa + size,
1519 dr->dr_sysbase + dr->dr_len) - pa);
1520 return (1);
1521 }
1522 if (pa < dr->dr_sysbase && dr->dr_sysbase < (pa + size)) {
1523 /*
1524 * End of region intersects with this range.
1525 */
1526 *pap = trunc_page(dr->dr_sysbase);
1527 *sizep = round_page(min((pa + size) - dr->dr_sysbase,
1528 dr->dr_len));
1529 return (1);
1530 }
1531 }
1532
1533 /* No intersection found. */
1534 return (0);
1535 }
1536
1537 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1538 static int
1539 _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
1540 bus_size_t size, int flags)
1541 {
1542 struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
1543 int error = 0;
1544
1545 #ifdef DIAGNOSTIC
1546 if (cookie == NULL)
1547 panic("_bus_dma_alloc_bouncebuf: no cookie");
1548 #endif
1549
1550 cookie->id_bouncebuflen = round_page(size);
1551 error = _bus_dmamem_alloc(t, cookie->id_bouncebuflen,
1552 PAGE_SIZE, map->_dm_boundary, cookie->id_bouncesegs,
1553 map->_dm_segcnt, &cookie->id_nbouncesegs, flags);
1554 if (error)
1555 goto out;
1556 error = _bus_dmamem_map(t, cookie->id_bouncesegs,
1557 cookie->id_nbouncesegs, cookie->id_bouncebuflen,
1558 (void **)&cookie->id_bouncebuf, flags);
1559
1560 out:
1561 if (error) {
1562 _bus_dmamem_free(t, cookie->id_bouncesegs,
1563 cookie->id_nbouncesegs);
1564 cookie->id_bouncebuflen = 0;
1565 cookie->id_nbouncesegs = 0;
1566 } else {
1567 cookie->id_flags |= _BUS_DMA_HAS_BOUNCE;
1568 }
1569
1570 return (error);
1571 }
1572
1573 static void
1574 _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map)
1575 {
1576 struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
1577
1578 #ifdef DIAGNOSTIC
1579 if (cookie == NULL)
1580 panic("_bus_dma_alloc_bouncebuf: no cookie");
1581 #endif
1582
1583 _bus_dmamem_unmap(t, cookie->id_bouncebuf, cookie->id_bouncebuflen);
1584 _bus_dmamem_free(t, cookie->id_bouncesegs,
1585 cookie->id_nbouncesegs);
1586 cookie->id_bouncebuflen = 0;
1587 cookie->id_nbouncesegs = 0;
1588 cookie->id_flags &= ~_BUS_DMA_HAS_BOUNCE;
1589 }
1590
1591 /*
1592 * This function does the same as uiomove, but takes an explicit
1593 * direction, and does not update the uio structure.
1594 */
1595 static int
1596 _bus_dma_uiomove(void *buf, struct uio *uio, size_t n, int direction)
1597 {
1598 struct iovec *iov;
1599 int error;
1600 struct vmspace *vm;
1601 char *cp;
1602 size_t resid, cnt;
1603 int i;
1604
1605 iov = uio->uio_iov;
1606 vm = uio->uio_vmspace;
1607 cp = buf;
1608 resid = n;
1609
1610 for (i = 0; i < uio->uio_iovcnt && resid > 0; i++) {
1611 iov = &uio->uio_iov[i];
1612 if (iov->iov_len == 0)
1613 continue;
1614 cnt = MIN(resid, iov->iov_len);
1615
1616 if (!VMSPACE_IS_KERNEL_P(vm) &&
1617 (curlwp->l_cpu->ci_schedstate.spc_flags & SPCF_SHOULDYIELD)
1618 != 0) {
1619 preempt();
1620 }
1621 if (direction == UIO_READ) {
1622 error = copyout_vmspace(vm, cp, iov->iov_base, cnt);
1623 } else {
1624 error = copyin_vmspace(vm, iov->iov_base, cp, cnt);
1625 }
1626 if (error)
1627 return (error);
1628 cp += cnt;
1629 resid -= cnt;
1630 }
1631 return (0);
1632 }
1633 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1634
1635 int
1636 _bus_dmatag_subregion(bus_dma_tag_t tag, bus_addr_t min_addr,
1637 bus_addr_t max_addr, bus_dma_tag_t *newtag, int flags)
1638 {
1639
1640 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1641 struct arm32_dma_range *dr;
1642 bool subset = false;
1643 size_t nranges = 0;
1644 size_t i;
1645 for (i = 0, dr = tag->_ranges; i < tag->_nranges; i++, dr++) {
1646 if (dr->dr_sysbase <= min_addr
1647 && max_addr <= dr->dr_sysbase + dr->dr_len - 1) {
1648 subset = true;
1649 }
1650 if (min_addr <= dr->dr_sysbase + dr->dr_len
1651 && max_addr >= dr->dr_sysbase) {
1652 nranges++;
1653 }
1654 }
1655 if (subset) {
1656 *newtag = tag;
1657 /* if the tag must be freed, add a reference */
1658 if (tag->_tag_needs_free)
1659 (tag->_tag_needs_free)++;
1660 return 0;
1661 }
1662 if (nranges == 0) {
1663 nranges = 1;
1664 }
1665
1666 size_t mallocsize = sizeof(*tag) + nranges * sizeof(*dr);
1667 if ((*newtag = malloc(mallocsize, M_DMAMAP,
1668 (flags & BUS_DMA_NOWAIT) ? M_NOWAIT : M_WAITOK)) == NULL)
1669 return ENOMEM;
1670
1671 dr = (void *)(*newtag + 1);
1672 **newtag = *tag;
1673 (*newtag)->_tag_needs_free = 1;
1674 (*newtag)->_ranges = dr;
1675 (*newtag)->_nranges = nranges;
1676
1677 if (tag->_ranges == NULL) {
1678 dr->dr_sysbase = min_addr;
1679 dr->dr_busbase = min_addr;
1680 dr->dr_len = max_addr + 1 - min_addr;
1681 } else {
1682 for (i = 0; i < nranges; i++) {
1683 if (min_addr > dr->dr_sysbase + dr->dr_len
1684 || max_addr < dr->dr_sysbase)
1685 continue;
1686 dr[0] = tag->_ranges[i];
1687 if (dr->dr_sysbase < min_addr) {
1688 psize_t diff = min_addr - dr->dr_sysbase;
1689 dr->dr_busbase += diff;
1690 dr->dr_len -= diff;
1691 dr->dr_sysbase += diff;
1692 }
1693 if (max_addr != 0xffffffff
1694 && max_addr + 1 < dr->dr_sysbase + dr->dr_len) {
1695 dr->dr_len = max_addr + 1 - dr->dr_sysbase;
1696 }
1697 dr++;
1698 }
1699 }
1700
1701 return 0;
1702 #else
1703 return EOPNOTSUPP;
1704 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1705 }
1706
1707 void
1708 _bus_dmatag_destroy(bus_dma_tag_t tag)
1709 {
1710 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1711 switch (tag->_tag_needs_free) {
1712 case 0:
1713 break; /* not allocated with malloc */
1714 case 1:
1715 free(tag, M_DMAMAP); /* last reference to tag */
1716 break;
1717 default:
1718 (tag->_tag_needs_free)--; /* one less reference */
1719 }
1720 #endif
1721 }
1722