Annotation of src/sys/arch/arm/arm32/bus_dma.c, Revision 1.113
1.113 ! skrll 1: /* $NetBSD: bus_dma.c,v 1.112 2018/09/03 16:29:23 riastradh Exp $ */
1.1 chris 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: */
1.33 lukem 32:
1.35 rearnsha 33: #define _ARM32_BUS_DMA_PRIVATE
34:
1.81 matt 35: #include "opt_arm_bus_space.h"
1.107 ryo 36: #include "opt_cputypes.h"
1.81 matt 37:
1.33 lukem 38: #include <sys/cdefs.h>
1.113 ! skrll 39: __KERNEL_RCSID(0, "$NetBSD: bus_dma.c,v 1.112 2018/09/03 16:29:23 riastradh Exp $");
1.1 chris 40:
41: #include <sys/param.h>
1.84 matt 42: #include <sys/bus.h>
43: #include <sys/cpu.h>
1.81 matt 44: #include <sys/kmem.h>
1.1 chris 45: #include <sys/mbuf.h>
46:
1.53 uebayasi 47: #include <uvm/uvm.h>
1.1 chris 48:
1.107 ryo 49: #include <arm/cpuconf.h>
1.84 matt 50: #include <arm/cpufunc.h>
1.4 thorpej 51:
1.84 matt 52: #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
53: #include <dev/mm.h>
54: #endif
1.1 chris 55:
1.76 matt 56: #ifdef BUSDMA_COUNTERS
1.58 matt 57: static struct evcnt bus_dma_creates =
58: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "creates");
59: static struct evcnt bus_dma_bounced_creates =
60: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced creates");
61: static struct evcnt bus_dma_loads =
62: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "loads");
63: static struct evcnt bus_dma_bounced_loads =
64: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced loads");
1.81 matt 65: static struct evcnt bus_dma_coherent_loads =
66: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "coherent loads");
1.58 matt 67: static struct evcnt bus_dma_read_bounces =
68: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "read bounces");
69: static struct evcnt bus_dma_write_bounces =
70: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "write bounces");
71: static struct evcnt bus_dma_bounced_unloads =
72: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced unloads");
73: static struct evcnt bus_dma_unloads =
74: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "unloads");
75: static struct evcnt bus_dma_bounced_destroys =
76: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced destroys");
77: static struct evcnt bus_dma_destroys =
78: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "destroys");
1.95 skrll 79: static struct evcnt bus_dma_sync_prereadwrite =
1.76 matt 80: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync prereadwrite");
81: static struct evcnt bus_dma_sync_preread_begin =
82: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync preread begin");
83: static struct evcnt bus_dma_sync_preread =
84: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync preread");
85: static struct evcnt bus_dma_sync_preread_tail =
86: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync preread tail");
1.95 skrll 87: static struct evcnt bus_dma_sync_prewrite =
1.76 matt 88: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync prewrite");
1.95 skrll 89: static struct evcnt bus_dma_sync_postread =
1.76 matt 90: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postread");
1.95 skrll 91: static struct evcnt bus_dma_sync_postreadwrite =
1.76 matt 92: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postreadwrite");
1.95 skrll 93: static struct evcnt bus_dma_sync_postwrite =
1.76 matt 94: EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postwrite");
1.58 matt 95:
96: EVCNT_ATTACH_STATIC(bus_dma_creates);
97: EVCNT_ATTACH_STATIC(bus_dma_bounced_creates);
98: EVCNT_ATTACH_STATIC(bus_dma_loads);
99: EVCNT_ATTACH_STATIC(bus_dma_bounced_loads);
1.81 matt 100: EVCNT_ATTACH_STATIC(bus_dma_coherent_loads);
1.58 matt 101: EVCNT_ATTACH_STATIC(bus_dma_read_bounces);
102: EVCNT_ATTACH_STATIC(bus_dma_write_bounces);
103: EVCNT_ATTACH_STATIC(bus_dma_unloads);
104: EVCNT_ATTACH_STATIC(bus_dma_bounced_unloads);
105: EVCNT_ATTACH_STATIC(bus_dma_destroys);
106: EVCNT_ATTACH_STATIC(bus_dma_bounced_destroys);
1.76 matt 107: EVCNT_ATTACH_STATIC(bus_dma_sync_prereadwrite);
108: EVCNT_ATTACH_STATIC(bus_dma_sync_preread_begin);
109: EVCNT_ATTACH_STATIC(bus_dma_sync_preread);
110: EVCNT_ATTACH_STATIC(bus_dma_sync_preread_tail);
111: EVCNT_ATTACH_STATIC(bus_dma_sync_prewrite);
112: EVCNT_ATTACH_STATIC(bus_dma_sync_postread);
113: EVCNT_ATTACH_STATIC(bus_dma_sync_postreadwrite);
114: EVCNT_ATTACH_STATIC(bus_dma_sync_postwrite);
1.58 matt 115:
116: #define STAT_INCR(x) (bus_dma_ ## x.ev_count++)
1.76 matt 117: #else
1.107 ryo 118: #define STAT_INCR(x) __nothing
1.76 matt 119: #endif
1.58 matt 120:
1.7 thorpej 121: int _bus_dmamap_load_buffer(bus_dma_tag_t, bus_dmamap_t, void *,
1.48 yamt 122: bus_size_t, struct vmspace *, int);
1.1 chris 123:
124: /*
1.19 briggs 125: * Check to see if the specified page is in an allowed DMA range.
126: */
1.105 skrll 127: static inline struct arm32_dma_range *
1.59 matt 128: _bus_dma_paddr_inrange(struct arm32_dma_range *ranges, int nranges,
1.19 briggs 129: bus_addr_t curaddr)
130: {
131: struct arm32_dma_range *dr;
132: int i;
133:
134: for (i = 0, dr = ranges; i < nranges; i++, dr++) {
135: if (curaddr >= dr->dr_sysbase &&
1.82 skrll 136: curaddr < (dr->dr_sysbase + dr->dr_len))
1.100 skrll 137: return dr;
1.19 briggs 138: }
139:
1.100 skrll 140: return NULL;
1.19 briggs 141: }
142:
143: /*
1.59 matt 144: * Check to see if the specified busaddr is in an allowed DMA range.
145: */
146: static inline paddr_t
147: _bus_dma_busaddr_to_paddr(bus_dma_tag_t t, bus_addr_t curaddr)
148: {
149: struct arm32_dma_range *dr;
150: u_int i;
151:
152: if (t->_nranges == 0)
153: return curaddr;
154:
155: for (i = 0, dr = t->_ranges; i < t->_nranges; i++, dr++) {
156: if (dr->dr_busbase <= curaddr
1.82 skrll 157: && curaddr < dr->dr_busbase + dr->dr_len)
1.59 matt 158: return curaddr - dr->dr_busbase + dr->dr_sysbase;
159: }
160: panic("%s: curaddr %#lx not in range", __func__, curaddr);
161: }
162:
163: /*
1.41 thorpej 164: * Common function to load the specified physical address into the
165: * DMA map, coalescing segments and boundary checking as necessary.
166: */
167: static int
168: _bus_dmamap_load_paddr(bus_dma_tag_t t, bus_dmamap_t map,
1.61 matt 169: bus_addr_t paddr, bus_size_t size, bool coherent)
1.41 thorpej 170: {
171: bus_dma_segment_t * const segs = map->dm_segs;
172: int nseg = map->dm_nsegs;
1.58 matt 173: bus_addr_t lastaddr;
1.41 thorpej 174: bus_addr_t bmask = ~(map->_dm_boundary - 1);
175: bus_addr_t curaddr;
176: bus_size_t sgsize;
1.61 matt 177: uint32_t _ds_flags = coherent ? _BUS_DMAMAP_COHERENT : 0;
1.41 thorpej 178:
179: if (nseg > 0)
1.101 skrll 180: lastaddr = segs[nseg - 1].ds_addr + segs[nseg - 1].ds_len;
1.58 matt 181: else
182: lastaddr = 0xdead;
1.95 skrll 183:
1.41 thorpej 184: again:
185: sgsize = size;
186:
187: /* Make sure we're in an allowed DMA range. */
188: if (t->_ranges != NULL) {
189: /* XXX cache last result? */
190: const struct arm32_dma_range * const dr =
1.59 matt 191: _bus_dma_paddr_inrange(t->_ranges, t->_nranges, paddr);
1.41 thorpej 192: if (dr == NULL)
1.100 skrll 193: return EINVAL;
1.61 matt 194:
195: /*
196: * If this region is coherent, mark the segment as coherent.
197: */
198: _ds_flags |= dr->dr_flags & _BUS_DMAMAP_COHERENT;
1.72 skrll 199:
1.41 thorpej 200: /*
201: * In a valid DMA range. Translate the physical
202: * memory address to an address in the DMA window.
203: */
204: curaddr = (paddr - dr->dr_sysbase) + dr->dr_busbase;
1.72 skrll 205: #if 0
206: printf("%p: %#lx: range %#lx/%#lx/%#lx/%#x: %#x <-- %#lx\n",
207: t, paddr, dr->dr_sysbase, dr->dr_busbase,
208: dr->dr_len, dr->dr_flags, _ds_flags, curaddr);
209: #endif
1.41 thorpej 210: } else
211: curaddr = paddr;
212:
213: /*
214: * Make sure we don't cross any boundaries.
215: */
216: if (map->_dm_boundary > 0) {
217: bus_addr_t baddr; /* next boundary address */
218:
219: baddr = (curaddr + map->_dm_boundary) & bmask;
220: if (sgsize > (baddr - curaddr))
221: sgsize = (baddr - curaddr);
222: }
223:
224: /*
225: * Insert chunk into a segment, coalescing with the
226: * previous segment if possible.
227: */
228: if (nseg > 0 && curaddr == lastaddr &&
1.101 skrll 229: segs[nseg - 1].ds_len + sgsize <= map->dm_maxsegsz &&
230: ((segs[nseg - 1]._ds_flags ^ _ds_flags) & _BUS_DMAMAP_COHERENT) == 0 &&
1.41 thorpej 231: (map->_dm_boundary == 0 ||
1.101 skrll 232: (segs[nseg - 1].ds_addr & bmask) == (curaddr & bmask))) {
1.41 thorpej 233: /* coalesce */
1.101 skrll 234: segs[nseg - 1].ds_len += sgsize;
1.41 thorpej 235: } else if (nseg >= map->_dm_segcnt) {
1.100 skrll 236: return EFBIG;
1.41 thorpej 237: } else {
238: /* new segment */
239: segs[nseg].ds_addr = curaddr;
240: segs[nseg].ds_len = sgsize;
1.61 matt 241: segs[nseg]._ds_flags = _ds_flags;
1.41 thorpej 242: nseg++;
243: }
244:
245: lastaddr = curaddr + sgsize;
246:
247: paddr += sgsize;
248: size -= sgsize;
249: if (size > 0)
250: goto again;
1.61 matt 251:
252: map->_dm_flags &= (_ds_flags & _BUS_DMAMAP_COHERENT);
1.41 thorpej 253: map->dm_nsegs = nseg;
1.100 skrll 254: return 0;
1.41 thorpej 255: }
256:
1.58 matt 257: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
258: static int _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
259: bus_size_t size, int flags);
260: static void _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map);
261: static int _bus_dma_uiomove(void *buf, struct uio *uio, size_t n,
262: int direction);
263:
264: static int
265: _bus_dma_load_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
266: size_t buflen, int buftype, int flags)
267: {
268: struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
269: struct vmspace * const vm = vmspace_kernel();
270: int error;
271:
272: KASSERT(cookie != NULL);
273: KASSERT(cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE);
274:
275: /*
276: * Allocate bounce pages, if necessary.
277: */
278: if ((cookie->id_flags & _BUS_DMA_HAS_BOUNCE) == 0) {
279: error = _bus_dma_alloc_bouncebuf(t, map, buflen, flags);
280: if (error)
1.100 skrll 281: return error;
1.58 matt 282: }
283:
284: /*
285: * Cache a pointer to the caller's buffer and load the DMA map
286: * with the bounce buffer.
287: */
288: cookie->id_origbuf = buf;
289: cookie->id_origbuflen = buflen;
290: error = _bus_dmamap_load_buffer(t, map, cookie->id_bouncebuf,
291: buflen, vm, flags);
292: if (error)
1.100 skrll 293: return error;
1.58 matt 294:
295: STAT_INCR(bounced_loads);
296: map->dm_mapsize = buflen;
297: map->_dm_vmspace = vm;
298: map->_dm_buftype = buftype;
299:
300: /* ...so _bus_dmamap_sync() knows we're bouncing */
1.63 matt 301: map->_dm_flags |= _BUS_DMAMAP_IS_BOUNCING;
1.58 matt 302: cookie->id_flags |= _BUS_DMA_IS_BOUNCING;
303: return 0;
304: }
305: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
306:
1.41 thorpej 307: /*
1.1 chris 308: * Common function for DMA map creation. May be called by bus-specific
309: * DMA map creation functions.
310: */
311: int
1.7 thorpej 312: _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
313: bus_size_t maxsegsz, bus_size_t boundary, int flags, bus_dmamap_t *dmamp)
1.1 chris 314: {
315: struct arm32_bus_dmamap *map;
316: void *mapstore;
317:
318: #ifdef DEBUG_DMA
1.103 skrll 319: printf("dmamap_create: t=%p size=%lx nseg=%x msegsz=%lx boundary=%lx"
320: " flags=%x\n", t, size, nsegments, maxsegsz, boundary, flags);
1.1 chris 321: #endif /* DEBUG_DMA */
322:
323: /*
324: * Allocate and initialize the DMA map. The end of the map
325: * is a variable-sized array of segments, so we allocate enough
326: * room for them in one shot.
327: *
328: * Note we don't preserve the WAITOK or NOWAIT flags. Preservation
329: * of ALLOCNOW notifies others that we've reserved these resources,
330: * and they are not to be freed.
331: *
332: * The bus_dmamap_t includes one bus_dma_segment_t, hence
333: * the (nsegments - 1).
334: */
1.81 matt 335: const size_t mapsize = sizeof(struct arm32_bus_dmamap) +
1.1 chris 336: (sizeof(bus_dma_segment_t) * (nsegments - 1));
1.81 matt 337: const int zallocflags = (flags & BUS_DMA_NOWAIT) ? KM_NOSLEEP : KM_SLEEP;
338: if ((mapstore = kmem_intr_zalloc(mapsize, zallocflags)) == NULL)
1.100 skrll 339: return ENOMEM;
1.1 chris 340:
341: map = (struct arm32_bus_dmamap *)mapstore;
342: map->_dm_size = size;
343: map->_dm_segcnt = nsegments;
1.43 matt 344: map->_dm_maxmaxsegsz = maxsegsz;
1.1 chris 345: map->_dm_boundary = boundary;
346: map->_dm_flags = flags & ~(BUS_DMA_WAITOK|BUS_DMA_NOWAIT);
1.14 thorpej 347: map->_dm_origbuf = NULL;
1.58 matt 348: map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
1.48 yamt 349: map->_dm_vmspace = vmspace_kernel();
1.58 matt 350: map->_dm_cookie = NULL;
1.43 matt 351: map->dm_maxsegsz = maxsegsz;
1.1 chris 352: map->dm_mapsize = 0; /* no valid mappings */
353: map->dm_nsegs = 0;
354:
1.58 matt 355: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
356: struct arm32_bus_dma_cookie *cookie;
357: int cookieflags;
358: void *cookiestore;
359: int error;
360:
361: cookieflags = 0;
362:
363: if (t->_may_bounce != NULL) {
364: error = (*t->_may_bounce)(t, map, flags, &cookieflags);
365: if (error != 0)
366: goto out;
367: }
368:
369: if (t->_ranges != NULL)
370: cookieflags |= _BUS_DMA_MIGHT_NEED_BOUNCE;
371:
372: if ((cookieflags & _BUS_DMA_MIGHT_NEED_BOUNCE) == 0) {
373: STAT_INCR(creates);
1.98 msaitoh 374: *dmamp = map;
1.58 matt 375: return 0;
376: }
377:
1.81 matt 378: const size_t cookiesize = sizeof(struct arm32_bus_dma_cookie) +
1.58 matt 379: (sizeof(bus_dma_segment_t) * map->_dm_segcnt);
380:
381: /*
382: * Allocate our cookie.
383: */
1.81 matt 384: if ((cookiestore = kmem_intr_zalloc(cookiesize, zallocflags)) == NULL) {
1.58 matt 385: error = ENOMEM;
386: goto out;
387: }
388: cookie = (struct arm32_bus_dma_cookie *)cookiestore;
389: cookie->id_flags = cookieflags;
390: map->_dm_cookie = cookie;
391: STAT_INCR(bounced_creates);
392:
393: error = _bus_dma_alloc_bouncebuf(t, map, size, flags);
394: out:
395: if (error)
396: _bus_dmamap_destroy(t, map);
1.98 msaitoh 397: else
398: *dmamp = map;
1.58 matt 399: #else
1.98 msaitoh 400: *dmamp = map;
1.58 matt 401: STAT_INCR(creates);
402: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1.1 chris 403: #ifdef DEBUG_DMA
404: printf("dmamap_create:map=%p\n", map);
405: #endif /* DEBUG_DMA */
1.100 skrll 406: return 0;
1.1 chris 407: }
408:
409: /*
410: * Common function for DMA map destruction. May be called by bus-specific
411: * DMA map destruction functions.
412: */
413: void
1.7 thorpej 414: _bus_dmamap_destroy(bus_dma_tag_t t, bus_dmamap_t map)
1.1 chris 415: {
416:
417: #ifdef DEBUG_DMA
418: printf("dmamap_destroy: t=%p map=%p\n", t, map);
419: #endif /* DEBUG_DMA */
1.58 matt 420: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
421: struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
1.13 briggs 422:
423: /*
1.58 matt 424: * Free any bounce pages this map might hold.
1.13 briggs 425: */
1.58 matt 426: if (cookie != NULL) {
1.81 matt 427: const size_t cookiesize = sizeof(struct arm32_bus_dma_cookie) +
428: (sizeof(bus_dma_segment_t) * map->_dm_segcnt);
429:
1.58 matt 430: if (cookie->id_flags & _BUS_DMA_IS_BOUNCING)
431: STAT_INCR(bounced_unloads);
432: map->dm_nsegs = 0;
433: if (cookie->id_flags & _BUS_DMA_HAS_BOUNCE)
434: _bus_dma_free_bouncebuf(t, map);
435: STAT_INCR(bounced_destroys);
1.81 matt 436: kmem_intr_free(cookie, cookiesize);
1.58 matt 437: } else
438: #endif
439: STAT_INCR(destroys);
440:
441: if (map->dm_nsegs > 0)
442: STAT_INCR(unloads);
1.13 briggs 443:
1.81 matt 444: const size_t mapsize = sizeof(struct arm32_bus_dmamap) +
445: (sizeof(bus_dma_segment_t) * (map->_dm_segcnt - 1));
446: kmem_intr_free(map, mapsize);
1.1 chris 447: }
448:
449: /*
450: * Common function for loading a DMA map with a linear buffer. May
451: * be called by bus-specific DMA map load functions.
452: */
453: int
1.7 thorpej 454: _bus_dmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
455: bus_size_t buflen, struct proc *p, int flags)
1.1 chris 456: {
1.58 matt 457: struct vmspace *vm;
1.41 thorpej 458: int error;
1.1 chris 459:
460: #ifdef DEBUG_DMA
461: printf("dmamap_load: t=%p map=%p buf=%p len=%lx p=%p f=%d\n",
462: t, map, buf, buflen, p, flags);
463: #endif /* DEBUG_DMA */
464:
1.58 matt 465: if (map->dm_nsegs > 0) {
466: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
467: struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
468: if (cookie != NULL) {
469: if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
470: STAT_INCR(bounced_unloads);
471: cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
1.63 matt 472: map->_dm_flags &= ~_BUS_DMAMAP_IS_BOUNCING;
1.58 matt 473: }
474: } else
475: #endif
476: STAT_INCR(unloads);
477: }
478:
1.1 chris 479: /*
480: * Make sure that on error condition we return "no valid mappings".
481: */
482: map->dm_mapsize = 0;
483: map->dm_nsegs = 0;
1.58 matt 484: map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
1.74 matt 485: KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
486: "dm_maxsegsz %lu _dm_maxmaxsegsz %lu",
487: map->dm_maxsegsz, map->_dm_maxmaxsegsz);
1.1 chris 488:
489: if (buflen > map->_dm_size)
1.100 skrll 490: return EINVAL;
1.1 chris 491:
1.48 yamt 492: if (p != NULL) {
493: vm = p->p_vmspace;
494: } else {
495: vm = vmspace_kernel();
496: }
497:
1.17 thorpej 498: /* _bus_dmamap_load_buffer() clears this if we're not... */
1.58 matt 499: map->_dm_flags |= _BUS_DMAMAP_COHERENT;
1.17 thorpej 500:
1.48 yamt 501: error = _bus_dmamap_load_buffer(t, map, buf, buflen, vm, flags);
1.1 chris 502: if (error == 0) {
503: map->dm_mapsize = buflen;
1.58 matt 504: map->_dm_vmspace = vm;
1.14 thorpej 505: map->_dm_origbuf = buf;
1.58 matt 506: map->_dm_buftype = _BUS_DMA_BUFTYPE_LINEAR;
1.81 matt 507: if (map->_dm_flags & _BUS_DMAMAP_COHERENT) {
508: STAT_INCR(coherent_loads);
509: } else {
510: STAT_INCR(loads);
511: }
1.58 matt 512: return 0;
1.1 chris 513: }
1.58 matt 514: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
515: struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
516: if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
517: error = _bus_dma_load_bouncebuf(t, map, buf, buflen,
518: _BUS_DMA_BUFTYPE_LINEAR, flags);
1.95 skrll 519: }
520: #endif
1.100 skrll 521: return error;
1.1 chris 522: }
523:
524: /*
525: * Like _bus_dmamap_load(), but for mbufs.
526: */
527: int
1.7 thorpej 528: _bus_dmamap_load_mbuf(bus_dma_tag_t t, bus_dmamap_t map, struct mbuf *m0,
529: int flags)
1.1 chris 530: {
1.105 skrll 531: struct mbuf *m;
1.41 thorpej 532: int error;
1.1 chris 533:
534: #ifdef DEBUG_DMA
535: printf("dmamap_load_mbuf: t=%p map=%p m0=%p f=%d\n",
536: t, map, m0, flags);
537: #endif /* DEBUG_DMA */
538:
1.58 matt 539: if (map->dm_nsegs > 0) {
540: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
541: struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
542: if (cookie != NULL) {
543: if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
544: STAT_INCR(bounced_unloads);
545: cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
1.63 matt 546: map->_dm_flags &= ~_BUS_DMAMAP_IS_BOUNCING;
1.58 matt 547: }
548: } else
549: #endif
550: STAT_INCR(unloads);
551: }
552:
1.1 chris 553: /*
554: * Make sure that on error condition we return "no valid mappings."
555: */
556: map->dm_mapsize = 0;
557: map->dm_nsegs = 0;
1.58 matt 558: map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
1.74 matt 559: KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
560: "dm_maxsegsz %lu _dm_maxmaxsegsz %lu",
561: map->dm_maxsegsz, map->_dm_maxmaxsegsz);
1.1 chris 562:
1.79 matt 563: KASSERT(m0->m_flags & M_PKTHDR);
1.1 chris 564:
565: if (m0->m_pkthdr.len > map->_dm_size)
1.100 skrll 566: return EINVAL;
1.1 chris 567:
1.61 matt 568: /* _bus_dmamap_load_paddr() clears this if we're not... */
569: map->_dm_flags |= _BUS_DMAMAP_COHERENT;
1.17 thorpej 570:
1.1 chris 571: error = 0;
572: for (m = m0; m != NULL && error == 0; m = m->m_next) {
1.41 thorpej 573: int offset;
574: int remainbytes;
575: const struct vm_page * const *pgs;
576: paddr_t paddr;
577: int size;
578:
1.28 thorpej 579: if (m->m_len == 0)
580: continue;
1.57 matt 581: /*
582: * Don't allow reads in read-only mbufs.
583: */
584: if (M_ROMAP(m) && (flags & BUS_DMA_READ)) {
585: error = EFAULT;
586: break;
587: }
1.108 maxv 588: switch (m->m_flags & (M_EXT|M_EXT_CLUSTER|M_EXT_PAGES)) {
589: case M_EXT|M_EXT_CLUSTER:
1.28 thorpej 590: /* XXX KDASSERT */
591: KASSERT(m->m_ext.ext_paddr != M_PADDR_INVALID);
1.41 thorpej 592: paddr = m->m_ext.ext_paddr +
1.28 thorpej 593: (m->m_data - m->m_ext.ext_buf);
1.41 thorpej 594: size = m->m_len;
1.61 matt 595: error = _bus_dmamap_load_paddr(t, map, paddr, size,
596: false);
1.41 thorpej 597: break;
1.95 skrll 598:
1.41 thorpej 599: case M_EXT|M_EXT_PAGES:
600: KASSERT(m->m_ext.ext_buf <= m->m_data);
601: KASSERT(m->m_data <=
602: m->m_ext.ext_buf + m->m_ext.ext_size);
1.95 skrll 603:
1.41 thorpej 604: offset = (vaddr_t)m->m_data -
605: trunc_page((vaddr_t)m->m_ext.ext_buf);
606: remainbytes = m->m_len;
607:
608: /* skip uninteresting pages */
609: pgs = (const struct vm_page * const *)
610: m->m_ext.ext_pgs + (offset >> PAGE_SHIFT);
1.95 skrll 611:
1.41 thorpej 612: offset &= PAGE_MASK; /* offset in the first page */
613:
614: /* load each page */
615: while (remainbytes > 0) {
616: const struct vm_page *pg;
617:
618: size = MIN(remainbytes, PAGE_SIZE - offset);
619:
620: pg = *pgs++;
621: KASSERT(pg);
622: paddr = VM_PAGE_TO_PHYS(pg) + offset;
623:
624: error = _bus_dmamap_load_paddr(t, map,
1.61 matt 625: paddr, size, false);
1.41 thorpej 626: if (error)
1.28 thorpej 627: break;
1.41 thorpej 628: offset = 0;
629: remainbytes -= size;
1.28 thorpej 630: }
631: break;
632:
633: case 0:
1.41 thorpej 634: paddr = m->m_paddr + M_BUFOFFSET(m) +
1.28 thorpej 635: (m->m_data - M_BUFADDR(m));
1.41 thorpej 636: size = m->m_len;
1.61 matt 637: error = _bus_dmamap_load_paddr(t, map, paddr, size,
638: false);
1.41 thorpej 639: break;
1.28 thorpej 640:
641: default:
642: error = _bus_dmamap_load_buffer(t, map, m->m_data,
1.48 yamt 643: m->m_len, vmspace_kernel(), flags);
1.28 thorpej 644: }
1.1 chris 645: }
646: if (error == 0) {
647: map->dm_mapsize = m0->m_pkthdr.len;
1.14 thorpej 648: map->_dm_origbuf = m0;
1.58 matt 649: map->_dm_buftype = _BUS_DMA_BUFTYPE_MBUF;
1.48 yamt 650: map->_dm_vmspace = vmspace_kernel(); /* always kernel */
1.81 matt 651: if (map->_dm_flags & _BUS_DMAMAP_COHERENT) {
652: STAT_INCR(coherent_loads);
653: } else {
654: STAT_INCR(loads);
655: }
1.58 matt 656: return 0;
1.1 chris 657: }
1.58 matt 658: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
659: struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
660: if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
661: error = _bus_dma_load_bouncebuf(t, map, m0, m0->m_pkthdr.len,
662: _BUS_DMA_BUFTYPE_MBUF, flags);
1.95 skrll 663: }
664: #endif
1.100 skrll 665: return error;
1.1 chris 666: }
667:
668: /*
669: * Like _bus_dmamap_load(), but for uios.
670: */
671: int
1.7 thorpej 672: _bus_dmamap_load_uio(bus_dma_tag_t t, bus_dmamap_t map, struct uio *uio,
673: int flags)
1.1 chris 674: {
675: bus_size_t minlen, resid;
676: struct iovec *iov;
1.50 christos 677: void *addr;
1.105 skrll 678: int i, error;
1.1 chris 679:
680: /*
681: * Make sure that on error condition we return "no valid mappings."
682: */
683: map->dm_mapsize = 0;
684: map->dm_nsegs = 0;
1.74 matt 685: KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
686: "dm_maxsegsz %lu _dm_maxmaxsegsz %lu",
687: map->dm_maxsegsz, map->_dm_maxmaxsegsz);
1.1 chris 688:
689: resid = uio->uio_resid;
690: iov = uio->uio_iov;
691:
1.17 thorpej 692: /* _bus_dmamap_load_buffer() clears this if we're not... */
1.58 matt 693: map->_dm_flags |= _BUS_DMAMAP_COHERENT;
1.17 thorpej 694:
1.1 chris 695: error = 0;
696: for (i = 0; i < uio->uio_iovcnt && resid != 0 && error == 0; i++) {
697: /*
698: * Now at the first iovec to load. Load each iovec
699: * until we have exhausted the residual count.
700: */
701: minlen = resid < iov[i].iov_len ? resid : iov[i].iov_len;
1.50 christos 702: addr = (void *)iov[i].iov_base;
1.1 chris 703:
704: error = _bus_dmamap_load_buffer(t, map, addr, minlen,
1.48 yamt 705: uio->uio_vmspace, flags);
1.1 chris 706:
707: resid -= minlen;
708: }
709: if (error == 0) {
710: map->dm_mapsize = uio->uio_resid;
1.14 thorpej 711: map->_dm_origbuf = uio;
1.58 matt 712: map->_dm_buftype = _BUS_DMA_BUFTYPE_UIO;
1.48 yamt 713: map->_dm_vmspace = uio->uio_vmspace;
1.81 matt 714: if (map->_dm_flags & _BUS_DMAMAP_COHERENT) {
715: STAT_INCR(coherent_loads);
716: } else {
717: STAT_INCR(loads);
718: }
1.1 chris 719: }
1.100 skrll 720: return error;
1.1 chris 721: }
722:
723: /*
724: * Like _bus_dmamap_load(), but for raw memory allocated with
725: * bus_dmamem_alloc().
726: */
727: int
1.7 thorpej 728: _bus_dmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
1.94 jmcneill 729: bus_dma_segment_t *segs, int nsegs, bus_size_t size0, int flags)
1.1 chris 730: {
731:
1.94 jmcneill 732: bus_size_t size;
733: int i, error = 0;
734:
735: /*
736: * Make sure that on error conditions we return "no valid mappings."
737: */
738: map->dm_mapsize = 0;
739: map->dm_nsegs = 0;
740: KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
741:
742: if (size0 > map->_dm_size)
743: return EINVAL;
744:
745: for (i = 0, size = size0; i < nsegs && size > 0; i++) {
746: bus_dma_segment_t *ds = &segs[i];
747: bus_size_t sgsize;
748:
749: sgsize = MIN(ds->ds_len, size);
750: if (sgsize == 0)
751: continue;
752: error = _bus_dmamap_load_paddr(t, map, ds->ds_addr,
753: sgsize, false);
754: if (error != 0)
755: break;
756: size -= sgsize;
757: }
758:
759: if (error != 0) {
760: map->dm_mapsize = 0;
761: map->dm_nsegs = 0;
762: return error;
763: }
764:
765: /* XXX TBD bounce */
766:
767: map->dm_mapsize = size0;
768: return 0;
1.1 chris 769: }
770:
771: /*
772: * Common function for unloading a DMA map. May be called by
773: * bus-specific DMA map unload functions.
774: */
775: void
1.7 thorpej 776: _bus_dmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
1.1 chris 777: {
778:
779: #ifdef DEBUG_DMA
780: printf("dmamap_unload: t=%p map=%p\n", t, map);
781: #endif /* DEBUG_DMA */
782:
783: /*
784: * No resources to free; just mark the mappings as
785: * invalid.
786: */
787: map->dm_mapsize = 0;
788: map->dm_nsegs = 0;
1.14 thorpej 789: map->_dm_origbuf = NULL;
1.58 matt 790: map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
1.48 yamt 791: map->_dm_vmspace = NULL;
1.1 chris 792: }
793:
1.57 matt 794: static void
1.103 skrll 795: _bus_dmamap_sync_segment(vaddr_t va, paddr_t pa, vsize_t len, int ops,
796: bool readonly_p)
1.14 thorpej 797: {
1.106 skrll 798:
1.113 ! skrll 799: #if defined(ARM_MMU_EXTENDED) || defined(CPU_CORTEX)
1.106 skrll 800: /*
801: * No optimisations are available for readonly mbufs on armv6+, so
802: * assume it's not readonly from here on.
803: *
804: * See the comment in _bus_dmamap_sync_mbuf
805: */
806: readonly_p = false;
807: #endif
808:
1.86 matt 809: KASSERTMSG((va & PAGE_MASK) == (pa & PAGE_MASK),
810: "va %#lx pa %#lx", va, pa);
1.62 matt 811: #if 0
812: printf("sync_segment: va=%#lx pa=%#lx len=%#lx ops=%#x ro=%d\n",
813: va, pa, len, ops, readonly_p);
814: #endif
1.14 thorpej 815:
816: switch (ops) {
817: case BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE:
1.57 matt 818: if (!readonly_p) {
1.76 matt 819: STAT_INCR(sync_prereadwrite);
1.57 matt 820: cpu_dcache_wbinv_range(va, len);
821: cpu_sdcache_wbinv_range(va, pa, len);
822: break;
823: }
824: /* FALLTHROUGH */
1.14 thorpej 825:
1.57 matt 826: case BUS_DMASYNC_PREREAD: {
1.59 matt 827: const size_t line_size = arm_dcache_align;
828: const size_t line_mask = arm_dcache_align_mask;
829: vsize_t misalignment = va & line_mask;
1.57 matt 830: if (misalignment) {
1.59 matt 831: va -= misalignment;
832: pa -= misalignment;
833: len += misalignment;
1.77 matt 834: STAT_INCR(sync_preread_begin);
1.59 matt 835: cpu_dcache_wbinv_range(va, line_size);
836: cpu_sdcache_wbinv_range(va, pa, line_size);
837: if (len <= line_size)
1.57 matt 838: break;
1.59 matt 839: va += line_size;
840: pa += line_size;
841: len -= line_size;
1.57 matt 842: }
1.59 matt 843: misalignment = len & line_mask;
1.57 matt 844: len -= misalignment;
1.65 matt 845: if (len > 0) {
1.77 matt 846: STAT_INCR(sync_preread);
1.65 matt 847: cpu_dcache_inv_range(va, len);
848: cpu_sdcache_inv_range(va, pa, len);
849: }
1.57 matt 850: if (misalignment) {
851: va += len;
852: pa += len;
1.77 matt 853: STAT_INCR(sync_preread_tail);
1.59 matt 854: cpu_dcache_wbinv_range(va, line_size);
855: cpu_sdcache_wbinv_range(va, pa, line_size);
1.57 matt 856: }
1.14 thorpej 857: break;
1.57 matt 858: }
1.14 thorpej 859:
860: case BUS_DMASYNC_PREWRITE:
1.76 matt 861: STAT_INCR(sync_prewrite);
1.57 matt 862: cpu_dcache_wb_range(va, len);
863: cpu_sdcache_wb_range(va, pa, len);
1.14 thorpej 864: break;
1.67 matt 865:
866: #ifdef CPU_CORTEX
867: /*
868: * Cortex CPUs can do speculative loads so we need to clean the cache
869: * after a DMA read to deal with any speculatively loaded cache lines.
870: * Since these can't be dirty, we can just invalidate them and don't
871: * have to worry about having to write back their contents.
872: */
873: case BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE:
1.76 matt 874: STAT_INCR(sync_postreadwrite);
875: cpu_dcache_inv_range(va, len);
876: cpu_sdcache_inv_range(va, pa, len);
877: break;
1.67 matt 878: case BUS_DMASYNC_POSTREAD:
1.76 matt 879: STAT_INCR(sync_postread);
1.67 matt 880: cpu_dcache_inv_range(va, len);
881: cpu_sdcache_inv_range(va, pa, len);
882: break;
883: #endif
1.14 thorpej 884: }
885: }
886:
1.47 perry 887: static inline void
1.57 matt 888: _bus_dmamap_sync_linear(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1.14 thorpej 889: bus_size_t len, int ops)
890: {
1.57 matt 891: bus_dma_segment_t *ds = map->dm_segs;
892: vaddr_t va = (vaddr_t) map->_dm_origbuf;
1.58 matt 893: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1.63 matt 894: if (map->_dm_flags & _BUS_DMAMAP_IS_BOUNCING) {
895: struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
1.58 matt 896: va = (vaddr_t) cookie->id_bouncebuf;
897: }
898: #endif
1.57 matt 899:
900: while (len > 0) {
901: while (offset >= ds->ds_len) {
902: offset -= ds->ds_len;
903: va += ds->ds_len;
904: ds++;
905: }
906:
1.59 matt 907: paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + offset);
1.112 riastrad 908: size_t seglen = uimin(len, ds->ds_len - offset);
1.57 matt 909:
1.61 matt 910: if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
911: _bus_dmamap_sync_segment(va + offset, pa, seglen, ops,
1.67 matt 912: false);
1.57 matt 913:
914: offset += seglen;
915: len -= seglen;
916: }
917: }
918:
919: static inline void
920: _bus_dmamap_sync_mbuf(bus_dma_tag_t t, bus_dmamap_t map, bus_size_t offset,
921: bus_size_t len, int ops)
922: {
923: bus_dma_segment_t *ds = map->dm_segs;
924: struct mbuf *m = map->_dm_origbuf;
925: bus_size_t voff = offset;
926: bus_size_t ds_off = offset;
927:
928: while (len > 0) {
929: /* Find the current dma segment */
930: while (ds_off >= ds->ds_len) {
931: ds_off -= ds->ds_len;
932: ds++;
933: }
934: /* Find the current mbuf. */
935: while (voff >= m->m_len) {
936: voff -= m->m_len;
937: m = m->m_next;
1.14 thorpej 938: }
939:
940: /*
941: * Now at the first mbuf to sync; nail each one until
942: * we have exhausted the length.
943: */
1.112 riastrad 944: vsize_t seglen = uimin(len, uimin(m->m_len - voff, ds->ds_len - ds_off));
1.57 matt 945: vaddr_t va = mtod(m, vaddr_t) + voff;
1.59 matt 946: paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
1.14 thorpej 947:
1.28 thorpej 948: /*
949: * We can save a lot of work here if we know the mapping
1.93 matt 950: * is read-only at the MMU and we aren't using the armv6+
951: * MMU:
1.28 thorpej 952: *
953: * If a mapping is read-only, no dirty cache blocks will
954: * exist for it. If a writable mapping was made read-only,
955: * we know any dirty cache lines for the range will have
956: * been cleaned for us already. Therefore, if the upper
957: * layer can tell us we have a read-only mapping, we can
958: * skip all cache cleaning.
959: *
960: * NOTE: This only works if we know the pmap cleans pages
961: * before making a read-write -> read-only transition. If
962: * this ever becomes non-true (e.g. Physically Indexed
963: * cache), this will have to be revisited.
964: */
1.14 thorpej 965:
1.92 matt 966: if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0) {
967: /*
968: * If we are doing preread (DMAing into the mbuf),
1.95 skrll 969: * this mbuf better not be readonly,
1.92 matt 970: */
971: KASSERT(!(ops & BUS_DMASYNC_PREREAD) || !M_ROMAP(m));
1.61 matt 972: _bus_dmamap_sync_segment(va, pa, seglen, ops,
973: M_ROMAP(m));
1.92 matt 974: }
1.57 matt 975: voff += seglen;
976: ds_off += seglen;
977: len -= seglen;
1.14 thorpej 978: }
979: }
980:
1.47 perry 981: static inline void
1.14 thorpej 982: _bus_dmamap_sync_uio(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
983: bus_size_t len, int ops)
984: {
1.57 matt 985: bus_dma_segment_t *ds = map->dm_segs;
1.14 thorpej 986: struct uio *uio = map->_dm_origbuf;
1.57 matt 987: struct iovec *iov = uio->uio_iov;
988: bus_size_t voff = offset;
989: bus_size_t ds_off = offset;
990:
991: while (len > 0) {
992: /* Find the current dma segment */
993: while (ds_off >= ds->ds_len) {
994: ds_off -= ds->ds_len;
995: ds++;
996: }
1.14 thorpej 997:
1.57 matt 998: /* Find the current iovec. */
999: while (voff >= iov->iov_len) {
1000: voff -= iov->iov_len;
1001: iov++;
1.14 thorpej 1002: }
1003:
1004: /*
1005: * Now at the first iovec to sync; nail each one until
1006: * we have exhausted the length.
1007: */
1.112 riastrad 1008: vsize_t seglen = uimin(len, uimin(iov->iov_len - voff, ds->ds_len - ds_off));
1.57 matt 1009: vaddr_t va = (vaddr_t) iov->iov_base + voff;
1.59 matt 1010: paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
1.57 matt 1011:
1.61 matt 1012: if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
1013: _bus_dmamap_sync_segment(va, pa, seglen, ops, false);
1.57 matt 1014:
1015: voff += seglen;
1016: ds_off += seglen;
1017: len -= seglen;
1.14 thorpej 1018: }
1019: }
1020:
1.1 chris 1021: /*
1022: * Common function for DMA map synchronization. May be called
1023: * by bus-specific DMA map synchronization functions.
1.8 thorpej 1024: *
1025: * This version works for the Virtually Indexed Virtually Tagged
1026: * cache found on 32-bit ARM processors.
1027: *
1028: * XXX Should have separate versions for write-through vs.
1029: * XXX write-back caches. We currently assume write-back
1030: * XXX here, which is not as efficient as it could be for
1031: * XXX the write-through case.
1.1 chris 1032: */
1033: void
1.7 thorpej 1034: _bus_dmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1035: bus_size_t len, int ops)
1.1 chris 1036: {
1037: #ifdef DEBUG_DMA
1038: printf("dmamap_sync: t=%p map=%p offset=%lx len=%lx ops=%x\n",
1039: t, map, offset, len, ops);
1040: #endif /* DEBUG_DMA */
1041:
1.8 thorpej 1042: /*
1043: * Mixing of PRE and POST operations is not allowed.
1044: */
1045: if ((ops & (BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE)) != 0 &&
1046: (ops & (BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE)) != 0)
1047: panic("_bus_dmamap_sync: mix PRE and POST");
1048:
1.79 matt 1049: KASSERTMSG(offset < map->dm_mapsize,
1050: "offset %lu mapsize %lu",
1051: offset, map->dm_mapsize);
1052: KASSERTMSG(len > 0 && offset + len <= map->dm_mapsize,
1053: "len %lu offset %lu mapsize %lu",
1054: len, offset, map->dm_mapsize);
1.8 thorpej 1055:
1056: /*
1057: * For a virtually-indexed write-back cache, we need
1058: * to do the following things:
1059: *
1060: * PREREAD -- Invalidate the D-cache. We do this
1061: * here in case a write-back is required by the back-end.
1062: *
1063: * PREWRITE -- Write-back the D-cache. Note that if
1064: * we are doing a PREREAD|PREWRITE, we can collapse
1065: * the whole thing into a single Wb-Inv.
1066: *
1.67 matt 1067: * POSTREAD -- Re-invalidate the D-cache in case speculative
1068: * memory accesses caused cachelines to become valid with now
1069: * invalid data.
1.8 thorpej 1070: *
1071: * POSTWRITE -- Nothing.
1072: */
1.58 matt 1073: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1.74 matt 1074: const bool bouncing = (map->_dm_flags & _BUS_DMAMAP_IS_BOUNCING);
1.63 matt 1075: #else
1076: const bool bouncing = false;
1.58 matt 1077: #endif
1.8 thorpej 1078:
1.58 matt 1079: const int pre_ops = ops & (BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1.67 matt 1080: #ifdef CPU_CORTEX
1081: const int post_ops = ops & (BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1082: #else
1083: const int post_ops = 0;
1084: #endif
1.96 skrll 1085: if (!bouncing) {
1086: if (pre_ops == 0 && post_ops == BUS_DMASYNC_POSTWRITE) {
1087: STAT_INCR(sync_postwrite);
1088: return;
1089: } else if (pre_ops == 0 && post_ops == 0) {
1090: return;
1091: }
1.61 matt 1092: }
1.74 matt 1093: KASSERTMSG(bouncing || pre_ops != 0 || (post_ops & BUS_DMASYNC_POSTREAD),
1094: "pre_ops %#x post_ops %#x", pre_ops, post_ops);
1.58 matt 1095: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1096: if (bouncing && (ops & BUS_DMASYNC_PREWRITE)) {
1.63 matt 1097: struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
1.58 matt 1098: STAT_INCR(write_bounces);
1099: char * const dataptr = (char *)cookie->id_bouncebuf + offset;
1100: /*
1101: * Copy the caller's buffer to the bounce buffer.
1102: */
1103: switch (map->_dm_buftype) {
1104: case _BUS_DMA_BUFTYPE_LINEAR:
1105: memcpy(dataptr, cookie->id_origlinearbuf + offset, len);
1106: break;
1107: case _BUS_DMA_BUFTYPE_MBUF:
1108: m_copydata(cookie->id_origmbuf, offset, len, dataptr);
1109: break;
1110: case _BUS_DMA_BUFTYPE_UIO:
1111: _bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_WRITE);
1112: break;
1113: #ifdef DIAGNOSTIC
1114: case _BUS_DMA_BUFTYPE_RAW:
1115: panic("_bus_dmamap_sync(pre): _BUS_DMA_BUFTYPE_RAW");
1116: break;
1117:
1118: case _BUS_DMA_BUFTYPE_INVALID:
1119: panic("_bus_dmamap_sync(pre): _BUS_DMA_BUFTYPE_INVALID");
1120: break;
1121:
1122: default:
1123: panic("_bus_dmamap_sync(pre): map %p: unknown buffer type %d\n",
1124: map, map->_dm_buftype);
1125: break;
1126: #endif /* DIAGNOSTIC */
1127: }
1128: }
1129: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1130:
1.17 thorpej 1131: /* Skip cache frobbing if mapping was COHERENT. */
1.75 matt 1132: if (!bouncing && (map->_dm_flags & _BUS_DMAMAP_COHERENT)) {
1.17 thorpej 1133: /* Drain the write buffer. */
1.75 matt 1134: if (pre_ops & BUS_DMASYNC_PREWRITE)
1135: cpu_drain_writebuf();
1.17 thorpej 1136: return;
1137: }
1.8 thorpej 1138:
1.58 matt 1139: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1140: if (bouncing && ((map->_dm_flags & _BUS_DMAMAP_COHERENT) || pre_ops == 0)) {
1141: goto bounce_it;
1142: }
1143: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1144:
1.80 matt 1145: #ifndef ARM_MMU_EXTENDED
1.8 thorpej 1146: /*
1.38 scw 1147: * If the mapping belongs to a non-kernel vmspace, and the
1148: * vmspace has not been active since the last time a full
1149: * cache flush was performed, we don't need to do anything.
1.8 thorpej 1150: */
1.48 yamt 1151: if (__predict_false(!VMSPACE_IS_KERNEL_P(map->_dm_vmspace) &&
1152: vm_map_pmap(&map->_dm_vmspace->vm_map)->pm_cstate.cs_cache_d == 0))
1.8 thorpej 1153: return;
1.80 matt 1154: #endif
1.8 thorpej 1155:
1.58 matt 1156: int buftype = map->_dm_buftype;
1157: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1158: if (bouncing) {
1159: buftype = _BUS_DMA_BUFTYPE_LINEAR;
1160: }
1161: #endif
1162:
1163: switch (buftype) {
1164: case _BUS_DMA_BUFTYPE_LINEAR:
1.14 thorpej 1165: _bus_dmamap_sync_linear(t, map, offset, len, ops);
1166: break;
1167:
1.58 matt 1168: case _BUS_DMA_BUFTYPE_MBUF:
1.14 thorpej 1169: _bus_dmamap_sync_mbuf(t, map, offset, len, ops);
1170: break;
1171:
1.58 matt 1172: case _BUS_DMA_BUFTYPE_UIO:
1.14 thorpej 1173: _bus_dmamap_sync_uio(t, map, offset, len, ops);
1174: break;
1175:
1.58 matt 1176: case _BUS_DMA_BUFTYPE_RAW:
1177: panic("_bus_dmamap_sync: _BUS_DMA_BUFTYPE_RAW");
1.14 thorpej 1178: break;
1179:
1.58 matt 1180: case _BUS_DMA_BUFTYPE_INVALID:
1181: panic("_bus_dmamap_sync: _BUS_DMA_BUFTYPE_INVALID");
1.14 thorpej 1182: break;
1183:
1184: default:
1.58 matt 1185: panic("_bus_dmamap_sync: map %p: unknown buffer type %d\n",
1186: map, map->_dm_buftype);
1.8 thorpej 1187: }
1.1 chris 1188:
1.8 thorpej 1189: /* Drain the write buffer. */
1190: cpu_drain_writebuf();
1.58 matt 1191:
1192: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1193: bounce_it:
1.76 matt 1194: if (!bouncing || (ops & BUS_DMASYNC_POSTREAD) == 0)
1.58 matt 1195: return;
1196:
1.63 matt 1197: struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
1.58 matt 1198: char * const dataptr = (char *)cookie->id_bouncebuf + offset;
1199: STAT_INCR(read_bounces);
1200: /*
1201: * Copy the bounce buffer to the caller's buffer.
1202: */
1203: switch (map->_dm_buftype) {
1204: case _BUS_DMA_BUFTYPE_LINEAR:
1205: memcpy(cookie->id_origlinearbuf + offset, dataptr, len);
1206: break;
1207:
1208: case _BUS_DMA_BUFTYPE_MBUF:
1209: m_copyback(cookie->id_origmbuf, offset, len, dataptr);
1210: break;
1211:
1212: case _BUS_DMA_BUFTYPE_UIO:
1213: _bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_READ);
1214: break;
1215: #ifdef DIAGNOSTIC
1216: case _BUS_DMA_BUFTYPE_RAW:
1217: panic("_bus_dmamap_sync(post): _BUS_DMA_BUFTYPE_RAW");
1218: break;
1219:
1220: case _BUS_DMA_BUFTYPE_INVALID:
1221: panic("_bus_dmamap_sync(post): _BUS_DMA_BUFTYPE_INVALID");
1222: break;
1223:
1224: default:
1225: panic("_bus_dmamap_sync(post): map %p: unknown buffer type %d\n",
1226: map, map->_dm_buftype);
1227: break;
1228: #endif
1229: }
1230: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1.1 chris 1231: }
1232:
1233: /*
1234: * Common function for DMA-safe memory allocation. May be called
1235: * by bus-specific DMA memory allocation functions.
1236: */
1237:
1.11 thorpej 1238: extern paddr_t physical_start;
1239: extern paddr_t physical_end;
1.1 chris 1240:
1241: int
1.7 thorpej 1242: _bus_dmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1243: bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1244: int flags)
1.1 chris 1245: {
1.15 thorpej 1246: struct arm32_dma_range *dr;
1.37 mycroft 1247: int error, i;
1.15 thorpej 1248:
1.1 chris 1249: #ifdef DEBUG_DMA
1.15 thorpej 1250: printf("dmamem_alloc t=%p size=%lx align=%lx boundary=%lx "
1251: "segs=%p nsegs=%x rsegs=%p flags=%x\n", t, size, alignment,
1252: boundary, segs, nsegs, rsegs, flags);
1253: #endif
1254:
1255: if ((dr = t->_ranges) != NULL) {
1.37 mycroft 1256: error = ENOMEM;
1.15 thorpej 1257: for (i = 0; i < t->_nranges; i++, dr++) {
1.70 matt 1258: if (dr->dr_len == 0
1259: || (dr->dr_flags & _BUS_DMAMAP_NOALLOC))
1.15 thorpej 1260: continue;
1261: error = _bus_dmamem_alloc_range(t, size, alignment,
1262: boundary, segs, nsegs, rsegs, flags,
1263: trunc_page(dr->dr_sysbase),
1264: trunc_page(dr->dr_sysbase + dr->dr_len));
1265: if (error == 0)
1266: break;
1267: }
1268: } else {
1269: error = _bus_dmamem_alloc_range(t, size, alignment, boundary,
1270: segs, nsegs, rsegs, flags, trunc_page(physical_start),
1271: trunc_page(physical_end));
1272: }
1273:
1.1 chris 1274: #ifdef DEBUG_DMA
1275: printf("dmamem_alloc: =%d\n", error);
1.15 thorpej 1276: #endif
1277:
1.100 skrll 1278: return error;
1.1 chris 1279: }
1280:
1281: /*
1282: * Common function for freeing DMA-safe memory. May be called by
1283: * bus-specific DMA memory free functions.
1284: */
1285: void
1.7 thorpej 1286: _bus_dmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
1.1 chris 1287: {
1288: struct vm_page *m;
1289: bus_addr_t addr;
1290: struct pglist mlist;
1291: int curseg;
1292:
1293: #ifdef DEBUG_DMA
1294: printf("dmamem_free: t=%p segs=%p nsegs=%x\n", t, segs, nsegs);
1295: #endif /* DEBUG_DMA */
1296:
1297: /*
1298: * Build a list of pages to free back to the VM system.
1299: */
1300: TAILQ_INIT(&mlist);
1301: for (curseg = 0; curseg < nsegs; curseg++) {
1302: for (addr = segs[curseg].ds_addr;
1303: addr < (segs[curseg].ds_addr + segs[curseg].ds_len);
1304: addr += PAGE_SIZE) {
1305: m = PHYS_TO_VM_PAGE(addr);
1.52 ad 1306: TAILQ_INSERT_TAIL(&mlist, m, pageq.queue);
1.1 chris 1307: }
1308: }
1309: uvm_pglistfree(&mlist);
1310: }
1311:
1312: /*
1313: * Common function for mapping DMA-safe memory. May be called by
1314: * bus-specific DMA memory map functions.
1315: */
1316: int
1.7 thorpej 1317: _bus_dmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1.50 christos 1318: size_t size, void **kvap, int flags)
1.1 chris 1319: {
1.11 thorpej 1320: vaddr_t va;
1.57 matt 1321: paddr_t pa;
1.1 chris 1322: int curseg;
1.65 matt 1323: const uvm_flag_t kmflags = UVM_KMF_VAONLY
1324: | ((flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0);
1325: vsize_t align = 0;
1.1 chris 1326:
1327: #ifdef DEBUG_DMA
1.3 rearnsha 1328: printf("dmamem_map: t=%p segs=%p nsegs=%x size=%lx flags=%x\n", t,
1329: segs, nsegs, (unsigned long)size, flags);
1.1 chris 1330: #endif /* DEBUG_DMA */
1331:
1.62 matt 1332: #ifdef PMAP_MAP_POOLPAGE
1333: /*
1334: * If all of memory is mapped, and we are mapping a single physically
1335: * contiguous area then this area is already mapped. Let's see if we
1336: * avoid having a separate mapping for it.
1337: */
1338: if (nsegs == 1) {
1339: /*
1340: * If this is a non-COHERENT mapping, then the existing kernel
1341: * mapping is already compatible with it.
1342: */
1.68 matt 1343: bool direct_mapable = (flags & BUS_DMA_COHERENT) == 0;
1344: pa = segs[0].ds_addr;
1345:
1.62 matt 1346: /*
1.68 matt 1347: * This is a COHERENT mapping which, unless this address is in
1.62 matt 1348: * a COHERENT dma range, will not be compatible.
1349: */
1350: if (t->_ranges != NULL) {
1351: const struct arm32_dma_range * const dr =
1.68 matt 1352: _bus_dma_paddr_inrange(t->_ranges, t->_nranges, pa);
1.71 matt 1353: if (dr != NULL
1354: && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
1355: direct_mapable = true;
1.68 matt 1356: }
1357: }
1358:
1.87 matt 1359: #ifdef PMAP_NEED_ALLOC_POOLPAGE
1360: /*
1361: * The page can only be direct mapped if was allocated out
1.95 skrll 1362: * of the arm poolpage vm freelist.
1.87 matt 1363: */
1.97 cherry 1364: uvm_physseg_t upm = uvm_physseg_find(atop(pa), NULL);
1365: KASSERT(uvm_physseg_valid_p(upm));
1.87 matt 1366: if (direct_mapable) {
1367: direct_mapable =
1.97 cherry 1368: (arm_poolpage_vmfreelist == uvm_physseg_get_free_list(upm));
1.87 matt 1369: }
1370: #endif
1371:
1.68 matt 1372: if (direct_mapable) {
1373: *kvap = (void *)PMAP_MAP_POOLPAGE(pa);
1.64 matt 1374: #ifdef DEBUG_DMA
1.68 matt 1375: printf("dmamem_map: =%p\n", *kvap);
1.64 matt 1376: #endif /* DEBUG_DMA */
1.68 matt 1377: return 0;
1.62 matt 1378: }
1379: }
1380: #endif
1381:
1.1 chris 1382: size = round_page(size);
1.107 ryo 1383:
1384: #ifdef PMAP_MAPSIZE1
1385: if (size >= PMAP_MAPSIZE1)
1386: align = PMAP_MAPSIZE1;
1387:
1388: #ifdef PMAP_MAPSIZE2
1389:
1390: #if PMAP_MAPSIZE1 > PMAP_MAPSIZE2
1391: #error PMAP_MAPSIZE1 must be smaller than PMAP_MAPSIZE2
1392: #endif
1393:
1394: if (size >= PMAP_MAPSIZE2)
1395: align = PMAP_MAPSIZE2;
1396:
1397: #ifdef PMAP_MAPSIZE3
1398:
1399: #if PMAP_MAPSIZE2 > PMAP_MAPSIZE3
1400: #error PMAP_MAPSIZE2 must be smaller than PMAP_MAPSIZE3
1401: #endif
1402:
1403: if (size >= PMAP_MAPSIZE3)
1404: align = PMAP_MAPSIZE3;
1405: #endif
1406: #endif
1407: #endif
1.65 matt 1408:
1409: va = uvm_km_alloc(kernel_map, size, align, kmflags);
1410: if (__predict_false(va == 0 && align > 0)) {
1411: align = 0;
1412: va = uvm_km_alloc(kernel_map, size, 0, kmflags);
1413: }
1.1 chris 1414:
1415: if (va == 0)
1.100 skrll 1416: return ENOMEM;
1.1 chris 1417:
1.50 christos 1418: *kvap = (void *)va;
1.1 chris 1419:
1420: for (curseg = 0; curseg < nsegs; curseg++) {
1.57 matt 1421: for (pa = segs[curseg].ds_addr;
1422: pa < (segs[curseg].ds_addr + segs[curseg].ds_len);
1423: pa += PAGE_SIZE, va += PAGE_SIZE, size -= PAGE_SIZE) {
1.68 matt 1424: bool uncached = (flags & BUS_DMA_COHERENT);
1.1 chris 1425: #ifdef DEBUG_DMA
1.57 matt 1426: printf("wiring p%lx to v%lx", pa, va);
1.1 chris 1427: #endif /* DEBUG_DMA */
1428: if (size == 0)
1429: panic("_bus_dmamem_map: size botch");
1.68 matt 1430:
1431: const struct arm32_dma_range * const dr =
1432: _bus_dma_paddr_inrange(t->_ranges, t->_nranges, pa);
1433: /*
1434: * If this dma region is coherent then there is
1435: * no need for an uncached mapping.
1436: */
1.71 matt 1437: if (dr != NULL
1438: && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
1439: uncached = false;
1.68 matt 1440: }
1.71 matt 1441:
1.81 matt 1442: pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE,
1443: PMAP_WIRED | (uncached ? PMAP_NOCACHE : 0));
1.1 chris 1444: }
1445: }
1.2 chris 1446: pmap_update(pmap_kernel());
1.1 chris 1447: #ifdef DEBUG_DMA
1448: printf("dmamem_map: =%p\n", *kvap);
1449: #endif /* DEBUG_DMA */
1.100 skrll 1450: return 0;
1.1 chris 1451: }
1452:
1453: /*
1454: * Common function for unmapping DMA-safe memory. May be called by
1455: * bus-specific DMA memory unmapping functions.
1456: */
1457: void
1.50 christos 1458: _bus_dmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
1.1 chris 1459: {
1460:
1461: #ifdef DEBUG_DMA
1.65 matt 1462: printf("dmamem_unmap: t=%p kva=%p size=%zx\n", t, kva, size);
1.1 chris 1463: #endif /* DEBUG_DMA */
1.79 matt 1464: KASSERTMSG(((uintptr_t)kva & PAGE_MASK) == 0,
1.83 christos 1465: "kva %p (%#"PRIxPTR")", kva, ((uintptr_t)kva & PAGE_MASK));
1.1 chris 1466:
1.84 matt 1467: #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
1468: /*
1.88 snj 1469: * Check to see if this used direct mapped memory. Get its physical
1.84 matt 1470: * address and try to map it. If the resultant matches the kva, then
1.99 skrll 1471: * it was and so we can just return since we have nothing to free up.
1.84 matt 1472: */
1473: paddr_t pa;
1474: vaddr_t va;
1475: (void)pmap_extract(pmap_kernel(), (vaddr_t)kva, &pa);
1476: if (mm_md_direct_mapped_phys(pa, &va) && va == (vaddr_t)kva)
1477: return;
1478: #endif
1479:
1.1 chris 1480: size = round_page(size);
1.65 matt 1481: pmap_kremove((vaddr_t)kva, size);
1.44 yamt 1482: pmap_update(pmap_kernel());
1483: uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
1.1 chris 1484: }
1485:
1486: /*
1487: * Common functin for mmap(2)'ing DMA-safe memory. May be called by
1488: * bus-specific DMA mmap(2)'ing functions.
1489: */
1490: paddr_t
1.7 thorpej 1491: _bus_dmamem_mmap(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1492: off_t off, int prot, int flags)
1.1 chris 1493: {
1.73 macallan 1494: paddr_t map_flags;
1.1 chris 1495: int i;
1496:
1497: for (i = 0; i < nsegs; i++) {
1.79 matt 1498: KASSERTMSG((off & PAGE_MASK) == 0,
1.111 christos 1499: "off %#jx (%#x)", (uintmax_t)off, (int)off & PAGE_MASK);
1.79 matt 1500: KASSERTMSG((segs[i].ds_addr & PAGE_MASK) == 0,
1501: "ds_addr %#lx (%#x)", segs[i].ds_addr,
1502: (int)segs[i].ds_addr & PAGE_MASK);
1503: KASSERTMSG((segs[i].ds_len & PAGE_MASK) == 0,
1504: "ds_len %#lx (%#x)", segs[i].ds_addr,
1505: (int)segs[i].ds_addr & PAGE_MASK);
1.1 chris 1506: if (off >= segs[i].ds_len) {
1507: off -= segs[i].ds_len;
1508: continue;
1509: }
1510:
1.73 macallan 1511: map_flags = 0;
1512: if (flags & BUS_DMA_PREFETCHABLE)
1.107 ryo 1513: map_flags |= ARM_MMAP_WRITECOMBINE;
1.73 macallan 1514:
1.100 skrll 1515: return arm_btop((u_long)segs[i].ds_addr + off) | map_flags;
1.95 skrll 1516:
1.1 chris 1517: }
1518:
1519: /* Page not found. */
1.100 skrll 1520: return -1;
1.1 chris 1521: }
1522:
1523: /**********************************************************************
1524: * DMA utility functions
1525: **********************************************************************/
1526:
1527: /*
1528: * Utility function to load a linear buffer. lastaddrp holds state
1529: * between invocations (for multiple-buffer loads). segp contains
1530: * the starting segment on entrace, and the ending segment on exit.
1531: * first indicates if this is the first invocation of this function.
1532: */
1533: int
1.7 thorpej 1534: _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
1.48 yamt 1535: bus_size_t buflen, struct vmspace *vm, int flags)
1.1 chris 1536: {
1537: bus_size_t sgsize;
1.41 thorpej 1538: bus_addr_t curaddr;
1.11 thorpej 1539: vaddr_t vaddr = (vaddr_t)buf;
1.41 thorpej 1540: int error;
1.1 chris 1541: pmap_t pmap;
1542:
1543: #ifdef DEBUG_DMA
1.40 scw 1544: printf("_bus_dmamem_load_buffer(buf=%p, len=%lx, flags=%d)\n",
1545: buf, buflen, flags);
1.1 chris 1546: #endif /* DEBUG_DMA */
1547:
1.48 yamt 1548: pmap = vm_map_pmap(&vm->vm_map);
1.1 chris 1549:
1.41 thorpej 1550: while (buflen > 0) {
1.1 chris 1551: /*
1552: * Get the physical address for this segment.
1.17 thorpej 1553: *
1.1 chris 1554: */
1.61 matt 1555: bool coherent;
1.107 ryo 1556: pmap_extract_coherency(pmap, vaddr, &curaddr, &coherent);
1557:
1.86 matt 1558: KASSERTMSG((vaddr & PAGE_MASK) == (curaddr & PAGE_MASK),
1559: "va %#lx curaddr %#lx", vaddr, curaddr);
1.1 chris 1560:
1561: /*
1562: * Compute the segment size, and adjust counts.
1563: */
1.27 thorpej 1564: sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
1.1 chris 1565: if (buflen < sgsize)
1566: sgsize = buflen;
1567:
1.61 matt 1568: error = _bus_dmamap_load_paddr(t, map, curaddr, sgsize,
1569: coherent);
1.41 thorpej 1570: if (error)
1.100 skrll 1571: return error;
1.1 chris 1572:
1573: vaddr += sgsize;
1574: buflen -= sgsize;
1575: }
1576:
1.100 skrll 1577: return 0;
1.1 chris 1578: }
1579:
1580: /*
1581: * Allocate physical memory from the given physical address range.
1582: * Called by DMA-safe memory allocation methods.
1583: */
1584: int
1.7 thorpej 1585: _bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1586: bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1.11 thorpej 1587: int flags, paddr_t low, paddr_t high)
1.1 chris 1588: {
1.11 thorpej 1589: paddr_t curaddr, lastaddr;
1.1 chris 1590: struct vm_page *m;
1591: struct pglist mlist;
1592: int curseg, error;
1593:
1.101 skrll 1594: KASSERTMSG(boundary == 0 || (boundary & (boundary - 1)) == 0,
1.76 matt 1595: "invalid boundary %#lx", boundary);
1596:
1.1 chris 1597: #ifdef DEBUG_DMA
1598: printf("alloc_range: t=%p size=%lx align=%lx boundary=%lx segs=%p nsegs=%x rsegs=%p flags=%x lo=%lx hi=%lx\n",
1599: t, size, alignment, boundary, segs, nsegs, rsegs, flags, low, high);
1600: #endif /* DEBUG_DMA */
1601:
1602: /* Always round the size. */
1603: size = round_page(size);
1604:
1605: /*
1.76 matt 1606: * We accept boundaries < size, splitting in multiple segments
1607: * if needed. uvm_pglistalloc does not, so compute an appropriate
1608: * boundary: next power of 2 >= size
1609: */
1610: bus_size_t uboundary = boundary;
1611: if (uboundary <= PAGE_SIZE) {
1612: uboundary = 0;
1613: } else {
1614: while (uboundary < size) {
1615: uboundary <<= 1;
1616: }
1617: }
1618:
1619: /*
1.1 chris 1620: * Allocate pages from the VM system.
1621: */
1.78 matt 1622: error = uvm_pglistalloc(size, low, high, alignment, uboundary,
1.1 chris 1623: &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
1624: if (error)
1.100 skrll 1625: return error;
1.1 chris 1626:
1627: /*
1628: * Compute the location, size, and number of segments actually
1629: * returned by the VM code.
1630: */
1.42 chris 1631: m = TAILQ_FIRST(&mlist);
1.1 chris 1632: curseg = 0;
1633: lastaddr = segs[curseg].ds_addr = VM_PAGE_TO_PHYS(m);
1634: segs[curseg].ds_len = PAGE_SIZE;
1635: #ifdef DEBUG_DMA
1636: printf("alloc: page %lx\n", lastaddr);
1637: #endif /* DEBUG_DMA */
1.52 ad 1638: m = TAILQ_NEXT(m, pageq.queue);
1.1 chris 1639:
1.52 ad 1640: for (; m != NULL; m = TAILQ_NEXT(m, pageq.queue)) {
1.1 chris 1641: curaddr = VM_PAGE_TO_PHYS(m);
1.76 matt 1642: KASSERTMSG(low <= curaddr && curaddr < high,
1643: "uvm_pglistalloc returned non-sensicaladdress %#lx "
1644: "(low=%#lx, high=%#lx\n", curaddr, low, high);
1.1 chris 1645: #ifdef DEBUG_DMA
1646: printf("alloc: page %lx\n", curaddr);
1647: #endif /* DEBUG_DMA */
1.76 matt 1648: if (curaddr == lastaddr + PAGE_SIZE
1649: && (lastaddr & boundary) == (curaddr & boundary))
1.1 chris 1650: segs[curseg].ds_len += PAGE_SIZE;
1651: else {
1652: curseg++;
1.76 matt 1653: if (curseg >= nsegs) {
1654: uvm_pglistfree(&mlist);
1655: return EFBIG;
1656: }
1.1 chris 1657: segs[curseg].ds_addr = curaddr;
1658: segs[curseg].ds_len = PAGE_SIZE;
1659: }
1660: lastaddr = curaddr;
1661: }
1662:
1663: *rsegs = curseg + 1;
1664:
1.100 skrll 1665: return 0;
1.15 thorpej 1666: }
1667:
1668: /*
1669: * Check if a memory region intersects with a DMA range, and return the
1670: * page-rounded intersection if it does.
1671: */
1672: int
1673: arm32_dma_range_intersect(struct arm32_dma_range *ranges, int nranges,
1674: paddr_t pa, psize_t size, paddr_t *pap, psize_t *sizep)
1675: {
1676: struct arm32_dma_range *dr;
1677: int i;
1678:
1679: if (ranges == NULL)
1.100 skrll 1680: return 0;
1.15 thorpej 1681:
1682: for (i = 0, dr = ranges; i < nranges; i++, dr++) {
1683: if (dr->dr_sysbase <= pa &&
1684: pa < (dr->dr_sysbase + dr->dr_len)) {
1685: /*
1686: * Beginning of region intersects with this range.
1687: */
1688: *pap = trunc_page(pa);
1.112 riastrad 1689: *sizep = round_page(uimin(pa + size,
1.15 thorpej 1690: dr->dr_sysbase + dr->dr_len) - pa);
1.100 skrll 1691: return 1;
1.15 thorpej 1692: }
1693: if (pa < dr->dr_sysbase && dr->dr_sysbase < (pa + size)) {
1694: /*
1695: * End of region intersects with this range.
1696: */
1697: *pap = trunc_page(dr->dr_sysbase);
1.112 riastrad 1698: *sizep = round_page(uimin((pa + size) - dr->dr_sysbase,
1.15 thorpej 1699: dr->dr_len));
1.100 skrll 1700: return 1;
1.15 thorpej 1701: }
1702: }
1703:
1704: /* No intersection found. */
1.100 skrll 1705: return 0;
1.1 chris 1706: }
1.58 matt 1707:
1708: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1709: static int
1710: _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
1711: bus_size_t size, int flags)
1712: {
1713: struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
1714: int error = 0;
1715:
1.79 matt 1716: KASSERT(cookie != NULL);
1.58 matt 1717:
1718: cookie->id_bouncebuflen = round_page(size);
1719: error = _bus_dmamem_alloc(t, cookie->id_bouncebuflen,
1720: PAGE_SIZE, map->_dm_boundary, cookie->id_bouncesegs,
1721: map->_dm_segcnt, &cookie->id_nbouncesegs, flags);
1.76 matt 1722: if (error == 0) {
1723: error = _bus_dmamem_map(t, cookie->id_bouncesegs,
1724: cookie->id_nbouncesegs, cookie->id_bouncebuflen,
1725: (void **)&cookie->id_bouncebuf, flags);
1726: if (error) {
1727: _bus_dmamem_free(t, cookie->id_bouncesegs,
1728: cookie->id_nbouncesegs);
1729: cookie->id_bouncebuflen = 0;
1730: cookie->id_nbouncesegs = 0;
1731: } else {
1732: cookie->id_flags |= _BUS_DMA_HAS_BOUNCE;
1733: }
1734: } else {
1.58 matt 1735: cookie->id_bouncebuflen = 0;
1736: cookie->id_nbouncesegs = 0;
1737: }
1738:
1.100 skrll 1739: return error;
1.58 matt 1740: }
1741:
1742: static void
1743: _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map)
1744: {
1745: struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
1746:
1.79 matt 1747: KASSERT(cookie != NULL);
1.58 matt 1748:
1749: _bus_dmamem_unmap(t, cookie->id_bouncebuf, cookie->id_bouncebuflen);
1.79 matt 1750: _bus_dmamem_free(t, cookie->id_bouncesegs, cookie->id_nbouncesegs);
1.58 matt 1751: cookie->id_bouncebuflen = 0;
1752: cookie->id_nbouncesegs = 0;
1753: cookie->id_flags &= ~_BUS_DMA_HAS_BOUNCE;
1754: }
1755:
1756: /*
1757: * This function does the same as uiomove, but takes an explicit
1758: * direction, and does not update the uio structure.
1759: */
1760: static int
1761: _bus_dma_uiomove(void *buf, struct uio *uio, size_t n, int direction)
1762: {
1763: struct iovec *iov;
1764: int error;
1765: struct vmspace *vm;
1766: char *cp;
1767: size_t resid, cnt;
1768: int i;
1769:
1770: iov = uio->uio_iov;
1771: vm = uio->uio_vmspace;
1772: cp = buf;
1773: resid = n;
1774:
1775: for (i = 0; i < uio->uio_iovcnt && resid > 0; i++) {
1776: iov = &uio->uio_iov[i];
1777: if (iov->iov_len == 0)
1778: continue;
1779: cnt = MIN(resid, iov->iov_len);
1780:
1781: if (!VMSPACE_IS_KERNEL_P(vm) &&
1782: (curlwp->l_cpu->ci_schedstate.spc_flags & SPCF_SHOULDYIELD)
1783: != 0) {
1784: preempt();
1785: }
1786: if (direction == UIO_READ) {
1787: error = copyout_vmspace(vm, cp, iov->iov_base, cnt);
1788: } else {
1789: error = copyin_vmspace(vm, iov->iov_base, cp, cnt);
1790: }
1791: if (error)
1.100 skrll 1792: return error;
1.58 matt 1793: cp += cnt;
1794: resid -= cnt;
1795: }
1.100 skrll 1796: return 0;
1.58 matt 1797: }
1798: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1799:
1800: int
1801: _bus_dmatag_subregion(bus_dma_tag_t tag, bus_addr_t min_addr,
1802: bus_addr_t max_addr, bus_dma_tag_t *newtag, int flags)
1803: {
1804:
1805: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1806: struct arm32_dma_range *dr;
1807: bool subset = false;
1808: size_t nranges = 0;
1809: size_t i;
1810: for (i = 0, dr = tag->_ranges; i < tag->_nranges; i++, dr++) {
1.95 skrll 1811: if (dr->dr_sysbase <= min_addr
1.58 matt 1812: && max_addr <= dr->dr_sysbase + dr->dr_len - 1) {
1813: subset = true;
1814: }
1815: if (min_addr <= dr->dr_sysbase + dr->dr_len
1816: && max_addr >= dr->dr_sysbase) {
1817: nranges++;
1818: }
1819: }
1820: if (subset) {
1821: *newtag = tag;
1822: /* if the tag must be freed, add a reference */
1823: if (tag->_tag_needs_free)
1824: (tag->_tag_needs_free)++;
1825: return 0;
1826: }
1827: if (nranges == 0) {
1828: nranges = 1;
1829: }
1830:
1.81 matt 1831: const size_t tagsize = sizeof(*tag) + nranges * sizeof(*dr);
1832: if ((*newtag = kmem_intr_zalloc(tagsize,
1833: (flags & BUS_DMA_NOWAIT) ? KM_NOSLEEP : KM_SLEEP)) == NULL)
1.58 matt 1834: return ENOMEM;
1835:
1836: dr = (void *)(*newtag + 1);
1837: **newtag = *tag;
1838: (*newtag)->_tag_needs_free = 1;
1839: (*newtag)->_ranges = dr;
1840: (*newtag)->_nranges = nranges;
1841:
1842: if (tag->_ranges == NULL) {
1843: dr->dr_sysbase = min_addr;
1844: dr->dr_busbase = min_addr;
1845: dr->dr_len = max_addr + 1 - min_addr;
1846: } else {
1847: for (i = 0; i < nranges; i++) {
1848: if (min_addr > dr->dr_sysbase + dr->dr_len
1849: || max_addr < dr->dr_sysbase)
1850: continue;
1851: dr[0] = tag->_ranges[i];
1852: if (dr->dr_sysbase < min_addr) {
1853: psize_t diff = min_addr - dr->dr_sysbase;
1854: dr->dr_busbase += diff;
1855: dr->dr_len -= diff;
1856: dr->dr_sysbase += diff;
1857: }
1858: if (max_addr != 0xffffffff
1859: && max_addr + 1 < dr->dr_sysbase + dr->dr_len) {
1860: dr->dr_len = max_addr + 1 - dr->dr_sysbase;
1861: }
1862: dr++;
1863: }
1864: }
1865:
1866: return 0;
1867: #else
1868: return EOPNOTSUPP;
1869: #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
1870: }
1871:
1872: void
1873: _bus_dmatag_destroy(bus_dma_tag_t tag)
1874: {
1875: #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
1876: switch (tag->_tag_needs_free) {
1877: case 0:
1.81 matt 1878: break; /* not allocated with kmem */
1879: case 1: {
1880: const size_t tagsize = sizeof(*tag)
1881: + tag->_nranges * sizeof(*tag->_ranges);
1882: kmem_intr_free(tag, tagsize); /* last reference to tag */
1.58 matt 1883: break;
1.81 matt 1884: }
1.58 matt 1885: default:
1886: (tag->_tag_needs_free)--; /* one less reference */
1887: }
1888: #endif
1889: }
CVSweb <webmaster@jp.NetBSD.org>