xref: /netbsd-src/sys/arch/riscv/riscv/bus_dma.c (revision fec46068b5c9dad78e835e987c10fecbc13f77f1)
1 /*	$NetBSD: bus_dma.c,v 1.9 2024/12/10 07:42:03 skrll Exp $	*/
2 
3 /*-
4  * Copyright (c) 1996, 1997, 1998, 2020 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 _RISCV_BUS_DMA_PRIVATE
34 #define _RISCV_NEED_BUS_DMA_BOUNCE
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: bus_dma.c,v 1.9 2024/12/10 07:42:03 skrll Exp $");
38 
39 #include <sys/param.h>
40 
41 #include <sys/bus.h>
42 #include <sys/cpu.h>
43 #include <sys/kmem.h>
44 #include <sys/mbuf.h>
45 
46 #include <uvm/uvm.h>
47 
48 #include <machine/cpufunc.h>
49 
50 #define BUSDMA_COUNTERS
51 #ifdef BUSDMA_COUNTERS
52 static struct evcnt bus_dma_creates =
53 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "creates");
54 static struct evcnt bus_dma_bounced_creates =
55 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced creates");
56 static struct evcnt bus_dma_loads =
57 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "loads");
58 static struct evcnt bus_dma_bounced_loads =
59 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced loads");
60 static struct evcnt bus_dma_coherent_loads =
61 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "coherent loads");
62 static struct evcnt bus_dma_read_bounces =
63 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "read bounces");
64 static struct evcnt bus_dma_write_bounces =
65 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "write bounces");
66 static struct evcnt bus_dma_bounced_unloads =
67 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced unloads");
68 static struct evcnt bus_dma_bounced_mbuf_loads =
69 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced mbuf loads");
70 static struct evcnt bus_dma_unloads =
71 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "unloads");
72 static struct evcnt bus_dma_bounced_destroys =
73 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced destroys");
74 static struct evcnt bus_dma_destroys =
75 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "destroys");
76 static struct evcnt bus_dma_sync_prereadwrite =
77 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync prereadwrite");
78 static struct evcnt bus_dma_sync_preread_begin =
79 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync preread begin");
80 static struct evcnt bus_dma_sync_preread =
81 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync preread");
82 static struct evcnt bus_dma_sync_preread_tail =
83 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync preread tail");
84 static struct evcnt bus_dma_sync_prewrite =
85 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync prewrite");
86 static struct evcnt bus_dma_sync_postread =
87 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postread");
88 static struct evcnt bus_dma_sync_postreadwrite =
89 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postreadwrite");
90 static struct evcnt bus_dma_sync_postwrite =
91 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postwrite");
92 static struct evcnt bus_dma_inrange_fail =
93 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "inrange check failed");
94 
95 static struct evcnt bus_dma_sync_coherent_prereadwrite =
96 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync coherent prereadwrite");
97 static struct evcnt bus_dma_sync_coherent_preread =
98 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync coherent preread");
99 static struct evcnt bus_dma_sync_coherent_prewrite =
100 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync coherent prewrite");
101 static struct evcnt bus_dma_sync_coherent_postread =
102 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync coherent postread");
103 static struct evcnt bus_dma_sync_coherent_postreadwrite =
104 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync coherent postreadwrite");
105 static struct evcnt bus_dma_sync_coherent_postwrite =
106 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync coherent postwrite");
107 
108 EVCNT_ATTACH_STATIC(bus_dma_creates);
109 EVCNT_ATTACH_STATIC(bus_dma_bounced_creates);
110 EVCNT_ATTACH_STATIC(bus_dma_loads);
111 EVCNT_ATTACH_STATIC(bus_dma_bounced_loads);
112 EVCNT_ATTACH_STATIC(bus_dma_coherent_loads);
113 EVCNT_ATTACH_STATIC(bus_dma_read_bounces);
114 EVCNT_ATTACH_STATIC(bus_dma_write_bounces);
115 EVCNT_ATTACH_STATIC(bus_dma_unloads);
116 EVCNT_ATTACH_STATIC(bus_dma_bounced_unloads);
117 EVCNT_ATTACH_STATIC(bus_dma_destroys);
118 EVCNT_ATTACH_STATIC(bus_dma_bounced_destroys);
119 EVCNT_ATTACH_STATIC(bus_dma_bounced_mbuf_loads);
120 EVCNT_ATTACH_STATIC(bus_dma_sync_prereadwrite);
121 EVCNT_ATTACH_STATIC(bus_dma_sync_preread_begin);
122 EVCNT_ATTACH_STATIC(bus_dma_sync_preread);
123 EVCNT_ATTACH_STATIC(bus_dma_sync_preread_tail);
124 EVCNT_ATTACH_STATIC(bus_dma_sync_prewrite);
125 EVCNT_ATTACH_STATIC(bus_dma_sync_postread);
126 EVCNT_ATTACH_STATIC(bus_dma_sync_postreadwrite);
127 EVCNT_ATTACH_STATIC(bus_dma_sync_postwrite);
128 EVCNT_ATTACH_STATIC(bus_dma_inrange_fail);
129 
130 EVCNT_ATTACH_STATIC(bus_dma_sync_coherent_prereadwrite);
131 EVCNT_ATTACH_STATIC(bus_dma_sync_coherent_preread);
132 EVCNT_ATTACH_STATIC(bus_dma_sync_coherent_prewrite);
133 EVCNT_ATTACH_STATIC(bus_dma_sync_coherent_postread);
134 EVCNT_ATTACH_STATIC(bus_dma_sync_coherent_postreadwrite);
135 EVCNT_ATTACH_STATIC(bus_dma_sync_coherent_postwrite);
136 
137 #define	STAT_INCR(x)	(bus_dma_ ## x.ev_count++)
138 #else
139 #define	STAT_INCR(x)	__nothing
140 #endif
141 
142 int	_bus_dmamap_load_buffer(bus_dma_tag_t, bus_dmamap_t, void *,
143 	    bus_size_t, struct vmspace *, int);
144 
145 /*
146  * Check to see if the specified page is in an allowed DMA range.
147  */
148 static inline struct riscv_dma_range *
149 _bus_dma_paddr_inrange(struct riscv_dma_range *ranges, int nranges,
150     bus_addr_t curaddr)
151 {
152 	struct riscv_dma_range *dr;
153 	int i;
154 
155 	for (i = 0, dr = ranges; i < nranges; i++, dr++) {
156 		if (curaddr >= dr->dr_sysbase &&
157 		    curaddr < (dr->dr_sysbase + dr->dr_len))
158 			return dr;
159 	}
160 
161 	return NULL;
162 }
163 
164 /*
165  * Check to see if the specified busaddr is in an allowed DMA range.
166  */
167 static inline paddr_t
168 _bus_dma_busaddr_to_paddr(bus_dma_tag_t t, bus_addr_t curaddr)
169 {
170 	struct riscv_dma_range *dr;
171 	u_int i;
172 
173 	if (t->_nranges == 0)
174 		return curaddr;
175 
176 	for (i = 0, dr = t->_ranges; i < t->_nranges; i++, dr++) {
177 		if (dr->dr_busbase <= curaddr
178 		    && curaddr < dr->dr_busbase + dr->dr_len)
179 			return curaddr - dr->dr_busbase + dr->dr_sysbase;
180 	}
181 	panic("%s: curaddr %#" PRIxBUSADDR "not in range", __func__, curaddr);
182 }
183 
184 /*
185  * Common function to load the specified physical address into the
186  * DMA map, coalescing segments and boundary checking as necessary.
187  */
188 static int
189 _bus_dmamap_load_paddr(bus_dma_tag_t t, bus_dmamap_t map,
190     bus_addr_t paddr, bus_size_t size, bool coherent)
191 {
192 	bus_dma_segment_t * const segs = map->dm_segs;
193 	int nseg = map->dm_nsegs;
194 	bus_addr_t lastaddr;
195 	bus_addr_t bmask = ~(map->_dm_boundary - 1);
196 	bus_addr_t curaddr;
197 	bus_size_t sgsize;
198 	uint32_t _ds_flags = coherent ? _BUS_DMAMAP_COHERENT : 0;
199 
200 	if (nseg > 0)
201 		lastaddr = segs[nseg - 1].ds_addr + segs[nseg - 1].ds_len;
202 	else
203 		lastaddr = 0xdead;
204 
205  again:
206 	sgsize = size;
207 
208 	/* Make sure we're in an allowed DMA range. */
209 	if (t->_ranges != NULL) {
210 		/* XXX cache last result? */
211 		const struct riscv_dma_range * const dr =
212 		    _bus_dma_paddr_inrange(t->_ranges, t->_nranges, paddr);
213 		if (__predict_false(dr == NULL)) {
214 			STAT_INCR(inrange_fail);
215 			return EINVAL;
216 		}
217 
218 		/*
219 		 * If this region is coherent, mark the segment as coherent.
220 		 */
221 		_ds_flags |= dr->dr_flags & _BUS_DMAMAP_COHERENT;
222 
223 		/*
224 		 * In a valid DMA range.  Translate the physical
225 		 * memory address to an address in the DMA window.
226 		 */
227 		curaddr = (paddr - dr->dr_sysbase) + dr->dr_busbase;
228 #if 0
229 		printf("%p: %#" PRIxPADDR
230 		    ": range %#" PRIxPADDR "/%#" PRIxBUSADDR
231 		    "/%#" PRIxBUSSIZE "/%#" PRIx32 ": %#" PRIx32
232 		    " <-- %#" PRIxBUSADDR "\n",
233 		    t, paddr, dr->dr_sysbase, dr->dr_busbase,
234 		    dr->dr_len, dr->dr_flags, _ds_flags, curaddr);
235 #endif
236 	} else
237 		curaddr = paddr;
238 
239 	/*
240 	 * Make sure we don't cross any boundaries.
241 	 */
242 	if (map->_dm_boundary > 0) {
243 		bus_addr_t baddr;	/* next boundary address */
244 
245 		baddr = (curaddr + map->_dm_boundary) & bmask;
246 		if (sgsize > (baddr - curaddr))
247 			sgsize = (baddr - curaddr);
248 	}
249 
250 	/*
251 	 * Insert chunk into a segment, coalescing with the
252 	 * previous segment if possible.
253 	 */
254 	if (nseg > 0 && curaddr == lastaddr &&
255 	    segs[nseg - 1].ds_len + sgsize <= map->dm_maxsegsz &&
256 	    ((segs[nseg - 1]._ds_flags ^ _ds_flags) & _BUS_DMAMAP_COHERENT) == 0 &&
257 	    (map->_dm_boundary == 0 ||
258 	     (segs[nseg - 1].ds_addr & bmask) == (curaddr & bmask))) {
259 		/* coalesce */
260 		segs[nseg - 1].ds_len += sgsize;
261 	} else if (__predict_false(nseg >= map->_dm_segcnt)) {
262 		return EFBIG;
263 	} else {
264 		/* new segment */
265 		segs[nseg].ds_addr = curaddr;
266 		segs[nseg].ds_len = sgsize;
267 		segs[nseg]._ds_paddr = curaddr;
268 		segs[nseg]._ds_flags = _ds_flags;
269 		nseg++;
270 	}
271 
272 	lastaddr = curaddr + sgsize;
273 
274 	paddr += sgsize;
275 	size -= sgsize;
276 	if (size > 0)
277 		goto again;
278 
279 	map->_dm_flags &= (_ds_flags & _BUS_DMAMAP_COHERENT);
280 	map->dm_nsegs = nseg;
281 	return 0;
282 }
283 
284 static int _bus_dma_uiomove(void *buf, struct uio *uio, size_t n,
285 	    int direction);
286 
287 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
288 static int _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
289 	    bus_size_t size, int flags);
290 static void _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map);
291 
292 static int
293 _bus_dma_load_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
294 	size_t buflen, int buftype, int flags)
295 {
296 	struct riscv_bus_dma_cookie * const cookie = map->_dm_cookie;
297 	struct vmspace * const vm = vmspace_kernel();
298 	int error;
299 
300 	KASSERT(cookie != NULL);
301 	KASSERT(cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE);
302 
303 	/*
304 	 * Allocate bounce pages, if necessary.
305 	 */
306 	if ((cookie->id_flags & _BUS_DMA_HAS_BOUNCE) == 0) {
307 		error = _bus_dma_alloc_bouncebuf(t, map, buflen, flags);
308 		if (__predict_false(error))
309 			return error;
310 	}
311 
312 	/*
313 	 * Since we're trying again, clear the previous attempt.
314 	 */
315 	map->dm_mapsize = 0;
316 	map->dm_nsegs = 0;
317 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
318 	/* _bus_dmamap_load_buffer() clears this if we're not... */
319 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
320 
321 	/*
322 	 * Cache a pointer to the caller's buffer and load the DMA map
323 	 * with the bounce buffer.
324 	 */
325 	cookie->id_origbuf = buf;
326 	cookie->id_origbuflen = buflen;
327 	error = _bus_dmamap_load_buffer(t, map, cookie->id_bouncebuf,
328 	    buflen, vm, flags);
329 	if (__predict_false(error))
330 		return error;
331 
332 	STAT_INCR(bounced_loads);
333 	map->dm_mapsize = buflen;
334 	map->_dm_vmspace = vm;
335 	map->_dm_buftype = buftype;
336 
337 	/* ...so _bus_dmamap_sync() knows we're bouncing */
338 	map->_dm_flags |= _BUS_DMAMAP_IS_BOUNCING;
339 	cookie->id_flags |= _BUS_DMA_IS_BOUNCING;
340 	return 0;
341 }
342 #endif /* _RISCV_NEED_BUS_DMA_BOUNCE */
343 
344 /*
345  * Common function for DMA map creation.  May be called by bus-specific
346  * DMA map creation functions.
347  */
348 int
349 _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
350     bus_size_t maxsegsz, bus_size_t boundary, int flags, bus_dmamap_t *dmamp)
351 {
352 	struct riscv_bus_dmamap *map;
353 	void *mapstore;
354 	int error = 0;
355 
356 #ifdef DEBUG_DMA
357 	printf("dmamap_create: t=%p size=%#" PRIxBUSSIZE
358 	    " nseg=%#x msegsz=%#" PRIxBUSSIZE
359 	    " boundary=%#" PRIxBUSSIZE
360 	    " flags=%#x\n", t, size, nsegments, maxsegsz, boundary, flags);
361 #endif	/* DEBUG_DMA */
362 
363 	/*
364 	 * Allocate and initialize the DMA map.  The end of the map
365 	 * is a variable-sized array of segments, so we allocate enough
366 	 * room for them in one shot.
367 	 *
368 	 * Note we don't preserve the WAITOK or NOWAIT flags.  Preservation
369 	 * of ALLOCNOW notifies others that we've reserved these resources,
370 	 * and they are not to be freed.
371 	 *
372 	 * The bus_dmamap_t includes one bus_dma_segment_t, hence
373 	 * the (nsegments - 1).
374 	 */
375 	const size_t mapsize = sizeof(struct riscv_bus_dmamap) +
376 	    (sizeof(bus_dma_segment_t) * (nsegments - 1));
377 	const int zallocflags = (flags & BUS_DMA_NOWAIT) ? KM_NOSLEEP : KM_SLEEP;
378 	if ((mapstore = kmem_intr_zalloc(mapsize, zallocflags)) == NULL)
379 		return ENOMEM;
380 
381 	map = (struct riscv_bus_dmamap *)mapstore;
382 	map->_dm_size = size;
383 	map->_dm_segcnt = nsegments;
384 	map->_dm_maxmaxsegsz = maxsegsz;
385 	map->_dm_boundary = boundary;
386 	map->_dm_flags = flags & ~(BUS_DMA_WAITOK | BUS_DMA_NOWAIT);
387 	map->_dm_origbuf = NULL;
388 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
389 	map->_dm_vmspace = vmspace_kernel();
390 	map->_dm_cookie = NULL;
391 	map->dm_maxsegsz = maxsegsz;
392 	map->dm_mapsize = 0;		/* no valid mappings */
393 	map->dm_nsegs = 0;
394 
395 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
396 	struct riscv_bus_dma_cookie *cookie;
397 	int cookieflags;
398 	void *cookiestore;
399 
400 	cookieflags = 0;
401 
402 	if (t->_may_bounce != NULL) {
403 		error = (*t->_may_bounce)(t, map, flags, &cookieflags);
404 		if (error != 0)
405 			goto out;
406 	}
407 
408 	if (t->_ranges != NULL) {
409 		/*
410 		 * If ranges are defined, we may have to bounce. The only
411 		 * exception is if there is exactly one range that covers
412 		 * all of physical memory.
413 		 */
414 		switch (t->_nranges) {
415 		case 1:
416 			if (t->_ranges[0].dr_sysbase == 0 &&
417 			    t->_ranges[0].dr_len == UINTPTR_MAX) {
418 				break;
419 			}
420 			/* FALLTHROUGH */
421 		default:
422 			cookieflags |= _BUS_DMA_MIGHT_NEED_BOUNCE;
423 		}
424 	}
425 
426 	if ((cookieflags & _BUS_DMA_MIGHT_NEED_BOUNCE) == 0) {
427 		STAT_INCR(creates);
428 		*dmamp = map;
429 		return 0;
430 	}
431 
432 	const size_t cookiesize = sizeof(struct riscv_bus_dma_cookie) +
433 	    (sizeof(bus_dma_segment_t) * map->_dm_segcnt);
434 
435 	/*
436 	 * Allocate our cookie.
437 	 */
438 	if ((cookiestore = kmem_intr_zalloc(cookiesize, zallocflags)) == NULL) {
439 		error = ENOMEM;
440 		goto out;
441 	}
442 	cookie = (struct riscv_bus_dma_cookie *)cookiestore;
443 	cookie->id_flags = cookieflags;
444 	map->_dm_cookie = cookie;
445 	STAT_INCR(bounced_creates);
446 
447 	error = _bus_dma_alloc_bouncebuf(t, map, size, flags);
448  out:
449 	if (error)
450 		_bus_dmamap_destroy(t, map);
451 	else
452 		*dmamp = map;
453 #else
454 	*dmamp = map;
455 	STAT_INCR(creates);
456 #endif /* _RISCV_NEED_BUS_DMA_BOUNCE */
457 #ifdef DEBUG_DMA
458 	printf("dmamap_create:map=%p\n", map);
459 #endif	/* DEBUG_DMA */
460 	return error;
461 }
462 
463 /*
464  * Common function for DMA map destruction.  May be called by bus-specific
465  * DMA map destruction functions.
466  */
467 void
468 _bus_dmamap_destroy(bus_dma_tag_t t, bus_dmamap_t map)
469 {
470 
471 #ifdef DEBUG_DMA
472 	printf("dmamap_destroy: t=%p map=%p\n", t, map);
473 #endif	/* DEBUG_DMA */
474 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
475 	struct riscv_bus_dma_cookie *cookie = map->_dm_cookie;
476 
477 	/*
478 	 * Free any bounce pages this map might hold.
479 	 */
480 	if (cookie != NULL) {
481 		const size_t cookiesize = sizeof(struct riscv_bus_dma_cookie) +
482 		    (sizeof(bus_dma_segment_t) * map->_dm_segcnt);
483 
484 		if (cookie->id_flags & _BUS_DMA_IS_BOUNCING)
485 			STAT_INCR(bounced_unloads);
486 		map->dm_nsegs = 0;
487 		if (cookie->id_flags & _BUS_DMA_HAS_BOUNCE)
488 			_bus_dma_free_bouncebuf(t, map);
489 		STAT_INCR(bounced_destroys);
490 		kmem_intr_free(cookie, cookiesize);
491 	} else
492 #endif
493 	STAT_INCR(destroys);
494 
495 	if (map->dm_nsegs > 0)
496 		STAT_INCR(unloads);
497 
498 	const size_t mapsize = sizeof(struct riscv_bus_dmamap) +
499 	    (sizeof(bus_dma_segment_t) * (map->_dm_segcnt - 1));
500 	kmem_intr_free(map, mapsize);
501 }
502 
503 /*
504  * Common function for loading a DMA map with a linear buffer.  May
505  * be called by bus-specific DMA map load functions.
506  */
507 int
508 _bus_dmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
509     bus_size_t buflen, struct proc *p, int flags)
510 {
511 	struct vmspace *vm;
512 	int error;
513 
514 #ifdef DEBUG_DMA
515 	printf("dmamap_load: t=%p map=%p buf=%p len=%#" PRIxBUSSIZE
516 	    " p=%p f=%#x\n", t, map, buf, buflen, p, flags);
517 #endif	/* DEBUG_DMA */
518 
519 	if (map->dm_nsegs > 0) {
520 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
521 		struct riscv_bus_dma_cookie *cookie = map->_dm_cookie;
522 		if (cookie != NULL) {
523 			if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
524 				STAT_INCR(bounced_unloads);
525 				cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
526 				map->_dm_flags &= ~_BUS_DMAMAP_IS_BOUNCING;
527 			}
528 		} else
529 #endif
530 		STAT_INCR(unloads);
531 	}
532 
533 	/*
534 	 * Make sure that on error condition we return "no valid mappings".
535 	 */
536 	map->dm_mapsize = 0;
537 	map->dm_nsegs = 0;
538 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
539 	KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
540 	    "dm_maxsegsz %" PRIuBUSSIZE " _dm_maxmaxsegsz %" PRIuBUSSIZE,
541 	    map->dm_maxsegsz, map->_dm_maxmaxsegsz);
542 
543 	if (__predict_false(buflen > map->_dm_size))
544 		return EINVAL;
545 
546 	if (p != NULL) {
547 		vm = p->p_vmspace;
548 	} else {
549 		vm = vmspace_kernel();
550 	}
551 
552 	/* _bus_dmamap_load_buffer() clears this if we're not... */
553 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
554 
555 	error = _bus_dmamap_load_buffer(t, map, buf, buflen, vm, flags);
556 	if (__predict_true(error == 0)) {
557 		map->dm_mapsize = buflen;
558 		map->_dm_vmspace = vm;
559 		map->_dm_origbuf = buf;
560 		map->_dm_buftype = _BUS_DMA_BUFTYPE_LINEAR;
561 		if (map->_dm_flags & _BUS_DMAMAP_COHERENT) {
562 			STAT_INCR(coherent_loads);
563 		} else {
564 			STAT_INCR(loads);
565 		}
566 		return 0;
567 	}
568 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
569 	struct riscv_bus_dma_cookie * const cookie = map->_dm_cookie;
570 	if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
571 		error = _bus_dma_load_bouncebuf(t, map, buf, buflen,
572 		    _BUS_DMA_BUFTYPE_LINEAR, flags);
573 	}
574 #endif
575 	return error;
576 }
577 
578 /*
579  * Like _bus_dmamap_load(), but for mbufs.
580  */
581 int
582 _bus_dmamap_load_mbuf(bus_dma_tag_t t, bus_dmamap_t map, struct mbuf *m0,
583     int flags)
584 {
585 	struct mbuf *m;
586 	int error;
587 
588 #ifdef DEBUG_DMA
589 	printf("dmamap_load_mbuf: t=%p map=%p m0=%p f=%#x\n",
590 	    t, map, m0, flags);
591 #endif	/* DEBUG_DMA */
592 
593 	if (map->dm_nsegs > 0) {
594 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
595 		struct riscv_bus_dma_cookie *cookie = map->_dm_cookie;
596 		if (cookie != NULL) {
597 			if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
598 				STAT_INCR(bounced_unloads);
599 				cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
600 				map->_dm_flags &= ~_BUS_DMAMAP_IS_BOUNCING;
601 			}
602 		} else
603 #endif
604 		STAT_INCR(unloads);
605 	}
606 
607 	/*
608 	 * Make sure that on error condition we return "no valid mappings."
609 	 */
610 	map->dm_mapsize = 0;
611 	map->dm_nsegs = 0;
612 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
613 	KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
614 	    "dm_maxsegsz %" PRIuBUSSIZE " _dm_maxmaxsegsz %" PRIuBUSSIZE,
615 	    map->dm_maxsegsz, map->_dm_maxmaxsegsz);
616 
617 	KASSERT(m0->m_flags & M_PKTHDR);
618 
619 	if (__predict_false(m0->m_pkthdr.len > map->_dm_size))
620 		return EINVAL;
621 
622 	/* _bus_dmamap_load_paddr() clears this if we're not... */
623 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
624 
625 	error = 0;
626 	for (m = m0; m != NULL && error == 0; m = m->m_next) {
627 		int offset;
628 		int remainbytes;
629 		const struct vm_page * const *pgs;
630 		paddr_t paddr;
631 		int size;
632 
633 		if (m->m_len == 0)
634 			continue;
635 		/*
636 		 * Don't allow reads in read-only mbufs.
637 		 */
638 		if (__predict_false(M_ROMAP(m) && (flags & BUS_DMA_READ))) {
639 			error = EFAULT;
640 			break;
641 		}
642 		switch (m->m_flags & (M_EXT | M_EXT_CLUSTER | M_EXT_PAGES)) {
643 		case M_EXT | M_EXT_CLUSTER:
644 			KASSERT(m->m_ext.ext_paddr != M_PADDR_INVALID);
645 			paddr = m->m_ext.ext_paddr +
646 			    (m->m_data - m->m_ext.ext_buf);
647 			size = m->m_len;
648 			error = _bus_dmamap_load_paddr(t, map, paddr, size,
649 			    false);
650 			break;
651 
652 		case M_EXT | M_EXT_PAGES:
653 			KASSERT(m->m_ext.ext_buf <= m->m_data);
654 			KASSERT(m->m_data <=
655 			    m->m_ext.ext_buf + m->m_ext.ext_size);
656 
657 			offset = (vaddr_t)m->m_data -
658 			    trunc_page((vaddr_t)m->m_ext.ext_buf);
659 			remainbytes = m->m_len;
660 
661 			/* skip uninteresting pages */
662 			pgs = (const struct vm_page * const *)
663 			    m->m_ext.ext_pgs + (offset >> PAGE_SHIFT);
664 
665 			offset &= PAGE_MASK;	/* offset in the first page */
666 
667 			/* load each page */
668 			while (remainbytes > 0) {
669 				const struct vm_page *pg;
670 
671 				size = MIN(remainbytes, PAGE_SIZE - offset);
672 
673 				pg = *pgs++;
674 				KASSERT(pg);
675 				paddr = VM_PAGE_TO_PHYS(pg) + offset;
676 
677 				error = _bus_dmamap_load_paddr(t, map,
678 				    paddr, size, false);
679 				if (__predict_false(error))
680 					break;
681 				offset = 0;
682 				remainbytes -= size;
683 			}
684 			break;
685 
686 		case 0:
687 			paddr = m->m_paddr + M_BUFOFFSET(m) +
688 			    (m->m_data - M_BUFADDR(m));
689 			size = m->m_len;
690 			error = _bus_dmamap_load_paddr(t, map, paddr, size,
691 			    false);
692 			break;
693 
694 		default:
695 			error = _bus_dmamap_load_buffer(t, map, m->m_data,
696 			    m->m_len, vmspace_kernel(), flags);
697 		}
698 	}
699 	if (__predict_true(error == 0)) {
700 		map->dm_mapsize = m0->m_pkthdr.len;
701 		map->_dm_origbuf = m0;
702 		map->_dm_buftype = _BUS_DMA_BUFTYPE_MBUF;
703 		map->_dm_vmspace = vmspace_kernel();	/* always kernel */
704 		if (map->_dm_flags & _BUS_DMAMAP_COHERENT) {
705 			STAT_INCR(coherent_loads);
706 		} else {
707 			STAT_INCR(loads);
708 		}
709 		return 0;
710 	}
711 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
712 	struct riscv_bus_dma_cookie * const cookie = map->_dm_cookie;
713 	if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
714 		error = _bus_dma_load_bouncebuf(t, map, m0, m0->m_pkthdr.len,
715 		    _BUS_DMA_BUFTYPE_MBUF, flags);
716 		STAT_INCR(bounced_mbuf_loads);
717 	}
718 #endif
719 	return error;
720 }
721 
722 /*
723  * Like _bus_dmamap_load(), but for uios.
724  */
725 int
726 _bus_dmamap_load_uio(bus_dma_tag_t t, bus_dmamap_t map, struct uio *uio,
727     int flags)
728 {
729 	bus_size_t minlen, resid;
730 	struct iovec *iov;
731 	void *addr;
732 	int i, error;
733 
734 	/*
735 	 * Make sure that on error condition we return "no valid mappings."
736 	 */
737 	map->dm_mapsize = 0;
738 	map->dm_nsegs = 0;
739 	KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
740 	    "dm_maxsegsz %" PRIuBUSSIZE " _dm_maxmaxsegsz %" PRIuBUSSIZE,
741 	    map->dm_maxsegsz, map->_dm_maxmaxsegsz);
742 
743 	resid = uio->uio_resid;
744 	iov = uio->uio_iov;
745 
746 	/* _bus_dmamap_load_buffer() clears this if we're not... */
747 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
748 
749 	error = 0;
750 	for (i = 0; i < uio->uio_iovcnt && resid != 0 && error == 0; i++) {
751 		/*
752 		 * Now at the first iovec to load.  Load each iovec
753 		 * until we have exhausted the residual count.
754 		 */
755 		minlen = resid < iov[i].iov_len ? resid : iov[i].iov_len;
756 		addr = (void *)iov[i].iov_base;
757 
758 		error = _bus_dmamap_load_buffer(t, map, addr, minlen,
759 		    uio->uio_vmspace, flags);
760 
761 		resid -= minlen;
762 	}
763 	if (__predict_true(error == 0)) {
764 		map->dm_mapsize = uio->uio_resid;
765 		map->_dm_origbuf = uio;
766 		map->_dm_buftype = _BUS_DMA_BUFTYPE_UIO;
767 		map->_dm_vmspace = uio->uio_vmspace;
768 		if (map->_dm_flags & _BUS_DMAMAP_COHERENT) {
769 			STAT_INCR(coherent_loads);
770 		} else {
771 			STAT_INCR(loads);
772 		}
773 	}
774 	return error;
775 }
776 
777 /*
778  * Like _bus_dmamap_load(), but for raw memory allocated with
779  * bus_dmamem_alloc().
780  */
781 int
782 _bus_dmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
783     bus_dma_segment_t *segs, int nsegs, bus_size_t size0, int flags)
784 {
785 
786 	bus_size_t size;
787 	int i, error = 0;
788 
789 	/*
790 	 * Make sure that on error conditions we return "no valid mappings."
791 	 */
792 	map->dm_mapsize = 0;
793 	map->dm_nsegs = 0;
794 	KASSERT(map->dm_maxsegsz <= map->_dm_maxmaxsegsz);
795 
796 	if (__predict_false(size0 > map->_dm_size))
797 		return EINVAL;
798 
799 	for (i = 0, size = size0; i < nsegs && size > 0; i++) {
800 		bus_dma_segment_t *ds = &segs[i];
801 		bus_size_t sgsize;
802 
803 		sgsize = MIN(ds->ds_len, size);
804 		if (sgsize == 0)
805 			continue;
806 		const bool coherent =
807 		    (ds->_ds_flags & _BUS_DMAMAP_COHERENT) != 0;
808 		error = _bus_dmamap_load_paddr(t, map, ds->ds_addr,
809 		    sgsize, coherent);
810 		if (__predict_false(error != 0))
811 			break;
812 		size -= sgsize;
813 	}
814 
815 	if (__predict_false(error != 0)) {
816 		map->dm_mapsize = 0;
817 		map->dm_nsegs = 0;
818 		return error;
819 	}
820 
821 	/* XXX TBD bounce */
822 
823 	map->dm_mapsize = size0;
824 	map->_dm_origbuf = NULL;
825 	map->_dm_buftype = _BUS_DMA_BUFTYPE_RAW;
826 	map->_dm_vmspace = NULL;
827 	return 0;
828 }
829 
830 /*
831  * Common function for unloading a DMA map.  May be called by
832  * bus-specific DMA map unload functions.
833  */
834 void
835 _bus_dmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
836 {
837 
838 #ifdef DEBUG_DMA
839 	printf("dmamap_unload: t=%p map=%p\n", t, map);
840 #endif	/* DEBUG_DMA */
841 
842 	/*
843 	 * No resources to free; just mark the mappings as
844 	 * invalid.
845 	 */
846 	map->dm_mapsize = 0;
847 	map->dm_nsegs = 0;
848 	map->_dm_origbuf = NULL;
849 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
850 	map->_dm_vmspace = NULL;
851 }
852 
853 static void
854 _bus_dmamap_sync_segment(vaddr_t va, paddr_t pa, vsize_t len, int ops)
855 {
856 
857 	KASSERTMSG((va & PAGE_MASK) == (pa & PAGE_MASK),
858 	    "va %#" PRIxVADDR " pa %#" PRIxPADDR, va, pa);
859 #if 0
860 	printf("sync_segment: va=%#" PRIxVADDR
861 	    " pa=%#" PRIxPADDR " len=%#" PRIxVSIZE " ops=%#x\n",
862 	    va, pa, len, ops);
863 #endif
864 	switch (ops) {
865 	case BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE:
866 		STAT_INCR(sync_prereadwrite);
867 		cpu_dcache_wbinv_range(va, len);
868 		cpu_sdcache_wbinv_range(va, pa, len);
869 		break;
870 
871 	case BUS_DMASYNC_PREREAD: {
872 		const vsize_t line_size = riscv_dcache_align;
873 		const vsize_t line_mask = riscv_dcache_align_mask;
874 		vsize_t misalignment = va & line_mask;
875 		if (misalignment) {
876 			va -= misalignment;
877 			pa -= misalignment;
878 			len += misalignment;
879 			STAT_INCR(sync_preread_begin);
880 			cpu_dcache_wbinv_range(va, line_size);
881 			cpu_sdcache_wbinv_range(va, pa, line_size);
882 			if (len <= line_size)
883 				break;
884 			va += line_size;
885 			pa += line_size;
886 			len -= line_size;
887 		}
888 		misalignment = len & line_mask;
889 		len -= misalignment;
890 		if (len > 0) {
891 			STAT_INCR(sync_preread);
892 			cpu_dcache_inv_range(va, len);
893 			cpu_sdcache_inv_range(va, pa, len);
894 		}
895 		if (misalignment) {
896 			va += len;
897 			pa += len;
898 			STAT_INCR(sync_preread_tail);
899 			cpu_dcache_wbinv_range(va, line_size);
900 			cpu_sdcache_wbinv_range(va, pa, line_size);
901 		}
902 		break;
903 	}
904 
905 	case BUS_DMASYNC_PREWRITE:
906 		STAT_INCR(sync_prewrite);
907 		cpu_dcache_wb_range(va, len);
908 		cpu_sdcache_wb_range(va, pa, len);
909 		break;
910 
911 	/*
912 	 * CPUs can do speculative loads so we need to clean the cache after
913 	 * a DMA read to deal with any speculatively loaded cache lines.
914 	 * Since these can't be dirty, we can just invalidate them and don't
915 	 * have to worry about having to write back their contents.
916 	 */
917 	case BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE:
918 		STAT_INCR(sync_postreadwrite);
919 		cpu_dcache_inv_range(va, len);
920 		cpu_sdcache_inv_range(va, pa, len);
921 		break;
922 	case BUS_DMASYNC_POSTREAD:
923 		STAT_INCR(sync_postread);
924 		cpu_dcache_inv_range(va, len);
925 		cpu_sdcache_inv_range(va, pa, len);
926 		break;
927 	}
928 }
929 
930 static inline void
931 _bus_dmamap_sync_linear(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
932     bus_size_t len, int ops)
933 {
934 	bus_dma_segment_t *ds = map->dm_segs;
935 	vaddr_t va = (vaddr_t) map->_dm_origbuf;
936 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
937 	if (map->_dm_flags & _BUS_DMAMAP_IS_BOUNCING) {
938 		struct riscv_bus_dma_cookie * const cookie = map->_dm_cookie;
939 		va = (vaddr_t) cookie->id_bouncebuf;
940 	}
941 #endif
942 
943 	while (len > 0) {
944 		while (offset >= ds->ds_len) {
945 			offset -= ds->ds_len;
946 			va += ds->ds_len;
947 			ds++;
948 		}
949 
950 		paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + offset);
951 		size_t seglen = uimin(len, ds->ds_len - offset);
952 
953 		if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
954 			_bus_dmamap_sync_segment(va + offset, pa, seglen, ops);
955 
956 		offset += seglen;
957 		len -= seglen;
958 	}
959 }
960 
961 static inline void
962 _bus_dmamap_sync_mbuf(bus_dma_tag_t t, bus_dmamap_t map, bus_size_t offset,
963     bus_size_t len, int ops)
964 {
965 	bus_dma_segment_t *ds = map->dm_segs;
966 	struct mbuf *m = map->_dm_origbuf;
967 	bus_size_t voff = offset;
968 	bus_size_t ds_off = offset;
969 
970 	while (len > 0) {
971 		/* Find the current dma segment */
972 		while (ds_off >= ds->ds_len) {
973 			ds_off -= ds->ds_len;
974 			ds++;
975 		}
976 		/* Find the current mbuf. */
977 		while (voff >= m->m_len) {
978 			voff -= m->m_len;
979 			m = m->m_next;
980 		}
981 
982 		/*
983 		 * Now at the first mbuf to sync; nail each one until
984 		 * we have exhausted the length.
985 		 */
986 		vsize_t seglen = uimin(len, uimin(m->m_len - voff, ds->ds_len - ds_off));
987 		vaddr_t va = mtod(m, vaddr_t) + voff;
988 		paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
989 
990 		/*
991 		 * If a mapping is read-only, no dirty cache blocks will
992 		 * exist for it.  If a writable mapping was made read-only,
993 		 * we know any dirty cache lines for the range will have
994 		 * been cleaned for us already.  Therefore, if the upper
995 		 * layer can tell us we have a read-only mapping, we can
996 		 * skip all cache cleaning.
997 		 *
998 		 * NOTE: This only works if we know the pmap cleans pages
999 		 * before making a read-write -> read-only transition.  Assume
1000 		 * this is not true here.
1001 		 *
1002 		 * XXXNH this will have to be revisited.
1003 		 */
1004 
1005 		if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0) {
1006 			/*
1007 			 * If we are doing preread (DMAing into the mbuf),
1008 			 * this mbuf better not be readonly,
1009 			 */
1010 			KASSERT(!(ops & BUS_DMASYNC_PREREAD) || !M_ROMAP(m));
1011 			_bus_dmamap_sync_segment(va, pa, seglen, ops);
1012 		}
1013 		voff += seglen;
1014 		ds_off += seglen;
1015 		len -= seglen;
1016 	}
1017 }
1018 
1019 static inline void
1020 _bus_dmamap_sync_uio(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1021     bus_size_t len, int ops)
1022 {
1023 	bus_dma_segment_t *ds = map->dm_segs;
1024 	struct uio *uio = map->_dm_origbuf;
1025 	struct iovec *iov = uio->uio_iov;
1026 	bus_size_t voff = offset;
1027 	bus_size_t ds_off = offset;
1028 
1029 	while (len > 0) {
1030 		/* Find the current dma segment */
1031 		while (ds_off >= ds->ds_len) {
1032 			ds_off -= ds->ds_len;
1033 			ds++;
1034 		}
1035 
1036 		/* Find the current iovec. */
1037 		while (voff >= iov->iov_len) {
1038 			voff -= iov->iov_len;
1039 			iov++;
1040 		}
1041 
1042 		/*
1043 		 * Now at the first iovec to sync; nail each one until
1044 		 * we have exhausted the length.
1045 		 */
1046 		vsize_t seglen = uimin(len, uimin(iov->iov_len - voff, ds->ds_len - ds_off));
1047 		vaddr_t va = (vaddr_t) iov->iov_base + voff;
1048 		paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
1049 
1050 		if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
1051 			_bus_dmamap_sync_segment(va, pa, seglen, ops);
1052 
1053 		voff += seglen;
1054 		ds_off += seglen;
1055 		len -= seglen;
1056 	}
1057 }
1058 
1059 /*
1060  * Common function for DMA map synchronization.  May be called
1061  * by bus-specific DMA map synchronization functions.
1062  *
1063  * XXX Should have separate versions for write-through vs.
1064  * XXX write-back caches.  We currently assume write-back
1065  * XXX here, which is not as efficient as it could be for
1066  * XXX the write-through case.
1067  */
1068 void
1069 _bus_dmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
1070     bus_size_t len, int ops)
1071 {
1072 #ifdef DEBUG_DMA
1073 	printf("dmamap_sync: t=%p map=%p offset=%#" PRIxBUSADDR
1074 	    " len=%#" PRIxBUSSIZE " ops=%#x\n", t, map, offset, len, ops);
1075 #endif	/* DEBUG_DMA */
1076 
1077 	/*
1078 	 * Mixing of PRE and POST operations is not allowed.
1079 	 */
1080 	KASSERTMSG((((ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) == 0)
1081 	    || ((ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE)) == 0)),
1082 	    "%s: mix PRE and POST", __func__);
1083 
1084 	KASSERTMSG(offset < map->dm_mapsize,
1085 	    "offset %" PRIxBUSADDR " mapsize %" PRIuBUSSIZE,
1086 	    offset, map->dm_mapsize);
1087 	KASSERTMSG(len > 0 && offset + len <= map->dm_mapsize,
1088 	    "len %" PRIuBUSSIZE " offset %" PRIxBUSADDR " mapsize %" PRIuBUSSIZE,
1089 	    len, offset, map->dm_mapsize);
1090 
1091 	/*
1092 	 * For a write-back cache, we need to do the following things:
1093 	 *
1094 	 *	PREREAD -- Invalidate the D-cache.  We do this
1095 	 *	here in case a write-back is required by the back-end.
1096 	 *
1097 	 *	PREWRITE -- Write-back the D-cache.  Note that if
1098 	 *	we are doing a PREREAD | PREWRITE, we can collapse
1099 	 *	the whole thing into a single Wb-Inv.
1100 	 *
1101 	 *	POSTREAD -- Re-invalidate the D-cache in case speculative
1102 	 *	memory accesses caused cachelines to become valid with now
1103 	 *	invalid data.
1104 	 *
1105 	 *	POSTWRITE -- Nothing.
1106 	 */
1107 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
1108 	const bool bouncing = (map->_dm_flags & _BUS_DMAMAP_IS_BOUNCING);
1109 #else
1110 	const bool bouncing = false;
1111 #endif
1112 
1113 	const int pre_ops = ops & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1114 	const int post_ops = ops & (BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1115 	if (pre_ops == 0 && post_ops == 0)
1116 		return;
1117 
1118 	if (post_ops == BUS_DMASYNC_POSTWRITE) {
1119 		KASSERT(pre_ops == 0);
1120 		if ((map->_dm_flags & _BUS_DMAMAP_COHERENT)) {
1121 			STAT_INCR(sync_coherent_postwrite);
1122 		} else {
1123 			STAT_INCR(sync_postwrite);
1124 		}
1125 		return;
1126 	}
1127 
1128 	KASSERTMSG(bouncing || pre_ops != 0 || (post_ops & BUS_DMASYNC_POSTREAD),
1129 	    "pre_ops %#x post_ops %#x", pre_ops, post_ops);
1130 
1131 	if (bouncing && (ops & BUS_DMASYNC_PREWRITE)) {
1132 		struct riscv_bus_dma_cookie * const cookie = map->_dm_cookie;
1133 		STAT_INCR(write_bounces);
1134 		char * const dataptr = (char *)cookie->id_bouncebuf + offset;
1135 		/*
1136 		 * Copy the caller's buffer to the bounce buffer.
1137 		 */
1138 		switch (map->_dm_buftype) {
1139 		case _BUS_DMA_BUFTYPE_LINEAR:
1140 			memcpy(dataptr, cookie->id_origlinearbuf + offset, len);
1141 			break;
1142 
1143 		case _BUS_DMA_BUFTYPE_MBUF:
1144 			m_copydata(cookie->id_origmbuf, offset, len, dataptr);
1145 			break;
1146 
1147 		case _BUS_DMA_BUFTYPE_UIO:
1148 			_bus_dma_uiomove(dataptr, cookie->id_origuio, len,
1149 			    UIO_WRITE);
1150 			break;
1151 
1152 #ifdef DIAGNOSTIC
1153 		case _BUS_DMA_BUFTYPE_RAW:
1154 			panic("%s:(pre): _BUS_DMA_BUFTYPE_RAW", __func__);
1155 			break;
1156 
1157 		case _BUS_DMA_BUFTYPE_INVALID:
1158 			panic("%s(pre): _BUS_DMA_BUFTYPE_INVALID", __func__);
1159 			break;
1160 
1161 		default:
1162 			panic("%s(pre): map %p: unknown buffer type %d\n",
1163 			    __func__, map, map->_dm_buftype);
1164 			break;
1165 #endif /* DIAGNOSTIC */
1166 		}
1167 	}
1168 
1169 	/* Skip cache frobbing if mapping was COHERENT */
1170 	if ((map->_dm_flags & _BUS_DMAMAP_COHERENT)) {
1171 		switch (ops) {
1172 		case BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE:
1173 			STAT_INCR(sync_coherent_prereadwrite);
1174 			break;
1175 
1176 		case BUS_DMASYNC_PREREAD:
1177 			STAT_INCR(sync_coherent_preread);
1178 			break;
1179 
1180 		case BUS_DMASYNC_PREWRITE:
1181 			STAT_INCR(sync_coherent_prewrite);
1182 			break;
1183 
1184 		case BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE:
1185 			STAT_INCR(sync_coherent_postreadwrite);
1186 			break;
1187 
1188 		case BUS_DMASYNC_POSTREAD:
1189 			STAT_INCR(sync_coherent_postread);
1190 			break;
1191 
1192 		/* BUS_DMASYNC_POSTWRITE was aleady handled as a fastpath */
1193 		}
1194 		/*
1195 		 * Drain the write buffer of DMA operators.
1196 		 * 1) when cpu->device (prewrite)
1197 		 * 2) when device->cpu (postread)
1198 		 */
1199 		if ((pre_ops & BUS_DMASYNC_PREWRITE) || (post_ops & BUS_DMASYNC_POSTREAD))
1200 			asm volatile ("fence iorw,iorw" ::: "memory");
1201 
1202 		/*
1203 		 * Only thing left to do for COHERENT mapping is copy from bounce
1204 		 * in the POSTREAD case.
1205 		 */
1206 		if (bouncing && (post_ops & BUS_DMASYNC_POSTREAD))
1207 			goto bounce_it;
1208 
1209 		return;
1210 	}
1211 
1212 	int buftype = map->_dm_buftype;
1213 	if (bouncing) {
1214 		buftype = _BUS_DMA_BUFTYPE_LINEAR;
1215 	}
1216 
1217 	switch (buftype) {
1218 	case _BUS_DMA_BUFTYPE_LINEAR:
1219 	case _BUS_DMA_BUFTYPE_RAW:
1220 		_bus_dmamap_sync_linear(t, map, offset, len, ops);
1221 		break;
1222 
1223 	case _BUS_DMA_BUFTYPE_MBUF:
1224 		_bus_dmamap_sync_mbuf(t, map, offset, len, ops);
1225 		break;
1226 
1227 	case _BUS_DMA_BUFTYPE_UIO:
1228 		_bus_dmamap_sync_uio(t, map, offset, len, ops);
1229 		break;
1230 
1231 	case _BUS_DMA_BUFTYPE_INVALID:
1232 		panic("%s: _BUS_DMA_BUFTYPE_INVALID", __func__);
1233 		break;
1234 
1235 	default:
1236 		panic("%s: map %p: unknown buffer type %d\n", __func__, map,
1237 		    map->_dm_buftype);
1238 	}
1239 
1240 	/* Drain the write buffer. */
1241 	asm volatile ("fence iorw,iorw" ::: "memory");
1242 
1243 	if (!bouncing || (ops & BUS_DMASYNC_POSTREAD) == 0)
1244 		return;
1245 
1246   bounce_it:
1247 	STAT_INCR(read_bounces);
1248 
1249 	struct riscv_bus_dma_cookie * const cookie = map->_dm_cookie;
1250 	char * const dataptr = (char *)cookie->id_bouncebuf + offset;
1251 	/*
1252 	 * Copy the bounce buffer to the caller's buffer.
1253 	 */
1254 	switch (map->_dm_buftype) {
1255 	case _BUS_DMA_BUFTYPE_LINEAR:
1256 		memcpy(cookie->id_origlinearbuf + offset, dataptr, len);
1257 		break;
1258 
1259 	case _BUS_DMA_BUFTYPE_MBUF:
1260 		m_copyback(cookie->id_origmbuf, offset, len, dataptr);
1261 		break;
1262 
1263 	case _BUS_DMA_BUFTYPE_UIO:
1264 		_bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_READ);
1265 		break;
1266 
1267 #ifdef DIAGNOSTIC
1268 	case _BUS_DMA_BUFTYPE_RAW:
1269 		panic("%s(post): _BUS_DMA_BUFTYPE_RAW", __func__);
1270 		break;
1271 
1272 	case _BUS_DMA_BUFTYPE_INVALID:
1273 		panic("%s(post): _BUS_DMA_BUFTYPE_INVALID", __func__);
1274 		break;
1275 
1276 	default:
1277 		panic("%s(post): map %p: unknown buffer type %d\n", __func__,
1278 		    map, map->_dm_buftype);
1279 		break;
1280 #endif
1281 	}
1282 }
1283 
1284 /*
1285  * Common function for DMA-safe memory allocation.  May be called
1286  * by bus-specific DMA memory allocation functions.
1287  */
1288 
1289 int
1290 _bus_dmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1291     bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1292     int flags)
1293 {
1294 	struct riscv_dma_range *dr;
1295 	int error, i;
1296 
1297 #ifdef DEBUG_DMA
1298 	printf("dmamem_alloc t=%p size=%#" PRIxBUSSIZE
1299 	    " align=%#" PRIxBUSSIZE
1300 	    " boundary=%#" PRIxBUSSIZE " "
1301 	    "segs=%p nsegs=%#x rsegs=%p flags=%#x\n", t, size, alignment,
1302 	    boundary, segs, nsegs, rsegs, flags);
1303 #endif
1304 
1305 	if ((dr = t->_ranges) != NULL) {
1306 		error = ENOMEM;
1307 		for (i = 0; i < t->_nranges; i++, dr++) {
1308 			if (dr->dr_len == 0
1309 			    || (dr->dr_flags & _BUS_DMAMAP_NOALLOC))
1310 				continue;
1311 			error = _bus_dmamem_alloc_range(t, size, alignment,
1312 			    boundary, segs, nsegs, rsegs, flags,
1313 			    trunc_page(dr->dr_sysbase),
1314 			    trunc_page(dr->dr_sysbase + dr->dr_len));
1315 			if (error == 0)
1316 				break;
1317 		}
1318 	} else {
1319 		error = _bus_dmamem_alloc_range(t, size, alignment, boundary,
1320 		    segs, nsegs, rsegs, flags, 0UL, ~0UL);
1321 	}
1322 
1323 #ifdef DEBUG_DMA
1324 	printf("dmamem_alloc: =%d\n", error);
1325 #endif
1326 
1327 	return error;
1328 }
1329 
1330 /*
1331  * Common function for freeing DMA-safe memory.  May be called by
1332  * bus-specific DMA memory free functions.
1333  */
1334 void
1335 _bus_dmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
1336 {
1337 	struct vm_page *m;
1338 	bus_addr_t addr;
1339 	struct pglist mlist;
1340 	int curseg;
1341 
1342 #ifdef DEBUG_DMA
1343 	printf("dmamem_free: t=%p segs=%p nsegs=%#x\n", t, segs, nsegs);
1344 #endif	/* DEBUG_DMA */
1345 
1346 	/*
1347 	 * Build a list of pages to free back to the VM system.
1348 	 */
1349 	TAILQ_INIT(&mlist);
1350 	for (curseg = 0; curseg < nsegs; curseg++) {
1351 		for (addr = segs[curseg].ds_addr;
1352 		    addr < (segs[curseg].ds_addr + segs[curseg].ds_len);
1353 		    addr += PAGE_SIZE) {
1354 			m = PHYS_TO_VM_PAGE(addr);
1355 			TAILQ_INSERT_TAIL(&mlist, m, pageq.queue);
1356 		}
1357 	}
1358 	uvm_pglistfree(&mlist);
1359 }
1360 
1361 /*
1362  * Common function for mapping DMA-safe memory.  May be called by
1363  * bus-specific DMA memory map functions.
1364  */
1365 int
1366 _bus_dmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1367     size_t size, void **kvap, int flags)
1368 {
1369 	vaddr_t va;
1370 	paddr_t pa;
1371 	int curseg;
1372 	const uvm_flag_t kmflags = UVM_KMF_VAONLY
1373 	    | ((flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0);
1374 	vsize_t align = 0;
1375 
1376 #ifdef DEBUG_DMA
1377 	printf("dmamem_map: t=%p segs=%p nsegs=%#x size=%#zx flags=%#x\n", t,
1378 	    segs, nsegs, size, flags);
1379 #endif	/* DEBUG_DMA */
1380 
1381 #ifdef PMAP_MAP_POOLPAGE
1382 	/*
1383 	 * If all of memory is mapped, and we are mapping a single physically
1384 	 * contiguous area then this area is already mapped.  Let's see if we
1385 	 * avoid having a separate mapping for it.
1386 	 */
1387 	if (nsegs == 1 && (flags & BUS_DMA_PREFETCHABLE) == 0) {
1388 		/*
1389 		 * If this is a non-COHERENT mapping, then the existing kernel
1390 		 * mapping is already compatible with it.
1391 		 */
1392 		bool direct_mapable = (flags & BUS_DMA_COHERENT) == 0;
1393 		pa = segs[0].ds_addr;
1394 
1395 		/*
1396 		 * This is a COHERENT mapping which, unless this address is in
1397 		 * a COHERENT dma range, will not be compatible.
1398 		 */
1399 		if (t->_ranges != NULL) {
1400 			const struct riscv_dma_range * const dr =
1401 			    _bus_dma_paddr_inrange(t->_ranges, t->_nranges, pa);
1402 			if (dr != NULL
1403 			    && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
1404 				direct_mapable = true;
1405 			}
1406 		}
1407 
1408 		if (direct_mapable) {
1409 			*kvap = (void *)PMAP_MAP_POOLPAGE(pa);
1410 #ifdef DEBUG_DMA
1411 			printf("dmamem_map: =%p\n", *kvap);
1412 #endif	/* DEBUG_DMA */
1413 			return 0;
1414 		}
1415 	}
1416 #endif
1417 
1418 	size = round_page(size);
1419 
1420 #ifdef PMAP_MAPSIZE1
1421 	if (size >= PMAP_MAPSIZE1)
1422 		align = PMAP_MAPSIZE1;
1423 
1424 #ifdef PMAP_MAPSIZE2
1425 
1426 #if PMAP_MAPSIZE1 > PMAP_MAPSIZE2
1427 #error PMAP_MAPSIZE1 must be smaller than PMAP_MAPSIZE2
1428 #endif
1429 
1430 	if (size >= PMAP_MAPSIZE2)
1431 		align = PMAP_MAPSIZE2;
1432 
1433 #ifdef PMAP_MAPSIZE3
1434 
1435 #if PMAP_MAPSIZE2 > PMAP_MAPSIZE3
1436 #error PMAP_MAPSIZE2 must be smaller than PMAP_MAPSIZE3
1437 #endif
1438 
1439 	if (size >= PMAP_MAPSIZE3)
1440 		align = PMAP_MAPSIZE3;
1441 #endif
1442 #endif
1443 #endif
1444 
1445 	va = uvm_km_alloc(kernel_map, size, align, kmflags);
1446 	if (__predict_false(va == 0 && align > 0)) {
1447 		align = 0;
1448 		va = uvm_km_alloc(kernel_map, size, 0, kmflags);
1449 	}
1450 
1451 	if (va == 0)
1452 		return ENOMEM;
1453 
1454 	*kvap = (void *)va;
1455 
1456 	for (curseg = 0; curseg < nsegs; curseg++) {
1457 		for (pa = segs[curseg].ds_addr;
1458 		    pa < (segs[curseg].ds_addr + segs[curseg].ds_len);
1459 		    pa += PAGE_SIZE, va += PAGE_SIZE, size -= PAGE_SIZE) {
1460 			bool uncached = (flags & BUS_DMA_COHERENT);
1461 			bool prefetchable = (flags & BUS_DMA_PREFETCHABLE);
1462 #ifdef DEBUG_DMA
1463 			printf("wiring P%#" PRIxPADDR
1464 			    " to V%#" PRIxVADDR "\n", pa, va);
1465 #endif	/* DEBUG_DMA */
1466 			if (size == 0)
1467 				panic("_bus_dmamem_map: size botch");
1468 
1469 			const struct riscv_dma_range * const dr =
1470 			    _bus_dma_paddr_inrange(t->_ranges, t->_nranges, pa);
1471 			/*
1472 			 * If this dma region is coherent then there is
1473 			 * no need for an uncached mapping.
1474 			 */
1475 			if (dr != NULL
1476 			    && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
1477 				uncached = false;
1478 			}
1479 
1480 			u_int pmap_flags = PMAP_WIRED;
1481 			if (prefetchable)
1482 				pmap_flags |= PMAP_WRITE_COMBINE;
1483 			else if (uncached)
1484 				pmap_flags |= PMAP_NOCACHE;
1485 
1486 			pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE,
1487 			    pmap_flags);
1488 		}
1489 	}
1490 	pmap_update(pmap_kernel());
1491 #ifdef DEBUG_DMA
1492 	printf("dmamem_map: =%p\n", *kvap);
1493 #endif	/* DEBUG_DMA */
1494 	return 0;
1495 }
1496 
1497 /*
1498  * Common function for unmapping DMA-safe memory.  May be called by
1499  * bus-specific DMA memory unmapping functions.
1500  */
1501 void
1502 _bus_dmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
1503 {
1504 
1505 #ifdef DEBUG_DMA
1506 	printf("dmamem_unmap: t=%p kva=%p size=%#zx\n", t, kva, size);
1507 #endif	/* DEBUG_DMA */
1508 	KASSERTMSG(((uintptr_t)kva & PAGE_MASK) == 0,
1509 	    "kva %p (%#"PRIxPTR")", kva, ((uintptr_t)kva & PAGE_MASK));
1510 
1511 	/*
1512 	 * Check to see if this used direct mapped memory.  If so we can
1513 	 * just return since we have nothing to free up.
1514 	 */
1515 	if (pmap_md_direct_mapped_vaddr_p((vaddr_t)kva))
1516 		return;
1517 
1518 	size = round_page(size);
1519 	pmap_kremove((vaddr_t)kva, size);
1520 	pmap_update(pmap_kernel());
1521 	uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
1522 }
1523 
1524 /*
1525  * Common function for mmap(2)'ing DMA-safe memory.  May be called by
1526  * bus-specific DMA mmap(2)'ing functions.
1527  */
1528 paddr_t
1529 _bus_dmamem_mmap(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
1530     off_t off, int prot, int flags)
1531 {
1532 	/* Page not found. */
1533 	return -1;
1534 }
1535 
1536 /**********************************************************************
1537  * DMA utility functions
1538  **********************************************************************/
1539 
1540 /*
1541  * Utility function to load a linear buffer.  lastaddrp holds state
1542  * between invocations (for multiple-buffer loads).  segp contains
1543  * the starting segment on entrance, and the ending segment on exit.
1544  * first indicates if this is the first invocation of this function.
1545  */
1546 int
1547 _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
1548     bus_size_t buflen, struct vmspace *vm, int flags)
1549 {
1550 	bus_size_t sgsize;
1551 	bus_addr_t curaddr;
1552 	vaddr_t vaddr = (vaddr_t)buf;
1553 	int error;
1554 	pmap_t pmap = vm_map_pmap(&vm->vm_map);
1555 
1556 #ifdef DEBUG_DMA
1557 	printf("_bus_dmamap_load_buffer(buf=%p, len=%#" PRIxBUSSIZE
1558 	    ", flags=%#x)\n", buf, buflen, flags);
1559 #endif	/* DEBUG_DMA */
1560 
1561 
1562 	while (buflen > 0) {
1563 		/*
1564 		 * Get the physical address for this segment.
1565 		 */
1566 		pmap_extract(pmap, vaddr, &curaddr);
1567 
1568 		KASSERTMSG((vaddr & PAGE_MASK) == (curaddr & PAGE_MASK),
1569 		    "va %#" PRIxVADDR " curaddr %#" PRIxBUSADDR, vaddr, curaddr);
1570 
1571 		/*
1572 		 * Compute the segment size, and adjust counts.
1573 		 */
1574 		sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
1575 		if (buflen < sgsize)
1576 			sgsize = buflen;
1577 
1578 		error = _bus_dmamap_load_paddr(t, map, curaddr, sgsize,
1579 		    false);
1580 		if (__predict_false(error))
1581 			return error;
1582 
1583 		vaddr += sgsize;
1584 		buflen -= sgsize;
1585 	}
1586 
1587 	return 0;
1588 }
1589 
1590 /*
1591  * Allocate physical memory from the given physical address range.
1592  * Called by DMA-safe memory allocation methods.
1593  */
1594 int
1595 _bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
1596     bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
1597     int flags, paddr_t low, paddr_t high)
1598 {
1599 	paddr_t curaddr, lastaddr;
1600 	struct vm_page *m;
1601 	struct pglist mlist;
1602 	int curseg, error;
1603 
1604 	KASSERTMSG(boundary == 0 || (boundary & (boundary - 1)) == 0,
1605 	    "invalid boundary %#" PRIxBUSSIZE, boundary);
1606 
1607 #ifdef DEBUG_DMA
1608 	printf("alloc_range: t=%p size=%#" PRIxBUSSIZE
1609 	    " align=%#" PRIxBUSSIZE " boundary=%#" PRIxBUSSIZE
1610 	    " segs=%p nsegs=%#x rsegs=%p flags=%#x"
1611 	    " lo=%#" PRIxPADDR " hi=%#" PRIxPADDR "\n",
1612 	    t, size, alignment, boundary, segs, nsegs, rsegs, flags, low, high);
1613 #endif	/* DEBUG_DMA */
1614 
1615 	/* Always round the size. */
1616 	size = round_page(size);
1617 
1618 	/*
1619 	 * We accept boundaries < size, splitting in multiple segments
1620 	 * if needed. uvm_pglistalloc does not, so compute an appropriate
1621 	 * boundary: next power of 2 >= size
1622 	 */
1623 	bus_size_t uboundary = boundary;
1624 	if (uboundary <= PAGE_SIZE) {
1625 		uboundary = 0;
1626 	} else {
1627 		while (uboundary < size) {
1628 			uboundary <<= 1;
1629 		}
1630 	}
1631 
1632 	/*
1633 	 * Allocate pages from the VM system.
1634 	 */
1635 	error = uvm_pglistalloc(size, low, high, alignment, uboundary,
1636 	    &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
1637 	if (error)
1638 		return error;
1639 
1640 	/*
1641 	 * Compute the location, size, and number of segments actually
1642 	 * returned by the VM code.
1643 	 */
1644 	m = TAILQ_FIRST(&mlist);
1645 	curseg = 0;
1646 	lastaddr = segs[curseg].ds_addr = segs[curseg]._ds_paddr =
1647 	    VM_PAGE_TO_PHYS(m);
1648 	segs[curseg].ds_len = PAGE_SIZE;
1649 #ifdef DEBUG_DMA
1650 	printf("alloc: page %#" PRIxPADDR "\n", lastaddr);
1651 #endif	/* DEBUG_DMA */
1652 	m = TAILQ_NEXT(m, pageq.queue);
1653 
1654 	for (; m != NULL; m = TAILQ_NEXT(m, pageq.queue)) {
1655 		curaddr = VM_PAGE_TO_PHYS(m);
1656 		KASSERTMSG(low <= curaddr && curaddr < high,
1657 		    "uvm_pglistalloc returned non-sensicaladdress %#" PRIxPADDR
1658 		    "(low=%#" PRIxPADDR ", high=%#" PRIxPADDR "\n",
1659 		    curaddr, low, high);
1660 #ifdef DEBUG_DMA
1661 		printf("alloc: page %#" PRIxPADDR "\n", curaddr);
1662 #endif	/* DEBUG_DMA */
1663 		if (curaddr == lastaddr + PAGE_SIZE
1664 		    && (lastaddr & boundary) == (curaddr & boundary))
1665 			segs[curseg].ds_len += PAGE_SIZE;
1666 		else {
1667 			curseg++;
1668 			if (curseg >= nsegs) {
1669 				uvm_pglistfree(&mlist);
1670 				return EFBIG;
1671 			}
1672 			segs[curseg].ds_addr = curaddr;
1673 			segs[curseg]._ds_paddr = curaddr;
1674 			segs[curseg].ds_len = PAGE_SIZE;
1675 		}
1676 		lastaddr = curaddr;
1677 	}
1678 
1679 	*rsegs = curseg + 1;
1680 
1681 	return 0;
1682 }
1683 
1684 /*
1685  * Check if a memory region intersects with a DMA range, and return the
1686  * page-rounded intersection if it does.
1687  */
1688 int
1689 riscv_dma_range_intersect(struct riscv_dma_range *ranges, int nranges,
1690     paddr_t pa, psize_t size, paddr_t *pap, psize_t *sizep)
1691 {
1692 	struct riscv_dma_range *dr;
1693 	int i;
1694 
1695 	if (ranges == NULL)
1696 		return 0;
1697 
1698 	for (i = 0, dr = ranges; i < nranges; i++, dr++) {
1699 		if (dr->dr_sysbase <= pa &&
1700 		    pa < (dr->dr_sysbase + dr->dr_len)) {
1701 			/*
1702 			 * Beginning of region intersects with this range.
1703 			 */
1704 			*pap = trunc_page(pa);
1705 			*sizep = round_page(uimin(pa + size,
1706 			    dr->dr_sysbase + dr->dr_len) - pa);
1707 			return 1;
1708 		}
1709 		if (pa < dr->dr_sysbase && dr->dr_sysbase < (pa + size)) {
1710 			/*
1711 			 * End of region intersects with this range.
1712 			 */
1713 			*pap = trunc_page(dr->dr_sysbase);
1714 			*sizep = round_page(uimin((pa + size) - dr->dr_sysbase,
1715 			    dr->dr_len));
1716 			return 1;
1717 		}
1718 	}
1719 
1720 	/* No intersection found. */
1721 	return 0;
1722 }
1723 
1724 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
1725 static int
1726 _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
1727     bus_size_t size, int flags)
1728 {
1729 	struct riscv_bus_dma_cookie *cookie = map->_dm_cookie;
1730 	int error = 0;
1731 
1732 	KASSERT(cookie != NULL);
1733 
1734 	cookie->id_bouncebuflen = round_page(size);
1735 	error = _bus_dmamem_alloc(t, cookie->id_bouncebuflen,
1736 	    PAGE_SIZE, map->_dm_boundary, cookie->id_bouncesegs,
1737 	    map->_dm_segcnt, &cookie->id_nbouncesegs, flags);
1738 	if (error == 0) {
1739 		error = _bus_dmamem_map(t, cookie->id_bouncesegs,
1740 		    cookie->id_nbouncesegs, cookie->id_bouncebuflen,
1741 		    (void **)&cookie->id_bouncebuf, flags);
1742 		if (error) {
1743 			_bus_dmamem_free(t, cookie->id_bouncesegs,
1744 			    cookie->id_nbouncesegs);
1745 			cookie->id_bouncebuflen = 0;
1746 			cookie->id_nbouncesegs = 0;
1747 		} else {
1748 			cookie->id_flags |= _BUS_DMA_HAS_BOUNCE;
1749 		}
1750 	} else {
1751 		cookie->id_bouncebuflen = 0;
1752 		cookie->id_nbouncesegs = 0;
1753 	}
1754 
1755 	return error;
1756 }
1757 
1758 static void
1759 _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map)
1760 {
1761 	struct riscv_bus_dma_cookie *cookie = map->_dm_cookie;
1762 
1763 	KASSERT(cookie != NULL);
1764 
1765 	_bus_dmamem_unmap(t, cookie->id_bouncebuf, cookie->id_bouncebuflen);
1766 	_bus_dmamem_free(t, cookie->id_bouncesegs, cookie->id_nbouncesegs);
1767 	cookie->id_bouncebuflen = 0;
1768 	cookie->id_nbouncesegs = 0;
1769 	cookie->id_flags &= ~_BUS_DMA_HAS_BOUNCE;
1770 }
1771 #endif /* _RISCV_NEED_BUS_DMA_BOUNCE */
1772 
1773 /*
1774  * This function does the same as uiomove, but takes an explicit
1775  * direction, and does not update the uio structure.
1776  */
1777 static int
1778 _bus_dma_uiomove(void *buf, struct uio *uio, size_t n, int direction)
1779 {
1780 	struct iovec *iov;
1781 	int error;
1782 	struct vmspace *vm;
1783 	char *cp;
1784 	size_t resid, cnt;
1785 	int i;
1786 
1787 	iov = uio->uio_iov;
1788 	vm = uio->uio_vmspace;
1789 	cp = buf;
1790 	resid = n;
1791 
1792 	for (i = 0; i < uio->uio_iovcnt && resid > 0; i++) {
1793 		iov = &uio->uio_iov[i];
1794 		if (iov->iov_len == 0)
1795 			continue;
1796 		cnt = MIN(resid, iov->iov_len);
1797 
1798 		if (!VMSPACE_IS_KERNEL_P(vm)) {
1799 			preempt_point();
1800 		}
1801 		if (direction == UIO_READ) {
1802 			error = copyout_vmspace(vm, cp, iov->iov_base, cnt);
1803 		} else {
1804 			error = copyin_vmspace(vm, iov->iov_base, cp, cnt);
1805 		}
1806 		if (error)
1807 			return error;
1808 		cp += cnt;
1809 		resid -= cnt;
1810 	}
1811 	return 0;
1812 }
1813 
1814 int
1815 _bus_dmatag_subregion(bus_dma_tag_t tag, bus_addr_t min_addr,
1816     bus_addr_t max_addr, bus_dma_tag_t *newtag, int flags)
1817 {
1818 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
1819 	if (min_addr >= max_addr)
1820 		return EOPNOTSUPP;
1821 
1822 	struct riscv_dma_range *dr;
1823 	bool psubset = true;
1824 	size_t nranges = 0;
1825 	size_t i;
1826 	for (i = 0, dr = tag->_ranges; i < tag->_nranges; i++, dr++) {
1827 		/*
1828 		 * If the new {min,max}_addr are narrower than any of the
1829 		 * ranges in the parent tag then we need a new tag;
1830 		 * otherwise the parent tag is a subset of the new
1831 		 * range and can continue to be used.
1832 		 */
1833 		if (min_addr > dr->dr_sysbase
1834 		    || max_addr < dr->dr_sysbase + dr->dr_len - 1) {
1835 			psubset = false;
1836 		}
1837 		if (min_addr <= dr->dr_sysbase + dr->dr_len
1838 		    && max_addr >= dr->dr_sysbase) {
1839 			nranges++;
1840 		}
1841 	}
1842 	if (nranges == 0) {
1843 		nranges = 1;
1844 		psubset = false;
1845 	}
1846 	if (psubset) {
1847 		*newtag = tag;
1848 		/* if the tag must be freed, add a reference */
1849 		if (tag->_tag_needs_free)
1850 			(tag->_tag_needs_free)++;
1851 		return 0;
1852 	}
1853 
1854 	const size_t tagsize = sizeof(*tag) + nranges * sizeof(*dr);
1855 	if ((*newtag = kmem_intr_zalloc(tagsize,
1856 	    (flags & BUS_DMA_NOWAIT) ? KM_NOSLEEP : KM_SLEEP)) == NULL)
1857 		return ENOMEM;
1858 
1859 	dr = (void *)(*newtag + 1);
1860 	**newtag = *tag;
1861 	(*newtag)->_tag_needs_free = 1;
1862 	(*newtag)->_ranges = dr;
1863 	(*newtag)->_nranges = nranges;
1864 
1865 	if (tag->_ranges == NULL) {
1866 		dr->dr_sysbase = min_addr;
1867 		dr->dr_busbase = min_addr;
1868 		dr->dr_len = max_addr + 1 - min_addr;
1869 	} else {
1870 		struct riscv_dma_range *pdr;
1871 
1872 		for (i = 0, pdr = tag->_ranges; i < tag->_nranges; i++, pdr++) {
1873 			KASSERT(nranges != 0);
1874 
1875 			if (min_addr > pdr->dr_sysbase + pdr->dr_len
1876 			    || max_addr < pdr->dr_sysbase) {
1877 				/*
1878 				 * this range doesn't overlap with new limits,
1879 				 * so skip.
1880 				 */
1881 				continue;
1882 			}
1883 			/*
1884 			 * Copy the range and adjust to fit within the new
1885 			 * limits
1886 			 */
1887 			dr[0] = pdr[0];
1888 			if (dr->dr_sysbase < min_addr) {
1889 				psize_t diff = min_addr - dr->dr_sysbase;
1890 				dr->dr_busbase += diff;
1891 				dr->dr_len -= diff;
1892 				dr->dr_sysbase += diff;
1893 			}
1894 			if (max_addr <= dr->dr_sysbase + dr->dr_len - 1) {
1895 				dr->dr_len = max_addr + 1 - dr->dr_sysbase;
1896 			}
1897 			dr++;
1898 			nranges--;
1899 		}
1900 	}
1901 
1902 	return 0;
1903 #else
1904 	return EOPNOTSUPP;
1905 #endif /* _RISCV_NEED_BUS_DMA_BOUNCE */
1906 }
1907 
1908 void
1909 _bus_dmatag_destroy(bus_dma_tag_t tag)
1910 {
1911 #ifdef _RISCV_NEED_BUS_DMA_BOUNCE
1912 	switch (tag->_tag_needs_free) {
1913 	case 0:
1914 		break;				/* not allocated with kmem */
1915 	case 1: {
1916 		const size_t tagsize = sizeof(*tag)
1917 		    + tag->_nranges * sizeof(*tag->_ranges);
1918 		kmem_intr_free(tag, tagsize);	/* last reference to tag */
1919 		break;
1920 	}
1921 	default:
1922 		(tag->_tag_needs_free)--;	/* one less reference */
1923 	}
1924 #endif
1925 }
1926