Home | History | Annotate | Line # | Download | only in arm32
bus_dma.c revision 1.78
      1 /*	$NetBSD: bus_dma.c,v 1.78 2013/02/15 01:03:43 matt Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1996, 1997, 1998 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
      9  * NASA Ames Research Center.
     10  *
     11  * Redistribution and use in source and binary forms, with or without
     12  * modification, are permitted provided that the following conditions
     13  * are met:
     14  * 1. Redistributions of source code must retain the above copyright
     15  *    notice, this list of conditions and the following disclaimer.
     16  * 2. Redistributions in binary form must reproduce the above copyright
     17  *    notice, this list of conditions and the following disclaimer in the
     18  *    documentation and/or other materials provided with the distribution.
     19  *
     20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     30  * POSSIBILITY OF SUCH DAMAGE.
     31  */
     32 
     33 #define _ARM32_BUS_DMA_PRIVATE
     34 
     35 #include <sys/cdefs.h>
     36 __KERNEL_RCSID(0, "$NetBSD: bus_dma.c,v 1.78 2013/02/15 01:03:43 matt Exp $");
     37 
     38 #include <sys/param.h>
     39 #include <sys/systm.h>
     40 #include <sys/kernel.h>
     41 #include <sys/proc.h>
     42 #include <sys/buf.h>
     43 #include <sys/reboot.h>
     44 #include <sys/conf.h>
     45 #include <sys/file.h>
     46 #include <sys/malloc.h>
     47 #include <sys/mbuf.h>
     48 #include <sys/vnode.h>
     49 #include <sys/device.h>
     50 
     51 #include <uvm/uvm.h>
     52 
     53 #include <sys/bus.h>
     54 #include <machine/cpu.h>
     55 
     56 #include <arm/cpufunc.h>
     57 
     58 #ifdef BUSDMA_COUNTERS
     59 static struct evcnt bus_dma_creates =
     60 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "creates");
     61 static struct evcnt bus_dma_bounced_creates =
     62 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced creates");
     63 static struct evcnt bus_dma_loads =
     64 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "loads");
     65 static struct evcnt bus_dma_bounced_loads =
     66 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "bounced loads");
     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");
     79 static struct evcnt bus_dma_sync_prereadwrite =
     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");
     87 static struct evcnt bus_dma_sync_prewrite =
     88 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync prewrite");
     89 static struct evcnt bus_dma_sync_postread =
     90 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postread");
     91 static struct evcnt bus_dma_sync_postreadwrite =
     92 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postreadwrite");
     93 static struct evcnt bus_dma_sync_postwrite =
     94 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "busdma", "sync postwrite");
     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);
    100 EVCNT_ATTACH_STATIC(bus_dma_read_bounces);
    101 EVCNT_ATTACH_STATIC(bus_dma_write_bounces);
    102 EVCNT_ATTACH_STATIC(bus_dma_unloads);
    103 EVCNT_ATTACH_STATIC(bus_dma_bounced_unloads);
    104 EVCNT_ATTACH_STATIC(bus_dma_destroys);
    105 EVCNT_ATTACH_STATIC(bus_dma_bounced_destroys);
    106 EVCNT_ATTACH_STATIC(bus_dma_sync_prereadwrite);
    107 EVCNT_ATTACH_STATIC(bus_dma_sync_preread_begin);
    108 EVCNT_ATTACH_STATIC(bus_dma_sync_preread);
    109 EVCNT_ATTACH_STATIC(bus_dma_sync_preread_tail);
    110 EVCNT_ATTACH_STATIC(bus_dma_sync_prewrite);
    111 EVCNT_ATTACH_STATIC(bus_dma_sync_postread);
    112 EVCNT_ATTACH_STATIC(bus_dma_sync_postreadwrite);
    113 EVCNT_ATTACH_STATIC(bus_dma_sync_postwrite);
    114 
    115 #define	STAT_INCR(x)	(bus_dma_ ## x.ev_count++)
    116 #else
    117 #define	STAT_INCR(x)	/*(bus_dma_ ## x.ev_count++)*/
    118 #endif
    119 
    120 int	_bus_dmamap_load_buffer(bus_dma_tag_t, bus_dmamap_t, void *,
    121 	    bus_size_t, struct vmspace *, int);
    122 static struct arm32_dma_range *
    123 	_bus_dma_paddr_inrange(struct arm32_dma_range *, int, paddr_t);
    124 
    125 /*
    126  * Check to see if the specified page is in an allowed DMA range.
    127  */
    128 inline struct arm32_dma_range *
    129 _bus_dma_paddr_inrange(struct arm32_dma_range *ranges, int nranges,
    130     bus_addr_t curaddr)
    131 {
    132 	struct arm32_dma_range *dr;
    133 	int i;
    134 
    135 	for (i = 0, dr = ranges; i < nranges; i++, dr++) {
    136 		if (curaddr >= dr->dr_sysbase &&
    137 		    round_page(curaddr) <= (dr->dr_sysbase + dr->dr_len))
    138 			return (dr);
    139 	}
    140 
    141 	return (NULL);
    142 }
    143 
    144 /*
    145  * Check to see if the specified busaddr is in an allowed DMA range.
    146  */
    147 static inline paddr_t
    148 _bus_dma_busaddr_to_paddr(bus_dma_tag_t t, bus_addr_t curaddr)
    149 {
    150 	struct arm32_dma_range *dr;
    151 	u_int i;
    152 
    153 	if (t->_nranges == 0)
    154 		return curaddr;
    155 
    156 	for (i = 0, dr = t->_ranges; i < t->_nranges; i++, dr++) {
    157 		if (dr->dr_busbase <= curaddr
    158 		    && round_page(curaddr) <= dr->dr_busbase + dr->dr_len)
    159 			return curaddr - dr->dr_busbase + dr->dr_sysbase;
    160 	}
    161 	panic("%s: curaddr %#lx not in range", __func__, curaddr);
    162 }
    163 
    164 /*
    165  * Common function to load the specified physical address into the
    166  * DMA map, coalescing segments and boundary checking as necessary.
    167  */
    168 static int
    169 _bus_dmamap_load_paddr(bus_dma_tag_t t, bus_dmamap_t map,
    170     bus_addr_t paddr, bus_size_t size, bool coherent)
    171 {
    172 	bus_dma_segment_t * const segs = map->dm_segs;
    173 	int nseg = map->dm_nsegs;
    174 	bus_addr_t lastaddr;
    175 	bus_addr_t bmask = ~(map->_dm_boundary - 1);
    176 	bus_addr_t curaddr;
    177 	bus_size_t sgsize;
    178 	uint32_t _ds_flags = coherent ? _BUS_DMAMAP_COHERENT : 0;
    179 
    180 	if (nseg > 0)
    181 		lastaddr = segs[nseg-1].ds_addr + segs[nseg-1].ds_len;
    182 	else
    183 		lastaddr = 0xdead;
    184 
    185  again:
    186 	sgsize = size;
    187 
    188 	/* Make sure we're in an allowed DMA range. */
    189 	if (t->_ranges != NULL) {
    190 		/* XXX cache last result? */
    191 		const struct arm32_dma_range * const dr =
    192 		    _bus_dma_paddr_inrange(t->_ranges, t->_nranges, paddr);
    193 		if (dr == NULL)
    194 			return (EINVAL);
    195 
    196 		/*
    197 		 * If this region is coherent, mark the segment as coherent.
    198 		 */
    199 		_ds_flags |= dr->dr_flags & _BUS_DMAMAP_COHERENT;
    200 
    201 		/*
    202 		 * In a valid DMA range.  Translate the physical
    203 		 * memory address to an address in the DMA window.
    204 		 */
    205 		curaddr = (paddr - dr->dr_sysbase) + dr->dr_busbase;
    206 #if 0
    207 		printf("%p: %#lx: range %#lx/%#lx/%#lx/%#x: %#x <-- %#lx\n",
    208 		    t, paddr, dr->dr_sysbase, dr->dr_busbase,
    209 		    dr->dr_len, dr->dr_flags, _ds_flags, curaddr);
    210 #endif
    211 	} else
    212 		curaddr = paddr;
    213 
    214 	/*
    215 	 * Make sure we don't cross any boundaries.
    216 	 */
    217 	if (map->_dm_boundary > 0) {
    218 		bus_addr_t baddr;	/* next boundary address */
    219 
    220 		baddr = (curaddr + map->_dm_boundary) & bmask;
    221 		if (sgsize > (baddr - curaddr))
    222 			sgsize = (baddr - curaddr);
    223 	}
    224 
    225 	/*
    226 	 * Insert chunk into a segment, coalescing with the
    227 	 * previous segment if possible.
    228 	 */
    229 	if (nseg > 0 && curaddr == lastaddr &&
    230 	    segs[nseg-1].ds_len + sgsize <= map->dm_maxsegsz &&
    231 	    ((segs[nseg-1]._ds_flags ^ _ds_flags) & _BUS_DMAMAP_COHERENT) == 0 &&
    232 	    (map->_dm_boundary == 0 ||
    233 	     (segs[nseg-1].ds_addr & bmask) == (curaddr & bmask))) {
    234 	     	/* coalesce */
    235 		segs[nseg-1].ds_len += sgsize;
    236 	} else if (nseg >= map->_dm_segcnt) {
    237 		return (EFBIG);
    238 	} else {
    239 		/* new segment */
    240 		segs[nseg].ds_addr = curaddr;
    241 		segs[nseg].ds_len = sgsize;
    242 		segs[nseg]._ds_flags = _ds_flags;
    243 		nseg++;
    244 	}
    245 
    246 	lastaddr = curaddr + sgsize;
    247 
    248 	paddr += sgsize;
    249 	size -= sgsize;
    250 	if (size > 0)
    251 		goto again;
    252 
    253 	map->_dm_flags &= (_ds_flags & _BUS_DMAMAP_COHERENT);
    254 	map->dm_nsegs = nseg;
    255 	return (0);
    256 }
    257 
    258 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    259 static int _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
    260 	    bus_size_t size, int flags);
    261 static void _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map);
    262 static int _bus_dma_uiomove(void *buf, struct uio *uio, size_t n,
    263 	    int direction);
    264 
    265 static int
    266 _bus_dma_load_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
    267 	size_t buflen, int buftype, int flags)
    268 {
    269 	struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
    270 	struct vmspace * const vm = vmspace_kernel();
    271 	int error;
    272 
    273 	KASSERT(cookie != NULL);
    274 	KASSERT(cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE);
    275 
    276 	/*
    277 	 * Allocate bounce pages, if necessary.
    278 	 */
    279 	if ((cookie->id_flags & _BUS_DMA_HAS_BOUNCE) == 0) {
    280 		error = _bus_dma_alloc_bouncebuf(t, map, buflen, flags);
    281 		if (error)
    282 			return (error);
    283 	}
    284 
    285 	/*
    286 	 * Cache a pointer to the caller's buffer and load the DMA map
    287 	 * with the bounce buffer.
    288 	 */
    289 	cookie->id_origbuf = buf;
    290 	cookie->id_origbuflen = buflen;
    291 	error = _bus_dmamap_load_buffer(t, map, cookie->id_bouncebuf,
    292 	    buflen, vm, flags);
    293 	if (error)
    294 		return (error);
    295 
    296 	STAT_INCR(bounced_loads);
    297 	map->dm_mapsize = buflen;
    298 	map->_dm_vmspace = vm;
    299 	map->_dm_buftype = buftype;
    300 
    301 	/* ...so _bus_dmamap_sync() knows we're bouncing */
    302 	map->_dm_flags |= _BUS_DMAMAP_IS_BOUNCING;
    303 	cookie->id_flags |= _BUS_DMA_IS_BOUNCING;
    304 	return 0;
    305 }
    306 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
    307 
    308 /*
    309  * Common function for DMA map creation.  May be called by bus-specific
    310  * DMA map creation functions.
    311  */
    312 int
    313 _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments,
    314     bus_size_t maxsegsz, bus_size_t boundary, int flags, bus_dmamap_t *dmamp)
    315 {
    316 	struct arm32_bus_dmamap *map;
    317 	void *mapstore;
    318 	size_t mapsize;
    319 
    320 #ifdef DEBUG_DMA
    321 	printf("dmamap_create: t=%p size=%lx nseg=%x msegsz=%lx boundary=%lx flags=%x\n",
    322 	    t, size, nsegments, maxsegsz, boundary, flags);
    323 #endif	/* DEBUG_DMA */
    324 
    325 	/*
    326 	 * Allocate and initialize the DMA map.  The end of the map
    327 	 * is a variable-sized array of segments, so we allocate enough
    328 	 * room for them in one shot.
    329 	 *
    330 	 * Note we don't preserve the WAITOK or NOWAIT flags.  Preservation
    331 	 * of ALLOCNOW notifies others that we've reserved these resources,
    332 	 * and they are not to be freed.
    333 	 *
    334 	 * The bus_dmamap_t includes one bus_dma_segment_t, hence
    335 	 * the (nsegments - 1).
    336 	 */
    337 	mapsize = sizeof(struct arm32_bus_dmamap) +
    338 	    (sizeof(bus_dma_segment_t) * (nsegments - 1));
    339 	const int mallocflags = M_ZERO|(flags & BUS_DMA_NOWAIT) ? M_NOWAIT : M_WAITOK;
    340 	if ((mapstore = malloc(mapsize, M_DMAMAP, mallocflags)) == NULL)
    341 		return (ENOMEM);
    342 
    343 	map = (struct arm32_bus_dmamap *)mapstore;
    344 	map->_dm_size = size;
    345 	map->_dm_segcnt = nsegments;
    346 	map->_dm_maxmaxsegsz = maxsegsz;
    347 	map->_dm_boundary = boundary;
    348 	map->_dm_flags = flags & ~(BUS_DMA_WAITOK|BUS_DMA_NOWAIT);
    349 	map->_dm_origbuf = NULL;
    350 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
    351 	map->_dm_vmspace = vmspace_kernel();
    352 	map->_dm_cookie = NULL;
    353 	map->dm_maxsegsz = maxsegsz;
    354 	map->dm_mapsize = 0;		/* no valid mappings */
    355 	map->dm_nsegs = 0;
    356 
    357 	*dmamp = map;
    358 
    359 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    360 	struct arm32_bus_dma_cookie *cookie;
    361 	int cookieflags;
    362 	void *cookiestore;
    363 	size_t cookiesize;
    364 	int error;
    365 
    366 	cookieflags = 0;
    367 
    368 	if (t->_may_bounce != NULL) {
    369 		error = (*t->_may_bounce)(t, map, flags, &cookieflags);
    370 		if (error != 0)
    371 			goto out;
    372 	}
    373 
    374 	if (t->_ranges != NULL)
    375 		cookieflags |= _BUS_DMA_MIGHT_NEED_BOUNCE;
    376 
    377 	if ((cookieflags & _BUS_DMA_MIGHT_NEED_BOUNCE) == 0) {
    378 		STAT_INCR(creates);
    379 		return 0;
    380 	}
    381 
    382 	cookiesize = sizeof(struct arm32_bus_dma_cookie) +
    383 	    (sizeof(bus_dma_segment_t) * map->_dm_segcnt);
    384 
    385 	/*
    386 	 * Allocate our cookie.
    387 	 */
    388 	if ((cookiestore = malloc(cookiesize, M_DMAMAP, mallocflags)) == NULL) {
    389 		error = ENOMEM;
    390 		goto out;
    391 	}
    392 	cookie = (struct arm32_bus_dma_cookie *)cookiestore;
    393 	cookie->id_flags = cookieflags;
    394 	map->_dm_cookie = cookie;
    395 	STAT_INCR(bounced_creates);
    396 
    397 	error = _bus_dma_alloc_bouncebuf(t, map, size, flags);
    398  out:
    399 	if (error)
    400 		_bus_dmamap_destroy(t, map);
    401 #else
    402 	STAT_INCR(creates);
    403 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
    404 
    405 #ifdef DEBUG_DMA
    406 	printf("dmamap_create:map=%p\n", map);
    407 #endif	/* DEBUG_DMA */
    408 	return (0);
    409 }
    410 
    411 /*
    412  * Common function for DMA map destruction.  May be called by bus-specific
    413  * DMA map destruction functions.
    414  */
    415 void
    416 _bus_dmamap_destroy(bus_dma_tag_t t, bus_dmamap_t map)
    417 {
    418 
    419 #ifdef DEBUG_DMA
    420 	printf("dmamap_destroy: t=%p map=%p\n", t, map);
    421 #endif	/* DEBUG_DMA */
    422 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    423 	struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
    424 
    425 	/*
    426 	 * Free any bounce pages this map might hold.
    427 	 */
    428 	if (cookie != NULL) {
    429 		if (cookie->id_flags & _BUS_DMA_IS_BOUNCING)
    430 			STAT_INCR(bounced_unloads);
    431 		map->dm_nsegs = 0;
    432 		if (cookie->id_flags & _BUS_DMA_HAS_BOUNCE)
    433 			_bus_dma_free_bouncebuf(t, map);
    434 		STAT_INCR(bounced_destroys);
    435 		free(cookie, M_DMAMAP);
    436 	} else
    437 #endif
    438 	STAT_INCR(destroys);
    439 
    440 	if (map->dm_nsegs > 0)
    441 		STAT_INCR(unloads);
    442 
    443 	free(map, M_DMAMAP);
    444 }
    445 
    446 /*
    447  * Common function for loading a DMA map with a linear buffer.  May
    448  * be called by bus-specific DMA map load functions.
    449  */
    450 int
    451 _bus_dmamap_load(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
    452     bus_size_t buflen, struct proc *p, int flags)
    453 {
    454 	struct vmspace *vm;
    455 	int error;
    456 
    457 #ifdef DEBUG_DMA
    458 	printf("dmamap_load: t=%p map=%p buf=%p len=%lx p=%p f=%d\n",
    459 	    t, map, buf, buflen, p, flags);
    460 #endif	/* DEBUG_DMA */
    461 
    462 	if (map->dm_nsegs > 0) {
    463 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    464 		struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
    465 		if (cookie != NULL) {
    466 			if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
    467 				STAT_INCR(bounced_unloads);
    468 				cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
    469 				map->_dm_flags &= ~_BUS_DMAMAP_IS_BOUNCING;
    470 			}
    471 		} else
    472 #endif
    473 		STAT_INCR(unloads);
    474 	}
    475 
    476 	/*
    477 	 * Make sure that on error condition we return "no valid mappings".
    478 	 */
    479 	map->dm_mapsize = 0;
    480 	map->dm_nsegs = 0;
    481 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
    482 	KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
    483 	    "dm_maxsegsz %lu _dm_maxmaxsegsz %lu",
    484 	    map->dm_maxsegsz, map->_dm_maxmaxsegsz);
    485 
    486 	if (buflen > map->_dm_size)
    487 		return (EINVAL);
    488 
    489 	if (p != NULL) {
    490 		vm = p->p_vmspace;
    491 	} else {
    492 		vm = vmspace_kernel();
    493 	}
    494 
    495 	/* _bus_dmamap_load_buffer() clears this if we're not... */
    496 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
    497 
    498 	error = _bus_dmamap_load_buffer(t, map, buf, buflen, vm, flags);
    499 	if (error == 0) {
    500 		map->dm_mapsize = buflen;
    501 		map->_dm_vmspace = vm;
    502 		map->_dm_origbuf = buf;
    503 		map->_dm_buftype = _BUS_DMA_BUFTYPE_LINEAR;
    504 		return 0;
    505 	}
    506 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    507 	struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
    508 	if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
    509 		error = _bus_dma_load_bouncebuf(t, map, buf, buflen,
    510 		    _BUS_DMA_BUFTYPE_LINEAR, flags);
    511 	}
    512 #endif
    513 	return (error);
    514 }
    515 
    516 /*
    517  * Like _bus_dmamap_load(), but for mbufs.
    518  */
    519 int
    520 _bus_dmamap_load_mbuf(bus_dma_tag_t t, bus_dmamap_t map, struct mbuf *m0,
    521     int flags)
    522 {
    523 	int error;
    524 	struct mbuf *m;
    525 
    526 #ifdef DEBUG_DMA
    527 	printf("dmamap_load_mbuf: t=%p map=%p m0=%p f=%d\n",
    528 	    t, map, m0, flags);
    529 #endif	/* DEBUG_DMA */
    530 
    531 	if (map->dm_nsegs > 0) {
    532 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    533 		struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
    534 		if (cookie != NULL) {
    535 			if (cookie->id_flags & _BUS_DMA_IS_BOUNCING) {
    536 				STAT_INCR(bounced_unloads);
    537 				cookie->id_flags &= ~_BUS_DMA_IS_BOUNCING;
    538 				map->_dm_flags &= ~_BUS_DMAMAP_IS_BOUNCING;
    539 			}
    540 		} else
    541 #endif
    542 		STAT_INCR(unloads);
    543 	}
    544 
    545 	/*
    546 	 * Make sure that on error condition we return "no valid mappings."
    547 	 */
    548 	map->dm_mapsize = 0;
    549 	map->dm_nsegs = 0;
    550 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
    551 	KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
    552 	    "dm_maxsegsz %lu _dm_maxmaxsegsz %lu",
    553 	    map->dm_maxsegsz, map->_dm_maxmaxsegsz);
    554 
    555 #ifdef DIAGNOSTIC
    556 	if ((m0->m_flags & M_PKTHDR) == 0)
    557 		panic("_bus_dmamap_load_mbuf: no packet header");
    558 #endif	/* DIAGNOSTIC */
    559 
    560 	if (m0->m_pkthdr.len > map->_dm_size)
    561 		return (EINVAL);
    562 
    563 	/* _bus_dmamap_load_paddr() clears this if we're not... */
    564 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
    565 
    566 	error = 0;
    567 	for (m = m0; m != NULL && error == 0; m = m->m_next) {
    568 		int offset;
    569 		int remainbytes;
    570 		const struct vm_page * const *pgs;
    571 		paddr_t paddr;
    572 		int size;
    573 
    574 		if (m->m_len == 0)
    575 			continue;
    576 		/*
    577 		 * Don't allow reads in read-only mbufs.
    578 		 */
    579 		if (M_ROMAP(m) && (flags & BUS_DMA_READ)) {
    580 			error = EFAULT;
    581 			break;
    582 		}
    583 		switch (m->m_flags & (M_EXT|M_CLUSTER|M_EXT_PAGES)) {
    584 		case M_EXT|M_CLUSTER:
    585 			/* XXX KDASSERT */
    586 			KASSERT(m->m_ext.ext_paddr != M_PADDR_INVALID);
    587 			paddr = m->m_ext.ext_paddr +
    588 			    (m->m_data - m->m_ext.ext_buf);
    589 			size = m->m_len;
    590 			error = _bus_dmamap_load_paddr(t, map, paddr, size,
    591 			    false);
    592 			break;
    593 
    594 		case M_EXT|M_EXT_PAGES:
    595 			KASSERT(m->m_ext.ext_buf <= m->m_data);
    596 			KASSERT(m->m_data <=
    597 			    m->m_ext.ext_buf + m->m_ext.ext_size);
    598 
    599 			offset = (vaddr_t)m->m_data -
    600 			    trunc_page((vaddr_t)m->m_ext.ext_buf);
    601 			remainbytes = m->m_len;
    602 
    603 			/* skip uninteresting pages */
    604 			pgs = (const struct vm_page * const *)
    605 			    m->m_ext.ext_pgs + (offset >> PAGE_SHIFT);
    606 
    607 			offset &= PAGE_MASK;	/* offset in the first page */
    608 
    609 			/* load each page */
    610 			while (remainbytes > 0) {
    611 				const struct vm_page *pg;
    612 
    613 				size = MIN(remainbytes, PAGE_SIZE - offset);
    614 
    615 				pg = *pgs++;
    616 				KASSERT(pg);
    617 				paddr = VM_PAGE_TO_PHYS(pg) + offset;
    618 
    619 				error = _bus_dmamap_load_paddr(t, map,
    620 				    paddr, size, false);
    621 				if (error)
    622 					break;
    623 				offset = 0;
    624 				remainbytes -= size;
    625 			}
    626 			break;
    627 
    628 		case 0:
    629 			paddr = m->m_paddr + M_BUFOFFSET(m) +
    630 			    (m->m_data - M_BUFADDR(m));
    631 			size = m->m_len;
    632 			error = _bus_dmamap_load_paddr(t, map, paddr, size,
    633 			    false);
    634 			break;
    635 
    636 		default:
    637 			error = _bus_dmamap_load_buffer(t, map, m->m_data,
    638 			    m->m_len, vmspace_kernel(), flags);
    639 		}
    640 	}
    641 	if (error == 0) {
    642 		map->dm_mapsize = m0->m_pkthdr.len;
    643 		map->_dm_origbuf = m0;
    644 		map->_dm_buftype = _BUS_DMA_BUFTYPE_MBUF;
    645 		map->_dm_vmspace = vmspace_kernel();	/* always kernel */
    646 		return 0;
    647 	}
    648 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    649 	struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
    650 	if (cookie != NULL && (cookie->id_flags & _BUS_DMA_MIGHT_NEED_BOUNCE)) {
    651 		error = _bus_dma_load_bouncebuf(t, map, m0, m0->m_pkthdr.len,
    652 		    _BUS_DMA_BUFTYPE_MBUF, flags);
    653 	}
    654 #endif
    655 	return (error);
    656 }
    657 
    658 /*
    659  * Like _bus_dmamap_load(), but for uios.
    660  */
    661 int
    662 _bus_dmamap_load_uio(bus_dma_tag_t t, bus_dmamap_t map, struct uio *uio,
    663     int flags)
    664 {
    665 	int i, error;
    666 	bus_size_t minlen, resid;
    667 	struct iovec *iov;
    668 	void *addr;
    669 
    670 	/*
    671 	 * Make sure that on error condition we return "no valid mappings."
    672 	 */
    673 	map->dm_mapsize = 0;
    674 	map->dm_nsegs = 0;
    675 	KASSERTMSG(map->dm_maxsegsz <= map->_dm_maxmaxsegsz,
    676 	    "dm_maxsegsz %lu _dm_maxmaxsegsz %lu",
    677 	    map->dm_maxsegsz, map->_dm_maxmaxsegsz);
    678 
    679 	resid = uio->uio_resid;
    680 	iov = uio->uio_iov;
    681 
    682 	/* _bus_dmamap_load_buffer() clears this if we're not... */
    683 	map->_dm_flags |= _BUS_DMAMAP_COHERENT;
    684 
    685 	error = 0;
    686 	for (i = 0; i < uio->uio_iovcnt && resid != 0 && error == 0; i++) {
    687 		/*
    688 		 * Now at the first iovec to load.  Load each iovec
    689 		 * until we have exhausted the residual count.
    690 		 */
    691 		minlen = resid < iov[i].iov_len ? resid : iov[i].iov_len;
    692 		addr = (void *)iov[i].iov_base;
    693 
    694 		error = _bus_dmamap_load_buffer(t, map, addr, minlen,
    695 		    uio->uio_vmspace, flags);
    696 
    697 		resid -= minlen;
    698 	}
    699 	if (error == 0) {
    700 		map->dm_mapsize = uio->uio_resid;
    701 		map->_dm_origbuf = uio;
    702 		map->_dm_buftype = _BUS_DMA_BUFTYPE_UIO;
    703 		map->_dm_vmspace = uio->uio_vmspace;
    704 	}
    705 	return (error);
    706 }
    707 
    708 /*
    709  * Like _bus_dmamap_load(), but for raw memory allocated with
    710  * bus_dmamem_alloc().
    711  */
    712 int
    713 _bus_dmamap_load_raw(bus_dma_tag_t t, bus_dmamap_t map,
    714     bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
    715 {
    716 
    717 	panic("_bus_dmamap_load_raw: not implemented");
    718 }
    719 
    720 /*
    721  * Common function for unloading a DMA map.  May be called by
    722  * bus-specific DMA map unload functions.
    723  */
    724 void
    725 _bus_dmamap_unload(bus_dma_tag_t t, bus_dmamap_t map)
    726 {
    727 
    728 #ifdef DEBUG_DMA
    729 	printf("dmamap_unload: t=%p map=%p\n", t, map);
    730 #endif	/* DEBUG_DMA */
    731 
    732 	/*
    733 	 * No resources to free; just mark the mappings as
    734 	 * invalid.
    735 	 */
    736 	map->dm_mapsize = 0;
    737 	map->dm_nsegs = 0;
    738 	map->_dm_origbuf = NULL;
    739 	map->_dm_buftype = _BUS_DMA_BUFTYPE_INVALID;
    740 	map->_dm_vmspace = NULL;
    741 }
    742 
    743 static void
    744 _bus_dmamap_sync_segment(vaddr_t va, paddr_t pa, vsize_t len, int ops, bool readonly_p)
    745 {
    746 	KASSERT((va & PAGE_MASK) == (pa & PAGE_MASK));
    747 #if 0
    748 	printf("sync_segment: va=%#lx pa=%#lx len=%#lx ops=%#x ro=%d\n",
    749 	    va, pa, len, ops, readonly_p);
    750 #endif
    751 
    752 	switch (ops) {
    753 	case BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE:
    754 		if (!readonly_p) {
    755 			STAT_INCR(sync_prereadwrite);
    756 			cpu_dcache_wbinv_range(va, len);
    757 			cpu_sdcache_wbinv_range(va, pa, len);
    758 			break;
    759 		}
    760 		/* FALLTHROUGH */
    761 
    762 	case BUS_DMASYNC_PREREAD: {
    763 		const size_t line_size = arm_dcache_align;
    764 		const size_t line_mask = arm_dcache_align_mask;
    765 		vsize_t misalignment = va & line_mask;
    766 		if (misalignment) {
    767 			va -= misalignment;
    768 			pa -= misalignment;
    769 			len += misalignment;
    770 			STAT_INCR(sync_preread_begin);
    771 			cpu_dcache_wbinv_range(va, line_size);
    772 			cpu_sdcache_wbinv_range(va, pa, line_size);
    773 			if (len <= line_size)
    774 				break;
    775 			va += line_size;
    776 			pa += line_size;
    777 			len -= line_size;
    778 		}
    779 		misalignment = len & line_mask;
    780 		len -= misalignment;
    781 		if (len > 0) {
    782 			STAT_INCR(sync_preread);
    783 			cpu_dcache_inv_range(va, len);
    784 			cpu_sdcache_inv_range(va, pa, len);
    785 		}
    786 		if (misalignment) {
    787 			va += len;
    788 			pa += len;
    789 			STAT_INCR(sync_preread_tail);
    790 			cpu_dcache_wbinv_range(va, line_size);
    791 			cpu_sdcache_wbinv_range(va, pa, line_size);
    792 		}
    793 		break;
    794 	}
    795 
    796 	case BUS_DMASYNC_PREWRITE:
    797 		STAT_INCR(sync_prewrite);
    798 		cpu_dcache_wb_range(va, len);
    799 		cpu_sdcache_wb_range(va, pa, len);
    800 		break;
    801 
    802 #ifdef CPU_CORTEX
    803 	/*
    804 	 * Cortex CPUs can do speculative loads so we need to clean the cache
    805 	 * after a DMA read to deal with any speculatively loaded cache lines.
    806 	 * Since these can't be dirty, we can just invalidate them and don't
    807 	 * have to worry about having to write back their contents.
    808 	 */
    809 	case BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE:
    810 		STAT_INCR(sync_postreadwrite);
    811 		cpu_dcache_inv_range(va, len);
    812 		cpu_sdcache_inv_range(va, pa, len);
    813 		break;
    814 	case BUS_DMASYNC_POSTREAD:
    815 		STAT_INCR(sync_postread);
    816 		cpu_dcache_inv_range(va, len);
    817 		cpu_sdcache_inv_range(va, pa, len);
    818 		break;
    819 #endif
    820 	}
    821 }
    822 
    823 static inline void
    824 _bus_dmamap_sync_linear(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
    825     bus_size_t len, int ops)
    826 {
    827 	bus_dma_segment_t *ds = map->dm_segs;
    828 	vaddr_t va = (vaddr_t) map->_dm_origbuf;
    829 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
    830 	if (map->_dm_flags & _BUS_DMAMAP_IS_BOUNCING) {
    831 		struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
    832 		va = (vaddr_t) cookie->id_bouncebuf;
    833 	}
    834 #endif
    835 
    836 	while (len > 0) {
    837 		while (offset >= ds->ds_len) {
    838 			offset -= ds->ds_len;
    839 			va += ds->ds_len;
    840 			ds++;
    841 		}
    842 
    843 		paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + offset);
    844 		size_t seglen = min(len, ds->ds_len - offset);
    845 
    846 		if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
    847 			_bus_dmamap_sync_segment(va + offset, pa, seglen, ops,
    848 			    false);
    849 
    850 		offset += seglen;
    851 		len -= seglen;
    852 	}
    853 }
    854 
    855 static inline void
    856 _bus_dmamap_sync_mbuf(bus_dma_tag_t t, bus_dmamap_t map, bus_size_t offset,
    857     bus_size_t len, int ops)
    858 {
    859 	bus_dma_segment_t *ds = map->dm_segs;
    860 	struct mbuf *m = map->_dm_origbuf;
    861 	bus_size_t voff = offset;
    862 	bus_size_t ds_off = offset;
    863 
    864 	while (len > 0) {
    865 		/* Find the current dma segment */
    866 		while (ds_off >= ds->ds_len) {
    867 			ds_off -= ds->ds_len;
    868 			ds++;
    869 		}
    870 		/* Find the current mbuf. */
    871 		while (voff >= m->m_len) {
    872 			voff -= m->m_len;
    873 			m = m->m_next;
    874 		}
    875 
    876 		/*
    877 		 * Now at the first mbuf to sync; nail each one until
    878 		 * we have exhausted the length.
    879 		 */
    880 		vsize_t seglen = min(len, min(m->m_len - voff, ds->ds_len - ds_off));
    881 		vaddr_t va = mtod(m, vaddr_t) + voff;
    882 		paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
    883 
    884 		/*
    885 		 * We can save a lot of work here if we know the mapping
    886 		 * is read-only at the MMU:
    887 		 *
    888 		 * If a mapping is read-only, no dirty cache blocks will
    889 		 * exist for it.  If a writable mapping was made read-only,
    890 		 * we know any dirty cache lines for the range will have
    891 		 * been cleaned for us already.  Therefore, if the upper
    892 		 * layer can tell us we have a read-only mapping, we can
    893 		 * skip all cache cleaning.
    894 		 *
    895 		 * NOTE: This only works if we know the pmap cleans pages
    896 		 * before making a read-write -> read-only transition.  If
    897 		 * this ever becomes non-true (e.g. Physically Indexed
    898 		 * cache), this will have to be revisited.
    899 		 */
    900 
    901 		if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
    902 			_bus_dmamap_sync_segment(va, pa, seglen, ops,
    903 			    M_ROMAP(m));
    904 		voff += seglen;
    905 		ds_off += seglen;
    906 		len -= seglen;
    907 	}
    908 }
    909 
    910 static inline void
    911 _bus_dmamap_sync_uio(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
    912     bus_size_t len, int ops)
    913 {
    914 	bus_dma_segment_t *ds = map->dm_segs;
    915 	struct uio *uio = map->_dm_origbuf;
    916 	struct iovec *iov = uio->uio_iov;
    917 	bus_size_t voff = offset;
    918 	bus_size_t ds_off = offset;
    919 
    920 	while (len > 0) {
    921 		/* Find the current dma segment */
    922 		while (ds_off >= ds->ds_len) {
    923 			ds_off -= ds->ds_len;
    924 			ds++;
    925 		}
    926 
    927 		/* Find the current iovec. */
    928 		while (voff >= iov->iov_len) {
    929 			voff -= iov->iov_len;
    930 			iov++;
    931 		}
    932 
    933 		/*
    934 		 * Now at the first iovec to sync; nail each one until
    935 		 * we have exhausted the length.
    936 		 */
    937 		vsize_t seglen = min(len, min(iov->iov_len - voff, ds->ds_len - ds_off));
    938 		vaddr_t va = (vaddr_t) iov->iov_base + voff;
    939 		paddr_t pa = _bus_dma_busaddr_to_paddr(t, ds->ds_addr + ds_off);
    940 
    941 		if ((ds->_ds_flags & _BUS_DMAMAP_COHERENT) == 0)
    942 			_bus_dmamap_sync_segment(va, pa, seglen, ops, false);
    943 
    944 		voff += seglen;
    945 		ds_off += seglen;
    946 		len -= seglen;
    947 	}
    948 }
    949 
    950 /*
    951  * Common function for DMA map synchronization.  May be called
    952  * by bus-specific DMA map synchronization functions.
    953  *
    954  * This version works for the Virtually Indexed Virtually Tagged
    955  * cache found on 32-bit ARM processors.
    956  *
    957  * XXX Should have separate versions for write-through vs.
    958  * XXX write-back caches.  We currently assume write-back
    959  * XXX here, which is not as efficient as it could be for
    960  * XXX the write-through case.
    961  */
    962 void
    963 _bus_dmamap_sync(bus_dma_tag_t t, bus_dmamap_t map, bus_addr_t offset,
    964     bus_size_t len, int ops)
    965 {
    966 #ifdef DEBUG_DMA
    967 	printf("dmamap_sync: t=%p map=%p offset=%lx len=%lx ops=%x\n",
    968 	    t, map, offset, len, ops);
    969 #endif	/* DEBUG_DMA */
    970 
    971 	/*
    972 	 * Mixing of PRE and POST operations is not allowed.
    973 	 */
    974 	if ((ops & (BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE)) != 0 &&
    975 	    (ops & (BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE)) != 0)
    976 		panic("_bus_dmamap_sync: mix PRE and POST");
    977 
    978 #ifdef DIAGNOSTIC
    979 	if (offset >= map->dm_mapsize)
    980 		panic("_bus_dmamap_sync: bad offset %lu (map size is %lu)",
    981 		    offset, map->dm_mapsize);
    982 	if (len == 0 || (offset + len) > map->dm_mapsize)
    983 		panic("_bus_dmamap_sync: bad length");
    984 #endif
    985 
    986 	/*
    987 	 * For a virtually-indexed write-back cache, we need
    988 	 * to do the following things:
    989 	 *
    990 	 *	PREREAD -- Invalidate the D-cache.  We do this
    991 	 *	here in case a write-back is required by the back-end.
    992 	 *
    993 	 *	PREWRITE -- Write-back the D-cache.  Note that if
    994 	 *	we are doing a PREREAD|PREWRITE, we can collapse
    995 	 *	the whole thing into a single Wb-Inv.
    996 	 *
    997 	 *	POSTREAD -- Re-invalidate the D-cache in case speculative
    998 	 *	memory accesses caused cachelines to become valid with now
    999 	 *	invalid data.
   1000 	 *
   1001 	 *	POSTWRITE -- Nothing.
   1002 	 */
   1003 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1004 	const bool bouncing = (map->_dm_flags & _BUS_DMAMAP_IS_BOUNCING);
   1005 #else
   1006 	const bool bouncing = false;
   1007 #endif
   1008 
   1009 	const int pre_ops = ops & (BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
   1010 #ifdef CPU_CORTEX
   1011 	const int post_ops = ops & (BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
   1012 #else
   1013 	const int post_ops = 0;
   1014 #endif
   1015 	if (!bouncing && pre_ops == 0 && post_ops == BUS_DMASYNC_POSTWRITE) {
   1016 		STAT_INCR(sync_postwrite);
   1017 		return;
   1018 	}
   1019 	KASSERTMSG(bouncing || pre_ops != 0 || (post_ops & BUS_DMASYNC_POSTREAD),
   1020 	    "pre_ops %#x post_ops %#x", pre_ops, post_ops);
   1021 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1022 	if (bouncing && (ops & BUS_DMASYNC_PREWRITE)) {
   1023 		struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
   1024 		STAT_INCR(write_bounces);
   1025 		char * const dataptr = (char *)cookie->id_bouncebuf + offset;
   1026 		/*
   1027 		 * Copy the caller's buffer to the bounce buffer.
   1028 		 */
   1029 		switch (map->_dm_buftype) {
   1030 		case _BUS_DMA_BUFTYPE_LINEAR:
   1031 			memcpy(dataptr, cookie->id_origlinearbuf + offset, len);
   1032 			break;
   1033 		case _BUS_DMA_BUFTYPE_MBUF:
   1034 			m_copydata(cookie->id_origmbuf, offset, len, dataptr);
   1035 			break;
   1036 		case _BUS_DMA_BUFTYPE_UIO:
   1037 			_bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_WRITE);
   1038 			break;
   1039 #ifdef DIAGNOSTIC
   1040 		case _BUS_DMA_BUFTYPE_RAW:
   1041 			panic("_bus_dmamap_sync(pre): _BUS_DMA_BUFTYPE_RAW");
   1042 			break;
   1043 
   1044 		case _BUS_DMA_BUFTYPE_INVALID:
   1045 			panic("_bus_dmamap_sync(pre): _BUS_DMA_BUFTYPE_INVALID");
   1046 			break;
   1047 
   1048 		default:
   1049 			panic("_bus_dmamap_sync(pre): map %p: unknown buffer type %d\n",
   1050 			    map, map->_dm_buftype);
   1051 			break;
   1052 #endif /* DIAGNOSTIC */
   1053 		}
   1054 	}
   1055 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
   1056 
   1057 	/* Skip cache frobbing if mapping was COHERENT. */
   1058 	if (!bouncing && (map->_dm_flags & _BUS_DMAMAP_COHERENT)) {
   1059 		/* Drain the write buffer. */
   1060 		if (pre_ops & BUS_DMASYNC_PREWRITE)
   1061 			cpu_drain_writebuf();
   1062 		return;
   1063 	}
   1064 
   1065 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1066 	if (bouncing && ((map->_dm_flags & _BUS_DMAMAP_COHERENT) || pre_ops == 0)) {
   1067 		goto bounce_it;
   1068 	}
   1069 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
   1070 
   1071 	/*
   1072 	 * If the mapping belongs to a non-kernel vmspace, and the
   1073 	 * vmspace has not been active since the last time a full
   1074 	 * cache flush was performed, we don't need to do anything.
   1075 	 */
   1076 	if (__predict_false(!VMSPACE_IS_KERNEL_P(map->_dm_vmspace) &&
   1077 	    vm_map_pmap(&map->_dm_vmspace->vm_map)->pm_cstate.cs_cache_d == 0))
   1078 		return;
   1079 
   1080 	int buftype = map->_dm_buftype;
   1081 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1082 	if (bouncing) {
   1083 		buftype = _BUS_DMA_BUFTYPE_LINEAR;
   1084 	}
   1085 #endif
   1086 
   1087 	switch (buftype) {
   1088 	case _BUS_DMA_BUFTYPE_LINEAR:
   1089 		_bus_dmamap_sync_linear(t, map, offset, len, ops);
   1090 		break;
   1091 
   1092 	case _BUS_DMA_BUFTYPE_MBUF:
   1093 		_bus_dmamap_sync_mbuf(t, map, offset, len, ops);
   1094 		break;
   1095 
   1096 	case _BUS_DMA_BUFTYPE_UIO:
   1097 		_bus_dmamap_sync_uio(t, map, offset, len, ops);
   1098 		break;
   1099 
   1100 	case _BUS_DMA_BUFTYPE_RAW:
   1101 		panic("_bus_dmamap_sync: _BUS_DMA_BUFTYPE_RAW");
   1102 		break;
   1103 
   1104 	case _BUS_DMA_BUFTYPE_INVALID:
   1105 		panic("_bus_dmamap_sync: _BUS_DMA_BUFTYPE_INVALID");
   1106 		break;
   1107 
   1108 	default:
   1109 		panic("_bus_dmamap_sync: map %p: unknown buffer type %d\n",
   1110 		    map, map->_dm_buftype);
   1111 	}
   1112 
   1113 	/* Drain the write buffer. */
   1114 	cpu_drain_writebuf();
   1115 
   1116 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1117   bounce_it:
   1118 	if (!bouncing || (ops & BUS_DMASYNC_POSTREAD) == 0)
   1119 		return;
   1120 
   1121 	struct arm32_bus_dma_cookie * const cookie = map->_dm_cookie;
   1122 	char * const dataptr = (char *)cookie->id_bouncebuf + offset;
   1123 	STAT_INCR(read_bounces);
   1124 	/*
   1125 	 * Copy the bounce buffer to the caller's buffer.
   1126 	 */
   1127 	switch (map->_dm_buftype) {
   1128 	case _BUS_DMA_BUFTYPE_LINEAR:
   1129 		memcpy(cookie->id_origlinearbuf + offset, dataptr, len);
   1130 		break;
   1131 
   1132 	case _BUS_DMA_BUFTYPE_MBUF:
   1133 		m_copyback(cookie->id_origmbuf, offset, len, dataptr);
   1134 		break;
   1135 
   1136 	case _BUS_DMA_BUFTYPE_UIO:
   1137 		_bus_dma_uiomove(dataptr, cookie->id_origuio, len, UIO_READ);
   1138 		break;
   1139 #ifdef DIAGNOSTIC
   1140 	case _BUS_DMA_BUFTYPE_RAW:
   1141 		panic("_bus_dmamap_sync(post): _BUS_DMA_BUFTYPE_RAW");
   1142 		break;
   1143 
   1144 	case _BUS_DMA_BUFTYPE_INVALID:
   1145 		panic("_bus_dmamap_sync(post): _BUS_DMA_BUFTYPE_INVALID");
   1146 		break;
   1147 
   1148 	default:
   1149 		panic("_bus_dmamap_sync(post): map %p: unknown buffer type %d\n",
   1150 		    map, map->_dm_buftype);
   1151 		break;
   1152 #endif
   1153 	}
   1154 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
   1155 }
   1156 
   1157 /*
   1158  * Common function for DMA-safe memory allocation.  May be called
   1159  * by bus-specific DMA memory allocation functions.
   1160  */
   1161 
   1162 extern paddr_t physical_start;
   1163 extern paddr_t physical_end;
   1164 
   1165 int
   1166 _bus_dmamem_alloc(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
   1167     bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
   1168     int flags)
   1169 {
   1170 	struct arm32_dma_range *dr;
   1171 	int error, i;
   1172 
   1173 #ifdef DEBUG_DMA
   1174 	printf("dmamem_alloc t=%p size=%lx align=%lx boundary=%lx "
   1175 	    "segs=%p nsegs=%x rsegs=%p flags=%x\n", t, size, alignment,
   1176 	    boundary, segs, nsegs, rsegs, flags);
   1177 #endif
   1178 
   1179 	if ((dr = t->_ranges) != NULL) {
   1180 		error = ENOMEM;
   1181 		for (i = 0; i < t->_nranges; i++, dr++) {
   1182 			if (dr->dr_len == 0
   1183 			    || (dr->dr_flags & _BUS_DMAMAP_NOALLOC))
   1184 				continue;
   1185 			error = _bus_dmamem_alloc_range(t, size, alignment,
   1186 			    boundary, segs, nsegs, rsegs, flags,
   1187 			    trunc_page(dr->dr_sysbase),
   1188 			    trunc_page(dr->dr_sysbase + dr->dr_len));
   1189 			if (error == 0)
   1190 				break;
   1191 		}
   1192 	} else {
   1193 		error = _bus_dmamem_alloc_range(t, size, alignment, boundary,
   1194 		    segs, nsegs, rsegs, flags, trunc_page(physical_start),
   1195 		    trunc_page(physical_end));
   1196 	}
   1197 
   1198 #ifdef DEBUG_DMA
   1199 	printf("dmamem_alloc: =%d\n", error);
   1200 #endif
   1201 
   1202 	return(error);
   1203 }
   1204 
   1205 /*
   1206  * Common function for freeing DMA-safe memory.  May be called by
   1207  * bus-specific DMA memory free functions.
   1208  */
   1209 void
   1210 _bus_dmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
   1211 {
   1212 	struct vm_page *m;
   1213 	bus_addr_t addr;
   1214 	struct pglist mlist;
   1215 	int curseg;
   1216 
   1217 #ifdef DEBUG_DMA
   1218 	printf("dmamem_free: t=%p segs=%p nsegs=%x\n", t, segs, nsegs);
   1219 #endif	/* DEBUG_DMA */
   1220 
   1221 	/*
   1222 	 * Build a list of pages to free back to the VM system.
   1223 	 */
   1224 	TAILQ_INIT(&mlist);
   1225 	for (curseg = 0; curseg < nsegs; curseg++) {
   1226 		for (addr = segs[curseg].ds_addr;
   1227 		    addr < (segs[curseg].ds_addr + segs[curseg].ds_len);
   1228 		    addr += PAGE_SIZE) {
   1229 			m = PHYS_TO_VM_PAGE(addr);
   1230 			TAILQ_INSERT_TAIL(&mlist, m, pageq.queue);
   1231 		}
   1232 	}
   1233 	uvm_pglistfree(&mlist);
   1234 }
   1235 
   1236 /*
   1237  * Common function for mapping DMA-safe memory.  May be called by
   1238  * bus-specific DMA memory map functions.
   1239  */
   1240 int
   1241 _bus_dmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
   1242     size_t size, void **kvap, int flags)
   1243 {
   1244 	vaddr_t va;
   1245 	paddr_t pa;
   1246 	int curseg;
   1247 	pt_entry_t *ptep;
   1248 	const uvm_flag_t kmflags = UVM_KMF_VAONLY
   1249 	    | ((flags & BUS_DMA_NOWAIT) != 0 ? UVM_KMF_NOWAIT : 0);
   1250 	vsize_t align = 0;
   1251 
   1252 #ifdef DEBUG_DMA
   1253 	printf("dmamem_map: t=%p segs=%p nsegs=%x size=%lx flags=%x\n", t,
   1254 	    segs, nsegs, (unsigned long)size, flags);
   1255 #endif	/* DEBUG_DMA */
   1256 
   1257 #ifdef PMAP_MAP_POOLPAGE
   1258 	/*
   1259 	 * If all of memory is mapped, and we are mapping a single physically
   1260 	 * contiguous area then this area is already mapped.  Let's see if we
   1261 	 * avoid having a separate mapping for it.
   1262 	 */
   1263 	if (nsegs == 1) {
   1264 		/*
   1265 		 * If this is a non-COHERENT mapping, then the existing kernel
   1266 		 * mapping is already compatible with it.
   1267 		 */
   1268 		bool direct_mapable = (flags & BUS_DMA_COHERENT) == 0;
   1269 		pa = segs[0].ds_addr;
   1270 
   1271 		/*
   1272 		 * This is a COHERENT mapping which, unless this address is in
   1273 		 * a COHERENT dma range, will not be compatible.
   1274 		 */
   1275 		if (t->_ranges != NULL) {
   1276 			const struct arm32_dma_range * const dr =
   1277 			    _bus_dma_paddr_inrange(t->_ranges, t->_nranges, pa);
   1278 			if (dr != NULL
   1279 			    && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
   1280 				direct_mapable = true;
   1281 			}
   1282 		}
   1283 
   1284 		if (direct_mapable) {
   1285 			*kvap = (void *)PMAP_MAP_POOLPAGE(pa);
   1286 #ifdef DEBUG_DMA
   1287 			printf("dmamem_map: =%p\n", *kvap);
   1288 #endif	/* DEBUG_DMA */
   1289 			return 0;
   1290 		}
   1291 	}
   1292 #endif
   1293 
   1294 	size = round_page(size);
   1295 	if (__predict_true(size > L2_L_SIZE)) {
   1296 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
   1297 		if (size >= L1_SS_SIZE)
   1298 			align = L1_SS_SIZE;
   1299 		else
   1300 #endif
   1301 		if (size >= L1_S_SIZE)
   1302 			align = L1_S_SIZE;
   1303 		else
   1304 			align = L2_S_SIZE;
   1305 	}
   1306 
   1307 	va = uvm_km_alloc(kernel_map, size, align, kmflags);
   1308 	if (__predict_false(va == 0 && align > 0)) {
   1309 		align = 0;
   1310 		va = uvm_km_alloc(kernel_map, size, 0, kmflags);
   1311 	}
   1312 
   1313 	if (va == 0)
   1314 		return (ENOMEM);
   1315 
   1316 	*kvap = (void *)va;
   1317 
   1318 	for (curseg = 0; curseg < nsegs; curseg++) {
   1319 		for (pa = segs[curseg].ds_addr;
   1320 		    pa < (segs[curseg].ds_addr + segs[curseg].ds_len);
   1321 		    pa += PAGE_SIZE, va += PAGE_SIZE, size -= PAGE_SIZE) {
   1322 			bool uncached = (flags & BUS_DMA_COHERENT);
   1323 #ifdef DEBUG_DMA
   1324 			printf("wiring p%lx to v%lx", pa, va);
   1325 #endif	/* DEBUG_DMA */
   1326 			if (size == 0)
   1327 				panic("_bus_dmamem_map: size botch");
   1328 
   1329 			const struct arm32_dma_range * const dr =
   1330 			    _bus_dma_paddr_inrange(t->_ranges, t->_nranges, pa);
   1331 			/*
   1332 			 * If this dma region is coherent then there is
   1333 			 * no need for an uncached mapping.
   1334 			 */
   1335 			if (dr != NULL
   1336 			    && (dr->dr_flags & _BUS_DMAMAP_COHERENT)) {
   1337 				uncached = false;
   1338 			}
   1339 
   1340 			pmap_kenter_pa(va, pa,
   1341 			    VM_PROT_READ | VM_PROT_WRITE, PMAP_WIRED);
   1342 
   1343 			/*
   1344 			 * If the memory must remain coherent with the
   1345 			 * cache then we must make the memory uncacheable
   1346 			 * in order to maintain virtual cache coherency.
   1347 			 * We must also guarantee the cache does not already
   1348 			 * contain the virtal addresses we are making
   1349 			 * uncacheable.
   1350 			 */
   1351 			if (uncached) {
   1352 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
   1353 				cpu_sdcache_wbinv_range(va, pa, PAGE_SIZE);
   1354 				cpu_drain_writebuf();
   1355 				ptep = vtopte(va);
   1356 				*ptep &= ~L2_S_CACHE_MASK;
   1357 				PTE_SYNC(ptep);
   1358 				tlb_flush();
   1359 			}
   1360 #ifdef DEBUG_DMA
   1361 			ptep = vtopte(va);
   1362 			printf(" pte=v%p *pte=%x\n", ptep, *ptep);
   1363 #endif	/* DEBUG_DMA */
   1364 		}
   1365 	}
   1366 	pmap_update(pmap_kernel());
   1367 #ifdef DEBUG_DMA
   1368 	printf("dmamem_map: =%p\n", *kvap);
   1369 #endif	/* DEBUG_DMA */
   1370 	return (0);
   1371 }
   1372 
   1373 /*
   1374  * Common function for unmapping DMA-safe memory.  May be called by
   1375  * bus-specific DMA memory unmapping functions.
   1376  */
   1377 void
   1378 _bus_dmamem_unmap(bus_dma_tag_t t, void *kva, size_t size)
   1379 {
   1380 
   1381 #ifdef DEBUG_DMA
   1382 	printf("dmamem_unmap: t=%p kva=%p size=%zx\n", t, kva, size);
   1383 #endif	/* DEBUG_DMA */
   1384 #ifdef DIAGNOSTIC
   1385 	if ((u_long)kva & PGOFSET)
   1386 		panic("_bus_dmamem_unmap");
   1387 #endif	/* DIAGNOSTIC */
   1388 
   1389 	size = round_page(size);
   1390 	pmap_kremove((vaddr_t)kva, size);
   1391 	pmap_update(pmap_kernel());
   1392 	uvm_km_free(kernel_map, (vaddr_t)kva, size, UVM_KMF_VAONLY);
   1393 }
   1394 
   1395 /*
   1396  * Common functin for mmap(2)'ing DMA-safe memory.  May be called by
   1397  * bus-specific DMA mmap(2)'ing functions.
   1398  */
   1399 paddr_t
   1400 _bus_dmamem_mmap(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
   1401     off_t off, int prot, int flags)
   1402 {
   1403 	paddr_t map_flags;
   1404 	int i;
   1405 
   1406 	for (i = 0; i < nsegs; i++) {
   1407 #ifdef DIAGNOSTIC
   1408 		if (off & PGOFSET)
   1409 			panic("_bus_dmamem_mmap: offset unaligned");
   1410 		if (segs[i].ds_addr & PGOFSET)
   1411 			panic("_bus_dmamem_mmap: segment unaligned");
   1412 		if (segs[i].ds_len & PGOFSET)
   1413 			panic("_bus_dmamem_mmap: segment size not multiple"
   1414 			    " of page size");
   1415 #endif	/* DIAGNOSTIC */
   1416 		if (off >= segs[i].ds_len) {
   1417 			off -= segs[i].ds_len;
   1418 			continue;
   1419 		}
   1420 
   1421 		map_flags = 0;
   1422 		if (flags & BUS_DMA_PREFETCHABLE)
   1423 			map_flags |= ARM32_MMAP_WRITECOMBINE;
   1424 
   1425 		return (arm_btop((u_long)segs[i].ds_addr + off) | map_flags);
   1426 
   1427 	}
   1428 
   1429 	/* Page not found. */
   1430 	return (-1);
   1431 }
   1432 
   1433 /**********************************************************************
   1434  * DMA utility functions
   1435  **********************************************************************/
   1436 
   1437 /*
   1438  * Utility function to load a linear buffer.  lastaddrp holds state
   1439  * between invocations (for multiple-buffer loads).  segp contains
   1440  * the starting segment on entrace, and the ending segment on exit.
   1441  * first indicates if this is the first invocation of this function.
   1442  */
   1443 int
   1444 _bus_dmamap_load_buffer(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
   1445     bus_size_t buflen, struct vmspace *vm, int flags)
   1446 {
   1447 	bus_size_t sgsize;
   1448 	bus_addr_t curaddr;
   1449 	vaddr_t vaddr = (vaddr_t)buf;
   1450 	int error;
   1451 	pmap_t pmap;
   1452 
   1453 #ifdef DEBUG_DMA
   1454 	printf("_bus_dmamem_load_buffer(buf=%p, len=%lx, flags=%d)\n",
   1455 	    buf, buflen, flags);
   1456 #endif	/* DEBUG_DMA */
   1457 
   1458 	pmap = vm_map_pmap(&vm->vm_map);
   1459 
   1460 	while (buflen > 0) {
   1461 		/*
   1462 		 * Get the physical address for this segment.
   1463 		 *
   1464 		 * XXX Doesn't support checking for coherent mappings
   1465 		 * XXX in user address space.
   1466 		 */
   1467 		bool coherent;
   1468 		if (__predict_true(pmap == pmap_kernel())) {
   1469 			pd_entry_t *pde;
   1470 			pt_entry_t *ptep;
   1471 			(void) pmap_get_pde_pte(pmap, vaddr, &pde, &ptep);
   1472 			if (__predict_false(pmap_pde_section(pde))) {
   1473 				paddr_t s_frame = L1_S_FRAME;
   1474 				paddr_t s_offset = L1_S_OFFSET;
   1475 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
   1476 				if (__predict_false(pmap_pde_supersection(pde))) {
   1477 					s_frame = L1_SS_FRAME;
   1478 					s_offset = L1_SS_OFFSET;
   1479 				}
   1480 #endif
   1481 				curaddr = (*pde & s_frame) | (vaddr & s_offset);
   1482 				coherent = (*pde & L1_S_CACHE_MASK) == 0;
   1483 			} else {
   1484 				pt_entry_t pte = *ptep;
   1485 				KDASSERTMSG((pte & L2_TYPE_MASK) != L2_TYPE_INV,
   1486 				    "va=%#"PRIxVADDR" pde=%#x ptep=%p pte=%#x",
   1487 				    vaddr, *pde, ptep, pte);
   1488 				if (__predict_false((pte & L2_TYPE_MASK)
   1489 						    == L2_TYPE_L)) {
   1490 					curaddr = (pte & L2_L_FRAME) |
   1491 					    (vaddr & L2_L_OFFSET);
   1492 					coherent = (pte & L2_L_CACHE_MASK) == 0;
   1493 				} else {
   1494 					curaddr = (pte & L2_S_FRAME) |
   1495 					    (vaddr & L2_S_OFFSET);
   1496 					coherent = (pte & L2_S_CACHE_MASK) == 0;
   1497 				}
   1498 			}
   1499 		} else {
   1500 			(void) pmap_extract(pmap, vaddr, &curaddr);
   1501 			coherent = false;
   1502 		}
   1503 
   1504 		/*
   1505 		 * Compute the segment size, and adjust counts.
   1506 		 */
   1507 		sgsize = PAGE_SIZE - ((u_long)vaddr & PGOFSET);
   1508 		if (buflen < sgsize)
   1509 			sgsize = buflen;
   1510 
   1511 		error = _bus_dmamap_load_paddr(t, map, curaddr, sgsize,
   1512 		    coherent);
   1513 		if (error)
   1514 			return (error);
   1515 
   1516 		vaddr += sgsize;
   1517 		buflen -= sgsize;
   1518 	}
   1519 
   1520 	return (0);
   1521 }
   1522 
   1523 /*
   1524  * Allocate physical memory from the given physical address range.
   1525  * Called by DMA-safe memory allocation methods.
   1526  */
   1527 int
   1528 _bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size, bus_size_t alignment,
   1529     bus_size_t boundary, bus_dma_segment_t *segs, int nsegs, int *rsegs,
   1530     int flags, paddr_t low, paddr_t high)
   1531 {
   1532 	paddr_t curaddr, lastaddr;
   1533 	struct vm_page *m;
   1534 	struct pglist mlist;
   1535 	int curseg, error;
   1536 
   1537 	KASSERTMSG(boundary == 0 || (boundary & (boundary-1)) == 0,
   1538 	    "invalid boundary %#lx", boundary);
   1539 
   1540 #ifdef DEBUG_DMA
   1541 	printf("alloc_range: t=%p size=%lx align=%lx boundary=%lx segs=%p nsegs=%x rsegs=%p flags=%x lo=%lx hi=%lx\n",
   1542 	    t, size, alignment, boundary, segs, nsegs, rsegs, flags, low, high);
   1543 #endif	/* DEBUG_DMA */
   1544 
   1545 	/* Always round the size. */
   1546 	size = round_page(size);
   1547 
   1548 	/*
   1549 	 * We accept boundaries < size, splitting in multiple segments
   1550 	 * if needed. uvm_pglistalloc does not, so compute an appropriate
   1551 	 * boundary: next power of 2 >= size
   1552 	 */
   1553 	bus_size_t uboundary = boundary;
   1554 	if (uboundary <= PAGE_SIZE) {
   1555 		uboundary = 0;
   1556 	} else {
   1557 		while (uboundary < size) {
   1558 			uboundary <<= 1;
   1559 		}
   1560 	}
   1561 
   1562 	/*
   1563 	 * Allocate pages from the VM system.
   1564 	 */
   1565 	error = uvm_pglistalloc(size, low, high, alignment, uboundary,
   1566 	    &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
   1567 	if (error)
   1568 		return (error);
   1569 
   1570 	/*
   1571 	 * Compute the location, size, and number of segments actually
   1572 	 * returned by the VM code.
   1573 	 */
   1574 	m = TAILQ_FIRST(&mlist);
   1575 	curseg = 0;
   1576 	lastaddr = segs[curseg].ds_addr = VM_PAGE_TO_PHYS(m);
   1577 	segs[curseg].ds_len = PAGE_SIZE;
   1578 #ifdef DEBUG_DMA
   1579 		printf("alloc: page %lx\n", lastaddr);
   1580 #endif	/* DEBUG_DMA */
   1581 	m = TAILQ_NEXT(m, pageq.queue);
   1582 
   1583 	for (; m != NULL; m = TAILQ_NEXT(m, pageq.queue)) {
   1584 		curaddr = VM_PAGE_TO_PHYS(m);
   1585 		KASSERTMSG(low <= curaddr && curaddr < high,
   1586 		    "uvm_pglistalloc returned non-sensicaladdress %#lx "
   1587 		    "(low=%#lx, high=%#lx\n", curaddr, low, high);
   1588 #ifdef DEBUG_DMA
   1589 		printf("alloc: page %lx\n", curaddr);
   1590 #endif	/* DEBUG_DMA */
   1591 		if (curaddr == lastaddr + PAGE_SIZE
   1592 		    && (lastaddr & boundary) == (curaddr & boundary))
   1593 			segs[curseg].ds_len += PAGE_SIZE;
   1594 		else {
   1595 			curseg++;
   1596 			if (curseg >= nsegs) {
   1597 				uvm_pglistfree(&mlist);
   1598 				return EFBIG;
   1599 			}
   1600 			segs[curseg].ds_addr = curaddr;
   1601 			segs[curseg].ds_len = PAGE_SIZE;
   1602 		}
   1603 		lastaddr = curaddr;
   1604 	}
   1605 
   1606 	*rsegs = curseg + 1;
   1607 
   1608 	return (0);
   1609 }
   1610 
   1611 /*
   1612  * Check if a memory region intersects with a DMA range, and return the
   1613  * page-rounded intersection if it does.
   1614  */
   1615 int
   1616 arm32_dma_range_intersect(struct arm32_dma_range *ranges, int nranges,
   1617     paddr_t pa, psize_t size, paddr_t *pap, psize_t *sizep)
   1618 {
   1619 	struct arm32_dma_range *dr;
   1620 	int i;
   1621 
   1622 	if (ranges == NULL)
   1623 		return (0);
   1624 
   1625 	for (i = 0, dr = ranges; i < nranges; i++, dr++) {
   1626 		if (dr->dr_sysbase <= pa &&
   1627 		    pa < (dr->dr_sysbase + dr->dr_len)) {
   1628 			/*
   1629 			 * Beginning of region intersects with this range.
   1630 			 */
   1631 			*pap = trunc_page(pa);
   1632 			*sizep = round_page(min(pa + size,
   1633 			    dr->dr_sysbase + dr->dr_len) - pa);
   1634 			return (1);
   1635 		}
   1636 		if (pa < dr->dr_sysbase && dr->dr_sysbase < (pa + size)) {
   1637 			/*
   1638 			 * End of region intersects with this range.
   1639 			 */
   1640 			*pap = trunc_page(dr->dr_sysbase);
   1641 			*sizep = round_page(min((pa + size) - dr->dr_sysbase,
   1642 			    dr->dr_len));
   1643 			return (1);
   1644 		}
   1645 	}
   1646 
   1647 	/* No intersection found. */
   1648 	return (0);
   1649 }
   1650 
   1651 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1652 static int
   1653 _bus_dma_alloc_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map,
   1654     bus_size_t size, int flags)
   1655 {
   1656 	struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
   1657 	int error = 0;
   1658 
   1659 #ifdef DIAGNOSTIC
   1660 	if (cookie == NULL)
   1661 		panic("_bus_dma_alloc_bouncebuf: no cookie");
   1662 #endif
   1663 
   1664 	cookie->id_bouncebuflen = round_page(size);
   1665 	error = _bus_dmamem_alloc(t, cookie->id_bouncebuflen,
   1666 	    PAGE_SIZE, map->_dm_boundary, cookie->id_bouncesegs,
   1667 	    map->_dm_segcnt, &cookie->id_nbouncesegs, flags);
   1668 	if (error == 0) {
   1669 		error = _bus_dmamem_map(t, cookie->id_bouncesegs,
   1670 		    cookie->id_nbouncesegs, cookie->id_bouncebuflen,
   1671 		    (void **)&cookie->id_bouncebuf, flags);
   1672 		if (error) {
   1673 			_bus_dmamem_free(t, cookie->id_bouncesegs,
   1674 			    cookie->id_nbouncesegs);
   1675 			cookie->id_bouncebuflen = 0;
   1676 			cookie->id_nbouncesegs = 0;
   1677 		} else {
   1678 			cookie->id_flags |= _BUS_DMA_HAS_BOUNCE;
   1679 		}
   1680 	} else {
   1681 		cookie->id_bouncebuflen = 0;
   1682 		cookie->id_nbouncesegs = 0;
   1683 	}
   1684 
   1685 	return (error);
   1686 }
   1687 
   1688 static void
   1689 _bus_dma_free_bouncebuf(bus_dma_tag_t t, bus_dmamap_t map)
   1690 {
   1691 	struct arm32_bus_dma_cookie *cookie = map->_dm_cookie;
   1692 
   1693 #ifdef DIAGNOSTIC
   1694 	if (cookie == NULL)
   1695 		panic("_bus_dma_alloc_bouncebuf: no cookie");
   1696 #endif
   1697 
   1698 	_bus_dmamem_unmap(t, cookie->id_bouncebuf, cookie->id_bouncebuflen);
   1699 	_bus_dmamem_free(t, cookie->id_bouncesegs,
   1700 	    cookie->id_nbouncesegs);
   1701 	cookie->id_bouncebuflen = 0;
   1702 	cookie->id_nbouncesegs = 0;
   1703 	cookie->id_flags &= ~_BUS_DMA_HAS_BOUNCE;
   1704 }
   1705 
   1706 /*
   1707  * This function does the same as uiomove, but takes an explicit
   1708  * direction, and does not update the uio structure.
   1709  */
   1710 static int
   1711 _bus_dma_uiomove(void *buf, struct uio *uio, size_t n, int direction)
   1712 {
   1713 	struct iovec *iov;
   1714 	int error;
   1715 	struct vmspace *vm;
   1716 	char *cp;
   1717 	size_t resid, cnt;
   1718 	int i;
   1719 
   1720 	iov = uio->uio_iov;
   1721 	vm = uio->uio_vmspace;
   1722 	cp = buf;
   1723 	resid = n;
   1724 
   1725 	for (i = 0; i < uio->uio_iovcnt && resid > 0; i++) {
   1726 		iov = &uio->uio_iov[i];
   1727 		if (iov->iov_len == 0)
   1728 			continue;
   1729 		cnt = MIN(resid, iov->iov_len);
   1730 
   1731 		if (!VMSPACE_IS_KERNEL_P(vm) &&
   1732 		    (curlwp->l_cpu->ci_schedstate.spc_flags & SPCF_SHOULDYIELD)
   1733 		    != 0) {
   1734 			preempt();
   1735 		}
   1736 		if (direction == UIO_READ) {
   1737 			error = copyout_vmspace(vm, cp, iov->iov_base, cnt);
   1738 		} else {
   1739 			error = copyin_vmspace(vm, iov->iov_base, cp, cnt);
   1740 		}
   1741 		if (error)
   1742 			return (error);
   1743 		cp += cnt;
   1744 		resid -= cnt;
   1745 	}
   1746 	return (0);
   1747 }
   1748 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
   1749 
   1750 int
   1751 _bus_dmatag_subregion(bus_dma_tag_t tag, bus_addr_t min_addr,
   1752     bus_addr_t max_addr, bus_dma_tag_t *newtag, int flags)
   1753 {
   1754 
   1755 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1756 	struct arm32_dma_range *dr;
   1757 	bool subset = false;
   1758 	size_t nranges = 0;
   1759 	size_t i;
   1760 	for (i = 0, dr = tag->_ranges; i < tag->_nranges; i++, dr++) {
   1761 		if (dr->dr_sysbase <= min_addr
   1762 		    && max_addr <= dr->dr_sysbase + dr->dr_len - 1) {
   1763 			subset = true;
   1764 		}
   1765 		if (min_addr <= dr->dr_sysbase + dr->dr_len
   1766 		    && max_addr >= dr->dr_sysbase) {
   1767 			nranges++;
   1768 		}
   1769 	}
   1770 	if (subset) {
   1771 		*newtag = tag;
   1772 		/* if the tag must be freed, add a reference */
   1773 		if (tag->_tag_needs_free)
   1774 			(tag->_tag_needs_free)++;
   1775 		return 0;
   1776 	}
   1777 	if (nranges == 0) {
   1778 		nranges = 1;
   1779 	}
   1780 
   1781 	size_t mallocsize = sizeof(*tag) + nranges * sizeof(*dr);
   1782 	if ((*newtag = malloc(mallocsize, M_DMAMAP,
   1783 	    (flags & BUS_DMA_NOWAIT) ? M_NOWAIT : M_WAITOK)) == NULL)
   1784 		return ENOMEM;
   1785 
   1786 	dr = (void *)(*newtag + 1);
   1787 	**newtag = *tag;
   1788 	(*newtag)->_tag_needs_free = 1;
   1789 	(*newtag)->_ranges = dr;
   1790 	(*newtag)->_nranges = nranges;
   1791 
   1792 	if (tag->_ranges == NULL) {
   1793 		dr->dr_sysbase = min_addr;
   1794 		dr->dr_busbase = min_addr;
   1795 		dr->dr_len = max_addr + 1 - min_addr;
   1796 	} else {
   1797 		for (i = 0; i < nranges; i++) {
   1798 			if (min_addr > dr->dr_sysbase + dr->dr_len
   1799 			    || max_addr < dr->dr_sysbase)
   1800 				continue;
   1801 			dr[0] = tag->_ranges[i];
   1802 			if (dr->dr_sysbase < min_addr) {
   1803 				psize_t diff = min_addr - dr->dr_sysbase;
   1804 				dr->dr_busbase += diff;
   1805 				dr->dr_len -= diff;
   1806 				dr->dr_sysbase += diff;
   1807 			}
   1808 			if (max_addr != 0xffffffff
   1809 			    && max_addr + 1 < dr->dr_sysbase + dr->dr_len) {
   1810 				dr->dr_len = max_addr + 1 - dr->dr_sysbase;
   1811 			}
   1812 			dr++;
   1813 		}
   1814 	}
   1815 
   1816 	return 0;
   1817 #else
   1818 	return EOPNOTSUPP;
   1819 #endif /* _ARM32_NEED_BUS_DMA_BOUNCE */
   1820 }
   1821 
   1822 void
   1823 _bus_dmatag_destroy(bus_dma_tag_t tag)
   1824 {
   1825 #ifdef _ARM32_NEED_BUS_DMA_BOUNCE
   1826 	switch (tag->_tag_needs_free) {
   1827 	case 0:
   1828 		break;				/* not allocated with malloc */
   1829 	case 1:
   1830 		free(tag, M_DMAMAP);		/* last reference to tag */
   1831 		break;
   1832 	default:
   1833 		(tag->_tag_needs_free)--;	/* one less reference */
   1834 	}
   1835 #endif
   1836 }
   1837