2 * Copyright (c) 1997, 1998 Justin T. Gibbs.
3 * Copyright (c) 2015-2016 The FreeBSD Foundation
6 * Portions of this software were developed by Andrew Turner
7 * under sponsorship of the FreeBSD Foundation.
9 * Portions of this software were developed by Semihalf
10 * under sponsorship of the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions, and the following disclaimer,
17 * without modification, immediately at the beginning of the file.
18 * 2. The name of the author may not be used to endorse or promote products
19 * derived from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
25 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/malloc.h>
41 #include <sys/interrupt.h>
42 #include <sys/kernel.h>
46 #include <sys/memdesc.h>
47 #include <sys/mutex.h>
48 #include <sys/sysctl.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_kern.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_map.h>
57 #include <machine/atomic.h>
58 #include <machine/bus.h>
59 #include <machine/md_var.h>
60 #include <arm64/include/bus_dma_impl.h>
62 #define MAX_BPAGES 4096
65 BF_COULD_BOUNCE = 0x01,
66 BF_MIN_ALLOC_COMP = 0x02,
74 struct bus_dma_tag_common common;
77 bus_dma_segment_t *segments;
78 struct bounce_zone *bounce_zone;
82 vm_offset_t vaddr; /* kva of bounce buffer */
83 bus_addr_t busaddr; /* Physical address */
84 vm_offset_t datavaddr; /* kva of client data */
85 vm_page_t datapage; /* physical page of client data */
86 vm_offset_t dataoffs; /* page offset of client data */
87 bus_size_t datacount; /* client data count */
88 STAILQ_ENTRY(bounce_page) links;
91 int busdma_swi_pending;
94 STAILQ_ENTRY(bounce_zone) links;
95 STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
103 bus_size_t alignment;
107 struct sysctl_ctx_list sysctl_tree;
108 struct sysctl_oid *sysctl_tree_top;
111 static struct mtx bounce_lock;
112 static int total_bpages;
113 static int busdma_zonecount;
114 static STAILQ_HEAD(, bounce_zone) bounce_zone_list;
116 static SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
117 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
118 "Total bounce pages");
121 vm_offset_t vaddr; /* kva of client data */
122 bus_addr_t paddr; /* physical address */
123 vm_page_t pages; /* starting page of client data */
124 bus_size_t datacount; /* client data count */
128 struct bp_list bpages;
133 bus_dmamap_callback_t *callback;
135 STAILQ_ENTRY(bus_dmamap) links;
137 #define DMAMAP_COULD_BOUNCE (1 << 0)
138 #define DMAMAP_FROM_DMAMEM (1 << 1)
140 struct sync_list slist[];
143 static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
144 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
146 static void init_bounce_pages(void *dummy);
147 static int alloc_bounce_zone(bus_dma_tag_t dmat);
148 static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
149 static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
151 static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
152 vm_offset_t vaddr, bus_addr_t addr, bus_size_t size);
153 static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
154 int run_filter(bus_dma_tag_t dmat, bus_addr_t paddr);
155 static bool _bus_dmamap_pagesneeded(bus_dma_tag_t dmat, vm_paddr_t buf,
156 bus_size_t buflen, int *pagesneeded);
157 static void _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
158 pmap_t pmap, void *buf, bus_size_t buflen, int flags);
159 static void _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
160 vm_paddr_t buf, bus_size_t buflen, int flags);
161 static int _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
165 * Allocate a device specific dma_tag.
168 bounce_bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
169 bus_addr_t boundary, bus_addr_t lowaddr, bus_addr_t highaddr,
170 bus_dma_filter_t *filter, void *filterarg, bus_size_t maxsize,
171 int nsegments, bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
172 void *lockfuncarg, bus_dma_tag_t *dmat)
174 bus_dma_tag_t newtag;
178 error = common_bus_dma_tag_create(parent != NULL ? &parent->common :
179 NULL, alignment, boundary, lowaddr, highaddr, filter, filterarg,
180 maxsize, nsegments, maxsegsz, flags, lockfunc, lockfuncarg,
181 sizeof (struct bus_dma_tag), (void **)&newtag);
185 newtag->common.impl = &bus_dma_bounce_impl;
186 newtag->map_count = 0;
187 newtag->segments = NULL;
189 if ((flags & BUS_DMA_COHERENT) != 0)
190 newtag->bounce_flags |= BF_COHERENT;
192 if (parent != NULL) {
193 if ((newtag->common.filter != NULL ||
194 (parent->bounce_flags & BF_COULD_BOUNCE) != 0))
195 newtag->bounce_flags |= BF_COULD_BOUNCE;
197 /* Copy some flags from the parent */
198 newtag->bounce_flags |= parent->bounce_flags & BF_COHERENT;
201 if (newtag->common.lowaddr < ptoa((vm_paddr_t)Maxmem) ||
202 newtag->common.alignment > 1)
203 newtag->bounce_flags |= BF_COULD_BOUNCE;
205 if (((newtag->bounce_flags & BF_COULD_BOUNCE) != 0) &&
206 (flags & BUS_DMA_ALLOCNOW) != 0) {
207 struct bounce_zone *bz;
210 if ((error = alloc_bounce_zone(newtag)) != 0) {
211 free(newtag, M_DEVBUF);
214 bz = newtag->bounce_zone;
216 if (ptoa(bz->total_bpages) < maxsize) {
219 pages = atop(round_page(maxsize)) - bz->total_bpages;
221 /* Add pages to our bounce pool */
222 if (alloc_bounce_pages(newtag, pages) < pages)
225 /* Performed initial allocation */
226 newtag->bounce_flags |= BF_MIN_ALLOC_COMP;
231 free(newtag, M_DEVBUF);
234 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
235 __func__, newtag, (newtag != NULL ? newtag->common.flags : 0),
241 bounce_bus_dma_tag_destroy(bus_dma_tag_t dmat)
243 bus_dma_tag_t dmat_copy, parent;
250 if (dmat->map_count != 0) {
254 while (dmat != NULL) {
255 parent = (bus_dma_tag_t)dmat->common.parent;
256 atomic_subtract_int(&dmat->common.ref_count, 1);
257 if (dmat->common.ref_count == 0) {
258 if (dmat->segments != NULL)
259 free(dmat->segments, M_DEVBUF);
260 free(dmat, M_DEVBUF);
262 * Last reference count, so
263 * release our reference
264 * count on our parent.
272 CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
277 bounce_bus_dma_id_mapped(bus_dma_tag_t dmat, vm_paddr_t buf, bus_size_t buflen)
280 if ((dmat->bounce_flags & BF_COULD_BOUNCE) == 0)
282 return (!_bus_dmamap_pagesneeded(dmat, buf, buflen, NULL));
286 alloc_dmamap(bus_dma_tag_t dmat, int flags)
291 mapsize = sizeof(*map);
292 mapsize += sizeof(struct sync_list) * dmat->common.nsegments;
293 map = malloc(mapsize, M_DEVBUF, flags | M_ZERO);
297 /* Initialize the new map */
298 STAILQ_INIT(&map->bpages);
304 * Allocate a handle for mapping from kva/uva/physical
305 * address space into bus device space.
308 bounce_bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
310 struct bounce_zone *bz;
311 int error, maxpages, pages;
315 if (dmat->segments == NULL) {
316 dmat->segments = (bus_dma_segment_t *)malloc(
317 sizeof(bus_dma_segment_t) * dmat->common.nsegments,
319 if (dmat->segments == NULL) {
320 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
321 __func__, dmat, ENOMEM);
326 *mapp = alloc_dmamap(dmat, M_NOWAIT);
328 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
329 __func__, dmat, ENOMEM);
334 * Bouncing might be required if the driver asks for an active
335 * exclusion region, a data alignment that is stricter than 1, and/or
336 * an active address boundary.
338 if (dmat->bounce_flags & BF_COULD_BOUNCE) {
340 if (dmat->bounce_zone == NULL) {
341 if ((error = alloc_bounce_zone(dmat)) != 0) {
342 free(*mapp, M_DEVBUF);
346 bz = dmat->bounce_zone;
348 (*mapp)->flags = DMAMAP_COULD_BOUNCE;
351 * Attempt to add pages to our pool on a per-instance
352 * basis up to a sane limit.
354 if (dmat->common.alignment > 1)
355 maxpages = MAX_BPAGES;
357 maxpages = MIN(MAX_BPAGES, Maxmem -
358 atop(dmat->common.lowaddr));
359 if ((dmat->bounce_flags & BF_MIN_ALLOC_COMP) == 0 ||
360 (bz->map_count > 0 && bz->total_bpages < maxpages)) {
361 pages = MAX(atop(dmat->common.maxsize), 1);
362 pages = MIN(maxpages - bz->total_bpages, pages);
363 pages = MAX(pages, 1);
364 if (alloc_bounce_pages(dmat, pages) < pages)
366 if ((dmat->bounce_flags & BF_MIN_ALLOC_COMP)
369 dmat->bounce_flags |=
380 free(*mapp, M_DEVBUF);
381 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
382 __func__, dmat, dmat->common.flags, error);
387 * Destroy a handle for mapping from kva/uva/physical
388 * address space into bus device space.
391 bounce_bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
394 /* Check we are destroying the correct map type */
395 if ((map->flags & DMAMAP_FROM_DMAMEM) != 0)
396 panic("bounce_bus_dmamap_destroy: Invalid map freed\n");
398 if (STAILQ_FIRST(&map->bpages) != NULL || map->sync_count != 0) {
399 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, EBUSY);
402 if (dmat->bounce_zone) {
403 KASSERT((map->flags & DMAMAP_COULD_BOUNCE) != 0,
404 ("%s: Bounce zone when cannot bounce", __func__));
405 dmat->bounce_zone->map_count--;
409 CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
415 * Allocate a piece of memory that can be efficiently mapped into
416 * bus device space based on the constraints lited in the dma tag.
417 * A dmamap to for use with dmamap_load is also allocated.
420 bounce_bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
425 * This bus_dma implementation requires IO-Coherent architecutre.
426 * If IO-Coherency is not guaranteed, the BUS_DMA_COHERENT flag has
427 * to be implented using non-cacheable memory.
433 if (flags & BUS_DMA_NOWAIT)
438 if (dmat->segments == NULL) {
439 dmat->segments = (bus_dma_segment_t *)malloc(
440 sizeof(bus_dma_segment_t) * dmat->common.nsegments,
442 if (dmat->segments == NULL) {
443 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
444 __func__, dmat, dmat->common.flags, ENOMEM);
448 if (flags & BUS_DMA_ZERO)
450 if (flags & BUS_DMA_NOCACHE)
451 attr = VM_MEMATTR_UNCACHEABLE;
452 else if ((flags & BUS_DMA_COHERENT) != 0 &&
453 (dmat->bounce_flags & BF_COHERENT) == 0)
455 * If we have a non-coherent tag, and are trying to allocate
456 * a coherent block of memory it needs to be uncached.
458 attr = VM_MEMATTR_UNCACHEABLE;
460 attr = VM_MEMATTR_DEFAULT;
463 * Create the map, but don't set the could bounce flag as
464 * this allocation should never bounce;
466 *mapp = alloc_dmamap(dmat, mflags);
468 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
469 __func__, dmat, dmat->common.flags, ENOMEM);
472 (*mapp)->flags = DMAMAP_FROM_DMAMEM;
475 * Allocate the buffer from the malloc(9) allocator if...
476 * - It's small enough to fit into a single power of two sized bucket.
477 * - The alignment is less than or equal to the maximum size
478 * - The low address requirement is fulfilled.
479 * else allocate non-contiguous pages if...
480 * - The page count that could get allocated doesn't exceed
481 * nsegments also when the maximum segment size is less
483 * - The alignment constraint isn't larger than a page boundary.
484 * - There are no boundary-crossing constraints.
485 * else allocate a block of contiguous pages because one or more of the
486 * constraints is something that only the contig allocator can fulfill.
488 * NOTE: The (dmat->common.alignment <= dmat->maxsize) check
489 * below is just a quick hack. The exact alignment guarantees
490 * of malloc(9) need to be nailed down, and the code below
491 * should be rewritten to take that into account.
493 * In the meantime warn the user if malloc gets it wrong.
495 if ((dmat->common.maxsize <= PAGE_SIZE) &&
496 (dmat->common.alignment <= dmat->common.maxsize) &&
497 dmat->common.lowaddr >= ptoa((vm_paddr_t)Maxmem) &&
498 attr == VM_MEMATTR_DEFAULT) {
499 *vaddr = malloc(dmat->common.maxsize, M_DEVBUF, mflags);
500 } else if (dmat->common.nsegments >=
501 howmany(dmat->common.maxsize, MIN(dmat->common.maxsegsz, PAGE_SIZE)) &&
502 dmat->common.alignment <= PAGE_SIZE &&
503 (dmat->common.boundary % PAGE_SIZE) == 0) {
504 /* Page-based multi-segment allocations allowed */
505 *vaddr = (void *)kmem_alloc_attr(dmat->common.maxsize, mflags,
506 0ul, dmat->common.lowaddr, attr);
507 dmat->bounce_flags |= BF_KMEM_ALLOC;
509 *vaddr = (void *)kmem_alloc_contig(dmat->common.maxsize, mflags,
510 0ul, dmat->common.lowaddr, dmat->common.alignment != 0 ?
511 dmat->common.alignment : 1ul, dmat->common.boundary, attr);
512 dmat->bounce_flags |= BF_KMEM_ALLOC;
514 if (*vaddr == NULL) {
515 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
516 __func__, dmat, dmat->common.flags, ENOMEM);
517 free(*mapp, M_DEVBUF);
519 } else if (vtophys(*vaddr) & (dmat->common.alignment - 1)) {
520 printf("bus_dmamem_alloc failed to align memory properly.\n");
523 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
524 __func__, dmat, dmat->common.flags, 0);
529 * Free a piece of memory and it's allociated dmamap, that was allocated
530 * via bus_dmamem_alloc. Make the same choice for free/contigfree.
533 bounce_bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
537 * Check the map came from bounce_bus_dmamem_alloc, so the map
538 * should be NULL and the BF_KMEM_ALLOC flag cleared if malloc()
539 * was used and set if kmem_alloc_contig() was used.
541 if ((map->flags & DMAMAP_FROM_DMAMEM) == 0)
542 panic("bus_dmamem_free: Invalid map freed\n");
543 if ((dmat->bounce_flags & BF_KMEM_ALLOC) == 0)
544 free(vaddr, M_DEVBUF);
546 kmem_free((vm_offset_t)vaddr, dmat->common.maxsize);
549 CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat,
554 _bus_dmamap_pagesneeded(bus_dma_tag_t dmat, vm_paddr_t buf, bus_size_t buflen,
562 * Count the number of bounce pages needed in order to
563 * complete this transfer
567 while (buflen != 0) {
568 sgsize = MIN(buflen, dmat->common.maxsegsz);
569 if (bus_dma_run_filter(&dmat->common, curaddr)) {
571 PAGE_SIZE - (curaddr & PAGE_MASK));
572 if (pagesneeded == NULL)
580 if (pagesneeded != NULL)
581 *pagesneeded = count;
586 _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
587 bus_size_t buflen, int flags)
590 if ((map->flags & DMAMAP_COULD_BOUNCE) != 0 && map->pagesneeded == 0) {
591 _bus_dmamap_pagesneeded(dmat, buf, buflen, &map->pagesneeded);
592 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
597 _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, pmap_t pmap,
598 void *buf, bus_size_t buflen, int flags)
601 vm_offset_t vendaddr;
605 if ((map->flags & DMAMAP_COULD_BOUNCE) != 0 && map->pagesneeded == 0) {
606 CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
607 "alignment= %d", dmat->common.lowaddr,
608 ptoa((vm_paddr_t)Maxmem),
609 dmat->common.boundary, dmat->common.alignment);
610 CTR2(KTR_BUSDMA, "map= %p, pagesneeded= %d", map,
613 * Count the number of bounce pages
614 * needed in order to complete this transfer
616 vaddr = (vm_offset_t)buf;
617 vendaddr = (vm_offset_t)buf + buflen;
619 while (vaddr < vendaddr) {
620 sg_len = PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK);
621 if (pmap == kernel_pmap)
622 paddr = pmap_kextract(vaddr);
624 paddr = pmap_extract(pmap, vaddr);
625 if (bus_dma_run_filter(&dmat->common, paddr) != 0) {
626 sg_len = roundup2(sg_len,
627 dmat->common.alignment);
632 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
637 _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags)
640 /* Reserve Necessary Bounce Pages */
641 mtx_lock(&bounce_lock);
642 if (flags & BUS_DMA_NOWAIT) {
643 if (reserve_bounce_pages(dmat, map, 0) != 0) {
644 mtx_unlock(&bounce_lock);
648 if (reserve_bounce_pages(dmat, map, 1) != 0) {
649 /* Queue us for resources */
650 STAILQ_INSERT_TAIL(&bounce_map_waitinglist, map, links);
651 mtx_unlock(&bounce_lock);
652 return (EINPROGRESS);
655 mtx_unlock(&bounce_lock);
661 * Add a single contiguous physical range to the segment list.
664 _bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr,
665 bus_size_t sgsize, bus_dma_segment_t *segs, int *segp)
667 bus_addr_t baddr, bmask;
671 * Make sure we don't cross any boundaries.
673 bmask = ~(dmat->common.boundary - 1);
674 if (dmat->common.boundary > 0) {
675 baddr = (curaddr + dmat->common.boundary) & bmask;
676 if (sgsize > (baddr - curaddr))
677 sgsize = (baddr - curaddr);
681 * Insert chunk into a segment, coalescing with
682 * previous segment if possible.
687 segs[seg].ds_addr = curaddr;
688 segs[seg].ds_len = sgsize;
690 if (curaddr == segs[seg].ds_addr + segs[seg].ds_len &&
691 (segs[seg].ds_len + sgsize) <= dmat->common.maxsegsz &&
692 (dmat->common.boundary == 0 ||
693 (segs[seg].ds_addr & bmask) == (curaddr & bmask)))
694 segs[seg].ds_len += sgsize;
696 if (++seg >= dmat->common.nsegments)
698 segs[seg].ds_addr = curaddr;
699 segs[seg].ds_len = sgsize;
707 * Utility function to load a physical buffer. segp contains
708 * the starting segment on entrace, and the ending segment on exit.
711 bounce_bus_dmamap_load_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
712 vm_paddr_t buf, bus_size_t buflen, int flags, bus_dma_segment_t *segs,
715 struct sync_list *sl;
717 bus_addr_t curaddr, sl_end;
721 segs = dmat->segments;
723 if ((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) {
724 _bus_dmamap_count_phys(dmat, map, buf, buflen, flags);
725 if (map->pagesneeded != 0) {
726 error = _bus_dmamap_reserve_pages(dmat, map, flags);
732 sl = map->slist + map->sync_count - 1;
737 sgsize = MIN(buflen, dmat->common.maxsegsz);
738 if (((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) &&
739 map->pagesneeded != 0 &&
740 bus_dma_run_filter(&dmat->common, curaddr)) {
741 sgsize = MIN(sgsize, PAGE_SIZE - (curaddr & PAGE_MASK));
742 curaddr = add_bounce_page(dmat, map, 0, curaddr,
744 } else if ((dmat->bounce_flags & BF_COHERENT) == 0) {
745 if (map->sync_count > 0)
746 sl_end = sl->paddr + sl->datacount;
748 if (map->sync_count == 0 || curaddr != sl_end) {
749 if (++map->sync_count > dmat->common.nsegments)
754 sl->datacount = sgsize;
755 sl->pages = PHYS_TO_VM_PAGE(curaddr);
756 KASSERT(sl->pages != NULL,
757 ("%s: page at PA:0x%08lx is not in "
758 "vm_page_array", __func__, curaddr));
760 sl->datacount += sgsize;
762 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
773 return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
777 * Utility function to load a linear buffer. segp contains
778 * the starting segment on entrace, and the ending segment on exit.
781 bounce_bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
782 bus_size_t buflen, pmap_t pmap, int flags, bus_dma_segment_t *segs,
785 struct sync_list *sl;
786 bus_size_t sgsize, max_sgsize;
787 bus_addr_t curaddr, sl_pend;
788 vm_offset_t kvaddr, vaddr, sl_vend;
792 segs = dmat->segments;
794 if ((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) {
795 _bus_dmamap_count_pages(dmat, map, pmap, buf, buflen, flags);
796 if (map->pagesneeded != 0) {
797 error = _bus_dmamap_reserve_pages(dmat, map, flags);
803 sl = map->slist + map->sync_count - 1;
804 vaddr = (vm_offset_t)buf;
810 * Get the physical address for this segment.
812 if (pmap == kernel_pmap) {
813 curaddr = pmap_kextract(vaddr);
816 curaddr = pmap_extract(pmap, vaddr);
821 * Compute the segment size, and adjust counts.
823 max_sgsize = MIN(buflen, dmat->common.maxsegsz);
824 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
825 if (((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) &&
826 map->pagesneeded != 0 &&
827 bus_dma_run_filter(&dmat->common, curaddr)) {
828 sgsize = roundup2(sgsize, dmat->common.alignment);
829 sgsize = MIN(sgsize, max_sgsize);
830 curaddr = add_bounce_page(dmat, map, kvaddr, curaddr,
832 } else if ((dmat->bounce_flags & BF_COHERENT) == 0) {
833 sgsize = MIN(sgsize, max_sgsize);
834 if (map->sync_count > 0) {
835 sl_pend = sl->paddr + sl->datacount;
836 sl_vend = sl->vaddr + sl->datacount;
839 if (map->sync_count == 0 ||
840 (kvaddr != 0 && kvaddr != sl_vend) ||
841 (curaddr != sl_pend)) {
843 if (++map->sync_count > dmat->common.nsegments)
851 sl->pages = PHYS_TO_VM_PAGE(curaddr);
852 KASSERT(sl->pages != NULL,
853 ("%s: page at PA:0x%08lx is not "
854 "in vm_page_array", __func__,
857 sl->datacount = sgsize;
859 sl->datacount += sgsize;
861 sgsize = MIN(sgsize, max_sgsize);
863 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
875 return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
879 bounce_bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map,
880 struct memdesc *mem, bus_dmamap_callback_t *callback, void *callback_arg)
883 if ((map->flags & DMAMAP_COULD_BOUNCE) == 0)
887 map->callback = callback;
888 map->callback_arg = callback_arg;
891 static bus_dma_segment_t *
892 bounce_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map,
893 bus_dma_segment_t *segs, int nsegs, int error)
897 segs = dmat->segments;
902 * Release the mapping held by map.
905 bounce_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
907 struct bounce_page *bpage;
909 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
910 STAILQ_REMOVE_HEAD(&map->bpages, links);
911 free_bounce_page(dmat, bpage);
918 dma_preread_safe(vm_offset_t va, vm_size_t size)
921 * Write back any partial cachelines immediately before and
922 * after the DMA region.
924 if (va & (dcache_line_size - 1))
925 cpu_dcache_wb_range(va, 1);
926 if ((va + size) & (dcache_line_size - 1))
927 cpu_dcache_wb_range(va + size, 1);
929 cpu_dcache_inv_range(va, size);
933 dma_dcache_sync(struct sync_list *sl, bus_dmasync_op_t op)
935 uint32_t len, offset;
938 vm_offset_t va, tempva;
941 offset = sl->paddr & PAGE_MASK;
943 size = sl->datacount;
946 for ( ; size != 0; size -= len, pa += len, offset = 0, ++m) {
948 if (sl->vaddr == 0) {
949 len = min(PAGE_SIZE - offset, size);
950 tempva = pmap_quick_enter_page(m);
951 va = tempva | offset;
952 KASSERT(pa == (VM_PAGE_TO_PHYS(m) | offset),
953 ("unexpected vm_page_t phys: 0x%16lx != 0x%16lx",
954 VM_PAGE_TO_PHYS(m) | offset, pa));
961 case BUS_DMASYNC_PREWRITE:
962 case BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD:
963 cpu_dcache_wb_range(va, len);
965 case BUS_DMASYNC_PREREAD:
967 * An mbuf may start in the middle of a cacheline. There
968 * will be no cpu writes to the beginning of that line
969 * (which contains the mbuf header) while dma is in
970 * progress. Handle that case by doing a writeback of
971 * just the first cacheline before invalidating the
972 * overall buffer. Any mbuf in a chain may have this
973 * misalignment. Buffers which are not mbufs bounce if
974 * they are not aligned to a cacheline.
976 dma_preread_safe(va, len);
978 case BUS_DMASYNC_POSTREAD:
979 case BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE:
980 cpu_dcache_inv_range(va, len);
983 panic("unsupported combination of sync operations: "
988 pmap_quick_remove_page(tempva);
993 bounce_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map,
996 struct bounce_page *bpage;
997 struct sync_list *sl, *end;
998 vm_offset_t datavaddr, tempvaddr;
1000 if (op == BUS_DMASYNC_POSTWRITE)
1003 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
1005 * Wait for any DMA operations to complete before the bcopy.
1010 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1011 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
1012 "performing bounce", __func__, dmat, dmat->common.flags,
1015 if ((op & BUS_DMASYNC_PREWRITE) != 0) {
1016 while (bpage != NULL) {
1018 datavaddr = bpage->datavaddr;
1019 if (datavaddr == 0) {
1020 tempvaddr = pmap_quick_enter_page(
1022 datavaddr = tempvaddr | bpage->dataoffs;
1025 bcopy((void *)datavaddr,
1026 (void *)bpage->vaddr, bpage->datacount);
1028 pmap_quick_remove_page(tempvaddr);
1029 if ((dmat->bounce_flags & BF_COHERENT) == 0)
1030 cpu_dcache_wb_range(bpage->vaddr,
1032 bpage = STAILQ_NEXT(bpage, links);
1034 dmat->bounce_zone->total_bounced++;
1035 } else if ((op & BUS_DMASYNC_PREREAD) != 0) {
1036 while (bpage != NULL) {
1037 if ((dmat->bounce_flags & BF_COHERENT) == 0)
1038 cpu_dcache_wbinv_range(bpage->vaddr,
1040 bpage = STAILQ_NEXT(bpage, links);
1044 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
1045 while (bpage != NULL) {
1046 if ((dmat->bounce_flags & BF_COHERENT) == 0)
1047 cpu_dcache_inv_range(bpage->vaddr,
1050 datavaddr = bpage->datavaddr;
1051 if (datavaddr == 0) {
1052 tempvaddr = pmap_quick_enter_page(
1054 datavaddr = tempvaddr | bpage->dataoffs;
1057 bcopy((void *)bpage->vaddr,
1058 (void *)datavaddr, bpage->datacount);
1061 pmap_quick_remove_page(tempvaddr);
1062 bpage = STAILQ_NEXT(bpage, links);
1064 dmat->bounce_zone->total_bounced++;
1069 * Cache maintenance for normal (non-COHERENT non-bounce) buffers.
1071 if (map->sync_count != 0) {
1072 sl = &map->slist[0];
1073 end = &map->slist[map->sync_count];
1074 CTR3(KTR_BUSDMA, "%s: tag %p op 0x%x "
1075 "performing sync", __func__, dmat, op);
1077 for ( ; sl != end; ++sl)
1078 dma_dcache_sync(sl, op);
1081 if ((op & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) != 0) {
1083 * Wait for the bcopy to complete before any DMA operations.
1090 init_bounce_pages(void *dummy __unused)
1094 STAILQ_INIT(&bounce_zone_list);
1095 STAILQ_INIT(&bounce_map_waitinglist);
1096 STAILQ_INIT(&bounce_map_callbacklist);
1097 mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
1099 SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
1101 static struct sysctl_ctx_list *
1102 busdma_sysctl_tree(struct bounce_zone *bz)
1105 return (&bz->sysctl_tree);
1108 static struct sysctl_oid *
1109 busdma_sysctl_tree_top(struct bounce_zone *bz)
1112 return (bz->sysctl_tree_top);
1116 alloc_bounce_zone(bus_dma_tag_t dmat)
1118 struct bounce_zone *bz;
1120 /* Check to see if we already have a suitable zone */
1121 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1122 if ((dmat->common.alignment <= bz->alignment) &&
1123 (dmat->common.lowaddr >= bz->lowaddr)) {
1124 dmat->bounce_zone = bz;
1129 if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
1130 M_NOWAIT | M_ZERO)) == NULL)
1133 STAILQ_INIT(&bz->bounce_page_list);
1134 bz->free_bpages = 0;
1135 bz->reserved_bpages = 0;
1136 bz->active_bpages = 0;
1137 bz->lowaddr = dmat->common.lowaddr;
1138 bz->alignment = MAX(dmat->common.alignment, PAGE_SIZE);
1140 snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1142 snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1143 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1144 dmat->bounce_zone = bz;
1146 sysctl_ctx_init(&bz->sysctl_tree);
1147 bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
1148 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1150 if (bz->sysctl_tree_top == NULL) {
1151 sysctl_ctx_free(&bz->sysctl_tree);
1152 return (0); /* XXX error code? */
1155 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1156 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1157 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1158 "Total bounce pages");
1159 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1160 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1161 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1162 "Free bounce pages");
1163 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1164 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1165 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1166 "Reserved bounce pages");
1167 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1168 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1169 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1170 "Active bounce pages");
1171 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1172 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1173 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1174 "Total bounce requests");
1175 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1176 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1177 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1178 "Total bounce requests that were deferred");
1179 SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
1180 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1181 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1182 SYSCTL_ADD_UAUTO(busdma_sysctl_tree(bz),
1183 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1184 "alignment", CTLFLAG_RD, &bz->alignment, "");
1190 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
1192 struct bounce_zone *bz;
1195 bz = dmat->bounce_zone;
1197 while (numpages > 0) {
1198 struct bounce_page *bpage;
1200 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
1205 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1206 M_NOWAIT, 0ul, bz->lowaddr, PAGE_SIZE, 0);
1207 if (bpage->vaddr == 0) {
1208 free(bpage, M_DEVBUF);
1211 bpage->busaddr = pmap_kextract(bpage->vaddr);
1212 mtx_lock(&bounce_lock);
1213 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1217 mtx_unlock(&bounce_lock);
1225 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1227 struct bounce_zone *bz;
1230 mtx_assert(&bounce_lock, MA_OWNED);
1231 bz = dmat->bounce_zone;
1232 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1233 if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
1234 return (map->pagesneeded - (map->pagesreserved + pages));
1235 bz->free_bpages -= pages;
1236 bz->reserved_bpages += pages;
1237 map->pagesreserved += pages;
1238 pages = map->pagesneeded - map->pagesreserved;
1244 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1245 bus_addr_t addr, bus_size_t size)
1247 struct bounce_zone *bz;
1248 struct bounce_page *bpage;
1250 KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
1251 KASSERT((map->flags & DMAMAP_COULD_BOUNCE) != 0,
1252 ("add_bounce_page: bad map %p", map));
1254 bz = dmat->bounce_zone;
1255 if (map->pagesneeded == 0)
1256 panic("add_bounce_page: map doesn't need any pages");
1259 if (map->pagesreserved == 0)
1260 panic("add_bounce_page: map doesn't need any pages");
1261 map->pagesreserved--;
1263 mtx_lock(&bounce_lock);
1264 bpage = STAILQ_FIRST(&bz->bounce_page_list);
1266 panic("add_bounce_page: free page list is empty");
1268 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1269 bz->reserved_bpages--;
1270 bz->active_bpages++;
1271 mtx_unlock(&bounce_lock);
1273 if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1274 /* Page offset needs to be preserved. */
1275 bpage->vaddr |= addr & PAGE_MASK;
1276 bpage->busaddr |= addr & PAGE_MASK;
1278 bpage->datavaddr = vaddr;
1279 bpage->datapage = PHYS_TO_VM_PAGE(addr);
1280 bpage->dataoffs = addr & PAGE_MASK;
1281 bpage->datacount = size;
1282 STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1283 return (bpage->busaddr);
1287 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1289 struct bus_dmamap *map;
1290 struct bounce_zone *bz;
1292 bz = dmat->bounce_zone;
1293 bpage->datavaddr = 0;
1294 bpage->datacount = 0;
1295 if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1297 * Reset the bounce page to start at offset 0. Other uses
1298 * of this bounce page may need to store a full page of
1299 * data and/or assume it starts on a page boundary.
1301 bpage->vaddr &= ~PAGE_MASK;
1302 bpage->busaddr &= ~PAGE_MASK;
1305 mtx_lock(&bounce_lock);
1306 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1308 bz->active_bpages--;
1309 if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1310 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1311 STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1312 STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1314 busdma_swi_pending = 1;
1315 bz->total_deferred++;
1316 swi_sched(vm_ih, 0);
1319 mtx_unlock(&bounce_lock);
1326 struct bus_dmamap *map;
1328 mtx_lock(&bounce_lock);
1329 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1330 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1331 mtx_unlock(&bounce_lock);
1333 (dmat->common.lockfunc)(dmat->common.lockfuncarg, BUS_DMA_LOCK);
1334 bus_dmamap_load_mem(map->dmat, map, &map->mem,
1335 map->callback, map->callback_arg, BUS_DMA_WAITOK);
1336 (dmat->common.lockfunc)(dmat->common.lockfuncarg,
1338 mtx_lock(&bounce_lock);
1340 mtx_unlock(&bounce_lock);
1343 struct bus_dma_impl bus_dma_bounce_impl = {
1344 .tag_create = bounce_bus_dma_tag_create,
1345 .tag_destroy = bounce_bus_dma_tag_destroy,
1346 .id_mapped = bounce_bus_dma_id_mapped,
1347 .map_create = bounce_bus_dmamap_create,
1348 .map_destroy = bounce_bus_dmamap_destroy,
1349 .mem_alloc = bounce_bus_dmamem_alloc,
1350 .mem_free = bounce_bus_dmamem_free,
1351 .load_phys = bounce_bus_dmamap_load_phys,
1352 .load_buffer = bounce_bus_dmamap_load_buffer,
1353 .load_ma = bus_dmamap_load_ma_triv,
1354 .map_waitok = bounce_bus_dmamap_waitok,
1355 .map_complete = bounce_bus_dmamap_complete,
1356 .map_unload = bounce_bus_dmamap_unload,
1357 .map_sync = bounce_bus_dmamap_sync