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 <machine/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 | CTLFLAG_MPSAFE, 0,
117 "Busdma parameters");
118 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
119 "Total bounce pages");
122 vm_offset_t vaddr; /* kva of client data */
123 bus_addr_t paddr; /* physical address */
124 vm_page_t pages; /* starting page of client data */
125 bus_size_t datacount; /* client data count */
129 struct bp_list bpages;
134 bus_dmamap_callback_t *callback;
136 STAILQ_ENTRY(bus_dmamap) links;
138 #define DMAMAP_COULD_BOUNCE (1 << 0)
139 #define DMAMAP_FROM_DMAMEM (1 << 1)
141 struct sync_list slist[];
144 static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
145 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
147 static void init_bounce_pages(void *dummy);
148 static int alloc_bounce_zone(bus_dma_tag_t dmat);
149 static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
150 static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
152 static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
153 vm_offset_t vaddr, bus_addr_t addr, bus_size_t size);
154 static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
155 int run_filter(bus_dma_tag_t dmat, bus_addr_t paddr);
156 static void _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
157 pmap_t pmap, void *buf, bus_size_t buflen, int flags);
158 static void _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
159 vm_paddr_t buf, bus_size_t buflen, int flags);
160 static int _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
164 * Allocate a device specific dma_tag.
167 bounce_bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
168 bus_addr_t boundary, bus_addr_t lowaddr, bus_addr_t highaddr,
169 bus_dma_filter_t *filter, void *filterarg, bus_size_t maxsize,
170 int nsegments, bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
171 void *lockfuncarg, bus_dma_tag_t *dmat)
173 bus_dma_tag_t newtag;
177 error = common_bus_dma_tag_create(parent != NULL ? &parent->common :
178 NULL, alignment, boundary, lowaddr, highaddr, filter, filterarg,
179 maxsize, nsegments, maxsegsz, flags, lockfunc, lockfuncarg,
180 sizeof (struct bus_dma_tag), (void **)&newtag);
184 newtag->common.impl = &bus_dma_bounce_impl;
185 newtag->map_count = 0;
186 newtag->segments = NULL;
188 if ((flags & BUS_DMA_COHERENT) != 0)
189 newtag->bounce_flags |= BF_COHERENT;
191 if (parent != NULL) {
192 if ((newtag->common.filter != NULL ||
193 (parent->bounce_flags & BF_COULD_BOUNCE) != 0))
194 newtag->bounce_flags |= BF_COULD_BOUNCE;
196 /* Copy some flags from the parent */
197 newtag->bounce_flags |= parent->bounce_flags & BF_COHERENT;
200 if (newtag->common.lowaddr < ptoa((vm_paddr_t)Maxmem) ||
201 newtag->common.alignment > 1)
202 newtag->bounce_flags |= BF_COULD_BOUNCE;
204 if (((newtag->bounce_flags & BF_COULD_BOUNCE) != 0) &&
205 (flags & BUS_DMA_ALLOCNOW) != 0) {
206 struct bounce_zone *bz;
209 if ((error = alloc_bounce_zone(newtag)) != 0) {
210 free(newtag, M_DEVBUF);
213 bz = newtag->bounce_zone;
215 if (ptoa(bz->total_bpages) < maxsize) {
218 pages = atop(round_page(maxsize)) - bz->total_bpages;
220 /* Add pages to our bounce pool */
221 if (alloc_bounce_pages(newtag, pages) < pages)
224 /* Performed initial allocation */
225 newtag->bounce_flags |= BF_MIN_ALLOC_COMP;
230 free(newtag, M_DEVBUF);
233 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
234 __func__, newtag, (newtag != NULL ? newtag->common.flags : 0),
240 bounce_bus_dma_tag_destroy(bus_dma_tag_t dmat)
242 bus_dma_tag_t dmat_copy, parent;
249 if (dmat->map_count != 0) {
253 while (dmat != NULL) {
254 parent = (bus_dma_tag_t)dmat->common.parent;
255 atomic_subtract_int(&dmat->common.ref_count, 1);
256 if (dmat->common.ref_count == 0) {
257 if (dmat->segments != NULL)
258 free(dmat->segments, M_DEVBUF);
259 free(dmat, M_DEVBUF);
261 * Last reference count, so
262 * release our reference
263 * count on our parent.
271 CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
276 alloc_dmamap(bus_dma_tag_t dmat, int flags)
281 mapsize = sizeof(*map);
282 mapsize += sizeof(struct sync_list) * dmat->common.nsegments;
283 map = malloc(mapsize, M_DEVBUF, flags | M_ZERO);
287 /* Initialize the new map */
288 STAILQ_INIT(&map->bpages);
294 * Allocate a handle for mapping from kva/uva/physical
295 * address space into bus device space.
298 bounce_bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
300 struct bounce_zone *bz;
301 int error, maxpages, pages;
305 if (dmat->segments == NULL) {
306 dmat->segments = (bus_dma_segment_t *)malloc(
307 sizeof(bus_dma_segment_t) * dmat->common.nsegments,
309 if (dmat->segments == NULL) {
310 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
311 __func__, dmat, ENOMEM);
316 *mapp = alloc_dmamap(dmat, M_NOWAIT);
318 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
319 __func__, dmat, ENOMEM);
324 * Bouncing might be required if the driver asks for an active
325 * exclusion region, a data alignment that is stricter than 1, and/or
326 * an active address boundary.
328 if (dmat->bounce_flags & BF_COULD_BOUNCE) {
330 if (dmat->bounce_zone == NULL) {
331 if ((error = alloc_bounce_zone(dmat)) != 0) {
332 free(*mapp, M_DEVBUF);
336 bz = dmat->bounce_zone;
338 (*mapp)->flags = DMAMAP_COULD_BOUNCE;
341 * Attempt to add pages to our pool on a per-instance
342 * basis up to a sane limit.
344 if (dmat->common.alignment > 1)
345 maxpages = MAX_BPAGES;
347 maxpages = MIN(MAX_BPAGES, Maxmem -
348 atop(dmat->common.lowaddr));
349 if ((dmat->bounce_flags & BF_MIN_ALLOC_COMP) == 0 ||
350 (bz->map_count > 0 && bz->total_bpages < maxpages)) {
351 pages = MAX(atop(dmat->common.maxsize), 1);
352 pages = MIN(maxpages - bz->total_bpages, pages);
353 pages = MAX(pages, 1);
354 if (alloc_bounce_pages(dmat, pages) < pages)
356 if ((dmat->bounce_flags & BF_MIN_ALLOC_COMP)
359 dmat->bounce_flags |=
370 free(*mapp, M_DEVBUF);
371 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
372 __func__, dmat, dmat->common.flags, error);
377 * Destroy a handle for mapping from kva/uva/physical
378 * address space into bus device space.
381 bounce_bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
384 /* Check we are destroying the correct map type */
385 if ((map->flags & DMAMAP_FROM_DMAMEM) != 0)
386 panic("bounce_bus_dmamap_destroy: Invalid map freed\n");
388 if (STAILQ_FIRST(&map->bpages) != NULL || map->sync_count != 0) {
389 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, EBUSY);
392 if (dmat->bounce_zone) {
393 KASSERT((map->flags & DMAMAP_COULD_BOUNCE) != 0,
394 ("%s: Bounce zone when cannot bounce", __func__));
395 dmat->bounce_zone->map_count--;
399 CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
404 * Allocate a piece of memory that can be efficiently mapped into
405 * bus device space based on the constraints lited in the dma tag.
406 * A dmamap to for use with dmamap_load is also allocated.
409 bounce_bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
414 * This bus_dma implementation requires IO-Coherent architecutre.
415 * If IO-Coherency is not guaranteed, the BUS_DMA_COHERENT flag has
416 * to be implented using non-cacheable memory.
422 if (flags & BUS_DMA_NOWAIT)
427 if (dmat->segments == NULL) {
428 dmat->segments = (bus_dma_segment_t *)malloc(
429 sizeof(bus_dma_segment_t) * dmat->common.nsegments,
431 if (dmat->segments == NULL) {
432 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
433 __func__, dmat, dmat->common.flags, ENOMEM);
437 if (flags & BUS_DMA_ZERO)
439 if (flags & BUS_DMA_NOCACHE)
440 attr = VM_MEMATTR_UNCACHEABLE;
441 else if ((flags & BUS_DMA_COHERENT) != 0 &&
442 (dmat->bounce_flags & BF_COHERENT) == 0)
444 * If we have a non-coherent tag, and are trying to allocate
445 * a coherent block of memory it needs to be uncached.
447 attr = VM_MEMATTR_UNCACHEABLE;
449 attr = VM_MEMATTR_DEFAULT;
452 * Create the map, but don't set the could bounce flag as
453 * this allocation should never bounce;
455 *mapp = alloc_dmamap(dmat, mflags);
457 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
458 __func__, dmat, dmat->common.flags, ENOMEM);
461 (*mapp)->flags = DMAMAP_FROM_DMAMEM;
464 * Allocate the buffer from the malloc(9) allocator if...
465 * - It's small enough to fit into a single power of two sized bucket.
466 * - The alignment is less than or equal to the maximum size
467 * - The low address requirement is fulfilled.
468 * else allocate non-contiguous pages if...
469 * - The page count that could get allocated doesn't exceed
470 * nsegments also when the maximum segment size is less
472 * - The alignment constraint isn't larger than a page boundary.
473 * - There are no boundary-crossing constraints.
474 * else allocate a block of contiguous pages because one or more of the
475 * constraints is something that only the contig allocator can fulfill.
477 * NOTE: The (dmat->common.alignment <= dmat->maxsize) check
478 * below is just a quick hack. The exact alignment guarantees
479 * of malloc(9) need to be nailed down, and the code below
480 * should be rewritten to take that into account.
482 * In the meantime warn the user if malloc gets it wrong.
484 if ((dmat->common.maxsize <= PAGE_SIZE) &&
485 (dmat->common.alignment <= dmat->common.maxsize) &&
486 dmat->common.lowaddr >= ptoa((vm_paddr_t)Maxmem) &&
487 attr == VM_MEMATTR_DEFAULT) {
488 *vaddr = malloc(dmat->common.maxsize, M_DEVBUF, mflags);
489 } else if (dmat->common.nsegments >=
490 howmany(dmat->common.maxsize, MIN(dmat->common.maxsegsz, PAGE_SIZE)) &&
491 dmat->common.alignment <= PAGE_SIZE &&
492 (dmat->common.boundary % PAGE_SIZE) == 0) {
493 /* Page-based multi-segment allocations allowed */
494 *vaddr = (void *)kmem_alloc_attr(dmat->common.maxsize, mflags,
495 0ul, dmat->common.lowaddr, attr);
496 dmat->bounce_flags |= BF_KMEM_ALLOC;
498 *vaddr = (void *)kmem_alloc_contig(dmat->common.maxsize, mflags,
499 0ul, dmat->common.lowaddr, dmat->common.alignment != 0 ?
500 dmat->common.alignment : 1ul, dmat->common.boundary, attr);
501 dmat->bounce_flags |= BF_KMEM_ALLOC;
503 if (*vaddr == NULL) {
504 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
505 __func__, dmat, dmat->common.flags, ENOMEM);
506 free(*mapp, M_DEVBUF);
508 } else if (vtophys(*vaddr) & (dmat->common.alignment - 1)) {
509 printf("bus_dmamem_alloc failed to align memory properly.\n");
512 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
513 __func__, dmat, dmat->common.flags, 0);
518 * Free a piece of memory and it's allociated dmamap, that was allocated
519 * via bus_dmamem_alloc. Make the same choice for free/contigfree.
522 bounce_bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
526 * Check the map came from bounce_bus_dmamem_alloc, so the map
527 * should be NULL and the BF_KMEM_ALLOC flag cleared if malloc()
528 * was used and set if kmem_alloc_contig() was used.
530 if ((map->flags & DMAMAP_FROM_DMAMEM) == 0)
531 panic("bus_dmamem_free: Invalid map freed\n");
532 if ((dmat->bounce_flags & BF_KMEM_ALLOC) == 0)
533 free(vaddr, M_DEVBUF);
535 kmem_free((vm_offset_t)vaddr, dmat->common.maxsize);
538 CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat,
543 _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
544 bus_size_t buflen, int flags)
549 if ((map->flags & DMAMAP_COULD_BOUNCE) != 0 && map->pagesneeded == 0) {
551 * Count the number of bounce pages
552 * needed in order to complete this transfer
555 while (buflen != 0) {
556 sgsize = MIN(buflen, dmat->common.maxsegsz);
557 if (bus_dma_run_filter(&dmat->common, curaddr)) {
559 PAGE_SIZE - (curaddr & PAGE_MASK));
565 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
570 _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, pmap_t pmap,
571 void *buf, bus_size_t buflen, int flags)
574 vm_offset_t vendaddr;
578 if ((map->flags & DMAMAP_COULD_BOUNCE) != 0 && map->pagesneeded == 0) {
579 CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
580 "alignment= %d", dmat->common.lowaddr,
581 ptoa((vm_paddr_t)Maxmem),
582 dmat->common.boundary, dmat->common.alignment);
583 CTR2(KTR_BUSDMA, "map= %p, pagesneeded= %d", map,
586 * Count the number of bounce pages
587 * needed in order to complete this transfer
589 vaddr = (vm_offset_t)buf;
590 vendaddr = (vm_offset_t)buf + buflen;
592 while (vaddr < vendaddr) {
593 sg_len = PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK);
594 if (pmap == kernel_pmap)
595 paddr = pmap_kextract(vaddr);
597 paddr = pmap_extract(pmap, vaddr);
598 if (bus_dma_run_filter(&dmat->common, paddr) != 0) {
599 sg_len = roundup2(sg_len,
600 dmat->common.alignment);
605 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
610 _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags)
613 /* Reserve Necessary Bounce Pages */
614 mtx_lock(&bounce_lock);
615 if (flags & BUS_DMA_NOWAIT) {
616 if (reserve_bounce_pages(dmat, map, 0) != 0) {
617 mtx_unlock(&bounce_lock);
621 if (reserve_bounce_pages(dmat, map, 1) != 0) {
622 /* Queue us for resources */
623 STAILQ_INSERT_TAIL(&bounce_map_waitinglist, map, links);
624 mtx_unlock(&bounce_lock);
625 return (EINPROGRESS);
628 mtx_unlock(&bounce_lock);
634 * Add a single contiguous physical range to the segment list.
637 _bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr,
638 bus_size_t sgsize, bus_dma_segment_t *segs, int *segp)
640 bus_addr_t baddr, bmask;
644 * Make sure we don't cross any boundaries.
646 bmask = ~(dmat->common.boundary - 1);
647 if (dmat->common.boundary > 0) {
648 baddr = (curaddr + dmat->common.boundary) & bmask;
649 if (sgsize > (baddr - curaddr))
650 sgsize = (baddr - curaddr);
654 * Insert chunk into a segment, coalescing with
655 * previous segment if possible.
660 segs[seg].ds_addr = curaddr;
661 segs[seg].ds_len = sgsize;
663 if (curaddr == segs[seg].ds_addr + segs[seg].ds_len &&
664 (segs[seg].ds_len + sgsize) <= dmat->common.maxsegsz &&
665 (dmat->common.boundary == 0 ||
666 (segs[seg].ds_addr & bmask) == (curaddr & bmask)))
667 segs[seg].ds_len += sgsize;
669 if (++seg >= dmat->common.nsegments)
671 segs[seg].ds_addr = curaddr;
672 segs[seg].ds_len = sgsize;
680 * Utility function to load a physical buffer. segp contains
681 * the starting segment on entrace, and the ending segment on exit.
684 bounce_bus_dmamap_load_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
685 vm_paddr_t buf, bus_size_t buflen, int flags, bus_dma_segment_t *segs,
688 struct sync_list *sl;
690 bus_addr_t curaddr, sl_end;
694 segs = dmat->segments;
696 if ((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) {
697 _bus_dmamap_count_phys(dmat, map, buf, buflen, flags);
698 if (map->pagesneeded != 0) {
699 error = _bus_dmamap_reserve_pages(dmat, map, flags);
705 sl = map->slist + map->sync_count - 1;
710 sgsize = MIN(buflen, dmat->common.maxsegsz);
711 if (((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) &&
712 map->pagesneeded != 0 &&
713 bus_dma_run_filter(&dmat->common, curaddr)) {
714 sgsize = MIN(sgsize, PAGE_SIZE - (curaddr & PAGE_MASK));
715 curaddr = add_bounce_page(dmat, map, 0, curaddr,
717 } else if ((dmat->bounce_flags & BF_COHERENT) == 0) {
718 if (map->sync_count > 0)
719 sl_end = sl->paddr + sl->datacount;
721 if (map->sync_count == 0 || curaddr != sl_end) {
722 if (++map->sync_count > dmat->common.nsegments)
727 sl->datacount = sgsize;
728 sl->pages = PHYS_TO_VM_PAGE(curaddr);
729 KASSERT(sl->pages != NULL,
730 ("%s: page at PA:0x%08lx is not in "
731 "vm_page_array", __func__, curaddr));
733 sl->datacount += sgsize;
735 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
746 return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
750 * Utility function to load a linear buffer. segp contains
751 * the starting segment on entrace, and the ending segment on exit.
754 bounce_bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
755 bus_size_t buflen, pmap_t pmap, int flags, bus_dma_segment_t *segs,
758 struct sync_list *sl;
759 bus_size_t sgsize, max_sgsize;
760 bus_addr_t curaddr, sl_pend;
761 vm_offset_t kvaddr, vaddr, sl_vend;
765 segs = dmat->segments;
767 if ((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) {
768 _bus_dmamap_count_pages(dmat, map, pmap, buf, buflen, flags);
769 if (map->pagesneeded != 0) {
770 error = _bus_dmamap_reserve_pages(dmat, map, flags);
776 sl = map->slist + map->sync_count - 1;
777 vaddr = (vm_offset_t)buf;
783 * Get the physical address for this segment.
785 if (pmap == kernel_pmap) {
786 curaddr = pmap_kextract(vaddr);
789 curaddr = pmap_extract(pmap, vaddr);
794 * Compute the segment size, and adjust counts.
796 max_sgsize = MIN(buflen, dmat->common.maxsegsz);
797 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
798 if (((dmat->bounce_flags & BF_COULD_BOUNCE) != 0) &&
799 map->pagesneeded != 0 &&
800 bus_dma_run_filter(&dmat->common, curaddr)) {
801 sgsize = roundup2(sgsize, dmat->common.alignment);
802 sgsize = MIN(sgsize, max_sgsize);
803 curaddr = add_bounce_page(dmat, map, kvaddr, curaddr,
805 } else if ((dmat->bounce_flags & BF_COHERENT) == 0) {
806 sgsize = MIN(sgsize, max_sgsize);
807 if (map->sync_count > 0) {
808 sl_pend = sl->paddr + sl->datacount;
809 sl_vend = sl->vaddr + sl->datacount;
812 if (map->sync_count == 0 ||
813 (kvaddr != 0 && kvaddr != sl_vend) ||
814 (curaddr != sl_pend)) {
815 if (++map->sync_count > dmat->common.nsegments)
823 sl->pages = PHYS_TO_VM_PAGE(curaddr);
824 KASSERT(sl->pages != NULL,
825 ("%s: page at PA:0x%08lx is not "
826 "in vm_page_array", __func__,
829 sl->datacount = sgsize;
831 sl->datacount += sgsize;
833 sgsize = MIN(sgsize, max_sgsize);
835 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
847 return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
851 bounce_bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map,
852 struct memdesc *mem, bus_dmamap_callback_t *callback, void *callback_arg)
855 if ((map->flags & DMAMAP_COULD_BOUNCE) == 0)
859 map->callback = callback;
860 map->callback_arg = callback_arg;
863 static bus_dma_segment_t *
864 bounce_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map,
865 bus_dma_segment_t *segs, int nsegs, int error)
869 segs = dmat->segments;
874 * Release the mapping held by map.
877 bounce_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
879 struct bounce_page *bpage;
881 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
882 STAILQ_REMOVE_HEAD(&map->bpages, links);
883 free_bounce_page(dmat, bpage);
890 dma_preread_safe(vm_offset_t va, vm_size_t size)
893 * Write back any partial cachelines immediately before and
894 * after the DMA region.
896 if (va & (dcache_line_size - 1))
897 cpu_dcache_wb_range(va, 1);
898 if ((va + size) & (dcache_line_size - 1))
899 cpu_dcache_wb_range(va + size, 1);
901 cpu_dcache_inv_range(va, size);
905 dma_dcache_sync(struct sync_list *sl, bus_dmasync_op_t op)
907 uint32_t len, offset;
910 vm_offset_t va, tempva;
913 offset = sl->paddr & PAGE_MASK;
915 size = sl->datacount;
918 for ( ; size != 0; size -= len, pa += len, offset = 0, ++m) {
920 if (sl->vaddr == 0) {
921 len = min(PAGE_SIZE - offset, size);
922 tempva = pmap_quick_enter_page(m);
923 va = tempva | offset;
924 KASSERT(pa == (VM_PAGE_TO_PHYS(m) | offset),
925 ("unexpected vm_page_t phys: 0x%16lx != 0x%16lx",
926 VM_PAGE_TO_PHYS(m) | offset, pa));
933 case BUS_DMASYNC_PREWRITE:
934 case BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD:
935 cpu_dcache_wb_range(va, len);
937 case BUS_DMASYNC_PREREAD:
939 * An mbuf may start in the middle of a cacheline. There
940 * will be no cpu writes to the beginning of that line
941 * (which contains the mbuf header) while dma is in
942 * progress. Handle that case by doing a writeback of
943 * just the first cacheline before invalidating the
944 * overall buffer. Any mbuf in a chain may have this
945 * misalignment. Buffers which are not mbufs bounce if
946 * they are not aligned to a cacheline.
948 dma_preread_safe(va, len);
950 case BUS_DMASYNC_POSTREAD:
951 case BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE:
952 cpu_dcache_inv_range(va, len);
955 panic("unsupported combination of sync operations: "
960 pmap_quick_remove_page(tempva);
965 bounce_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map,
968 struct bounce_page *bpage;
969 struct sync_list *sl, *end;
970 vm_offset_t datavaddr, tempvaddr;
972 if (op == BUS_DMASYNC_POSTWRITE)
975 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
977 * Wait for any DMA operations to complete before the bcopy.
982 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
983 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
984 "performing bounce", __func__, dmat, dmat->common.flags,
987 if ((op & BUS_DMASYNC_PREWRITE) != 0) {
988 while (bpage != NULL) {
990 datavaddr = bpage->datavaddr;
991 if (datavaddr == 0) {
992 tempvaddr = pmap_quick_enter_page(
994 datavaddr = tempvaddr | bpage->dataoffs;
997 bcopy((void *)datavaddr,
998 (void *)bpage->vaddr, bpage->datacount);
1000 pmap_quick_remove_page(tempvaddr);
1001 if ((dmat->bounce_flags & BF_COHERENT) == 0)
1002 cpu_dcache_wb_range(bpage->vaddr,
1004 bpage = STAILQ_NEXT(bpage, links);
1006 dmat->bounce_zone->total_bounced++;
1007 } else if ((op & BUS_DMASYNC_PREREAD) != 0) {
1008 while (bpage != NULL) {
1009 if ((dmat->bounce_flags & BF_COHERENT) == 0)
1010 cpu_dcache_wbinv_range(bpage->vaddr,
1012 bpage = STAILQ_NEXT(bpage, links);
1016 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
1017 while (bpage != NULL) {
1018 if ((dmat->bounce_flags & BF_COHERENT) == 0)
1019 cpu_dcache_inv_range(bpage->vaddr,
1022 datavaddr = bpage->datavaddr;
1023 if (datavaddr == 0) {
1024 tempvaddr = pmap_quick_enter_page(
1026 datavaddr = tempvaddr | bpage->dataoffs;
1029 bcopy((void *)bpage->vaddr,
1030 (void *)datavaddr, bpage->datacount);
1033 pmap_quick_remove_page(tempvaddr);
1034 bpage = STAILQ_NEXT(bpage, links);
1036 dmat->bounce_zone->total_bounced++;
1041 * Cache maintenance for normal (non-COHERENT non-bounce) buffers.
1043 if (map->sync_count != 0) {
1044 sl = &map->slist[0];
1045 end = &map->slist[map->sync_count];
1046 CTR3(KTR_BUSDMA, "%s: tag %p op 0x%x "
1047 "performing sync", __func__, dmat, op);
1049 for ( ; sl != end; ++sl)
1050 dma_dcache_sync(sl, op);
1053 if ((op & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) != 0) {
1055 * Wait for the bcopy to complete before any DMA operations.
1062 init_bounce_pages(void *dummy __unused)
1066 STAILQ_INIT(&bounce_zone_list);
1067 STAILQ_INIT(&bounce_map_waitinglist);
1068 STAILQ_INIT(&bounce_map_callbacklist);
1069 mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
1071 SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
1073 static struct sysctl_ctx_list *
1074 busdma_sysctl_tree(struct bounce_zone *bz)
1077 return (&bz->sysctl_tree);
1080 static struct sysctl_oid *
1081 busdma_sysctl_tree_top(struct bounce_zone *bz)
1084 return (bz->sysctl_tree_top);
1088 alloc_bounce_zone(bus_dma_tag_t dmat)
1090 struct bounce_zone *bz;
1092 /* Check to see if we already have a suitable zone */
1093 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1094 if ((dmat->common.alignment <= bz->alignment) &&
1095 (dmat->common.lowaddr >= bz->lowaddr)) {
1096 dmat->bounce_zone = bz;
1101 if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
1102 M_NOWAIT | M_ZERO)) == NULL)
1105 STAILQ_INIT(&bz->bounce_page_list);
1106 bz->free_bpages = 0;
1107 bz->reserved_bpages = 0;
1108 bz->active_bpages = 0;
1109 bz->lowaddr = dmat->common.lowaddr;
1110 bz->alignment = MAX(dmat->common.alignment, PAGE_SIZE);
1112 snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1114 snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1115 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1116 dmat->bounce_zone = bz;
1118 sysctl_ctx_init(&bz->sysctl_tree);
1119 bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
1120 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1121 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
1122 if (bz->sysctl_tree_top == NULL) {
1123 sysctl_ctx_free(&bz->sysctl_tree);
1124 return (0); /* XXX error code? */
1127 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1128 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1129 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1130 "Total bounce pages");
1131 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1132 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1133 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1134 "Free bounce pages");
1135 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1136 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1137 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1138 "Reserved bounce pages");
1139 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1140 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1141 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1142 "Active bounce pages");
1143 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1144 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1145 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1146 "Total bounce requests");
1147 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1148 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1149 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1150 "Total bounce requests that were deferred");
1151 SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
1152 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1153 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1154 SYSCTL_ADD_UAUTO(busdma_sysctl_tree(bz),
1155 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1156 "alignment", CTLFLAG_RD, &bz->alignment, "");
1162 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
1164 struct bounce_zone *bz;
1167 bz = dmat->bounce_zone;
1169 while (numpages > 0) {
1170 struct bounce_page *bpage;
1172 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
1177 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1178 M_NOWAIT, 0ul, bz->lowaddr, PAGE_SIZE, 0);
1179 if (bpage->vaddr == 0) {
1180 free(bpage, M_DEVBUF);
1183 bpage->busaddr = pmap_kextract(bpage->vaddr);
1184 mtx_lock(&bounce_lock);
1185 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1189 mtx_unlock(&bounce_lock);
1197 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1199 struct bounce_zone *bz;
1202 mtx_assert(&bounce_lock, MA_OWNED);
1203 bz = dmat->bounce_zone;
1204 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1205 if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
1206 return (map->pagesneeded - (map->pagesreserved + pages));
1207 bz->free_bpages -= pages;
1208 bz->reserved_bpages += pages;
1209 map->pagesreserved += pages;
1210 pages = map->pagesneeded - map->pagesreserved;
1216 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1217 bus_addr_t addr, bus_size_t size)
1219 struct bounce_zone *bz;
1220 struct bounce_page *bpage;
1222 KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
1223 KASSERT((map->flags & DMAMAP_COULD_BOUNCE) != 0,
1224 ("add_bounce_page: bad map %p", map));
1226 bz = dmat->bounce_zone;
1227 if (map->pagesneeded == 0)
1228 panic("add_bounce_page: map doesn't need any pages");
1231 if (map->pagesreserved == 0)
1232 panic("add_bounce_page: map doesn't need any pages");
1233 map->pagesreserved--;
1235 mtx_lock(&bounce_lock);
1236 bpage = STAILQ_FIRST(&bz->bounce_page_list);
1238 panic("add_bounce_page: free page list is empty");
1240 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1241 bz->reserved_bpages--;
1242 bz->active_bpages++;
1243 mtx_unlock(&bounce_lock);
1245 if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1246 /* Page offset needs to be preserved. */
1247 bpage->vaddr |= addr & PAGE_MASK;
1248 bpage->busaddr |= addr & PAGE_MASK;
1250 bpage->datavaddr = vaddr;
1251 bpage->datapage = PHYS_TO_VM_PAGE(addr);
1252 bpage->dataoffs = addr & PAGE_MASK;
1253 bpage->datacount = size;
1254 STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1255 return (bpage->busaddr);
1259 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1261 struct bus_dmamap *map;
1262 struct bounce_zone *bz;
1264 bz = dmat->bounce_zone;
1265 bpage->datavaddr = 0;
1266 bpage->datacount = 0;
1267 if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1269 * Reset the bounce page to start at offset 0. Other uses
1270 * of this bounce page may need to store a full page of
1271 * data and/or assume it starts on a page boundary.
1273 bpage->vaddr &= ~PAGE_MASK;
1274 bpage->busaddr &= ~PAGE_MASK;
1277 mtx_lock(&bounce_lock);
1278 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1280 bz->active_bpages--;
1281 if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1282 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1283 STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1284 STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1286 busdma_swi_pending = 1;
1287 bz->total_deferred++;
1288 swi_sched(vm_ih, 0);
1291 mtx_unlock(&bounce_lock);
1298 struct bus_dmamap *map;
1300 mtx_lock(&bounce_lock);
1301 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1302 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1303 mtx_unlock(&bounce_lock);
1305 (dmat->common.lockfunc)(dmat->common.lockfuncarg, BUS_DMA_LOCK);
1306 bus_dmamap_load_mem(map->dmat, map, &map->mem,
1307 map->callback, map->callback_arg, BUS_DMA_WAITOK);
1308 (dmat->common.lockfunc)(dmat->common.lockfuncarg,
1310 mtx_lock(&bounce_lock);
1312 mtx_unlock(&bounce_lock);
1315 struct bus_dma_impl bus_dma_bounce_impl = {
1316 .tag_create = bounce_bus_dma_tag_create,
1317 .tag_destroy = bounce_bus_dma_tag_destroy,
1318 .map_create = bounce_bus_dmamap_create,
1319 .map_destroy = bounce_bus_dmamap_destroy,
1320 .mem_alloc = bounce_bus_dmamem_alloc,
1321 .mem_free = bounce_bus_dmamem_free,
1322 .load_phys = bounce_bus_dmamap_load_phys,
1323 .load_buffer = bounce_bus_dmamap_load_buffer,
1324 .load_ma = bus_dmamap_load_ma_triv,
1325 .map_waitok = bounce_bus_dmamap_waitok,
1326 .map_complete = bounce_bus_dmamap_complete,
1327 .map_unload = bounce_bus_dmamap_unload,
1328 .map_sync = bounce_bus_dmamap_sync