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;
76 size_t alloc_alignment;
79 bus_dma_segment_t *segments;
80 struct bounce_zone *bounce_zone;
84 vm_offset_t vaddr; /* kva of bounce buffer */
85 bus_addr_t busaddr; /* Physical address */
86 vm_offset_t datavaddr; /* kva of client data */
87 vm_page_t datapage; /* physical page of client data */
88 vm_offset_t dataoffs; /* page offset of client data */
89 bus_size_t datacount; /* client data count */
90 STAILQ_ENTRY(bounce_page) links;
93 int busdma_swi_pending;
96 STAILQ_ENTRY(bounce_zone) links;
97 STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
105 bus_size_t alignment;
109 struct sysctl_ctx_list sysctl_tree;
110 struct sysctl_oid *sysctl_tree_top;
113 static struct mtx bounce_lock;
114 static int total_bpages;
115 static int busdma_zonecount;
116 static STAILQ_HEAD(, bounce_zone) bounce_zone_list;
118 static SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
119 "Busdma parameters");
120 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
121 "Total bounce pages");
124 vm_offset_t vaddr; /* kva of client data */
125 bus_addr_t paddr; /* physical address */
126 vm_page_t pages; /* starting page of client data */
127 bus_size_t datacount; /* client data count */
131 struct bp_list bpages;
136 bus_dmamap_callback_t *callback;
138 STAILQ_ENTRY(bus_dmamap) links;
140 #define DMAMAP_COHERENT (1 << 0)
141 #define DMAMAP_FROM_DMAMEM (1 << 1)
142 #define DMAMAP_MBUF (1 << 2)
144 struct sync_list slist[];
147 static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
148 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
150 static void init_bounce_pages(void *dummy);
151 static int alloc_bounce_zone(bus_dma_tag_t dmat);
152 static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
153 static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
155 static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
156 vm_offset_t vaddr, bus_addr_t addr, bus_size_t size);
157 static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
158 int run_filter(bus_dma_tag_t dmat, bus_addr_t paddr);
159 static bool _bus_dmamap_pagesneeded(bus_dma_tag_t dmat, bus_dmamap_t map,
160 vm_paddr_t buf, bus_size_t buflen, int *pagesneeded);
161 static void _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
162 pmap_t pmap, void *buf, bus_size_t buflen, int flags);
163 static void _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
164 vm_paddr_t buf, bus_size_t buflen, int flags);
165 static int _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
169 * Return true if the DMA should bounce because the start or end does not fall
170 * on a cacheline boundary (which would require a partial cacheline flush).
171 * COHERENT memory doesn't trigger cacheline flushes. Memory allocated by
172 * bus_dmamem_alloc() is always aligned to cacheline boundaries, and there's a
173 * strict rule that such memory cannot be accessed by the CPU while DMA is in
174 * progress (or by multiple DMA engines at once), so that it's always safe to do
175 * full cacheline flushes even if that affects memory outside the range of a
176 * given DMA operation that doesn't involve the full allocated buffer. If we're
177 * mapping an mbuf, that follows the same rules as a buffer we allocated.
180 cacheline_bounce(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t paddr,
184 #define DMAMAP_CACHELINE_FLAGS \
185 (DMAMAP_FROM_DMAMEM | DMAMAP_COHERENT | DMAMAP_MBUF)
186 if ((dmat->bounce_flags & BF_COHERENT) != 0)
188 if (map != NULL && (map->flags & DMAMAP_CACHELINE_FLAGS) != 0)
190 return (((paddr | size) & (dcache_line_size - 1)) != 0);
191 #undef DMAMAP_CACHELINE_FLAGS
195 * Return true if the given address does not fall on the alignment boundary.
198 alignment_bounce(bus_dma_tag_t dmat, bus_addr_t addr)
201 return ((addr & (dmat->common.alignment - 1)) != 0);
205 might_bounce(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t paddr,
209 /* Memory allocated by bounce_bus_dmamem_alloc won't bounce */
210 if ((map->flags & DMAMAP_FROM_DMAMEM) != 0)
213 if ((dmat->bounce_flags & BF_COULD_BOUNCE) != 0)
216 if (cacheline_bounce(dmat, map, paddr, size))
219 if (alignment_bounce(dmat, paddr))
226 must_bounce(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t paddr,
230 if (cacheline_bounce(dmat, map, paddr, size))
233 if (alignment_bounce(dmat, paddr))
236 if ((dmat->bounce_flags & BF_COULD_BOUNCE) != 0 &&
237 bus_dma_run_filter(&dmat->common, paddr))
244 * Allocate a device specific dma_tag.
247 bounce_bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
248 bus_addr_t boundary, bus_addr_t lowaddr, bus_addr_t highaddr,
249 bus_dma_filter_t *filter, void *filterarg, bus_size_t maxsize,
250 int nsegments, bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
251 void *lockfuncarg, bus_dma_tag_t *dmat)
253 bus_dma_tag_t newtag;
257 error = common_bus_dma_tag_create(parent != NULL ? &parent->common :
258 NULL, alignment, boundary, lowaddr, highaddr, filter, filterarg,
259 maxsize, nsegments, maxsegsz, flags, lockfunc, lockfuncarg,
260 sizeof (struct bus_dma_tag), (void **)&newtag);
264 newtag->common.impl = &bus_dma_bounce_impl;
265 newtag->map_count = 0;
266 newtag->segments = NULL;
268 if ((flags & BUS_DMA_COHERENT) != 0) {
269 newtag->bounce_flags |= BF_COHERENT;
270 newtag->alloc_alignment = newtag->common.alignment;
271 newtag->alloc_size = newtag->common.maxsize;
274 * Ensure the buffer is aligned to a cacheline when allocating
275 * a non-coherent buffer. This is so we don't have any data
276 * that another CPU may be accessing around DMA buffer
277 * causing the cache to become dirty.
279 newtag->alloc_alignment = MAX(newtag->common.alignment,
281 newtag->alloc_size = roundup2(newtag->common.maxsize,
285 if (parent != NULL) {
286 if ((newtag->common.filter != NULL ||
287 (parent->bounce_flags & BF_COULD_BOUNCE) != 0))
288 newtag->bounce_flags |= BF_COULD_BOUNCE;
290 /* Copy some flags from the parent */
291 newtag->bounce_flags |= parent->bounce_flags & BF_COHERENT;
294 if (newtag->common.lowaddr < ptoa((vm_paddr_t)Maxmem) ||
295 newtag->common.alignment > 1)
296 newtag->bounce_flags |= BF_COULD_BOUNCE;
298 if ((flags & BUS_DMA_ALLOCNOW) != 0) {
299 struct bounce_zone *bz;
302 if ((error = alloc_bounce_zone(newtag)) != 0) {
303 free(newtag, M_DEVBUF);
306 bz = newtag->bounce_zone;
308 if (ptoa(bz->total_bpages) < maxsize) {
311 pages = atop(round_page(maxsize)) - bz->total_bpages;
313 /* Add pages to our bounce pool */
314 if (alloc_bounce_pages(newtag, pages) < pages)
317 /* Performed initial allocation */
318 newtag->bounce_flags |= BF_MIN_ALLOC_COMP;
323 free(newtag, M_DEVBUF);
326 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
327 __func__, newtag, (newtag != NULL ? newtag->common.flags : 0),
333 bounce_bus_dma_tag_destroy(bus_dma_tag_t dmat)
335 bus_dma_tag_t dmat_copy, parent;
342 if (dmat->map_count != 0) {
346 while (dmat != NULL) {
347 parent = (bus_dma_tag_t)dmat->common.parent;
348 atomic_subtract_int(&dmat->common.ref_count, 1);
349 if (dmat->common.ref_count == 0) {
350 if (dmat->segments != NULL)
351 free(dmat->segments, M_DEVBUF);
352 free(dmat, M_DEVBUF);
354 * Last reference count, so
355 * release our reference
356 * count on our parent.
364 CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
369 bounce_bus_dma_id_mapped(bus_dma_tag_t dmat, vm_paddr_t buf, bus_size_t buflen)
372 if (!might_bounce(dmat, NULL, buf, buflen))
374 return (!_bus_dmamap_pagesneeded(dmat, NULL, buf, buflen, NULL));
378 alloc_dmamap(bus_dma_tag_t dmat, int flags)
383 mapsize = sizeof(*map);
384 mapsize += sizeof(struct sync_list) * dmat->common.nsegments;
385 map = malloc(mapsize, M_DEVBUF, flags | M_ZERO);
389 /* Initialize the new map */
390 STAILQ_INIT(&map->bpages);
396 * Allocate a handle for mapping from kva/uva/physical
397 * address space into bus device space.
400 bounce_bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
402 struct bounce_zone *bz;
403 int error, maxpages, pages;
407 if (dmat->segments == NULL) {
408 dmat->segments = (bus_dma_segment_t *)malloc(
409 sizeof(bus_dma_segment_t) * dmat->common.nsegments,
411 if (dmat->segments == NULL) {
412 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
413 __func__, dmat, ENOMEM);
418 *mapp = alloc_dmamap(dmat, M_NOWAIT);
420 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
421 __func__, dmat, ENOMEM);
426 * Bouncing might be required if the driver asks for an active
427 * exclusion region, a data alignment that is stricter than 1, and/or
428 * an active address boundary.
430 if (dmat->bounce_zone == NULL) {
431 if ((error = alloc_bounce_zone(dmat)) != 0) {
432 free(*mapp, M_DEVBUF);
436 bz = dmat->bounce_zone;
439 * Attempt to add pages to our pool on a per-instance
440 * basis up to a sane limit.
442 if (dmat->common.alignment > 1)
443 maxpages = MAX_BPAGES;
445 maxpages = MIN(MAX_BPAGES, Maxmem -
446 atop(dmat->common.lowaddr));
447 if ((dmat->bounce_flags & BF_MIN_ALLOC_COMP) == 0 ||
448 (bz->map_count > 0 && bz->total_bpages < maxpages)) {
449 pages = MAX(atop(dmat->common.maxsize), 1);
450 pages = MIN(maxpages - bz->total_bpages, pages);
451 pages = MAX(pages, 1);
452 if (alloc_bounce_pages(dmat, pages) < pages)
454 if ((dmat->bounce_flags & BF_MIN_ALLOC_COMP) == 0) {
456 dmat->bounce_flags |= BF_MIN_ALLOC_COMP;
465 if ((dmat->bounce_flags & BF_COHERENT) != 0)
466 (*mapp)->flags |= DMAMAP_COHERENT;
468 free(*mapp, M_DEVBUF);
470 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
471 __func__, dmat, dmat->common.flags, error);
476 * Destroy a handle for mapping from kva/uva/physical
477 * address space into bus device space.
480 bounce_bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
483 /* Check we are destroying the correct map type */
484 if ((map->flags & DMAMAP_FROM_DMAMEM) != 0)
485 panic("bounce_bus_dmamap_destroy: Invalid map freed\n");
487 if (STAILQ_FIRST(&map->bpages) != NULL || map->sync_count != 0) {
488 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, EBUSY);
491 if (dmat->bounce_zone)
492 dmat->bounce_zone->map_count--;
495 CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
500 * Allocate a piece of memory that can be efficiently mapped into
501 * bus device space based on the constraints lited in the dma tag.
502 * A dmamap to for use with dmamap_load is also allocated.
505 bounce_bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
511 if (flags & BUS_DMA_NOWAIT)
516 if (dmat->segments == NULL) {
517 dmat->segments = (bus_dma_segment_t *)malloc(
518 sizeof(bus_dma_segment_t) * dmat->common.nsegments,
520 if (dmat->segments == NULL) {
521 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
522 __func__, dmat, dmat->common.flags, ENOMEM);
526 if (flags & BUS_DMA_ZERO)
528 if (flags & BUS_DMA_NOCACHE)
529 attr = VM_MEMATTR_UNCACHEABLE;
530 else if ((flags & BUS_DMA_COHERENT) != 0 &&
531 (dmat->bounce_flags & BF_COHERENT) == 0)
533 * If we have a non-coherent tag, and are trying to allocate
534 * a coherent block of memory it needs to be uncached.
536 attr = VM_MEMATTR_UNCACHEABLE;
538 attr = VM_MEMATTR_DEFAULT;
541 * Create the map, but don't set the could bounce flag as
542 * this allocation should never bounce;
544 *mapp = alloc_dmamap(dmat, mflags);
546 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
547 __func__, dmat, dmat->common.flags, ENOMEM);
552 * Mark the map as coherent if we used uncacheable memory or the
553 * tag was already marked as coherent.
555 if (attr == VM_MEMATTR_UNCACHEABLE ||
556 (dmat->bounce_flags & BF_COHERENT) != 0)
557 (*mapp)->flags |= DMAMAP_COHERENT;
559 (*mapp)->flags |= DMAMAP_FROM_DMAMEM;
562 * Allocate the buffer from the malloc(9) allocator if...
563 * - It's small enough to fit into a single power of two sized bucket.
564 * - The alignment is less than or equal to the maximum size
565 * - The low address requirement is fulfilled.
566 * else allocate non-contiguous pages if...
567 * - The page count that could get allocated doesn't exceed
568 * nsegments also when the maximum segment size is less
570 * - The alignment constraint isn't larger than a page boundary.
571 * - There are no boundary-crossing constraints.
572 * else allocate a block of contiguous pages because one or more of the
573 * constraints is something that only the contig allocator can fulfill.
575 * NOTE: The (dmat->common.alignment <= dmat->maxsize) check
576 * below is just a quick hack. The exact alignment guarantees
577 * of malloc(9) need to be nailed down, and the code below
578 * should be rewritten to take that into account.
580 * In the meantime warn the user if malloc gets it wrong.
582 if ((dmat->alloc_size <= PAGE_SIZE) &&
583 (dmat->alloc_alignment <= dmat->alloc_size) &&
584 dmat->common.lowaddr >= ptoa((vm_paddr_t)Maxmem) &&
585 attr == VM_MEMATTR_DEFAULT) {
586 *vaddr = malloc(dmat->alloc_size, M_DEVBUF, mflags);
587 } else if (dmat->common.nsegments >=
588 howmany(dmat->alloc_size, MIN(dmat->common.maxsegsz, PAGE_SIZE)) &&
589 dmat->alloc_alignment <= PAGE_SIZE &&
590 (dmat->common.boundary % PAGE_SIZE) == 0) {
591 /* Page-based multi-segment allocations allowed */
592 *vaddr = (void *)kmem_alloc_attr(dmat->alloc_size, mflags,
593 0ul, dmat->common.lowaddr, attr);
594 dmat->bounce_flags |= BF_KMEM_ALLOC;
596 *vaddr = (void *)kmem_alloc_contig(dmat->alloc_size, mflags,
597 0ul, dmat->common.lowaddr, dmat->alloc_alignment != 0 ?
598 dmat->alloc_alignment : 1ul, dmat->common.boundary, attr);
599 dmat->bounce_flags |= BF_KMEM_ALLOC;
601 if (*vaddr == NULL) {
602 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
603 __func__, dmat, dmat->common.flags, ENOMEM);
604 free(*mapp, M_DEVBUF);
606 } else if (vtophys(*vaddr) & (dmat->alloc_alignment - 1)) {
607 printf("bus_dmamem_alloc failed to align memory properly.\n");
610 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
611 __func__, dmat, dmat->common.flags, 0);
616 * Free a piece of memory and it's allociated dmamap, that was allocated
617 * via bus_dmamem_alloc. Make the same choice for free/contigfree.
620 bounce_bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
624 * Check the map came from bounce_bus_dmamem_alloc, so the map
625 * should be NULL and the BF_KMEM_ALLOC flag cleared if malloc()
626 * was used and set if kmem_alloc_contig() was used.
628 if ((map->flags & DMAMAP_FROM_DMAMEM) == 0)
629 panic("bus_dmamem_free: Invalid map freed\n");
630 if ((dmat->bounce_flags & BF_KMEM_ALLOC) == 0)
631 free(vaddr, M_DEVBUF);
633 kmem_free((vm_offset_t)vaddr, dmat->alloc_size);
636 CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat,
641 _bus_dmamap_pagesneeded(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
642 bus_size_t buflen, int *pagesneeded)
649 * Count the number of bounce pages needed in order to
650 * complete this transfer
654 while (buflen != 0) {
655 sgsize = MIN(buflen, dmat->common.maxsegsz);
656 if (must_bounce(dmat, map, curaddr, sgsize)) {
658 PAGE_SIZE - (curaddr & PAGE_MASK));
659 if (pagesneeded == NULL)
667 if (pagesneeded != NULL)
668 *pagesneeded = count;
673 _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
674 bus_size_t buflen, int flags)
677 if (map->pagesneeded == 0) {
678 _bus_dmamap_pagesneeded(dmat, map, buf, buflen,
680 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
685 _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, pmap_t pmap,
686 void *buf, bus_size_t buflen, int flags)
689 vm_offset_t vendaddr;
693 if (map->pagesneeded == 0) {
694 CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
695 "alignment= %d", dmat->common.lowaddr,
696 ptoa((vm_paddr_t)Maxmem),
697 dmat->common.boundary, dmat->common.alignment);
698 CTR2(KTR_BUSDMA, "map= %p, pagesneeded= %d", map,
701 * Count the number of bounce pages
702 * needed in order to complete this transfer
704 vaddr = (vm_offset_t)buf;
705 vendaddr = (vm_offset_t)buf + buflen;
707 while (vaddr < vendaddr) {
708 sg_len = PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK);
709 if (pmap == kernel_pmap)
710 paddr = pmap_kextract(vaddr);
712 paddr = pmap_extract(pmap, vaddr);
713 if (must_bounce(dmat, map, paddr,
714 min(vendaddr - vaddr, (PAGE_SIZE - ((vm_offset_t)vaddr &
715 PAGE_MASK)))) != 0) {
716 sg_len = roundup2(sg_len,
717 dmat->common.alignment);
722 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
727 _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags)
730 /* Reserve Necessary Bounce Pages */
731 mtx_lock(&bounce_lock);
732 if (flags & BUS_DMA_NOWAIT) {
733 if (reserve_bounce_pages(dmat, map, 0) != 0) {
734 mtx_unlock(&bounce_lock);
738 if (reserve_bounce_pages(dmat, map, 1) != 0) {
739 /* Queue us for resources */
740 STAILQ_INSERT_TAIL(&bounce_map_waitinglist, map, links);
741 mtx_unlock(&bounce_lock);
742 return (EINPROGRESS);
745 mtx_unlock(&bounce_lock);
751 * Add a single contiguous physical range to the segment list.
754 _bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr,
755 bus_size_t sgsize, bus_dma_segment_t *segs, int *segp)
757 bus_addr_t baddr, bmask;
761 * Make sure we don't cross any boundaries.
763 bmask = ~(dmat->common.boundary - 1);
764 if (dmat->common.boundary > 0) {
765 baddr = (curaddr + dmat->common.boundary) & bmask;
766 if (sgsize > (baddr - curaddr))
767 sgsize = (baddr - curaddr);
771 * Insert chunk into a segment, coalescing with
772 * previous segment if possible.
777 segs[seg].ds_addr = curaddr;
778 segs[seg].ds_len = sgsize;
780 if (curaddr == segs[seg].ds_addr + segs[seg].ds_len &&
781 (segs[seg].ds_len + sgsize) <= dmat->common.maxsegsz &&
782 (dmat->common.boundary == 0 ||
783 (segs[seg].ds_addr & bmask) == (curaddr & bmask)))
784 segs[seg].ds_len += sgsize;
786 if (++seg >= dmat->common.nsegments)
788 segs[seg].ds_addr = curaddr;
789 segs[seg].ds_len = sgsize;
797 * Utility function to load a physical buffer. segp contains
798 * the starting segment on entrace, and the ending segment on exit.
801 bounce_bus_dmamap_load_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
802 vm_paddr_t buf, bus_size_t buflen, int flags, bus_dma_segment_t *segs,
805 struct sync_list *sl;
807 bus_addr_t curaddr, sl_end;
811 segs = dmat->segments;
813 if (might_bounce(dmat, map, (bus_addr_t)buf, buflen)) {
814 _bus_dmamap_count_phys(dmat, map, buf, buflen, flags);
815 if (map->pagesneeded != 0) {
816 error = _bus_dmamap_reserve_pages(dmat, map, flags);
822 sl = map->slist + map->sync_count - 1;
827 sgsize = MIN(buflen, dmat->common.maxsegsz);
828 if (map->pagesneeded != 0 &&
829 must_bounce(dmat, map, curaddr, sgsize)) {
831 * The attempt to split a physically continuous buffer
832 * seems very controversial, it's unclear whether we
833 * can do this in all cases. Also, memory for bounced
834 * buffers is allocated as pages, so we cannot
835 * guarantee multipage alignment.
837 KASSERT(dmat->common.alignment <= PAGE_SIZE,
838 ("bounced buffer cannot have alignment bigger "
839 "than PAGE_SIZE: %lu", dmat->common.alignment));
840 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
841 sgsize = roundup2(sgsize, dmat->common.alignment);
842 sgsize = MIN(sgsize, dmat->common.maxsegsz);
843 curaddr = add_bounce_page(dmat, map, 0, curaddr,
845 } else if ((map->flags & DMAMAP_COHERENT) == 0) {
846 if (map->sync_count > 0)
847 sl_end = sl->paddr + sl->datacount;
849 if (map->sync_count == 0 || curaddr != sl_end) {
850 if (++map->sync_count > dmat->common.nsegments)
855 sl->pages = PHYS_TO_VM_PAGE(curaddr);
856 KASSERT(sl->pages != NULL,
857 ("%s: page at PA:0x%08lx is not in "
858 "vm_page_array", __func__, curaddr));
859 sl->datacount = sgsize;
861 sl->datacount += sgsize;
863 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
874 return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
878 * Utility function to load a linear buffer. segp contains
879 * the starting segment on entrace, and the ending segment on exit.
882 bounce_bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
883 bus_size_t buflen, pmap_t pmap, int flags, bus_dma_segment_t *segs,
886 struct sync_list *sl;
887 bus_size_t sgsize, max_sgsize;
888 bus_addr_t curaddr, sl_pend;
889 vm_offset_t kvaddr, vaddr, sl_vend;
892 KASSERT((map->flags & DMAMAP_FROM_DMAMEM) != 0 ||
893 dmat->common.alignment <= PAGE_SIZE,
894 ("loading user buffer with alignment bigger than PAGE_SIZE is not "
898 segs = dmat->segments;
900 if (flags & BUS_DMA_LOAD_MBUF)
901 map->flags |= DMAMAP_MBUF;
903 if (might_bounce(dmat, map, (bus_addr_t)buf, buflen)) {
904 _bus_dmamap_count_pages(dmat, map, pmap, buf, buflen, flags);
905 if (map->pagesneeded != 0) {
906 error = _bus_dmamap_reserve_pages(dmat, map, flags);
913 * XXX Optimally we should parse input buffer for physically
914 * continuous segments first and then pass these segment into
917 sl = map->slist + map->sync_count - 1;
918 vaddr = (vm_offset_t)buf;
924 * Get the physical address for this segment.
926 if (pmap == kernel_pmap) {
927 curaddr = pmap_kextract(vaddr);
930 curaddr = pmap_extract(pmap, vaddr);
935 * Compute the segment size, and adjust counts.
937 max_sgsize = MIN(buflen, dmat->common.maxsegsz);
938 if ((map->flags & DMAMAP_FROM_DMAMEM) != 0) {
941 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
942 sgsize = MIN(sgsize, max_sgsize);
945 if (map->pagesneeded != 0 &&
946 must_bounce(dmat, map, curaddr, sgsize)) {
947 /* See comment in bounce_bus_dmamap_load_phys */
948 KASSERT(dmat->common.alignment <= PAGE_SIZE,
949 ("bounced buffer cannot have alignment bigger "
950 "than PAGE_SIZE: %lu", dmat->common.alignment));
951 sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
952 sgsize = roundup2(sgsize, dmat->common.alignment);
953 sgsize = MIN(sgsize, max_sgsize);
954 curaddr = add_bounce_page(dmat, map, kvaddr, curaddr,
956 } else if ((map->flags & DMAMAP_COHERENT) == 0) {
957 if (map->sync_count > 0) {
958 sl_pend = sl->paddr + sl->datacount;
959 sl_vend = sl->vaddr + sl->datacount;
962 if (map->sync_count == 0 ||
963 (kvaddr != 0 && kvaddr != sl_vend) ||
964 (curaddr != sl_pend)) {
965 if (++map->sync_count > dmat->common.nsegments)
973 sl->pages = PHYS_TO_VM_PAGE(curaddr);
974 KASSERT(sl->pages != NULL,
975 ("%s: page at PA:0x%08lx is not "
976 "in vm_page_array", __func__,
979 sl->datacount = sgsize;
981 sl->datacount += sgsize;
983 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
994 return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
998 bounce_bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map,
999 struct memdesc *mem, bus_dmamap_callback_t *callback, void *callback_arg)
1004 map->callback = callback;
1005 map->callback_arg = callback_arg;
1008 static bus_dma_segment_t *
1009 bounce_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map,
1010 bus_dma_segment_t *segs, int nsegs, int error)
1014 segs = dmat->segments;
1019 * Release the mapping held by map.
1022 bounce_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1024 struct bounce_page *bpage;
1026 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1027 STAILQ_REMOVE_HEAD(&map->bpages, links);
1028 free_bounce_page(dmat, bpage);
1031 map->sync_count = 0;
1032 map->flags &= ~DMAMAP_MBUF;
1036 dma_preread_safe(vm_offset_t va, vm_size_t size)
1039 * Write back any partial cachelines immediately before and
1040 * after the DMA region.
1042 if (va & (dcache_line_size - 1))
1043 cpu_dcache_wb_range(va, 1);
1044 if ((va + size) & (dcache_line_size - 1))
1045 cpu_dcache_wb_range(va + size, 1);
1047 cpu_dcache_inv_range(va, size);
1051 dma_dcache_sync(struct sync_list *sl, bus_dmasync_op_t op)
1053 uint32_t len, offset;
1056 vm_offset_t va, tempva;
1059 offset = sl->paddr & PAGE_MASK;
1061 size = sl->datacount;
1064 for ( ; size != 0; size -= len, pa += len, offset = 0, ++m) {
1066 if (sl->vaddr == 0) {
1067 len = min(PAGE_SIZE - offset, size);
1068 tempva = pmap_quick_enter_page(m);
1069 va = tempva | offset;
1070 KASSERT(pa == (VM_PAGE_TO_PHYS(m) | offset),
1071 ("unexpected vm_page_t phys: 0x%16lx != 0x%16lx",
1072 VM_PAGE_TO_PHYS(m) | offset, pa));
1074 len = sl->datacount;
1079 case BUS_DMASYNC_PREWRITE:
1080 case BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD:
1081 cpu_dcache_wb_range(va, len);
1083 case BUS_DMASYNC_PREREAD:
1085 * An mbuf may start in the middle of a cacheline. There
1086 * will be no cpu writes to the beginning of that line
1087 * (which contains the mbuf header) while dma is in
1088 * progress. Handle that case by doing a writeback of
1089 * just the first cacheline before invalidating the
1090 * overall buffer. Any mbuf in a chain may have this
1091 * misalignment. Buffers which are not mbufs bounce if
1092 * they are not aligned to a cacheline.
1094 dma_preread_safe(va, len);
1096 case BUS_DMASYNC_POSTREAD:
1097 case BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE:
1098 cpu_dcache_inv_range(va, len);
1101 panic("unsupported combination of sync operations: "
1106 pmap_quick_remove_page(tempva);
1111 bounce_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map,
1112 bus_dmasync_op_t op)
1114 struct bounce_page *bpage;
1115 struct sync_list *sl, *end;
1116 vm_offset_t datavaddr, tempvaddr;
1118 if (op == BUS_DMASYNC_POSTWRITE)
1121 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
1123 * Wait for any DMA operations to complete before the bcopy.
1128 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1129 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
1130 "performing bounce", __func__, dmat, dmat->common.flags,
1133 if ((op & BUS_DMASYNC_PREWRITE) != 0) {
1134 while (bpage != NULL) {
1136 datavaddr = bpage->datavaddr;
1137 if (datavaddr == 0) {
1138 tempvaddr = pmap_quick_enter_page(
1140 datavaddr = tempvaddr | bpage->dataoffs;
1143 bcopy((void *)datavaddr,
1144 (void *)bpage->vaddr, bpage->datacount);
1146 pmap_quick_remove_page(tempvaddr);
1147 if ((map->flags & DMAMAP_COHERENT) == 0)
1148 cpu_dcache_wb_range(bpage->vaddr,
1150 bpage = STAILQ_NEXT(bpage, links);
1152 dmat->bounce_zone->total_bounced++;
1153 } else if ((op & BUS_DMASYNC_PREREAD) != 0) {
1154 while (bpage != NULL) {
1155 if ((map->flags & DMAMAP_COHERENT) == 0)
1156 cpu_dcache_wbinv_range(bpage->vaddr,
1158 bpage = STAILQ_NEXT(bpage, links);
1162 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
1163 while (bpage != NULL) {
1164 if ((map->flags & DMAMAP_COHERENT) == 0)
1165 cpu_dcache_inv_range(bpage->vaddr,
1168 datavaddr = bpage->datavaddr;
1169 if (datavaddr == 0) {
1170 tempvaddr = pmap_quick_enter_page(
1172 datavaddr = tempvaddr | bpage->dataoffs;
1175 bcopy((void *)bpage->vaddr,
1176 (void *)datavaddr, bpage->datacount);
1179 pmap_quick_remove_page(tempvaddr);
1180 bpage = STAILQ_NEXT(bpage, links);
1182 dmat->bounce_zone->total_bounced++;
1187 * Cache maintenance for normal (non-COHERENT non-bounce) buffers.
1189 if (map->sync_count != 0) {
1190 sl = &map->slist[0];
1191 end = &map->slist[map->sync_count];
1192 CTR3(KTR_BUSDMA, "%s: tag %p op 0x%x "
1193 "performing sync", __func__, dmat, op);
1195 for ( ; sl != end; ++sl)
1196 dma_dcache_sync(sl, op);
1199 if ((op & (BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE)) != 0) {
1201 * Wait for the bcopy to complete before any DMA operations.
1208 init_bounce_pages(void *dummy __unused)
1212 STAILQ_INIT(&bounce_zone_list);
1213 STAILQ_INIT(&bounce_map_waitinglist);
1214 STAILQ_INIT(&bounce_map_callbacklist);
1215 mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
1217 SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
1219 static struct sysctl_ctx_list *
1220 busdma_sysctl_tree(struct bounce_zone *bz)
1223 return (&bz->sysctl_tree);
1226 static struct sysctl_oid *
1227 busdma_sysctl_tree_top(struct bounce_zone *bz)
1230 return (bz->sysctl_tree_top);
1234 alloc_bounce_zone(bus_dma_tag_t dmat)
1236 struct bounce_zone *bz;
1238 /* Check to see if we already have a suitable zone */
1239 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1240 if ((dmat->common.alignment <= bz->alignment) &&
1241 (dmat->common.lowaddr >= bz->lowaddr)) {
1242 dmat->bounce_zone = bz;
1247 if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
1248 M_NOWAIT | M_ZERO)) == NULL)
1251 STAILQ_INIT(&bz->bounce_page_list);
1252 bz->free_bpages = 0;
1253 bz->reserved_bpages = 0;
1254 bz->active_bpages = 0;
1255 bz->lowaddr = dmat->common.lowaddr;
1256 bz->alignment = MAX(dmat->common.alignment, PAGE_SIZE);
1258 snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1260 snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1261 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1262 dmat->bounce_zone = bz;
1264 sysctl_ctx_init(&bz->sysctl_tree);
1265 bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
1266 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1267 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
1268 if (bz->sysctl_tree_top == NULL) {
1269 sysctl_ctx_free(&bz->sysctl_tree);
1270 return (0); /* XXX error code? */
1273 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1274 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1275 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1276 "Total bounce pages");
1277 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1278 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1279 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1280 "Free bounce pages");
1281 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1282 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1283 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1284 "Reserved bounce pages");
1285 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1286 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1287 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1288 "Active bounce pages");
1289 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1290 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1291 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1292 "Total bounce requests");
1293 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1294 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1295 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1296 "Total bounce requests that were deferred");
1297 SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
1298 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1299 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1300 SYSCTL_ADD_UAUTO(busdma_sysctl_tree(bz),
1301 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1302 "alignment", CTLFLAG_RD, &bz->alignment, "");
1308 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
1310 struct bounce_zone *bz;
1313 bz = dmat->bounce_zone;
1315 while (numpages > 0) {
1316 struct bounce_page *bpage;
1318 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
1323 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1324 M_NOWAIT, 0ul, bz->lowaddr, PAGE_SIZE, 0);
1325 if (bpage->vaddr == 0) {
1326 free(bpage, M_DEVBUF);
1329 bpage->busaddr = pmap_kextract(bpage->vaddr);
1330 mtx_lock(&bounce_lock);
1331 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1335 mtx_unlock(&bounce_lock);
1343 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1345 struct bounce_zone *bz;
1348 mtx_assert(&bounce_lock, MA_OWNED);
1349 bz = dmat->bounce_zone;
1350 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1351 if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
1352 return (map->pagesneeded - (map->pagesreserved + pages));
1353 bz->free_bpages -= pages;
1354 bz->reserved_bpages += pages;
1355 map->pagesreserved += pages;
1356 pages = map->pagesneeded - map->pagesreserved;
1362 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1363 bus_addr_t addr, bus_size_t size)
1365 struct bounce_zone *bz;
1366 struct bounce_page *bpage;
1368 KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
1370 bz = dmat->bounce_zone;
1371 if (map->pagesneeded == 0)
1372 panic("add_bounce_page: map doesn't need any pages");
1375 if (map->pagesreserved == 0)
1376 panic("add_bounce_page: map doesn't need any pages");
1377 map->pagesreserved--;
1379 mtx_lock(&bounce_lock);
1380 bpage = STAILQ_FIRST(&bz->bounce_page_list);
1382 panic("add_bounce_page: free page list is empty");
1384 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1385 bz->reserved_bpages--;
1386 bz->active_bpages++;
1387 mtx_unlock(&bounce_lock);
1389 if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1390 /* Page offset needs to be preserved. */
1391 bpage->vaddr |= addr & PAGE_MASK;
1392 bpage->busaddr |= addr & PAGE_MASK;
1394 bpage->datavaddr = vaddr;
1395 bpage->datapage = PHYS_TO_VM_PAGE(addr);
1396 bpage->dataoffs = addr & PAGE_MASK;
1397 bpage->datacount = size;
1398 STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1399 return (bpage->busaddr);
1403 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1405 struct bus_dmamap *map;
1406 struct bounce_zone *bz;
1408 bz = dmat->bounce_zone;
1409 bpage->datavaddr = 0;
1410 bpage->datacount = 0;
1411 if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1413 * Reset the bounce page to start at offset 0. Other uses
1414 * of this bounce page may need to store a full page of
1415 * data and/or assume it starts on a page boundary.
1417 bpage->vaddr &= ~PAGE_MASK;
1418 bpage->busaddr &= ~PAGE_MASK;
1421 mtx_lock(&bounce_lock);
1422 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1424 bz->active_bpages--;
1425 if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1426 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1427 STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1428 STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1430 busdma_swi_pending = 1;
1431 bz->total_deferred++;
1432 swi_sched(vm_ih, 0);
1435 mtx_unlock(&bounce_lock);
1442 struct bus_dmamap *map;
1444 mtx_lock(&bounce_lock);
1445 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1446 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1447 mtx_unlock(&bounce_lock);
1449 (dmat->common.lockfunc)(dmat->common.lockfuncarg, BUS_DMA_LOCK);
1450 bus_dmamap_load_mem(map->dmat, map, &map->mem,
1451 map->callback, map->callback_arg, BUS_DMA_WAITOK);
1452 (dmat->common.lockfunc)(dmat->common.lockfuncarg,
1454 mtx_lock(&bounce_lock);
1456 mtx_unlock(&bounce_lock);
1459 struct bus_dma_impl bus_dma_bounce_impl = {
1460 .tag_create = bounce_bus_dma_tag_create,
1461 .tag_destroy = bounce_bus_dma_tag_destroy,
1462 .id_mapped = bounce_bus_dma_id_mapped,
1463 .map_create = bounce_bus_dmamap_create,
1464 .map_destroy = bounce_bus_dmamap_destroy,
1465 .mem_alloc = bounce_bus_dmamem_alloc,
1466 .mem_free = bounce_bus_dmamem_free,
1467 .load_phys = bounce_bus_dmamap_load_phys,
1468 .load_buffer = bounce_bus_dmamap_load_buffer,
1469 .load_ma = bus_dmamap_load_ma_triv,
1470 .map_waitok = bounce_bus_dmamap_waitok,
1471 .map_complete = bounce_bus_dmamap_complete,
1472 .map_unload = bounce_bus_dmamap_unload,
1473 .map_sync = bounce_bus_dmamap_sync