2 * Copyright (c) 2012 Ian Lepore
3 * Copyright (c) 2010 Mark Tinguely
4 * Copyright (c) 2004 Olivier Houchard
5 * Copyright (c) 2002 Peter Grehan
6 * Copyright (c) 1997, 1998 Justin T. Gibbs.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions, and the following disclaimer,
14 * without modification, immediately at the beginning of the file.
15 * 2. The name of the author may not be used to endorse or promote products
16 * derived from this software without specific prior written permission.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
22 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * From i386/busdma_machdep.c 191438 2009-04-23 20:24:19Z jhb
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
36 #define _ARM32_BUS_DMA_PRIVATE
37 #include <sys/param.h>
40 #include <ddb/db_output.h>
41 #include <sys/systm.h>
42 #include <sys/malloc.h>
44 #include <sys/busdma_bufalloc.h>
45 #include <sys/interrupt.h>
46 #include <sys/kernel.h>
49 #include <sys/memdesc.h>
51 #include <sys/mutex.h>
52 #include <sys/sysctl.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_map.h>
58 #include <vm/vm_extern.h>
59 #include <vm/vm_kern.h>
61 #include <machine/atomic.h>
62 #include <machine/bus.h>
63 #include <machine/cpufunc.h>
64 #include <machine/md_var.h>
67 #define BUS_DMA_COULD_BOUNCE BUS_DMA_BUS3
68 #define BUS_DMA_MIN_ALLOC_COMP BUS_DMA_BUS4
78 bus_dma_filter_t *filter;
86 bus_dma_lock_t *lockfunc;
88 struct bounce_zone *bounce_zone;
90 * DMA range for this tag. If the page doesn't fall within
91 * one of these ranges, an error is returned. The caller
92 * may then decide what to do with the transfer. If the
93 * range pointer is NULL, it is ignored.
95 struct arm32_dma_range *ranges;
98 * Most tags need one or two segments, and can use the local tagsegs
99 * array. For tags with a larger limit, we'll allocate a bigger array
102 bus_dma_segment_t *segments;
103 bus_dma_segment_t tagsegs[2];
109 vm_offset_t vaddr; /* kva of bounce buffer */
110 bus_addr_t busaddr; /* Physical address */
111 vm_offset_t datavaddr; /* kva of client data */
112 bus_addr_t dataaddr; /* client physical address */
113 bus_size_t datacount; /* client data count */
114 STAILQ_ENTRY(bounce_page) links;
118 vm_offset_t vaddr; /* kva of bounce buffer */
119 bus_addr_t busaddr; /* Physical address */
120 bus_size_t datacount; /* client data count */
123 int busdma_swi_pending;
126 STAILQ_ENTRY(bounce_zone) links;
127 STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
135 bus_size_t alignment;
139 struct sysctl_ctx_list sysctl_tree;
140 struct sysctl_oid *sysctl_tree_top;
143 static struct mtx bounce_lock;
144 static int total_bpages;
145 static int busdma_zonecount;
146 static STAILQ_HEAD(, bounce_zone) bounce_zone_list;
148 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
149 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
150 "Total bounce pages");
153 struct bp_list bpages;
159 bus_dmamap_callback_t *callback;
162 #define DMAMAP_COHERENT (1 << 0)
163 STAILQ_ENTRY(bus_dmamap) links;
165 struct sync_list slist[];
168 static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
169 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
171 static void init_bounce_pages(void *dummy);
172 static int alloc_bounce_zone(bus_dma_tag_t dmat);
173 static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
174 static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
176 static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
177 vm_offset_t vaddr, bus_addr_t addr,
179 static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
180 int run_filter(bus_dma_tag_t dmat, bus_addr_t paddr, bus_size_t size, int coherent);
181 static void _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
182 void *buf, bus_size_t buflen, int flags);
183 static void _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
184 vm_paddr_t buf, bus_size_t buflen, int flags);
185 static int _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
188 static busdma_bufalloc_t coherent_allocator; /* Cache of coherent buffers */
189 static busdma_bufalloc_t standard_allocator; /* Cache of standard buffers */
191 busdma_init(void *dummy)
197 /* Create a cache of buffers in standard (cacheable) memory. */
198 standard_allocator = busdma_bufalloc_create("buffer",
199 arm_dcache_align, /* minimum_alignment */
200 NULL, /* uma_alloc func */
201 NULL, /* uma_free func */
202 uma_flags); /* uma_zcreate_flags */
206 * Force UMA zone to allocate service structures like
207 * slabs using own allocator. uma_debug code performs
208 * atomic ops on uma_slab_t fields and safety of this
209 * operation is not guaranteed for write-back caches
211 uma_flags = UMA_ZONE_OFFPAGE;
214 * Create a cache of buffers in uncacheable memory, to implement the
215 * BUS_DMA_COHERENT (and potentially BUS_DMA_NOCACHE) flag.
217 coherent_allocator = busdma_bufalloc_create("coherent",
218 arm_dcache_align, /* minimum_alignment */
219 busdma_bufalloc_alloc_uncacheable,
220 busdma_bufalloc_free_uncacheable,
221 uma_flags); /* uma_zcreate_flags */
225 * This init historically used SI_SUB_VM, but now the init code requires
226 * malloc(9) using M_DEVBUF memory, which is set up later than SI_SUB_VM, by
227 * SI_SUB_KMEM and SI_ORDER_SECOND, so we'll go right after that by using
228 * SI_SUB_KMEM and SI_ORDER_THIRD.
230 SYSINIT(busdma, SI_SUB_KMEM, SI_ORDER_THIRD, busdma_init, NULL);
233 _bus_dma_can_bounce(vm_offset_t lowaddr, vm_offset_t highaddr)
236 for (i = 0; phys_avail[i] && phys_avail[i + 1]; i += 2) {
237 if ((lowaddr >= phys_avail[i] && lowaddr <= phys_avail[i + 1])
238 || (lowaddr < phys_avail[i] &&
239 highaddr > phys_avail[i]))
245 static __inline struct arm32_dma_range *
246 _bus_dma_inrange(struct arm32_dma_range *ranges, int nranges,
249 struct arm32_dma_range *dr;
252 for (i = 0, dr = ranges; i < nranges; i++, dr++) {
253 if (curaddr >= dr->dr_sysbase &&
254 round_page(curaddr) <= (dr->dr_sysbase + dr->dr_len))
262 * Return true if a match is made.
264 * To find a match walk the chain of bus_dma_tag_t's looking for 'paddr'.
266 * If paddr is within the bounds of the dma tag then call the filter callback
267 * to check for a match, if there is no filter callback then assume a match.
270 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr, bus_size_t size, int coherent)
277 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr)
278 || ((paddr & (dmat->alignment - 1)) != 0) ||
279 (!coherent && (size & arm_dcache_align_mask)) ||
280 (!coherent && (paddr & arm_dcache_align_mask)))
281 && (dmat->filter == NULL
282 || (*dmat->filter)(dmat->filterarg, paddr) != 0))
286 } while (retval == 0 && dmat != NULL);
291 * Convenience function for manipulating driver locks from busdma (during
292 * busdma_swi, for example). Drivers that don't provide their own locks
293 * should specify &Giant to dmat->lockfuncarg. Drivers that use their own
294 * non-mutex locking scheme don't have to use this at all.
297 busdma_lock_mutex(void *arg, bus_dma_lock_op_t op)
301 dmtx = (struct mtx *)arg;
310 panic("Unknown operation 0x%x for busdma_lock_mutex!", op);
315 * dflt_lock should never get called. It gets put into the dma tag when
316 * lockfunc == NULL, which is only valid if the maps that are associated
317 * with the tag are meant to never be defered.
318 * XXX Should have a way to identify which driver is responsible here.
321 dflt_lock(void *arg, bus_dma_lock_op_t op)
323 panic("driver error: busdma dflt_lock called");
327 * Allocate a device specific dma_tag.
330 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
331 bus_size_t boundary, bus_addr_t lowaddr,
332 bus_addr_t highaddr, bus_dma_filter_t *filter,
333 void *filterarg, bus_size_t maxsize, int nsegments,
334 bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
335 void *lockfuncarg, bus_dma_tag_t *dmat)
337 bus_dma_tag_t newtag;
342 parent = arm_root_dma_tag;
345 /* Basic sanity checking */
346 if (boundary != 0 && boundary < maxsegsz)
349 /* Return a NULL tag on failure */
356 newtag = (bus_dma_tag_t)malloc(sizeof(*newtag), M_DEVBUF,
358 if (newtag == NULL) {
359 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
360 __func__, newtag, 0, error);
364 newtag->parent = parent;
365 newtag->alignment = alignment;
366 newtag->boundary = boundary;
367 newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
368 newtag->highaddr = trunc_page((vm_paddr_t)highaddr) +
370 newtag->filter = filter;
371 newtag->filterarg = filterarg;
372 newtag->maxsize = maxsize;
373 newtag->nsegments = nsegments;
374 newtag->maxsegsz = maxsegsz;
375 newtag->flags = flags;
376 newtag->ref_count = 1; /* Count ourself */
377 newtag->map_count = 0;
378 newtag->ranges = bus_dma_get_range();
379 newtag->_nranges = bus_dma_get_range_nb();
380 if (lockfunc != NULL) {
381 newtag->lockfunc = lockfunc;
382 newtag->lockfuncarg = lockfuncarg;
384 newtag->lockfunc = dflt_lock;
385 newtag->lockfuncarg = NULL;
388 * If all the segments we need fit into the local tagsegs array, set the
389 * pointer now. Otherwise NULL the pointer and an array of segments
390 * will be allocated later, on first use. We don't pre-allocate now
391 * because some tags exist just to pass contraints to children in the
392 * device hierarchy, and they tend to use BUS_SPACE_UNRESTRICTED and we
393 * sure don't want to try to allocate an array for that.
395 if (newtag->nsegments <= nitems(newtag->tagsegs))
396 newtag->segments = newtag->tagsegs;
398 newtag->segments = NULL;
400 /* Take into account any restrictions imposed by our parent tag */
401 if (parent != NULL) {
402 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
403 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
404 if (newtag->boundary == 0)
405 newtag->boundary = parent->boundary;
406 else if (parent->boundary != 0)
407 newtag->boundary = MIN(parent->boundary,
409 if ((newtag->filter != NULL) ||
410 ((parent->flags & BUS_DMA_COULD_BOUNCE) != 0))
411 newtag->flags |= BUS_DMA_COULD_BOUNCE;
412 if (newtag->filter == NULL) {
414 * Short circuit looking at our parent directly
415 * since we have encapsulated all of its information
417 newtag->filter = parent->filter;
418 newtag->filterarg = parent->filterarg;
419 newtag->parent = parent->parent;
421 if (newtag->parent != NULL)
422 atomic_add_int(&parent->ref_count, 1);
425 if (_bus_dma_can_bounce(newtag->lowaddr, newtag->highaddr)
426 || newtag->alignment > 1)
427 newtag->flags |= BUS_DMA_COULD_BOUNCE;
429 maxsize = 2; /* Need at most 2 bounce pages for unaligned access on cache line boundaries */
431 if ((flags & BUS_DMA_ALLOCNOW) != 0) {
432 struct bounce_zone *bz;
436 if ((error = alloc_bounce_zone(newtag)) != 0) {
437 free(newtag, M_DEVBUF);
440 bz = newtag->bounce_zone;
442 if (ptoa(bz->total_bpages) < maxsize) {
445 pages = atop(maxsize) - bz->total_bpages;
447 /* Add pages to our bounce pool */
448 if (alloc_bounce_pages(newtag, pages) < pages)
451 /* Performed initial allocation */
452 newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
454 newtag->bounce_zone = NULL;
457 free(newtag, M_DEVBUF);
461 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
462 __func__, newtag, (newtag != NULL ? newtag->flags : 0), error);
467 bus_dma_tag_destroy(bus_dma_tag_t dmat)
469 bus_dma_tag_t dmat_copy;
477 if (dmat->map_count != 0) {
482 while (dmat != NULL) {
483 bus_dma_tag_t parent;
485 parent = dmat->parent;
486 atomic_subtract_int(&dmat->ref_count, 1);
487 if (dmat->ref_count == 0) {
488 if (dmat->segments != NULL &&
489 dmat->segments != dmat->tagsegs)
490 free(dmat->segments, M_DEVBUF);
491 free(dmat, M_DEVBUF);
493 * Last reference count, so
494 * release our reference
495 * count on our parent.
503 CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
507 static int allocate_bz_and_pages(bus_dma_tag_t dmat, bus_dmamap_t mapp)
509 struct bounce_zone *bz;
513 if (dmat->bounce_zone == NULL)
514 if ((error = alloc_bounce_zone(dmat)) != 0)
516 bz = dmat->bounce_zone;
517 /* Initialize the new map */
518 STAILQ_INIT(&(mapp->bpages));
521 * Attempt to add pages to our pool on a per-instance
522 * basis up to a sane limit.
524 if (dmat->flags & BUS_DMA_COULD_BOUNCE)
525 maxpages = MAX_BPAGES;
527 maxpages = 2 * bz->map_count; /* Only need at most 2 pages for buffers unaligned on cache line boundaries */
528 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0
529 || (bz->map_count > 0 && bz->total_bpages < maxpages)) {
532 pages = MAX(atop(dmat->maxsize), 1);
533 pages = MIN(maxpages - bz->total_bpages, pages);
534 pages = MAX(pages, 1);
535 if (alloc_bounce_pages(dmat, pages) < pages)
538 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0)
539 dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
546 * Allocate a handle for mapping from kva/uva/physical
547 * address space into bus device space.
550 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
555 mapsize = sizeof(**mapp) + (sizeof(struct sync_list) * dmat->nsegments);
556 *mapp = (bus_dmamap_t)malloc(mapsize, M_DEVBUF, M_NOWAIT | M_ZERO);
558 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, ENOMEM);
561 (*mapp)->sync_count = 0;
563 if (dmat->segments == NULL) {
564 dmat->segments = (bus_dma_segment_t *)malloc(
565 sizeof(bus_dma_segment_t) * dmat->nsegments, M_DEVBUF,
567 if (dmat->segments == NULL) {
568 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
569 __func__, dmat, ENOMEM);
570 free(*mapp, M_DEVBUF);
576 * Bouncing might be required if the driver asks for an active
577 * exclusion region, a data alignment that is stricter than 1, and/or
578 * an active address boundary.
580 error = allocate_bz_and_pages(dmat, *mapp);
582 free(*mapp, M_DEVBUF);
590 * Destroy a handle for mapping from kva/uva/physical
591 * address space into bus device space.
594 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
596 if (STAILQ_FIRST(&map->bpages) != NULL || map->sync_count != 0) {
597 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
598 __func__, dmat, EBUSY);
601 if (dmat->bounce_zone)
602 dmat->bounce_zone->map_count--;
605 CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
611 * Allocate a piece of memory that can be efficiently mapped into
612 * bus device space based on the constraints lited in the dma tag.
613 * A dmamap to for use with dmamap_load is also allocated.
616 bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
619 busdma_bufalloc_t ba;
620 struct busdma_bufzone *bufzone;
621 vm_memattr_t memattr;
626 if (flags & BUS_DMA_NOWAIT)
631 /* ARM non-snooping caches need a map for the VA cache sync structure */
633 mapsize = sizeof(**mapp) + (sizeof(struct sync_list) * dmat->nsegments);
634 *mapp = (bus_dmamap_t)malloc(mapsize, M_DEVBUF, M_NOWAIT | M_ZERO);
636 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
637 __func__, dmat, dmat->flags, ENOMEM);
641 (*mapp)->sync_count = 0;
642 /* We may need bounce pages, even for allocated memory */
643 error = allocate_bz_and_pages(dmat, *mapp);
645 free(*mapp, M_DEVBUF);
650 if (dmat->segments == NULL) {
651 dmat->segments = (bus_dma_segment_t *)malloc(
652 sizeof(bus_dma_segment_t) * dmat->nsegments, M_DEVBUF,
654 if (dmat->segments == NULL) {
655 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
656 __func__, dmat, dmat->flags, ENOMEM);
657 free(*mapp, M_DEVBUF);
663 if (flags & BUS_DMA_ZERO)
665 if (flags & BUS_DMA_COHERENT) {
666 memattr = VM_MEMATTR_UNCACHEABLE;
667 ba = coherent_allocator;
668 (*mapp)->flags |= DMAMAP_COHERENT;
670 memattr = VM_MEMATTR_DEFAULT;
671 ba = standard_allocator;
676 * Try to find a bufzone in the allocator that holds a cache of buffers
677 * of the right size for this request. If the buffer is too big to be
678 * held in the allocator cache, this returns NULL.
680 bufzone = busdma_bufalloc_findzone(ba, dmat->maxsize);
683 * Allocate the buffer from the uma(9) allocator if...
684 * - It's small enough to be in the allocator (bufzone not NULL).
685 * - The alignment constraint isn't larger than the allocation size
686 * (the allocator aligns buffers to their size boundaries).
687 * - There's no need to handle lowaddr/highaddr exclusion zones.
688 * else allocate non-contiguous pages if...
689 * - The page count that could get allocated doesn't exceed nsegments.
690 * - The alignment constraint isn't larger than a page boundary.
691 * - There are no boundary-crossing constraints.
692 * else allocate a block of contiguous pages because one or more of the
693 * constraints is something that only the contig allocator can fulfill.
695 if (bufzone != NULL && dmat->alignment <= bufzone->size &&
696 !_bus_dma_can_bounce(dmat->lowaddr, dmat->highaddr)) {
697 *vaddr = uma_zalloc(bufzone->umazone, mflags);
698 } else if (dmat->nsegments >= btoc(dmat->maxsize) &&
699 dmat->alignment <= PAGE_SIZE && dmat->boundary == 0) {
700 *vaddr = (void *)kmem_alloc_attr(kernel_arena, dmat->maxsize,
701 mflags, 0, dmat->lowaddr, memattr);
703 *vaddr = (void *)kmem_alloc_contig(kernel_arena, dmat->maxsize,
704 mflags, 0, dmat->lowaddr, dmat->alignment, dmat->boundary,
709 if (*vaddr == NULL) {
710 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
711 __func__, dmat, dmat->flags, ENOMEM);
712 free(*mapp, M_DEVBUF);
715 } else if ((uintptr_t)*vaddr & (dmat->alignment - 1)) {
716 printf("bus_dmamem_alloc failed to align memory properly.\n");
720 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
721 __func__, dmat, dmat->flags, 0);
726 * Free a piece of memory and it's allociated dmamap, that was allocated
727 * via bus_dmamem_alloc. Make the same choice for free/contigfree.
730 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
732 struct busdma_bufzone *bufzone;
733 busdma_bufalloc_t ba;
735 if (map->flags & DMAMAP_COHERENT)
736 ba = coherent_allocator;
738 ba = standard_allocator;
740 /* Be careful not to access map from here on. */
742 bufzone = busdma_bufalloc_findzone(ba, dmat->maxsize);
744 if (bufzone != NULL && dmat->alignment <= bufzone->size &&
745 !_bus_dma_can_bounce(dmat->lowaddr, dmat->highaddr))
746 uma_zfree(bufzone->umazone, vaddr);
748 kmem_free(kernel_arena, (vm_offset_t)vaddr, dmat->maxsize);
752 CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat, dmat->flags);
756 _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
757 bus_size_t buflen, int flags)
762 if (map->pagesneeded == 0) {
763 CTR5(KTR_BUSDMA, "lowaddr= %d, boundary= %d, alignment= %d"
764 " map= %p, pagesneeded= %d",
765 dmat->lowaddr, dmat->boundary, dmat->alignment,
766 map, map->pagesneeded);
768 * Count the number of bounce pages
769 * needed in order to complete this transfer
772 while (buflen != 0) {
773 sgsize = MIN(buflen, dmat->maxsegsz);
774 if (run_filter(dmat, curaddr, sgsize,
775 map->flags & DMAMAP_COHERENT) != 0) {
776 sgsize = MIN(sgsize, PAGE_SIZE);
782 CTR1(KTR_BUSDMA, "pagesneeded= %d", map->pagesneeded);
787 _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
788 void *buf, bus_size_t buflen, int flags)
791 vm_offset_t vendaddr;
794 if (map->pagesneeded == 0) {
795 CTR5(KTR_BUSDMA, "lowaddr= %d, boundary= %d, alignment= %d"
796 " map= %p, pagesneeded= %d",
797 dmat->lowaddr, dmat->boundary, dmat->alignment,
798 map, map->pagesneeded);
800 * Count the number of bounce pages
801 * needed in order to complete this transfer
803 vaddr = (vm_offset_t)buf;
804 vendaddr = (vm_offset_t)buf + buflen;
806 while (vaddr < vendaddr) {
807 if (__predict_true(map->pmap == kernel_pmap))
808 paddr = pmap_kextract(vaddr);
810 paddr = pmap_extract(map->pmap, vaddr);
811 if (run_filter(dmat, paddr,
812 min(vendaddr - vaddr,
813 (PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK))),
814 map->flags & DMAMAP_COHERENT) != 0) {
817 vaddr += (PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK));
820 CTR1(KTR_BUSDMA, "pagesneeded= %d", map->pagesneeded);
825 _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags)
828 /* Reserve Necessary Bounce Pages */
829 mtx_lock(&bounce_lock);
830 if (flags & BUS_DMA_NOWAIT) {
831 if (reserve_bounce_pages(dmat, map, 0) != 0) {
832 map->pagesneeded = 0;
833 mtx_unlock(&bounce_lock);
837 if (reserve_bounce_pages(dmat, map, 1) != 0) {
838 /* Queue us for resources */
839 STAILQ_INSERT_TAIL(&bounce_map_waitinglist, map, links);
840 mtx_unlock(&bounce_lock);
841 return (EINPROGRESS);
844 mtx_unlock(&bounce_lock);
850 * Add a single contiguous physical range to the segment list.
853 _bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr,
854 bus_size_t sgsize, bus_dma_segment_t *segs, int *segp)
856 bus_addr_t baddr, bmask;
860 * Make sure we don't cross any boundaries.
862 bmask = ~(dmat->boundary - 1);
863 if (dmat->boundary > 0) {
864 baddr = (curaddr + dmat->boundary) & bmask;
865 if (sgsize > (baddr - curaddr))
866 sgsize = (baddr - curaddr);
870 struct arm32_dma_range *dr;
872 dr = _bus_dma_inrange(dmat->ranges, dmat->_nranges,
875 _bus_dmamap_unload(dmat, map);
879 * In a valid DMA range. Translate the physical
880 * memory address to an address in the DMA window.
882 curaddr = (curaddr - dr->dr_sysbase) + dr->dr_busbase;
886 * Insert chunk into a segment, coalescing with
887 * previous segment if possible.
892 segs[seg].ds_addr = curaddr;
893 segs[seg].ds_len = sgsize;
895 if (curaddr == segs[seg].ds_addr + segs[seg].ds_len &&
896 (segs[seg].ds_len + sgsize) <= dmat->maxsegsz &&
897 (dmat->boundary == 0 ||
898 (segs[seg].ds_addr & bmask) == (curaddr & bmask)))
899 segs[seg].ds_len += sgsize;
901 if (++seg >= dmat->nsegments)
903 segs[seg].ds_addr = curaddr;
904 segs[seg].ds_len = sgsize;
912 * Utility function to load a physical buffer. segp contains
913 * the starting segment on entrace, and the ending segment on exit.
916 _bus_dmamap_load_phys(bus_dma_tag_t dmat,
918 vm_paddr_t buf, bus_size_t buflen,
920 bus_dma_segment_t *segs,
928 segs = dmat->segments;
930 if (((map->flags & DMAMAP_COHERENT) == 0) ||
931 (dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) {
932 _bus_dmamap_count_phys(dmat, map, buf, buflen, flags);
933 if (map->pagesneeded != 0) {
934 error = _bus_dmamap_reserve_pages(dmat, map, flags);
942 sgsize = MIN(buflen, dmat->maxsegsz);
943 if ((((map->flags & DMAMAP_COHERENT) == 0) ||
944 ((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0)) &&
945 map->pagesneeded != 0 && run_filter(dmat, curaddr,
946 sgsize, map->flags & DMAMAP_COHERENT)) {
947 sgsize = MIN(sgsize, PAGE_SIZE);
948 curaddr = add_bounce_page(dmat, map, 0, curaddr,
951 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
963 _bus_dmamap_unload(dmat, map);
964 return (EFBIG); /* XXX better return value here? */
970 * Utility function to load a linear buffer. segp contains
971 * the starting segment on entrace, and the ending segment on exit.
974 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
976 void *buf, bus_size_t buflen,
979 bus_dma_segment_t *segs,
985 struct sync_list *sl;
989 segs = dmat->segments;
993 if (!(map->flags & DMAMAP_COHERENT) ||
994 (dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) {
995 _bus_dmamap_count_pages(dmat, map, buf, buflen, flags);
996 if (map->pagesneeded != 0) {
997 error = _bus_dmamap_reserve_pages(dmat, map, flags);
1004 vaddr = (vm_offset_t)buf;
1006 while (buflen > 0) {
1008 * Get the physical address for this segment.
1010 if (__predict_true(map->pmap == kernel_pmap))
1011 curaddr = pmap_kextract(vaddr);
1013 curaddr = pmap_extract(map->pmap, vaddr);
1016 * Compute the segment size, and adjust counts.
1018 sgsize = PAGE_SIZE - ((u_long)curaddr & PAGE_MASK);
1019 if (sgsize > dmat->maxsegsz)
1020 sgsize = dmat->maxsegsz;
1021 if (buflen < sgsize)
1024 if ((((map->flags & DMAMAP_COHERENT) == 0) ||
1025 ((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0)) &&
1026 map->pagesneeded != 0 && run_filter(dmat, curaddr,
1027 sgsize, map->flags & DMAMAP_COHERENT)) {
1028 curaddr = add_bounce_page(dmat, map, vaddr, curaddr,
1031 sl = &map->slist[map->sync_count - 1];
1032 if (map->sync_count == 0 ||
1034 curaddr != sl->busaddr + sl->datacount ||
1036 vaddr != sl->vaddr + sl->datacount) {
1037 if (++map->sync_count > dmat->nsegments)
1041 sl->datacount = sgsize;
1042 sl->busaddr = curaddr;
1044 sl->datacount += sgsize;
1046 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
1059 _bus_dmamap_unload(dmat, map);
1060 return (EFBIG); /* XXX better return value here? */
1067 __bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map,
1068 struct memdesc *mem, bus_dmamap_callback_t *callback,
1074 map->callback = callback;
1075 map->callback_arg = callback_arg;
1079 _bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map,
1080 bus_dma_segment_t *segs, int nsegs, int error)
1084 segs = dmat->segments;
1089 * Release the mapping held by map.
1092 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1094 struct bounce_page *bpage;
1095 struct bounce_zone *bz;
1097 if ((bz = dmat->bounce_zone) != NULL) {
1098 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1099 STAILQ_REMOVE_HEAD(&map->bpages, links);
1100 free_bounce_page(dmat, bpage);
1103 bz = dmat->bounce_zone;
1104 bz->free_bpages += map->pagesreserved;
1105 bz->reserved_bpages -= map->pagesreserved;
1106 map->pagesreserved = 0;
1107 map->pagesneeded = 0;
1109 map->sync_count = 0;
1112 #ifdef notyetbounceuser
1113 /* If busdma uses user pages, then the interrupt handler could
1114 * be use the kernel vm mapping. Both bounce pages and sync list
1115 * do not cross page boundaries.
1116 * Below is a rough sequence that a person would do to fix the
1117 * user page reference in the kernel vmspace. This would be
1118 * done in the dma post routine.
1121 _bus_dmamap_fix_user(vm_offset_t buf, bus_size_t len,
1122 pmap_t pmap, int op)
1128 /* each synclist entry is contained within a single page.
1130 * this would be needed if BUS_DMASYNC_POSTxxxx was implemented
1132 curaddr = pmap_extract(pmap, buf);
1133 va = pmap_dma_map(curaddr);
1136 cpu_dcache_wb_range(va, sgsize);
1139 case SYNC_USER_COPYTO:
1140 bcopy((void *)va, (void *)bounce, sgsize);
1143 case SYNC_USER_COPYFROM:
1144 bcopy((void *) bounce, (void *)va, sgsize);
1156 #define l2cache_wb_range(va, pa, size) cpu_l2cache_wb_range(pa, size)
1157 #define l2cache_wbinv_range(va, pa, size) cpu_l2cache_wbinv_range(pa, size)
1158 #define l2cache_inv_range(va, pa, size) cpu_l2cache_inv_range(pa, size)
1160 #define l2cache_wb_range(va, pa, size) cpu_l2cache_wb_range(va, size)
1161 #define l2cache_wbinv_range(va, pa, size) cpu_l2cache_wbinv_range(va, size)
1162 #define l2cache_inv_range(va, pa, size) cpu_l2cache_inv_range(va, size)
1166 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1168 struct bounce_page *bpage;
1169 struct sync_list *sl, *end;
1171 * If the buffer was from user space, it is possible that this is not
1172 * the same vm map, especially on a POST operation. It's not clear that
1173 * dma on userland buffers can work at all right now, certainly not if a
1174 * partial cacheline flush has to be handled. To be safe, until we're
1175 * able to test direct userland dma, panic on a map mismatch.
1177 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1178 if (!pmap_dmap_iscurrent(map->pmap))
1179 panic("_bus_dmamap_sync: wrong user map for bounce sync.");
1180 /* Handle data bouncing. */
1181 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
1182 "performing bounce", __func__, dmat, dmat->flags, op);
1184 if (op & BUS_DMASYNC_PREWRITE) {
1185 while (bpage != NULL) {
1186 if (bpage->datavaddr != 0)
1187 bcopy((void *)bpage->datavaddr,
1188 (void *)bpage->vaddr,
1191 physcopyout(bpage->dataaddr,
1192 (void *)bpage->vaddr,
1194 cpu_dcache_wb_range((vm_offset_t)bpage->vaddr,
1196 l2cache_wb_range((vm_offset_t)bpage->vaddr,
1197 (vm_offset_t)bpage->busaddr,
1199 bpage = STAILQ_NEXT(bpage, links);
1201 dmat->bounce_zone->total_bounced++;
1204 if (op & BUS_DMASYNC_POSTREAD) {
1205 while (bpage != NULL) {
1210 startv = bpage->vaddr &~ arm_dcache_align_mask;
1211 startp = bpage->busaddr &~ arm_dcache_align_mask;
1212 len = bpage->datacount;
1214 if (startv != bpage->vaddr)
1215 len += bpage->vaddr & arm_dcache_align_mask;
1216 if (len & arm_dcache_align_mask)
1218 (len & arm_dcache_align_mask)) +
1220 cpu_dcache_inv_range(startv, len);
1221 l2cache_inv_range(startv, startp, len);
1222 if (bpage->datavaddr != 0)
1223 bcopy((void *)bpage->vaddr,
1224 (void *)bpage->datavaddr,
1227 physcopyin((void *)bpage->vaddr,
1230 bpage = STAILQ_NEXT(bpage, links);
1232 dmat->bounce_zone->total_bounced++;
1235 if (map->flags & DMAMAP_COHERENT)
1238 if (map->sync_count != 0) {
1239 if (!pmap_dmap_iscurrent(map->pmap))
1240 panic("_bus_dmamap_sync: wrong user map for sync.");
1241 /* ARM caches are not self-snooping for dma */
1243 sl = &map->slist[0];
1244 end = &map->slist[map->sync_count];
1245 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
1246 "performing sync", __func__, dmat, dmat->flags, op);
1249 case BUS_DMASYNC_PREWRITE:
1251 cpu_dcache_wb_range(sl->vaddr, sl->datacount);
1252 l2cache_wb_range(sl->vaddr, sl->busaddr,
1258 case BUS_DMASYNC_PREREAD:
1260 cpu_dcache_inv_range(sl->vaddr, sl->datacount);
1261 l2cache_inv_range(sl->vaddr, sl->busaddr,
1267 case BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD:
1269 cpu_dcache_wbinv_range(sl->vaddr, sl->datacount);
1270 l2cache_wbinv_range(sl->vaddr,
1271 sl->busaddr, sl->datacount);
1283 init_bounce_pages(void *dummy __unused)
1287 STAILQ_INIT(&bounce_zone_list);
1288 STAILQ_INIT(&bounce_map_waitinglist);
1289 STAILQ_INIT(&bounce_map_callbacklist);
1290 mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
1292 SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
1294 static struct sysctl_ctx_list *
1295 busdma_sysctl_tree(struct bounce_zone *bz)
1297 return (&bz->sysctl_tree);
1300 static struct sysctl_oid *
1301 busdma_sysctl_tree_top(struct bounce_zone *bz)
1303 return (bz->sysctl_tree_top);
1307 alloc_bounce_zone(bus_dma_tag_t dmat)
1309 struct bounce_zone *bz;
1311 /* Check to see if we already have a suitable zone */
1312 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1313 if ((dmat->alignment <= bz->alignment)
1314 && (dmat->lowaddr >= bz->lowaddr)) {
1315 dmat->bounce_zone = bz;
1320 if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
1321 M_NOWAIT | M_ZERO)) == NULL)
1324 STAILQ_INIT(&bz->bounce_page_list);
1325 bz->free_bpages = 0;
1326 bz->reserved_bpages = 0;
1327 bz->active_bpages = 0;
1328 bz->lowaddr = dmat->lowaddr;
1329 bz->alignment = MAX(dmat->alignment, PAGE_SIZE);
1331 snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1333 snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1334 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1335 dmat->bounce_zone = bz;
1337 sysctl_ctx_init(&bz->sysctl_tree);
1338 bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
1339 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1341 if (bz->sysctl_tree_top == NULL) {
1342 sysctl_ctx_free(&bz->sysctl_tree);
1343 return (0); /* XXX error code? */
1346 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1347 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1348 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1349 "Total bounce pages");
1350 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1351 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1352 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1353 "Free bounce pages");
1354 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1355 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1356 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1357 "Reserved bounce pages");
1358 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1359 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1360 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1361 "Active bounce pages");
1362 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1363 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1364 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1365 "Total bounce requests");
1366 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1367 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1368 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1369 "Total bounce requests that were deferred");
1370 SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
1371 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1372 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1373 SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1374 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1375 "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1381 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
1383 struct bounce_zone *bz;
1386 bz = dmat->bounce_zone;
1388 while (numpages > 0) {
1389 struct bounce_page *bpage;
1391 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
1396 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1401 if (bpage->vaddr == 0) {
1402 free(bpage, M_DEVBUF);
1405 bpage->busaddr = pmap_kextract(bpage->vaddr);
1406 mtx_lock(&bounce_lock);
1407 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1411 mtx_unlock(&bounce_lock);
1419 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1421 struct bounce_zone *bz;
1424 mtx_assert(&bounce_lock, MA_OWNED);
1425 bz = dmat->bounce_zone;
1426 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1427 if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
1428 return (map->pagesneeded - (map->pagesreserved + pages));
1429 bz->free_bpages -= pages;
1430 bz->reserved_bpages += pages;
1431 map->pagesreserved += pages;
1432 pages = map->pagesneeded - map->pagesreserved;
1438 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1439 bus_addr_t addr, bus_size_t size)
1441 struct bounce_zone *bz;
1442 struct bounce_page *bpage;
1444 KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
1445 KASSERT(map != NULL,
1446 ("add_bounce_page: bad map %p", map));
1448 bz = dmat->bounce_zone;
1449 if (map->pagesneeded == 0)
1450 panic("add_bounce_page: map doesn't need any pages");
1453 if (map->pagesreserved == 0)
1454 panic("add_bounce_page: map doesn't need any pages");
1455 map->pagesreserved--;
1457 mtx_lock(&bounce_lock);
1458 bpage = STAILQ_FIRST(&bz->bounce_page_list);
1460 panic("add_bounce_page: free page list is empty");
1462 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1463 bz->reserved_bpages--;
1464 bz->active_bpages++;
1465 mtx_unlock(&bounce_lock);
1467 if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1468 /* Page offset needs to be preserved. */
1469 bpage->vaddr |= vaddr & PAGE_MASK;
1470 bpage->busaddr |= vaddr & PAGE_MASK;
1472 bpage->datavaddr = vaddr;
1473 bpage->dataaddr = addr;
1474 bpage->datacount = size;
1475 STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1476 return (bpage->busaddr);
1480 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1482 struct bus_dmamap *map;
1483 struct bounce_zone *bz;
1485 bz = dmat->bounce_zone;
1486 bpage->datavaddr = 0;
1487 bpage->datacount = 0;
1488 if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1490 * Reset the bounce page to start at offset 0. Other uses
1491 * of this bounce page may need to store a full page of
1492 * data and/or assume it starts on a page boundary.
1494 bpage->vaddr &= ~PAGE_MASK;
1495 bpage->busaddr &= ~PAGE_MASK;
1498 mtx_lock(&bounce_lock);
1499 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1501 bz->active_bpages--;
1502 if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1503 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1504 STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1505 STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1507 busdma_swi_pending = 1;
1508 bz->total_deferred++;
1509 swi_sched(vm_ih, 0);
1512 mtx_unlock(&bounce_lock);
1519 struct bus_dmamap *map;
1521 mtx_lock(&bounce_lock);
1522 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1523 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1524 mtx_unlock(&bounce_lock);
1526 (dmat->lockfunc)(dmat->lockfuncarg, BUS_DMA_LOCK);
1527 bus_dmamap_load_mem(map->dmat, map, &map->mem, map->callback,
1528 map->callback_arg, BUS_DMA_WAITOK);
1529 (dmat->lockfunc)(dmat->lockfuncarg, BUS_DMA_UNLOCK);
1530 mtx_lock(&bounce_lock);
1532 mtx_unlock(&bounce_lock);