1/*-
2 * Copyright (c) 1997, 1998 Justin T. Gibbs.
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions, and the following disclaimer,
10 *    without modification, immediately at the beginning of the file.
11 * 2. The name of the author may not be used to endorse or promote products
12 *    derived from this software without specific prior written permission.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
18 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27/*
28 * From amd64/busdma_machdep.c, r204214
29 */
30
31#include <sys/cdefs.h>
32__FBSDID("$FreeBSD$");
33
34#include <sys/param.h>
35#include <sys/systm.h>
36#include <sys/malloc.h>
37#include <sys/bus.h>
38#include <sys/interrupt.h>
39#include <sys/kernel.h>
40#include <sys/ktr.h>
41#include <sys/lock.h>
42#include <sys/proc.h>
43#include <sys/memdesc.h>
44#include <sys/mutex.h>
45#include <sys/sysctl.h>
46#include <sys/uio.h>
47
48#include <vm/vm.h>
49#include <vm/vm_extern.h>
50#include <vm/vm_kern.h>
51#include <vm/vm_page.h>
52#include <vm/vm_map.h>
53
54#include <machine/atomic.h>
55#include <machine/bus.h>
56#include <machine/md_var.h>
57
58#include "iommu_if.h"
59
60#define MAX_BPAGES MIN(8192, physmem/40)
61
62struct bounce_zone;
63
64struct bus_dma_tag {
65	bus_dma_tag_t	  parent;
66	bus_size_t	  alignment;
67	bus_size_t	  boundary;
68	bus_addr_t	  lowaddr;
69	bus_addr_t	  highaddr;
70	bus_dma_filter_t *filter;
71	void		 *filterarg;
72	bus_size_t	  maxsize;
73	u_int		  nsegments;
74	bus_size_t	  maxsegsz;
75	int		  flags;
76	int		  ref_count;
77	int		  map_count;
78	bus_dma_lock_t	 *lockfunc;
79	void		 *lockfuncarg;
80	struct bounce_zone *bounce_zone;
81	device_t	  iommu;
82	void		 *iommu_cookie;
83};
84
85struct bounce_page {
86	vm_offset_t	vaddr;		/* kva of bounce buffer */
87	bus_addr_t	busaddr;	/* Physical address */
88	vm_offset_t	datavaddr;	/* kva of client data */
89	bus_addr_t	dataaddr;	/* client physical address */
90	bus_size_t	datacount;	/* client data count */
91	STAILQ_ENTRY(bounce_page) links;
92};
93
94int busdma_swi_pending;
95
96struct bounce_zone {
97	STAILQ_ENTRY(bounce_zone) links;
98	STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
99	int		total_bpages;
100	int		free_bpages;
101	int		reserved_bpages;
102	int		active_bpages;
103	int		total_bounced;
104	int		total_deferred;
105	int		map_count;
106	bus_size_t	alignment;
107	bus_addr_t	lowaddr;
108	char		zoneid[8];
109	char		lowaddrid[20];
110	struct sysctl_ctx_list sysctl_tree;
111	struct sysctl_oid *sysctl_tree_top;
112};
113
114static struct mtx bounce_lock;
115static int total_bpages;
116static int busdma_zonecount;
117static STAILQ_HEAD(, bounce_zone) bounce_zone_list;
118
119static SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
120SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
121	   "Total bounce pages");
122
123struct bus_dmamap {
124	struct bp_list	       bpages;
125	int		       pagesneeded;
126	int		       pagesreserved;
127	bus_dma_tag_t	       dmat;
128	struct memdesc	       mem;
129	bus_dma_segment_t     *segments;
130	int		       nsegs;
131	bus_dmamap_callback_t *callback;
132	void		      *callback_arg;
133	STAILQ_ENTRY(bus_dmamap) links;
134	int		       contigalloc;
135};
136
137static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
138static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
139
140static void init_bounce_pages(void *dummy);
141static int alloc_bounce_zone(bus_dma_tag_t dmat);
142static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
143static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
144				int commit);
145static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
146				  vm_offset_t vaddr, bus_addr_t addr,
147				  bus_size_t size);
148static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
149static __inline int run_filter(bus_dma_tag_t dmat, bus_addr_t paddr);
150
151/*
152 * Return true if a match is made.
153 *
154 * To find a match walk the chain of bus_dma_tag_t's looking for 'paddr'.
155 *
156 * If paddr is within the bounds of the dma tag then call the filter callback
157 * to check for a match, if there is no filter callback then assume a match.
158 */
159static __inline int
160run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
161{
162	int retval;
163
164	retval = 0;
165
166	do {
167		if (dmat->filter == NULL && dmat->iommu == NULL &&
168		    paddr > dmat->lowaddr && paddr <= dmat->highaddr)
169			retval = 1;
170		if (dmat->filter == NULL &&
171		    (paddr & (dmat->alignment - 1)) != 0)
172			retval = 1;
173		if (dmat->filter != NULL &&
174		    (*dmat->filter)(dmat->filterarg, paddr) != 0)
175			retval = 1;
176
177		dmat = dmat->parent;
178	} while (retval == 0 && dmat != NULL);
179	return (retval);
180}
181
182/*
183 * Convenience function for manipulating driver locks from busdma (during
184 * busdma_swi, for example).  Drivers that don't provide their own locks
185 * should specify &Giant to dmat->lockfuncarg.  Drivers that use their own
186 * non-mutex locking scheme don't have to use this at all.
187 */
188void
189busdma_lock_mutex(void *arg, bus_dma_lock_op_t op)
190{
191	struct mtx *dmtx;
192
193	dmtx = (struct mtx *)arg;
194	switch (op) {
195	case BUS_DMA_LOCK:
196		mtx_lock(dmtx);
197		break;
198	case BUS_DMA_UNLOCK:
199		mtx_unlock(dmtx);
200		break;
201	default:
202		panic("Unknown operation 0x%x for busdma_lock_mutex!", op);
203	}
204}
205
206/*
207 * dflt_lock should never get called.  It gets put into the dma tag when
208 * lockfunc == NULL, which is only valid if the maps that are associated
209 * with the tag are meant to never be defered.
210 * XXX Should have a way to identify which driver is responsible here.
211 */
212static void
213dflt_lock(void *arg, bus_dma_lock_op_t op)
214{
215	panic("driver error: busdma dflt_lock called");
216}
217
218#define BUS_DMA_COULD_BOUNCE	BUS_DMA_BUS3
219#define BUS_DMA_MIN_ALLOC_COMP	BUS_DMA_BUS4
220/*
221 * Allocate a device specific dma_tag.
222 */
223int
224bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
225		   bus_size_t boundary, bus_addr_t lowaddr,
226		   bus_addr_t highaddr, bus_dma_filter_t *filter,
227		   void *filterarg, bus_size_t maxsize, int nsegments,
228		   bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
229		   void *lockfuncarg, bus_dma_tag_t *dmat)
230{
231	bus_dma_tag_t newtag;
232	int error = 0;
233
234	/* Basic sanity checking */
235	if (boundary != 0 && boundary < maxsegsz)
236		maxsegsz = boundary;
237
238	if (maxsegsz == 0) {
239		return (EINVAL);
240	}
241
242	/* Return a NULL tag on failure */
243	*dmat = NULL;
244
245	newtag = (bus_dma_tag_t)malloc(sizeof(*newtag), M_DEVBUF,
246	    M_ZERO | M_NOWAIT);
247	if (newtag == NULL) {
248		CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
249		    __func__, newtag, 0, error);
250		return (ENOMEM);
251	}
252
253	newtag->parent = parent;
254	newtag->alignment = alignment;
255	newtag->boundary = boundary;
256	newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
257	newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
258	newtag->filter = filter;
259	newtag->filterarg = filterarg;
260	newtag->maxsize = maxsize;
261	newtag->nsegments = nsegments;
262	newtag->maxsegsz = maxsegsz;
263	newtag->flags = flags;
264	newtag->ref_count = 1; /* Count ourself */
265	newtag->map_count = 0;
266	if (lockfunc != NULL) {
267		newtag->lockfunc = lockfunc;
268		newtag->lockfuncarg = lockfuncarg;
269	} else {
270		newtag->lockfunc = dflt_lock;
271		newtag->lockfuncarg = NULL;
272	}
273
274	/* Take into account any restrictions imposed by our parent tag */
275	if (parent != NULL) {
276		newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
277		newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
278		if (newtag->boundary == 0)
279			newtag->boundary = parent->boundary;
280		else if (parent->boundary != 0)
281			newtag->boundary = MIN(parent->boundary,
282					       newtag->boundary);
283		if (newtag->filter == NULL) {
284			/*
285			 * Short circuit looking at our parent directly
286			 * since we have encapsulated all of its information
287			 */
288			newtag->filter = parent->filter;
289			newtag->filterarg = parent->filterarg;
290			newtag->parent = parent->parent;
291		}
292		if (newtag->parent != NULL)
293			atomic_add_int(&parent->ref_count, 1);
294		newtag->iommu = parent->iommu;
295		newtag->iommu_cookie = parent->iommu_cookie;
296	}
297
298	if (newtag->lowaddr < ptoa((vm_paddr_t)Maxmem) && newtag->iommu == NULL)
299		newtag->flags |= BUS_DMA_COULD_BOUNCE;
300
301	if (newtag->alignment > 1)
302		newtag->flags |= BUS_DMA_COULD_BOUNCE;
303
304	if (((newtag->flags & BUS_DMA_COULD_BOUNCE) != 0) &&
305	    (flags & BUS_DMA_ALLOCNOW) != 0) {
306		struct bounce_zone *bz;
307
308		/* Must bounce */
309
310		if ((error = alloc_bounce_zone(newtag)) != 0) {
311			free(newtag, M_DEVBUF);
312			return (error);
313		}
314		bz = newtag->bounce_zone;
315
316		if (ptoa(bz->total_bpages) < maxsize) {
317			int pages;
318
319			pages = atop(maxsize) - bz->total_bpages;
320
321			/* Add pages to our bounce pool */
322			if (alloc_bounce_pages(newtag, pages) < pages)
323				error = ENOMEM;
324		}
325		/* Performed initial allocation */
326		newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
327	}
328
329	if (error != 0) {
330		free(newtag, M_DEVBUF);
331	} else {
332		*dmat = newtag;
333	}
334	CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
335	    __func__, newtag, (newtag != NULL ? newtag->flags : 0), error);
336	return (error);
337}
338
339int
340bus_dma_tag_destroy(bus_dma_tag_t dmat)
341{
342	bus_dma_tag_t dmat_copy;
343	int error;
344
345	error = 0;
346	dmat_copy = dmat;
347
348	if (dmat != NULL) {
349
350		if (dmat->map_count != 0) {
351			error = EBUSY;
352			goto out;
353		}
354
355		while (dmat != NULL) {
356			bus_dma_tag_t parent;
357
358			parent = dmat->parent;
359			atomic_subtract_int(&dmat->ref_count, 1);
360			if (dmat->ref_count == 0) {
361				free(dmat, M_DEVBUF);
362				/*
363				 * Last reference count, so
364				 * release our reference
365				 * count on our parent.
366				 */
367				dmat = parent;
368			} else
369				dmat = NULL;
370		}
371	}
372out:
373	CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
374	return (error);
375}
376
377/*
378 * Allocate a handle for mapping from kva/uva/physical
379 * address space into bus device space.
380 */
381int
382bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
383{
384	int error;
385
386	error = 0;
387
388	*mapp = (bus_dmamap_t)malloc(sizeof(**mapp), M_DEVBUF,
389				     M_NOWAIT | M_ZERO);
390	if (*mapp == NULL) {
391		CTR3(KTR_BUSDMA, "%s: tag %p error %d",
392		    __func__, dmat, ENOMEM);
393		return (ENOMEM);
394	}
395
396
397	/*
398	 * Bouncing might be required if the driver asks for an active
399	 * exclusion region, a data alignment that is stricter than 1, and/or
400	 * an active address boundary.
401	 */
402	if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
403
404		/* Must bounce */
405		struct bounce_zone *bz;
406		int maxpages;
407
408		if (dmat->bounce_zone == NULL) {
409			if ((error = alloc_bounce_zone(dmat)) != 0)
410				return (error);
411		}
412		bz = dmat->bounce_zone;
413
414		/* Initialize the new map */
415		STAILQ_INIT(&((*mapp)->bpages));
416
417		/*
418		 * Attempt to add pages to our pool on a per-instance
419		 * basis up to a sane limit.
420		 */
421		if (dmat->alignment > 1)
422			maxpages = MAX_BPAGES;
423		else
424			maxpages = MIN(MAX_BPAGES, Maxmem -atop(dmat->lowaddr));
425		if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0
426		 || (bz->map_count > 0 && bz->total_bpages < maxpages)) {
427			int pages;
428
429			pages = MAX(atop(dmat->maxsize), 1);
430			pages = MIN(maxpages - bz->total_bpages, pages);
431			pages = MAX(pages, 1);
432			if (alloc_bounce_pages(dmat, pages) < pages)
433				error = ENOMEM;
434
435			if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
436				if (error == 0)
437					dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
438			} else {
439				error = 0;
440			}
441		}
442		bz->map_count++;
443	}
444
445	(*mapp)->nsegs = 0;
446	(*mapp)->segments = (bus_dma_segment_t *)malloc(
447	    sizeof(bus_dma_segment_t) * dmat->nsegments, M_DEVBUF,
448	    M_NOWAIT);
449	if ((*mapp)->segments == NULL) {
450		CTR3(KTR_BUSDMA, "%s: tag %p error %d",
451		    __func__, dmat, ENOMEM);
452		return (ENOMEM);
453	}
454
455	if (error == 0)
456		dmat->map_count++;
457	CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
458	    __func__, dmat, dmat->flags, error);
459	return (error);
460}
461
462/*
463 * Destroy a handle for mapping from kva/uva/physical
464 * address space into bus device space.
465 */
466int
467bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
468{
469	if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
470		if (STAILQ_FIRST(&map->bpages) != NULL) {
471			CTR3(KTR_BUSDMA, "%s: tag %p error %d",
472			    __func__, dmat, EBUSY);
473			return (EBUSY);
474		}
475		if (dmat->bounce_zone)
476			dmat->bounce_zone->map_count--;
477	}
478	free(map->segments, M_DEVBUF);
479	free(map, M_DEVBUF);
480	dmat->map_count--;
481	CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
482	return (0);
483}
484
485
486/*
487 * Allocate a piece of memory that can be efficiently mapped into
488 * bus device space based on the constraints lited in the dma tag.
489 * A dmamap to for use with dmamap_load is also allocated.
490 */
491int
492bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
493		 bus_dmamap_t *mapp)
494{
495	vm_memattr_t attr;
496	int mflags;
497
498	if (flags & BUS_DMA_NOWAIT)
499		mflags = M_NOWAIT;
500	else
501		mflags = M_WAITOK;
502
503	bus_dmamap_create(dmat, flags, mapp);
504
505	if (flags & BUS_DMA_ZERO)
506		mflags |= M_ZERO;
507#ifdef NOTYET
508	if (flags & BUS_DMA_NOCACHE)
509		attr = VM_MEMATTR_UNCACHEABLE;
510	else
511#endif
512		attr = VM_MEMATTR_DEFAULT;
513
514	/*
515	 * XXX:
516	 * (dmat->alignment < dmat->maxsize) is just a quick hack; the exact
517	 * alignment guarantees of malloc need to be nailed down, and the
518	 * code below should be rewritten to take that into account.
519	 *
520	 * In the meantime, we'll warn the user if malloc gets it wrong.
521	 */
522	if ((dmat->maxsize <= PAGE_SIZE) &&
523	   (dmat->alignment < dmat->maxsize) &&
524	    dmat->lowaddr >= ptoa((vm_paddr_t)Maxmem) &&
525	    attr == VM_MEMATTR_DEFAULT) {
526		*vaddr = malloc(dmat->maxsize, M_DEVBUF, mflags);
527	} else {
528		/*
529		 * XXX Use Contigmalloc until it is merged into this facility
530		 *     and handles multi-seg allocations.  Nobody is doing
531		 *     multi-seg allocations yet though.
532		 * XXX Certain AGP hardware does.
533		 */
534		*vaddr = (void *)kmem_alloc_contig(kernel_map, dmat->maxsize,
535		    mflags, 0ul, dmat->lowaddr, dmat->alignment ?
536		    dmat->alignment : 1ul, dmat->boundary, attr);
537		(*mapp)->contigalloc = 1;
538	}
539	if (*vaddr == NULL) {
540		CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
541		    __func__, dmat, dmat->flags, ENOMEM);
542		return (ENOMEM);
543	} else if (vtophys(*vaddr) & (dmat->alignment - 1)) {
544		printf("bus_dmamem_alloc failed to align memory properly.\n");
545	}
546	CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
547	    __func__, dmat, dmat->flags, 0);
548	return (0);
549}
550
551/*
552 * Free a piece of memory and it's allociated dmamap, that was allocated
553 * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
554 */
555void
556bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
557{
558
559	if (!map->contigalloc)
560		free(vaddr, M_DEVBUF);
561	else
562		kmem_free(kernel_map, (vm_offset_t)vaddr, dmat->maxsize);
563	bus_dmamap_destroy(dmat, map);
564	CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat, dmat->flags);
565}
566
567static void
568_bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
569    bus_size_t buflen, int flags)
570{
571	bus_addr_t curaddr;
572	bus_size_t sgsize;
573
574	if (map->pagesneeded == 0) {
575		CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
576		    "alignment= %d", dmat->lowaddr, ptoa((vm_paddr_t)Maxmem),
577		    dmat->boundary, dmat->alignment);
578		CTR2(KTR_BUSDMA, "map= %p, pagesneeded= %d", map, map->pagesneeded);
579		/*
580		 * Count the number of bounce pages
581		 * needed in order to complete this transfer
582		 */
583		curaddr = buf;
584		while (buflen != 0) {
585			sgsize = MIN(buflen, dmat->maxsegsz);
586			if (run_filter(dmat, curaddr) != 0) {
587				sgsize = MIN(sgsize, PAGE_SIZE);
588				map->pagesneeded++;
589			}
590			curaddr += sgsize;
591			buflen -= sgsize;
592		}
593		CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
594	}
595}
596
597static void
598_bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, pmap_t pmap,
599    void *buf, bus_size_t buflen, int flags)
600{
601        vm_offset_t vaddr;
602        vm_offset_t vendaddr;
603        bus_addr_t paddr;
604
605	if (map->pagesneeded == 0) {
606		CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
607		    "alignment= %d", dmat->lowaddr, ptoa((vm_paddr_t)Maxmem),
608		    dmat->boundary, dmat->alignment);
609		CTR2(KTR_BUSDMA, "map= %p, pagesneeded= %d", map, map->pagesneeded);
610		/*
611		 * Count the number of bounce pages
612		 * needed in order to complete this transfer
613		 */
614		vaddr = (vm_offset_t)buf;
615		vendaddr = (vm_offset_t)buf + buflen;
616
617		while (vaddr < vendaddr) {
618			bus_size_t sg_len;
619
620			sg_len = PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK);
621			if (pmap == kernel_pmap)
622				paddr = pmap_kextract(vaddr);
623			else
624				paddr = pmap_extract(pmap, vaddr);
625			if (run_filter(dmat, paddr) != 0) {
626				sg_len = roundup2(sg_len, dmat->alignment);
627				map->pagesneeded++;
628			}
629			vaddr += sg_len;
630		}
631		CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
632	}
633}
634
635static int
636_bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags)
637{
638
639	/* Reserve Necessary Bounce Pages */
640	mtx_lock(&bounce_lock);
641	if (flags & BUS_DMA_NOWAIT) {
642		if (reserve_bounce_pages(dmat, map, 0) != 0) {
643			mtx_unlock(&bounce_lock);
644			return (ENOMEM);
645		}
646	} else {
647		if (reserve_bounce_pages(dmat, map, 1) != 0) {
648			/* Queue us for resources */
649			STAILQ_INSERT_TAIL(&bounce_map_waitinglist,
650			    map, links);
651			mtx_unlock(&bounce_lock);
652			return (EINPROGRESS);
653		}
654	}
655	mtx_unlock(&bounce_lock);
656
657	return (0);
658}
659
660/*
661 * Add a single contiguous physical range to the segment list.
662 */
663static int
664_bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr,
665		   bus_size_t sgsize, bus_dma_segment_t *segs, int *segp)
666{
667	bus_addr_t baddr, bmask;
668	int seg;
669
670	/*
671	 * Make sure we don't cross any boundaries.
672	 */
673	bmask = ~(dmat->boundary - 1);
674	if (dmat->boundary > 0) {
675		baddr = (curaddr + dmat->boundary) & bmask;
676		if (sgsize > (baddr - curaddr))
677			sgsize = (baddr - curaddr);
678	}
679
680	/*
681	 * Insert chunk into a segment, coalescing with
682	 * previous segment if possible.
683	 */
684	seg = *segp;
685	if (seg == -1) {
686		seg = 0;
687		segs[seg].ds_addr = curaddr;
688		segs[seg].ds_len = sgsize;
689	} else {
690		if (curaddr == segs[seg].ds_addr + segs[seg].ds_len &&
691		    (segs[seg].ds_len + sgsize) <= dmat->maxsegsz &&
692		    (dmat->boundary == 0 ||
693		     (segs[seg].ds_addr & bmask) == (curaddr & bmask)))
694			segs[seg].ds_len += sgsize;
695		else {
696			if (++seg >= dmat->nsegments)
697				return (0);
698			segs[seg].ds_addr = curaddr;
699			segs[seg].ds_len = sgsize;
700		}
701	}
702	*segp = seg;
703	return (sgsize);
704}
705
706/*
707 * Utility function to load a physical buffer.  segp contains
708 * the starting segment on entrace, and the ending segment on exit.
709 */
710int
711_bus_dmamap_load_phys(bus_dma_tag_t dmat,
712		      bus_dmamap_t map,
713		      vm_paddr_t buf, bus_size_t buflen,
714		      int flags,
715		      bus_dma_segment_t *segs,
716		      int *segp)
717{
718	bus_addr_t curaddr;
719	bus_size_t sgsize;
720	int error;
721
722	if (segs == NULL)
723		segs = map->segments;
724
725	if ((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) {
726		_bus_dmamap_count_phys(dmat, map, buf, buflen, flags);
727		if (map->pagesneeded != 0) {
728			error = _bus_dmamap_reserve_pages(dmat, map, flags);
729			if (error)
730				return (error);
731		}
732	}
733
734	while (buflen > 0) {
735		curaddr = buf;
736		sgsize = MIN(buflen, dmat->maxsegsz);
737		if (map->pagesneeded != 0 && run_filter(dmat, curaddr)) {
738			sgsize = MIN(sgsize, PAGE_SIZE);
739			curaddr = add_bounce_page(dmat, map, 0, curaddr,
740			    sgsize);
741		}
742		sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
743		    segp);
744		if (sgsize == 0)
745			break;
746		buf += sgsize;
747		buflen -= sgsize;
748	}
749
750	/*
751	 * Did we fit?
752	 */
753	return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
754}
755
756/*
757 * Utility function to load a linear buffer.  segp contains
758 * the starting segment on entrance, and the ending segment on exit.
759 */
760int
761_bus_dmamap_load_buffer(bus_dma_tag_t dmat,
762    			bus_dmamap_t map,
763			void *buf, bus_size_t buflen,
764			pmap_t pmap,
765			int flags,
766			bus_dma_segment_t *segs,
767			int *segp)
768{
769	bus_size_t sgsize;
770	bus_addr_t curaddr;
771	vm_offset_t vaddr;
772	int error;
773
774	if (segs == NULL)
775		segs = map->segments;
776
777	if ((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) {
778		_bus_dmamap_count_pages(dmat, map, pmap, buf, buflen, flags);
779		if (map->pagesneeded != 0) {
780			error = _bus_dmamap_reserve_pages(dmat, map, flags);
781			if (error)
782				return (error);
783		}
784	}
785
786	vaddr = (vm_offset_t)buf;
787
788	while (buflen > 0) {
789		bus_size_t max_sgsize;
790
791		/*
792		 * Get the physical address for this segment.
793		 */
794		if (pmap == kernel_pmap)
795			curaddr = pmap_kextract(vaddr);
796		else
797			curaddr = pmap_extract(pmap, vaddr);
798
799		/*
800		 * Compute the segment size, and adjust counts.
801		 */
802		max_sgsize = MIN(buflen, dmat->maxsegsz);
803		sgsize = PAGE_SIZE - ((vm_offset_t)curaddr & PAGE_MASK);
804		if (map->pagesneeded != 0 && run_filter(dmat, curaddr)) {
805			sgsize = roundup2(sgsize, dmat->alignment);
806			sgsize = MIN(sgsize, max_sgsize);
807			curaddr = add_bounce_page(dmat, map, vaddr, curaddr,
808			    sgsize);
809		} else {
810			sgsize = MIN(sgsize, max_sgsize);
811		}
812
813		sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
814		    segp);
815		if (sgsize == 0)
816			break;
817		vaddr += sgsize;
818		buflen -= sgsize;
819	}
820
821	/*
822	 * Did we fit?
823	 */
824	return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
825}
826
827void
828__bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map,
829		    struct memdesc *mem, bus_dmamap_callback_t *callback,
830		    void *callback_arg)
831{
832
833	if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
834		map->dmat = dmat;
835		map->mem = *mem;
836		map->callback = callback;
837		map->callback_arg = callback_arg;
838	}
839}
840
841bus_dma_segment_t *
842_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map,
843		     bus_dma_segment_t *segs, int nsegs, int error)
844{
845
846	map->nsegs = nsegs;
847	if (segs != NULL)
848		memcpy(map->segments, segs, map->nsegs*sizeof(segs[0]));
849	if (dmat->iommu != NULL)
850		IOMMU_MAP(dmat->iommu, map->segments, &map->nsegs,
851		    dmat->lowaddr, dmat->highaddr, dmat->alignment,
852		    dmat->boundary, dmat->iommu_cookie);
853
854	if (segs != NULL)
855		memcpy(segs, map->segments, map->nsegs*sizeof(segs[0]));
856	else
857		segs = map->segments;
858
859	return (segs);
860}
861
862/*
863 * Release the mapping held by map.
864 */
865void
866_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
867{
868	struct bounce_page *bpage;
869
870	if (dmat->iommu) {
871		IOMMU_UNMAP(dmat->iommu, map->segments, map->nsegs, dmat->iommu_cookie);
872		map->nsegs = 0;
873	}
874
875	while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
876		STAILQ_REMOVE_HEAD(&map->bpages, links);
877		free_bounce_page(dmat, bpage);
878	}
879}
880
881void
882_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
883{
884	struct bounce_page *bpage;
885
886	if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
887		/*
888		 * Handle data bouncing.  We might also
889		 * want to add support for invalidating
890		 * the caches on broken hardware
891		 */
892		CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
893		    "performing bounce", __func__, op, dmat, dmat->flags);
894
895		if (op & BUS_DMASYNC_PREWRITE) {
896			while (bpage != NULL) {
897				if (bpage->datavaddr != 0)
898					bcopy((void *)bpage->datavaddr,
899					      (void *)bpage->vaddr,
900					      bpage->datacount);
901				else
902					physcopyout(bpage->dataaddr,
903					    (void *)bpage->vaddr,
904					    bpage->datacount);
905				bpage = STAILQ_NEXT(bpage, links);
906			}
907			dmat->bounce_zone->total_bounced++;
908		}
909
910		if (op & BUS_DMASYNC_POSTREAD) {
911			while (bpage != NULL) {
912				if (bpage->datavaddr != 0)
913					bcopy((void *)bpage->vaddr,
914					      (void *)bpage->datavaddr,
915					      bpage->datacount);
916				else
917					physcopyin((void *)bpage->vaddr,
918					    bpage->dataaddr, bpage->datacount);
919				bpage = STAILQ_NEXT(bpage, links);
920			}
921			dmat->bounce_zone->total_bounced++;
922		}
923	}
924}
925
926static void
927init_bounce_pages(void *dummy __unused)
928{
929
930	total_bpages = 0;
931	STAILQ_INIT(&bounce_zone_list);
932	STAILQ_INIT(&bounce_map_waitinglist);
933	STAILQ_INIT(&bounce_map_callbacklist);
934	mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
935}
936SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
937
938static struct sysctl_ctx_list *
939busdma_sysctl_tree(struct bounce_zone *bz)
940{
941	return (&bz->sysctl_tree);
942}
943
944static struct sysctl_oid *
945busdma_sysctl_tree_top(struct bounce_zone *bz)
946{
947	return (bz->sysctl_tree_top);
948}
949
950static int
951alloc_bounce_zone(bus_dma_tag_t dmat)
952{
953	struct bounce_zone *bz;
954
955	/* Check to see if we already have a suitable zone */
956	STAILQ_FOREACH(bz, &bounce_zone_list, links) {
957		if ((dmat->alignment <= bz->alignment)
958		 && (dmat->lowaddr >= bz->lowaddr)) {
959			dmat->bounce_zone = bz;
960			return (0);
961		}
962	}
963
964	if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
965	    M_NOWAIT | M_ZERO)) == NULL)
966		return (ENOMEM);
967
968	STAILQ_INIT(&bz->bounce_page_list);
969	bz->free_bpages = 0;
970	bz->reserved_bpages = 0;
971	bz->active_bpages = 0;
972	bz->lowaddr = dmat->lowaddr;
973	bz->alignment = MAX(dmat->alignment, PAGE_SIZE);
974	bz->map_count = 0;
975	snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
976	busdma_zonecount++;
977	snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
978	STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
979	dmat->bounce_zone = bz;
980
981	sysctl_ctx_init(&bz->sysctl_tree);
982	bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
983	    SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
984	    CTLFLAG_RD, 0, "");
985	if (bz->sysctl_tree_top == NULL) {
986		sysctl_ctx_free(&bz->sysctl_tree);
987		return (0);	/* XXX error code? */
988	}
989
990	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
991	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
992	    "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
993	    "Total bounce pages");
994	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
995	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
996	    "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
997	    "Free bounce pages");
998	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
999	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1000	    "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1001	    "Reserved bounce pages");
1002	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1003	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1004	    "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1005	    "Active bounce pages");
1006	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1007	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1008	    "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1009	    "Total bounce requests");
1010	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1011	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1012	    "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1013	    "Total bounce requests that were deferred");
1014	SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
1015	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1016	    "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1017	SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1018	    SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1019	    "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1020
1021	return (0);
1022}
1023
1024static int
1025alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
1026{
1027	struct bounce_zone *bz;
1028	int count;
1029
1030	bz = dmat->bounce_zone;
1031	count = 0;
1032	while (numpages > 0) {
1033		struct bounce_page *bpage;
1034
1035		bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
1036						     M_NOWAIT | M_ZERO);
1037
1038		if (bpage == NULL)
1039			break;
1040		bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1041							 M_NOWAIT, 0ul,
1042							 bz->lowaddr,
1043							 PAGE_SIZE,
1044							 0);
1045		if (bpage->vaddr == 0) {
1046			free(bpage, M_DEVBUF);
1047			break;
1048		}
1049		bpage->busaddr = pmap_kextract(bpage->vaddr);
1050		mtx_lock(&bounce_lock);
1051		STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1052		total_bpages++;
1053		bz->total_bpages++;
1054		bz->free_bpages++;
1055		mtx_unlock(&bounce_lock);
1056		count++;
1057		numpages--;
1058	}
1059	return (count);
1060}
1061
1062static int
1063reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1064{
1065	struct bounce_zone *bz;
1066	int pages;
1067
1068	mtx_assert(&bounce_lock, MA_OWNED);
1069	bz = dmat->bounce_zone;
1070	pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1071	if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
1072		return (map->pagesneeded - (map->pagesreserved + pages));
1073	bz->free_bpages -= pages;
1074	bz->reserved_bpages += pages;
1075	map->pagesreserved += pages;
1076	pages = map->pagesneeded - map->pagesreserved;
1077
1078	return (pages);
1079}
1080
1081static bus_addr_t
1082add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1083		bus_addr_t addr, bus_size_t size)
1084{
1085	struct bounce_zone *bz;
1086	struct bounce_page *bpage;
1087
1088	KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
1089
1090	bz = dmat->bounce_zone;
1091	if (map->pagesneeded == 0)
1092		panic("add_bounce_page: map doesn't need any pages");
1093	map->pagesneeded--;
1094
1095	if (map->pagesreserved == 0)
1096		panic("add_bounce_page: map doesn't need any pages");
1097	map->pagesreserved--;
1098
1099	mtx_lock(&bounce_lock);
1100	bpage = STAILQ_FIRST(&bz->bounce_page_list);
1101	if (bpage == NULL)
1102		panic("add_bounce_page: free page list is empty");
1103
1104	STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1105	bz->reserved_bpages--;
1106	bz->active_bpages++;
1107	mtx_unlock(&bounce_lock);
1108
1109	if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1110		/* Page offset needs to be preserved. */
1111		bpage->vaddr |= vaddr & PAGE_MASK;
1112		bpage->busaddr |= vaddr & PAGE_MASK;
1113	}
1114	bpage->datavaddr = vaddr;
1115	bpage->dataaddr = addr;
1116	bpage->datacount = size;
1117	STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1118	return (bpage->busaddr);
1119}
1120
1121static void
1122free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1123{
1124	struct bus_dmamap *map;
1125	struct bounce_zone *bz;
1126
1127	bz = dmat->bounce_zone;
1128	bpage->datavaddr = 0;
1129	bpage->datacount = 0;
1130	if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1131		/*
1132		 * Reset the bounce page to start at offset 0.  Other uses
1133		 * of this bounce page may need to store a full page of
1134		 * data and/or assume it starts on a page boundary.
1135		 */
1136		bpage->vaddr &= ~PAGE_MASK;
1137		bpage->busaddr &= ~PAGE_MASK;
1138	}
1139
1140	mtx_lock(&bounce_lock);
1141	STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1142	bz->free_bpages++;
1143	bz->active_bpages--;
1144	if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1145		if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1146			STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1147			STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1148					   map, links);
1149			busdma_swi_pending = 1;
1150			bz->total_deferred++;
1151			swi_sched(vm_ih, 0);
1152		}
1153	}
1154	mtx_unlock(&bounce_lock);
1155}
1156
1157void
1158busdma_swi(void)
1159{
1160	bus_dma_tag_t dmat;
1161	struct bus_dmamap *map;
1162
1163	mtx_lock(&bounce_lock);
1164	while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1165		STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1166		mtx_unlock(&bounce_lock);
1167		dmat = map->dmat;
1168		(dmat->lockfunc)(dmat->lockfuncarg, BUS_DMA_LOCK);
1169		bus_dmamap_load_mem(map->dmat, map, &map->mem,
1170				    map->callback, map->callback_arg,
1171				    BUS_DMA_WAITOK);
1172		(dmat->lockfunc)(dmat->lockfuncarg, BUS_DMA_UNLOCK);
1173		mtx_lock(&bounce_lock);
1174	}
1175	mtx_unlock(&bounce_lock);
1176}
1177
1178int
1179bus_dma_tag_set_iommu(bus_dma_tag_t tag, struct device *iommu, void *cookie)
1180{
1181	tag->iommu = iommu;
1182	tag->iommu_cookie = cookie;
1183
1184	return (0);
1185}
1186
1187