a8cff4ed937b655f17c5c71bac0535b8e0770fb3
[dragonfly.git] / sys / platform / pc32 / i386 / busdma_machdep.c
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  * $FreeBSD: src/sys/i386/i386/busdma_machdep.c,v 1.94 2008/08/15 20:51:31 kmacy Exp $
27  */
28
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/malloc.h>
32 #include <sys/mbuf.h>
33 #include <sys/uio.h>
34 #include <sys/bus_dma.h>
35 #include <sys/kernel.h>
36 #include <sys/sysctl.h>
37 #include <sys/lock.h>
38
39 #include <sys/thread2.h>
40 #include <sys/spinlock2.h>
41 #include <sys/mplock2.h>
42
43 #include <vm/vm.h>
44 #include <vm/vm_page.h>
45
46 /* XXX needed for to access pmap to convert per-proc virtual to physical */
47 #include <sys/proc.h>
48 #include <sys/lock.h>
49 #include <vm/vm_map.h>
50
51 #include <machine/md_var.h>
52
53 #define MAX_BPAGES      1024
54
55 /*
56  * 16 x N declared on stack.
57  */
58 #define BUS_DMA_CACHE_SEGMENTS  8
59
60 struct bounce_zone;
61 struct bus_dmamap;
62
63 struct bus_dma_tag {
64         bus_dma_tag_t   parent;
65         bus_size_t      alignment;
66         bus_size_t      boundary;
67         bus_addr_t      lowaddr;
68         bus_addr_t      highaddr;
69         bus_dma_filter_t *filter;
70         void            *filterarg;
71         bus_size_t      maxsize;
72         u_int           nsegments;
73         bus_size_t      maxsegsz;
74         int             flags;
75         int             ref_count;
76         int             map_count;
77         bus_dma_segment_t *segments;
78         struct bounce_zone *bounce_zone;
79 #ifdef SMP
80         struct spinlock spin;
81 #else
82         int             unused0;
83 #endif
84 };
85
86 /*
87  * bus_dma_tag private flags
88  */
89 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
90 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
91 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
92
93 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
94
95 #define BUS_DMAMEM_KMALLOC(dmat) \
96         ((dmat)->maxsize <= PAGE_SIZE && \
97          (dmat)->alignment <= PAGE_SIZE && \
98          (dmat)->lowaddr >= ptoa(Maxmem))
99
100 struct bounce_page {
101         vm_offset_t     vaddr;          /* kva of bounce buffer */
102         bus_addr_t      busaddr;        /* Physical address */
103         vm_offset_t     datavaddr;      /* kva of client data */
104         bus_size_t      datacount;      /* client data count */
105         STAILQ_ENTRY(bounce_page) links;
106 };
107
108 struct bounce_zone {
109         STAILQ_ENTRY(bounce_zone) links;
110         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
111         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
112 #ifdef SMP
113         struct spinlock spin;
114 #else
115         int             unused0;
116 #endif
117         int             total_bpages;
118         int             free_bpages;
119         int             reserved_bpages;
120         int             active_bpages;
121         int             total_bounced;
122         int             total_deferred;
123         int             reserve_failed;
124         bus_size_t      alignment;
125         bus_addr_t      lowaddr;
126         char            zoneid[8];
127         char            lowaddrid[20];
128         struct sysctl_ctx_list sysctl_ctx;
129         struct sysctl_oid *sysctl_tree;
130 };
131
132 #ifdef SMP
133 #define BZ_LOCK(bz)     spin_lock(&(bz)->spin)
134 #define BZ_UNLOCK(bz)   spin_unlock(&(bz)->spin)
135 #else
136 #define BZ_LOCK(bz)     crit_enter()
137 #define BZ_UNLOCK(bz)   crit_exit()
138 #endif
139
140 static struct lwkt_token bounce_zone_tok =
141         LWKT_TOKEN_INITIALIZER(bounce_zone_tok);
142 static int busdma_zonecount;
143 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
144         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
145
146 static int busdma_priv_zonecount = -1;
147
148 int busdma_swi_pending;
149 static int total_bounce_pages;
150 static int max_bounce_pages = MAX_BPAGES;
151 static int bounce_alignment = 1; /* XXX temporary */
152
153 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
154 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
155
156 struct bus_dmamap {
157         struct bp_list  bpages;
158         int             pagesneeded;
159         int             pagesreserved;
160         bus_dma_tag_t   dmat;
161         void            *buf;           /* unmapped buffer pointer */
162         bus_size_t      buflen;         /* unmapped buffer length */
163         bus_dmamap_callback_t *callback;
164         void            *callback_arg;
165         STAILQ_ENTRY(bus_dmamap) links;
166 };
167
168 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
169         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
170 static struct spinlock bounce_map_list_spin =
171         SPINLOCK_INITIALIZER(&bounce_map_list_spin);
172
173 static struct bus_dmamap nobounce_dmamap;
174
175 static int              alloc_bounce_zone(bus_dma_tag_t);
176 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
177 static void             free_bounce_pages_all(bus_dma_tag_t);
178 static void             free_bounce_zone(bus_dma_tag_t);
179 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
180 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
181 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
182                             vm_offset_t, bus_size_t);
183 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
184
185 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
186 static void             add_map_callback(bus_dmamap_t);
187
188 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
189 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
190            0, "Total bounce pages");
191 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
192            0, "Max bounce pages per bounce zone");
193 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
194            &bounce_alignment, 0, "Obey alignment constraint");
195
196 static __inline int
197 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
198 {
199         int retval;
200
201         retval = 0;
202         do {
203                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
204                      (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0))
205                  && (dmat->filter == NULL ||
206                      dmat->filter(dmat->filterarg, paddr) != 0))
207                         retval = 1;
208
209                 dmat = dmat->parent;
210         } while (retval == 0 && dmat != NULL);
211         return (retval);
212 }
213
214 static __inline
215 bus_dma_segment_t *
216 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
217 {
218         if (tag->flags & BUS_DMA_PROTECTED)
219                 return(tag->segments);
220
221         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
222                 return(cache);
223 #ifdef SMP
224         spin_lock(&tag->spin);
225 #endif
226         return(tag->segments);
227 }
228
229 static __inline
230 void
231 bus_dma_tag_unlock(bus_dma_tag_t tag)
232 {
233 #ifdef SMP
234         if (tag->flags & BUS_DMA_PROTECTED)
235                 return;
236
237         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
238                 spin_unlock(&tag->spin);
239 #endif
240 }
241
242 /*
243  * Allocate a device specific dma_tag.
244  */
245 int
246 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
247                    bus_size_t boundary, bus_addr_t lowaddr,
248                    bus_addr_t highaddr, bus_dma_filter_t *filter,
249                    void *filterarg, bus_size_t maxsize, int nsegments,
250                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
251 {
252         bus_dma_tag_t newtag;
253         int error = 0;
254
255         /*
256          * Sanity checks
257          */
258
259         if (alignment == 0)
260                 alignment = 1;
261         if (alignment & (alignment - 1))
262                 panic("alignment must be power of 2");
263
264         if (boundary != 0) {
265                 if (boundary & (boundary - 1))
266                         panic("boundary must be power of 2");
267                 if (boundary < maxsegsz) {
268                         kprintf("boundary < maxsegsz:\n");
269                         print_backtrace(-1);
270                         maxsegsz = boundary;
271                 }
272         }
273
274         /* Return a NULL tag on failure */
275         *dmat = NULL;
276
277         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
278
279 #ifdef SMP
280         spin_init(&newtag->spin);
281 #endif
282         newtag->parent = parent;
283         newtag->alignment = alignment;
284         newtag->boundary = boundary;
285         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
286         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
287         newtag->filter = filter;
288         newtag->filterarg = filterarg;
289         newtag->maxsize = maxsize;
290         newtag->nsegments = nsegments;
291         newtag->maxsegsz = maxsegsz;
292         newtag->flags = flags;
293         newtag->ref_count = 1; /* Count ourself */
294         newtag->map_count = 0;
295         newtag->segments = NULL;
296         newtag->bounce_zone = NULL;
297
298         /* Take into account any restrictions imposed by our parent tag */
299         if (parent != NULL) {
300                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
301                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
302
303                 if (newtag->boundary == 0) {
304                         newtag->boundary = parent->boundary;
305                 } else if (parent->boundary != 0) {
306                         newtag->boundary = MIN(parent->boundary,
307                                                newtag->boundary);
308                 }
309
310 #ifdef notyet
311                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
312 #endif
313
314                 if (newtag->filter == NULL) {
315                         /*
316                          * Short circuit looking at our parent directly
317                          * since we have encapsulated all of its information
318                          */
319                         newtag->filter = parent->filter;
320                         newtag->filterarg = parent->filterarg;
321                         newtag->parent = parent->parent;
322                 }
323                 if (newtag->parent != NULL)
324                         parent->ref_count++;
325         }
326
327         if (newtag->lowaddr < ptoa(Maxmem))
328                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
329         if (bounce_alignment && newtag->alignment > 1 &&
330             !(newtag->flags & BUS_DMA_ALIGNED))
331                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
332
333         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
334             (flags & BUS_DMA_ALLOCNOW) != 0) {
335                 struct bounce_zone *bz;
336
337                 /* Must bounce */
338
339                 error = alloc_bounce_zone(newtag);
340                 if (error)
341                         goto back;
342                 bz = newtag->bounce_zone;
343
344                 if ((newtag->flags & BUS_DMA_ALLOCALL) == 0 &&
345                     ptoa(bz->total_bpages) < maxsize) {
346                         int pages;
347
348                         if (flags & BUS_DMA_ONEBPAGE) {
349                                 pages = 1;
350                         } else {
351                                 pages = atop(round_page(maxsize)) -
352                                         bz->total_bpages;
353                                 pages = MAX(pages, 1);
354                         }
355
356                         /* Add pages to our bounce pool */
357                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
358                                 error = ENOMEM;
359
360                         /* Performed initial allocation */
361                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
362                 }
363         }
364 back:
365         if (error) {
366                 free_bounce_zone(newtag);
367                 kfree(newtag, M_DEVBUF);
368         } else {
369                 *dmat = newtag;
370         }
371         return error;
372 }
373
374 int
375 bus_dma_tag_destroy(bus_dma_tag_t dmat)
376 {
377         if (dmat != NULL) {
378                 if (dmat->map_count != 0)
379                         return (EBUSY);
380
381                 while (dmat != NULL) {
382                         bus_dma_tag_t parent;
383
384                         parent = dmat->parent;
385                         dmat->ref_count--;
386                         if (dmat->ref_count == 0) {
387                                 free_bounce_zone(dmat);
388                                 if (dmat->segments != NULL)
389                                         kfree(dmat->segments, M_DEVBUF);
390                                 kfree(dmat, M_DEVBUF);
391                                 /*
392                                  * Last reference count, so
393                                  * release our reference
394                                  * count on our parent.
395                                  */
396                                 dmat = parent;
397                         } else
398                                 dmat = NULL;
399                 }
400         }
401         return (0);
402 }
403
404 bus_size_t
405 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
406 {
407         return(tag->maxsize);
408 }
409
410 /*
411  * Allocate a handle for mapping from kva/uva/physical
412  * address space into bus device space.
413  */
414 int
415 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
416 {
417         int error;
418
419         error = 0;
420
421         if (dmat->segments == NULL) {
422                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
423                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) * 
424                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
425         }
426
427         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
428                 struct bounce_zone *bz;
429                 int maxpages;
430
431                 /* Must bounce */
432
433                 if (dmat->bounce_zone == NULL) {
434                         error = alloc_bounce_zone(dmat);
435                         if (error)
436                                 return error;
437                 }
438                 bz = dmat->bounce_zone;
439
440                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
441
442                 /* Initialize the new map */
443                 STAILQ_INIT(&((*mapp)->bpages));
444
445                 /*
446                  * Attempt to add pages to our pool on a per-instance
447                  * basis up to a sane limit.
448                  */
449                 if (dmat->flags & BUS_DMA_ALLOCALL) {
450                         maxpages = Maxmem - atop(dmat->lowaddr);
451                 } else if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
452                         maxpages = max_bounce_pages;
453                 } else {
454                         maxpages = MIN(max_bounce_pages,
455                                        Maxmem - atop(dmat->lowaddr));
456                 }
457                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 ||
458                     (dmat->map_count > 0 && bz->total_bpages < maxpages)) {
459                         int pages;
460
461                         if (flags & BUS_DMA_ONEBPAGE) {
462                                 pages = 1;
463                         } else {
464                                 pages = atop(round_page(dmat->maxsize));
465                                 pages = MIN(maxpages - bz->total_bpages, pages);
466                                 pages = MAX(pages, 1);
467                         }
468                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
469                                 error = ENOMEM;
470
471                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
472                                 if (!error &&
473                                     (dmat->flags & BUS_DMA_ALLOCALL) == 0)
474                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
475                         } else {
476                                 error = 0;
477                         }
478                 }
479         } else {
480                 *mapp = NULL;
481         }
482         if (!error) {
483                 dmat->map_count++;
484         } else {
485                 kfree(*mapp, M_DEVBUF);
486                 *mapp = NULL;
487         }
488         return error;
489 }
490
491 /*
492  * Destroy a handle for mapping from kva/uva/physical
493  * address space into bus device space.
494  */
495 int
496 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
497 {
498         if (map != NULL) {
499                 if (STAILQ_FIRST(&map->bpages) != NULL)
500                         return (EBUSY);
501                 kfree(map, M_DEVBUF);
502         }
503         dmat->map_count--;
504         return (0);
505 }
506
507 static __inline bus_size_t
508 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
509 {
510         bus_size_t maxsize = 0;
511         uintptr_t vaddr = (uintptr_t)vaddr0;
512
513         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
514                 if (verify || bootverbose)
515                         kprintf("boundary check failed\n");
516                 if (verify)
517                         print_backtrace(-1); /* XXX panic */
518                 maxsize = dmat->maxsize;
519         }
520         if (vaddr & (dmat->alignment - 1)) {
521                 if (verify || bootverbose)
522                         kprintf("alignment check failed\n");
523                 if (verify)
524                         print_backtrace(-1); /* XXX panic */
525                 if (dmat->maxsize < dmat->alignment)
526                         maxsize = dmat->alignment;
527                 else
528                         maxsize = dmat->maxsize;
529         }
530         return maxsize;
531 }
532
533 /*
534  * Allocate a piece of memory that can be efficiently mapped into
535  * bus device space based on the constraints lited in the dma tag.
536  *
537  * mapp is degenerate.  By definition this allocation should not require
538  * bounce buffers so do not allocate a dma map.
539  */
540 int
541 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
542                  bus_dmamap_t *mapp)
543 {
544         int mflags;
545
546         /* If we succeed, no mapping/bouncing will be required */
547         *mapp = NULL;
548
549         if (dmat->segments == NULL) {
550                 KKASSERT(dmat->nsegments < 16384);
551                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) * 
552                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
553         }
554
555         if (flags & BUS_DMA_NOWAIT)
556                 mflags = M_NOWAIT;
557         else
558                 mflags = M_WAITOK;
559         if (flags & BUS_DMA_ZERO)
560                 mflags |= M_ZERO;
561
562         if (BUS_DMAMEM_KMALLOC(dmat)) {
563                 bus_size_t maxsize;
564
565                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
566
567                 /*
568                  * XXX
569                  * Check whether the allocation
570                  * - crossed a page boundary
571                  * - was not aligned
572                  * Retry with power-of-2 alignment in the above cases.
573                  */
574                 maxsize = check_kmalloc(dmat, *vaddr, 0);
575                 if (maxsize) {
576                         size_t size;
577
578                         kfree(*vaddr, M_DEVBUF);
579                         /* XXX check for overflow? */
580                         for (size = 1; size <= maxsize; size <<= 1)
581                                 ;
582                         *vaddr = kmalloc(size, M_DEVBUF, mflags);
583                         check_kmalloc(dmat, *vaddr, 1);
584                 }
585         } else {
586                 /*
587                  * XXX Use Contigmalloc until it is merged into this facility
588                  *     and handles multi-seg allocations.  Nobody is doing
589                  *     multi-seg allocations yet though.
590                  */
591                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
592                     0ul, dmat->lowaddr, dmat->alignment, dmat->boundary);
593         }
594         if (*vaddr == NULL)
595                 return (ENOMEM);
596         return (0);
597 }
598
599 /*
600  * Free a piece of memory and it's allociated dmamap, that was allocated
601  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
602  */
603 void
604 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
605 {
606         /*
607          * dmamem does not need to be bounced, so the map should be
608          * NULL
609          */
610         if (map != NULL)
611                 panic("bus_dmamem_free: Invalid map freed");
612         if (BUS_DMAMEM_KMALLOC(dmat))
613                 kfree(vaddr, M_DEVBUF);
614         else
615                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
616 }
617
618 static __inline vm_paddr_t
619 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
620 {
621         if (pmap)
622                 return pmap_extract(pmap, vaddr);
623         else
624                 return pmap_kextract(vaddr);
625 }
626
627 /*
628  * Utility function to load a linear buffer.  lastaddrp holds state
629  * between invocations (for multiple-buffer loads).  segp contains
630  * the segment following the starting one on entrace, and the ending
631  * segment on exit.  first indicates if this is the first invocation
632  * of this function.
633  */
634 static int
635 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
636                         bus_dmamap_t map,
637                         void *buf, bus_size_t buflen,
638                         bus_dma_segment_t *segments,
639                         int nsegments,
640                         pmap_t pmap,
641                         int flags,
642                         vm_paddr_t *lastpaddrp,
643                         int *segp,
644                         int first)
645 {
646         vm_offset_t vaddr;
647         vm_paddr_t paddr, nextpaddr;
648         bus_dma_segment_t *sg;
649         bus_addr_t bmask;
650         int seg, error = 0;
651
652         if (map == NULL)
653                 map = &nobounce_dmamap;
654
655 #ifdef INVARIANTS
656         if (dmat->flags & BUS_DMA_ALIGNED)
657                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
658 #endif
659
660         /*
661          * If we are being called during a callback, pagesneeded will
662          * be non-zero, so we can avoid doing the work twice.
663          */
664         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
665             map != &nobounce_dmamap && map->pagesneeded == 0) {
666                 vm_offset_t vendaddr;
667
668                 /*
669                  * Count the number of bounce pages
670                  * needed in order to complete this transfer
671                  */
672                 vaddr = (vm_offset_t)buf;
673                 vendaddr = (vm_offset_t)buf + buflen;
674
675                 while (vaddr < vendaddr) {
676                         paddr = _bus_dma_extract(pmap, vaddr);
677                         if (run_filter(dmat, paddr) != 0)
678                                 map->pagesneeded++;
679                         vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK));
680                 }
681         }
682
683         /* Reserve Necessary Bounce Pages */
684         if (map->pagesneeded != 0) {
685                 struct bounce_zone *bz;
686
687                 bz = dmat->bounce_zone;
688                 BZ_LOCK(bz);
689                 if (flags & BUS_DMA_NOWAIT) {
690                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
691                                 BZ_UNLOCK(bz);
692                                 error = ENOMEM;
693                                 goto free_bounce;
694                         }
695                 } else {
696                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
697                                 /* Queue us for resources */
698                                 map->dmat = dmat;
699                                 map->buf = buf;
700                                 map->buflen = buflen;
701
702                                 STAILQ_INSERT_TAIL(
703                                     &dmat->bounce_zone->bounce_map_waitinglist,
704                                     map, links);
705                                 BZ_UNLOCK(bz);
706
707                                 return (EINPROGRESS);
708                         }
709                 }
710                 BZ_UNLOCK(bz);
711         }
712
713         KKASSERT(*segp >= 1 && *segp <= nsegments);
714         seg = *segp;
715         sg = &segments[seg - 1];
716
717         vaddr = (vm_offset_t)buf;
718         nextpaddr = *lastpaddrp;
719         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
720
721         /* force at least one segment */
722         do {
723                 bus_size_t size;
724
725                 /*
726                  * Per-page main loop
727                  */
728                 paddr = _bus_dma_extract(pmap, vaddr);
729                 size = PAGE_SIZE - (paddr & PAGE_MASK);
730                 if (size > buflen)
731                         size = buflen;
732                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
733                         /*
734                          * note: this paddr has the same in-page offset
735                          * as vaddr and thus the paddr above, so the
736                          * size does not have to be recalculated
737                          */
738                         paddr = add_bounce_page(dmat, map, vaddr, size);
739                 }
740
741                 /*
742                  * Fill in the bus_dma_segment
743                  */
744                 if (first) {
745                         sg->ds_addr = paddr;
746                         sg->ds_len = size;
747                         first = 0;
748                 } else if (paddr == nextpaddr) {
749                         sg->ds_len += size;
750                 } else {
751                         sg++;
752                         seg++;
753                         if (seg > nsegments)
754                                 break;
755                         sg->ds_addr = paddr;
756                         sg->ds_len = size;
757                 }
758                 nextpaddr = paddr + size;
759
760                 /*
761                  * Handle maxsegsz and boundary issues with a nested loop
762                  */
763                 for (;;) {
764                         bus_size_t tmpsize;
765
766                         /*
767                          * Limit to the boundary and maximum segment size
768                          */
769                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
770                                 tmpsize = dmat->boundary -
771                                           (sg->ds_addr & ~bmask);
772                                 if (tmpsize > dmat->maxsegsz)
773                                         tmpsize = dmat->maxsegsz;
774                                 KKASSERT(tmpsize < sg->ds_len);
775                         } else if (sg->ds_len > dmat->maxsegsz) {
776                                 tmpsize = dmat->maxsegsz;
777                         } else {
778                                 break;
779                         }
780
781                         /*
782                          * Futz, split the data into a new segment.
783                          */
784                         if (seg >= nsegments)
785                                 goto fail;
786                         sg[1].ds_len = sg[0].ds_len - tmpsize;
787                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
788                         sg[0].ds_len = tmpsize;
789                         sg++;
790                         seg++;
791                 }
792
793                 /*
794                  * Adjust for loop
795                  */
796                 buflen -= size;
797                 vaddr += size;
798         } while (buflen > 0);
799 fail:
800         if (buflen != 0)
801                 error = EFBIG;
802
803         *segp = seg;
804         *lastpaddrp = nextpaddr;
805
806 free_bounce:
807         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
808             map != &nobounce_dmamap) {
809                 _bus_dmamap_unload(dmat, map);
810                 return_bounce_pages(dmat, map);
811         }
812         return error;
813 }
814
815 /*
816  * Map the buffer buf into bus space using the dmamap map.
817  */
818 int
819 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
820                 bus_size_t buflen, bus_dmamap_callback_t *callback,
821                 void *callback_arg, int flags)
822 {
823         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
824         bus_dma_segment_t *segments;
825         vm_paddr_t lastaddr = 0;
826         int error, nsegs = 1;
827
828         if (map != NULL) {
829                 /*
830                  * XXX
831                  * Follow old semantics.  Once all of the callers are fixed,
832                  * we should get rid of these internal flag "adjustment".
833                  */
834                 flags &= ~BUS_DMA_NOWAIT;
835                 flags |= BUS_DMA_WAITOK;
836
837                 map->callback = callback;
838                 map->callback_arg = callback_arg;
839         }
840
841         segments = bus_dma_tag_lock(dmat, cache_segments);
842         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
843                         segments, dmat->nsegments,
844                         NULL, flags, &lastaddr, &nsegs, 1);
845         if (error == EINPROGRESS) {
846                 KKASSERT((dmat->flags &
847                           (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL)) !=
848                          (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL));
849
850                 if (dmat->flags & BUS_DMA_PROTECTED)
851                         panic("protected dmamap callback will be defered");
852
853                 bus_dma_tag_unlock(dmat);
854                 return error;
855         }
856         callback(callback_arg, segments, nsegs, error);
857         bus_dma_tag_unlock(dmat);
858         return 0;
859 }
860
861 /*
862  * Like _bus_dmamap_load(), but for mbufs.
863  */
864 int
865 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
866                      struct mbuf *m0,
867                      bus_dmamap_callback2_t *callback, void *callback_arg,
868                      int flags)
869 {
870         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
871         bus_dma_segment_t *segments;
872         int nsegs, error;
873
874         /*
875          * XXX
876          * Follow old semantics.  Once all of the callers are fixed,
877          * we should get rid of these internal flag "adjustment".
878          */
879         flags &= ~BUS_DMA_WAITOK;
880         flags |= BUS_DMA_NOWAIT;
881
882         segments = bus_dma_tag_lock(dmat, cache_segments);
883         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
884                         segments, dmat->nsegments, &nsegs, flags);
885         if (error) {
886                 /* force "no valid mappings" in callback */
887                 callback(callback_arg, segments, 0,
888                          0, error);
889         } else {
890                 callback(callback_arg, segments, nsegs,
891                          m0->m_pkthdr.len, error);
892         }
893         bus_dma_tag_unlock(dmat);
894         return error;
895 }
896
897 int
898 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
899                              struct mbuf *m0,
900                              bus_dma_segment_t *segs, int maxsegs,
901                              int *nsegs, int flags)
902 {
903         int error;
904
905         M_ASSERTPKTHDR(m0);
906
907         KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs));
908         KASSERT(maxsegs <= dmat->nsegments,
909                 ("%d too many segments, dmat only supports %d segments",
910                  maxsegs, dmat->nsegments));
911         KASSERT(flags & BUS_DMA_NOWAIT,
912                 ("only BUS_DMA_NOWAIT is supported"));
913
914         if (m0->m_pkthdr.len <= dmat->maxsize) {
915                 int first = 1;
916                 vm_paddr_t lastaddr = 0;
917                 struct mbuf *m;
918
919                 *nsegs = 1;
920                 error = 0;
921                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
922                         if (m->m_len == 0)
923                                 continue;
924
925                         error = _bus_dmamap_load_buffer(dmat, map,
926                                         m->m_data, m->m_len,
927                                         segs, maxsegs,
928                                         NULL, flags, &lastaddr,
929                                         nsegs, first);
930                         if (error == ENOMEM && !first) {
931                                 /*
932                                  * Out of bounce pages due to too many
933                                  * fragments in the mbuf chain; return
934                                  * EFBIG instead.
935                                  */
936                                 error = EFBIG;
937                         }
938                         first = 0;
939                 }
940 #ifdef INVARIANTS
941                 if (!error)
942                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
943 #endif
944         } else {
945                 *nsegs = 0;
946                 error = EINVAL;
947         }
948         KKASSERT(error != EINPROGRESS);
949         return error;
950 }
951
952 /*
953  * Like _bus_dmamap_load(), but for uios.
954  */
955 int
956 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
957                     struct uio *uio,
958                     bus_dmamap_callback2_t *callback, void *callback_arg,
959                     int flags)
960 {
961         vm_paddr_t lastaddr;
962         int nsegs, error, first, i;
963         bus_size_t resid;
964         struct iovec *iov;
965         pmap_t pmap;
966         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
967         bus_dma_segment_t *segments;
968         bus_dma_segment_t *segs;
969         int nsegs_left;
970
971         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
972                 segments = cache_segments;
973         else
974                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
975                                    M_DEVBUF, M_WAITOK | M_ZERO);
976
977         /*
978          * XXX
979          * Follow old semantics.  Once all of the callers are fixed,
980          * we should get rid of these internal flag "adjustment".
981          */
982         flags &= ~BUS_DMA_WAITOK;
983         flags |= BUS_DMA_NOWAIT;
984
985         resid = (bus_size_t)uio->uio_resid;
986         iov = uio->uio_iov;
987
988         segs = segments;
989         nsegs_left = dmat->nsegments;
990
991         if (uio->uio_segflg == UIO_USERSPACE) {
992                 struct thread *td;
993
994                 td = uio->uio_td;
995                 KASSERT(td != NULL && td->td_proc != NULL,
996                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
997                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
998         } else {
999                 pmap = NULL;
1000         }
1001
1002         error = 0;
1003         nsegs = 1;
1004         first = 1;
1005         lastaddr = 0;
1006         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
1007                 /*
1008                  * Now at the first iovec to load.  Load each iovec
1009                  * until we have exhausted the residual count.
1010                  */
1011                 bus_size_t minlen =
1012                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
1013                 caddr_t addr = (caddr_t) iov[i].iov_base;
1014
1015                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
1016                                 segs, nsegs_left,
1017                                 pmap, flags, &lastaddr, &nsegs, first);
1018                 first = 0;
1019
1020                 resid -= minlen;
1021                 if (error == 0) {
1022                         nsegs_left -= nsegs;
1023                         segs += nsegs;
1024                 }
1025         }
1026
1027         /*
1028          * Minimum one DMA segment, even if 0-length buffer.
1029          */
1030         if (nsegs_left == dmat->nsegments)
1031                 --nsegs_left;
1032
1033         if (error) {
1034                 /* force "no valid mappings" in callback */
1035                 callback(callback_arg, segments, 0,
1036                          0, error);
1037         } else {
1038                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1039                          (bus_size_t)uio->uio_resid, error);
1040         }
1041         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1042                 kfree(segments, M_DEVBUF);
1043         return error;
1044 }
1045
1046 /*
1047  * Release the mapping held by map.
1048  */
1049 void
1050 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1051 {
1052         struct bounce_page *bpage;
1053
1054         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1055                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1056                 free_bounce_page(dmat, bpage);
1057         }
1058 }
1059
1060 void
1061 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1062 {
1063         struct bounce_page *bpage;
1064
1065         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1066                 /*
1067                  * Handle data bouncing.  We might also
1068                  * want to add support for invalidating
1069                  * the caches on broken hardware
1070                  */
1071                 switch (op) {
1072                 case BUS_DMASYNC_PREWRITE:
1073                         while (bpage != NULL) {
1074                                 bcopy((void *)bpage->datavaddr,
1075                                       (void *)bpage->vaddr,
1076                                       bpage->datacount);
1077                                 bpage = STAILQ_NEXT(bpage, links);
1078                         }
1079                         dmat->bounce_zone->total_bounced++;
1080                         break;
1081
1082                 case BUS_DMASYNC_POSTREAD:
1083                         while (bpage != NULL) {
1084                                 bcopy((void *)bpage->vaddr,
1085                                       (void *)bpage->datavaddr,
1086                                       bpage->datacount);
1087                                 bpage = STAILQ_NEXT(bpage, links);
1088                         }
1089                         dmat->bounce_zone->total_bounced++;
1090                         break;
1091
1092                 case BUS_DMASYNC_PREREAD:
1093                 case BUS_DMASYNC_POSTWRITE:
1094                         /* No-ops */
1095                         break;
1096                 }
1097         }
1098 }
1099
1100 static int
1101 alloc_bounce_zone(bus_dma_tag_t dmat)
1102 {
1103         struct bounce_zone *bz, *new_bz;
1104
1105         KASSERT(dmat->bounce_zone == NULL,
1106                 ("bounce zone was already assigned"));
1107
1108         new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO);
1109
1110         lwkt_gettoken(&bounce_zone_tok);
1111
1112         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1113                 /*
1114                  * For shared bounce zone, check to see
1115                  * if we already have a suitable zone
1116                  */
1117                 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1118                         if (dmat->alignment <= bz->alignment &&
1119                             dmat->lowaddr >= bz->lowaddr) {
1120                                 lwkt_reltoken(&bounce_zone_tok);
1121
1122                                 dmat->bounce_zone = bz;
1123                                 kfree(new_bz, M_DEVBUF);
1124                                 return 0;
1125                         }
1126                 }
1127         }
1128         bz = new_bz;
1129
1130 #ifdef SMP
1131         spin_init(&bz->spin);
1132 #endif
1133         STAILQ_INIT(&bz->bounce_page_list);
1134         STAILQ_INIT(&bz->bounce_map_waitinglist);
1135         bz->free_bpages = 0;
1136         bz->reserved_bpages = 0;
1137         bz->active_bpages = 0;
1138         bz->lowaddr = dmat->lowaddr;
1139         bz->alignment = round_page(dmat->alignment);
1140         ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1141
1142         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1143                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1144                 busdma_zonecount++;
1145                 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1146         } else {
1147                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_priv_zonecount);
1148                 busdma_priv_zonecount--;
1149         }
1150
1151         lwkt_reltoken(&bounce_zone_tok);
1152
1153         dmat->bounce_zone = bz;
1154
1155         sysctl_ctx_init(&bz->sysctl_ctx);
1156         bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx,
1157             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1158             CTLFLAG_RD, 0, "");
1159         if (bz->sysctl_tree == NULL) {
1160                 sysctl_ctx_free(&bz->sysctl_ctx);
1161                 return 0;       /* XXX error code? */
1162         }
1163
1164         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1165             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1166             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1167             "Total bounce pages");
1168         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1169             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1170             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1171             "Free bounce pages");
1172         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1173             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1174             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1175             "Reserved bounce pages");
1176         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1177             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1178             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1179             "Active bounce pages");
1180         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1181             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1182             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1183             "Total bounce requests");
1184         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1185             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1186             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1187             "Total bounce requests that were deferred");
1188         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1189             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1190             "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0,
1191             "Total bounce page reservations that were failed");
1192         SYSCTL_ADD_STRING(&bz->sysctl_ctx,
1193             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1194             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1195         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1196             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1197             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1198
1199         return 0;
1200 }
1201
1202 static int
1203 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags)
1204 {
1205         struct bounce_zone *bz = dmat->bounce_zone;
1206         int count = 0, mflags;
1207
1208         if (flags & BUS_DMA_NOWAIT)
1209                 mflags = M_NOWAIT;
1210         else
1211                 mflags = M_WAITOK;
1212
1213         while (numpages > 0) {
1214                 struct bounce_page *bpage;
1215
1216                 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO);
1217
1218                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1219                                                          mflags, 0ul,
1220                                                          bz->lowaddr,
1221                                                          bz->alignment, 0);
1222                 if (bpage->vaddr == 0) {
1223                         kfree(bpage, M_DEVBUF);
1224                         break;
1225                 }
1226                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1227
1228                 BZ_LOCK(bz);
1229                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1230                 total_bounce_pages++;
1231                 bz->total_bpages++;
1232                 bz->free_bpages++;
1233                 BZ_UNLOCK(bz);
1234
1235                 count++;
1236                 numpages--;
1237         }
1238         return count;
1239 }
1240
1241 static void
1242 free_bounce_pages_all(bus_dma_tag_t dmat)
1243 {
1244         struct bounce_zone *bz = dmat->bounce_zone;
1245         struct bounce_page *bpage;
1246
1247         BZ_LOCK(bz);
1248
1249         while ((bpage = STAILQ_FIRST(&bz->bounce_page_list)) != NULL) {
1250                 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1251
1252                 KKASSERT(total_bounce_pages > 0);
1253                 total_bounce_pages--;
1254
1255                 KKASSERT(bz->total_bpages > 0);
1256                 bz->total_bpages--;
1257
1258                 KKASSERT(bz->free_bpages > 0);
1259                 bz->free_bpages--;
1260
1261                 BZ_UNLOCK(bz);
1262                 contigfree((void *)bpage->vaddr, PAGE_SIZE, M_DEVBUF);
1263                 kfree(bpage, M_DEVBUF);
1264                 BZ_LOCK(bz);
1265         }
1266         if (bz->total_bpages) {
1267                 kprintf("#%d bounce pages are still in use\n",
1268                         bz->total_bpages);
1269                 print_backtrace(-1);
1270         }
1271
1272         BZ_UNLOCK(bz);
1273 }
1274
1275 static void
1276 free_bounce_zone(bus_dma_tag_t dmat)
1277 {
1278         struct bounce_zone *bz = dmat->bounce_zone;
1279
1280         if (bz == NULL)
1281                 return;
1282
1283         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0)
1284                 return;
1285
1286         free_bounce_pages_all(dmat);
1287         dmat->bounce_zone = NULL;
1288
1289         if (bz->sysctl_tree != NULL)
1290                 sysctl_ctx_free(&bz->sysctl_ctx);
1291         kfree(bz, M_DEVBUF);
1292 }
1293
1294 /* Assume caller holds bounce zone spinlock */
1295 static int
1296 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1297 {
1298         struct bounce_zone *bz = dmat->bounce_zone;
1299         int pages;
1300
1301         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1302         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1303                 bz->reserve_failed++;
1304                 return (map->pagesneeded - (map->pagesreserved + pages));
1305         }
1306
1307         bz->free_bpages -= pages;
1308
1309         bz->reserved_bpages += pages;
1310         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1311
1312         map->pagesreserved += pages;
1313         pages = map->pagesneeded - map->pagesreserved;
1314
1315         return pages;
1316 }
1317
1318 static void
1319 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1320 {
1321         struct bounce_zone *bz = dmat->bounce_zone;
1322         int reserved = map->pagesreserved;
1323         bus_dmamap_t wait_map;
1324
1325         map->pagesreserved = 0;
1326         map->pagesneeded = 0;
1327
1328         if (reserved == 0)
1329                 return;
1330
1331         BZ_LOCK(bz);
1332
1333         bz->free_bpages += reserved;
1334         KKASSERT(bz->free_bpages <= bz->total_bpages);
1335
1336         KKASSERT(bz->reserved_bpages >= reserved);
1337         bz->reserved_bpages -= reserved;
1338
1339         wait_map = get_map_waiting(dmat);
1340
1341         BZ_UNLOCK(bz);
1342
1343         if (wait_map != NULL)
1344                 add_map_callback(map);
1345 }
1346
1347 static bus_addr_t
1348 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1349                 bus_size_t size)
1350 {
1351         struct bounce_zone *bz = dmat->bounce_zone;
1352         struct bounce_page *bpage;
1353
1354         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1355         map->pagesneeded--;
1356
1357         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1358         map->pagesreserved--;
1359
1360         BZ_LOCK(bz);
1361
1362         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1363         KASSERT(bpage != NULL, ("free page list is empty"));
1364         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1365
1366         KKASSERT(bz->reserved_bpages > 0);
1367         bz->reserved_bpages--;
1368
1369         bz->active_bpages++;
1370         KKASSERT(bz->active_bpages <= bz->total_bpages);
1371
1372         BZ_UNLOCK(bz);
1373
1374         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1375                 /* Page offset needs to be preserved. */
1376                 bpage->vaddr |= vaddr & PAGE_MASK;
1377                 bpage->busaddr |= vaddr & PAGE_MASK;
1378         }
1379         bpage->datavaddr = vaddr;
1380         bpage->datacount = size;
1381         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1382         return bpage->busaddr;
1383 }
1384
1385 static void
1386 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1387 {
1388         struct bounce_zone *bz = dmat->bounce_zone;
1389         bus_dmamap_t map;
1390
1391         bpage->datavaddr = 0;
1392         bpage->datacount = 0;
1393
1394         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1395                 /*
1396                  * Reset the bounce page to start at offset 0.  Other uses
1397                  * of this bounce page may need to store a full page of
1398                  * data and/or assume it starts on a page boundary.
1399                  */
1400                 bpage->vaddr &= ~PAGE_MASK;
1401                 bpage->busaddr &= ~PAGE_MASK;
1402         }
1403
1404         BZ_LOCK(bz);
1405
1406         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1407
1408         bz->free_bpages++;
1409         KKASSERT(bz->free_bpages <= bz->total_bpages);
1410
1411         KKASSERT(bz->active_bpages > 0);
1412         bz->active_bpages--;
1413
1414         map = get_map_waiting(dmat);
1415
1416         BZ_UNLOCK(bz);
1417
1418         if (map != NULL)
1419                 add_map_callback(map);
1420 }
1421
1422 /* Assume caller holds bounce zone spinlock */
1423 static bus_dmamap_t
1424 get_map_waiting(bus_dma_tag_t dmat)
1425 {
1426         struct bounce_zone *bz = dmat->bounce_zone;
1427         bus_dmamap_t map;
1428
1429         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1430         if (map != NULL) {
1431                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1432                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1433                         bz->total_deferred++;
1434                 } else {
1435                         map = NULL;
1436                 }
1437         }
1438         return map;
1439 }
1440
1441 static void
1442 add_map_callback(bus_dmamap_t map)
1443 {
1444         spin_lock(&bounce_map_list_spin);
1445         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1446         busdma_swi_pending = 1;
1447         setsoftvm();
1448         spin_unlock(&bounce_map_list_spin);
1449 }
1450
1451 void
1452 busdma_swi(void)
1453 {
1454         bus_dmamap_t map;
1455
1456         spin_lock(&bounce_map_list_spin);
1457         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1458                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1459                 spin_unlock(&bounce_map_list_spin);
1460                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1461                                 map->callback, map->callback_arg, /*flags*/0);
1462                 spin_lock(&bounce_map_list_spin);
1463         }
1464         spin_unlock(&bounce_map_list_spin);
1465 }