sys/arm/arm/busdma_machdep.c

   1 /*-
   2  * Copyright (c) 2004 Olivier Houchard
   3  * Copyright (c) 2002 Peter Grehan
   4  * Copyright (c) 1997, 1998 Justin T. Gibbs.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions, and the following disclaimer,
  12  *    without modification, immediately at the beginning of the file.
  13  * 2. The name of the author may not be used to endorse or promote products
  14  *    derived from this software without specific prior written permission.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  20  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  *
  28  *   From i386/busdma_machdep.c,v 1.26 2002/04/19 22:58:09 alfred
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 /*
  35  * ARM bus dma support routines
  36  */
  37
  38 #define _ARM32_BUS_DMA_PRIVATE
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/malloc.h>
  42 #include <sys/bus.h>
  43 #include <sys/interrupt.h>
  44 #include <sys/lock.h>
  45 #include <sys/proc.h>
  46 #include <sys/mutex.h>
  47 #include <sys/mbuf.h>
  48 #include <sys/uio.h>
  49 #include <sys/ktr.h>
  50 #include <sys/kernel.h>
  51 #include <sys/sysctl.h>
  52
  53 #include <vm/vm.h>
  54 #include <vm/vm_page.h>
  55 #include <vm/vm_map.h>
  56
  57 #include <machine/atomic.h>
  58 #include <machine/bus.h>
  59 #include <machine/cpufunc.h>
  60 #include <machine/md_var.h>
  61
  62 #define MAX_BPAGES 64
  63 #define BUS_DMA_COULD_BOUNCE    BUS_DMA_BUS3
  64 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  65
  66 struct bounce_zone;
  67
  68 struct bus_dma_tag {
  69         bus_dma_tag_t           parent;
  70         bus_size_t              alignment;
  71         bus_size_t              boundary;
  72         bus_addr_t              lowaddr;
  73         bus_addr_t              highaddr;
  74         bus_dma_filter_t        *filter;
  75         void                    *filterarg;
  76         bus_size_t              maxsize;
  77         u_int                   nsegments;
  78         bus_size_t              maxsegsz;
  79         int                     flags;
  80         int                     ref_count;
  81         int                     map_count;
  82         bus_dma_lock_t          *lockfunc;
  83         void                    *lockfuncarg;
  84         /*
  85          * DMA range for this tag.  If the page doesn't fall within
  86          * one of these ranges, an error is returned.  The caller
  87          * may then decide what to do with the transfer.  If the
  88          * range pointer is NULL, it is ignored.
  89          */
  90         struct arm32_dma_range  *ranges;
  91         int                     _nranges;
  92         struct bounce_zone *bounce_zone;
  93 };
  94
  95 struct bounce_page {
  96         vm_offset_t     vaddr;          /* kva of bounce buffer */
  97         vm_offset_t     vaddr_nocache;  /* kva of bounce buffer uncached */
  98         bus_addr_t      busaddr;        /* Physical address */
  99         vm_offset_t     datavaddr;      /* kva of client data */
 100         bus_size_t      datacount;      /* client data count */
 101         STAILQ_ENTRY(bounce_page) links;
 102 };
 103
 104 int busdma_swi_pending;
 105
 106 struct bounce_zone {
 107         STAILQ_ENTRY(bounce_zone) links;
 108         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 109         int             total_bpages;
 110         int             free_bpages;
 111         int             reserved_bpages;
 112         int             active_bpages;
 113         int             total_bounced;
 114         int             total_deferred;
 115         bus_size_t      alignment;
 116         bus_size_t      boundary;
 117         bus_addr_t      lowaddr;
 118         char            zoneid[8];
 119         char            lowaddrid[20];
 120         struct sysctl_ctx_list sysctl_tree;
 121         struct sysctl_oid *sysctl_tree_top;
 122 };
 123
 124 static struct mtx bounce_lock;
 125 static int total_bpages;
 126 static int busdma_zonecount;
 127 static STAILQ_HEAD(, bounce_zone) bounce_zone_list;
 128
 129 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 130 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
 131            "Total bounce pages");
 132
 133 #define DMAMAP_LINEAR           0x1
 134 #define DMAMAP_MBUF             0x2
 135 #define DMAMAP_UIO              0x4
 136 #define DMAMAP_ALLOCATED        0x10
 137 #define DMAMAP_TYPE_MASK        (DMAMAP_LINEAR|DMAMAP_MBUF|DMAMAP_UIO)
 138 #define DMAMAP_COHERENT         0x8
 139 struct bus_dmamap {
 140         struct bp_list  bpages;
 141         int             pagesneeded;
 142         int             pagesreserved;
 143         bus_dma_tag_t   dmat;
 144         int             flags;
 145         void            *buffer;
 146         void            *origbuffer;
 147         void            *allocbuffer;
 148         TAILQ_ENTRY(bus_dmamap) freelist;
 149         int             len;
 150         STAILQ_ENTRY(bus_dmamap) links;
 151         bus_dmamap_callback_t *callback;
 152         void                  *callback_arg;
 153
 154 };
 155
 156 static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 157 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
 158
 159 static TAILQ_HEAD(,bus_dmamap) dmamap_freelist =
 160         TAILQ_HEAD_INITIALIZER(dmamap_freelist);
 161
 162 #define BUSDMA_STATIC_MAPS      500
 163 static struct bus_dmamap map_pool[BUSDMA_STATIC_MAPS];
 164
 165 static struct mtx busdma_mtx;
 166
 167 MTX_SYSINIT(busdma_mtx, &busdma_mtx, "busdma lock", MTX_DEF);
 168
 169 static void init_bounce_pages(void *dummy);
 170 static int alloc_bounce_zone(bus_dma_tag_t dmat);
 171 static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
 172 static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
 173                                 int commit);
 174 static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
 175                                    vm_offset_t vaddr, bus_size_t size);
 176 static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
 177
 178 /* Default tag, as most drivers provide no parent tag. */
 179 bus_dma_tag_t arm_root_dma_tag;
 180
 181 /*
 182  * Return true if a match is made.
 183  *
 184  * To find a match walk the chain of bus_dma_tag_t's looking for 'paddr'.
 185  *
 186  * If paddr is within the bounds of the dma tag then call the filter callback
 187  * to check for a match, if there is no filter callback then assume a match.
 188  */
 189 static int
 190 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 191 {
 192         int retval;
 193
 194         retval = 0;
 195
 196         do {
 197                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr)
 198                  || ((paddr & (dmat->alignment - 1)) != 0))
 199                  && (dmat->filter == NULL
 200                   || (*dmat->filter)(dmat->filterarg, paddr) != 0))
 201                         retval = 1;
 202
 203                 dmat = dmat->parent;
 204         } while (retval == 0 && dmat != NULL);
 205         return (retval);
 206 }
 207
 208 static void
 209 arm_dmamap_freelist_init(void *dummy)
 210 {
 211         int i;
 212
 213         for (i = 0; i < BUSDMA_STATIC_MAPS; i++)
 214                 TAILQ_INSERT_HEAD(&dmamap_freelist, &map_pool[i], freelist);
 215 }
 216
 217 SYSINIT(busdma, SI_SUB_VM, SI_ORDER_ANY, arm_dmamap_freelist_init, NULL);
 218
 219 /*
 220  * Check to see if the specified page is in an allowed DMA range.
 221  */
 222
 223 static __inline int
 224 bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dma_segment_t *segs,
 225     bus_dmamap_t map, void *buf, bus_size_t buflen, struct pmap *pmap,
 226     int flags, vm_offset_t *lastaddrp, int *segp);
 227
 228 static __inline int
 229 _bus_dma_can_bounce(vm_offset_t lowaddr, vm_offset_t highaddr)
 230 {
 231         int i;
 232         for (i = 0; phys_avail[i] && phys_avail[i + 1]; i += 2) {
 233                 if ((lowaddr >= phys_avail[i] && lowaddr <= phys_avail[i + 1])
 234                     || (lowaddr < phys_avail[i] &&
 235                     highaddr > phys_avail[i]))
 236                         return (1);
 237         }
 238         return (0);
 239 }
 240
 241 static __inline struct arm32_dma_range *
 242 _bus_dma_inrange(struct arm32_dma_range *ranges, int nranges,
 243     bus_addr_t curaddr)
 244 {
 245         struct arm32_dma_range *dr;
 246         int i;
 247
 248         for (i = 0, dr = ranges; i < nranges; i++, dr++) {
 249                 if (curaddr >= dr->dr_sysbase &&
 250                     round_page(curaddr) <= (dr->dr_sysbase + dr->dr_len))
 251                         return (dr);
 252         }
 253
 254         return (NULL);
 255 }
 256 /*
 257  * Convenience function for manipulating driver locks from busdma (during
 258  * busdma_swi, for example).  Drivers that don't provide their own locks
 259  * should specify &Giant to dmat->lockfuncarg.  Drivers that use their own
 260  * non-mutex locking scheme don't have to use this at all.
 261  */
 262 void
 263 busdma_lock_mutex(void *arg, bus_dma_lock_op_t op)
 264 {
 265         struct mtx *dmtx;
 266
 267         dmtx = (struct mtx *)arg;
 268         switch (op) {
 269         case BUS_DMA_LOCK:
 270                 mtx_lock(dmtx);
 271                 break;
 272         case BUS_DMA_UNLOCK:
 273                 mtx_unlock(dmtx);
 274                 break;
 275         default:
 276                 panic("Unknown operation 0x%x for busdma_lock_mutex!", op);
 277         }
 278 }
 279
 280 /*
 281  * dflt_lock should never get called.  It gets put into the dma tag when
 282  * lockfunc == NULL, which is only valid if the maps that are associated
 283  * with the tag are meant to never be defered.
 284  * XXX Should have a way to identify which driver is responsible here.
 285  */
 286 static void
 287 dflt_lock(void *arg, bus_dma_lock_op_t op)
 288 {
 289 #ifdef INVARIANTS
 290         panic("driver error: busdma dflt_lock called");
 291 #else
 292         printf("DRIVER_ERROR: busdma dflt_lock called\n");
 293 #endif
 294 }
 295
 296 static __inline bus_dmamap_t
 297 _busdma_alloc_dmamap(void)
 298 {
 299         bus_dmamap_t map;
 300
 301         mtx_lock(&busdma_mtx);
 302         map = TAILQ_FIRST(&dmamap_freelist);
 303         if (map)
 304                 TAILQ_REMOVE(&dmamap_freelist, map, freelist);
 305         mtx_unlock(&busdma_mtx);
 306         if (!map) {
 307                 map = malloc(sizeof(*map), M_DEVBUF, M_NOWAIT | M_ZERO);
 308                 if (map)
 309                         map->flags = DMAMAP_ALLOCATED;
 310         } else
 311                 map->flags = 0;
 312         STAILQ_INIT(&map->bpages);
 313         return (map);
 314 }
 315
 316 static __inline void
 317 _busdma_free_dmamap(bus_dmamap_t map)
 318 {
 319         if (map->flags & DMAMAP_ALLOCATED)
 320                 free(map, M_DEVBUF);
 321         else {
 322                 mtx_lock(&busdma_mtx);
 323                 TAILQ_INSERT_HEAD(&dmamap_freelist, map, freelist);
 324                 mtx_unlock(&busdma_mtx);
 325         }
 326 }
 327
 328 /*
 329  * Allocate a device specific dma_tag.
 330  */
 331 #define SEG_NB 1024
 332
 333 int
 334 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 335                    bus_size_t boundary, bus_addr_t lowaddr,
 336                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 337                    void *filterarg, bus_size_t maxsize, int nsegments,
 338                    bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
 339                    void *lockfuncarg, bus_dma_tag_t *dmat)
 340 {
 341         bus_dma_tag_t newtag;
 342         int error = 0;
 343         /* Return a NULL tag on failure */
 344         *dmat = NULL;
 345         if (!parent)
 346                 parent = arm_root_dma_tag;
 347
 348         newtag = (bus_dma_tag_t)malloc(sizeof(*newtag), M_DEVBUF, M_NOWAIT);
 349         if (newtag == NULL) {
 350                 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
 351                     __func__, newtag, 0, error);
 352                 return (ENOMEM);
 353         }
 354
 355         newtag->parent = parent;
 356         newtag->alignment = alignment;
 357         newtag->boundary = boundary;
 358         newtag->lowaddr = trunc_page((vm_offset_t)lowaddr) + (PAGE_SIZE - 1);
 359         newtag->highaddr = trunc_page((vm_offset_t)highaddr) + (PAGE_SIZE - 1);
 360         newtag->filter = filter;
 361         newtag->filterarg = filterarg;
 362         newtag->maxsize = maxsize;
 363         newtag->nsegments = nsegments;
 364         newtag->maxsegsz = maxsegsz;
 365         newtag->flags = flags;
 366         newtag->ref_count = 1; /* Count ourself */
 367         newtag->map_count = 0;
 368         newtag->ranges = bus_dma_get_range();
 369         newtag->_nranges = bus_dma_get_range_nb();
 370         if (lockfunc != NULL) {
 371                 newtag->lockfunc = lockfunc;
 372                 newtag->lockfuncarg = lockfuncarg;
 373         } else {
 374                 newtag->lockfunc = dflt_lock;
 375                 newtag->lockfuncarg = NULL;
 376         }
 377         /*
 378          * Take into account any restrictions imposed by our parent tag
 379          */
 380         if (parent != NULL) {
 381                 newtag->lowaddr = min(parent->lowaddr, newtag->lowaddr);
 382                 newtag->highaddr = max(parent->highaddr, newtag->highaddr);
 383                 if (newtag->boundary == 0)
 384                         newtag->boundary = parent->boundary;
 385                 else if (parent->boundary != 0)
 386                         newtag->boundary = min(parent->boundary,
 387                                                newtag->boundary);
 388                 if ((newtag->filter != NULL) ||
 389                     ((parent->flags & BUS_DMA_COULD_BOUNCE) != 0))
 390                         newtag->flags |= BUS_DMA_COULD_BOUNCE;
 391                 if (newtag->filter == NULL) {
 392                         /*
 393                          * Short circuit looking at our parent directly
 394                          * since we have encapsulated all of its information
 395                          */
 396                         newtag->filter = parent->filter;
 397                         newtag->filterarg = parent->filterarg;
 398                         newtag->parent = parent->parent;
 399                 }
 400                 if (newtag->parent != NULL)
 401                         atomic_add_int(&parent->ref_count, 1);
 402         }
 403         if (_bus_dma_can_bounce(newtag->lowaddr, newtag->highaddr)
 404          || newtag->alignment > 1)
 405                 newtag->flags |= BUS_DMA_COULD_BOUNCE;
 406
 407         if (((newtag->flags & BUS_DMA_COULD_BOUNCE) != 0) &&
 408             (flags & BUS_DMA_ALLOCNOW) != 0) {
 409                 struct bounce_zone *bz;
 410
 411                 /* Must bounce */
 412
 413                 if ((error = alloc_bounce_zone(newtag)) != 0) {
 414                         free(newtag, M_DEVBUF);
 415                         return (error);
 416                 }
 417                 bz = newtag->bounce_zone;
 418
 419                 if (ptoa(bz->total_bpages) < maxsize) {
 420                         int pages;
 421
 422                         pages = atop(maxsize) - bz->total_bpages;
 423
 424                         /* Add pages to our bounce pool */
 425                         if (alloc_bounce_pages(newtag, pages) < pages)
 426                                 error = ENOMEM;
 427                 }
 428                 /* Performed initial allocation */
 429                 newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 430         }
 431         if (error != 0)
 432                 free(newtag, M_DEVBUF);
 433         else
 434                 *dmat = newtag;
 435         CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
 436             __func__, newtag, (newtag != NULL ? newtag->flags : 0), error);
 437
 438         return (error);
 439 }
 440
 441 int
 442 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 443 {
 444 #ifdef KTR
 445         bus_dma_tag_t dmat_copy = dmat;
 446 #endif
 447
 448         if (dmat != NULL) {
 449
 450                 if (dmat->map_count != 0)
 451                         return (EBUSY);
 452
 453                 while (dmat != NULL) {
 454                         bus_dma_tag_t parent;
 455
 456                         parent = dmat->parent;
 457                         atomic_subtract_int(&dmat->ref_count, 1);
 458                         if (dmat->ref_count == 0) {
 459                                 free(dmat, M_DEVBUF);
 460                                 /*
 461                                  * Last reference count, so
 462                                  * release our reference
 463                                  * count on our parent.
 464                                  */
 465                                 dmat = parent;
 466                         } else
 467                                 dmat = NULL;
 468                 }
 469         }
 470         CTR2(KTR_BUSDMA, "%s tag %p", __func__, dmat_copy);
 471
 472         return (0);
 473 }
 474
 475 #include <sys/kdb.h>
 476 /*
 477  * Allocate a handle for mapping from kva/uva/physical
 478  * address space into bus device space.
 479  */
 480 int
 481 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 482 {
 483         bus_dmamap_t newmap;
 484         int error = 0;
 485
 486         newmap = _busdma_alloc_dmamap();
 487         if (newmap == NULL) {
 488                 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, ENOMEM);
 489                 return (ENOMEM);
 490         }
 491         *mapp = newmap;
 492         newmap->dmat = dmat;
 493         newmap->allocbuffer = NULL;
 494         dmat->map_count++;
 495
 496         /*
 497          * Bouncing might be required if the driver asks for an active
 498          * exclusion region, a data alignment that is stricter than 1, and/or
 499          * an active address boundary.
 500          */
 501         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 502
 503                 /* Must bounce */
 504                 struct bounce_zone *bz;
 505                 int maxpages;
 506
 507                 if (dmat->bounce_zone == NULL) {
 508                         if ((error = alloc_bounce_zone(dmat)) != 0) {
 509                                 _busdma_free_dmamap(newmap);
 510                                 *mapp = NULL;
 511                                 return (error);
 512                         }
 513                 }
 514                 bz = dmat->bounce_zone;
 515
 516                 /* Initialize the new map */
 517                 STAILQ_INIT(&((*mapp)->bpages));
 518
 519                 /*
 520                  * Attempt to add pages to our pool on a per-instance
 521                  * basis up to a sane limit.
 522                  */
 523                 maxpages = MAX_BPAGES;
 524                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0
 525                  || (dmat->map_count > 0 && bz->total_bpages < maxpages)) {
 526                         int pages;
 527
 528                         pages = MAX(atop(dmat->maxsize), 1);
 529                         pages = MIN(maxpages - bz->total_bpages, pages);
 530                         pages = MAX(pages, 1);
 531                         if (alloc_bounce_pages(dmat, pages) < pages)
 532                                 error = ENOMEM;
 533
 534                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 535                                 if (error == 0)
 536                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 537                         } else {
 538                                 error = 0;
 539                         }
 540                 }
 541         }
 542         CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
 543             __func__, dmat, dmat->flags, error);
 544
 545         return (0);
 546 }
 547
 548 /*
 549  * Destroy a handle for mapping from kva/uva/physical
 550  * address space into bus device space.
 551  */
 552 int
 553 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 554 {
 555
 556         _busdma_free_dmamap(map);
 557         if (STAILQ_FIRST(&map->bpages) != NULL) {
 558                 CTR3(KTR_BUSDMA, "%s: tag %p error %d",
 559                     __func__, dmat, EBUSY);
 560                 return (EBUSY);
 561         }
 562         dmat->map_count--;
 563         CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
 564         return (0);
 565 }
 566
 567 /*
 568  * Allocate a piece of memory that can be efficiently mapped into
 569  * bus device space based on the constraints lited in the dma tag.
 570  * A dmamap to for use with dmamap_load is also allocated.
 571  */
 572 int
 573 bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
 574                  bus_dmamap_t *mapp)
 575 {
 576         bus_dmamap_t newmap = NULL;
 577
 578         int mflags;
 579
 580         if (flags & BUS_DMA_NOWAIT)
 581                 mflags = M_NOWAIT;
 582         else
 583                 mflags = M_WAITOK;
 584         if (flags & BUS_DMA_ZERO)
 585                 mflags |= M_ZERO;
 586
 587         newmap = _busdma_alloc_dmamap();
 588         if (newmap == NULL) {
 589                 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
 590                     __func__, dmat, dmat->flags, ENOMEM);
 591                 return (ENOMEM);
 592         }
 593         dmat->map_count++;
 594         *mapp = newmap;
 595         newmap->dmat = dmat;
 596
 597         if (dmat->maxsize <= PAGE_SIZE &&
 598            (dmat->alignment < dmat->maxsize) &&
 599            !_bus_dma_can_bounce(dmat->lowaddr, dmat->highaddr)) {
 600                 *vaddr = malloc(dmat->maxsize, M_DEVBUF, mflags);
 601         } else {
 602                 /*
 603                  * XXX Use Contigmalloc until it is merged into this facility
 604                  *     and handles multi-seg allocations.  Nobody is doing
 605                  *     multi-seg allocations yet though.
 606                  */
 607                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 608                     0ul, dmat->lowaddr, dmat->alignment? dmat->alignment : 1ul,
 609                     dmat->boundary);
 610         }
 611         if (*vaddr == NULL) {
 612                 if (newmap != NULL) {
 613                         _busdma_free_dmamap(newmap);
 614                         dmat->map_count--;
 615                 }
 616                 *mapp = NULL;
 617                 return (ENOMEM);
 618         }
 619         if (flags & BUS_DMA_COHERENT) {
 620                 void *tmpaddr = arm_remap_nocache(
 621                     (void *)((vm_offset_t)*vaddr &~ PAGE_MASK),
 622                     dmat->maxsize + ((vm_offset_t)*vaddr & PAGE_MASK));
 623
 624                 if (tmpaddr) {
 625                         tmpaddr = (void *)((vm_offset_t)(tmpaddr) +
 626                             ((vm_offset_t)*vaddr & PAGE_MASK));
 627                         newmap->origbuffer = *vaddr;
 628                         newmap->allocbuffer = tmpaddr;
 629                         cpu_idcache_wbinv_range((vm_offset_t)*vaddr,
 630                             dmat->maxsize);
 631                         *vaddr = tmpaddr;
 632                 } else
 633                         newmap->origbuffer = newmap->allocbuffer = NULL;
 634         } else
 635                 newmap->origbuffer = newmap->allocbuffer = NULL;
 636         return (0);
 637 }
 638
 639 /*
 640  * Free a piece of memory and it's allocated dmamap, that was allocated
 641  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 642  */
 643 void
 644 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 645 {
 646         if (map->allocbuffer) {
 647                 KASSERT(map->allocbuffer == vaddr,
 648                     ("Trying to freeing the wrong DMA buffer"));
 649                 vaddr = map->origbuffer;
 650                 arm_unmap_nocache(map->allocbuffer, dmat->maxsize);
 651         }
 652         if (dmat->maxsize <= PAGE_SIZE &&
 653            dmat->alignment < dmat->maxsize &&
 654             !_bus_dma_can_bounce(dmat->lowaddr, dmat->highaddr))
 655                 free(vaddr, M_DEVBUF);
 656         else {
 657                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 658         }
 659         dmat->map_count--;
 660         _busdma_free_dmamap(map);
 661         CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat, dmat->flags);
 662 }
 663
 664 static int
 665 _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 666                         bus_size_t buflen, int flags, int *nb)
 667 {
 668         vm_offset_t vaddr;
 669         vm_offset_t vendaddr;
 670         bus_addr_t paddr;
 671         int needbounce = *nb;
 672
 673         if ((map->pagesneeded == 0)) {
 674                 CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
 675                     "alignment= %d", dmat->lowaddr, ptoa((vm_paddr_t)Maxmem),
 676                     dmat->boundary, dmat->alignment);
 677                 CTR3(KTR_BUSDMA, "map= %p, nobouncemap= %p, pagesneeded= %d",
 678                     map, &nobounce_dmamap, map->pagesneeded);
 679                 /*
 680                  * Count the number of bounce pages
 681                  * needed in order to complete this transfer
 682                  */
 683                 vaddr = trunc_page((vm_offset_t)buf);
 684                 vendaddr = (vm_offset_t)buf + buflen;
 685
 686                 while (vaddr < vendaddr) {
 687                         paddr = pmap_kextract(vaddr);
 688                         if (((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) &&
 689                             run_filter(dmat, paddr) != 0) {
 690                                 needbounce = 1;
 691                                 map->pagesneeded++;
 692                         }
 693                         vaddr += PAGE_SIZE;
 694                 }
 695                 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
 696         }
 697
 698         /* Reserve Necessary Bounce Pages */
 699         if (map->pagesneeded != 0) {
 700                 mtx_lock(&bounce_lock);
 701                 if (flags & BUS_DMA_NOWAIT) {
 702                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 703                                 mtx_unlock(&bounce_lock);
 704                                 return (ENOMEM);
 705                         }
 706                 } else {
 707                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 708                                 /* Queue us for resources */
 709                                 STAILQ_INSERT_TAIL(&bounce_map_waitinglist,
 710                                     map, links);
 711                                 mtx_unlock(&bounce_lock);
 712                                 return (EINPROGRESS);
 713                         }
 714                 }
 715                 mtx_unlock(&bounce_lock);
 716         }
 717
 718         *nb = needbounce;
 719         return (0);
 720 }
 721
 722 /*
 723  * Utility function to load a linear buffer.  lastaddrp holds state
 724  * between invocations (for multiple-buffer loads).  segp contains
 725  * the starting segment on entrance, and the ending segment on exit.
 726  * first indicates if this is the first invocation of this function.
 727  */
 728 static __inline int
 729 bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dma_segment_t *segs,
 730     bus_dmamap_t map, void *buf, bus_size_t buflen, struct pmap *pmap,
 731     int flags, vm_offset_t *lastaddrp, int *segp)
 732 {
 733         bus_size_t sgsize;
 734         bus_addr_t curaddr, lastaddr, baddr, bmask;
 735         vm_offset_t vaddr = (vm_offset_t)buf;
 736         int seg;
 737         int error = 0;
 738         pd_entry_t *pde;
 739         pt_entry_t pte;
 740         pt_entry_t *ptep;
 741         int needbounce = 0;
 742
 743         lastaddr = *lastaddrp;
 744         bmask = ~(dmat->boundary - 1);
 745
 746         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) {
 747                 error = _bus_dmamap_count_pages(dmat, map, buf, buflen, flags,
 748                     &needbounce);
 749                 if (error)
 750                         return (error);
 751         }
 752         CTR3(KTR_BUSDMA, "lowaddr= %d boundary= %d, "
 753             "alignment= %d", dmat->lowaddr, dmat->boundary, dmat->alignment);
 754
 755         for (seg = *segp; buflen > 0 ; ) {
 756                 /*
 757                  * Get the physical address for this segment.
 758                  *
 759                  * XXX Don't support checking for coherent mappings
 760                  * XXX in user address space.
 761                  */
 762                 if (__predict_true(pmap == pmap_kernel())) {
 763                         (void) pmap_get_pde_pte(pmap, vaddr, &pde, &ptep);
 764                         if (__predict_false(pmap_pde_section(pde))) {
 765                                 curaddr = (*pde & L1_S_FRAME) |
 766                                     (vaddr & L1_S_OFFSET);
 767                                 if (*pde & L1_S_CACHE_MASK) {
 768                                         map->flags &=
 769                                             ~DMAMAP_COHERENT;
 770                                 }
 771                         } else {
 772                                 pte = *ptep;
 773                                 KASSERT((pte & L2_TYPE_MASK) != L2_TYPE_INV,
 774                                     ("INV type"));
 775                                 if (__predict_false((pte & L2_TYPE_MASK)
 776                                                     == L2_TYPE_L)) {
 777                                         curaddr = (pte & L2_L_FRAME) |
 778                                             (vaddr & L2_L_OFFSET);
 779                                         if (pte & L2_L_CACHE_MASK) {
 780                                                 map->flags &=
 781                                                     ~DMAMAP_COHERENT;
 782
 783                                         }
 784                                 } else {
 785                                         curaddr = (pte & L2_S_FRAME) |
 786                                             (vaddr & L2_S_OFFSET);
 787                                         if (pte & L2_S_CACHE_MASK) {
 788                                                 map->flags &=
 789                                                     ~DMAMAP_COHERENT;
 790                                         }
 791                                 }
 792                         }
 793                 } else {
 794                         curaddr = pmap_extract(pmap, vaddr);
 795                         map->flags &= ~DMAMAP_COHERENT;
 796                 }
 797
 798                 /*
 799                  * Compute the segment size, and adjust counts.
 800                  */
 801                 sgsize = PAGE_SIZE - ((u_long)curaddr & PAGE_MASK);
 802                 if (buflen < sgsize)
 803                         sgsize = buflen;
 804
 805                 /*
 806                  * Make sure we don't cross any boundaries.
 807                  */
 808                 if (dmat->boundary > 0) {
 809                         baddr = (curaddr + dmat->boundary) & bmask;
 810                         if (sgsize > (baddr - curaddr))
 811                                 sgsize = (baddr - curaddr);
 812                 }
 813                 if (((dmat->flags & BUS_DMA_COULD_BOUNCE) != 0) &&
 814                     map->pagesneeded != 0 && run_filter(dmat, curaddr))
 815                         curaddr = add_bounce_page(dmat, map, vaddr, sgsize);
 816
 817                 if (dmat->ranges) {
 818                         struct arm32_dma_range *dr;
 819
 820                         dr = _bus_dma_inrange(dmat->ranges, dmat->_nranges,
 821                             curaddr);
 822                         if (dr == NULL)
 823                                 return (EINVAL);
 824                         /*
 825                          * In a valid DMA range.  Translate the physical
 826                          * memory address to an address in the DMA window.
 827                          */
 828                         curaddr = (curaddr - dr->dr_sysbase) + dr->dr_busbase;
 829
 830                 }
 831
 832                 /*
 833                  * Insert chunk into a segment, coalescing with
 834                  * the previous segment if possible.
 835                  */
 836                 if (needbounce == 0 && seg >= 0 && curaddr == lastaddr &&
 837                     (segs[seg].ds_len + sgsize) <= dmat->maxsegsz &&
 838                     (dmat->boundary == 0 ||
 839                      (segs[seg].ds_addr & bmask) ==
 840                      (curaddr & bmask))) {
 841                         segs[seg].ds_len += sgsize;
 842                         goto segdone;
 843                 } else {
 844                         if (++seg >= dmat->nsegments)
 845                                 break;
 846                         segs[seg].ds_addr = curaddr;
 847                         segs[seg].ds_len = sgsize;
 848                 }
 849                 if (error)
 850                         break;
 851 segdone:
 852                 lastaddr = curaddr + sgsize;
 853                 vaddr += sgsize;
 854                 buflen -= sgsize;
 855         }
 856
 857         *segp = seg;
 858         *lastaddrp = lastaddr;
 859
 860         /*
 861          * Did we fit?
 862          */
 863         if (buflen != 0)
 864                 error = EFBIG; /* XXX better return value here? */
 865         return (error);
 866 }
 867
 868 /*
 869  * Map the buffer buf into bus space using the dmamap map.
 870  */
 871 int
 872 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 873                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 874                 void *callback_arg, int flags)
 875 {
 876         vm_offset_t     lastaddr = 0;
 877         int             error, nsegs = -1;
 878 #ifdef __CC_SUPPORTS_DYNAMIC_ARRAY_INIT
 879         bus_dma_segment_t dm_segments[dmat->nsegments];
 880 #else
 881         bus_dma_segment_t dm_segments[BUS_DMAMAP_NSEGS];
 882 #endif
 883
 884         KASSERT(dmat != NULL, ("dmatag is NULL"));
 885         KASSERT(map != NULL, ("dmamap is NULL"));
 886         map->callback = callback;
 887         map->callback_arg = callback_arg;
 888         map->flags &= ~DMAMAP_TYPE_MASK;
 889         map->flags |= DMAMAP_LINEAR|DMAMAP_COHERENT;
 890         map->buffer = buf;
 891         map->len = buflen;
 892         error = bus_dmamap_load_buffer(dmat,
 893             dm_segments, map, buf, buflen, kernel_pmap,
 894             flags, &lastaddr, &nsegs);
 895         if (error == EINPROGRESS)
 896                 return (error);
 897         if (error)
 898                 (*callback)(callback_arg, NULL, 0, error);
 899         else
 900                 (*callback)(callback_arg, dm_segments, nsegs + 1, error);
 901
 902         CTR5(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d nsegs %d",
 903             __func__, dmat, dmat->flags, nsegs + 1, error);
 904
 905         return (0);
 906 }
 907
 908 /*
 909  * Like bus_dmamap_load(), but for mbufs.
 910  */
 911 int
 912 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map, struct mbuf *m0,
 913                      bus_dmamap_callback2_t *callback, void *callback_arg,
 914                      int flags)
 915 {
 916 #ifdef __CC_SUPPORTS_DYNAMIC_ARRAY_INIT
 917         bus_dma_segment_t dm_segments[dmat->nsegments];
 918 #else
 919         bus_dma_segment_t dm_segments[BUS_DMAMAP_NSEGS];
 920 #endif
 921         int nsegs = -1, error = 0;
 922
 923         M_ASSERTPKTHDR(m0);
 924
 925         map->flags &= ~DMAMAP_TYPE_MASK;
 926         map->flags |= DMAMAP_MBUF | DMAMAP_COHERENT;
 927         map->buffer = m0;
 928         map->len = 0;
 929         if (m0->m_pkthdr.len <= dmat->maxsize) {
 930                 vm_offset_t lastaddr = 0;
 931                 struct mbuf *m;
 932
 933                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 934                         if (m->m_len > 0) {
 935                                 error = bus_dmamap_load_buffer(dmat,
 936                                     dm_segments, map, m->m_data, m->m_len,
 937                                     pmap_kernel(), flags, &lastaddr, &nsegs);
 938                                 map->len += m->m_len;
 939                         }
 940                 }
 941         } else {
 942                 error = EINVAL;
 943         }
 944
 945         if (error) {
 946                 /*
 947                  * force "no valid mappings" on error in callback.
 948                  */
 949                 (*callback)(callback_arg, dm_segments, 0, 0, error);
 950         } else {
 951                 (*callback)(callback_arg, dm_segments, nsegs + 1,
 952                     m0->m_pkthdr.len, error);
 953         }
 954         CTR5(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d nsegs %d",
 955             __func__, dmat, dmat->flags, error, nsegs + 1);
 956
 957         return (error);
 958 }
 959
 960 int
 961 bus_dmamap_load_mbuf_sg(bus_dma_tag_t dmat, bus_dmamap_t map,
 962                         struct mbuf *m0, bus_dma_segment_t *segs, int *nsegs,
 963                         int flags)
 964 {
 965         int error = 0;
 966         M_ASSERTPKTHDR(m0);
 967
 968         flags |= BUS_DMA_NOWAIT;
 969         *nsegs = -1;
 970         map->flags &= ~DMAMAP_TYPE_MASK;
 971         map->flags |= DMAMAP_MBUF | DMAMAP_COHERENT;
 972         map->buffer = m0;
 973         map->len = 0;
 974         if (m0->m_pkthdr.len <= dmat->maxsize) {
 975                 vm_offset_t lastaddr = 0;
 976                 struct mbuf *m;
 977
 978                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 979                         if (m->m_len > 0) {
 980                                 error = bus_dmamap_load_buffer(dmat, segs, map,
 981                                                 m->m_data, m->m_len,
 982                                                 pmap_kernel(), flags, &lastaddr,
 983                                                 nsegs);
 984                                 map->len += m->m_len;
 985                         }
 986                 }
 987         } else {
 988                 error = EINVAL;
 989         }
 990
 991         /* XXX FIXME: Having to increment nsegs is really annoying */
 992         ++*nsegs;
 993         CTR5(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d nsegs %d",
 994             __func__, dmat, dmat->flags, error, *nsegs);
 995         return (error);
 996 }
 997
 998 /*
 999  * Like bus_dmamap_load(), but for uios.
1000  */
1001 int
1002 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map, struct uio *uio,
1003     bus_dmamap_callback2_t *callback, void *callback_arg,
1004     int flags)
1005 {
1006         vm_offset_t lastaddr;
1007 #ifdef __CC_SUPPORTS_DYNAMIC_ARRAY_INIT
1008         bus_dma_segment_t dm_segments[dmat->nsegments];
1009 #else
1010         bus_dma_segment_t dm_segments[BUS_DMAMAP_NSEGS];
1011 #endif
1012         int nsegs, i, error;
1013         bus_size_t resid;
1014         struct iovec *iov;
1015         struct pmap *pmap;
1016
1017         resid = uio->uio_resid;
1018         iov = uio->uio_iov;
1019         map->flags &= ~DMAMAP_TYPE_MASK;
1020         map->flags |= DMAMAP_UIO|DMAMAP_COHERENT;
1021         map->buffer = uio;
1022         map->len = 0;
1023
1024         if (uio->uio_segflg == UIO_USERSPACE) {
1025                 KASSERT(uio->uio_td != NULL,
1026                     ("bus_dmamap_load_uio: USERSPACE but no proc"));
1027                 pmap = vmspace_pmap(uio->uio_td->td_proc->p_vmspace);
1028         } else
1029                 pmap = kernel_pmap;
1030
1031         error = 0;
1032         nsegs = -1;
1033         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
1034                 /*
1035                  * Now at the first iovec to load.  Load each iovec
1036                  * until we have exhausted the residual count.
1037                  */
1038                 bus_size_t minlen =
1039                     resid < iov[i].iov_len ? resid : iov[i].iov_len;
1040                 caddr_t addr = (caddr_t) iov[i].iov_base;
1041
1042                 if (minlen > 0) {
1043                         error = bus_dmamap_load_buffer(dmat, dm_segments, map,
1044                             addr, minlen, pmap, flags, &lastaddr, &nsegs);
1045
1046                         map->len += minlen;
1047                         resid -= minlen;
1048                 }
1049         }
1050
1051         if (error) {
1052                 /*
1053                  * force "no valid mappings" on error in callback.
1054                  */
1055                 (*callback)(callback_arg, dm_segments, 0, 0, error);
1056         } else {
1057                 (*callback)(callback_arg, dm_segments, nsegs+1,
1058                     uio->uio_resid, error);
1059         }
1060
1061         CTR5(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d nsegs %d",
1062             __func__, dmat, dmat->flags, error, nsegs + 1);
1063         return (error);
1064 }
1065
1066 /*
1067  * Release the mapping held by map.
1068  */
1069 void
1070 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1071 {
1072         struct bounce_page *bpage;
1073
1074         map->flags &= ~DMAMAP_TYPE_MASK;
1075         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1076                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1077                 free_bounce_page(dmat, bpage);
1078         }
1079         return;
1080 }
1081
1082 static __inline void
1083 bus_dmamap_sync_buf(void *buf, int len, bus_dmasync_op_t op)
1084 {
1085         char _tmp_cl[arm_dcache_align], _tmp_clend[arm_dcache_align];
1086
1087         if (op & BUS_DMASYNC_PREWRITE)
1088                 cpu_dcache_wb_range((vm_offset_t)buf, len);
1089         if (op & BUS_DMASYNC_POSTREAD) {
1090                 if ((vm_offset_t)buf & arm_dcache_align_mask)
1091                         memcpy(_tmp_cl, (void *)((vm_offset_t)buf & ~
1092                             arm_dcache_align_mask),
1093                             (vm_offset_t)buf - ((vm_offset_t)buf &~
1094                             arm_dcache_align_mask));
1095                 if (((vm_offset_t)buf + len) & arm_dcache_align_mask)
1096                         memcpy(_tmp_cl, (void *)((vm_offset_t)buf & ~
1097                             arm_dcache_align_mask),
1098                             (vm_offset_t)buf - ((vm_offset_t)buf &~
1099                             arm_dcache_align_mask));
1100                 if (((vm_offset_t)buf + len) & arm_dcache_align_mask)
1101                         memcpy(_tmp_clend, (void *)(((vm_offset_t)buf + len) & ~
1102                             arm_dcache_align_mask),
1103                             (vm_offset_t)buf +len - (((vm_offset_t)buf + len) &~
1104                             arm_dcache_align_mask));
1105                 cpu_dcache_inv_range((vm_offset_t)buf, len);
1106                 if ((vm_offset_t)buf & arm_dcache_align_mask)
1107                         memcpy((void *)((vm_offset_t)buf &
1108                             ~arm_dcache_align_mask),
1109                             _tmp_cl,
1110                             (vm_offset_t)buf - ((vm_offset_t)buf &~
1111                             arm_dcache_align_mask));
1112                 if (((vm_offset_t)buf + len) & arm_dcache_align_mask)
1113                         memcpy((void *)(((vm_offset_t)buf + len) & ~
1114                             arm_dcache_align_mask), _tmp_clend,
1115                             (vm_offset_t)buf +len - (((vm_offset_t)buf + len) &~
1116                             arm_dcache_align_mask));
1117         }
1118 }
1119
1120 static void
1121 _bus_dmamap_sync_bp(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1122 {
1123         struct bounce_page *bpage;
1124
1125         STAILQ_FOREACH(bpage, &map->bpages, links) {
1126                 if (op & BUS_DMASYNC_PREWRITE) {
1127                         bcopy((void *)bpage->datavaddr,
1128                             (void *)(bpage->vaddr_nocache != 0 ?
1129                                      bpage->vaddr_nocache : bpage->vaddr),
1130                             bpage->datacount);
1131                         if (bpage->vaddr_nocache == 0)
1132                                 cpu_dcache_wb_range(bpage->vaddr,
1133                                     bpage->datacount);
1134                 }
1135                 if (op & BUS_DMASYNC_POSTREAD) {
1136                         if (bpage->vaddr_nocache == 0)
1137                                 cpu_dcache_inv_range(bpage->vaddr,
1138                                     bpage->datacount);
1139                         bcopy((void *)(bpage->vaddr_nocache != 0 ?
1140                             bpage->vaddr_nocache : bpage->vaddr),
1141                             (void *)bpage->datavaddr, bpage->datacount);
1142                 }
1143         }
1144 }
1145
1146 static __inline int
1147 _bus_dma_buf_is_in_bp(bus_dmamap_t map, void *buf, int len)
1148 {
1149         struct bounce_page *bpage;
1150
1151         STAILQ_FOREACH(bpage, &map->bpages, links) {
1152                 if ((vm_offset_t)buf >= bpage->datavaddr &&
1153                     (vm_offset_t)buf + len < bpage->datavaddr +
1154                     bpage->datacount)
1155                         return (1);
1156         }
1157         return (0);
1158
1159 }
1160
1161 void
1162 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1163 {
1164         struct mbuf *m;
1165         struct uio *uio;
1166         int resid;
1167         struct iovec *iov;
1168
1169         if (op == BUS_DMASYNC_POSTWRITE)
1170                 return;
1171         if (STAILQ_FIRST(&map->bpages))
1172                 _bus_dmamap_sync_bp(dmat, map, op);
1173         if (map->flags & DMAMAP_COHERENT)
1174                 return;
1175         if ((op && BUS_DMASYNC_POSTREAD) && (map->len >= 2 * PAGE_SIZE)) {
1176                 cpu_dcache_wbinv_all();
1177                 return;
1178         }
1179         CTR3(KTR_BUSDMA, "%s: op %x flags %x", __func__, op, map->flags);
1180         switch(map->flags & DMAMAP_TYPE_MASK) {
1181         case DMAMAP_LINEAR:
1182                 if (!(_bus_dma_buf_is_in_bp(map, map->buffer, map->len)))
1183                         bus_dmamap_sync_buf(map->buffer, map->len, op);
1184                 break;
1185         case DMAMAP_MBUF:
1186                 m = map->buffer;
1187                 while (m) {
1188                         if (m->m_len > 0 &&
1189                             !(_bus_dma_buf_is_in_bp(map, m->m_data, m->m_len)))
1190                                 bus_dmamap_sync_buf(m->m_data, m->m_len, op);
1191                         m = m->m_next;
1192                 }
1193                 break;
1194         case DMAMAP_UIO:
1195                 uio = map->buffer;
1196                 iov = uio->uio_iov;
1197                 resid = uio->uio_resid;
1198                 for (int i = 0; i < uio->uio_iovcnt && resid != 0; i++) {
1199                         bus_size_t minlen = resid < iov[i].iov_len ? resid :
1200                             iov[i].iov_len;
1201                         if (minlen > 0) {
1202                                 if (!_bus_dma_buf_is_in_bp(map, iov[i].iov_base,
1203                                     minlen))
1204                                         bus_dmamap_sync_buf(iov[i].iov_base,
1205                                             minlen, op);
1206                                 resid -= minlen;
1207                         }
1208                 }
1209                 break;
1210         default:
1211                 break;
1212         }
1213         cpu_drain_writebuf();
1214 }
1215
1216 static void
1217 init_bounce_pages(void *dummy __unused)
1218 {
1219
1220         total_bpages = 0;
1221         STAILQ_INIT(&bounce_zone_list);
1222         STAILQ_INIT(&bounce_map_waitinglist);
1223         STAILQ_INIT(&bounce_map_callbacklist);
1224         mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
1225 }
1226 SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
1227
1228 static struct sysctl_ctx_list *
1229 busdma_sysctl_tree(struct bounce_zone *bz)
1230 {
1231         return (&bz->sysctl_tree);
1232 }
1233
1234 static struct sysctl_oid *
1235 busdma_sysctl_tree_top(struct bounce_zone *bz)
1236 {
1237         return (bz->sysctl_tree_top);
1238 }
1239
1240 static int
1241 alloc_bounce_zone(bus_dma_tag_t dmat)
1242 {
1243         struct bounce_zone *bz;
1244
1245         /* Check to see if we already have a suitable zone */
1246         STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1247                 if ((dmat->alignment <= bz->alignment)
1248                  && (dmat->boundary <= bz->boundary)
1249                  && (dmat->lowaddr >= bz->lowaddr)) {
1250                         dmat->bounce_zone = bz;
1251                         return (0);
1252                 }
1253         }
1254
1255         if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
1256             M_NOWAIT | M_ZERO)) == NULL)
1257                 return (ENOMEM);
1258
1259         STAILQ_INIT(&bz->bounce_page_list);
1260         bz->free_bpages = 0;
1261         bz->reserved_bpages = 0;
1262         bz->active_bpages = 0;
1263         bz->lowaddr = dmat->lowaddr;
1264         bz->alignment = dmat->alignment;
1265         bz->boundary = dmat->boundary;
1266         snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1267         busdma_zonecount++;
1268         snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1269         STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1270         dmat->bounce_zone = bz;
1271
1272         sysctl_ctx_init(&bz->sysctl_tree);
1273         bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
1274             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1275             CTLFLAG_RD, 0, "");
1276         if (bz->sysctl_tree_top == NULL) {
1277                 sysctl_ctx_free(&bz->sysctl_tree);
1278                 return (0);     /* XXX error code? */
1279         }
1280
1281         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1282             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1283             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1284             "Total bounce pages");
1285         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1286             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1287             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1288             "Free bounce pages");
1289         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1290             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1291             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1292             "Reserved bounce pages");
1293         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1294             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1295             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1296             "Active bounce pages");
1297         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1298             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1299             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1300             "Total bounce requests");
1301         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1302             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1303             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1304             "Total bounce requests that were deferred");
1305         SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
1306             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1307             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1308         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1309             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1310             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1311         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
1312             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
1313             "boundary", CTLFLAG_RD, &bz->boundary, 0, "");
1314
1315         return (0);
1316 }
1317
1318 static int
1319 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
1320 {
1321         struct bounce_zone *bz;
1322         int count;
1323
1324         bz = dmat->bounce_zone;
1325         count = 0;
1326         while (numpages > 0) {
1327                 struct bounce_page *bpage;
1328
1329                 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
1330                                                      M_NOWAIT | M_ZERO);
1331
1332                 if (bpage == NULL)
1333                         break;
1334                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1335                                                          M_NOWAIT, 0ul,
1336                                                          bz->lowaddr,
1337                                                          PAGE_SIZE,
1338                                                          bz->boundary);
1339                 if (bpage->vaddr == 0) {
1340                         free(bpage, M_DEVBUF);
1341                         break;
1342                 }
1343                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1344                 bpage->vaddr_nocache = (vm_offset_t)arm_remap_nocache(
1345                     (void *)bpage->vaddr, PAGE_SIZE);
1346                 mtx_lock(&bounce_lock);
1347                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1348                 total_bpages++;
1349                 bz->total_bpages++;
1350                 bz->free_bpages++;
1351                 mtx_unlock(&bounce_lock);
1352                 count++;
1353                 numpages--;
1354         }
1355         return (count);
1356 }
1357
1358 static int
1359 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1360 {
1361         struct bounce_zone *bz;
1362         int pages;
1363
1364         mtx_assert(&bounce_lock, MA_OWNED);
1365         bz = dmat->bounce_zone;
1366         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1367         if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
1368                 return (map->pagesneeded - (map->pagesreserved + pages));
1369         bz->free_bpages -= pages;
1370         bz->reserved_bpages += pages;
1371         map->pagesreserved += pages;
1372         pages = map->pagesneeded - map->pagesreserved;
1373
1374         return (pages);
1375 }
1376
1377 static bus_addr_t
1378 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1379                 bus_size_t size)
1380 {
1381         struct bounce_zone *bz;
1382         struct bounce_page *bpage;
1383
1384         KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
1385         KASSERT(map != NULL && map != &nobounce_dmamap,
1386             ("add_bounce_page: bad map %p", map));
1387
1388         bz = dmat->bounce_zone;
1389         if (map->pagesneeded == 0)
1390                 panic("add_bounce_page: map doesn't need any pages");
1391         map->pagesneeded--;
1392
1393         if (map->pagesreserved == 0)
1394                 panic("add_bounce_page: map doesn't need any pages");
1395         map->pagesreserved--;
1396
1397         mtx_lock(&bounce_lock);
1398         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1399         if (bpage == NULL)
1400                 panic("add_bounce_page: free page list is empty");
1401
1402         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1403         bz->reserved_bpages--;
1404         bz->active_bpages++;
1405         mtx_unlock(&bounce_lock);
1406
1407         bpage->datavaddr = vaddr;
1408         bpage->datacount = size;
1409         STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1410         return (bpage->busaddr);
1411 }
1412
1413 static void
1414 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1415 {
1416         struct bus_dmamap *map;
1417         struct bounce_zone *bz;
1418
1419         bz = dmat->bounce_zone;
1420         bpage->datavaddr = 0;
1421         bpage->datacount = 0;
1422
1423         mtx_lock(&bounce_lock);
1424         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1425         bz->free_bpages++;
1426         bz->active_bpages--;
1427         if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1428                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1429                         STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1430                         STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1431                                            map, links);
1432                         busdma_swi_pending = 1;
1433                         bz->total_deferred++;
1434                         swi_sched(vm_ih, 0);
1435                 }
1436         }
1437         mtx_unlock(&bounce_lock);
1438 }
1439
1440 void
1441 busdma_swi(void)
1442 {
1443         bus_dma_tag_t dmat;
1444         struct bus_dmamap *map;
1445
1446         mtx_lock(&bounce_lock);
1447         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1448                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1449                 mtx_unlock(&bounce_lock);
1450                 dmat = map->dmat;
1451                 (dmat->lockfunc)(dmat->lockfuncarg, BUS_DMA_LOCK);
1452                 bus_dmamap_load(map->dmat, map, map->buffer, map->len,
1453                     map->callback, map->callback_arg, /*flags*/0);
1454                 (dmat->lockfunc)(dmat->lockfuncarg, BUS_DMA_UNLOCK);
1455                 mtx_lock(&bounce_lock);
1456         }
1457         mtx_unlock(&bounce_lock);
1458 }