sys/platform/pc64/x86_64/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  * $DragonFly: src/sys/platform/pc32/i386/busdma_machdep.c,v 1.23 2008/06/05 18:06:32 swildner Exp $
  28  */
  29
  30 #include <sys/param.h>
  31 #include <sys/systm.h>
  32 #include <sys/malloc.h>
  33 #include <sys/mbuf.h>
  34 #include <sys/uio.h>
  35 #include <sys/thread2.h>
  36 #include <sys/bus_dma.h>
  37 #include <sys/kernel.h>
  38 #include <sys/sysctl.h>
  39 #include <sys/lock.h>
  40 #include <sys/spinlock2.h>
  41
  42 #include <vm/vm.h>
  43 #include <vm/vm_page.h>
  44
  45 /* XXX needed for to access pmap to convert per-proc virtual to physical */
  46 #include <sys/proc.h>
  47 #include <sys/lock.h>
  48 #include <vm/vm_map.h>
  49
  50 #include <machine/md_var.h>
  51
  52 #define MAX_BPAGES      1024
  53
  54 /*
  55  * 16 x N declared on stack.
  56  */
  57 #define BUS_DMA_CACHE_SEGMENTS  8
  58
  59 struct bounce_zone;
  60 struct bus_dmamap;
  61
  62 struct bus_dma_tag {
  63         bus_dma_tag_t   parent;
  64         bus_size_t      alignment;
  65         bus_size_t      boundary;
  66         bus_addr_t      lowaddr;
  67         bus_addr_t      highaddr;
  68         bus_dma_filter_t *filter;
  69         void            *filterarg;
  70         bus_size_t      maxsize;
  71         u_int           nsegments;
  72         bus_size_t      maxsegsz;
  73         int             flags;
  74         int             ref_count;
  75         int             map_count;
  76         bus_dma_segment_t *segments;
  77         struct bounce_zone *bounce_zone;
  78 #ifdef SMP
  79         struct spinlock spin;
  80 #else
  81         int             unused0;
  82 #endif
  83 };
  84
  85 /*
  86  * bus_dma_tag private flags
  87  */
  88 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
  89 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
  90 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  91
  92 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
  93
  94 #define BUS_DMAMEM_KMALLOC(dmat) \
  95         ((dmat)->maxsize <= PAGE_SIZE && \
  96          (dmat)->alignment <= PAGE_SIZE && \
  97          (dmat)->lowaddr >= ptoa(Maxmem))
  98
  99 struct bounce_page {
 100         vm_offset_t     vaddr;          /* kva of bounce buffer */
 101         bus_addr_t      busaddr;        /* Physical address */
 102         vm_offset_t     datavaddr;      /* kva of client data */
 103         bus_size_t      datacount;      /* client data count */
 104         STAILQ_ENTRY(bounce_page) links;
 105 };
 106
 107 struct bounce_zone {
 108         STAILQ_ENTRY(bounce_zone) links;
 109         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 110         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 111 #ifdef SMP
 112         struct spinlock spin;
 113 #else
 114         int             unused0;
 115 #endif
 116         int             total_bpages;
 117         int             free_bpages;
 118         int             reserved_bpages;
 119         int             active_bpages;
 120         int             total_bounced;
 121         int             total_deferred;
 122         int             reserve_failed;
 123         bus_size_t      alignment;
 124         bus_addr_t      lowaddr;
 125         char            zoneid[8];
 126         char            lowaddrid[20];
 127         struct sysctl_ctx_list sysctl_ctx;
 128         struct sysctl_oid *sysctl_tree;
 129 };
 130
 131 #ifdef SMP
 132 #define BZ_LOCK(bz)     spin_lock_wr(&(bz)->spin)
 133 #define BZ_UNLOCK(bz)   spin_unlock_wr(&(bz)->spin)
 134 #else
 135 #define BZ_LOCK(bz)     crit_enter()
 136 #define BZ_UNLOCK(bz)   crit_exit()
 137 #endif
 138
 139 static struct lwkt_token bounce_zone_tok =
 140         LWKT_TOKEN_INITIALIZER(bounce_zone_tok);
 141 static int busdma_zonecount;
 142 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
 143         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
 144
 145 int busdma_swi_pending;
 146 static int total_bounce_pages;
 147 static int max_bounce_pages = MAX_BPAGES;
 148 static int bounce_alignment = 1; /* XXX temporary */
 149
 150 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
 151 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
 152
 153 struct bus_dmamap {
 154         struct bp_list  bpages;
 155         int             pagesneeded;
 156         int             pagesreserved;
 157         bus_dma_tag_t   dmat;
 158         void            *buf;           /* unmapped buffer pointer */
 159         bus_size_t      buflen;         /* unmapped buffer length */
 160         bus_dmamap_callback_t *callback;
 161         void            *callback_arg;
 162         STAILQ_ENTRY(bus_dmamap) links;
 163 };
 164
 165 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
 166         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
 167
 168 static struct bus_dmamap nobounce_dmamap;
 169
 170 static int              alloc_bounce_zone(bus_dma_tag_t);
 171 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
 172 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
 173 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
 174 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
 175                             vm_offset_t, bus_size_t);
 176 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
 177
 178 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
 179 static void             add_map_callback(bus_dmamap_t);
 180
 181 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 182 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
 183            0, "Total bounce pages");
 184 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
 185            0, "Max bounce pages per bounce zone");
 186 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
 187            &bounce_alignment, 0, "Obey alignment constraint");
 188
 189 static __inline int
 190 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 191 {
 192         int retval;
 193
 194         retval = 0;
 195         do {
 196                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
 197                      (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0))
 198                  && (dmat->filter == NULL ||
 199                      dmat->filter(dmat->filterarg, paddr) != 0))
 200                         retval = 1;
 201
 202                 dmat = dmat->parent;
 203         } while (retval == 0 && dmat != NULL);
 204         return (retval);
 205 }
 206
 207 static __inline
 208 bus_dma_segment_t *
 209 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
 210 {
 211         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 212                 return(cache);
 213 #ifdef SMP
 214         spin_lock_wr(&tag->spin);
 215 #endif
 216         return(tag->segments);
 217 }
 218
 219 static __inline
 220 void
 221 bus_dma_tag_unlock(bus_dma_tag_t tag)
 222 {
 223 #ifdef SMP
 224         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
 225                 spin_unlock_wr(&tag->spin);
 226 #endif
 227 }
 228
 229 /*
 230  * Allocate a device specific dma_tag.
 231  */
 232 int
 233 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 234                    bus_size_t boundary, bus_addr_t lowaddr,
 235                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 236                    void *filterarg, bus_size_t maxsize, int nsegments,
 237                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
 238 {
 239         bus_dma_tag_t newtag;
 240         int error = 0;
 241
 242         /*
 243          * Sanity checks
 244          */
 245
 246         if (alignment == 0)
 247                 alignment = 1;
 248         if (alignment & (alignment - 1))
 249                 panic("alignment must be power of 2\n");
 250
 251         if (boundary != 0) {
 252                 if (boundary & (boundary - 1))
 253                         panic("boundary must be power of 2\n");
 254                 if (boundary < maxsegsz) {
 255                         kprintf("boundary < maxsegsz:\n");
 256                         print_backtrace();
 257                         maxsegsz = boundary;
 258                 }
 259         }
 260
 261         /* Return a NULL tag on failure */
 262         *dmat = NULL;
 263
 264         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
 265
 266 #ifdef SMP
 267         spin_init(&newtag->spin);
 268 #endif
 269         newtag->parent = parent;
 270         newtag->alignment = alignment;
 271         newtag->boundary = boundary;
 272         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
 273         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
 274         newtag->filter = filter;
 275         newtag->filterarg = filterarg;
 276         newtag->maxsize = maxsize;
 277         newtag->nsegments = nsegments;
 278         newtag->maxsegsz = maxsegsz;
 279         newtag->flags = flags;
 280         newtag->ref_count = 1; /* Count ourself */
 281         newtag->map_count = 0;
 282         newtag->segments = NULL;
 283         newtag->bounce_zone = NULL;
 284
 285         /* Take into account any restrictions imposed by our parent tag */
 286         if (parent != NULL) {
 287                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
 288                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 289
 290                 if (newtag->boundary == 0) {
 291                         newtag->boundary = parent->boundary;
 292                 } else if (parent->boundary != 0) {
 293                         newtag->boundary = MIN(parent->boundary,
 294                                                newtag->boundary);
 295                 }
 296
 297 #ifdef notyet
 298                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
 299 #endif
 300
 301                 if (newtag->filter == NULL) {
 302                         /*
 303                          * Short circuit looking at our parent directly
 304                          * since we have encapsulated all of its information
 305                          */
 306                         newtag->filter = parent->filter;
 307                         newtag->filterarg = parent->filterarg;
 308                         newtag->parent = parent->parent;
 309                 }
 310                 if (newtag->parent != NULL)
 311                         parent->ref_count++;
 312         }
 313
 314         if (newtag->lowaddr < ptoa(Maxmem))
 315                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
 316         if (bounce_alignment && newtag->alignment > 1 &&
 317             !(newtag->flags & BUS_DMA_ALIGNED))
 318                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
 319
 320         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
 321             (flags & BUS_DMA_ALLOCNOW) != 0) {
 322                 struct bounce_zone *bz;
 323
 324                 /* Must bounce */
 325
 326                 error = alloc_bounce_zone(newtag);
 327                 if (error)
 328                         goto back;
 329                 bz = newtag->bounce_zone;
 330
 331                 if (ptoa(bz->total_bpages) < maxsize) {
 332                         int pages;
 333
 334                         if (flags & BUS_DMA_ONEBPAGE) {
 335                                 pages = 1;
 336                         } else {
 337                                 pages = atop(round_page(maxsize)) -
 338                                         bz->total_bpages;
 339                                 pages = MAX(pages, 1);
 340                         }
 341
 342                         /* Add pages to our bounce pool */
 343                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
 344                                 error = ENOMEM;
 345
 346                         /* Performed initial allocation */
 347                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 348                 }
 349         }
 350 back:
 351         if (error)
 352                 kfree(newtag, M_DEVBUF);
 353         else
 354                 *dmat = newtag;
 355         return error;
 356 }
 357
 358 int
 359 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 360 {
 361         if (dmat != NULL) {
 362                 if (dmat->map_count != 0)
 363                         return (EBUSY);
 364
 365                 while (dmat != NULL) {
 366                         bus_dma_tag_t parent;
 367
 368                         parent = dmat->parent;
 369                         dmat->ref_count--;
 370                         if (dmat->ref_count == 0) {
 371                                 if (dmat->segments != NULL)
 372                                         kfree(dmat->segments, M_DEVBUF);
 373                                 kfree(dmat, M_DEVBUF);
 374                                 /*
 375                                  * Last reference count, so
 376                                  * release our reference
 377                                  * count on our parent.
 378                                  */
 379                                 dmat = parent;
 380                         } else
 381                                 dmat = NULL;
 382                 }
 383         }
 384         return (0);
 385 }
 386
 387 bus_size_t
 388 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
 389 {
 390         return(tag->maxsize);
 391 }
 392
 393 /*
 394  * Allocate a handle for mapping from kva/uva/physical
 395  * address space into bus device space.
 396  */
 397 int
 398 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 399 {
 400         int error;
 401
 402         error = 0;
 403
 404         if (dmat->segments == NULL) {
 405                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
 406                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 407                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 408         }
 409
 410         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 411                 struct bounce_zone *bz;
 412                 int maxpages;
 413
 414                 /* Must bounce */
 415
 416                 if (dmat->bounce_zone == NULL) {
 417                         error = alloc_bounce_zone(dmat);
 418                         if (error)
 419                                 return error;
 420                 }
 421                 bz = dmat->bounce_zone;
 422
 423                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
 424
 425                 /* Initialize the new map */
 426                 STAILQ_INIT(&((*mapp)->bpages));
 427
 428                 /*
 429                  * Attempt to add pages to our pool on a per-instance
 430                  * basis up to a sane limit.
 431                  */
 432                 if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
 433                         maxpages = max_bounce_pages;
 434                 } else {
 435                         maxpages = MIN(max_bounce_pages,
 436                                        Maxmem - atop(dmat->lowaddr));
 437                 }
 438                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0
 439                  || (dmat->map_count > 0
 440                   && bz->total_bpages < maxpages)) {
 441                         int pages;
 442
 443                         if (flags & BUS_DMA_ONEBPAGE) {
 444                                 pages = 1;
 445                         } else {
 446                                 pages = atop(round_page(dmat->maxsize));
 447                                 pages = MIN(maxpages - bz->total_bpages, pages);
 448                                 pages = MAX(pages, 1);
 449                         }
 450                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
 451                                 error = ENOMEM;
 452
 453                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 454                                 if (!error)
 455                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 456                         } else {
 457                                 error = 0;
 458                         }
 459                 }
 460         } else {
 461                 *mapp = NULL;
 462         }
 463         if (!error)
 464                 dmat->map_count++;
 465         return error;
 466 }
 467
 468 /*
 469  * Destroy a handle for mapping from kva/uva/physical
 470  * address space into bus device space.
 471  */
 472 int
 473 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 474 {
 475         if (map != NULL) {
 476                 if (STAILQ_FIRST(&map->bpages) != NULL)
 477                         return (EBUSY);
 478                 kfree(map, M_DEVBUF);
 479         }
 480         dmat->map_count--;
 481         return (0);
 482 }
 483
 484 static __inline bus_size_t
 485 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
 486 {
 487         bus_size_t maxsize = 0;
 488         uintptr_t vaddr = (uintptr_t)vaddr0;
 489
 490         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
 491                 if (verify || bootverbose)
 492                         kprintf("boundary check failed\n");
 493                 if (verify)
 494                         print_backtrace(); /* XXX panic */
 495                 maxsize = dmat->maxsize;
 496         }
 497         if (vaddr & (dmat->alignment - 1)) {
 498                 if (verify || bootverbose)
 499                         kprintf("alignment check failed\n");
 500                 if (verify)
 501                         print_backtrace(); /* XXX panic */
 502                 if (dmat->maxsize < dmat->alignment)
 503                         maxsize = dmat->alignment;
 504                 else
 505                         maxsize = dmat->maxsize;
 506         }
 507         return maxsize;
 508 }
 509
 510 /*
 511  * Allocate a piece of memory that can be efficiently mapped into
 512  * bus device space based on the constraints lited in the dma tag.
 513  *
 514  * mapp is degenerate.  By definition this allocation should not require
 515  * bounce buffers so do not allocate a dma map.
 516  */
 517 int
 518 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
 519                  bus_dmamap_t *mapp)
 520 {
 521         int mflags;
 522
 523         /* If we succeed, no mapping/bouncing will be required */
 524         *mapp = NULL;
 525
 526         if (dmat->segments == NULL) {
 527                 KKASSERT(dmat->nsegments < 16384);
 528                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 529                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 530         }
 531
 532         if (flags & BUS_DMA_NOWAIT)
 533                 mflags = M_NOWAIT;
 534         else
 535                 mflags = M_WAITOK;
 536         if (flags & BUS_DMA_ZERO)
 537                 mflags |= M_ZERO;
 538
 539         if (BUS_DMAMEM_KMALLOC(dmat)) {
 540                 bus_size_t maxsize;
 541
 542                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 543
 544                 /*
 545                  * XXX
 546                  * Check whether the allocation
 547                  * - crossed a page boundary
 548                  * - was not aligned
 549                  * Retry with power-of-2 alignment in the above cases.
 550                  */
 551                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 552                 if (maxsize) {
 553                         size_t size;
 554
 555                         kfree(*vaddr, M_DEVBUF);
 556                         /* XXX check for overflow? */
 557                         for (size = 1; size <= maxsize; size <<= 1)
 558                                 ;
 559                         *vaddr = kmalloc(size, M_DEVBUF, mflags);
 560                         check_kmalloc(dmat, *vaddr, 1);
 561                 }
 562         } else {
 563                 /*
 564                  * XXX Use Contigmalloc until it is merged into this facility
 565                  *     and handles multi-seg allocations.  Nobody is doing
 566                  *     multi-seg allocations yet though.
 567                  */
 568                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 569                     0ul, dmat->lowaddr, dmat->alignment, dmat->boundary);
 570         }
 571         if (*vaddr == NULL)
 572                 return (ENOMEM);
 573         return (0);
 574 }
 575
 576 /*
 577  * Free a piece of memory and it's allociated dmamap, that was allocated
 578  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 579  */
 580 void
 581 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 582 {
 583         /*
 584          * dmamem does not need to be bounced, so the map should be
 585          * NULL
 586          */
 587         if (map != NULL)
 588                 panic("bus_dmamem_free: Invalid map freed\n");
 589         if (BUS_DMAMEM_KMALLOC(dmat))
 590                 kfree(vaddr, M_DEVBUF);
 591         else
 592                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 593 }
 594
 595 static __inline vm_paddr_t
 596 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
 597 {
 598         if (pmap)
 599                 return pmap_extract(pmap, vaddr);
 600         else
 601                 return pmap_kextract(vaddr);
 602 }
 603
 604 /*
 605  * Utility function to load a linear buffer.  lastaddrp holds state
 606  * between invocations (for multiple-buffer loads).  segp contains
 607  * the segment following the starting one on entrace, and the ending
 608  * segment on exit.  first indicates if this is the first invocation
 609  * of this function.
 610  */
 611 static int
 612 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
 613                         bus_dmamap_t map,
 614                         void *buf, bus_size_t buflen,
 615                         bus_dma_segment_t *segments,
 616                         int nsegments,
 617                         pmap_t pmap,
 618                         int flags,
 619                         vm_paddr_t *lastpaddrp,
 620                         int *segp,
 621                         int first)
 622 {
 623         vm_offset_t vaddr;
 624         vm_paddr_t paddr, nextpaddr;
 625         bus_dma_segment_t *sg;
 626         bus_addr_t bmask;
 627         int seg, error = 0;
 628
 629         if (map == NULL)
 630                 map = &nobounce_dmamap;
 631
 632 #ifdef INVARIANTS
 633         if (dmat->flags & BUS_DMA_ALIGNED)
 634                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
 635 #endif
 636
 637         /*
 638          * If we are being called during a callback, pagesneeded will
 639          * be non-zero, so we can avoid doing the work twice.
 640          */
 641         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 642             map != &nobounce_dmamap && map->pagesneeded == 0) {
 643                 vm_offset_t vendaddr;
 644
 645                 /*
 646                  * Count the number of bounce pages
 647                  * needed in order to complete this transfer
 648                  */
 649                 vaddr = (vm_offset_t)buf;
 650                 vendaddr = (vm_offset_t)buf + buflen;
 651
 652                 while (vaddr < vendaddr) {
 653                         paddr = _bus_dma_extract(pmap, vaddr);
 654                         if (run_filter(dmat, paddr) != 0)
 655                                 map->pagesneeded++;
 656                         vaddr += (PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK));
 657                 }
 658         }
 659
 660         /* Reserve Necessary Bounce Pages */
 661         if (map->pagesneeded != 0) {
 662                 struct bounce_zone *bz;
 663
 664                 bz = dmat->bounce_zone;
 665                 BZ_LOCK(bz);
 666                 if (flags & BUS_DMA_NOWAIT) {
 667                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 668                                 BZ_UNLOCK(bz);
 669                                 error = ENOMEM;
 670                                 goto free_bounce;
 671                         }
 672                 } else {
 673                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 674                                 /* Queue us for resources */
 675                                 map->dmat = dmat;
 676                                 map->buf = buf;
 677                                 map->buflen = buflen;
 678
 679                                 STAILQ_INSERT_TAIL(
 680                                     &dmat->bounce_zone->bounce_map_waitinglist,
 681                                     map, links);
 682                                 BZ_UNLOCK(bz);
 683
 684                                 return (EINPROGRESS);
 685                         }
 686                 }
 687                 BZ_UNLOCK(bz);
 688         }
 689
 690         KKASSERT(*segp >= 1 && *segp <= nsegments);
 691         seg = *segp;
 692         sg = &segments[seg - 1];
 693
 694         vaddr = (vm_offset_t)buf;
 695         nextpaddr = *lastpaddrp;
 696         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
 697
 698         /* force at least one segment */
 699         do {
 700                 bus_size_t size;
 701
 702                 /*
 703                  * Per-page main loop
 704                  */
 705                 paddr = _bus_dma_extract(pmap, vaddr);
 706                 size = PAGE_SIZE - (paddr & PAGE_MASK);
 707                 if (size > buflen)
 708                         size = buflen;
 709                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
 710                         /*
 711                          * note: this paddr has the same in-page offset
 712                          * as vaddr and thus the paddr above, so the
 713                          * size does not have to be recalculated
 714                          */
 715                         paddr = add_bounce_page(dmat, map, vaddr, size);
 716                 }
 717
 718                 /*
 719                  * Fill in the bus_dma_segment
 720                  */
 721                 if (first) {
 722                         sg->ds_addr = paddr;
 723                         sg->ds_len = size;
 724                         first = 0;
 725                 } else if (paddr == nextpaddr) {
 726                         sg->ds_len += size;
 727                 } else {
 728                         sg++;
 729                         seg++;
 730                         if (seg > nsegments)
 731                                 break;
 732                         sg->ds_addr = paddr;
 733                         sg->ds_len = size;
 734                 }
 735                 nextpaddr = paddr + size;
 736
 737                 /*
 738                  * Handle maxsegsz and boundary issues with a nested loop
 739                  */
 740                 for (;;) {
 741                         bus_size_t tmpsize;
 742
 743                         /*
 744                          * Limit to the boundary and maximum segment size
 745                          */
 746                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
 747                                 tmpsize = dmat->boundary -
 748                                           (sg->ds_addr & ~bmask);
 749                                 if (tmpsize > dmat->maxsegsz)
 750                                         tmpsize = dmat->maxsegsz;
 751                                 KKASSERT(tmpsize < sg->ds_len);
 752                         } else if (sg->ds_len > dmat->maxsegsz) {
 753                                 tmpsize = dmat->maxsegsz;
 754                         } else {
 755                                 break;
 756                         }
 757
 758                         /*
 759                          * Futz, split the data into a new segment.
 760                          */
 761                         if (seg >= nsegments)
 762                                 goto fail;
 763                         sg[1].ds_len = sg[0].ds_len - tmpsize;
 764                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
 765                         sg[0].ds_len = tmpsize;
 766                         sg++;
 767                         seg++;
 768                 }
 769
 770                 /*
 771                  * Adjust for loop
 772                  */
 773                 buflen -= size;
 774                 vaddr += size;
 775         } while (buflen > 0);
 776 fail:
 777         if (buflen != 0)
 778                 error = EFBIG;
 779
 780         *segp = seg;
 781         *lastpaddrp = nextpaddr;
 782
 783 free_bounce:
 784         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 785             map != &nobounce_dmamap) {
 786                 _bus_dmamap_unload(dmat, map);
 787                 return_bounce_pages(dmat, map);
 788         }
 789         return error;
 790 }
 791
 792 /*
 793  * Map the buffer buf into bus space using the dmamap map.
 794  */
 795 int
 796 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 797                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 798                 void *callback_arg, int flags)
 799 {
 800         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 801         bus_dma_segment_t *segments;
 802         vm_paddr_t lastaddr = 0;
 803         int error, nsegs = 1;
 804
 805         if (map != NULL) {
 806                 /*
 807                  * XXX
 808                  * Follow old semantics.  Once all of the callers are fixed,
 809                  * we should get rid of these internal flag "adjustment".
 810                  */
 811                 flags &= ~BUS_DMA_NOWAIT;
 812                 flags |= BUS_DMA_WAITOK;
 813
 814                 map->callback = callback;
 815                 map->callback_arg = callback_arg;
 816         }
 817
 818         segments = bus_dma_tag_lock(dmat, cache_segments);
 819         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
 820                         segments, dmat->nsegments,
 821                         NULL, flags, &lastaddr, &nsegs, 1);
 822         if (error == EINPROGRESS) {
 823                 bus_dma_tag_unlock(dmat);
 824                 return error;
 825         }
 826         callback(callback_arg, segments, nsegs, error);
 827         bus_dma_tag_unlock(dmat);
 828         return 0;
 829 }
 830
 831 /*
 832  * Like _bus_dmamap_load(), but for mbufs.
 833  */
 834 int
 835 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
 836                      struct mbuf *m0,
 837                      bus_dmamap_callback2_t *callback, void *callback_arg,
 838                      int flags)
 839 {
 840         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 841         bus_dma_segment_t *segments;
 842         int nsegs, error;
 843
 844         /*
 845          * XXX
 846          * Follow old semantics.  Once all of the callers are fixed,
 847          * we should get rid of these internal flag "adjustment".
 848          */
 849         flags &= ~BUS_DMA_WAITOK;
 850         flags |= BUS_DMA_NOWAIT;
 851
 852         segments = bus_dma_tag_lock(dmat, cache_segments);
 853         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
 854                         segments, dmat->nsegments, &nsegs, flags);
 855         if (error) {
 856                 /* force "no valid mappings" in callback */
 857                 callback(callback_arg, segments, 0,
 858                          0, error);
 859         } else {
 860                 callback(callback_arg, segments, nsegs,
 861                          m0->m_pkthdr.len, error);
 862         }
 863         bus_dma_tag_unlock(dmat);
 864         return error;
 865 }
 866
 867 int
 868 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
 869                              struct mbuf *m0,
 870                              bus_dma_segment_t *segs, int maxsegs,
 871                              int *nsegs, int flags)
 872 {
 873         int error;
 874
 875         M_ASSERTPKTHDR(m0);
 876
 877         KASSERT(maxsegs >= 1, ("invalid maxsegs %d\n", maxsegs));
 878         KASSERT(maxsegs <= dmat->nsegments,
 879                 ("%d too many segments, dmat only support %d segments\n",
 880                  maxsegs, dmat->nsegments));
 881         KASSERT(flags & BUS_DMA_NOWAIT,
 882                 ("only BUS_DMA_NOWAIT is supported\n"));
 883
 884         if (m0->m_pkthdr.len <= dmat->maxsize) {
 885                 int first = 1;
 886                 vm_paddr_t lastaddr = 0;
 887                 struct mbuf *m;
 888
 889                 *nsegs = 1;
 890                 error = 0;
 891                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 892                         if (m->m_len == 0)
 893                                 continue;
 894
 895                         error = _bus_dmamap_load_buffer(dmat, map,
 896                                         m->m_data, m->m_len,
 897                                         segs, maxsegs,
 898                                         NULL, flags, &lastaddr,
 899                                         nsegs, first);
 900                         if (error == ENOMEM && !first) {
 901                                 /*
 902                                  * Out of bounce pages due to too many
 903                                  * fragments in the mbuf chain; return
 904                                  * EFBIG instead.
 905                                  */
 906                                 error = EFBIG;
 907                         }
 908                         first = 0;
 909                 }
 910 #ifdef INVARIANTS
 911                 if (!error)
 912                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
 913 #endif
 914         } else {
 915                 *nsegs = 0;
 916                 error = EINVAL;
 917         }
 918         KKASSERT(error != EINPROGRESS);
 919         return error;
 920 }
 921
 922 /*
 923  * Like _bus_dmamap_load(), but for uios.
 924  */
 925 int
 926 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
 927                     struct uio *uio,
 928                     bus_dmamap_callback2_t *callback, void *callback_arg,
 929                     int flags)
 930 {
 931         vm_paddr_t lastaddr;
 932         int nsegs, error, first, i;
 933         bus_size_t resid;
 934         struct iovec *iov;
 935         pmap_t pmap;
 936         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 937         bus_dma_segment_t *segments;
 938         bus_dma_segment_t *segs;
 939         int nsegs_left;
 940
 941         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 942                 segments = cache_segments;
 943         else
 944                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
 945                                    M_DEVBUF, M_WAITOK | M_ZERO);
 946
 947         /*
 948          * XXX
 949          * Follow old semantics.  Once all of the callers are fixed,
 950          * we should get rid of these internal flag "adjustment".
 951          */
 952         flags &= ~BUS_DMA_WAITOK;
 953         flags |= BUS_DMA_NOWAIT;
 954
 955         resid = (bus_size_t)uio->uio_resid;
 956         iov = uio->uio_iov;
 957
 958         segs = segments;
 959         nsegs_left = dmat->nsegments;
 960
 961         if (uio->uio_segflg == UIO_USERSPACE) {
 962                 struct thread *td;
 963
 964                 td = uio->uio_td;
 965                 KASSERT(td != NULL && td->td_proc != NULL,
 966                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
 967                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
 968         } else {
 969                 pmap = NULL;
 970         }
 971
 972         error = 0;
 973         nsegs = 1;
 974         first = 1;
 975         lastaddr = 0;
 976         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
 977                 /*
 978                  * Now at the first iovec to load.  Load each iovec
 979                  * until we have exhausted the residual count.
 980                  */
 981                 bus_size_t minlen =
 982                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
 983                 caddr_t addr = (caddr_t) iov[i].iov_base;
 984
 985                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
 986                                 segs, nsegs_left,
 987                                 pmap, flags, &lastaddr, &nsegs, first);
 988                 first = 0;
 989
 990                 resid -= minlen;
 991                 if (error == 0) {
 992                         nsegs_left -= nsegs;
 993                         segs += nsegs;
 994                 }
 995         }
 996
 997         /*
 998          * Minimum one DMA segment, even if 0-length buffer.
 999          */
1000         if (nsegs_left == dmat->nsegments)
1001                 --nsegs_left;
1002
1003         if (error) {
1004                 /* force "no valid mappings" in callback */
1005                 callback(callback_arg, segments, 0,
1006                          0, error);
1007         } else {
1008                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1009                          (bus_size_t)uio->uio_resid, error);
1010         }
1011         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1012                 kfree(segments, M_DEVBUF);
1013         return error;
1014 }
1015
1016 /*
1017  * Release the mapping held by map.
1018  */
1019 void
1020 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1021 {
1022         struct bounce_page *bpage;
1023
1024         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1025                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1026                 free_bounce_page(dmat, bpage);
1027         }
1028 }
1029
1030 void
1031 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1032 {
1033         struct bounce_page *bpage;
1034
1035         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1036                 /*
1037                  * Handle data bouncing.  We might also
1038                  * want to add support for invalidating
1039                  * the caches on broken hardware
1040                  */
1041                 switch (op) {
1042                 case BUS_DMASYNC_PREWRITE:
1043                         while (bpage != NULL) {
1044                                 bcopy((void *)bpage->datavaddr,
1045                                       (void *)bpage->vaddr,
1046                                       bpage->datacount);
1047                                 bpage = STAILQ_NEXT(bpage, links);
1048                         }
1049                         dmat->bounce_zone->total_bounced++;
1050                         break;
1051
1052                 case BUS_DMASYNC_POSTREAD:
1053                         while (bpage != NULL) {
1054                                 bcopy((void *)bpage->vaddr,
1055                                       (void *)bpage->datavaddr,
1056                                       bpage->datacount);
1057                                 bpage = STAILQ_NEXT(bpage, links);
1058                         }
1059                         dmat->bounce_zone->total_bounced++;
1060                         break;
1061
1062                 case BUS_DMASYNC_PREREAD:
1063                 case BUS_DMASYNC_POSTWRITE:
1064                         /* No-ops */
1065                         break;
1066                 }
1067         }
1068 }
1069
1070 static int
1071 alloc_bounce_zone(bus_dma_tag_t dmat)
1072 {
1073         struct bounce_zone *bz, *new_bz;
1074         lwkt_tokref ref;
1075
1076         KASSERT(dmat->bounce_zone == NULL,
1077                 ("bounce zone was already assigned\n"));
1078
1079         new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO);
1080
1081         lwkt_gettoken(&ref, &bounce_zone_tok);
1082
1083         /* Check to see if we already have a suitable zone */
1084         STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1085                 if (dmat->alignment <= bz->alignment &&
1086                     dmat->lowaddr >= bz->lowaddr) {
1087                         lwkt_reltoken(&ref);
1088
1089                         dmat->bounce_zone = bz;
1090                         kfree(new_bz, M_DEVBUF);
1091                         return 0;
1092                 }
1093         }
1094         bz = new_bz;
1095
1096 #ifdef SMP
1097         spin_init(&bz->spin);
1098 #endif
1099         STAILQ_INIT(&bz->bounce_page_list);
1100         STAILQ_INIT(&bz->bounce_map_waitinglist);
1101         bz->free_bpages = 0;
1102         bz->reserved_bpages = 0;
1103         bz->active_bpages = 0;
1104         bz->lowaddr = dmat->lowaddr;
1105         bz->alignment = round_page(dmat->alignment);
1106         ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1107         busdma_zonecount++;
1108         ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1109         STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1110
1111         lwkt_reltoken(&ref);
1112
1113         dmat->bounce_zone = bz;
1114
1115         sysctl_ctx_init(&bz->sysctl_ctx);
1116         bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx,
1117             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1118             CTLFLAG_RD, 0, "");
1119         if (bz->sysctl_tree == NULL) {
1120                 sysctl_ctx_free(&bz->sysctl_ctx);
1121                 return 0;       /* XXX error code? */
1122         }
1123
1124         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1125             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1126             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1127             "Total bounce pages");
1128         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1129             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1130             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1131             "Free bounce pages");
1132         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1133             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1134             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1135             "Reserved bounce pages");
1136         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1137             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1138             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1139             "Active bounce pages");
1140         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1141             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1142             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1143             "Total bounce requests");
1144         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1145             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1146             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1147             "Total bounce requests that were deferred");
1148         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1149             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1150             "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0,
1151             "Total bounce page reservations that were failed");
1152         SYSCTL_ADD_STRING(&bz->sysctl_ctx,
1153             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1154             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1155         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1156             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1157             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1158
1159         return 0;
1160 }
1161
1162 static int
1163 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags)
1164 {
1165         struct bounce_zone *bz = dmat->bounce_zone;
1166         int count = 0, mflags;
1167
1168         if (flags & BUS_DMA_NOWAIT)
1169                 mflags = M_NOWAIT;
1170         else
1171                 mflags = M_WAITOK;
1172
1173         while (numpages > 0) {
1174                 struct bounce_page *bpage;
1175
1176                 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO);
1177
1178                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1179                                                          mflags, 0ul,
1180                                                          bz->lowaddr,
1181                                                          bz->alignment, 0);
1182                 if (bpage->vaddr == 0) {
1183                         kfree(bpage, M_DEVBUF);
1184                         break;
1185                 }
1186                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1187
1188                 BZ_LOCK(bz);
1189                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1190                 total_bounce_pages++;
1191                 bz->total_bpages++;
1192                 bz->free_bpages++;
1193                 BZ_UNLOCK(bz);
1194
1195                 count++;
1196                 numpages--;
1197         }
1198         return count;
1199 }
1200
1201 /* Assume caller holds bounce zone spinlock */
1202 static int
1203 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1204 {
1205         struct bounce_zone *bz = dmat->bounce_zone;
1206         int pages;
1207
1208         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1209         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1210                 bz->reserve_failed++;
1211                 return (map->pagesneeded - (map->pagesreserved + pages));
1212         }
1213
1214         bz->free_bpages -= pages;
1215
1216         bz->reserved_bpages += pages;
1217         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1218
1219         map->pagesreserved += pages;
1220         pages = map->pagesneeded - map->pagesreserved;
1221
1222         return pages;
1223 }
1224
1225 static void
1226 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1227 {
1228         struct bounce_zone *bz = dmat->bounce_zone;
1229         int reserved = map->pagesreserved;
1230         bus_dmamap_t wait_map;
1231
1232         map->pagesreserved = 0;
1233         map->pagesneeded = 0;
1234
1235         if (reserved == 0)
1236                 return;
1237
1238         BZ_LOCK(bz);
1239
1240         bz->free_bpages += reserved;
1241         KKASSERT(bz->free_bpages <= bz->total_bpages);
1242
1243         KKASSERT(bz->reserved_bpages >= reserved);
1244         bz->reserved_bpages -= reserved;
1245
1246         wait_map = get_map_waiting(dmat);
1247
1248         BZ_UNLOCK(bz);
1249
1250         if (wait_map != NULL)
1251                 add_map_callback(map);
1252 }
1253
1254 static bus_addr_t
1255 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1256                 bus_size_t size)
1257 {
1258         struct bounce_zone *bz = dmat->bounce_zone;
1259         struct bounce_page *bpage;
1260
1261         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1262         map->pagesneeded--;
1263
1264         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1265         map->pagesreserved--;
1266
1267         BZ_LOCK(bz);
1268
1269         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1270         KASSERT(bpage != NULL, ("free page list is empty"));
1271         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1272
1273         KKASSERT(bz->reserved_bpages > 0);
1274         bz->reserved_bpages--;
1275
1276         bz->active_bpages++;
1277         KKASSERT(bz->active_bpages <= bz->total_bpages);
1278
1279         BZ_UNLOCK(bz);
1280
1281         bpage->datavaddr = vaddr;
1282         bpage->datacount = size;
1283         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1284         return bpage->busaddr;
1285 }
1286
1287 static void
1288 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1289 {
1290         struct bounce_zone *bz = dmat->bounce_zone;
1291         bus_dmamap_t map;
1292
1293         bpage->datavaddr = 0;
1294         bpage->datacount = 0;
1295
1296         BZ_LOCK(bz);
1297
1298         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1299
1300         bz->free_bpages++;
1301         KKASSERT(bz->free_bpages <= bz->total_bpages);
1302
1303         KKASSERT(bz->active_bpages > 0);
1304         bz->active_bpages--;
1305
1306         map = get_map_waiting(dmat);
1307
1308         BZ_UNLOCK(bz);
1309
1310         if (map != NULL)
1311                 add_map_callback(map);
1312 }
1313
1314 /* Assume caller holds bounce zone spinlock */
1315 static bus_dmamap_t
1316 get_map_waiting(bus_dma_tag_t dmat)
1317 {
1318         struct bounce_zone *bz = dmat->bounce_zone;
1319         bus_dmamap_t map;
1320
1321         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1322         if (map != NULL) {
1323                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1324                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1325                         bz->total_deferred++;
1326                 } else {
1327                         map = NULL;
1328                 }
1329         }
1330         return map;
1331 }
1332
1333 static void
1334 add_map_callback(bus_dmamap_t map)
1335 {
1336         /* XXX callbacklist is not MPSAFE */
1337         crit_enter();
1338         get_mplock();
1339         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1340         busdma_swi_pending = 1;
1341         setsoftvm();
1342         rel_mplock();
1343         crit_exit();
1344 }
1345
1346 void
1347 busdma_swi(void)
1348 {
1349         bus_dmamap_t map;
1350
1351         crit_enter();
1352         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1353                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1354                 crit_exit();
1355                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1356                                 map->callback, map->callback_arg, /*flags*/0);
1357                 crit_enter();
1358         }
1359         crit_exit();
1360 }