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/bus_dma.h>
  36 #include <sys/kernel.h>
  37 #include <sys/sysctl.h>
  38 #include <sys/lock.h>
  39
  40 #include <sys/thread2.h>
  41 #include <sys/spinlock2.h>
  42 #include <sys/mplock2.h>
  43
  44 #include <vm/vm.h>
  45 #include <vm/vm_page.h>
  46
  47 /* XXX needed for to access pmap to convert per-proc virtual to physical */
  48 #include <sys/proc.h>
  49 #include <sys/lock.h>
  50 #include <vm/vm_map.h>
  51
  52 #include <machine/md_var.h>
  53
  54 #define MAX_BPAGES      1024
  55
  56 /*
  57  * 16 x N declared on stack.
  58  */
  59 #define BUS_DMA_CACHE_SEGMENTS  8
  60
  61 struct bounce_zone;
  62 struct bus_dmamap;
  63
  64 struct bus_dma_tag {
  65         bus_dma_tag_t   parent;
  66         bus_size_t      alignment;
  67         bus_size_t      boundary;
  68         bus_addr_t      lowaddr;
  69         bus_addr_t      highaddr;
  70         bus_dma_filter_t *filter;
  71         void            *filterarg;
  72         bus_size_t      maxsize;
  73         u_int           nsegments;
  74         bus_size_t      maxsegsz;
  75         int             flags;
  76         int             ref_count;
  77         int             map_count;
  78         bus_dma_segment_t *segments;
  79         struct bounce_zone *bounce_zone;
  80 #ifdef SMP
  81         struct spinlock spin;
  82 #else
  83         int             unused0;
  84 #endif
  85 };
  86
  87 /*
  88  * bus_dma_tag private flags
  89  */
  90 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
  91 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
  92 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  93
  94 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
  95
  96 #define BUS_DMAMEM_KMALLOC(dmat) \
  97         ((dmat)->maxsize <= PAGE_SIZE && \
  98          (dmat)->alignment <= PAGE_SIZE && \
  99          (dmat)->lowaddr >= ptoa(Maxmem))
 100
 101 struct bounce_page {
 102         vm_offset_t     vaddr;          /* kva of bounce buffer */
 103         bus_addr_t      busaddr;        /* Physical address */
 104         vm_offset_t     datavaddr;      /* kva of client data */
 105         bus_size_t      datacount;      /* client data count */
 106         STAILQ_ENTRY(bounce_page) links;
 107 };
 108
 109 struct bounce_zone {
 110         STAILQ_ENTRY(bounce_zone) links;
 111         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 112         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 113 #ifdef SMP
 114         struct spinlock spin;
 115 #else
 116         int             unused0;
 117 #endif
 118         int             total_bpages;
 119         int             free_bpages;
 120         int             reserved_bpages;
 121         int             active_bpages;
 122         int             total_bounced;
 123         int             total_deferred;
 124         int             reserve_failed;
 125         bus_size_t      alignment;
 126         bus_addr_t      lowaddr;
 127         char            zoneid[8];
 128         char            lowaddrid[20];
 129         struct sysctl_ctx_list sysctl_ctx;
 130         struct sysctl_oid *sysctl_tree;
 131 };
 132
 133 #ifdef SMP
 134 #define BZ_LOCK(bz)     spin_lock_wr(&(bz)->spin)
 135 #define BZ_UNLOCK(bz)   spin_unlock_wr(&(bz)->spin)
 136 #else
 137 #define BZ_LOCK(bz)     crit_enter()
 138 #define BZ_UNLOCK(bz)   crit_exit()
 139 #endif
 140
 141 static struct lwkt_token bounce_zone_tok = LWKT_TOKEN_MP_INITIALIZER;
 142 static int busdma_zonecount;
 143 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
 144         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
 145
 146 int busdma_swi_pending;
 147 static int total_bounce_pages;
 148 static int max_bounce_pages = MAX_BPAGES;
 149 static int bounce_alignment = 1; /* XXX temporary */
 150
 151 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
 152 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
 153
 154 struct bus_dmamap {
 155         struct bp_list  bpages;
 156         int             pagesneeded;
 157         int             pagesreserved;
 158         bus_dma_tag_t   dmat;
 159         void            *buf;           /* unmapped buffer pointer */
 160         bus_size_t      buflen;         /* unmapped buffer length */
 161         bus_dmamap_callback_t *callback;
 162         void            *callback_arg;
 163         STAILQ_ENTRY(bus_dmamap) links;
 164 };
 165
 166 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
 167         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
 168
 169 static struct bus_dmamap nobounce_dmamap;
 170
 171 static int              alloc_bounce_zone(bus_dma_tag_t);
 172 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
 173 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
 174 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
 175 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
 176                             vm_offset_t, bus_size_t);
 177 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
 178
 179 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
 180 static void             add_map_callback(bus_dmamap_t);
 181
 182 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 183 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
 184            0, "Total bounce pages");
 185 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
 186            0, "Max bounce pages per bounce zone");
 187 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
 188            &bounce_alignment, 0, "Obey alignment constraint");
 189
 190 static __inline int
 191 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 192 {
 193         int retval;
 194
 195         retval = 0;
 196         do {
 197                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
 198                      (bounce_alignment && (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 __inline
 209 bus_dma_segment_t *
 210 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
 211 {
 212         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 213                 return(cache);
 214 #ifdef SMP
 215         spin_lock_wr(&tag->spin);
 216 #endif
 217         return(tag->segments);
 218 }
 219
 220 static __inline
 221 void
 222 bus_dma_tag_unlock(bus_dma_tag_t tag)
 223 {
 224 #ifdef SMP
 225         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
 226                 spin_unlock_wr(&tag->spin);
 227 #endif
 228 }
 229
 230 /*
 231  * Allocate a device specific dma_tag.
 232  */
 233 int
 234 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 235                    bus_size_t boundary, bus_addr_t lowaddr,
 236                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 237                    void *filterarg, bus_size_t maxsize, int nsegments,
 238                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
 239 {
 240         bus_dma_tag_t newtag;
 241         int error = 0;
 242
 243         /*
 244          * Sanity checks
 245          */
 246
 247         if (alignment == 0)
 248                 alignment = 1;
 249         if (alignment & (alignment - 1))
 250                 panic("alignment must be power of 2\n");
 251
 252         if (boundary != 0) {
 253                 if (boundary & (boundary - 1))
 254                         panic("boundary must be power of 2\n");
 255                 if (boundary < maxsegsz) {
 256                         kprintf("boundary < maxsegsz:\n");
 257                         print_backtrace(-1);
 258                         maxsegsz = boundary;
 259                 }
 260         }
 261
 262         /* Return a NULL tag on failure */
 263         *dmat = NULL;
 264
 265         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
 266
 267 #ifdef SMP
 268         spin_init(&newtag->spin);
 269 #endif
 270         newtag->parent = parent;
 271         newtag->alignment = alignment;
 272         newtag->boundary = boundary;
 273         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
 274         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
 275         newtag->filter = filter;
 276         newtag->filterarg = filterarg;
 277         newtag->maxsize = maxsize;
 278         newtag->nsegments = nsegments;
 279         newtag->maxsegsz = maxsegsz;
 280         newtag->flags = flags;
 281         newtag->ref_count = 1; /* Count ourself */
 282         newtag->map_count = 0;
 283         newtag->segments = NULL;
 284         newtag->bounce_zone = NULL;
 285
 286         /* Take into account any restrictions imposed by our parent tag */
 287         if (parent != NULL) {
 288                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
 289                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 290
 291                 if (newtag->boundary == 0) {
 292                         newtag->boundary = parent->boundary;
 293                 } else if (parent->boundary != 0) {
 294                         newtag->boundary = MIN(parent->boundary,
 295                                                newtag->boundary);
 296                 }
 297
 298 #ifdef notyet
 299                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
 300 #endif
 301
 302                 if (newtag->filter == NULL) {
 303                         /*
 304                          * Short circuit looking at our parent directly
 305                          * since we have encapsulated all of its information
 306                          */
 307                         newtag->filter = parent->filter;
 308                         newtag->filterarg = parent->filterarg;
 309                         newtag->parent = parent->parent;
 310                 }
 311                 if (newtag->parent != NULL)
 312                         parent->ref_count++;
 313         }
 314
 315         if (newtag->lowaddr < ptoa(Maxmem))
 316                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
 317         if (bounce_alignment && newtag->alignment > 1 &&
 318             !(newtag->flags & BUS_DMA_ALIGNED))
 319                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
 320
 321         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
 322             (flags & BUS_DMA_ALLOCNOW) != 0) {
 323                 struct bounce_zone *bz;
 324
 325                 /* Must bounce */
 326
 327                 error = alloc_bounce_zone(newtag);
 328                 if (error)
 329                         goto back;
 330                 bz = newtag->bounce_zone;
 331
 332                 if (ptoa(bz->total_bpages) < maxsize) {
 333                         int pages;
 334
 335                         if (flags & BUS_DMA_ONEBPAGE) {
 336                                 pages = 1;
 337                         } else {
 338                                 pages = atop(round_page(maxsize)) -
 339                                         bz->total_bpages;
 340                                 pages = MAX(pages, 1);
 341                         }
 342
 343                         /* Add pages to our bounce pool */
 344                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
 345                                 error = ENOMEM;
 346
 347                         /* Performed initial allocation */
 348                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 349                 }
 350         }
 351 back:
 352         if (error)
 353                 kfree(newtag, M_DEVBUF);
 354         else
 355                 *dmat = newtag;
 356         return error;
 357 }
 358
 359 int
 360 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 361 {
 362         if (dmat != NULL) {
 363                 if (dmat->map_count != 0)
 364                         return (EBUSY);
 365
 366                 while (dmat != NULL) {
 367                         bus_dma_tag_t parent;
 368
 369                         parent = dmat->parent;
 370                         dmat->ref_count--;
 371                         if (dmat->ref_count == 0) {
 372                                 if (dmat->segments != NULL)
 373                                         kfree(dmat->segments, M_DEVBUF);
 374                                 kfree(dmat, M_DEVBUF);
 375                                 /*
 376                                  * Last reference count, so
 377                                  * release our reference
 378                                  * count on our parent.
 379                                  */
 380                                 dmat = parent;
 381                         } else
 382                                 dmat = NULL;
 383                 }
 384         }
 385         return (0);
 386 }
 387
 388 bus_size_t
 389 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
 390 {
 391         return(tag->maxsize);
 392 }
 393
 394 /*
 395  * Allocate a handle for mapping from kva/uva/physical
 396  * address space into bus device space.
 397  */
 398 int
 399 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 400 {
 401         int error;
 402
 403         error = 0;
 404
 405         if (dmat->segments == NULL) {
 406                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
 407                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 408                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 409         }
 410
 411         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 412                 struct bounce_zone *bz;
 413                 int maxpages;
 414
 415                 /* Must bounce */
 416
 417                 if (dmat->bounce_zone == NULL) {
 418                         error = alloc_bounce_zone(dmat);
 419                         if (error)
 420                                 return error;
 421                 }
 422                 bz = dmat->bounce_zone;
 423
 424                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
 425
 426                 /* Initialize the new map */
 427                 STAILQ_INIT(&((*mapp)->bpages));
 428
 429                 /*
 430                  * Attempt to add pages to our pool on a per-instance
 431                  * basis up to a sane limit.
 432                  */
 433                 if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
 434                         maxpages = max_bounce_pages;
 435                 } else {
 436                         maxpages = MIN(max_bounce_pages,
 437                                        Maxmem - atop(dmat->lowaddr));
 438                 }
 439                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0
 440                  || (dmat->map_count > 0
 441                   && bz->total_bpages < maxpages)) {
 442                         int pages;
 443
 444                         if (flags & BUS_DMA_ONEBPAGE) {
 445                                 pages = 1;
 446                         } else {
 447                                 pages = atop(round_page(dmat->maxsize));
 448                                 pages = MIN(maxpages - bz->total_bpages, pages);
 449                                 pages = MAX(pages, 1);
 450                         }
 451                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
 452                                 error = ENOMEM;
 453
 454                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 455                                 if (!error)
 456                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 457                         } else {
 458                                 error = 0;
 459                         }
 460                 }
 461         } else {
 462                 *mapp = NULL;
 463         }
 464         if (!error)
 465                 dmat->map_count++;
 466         return error;
 467 }
 468
 469 /*
 470  * Destroy a handle for mapping from kva/uva/physical
 471  * address space into bus device space.
 472  */
 473 int
 474 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 475 {
 476         if (map != NULL) {
 477                 if (STAILQ_FIRST(&map->bpages) != NULL)
 478                         return (EBUSY);
 479                 kfree(map, M_DEVBUF);
 480         }
 481         dmat->map_count--;
 482         return (0);
 483 }
 484
 485 static __inline bus_size_t
 486 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
 487 {
 488         bus_size_t maxsize = 0;
 489         uintptr_t vaddr = (uintptr_t)vaddr0;
 490
 491         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
 492                 if (verify || bootverbose)
 493                         kprintf("boundary check failed\n");
 494                 if (verify)
 495                         print_backtrace(-1); /* XXX panic */
 496                 maxsize = dmat->maxsize;
 497         }
 498         if (vaddr & (dmat->alignment - 1)) {
 499                 if (verify || bootverbose)
 500                         kprintf("alignment check failed\n");
 501                 if (verify)
 502                         print_backtrace(-1); /* XXX panic */
 503                 if (dmat->maxsize < dmat->alignment)
 504                         maxsize = dmat->alignment;
 505                 else
 506                         maxsize = dmat->maxsize;
 507         }
 508         return maxsize;
 509 }
 510
 511 /*
 512  * Allocate a piece of memory that can be efficiently mapped into
 513  * bus device space based on the constraints lited in the dma tag.
 514  *
 515  * mapp is degenerate.  By definition this allocation should not require
 516  * bounce buffers so do not allocate a dma map.
 517  */
 518 int
 519 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
 520                  bus_dmamap_t *mapp)
 521 {
 522         int mflags;
 523
 524         /* If we succeed, no mapping/bouncing will be required */
 525         *mapp = NULL;
 526
 527         if (dmat->segments == NULL) {
 528                 KKASSERT(dmat->nsegments < 16384);
 529                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 530                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 531         }
 532
 533         if (flags & BUS_DMA_NOWAIT)
 534                 mflags = M_NOWAIT;
 535         else
 536                 mflags = M_WAITOK;
 537         if (flags & BUS_DMA_ZERO)
 538                 mflags |= M_ZERO;
 539
 540         if (BUS_DMAMEM_KMALLOC(dmat)) {
 541                 bus_size_t maxsize;
 542
 543                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 544
 545                 /*
 546                  * XXX
 547                  * Check whether the allocation
 548                  * - crossed a page boundary
 549                  * - was not aligned
 550                  * Retry with power-of-2 alignment in the above cases.
 551                  */
 552                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 553                 if (maxsize) {
 554                         size_t size;
 555
 556                         kfree(*vaddr, M_DEVBUF);
 557                         /* XXX check for overflow? */
 558                         for (size = 1; size <= maxsize; size <<= 1)
 559                                 ;
 560                         *vaddr = kmalloc(size, M_DEVBUF, mflags);
 561                         check_kmalloc(dmat, *vaddr, 1);
 562                 }
 563         } else {
 564                 /*
 565                  * XXX Use Contigmalloc until it is merged into this facility
 566                  *     and handles multi-seg allocations.  Nobody is doing
 567                  *     multi-seg allocations yet though.
 568                  */
 569                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 570                     0ul, dmat->lowaddr, dmat->alignment, dmat->boundary);
 571         }
 572         if (*vaddr == NULL)
 573                 return (ENOMEM);
 574         return (0);
 575 }
 576
 577 /*
 578  * Free a piece of memory and it's allociated dmamap, that was allocated
 579  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 580  */
 581 void
 582 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 583 {
 584         /*
 585          * dmamem does not need to be bounced, so the map should be
 586          * NULL
 587          */
 588         if (map != NULL)
 589                 panic("bus_dmamem_free: Invalid map freed\n");
 590         if (BUS_DMAMEM_KMALLOC(dmat))
 591                 kfree(vaddr, M_DEVBUF);
 592         else
 593                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 594 }
 595
 596 static __inline vm_paddr_t
 597 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
 598 {
 599         if (pmap)
 600                 return pmap_extract(pmap, vaddr);
 601         else
 602                 return pmap_kextract(vaddr);
 603 }
 604
 605 /*
 606  * Utility function to load a linear buffer.  lastaddrp holds state
 607  * between invocations (for multiple-buffer loads).  segp contains
 608  * the segment following the starting one on entrace, and the ending
 609  * segment on exit.  first indicates if this is the first invocation
 610  * of this function.
 611  */
 612 static int
 613 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
 614                         bus_dmamap_t map,
 615                         void *buf, bus_size_t buflen,
 616                         bus_dma_segment_t *segments,
 617                         int nsegments,
 618                         pmap_t pmap,
 619                         int flags,
 620                         vm_paddr_t *lastpaddrp,
 621                         int *segp,
 622                         int first)
 623 {
 624         vm_offset_t vaddr;
 625         vm_paddr_t paddr, nextpaddr;
 626         bus_dma_segment_t *sg;
 627         bus_addr_t bmask;
 628         int seg, error = 0;
 629
 630         if (map == NULL)
 631                 map = &nobounce_dmamap;
 632
 633 #ifdef INVARIANTS
 634         if (dmat->flags & BUS_DMA_ALIGNED)
 635                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
 636 #endif
 637
 638         /*
 639          * If we are being called during a callback, pagesneeded will
 640          * be non-zero, so we can avoid doing the work twice.
 641          */
 642         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 643             map != &nobounce_dmamap && map->pagesneeded == 0) {
 644                 vm_offset_t vendaddr;
 645
 646                 /*
 647                  * Count the number of bounce pages
 648                  * needed in order to complete this transfer
 649                  */
 650                 vaddr = (vm_offset_t)buf;
 651                 vendaddr = (vm_offset_t)buf + buflen;
 652
 653                 while (vaddr < vendaddr) {
 654                         paddr = _bus_dma_extract(pmap, vaddr);
 655                         if (run_filter(dmat, paddr) != 0)
 656                                 map->pagesneeded++;
 657                         vaddr += (PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK));
 658                 }
 659         }
 660
 661         /* Reserve Necessary Bounce Pages */
 662         if (map->pagesneeded != 0) {
 663                 struct bounce_zone *bz;
 664
 665                 bz = dmat->bounce_zone;
 666                 BZ_LOCK(bz);
 667                 if (flags & BUS_DMA_NOWAIT) {
 668                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 669                                 BZ_UNLOCK(bz);
 670                                 error = ENOMEM;
 671                                 goto free_bounce;
 672                         }
 673                 } else {
 674                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 675                                 /* Queue us for resources */
 676                                 map->dmat = dmat;
 677                                 map->buf = buf;
 678                                 map->buflen = buflen;
 679
 680                                 STAILQ_INSERT_TAIL(
 681                                     &dmat->bounce_zone->bounce_map_waitinglist,
 682                                     map, links);
 683                                 BZ_UNLOCK(bz);
 684
 685                                 return (EINPROGRESS);
 686                         }
 687                 }
 688                 BZ_UNLOCK(bz);
 689         }
 690
 691         KKASSERT(*segp >= 1 && *segp <= nsegments);
 692         seg = *segp;
 693         sg = &segments[seg - 1];
 694
 695         vaddr = (vm_offset_t)buf;
 696         nextpaddr = *lastpaddrp;
 697         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
 698
 699         /* force at least one segment */
 700         do {
 701                 bus_size_t size;
 702
 703                 /*
 704                  * Per-page main loop
 705                  */
 706                 paddr = _bus_dma_extract(pmap, vaddr);
 707                 size = PAGE_SIZE - (paddr & PAGE_MASK);
 708                 if (size > buflen)
 709                         size = buflen;
 710                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
 711                         /*
 712                          * note: this paddr has the same in-page offset
 713                          * as vaddr and thus the paddr above, so the
 714                          * size does not have to be recalculated
 715                          */
 716                         paddr = add_bounce_page(dmat, map, vaddr, size);
 717                 }
 718
 719                 /*
 720                  * Fill in the bus_dma_segment
 721                  */
 722                 if (first) {
 723                         sg->ds_addr = paddr;
 724                         sg->ds_len = size;
 725                         first = 0;
 726                 } else if (paddr == nextpaddr) {
 727                         sg->ds_len += size;
 728                 } else {
 729                         sg++;
 730                         seg++;
 731                         if (seg > nsegments)
 732                                 break;
 733                         sg->ds_addr = paddr;
 734                         sg->ds_len = size;
 735                 }
 736                 nextpaddr = paddr + size;
 737
 738                 /*
 739                  * Handle maxsegsz and boundary issues with a nested loop
 740                  */
 741                 for (;;) {
 742                         bus_size_t tmpsize;
 743
 744                         /*
 745                          * Limit to the boundary and maximum segment size
 746                          */
 747                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
 748                                 tmpsize = dmat->boundary -
 749                                           (sg->ds_addr & ~bmask);
 750                                 if (tmpsize > dmat->maxsegsz)
 751                                         tmpsize = dmat->maxsegsz;
 752                                 KKASSERT(tmpsize < sg->ds_len);
 753                         } else if (sg->ds_len > dmat->maxsegsz) {
 754                                 tmpsize = dmat->maxsegsz;
 755                         } else {
 756                                 break;
 757                         }
 758
 759                         /*
 760                          * Futz, split the data into a new segment.
 761                          */
 762                         if (seg >= nsegments)
 763                                 goto fail;
 764                         sg[1].ds_len = sg[0].ds_len - tmpsize;
 765                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
 766                         sg[0].ds_len = tmpsize;
 767                         sg++;
 768                         seg++;
 769                 }
 770
 771                 /*
 772                  * Adjust for loop
 773                  */
 774                 buflen -= size;
 775                 vaddr += size;
 776         } while (buflen > 0);
 777 fail:
 778         if (buflen != 0)
 779                 error = EFBIG;
 780
 781         *segp = seg;
 782         *lastpaddrp = nextpaddr;
 783
 784 free_bounce:
 785         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 786             map != &nobounce_dmamap) {
 787                 _bus_dmamap_unload(dmat, map);
 788                 return_bounce_pages(dmat, map);
 789         }
 790         return error;
 791 }
 792
 793 /*
 794  * Map the buffer buf into bus space using the dmamap map.
 795  */
 796 int
 797 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 798                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 799                 void *callback_arg, int flags)
 800 {
 801         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 802         bus_dma_segment_t *segments;
 803         vm_paddr_t lastaddr = 0;
 804         int error, nsegs = 1;
 805
 806         if (map != NULL) {
 807                 /*
 808                  * XXX
 809                  * Follow old semantics.  Once all of the callers are fixed,
 810                  * we should get rid of these internal flag "adjustment".
 811                  */
 812                 flags &= ~BUS_DMA_NOWAIT;
 813                 flags |= BUS_DMA_WAITOK;
 814
 815                 map->callback = callback;
 816                 map->callback_arg = callback_arg;
 817         }
 818
 819         segments = bus_dma_tag_lock(dmat, cache_segments);
 820         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
 821                         segments, dmat->nsegments,
 822                         NULL, flags, &lastaddr, &nsegs, 1);
 823         if (error == EINPROGRESS) {
 824                 bus_dma_tag_unlock(dmat);
 825                 return error;
 826         }
 827         callback(callback_arg, segments, nsegs, error);
 828         bus_dma_tag_unlock(dmat);
 829         return 0;
 830 }
 831
 832 /*
 833  * Like _bus_dmamap_load(), but for mbufs.
 834  */
 835 int
 836 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
 837                      struct mbuf *m0,
 838                      bus_dmamap_callback2_t *callback, void *callback_arg,
 839                      int flags)
 840 {
 841         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 842         bus_dma_segment_t *segments;
 843         int nsegs, error;
 844
 845         /*
 846          * XXX
 847          * Follow old semantics.  Once all of the callers are fixed,
 848          * we should get rid of these internal flag "adjustment".
 849          */
 850         flags &= ~BUS_DMA_WAITOK;
 851         flags |= BUS_DMA_NOWAIT;
 852
 853         segments = bus_dma_tag_lock(dmat, cache_segments);
 854         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
 855                         segments, dmat->nsegments, &nsegs, flags);
 856         if (error) {
 857                 /* force "no valid mappings" in callback */
 858                 callback(callback_arg, segments, 0,
 859                          0, error);
 860         } else {
 861                 callback(callback_arg, segments, nsegs,
 862                          m0->m_pkthdr.len, error);
 863         }
 864         bus_dma_tag_unlock(dmat);
 865         return error;
 866 }
 867
 868 int
 869 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
 870                              struct mbuf *m0,
 871                              bus_dma_segment_t *segs, int maxsegs,
 872                              int *nsegs, int flags)
 873 {
 874         int error;
 875
 876         M_ASSERTPKTHDR(m0);
 877
 878         KASSERT(maxsegs >= 1, ("invalid maxsegs %d\n", maxsegs));
 879         KASSERT(maxsegs <= dmat->nsegments,
 880                 ("%d too many segments, dmat only support %d segments\n",
 881                  maxsegs, dmat->nsegments));
 882         KASSERT(flags & BUS_DMA_NOWAIT,
 883                 ("only BUS_DMA_NOWAIT is supported\n"));
 884
 885         if (m0->m_pkthdr.len <= dmat->maxsize) {
 886                 int first = 1;
 887                 vm_paddr_t lastaddr = 0;
 888                 struct mbuf *m;
 889
 890                 *nsegs = 1;
 891                 error = 0;
 892                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 893                         if (m->m_len == 0)
 894                                 continue;
 895
 896                         error = _bus_dmamap_load_buffer(dmat, map,
 897                                         m->m_data, m->m_len,
 898                                         segs, maxsegs,
 899                                         NULL, flags, &lastaddr,
 900                                         nsegs, first);
 901                         if (error == ENOMEM && !first) {
 902                                 /*
 903                                  * Out of bounce pages due to too many
 904                                  * fragments in the mbuf chain; return
 905                                  * EFBIG instead.
 906                                  */
 907                                 error = EFBIG;
 908                         }
 909                         first = 0;
 910                 }
 911 #ifdef INVARIANTS
 912                 if (!error)
 913                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
 914 #endif
 915         } else {
 916                 *nsegs = 0;
 917                 error = EINVAL;
 918         }
 919         KKASSERT(error != EINPROGRESS);
 920         return error;
 921 }
 922
 923 /*
 924  * Like _bus_dmamap_load(), but for uios.
 925  */
 926 int
 927 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
 928                     struct uio *uio,
 929                     bus_dmamap_callback2_t *callback, void *callback_arg,
 930                     int flags)
 931 {
 932         vm_paddr_t lastaddr;
 933         int nsegs, error, first, i;
 934         bus_size_t resid;
 935         struct iovec *iov;
 936         pmap_t pmap;
 937         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 938         bus_dma_segment_t *segments;
 939         bus_dma_segment_t *segs;
 940         int nsegs_left;
 941
 942         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 943                 segments = cache_segments;
 944         else
 945                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
 946                                    M_DEVBUF, M_WAITOK | M_ZERO);
 947
 948         /*
 949          * XXX
 950          * Follow old semantics.  Once all of the callers are fixed,
 951          * we should get rid of these internal flag "adjustment".
 952          */
 953         flags &= ~BUS_DMA_WAITOK;
 954         flags |= BUS_DMA_NOWAIT;
 955
 956         resid = (bus_size_t)uio->uio_resid;
 957         iov = uio->uio_iov;
 958
 959         segs = segments;
 960         nsegs_left = dmat->nsegments;
 961
 962         if (uio->uio_segflg == UIO_USERSPACE) {
 963                 struct thread *td;
 964
 965                 td = uio->uio_td;
 966                 KASSERT(td != NULL && td->td_proc != NULL,
 967                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
 968                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
 969         } else {
 970                 pmap = NULL;
 971         }
 972
 973         error = 0;
 974         nsegs = 1;
 975         first = 1;
 976         lastaddr = 0;
 977         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
 978                 /*
 979                  * Now at the first iovec to load.  Load each iovec
 980                  * until we have exhausted the residual count.
 981                  */
 982                 bus_size_t minlen =
 983                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
 984                 caddr_t addr = (caddr_t) iov[i].iov_base;
 985
 986                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
 987                                 segs, nsegs_left,
 988                                 pmap, flags, &lastaddr, &nsegs, first);
 989                 first = 0;
 990
 991                 resid -= minlen;
 992                 if (error == 0) {
 993                         nsegs_left -= nsegs;
 994                         segs += nsegs;
 995                 }
 996         }
 997
 998         /*
 999          * Minimum one DMA segment, even if 0-length buffer.
1000          */
1001         if (nsegs_left == dmat->nsegments)
1002                 --nsegs_left;
1003
1004         if (error) {
1005                 /* force "no valid mappings" in callback */
1006                 callback(callback_arg, segments, 0,
1007                          0, error);
1008         } else {
1009                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1010                          (bus_size_t)uio->uio_resid, error);
1011         }
1012         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1013                 kfree(segments, M_DEVBUF);
1014         return error;
1015 }
1016
1017 /*
1018  * Release the mapping held by map.
1019  */
1020 void
1021 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1022 {
1023         struct bounce_page *bpage;
1024
1025         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1026                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1027                 free_bounce_page(dmat, bpage);
1028         }
1029 }
1030
1031 void
1032 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1033 {
1034         struct bounce_page *bpage;
1035
1036         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1037                 /*
1038                  * Handle data bouncing.  We might also
1039                  * want to add support for invalidating
1040                  * the caches on broken hardware
1041                  */
1042                 switch (op) {
1043                 case BUS_DMASYNC_PREWRITE:
1044                         while (bpage != NULL) {
1045                                 bcopy((void *)bpage->datavaddr,
1046                                       (void *)bpage->vaddr,
1047                                       bpage->datacount);
1048                                 bpage = STAILQ_NEXT(bpage, links);
1049                         }
1050                         dmat->bounce_zone->total_bounced++;
1051                         break;
1052
1053                 case BUS_DMASYNC_POSTREAD:
1054                         while (bpage != NULL) {
1055                                 bcopy((void *)bpage->vaddr,
1056                                       (void *)bpage->datavaddr,
1057                                       bpage->datacount);
1058                                 bpage = STAILQ_NEXT(bpage, links);
1059                         }
1060                         dmat->bounce_zone->total_bounced++;
1061                         break;
1062
1063                 case BUS_DMASYNC_PREREAD:
1064                 case BUS_DMASYNC_POSTWRITE:
1065                         /* No-ops */
1066                         break;
1067                 }
1068         }
1069 }
1070
1071 static int
1072 alloc_bounce_zone(bus_dma_tag_t dmat)
1073 {
1074         struct bounce_zone *bz, *new_bz;
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(&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(&bounce_zone_tok);
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(&bounce_zone_tok);
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 }