sys/platform/pc32/i386/busdma_machdep.c

   1 /*
   2  * Copyright (c) 1997, 1998 Justin T. Gibbs.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions, and the following disclaimer,
  10  *    without modification, immediately at the beginning of the file.
  11  * 2. The name of the author may not be used to endorse or promote products
  12  *    derived from this software without specific prior written permission.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  18  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  *
  26  * $FreeBSD: src/sys/i386/i386/busdma_machdep.c,v 1.94 2008/08/15 20:51:31 kmacy Exp $
  27  */
  28
  29 #include <sys/param.h>
  30 #include <sys/systm.h>
  31 #include <sys/malloc.h>
  32 #include <sys/mbuf.h>
  33 #include <sys/uio.h>
  34 #include <sys/bus_dma.h>
  35 #include <sys/kernel.h>
  36 #include <sys/sysctl.h>
  37 #include <sys/lock.h>
  38
  39 #include <sys/thread2.h>
  40 #include <sys/spinlock2.h>
  41 #include <sys/mplock2.h>
  42
  43 #include <vm/vm.h>
  44 #include <vm/vm_page.h>
  45
  46 /* XXX needed for to access pmap to convert per-proc virtual to physical */
  47 #include <sys/proc.h>
  48 #include <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         struct spinlock spin;
  79 };
  80
  81 /*
  82  * bus_dma_tag private flags
  83  */
  84 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
  85 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
  86 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  87
  88 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
  89
  90 #define BUS_DMAMEM_KMALLOC(dmat) \
  91         ((dmat)->maxsize <= PAGE_SIZE && \
  92          (dmat)->alignment <= PAGE_SIZE && \
  93          (dmat)->lowaddr >= ptoa(Maxmem))
  94
  95 struct bounce_page {
  96         vm_offset_t     vaddr;          /* kva of bounce buffer */
  97         bus_addr_t      busaddr;        /* Physical address */
  98         vm_offset_t     datavaddr;      /* kva of client data */
  99         bus_size_t      datacount;      /* client data count */
 100         STAILQ_ENTRY(bounce_page) links;
 101 };
 102
 103 struct bounce_zone {
 104         STAILQ_ENTRY(bounce_zone) links;
 105         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 106         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 107         struct spinlock spin;
 108         int             total_bpages;
 109         int             free_bpages;
 110         int             reserved_bpages;
 111         int             active_bpages;
 112         int             total_bounced;
 113         int             total_deferred;
 114         int             reserve_failed;
 115         bus_size_t      alignment;
 116         bus_addr_t      lowaddr;
 117         char            zoneid[8];
 118         char            lowaddrid[20];
 119         struct sysctl_ctx_list sysctl_ctx;
 120         struct sysctl_oid *sysctl_tree;
 121 };
 122
 123 #define BZ_LOCK(bz)     spin_lock(&(bz)->spin)
 124 #define BZ_UNLOCK(bz)   spin_unlock(&(bz)->spin)
 125
 126 static struct lwkt_token bounce_zone_tok =
 127         LWKT_TOKEN_INITIALIZER(bounce_zone_tok);
 128 static int busdma_zonecount;
 129 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
 130         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
 131
 132 static int busdma_priv_zonecount = -1;
 133
 134 int busdma_swi_pending;
 135 static int total_bounce_pages;
 136 static int max_bounce_pages = MAX_BPAGES;
 137 static int bounce_alignment = 1; /* XXX temporary */
 138
 139 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
 140 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
 141
 142 struct bus_dmamap {
 143         struct bp_list  bpages;
 144         int             pagesneeded;
 145         int             pagesreserved;
 146         bus_dma_tag_t   dmat;
 147         void            *buf;           /* unmapped buffer pointer */
 148         bus_size_t      buflen;         /* unmapped buffer length */
 149         bus_dmamap_callback_t *callback;
 150         void            *callback_arg;
 151         STAILQ_ENTRY(bus_dmamap) links;
 152 };
 153
 154 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
 155         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
 156 static struct spinlock bounce_map_list_spin =
 157         SPINLOCK_INITIALIZER(&bounce_map_list_spin, "bounce_map_list_spin");
 158
 159 static struct bus_dmamap nobounce_dmamap;
 160
 161 static int              alloc_bounce_zone(bus_dma_tag_t);
 162 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
 163 static void             free_bounce_pages_all(bus_dma_tag_t);
 164 static void             free_bounce_zone(bus_dma_tag_t);
 165 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
 166 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
 167 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
 168                             vm_offset_t, bus_size_t);
 169 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
 170
 171 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
 172 static void             add_map_callback(bus_dmamap_t);
 173
 174 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 175 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
 176            0, "Total bounce pages");
 177 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
 178            0, "Max bounce pages per bounce zone");
 179 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
 180            &bounce_alignment, 0, "Obey alignment constraint");
 181
 182 static __inline int
 183 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 184 {
 185         int retval;
 186
 187         retval = 0;
 188         do {
 189                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
 190                      (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0))
 191                  && (dmat->filter == NULL ||
 192                      dmat->filter(dmat->filterarg, paddr) != 0))
 193                         retval = 1;
 194
 195                 dmat = dmat->parent;
 196         } while (retval == 0 && dmat != NULL);
 197         return (retval);
 198 }
 199
 200 static __inline
 201 bus_dma_segment_t *
 202 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
 203 {
 204         if (tag->flags & BUS_DMA_PROTECTED)
 205                 return(tag->segments);
 206
 207         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 208                 return(cache);
 209         spin_lock(&tag->spin);
 210         return(tag->segments);
 211 }
 212
 213 static __inline
 214 void
 215 bus_dma_tag_unlock(bus_dma_tag_t tag)
 216 {
 217         if (tag->flags & BUS_DMA_PROTECTED)
 218                 return;
 219
 220         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
 221                 spin_unlock(&tag->spin);
 222 }
 223
 224 /*
 225  * Allocate a device specific dma_tag.
 226  */
 227 int
 228 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 229                    bus_size_t boundary, bus_addr_t lowaddr,
 230                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 231                    void *filterarg, bus_size_t maxsize, int nsegments,
 232                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
 233 {
 234         bus_dma_tag_t newtag;
 235         int error = 0;
 236
 237         /*
 238          * Sanity checks
 239          */
 240
 241         if (alignment == 0)
 242                 alignment = 1;
 243         if (alignment & (alignment - 1))
 244                 panic("alignment must be power of 2");
 245
 246         if (boundary != 0) {
 247                 if (boundary & (boundary - 1))
 248                         panic("boundary must be power of 2");
 249                 if (boundary < maxsegsz) {
 250                         kprintf("boundary < maxsegsz:\n");
 251                         print_backtrace(-1);
 252                         maxsegsz = boundary;
 253                 }
 254         }
 255
 256         /* Return a NULL tag on failure */
 257         *dmat = NULL;
 258
 259         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
 260
 261         spin_init(&newtag->spin, "busdmacreate");
 262         newtag->parent = parent;
 263         newtag->alignment = alignment;
 264         newtag->boundary = boundary;
 265         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
 266         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
 267         newtag->filter = filter;
 268         newtag->filterarg = filterarg;
 269         newtag->maxsize = maxsize;
 270         newtag->nsegments = nsegments;
 271         newtag->maxsegsz = maxsegsz;
 272         newtag->flags = flags;
 273         newtag->ref_count = 1; /* Count ourself */
 274         newtag->map_count = 0;
 275         newtag->segments = NULL;
 276         newtag->bounce_zone = NULL;
 277
 278         /* Take into account any restrictions imposed by our parent tag */
 279         if (parent != NULL) {
 280                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
 281                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 282
 283                 if (newtag->boundary == 0) {
 284                         newtag->boundary = parent->boundary;
 285                 } else if (parent->boundary != 0) {
 286                         newtag->boundary = MIN(parent->boundary,
 287                                                newtag->boundary);
 288                 }
 289
 290 #ifdef notyet
 291                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
 292 #endif
 293
 294                 if (newtag->filter == NULL) {
 295                         /*
 296                          * Short circuit looking at our parent directly
 297                          * since we have encapsulated all of its information
 298                          */
 299                         newtag->filter = parent->filter;
 300                         newtag->filterarg = parent->filterarg;
 301                         newtag->parent = parent->parent;
 302                 }
 303                 if (newtag->parent != NULL)
 304                         parent->ref_count++;
 305         }
 306
 307         if (newtag->lowaddr < ptoa(Maxmem))
 308                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
 309         if (bounce_alignment && newtag->alignment > 1 &&
 310             !(newtag->flags & BUS_DMA_ALIGNED))
 311                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
 312
 313         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
 314             (flags & BUS_DMA_ALLOCNOW) != 0) {
 315                 struct bounce_zone *bz;
 316
 317                 /* Must bounce */
 318
 319                 error = alloc_bounce_zone(newtag);
 320                 if (error)
 321                         goto back;
 322                 bz = newtag->bounce_zone;
 323
 324                 if ((newtag->flags & BUS_DMA_ALLOCALL) == 0 &&
 325                     ptoa(bz->total_bpages) < maxsize) {
 326                         int pages;
 327
 328                         if (flags & BUS_DMA_ONEBPAGE) {
 329                                 pages = 1;
 330                         } else {
 331                                 pages = atop(round_page(maxsize)) -
 332                                         bz->total_bpages;
 333                                 pages = MAX(pages, 1);
 334                         }
 335
 336                         /* Add pages to our bounce pool */
 337                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
 338                                 error = ENOMEM;
 339
 340                         /* Performed initial allocation */
 341                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 342                 }
 343         }
 344 back:
 345         if (error) {
 346                 free_bounce_zone(newtag);
 347                 kfree(newtag, M_DEVBUF);
 348         } else {
 349                 *dmat = newtag;
 350         }
 351         return error;
 352 }
 353
 354 int
 355 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 356 {
 357         if (dmat != NULL) {
 358                 if (dmat->map_count != 0)
 359                         return (EBUSY);
 360
 361                 while (dmat != NULL) {
 362                         bus_dma_tag_t parent;
 363
 364                         parent = dmat->parent;
 365                         dmat->ref_count--;
 366                         if (dmat->ref_count == 0) {
 367                                 free_bounce_zone(dmat);
 368                                 if (dmat->segments != NULL)
 369                                         kfree(dmat->segments, M_DEVBUF);
 370                                 kfree(dmat, M_DEVBUF);
 371                                 /*
 372                                  * Last reference count, so
 373                                  * release our reference
 374                                  * count on our parent.
 375                                  */
 376                                 dmat = parent;
 377                         } else
 378                                 dmat = NULL;
 379                 }
 380         }
 381         return (0);
 382 }
 383
 384 bus_size_t
 385 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
 386 {
 387         return(tag->maxsize);
 388 }
 389
 390 /*
 391  * Allocate a handle for mapping from kva/uva/physical
 392  * address space into bus device space.
 393  */
 394 int
 395 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 396 {
 397         int error;
 398
 399         error = 0;
 400
 401         if (dmat->segments == NULL) {
 402                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
 403                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 404                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 405         }
 406
 407         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 408                 struct bounce_zone *bz;
 409                 int maxpages;
 410
 411                 /* Must bounce */
 412
 413                 if (dmat->bounce_zone == NULL) {
 414                         error = alloc_bounce_zone(dmat);
 415                         if (error)
 416                                 return error;
 417                 }
 418                 bz = dmat->bounce_zone;
 419
 420                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
 421
 422                 /* Initialize the new map */
 423                 STAILQ_INIT(&((*mapp)->bpages));
 424
 425                 /*
 426                  * Attempt to add pages to our pool on a per-instance
 427                  * basis up to a sane limit.
 428                  */
 429                 if (dmat->flags & BUS_DMA_ALLOCALL) {
 430                         maxpages = Maxmem - atop(dmat->lowaddr);
 431                 } else if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
 432                         maxpages = max_bounce_pages;
 433                 } else {
 434                         maxpages = MIN(max_bounce_pages,
 435                                        Maxmem - atop(dmat->lowaddr));
 436                 }
 437                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 ||
 438                     (dmat->map_count > 0 && bz->total_bpages < maxpages)) {
 439                         int pages;
 440
 441                         if (flags & BUS_DMA_ONEBPAGE) {
 442                                 pages = 1;
 443                         } else {
 444                                 pages = atop(round_page(dmat->maxsize));
 445                                 pages = MIN(maxpages - bz->total_bpages, pages);
 446                                 pages = MAX(pages, 1);
 447                         }
 448                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
 449                                 error = ENOMEM;
 450
 451                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 452                                 if (!error &&
 453                                     (dmat->flags & BUS_DMA_ALLOCALL) == 0)
 454                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 455                         } else {
 456                                 error = 0;
 457                         }
 458                 }
 459         } else {
 460                 *mapp = NULL;
 461         }
 462         if (!error) {
 463                 dmat->map_count++;
 464         } else {
 465                 kfree(*mapp, M_DEVBUF);
 466                 *mapp = NULL;
 467         }
 468         return error;
 469 }
 470
 471 /*
 472  * Destroy a handle for mapping from kva/uva/physical
 473  * address space into bus device space.
 474  */
 475 int
 476 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 477 {
 478         if (map != NULL) {
 479                 if (STAILQ_FIRST(&map->bpages) != NULL)
 480                         return (EBUSY);
 481                 kfree(map, M_DEVBUF);
 482         }
 483         dmat->map_count--;
 484         return (0);
 485 }
 486
 487 static __inline bus_size_t
 488 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
 489 {
 490         bus_size_t maxsize = 0;
 491         uintptr_t vaddr = (uintptr_t)vaddr0;
 492
 493         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
 494                 if (verify)
 495                         panic("boundary check failed\n");
 496                 if (bootverbose)
 497                         kprintf("boundary check failed\n");
 498                 maxsize = dmat->maxsize;
 499         }
 500         if (vaddr & (dmat->alignment - 1)) {
 501                 if (verify)
 502                         panic("alignment check failed\n");
 503                 if (bootverbose)
 504                         kprintf("alignment check failed\n");
 505                 if (dmat->maxsize < dmat->alignment)
 506                         maxsize = dmat->alignment;
 507                 else
 508                         maxsize = dmat->maxsize;
 509         }
 510         return maxsize;
 511 }
 512
 513 /*
 514  * Allocate a piece of memory that can be efficiently mapped into
 515  * bus device space based on the constraints lited in the dma tag.
 516  *
 517  * mapp is degenerate.  By definition this allocation should not require
 518  * bounce buffers so do not allocate a dma map.
 519  */
 520 int
 521 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
 522                  bus_dmamap_t *mapp)
 523 {
 524         int mflags;
 525
 526         /* If we succeed, no mapping/bouncing will be required */
 527         *mapp = NULL;
 528
 529         if (dmat->segments == NULL) {
 530                 KKASSERT(dmat->nsegments < 16384);
 531                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 532                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 533         }
 534
 535         if (flags & BUS_DMA_NOWAIT)
 536                 mflags = M_NOWAIT;
 537         else
 538                 mflags = M_WAITOK;
 539         if (flags & BUS_DMA_ZERO)
 540                 mflags |= M_ZERO;
 541
 542         if (BUS_DMAMEM_KMALLOC(dmat)) {
 543                 bus_size_t maxsize;
 544
 545                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 546
 547                 /*
 548                  * XXX
 549                  * Check whether the allocation
 550                  * - crossed a page boundary
 551                  * - was not aligned
 552                  * Retry with power-of-2 alignment in the above cases.
 553                  */
 554                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 555                 if (maxsize) {
 556                         kfree(*vaddr, M_DEVBUF);
 557                         *vaddr = kmalloc(maxsize, M_DEVBUF,
 558                             mflags | M_POWEROF2);
 559                         check_kmalloc(dmat, *vaddr, 1);
 560                 }
 561         } else {
 562                 /*
 563                  * XXX Use Contigmalloc until it is merged into this facility
 564                  *     and handles multi-seg allocations.  Nobody is doing
 565                  *     multi-seg allocations yet though.
 566                  */
 567                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 568                     0ul, dmat->lowaddr, dmat->alignment, dmat->boundary);
 569         }
 570         if (*vaddr == NULL)
 571                 return (ENOMEM);
 572         return (0);
 573 }
 574
 575 /*
 576  * Free a piece of memory and it's allociated dmamap, that was allocated
 577  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 578  */
 579 void
 580 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 581 {
 582         /*
 583          * dmamem does not need to be bounced, so the map should be
 584          * NULL
 585          */
 586         if (map != NULL)
 587                 panic("bus_dmamem_free: Invalid map freed");
 588         if (BUS_DMAMEM_KMALLOC(dmat))
 589                 kfree(vaddr, M_DEVBUF);
 590         else
 591                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 592 }
 593
 594 static __inline vm_paddr_t
 595 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
 596 {
 597         if (pmap)
 598                 return pmap_extract(pmap, vaddr);
 599         else
 600                 return pmap_kextract(vaddr);
 601 }
 602
 603 /*
 604  * Utility function to load a linear buffer.  lastaddrp holds state
 605  * between invocations (for multiple-buffer loads).  segp contains
 606  * the segment following the starting one on entrace, and the ending
 607  * segment on exit.  first indicates if this is the first invocation
 608  * of this function.
 609  */
 610 static int
 611 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
 612                         bus_dmamap_t map,
 613                         void *buf, bus_size_t buflen,
 614                         bus_dma_segment_t *segments,
 615                         int nsegments,
 616                         pmap_t pmap,
 617                         int flags,
 618                         vm_paddr_t *lastpaddrp,
 619                         int *segp,
 620                         int first)
 621 {
 622         vm_offset_t vaddr;
 623         vm_paddr_t paddr, nextpaddr;
 624         bus_dma_segment_t *sg;
 625         bus_addr_t bmask;
 626         int seg, error = 0;
 627
 628         if (map == NULL)
 629                 map = &nobounce_dmamap;
 630
 631 #ifdef INVARIANTS
 632         if (dmat->flags & BUS_DMA_ALIGNED)
 633                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
 634 #endif
 635
 636         /*
 637          * If we are being called during a callback, pagesneeded will
 638          * be non-zero, so we can avoid doing the work twice.
 639          */
 640         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 641             map != &nobounce_dmamap && map->pagesneeded == 0) {
 642                 vm_offset_t vendaddr;
 643
 644                 /*
 645                  * Count the number of bounce pages
 646                  * needed in order to complete this transfer
 647                  */
 648                 vaddr = (vm_offset_t)buf;
 649                 vendaddr = (vm_offset_t)buf + buflen;
 650
 651                 while (vaddr < vendaddr) {
 652                         paddr = _bus_dma_extract(pmap, vaddr);
 653                         if (run_filter(dmat, paddr) != 0)
 654                                 map->pagesneeded++;
 655                         vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK));
 656                 }
 657         }
 658
 659         /* Reserve Necessary Bounce Pages */
 660         if (map->pagesneeded != 0) {
 661                 struct bounce_zone *bz;
 662
 663                 bz = dmat->bounce_zone;
 664                 BZ_LOCK(bz);
 665                 if (flags & BUS_DMA_NOWAIT) {
 666                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 667                                 BZ_UNLOCK(bz);
 668                                 error = ENOMEM;
 669                                 goto free_bounce;
 670                         }
 671                 } else {
 672                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 673                                 /* Queue us for resources */
 674                                 map->dmat = dmat;
 675                                 map->buf = buf;
 676                                 map->buflen = buflen;
 677
 678                                 STAILQ_INSERT_TAIL(
 679                                     &dmat->bounce_zone->bounce_map_waitinglist,
 680                                     map, links);
 681                                 BZ_UNLOCK(bz);
 682
 683                                 return (EINPROGRESS);
 684                         }
 685                 }
 686                 BZ_UNLOCK(bz);
 687         }
 688
 689         KKASSERT(*segp >= 1 && *segp <= nsegments);
 690         seg = *segp;
 691         sg = &segments[seg - 1];
 692
 693         vaddr = (vm_offset_t)buf;
 694         nextpaddr = *lastpaddrp;
 695         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
 696
 697         /* force at least one segment */
 698         do {
 699                 bus_size_t size;
 700
 701                 /*
 702                  * Per-page main loop
 703                  */
 704                 paddr = _bus_dma_extract(pmap, vaddr);
 705                 size = PAGE_SIZE - (paddr & PAGE_MASK);
 706                 if (size > buflen)
 707                         size = buflen;
 708                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
 709                         /*
 710                          * note: this paddr has the same in-page offset
 711                          * as vaddr and thus the paddr above, so the
 712                          * size does not have to be recalculated
 713                          */
 714                         paddr = add_bounce_page(dmat, map, vaddr, size);
 715                 }
 716
 717                 /*
 718                  * Fill in the bus_dma_segment
 719                  */
 720                 if (first) {
 721                         sg->ds_addr = paddr;
 722                         sg->ds_len = size;
 723                         first = 0;
 724                 } else if (paddr == nextpaddr) {
 725                         sg->ds_len += size;
 726                 } else {
 727                         sg++;
 728                         seg++;
 729                         if (seg > nsegments)
 730                                 break;
 731                         sg->ds_addr = paddr;
 732                         sg->ds_len = size;
 733                 }
 734                 nextpaddr = paddr + size;
 735
 736                 /*
 737                  * Handle maxsegsz and boundary issues with a nested loop
 738                  */
 739                 for (;;) {
 740                         bus_size_t tmpsize;
 741
 742                         /*
 743                          * Limit to the boundary and maximum segment size
 744                          */
 745                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
 746                                 tmpsize = dmat->boundary -
 747                                           (sg->ds_addr & ~bmask);
 748                                 if (tmpsize > dmat->maxsegsz)
 749                                         tmpsize = dmat->maxsegsz;
 750                                 KKASSERT(tmpsize < sg->ds_len);
 751                         } else if (sg->ds_len > dmat->maxsegsz) {
 752                                 tmpsize = dmat->maxsegsz;
 753                         } else {
 754                                 break;
 755                         }
 756
 757                         /*
 758                          * Futz, split the data into a new segment.
 759                          */
 760                         if (seg >= nsegments)
 761                                 goto fail;
 762                         sg[1].ds_len = sg[0].ds_len - tmpsize;
 763                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
 764                         sg[0].ds_len = tmpsize;
 765                         sg++;
 766                         seg++;
 767                 }
 768
 769                 /*
 770                  * Adjust for loop
 771                  */
 772                 buflen -= size;
 773                 vaddr += size;
 774         } while (buflen > 0);
 775 fail:
 776         if (buflen != 0)
 777                 error = EFBIG;
 778
 779         *segp = seg;
 780         *lastpaddrp = nextpaddr;
 781
 782 free_bounce:
 783         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 784             map != &nobounce_dmamap) {
 785                 _bus_dmamap_unload(dmat, map);
 786                 return_bounce_pages(dmat, map);
 787         }
 788         return error;
 789 }
 790
 791 /*
 792  * Map the buffer buf into bus space using the dmamap map.
 793  */
 794 int
 795 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 796                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 797                 void *callback_arg, int flags)
 798 {
 799         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 800         bus_dma_segment_t *segments;
 801         vm_paddr_t lastaddr = 0;
 802         int error, nsegs = 1;
 803
 804         if (map != NULL) {
 805                 /*
 806                  * XXX
 807                  * Follow old semantics.  Once all of the callers are fixed,
 808                  * we should get rid of these internal flag "adjustment".
 809                  */
 810                 flags &= ~BUS_DMA_NOWAIT;
 811                 flags |= BUS_DMA_WAITOK;
 812
 813                 map->callback = callback;
 814                 map->callback_arg = callback_arg;
 815         }
 816
 817         segments = bus_dma_tag_lock(dmat, cache_segments);
 818         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
 819                         segments, dmat->nsegments,
 820                         NULL, flags, &lastaddr, &nsegs, 1);
 821         if (error == EINPROGRESS) {
 822                 KKASSERT((dmat->flags &
 823                           (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL)) !=
 824                          (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL));
 825
 826                 if (dmat->flags & BUS_DMA_PROTECTED)
 827                         panic("protected dmamap callback will be defered");
 828
 829                 bus_dma_tag_unlock(dmat);
 830                 return error;
 831         }
 832         callback(callback_arg, segments, nsegs, error);
 833         bus_dma_tag_unlock(dmat);
 834         return 0;
 835 }
 836
 837 /*
 838  * Like _bus_dmamap_load(), but for mbufs.
 839  */
 840 int
 841 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
 842                      struct mbuf *m0,
 843                      bus_dmamap_callback2_t *callback, void *callback_arg,
 844                      int flags)
 845 {
 846         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 847         bus_dma_segment_t *segments;
 848         int nsegs, error;
 849
 850         /*
 851          * XXX
 852          * Follow old semantics.  Once all of the callers are fixed,
 853          * we should get rid of these internal flag "adjustment".
 854          */
 855         flags &= ~BUS_DMA_WAITOK;
 856         flags |= BUS_DMA_NOWAIT;
 857
 858         segments = bus_dma_tag_lock(dmat, cache_segments);
 859         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
 860                         segments, dmat->nsegments, &nsegs, flags);
 861         if (error) {
 862                 /* force "no valid mappings" in callback */
 863                 callback(callback_arg, segments, 0,
 864                          0, error);
 865         } else {
 866                 callback(callback_arg, segments, nsegs,
 867                          m0->m_pkthdr.len, error);
 868         }
 869         bus_dma_tag_unlock(dmat);
 870         return error;
 871 }
 872
 873 int
 874 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
 875                              struct mbuf *m0,
 876                              bus_dma_segment_t *segs, int maxsegs,
 877                              int *nsegs, int flags)
 878 {
 879         int error;
 880
 881         M_ASSERTPKTHDR(m0);
 882
 883         KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs));
 884         KASSERT(maxsegs <= dmat->nsegments,
 885                 ("%d too many segments, dmat only supports %d segments",
 886                  maxsegs, dmat->nsegments));
 887         KASSERT(flags & BUS_DMA_NOWAIT,
 888                 ("only BUS_DMA_NOWAIT is supported"));
 889
 890         if (m0->m_pkthdr.len <= dmat->maxsize) {
 891                 int first = 1;
 892                 vm_paddr_t lastaddr = 0;
 893                 struct mbuf *m;
 894
 895                 *nsegs = 1;
 896                 error = 0;
 897                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 898                         if (m->m_len == 0)
 899                                 continue;
 900
 901                         error = _bus_dmamap_load_buffer(dmat, map,
 902                                         m->m_data, m->m_len,
 903                                         segs, maxsegs,
 904                                         NULL, flags, &lastaddr,
 905                                         nsegs, first);
 906                         if (error == ENOMEM && !first) {
 907                                 /*
 908                                  * Out of bounce pages due to too many
 909                                  * fragments in the mbuf chain; return
 910                                  * EFBIG instead.
 911                                  */
 912                                 error = EFBIG;
 913                         }
 914                         first = 0;
 915                 }
 916 #ifdef INVARIANTS
 917                 if (!error)
 918                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
 919 #endif
 920         } else {
 921                 *nsegs = 0;
 922                 error = EINVAL;
 923         }
 924         KKASSERT(error != EINPROGRESS);
 925         return error;
 926 }
 927
 928 /*
 929  * Like _bus_dmamap_load(), but for uios.
 930  */
 931 int
 932 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
 933                     struct uio *uio,
 934                     bus_dmamap_callback2_t *callback, void *callback_arg,
 935                     int flags)
 936 {
 937         vm_paddr_t lastaddr;
 938         int nsegs, error, first, i;
 939         bus_size_t resid;
 940         struct iovec *iov;
 941         pmap_t pmap;
 942         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 943         bus_dma_segment_t *segments;
 944         bus_dma_segment_t *segs;
 945         int nsegs_left;
 946
 947         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 948                 segments = cache_segments;
 949         else
 950                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
 951                                    M_DEVBUF, M_WAITOK | M_ZERO);
 952
 953         /*
 954          * XXX
 955          * Follow old semantics.  Once all of the callers are fixed,
 956          * we should get rid of these internal flag "adjustment".
 957          */
 958         flags &= ~BUS_DMA_WAITOK;
 959         flags |= BUS_DMA_NOWAIT;
 960
 961         resid = (bus_size_t)uio->uio_resid;
 962         iov = uio->uio_iov;
 963
 964         segs = segments;
 965         nsegs_left = dmat->nsegments;
 966
 967         if (uio->uio_segflg == UIO_USERSPACE) {
 968                 struct thread *td;
 969
 970                 td = uio->uio_td;
 971                 KASSERT(td != NULL && td->td_proc != NULL,
 972                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
 973                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
 974         } else {
 975                 pmap = NULL;
 976         }
 977
 978         error = 0;
 979         nsegs = 1;
 980         first = 1;
 981         lastaddr = 0;
 982         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
 983                 /*
 984                  * Now at the first iovec to load.  Load each iovec
 985                  * until we have exhausted the residual count.
 986                  */
 987                 bus_size_t minlen =
 988                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
 989                 caddr_t addr = (caddr_t) iov[i].iov_base;
 990
 991                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
 992                                 segs, nsegs_left,
 993                                 pmap, flags, &lastaddr, &nsegs, first);
 994                 first = 0;
 995
 996                 resid -= minlen;
 997                 if (error == 0) {
 998                         nsegs_left -= nsegs;
 999                         segs += nsegs;
1000                 }
1001         }
1002
1003         /*
1004          * Minimum one DMA segment, even if 0-length buffer.
1005          */
1006         if (nsegs_left == dmat->nsegments)
1007                 --nsegs_left;
1008
1009         if (error) {
1010                 /* force "no valid mappings" in callback */
1011                 callback(callback_arg, segments, 0,
1012                          0, error);
1013         } else {
1014                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1015                          (bus_size_t)uio->uio_resid, error);
1016         }
1017         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1018                 kfree(segments, M_DEVBUF);
1019         return error;
1020 }
1021
1022 /*
1023  * Release the mapping held by map.
1024  */
1025 void
1026 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1027 {
1028         struct bounce_page *bpage;
1029
1030         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1031                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1032                 free_bounce_page(dmat, bpage);
1033         }
1034 }
1035
1036 void
1037 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1038 {
1039         struct bounce_page *bpage;
1040
1041         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1042                 /*
1043                  * Handle data bouncing.  We might also
1044                  * want to add support for invalidating
1045                  * the caches on broken hardware
1046                  */
1047                 if (op & BUS_DMASYNC_PREWRITE) {
1048                         while (bpage != NULL) {
1049                                 bcopy((void *)bpage->datavaddr,
1050                                       (void *)bpage->vaddr,
1051                                       bpage->datacount);
1052                                 bpage = STAILQ_NEXT(bpage, links);
1053                         }
1054                         dmat->bounce_zone->total_bounced++;
1055                 }
1056                 if (op & BUS_DMASYNC_POSTREAD) {
1057                         while (bpage != NULL) {
1058                                 bcopy((void *)bpage->vaddr,
1059                                       (void *)bpage->datavaddr,
1060                                       bpage->datacount);
1061                                 bpage = STAILQ_NEXT(bpage, links);
1062                         }
1063                         dmat->bounce_zone->total_bounced++;
1064                 }
1065         }
1066 }
1067
1068 static int
1069 alloc_bounce_zone(bus_dma_tag_t dmat)
1070 {
1071         struct bounce_zone *bz, *new_bz;
1072
1073         KASSERT(dmat->bounce_zone == NULL,
1074                 ("bounce zone was already assigned"));
1075
1076         new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO);
1077
1078         lwkt_gettoken(&bounce_zone_tok);
1079
1080         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1081                 /*
1082                  * For shared bounce zone, check to see
1083                  * if we already have a suitable zone
1084                  */
1085                 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1086                         if (dmat->alignment <= bz->alignment &&
1087                             dmat->lowaddr >= bz->lowaddr) {
1088                                 lwkt_reltoken(&bounce_zone_tok);
1089
1090                                 dmat->bounce_zone = bz;
1091                                 kfree(new_bz, M_DEVBUF);
1092                                 return 0;
1093                         }
1094                 }
1095         }
1096         bz = new_bz;
1097
1098         spin_init(&bz->spin, "allocbouncezone");
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->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1107
1108         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1109                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1110                 busdma_zonecount++;
1111                 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1112         } else {
1113                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_priv_zonecount);
1114                 busdma_priv_zonecount--;
1115         }
1116
1117         lwkt_reltoken(&bounce_zone_tok);
1118
1119         dmat->bounce_zone = bz;
1120
1121         sysctl_ctx_init(&bz->sysctl_ctx);
1122         bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx,
1123             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1124             CTLFLAG_RD, 0, "");
1125         if (bz->sysctl_tree == NULL) {
1126                 sysctl_ctx_free(&bz->sysctl_ctx);
1127                 return 0;       /* XXX error code? */
1128         }
1129
1130         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1131             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1132             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1133             "Total bounce pages");
1134         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1135             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1136             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1137             "Free bounce pages");
1138         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1139             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1140             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1141             "Reserved bounce pages");
1142         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1143             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1144             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1145             "Active bounce pages");
1146         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1147             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1148             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1149             "Total bounce requests");
1150         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1151             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1152             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1153             "Total bounce requests that were deferred");
1154         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1155             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1156             "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0,
1157             "Total bounce page reservations that were failed");
1158         SYSCTL_ADD_STRING(&bz->sysctl_ctx,
1159             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1160             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1161         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1162             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1163             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1164
1165         return 0;
1166 }
1167
1168 static int
1169 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags)
1170 {
1171         struct bounce_zone *bz = dmat->bounce_zone;
1172         int count = 0, mflags;
1173
1174         if (flags & BUS_DMA_NOWAIT)
1175                 mflags = M_NOWAIT;
1176         else
1177                 mflags = M_WAITOK;
1178
1179         while (numpages > 0) {
1180                 struct bounce_page *bpage;
1181
1182                 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO);
1183
1184                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1185                                                          mflags, 0ul,
1186                                                          bz->lowaddr,
1187                                                          bz->alignment, 0);
1188                 if (bpage->vaddr == 0) {
1189                         kfree(bpage, M_DEVBUF);
1190                         break;
1191                 }
1192                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1193
1194                 BZ_LOCK(bz);
1195                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1196                 total_bounce_pages++;
1197                 bz->total_bpages++;
1198                 bz->free_bpages++;
1199                 BZ_UNLOCK(bz);
1200
1201                 count++;
1202                 numpages--;
1203         }
1204         return count;
1205 }
1206
1207 static void
1208 free_bounce_pages_all(bus_dma_tag_t dmat)
1209 {
1210         struct bounce_zone *bz = dmat->bounce_zone;
1211         struct bounce_page *bpage;
1212
1213         BZ_LOCK(bz);
1214
1215         while ((bpage = STAILQ_FIRST(&bz->bounce_page_list)) != NULL) {
1216                 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1217
1218                 KKASSERT(total_bounce_pages > 0);
1219                 total_bounce_pages--;
1220
1221                 KKASSERT(bz->total_bpages > 0);
1222                 bz->total_bpages--;
1223
1224                 KKASSERT(bz->free_bpages > 0);
1225                 bz->free_bpages--;
1226
1227                 BZ_UNLOCK(bz);
1228                 contigfree((void *)bpage->vaddr, PAGE_SIZE, M_DEVBUF);
1229                 kfree(bpage, M_DEVBUF);
1230                 BZ_LOCK(bz);
1231         }
1232         if (bz->total_bpages) {
1233                 kprintf("#%d bounce pages are still in use\n",
1234                         bz->total_bpages);
1235                 print_backtrace(-1);
1236         }
1237
1238         BZ_UNLOCK(bz);
1239 }
1240
1241 static void
1242 free_bounce_zone(bus_dma_tag_t dmat)
1243 {
1244         struct bounce_zone *bz = dmat->bounce_zone;
1245
1246         if (bz == NULL)
1247                 return;
1248
1249         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0)
1250                 return;
1251
1252         free_bounce_pages_all(dmat);
1253         dmat->bounce_zone = NULL;
1254
1255         if (bz->sysctl_tree != NULL)
1256                 sysctl_ctx_free(&bz->sysctl_ctx);
1257         kfree(bz, M_DEVBUF);
1258 }
1259
1260 /* Assume caller holds bounce zone spinlock */
1261 static int
1262 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1263 {
1264         struct bounce_zone *bz = dmat->bounce_zone;
1265         int pages;
1266
1267         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1268         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1269                 bz->reserve_failed++;
1270                 return (map->pagesneeded - (map->pagesreserved + pages));
1271         }
1272
1273         bz->free_bpages -= pages;
1274
1275         bz->reserved_bpages += pages;
1276         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1277
1278         map->pagesreserved += pages;
1279         pages = map->pagesneeded - map->pagesreserved;
1280
1281         return pages;
1282 }
1283
1284 static void
1285 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1286 {
1287         struct bounce_zone *bz = dmat->bounce_zone;
1288         int reserved = map->pagesreserved;
1289         bus_dmamap_t wait_map;
1290
1291         map->pagesreserved = 0;
1292         map->pagesneeded = 0;
1293
1294         if (reserved == 0)
1295                 return;
1296
1297         BZ_LOCK(bz);
1298
1299         bz->free_bpages += reserved;
1300         KKASSERT(bz->free_bpages <= bz->total_bpages);
1301
1302         KKASSERT(bz->reserved_bpages >= reserved);
1303         bz->reserved_bpages -= reserved;
1304
1305         wait_map = get_map_waiting(dmat);
1306
1307         BZ_UNLOCK(bz);
1308
1309         if (wait_map != NULL)
1310                 add_map_callback(map);
1311 }
1312
1313 static bus_addr_t
1314 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1315                 bus_size_t size)
1316 {
1317         struct bounce_zone *bz = dmat->bounce_zone;
1318         struct bounce_page *bpage;
1319
1320         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1321         map->pagesneeded--;
1322
1323         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1324         map->pagesreserved--;
1325
1326         BZ_LOCK(bz);
1327
1328         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1329         KASSERT(bpage != NULL, ("free page list is empty"));
1330         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1331
1332         KKASSERT(bz->reserved_bpages > 0);
1333         bz->reserved_bpages--;
1334
1335         bz->active_bpages++;
1336         KKASSERT(bz->active_bpages <= bz->total_bpages);
1337
1338         BZ_UNLOCK(bz);
1339
1340         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1341                 /* Page offset needs to be preserved. */
1342                 bpage->vaddr |= vaddr & PAGE_MASK;
1343                 bpage->busaddr |= vaddr & PAGE_MASK;
1344         }
1345         bpage->datavaddr = vaddr;
1346         bpage->datacount = size;
1347         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1348         return bpage->busaddr;
1349 }
1350
1351 static void
1352 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1353 {
1354         struct bounce_zone *bz = dmat->bounce_zone;
1355         bus_dmamap_t map;
1356
1357         bpage->datavaddr = 0;
1358         bpage->datacount = 0;
1359
1360         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1361                 /*
1362                  * Reset the bounce page to start at offset 0.  Other uses
1363                  * of this bounce page may need to store a full page of
1364                  * data and/or assume it starts on a page boundary.
1365                  */
1366                 bpage->vaddr &= ~PAGE_MASK;
1367                 bpage->busaddr &= ~PAGE_MASK;
1368         }
1369
1370         BZ_LOCK(bz);
1371
1372         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1373
1374         bz->free_bpages++;
1375         KKASSERT(bz->free_bpages <= bz->total_bpages);
1376
1377         KKASSERT(bz->active_bpages > 0);
1378         bz->active_bpages--;
1379
1380         map = get_map_waiting(dmat);
1381
1382         BZ_UNLOCK(bz);
1383
1384         if (map != NULL)
1385                 add_map_callback(map);
1386 }
1387
1388 /* Assume caller holds bounce zone spinlock */
1389 static bus_dmamap_t
1390 get_map_waiting(bus_dma_tag_t dmat)
1391 {
1392         struct bounce_zone *bz = dmat->bounce_zone;
1393         bus_dmamap_t map;
1394
1395         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1396         if (map != NULL) {
1397                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1398                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1399                         bz->total_deferred++;
1400                 } else {
1401                         map = NULL;
1402                 }
1403         }
1404         return map;
1405 }
1406
1407 static void
1408 add_map_callback(bus_dmamap_t map)
1409 {
1410         spin_lock(&bounce_map_list_spin);
1411         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1412         busdma_swi_pending = 1;
1413         setsoftvm();
1414         spin_unlock(&bounce_map_list_spin);
1415 }
1416
1417 void
1418 busdma_swi(void)
1419 {
1420         bus_dmamap_t map;
1421
1422         spin_lock(&bounce_map_list_spin);
1423         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1424                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1425                 spin_unlock(&bounce_map_list_spin);
1426                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1427                                 map->callback, map->callback_arg, /*flags*/0);
1428                 spin_lock(&bounce_map_list_spin);
1429         }
1430         spin_unlock(&bounce_map_list_spin);
1431 }