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  */
  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_token);
 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         vm_memattr_t attr;
 525         int mflags;
 526
 527         /* If we succeed, no mapping/bouncing will be required */
 528         *mapp = NULL;
 529
 530         if (dmat->segments == NULL) {
 531                 KKASSERT(dmat->nsegments < 16384);
 532                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 533                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 534         }
 535
 536         if (flags & BUS_DMA_NOWAIT)
 537                 mflags = M_NOWAIT;
 538         else
 539                 mflags = M_WAITOK;
 540         if (flags & BUS_DMA_ZERO)
 541                 mflags |= M_ZERO;
 542         if (flags & BUS_DMA_NOCACHE)
 543                 attr = VM_MEMATTR_UNCACHEABLE;
 544         else
 545                 attr = VM_MEMATTR_DEFAULT;
 546
 547         /* XXX must alloc with correct mem attribute here */
 548         if (BUS_DMAMEM_KMALLOC(dmat)) {
 549                 bus_size_t maxsize;
 550
 551                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 552
 553                 /*
 554                  * XXX
 555                  * Check whether the allocation
 556                  * - crossed a page boundary
 557                  * - was not aligned
 558                  * Retry with power-of-2 alignment in the above cases.
 559                  */
 560                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 561                 if (maxsize) {
 562                         kfree(*vaddr, M_DEVBUF);
 563                         *vaddr = kmalloc(maxsize, M_DEVBUF,
 564                             mflags | M_POWEROF2);
 565                         check_kmalloc(dmat, *vaddr, 1);
 566                 }
 567         } else {
 568                 /*
 569                  * XXX Use Contigmalloc until it is merged into this facility
 570                  *     and handles multi-seg allocations.  Nobody is doing
 571                  *     multi-seg allocations yet though.
 572                  */
 573                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 574                     0ul, dmat->lowaddr, dmat->alignment, dmat->boundary);
 575         }
 576         if (*vaddr == NULL)
 577                 return (ENOMEM);
 578
 579         if (attr != VM_MEMATTR_DEFAULT)
 580                 pmap_change_attr((vm_offset_t)(*vaddr), dmat->maxsize / PAGE_SIZE, attr);
 581         return (0);
 582 }
 583
 584 /*
 585  * Free a piece of memory and it's allociated dmamap, that was allocated
 586  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 587  */
 588 void
 589 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 590 {
 591         /*
 592          * dmamem does not need to be bounced, so the map should be
 593          * NULL
 594          */
 595         if (map != NULL)
 596                 panic("bus_dmamem_free: Invalid map freed");
 597         if (BUS_DMAMEM_KMALLOC(dmat))
 598                 kfree(vaddr, M_DEVBUF);
 599         else
 600                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 601 }
 602
 603 static __inline vm_paddr_t
 604 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
 605 {
 606         if (pmap)
 607                 return pmap_extract(pmap, vaddr);
 608         else
 609                 return pmap_kextract(vaddr);
 610 }
 611
 612 /*
 613  * Utility function to load a linear buffer.  lastaddrp holds state
 614  * between invocations (for multiple-buffer loads).  segp contains
 615  * the segment following the starting one on entrace, and the ending
 616  * segment on exit.  first indicates if this is the first invocation
 617  * of this function.
 618  */
 619 static int
 620 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
 621                         bus_dmamap_t map,
 622                         void *buf, bus_size_t buflen,
 623                         bus_dma_segment_t *segments,
 624                         int nsegments,
 625                         pmap_t pmap,
 626                         int flags,
 627                         vm_paddr_t *lastpaddrp,
 628                         int *segp,
 629                         int first)
 630 {
 631         vm_offset_t vaddr;
 632         vm_paddr_t paddr, nextpaddr;
 633         bus_dma_segment_t *sg;
 634         bus_addr_t bmask;
 635         int seg, error = 0;
 636
 637         if (map == NULL)
 638                 map = &nobounce_dmamap;
 639
 640 #ifdef INVARIANTS
 641         if (dmat->flags & BUS_DMA_ALIGNED)
 642                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
 643 #endif
 644
 645         /*
 646          * If we are being called during a callback, pagesneeded will
 647          * be non-zero, so we can avoid doing the work twice.
 648          */
 649         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 650             map != &nobounce_dmamap && map->pagesneeded == 0) {
 651                 vm_offset_t vendaddr;
 652
 653                 /*
 654                  * Count the number of bounce pages
 655                  * needed in order to complete this transfer
 656                  */
 657                 vaddr = (vm_offset_t)buf;
 658                 vendaddr = (vm_offset_t)buf + buflen;
 659
 660                 while (vaddr < vendaddr) {
 661                         paddr = _bus_dma_extract(pmap, vaddr);
 662                         if (run_filter(dmat, paddr) != 0)
 663                                 map->pagesneeded++;
 664                         vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK));
 665                 }
 666         }
 667
 668         /* Reserve Necessary Bounce Pages */
 669         if (map->pagesneeded != 0) {
 670                 struct bounce_zone *bz;
 671
 672                 bz = dmat->bounce_zone;
 673                 BZ_LOCK(bz);
 674                 if (flags & BUS_DMA_NOWAIT) {
 675                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 676                                 BZ_UNLOCK(bz);
 677                                 error = ENOMEM;
 678                                 goto free_bounce;
 679                         }
 680                 } else {
 681                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 682                                 /* Queue us for resources */
 683                                 map->dmat = dmat;
 684                                 map->buf = buf;
 685                                 map->buflen = buflen;
 686
 687                                 STAILQ_INSERT_TAIL(
 688                                     &dmat->bounce_zone->bounce_map_waitinglist,
 689                                     map, links);
 690                                 BZ_UNLOCK(bz);
 691
 692                                 return (EINPROGRESS);
 693                         }
 694                 }
 695                 BZ_UNLOCK(bz);
 696         }
 697
 698         KKASSERT(*segp >= 1 && *segp <= nsegments);
 699         seg = *segp;
 700         sg = &segments[seg - 1];
 701
 702         vaddr = (vm_offset_t)buf;
 703         nextpaddr = *lastpaddrp;
 704         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
 705
 706         /* force at least one segment */
 707         do {
 708                 bus_size_t size;
 709
 710                 /*
 711                  * Per-page main loop
 712                  */
 713                 paddr = _bus_dma_extract(pmap, vaddr);
 714                 size = PAGE_SIZE - (paddr & PAGE_MASK);
 715                 if (size > buflen)
 716                         size = buflen;
 717                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
 718                         /*
 719                          * NOTE: paddr may have different in-page offset,
 720                          *       unless BUS_DMA_KEEP_PG_OFFSET is set.
 721                          */
 722                         paddr = add_bounce_page(dmat, map, vaddr, &size);
 723                 }
 724
 725                 /*
 726                  * Fill in the bus_dma_segment
 727                  */
 728                 if (first) {
 729                         sg->ds_addr = paddr;
 730                         sg->ds_len = size;
 731                         first = 0;
 732                 } else if (paddr == nextpaddr) {
 733                         sg->ds_len += size;
 734                 } else {
 735                         sg++;
 736                         seg++;
 737                         if (seg > nsegments)
 738                                 break;
 739                         sg->ds_addr = paddr;
 740                         sg->ds_len = size;
 741                 }
 742                 nextpaddr = paddr + size;
 743
 744                 /*
 745                  * Handle maxsegsz and boundary issues with a nested loop
 746                  */
 747                 for (;;) {
 748                         bus_size_t tmpsize;
 749
 750                         /*
 751                          * Limit to the boundary and maximum segment size
 752                          */
 753                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
 754                                 tmpsize = dmat->boundary -
 755                                           (sg->ds_addr & ~bmask);
 756                                 if (tmpsize > dmat->maxsegsz)
 757                                         tmpsize = dmat->maxsegsz;
 758                                 KKASSERT(tmpsize < sg->ds_len);
 759                         } else if (sg->ds_len > dmat->maxsegsz) {
 760                                 tmpsize = dmat->maxsegsz;
 761                         } else {
 762                                 break;
 763                         }
 764
 765                         /*
 766                          * Futz, split the data into a new segment.
 767                          */
 768                         if (seg >= nsegments)
 769                                 goto fail;
 770                         sg[1].ds_len = sg[0].ds_len - tmpsize;
 771                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
 772                         sg[0].ds_len = tmpsize;
 773                         sg++;
 774                         seg++;
 775                 }
 776
 777                 /*
 778                  * Adjust for loop
 779                  */
 780                 buflen -= size;
 781                 vaddr += size;
 782         } while (buflen > 0);
 783 fail:
 784         if (buflen != 0)
 785                 error = EFBIG;
 786
 787         *segp = seg;
 788         *lastpaddrp = nextpaddr;
 789
 790 free_bounce:
 791         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 792             map != &nobounce_dmamap) {
 793                 _bus_dmamap_unload(dmat, map);
 794                 return_bounce_pages(dmat, map);
 795         }
 796         return error;
 797 }
 798
 799 /*
 800  * Map the buffer buf into bus space using the dmamap map.
 801  */
 802 int
 803 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 804                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 805                 void *callback_arg, int flags)
 806 {
 807         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 808         bus_dma_segment_t *segments;
 809         vm_paddr_t lastaddr = 0;
 810         int error, nsegs = 1;
 811
 812         if (map != NULL) {
 813                 /*
 814                  * XXX
 815                  * Follow old semantics.  Once all of the callers are fixed,
 816                  * we should get rid of these internal flag "adjustment".
 817                  */
 818                 flags &= ~BUS_DMA_NOWAIT;
 819                 flags |= BUS_DMA_WAITOK;
 820
 821                 map->callback = callback;
 822                 map->callback_arg = callback_arg;
 823         }
 824
 825         segments = bus_dma_tag_lock(dmat, cache_segments);
 826         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
 827                         segments, dmat->nsegments,
 828                         NULL, flags, &lastaddr, &nsegs, 1);
 829         if (error == EINPROGRESS) {
 830                 KKASSERT((dmat->flags &
 831                           (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL)) !=
 832                          (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL));
 833
 834                 if (dmat->flags & BUS_DMA_PROTECTED)
 835                         panic("protected dmamap callback will be defered");
 836
 837                 bus_dma_tag_unlock(dmat);
 838                 return error;
 839         }
 840         callback(callback_arg, segments, nsegs, error);
 841         bus_dma_tag_unlock(dmat);
 842         return 0;
 843 }
 844
 845 /*
 846  * Like _bus_dmamap_load(), but for mbufs.
 847  */
 848 int
 849 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
 850                      struct mbuf *m0,
 851                      bus_dmamap_callback2_t *callback, void *callback_arg,
 852                      int flags)
 853 {
 854         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 855         bus_dma_segment_t *segments;
 856         int nsegs, error;
 857
 858         /*
 859          * XXX
 860          * Follow old semantics.  Once all of the callers are fixed,
 861          * we should get rid of these internal flag "adjustment".
 862          */
 863         flags &= ~BUS_DMA_WAITOK;
 864         flags |= BUS_DMA_NOWAIT;
 865
 866         segments = bus_dma_tag_lock(dmat, cache_segments);
 867         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
 868                         segments, dmat->nsegments, &nsegs, flags);
 869         if (error) {
 870                 /* force "no valid mappings" in callback */
 871                 callback(callback_arg, segments, 0,
 872                          0, error);
 873         } else {
 874                 callback(callback_arg, segments, nsegs,
 875                          m0->m_pkthdr.len, error);
 876         }
 877         bus_dma_tag_unlock(dmat);
 878         return error;
 879 }
 880
 881 int
 882 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
 883                              struct mbuf *m0,
 884                              bus_dma_segment_t *segs, int maxsegs,
 885                              int *nsegs, int flags)
 886 {
 887         int error;
 888
 889         M_ASSERTPKTHDR(m0);
 890
 891         KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs));
 892         KASSERT(maxsegs <= dmat->nsegments,
 893                 ("%d too many segments, dmat only supports %d segments",
 894                  maxsegs, dmat->nsegments));
 895         KASSERT(flags & BUS_DMA_NOWAIT,
 896                 ("only BUS_DMA_NOWAIT is supported"));
 897
 898         if (m0->m_pkthdr.len <= dmat->maxsize) {
 899                 int first = 1;
 900                 vm_paddr_t lastaddr = 0;
 901                 struct mbuf *m;
 902
 903                 *nsegs = 1;
 904                 error = 0;
 905                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 906                         if (m->m_len == 0)
 907                                 continue;
 908
 909                         error = _bus_dmamap_load_buffer(dmat, map,
 910                                         m->m_data, m->m_len,
 911                                         segs, maxsegs,
 912                                         NULL, flags, &lastaddr,
 913                                         nsegs, first);
 914                         if (error == ENOMEM && !first) {
 915                                 /*
 916                                  * Out of bounce pages due to too many
 917                                  * fragments in the mbuf chain; return
 918                                  * EFBIG instead.
 919                                  */
 920                                 error = EFBIG;
 921                                 break;
 922                         }
 923                         first = 0;
 924                 }
 925 #ifdef INVARIANTS
 926                 if (!error)
 927                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
 928 #endif
 929         } else {
 930                 *nsegs = 0;
 931                 error = EINVAL;
 932         }
 933         KKASSERT(error != EINPROGRESS);
 934         return error;
 935 }
 936
 937 /*
 938  * Like _bus_dmamap_load(), but for uios.
 939  */
 940 int
 941 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
 942                     struct uio *uio,
 943                     bus_dmamap_callback2_t *callback, void *callback_arg,
 944                     int flags)
 945 {
 946         vm_paddr_t lastaddr;
 947         int nsegs, error, first, i;
 948         bus_size_t resid;
 949         struct iovec *iov;
 950         pmap_t pmap;
 951         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 952         bus_dma_segment_t *segments;
 953         bus_dma_segment_t *segs;
 954         int nsegs_left;
 955
 956         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 957                 segments = cache_segments;
 958         else
 959                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
 960                                    M_DEVBUF, M_WAITOK | M_ZERO);
 961
 962         /*
 963          * XXX
 964          * Follow old semantics.  Once all of the callers are fixed,
 965          * we should get rid of these internal flag "adjustment".
 966          */
 967         flags &= ~BUS_DMA_WAITOK;
 968         flags |= BUS_DMA_NOWAIT;
 969
 970         resid = (bus_size_t)uio->uio_resid;
 971         iov = uio->uio_iov;
 972
 973         segs = segments;
 974         nsegs_left = dmat->nsegments;
 975
 976         if (uio->uio_segflg == UIO_USERSPACE) {
 977                 struct thread *td;
 978
 979                 td = uio->uio_td;
 980                 KASSERT(td != NULL && td->td_proc != NULL,
 981                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
 982                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
 983         } else {
 984                 pmap = NULL;
 985         }
 986
 987         error = 0;
 988         nsegs = 1;
 989         first = 1;
 990         lastaddr = 0;
 991         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
 992                 /*
 993                  * Now at the first iovec to load.  Load each iovec
 994                  * until we have exhausted the residual count.
 995                  */
 996                 bus_size_t minlen =
 997                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
 998                 caddr_t addr = (caddr_t) iov[i].iov_base;
 999
1000                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
1001                                 segs, nsegs_left,
1002                                 pmap, flags, &lastaddr, &nsegs, first);
1003                 first = 0;
1004
1005                 resid -= minlen;
1006                 if (error == 0) {
1007                         nsegs_left -= nsegs;
1008                         segs += nsegs;
1009                 }
1010         }
1011
1012         /*
1013          * Minimum one DMA segment, even if 0-length buffer.
1014          */
1015         if (nsegs_left == dmat->nsegments)
1016                 --nsegs_left;
1017
1018         if (error) {
1019                 /* force "no valid mappings" in callback */
1020                 callback(callback_arg, segments, 0,
1021                          0, error);
1022         } else {
1023                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1024                          (bus_size_t)uio->uio_resid, error);
1025         }
1026         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1027                 kfree(segments, M_DEVBUF);
1028         return error;
1029 }
1030
1031 /*
1032  * Release the mapping held by map.
1033  */
1034 void
1035 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1036 {
1037         struct bounce_page *bpage;
1038
1039         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1040                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1041                 free_bounce_page(dmat, bpage);
1042         }
1043 }
1044
1045 void
1046 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1047 {
1048         struct bounce_page *bpage;
1049
1050         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1051                 /*
1052                  * Handle data bouncing.  We might also
1053                  * want to add support for invalidating
1054                  * the caches on broken hardware
1055                  */
1056                 if (op & BUS_DMASYNC_PREWRITE) {
1057                         while (bpage != NULL) {
1058                                 bcopy((void *)bpage->datavaddr,
1059                                       (void *)bpage->vaddr,
1060                                       bpage->datacount);
1061                                 bpage = STAILQ_NEXT(bpage, links);
1062                         }
1063                         cpu_sfence();
1064                         dmat->bounce_zone->total_bounced++;
1065                 }
1066                 if (op & BUS_DMASYNC_POSTREAD) {
1067                         cpu_lfence();
1068                         while (bpage != NULL) {
1069                                 bcopy((void *)bpage->vaddr,
1070                                       (void *)bpage->datavaddr,
1071                                       bpage->datacount);
1072                                 bpage = STAILQ_NEXT(bpage, links);
1073                         }
1074                         dmat->bounce_zone->total_bounced++;
1075                 }
1076                 /* BUS_DMASYNC_PREREAD          - no operation on intel */
1077                 /* BUS_DMASYNC_POSTWRITE        - no operation on intel */
1078         }
1079 }
1080
1081 static int
1082 alloc_bounce_zone(bus_dma_tag_t dmat)
1083 {
1084         struct bounce_zone *bz, *new_bz;
1085
1086         KASSERT(dmat->bounce_zone == NULL,
1087                 ("bounce zone was already assigned"));
1088
1089         new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO);
1090
1091         lwkt_gettoken(&bounce_zone_tok);
1092
1093         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1094                 /*
1095                  * For shared bounce zone, check to see
1096                  * if we already have a suitable zone
1097                  */
1098                 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1099                         if (dmat->alignment <= bz->alignment &&
1100                             dmat->lowaddr >= bz->lowaddr) {
1101                                 lwkt_reltoken(&bounce_zone_tok);
1102
1103                                 dmat->bounce_zone = bz;
1104                                 kfree(new_bz, M_DEVBUF);
1105                                 return 0;
1106                         }
1107                 }
1108         }
1109         bz = new_bz;
1110
1111         spin_init(&bz->spin, "allocbouncezone");
1112         STAILQ_INIT(&bz->bounce_page_list);
1113         STAILQ_INIT(&bz->bounce_map_waitinglist);
1114         bz->free_bpages = 0;
1115         bz->reserved_bpages = 0;
1116         bz->active_bpages = 0;
1117         bz->lowaddr = dmat->lowaddr;
1118         bz->alignment = round_page(dmat->alignment);
1119         ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1120
1121         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1122                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1123                 busdma_zonecount++;
1124                 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1125         } else {
1126                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_priv_zonecount);
1127                 busdma_priv_zonecount--;
1128         }
1129
1130         lwkt_reltoken(&bounce_zone_tok);
1131
1132         dmat->bounce_zone = bz;
1133
1134         sysctl_ctx_init(&bz->sysctl_ctx);
1135         bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx,
1136             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1137             CTLFLAG_RD, 0, "");
1138         if (bz->sysctl_tree == NULL) {
1139                 sysctl_ctx_free(&bz->sysctl_ctx);
1140                 return 0;       /* XXX error code? */
1141         }
1142
1143         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1144             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1145             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1146             "Total bounce pages");
1147         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1148             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1149             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1150             "Free bounce pages");
1151         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1152             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1153             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1154             "Reserved bounce pages");
1155         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1156             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1157             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1158             "Active bounce pages");
1159         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1160             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1161             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1162             "Total bounce requests");
1163         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1164             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1165             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1166             "Total bounce requests that were deferred");
1167         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1168             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1169             "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0,
1170             "Total bounce page reservations that were failed");
1171         SYSCTL_ADD_STRING(&bz->sysctl_ctx,
1172             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1173             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1174         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1175             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1176             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1177
1178         return 0;
1179 }
1180
1181 static int
1182 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags)
1183 {
1184         struct bounce_zone *bz = dmat->bounce_zone;
1185         int count = 0, mflags;
1186
1187         if (flags & BUS_DMA_NOWAIT)
1188                 mflags = M_NOWAIT;
1189         else
1190                 mflags = M_WAITOK;
1191
1192         while (numpages > 0) {
1193                 struct bounce_page *bpage;
1194
1195                 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO);
1196
1197                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1198                                                          mflags, 0ul,
1199                                                          bz->lowaddr,
1200                                                          bz->alignment, 0);
1201                 if (bpage->vaddr == 0) {
1202                         kfree(bpage, M_DEVBUF);
1203                         break;
1204                 }
1205                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1206
1207                 BZ_LOCK(bz);
1208                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1209                 total_bounce_pages++;
1210                 bz->total_bpages++;
1211                 bz->free_bpages++;
1212                 BZ_UNLOCK(bz);
1213
1214                 count++;
1215                 numpages--;
1216         }
1217         return count;
1218 }
1219
1220 static void
1221 free_bounce_pages_all(bus_dma_tag_t dmat)
1222 {
1223         struct bounce_zone *bz = dmat->bounce_zone;
1224         struct bounce_page *bpage;
1225
1226         BZ_LOCK(bz);
1227
1228         while ((bpage = STAILQ_FIRST(&bz->bounce_page_list)) != NULL) {
1229                 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1230
1231                 KKASSERT(total_bounce_pages > 0);
1232                 total_bounce_pages--;
1233
1234                 KKASSERT(bz->total_bpages > 0);
1235                 bz->total_bpages--;
1236
1237                 KKASSERT(bz->free_bpages > 0);
1238                 bz->free_bpages--;
1239
1240                 BZ_UNLOCK(bz);
1241                 contigfree((void *)bpage->vaddr, PAGE_SIZE, M_DEVBUF);
1242                 kfree(bpage, M_DEVBUF);
1243                 BZ_LOCK(bz);
1244         }
1245         if (bz->total_bpages) {
1246                 kprintf("#%d bounce pages are still in use\n",
1247                         bz->total_bpages);
1248                 print_backtrace(-1);
1249         }
1250
1251         BZ_UNLOCK(bz);
1252 }
1253
1254 static void
1255 free_bounce_zone(bus_dma_tag_t dmat)
1256 {
1257         struct bounce_zone *bz = dmat->bounce_zone;
1258
1259         if (bz == NULL)
1260                 return;
1261
1262         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0)
1263                 return;
1264
1265         free_bounce_pages_all(dmat);
1266         dmat->bounce_zone = NULL;
1267
1268         if (bz->sysctl_tree != NULL)
1269                 sysctl_ctx_free(&bz->sysctl_ctx);
1270         kfree(bz, M_DEVBUF);
1271 }
1272
1273 /* Assume caller holds bounce zone spinlock */
1274 static int
1275 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1276 {
1277         struct bounce_zone *bz = dmat->bounce_zone;
1278         int pages;
1279
1280         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1281         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1282                 bz->reserve_failed++;
1283                 return (map->pagesneeded - (map->pagesreserved + pages));
1284         }
1285
1286         bz->free_bpages -= pages;
1287
1288         bz->reserved_bpages += pages;
1289         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1290
1291         map->pagesreserved += pages;
1292         pages = map->pagesneeded - map->pagesreserved;
1293
1294         return pages;
1295 }
1296
1297 static void
1298 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1299 {
1300         struct bounce_zone *bz = dmat->bounce_zone;
1301         int reserved = map->pagesreserved;
1302         bus_dmamap_t wait_map;
1303
1304         map->pagesreserved = 0;
1305         map->pagesneeded = 0;
1306
1307         if (reserved == 0)
1308                 return;
1309
1310         BZ_LOCK(bz);
1311
1312         bz->free_bpages += reserved;
1313         KKASSERT(bz->free_bpages <= bz->total_bpages);
1314
1315         KKASSERT(bz->reserved_bpages >= reserved);
1316         bz->reserved_bpages -= reserved;
1317
1318         wait_map = get_map_waiting(dmat);
1319
1320         BZ_UNLOCK(bz);
1321
1322         if (wait_map != NULL)
1323                 add_map_callback(map);
1324 }
1325
1326 static bus_addr_t
1327 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1328                 bus_size_t *sizep)
1329 {
1330         struct bounce_zone *bz = dmat->bounce_zone;
1331         struct bounce_page *bpage;
1332         bus_size_t size;
1333
1334         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1335         map->pagesneeded--;
1336
1337         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1338         map->pagesreserved--;
1339
1340         BZ_LOCK(bz);
1341
1342         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1343         KASSERT(bpage != NULL, ("free page list is empty"));
1344         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1345
1346         KKASSERT(bz->reserved_bpages > 0);
1347         bz->reserved_bpages--;
1348
1349         bz->active_bpages++;
1350         KKASSERT(bz->active_bpages <= bz->total_bpages);
1351
1352         BZ_UNLOCK(bz);
1353
1354         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1355                 /*
1356                  * Page offset needs to be preserved.  No size adjustments
1357                  * needed.
1358                  */
1359                 bpage->vaddr |= vaddr & PAGE_MASK;
1360                 bpage->busaddr |= vaddr & PAGE_MASK;
1361                 size = *sizep;
1362         } else {
1363                 /*
1364                  * Realign to bounce page base address, reduce size if
1365                  * necessary.  Bounce pages are typically already
1366                  * page-aligned.
1367                  */
1368                 size = PAGE_SIZE - (bpage->busaddr & PAGE_MASK);
1369                 if (size < *sizep) {
1370                         *sizep = size;
1371                 } else {
1372                         size = *sizep;
1373                 }
1374         }
1375
1376         bpage->datavaddr = vaddr;
1377         bpage->datacount = size;
1378         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1379         return bpage->busaddr;
1380 }
1381
1382 static void
1383 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1384 {
1385         struct bounce_zone *bz = dmat->bounce_zone;
1386         bus_dmamap_t map;
1387
1388         bpage->datavaddr = 0;
1389         bpage->datacount = 0;
1390
1391         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1392                 /*
1393                  * Reset the bounce page to start at offset 0.  Other uses
1394                  * of this bounce page may need to store a full page of
1395                  * data and/or assume it starts on a page boundary.
1396                  */
1397                 bpage->vaddr &= ~PAGE_MASK;
1398                 bpage->busaddr &= ~PAGE_MASK;
1399         }
1400
1401         BZ_LOCK(bz);
1402
1403         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1404
1405         bz->free_bpages++;
1406         KKASSERT(bz->free_bpages <= bz->total_bpages);
1407
1408         KKASSERT(bz->active_bpages > 0);
1409         bz->active_bpages--;
1410
1411         map = get_map_waiting(dmat);
1412
1413         BZ_UNLOCK(bz);
1414
1415         if (map != NULL)
1416                 add_map_callback(map);
1417 }
1418
1419 /* Assume caller holds bounce zone spinlock */
1420 static bus_dmamap_t
1421 get_map_waiting(bus_dma_tag_t dmat)
1422 {
1423         struct bounce_zone *bz = dmat->bounce_zone;
1424         bus_dmamap_t map;
1425
1426         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1427         if (map != NULL) {
1428                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1429                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1430                         bz->total_deferred++;
1431                 } else {
1432                         map = NULL;
1433                 }
1434         }
1435         return map;
1436 }
1437
1438 static void
1439 add_map_callback(bus_dmamap_t map)
1440 {
1441         spin_lock(&bounce_map_list_spin);
1442         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1443         busdma_swi_pending = 1;
1444         setsoftvm();
1445         spin_unlock(&bounce_map_list_spin);
1446 }
1447
1448 void
1449 busdma_swi(void)
1450 {
1451         bus_dmamap_t map;
1452
1453         spin_lock(&bounce_map_list_spin);
1454         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1455                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1456                 spin_unlock(&bounce_map_list_spin);
1457                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1458                                 map->callback, map->callback_arg, /*flags*/0);
1459                 spin_lock(&bounce_map_list_spin);
1460         }
1461         spin_unlock(&bounce_map_list_spin);
1462 }