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 <sys/lock.h>
  49 #include <vm/vm_map.h>
  50
  51 #include <machine/md_var.h>
  52
  53 #define MAX_BPAGES      1024
  54
  55 /*
  56  * 16 x N declared on stack.
  57  */
  58 #define BUS_DMA_CACHE_SEGMENTS  8
  59
  60 struct bounce_zone;
  61 struct bus_dmamap;
  62
  63 struct bus_dma_tag {
  64         bus_dma_tag_t   parent;
  65         bus_size_t      alignment;
  66         bus_size_t      boundary;
  67         bus_addr_t      lowaddr;
  68         bus_addr_t      highaddr;
  69         bus_dma_filter_t *filter;
  70         void            *filterarg;
  71         bus_size_t      maxsize;
  72         u_int           nsegments;
  73         bus_size_t      maxsegsz;
  74         int             flags;
  75         int             ref_count;
  76         int             map_count;
  77         bus_dma_segment_t *segments;
  78         struct bounce_zone *bounce_zone;
  79 #ifdef SMP
  80         struct spinlock spin;
  81 #else
  82         int             unused0;
  83 #endif
  84 };
  85
  86 /*
  87  * bus_dma_tag private flags
  88  */
  89 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
  90 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
  91 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  92
  93 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
  94
  95 #define BUS_DMAMEM_KMALLOC(dmat) \
  96         ((dmat)->maxsize <= PAGE_SIZE && \
  97          (dmat)->alignment <= PAGE_SIZE && \
  98          (dmat)->lowaddr >= ptoa(Maxmem))
  99
 100 struct bounce_page {
 101         vm_offset_t     vaddr;          /* kva of bounce buffer */
 102         bus_addr_t      busaddr;        /* Physical address */
 103         vm_offset_t     datavaddr;      /* kva of client data */
 104         bus_size_t      datacount;      /* client data count */
 105         STAILQ_ENTRY(bounce_page) links;
 106 };
 107
 108 struct bounce_zone {
 109         STAILQ_ENTRY(bounce_zone) links;
 110         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 111         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 112 #ifdef SMP
 113         struct spinlock spin;
 114 #else
 115         int             unused0;
 116 #endif
 117         int             total_bpages;
 118         int             free_bpages;
 119         int             reserved_bpages;
 120         int             active_bpages;
 121         int             total_bounced;
 122         int             total_deferred;
 123         int             reserve_failed;
 124         bus_size_t      alignment;
 125         bus_addr_t      lowaddr;
 126         char            zoneid[8];
 127         char            lowaddrid[20];
 128         struct sysctl_ctx_list sysctl_ctx;
 129         struct sysctl_oid *sysctl_tree;
 130 };
 131
 132 #ifdef SMP
 133 #define BZ_LOCK(bz)     spin_lock(&(bz)->spin)
 134 #define BZ_UNLOCK(bz)   spin_unlock(&(bz)->spin)
 135 #else
 136 #define BZ_LOCK(bz)     crit_enter()
 137 #define BZ_UNLOCK(bz)   crit_exit()
 138 #endif
 139
 140 static struct lwkt_token bounce_zone_tok =
 141         LWKT_TOKEN_INITIALIZER(bounce_zone_tok);
 142 static int busdma_zonecount;
 143 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
 144         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
 145
 146 static int busdma_priv_zonecount = -1;
 147
 148 int busdma_swi_pending;
 149 static int total_bounce_pages;
 150 static int max_bounce_pages = MAX_BPAGES;
 151 static int bounce_alignment = 1; /* XXX temporary */
 152
 153 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
 154 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
 155
 156 struct bus_dmamap {
 157         struct bp_list  bpages;
 158         int             pagesneeded;
 159         int             pagesreserved;
 160         bus_dma_tag_t   dmat;
 161         void            *buf;           /* unmapped buffer pointer */
 162         bus_size_t      buflen;         /* unmapped buffer length */
 163         bus_dmamap_callback_t *callback;
 164         void            *callback_arg;
 165         STAILQ_ENTRY(bus_dmamap) links;
 166 };
 167
 168 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
 169         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
 170 static struct spinlock bounce_map_list_spin =
 171         SPINLOCK_INITIALIZER(&bounce_map_list_spin);
 172
 173 static struct bus_dmamap nobounce_dmamap;
 174
 175 static int              alloc_bounce_zone(bus_dma_tag_t);
 176 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
 177 static void             free_bounce_pages_all(bus_dma_tag_t);
 178 static void             free_bounce_zone(bus_dma_tag_t);
 179 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
 180 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
 181 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
 182                             vm_offset_t, bus_size_t);
 183 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
 184
 185 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
 186 static void             add_map_callback(bus_dmamap_t);
 187
 188 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 189 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
 190            0, "Total bounce pages");
 191 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
 192            0, "Max bounce pages per bounce zone");
 193 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
 194            &bounce_alignment, 0, "Obey alignment constraint");
 195
 196 static __inline int
 197 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 198 {
 199         int retval;
 200
 201         retval = 0;
 202         do {
 203                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
 204                      (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0))
 205                  && (dmat->filter == NULL ||
 206                      dmat->filter(dmat->filterarg, paddr) != 0))
 207                         retval = 1;
 208
 209                 dmat = dmat->parent;
 210         } while (retval == 0 && dmat != NULL);
 211         return (retval);
 212 }
 213
 214 static __inline
 215 bus_dma_segment_t *
 216 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
 217 {
 218         if (tag->flags & BUS_DMA_PROTECTED)
 219                 return(tag->segments);
 220
 221         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 222                 return(cache);
 223 #ifdef SMP
 224         spin_lock(&tag->spin);
 225 #endif
 226         return(tag->segments);
 227 }
 228
 229 static __inline
 230 void
 231 bus_dma_tag_unlock(bus_dma_tag_t tag)
 232 {
 233 #ifdef SMP
 234         if (tag->flags & BUS_DMA_PROTECTED)
 235                 return;
 236
 237         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
 238                 spin_unlock(&tag->spin);
 239 #endif
 240 }
 241
 242 /*
 243  * Allocate a device specific dma_tag.
 244  */
 245 int
 246 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 247                    bus_size_t boundary, bus_addr_t lowaddr,
 248                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 249                    void *filterarg, bus_size_t maxsize, int nsegments,
 250                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
 251 {
 252         bus_dma_tag_t newtag;
 253         int error = 0;
 254
 255         /*
 256          * Sanity checks
 257          */
 258
 259         if (alignment == 0)
 260                 alignment = 1;
 261         if (alignment & (alignment - 1))
 262                 panic("alignment must be power of 2");
 263
 264         if (boundary != 0) {
 265                 if (boundary & (boundary - 1))
 266                         panic("boundary must be power of 2");
 267                 if (boundary < maxsegsz) {
 268                         kprintf("boundary < maxsegsz:\n");
 269                         print_backtrace(-1);
 270                         maxsegsz = boundary;
 271                 }
 272         }
 273
 274         /* Return a NULL tag on failure */
 275         *dmat = NULL;
 276
 277         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
 278
 279 #ifdef SMP
 280         spin_init(&newtag->spin);
 281 #endif
 282         newtag->parent = parent;
 283         newtag->alignment = alignment;
 284         newtag->boundary = boundary;
 285         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
 286         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
 287         newtag->filter = filter;
 288         newtag->filterarg = filterarg;
 289         newtag->maxsize = maxsize;
 290         newtag->nsegments = nsegments;
 291         newtag->maxsegsz = maxsegsz;
 292         newtag->flags = flags;
 293         newtag->ref_count = 1; /* Count ourself */
 294         newtag->map_count = 0;
 295         newtag->segments = NULL;
 296         newtag->bounce_zone = NULL;
 297
 298         /* Take into account any restrictions imposed by our parent tag */
 299         if (parent != NULL) {
 300                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
 301                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 302
 303                 if (newtag->boundary == 0) {
 304                         newtag->boundary = parent->boundary;
 305                 } else if (parent->boundary != 0) {
 306                         newtag->boundary = MIN(parent->boundary,
 307                                                newtag->boundary);
 308                 }
 309
 310 #ifdef notyet
 311                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
 312 #endif
 313
 314                 if (newtag->filter == NULL) {
 315                         /*
 316                          * Short circuit looking at our parent directly
 317                          * since we have encapsulated all of its information
 318                          */
 319                         newtag->filter = parent->filter;
 320                         newtag->filterarg = parent->filterarg;
 321                         newtag->parent = parent->parent;
 322                 }
 323                 if (newtag->parent != NULL)
 324                         parent->ref_count++;
 325         }
 326
 327         if (newtag->lowaddr < ptoa(Maxmem))
 328                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
 329         if (bounce_alignment && newtag->alignment > 1 &&
 330             !(newtag->flags & BUS_DMA_ALIGNED))
 331                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
 332
 333         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
 334             (flags & BUS_DMA_ALLOCNOW) != 0) {
 335                 struct bounce_zone *bz;
 336
 337                 /* Must bounce */
 338
 339                 error = alloc_bounce_zone(newtag);
 340                 if (error)
 341                         goto back;
 342                 bz = newtag->bounce_zone;
 343
 344                 if ((newtag->flags & BUS_DMA_ALLOCALL) == 0 &&
 345                     ptoa(bz->total_bpages) < maxsize) {
 346                         int pages;
 347
 348                         if (flags & BUS_DMA_ONEBPAGE) {
 349                                 pages = 1;
 350                         } else {
 351                                 pages = atop(round_page(maxsize)) -
 352                                         bz->total_bpages;
 353                                 pages = MAX(pages, 1);
 354                         }
 355
 356                         /* Add pages to our bounce pool */
 357                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
 358                                 error = ENOMEM;
 359
 360                         /* Performed initial allocation */
 361                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 362                 }
 363         }
 364 back:
 365         if (error) {
 366                 free_bounce_zone(newtag);
 367                 kfree(newtag, M_DEVBUF);
 368         } else {
 369                 *dmat = newtag;
 370         }
 371         return error;
 372 }
 373
 374 int
 375 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 376 {
 377         if (dmat != NULL) {
 378                 if (dmat->map_count != 0)
 379                         return (EBUSY);
 380
 381                 while (dmat != NULL) {
 382                         bus_dma_tag_t parent;
 383
 384                         parent = dmat->parent;
 385                         dmat->ref_count--;
 386                         if (dmat->ref_count == 0) {
 387                                 free_bounce_zone(dmat);
 388                                 if (dmat->segments != NULL)
 389                                         kfree(dmat->segments, M_DEVBUF);
 390                                 kfree(dmat, M_DEVBUF);
 391                                 /*
 392                                  * Last reference count, so
 393                                  * release our reference
 394                                  * count on our parent.
 395                                  */
 396                                 dmat = parent;
 397                         } else
 398                                 dmat = NULL;
 399                 }
 400         }
 401         return (0);
 402 }
 403
 404 bus_size_t
 405 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
 406 {
 407         return(tag->maxsize);
 408 }
 409
 410 /*
 411  * Allocate a handle for mapping from kva/uva/physical
 412  * address space into bus device space.
 413  */
 414 int
 415 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 416 {
 417         int error;
 418
 419         error = 0;
 420
 421         if (dmat->segments == NULL) {
 422                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
 423                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 424                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 425         }
 426
 427         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 428                 struct bounce_zone *bz;
 429                 int maxpages;
 430
 431                 /* Must bounce */
 432
 433                 if (dmat->bounce_zone == NULL) {
 434                         error = alloc_bounce_zone(dmat);
 435                         if (error)
 436                                 return error;
 437                 }
 438                 bz = dmat->bounce_zone;
 439
 440                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
 441
 442                 /* Initialize the new map */
 443                 STAILQ_INIT(&((*mapp)->bpages));
 444
 445                 /*
 446                  * Attempt to add pages to our pool on a per-instance
 447                  * basis up to a sane limit.
 448                  */
 449                 if (dmat->flags & BUS_DMA_ALLOCALL) {
 450                         maxpages = Maxmem - atop(dmat->lowaddr);
 451                 } else if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
 452                         maxpages = max_bounce_pages;
 453                 } else {
 454                         maxpages = MIN(max_bounce_pages,
 455                                        Maxmem - atop(dmat->lowaddr));
 456                 }
 457                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 ||
 458                     (dmat->map_count > 0 && bz->total_bpages < maxpages)) {
 459                         int pages;
 460
 461                         if (flags & BUS_DMA_ONEBPAGE) {
 462                                 pages = 1;
 463                         } else {
 464                                 pages = atop(round_page(dmat->maxsize));
 465                                 pages = MIN(maxpages - bz->total_bpages, pages);
 466                                 pages = MAX(pages, 1);
 467                         }
 468                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
 469                                 error = ENOMEM;
 470
 471                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 472                                 if (!error &&
 473                                     (dmat->flags & BUS_DMA_ALLOCALL) == 0)
 474                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 475                         } else {
 476                                 error = 0;
 477                         }
 478                 }
 479         } else {
 480                 *mapp = NULL;
 481         }
 482         if (!error) {
 483                 dmat->map_count++;
 484         } else {
 485                 kfree(*mapp, M_DEVBUF);
 486                 *mapp = NULL;
 487         }
 488         return error;
 489 }
 490
 491 /*
 492  * Destroy a handle for mapping from kva/uva/physical
 493  * address space into bus device space.
 494  */
 495 int
 496 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 497 {
 498         if (map != NULL) {
 499                 if (STAILQ_FIRST(&map->bpages) != NULL)
 500                         return (EBUSY);
 501                 kfree(map, M_DEVBUF);
 502         }
 503         dmat->map_count--;
 504         return (0);
 505 }
 506
 507 static __inline bus_size_t
 508 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
 509 {
 510         bus_size_t maxsize = 0;
 511         uintptr_t vaddr = (uintptr_t)vaddr0;
 512
 513         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
 514                 if (verify || bootverbose)
 515                         kprintf("boundary check failed\n");
 516                 if (verify)
 517                         print_backtrace(-1); /* XXX panic */
 518                 maxsize = dmat->maxsize;
 519         }
 520         if (vaddr & (dmat->alignment - 1)) {
 521                 if (verify || bootverbose)
 522                         kprintf("alignment check failed\n");
 523                 if (verify)
 524                         print_backtrace(-1); /* XXX panic */
 525                 if (dmat->maxsize < dmat->alignment)
 526                         maxsize = dmat->alignment;
 527                 else
 528                         maxsize = dmat->maxsize;
 529         }
 530         return maxsize;
 531 }
 532
 533 /*
 534  * Allocate a piece of memory that can be efficiently mapped into
 535  * bus device space based on the constraints lited in the dma tag.
 536  *
 537  * mapp is degenerate.  By definition this allocation should not require
 538  * bounce buffers so do not allocate a dma map.
 539  */
 540 int
 541 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
 542                  bus_dmamap_t *mapp)
 543 {
 544         int mflags;
 545
 546         /* If we succeed, no mapping/bouncing will be required */
 547         *mapp = NULL;
 548
 549         if (dmat->segments == NULL) {
 550                 KKASSERT(dmat->nsegments < 16384);
 551                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 552                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 553         }
 554
 555         if (flags & BUS_DMA_NOWAIT)
 556                 mflags = M_NOWAIT;
 557         else
 558                 mflags = M_WAITOK;
 559         if (flags & BUS_DMA_ZERO)
 560                 mflags |= M_ZERO;
 561
 562         if (BUS_DMAMEM_KMALLOC(dmat)) {
 563                 bus_size_t maxsize;
 564
 565                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 566
 567                 /*
 568                  * XXX
 569                  * Check whether the allocation
 570                  * - crossed a page boundary
 571                  * - was not aligned
 572                  * Retry with power-of-2 alignment in the above cases.
 573                  */
 574                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 575                 if (maxsize) {
 576                         kfree(*vaddr, M_DEVBUF);
 577                         *vaddr = kmalloc_powerof2(maxsize, M_DEVBUF, mflags);
 578                         check_kmalloc(dmat, *vaddr, 1);
 579                 }
 580         } else {
 581                 /*
 582                  * XXX Use Contigmalloc until it is merged into this facility
 583                  *     and handles multi-seg allocations.  Nobody is doing
 584                  *     multi-seg allocations yet though.
 585                  */
 586                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 587                     0ul, dmat->lowaddr, dmat->alignment, dmat->boundary);
 588         }
 589         if (*vaddr == NULL)
 590                 return (ENOMEM);
 591         return (0);
 592 }
 593
 594 /*
 595  * Free a piece of memory and it's allociated dmamap, that was allocated
 596  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 597  */
 598 void
 599 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 600 {
 601         /*
 602          * dmamem does not need to be bounced, so the map should be
 603          * NULL
 604          */
 605         if (map != NULL)
 606                 panic("bus_dmamem_free: Invalid map freed");
 607         if (BUS_DMAMEM_KMALLOC(dmat))
 608                 kfree(vaddr, M_DEVBUF);
 609         else
 610                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 611 }
 612
 613 static __inline vm_paddr_t
 614 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
 615 {
 616         if (pmap)
 617                 return pmap_extract(pmap, vaddr);
 618         else
 619                 return pmap_kextract(vaddr);
 620 }
 621
 622 /*
 623  * Utility function to load a linear buffer.  lastaddrp holds state
 624  * between invocations (for multiple-buffer loads).  segp contains
 625  * the segment following the starting one on entrace, and the ending
 626  * segment on exit.  first indicates if this is the first invocation
 627  * of this function.
 628  */
 629 static int
 630 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
 631                         bus_dmamap_t map,
 632                         void *buf, bus_size_t buflen,
 633                         bus_dma_segment_t *segments,
 634                         int nsegments,
 635                         pmap_t pmap,
 636                         int flags,
 637                         vm_paddr_t *lastpaddrp,
 638                         int *segp,
 639                         int first)
 640 {
 641         vm_offset_t vaddr;
 642         vm_paddr_t paddr, nextpaddr;
 643         bus_dma_segment_t *sg;
 644         bus_addr_t bmask;
 645         int seg, error = 0;
 646
 647         if (map == NULL)
 648                 map = &nobounce_dmamap;
 649
 650 #ifdef INVARIANTS
 651         if (dmat->flags & BUS_DMA_ALIGNED)
 652                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
 653 #endif
 654
 655         /*
 656          * If we are being called during a callback, pagesneeded will
 657          * be non-zero, so we can avoid doing the work twice.
 658          */
 659         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 660             map != &nobounce_dmamap && map->pagesneeded == 0) {
 661                 vm_offset_t vendaddr;
 662
 663                 /*
 664                  * Count the number of bounce pages
 665                  * needed in order to complete this transfer
 666                  */
 667                 vaddr = (vm_offset_t)buf;
 668                 vendaddr = (vm_offset_t)buf + buflen;
 669
 670                 while (vaddr < vendaddr) {
 671                         paddr = _bus_dma_extract(pmap, vaddr);
 672                         if (run_filter(dmat, paddr) != 0)
 673                                 map->pagesneeded++;
 674                         vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK));
 675                 }
 676         }
 677
 678         /* Reserve Necessary Bounce Pages */
 679         if (map->pagesneeded != 0) {
 680                 struct bounce_zone *bz;
 681
 682                 bz = dmat->bounce_zone;
 683                 BZ_LOCK(bz);
 684                 if (flags & BUS_DMA_NOWAIT) {
 685                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 686                                 BZ_UNLOCK(bz);
 687                                 error = ENOMEM;
 688                                 goto free_bounce;
 689                         }
 690                 } else {
 691                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 692                                 /* Queue us for resources */
 693                                 map->dmat = dmat;
 694                                 map->buf = buf;
 695                                 map->buflen = buflen;
 696
 697                                 STAILQ_INSERT_TAIL(
 698                                     &dmat->bounce_zone->bounce_map_waitinglist,
 699                                     map, links);
 700                                 BZ_UNLOCK(bz);
 701
 702                                 return (EINPROGRESS);
 703                         }
 704                 }
 705                 BZ_UNLOCK(bz);
 706         }
 707
 708         KKASSERT(*segp >= 1 && *segp <= nsegments);
 709         seg = *segp;
 710         sg = &segments[seg - 1];
 711
 712         vaddr = (vm_offset_t)buf;
 713         nextpaddr = *lastpaddrp;
 714         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
 715
 716         /* force at least one segment */
 717         do {
 718                 bus_size_t size;
 719
 720                 /*
 721                  * Per-page main loop
 722                  */
 723                 paddr = _bus_dma_extract(pmap, vaddr);
 724                 size = PAGE_SIZE - (paddr & PAGE_MASK);
 725                 if (size > buflen)
 726                         size = buflen;
 727                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
 728                         /*
 729                          * note: this paddr has the same in-page offset
 730                          * as vaddr and thus the paddr above, so the
 731                          * size does not have to be recalculated
 732                          */
 733                         paddr = add_bounce_page(dmat, map, vaddr, size);
 734                 }
 735
 736                 /*
 737                  * Fill in the bus_dma_segment
 738                  */
 739                 if (first) {
 740                         sg->ds_addr = paddr;
 741                         sg->ds_len = size;
 742                         first = 0;
 743                 } else if (paddr == nextpaddr) {
 744                         sg->ds_len += size;
 745                 } else {
 746                         sg++;
 747                         seg++;
 748                         if (seg > nsegments)
 749                                 break;
 750                         sg->ds_addr = paddr;
 751                         sg->ds_len = size;
 752                 }
 753                 nextpaddr = paddr + size;
 754
 755                 /*
 756                  * Handle maxsegsz and boundary issues with a nested loop
 757                  */
 758                 for (;;) {
 759                         bus_size_t tmpsize;
 760
 761                         /*
 762                          * Limit to the boundary and maximum segment size
 763                          */
 764                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
 765                                 tmpsize = dmat->boundary -
 766                                           (sg->ds_addr & ~bmask);
 767                                 if (tmpsize > dmat->maxsegsz)
 768                                         tmpsize = dmat->maxsegsz;
 769                                 KKASSERT(tmpsize < sg->ds_len);
 770                         } else if (sg->ds_len > dmat->maxsegsz) {
 771                                 tmpsize = dmat->maxsegsz;
 772                         } else {
 773                                 break;
 774                         }
 775
 776                         /*
 777                          * Futz, split the data into a new segment.
 778                          */
 779                         if (seg >= nsegments)
 780                                 goto fail;
 781                         sg[1].ds_len = sg[0].ds_len - tmpsize;
 782                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
 783                         sg[0].ds_len = tmpsize;
 784                         sg++;
 785                         seg++;
 786                 }
 787
 788                 /*
 789                  * Adjust for loop
 790                  */
 791                 buflen -= size;
 792                 vaddr += size;
 793         } while (buflen > 0);
 794 fail:
 795         if (buflen != 0)
 796                 error = EFBIG;
 797
 798         *segp = seg;
 799         *lastpaddrp = nextpaddr;
 800
 801 free_bounce:
 802         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 803             map != &nobounce_dmamap) {
 804                 _bus_dmamap_unload(dmat, map);
 805                 return_bounce_pages(dmat, map);
 806         }
 807         return error;
 808 }
 809
 810 /*
 811  * Map the buffer buf into bus space using the dmamap map.
 812  */
 813 int
 814 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 815                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 816                 void *callback_arg, int flags)
 817 {
 818         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 819         bus_dma_segment_t *segments;
 820         vm_paddr_t lastaddr = 0;
 821         int error, nsegs = 1;
 822
 823         if (map != NULL) {
 824                 /*
 825                  * XXX
 826                  * Follow old semantics.  Once all of the callers are fixed,
 827                  * we should get rid of these internal flag "adjustment".
 828                  */
 829                 flags &= ~BUS_DMA_NOWAIT;
 830                 flags |= BUS_DMA_WAITOK;
 831
 832                 map->callback = callback;
 833                 map->callback_arg = callback_arg;
 834         }
 835
 836         segments = bus_dma_tag_lock(dmat, cache_segments);
 837         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
 838                         segments, dmat->nsegments,
 839                         NULL, flags, &lastaddr, &nsegs, 1);
 840         if (error == EINPROGRESS) {
 841                 KKASSERT((dmat->flags &
 842                           (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL)) !=
 843                          (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL));
 844
 845                 if (dmat->flags & BUS_DMA_PROTECTED)
 846                         panic("protected dmamap callback will be defered");
 847
 848                 bus_dma_tag_unlock(dmat);
 849                 return error;
 850         }
 851         callback(callback_arg, segments, nsegs, error);
 852         bus_dma_tag_unlock(dmat);
 853         return 0;
 854 }
 855
 856 /*
 857  * Like _bus_dmamap_load(), but for mbufs.
 858  */
 859 int
 860 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
 861                      struct mbuf *m0,
 862                      bus_dmamap_callback2_t *callback, void *callback_arg,
 863                      int flags)
 864 {
 865         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 866         bus_dma_segment_t *segments;
 867         int nsegs, error;
 868
 869         /*
 870          * XXX
 871          * Follow old semantics.  Once all of the callers are fixed,
 872          * we should get rid of these internal flag "adjustment".
 873          */
 874         flags &= ~BUS_DMA_WAITOK;
 875         flags |= BUS_DMA_NOWAIT;
 876
 877         segments = bus_dma_tag_lock(dmat, cache_segments);
 878         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
 879                         segments, dmat->nsegments, &nsegs, flags);
 880         if (error) {
 881                 /* force "no valid mappings" in callback */
 882                 callback(callback_arg, segments, 0,
 883                          0, error);
 884         } else {
 885                 callback(callback_arg, segments, nsegs,
 886                          m0->m_pkthdr.len, error);
 887         }
 888         bus_dma_tag_unlock(dmat);
 889         return error;
 890 }
 891
 892 int
 893 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
 894                              struct mbuf *m0,
 895                              bus_dma_segment_t *segs, int maxsegs,
 896                              int *nsegs, int flags)
 897 {
 898         int error;
 899
 900         M_ASSERTPKTHDR(m0);
 901
 902         KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs));
 903         KASSERT(maxsegs <= dmat->nsegments,
 904                 ("%d too many segments, dmat only supports %d segments",
 905                  maxsegs, dmat->nsegments));
 906         KASSERT(flags & BUS_DMA_NOWAIT,
 907                 ("only BUS_DMA_NOWAIT is supported"));
 908
 909         if (m0->m_pkthdr.len <= dmat->maxsize) {
 910                 int first = 1;
 911                 vm_paddr_t lastaddr = 0;
 912                 struct mbuf *m;
 913
 914                 *nsegs = 1;
 915                 error = 0;
 916                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 917                         if (m->m_len == 0)
 918                                 continue;
 919
 920                         error = _bus_dmamap_load_buffer(dmat, map,
 921                                         m->m_data, m->m_len,
 922                                         segs, maxsegs,
 923                                         NULL, flags, &lastaddr,
 924                                         nsegs, first);
 925                         if (error == ENOMEM && !first) {
 926                                 /*
 927                                  * Out of bounce pages due to too many
 928                                  * fragments in the mbuf chain; return
 929                                  * EFBIG instead.
 930                                  */
 931                                 error = EFBIG;
 932                         }
 933                         first = 0;
 934                 }
 935 #ifdef INVARIANTS
 936                 if (!error)
 937                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
 938 #endif
 939         } else {
 940                 *nsegs = 0;
 941                 error = EINVAL;
 942         }
 943         KKASSERT(error != EINPROGRESS);
 944         return error;
 945 }
 946
 947 /*
 948  * Like _bus_dmamap_load(), but for uios.
 949  */
 950 int
 951 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
 952                     struct uio *uio,
 953                     bus_dmamap_callback2_t *callback, void *callback_arg,
 954                     int flags)
 955 {
 956         vm_paddr_t lastaddr;
 957         int nsegs, error, first, i;
 958         bus_size_t resid;
 959         struct iovec *iov;
 960         pmap_t pmap;
 961         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 962         bus_dma_segment_t *segments;
 963         bus_dma_segment_t *segs;
 964         int nsegs_left;
 965
 966         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 967                 segments = cache_segments;
 968         else
 969                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
 970                                    M_DEVBUF, M_WAITOK | M_ZERO);
 971
 972         /*
 973          * XXX
 974          * Follow old semantics.  Once all of the callers are fixed,
 975          * we should get rid of these internal flag "adjustment".
 976          */
 977         flags &= ~BUS_DMA_WAITOK;
 978         flags |= BUS_DMA_NOWAIT;
 979
 980         resid = (bus_size_t)uio->uio_resid;
 981         iov = uio->uio_iov;
 982
 983         segs = segments;
 984         nsegs_left = dmat->nsegments;
 985
 986         if (uio->uio_segflg == UIO_USERSPACE) {
 987                 struct thread *td;
 988
 989                 td = uio->uio_td;
 990                 KASSERT(td != NULL && td->td_proc != NULL,
 991                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
 992                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
 993         } else {
 994                 pmap = NULL;
 995         }
 996
 997         error = 0;
 998         nsegs = 1;
 999         first = 1;
1000         lastaddr = 0;
1001         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
1002                 /*
1003                  * Now at the first iovec to load.  Load each iovec
1004                  * until we have exhausted the residual count.
1005                  */
1006                 bus_size_t minlen =
1007                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
1008                 caddr_t addr = (caddr_t) iov[i].iov_base;
1009
1010                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
1011                                 segs, nsegs_left,
1012                                 pmap, flags, &lastaddr, &nsegs, first);
1013                 first = 0;
1014
1015                 resid -= minlen;
1016                 if (error == 0) {
1017                         nsegs_left -= nsegs;
1018                         segs += nsegs;
1019                 }
1020         }
1021
1022         /*
1023          * Minimum one DMA segment, even if 0-length buffer.
1024          */
1025         if (nsegs_left == dmat->nsegments)
1026                 --nsegs_left;
1027
1028         if (error) {
1029                 /* force "no valid mappings" in callback */
1030                 callback(callback_arg, segments, 0,
1031                          0, error);
1032         } else {
1033                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1034                          (bus_size_t)uio->uio_resid, error);
1035         }
1036         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1037                 kfree(segments, M_DEVBUF);
1038         return error;
1039 }
1040
1041 /*
1042  * Release the mapping held by map.
1043  */
1044 void
1045 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1046 {
1047         struct bounce_page *bpage;
1048
1049         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1050                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1051                 free_bounce_page(dmat, bpage);
1052         }
1053 }
1054
1055 void
1056 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1057 {
1058         struct bounce_page *bpage;
1059
1060         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1061                 /*
1062                  * Handle data bouncing.  We might also
1063                  * want to add support for invalidating
1064                  * the caches on broken hardware
1065                  */
1066                 switch (op) {
1067                 case BUS_DMASYNC_PREWRITE:
1068                         while (bpage != NULL) {
1069                                 bcopy((void *)bpage->datavaddr,
1070                                       (void *)bpage->vaddr,
1071                                       bpage->datacount);
1072                                 bpage = STAILQ_NEXT(bpage, links);
1073                         }
1074                         dmat->bounce_zone->total_bounced++;
1075                         break;
1076
1077                 case BUS_DMASYNC_POSTREAD:
1078                         while (bpage != NULL) {
1079                                 bcopy((void *)bpage->vaddr,
1080                                       (void *)bpage->datavaddr,
1081                                       bpage->datacount);
1082                                 bpage = STAILQ_NEXT(bpage, links);
1083                         }
1084                         dmat->bounce_zone->total_bounced++;
1085                         break;
1086
1087                 case BUS_DMASYNC_PREREAD:
1088                 case BUS_DMASYNC_POSTWRITE:
1089                         /* No-ops */
1090                         break;
1091                 }
1092         }
1093 }
1094
1095 static int
1096 alloc_bounce_zone(bus_dma_tag_t dmat)
1097 {
1098         struct bounce_zone *bz, *new_bz;
1099
1100         KASSERT(dmat->bounce_zone == NULL,
1101                 ("bounce zone was already assigned"));
1102
1103         new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO);
1104
1105         lwkt_gettoken(&bounce_zone_tok);
1106
1107         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1108                 /*
1109                  * For shared bounce zone, check to see
1110                  * if we already have a suitable zone
1111                  */
1112                 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1113                         if (dmat->alignment <= bz->alignment &&
1114                             dmat->lowaddr >= bz->lowaddr) {
1115                                 lwkt_reltoken(&bounce_zone_tok);
1116
1117                                 dmat->bounce_zone = bz;
1118                                 kfree(new_bz, M_DEVBUF);
1119                                 return 0;
1120                         }
1121                 }
1122         }
1123         bz = new_bz;
1124
1125 #ifdef SMP
1126         spin_init(&bz->spin);
1127 #endif
1128         STAILQ_INIT(&bz->bounce_page_list);
1129         STAILQ_INIT(&bz->bounce_map_waitinglist);
1130         bz->free_bpages = 0;
1131         bz->reserved_bpages = 0;
1132         bz->active_bpages = 0;
1133         bz->lowaddr = dmat->lowaddr;
1134         bz->alignment = round_page(dmat->alignment);
1135         ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1136
1137         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1138                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1139                 busdma_zonecount++;
1140                 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1141         } else {
1142                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_priv_zonecount);
1143                 busdma_priv_zonecount--;
1144         }
1145
1146         lwkt_reltoken(&bounce_zone_tok);
1147
1148         dmat->bounce_zone = bz;
1149
1150         sysctl_ctx_init(&bz->sysctl_ctx);
1151         bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx,
1152             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1153             CTLFLAG_RD, 0, "");
1154         if (bz->sysctl_tree == NULL) {
1155                 sysctl_ctx_free(&bz->sysctl_ctx);
1156                 return 0;       /* XXX error code? */
1157         }
1158
1159         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1160             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1161             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1162             "Total bounce pages");
1163         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1164             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1165             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1166             "Free bounce pages");
1167         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1168             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1169             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1170             "Reserved bounce pages");
1171         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1172             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1173             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1174             "Active bounce pages");
1175         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1176             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1177             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1178             "Total bounce requests");
1179         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1180             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1181             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1182             "Total bounce requests that were deferred");
1183         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1184             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1185             "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0,
1186             "Total bounce page reservations that were failed");
1187         SYSCTL_ADD_STRING(&bz->sysctl_ctx,
1188             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1189             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1190         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1191             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1192             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1193
1194         return 0;
1195 }
1196
1197 static int
1198 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags)
1199 {
1200         struct bounce_zone *bz = dmat->bounce_zone;
1201         int count = 0, mflags;
1202
1203         if (flags & BUS_DMA_NOWAIT)
1204                 mflags = M_NOWAIT;
1205         else
1206                 mflags = M_WAITOK;
1207
1208         while (numpages > 0) {
1209                 struct bounce_page *bpage;
1210
1211                 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO);
1212
1213                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1214                                                          mflags, 0ul,
1215                                                          bz->lowaddr,
1216                                                          bz->alignment, 0);
1217                 if (bpage->vaddr == 0) {
1218                         kfree(bpage, M_DEVBUF);
1219                         break;
1220                 }
1221                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1222
1223                 BZ_LOCK(bz);
1224                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1225                 total_bounce_pages++;
1226                 bz->total_bpages++;
1227                 bz->free_bpages++;
1228                 BZ_UNLOCK(bz);
1229
1230                 count++;
1231                 numpages--;
1232         }
1233         return count;
1234 }
1235
1236 static void
1237 free_bounce_pages_all(bus_dma_tag_t dmat)
1238 {
1239         struct bounce_zone *bz = dmat->bounce_zone;
1240         struct bounce_page *bpage;
1241
1242         BZ_LOCK(bz);
1243
1244         while ((bpage = STAILQ_FIRST(&bz->bounce_page_list)) != NULL) {
1245                 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1246
1247                 KKASSERT(total_bounce_pages > 0);
1248                 total_bounce_pages--;
1249
1250                 KKASSERT(bz->total_bpages > 0);
1251                 bz->total_bpages--;
1252
1253                 KKASSERT(bz->free_bpages > 0);
1254                 bz->free_bpages--;
1255
1256                 BZ_UNLOCK(bz);
1257                 contigfree((void *)bpage->vaddr, PAGE_SIZE, M_DEVBUF);
1258                 kfree(bpage, M_DEVBUF);
1259                 BZ_LOCK(bz);
1260         }
1261         if (bz->total_bpages) {
1262                 kprintf("#%d bounce pages are still in use\n",
1263                         bz->total_bpages);
1264                 print_backtrace(-1);
1265         }
1266
1267         BZ_UNLOCK(bz);
1268 }
1269
1270 static void
1271 free_bounce_zone(bus_dma_tag_t dmat)
1272 {
1273         struct bounce_zone *bz = dmat->bounce_zone;
1274
1275         if (bz == NULL)
1276                 return;
1277
1278         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0)
1279                 return;
1280
1281         free_bounce_pages_all(dmat);
1282         dmat->bounce_zone = NULL;
1283
1284         if (bz->sysctl_tree != NULL)
1285                 sysctl_ctx_free(&bz->sysctl_ctx);
1286         kfree(bz, M_DEVBUF);
1287 }
1288
1289 /* Assume caller holds bounce zone spinlock */
1290 static int
1291 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1292 {
1293         struct bounce_zone *bz = dmat->bounce_zone;
1294         int pages;
1295
1296         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1297         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1298                 bz->reserve_failed++;
1299                 return (map->pagesneeded - (map->pagesreserved + pages));
1300         }
1301
1302         bz->free_bpages -= pages;
1303
1304         bz->reserved_bpages += pages;
1305         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1306
1307         map->pagesreserved += pages;
1308         pages = map->pagesneeded - map->pagesreserved;
1309
1310         return pages;
1311 }
1312
1313 static void
1314 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1315 {
1316         struct bounce_zone *bz = dmat->bounce_zone;
1317         int reserved = map->pagesreserved;
1318         bus_dmamap_t wait_map;
1319
1320         map->pagesreserved = 0;
1321         map->pagesneeded = 0;
1322
1323         if (reserved == 0)
1324                 return;
1325
1326         BZ_LOCK(bz);
1327
1328         bz->free_bpages += reserved;
1329         KKASSERT(bz->free_bpages <= bz->total_bpages);
1330
1331         KKASSERT(bz->reserved_bpages >= reserved);
1332         bz->reserved_bpages -= reserved;
1333
1334         wait_map = get_map_waiting(dmat);
1335
1336         BZ_UNLOCK(bz);
1337
1338         if (wait_map != NULL)
1339                 add_map_callback(map);
1340 }
1341
1342 static bus_addr_t
1343 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1344                 bus_size_t size)
1345 {
1346         struct bounce_zone *bz = dmat->bounce_zone;
1347         struct bounce_page *bpage;
1348
1349         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1350         map->pagesneeded--;
1351
1352         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1353         map->pagesreserved--;
1354
1355         BZ_LOCK(bz);
1356
1357         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1358         KASSERT(bpage != NULL, ("free page list is empty"));
1359         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1360
1361         KKASSERT(bz->reserved_bpages > 0);
1362         bz->reserved_bpages--;
1363
1364         bz->active_bpages++;
1365         KKASSERT(bz->active_bpages <= bz->total_bpages);
1366
1367         BZ_UNLOCK(bz);
1368
1369         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1370                 /* Page offset needs to be preserved. */
1371                 bpage->vaddr |= vaddr & PAGE_MASK;
1372                 bpage->busaddr |= vaddr & PAGE_MASK;
1373         }
1374         bpage->datavaddr = vaddr;
1375         bpage->datacount = size;
1376         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1377         return bpage->busaddr;
1378 }
1379
1380 static void
1381 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1382 {
1383         struct bounce_zone *bz = dmat->bounce_zone;
1384         bus_dmamap_t map;
1385
1386         bpage->datavaddr = 0;
1387         bpage->datacount = 0;
1388
1389         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1390                 /*
1391                  * Reset the bounce page to start at offset 0.  Other uses
1392                  * of this bounce page may need to store a full page of
1393                  * data and/or assume it starts on a page boundary.
1394                  */
1395                 bpage->vaddr &= ~PAGE_MASK;
1396                 bpage->busaddr &= ~PAGE_MASK;
1397         }
1398
1399         BZ_LOCK(bz);
1400
1401         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1402
1403         bz->free_bpages++;
1404         KKASSERT(bz->free_bpages <= bz->total_bpages);
1405
1406         KKASSERT(bz->active_bpages > 0);
1407         bz->active_bpages--;
1408
1409         map = get_map_waiting(dmat);
1410
1411         BZ_UNLOCK(bz);
1412
1413         if (map != NULL)
1414                 add_map_callback(map);
1415 }
1416
1417 /* Assume caller holds bounce zone spinlock */
1418 static bus_dmamap_t
1419 get_map_waiting(bus_dma_tag_t dmat)
1420 {
1421         struct bounce_zone *bz = dmat->bounce_zone;
1422         bus_dmamap_t map;
1423
1424         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1425         if (map != NULL) {
1426                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1427                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1428                         bz->total_deferred++;
1429                 } else {
1430                         map = NULL;
1431                 }
1432         }
1433         return map;
1434 }
1435
1436 static void
1437 add_map_callback(bus_dmamap_t map)
1438 {
1439         spin_lock(&bounce_map_list_spin);
1440         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1441         busdma_swi_pending = 1;
1442         setsoftvm();
1443         spin_unlock(&bounce_map_list_spin);
1444 }
1445
1446 void
1447 busdma_swi(void)
1448 {
1449         bus_dmamap_t map;
1450
1451         spin_lock(&bounce_map_list_spin);
1452         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1453                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1454                 spin_unlock(&bounce_map_list_spin);
1455                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1456                                 map->callback, map->callback_arg, /*flags*/0);
1457                 spin_lock(&bounce_map_list_spin);
1458         }
1459         spin_unlock(&bounce_map_list_spin);
1460 }