sys/platform/pc64/x86_64/busdma_machdep.c

   1 /*
   2  * Copyright (c) 1997, 1998 Justin T. Gibbs.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions, and the following disclaimer,
  10  *    without modification, immediately at the beginning of the file.
  11  * 2. The name of the author may not be used to endorse or promote products
  12  *    derived from this software without specific prior written permission.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  18  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  *
  26  * $FreeBSD: src/sys/i386/i386/busdma_machdep.c,v 1.94 2008/08/15 20:51:31 kmacy Exp $
  27  * $DragonFly: src/sys/platform/pc32/i386/busdma_machdep.c,v 1.23 2008/06/05 18:06:32 swildner Exp $
  28  */
  29
  30 #include <sys/param.h>
  31 #include <sys/systm.h>
  32 #include <sys/malloc.h>
  33 #include <sys/mbuf.h>
  34 #include <sys/uio.h>
  35 #include <sys/bus_dma.h>
  36 #include <sys/kernel.h>
  37 #include <sys/sysctl.h>
  38 #include <sys/lock.h>
  39
  40 #include <sys/thread2.h>
  41 #include <sys/spinlock2.h>
  42 #include <sys/mplock2.h>
  43
  44 #include <vm/vm.h>
  45 #include <vm/vm_page.h>
  46
  47 /* XXX needed for to access pmap to convert per-proc virtual to physical */
  48 #include <sys/proc.h>
  49 #include <sys/lock.h>
  50 #include <vm/vm_map.h>
  51
  52 #include <machine/md_var.h>
  53
  54 #define MAX_BPAGES      1024
  55
  56 /*
  57  * 16 x N declared on stack.
  58  */
  59 #define BUS_DMA_CACHE_SEGMENTS  8
  60
  61 struct bounce_zone;
  62 struct bus_dmamap;
  63
  64 struct bus_dma_tag {
  65         bus_dma_tag_t   parent;
  66         bus_size_t      alignment;
  67         bus_size_t      boundary;
  68         bus_addr_t      lowaddr;
  69         bus_addr_t      highaddr;
  70         bus_dma_filter_t *filter;
  71         void            *filterarg;
  72         bus_size_t      maxsize;
  73         u_int           nsegments;
  74         bus_size_t      maxsegsz;
  75         int             flags;
  76         int             ref_count;
  77         int             map_count;
  78         bus_dma_segment_t *segments;
  79         struct bounce_zone *bounce_zone;
  80 #ifdef SMP
  81         struct spinlock spin;
  82 #else
  83         int             unused0;
  84 #endif
  85 };
  86
  87 /*
  88  * bus_dma_tag private flags
  89  */
  90 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
  91 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
  92 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  93
  94 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
  95
  96 #define BUS_DMAMEM_KMALLOC(dmat) \
  97         ((dmat)->maxsize <= PAGE_SIZE && \
  98          (dmat)->alignment <= PAGE_SIZE && \
  99          (dmat)->lowaddr >= ptoa(Maxmem))
 100
 101 struct bounce_page {
 102         vm_offset_t     vaddr;          /* kva of bounce buffer */
 103         bus_addr_t      busaddr;        /* Physical address */
 104         vm_offset_t     datavaddr;      /* kva of client data */
 105         bus_size_t      datacount;      /* client data count */
 106         STAILQ_ENTRY(bounce_page) links;
 107 };
 108
 109 struct bounce_zone {
 110         STAILQ_ENTRY(bounce_zone) links;
 111         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 112         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 113 #ifdef SMP
 114         struct spinlock spin;
 115 #else
 116         int             unused0;
 117 #endif
 118         int             total_bpages;
 119         int             free_bpages;
 120         int             reserved_bpages;
 121         int             active_bpages;
 122         int             total_bounced;
 123         int             total_deferred;
 124         int             reserve_failed;
 125         bus_size_t      alignment;
 126         bus_addr_t      lowaddr;
 127         char            zoneid[8];
 128         char            lowaddrid[20];
 129         struct sysctl_ctx_list sysctl_ctx;
 130         struct sysctl_oid *sysctl_tree;
 131 };
 132
 133 #ifdef SMP
 134 #define BZ_LOCK(bz)     spin_lock(&(bz)->spin)
 135 #define BZ_UNLOCK(bz)   spin_unlock(&(bz)->spin)
 136 #else
 137 #define BZ_LOCK(bz)     crit_enter()
 138 #define BZ_UNLOCK(bz)   crit_exit()
 139 #endif
 140
 141 static struct lwkt_token bounce_zone_tok =
 142         LWKT_TOKEN_MP_INITIALIZER(bounce_zone_token);
 143 static int busdma_zonecount;
 144 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
 145         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
 146
 147 static int busdma_priv_zonecount = -1;
 148
 149 int busdma_swi_pending;
 150 static int total_bounce_pages;
 151 static int max_bounce_pages = MAX_BPAGES;
 152 static int bounce_alignment = 1; /* XXX temporary */
 153
 154 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
 155 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
 156
 157 struct bus_dmamap {
 158         struct bp_list  bpages;
 159         int             pagesneeded;
 160         int             pagesreserved;
 161         bus_dma_tag_t   dmat;
 162         void            *buf;           /* unmapped buffer pointer */
 163         bus_size_t      buflen;         /* unmapped buffer length */
 164         bus_dmamap_callback_t *callback;
 165         void            *callback_arg;
 166         STAILQ_ENTRY(bus_dmamap) links;
 167 };
 168
 169 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
 170         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
 171 static struct spinlock bounce_map_list_spin =
 172         SPINLOCK_INITIALIZER(&bounce_map_list_spin);
 173
 174 static struct bus_dmamap nobounce_dmamap;
 175
 176 static int              alloc_bounce_zone(bus_dma_tag_t);
 177 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
 178 static void             free_bounce_pages_all(bus_dma_tag_t);
 179 static void             free_bounce_zone(bus_dma_tag_t);
 180 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
 181 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
 182 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
 183                             vm_offset_t, bus_size_t);
 184 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
 185
 186 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
 187 static void             add_map_callback(bus_dmamap_t);
 188
 189 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 190 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
 191            0, "Total bounce pages");
 192 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
 193            0, "Max bounce pages per bounce zone");
 194 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
 195            &bounce_alignment, 0, "Obey alignment constraint");
 196
 197 static __inline int
 198 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 199 {
 200         int retval;
 201
 202         retval = 0;
 203         do {
 204                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
 205                      (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0))
 206                  && (dmat->filter == NULL ||
 207                      dmat->filter(dmat->filterarg, paddr) != 0))
 208                         retval = 1;
 209
 210                 dmat = dmat->parent;
 211         } while (retval == 0 && dmat != NULL);
 212         return (retval);
 213 }
 214
 215 static __inline
 216 bus_dma_segment_t *
 217 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
 218 {
 219         if (tag->flags & BUS_DMA_PROTECTED)
 220                 return(tag->segments);
 221
 222         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 223                 return(cache);
 224 #ifdef SMP
 225         spin_lock(&tag->spin);
 226 #endif
 227         return(tag->segments);
 228 }
 229
 230 static __inline
 231 void
 232 bus_dma_tag_unlock(bus_dma_tag_t tag)
 233 {
 234 #ifdef SMP
 235         if (tag->flags & BUS_DMA_PROTECTED)
 236                 return;
 237
 238         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
 239                 spin_unlock(&tag->spin);
 240 #endif
 241 }
 242
 243 /*
 244  * Allocate a device specific dma_tag.
 245  */
 246 int
 247 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 248                    bus_size_t boundary, bus_addr_t lowaddr,
 249                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 250                    void *filterarg, bus_size_t maxsize, int nsegments,
 251                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
 252 {
 253         bus_dma_tag_t newtag;
 254         int error = 0;
 255
 256         /*
 257          * Sanity checks
 258          */
 259
 260         if (alignment == 0)
 261                 alignment = 1;
 262         if (alignment & (alignment - 1))
 263                 panic("alignment must be power of 2\n");
 264
 265         if (boundary != 0) {
 266                 if (boundary & (boundary - 1))
 267                         panic("boundary must be power of 2\n");
 268                 if (boundary < maxsegsz) {
 269                         kprintf("boundary < maxsegsz:\n");
 270                         print_backtrace(-1);
 271                         maxsegsz = boundary;
 272                 }
 273         }
 274
 275         /* Return a NULL tag on failure */
 276         *dmat = NULL;
 277
 278         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
 279
 280 #ifdef SMP
 281         spin_init(&newtag->spin);
 282 #endif
 283         newtag->parent = parent;
 284         newtag->alignment = alignment;
 285         newtag->boundary = boundary;
 286         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
 287         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
 288         newtag->filter = filter;
 289         newtag->filterarg = filterarg;
 290         newtag->maxsize = maxsize;
 291         newtag->nsegments = nsegments;
 292         newtag->maxsegsz = maxsegsz;
 293         newtag->flags = flags;
 294         newtag->ref_count = 1; /* Count ourself */
 295         newtag->map_count = 0;
 296         newtag->segments = NULL;
 297         newtag->bounce_zone = NULL;
 298
 299         /* Take into account any restrictions imposed by our parent tag */
 300         if (parent != NULL) {
 301                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
 302                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 303
 304                 if (newtag->boundary == 0) {
 305                         newtag->boundary = parent->boundary;
 306                 } else if (parent->boundary != 0) {
 307                         newtag->boundary = MIN(parent->boundary,
 308                                                newtag->boundary);
 309                 }
 310
 311 #ifdef notyet
 312                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
 313 #endif
 314
 315                 if (newtag->filter == NULL) {
 316                         /*
 317                          * Short circuit looking at our parent directly
 318                          * since we have encapsulated all of its information
 319                          */
 320                         newtag->filter = parent->filter;
 321                         newtag->filterarg = parent->filterarg;
 322                         newtag->parent = parent->parent;
 323                 }
 324                 if (newtag->parent != NULL)
 325                         parent->ref_count++;
 326         }
 327
 328         if (newtag->lowaddr < ptoa(Maxmem))
 329                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
 330         if (bounce_alignment && newtag->alignment > 1 &&
 331             !(newtag->flags & BUS_DMA_ALIGNED))
 332                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
 333
 334         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
 335             (flags & BUS_DMA_ALLOCNOW) != 0) {
 336                 struct bounce_zone *bz;
 337
 338                 /* Must bounce */
 339
 340                 error = alloc_bounce_zone(newtag);
 341                 if (error)
 342                         goto back;
 343                 bz = newtag->bounce_zone;
 344
 345                 if ((newtag->flags & BUS_DMA_ALLOCALL) == 0 &&
 346                     ptoa(bz->total_bpages) < maxsize) {
 347                         int pages;
 348
 349                         if (flags & BUS_DMA_ONEBPAGE) {
 350                                 pages = 1;
 351                         } else {
 352                                 pages = atop(round_page(maxsize)) -
 353                                         bz->total_bpages;
 354                                 pages = MAX(pages, 1);
 355                         }
 356
 357                         /* Add pages to our bounce pool */
 358                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
 359                                 error = ENOMEM;
 360
 361                         /* Performed initial allocation */
 362                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 363                 }
 364         }
 365 back:
 366         if (error)
 367                 kfree(newtag, M_DEVBUF);
 368         else
 369                 *dmat = newtag;
 370         return error;
 371 }
 372
 373 int
 374 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 375 {
 376         if (dmat != NULL) {
 377                 if (dmat->map_count != 0)
 378                         return (EBUSY);
 379
 380                 while (dmat != NULL) {
 381                         bus_dma_tag_t parent;
 382
 383                         parent = dmat->parent;
 384                         dmat->ref_count--;
 385                         if (dmat->ref_count == 0) {
 386                                 free_bounce_zone(dmat);
 387                                 if (dmat->segments != NULL)
 388                                         kfree(dmat->segments, M_DEVBUF);
 389                                 kfree(dmat, M_DEVBUF);
 390                                 /*
 391                                  * Last reference count, so
 392                                  * release our reference
 393                                  * count on our parent.
 394                                  */
 395                                 dmat = parent;
 396                         } else
 397                                 dmat = NULL;
 398                 }
 399         }
 400         return (0);
 401 }
 402
 403 bus_size_t
 404 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
 405 {
 406         return(tag->maxsize);
 407 }
 408
 409 /*
 410  * Allocate a handle for mapping from kva/uva/physical
 411  * address space into bus device space.
 412  */
 413 int
 414 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 415 {
 416         int error;
 417
 418         error = 0;
 419
 420         if (dmat->segments == NULL) {
 421                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
 422                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 423                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 424         }
 425
 426         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 427                 struct bounce_zone *bz;
 428                 int maxpages;
 429
 430                 /* Must bounce */
 431
 432                 if (dmat->bounce_zone == NULL) {
 433                         error = alloc_bounce_zone(dmat);
 434                         if (error)
 435                                 return error;
 436                 }
 437                 bz = dmat->bounce_zone;
 438
 439                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
 440
 441                 /* Initialize the new map */
 442                 STAILQ_INIT(&((*mapp)->bpages));
 443
 444                 /*
 445                  * Attempt to add pages to our pool on a per-instance
 446                  * basis up to a sane limit.
 447                  */
 448                 if (dmat->flags & BUS_DMA_ALLOCALL) {
 449                         maxpages = Maxmem - atop(dmat->lowaddr);
 450                 } else if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
 451                         maxpages = max_bounce_pages;
 452                 } else {
 453                         maxpages = MIN(max_bounce_pages,
 454                                        Maxmem - atop(dmat->lowaddr));
 455                 }
 456                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 ||
 457                     (dmat->map_count > 0 && bz->total_bpages < maxpages)) {
 458                         int pages;
 459
 460                         if (flags & BUS_DMA_ONEBPAGE) {
 461                                 pages = 1;
 462                         } else {
 463                                 pages = atop(round_page(dmat->maxsize));
 464                                 pages = MIN(maxpages - bz->total_bpages, pages);
 465                                 pages = MAX(pages, 1);
 466                         }
 467                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
 468                                 error = ENOMEM;
 469
 470                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 471                                 if (!error &&
 472                                     (dmat->flags & BUS_DMA_ALLOCALL) == 0)
 473                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 474                         } else {
 475                                 error = 0;
 476                         }
 477                 }
 478         } else {
 479                 *mapp = NULL;
 480         }
 481         if (!error)
 482                 dmat->map_count++;
 483         return error;
 484 }
 485
 486 /*
 487  * Destroy a handle for mapping from kva/uva/physical
 488  * address space into bus device space.
 489  */
 490 int
 491 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 492 {
 493         if (map != NULL) {
 494                 if (STAILQ_FIRST(&map->bpages) != NULL)
 495                         return (EBUSY);
 496                 kfree(map, M_DEVBUF);
 497         }
 498         dmat->map_count--;
 499         return (0);
 500 }
 501
 502 static __inline bus_size_t
 503 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
 504 {
 505         bus_size_t maxsize = 0;
 506         uintptr_t vaddr = (uintptr_t)vaddr0;
 507
 508         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
 509                 if (verify || bootverbose)
 510                         kprintf("boundary check failed\n");
 511                 if (verify)
 512                         print_backtrace(-1); /* XXX panic */
 513                 maxsize = dmat->maxsize;
 514         }
 515         if (vaddr & (dmat->alignment - 1)) {
 516                 if (verify || bootverbose)
 517                         kprintf("alignment check failed\n");
 518                 if (verify)
 519                         print_backtrace(-1); /* XXX panic */
 520                 if (dmat->maxsize < dmat->alignment)
 521                         maxsize = dmat->alignment;
 522                 else
 523                         maxsize = dmat->maxsize;
 524         }
 525         return maxsize;
 526 }
 527
 528 /*
 529  * Allocate a piece of memory that can be efficiently mapped into
 530  * bus device space based on the constraints lited in the dma tag.
 531  *
 532  * mapp is degenerate.  By definition this allocation should not require
 533  * bounce buffers so do not allocate a dma map.
 534  */
 535 int
 536 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
 537                  bus_dmamap_t *mapp)
 538 {
 539         int mflags;
 540
 541         /* If we succeed, no mapping/bouncing will be required */
 542         *mapp = NULL;
 543
 544         if (dmat->segments == NULL) {
 545                 KKASSERT(dmat->nsegments < 16384);
 546                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 547                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 548         }
 549
 550         if (flags & BUS_DMA_NOWAIT)
 551                 mflags = M_NOWAIT;
 552         else
 553                 mflags = M_WAITOK;
 554         if (flags & BUS_DMA_ZERO)
 555                 mflags |= M_ZERO;
 556
 557         if (BUS_DMAMEM_KMALLOC(dmat)) {
 558                 bus_size_t maxsize;
 559
 560                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 561
 562                 /*
 563                  * XXX
 564                  * Check whether the allocation
 565                  * - crossed a page boundary
 566                  * - was not aligned
 567                  * Retry with power-of-2 alignment in the above cases.
 568                  */
 569                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 570                 if (maxsize) {
 571                         size_t size;
 572
 573                         kfree(*vaddr, M_DEVBUF);
 574                         /* XXX check for overflow? */
 575                         for (size = 1; size <= maxsize; size <<= 1)
 576                                 ;
 577                         *vaddr = kmalloc(size, 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\n");
 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 - ((vm_offset_t)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\n");
 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\n", maxsegs));
 903         KASSERT(maxsegs <= dmat->nsegments,
 904                 ("%d too many segments, dmat only support %d segments\n",
 905                  maxsegs, dmat->nsegments));
 906         KASSERT(flags & BUS_DMA_NOWAIT,
 907                 ("only BUS_DMA_NOWAIT is supported\n"));
 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\n"));
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                 contigfree((void *)bpage->vaddr, PAGE_SIZE, M_DEVBUF);
1257                 kfree(bpage, M_DEVBUF);
1258         }
1259         if (bz->total_bpages) {
1260                 kprintf("#%d bounce pages are still in use\n",
1261                         bz->total_bpages);
1262                 print_backtrace(-1);
1263         }
1264
1265         BZ_UNLOCK(bz);
1266 }
1267
1268 static void
1269 free_bounce_zone(bus_dma_tag_t dmat)
1270 {
1271         struct bounce_zone *bz = dmat->bounce_zone;
1272
1273         if (bz == NULL)
1274                 return;
1275
1276         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0)
1277                 return;
1278
1279         free_bounce_pages_all(dmat);
1280         dmat->bounce_zone = NULL;
1281
1282         sysctl_ctx_free(&bz->sysctl_ctx);
1283         kfree(bz, M_DEVBUF);
1284 }
1285
1286 /* Assume caller holds bounce zone spinlock */
1287 static int
1288 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1289 {
1290         struct bounce_zone *bz = dmat->bounce_zone;
1291         int pages;
1292
1293         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1294         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1295                 bz->reserve_failed++;
1296                 return (map->pagesneeded - (map->pagesreserved + pages));
1297         }
1298
1299         bz->free_bpages -= pages;
1300
1301         bz->reserved_bpages += pages;
1302         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1303
1304         map->pagesreserved += pages;
1305         pages = map->pagesneeded - map->pagesreserved;
1306
1307         return pages;
1308 }
1309
1310 static void
1311 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1312 {
1313         struct bounce_zone *bz = dmat->bounce_zone;
1314         int reserved = map->pagesreserved;
1315         bus_dmamap_t wait_map;
1316
1317         map->pagesreserved = 0;
1318         map->pagesneeded = 0;
1319
1320         if (reserved == 0)
1321                 return;
1322
1323         BZ_LOCK(bz);
1324
1325         bz->free_bpages += reserved;
1326         KKASSERT(bz->free_bpages <= bz->total_bpages);
1327
1328         KKASSERT(bz->reserved_bpages >= reserved);
1329         bz->reserved_bpages -= reserved;
1330
1331         wait_map = get_map_waiting(dmat);
1332
1333         BZ_UNLOCK(bz);
1334
1335         if (wait_map != NULL)
1336                 add_map_callback(map);
1337 }
1338
1339 static bus_addr_t
1340 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1341                 bus_size_t size)
1342 {
1343         struct bounce_zone *bz = dmat->bounce_zone;
1344         struct bounce_page *bpage;
1345
1346         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1347         map->pagesneeded--;
1348
1349         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1350         map->pagesreserved--;
1351
1352         BZ_LOCK(bz);
1353
1354         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1355         KASSERT(bpage != NULL, ("free page list is empty"));
1356         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1357
1358         KKASSERT(bz->reserved_bpages > 0);
1359         bz->reserved_bpages--;
1360
1361         bz->active_bpages++;
1362         KKASSERT(bz->active_bpages <= bz->total_bpages);
1363
1364         BZ_UNLOCK(bz);
1365
1366         bpage->datavaddr = vaddr;
1367         bpage->datacount = size;
1368         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1369         return bpage->busaddr;
1370 }
1371
1372 static void
1373 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1374 {
1375         struct bounce_zone *bz = dmat->bounce_zone;
1376         bus_dmamap_t map;
1377
1378         bpage->datavaddr = 0;
1379         bpage->datacount = 0;
1380
1381         BZ_LOCK(bz);
1382
1383         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1384
1385         bz->free_bpages++;
1386         KKASSERT(bz->free_bpages <= bz->total_bpages);
1387
1388         KKASSERT(bz->active_bpages > 0);
1389         bz->active_bpages--;
1390
1391         map = get_map_waiting(dmat);
1392
1393         BZ_UNLOCK(bz);
1394
1395         if (map != NULL)
1396                 add_map_callback(map);
1397 }
1398
1399 /* Assume caller holds bounce zone spinlock */
1400 static bus_dmamap_t
1401 get_map_waiting(bus_dma_tag_t dmat)
1402 {
1403         struct bounce_zone *bz = dmat->bounce_zone;
1404         bus_dmamap_t map;
1405
1406         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1407         if (map != NULL) {
1408                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1409                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1410                         bz->total_deferred++;
1411                 } else {
1412                         map = NULL;
1413                 }
1414         }
1415         return map;
1416 }
1417
1418 static void
1419 add_map_callback(bus_dmamap_t map)
1420 {
1421         spin_lock(&bounce_map_list_spin);
1422         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1423         busdma_swi_pending = 1;
1424         setsoftvm();
1425         spin_unlock(&bounce_map_list_spin);
1426 }
1427
1428 void
1429 busdma_swi(void)
1430 {
1431         bus_dmamap_t map;
1432
1433         spin_lock(&bounce_map_list_spin);
1434         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1435                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1436                 spin_unlock(&bounce_map_list_spin);
1437                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1438                                 map->callback, map->callback_arg, /*flags*/0);
1439                 spin_lock(&bounce_map_list_spin);
1440         }
1441         spin_unlock(&bounce_map_list_spin);
1442 }