sys/platform/pc64/x86_64/busdma_machdep.c

   1 /*
   2  * Copyright (c) 1997, 1998 Justin T. Gibbs.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions, and the following disclaimer,
  10  *    without modification, immediately at the beginning of the file.
  11  * 2. The name of the author may not be used to endorse or promote products
  12  *    derived from this software without specific prior written permission.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  18  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  *
  26  * $FreeBSD: src/sys/i386/i386/busdma_machdep.c,v 1.94 2008/08/15 20:51:31 kmacy Exp $
  27  */
  28
  29 #include <sys/param.h>
  30 #include <sys/systm.h>
  31 #include <sys/malloc.h>
  32 #include <sys/mbuf.h>
  33 #include <sys/uio.h>
  34 #include <sys/bus_dma.h>
  35 #include <sys/kernel.h>
  36 #include <sys/sysctl.h>
  37 #include <sys/lock.h>
  38
  39 #include <sys/spinlock2.h>
  40 #include <sys/mplock2.h>
  41
  42 #include <vm/vm.h>
  43 #include <vm/vm_page.h>
  44
  45 /* XXX needed for to access pmap to convert per-proc virtual to physical */
  46 #include <sys/proc.h>
  47 #include <vm/vm_map.h>
  48
  49 #include <machine/md_var.h>
  50 #include <machine/pmap.h>
  51
  52 #include <bus/cam/cam.h>
  53 #include <bus/cam/cam_ccb.h>
  54
  55 #define MAX_BPAGES      1024
  56
  57 /*
  58  * 16 x N declared on stack.
  59  */
  60 #define BUS_DMA_CACHE_SEGMENTS  8
  61
  62 struct bounce_zone;
  63 struct bus_dmamap;
  64
  65 struct bus_dma_tag {
  66         bus_dma_tag_t   parent;
  67         bus_size_t      alignment;
  68         bus_size_t      boundary;
  69         bus_addr_t      lowaddr;
  70         bus_addr_t      highaddr;
  71         bus_dma_filter_t *filter;
  72         void            *filterarg;
  73         bus_size_t      maxsize;
  74         u_int           nsegments;
  75         bus_size_t      maxsegsz;
  76         int             flags;
  77         int             ref_count;
  78         int             map_count;
  79         bus_dma_segment_t *segments;
  80         struct bounce_zone *bounce_zone;
  81         struct spinlock spin;
  82 };
  83
  84 /*
  85  * bus_dma_tag private flags
  86  */
  87 #define BUS_DMA_BOUNCE_ALIGN    BUS_DMA_BUS2
  88 #define BUS_DMA_BOUNCE_LOWADDR  BUS_DMA_BUS3
  89 #define BUS_DMA_MIN_ALLOC_COMP  BUS_DMA_BUS4
  90
  91 #define BUS_DMA_COULD_BOUNCE    (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN)
  92
  93 #define BUS_DMAMEM_KMALLOC(dmat) \
  94         ((dmat)->maxsize <= PAGE_SIZE && \
  95          (dmat)->alignment <= PAGE_SIZE && \
  96          (dmat)->lowaddr >= ptoa(Maxmem))
  97
  98 struct bounce_page {
  99         vm_offset_t     vaddr;          /* kva of bounce buffer */
 100         bus_addr_t      busaddr;        /* Physical address */
 101         vm_offset_t     datavaddr;      /* kva of client data */
 102         bus_size_t      datacount;      /* client data count */
 103         STAILQ_ENTRY(bounce_page) links;
 104 };
 105
 106 struct bounce_zone {
 107         STAILQ_ENTRY(bounce_zone) links;
 108         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
 109         STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 110         struct spinlock spin;
 111         int             total_bpages;
 112         int             free_bpages;
 113         int             reserved_bpages;
 114         int             active_bpages;
 115         int             total_bounced;
 116         int             total_deferred;
 117         int             reserve_failed;
 118         bus_size_t      alignment;
 119         bus_addr_t      lowaddr;
 120         char            zoneid[8];
 121         char            lowaddrid[20];
 122         struct sysctl_ctx_list sysctl_ctx;
 123         struct sysctl_oid *sysctl_tree;
 124 };
 125
 126 #define BZ_LOCK(bz)     spin_lock(&(bz)->spin)
 127 #define BZ_UNLOCK(bz)   spin_unlock(&(bz)->spin)
 128
 129 static struct lwkt_token bounce_zone_tok =
 130         LWKT_TOKEN_INITIALIZER(bounce_zone_token);
 131 static int busdma_zonecount;
 132 static STAILQ_HEAD(, bounce_zone) bounce_zone_list =
 133         STAILQ_HEAD_INITIALIZER(bounce_zone_list);
 134
 135 static int busdma_priv_zonecount = -1;
 136
 137 int busdma_swi_pending;
 138 static int total_bounce_pages;
 139 static int max_bounce_pages = MAX_BPAGES;
 140 static int bounce_alignment = 1; /* XXX temporary */
 141
 142 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages);
 143 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment);
 144
 145 struct bus_dmamap {
 146         struct bp_list  bpages;
 147         int             pagesneeded;
 148         int             pagesreserved;
 149         bus_dma_tag_t   dmat;
 150         void            *buf;           /* unmapped buffer pointer */
 151         bus_size_t      buflen;         /* unmapped buffer length */
 152         bus_dmamap_callback_t *callback;
 153         void            *callback_arg;
 154         STAILQ_ENTRY(bus_dmamap) links;
 155 };
 156
 157 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist =
 158         STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist);
 159 static struct spinlock bounce_map_list_spin =
 160         SPINLOCK_INITIALIZER(&bounce_map_list_spin, "bounce_map_list_spin");
 161
 162 static struct bus_dmamap nobounce_dmamap;
 163
 164 static int              alloc_bounce_zone(bus_dma_tag_t);
 165 static int              alloc_bounce_pages(bus_dma_tag_t, u_int, int);
 166 static void             free_bounce_pages_all(bus_dma_tag_t);
 167 static void             free_bounce_zone(bus_dma_tag_t);
 168 static int              reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int);
 169 static void             return_bounce_pages(bus_dma_tag_t, bus_dmamap_t);
 170 static bus_addr_t       add_bounce_page(bus_dma_tag_t, bus_dmamap_t,
 171                             vm_offset_t, bus_size_t *);
 172 static void             free_bounce_page(bus_dma_tag_t, struct bounce_page *);
 173
 174 static bus_dmamap_t     get_map_waiting(bus_dma_tag_t);
 175 static void             add_map_callback(bus_dmamap_t);
 176
 177 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 178 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages,
 179            0, "Total bounce pages");
 180 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages,
 181            0, "Max bounce pages per bounce zone");
 182 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD,
 183            &bounce_alignment, 0, "Obey alignment constraint");
 184
 185 static __inline int
 186 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr)
 187 {
 188         int retval;
 189
 190         retval = 0;
 191         do {
 192                 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) ||
 193                      (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0))
 194                  && (dmat->filter == NULL ||
 195                      dmat->filter(dmat->filterarg, paddr) != 0))
 196                         retval = 1;
 197
 198                 dmat = dmat->parent;
 199         } while (retval == 0 && dmat != NULL);
 200         return (retval);
 201 }
 202
 203 static __inline
 204 bus_dma_segment_t *
 205 bus_dma_tag_lock(bus_dma_tag_t tag, bus_dma_segment_t *cache)
 206 {
 207         if (tag->flags & BUS_DMA_PROTECTED)
 208                 return(tag->segments);
 209
 210         if (tag->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 211                 return(cache);
 212         spin_lock(&tag->spin);
 213         return(tag->segments);
 214 }
 215
 216 static __inline
 217 void
 218 bus_dma_tag_unlock(bus_dma_tag_t tag)
 219 {
 220         if (tag->flags & BUS_DMA_PROTECTED)
 221                 return;
 222
 223         if (tag->nsegments > BUS_DMA_CACHE_SEGMENTS)
 224                 spin_unlock(&tag->spin);
 225 }
 226
 227 /*
 228  * Allocate a device specific dma_tag.
 229  */
 230 int
 231 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 232                    bus_size_t boundary, bus_addr_t lowaddr,
 233                    bus_addr_t highaddr, bus_dma_filter_t *filter,
 234                    void *filterarg, bus_size_t maxsize, int nsegments,
 235                    bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat)
 236 {
 237         bus_dma_tag_t newtag;
 238         int error = 0;
 239
 240         /*
 241          * Sanity checks
 242          */
 243
 244         if (alignment == 0)
 245                 alignment = 1;
 246         if (alignment & (alignment - 1))
 247                 panic("alignment must be power of 2");
 248
 249         if (boundary != 0) {
 250                 if (boundary & (boundary - 1))
 251                         panic("boundary must be power of 2");
 252                 if (boundary < maxsegsz) {
 253                         kprintf("boundary < maxsegsz:\n");
 254                         print_backtrace(-1);
 255                         maxsegsz = boundary;
 256                 }
 257         }
 258
 259         /* Return a NULL tag on failure */
 260         *dmat = NULL;
 261
 262         newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT | M_ZERO);
 263
 264         spin_init(&newtag->spin, "busdmacreate");
 265         newtag->parent = parent;
 266         newtag->alignment = alignment;
 267         newtag->boundary = boundary;
 268         newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1);
 269         newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1);
 270         newtag->filter = filter;
 271         newtag->filterarg = filterarg;
 272         newtag->maxsize = maxsize;
 273         newtag->nsegments = nsegments;
 274         newtag->maxsegsz = maxsegsz;
 275         newtag->flags = flags;
 276         newtag->ref_count = 1; /* Count ourself */
 277         newtag->map_count = 0;
 278         newtag->segments = NULL;
 279         newtag->bounce_zone = NULL;
 280
 281         /* Take into account any restrictions imposed by our parent tag */
 282         if (parent != NULL) {
 283                 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr);
 284                 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr);
 285
 286                 if (newtag->boundary == 0) {
 287                         newtag->boundary = parent->boundary;
 288                 } else if (parent->boundary != 0) {
 289                         newtag->boundary = MIN(parent->boundary,
 290                                                newtag->boundary);
 291                 }
 292
 293 #ifdef notyet
 294                 newtag->alignment = MAX(parent->alignment, newtag->alignment);
 295 #endif
 296
 297                 if (newtag->filter == NULL) {
 298                         /*
 299                          * Short circuit looking at our parent directly
 300                          * since we have encapsulated all of its information
 301                          */
 302                         newtag->filter = parent->filter;
 303                         newtag->filterarg = parent->filterarg;
 304                         newtag->parent = parent->parent;
 305                 }
 306                 if (newtag->parent != NULL)
 307                         parent->ref_count++;
 308         }
 309
 310         if (newtag->lowaddr < ptoa(Maxmem))
 311                 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR;
 312         if (bounce_alignment && newtag->alignment > 1 &&
 313             !(newtag->flags & BUS_DMA_ALIGNED))
 314                 newtag->flags |= BUS_DMA_BOUNCE_ALIGN;
 315
 316         if ((newtag->flags & BUS_DMA_COULD_BOUNCE) &&
 317             (flags & BUS_DMA_ALLOCNOW) != 0) {
 318                 struct bounce_zone *bz;
 319
 320                 /* Must bounce */
 321
 322                 error = alloc_bounce_zone(newtag);
 323                 if (error)
 324                         goto back;
 325                 bz = newtag->bounce_zone;
 326
 327                 if ((newtag->flags & BUS_DMA_ALLOCALL) == 0 &&
 328                     ptoa(bz->total_bpages) < maxsize) {
 329                         int pages;
 330
 331                         if (flags & BUS_DMA_ONEBPAGE) {
 332                                 pages = 1;
 333                         } else {
 334                                 pages = atop(round_page(maxsize)) -
 335                                         bz->total_bpages;
 336                                 pages = MAX(pages, 1);
 337                         }
 338
 339                         /* Add pages to our bounce pool */
 340                         if (alloc_bounce_pages(newtag, pages, flags) < pages)
 341                                 error = ENOMEM;
 342
 343                         /* Performed initial allocation */
 344                         newtag->flags |= BUS_DMA_MIN_ALLOC_COMP;
 345                 }
 346         }
 347 back:
 348         if (error) {
 349                 free_bounce_zone(newtag);
 350                 kfree(newtag, M_DEVBUF);
 351         } else {
 352                 *dmat = newtag;
 353         }
 354         return error;
 355 }
 356
 357 int
 358 bus_dma_tag_destroy(bus_dma_tag_t dmat)
 359 {
 360         if (dmat != NULL) {
 361                 if (dmat->map_count != 0)
 362                         return (EBUSY);
 363
 364                 while (dmat != NULL) {
 365                         bus_dma_tag_t parent;
 366
 367                         parent = dmat->parent;
 368                         dmat->ref_count--;
 369                         if (dmat->ref_count == 0) {
 370                                 free_bounce_zone(dmat);
 371                                 if (dmat->segments != NULL)
 372                                         kfree(dmat->segments, M_DEVBUF);
 373                                 kfree(dmat, M_DEVBUF);
 374                                 /*
 375                                  * Last reference count, so
 376                                  * release our reference
 377                                  * count on our parent.
 378                                  */
 379                                 dmat = parent;
 380                         } else
 381                                 dmat = NULL;
 382                 }
 383         }
 384         return (0);
 385 }
 386
 387 bus_size_t
 388 bus_dma_tag_getmaxsize(bus_dma_tag_t tag)
 389 {
 390         return(tag->maxsize);
 391 }
 392
 393 /*
 394  * Allocate a handle for mapping from kva/uva/physical
 395  * address space into bus device space.
 396  */
 397 int
 398 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 399 {
 400         int error;
 401
 402         error = 0;
 403
 404         if (dmat->segments == NULL) {
 405                 KKASSERT(dmat->nsegments && dmat->nsegments < 16384);
 406                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 407                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 408         }
 409
 410         if (dmat->flags & BUS_DMA_COULD_BOUNCE) {
 411                 struct bounce_zone *bz;
 412                 int maxpages;
 413
 414                 /* Must bounce */
 415
 416                 if (dmat->bounce_zone == NULL) {
 417                         error = alloc_bounce_zone(dmat);
 418                         if (error)
 419                                 return error;
 420                 }
 421                 bz = dmat->bounce_zone;
 422
 423                 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO);
 424
 425                 /* Initialize the new map */
 426                 STAILQ_INIT(&((*mapp)->bpages));
 427
 428                 /*
 429                  * Attempt to add pages to our pool on a per-instance
 430                  * basis up to a sane limit.
 431                  */
 432                 if (dmat->flags & BUS_DMA_ALLOCALL) {
 433                         maxpages = Maxmem - atop(dmat->lowaddr);
 434                 } else if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) {
 435                         maxpages = max_bounce_pages;
 436                 } else {
 437                         maxpages = MIN(max_bounce_pages,
 438                                        Maxmem - atop(dmat->lowaddr));
 439                 }
 440                 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 ||
 441                     (dmat->map_count > 0 && bz->total_bpages < maxpages)) {
 442                         int pages;
 443
 444                         if (flags & BUS_DMA_ONEBPAGE) {
 445                                 pages = 1;
 446                         } else {
 447                                 pages = atop(round_page(dmat->maxsize));
 448                                 pages = MIN(maxpages - bz->total_bpages, pages);
 449                                 pages = MAX(pages, 1);
 450                         }
 451                         if (alloc_bounce_pages(dmat, pages, flags) < pages)
 452                                 error = ENOMEM;
 453
 454                         if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) {
 455                                 if (!error &&
 456                                     (dmat->flags & BUS_DMA_ALLOCALL) == 0)
 457                                         dmat->flags |= BUS_DMA_MIN_ALLOC_COMP;
 458                         } else {
 459                                 error = 0;
 460                         }
 461                 }
 462         } else {
 463                 *mapp = NULL;
 464         }
 465         if (!error) {
 466                 dmat->map_count++;
 467         } else {
 468                 kfree(*mapp, M_DEVBUF);
 469                 *mapp = NULL;
 470         }
 471         return error;
 472 }
 473
 474 /*
 475  * Destroy a handle for mapping from kva/uva/physical
 476  * address space into bus device space.
 477  */
 478 int
 479 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 480 {
 481         if (map != NULL && map != (void *)-1) {
 482                 if (STAILQ_FIRST(&map->bpages) != NULL)
 483                         return (EBUSY);
 484                 kfree(map, M_DEVBUF);
 485         }
 486         dmat->map_count--;
 487         return (0);
 488 }
 489
 490 static __inline bus_size_t
 491 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify)
 492 {
 493         bus_size_t maxsize = 0;
 494         uintptr_t vaddr = (uintptr_t)vaddr0;
 495
 496         if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) {
 497                 if (verify)
 498                         panic("boundary check failed\n");
 499                 maxsize = dmat->maxsize;
 500         }
 501         if (vaddr & (dmat->alignment - 1)) {
 502                 if (verify)
 503                         panic("alignment check failed\n");
 504                 if (dmat->maxsize < dmat->alignment)
 505                         maxsize = dmat->alignment;
 506                 else
 507                         maxsize = dmat->maxsize;
 508         }
 509         return maxsize;
 510 }
 511
 512 /*
 513  * Allocate a piece of memory that can be efficiently mapped into
 514  * bus device space based on the constraints lited in the dma tag.
 515  *
 516  * Use *mapp to record whether we were able to use kmalloc()
 517  * or whether we had to use contigmalloc().
 518  */
 519 int
 520 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags,
 521                  bus_dmamap_t *mapp)
 522 {
 523         vm_memattr_t attr;
 524         int mflags;
 525
 526         /* If we succeed, no mapping/bouncing will be required */
 527         *mapp = NULL;
 528
 529         if (dmat->segments == NULL) {
 530                 KKASSERT(dmat->nsegments < 16384);
 531                 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) *
 532                                         dmat->nsegments, M_DEVBUF, M_INTWAIT);
 533         }
 534
 535         if (flags & BUS_DMA_NOWAIT)
 536                 mflags = M_NOWAIT;
 537         else
 538                 mflags = M_WAITOK;
 539         if (flags & BUS_DMA_ZERO)
 540                 mflags |= M_ZERO;
 541         if (flags & BUS_DMA_NOCACHE)
 542                 attr = VM_MEMATTR_UNCACHEABLE;
 543         else
 544                 attr = VM_MEMATTR_DEFAULT;
 545
 546         /* XXX must alloc with correct mem attribute here */
 547         if (BUS_DMAMEM_KMALLOC(dmat) && attr == VM_MEMATTR_DEFAULT) {
 548                 bus_size_t maxsize;
 549
 550                 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags);
 551
 552                 /*
 553                  * XXX
 554                  * Check whether the allocation
 555                  * - crossed a page boundary
 556                  * - was not aligned
 557                  * Retry with power-of-2 alignment in the above cases.
 558                  */
 559                 maxsize = check_kmalloc(dmat, *vaddr, 0);
 560                 if (maxsize) {
 561                         kfree(*vaddr, M_DEVBUF);
 562                         *vaddr = kmalloc(maxsize, M_DEVBUF,
 563                                          mflags | M_POWEROF2);
 564                         check_kmalloc(dmat, *vaddr, 1);
 565                 }
 566         } else {
 567                 /*
 568                  * XXX Use Contigmalloc until it is merged into this facility
 569                  *     and handles multi-seg allocations.  Nobody is doing
 570                  *     multi-seg allocations yet though.
 571                  */
 572                 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags,
 573                                       0ul, dmat->lowaddr,
 574                                       dmat->alignment, dmat->boundary);
 575                 *mapp = (void  *)-1;
 576         }
 577         if (*vaddr == NULL)
 578                 return (ENOMEM);
 579
 580         if (attr != VM_MEMATTR_DEFAULT) {
 581                 pmap_change_attr((vm_offset_t)(*vaddr),
 582                                  dmat->maxsize / PAGE_SIZE, attr);
 583         }
 584         return (0);
 585 }
 586
 587 /*
 588  * Free a piece of memory and it's allociated dmamap, that was allocated
 589  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 590  */
 591 void
 592 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 593 {
 594         /*
 595          * dmamem does not need to be bounced, so the map should be
 596          * NULL
 597          */
 598         if (map != NULL && map != (void *)-1)
 599                 panic("bus_dmamem_free: Invalid map freed");
 600         if (map == NULL)
 601                 kfree(vaddr, M_DEVBUF);
 602         else
 603                 contigfree(vaddr, dmat->maxsize, M_DEVBUF);
 604 }
 605
 606 static __inline vm_paddr_t
 607 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr)
 608 {
 609         if (pmap)
 610                 return pmap_extract(pmap, vaddr, NULL);
 611         else
 612                 return pmap_kextract(vaddr);
 613 }
 614
 615 /*
 616  * Utility function to load a linear buffer.  lastaddrp holds state
 617  * between invocations (for multiple-buffer loads).  segp contains
 618  * the segment following the starting one on entrace, and the ending
 619  * segment on exit.  first indicates if this is the first invocation
 620  * of this function.
 621  */
 622 static int
 623 _bus_dmamap_load_buffer(bus_dma_tag_t dmat,
 624                         bus_dmamap_t map,
 625                         void *buf, bus_size_t buflen,
 626                         bus_dma_segment_t *segments,
 627                         int nsegments,
 628                         pmap_t pmap,
 629                         int flags,
 630                         vm_paddr_t *lastpaddrp,
 631                         int *segp,
 632                         int first)
 633 {
 634         vm_offset_t vaddr;
 635         vm_paddr_t paddr, nextpaddr;
 636         bus_dma_segment_t *sg;
 637         bus_addr_t bmask;
 638         int seg, error = 0;
 639
 640         if (map == NULL || map == (void *)-1)
 641                 map = &nobounce_dmamap;
 642
 643 #ifdef INVARIANTS
 644         if (dmat->flags & BUS_DMA_ALIGNED)
 645                 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0);
 646 #endif
 647
 648         /*
 649          * If we are being called during a callback, pagesneeded will
 650          * be non-zero, so we can avoid doing the work twice.
 651          */
 652         if ((dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 653             map != &nobounce_dmamap && map->pagesneeded == 0) {
 654                 vm_offset_t vendaddr;
 655
 656                 /*
 657                  * Count the number of bounce pages
 658                  * needed in order to complete this transfer
 659                  */
 660                 vaddr = (vm_offset_t)buf;
 661                 vendaddr = (vm_offset_t)buf + buflen;
 662
 663                 while (vaddr < vendaddr) {
 664                         paddr = _bus_dma_extract(pmap, vaddr);
 665                         if (run_filter(dmat, paddr) != 0)
 666                                 map->pagesneeded++;
 667                         vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK));
 668                 }
 669         }
 670
 671         /* Reserve Necessary Bounce Pages */
 672         if (map->pagesneeded != 0) {
 673                 struct bounce_zone *bz;
 674
 675                 bz = dmat->bounce_zone;
 676                 BZ_LOCK(bz);
 677                 if (flags & BUS_DMA_NOWAIT) {
 678                         if (reserve_bounce_pages(dmat, map, 0) != 0) {
 679                                 BZ_UNLOCK(bz);
 680                                 error = ENOMEM;
 681                                 goto free_bounce;
 682                         }
 683                 } else {
 684                         if (reserve_bounce_pages(dmat, map, 1) != 0) {
 685                                 /* Queue us for resources */
 686                                 map->dmat = dmat;
 687                                 map->buf = buf;
 688                                 map->buflen = buflen;
 689
 690                                 STAILQ_INSERT_TAIL(
 691                                     &dmat->bounce_zone->bounce_map_waitinglist,
 692                                     map, links);
 693                                 BZ_UNLOCK(bz);
 694
 695                                 return (EINPROGRESS);
 696                         }
 697                 }
 698                 BZ_UNLOCK(bz);
 699         }
 700
 701         KKASSERT(*segp >= 1 && *segp <= nsegments);
 702         seg = *segp;
 703         sg = &segments[seg - 1];
 704
 705         vaddr = (vm_offset_t)buf;
 706         nextpaddr = *lastpaddrp;
 707         bmask = ~(dmat->boundary - 1);  /* note: will be 0 if boundary is 0 */
 708
 709         /* force at least one segment */
 710         do {
 711                 bus_size_t size;
 712
 713                 /*
 714                  * Per-page main loop
 715                  */
 716                 paddr = _bus_dma_extract(pmap, vaddr);
 717                 size = PAGE_SIZE - (paddr & PAGE_MASK);
 718                 if (size > buflen)
 719                         size = buflen;
 720                 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) {
 721                         /*
 722                          * NOTE: paddr may have different in-page offset,
 723                          *       unless BUS_DMA_KEEP_PG_OFFSET is set.
 724                          */
 725                         paddr = add_bounce_page(dmat, map, vaddr, &size);
 726                 }
 727
 728                 /*
 729                  * Fill in the bus_dma_segment
 730                  */
 731                 if (first) {
 732                         sg->ds_addr = paddr;
 733                         sg->ds_len = size;
 734                         first = 0;
 735                 } else if (paddr == nextpaddr) {
 736                         sg->ds_len += size;
 737                 } else {
 738                         sg++;
 739                         seg++;
 740                         if (seg > nsegments)
 741                                 break;
 742                         sg->ds_addr = paddr;
 743                         sg->ds_len = size;
 744                 }
 745                 nextpaddr = paddr + size;
 746
 747                 /*
 748                  * Handle maxsegsz and boundary issues with a nested loop
 749                  */
 750                 for (;;) {
 751                         bus_size_t tmpsize;
 752
 753                         /*
 754                          * Limit to the boundary and maximum segment size
 755                          */
 756                         if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) {
 757                                 tmpsize = dmat->boundary -
 758                                           (sg->ds_addr & ~bmask);
 759                                 if (tmpsize > dmat->maxsegsz)
 760                                         tmpsize = dmat->maxsegsz;
 761                                 KKASSERT(tmpsize < sg->ds_len);
 762                         } else if (sg->ds_len > dmat->maxsegsz) {
 763                                 tmpsize = dmat->maxsegsz;
 764                         } else {
 765                                 break;
 766                         }
 767
 768                         /*
 769                          * Futz, split the data into a new segment.
 770                          */
 771                         if (seg >= nsegments)
 772                                 goto fail;
 773                         sg[1].ds_len = sg[0].ds_len - tmpsize;
 774                         sg[1].ds_addr = sg[0].ds_addr + tmpsize;
 775                         sg[0].ds_len = tmpsize;
 776                         sg++;
 777                         seg++;
 778                 }
 779
 780                 /*
 781                  * Adjust for loop
 782                  */
 783                 buflen -= size;
 784                 vaddr += size;
 785         } while (buflen > 0);
 786 fail:
 787         if (buflen != 0)
 788                 error = EFBIG;
 789
 790         *segp = seg;
 791         *lastpaddrp = nextpaddr;
 792
 793 free_bounce:
 794         if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) &&
 795             map != &nobounce_dmamap) {
 796                 _bus_dmamap_unload(dmat, map);
 797                 return_bounce_pages(dmat, map);
 798         }
 799         return error;
 800 }
 801
 802 /*
 803  * Map the buffer buf into bus space using the dmamap map.
 804  */
 805 int
 806 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 807                 bus_size_t buflen, bus_dmamap_callback_t *callback,
 808                 void *callback_arg, int flags)
 809 {
 810         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 811         bus_dma_segment_t *segments;
 812         vm_paddr_t lastaddr = 0;
 813         int error, nsegs = 1;
 814
 815         if (map != NULL && map != (void *)-1) {
 816                 /*
 817                  * XXX
 818                  * Follow old semantics.  Once all of the callers are fixed,
 819                  * we should get rid of these internal flag "adjustment".
 820                  */
 821                 flags &= ~BUS_DMA_NOWAIT;
 822                 flags |= BUS_DMA_WAITOK;
 823
 824                 map->callback = callback;
 825                 map->callback_arg = callback_arg;
 826         }
 827
 828         segments = bus_dma_tag_lock(dmat, cache_segments);
 829         error = _bus_dmamap_load_buffer(dmat, map, buf, buflen,
 830                         segments, dmat->nsegments,
 831                         NULL, flags, &lastaddr, &nsegs, 1);
 832         if (error == EINPROGRESS) {
 833                 KKASSERT((dmat->flags &
 834                           (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL)) !=
 835                          (BUS_DMA_PRIVBZONE | BUS_DMA_ALLOCALL));
 836
 837                 if (dmat->flags & BUS_DMA_PROTECTED)
 838                         panic("protected dmamap callback will be defered");
 839
 840                 bus_dma_tag_unlock(dmat);
 841                 return error;
 842         }
 843         callback(callback_arg, segments, nsegs, error);
 844         bus_dma_tag_unlock(dmat);
 845         return 0;
 846 }
 847
 848 /*
 849  * Like _bus_dmamap_load(), but for ccb.
 850  */
 851 int
 852 bus_dmamap_load_ccb(bus_dma_tag_t dmat, bus_dmamap_t map, union ccb *ccb,
 853     bus_dmamap_callback_t *callback, void *callback_arg, int flags)
 854 {
 855         const struct ccb_scsiio *csio;
 856
 857         KASSERT(ccb->ccb_h.func_code == XPT_SCSI_IO ||
 858             ccb->ccb_h.func_code == XPT_CONT_TARGET_IO,
 859             ("invalid ccb func_code %u", ccb->ccb_h.func_code));
 860         csio = &ccb->csio;
 861
 862         return (bus_dmamap_load(dmat, map, csio->data_ptr, csio->dxfer_len,
 863             callback, callback_arg, flags));
 864 }
 865
 866 /*
 867  * Like _bus_dmamap_load(), but for mbufs.
 868  */
 869 int
 870 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map,
 871                      struct mbuf *m0,
 872                      bus_dmamap_callback2_t *callback, void *callback_arg,
 873                      int flags)
 874 {
 875         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 876         bus_dma_segment_t *segments;
 877         int nsegs, error;
 878
 879         /*
 880          * XXX
 881          * Follow old semantics.  Once all of the callers are fixed,
 882          * we should get rid of these internal flag "adjustment".
 883          */
 884         flags &= ~BUS_DMA_WAITOK;
 885         flags |= BUS_DMA_NOWAIT;
 886
 887         segments = bus_dma_tag_lock(dmat, cache_segments);
 888         error = bus_dmamap_load_mbuf_segment(dmat, map, m0,
 889                         segments, dmat->nsegments, &nsegs, flags);
 890         if (error) {
 891                 /* force "no valid mappings" in callback */
 892                 callback(callback_arg, segments, 0,
 893                          0, error);
 894         } else {
 895                 callback(callback_arg, segments, nsegs,
 896                          m0->m_pkthdr.len, error);
 897         }
 898         bus_dma_tag_unlock(dmat);
 899         return error;
 900 }
 901
 902 int
 903 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map,
 904                              struct mbuf *m0,
 905                              bus_dma_segment_t *segs, int maxsegs,
 906                              int *nsegs, int flags)
 907 {
 908         int error;
 909
 910         M_ASSERTPKTHDR(m0);
 911
 912         KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs));
 913         KASSERT(maxsegs <= dmat->nsegments,
 914                 ("%d too many segments, dmat only supports %d segments",
 915                  maxsegs, dmat->nsegments));
 916         KASSERT(flags & BUS_DMA_NOWAIT,
 917                 ("only BUS_DMA_NOWAIT is supported"));
 918
 919         if (m0->m_pkthdr.len <= dmat->maxsize) {
 920                 int first = 1;
 921                 vm_paddr_t lastaddr = 0;
 922                 struct mbuf *m;
 923
 924                 *nsegs = 1;
 925                 error = 0;
 926                 for (m = m0; m != NULL && error == 0; m = m->m_next) {
 927                         if (m->m_len == 0)
 928                                 continue;
 929
 930                         error = _bus_dmamap_load_buffer(dmat, map,
 931                                         m->m_data, m->m_len,
 932                                         segs, maxsegs,
 933                                         NULL, flags, &lastaddr,
 934                                         nsegs, first);
 935                         if (error == ENOMEM && !first) {
 936                                 /*
 937                                  * Out of bounce pages due to too many
 938                                  * fragments in the mbuf chain; return
 939                                  * EFBIG instead.
 940                                  */
 941                                 error = EFBIG;
 942                                 break;
 943                         }
 944                         first = 0;
 945                 }
 946 #ifdef INVARIANTS
 947                 if (!error)
 948                         KKASSERT(*nsegs <= maxsegs && *nsegs >= 1);
 949 #endif
 950         } else {
 951                 *nsegs = 0;
 952                 error = EINVAL;
 953         }
 954         KKASSERT(error != EINPROGRESS);
 955         return error;
 956 }
 957
 958 /*
 959  * Like _bus_dmamap_load(), but for uios.
 960  */
 961 int
 962 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map,
 963                     struct uio *uio,
 964                     bus_dmamap_callback2_t *callback, void *callback_arg,
 965                     int flags)
 966 {
 967         vm_paddr_t lastaddr;
 968         int nsegs, error, first, i;
 969         bus_size_t resid;
 970         struct iovec *iov;
 971         pmap_t pmap;
 972         bus_dma_segment_t cache_segments[BUS_DMA_CACHE_SEGMENTS];
 973         bus_dma_segment_t *segments;
 974         bus_dma_segment_t *segs;
 975         int nsegs_left;
 976
 977         if (dmat->nsegments <= BUS_DMA_CACHE_SEGMENTS)
 978                 segments = cache_segments;
 979         else
 980                 segments = kmalloc(sizeof(bus_dma_segment_t) * dmat->nsegments,
 981                                    M_DEVBUF, M_WAITOK | M_ZERO);
 982
 983         /*
 984          * XXX
 985          * Follow old semantics.  Once all of the callers are fixed,
 986          * we should get rid of these internal flag "adjustment".
 987          */
 988         flags &= ~BUS_DMA_WAITOK;
 989         flags |= BUS_DMA_NOWAIT;
 990
 991         resid = (bus_size_t)uio->uio_resid;
 992         iov = uio->uio_iov;
 993
 994         segs = segments;
 995         nsegs_left = dmat->nsegments;
 996
 997         if (uio->uio_segflg == UIO_USERSPACE) {
 998                 struct thread *td;
 999
1000                 td = uio->uio_td;
1001                 KASSERT(td != NULL && td->td_proc != NULL,
1002                         ("bus_dmamap_load_uio: USERSPACE but no proc"));
1003                 pmap = vmspace_pmap(td->td_proc->p_vmspace);
1004         } else {
1005                 pmap = NULL;
1006         }
1007
1008         error = 0;
1009         nsegs = 1;
1010         first = 1;
1011         lastaddr = 0;
1012         for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) {
1013                 /*
1014                  * Now at the first iovec to load.  Load each iovec
1015                  * until we have exhausted the residual count.
1016                  */
1017                 bus_size_t minlen =
1018                         resid < iov[i].iov_len ? resid : iov[i].iov_len;
1019                 caddr_t addr = (caddr_t) iov[i].iov_base;
1020
1021                 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen,
1022                                 segs, nsegs_left,
1023                                 pmap, flags, &lastaddr, &nsegs, first);
1024                 first = 0;
1025
1026                 resid -= minlen;
1027                 if (error == 0) {
1028                         nsegs_left -= nsegs;
1029                         segs += nsegs;
1030                 }
1031         }
1032
1033         /*
1034          * Minimum one DMA segment, even if 0-length buffer.
1035          */
1036         if (nsegs_left == dmat->nsegments)
1037                 --nsegs_left;
1038
1039         if (error) {
1040                 /* force "no valid mappings" in callback */
1041                 callback(callback_arg, segments, 0,
1042                          0, error);
1043         } else {
1044                 callback(callback_arg, segments, dmat->nsegments - nsegs_left,
1045                          (bus_size_t)uio->uio_resid, error);
1046         }
1047         if (dmat->nsegments > BUS_DMA_CACHE_SEGMENTS)
1048                 kfree(segments, M_DEVBUF);
1049         return error;
1050 }
1051
1052 /*
1053  * Release the mapping held by map.
1054  */
1055 void
1056 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
1057 {
1058         struct bounce_page *bpage;
1059
1060         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1061                 STAILQ_REMOVE_HEAD(&map->bpages, links);
1062                 free_bounce_page(dmat, bpage);
1063         }
1064 }
1065
1066 void
1067 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op)
1068 {
1069         struct bounce_page *bpage;
1070
1071         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
1072                 /*
1073                  * Handle data bouncing.  We might also
1074                  * want to add support for invalidating
1075                  * the caches on broken hardware
1076                  */
1077                 if (op & BUS_DMASYNC_PREWRITE) {
1078                         while (bpage != NULL) {
1079                                 bcopy((void *)bpage->datavaddr,
1080                                       (void *)bpage->vaddr,
1081                                       bpage->datacount);
1082                                 bpage = STAILQ_NEXT(bpage, links);
1083                         }
1084                         cpu_sfence();
1085                         dmat->bounce_zone->total_bounced++;
1086                 }
1087                 if (op & BUS_DMASYNC_POSTREAD) {
1088                         cpu_lfence();
1089                         while (bpage != NULL) {
1090                                 bcopy((void *)bpage->vaddr,
1091                                       (void *)bpage->datavaddr,
1092                                       bpage->datacount);
1093                                 bpage = STAILQ_NEXT(bpage, links);
1094                         }
1095                         dmat->bounce_zone->total_bounced++;
1096                 }
1097                 /* BUS_DMASYNC_PREREAD          - no operation on intel */
1098                 /* BUS_DMASYNC_POSTWRITE        - no operation on intel */
1099         }
1100 }
1101
1102 static int
1103 alloc_bounce_zone(bus_dma_tag_t dmat)
1104 {
1105         struct bounce_zone *bz, *new_bz;
1106
1107         KASSERT(dmat->bounce_zone == NULL,
1108                 ("bounce zone was already assigned"));
1109
1110         new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO);
1111
1112         lwkt_gettoken(&bounce_zone_tok);
1113
1114         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1115                 /*
1116                  * For shared bounce zone, check to see
1117                  * if we already have a suitable zone
1118                  */
1119                 STAILQ_FOREACH(bz, &bounce_zone_list, links) {
1120                         if (dmat->alignment <= bz->alignment &&
1121                             dmat->lowaddr >= bz->lowaddr) {
1122                                 lwkt_reltoken(&bounce_zone_tok);
1123
1124                                 dmat->bounce_zone = bz;
1125                                 kfree(new_bz, M_DEVBUF);
1126                                 return 0;
1127                         }
1128                 }
1129         }
1130         bz = new_bz;
1131
1132         spin_init(&bz->spin, "allocbouncezone");
1133         STAILQ_INIT(&bz->bounce_page_list);
1134         STAILQ_INIT(&bz->bounce_map_waitinglist);
1135         bz->free_bpages = 0;
1136         bz->reserved_bpages = 0;
1137         bz->active_bpages = 0;
1138         bz->lowaddr = dmat->lowaddr;
1139         bz->alignment = round_page(dmat->alignment);
1140         ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
1141
1142         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0) {
1143                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
1144                 busdma_zonecount++;
1145                 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
1146         } else {
1147                 ksnprintf(bz->zoneid, 8, "zone%d", busdma_priv_zonecount);
1148                 busdma_priv_zonecount--;
1149         }
1150
1151         lwkt_reltoken(&bounce_zone_tok);
1152
1153         dmat->bounce_zone = bz;
1154
1155         sysctl_ctx_init(&bz->sysctl_ctx);
1156         bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx,
1157             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
1158             CTLFLAG_RD, 0, "");
1159         if (bz->sysctl_tree == NULL) {
1160                 sysctl_ctx_free(&bz->sysctl_ctx);
1161                 return 0;       /* XXX error code? */
1162         }
1163
1164         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1165             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1166             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
1167             "Total bounce pages");
1168         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1169             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1170             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
1171             "Free bounce pages");
1172         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1173             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1174             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
1175             "Reserved bounce pages");
1176         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1177             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1178             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
1179             "Active bounce pages");
1180         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1181             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1182             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
1183             "Total bounce requests");
1184         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1185             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1186             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
1187             "Total bounce requests that were deferred");
1188         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1189             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1190             "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0,
1191             "Total bounce page reservations that were failed");
1192         SYSCTL_ADD_STRING(&bz->sysctl_ctx,
1193             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1194             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
1195         SYSCTL_ADD_INT(&bz->sysctl_ctx,
1196             SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO,
1197             "alignment", CTLFLAG_RD, &bz->alignment, 0, "");
1198
1199         return 0;
1200 }
1201
1202 static int
1203 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags)
1204 {
1205         struct bounce_zone *bz = dmat->bounce_zone;
1206         int count = 0, mflags;
1207
1208         if (flags & BUS_DMA_NOWAIT)
1209                 mflags = M_NOWAIT;
1210         else
1211                 mflags = M_WAITOK;
1212
1213         while (numpages > 0) {
1214                 struct bounce_page *bpage;
1215
1216                 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO);
1217
1218                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
1219                                                          mflags, 0ul,
1220                                                          bz->lowaddr,
1221                                                          bz->alignment, 0);
1222                 if (bpage->vaddr == 0) {
1223                         kfree(bpage, M_DEVBUF);
1224                         break;
1225                 }
1226                 bpage->busaddr = pmap_kextract(bpage->vaddr);
1227
1228                 BZ_LOCK(bz);
1229                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
1230                 total_bounce_pages++;
1231                 bz->total_bpages++;
1232                 bz->free_bpages++;
1233                 BZ_UNLOCK(bz);
1234
1235                 count++;
1236                 numpages--;
1237         }
1238         return count;
1239 }
1240
1241 static void
1242 free_bounce_pages_all(bus_dma_tag_t dmat)
1243 {
1244         struct bounce_zone *bz = dmat->bounce_zone;
1245         struct bounce_page *bpage;
1246
1247         BZ_LOCK(bz);
1248
1249         while ((bpage = STAILQ_FIRST(&bz->bounce_page_list)) != NULL) {
1250                 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1251
1252                 KKASSERT(total_bounce_pages > 0);
1253                 total_bounce_pages--;
1254
1255                 KKASSERT(bz->total_bpages > 0);
1256                 bz->total_bpages--;
1257
1258                 KKASSERT(bz->free_bpages > 0);
1259                 bz->free_bpages--;
1260
1261                 BZ_UNLOCK(bz);
1262                 contigfree((void *)bpage->vaddr, PAGE_SIZE, M_DEVBUF);
1263                 kfree(bpage, M_DEVBUF);
1264                 BZ_LOCK(bz);
1265         }
1266         if (bz->total_bpages) {
1267                 kprintf("#%d bounce pages are still in use\n",
1268                         bz->total_bpages);
1269                 print_backtrace(-1);
1270         }
1271
1272         BZ_UNLOCK(bz);
1273 }
1274
1275 static void
1276 free_bounce_zone(bus_dma_tag_t dmat)
1277 {
1278         struct bounce_zone *bz = dmat->bounce_zone;
1279
1280         if (bz == NULL)
1281                 return;
1282
1283         if ((dmat->flags & BUS_DMA_PRIVBZONE) == 0)
1284                 return;
1285
1286         free_bounce_pages_all(dmat);
1287         dmat->bounce_zone = NULL;
1288
1289         if (bz->sysctl_tree != NULL)
1290                 sysctl_ctx_free(&bz->sysctl_ctx);
1291         kfree(bz, M_DEVBUF);
1292 }
1293
1294 /* Assume caller holds bounce zone spinlock */
1295 static int
1296 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
1297 {
1298         struct bounce_zone *bz = dmat->bounce_zone;
1299         int pages;
1300
1301         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
1302         if (!commit && map->pagesneeded > (map->pagesreserved + pages)) {
1303                 bz->reserve_failed++;
1304                 return (map->pagesneeded - (map->pagesreserved + pages));
1305         }
1306
1307         bz->free_bpages -= pages;
1308
1309         bz->reserved_bpages += pages;
1310         KKASSERT(bz->reserved_bpages <= bz->total_bpages);
1311
1312         map->pagesreserved += pages;
1313         pages = map->pagesneeded - map->pagesreserved;
1314
1315         return pages;
1316 }
1317
1318 static void
1319 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map)
1320 {
1321         struct bounce_zone *bz = dmat->bounce_zone;
1322         int reserved = map->pagesreserved;
1323         bus_dmamap_t wait_map;
1324
1325         map->pagesreserved = 0;
1326         map->pagesneeded = 0;
1327
1328         if (reserved == 0)
1329                 return;
1330
1331         BZ_LOCK(bz);
1332
1333         bz->free_bpages += reserved;
1334         KKASSERT(bz->free_bpages <= bz->total_bpages);
1335
1336         KKASSERT(bz->reserved_bpages >= reserved);
1337         bz->reserved_bpages -= reserved;
1338
1339         wait_map = get_map_waiting(dmat);
1340
1341         BZ_UNLOCK(bz);
1342
1343         if (wait_map != NULL)
1344                 add_map_callback(map);
1345 }
1346
1347 static bus_addr_t
1348 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
1349                 bus_size_t *sizep)
1350 {
1351         struct bounce_zone *bz = dmat->bounce_zone;
1352         struct bounce_page *bpage;
1353         bus_size_t size;
1354
1355         KASSERT(map->pagesneeded > 0, ("map doesn't need any pages"));
1356         map->pagesneeded--;
1357
1358         KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages"));
1359         map->pagesreserved--;
1360
1361         BZ_LOCK(bz);
1362
1363         bpage = STAILQ_FIRST(&bz->bounce_page_list);
1364         KASSERT(bpage != NULL, ("free page list is empty"));
1365         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
1366
1367         KKASSERT(bz->reserved_bpages > 0);
1368         bz->reserved_bpages--;
1369
1370         bz->active_bpages++;
1371         KKASSERT(bz->active_bpages <= bz->total_bpages);
1372
1373         BZ_UNLOCK(bz);
1374
1375         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1376                 /*
1377                  * Page offset needs to be preserved.  No size adjustments
1378                  * needed.
1379                  */
1380                 bpage->vaddr |= vaddr & PAGE_MASK;
1381                 bpage->busaddr |= vaddr & PAGE_MASK;
1382                 size = *sizep;
1383         } else {
1384                 /*
1385                  * Realign to bounce page base address, reduce size if
1386                  * necessary.  Bounce pages are typically already
1387                  * page-aligned.
1388                  */
1389                 size = PAGE_SIZE - (bpage->busaddr & PAGE_MASK);
1390                 if (size < *sizep) {
1391                         *sizep = size;
1392                 } else {
1393                         size = *sizep;
1394                 }
1395         }
1396
1397         bpage->datavaddr = vaddr;
1398         bpage->datacount = size;
1399         STAILQ_INSERT_TAIL(&map->bpages, bpage, links);
1400         return bpage->busaddr;
1401 }
1402
1403 static void
1404 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1405 {
1406         struct bounce_zone *bz = dmat->bounce_zone;
1407         bus_dmamap_t map;
1408
1409         bpage->datavaddr = 0;
1410         bpage->datacount = 0;
1411
1412         if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) {
1413                 /*
1414                  * Reset the bounce page to start at offset 0.  Other uses
1415                  * of this bounce page may need to store a full page of
1416                  * data and/or assume it starts on a page boundary.
1417                  */
1418                 bpage->vaddr &= ~PAGE_MASK;
1419                 bpage->busaddr &= ~PAGE_MASK;
1420         }
1421
1422         BZ_LOCK(bz);
1423
1424         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1425
1426         bz->free_bpages++;
1427         KKASSERT(bz->free_bpages <= bz->total_bpages);
1428
1429         KKASSERT(bz->active_bpages > 0);
1430         bz->active_bpages--;
1431
1432         map = get_map_waiting(dmat);
1433
1434         BZ_UNLOCK(bz);
1435
1436         if (map != NULL && map != (void *)-1)
1437                 add_map_callback(map);
1438 }
1439
1440 /* Assume caller holds bounce zone spinlock */
1441 static bus_dmamap_t
1442 get_map_waiting(bus_dma_tag_t dmat)
1443 {
1444         struct bounce_zone *bz = dmat->bounce_zone;
1445         bus_dmamap_t map;
1446
1447         map = STAILQ_FIRST(&bz->bounce_map_waitinglist);
1448         if (map != NULL && map != (void *)-1) {
1449                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1450                         STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links);
1451                         bz->total_deferred++;
1452                 } else {
1453                         map = NULL;
1454                 }
1455         }
1456         return map;
1457 }
1458
1459 static void
1460 add_map_callback(bus_dmamap_t map)
1461 {
1462         spin_lock(&bounce_map_list_spin);
1463         STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links);
1464         busdma_swi_pending = 1;
1465         setsoftvm();
1466         spin_unlock(&bounce_map_list_spin);
1467 }
1468
1469 void
1470 busdma_swi(void)
1471 {
1472         bus_dmamap_t map;
1473
1474         spin_lock(&bounce_map_list_spin);
1475         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1476                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1477                 spin_unlock(&bounce_map_list_spin);
1478                 bus_dmamap_load(map->dmat, map, map->buf, map->buflen,
1479                                 map->callback, map->callback_arg, /*flags*/0);
1480                 spin_lock(&bounce_map_list_spin);
1481         }
1482         spin_unlock(&bounce_map_list_spin);
1483 }
1484
1485 int
1486 bus_space_map(bus_space_tag_t t __unused, bus_addr_t addr, bus_size_t size,
1487     int flags __unused, bus_space_handle_t *bshp)
1488 {
1489
1490         if (t == X86_64_BUS_SPACE_MEM)
1491                 *bshp = (uintptr_t)pmap_mapdev(addr, size);
1492         else
1493                 *bshp = addr;
1494         return (0);
1495 }
1496
1497 void
1498 bus_space_unmap(bus_space_tag_t t, bus_space_handle_t bsh, bus_size_t size)
1499 {
1500         if (t == X86_64_BUS_SPACE_MEM)
1501                 pmap_unmapdev(bsh, size);
1502 }