sys/arm64/arm64/busdma_bounce.c

   1 /*-
   2  * Copyright (c) 1997, 1998 Justin T. Gibbs.
   3  * Copyright (c) 2015 The FreeBSD Foundation
   4  * All rights reserved.
   5  *
   6  * Portions of this software were developed by Semihalf
   7  * under sponsorship of the FreeBSD Foundation.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions, and the following disclaimer,
  14  *    without modification, immediately at the beginning of the file.
  15  * 2. The name of the author may not be used to endorse or promote products
  16  *    derived from this software without specific prior written permission.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  22  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/malloc.h>
  37 #include <sys/bus.h>
  38 #include <sys/interrupt.h>
  39 #include <sys/kernel.h>
  40 #include <sys/ktr.h>
  41 #include <sys/lock.h>
  42 #include <sys/proc.h>
  43 #include <sys/memdesc.h>
  44 #include <sys/mutex.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/uio.h>
  47
  48 #include <vm/vm.h>
  49 #include <vm/vm_extern.h>
  50 #include <vm/vm_kern.h>
  51 #include <vm/vm_page.h>
  52 #include <vm/vm_map.h>
  53
  54 #include <machine/atomic.h>
  55 #include <machine/bus.h>
  56 #include <machine/md_var.h>
  57 #include <arm64/include/bus_dma_impl.h>
  58
  59 #define MAX_BPAGES 4096
  60
  61 enum {
  62         BUS_DMA_COULD_BOUNCE    = 0x01,
  63         BUS_DMA_MIN_ALLOC_COMP  = 0x02,
  64         BUS_DMA_KMEM_ALLOC      = 0x04,
  65 };
  66
  67 struct bounce_zone;
  68
  69 struct bus_dma_tag {
  70         struct bus_dma_tag_common common;
  71         int                     map_count;
  72         int                     bounce_flags;
  73         bus_dma_segment_t       *segments;
  74         struct bounce_zone      *bounce_zone;
  75 };
  76
  77 struct bounce_page {
  78         vm_offset_t     vaddr;          /* kva of bounce buffer */
  79         bus_addr_t      busaddr;        /* Physical address */
  80         vm_offset_t     datavaddr;      /* kva of client data */
  81         bus_addr_t      dataaddr;       /* client physical address */
  82         bus_size_t      datacount;      /* client data count */
  83         STAILQ_ENTRY(bounce_page) links;
  84 };
  85
  86 int busdma_swi_pending;
  87
  88 struct bounce_zone {
  89         STAILQ_ENTRY(bounce_zone) links;
  90         STAILQ_HEAD(bp_list, bounce_page) bounce_page_list;
  91         int             total_bpages;
  92         int             free_bpages;
  93         int             reserved_bpages;
  94         int             active_bpages;
  95         int             total_bounced;
  96         int             total_deferred;
  97         int             map_count;
  98         bus_size_t      alignment;
  99         bus_addr_t      lowaddr;
 100         char            zoneid[8];
 101         char            lowaddrid[20];
 102         struct sysctl_ctx_list sysctl_tree;
 103         struct sysctl_oid *sysctl_tree_top;
 104 };
 105
 106 static struct mtx bounce_lock;
 107 static int total_bpages;
 108 static int busdma_zonecount;
 109 static STAILQ_HEAD(, bounce_zone) bounce_zone_list;
 110
 111 static SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters");
 112 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0,
 113            "Total bounce pages");
 114
 115 struct bus_dmamap {
 116         struct bp_list         bpages;
 117         int                    pagesneeded;
 118         int                    pagesreserved;
 119         bus_dma_tag_t          dmat;
 120         struct memdesc         mem;
 121         bus_dmamap_callback_t *callback;
 122         void                  *callback_arg;
 123         STAILQ_ENTRY(bus_dmamap) links;
 124 };
 125
 126 static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist;
 127 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist;
 128 static struct bus_dmamap nobounce_dmamap;
 129
 130 static void init_bounce_pages(void *dummy);
 131 static int alloc_bounce_zone(bus_dma_tag_t dmat);
 132 static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages);
 133 static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
 134     int commit);
 135 static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map,
 136     vm_offset_t vaddr, bus_addr_t addr, bus_size_t size);
 137 static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage);
 138 int run_filter(bus_dma_tag_t dmat, bus_addr_t paddr);
 139 static void _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
 140     pmap_t pmap, void *buf, bus_size_t buflen, int flags);
 141 static void _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
 142     vm_paddr_t buf, bus_size_t buflen, int flags);
 143 static int _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map,
 144     int flags);
 145
 146 /*
 147  * Allocate a device specific dma_tag.
 148  */
 149 static int
 150 bounce_bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
 151     bus_addr_t boundary, bus_addr_t lowaddr, bus_addr_t highaddr,
 152     bus_dma_filter_t *filter, void *filterarg, bus_size_t maxsize,
 153     int nsegments, bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
 154     void *lockfuncarg, bus_dma_tag_t *dmat)
 155 {
 156         bus_dma_tag_t newtag;
 157         int error;
 158
 159         *dmat = NULL;
 160         error = common_bus_dma_tag_create(parent != NULL ? &parent->common :
 161             NULL, alignment, boundary, lowaddr, highaddr, filter, filterarg,
 162             maxsize, nsegments, maxsegsz, flags, lockfunc, lockfuncarg,
 163             sizeof (struct bus_dma_tag), (void **)&newtag);
 164         if (error != 0)
 165                 return (error);
 166
 167         newtag->common.impl = &bus_dma_bounce_impl;
 168         newtag->map_count = 0;
 169         newtag->segments = NULL;
 170
 171         if (parent != NULL && ((newtag->common.filter != NULL) ||
 172             ((parent->bounce_flags & BUS_DMA_COULD_BOUNCE) != 0)))
 173                 newtag->bounce_flags |= BUS_DMA_COULD_BOUNCE;
 174
 175         if (newtag->common.lowaddr < ptoa((vm_paddr_t)Maxmem) ||
 176             newtag->common.alignment > 1)
 177                 newtag->bounce_flags |= BUS_DMA_COULD_BOUNCE;
 178
 179         if (((newtag->bounce_flags & BUS_DMA_COULD_BOUNCE) != 0) &&
 180             (flags & BUS_DMA_ALLOCNOW) != 0) {
 181                 struct bounce_zone *bz;
 182
 183                 /* Must bounce */
 184                 if ((error = alloc_bounce_zone(newtag)) != 0) {
 185                         free(newtag, M_DEVBUF);
 186                         return (error);
 187                 }
 188                 bz = newtag->bounce_zone;
 189
 190                 if (ptoa(bz->total_bpages) < maxsize) {
 191                         int pages;
 192
 193                         pages = atop(maxsize) - bz->total_bpages;
 194
 195                         /* Add pages to our bounce pool */
 196                         if (alloc_bounce_pages(newtag, pages) < pages)
 197                                 error = ENOMEM;
 198                 }
 199                 /* Performed initial allocation */
 200                 newtag->bounce_flags |= BUS_DMA_MIN_ALLOC_COMP;
 201         } else
 202                 error = 0;
 203
 204         if (error != 0)
 205                 free(newtag, M_DEVBUF);
 206         else
 207                 *dmat = newtag;
 208         CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
 209             __func__, newtag, (newtag != NULL ? newtag->common.flags : 0),
 210             error);
 211         return (error);
 212 }
 213
 214 static int
 215 bounce_bus_dma_tag_destroy(bus_dma_tag_t dmat)
 216 {
 217         bus_dma_tag_t dmat_copy, parent;
 218         int error;
 219
 220         error = 0;
 221         dmat_copy = dmat;
 222
 223         if (dmat != NULL) {
 224                 if (dmat->map_count != 0) {
 225                         error = EBUSY;
 226                         goto out;
 227                 }
 228                 while (dmat != NULL) {
 229                         parent = (bus_dma_tag_t)dmat->common.parent;
 230                         atomic_subtract_int(&dmat->common.ref_count, 1);
 231                         if (dmat->common.ref_count == 0) {
 232                                 if (dmat->segments != NULL)
 233                                         free(dmat->segments, M_DEVBUF);
 234                                 free(dmat, M_DEVBUF);
 235                                 /*
 236                                  * Last reference count, so
 237                                  * release our reference
 238                                  * count on our parent.
 239                                  */
 240                                 dmat = parent;
 241                         } else
 242                                 dmat = NULL;
 243                 }
 244         }
 245 out:
 246         CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
 247         return (error);
 248 }
 249
 250 /*
 251  * Allocate a handle for mapping from kva/uva/physical
 252  * address space into bus device space.
 253  */
 254 static int
 255 bounce_bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
 256 {
 257         struct bounce_zone *bz;
 258         int error, maxpages, pages;
 259
 260         error = 0;
 261
 262         if (dmat->segments == NULL) {
 263                 dmat->segments = (bus_dma_segment_t *)malloc(
 264                     sizeof(bus_dma_segment_t) * dmat->common.nsegments,
 265                     M_DEVBUF, M_NOWAIT);
 266                 if (dmat->segments == NULL) {
 267                         CTR3(KTR_BUSDMA, "%s: tag %p error %d",
 268                             __func__, dmat, ENOMEM);
 269                         return (ENOMEM);
 270                 }
 271         }
 272
 273         /*
 274          * Bouncing might be required if the driver asks for an active
 275          * exclusion region, a data alignment that is stricter than 1, and/or
 276          * an active address boundary.
 277          */
 278         if (dmat->bounce_flags & BUS_DMA_COULD_BOUNCE) {
 279                 /* Must bounce */
 280                 if (dmat->bounce_zone == NULL) {
 281                         if ((error = alloc_bounce_zone(dmat)) != 0)
 282                                 return (error);
 283                 }
 284                 bz = dmat->bounce_zone;
 285
 286                 *mapp = (bus_dmamap_t)malloc(sizeof(**mapp), M_DEVBUF,
 287                     M_NOWAIT | M_ZERO);
 288                 if (*mapp == NULL) {
 289                         CTR3(KTR_BUSDMA, "%s: tag %p error %d",
 290                             __func__, dmat, ENOMEM);
 291                         return (ENOMEM);
 292                 }
 293
 294                 /* Initialize the new map */
 295                 STAILQ_INIT(&((*mapp)->bpages));
 296
 297                 /*
 298                  * Attempt to add pages to our pool on a per-instance
 299                  * basis up to a sane limit.
 300                  */
 301                 if (dmat->common.alignment > 1)
 302                         maxpages = MAX_BPAGES;
 303                 else
 304                         maxpages = MIN(MAX_BPAGES, Maxmem -
 305                             atop(dmat->common.lowaddr));
 306                 if ((dmat->bounce_flags & BUS_DMA_MIN_ALLOC_COMP) == 0 ||
 307                     (bz->map_count > 0 && bz->total_bpages < maxpages)) {
 308                         pages = MAX(atop(dmat->common.maxsize), 1);
 309                         pages = MIN(maxpages - bz->total_bpages, pages);
 310                         pages = MAX(pages, 1);
 311                         if (alloc_bounce_pages(dmat, pages) < pages)
 312                                 error = ENOMEM;
 313                         if ((dmat->bounce_flags & BUS_DMA_MIN_ALLOC_COMP)
 314                             == 0) {
 315                                 if (error == 0) {
 316                                         dmat->bounce_flags |=
 317                                             BUS_DMA_MIN_ALLOC_COMP;
 318                                 }
 319                         } else
 320                                 error = 0;
 321                 }
 322                 bz->map_count++;
 323         } else {
 324                 *mapp = NULL;
 325         }
 326         if (error == 0)
 327                 dmat->map_count++;
 328         CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
 329             __func__, dmat, dmat->common.flags, error);
 330         return (error);
 331 }
 332
 333 /*
 334  * Destroy a handle for mapping from kva/uva/physical
 335  * address space into bus device space.
 336  */
 337 static int
 338 bounce_bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map)
 339 {
 340
 341         if (map != NULL && map != &nobounce_dmamap) {
 342                 if (STAILQ_FIRST(&map->bpages) != NULL) {
 343                         CTR3(KTR_BUSDMA, "%s: tag %p error %d",
 344                             __func__, dmat, EBUSY);
 345                         return (EBUSY);
 346                 }
 347                 if (dmat->bounce_zone)
 348                         dmat->bounce_zone->map_count--;
 349                 free(map, M_DEVBUF);
 350         }
 351         dmat->map_count--;
 352         CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat);
 353         return (0);
 354 }
 355
 356
 357 /*
 358  * Allocate a piece of memory that can be efficiently mapped into
 359  * bus device space based on the constraints lited in the dma tag.
 360  * A dmamap to for use with dmamap_load is also allocated.
 361  */
 362 static int
 363 bounce_bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
 364     bus_dmamap_t *mapp)
 365 {
 366         /*
 367          * XXX ARM64TODO:
 368          * This bus_dma implementation requires IO-Coherent architecutre.
 369          * If IO-Coherency is not guaranteed, the BUS_DMA_COHERENT flag has
 370          * to be implented using non-cacheable memory.
 371          */
 372
 373         vm_memattr_t attr;
 374         int mflags;
 375
 376         if (flags & BUS_DMA_NOWAIT)
 377                 mflags = M_NOWAIT;
 378         else
 379                 mflags = M_WAITOK;
 380
 381         /* If we succeed, no mapping/bouncing will be required */
 382         *mapp = NULL;
 383
 384         if (dmat->segments == NULL) {
 385                 dmat->segments = (bus_dma_segment_t *)malloc(
 386                     sizeof(bus_dma_segment_t) * dmat->common.nsegments,
 387                     M_DEVBUF, mflags);
 388                 if (dmat->segments == NULL) {
 389                         CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
 390                             __func__, dmat, dmat->common.flags, ENOMEM);
 391                         return (ENOMEM);
 392                 }
 393         }
 394         if (flags & BUS_DMA_ZERO)
 395                 mflags |= M_ZERO;
 396         if (flags & BUS_DMA_NOCACHE)
 397                 attr = VM_MEMATTR_UNCACHEABLE;
 398         else
 399                 attr = VM_MEMATTR_DEFAULT;
 400
 401         /*
 402          * XXX:
 403          * (dmat->alignment < dmat->maxsize) is just a quick hack; the exact
 404          * alignment guarantees of malloc need to be nailed down, and the
 405          * code below should be rewritten to take that into account.
 406          *
 407          * In the meantime, we'll warn the user if malloc gets it wrong.
 408          */
 409         if ((dmat->common.maxsize <= PAGE_SIZE) &&
 410            (dmat->common.alignment < dmat->common.maxsize) &&
 411             dmat->common.lowaddr >= ptoa((vm_paddr_t)Maxmem) &&
 412             attr == VM_MEMATTR_DEFAULT) {
 413                 *vaddr = malloc(dmat->common.maxsize, M_DEVBUF, mflags);
 414         } else if (dmat->common.nsegments >= btoc(dmat->common.maxsize) &&
 415             dmat->common.alignment <= PAGE_SIZE &&
 416             (dmat->common.boundary == 0 ||
 417             dmat->common.boundary >= dmat->common.lowaddr)) {
 418                 /* Page-based multi-segment allocations allowed */
 419                 *vaddr = (void *)kmem_alloc_attr(kernel_arena,
 420                     dmat->common.maxsize, mflags, 0ul, dmat->common.lowaddr,
 421                     attr);
 422                 dmat->bounce_flags |= BUS_DMA_KMEM_ALLOC;
 423         } else {
 424                 *vaddr = (void *)kmem_alloc_contig(kernel_arena,
 425                     dmat->common.maxsize, mflags, 0ul, dmat->common.lowaddr,
 426                     dmat->common.alignment != 0 ? dmat->common.alignment : 1ul,
 427                     dmat->common.boundary, attr);
 428                 dmat->bounce_flags |= BUS_DMA_KMEM_ALLOC;
 429         }
 430         if (*vaddr == NULL) {
 431                 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
 432                     __func__, dmat, dmat->common.flags, ENOMEM);
 433                 return (ENOMEM);
 434         } else if (vtophys(*vaddr) & (dmat->common.alignment - 1)) {
 435                 printf("bus_dmamem_alloc failed to align memory properly.\n");
 436         }
 437         CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d",
 438             __func__, dmat, dmat->common.flags, 0);
 439         return (0);
 440 }
 441
 442 /*
 443  * Free a piece of memory and it's allociated dmamap, that was allocated
 444  * via bus_dmamem_alloc.  Make the same choice for free/contigfree.
 445  */
 446 static void
 447 bounce_bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map)
 448 {
 449         /*
 450          * dmamem does not need to be bounced, so the map should be
 451          * NULL and the BUS_DMA_KMEM_ALLOC flag cleared if malloc()
 452          * was used and set if kmem_alloc_contig() was used.
 453          */
 454         if (map != NULL)
 455                 panic("bus_dmamem_free: Invalid map freed\n");
 456         if ((dmat->bounce_flags & BUS_DMA_KMEM_ALLOC) == 0)
 457                 free(vaddr, M_DEVBUF);
 458         else
 459                 kmem_free(kernel_arena, (vm_offset_t)vaddr,
 460                     dmat->common.maxsize);
 461         CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat,
 462             dmat->bounce_flags);
 463 }
 464
 465 static void
 466 _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf,
 467     bus_size_t buflen, int flags)
 468 {
 469         bus_addr_t curaddr;
 470         bus_size_t sgsize;
 471
 472         if ((map != &nobounce_dmamap && map->pagesneeded == 0)) {
 473                 /*
 474                  * Count the number of bounce pages
 475                  * needed in order to complete this transfer
 476                  */
 477                 curaddr = buf;
 478                 while (buflen != 0) {
 479                         sgsize = MIN(buflen, dmat->common.maxsegsz);
 480                         if (bus_dma_run_filter(&dmat->common, curaddr)) {
 481                                 sgsize = MIN(sgsize, PAGE_SIZE);
 482                                 map->pagesneeded++;
 483                         }
 484                         curaddr += sgsize;
 485                         buflen -= sgsize;
 486                 }
 487                 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
 488         }
 489 }
 490
 491 static void
 492 _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, pmap_t pmap,
 493     void *buf, bus_size_t buflen, int flags)
 494 {
 495         vm_offset_t vaddr;
 496         vm_offset_t vendaddr;
 497         bus_addr_t paddr;
 498         bus_size_t sg_len;
 499
 500         if ((map != &nobounce_dmamap && map->pagesneeded == 0)) {
 501                 CTR4(KTR_BUSDMA, "lowaddr= %d Maxmem= %d, boundary= %d, "
 502                     "alignment= %d", dmat->common.lowaddr,
 503                     ptoa((vm_paddr_t)Maxmem),
 504                     dmat->common.boundary, dmat->common.alignment);
 505                 CTR3(KTR_BUSDMA, "map= %p, nobouncemap= %p, pagesneeded= %d",
 506                     map, &nobounce_dmamap, map->pagesneeded);
 507                 /*
 508                  * Count the number of bounce pages
 509                  * needed in order to complete this transfer
 510                  */
 511                 vaddr = (vm_offset_t)buf;
 512                 vendaddr = (vm_offset_t)buf + buflen;
 513
 514                 while (vaddr < vendaddr) {
 515                         sg_len = PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK);
 516                         if (pmap == kernel_pmap)
 517                                 paddr = pmap_kextract(vaddr);
 518                         else
 519                                 paddr = pmap_extract(pmap, vaddr);
 520                         if (bus_dma_run_filter(&dmat->common, paddr) != 0) {
 521                                 sg_len = roundup2(sg_len,
 522                                     dmat->common.alignment);
 523                                 map->pagesneeded++;
 524                         }
 525                         vaddr += sg_len;
 526                 }
 527                 CTR1(KTR_BUSDMA, "pagesneeded= %d\n", map->pagesneeded);
 528         }
 529 }
 530
 531 static int
 532 _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags)
 533 {
 534
 535         /* Reserve Necessary Bounce Pages */
 536         mtx_lock(&bounce_lock);
 537         if (flags & BUS_DMA_NOWAIT) {
 538                 if (reserve_bounce_pages(dmat, map, 0) != 0) {
 539                         mtx_unlock(&bounce_lock);
 540                         return (ENOMEM);
 541                 }
 542         } else {
 543                 if (reserve_bounce_pages(dmat, map, 1) != 0) {
 544                         /* Queue us for resources */
 545                         STAILQ_INSERT_TAIL(&bounce_map_waitinglist, map, links);
 546                         mtx_unlock(&bounce_lock);
 547                         return (EINPROGRESS);
 548                 }
 549         }
 550         mtx_unlock(&bounce_lock);
 551
 552         return (0);
 553 }
 554
 555 /*
 556  * Add a single contiguous physical range to the segment list.
 557  */
 558 static int
 559 _bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr,
 560     bus_size_t sgsize, bus_dma_segment_t *segs, int *segp)
 561 {
 562         bus_addr_t baddr, bmask;
 563         int seg;
 564
 565         /*
 566          * Make sure we don't cross any boundaries.
 567          */
 568         bmask = ~(dmat->common.boundary - 1);
 569         if (dmat->common.boundary > 0) {
 570                 baddr = (curaddr + dmat->common.boundary) & bmask;
 571                 if (sgsize > (baddr - curaddr))
 572                         sgsize = (baddr - curaddr);
 573         }
 574
 575         /*
 576          * Insert chunk into a segment, coalescing with
 577          * previous segment if possible.
 578          */
 579         seg = *segp;
 580         if (seg == -1) {
 581                 seg = 0;
 582                 segs[seg].ds_addr = curaddr;
 583                 segs[seg].ds_len = sgsize;
 584         } else {
 585                 if (curaddr == segs[seg].ds_addr + segs[seg].ds_len &&
 586                     (segs[seg].ds_len + sgsize) <= dmat->common.maxsegsz &&
 587                     (dmat->common.boundary == 0 ||
 588                      (segs[seg].ds_addr & bmask) == (curaddr & bmask)))
 589                         segs[seg].ds_len += sgsize;
 590                 else {
 591                         if (++seg >= dmat->common.nsegments)
 592                                 return (0);
 593                         segs[seg].ds_addr = curaddr;
 594                         segs[seg].ds_len = sgsize;
 595                 }
 596         }
 597         *segp = seg;
 598         return (sgsize);
 599 }
 600
 601 /*
 602  * Utility function to load a physical buffer.  segp contains
 603  * the starting segment on entrace, and the ending segment on exit.
 604  */
 605 static int
 606 bounce_bus_dmamap_load_phys(bus_dma_tag_t dmat, bus_dmamap_t map,
 607     vm_paddr_t buf, bus_size_t buflen, int flags, bus_dma_segment_t *segs,
 608     int *segp)
 609 {
 610         bus_size_t sgsize;
 611         bus_addr_t curaddr;
 612         int error;
 613
 614         if (map == NULL)
 615                 map = &nobounce_dmamap;
 616
 617         if (segs == NULL)
 618                 segs = dmat->segments;
 619
 620         if ((dmat->bounce_flags & BUS_DMA_COULD_BOUNCE) != 0) {
 621                 _bus_dmamap_count_phys(dmat, map, buf, buflen, flags);
 622                 if (map->pagesneeded != 0) {
 623                         error = _bus_dmamap_reserve_pages(dmat, map, flags);
 624                         if (error)
 625                                 return (error);
 626                 }
 627         }
 628
 629         while (buflen > 0) {
 630                 curaddr = buf;
 631                 sgsize = MIN(buflen, dmat->common.maxsegsz);
 632                 if (((dmat->bounce_flags & BUS_DMA_COULD_BOUNCE) != 0) &&
 633                     map->pagesneeded != 0 &&
 634                     bus_dma_run_filter(&dmat->common, curaddr)) {
 635                         sgsize = MIN(sgsize, PAGE_SIZE);
 636                         curaddr = add_bounce_page(dmat, map, 0, curaddr,
 637                             sgsize);
 638                 }
 639                 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
 640                     segp);
 641                 if (sgsize == 0)
 642                         break;
 643                 buf += sgsize;
 644                 buflen -= sgsize;
 645         }
 646
 647         /*
 648          * Did we fit?
 649          */
 650         return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
 651 }
 652
 653 /*
 654  * Utility function to load a linear buffer.  segp contains
 655  * the starting segment on entrace, and the ending segment on exit.
 656  */
 657 static int
 658 bounce_bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf,
 659     bus_size_t buflen, pmap_t pmap, int flags, bus_dma_segment_t *segs,
 660     int *segp)
 661 {
 662         bus_size_t sgsize, max_sgsize;
 663         bus_addr_t curaddr;
 664         vm_offset_t vaddr;
 665         int error;
 666
 667         if (map == NULL)
 668                 map = &nobounce_dmamap;
 669
 670         if (segs == NULL)
 671                 segs = dmat->segments;
 672
 673         if ((dmat->bounce_flags & BUS_DMA_COULD_BOUNCE) != 0) {
 674                 _bus_dmamap_count_pages(dmat, map, pmap, buf, buflen, flags);
 675                 if (map->pagesneeded != 0) {
 676                         error = _bus_dmamap_reserve_pages(dmat, map, flags);
 677                         if (error)
 678                                 return (error);
 679                 }
 680         }
 681
 682         vaddr = (vm_offset_t)buf;
 683         while (buflen > 0) {
 684                 /*
 685                  * Get the physical address for this segment.
 686                  */
 687                 if (pmap == kernel_pmap)
 688                         curaddr = pmap_kextract(vaddr);
 689                 else
 690                         curaddr = pmap_extract(pmap, vaddr);
 691
 692                 /*
 693                  * Compute the segment size, and adjust counts.
 694                  */
 695                 max_sgsize = MIN(buflen, dmat->common.maxsegsz);
 696                 sgsize = PAGE_SIZE - ((vm_offset_t)curaddr & PAGE_MASK);
 697                 if (((dmat->bounce_flags & BUS_DMA_COULD_BOUNCE) != 0) &&
 698                     map->pagesneeded != 0 &&
 699                     bus_dma_run_filter(&dmat->common, curaddr)) {
 700                         sgsize = roundup2(sgsize, dmat->common.alignment);
 701                         sgsize = MIN(sgsize, max_sgsize);
 702                         curaddr = add_bounce_page(dmat, map, vaddr, curaddr,
 703                             sgsize);
 704                 } else {
 705                         sgsize = MIN(sgsize, max_sgsize);
 706                 }
 707                 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs,
 708                     segp);
 709                 if (sgsize == 0)
 710                         break;
 711                 vaddr += sgsize;
 712                 buflen -= sgsize;
 713         }
 714
 715         /*
 716          * Did we fit?
 717          */
 718         return (buflen != 0 ? EFBIG : 0); /* XXX better return value here? */
 719 }
 720
 721 static void
 722 bounce_bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map,
 723     struct memdesc *mem, bus_dmamap_callback_t *callback, void *callback_arg)
 724 {
 725
 726         if (map == NULL)
 727                 return;
 728         map->mem = *mem;
 729         map->dmat = dmat;
 730         map->callback = callback;
 731         map->callback_arg = callback_arg;
 732 }
 733
 734 static bus_dma_segment_t *
 735 bounce_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map,
 736     bus_dma_segment_t *segs, int nsegs, int error)
 737 {
 738
 739         if (segs == NULL)
 740                 segs = dmat->segments;
 741         return (segs);
 742 }
 743
 744 /*
 745  * Release the mapping held by map.
 746  */
 747 static void
 748 bounce_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map)
 749 {
 750         struct bounce_page *bpage;
 751
 752         while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
 753                 STAILQ_REMOVE_HEAD(&map->bpages, links);
 754                 free_bounce_page(dmat, bpage);
 755         }
 756 }
 757
 758 static void
 759 bounce_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map,
 760     bus_dmasync_op_t op)
 761 {
 762         struct bounce_page *bpage;
 763
 764         /*
 765          * XXX ARM64TODO:
 766          * This bus_dma implementation requires IO-Coherent architecutre.
 767          * If IO-Coherency is not guaranteed, cache operations have to be
 768          * added to this function.
 769          */
 770
 771         if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) {
 772                 /*
 773                  * Handle data bouncing.  We might also
 774                  * want to add support for invalidating
 775                  * the caches on broken hardware
 776                  */
 777                 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x "
 778                     "performing bounce", __func__, dmat,
 779                     dmat->common.flags, op);
 780
 781                 if ((op & BUS_DMASYNC_PREWRITE) != 0) {
 782                         while (bpage != NULL) {
 783                                 if (bpage->datavaddr != 0) {
 784                                         bcopy((void *)bpage->datavaddr,
 785                                             (void *)bpage->vaddr,
 786                                             bpage->datacount);
 787                                 } else {
 788                                         physcopyout(bpage->dataaddr,
 789                                             (void *)bpage->vaddr,
 790                                             bpage->datacount);
 791                                 }
 792                                 bpage = STAILQ_NEXT(bpage, links);
 793                         }
 794                         dmat->bounce_zone->total_bounced++;
 795                 }
 796
 797                 if ((op & BUS_DMASYNC_POSTREAD) != 0) {
 798                         while (bpage != NULL) {
 799                                 if (bpage->datavaddr != 0) {
 800                                         bcopy((void *)bpage->vaddr,
 801                                             (void *)bpage->datavaddr,
 802                                             bpage->datacount);
 803                                 } else {
 804                                         physcopyin((void *)bpage->vaddr,
 805                                             bpage->dataaddr,
 806                                             bpage->datacount);
 807                                 }
 808                                 bpage = STAILQ_NEXT(bpage, links);
 809                         }
 810                         dmat->bounce_zone->total_bounced++;
 811                 }
 812         }
 813 }
 814
 815 static void
 816 init_bounce_pages(void *dummy __unused)
 817 {
 818
 819         total_bpages = 0;
 820         STAILQ_INIT(&bounce_zone_list);
 821         STAILQ_INIT(&bounce_map_waitinglist);
 822         STAILQ_INIT(&bounce_map_callbacklist);
 823         mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF);
 824 }
 825 SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL);
 826
 827 static struct sysctl_ctx_list *
 828 busdma_sysctl_tree(struct bounce_zone *bz)
 829 {
 830         return (&bz->sysctl_tree);
 831 }
 832
 833 static struct sysctl_oid *
 834 busdma_sysctl_tree_top(struct bounce_zone *bz)
 835 {
 836         return (bz->sysctl_tree_top);
 837 }
 838
 839 static int
 840 alloc_bounce_zone(bus_dma_tag_t dmat)
 841 {
 842         struct bounce_zone *bz;
 843
 844         /* Check to see if we already have a suitable zone */
 845         STAILQ_FOREACH(bz, &bounce_zone_list, links) {
 846                 if ((dmat->common.alignment <= bz->alignment) &&
 847                     (dmat->common.lowaddr >= bz->lowaddr)) {
 848                         dmat->bounce_zone = bz;
 849                         return (0);
 850                 }
 851         }
 852
 853         if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF,
 854             M_NOWAIT | M_ZERO)) == NULL)
 855                 return (ENOMEM);
 856
 857         STAILQ_INIT(&bz->bounce_page_list);
 858         bz->free_bpages = 0;
 859         bz->reserved_bpages = 0;
 860         bz->active_bpages = 0;
 861         bz->lowaddr = dmat->common.lowaddr;
 862         bz->alignment = MAX(dmat->common.alignment, PAGE_SIZE);
 863         bz->map_count = 0;
 864         snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount);
 865         busdma_zonecount++;
 866         snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr);
 867         STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links);
 868         dmat->bounce_zone = bz;
 869
 870         sysctl_ctx_init(&bz->sysctl_tree);
 871         bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree,
 872             SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid,
 873             CTLFLAG_RD, 0, "");
 874         if (bz->sysctl_tree_top == NULL) {
 875                 sysctl_ctx_free(&bz->sysctl_tree);
 876                 return (0);     /* XXX error code? */
 877         }
 878
 879         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
 880             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 881             "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0,
 882             "Total bounce pages");
 883         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
 884             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 885             "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0,
 886             "Free bounce pages");
 887         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
 888             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 889             "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0,
 890             "Reserved bounce pages");
 891         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
 892             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 893             "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0,
 894             "Active bounce pages");
 895         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
 896             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 897             "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0,
 898             "Total bounce requests");
 899         SYSCTL_ADD_INT(busdma_sysctl_tree(bz),
 900             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 901             "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0,
 902             "Total bounce requests that were deferred");
 903         SYSCTL_ADD_STRING(busdma_sysctl_tree(bz),
 904             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 905             "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, "");
 906         SYSCTL_ADD_UAUTO(busdma_sysctl_tree(bz),
 907             SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO,
 908             "alignment", CTLFLAG_RD, &bz->alignment, "");
 909
 910         return (0);
 911 }
 912
 913 static int
 914 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages)
 915 {
 916         struct bounce_zone *bz;
 917         int count;
 918
 919         bz = dmat->bounce_zone;
 920         count = 0;
 921         while (numpages > 0) {
 922                 struct bounce_page *bpage;
 923
 924                 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF,
 925                                                      M_NOWAIT | M_ZERO);
 926
 927                 if (bpage == NULL)
 928                         break;
 929                 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF,
 930                     M_NOWAIT, 0ul, bz->lowaddr, PAGE_SIZE, 0);
 931                 if (bpage->vaddr == 0) {
 932                         free(bpage, M_DEVBUF);
 933                         break;
 934                 }
 935                 bpage->busaddr = pmap_kextract(bpage->vaddr);
 936                 mtx_lock(&bounce_lock);
 937                 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links);
 938                 total_bpages++;
 939                 bz->total_bpages++;
 940                 bz->free_bpages++;
 941                 mtx_unlock(&bounce_lock);
 942                 count++;
 943                 numpages--;
 944         }
 945         return (count);
 946 }
 947
 948 static int
 949 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit)
 950 {
 951         struct bounce_zone *bz;
 952         int pages;
 953
 954         mtx_assert(&bounce_lock, MA_OWNED);
 955         bz = dmat->bounce_zone;
 956         pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved);
 957         if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages))
 958                 return (map->pagesneeded - (map->pagesreserved + pages));
 959         bz->free_bpages -= pages;
 960         bz->reserved_bpages += pages;
 961         map->pagesreserved += pages;
 962         pages = map->pagesneeded - map->pagesreserved;
 963
 964         return (pages);
 965 }
 966
 967 static bus_addr_t
 968 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr,
 969                 bus_addr_t addr, bus_size_t size)
 970 {
 971         struct bounce_zone *bz;
 972         struct bounce_page *bpage;
 973
 974         KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag"));
 975         KASSERT(map != NULL && map != &nobounce_dmamap,
 976             ("add_bounce_page: bad map %p", map));
 977
 978         bz = dmat->bounce_zone;
 979         if (map->pagesneeded == 0)
 980                 panic("add_bounce_page: map doesn't need any pages");
 981         map->pagesneeded--;
 982
 983         if (map->pagesreserved == 0)
 984                 panic("add_bounce_page: map doesn't need any pages");
 985         map->pagesreserved--;
 986
 987         mtx_lock(&bounce_lock);
 988         bpage = STAILQ_FIRST(&bz->bounce_page_list);
 989         if (bpage == NULL)
 990                 panic("add_bounce_page: free page list is empty");
 991
 992         STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links);
 993         bz->reserved_bpages--;
 994         bz->active_bpages++;
 995         mtx_unlock(&bounce_lock);
 996
 997         if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
 998                 /* Page offset needs to be preserved. */
 999                 bpage->vaddr |= addr & PAGE_MASK;
1000                 bpage->busaddr |= addr & PAGE_MASK;
1001         }
1002         bpage->datavaddr = vaddr;
1003         bpage->dataaddr = addr;
1004         bpage->datacount = size;
1005         STAILQ_INSERT_TAIL(&(map->bpages), bpage, links);
1006         return (bpage->busaddr);
1007 }
1008
1009 static void
1010 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage)
1011 {
1012         struct bus_dmamap *map;
1013         struct bounce_zone *bz;
1014
1015         bz = dmat->bounce_zone;
1016         bpage->datavaddr = 0;
1017         bpage->datacount = 0;
1018         if (dmat->common.flags & BUS_DMA_KEEP_PG_OFFSET) {
1019                 /*
1020                  * Reset the bounce page to start at offset 0.  Other uses
1021                  * of this bounce page may need to store a full page of
1022                  * data and/or assume it starts on a page boundary.
1023                  */
1024                 bpage->vaddr &= ~PAGE_MASK;
1025                 bpage->busaddr &= ~PAGE_MASK;
1026         }
1027
1028         mtx_lock(&bounce_lock);
1029         STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links);
1030         bz->free_bpages++;
1031         bz->active_bpages--;
1032         if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) {
1033                 if (reserve_bounce_pages(map->dmat, map, 1) == 0) {
1034                         STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links);
1035                         STAILQ_INSERT_TAIL(&bounce_map_callbacklist,
1036                             map, links);
1037                         busdma_swi_pending = 1;
1038                         bz->total_deferred++;
1039                         swi_sched(vm_ih, 0);
1040                 }
1041         }
1042         mtx_unlock(&bounce_lock);
1043 }
1044
1045 void
1046 busdma_swi(void)
1047 {
1048         bus_dma_tag_t dmat;
1049         struct bus_dmamap *map;
1050
1051         mtx_lock(&bounce_lock);
1052         while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) {
1053                 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links);
1054                 mtx_unlock(&bounce_lock);
1055                 dmat = map->dmat;
1056                 (dmat->common.lockfunc)(dmat->common.lockfuncarg, BUS_DMA_LOCK);
1057                 bus_dmamap_load_mem(map->dmat, map, &map->mem,
1058                     map->callback, map->callback_arg, BUS_DMA_WAITOK);
1059                 (dmat->common.lockfunc)(dmat->common.lockfuncarg,
1060                     BUS_DMA_UNLOCK);
1061                 mtx_lock(&bounce_lock);
1062         }
1063         mtx_unlock(&bounce_lock);
1064 }
1065
1066 struct bus_dma_impl bus_dma_bounce_impl = {
1067         .tag_create = bounce_bus_dma_tag_create,
1068         .tag_destroy = bounce_bus_dma_tag_destroy,
1069         .map_create = bounce_bus_dmamap_create,
1070         .map_destroy = bounce_bus_dmamap_destroy,
1071         .mem_alloc = bounce_bus_dmamem_alloc,
1072         .mem_free = bounce_bus_dmamem_free,
1073         .load_phys = bounce_bus_dmamap_load_phys,
1074         .load_buffer = bounce_bus_dmamap_load_buffer,
1075         .load_ma = bus_dmamap_load_ma_triv,
1076         .map_waitok = bounce_bus_dmamap_waitok,
1077         .map_complete = bounce_bus_dmamap_complete,
1078         .map_unload = bounce_bus_dmamap_unload,
1079         .map_sync = bounce_bus_dmamap_sync
1080 };