busdma_machdep-v6.c revision 318977
1/*- 2 * Copyright (c) 2012-2014 Ian Lepore 3 * Copyright (c) 2010 Mark Tinguely 4 * Copyright (c) 2004 Olivier Houchard 5 * Copyright (c) 2002 Peter Grehan 6 * Copyright (c) 1997, 1998 Justin T. Gibbs. 7 * All rights reserved. 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 * From i386/busdma_machdep.c 191438 2009-04-23 20:24:19Z jhb 31 */ 32 33#include <sys/cdefs.h> 34__FBSDID("$FreeBSD: stable/10/sys/arm/arm/busdma_machdep-v6.c 318977 2017-05-27 08:17:59Z hselasky $"); 35 36#define _ARM32_BUS_DMA_PRIVATE 37#include <sys/param.h> 38#include <sys/kdb.h> 39#include <ddb/ddb.h> 40#include <ddb/db_output.h> 41#include <sys/systm.h> 42#include <sys/malloc.h> 43#include <sys/bus.h> 44#include <sys/busdma_bufalloc.h> 45#include <sys/interrupt.h> 46#include <sys/kernel.h> 47#include <sys/ktr.h> 48#include <sys/lock.h> 49#include <sys/memdesc.h> 50#include <sys/proc.h> 51#include <sys/mutex.h> 52#include <sys/sysctl.h> 53#include <sys/uio.h> 54 55#include <vm/vm.h> 56#include <vm/vm_page.h> 57#include <vm/vm_map.h> 58#include <vm/vm_extern.h> 59#include <vm/vm_kern.h> 60 61#include <machine/atomic.h> 62#include <machine/bus.h> 63#include <machine/cpufunc.h> 64#include <machine/md_var.h> 65 66#define MAX_BPAGES 64 67#define MAX_DMA_SEGMENTS 4096 68#define BUS_DMA_EXCL_BOUNCE BUS_DMA_BUS2 69#define BUS_DMA_ALIGN_BOUNCE BUS_DMA_BUS3 70#define BUS_DMA_COULD_BOUNCE (BUS_DMA_EXCL_BOUNCE | BUS_DMA_ALIGN_BOUNCE) 71#define BUS_DMA_MIN_ALLOC_COMP BUS_DMA_BUS4 72 73struct bounce_zone; 74 75struct bus_dma_tag { 76 bus_dma_tag_t parent; 77 bus_size_t alignment; 78 bus_size_t boundary; 79 bus_addr_t lowaddr; 80 bus_addr_t highaddr; 81 bus_dma_filter_t *filter; 82 void *filterarg; 83 bus_size_t maxsize; 84 u_int nsegments; 85 bus_size_t maxsegsz; 86 int flags; 87 int ref_count; 88 int map_count; 89 bus_dma_lock_t *lockfunc; 90 void *lockfuncarg; 91 struct bounce_zone *bounce_zone; 92 /* 93 * DMA range for this tag. If the page doesn't fall within 94 * one of these ranges, an error is returned. The caller 95 * may then decide what to do with the transfer. If the 96 * range pointer is NULL, it is ignored. 97 */ 98 struct arm32_dma_range *ranges; 99 int _nranges; 100}; 101 102struct bounce_page { 103 vm_offset_t vaddr; /* kva of bounce buffer */ 104 bus_addr_t busaddr; /* Physical address */ 105 vm_offset_t datavaddr; /* kva of client data */ 106 bus_addr_t dataaddr; /* client physical address */ 107 bus_size_t datacount; /* client data count */ 108 STAILQ_ENTRY(bounce_page) links; 109}; 110 111struct sync_list { 112 vm_offset_t vaddr; /* kva of bounce buffer */ 113 bus_addr_t busaddr; /* Physical address */ 114 bus_size_t datacount; /* client data count */ 115}; 116 117int busdma_swi_pending; 118 119struct bounce_zone { 120 STAILQ_ENTRY(bounce_zone) links; 121 STAILQ_HEAD(bp_list, bounce_page) bounce_page_list; 122 int total_bpages; 123 int free_bpages; 124 int reserved_bpages; 125 int active_bpages; 126 int total_bounced; 127 int total_deferred; 128 int map_count; 129 bus_size_t alignment; 130 bus_addr_t lowaddr; 131 char zoneid[8]; 132 char lowaddrid[20]; 133 struct sysctl_ctx_list sysctl_tree; 134 struct sysctl_oid *sysctl_tree_top; 135}; 136 137static struct mtx bounce_lock; 138static int total_bpages; 139static int busdma_zonecount; 140static uint32_t tags_total; 141static uint32_t maps_total; 142static uint32_t maps_dmamem; 143static uint32_t maps_coherent; 144static uint64_t maploads_total; 145static uint64_t maploads_bounced; 146static uint64_t maploads_coherent; 147static uint64_t maploads_dmamem; 148static uint64_t maploads_mbuf; 149static uint64_t maploads_physmem; 150 151static STAILQ_HEAD(, bounce_zone) bounce_zone_list; 152 153SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters"); 154SYSCTL_UINT(_hw_busdma, OID_AUTO, tags_total, CTLFLAG_RD, &tags_total, 0, 155 "Number of active tags"); 156SYSCTL_UINT(_hw_busdma, OID_AUTO, maps_total, CTLFLAG_RD, &maps_total, 0, 157 "Number of active maps"); 158SYSCTL_UINT(_hw_busdma, OID_AUTO, maps_dmamem, CTLFLAG_RD, &maps_dmamem, 0, 159 "Number of active maps for bus_dmamem_alloc buffers"); 160SYSCTL_UINT(_hw_busdma, OID_AUTO, maps_coherent, CTLFLAG_RD, &maps_coherent, 0, 161 "Number of active maps with BUS_DMA_COHERENT flag set"); 162SYSCTL_UQUAD(_hw_busdma, OID_AUTO, maploads_total, CTLFLAG_RD, &maploads_total, 0, 163 "Number of load operations performed"); 164SYSCTL_UQUAD(_hw_busdma, OID_AUTO, maploads_bounced, CTLFLAG_RD, &maploads_bounced, 0, 165 "Number of load operations that used bounce buffers"); 166SYSCTL_UQUAD(_hw_busdma, OID_AUTO, maploads_coherent, CTLFLAG_RD, &maploads_dmamem, 0, 167 "Number of load operations on BUS_DMA_COHERENT memory"); 168SYSCTL_UQUAD(_hw_busdma, OID_AUTO, maploads_dmamem, CTLFLAG_RD, &maploads_dmamem, 0, 169 "Number of load operations on bus_dmamem_alloc buffers"); 170SYSCTL_UQUAD(_hw_busdma, OID_AUTO, maploads_mbuf, CTLFLAG_RD, &maploads_mbuf, 0, 171 "Number of load operations for mbufs"); 172SYSCTL_UQUAD(_hw_busdma, OID_AUTO, maploads_physmem, CTLFLAG_RD, &maploads_physmem, 0, 173 "Number of load operations on physical buffers"); 174SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bpages, 0, 175 "Total bounce pages"); 176 177struct bus_dmamap { 178 struct bp_list bpages; 179 int pagesneeded; 180 int pagesreserved; 181 bus_dma_tag_t dmat; 182 struct memdesc mem; 183 pmap_t pmap; 184 bus_dmamap_callback_t *callback; 185 void *callback_arg; 186 int flags; 187#define DMAMAP_COHERENT (1 << 0) 188#define DMAMAP_DMAMEM_ALLOC (1 << 1) 189#define DMAMAP_MBUF (1 << 2) 190 STAILQ_ENTRY(bus_dmamap) links; 191 bus_dma_segment_t *segments; 192 int sync_count; 193 struct sync_list slist[]; 194}; 195 196static STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist; 197static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist; 198 199static void init_bounce_pages(void *dummy); 200static int alloc_bounce_zone(bus_dma_tag_t dmat); 201static int alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages); 202static int reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, 203 int commit); 204static bus_addr_t add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, 205 vm_offset_t vaddr, bus_addr_t addr, 206 bus_size_t size); 207static void free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage); 208static void _bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, 209 void *buf, bus_size_t buflen, int flags); 210static void _bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, 211 vm_paddr_t buf, bus_size_t buflen, int flags); 212static int _bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, 213 int flags); 214 215static busdma_bufalloc_t coherent_allocator; /* Cache of coherent buffers */ 216static busdma_bufalloc_t standard_allocator; /* Cache of standard buffers */ 217static void 218busdma_init(void *dummy) 219{ 220 int uma_flags; 221 222 uma_flags = 0; 223 224 /* Create a cache of buffers in standard (cacheable) memory. */ 225 standard_allocator = busdma_bufalloc_create("buffer", 226 arm_dcache_align, /* minimum_alignment */ 227 NULL, /* uma_alloc func */ 228 NULL, /* uma_free func */ 229 uma_flags); /* uma_zcreate_flags */ 230 231#ifdef INVARIANTS 232 /* 233 * Force UMA zone to allocate service structures like 234 * slabs using own allocator. uma_debug code performs 235 * atomic ops on uma_slab_t fields and safety of this 236 * operation is not guaranteed for write-back caches 237 */ 238 uma_flags = UMA_ZONE_OFFPAGE; 239#endif 240 /* 241 * Create a cache of buffers in uncacheable memory, to implement the 242 * BUS_DMA_COHERENT (and potentially BUS_DMA_NOCACHE) flag. 243 */ 244 coherent_allocator = busdma_bufalloc_create("coherent", 245 arm_dcache_align, /* minimum_alignment */ 246 busdma_bufalloc_alloc_uncacheable, 247 busdma_bufalloc_free_uncacheable, 248 uma_flags); /* uma_zcreate_flags */ 249} 250 251/* 252 * This init historically used SI_SUB_VM, but now the init code requires 253 * malloc(9) using M_DEVBUF memory, which is set up later than SI_SUB_VM, by 254 * SI_SUB_KMEM and SI_ORDER_SECOND, so we'll go right after that by using 255 * SI_SUB_KMEM and SI_ORDER_THIRD. 256 */ 257SYSINIT(busdma, SI_SUB_KMEM, SI_ORDER_THIRD, busdma_init, NULL); 258 259static int 260exclusion_bounce_check(vm_offset_t lowaddr, vm_offset_t highaddr) 261{ 262 int i; 263 for (i = 0; phys_avail[i] && phys_avail[i + 1]; i += 2) { 264 if ((lowaddr >= phys_avail[i] && lowaddr < phys_avail[i + 1]) || 265 (lowaddr < phys_avail[i] && highaddr >= phys_avail[i])) 266 return (1); 267 } 268 return (0); 269} 270 271/* 272 * Return true if the tag has an exclusion zone that could lead to bouncing. 273 */ 274static __inline int 275exclusion_bounce(bus_dma_tag_t dmat) 276{ 277 278 return (dmat->flags & BUS_DMA_EXCL_BOUNCE); 279} 280 281/* 282 * Return true if the given address does not fall on the alignment boundary. 283 */ 284static __inline int 285alignment_bounce(bus_dma_tag_t dmat, bus_addr_t addr) 286{ 287 288 return (addr & (dmat->alignment - 1)); 289} 290 291/* 292 * Return true if the DMA should bounce because the start or end does not fall 293 * on a cacheline boundary (which would require a partial cacheline flush). 294 * COHERENT memory doesn't trigger cacheline flushes. Memory allocated by 295 * bus_dmamem_alloc() is always aligned to cacheline boundaries, and there's a 296 * strict rule that such memory cannot be accessed by the CPU while DMA is in 297 * progress (or by multiple DMA engines at once), so that it's always safe to do 298 * full cacheline flushes even if that affects memory outside the range of a 299 * given DMA operation that doesn't involve the full allocated buffer. If we're 300 * mapping an mbuf, that follows the same rules as a buffer we allocated. 301 */ 302static __inline int 303cacheline_bounce(bus_dmamap_t map, bus_addr_t addr, bus_size_t size) 304{ 305 306 if (map->flags & (DMAMAP_DMAMEM_ALLOC | DMAMAP_COHERENT | DMAMAP_MBUF)) 307 return (0); 308 return ((addr | size) & arm_dcache_align_mask); 309} 310 311/* 312 * Return true if we might need to bounce the DMA described by addr and size. 313 * 314 * This is used to quick-check whether we need to do the more expensive work of 315 * checking the DMA page-by-page looking for alignment and exclusion bounces. 316 * 317 * Note that the addr argument might be either virtual or physical. It doesn't 318 * matter because we only look at the low-order bits, which are the same in both 319 * address spaces. 320 */ 321static __inline int 322might_bounce(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t addr, 323 bus_size_t size) 324{ 325 326 return ((dmat->flags & BUS_DMA_EXCL_BOUNCE) || 327 alignment_bounce(dmat, addr) || 328 cacheline_bounce(map, addr, size)); 329} 330 331/* 332 * Return true if we must bounce the DMA described by paddr and size. 333 * 334 * Bouncing can be triggered by DMA that doesn't begin and end on cacheline 335 * boundaries, or doesn't begin on an alignment boundary, or falls within the 336 * exclusion zone of any tag in the ancestry chain. 337 * 338 * For exclusions, walk the chain of tags comparing paddr to the exclusion zone 339 * within each tag. If the tag has a filter function, use it to decide whether 340 * the DMA needs to bounce, otherwise any DMA within the zone bounces. 341 */ 342static int 343must_bounce(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t paddr, 344 bus_size_t size) 345{ 346 347 if (cacheline_bounce(map, paddr, size)) 348 return (1); 349 350 /* 351 * The tag already contains ancestors' alignment restrictions so this 352 * check doesn't need to be inside the loop. 353 */ 354 if (alignment_bounce(dmat, paddr)) 355 return (1); 356 357 /* 358 * Even though each tag has an exclusion zone that is a superset of its 359 * own and all its ancestors' exclusions, the exclusion zone of each tag 360 * up the chain must be checked within the loop, because the busdma 361 * rules say the filter function is called only when the address lies 362 * within the low-highaddr range of the tag that filterfunc belongs to. 363 */ 364 while (dmat != NULL && exclusion_bounce(dmat)) { 365 if ((paddr >= dmat->lowaddr && paddr <= dmat->highaddr) && 366 (dmat->filter == NULL || 367 dmat->filter(dmat->filterarg, paddr) != 0)) 368 return (1); 369 dmat = dmat->parent; 370 } 371 372 return (0); 373} 374 375static __inline struct arm32_dma_range * 376_bus_dma_inrange(struct arm32_dma_range *ranges, int nranges, 377 bus_addr_t curaddr) 378{ 379 struct arm32_dma_range *dr; 380 int i; 381 382 for (i = 0, dr = ranges; i < nranges; i++, dr++) { 383 if (curaddr >= dr->dr_sysbase && 384 round_page(curaddr) <= (dr->dr_sysbase + dr->dr_len)) 385 return (dr); 386 } 387 388 return (NULL); 389} 390 391/* 392 * Convenience function for manipulating driver locks from busdma (during 393 * busdma_swi, for example). Drivers that don't provide their own locks 394 * should specify &Giant to dmat->lockfuncarg. Drivers that use their own 395 * non-mutex locking scheme don't have to use this at all. 396 */ 397void 398busdma_lock_mutex(void *arg, bus_dma_lock_op_t op) 399{ 400 struct mtx *dmtx; 401 402 dmtx = (struct mtx *)arg; 403 switch (op) { 404 case BUS_DMA_LOCK: 405 mtx_lock(dmtx); 406 break; 407 case BUS_DMA_UNLOCK: 408 mtx_unlock(dmtx); 409 break; 410 default: 411 panic("Unknown operation 0x%x for busdma_lock_mutex!", op); 412 } 413} 414 415/* 416 * dflt_lock should never get called. It gets put into the dma tag when 417 * lockfunc == NULL, which is only valid if the maps that are associated 418 * with the tag are meant to never be defered. 419 * XXX Should have a way to identify which driver is responsible here. 420 */ 421static void 422dflt_lock(void *arg, bus_dma_lock_op_t op) 423{ 424 425 panic("driver error: busdma dflt_lock called"); 426} 427 428/* 429 * Allocate a device specific dma_tag. 430 */ 431int 432bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment, 433 bus_size_t boundary, bus_addr_t lowaddr, 434 bus_addr_t highaddr, bus_dma_filter_t *filter, 435 void *filterarg, bus_size_t maxsize, int nsegments, 436 bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc, 437 void *lockfuncarg, bus_dma_tag_t *dmat) 438{ 439 bus_dma_tag_t newtag; 440 int error = 0; 441 442#if 0 443 if (!parent) 444 parent = arm_root_dma_tag; 445#endif 446 447 /* Basic sanity checking */ 448 if (boundary != 0 && boundary < maxsegsz) 449 maxsegsz = boundary; 450 451 /* Return a NULL tag on failure */ 452 *dmat = NULL; 453 454 if (maxsegsz == 0) { 455 return (EINVAL); 456 } 457 458 newtag = (bus_dma_tag_t)malloc(sizeof(*newtag), M_DEVBUF, 459 M_ZERO | M_NOWAIT); 460 if (newtag == NULL) { 461 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d", 462 __func__, newtag, 0, error); 463 return (ENOMEM); 464 } 465 466 newtag->parent = parent; 467 newtag->alignment = alignment; 468 newtag->boundary = boundary; 469 newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1); 470 newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + 471 (PAGE_SIZE - 1); 472 newtag->filter = filter; 473 newtag->filterarg = filterarg; 474 newtag->maxsize = maxsize; 475 newtag->nsegments = nsegments; 476 newtag->maxsegsz = maxsegsz; 477 newtag->flags = flags; 478 newtag->ref_count = 1; /* Count ourself */ 479 newtag->map_count = 0; 480 newtag->ranges = bus_dma_get_range(); 481 newtag->_nranges = bus_dma_get_range_nb(); 482 if (lockfunc != NULL) { 483 newtag->lockfunc = lockfunc; 484 newtag->lockfuncarg = lockfuncarg; 485 } else { 486 newtag->lockfunc = dflt_lock; 487 newtag->lockfuncarg = NULL; 488 } 489 490 /* Take into account any restrictions imposed by our parent tag */ 491 if (parent != NULL) { 492 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr); 493 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr); 494 newtag->alignment = MAX(parent->alignment, newtag->alignment); 495 newtag->flags |= parent->flags & BUS_DMA_COULD_BOUNCE; 496 if (newtag->boundary == 0) 497 newtag->boundary = parent->boundary; 498 else if (parent->boundary != 0) 499 newtag->boundary = MIN(parent->boundary, 500 newtag->boundary); 501 if (newtag->filter == NULL) { 502 /* 503 * Short circuit to looking at our parent directly 504 * since we have encapsulated all of its information 505 */ 506 newtag->filter = parent->filter; 507 newtag->filterarg = parent->filterarg; 508 newtag->parent = parent->parent; 509 } 510 if (newtag->parent != NULL) 511 atomic_add_int(&parent->ref_count, 1); 512 } 513 514 if (exclusion_bounce_check(newtag->lowaddr, newtag->highaddr)) 515 newtag->flags |= BUS_DMA_EXCL_BOUNCE; 516 if (alignment_bounce(newtag, 1)) 517 newtag->flags |= BUS_DMA_ALIGN_BOUNCE; 518 519 /* 520 * Any request can auto-bounce due to cacheline alignment, in addition 521 * to any alignment or boundary specifications in the tag, so if the 522 * ALLOCNOW flag is set, there's always work to do. 523 */ 524 if ((flags & BUS_DMA_ALLOCNOW) != 0) { 525 struct bounce_zone *bz; 526 /* 527 * Round size up to a full page, and add one more page because 528 * there can always be one more boundary crossing than the 529 * number of pages in a transfer. 530 */ 531 maxsize = roundup2(maxsize, PAGE_SIZE) + PAGE_SIZE; 532 533 if ((error = alloc_bounce_zone(newtag)) != 0) { 534 free(newtag, M_DEVBUF); 535 return (error); 536 } 537 bz = newtag->bounce_zone; 538 539 if (ptoa(bz->total_bpages) < maxsize) { 540 int pages; 541 542 pages = atop(maxsize) - bz->total_bpages; 543 544 /* Add pages to our bounce pool */ 545 if (alloc_bounce_pages(newtag, pages) < pages) 546 error = ENOMEM; 547 } 548 /* Performed initial allocation */ 549 newtag->flags |= BUS_DMA_MIN_ALLOC_COMP; 550 } else 551 newtag->bounce_zone = NULL; 552 553 if (error != 0) { 554 free(newtag, M_DEVBUF); 555 } else { 556 atomic_add_32(&tags_total, 1); 557 *dmat = newtag; 558 } 559 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d", 560 __func__, newtag, (newtag != NULL ? newtag->flags : 0), error); 561 return (error); 562} 563 564int 565bus_dma_tag_destroy(bus_dma_tag_t dmat) 566{ 567 bus_dma_tag_t dmat_copy; 568 int error; 569 570 error = 0; 571 dmat_copy = dmat; 572 573 if (dmat != NULL) { 574 575 if (dmat->map_count != 0) { 576 error = EBUSY; 577 goto out; 578 } 579 580 while (dmat != NULL) { 581 bus_dma_tag_t parent; 582 583 parent = dmat->parent; 584 atomic_subtract_int(&dmat->ref_count, 1); 585 if (dmat->ref_count == 0) { 586 atomic_subtract_32(&tags_total, 1); 587 free(dmat, M_DEVBUF); 588 /* 589 * Last reference count, so 590 * release our reference 591 * count on our parent. 592 */ 593 dmat = parent; 594 } else 595 dmat = NULL; 596 } 597 } 598out: 599 CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error); 600 return (error); 601} 602 603static int allocate_bz_and_pages(bus_dma_tag_t dmat, bus_dmamap_t mapp) 604{ 605 struct bounce_zone *bz; 606 int maxpages; 607 int error; 608 609 if (dmat->bounce_zone == NULL) 610 if ((error = alloc_bounce_zone(dmat)) != 0) 611 return (error); 612 bz = dmat->bounce_zone; 613 /* Initialize the new map */ 614 STAILQ_INIT(&(mapp->bpages)); 615 616 /* 617 * Attempt to add pages to our pool on a per-instance basis up to a sane 618 * limit. Even if the tag isn't flagged as COULD_BOUNCE due to 619 * alignment and boundary constraints, it could still auto-bounce due to 620 * cacheline alignment, which requires at most two bounce pages. 621 */ 622 if (dmat->flags & BUS_DMA_COULD_BOUNCE) 623 maxpages = MAX_BPAGES; 624 else 625 maxpages = 2 * bz->map_count; 626 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 || 627 (bz->map_count > 0 && bz->total_bpages < maxpages)) { 628 int pages; 629 630 pages = atop(roundup2(dmat->maxsize, PAGE_SIZE)) + 1; 631 pages = MIN(maxpages - bz->total_bpages, pages); 632 pages = MAX(pages, 2); 633 if (alloc_bounce_pages(dmat, pages) < pages) 634 return (ENOMEM); 635 636 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) 637 dmat->flags |= BUS_DMA_MIN_ALLOC_COMP; 638 } 639 bz->map_count++; 640 return (0); 641} 642 643static bus_dmamap_t 644allocate_map(bus_dma_tag_t dmat, int mflags) 645{ 646 int mapsize, segsize; 647 bus_dmamap_t map; 648 649 /* 650 * Allocate the map. The map structure ends with an embedded 651 * variable-sized array of sync_list structures. Following that 652 * we allocate enough extra space to hold the array of bus_dma_segments. 653 */ 654 KASSERT(dmat->nsegments <= MAX_DMA_SEGMENTS, 655 ("cannot allocate %u dma segments (max is %u)", 656 dmat->nsegments, MAX_DMA_SEGMENTS)); 657 segsize = sizeof(struct bus_dma_segment) * dmat->nsegments; 658 mapsize = sizeof(*map) + sizeof(struct sync_list) * dmat->nsegments; 659 map = malloc(mapsize + segsize, M_DEVBUF, mflags | M_ZERO); 660 if (map == NULL) { 661 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, ENOMEM); 662 return (NULL); 663 } 664 map->segments = (bus_dma_segment_t *)((uintptr_t)map + mapsize); 665 return (map); 666} 667 668/* 669 * Allocate a handle for mapping from kva/uva/physical 670 * address space into bus device space. 671 */ 672int 673bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp) 674{ 675 bus_dmamap_t map; 676 int error = 0; 677 678 *mapp = map = allocate_map(dmat, M_NOWAIT); 679 if (map == NULL) { 680 CTR3(KTR_BUSDMA, "%s: tag %p error %d", __func__, dmat, ENOMEM); 681 return (ENOMEM); 682 } 683 684 /* 685 * Bouncing might be required if the driver asks for an exclusion 686 * region, a data alignment that is stricter than 1, or DMA that begins 687 * or ends with a partial cacheline. Whether bouncing will actually 688 * happen can't be known until mapping time, but we need to pre-allocate 689 * resources now because we might not be allowed to at mapping time. 690 */ 691 error = allocate_bz_and_pages(dmat, map); 692 if (error != 0) { 693 free(map, M_DEVBUF); 694 *mapp = NULL; 695 return (error); 696 } 697 if (map->flags & DMAMAP_COHERENT) 698 atomic_add_32(&maps_coherent, 1); 699 atomic_add_32(&maps_total, 1); 700 dmat->map_count++; 701 702 return (0); 703} 704 705/* 706 * Destroy a handle for mapping from kva/uva/physical 707 * address space into bus device space. 708 */ 709int 710bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map) 711{ 712 if (STAILQ_FIRST(&map->bpages) != NULL || map->sync_count != 0) { 713 CTR3(KTR_BUSDMA, "%s: tag %p error %d", 714 __func__, dmat, EBUSY); 715 return (EBUSY); 716 } 717 if (dmat->bounce_zone) 718 dmat->bounce_zone->map_count--; 719 if (map->flags & DMAMAP_COHERENT) 720 atomic_subtract_32(&maps_coherent, 1); 721 atomic_subtract_32(&maps_total, 1); 722 free(map, M_DEVBUF); 723 dmat->map_count--; 724 CTR2(KTR_BUSDMA, "%s: tag %p error 0", __func__, dmat); 725 return (0); 726} 727 728 729/* 730 * Allocate a piece of memory that can be efficiently mapped into 731 * bus device space based on the constraints lited in the dma tag. 732 * A dmamap to for use with dmamap_load is also allocated. 733 */ 734int 735bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags, 736 bus_dmamap_t *mapp) 737{ 738 busdma_bufalloc_t ba; 739 struct busdma_bufzone *bufzone; 740 bus_dmamap_t map; 741 vm_memattr_t memattr; 742 int mflags; 743 744 if (flags & BUS_DMA_NOWAIT) 745 mflags = M_NOWAIT; 746 else 747 mflags = M_WAITOK; 748 if (flags & BUS_DMA_ZERO) 749 mflags |= M_ZERO; 750 751 *mapp = map = allocate_map(dmat, mflags); 752 if (map == NULL) { 753 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d", 754 __func__, dmat, dmat->flags, ENOMEM); 755 return (ENOMEM); 756 } 757 map->flags = DMAMAP_DMAMEM_ALLOC; 758 759 /* Choose a busdma buffer allocator based on memory type flags. */ 760 if (flags & BUS_DMA_COHERENT) { 761 memattr = VM_MEMATTR_UNCACHEABLE; 762 ba = coherent_allocator; 763 map->flags |= DMAMAP_COHERENT; 764 } else { 765 memattr = VM_MEMATTR_DEFAULT; 766 ba = standard_allocator; 767 } 768 769 /* 770 * Try to find a bufzone in the allocator that holds a cache of buffers 771 * of the right size for this request. If the buffer is too big to be 772 * held in the allocator cache, this returns NULL. 773 */ 774 bufzone = busdma_bufalloc_findzone(ba, dmat->maxsize); 775 776 /* 777 * Allocate the buffer from the uma(9) allocator if... 778 * - It's small enough to be in the allocator (bufzone not NULL). 779 * - The alignment constraint isn't larger than the allocation size 780 * (the allocator aligns buffers to their size boundaries). 781 * - There's no need to handle lowaddr/highaddr exclusion zones. 782 * else allocate non-contiguous pages if... 783 * - The page count that could get allocated doesn't exceed 784 * nsegments also when the maximum segment size is less 785 * than PAGE_SIZE. 786 * - The alignment constraint isn't larger than a page boundary. 787 * - There are no boundary-crossing constraints. 788 * else allocate a block of contiguous pages because one or more of the 789 * constraints is something that only the contig allocator can fulfill. 790 */ 791 if (bufzone != NULL && dmat->alignment <= bufzone->size && 792 !exclusion_bounce(dmat)) { 793 *vaddr = uma_zalloc(bufzone->umazone, mflags); 794 } else if (dmat->nsegments >= 795 howmany(dmat->maxsize, MIN(dmat->maxsegsz, PAGE_SIZE)) && 796 dmat->alignment <= PAGE_SIZE && 797 (dmat->boundary % PAGE_SIZE) == 0) { 798 *vaddr = (void *)kmem_alloc_attr(kernel_arena, dmat->maxsize, 799 mflags, 0, dmat->lowaddr, memattr); 800 } else { 801 *vaddr = (void *)kmem_alloc_contig(kernel_arena, dmat->maxsize, 802 mflags, 0, dmat->lowaddr, dmat->alignment, dmat->boundary, 803 memattr); 804 } 805 806 807 if (*vaddr == NULL) { 808 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d", 809 __func__, dmat, dmat->flags, ENOMEM); 810 free(map, M_DEVBUF); 811 *mapp = NULL; 812 return (ENOMEM); 813 } else if ((uintptr_t)*vaddr & (dmat->alignment - 1)) { 814 printf("bus_dmamem_alloc failed to align memory properly.\n"); 815 } 816 if (map->flags & DMAMAP_COHERENT) 817 atomic_add_32(&maps_coherent, 1); 818 atomic_add_32(&maps_dmamem, 1); 819 atomic_add_32(&maps_total, 1); 820 dmat->map_count++; 821 822 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x error %d", 823 __func__, dmat, dmat->flags, 0); 824 return (0); 825} 826 827/* 828 * Free a piece of memory and it's allociated dmamap, that was allocated 829 * via bus_dmamem_alloc. Make the same choice for free/contigfree. 830 */ 831void 832bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map) 833{ 834 struct busdma_bufzone *bufzone; 835 busdma_bufalloc_t ba; 836 837 if (map->flags & DMAMAP_COHERENT) 838 ba = coherent_allocator; 839 else 840 ba = standard_allocator; 841 842 bufzone = busdma_bufalloc_findzone(ba, dmat->maxsize); 843 844 if (bufzone != NULL && dmat->alignment <= bufzone->size && 845 !exclusion_bounce(dmat)) 846 uma_zfree(bufzone->umazone, vaddr); 847 else 848 kmem_free(kernel_arena, (vm_offset_t)vaddr, dmat->maxsize); 849 850 dmat->map_count--; 851 if (map->flags & DMAMAP_COHERENT) 852 atomic_subtract_32(&maps_coherent, 1); 853 atomic_subtract_32(&maps_total, 1); 854 atomic_subtract_32(&maps_dmamem, 1); 855 free(map, M_DEVBUF); 856 CTR3(KTR_BUSDMA, "%s: tag %p flags 0x%x", __func__, dmat, dmat->flags); 857} 858 859static void 860_bus_dmamap_count_phys(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t buf, 861 bus_size_t buflen, int flags) 862{ 863 bus_addr_t curaddr; 864 bus_size_t sgsize; 865 866 if (map->pagesneeded == 0) { 867 CTR5(KTR_BUSDMA, "lowaddr= %d, boundary= %d, alignment= %d" 868 " map= %p, pagesneeded= %d", 869 dmat->lowaddr, dmat->boundary, dmat->alignment, 870 map, map->pagesneeded); 871 /* 872 * Count the number of bounce pages 873 * needed in order to complete this transfer 874 */ 875 curaddr = buf; 876 while (buflen != 0) { 877 sgsize = MIN(buflen, dmat->maxsegsz); 878 if (must_bounce(dmat, map, curaddr, sgsize) != 0) { 879 sgsize = MIN(sgsize, PAGE_SIZE); 880 map->pagesneeded++; 881 } 882 curaddr += sgsize; 883 buflen -= sgsize; 884 } 885 CTR1(KTR_BUSDMA, "pagesneeded= %d", map->pagesneeded); 886 } 887} 888 889static void 890_bus_dmamap_count_pages(bus_dma_tag_t dmat, bus_dmamap_t map, 891 void *buf, bus_size_t buflen, int flags) 892{ 893 vm_offset_t vaddr; 894 vm_offset_t vendaddr; 895 bus_addr_t paddr; 896 897 if (map->pagesneeded == 0) { 898 CTR5(KTR_BUSDMA, "lowaddr= %d, boundary= %d, alignment= %d" 899 " map= %p, pagesneeded= %d", 900 dmat->lowaddr, dmat->boundary, dmat->alignment, 901 map, map->pagesneeded); 902 /* 903 * Count the number of bounce pages 904 * needed in order to complete this transfer 905 */ 906 vaddr = (vm_offset_t)buf; 907 vendaddr = (vm_offset_t)buf + buflen; 908 909 while (vaddr < vendaddr) { 910 if (__predict_true(map->pmap == kernel_pmap)) 911 paddr = pmap_kextract(vaddr); 912 else 913 paddr = pmap_extract(map->pmap, vaddr); 914 if (must_bounce(dmat, map, paddr, 915 min(vendaddr - vaddr, (PAGE_SIZE - ((vm_offset_t)vaddr & 916 PAGE_MASK)))) != 0) { 917 map->pagesneeded++; 918 } 919 vaddr += (PAGE_SIZE - ((vm_offset_t)vaddr & PAGE_MASK)); 920 921 } 922 CTR1(KTR_BUSDMA, "pagesneeded= %d", map->pagesneeded); 923 } 924} 925 926static int 927_bus_dmamap_reserve_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int flags) 928{ 929 930 /* Reserve Necessary Bounce Pages */ 931 mtx_lock(&bounce_lock); 932 if (flags & BUS_DMA_NOWAIT) { 933 if (reserve_bounce_pages(dmat, map, 0) != 0) { 934 map->pagesneeded = 0; 935 mtx_unlock(&bounce_lock); 936 return (ENOMEM); 937 } 938 } else { 939 if (reserve_bounce_pages(dmat, map, 1) != 0) { 940 /* Queue us for resources */ 941 STAILQ_INSERT_TAIL(&bounce_map_waitinglist, map, links); 942 mtx_unlock(&bounce_lock); 943 return (EINPROGRESS); 944 } 945 } 946 mtx_unlock(&bounce_lock); 947 948 return (0); 949} 950 951/* 952 * Add a single contiguous physical range to the segment list. 953 */ 954static int 955_bus_dmamap_addseg(bus_dma_tag_t dmat, bus_dmamap_t map, bus_addr_t curaddr, 956 bus_size_t sgsize, bus_dma_segment_t *segs, int *segp) 957{ 958 bus_addr_t baddr, bmask; 959 int seg; 960 961 /* 962 * Make sure we don't cross any boundaries. 963 */ 964 bmask = ~(dmat->boundary - 1); 965 if (dmat->boundary > 0) { 966 baddr = (curaddr + dmat->boundary) & bmask; 967 if (sgsize > (baddr - curaddr)) 968 sgsize = (baddr - curaddr); 969 } 970 971 if (dmat->ranges) { 972 struct arm32_dma_range *dr; 973 974 dr = _bus_dma_inrange(dmat->ranges, dmat->_nranges, 975 curaddr); 976 if (dr == NULL) { 977 _bus_dmamap_unload(dmat, map); 978 return (0); 979 } 980 /* 981 * In a valid DMA range. Translate the physical 982 * memory address to an address in the DMA window. 983 */ 984 curaddr = (curaddr - dr->dr_sysbase) + dr->dr_busbase; 985 } 986 987 /* 988 * Insert chunk into a segment, coalescing with 989 * previous segment if possible. 990 */ 991 seg = *segp; 992 if (seg == -1) { 993 seg = 0; 994 segs[seg].ds_addr = curaddr; 995 segs[seg].ds_len = sgsize; 996 } else { 997 if (curaddr == segs[seg].ds_addr + segs[seg].ds_len && 998 (segs[seg].ds_len + sgsize) <= dmat->maxsegsz && 999 (dmat->boundary == 0 || 1000 (segs[seg].ds_addr & bmask) == (curaddr & bmask))) 1001 segs[seg].ds_len += sgsize; 1002 else { 1003 if (++seg >= dmat->nsegments) 1004 return (0); 1005 segs[seg].ds_addr = curaddr; 1006 segs[seg].ds_len = sgsize; 1007 } 1008 } 1009 *segp = seg; 1010 return (sgsize); 1011} 1012 1013/* 1014 * Utility function to load a physical buffer. segp contains 1015 * the starting segment on entrace, and the ending segment on exit. 1016 */ 1017int 1018_bus_dmamap_load_phys(bus_dma_tag_t dmat, 1019 bus_dmamap_t map, 1020 vm_paddr_t buf, bus_size_t buflen, 1021 int flags, 1022 bus_dma_segment_t *segs, 1023 int *segp) 1024{ 1025 bus_addr_t curaddr; 1026 bus_size_t sgsize; 1027 int error; 1028 1029 if (segs == NULL) 1030 segs = map->segments; 1031 1032 maploads_total++; 1033 maploads_physmem++; 1034 1035 if (might_bounce(dmat, map, buflen, buflen)) { 1036 _bus_dmamap_count_phys(dmat, map, buf, buflen, flags); 1037 if (map->pagesneeded != 0) { 1038 maploads_bounced++; 1039 error = _bus_dmamap_reserve_pages(dmat, map, flags); 1040 if (error) 1041 return (error); 1042 } 1043 } 1044 1045 while (buflen > 0) { 1046 curaddr = buf; 1047 sgsize = MIN(buflen, dmat->maxsegsz); 1048 if (map->pagesneeded != 0 && must_bounce(dmat, map, curaddr, 1049 sgsize)) { 1050 sgsize = MIN(sgsize, PAGE_SIZE); 1051 curaddr = add_bounce_page(dmat, map, 0, curaddr, 1052 sgsize); 1053 } 1054 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs, 1055 segp); 1056 if (sgsize == 0) 1057 break; 1058 buf += sgsize; 1059 buflen -= sgsize; 1060 } 1061 1062 /* 1063 * Did we fit? 1064 */ 1065 if (buflen != 0) { 1066 _bus_dmamap_unload(dmat, map); 1067 return (EFBIG); /* XXX better return value here? */ 1068 } 1069 return (0); 1070} 1071 1072int 1073_bus_dmamap_load_ma(bus_dma_tag_t dmat, bus_dmamap_t map, 1074 struct vm_page **ma, bus_size_t tlen, int ma_offs, int flags, 1075 bus_dma_segment_t *segs, int *segp) 1076{ 1077 1078 return (bus_dmamap_load_ma_triv(dmat, map, ma, tlen, ma_offs, flags, 1079 segs, segp)); 1080} 1081 1082/* 1083 * Utility function to load a linear buffer. segp contains 1084 * the starting segment on entrace, and the ending segment on exit. 1085 */ 1086int 1087_bus_dmamap_load_buffer(bus_dma_tag_t dmat, 1088 bus_dmamap_t map, 1089 void *buf, bus_size_t buflen, 1090 pmap_t pmap, 1091 int flags, 1092 bus_dma_segment_t *segs, 1093 int *segp) 1094{ 1095 bus_size_t sgsize; 1096 bus_addr_t curaddr; 1097 vm_offset_t vaddr; 1098 struct sync_list *sl; 1099 int error; 1100 1101 maploads_total++; 1102 if (map->flags & DMAMAP_COHERENT) 1103 maploads_coherent++; 1104 if (map->flags & DMAMAP_DMAMEM_ALLOC) 1105 maploads_dmamem++; 1106 1107 if (segs == NULL) 1108 segs = map->segments; 1109 1110 if (flags & BUS_DMA_LOAD_MBUF) { 1111 maploads_mbuf++; 1112 map->flags |= DMAMAP_MBUF; 1113 } 1114 1115 map->pmap = pmap; 1116 1117 if (might_bounce(dmat, map, (bus_addr_t)buf, buflen)) { 1118 _bus_dmamap_count_pages(dmat, map, buf, buflen, flags); 1119 if (map->pagesneeded != 0) { 1120 maploads_bounced++; 1121 error = _bus_dmamap_reserve_pages(dmat, map, flags); 1122 if (error) 1123 return (error); 1124 } 1125 } 1126 1127 sl = NULL; 1128 vaddr = (vm_offset_t)buf; 1129 1130 while (buflen > 0) { 1131 /* 1132 * Get the physical address for this segment. 1133 */ 1134 if (__predict_true(map->pmap == kernel_pmap)) 1135 curaddr = pmap_kextract(vaddr); 1136 else 1137 curaddr = pmap_extract(map->pmap, vaddr); 1138 1139 /* 1140 * Compute the segment size, and adjust counts. 1141 */ 1142 sgsize = PAGE_SIZE - ((u_long)curaddr & PAGE_MASK); 1143 if (sgsize > dmat->maxsegsz) 1144 sgsize = dmat->maxsegsz; 1145 if (buflen < sgsize) 1146 sgsize = buflen; 1147 1148 if (map->pagesneeded != 0 && must_bounce(dmat, map, curaddr, 1149 sgsize)) { 1150 curaddr = add_bounce_page(dmat, map, vaddr, curaddr, 1151 sgsize); 1152 } else { 1153 sl = &map->slist[map->sync_count - 1]; 1154 if (map->sync_count == 0 || 1155#ifdef ARM_L2_PIPT 1156 curaddr != sl->busaddr + sl->datacount || 1157#endif 1158 vaddr != sl->vaddr + sl->datacount) { 1159 if (++map->sync_count > dmat->nsegments) 1160 goto cleanup; 1161 sl++; 1162 sl->vaddr = vaddr; 1163 sl->datacount = sgsize; 1164 sl->busaddr = curaddr; 1165 } else 1166 sl->datacount += sgsize; 1167 } 1168 sgsize = _bus_dmamap_addseg(dmat, map, curaddr, sgsize, segs, 1169 segp); 1170 if (sgsize == 0) 1171 break; 1172 vaddr += sgsize; 1173 buflen -= sgsize; 1174 } 1175 1176cleanup: 1177 /* 1178 * Did we fit? 1179 */ 1180 if (buflen != 0) { 1181 _bus_dmamap_unload(dmat, map); 1182 return (EFBIG); /* XXX better return value here? */ 1183 } 1184 return (0); 1185} 1186 1187 1188void 1189__bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map, 1190 struct memdesc *mem, bus_dmamap_callback_t *callback, 1191 void *callback_arg) 1192{ 1193 1194 map->mem = *mem; 1195 map->dmat = dmat; 1196 map->callback = callback; 1197 map->callback_arg = callback_arg; 1198} 1199 1200bus_dma_segment_t * 1201_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map, 1202 bus_dma_segment_t *segs, int nsegs, int error) 1203{ 1204 1205 if (segs == NULL) 1206 segs = map->segments; 1207 return (segs); 1208} 1209 1210/* 1211 * Release the mapping held by map. 1212 */ 1213void 1214_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map) 1215{ 1216 struct bounce_page *bpage; 1217 struct bounce_zone *bz; 1218 1219 if ((bz = dmat->bounce_zone) != NULL) { 1220 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) { 1221 STAILQ_REMOVE_HEAD(&map->bpages, links); 1222 free_bounce_page(dmat, bpage); 1223 } 1224 1225 bz = dmat->bounce_zone; 1226 bz->free_bpages += map->pagesreserved; 1227 bz->reserved_bpages -= map->pagesreserved; 1228 map->pagesreserved = 0; 1229 map->pagesneeded = 0; 1230 } 1231 map->sync_count = 0; 1232 map->flags &= ~DMAMAP_MBUF; 1233} 1234 1235#ifdef notyetbounceuser 1236/* If busdma uses user pages, then the interrupt handler could 1237 * be use the kernel vm mapping. Both bounce pages and sync list 1238 * do not cross page boundaries. 1239 * Below is a rough sequence that a person would do to fix the 1240 * user page reference in the kernel vmspace. This would be 1241 * done in the dma post routine. 1242 */ 1243void 1244_bus_dmamap_fix_user(vm_offset_t buf, bus_size_t len, 1245 pmap_t pmap, int op) 1246{ 1247 bus_size_t sgsize; 1248 bus_addr_t curaddr; 1249 vm_offset_t va; 1250 1251 /* 1252 * each synclist entry is contained within a single page. 1253 * this would be needed if BUS_DMASYNC_POSTxxxx was implemented 1254 */ 1255 curaddr = pmap_extract(pmap, buf); 1256 va = pmap_dma_map(curaddr); 1257 switch (op) { 1258 case SYNC_USER_INV: 1259 cpu_dcache_wb_range(va, sgsize); 1260 break; 1261 1262 case SYNC_USER_COPYTO: 1263 bcopy((void *)va, (void *)bounce, sgsize); 1264 break; 1265 1266 case SYNC_USER_COPYFROM: 1267 bcopy((void *) bounce, (void *)va, sgsize); 1268 break; 1269 1270 default: 1271 break; 1272 } 1273 1274 pmap_dma_unmap(va); 1275} 1276#endif 1277 1278#ifdef ARM_L2_PIPT 1279#define l2cache_wb_range(va, pa, size) cpu_l2cache_wb_range(pa, size) 1280#define l2cache_wbinv_range(va, pa, size) cpu_l2cache_wbinv_range(pa, size) 1281#define l2cache_inv_range(va, pa, size) cpu_l2cache_inv_range(pa, size) 1282#else 1283#define l2cache_wb_range(va, pa, size) cpu_l2cache_wb_range(va, size) 1284#define l2cache_wbinv_range(va, pa, size) cpu_l2cache_wbinv_range(va, size) 1285#define l2cache_inv_range(va, pa, size) cpu_l2cache_inv_range(va, size) 1286#endif 1287 1288void 1289_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op) 1290{ 1291 struct bounce_page *bpage; 1292 struct sync_list *sl, *end; 1293 /* 1294 * If the buffer was from user space, it is possible that this is not 1295 * the same vm map, especially on a POST operation. It's not clear that 1296 * dma on userland buffers can work at all right now. To be safe, until 1297 * we're able to test direct userland dma, panic on a map mismatch. 1298 */ 1299 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) { 1300 if (!pmap_dmap_iscurrent(map->pmap)) 1301 panic("_bus_dmamap_sync: wrong user map for bounce sync."); 1302 1303 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x " 1304 "performing bounce", __func__, dmat, dmat->flags, op); 1305 1306 /* 1307 * For PREWRITE do a writeback. Clean the caches from the 1308 * innermost to the outermost levels. 1309 */ 1310 if (op & BUS_DMASYNC_PREWRITE) { 1311 while (bpage != NULL) { 1312 if (bpage->datavaddr != 0) 1313 bcopy((void *)bpage->datavaddr, 1314 (void *)bpage->vaddr, 1315 bpage->datacount); 1316 else 1317 physcopyout(bpage->dataaddr, 1318 (void *)bpage->vaddr, 1319 bpage->datacount); 1320 cpu_dcache_wb_range((vm_offset_t)bpage->vaddr, 1321 bpage->datacount); 1322 l2cache_wb_range((vm_offset_t)bpage->vaddr, 1323 (vm_offset_t)bpage->busaddr, 1324 bpage->datacount); 1325 bpage = STAILQ_NEXT(bpage, links); 1326 } 1327 dmat->bounce_zone->total_bounced++; 1328 } 1329 1330 /* 1331 * Do an invalidate for PREREAD unless a writeback was already 1332 * done above due to PREWRITE also being set. The reason for a 1333 * PREREAD invalidate is to prevent dirty lines currently in the 1334 * cache from being evicted during the DMA. If a writeback was 1335 * done due to PREWRITE also being set there will be no dirty 1336 * lines and the POSTREAD invalidate handles the rest. The 1337 * invalidate is done from the innermost to outermost level. If 1338 * L2 were done first, a dirty cacheline could be automatically 1339 * evicted from L1 before we invalidated it, re-dirtying the L2. 1340 */ 1341 if ((op & BUS_DMASYNC_PREREAD) && !(op & BUS_DMASYNC_PREWRITE)) { 1342 bpage = STAILQ_FIRST(&map->bpages); 1343 while (bpage != NULL) { 1344 cpu_dcache_inv_range((vm_offset_t)bpage->vaddr, 1345 bpage->datacount); 1346 l2cache_inv_range((vm_offset_t)bpage->vaddr, 1347 (vm_offset_t)bpage->busaddr, 1348 bpage->datacount); 1349 bpage = STAILQ_NEXT(bpage, links); 1350 } 1351 } 1352 1353 /* 1354 * Re-invalidate the caches on a POSTREAD, even though they were 1355 * already invalidated at PREREAD time. Aggressive prefetching 1356 * due to accesses to other data near the dma buffer could have 1357 * brought buffer data into the caches which is now stale. The 1358 * caches are invalidated from the outermost to innermost; the 1359 * prefetches could be happening right now, and if L1 were 1360 * invalidated first, stale L2 data could be prefetched into L1. 1361 */ 1362 if (op & BUS_DMASYNC_POSTREAD) { 1363 while (bpage != NULL) { 1364 vm_offset_t startv; 1365 vm_paddr_t startp; 1366 int len; 1367 1368 startv = bpage->vaddr &~ arm_dcache_align_mask; 1369 startp = bpage->busaddr &~ arm_dcache_align_mask; 1370 len = bpage->datacount; 1371 1372 if (startv != bpage->vaddr) 1373 len += bpage->vaddr & arm_dcache_align_mask; 1374 if (len & arm_dcache_align_mask) 1375 len = (len - 1376 (len & arm_dcache_align_mask)) + 1377 arm_dcache_align; 1378 l2cache_inv_range(startv, startp, len); 1379 cpu_dcache_inv_range(startv, len); 1380 if (bpage->datavaddr != 0) 1381 bcopy((void *)bpage->vaddr, 1382 (void *)bpage->datavaddr, 1383 bpage->datacount); 1384 else 1385 physcopyin((void *)bpage->vaddr, 1386 bpage->dataaddr, 1387 bpage->datacount); 1388 bpage = STAILQ_NEXT(bpage, links); 1389 } 1390 dmat->bounce_zone->total_bounced++; 1391 } 1392 } 1393 1394 /* 1395 * For COHERENT memory no cache maintenance is necessary, but ensure all 1396 * writes have reached memory for the PREWRITE case. No action is 1397 * needed for a PREREAD without PREWRITE also set, because that would 1398 * imply that the cpu had written to the COHERENT buffer and expected 1399 * the dma device to see that change, and by definition a PREWRITE sync 1400 * is required to make that happen. 1401 */ 1402 if (map->flags & DMAMAP_COHERENT) { 1403 if (op & BUS_DMASYNC_PREWRITE) { 1404 dsb(); 1405 cpu_l2cache_drain_writebuf(); 1406 } 1407 return; 1408 } 1409 1410 /* 1411 * Cache maintenance for normal (non-COHERENT non-bounce) buffers. All 1412 * the comments about the sequences for flushing cache levels in the 1413 * bounce buffer code above apply here as well. In particular, the fact 1414 * that the sequence is inner-to-outer for PREREAD invalidation and 1415 * outer-to-inner for POSTREAD invalidation is not a mistake. 1416 */ 1417 if (map->sync_count != 0) { 1418 if (!pmap_dmap_iscurrent(map->pmap)) 1419 panic("_bus_dmamap_sync: wrong user map for sync."); 1420 1421 sl = &map->slist[0]; 1422 end = &map->slist[map->sync_count]; 1423 CTR4(KTR_BUSDMA, "%s: tag %p tag flags 0x%x op 0x%x " 1424 "performing sync", __func__, dmat, dmat->flags, op); 1425 1426 switch (op) { 1427 case BUS_DMASYNC_PREWRITE: 1428 case BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD: 1429 while (sl != end) { 1430 cpu_dcache_wb_range(sl->vaddr, sl->datacount); 1431 l2cache_wb_range(sl->vaddr, sl->busaddr, 1432 sl->datacount); 1433 sl++; 1434 } 1435 break; 1436 1437 case BUS_DMASYNC_PREREAD: 1438 /* 1439 * An mbuf may start in the middle of a cacheline. There 1440 * will be no cpu writes to the beginning of that line 1441 * (which contains the mbuf header) while dma is in 1442 * progress. Handle that case by doing a writeback of 1443 * just the first cacheline before invalidating the 1444 * overall buffer. Any mbuf in a chain may have this 1445 * misalignment. Buffers which are not mbufs bounce if 1446 * they are not aligned to a cacheline. 1447 */ 1448 while (sl != end) { 1449 if (sl->vaddr & arm_dcache_align_mask) { 1450 KASSERT(map->flags & DMAMAP_MBUF, 1451 ("unaligned buffer is not an mbuf")); 1452 cpu_dcache_wb_range(sl->vaddr, 1); 1453 l2cache_wb_range(sl->vaddr, 1454 sl->busaddr, 1); 1455 } 1456 cpu_dcache_inv_range(sl->vaddr, sl->datacount); 1457 l2cache_inv_range(sl->vaddr, sl->busaddr, 1458 sl->datacount); 1459 sl++; 1460 } 1461 break; 1462 1463 case BUS_DMASYNC_POSTWRITE: 1464 break; 1465 1466 case BUS_DMASYNC_POSTREAD: 1467 case BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE: 1468 while (sl != end) { 1469 l2cache_inv_range(sl->vaddr, sl->busaddr, 1470 sl->datacount); 1471 cpu_dcache_inv_range(sl->vaddr, sl->datacount); 1472 sl++; 1473 } 1474 break; 1475 1476 default: 1477 panic("unsupported combination of sync operations: 0x%08x\n", op); 1478 break; 1479 } 1480 } 1481} 1482 1483static void 1484init_bounce_pages(void *dummy __unused) 1485{ 1486 1487 total_bpages = 0; 1488 STAILQ_INIT(&bounce_zone_list); 1489 STAILQ_INIT(&bounce_map_waitinglist); 1490 STAILQ_INIT(&bounce_map_callbacklist); 1491 mtx_init(&bounce_lock, "bounce pages lock", NULL, MTX_DEF); 1492} 1493SYSINIT(bpages, SI_SUB_LOCK, SI_ORDER_ANY, init_bounce_pages, NULL); 1494 1495static struct sysctl_ctx_list * 1496busdma_sysctl_tree(struct bounce_zone *bz) 1497{ 1498 1499 return (&bz->sysctl_tree); 1500} 1501 1502static struct sysctl_oid * 1503busdma_sysctl_tree_top(struct bounce_zone *bz) 1504{ 1505 1506 return (bz->sysctl_tree_top); 1507} 1508 1509static int 1510alloc_bounce_zone(bus_dma_tag_t dmat) 1511{ 1512 struct bounce_zone *bz; 1513 1514 /* Check to see if we already have a suitable zone */ 1515 STAILQ_FOREACH(bz, &bounce_zone_list, links) { 1516 if ((dmat->alignment <= bz->alignment) && 1517 (dmat->lowaddr >= bz->lowaddr)) { 1518 dmat->bounce_zone = bz; 1519 return (0); 1520 } 1521 } 1522 1523 if ((bz = (struct bounce_zone *)malloc(sizeof(*bz), M_DEVBUF, 1524 M_NOWAIT | M_ZERO)) == NULL) 1525 return (ENOMEM); 1526 1527 STAILQ_INIT(&bz->bounce_page_list); 1528 bz->free_bpages = 0; 1529 bz->reserved_bpages = 0; 1530 bz->active_bpages = 0; 1531 bz->lowaddr = dmat->lowaddr; 1532 bz->alignment = MAX(dmat->alignment, PAGE_SIZE); 1533 bz->map_count = 0; 1534 snprintf(bz->zoneid, 8, "zone%d", busdma_zonecount); 1535 busdma_zonecount++; 1536 snprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr); 1537 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links); 1538 dmat->bounce_zone = bz; 1539 1540 sysctl_ctx_init(&bz->sysctl_tree); 1541 bz->sysctl_tree_top = SYSCTL_ADD_NODE(&bz->sysctl_tree, 1542 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid, 1543 CTLFLAG_RD, 0, ""); 1544 if (bz->sysctl_tree_top == NULL) { 1545 sysctl_ctx_free(&bz->sysctl_tree); 1546 return (0); /* XXX error code? */ 1547 } 1548 1549 SYSCTL_ADD_INT(busdma_sysctl_tree(bz), 1550 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1551 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0, 1552 "Total bounce pages"); 1553 SYSCTL_ADD_INT(busdma_sysctl_tree(bz), 1554 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1555 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0, 1556 "Free bounce pages"); 1557 SYSCTL_ADD_INT(busdma_sysctl_tree(bz), 1558 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1559 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0, 1560 "Reserved bounce pages"); 1561 SYSCTL_ADD_INT(busdma_sysctl_tree(bz), 1562 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1563 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0, 1564 "Active bounce pages"); 1565 SYSCTL_ADD_INT(busdma_sysctl_tree(bz), 1566 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1567 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0, 1568 "Total bounce requests (pages bounced)"); 1569 SYSCTL_ADD_INT(busdma_sysctl_tree(bz), 1570 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1571 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0, 1572 "Total bounce requests that were deferred"); 1573 SYSCTL_ADD_STRING(busdma_sysctl_tree(bz), 1574 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1575 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, ""); 1576 SYSCTL_ADD_ULONG(busdma_sysctl_tree(bz), 1577 SYSCTL_CHILDREN(busdma_sysctl_tree_top(bz)), OID_AUTO, 1578 "alignment", CTLFLAG_RD, &bz->alignment, ""); 1579 1580 return (0); 1581} 1582 1583static int 1584alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages) 1585{ 1586 struct bounce_zone *bz; 1587 int count; 1588 1589 bz = dmat->bounce_zone; 1590 count = 0; 1591 while (numpages > 0) { 1592 struct bounce_page *bpage; 1593 1594 bpage = (struct bounce_page *)malloc(sizeof(*bpage), M_DEVBUF, 1595 M_NOWAIT | M_ZERO); 1596 1597 if (bpage == NULL) 1598 break; 1599 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF, 1600 M_NOWAIT, 0ul, bz->lowaddr, PAGE_SIZE, 0); 1601 if (bpage->vaddr == 0) { 1602 free(bpage, M_DEVBUF); 1603 break; 1604 } 1605 bpage->busaddr = pmap_kextract(bpage->vaddr); 1606 mtx_lock(&bounce_lock); 1607 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links); 1608 total_bpages++; 1609 bz->total_bpages++; 1610 bz->free_bpages++; 1611 mtx_unlock(&bounce_lock); 1612 count++; 1613 numpages--; 1614 } 1615 return (count); 1616} 1617 1618static int 1619reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit) 1620{ 1621 struct bounce_zone *bz; 1622 int pages; 1623 1624 mtx_assert(&bounce_lock, MA_OWNED); 1625 bz = dmat->bounce_zone; 1626 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved); 1627 if (commit == 0 && map->pagesneeded > (map->pagesreserved + pages)) 1628 return (map->pagesneeded - (map->pagesreserved + pages)); 1629 bz->free_bpages -= pages; 1630 bz->reserved_bpages += pages; 1631 map->pagesreserved += pages; 1632 pages = map->pagesneeded - map->pagesreserved; 1633 1634 return (pages); 1635} 1636 1637static bus_addr_t 1638add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr, 1639 bus_addr_t addr, bus_size_t size) 1640{ 1641 struct bounce_zone *bz; 1642 struct bounce_page *bpage; 1643 1644 KASSERT(dmat->bounce_zone != NULL, ("no bounce zone in dma tag")); 1645 KASSERT(map != NULL, 1646 ("add_bounce_page: bad map %p", map)); 1647 1648 bz = dmat->bounce_zone; 1649 if (map->pagesneeded == 0) 1650 panic("add_bounce_page: map doesn't need any pages"); 1651 map->pagesneeded--; 1652 1653 if (map->pagesreserved == 0) 1654 panic("add_bounce_page: map doesn't need any pages"); 1655 map->pagesreserved--; 1656 1657 mtx_lock(&bounce_lock); 1658 bpage = STAILQ_FIRST(&bz->bounce_page_list); 1659 if (bpage == NULL) 1660 panic("add_bounce_page: free page list is empty"); 1661 1662 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links); 1663 bz->reserved_bpages--; 1664 bz->active_bpages++; 1665 mtx_unlock(&bounce_lock); 1666 1667 if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) { 1668 /* Page offset needs to be preserved. */ 1669 bpage->vaddr |= addr & PAGE_MASK; 1670 bpage->busaddr |= addr & PAGE_MASK; 1671 } 1672 bpage->datavaddr = vaddr; 1673 bpage->dataaddr = addr; 1674 bpage->datacount = size; 1675 STAILQ_INSERT_TAIL(&(map->bpages), bpage, links); 1676 return (bpage->busaddr); 1677} 1678 1679static void 1680free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage) 1681{ 1682 struct bus_dmamap *map; 1683 struct bounce_zone *bz; 1684 1685 bz = dmat->bounce_zone; 1686 bpage->datavaddr = 0; 1687 bpage->datacount = 0; 1688 if (dmat->flags & BUS_DMA_KEEP_PG_OFFSET) { 1689 /* 1690 * Reset the bounce page to start at offset 0. Other uses 1691 * of this bounce page may need to store a full page of 1692 * data and/or assume it starts on a page boundary. 1693 */ 1694 bpage->vaddr &= ~PAGE_MASK; 1695 bpage->busaddr &= ~PAGE_MASK; 1696 } 1697 1698 mtx_lock(&bounce_lock); 1699 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links); 1700 bz->free_bpages++; 1701 bz->active_bpages--; 1702 if ((map = STAILQ_FIRST(&bounce_map_waitinglist)) != NULL) { 1703 if (reserve_bounce_pages(map->dmat, map, 1) == 0) { 1704 STAILQ_REMOVE_HEAD(&bounce_map_waitinglist, links); 1705 STAILQ_INSERT_TAIL(&bounce_map_callbacklist, 1706 map, links); 1707 busdma_swi_pending = 1; 1708 bz->total_deferred++; 1709 swi_sched(vm_ih, 0); 1710 } 1711 } 1712 mtx_unlock(&bounce_lock); 1713} 1714 1715void 1716busdma_swi(void) 1717{ 1718 bus_dma_tag_t dmat; 1719 struct bus_dmamap *map; 1720 1721 mtx_lock(&bounce_lock); 1722 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) { 1723 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links); 1724 mtx_unlock(&bounce_lock); 1725 dmat = map->dmat; 1726 dmat->lockfunc(dmat->lockfuncarg, BUS_DMA_LOCK); 1727 bus_dmamap_load_mem(map->dmat, map, &map->mem, map->callback, 1728 map->callback_arg, BUS_DMA_WAITOK); 1729 dmat->lockfunc(dmat->lockfuncarg, BUS_DMA_UNLOCK); 1730 mtx_lock(&bounce_lock); 1731 } 1732 mtx_unlock(&bounce_lock); 1733} 1734