1 /* 2 * Copyright (c) 1997, 1998 Justin T. Gibbs. 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions, and the following disclaimer, 10 * without modification, immediately at the beginning of the file. 11 * 2. The name of the author may not be used to endorse or promote products 12 * derived from this software without specific prior written permission. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR 18 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 * $FreeBSD: src/sys/i386/i386/busdma_machdep.c,v 1.94 2008/08/15 20:51:31 kmacy Exp $ 27 */ 28 29 #include <sys/param.h> 30 #include <sys/systm.h> 31 #include <sys/malloc.h> 32 #include <sys/mbuf.h> 33 #include <sys/uio.h> 34 #include <sys/thread2.h> 35 #include <sys/bus_dma.h> 36 #include <sys/kernel.h> 37 #include <sys/sysctl.h> 38 #include <sys/lock.h> 39 #include <sys/spinlock2.h> 40 41 #include <vm/vm.h> 42 #include <vm/vm_page.h> 43 44 /* XXX needed for to access pmap to convert per-proc virtual to physical */ 45 #include <sys/proc.h> 46 #include <vm/vm_map.h> 47 48 #include <machine/md_var.h> 49 50 #define MAX_BPAGES 1024 51 52 struct bounce_zone; 53 struct bus_dmamap; 54 55 struct bus_dma_tag { 56 bus_dma_tag_t parent; 57 bus_size_t alignment; 58 bus_size_t boundary; 59 bus_addr_t lowaddr; 60 bus_addr_t highaddr; 61 bus_dma_filter_t *filter; 62 void *filterarg; 63 bus_size_t maxsize; 64 u_int nsegments; 65 bus_size_t maxsegsz; 66 int flags; 67 int ref_count; 68 int map_count; 69 bus_dma_segment_t *segments; 70 struct bounce_zone *bounce_zone; 71 }; 72 73 /* 74 * bus_dma_tag private flags 75 */ 76 #define BUS_DMA_BOUNCE_ALIGN BUS_DMA_BUS2 77 #define BUS_DMA_BOUNCE_LOWADDR BUS_DMA_BUS3 78 #define BUS_DMA_MIN_ALLOC_COMP BUS_DMA_BUS4 79 80 #define BUS_DMA_COULD_BOUNCE (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN) 81 82 #define BUS_DMAMEM_KMALLOC(dmat) \ 83 ((dmat)->maxsize <= PAGE_SIZE && \ 84 (dmat)->alignment <= PAGE_SIZE && \ 85 (dmat)->lowaddr >= ptoa(Maxmem)) 86 87 struct bounce_page { 88 vm_offset_t vaddr; /* kva of bounce buffer */ 89 bus_addr_t busaddr; /* Physical address */ 90 vm_offset_t datavaddr; /* kva of client data */ 91 bus_size_t datacount; /* client data count */ 92 STAILQ_ENTRY(bounce_page) links; 93 }; 94 95 struct bounce_zone { 96 STAILQ_ENTRY(bounce_zone) links; 97 STAILQ_HEAD(bp_list, bounce_page) bounce_page_list; 98 STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist; 99 struct spinlock spin; 100 int total_bpages; 101 int free_bpages; 102 int reserved_bpages; 103 int active_bpages; 104 int total_bounced; 105 int total_deferred; 106 int reserve_failed; 107 bus_size_t alignment; 108 bus_addr_t lowaddr; 109 char zoneid[8]; 110 char lowaddrid[20]; 111 struct sysctl_ctx_list sysctl_ctx; 112 struct sysctl_oid *sysctl_tree; 113 }; 114 115 #define BZ_LOCK(bz) spin_lock(&(bz)->spin) 116 #define BZ_UNLOCK(bz) spin_unlock(&(bz)->spin) 117 118 static struct lwkt_token bounce_zone_tok = 119 LWKT_TOKEN_INITIALIZER(bounce_zone_token); 120 static int busdma_zonecount; 121 static STAILQ_HEAD(, bounce_zone) bounce_zone_list = 122 STAILQ_HEAD_INITIALIZER(bounce_zone_list); 123 124 int busdma_swi_pending; 125 static int total_bounce_pages; 126 static int max_bounce_pages = MAX_BPAGES; 127 static int bounce_alignment = 1; /* XXX temporary */ 128 129 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages); 130 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment); 131 132 struct bus_dmamap { 133 struct bp_list bpages; 134 int pagesneeded; 135 int pagesreserved; 136 bus_dma_tag_t dmat; 137 void *buf; /* unmapped buffer pointer */ 138 bus_size_t buflen; /* unmapped buffer length */ 139 bus_dmamap_callback_t *callback; 140 void *callback_arg; 141 STAILQ_ENTRY(bus_dmamap) links; 142 }; 143 144 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist = 145 STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist); 146 147 static struct bus_dmamap nobounce_dmamap; 148 149 static int alloc_bounce_zone(bus_dma_tag_t); 150 static int alloc_bounce_pages(bus_dma_tag_t, u_int, int); 151 static int reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int); 152 static void return_bounce_pages(bus_dma_tag_t, bus_dmamap_t); 153 static bus_addr_t add_bounce_page(bus_dma_tag_t, bus_dmamap_t, 154 vm_offset_t, bus_size_t); 155 static void free_bounce_page(bus_dma_tag_t, struct bounce_page *); 156 157 static bus_dmamap_t get_map_waiting(bus_dma_tag_t); 158 static void add_map_callback(bus_dmamap_t); 159 160 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters"); 161 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages, 162 0, "Total bounce pages"); 163 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages, 164 0, "Max bounce pages per bounce zone"); 165 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD, 166 &bounce_alignment, 0, "Obey alignment constraint"); 167 168 static __inline int 169 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr) 170 { 171 int retval; 172 173 retval = 0; 174 do { 175 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) || 176 (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0)) 177 && (dmat->filter == NULL || 178 dmat->filter(dmat->filterarg, paddr) != 0)) 179 retval = 1; 180 181 dmat = dmat->parent; 182 } while (retval == 0 && dmat != NULL); 183 return (retval); 184 } 185 186 /* 187 * Allocate a device specific dma_tag. 188 */ 189 int 190 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment, 191 bus_size_t boundary, bus_addr_t lowaddr, 192 bus_addr_t highaddr, bus_dma_filter_t *filter, 193 void *filterarg, bus_size_t maxsize, int nsegments, 194 bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat) 195 { 196 bus_dma_tag_t newtag; 197 int error = 0; 198 199 /* 200 * Sanity checks 201 */ 202 203 if (alignment == 0) 204 alignment = 1; 205 if (alignment & (alignment - 1)) 206 panic("alignment must be power of 2"); 207 208 if (boundary != 0) { 209 if (boundary & (boundary - 1)) 210 panic("boundary must be power of 2"); 211 if (boundary < maxsegsz) { 212 kprintf("boundary < maxsegsz:\n"); 213 print_backtrace(-1); 214 maxsegsz = boundary; 215 } 216 } 217 218 /* Return a NULL tag on failure */ 219 *dmat = NULL; 220 221 newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT); 222 223 newtag->parent = parent; 224 newtag->alignment = alignment; 225 newtag->boundary = boundary; 226 newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1); 227 newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1); 228 newtag->filter = filter; 229 newtag->filterarg = filterarg; 230 newtag->maxsize = maxsize; 231 newtag->nsegments = nsegments; 232 newtag->maxsegsz = maxsegsz; 233 newtag->flags = flags; 234 newtag->ref_count = 1; /* Count ourself */ 235 newtag->map_count = 0; 236 newtag->segments = NULL; 237 newtag->bounce_zone = NULL; 238 239 /* Take into account any restrictions imposed by our parent tag */ 240 if (parent != NULL) { 241 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr); 242 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr); 243 244 if (newtag->boundary == 0) { 245 newtag->boundary = parent->boundary; 246 } else if (parent->boundary != 0) { 247 newtag->boundary = MIN(parent->boundary, 248 newtag->boundary); 249 } 250 251 #ifdef notyet 252 newtag->alignment = MAX(parent->alignment, newtag->alignment); 253 #endif 254 255 if (newtag->filter == NULL) { 256 /* 257 * Short circuit looking at our parent directly 258 * since we have encapsulated all of its information 259 */ 260 newtag->filter = parent->filter; 261 newtag->filterarg = parent->filterarg; 262 newtag->parent = parent->parent; 263 } 264 if (newtag->parent != NULL) 265 parent->ref_count++; 266 } 267 268 if (newtag->lowaddr < ptoa(Maxmem)) 269 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR; 270 if (bounce_alignment && newtag->alignment > 1 && 271 !(newtag->flags & BUS_DMA_ALIGNED)) 272 newtag->flags |= BUS_DMA_BOUNCE_ALIGN; 273 274 if ((newtag->flags & BUS_DMA_COULD_BOUNCE) && 275 (flags & BUS_DMA_ALLOCNOW) != 0) { 276 struct bounce_zone *bz; 277 278 /* Must bounce */ 279 280 error = alloc_bounce_zone(newtag); 281 if (error) 282 goto back; 283 bz = newtag->bounce_zone; 284 285 if (ptoa(bz->total_bpages) < maxsize) { 286 int pages; 287 288 if (flags & BUS_DMA_ONEBPAGE) { 289 pages = 1; 290 } else { 291 pages = atop(round_page(maxsize)) - 292 bz->total_bpages; 293 pages = MAX(pages, 1); 294 } 295 296 /* Add pages to our bounce pool */ 297 if (alloc_bounce_pages(newtag, pages, flags) < pages) 298 error = ENOMEM; 299 300 /* Performed initial allocation */ 301 newtag->flags |= BUS_DMA_MIN_ALLOC_COMP; 302 } 303 } 304 back: 305 if (error) 306 kfree(newtag, M_DEVBUF); 307 else 308 *dmat = newtag; 309 return error; 310 } 311 312 int 313 bus_dma_tag_destroy(bus_dma_tag_t dmat) 314 { 315 if (dmat != NULL) { 316 if (dmat->map_count != 0) 317 return (EBUSY); 318 319 while (dmat != NULL) { 320 bus_dma_tag_t parent; 321 322 parent = dmat->parent; 323 dmat->ref_count--; 324 if (dmat->ref_count == 0) { 325 if (dmat->segments != NULL) 326 kfree(dmat->segments, M_DEVBUF); 327 kfree(dmat, M_DEVBUF); 328 /* 329 * Last reference count, so 330 * release our reference 331 * count on our parent. 332 */ 333 dmat = parent; 334 } else 335 dmat = NULL; 336 } 337 } 338 return (0); 339 } 340 341 /* 342 * Allocate a handle for mapping from kva/uva/physical 343 * address space into bus device space. 344 */ 345 int 346 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp) 347 { 348 int error; 349 350 error = 0; 351 352 if (dmat->segments == NULL) { 353 KKASSERT(dmat->nsegments && dmat->nsegments < 16384); 354 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) * 355 dmat->nsegments, M_DEVBUF, M_INTWAIT); 356 } 357 358 if (dmat->flags & BUS_DMA_COULD_BOUNCE) { 359 struct bounce_zone *bz; 360 int maxpages; 361 362 /* Must bounce */ 363 364 if (dmat->bounce_zone == NULL) { 365 error = alloc_bounce_zone(dmat); 366 if (error) 367 return error; 368 } 369 bz = dmat->bounce_zone; 370 371 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO); 372 373 /* Initialize the new map */ 374 STAILQ_INIT(&((*mapp)->bpages)); 375 376 /* 377 * Attempt to add pages to our pool on a per-instance 378 * basis up to a sane limit. 379 */ 380 if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) { 381 maxpages = max_bounce_pages; 382 } else { 383 maxpages = MIN(max_bounce_pages, 384 Maxmem - atop(dmat->lowaddr)); 385 } 386 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 387 || (dmat->map_count > 0 388 && bz->total_bpages < maxpages)) { 389 int pages; 390 391 if (flags & BUS_DMA_ONEBPAGE) { 392 pages = 1; 393 } else { 394 pages = atop(round_page(dmat->maxsize)); 395 pages = MIN(maxpages - bz->total_bpages, pages); 396 pages = MAX(pages, 1); 397 } 398 if (alloc_bounce_pages(dmat, pages, flags) < pages) 399 error = ENOMEM; 400 401 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) { 402 if (!error) 403 dmat->flags |= BUS_DMA_MIN_ALLOC_COMP; 404 } else { 405 error = 0; 406 } 407 } 408 } else { 409 *mapp = NULL; 410 } 411 if (!error) 412 dmat->map_count++; 413 return error; 414 } 415 416 /* 417 * Destroy a handle for mapping from kva/uva/physical 418 * address space into bus device space. 419 */ 420 int 421 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map) 422 { 423 if (map != NULL) { 424 if (STAILQ_FIRST(&map->bpages) != NULL) 425 return (EBUSY); 426 kfree(map, M_DEVBUF); 427 } 428 dmat->map_count--; 429 return (0); 430 } 431 432 static __inline bus_size_t 433 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify) 434 { 435 bus_size_t maxsize = 0; 436 uintptr_t vaddr = (uintptr_t)vaddr0; 437 438 if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) { 439 kprintf("boundary check failed\n"); 440 if (verify) 441 print_backtrace(-1); /* XXX panic */ 442 maxsize = dmat->maxsize; 443 } 444 if (vaddr & (dmat->alignment - 1)) { 445 kprintf("alignment check failed\n"); 446 if (verify) 447 print_backtrace(-1); /* XXX panic */ 448 if (dmat->maxsize < dmat->alignment) 449 maxsize = dmat->alignment; 450 else 451 maxsize = dmat->maxsize; 452 } 453 return maxsize; 454 } 455 456 /* 457 * Allocate a piece of memory that can be efficiently mapped into 458 * bus device space based on the constraints lited in the dma tag. 459 * 460 * mapp is degenerate. By definition this allocation should not require 461 * bounce buffers so do not allocate a dma map. 462 */ 463 int 464 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags, 465 bus_dmamap_t *mapp) 466 { 467 int mflags; 468 469 /* If we succeed, no mapping/bouncing will be required */ 470 *mapp = NULL; 471 472 if (dmat->segments == NULL) { 473 KKASSERT(dmat->nsegments < 16384); 474 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) * 475 dmat->nsegments, M_DEVBUF, M_INTWAIT); 476 } 477 478 if (flags & BUS_DMA_NOWAIT) 479 mflags = M_NOWAIT; 480 else 481 mflags = M_WAITOK; 482 if (flags & BUS_DMA_ZERO) 483 mflags |= M_ZERO; 484 485 if (BUS_DMAMEM_KMALLOC(dmat)) { 486 bus_size_t maxsize; 487 488 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags); 489 490 /* 491 * XXX 492 * Check whether the allocation 493 * - crossed a page boundary 494 * - was not aligned 495 * Retry with power-of-2 alignment in the above cases. 496 */ 497 maxsize = check_kmalloc(dmat, *vaddr, 0); 498 if (maxsize) { 499 size_t size; 500 501 kfree(*vaddr, M_DEVBUF); 502 /* XXX check for overflow? */ 503 for (size = 1; size <= maxsize; size <<= 1) 504 ; 505 *vaddr = kmalloc(size, M_DEVBUF, mflags); 506 check_kmalloc(dmat, *vaddr, 1); 507 } 508 } else { 509 /* 510 * XXX Use Contigmalloc until it is merged into this facility 511 * and handles multi-seg allocations. Nobody is doing 512 * multi-seg allocations yet though. 513 */ 514 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags, 515 0ul, dmat->lowaddr, dmat->alignment, dmat->boundary); 516 } 517 if (*vaddr == NULL) 518 return (ENOMEM); 519 return (0); 520 } 521 522 /* 523 * Free a piece of memory and it's allociated dmamap, that was allocated 524 * via bus_dmamem_alloc. Make the same choice for free/contigfree. 525 */ 526 void 527 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map) 528 { 529 /* 530 * dmamem does not need to be bounced, so the map should be 531 * NULL 532 */ 533 if (map != NULL) 534 panic("bus_dmamem_free: Invalid map freed"); 535 if (BUS_DMAMEM_KMALLOC(dmat)) 536 kfree(vaddr, M_DEVBUF); 537 else 538 contigfree(vaddr, dmat->maxsize, M_DEVBUF); 539 } 540 541 static __inline vm_paddr_t 542 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr) 543 { 544 if (pmap) 545 return pmap_extract(pmap, vaddr, NULL); 546 else 547 return pmap_kextract(vaddr); 548 } 549 550 /* 551 * Utility function to load a linear buffer. lastaddrp holds state 552 * between invocations (for multiple-buffer loads). segp contains 553 * the segment following the starting one on entrace, and the ending 554 * segment on exit. first indicates if this is the first invocation 555 * of this function. 556 */ 557 static int 558 _bus_dmamap_load_buffer(bus_dma_tag_t dmat, 559 bus_dmamap_t map, 560 void *buf, bus_size_t buflen, 561 bus_dma_segment_t *segments, 562 int nsegments, 563 pmap_t pmap, 564 int flags, 565 vm_paddr_t *lastpaddrp, 566 int *segp, 567 int first) 568 { 569 vm_offset_t vaddr; 570 vm_paddr_t paddr, nextpaddr; 571 bus_dma_segment_t *sg; 572 bus_addr_t bmask; 573 int seg, error = 0; 574 575 if (map == NULL) 576 map = &nobounce_dmamap; 577 578 #ifdef INVARIANTS 579 if (dmat->flags & BUS_DMA_ALIGNED) 580 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0); 581 #endif 582 583 /* 584 * If we are being called during a callback, pagesneeded will 585 * be non-zero, so we can avoid doing the work twice. 586 */ 587 if ((dmat->flags & BUS_DMA_COULD_BOUNCE) && 588 map != &nobounce_dmamap && map->pagesneeded == 0) { 589 vm_offset_t vendaddr; 590 591 /* 592 * Count the number of bounce pages 593 * needed in order to complete this transfer 594 */ 595 vaddr = (vm_offset_t)buf; 596 vendaddr = (vm_offset_t)buf + buflen; 597 598 while (vaddr < vendaddr) { 599 paddr = _bus_dma_extract(pmap, vaddr); 600 if (run_filter(dmat, paddr) != 0) 601 map->pagesneeded++; 602 vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK)); 603 } 604 } 605 606 /* Reserve Necessary Bounce Pages */ 607 if (map->pagesneeded != 0) { 608 struct bounce_zone *bz; 609 610 bz = dmat->bounce_zone; 611 BZ_LOCK(bz); 612 if (flags & BUS_DMA_NOWAIT) { 613 if (reserve_bounce_pages(dmat, map, 0) != 0) { 614 BZ_UNLOCK(bz); 615 error = ENOMEM; 616 goto free_bounce; 617 } 618 } else { 619 if (reserve_bounce_pages(dmat, map, 1) != 0) { 620 /* Queue us for resources */ 621 map->dmat = dmat; 622 map->buf = buf; 623 map->buflen = buflen; 624 625 STAILQ_INSERT_TAIL( 626 &dmat->bounce_zone->bounce_map_waitinglist, 627 map, links); 628 BZ_UNLOCK(bz); 629 630 return (EINPROGRESS); 631 } 632 } 633 BZ_UNLOCK(bz); 634 } 635 636 KKASSERT(*segp >= 1 && *segp <= nsegments); 637 seg = *segp; 638 sg = &segments[seg - 1]; 639 640 vaddr = (vm_offset_t)buf; 641 nextpaddr = *lastpaddrp; 642 bmask = ~(dmat->boundary - 1); /* note: will be 0 if boundary is 0 */ 643 644 /* force at least one segment */ 645 do { 646 bus_size_t size; 647 648 /* 649 * Per-page main loop 650 */ 651 paddr = _bus_dma_extract(pmap, vaddr); 652 size = PAGE_SIZE - (paddr & PAGE_MASK); 653 if (size > buflen) 654 size = buflen; 655 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) { 656 /* 657 * note: this paddr has the same in-page offset 658 * as vaddr and thus the paddr above, so the 659 * size does not have to be recalculated 660 */ 661 paddr = add_bounce_page(dmat, map, vaddr, size); 662 } 663 664 /* 665 * Fill in the bus_dma_segment 666 */ 667 if (first) { 668 sg->ds_addr = paddr; 669 sg->ds_len = size; 670 first = 0; 671 } else if (paddr == nextpaddr) { 672 sg->ds_len += size; 673 } else { 674 sg++; 675 seg++; 676 if (seg > nsegments) 677 break; 678 sg->ds_addr = paddr; 679 sg->ds_len = size; 680 } 681 nextpaddr = paddr + size; 682 683 /* 684 * Handle maxsegsz and boundary issues with a nested loop 685 */ 686 for (;;) { 687 bus_size_t tmpsize; 688 689 /* 690 * Limit to the boundary and maximum segment size 691 */ 692 if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) { 693 tmpsize = dmat->boundary - 694 (sg->ds_addr & ~bmask); 695 if (tmpsize > dmat->maxsegsz) 696 tmpsize = dmat->maxsegsz; 697 KKASSERT(tmpsize < sg->ds_len); 698 } else if (sg->ds_len > dmat->maxsegsz) { 699 tmpsize = dmat->maxsegsz; 700 } else { 701 break; 702 } 703 704 /* 705 * Futz, split the data into a new segment. 706 */ 707 if (seg >= nsegments) 708 goto fail; 709 sg[1].ds_len = sg[0].ds_len - tmpsize; 710 sg[1].ds_addr = sg[0].ds_addr + tmpsize; 711 sg[0].ds_len = tmpsize; 712 sg++; 713 seg++; 714 } 715 716 /* 717 * Adjust for loop 718 */ 719 buflen -= size; 720 vaddr += size; 721 } while (buflen > 0); 722 fail: 723 if (buflen != 0) 724 error = EFBIG; 725 726 *segp = seg; 727 *lastpaddrp = nextpaddr; 728 729 free_bounce: 730 if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) && 731 map != &nobounce_dmamap) { 732 _bus_dmamap_unload(dmat, map); 733 return_bounce_pages(dmat, map); 734 } 735 return error; 736 } 737 738 /* 739 * Map the buffer buf into bus space using the dmamap map. 740 */ 741 int 742 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf, 743 bus_size_t buflen, bus_dmamap_callback_t *callback, 744 void *callback_arg, int flags) 745 { 746 vm_paddr_t lastaddr = 0; 747 int error, nsegs = 1; 748 749 if (map != NULL) { 750 /* 751 * XXX 752 * Follow old semantics. Once all of the callers are fixed, 753 * we should get rid of these internal flag "adjustment". 754 */ 755 flags &= ~BUS_DMA_NOWAIT; 756 flags |= BUS_DMA_WAITOK; 757 758 map->callback = callback; 759 map->callback_arg = callback_arg; 760 } 761 762 error = _bus_dmamap_load_buffer(dmat, map, buf, buflen, 763 dmat->segments, dmat->nsegments, 764 NULL, flags, &lastaddr, &nsegs, 1); 765 if (error == EINPROGRESS) 766 return error; 767 768 callback(callback_arg, dmat->segments, nsegs, error); 769 return 0; 770 } 771 772 /* 773 * Like _bus_dmamap_load(), but for mbufs. 774 */ 775 int 776 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map, 777 struct mbuf *m0, 778 bus_dmamap_callback2_t *callback, void *callback_arg, 779 int flags) 780 { 781 int nsegs, error; 782 783 /* 784 * XXX 785 * Follow old semantics. Once all of the callers are fixed, 786 * we should get rid of these internal flag "adjustment". 787 */ 788 flags &= ~BUS_DMA_WAITOK; 789 flags |= BUS_DMA_NOWAIT; 790 791 error = bus_dmamap_load_mbuf_segment(dmat, map, m0, 792 dmat->segments, dmat->nsegments, &nsegs, flags); 793 if (error) { 794 /* force "no valid mappings" in callback */ 795 callback(callback_arg, dmat->segments, 0, 0, error); 796 } else { 797 callback(callback_arg, dmat->segments, nsegs, 798 m0->m_pkthdr.len, error); 799 } 800 return error; 801 } 802 803 int 804 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map, 805 struct mbuf *m0, 806 bus_dma_segment_t *segs, int maxsegs, 807 int *nsegs, int flags) 808 { 809 int error; 810 811 M_ASSERTPKTHDR(m0); 812 813 KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs)); 814 KASSERT(maxsegs <= dmat->nsegments, 815 ("%d too many segments, dmat only supports %d segments", 816 maxsegs, dmat->nsegments)); 817 KASSERT(flags & BUS_DMA_NOWAIT, 818 ("only BUS_DMA_NOWAIT is supported")); 819 820 if (m0->m_pkthdr.len <= dmat->maxsize) { 821 int first = 1; 822 vm_paddr_t lastaddr = 0; 823 struct mbuf *m; 824 825 *nsegs = 1; 826 error = 0; 827 for (m = m0; m != NULL && error == 0; m = m->m_next) { 828 if (m->m_len == 0) 829 continue; 830 831 error = _bus_dmamap_load_buffer(dmat, map, 832 m->m_data, m->m_len, 833 segs, maxsegs, 834 NULL, flags, &lastaddr, 835 nsegs, first); 836 if (error == ENOMEM && !first) { 837 /* 838 * Out of bounce pages due to too many 839 * fragments in the mbuf chain; return 840 * EFBIG instead. 841 */ 842 error = EFBIG; 843 } 844 first = 0; 845 } 846 #ifdef INVARIANTS 847 if (!error) 848 KKASSERT(*nsegs <= maxsegs && *nsegs >= 1); 849 #endif 850 } else { 851 *nsegs = 0; 852 error = EINVAL; 853 } 854 KKASSERT(error != EINPROGRESS); 855 return error; 856 } 857 858 /* 859 * Like _bus_dmamap_load(), but for uios. 860 */ 861 int 862 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map, 863 struct uio *uio, 864 bus_dmamap_callback2_t *callback, void *callback_arg, 865 int flags) 866 { 867 vm_paddr_t lastaddr; 868 int nsegs, error, first, i; 869 bus_size_t resid; 870 struct iovec *iov; 871 pmap_t pmap; 872 873 /* 874 * XXX 875 * Follow old semantics. Once all of the callers are fixed, 876 * we should get rid of these internal flag "adjustment". 877 */ 878 flags &= ~BUS_DMA_WAITOK; 879 flags |= BUS_DMA_NOWAIT; 880 881 resid = (bus_size_t)uio->uio_resid; 882 iov = uio->uio_iov; 883 884 if (uio->uio_segflg == UIO_USERSPACE) { 885 struct thread *td; 886 887 td = uio->uio_td; 888 KASSERT(td != NULL && td->td_proc != NULL, 889 ("bus_dmamap_load_uio: USERSPACE but no proc")); 890 pmap = vmspace_pmap(td->td_proc->p_vmspace); 891 } else { 892 pmap = NULL; 893 } 894 895 error = 0; 896 nsegs = 1; 897 first = 1; 898 lastaddr = 0; 899 for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) { 900 /* 901 * Now at the first iovec to load. Load each iovec 902 * until we have exhausted the residual count. 903 */ 904 bus_size_t minlen = 905 resid < iov[i].iov_len ? resid : iov[i].iov_len; 906 caddr_t addr = (caddr_t) iov[i].iov_base; 907 908 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen, 909 dmat->segments, dmat->nsegments, 910 pmap, flags, &lastaddr, &nsegs, first); 911 first = 0; 912 913 resid -= minlen; 914 } 915 916 if (error) { 917 /* force "no valid mappings" in callback */ 918 callback(callback_arg, dmat->segments, 0, 0, error); 919 } else { 920 callback(callback_arg, dmat->segments, nsegs, 921 (bus_size_t)uio->uio_resid, error); 922 } 923 return error; 924 } 925 926 /* 927 * Release the mapping held by map. 928 */ 929 void 930 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map) 931 { 932 struct bounce_page *bpage; 933 934 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) { 935 STAILQ_REMOVE_HEAD(&map->bpages, links); 936 free_bounce_page(dmat, bpage); 937 } 938 } 939 940 void 941 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op) 942 { 943 struct bounce_page *bpage; 944 945 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) { 946 /* 947 * Handle data bouncing. We might also 948 * want to add support for invalidating 949 * the caches on broken hardware 950 */ 951 switch (op) { 952 case BUS_DMASYNC_PREWRITE: 953 while (bpage != NULL) { 954 bcopy((void *)bpage->datavaddr, 955 (void *)bpage->vaddr, 956 bpage->datacount); 957 bpage = STAILQ_NEXT(bpage, links); 958 } 959 dmat->bounce_zone->total_bounced++; 960 break; 961 962 case BUS_DMASYNC_POSTREAD: 963 while (bpage != NULL) { 964 bcopy((void *)bpage->vaddr, 965 (void *)bpage->datavaddr, 966 bpage->datacount); 967 bpage = STAILQ_NEXT(bpage, links); 968 } 969 dmat->bounce_zone->total_bounced++; 970 break; 971 972 case BUS_DMASYNC_PREREAD: 973 case BUS_DMASYNC_POSTWRITE: 974 /* No-ops */ 975 break; 976 } 977 } 978 } 979 980 static int 981 alloc_bounce_zone(bus_dma_tag_t dmat) 982 { 983 struct bounce_zone *bz, *new_bz; 984 985 KASSERT(dmat->bounce_zone == NULL, 986 ("bounce zone was already assigned")); 987 988 new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO); 989 990 lwkt_gettoken(&bounce_zone_tok); 991 992 /* Check to see if we already have a suitable zone */ 993 STAILQ_FOREACH(bz, &bounce_zone_list, links) { 994 if (dmat->alignment <= bz->alignment && 995 dmat->lowaddr >= bz->lowaddr) { 996 lwkt_reltoken(&bounce_zone_tok); 997 998 dmat->bounce_zone = bz; 999 kfree(new_bz, M_DEVBUF); 1000 return 0; 1001 } 1002 } 1003 bz = new_bz; 1004 1005 spin_init(&bz->spin, "allocbouncezone"); 1006 STAILQ_INIT(&bz->bounce_page_list); 1007 STAILQ_INIT(&bz->bounce_map_waitinglist); 1008 bz->free_bpages = 0; 1009 bz->reserved_bpages = 0; 1010 bz->active_bpages = 0; 1011 bz->lowaddr = dmat->lowaddr; 1012 bz->alignment = round_page(dmat->alignment); 1013 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount); 1014 busdma_zonecount++; 1015 ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr); 1016 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links); 1017 1018 lwkt_reltoken(&bounce_zone_tok); 1019 1020 dmat->bounce_zone = bz; 1021 1022 sysctl_ctx_init(&bz->sysctl_ctx); 1023 bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx, 1024 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid, 1025 CTLFLAG_RD, 0, ""); 1026 if (bz->sysctl_tree == NULL) { 1027 sysctl_ctx_free(&bz->sysctl_ctx); 1028 return 0; /* XXX error code? */ 1029 } 1030 1031 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1032 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1033 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0, 1034 "Total bounce pages"); 1035 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1036 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1037 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0, 1038 "Free bounce pages"); 1039 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1040 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1041 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0, 1042 "Reserved bounce pages"); 1043 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1044 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1045 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0, 1046 "Active bounce pages"); 1047 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1048 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1049 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0, 1050 "Total bounce requests"); 1051 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1052 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1053 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0, 1054 "Total bounce requests that were deferred"); 1055 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1056 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1057 "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0, 1058 "Total bounce page reservations that were failed"); 1059 SYSCTL_ADD_STRING(&bz->sysctl_ctx, 1060 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1061 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, ""); 1062 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1063 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1064 "alignment", CTLFLAG_RD, &bz->alignment, 0, ""); 1065 1066 return 0; 1067 } 1068 1069 static int 1070 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags) 1071 { 1072 struct bounce_zone *bz = dmat->bounce_zone; 1073 int count = 0, mflags; 1074 1075 if (flags & BUS_DMA_NOWAIT) 1076 mflags = M_NOWAIT; 1077 else 1078 mflags = M_WAITOK; 1079 1080 while (numpages > 0) { 1081 struct bounce_page *bpage; 1082 1083 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO); 1084 1085 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF, 1086 mflags, 0ul, 1087 bz->lowaddr, 1088 bz->alignment, 0); 1089 if (bpage->vaddr == 0) { 1090 kfree(bpage, M_DEVBUF); 1091 break; 1092 } 1093 bpage->busaddr = pmap_kextract(bpage->vaddr); 1094 1095 BZ_LOCK(bz); 1096 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links); 1097 total_bounce_pages++; 1098 bz->total_bpages++; 1099 bz->free_bpages++; 1100 BZ_UNLOCK(bz); 1101 1102 count++; 1103 numpages--; 1104 } 1105 return count; 1106 } 1107 1108 /* Assume caller holds bounce zone spinlock */ 1109 static int 1110 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit) 1111 { 1112 struct bounce_zone *bz = dmat->bounce_zone; 1113 int pages; 1114 1115 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved); 1116 if (!commit && map->pagesneeded > (map->pagesreserved + pages)) { 1117 bz->reserve_failed++; 1118 return (map->pagesneeded - (map->pagesreserved + pages)); 1119 } 1120 1121 bz->free_bpages -= pages; 1122 1123 bz->reserved_bpages += pages; 1124 KKASSERT(bz->reserved_bpages <= bz->total_bpages); 1125 1126 map->pagesreserved += pages; 1127 pages = map->pagesneeded - map->pagesreserved; 1128 1129 return pages; 1130 } 1131 1132 static void 1133 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map) 1134 { 1135 struct bounce_zone *bz = dmat->bounce_zone; 1136 int reserved = map->pagesreserved; 1137 bus_dmamap_t wait_map; 1138 1139 map->pagesreserved = 0; 1140 map->pagesneeded = 0; 1141 1142 if (reserved == 0) 1143 return; 1144 1145 BZ_LOCK(bz); 1146 1147 bz->free_bpages += reserved; 1148 KKASSERT(bz->free_bpages <= bz->total_bpages); 1149 1150 KKASSERT(bz->reserved_bpages >= reserved); 1151 bz->reserved_bpages -= reserved; 1152 1153 wait_map = get_map_waiting(dmat); 1154 1155 BZ_UNLOCK(bz); 1156 1157 if (wait_map != NULL) 1158 add_map_callback(map); 1159 } 1160 1161 static bus_addr_t 1162 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr, 1163 bus_size_t size) 1164 { 1165 struct bounce_zone *bz = dmat->bounce_zone; 1166 struct bounce_page *bpage; 1167 1168 KASSERT(map->pagesneeded > 0, ("map doesn't need any pages")); 1169 map->pagesneeded--; 1170 1171 KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages")); 1172 map->pagesreserved--; 1173 1174 BZ_LOCK(bz); 1175 1176 bpage = STAILQ_FIRST(&bz->bounce_page_list); 1177 KASSERT(bpage != NULL, ("free page list is empty")); 1178 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links); 1179 1180 KKASSERT(bz->reserved_bpages > 0); 1181 bz->reserved_bpages--; 1182 1183 bz->active_bpages++; 1184 KKASSERT(bz->active_bpages <= bz->total_bpages); 1185 1186 BZ_UNLOCK(bz); 1187 1188 bpage->datavaddr = vaddr; 1189 bpage->datacount = size; 1190 STAILQ_INSERT_TAIL(&map->bpages, bpage, links); 1191 return bpage->busaddr; 1192 } 1193 1194 static void 1195 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage) 1196 { 1197 struct bounce_zone *bz = dmat->bounce_zone; 1198 bus_dmamap_t map; 1199 1200 bpage->datavaddr = 0; 1201 bpage->datacount = 0; 1202 1203 BZ_LOCK(bz); 1204 1205 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links); 1206 1207 bz->free_bpages++; 1208 KKASSERT(bz->free_bpages <= bz->total_bpages); 1209 1210 KKASSERT(bz->active_bpages > 0); 1211 bz->active_bpages--; 1212 1213 map = get_map_waiting(dmat); 1214 1215 BZ_UNLOCK(bz); 1216 1217 if (map != NULL) 1218 add_map_callback(map); 1219 } 1220 1221 /* Assume caller holds bounce zone spinlock */ 1222 static bus_dmamap_t 1223 get_map_waiting(bus_dma_tag_t dmat) 1224 { 1225 struct bounce_zone *bz = dmat->bounce_zone; 1226 bus_dmamap_t map; 1227 1228 map = STAILQ_FIRST(&bz->bounce_map_waitinglist); 1229 if (map != NULL) { 1230 if (reserve_bounce_pages(map->dmat, map, 1) == 0) { 1231 STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links); 1232 bz->total_deferred++; 1233 } else { 1234 map = NULL; 1235 } 1236 } 1237 return map; 1238 } 1239 1240 static void 1241 add_map_callback(bus_dmamap_t map) 1242 { 1243 #ifdef notyet 1244 /* XXX callbacklist is not MPSAFE */ 1245 crit_enter(); 1246 get_mplock(); 1247 STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links); 1248 busdma_swi_pending = 1; 1249 setsoftvm(); 1250 rel_mplock(); 1251 crit_exit(); 1252 #else 1253 panic("%s uncoded", __func__); 1254 #endif 1255 } 1256 1257 #ifdef notyet 1258 void 1259 busdma_swi(void) 1260 { 1261 bus_dmamap_t map; 1262 1263 crit_enter(); 1264 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) { 1265 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links); 1266 crit_exit(); 1267 bus_dmamap_load(map->dmat, map, map->buf, map->buflen, 1268 map->callback, map->callback_arg, /*flags*/0); 1269 crit_enter(); 1270 } 1271 crit_exit(); 1272 } 1273 #endif 1274