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 <sys/lock.h> 47 #include <vm/vm_map.h> 48 49 #include <machine/md_var.h> 50 51 #define MAX_BPAGES 1024 52 53 struct bounce_zone; 54 struct bus_dmamap; 55 56 struct bus_dma_tag { 57 bus_dma_tag_t parent; 58 bus_size_t alignment; 59 bus_size_t boundary; 60 bus_addr_t lowaddr; 61 bus_addr_t highaddr; 62 bus_dma_filter_t *filter; 63 void *filterarg; 64 bus_size_t maxsize; 65 u_int nsegments; 66 bus_size_t maxsegsz; 67 int flags; 68 int ref_count; 69 int map_count; 70 bus_dma_segment_t *segments; 71 struct bounce_zone *bounce_zone; 72 }; 73 74 /* 75 * bus_dma_tag private flags 76 */ 77 #define BUS_DMA_BOUNCE_ALIGN BUS_DMA_BUS2 78 #define BUS_DMA_BOUNCE_LOWADDR BUS_DMA_BUS3 79 #define BUS_DMA_MIN_ALLOC_COMP BUS_DMA_BUS4 80 81 #define BUS_DMA_COULD_BOUNCE (BUS_DMA_BOUNCE_LOWADDR | BUS_DMA_BOUNCE_ALIGN) 82 83 #define BUS_DMAMEM_KMALLOC(dmat) \ 84 ((dmat)->maxsize <= PAGE_SIZE && \ 85 (dmat)->alignment <= PAGE_SIZE && \ 86 (dmat)->lowaddr >= ptoa(Maxmem)) 87 88 struct bounce_page { 89 vm_offset_t vaddr; /* kva of bounce buffer */ 90 bus_addr_t busaddr; /* Physical address */ 91 vm_offset_t datavaddr; /* kva of client data */ 92 bus_size_t datacount; /* client data count */ 93 STAILQ_ENTRY(bounce_page) links; 94 }; 95 96 struct bounce_zone { 97 STAILQ_ENTRY(bounce_zone) links; 98 STAILQ_HEAD(bp_list, bounce_page) bounce_page_list; 99 STAILQ_HEAD(, bus_dmamap) bounce_map_waitinglist; 100 struct spinlock spin; 101 int total_bpages; 102 int free_bpages; 103 int reserved_bpages; 104 int active_bpages; 105 int total_bounced; 106 int total_deferred; 107 int reserve_failed; 108 bus_size_t alignment; 109 bus_addr_t lowaddr; 110 char zoneid[8]; 111 char lowaddrid[20]; 112 struct sysctl_ctx_list sysctl_ctx; 113 struct sysctl_oid *sysctl_tree; 114 }; 115 116 #define BZ_LOCK(bz) spin_lock(&(bz)->spin) 117 #define BZ_UNLOCK(bz) spin_unlock(&(bz)->spin) 118 119 static struct lwkt_token bounce_zone_tok = 120 LWKT_TOKEN_INITIALIZER(bounce_zone_token); 121 static int busdma_zonecount; 122 static STAILQ_HEAD(, bounce_zone) bounce_zone_list = 123 STAILQ_HEAD_INITIALIZER(bounce_zone_list); 124 125 int busdma_swi_pending; 126 static int total_bounce_pages; 127 static int max_bounce_pages = MAX_BPAGES; 128 static int bounce_alignment = 1; /* XXX temporary */ 129 130 TUNABLE_INT("hw.busdma.max_bpages", &max_bounce_pages); 131 TUNABLE_INT("hw.busdma.bounce_alignment", &bounce_alignment); 132 133 struct bus_dmamap { 134 struct bp_list bpages; 135 int pagesneeded; 136 int pagesreserved; 137 bus_dma_tag_t dmat; 138 void *buf; /* unmapped buffer pointer */ 139 bus_size_t buflen; /* unmapped buffer length */ 140 bus_dmamap_callback_t *callback; 141 void *callback_arg; 142 STAILQ_ENTRY(bus_dmamap) links; 143 }; 144 145 static STAILQ_HEAD(, bus_dmamap) bounce_map_callbacklist = 146 STAILQ_HEAD_INITIALIZER(bounce_map_callbacklist); 147 148 static struct bus_dmamap nobounce_dmamap; 149 150 static int alloc_bounce_zone(bus_dma_tag_t); 151 static int alloc_bounce_pages(bus_dma_tag_t, u_int, int); 152 static int reserve_bounce_pages(bus_dma_tag_t, bus_dmamap_t, int); 153 static void return_bounce_pages(bus_dma_tag_t, bus_dmamap_t); 154 static bus_addr_t add_bounce_page(bus_dma_tag_t, bus_dmamap_t, 155 vm_offset_t, bus_size_t); 156 static void free_bounce_page(bus_dma_tag_t, struct bounce_page *); 157 158 static bus_dmamap_t get_map_waiting(bus_dma_tag_t); 159 static void add_map_callback(bus_dmamap_t); 160 161 SYSCTL_NODE(_hw, OID_AUTO, busdma, CTLFLAG_RD, 0, "Busdma parameters"); 162 SYSCTL_INT(_hw_busdma, OID_AUTO, total_bpages, CTLFLAG_RD, &total_bounce_pages, 163 0, "Total bounce pages"); 164 SYSCTL_INT(_hw_busdma, OID_AUTO, max_bpages, CTLFLAG_RD, &max_bounce_pages, 165 0, "Max bounce pages per bounce zone"); 166 SYSCTL_INT(_hw_busdma, OID_AUTO, bounce_alignment, CTLFLAG_RD, 167 &bounce_alignment, 0, "Obey alignment constraint"); 168 169 static __inline int 170 run_filter(bus_dma_tag_t dmat, bus_addr_t paddr) 171 { 172 int retval; 173 174 retval = 0; 175 do { 176 if (((paddr > dmat->lowaddr && paddr <= dmat->highaddr) || 177 (bounce_alignment && (paddr & (dmat->alignment - 1)) != 0)) 178 && (dmat->filter == NULL || 179 dmat->filter(dmat->filterarg, paddr) != 0)) 180 retval = 1; 181 182 dmat = dmat->parent; 183 } while (retval == 0 && dmat != NULL); 184 return (retval); 185 } 186 187 /* 188 * Allocate a device specific dma_tag. 189 */ 190 int 191 bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment, 192 bus_size_t boundary, bus_addr_t lowaddr, 193 bus_addr_t highaddr, bus_dma_filter_t *filter, 194 void *filterarg, bus_size_t maxsize, int nsegments, 195 bus_size_t maxsegsz, int flags, bus_dma_tag_t *dmat) 196 { 197 bus_dma_tag_t newtag; 198 int error = 0; 199 200 /* 201 * Sanity checks 202 */ 203 204 if (alignment == 0) 205 alignment = 1; 206 if (alignment & (alignment - 1)) 207 panic("alignment must be power of 2"); 208 209 if (boundary != 0) { 210 if (boundary & (boundary - 1)) 211 panic("boundary must be power of 2"); 212 if (boundary < maxsegsz) { 213 kprintf("boundary < maxsegsz:\n"); 214 print_backtrace(-1); 215 maxsegsz = boundary; 216 } 217 } 218 219 /* Return a NULL tag on failure */ 220 *dmat = NULL; 221 222 newtag = kmalloc(sizeof(*newtag), M_DEVBUF, M_INTWAIT); 223 224 newtag->parent = parent; 225 newtag->alignment = alignment; 226 newtag->boundary = boundary; 227 newtag->lowaddr = trunc_page((vm_paddr_t)lowaddr) + (PAGE_SIZE - 1); 228 newtag->highaddr = trunc_page((vm_paddr_t)highaddr) + (PAGE_SIZE - 1); 229 newtag->filter = filter; 230 newtag->filterarg = filterarg; 231 newtag->maxsize = maxsize; 232 newtag->nsegments = nsegments; 233 newtag->maxsegsz = maxsegsz; 234 newtag->flags = flags; 235 newtag->ref_count = 1; /* Count ourself */ 236 newtag->map_count = 0; 237 newtag->segments = NULL; 238 newtag->bounce_zone = NULL; 239 240 /* Take into account any restrictions imposed by our parent tag */ 241 if (parent != NULL) { 242 newtag->lowaddr = MIN(parent->lowaddr, newtag->lowaddr); 243 newtag->highaddr = MAX(parent->highaddr, newtag->highaddr); 244 245 if (newtag->boundary == 0) { 246 newtag->boundary = parent->boundary; 247 } else if (parent->boundary != 0) { 248 newtag->boundary = MIN(parent->boundary, 249 newtag->boundary); 250 } 251 252 #ifdef notyet 253 newtag->alignment = MAX(parent->alignment, newtag->alignment); 254 #endif 255 256 if (newtag->filter == NULL) { 257 /* 258 * Short circuit looking at our parent directly 259 * since we have encapsulated all of its information 260 */ 261 newtag->filter = parent->filter; 262 newtag->filterarg = parent->filterarg; 263 newtag->parent = parent->parent; 264 } 265 if (newtag->parent != NULL) 266 parent->ref_count++; 267 } 268 269 if (newtag->lowaddr < ptoa(Maxmem)) 270 newtag->flags |= BUS_DMA_BOUNCE_LOWADDR; 271 if (bounce_alignment && newtag->alignment > 1 && 272 !(newtag->flags & BUS_DMA_ALIGNED)) 273 newtag->flags |= BUS_DMA_BOUNCE_ALIGN; 274 275 if ((newtag->flags & BUS_DMA_COULD_BOUNCE) && 276 (flags & BUS_DMA_ALLOCNOW) != 0) { 277 struct bounce_zone *bz; 278 279 /* Must bounce */ 280 281 error = alloc_bounce_zone(newtag); 282 if (error) 283 goto back; 284 bz = newtag->bounce_zone; 285 286 if (ptoa(bz->total_bpages) < maxsize) { 287 int pages; 288 289 if (flags & BUS_DMA_ONEBPAGE) { 290 pages = 1; 291 } else { 292 pages = atop(round_page(maxsize)) - 293 bz->total_bpages; 294 pages = MAX(pages, 1); 295 } 296 297 /* Add pages to our bounce pool */ 298 if (alloc_bounce_pages(newtag, pages, flags) < pages) 299 error = ENOMEM; 300 301 /* Performed initial allocation */ 302 newtag->flags |= BUS_DMA_MIN_ALLOC_COMP; 303 } 304 } 305 back: 306 if (error) 307 kfree(newtag, M_DEVBUF); 308 else 309 *dmat = newtag; 310 return error; 311 } 312 313 int 314 bus_dma_tag_destroy(bus_dma_tag_t dmat) 315 { 316 if (dmat != NULL) { 317 if (dmat->map_count != 0) 318 return (EBUSY); 319 320 while (dmat != NULL) { 321 bus_dma_tag_t parent; 322 323 parent = dmat->parent; 324 dmat->ref_count--; 325 if (dmat->ref_count == 0) { 326 if (dmat->segments != NULL) 327 kfree(dmat->segments, M_DEVBUF); 328 kfree(dmat, M_DEVBUF); 329 /* 330 * Last reference count, so 331 * release our reference 332 * count on our parent. 333 */ 334 dmat = parent; 335 } else 336 dmat = NULL; 337 } 338 } 339 return (0); 340 } 341 342 /* 343 * Allocate a handle for mapping from kva/uva/physical 344 * address space into bus device space. 345 */ 346 int 347 bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp) 348 { 349 int error; 350 351 error = 0; 352 353 if (dmat->segments == NULL) { 354 KKASSERT(dmat->nsegments && dmat->nsegments < 16384); 355 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) * 356 dmat->nsegments, M_DEVBUF, M_INTWAIT); 357 } 358 359 if (dmat->flags & BUS_DMA_COULD_BOUNCE) { 360 struct bounce_zone *bz; 361 int maxpages; 362 363 /* Must bounce */ 364 365 if (dmat->bounce_zone == NULL) { 366 error = alloc_bounce_zone(dmat); 367 if (error) 368 return error; 369 } 370 bz = dmat->bounce_zone; 371 372 *mapp = kmalloc(sizeof(**mapp), M_DEVBUF, M_INTWAIT | M_ZERO); 373 374 /* Initialize the new map */ 375 STAILQ_INIT(&((*mapp)->bpages)); 376 377 /* 378 * Attempt to add pages to our pool on a per-instance 379 * basis up to a sane limit. 380 */ 381 if (dmat->flags & BUS_DMA_BOUNCE_ALIGN) { 382 maxpages = max_bounce_pages; 383 } else { 384 maxpages = MIN(max_bounce_pages, 385 Maxmem - atop(dmat->lowaddr)); 386 } 387 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0 388 || (dmat->map_count > 0 389 && bz->total_bpages < maxpages)) { 390 int pages; 391 392 if (flags & BUS_DMA_ONEBPAGE) { 393 pages = 1; 394 } else { 395 pages = atop(round_page(dmat->maxsize)); 396 pages = MIN(maxpages - bz->total_bpages, pages); 397 pages = MAX(pages, 1); 398 } 399 if (alloc_bounce_pages(dmat, pages, flags) < pages) 400 error = ENOMEM; 401 402 if ((dmat->flags & BUS_DMA_MIN_ALLOC_COMP) == 0) { 403 if (!error) 404 dmat->flags |= BUS_DMA_MIN_ALLOC_COMP; 405 } else { 406 error = 0; 407 } 408 } 409 } else { 410 *mapp = NULL; 411 } 412 if (!error) 413 dmat->map_count++; 414 return error; 415 } 416 417 /* 418 * Destroy a handle for mapping from kva/uva/physical 419 * address space into bus device space. 420 */ 421 int 422 bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map) 423 { 424 if (map != NULL) { 425 if (STAILQ_FIRST(&map->bpages) != NULL) 426 return (EBUSY); 427 kfree(map, M_DEVBUF); 428 } 429 dmat->map_count--; 430 return (0); 431 } 432 433 static __inline bus_size_t 434 check_kmalloc(bus_dma_tag_t dmat, const void *vaddr0, int verify) 435 { 436 bus_size_t maxsize = 0; 437 uintptr_t vaddr = (uintptr_t)vaddr0; 438 439 if ((vaddr ^ (vaddr + dmat->maxsize - 1)) & ~PAGE_MASK) { 440 kprintf("boundary check failed\n"); 441 if (verify) 442 print_backtrace(-1); /* XXX panic */ 443 maxsize = dmat->maxsize; 444 } 445 if (vaddr & (dmat->alignment - 1)) { 446 kprintf("alignment check failed\n"); 447 if (verify) 448 print_backtrace(-1); /* XXX panic */ 449 if (dmat->maxsize < dmat->alignment) 450 maxsize = dmat->alignment; 451 else 452 maxsize = dmat->maxsize; 453 } 454 return maxsize; 455 } 456 457 /* 458 * Allocate a piece of memory that can be efficiently mapped into 459 * bus device space based on the constraints lited in the dma tag. 460 * 461 * mapp is degenerate. By definition this allocation should not require 462 * bounce buffers so do not allocate a dma map. 463 */ 464 int 465 bus_dmamem_alloc(bus_dma_tag_t dmat, void **vaddr, int flags, 466 bus_dmamap_t *mapp) 467 { 468 int mflags; 469 470 /* If we succeed, no mapping/bouncing will be required */ 471 *mapp = NULL; 472 473 if (dmat->segments == NULL) { 474 KKASSERT(dmat->nsegments < 16384); 475 dmat->segments = kmalloc(sizeof(bus_dma_segment_t) * 476 dmat->nsegments, M_DEVBUF, M_INTWAIT); 477 } 478 479 if (flags & BUS_DMA_NOWAIT) 480 mflags = M_NOWAIT; 481 else 482 mflags = M_WAITOK; 483 if (flags & BUS_DMA_ZERO) 484 mflags |= M_ZERO; 485 486 if (BUS_DMAMEM_KMALLOC(dmat)) { 487 bus_size_t maxsize; 488 489 *vaddr = kmalloc(dmat->maxsize, M_DEVBUF, mflags); 490 491 /* 492 * XXX 493 * Check whether the allocation 494 * - crossed a page boundary 495 * - was not aligned 496 * Retry with power-of-2 alignment in the above cases. 497 */ 498 maxsize = check_kmalloc(dmat, *vaddr, 0); 499 if (maxsize) { 500 size_t size; 501 502 kfree(*vaddr, M_DEVBUF); 503 /* XXX check for overflow? */ 504 for (size = 1; size <= maxsize; size <<= 1) 505 ; 506 *vaddr = kmalloc(size, M_DEVBUF, mflags); 507 check_kmalloc(dmat, *vaddr, 1); 508 } 509 } else { 510 /* 511 * XXX Use Contigmalloc until it is merged into this facility 512 * and handles multi-seg allocations. Nobody is doing 513 * multi-seg allocations yet though. 514 */ 515 *vaddr = contigmalloc(dmat->maxsize, M_DEVBUF, mflags, 516 0ul, dmat->lowaddr, dmat->alignment, dmat->boundary); 517 } 518 if (*vaddr == NULL) 519 return (ENOMEM); 520 return (0); 521 } 522 523 /* 524 * Free a piece of memory and it's allociated dmamap, that was allocated 525 * via bus_dmamem_alloc. Make the same choice for free/contigfree. 526 */ 527 void 528 bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map) 529 { 530 /* 531 * dmamem does not need to be bounced, so the map should be 532 * NULL 533 */ 534 if (map != NULL) 535 panic("bus_dmamem_free: Invalid map freed"); 536 if (BUS_DMAMEM_KMALLOC(dmat)) 537 kfree(vaddr, M_DEVBUF); 538 else 539 contigfree(vaddr, dmat->maxsize, M_DEVBUF); 540 } 541 542 static __inline vm_paddr_t 543 _bus_dma_extract(pmap_t pmap, vm_offset_t vaddr) 544 { 545 if (pmap) 546 return pmap_extract(pmap, vaddr); 547 else 548 return pmap_kextract(vaddr); 549 } 550 551 /* 552 * Utility function to load a linear buffer. lastaddrp holds state 553 * between invocations (for multiple-buffer loads). segp contains 554 * the segment following the starting one on entrace, and the ending 555 * segment on exit. first indicates if this is the first invocation 556 * of this function. 557 */ 558 static int 559 _bus_dmamap_load_buffer(bus_dma_tag_t dmat, 560 bus_dmamap_t map, 561 void *buf, bus_size_t buflen, 562 bus_dma_segment_t *segments, 563 int nsegments, 564 pmap_t pmap, 565 int flags, 566 vm_paddr_t *lastpaddrp, 567 int *segp, 568 int first) 569 { 570 vm_offset_t vaddr; 571 vm_paddr_t paddr, nextpaddr; 572 bus_dma_segment_t *sg; 573 bus_addr_t bmask; 574 int seg, error = 0; 575 576 if (map == NULL) 577 map = &nobounce_dmamap; 578 579 #ifdef INVARIANTS 580 if (dmat->flags & BUS_DMA_ALIGNED) 581 KKASSERT(((uintptr_t)buf & (dmat->alignment - 1)) == 0); 582 #endif 583 584 /* 585 * If we are being called during a callback, pagesneeded will 586 * be non-zero, so we can avoid doing the work twice. 587 */ 588 if ((dmat->flags & BUS_DMA_COULD_BOUNCE) && 589 map != &nobounce_dmamap && map->pagesneeded == 0) { 590 vm_offset_t vendaddr; 591 592 /* 593 * Count the number of bounce pages 594 * needed in order to complete this transfer 595 */ 596 vaddr = (vm_offset_t)buf; 597 vendaddr = (vm_offset_t)buf + buflen; 598 599 while (vaddr < vendaddr) { 600 paddr = _bus_dma_extract(pmap, vaddr); 601 if (run_filter(dmat, paddr) != 0) 602 map->pagesneeded++; 603 vaddr += (PAGE_SIZE - (vaddr & PAGE_MASK)); 604 } 605 } 606 607 /* Reserve Necessary Bounce Pages */ 608 if (map->pagesneeded != 0) { 609 struct bounce_zone *bz; 610 611 bz = dmat->bounce_zone; 612 BZ_LOCK(bz); 613 if (flags & BUS_DMA_NOWAIT) { 614 if (reserve_bounce_pages(dmat, map, 0) != 0) { 615 BZ_UNLOCK(bz); 616 error = ENOMEM; 617 goto free_bounce; 618 } 619 } else { 620 if (reserve_bounce_pages(dmat, map, 1) != 0) { 621 /* Queue us for resources */ 622 map->dmat = dmat; 623 map->buf = buf; 624 map->buflen = buflen; 625 626 STAILQ_INSERT_TAIL( 627 &dmat->bounce_zone->bounce_map_waitinglist, 628 map, links); 629 BZ_UNLOCK(bz); 630 631 return (EINPROGRESS); 632 } 633 } 634 BZ_UNLOCK(bz); 635 } 636 637 KKASSERT(*segp >= 1 && *segp <= nsegments); 638 seg = *segp; 639 sg = &segments[seg - 1]; 640 641 vaddr = (vm_offset_t)buf; 642 nextpaddr = *lastpaddrp; 643 bmask = ~(dmat->boundary - 1); /* note: will be 0 if boundary is 0 */ 644 645 /* force at least one segment */ 646 do { 647 bus_size_t size; 648 649 /* 650 * Per-page main loop 651 */ 652 paddr = _bus_dma_extract(pmap, vaddr); 653 size = PAGE_SIZE - (paddr & PAGE_MASK); 654 if (size > buflen) 655 size = buflen; 656 if (map->pagesneeded != 0 && run_filter(dmat, paddr)) { 657 /* 658 * note: this paddr has the same in-page offset 659 * as vaddr and thus the paddr above, so the 660 * size does not have to be recalculated 661 */ 662 paddr = add_bounce_page(dmat, map, vaddr, size); 663 } 664 665 /* 666 * Fill in the bus_dma_segment 667 */ 668 if (first) { 669 sg->ds_addr = paddr; 670 sg->ds_len = size; 671 first = 0; 672 } else if (paddr == nextpaddr) { 673 sg->ds_len += size; 674 } else { 675 sg++; 676 seg++; 677 if (seg > nsegments) 678 break; 679 sg->ds_addr = paddr; 680 sg->ds_len = size; 681 } 682 nextpaddr = paddr + size; 683 684 /* 685 * Handle maxsegsz and boundary issues with a nested loop 686 */ 687 for (;;) { 688 bus_size_t tmpsize; 689 690 /* 691 * Limit to the boundary and maximum segment size 692 */ 693 if (((nextpaddr - 1) ^ sg->ds_addr) & bmask) { 694 tmpsize = dmat->boundary - 695 (sg->ds_addr & ~bmask); 696 if (tmpsize > dmat->maxsegsz) 697 tmpsize = dmat->maxsegsz; 698 KKASSERT(tmpsize < sg->ds_len); 699 } else if (sg->ds_len > dmat->maxsegsz) { 700 tmpsize = dmat->maxsegsz; 701 } else { 702 break; 703 } 704 705 /* 706 * Futz, split the data into a new segment. 707 */ 708 if (seg >= nsegments) 709 goto fail; 710 sg[1].ds_len = sg[0].ds_len - tmpsize; 711 sg[1].ds_addr = sg[0].ds_addr + tmpsize; 712 sg[0].ds_len = tmpsize; 713 sg++; 714 seg++; 715 } 716 717 /* 718 * Adjust for loop 719 */ 720 buflen -= size; 721 vaddr += size; 722 } while (buflen > 0); 723 fail: 724 if (buflen != 0) 725 error = EFBIG; 726 727 *segp = seg; 728 *lastpaddrp = nextpaddr; 729 730 free_bounce: 731 if (error && (dmat->flags & BUS_DMA_COULD_BOUNCE) && 732 map != &nobounce_dmamap) { 733 _bus_dmamap_unload(dmat, map); 734 return_bounce_pages(dmat, map); 735 } 736 return error; 737 } 738 739 /* 740 * Map the buffer buf into bus space using the dmamap map. 741 */ 742 int 743 bus_dmamap_load(bus_dma_tag_t dmat, bus_dmamap_t map, void *buf, 744 bus_size_t buflen, bus_dmamap_callback_t *callback, 745 void *callback_arg, int flags) 746 { 747 vm_paddr_t lastaddr = 0; 748 int error, nsegs = 1; 749 750 if (map != NULL) { 751 /* 752 * XXX 753 * Follow old semantics. Once all of the callers are fixed, 754 * we should get rid of these internal flag "adjustment". 755 */ 756 flags &= ~BUS_DMA_NOWAIT; 757 flags |= BUS_DMA_WAITOK; 758 759 map->callback = callback; 760 map->callback_arg = callback_arg; 761 } 762 763 error = _bus_dmamap_load_buffer(dmat, map, buf, buflen, 764 dmat->segments, dmat->nsegments, 765 NULL, flags, &lastaddr, &nsegs, 1); 766 if (error == EINPROGRESS) 767 return error; 768 769 callback(callback_arg, dmat->segments, nsegs, error); 770 return 0; 771 } 772 773 /* 774 * Like _bus_dmamap_load(), but for mbufs. 775 */ 776 int 777 bus_dmamap_load_mbuf(bus_dma_tag_t dmat, bus_dmamap_t map, 778 struct mbuf *m0, 779 bus_dmamap_callback2_t *callback, void *callback_arg, 780 int flags) 781 { 782 int nsegs, error; 783 784 /* 785 * XXX 786 * Follow old semantics. Once all of the callers are fixed, 787 * we should get rid of these internal flag "adjustment". 788 */ 789 flags &= ~BUS_DMA_WAITOK; 790 flags |= BUS_DMA_NOWAIT; 791 792 error = bus_dmamap_load_mbuf_segment(dmat, map, m0, 793 dmat->segments, dmat->nsegments, &nsegs, flags); 794 if (error) { 795 /* force "no valid mappings" in callback */ 796 callback(callback_arg, dmat->segments, 0, 0, error); 797 } else { 798 callback(callback_arg, dmat->segments, nsegs, 799 m0->m_pkthdr.len, error); 800 } 801 return error; 802 } 803 804 int 805 bus_dmamap_load_mbuf_segment(bus_dma_tag_t dmat, bus_dmamap_t map, 806 struct mbuf *m0, 807 bus_dma_segment_t *segs, int maxsegs, 808 int *nsegs, int flags) 809 { 810 int error; 811 812 M_ASSERTPKTHDR(m0); 813 814 KASSERT(maxsegs >= 1, ("invalid maxsegs %d", maxsegs)); 815 KASSERT(maxsegs <= dmat->nsegments, 816 ("%d too many segments, dmat only supports %d segments", 817 maxsegs, dmat->nsegments)); 818 KASSERT(flags & BUS_DMA_NOWAIT, 819 ("only BUS_DMA_NOWAIT is supported")); 820 821 if (m0->m_pkthdr.len <= dmat->maxsize) { 822 int first = 1; 823 vm_paddr_t lastaddr = 0; 824 struct mbuf *m; 825 826 *nsegs = 1; 827 error = 0; 828 for (m = m0; m != NULL && error == 0; m = m->m_next) { 829 if (m->m_len == 0) 830 continue; 831 832 error = _bus_dmamap_load_buffer(dmat, map, 833 m->m_data, m->m_len, 834 segs, maxsegs, 835 NULL, flags, &lastaddr, 836 nsegs, first); 837 if (error == ENOMEM && !first) { 838 /* 839 * Out of bounce pages due to too many 840 * fragments in the mbuf chain; return 841 * EFBIG instead. 842 */ 843 error = EFBIG; 844 } 845 first = 0; 846 } 847 #ifdef INVARIANTS 848 if (!error) 849 KKASSERT(*nsegs <= maxsegs && *nsegs >= 1); 850 #endif 851 } else { 852 *nsegs = 0; 853 error = EINVAL; 854 } 855 KKASSERT(error != EINPROGRESS); 856 return error; 857 } 858 859 /* 860 * Like _bus_dmamap_load(), but for uios. 861 */ 862 int 863 bus_dmamap_load_uio(bus_dma_tag_t dmat, bus_dmamap_t map, 864 struct uio *uio, 865 bus_dmamap_callback2_t *callback, void *callback_arg, 866 int flags) 867 { 868 vm_paddr_t lastaddr; 869 int nsegs, error, first, i; 870 bus_size_t resid; 871 struct iovec *iov; 872 pmap_t pmap; 873 874 /* 875 * XXX 876 * Follow old semantics. Once all of the callers are fixed, 877 * we should get rid of these internal flag "adjustment". 878 */ 879 flags &= ~BUS_DMA_WAITOK; 880 flags |= BUS_DMA_NOWAIT; 881 882 resid = (bus_size_t)uio->uio_resid; 883 iov = uio->uio_iov; 884 885 if (uio->uio_segflg == UIO_USERSPACE) { 886 struct thread *td; 887 888 td = uio->uio_td; 889 KASSERT(td != NULL && td->td_proc != NULL, 890 ("bus_dmamap_load_uio: USERSPACE but no proc")); 891 pmap = vmspace_pmap(td->td_proc->p_vmspace); 892 } else { 893 pmap = NULL; 894 } 895 896 error = 0; 897 nsegs = 1; 898 first = 1; 899 lastaddr = 0; 900 for (i = 0; i < uio->uio_iovcnt && resid != 0 && !error; i++) { 901 /* 902 * Now at the first iovec to load. Load each iovec 903 * until we have exhausted the residual count. 904 */ 905 bus_size_t minlen = 906 resid < iov[i].iov_len ? resid : iov[i].iov_len; 907 caddr_t addr = (caddr_t) iov[i].iov_base; 908 909 error = _bus_dmamap_load_buffer(dmat, map, addr, minlen, 910 dmat->segments, dmat->nsegments, 911 pmap, flags, &lastaddr, &nsegs, first); 912 first = 0; 913 914 resid -= minlen; 915 } 916 917 if (error) { 918 /* force "no valid mappings" in callback */ 919 callback(callback_arg, dmat->segments, 0, 0, error); 920 } else { 921 callback(callback_arg, dmat->segments, nsegs, 922 (bus_size_t)uio->uio_resid, error); 923 } 924 return error; 925 } 926 927 /* 928 * Release the mapping held by map. 929 */ 930 void 931 _bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map) 932 { 933 struct bounce_page *bpage; 934 935 while ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) { 936 STAILQ_REMOVE_HEAD(&map->bpages, links); 937 free_bounce_page(dmat, bpage); 938 } 939 } 940 941 void 942 _bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op) 943 { 944 struct bounce_page *bpage; 945 946 if ((bpage = STAILQ_FIRST(&map->bpages)) != NULL) { 947 /* 948 * Handle data bouncing. We might also 949 * want to add support for invalidating 950 * the caches on broken hardware 951 */ 952 switch (op) { 953 case BUS_DMASYNC_PREWRITE: 954 while (bpage != NULL) { 955 bcopy((void *)bpage->datavaddr, 956 (void *)bpage->vaddr, 957 bpage->datacount); 958 bpage = STAILQ_NEXT(bpage, links); 959 } 960 dmat->bounce_zone->total_bounced++; 961 break; 962 963 case BUS_DMASYNC_POSTREAD: 964 while (bpage != NULL) { 965 bcopy((void *)bpage->vaddr, 966 (void *)bpage->datavaddr, 967 bpage->datacount); 968 bpage = STAILQ_NEXT(bpage, links); 969 } 970 dmat->bounce_zone->total_bounced++; 971 break; 972 973 case BUS_DMASYNC_PREREAD: 974 case BUS_DMASYNC_POSTWRITE: 975 /* No-ops */ 976 break; 977 } 978 } 979 } 980 981 static int 982 alloc_bounce_zone(bus_dma_tag_t dmat) 983 { 984 struct bounce_zone *bz, *new_bz; 985 986 KASSERT(dmat->bounce_zone == NULL, 987 ("bounce zone was already assigned")); 988 989 new_bz = kmalloc(sizeof(*new_bz), M_DEVBUF, M_INTWAIT | M_ZERO); 990 991 lwkt_gettoken(&bounce_zone_tok); 992 993 /* Check to see if we already have a suitable zone */ 994 STAILQ_FOREACH(bz, &bounce_zone_list, links) { 995 if (dmat->alignment <= bz->alignment && 996 dmat->lowaddr >= bz->lowaddr) { 997 lwkt_reltoken(&bounce_zone_tok); 998 999 dmat->bounce_zone = bz; 1000 kfree(new_bz, M_DEVBUF); 1001 return 0; 1002 } 1003 } 1004 bz = new_bz; 1005 1006 spin_init(&bz->spin); 1007 STAILQ_INIT(&bz->bounce_page_list); 1008 STAILQ_INIT(&bz->bounce_map_waitinglist); 1009 bz->free_bpages = 0; 1010 bz->reserved_bpages = 0; 1011 bz->active_bpages = 0; 1012 bz->lowaddr = dmat->lowaddr; 1013 bz->alignment = round_page(dmat->alignment); 1014 ksnprintf(bz->zoneid, 8, "zone%d", busdma_zonecount); 1015 busdma_zonecount++; 1016 ksnprintf(bz->lowaddrid, 18, "%#jx", (uintmax_t)bz->lowaddr); 1017 STAILQ_INSERT_TAIL(&bounce_zone_list, bz, links); 1018 1019 lwkt_reltoken(&bounce_zone_tok); 1020 1021 dmat->bounce_zone = bz; 1022 1023 sysctl_ctx_init(&bz->sysctl_ctx); 1024 bz->sysctl_tree = SYSCTL_ADD_NODE(&bz->sysctl_ctx, 1025 SYSCTL_STATIC_CHILDREN(_hw_busdma), OID_AUTO, bz->zoneid, 1026 CTLFLAG_RD, 0, ""); 1027 if (bz->sysctl_tree == NULL) { 1028 sysctl_ctx_free(&bz->sysctl_ctx); 1029 return 0; /* XXX error code? */ 1030 } 1031 1032 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1033 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1034 "total_bpages", CTLFLAG_RD, &bz->total_bpages, 0, 1035 "Total bounce pages"); 1036 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1037 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1038 "free_bpages", CTLFLAG_RD, &bz->free_bpages, 0, 1039 "Free bounce pages"); 1040 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1041 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1042 "reserved_bpages", CTLFLAG_RD, &bz->reserved_bpages, 0, 1043 "Reserved bounce pages"); 1044 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1045 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1046 "active_bpages", CTLFLAG_RD, &bz->active_bpages, 0, 1047 "Active bounce pages"); 1048 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1049 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1050 "total_bounced", CTLFLAG_RD, &bz->total_bounced, 0, 1051 "Total bounce requests"); 1052 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1053 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1054 "total_deferred", CTLFLAG_RD, &bz->total_deferred, 0, 1055 "Total bounce requests that were deferred"); 1056 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1057 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1058 "reserve_failed", CTLFLAG_RD, &bz->reserve_failed, 0, 1059 "Total bounce page reservations that were failed"); 1060 SYSCTL_ADD_STRING(&bz->sysctl_ctx, 1061 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1062 "lowaddr", CTLFLAG_RD, bz->lowaddrid, 0, ""); 1063 SYSCTL_ADD_INT(&bz->sysctl_ctx, 1064 SYSCTL_CHILDREN(bz->sysctl_tree), OID_AUTO, 1065 "alignment", CTLFLAG_RD, &bz->alignment, 0, ""); 1066 1067 return 0; 1068 } 1069 1070 static int 1071 alloc_bounce_pages(bus_dma_tag_t dmat, u_int numpages, int flags) 1072 { 1073 struct bounce_zone *bz = dmat->bounce_zone; 1074 int count = 0, mflags; 1075 1076 if (flags & BUS_DMA_NOWAIT) 1077 mflags = M_NOWAIT; 1078 else 1079 mflags = M_WAITOK; 1080 1081 while (numpages > 0) { 1082 struct bounce_page *bpage; 1083 1084 bpage = kmalloc(sizeof(*bpage), M_DEVBUF, M_INTWAIT | M_ZERO); 1085 1086 bpage->vaddr = (vm_offset_t)contigmalloc(PAGE_SIZE, M_DEVBUF, 1087 mflags, 0ul, 1088 bz->lowaddr, 1089 bz->alignment, 0); 1090 if (bpage->vaddr == 0) { 1091 kfree(bpage, M_DEVBUF); 1092 break; 1093 } 1094 bpage->busaddr = pmap_kextract(bpage->vaddr); 1095 1096 BZ_LOCK(bz); 1097 STAILQ_INSERT_TAIL(&bz->bounce_page_list, bpage, links); 1098 total_bounce_pages++; 1099 bz->total_bpages++; 1100 bz->free_bpages++; 1101 BZ_UNLOCK(bz); 1102 1103 count++; 1104 numpages--; 1105 } 1106 return count; 1107 } 1108 1109 /* Assume caller holds bounce zone spinlock */ 1110 static int 1111 reserve_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map, int commit) 1112 { 1113 struct bounce_zone *bz = dmat->bounce_zone; 1114 int pages; 1115 1116 pages = MIN(bz->free_bpages, map->pagesneeded - map->pagesreserved); 1117 if (!commit && map->pagesneeded > (map->pagesreserved + pages)) { 1118 bz->reserve_failed++; 1119 return (map->pagesneeded - (map->pagesreserved + pages)); 1120 } 1121 1122 bz->free_bpages -= pages; 1123 1124 bz->reserved_bpages += pages; 1125 KKASSERT(bz->reserved_bpages <= bz->total_bpages); 1126 1127 map->pagesreserved += pages; 1128 pages = map->pagesneeded - map->pagesreserved; 1129 1130 return pages; 1131 } 1132 1133 static void 1134 return_bounce_pages(bus_dma_tag_t dmat, bus_dmamap_t map) 1135 { 1136 struct bounce_zone *bz = dmat->bounce_zone; 1137 int reserved = map->pagesreserved; 1138 bus_dmamap_t wait_map; 1139 1140 map->pagesreserved = 0; 1141 map->pagesneeded = 0; 1142 1143 if (reserved == 0) 1144 return; 1145 1146 BZ_LOCK(bz); 1147 1148 bz->free_bpages += reserved; 1149 KKASSERT(bz->free_bpages <= bz->total_bpages); 1150 1151 KKASSERT(bz->reserved_bpages >= reserved); 1152 bz->reserved_bpages -= reserved; 1153 1154 wait_map = get_map_waiting(dmat); 1155 1156 BZ_UNLOCK(bz); 1157 1158 if (wait_map != NULL) 1159 add_map_callback(map); 1160 } 1161 1162 static bus_addr_t 1163 add_bounce_page(bus_dma_tag_t dmat, bus_dmamap_t map, vm_offset_t vaddr, 1164 bus_size_t size) 1165 { 1166 struct bounce_zone *bz = dmat->bounce_zone; 1167 struct bounce_page *bpage; 1168 1169 KASSERT(map->pagesneeded > 0, ("map doesn't need any pages")); 1170 map->pagesneeded--; 1171 1172 KASSERT(map->pagesreserved > 0, ("map doesn't reserve any pages")); 1173 map->pagesreserved--; 1174 1175 BZ_LOCK(bz); 1176 1177 bpage = STAILQ_FIRST(&bz->bounce_page_list); 1178 KASSERT(bpage != NULL, ("free page list is empty")); 1179 STAILQ_REMOVE_HEAD(&bz->bounce_page_list, links); 1180 1181 KKASSERT(bz->reserved_bpages > 0); 1182 bz->reserved_bpages--; 1183 1184 bz->active_bpages++; 1185 KKASSERT(bz->active_bpages <= bz->total_bpages); 1186 1187 BZ_UNLOCK(bz); 1188 1189 bpage->datavaddr = vaddr; 1190 bpage->datacount = size; 1191 STAILQ_INSERT_TAIL(&map->bpages, bpage, links); 1192 return bpage->busaddr; 1193 } 1194 1195 static void 1196 free_bounce_page(bus_dma_tag_t dmat, struct bounce_page *bpage) 1197 { 1198 struct bounce_zone *bz = dmat->bounce_zone; 1199 bus_dmamap_t map; 1200 1201 bpage->datavaddr = 0; 1202 bpage->datacount = 0; 1203 1204 BZ_LOCK(bz); 1205 1206 STAILQ_INSERT_HEAD(&bz->bounce_page_list, bpage, links); 1207 1208 bz->free_bpages++; 1209 KKASSERT(bz->free_bpages <= bz->total_bpages); 1210 1211 KKASSERT(bz->active_bpages > 0); 1212 bz->active_bpages--; 1213 1214 map = get_map_waiting(dmat); 1215 1216 BZ_UNLOCK(bz); 1217 1218 if (map != NULL) 1219 add_map_callback(map); 1220 } 1221 1222 /* Assume caller holds bounce zone spinlock */ 1223 static bus_dmamap_t 1224 get_map_waiting(bus_dma_tag_t dmat) 1225 { 1226 struct bounce_zone *bz = dmat->bounce_zone; 1227 bus_dmamap_t map; 1228 1229 map = STAILQ_FIRST(&bz->bounce_map_waitinglist); 1230 if (map != NULL) { 1231 if (reserve_bounce_pages(map->dmat, map, 1) == 0) { 1232 STAILQ_REMOVE_HEAD(&bz->bounce_map_waitinglist, links); 1233 bz->total_deferred++; 1234 } else { 1235 map = NULL; 1236 } 1237 } 1238 return map; 1239 } 1240 1241 static void 1242 add_map_callback(bus_dmamap_t map) 1243 { 1244 #ifdef notyet 1245 /* XXX callbacklist is not MPSAFE */ 1246 crit_enter(); 1247 get_mplock(); 1248 STAILQ_INSERT_TAIL(&bounce_map_callbacklist, map, links); 1249 busdma_swi_pending = 1; 1250 setsoftvm(); 1251 rel_mplock(); 1252 crit_exit(); 1253 #else 1254 panic("%s uncoded", __func__); 1255 #endif 1256 } 1257 1258 #ifdef notyet 1259 void 1260 busdma_swi(void) 1261 { 1262 bus_dmamap_t map; 1263 1264 crit_enter(); 1265 while ((map = STAILQ_FIRST(&bounce_map_callbacklist)) != NULL) { 1266 STAILQ_REMOVE_HEAD(&bounce_map_callbacklist, links); 1267 crit_exit(); 1268 bus_dmamap_load(map->dmat, map, map->buf, map->buflen, 1269 map->callback, map->callback_arg, /*flags*/0); 1270 crit_enter(); 1271 } 1272 crit_exit(); 1273 } 1274 #endif 1275