1 /* 2 ** PARISC 1.1 Dynamic DMA mapping support. 3 ** This implementation is for PA-RISC platforms that do not support 4 ** I/O TLBs (aka DMA address translation hardware). 5 ** See Documentation/DMA-mapping.txt for interface definitions. 6 ** 7 ** (c) Copyright 1999,2000 Hewlett-Packard Company 8 ** (c) Copyright 2000 Grant Grundler 9 ** (c) Copyright 2000 Philipp Rumpf <prumpf@tux.org> 10 ** (c) Copyright 2000 John Marvin 11 ** 12 ** "leveraged" from 2.3.47: arch/ia64/kernel/pci-dma.c. 13 ** (I assume it's from David Mosberger-Tang but there was no Copyright) 14 ** 15 ** AFAIK, all PA7100LC and PA7300LC platforms can use this code. 16 ** 17 ** - ggg 18 */ 19 20 #include <linux/init.h> 21 #include <linux/mm.h> 22 #include <linux/pci.h> 23 #include <linux/proc_fs.h> 24 #include <linux/seq_file.h> 25 #include <linux/slab.h> 26 #include <linux/string.h> 27 #include <linux/types.h> 28 #include <linux/scatterlist.h> 29 30 #include <asm/cacheflush.h> 31 #include <asm/dma.h> /* for DMA_CHUNK_SIZE */ 32 #include <asm/io.h> 33 #include <asm/page.h> /* get_order */ 34 #include <asm/pgalloc.h> 35 #include <asm/uaccess.h> 36 #include <asm/tlbflush.h> /* for purge_tlb_*() macros */ 37 38 static struct proc_dir_entry * proc_gsc_root __read_mostly = NULL; 39 static unsigned long pcxl_used_bytes __read_mostly = 0; 40 static unsigned long pcxl_used_pages __read_mostly = 0; 41 42 extern unsigned long pcxl_dma_start; /* Start of pcxl dma mapping area */ 43 static spinlock_t pcxl_res_lock; 44 static char *pcxl_res_map; 45 static int pcxl_res_hint; 46 static int pcxl_res_size; 47 48 #ifdef DEBUG_PCXL_RESOURCE 49 #define DBG_RES(x...) printk(x) 50 #else 51 #define DBG_RES(x...) 52 #endif 53 54 55 /* 56 ** Dump a hex representation of the resource map. 57 */ 58 59 #ifdef DUMP_RESMAP 60 static 61 void dump_resmap(void) 62 { 63 u_long *res_ptr = (unsigned long *)pcxl_res_map; 64 u_long i = 0; 65 66 printk("res_map: "); 67 for(; i < (pcxl_res_size / sizeof(unsigned long)); ++i, ++res_ptr) 68 printk("%08lx ", *res_ptr); 69 70 printk("\n"); 71 } 72 #else 73 static inline void dump_resmap(void) {;} 74 #endif 75 76 static int pa11_dma_supported( struct device *dev, u64 mask) 77 { 78 return 1; 79 } 80 81 static inline int map_pte_uncached(pte_t * pte, 82 unsigned long vaddr, 83 unsigned long size, unsigned long *paddr_ptr) 84 { 85 unsigned long end; 86 unsigned long orig_vaddr = vaddr; 87 88 vaddr &= ~PMD_MASK; 89 end = vaddr + size; 90 if (end > PMD_SIZE) 91 end = PMD_SIZE; 92 do { 93 if (!pte_none(*pte)) 94 printk(KERN_ERR "map_pte_uncached: page already exists\n"); 95 set_pte(pte, __mk_pte(*paddr_ptr, PAGE_KERNEL_UNC)); 96 purge_tlb_start(); 97 pdtlb_kernel(orig_vaddr); 98 purge_tlb_end(); 99 vaddr += PAGE_SIZE; 100 orig_vaddr += PAGE_SIZE; 101 (*paddr_ptr) += PAGE_SIZE; 102 pte++; 103 } while (vaddr < end); 104 return 0; 105 } 106 107 static inline int map_pmd_uncached(pmd_t * pmd, unsigned long vaddr, 108 unsigned long size, unsigned long *paddr_ptr) 109 { 110 unsigned long end; 111 unsigned long orig_vaddr = vaddr; 112 113 vaddr &= ~PGDIR_MASK; 114 end = vaddr + size; 115 if (end > PGDIR_SIZE) 116 end = PGDIR_SIZE; 117 do { 118 pte_t * pte = pte_alloc_kernel(pmd, vaddr); 119 if (!pte) 120 return -ENOMEM; 121 if (map_pte_uncached(pte, orig_vaddr, end - vaddr, paddr_ptr)) 122 return -ENOMEM; 123 vaddr = (vaddr + PMD_SIZE) & PMD_MASK; 124 orig_vaddr += PMD_SIZE; 125 pmd++; 126 } while (vaddr < end); 127 return 0; 128 } 129 130 static inline int map_uncached_pages(unsigned long vaddr, unsigned long size, 131 unsigned long paddr) 132 { 133 pgd_t * dir; 134 unsigned long end = vaddr + size; 135 136 dir = pgd_offset_k(vaddr); 137 do { 138 pmd_t *pmd; 139 140 pmd = pmd_alloc(NULL, dir, vaddr); 141 if (!pmd) 142 return -ENOMEM; 143 if (map_pmd_uncached(pmd, vaddr, end - vaddr, &paddr)) 144 return -ENOMEM; 145 vaddr = vaddr + PGDIR_SIZE; 146 dir++; 147 } while (vaddr && (vaddr < end)); 148 return 0; 149 } 150 151 static inline void unmap_uncached_pte(pmd_t * pmd, unsigned long vaddr, 152 unsigned long size) 153 { 154 pte_t * pte; 155 unsigned long end; 156 unsigned long orig_vaddr = vaddr; 157 158 if (pmd_none(*pmd)) 159 return; 160 if (pmd_bad(*pmd)) { 161 pmd_ERROR(*pmd); 162 pmd_clear(pmd); 163 return; 164 } 165 pte = pte_offset_map(pmd, vaddr); 166 vaddr &= ~PMD_MASK; 167 end = vaddr + size; 168 if (end > PMD_SIZE) 169 end = PMD_SIZE; 170 do { 171 pte_t page = *pte; 172 pte_clear(&init_mm, vaddr, pte); 173 purge_tlb_start(); 174 pdtlb_kernel(orig_vaddr); 175 purge_tlb_end(); 176 vaddr += PAGE_SIZE; 177 orig_vaddr += PAGE_SIZE; 178 pte++; 179 if (pte_none(page) || pte_present(page)) 180 continue; 181 printk(KERN_CRIT "Whee.. Swapped out page in kernel page table\n"); 182 } while (vaddr < end); 183 } 184 185 static inline void unmap_uncached_pmd(pgd_t * dir, unsigned long vaddr, 186 unsigned long size) 187 { 188 pmd_t * pmd; 189 unsigned long end; 190 unsigned long orig_vaddr = vaddr; 191 192 if (pgd_none(*dir)) 193 return; 194 if (pgd_bad(*dir)) { 195 pgd_ERROR(*dir); 196 pgd_clear(dir); 197 return; 198 } 199 pmd = pmd_offset(dir, vaddr); 200 vaddr &= ~PGDIR_MASK; 201 end = vaddr + size; 202 if (end > PGDIR_SIZE) 203 end = PGDIR_SIZE; 204 do { 205 unmap_uncached_pte(pmd, orig_vaddr, end - vaddr); 206 vaddr = (vaddr + PMD_SIZE) & PMD_MASK; 207 orig_vaddr += PMD_SIZE; 208 pmd++; 209 } while (vaddr < end); 210 } 211 212 static void unmap_uncached_pages(unsigned long vaddr, unsigned long size) 213 { 214 pgd_t * dir; 215 unsigned long end = vaddr + size; 216 217 dir = pgd_offset_k(vaddr); 218 do { 219 unmap_uncached_pmd(dir, vaddr, end - vaddr); 220 vaddr = vaddr + PGDIR_SIZE; 221 dir++; 222 } while (vaddr && (vaddr < end)); 223 } 224 225 #define PCXL_SEARCH_LOOP(idx, mask, size) \ 226 for(; res_ptr < res_end; ++res_ptr) \ 227 { \ 228 if(0 == ((*res_ptr) & mask)) { \ 229 *res_ptr |= mask; \ 230 idx = (int)((u_long)res_ptr - (u_long)pcxl_res_map); \ 231 pcxl_res_hint = idx + (size >> 3); \ 232 goto resource_found; \ 233 } \ 234 } 235 236 #define PCXL_FIND_FREE_MAPPING(idx, mask, size) { \ 237 u##size *res_ptr = (u##size *)&(pcxl_res_map[pcxl_res_hint & ~((size >> 3) - 1)]); \ 238 u##size *res_end = (u##size *)&pcxl_res_map[pcxl_res_size]; \ 239 PCXL_SEARCH_LOOP(idx, mask, size); \ 240 res_ptr = (u##size *)&pcxl_res_map[0]; \ 241 PCXL_SEARCH_LOOP(idx, mask, size); \ 242 } 243 244 unsigned long 245 pcxl_alloc_range(size_t size) 246 { 247 int res_idx; 248 u_long mask, flags; 249 unsigned int pages_needed = size >> PAGE_SHIFT; 250 251 mask = (u_long) -1L; 252 mask >>= BITS_PER_LONG - pages_needed; 253 254 DBG_RES("pcxl_alloc_range() size: %d pages_needed %d pages_mask 0x%08lx\n", 255 size, pages_needed, mask); 256 257 spin_lock_irqsave(&pcxl_res_lock, flags); 258 259 if(pages_needed <= 8) { 260 PCXL_FIND_FREE_MAPPING(res_idx, mask, 8); 261 } else if(pages_needed <= 16) { 262 PCXL_FIND_FREE_MAPPING(res_idx, mask, 16); 263 } else if(pages_needed <= 32) { 264 PCXL_FIND_FREE_MAPPING(res_idx, mask, 32); 265 } else { 266 panic("%s: pcxl_alloc_range() Too many pages to map.\n", 267 __FILE__); 268 } 269 270 dump_resmap(); 271 panic("%s: pcxl_alloc_range() out of dma mapping resources\n", 272 __FILE__); 273 274 resource_found: 275 276 DBG_RES("pcxl_alloc_range() res_idx %d mask 0x%08lx res_hint: %d\n", 277 res_idx, mask, pcxl_res_hint); 278 279 pcxl_used_pages += pages_needed; 280 pcxl_used_bytes += ((pages_needed >> 3) ? (pages_needed >> 3) : 1); 281 282 spin_unlock_irqrestore(&pcxl_res_lock, flags); 283 284 dump_resmap(); 285 286 /* 287 ** return the corresponding vaddr in the pcxl dma map 288 */ 289 return (pcxl_dma_start + (res_idx << (PAGE_SHIFT + 3))); 290 } 291 292 #define PCXL_FREE_MAPPINGS(idx, m, size) \ 293 u##size *res_ptr = (u##size *)&(pcxl_res_map[(idx) + (((size >> 3) - 1) & (~((size >> 3) - 1)))]); \ 294 /* BUG_ON((*res_ptr & m) != m); */ \ 295 *res_ptr &= ~m; 296 297 /* 298 ** clear bits in the pcxl resource map 299 */ 300 static void 301 pcxl_free_range(unsigned long vaddr, size_t size) 302 { 303 u_long mask, flags; 304 unsigned int res_idx = (vaddr - pcxl_dma_start) >> (PAGE_SHIFT + 3); 305 unsigned int pages_mapped = size >> PAGE_SHIFT; 306 307 mask = (u_long) -1L; 308 mask >>= BITS_PER_LONG - pages_mapped; 309 310 DBG_RES("pcxl_free_range() res_idx: %d size: %d pages_mapped %d mask 0x%08lx\n", 311 res_idx, size, pages_mapped, mask); 312 313 spin_lock_irqsave(&pcxl_res_lock, flags); 314 315 if(pages_mapped <= 8) { 316 PCXL_FREE_MAPPINGS(res_idx, mask, 8); 317 } else if(pages_mapped <= 16) { 318 PCXL_FREE_MAPPINGS(res_idx, mask, 16); 319 } else if(pages_mapped <= 32) { 320 PCXL_FREE_MAPPINGS(res_idx, mask, 32); 321 } else { 322 panic("%s: pcxl_free_range() Too many pages to unmap.\n", 323 __FILE__); 324 } 325 326 pcxl_used_pages -= (pages_mapped ? pages_mapped : 1); 327 pcxl_used_bytes -= ((pages_mapped >> 3) ? (pages_mapped >> 3) : 1); 328 329 spin_unlock_irqrestore(&pcxl_res_lock, flags); 330 331 dump_resmap(); 332 } 333 334 static int proc_pcxl_dma_show(struct seq_file *m, void *v) 335 { 336 #if 0 337 u_long i = 0; 338 unsigned long *res_ptr = (u_long *)pcxl_res_map; 339 #endif 340 unsigned long total_pages = pcxl_res_size << 3; /* 8 bits per byte */ 341 342 seq_printf(m, "\nDMA Mapping Area size : %d bytes (%ld pages)\n", 343 PCXL_DMA_MAP_SIZE, total_pages); 344 345 seq_printf(m, "Resource bitmap : %d bytes\n", pcxl_res_size); 346 347 seq_puts(m, " total: free: used: % used:\n"); 348 seq_printf(m, "blocks %8d %8ld %8ld %8ld%%\n", pcxl_res_size, 349 pcxl_res_size - pcxl_used_bytes, pcxl_used_bytes, 350 (pcxl_used_bytes * 100) / pcxl_res_size); 351 352 seq_printf(m, "pages %8ld %8ld %8ld %8ld%%\n", total_pages, 353 total_pages - pcxl_used_pages, pcxl_used_pages, 354 (pcxl_used_pages * 100 / total_pages)); 355 356 #if 0 357 seq_puts(m, "\nResource bitmap:"); 358 359 for(; i < (pcxl_res_size / sizeof(u_long)); ++i, ++res_ptr) { 360 if ((i & 7) == 0) 361 seq_puts(m,"\n "); 362 seq_printf(m, "%s %08lx", buf, *res_ptr); 363 } 364 #endif 365 seq_putc(m, '\n'); 366 return 0; 367 } 368 369 static int proc_pcxl_dma_open(struct inode *inode, struct file *file) 370 { 371 return single_open(file, proc_pcxl_dma_show, NULL); 372 } 373 374 static const struct file_operations proc_pcxl_dma_ops = { 375 .owner = THIS_MODULE, 376 .open = proc_pcxl_dma_open, 377 .read = seq_read, 378 .llseek = seq_lseek, 379 .release = single_release, 380 }; 381 382 static int __init 383 pcxl_dma_init(void) 384 { 385 if (pcxl_dma_start == 0) 386 return 0; 387 388 spin_lock_init(&pcxl_res_lock); 389 pcxl_res_size = PCXL_DMA_MAP_SIZE >> (PAGE_SHIFT + 3); 390 pcxl_res_hint = 0; 391 pcxl_res_map = (char *)__get_free_pages(GFP_KERNEL, 392 get_order(pcxl_res_size)); 393 memset(pcxl_res_map, 0, pcxl_res_size); 394 proc_gsc_root = proc_mkdir("gsc", NULL); 395 if (!proc_gsc_root) 396 printk(KERN_WARNING 397 "pcxl_dma_init: Unable to create gsc /proc dir entry\n"); 398 else { 399 struct proc_dir_entry* ent; 400 ent = create_proc_entry("pcxl_dma", 0, proc_gsc_root); 401 if (ent) 402 ent->proc_fops = &proc_pcxl_dma_ops; 403 else 404 printk(KERN_WARNING 405 "pci-dma.c: Unable to create pcxl_dma /proc entry.\n"); 406 } 407 return 0; 408 } 409 410 __initcall(pcxl_dma_init); 411 412 static void * pa11_dma_alloc_consistent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag) 413 { 414 unsigned long vaddr; 415 unsigned long paddr; 416 int order; 417 418 order = get_order(size); 419 size = 1 << (order + PAGE_SHIFT); 420 vaddr = pcxl_alloc_range(size); 421 paddr = __get_free_pages(flag, order); 422 flush_kernel_dcache_range(paddr, size); 423 paddr = __pa(paddr); 424 map_uncached_pages(vaddr, size, paddr); 425 *dma_handle = (dma_addr_t) paddr; 426 427 #if 0 428 /* This probably isn't needed to support EISA cards. 429 ** ISA cards will certainly only support 24-bit DMA addressing. 430 ** Not clear if we can, want, or need to support ISA. 431 */ 432 if (!dev || *dev->coherent_dma_mask < 0xffffffff) 433 gfp |= GFP_DMA; 434 #endif 435 return (void *)vaddr; 436 } 437 438 static void pa11_dma_free_consistent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle) 439 { 440 int order; 441 442 order = get_order(size); 443 size = 1 << (order + PAGE_SHIFT); 444 unmap_uncached_pages((unsigned long)vaddr, size); 445 pcxl_free_range((unsigned long)vaddr, size); 446 free_pages((unsigned long)__va(dma_handle), order); 447 } 448 449 static dma_addr_t pa11_dma_map_single(struct device *dev, void *addr, size_t size, enum dma_data_direction direction) 450 { 451 if (direction == DMA_NONE) { 452 printk(KERN_ERR "pa11_dma_map_single(PCI_DMA_NONE) called by %p\n", __builtin_return_address(0)); 453 BUG(); 454 } 455 456 flush_kernel_dcache_range((unsigned long) addr, size); 457 return virt_to_phys(addr); 458 } 459 460 static void pa11_dma_unmap_single(struct device *dev, dma_addr_t dma_handle, size_t size, enum dma_data_direction direction) 461 { 462 if (direction == DMA_NONE) { 463 printk(KERN_ERR "pa11_dma_unmap_single(PCI_DMA_NONE) called by %p\n", __builtin_return_address(0)); 464 BUG(); 465 } 466 467 if (direction == DMA_TO_DEVICE) 468 return; 469 470 /* 471 * For PCI_DMA_FROMDEVICE this flush is not necessary for the 472 * simple map/unmap case. However, it IS necessary if if 473 * pci_dma_sync_single_* has been called and the buffer reused. 474 */ 475 476 flush_kernel_dcache_range((unsigned long) phys_to_virt(dma_handle), size); 477 return; 478 } 479 480 static int pa11_dma_map_sg(struct device *dev, struct scatterlist *sglist, int nents, enum dma_data_direction direction) 481 { 482 int i; 483 484 if (direction == DMA_NONE) 485 BUG(); 486 487 for (i = 0; i < nents; i++, sglist++ ) { 488 unsigned long vaddr = sg_virt_addr(sglist); 489 sg_dma_address(sglist) = (dma_addr_t) virt_to_phys(vaddr); 490 sg_dma_len(sglist) = sglist->length; 491 flush_kernel_dcache_range(vaddr, sglist->length); 492 } 493 return nents; 494 } 495 496 static void pa11_dma_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents, enum dma_data_direction direction) 497 { 498 int i; 499 500 if (direction == DMA_NONE) 501 BUG(); 502 503 if (direction == DMA_TO_DEVICE) 504 return; 505 506 /* once we do combining we'll need to use phys_to_virt(sg_dma_address(sglist)) */ 507 508 for (i = 0; i < nents; i++, sglist++ ) 509 flush_kernel_dcache_range(sg_virt_addr(sglist), sglist->length); 510 return; 511 } 512 513 static void pa11_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, unsigned long offset, size_t size, enum dma_data_direction direction) 514 { 515 if (direction == DMA_NONE) 516 BUG(); 517 518 flush_kernel_dcache_range((unsigned long) phys_to_virt(dma_handle) + offset, size); 519 } 520 521 static void pa11_dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle, unsigned long offset, size_t size, enum dma_data_direction direction) 522 { 523 if (direction == DMA_NONE) 524 BUG(); 525 526 flush_kernel_dcache_range((unsigned long) phys_to_virt(dma_handle) + offset, size); 527 } 528 529 static void pa11_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sglist, int nents, enum dma_data_direction direction) 530 { 531 int i; 532 533 /* once we do combining we'll need to use phys_to_virt(sg_dma_address(sglist)) */ 534 535 for (i = 0; i < nents; i++, sglist++ ) 536 flush_kernel_dcache_range(sg_virt_addr(sglist), sglist->length); 537 } 538 539 static void pa11_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sglist, int nents, enum dma_data_direction direction) 540 { 541 int i; 542 543 /* once we do combining we'll need to use phys_to_virt(sg_dma_address(sglist)) */ 544 545 for (i = 0; i < nents; i++, sglist++ ) 546 flush_kernel_dcache_range(sg_virt_addr(sglist), sglist->length); 547 } 548 549 struct hppa_dma_ops pcxl_dma_ops = { 550 .dma_supported = pa11_dma_supported, 551 .alloc_consistent = pa11_dma_alloc_consistent, 552 .alloc_noncoherent = pa11_dma_alloc_consistent, 553 .free_consistent = pa11_dma_free_consistent, 554 .map_single = pa11_dma_map_single, 555 .unmap_single = pa11_dma_unmap_single, 556 .map_sg = pa11_dma_map_sg, 557 .unmap_sg = pa11_dma_unmap_sg, 558 .dma_sync_single_for_cpu = pa11_dma_sync_single_for_cpu, 559 .dma_sync_single_for_device = pa11_dma_sync_single_for_device, 560 .dma_sync_sg_for_cpu = pa11_dma_sync_sg_for_cpu, 561 .dma_sync_sg_for_device = pa11_dma_sync_sg_for_device, 562 }; 563 564 static void *fail_alloc_consistent(struct device *dev, size_t size, 565 dma_addr_t *dma_handle, gfp_t flag) 566 { 567 return NULL; 568 } 569 570 static void *pa11_dma_alloc_noncoherent(struct device *dev, size_t size, 571 dma_addr_t *dma_handle, gfp_t flag) 572 { 573 void *addr; 574 575 addr = (void *)__get_free_pages(flag, get_order(size)); 576 if (addr) 577 *dma_handle = (dma_addr_t)virt_to_phys(addr); 578 579 return addr; 580 } 581 582 static void pa11_dma_free_noncoherent(struct device *dev, size_t size, 583 void *vaddr, dma_addr_t iova) 584 { 585 free_pages((unsigned long)vaddr, get_order(size)); 586 return; 587 } 588 589 struct hppa_dma_ops pcx_dma_ops = { 590 .dma_supported = pa11_dma_supported, 591 .alloc_consistent = fail_alloc_consistent, 592 .alloc_noncoherent = pa11_dma_alloc_noncoherent, 593 .free_consistent = pa11_dma_free_noncoherent, 594 .map_single = pa11_dma_map_single, 595 .unmap_single = pa11_dma_unmap_single, 596 .map_sg = pa11_dma_map_sg, 597 .unmap_sg = pa11_dma_unmap_sg, 598 .dma_sync_single_for_cpu = pa11_dma_sync_single_for_cpu, 599 .dma_sync_single_for_device = pa11_dma_sync_single_for_device, 600 .dma_sync_sg_for_cpu = pa11_dma_sync_sg_for_cpu, 601 .dma_sync_sg_for_device = pa11_dma_sync_sg_for_device, 602 }; 603