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