b.liu | e958203 | 2025-04-17 19:18:16 +0800 | [diff] [blame^] | 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * linux/arch/arm/mm/ioremap.c |
| 4 | * |
| 5 | * Re-map IO memory to kernel address space so that we can access it. |
| 6 | * |
| 7 | * (C) Copyright 1995 1996 Linus Torvalds |
| 8 | * |
| 9 | * Hacked for ARM by Phil Blundell <philb@gnu.org> |
| 10 | * Hacked to allow all architectures to build, and various cleanups |
| 11 | * by Russell King |
| 12 | * |
| 13 | * This allows a driver to remap an arbitrary region of bus memory into |
| 14 | * virtual space. One should *only* use readl, writel, memcpy_toio and |
| 15 | * so on with such remapped areas. |
| 16 | * |
| 17 | * Because the ARM only has a 32-bit address space we can't address the |
| 18 | * whole of the (physical) PCI space at once. PCI huge-mode addressing |
| 19 | * allows us to circumvent this restriction by splitting PCI space into |
| 20 | * two 2GB chunks and mapping only one at a time into processor memory. |
| 21 | * We use MMU protection domains to trap any attempt to access the bank |
| 22 | * that is not currently mapped. (This isn't fully implemented yet.) |
| 23 | */ |
| 24 | #include <linux/module.h> |
| 25 | #include <linux/errno.h> |
| 26 | #include <linux/mm.h> |
| 27 | #include <linux/vmalloc.h> |
| 28 | #include <linux/io.h> |
| 29 | #include <linux/sizes.h> |
| 30 | #include <linux/memblock.h> |
| 31 | |
| 32 | #include <asm/cp15.h> |
| 33 | #include <asm/cputype.h> |
| 34 | #include <asm/cacheflush.h> |
| 35 | #include <asm/early_ioremap.h> |
| 36 | #include <asm/mmu_context.h> |
| 37 | #include <asm/pgalloc.h> |
| 38 | #include <asm/tlbflush.h> |
| 39 | #include <asm/system_info.h> |
| 40 | |
| 41 | #include <asm/mach/map.h> |
| 42 | #include <asm/mach/pci.h> |
| 43 | #include "mm.h" |
| 44 | |
| 45 | |
| 46 | LIST_HEAD(static_vmlist); |
| 47 | |
| 48 | static struct static_vm *find_static_vm_paddr(phys_addr_t paddr, |
| 49 | size_t size, unsigned int mtype) |
| 50 | { |
| 51 | struct static_vm *svm; |
| 52 | struct vm_struct *vm; |
| 53 | |
| 54 | list_for_each_entry(svm, &static_vmlist, list) { |
| 55 | vm = &svm->vm; |
| 56 | if (!(vm->flags & VM_ARM_STATIC_MAPPING)) |
| 57 | continue; |
| 58 | if ((vm->flags & VM_ARM_MTYPE_MASK) != VM_ARM_MTYPE(mtype)) |
| 59 | continue; |
| 60 | |
| 61 | if (vm->phys_addr > paddr || |
| 62 | paddr + size - 1 > vm->phys_addr + vm->size - 1) |
| 63 | continue; |
| 64 | |
| 65 | return svm; |
| 66 | } |
| 67 | |
| 68 | return NULL; |
| 69 | } |
| 70 | |
| 71 | struct static_vm *find_static_vm_vaddr(void *vaddr) |
| 72 | { |
| 73 | struct static_vm *svm; |
| 74 | struct vm_struct *vm; |
| 75 | |
| 76 | list_for_each_entry(svm, &static_vmlist, list) { |
| 77 | vm = &svm->vm; |
| 78 | |
| 79 | /* static_vmlist is ascending order */ |
| 80 | if (vm->addr > vaddr) |
| 81 | break; |
| 82 | |
| 83 | if (vm->addr <= vaddr && vm->addr + vm->size > vaddr) |
| 84 | return svm; |
| 85 | } |
| 86 | |
| 87 | return NULL; |
| 88 | } |
| 89 | |
| 90 | void __init add_static_vm_early(struct static_vm *svm) |
| 91 | { |
| 92 | struct static_vm *curr_svm; |
| 93 | struct vm_struct *vm; |
| 94 | void *vaddr; |
| 95 | |
| 96 | vm = &svm->vm; |
| 97 | vm_area_add_early(vm); |
| 98 | vaddr = vm->addr; |
| 99 | |
| 100 | list_for_each_entry(curr_svm, &static_vmlist, list) { |
| 101 | vm = &curr_svm->vm; |
| 102 | |
| 103 | if (vm->addr > vaddr) |
| 104 | break; |
| 105 | } |
| 106 | list_add_tail(&svm->list, &curr_svm->list); |
| 107 | } |
| 108 | |
| 109 | int ioremap_page(unsigned long virt, unsigned long phys, |
| 110 | const struct mem_type *mtype) |
| 111 | { |
| 112 | return ioremap_page_range(virt, virt + PAGE_SIZE, phys, |
| 113 | __pgprot(mtype->prot_pte)); |
| 114 | } |
| 115 | EXPORT_SYMBOL(ioremap_page); |
| 116 | |
| 117 | void __check_vmalloc_seq(struct mm_struct *mm) |
| 118 | { |
| 119 | unsigned int seq; |
| 120 | |
| 121 | do { |
| 122 | seq = init_mm.context.vmalloc_seq; |
| 123 | memcpy(pgd_offset(mm, VMALLOC_START), |
| 124 | pgd_offset_k(VMALLOC_START), |
| 125 | sizeof(pgd_t) * (pgd_index(VMALLOC_END) - |
| 126 | pgd_index(VMALLOC_START))); |
| 127 | mm->context.vmalloc_seq = seq; |
| 128 | } while (seq != init_mm.context.vmalloc_seq); |
| 129 | } |
| 130 | |
| 131 | #if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE) |
| 132 | /* |
| 133 | * Section support is unsafe on SMP - If you iounmap and ioremap a region, |
| 134 | * the other CPUs will not see this change until their next context switch. |
| 135 | * Meanwhile, (eg) if an interrupt comes in on one of those other CPUs |
| 136 | * which requires the new ioremap'd region to be referenced, the CPU will |
| 137 | * reference the _old_ region. |
| 138 | * |
| 139 | * Note that get_vm_area_caller() allocates a guard 4K page, so we need to |
| 140 | * mask the size back to 1MB aligned or we will overflow in the loop below. |
| 141 | */ |
| 142 | static void unmap_area_sections(unsigned long virt, unsigned long size) |
| 143 | { |
| 144 | unsigned long addr = virt, end = virt + (size & ~(SZ_1M - 1)); |
| 145 | pgd_t *pgd; |
| 146 | pud_t *pud; |
| 147 | pmd_t *pmdp; |
| 148 | |
| 149 | flush_cache_vunmap(addr, end); |
| 150 | pgd = pgd_offset_k(addr); |
| 151 | pud = pud_offset(pgd, addr); |
| 152 | pmdp = pmd_offset(pud, addr); |
| 153 | do { |
| 154 | pmd_t pmd = *pmdp; |
| 155 | |
| 156 | if (!pmd_none(pmd)) { |
| 157 | /* |
| 158 | * Clear the PMD from the page table, and |
| 159 | * increment the vmalloc sequence so others |
| 160 | * notice this change. |
| 161 | * |
| 162 | * Note: this is still racy on SMP machines. |
| 163 | */ |
| 164 | pmd_clear(pmdp); |
| 165 | init_mm.context.vmalloc_seq++; |
| 166 | |
| 167 | /* |
| 168 | * Free the page table, if there was one. |
| 169 | */ |
| 170 | if ((pmd_val(pmd) & PMD_TYPE_MASK) == PMD_TYPE_TABLE) |
| 171 | pte_free_kernel(&init_mm, pmd_page_vaddr(pmd)); |
| 172 | } |
| 173 | |
| 174 | addr += PMD_SIZE; |
| 175 | pmdp += 2; |
| 176 | } while (addr < end); |
| 177 | |
| 178 | /* |
| 179 | * Ensure that the active_mm is up to date - we want to |
| 180 | * catch any use-after-iounmap cases. |
| 181 | */ |
| 182 | if (current->active_mm->context.vmalloc_seq != init_mm.context.vmalloc_seq) |
| 183 | __check_vmalloc_seq(current->active_mm); |
| 184 | |
| 185 | flush_tlb_kernel_range(virt, end); |
| 186 | } |
| 187 | |
| 188 | static int |
| 189 | remap_area_sections(unsigned long virt, unsigned long pfn, |
| 190 | size_t size, const struct mem_type *type) |
| 191 | { |
| 192 | unsigned long addr = virt, end = virt + size; |
| 193 | pgd_t *pgd; |
| 194 | pud_t *pud; |
| 195 | pmd_t *pmd; |
| 196 | |
| 197 | /* |
| 198 | * Remove and free any PTE-based mapping, and |
| 199 | * sync the current kernel mapping. |
| 200 | */ |
| 201 | unmap_area_sections(virt, size); |
| 202 | |
| 203 | pgd = pgd_offset_k(addr); |
| 204 | pud = pud_offset(pgd, addr); |
| 205 | pmd = pmd_offset(pud, addr); |
| 206 | do { |
| 207 | pmd[0] = __pmd(__pfn_to_phys(pfn) | type->prot_sect); |
| 208 | pfn += SZ_1M >> PAGE_SHIFT; |
| 209 | pmd[1] = __pmd(__pfn_to_phys(pfn) | type->prot_sect); |
| 210 | pfn += SZ_1M >> PAGE_SHIFT; |
| 211 | flush_pmd_entry(pmd); |
| 212 | |
| 213 | addr += PMD_SIZE; |
| 214 | pmd += 2; |
| 215 | } while (addr < end); |
| 216 | |
| 217 | return 0; |
| 218 | } |
| 219 | |
| 220 | static int |
| 221 | remap_area_supersections(unsigned long virt, unsigned long pfn, |
| 222 | size_t size, const struct mem_type *type) |
| 223 | { |
| 224 | unsigned long addr = virt, end = virt + size; |
| 225 | pgd_t *pgd; |
| 226 | pud_t *pud; |
| 227 | pmd_t *pmd; |
| 228 | |
| 229 | /* |
| 230 | * Remove and free any PTE-based mapping, and |
| 231 | * sync the current kernel mapping. |
| 232 | */ |
| 233 | unmap_area_sections(virt, size); |
| 234 | |
| 235 | pgd = pgd_offset_k(virt); |
| 236 | pud = pud_offset(pgd, addr); |
| 237 | pmd = pmd_offset(pud, addr); |
| 238 | do { |
| 239 | unsigned long super_pmd_val, i; |
| 240 | |
| 241 | super_pmd_val = __pfn_to_phys(pfn) | type->prot_sect | |
| 242 | PMD_SECT_SUPER; |
| 243 | super_pmd_val |= ((pfn >> (32 - PAGE_SHIFT)) & 0xf) << 20; |
| 244 | |
| 245 | for (i = 0; i < 8; i++) { |
| 246 | pmd[0] = __pmd(super_pmd_val); |
| 247 | pmd[1] = __pmd(super_pmd_val); |
| 248 | flush_pmd_entry(pmd); |
| 249 | |
| 250 | addr += PMD_SIZE; |
| 251 | pmd += 2; |
| 252 | } |
| 253 | |
| 254 | pfn += SUPERSECTION_SIZE >> PAGE_SHIFT; |
| 255 | } while (addr < end); |
| 256 | |
| 257 | return 0; |
| 258 | } |
| 259 | #endif |
| 260 | |
| 261 | static void __iomem * __arm_ioremap_pfn_caller(unsigned long pfn, |
| 262 | unsigned long offset, size_t size, unsigned int mtype, void *caller) |
| 263 | { |
| 264 | const struct mem_type *type; |
| 265 | int err; |
| 266 | unsigned long addr; |
| 267 | struct vm_struct *area; |
| 268 | phys_addr_t paddr = __pfn_to_phys(pfn); |
| 269 | |
| 270 | #ifndef CONFIG_ARM_LPAE |
| 271 | /* |
| 272 | * High mappings must be supersection aligned |
| 273 | */ |
| 274 | if (pfn >= 0x100000 && (paddr & ~SUPERSECTION_MASK)) |
| 275 | return NULL; |
| 276 | #endif |
| 277 | |
| 278 | type = get_mem_type(mtype); |
| 279 | if (!type) |
| 280 | return NULL; |
| 281 | |
| 282 | /* |
| 283 | * Page align the mapping size, taking account of any offset. |
| 284 | */ |
| 285 | size = PAGE_ALIGN(offset + size); |
| 286 | |
| 287 | /* |
| 288 | * Try to reuse one of the static mapping whenever possible. |
| 289 | */ |
| 290 | if (size && !(sizeof(phys_addr_t) == 4 && pfn >= 0x100000)) { |
| 291 | struct static_vm *svm; |
| 292 | |
| 293 | svm = find_static_vm_paddr(paddr, size, mtype); |
| 294 | if (svm) { |
| 295 | addr = (unsigned long)svm->vm.addr; |
| 296 | addr += paddr - svm->vm.phys_addr; |
| 297 | return (void __iomem *) (offset + addr); |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | /* |
| 302 | * Don't allow RAM to be mapped with mismatched attributes - this |
| 303 | * causes problems with ARMv6+ |
| 304 | */ |
| 305 | if (WARN_ON(memblock_is_map_memory(PFN_PHYS(pfn)) && |
| 306 | mtype != MT_MEMORY_RW)) |
| 307 | return NULL; |
| 308 | |
| 309 | area = get_vm_area_caller(size, VM_IOREMAP, caller); |
| 310 | if (!area) |
| 311 | return NULL; |
| 312 | addr = (unsigned long)area->addr; |
| 313 | area->phys_addr = paddr; |
| 314 | |
| 315 | #if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE) |
| 316 | if (DOMAIN_IO == 0 && |
| 317 | (((cpu_architecture() >= CPU_ARCH_ARMv6) && (get_cr() & CR_XP)) || |
| 318 | cpu_is_xsc3()) && pfn >= 0x100000 && |
| 319 | !((paddr | size | addr) & ~SUPERSECTION_MASK)) { |
| 320 | area->flags |= VM_ARM_SECTION_MAPPING; |
| 321 | err = remap_area_supersections(addr, pfn, size, type); |
| 322 | } else if (!((paddr | size | addr) & ~PMD_MASK)) { |
| 323 | area->flags |= VM_ARM_SECTION_MAPPING; |
| 324 | err = remap_area_sections(addr, pfn, size, type); |
| 325 | } else |
| 326 | #endif |
| 327 | err = ioremap_page_range(addr, addr + size, paddr, |
| 328 | __pgprot(type->prot_pte)); |
| 329 | |
| 330 | if (err) { |
| 331 | vunmap((void *)addr); |
| 332 | return NULL; |
| 333 | } |
| 334 | |
| 335 | flush_cache_vmap(addr, addr + size); |
| 336 | return (void __iomem *) (offset + addr); |
| 337 | } |
| 338 | |
| 339 | void __iomem *__arm_ioremap_caller(phys_addr_t phys_addr, size_t size, |
| 340 | unsigned int mtype, void *caller) |
| 341 | { |
| 342 | phys_addr_t last_addr; |
| 343 | unsigned long offset = phys_addr & ~PAGE_MASK; |
| 344 | unsigned long pfn = __phys_to_pfn(phys_addr); |
| 345 | |
| 346 | /* |
| 347 | * Don't allow wraparound or zero size |
| 348 | */ |
| 349 | last_addr = phys_addr + size - 1; |
| 350 | if (!size || last_addr < phys_addr) |
| 351 | return NULL; |
| 352 | |
| 353 | return __arm_ioremap_pfn_caller(pfn, offset, size, mtype, |
| 354 | caller); |
| 355 | } |
| 356 | |
| 357 | /* |
| 358 | * Remap an arbitrary physical address space into the kernel virtual |
| 359 | * address space. Needed when the kernel wants to access high addresses |
| 360 | * directly. |
| 361 | * |
| 362 | * NOTE! We need to allow non-page-aligned mappings too: we will obviously |
| 363 | * have to convert them into an offset in a page-aligned mapping, but the |
| 364 | * caller shouldn't need to know that small detail. |
| 365 | */ |
| 366 | void __iomem * |
| 367 | __arm_ioremap_pfn(unsigned long pfn, unsigned long offset, size_t size, |
| 368 | unsigned int mtype) |
| 369 | { |
| 370 | return __arm_ioremap_pfn_caller(pfn, offset, size, mtype, |
| 371 | __builtin_return_address(0)); |
| 372 | } |
| 373 | EXPORT_SYMBOL(__arm_ioremap_pfn); |
| 374 | |
| 375 | void __iomem * (*arch_ioremap_caller)(phys_addr_t, size_t, |
| 376 | unsigned int, void *) = |
| 377 | __arm_ioremap_caller; |
| 378 | |
| 379 | void __iomem *ioremap(resource_size_t res_cookie, size_t size) |
| 380 | { |
| 381 | return arch_ioremap_caller(res_cookie, size, MT_DEVICE, |
| 382 | __builtin_return_address(0)); |
| 383 | } |
| 384 | EXPORT_SYMBOL(ioremap); |
| 385 | |
| 386 | void __iomem *ioremap_cache(resource_size_t res_cookie, size_t size) |
| 387 | __alias(ioremap_cached); |
| 388 | |
| 389 | void __iomem *ioremap_cached(resource_size_t res_cookie, size_t size) |
| 390 | { |
| 391 | return arch_ioremap_caller(res_cookie, size, MT_DEVICE_CACHED, |
| 392 | __builtin_return_address(0)); |
| 393 | } |
| 394 | EXPORT_SYMBOL(ioremap_cache); |
| 395 | EXPORT_SYMBOL(ioremap_cached); |
| 396 | |
| 397 | void __iomem *ioremap_wc(resource_size_t res_cookie, size_t size) |
| 398 | { |
| 399 | return arch_ioremap_caller(res_cookie, size, MT_DEVICE_WC, |
| 400 | __builtin_return_address(0)); |
| 401 | } |
| 402 | EXPORT_SYMBOL(ioremap_wc); |
| 403 | |
| 404 | /* |
| 405 | * Remap an arbitrary physical address space into the kernel virtual |
| 406 | * address space as memory. Needed when the kernel wants to execute |
| 407 | * code in external memory. This is needed for reprogramming source |
| 408 | * clocks that would affect normal memory for example. Please see |
| 409 | * CONFIG_GENERIC_ALLOCATOR for allocating external memory. |
| 410 | */ |
| 411 | void __iomem * |
| 412 | __arm_ioremap_exec(phys_addr_t phys_addr, size_t size, bool cached) |
| 413 | { |
| 414 | unsigned int mtype; |
| 415 | |
| 416 | if (cached) |
| 417 | mtype = MT_MEMORY_RWX; |
| 418 | else |
| 419 | mtype = MT_MEMORY_RWX_NONCACHED; |
| 420 | |
| 421 | return __arm_ioremap_caller(phys_addr, size, mtype, |
| 422 | __builtin_return_address(0)); |
| 423 | } |
| 424 | |
| 425 | void *arch_memremap_wb(phys_addr_t phys_addr, size_t size) |
| 426 | { |
| 427 | return (__force void *)arch_ioremap_caller(phys_addr, size, |
| 428 | MT_MEMORY_RW, |
| 429 | __builtin_return_address(0)); |
| 430 | } |
| 431 | |
| 432 | void __iounmap(volatile void __iomem *io_addr) |
| 433 | { |
| 434 | void *addr = (void *)(PAGE_MASK & (unsigned long)io_addr); |
| 435 | struct static_vm *svm; |
| 436 | |
| 437 | /* If this is a static mapping, we must leave it alone */ |
| 438 | svm = find_static_vm_vaddr(addr); |
| 439 | if (svm) |
| 440 | return; |
| 441 | |
| 442 | #if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE) |
| 443 | { |
| 444 | struct vm_struct *vm; |
| 445 | |
| 446 | vm = find_vm_area(addr); |
| 447 | |
| 448 | /* |
| 449 | * If this is a section based mapping we need to handle it |
| 450 | * specially as the VM subsystem does not know how to handle |
| 451 | * such a beast. |
| 452 | */ |
| 453 | if (vm && (vm->flags & VM_ARM_SECTION_MAPPING)) |
| 454 | unmap_area_sections((unsigned long)vm->addr, vm->size); |
| 455 | } |
| 456 | #endif |
| 457 | |
| 458 | vunmap(addr); |
| 459 | } |
| 460 | |
| 461 | void (*arch_iounmap)(volatile void __iomem *) = __iounmap; |
| 462 | |
| 463 | void iounmap(volatile void __iomem *cookie) |
| 464 | { |
| 465 | arch_iounmap(cookie); |
| 466 | } |
| 467 | EXPORT_SYMBOL(iounmap); |
| 468 | |
| 469 | #ifdef CONFIG_PCI |
| 470 | static int pci_ioremap_mem_type = MT_DEVICE; |
| 471 | |
| 472 | void pci_ioremap_set_mem_type(int mem_type) |
| 473 | { |
| 474 | pci_ioremap_mem_type = mem_type; |
| 475 | } |
| 476 | |
| 477 | int pci_ioremap_io(unsigned int offset, phys_addr_t phys_addr) |
| 478 | { |
| 479 | BUG_ON(offset + SZ_64K - 1 > IO_SPACE_LIMIT); |
| 480 | |
| 481 | return ioremap_page_range(PCI_IO_VIRT_BASE + offset, |
| 482 | PCI_IO_VIRT_BASE + offset + SZ_64K, |
| 483 | phys_addr, |
| 484 | __pgprot(get_mem_type(pci_ioremap_mem_type)->prot_pte)); |
| 485 | } |
| 486 | EXPORT_SYMBOL_GPL(pci_ioremap_io); |
| 487 | |
| 488 | void __iomem *pci_remap_cfgspace(resource_size_t res_cookie, size_t size) |
| 489 | { |
| 490 | return arch_ioremap_caller(res_cookie, size, MT_UNCACHED, |
| 491 | __builtin_return_address(0)); |
| 492 | } |
| 493 | EXPORT_SYMBOL_GPL(pci_remap_cfgspace); |
| 494 | #endif |
| 495 | |
| 496 | /* |
| 497 | * Must be called after early_fixmap_init |
| 498 | */ |
| 499 | void __init early_ioremap_init(void) |
| 500 | { |
| 501 | early_ioremap_setup(); |
| 502 | } |
| 503 | |
| 504 | bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size, |
| 505 | unsigned long flags) |
| 506 | { |
| 507 | unsigned long pfn = PHYS_PFN(offset); |
| 508 | |
| 509 | return memblock_is_map_memory(pfn); |
| 510 | } |
| 511 | |
| 512 | DEFINE_SPINLOCK(addr_map_lock); |
| 513 | unsigned mv_cp_virtual_to_physical(unsigned va) |
| 514 | { |
| 515 | unsigned pa; |
| 516 | unsigned old_bp = 0; |
| 517 | unsigned long flags; |
| 518 | |
| 519 | spin_lock_irqsave(&addr_map_lock, flags); |
| 520 | |
| 521 | __asm__ __volatile__ ( "mrc p15, 0x0, %0, c2, c0, 0\n" : "=r" (old_bp) : : "memory"); |
| 522 | __asm__ __volatile__ ( "mcr p15, 0x0, %0, c2, c0, 0\n" : : "r" (__pa(swapper_pg_dir)) : "memory"); |
| 523 | |
| 524 | __asm__ __volatile__ ( "mcr p15, 0x0, %0, c7, c8, 0\nisb\n" : : "r" (va) : "memory"); |
| 525 | __asm__ __volatile__ ( "mrc p15, 0x0, %0, c7, c4, 0\n" : "=r" (pa) : : "memory"); |
| 526 | |
| 527 | __asm__ __volatile__ ( "mcr p15, 0x0, %0, c2, c0, 0\n" : : "r" (old_bp) : "memory"); |
| 528 | |
| 529 | spin_unlock_irqrestore(&addr_map_lock, flags); |
| 530 | |
| 531 | if ((pa & 1) == 0) |
| 532 | pa = (pa & ~0xfff) | (va & 0xfff); |
| 533 | else |
| 534 | pa = va; /* translation failed */ |
| 535 | |
| 536 | return pa; |
| 537 | } |
| 538 | EXPORT_SYMBOL_GPL(mv_cp_virtual_to_physical); |