1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/arch/arm/mm/mmu.c
5 * Copyright (C) 1995-2005 Russell King
7 #include <linux/module.h>
8 #include <linux/kernel.h>
9 #include <linux/errno.h>
10 #include <linux/init.h>
11 #include <linux/mman.h>
12 #include <linux/nodemask.h>
13 #include <linux/memblock.h>
15 #include <linux/vmalloc.h>
16 #include <linux/sizes.h>
19 #include <asm/cputype.h>
20 #include <asm/cachetype.h>
21 #include <asm/sections.h>
22 #include <asm/setup.h>
23 #include <asm/smp_plat.h>
26 #include <asm/highmem.h>
27 #include <asm/system_info.h>
28 #include <asm/traps.h>
29 #include <asm/procinfo.h>
31 #include <asm/pgalloc.h>
32 #include <asm/kasan_def.h>
34 #include <asm/mach/arch.h>
35 #include <asm/mach/map.h>
36 #include <asm/mach/pci.h>
37 #include <asm/fixmap.h>
42 extern unsigned long __atags_pointer;
45 * empty_zero_page is a special page that is used for
46 * zero-initialized data and COW.
48 struct page *empty_zero_page;
49 EXPORT_SYMBOL(empty_zero_page);
52 * The pmd table for the upper-most set of pages.
56 pmdval_t user_pmd_table = _PAGE_USER_TABLE;
58 #define CPOLICY_UNCACHED 0
59 #define CPOLICY_BUFFERED 1
60 #define CPOLICY_WRITETHROUGH 2
61 #define CPOLICY_WRITEBACK 3
62 #define CPOLICY_WRITEALLOC 4
64 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
65 static unsigned int ecc_mask __initdata = 0;
67 pgprot_t pgprot_kernel;
69 EXPORT_SYMBOL(pgprot_user);
70 EXPORT_SYMBOL(pgprot_kernel);
73 const char policy[16];
79 static struct cachepolicy cache_policies[] __initdata = {
83 .pmd = PMD_SECT_UNCACHED,
84 .pte = L_PTE_MT_UNCACHED,
88 .pmd = PMD_SECT_BUFFERED,
89 .pte = L_PTE_MT_BUFFERABLE,
91 .policy = "writethrough",
94 .pte = L_PTE_MT_WRITETHROUGH,
96 .policy = "writeback",
99 .pte = L_PTE_MT_WRITEBACK,
101 .policy = "writealloc",
103 .pmd = PMD_SECT_WBWA,
104 .pte = L_PTE_MT_WRITEALLOC,
108 #ifdef CONFIG_CPU_CP15
109 static unsigned long initial_pmd_value __initdata = 0;
112 * Initialise the cache_policy variable with the initial state specified
113 * via the "pmd" value. This is used to ensure that on ARMv6 and later,
114 * the C code sets the page tables up with the same policy as the head
115 * assembly code, which avoids an illegal state where the TLBs can get
116 * confused. See comments in early_cachepolicy() for more information.
118 void __init init_default_cache_policy(unsigned long pmd)
122 initial_pmd_value = pmd;
124 pmd &= PMD_SECT_CACHE_MASK;
126 for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
127 if (cache_policies[i].pmd == pmd) {
132 if (i == ARRAY_SIZE(cache_policies))
133 pr_err("ERROR: could not find cache policy\n");
137 * These are useful for identifying cache coherency problems by allowing
138 * the cache or the cache and writebuffer to be turned off. (Note: the
139 * write buffer should not be on and the cache off).
141 static int __init early_cachepolicy(char *p)
143 int i, selected = -1;
145 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
146 int len = strlen(cache_policies[i].policy);
148 if (memcmp(p, cache_policies[i].policy, len) == 0) {
155 pr_err("ERROR: unknown or unsupported cache policy\n");
158 * This restriction is partly to do with the way we boot; it is
159 * unpredictable to have memory mapped using two different sets of
160 * memory attributes (shared, type, and cache attribs). We can not
161 * change these attributes once the initial assembly has setup the
164 if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) {
165 pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n",
166 cache_policies[cachepolicy].policy);
170 if (selected != cachepolicy) {
171 unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
172 cachepolicy = selected;
178 early_param("cachepolicy", early_cachepolicy);
180 static int __init early_nocache(char *__unused)
182 char *p = "buffered";
183 pr_warn("nocache is deprecated; use cachepolicy=%s\n", p);
184 early_cachepolicy(p);
187 early_param("nocache", early_nocache);
189 static int __init early_nowrite(char *__unused)
191 char *p = "uncached";
192 pr_warn("nowb is deprecated; use cachepolicy=%s\n", p);
193 early_cachepolicy(p);
196 early_param("nowb", early_nowrite);
198 #ifndef CONFIG_ARM_LPAE
199 static int __init early_ecc(char *p)
201 if (memcmp(p, "on", 2) == 0)
202 ecc_mask = PMD_PROTECTION;
203 else if (memcmp(p, "off", 3) == 0)
207 early_param("ecc", early_ecc);
210 #else /* ifdef CONFIG_CPU_CP15 */
212 static int __init early_cachepolicy(char *p)
214 pr_warn("cachepolicy kernel parameter not supported without cp15\n");
217 early_param("cachepolicy", early_cachepolicy);
219 static int __init noalign_setup(char *__unused)
221 pr_warn("noalign kernel parameter not supported without cp15\n");
224 __setup("noalign", noalign_setup);
226 #endif /* ifdef CONFIG_CPU_CP15 / else */
228 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
229 #define PROT_PTE_S2_DEVICE PROT_PTE_DEVICE
230 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE
232 static struct mem_type mem_types[] __ro_after_init = {
233 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */
234 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
236 .prot_l1 = PMD_TYPE_TABLE,
237 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S,
240 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
241 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
242 .prot_l1 = PMD_TYPE_TABLE,
243 .prot_sect = PROT_SECT_DEVICE,
246 [MT_DEVICE_CACHED] = { /* ioremap_cache */
247 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
248 .prot_l1 = PMD_TYPE_TABLE,
249 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB,
252 [MT_DEVICE_WC] = { /* ioremap_wc */
253 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
254 .prot_l1 = PMD_TYPE_TABLE,
255 .prot_sect = PROT_SECT_DEVICE,
259 .prot_pte = PROT_PTE_DEVICE,
260 .prot_l1 = PMD_TYPE_TABLE,
261 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
265 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
266 .domain = DOMAIN_KERNEL,
268 #ifndef CONFIG_ARM_LPAE
270 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
271 .domain = DOMAIN_KERNEL,
275 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
277 .prot_l1 = PMD_TYPE_TABLE,
278 .domain = DOMAIN_VECTORS,
280 [MT_HIGH_VECTORS] = {
281 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
282 L_PTE_USER | L_PTE_RDONLY,
283 .prot_l1 = PMD_TYPE_TABLE,
284 .domain = DOMAIN_VECTORS,
287 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
288 .prot_l1 = PMD_TYPE_TABLE,
289 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
290 .domain = DOMAIN_KERNEL,
293 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
295 .prot_l1 = PMD_TYPE_TABLE,
296 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
297 .domain = DOMAIN_KERNEL,
300 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
301 L_PTE_XN | L_PTE_RDONLY,
302 .prot_l1 = PMD_TYPE_TABLE,
303 #ifdef CONFIG_ARM_LPAE
304 .prot_sect = PMD_TYPE_SECT | L_PMD_SECT_RDONLY | PMD_SECT_AP2,
306 .prot_sect = PMD_TYPE_SECT,
308 .domain = DOMAIN_KERNEL,
311 .prot_sect = PMD_TYPE_SECT,
312 .domain = DOMAIN_KERNEL,
314 [MT_MEMORY_RWX_NONCACHED] = {
315 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
317 .prot_l1 = PMD_TYPE_TABLE,
318 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
319 .domain = DOMAIN_KERNEL,
321 [MT_MEMORY_RW_DTCM] = {
322 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
324 .prot_l1 = PMD_TYPE_TABLE,
325 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
326 .domain = DOMAIN_KERNEL,
328 [MT_MEMORY_RWX_ITCM] = {
329 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
330 .prot_l1 = PMD_TYPE_TABLE,
331 .domain = DOMAIN_KERNEL,
333 [MT_MEMORY_RW_SO] = {
334 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
335 L_PTE_MT_UNCACHED | L_PTE_XN,
336 .prot_l1 = PMD_TYPE_TABLE,
337 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
338 PMD_SECT_UNCACHED | PMD_SECT_XN,
339 .domain = DOMAIN_KERNEL,
341 [MT_MEMORY_DMA_READY] = {
342 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
344 .prot_l1 = PMD_TYPE_TABLE,
345 .domain = DOMAIN_KERNEL,
349 const struct mem_type *get_mem_type(unsigned int type)
351 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
353 EXPORT_SYMBOL(get_mem_type);
355 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr);
357 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS]
358 __aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata;
360 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr)
362 return &bm_pte[pte_index(addr)];
365 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr)
367 return pte_offset_kernel(dir, addr);
370 static inline pmd_t * __init fixmap_pmd(unsigned long addr)
372 return pmd_off_k(addr);
375 void __init early_fixmap_init(void)
380 * The early fixmap range spans multiple pmds, for which
381 * we are not prepared:
383 BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT)
384 != FIXADDR_TOP >> PMD_SHIFT);
386 pmd = fixmap_pmd(FIXADDR_TOP);
387 pmd_populate_kernel(&init_mm, pmd, bm_pte);
389 pte_offset_fixmap = pte_offset_early_fixmap;
393 * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
394 * As a result, this can only be called with preemption disabled, as under
397 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
399 unsigned long vaddr = __fix_to_virt(idx);
400 pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr);
402 /* Make sure fixmap region does not exceed available allocation. */
403 BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) < FIXADDR_START);
404 BUG_ON(idx >= __end_of_fixed_addresses);
406 /* We support only device mappings before pgprot_kernel is set. */
407 if (WARN_ON(pgprot_val(prot) != pgprot_val(FIXMAP_PAGE_IO) &&
408 pgprot_val(prot) && pgprot_val(pgprot_kernel) == 0))
411 if (pgprot_val(prot))
412 set_pte_at(NULL, vaddr, pte,
413 pfn_pte(phys >> PAGE_SHIFT, prot));
415 pte_clear(NULL, vaddr, pte);
416 local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
419 static pgprot_t protection_map[16] __ro_after_init = {
420 [VM_NONE] = __PAGE_NONE,
421 [VM_READ] = __PAGE_READONLY,
422 [VM_WRITE] = __PAGE_COPY,
423 [VM_WRITE | VM_READ] = __PAGE_COPY,
424 [VM_EXEC] = __PAGE_READONLY_EXEC,
425 [VM_EXEC | VM_READ] = __PAGE_READONLY_EXEC,
426 [VM_EXEC | VM_WRITE] = __PAGE_COPY_EXEC,
427 [VM_EXEC | VM_WRITE | VM_READ] = __PAGE_COPY_EXEC,
428 [VM_SHARED] = __PAGE_NONE,
429 [VM_SHARED | VM_READ] = __PAGE_READONLY,
430 [VM_SHARED | VM_WRITE] = __PAGE_SHARED,
431 [VM_SHARED | VM_WRITE | VM_READ] = __PAGE_SHARED,
432 [VM_SHARED | VM_EXEC] = __PAGE_READONLY_EXEC,
433 [VM_SHARED | VM_EXEC | VM_READ] = __PAGE_READONLY_EXEC,
434 [VM_SHARED | VM_EXEC | VM_WRITE] = __PAGE_SHARED_EXEC,
435 [VM_SHARED | VM_EXEC | VM_WRITE | VM_READ] = __PAGE_SHARED_EXEC
437 DECLARE_VM_GET_PAGE_PROT
440 * Adjust the PMD section entries according to the CPU in use.
442 static void __init build_mem_type_table(void)
444 struct cachepolicy *cp;
445 unsigned int cr = get_cr();
446 pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
447 int cpu_arch = cpu_architecture();
450 if (cpu_arch < CPU_ARCH_ARMv6) {
451 #if defined(CONFIG_CPU_DCACHE_DISABLE)
452 if (cachepolicy > CPOLICY_BUFFERED)
453 cachepolicy = CPOLICY_BUFFERED;
454 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
455 if (cachepolicy > CPOLICY_WRITETHROUGH)
456 cachepolicy = CPOLICY_WRITETHROUGH;
459 if (cpu_arch < CPU_ARCH_ARMv5) {
460 if (cachepolicy >= CPOLICY_WRITEALLOC)
461 cachepolicy = CPOLICY_WRITEBACK;
466 if (cachepolicy != CPOLICY_WRITEALLOC) {
467 pr_warn("Forcing write-allocate cache policy for SMP\n");
468 cachepolicy = CPOLICY_WRITEALLOC;
470 if (!(initial_pmd_value & PMD_SECT_S)) {
471 pr_warn("Forcing shared mappings for SMP\n");
472 initial_pmd_value |= PMD_SECT_S;
477 * Strip out features not present on earlier architectures.
478 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those
479 * without extended page tables don't have the 'Shared' bit.
481 if (cpu_arch < CPU_ARCH_ARMv5)
482 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
483 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
484 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
485 for (i = 0; i < ARRAY_SIZE(mem_types); i++)
486 mem_types[i].prot_sect &= ~PMD_SECT_S;
489 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
490 * "update-able on write" bit on ARM610). However, Xscale and
491 * Xscale3 require this bit to be cleared.
493 if (cpu_is_xscale_family()) {
494 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
495 mem_types[i].prot_sect &= ~PMD_BIT4;
496 mem_types[i].prot_l1 &= ~PMD_BIT4;
498 } else if (cpu_arch < CPU_ARCH_ARMv6) {
499 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
500 if (mem_types[i].prot_l1)
501 mem_types[i].prot_l1 |= PMD_BIT4;
502 if (mem_types[i].prot_sect)
503 mem_types[i].prot_sect |= PMD_BIT4;
508 * Mark the device areas according to the CPU/architecture.
510 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
511 if (!cpu_is_xsc3()) {
513 * Mark device regions on ARMv6+ as execute-never
514 * to prevent speculative instruction fetches.
516 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
517 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
518 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
519 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
521 /* Also setup NX memory mapping */
522 mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
523 mem_types[MT_MEMORY_RO].prot_sect |= PMD_SECT_XN;
525 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
527 * For ARMv7 with TEX remapping,
528 * - shared device is SXCB=1100
529 * - nonshared device is SXCB=0100
530 * - write combine device mem is SXCB=0001
531 * (Uncached Normal memory)
533 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
534 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
535 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
536 } else if (cpu_is_xsc3()) {
539 * - shared device is TEXCB=00101
540 * - nonshared device is TEXCB=01000
541 * - write combine device mem is TEXCB=00100
542 * (Inner/Outer Uncacheable in xsc3 parlance)
544 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
545 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
546 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
549 * For ARMv6 and ARMv7 without TEX remapping,
550 * - shared device is TEXCB=00001
551 * - nonshared device is TEXCB=01000
552 * - write combine device mem is TEXCB=00100
553 * (Uncached Normal in ARMv6 parlance).
555 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
556 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
557 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
561 * On others, write combining is "Uncached/Buffered"
563 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
567 * Now deal with the memory-type mappings
569 cp = &cache_policies[cachepolicy];
570 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
572 #ifndef CONFIG_ARM_LPAE
574 * We don't use domains on ARMv6 (since this causes problems with
575 * v6/v7 kernels), so we must use a separate memory type for user
576 * r/o, kernel r/w to map the vectors page.
578 if (cpu_arch == CPU_ARCH_ARMv6)
579 vecs_pgprot |= L_PTE_MT_VECTORS;
582 * Check is it with support for the PXN bit
583 * in the Short-descriptor translation table format descriptors.
585 if (cpu_arch == CPU_ARCH_ARMv7 &&
586 (read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
587 user_pmd_table |= PMD_PXNTABLE;
592 * ARMv6 and above have extended page tables.
594 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
595 #ifndef CONFIG_ARM_LPAE
597 * Mark cache clean areas and XIP ROM read only
598 * from SVC mode and no access from userspace.
600 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
601 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
602 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
603 mem_types[MT_MEMORY_RO].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
607 * If the initial page tables were created with the S bit
608 * set, then we need to do the same here for the same
609 * reasons given in early_cachepolicy().
611 if (initial_pmd_value & PMD_SECT_S) {
612 user_pgprot |= L_PTE_SHARED;
613 kern_pgprot |= L_PTE_SHARED;
614 vecs_pgprot |= L_PTE_SHARED;
615 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
616 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
617 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
618 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
619 mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
620 mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
621 mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
622 mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
623 mem_types[MT_MEMORY_RO].prot_sect |= PMD_SECT_S;
624 mem_types[MT_MEMORY_RO].prot_pte |= L_PTE_SHARED;
625 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
626 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
627 mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
632 * Non-cacheable Normal - intended for memory areas that must
633 * not cause dirty cache line writebacks when used
635 if (cpu_arch >= CPU_ARCH_ARMv6) {
636 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
637 /* Non-cacheable Normal is XCB = 001 */
638 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
641 /* For both ARMv6 and non-TEX-remapping ARMv7 */
642 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
646 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
649 #ifdef CONFIG_ARM_LPAE
651 * Do not generate access flag faults for the kernel mappings.
653 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
654 mem_types[i].prot_pte |= PTE_EXT_AF;
655 if (mem_types[i].prot_sect)
656 mem_types[i].prot_sect |= PMD_SECT_AF;
658 kern_pgprot |= PTE_EXT_AF;
659 vecs_pgprot |= PTE_EXT_AF;
662 * Set PXN for user mappings
664 user_pgprot |= PTE_EXT_PXN;
667 for (i = 0; i < 16; i++) {
668 pteval_t v = pgprot_val(protection_map[i]);
669 protection_map[i] = __pgprot(v | user_pgprot);
672 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
673 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
675 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
676 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
677 L_PTE_DIRTY | kern_pgprot);
679 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
680 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
681 mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
682 mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
683 mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
684 mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
685 mem_types[MT_MEMORY_RO].prot_sect |= ecc_mask | cp->pmd;
686 mem_types[MT_MEMORY_RO].prot_pte |= kern_pgprot;
687 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
688 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
689 mem_types[MT_ROM].prot_sect |= cp->pmd;
693 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
697 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
700 pr_info("Memory policy: %sData cache %s\n",
701 ecc_mask ? "ECC enabled, " : "", cp->policy);
703 for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
704 struct mem_type *t = &mem_types[i];
706 t->prot_l1 |= PMD_DOMAIN(t->domain);
708 t->prot_sect |= PMD_DOMAIN(t->domain);
712 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
713 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
714 unsigned long size, pgprot_t vma_prot)
717 return pgprot_noncached(vma_prot);
718 else if (file->f_flags & O_SYNC)
719 return pgprot_writecombine(vma_prot);
722 EXPORT_SYMBOL(phys_mem_access_prot);
725 #define vectors_base() (vectors_high() ? 0xffff0000 : 0)
727 static void __init *early_alloc(unsigned long sz)
729 void *ptr = memblock_alloc(sz, sz);
732 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
738 static void *__init late_alloc(unsigned long sz)
740 void *ptdesc = pagetable_alloc(GFP_PGTABLE_KERNEL & ~__GFP_HIGHMEM,
743 if (!ptdesc || !pagetable_pte_ctor(ptdesc))
745 return ptdesc_to_virt(ptdesc);
748 static pte_t * __init arm_pte_alloc(pmd_t *pmd, unsigned long addr,
750 void *(*alloc)(unsigned long sz))
752 if (pmd_none(*pmd)) {
753 pte_t *pte = alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
754 __pmd_populate(pmd, __pa(pte), prot);
756 BUG_ON(pmd_bad(*pmd));
757 return pte_offset_kernel(pmd, addr);
760 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr,
763 return arm_pte_alloc(pmd, addr, prot, early_alloc);
766 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
767 unsigned long end, unsigned long pfn,
768 const struct mem_type *type,
769 void *(*alloc)(unsigned long sz),
772 pte_t *pte = arm_pte_alloc(pmd, addr, type->prot_l1, alloc);
774 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)),
775 ng ? PTE_EXT_NG : 0);
777 } while (pte++, addr += PAGE_SIZE, addr != end);
780 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
781 unsigned long end, phys_addr_t phys,
782 const struct mem_type *type, bool ng)
786 #ifndef CONFIG_ARM_LPAE
788 * In classic MMU format, puds and pmds are folded in to
789 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
790 * group of L1 entries making up one logical pointer to
791 * an L2 table (2MB), where as PMDs refer to the individual
792 * L1 entries (1MB). Hence increment to get the correct
793 * offset for odd 1MB sections.
794 * (See arch/arm/include/asm/pgtable-2level.h)
796 if (addr & SECTION_SIZE)
800 *pmd = __pmd(phys | type->prot_sect | (ng ? PMD_SECT_nG : 0));
801 phys += SECTION_SIZE;
802 } while (pmd++, addr += SECTION_SIZE, addr != end);
807 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
808 unsigned long end, phys_addr_t phys,
809 const struct mem_type *type,
810 void *(*alloc)(unsigned long sz), bool ng)
812 pmd_t *pmd = pmd_offset(pud, addr);
817 * With LPAE, we must loop over to map
818 * all the pmds for the given range.
820 next = pmd_addr_end(addr, end);
823 * Try a section mapping - addr, next and phys must all be
824 * aligned to a section boundary.
826 if (type->prot_sect &&
827 ((addr | next | phys) & ~SECTION_MASK) == 0) {
828 __map_init_section(pmd, addr, next, phys, type, ng);
830 alloc_init_pte(pmd, addr, next,
831 __phys_to_pfn(phys), type, alloc, ng);
836 } while (pmd++, addr = next, addr != end);
839 static void __init alloc_init_pud(p4d_t *p4d, unsigned long addr,
840 unsigned long end, phys_addr_t phys,
841 const struct mem_type *type,
842 void *(*alloc)(unsigned long sz), bool ng)
844 pud_t *pud = pud_offset(p4d, addr);
848 next = pud_addr_end(addr, end);
849 alloc_init_pmd(pud, addr, next, phys, type, alloc, ng);
851 } while (pud++, addr = next, addr != end);
854 static void __init alloc_init_p4d(pgd_t *pgd, unsigned long addr,
855 unsigned long end, phys_addr_t phys,
856 const struct mem_type *type,
857 void *(*alloc)(unsigned long sz), bool ng)
859 p4d_t *p4d = p4d_offset(pgd, addr);
863 next = p4d_addr_end(addr, end);
864 alloc_init_pud(p4d, addr, next, phys, type, alloc, ng);
866 } while (p4d++, addr = next, addr != end);
869 #ifndef CONFIG_ARM_LPAE
870 static void __init create_36bit_mapping(struct mm_struct *mm,
872 const struct mem_type *type,
875 unsigned long addr, length, end;
880 phys = __pfn_to_phys(md->pfn);
881 length = PAGE_ALIGN(md->length);
883 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
884 pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
885 (long long)__pfn_to_phys((u64)md->pfn), addr);
889 /* N.B. ARMv6 supersections are only defined to work with domain 0.
890 * Since domain assignments can in fact be arbitrary, the
891 * 'domain == 0' check below is required to insure that ARMv6
892 * supersections are only allocated for domain 0 regardless
893 * of the actual domain assignments in use.
896 pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
897 (long long)__pfn_to_phys((u64)md->pfn), addr);
901 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
902 pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
903 (long long)__pfn_to_phys((u64)md->pfn), addr);
908 * Shift bits [35:32] of address into bits [23:20] of PMD
911 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
913 pgd = pgd_offset(mm, addr);
916 p4d_t *p4d = p4d_offset(pgd, addr);
917 pud_t *pud = pud_offset(p4d, addr);
918 pmd_t *pmd = pmd_offset(pud, addr);
921 for (i = 0; i < 16; i++)
922 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER |
923 (ng ? PMD_SECT_nG : 0));
925 addr += SUPERSECTION_SIZE;
926 phys += SUPERSECTION_SIZE;
927 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
928 } while (addr != end);
930 #endif /* !CONFIG_ARM_LPAE */
932 static void __init __create_mapping(struct mm_struct *mm, struct map_desc *md,
933 void *(*alloc)(unsigned long sz),
936 unsigned long addr, length, end;
938 const struct mem_type *type;
941 type = &mem_types[md->type];
943 #ifndef CONFIG_ARM_LPAE
945 * Catch 36-bit addresses
947 if (md->pfn >= 0x100000) {
948 create_36bit_mapping(mm, md, type, ng);
953 addr = md->virtual & PAGE_MASK;
954 phys = __pfn_to_phys(md->pfn);
955 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
957 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
958 pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
959 (long long)__pfn_to_phys(md->pfn), addr);
963 pgd = pgd_offset(mm, addr);
966 unsigned long next = pgd_addr_end(addr, end);
968 alloc_init_p4d(pgd, addr, next, phys, type, alloc, ng);
972 } while (pgd++, addr != end);
976 * Create the page directory entries and any necessary
977 * page tables for the mapping specified by `md'. We
978 * are able to cope here with varying sizes and address
979 * offsets, and we take full advantage of sections and
982 static void __init create_mapping(struct map_desc *md)
984 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
985 pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
986 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
990 if (md->type == MT_DEVICE &&
991 md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START &&
992 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
993 pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
994 (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
997 __create_mapping(&init_mm, md, early_alloc, false);
1000 void __init create_mapping_late(struct mm_struct *mm, struct map_desc *md,
1003 #ifdef CONFIG_ARM_LPAE
1007 p4d = p4d_alloc(mm, pgd_offset(mm, md->virtual), md->virtual);
1010 pud = pud_alloc(mm, p4d, md->virtual);
1013 pmd_alloc(mm, pud, 0);
1015 __create_mapping(mm, md, late_alloc, ng);
1019 * Create the architecture specific mappings
1021 void __init iotable_init(struct map_desc *io_desc, int nr)
1023 struct map_desc *md;
1024 struct vm_struct *vm;
1025 struct static_vm *svm;
1030 svm = memblock_alloc(sizeof(*svm) * nr, __alignof__(*svm));
1032 panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1033 __func__, sizeof(*svm) * nr, __alignof__(*svm));
1035 for (md = io_desc; nr; md++, nr--) {
1039 vm->addr = (void *)(md->virtual & PAGE_MASK);
1040 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
1041 vm->phys_addr = __pfn_to_phys(md->pfn);
1042 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
1043 vm->flags |= VM_ARM_MTYPE(md->type);
1044 vm->caller = iotable_init;
1045 add_static_vm_early(svm++);
1049 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
1052 struct vm_struct *vm;
1053 struct static_vm *svm;
1055 svm = memblock_alloc(sizeof(*svm), __alignof__(*svm));
1057 panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1058 __func__, sizeof(*svm), __alignof__(*svm));
1061 vm->addr = (void *)addr;
1063 vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
1064 vm->caller = caller;
1065 add_static_vm_early(svm);
1068 #ifndef CONFIG_ARM_LPAE
1071 * The Linux PMD is made of two consecutive section entries covering 2MB
1072 * (see definition in include/asm/pgtable-2level.h). However a call to
1073 * create_mapping() may optimize static mappings by using individual
1074 * 1MB section mappings. This leaves the actual PMD potentially half
1075 * initialized if the top or bottom section entry isn't used, leaving it
1076 * open to problems if a subsequent ioremap() or vmalloc() tries to use
1077 * the virtual space left free by that unused section entry.
1079 * Let's avoid the issue by inserting dummy vm entries covering the unused
1080 * PMD halves once the static mappings are in place.
1083 static void __init pmd_empty_section_gap(unsigned long addr)
1085 vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1088 static void __init fill_pmd_gaps(void)
1090 struct static_vm *svm;
1091 struct vm_struct *vm;
1092 unsigned long addr, next = 0;
1095 list_for_each_entry(svm, &static_vmlist, list) {
1097 addr = (unsigned long)vm->addr;
1102 * Check if this vm starts on an odd section boundary.
1103 * If so and the first section entry for this PMD is free
1104 * then we block the corresponding virtual address.
1106 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1107 pmd = pmd_off_k(addr);
1109 pmd_empty_section_gap(addr & PMD_MASK);
1113 * Then check if this vm ends on an odd section boundary.
1114 * If so and the second section entry for this PMD is empty
1115 * then we block the corresponding virtual address.
1118 if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1119 pmd = pmd_off_k(addr) + 1;
1121 pmd_empty_section_gap(addr);
1124 /* no need to look at any vm entry until we hit the next PMD */
1125 next = (addr + PMD_SIZE - 1) & PMD_MASK;
1130 #define fill_pmd_gaps() do { } while (0)
1133 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1134 static void __init pci_reserve_io(void)
1136 struct static_vm *svm;
1138 svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1142 vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1145 #define pci_reserve_io() do { } while (0)
1148 #ifdef CONFIG_DEBUG_LL
1149 void __init debug_ll_io_init(void)
1151 struct map_desc map;
1153 debug_ll_addr(&map.pfn, &map.virtual);
1154 if (!map.pfn || !map.virtual)
1156 map.pfn = __phys_to_pfn(map.pfn);
1157 map.virtual &= PAGE_MASK;
1158 map.length = PAGE_SIZE;
1159 map.type = MT_DEVICE;
1160 iotable_init(&map, 1);
1164 static unsigned long __initdata vmalloc_size = 240 * SZ_1M;
1167 * vmalloc=size forces the vmalloc area to be exactly 'size'
1168 * bytes. This can be used to increase (or decrease) the vmalloc
1169 * area - the default is 240MiB.
1171 static int __init early_vmalloc(char *arg)
1173 unsigned long vmalloc_reserve = memparse(arg, NULL);
1174 unsigned long vmalloc_max;
1176 if (vmalloc_reserve < SZ_16M) {
1177 vmalloc_reserve = SZ_16M;
1178 pr_warn("vmalloc area is too small, limiting to %luMiB\n",
1179 vmalloc_reserve >> 20);
1182 vmalloc_max = VMALLOC_END - (PAGE_OFFSET + SZ_32M + VMALLOC_OFFSET);
1183 if (vmalloc_reserve > vmalloc_max) {
1184 vmalloc_reserve = vmalloc_max;
1185 pr_warn("vmalloc area is too big, limiting to %luMiB\n",
1186 vmalloc_reserve >> 20);
1189 vmalloc_size = vmalloc_reserve;
1192 early_param("vmalloc", early_vmalloc);
1194 phys_addr_t arm_lowmem_limit __initdata = 0;
1196 void __init adjust_lowmem_bounds(void)
1198 phys_addr_t block_start, block_end, memblock_limit = 0;
1199 u64 vmalloc_limit, i;
1200 phys_addr_t lowmem_limit = 0;
1203 * Let's use our own (unoptimized) equivalent of __pa() that is
1204 * not affected by wrap-arounds when sizeof(phys_addr_t) == 4.
1205 * The result is used as the upper bound on physical memory address
1206 * and may itself be outside the valid range for which phys_addr_t
1207 * and therefore __pa() is defined.
1209 vmalloc_limit = (u64)VMALLOC_END - vmalloc_size - VMALLOC_OFFSET -
1210 PAGE_OFFSET + PHYS_OFFSET;
1213 * The first usable region must be PMD aligned. Mark its start
1214 * as MEMBLOCK_NOMAP if it isn't
1216 for_each_mem_range(i, &block_start, &block_end) {
1217 if (!IS_ALIGNED(block_start, PMD_SIZE)) {
1220 len = round_up(block_start, PMD_SIZE) - block_start;
1221 memblock_mark_nomap(block_start, len);
1226 for_each_mem_range(i, &block_start, &block_end) {
1227 if (block_start < vmalloc_limit) {
1228 if (block_end > lowmem_limit)
1230 * Compare as u64 to ensure vmalloc_limit does
1231 * not get truncated. block_end should always
1232 * fit in phys_addr_t so there should be no
1233 * issue with assignment.
1235 lowmem_limit = min_t(u64,
1240 * Find the first non-pmd-aligned page, and point
1241 * memblock_limit at it. This relies on rounding the
1242 * limit down to be pmd-aligned, which happens at the
1243 * end of this function.
1245 * With this algorithm, the start or end of almost any
1246 * bank can be non-pmd-aligned. The only exception is
1247 * that the start of the bank 0 must be section-
1248 * aligned, since otherwise memory would need to be
1249 * allocated when mapping the start of bank 0, which
1250 * occurs before any free memory is mapped.
1252 if (!memblock_limit) {
1253 if (!IS_ALIGNED(block_start, PMD_SIZE))
1254 memblock_limit = block_start;
1255 else if (!IS_ALIGNED(block_end, PMD_SIZE))
1256 memblock_limit = lowmem_limit;
1262 arm_lowmem_limit = lowmem_limit;
1264 high_memory = __va(arm_lowmem_limit - 1) + 1;
1266 if (!memblock_limit)
1267 memblock_limit = arm_lowmem_limit;
1270 * Round the memblock limit down to a pmd size. This
1271 * helps to ensure that we will allocate memory from the
1272 * last full pmd, which should be mapped.
1274 memblock_limit = round_down(memblock_limit, PMD_SIZE);
1276 if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1277 if (memblock_end_of_DRAM() > arm_lowmem_limit) {
1278 phys_addr_t end = memblock_end_of_DRAM();
1280 pr_notice("Ignoring RAM at %pa-%pa\n",
1281 &memblock_limit, &end);
1282 pr_notice("Consider using a HIGHMEM enabled kernel.\n");
1284 memblock_remove(memblock_limit, end - memblock_limit);
1288 memblock_set_current_limit(memblock_limit);
1291 static __init void prepare_page_table(void)
1297 * Clear out all the mappings below the kernel image.
1301 * KASan's shadow memory inserts itself between the TASK_SIZE
1302 * and MODULES_VADDR. Do not clear the KASan shadow memory mappings.
1304 for (addr = 0; addr < KASAN_SHADOW_START; addr += PMD_SIZE)
1305 pmd_clear(pmd_off_k(addr));
1307 * Skip over the KASan shadow area. KASAN_SHADOW_END is sometimes
1308 * equal to MODULES_VADDR and then we exit the pmd clearing. If we
1309 * are using a thumb-compiled kernel, there there will be 8MB more
1310 * to clear as KASan always offset to 16 MB below MODULES_VADDR.
1312 for (addr = KASAN_SHADOW_END; addr < MODULES_VADDR; addr += PMD_SIZE)
1313 pmd_clear(pmd_off_k(addr));
1315 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1316 pmd_clear(pmd_off_k(addr));
1319 #ifdef CONFIG_XIP_KERNEL
1320 /* The XIP kernel is mapped in the module area -- skip over it */
1321 addr = ((unsigned long)_exiprom + PMD_SIZE - 1) & PMD_MASK;
1323 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1324 pmd_clear(pmd_off_k(addr));
1327 * Find the end of the first block of lowmem.
1329 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1330 if (end >= arm_lowmem_limit)
1331 end = arm_lowmem_limit;
1334 * Clear out all the kernel space mappings, except for the first
1335 * memory bank, up to the vmalloc region.
1337 for (addr = __phys_to_virt(end);
1338 addr < VMALLOC_START; addr += PMD_SIZE)
1339 pmd_clear(pmd_off_k(addr));
1342 #ifdef CONFIG_ARM_LPAE
1343 /* the first page is reserved for pgd */
1344 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \
1345 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1347 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
1351 * Reserve the special regions of memory
1353 void __init arm_mm_memblock_reserve(void)
1356 * Reserve the page tables. These are already in use,
1357 * and can only be in node 0.
1359 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1361 #ifdef CONFIG_SA1111
1363 * Because of the SA1111 DMA bug, we want to preserve our
1364 * precious DMA-able memory...
1366 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1371 * Set up the device mappings. Since we clear out the page tables for all
1372 * mappings above VMALLOC_START, except early fixmap, we might remove debug
1373 * device mappings. This means earlycon can be used to debug this function
1374 * Any other function or debugging method which may touch any device _will_
1377 static void __init devicemaps_init(const struct machine_desc *mdesc)
1379 struct map_desc map;
1384 * Allocate the vector page early.
1386 vectors = early_alloc(PAGE_SIZE * 2);
1388 early_trap_init(vectors);
1391 * Clear page table except top pmd used by early fixmaps
1393 for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE)
1394 pmd_clear(pmd_off_k(addr));
1396 if (__atags_pointer) {
1397 /* create a read-only mapping of the device tree */
1398 map.pfn = __phys_to_pfn(__atags_pointer & SECTION_MASK);
1399 map.virtual = FDT_FIXED_BASE;
1400 map.length = FDT_FIXED_SIZE;
1401 map.type = MT_MEMORY_RO;
1402 create_mapping(&map);
1406 * Map the kernel if it is XIP.
1407 * It is always first in the modulearea.
1409 #ifdef CONFIG_XIP_KERNEL
1410 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1411 map.virtual = MODULES_VADDR;
1412 map.length = ((unsigned long)_exiprom - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1414 create_mapping(&map);
1418 * Map the cache flushing regions.
1421 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1422 map.virtual = FLUSH_BASE;
1424 map.type = MT_CACHECLEAN;
1425 create_mapping(&map);
1427 #ifdef FLUSH_BASE_MINICACHE
1428 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1429 map.virtual = FLUSH_BASE_MINICACHE;
1431 map.type = MT_MINICLEAN;
1432 create_mapping(&map);
1436 * Create a mapping for the machine vectors at the high-vectors
1437 * location (0xffff0000). If we aren't using high-vectors, also
1438 * create a mapping at the low-vectors virtual address.
1440 map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1441 map.virtual = 0xffff0000;
1442 map.length = PAGE_SIZE;
1443 #ifdef CONFIG_KUSER_HELPERS
1444 map.type = MT_HIGH_VECTORS;
1446 map.type = MT_LOW_VECTORS;
1448 create_mapping(&map);
1450 if (!vectors_high()) {
1452 map.length = PAGE_SIZE * 2;
1453 map.type = MT_LOW_VECTORS;
1454 create_mapping(&map);
1457 /* Now create a kernel read-only mapping */
1459 map.virtual = 0xffff0000 + PAGE_SIZE;
1460 map.length = PAGE_SIZE;
1461 map.type = MT_LOW_VECTORS;
1462 create_mapping(&map);
1465 * Ask the machine support to map in the statically mapped devices.
1473 /* Reserve fixed i/o space in VMALLOC region */
1477 * Finally flush the caches and tlb to ensure that we're in a
1478 * consistent state wrt the writebuffer. This also ensures that
1479 * any write-allocated cache lines in the vector page are written
1480 * back. After this point, we can start to touch devices again.
1482 local_flush_tlb_all();
1485 /* Enable asynchronous aborts */
1489 static void __init kmap_init(void)
1491 #ifdef CONFIG_HIGHMEM
1492 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1493 PKMAP_BASE, _PAGE_KERNEL_TABLE);
1496 early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1497 _PAGE_KERNEL_TABLE);
1500 static void __init map_lowmem(void)
1502 phys_addr_t start, end;
1505 /* Map all the lowmem memory banks. */
1506 for_each_mem_range(i, &start, &end) {
1507 struct map_desc map;
1509 pr_debug("map lowmem start: 0x%08llx, end: 0x%08llx\n",
1510 (long long)start, (long long)end);
1511 if (end > arm_lowmem_limit)
1512 end = arm_lowmem_limit;
1517 * If our kernel image is in the VMALLOC area we need to remove
1518 * the kernel physical memory from lowmem since the kernel will
1519 * be mapped separately.
1521 * The kernel will typically be at the very start of lowmem,
1522 * but any placement relative to memory ranges is possible.
1524 * If the memblock contains the kernel, we have to chisel out
1525 * the kernel memory from it and map each part separately. We
1526 * get 6 different theoretical cases:
1528 * +--------+ +--------+
1529 * +-- start --+ +--------+ | Kernel | | Kernel |
1530 * | | | Kernel | | case 2 | | case 5 |
1531 * | | | case 1 | +--------+ | | +--------+
1532 * | Memory | +--------+ | | | Kernel |
1533 * | range | +--------+ | | | case 6 |
1534 * | | | Kernel | +--------+ | | +--------+
1535 * | | | case 3 | | Kernel | | |
1536 * +-- end ----+ +--------+ | case 4 | | |
1537 * +--------+ +--------+
1540 /* Case 5: kernel covers range, don't map anything, should be rare */
1541 if ((start > kernel_sec_start) && (end < kernel_sec_end))
1544 /* Cases where the kernel is starting inside the range */
1545 if ((kernel_sec_start >= start) && (kernel_sec_start <= end)) {
1546 /* Case 6: kernel is embedded in the range, we need two mappings */
1547 if ((start < kernel_sec_start) && (end > kernel_sec_end)) {
1548 /* Map memory below the kernel */
1549 map.pfn = __phys_to_pfn(start);
1550 map.virtual = __phys_to_virt(start);
1551 map.length = kernel_sec_start - start;
1552 map.type = MT_MEMORY_RW;
1553 create_mapping(&map);
1554 /* Map memory above the kernel */
1555 map.pfn = __phys_to_pfn(kernel_sec_end);
1556 map.virtual = __phys_to_virt(kernel_sec_end);
1557 map.length = end - kernel_sec_end;
1558 map.type = MT_MEMORY_RW;
1559 create_mapping(&map);
1562 /* Case 1: kernel and range start at the same address, should be common */
1563 if (kernel_sec_start == start)
1564 start = kernel_sec_end;
1565 /* Case 3: kernel and range end at the same address, should be rare */
1566 if (kernel_sec_end == end)
1567 end = kernel_sec_start;
1568 } else if ((kernel_sec_start < start) && (kernel_sec_end > start) && (kernel_sec_end < end)) {
1569 /* Case 2: kernel ends inside range, starts below it */
1570 start = kernel_sec_end;
1571 } else if ((kernel_sec_start > start) && (kernel_sec_start < end) && (kernel_sec_end > end)) {
1572 /* Case 4: kernel starts inside range, ends above it */
1573 end = kernel_sec_start;
1575 map.pfn = __phys_to_pfn(start);
1576 map.virtual = __phys_to_virt(start);
1577 map.length = end - start;
1578 map.type = MT_MEMORY_RW;
1579 create_mapping(&map);
1583 static void __init map_kernel(void)
1586 * We use the well known kernel section start and end and split the area in the
1590 * +----------------+ kernel_x_start
1593 * +----------------+ kernel_x_end / kernel_nx_start
1594 * | Non-executable |
1596 * +----------------+ kernel_nx_end
1600 * Notice that we are dealing with section sized mappings here so all of this
1601 * will be bumped to the closest section boundary. This means that some of the
1602 * non-executable part of the kernel memory is actually mapped as executable.
1603 * This will only persist until we turn on proper memory management later on
1604 * and we remap the whole kernel with page granularity.
1606 phys_addr_t kernel_x_start = kernel_sec_start;
1607 phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1608 phys_addr_t kernel_nx_start = kernel_x_end;
1609 phys_addr_t kernel_nx_end = kernel_sec_end;
1610 struct map_desc map;
1612 map.pfn = __phys_to_pfn(kernel_x_start);
1613 map.virtual = __phys_to_virt(kernel_x_start);
1614 map.length = kernel_x_end - kernel_x_start;
1615 map.type = MT_MEMORY_RWX;
1616 create_mapping(&map);
1618 /* If the nx part is small it may end up covered by the tail of the RWX section */
1619 if (kernel_x_end == kernel_nx_end)
1622 map.pfn = __phys_to_pfn(kernel_nx_start);
1623 map.virtual = __phys_to_virt(kernel_nx_start);
1624 map.length = kernel_nx_end - kernel_nx_start;
1625 map.type = MT_MEMORY_RW;
1626 create_mapping(&map);
1629 #ifdef CONFIG_ARM_PV_FIXUP
1630 typedef void pgtables_remap(long long offset, unsigned long pgd);
1631 pgtables_remap lpae_pgtables_remap_asm;
1634 * early_paging_init() recreates boot time page table setup, allowing machines
1635 * to switch over to a high (>4G) address space on LPAE systems
1637 static void __init early_paging_init(const struct machine_desc *mdesc)
1639 pgtables_remap *lpae_pgtables_remap;
1640 unsigned long pa_pgd;
1641 unsigned int cr, ttbcr;
1644 if (!mdesc->pv_fixup)
1647 offset = mdesc->pv_fixup();
1652 * Offset the kernel section physical offsets so that the kernel
1653 * mapping will work out later on.
1655 kernel_sec_start += offset;
1656 kernel_sec_end += offset;
1659 * Get the address of the remap function in the 1:1 identity
1660 * mapping setup by the early page table assembly code. We
1661 * must get this prior to the pv update. The following barrier
1662 * ensures that this is complete before we fixup any P:V offsets.
1664 lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1665 pa_pgd = __pa(swapper_pg_dir);
1668 pr_info("Switching physical address space to 0x%08llx\n",
1669 (u64)PHYS_OFFSET + offset);
1671 /* Re-set the phys pfn offset, and the pv offset */
1672 __pv_offset += offset;
1673 __pv_phys_pfn_offset += PFN_DOWN(offset);
1675 /* Run the patch stub to update the constants */
1676 fixup_pv_table(&__pv_table_begin,
1677 (&__pv_table_end - &__pv_table_begin) << 2);
1680 * We changing not only the virtual to physical mapping, but also
1681 * the physical addresses used to access memory. We need to flush
1682 * all levels of cache in the system with caching disabled to
1683 * ensure that all data is written back, and nothing is prefetched
1684 * into the caches. We also need to prevent the TLB walkers
1685 * allocating into the caches too. Note that this is ARMv7 LPAE
1689 set_cr(cr & ~(CR_I | CR_C));
1690 asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1691 asm volatile("mcr p15, 0, %0, c2, c0, 2"
1692 : : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1696 * Fixup the page tables - this must be in the idmap region as
1697 * we need to disable the MMU to do this safely, and hence it
1698 * needs to be assembly. It's fairly simple, as we're using the
1699 * temporary tables setup by the initial assembly code.
1701 lpae_pgtables_remap(offset, pa_pgd);
1703 /* Re-enable the caches and cacheable TLB walks */
1704 asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1710 static void __init early_paging_init(const struct machine_desc *mdesc)
1714 if (!mdesc->pv_fixup)
1717 offset = mdesc->pv_fixup();
1721 pr_crit("Physical address space modification is only to support Keystone2.\n");
1722 pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1723 pr_crit("feature. Your kernel may crash now, have a good day.\n");
1724 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1729 static void __init early_fixmap_shutdown(void)
1732 unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1);
1734 pte_offset_fixmap = pte_offset_late_fixmap;
1735 pmd_clear(fixmap_pmd(va));
1736 local_flush_tlb_kernel_page(va);
1738 for (i = 0; i < __end_of_permanent_fixed_addresses; i++) {
1740 struct map_desc map;
1742 map.virtual = fix_to_virt(i);
1743 pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual);
1745 /* Only i/o device mappings are supported ATM */
1746 if (pte_none(*pte) ||
1747 (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED)
1750 map.pfn = pte_pfn(*pte);
1751 map.type = MT_DEVICE;
1752 map.length = PAGE_SIZE;
1754 create_mapping(&map);
1759 * paging_init() sets up the page tables, initialises the zone memory
1760 * maps, and sets up the zero page, bad page and bad page tables.
1762 void __init paging_init(const struct machine_desc *mdesc)
1766 pr_debug("physical kernel sections: 0x%08llx-0x%08llx\n",
1767 kernel_sec_start, kernel_sec_end);
1769 prepare_page_table();
1771 memblock_set_current_limit(arm_lowmem_limit);
1772 pr_debug("lowmem limit is %08llx\n", (long long)arm_lowmem_limit);
1774 * After this point early_alloc(), i.e. the memblock allocator, can
1778 dma_contiguous_remap();
1779 early_fixmap_shutdown();
1780 devicemaps_init(mdesc);
1784 top_pmd = pmd_off_k(0xffff0000);
1786 /* allocate the zero page. */
1787 zero_page = early_alloc(PAGE_SIZE);
1791 empty_zero_page = virt_to_page(zero_page);
1792 __flush_dcache_page(NULL, empty_zero_page);
1795 void __init early_mm_init(const struct machine_desc *mdesc)
1797 build_mem_type_table();
1798 early_paging_init(mdesc);
1801 void set_pte_at(struct mm_struct *mm, unsigned long addr,
1802 pte_t *ptep, pte_t pteval)
1804 unsigned long ext = 0;
1806 if (addr < TASK_SIZE && pte_valid_user(pteval)) {
1807 if (!pte_special(pteval))
1808 __sync_icache_dcache(pteval);
1812 set_pte_ext(ptep, pteval, ext);