xref: /linux/arch/arm/mm/mmu.c (revision 165f2d2858013253042809df082b8df7e34e86d7)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/arch/arm/mm/mmu.c
4  *
5  *  Copyright (C) 1995-2005 Russell King
6  */
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>
14 #include <linux/fs.h>
15 #include <linux/vmalloc.h>
16 #include <linux/sizes.h>
17 
18 #include <asm/cp15.h>
19 #include <asm/cputype.h>
20 #include <asm/sections.h>
21 #include <asm/cachetype.h>
22 #include <asm/fixmap.h>
23 #include <asm/sections.h>
24 #include <asm/setup.h>
25 #include <asm/smp_plat.h>
26 #include <asm/tlb.h>
27 #include <asm/highmem.h>
28 #include <asm/system_info.h>
29 #include <asm/traps.h>
30 #include <asm/procinfo.h>
31 #include <asm/memory.h>
32 
33 #include <asm/mach/arch.h>
34 #include <asm/mach/map.h>
35 #include <asm/mach/pci.h>
36 #include <asm/fixmap.h>
37 
38 #include "fault.h"
39 #include "mm.h"
40 #include "tcm.h"
41 
42 /*
43  * empty_zero_page is a special page that is used for
44  * zero-initialized data and COW.
45  */
46 struct page *empty_zero_page;
47 EXPORT_SYMBOL(empty_zero_page);
48 
49 /*
50  * The pmd table for the upper-most set of pages.
51  */
52 pmd_t *top_pmd;
53 
54 pmdval_t user_pmd_table = _PAGE_USER_TABLE;
55 
56 #define CPOLICY_UNCACHED	0
57 #define CPOLICY_BUFFERED	1
58 #define CPOLICY_WRITETHROUGH	2
59 #define CPOLICY_WRITEBACK	3
60 #define CPOLICY_WRITEALLOC	4
61 
62 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
63 static unsigned int ecc_mask __initdata = 0;
64 pgprot_t pgprot_user;
65 pgprot_t pgprot_kernel;
66 
67 EXPORT_SYMBOL(pgprot_user);
68 EXPORT_SYMBOL(pgprot_kernel);
69 
70 struct cachepolicy {
71 	const char	policy[16];
72 	unsigned int	cr_mask;
73 	pmdval_t	pmd;
74 	pteval_t	pte;
75 };
76 
77 unsigned long kimage_voffset __ro_after_init;
78 
79 static struct cachepolicy cache_policies[] __initdata = {
80 	{
81 		.policy		= "uncached",
82 		.cr_mask	= CR_W|CR_C,
83 		.pmd		= PMD_SECT_UNCACHED,
84 		.pte		= L_PTE_MT_UNCACHED,
85 	}, {
86 		.policy		= "buffered",
87 		.cr_mask	= CR_C,
88 		.pmd		= PMD_SECT_BUFFERED,
89 		.pte		= L_PTE_MT_BUFFERABLE,
90 	}, {
91 		.policy		= "writethrough",
92 		.cr_mask	= 0,
93 		.pmd		= PMD_SECT_WT,
94 		.pte		= L_PTE_MT_WRITETHROUGH,
95 	}, {
96 		.policy		= "writeback",
97 		.cr_mask	= 0,
98 		.pmd		= PMD_SECT_WB,
99 		.pte		= L_PTE_MT_WRITEBACK,
100 	}, {
101 		.policy		= "writealloc",
102 		.cr_mask	= 0,
103 		.pmd		= PMD_SECT_WBWA,
104 		.pte		= L_PTE_MT_WRITEALLOC,
105 	}
106 };
107 
108 #ifdef CONFIG_CPU_CP15
109 static unsigned long initial_pmd_value __initdata = 0;
110 
111 /*
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.
117  */
118 void __init init_default_cache_policy(unsigned long pmd)
119 {
120 	int i;
121 
122 	initial_pmd_value = pmd;
123 
124 	pmd &= PMD_SECT_CACHE_MASK;
125 
126 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
127 		if (cache_policies[i].pmd == pmd) {
128 			cachepolicy = i;
129 			break;
130 		}
131 
132 	if (i == ARRAY_SIZE(cache_policies))
133 		pr_err("ERROR: could not find cache policy\n");
134 }
135 
136 /*
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).
140  */
141 static int __init early_cachepolicy(char *p)
142 {
143 	int i, selected = -1;
144 
145 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
146 		int len = strlen(cache_policies[i].policy);
147 
148 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
149 			selected = i;
150 			break;
151 		}
152 	}
153 
154 	if (selected == -1)
155 		pr_err("ERROR: unknown or unsupported cache policy\n");
156 
157 	/*
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
162 	 * page tables.
163 	 */
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);
167 		return 0;
168 	}
169 
170 	if (selected != cachepolicy) {
171 		unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
172 		cachepolicy = selected;
173 		flush_cache_all();
174 		set_cr(cr);
175 	}
176 	return 0;
177 }
178 early_param("cachepolicy", early_cachepolicy);
179 
180 static int __init early_nocache(char *__unused)
181 {
182 	char *p = "buffered";
183 	pr_warn("nocache is deprecated; use cachepolicy=%s\n", p);
184 	early_cachepolicy(p);
185 	return 0;
186 }
187 early_param("nocache", early_nocache);
188 
189 static int __init early_nowrite(char *__unused)
190 {
191 	char *p = "uncached";
192 	pr_warn("nowb is deprecated; use cachepolicy=%s\n", p);
193 	early_cachepolicy(p);
194 	return 0;
195 }
196 early_param("nowb", early_nowrite);
197 
198 #ifndef CONFIG_ARM_LPAE
199 static int __init early_ecc(char *p)
200 {
201 	if (memcmp(p, "on", 2) == 0)
202 		ecc_mask = PMD_PROTECTION;
203 	else if (memcmp(p, "off", 3) == 0)
204 		ecc_mask = 0;
205 	return 0;
206 }
207 early_param("ecc", early_ecc);
208 #endif
209 
210 #else /* ifdef CONFIG_CPU_CP15 */
211 
212 static int __init early_cachepolicy(char *p)
213 {
214 	pr_warn("cachepolicy kernel parameter not supported without cp15\n");
215 }
216 early_param("cachepolicy", early_cachepolicy);
217 
218 static int __init noalign_setup(char *__unused)
219 {
220 	pr_warn("noalign kernel parameter not supported without cp15\n");
221 }
222 __setup("noalign", noalign_setup);
223 
224 #endif /* ifdef CONFIG_CPU_CP15 / else */
225 
226 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
227 #define PROT_PTE_S2_DEVICE	PROT_PTE_DEVICE
228 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
229 
230 static struct mem_type mem_types[] __ro_after_init = {
231 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
232 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
233 				  L_PTE_SHARED,
234 		.prot_l1	= PMD_TYPE_TABLE,
235 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
236 		.domain		= DOMAIN_IO,
237 	},
238 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
239 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
240 		.prot_l1	= PMD_TYPE_TABLE,
241 		.prot_sect	= PROT_SECT_DEVICE,
242 		.domain		= DOMAIN_IO,
243 	},
244 	[MT_DEVICE_CACHED] = {	  /* ioremap_cache */
245 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
246 		.prot_l1	= PMD_TYPE_TABLE,
247 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
248 		.domain		= DOMAIN_IO,
249 	},
250 	[MT_DEVICE_WC] = {	/* ioremap_wc */
251 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
252 		.prot_l1	= PMD_TYPE_TABLE,
253 		.prot_sect	= PROT_SECT_DEVICE,
254 		.domain		= DOMAIN_IO,
255 	},
256 	[MT_UNCACHED] = {
257 		.prot_pte	= PROT_PTE_DEVICE,
258 		.prot_l1	= PMD_TYPE_TABLE,
259 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
260 		.domain		= DOMAIN_IO,
261 	},
262 	[MT_CACHECLEAN] = {
263 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
264 		.domain    = DOMAIN_KERNEL,
265 	},
266 #ifndef CONFIG_ARM_LPAE
267 	[MT_MINICLEAN] = {
268 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
269 		.domain    = DOMAIN_KERNEL,
270 	},
271 #endif
272 	[MT_LOW_VECTORS] = {
273 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
274 				L_PTE_RDONLY,
275 		.prot_l1   = PMD_TYPE_TABLE,
276 		.domain    = DOMAIN_VECTORS,
277 	},
278 	[MT_HIGH_VECTORS] = {
279 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
280 				L_PTE_USER | L_PTE_RDONLY,
281 		.prot_l1   = PMD_TYPE_TABLE,
282 		.domain    = DOMAIN_VECTORS,
283 	},
284 	[MT_MEMORY_RWX] = {
285 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
286 		.prot_l1   = PMD_TYPE_TABLE,
287 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
288 		.domain    = DOMAIN_KERNEL,
289 	},
290 	[MT_MEMORY_RW] = {
291 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
292 			     L_PTE_XN,
293 		.prot_l1   = PMD_TYPE_TABLE,
294 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
295 		.domain    = DOMAIN_KERNEL,
296 	},
297 	[MT_ROM] = {
298 		.prot_sect = PMD_TYPE_SECT,
299 		.domain    = DOMAIN_KERNEL,
300 	},
301 	[MT_MEMORY_RWX_NONCACHED] = {
302 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
303 				L_PTE_MT_BUFFERABLE,
304 		.prot_l1   = PMD_TYPE_TABLE,
305 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
306 		.domain    = DOMAIN_KERNEL,
307 	},
308 	[MT_MEMORY_RW_DTCM] = {
309 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
310 				L_PTE_XN,
311 		.prot_l1   = PMD_TYPE_TABLE,
312 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
313 		.domain    = DOMAIN_KERNEL,
314 	},
315 	[MT_MEMORY_RWX_ITCM] = {
316 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
317 		.prot_l1   = PMD_TYPE_TABLE,
318 		.domain    = DOMAIN_KERNEL,
319 	},
320 	[MT_MEMORY_RW_SO] = {
321 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
322 				L_PTE_MT_UNCACHED | L_PTE_XN,
323 		.prot_l1   = PMD_TYPE_TABLE,
324 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
325 				PMD_SECT_UNCACHED | PMD_SECT_XN,
326 		.domain    = DOMAIN_KERNEL,
327 	},
328 	[MT_MEMORY_DMA_READY] = {
329 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
330 				L_PTE_XN,
331 		.prot_l1   = PMD_TYPE_TABLE,
332 		.domain    = DOMAIN_KERNEL,
333 	},
334 };
335 
336 const struct mem_type *get_mem_type(unsigned int type)
337 {
338 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
339 }
340 EXPORT_SYMBOL(get_mem_type);
341 
342 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr);
343 
344 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS]
345 	__aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata;
346 
347 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr)
348 {
349 	return &bm_pte[pte_index(addr)];
350 }
351 
352 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr)
353 {
354 	return pte_offset_kernel(dir, addr);
355 }
356 
357 static inline pmd_t * __init fixmap_pmd(unsigned long addr)
358 {
359 	pgd_t *pgd = pgd_offset_k(addr);
360 	pud_t *pud = pud_offset(pgd, addr);
361 	pmd_t *pmd = pmd_offset(pud, addr);
362 
363 	return pmd;
364 }
365 
366 void __init early_fixmap_init(void)
367 {
368 	pmd_t *pmd;
369 
370 	/*
371 	 * The early fixmap range spans multiple pmds, for which
372 	 * we are not prepared:
373 	 */
374 	BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT)
375 		     != FIXADDR_TOP >> PMD_SHIFT);
376 
377 	pmd = fixmap_pmd(FIXADDR_TOP);
378 	pmd_populate_kernel(&init_mm, pmd, bm_pte);
379 
380 	pte_offset_fixmap = pte_offset_early_fixmap;
381 }
382 
383 /*
384  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
385  * As a result, this can only be called with preemption disabled, as under
386  * stop_machine().
387  */
388 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
389 {
390 	unsigned long vaddr = __fix_to_virt(idx);
391 	pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr);
392 
393 	/* Make sure fixmap region does not exceed available allocation. */
394 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
395 		     FIXADDR_END);
396 	BUG_ON(idx >= __end_of_fixed_addresses);
397 
398 	/* we only support device mappings until pgprot_kernel has been set */
399 	if (WARN_ON(pgprot_val(prot) != pgprot_val(FIXMAP_PAGE_IO) &&
400 		    pgprot_val(pgprot_kernel) == 0))
401 		return;
402 
403 	if (pgprot_val(prot))
404 		set_pte_at(NULL, vaddr, pte,
405 			pfn_pte(phys >> PAGE_SHIFT, prot));
406 	else
407 		pte_clear(NULL, vaddr, pte);
408 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
409 }
410 
411 /*
412  * Adjust the PMD section entries according to the CPU in use.
413  */
414 static void __init build_mem_type_table(void)
415 {
416 	struct cachepolicy *cp;
417 	unsigned int cr = get_cr();
418 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
419 	int cpu_arch = cpu_architecture();
420 	int i;
421 
422 	if (cpu_arch < CPU_ARCH_ARMv6) {
423 #if defined(CONFIG_CPU_DCACHE_DISABLE)
424 		if (cachepolicy > CPOLICY_BUFFERED)
425 			cachepolicy = CPOLICY_BUFFERED;
426 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
427 		if (cachepolicy > CPOLICY_WRITETHROUGH)
428 			cachepolicy = CPOLICY_WRITETHROUGH;
429 #endif
430 	}
431 	if (cpu_arch < CPU_ARCH_ARMv5) {
432 		if (cachepolicy >= CPOLICY_WRITEALLOC)
433 			cachepolicy = CPOLICY_WRITEBACK;
434 		ecc_mask = 0;
435 	}
436 
437 	if (is_smp()) {
438 		if (cachepolicy != CPOLICY_WRITEALLOC) {
439 			pr_warn("Forcing write-allocate cache policy for SMP\n");
440 			cachepolicy = CPOLICY_WRITEALLOC;
441 		}
442 		if (!(initial_pmd_value & PMD_SECT_S)) {
443 			pr_warn("Forcing shared mappings for SMP\n");
444 			initial_pmd_value |= PMD_SECT_S;
445 		}
446 	}
447 
448 	/*
449 	 * Strip out features not present on earlier architectures.
450 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
451 	 * without extended page tables don't have the 'Shared' bit.
452 	 */
453 	if (cpu_arch < CPU_ARCH_ARMv5)
454 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
455 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
456 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
457 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
458 			mem_types[i].prot_sect &= ~PMD_SECT_S;
459 
460 	/*
461 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
462 	 * "update-able on write" bit on ARM610).  However, Xscale and
463 	 * Xscale3 require this bit to be cleared.
464 	 */
465 	if (cpu_is_xscale_family()) {
466 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
467 			mem_types[i].prot_sect &= ~PMD_BIT4;
468 			mem_types[i].prot_l1 &= ~PMD_BIT4;
469 		}
470 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
471 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
472 			if (mem_types[i].prot_l1)
473 				mem_types[i].prot_l1 |= PMD_BIT4;
474 			if (mem_types[i].prot_sect)
475 				mem_types[i].prot_sect |= PMD_BIT4;
476 		}
477 	}
478 
479 	/*
480 	 * Mark the device areas according to the CPU/architecture.
481 	 */
482 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
483 		if (!cpu_is_xsc3()) {
484 			/*
485 			 * Mark device regions on ARMv6+ as execute-never
486 			 * to prevent speculative instruction fetches.
487 			 */
488 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
489 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
490 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
491 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
492 
493 			/* Also setup NX memory mapping */
494 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
495 		}
496 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
497 			/*
498 			 * For ARMv7 with TEX remapping,
499 			 * - shared device is SXCB=1100
500 			 * - nonshared device is SXCB=0100
501 			 * - write combine device mem is SXCB=0001
502 			 * (Uncached Normal memory)
503 			 */
504 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
505 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
506 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
507 		} else if (cpu_is_xsc3()) {
508 			/*
509 			 * For Xscale3,
510 			 * - shared device is TEXCB=00101
511 			 * - nonshared device is TEXCB=01000
512 			 * - write combine device mem is TEXCB=00100
513 			 * (Inner/Outer Uncacheable in xsc3 parlance)
514 			 */
515 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
516 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
517 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
518 		} else {
519 			/*
520 			 * For ARMv6 and ARMv7 without TEX remapping,
521 			 * - shared device is TEXCB=00001
522 			 * - nonshared device is TEXCB=01000
523 			 * - write combine device mem is TEXCB=00100
524 			 * (Uncached Normal in ARMv6 parlance).
525 			 */
526 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
527 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
528 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
529 		}
530 	} else {
531 		/*
532 		 * On others, write combining is "Uncached/Buffered"
533 		 */
534 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
535 	}
536 
537 	/*
538 	 * Now deal with the memory-type mappings
539 	 */
540 	cp = &cache_policies[cachepolicy];
541 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
542 
543 #ifndef CONFIG_ARM_LPAE
544 	/*
545 	 * We don't use domains on ARMv6 (since this causes problems with
546 	 * v6/v7 kernels), so we must use a separate memory type for user
547 	 * r/o, kernel r/w to map the vectors page.
548 	 */
549 	if (cpu_arch == CPU_ARCH_ARMv6)
550 		vecs_pgprot |= L_PTE_MT_VECTORS;
551 
552 	/*
553 	 * Check is it with support for the PXN bit
554 	 * in the Short-descriptor translation table format descriptors.
555 	 */
556 	if (cpu_arch == CPU_ARCH_ARMv7 &&
557 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
558 		user_pmd_table |= PMD_PXNTABLE;
559 	}
560 #endif
561 
562 	/*
563 	 * ARMv6 and above have extended page tables.
564 	 */
565 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
566 #ifndef CONFIG_ARM_LPAE
567 		/*
568 		 * Mark cache clean areas and XIP ROM read only
569 		 * from SVC mode and no access from userspace.
570 		 */
571 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
572 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
573 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
574 #endif
575 
576 		/*
577 		 * If the initial page tables were created with the S bit
578 		 * set, then we need to do the same here for the same
579 		 * reasons given in early_cachepolicy().
580 		 */
581 		if (initial_pmd_value & PMD_SECT_S) {
582 			user_pgprot |= L_PTE_SHARED;
583 			kern_pgprot |= L_PTE_SHARED;
584 			vecs_pgprot |= L_PTE_SHARED;
585 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
586 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
587 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
588 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
589 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
590 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
591 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
592 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
593 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
594 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
595 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
596 		}
597 	}
598 
599 	/*
600 	 * Non-cacheable Normal - intended for memory areas that must
601 	 * not cause dirty cache line writebacks when used
602 	 */
603 	if (cpu_arch >= CPU_ARCH_ARMv6) {
604 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
605 			/* Non-cacheable Normal is XCB = 001 */
606 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
607 				PMD_SECT_BUFFERED;
608 		} else {
609 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
610 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
611 				PMD_SECT_TEX(1);
612 		}
613 	} else {
614 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
615 	}
616 
617 #ifdef CONFIG_ARM_LPAE
618 	/*
619 	 * Do not generate access flag faults for the kernel mappings.
620 	 */
621 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
622 		mem_types[i].prot_pte |= PTE_EXT_AF;
623 		if (mem_types[i].prot_sect)
624 			mem_types[i].prot_sect |= PMD_SECT_AF;
625 	}
626 	kern_pgprot |= PTE_EXT_AF;
627 	vecs_pgprot |= PTE_EXT_AF;
628 
629 	/*
630 	 * Set PXN for user mappings
631 	 */
632 	user_pgprot |= PTE_EXT_PXN;
633 #endif
634 
635 	for (i = 0; i < 16; i++) {
636 		pteval_t v = pgprot_val(protection_map[i]);
637 		protection_map[i] = __pgprot(v | user_pgprot);
638 	}
639 
640 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
641 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
642 
643 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
644 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
645 				 L_PTE_DIRTY | kern_pgprot);
646 
647 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
648 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
649 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
650 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
651 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
652 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
653 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
654 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
655 	mem_types[MT_ROM].prot_sect |= cp->pmd;
656 
657 	switch (cp->pmd) {
658 	case PMD_SECT_WT:
659 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
660 		break;
661 	case PMD_SECT_WB:
662 	case PMD_SECT_WBWA:
663 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
664 		break;
665 	}
666 	pr_info("Memory policy: %sData cache %s\n",
667 		ecc_mask ? "ECC enabled, " : "", cp->policy);
668 
669 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
670 		struct mem_type *t = &mem_types[i];
671 		if (t->prot_l1)
672 			t->prot_l1 |= PMD_DOMAIN(t->domain);
673 		if (t->prot_sect)
674 			t->prot_sect |= PMD_DOMAIN(t->domain);
675 	}
676 }
677 
678 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
679 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
680 			      unsigned long size, pgprot_t vma_prot)
681 {
682 	if (!pfn_valid(pfn))
683 		return pgprot_noncached(vma_prot);
684 	else if (file->f_flags & O_SYNC)
685 		return pgprot_writecombine(vma_prot);
686 	return vma_prot;
687 }
688 EXPORT_SYMBOL(phys_mem_access_prot);
689 #endif
690 
691 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
692 
693 static void __init *early_alloc(unsigned long sz)
694 {
695 	void *ptr = memblock_alloc(sz, sz);
696 
697 	if (!ptr)
698 		panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
699 		      __func__, sz, sz);
700 
701 	return ptr;
702 }
703 
704 static void *__init late_alloc(unsigned long sz)
705 {
706 	void *ptr = (void *)__get_free_pages(GFP_PGTABLE_KERNEL, get_order(sz));
707 
708 	if (!ptr || !pgtable_pte_page_ctor(virt_to_page(ptr)))
709 		BUG();
710 	return ptr;
711 }
712 
713 static pte_t * __init arm_pte_alloc(pmd_t *pmd, unsigned long addr,
714 				unsigned long prot,
715 				void *(*alloc)(unsigned long sz))
716 {
717 	if (pmd_none(*pmd)) {
718 		pte_t *pte = alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
719 		__pmd_populate(pmd, __pa(pte), prot);
720 	}
721 	BUG_ON(pmd_bad(*pmd));
722 	return pte_offset_kernel(pmd, addr);
723 }
724 
725 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr,
726 				      unsigned long prot)
727 {
728 	return arm_pte_alloc(pmd, addr, prot, early_alloc);
729 }
730 
731 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
732 				  unsigned long end, unsigned long pfn,
733 				  const struct mem_type *type,
734 				  void *(*alloc)(unsigned long sz),
735 				  bool ng)
736 {
737 	pte_t *pte = arm_pte_alloc(pmd, addr, type->prot_l1, alloc);
738 	do {
739 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)),
740 			    ng ? PTE_EXT_NG : 0);
741 		pfn++;
742 	} while (pte++, addr += PAGE_SIZE, addr != end);
743 }
744 
745 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
746 			unsigned long end, phys_addr_t phys,
747 			const struct mem_type *type, bool ng)
748 {
749 	pmd_t *p = pmd;
750 
751 #ifndef CONFIG_ARM_LPAE
752 	/*
753 	 * In classic MMU format, puds and pmds are folded in to
754 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
755 	 * group of L1 entries making up one logical pointer to
756 	 * an L2 table (2MB), where as PMDs refer to the individual
757 	 * L1 entries (1MB). Hence increment to get the correct
758 	 * offset for odd 1MB sections.
759 	 * (See arch/arm/include/asm/pgtable-2level.h)
760 	 */
761 	if (addr & SECTION_SIZE)
762 		pmd++;
763 #endif
764 	do {
765 		*pmd = __pmd(phys | type->prot_sect | (ng ? PMD_SECT_nG : 0));
766 		phys += SECTION_SIZE;
767 	} while (pmd++, addr += SECTION_SIZE, addr != end);
768 
769 	flush_pmd_entry(p);
770 }
771 
772 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
773 				      unsigned long end, phys_addr_t phys,
774 				      const struct mem_type *type,
775 				      void *(*alloc)(unsigned long sz), bool ng)
776 {
777 	pmd_t *pmd = pmd_offset(pud, addr);
778 	unsigned long next;
779 
780 	do {
781 		/*
782 		 * With LPAE, we must loop over to map
783 		 * all the pmds for the given range.
784 		 */
785 		next = pmd_addr_end(addr, end);
786 
787 		/*
788 		 * Try a section mapping - addr, next and phys must all be
789 		 * aligned to a section boundary.
790 		 */
791 		if (type->prot_sect &&
792 				((addr | next | phys) & ~SECTION_MASK) == 0) {
793 			__map_init_section(pmd, addr, next, phys, type, ng);
794 		} else {
795 			alloc_init_pte(pmd, addr, next,
796 				       __phys_to_pfn(phys), type, alloc, ng);
797 		}
798 
799 		phys += next - addr;
800 
801 	} while (pmd++, addr = next, addr != end);
802 }
803 
804 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
805 				  unsigned long end, phys_addr_t phys,
806 				  const struct mem_type *type,
807 				  void *(*alloc)(unsigned long sz), bool ng)
808 {
809 	pud_t *pud = pud_offset(pgd, addr);
810 	unsigned long next;
811 
812 	do {
813 		next = pud_addr_end(addr, end);
814 		alloc_init_pmd(pud, addr, next, phys, type, alloc, ng);
815 		phys += next - addr;
816 	} while (pud++, addr = next, addr != end);
817 }
818 
819 #ifndef CONFIG_ARM_LPAE
820 static void __init create_36bit_mapping(struct mm_struct *mm,
821 					struct map_desc *md,
822 					const struct mem_type *type,
823 					bool ng)
824 {
825 	unsigned long addr, length, end;
826 	phys_addr_t phys;
827 	pgd_t *pgd;
828 
829 	addr = md->virtual;
830 	phys = __pfn_to_phys(md->pfn);
831 	length = PAGE_ALIGN(md->length);
832 
833 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
834 		pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
835 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
836 		return;
837 	}
838 
839 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
840 	 *	Since domain assignments can in fact be arbitrary, the
841 	 *	'domain == 0' check below is required to insure that ARMv6
842 	 *	supersections are only allocated for domain 0 regardless
843 	 *	of the actual domain assignments in use.
844 	 */
845 	if (type->domain) {
846 		pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
847 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
848 		return;
849 	}
850 
851 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
852 		pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
853 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
854 		return;
855 	}
856 
857 	/*
858 	 * Shift bits [35:32] of address into bits [23:20] of PMD
859 	 * (See ARMv6 spec).
860 	 */
861 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
862 
863 	pgd = pgd_offset(mm, addr);
864 	end = addr + length;
865 	do {
866 		pud_t *pud = pud_offset(pgd, addr);
867 		pmd_t *pmd = pmd_offset(pud, addr);
868 		int i;
869 
870 		for (i = 0; i < 16; i++)
871 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER |
872 				       (ng ? PMD_SECT_nG : 0));
873 
874 		addr += SUPERSECTION_SIZE;
875 		phys += SUPERSECTION_SIZE;
876 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
877 	} while (addr != end);
878 }
879 #endif	/* !CONFIG_ARM_LPAE */
880 
881 static void __init __create_mapping(struct mm_struct *mm, struct map_desc *md,
882 				    void *(*alloc)(unsigned long sz),
883 				    bool ng)
884 {
885 	unsigned long addr, length, end;
886 	phys_addr_t phys;
887 	const struct mem_type *type;
888 	pgd_t *pgd;
889 
890 	type = &mem_types[md->type];
891 
892 #ifndef CONFIG_ARM_LPAE
893 	/*
894 	 * Catch 36-bit addresses
895 	 */
896 	if (md->pfn >= 0x100000) {
897 		create_36bit_mapping(mm, md, type, ng);
898 		return;
899 	}
900 #endif
901 
902 	addr = md->virtual & PAGE_MASK;
903 	phys = __pfn_to_phys(md->pfn);
904 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
905 
906 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
907 		pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
908 			(long long)__pfn_to_phys(md->pfn), addr);
909 		return;
910 	}
911 
912 	pgd = pgd_offset(mm, addr);
913 	end = addr + length;
914 	do {
915 		unsigned long next = pgd_addr_end(addr, end);
916 
917 		alloc_init_pud(pgd, addr, next, phys, type, alloc, ng);
918 
919 		phys += next - addr;
920 		addr = next;
921 	} while (pgd++, addr != end);
922 }
923 
924 /*
925  * Create the page directory entries and any necessary
926  * page tables for the mapping specified by `md'.  We
927  * are able to cope here with varying sizes and address
928  * offsets, and we take full advantage of sections and
929  * supersections.
930  */
931 static void __init create_mapping(struct map_desc *md)
932 {
933 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
934 		pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
935 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
936 		return;
937 	}
938 
939 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
940 	    md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START &&
941 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
942 		pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
943 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
944 	}
945 
946 	__create_mapping(&init_mm, md, early_alloc, false);
947 }
948 
949 void __init create_mapping_late(struct mm_struct *mm, struct map_desc *md,
950 				bool ng)
951 {
952 #ifdef CONFIG_ARM_LPAE
953 	pud_t *pud = pud_alloc(mm, pgd_offset(mm, md->virtual), md->virtual);
954 	if (WARN_ON(!pud))
955 		return;
956 	pmd_alloc(mm, pud, 0);
957 #endif
958 	__create_mapping(mm, md, late_alloc, ng);
959 }
960 
961 /*
962  * Create the architecture specific mappings
963  */
964 void __init iotable_init(struct map_desc *io_desc, int nr)
965 {
966 	struct map_desc *md;
967 	struct vm_struct *vm;
968 	struct static_vm *svm;
969 
970 	if (!nr)
971 		return;
972 
973 	svm = memblock_alloc(sizeof(*svm) * nr, __alignof__(*svm));
974 	if (!svm)
975 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
976 		      __func__, sizeof(*svm) * nr, __alignof__(*svm));
977 
978 	for (md = io_desc; nr; md++, nr--) {
979 		create_mapping(md);
980 
981 		vm = &svm->vm;
982 		vm->addr = (void *)(md->virtual & PAGE_MASK);
983 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
984 		vm->phys_addr = __pfn_to_phys(md->pfn);
985 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
986 		vm->flags |= VM_ARM_MTYPE(md->type);
987 		vm->caller = iotable_init;
988 		add_static_vm_early(svm++);
989 	}
990 }
991 
992 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
993 				  void *caller)
994 {
995 	struct vm_struct *vm;
996 	struct static_vm *svm;
997 
998 	svm = memblock_alloc(sizeof(*svm), __alignof__(*svm));
999 	if (!svm)
1000 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1001 		      __func__, sizeof(*svm), __alignof__(*svm));
1002 
1003 	vm = &svm->vm;
1004 	vm->addr = (void *)addr;
1005 	vm->size = size;
1006 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
1007 	vm->caller = caller;
1008 	add_static_vm_early(svm);
1009 }
1010 
1011 #ifndef CONFIG_ARM_LPAE
1012 
1013 /*
1014  * The Linux PMD is made of two consecutive section entries covering 2MB
1015  * (see definition in include/asm/pgtable-2level.h).  However a call to
1016  * create_mapping() may optimize static mappings by using individual
1017  * 1MB section mappings.  This leaves the actual PMD potentially half
1018  * initialized if the top or bottom section entry isn't used, leaving it
1019  * open to problems if a subsequent ioremap() or vmalloc() tries to use
1020  * the virtual space left free by that unused section entry.
1021  *
1022  * Let's avoid the issue by inserting dummy vm entries covering the unused
1023  * PMD halves once the static mappings are in place.
1024  */
1025 
1026 static void __init pmd_empty_section_gap(unsigned long addr)
1027 {
1028 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1029 }
1030 
1031 static void __init fill_pmd_gaps(void)
1032 {
1033 	struct static_vm *svm;
1034 	struct vm_struct *vm;
1035 	unsigned long addr, next = 0;
1036 	pmd_t *pmd;
1037 
1038 	list_for_each_entry(svm, &static_vmlist, list) {
1039 		vm = &svm->vm;
1040 		addr = (unsigned long)vm->addr;
1041 		if (addr < next)
1042 			continue;
1043 
1044 		/*
1045 		 * Check if this vm starts on an odd section boundary.
1046 		 * If so and the first section entry for this PMD is free
1047 		 * then we block the corresponding virtual address.
1048 		 */
1049 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1050 			pmd = pmd_off_k(addr);
1051 			if (pmd_none(*pmd))
1052 				pmd_empty_section_gap(addr & PMD_MASK);
1053 		}
1054 
1055 		/*
1056 		 * Then check if this vm ends on an odd section boundary.
1057 		 * If so and the second section entry for this PMD is empty
1058 		 * then we block the corresponding virtual address.
1059 		 */
1060 		addr += vm->size;
1061 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1062 			pmd = pmd_off_k(addr) + 1;
1063 			if (pmd_none(*pmd))
1064 				pmd_empty_section_gap(addr);
1065 		}
1066 
1067 		/* no need to look at any vm entry until we hit the next PMD */
1068 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1069 	}
1070 }
1071 
1072 #else
1073 #define fill_pmd_gaps() do { } while (0)
1074 #endif
1075 
1076 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1077 static void __init pci_reserve_io(void)
1078 {
1079 	struct static_vm *svm;
1080 
1081 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1082 	if (svm)
1083 		return;
1084 
1085 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1086 }
1087 #else
1088 #define pci_reserve_io() do { } while (0)
1089 #endif
1090 
1091 #ifdef CONFIG_DEBUG_LL
1092 void __init debug_ll_io_init(void)
1093 {
1094 	struct map_desc map;
1095 
1096 	debug_ll_addr(&map.pfn, &map.virtual);
1097 	if (!map.pfn || !map.virtual)
1098 		return;
1099 	map.pfn = __phys_to_pfn(map.pfn);
1100 	map.virtual &= PAGE_MASK;
1101 	map.length = PAGE_SIZE;
1102 	map.type = MT_DEVICE;
1103 	iotable_init(&map, 1);
1104 }
1105 #endif
1106 
1107 static void * __initdata vmalloc_min =
1108 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1109 
1110 /*
1111  * vmalloc=size forces the vmalloc area to be exactly 'size'
1112  * bytes. This can be used to increase (or decrease) the vmalloc
1113  * area - the default is 240m.
1114  */
1115 static int __init early_vmalloc(char *arg)
1116 {
1117 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1118 
1119 	if (vmalloc_reserve < SZ_16M) {
1120 		vmalloc_reserve = SZ_16M;
1121 		pr_warn("vmalloc area too small, limiting to %luMB\n",
1122 			vmalloc_reserve >> 20);
1123 	}
1124 
1125 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1126 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1127 		pr_warn("vmalloc area is too big, limiting to %luMB\n",
1128 			vmalloc_reserve >> 20);
1129 	}
1130 
1131 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1132 	return 0;
1133 }
1134 early_param("vmalloc", early_vmalloc);
1135 
1136 phys_addr_t arm_lowmem_limit __initdata = 0;
1137 
1138 void __init adjust_lowmem_bounds(void)
1139 {
1140 	phys_addr_t memblock_limit = 0;
1141 	u64 vmalloc_limit;
1142 	struct memblock_region *reg;
1143 	phys_addr_t lowmem_limit = 0;
1144 
1145 	/*
1146 	 * Let's use our own (unoptimized) equivalent of __pa() that is
1147 	 * not affected by wrap-arounds when sizeof(phys_addr_t) == 4.
1148 	 * The result is used as the upper bound on physical memory address
1149 	 * and may itself be outside the valid range for which phys_addr_t
1150 	 * and therefore __pa() is defined.
1151 	 */
1152 	vmalloc_limit = (u64)(uintptr_t)vmalloc_min - PAGE_OFFSET + PHYS_OFFSET;
1153 
1154 	/*
1155 	 * The first usable region must be PMD aligned. Mark its start
1156 	 * as MEMBLOCK_NOMAP if it isn't
1157 	 */
1158 	for_each_memblock(memory, reg) {
1159 		if (!memblock_is_nomap(reg)) {
1160 			if (!IS_ALIGNED(reg->base, PMD_SIZE)) {
1161 				phys_addr_t len;
1162 
1163 				len = round_up(reg->base, PMD_SIZE) - reg->base;
1164 				memblock_mark_nomap(reg->base, len);
1165 			}
1166 			break;
1167 		}
1168 	}
1169 
1170 	for_each_memblock(memory, reg) {
1171 		phys_addr_t block_start = reg->base;
1172 		phys_addr_t block_end = reg->base + reg->size;
1173 
1174 		if (memblock_is_nomap(reg))
1175 			continue;
1176 
1177 		if (reg->base < vmalloc_limit) {
1178 			if (block_end > lowmem_limit)
1179 				/*
1180 				 * Compare as u64 to ensure vmalloc_limit does
1181 				 * not get truncated. block_end should always
1182 				 * fit in phys_addr_t so there should be no
1183 				 * issue with assignment.
1184 				 */
1185 				lowmem_limit = min_t(u64,
1186 							 vmalloc_limit,
1187 							 block_end);
1188 
1189 			/*
1190 			 * Find the first non-pmd-aligned page, and point
1191 			 * memblock_limit at it. This relies on rounding the
1192 			 * limit down to be pmd-aligned, which happens at the
1193 			 * end of this function.
1194 			 *
1195 			 * With this algorithm, the start or end of almost any
1196 			 * bank can be non-pmd-aligned. The only exception is
1197 			 * that the start of the bank 0 must be section-
1198 			 * aligned, since otherwise memory would need to be
1199 			 * allocated when mapping the start of bank 0, which
1200 			 * occurs before any free memory is mapped.
1201 			 */
1202 			if (!memblock_limit) {
1203 				if (!IS_ALIGNED(block_start, PMD_SIZE))
1204 					memblock_limit = block_start;
1205 				else if (!IS_ALIGNED(block_end, PMD_SIZE))
1206 					memblock_limit = lowmem_limit;
1207 			}
1208 
1209 		}
1210 	}
1211 
1212 	arm_lowmem_limit = lowmem_limit;
1213 
1214 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1215 
1216 	if (!memblock_limit)
1217 		memblock_limit = arm_lowmem_limit;
1218 
1219 	/*
1220 	 * Round the memblock limit down to a pmd size.  This
1221 	 * helps to ensure that we will allocate memory from the
1222 	 * last full pmd, which should be mapped.
1223 	 */
1224 	memblock_limit = round_down(memblock_limit, PMD_SIZE);
1225 
1226 	if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1227 		if (memblock_end_of_DRAM() > arm_lowmem_limit) {
1228 			phys_addr_t end = memblock_end_of_DRAM();
1229 
1230 			pr_notice("Ignoring RAM at %pa-%pa\n",
1231 				  &memblock_limit, &end);
1232 			pr_notice("Consider using a HIGHMEM enabled kernel.\n");
1233 
1234 			memblock_remove(memblock_limit, end - memblock_limit);
1235 		}
1236 	}
1237 
1238 	memblock_set_current_limit(memblock_limit);
1239 }
1240 
1241 static inline void prepare_page_table(void)
1242 {
1243 	unsigned long addr;
1244 	phys_addr_t end;
1245 
1246 	/*
1247 	 * Clear out all the mappings below the kernel image.
1248 	 */
1249 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1250 		pmd_clear(pmd_off_k(addr));
1251 
1252 #ifdef CONFIG_XIP_KERNEL
1253 	/* The XIP kernel is mapped in the module area -- skip over it */
1254 	addr = ((unsigned long)_exiprom + PMD_SIZE - 1) & PMD_MASK;
1255 #endif
1256 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1257 		pmd_clear(pmd_off_k(addr));
1258 
1259 	/*
1260 	 * Find the end of the first block of lowmem.
1261 	 */
1262 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1263 	if (end >= arm_lowmem_limit)
1264 		end = arm_lowmem_limit;
1265 
1266 	/*
1267 	 * Clear out all the kernel space mappings, except for the first
1268 	 * memory bank, up to the vmalloc region.
1269 	 */
1270 	for (addr = __phys_to_virt(end);
1271 	     addr < VMALLOC_START; addr += PMD_SIZE)
1272 		pmd_clear(pmd_off_k(addr));
1273 }
1274 
1275 #ifdef CONFIG_ARM_LPAE
1276 /* the first page is reserved for pgd */
1277 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1278 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1279 #else
1280 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1281 #endif
1282 
1283 /*
1284  * Reserve the special regions of memory
1285  */
1286 void __init arm_mm_memblock_reserve(void)
1287 {
1288 	/*
1289 	 * Reserve the page tables.  These are already in use,
1290 	 * and can only be in node 0.
1291 	 */
1292 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1293 
1294 #ifdef CONFIG_SA1111
1295 	/*
1296 	 * Because of the SA1111 DMA bug, we want to preserve our
1297 	 * precious DMA-able memory...
1298 	 */
1299 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1300 #endif
1301 }
1302 
1303 /*
1304  * Set up the device mappings.  Since we clear out the page tables for all
1305  * mappings above VMALLOC_START, except early fixmap, we might remove debug
1306  * device mappings.  This means earlycon can be used to debug this function
1307  * Any other function or debugging method which may touch any device _will_
1308  * crash the kernel.
1309  */
1310 static void __init devicemaps_init(const struct machine_desc *mdesc)
1311 {
1312 	struct map_desc map;
1313 	unsigned long addr;
1314 	void *vectors;
1315 
1316 	/*
1317 	 * Allocate the vector page early.
1318 	 */
1319 	vectors = early_alloc(PAGE_SIZE * 2);
1320 
1321 	early_trap_init(vectors);
1322 
1323 	/*
1324 	 * Clear page table except top pmd used by early fixmaps
1325 	 */
1326 	for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE)
1327 		pmd_clear(pmd_off_k(addr));
1328 
1329 	/*
1330 	 * Map the kernel if it is XIP.
1331 	 * It is always first in the modulearea.
1332 	 */
1333 #ifdef CONFIG_XIP_KERNEL
1334 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1335 	map.virtual = MODULES_VADDR;
1336 	map.length = ((unsigned long)_exiprom - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1337 	map.type = MT_ROM;
1338 	create_mapping(&map);
1339 #endif
1340 
1341 	/*
1342 	 * Map the cache flushing regions.
1343 	 */
1344 #ifdef FLUSH_BASE
1345 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1346 	map.virtual = FLUSH_BASE;
1347 	map.length = SZ_1M;
1348 	map.type = MT_CACHECLEAN;
1349 	create_mapping(&map);
1350 #endif
1351 #ifdef FLUSH_BASE_MINICACHE
1352 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1353 	map.virtual = FLUSH_BASE_MINICACHE;
1354 	map.length = SZ_1M;
1355 	map.type = MT_MINICLEAN;
1356 	create_mapping(&map);
1357 #endif
1358 
1359 	/*
1360 	 * Create a mapping for the machine vectors at the high-vectors
1361 	 * location (0xffff0000).  If we aren't using high-vectors, also
1362 	 * create a mapping at the low-vectors virtual address.
1363 	 */
1364 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1365 	map.virtual = 0xffff0000;
1366 	map.length = PAGE_SIZE;
1367 #ifdef CONFIG_KUSER_HELPERS
1368 	map.type = MT_HIGH_VECTORS;
1369 #else
1370 	map.type = MT_LOW_VECTORS;
1371 #endif
1372 	create_mapping(&map);
1373 
1374 	if (!vectors_high()) {
1375 		map.virtual = 0;
1376 		map.length = PAGE_SIZE * 2;
1377 		map.type = MT_LOW_VECTORS;
1378 		create_mapping(&map);
1379 	}
1380 
1381 	/* Now create a kernel read-only mapping */
1382 	map.pfn += 1;
1383 	map.virtual = 0xffff0000 + PAGE_SIZE;
1384 	map.length = PAGE_SIZE;
1385 	map.type = MT_LOW_VECTORS;
1386 	create_mapping(&map);
1387 
1388 	/*
1389 	 * Ask the machine support to map in the statically mapped devices.
1390 	 */
1391 	if (mdesc->map_io)
1392 		mdesc->map_io();
1393 	else
1394 		debug_ll_io_init();
1395 	fill_pmd_gaps();
1396 
1397 	/* Reserve fixed i/o space in VMALLOC region */
1398 	pci_reserve_io();
1399 
1400 	/*
1401 	 * Finally flush the caches and tlb to ensure that we're in a
1402 	 * consistent state wrt the writebuffer.  This also ensures that
1403 	 * any write-allocated cache lines in the vector page are written
1404 	 * back.  After this point, we can start to touch devices again.
1405 	 */
1406 	local_flush_tlb_all();
1407 	flush_cache_all();
1408 
1409 	/* Enable asynchronous aborts */
1410 	early_abt_enable();
1411 }
1412 
1413 static void __init kmap_init(void)
1414 {
1415 #ifdef CONFIG_HIGHMEM
1416 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1417 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1418 #endif
1419 
1420 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1421 			_PAGE_KERNEL_TABLE);
1422 }
1423 
1424 static void __init map_lowmem(void)
1425 {
1426 	struct memblock_region *reg;
1427 	phys_addr_t kernel_x_start = round_down(__pa(KERNEL_START), SECTION_SIZE);
1428 	phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1429 
1430 	/* Map all the lowmem memory banks. */
1431 	for_each_memblock(memory, reg) {
1432 		phys_addr_t start = reg->base;
1433 		phys_addr_t end = start + reg->size;
1434 		struct map_desc map;
1435 
1436 		if (memblock_is_nomap(reg))
1437 			continue;
1438 
1439 		if (end > arm_lowmem_limit)
1440 			end = arm_lowmem_limit;
1441 		if (start >= end)
1442 			break;
1443 
1444 		if (end < kernel_x_start) {
1445 			map.pfn = __phys_to_pfn(start);
1446 			map.virtual = __phys_to_virt(start);
1447 			map.length = end - start;
1448 			map.type = MT_MEMORY_RWX;
1449 
1450 			create_mapping(&map);
1451 		} else if (start >= kernel_x_end) {
1452 			map.pfn = __phys_to_pfn(start);
1453 			map.virtual = __phys_to_virt(start);
1454 			map.length = end - start;
1455 			map.type = MT_MEMORY_RW;
1456 
1457 			create_mapping(&map);
1458 		} else {
1459 			/* This better cover the entire kernel */
1460 			if (start < kernel_x_start) {
1461 				map.pfn = __phys_to_pfn(start);
1462 				map.virtual = __phys_to_virt(start);
1463 				map.length = kernel_x_start - start;
1464 				map.type = MT_MEMORY_RW;
1465 
1466 				create_mapping(&map);
1467 			}
1468 
1469 			map.pfn = __phys_to_pfn(kernel_x_start);
1470 			map.virtual = __phys_to_virt(kernel_x_start);
1471 			map.length = kernel_x_end - kernel_x_start;
1472 			map.type = MT_MEMORY_RWX;
1473 
1474 			create_mapping(&map);
1475 
1476 			if (kernel_x_end < end) {
1477 				map.pfn = __phys_to_pfn(kernel_x_end);
1478 				map.virtual = __phys_to_virt(kernel_x_end);
1479 				map.length = end - kernel_x_end;
1480 				map.type = MT_MEMORY_RW;
1481 
1482 				create_mapping(&map);
1483 			}
1484 		}
1485 	}
1486 }
1487 
1488 #ifdef CONFIG_ARM_PV_FIXUP
1489 extern unsigned long __atags_pointer;
1490 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata);
1491 pgtables_remap lpae_pgtables_remap_asm;
1492 
1493 /*
1494  * early_paging_init() recreates boot time page table setup, allowing machines
1495  * to switch over to a high (>4G) address space on LPAE systems
1496  */
1497 static void __init early_paging_init(const struct machine_desc *mdesc)
1498 {
1499 	pgtables_remap *lpae_pgtables_remap;
1500 	unsigned long pa_pgd;
1501 	unsigned int cr, ttbcr;
1502 	long long offset;
1503 	void *boot_data;
1504 
1505 	if (!mdesc->pv_fixup)
1506 		return;
1507 
1508 	offset = mdesc->pv_fixup();
1509 	if (offset == 0)
1510 		return;
1511 
1512 	/*
1513 	 * Get the address of the remap function in the 1:1 identity
1514 	 * mapping setup by the early page table assembly code.  We
1515 	 * must get this prior to the pv update.  The following barrier
1516 	 * ensures that this is complete before we fixup any P:V offsets.
1517 	 */
1518 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1519 	pa_pgd = __pa(swapper_pg_dir);
1520 	boot_data = __va(__atags_pointer);
1521 	barrier();
1522 
1523 	pr_info("Switching physical address space to 0x%08llx\n",
1524 		(u64)PHYS_OFFSET + offset);
1525 
1526 	/* Re-set the phys pfn offset, and the pv offset */
1527 	__pv_offset += offset;
1528 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1529 
1530 	/* Run the patch stub to update the constants */
1531 	fixup_pv_table(&__pv_table_begin,
1532 		(&__pv_table_end - &__pv_table_begin) << 2);
1533 
1534 	/*
1535 	 * We changing not only the virtual to physical mapping, but also
1536 	 * the physical addresses used to access memory.  We need to flush
1537 	 * all levels of cache in the system with caching disabled to
1538 	 * ensure that all data is written back, and nothing is prefetched
1539 	 * into the caches.  We also need to prevent the TLB walkers
1540 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1541 	 * specific.
1542 	 */
1543 	cr = get_cr();
1544 	set_cr(cr & ~(CR_I | CR_C));
1545 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1546 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1547 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1548 	flush_cache_all();
1549 
1550 	/*
1551 	 * Fixup the page tables - this must be in the idmap region as
1552 	 * we need to disable the MMU to do this safely, and hence it
1553 	 * needs to be assembly.  It's fairly simple, as we're using the
1554 	 * temporary tables setup by the initial assembly code.
1555 	 */
1556 	lpae_pgtables_remap(offset, pa_pgd, boot_data);
1557 
1558 	/* Re-enable the caches and cacheable TLB walks */
1559 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1560 	set_cr(cr);
1561 }
1562 
1563 #else
1564 
1565 static void __init early_paging_init(const struct machine_desc *mdesc)
1566 {
1567 	long long offset;
1568 
1569 	if (!mdesc->pv_fixup)
1570 		return;
1571 
1572 	offset = mdesc->pv_fixup();
1573 	if (offset == 0)
1574 		return;
1575 
1576 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1577 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1578 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1579 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1580 }
1581 
1582 #endif
1583 
1584 static void __init early_fixmap_shutdown(void)
1585 {
1586 	int i;
1587 	unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1);
1588 
1589 	pte_offset_fixmap = pte_offset_late_fixmap;
1590 	pmd_clear(fixmap_pmd(va));
1591 	local_flush_tlb_kernel_page(va);
1592 
1593 	for (i = 0; i < __end_of_permanent_fixed_addresses; i++) {
1594 		pte_t *pte;
1595 		struct map_desc map;
1596 
1597 		map.virtual = fix_to_virt(i);
1598 		pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual);
1599 
1600 		/* Only i/o device mappings are supported ATM */
1601 		if (pte_none(*pte) ||
1602 		    (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED)
1603 			continue;
1604 
1605 		map.pfn = pte_pfn(*pte);
1606 		map.type = MT_DEVICE;
1607 		map.length = PAGE_SIZE;
1608 
1609 		create_mapping(&map);
1610 	}
1611 }
1612 
1613 /*
1614  * paging_init() sets up the page tables, initialises the zone memory
1615  * maps, and sets up the zero page, bad page and bad page tables.
1616  */
1617 void __init paging_init(const struct machine_desc *mdesc)
1618 {
1619 	void *zero_page;
1620 
1621 	prepare_page_table();
1622 	map_lowmem();
1623 	memblock_set_current_limit(arm_lowmem_limit);
1624 	dma_contiguous_remap();
1625 	early_fixmap_shutdown();
1626 	devicemaps_init(mdesc);
1627 	kmap_init();
1628 	tcm_init();
1629 
1630 	top_pmd = pmd_off_k(0xffff0000);
1631 
1632 	/* allocate the zero page. */
1633 	zero_page = early_alloc(PAGE_SIZE);
1634 
1635 	bootmem_init();
1636 
1637 	empty_zero_page = virt_to_page(zero_page);
1638 	__flush_dcache_page(NULL, empty_zero_page);
1639 
1640 	/* Compute the virt/idmap offset, mostly for the sake of KVM */
1641 	kimage_voffset = (unsigned long)&kimage_voffset - virt_to_idmap(&kimage_voffset);
1642 }
1643 
1644 void __init early_mm_init(const struct machine_desc *mdesc)
1645 {
1646 	build_mem_type_table();
1647 	early_paging_init(mdesc);
1648 }
1649 
1650 void set_pte_at(struct mm_struct *mm, unsigned long addr,
1651 			      pte_t *ptep, pte_t pteval)
1652 {
1653 	unsigned long ext = 0;
1654 
1655 	if (addr < TASK_SIZE && pte_valid_user(pteval)) {
1656 		if (!pte_special(pteval))
1657 			__sync_icache_dcache(pteval);
1658 		ext |= PTE_EXT_NG;
1659 	}
1660 
1661 	set_pte_ext(ptep, pteval, ext);
1662 }
1663