xref: /linux/arch/arm/mm/mmu.c (revision c0e297dc61f8d4453e07afbea1fa8d0e67cd4a34)
1 /*
2  *  linux/arch/arm/mm/mmu.c
3  *
4  *  Copyright (C) 1995-2005 Russell King
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  */
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/mman.h>
15 #include <linux/nodemask.h>
16 #include <linux/memblock.h>
17 #include <linux/fs.h>
18 #include <linux/vmalloc.h>
19 #include <linux/sizes.h>
20 
21 #include <asm/cp15.h>
22 #include <asm/cputype.h>
23 #include <asm/sections.h>
24 #include <asm/cachetype.h>
25 #include <asm/fixmap.h>
26 #include <asm/sections.h>
27 #include <asm/setup.h>
28 #include <asm/smp_plat.h>
29 #include <asm/tlb.h>
30 #include <asm/highmem.h>
31 #include <asm/system_info.h>
32 #include <asm/traps.h>
33 #include <asm/procinfo.h>
34 #include <asm/memory.h>
35 
36 #include <asm/mach/arch.h>
37 #include <asm/mach/map.h>
38 #include <asm/mach/pci.h>
39 #include <asm/fixmap.h>
40 
41 #include "mm.h"
42 #include "tcm.h"
43 
44 /*
45  * empty_zero_page is a special page that is used for
46  * zero-initialized data and COW.
47  */
48 struct page *empty_zero_page;
49 EXPORT_SYMBOL(empty_zero_page);
50 
51 /*
52  * The pmd table for the upper-most set of pages.
53  */
54 pmd_t *top_pmd;
55 
56 pmdval_t user_pmd_table = _PAGE_USER_TABLE;
57 
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
63 
64 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
65 static unsigned int ecc_mask __initdata = 0;
66 pgprot_t pgprot_user;
67 pgprot_t pgprot_kernel;
68 pgprot_t pgprot_hyp_device;
69 pgprot_t pgprot_s2;
70 pgprot_t pgprot_s2_device;
71 
72 EXPORT_SYMBOL(pgprot_user);
73 EXPORT_SYMBOL(pgprot_kernel);
74 
75 struct cachepolicy {
76 	const char	policy[16];
77 	unsigned int	cr_mask;
78 	pmdval_t	pmd;
79 	pteval_t	pte;
80 	pteval_t	pte_s2;
81 };
82 
83 #ifdef CONFIG_ARM_LPAE
84 #define s2_policy(policy)	policy
85 #else
86 #define s2_policy(policy)	0
87 #endif
88 
89 static struct cachepolicy cache_policies[] __initdata = {
90 	{
91 		.policy		= "uncached",
92 		.cr_mask	= CR_W|CR_C,
93 		.pmd		= PMD_SECT_UNCACHED,
94 		.pte		= L_PTE_MT_UNCACHED,
95 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
96 	}, {
97 		.policy		= "buffered",
98 		.cr_mask	= CR_C,
99 		.pmd		= PMD_SECT_BUFFERED,
100 		.pte		= L_PTE_MT_BUFFERABLE,
101 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
102 	}, {
103 		.policy		= "writethrough",
104 		.cr_mask	= 0,
105 		.pmd		= PMD_SECT_WT,
106 		.pte		= L_PTE_MT_WRITETHROUGH,
107 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITETHROUGH),
108 	}, {
109 		.policy		= "writeback",
110 		.cr_mask	= 0,
111 		.pmd		= PMD_SECT_WB,
112 		.pte		= L_PTE_MT_WRITEBACK,
113 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
114 	}, {
115 		.policy		= "writealloc",
116 		.cr_mask	= 0,
117 		.pmd		= PMD_SECT_WBWA,
118 		.pte		= L_PTE_MT_WRITEALLOC,
119 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
120 	}
121 };
122 
123 #ifdef CONFIG_CPU_CP15
124 static unsigned long initial_pmd_value __initdata = 0;
125 
126 /*
127  * Initialise the cache_policy variable with the initial state specified
128  * via the "pmd" value.  This is used to ensure that on ARMv6 and later,
129  * the C code sets the page tables up with the same policy as the head
130  * assembly code, which avoids an illegal state where the TLBs can get
131  * confused.  See comments in early_cachepolicy() for more information.
132  */
133 void __init init_default_cache_policy(unsigned long pmd)
134 {
135 	int i;
136 
137 	initial_pmd_value = pmd;
138 
139 	pmd &= PMD_SECT_TEX(1) | PMD_SECT_BUFFERABLE | PMD_SECT_CACHEABLE;
140 
141 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
142 		if (cache_policies[i].pmd == pmd) {
143 			cachepolicy = i;
144 			break;
145 		}
146 
147 	if (i == ARRAY_SIZE(cache_policies))
148 		pr_err("ERROR: could not find cache policy\n");
149 }
150 
151 /*
152  * These are useful for identifying cache coherency problems by allowing
153  * the cache or the cache and writebuffer to be turned off.  (Note: the
154  * write buffer should not be on and the cache off).
155  */
156 static int __init early_cachepolicy(char *p)
157 {
158 	int i, selected = -1;
159 
160 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
161 		int len = strlen(cache_policies[i].policy);
162 
163 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
164 			selected = i;
165 			break;
166 		}
167 	}
168 
169 	if (selected == -1)
170 		pr_err("ERROR: unknown or unsupported cache policy\n");
171 
172 	/*
173 	 * This restriction is partly to do with the way we boot; it is
174 	 * unpredictable to have memory mapped using two different sets of
175 	 * memory attributes (shared, type, and cache attribs).  We can not
176 	 * change these attributes once the initial assembly has setup the
177 	 * page tables.
178 	 */
179 	if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) {
180 		pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n",
181 			cache_policies[cachepolicy].policy);
182 		return 0;
183 	}
184 
185 	if (selected != cachepolicy) {
186 		unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
187 		cachepolicy = selected;
188 		flush_cache_all();
189 		set_cr(cr);
190 	}
191 	return 0;
192 }
193 early_param("cachepolicy", early_cachepolicy);
194 
195 static int __init early_nocache(char *__unused)
196 {
197 	char *p = "buffered";
198 	pr_warn("nocache is deprecated; use cachepolicy=%s\n", p);
199 	early_cachepolicy(p);
200 	return 0;
201 }
202 early_param("nocache", early_nocache);
203 
204 static int __init early_nowrite(char *__unused)
205 {
206 	char *p = "uncached";
207 	pr_warn("nowb is deprecated; use cachepolicy=%s\n", p);
208 	early_cachepolicy(p);
209 	return 0;
210 }
211 early_param("nowb", early_nowrite);
212 
213 #ifndef CONFIG_ARM_LPAE
214 static int __init early_ecc(char *p)
215 {
216 	if (memcmp(p, "on", 2) == 0)
217 		ecc_mask = PMD_PROTECTION;
218 	else if (memcmp(p, "off", 3) == 0)
219 		ecc_mask = 0;
220 	return 0;
221 }
222 early_param("ecc", early_ecc);
223 #endif
224 
225 #else /* ifdef CONFIG_CPU_CP15 */
226 
227 static int __init early_cachepolicy(char *p)
228 {
229 	pr_warn("cachepolicy kernel parameter not supported without cp15\n");
230 }
231 early_param("cachepolicy", early_cachepolicy);
232 
233 static int __init noalign_setup(char *__unused)
234 {
235 	pr_warn("noalign kernel parameter not supported without cp15\n");
236 }
237 __setup("noalign", noalign_setup);
238 
239 #endif /* ifdef CONFIG_CPU_CP15 / else */
240 
241 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
242 #define PROT_PTE_S2_DEVICE	PROT_PTE_DEVICE
243 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
244 
245 static struct mem_type mem_types[] = {
246 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
247 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
248 				  L_PTE_SHARED,
249 		.prot_pte_s2	= s2_policy(PROT_PTE_S2_DEVICE) |
250 				  s2_policy(L_PTE_S2_MT_DEV_SHARED) |
251 				  L_PTE_SHARED,
252 		.prot_l1	= PMD_TYPE_TABLE,
253 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
254 		.domain		= DOMAIN_IO,
255 	},
256 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
257 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
258 		.prot_l1	= PMD_TYPE_TABLE,
259 		.prot_sect	= PROT_SECT_DEVICE,
260 		.domain		= DOMAIN_IO,
261 	},
262 	[MT_DEVICE_CACHED] = {	  /* ioremap_cached */
263 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
264 		.prot_l1	= PMD_TYPE_TABLE,
265 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
266 		.domain		= DOMAIN_IO,
267 	},
268 	[MT_DEVICE_WC] = {	/* ioremap_wc */
269 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
270 		.prot_l1	= PMD_TYPE_TABLE,
271 		.prot_sect	= PROT_SECT_DEVICE,
272 		.domain		= DOMAIN_IO,
273 	},
274 	[MT_UNCACHED] = {
275 		.prot_pte	= PROT_PTE_DEVICE,
276 		.prot_l1	= PMD_TYPE_TABLE,
277 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
278 		.domain		= DOMAIN_IO,
279 	},
280 	[MT_CACHECLEAN] = {
281 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
282 		.domain    = DOMAIN_KERNEL,
283 	},
284 #ifndef CONFIG_ARM_LPAE
285 	[MT_MINICLEAN] = {
286 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
287 		.domain    = DOMAIN_KERNEL,
288 	},
289 #endif
290 	[MT_LOW_VECTORS] = {
291 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
292 				L_PTE_RDONLY,
293 		.prot_l1   = PMD_TYPE_TABLE,
294 		.domain    = DOMAIN_USER,
295 	},
296 	[MT_HIGH_VECTORS] = {
297 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
298 				L_PTE_USER | L_PTE_RDONLY,
299 		.prot_l1   = PMD_TYPE_TABLE,
300 		.domain    = DOMAIN_USER,
301 	},
302 	[MT_MEMORY_RWX] = {
303 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
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] = {
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_AP_WRITE,
313 		.domain    = DOMAIN_KERNEL,
314 	},
315 	[MT_ROM] = {
316 		.prot_sect = PMD_TYPE_SECT,
317 		.domain    = DOMAIN_KERNEL,
318 	},
319 	[MT_MEMORY_RWX_NONCACHED] = {
320 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
321 				L_PTE_MT_BUFFERABLE,
322 		.prot_l1   = PMD_TYPE_TABLE,
323 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
324 		.domain    = DOMAIN_KERNEL,
325 	},
326 	[MT_MEMORY_RW_DTCM] = {
327 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
328 				L_PTE_XN,
329 		.prot_l1   = PMD_TYPE_TABLE,
330 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
331 		.domain    = DOMAIN_KERNEL,
332 	},
333 	[MT_MEMORY_RWX_ITCM] = {
334 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
335 		.prot_l1   = PMD_TYPE_TABLE,
336 		.domain    = DOMAIN_KERNEL,
337 	},
338 	[MT_MEMORY_RW_SO] = {
339 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
340 				L_PTE_MT_UNCACHED | L_PTE_XN,
341 		.prot_l1   = PMD_TYPE_TABLE,
342 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
343 				PMD_SECT_UNCACHED | PMD_SECT_XN,
344 		.domain    = DOMAIN_KERNEL,
345 	},
346 	[MT_MEMORY_DMA_READY] = {
347 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
348 				L_PTE_XN,
349 		.prot_l1   = PMD_TYPE_TABLE,
350 		.domain    = DOMAIN_KERNEL,
351 	},
352 };
353 
354 const struct mem_type *get_mem_type(unsigned int type)
355 {
356 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
357 }
358 EXPORT_SYMBOL(get_mem_type);
359 
360 /*
361  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
362  * As a result, this can only be called with preemption disabled, as under
363  * stop_machine().
364  */
365 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
366 {
367 	unsigned long vaddr = __fix_to_virt(idx);
368 	pte_t *pte = pte_offset_kernel(pmd_off_k(vaddr), vaddr);
369 
370 	/* Make sure fixmap region does not exceed available allocation. */
371 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
372 		     FIXADDR_END);
373 	BUG_ON(idx >= __end_of_fixed_addresses);
374 
375 	if (pgprot_val(prot))
376 		set_pte_at(NULL, vaddr, pte,
377 			pfn_pte(phys >> PAGE_SHIFT, prot));
378 	else
379 		pte_clear(NULL, vaddr, pte);
380 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
381 }
382 
383 /*
384  * Adjust the PMD section entries according to the CPU in use.
385  */
386 static void __init build_mem_type_table(void)
387 {
388 	struct cachepolicy *cp;
389 	unsigned int cr = get_cr();
390 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
391 	pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
392 	int cpu_arch = cpu_architecture();
393 	int i;
394 
395 	if (cpu_arch < CPU_ARCH_ARMv6) {
396 #if defined(CONFIG_CPU_DCACHE_DISABLE)
397 		if (cachepolicy > CPOLICY_BUFFERED)
398 			cachepolicy = CPOLICY_BUFFERED;
399 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
400 		if (cachepolicy > CPOLICY_WRITETHROUGH)
401 			cachepolicy = CPOLICY_WRITETHROUGH;
402 #endif
403 	}
404 	if (cpu_arch < CPU_ARCH_ARMv5) {
405 		if (cachepolicy >= CPOLICY_WRITEALLOC)
406 			cachepolicy = CPOLICY_WRITEBACK;
407 		ecc_mask = 0;
408 	}
409 
410 	if (is_smp()) {
411 		if (cachepolicy != CPOLICY_WRITEALLOC) {
412 			pr_warn("Forcing write-allocate cache policy for SMP\n");
413 			cachepolicy = CPOLICY_WRITEALLOC;
414 		}
415 		if (!(initial_pmd_value & PMD_SECT_S)) {
416 			pr_warn("Forcing shared mappings for SMP\n");
417 			initial_pmd_value |= PMD_SECT_S;
418 		}
419 	}
420 
421 	/*
422 	 * Strip out features not present on earlier architectures.
423 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
424 	 * without extended page tables don't have the 'Shared' bit.
425 	 */
426 	if (cpu_arch < CPU_ARCH_ARMv5)
427 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
428 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
429 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
430 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
431 			mem_types[i].prot_sect &= ~PMD_SECT_S;
432 
433 	/*
434 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
435 	 * "update-able on write" bit on ARM610).  However, Xscale and
436 	 * Xscale3 require this bit to be cleared.
437 	 */
438 	if (cpu_is_xscale() || cpu_is_xsc3()) {
439 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
440 			mem_types[i].prot_sect &= ~PMD_BIT4;
441 			mem_types[i].prot_l1 &= ~PMD_BIT4;
442 		}
443 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
444 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
445 			if (mem_types[i].prot_l1)
446 				mem_types[i].prot_l1 |= PMD_BIT4;
447 			if (mem_types[i].prot_sect)
448 				mem_types[i].prot_sect |= PMD_BIT4;
449 		}
450 	}
451 
452 	/*
453 	 * Mark the device areas according to the CPU/architecture.
454 	 */
455 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
456 		if (!cpu_is_xsc3()) {
457 			/*
458 			 * Mark device regions on ARMv6+ as execute-never
459 			 * to prevent speculative instruction fetches.
460 			 */
461 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
462 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
463 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
464 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
465 
466 			/* Also setup NX memory mapping */
467 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
468 		}
469 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
470 			/*
471 			 * For ARMv7 with TEX remapping,
472 			 * - shared device is SXCB=1100
473 			 * - nonshared device is SXCB=0100
474 			 * - write combine device mem is SXCB=0001
475 			 * (Uncached Normal memory)
476 			 */
477 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
478 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
479 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
480 		} else if (cpu_is_xsc3()) {
481 			/*
482 			 * For Xscale3,
483 			 * - shared device is TEXCB=00101
484 			 * - nonshared device is TEXCB=01000
485 			 * - write combine device mem is TEXCB=00100
486 			 * (Inner/Outer Uncacheable in xsc3 parlance)
487 			 */
488 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
489 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
490 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
491 		} else {
492 			/*
493 			 * For ARMv6 and ARMv7 without TEX remapping,
494 			 * - shared device is TEXCB=00001
495 			 * - nonshared device is TEXCB=01000
496 			 * - write combine device mem is TEXCB=00100
497 			 * (Uncached Normal in ARMv6 parlance).
498 			 */
499 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
500 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
501 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
502 		}
503 	} else {
504 		/*
505 		 * On others, write combining is "Uncached/Buffered"
506 		 */
507 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
508 	}
509 
510 	/*
511 	 * Now deal with the memory-type mappings
512 	 */
513 	cp = &cache_policies[cachepolicy];
514 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
515 	s2_pgprot = cp->pte_s2;
516 	hyp_device_pgprot = mem_types[MT_DEVICE].prot_pte;
517 	s2_device_pgprot = mem_types[MT_DEVICE].prot_pte_s2;
518 
519 #ifndef CONFIG_ARM_LPAE
520 	/*
521 	 * We don't use domains on ARMv6 (since this causes problems with
522 	 * v6/v7 kernels), so we must use a separate memory type for user
523 	 * r/o, kernel r/w to map the vectors page.
524 	 */
525 	if (cpu_arch == CPU_ARCH_ARMv6)
526 		vecs_pgprot |= L_PTE_MT_VECTORS;
527 
528 	/*
529 	 * Check is it with support for the PXN bit
530 	 * in the Short-descriptor translation table format descriptors.
531 	 */
532 	if (cpu_arch == CPU_ARCH_ARMv7 &&
533 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) == 4) {
534 		user_pmd_table |= PMD_PXNTABLE;
535 	}
536 #endif
537 
538 	/*
539 	 * ARMv6 and above have extended page tables.
540 	 */
541 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
542 #ifndef CONFIG_ARM_LPAE
543 		/*
544 		 * Mark cache clean areas and XIP ROM read only
545 		 * from SVC mode and no access from userspace.
546 		 */
547 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
548 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
549 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
550 #endif
551 
552 		/*
553 		 * If the initial page tables were created with the S bit
554 		 * set, then we need to do the same here for the same
555 		 * reasons given in early_cachepolicy().
556 		 */
557 		if (initial_pmd_value & PMD_SECT_S) {
558 			user_pgprot |= L_PTE_SHARED;
559 			kern_pgprot |= L_PTE_SHARED;
560 			vecs_pgprot |= L_PTE_SHARED;
561 			s2_pgprot |= L_PTE_SHARED;
562 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
563 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
564 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
565 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
566 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
567 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
568 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
569 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
570 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
571 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
572 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
573 		}
574 	}
575 
576 	/*
577 	 * Non-cacheable Normal - intended for memory areas that must
578 	 * not cause dirty cache line writebacks when used
579 	 */
580 	if (cpu_arch >= CPU_ARCH_ARMv6) {
581 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
582 			/* Non-cacheable Normal is XCB = 001 */
583 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
584 				PMD_SECT_BUFFERED;
585 		} else {
586 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
587 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
588 				PMD_SECT_TEX(1);
589 		}
590 	} else {
591 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
592 	}
593 
594 #ifdef CONFIG_ARM_LPAE
595 	/*
596 	 * Do not generate access flag faults for the kernel mappings.
597 	 */
598 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
599 		mem_types[i].prot_pte |= PTE_EXT_AF;
600 		if (mem_types[i].prot_sect)
601 			mem_types[i].prot_sect |= PMD_SECT_AF;
602 	}
603 	kern_pgprot |= PTE_EXT_AF;
604 	vecs_pgprot |= PTE_EXT_AF;
605 
606 	/*
607 	 * Set PXN for user mappings
608 	 */
609 	user_pgprot |= PTE_EXT_PXN;
610 #endif
611 
612 	for (i = 0; i < 16; i++) {
613 		pteval_t v = pgprot_val(protection_map[i]);
614 		protection_map[i] = __pgprot(v | user_pgprot);
615 	}
616 
617 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
618 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
619 
620 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
621 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
622 				 L_PTE_DIRTY | kern_pgprot);
623 	pgprot_s2  = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
624 	pgprot_s2_device  = __pgprot(s2_device_pgprot);
625 	pgprot_hyp_device  = __pgprot(hyp_device_pgprot);
626 
627 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
628 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
629 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
630 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
631 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
632 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
633 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
634 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
635 	mem_types[MT_ROM].prot_sect |= cp->pmd;
636 
637 	switch (cp->pmd) {
638 	case PMD_SECT_WT:
639 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
640 		break;
641 	case PMD_SECT_WB:
642 	case PMD_SECT_WBWA:
643 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
644 		break;
645 	}
646 	pr_info("Memory policy: %sData cache %s\n",
647 		ecc_mask ? "ECC enabled, " : "", cp->policy);
648 
649 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
650 		struct mem_type *t = &mem_types[i];
651 		if (t->prot_l1)
652 			t->prot_l1 |= PMD_DOMAIN(t->domain);
653 		if (t->prot_sect)
654 			t->prot_sect |= PMD_DOMAIN(t->domain);
655 	}
656 }
657 
658 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
659 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
660 			      unsigned long size, pgprot_t vma_prot)
661 {
662 	if (!pfn_valid(pfn))
663 		return pgprot_noncached(vma_prot);
664 	else if (file->f_flags & O_SYNC)
665 		return pgprot_writecombine(vma_prot);
666 	return vma_prot;
667 }
668 EXPORT_SYMBOL(phys_mem_access_prot);
669 #endif
670 
671 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
672 
673 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
674 {
675 	void *ptr = __va(memblock_alloc(sz, align));
676 	memset(ptr, 0, sz);
677 	return ptr;
678 }
679 
680 static void __init *early_alloc(unsigned long sz)
681 {
682 	return early_alloc_aligned(sz, sz);
683 }
684 
685 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
686 {
687 	if (pmd_none(*pmd)) {
688 		pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
689 		__pmd_populate(pmd, __pa(pte), prot);
690 	}
691 	BUG_ON(pmd_bad(*pmd));
692 	return pte_offset_kernel(pmd, addr);
693 }
694 
695 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
696 				  unsigned long end, unsigned long pfn,
697 				  const struct mem_type *type)
698 {
699 	pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
700 	do {
701 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
702 		pfn++;
703 	} while (pte++, addr += PAGE_SIZE, addr != end);
704 }
705 
706 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
707 			unsigned long end, phys_addr_t phys,
708 			const struct mem_type *type)
709 {
710 	pmd_t *p = pmd;
711 
712 #ifndef CONFIG_ARM_LPAE
713 	/*
714 	 * In classic MMU format, puds and pmds are folded in to
715 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
716 	 * group of L1 entries making up one logical pointer to
717 	 * an L2 table (2MB), where as PMDs refer to the individual
718 	 * L1 entries (1MB). Hence increment to get the correct
719 	 * offset for odd 1MB sections.
720 	 * (See arch/arm/include/asm/pgtable-2level.h)
721 	 */
722 	if (addr & SECTION_SIZE)
723 		pmd++;
724 #endif
725 	do {
726 		*pmd = __pmd(phys | type->prot_sect);
727 		phys += SECTION_SIZE;
728 	} while (pmd++, addr += SECTION_SIZE, addr != end);
729 
730 	flush_pmd_entry(p);
731 }
732 
733 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
734 				      unsigned long end, phys_addr_t phys,
735 				      const struct mem_type *type)
736 {
737 	pmd_t *pmd = pmd_offset(pud, addr);
738 	unsigned long next;
739 
740 	do {
741 		/*
742 		 * With LPAE, we must loop over to map
743 		 * all the pmds for the given range.
744 		 */
745 		next = pmd_addr_end(addr, end);
746 
747 		/*
748 		 * Try a section mapping - addr, next and phys must all be
749 		 * aligned to a section boundary.
750 		 */
751 		if (type->prot_sect &&
752 				((addr | next | phys) & ~SECTION_MASK) == 0) {
753 			__map_init_section(pmd, addr, next, phys, type);
754 		} else {
755 			alloc_init_pte(pmd, addr, next,
756 						__phys_to_pfn(phys), type);
757 		}
758 
759 		phys += next - addr;
760 
761 	} while (pmd++, addr = next, addr != end);
762 }
763 
764 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
765 				  unsigned long end, phys_addr_t phys,
766 				  const struct mem_type *type)
767 {
768 	pud_t *pud = pud_offset(pgd, addr);
769 	unsigned long next;
770 
771 	do {
772 		next = pud_addr_end(addr, end);
773 		alloc_init_pmd(pud, addr, next, phys, type);
774 		phys += next - addr;
775 	} while (pud++, addr = next, addr != end);
776 }
777 
778 #ifndef CONFIG_ARM_LPAE
779 static void __init create_36bit_mapping(struct map_desc *md,
780 					const struct mem_type *type)
781 {
782 	unsigned long addr, length, end;
783 	phys_addr_t phys;
784 	pgd_t *pgd;
785 
786 	addr = md->virtual;
787 	phys = __pfn_to_phys(md->pfn);
788 	length = PAGE_ALIGN(md->length);
789 
790 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
791 		pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
792 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
793 		return;
794 	}
795 
796 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
797 	 *	Since domain assignments can in fact be arbitrary, the
798 	 *	'domain == 0' check below is required to insure that ARMv6
799 	 *	supersections are only allocated for domain 0 regardless
800 	 *	of the actual domain assignments in use.
801 	 */
802 	if (type->domain) {
803 		pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
804 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
805 		return;
806 	}
807 
808 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
809 		pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
810 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
811 		return;
812 	}
813 
814 	/*
815 	 * Shift bits [35:32] of address into bits [23:20] of PMD
816 	 * (See ARMv6 spec).
817 	 */
818 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
819 
820 	pgd = pgd_offset_k(addr);
821 	end = addr + length;
822 	do {
823 		pud_t *pud = pud_offset(pgd, addr);
824 		pmd_t *pmd = pmd_offset(pud, addr);
825 		int i;
826 
827 		for (i = 0; i < 16; i++)
828 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
829 
830 		addr += SUPERSECTION_SIZE;
831 		phys += SUPERSECTION_SIZE;
832 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
833 	} while (addr != end);
834 }
835 #endif	/* !CONFIG_ARM_LPAE */
836 
837 /*
838  * Create the page directory entries and any necessary
839  * page tables for the mapping specified by `md'.  We
840  * are able to cope here with varying sizes and address
841  * offsets, and we take full advantage of sections and
842  * supersections.
843  */
844 static void __init create_mapping(struct map_desc *md)
845 {
846 	unsigned long addr, length, end;
847 	phys_addr_t phys;
848 	const struct mem_type *type;
849 	pgd_t *pgd;
850 
851 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
852 		pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
853 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
854 		return;
855 	}
856 
857 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
858 	    md->virtual >= PAGE_OFFSET &&
859 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
860 		pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
861 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
862 	}
863 
864 	type = &mem_types[md->type];
865 
866 #ifndef CONFIG_ARM_LPAE
867 	/*
868 	 * Catch 36-bit addresses
869 	 */
870 	if (md->pfn >= 0x100000) {
871 		create_36bit_mapping(md, type);
872 		return;
873 	}
874 #endif
875 
876 	addr = md->virtual & PAGE_MASK;
877 	phys = __pfn_to_phys(md->pfn);
878 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
879 
880 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
881 		pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
882 			(long long)__pfn_to_phys(md->pfn), addr);
883 		return;
884 	}
885 
886 	pgd = pgd_offset_k(addr);
887 	end = addr + length;
888 	do {
889 		unsigned long next = pgd_addr_end(addr, end);
890 
891 		alloc_init_pud(pgd, addr, next, phys, type);
892 
893 		phys += next - addr;
894 		addr = next;
895 	} while (pgd++, addr != end);
896 }
897 
898 /*
899  * Create the architecture specific mappings
900  */
901 void __init iotable_init(struct map_desc *io_desc, int nr)
902 {
903 	struct map_desc *md;
904 	struct vm_struct *vm;
905 	struct static_vm *svm;
906 
907 	if (!nr)
908 		return;
909 
910 	svm = early_alloc_aligned(sizeof(*svm) * nr, __alignof__(*svm));
911 
912 	for (md = io_desc; nr; md++, nr--) {
913 		create_mapping(md);
914 
915 		vm = &svm->vm;
916 		vm->addr = (void *)(md->virtual & PAGE_MASK);
917 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
918 		vm->phys_addr = __pfn_to_phys(md->pfn);
919 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
920 		vm->flags |= VM_ARM_MTYPE(md->type);
921 		vm->caller = iotable_init;
922 		add_static_vm_early(svm++);
923 	}
924 }
925 
926 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
927 				  void *caller)
928 {
929 	struct vm_struct *vm;
930 	struct static_vm *svm;
931 
932 	svm = early_alloc_aligned(sizeof(*svm), __alignof__(*svm));
933 
934 	vm = &svm->vm;
935 	vm->addr = (void *)addr;
936 	vm->size = size;
937 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
938 	vm->caller = caller;
939 	add_static_vm_early(svm);
940 }
941 
942 #ifndef CONFIG_ARM_LPAE
943 
944 /*
945  * The Linux PMD is made of two consecutive section entries covering 2MB
946  * (see definition in include/asm/pgtable-2level.h).  However a call to
947  * create_mapping() may optimize static mappings by using individual
948  * 1MB section mappings.  This leaves the actual PMD potentially half
949  * initialized if the top or bottom section entry isn't used, leaving it
950  * open to problems if a subsequent ioremap() or vmalloc() tries to use
951  * the virtual space left free by that unused section entry.
952  *
953  * Let's avoid the issue by inserting dummy vm entries covering the unused
954  * PMD halves once the static mappings are in place.
955  */
956 
957 static void __init pmd_empty_section_gap(unsigned long addr)
958 {
959 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
960 }
961 
962 static void __init fill_pmd_gaps(void)
963 {
964 	struct static_vm *svm;
965 	struct vm_struct *vm;
966 	unsigned long addr, next = 0;
967 	pmd_t *pmd;
968 
969 	list_for_each_entry(svm, &static_vmlist, list) {
970 		vm = &svm->vm;
971 		addr = (unsigned long)vm->addr;
972 		if (addr < next)
973 			continue;
974 
975 		/*
976 		 * Check if this vm starts on an odd section boundary.
977 		 * If so and the first section entry for this PMD is free
978 		 * then we block the corresponding virtual address.
979 		 */
980 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
981 			pmd = pmd_off_k(addr);
982 			if (pmd_none(*pmd))
983 				pmd_empty_section_gap(addr & PMD_MASK);
984 		}
985 
986 		/*
987 		 * Then check if this vm ends on an odd section boundary.
988 		 * If so and the second section entry for this PMD is empty
989 		 * then we block the corresponding virtual address.
990 		 */
991 		addr += vm->size;
992 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
993 			pmd = pmd_off_k(addr) + 1;
994 			if (pmd_none(*pmd))
995 				pmd_empty_section_gap(addr);
996 		}
997 
998 		/* no need to look at any vm entry until we hit the next PMD */
999 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1000 	}
1001 }
1002 
1003 #else
1004 #define fill_pmd_gaps() do { } while (0)
1005 #endif
1006 
1007 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1008 static void __init pci_reserve_io(void)
1009 {
1010 	struct static_vm *svm;
1011 
1012 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1013 	if (svm)
1014 		return;
1015 
1016 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1017 }
1018 #else
1019 #define pci_reserve_io() do { } while (0)
1020 #endif
1021 
1022 #ifdef CONFIG_DEBUG_LL
1023 void __init debug_ll_io_init(void)
1024 {
1025 	struct map_desc map;
1026 
1027 	debug_ll_addr(&map.pfn, &map.virtual);
1028 	if (!map.pfn || !map.virtual)
1029 		return;
1030 	map.pfn = __phys_to_pfn(map.pfn);
1031 	map.virtual &= PAGE_MASK;
1032 	map.length = PAGE_SIZE;
1033 	map.type = MT_DEVICE;
1034 	iotable_init(&map, 1);
1035 }
1036 #endif
1037 
1038 static void * __initdata vmalloc_min =
1039 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1040 
1041 /*
1042  * vmalloc=size forces the vmalloc area to be exactly 'size'
1043  * bytes. This can be used to increase (or decrease) the vmalloc
1044  * area - the default is 240m.
1045  */
1046 static int __init early_vmalloc(char *arg)
1047 {
1048 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1049 
1050 	if (vmalloc_reserve < SZ_16M) {
1051 		vmalloc_reserve = SZ_16M;
1052 		pr_warn("vmalloc area too small, limiting to %luMB\n",
1053 			vmalloc_reserve >> 20);
1054 	}
1055 
1056 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1057 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1058 		pr_warn("vmalloc area is too big, limiting to %luMB\n",
1059 			vmalloc_reserve >> 20);
1060 	}
1061 
1062 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1063 	return 0;
1064 }
1065 early_param("vmalloc", early_vmalloc);
1066 
1067 phys_addr_t arm_lowmem_limit __initdata = 0;
1068 
1069 void __init sanity_check_meminfo(void)
1070 {
1071 	phys_addr_t memblock_limit = 0;
1072 	int highmem = 0;
1073 	phys_addr_t vmalloc_limit = __pa(vmalloc_min - 1) + 1;
1074 	struct memblock_region *reg;
1075 	bool should_use_highmem = false;
1076 
1077 	for_each_memblock(memory, reg) {
1078 		phys_addr_t block_start = reg->base;
1079 		phys_addr_t block_end = reg->base + reg->size;
1080 		phys_addr_t size_limit = reg->size;
1081 
1082 		if (reg->base >= vmalloc_limit)
1083 			highmem = 1;
1084 		else
1085 			size_limit = vmalloc_limit - reg->base;
1086 
1087 
1088 		if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1089 
1090 			if (highmem) {
1091 				pr_notice("Ignoring RAM at %pa-%pa (!CONFIG_HIGHMEM)\n",
1092 					  &block_start, &block_end);
1093 				memblock_remove(reg->base, reg->size);
1094 				should_use_highmem = true;
1095 				continue;
1096 			}
1097 
1098 			if (reg->size > size_limit) {
1099 				phys_addr_t overlap_size = reg->size - size_limit;
1100 
1101 				pr_notice("Truncating RAM at %pa-%pa to -%pa",
1102 					  &block_start, &block_end, &vmalloc_limit);
1103 				memblock_remove(vmalloc_limit, overlap_size);
1104 				block_end = vmalloc_limit;
1105 				should_use_highmem = true;
1106 			}
1107 		}
1108 
1109 		if (!highmem) {
1110 			if (block_end > arm_lowmem_limit) {
1111 				if (reg->size > size_limit)
1112 					arm_lowmem_limit = vmalloc_limit;
1113 				else
1114 					arm_lowmem_limit = block_end;
1115 			}
1116 
1117 			/*
1118 			 * Find the first non-pmd-aligned page, and point
1119 			 * memblock_limit at it. This relies on rounding the
1120 			 * limit down to be pmd-aligned, which happens at the
1121 			 * end of this function.
1122 			 *
1123 			 * With this algorithm, the start or end of almost any
1124 			 * bank can be non-pmd-aligned. The only exception is
1125 			 * that the start of the bank 0 must be section-
1126 			 * aligned, since otherwise memory would need to be
1127 			 * allocated when mapping the start of bank 0, which
1128 			 * occurs before any free memory is mapped.
1129 			 */
1130 			if (!memblock_limit) {
1131 				if (!IS_ALIGNED(block_start, PMD_SIZE))
1132 					memblock_limit = block_start;
1133 				else if (!IS_ALIGNED(block_end, PMD_SIZE))
1134 					memblock_limit = arm_lowmem_limit;
1135 			}
1136 
1137 		}
1138 	}
1139 
1140 	if (should_use_highmem)
1141 		pr_notice("Consider using a HIGHMEM enabled kernel.\n");
1142 
1143 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1144 
1145 	/*
1146 	 * Round the memblock limit down to a pmd size.  This
1147 	 * helps to ensure that we will allocate memory from the
1148 	 * last full pmd, which should be mapped.
1149 	 */
1150 	if (memblock_limit)
1151 		memblock_limit = round_down(memblock_limit, PMD_SIZE);
1152 	if (!memblock_limit)
1153 		memblock_limit = arm_lowmem_limit;
1154 
1155 	memblock_set_current_limit(memblock_limit);
1156 }
1157 
1158 static inline void prepare_page_table(void)
1159 {
1160 	unsigned long addr;
1161 	phys_addr_t end;
1162 
1163 	/*
1164 	 * Clear out all the mappings below the kernel image.
1165 	 */
1166 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1167 		pmd_clear(pmd_off_k(addr));
1168 
1169 #ifdef CONFIG_XIP_KERNEL
1170 	/* The XIP kernel is mapped in the module area -- skip over it */
1171 	addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1172 #endif
1173 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1174 		pmd_clear(pmd_off_k(addr));
1175 
1176 	/*
1177 	 * Find the end of the first block of lowmem.
1178 	 */
1179 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1180 	if (end >= arm_lowmem_limit)
1181 		end = arm_lowmem_limit;
1182 
1183 	/*
1184 	 * Clear out all the kernel space mappings, except for the first
1185 	 * memory bank, up to the vmalloc region.
1186 	 */
1187 	for (addr = __phys_to_virt(end);
1188 	     addr < VMALLOC_START; addr += PMD_SIZE)
1189 		pmd_clear(pmd_off_k(addr));
1190 }
1191 
1192 #ifdef CONFIG_ARM_LPAE
1193 /* the first page is reserved for pgd */
1194 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1195 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1196 #else
1197 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1198 #endif
1199 
1200 /*
1201  * Reserve the special regions of memory
1202  */
1203 void __init arm_mm_memblock_reserve(void)
1204 {
1205 	/*
1206 	 * Reserve the page tables.  These are already in use,
1207 	 * and can only be in node 0.
1208 	 */
1209 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1210 
1211 #ifdef CONFIG_SA1111
1212 	/*
1213 	 * Because of the SA1111 DMA bug, we want to preserve our
1214 	 * precious DMA-able memory...
1215 	 */
1216 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1217 #endif
1218 }
1219 
1220 /*
1221  * Set up the device mappings.  Since we clear out the page tables for all
1222  * mappings above VMALLOC_START, we will remove any debug device mappings.
1223  * This means you have to be careful how you debug this function, or any
1224  * called function.  This means you can't use any function or debugging
1225  * method which may touch any device, otherwise the kernel _will_ crash.
1226  */
1227 static void __init devicemaps_init(const struct machine_desc *mdesc)
1228 {
1229 	struct map_desc map;
1230 	unsigned long addr;
1231 	void *vectors;
1232 
1233 	/*
1234 	 * Allocate the vector page early.
1235 	 */
1236 	vectors = early_alloc(PAGE_SIZE * 2);
1237 
1238 	early_trap_init(vectors);
1239 
1240 	for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1241 		pmd_clear(pmd_off_k(addr));
1242 
1243 	/*
1244 	 * Map the kernel if it is XIP.
1245 	 * It is always first in the modulearea.
1246 	 */
1247 #ifdef CONFIG_XIP_KERNEL
1248 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1249 	map.virtual = MODULES_VADDR;
1250 	map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1251 	map.type = MT_ROM;
1252 	create_mapping(&map);
1253 #endif
1254 
1255 	/*
1256 	 * Map the cache flushing regions.
1257 	 */
1258 #ifdef FLUSH_BASE
1259 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1260 	map.virtual = FLUSH_BASE;
1261 	map.length = SZ_1M;
1262 	map.type = MT_CACHECLEAN;
1263 	create_mapping(&map);
1264 #endif
1265 #ifdef FLUSH_BASE_MINICACHE
1266 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1267 	map.virtual = FLUSH_BASE_MINICACHE;
1268 	map.length = SZ_1M;
1269 	map.type = MT_MINICLEAN;
1270 	create_mapping(&map);
1271 #endif
1272 
1273 	/*
1274 	 * Create a mapping for the machine vectors at the high-vectors
1275 	 * location (0xffff0000).  If we aren't using high-vectors, also
1276 	 * create a mapping at the low-vectors virtual address.
1277 	 */
1278 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1279 	map.virtual = 0xffff0000;
1280 	map.length = PAGE_SIZE;
1281 #ifdef CONFIG_KUSER_HELPERS
1282 	map.type = MT_HIGH_VECTORS;
1283 #else
1284 	map.type = MT_LOW_VECTORS;
1285 #endif
1286 	create_mapping(&map);
1287 
1288 	if (!vectors_high()) {
1289 		map.virtual = 0;
1290 		map.length = PAGE_SIZE * 2;
1291 		map.type = MT_LOW_VECTORS;
1292 		create_mapping(&map);
1293 	}
1294 
1295 	/* Now create a kernel read-only mapping */
1296 	map.pfn += 1;
1297 	map.virtual = 0xffff0000 + PAGE_SIZE;
1298 	map.length = PAGE_SIZE;
1299 	map.type = MT_LOW_VECTORS;
1300 	create_mapping(&map);
1301 
1302 	/*
1303 	 * Ask the machine support to map in the statically mapped devices.
1304 	 */
1305 	if (mdesc->map_io)
1306 		mdesc->map_io();
1307 	else
1308 		debug_ll_io_init();
1309 	fill_pmd_gaps();
1310 
1311 	/* Reserve fixed i/o space in VMALLOC region */
1312 	pci_reserve_io();
1313 
1314 	/*
1315 	 * Finally flush the caches and tlb to ensure that we're in a
1316 	 * consistent state wrt the writebuffer.  This also ensures that
1317 	 * any write-allocated cache lines in the vector page are written
1318 	 * back.  After this point, we can start to touch devices again.
1319 	 */
1320 	local_flush_tlb_all();
1321 	flush_cache_all();
1322 }
1323 
1324 static void __init kmap_init(void)
1325 {
1326 #ifdef CONFIG_HIGHMEM
1327 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1328 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1329 #endif
1330 
1331 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1332 			_PAGE_KERNEL_TABLE);
1333 }
1334 
1335 static void __init map_lowmem(void)
1336 {
1337 	struct memblock_region *reg;
1338 	phys_addr_t kernel_x_start = round_down(__pa(_stext), SECTION_SIZE);
1339 	phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1340 
1341 	/* Map all the lowmem memory banks. */
1342 	for_each_memblock(memory, reg) {
1343 		phys_addr_t start = reg->base;
1344 		phys_addr_t end = start + reg->size;
1345 		struct map_desc map;
1346 
1347 		if (end > arm_lowmem_limit)
1348 			end = arm_lowmem_limit;
1349 		if (start >= end)
1350 			break;
1351 
1352 		if (end < kernel_x_start) {
1353 			map.pfn = __phys_to_pfn(start);
1354 			map.virtual = __phys_to_virt(start);
1355 			map.length = end - start;
1356 			map.type = MT_MEMORY_RWX;
1357 
1358 			create_mapping(&map);
1359 		} else if (start >= kernel_x_end) {
1360 			map.pfn = __phys_to_pfn(start);
1361 			map.virtual = __phys_to_virt(start);
1362 			map.length = end - start;
1363 			map.type = MT_MEMORY_RW;
1364 
1365 			create_mapping(&map);
1366 		} else {
1367 			/* This better cover the entire kernel */
1368 			if (start < kernel_x_start) {
1369 				map.pfn = __phys_to_pfn(start);
1370 				map.virtual = __phys_to_virt(start);
1371 				map.length = kernel_x_start - start;
1372 				map.type = MT_MEMORY_RW;
1373 
1374 				create_mapping(&map);
1375 			}
1376 
1377 			map.pfn = __phys_to_pfn(kernel_x_start);
1378 			map.virtual = __phys_to_virt(kernel_x_start);
1379 			map.length = kernel_x_end - kernel_x_start;
1380 			map.type = MT_MEMORY_RWX;
1381 
1382 			create_mapping(&map);
1383 
1384 			if (kernel_x_end < end) {
1385 				map.pfn = __phys_to_pfn(kernel_x_end);
1386 				map.virtual = __phys_to_virt(kernel_x_end);
1387 				map.length = end - kernel_x_end;
1388 				map.type = MT_MEMORY_RW;
1389 
1390 				create_mapping(&map);
1391 			}
1392 		}
1393 	}
1394 }
1395 
1396 #ifdef CONFIG_ARM_PV_FIXUP
1397 extern unsigned long __atags_pointer;
1398 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata);
1399 pgtables_remap lpae_pgtables_remap_asm;
1400 
1401 /*
1402  * early_paging_init() recreates boot time page table setup, allowing machines
1403  * to switch over to a high (>4G) address space on LPAE systems
1404  */
1405 void __init early_paging_init(const struct machine_desc *mdesc)
1406 {
1407 	pgtables_remap *lpae_pgtables_remap;
1408 	unsigned long pa_pgd;
1409 	unsigned int cr, ttbcr;
1410 	long long offset;
1411 	void *boot_data;
1412 
1413 	if (!mdesc->pv_fixup)
1414 		return;
1415 
1416 	offset = mdesc->pv_fixup();
1417 	if (offset == 0)
1418 		return;
1419 
1420 	/*
1421 	 * Get the address of the remap function in the 1:1 identity
1422 	 * mapping setup by the early page table assembly code.  We
1423 	 * must get this prior to the pv update.  The following barrier
1424 	 * ensures that this is complete before we fixup any P:V offsets.
1425 	 */
1426 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1427 	pa_pgd = __pa(swapper_pg_dir);
1428 	boot_data = __va(__atags_pointer);
1429 	barrier();
1430 
1431 	pr_info("Switching physical address space to 0x%08llx\n",
1432 		(u64)PHYS_OFFSET + offset);
1433 
1434 	/* Re-set the phys pfn offset, and the pv offset */
1435 	__pv_offset += offset;
1436 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1437 
1438 	/* Run the patch stub to update the constants */
1439 	fixup_pv_table(&__pv_table_begin,
1440 		(&__pv_table_end - &__pv_table_begin) << 2);
1441 
1442 	/*
1443 	 * We changing not only the virtual to physical mapping, but also
1444 	 * the physical addresses used to access memory.  We need to flush
1445 	 * all levels of cache in the system with caching disabled to
1446 	 * ensure that all data is written back, and nothing is prefetched
1447 	 * into the caches.  We also need to prevent the TLB walkers
1448 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1449 	 * specific.
1450 	 */
1451 	cr = get_cr();
1452 	set_cr(cr & ~(CR_I | CR_C));
1453 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1454 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1455 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1456 	flush_cache_all();
1457 
1458 	/*
1459 	 * Fixup the page tables - this must be in the idmap region as
1460 	 * we need to disable the MMU to do this safely, and hence it
1461 	 * needs to be assembly.  It's fairly simple, as we're using the
1462 	 * temporary tables setup by the initial assembly code.
1463 	 */
1464 	lpae_pgtables_remap(offset, pa_pgd, boot_data);
1465 
1466 	/* Re-enable the caches and cacheable TLB walks */
1467 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1468 	set_cr(cr);
1469 }
1470 
1471 #else
1472 
1473 void __init early_paging_init(const struct machine_desc *mdesc)
1474 {
1475 	long long offset;
1476 
1477 	if (!mdesc->pv_fixup)
1478 		return;
1479 
1480 	offset = mdesc->pv_fixup();
1481 	if (offset == 0)
1482 		return;
1483 
1484 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1485 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1486 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1487 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1488 }
1489 
1490 #endif
1491 
1492 /*
1493  * paging_init() sets up the page tables, initialises the zone memory
1494  * maps, and sets up the zero page, bad page and bad page tables.
1495  */
1496 void __init paging_init(const struct machine_desc *mdesc)
1497 {
1498 	void *zero_page;
1499 
1500 	build_mem_type_table();
1501 	prepare_page_table();
1502 	map_lowmem();
1503 	memblock_set_current_limit(arm_lowmem_limit);
1504 	dma_contiguous_remap();
1505 	devicemaps_init(mdesc);
1506 	kmap_init();
1507 	tcm_init();
1508 
1509 	top_pmd = pmd_off_k(0xffff0000);
1510 
1511 	/* allocate the zero page. */
1512 	zero_page = early_alloc(PAGE_SIZE);
1513 
1514 	bootmem_init();
1515 
1516 	empty_zero_page = virt_to_page(zero_page);
1517 	__flush_dcache_page(NULL, empty_zero_page);
1518 }
1519