xref: /openbmc/linux/arch/arm/mm/mmu.c (revision 6c33a6f4)
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 pgprot_t pgprot_hyp_device;
67 pgprot_t pgprot_s2;
68 pgprot_t pgprot_s2_device;
69 
70 EXPORT_SYMBOL(pgprot_user);
71 EXPORT_SYMBOL(pgprot_kernel);
72 
73 struct cachepolicy {
74 	const char	policy[16];
75 	unsigned int	cr_mask;
76 	pmdval_t	pmd;
77 	pteval_t	pte;
78 	pteval_t	pte_s2;
79 };
80 
81 #ifdef CONFIG_ARM_LPAE
82 #define s2_policy(policy)	policy
83 #else
84 #define s2_policy(policy)	0
85 #endif
86 
87 unsigned long kimage_voffset __ro_after_init;
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_CACHE_MASK;
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[] __ro_after_init = {
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_cache */
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_VECTORS,
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_VECTORS,
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 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr);
361 
362 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS]
363 	__aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata;
364 
365 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr)
366 {
367 	return &bm_pte[pte_index(addr)];
368 }
369 
370 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr)
371 {
372 	return pte_offset_kernel(dir, addr);
373 }
374 
375 static inline pmd_t * __init fixmap_pmd(unsigned long addr)
376 {
377 	pgd_t *pgd = pgd_offset_k(addr);
378 	pud_t *pud = pud_offset(pgd, addr);
379 	pmd_t *pmd = pmd_offset(pud, addr);
380 
381 	return pmd;
382 }
383 
384 void __init early_fixmap_init(void)
385 {
386 	pmd_t *pmd;
387 
388 	/*
389 	 * The early fixmap range spans multiple pmds, for which
390 	 * we are not prepared:
391 	 */
392 	BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT)
393 		     != FIXADDR_TOP >> PMD_SHIFT);
394 
395 	pmd = fixmap_pmd(FIXADDR_TOP);
396 	pmd_populate_kernel(&init_mm, pmd, bm_pte);
397 
398 	pte_offset_fixmap = pte_offset_early_fixmap;
399 }
400 
401 /*
402  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
403  * As a result, this can only be called with preemption disabled, as under
404  * stop_machine().
405  */
406 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
407 {
408 	unsigned long vaddr = __fix_to_virt(idx);
409 	pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr);
410 
411 	/* Make sure fixmap region does not exceed available allocation. */
412 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
413 		     FIXADDR_END);
414 	BUG_ON(idx >= __end_of_fixed_addresses);
415 
416 	/* we only support device mappings until pgprot_kernel has been set */
417 	if (WARN_ON(pgprot_val(prot) != pgprot_val(FIXMAP_PAGE_IO) &&
418 		    pgprot_val(pgprot_kernel) == 0))
419 		return;
420 
421 	if (pgprot_val(prot))
422 		set_pte_at(NULL, vaddr, pte,
423 			pfn_pte(phys >> PAGE_SHIFT, prot));
424 	else
425 		pte_clear(NULL, vaddr, pte);
426 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
427 }
428 
429 /*
430  * Adjust the PMD section entries according to the CPU in use.
431  */
432 static void __init build_mem_type_table(void)
433 {
434 	struct cachepolicy *cp;
435 	unsigned int cr = get_cr();
436 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
437 	pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
438 	int cpu_arch = cpu_architecture();
439 	int i;
440 
441 	if (cpu_arch < CPU_ARCH_ARMv6) {
442 #if defined(CONFIG_CPU_DCACHE_DISABLE)
443 		if (cachepolicy > CPOLICY_BUFFERED)
444 			cachepolicy = CPOLICY_BUFFERED;
445 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
446 		if (cachepolicy > CPOLICY_WRITETHROUGH)
447 			cachepolicy = CPOLICY_WRITETHROUGH;
448 #endif
449 	}
450 	if (cpu_arch < CPU_ARCH_ARMv5) {
451 		if (cachepolicy >= CPOLICY_WRITEALLOC)
452 			cachepolicy = CPOLICY_WRITEBACK;
453 		ecc_mask = 0;
454 	}
455 
456 	if (is_smp()) {
457 		if (cachepolicy != CPOLICY_WRITEALLOC) {
458 			pr_warn("Forcing write-allocate cache policy for SMP\n");
459 			cachepolicy = CPOLICY_WRITEALLOC;
460 		}
461 		if (!(initial_pmd_value & PMD_SECT_S)) {
462 			pr_warn("Forcing shared mappings for SMP\n");
463 			initial_pmd_value |= PMD_SECT_S;
464 		}
465 	}
466 
467 	/*
468 	 * Strip out features not present on earlier architectures.
469 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
470 	 * without extended page tables don't have the 'Shared' bit.
471 	 */
472 	if (cpu_arch < CPU_ARCH_ARMv5)
473 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
474 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
475 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
476 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
477 			mem_types[i].prot_sect &= ~PMD_SECT_S;
478 
479 	/*
480 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
481 	 * "update-able on write" bit on ARM610).  However, Xscale and
482 	 * Xscale3 require this bit to be cleared.
483 	 */
484 	if (cpu_is_xscale_family()) {
485 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
486 			mem_types[i].prot_sect &= ~PMD_BIT4;
487 			mem_types[i].prot_l1 &= ~PMD_BIT4;
488 		}
489 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
490 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
491 			if (mem_types[i].prot_l1)
492 				mem_types[i].prot_l1 |= PMD_BIT4;
493 			if (mem_types[i].prot_sect)
494 				mem_types[i].prot_sect |= PMD_BIT4;
495 		}
496 	}
497 
498 	/*
499 	 * Mark the device areas according to the CPU/architecture.
500 	 */
501 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
502 		if (!cpu_is_xsc3()) {
503 			/*
504 			 * Mark device regions on ARMv6+ as execute-never
505 			 * to prevent speculative instruction fetches.
506 			 */
507 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
508 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
509 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
510 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
511 
512 			/* Also setup NX memory mapping */
513 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
514 		}
515 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
516 			/*
517 			 * For ARMv7 with TEX remapping,
518 			 * - shared device is SXCB=1100
519 			 * - nonshared device is SXCB=0100
520 			 * - write combine device mem is SXCB=0001
521 			 * (Uncached Normal memory)
522 			 */
523 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
524 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
525 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
526 		} else if (cpu_is_xsc3()) {
527 			/*
528 			 * For Xscale3,
529 			 * - shared device is TEXCB=00101
530 			 * - nonshared device is TEXCB=01000
531 			 * - write combine device mem is TEXCB=00100
532 			 * (Inner/Outer Uncacheable in xsc3 parlance)
533 			 */
534 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
535 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
536 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
537 		} else {
538 			/*
539 			 * For ARMv6 and ARMv7 without TEX remapping,
540 			 * - shared device is TEXCB=00001
541 			 * - nonshared device is TEXCB=01000
542 			 * - write combine device mem is TEXCB=00100
543 			 * (Uncached Normal in ARMv6 parlance).
544 			 */
545 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
546 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
547 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
548 		}
549 	} else {
550 		/*
551 		 * On others, write combining is "Uncached/Buffered"
552 		 */
553 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
554 	}
555 
556 	/*
557 	 * Now deal with the memory-type mappings
558 	 */
559 	cp = &cache_policies[cachepolicy];
560 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
561 	s2_pgprot = cp->pte_s2;
562 	hyp_device_pgprot = mem_types[MT_DEVICE].prot_pte;
563 	s2_device_pgprot = mem_types[MT_DEVICE].prot_pte_s2;
564 
565 #ifndef CONFIG_ARM_LPAE
566 	/*
567 	 * We don't use domains on ARMv6 (since this causes problems with
568 	 * v6/v7 kernels), so we must use a separate memory type for user
569 	 * r/o, kernel r/w to map the vectors page.
570 	 */
571 	if (cpu_arch == CPU_ARCH_ARMv6)
572 		vecs_pgprot |= L_PTE_MT_VECTORS;
573 
574 	/*
575 	 * Check is it with support for the PXN bit
576 	 * in the Short-descriptor translation table format descriptors.
577 	 */
578 	if (cpu_arch == CPU_ARCH_ARMv7 &&
579 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
580 		user_pmd_table |= PMD_PXNTABLE;
581 	}
582 #endif
583 
584 	/*
585 	 * ARMv6 and above have extended page tables.
586 	 */
587 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
588 #ifndef CONFIG_ARM_LPAE
589 		/*
590 		 * Mark cache clean areas and XIP ROM read only
591 		 * from SVC mode and no access from userspace.
592 		 */
593 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
594 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
595 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
596 #endif
597 
598 		/*
599 		 * If the initial page tables were created with the S bit
600 		 * set, then we need to do the same here for the same
601 		 * reasons given in early_cachepolicy().
602 		 */
603 		if (initial_pmd_value & PMD_SECT_S) {
604 			user_pgprot |= L_PTE_SHARED;
605 			kern_pgprot |= L_PTE_SHARED;
606 			vecs_pgprot |= L_PTE_SHARED;
607 			s2_pgprot |= L_PTE_SHARED;
608 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
609 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
610 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
611 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
612 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
613 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
614 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
615 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
616 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
617 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
618 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
619 		}
620 	}
621 
622 	/*
623 	 * Non-cacheable Normal - intended for memory areas that must
624 	 * not cause dirty cache line writebacks when used
625 	 */
626 	if (cpu_arch >= CPU_ARCH_ARMv6) {
627 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
628 			/* Non-cacheable Normal is XCB = 001 */
629 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
630 				PMD_SECT_BUFFERED;
631 		} else {
632 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
633 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
634 				PMD_SECT_TEX(1);
635 		}
636 	} else {
637 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
638 	}
639 
640 #ifdef CONFIG_ARM_LPAE
641 	/*
642 	 * Do not generate access flag faults for the kernel mappings.
643 	 */
644 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
645 		mem_types[i].prot_pte |= PTE_EXT_AF;
646 		if (mem_types[i].prot_sect)
647 			mem_types[i].prot_sect |= PMD_SECT_AF;
648 	}
649 	kern_pgprot |= PTE_EXT_AF;
650 	vecs_pgprot |= PTE_EXT_AF;
651 
652 	/*
653 	 * Set PXN for user mappings
654 	 */
655 	user_pgprot |= PTE_EXT_PXN;
656 #endif
657 
658 	for (i = 0; i < 16; i++) {
659 		pteval_t v = pgprot_val(protection_map[i]);
660 		protection_map[i] = __pgprot(v | user_pgprot);
661 	}
662 
663 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
664 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
665 
666 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
667 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
668 				 L_PTE_DIRTY | kern_pgprot);
669 	pgprot_s2  = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
670 	pgprot_s2_device  = __pgprot(s2_device_pgprot);
671 	pgprot_hyp_device  = __pgprot(hyp_device_pgprot);
672 
673 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
674 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
675 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
676 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
677 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
678 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
679 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
680 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
681 	mem_types[MT_ROM].prot_sect |= cp->pmd;
682 
683 	switch (cp->pmd) {
684 	case PMD_SECT_WT:
685 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
686 		break;
687 	case PMD_SECT_WB:
688 	case PMD_SECT_WBWA:
689 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
690 		break;
691 	}
692 	pr_info("Memory policy: %sData cache %s\n",
693 		ecc_mask ? "ECC enabled, " : "", cp->policy);
694 
695 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
696 		struct mem_type *t = &mem_types[i];
697 		if (t->prot_l1)
698 			t->prot_l1 |= PMD_DOMAIN(t->domain);
699 		if (t->prot_sect)
700 			t->prot_sect |= PMD_DOMAIN(t->domain);
701 	}
702 }
703 
704 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
705 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
706 			      unsigned long size, pgprot_t vma_prot)
707 {
708 	if (!pfn_valid(pfn))
709 		return pgprot_noncached(vma_prot);
710 	else if (file->f_flags & O_SYNC)
711 		return pgprot_writecombine(vma_prot);
712 	return vma_prot;
713 }
714 EXPORT_SYMBOL(phys_mem_access_prot);
715 #endif
716 
717 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
718 
719 static void __init *early_alloc(unsigned long sz)
720 {
721 	void *ptr = memblock_alloc(sz, sz);
722 
723 	if (!ptr)
724 		panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
725 		      __func__, sz, sz);
726 
727 	return ptr;
728 }
729 
730 static void *__init late_alloc(unsigned long sz)
731 {
732 	void *ptr = (void *)__get_free_pages(GFP_PGTABLE_KERNEL, get_order(sz));
733 
734 	if (!ptr || !pgtable_pte_page_ctor(virt_to_page(ptr)))
735 		BUG();
736 	return ptr;
737 }
738 
739 static pte_t * __init arm_pte_alloc(pmd_t *pmd, unsigned long addr,
740 				unsigned long prot,
741 				void *(*alloc)(unsigned long sz))
742 {
743 	if (pmd_none(*pmd)) {
744 		pte_t *pte = alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
745 		__pmd_populate(pmd, __pa(pte), prot);
746 	}
747 	BUG_ON(pmd_bad(*pmd));
748 	return pte_offset_kernel(pmd, addr);
749 }
750 
751 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr,
752 				      unsigned long prot)
753 {
754 	return arm_pte_alloc(pmd, addr, prot, early_alloc);
755 }
756 
757 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
758 				  unsigned long end, unsigned long pfn,
759 				  const struct mem_type *type,
760 				  void *(*alloc)(unsigned long sz),
761 				  bool ng)
762 {
763 	pte_t *pte = arm_pte_alloc(pmd, addr, type->prot_l1, alloc);
764 	do {
765 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)),
766 			    ng ? PTE_EXT_NG : 0);
767 		pfn++;
768 	} while (pte++, addr += PAGE_SIZE, addr != end);
769 }
770 
771 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
772 			unsigned long end, phys_addr_t phys,
773 			const struct mem_type *type, bool ng)
774 {
775 	pmd_t *p = pmd;
776 
777 #ifndef CONFIG_ARM_LPAE
778 	/*
779 	 * In classic MMU format, puds and pmds are folded in to
780 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
781 	 * group of L1 entries making up one logical pointer to
782 	 * an L2 table (2MB), where as PMDs refer to the individual
783 	 * L1 entries (1MB). Hence increment to get the correct
784 	 * offset for odd 1MB sections.
785 	 * (See arch/arm/include/asm/pgtable-2level.h)
786 	 */
787 	if (addr & SECTION_SIZE)
788 		pmd++;
789 #endif
790 	do {
791 		*pmd = __pmd(phys | type->prot_sect | (ng ? PMD_SECT_nG : 0));
792 		phys += SECTION_SIZE;
793 	} while (pmd++, addr += SECTION_SIZE, addr != end);
794 
795 	flush_pmd_entry(p);
796 }
797 
798 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
799 				      unsigned long end, phys_addr_t phys,
800 				      const struct mem_type *type,
801 				      void *(*alloc)(unsigned long sz), bool ng)
802 {
803 	pmd_t *pmd = pmd_offset(pud, addr);
804 	unsigned long next;
805 
806 	do {
807 		/*
808 		 * With LPAE, we must loop over to map
809 		 * all the pmds for the given range.
810 		 */
811 		next = pmd_addr_end(addr, end);
812 
813 		/*
814 		 * Try a section mapping - addr, next and phys must all be
815 		 * aligned to a section boundary.
816 		 */
817 		if (type->prot_sect &&
818 				((addr | next | phys) & ~SECTION_MASK) == 0) {
819 			__map_init_section(pmd, addr, next, phys, type, ng);
820 		} else {
821 			alloc_init_pte(pmd, addr, next,
822 				       __phys_to_pfn(phys), type, alloc, ng);
823 		}
824 
825 		phys += next - addr;
826 
827 	} while (pmd++, addr = next, addr != end);
828 }
829 
830 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
831 				  unsigned long end, phys_addr_t phys,
832 				  const struct mem_type *type,
833 				  void *(*alloc)(unsigned long sz), bool ng)
834 {
835 	pud_t *pud = pud_offset(pgd, addr);
836 	unsigned long next;
837 
838 	do {
839 		next = pud_addr_end(addr, end);
840 		alloc_init_pmd(pud, addr, next, phys, type, alloc, ng);
841 		phys += next - addr;
842 	} while (pud++, addr = next, addr != end);
843 }
844 
845 #ifndef CONFIG_ARM_LPAE
846 static void __init create_36bit_mapping(struct mm_struct *mm,
847 					struct map_desc *md,
848 					const struct mem_type *type,
849 					bool ng)
850 {
851 	unsigned long addr, length, end;
852 	phys_addr_t phys;
853 	pgd_t *pgd;
854 
855 	addr = md->virtual;
856 	phys = __pfn_to_phys(md->pfn);
857 	length = PAGE_ALIGN(md->length);
858 
859 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
860 		pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
861 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
862 		return;
863 	}
864 
865 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
866 	 *	Since domain assignments can in fact be arbitrary, the
867 	 *	'domain == 0' check below is required to insure that ARMv6
868 	 *	supersections are only allocated for domain 0 regardless
869 	 *	of the actual domain assignments in use.
870 	 */
871 	if (type->domain) {
872 		pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
873 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
874 		return;
875 	}
876 
877 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
878 		pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
879 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
880 		return;
881 	}
882 
883 	/*
884 	 * Shift bits [35:32] of address into bits [23:20] of PMD
885 	 * (See ARMv6 spec).
886 	 */
887 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
888 
889 	pgd = pgd_offset(mm, addr);
890 	end = addr + length;
891 	do {
892 		pud_t *pud = pud_offset(pgd, addr);
893 		pmd_t *pmd = pmd_offset(pud, addr);
894 		int i;
895 
896 		for (i = 0; i < 16; i++)
897 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER |
898 				       (ng ? PMD_SECT_nG : 0));
899 
900 		addr += SUPERSECTION_SIZE;
901 		phys += SUPERSECTION_SIZE;
902 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
903 	} while (addr != end);
904 }
905 #endif	/* !CONFIG_ARM_LPAE */
906 
907 static void __init __create_mapping(struct mm_struct *mm, struct map_desc *md,
908 				    void *(*alloc)(unsigned long sz),
909 				    bool ng)
910 {
911 	unsigned long addr, length, end;
912 	phys_addr_t phys;
913 	const struct mem_type *type;
914 	pgd_t *pgd;
915 
916 	type = &mem_types[md->type];
917 
918 #ifndef CONFIG_ARM_LPAE
919 	/*
920 	 * Catch 36-bit addresses
921 	 */
922 	if (md->pfn >= 0x100000) {
923 		create_36bit_mapping(mm, md, type, ng);
924 		return;
925 	}
926 #endif
927 
928 	addr = md->virtual & PAGE_MASK;
929 	phys = __pfn_to_phys(md->pfn);
930 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
931 
932 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
933 		pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
934 			(long long)__pfn_to_phys(md->pfn), addr);
935 		return;
936 	}
937 
938 	pgd = pgd_offset(mm, addr);
939 	end = addr + length;
940 	do {
941 		unsigned long next = pgd_addr_end(addr, end);
942 
943 		alloc_init_pud(pgd, addr, next, phys, type, alloc, ng);
944 
945 		phys += next - addr;
946 		addr = next;
947 	} while (pgd++, addr != end);
948 }
949 
950 /*
951  * Create the page directory entries and any necessary
952  * page tables for the mapping specified by `md'.  We
953  * are able to cope here with varying sizes and address
954  * offsets, and we take full advantage of sections and
955  * supersections.
956  */
957 static void __init create_mapping(struct map_desc *md)
958 {
959 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
960 		pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
961 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
962 		return;
963 	}
964 
965 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
966 	    md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START &&
967 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
968 		pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
969 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
970 	}
971 
972 	__create_mapping(&init_mm, md, early_alloc, false);
973 }
974 
975 void __init create_mapping_late(struct mm_struct *mm, struct map_desc *md,
976 				bool ng)
977 {
978 #ifdef CONFIG_ARM_LPAE
979 	pud_t *pud = pud_alloc(mm, pgd_offset(mm, md->virtual), md->virtual);
980 	if (WARN_ON(!pud))
981 		return;
982 	pmd_alloc(mm, pud, 0);
983 #endif
984 	__create_mapping(mm, md, late_alloc, ng);
985 }
986 
987 /*
988  * Create the architecture specific mappings
989  */
990 void __init iotable_init(struct map_desc *io_desc, int nr)
991 {
992 	struct map_desc *md;
993 	struct vm_struct *vm;
994 	struct static_vm *svm;
995 
996 	if (!nr)
997 		return;
998 
999 	svm = memblock_alloc(sizeof(*svm) * nr, __alignof__(*svm));
1000 	if (!svm)
1001 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1002 		      __func__, sizeof(*svm) * nr, __alignof__(*svm));
1003 
1004 	for (md = io_desc; nr; md++, nr--) {
1005 		create_mapping(md);
1006 
1007 		vm = &svm->vm;
1008 		vm->addr = (void *)(md->virtual & PAGE_MASK);
1009 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
1010 		vm->phys_addr = __pfn_to_phys(md->pfn);
1011 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
1012 		vm->flags |= VM_ARM_MTYPE(md->type);
1013 		vm->caller = iotable_init;
1014 		add_static_vm_early(svm++);
1015 	}
1016 }
1017 
1018 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
1019 				  void *caller)
1020 {
1021 	struct vm_struct *vm;
1022 	struct static_vm *svm;
1023 
1024 	svm = memblock_alloc(sizeof(*svm), __alignof__(*svm));
1025 	if (!svm)
1026 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1027 		      __func__, sizeof(*svm), __alignof__(*svm));
1028 
1029 	vm = &svm->vm;
1030 	vm->addr = (void *)addr;
1031 	vm->size = size;
1032 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
1033 	vm->caller = caller;
1034 	add_static_vm_early(svm);
1035 }
1036 
1037 #ifndef CONFIG_ARM_LPAE
1038 
1039 /*
1040  * The Linux PMD is made of two consecutive section entries covering 2MB
1041  * (see definition in include/asm/pgtable-2level.h).  However a call to
1042  * create_mapping() may optimize static mappings by using individual
1043  * 1MB section mappings.  This leaves the actual PMD potentially half
1044  * initialized if the top or bottom section entry isn't used, leaving it
1045  * open to problems if a subsequent ioremap() or vmalloc() tries to use
1046  * the virtual space left free by that unused section entry.
1047  *
1048  * Let's avoid the issue by inserting dummy vm entries covering the unused
1049  * PMD halves once the static mappings are in place.
1050  */
1051 
1052 static void __init pmd_empty_section_gap(unsigned long addr)
1053 {
1054 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1055 }
1056 
1057 static void __init fill_pmd_gaps(void)
1058 {
1059 	struct static_vm *svm;
1060 	struct vm_struct *vm;
1061 	unsigned long addr, next = 0;
1062 	pmd_t *pmd;
1063 
1064 	list_for_each_entry(svm, &static_vmlist, list) {
1065 		vm = &svm->vm;
1066 		addr = (unsigned long)vm->addr;
1067 		if (addr < next)
1068 			continue;
1069 
1070 		/*
1071 		 * Check if this vm starts on an odd section boundary.
1072 		 * If so and the first section entry for this PMD is free
1073 		 * then we block the corresponding virtual address.
1074 		 */
1075 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1076 			pmd = pmd_off_k(addr);
1077 			if (pmd_none(*pmd))
1078 				pmd_empty_section_gap(addr & PMD_MASK);
1079 		}
1080 
1081 		/*
1082 		 * Then check if this vm ends on an odd section boundary.
1083 		 * If so and the second section entry for this PMD is empty
1084 		 * then we block the corresponding virtual address.
1085 		 */
1086 		addr += vm->size;
1087 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1088 			pmd = pmd_off_k(addr) + 1;
1089 			if (pmd_none(*pmd))
1090 				pmd_empty_section_gap(addr);
1091 		}
1092 
1093 		/* no need to look at any vm entry until we hit the next PMD */
1094 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1095 	}
1096 }
1097 
1098 #else
1099 #define fill_pmd_gaps() do { } while (0)
1100 #endif
1101 
1102 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1103 static void __init pci_reserve_io(void)
1104 {
1105 	struct static_vm *svm;
1106 
1107 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1108 	if (svm)
1109 		return;
1110 
1111 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1112 }
1113 #else
1114 #define pci_reserve_io() do { } while (0)
1115 #endif
1116 
1117 #ifdef CONFIG_DEBUG_LL
1118 void __init debug_ll_io_init(void)
1119 {
1120 	struct map_desc map;
1121 
1122 	debug_ll_addr(&map.pfn, &map.virtual);
1123 	if (!map.pfn || !map.virtual)
1124 		return;
1125 	map.pfn = __phys_to_pfn(map.pfn);
1126 	map.virtual &= PAGE_MASK;
1127 	map.length = PAGE_SIZE;
1128 	map.type = MT_DEVICE;
1129 	iotable_init(&map, 1);
1130 }
1131 #endif
1132 
1133 static void * __initdata vmalloc_min =
1134 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1135 
1136 /*
1137  * vmalloc=size forces the vmalloc area to be exactly 'size'
1138  * bytes. This can be used to increase (or decrease) the vmalloc
1139  * area - the default is 240m.
1140  */
1141 static int __init early_vmalloc(char *arg)
1142 {
1143 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1144 
1145 	if (vmalloc_reserve < SZ_16M) {
1146 		vmalloc_reserve = SZ_16M;
1147 		pr_warn("vmalloc area too small, limiting to %luMB\n",
1148 			vmalloc_reserve >> 20);
1149 	}
1150 
1151 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1152 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1153 		pr_warn("vmalloc area is too big, limiting to %luMB\n",
1154 			vmalloc_reserve >> 20);
1155 	}
1156 
1157 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1158 	return 0;
1159 }
1160 early_param("vmalloc", early_vmalloc);
1161 
1162 phys_addr_t arm_lowmem_limit __initdata = 0;
1163 
1164 void __init adjust_lowmem_bounds(void)
1165 {
1166 	phys_addr_t memblock_limit = 0;
1167 	u64 vmalloc_limit;
1168 	struct memblock_region *reg;
1169 	phys_addr_t lowmem_limit = 0;
1170 
1171 	/*
1172 	 * Let's use our own (unoptimized) equivalent of __pa() that is
1173 	 * not affected by wrap-arounds when sizeof(phys_addr_t) == 4.
1174 	 * The result is used as the upper bound on physical memory address
1175 	 * and may itself be outside the valid range for which phys_addr_t
1176 	 * and therefore __pa() is defined.
1177 	 */
1178 	vmalloc_limit = (u64)(uintptr_t)vmalloc_min - PAGE_OFFSET + PHYS_OFFSET;
1179 
1180 	/*
1181 	 * The first usable region must be PMD aligned. Mark its start
1182 	 * as MEMBLOCK_NOMAP if it isn't
1183 	 */
1184 	for_each_memblock(memory, reg) {
1185 		if (!memblock_is_nomap(reg)) {
1186 			if (!IS_ALIGNED(reg->base, PMD_SIZE)) {
1187 				phys_addr_t len;
1188 
1189 				len = round_up(reg->base, PMD_SIZE) - reg->base;
1190 				memblock_mark_nomap(reg->base, len);
1191 			}
1192 			break;
1193 		}
1194 	}
1195 
1196 	for_each_memblock(memory, reg) {
1197 		phys_addr_t block_start = reg->base;
1198 		phys_addr_t block_end = reg->base + reg->size;
1199 
1200 		if (memblock_is_nomap(reg))
1201 			continue;
1202 
1203 		if (reg->base < vmalloc_limit) {
1204 			if (block_end > lowmem_limit)
1205 				/*
1206 				 * Compare as u64 to ensure vmalloc_limit does
1207 				 * not get truncated. block_end should always
1208 				 * fit in phys_addr_t so there should be no
1209 				 * issue with assignment.
1210 				 */
1211 				lowmem_limit = min_t(u64,
1212 							 vmalloc_limit,
1213 							 block_end);
1214 
1215 			/*
1216 			 * Find the first non-pmd-aligned page, and point
1217 			 * memblock_limit at it. This relies on rounding the
1218 			 * limit down to be pmd-aligned, which happens at the
1219 			 * end of this function.
1220 			 *
1221 			 * With this algorithm, the start or end of almost any
1222 			 * bank can be non-pmd-aligned. The only exception is
1223 			 * that the start of the bank 0 must be section-
1224 			 * aligned, since otherwise memory would need to be
1225 			 * allocated when mapping the start of bank 0, which
1226 			 * occurs before any free memory is mapped.
1227 			 */
1228 			if (!memblock_limit) {
1229 				if (!IS_ALIGNED(block_start, PMD_SIZE))
1230 					memblock_limit = block_start;
1231 				else if (!IS_ALIGNED(block_end, PMD_SIZE))
1232 					memblock_limit = lowmem_limit;
1233 			}
1234 
1235 		}
1236 	}
1237 
1238 	arm_lowmem_limit = lowmem_limit;
1239 
1240 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1241 
1242 	if (!memblock_limit)
1243 		memblock_limit = arm_lowmem_limit;
1244 
1245 	/*
1246 	 * Round the memblock limit down to a pmd size.  This
1247 	 * helps to ensure that we will allocate memory from the
1248 	 * last full pmd, which should be mapped.
1249 	 */
1250 	memblock_limit = round_down(memblock_limit, PMD_SIZE);
1251 
1252 	if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1253 		if (memblock_end_of_DRAM() > arm_lowmem_limit) {
1254 			phys_addr_t end = memblock_end_of_DRAM();
1255 
1256 			pr_notice("Ignoring RAM at %pa-%pa\n",
1257 				  &memblock_limit, &end);
1258 			pr_notice("Consider using a HIGHMEM enabled kernel.\n");
1259 
1260 			memblock_remove(memblock_limit, end - memblock_limit);
1261 		}
1262 	}
1263 
1264 	memblock_set_current_limit(memblock_limit);
1265 }
1266 
1267 static inline void prepare_page_table(void)
1268 {
1269 	unsigned long addr;
1270 	phys_addr_t end;
1271 
1272 	/*
1273 	 * Clear out all the mappings below the kernel image.
1274 	 */
1275 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1276 		pmd_clear(pmd_off_k(addr));
1277 
1278 #ifdef CONFIG_XIP_KERNEL
1279 	/* The XIP kernel is mapped in the module area -- skip over it */
1280 	addr = ((unsigned long)_exiprom + PMD_SIZE - 1) & PMD_MASK;
1281 #endif
1282 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1283 		pmd_clear(pmd_off_k(addr));
1284 
1285 	/*
1286 	 * Find the end of the first block of lowmem.
1287 	 */
1288 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1289 	if (end >= arm_lowmem_limit)
1290 		end = arm_lowmem_limit;
1291 
1292 	/*
1293 	 * Clear out all the kernel space mappings, except for the first
1294 	 * memory bank, up to the vmalloc region.
1295 	 */
1296 	for (addr = __phys_to_virt(end);
1297 	     addr < VMALLOC_START; addr += PMD_SIZE)
1298 		pmd_clear(pmd_off_k(addr));
1299 }
1300 
1301 #ifdef CONFIG_ARM_LPAE
1302 /* the first page is reserved for pgd */
1303 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1304 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1305 #else
1306 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1307 #endif
1308 
1309 /*
1310  * Reserve the special regions of memory
1311  */
1312 void __init arm_mm_memblock_reserve(void)
1313 {
1314 	/*
1315 	 * Reserve the page tables.  These are already in use,
1316 	 * and can only be in node 0.
1317 	 */
1318 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1319 
1320 #ifdef CONFIG_SA1111
1321 	/*
1322 	 * Because of the SA1111 DMA bug, we want to preserve our
1323 	 * precious DMA-able memory...
1324 	 */
1325 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1326 #endif
1327 }
1328 
1329 /*
1330  * Set up the device mappings.  Since we clear out the page tables for all
1331  * mappings above VMALLOC_START, except early fixmap, we might remove debug
1332  * device mappings.  This means earlycon can be used to debug this function
1333  * Any other function or debugging method which may touch any device _will_
1334  * crash the kernel.
1335  */
1336 static void __init devicemaps_init(const struct machine_desc *mdesc)
1337 {
1338 	struct map_desc map;
1339 	unsigned long addr;
1340 	void *vectors;
1341 
1342 	/*
1343 	 * Allocate the vector page early.
1344 	 */
1345 	vectors = early_alloc(PAGE_SIZE * 2);
1346 
1347 	early_trap_init(vectors);
1348 
1349 	/*
1350 	 * Clear page table except top pmd used by early fixmaps
1351 	 */
1352 	for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE)
1353 		pmd_clear(pmd_off_k(addr));
1354 
1355 	/*
1356 	 * Map the kernel if it is XIP.
1357 	 * It is always first in the modulearea.
1358 	 */
1359 #ifdef CONFIG_XIP_KERNEL
1360 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1361 	map.virtual = MODULES_VADDR;
1362 	map.length = ((unsigned long)_exiprom - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1363 	map.type = MT_ROM;
1364 	create_mapping(&map);
1365 #endif
1366 
1367 	/*
1368 	 * Map the cache flushing regions.
1369 	 */
1370 #ifdef FLUSH_BASE
1371 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1372 	map.virtual = FLUSH_BASE;
1373 	map.length = SZ_1M;
1374 	map.type = MT_CACHECLEAN;
1375 	create_mapping(&map);
1376 #endif
1377 #ifdef FLUSH_BASE_MINICACHE
1378 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1379 	map.virtual = FLUSH_BASE_MINICACHE;
1380 	map.length = SZ_1M;
1381 	map.type = MT_MINICLEAN;
1382 	create_mapping(&map);
1383 #endif
1384 
1385 	/*
1386 	 * Create a mapping for the machine vectors at the high-vectors
1387 	 * location (0xffff0000).  If we aren't using high-vectors, also
1388 	 * create a mapping at the low-vectors virtual address.
1389 	 */
1390 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1391 	map.virtual = 0xffff0000;
1392 	map.length = PAGE_SIZE;
1393 #ifdef CONFIG_KUSER_HELPERS
1394 	map.type = MT_HIGH_VECTORS;
1395 #else
1396 	map.type = MT_LOW_VECTORS;
1397 #endif
1398 	create_mapping(&map);
1399 
1400 	if (!vectors_high()) {
1401 		map.virtual = 0;
1402 		map.length = PAGE_SIZE * 2;
1403 		map.type = MT_LOW_VECTORS;
1404 		create_mapping(&map);
1405 	}
1406 
1407 	/* Now create a kernel read-only mapping */
1408 	map.pfn += 1;
1409 	map.virtual = 0xffff0000 + PAGE_SIZE;
1410 	map.length = PAGE_SIZE;
1411 	map.type = MT_LOW_VECTORS;
1412 	create_mapping(&map);
1413 
1414 	/*
1415 	 * Ask the machine support to map in the statically mapped devices.
1416 	 */
1417 	if (mdesc->map_io)
1418 		mdesc->map_io();
1419 	else
1420 		debug_ll_io_init();
1421 	fill_pmd_gaps();
1422 
1423 	/* Reserve fixed i/o space in VMALLOC region */
1424 	pci_reserve_io();
1425 
1426 	/*
1427 	 * Finally flush the caches and tlb to ensure that we're in a
1428 	 * consistent state wrt the writebuffer.  This also ensures that
1429 	 * any write-allocated cache lines in the vector page are written
1430 	 * back.  After this point, we can start to touch devices again.
1431 	 */
1432 	local_flush_tlb_all();
1433 	flush_cache_all();
1434 
1435 	/* Enable asynchronous aborts */
1436 	early_abt_enable();
1437 }
1438 
1439 static void __init kmap_init(void)
1440 {
1441 #ifdef CONFIG_HIGHMEM
1442 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1443 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1444 #endif
1445 
1446 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1447 			_PAGE_KERNEL_TABLE);
1448 }
1449 
1450 static void __init map_lowmem(void)
1451 {
1452 	struct memblock_region *reg;
1453 	phys_addr_t kernel_x_start = round_down(__pa(KERNEL_START), SECTION_SIZE);
1454 	phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1455 
1456 	/* Map all the lowmem memory banks. */
1457 	for_each_memblock(memory, reg) {
1458 		phys_addr_t start = reg->base;
1459 		phys_addr_t end = start + reg->size;
1460 		struct map_desc map;
1461 
1462 		if (memblock_is_nomap(reg))
1463 			continue;
1464 
1465 		if (end > arm_lowmem_limit)
1466 			end = arm_lowmem_limit;
1467 		if (start >= end)
1468 			break;
1469 
1470 		if (end < kernel_x_start) {
1471 			map.pfn = __phys_to_pfn(start);
1472 			map.virtual = __phys_to_virt(start);
1473 			map.length = end - start;
1474 			map.type = MT_MEMORY_RWX;
1475 
1476 			create_mapping(&map);
1477 		} else if (start >= kernel_x_end) {
1478 			map.pfn = __phys_to_pfn(start);
1479 			map.virtual = __phys_to_virt(start);
1480 			map.length = end - start;
1481 			map.type = MT_MEMORY_RW;
1482 
1483 			create_mapping(&map);
1484 		} else {
1485 			/* This better cover the entire kernel */
1486 			if (start < kernel_x_start) {
1487 				map.pfn = __phys_to_pfn(start);
1488 				map.virtual = __phys_to_virt(start);
1489 				map.length = kernel_x_start - start;
1490 				map.type = MT_MEMORY_RW;
1491 
1492 				create_mapping(&map);
1493 			}
1494 
1495 			map.pfn = __phys_to_pfn(kernel_x_start);
1496 			map.virtual = __phys_to_virt(kernel_x_start);
1497 			map.length = kernel_x_end - kernel_x_start;
1498 			map.type = MT_MEMORY_RWX;
1499 
1500 			create_mapping(&map);
1501 
1502 			if (kernel_x_end < end) {
1503 				map.pfn = __phys_to_pfn(kernel_x_end);
1504 				map.virtual = __phys_to_virt(kernel_x_end);
1505 				map.length = end - kernel_x_end;
1506 				map.type = MT_MEMORY_RW;
1507 
1508 				create_mapping(&map);
1509 			}
1510 		}
1511 	}
1512 }
1513 
1514 #ifdef CONFIG_ARM_PV_FIXUP
1515 extern unsigned long __atags_pointer;
1516 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata);
1517 pgtables_remap lpae_pgtables_remap_asm;
1518 
1519 /*
1520  * early_paging_init() recreates boot time page table setup, allowing machines
1521  * to switch over to a high (>4G) address space on LPAE systems
1522  */
1523 static void __init early_paging_init(const struct machine_desc *mdesc)
1524 {
1525 	pgtables_remap *lpae_pgtables_remap;
1526 	unsigned long pa_pgd;
1527 	unsigned int cr, ttbcr;
1528 	long long offset;
1529 	void *boot_data;
1530 
1531 	if (!mdesc->pv_fixup)
1532 		return;
1533 
1534 	offset = mdesc->pv_fixup();
1535 	if (offset == 0)
1536 		return;
1537 
1538 	/*
1539 	 * Get the address of the remap function in the 1:1 identity
1540 	 * mapping setup by the early page table assembly code.  We
1541 	 * must get this prior to the pv update.  The following barrier
1542 	 * ensures that this is complete before we fixup any P:V offsets.
1543 	 */
1544 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1545 	pa_pgd = __pa(swapper_pg_dir);
1546 	boot_data = __va(__atags_pointer);
1547 	barrier();
1548 
1549 	pr_info("Switching physical address space to 0x%08llx\n",
1550 		(u64)PHYS_OFFSET + offset);
1551 
1552 	/* Re-set the phys pfn offset, and the pv offset */
1553 	__pv_offset += offset;
1554 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1555 
1556 	/* Run the patch stub to update the constants */
1557 	fixup_pv_table(&__pv_table_begin,
1558 		(&__pv_table_end - &__pv_table_begin) << 2);
1559 
1560 	/*
1561 	 * We changing not only the virtual to physical mapping, but also
1562 	 * the physical addresses used to access memory.  We need to flush
1563 	 * all levels of cache in the system with caching disabled to
1564 	 * ensure that all data is written back, and nothing is prefetched
1565 	 * into the caches.  We also need to prevent the TLB walkers
1566 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1567 	 * specific.
1568 	 */
1569 	cr = get_cr();
1570 	set_cr(cr & ~(CR_I | CR_C));
1571 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1572 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1573 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1574 	flush_cache_all();
1575 
1576 	/*
1577 	 * Fixup the page tables - this must be in the idmap region as
1578 	 * we need to disable the MMU to do this safely, and hence it
1579 	 * needs to be assembly.  It's fairly simple, as we're using the
1580 	 * temporary tables setup by the initial assembly code.
1581 	 */
1582 	lpae_pgtables_remap(offset, pa_pgd, boot_data);
1583 
1584 	/* Re-enable the caches and cacheable TLB walks */
1585 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1586 	set_cr(cr);
1587 }
1588 
1589 #else
1590 
1591 static void __init early_paging_init(const struct machine_desc *mdesc)
1592 {
1593 	long long offset;
1594 
1595 	if (!mdesc->pv_fixup)
1596 		return;
1597 
1598 	offset = mdesc->pv_fixup();
1599 	if (offset == 0)
1600 		return;
1601 
1602 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1603 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1604 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1605 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1606 }
1607 
1608 #endif
1609 
1610 static void __init early_fixmap_shutdown(void)
1611 {
1612 	int i;
1613 	unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1);
1614 
1615 	pte_offset_fixmap = pte_offset_late_fixmap;
1616 	pmd_clear(fixmap_pmd(va));
1617 	local_flush_tlb_kernel_page(va);
1618 
1619 	for (i = 0; i < __end_of_permanent_fixed_addresses; i++) {
1620 		pte_t *pte;
1621 		struct map_desc map;
1622 
1623 		map.virtual = fix_to_virt(i);
1624 		pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual);
1625 
1626 		/* Only i/o device mappings are supported ATM */
1627 		if (pte_none(*pte) ||
1628 		    (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED)
1629 			continue;
1630 
1631 		map.pfn = pte_pfn(*pte);
1632 		map.type = MT_DEVICE;
1633 		map.length = PAGE_SIZE;
1634 
1635 		create_mapping(&map);
1636 	}
1637 }
1638 
1639 /*
1640  * paging_init() sets up the page tables, initialises the zone memory
1641  * maps, and sets up the zero page, bad page and bad page tables.
1642  */
1643 void __init paging_init(const struct machine_desc *mdesc)
1644 {
1645 	void *zero_page;
1646 
1647 	prepare_page_table();
1648 	map_lowmem();
1649 	memblock_set_current_limit(arm_lowmem_limit);
1650 	dma_contiguous_remap();
1651 	early_fixmap_shutdown();
1652 	devicemaps_init(mdesc);
1653 	kmap_init();
1654 	tcm_init();
1655 
1656 	top_pmd = pmd_off_k(0xffff0000);
1657 
1658 	/* allocate the zero page. */
1659 	zero_page = early_alloc(PAGE_SIZE);
1660 
1661 	bootmem_init();
1662 
1663 	empty_zero_page = virt_to_page(zero_page);
1664 	__flush_dcache_page(NULL, empty_zero_page);
1665 
1666 	/* Compute the virt/idmap offset, mostly for the sake of KVM */
1667 	kimage_voffset = (unsigned long)&kimage_voffset - virt_to_idmap(&kimage_voffset);
1668 }
1669 
1670 void __init early_mm_init(const struct machine_desc *mdesc)
1671 {
1672 	build_mem_type_table();
1673 	early_paging_init(mdesc);
1674 }
1675