xref: /openbmc/linux/arch/arm/mm/mmu.c (revision 9b93eb47)
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 "fault.h"
42 #include "mm.h"
43 #include "tcm.h"
44 
45 /*
46  * empty_zero_page is a special page that is used for
47  * zero-initialized data and COW.
48  */
49 struct page *empty_zero_page;
50 EXPORT_SYMBOL(empty_zero_page);
51 
52 /*
53  * The pmd table for the upper-most set of pages.
54  */
55 pmd_t *top_pmd;
56 
57 pmdval_t user_pmd_table = _PAGE_USER_TABLE;
58 
59 #define CPOLICY_UNCACHED	0
60 #define CPOLICY_BUFFERED	1
61 #define CPOLICY_WRITETHROUGH	2
62 #define CPOLICY_WRITEBACK	3
63 #define CPOLICY_WRITEALLOC	4
64 
65 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
66 static unsigned int ecc_mask __initdata = 0;
67 pgprot_t pgprot_user;
68 pgprot_t pgprot_kernel;
69 pgprot_t pgprot_hyp_device;
70 pgprot_t pgprot_s2;
71 pgprot_t pgprot_s2_device;
72 
73 EXPORT_SYMBOL(pgprot_user);
74 EXPORT_SYMBOL(pgprot_kernel);
75 
76 struct cachepolicy {
77 	const char	policy[16];
78 	unsigned int	cr_mask;
79 	pmdval_t	pmd;
80 	pteval_t	pte;
81 	pteval_t	pte_s2;
82 };
83 
84 #ifdef CONFIG_ARM_LPAE
85 #define s2_policy(policy)	policy
86 #else
87 #define s2_policy(policy)	0
88 #endif
89 
90 unsigned long kimage_voffset __ro_after_init;
91 
92 static struct cachepolicy cache_policies[] __initdata = {
93 	{
94 		.policy		= "uncached",
95 		.cr_mask	= CR_W|CR_C,
96 		.pmd		= PMD_SECT_UNCACHED,
97 		.pte		= L_PTE_MT_UNCACHED,
98 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
99 	}, {
100 		.policy		= "buffered",
101 		.cr_mask	= CR_C,
102 		.pmd		= PMD_SECT_BUFFERED,
103 		.pte		= L_PTE_MT_BUFFERABLE,
104 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
105 	}, {
106 		.policy		= "writethrough",
107 		.cr_mask	= 0,
108 		.pmd		= PMD_SECT_WT,
109 		.pte		= L_PTE_MT_WRITETHROUGH,
110 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITETHROUGH),
111 	}, {
112 		.policy		= "writeback",
113 		.cr_mask	= 0,
114 		.pmd		= PMD_SECT_WB,
115 		.pte		= L_PTE_MT_WRITEBACK,
116 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
117 	}, {
118 		.policy		= "writealloc",
119 		.cr_mask	= 0,
120 		.pmd		= PMD_SECT_WBWA,
121 		.pte		= L_PTE_MT_WRITEALLOC,
122 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
123 	}
124 };
125 
126 #ifdef CONFIG_CPU_CP15
127 static unsigned long initial_pmd_value __initdata = 0;
128 
129 /*
130  * Initialise the cache_policy variable with the initial state specified
131  * via the "pmd" value.  This is used to ensure that on ARMv6 and later,
132  * the C code sets the page tables up with the same policy as the head
133  * assembly code, which avoids an illegal state where the TLBs can get
134  * confused.  See comments in early_cachepolicy() for more information.
135  */
136 void __init init_default_cache_policy(unsigned long pmd)
137 {
138 	int i;
139 
140 	initial_pmd_value = pmd;
141 
142 	pmd &= PMD_SECT_CACHE_MASK;
143 
144 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
145 		if (cache_policies[i].pmd == pmd) {
146 			cachepolicy = i;
147 			break;
148 		}
149 
150 	if (i == ARRAY_SIZE(cache_policies))
151 		pr_err("ERROR: could not find cache policy\n");
152 }
153 
154 /*
155  * These are useful for identifying cache coherency problems by allowing
156  * the cache or the cache and writebuffer to be turned off.  (Note: the
157  * write buffer should not be on and the cache off).
158  */
159 static int __init early_cachepolicy(char *p)
160 {
161 	int i, selected = -1;
162 
163 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
164 		int len = strlen(cache_policies[i].policy);
165 
166 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
167 			selected = i;
168 			break;
169 		}
170 	}
171 
172 	if (selected == -1)
173 		pr_err("ERROR: unknown or unsupported cache policy\n");
174 
175 	/*
176 	 * This restriction is partly to do with the way we boot; it is
177 	 * unpredictable to have memory mapped using two different sets of
178 	 * memory attributes (shared, type, and cache attribs).  We can not
179 	 * change these attributes once the initial assembly has setup the
180 	 * page tables.
181 	 */
182 	if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) {
183 		pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n",
184 			cache_policies[cachepolicy].policy);
185 		return 0;
186 	}
187 
188 	if (selected != cachepolicy) {
189 		unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
190 		cachepolicy = selected;
191 		flush_cache_all();
192 		set_cr(cr);
193 	}
194 	return 0;
195 }
196 early_param("cachepolicy", early_cachepolicy);
197 
198 static int __init early_nocache(char *__unused)
199 {
200 	char *p = "buffered";
201 	pr_warn("nocache is deprecated; use cachepolicy=%s\n", p);
202 	early_cachepolicy(p);
203 	return 0;
204 }
205 early_param("nocache", early_nocache);
206 
207 static int __init early_nowrite(char *__unused)
208 {
209 	char *p = "uncached";
210 	pr_warn("nowb is deprecated; use cachepolicy=%s\n", p);
211 	early_cachepolicy(p);
212 	return 0;
213 }
214 early_param("nowb", early_nowrite);
215 
216 #ifndef CONFIG_ARM_LPAE
217 static int __init early_ecc(char *p)
218 {
219 	if (memcmp(p, "on", 2) == 0)
220 		ecc_mask = PMD_PROTECTION;
221 	else if (memcmp(p, "off", 3) == 0)
222 		ecc_mask = 0;
223 	return 0;
224 }
225 early_param("ecc", early_ecc);
226 #endif
227 
228 #else /* ifdef CONFIG_CPU_CP15 */
229 
230 static int __init early_cachepolicy(char *p)
231 {
232 	pr_warn("cachepolicy kernel parameter not supported without cp15\n");
233 }
234 early_param("cachepolicy", early_cachepolicy);
235 
236 static int __init noalign_setup(char *__unused)
237 {
238 	pr_warn("noalign kernel parameter not supported without cp15\n");
239 }
240 __setup("noalign", noalign_setup);
241 
242 #endif /* ifdef CONFIG_CPU_CP15 / else */
243 
244 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
245 #define PROT_PTE_S2_DEVICE	PROT_PTE_DEVICE
246 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
247 
248 static struct mem_type mem_types[] __ro_after_init = {
249 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
250 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
251 				  L_PTE_SHARED,
252 		.prot_pte_s2	= s2_policy(PROT_PTE_S2_DEVICE) |
253 				  s2_policy(L_PTE_S2_MT_DEV_SHARED) |
254 				  L_PTE_SHARED,
255 		.prot_l1	= PMD_TYPE_TABLE,
256 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
257 		.domain		= DOMAIN_IO,
258 	},
259 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
260 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
261 		.prot_l1	= PMD_TYPE_TABLE,
262 		.prot_sect	= PROT_SECT_DEVICE,
263 		.domain		= DOMAIN_IO,
264 	},
265 	[MT_DEVICE_CACHED] = {	  /* ioremap_cached */
266 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
267 		.prot_l1	= PMD_TYPE_TABLE,
268 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
269 		.domain		= DOMAIN_IO,
270 	},
271 	[MT_DEVICE_WC] = {	/* ioremap_wc */
272 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
273 		.prot_l1	= PMD_TYPE_TABLE,
274 		.prot_sect	= PROT_SECT_DEVICE,
275 		.domain		= DOMAIN_IO,
276 	},
277 	[MT_UNCACHED] = {
278 		.prot_pte	= PROT_PTE_DEVICE,
279 		.prot_l1	= PMD_TYPE_TABLE,
280 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
281 		.domain		= DOMAIN_IO,
282 	},
283 	[MT_CACHECLEAN] = {
284 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
285 		.domain    = DOMAIN_KERNEL,
286 	},
287 #ifndef CONFIG_ARM_LPAE
288 	[MT_MINICLEAN] = {
289 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
290 		.domain    = DOMAIN_KERNEL,
291 	},
292 #endif
293 	[MT_LOW_VECTORS] = {
294 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
295 				L_PTE_RDONLY,
296 		.prot_l1   = PMD_TYPE_TABLE,
297 		.domain    = DOMAIN_VECTORS,
298 	},
299 	[MT_HIGH_VECTORS] = {
300 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
301 				L_PTE_USER | L_PTE_RDONLY,
302 		.prot_l1   = PMD_TYPE_TABLE,
303 		.domain    = DOMAIN_VECTORS,
304 	},
305 	[MT_MEMORY_RWX] = {
306 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
307 		.prot_l1   = PMD_TYPE_TABLE,
308 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
309 		.domain    = DOMAIN_KERNEL,
310 	},
311 	[MT_MEMORY_RW] = {
312 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
313 			     L_PTE_XN,
314 		.prot_l1   = PMD_TYPE_TABLE,
315 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
316 		.domain    = DOMAIN_KERNEL,
317 	},
318 	[MT_ROM] = {
319 		.prot_sect = PMD_TYPE_SECT,
320 		.domain    = DOMAIN_KERNEL,
321 	},
322 	[MT_MEMORY_RWX_NONCACHED] = {
323 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
324 				L_PTE_MT_BUFFERABLE,
325 		.prot_l1   = PMD_TYPE_TABLE,
326 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
327 		.domain    = DOMAIN_KERNEL,
328 	},
329 	[MT_MEMORY_RW_DTCM] = {
330 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
331 				L_PTE_XN,
332 		.prot_l1   = PMD_TYPE_TABLE,
333 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
334 		.domain    = DOMAIN_KERNEL,
335 	},
336 	[MT_MEMORY_RWX_ITCM] = {
337 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
338 		.prot_l1   = PMD_TYPE_TABLE,
339 		.domain    = DOMAIN_KERNEL,
340 	},
341 	[MT_MEMORY_RW_SO] = {
342 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
343 				L_PTE_MT_UNCACHED | L_PTE_XN,
344 		.prot_l1   = PMD_TYPE_TABLE,
345 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
346 				PMD_SECT_UNCACHED | PMD_SECT_XN,
347 		.domain    = DOMAIN_KERNEL,
348 	},
349 	[MT_MEMORY_DMA_READY] = {
350 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
351 				L_PTE_XN,
352 		.prot_l1   = PMD_TYPE_TABLE,
353 		.domain    = DOMAIN_KERNEL,
354 	},
355 };
356 
357 const struct mem_type *get_mem_type(unsigned int type)
358 {
359 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
360 }
361 EXPORT_SYMBOL(get_mem_type);
362 
363 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr);
364 
365 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS]
366 	__aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata;
367 
368 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr)
369 {
370 	return &bm_pte[pte_index(addr)];
371 }
372 
373 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr)
374 {
375 	return pte_offset_kernel(dir, addr);
376 }
377 
378 static inline pmd_t * __init fixmap_pmd(unsigned long addr)
379 {
380 	pgd_t *pgd = pgd_offset_k(addr);
381 	pud_t *pud = pud_offset(pgd, addr);
382 	pmd_t *pmd = pmd_offset(pud, addr);
383 
384 	return pmd;
385 }
386 
387 void __init early_fixmap_init(void)
388 {
389 	pmd_t *pmd;
390 
391 	/*
392 	 * The early fixmap range spans multiple pmds, for which
393 	 * we are not prepared:
394 	 */
395 	BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT)
396 		     != FIXADDR_TOP >> PMD_SHIFT);
397 
398 	pmd = fixmap_pmd(FIXADDR_TOP);
399 	pmd_populate_kernel(&init_mm, pmd, bm_pte);
400 
401 	pte_offset_fixmap = pte_offset_early_fixmap;
402 }
403 
404 /*
405  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
406  * As a result, this can only be called with preemption disabled, as under
407  * stop_machine().
408  */
409 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
410 {
411 	unsigned long vaddr = __fix_to_virt(idx);
412 	pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr);
413 
414 	/* Make sure fixmap region does not exceed available allocation. */
415 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
416 		     FIXADDR_END);
417 	BUG_ON(idx >= __end_of_fixed_addresses);
418 
419 	/* we only support device mappings until pgprot_kernel has been set */
420 	if (WARN_ON(pgprot_val(prot) != pgprot_val(FIXMAP_PAGE_IO) &&
421 		    pgprot_val(pgprot_kernel) == 0))
422 		return;
423 
424 	if (pgprot_val(prot))
425 		set_pte_at(NULL, vaddr, pte,
426 			pfn_pte(phys >> PAGE_SHIFT, prot));
427 	else
428 		pte_clear(NULL, vaddr, pte);
429 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
430 }
431 
432 /*
433  * Adjust the PMD section entries according to the CPU in use.
434  */
435 static void __init build_mem_type_table(void)
436 {
437 	struct cachepolicy *cp;
438 	unsigned int cr = get_cr();
439 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
440 	pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
441 	int cpu_arch = cpu_architecture();
442 	int i;
443 
444 	if (cpu_arch < CPU_ARCH_ARMv6) {
445 #if defined(CONFIG_CPU_DCACHE_DISABLE)
446 		if (cachepolicy > CPOLICY_BUFFERED)
447 			cachepolicy = CPOLICY_BUFFERED;
448 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
449 		if (cachepolicy > CPOLICY_WRITETHROUGH)
450 			cachepolicy = CPOLICY_WRITETHROUGH;
451 #endif
452 	}
453 	if (cpu_arch < CPU_ARCH_ARMv5) {
454 		if (cachepolicy >= CPOLICY_WRITEALLOC)
455 			cachepolicy = CPOLICY_WRITEBACK;
456 		ecc_mask = 0;
457 	}
458 
459 	if (is_smp()) {
460 		if (cachepolicy != CPOLICY_WRITEALLOC) {
461 			pr_warn("Forcing write-allocate cache policy for SMP\n");
462 			cachepolicy = CPOLICY_WRITEALLOC;
463 		}
464 		if (!(initial_pmd_value & PMD_SECT_S)) {
465 			pr_warn("Forcing shared mappings for SMP\n");
466 			initial_pmd_value |= PMD_SECT_S;
467 		}
468 	}
469 
470 	/*
471 	 * Strip out features not present on earlier architectures.
472 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
473 	 * without extended page tables don't have the 'Shared' bit.
474 	 */
475 	if (cpu_arch < CPU_ARCH_ARMv5)
476 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
477 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
478 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
479 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
480 			mem_types[i].prot_sect &= ~PMD_SECT_S;
481 
482 	/*
483 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
484 	 * "update-able on write" bit on ARM610).  However, Xscale and
485 	 * Xscale3 require this bit to be cleared.
486 	 */
487 	if (cpu_is_xscale_family()) {
488 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
489 			mem_types[i].prot_sect &= ~PMD_BIT4;
490 			mem_types[i].prot_l1 &= ~PMD_BIT4;
491 		}
492 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
493 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
494 			if (mem_types[i].prot_l1)
495 				mem_types[i].prot_l1 |= PMD_BIT4;
496 			if (mem_types[i].prot_sect)
497 				mem_types[i].prot_sect |= PMD_BIT4;
498 		}
499 	}
500 
501 	/*
502 	 * Mark the device areas according to the CPU/architecture.
503 	 */
504 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
505 		if (!cpu_is_xsc3()) {
506 			/*
507 			 * Mark device regions on ARMv6+ as execute-never
508 			 * to prevent speculative instruction fetches.
509 			 */
510 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
511 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
512 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
513 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
514 
515 			/* Also setup NX memory mapping */
516 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
517 		}
518 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
519 			/*
520 			 * For ARMv7 with TEX remapping,
521 			 * - shared device is SXCB=1100
522 			 * - nonshared device is SXCB=0100
523 			 * - write combine device mem is SXCB=0001
524 			 * (Uncached Normal memory)
525 			 */
526 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
527 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
528 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
529 		} else if (cpu_is_xsc3()) {
530 			/*
531 			 * For Xscale3,
532 			 * - shared device is TEXCB=00101
533 			 * - nonshared device is TEXCB=01000
534 			 * - write combine device mem is TEXCB=00100
535 			 * (Inner/Outer Uncacheable in xsc3 parlance)
536 			 */
537 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
538 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
539 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
540 		} else {
541 			/*
542 			 * For ARMv6 and ARMv7 without TEX remapping,
543 			 * - shared device is TEXCB=00001
544 			 * - nonshared device is TEXCB=01000
545 			 * - write combine device mem is TEXCB=00100
546 			 * (Uncached Normal in ARMv6 parlance).
547 			 */
548 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
549 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
550 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
551 		}
552 	} else {
553 		/*
554 		 * On others, write combining is "Uncached/Buffered"
555 		 */
556 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
557 	}
558 
559 	/*
560 	 * Now deal with the memory-type mappings
561 	 */
562 	cp = &cache_policies[cachepolicy];
563 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
564 	s2_pgprot = cp->pte_s2;
565 	hyp_device_pgprot = mem_types[MT_DEVICE].prot_pte;
566 	s2_device_pgprot = mem_types[MT_DEVICE].prot_pte_s2;
567 
568 #ifndef CONFIG_ARM_LPAE
569 	/*
570 	 * We don't use domains on ARMv6 (since this causes problems with
571 	 * v6/v7 kernels), so we must use a separate memory type for user
572 	 * r/o, kernel r/w to map the vectors page.
573 	 */
574 	if (cpu_arch == CPU_ARCH_ARMv6)
575 		vecs_pgprot |= L_PTE_MT_VECTORS;
576 
577 	/*
578 	 * Check is it with support for the PXN bit
579 	 * in the Short-descriptor translation table format descriptors.
580 	 */
581 	if (cpu_arch == CPU_ARCH_ARMv7 &&
582 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
583 		user_pmd_table |= PMD_PXNTABLE;
584 	}
585 #endif
586 
587 	/*
588 	 * ARMv6 and above have extended page tables.
589 	 */
590 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
591 #ifndef CONFIG_ARM_LPAE
592 		/*
593 		 * Mark cache clean areas and XIP ROM read only
594 		 * from SVC mode and no access from userspace.
595 		 */
596 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
597 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
598 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
599 #endif
600 
601 		/*
602 		 * If the initial page tables were created with the S bit
603 		 * set, then we need to do the same here for the same
604 		 * reasons given in early_cachepolicy().
605 		 */
606 		if (initial_pmd_value & PMD_SECT_S) {
607 			user_pgprot |= L_PTE_SHARED;
608 			kern_pgprot |= L_PTE_SHARED;
609 			vecs_pgprot |= L_PTE_SHARED;
610 			s2_pgprot |= L_PTE_SHARED;
611 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
612 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
613 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
614 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
615 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
616 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
617 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
618 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
619 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
620 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
621 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
622 		}
623 	}
624 
625 	/*
626 	 * Non-cacheable Normal - intended for memory areas that must
627 	 * not cause dirty cache line writebacks when used
628 	 */
629 	if (cpu_arch >= CPU_ARCH_ARMv6) {
630 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
631 			/* Non-cacheable Normal is XCB = 001 */
632 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
633 				PMD_SECT_BUFFERED;
634 		} else {
635 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
636 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
637 				PMD_SECT_TEX(1);
638 		}
639 	} else {
640 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
641 	}
642 
643 #ifdef CONFIG_ARM_LPAE
644 	/*
645 	 * Do not generate access flag faults for the kernel mappings.
646 	 */
647 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
648 		mem_types[i].prot_pte |= PTE_EXT_AF;
649 		if (mem_types[i].prot_sect)
650 			mem_types[i].prot_sect |= PMD_SECT_AF;
651 	}
652 	kern_pgprot |= PTE_EXT_AF;
653 	vecs_pgprot |= PTE_EXT_AF;
654 
655 	/*
656 	 * Set PXN for user mappings
657 	 */
658 	user_pgprot |= PTE_EXT_PXN;
659 #endif
660 
661 	for (i = 0; i < 16; i++) {
662 		pteval_t v = pgprot_val(protection_map[i]);
663 		protection_map[i] = __pgprot(v | user_pgprot);
664 	}
665 
666 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
667 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
668 
669 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
670 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
671 				 L_PTE_DIRTY | kern_pgprot);
672 	pgprot_s2  = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
673 	pgprot_s2_device  = __pgprot(s2_device_pgprot);
674 	pgprot_hyp_device  = __pgprot(hyp_device_pgprot);
675 
676 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
677 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
678 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
679 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
680 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
681 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
682 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
683 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
684 	mem_types[MT_ROM].prot_sect |= cp->pmd;
685 
686 	switch (cp->pmd) {
687 	case PMD_SECT_WT:
688 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
689 		break;
690 	case PMD_SECT_WB:
691 	case PMD_SECT_WBWA:
692 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
693 		break;
694 	}
695 	pr_info("Memory policy: %sData cache %s\n",
696 		ecc_mask ? "ECC enabled, " : "", cp->policy);
697 
698 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
699 		struct mem_type *t = &mem_types[i];
700 		if (t->prot_l1)
701 			t->prot_l1 |= PMD_DOMAIN(t->domain);
702 		if (t->prot_sect)
703 			t->prot_sect |= PMD_DOMAIN(t->domain);
704 	}
705 }
706 
707 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
708 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
709 			      unsigned long size, pgprot_t vma_prot)
710 {
711 	if (!pfn_valid(pfn))
712 		return pgprot_noncached(vma_prot);
713 	else if (file->f_flags & O_SYNC)
714 		return pgprot_writecombine(vma_prot);
715 	return vma_prot;
716 }
717 EXPORT_SYMBOL(phys_mem_access_prot);
718 #endif
719 
720 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
721 
722 static void __init *early_alloc(unsigned long sz)
723 {
724 	void *ptr = memblock_alloc(sz, sz);
725 
726 	if (!ptr)
727 		panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
728 		      __func__, sz, sz);
729 
730 	return ptr;
731 }
732 
733 static void *__init late_alloc(unsigned long sz)
734 {
735 	void *ptr = (void *)__get_free_pages(PGALLOC_GFP, get_order(sz));
736 
737 	if (!ptr || !pgtable_page_ctor(virt_to_page(ptr)))
738 		BUG();
739 	return ptr;
740 }
741 
742 static pte_t * __init arm_pte_alloc(pmd_t *pmd, unsigned long addr,
743 				unsigned long prot,
744 				void *(*alloc)(unsigned long sz))
745 {
746 	if (pmd_none(*pmd)) {
747 		pte_t *pte = alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
748 		__pmd_populate(pmd, __pa(pte), prot);
749 	}
750 	BUG_ON(pmd_bad(*pmd));
751 	return pte_offset_kernel(pmd, addr);
752 }
753 
754 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr,
755 				      unsigned long prot)
756 {
757 	return arm_pte_alloc(pmd, addr, prot, early_alloc);
758 }
759 
760 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
761 				  unsigned long end, unsigned long pfn,
762 				  const struct mem_type *type,
763 				  void *(*alloc)(unsigned long sz),
764 				  bool ng)
765 {
766 	pte_t *pte = arm_pte_alloc(pmd, addr, type->prot_l1, alloc);
767 	do {
768 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)),
769 			    ng ? PTE_EXT_NG : 0);
770 		pfn++;
771 	} while (pte++, addr += PAGE_SIZE, addr != end);
772 }
773 
774 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
775 			unsigned long end, phys_addr_t phys,
776 			const struct mem_type *type, bool ng)
777 {
778 	pmd_t *p = pmd;
779 
780 #ifndef CONFIG_ARM_LPAE
781 	/*
782 	 * In classic MMU format, puds and pmds are folded in to
783 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
784 	 * group of L1 entries making up one logical pointer to
785 	 * an L2 table (2MB), where as PMDs refer to the individual
786 	 * L1 entries (1MB). Hence increment to get the correct
787 	 * offset for odd 1MB sections.
788 	 * (See arch/arm/include/asm/pgtable-2level.h)
789 	 */
790 	if (addr & SECTION_SIZE)
791 		pmd++;
792 #endif
793 	do {
794 		*pmd = __pmd(phys | type->prot_sect | (ng ? PMD_SECT_nG : 0));
795 		phys += SECTION_SIZE;
796 	} while (pmd++, addr += SECTION_SIZE, addr != end);
797 
798 	flush_pmd_entry(p);
799 }
800 
801 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
802 				      unsigned long end, phys_addr_t phys,
803 				      const struct mem_type *type,
804 				      void *(*alloc)(unsigned long sz), bool ng)
805 {
806 	pmd_t *pmd = pmd_offset(pud, addr);
807 	unsigned long next;
808 
809 	do {
810 		/*
811 		 * With LPAE, we must loop over to map
812 		 * all the pmds for the given range.
813 		 */
814 		next = pmd_addr_end(addr, end);
815 
816 		/*
817 		 * Try a section mapping - addr, next and phys must all be
818 		 * aligned to a section boundary.
819 		 */
820 		if (type->prot_sect &&
821 				((addr | next | phys) & ~SECTION_MASK) == 0) {
822 			__map_init_section(pmd, addr, next, phys, type, ng);
823 		} else {
824 			alloc_init_pte(pmd, addr, next,
825 				       __phys_to_pfn(phys), type, alloc, ng);
826 		}
827 
828 		phys += next - addr;
829 
830 	} while (pmd++, addr = next, addr != end);
831 }
832 
833 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
834 				  unsigned long end, phys_addr_t phys,
835 				  const struct mem_type *type,
836 				  void *(*alloc)(unsigned long sz), bool ng)
837 {
838 	pud_t *pud = pud_offset(pgd, addr);
839 	unsigned long next;
840 
841 	do {
842 		next = pud_addr_end(addr, end);
843 		alloc_init_pmd(pud, addr, next, phys, type, alloc, ng);
844 		phys += next - addr;
845 	} while (pud++, addr = next, addr != end);
846 }
847 
848 #ifndef CONFIG_ARM_LPAE
849 static void __init create_36bit_mapping(struct mm_struct *mm,
850 					struct map_desc *md,
851 					const struct mem_type *type,
852 					bool ng)
853 {
854 	unsigned long addr, length, end;
855 	phys_addr_t phys;
856 	pgd_t *pgd;
857 
858 	addr = md->virtual;
859 	phys = __pfn_to_phys(md->pfn);
860 	length = PAGE_ALIGN(md->length);
861 
862 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
863 		pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
864 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
865 		return;
866 	}
867 
868 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
869 	 *	Since domain assignments can in fact be arbitrary, the
870 	 *	'domain == 0' check below is required to insure that ARMv6
871 	 *	supersections are only allocated for domain 0 regardless
872 	 *	of the actual domain assignments in use.
873 	 */
874 	if (type->domain) {
875 		pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
876 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
877 		return;
878 	}
879 
880 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
881 		pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
882 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
883 		return;
884 	}
885 
886 	/*
887 	 * Shift bits [35:32] of address into bits [23:20] of PMD
888 	 * (See ARMv6 spec).
889 	 */
890 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
891 
892 	pgd = pgd_offset(mm, addr);
893 	end = addr + length;
894 	do {
895 		pud_t *pud = pud_offset(pgd, addr);
896 		pmd_t *pmd = pmd_offset(pud, addr);
897 		int i;
898 
899 		for (i = 0; i < 16; i++)
900 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER |
901 				       (ng ? PMD_SECT_nG : 0));
902 
903 		addr += SUPERSECTION_SIZE;
904 		phys += SUPERSECTION_SIZE;
905 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
906 	} while (addr != end);
907 }
908 #endif	/* !CONFIG_ARM_LPAE */
909 
910 static void __init __create_mapping(struct mm_struct *mm, struct map_desc *md,
911 				    void *(*alloc)(unsigned long sz),
912 				    bool ng)
913 {
914 	unsigned long addr, length, end;
915 	phys_addr_t phys;
916 	const struct mem_type *type;
917 	pgd_t *pgd;
918 
919 	type = &mem_types[md->type];
920 
921 #ifndef CONFIG_ARM_LPAE
922 	/*
923 	 * Catch 36-bit addresses
924 	 */
925 	if (md->pfn >= 0x100000) {
926 		create_36bit_mapping(mm, md, type, ng);
927 		return;
928 	}
929 #endif
930 
931 	addr = md->virtual & PAGE_MASK;
932 	phys = __pfn_to_phys(md->pfn);
933 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
934 
935 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
936 		pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
937 			(long long)__pfn_to_phys(md->pfn), addr);
938 		return;
939 	}
940 
941 	pgd = pgd_offset(mm, addr);
942 	end = addr + length;
943 	do {
944 		unsigned long next = pgd_addr_end(addr, end);
945 
946 		alloc_init_pud(pgd, addr, next, phys, type, alloc, ng);
947 
948 		phys += next - addr;
949 		addr = next;
950 	} while (pgd++, addr != end);
951 }
952 
953 /*
954  * Create the page directory entries and any necessary
955  * page tables for the mapping specified by `md'.  We
956  * are able to cope here with varying sizes and address
957  * offsets, and we take full advantage of sections and
958  * supersections.
959  */
960 static void __init create_mapping(struct map_desc *md)
961 {
962 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
963 		pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
964 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
965 		return;
966 	}
967 
968 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
969 	    md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START &&
970 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
971 		pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
972 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
973 	}
974 
975 	__create_mapping(&init_mm, md, early_alloc, false);
976 }
977 
978 void __init create_mapping_late(struct mm_struct *mm, struct map_desc *md,
979 				bool ng)
980 {
981 #ifdef CONFIG_ARM_LPAE
982 	pud_t *pud = pud_alloc(mm, pgd_offset(mm, md->virtual), md->virtual);
983 	if (WARN_ON(!pud))
984 		return;
985 	pmd_alloc(mm, pud, 0);
986 #endif
987 	__create_mapping(mm, md, late_alloc, ng);
988 }
989 
990 /*
991  * Create the architecture specific mappings
992  */
993 void __init iotable_init(struct map_desc *io_desc, int nr)
994 {
995 	struct map_desc *md;
996 	struct vm_struct *vm;
997 	struct static_vm *svm;
998 
999 	if (!nr)
1000 		return;
1001 
1002 	svm = memblock_alloc(sizeof(*svm) * nr, __alignof__(*svm));
1003 	if (!svm)
1004 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1005 		      __func__, sizeof(*svm) * nr, __alignof__(*svm));
1006 
1007 	for (md = io_desc; nr; md++, nr--) {
1008 		create_mapping(md);
1009 
1010 		vm = &svm->vm;
1011 		vm->addr = (void *)(md->virtual & PAGE_MASK);
1012 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
1013 		vm->phys_addr = __pfn_to_phys(md->pfn);
1014 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
1015 		vm->flags |= VM_ARM_MTYPE(md->type);
1016 		vm->caller = iotable_init;
1017 		add_static_vm_early(svm++);
1018 	}
1019 }
1020 
1021 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
1022 				  void *caller)
1023 {
1024 	struct vm_struct *vm;
1025 	struct static_vm *svm;
1026 
1027 	svm = memblock_alloc(sizeof(*svm), __alignof__(*svm));
1028 	if (!svm)
1029 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1030 		      __func__, sizeof(*svm), __alignof__(*svm));
1031 
1032 	vm = &svm->vm;
1033 	vm->addr = (void *)addr;
1034 	vm->size = size;
1035 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
1036 	vm->caller = caller;
1037 	add_static_vm_early(svm);
1038 }
1039 
1040 #ifndef CONFIG_ARM_LPAE
1041 
1042 /*
1043  * The Linux PMD is made of two consecutive section entries covering 2MB
1044  * (see definition in include/asm/pgtable-2level.h).  However a call to
1045  * create_mapping() may optimize static mappings by using individual
1046  * 1MB section mappings.  This leaves the actual PMD potentially half
1047  * initialized if the top or bottom section entry isn't used, leaving it
1048  * open to problems if a subsequent ioremap() or vmalloc() tries to use
1049  * the virtual space left free by that unused section entry.
1050  *
1051  * Let's avoid the issue by inserting dummy vm entries covering the unused
1052  * PMD halves once the static mappings are in place.
1053  */
1054 
1055 static void __init pmd_empty_section_gap(unsigned long addr)
1056 {
1057 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1058 }
1059 
1060 static void __init fill_pmd_gaps(void)
1061 {
1062 	struct static_vm *svm;
1063 	struct vm_struct *vm;
1064 	unsigned long addr, next = 0;
1065 	pmd_t *pmd;
1066 
1067 	list_for_each_entry(svm, &static_vmlist, list) {
1068 		vm = &svm->vm;
1069 		addr = (unsigned long)vm->addr;
1070 		if (addr < next)
1071 			continue;
1072 
1073 		/*
1074 		 * Check if this vm starts on an odd section boundary.
1075 		 * If so and the first section entry for this PMD is free
1076 		 * then we block the corresponding virtual address.
1077 		 */
1078 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1079 			pmd = pmd_off_k(addr);
1080 			if (pmd_none(*pmd))
1081 				pmd_empty_section_gap(addr & PMD_MASK);
1082 		}
1083 
1084 		/*
1085 		 * Then check if this vm ends on an odd section boundary.
1086 		 * If so and the second section entry for this PMD is empty
1087 		 * then we block the corresponding virtual address.
1088 		 */
1089 		addr += vm->size;
1090 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1091 			pmd = pmd_off_k(addr) + 1;
1092 			if (pmd_none(*pmd))
1093 				pmd_empty_section_gap(addr);
1094 		}
1095 
1096 		/* no need to look at any vm entry until we hit the next PMD */
1097 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1098 	}
1099 }
1100 
1101 #else
1102 #define fill_pmd_gaps() do { } while (0)
1103 #endif
1104 
1105 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1106 static void __init pci_reserve_io(void)
1107 {
1108 	struct static_vm *svm;
1109 
1110 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1111 	if (svm)
1112 		return;
1113 
1114 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1115 }
1116 #else
1117 #define pci_reserve_io() do { } while (0)
1118 #endif
1119 
1120 #ifdef CONFIG_DEBUG_LL
1121 void __init debug_ll_io_init(void)
1122 {
1123 	struct map_desc map;
1124 
1125 	debug_ll_addr(&map.pfn, &map.virtual);
1126 	if (!map.pfn || !map.virtual)
1127 		return;
1128 	map.pfn = __phys_to_pfn(map.pfn);
1129 	map.virtual &= PAGE_MASK;
1130 	map.length = PAGE_SIZE;
1131 	map.type = MT_DEVICE;
1132 	iotable_init(&map, 1);
1133 }
1134 #endif
1135 
1136 static void * __initdata vmalloc_min =
1137 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1138 
1139 /*
1140  * vmalloc=size forces the vmalloc area to be exactly 'size'
1141  * bytes. This can be used to increase (or decrease) the vmalloc
1142  * area - the default is 240m.
1143  */
1144 static int __init early_vmalloc(char *arg)
1145 {
1146 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1147 
1148 	if (vmalloc_reserve < SZ_16M) {
1149 		vmalloc_reserve = SZ_16M;
1150 		pr_warn("vmalloc area too small, limiting to %luMB\n",
1151 			vmalloc_reserve >> 20);
1152 	}
1153 
1154 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1155 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1156 		pr_warn("vmalloc area is too big, limiting to %luMB\n",
1157 			vmalloc_reserve >> 20);
1158 	}
1159 
1160 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1161 	return 0;
1162 }
1163 early_param("vmalloc", early_vmalloc);
1164 
1165 phys_addr_t arm_lowmem_limit __initdata = 0;
1166 
1167 void __init adjust_lowmem_bounds(void)
1168 {
1169 	phys_addr_t memblock_limit = 0;
1170 	u64 vmalloc_limit;
1171 	struct memblock_region *reg;
1172 	phys_addr_t lowmem_limit = 0;
1173 
1174 	/*
1175 	 * Let's use our own (unoptimized) equivalent of __pa() that is
1176 	 * not affected by wrap-arounds when sizeof(phys_addr_t) == 4.
1177 	 * The result is used as the upper bound on physical memory address
1178 	 * and may itself be outside the valid range for which phys_addr_t
1179 	 * and therefore __pa() is defined.
1180 	 */
1181 	vmalloc_limit = (u64)(uintptr_t)vmalloc_min - PAGE_OFFSET + PHYS_OFFSET;
1182 
1183 	for_each_memblock(memory, reg) {
1184 		phys_addr_t block_start = reg->base;
1185 		phys_addr_t block_end = reg->base + reg->size;
1186 
1187 		if (reg->base < vmalloc_limit) {
1188 			if (block_end > lowmem_limit)
1189 				/*
1190 				 * Compare as u64 to ensure vmalloc_limit does
1191 				 * not get truncated. block_end should always
1192 				 * fit in phys_addr_t so there should be no
1193 				 * issue with assignment.
1194 				 */
1195 				lowmem_limit = min_t(u64,
1196 							 vmalloc_limit,
1197 							 block_end);
1198 
1199 			/*
1200 			 * Find the first non-pmd-aligned page, and point
1201 			 * memblock_limit at it. This relies on rounding the
1202 			 * limit down to be pmd-aligned, which happens at the
1203 			 * end of this function.
1204 			 *
1205 			 * With this algorithm, the start or end of almost any
1206 			 * bank can be non-pmd-aligned. The only exception is
1207 			 * that the start of the bank 0 must be section-
1208 			 * aligned, since otherwise memory would need to be
1209 			 * allocated when mapping the start of bank 0, which
1210 			 * occurs before any free memory is mapped.
1211 			 */
1212 			if (!memblock_limit) {
1213 				if (!IS_ALIGNED(block_start, PMD_SIZE))
1214 					memblock_limit = block_start;
1215 				else if (!IS_ALIGNED(block_end, PMD_SIZE))
1216 					memblock_limit = lowmem_limit;
1217 			}
1218 
1219 		}
1220 	}
1221 
1222 	arm_lowmem_limit = lowmem_limit;
1223 
1224 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1225 
1226 	if (!memblock_limit)
1227 		memblock_limit = arm_lowmem_limit;
1228 
1229 	/*
1230 	 * Round the memblock limit down to a pmd size.  This
1231 	 * helps to ensure that we will allocate memory from the
1232 	 * last full pmd, which should be mapped.
1233 	 */
1234 	memblock_limit = round_down(memblock_limit, PMD_SIZE);
1235 
1236 	if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1237 		if (memblock_end_of_DRAM() > arm_lowmem_limit) {
1238 			phys_addr_t end = memblock_end_of_DRAM();
1239 
1240 			pr_notice("Ignoring RAM at %pa-%pa\n",
1241 				  &memblock_limit, &end);
1242 			pr_notice("Consider using a HIGHMEM enabled kernel.\n");
1243 
1244 			memblock_remove(memblock_limit, end - memblock_limit);
1245 		}
1246 	}
1247 
1248 	memblock_set_current_limit(memblock_limit);
1249 }
1250 
1251 static inline void prepare_page_table(void)
1252 {
1253 	unsigned long addr;
1254 	phys_addr_t end;
1255 
1256 	/*
1257 	 * Clear out all the mappings below the kernel image.
1258 	 */
1259 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1260 		pmd_clear(pmd_off_k(addr));
1261 
1262 #ifdef CONFIG_XIP_KERNEL
1263 	/* The XIP kernel is mapped in the module area -- skip over it */
1264 	addr = ((unsigned long)_exiprom + PMD_SIZE - 1) & PMD_MASK;
1265 #endif
1266 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1267 		pmd_clear(pmd_off_k(addr));
1268 
1269 	/*
1270 	 * Find the end of the first block of lowmem.
1271 	 */
1272 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1273 	if (end >= arm_lowmem_limit)
1274 		end = arm_lowmem_limit;
1275 
1276 	/*
1277 	 * Clear out all the kernel space mappings, except for the first
1278 	 * memory bank, up to the vmalloc region.
1279 	 */
1280 	for (addr = __phys_to_virt(end);
1281 	     addr < VMALLOC_START; addr += PMD_SIZE)
1282 		pmd_clear(pmd_off_k(addr));
1283 }
1284 
1285 #ifdef CONFIG_ARM_LPAE
1286 /* the first page is reserved for pgd */
1287 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1288 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1289 #else
1290 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1291 #endif
1292 
1293 /*
1294  * Reserve the special regions of memory
1295  */
1296 void __init arm_mm_memblock_reserve(void)
1297 {
1298 	/*
1299 	 * Reserve the page tables.  These are already in use,
1300 	 * and can only be in node 0.
1301 	 */
1302 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1303 
1304 #ifdef CONFIG_SA1111
1305 	/*
1306 	 * Because of the SA1111 DMA bug, we want to preserve our
1307 	 * precious DMA-able memory...
1308 	 */
1309 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1310 #endif
1311 }
1312 
1313 /*
1314  * Set up the device mappings.  Since we clear out the page tables for all
1315  * mappings above VMALLOC_START, except early fixmap, we might remove debug
1316  * device mappings.  This means earlycon can be used to debug this function
1317  * Any other function or debugging method which may touch any device _will_
1318  * crash the kernel.
1319  */
1320 static void __init devicemaps_init(const struct machine_desc *mdesc)
1321 {
1322 	struct map_desc map;
1323 	unsigned long addr;
1324 	void *vectors;
1325 
1326 	/*
1327 	 * Allocate the vector page early.
1328 	 */
1329 	vectors = early_alloc(PAGE_SIZE * 2);
1330 
1331 	early_trap_init(vectors);
1332 
1333 	/*
1334 	 * Clear page table except top pmd used by early fixmaps
1335 	 */
1336 	for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE)
1337 		pmd_clear(pmd_off_k(addr));
1338 
1339 	/*
1340 	 * Map the kernel if it is XIP.
1341 	 * It is always first in the modulearea.
1342 	 */
1343 #ifdef CONFIG_XIP_KERNEL
1344 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1345 	map.virtual = MODULES_VADDR;
1346 	map.length = ((unsigned long)_exiprom - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1347 	map.type = MT_ROM;
1348 	create_mapping(&map);
1349 #endif
1350 
1351 	/*
1352 	 * Map the cache flushing regions.
1353 	 */
1354 #ifdef FLUSH_BASE
1355 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1356 	map.virtual = FLUSH_BASE;
1357 	map.length = SZ_1M;
1358 	map.type = MT_CACHECLEAN;
1359 	create_mapping(&map);
1360 #endif
1361 #ifdef FLUSH_BASE_MINICACHE
1362 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1363 	map.virtual = FLUSH_BASE_MINICACHE;
1364 	map.length = SZ_1M;
1365 	map.type = MT_MINICLEAN;
1366 	create_mapping(&map);
1367 #endif
1368 
1369 	/*
1370 	 * Create a mapping for the machine vectors at the high-vectors
1371 	 * location (0xffff0000).  If we aren't using high-vectors, also
1372 	 * create a mapping at the low-vectors virtual address.
1373 	 */
1374 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1375 	map.virtual = 0xffff0000;
1376 	map.length = PAGE_SIZE;
1377 #ifdef CONFIG_KUSER_HELPERS
1378 	map.type = MT_HIGH_VECTORS;
1379 #else
1380 	map.type = MT_LOW_VECTORS;
1381 #endif
1382 	create_mapping(&map);
1383 
1384 	if (!vectors_high()) {
1385 		map.virtual = 0;
1386 		map.length = PAGE_SIZE * 2;
1387 		map.type = MT_LOW_VECTORS;
1388 		create_mapping(&map);
1389 	}
1390 
1391 	/* Now create a kernel read-only mapping */
1392 	map.pfn += 1;
1393 	map.virtual = 0xffff0000 + PAGE_SIZE;
1394 	map.length = PAGE_SIZE;
1395 	map.type = MT_LOW_VECTORS;
1396 	create_mapping(&map);
1397 
1398 	/*
1399 	 * Ask the machine support to map in the statically mapped devices.
1400 	 */
1401 	if (mdesc->map_io)
1402 		mdesc->map_io();
1403 	else
1404 		debug_ll_io_init();
1405 	fill_pmd_gaps();
1406 
1407 	/* Reserve fixed i/o space in VMALLOC region */
1408 	pci_reserve_io();
1409 
1410 	/*
1411 	 * Finally flush the caches and tlb to ensure that we're in a
1412 	 * consistent state wrt the writebuffer.  This also ensures that
1413 	 * any write-allocated cache lines in the vector page are written
1414 	 * back.  After this point, we can start to touch devices again.
1415 	 */
1416 	local_flush_tlb_all();
1417 	flush_cache_all();
1418 
1419 	/* Enable asynchronous aborts */
1420 	early_abt_enable();
1421 }
1422 
1423 static void __init kmap_init(void)
1424 {
1425 #ifdef CONFIG_HIGHMEM
1426 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1427 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1428 #endif
1429 
1430 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1431 			_PAGE_KERNEL_TABLE);
1432 }
1433 
1434 static void __init map_lowmem(void)
1435 {
1436 	struct memblock_region *reg;
1437 	phys_addr_t kernel_x_start = round_down(__pa(KERNEL_START), SECTION_SIZE);
1438 	phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1439 
1440 	/* Map all the lowmem memory banks. */
1441 	for_each_memblock(memory, reg) {
1442 		phys_addr_t start = reg->base;
1443 		phys_addr_t end = start + reg->size;
1444 		struct map_desc map;
1445 
1446 		if (memblock_is_nomap(reg))
1447 			continue;
1448 
1449 		if (end > arm_lowmem_limit)
1450 			end = arm_lowmem_limit;
1451 		if (start >= end)
1452 			break;
1453 
1454 		if (end < kernel_x_start) {
1455 			map.pfn = __phys_to_pfn(start);
1456 			map.virtual = __phys_to_virt(start);
1457 			map.length = end - start;
1458 			map.type = MT_MEMORY_RWX;
1459 
1460 			create_mapping(&map);
1461 		} else if (start >= kernel_x_end) {
1462 			map.pfn = __phys_to_pfn(start);
1463 			map.virtual = __phys_to_virt(start);
1464 			map.length = end - start;
1465 			map.type = MT_MEMORY_RW;
1466 
1467 			create_mapping(&map);
1468 		} else {
1469 			/* This better cover the entire kernel */
1470 			if (start < kernel_x_start) {
1471 				map.pfn = __phys_to_pfn(start);
1472 				map.virtual = __phys_to_virt(start);
1473 				map.length = kernel_x_start - start;
1474 				map.type = MT_MEMORY_RW;
1475 
1476 				create_mapping(&map);
1477 			}
1478 
1479 			map.pfn = __phys_to_pfn(kernel_x_start);
1480 			map.virtual = __phys_to_virt(kernel_x_start);
1481 			map.length = kernel_x_end - kernel_x_start;
1482 			map.type = MT_MEMORY_RWX;
1483 
1484 			create_mapping(&map);
1485 
1486 			if (kernel_x_end < end) {
1487 				map.pfn = __phys_to_pfn(kernel_x_end);
1488 				map.virtual = __phys_to_virt(kernel_x_end);
1489 				map.length = end - kernel_x_end;
1490 				map.type = MT_MEMORY_RW;
1491 
1492 				create_mapping(&map);
1493 			}
1494 		}
1495 	}
1496 }
1497 
1498 #ifdef CONFIG_ARM_PV_FIXUP
1499 extern unsigned long __atags_pointer;
1500 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata);
1501 pgtables_remap lpae_pgtables_remap_asm;
1502 
1503 /*
1504  * early_paging_init() recreates boot time page table setup, allowing machines
1505  * to switch over to a high (>4G) address space on LPAE systems
1506  */
1507 static void __init early_paging_init(const struct machine_desc *mdesc)
1508 {
1509 	pgtables_remap *lpae_pgtables_remap;
1510 	unsigned long pa_pgd;
1511 	unsigned int cr, ttbcr;
1512 	long long offset;
1513 	void *boot_data;
1514 
1515 	if (!mdesc->pv_fixup)
1516 		return;
1517 
1518 	offset = mdesc->pv_fixup();
1519 	if (offset == 0)
1520 		return;
1521 
1522 	/*
1523 	 * Get the address of the remap function in the 1:1 identity
1524 	 * mapping setup by the early page table assembly code.  We
1525 	 * must get this prior to the pv update.  The following barrier
1526 	 * ensures that this is complete before we fixup any P:V offsets.
1527 	 */
1528 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1529 	pa_pgd = __pa(swapper_pg_dir);
1530 	boot_data = __va(__atags_pointer);
1531 	barrier();
1532 
1533 	pr_info("Switching physical address space to 0x%08llx\n",
1534 		(u64)PHYS_OFFSET + offset);
1535 
1536 	/* Re-set the phys pfn offset, and the pv offset */
1537 	__pv_offset += offset;
1538 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1539 
1540 	/* Run the patch stub to update the constants */
1541 	fixup_pv_table(&__pv_table_begin,
1542 		(&__pv_table_end - &__pv_table_begin) << 2);
1543 
1544 	/*
1545 	 * We changing not only the virtual to physical mapping, but also
1546 	 * the physical addresses used to access memory.  We need to flush
1547 	 * all levels of cache in the system with caching disabled to
1548 	 * ensure that all data is written back, and nothing is prefetched
1549 	 * into the caches.  We also need to prevent the TLB walkers
1550 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1551 	 * specific.
1552 	 */
1553 	cr = get_cr();
1554 	set_cr(cr & ~(CR_I | CR_C));
1555 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1556 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1557 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1558 	flush_cache_all();
1559 
1560 	/*
1561 	 * Fixup the page tables - this must be in the idmap region as
1562 	 * we need to disable the MMU to do this safely, and hence it
1563 	 * needs to be assembly.  It's fairly simple, as we're using the
1564 	 * temporary tables setup by the initial assembly code.
1565 	 */
1566 	lpae_pgtables_remap(offset, pa_pgd, boot_data);
1567 
1568 	/* Re-enable the caches and cacheable TLB walks */
1569 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1570 	set_cr(cr);
1571 }
1572 
1573 #else
1574 
1575 static void __init early_paging_init(const struct machine_desc *mdesc)
1576 {
1577 	long long offset;
1578 
1579 	if (!mdesc->pv_fixup)
1580 		return;
1581 
1582 	offset = mdesc->pv_fixup();
1583 	if (offset == 0)
1584 		return;
1585 
1586 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1587 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1588 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1589 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1590 }
1591 
1592 #endif
1593 
1594 static void __init early_fixmap_shutdown(void)
1595 {
1596 	int i;
1597 	unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1);
1598 
1599 	pte_offset_fixmap = pte_offset_late_fixmap;
1600 	pmd_clear(fixmap_pmd(va));
1601 	local_flush_tlb_kernel_page(va);
1602 
1603 	for (i = 0; i < __end_of_permanent_fixed_addresses; i++) {
1604 		pte_t *pte;
1605 		struct map_desc map;
1606 
1607 		map.virtual = fix_to_virt(i);
1608 		pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual);
1609 
1610 		/* Only i/o device mappings are supported ATM */
1611 		if (pte_none(*pte) ||
1612 		    (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED)
1613 			continue;
1614 
1615 		map.pfn = pte_pfn(*pte);
1616 		map.type = MT_DEVICE;
1617 		map.length = PAGE_SIZE;
1618 
1619 		create_mapping(&map);
1620 	}
1621 }
1622 
1623 /*
1624  * paging_init() sets up the page tables, initialises the zone memory
1625  * maps, and sets up the zero page, bad page and bad page tables.
1626  */
1627 void __init paging_init(const struct machine_desc *mdesc)
1628 {
1629 	void *zero_page;
1630 
1631 	prepare_page_table();
1632 	map_lowmem();
1633 	memblock_set_current_limit(arm_lowmem_limit);
1634 	dma_contiguous_remap();
1635 	early_fixmap_shutdown();
1636 	devicemaps_init(mdesc);
1637 	kmap_init();
1638 	tcm_init();
1639 
1640 	top_pmd = pmd_off_k(0xffff0000);
1641 
1642 	/* allocate the zero page. */
1643 	zero_page = early_alloc(PAGE_SIZE);
1644 
1645 	bootmem_init();
1646 
1647 	empty_zero_page = virt_to_page(zero_page);
1648 	__flush_dcache_page(NULL, empty_zero_page);
1649 
1650 	/* Compute the virt/idmap offset, mostly for the sake of KVM */
1651 	kimage_voffset = (unsigned long)&kimage_voffset - virt_to_idmap(&kimage_voffset);
1652 }
1653 
1654 void __init early_mm_init(const struct machine_desc *mdesc)
1655 {
1656 	build_mem_type_table();
1657 	early_paging_init(mdesc);
1658 }
1659