xref: /linux/arch/arm/mm/mmu.c (revision 2634682fdffd9ba6e74b76be8aa91cf8b2e05c41)
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 #include <asm/pgalloc.h>
33 
34 #include <asm/mach/arch.h>
35 #include <asm/mach/map.h>
36 #include <asm/mach/pci.h>
37 #include <asm/fixmap.h>
38 
39 #include "fault.h"
40 #include "mm.h"
41 #include "tcm.h"
42 
43 /*
44  * empty_zero_page is a special page that is used for
45  * zero-initialized data and COW.
46  */
47 struct page *empty_zero_page;
48 EXPORT_SYMBOL(empty_zero_page);
49 
50 /*
51  * The pmd table for the upper-most set of pages.
52  */
53 pmd_t *top_pmd;
54 
55 pmdval_t user_pmd_table = _PAGE_USER_TABLE;
56 
57 #define CPOLICY_UNCACHED	0
58 #define CPOLICY_BUFFERED	1
59 #define CPOLICY_WRITETHROUGH	2
60 #define CPOLICY_WRITEBACK	3
61 #define CPOLICY_WRITEALLOC	4
62 
63 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
64 static unsigned int ecc_mask __initdata = 0;
65 pgprot_t pgprot_user;
66 pgprot_t pgprot_kernel;
67 
68 EXPORT_SYMBOL(pgprot_user);
69 EXPORT_SYMBOL(pgprot_kernel);
70 
71 struct cachepolicy {
72 	const char	policy[16];
73 	unsigned int	cr_mask;
74 	pmdval_t	pmd;
75 	pteval_t	pte;
76 };
77 
78 static struct cachepolicy cache_policies[] __initdata = {
79 	{
80 		.policy		= "uncached",
81 		.cr_mask	= CR_W|CR_C,
82 		.pmd		= PMD_SECT_UNCACHED,
83 		.pte		= L_PTE_MT_UNCACHED,
84 	}, {
85 		.policy		= "buffered",
86 		.cr_mask	= CR_C,
87 		.pmd		= PMD_SECT_BUFFERED,
88 		.pte		= L_PTE_MT_BUFFERABLE,
89 	}, {
90 		.policy		= "writethrough",
91 		.cr_mask	= 0,
92 		.pmd		= PMD_SECT_WT,
93 		.pte		= L_PTE_MT_WRITETHROUGH,
94 	}, {
95 		.policy		= "writeback",
96 		.cr_mask	= 0,
97 		.pmd		= PMD_SECT_WB,
98 		.pte		= L_PTE_MT_WRITEBACK,
99 	}, {
100 		.policy		= "writealloc",
101 		.cr_mask	= 0,
102 		.pmd		= PMD_SECT_WBWA,
103 		.pte		= L_PTE_MT_WRITEALLOC,
104 	}
105 };
106 
107 #ifdef CONFIG_CPU_CP15
108 static unsigned long initial_pmd_value __initdata = 0;
109 
110 /*
111  * Initialise the cache_policy variable with the initial state specified
112  * via the "pmd" value.  This is used to ensure that on ARMv6 and later,
113  * the C code sets the page tables up with the same policy as the head
114  * assembly code, which avoids an illegal state where the TLBs can get
115  * confused.  See comments in early_cachepolicy() for more information.
116  */
117 void __init init_default_cache_policy(unsigned long pmd)
118 {
119 	int i;
120 
121 	initial_pmd_value = pmd;
122 
123 	pmd &= PMD_SECT_CACHE_MASK;
124 
125 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
126 		if (cache_policies[i].pmd == pmd) {
127 			cachepolicy = i;
128 			break;
129 		}
130 
131 	if (i == ARRAY_SIZE(cache_policies))
132 		pr_err("ERROR: could not find cache policy\n");
133 }
134 
135 /*
136  * These are useful for identifying cache coherency problems by allowing
137  * the cache or the cache and writebuffer to be turned off.  (Note: the
138  * write buffer should not be on and the cache off).
139  */
140 static int __init early_cachepolicy(char *p)
141 {
142 	int i, selected = -1;
143 
144 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
145 		int len = strlen(cache_policies[i].policy);
146 
147 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
148 			selected = i;
149 			break;
150 		}
151 	}
152 
153 	if (selected == -1)
154 		pr_err("ERROR: unknown or unsupported cache policy\n");
155 
156 	/*
157 	 * This restriction is partly to do with the way we boot; it is
158 	 * unpredictable to have memory mapped using two different sets of
159 	 * memory attributes (shared, type, and cache attribs).  We can not
160 	 * change these attributes once the initial assembly has setup the
161 	 * page tables.
162 	 */
163 	if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) {
164 		pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n",
165 			cache_policies[cachepolicy].policy);
166 		return 0;
167 	}
168 
169 	if (selected != cachepolicy) {
170 		unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
171 		cachepolicy = selected;
172 		flush_cache_all();
173 		set_cr(cr);
174 	}
175 	return 0;
176 }
177 early_param("cachepolicy", early_cachepolicy);
178 
179 static int __init early_nocache(char *__unused)
180 {
181 	char *p = "buffered";
182 	pr_warn("nocache is deprecated; use cachepolicy=%s\n", p);
183 	early_cachepolicy(p);
184 	return 0;
185 }
186 early_param("nocache", early_nocache);
187 
188 static int __init early_nowrite(char *__unused)
189 {
190 	char *p = "uncached";
191 	pr_warn("nowb is deprecated; use cachepolicy=%s\n", p);
192 	early_cachepolicy(p);
193 	return 0;
194 }
195 early_param("nowb", early_nowrite);
196 
197 #ifndef CONFIG_ARM_LPAE
198 static int __init early_ecc(char *p)
199 {
200 	if (memcmp(p, "on", 2) == 0)
201 		ecc_mask = PMD_PROTECTION;
202 	else if (memcmp(p, "off", 3) == 0)
203 		ecc_mask = 0;
204 	return 0;
205 }
206 early_param("ecc", early_ecc);
207 #endif
208 
209 #else /* ifdef CONFIG_CPU_CP15 */
210 
211 static int __init early_cachepolicy(char *p)
212 {
213 	pr_warn("cachepolicy kernel parameter not supported without cp15\n");
214 }
215 early_param("cachepolicy", early_cachepolicy);
216 
217 static int __init noalign_setup(char *__unused)
218 {
219 	pr_warn("noalign kernel parameter not supported without cp15\n");
220 }
221 __setup("noalign", noalign_setup);
222 
223 #endif /* ifdef CONFIG_CPU_CP15 / else */
224 
225 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
226 #define PROT_PTE_S2_DEVICE	PROT_PTE_DEVICE
227 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
228 
229 static struct mem_type mem_types[] __ro_after_init = {
230 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
231 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
232 				  L_PTE_SHARED,
233 		.prot_l1	= PMD_TYPE_TABLE,
234 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
235 		.domain		= DOMAIN_IO,
236 	},
237 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
238 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
239 		.prot_l1	= PMD_TYPE_TABLE,
240 		.prot_sect	= PROT_SECT_DEVICE,
241 		.domain		= DOMAIN_IO,
242 	},
243 	[MT_DEVICE_CACHED] = {	  /* ioremap_cache */
244 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
245 		.prot_l1	= PMD_TYPE_TABLE,
246 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
247 		.domain		= DOMAIN_IO,
248 	},
249 	[MT_DEVICE_WC] = {	/* ioremap_wc */
250 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
251 		.prot_l1	= PMD_TYPE_TABLE,
252 		.prot_sect	= PROT_SECT_DEVICE,
253 		.domain		= DOMAIN_IO,
254 	},
255 	[MT_UNCACHED] = {
256 		.prot_pte	= PROT_PTE_DEVICE,
257 		.prot_l1	= PMD_TYPE_TABLE,
258 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
259 		.domain		= DOMAIN_IO,
260 	},
261 	[MT_CACHECLEAN] = {
262 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
263 		.domain    = DOMAIN_KERNEL,
264 	},
265 #ifndef CONFIG_ARM_LPAE
266 	[MT_MINICLEAN] = {
267 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
268 		.domain    = DOMAIN_KERNEL,
269 	},
270 #endif
271 	[MT_LOW_VECTORS] = {
272 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
273 				L_PTE_RDONLY,
274 		.prot_l1   = PMD_TYPE_TABLE,
275 		.domain    = DOMAIN_VECTORS,
276 	},
277 	[MT_HIGH_VECTORS] = {
278 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
279 				L_PTE_USER | L_PTE_RDONLY,
280 		.prot_l1   = PMD_TYPE_TABLE,
281 		.domain    = DOMAIN_VECTORS,
282 	},
283 	[MT_MEMORY_RWX] = {
284 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
285 		.prot_l1   = PMD_TYPE_TABLE,
286 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
287 		.domain    = DOMAIN_KERNEL,
288 	},
289 	[MT_MEMORY_RW] = {
290 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
291 			     L_PTE_XN,
292 		.prot_l1   = PMD_TYPE_TABLE,
293 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
294 		.domain    = DOMAIN_KERNEL,
295 	},
296 	[MT_ROM] = {
297 		.prot_sect = PMD_TYPE_SECT,
298 		.domain    = DOMAIN_KERNEL,
299 	},
300 	[MT_MEMORY_RWX_NONCACHED] = {
301 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
302 				L_PTE_MT_BUFFERABLE,
303 		.prot_l1   = PMD_TYPE_TABLE,
304 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
305 		.domain    = DOMAIN_KERNEL,
306 	},
307 	[MT_MEMORY_RW_DTCM] = {
308 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
309 				L_PTE_XN,
310 		.prot_l1   = PMD_TYPE_TABLE,
311 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
312 		.domain    = DOMAIN_KERNEL,
313 	},
314 	[MT_MEMORY_RWX_ITCM] = {
315 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
316 		.prot_l1   = PMD_TYPE_TABLE,
317 		.domain    = DOMAIN_KERNEL,
318 	},
319 	[MT_MEMORY_RW_SO] = {
320 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
321 				L_PTE_MT_UNCACHED | L_PTE_XN,
322 		.prot_l1   = PMD_TYPE_TABLE,
323 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
324 				PMD_SECT_UNCACHED | PMD_SECT_XN,
325 		.domain    = DOMAIN_KERNEL,
326 	},
327 	[MT_MEMORY_DMA_READY] = {
328 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
329 				L_PTE_XN,
330 		.prot_l1   = PMD_TYPE_TABLE,
331 		.domain    = DOMAIN_KERNEL,
332 	},
333 };
334 
335 const struct mem_type *get_mem_type(unsigned int type)
336 {
337 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
338 }
339 EXPORT_SYMBOL(get_mem_type);
340 
341 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr);
342 
343 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS]
344 	__aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata;
345 
346 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr)
347 {
348 	return &bm_pte[pte_index(addr)];
349 }
350 
351 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr)
352 {
353 	return pte_offset_kernel(dir, addr);
354 }
355 
356 static inline pmd_t * __init fixmap_pmd(unsigned long addr)
357 {
358 	return pmd_off_k(addr);
359 }
360 
361 void __init early_fixmap_init(void)
362 {
363 	pmd_t *pmd;
364 
365 	/*
366 	 * The early fixmap range spans multiple pmds, for which
367 	 * we are not prepared:
368 	 */
369 	BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT)
370 		     != FIXADDR_TOP >> PMD_SHIFT);
371 
372 	pmd = fixmap_pmd(FIXADDR_TOP);
373 	pmd_populate_kernel(&init_mm, pmd, bm_pte);
374 
375 	pte_offset_fixmap = pte_offset_early_fixmap;
376 }
377 
378 /*
379  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
380  * As a result, this can only be called with preemption disabled, as under
381  * stop_machine().
382  */
383 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
384 {
385 	unsigned long vaddr = __fix_to_virt(idx);
386 	pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr);
387 
388 	/* Make sure fixmap region does not exceed available allocation. */
389 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
390 		     FIXADDR_END);
391 	BUG_ON(idx >= __end_of_fixed_addresses);
392 
393 	/* we only support device mappings until pgprot_kernel has been set */
394 	if (WARN_ON(pgprot_val(prot) != pgprot_val(FIXMAP_PAGE_IO) &&
395 		    pgprot_val(pgprot_kernel) == 0))
396 		return;
397 
398 	if (pgprot_val(prot))
399 		set_pte_at(NULL, vaddr, pte,
400 			pfn_pte(phys >> PAGE_SHIFT, prot));
401 	else
402 		pte_clear(NULL, vaddr, pte);
403 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
404 }
405 
406 /*
407  * Adjust the PMD section entries according to the CPU in use.
408  */
409 static void __init build_mem_type_table(void)
410 {
411 	struct cachepolicy *cp;
412 	unsigned int cr = get_cr();
413 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
414 	int cpu_arch = cpu_architecture();
415 	int i;
416 
417 	if (cpu_arch < CPU_ARCH_ARMv6) {
418 #if defined(CONFIG_CPU_DCACHE_DISABLE)
419 		if (cachepolicy > CPOLICY_BUFFERED)
420 			cachepolicy = CPOLICY_BUFFERED;
421 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
422 		if (cachepolicy > CPOLICY_WRITETHROUGH)
423 			cachepolicy = CPOLICY_WRITETHROUGH;
424 #endif
425 	}
426 	if (cpu_arch < CPU_ARCH_ARMv5) {
427 		if (cachepolicy >= CPOLICY_WRITEALLOC)
428 			cachepolicy = CPOLICY_WRITEBACK;
429 		ecc_mask = 0;
430 	}
431 
432 	if (is_smp()) {
433 		if (cachepolicy != CPOLICY_WRITEALLOC) {
434 			pr_warn("Forcing write-allocate cache policy for SMP\n");
435 			cachepolicy = CPOLICY_WRITEALLOC;
436 		}
437 		if (!(initial_pmd_value & PMD_SECT_S)) {
438 			pr_warn("Forcing shared mappings for SMP\n");
439 			initial_pmd_value |= PMD_SECT_S;
440 		}
441 	}
442 
443 	/*
444 	 * Strip out features not present on earlier architectures.
445 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
446 	 * without extended page tables don't have the 'Shared' bit.
447 	 */
448 	if (cpu_arch < CPU_ARCH_ARMv5)
449 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
450 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
451 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
452 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
453 			mem_types[i].prot_sect &= ~PMD_SECT_S;
454 
455 	/*
456 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
457 	 * "update-able on write" bit on ARM610).  However, Xscale and
458 	 * Xscale3 require this bit to be cleared.
459 	 */
460 	if (cpu_is_xscale_family()) {
461 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
462 			mem_types[i].prot_sect &= ~PMD_BIT4;
463 			mem_types[i].prot_l1 &= ~PMD_BIT4;
464 		}
465 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
466 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
467 			if (mem_types[i].prot_l1)
468 				mem_types[i].prot_l1 |= PMD_BIT4;
469 			if (mem_types[i].prot_sect)
470 				mem_types[i].prot_sect |= PMD_BIT4;
471 		}
472 	}
473 
474 	/*
475 	 * Mark the device areas according to the CPU/architecture.
476 	 */
477 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
478 		if (!cpu_is_xsc3()) {
479 			/*
480 			 * Mark device regions on ARMv6+ as execute-never
481 			 * to prevent speculative instruction fetches.
482 			 */
483 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
484 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
485 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
486 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
487 
488 			/* Also setup NX memory mapping */
489 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
490 		}
491 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
492 			/*
493 			 * For ARMv7 with TEX remapping,
494 			 * - shared device is SXCB=1100
495 			 * - nonshared device is SXCB=0100
496 			 * - write combine device mem is SXCB=0001
497 			 * (Uncached Normal memory)
498 			 */
499 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
500 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
501 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
502 		} else if (cpu_is_xsc3()) {
503 			/*
504 			 * For Xscale3,
505 			 * - shared device is TEXCB=00101
506 			 * - nonshared device is TEXCB=01000
507 			 * - write combine device mem is TEXCB=00100
508 			 * (Inner/Outer Uncacheable in xsc3 parlance)
509 			 */
510 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
511 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
512 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
513 		} else {
514 			/*
515 			 * For ARMv6 and ARMv7 without TEX remapping,
516 			 * - shared device is TEXCB=00001
517 			 * - nonshared device is TEXCB=01000
518 			 * - write combine device mem is TEXCB=00100
519 			 * (Uncached Normal in ARMv6 parlance).
520 			 */
521 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
522 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
523 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
524 		}
525 	} else {
526 		/*
527 		 * On others, write combining is "Uncached/Buffered"
528 		 */
529 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
530 	}
531 
532 	/*
533 	 * Now deal with the memory-type mappings
534 	 */
535 	cp = &cache_policies[cachepolicy];
536 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
537 
538 #ifndef CONFIG_ARM_LPAE
539 	/*
540 	 * We don't use domains on ARMv6 (since this causes problems with
541 	 * v6/v7 kernels), so we must use a separate memory type for user
542 	 * r/o, kernel r/w to map the vectors page.
543 	 */
544 	if (cpu_arch == CPU_ARCH_ARMv6)
545 		vecs_pgprot |= L_PTE_MT_VECTORS;
546 
547 	/*
548 	 * Check is it with support for the PXN bit
549 	 * in the Short-descriptor translation table format descriptors.
550 	 */
551 	if (cpu_arch == CPU_ARCH_ARMv7 &&
552 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
553 		user_pmd_table |= PMD_PXNTABLE;
554 	}
555 #endif
556 
557 	/*
558 	 * ARMv6 and above have extended page tables.
559 	 */
560 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
561 #ifndef CONFIG_ARM_LPAE
562 		/*
563 		 * Mark cache clean areas and XIP ROM read only
564 		 * from SVC mode and no access from userspace.
565 		 */
566 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
567 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
568 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
569 #endif
570 
571 		/*
572 		 * If the initial page tables were created with the S bit
573 		 * set, then we need to do the same here for the same
574 		 * reasons given in early_cachepolicy().
575 		 */
576 		if (initial_pmd_value & PMD_SECT_S) {
577 			user_pgprot |= L_PTE_SHARED;
578 			kern_pgprot |= L_PTE_SHARED;
579 			vecs_pgprot |= L_PTE_SHARED;
580 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
581 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
582 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
583 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
584 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
585 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
586 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
587 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
588 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
589 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
590 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
591 		}
592 	}
593 
594 	/*
595 	 * Non-cacheable Normal - intended for memory areas that must
596 	 * not cause dirty cache line writebacks when used
597 	 */
598 	if (cpu_arch >= CPU_ARCH_ARMv6) {
599 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
600 			/* Non-cacheable Normal is XCB = 001 */
601 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
602 				PMD_SECT_BUFFERED;
603 		} else {
604 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
605 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
606 				PMD_SECT_TEX(1);
607 		}
608 	} else {
609 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
610 	}
611 
612 #ifdef CONFIG_ARM_LPAE
613 	/*
614 	 * Do not generate access flag faults for the kernel mappings.
615 	 */
616 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
617 		mem_types[i].prot_pte |= PTE_EXT_AF;
618 		if (mem_types[i].prot_sect)
619 			mem_types[i].prot_sect |= PMD_SECT_AF;
620 	}
621 	kern_pgprot |= PTE_EXT_AF;
622 	vecs_pgprot |= PTE_EXT_AF;
623 
624 	/*
625 	 * Set PXN for user mappings
626 	 */
627 	user_pgprot |= PTE_EXT_PXN;
628 #endif
629 
630 	for (i = 0; i < 16; i++) {
631 		pteval_t v = pgprot_val(protection_map[i]);
632 		protection_map[i] = __pgprot(v | user_pgprot);
633 	}
634 
635 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
636 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
637 
638 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
639 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
640 				 L_PTE_DIRTY | kern_pgprot);
641 
642 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
643 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
644 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
645 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
646 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
647 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
648 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
649 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
650 	mem_types[MT_ROM].prot_sect |= cp->pmd;
651 
652 	switch (cp->pmd) {
653 	case PMD_SECT_WT:
654 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
655 		break;
656 	case PMD_SECT_WB:
657 	case PMD_SECT_WBWA:
658 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
659 		break;
660 	}
661 	pr_info("Memory policy: %sData cache %s\n",
662 		ecc_mask ? "ECC enabled, " : "", cp->policy);
663 
664 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
665 		struct mem_type *t = &mem_types[i];
666 		if (t->prot_l1)
667 			t->prot_l1 |= PMD_DOMAIN(t->domain);
668 		if (t->prot_sect)
669 			t->prot_sect |= PMD_DOMAIN(t->domain);
670 	}
671 }
672 
673 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
674 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
675 			      unsigned long size, pgprot_t vma_prot)
676 {
677 	if (!pfn_valid(pfn))
678 		return pgprot_noncached(vma_prot);
679 	else if (file->f_flags & O_SYNC)
680 		return pgprot_writecombine(vma_prot);
681 	return vma_prot;
682 }
683 EXPORT_SYMBOL(phys_mem_access_prot);
684 #endif
685 
686 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
687 
688 static void __init *early_alloc(unsigned long sz)
689 {
690 	void *ptr = memblock_alloc(sz, sz);
691 
692 	if (!ptr)
693 		panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
694 		      __func__, sz, sz);
695 
696 	return ptr;
697 }
698 
699 static void *__init late_alloc(unsigned long sz)
700 {
701 	void *ptr = (void *)__get_free_pages(GFP_PGTABLE_KERNEL, get_order(sz));
702 
703 	if (!ptr || !pgtable_pte_page_ctor(virt_to_page(ptr)))
704 		BUG();
705 	return ptr;
706 }
707 
708 static pte_t * __init arm_pte_alloc(pmd_t *pmd, unsigned long addr,
709 				unsigned long prot,
710 				void *(*alloc)(unsigned long sz))
711 {
712 	if (pmd_none(*pmd)) {
713 		pte_t *pte = alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
714 		__pmd_populate(pmd, __pa(pte), prot);
715 	}
716 	BUG_ON(pmd_bad(*pmd));
717 	return pte_offset_kernel(pmd, addr);
718 }
719 
720 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr,
721 				      unsigned long prot)
722 {
723 	return arm_pte_alloc(pmd, addr, prot, early_alloc);
724 }
725 
726 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
727 				  unsigned long end, unsigned long pfn,
728 				  const struct mem_type *type,
729 				  void *(*alloc)(unsigned long sz),
730 				  bool ng)
731 {
732 	pte_t *pte = arm_pte_alloc(pmd, addr, type->prot_l1, alloc);
733 	do {
734 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)),
735 			    ng ? PTE_EXT_NG : 0);
736 		pfn++;
737 	} while (pte++, addr += PAGE_SIZE, addr != end);
738 }
739 
740 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
741 			unsigned long end, phys_addr_t phys,
742 			const struct mem_type *type, bool ng)
743 {
744 	pmd_t *p = pmd;
745 
746 #ifndef CONFIG_ARM_LPAE
747 	/*
748 	 * In classic MMU format, puds and pmds are folded in to
749 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
750 	 * group of L1 entries making up one logical pointer to
751 	 * an L2 table (2MB), where as PMDs refer to the individual
752 	 * L1 entries (1MB). Hence increment to get the correct
753 	 * offset for odd 1MB sections.
754 	 * (See arch/arm/include/asm/pgtable-2level.h)
755 	 */
756 	if (addr & SECTION_SIZE)
757 		pmd++;
758 #endif
759 	do {
760 		*pmd = __pmd(phys | type->prot_sect | (ng ? PMD_SECT_nG : 0));
761 		phys += SECTION_SIZE;
762 	} while (pmd++, addr += SECTION_SIZE, addr != end);
763 
764 	flush_pmd_entry(p);
765 }
766 
767 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
768 				      unsigned long end, phys_addr_t phys,
769 				      const struct mem_type *type,
770 				      void *(*alloc)(unsigned long sz), bool ng)
771 {
772 	pmd_t *pmd = pmd_offset(pud, addr);
773 	unsigned long next;
774 
775 	do {
776 		/*
777 		 * With LPAE, we must loop over to map
778 		 * all the pmds for the given range.
779 		 */
780 		next = pmd_addr_end(addr, end);
781 
782 		/*
783 		 * Try a section mapping - addr, next and phys must all be
784 		 * aligned to a section boundary.
785 		 */
786 		if (type->prot_sect &&
787 				((addr | next | phys) & ~SECTION_MASK) == 0) {
788 			__map_init_section(pmd, addr, next, phys, type, ng);
789 		} else {
790 			alloc_init_pte(pmd, addr, next,
791 				       __phys_to_pfn(phys), type, alloc, ng);
792 		}
793 
794 		phys += next - addr;
795 
796 	} while (pmd++, addr = next, addr != end);
797 }
798 
799 static void __init alloc_init_pud(p4d_t *p4d, unsigned long addr,
800 				  unsigned long end, phys_addr_t phys,
801 				  const struct mem_type *type,
802 				  void *(*alloc)(unsigned long sz), bool ng)
803 {
804 	pud_t *pud = pud_offset(p4d, addr);
805 	unsigned long next;
806 
807 	do {
808 		next = pud_addr_end(addr, end);
809 		alloc_init_pmd(pud, addr, next, phys, type, alloc, ng);
810 		phys += next - addr;
811 	} while (pud++, addr = next, addr != end);
812 }
813 
814 static void __init alloc_init_p4d(pgd_t *pgd, unsigned long addr,
815 				  unsigned long end, phys_addr_t phys,
816 				  const struct mem_type *type,
817 				  void *(*alloc)(unsigned long sz), bool ng)
818 {
819 	p4d_t *p4d = p4d_offset(pgd, addr);
820 	unsigned long next;
821 
822 	do {
823 		next = p4d_addr_end(addr, end);
824 		alloc_init_pud(p4d, addr, next, phys, type, alloc, ng);
825 		phys += next - addr;
826 	} while (p4d++, addr = next, addr != end);
827 }
828 
829 #ifndef CONFIG_ARM_LPAE
830 static void __init create_36bit_mapping(struct mm_struct *mm,
831 					struct map_desc *md,
832 					const struct mem_type *type,
833 					bool ng)
834 {
835 	unsigned long addr, length, end;
836 	phys_addr_t phys;
837 	pgd_t *pgd;
838 
839 	addr = md->virtual;
840 	phys = __pfn_to_phys(md->pfn);
841 	length = PAGE_ALIGN(md->length);
842 
843 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
844 		pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
845 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
846 		return;
847 	}
848 
849 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
850 	 *	Since domain assignments can in fact be arbitrary, the
851 	 *	'domain == 0' check below is required to insure that ARMv6
852 	 *	supersections are only allocated for domain 0 regardless
853 	 *	of the actual domain assignments in use.
854 	 */
855 	if (type->domain) {
856 		pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
857 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
858 		return;
859 	}
860 
861 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
862 		pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
863 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
864 		return;
865 	}
866 
867 	/*
868 	 * Shift bits [35:32] of address into bits [23:20] of PMD
869 	 * (See ARMv6 spec).
870 	 */
871 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
872 
873 	pgd = pgd_offset(mm, addr);
874 	end = addr + length;
875 	do {
876 		p4d_t *p4d = p4d_offset(pgd, addr);
877 		pud_t *pud = pud_offset(p4d, addr);
878 		pmd_t *pmd = pmd_offset(pud, addr);
879 		int i;
880 
881 		for (i = 0; i < 16; i++)
882 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER |
883 				       (ng ? PMD_SECT_nG : 0));
884 
885 		addr += SUPERSECTION_SIZE;
886 		phys += SUPERSECTION_SIZE;
887 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
888 	} while (addr != end);
889 }
890 #endif	/* !CONFIG_ARM_LPAE */
891 
892 static void __init __create_mapping(struct mm_struct *mm, struct map_desc *md,
893 				    void *(*alloc)(unsigned long sz),
894 				    bool ng)
895 {
896 	unsigned long addr, length, end;
897 	phys_addr_t phys;
898 	const struct mem_type *type;
899 	pgd_t *pgd;
900 
901 	type = &mem_types[md->type];
902 
903 #ifndef CONFIG_ARM_LPAE
904 	/*
905 	 * Catch 36-bit addresses
906 	 */
907 	if (md->pfn >= 0x100000) {
908 		create_36bit_mapping(mm, md, type, ng);
909 		return;
910 	}
911 #endif
912 
913 	addr = md->virtual & PAGE_MASK;
914 	phys = __pfn_to_phys(md->pfn);
915 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
916 
917 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
918 		pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
919 			(long long)__pfn_to_phys(md->pfn), addr);
920 		return;
921 	}
922 
923 	pgd = pgd_offset(mm, addr);
924 	end = addr + length;
925 	do {
926 		unsigned long next = pgd_addr_end(addr, end);
927 
928 		alloc_init_p4d(pgd, addr, next, phys, type, alloc, ng);
929 
930 		phys += next - addr;
931 		addr = next;
932 	} while (pgd++, addr != end);
933 }
934 
935 /*
936  * Create the page directory entries and any necessary
937  * page tables for the mapping specified by `md'.  We
938  * are able to cope here with varying sizes and address
939  * offsets, and we take full advantage of sections and
940  * supersections.
941  */
942 static void __init create_mapping(struct map_desc *md)
943 {
944 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
945 		pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
946 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
947 		return;
948 	}
949 
950 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
951 	    md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START &&
952 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
953 		pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
954 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
955 	}
956 
957 	__create_mapping(&init_mm, md, early_alloc, false);
958 }
959 
960 void __init create_mapping_late(struct mm_struct *mm, struct map_desc *md,
961 				bool ng)
962 {
963 #ifdef CONFIG_ARM_LPAE
964 	p4d_t *p4d;
965 	pud_t *pud;
966 
967 	p4d = p4d_alloc(mm, pgd_offset(mm, md->virtual), md->virtual);
968 	if (WARN_ON(!p4d))
969 		return;
970 	pud = pud_alloc(mm, p4d, md->virtual);
971 	if (WARN_ON(!pud))
972 		return;
973 	pmd_alloc(mm, pud, 0);
974 #endif
975 	__create_mapping(mm, md, late_alloc, ng);
976 }
977 
978 /*
979  * Create the architecture specific mappings
980  */
981 void __init iotable_init(struct map_desc *io_desc, int nr)
982 {
983 	struct map_desc *md;
984 	struct vm_struct *vm;
985 	struct static_vm *svm;
986 
987 	if (!nr)
988 		return;
989 
990 	svm = memblock_alloc(sizeof(*svm) * nr, __alignof__(*svm));
991 	if (!svm)
992 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
993 		      __func__, sizeof(*svm) * nr, __alignof__(*svm));
994 
995 	for (md = io_desc; nr; md++, nr--) {
996 		create_mapping(md);
997 
998 		vm = &svm->vm;
999 		vm->addr = (void *)(md->virtual & PAGE_MASK);
1000 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
1001 		vm->phys_addr = __pfn_to_phys(md->pfn);
1002 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
1003 		vm->flags |= VM_ARM_MTYPE(md->type);
1004 		vm->caller = iotable_init;
1005 		add_static_vm_early(svm++);
1006 	}
1007 }
1008 
1009 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
1010 				  void *caller)
1011 {
1012 	struct vm_struct *vm;
1013 	struct static_vm *svm;
1014 
1015 	svm = memblock_alloc(sizeof(*svm), __alignof__(*svm));
1016 	if (!svm)
1017 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1018 		      __func__, sizeof(*svm), __alignof__(*svm));
1019 
1020 	vm = &svm->vm;
1021 	vm->addr = (void *)addr;
1022 	vm->size = size;
1023 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
1024 	vm->caller = caller;
1025 	add_static_vm_early(svm);
1026 }
1027 
1028 #ifndef CONFIG_ARM_LPAE
1029 
1030 /*
1031  * The Linux PMD is made of two consecutive section entries covering 2MB
1032  * (see definition in include/asm/pgtable-2level.h).  However a call to
1033  * create_mapping() may optimize static mappings by using individual
1034  * 1MB section mappings.  This leaves the actual PMD potentially half
1035  * initialized if the top or bottom section entry isn't used, leaving it
1036  * open to problems if a subsequent ioremap() or vmalloc() tries to use
1037  * the virtual space left free by that unused section entry.
1038  *
1039  * Let's avoid the issue by inserting dummy vm entries covering the unused
1040  * PMD halves once the static mappings are in place.
1041  */
1042 
1043 static void __init pmd_empty_section_gap(unsigned long addr)
1044 {
1045 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1046 }
1047 
1048 static void __init fill_pmd_gaps(void)
1049 {
1050 	struct static_vm *svm;
1051 	struct vm_struct *vm;
1052 	unsigned long addr, next = 0;
1053 	pmd_t *pmd;
1054 
1055 	list_for_each_entry(svm, &static_vmlist, list) {
1056 		vm = &svm->vm;
1057 		addr = (unsigned long)vm->addr;
1058 		if (addr < next)
1059 			continue;
1060 
1061 		/*
1062 		 * Check if this vm starts on an odd section boundary.
1063 		 * If so and the first section entry for this PMD is free
1064 		 * then we block the corresponding virtual address.
1065 		 */
1066 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1067 			pmd = pmd_off_k(addr);
1068 			if (pmd_none(*pmd))
1069 				pmd_empty_section_gap(addr & PMD_MASK);
1070 		}
1071 
1072 		/*
1073 		 * Then check if this vm ends on an odd section boundary.
1074 		 * If so and the second section entry for this PMD is empty
1075 		 * then we block the corresponding virtual address.
1076 		 */
1077 		addr += vm->size;
1078 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1079 			pmd = pmd_off_k(addr) + 1;
1080 			if (pmd_none(*pmd))
1081 				pmd_empty_section_gap(addr);
1082 		}
1083 
1084 		/* no need to look at any vm entry until we hit the next PMD */
1085 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1086 	}
1087 }
1088 
1089 #else
1090 #define fill_pmd_gaps() do { } while (0)
1091 #endif
1092 
1093 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1094 static void __init pci_reserve_io(void)
1095 {
1096 	struct static_vm *svm;
1097 
1098 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1099 	if (svm)
1100 		return;
1101 
1102 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1103 }
1104 #else
1105 #define pci_reserve_io() do { } while (0)
1106 #endif
1107 
1108 #ifdef CONFIG_DEBUG_LL
1109 void __init debug_ll_io_init(void)
1110 {
1111 	struct map_desc map;
1112 
1113 	debug_ll_addr(&map.pfn, &map.virtual);
1114 	if (!map.pfn || !map.virtual)
1115 		return;
1116 	map.pfn = __phys_to_pfn(map.pfn);
1117 	map.virtual &= PAGE_MASK;
1118 	map.length = PAGE_SIZE;
1119 	map.type = MT_DEVICE;
1120 	iotable_init(&map, 1);
1121 }
1122 #endif
1123 
1124 static void * __initdata vmalloc_min =
1125 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1126 
1127 /*
1128  * vmalloc=size forces the vmalloc area to be exactly 'size'
1129  * bytes. This can be used to increase (or decrease) the vmalloc
1130  * area - the default is 240m.
1131  */
1132 static int __init early_vmalloc(char *arg)
1133 {
1134 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1135 
1136 	if (vmalloc_reserve < SZ_16M) {
1137 		vmalloc_reserve = SZ_16M;
1138 		pr_warn("vmalloc area too small, limiting to %luMB\n",
1139 			vmalloc_reserve >> 20);
1140 	}
1141 
1142 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1143 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1144 		pr_warn("vmalloc area is too big, limiting to %luMB\n",
1145 			vmalloc_reserve >> 20);
1146 	}
1147 
1148 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1149 	return 0;
1150 }
1151 early_param("vmalloc", early_vmalloc);
1152 
1153 phys_addr_t arm_lowmem_limit __initdata = 0;
1154 
1155 void __init adjust_lowmem_bounds(void)
1156 {
1157 	phys_addr_t memblock_limit = 0;
1158 	u64 vmalloc_limit;
1159 	struct memblock_region *reg;
1160 	phys_addr_t lowmem_limit = 0;
1161 
1162 	/*
1163 	 * Let's use our own (unoptimized) equivalent of __pa() that is
1164 	 * not affected by wrap-arounds when sizeof(phys_addr_t) == 4.
1165 	 * The result is used as the upper bound on physical memory address
1166 	 * and may itself be outside the valid range for which phys_addr_t
1167 	 * and therefore __pa() is defined.
1168 	 */
1169 	vmalloc_limit = (u64)(uintptr_t)vmalloc_min - PAGE_OFFSET + PHYS_OFFSET;
1170 
1171 	/*
1172 	 * The first usable region must be PMD aligned. Mark its start
1173 	 * as MEMBLOCK_NOMAP if it isn't
1174 	 */
1175 	for_each_memblock(memory, reg) {
1176 		if (!memblock_is_nomap(reg)) {
1177 			if (!IS_ALIGNED(reg->base, PMD_SIZE)) {
1178 				phys_addr_t len;
1179 
1180 				len = round_up(reg->base, PMD_SIZE) - reg->base;
1181 				memblock_mark_nomap(reg->base, len);
1182 			}
1183 			break;
1184 		}
1185 	}
1186 
1187 	for_each_memblock(memory, reg) {
1188 		phys_addr_t block_start = reg->base;
1189 		phys_addr_t block_end = reg->base + reg->size;
1190 
1191 		if (memblock_is_nomap(reg))
1192 			continue;
1193 
1194 		if (reg->base < vmalloc_limit) {
1195 			if (block_end > lowmem_limit)
1196 				/*
1197 				 * Compare as u64 to ensure vmalloc_limit does
1198 				 * not get truncated. block_end should always
1199 				 * fit in phys_addr_t so there should be no
1200 				 * issue with assignment.
1201 				 */
1202 				lowmem_limit = min_t(u64,
1203 							 vmalloc_limit,
1204 							 block_end);
1205 
1206 			/*
1207 			 * Find the first non-pmd-aligned page, and point
1208 			 * memblock_limit at it. This relies on rounding the
1209 			 * limit down to be pmd-aligned, which happens at the
1210 			 * end of this function.
1211 			 *
1212 			 * With this algorithm, the start or end of almost any
1213 			 * bank can be non-pmd-aligned. The only exception is
1214 			 * that the start of the bank 0 must be section-
1215 			 * aligned, since otherwise memory would need to be
1216 			 * allocated when mapping the start of bank 0, which
1217 			 * occurs before any free memory is mapped.
1218 			 */
1219 			if (!memblock_limit) {
1220 				if (!IS_ALIGNED(block_start, PMD_SIZE))
1221 					memblock_limit = block_start;
1222 				else if (!IS_ALIGNED(block_end, PMD_SIZE))
1223 					memblock_limit = lowmem_limit;
1224 			}
1225 
1226 		}
1227 	}
1228 
1229 	arm_lowmem_limit = lowmem_limit;
1230 
1231 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1232 
1233 	if (!memblock_limit)
1234 		memblock_limit = arm_lowmem_limit;
1235 
1236 	/*
1237 	 * Round the memblock limit down to a pmd size.  This
1238 	 * helps to ensure that we will allocate memory from the
1239 	 * last full pmd, which should be mapped.
1240 	 */
1241 	memblock_limit = round_down(memblock_limit, PMD_SIZE);
1242 
1243 	if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1244 		if (memblock_end_of_DRAM() > arm_lowmem_limit) {
1245 			phys_addr_t end = memblock_end_of_DRAM();
1246 
1247 			pr_notice("Ignoring RAM at %pa-%pa\n",
1248 				  &memblock_limit, &end);
1249 			pr_notice("Consider using a HIGHMEM enabled kernel.\n");
1250 
1251 			memblock_remove(memblock_limit, end - memblock_limit);
1252 		}
1253 	}
1254 
1255 	memblock_set_current_limit(memblock_limit);
1256 }
1257 
1258 static inline void prepare_page_table(void)
1259 {
1260 	unsigned long addr;
1261 	phys_addr_t end;
1262 
1263 	/*
1264 	 * Clear out all the mappings below the kernel image.
1265 	 */
1266 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1267 		pmd_clear(pmd_off_k(addr));
1268 
1269 #ifdef CONFIG_XIP_KERNEL
1270 	/* The XIP kernel is mapped in the module area -- skip over it */
1271 	addr = ((unsigned long)_exiprom + PMD_SIZE - 1) & PMD_MASK;
1272 #endif
1273 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1274 		pmd_clear(pmd_off_k(addr));
1275 
1276 	/*
1277 	 * Find the end of the first block of lowmem.
1278 	 */
1279 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1280 	if (end >= arm_lowmem_limit)
1281 		end = arm_lowmem_limit;
1282 
1283 	/*
1284 	 * Clear out all the kernel space mappings, except for the first
1285 	 * memory bank, up to the vmalloc region.
1286 	 */
1287 	for (addr = __phys_to_virt(end);
1288 	     addr < VMALLOC_START; addr += PMD_SIZE)
1289 		pmd_clear(pmd_off_k(addr));
1290 }
1291 
1292 #ifdef CONFIG_ARM_LPAE
1293 /* the first page is reserved for pgd */
1294 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1295 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1296 #else
1297 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1298 #endif
1299 
1300 /*
1301  * Reserve the special regions of memory
1302  */
1303 void __init arm_mm_memblock_reserve(void)
1304 {
1305 	/*
1306 	 * Reserve the page tables.  These are already in use,
1307 	 * and can only be in node 0.
1308 	 */
1309 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1310 
1311 #ifdef CONFIG_SA1111
1312 	/*
1313 	 * Because of the SA1111 DMA bug, we want to preserve our
1314 	 * precious DMA-able memory...
1315 	 */
1316 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1317 #endif
1318 }
1319 
1320 /*
1321  * Set up the device mappings.  Since we clear out the page tables for all
1322  * mappings above VMALLOC_START, except early fixmap, we might remove debug
1323  * device mappings.  This means earlycon can be used to debug this function
1324  * Any other function or debugging method which may touch any device _will_
1325  * crash the kernel.
1326  */
1327 static void __init devicemaps_init(const struct machine_desc *mdesc)
1328 {
1329 	struct map_desc map;
1330 	unsigned long addr;
1331 	void *vectors;
1332 
1333 	/*
1334 	 * Allocate the vector page early.
1335 	 */
1336 	vectors = early_alloc(PAGE_SIZE * 2);
1337 
1338 	early_trap_init(vectors);
1339 
1340 	/*
1341 	 * Clear page table except top pmd used by early fixmaps
1342 	 */
1343 	for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE)
1344 		pmd_clear(pmd_off_k(addr));
1345 
1346 	/*
1347 	 * Map the kernel if it is XIP.
1348 	 * It is always first in the modulearea.
1349 	 */
1350 #ifdef CONFIG_XIP_KERNEL
1351 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1352 	map.virtual = MODULES_VADDR;
1353 	map.length = ((unsigned long)_exiprom - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1354 	map.type = MT_ROM;
1355 	create_mapping(&map);
1356 #endif
1357 
1358 	/*
1359 	 * Map the cache flushing regions.
1360 	 */
1361 #ifdef FLUSH_BASE
1362 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1363 	map.virtual = FLUSH_BASE;
1364 	map.length = SZ_1M;
1365 	map.type = MT_CACHECLEAN;
1366 	create_mapping(&map);
1367 #endif
1368 #ifdef FLUSH_BASE_MINICACHE
1369 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1370 	map.virtual = FLUSH_BASE_MINICACHE;
1371 	map.length = SZ_1M;
1372 	map.type = MT_MINICLEAN;
1373 	create_mapping(&map);
1374 #endif
1375 
1376 	/*
1377 	 * Create a mapping for the machine vectors at the high-vectors
1378 	 * location (0xffff0000).  If we aren't using high-vectors, also
1379 	 * create a mapping at the low-vectors virtual address.
1380 	 */
1381 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1382 	map.virtual = 0xffff0000;
1383 	map.length = PAGE_SIZE;
1384 #ifdef CONFIG_KUSER_HELPERS
1385 	map.type = MT_HIGH_VECTORS;
1386 #else
1387 	map.type = MT_LOW_VECTORS;
1388 #endif
1389 	create_mapping(&map);
1390 
1391 	if (!vectors_high()) {
1392 		map.virtual = 0;
1393 		map.length = PAGE_SIZE * 2;
1394 		map.type = MT_LOW_VECTORS;
1395 		create_mapping(&map);
1396 	}
1397 
1398 	/* Now create a kernel read-only mapping */
1399 	map.pfn += 1;
1400 	map.virtual = 0xffff0000 + PAGE_SIZE;
1401 	map.length = PAGE_SIZE;
1402 	map.type = MT_LOW_VECTORS;
1403 	create_mapping(&map);
1404 
1405 	/*
1406 	 * Ask the machine support to map in the statically mapped devices.
1407 	 */
1408 	if (mdesc->map_io)
1409 		mdesc->map_io();
1410 	else
1411 		debug_ll_io_init();
1412 	fill_pmd_gaps();
1413 
1414 	/* Reserve fixed i/o space in VMALLOC region */
1415 	pci_reserve_io();
1416 
1417 	/*
1418 	 * Finally flush the caches and tlb to ensure that we're in a
1419 	 * consistent state wrt the writebuffer.  This also ensures that
1420 	 * any write-allocated cache lines in the vector page are written
1421 	 * back.  After this point, we can start to touch devices again.
1422 	 */
1423 	local_flush_tlb_all();
1424 	flush_cache_all();
1425 
1426 	/* Enable asynchronous aborts */
1427 	early_abt_enable();
1428 }
1429 
1430 static void __init kmap_init(void)
1431 {
1432 #ifdef CONFIG_HIGHMEM
1433 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1434 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1435 #endif
1436 
1437 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1438 			_PAGE_KERNEL_TABLE);
1439 }
1440 
1441 static void __init map_lowmem(void)
1442 {
1443 	struct memblock_region *reg;
1444 	phys_addr_t kernel_x_start = round_down(__pa(KERNEL_START), SECTION_SIZE);
1445 	phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1446 
1447 	/* Map all the lowmem memory banks. */
1448 	for_each_memblock(memory, reg) {
1449 		phys_addr_t start = reg->base;
1450 		phys_addr_t end = start + reg->size;
1451 		struct map_desc map;
1452 
1453 		if (memblock_is_nomap(reg))
1454 			continue;
1455 
1456 		if (end > arm_lowmem_limit)
1457 			end = arm_lowmem_limit;
1458 		if (start >= end)
1459 			break;
1460 
1461 		if (end < kernel_x_start) {
1462 			map.pfn = __phys_to_pfn(start);
1463 			map.virtual = __phys_to_virt(start);
1464 			map.length = end - start;
1465 			map.type = MT_MEMORY_RWX;
1466 
1467 			create_mapping(&map);
1468 		} else if (start >= kernel_x_end) {
1469 			map.pfn = __phys_to_pfn(start);
1470 			map.virtual = __phys_to_virt(start);
1471 			map.length = end - start;
1472 			map.type = MT_MEMORY_RW;
1473 
1474 			create_mapping(&map);
1475 		} else {
1476 			/* This better cover the entire kernel */
1477 			if (start < kernel_x_start) {
1478 				map.pfn = __phys_to_pfn(start);
1479 				map.virtual = __phys_to_virt(start);
1480 				map.length = kernel_x_start - start;
1481 				map.type = MT_MEMORY_RW;
1482 
1483 				create_mapping(&map);
1484 			}
1485 
1486 			map.pfn = __phys_to_pfn(kernel_x_start);
1487 			map.virtual = __phys_to_virt(kernel_x_start);
1488 			map.length = kernel_x_end - kernel_x_start;
1489 			map.type = MT_MEMORY_RWX;
1490 
1491 			create_mapping(&map);
1492 
1493 			if (kernel_x_end < end) {
1494 				map.pfn = __phys_to_pfn(kernel_x_end);
1495 				map.virtual = __phys_to_virt(kernel_x_end);
1496 				map.length = end - kernel_x_end;
1497 				map.type = MT_MEMORY_RW;
1498 
1499 				create_mapping(&map);
1500 			}
1501 		}
1502 	}
1503 }
1504 
1505 #ifdef CONFIG_ARM_PV_FIXUP
1506 extern unsigned long __atags_pointer;
1507 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata);
1508 pgtables_remap lpae_pgtables_remap_asm;
1509 
1510 /*
1511  * early_paging_init() recreates boot time page table setup, allowing machines
1512  * to switch over to a high (>4G) address space on LPAE systems
1513  */
1514 static void __init early_paging_init(const struct machine_desc *mdesc)
1515 {
1516 	pgtables_remap *lpae_pgtables_remap;
1517 	unsigned long pa_pgd;
1518 	unsigned int cr, ttbcr;
1519 	long long offset;
1520 	void *boot_data;
1521 
1522 	if (!mdesc->pv_fixup)
1523 		return;
1524 
1525 	offset = mdesc->pv_fixup();
1526 	if (offset == 0)
1527 		return;
1528 
1529 	/*
1530 	 * Get the address of the remap function in the 1:1 identity
1531 	 * mapping setup by the early page table assembly code.  We
1532 	 * must get this prior to the pv update.  The following barrier
1533 	 * ensures that this is complete before we fixup any P:V offsets.
1534 	 */
1535 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1536 	pa_pgd = __pa(swapper_pg_dir);
1537 	boot_data = __va(__atags_pointer);
1538 	barrier();
1539 
1540 	pr_info("Switching physical address space to 0x%08llx\n",
1541 		(u64)PHYS_OFFSET + offset);
1542 
1543 	/* Re-set the phys pfn offset, and the pv offset */
1544 	__pv_offset += offset;
1545 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1546 
1547 	/* Run the patch stub to update the constants */
1548 	fixup_pv_table(&__pv_table_begin,
1549 		(&__pv_table_end - &__pv_table_begin) << 2);
1550 
1551 	/*
1552 	 * We changing not only the virtual to physical mapping, but also
1553 	 * the physical addresses used to access memory.  We need to flush
1554 	 * all levels of cache in the system with caching disabled to
1555 	 * ensure that all data is written back, and nothing is prefetched
1556 	 * into the caches.  We also need to prevent the TLB walkers
1557 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1558 	 * specific.
1559 	 */
1560 	cr = get_cr();
1561 	set_cr(cr & ~(CR_I | CR_C));
1562 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1563 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1564 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1565 	flush_cache_all();
1566 
1567 	/*
1568 	 * Fixup the page tables - this must be in the idmap region as
1569 	 * we need to disable the MMU to do this safely, and hence it
1570 	 * needs to be assembly.  It's fairly simple, as we're using the
1571 	 * temporary tables setup by the initial assembly code.
1572 	 */
1573 	lpae_pgtables_remap(offset, pa_pgd, boot_data);
1574 
1575 	/* Re-enable the caches and cacheable TLB walks */
1576 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1577 	set_cr(cr);
1578 }
1579 
1580 #else
1581 
1582 static void __init early_paging_init(const struct machine_desc *mdesc)
1583 {
1584 	long long offset;
1585 
1586 	if (!mdesc->pv_fixup)
1587 		return;
1588 
1589 	offset = mdesc->pv_fixup();
1590 	if (offset == 0)
1591 		return;
1592 
1593 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1594 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1595 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1596 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1597 }
1598 
1599 #endif
1600 
1601 static void __init early_fixmap_shutdown(void)
1602 {
1603 	int i;
1604 	unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1);
1605 
1606 	pte_offset_fixmap = pte_offset_late_fixmap;
1607 	pmd_clear(fixmap_pmd(va));
1608 	local_flush_tlb_kernel_page(va);
1609 
1610 	for (i = 0; i < __end_of_permanent_fixed_addresses; i++) {
1611 		pte_t *pte;
1612 		struct map_desc map;
1613 
1614 		map.virtual = fix_to_virt(i);
1615 		pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual);
1616 
1617 		/* Only i/o device mappings are supported ATM */
1618 		if (pte_none(*pte) ||
1619 		    (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED)
1620 			continue;
1621 
1622 		map.pfn = pte_pfn(*pte);
1623 		map.type = MT_DEVICE;
1624 		map.length = PAGE_SIZE;
1625 
1626 		create_mapping(&map);
1627 	}
1628 }
1629 
1630 /*
1631  * paging_init() sets up the page tables, initialises the zone memory
1632  * maps, and sets up the zero page, bad page and bad page tables.
1633  */
1634 void __init paging_init(const struct machine_desc *mdesc)
1635 {
1636 	void *zero_page;
1637 
1638 	prepare_page_table();
1639 	map_lowmem();
1640 	memblock_set_current_limit(arm_lowmem_limit);
1641 	dma_contiguous_remap();
1642 	early_fixmap_shutdown();
1643 	devicemaps_init(mdesc);
1644 	kmap_init();
1645 	tcm_init();
1646 
1647 	top_pmd = pmd_off_k(0xffff0000);
1648 
1649 	/* allocate the zero page. */
1650 	zero_page = early_alloc(PAGE_SIZE);
1651 
1652 	bootmem_init();
1653 
1654 	empty_zero_page = virt_to_page(zero_page);
1655 	__flush_dcache_page(NULL, empty_zero_page);
1656 }
1657 
1658 void __init early_mm_init(const struct machine_desc *mdesc)
1659 {
1660 	build_mem_type_table();
1661 	early_paging_init(mdesc);
1662 }
1663 
1664 void set_pte_at(struct mm_struct *mm, unsigned long addr,
1665 			      pte_t *ptep, pte_t pteval)
1666 {
1667 	unsigned long ext = 0;
1668 
1669 	if (addr < TASK_SIZE && pte_valid_user(pteval)) {
1670 		if (!pte_special(pteval))
1671 			__sync_icache_dcache(pteval);
1672 		ext |= PTE_EXT_NG;
1673 	}
1674 
1675 	set_pte_ext(ptep, pteval, ext);
1676 }
1677