xref: /openbmc/linux/arch/arm/mm/mmu.c (revision 36acd5e2)
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/cachetype.h>
21 #include <asm/fixmap.h>
22 #include <asm/sections.h>
23 #include <asm/setup.h>
24 #include <asm/smp_plat.h>
25 #include <asm/tlb.h>
26 #include <asm/highmem.h>
27 #include <asm/system_info.h>
28 #include <asm/traps.h>
29 #include <asm/procinfo.h>
30 #include <asm/memory.h>
31 #include <asm/pgalloc.h>
32 #include <asm/kasan_def.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 extern unsigned long __atags_pointer;
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 
70 EXPORT_SYMBOL(pgprot_user);
71 EXPORT_SYMBOL(pgprot_kernel);
72 
73 struct cachepolicy {
74 	const char	policy[16];
75 	unsigned int	cr_mask;
76 	pmdval_t	pmd;
77 	pteval_t	pte;
78 };
79 
80 static struct cachepolicy cache_policies[] __initdata = {
81 	{
82 		.policy		= "uncached",
83 		.cr_mask	= CR_W|CR_C,
84 		.pmd		= PMD_SECT_UNCACHED,
85 		.pte		= L_PTE_MT_UNCACHED,
86 	}, {
87 		.policy		= "buffered",
88 		.cr_mask	= CR_C,
89 		.pmd		= PMD_SECT_BUFFERED,
90 		.pte		= L_PTE_MT_BUFFERABLE,
91 	}, {
92 		.policy		= "writethrough",
93 		.cr_mask	= 0,
94 		.pmd		= PMD_SECT_WT,
95 		.pte		= L_PTE_MT_WRITETHROUGH,
96 	}, {
97 		.policy		= "writeback",
98 		.cr_mask	= 0,
99 		.pmd		= PMD_SECT_WB,
100 		.pte		= L_PTE_MT_WRITEBACK,
101 	}, {
102 		.policy		= "writealloc",
103 		.cr_mask	= 0,
104 		.pmd		= PMD_SECT_WBWA,
105 		.pte		= L_PTE_MT_WRITEALLOC,
106 	}
107 };
108 
109 #ifdef CONFIG_CPU_CP15
110 static unsigned long initial_pmd_value __initdata = 0;
111 
112 /*
113  * Initialise the cache_policy variable with the initial state specified
114  * via the "pmd" value.  This is used to ensure that on ARMv6 and later,
115  * the C code sets the page tables up with the same policy as the head
116  * assembly code, which avoids an illegal state where the TLBs can get
117  * confused.  See comments in early_cachepolicy() for more information.
118  */
119 void __init init_default_cache_policy(unsigned long pmd)
120 {
121 	int i;
122 
123 	initial_pmd_value = pmd;
124 
125 	pmd &= PMD_SECT_CACHE_MASK;
126 
127 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
128 		if (cache_policies[i].pmd == pmd) {
129 			cachepolicy = i;
130 			break;
131 		}
132 
133 	if (i == ARRAY_SIZE(cache_policies))
134 		pr_err("ERROR: could not find cache policy\n");
135 }
136 
137 /*
138  * These are useful for identifying cache coherency problems by allowing
139  * the cache or the cache and writebuffer to be turned off.  (Note: the
140  * write buffer should not be on and the cache off).
141  */
142 static int __init early_cachepolicy(char *p)
143 {
144 	int i, selected = -1;
145 
146 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
147 		int len = strlen(cache_policies[i].policy);
148 
149 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
150 			selected = i;
151 			break;
152 		}
153 	}
154 
155 	if (selected == -1)
156 		pr_err("ERROR: unknown or unsupported cache policy\n");
157 
158 	/*
159 	 * This restriction is partly to do with the way we boot; it is
160 	 * unpredictable to have memory mapped using two different sets of
161 	 * memory attributes (shared, type, and cache attribs).  We can not
162 	 * change these attributes once the initial assembly has setup the
163 	 * page tables.
164 	 */
165 	if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) {
166 		pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n",
167 			cache_policies[cachepolicy].policy);
168 		return 0;
169 	}
170 
171 	if (selected != cachepolicy) {
172 		unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
173 		cachepolicy = selected;
174 		flush_cache_all();
175 		set_cr(cr);
176 	}
177 	return 0;
178 }
179 early_param("cachepolicy", early_cachepolicy);
180 
181 static int __init early_nocache(char *__unused)
182 {
183 	char *p = "buffered";
184 	pr_warn("nocache is deprecated; use cachepolicy=%s\n", p);
185 	early_cachepolicy(p);
186 	return 0;
187 }
188 early_param("nocache", early_nocache);
189 
190 static int __init early_nowrite(char *__unused)
191 {
192 	char *p = "uncached";
193 	pr_warn("nowb is deprecated; use cachepolicy=%s\n", p);
194 	early_cachepolicy(p);
195 	return 0;
196 }
197 early_param("nowb", early_nowrite);
198 
199 #ifndef CONFIG_ARM_LPAE
200 static int __init early_ecc(char *p)
201 {
202 	if (memcmp(p, "on", 2) == 0)
203 		ecc_mask = PMD_PROTECTION;
204 	else if (memcmp(p, "off", 3) == 0)
205 		ecc_mask = 0;
206 	return 0;
207 }
208 early_param("ecc", early_ecc);
209 #endif
210 
211 #else /* ifdef CONFIG_CPU_CP15 */
212 
213 static int __init early_cachepolicy(char *p)
214 {
215 	pr_warn("cachepolicy kernel parameter not supported without cp15\n");
216 }
217 early_param("cachepolicy", early_cachepolicy);
218 
219 static int __init noalign_setup(char *__unused)
220 {
221 	pr_warn("noalign kernel parameter not supported without cp15\n");
222 }
223 __setup("noalign", noalign_setup);
224 
225 #endif /* ifdef CONFIG_CPU_CP15 / else */
226 
227 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
228 #define PROT_PTE_S2_DEVICE	PROT_PTE_DEVICE
229 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
230 
231 static struct mem_type mem_types[] __ro_after_init = {
232 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
233 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
234 				  L_PTE_SHARED,
235 		.prot_l1	= PMD_TYPE_TABLE,
236 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
237 		.domain		= DOMAIN_IO,
238 	},
239 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
240 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
241 		.prot_l1	= PMD_TYPE_TABLE,
242 		.prot_sect	= PROT_SECT_DEVICE,
243 		.domain		= DOMAIN_IO,
244 	},
245 	[MT_DEVICE_CACHED] = {	  /* ioremap_cache */
246 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
247 		.prot_l1	= PMD_TYPE_TABLE,
248 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
249 		.domain		= DOMAIN_IO,
250 	},
251 	[MT_DEVICE_WC] = {	/* ioremap_wc */
252 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
253 		.prot_l1	= PMD_TYPE_TABLE,
254 		.prot_sect	= PROT_SECT_DEVICE,
255 		.domain		= DOMAIN_IO,
256 	},
257 	[MT_UNCACHED] = {
258 		.prot_pte	= PROT_PTE_DEVICE,
259 		.prot_l1	= PMD_TYPE_TABLE,
260 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
261 		.domain		= DOMAIN_IO,
262 	},
263 	[MT_CACHECLEAN] = {
264 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
265 		.domain    = DOMAIN_KERNEL,
266 	},
267 #ifndef CONFIG_ARM_LPAE
268 	[MT_MINICLEAN] = {
269 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
270 		.domain    = DOMAIN_KERNEL,
271 	},
272 #endif
273 	[MT_LOW_VECTORS] = {
274 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
275 				L_PTE_RDONLY,
276 		.prot_l1   = PMD_TYPE_TABLE,
277 		.domain    = DOMAIN_VECTORS,
278 	},
279 	[MT_HIGH_VECTORS] = {
280 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
281 				L_PTE_USER | L_PTE_RDONLY,
282 		.prot_l1   = PMD_TYPE_TABLE,
283 		.domain    = DOMAIN_VECTORS,
284 	},
285 	[MT_MEMORY_RWX] = {
286 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
287 		.prot_l1   = PMD_TYPE_TABLE,
288 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
289 		.domain    = DOMAIN_KERNEL,
290 	},
291 	[MT_MEMORY_RW] = {
292 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
293 			     L_PTE_XN,
294 		.prot_l1   = PMD_TYPE_TABLE,
295 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
296 		.domain    = DOMAIN_KERNEL,
297 	},
298 	[MT_ROM] = {
299 		.prot_sect = PMD_TYPE_SECT,
300 		.domain    = DOMAIN_KERNEL,
301 	},
302 	[MT_MEMORY_RWX_NONCACHED] = {
303 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
304 				L_PTE_MT_BUFFERABLE,
305 		.prot_l1   = PMD_TYPE_TABLE,
306 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
307 		.domain    = DOMAIN_KERNEL,
308 	},
309 	[MT_MEMORY_RW_DTCM] = {
310 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
311 				L_PTE_XN,
312 		.prot_l1   = PMD_TYPE_TABLE,
313 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
314 		.domain    = DOMAIN_KERNEL,
315 	},
316 	[MT_MEMORY_RWX_ITCM] = {
317 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
318 		.prot_l1   = PMD_TYPE_TABLE,
319 		.domain    = DOMAIN_KERNEL,
320 	},
321 	[MT_MEMORY_RW_SO] = {
322 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
323 				L_PTE_MT_UNCACHED | L_PTE_XN,
324 		.prot_l1   = PMD_TYPE_TABLE,
325 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
326 				PMD_SECT_UNCACHED | PMD_SECT_XN,
327 		.domain    = DOMAIN_KERNEL,
328 	},
329 	[MT_MEMORY_DMA_READY] = {
330 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
331 				L_PTE_XN,
332 		.prot_l1   = PMD_TYPE_TABLE,
333 		.domain    = DOMAIN_KERNEL,
334 	},
335 };
336 
337 const struct mem_type *get_mem_type(unsigned int type)
338 {
339 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
340 }
341 EXPORT_SYMBOL(get_mem_type);
342 
343 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr);
344 
345 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS]
346 	__aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata;
347 
348 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr)
349 {
350 	return &bm_pte[pte_index(addr)];
351 }
352 
353 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr)
354 {
355 	return pte_offset_kernel(dir, addr);
356 }
357 
358 static inline pmd_t * __init fixmap_pmd(unsigned long addr)
359 {
360 	return pmd_off_k(addr);
361 }
362 
363 void __init early_fixmap_init(void)
364 {
365 	pmd_t *pmd;
366 
367 	/*
368 	 * The early fixmap range spans multiple pmds, for which
369 	 * we are not prepared:
370 	 */
371 	BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT)
372 		     != FIXADDR_TOP >> PMD_SHIFT);
373 
374 	pmd = fixmap_pmd(FIXADDR_TOP);
375 	pmd_populate_kernel(&init_mm, pmd, bm_pte);
376 
377 	pte_offset_fixmap = pte_offset_early_fixmap;
378 }
379 
380 /*
381  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
382  * As a result, this can only be called with preemption disabled, as under
383  * stop_machine().
384  */
385 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
386 {
387 	unsigned long vaddr = __fix_to_virt(idx);
388 	pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr);
389 
390 	/* Make sure fixmap region does not exceed available allocation. */
391 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
392 		     FIXADDR_END);
393 	BUG_ON(idx >= __end_of_fixed_addresses);
394 
395 	/* we only support device mappings until pgprot_kernel has been set */
396 	if (WARN_ON(pgprot_val(prot) != pgprot_val(FIXMAP_PAGE_IO) &&
397 		    pgprot_val(pgprot_kernel) == 0))
398 		return;
399 
400 	if (pgprot_val(prot))
401 		set_pte_at(NULL, vaddr, pte,
402 			pfn_pte(phys >> PAGE_SHIFT, prot));
403 	else
404 		pte_clear(NULL, vaddr, pte);
405 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
406 }
407 
408 /*
409  * Adjust the PMD section entries according to the CPU in use.
410  */
411 static void __init build_mem_type_table(void)
412 {
413 	struct cachepolicy *cp;
414 	unsigned int cr = get_cr();
415 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
416 	int cpu_arch = cpu_architecture();
417 	int i;
418 
419 	if (cpu_arch < CPU_ARCH_ARMv6) {
420 #if defined(CONFIG_CPU_DCACHE_DISABLE)
421 		if (cachepolicy > CPOLICY_BUFFERED)
422 			cachepolicy = CPOLICY_BUFFERED;
423 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
424 		if (cachepolicy > CPOLICY_WRITETHROUGH)
425 			cachepolicy = CPOLICY_WRITETHROUGH;
426 #endif
427 	}
428 	if (cpu_arch < CPU_ARCH_ARMv5) {
429 		if (cachepolicy >= CPOLICY_WRITEALLOC)
430 			cachepolicy = CPOLICY_WRITEBACK;
431 		ecc_mask = 0;
432 	}
433 
434 	if (is_smp()) {
435 		if (cachepolicy != CPOLICY_WRITEALLOC) {
436 			pr_warn("Forcing write-allocate cache policy for SMP\n");
437 			cachepolicy = CPOLICY_WRITEALLOC;
438 		}
439 		if (!(initial_pmd_value & PMD_SECT_S)) {
440 			pr_warn("Forcing shared mappings for SMP\n");
441 			initial_pmd_value |= PMD_SECT_S;
442 		}
443 	}
444 
445 	/*
446 	 * Strip out features not present on earlier architectures.
447 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
448 	 * without extended page tables don't have the 'Shared' bit.
449 	 */
450 	if (cpu_arch < CPU_ARCH_ARMv5)
451 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
452 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
453 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
454 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
455 			mem_types[i].prot_sect &= ~PMD_SECT_S;
456 
457 	/*
458 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
459 	 * "update-able on write" bit on ARM610).  However, Xscale and
460 	 * Xscale3 require this bit to be cleared.
461 	 */
462 	if (cpu_is_xscale_family()) {
463 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
464 			mem_types[i].prot_sect &= ~PMD_BIT4;
465 			mem_types[i].prot_l1 &= ~PMD_BIT4;
466 		}
467 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
468 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
469 			if (mem_types[i].prot_l1)
470 				mem_types[i].prot_l1 |= PMD_BIT4;
471 			if (mem_types[i].prot_sect)
472 				mem_types[i].prot_sect |= PMD_BIT4;
473 		}
474 	}
475 
476 	/*
477 	 * Mark the device areas according to the CPU/architecture.
478 	 */
479 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
480 		if (!cpu_is_xsc3()) {
481 			/*
482 			 * Mark device regions on ARMv6+ as execute-never
483 			 * to prevent speculative instruction fetches.
484 			 */
485 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
486 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
487 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
488 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
489 
490 			/* Also setup NX memory mapping */
491 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
492 		}
493 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
494 			/*
495 			 * For ARMv7 with TEX remapping,
496 			 * - shared device is SXCB=1100
497 			 * - nonshared device is SXCB=0100
498 			 * - write combine device mem is SXCB=0001
499 			 * (Uncached Normal memory)
500 			 */
501 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
502 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
503 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
504 		} else if (cpu_is_xsc3()) {
505 			/*
506 			 * For Xscale3,
507 			 * - shared device is TEXCB=00101
508 			 * - nonshared device is TEXCB=01000
509 			 * - write combine device mem is TEXCB=00100
510 			 * (Inner/Outer Uncacheable in xsc3 parlance)
511 			 */
512 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
513 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
514 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
515 		} else {
516 			/*
517 			 * For ARMv6 and ARMv7 without TEX remapping,
518 			 * - shared device is TEXCB=00001
519 			 * - nonshared device is TEXCB=01000
520 			 * - write combine device mem is TEXCB=00100
521 			 * (Uncached Normal in ARMv6 parlance).
522 			 */
523 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
524 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
525 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
526 		}
527 	} else {
528 		/*
529 		 * On others, write combining is "Uncached/Buffered"
530 		 */
531 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
532 	}
533 
534 	/*
535 	 * Now deal with the memory-type mappings
536 	 */
537 	cp = &cache_policies[cachepolicy];
538 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
539 
540 #ifndef CONFIG_ARM_LPAE
541 	/*
542 	 * We don't use domains on ARMv6 (since this causes problems with
543 	 * v6/v7 kernels), so we must use a separate memory type for user
544 	 * r/o, kernel r/w to map the vectors page.
545 	 */
546 	if (cpu_arch == CPU_ARCH_ARMv6)
547 		vecs_pgprot |= L_PTE_MT_VECTORS;
548 
549 	/*
550 	 * Check is it with support for the PXN bit
551 	 * in the Short-descriptor translation table format descriptors.
552 	 */
553 	if (cpu_arch == CPU_ARCH_ARMv7 &&
554 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
555 		user_pmd_table |= PMD_PXNTABLE;
556 	}
557 #endif
558 
559 	/*
560 	 * ARMv6 and above have extended page tables.
561 	 */
562 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
563 #ifndef CONFIG_ARM_LPAE
564 		/*
565 		 * Mark cache clean areas and XIP ROM read only
566 		 * from SVC mode and no access from userspace.
567 		 */
568 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
569 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
570 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
571 #endif
572 
573 		/*
574 		 * If the initial page tables were created with the S bit
575 		 * set, then we need to do the same here for the same
576 		 * reasons given in early_cachepolicy().
577 		 */
578 		if (initial_pmd_value & PMD_SECT_S) {
579 			user_pgprot |= L_PTE_SHARED;
580 			kern_pgprot |= L_PTE_SHARED;
581 			vecs_pgprot |= L_PTE_SHARED;
582 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
583 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
584 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
585 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
586 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
587 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
588 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
589 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
590 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
591 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
592 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
593 		}
594 	}
595 
596 	/*
597 	 * Non-cacheable Normal - intended for memory areas that must
598 	 * not cause dirty cache line writebacks when used
599 	 */
600 	if (cpu_arch >= CPU_ARCH_ARMv6) {
601 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
602 			/* Non-cacheable Normal is XCB = 001 */
603 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
604 				PMD_SECT_BUFFERED;
605 		} else {
606 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
607 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
608 				PMD_SECT_TEX(1);
609 		}
610 	} else {
611 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
612 	}
613 
614 #ifdef CONFIG_ARM_LPAE
615 	/*
616 	 * Do not generate access flag faults for the kernel mappings.
617 	 */
618 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
619 		mem_types[i].prot_pte |= PTE_EXT_AF;
620 		if (mem_types[i].prot_sect)
621 			mem_types[i].prot_sect |= PMD_SECT_AF;
622 	}
623 	kern_pgprot |= PTE_EXT_AF;
624 	vecs_pgprot |= PTE_EXT_AF;
625 
626 	/*
627 	 * Set PXN for user mappings
628 	 */
629 	user_pgprot |= PTE_EXT_PXN;
630 #endif
631 
632 	for (i = 0; i < 16; i++) {
633 		pteval_t v = pgprot_val(protection_map[i]);
634 		protection_map[i] = __pgprot(v | user_pgprot);
635 	}
636 
637 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
638 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
639 
640 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
641 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
642 				 L_PTE_DIRTY | kern_pgprot);
643 
644 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
645 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
646 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
647 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
648 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
649 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
650 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
651 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
652 	mem_types[MT_ROM].prot_sect |= cp->pmd;
653 
654 	switch (cp->pmd) {
655 	case PMD_SECT_WT:
656 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
657 		break;
658 	case PMD_SECT_WB:
659 	case PMD_SECT_WBWA:
660 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
661 		break;
662 	}
663 	pr_info("Memory policy: %sData cache %s\n",
664 		ecc_mask ? "ECC enabled, " : "", cp->policy);
665 
666 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
667 		struct mem_type *t = &mem_types[i];
668 		if (t->prot_l1)
669 			t->prot_l1 |= PMD_DOMAIN(t->domain);
670 		if (t->prot_sect)
671 			t->prot_sect |= PMD_DOMAIN(t->domain);
672 	}
673 }
674 
675 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
676 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
677 			      unsigned long size, pgprot_t vma_prot)
678 {
679 	if (!pfn_valid(pfn))
680 		return pgprot_noncached(vma_prot);
681 	else if (file->f_flags & O_SYNC)
682 		return pgprot_writecombine(vma_prot);
683 	return vma_prot;
684 }
685 EXPORT_SYMBOL(phys_mem_access_prot);
686 #endif
687 
688 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
689 
690 static void __init *early_alloc(unsigned long sz)
691 {
692 	void *ptr = memblock_alloc(sz, sz);
693 
694 	if (!ptr)
695 		panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
696 		      __func__, sz, sz);
697 
698 	return ptr;
699 }
700 
701 static void *__init late_alloc(unsigned long sz)
702 {
703 	void *ptr = (void *)__get_free_pages(GFP_PGTABLE_KERNEL, get_order(sz));
704 
705 	if (!ptr || !pgtable_pte_page_ctor(virt_to_page(ptr)))
706 		BUG();
707 	return ptr;
708 }
709 
710 static pte_t * __init arm_pte_alloc(pmd_t *pmd, unsigned long addr,
711 				unsigned long prot,
712 				void *(*alloc)(unsigned long sz))
713 {
714 	if (pmd_none(*pmd)) {
715 		pte_t *pte = alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
716 		__pmd_populate(pmd, __pa(pte), prot);
717 	}
718 	BUG_ON(pmd_bad(*pmd));
719 	return pte_offset_kernel(pmd, addr);
720 }
721 
722 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr,
723 				      unsigned long prot)
724 {
725 	return arm_pte_alloc(pmd, addr, prot, early_alloc);
726 }
727 
728 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
729 				  unsigned long end, unsigned long pfn,
730 				  const struct mem_type *type,
731 				  void *(*alloc)(unsigned long sz),
732 				  bool ng)
733 {
734 	pte_t *pte = arm_pte_alloc(pmd, addr, type->prot_l1, alloc);
735 	do {
736 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)),
737 			    ng ? PTE_EXT_NG : 0);
738 		pfn++;
739 	} while (pte++, addr += PAGE_SIZE, addr != end);
740 }
741 
742 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
743 			unsigned long end, phys_addr_t phys,
744 			const struct mem_type *type, bool ng)
745 {
746 	pmd_t *p = pmd;
747 
748 #ifndef CONFIG_ARM_LPAE
749 	/*
750 	 * In classic MMU format, puds and pmds are folded in to
751 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
752 	 * group of L1 entries making up one logical pointer to
753 	 * an L2 table (2MB), where as PMDs refer to the individual
754 	 * L1 entries (1MB). Hence increment to get the correct
755 	 * offset for odd 1MB sections.
756 	 * (See arch/arm/include/asm/pgtable-2level.h)
757 	 */
758 	if (addr & SECTION_SIZE)
759 		pmd++;
760 #endif
761 	do {
762 		*pmd = __pmd(phys | type->prot_sect | (ng ? PMD_SECT_nG : 0));
763 		phys += SECTION_SIZE;
764 	} while (pmd++, addr += SECTION_SIZE, addr != end);
765 
766 	flush_pmd_entry(p);
767 }
768 
769 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
770 				      unsigned long end, phys_addr_t phys,
771 				      const struct mem_type *type,
772 				      void *(*alloc)(unsigned long sz), bool ng)
773 {
774 	pmd_t *pmd = pmd_offset(pud, addr);
775 	unsigned long next;
776 
777 	do {
778 		/*
779 		 * With LPAE, we must loop over to map
780 		 * all the pmds for the given range.
781 		 */
782 		next = pmd_addr_end(addr, end);
783 
784 		/*
785 		 * Try a section mapping - addr, next and phys must all be
786 		 * aligned to a section boundary.
787 		 */
788 		if (type->prot_sect &&
789 				((addr | next | phys) & ~SECTION_MASK) == 0) {
790 			__map_init_section(pmd, addr, next, phys, type, ng);
791 		} else {
792 			alloc_init_pte(pmd, addr, next,
793 				       __phys_to_pfn(phys), type, alloc, ng);
794 		}
795 
796 		phys += next - addr;
797 
798 	} while (pmd++, addr = next, addr != end);
799 }
800 
801 static void __init alloc_init_pud(p4d_t *p4d, 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 	pud_t *pud = pud_offset(p4d, addr);
807 	unsigned long next;
808 
809 	do {
810 		next = pud_addr_end(addr, end);
811 		alloc_init_pmd(pud, addr, next, phys, type, alloc, ng);
812 		phys += next - addr;
813 	} while (pud++, addr = next, addr != end);
814 }
815 
816 static void __init alloc_init_p4d(pgd_t *pgd, unsigned long addr,
817 				  unsigned long end, phys_addr_t phys,
818 				  const struct mem_type *type,
819 				  void *(*alloc)(unsigned long sz), bool ng)
820 {
821 	p4d_t *p4d = p4d_offset(pgd, addr);
822 	unsigned long next;
823 
824 	do {
825 		next = p4d_addr_end(addr, end);
826 		alloc_init_pud(p4d, addr, next, phys, type, alloc, ng);
827 		phys += next - addr;
828 	} while (p4d++, addr = next, addr != end);
829 }
830 
831 #ifndef CONFIG_ARM_LPAE
832 static void __init create_36bit_mapping(struct mm_struct *mm,
833 					struct map_desc *md,
834 					const struct mem_type *type,
835 					bool ng)
836 {
837 	unsigned long addr, length, end;
838 	phys_addr_t phys;
839 	pgd_t *pgd;
840 
841 	addr = md->virtual;
842 	phys = __pfn_to_phys(md->pfn);
843 	length = PAGE_ALIGN(md->length);
844 
845 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
846 		pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n",
847 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
848 		return;
849 	}
850 
851 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
852 	 *	Since domain assignments can in fact be arbitrary, the
853 	 *	'domain == 0' check below is required to insure that ARMv6
854 	 *	supersections are only allocated for domain 0 regardless
855 	 *	of the actual domain assignments in use.
856 	 */
857 	if (type->domain) {
858 		pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n",
859 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
860 		return;
861 	}
862 
863 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
864 		pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n",
865 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
866 		return;
867 	}
868 
869 	/*
870 	 * Shift bits [35:32] of address into bits [23:20] of PMD
871 	 * (See ARMv6 spec).
872 	 */
873 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
874 
875 	pgd = pgd_offset(mm, addr);
876 	end = addr + length;
877 	do {
878 		p4d_t *p4d = p4d_offset(pgd, addr);
879 		pud_t *pud = pud_offset(p4d, addr);
880 		pmd_t *pmd = pmd_offset(pud, addr);
881 		int i;
882 
883 		for (i = 0; i < 16; i++)
884 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER |
885 				       (ng ? PMD_SECT_nG : 0));
886 
887 		addr += SUPERSECTION_SIZE;
888 		phys += SUPERSECTION_SIZE;
889 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
890 	} while (addr != end);
891 }
892 #endif	/* !CONFIG_ARM_LPAE */
893 
894 static void __init __create_mapping(struct mm_struct *mm, struct map_desc *md,
895 				    void *(*alloc)(unsigned long sz),
896 				    bool ng)
897 {
898 	unsigned long addr, length, end;
899 	phys_addr_t phys;
900 	const struct mem_type *type;
901 	pgd_t *pgd;
902 
903 	type = &mem_types[md->type];
904 
905 #ifndef CONFIG_ARM_LPAE
906 	/*
907 	 * Catch 36-bit addresses
908 	 */
909 	if (md->pfn >= 0x100000) {
910 		create_36bit_mapping(mm, md, type, ng);
911 		return;
912 	}
913 #endif
914 
915 	addr = md->virtual & PAGE_MASK;
916 	phys = __pfn_to_phys(md->pfn);
917 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
918 
919 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
920 		pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n",
921 			(long long)__pfn_to_phys(md->pfn), addr);
922 		return;
923 	}
924 
925 	pgd = pgd_offset(mm, addr);
926 	end = addr + length;
927 	do {
928 		unsigned long next = pgd_addr_end(addr, end);
929 
930 		alloc_init_p4d(pgd, addr, next, phys, type, alloc, ng);
931 
932 		phys += next - addr;
933 		addr = next;
934 	} while (pgd++, addr != end);
935 }
936 
937 /*
938  * Create the page directory entries and any necessary
939  * page tables for the mapping specified by `md'.  We
940  * are able to cope here with varying sizes and address
941  * offsets, and we take full advantage of sections and
942  * supersections.
943  */
944 static void __init create_mapping(struct map_desc *md)
945 {
946 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
947 		pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n",
948 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
949 		return;
950 	}
951 
952 	if (md->type == MT_DEVICE &&
953 	    md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START &&
954 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
955 		pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n",
956 			(long long)__pfn_to_phys((u64)md->pfn), md->virtual);
957 	}
958 
959 	__create_mapping(&init_mm, md, early_alloc, false);
960 }
961 
962 void __init create_mapping_late(struct mm_struct *mm, struct map_desc *md,
963 				bool ng)
964 {
965 #ifdef CONFIG_ARM_LPAE
966 	p4d_t *p4d;
967 	pud_t *pud;
968 
969 	p4d = p4d_alloc(mm, pgd_offset(mm, md->virtual), md->virtual);
970 	if (WARN_ON(!p4d))
971 		return;
972 	pud = pud_alloc(mm, p4d, md->virtual);
973 	if (WARN_ON(!pud))
974 		return;
975 	pmd_alloc(mm, pud, 0);
976 #endif
977 	__create_mapping(mm, md, late_alloc, ng);
978 }
979 
980 /*
981  * Create the architecture specific mappings
982  */
983 void __init iotable_init(struct map_desc *io_desc, int nr)
984 {
985 	struct map_desc *md;
986 	struct vm_struct *vm;
987 	struct static_vm *svm;
988 
989 	if (!nr)
990 		return;
991 
992 	svm = memblock_alloc(sizeof(*svm) * nr, __alignof__(*svm));
993 	if (!svm)
994 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
995 		      __func__, sizeof(*svm) * nr, __alignof__(*svm));
996 
997 	for (md = io_desc; nr; md++, nr--) {
998 		create_mapping(md);
999 
1000 		vm = &svm->vm;
1001 		vm->addr = (void *)(md->virtual & PAGE_MASK);
1002 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
1003 		vm->phys_addr = __pfn_to_phys(md->pfn);
1004 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
1005 		vm->flags |= VM_ARM_MTYPE(md->type);
1006 		vm->caller = iotable_init;
1007 		add_static_vm_early(svm++);
1008 	}
1009 }
1010 
1011 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
1012 				  void *caller)
1013 {
1014 	struct vm_struct *vm;
1015 	struct static_vm *svm;
1016 
1017 	svm = memblock_alloc(sizeof(*svm), __alignof__(*svm));
1018 	if (!svm)
1019 		panic("%s: Failed to allocate %zu bytes align=0x%zx\n",
1020 		      __func__, sizeof(*svm), __alignof__(*svm));
1021 
1022 	vm = &svm->vm;
1023 	vm->addr = (void *)addr;
1024 	vm->size = size;
1025 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
1026 	vm->caller = caller;
1027 	add_static_vm_early(svm);
1028 }
1029 
1030 #ifndef CONFIG_ARM_LPAE
1031 
1032 /*
1033  * The Linux PMD is made of two consecutive section entries covering 2MB
1034  * (see definition in include/asm/pgtable-2level.h).  However a call to
1035  * create_mapping() may optimize static mappings by using individual
1036  * 1MB section mappings.  This leaves the actual PMD potentially half
1037  * initialized if the top or bottom section entry isn't used, leaving it
1038  * open to problems if a subsequent ioremap() or vmalloc() tries to use
1039  * the virtual space left free by that unused section entry.
1040  *
1041  * Let's avoid the issue by inserting dummy vm entries covering the unused
1042  * PMD halves once the static mappings are in place.
1043  */
1044 
1045 static void __init pmd_empty_section_gap(unsigned long addr)
1046 {
1047 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1048 }
1049 
1050 static void __init fill_pmd_gaps(void)
1051 {
1052 	struct static_vm *svm;
1053 	struct vm_struct *vm;
1054 	unsigned long addr, next = 0;
1055 	pmd_t *pmd;
1056 
1057 	list_for_each_entry(svm, &static_vmlist, list) {
1058 		vm = &svm->vm;
1059 		addr = (unsigned long)vm->addr;
1060 		if (addr < next)
1061 			continue;
1062 
1063 		/*
1064 		 * Check if this vm starts on an odd section boundary.
1065 		 * If so and the first section entry for this PMD is free
1066 		 * then we block the corresponding virtual address.
1067 		 */
1068 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1069 			pmd = pmd_off_k(addr);
1070 			if (pmd_none(*pmd))
1071 				pmd_empty_section_gap(addr & PMD_MASK);
1072 		}
1073 
1074 		/*
1075 		 * Then check if this vm ends on an odd section boundary.
1076 		 * If so and the second section entry for this PMD is empty
1077 		 * then we block the corresponding virtual address.
1078 		 */
1079 		addr += vm->size;
1080 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1081 			pmd = pmd_off_k(addr) + 1;
1082 			if (pmd_none(*pmd))
1083 				pmd_empty_section_gap(addr);
1084 		}
1085 
1086 		/* no need to look at any vm entry until we hit the next PMD */
1087 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1088 	}
1089 }
1090 
1091 #else
1092 #define fill_pmd_gaps() do { } while (0)
1093 #endif
1094 
1095 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
1096 static void __init pci_reserve_io(void)
1097 {
1098 	struct static_vm *svm;
1099 
1100 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1101 	if (svm)
1102 		return;
1103 
1104 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1105 }
1106 #else
1107 #define pci_reserve_io() do { } while (0)
1108 #endif
1109 
1110 #ifdef CONFIG_DEBUG_LL
1111 void __init debug_ll_io_init(void)
1112 {
1113 	struct map_desc map;
1114 
1115 	debug_ll_addr(&map.pfn, &map.virtual);
1116 	if (!map.pfn || !map.virtual)
1117 		return;
1118 	map.pfn = __phys_to_pfn(map.pfn);
1119 	map.virtual &= PAGE_MASK;
1120 	map.length = PAGE_SIZE;
1121 	map.type = MT_DEVICE;
1122 	iotable_init(&map, 1);
1123 }
1124 #endif
1125 
1126 static void * __initdata vmalloc_min =
1127 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1128 
1129 /*
1130  * vmalloc=size forces the vmalloc area to be exactly 'size'
1131  * bytes. This can be used to increase (or decrease) the vmalloc
1132  * area - the default is 240m.
1133  */
1134 static int __init early_vmalloc(char *arg)
1135 {
1136 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1137 
1138 	if (vmalloc_reserve < SZ_16M) {
1139 		vmalloc_reserve = SZ_16M;
1140 		pr_warn("vmalloc area too small, limiting to %luMB\n",
1141 			vmalloc_reserve >> 20);
1142 	}
1143 
1144 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1145 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1146 		pr_warn("vmalloc area is too big, limiting to %luMB\n",
1147 			vmalloc_reserve >> 20);
1148 	}
1149 
1150 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1151 	return 0;
1152 }
1153 early_param("vmalloc", early_vmalloc);
1154 
1155 phys_addr_t arm_lowmem_limit __initdata = 0;
1156 
1157 void __init adjust_lowmem_bounds(void)
1158 {
1159 	phys_addr_t block_start, block_end, memblock_limit = 0;
1160 	u64 vmalloc_limit, i;
1161 	phys_addr_t lowmem_limit = 0;
1162 
1163 	/*
1164 	 * Let's use our own (unoptimized) equivalent of __pa() that is
1165 	 * not affected by wrap-arounds when sizeof(phys_addr_t) == 4.
1166 	 * The result is used as the upper bound on physical memory address
1167 	 * and may itself be outside the valid range for which phys_addr_t
1168 	 * and therefore __pa() is defined.
1169 	 */
1170 	vmalloc_limit = (u64)(uintptr_t)vmalloc_min - PAGE_OFFSET + PHYS_OFFSET;
1171 
1172 	/*
1173 	 * The first usable region must be PMD aligned. Mark its start
1174 	 * as MEMBLOCK_NOMAP if it isn't
1175 	 */
1176 	for_each_mem_range(i, &block_start, &block_end) {
1177 		if (!IS_ALIGNED(block_start, PMD_SIZE)) {
1178 			phys_addr_t len;
1179 
1180 			len = round_up(block_start, PMD_SIZE) - block_start;
1181 			memblock_mark_nomap(block_start, len);
1182 		}
1183 		break;
1184 	}
1185 
1186 	for_each_mem_range(i, &block_start, &block_end) {
1187 		if (block_start < 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 #ifdef CONFIG_KASAN
1260 	/*
1261 	 * KASan's shadow memory inserts itself between the TASK_SIZE
1262 	 * and MODULES_VADDR. Do not clear the KASan shadow memory mappings.
1263 	 */
1264 	for (addr = 0; addr < KASAN_SHADOW_START; addr += PMD_SIZE)
1265 		pmd_clear(pmd_off_k(addr));
1266 	/*
1267 	 * Skip over the KASan shadow area. KASAN_SHADOW_END is sometimes
1268 	 * equal to MODULES_VADDR and then we exit the pmd clearing. If we
1269 	 * are using a thumb-compiled kernel, there there will be 8MB more
1270 	 * to clear as KASan always offset to 16 MB below MODULES_VADDR.
1271 	 */
1272 	for (addr = KASAN_SHADOW_END; addr < MODULES_VADDR; addr += PMD_SIZE)
1273 		pmd_clear(pmd_off_k(addr));
1274 #else
1275 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1276 		pmd_clear(pmd_off_k(addr));
1277 #endif
1278 
1279 #ifdef CONFIG_XIP_KERNEL
1280 	/* The XIP kernel is mapped in the module area -- skip over it */
1281 	addr = ((unsigned long)_exiprom + PMD_SIZE - 1) & PMD_MASK;
1282 #endif
1283 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1284 		pmd_clear(pmd_off_k(addr));
1285 
1286 	/*
1287 	 * Find the end of the first block of lowmem.
1288 	 */
1289 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1290 	if (end >= arm_lowmem_limit)
1291 		end = arm_lowmem_limit;
1292 
1293 	/*
1294 	 * Clear out all the kernel space mappings, except for the first
1295 	 * memory bank, up to the vmalloc region.
1296 	 */
1297 	for (addr = __phys_to_virt(end);
1298 	     addr < VMALLOC_START; addr += PMD_SIZE)
1299 		pmd_clear(pmd_off_k(addr));
1300 }
1301 
1302 #ifdef CONFIG_ARM_LPAE
1303 /* the first page is reserved for pgd */
1304 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1305 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1306 #else
1307 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1308 #endif
1309 
1310 /*
1311  * Reserve the special regions of memory
1312  */
1313 void __init arm_mm_memblock_reserve(void)
1314 {
1315 	/*
1316 	 * Reserve the page tables.  These are already in use,
1317 	 * and can only be in node 0.
1318 	 */
1319 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1320 
1321 #ifdef CONFIG_SA1111
1322 	/*
1323 	 * Because of the SA1111 DMA bug, we want to preserve our
1324 	 * precious DMA-able memory...
1325 	 */
1326 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1327 #endif
1328 }
1329 
1330 /*
1331  * Set up the device mappings.  Since we clear out the page tables for all
1332  * mappings above VMALLOC_START, except early fixmap, we might remove debug
1333  * device mappings.  This means earlycon can be used to debug this function
1334  * Any other function or debugging method which may touch any device _will_
1335  * crash the kernel.
1336  */
1337 static void __init devicemaps_init(const struct machine_desc *mdesc)
1338 {
1339 	struct map_desc map;
1340 	unsigned long addr;
1341 	void *vectors;
1342 
1343 	/*
1344 	 * Allocate the vector page early.
1345 	 */
1346 	vectors = early_alloc(PAGE_SIZE * 2);
1347 
1348 	early_trap_init(vectors);
1349 
1350 	/*
1351 	 * Clear page table except top pmd used by early fixmaps
1352 	 */
1353 	for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE)
1354 		pmd_clear(pmd_off_k(addr));
1355 
1356 	if (__atags_pointer) {
1357 		/* create a read-only mapping of the device tree */
1358 		map.pfn = __phys_to_pfn(__atags_pointer & SECTION_MASK);
1359 		map.virtual = FDT_FIXED_BASE;
1360 		map.length = FDT_FIXED_SIZE;
1361 		map.type = MT_ROM;
1362 		create_mapping(&map);
1363 	}
1364 
1365 	/*
1366 	 * Map the kernel if it is XIP.
1367 	 * It is always first in the modulearea.
1368 	 */
1369 #ifdef CONFIG_XIP_KERNEL
1370 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1371 	map.virtual = MODULES_VADDR;
1372 	map.length = ((unsigned long)_exiprom - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1373 	map.type = MT_ROM;
1374 	create_mapping(&map);
1375 #endif
1376 
1377 	/*
1378 	 * Map the cache flushing regions.
1379 	 */
1380 #ifdef FLUSH_BASE
1381 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1382 	map.virtual = FLUSH_BASE;
1383 	map.length = SZ_1M;
1384 	map.type = MT_CACHECLEAN;
1385 	create_mapping(&map);
1386 #endif
1387 #ifdef FLUSH_BASE_MINICACHE
1388 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1389 	map.virtual = FLUSH_BASE_MINICACHE;
1390 	map.length = SZ_1M;
1391 	map.type = MT_MINICLEAN;
1392 	create_mapping(&map);
1393 #endif
1394 
1395 	/*
1396 	 * Create a mapping for the machine vectors at the high-vectors
1397 	 * location (0xffff0000).  If we aren't using high-vectors, also
1398 	 * create a mapping at the low-vectors virtual address.
1399 	 */
1400 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1401 	map.virtual = 0xffff0000;
1402 	map.length = PAGE_SIZE;
1403 #ifdef CONFIG_KUSER_HELPERS
1404 	map.type = MT_HIGH_VECTORS;
1405 #else
1406 	map.type = MT_LOW_VECTORS;
1407 #endif
1408 	create_mapping(&map);
1409 
1410 	if (!vectors_high()) {
1411 		map.virtual = 0;
1412 		map.length = PAGE_SIZE * 2;
1413 		map.type = MT_LOW_VECTORS;
1414 		create_mapping(&map);
1415 	}
1416 
1417 	/* Now create a kernel read-only mapping */
1418 	map.pfn += 1;
1419 	map.virtual = 0xffff0000 + PAGE_SIZE;
1420 	map.length = PAGE_SIZE;
1421 	map.type = MT_LOW_VECTORS;
1422 	create_mapping(&map);
1423 
1424 	/*
1425 	 * Ask the machine support to map in the statically mapped devices.
1426 	 */
1427 	if (mdesc->map_io)
1428 		mdesc->map_io();
1429 	else
1430 		debug_ll_io_init();
1431 	fill_pmd_gaps();
1432 
1433 	/* Reserve fixed i/o space in VMALLOC region */
1434 	pci_reserve_io();
1435 
1436 	/*
1437 	 * Finally flush the caches and tlb to ensure that we're in a
1438 	 * consistent state wrt the writebuffer.  This also ensures that
1439 	 * any write-allocated cache lines in the vector page are written
1440 	 * back.  After this point, we can start to touch devices again.
1441 	 */
1442 	local_flush_tlb_all();
1443 	flush_cache_all();
1444 
1445 	/* Enable asynchronous aborts */
1446 	early_abt_enable();
1447 }
1448 
1449 static void __init kmap_init(void)
1450 {
1451 #ifdef CONFIG_HIGHMEM
1452 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1453 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1454 #endif
1455 
1456 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1457 			_PAGE_KERNEL_TABLE);
1458 }
1459 
1460 static void __init map_lowmem(void)
1461 {
1462 	phys_addr_t kernel_x_start = round_down(__pa(KERNEL_START), SECTION_SIZE);
1463 	phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1464 	phys_addr_t start, end;
1465 	u64 i;
1466 
1467 	/* Map all the lowmem memory banks. */
1468 	for_each_mem_range(i, &start, &end) {
1469 		struct map_desc map;
1470 
1471 		if (end > arm_lowmem_limit)
1472 			end = arm_lowmem_limit;
1473 		if (start >= end)
1474 			break;
1475 
1476 		if (end < kernel_x_start) {
1477 			map.pfn = __phys_to_pfn(start);
1478 			map.virtual = __phys_to_virt(start);
1479 			map.length = end - start;
1480 			map.type = MT_MEMORY_RWX;
1481 
1482 			create_mapping(&map);
1483 		} else if (start >= kernel_x_end) {
1484 			map.pfn = __phys_to_pfn(start);
1485 			map.virtual = __phys_to_virt(start);
1486 			map.length = end - start;
1487 			map.type = MT_MEMORY_RW;
1488 
1489 			create_mapping(&map);
1490 		} else {
1491 			/* This better cover the entire kernel */
1492 			if (start < kernel_x_start) {
1493 				map.pfn = __phys_to_pfn(start);
1494 				map.virtual = __phys_to_virt(start);
1495 				map.length = kernel_x_start - start;
1496 				map.type = MT_MEMORY_RW;
1497 
1498 				create_mapping(&map);
1499 			}
1500 
1501 			map.pfn = __phys_to_pfn(kernel_x_start);
1502 			map.virtual = __phys_to_virt(kernel_x_start);
1503 			map.length = kernel_x_end - kernel_x_start;
1504 			map.type = MT_MEMORY_RWX;
1505 
1506 			create_mapping(&map);
1507 
1508 			if (kernel_x_end < end) {
1509 				map.pfn = __phys_to_pfn(kernel_x_end);
1510 				map.virtual = __phys_to_virt(kernel_x_end);
1511 				map.length = end - kernel_x_end;
1512 				map.type = MT_MEMORY_RW;
1513 
1514 				create_mapping(&map);
1515 			}
1516 		}
1517 	}
1518 }
1519 
1520 #ifdef CONFIG_ARM_PV_FIXUP
1521 typedef void pgtables_remap(long long offset, unsigned long pgd);
1522 pgtables_remap lpae_pgtables_remap_asm;
1523 
1524 /*
1525  * early_paging_init() recreates boot time page table setup, allowing machines
1526  * to switch over to a high (>4G) address space on LPAE systems
1527  */
1528 static void __init early_paging_init(const struct machine_desc *mdesc)
1529 {
1530 	pgtables_remap *lpae_pgtables_remap;
1531 	unsigned long pa_pgd;
1532 	unsigned int cr, ttbcr;
1533 	long long offset;
1534 
1535 	if (!mdesc->pv_fixup)
1536 		return;
1537 
1538 	offset = mdesc->pv_fixup();
1539 	if (offset == 0)
1540 		return;
1541 
1542 	/*
1543 	 * Get the address of the remap function in the 1:1 identity
1544 	 * mapping setup by the early page table assembly code.  We
1545 	 * must get this prior to the pv update.  The following barrier
1546 	 * ensures that this is complete before we fixup any P:V offsets.
1547 	 */
1548 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1549 	pa_pgd = __pa(swapper_pg_dir);
1550 	barrier();
1551 
1552 	pr_info("Switching physical address space to 0x%08llx\n",
1553 		(u64)PHYS_OFFSET + offset);
1554 
1555 	/* Re-set the phys pfn offset, and the pv offset */
1556 	__pv_offset += offset;
1557 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1558 
1559 	/* Run the patch stub to update the constants */
1560 	fixup_pv_table(&__pv_table_begin,
1561 		(&__pv_table_end - &__pv_table_begin) << 2);
1562 
1563 	/*
1564 	 * We changing not only the virtual to physical mapping, but also
1565 	 * the physical addresses used to access memory.  We need to flush
1566 	 * all levels of cache in the system with caching disabled to
1567 	 * ensure that all data is written back, and nothing is prefetched
1568 	 * into the caches.  We also need to prevent the TLB walkers
1569 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1570 	 * specific.
1571 	 */
1572 	cr = get_cr();
1573 	set_cr(cr & ~(CR_I | CR_C));
1574 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1575 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1576 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1577 	flush_cache_all();
1578 
1579 	/*
1580 	 * Fixup the page tables - this must be in the idmap region as
1581 	 * we need to disable the MMU to do this safely, and hence it
1582 	 * needs to be assembly.  It's fairly simple, as we're using the
1583 	 * temporary tables setup by the initial assembly code.
1584 	 */
1585 	lpae_pgtables_remap(offset, pa_pgd);
1586 
1587 	/* Re-enable the caches and cacheable TLB walks */
1588 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1589 	set_cr(cr);
1590 }
1591 
1592 #else
1593 
1594 static void __init early_paging_init(const struct machine_desc *mdesc)
1595 {
1596 	long long offset;
1597 
1598 	if (!mdesc->pv_fixup)
1599 		return;
1600 
1601 	offset = mdesc->pv_fixup();
1602 	if (offset == 0)
1603 		return;
1604 
1605 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1606 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1607 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1608 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1609 }
1610 
1611 #endif
1612 
1613 static void __init early_fixmap_shutdown(void)
1614 {
1615 	int i;
1616 	unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1);
1617 
1618 	pte_offset_fixmap = pte_offset_late_fixmap;
1619 	pmd_clear(fixmap_pmd(va));
1620 	local_flush_tlb_kernel_page(va);
1621 
1622 	for (i = 0; i < __end_of_permanent_fixed_addresses; i++) {
1623 		pte_t *pte;
1624 		struct map_desc map;
1625 
1626 		map.virtual = fix_to_virt(i);
1627 		pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual);
1628 
1629 		/* Only i/o device mappings are supported ATM */
1630 		if (pte_none(*pte) ||
1631 		    (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED)
1632 			continue;
1633 
1634 		map.pfn = pte_pfn(*pte);
1635 		map.type = MT_DEVICE;
1636 		map.length = PAGE_SIZE;
1637 
1638 		create_mapping(&map);
1639 	}
1640 }
1641 
1642 /*
1643  * paging_init() sets up the page tables, initialises the zone memory
1644  * maps, and sets up the zero page, bad page and bad page tables.
1645  */
1646 void __init paging_init(const struct machine_desc *mdesc)
1647 {
1648 	void *zero_page;
1649 
1650 	prepare_page_table();
1651 	map_lowmem();
1652 	memblock_set_current_limit(arm_lowmem_limit);
1653 	dma_contiguous_remap();
1654 	early_fixmap_shutdown();
1655 	devicemaps_init(mdesc);
1656 	kmap_init();
1657 	tcm_init();
1658 
1659 	top_pmd = pmd_off_k(0xffff0000);
1660 
1661 	/* allocate the zero page. */
1662 	zero_page = early_alloc(PAGE_SIZE);
1663 
1664 	bootmem_init();
1665 
1666 	empty_zero_page = virt_to_page(zero_page);
1667 	__flush_dcache_page(NULL, empty_zero_page);
1668 }
1669 
1670 void __init early_mm_init(const struct machine_desc *mdesc)
1671 {
1672 	build_mem_type_table();
1673 	early_paging_init(mdesc);
1674 }
1675 
1676 void set_pte_at(struct mm_struct *mm, unsigned long addr,
1677 			      pte_t *ptep, pte_t pteval)
1678 {
1679 	unsigned long ext = 0;
1680 
1681 	if (addr < TASK_SIZE && pte_valid_user(pteval)) {
1682 		if (!pte_special(pteval))
1683 			__sync_icache_dcache(pteval);
1684 		ext |= PTE_EXT_NG;
1685 	}
1686 
1687 	set_pte_ext(ptep, pteval, ext);
1688 }
1689