• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  linux/arch/arm/mm/mmu.c
3  *
4  *  Copyright (C) 1995-2005 Russell King
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  */
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/mman.h>
15 #include <linux/nodemask.h>
16 #include <linux/memblock.h>
17 #include <linux/fs.h>
18 #include <linux/vmalloc.h>
19 #include <linux/sizes.h>
20 
21 #include <asm/cp15.h>
22 #include <asm/cputype.h>
23 #include <asm/sections.h>
24 #include <asm/cachetype.h>
25 #include <asm/fixmap.h>
26 #include <asm/sections.h>
27 #include <asm/setup.h>
28 #include <asm/smp_plat.h>
29 #include <asm/tlb.h>
30 #include <asm/highmem.h>
31 #include <asm/system_info.h>
32 #include <asm/traps.h>
33 #include <asm/procinfo.h>
34 #include <asm/memory.h>
35 
36 #include <asm/mach/arch.h>
37 #include <asm/mach/map.h>
38 #include <asm/mach/pci.h>
39 #include <asm/fixmap.h>
40 
41 #include "mm.h"
42 #include "tcm.h"
43 
44 /*
45  * empty_zero_page is a special page that is used for
46  * zero-initialized data and COW.
47  */
48 struct page *empty_zero_page;
49 EXPORT_SYMBOL(empty_zero_page);
50 
51 /*
52  * The pmd table for the upper-most set of pages.
53  */
54 pmd_t *top_pmd;
55 
56 pmdval_t user_pmd_table = _PAGE_USER_TABLE;
57 
58 #define CPOLICY_UNCACHED	0
59 #define CPOLICY_BUFFERED	1
60 #define CPOLICY_WRITETHROUGH	2
61 #define CPOLICY_WRITEBACK	3
62 #define CPOLICY_WRITEALLOC	4
63 
64 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK;
65 static unsigned int ecc_mask __initdata = 0;
66 pgprot_t pgprot_user;
67 pgprot_t pgprot_kernel;
68 pgprot_t pgprot_hyp_device;
69 pgprot_t pgprot_s2;
70 pgprot_t pgprot_s2_device;
71 
72 EXPORT_SYMBOL(pgprot_user);
73 EXPORT_SYMBOL(pgprot_kernel);
74 
75 struct cachepolicy {
76 	const char	policy[16];
77 	unsigned int	cr_mask;
78 	pmdval_t	pmd;
79 	pteval_t	pte;
80 	pteval_t	pte_s2;
81 };
82 
83 #ifdef CONFIG_ARM_LPAE
84 #define s2_policy(policy)	policy
85 #else
86 #define s2_policy(policy)	0
87 #endif
88 
89 static struct cachepolicy cache_policies[] __initdata = {
90 	{
91 		.policy		= "uncached",
92 		.cr_mask	= CR_W|CR_C,
93 		.pmd		= PMD_SECT_UNCACHED,
94 		.pte		= L_PTE_MT_UNCACHED,
95 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
96 	}, {
97 		.policy		= "buffered",
98 		.cr_mask	= CR_C,
99 		.pmd		= PMD_SECT_BUFFERED,
100 		.pte		= L_PTE_MT_BUFFERABLE,
101 		.pte_s2		= s2_policy(L_PTE_S2_MT_UNCACHED),
102 	}, {
103 		.policy		= "writethrough",
104 		.cr_mask	= 0,
105 		.pmd		= PMD_SECT_WT,
106 		.pte		= L_PTE_MT_WRITETHROUGH,
107 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITETHROUGH),
108 	}, {
109 		.policy		= "writeback",
110 		.cr_mask	= 0,
111 		.pmd		= PMD_SECT_WB,
112 		.pte		= L_PTE_MT_WRITEBACK,
113 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
114 	}, {
115 		.policy		= "writealloc",
116 		.cr_mask	= 0,
117 		.pmd		= PMD_SECT_WBWA,
118 		.pte		= L_PTE_MT_WRITEALLOC,
119 		.pte_s2		= s2_policy(L_PTE_S2_MT_WRITEBACK),
120 	}
121 };
122 
123 #ifdef CONFIG_CPU_CP15
124 static unsigned long initial_pmd_value __initdata = 0;
125 
126 /*
127  * Initialise the cache_policy variable with the initial state specified
128  * via the "pmd" value.  This is used to ensure that on ARMv6 and later,
129  * the C code sets the page tables up with the same policy as the head
130  * assembly code, which avoids an illegal state where the TLBs can get
131  * confused.  See comments in early_cachepolicy() for more information.
132  */
init_default_cache_policy(unsigned long pmd)133 void __init init_default_cache_policy(unsigned long pmd)
134 {
135 	int i;
136 
137 	initial_pmd_value = pmd;
138 
139 	pmd &= PMD_SECT_TEX(1) | PMD_SECT_BUFFERABLE | PMD_SECT_CACHEABLE;
140 
141 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++)
142 		if (cache_policies[i].pmd == pmd) {
143 			cachepolicy = i;
144 			break;
145 		}
146 
147 	if (i == ARRAY_SIZE(cache_policies))
148 		pr_err("ERROR: could not find cache policy\n");
149 }
150 
151 /*
152  * These are useful for identifying cache coherency problems by allowing
153  * the cache or the cache and writebuffer to be turned off.  (Note: the
154  * write buffer should not be on and the cache off).
155  */
early_cachepolicy(char * p)156 static int __init early_cachepolicy(char *p)
157 {
158 	int i, selected = -1;
159 
160 	for (i = 0; i < ARRAY_SIZE(cache_policies); i++) {
161 		int len = strlen(cache_policies[i].policy);
162 
163 		if (memcmp(p, cache_policies[i].policy, len) == 0) {
164 			selected = i;
165 			break;
166 		}
167 	}
168 
169 	if (selected == -1)
170 		pr_err("ERROR: unknown or unsupported cache policy\n");
171 
172 	/*
173 	 * This restriction is partly to do with the way we boot; it is
174 	 * unpredictable to have memory mapped using two different sets of
175 	 * memory attributes (shared, type, and cache attribs).  We can not
176 	 * change these attributes once the initial assembly has setup the
177 	 * page tables.
178 	 */
179 	if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) {
180 		pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n",
181 			cache_policies[cachepolicy].policy);
182 		return 0;
183 	}
184 
185 	if (selected != cachepolicy) {
186 		unsigned long cr = __clear_cr(cache_policies[selected].cr_mask);
187 		cachepolicy = selected;
188 		flush_cache_all();
189 		set_cr(cr);
190 	}
191 	return 0;
192 }
193 early_param("cachepolicy", early_cachepolicy);
194 
early_nocache(char * __unused)195 static int __init early_nocache(char *__unused)
196 {
197 	char *p = "buffered";
198 	printk(KERN_WARNING "nocache is deprecated; use cachepolicy=%s\n", p);
199 	early_cachepolicy(p);
200 	return 0;
201 }
202 early_param("nocache", early_nocache);
203 
early_nowrite(char * __unused)204 static int __init early_nowrite(char *__unused)
205 {
206 	char *p = "uncached";
207 	printk(KERN_WARNING "nowb is deprecated; use cachepolicy=%s\n", p);
208 	early_cachepolicy(p);
209 	return 0;
210 }
211 early_param("nowb", early_nowrite);
212 
213 #ifndef CONFIG_ARM_LPAE
early_ecc(char * p)214 static int __init early_ecc(char *p)
215 {
216 	if (memcmp(p, "on", 2) == 0)
217 		ecc_mask = PMD_PROTECTION;
218 	else if (memcmp(p, "off", 3) == 0)
219 		ecc_mask = 0;
220 	return 0;
221 }
222 early_param("ecc", early_ecc);
223 #endif
224 
225 #else /* ifdef CONFIG_CPU_CP15 */
226 
early_cachepolicy(char * p)227 static int __init early_cachepolicy(char *p)
228 {
229 	pr_warn("cachepolicy kernel parameter not supported without cp15\n");
230 }
231 early_param("cachepolicy", early_cachepolicy);
232 
noalign_setup(char * __unused)233 static int __init noalign_setup(char *__unused)
234 {
235 	pr_warn("noalign kernel parameter not supported without cp15\n");
236 }
237 __setup("noalign", noalign_setup);
238 
239 #endif /* ifdef CONFIG_CPU_CP15 / else */
240 
241 #define PROT_PTE_DEVICE		L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN
242 #define PROT_PTE_S2_DEVICE	PROT_PTE_DEVICE
243 #define PROT_SECT_DEVICE	PMD_TYPE_SECT|PMD_SECT_AP_WRITE
244 
245 static struct mem_type mem_types[] = {
246 	[MT_DEVICE] = {		  /* Strongly ordered / ARMv6 shared device */
247 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED |
248 				  L_PTE_SHARED,
249 		.prot_pte_s2	= s2_policy(PROT_PTE_S2_DEVICE) |
250 				  s2_policy(L_PTE_S2_MT_DEV_SHARED) |
251 				  L_PTE_SHARED,
252 		.prot_l1	= PMD_TYPE_TABLE,
253 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_S,
254 		.domain		= DOMAIN_IO,
255 	},
256 	[MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */
257 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED,
258 		.prot_l1	= PMD_TYPE_TABLE,
259 		.prot_sect	= PROT_SECT_DEVICE,
260 		.domain		= DOMAIN_IO,
261 	},
262 	[MT_DEVICE_CACHED] = {	  /* ioremap_cached */
263 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED,
264 		.prot_l1	= PMD_TYPE_TABLE,
265 		.prot_sect	= PROT_SECT_DEVICE | PMD_SECT_WB,
266 		.domain		= DOMAIN_IO,
267 	},
268 	[MT_DEVICE_WC] = {	/* ioremap_wc */
269 		.prot_pte	= PROT_PTE_DEVICE | L_PTE_MT_DEV_WC,
270 		.prot_l1	= PMD_TYPE_TABLE,
271 		.prot_sect	= PROT_SECT_DEVICE,
272 		.domain		= DOMAIN_IO,
273 	},
274 	[MT_UNCACHED] = {
275 		.prot_pte	= PROT_PTE_DEVICE,
276 		.prot_l1	= PMD_TYPE_TABLE,
277 		.prot_sect	= PMD_TYPE_SECT | PMD_SECT_XN,
278 		.domain		= DOMAIN_IO,
279 	},
280 	[MT_CACHECLEAN] = {
281 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
282 		.domain    = DOMAIN_KERNEL,
283 	},
284 #ifndef CONFIG_ARM_LPAE
285 	[MT_MINICLEAN] = {
286 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE,
287 		.domain    = DOMAIN_KERNEL,
288 	},
289 #endif
290 	[MT_LOW_VECTORS] = {
291 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
292 				L_PTE_RDONLY,
293 		.prot_l1   = PMD_TYPE_TABLE,
294 		.domain    = DOMAIN_VECTORS,
295 	},
296 	[MT_HIGH_VECTORS] = {
297 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
298 				L_PTE_USER | L_PTE_RDONLY,
299 		.prot_l1   = PMD_TYPE_TABLE,
300 		.domain    = DOMAIN_VECTORS,
301 	},
302 	[MT_MEMORY_RWX] = {
303 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
304 		.prot_l1   = PMD_TYPE_TABLE,
305 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
306 		.domain    = DOMAIN_KERNEL,
307 	},
308 	[MT_MEMORY_RW] = {
309 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
310 			     L_PTE_XN,
311 		.prot_l1   = PMD_TYPE_TABLE,
312 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
313 		.domain    = DOMAIN_KERNEL,
314 	},
315 	[MT_ROM] = {
316 		.prot_sect = PMD_TYPE_SECT,
317 		.domain    = DOMAIN_KERNEL,
318 	},
319 	[MT_MEMORY_RWX_NONCACHED] = {
320 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
321 				L_PTE_MT_BUFFERABLE,
322 		.prot_l1   = PMD_TYPE_TABLE,
323 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE,
324 		.domain    = DOMAIN_KERNEL,
325 	},
326 	[MT_MEMORY_RW_DTCM] = {
327 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
328 				L_PTE_XN,
329 		.prot_l1   = PMD_TYPE_TABLE,
330 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_XN,
331 		.domain    = DOMAIN_KERNEL,
332 	},
333 	[MT_MEMORY_RWX_ITCM] = {
334 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY,
335 		.prot_l1   = PMD_TYPE_TABLE,
336 		.domain    = DOMAIN_KERNEL,
337 	},
338 	[MT_MEMORY_RW_SO] = {
339 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
340 				L_PTE_MT_UNCACHED | L_PTE_XN,
341 		.prot_l1   = PMD_TYPE_TABLE,
342 		.prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S |
343 				PMD_SECT_UNCACHED | PMD_SECT_XN,
344 		.domain    = DOMAIN_KERNEL,
345 	},
346 	[MT_MEMORY_DMA_READY] = {
347 		.prot_pte  = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY |
348 				L_PTE_XN,
349 		.prot_l1   = PMD_TYPE_TABLE,
350 		.domain    = DOMAIN_KERNEL,
351 	},
352 };
353 
get_mem_type(unsigned int type)354 const struct mem_type *get_mem_type(unsigned int type)
355 {
356 	return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL;
357 }
358 EXPORT_SYMBOL(get_mem_type);
359 
360 #define PTE_SET_FN(_name, pteop) \
361 static int pte_set_##_name(pte_t *ptep, pgtable_t token, unsigned long addr, \
362 			void *data) \
363 { \
364 	pte_t pte = pteop(*ptep); \
365 \
366 	set_pte_ext(ptep, pte, 0); \
367 	return 0; \
368 } \
369 
370 #define SET_MEMORY_FN(_name, callback) \
371 int set_memory_##_name(unsigned long addr, int numpages) \
372 { \
373 	unsigned long start = addr; \
374 	unsigned long size = PAGE_SIZE*numpages; \
375 	unsigned end = start + size; \
376 \
377 	if (start < MODULES_VADDR || start >= MODULES_END) \
378 		return -EINVAL;\
379 \
380 	if (end < MODULES_VADDR || end >= MODULES_END) \
381 		return -EINVAL; \
382 \
383 	apply_to_page_range(&init_mm, start, size, callback, NULL); \
384 	flush_tlb_kernel_range(start, end); \
385 	return 0;\
386 }
387 
PTE_SET_FN(ro,pte_wrprotect)388 PTE_SET_FN(ro, pte_wrprotect)
389 PTE_SET_FN(rw, pte_mkwrite)
390 PTE_SET_FN(x, pte_mkexec)
391 PTE_SET_FN(nx, pte_mknexec)
392 
393 SET_MEMORY_FN(ro, pte_set_ro)
394 SET_MEMORY_FN(rw, pte_set_rw)
395 SET_MEMORY_FN(x, pte_set_x)
396 SET_MEMORY_FN(nx, pte_set_nx)
397 
398 /*
399  * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range().
400  * As a result, this can only be called with preemption disabled, as under
401  * stop_machine().
402  */
403 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
404 {
405 	unsigned long vaddr = __fix_to_virt(idx);
406 	pte_t *pte = pte_offset_kernel(pmd_off_k(vaddr), vaddr);
407 
408 	/* Make sure fixmap region does not exceed available allocation. */
409 	BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) >
410 		     FIXADDR_END);
411 	BUG_ON(idx >= __end_of_fixed_addresses);
412 
413 	if (pgprot_val(prot))
414 		set_pte_at(NULL, vaddr, pte,
415 			pfn_pte(phys >> PAGE_SHIFT, prot));
416 	else
417 		pte_clear(NULL, vaddr, pte);
418 	local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE);
419 }
420 
421 /*
422  * Adjust the PMD section entries according to the CPU in use.
423  */
build_mem_type_table(void)424 static void __init build_mem_type_table(void)
425 {
426 	struct cachepolicy *cp;
427 	unsigned int cr = get_cr();
428 	pteval_t user_pgprot, kern_pgprot, vecs_pgprot;
429 	pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot;
430 	int cpu_arch = cpu_architecture();
431 	int i;
432 
433 	if (cpu_arch < CPU_ARCH_ARMv6) {
434 #if defined(CONFIG_CPU_DCACHE_DISABLE)
435 		if (cachepolicy > CPOLICY_BUFFERED)
436 			cachepolicy = CPOLICY_BUFFERED;
437 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH)
438 		if (cachepolicy > CPOLICY_WRITETHROUGH)
439 			cachepolicy = CPOLICY_WRITETHROUGH;
440 #endif
441 	}
442 	if (cpu_arch < CPU_ARCH_ARMv5) {
443 		if (cachepolicy >= CPOLICY_WRITEALLOC)
444 			cachepolicy = CPOLICY_WRITEBACK;
445 		ecc_mask = 0;
446 	}
447 
448 	if (is_smp()) {
449 		if (cachepolicy != CPOLICY_WRITEALLOC) {
450 			pr_warn("Forcing write-allocate cache policy for SMP\n");
451 			cachepolicy = CPOLICY_WRITEALLOC;
452 		}
453 		if (!(initial_pmd_value & PMD_SECT_S)) {
454 			pr_warn("Forcing shared mappings for SMP\n");
455 			initial_pmd_value |= PMD_SECT_S;
456 		}
457 	}
458 
459 	/*
460 	 * Strip out features not present on earlier architectures.
461 	 * Pre-ARMv5 CPUs don't have TEX bits.  Pre-ARMv6 CPUs or those
462 	 * without extended page tables don't have the 'Shared' bit.
463 	 */
464 	if (cpu_arch < CPU_ARCH_ARMv5)
465 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
466 			mem_types[i].prot_sect &= ~PMD_SECT_TEX(7);
467 	if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3())
468 		for (i = 0; i < ARRAY_SIZE(mem_types); i++)
469 			mem_types[i].prot_sect &= ~PMD_SECT_S;
470 
471 	/*
472 	 * ARMv5 and lower, bit 4 must be set for page tables (was: cache
473 	 * "update-able on write" bit on ARM610).  However, Xscale and
474 	 * Xscale3 require this bit to be cleared.
475 	 */
476 	if (cpu_is_xscale() || cpu_is_xsc3()) {
477 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
478 			mem_types[i].prot_sect &= ~PMD_BIT4;
479 			mem_types[i].prot_l1 &= ~PMD_BIT4;
480 		}
481 	} else if (cpu_arch < CPU_ARCH_ARMv6) {
482 		for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
483 			if (mem_types[i].prot_l1)
484 				mem_types[i].prot_l1 |= PMD_BIT4;
485 			if (mem_types[i].prot_sect)
486 				mem_types[i].prot_sect |= PMD_BIT4;
487 		}
488 	}
489 
490 	/*
491 	 * Mark the device areas according to the CPU/architecture.
492 	 */
493 	if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) {
494 		if (!cpu_is_xsc3()) {
495 			/*
496 			 * Mark device regions on ARMv6+ as execute-never
497 			 * to prevent speculative instruction fetches.
498 			 */
499 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN;
500 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN;
501 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN;
502 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN;
503 
504 			/* Also setup NX memory mapping */
505 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN;
506 		}
507 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
508 			/*
509 			 * For ARMv7 with TEX remapping,
510 			 * - shared device is SXCB=1100
511 			 * - nonshared device is SXCB=0100
512 			 * - write combine device mem is SXCB=0001
513 			 * (Uncached Normal memory)
514 			 */
515 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1);
516 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1);
517 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
518 		} else if (cpu_is_xsc3()) {
519 			/*
520 			 * For Xscale3,
521 			 * - shared device is TEXCB=00101
522 			 * - nonshared device is TEXCB=01000
523 			 * - write combine device mem is TEXCB=00100
524 			 * (Inner/Outer Uncacheable in xsc3 parlance)
525 			 */
526 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED;
527 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
528 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
529 		} else {
530 			/*
531 			 * For ARMv6 and ARMv7 without TEX remapping,
532 			 * - shared device is TEXCB=00001
533 			 * - nonshared device is TEXCB=01000
534 			 * - write combine device mem is TEXCB=00100
535 			 * (Uncached Normal in ARMv6 parlance).
536 			 */
537 			mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED;
538 			mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2);
539 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1);
540 		}
541 	} else {
542 		/*
543 		 * On others, write combining is "Uncached/Buffered"
544 		 */
545 		mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE;
546 	}
547 
548 	/*
549 	 * Now deal with the memory-type mappings
550 	 */
551 	cp = &cache_policies[cachepolicy];
552 	vecs_pgprot = kern_pgprot = user_pgprot = cp->pte;
553 	s2_pgprot = cp->pte_s2;
554 	hyp_device_pgprot = mem_types[MT_DEVICE].prot_pte;
555 	s2_device_pgprot = mem_types[MT_DEVICE].prot_pte_s2;
556 
557 	/*
558 	 * We don't use domains on ARMv6 (since this causes problems with
559 	 * v6/v7 kernels), so we must use a separate memory type for user
560 	 * r/o, kernel r/w to map the vectors page.
561 	 */
562 #ifndef CONFIG_ARM_LPAE
563 	if (cpu_arch == CPU_ARCH_ARMv6)
564 		vecs_pgprot |= L_PTE_MT_VECTORS;
565 #endif
566 
567 #ifndef CONFIG_ARM_LPAE
568 	/*
569 	 * We don't use domains on ARMv6 (since this causes problems with
570 	 * v6/v7 kernels), so we must use a separate memory type for user
571 	 * r/o, kernel r/w to map the vectors page.
572 	 */
573 	if (cpu_arch == CPU_ARCH_ARMv6)
574 		vecs_pgprot |= L_PTE_MT_VECTORS;
575 
576 	/*
577 	 * Check is it with support for the PXN bit
578 	 * in the Short-descriptor translation table format descriptors.
579 	 */
580 	if (cpu_arch == CPU_ARCH_ARMv7 &&
581 		(read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) >= 4) {
582 		user_pmd_table |= PMD_PXNTABLE;
583 	}
584 #endif
585 
586 	/*
587 	 * ARMv6 and above have extended page tables.
588 	 */
589 	if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) {
590 #ifndef CONFIG_ARM_LPAE
591 		/*
592 		 * Mark cache clean areas and XIP ROM read only
593 		 * from SVC mode and no access from userspace.
594 		 */
595 		mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
596 		mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
597 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE;
598 #endif
599 
600 		/*
601 		 * If the initial page tables were created with the S bit
602 		 * set, then we need to do the same here for the same
603 		 * reasons given in early_cachepolicy().
604 		 */
605 		if (initial_pmd_value & PMD_SECT_S) {
606 			user_pgprot |= L_PTE_SHARED;
607 			kern_pgprot |= L_PTE_SHARED;
608 			vecs_pgprot |= L_PTE_SHARED;
609 			s2_pgprot |= L_PTE_SHARED;
610 			mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S;
611 			mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED;
612 			mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S;
613 			mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED;
614 			mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S;
615 			mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED;
616 			mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S;
617 			mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED;
618 			mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED;
619 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S;
620 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED;
621 		}
622 	}
623 
624 	/*
625 	 * Non-cacheable Normal - intended for memory areas that must
626 	 * not cause dirty cache line writebacks when used
627 	 */
628 	if (cpu_arch >= CPU_ARCH_ARMv6) {
629 		if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) {
630 			/* Non-cacheable Normal is XCB = 001 */
631 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
632 				PMD_SECT_BUFFERED;
633 		} else {
634 			/* For both ARMv6 and non-TEX-remapping ARMv7 */
635 			mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |=
636 				PMD_SECT_TEX(1);
637 		}
638 	} else {
639 		mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE;
640 	}
641 
642 #ifdef CONFIG_ARM_LPAE
643 	/*
644 	 * Do not generate access flag faults for the kernel mappings.
645 	 */
646 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
647 		mem_types[i].prot_pte |= PTE_EXT_AF;
648 		if (mem_types[i].prot_sect)
649 			mem_types[i].prot_sect |= PMD_SECT_AF;
650 	}
651 	kern_pgprot |= PTE_EXT_AF;
652 	vecs_pgprot |= PTE_EXT_AF;
653 
654 	/*
655 	 * Set PXN for user mappings
656 	 */
657 	user_pgprot |= PTE_EXT_PXN;
658 #endif
659 
660 	for (i = 0; i < 16; i++) {
661 		pteval_t v = pgprot_val(protection_map[i]);
662 		protection_map[i] = __pgprot(v | user_pgprot);
663 	}
664 
665 	mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot;
666 	mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot;
667 
668 	pgprot_user   = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot);
669 	pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG |
670 				 L_PTE_DIRTY | kern_pgprot);
671 	pgprot_s2  = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot);
672 	pgprot_s2_device  = __pgprot(s2_device_pgprot);
673 	pgprot_hyp_device  = __pgprot(hyp_device_pgprot);
674 
675 	mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask;
676 	mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask;
677 	mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd;
678 	mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot;
679 	mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd;
680 	mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot;
681 	mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot;
682 	mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask;
683 	mem_types[MT_ROM].prot_sect |= cp->pmd;
684 
685 	switch (cp->pmd) {
686 	case PMD_SECT_WT:
687 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT;
688 		break;
689 	case PMD_SECT_WB:
690 	case PMD_SECT_WBWA:
691 		mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB;
692 		break;
693 	}
694 	pr_info("Memory policy: %sData cache %s\n",
695 		ecc_mask ? "ECC enabled, " : "", cp->policy);
696 
697 	for (i = 0; i < ARRAY_SIZE(mem_types); i++) {
698 		struct mem_type *t = &mem_types[i];
699 		if (t->prot_l1)
700 			t->prot_l1 |= PMD_DOMAIN(t->domain);
701 		if (t->prot_sect)
702 			t->prot_sect |= PMD_DOMAIN(t->domain);
703 	}
704 }
705 
706 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
phys_mem_access_prot(struct file * file,unsigned long pfn,unsigned long size,pgprot_t vma_prot)707 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
708 			      unsigned long size, pgprot_t vma_prot)
709 {
710 	if (!pfn_valid(pfn))
711 		return pgprot_noncached(vma_prot);
712 	else if (file->f_flags & O_SYNC)
713 		return pgprot_writecombine(vma_prot);
714 	return vma_prot;
715 }
716 EXPORT_SYMBOL(phys_mem_access_prot);
717 #endif
718 
719 #define vectors_base()	(vectors_high() ? 0xffff0000 : 0)
720 
early_alloc_aligned(unsigned long sz,unsigned long align)721 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align)
722 {
723 	void *ptr = __va(memblock_alloc(sz, align));
724 	memset(ptr, 0, sz);
725 	return ptr;
726 }
727 
early_alloc(unsigned long sz)728 static void __init *early_alloc(unsigned long sz)
729 {
730 	return early_alloc_aligned(sz, sz);
731 }
732 
early_pte_alloc(pmd_t * pmd,unsigned long addr,unsigned long prot)733 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot)
734 {
735 	if (pmd_none(*pmd)) {
736 		pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE);
737 		__pmd_populate(pmd, __pa(pte), prot);
738 	}
739 	BUG_ON(pmd_bad(*pmd));
740 	return pte_offset_kernel(pmd, addr);
741 }
742 
alloc_init_pte(pmd_t * pmd,unsigned long addr,unsigned long end,unsigned long pfn,const struct mem_type * type)743 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr,
744 				  unsigned long end, unsigned long pfn,
745 				  const struct mem_type *type)
746 {
747 	pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1);
748 	do {
749 		set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0);
750 		pfn++;
751 	} while (pte++, addr += PAGE_SIZE, addr != end);
752 }
753 
__map_init_section(pmd_t * pmd,unsigned long addr,unsigned long end,phys_addr_t phys,const struct mem_type * type)754 static void __init __map_init_section(pmd_t *pmd, unsigned long addr,
755 			unsigned long end, phys_addr_t phys,
756 			const struct mem_type *type)
757 {
758 	pmd_t *p = pmd;
759 
760 #ifndef CONFIG_ARM_LPAE
761 	/*
762 	 * In classic MMU format, puds and pmds are folded in to
763 	 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a
764 	 * group of L1 entries making up one logical pointer to
765 	 * an L2 table (2MB), where as PMDs refer to the individual
766 	 * L1 entries (1MB). Hence increment to get the correct
767 	 * offset for odd 1MB sections.
768 	 * (See arch/arm/include/asm/pgtable-2level.h)
769 	 */
770 	if (addr & SECTION_SIZE)
771 		pmd++;
772 #endif
773 	do {
774 		*pmd = __pmd(phys | type->prot_sect);
775 		phys += SECTION_SIZE;
776 	} while (pmd++, addr += SECTION_SIZE, addr != end);
777 
778 	flush_pmd_entry(p);
779 }
780 
alloc_init_pmd(pud_t * pud,unsigned long addr,unsigned long end,phys_addr_t phys,const struct mem_type * type)781 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr,
782 				      unsigned long end, phys_addr_t phys,
783 				      const struct mem_type *type)
784 {
785 	pmd_t *pmd = pmd_offset(pud, addr);
786 	unsigned long next;
787 
788 	do {
789 		/*
790 		 * With LPAE, we must loop over to map
791 		 * all the pmds for the given range.
792 		 */
793 		next = pmd_addr_end(addr, end);
794 
795 		/*
796 		 * Try a section mapping - addr, next and phys must all be
797 		 * aligned to a section boundary.
798 		 */
799 		if (type->prot_sect &&
800 				((addr | next | phys) & ~SECTION_MASK) == 0) {
801 			__map_init_section(pmd, addr, next, phys, type);
802 		} else {
803 			alloc_init_pte(pmd, addr, next,
804 						__phys_to_pfn(phys), type);
805 		}
806 
807 		phys += next - addr;
808 
809 	} while (pmd++, addr = next, addr != end);
810 }
811 
alloc_init_pud(pgd_t * pgd,unsigned long addr,unsigned long end,phys_addr_t phys,const struct mem_type * type)812 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr,
813 				  unsigned long end, phys_addr_t phys,
814 				  const struct mem_type *type)
815 {
816 	pud_t *pud = pud_offset(pgd, addr);
817 	unsigned long next;
818 
819 	do {
820 		next = pud_addr_end(addr, end);
821 		alloc_init_pmd(pud, addr, next, phys, type);
822 		phys += next - addr;
823 	} while (pud++, addr = next, addr != end);
824 }
825 
826 #ifndef CONFIG_ARM_LPAE
create_36bit_mapping(struct map_desc * md,const struct mem_type * type)827 static void __init create_36bit_mapping(struct map_desc *md,
828 					const struct mem_type *type)
829 {
830 	unsigned long addr, length, end;
831 	phys_addr_t phys;
832 	pgd_t *pgd;
833 
834 	addr = md->virtual;
835 	phys = __pfn_to_phys(md->pfn);
836 	length = PAGE_ALIGN(md->length);
837 
838 	if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) {
839 		printk(KERN_ERR "MM: CPU does not support supersection "
840 		       "mapping for 0x%08llx at 0x%08lx\n",
841 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
842 		return;
843 	}
844 
845 	/* N.B.	ARMv6 supersections are only defined to work with domain 0.
846 	 *	Since domain assignments can in fact be arbitrary, the
847 	 *	'domain == 0' check below is required to insure that ARMv6
848 	 *	supersections are only allocated for domain 0 regardless
849 	 *	of the actual domain assignments in use.
850 	 */
851 	if (type->domain) {
852 		printk(KERN_ERR "MM: invalid domain in supersection "
853 		       "mapping for 0x%08llx at 0x%08lx\n",
854 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
855 		return;
856 	}
857 
858 	if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) {
859 		printk(KERN_ERR "MM: cannot create mapping for 0x%08llx"
860 		       " at 0x%08lx invalid alignment\n",
861 		       (long long)__pfn_to_phys((u64)md->pfn), addr);
862 		return;
863 	}
864 
865 	/*
866 	 * Shift bits [35:32] of address into bits [23:20] of PMD
867 	 * (See ARMv6 spec).
868 	 */
869 	phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20);
870 
871 	pgd = pgd_offset_k(addr);
872 	end = addr + length;
873 	do {
874 		pud_t *pud = pud_offset(pgd, addr);
875 		pmd_t *pmd = pmd_offset(pud, addr);
876 		int i;
877 
878 		for (i = 0; i < 16; i++)
879 			*pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER);
880 
881 		addr += SUPERSECTION_SIZE;
882 		phys += SUPERSECTION_SIZE;
883 		pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT;
884 	} while (addr != end);
885 }
886 #endif	/* !CONFIG_ARM_LPAE */
887 
888 /*
889  * Create the page directory entries and any necessary
890  * page tables for the mapping specified by `md'.  We
891  * are able to cope here with varying sizes and address
892  * offsets, and we take full advantage of sections and
893  * supersections.
894  */
create_mapping(struct map_desc * md)895 static void __init create_mapping(struct map_desc *md)
896 {
897 	unsigned long addr, length, end;
898 	phys_addr_t phys;
899 	const struct mem_type *type;
900 	pgd_t *pgd;
901 
902 	if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) {
903 		printk(KERN_WARNING "BUG: not creating mapping for 0x%08llx"
904 		       " at 0x%08lx in user region\n",
905 		       (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
906 		return;
907 	}
908 
909 	if ((md->type == MT_DEVICE || md->type == MT_ROM) &&
910 	    md->virtual >= PAGE_OFFSET &&
911 	    (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) {
912 		printk(KERN_WARNING "BUG: mapping for 0x%08llx"
913 		       " at 0x%08lx out of vmalloc space\n",
914 		       (long long)__pfn_to_phys((u64)md->pfn), md->virtual);
915 	}
916 
917 	type = &mem_types[md->type];
918 
919 #ifndef CONFIG_ARM_LPAE
920 	/*
921 	 * Catch 36-bit addresses
922 	 */
923 	if (md->pfn >= 0x100000) {
924 		create_36bit_mapping(md, type);
925 		return;
926 	}
927 #endif
928 
929 	addr = md->virtual & PAGE_MASK;
930 	phys = __pfn_to_phys(md->pfn);
931 	length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
932 
933 	if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) {
934 		printk(KERN_WARNING "BUG: map for 0x%08llx at 0x%08lx can not "
935 		       "be mapped using pages, ignoring.\n",
936 		       (long long)__pfn_to_phys(md->pfn), addr);
937 		return;
938 	}
939 
940 	pgd = pgd_offset_k(addr);
941 	end = addr + length;
942 	do {
943 		unsigned long next = pgd_addr_end(addr, end);
944 
945 		alloc_init_pud(pgd, addr, next, phys, type);
946 
947 		phys += next - addr;
948 		addr = next;
949 	} while (pgd++, addr != end);
950 }
951 
952 /*
953  * Create the architecture specific mappings
954  */
iotable_init(struct map_desc * io_desc,int nr)955 void __init iotable_init(struct map_desc *io_desc, int nr)
956 {
957 	struct map_desc *md;
958 	struct vm_struct *vm;
959 	struct static_vm *svm;
960 
961 	if (!nr)
962 		return;
963 
964 	svm = early_alloc_aligned(sizeof(*svm) * nr, __alignof__(*svm));
965 
966 	for (md = io_desc; nr; md++, nr--) {
967 		create_mapping(md);
968 
969 		vm = &svm->vm;
970 		vm->addr = (void *)(md->virtual & PAGE_MASK);
971 		vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK));
972 		vm->phys_addr = __pfn_to_phys(md->pfn);
973 		vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING;
974 		vm->flags |= VM_ARM_MTYPE(md->type);
975 		vm->caller = iotable_init;
976 		add_static_vm_early(svm++);
977 	}
978 }
979 
vm_reserve_area_early(unsigned long addr,unsigned long size,void * caller)980 void __init vm_reserve_area_early(unsigned long addr, unsigned long size,
981 				  void *caller)
982 {
983 	struct vm_struct *vm;
984 	struct static_vm *svm;
985 
986 	svm = early_alloc_aligned(sizeof(*svm), __alignof__(*svm));
987 
988 	vm = &svm->vm;
989 	vm->addr = (void *)addr;
990 	vm->size = size;
991 	vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING;
992 	vm->caller = caller;
993 	add_static_vm_early(svm);
994 }
995 
996 #ifndef CONFIG_ARM_LPAE
997 
998 /*
999  * The Linux PMD is made of two consecutive section entries covering 2MB
1000  * (see definition in include/asm/pgtable-2level.h).  However a call to
1001  * create_mapping() may optimize static mappings by using individual
1002  * 1MB section mappings.  This leaves the actual PMD potentially half
1003  * initialized if the top or bottom section entry isn't used, leaving it
1004  * open to problems if a subsequent ioremap() or vmalloc() tries to use
1005  * the virtual space left free by that unused section entry.
1006  *
1007  * Let's avoid the issue by inserting dummy vm entries covering the unused
1008  * PMD halves once the static mappings are in place.
1009  */
1010 
pmd_empty_section_gap(unsigned long addr)1011 static void __init pmd_empty_section_gap(unsigned long addr)
1012 {
1013 	vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap);
1014 }
1015 
fill_pmd_gaps(void)1016 static void __init fill_pmd_gaps(void)
1017 {
1018 	struct static_vm *svm;
1019 	struct vm_struct *vm;
1020 	unsigned long addr, next = 0;
1021 	pmd_t *pmd;
1022 
1023 	list_for_each_entry(svm, &static_vmlist, list) {
1024 		vm = &svm->vm;
1025 		addr = (unsigned long)vm->addr;
1026 		if (addr < next)
1027 			continue;
1028 
1029 		/*
1030 		 * Check if this vm starts on an odd section boundary.
1031 		 * If so and the first section entry for this PMD is free
1032 		 * then we block the corresponding virtual address.
1033 		 */
1034 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1035 			pmd = pmd_off_k(addr);
1036 			if (pmd_none(*pmd))
1037 				pmd_empty_section_gap(addr & PMD_MASK);
1038 		}
1039 
1040 		/*
1041 		 * Then check if this vm ends on an odd section boundary.
1042 		 * If so and the second section entry for this PMD is empty
1043 		 * then we block the corresponding virtual address.
1044 		 */
1045 		addr += vm->size;
1046 		if ((addr & ~PMD_MASK) == SECTION_SIZE) {
1047 			pmd = pmd_off_k(addr) + 1;
1048 			if (pmd_none(*pmd))
1049 				pmd_empty_section_gap(addr);
1050 		}
1051 
1052 		/* no need to look at any vm entry until we hit the next PMD */
1053 		next = (addr + PMD_SIZE - 1) & PMD_MASK;
1054 	}
1055 }
1056 
1057 #else
1058 #define fill_pmd_gaps() do { } while (0)
1059 #endif
1060 
1061 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H)
pci_reserve_io(void)1062 static void __init pci_reserve_io(void)
1063 {
1064 	struct static_vm *svm;
1065 
1066 	svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE);
1067 	if (svm)
1068 		return;
1069 
1070 	vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io);
1071 }
1072 #else
1073 #define pci_reserve_io() do { } while (0)
1074 #endif
1075 
1076 #ifdef CONFIG_DEBUG_LL
debug_ll_io_init(void)1077 void __init debug_ll_io_init(void)
1078 {
1079 	struct map_desc map;
1080 
1081 	debug_ll_addr(&map.pfn, &map.virtual);
1082 	if (!map.pfn || !map.virtual)
1083 		return;
1084 	map.pfn = __phys_to_pfn(map.pfn);
1085 	map.virtual &= PAGE_MASK;
1086 	map.length = PAGE_SIZE;
1087 	map.type = MT_DEVICE;
1088 	iotable_init(&map, 1);
1089 }
1090 #endif
1091 
1092 static void * __initdata vmalloc_min =
1093 	(void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET);
1094 
1095 /*
1096  * vmalloc=size forces the vmalloc area to be exactly 'size'
1097  * bytes. This can be used to increase (or decrease) the vmalloc
1098  * area - the default is 240m.
1099  */
early_vmalloc(char * arg)1100 static int __init early_vmalloc(char *arg)
1101 {
1102 	unsigned long vmalloc_reserve = memparse(arg, NULL);
1103 
1104 	if (vmalloc_reserve < SZ_16M) {
1105 		vmalloc_reserve = SZ_16M;
1106 		printk(KERN_WARNING
1107 			"vmalloc area too small, limiting to %luMB\n",
1108 			vmalloc_reserve >> 20);
1109 	}
1110 
1111 	if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) {
1112 		vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M);
1113 		printk(KERN_WARNING
1114 			"vmalloc area is too big, limiting to %luMB\n",
1115 			vmalloc_reserve >> 20);
1116 	}
1117 
1118 	vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve);
1119 	return 0;
1120 }
1121 early_param("vmalloc", early_vmalloc);
1122 
1123 phys_addr_t arm_lowmem_limit __initdata = 0;
1124 
sanity_check_meminfo(void)1125 void __init sanity_check_meminfo(void)
1126 {
1127 	phys_addr_t memblock_limit = 0;
1128 	int highmem = 0;
1129 	phys_addr_t vmalloc_limit = __pa(vmalloc_min - 1) + 1;
1130 	struct memblock_region *reg;
1131 
1132 	for_each_memblock(memory, reg) {
1133 		phys_addr_t block_start = reg->base;
1134 		phys_addr_t block_end = reg->base + reg->size;
1135 		phys_addr_t size_limit = reg->size;
1136 
1137 		if (reg->base >= vmalloc_limit)
1138 			highmem = 1;
1139 		else
1140 			size_limit = vmalloc_limit - reg->base;
1141 
1142 
1143 		if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) {
1144 
1145 			if (highmem) {
1146 				pr_notice("Ignoring RAM at %pa-%pa (!CONFIG_HIGHMEM)\n",
1147 					&block_start, &block_end);
1148 				memblock_remove(reg->base, reg->size);
1149 				continue;
1150 			}
1151 
1152 			if (reg->size > size_limit) {
1153 				phys_addr_t overlap_size = reg->size - size_limit;
1154 
1155 				pr_notice("Truncating RAM at %pa-%pa to -%pa",
1156 				      &block_start, &block_end, &vmalloc_limit);
1157 				memblock_remove(vmalloc_limit, overlap_size);
1158 				block_end = vmalloc_limit;
1159 			}
1160 		}
1161 
1162 		if (!highmem) {
1163 			if (block_end > arm_lowmem_limit) {
1164 				if (reg->size > size_limit)
1165 					arm_lowmem_limit = vmalloc_limit;
1166 				else
1167 					arm_lowmem_limit = block_end;
1168 			}
1169 
1170 			/*
1171 			 * Find the first non-pmd-aligned page, and point
1172 			 * memblock_limit at it. This relies on rounding the
1173 			 * limit down to be pmd-aligned, which happens at the
1174 			 * end of this function.
1175 			 *
1176 			 * With this algorithm, the start or end of almost any
1177 			 * bank can be non-pmd-aligned. The only exception is
1178 			 * that the start of the bank 0 must be section-
1179 			 * aligned, since otherwise memory would need to be
1180 			 * allocated when mapping the start of bank 0, which
1181 			 * occurs before any free memory is mapped.
1182 			 */
1183 			if (!memblock_limit) {
1184 				if (!IS_ALIGNED(block_start, PMD_SIZE))
1185 					memblock_limit = block_start;
1186 				else if (!IS_ALIGNED(block_end, PMD_SIZE))
1187 					memblock_limit = arm_lowmem_limit;
1188 			}
1189 
1190 		}
1191 	}
1192 
1193 	high_memory = __va(arm_lowmem_limit - 1) + 1;
1194 
1195 	if (!memblock_limit)
1196 		memblock_limit = arm_lowmem_limit;
1197 
1198 	/*
1199 	 * Round the memblock limit down to a pmd size.  This
1200 	 * helps to ensure that we will allocate memory from the
1201 	 * last full pmd, which should be mapped.
1202 	 */
1203 	memblock_limit = round_down(memblock_limit, PMD_SIZE);
1204 
1205 	memblock_set_current_limit(memblock_limit);
1206 }
1207 
prepare_page_table(void)1208 static inline void prepare_page_table(void)
1209 {
1210 	unsigned long addr;
1211 	phys_addr_t end;
1212 
1213 	/*
1214 	 * Clear out all the mappings below the kernel image.
1215 	 */
1216 	for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE)
1217 		pmd_clear(pmd_off_k(addr));
1218 
1219 #ifdef CONFIG_XIP_KERNEL
1220 	/* The XIP kernel is mapped in the module area -- skip over it */
1221 	addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK;
1222 #endif
1223 	for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE)
1224 		pmd_clear(pmd_off_k(addr));
1225 
1226 	/*
1227 	 * Find the end of the first block of lowmem.
1228 	 */
1229 	end = memblock.memory.regions[0].base + memblock.memory.regions[0].size;
1230 	if (end >= arm_lowmem_limit)
1231 		end = arm_lowmem_limit;
1232 
1233 	/*
1234 	 * Clear out all the kernel space mappings, except for the first
1235 	 * memory bank, up to the vmalloc region.
1236 	 */
1237 	for (addr = __phys_to_virt(end);
1238 	     addr < VMALLOC_START; addr += PMD_SIZE)
1239 		pmd_clear(pmd_off_k(addr));
1240 }
1241 
1242 #ifdef CONFIG_ARM_LPAE
1243 /* the first page is reserved for pgd */
1244 #define SWAPPER_PG_DIR_SIZE	(PAGE_SIZE + \
1245 				 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t))
1246 #else
1247 #define SWAPPER_PG_DIR_SIZE	(PTRS_PER_PGD * sizeof(pgd_t))
1248 #endif
1249 
1250 /*
1251  * Reserve the special regions of memory
1252  */
arm_mm_memblock_reserve(void)1253 void __init arm_mm_memblock_reserve(void)
1254 {
1255 	/*
1256 	 * Reserve the page tables.  These are already in use,
1257 	 * and can only be in node 0.
1258 	 */
1259 	memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE);
1260 
1261 #ifdef CONFIG_SA1111
1262 	/*
1263 	 * Because of the SA1111 DMA bug, we want to preserve our
1264 	 * precious DMA-able memory...
1265 	 */
1266 	memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET);
1267 #endif
1268 }
1269 
1270 /*
1271  * Set up the device mappings.  Since we clear out the page tables for all
1272  * mappings above VMALLOC_START, we will remove any debug device mappings.
1273  * This means you have to be careful how you debug this function, or any
1274  * called function.  This means you can't use any function or debugging
1275  * method which may touch any device, otherwise the kernel _will_ crash.
1276  */
devicemaps_init(const struct machine_desc * mdesc)1277 static void __init devicemaps_init(const struct machine_desc *mdesc)
1278 {
1279 	struct map_desc map;
1280 	unsigned long addr;
1281 	void *vectors;
1282 
1283 	/*
1284 	 * Allocate the vector page early.
1285 	 */
1286 	vectors = early_alloc(PAGE_SIZE * 2);
1287 
1288 	early_trap_init(vectors);
1289 
1290 	for (addr = VMALLOC_START; addr; addr += PMD_SIZE)
1291 		pmd_clear(pmd_off_k(addr));
1292 
1293 	/*
1294 	 * Map the kernel if it is XIP.
1295 	 * It is always first in the modulearea.
1296 	 */
1297 #ifdef CONFIG_XIP_KERNEL
1298 	map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK);
1299 	map.virtual = MODULES_VADDR;
1300 	map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK;
1301 	map.type = MT_ROM;
1302 	create_mapping(&map);
1303 #endif
1304 
1305 	/*
1306 	 * Map the cache flushing regions.
1307 	 */
1308 #ifdef FLUSH_BASE
1309 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS);
1310 	map.virtual = FLUSH_BASE;
1311 	map.length = SZ_1M;
1312 	map.type = MT_CACHECLEAN;
1313 	create_mapping(&map);
1314 #endif
1315 #ifdef FLUSH_BASE_MINICACHE
1316 	map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M);
1317 	map.virtual = FLUSH_BASE_MINICACHE;
1318 	map.length = SZ_1M;
1319 	map.type = MT_MINICLEAN;
1320 	create_mapping(&map);
1321 #endif
1322 
1323 	/*
1324 	 * Create a mapping for the machine vectors at the high-vectors
1325 	 * location (0xffff0000).  If we aren't using high-vectors, also
1326 	 * create a mapping at the low-vectors virtual address.
1327 	 */
1328 	map.pfn = __phys_to_pfn(virt_to_phys(vectors));
1329 	map.virtual = 0xffff0000;
1330 	map.length = PAGE_SIZE;
1331 #ifdef CONFIG_KUSER_HELPERS
1332 	map.type = MT_HIGH_VECTORS;
1333 #else
1334 	map.type = MT_LOW_VECTORS;
1335 #endif
1336 	create_mapping(&map);
1337 
1338 	if (!vectors_high()) {
1339 		map.virtual = 0;
1340 		map.length = PAGE_SIZE * 2;
1341 		map.type = MT_LOW_VECTORS;
1342 		create_mapping(&map);
1343 	}
1344 
1345 	/* Now create a kernel read-only mapping */
1346 	map.pfn += 1;
1347 	map.virtual = 0xffff0000 + PAGE_SIZE;
1348 	map.length = PAGE_SIZE;
1349 	map.type = MT_LOW_VECTORS;
1350 	create_mapping(&map);
1351 
1352 	/*
1353 	 * Ask the machine support to map in the statically mapped devices.
1354 	 */
1355 	if (mdesc->map_io)
1356 		mdesc->map_io();
1357 	else
1358 		debug_ll_io_init();
1359 	fill_pmd_gaps();
1360 
1361 	/* Reserve fixed i/o space in VMALLOC region */
1362 	pci_reserve_io();
1363 
1364 	/*
1365 	 * Finally flush the caches and tlb to ensure that we're in a
1366 	 * consistent state wrt the writebuffer.  This also ensures that
1367 	 * any write-allocated cache lines in the vector page are written
1368 	 * back.  After this point, we can start to touch devices again.
1369 	 */
1370 	local_flush_tlb_all();
1371 	flush_cache_all();
1372 }
1373 
kmap_init(void)1374 static void __init kmap_init(void)
1375 {
1376 #ifdef CONFIG_HIGHMEM
1377 	pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE),
1378 		PKMAP_BASE, _PAGE_KERNEL_TABLE);
1379 #endif
1380 
1381 	early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START,
1382 			_PAGE_KERNEL_TABLE);
1383 }
1384 
map_lowmem(void)1385 static void __init map_lowmem(void)
1386 {
1387 	struct memblock_region *reg;
1388 	unsigned long kernel_x_start = round_down(__pa(_stext), SECTION_SIZE);
1389 	unsigned long kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE);
1390 
1391 	/* Map all the lowmem memory banks. */
1392 	for_each_memblock(memory, reg) {
1393 		phys_addr_t start = reg->base;
1394 		phys_addr_t end = start + reg->size;
1395 		struct map_desc map;
1396 
1397 		if (end > arm_lowmem_limit)
1398 			end = arm_lowmem_limit;
1399 		if (start >= end)
1400 			break;
1401 
1402 		if (end < kernel_x_start) {
1403 			map.pfn = __phys_to_pfn(start);
1404 			map.virtual = __phys_to_virt(start);
1405 			map.length = end - start;
1406 			map.type = MT_MEMORY_RWX;
1407 
1408 			create_mapping(&map);
1409 		} else if (start >= kernel_x_end) {
1410 			map.pfn = __phys_to_pfn(start);
1411 			map.virtual = __phys_to_virt(start);
1412 			map.length = end - start;
1413 			map.type = MT_MEMORY_RW;
1414 
1415 			create_mapping(&map);
1416 		} else {
1417 			/* This better cover the entire kernel */
1418 			if (start < kernel_x_start) {
1419 				map.pfn = __phys_to_pfn(start);
1420 				map.virtual = __phys_to_virt(start);
1421 				map.length = kernel_x_start - start;
1422 				map.type = MT_MEMORY_RW;
1423 
1424 				create_mapping(&map);
1425 			}
1426 
1427 			map.pfn = __phys_to_pfn(kernel_x_start);
1428 			map.virtual = __phys_to_virt(kernel_x_start);
1429 			map.length = kernel_x_end - kernel_x_start;
1430 			map.type = MT_MEMORY_RWX;
1431 
1432 			create_mapping(&map);
1433 
1434 			if (kernel_x_end < end) {
1435 				map.pfn = __phys_to_pfn(kernel_x_end);
1436 				map.virtual = __phys_to_virt(kernel_x_end);
1437 				map.length = end - kernel_x_end;
1438 				map.type = MT_MEMORY_RW;
1439 
1440 				create_mapping(&map);
1441 			}
1442 		}
1443 	}
1444 }
1445 
1446 #ifdef CONFIG_ARM_PV_FIXUP
1447 extern unsigned long __atags_pointer;
1448 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata);
1449 pgtables_remap lpae_pgtables_remap_asm;
1450 
1451 /*
1452  * early_paging_init() recreates boot time page table setup, allowing machines
1453  * to switch over to a high (>4G) address space on LPAE systems
1454  */
early_paging_init(const struct machine_desc * mdesc,struct proc_info_list * procinfo)1455 void __init early_paging_init(const struct machine_desc *mdesc,
1456 			      struct proc_info_list *procinfo)
1457 {
1458 	pgtables_remap *lpae_pgtables_remap;
1459 	unsigned long pa_pgd;
1460 	unsigned int cr, ttbcr;
1461 	long long offset;
1462 	void *boot_data;
1463 
1464 	if (!(mdesc->init_meminfo))
1465 		return;
1466 
1467 	offset = mdesc->init_meminfo();
1468 	if (offset == 0)
1469 		return;
1470 
1471 	/* Re-set the phys pfn offset, and the pv offset */
1472 	__pv_offset += offset;
1473 	__pv_phys_pfn_offset += PFN_DOWN(offset);
1474 
1475 	/*
1476 	 * Get the address of the remap function in the 1:1 identity
1477 	 * mapping setup by the early page table assembly code.  We
1478 	 * must get this prior to the pv update.  The following barrier
1479 	 * ensures that this is complete before we fixup any P:V offsets.
1480 	 */
1481 	lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm);
1482 	pa_pgd = __pa(swapper_pg_dir);
1483 	boot_data = __va(__atags_pointer);
1484 	barrier();
1485 
1486 	/* Run the patch stub to update the constants */
1487 	fixup_pv_table(&__pv_table_begin,
1488 		(&__pv_table_end - &__pv_table_begin) << 2);
1489 
1490 	/*
1491 	 * We changing not only the virtual to physical mapping, but also
1492 	 * the physical addresses used to access memory.  We need to flush
1493 	 * all levels of cache in the system with caching disabled to
1494 	 * ensure that all data is written back, and nothing is prefetched
1495 	 * into the caches.  We also need to prevent the TLB walkers
1496 	 * allocating into the caches too.  Note that this is ARMv7 LPAE
1497 	 * specific.
1498 	 */
1499 	cr = get_cr();
1500 	set_cr(cr & ~(CR_I | CR_C));
1501 	asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr));
1502 	asm volatile("mcr p15, 0, %0, c2, c0, 2"
1503 		: : "r" (ttbcr & ~(3 << 8 | 3 << 10)));
1504 	flush_cache_all();
1505 
1506 	/*
1507 	 * Fixup the page tables - this must be in the idmap region as
1508 	 * we need to disable the MMU to do this safely, and hence it
1509 	 * needs to be assembly.  It's fairly simple, as we're using the
1510 	 * temporary tables setup by the initial assembly code.
1511 	 */
1512 	lpae_pgtables_remap(offset, pa_pgd, boot_data);
1513 
1514 	/* Re-enable the caches and cacheable TLB walks */
1515 	asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr));
1516 	set_cr(cr);
1517 }
1518 
1519 #else
1520 
early_paging_init(const struct machine_desc * mdesc,struct proc_info_list * procinfo)1521 void __init early_paging_init(const struct machine_desc *mdesc,
1522 			      struct proc_info_list *procinfo)
1523 {
1524 	long long offset;
1525 
1526 	if (!mdesc->init_meminfo)
1527 		return;
1528 
1529 	offset = mdesc->init_meminfo();
1530 	if (offset == 0)
1531 		return;
1532 
1533 	pr_crit("Physical address space modification is only to support Keystone2.\n");
1534 	pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n");
1535 	pr_crit("feature. Your kernel may crash now, have a good day.\n");
1536 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1537 }
1538 
1539 #endif
1540 
1541 /*
1542  * paging_init() sets up the page tables, initialises the zone memory
1543  * maps, and sets up the zero page, bad page and bad page tables.
1544  */
paging_init(const struct machine_desc * mdesc)1545 void __init paging_init(const struct machine_desc *mdesc)
1546 {
1547 	void *zero_page;
1548 
1549 	build_mem_type_table();
1550 	prepare_page_table();
1551 	map_lowmem();
1552 	dma_contiguous_remap();
1553 	devicemaps_init(mdesc);
1554 	kmap_init();
1555 	tcm_init();
1556 
1557 	top_pmd = pmd_off_k(0xffff0000);
1558 
1559 	/* allocate the zero page. */
1560 	zero_page = early_alloc(PAGE_SIZE);
1561 
1562 	bootmem_init();
1563 
1564 	empty_zero_page = virt_to_page(zero_page);
1565 	__flush_dcache_page(NULL, empty_zero_page);
1566 }
1567