Loading...
1/*
2 *
3 * Copyright (C) 1995 Linus Torvalds
4 *
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6 */
7
8#include <linux/module.h>
9#include <linux/signal.h>
10#include <linux/sched.h>
11#include <linux/kernel.h>
12#include <linux/errno.h>
13#include <linux/string.h>
14#include <linux/types.h>
15#include <linux/ptrace.h>
16#include <linux/mman.h>
17#include <linux/mm.h>
18#include <linux/hugetlb.h>
19#include <linux/swap.h>
20#include <linux/smp.h>
21#include <linux/init.h>
22#include <linux/highmem.h>
23#include <linux/pagemap.h>
24#include <linux/pci.h>
25#include <linux/pfn.h>
26#include <linux/poison.h>
27#include <linux/bootmem.h>
28#include <linux/memblock.h>
29#include <linux/proc_fs.h>
30#include <linux/memory_hotplug.h>
31#include <linux/initrd.h>
32#include <linux/cpumask.h>
33#include <linux/gfp.h>
34
35#include <asm/asm.h>
36#include <asm/bios_ebda.h>
37#include <asm/processor.h>
38#include <asm/uaccess.h>
39#include <asm/pgtable.h>
40#include <asm/dma.h>
41#include <asm/fixmap.h>
42#include <asm/e820.h>
43#include <asm/apic.h>
44#include <asm/bugs.h>
45#include <asm/tlb.h>
46#include <asm/tlbflush.h>
47#include <asm/olpc_ofw.h>
48#include <asm/pgalloc.h>
49#include <asm/sections.h>
50#include <asm/paravirt.h>
51#include <asm/setup.h>
52#include <asm/cacheflush.h>
53#include <asm/page_types.h>
54#include <asm/init.h>
55
56#include "mm_internal.h"
57
58unsigned long highstart_pfn, highend_pfn;
59
60static noinline int do_test_wp_bit(void);
61
62bool __read_mostly __vmalloc_start_set = false;
63
64/*
65 * Creates a middle page table and puts a pointer to it in the
66 * given global directory entry. This only returns the gd entry
67 * in non-PAE compilation mode, since the middle layer is folded.
68 */
69static pmd_t * __init one_md_table_init(pgd_t *pgd)
70{
71 pud_t *pud;
72 pmd_t *pmd_table;
73
74#ifdef CONFIG_X86_PAE
75 if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
76 pmd_table = (pmd_t *)alloc_low_page();
77 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
78 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
79 pud = pud_offset(pgd, 0);
80 BUG_ON(pmd_table != pmd_offset(pud, 0));
81
82 return pmd_table;
83 }
84#endif
85 pud = pud_offset(pgd, 0);
86 pmd_table = pmd_offset(pud, 0);
87
88 return pmd_table;
89}
90
91/*
92 * Create a page table and place a pointer to it in a middle page
93 * directory entry:
94 */
95static pte_t * __init one_page_table_init(pmd_t *pmd)
96{
97 if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
98 pte_t *page_table = (pte_t *)alloc_low_page();
99
100 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
101 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
102 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
103 }
104
105 return pte_offset_kernel(pmd, 0);
106}
107
108pmd_t * __init populate_extra_pmd(unsigned long vaddr)
109{
110 int pgd_idx = pgd_index(vaddr);
111 int pmd_idx = pmd_index(vaddr);
112
113 return one_md_table_init(swapper_pg_dir + pgd_idx) + pmd_idx;
114}
115
116pte_t * __init populate_extra_pte(unsigned long vaddr)
117{
118 int pte_idx = pte_index(vaddr);
119 pmd_t *pmd;
120
121 pmd = populate_extra_pmd(vaddr);
122 return one_page_table_init(pmd) + pte_idx;
123}
124
125static unsigned long __init
126page_table_range_init_count(unsigned long start, unsigned long end)
127{
128 unsigned long count = 0;
129#ifdef CONFIG_HIGHMEM
130 int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
131 int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
132 int pgd_idx, pmd_idx;
133 unsigned long vaddr;
134
135 if (pmd_idx_kmap_begin == pmd_idx_kmap_end)
136 return 0;
137
138 vaddr = start;
139 pgd_idx = pgd_index(vaddr);
140 pmd_idx = pmd_index(vaddr);
141
142 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd_idx++) {
143 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
144 pmd_idx++) {
145 if ((vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin &&
146 (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end)
147 count++;
148 vaddr += PMD_SIZE;
149 }
150 pmd_idx = 0;
151 }
152#endif
153 return count;
154}
155
156static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
157 unsigned long vaddr, pte_t *lastpte,
158 void **adr)
159{
160#ifdef CONFIG_HIGHMEM
161 /*
162 * Something (early fixmap) may already have put a pte
163 * page here, which causes the page table allocation
164 * to become nonlinear. Attempt to fix it, and if it
165 * is still nonlinear then we have to bug.
166 */
167 int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
168 int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
169
170 if (pmd_idx_kmap_begin != pmd_idx_kmap_end
171 && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
172 && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end) {
173 pte_t *newpte;
174 int i;
175
176 BUG_ON(after_bootmem);
177 newpte = *adr;
178 for (i = 0; i < PTRS_PER_PTE; i++)
179 set_pte(newpte + i, pte[i]);
180 *adr = (void *)(((unsigned long)(*adr)) + PAGE_SIZE);
181
182 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
183 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
184 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
185 __flush_tlb_all();
186
187 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
188 pte = newpte;
189 }
190 BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
191 && vaddr > fix_to_virt(FIX_KMAP_END)
192 && lastpte && lastpte + PTRS_PER_PTE != pte);
193#endif
194 return pte;
195}
196
197/*
198 * This function initializes a certain range of kernel virtual memory
199 * with new bootmem page tables, everywhere page tables are missing in
200 * the given range.
201 *
202 * NOTE: The pagetables are allocated contiguous on the physical space
203 * so we can cache the place of the first one and move around without
204 * checking the pgd every time.
205 */
206static void __init
207page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
208{
209 int pgd_idx, pmd_idx;
210 unsigned long vaddr;
211 pgd_t *pgd;
212 pmd_t *pmd;
213 pte_t *pte = NULL;
214 unsigned long count = page_table_range_init_count(start, end);
215 void *adr = NULL;
216
217 if (count)
218 adr = alloc_low_pages(count);
219
220 vaddr = start;
221 pgd_idx = pgd_index(vaddr);
222 pmd_idx = pmd_index(vaddr);
223 pgd = pgd_base + pgd_idx;
224
225 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
226 pmd = one_md_table_init(pgd);
227 pmd = pmd + pmd_index(vaddr);
228 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
229 pmd++, pmd_idx++) {
230 pte = page_table_kmap_check(one_page_table_init(pmd),
231 pmd, vaddr, pte, &adr);
232
233 vaddr += PMD_SIZE;
234 }
235 pmd_idx = 0;
236 }
237}
238
239static inline int is_kernel_text(unsigned long addr)
240{
241 if (addr >= (unsigned long)_text && addr <= (unsigned long)__init_end)
242 return 1;
243 return 0;
244}
245
246/*
247 * This maps the physical memory to kernel virtual address space, a total
248 * of max_low_pfn pages, by creating page tables starting from address
249 * PAGE_OFFSET:
250 */
251unsigned long __init
252kernel_physical_mapping_init(unsigned long start,
253 unsigned long end,
254 unsigned long page_size_mask)
255{
256 int use_pse = page_size_mask == (1<<PG_LEVEL_2M);
257 unsigned long last_map_addr = end;
258 unsigned long start_pfn, end_pfn;
259 pgd_t *pgd_base = swapper_pg_dir;
260 int pgd_idx, pmd_idx, pte_ofs;
261 unsigned long pfn;
262 pgd_t *pgd;
263 pmd_t *pmd;
264 pte_t *pte;
265 unsigned pages_2m, pages_4k;
266 int mapping_iter;
267
268 start_pfn = start >> PAGE_SHIFT;
269 end_pfn = end >> PAGE_SHIFT;
270
271 /*
272 * First iteration will setup identity mapping using large/small pages
273 * based on use_pse, with other attributes same as set by
274 * the early code in head_32.S
275 *
276 * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
277 * as desired for the kernel identity mapping.
278 *
279 * This two pass mechanism conforms to the TLB app note which says:
280 *
281 * "Software should not write to a paging-structure entry in a way
282 * that would change, for any linear address, both the page size
283 * and either the page frame or attributes."
284 */
285 mapping_iter = 1;
286
287 if (!cpu_has_pse)
288 use_pse = 0;
289
290repeat:
291 pages_2m = pages_4k = 0;
292 pfn = start_pfn;
293 pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
294 pgd = pgd_base + pgd_idx;
295 for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
296 pmd = one_md_table_init(pgd);
297
298 if (pfn >= end_pfn)
299 continue;
300#ifdef CONFIG_X86_PAE
301 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
302 pmd += pmd_idx;
303#else
304 pmd_idx = 0;
305#endif
306 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
307 pmd++, pmd_idx++) {
308 unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
309
310 /*
311 * Map with big pages if possible, otherwise
312 * create normal page tables:
313 */
314 if (use_pse) {
315 unsigned int addr2;
316 pgprot_t prot = PAGE_KERNEL_LARGE;
317 /*
318 * first pass will use the same initial
319 * identity mapping attribute + _PAGE_PSE.
320 */
321 pgprot_t init_prot =
322 __pgprot(PTE_IDENT_ATTR |
323 _PAGE_PSE);
324
325 pfn &= PMD_MASK >> PAGE_SHIFT;
326 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
327 PAGE_OFFSET + PAGE_SIZE-1;
328
329 if (is_kernel_text(addr) ||
330 is_kernel_text(addr2))
331 prot = PAGE_KERNEL_LARGE_EXEC;
332
333 pages_2m++;
334 if (mapping_iter == 1)
335 set_pmd(pmd, pfn_pmd(pfn, init_prot));
336 else
337 set_pmd(pmd, pfn_pmd(pfn, prot));
338
339 pfn += PTRS_PER_PTE;
340 continue;
341 }
342 pte = one_page_table_init(pmd);
343
344 pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
345 pte += pte_ofs;
346 for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
347 pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
348 pgprot_t prot = PAGE_KERNEL;
349 /*
350 * first pass will use the same initial
351 * identity mapping attribute.
352 */
353 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
354
355 if (is_kernel_text(addr))
356 prot = PAGE_KERNEL_EXEC;
357
358 pages_4k++;
359 if (mapping_iter == 1) {
360 set_pte(pte, pfn_pte(pfn, init_prot));
361 last_map_addr = (pfn << PAGE_SHIFT) + PAGE_SIZE;
362 } else
363 set_pte(pte, pfn_pte(pfn, prot));
364 }
365 }
366 }
367 if (mapping_iter == 1) {
368 /*
369 * update direct mapping page count only in the first
370 * iteration.
371 */
372 update_page_count(PG_LEVEL_2M, pages_2m);
373 update_page_count(PG_LEVEL_4K, pages_4k);
374
375 /*
376 * local global flush tlb, which will flush the previous
377 * mappings present in both small and large page TLB's.
378 */
379 __flush_tlb_all();
380
381 /*
382 * Second iteration will set the actual desired PTE attributes.
383 */
384 mapping_iter = 2;
385 goto repeat;
386 }
387 return last_map_addr;
388}
389
390pte_t *kmap_pte;
391
392static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
393{
394 return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
395 vaddr), vaddr), vaddr);
396}
397
398static void __init kmap_init(void)
399{
400 unsigned long kmap_vstart;
401
402 /*
403 * Cache the first kmap pte:
404 */
405 kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
406 kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
407}
408
409#ifdef CONFIG_HIGHMEM
410static void __init permanent_kmaps_init(pgd_t *pgd_base)
411{
412 unsigned long vaddr;
413 pgd_t *pgd;
414 pud_t *pud;
415 pmd_t *pmd;
416 pte_t *pte;
417
418 vaddr = PKMAP_BASE;
419 page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
420
421 pgd = swapper_pg_dir + pgd_index(vaddr);
422 pud = pud_offset(pgd, vaddr);
423 pmd = pmd_offset(pud, vaddr);
424 pte = pte_offset_kernel(pmd, vaddr);
425 pkmap_page_table = pte;
426}
427
428void __init add_highpages_with_active_regions(int nid,
429 unsigned long start_pfn, unsigned long end_pfn)
430{
431 phys_addr_t start, end;
432 u64 i;
433
434 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &start, &end, NULL) {
435 unsigned long pfn = clamp_t(unsigned long, PFN_UP(start),
436 start_pfn, end_pfn);
437 unsigned long e_pfn = clamp_t(unsigned long, PFN_DOWN(end),
438 start_pfn, end_pfn);
439 for ( ; pfn < e_pfn; pfn++)
440 if (pfn_valid(pfn))
441 free_highmem_page(pfn_to_page(pfn));
442 }
443}
444#else
445static inline void permanent_kmaps_init(pgd_t *pgd_base)
446{
447}
448#endif /* CONFIG_HIGHMEM */
449
450void __init native_pagetable_init(void)
451{
452 unsigned long pfn, va;
453 pgd_t *pgd, *base = swapper_pg_dir;
454 pud_t *pud;
455 pmd_t *pmd;
456 pte_t *pte;
457
458 /*
459 * Remove any mappings which extend past the end of physical
460 * memory from the boot time page table.
461 * In virtual address space, we should have at least two pages
462 * from VMALLOC_END to pkmap or fixmap according to VMALLOC_END
463 * definition. And max_low_pfn is set to VMALLOC_END physical
464 * address. If initial memory mapping is doing right job, we
465 * should have pte used near max_low_pfn or one pmd is not present.
466 */
467 for (pfn = max_low_pfn; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
468 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
469 pgd = base + pgd_index(va);
470 if (!pgd_present(*pgd))
471 break;
472
473 pud = pud_offset(pgd, va);
474 pmd = pmd_offset(pud, va);
475 if (!pmd_present(*pmd))
476 break;
477
478 /* should not be large page here */
479 if (pmd_large(*pmd)) {
480 pr_warn("try to clear pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx, but pmd is big page and is not using pte !\n",
481 pfn, pmd, __pa(pmd));
482 BUG_ON(1);
483 }
484
485 pte = pte_offset_kernel(pmd, va);
486 if (!pte_present(*pte))
487 break;
488
489 printk(KERN_DEBUG "clearing pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx pte: %p pte phys: %lx\n",
490 pfn, pmd, __pa(pmd), pte, __pa(pte));
491 pte_clear(NULL, va, pte);
492 }
493 paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
494 paging_init();
495}
496
497/*
498 * Build a proper pagetable for the kernel mappings. Up until this
499 * point, we've been running on some set of pagetables constructed by
500 * the boot process.
501 *
502 * If we're booting on native hardware, this will be a pagetable
503 * constructed in arch/x86/kernel/head_32.S. The root of the
504 * pagetable will be swapper_pg_dir.
505 *
506 * If we're booting paravirtualized under a hypervisor, then there are
507 * more options: we may already be running PAE, and the pagetable may
508 * or may not be based in swapper_pg_dir. In any case,
509 * paravirt_pagetable_init() will set up swapper_pg_dir
510 * appropriately for the rest of the initialization to work.
511 *
512 * In general, pagetable_init() assumes that the pagetable may already
513 * be partially populated, and so it avoids stomping on any existing
514 * mappings.
515 */
516void __init early_ioremap_page_table_range_init(void)
517{
518 pgd_t *pgd_base = swapper_pg_dir;
519 unsigned long vaddr, end;
520
521 /*
522 * Fixed mappings, only the page table structure has to be
523 * created - mappings will be set by set_fixmap():
524 */
525 vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
526 end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
527 page_table_range_init(vaddr, end, pgd_base);
528 early_ioremap_reset();
529}
530
531static void __init pagetable_init(void)
532{
533 pgd_t *pgd_base = swapper_pg_dir;
534
535 permanent_kmaps_init(pgd_base);
536}
537
538pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL);
539EXPORT_SYMBOL_GPL(__supported_pte_mask);
540
541/* user-defined highmem size */
542static unsigned int highmem_pages = -1;
543
544/*
545 * highmem=size forces highmem to be exactly 'size' bytes.
546 * This works even on boxes that have no highmem otherwise.
547 * This also works to reduce highmem size on bigger boxes.
548 */
549static int __init parse_highmem(char *arg)
550{
551 if (!arg)
552 return -EINVAL;
553
554 highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
555 return 0;
556}
557early_param("highmem", parse_highmem);
558
559#define MSG_HIGHMEM_TOO_BIG \
560 "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
561
562#define MSG_LOWMEM_TOO_SMALL \
563 "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
564/*
565 * All of RAM fits into lowmem - but if user wants highmem
566 * artificially via the highmem=x boot parameter then create
567 * it:
568 */
569static void __init lowmem_pfn_init(void)
570{
571 /* max_low_pfn is 0, we already have early_res support */
572 max_low_pfn = max_pfn;
573
574 if (highmem_pages == -1)
575 highmem_pages = 0;
576#ifdef CONFIG_HIGHMEM
577 if (highmem_pages >= max_pfn) {
578 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
579 pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
580 highmem_pages = 0;
581 }
582 if (highmem_pages) {
583 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
584 printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
585 pages_to_mb(highmem_pages));
586 highmem_pages = 0;
587 }
588 max_low_pfn -= highmem_pages;
589 }
590#else
591 if (highmem_pages)
592 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
593#endif
594}
595
596#define MSG_HIGHMEM_TOO_SMALL \
597 "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
598
599#define MSG_HIGHMEM_TRIMMED \
600 "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
601/*
602 * We have more RAM than fits into lowmem - we try to put it into
603 * highmem, also taking the highmem=x boot parameter into account:
604 */
605static void __init highmem_pfn_init(void)
606{
607 max_low_pfn = MAXMEM_PFN;
608
609 if (highmem_pages == -1)
610 highmem_pages = max_pfn - MAXMEM_PFN;
611
612 if (highmem_pages + MAXMEM_PFN < max_pfn)
613 max_pfn = MAXMEM_PFN + highmem_pages;
614
615 if (highmem_pages + MAXMEM_PFN > max_pfn) {
616 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
617 pages_to_mb(max_pfn - MAXMEM_PFN),
618 pages_to_mb(highmem_pages));
619 highmem_pages = 0;
620 }
621#ifndef CONFIG_HIGHMEM
622 /* Maximum memory usable is what is directly addressable */
623 printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
624 if (max_pfn > MAX_NONPAE_PFN)
625 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
626 else
627 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
628 max_pfn = MAXMEM_PFN;
629#else /* !CONFIG_HIGHMEM */
630#ifndef CONFIG_HIGHMEM64G
631 if (max_pfn > MAX_NONPAE_PFN) {
632 max_pfn = MAX_NONPAE_PFN;
633 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
634 }
635#endif /* !CONFIG_HIGHMEM64G */
636#endif /* !CONFIG_HIGHMEM */
637}
638
639/*
640 * Determine low and high memory ranges:
641 */
642void __init find_low_pfn_range(void)
643{
644 /* it could update max_pfn */
645
646 if (max_pfn <= MAXMEM_PFN)
647 lowmem_pfn_init();
648 else
649 highmem_pfn_init();
650}
651
652#ifndef CONFIG_NEED_MULTIPLE_NODES
653void __init initmem_init(void)
654{
655#ifdef CONFIG_HIGHMEM
656 highstart_pfn = highend_pfn = max_pfn;
657 if (max_pfn > max_low_pfn)
658 highstart_pfn = max_low_pfn;
659 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
660 pages_to_mb(highend_pfn - highstart_pfn));
661 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
662#else
663 high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
664#endif
665
666 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
667 sparse_memory_present_with_active_regions(0);
668
669#ifdef CONFIG_FLATMEM
670 max_mapnr = IS_ENABLED(CONFIG_HIGHMEM) ? highend_pfn : max_low_pfn;
671#endif
672 __vmalloc_start_set = true;
673
674 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
675 pages_to_mb(max_low_pfn));
676
677 setup_bootmem_allocator();
678}
679#endif /* !CONFIG_NEED_MULTIPLE_NODES */
680
681void __init setup_bootmem_allocator(void)
682{
683 printk(KERN_INFO " mapped low ram: 0 - %08lx\n",
684 max_pfn_mapped<<PAGE_SHIFT);
685 printk(KERN_INFO " low ram: 0 - %08lx\n", max_low_pfn<<PAGE_SHIFT);
686}
687
688/*
689 * paging_init() sets up the page tables - note that the first 8MB are
690 * already mapped by head.S.
691 *
692 * This routines also unmaps the page at virtual kernel address 0, so
693 * that we can trap those pesky NULL-reference errors in the kernel.
694 */
695void __init paging_init(void)
696{
697 pagetable_init();
698
699 __flush_tlb_all();
700
701 kmap_init();
702
703 /*
704 * NOTE: at this point the bootmem allocator is fully available.
705 */
706 olpc_dt_build_devicetree();
707 sparse_memory_present_with_active_regions(MAX_NUMNODES);
708 sparse_init();
709 zone_sizes_init();
710}
711
712/*
713 * Test if the WP bit works in supervisor mode. It isn't supported on 386's
714 * and also on some strange 486's. All 586+'s are OK. This used to involve
715 * black magic jumps to work around some nasty CPU bugs, but fortunately the
716 * switch to using exceptions got rid of all that.
717 */
718static void __init test_wp_bit(void)
719{
720 printk(KERN_INFO
721 "Checking if this processor honours the WP bit even in supervisor mode...");
722
723 /* Any page-aligned address will do, the test is non-destructive */
724 __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_KERNEL_RO);
725 boot_cpu_data.wp_works_ok = do_test_wp_bit();
726 clear_fixmap(FIX_WP_TEST);
727
728 if (!boot_cpu_data.wp_works_ok) {
729 printk(KERN_CONT "No.\n");
730 panic("Linux doesn't support CPUs with broken WP.");
731 } else {
732 printk(KERN_CONT "Ok.\n");
733 }
734}
735
736void __init mem_init(void)
737{
738 pci_iommu_alloc();
739
740#ifdef CONFIG_FLATMEM
741 BUG_ON(!mem_map);
742#endif
743 /*
744 * With CONFIG_DEBUG_PAGEALLOC initialization of highmem pages has to
745 * be done before free_all_bootmem(). Memblock use free low memory for
746 * temporary data (see find_range_array()) and for this purpose can use
747 * pages that was already passed to the buddy allocator, hence marked as
748 * not accessible in the page tables when compiled with
749 * CONFIG_DEBUG_PAGEALLOC. Otherwise order of initialization is not
750 * important here.
751 */
752 set_highmem_pages_init();
753
754 /* this will put all low memory onto the freelists */
755 free_all_bootmem();
756
757 after_bootmem = 1;
758
759 mem_init_print_info(NULL);
760 printk(KERN_INFO "virtual kernel memory layout:\n"
761 " fixmap : 0x%08lx - 0x%08lx (%4ld kB)\n"
762#ifdef CONFIG_HIGHMEM
763 " pkmap : 0x%08lx - 0x%08lx (%4ld kB)\n"
764#endif
765 " vmalloc : 0x%08lx - 0x%08lx (%4ld MB)\n"
766 " lowmem : 0x%08lx - 0x%08lx (%4ld MB)\n"
767 " .init : 0x%08lx - 0x%08lx (%4ld kB)\n"
768 " .data : 0x%08lx - 0x%08lx (%4ld kB)\n"
769 " .text : 0x%08lx - 0x%08lx (%4ld kB)\n",
770 FIXADDR_START, FIXADDR_TOP,
771 (FIXADDR_TOP - FIXADDR_START) >> 10,
772
773#ifdef CONFIG_HIGHMEM
774 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
775 (LAST_PKMAP*PAGE_SIZE) >> 10,
776#endif
777
778 VMALLOC_START, VMALLOC_END,
779 (VMALLOC_END - VMALLOC_START) >> 20,
780
781 (unsigned long)__va(0), (unsigned long)high_memory,
782 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
783
784 (unsigned long)&__init_begin, (unsigned long)&__init_end,
785 ((unsigned long)&__init_end -
786 (unsigned long)&__init_begin) >> 10,
787
788 (unsigned long)&_etext, (unsigned long)&_edata,
789 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
790
791 (unsigned long)&_text, (unsigned long)&_etext,
792 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
793
794 /*
795 * Check boundaries twice: Some fundamental inconsistencies can
796 * be detected at build time already.
797 */
798#define __FIXADDR_TOP (-PAGE_SIZE)
799#ifdef CONFIG_HIGHMEM
800 BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE > FIXADDR_START);
801 BUILD_BUG_ON(VMALLOC_END > PKMAP_BASE);
802#endif
803#define high_memory (-128UL << 20)
804 BUILD_BUG_ON(VMALLOC_START >= VMALLOC_END);
805#undef high_memory
806#undef __FIXADDR_TOP
807#ifdef CONFIG_RANDOMIZE_BASE
808 BUILD_BUG_ON(CONFIG_RANDOMIZE_BASE_MAX_OFFSET > KERNEL_IMAGE_SIZE);
809#endif
810
811#ifdef CONFIG_HIGHMEM
812 BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE > FIXADDR_START);
813 BUG_ON(VMALLOC_END > PKMAP_BASE);
814#endif
815 BUG_ON(VMALLOC_START >= VMALLOC_END);
816 BUG_ON((unsigned long)high_memory > VMALLOC_START);
817
818 if (boot_cpu_data.wp_works_ok < 0)
819 test_wp_bit();
820}
821
822#ifdef CONFIG_MEMORY_HOTPLUG
823int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
824{
825 struct pglist_data *pgdata = NODE_DATA(nid);
826 struct zone *zone = pgdata->node_zones +
827 zone_for_memory(nid, start, size, ZONE_HIGHMEM, for_device);
828 unsigned long start_pfn = start >> PAGE_SHIFT;
829 unsigned long nr_pages = size >> PAGE_SHIFT;
830
831 return __add_pages(nid, zone, start_pfn, nr_pages);
832}
833
834#ifdef CONFIG_MEMORY_HOTREMOVE
835int arch_remove_memory(u64 start, u64 size)
836{
837 unsigned long start_pfn = start >> PAGE_SHIFT;
838 unsigned long nr_pages = size >> PAGE_SHIFT;
839 struct zone *zone;
840
841 zone = page_zone(pfn_to_page(start_pfn));
842 return __remove_pages(zone, start_pfn, nr_pages);
843}
844#endif
845#endif
846
847/*
848 * This function cannot be __init, since exceptions don't work in that
849 * section. Put this after the callers, so that it cannot be inlined.
850 */
851static noinline int do_test_wp_bit(void)
852{
853 char tmp_reg;
854 int flag;
855
856 __asm__ __volatile__(
857 " movb %0, %1 \n"
858 "1: movb %1, %0 \n"
859 " xorl %2, %2 \n"
860 "2: \n"
861 _ASM_EXTABLE(1b,2b)
862 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
863 "=q" (tmp_reg),
864 "=r" (flag)
865 :"2" (1)
866 :"memory");
867
868 return flag;
869}
870
871const int rodata_test_data = 0xC3;
872EXPORT_SYMBOL_GPL(rodata_test_data);
873
874int kernel_set_to_readonly __read_mostly;
875
876void set_kernel_text_rw(void)
877{
878 unsigned long start = PFN_ALIGN(_text);
879 unsigned long size = PFN_ALIGN(_etext) - start;
880
881 if (!kernel_set_to_readonly)
882 return;
883
884 pr_debug("Set kernel text: %lx - %lx for read write\n",
885 start, start+size);
886
887 set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
888}
889
890void set_kernel_text_ro(void)
891{
892 unsigned long start = PFN_ALIGN(_text);
893 unsigned long size = PFN_ALIGN(_etext) - start;
894
895 if (!kernel_set_to_readonly)
896 return;
897
898 pr_debug("Set kernel text: %lx - %lx for read only\n",
899 start, start+size);
900
901 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
902}
903
904static void mark_nxdata_nx(void)
905{
906 /*
907 * When this called, init has already been executed and released,
908 * so everything past _etext should be NX.
909 */
910 unsigned long start = PFN_ALIGN(_etext);
911 /*
912 * This comes from is_kernel_text upper limit. Also HPAGE where used:
913 */
914 unsigned long size = (((unsigned long)__init_end + HPAGE_SIZE) & HPAGE_MASK) - start;
915
916 if (__supported_pte_mask & _PAGE_NX)
917 printk(KERN_INFO "NX-protecting the kernel data: %luk\n", size >> 10);
918 set_pages_nx(virt_to_page(start), size >> PAGE_SHIFT);
919}
920
921void mark_rodata_ro(void)
922{
923 unsigned long start = PFN_ALIGN(_text);
924 unsigned long size = PFN_ALIGN(_etext) - start;
925
926 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
927 printk(KERN_INFO "Write protecting the kernel text: %luk\n",
928 size >> 10);
929
930 kernel_set_to_readonly = 1;
931
932#ifdef CONFIG_CPA_DEBUG
933 printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
934 start, start+size);
935 set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
936
937 printk(KERN_INFO "Testing CPA: write protecting again\n");
938 set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
939#endif
940
941 start += size;
942 size = (unsigned long)__end_rodata - start;
943 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
944 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
945 size >> 10);
946 rodata_test();
947
948#ifdef CONFIG_CPA_DEBUG
949 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
950 set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
951
952 printk(KERN_INFO "Testing CPA: write protecting again\n");
953 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
954#endif
955 mark_nxdata_nx();
956 if (__supported_pte_mask & _PAGE_NX)
957 debug_checkwx();
958}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 *
4 * Copyright (C) 1995 Linus Torvalds
5 *
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
7 */
8
9#include <linux/signal.h>
10#include <linux/sched.h>
11#include <linux/kernel.h>
12#include <linux/errno.h>
13#include <linux/string.h>
14#include <linux/types.h>
15#include <linux/ptrace.h>
16#include <linux/mman.h>
17#include <linux/mm.h>
18#include <linux/hugetlb.h>
19#include <linux/swap.h>
20#include <linux/smp.h>
21#include <linux/init.h>
22#include <linux/highmem.h>
23#include <linux/pagemap.h>
24#include <linux/pci.h>
25#include <linux/pfn.h>
26#include <linux/poison.h>
27#include <linux/memblock.h>
28#include <linux/proc_fs.h>
29#include <linux/memory_hotplug.h>
30#include <linux/initrd.h>
31#include <linux/cpumask.h>
32#include <linux/gfp.h>
33
34#include <asm/asm.h>
35#include <asm/bios_ebda.h>
36#include <asm/processor.h>
37#include <linux/uaccess.h>
38#include <asm/dma.h>
39#include <asm/fixmap.h>
40#include <asm/e820/api.h>
41#include <asm/apic.h>
42#include <asm/bugs.h>
43#include <asm/tlb.h>
44#include <asm/tlbflush.h>
45#include <asm/olpc_ofw.h>
46#include <asm/pgalloc.h>
47#include <asm/sections.h>
48#include <asm/paravirt.h>
49#include <asm/setup.h>
50#include <asm/set_memory.h>
51#include <asm/page_types.h>
52#include <asm/cpu_entry_area.h>
53#include <asm/init.h>
54#include <asm/pgtable_areas.h>
55#include <asm/numa.h>
56
57#include "mm_internal.h"
58
59unsigned long highstart_pfn, highend_pfn;
60
61bool __read_mostly __vmalloc_start_set = false;
62
63/*
64 * Creates a middle page table and puts a pointer to it in the
65 * given global directory entry. This only returns the gd entry
66 * in non-PAE compilation mode, since the middle layer is folded.
67 */
68static pmd_t * __init one_md_table_init(pgd_t *pgd)
69{
70 p4d_t *p4d;
71 pud_t *pud;
72 pmd_t *pmd_table;
73
74#ifdef CONFIG_X86_PAE
75 if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
76 pmd_table = (pmd_t *)alloc_low_page();
77 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
78 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
79 p4d = p4d_offset(pgd, 0);
80 pud = pud_offset(p4d, 0);
81 BUG_ON(pmd_table != pmd_offset(pud, 0));
82
83 return pmd_table;
84 }
85#endif
86 p4d = p4d_offset(pgd, 0);
87 pud = pud_offset(p4d, 0);
88 pmd_table = pmd_offset(pud, 0);
89
90 return pmd_table;
91}
92
93/*
94 * Create a page table and place a pointer to it in a middle page
95 * directory entry:
96 */
97static pte_t * __init one_page_table_init(pmd_t *pmd)
98{
99 if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
100 pte_t *page_table = (pte_t *)alloc_low_page();
101
102 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
103 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
104 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
105 }
106
107 return pte_offset_kernel(pmd, 0);
108}
109
110pmd_t * __init populate_extra_pmd(unsigned long vaddr)
111{
112 int pgd_idx = pgd_index(vaddr);
113 int pmd_idx = pmd_index(vaddr);
114
115 return one_md_table_init(swapper_pg_dir + pgd_idx) + pmd_idx;
116}
117
118pte_t * __init populate_extra_pte(unsigned long vaddr)
119{
120 int pte_idx = pte_index(vaddr);
121 pmd_t *pmd;
122
123 pmd = populate_extra_pmd(vaddr);
124 return one_page_table_init(pmd) + pte_idx;
125}
126
127static unsigned long __init
128page_table_range_init_count(unsigned long start, unsigned long end)
129{
130 unsigned long count = 0;
131#ifdef CONFIG_HIGHMEM
132 int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
133 int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
134 int pgd_idx, pmd_idx;
135 unsigned long vaddr;
136
137 if (pmd_idx_kmap_begin == pmd_idx_kmap_end)
138 return 0;
139
140 vaddr = start;
141 pgd_idx = pgd_index(vaddr);
142 pmd_idx = pmd_index(vaddr);
143
144 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd_idx++) {
145 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
146 pmd_idx++) {
147 if ((vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin &&
148 (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end)
149 count++;
150 vaddr += PMD_SIZE;
151 }
152 pmd_idx = 0;
153 }
154#endif
155 return count;
156}
157
158static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
159 unsigned long vaddr, pte_t *lastpte,
160 void **adr)
161{
162#ifdef CONFIG_HIGHMEM
163 /*
164 * Something (early fixmap) may already have put a pte
165 * page here, which causes the page table allocation
166 * to become nonlinear. Attempt to fix it, and if it
167 * is still nonlinear then we have to bug.
168 */
169 int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
170 int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
171
172 if (pmd_idx_kmap_begin != pmd_idx_kmap_end
173 && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
174 && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end) {
175 pte_t *newpte;
176 int i;
177
178 BUG_ON(after_bootmem);
179 newpte = *adr;
180 for (i = 0; i < PTRS_PER_PTE; i++)
181 set_pte(newpte + i, pte[i]);
182 *adr = (void *)(((unsigned long)(*adr)) + PAGE_SIZE);
183
184 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
185 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
186 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
187 __flush_tlb_all();
188
189 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
190 pte = newpte;
191 }
192 BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
193 && vaddr > fix_to_virt(FIX_KMAP_END)
194 && lastpte && lastpte + PTRS_PER_PTE != pte);
195#endif
196 return pte;
197}
198
199/*
200 * This function initializes a certain range of kernel virtual memory
201 * with new bootmem page tables, everywhere page tables are missing in
202 * the given range.
203 *
204 * NOTE: The pagetables are allocated contiguous on the physical space
205 * so we can cache the place of the first one and move around without
206 * checking the pgd every time.
207 */
208static void __init
209page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
210{
211 int pgd_idx, pmd_idx;
212 unsigned long vaddr;
213 pgd_t *pgd;
214 pmd_t *pmd;
215 pte_t *pte = NULL;
216 unsigned long count = page_table_range_init_count(start, end);
217 void *adr = NULL;
218
219 if (count)
220 adr = alloc_low_pages(count);
221
222 vaddr = start;
223 pgd_idx = pgd_index(vaddr);
224 pmd_idx = pmd_index(vaddr);
225 pgd = pgd_base + pgd_idx;
226
227 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
228 pmd = one_md_table_init(pgd);
229 pmd = pmd + pmd_index(vaddr);
230 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
231 pmd++, pmd_idx++) {
232 pte = page_table_kmap_check(one_page_table_init(pmd),
233 pmd, vaddr, pte, &adr);
234
235 vaddr += PMD_SIZE;
236 }
237 pmd_idx = 0;
238 }
239}
240
241/*
242 * The <linux/kallsyms.h> already defines is_kernel_text,
243 * using '__' prefix not to get in conflict.
244 */
245static inline int __is_kernel_text(unsigned long addr)
246{
247 if (addr >= (unsigned long)_text && addr <= (unsigned long)__init_end)
248 return 1;
249 return 0;
250}
251
252/*
253 * This maps the physical memory to kernel virtual address space, a total
254 * of max_low_pfn pages, by creating page tables starting from address
255 * PAGE_OFFSET:
256 */
257unsigned long __init
258kernel_physical_mapping_init(unsigned long start,
259 unsigned long end,
260 unsigned long page_size_mask,
261 pgprot_t prot)
262{
263 int use_pse = page_size_mask == (1<<PG_LEVEL_2M);
264 unsigned long last_map_addr = end;
265 unsigned long start_pfn, end_pfn;
266 pgd_t *pgd_base = swapper_pg_dir;
267 int pgd_idx, pmd_idx, pte_ofs;
268 unsigned long pfn;
269 pgd_t *pgd;
270 pmd_t *pmd;
271 pte_t *pte;
272 unsigned pages_2m, pages_4k;
273 int mapping_iter;
274
275 start_pfn = start >> PAGE_SHIFT;
276 end_pfn = end >> PAGE_SHIFT;
277
278 /*
279 * First iteration will setup identity mapping using large/small pages
280 * based on use_pse, with other attributes same as set by
281 * the early code in head_32.S
282 *
283 * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
284 * as desired for the kernel identity mapping.
285 *
286 * This two pass mechanism conforms to the TLB app note which says:
287 *
288 * "Software should not write to a paging-structure entry in a way
289 * that would change, for any linear address, both the page size
290 * and either the page frame or attributes."
291 */
292 mapping_iter = 1;
293
294 if (!boot_cpu_has(X86_FEATURE_PSE))
295 use_pse = 0;
296
297repeat:
298 pages_2m = pages_4k = 0;
299 pfn = start_pfn;
300 pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
301 pgd = pgd_base + pgd_idx;
302 for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
303 pmd = one_md_table_init(pgd);
304
305 if (pfn >= end_pfn)
306 continue;
307#ifdef CONFIG_X86_PAE
308 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
309 pmd += pmd_idx;
310#else
311 pmd_idx = 0;
312#endif
313 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
314 pmd++, pmd_idx++) {
315 unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
316
317 /*
318 * Map with big pages if possible, otherwise
319 * create normal page tables:
320 */
321 if (use_pse) {
322 unsigned int addr2;
323 pgprot_t prot = PAGE_KERNEL_LARGE;
324 /*
325 * first pass will use the same initial
326 * identity mapping attribute + _PAGE_PSE.
327 */
328 pgprot_t init_prot =
329 __pgprot(PTE_IDENT_ATTR |
330 _PAGE_PSE);
331
332 pfn &= PMD_MASK >> PAGE_SHIFT;
333 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
334 PAGE_OFFSET + PAGE_SIZE-1;
335
336 if (__is_kernel_text(addr) ||
337 __is_kernel_text(addr2))
338 prot = PAGE_KERNEL_LARGE_EXEC;
339
340 pages_2m++;
341 if (mapping_iter == 1)
342 set_pmd(pmd, pfn_pmd(pfn, init_prot));
343 else
344 set_pmd(pmd, pfn_pmd(pfn, prot));
345
346 pfn += PTRS_PER_PTE;
347 continue;
348 }
349 pte = one_page_table_init(pmd);
350
351 pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
352 pte += pte_ofs;
353 for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
354 pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
355 pgprot_t prot = PAGE_KERNEL;
356 /*
357 * first pass will use the same initial
358 * identity mapping attribute.
359 */
360 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
361
362 if (__is_kernel_text(addr))
363 prot = PAGE_KERNEL_EXEC;
364
365 pages_4k++;
366 if (mapping_iter == 1) {
367 set_pte(pte, pfn_pte(pfn, init_prot));
368 last_map_addr = (pfn << PAGE_SHIFT) + PAGE_SIZE;
369 } else
370 set_pte(pte, pfn_pte(pfn, prot));
371 }
372 }
373 }
374 if (mapping_iter == 1) {
375 /*
376 * update direct mapping page count only in the first
377 * iteration.
378 */
379 update_page_count(PG_LEVEL_2M, pages_2m);
380 update_page_count(PG_LEVEL_4K, pages_4k);
381
382 /*
383 * local global flush tlb, which will flush the previous
384 * mappings present in both small and large page TLB's.
385 */
386 __flush_tlb_all();
387
388 /*
389 * Second iteration will set the actual desired PTE attributes.
390 */
391 mapping_iter = 2;
392 goto repeat;
393 }
394 return last_map_addr;
395}
396
397pte_t *kmap_pte;
398
399static void __init kmap_init(void)
400{
401 unsigned long kmap_vstart;
402
403 /*
404 * Cache the first kmap pte:
405 */
406 kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
407 kmap_pte = virt_to_kpte(kmap_vstart);
408}
409
410#ifdef CONFIG_HIGHMEM
411static void __init permanent_kmaps_init(pgd_t *pgd_base)
412{
413 unsigned long vaddr = PKMAP_BASE;
414
415 page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
416
417 pkmap_page_table = virt_to_kpte(vaddr);
418}
419
420void __init add_highpages_with_active_regions(int nid,
421 unsigned long start_pfn, unsigned long end_pfn)
422{
423 phys_addr_t start, end;
424 u64 i;
425
426 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &start, &end, NULL) {
427 unsigned long pfn = clamp_t(unsigned long, PFN_UP(start),
428 start_pfn, end_pfn);
429 unsigned long e_pfn = clamp_t(unsigned long, PFN_DOWN(end),
430 start_pfn, end_pfn);
431 for ( ; pfn < e_pfn; pfn++)
432 if (pfn_valid(pfn))
433 free_highmem_page(pfn_to_page(pfn));
434 }
435}
436#else
437static inline void permanent_kmaps_init(pgd_t *pgd_base)
438{
439}
440#endif /* CONFIG_HIGHMEM */
441
442void __init sync_initial_page_table(void)
443{
444 clone_pgd_range(initial_page_table + KERNEL_PGD_BOUNDARY,
445 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
446 KERNEL_PGD_PTRS);
447
448 /*
449 * sync back low identity map too. It is used for example
450 * in the 32-bit EFI stub.
451 */
452 clone_pgd_range(initial_page_table,
453 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
454 min(KERNEL_PGD_PTRS, KERNEL_PGD_BOUNDARY));
455}
456
457void __init native_pagetable_init(void)
458{
459 unsigned long pfn, va;
460 pgd_t *pgd, *base = swapper_pg_dir;
461 p4d_t *p4d;
462 pud_t *pud;
463 pmd_t *pmd;
464 pte_t *pte;
465
466 /*
467 * Remove any mappings which extend past the end of physical
468 * memory from the boot time page table.
469 * In virtual address space, we should have at least two pages
470 * from VMALLOC_END to pkmap or fixmap according to VMALLOC_END
471 * definition. And max_low_pfn is set to VMALLOC_END physical
472 * address. If initial memory mapping is doing right job, we
473 * should have pte used near max_low_pfn or one pmd is not present.
474 */
475 for (pfn = max_low_pfn; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
476 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
477 pgd = base + pgd_index(va);
478 if (!pgd_present(*pgd))
479 break;
480
481 p4d = p4d_offset(pgd, va);
482 pud = pud_offset(p4d, va);
483 pmd = pmd_offset(pud, va);
484 if (!pmd_present(*pmd))
485 break;
486
487 /* should not be large page here */
488 if (pmd_large(*pmd)) {
489 pr_warn("try to clear pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx, but pmd is big page and is not using pte !\n",
490 pfn, pmd, __pa(pmd));
491 BUG_ON(1);
492 }
493
494 pte = pte_offset_kernel(pmd, va);
495 if (!pte_present(*pte))
496 break;
497
498 printk(KERN_DEBUG "clearing pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx pte: %p pte phys: %lx\n",
499 pfn, pmd, __pa(pmd), pte, __pa(pte));
500 pte_clear(NULL, va, pte);
501 }
502 paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
503 paging_init();
504}
505
506/*
507 * Build a proper pagetable for the kernel mappings. Up until this
508 * point, we've been running on some set of pagetables constructed by
509 * the boot process.
510 *
511 * If we're booting on native hardware, this will be a pagetable
512 * constructed in arch/x86/kernel/head_32.S. The root of the
513 * pagetable will be swapper_pg_dir.
514 *
515 * If we're booting paravirtualized under a hypervisor, then there are
516 * more options: we may already be running PAE, and the pagetable may
517 * or may not be based in swapper_pg_dir. In any case,
518 * paravirt_pagetable_init() will set up swapper_pg_dir
519 * appropriately for the rest of the initialization to work.
520 *
521 * In general, pagetable_init() assumes that the pagetable may already
522 * be partially populated, and so it avoids stomping on any existing
523 * mappings.
524 */
525void __init early_ioremap_page_table_range_init(void)
526{
527 pgd_t *pgd_base = swapper_pg_dir;
528 unsigned long vaddr, end;
529
530 /*
531 * Fixed mappings, only the page table structure has to be
532 * created - mappings will be set by set_fixmap():
533 */
534 vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
535 end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
536 page_table_range_init(vaddr, end, pgd_base);
537 early_ioremap_reset();
538}
539
540static void __init pagetable_init(void)
541{
542 pgd_t *pgd_base = swapper_pg_dir;
543
544 permanent_kmaps_init(pgd_base);
545}
546
547#define DEFAULT_PTE_MASK ~(_PAGE_NX | _PAGE_GLOBAL)
548/* Bits supported by the hardware: */
549pteval_t __supported_pte_mask __read_mostly = DEFAULT_PTE_MASK;
550/* Bits allowed in normal kernel mappings: */
551pteval_t __default_kernel_pte_mask __read_mostly = DEFAULT_PTE_MASK;
552EXPORT_SYMBOL_GPL(__supported_pte_mask);
553/* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
554EXPORT_SYMBOL(__default_kernel_pte_mask);
555
556/* user-defined highmem size */
557static unsigned int highmem_pages = -1;
558
559/*
560 * highmem=size forces highmem to be exactly 'size' bytes.
561 * This works even on boxes that have no highmem otherwise.
562 * This also works to reduce highmem size on bigger boxes.
563 */
564static int __init parse_highmem(char *arg)
565{
566 if (!arg)
567 return -EINVAL;
568
569 highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
570 return 0;
571}
572early_param("highmem", parse_highmem);
573
574#define MSG_HIGHMEM_TOO_BIG \
575 "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
576
577#define MSG_LOWMEM_TOO_SMALL \
578 "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
579/*
580 * All of RAM fits into lowmem - but if user wants highmem
581 * artificially via the highmem=x boot parameter then create
582 * it:
583 */
584static void __init lowmem_pfn_init(void)
585{
586 /* max_low_pfn is 0, we already have early_res support */
587 max_low_pfn = max_pfn;
588
589 if (highmem_pages == -1)
590 highmem_pages = 0;
591#ifdef CONFIG_HIGHMEM
592 if (highmem_pages >= max_pfn) {
593 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
594 pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
595 highmem_pages = 0;
596 }
597 if (highmem_pages) {
598 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
599 printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
600 pages_to_mb(highmem_pages));
601 highmem_pages = 0;
602 }
603 max_low_pfn -= highmem_pages;
604 }
605#else
606 if (highmem_pages)
607 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
608#endif
609}
610
611#define MSG_HIGHMEM_TOO_SMALL \
612 "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
613
614#define MSG_HIGHMEM_TRIMMED \
615 "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
616/*
617 * We have more RAM than fits into lowmem - we try to put it into
618 * highmem, also taking the highmem=x boot parameter into account:
619 */
620static void __init highmem_pfn_init(void)
621{
622 max_low_pfn = MAXMEM_PFN;
623
624 if (highmem_pages == -1)
625 highmem_pages = max_pfn - MAXMEM_PFN;
626
627 if (highmem_pages + MAXMEM_PFN < max_pfn)
628 max_pfn = MAXMEM_PFN + highmem_pages;
629
630 if (highmem_pages + MAXMEM_PFN > max_pfn) {
631 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
632 pages_to_mb(max_pfn - MAXMEM_PFN),
633 pages_to_mb(highmem_pages));
634 highmem_pages = 0;
635 }
636#ifndef CONFIG_HIGHMEM
637 /* Maximum memory usable is what is directly addressable */
638 printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
639 if (max_pfn > MAX_NONPAE_PFN)
640 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
641 else
642 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
643 max_pfn = MAXMEM_PFN;
644#else /* !CONFIG_HIGHMEM */
645#ifndef CONFIG_HIGHMEM64G
646 if (max_pfn > MAX_NONPAE_PFN) {
647 max_pfn = MAX_NONPAE_PFN;
648 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
649 }
650#endif /* !CONFIG_HIGHMEM64G */
651#endif /* !CONFIG_HIGHMEM */
652}
653
654/*
655 * Determine low and high memory ranges:
656 */
657void __init find_low_pfn_range(void)
658{
659 /* it could update max_pfn */
660
661 if (max_pfn <= MAXMEM_PFN)
662 lowmem_pfn_init();
663 else
664 highmem_pfn_init();
665}
666
667#ifndef CONFIG_NEED_MULTIPLE_NODES
668void __init initmem_init(void)
669{
670#ifdef CONFIG_HIGHMEM
671 highstart_pfn = highend_pfn = max_pfn;
672 if (max_pfn > max_low_pfn)
673 highstart_pfn = max_low_pfn;
674 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
675 pages_to_mb(highend_pfn - highstart_pfn));
676 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
677#else
678 high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
679#endif
680
681 memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
682
683#ifdef CONFIG_FLATMEM
684 max_mapnr = IS_ENABLED(CONFIG_HIGHMEM) ? highend_pfn : max_low_pfn;
685#endif
686 __vmalloc_start_set = true;
687
688 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
689 pages_to_mb(max_low_pfn));
690
691 setup_bootmem_allocator();
692}
693#endif /* !CONFIG_NEED_MULTIPLE_NODES */
694
695void __init setup_bootmem_allocator(void)
696{
697 printk(KERN_INFO " mapped low ram: 0 - %08lx\n",
698 max_pfn_mapped<<PAGE_SHIFT);
699 printk(KERN_INFO " low ram: 0 - %08lx\n", max_low_pfn<<PAGE_SHIFT);
700}
701
702/*
703 * paging_init() sets up the page tables - note that the first 8MB are
704 * already mapped by head.S.
705 *
706 * This routines also unmaps the page at virtual kernel address 0, so
707 * that we can trap those pesky NULL-reference errors in the kernel.
708 */
709void __init paging_init(void)
710{
711 pagetable_init();
712
713 __flush_tlb_all();
714
715 kmap_init();
716
717 /*
718 * NOTE: at this point the bootmem allocator is fully available.
719 */
720 olpc_dt_build_devicetree();
721 sparse_init();
722 zone_sizes_init();
723}
724
725/*
726 * Test if the WP bit works in supervisor mode. It isn't supported on 386's
727 * and also on some strange 486's. All 586+'s are OK. This used to involve
728 * black magic jumps to work around some nasty CPU bugs, but fortunately the
729 * switch to using exceptions got rid of all that.
730 */
731static void __init test_wp_bit(void)
732{
733 char z = 0;
734
735 printk(KERN_INFO "Checking if this processor honours the WP bit even in supervisor mode...");
736
737 __set_fixmap(FIX_WP_TEST, __pa_symbol(empty_zero_page), PAGE_KERNEL_RO);
738
739 if (copy_to_kernel_nofault((char *)fix_to_virt(FIX_WP_TEST), &z, 1)) {
740 clear_fixmap(FIX_WP_TEST);
741 printk(KERN_CONT "Ok.\n");
742 return;
743 }
744
745 printk(KERN_CONT "No.\n");
746 panic("Linux doesn't support CPUs with broken WP.");
747}
748
749void __init mem_init(void)
750{
751 pci_iommu_alloc();
752
753#ifdef CONFIG_FLATMEM
754 BUG_ON(!mem_map);
755#endif
756 /*
757 * With CONFIG_DEBUG_PAGEALLOC initialization of highmem pages has to
758 * be done before memblock_free_all(). Memblock use free low memory for
759 * temporary data (see find_range_array()) and for this purpose can use
760 * pages that was already passed to the buddy allocator, hence marked as
761 * not accessible in the page tables when compiled with
762 * CONFIG_DEBUG_PAGEALLOC. Otherwise order of initialization is not
763 * important here.
764 */
765 set_highmem_pages_init();
766
767 /* this will put all low memory onto the freelists */
768 memblock_free_all();
769
770 after_bootmem = 1;
771 x86_init.hyper.init_after_bootmem();
772
773 mem_init_print_info(NULL);
774
775 /*
776 * Check boundaries twice: Some fundamental inconsistencies can
777 * be detected at build time already.
778 */
779#define __FIXADDR_TOP (-PAGE_SIZE)
780#ifdef CONFIG_HIGHMEM
781 BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE > FIXADDR_START);
782 BUILD_BUG_ON(VMALLOC_END > PKMAP_BASE);
783#endif
784#define high_memory (-128UL << 20)
785 BUILD_BUG_ON(VMALLOC_START >= VMALLOC_END);
786#undef high_memory
787#undef __FIXADDR_TOP
788
789#ifdef CONFIG_HIGHMEM
790 BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE > FIXADDR_START);
791 BUG_ON(VMALLOC_END > PKMAP_BASE);
792#endif
793 BUG_ON(VMALLOC_START >= VMALLOC_END);
794 BUG_ON((unsigned long)high_memory > VMALLOC_START);
795
796 test_wp_bit();
797}
798
799#ifdef CONFIG_MEMORY_HOTPLUG
800int arch_add_memory(int nid, u64 start, u64 size,
801 struct mhp_params *params)
802{
803 unsigned long start_pfn = start >> PAGE_SHIFT;
804 unsigned long nr_pages = size >> PAGE_SHIFT;
805 int ret;
806
807 /*
808 * The page tables were already mapped at boot so if the caller
809 * requests a different mapping type then we must change all the
810 * pages with __set_memory_prot().
811 */
812 if (params->pgprot.pgprot != PAGE_KERNEL.pgprot) {
813 ret = __set_memory_prot(start, nr_pages, params->pgprot);
814 if (ret)
815 return ret;
816 }
817
818 return __add_pages(nid, start_pfn, nr_pages, params);
819}
820
821void arch_remove_memory(int nid, u64 start, u64 size,
822 struct vmem_altmap *altmap)
823{
824 unsigned long start_pfn = start >> PAGE_SHIFT;
825 unsigned long nr_pages = size >> PAGE_SHIFT;
826
827 __remove_pages(start_pfn, nr_pages, altmap);
828}
829#endif
830
831int kernel_set_to_readonly __read_mostly;
832
833static void mark_nxdata_nx(void)
834{
835 /*
836 * When this called, init has already been executed and released,
837 * so everything past _etext should be NX.
838 */
839 unsigned long start = PFN_ALIGN(_etext);
840 /*
841 * This comes from __is_kernel_text upper limit. Also HPAGE where used:
842 */
843 unsigned long size = (((unsigned long)__init_end + HPAGE_SIZE) & HPAGE_MASK) - start;
844
845 if (__supported_pte_mask & _PAGE_NX)
846 printk(KERN_INFO "NX-protecting the kernel data: %luk\n", size >> 10);
847 set_memory_nx(start, size >> PAGE_SHIFT);
848}
849
850void mark_rodata_ro(void)
851{
852 unsigned long start = PFN_ALIGN(_text);
853 unsigned long size = (unsigned long)__end_rodata - start;
854
855 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
856 pr_info("Write protecting kernel text and read-only data: %luk\n",
857 size >> 10);
858
859 kernel_set_to_readonly = 1;
860
861#ifdef CONFIG_CPA_DEBUG
862 pr_info("Testing CPA: Reverting %lx-%lx\n", start, start + size);
863 set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
864
865 pr_info("Testing CPA: write protecting again\n");
866 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
867#endif
868 mark_nxdata_nx();
869 if (__supported_pte_mask & _PAGE_NX)
870 debug_checkwx();
871}