Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright IBM Corp. 2006
4 */
5
6#include <linux/memory_hotplug.h>
7#include <linux/memblock.h>
8#include <linux/pfn.h>
9#include <linux/mm.h>
10#include <linux/init.h>
11#include <linux/list.h>
12#include <linux/hugetlb.h>
13#include <linux/slab.h>
14#include <linux/sort.h>
15#include <asm/page-states.h>
16#include <asm/cacheflush.h>
17#include <asm/nospec-branch.h>
18#include <asm/ctlreg.h>
19#include <asm/pgalloc.h>
20#include <asm/setup.h>
21#include <asm/tlbflush.h>
22#include <asm/sections.h>
23#include <asm/set_memory.h>
24
25static DEFINE_MUTEX(vmem_mutex);
26
27static void __ref *vmem_alloc_pages(unsigned int order)
28{
29 unsigned long size = PAGE_SIZE << order;
30
31 if (slab_is_available())
32 return (void *)__get_free_pages(GFP_KERNEL, order);
33 return memblock_alloc(size, size);
34}
35
36static void vmem_free_pages(unsigned long addr, int order)
37{
38 /* We don't expect boot memory to be removed ever. */
39 if (!slab_is_available() ||
40 WARN_ON_ONCE(PageReserved(virt_to_page((void *)addr))))
41 return;
42 free_pages(addr, order);
43}
44
45void *vmem_crst_alloc(unsigned long val)
46{
47 unsigned long *table;
48
49 table = vmem_alloc_pages(CRST_ALLOC_ORDER);
50 if (!table)
51 return NULL;
52 crst_table_init(table, val);
53 __arch_set_page_dat(table, 1UL << CRST_ALLOC_ORDER);
54 return table;
55}
56
57pte_t __ref *vmem_pte_alloc(void)
58{
59 unsigned long size = PTRS_PER_PTE * sizeof(pte_t);
60 pte_t *pte;
61
62 if (slab_is_available())
63 pte = (pte_t *) page_table_alloc(&init_mm);
64 else
65 pte = (pte_t *) memblock_alloc(size, size);
66 if (!pte)
67 return NULL;
68 memset64((u64 *)pte, _PAGE_INVALID, PTRS_PER_PTE);
69 __arch_set_page_dat(pte, 1);
70 return pte;
71}
72
73static void vmem_pte_free(unsigned long *table)
74{
75 /* We don't expect boot memory to be removed ever. */
76 if (!slab_is_available() ||
77 WARN_ON_ONCE(PageReserved(virt_to_page(table))))
78 return;
79 page_table_free(&init_mm, table);
80}
81
82#define PAGE_UNUSED 0xFD
83
84/*
85 * The unused vmemmap range, which was not yet memset(PAGE_UNUSED) ranges
86 * from unused_sub_pmd_start to next PMD_SIZE boundary.
87 */
88static unsigned long unused_sub_pmd_start;
89
90static void vmemmap_flush_unused_sub_pmd(void)
91{
92 if (!unused_sub_pmd_start)
93 return;
94 memset((void *)unused_sub_pmd_start, PAGE_UNUSED,
95 ALIGN(unused_sub_pmd_start, PMD_SIZE) - unused_sub_pmd_start);
96 unused_sub_pmd_start = 0;
97}
98
99static void vmemmap_mark_sub_pmd_used(unsigned long start, unsigned long end)
100{
101 /*
102 * As we expect to add in the same granularity as we remove, it's
103 * sufficient to mark only some piece used to block the memmap page from
104 * getting removed (just in case the memmap never gets initialized,
105 * e.g., because the memory block never gets onlined).
106 */
107 memset((void *)start, 0, sizeof(struct page));
108}
109
110static void vmemmap_use_sub_pmd(unsigned long start, unsigned long end)
111{
112 /*
113 * We only optimize if the new used range directly follows the
114 * previously unused range (esp., when populating consecutive sections).
115 */
116 if (unused_sub_pmd_start == start) {
117 unused_sub_pmd_start = end;
118 if (likely(IS_ALIGNED(unused_sub_pmd_start, PMD_SIZE)))
119 unused_sub_pmd_start = 0;
120 return;
121 }
122 vmemmap_flush_unused_sub_pmd();
123 vmemmap_mark_sub_pmd_used(start, end);
124}
125
126static void vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end)
127{
128 unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
129
130 vmemmap_flush_unused_sub_pmd();
131
132 /* Could be our memmap page is filled with PAGE_UNUSED already ... */
133 vmemmap_mark_sub_pmd_used(start, end);
134
135 /* Mark the unused parts of the new memmap page PAGE_UNUSED. */
136 if (!IS_ALIGNED(start, PMD_SIZE))
137 memset((void *)page, PAGE_UNUSED, start - page);
138 /*
139 * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of
140 * consecutive sections. Remember for the last added PMD the last
141 * unused range in the populated PMD.
142 */
143 if (!IS_ALIGNED(end, PMD_SIZE))
144 unused_sub_pmd_start = end;
145}
146
147/* Returns true if the PMD is completely unused and can be freed. */
148static bool vmemmap_unuse_sub_pmd(unsigned long start, unsigned long end)
149{
150 unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
151
152 vmemmap_flush_unused_sub_pmd();
153 memset((void *)start, PAGE_UNUSED, end - start);
154 return !memchr_inv((void *)page, PAGE_UNUSED, PMD_SIZE);
155}
156
157/* __ref: we'll only call vmemmap_alloc_block() via vmemmap_populate() */
158static int __ref modify_pte_table(pmd_t *pmd, unsigned long addr,
159 unsigned long end, bool add, bool direct)
160{
161 unsigned long prot, pages = 0;
162 int ret = -ENOMEM;
163 pte_t *pte;
164
165 prot = pgprot_val(PAGE_KERNEL);
166 if (!MACHINE_HAS_NX)
167 prot &= ~_PAGE_NOEXEC;
168
169 pte = pte_offset_kernel(pmd, addr);
170 for (; addr < end; addr += PAGE_SIZE, pte++) {
171 if (!add) {
172 if (pte_none(*pte))
173 continue;
174 if (!direct)
175 vmem_free_pages((unsigned long) pfn_to_virt(pte_pfn(*pte)), 0);
176 pte_clear(&init_mm, addr, pte);
177 } else if (pte_none(*pte)) {
178 if (!direct) {
179 void *new_page = vmemmap_alloc_block(PAGE_SIZE, NUMA_NO_NODE);
180
181 if (!new_page)
182 goto out;
183 set_pte(pte, __pte(__pa(new_page) | prot));
184 } else {
185 set_pte(pte, __pte(__pa(addr) | prot));
186 }
187 } else {
188 continue;
189 }
190 pages++;
191 }
192 ret = 0;
193out:
194 if (direct)
195 update_page_count(PG_DIRECT_MAP_4K, add ? pages : -pages);
196 return ret;
197}
198
199static void try_free_pte_table(pmd_t *pmd, unsigned long start)
200{
201 pte_t *pte;
202 int i;
203
204 /* We can safely assume this is fully in 1:1 mapping & vmemmap area */
205 pte = pte_offset_kernel(pmd, start);
206 for (i = 0; i < PTRS_PER_PTE; i++, pte++) {
207 if (!pte_none(*pte))
208 return;
209 }
210 vmem_pte_free((unsigned long *) pmd_deref(*pmd));
211 pmd_clear(pmd);
212}
213
214/* __ref: we'll only call vmemmap_alloc_block() via vmemmap_populate() */
215static int __ref modify_pmd_table(pud_t *pud, unsigned long addr,
216 unsigned long end, bool add, bool direct)
217{
218 unsigned long next, prot, pages = 0;
219 int ret = -ENOMEM;
220 pmd_t *pmd;
221 pte_t *pte;
222
223 prot = pgprot_val(SEGMENT_KERNEL);
224 if (!MACHINE_HAS_NX)
225 prot &= ~_SEGMENT_ENTRY_NOEXEC;
226
227 pmd = pmd_offset(pud, addr);
228 for (; addr < end; addr = next, pmd++) {
229 next = pmd_addr_end(addr, end);
230 if (!add) {
231 if (pmd_none(*pmd))
232 continue;
233 if (pmd_large(*pmd)) {
234 if (IS_ALIGNED(addr, PMD_SIZE) &&
235 IS_ALIGNED(next, PMD_SIZE)) {
236 if (!direct)
237 vmem_free_pages(pmd_deref(*pmd), get_order(PMD_SIZE));
238 pmd_clear(pmd);
239 pages++;
240 } else if (!direct && vmemmap_unuse_sub_pmd(addr, next)) {
241 vmem_free_pages(pmd_deref(*pmd), get_order(PMD_SIZE));
242 pmd_clear(pmd);
243 }
244 continue;
245 }
246 } else if (pmd_none(*pmd)) {
247 if (IS_ALIGNED(addr, PMD_SIZE) &&
248 IS_ALIGNED(next, PMD_SIZE) &&
249 MACHINE_HAS_EDAT1 && direct &&
250 !debug_pagealloc_enabled()) {
251 set_pmd(pmd, __pmd(__pa(addr) | prot));
252 pages++;
253 continue;
254 } else if (!direct && MACHINE_HAS_EDAT1) {
255 void *new_page;
256
257 /*
258 * Use 1MB frames for vmemmap if available. We
259 * always use large frames even if they are only
260 * partially used. Otherwise we would have also
261 * page tables since vmemmap_populate gets
262 * called for each section separately.
263 */
264 new_page = vmemmap_alloc_block(PMD_SIZE, NUMA_NO_NODE);
265 if (new_page) {
266 set_pmd(pmd, __pmd(__pa(new_page) | prot));
267 if (!IS_ALIGNED(addr, PMD_SIZE) ||
268 !IS_ALIGNED(next, PMD_SIZE)) {
269 vmemmap_use_new_sub_pmd(addr, next);
270 }
271 continue;
272 }
273 }
274 pte = vmem_pte_alloc();
275 if (!pte)
276 goto out;
277 pmd_populate(&init_mm, pmd, pte);
278 } else if (pmd_large(*pmd)) {
279 if (!direct)
280 vmemmap_use_sub_pmd(addr, next);
281 continue;
282 }
283 ret = modify_pte_table(pmd, addr, next, add, direct);
284 if (ret)
285 goto out;
286 if (!add)
287 try_free_pte_table(pmd, addr & PMD_MASK);
288 }
289 ret = 0;
290out:
291 if (direct)
292 update_page_count(PG_DIRECT_MAP_1M, add ? pages : -pages);
293 return ret;
294}
295
296static void try_free_pmd_table(pud_t *pud, unsigned long start)
297{
298 pmd_t *pmd;
299 int i;
300
301 pmd = pmd_offset(pud, start);
302 for (i = 0; i < PTRS_PER_PMD; i++, pmd++)
303 if (!pmd_none(*pmd))
304 return;
305 vmem_free_pages(pud_deref(*pud), CRST_ALLOC_ORDER);
306 pud_clear(pud);
307}
308
309static int modify_pud_table(p4d_t *p4d, unsigned long addr, unsigned long end,
310 bool add, bool direct)
311{
312 unsigned long next, prot, pages = 0;
313 int ret = -ENOMEM;
314 pud_t *pud;
315 pmd_t *pmd;
316
317 prot = pgprot_val(REGION3_KERNEL);
318 if (!MACHINE_HAS_NX)
319 prot &= ~_REGION_ENTRY_NOEXEC;
320 pud = pud_offset(p4d, addr);
321 for (; addr < end; addr = next, pud++) {
322 next = pud_addr_end(addr, end);
323 if (!add) {
324 if (pud_none(*pud))
325 continue;
326 if (pud_large(*pud)) {
327 if (IS_ALIGNED(addr, PUD_SIZE) &&
328 IS_ALIGNED(next, PUD_SIZE)) {
329 pud_clear(pud);
330 pages++;
331 }
332 continue;
333 }
334 } else if (pud_none(*pud)) {
335 if (IS_ALIGNED(addr, PUD_SIZE) &&
336 IS_ALIGNED(next, PUD_SIZE) &&
337 MACHINE_HAS_EDAT2 && direct &&
338 !debug_pagealloc_enabled()) {
339 set_pud(pud, __pud(__pa(addr) | prot));
340 pages++;
341 continue;
342 }
343 pmd = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY);
344 if (!pmd)
345 goto out;
346 pud_populate(&init_mm, pud, pmd);
347 } else if (pud_large(*pud)) {
348 continue;
349 }
350 ret = modify_pmd_table(pud, addr, next, add, direct);
351 if (ret)
352 goto out;
353 if (!add)
354 try_free_pmd_table(pud, addr & PUD_MASK);
355 }
356 ret = 0;
357out:
358 if (direct)
359 update_page_count(PG_DIRECT_MAP_2G, add ? pages : -pages);
360 return ret;
361}
362
363static void try_free_pud_table(p4d_t *p4d, unsigned long start)
364{
365 pud_t *pud;
366 int i;
367
368 pud = pud_offset(p4d, start);
369 for (i = 0; i < PTRS_PER_PUD; i++, pud++) {
370 if (!pud_none(*pud))
371 return;
372 }
373 vmem_free_pages(p4d_deref(*p4d), CRST_ALLOC_ORDER);
374 p4d_clear(p4d);
375}
376
377static int modify_p4d_table(pgd_t *pgd, unsigned long addr, unsigned long end,
378 bool add, bool direct)
379{
380 unsigned long next;
381 int ret = -ENOMEM;
382 p4d_t *p4d;
383 pud_t *pud;
384
385 p4d = p4d_offset(pgd, addr);
386 for (; addr < end; addr = next, p4d++) {
387 next = p4d_addr_end(addr, end);
388 if (!add) {
389 if (p4d_none(*p4d))
390 continue;
391 } else if (p4d_none(*p4d)) {
392 pud = vmem_crst_alloc(_REGION3_ENTRY_EMPTY);
393 if (!pud)
394 goto out;
395 p4d_populate(&init_mm, p4d, pud);
396 }
397 ret = modify_pud_table(p4d, addr, next, add, direct);
398 if (ret)
399 goto out;
400 if (!add)
401 try_free_pud_table(p4d, addr & P4D_MASK);
402 }
403 ret = 0;
404out:
405 return ret;
406}
407
408static void try_free_p4d_table(pgd_t *pgd, unsigned long start)
409{
410 p4d_t *p4d;
411 int i;
412
413 p4d = p4d_offset(pgd, start);
414 for (i = 0; i < PTRS_PER_P4D; i++, p4d++) {
415 if (!p4d_none(*p4d))
416 return;
417 }
418 vmem_free_pages(pgd_deref(*pgd), CRST_ALLOC_ORDER);
419 pgd_clear(pgd);
420}
421
422static int modify_pagetable(unsigned long start, unsigned long end, bool add,
423 bool direct)
424{
425 unsigned long addr, next;
426 int ret = -ENOMEM;
427 pgd_t *pgd;
428 p4d_t *p4d;
429
430 if (WARN_ON_ONCE(!PAGE_ALIGNED(start | end)))
431 return -EINVAL;
432 /* Don't mess with any tables not fully in 1:1 mapping & vmemmap area */
433 if (WARN_ON_ONCE(end > VMALLOC_START))
434 return -EINVAL;
435 for (addr = start; addr < end; addr = next) {
436 next = pgd_addr_end(addr, end);
437 pgd = pgd_offset_k(addr);
438
439 if (!add) {
440 if (pgd_none(*pgd))
441 continue;
442 } else if (pgd_none(*pgd)) {
443 p4d = vmem_crst_alloc(_REGION2_ENTRY_EMPTY);
444 if (!p4d)
445 goto out;
446 pgd_populate(&init_mm, pgd, p4d);
447 }
448 ret = modify_p4d_table(pgd, addr, next, add, direct);
449 if (ret)
450 goto out;
451 if (!add)
452 try_free_p4d_table(pgd, addr & PGDIR_MASK);
453 }
454 ret = 0;
455out:
456 if (!add)
457 flush_tlb_kernel_range(start, end);
458 return ret;
459}
460
461static int add_pagetable(unsigned long start, unsigned long end, bool direct)
462{
463 return modify_pagetable(start, end, true, direct);
464}
465
466static int remove_pagetable(unsigned long start, unsigned long end, bool direct)
467{
468 return modify_pagetable(start, end, false, direct);
469}
470
471/*
472 * Add a physical memory range to the 1:1 mapping.
473 */
474static int vmem_add_range(unsigned long start, unsigned long size)
475{
476 start = (unsigned long)__va(start);
477 return add_pagetable(start, start + size, true);
478}
479
480/*
481 * Remove a physical memory range from the 1:1 mapping.
482 */
483static void vmem_remove_range(unsigned long start, unsigned long size)
484{
485 start = (unsigned long)__va(start);
486 remove_pagetable(start, start + size, true);
487}
488
489/*
490 * Add a backed mem_map array to the virtual mem_map array.
491 */
492int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
493 struct vmem_altmap *altmap)
494{
495 int ret;
496
497 mutex_lock(&vmem_mutex);
498 /* We don't care about the node, just use NUMA_NO_NODE on allocations */
499 ret = add_pagetable(start, end, false);
500 if (ret)
501 remove_pagetable(start, end, false);
502 mutex_unlock(&vmem_mutex);
503 return ret;
504}
505
506#ifdef CONFIG_MEMORY_HOTPLUG
507
508void vmemmap_free(unsigned long start, unsigned long end,
509 struct vmem_altmap *altmap)
510{
511 mutex_lock(&vmem_mutex);
512 remove_pagetable(start, end, false);
513 mutex_unlock(&vmem_mutex);
514}
515
516#endif
517
518void vmem_remove_mapping(unsigned long start, unsigned long size)
519{
520 mutex_lock(&vmem_mutex);
521 vmem_remove_range(start, size);
522 mutex_unlock(&vmem_mutex);
523}
524
525struct range arch_get_mappable_range(void)
526{
527 struct range mhp_range;
528
529 mhp_range.start = 0;
530 mhp_range.end = max_mappable - 1;
531 return mhp_range;
532}
533
534int vmem_add_mapping(unsigned long start, unsigned long size)
535{
536 struct range range = arch_get_mappable_range();
537 int ret;
538
539 if (start < range.start ||
540 start + size > range.end + 1 ||
541 start + size < start)
542 return -ERANGE;
543
544 mutex_lock(&vmem_mutex);
545 ret = vmem_add_range(start, size);
546 if (ret)
547 vmem_remove_range(start, size);
548 mutex_unlock(&vmem_mutex);
549 return ret;
550}
551
552/*
553 * Allocate new or return existing page-table entry, but do not map it
554 * to any physical address. If missing, allocate segment- and region-
555 * table entries along. Meeting a large segment- or region-table entry
556 * while traversing is an error, since the function is expected to be
557 * called against virtual regions reserved for 4KB mappings only.
558 */
559pte_t *vmem_get_alloc_pte(unsigned long addr, bool alloc)
560{
561 pte_t *ptep = NULL;
562 pgd_t *pgd;
563 p4d_t *p4d;
564 pud_t *pud;
565 pmd_t *pmd;
566 pte_t *pte;
567
568 pgd = pgd_offset_k(addr);
569 if (pgd_none(*pgd)) {
570 if (!alloc)
571 goto out;
572 p4d = vmem_crst_alloc(_REGION2_ENTRY_EMPTY);
573 if (!p4d)
574 goto out;
575 pgd_populate(&init_mm, pgd, p4d);
576 }
577 p4d = p4d_offset(pgd, addr);
578 if (p4d_none(*p4d)) {
579 if (!alloc)
580 goto out;
581 pud = vmem_crst_alloc(_REGION3_ENTRY_EMPTY);
582 if (!pud)
583 goto out;
584 p4d_populate(&init_mm, p4d, pud);
585 }
586 pud = pud_offset(p4d, addr);
587 if (pud_none(*pud)) {
588 if (!alloc)
589 goto out;
590 pmd = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY);
591 if (!pmd)
592 goto out;
593 pud_populate(&init_mm, pud, pmd);
594 } else if (WARN_ON_ONCE(pud_large(*pud))) {
595 goto out;
596 }
597 pmd = pmd_offset(pud, addr);
598 if (pmd_none(*pmd)) {
599 if (!alloc)
600 goto out;
601 pte = vmem_pte_alloc();
602 if (!pte)
603 goto out;
604 pmd_populate(&init_mm, pmd, pte);
605 } else if (WARN_ON_ONCE(pmd_large(*pmd))) {
606 goto out;
607 }
608 ptep = pte_offset_kernel(pmd, addr);
609out:
610 return ptep;
611}
612
613int __vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot, bool alloc)
614{
615 pte_t *ptep, pte;
616
617 if (!IS_ALIGNED(addr, PAGE_SIZE))
618 return -EINVAL;
619 ptep = vmem_get_alloc_pte(addr, alloc);
620 if (!ptep)
621 return -ENOMEM;
622 __ptep_ipte(addr, ptep, 0, 0, IPTE_GLOBAL);
623 pte = mk_pte_phys(phys, prot);
624 set_pte(ptep, pte);
625 return 0;
626}
627
628int vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot)
629{
630 int rc;
631
632 mutex_lock(&vmem_mutex);
633 rc = __vmem_map_4k_page(addr, phys, prot, true);
634 mutex_unlock(&vmem_mutex);
635 return rc;
636}
637
638void vmem_unmap_4k_page(unsigned long addr)
639{
640 pte_t *ptep;
641
642 mutex_lock(&vmem_mutex);
643 ptep = virt_to_kpte(addr);
644 __ptep_ipte(addr, ptep, 0, 0, IPTE_GLOBAL);
645 pte_clear(&init_mm, addr, ptep);
646 mutex_unlock(&vmem_mutex);
647}
648
649void __init vmem_map_init(void)
650{
651 __set_memory_rox(_stext, _etext);
652 __set_memory_ro(_etext, __end_rodata);
653 __set_memory_rox(_sinittext, _einittext);
654 __set_memory_rox(__stext_amode31, __etext_amode31);
655 /*
656 * If the BEAR-enhancement facility is not installed the first
657 * prefix page is used to return to the previous context with
658 * an LPSWE instruction and therefore must be executable.
659 */
660 if (!static_key_enabled(&cpu_has_bear))
661 set_memory_x(0, 1);
662 if (debug_pagealloc_enabled()) {
663 /*
664 * Use RELOC_HIDE() as long as __va(0) translates to NULL,
665 * since performing pointer arithmetic on a NULL pointer
666 * has undefined behavior and generates compiler warnings.
667 */
668 __set_memory_4k(__va(0), RELOC_HIDE(__va(0), ident_map_size));
669 }
670 if (MACHINE_HAS_NX)
671 system_ctl_set_bit(0, CR0_INSTRUCTION_EXEC_PROTECTION_BIT);
672 pr_info("Write protected kernel read-only data: %luk\n",
673 (unsigned long)(__end_rodata - _stext) >> 10);
674}
1/*
2 * Copyright IBM Corp. 2006
3 * Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
4 */
5
6#include <linux/bootmem.h>
7#include <linux/pfn.h>
8#include <linux/mm.h>
9#include <linux/module.h>
10#include <linux/list.h>
11#include <linux/hugetlb.h>
12#include <linux/slab.h>
13#include <asm/pgalloc.h>
14#include <asm/pgtable.h>
15#include <asm/setup.h>
16#include <asm/tlbflush.h>
17#include <asm/sections.h>
18
19static DEFINE_MUTEX(vmem_mutex);
20
21struct memory_segment {
22 struct list_head list;
23 unsigned long start;
24 unsigned long size;
25};
26
27static LIST_HEAD(mem_segs);
28
29static void __ref *vmem_alloc_pages(unsigned int order)
30{
31 if (slab_is_available())
32 return (void *)__get_free_pages(GFP_KERNEL, order);
33 return alloc_bootmem_pages((1 << order) * PAGE_SIZE);
34}
35
36static inline pud_t *vmem_pud_alloc(void)
37{
38 pud_t *pud = NULL;
39
40#ifdef CONFIG_64BIT
41 pud = vmem_alloc_pages(2);
42 if (!pud)
43 return NULL;
44 clear_table((unsigned long *) pud, _REGION3_ENTRY_EMPTY, PAGE_SIZE * 4);
45#endif
46 return pud;
47}
48
49static inline pmd_t *vmem_pmd_alloc(void)
50{
51 pmd_t *pmd = NULL;
52
53#ifdef CONFIG_64BIT
54 pmd = vmem_alloc_pages(2);
55 if (!pmd)
56 return NULL;
57 clear_table((unsigned long *) pmd, _SEGMENT_ENTRY_EMPTY, PAGE_SIZE * 4);
58#endif
59 return pmd;
60}
61
62static pte_t __ref *vmem_pte_alloc(unsigned long address)
63{
64 pte_t *pte;
65
66 if (slab_is_available())
67 pte = (pte_t *) page_table_alloc(&init_mm, address);
68 else
69 pte = alloc_bootmem(PTRS_PER_PTE * sizeof(pte_t));
70 if (!pte)
71 return NULL;
72 clear_table((unsigned long *) pte, _PAGE_INVALID,
73 PTRS_PER_PTE * sizeof(pte_t));
74 return pte;
75}
76
77/*
78 * Add a physical memory range to the 1:1 mapping.
79 */
80static int vmem_add_mem(unsigned long start, unsigned long size, int ro)
81{
82 unsigned long end = start + size;
83 unsigned long address = start;
84 pgd_t *pg_dir;
85 pud_t *pu_dir;
86 pmd_t *pm_dir;
87 pte_t *pt_dir;
88 int ret = -ENOMEM;
89
90 while (address < end) {
91 pg_dir = pgd_offset_k(address);
92 if (pgd_none(*pg_dir)) {
93 pu_dir = vmem_pud_alloc();
94 if (!pu_dir)
95 goto out;
96 pgd_populate(&init_mm, pg_dir, pu_dir);
97 }
98 pu_dir = pud_offset(pg_dir, address);
99#if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
100 if (MACHINE_HAS_EDAT2 && pud_none(*pu_dir) && address &&
101 !(address & ~PUD_MASK) && (address + PUD_SIZE <= end)) {
102 pud_val(*pu_dir) = __pa(address) |
103 _REGION_ENTRY_TYPE_R3 | _REGION3_ENTRY_LARGE |
104 (ro ? _REGION_ENTRY_PROTECT : 0);
105 address += PUD_SIZE;
106 continue;
107 }
108#endif
109 if (pud_none(*pu_dir)) {
110 pm_dir = vmem_pmd_alloc();
111 if (!pm_dir)
112 goto out;
113 pud_populate(&init_mm, pu_dir, pm_dir);
114 }
115 pm_dir = pmd_offset(pu_dir, address);
116#if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
117 if (MACHINE_HAS_EDAT1 && pmd_none(*pm_dir) && address &&
118 !(address & ~PMD_MASK) && (address + PMD_SIZE <= end)) {
119 pmd_val(*pm_dir) = __pa(address) |
120 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
121 _SEGMENT_ENTRY_YOUNG |
122 (ro ? _SEGMENT_ENTRY_PROTECT : 0);
123 address += PMD_SIZE;
124 continue;
125 }
126#endif
127 if (pmd_none(*pm_dir)) {
128 pt_dir = vmem_pte_alloc(address);
129 if (!pt_dir)
130 goto out;
131 pmd_populate(&init_mm, pm_dir, pt_dir);
132 }
133
134 pt_dir = pte_offset_kernel(pm_dir, address);
135 pte_val(*pt_dir) = __pa(address) |
136 pgprot_val(ro ? PAGE_KERNEL_RO : PAGE_KERNEL);
137 address += PAGE_SIZE;
138 }
139 ret = 0;
140out:
141 return ret;
142}
143
144/*
145 * Remove a physical memory range from the 1:1 mapping.
146 * Currently only invalidates page table entries.
147 */
148static void vmem_remove_range(unsigned long start, unsigned long size)
149{
150 unsigned long end = start + size;
151 unsigned long address = start;
152 pgd_t *pg_dir;
153 pud_t *pu_dir;
154 pmd_t *pm_dir;
155 pte_t *pt_dir;
156 pte_t pte;
157
158 pte_val(pte) = _PAGE_INVALID;
159 while (address < end) {
160 pg_dir = pgd_offset_k(address);
161 if (pgd_none(*pg_dir)) {
162 address += PGDIR_SIZE;
163 continue;
164 }
165 pu_dir = pud_offset(pg_dir, address);
166 if (pud_none(*pu_dir)) {
167 address += PUD_SIZE;
168 continue;
169 }
170 if (pud_large(*pu_dir)) {
171 pud_clear(pu_dir);
172 address += PUD_SIZE;
173 continue;
174 }
175 pm_dir = pmd_offset(pu_dir, address);
176 if (pmd_none(*pm_dir)) {
177 address += PMD_SIZE;
178 continue;
179 }
180 if (pmd_large(*pm_dir)) {
181 pmd_clear(pm_dir);
182 address += PMD_SIZE;
183 continue;
184 }
185 pt_dir = pte_offset_kernel(pm_dir, address);
186 *pt_dir = pte;
187 address += PAGE_SIZE;
188 }
189 flush_tlb_kernel_range(start, end);
190}
191
192/*
193 * Add a backed mem_map array to the virtual mem_map array.
194 */
195int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
196{
197 unsigned long address = start;
198 pgd_t *pg_dir;
199 pud_t *pu_dir;
200 pmd_t *pm_dir;
201 pte_t *pt_dir;
202 int ret = -ENOMEM;
203
204 for (address = start; address < end;) {
205 pg_dir = pgd_offset_k(address);
206 if (pgd_none(*pg_dir)) {
207 pu_dir = vmem_pud_alloc();
208 if (!pu_dir)
209 goto out;
210 pgd_populate(&init_mm, pg_dir, pu_dir);
211 }
212
213 pu_dir = pud_offset(pg_dir, address);
214 if (pud_none(*pu_dir)) {
215 pm_dir = vmem_pmd_alloc();
216 if (!pm_dir)
217 goto out;
218 pud_populate(&init_mm, pu_dir, pm_dir);
219 }
220
221 pm_dir = pmd_offset(pu_dir, address);
222 if (pmd_none(*pm_dir)) {
223#ifdef CONFIG_64BIT
224 /* Use 1MB frames for vmemmap if available. We always
225 * use large frames even if they are only partially
226 * used.
227 * Otherwise we would have also page tables since
228 * vmemmap_populate gets called for each section
229 * separately. */
230 if (MACHINE_HAS_EDAT1) {
231 void *new_page;
232
233 new_page = vmemmap_alloc_block(PMD_SIZE, node);
234 if (!new_page)
235 goto out;
236 pmd_val(*pm_dir) = __pa(new_page) |
237 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
238 _SEGMENT_ENTRY_CO;
239 address = (address + PMD_SIZE) & PMD_MASK;
240 continue;
241 }
242#endif
243 pt_dir = vmem_pte_alloc(address);
244 if (!pt_dir)
245 goto out;
246 pmd_populate(&init_mm, pm_dir, pt_dir);
247 } else if (pmd_large(*pm_dir)) {
248 address = (address + PMD_SIZE) & PMD_MASK;
249 continue;
250 }
251
252 pt_dir = pte_offset_kernel(pm_dir, address);
253 if (pte_none(*pt_dir)) {
254 unsigned long new_page;
255
256 new_page =__pa(vmem_alloc_pages(0));
257 if (!new_page)
258 goto out;
259 pte_val(*pt_dir) =
260 __pa(new_page) | pgprot_val(PAGE_KERNEL);
261 }
262 address += PAGE_SIZE;
263 }
264 memset((void *)start, 0, end - start);
265 ret = 0;
266out:
267 return ret;
268}
269
270void vmemmap_free(unsigned long start, unsigned long end)
271{
272}
273
274/*
275 * Add memory segment to the segment list if it doesn't overlap with
276 * an already present segment.
277 */
278static int insert_memory_segment(struct memory_segment *seg)
279{
280 struct memory_segment *tmp;
281
282 if (seg->start + seg->size > VMEM_MAX_PHYS ||
283 seg->start + seg->size < seg->start)
284 return -ERANGE;
285
286 list_for_each_entry(tmp, &mem_segs, list) {
287 if (seg->start >= tmp->start + tmp->size)
288 continue;
289 if (seg->start + seg->size <= tmp->start)
290 continue;
291 return -ENOSPC;
292 }
293 list_add(&seg->list, &mem_segs);
294 return 0;
295}
296
297/*
298 * Remove memory segment from the segment list.
299 */
300static void remove_memory_segment(struct memory_segment *seg)
301{
302 list_del(&seg->list);
303}
304
305static void __remove_shared_memory(struct memory_segment *seg)
306{
307 remove_memory_segment(seg);
308 vmem_remove_range(seg->start, seg->size);
309}
310
311int vmem_remove_mapping(unsigned long start, unsigned long size)
312{
313 struct memory_segment *seg;
314 int ret;
315
316 mutex_lock(&vmem_mutex);
317
318 ret = -ENOENT;
319 list_for_each_entry(seg, &mem_segs, list) {
320 if (seg->start == start && seg->size == size)
321 break;
322 }
323
324 if (seg->start != start || seg->size != size)
325 goto out;
326
327 ret = 0;
328 __remove_shared_memory(seg);
329 kfree(seg);
330out:
331 mutex_unlock(&vmem_mutex);
332 return ret;
333}
334
335int vmem_add_mapping(unsigned long start, unsigned long size)
336{
337 struct memory_segment *seg;
338 int ret;
339
340 mutex_lock(&vmem_mutex);
341 ret = -ENOMEM;
342 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
343 if (!seg)
344 goto out;
345 seg->start = start;
346 seg->size = size;
347
348 ret = insert_memory_segment(seg);
349 if (ret)
350 goto out_free;
351
352 ret = vmem_add_mem(start, size, 0);
353 if (ret)
354 goto out_remove;
355 goto out;
356
357out_remove:
358 __remove_shared_memory(seg);
359out_free:
360 kfree(seg);
361out:
362 mutex_unlock(&vmem_mutex);
363 return ret;
364}
365
366/*
367 * map whole physical memory to virtual memory (identity mapping)
368 * we reserve enough space in the vmalloc area for vmemmap to hotplug
369 * additional memory segments.
370 */
371void __init vmem_map_init(void)
372{
373 unsigned long ro_start, ro_end;
374 unsigned long start, end;
375 int i;
376
377 ro_start = PFN_ALIGN((unsigned long)&_stext);
378 ro_end = (unsigned long)&_eshared & PAGE_MASK;
379 for (i = 0; i < MEMORY_CHUNKS; i++) {
380 if (!memory_chunk[i].size)
381 continue;
382 start = memory_chunk[i].addr;
383 end = memory_chunk[i].addr + memory_chunk[i].size;
384 if (start >= ro_end || end <= ro_start)
385 vmem_add_mem(start, end - start, 0);
386 else if (start >= ro_start && end <= ro_end)
387 vmem_add_mem(start, end - start, 1);
388 else if (start >= ro_start) {
389 vmem_add_mem(start, ro_end - start, 1);
390 vmem_add_mem(ro_end, end - ro_end, 0);
391 } else if (end < ro_end) {
392 vmem_add_mem(start, ro_start - start, 0);
393 vmem_add_mem(ro_start, end - ro_start, 1);
394 } else {
395 vmem_add_mem(start, ro_start - start, 0);
396 vmem_add_mem(ro_start, ro_end - ro_start, 1);
397 vmem_add_mem(ro_end, end - ro_end, 0);
398 }
399 }
400}
401
402/*
403 * Convert memory chunk array to a memory segment list so there is a single
404 * list that contains both r/w memory and shared memory segments.
405 */
406static int __init vmem_convert_memory_chunk(void)
407{
408 struct memory_segment *seg;
409 int i;
410
411 mutex_lock(&vmem_mutex);
412 for (i = 0; i < MEMORY_CHUNKS; i++) {
413 if (!memory_chunk[i].size)
414 continue;
415 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
416 if (!seg)
417 panic("Out of memory...\n");
418 seg->start = memory_chunk[i].addr;
419 seg->size = memory_chunk[i].size;
420 insert_memory_segment(seg);
421 }
422 mutex_unlock(&vmem_mutex);
423 return 0;
424}
425
426core_initcall(vmem_convert_memory_chunk);