Commit | Line | Data |
---|---|---|
14cf11af PM |
1 | /* |
2 | * PowerPC version | |
3 | * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org) | |
4 | * | |
5 | * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au) | |
6 | * and Cort Dougan (PReP) (cort@cs.nmt.edu) | |
7 | * Copyright (C) 1996 Paul Mackerras | |
14cf11af PM |
8 | * |
9 | * Derived from "arch/i386/mm/init.c" | |
10 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
11 | * | |
12 | * Dave Engebretsen <engebret@us.ibm.com> | |
13 | * Rework for PPC64 port. | |
14 | * | |
15 | * This program is free software; you can redistribute it and/or | |
16 | * modify it under the terms of the GNU General Public License | |
17 | * as published by the Free Software Foundation; either version | |
18 | * 2 of the License, or (at your option) any later version. | |
19 | * | |
20 | */ | |
21 | ||
cec08e7a BH |
22 | #undef DEBUG |
23 | ||
14cf11af PM |
24 | #include <linux/signal.h> |
25 | #include <linux/sched.h> | |
26 | #include <linux/kernel.h> | |
27 | #include <linux/errno.h> | |
28 | #include <linux/string.h> | |
29 | #include <linux/types.h> | |
30 | #include <linux/mman.h> | |
31 | #include <linux/mm.h> | |
32 | #include <linux/swap.h> | |
33 | #include <linux/stddef.h> | |
34 | #include <linux/vmalloc.h> | |
35 | #include <linux/init.h> | |
36 | #include <linux/delay.h> | |
14cf11af PM |
37 | #include <linux/highmem.h> |
38 | #include <linux/idr.h> | |
39 | #include <linux/nodemask.h> | |
40 | #include <linux/module.h> | |
c9cf5528 | 41 | #include <linux/poison.h> |
95f72d1e | 42 | #include <linux/memblock.h> |
a4fe3ce7 | 43 | #include <linux/hugetlb.h> |
5a0e3ad6 | 44 | #include <linux/slab.h> |
14cf11af PM |
45 | |
46 | #include <asm/pgalloc.h> | |
47 | #include <asm/page.h> | |
48 | #include <asm/prom.h> | |
14cf11af PM |
49 | #include <asm/rtas.h> |
50 | #include <asm/io.h> | |
51 | #include <asm/mmu_context.h> | |
52 | #include <asm/pgtable.h> | |
53 | #include <asm/mmu.h> | |
54 | #include <asm/uaccess.h> | |
55 | #include <asm/smp.h> | |
56 | #include <asm/machdep.h> | |
57 | #include <asm/tlb.h> | |
58 | #include <asm/eeh.h> | |
59 | #include <asm/processor.h> | |
60 | #include <asm/mmzone.h> | |
61 | #include <asm/cputable.h> | |
14cf11af | 62 | #include <asm/sections.h> |
14cf11af | 63 | #include <asm/iommu.h> |
14cf11af | 64 | #include <asm/vdso.h> |
800fc3ee DG |
65 | |
66 | #include "mmu_decl.h" | |
14cf11af | 67 | |
94491685 | 68 | #ifdef CONFIG_PPC_STD_MMU_64 |
14cf11af PM |
69 | #if PGTABLE_RANGE > USER_VSID_RANGE |
70 | #warning Limited user VSID range means pagetable space is wasted | |
71 | #endif | |
72 | ||
73 | #if (TASK_SIZE_USER64 < PGTABLE_RANGE) && (TASK_SIZE_USER64 < USER_VSID_RANGE) | |
74 | #warning TASK_SIZE is smaller than it needs to be. | |
75 | #endif | |
94491685 | 76 | #endif /* CONFIG_PPC_STD_MMU_64 */ |
14cf11af | 77 | |
37dd2bad | 78 | phys_addr_t memstart_addr = ~0; |
79c3095f | 79 | EXPORT_SYMBOL_GPL(memstart_addr); |
37dd2bad | 80 | phys_addr_t kernstart_addr; |
79c3095f | 81 | EXPORT_SYMBOL_GPL(kernstart_addr); |
d7917ba7 | 82 | |
51cc5068 | 83 | static void pgd_ctor(void *addr) |
14cf11af | 84 | { |
51cc5068 AD |
85 | memset(addr, 0, PGD_TABLE_SIZE); |
86 | } | |
87 | ||
368ced78 AK |
88 | static void pud_ctor(void *addr) |
89 | { | |
90 | memset(addr, 0, PUD_TABLE_SIZE); | |
91 | } | |
92 | ||
51cc5068 AD |
93 | static void pmd_ctor(void *addr) |
94 | { | |
95 | memset(addr, 0, PMD_TABLE_SIZE); | |
14cf11af PM |
96 | } |
97 | ||
a0668cdc DG |
98 | struct kmem_cache *pgtable_cache[MAX_PGTABLE_INDEX_SIZE]; |
99 | ||
100 | /* | |
101 | * Create a kmem_cache() for pagetables. This is not used for PTE | |
102 | * pages - they're linked to struct page, come from the normal free | |
103 | * pages pool and have a different entry size (see real_pte_t) to | |
104 | * everything else. Caches created by this function are used for all | |
105 | * the higher level pagetables, and for hugepage pagetables. | |
106 | */ | |
107 | void pgtable_cache_add(unsigned shift, void (*ctor)(void *)) | |
108 | { | |
109 | char *name; | |
110 | unsigned long table_size = sizeof(void *) << shift; | |
111 | unsigned long align = table_size; | |
112 | ||
113 | /* When batching pgtable pointers for RCU freeing, we store | |
114 | * the index size in the low bits. Table alignment must be | |
a4fe3ce7 DG |
115 | * big enough to fit it. |
116 | * | |
117 | * Likewise, hugeapge pagetable pointers contain a (different) | |
118 | * shift value in the low bits. All tables must be aligned so | |
119 | * as to leave enough 0 bits in the address to contain it. */ | |
120 | unsigned long minalign = max(MAX_PGTABLE_INDEX_SIZE + 1, | |
121 | HUGEPD_SHIFT_MASK + 1); | |
a0668cdc DG |
122 | struct kmem_cache *new; |
123 | ||
124 | /* It would be nice if this was a BUILD_BUG_ON(), but at the | |
125 | * moment, gcc doesn't seem to recognize is_power_of_2 as a | |
126 | * constant expression, so so much for that. */ | |
127 | BUG_ON(!is_power_of_2(minalign)); | |
128 | BUG_ON((shift < 1) || (shift > MAX_PGTABLE_INDEX_SIZE)); | |
129 | ||
130 | if (PGT_CACHE(shift)) | |
131 | return; /* Already have a cache of this size */ | |
132 | ||
133 | align = max_t(unsigned long, align, minalign); | |
134 | name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift); | |
135 | new = kmem_cache_create(name, table_size, align, 0, ctor); | |
e77553cb | 136 | kfree(name); |
cf9427b8 | 137 | pgtable_cache[shift - 1] = new; |
a0668cdc DG |
138 | pr_debug("Allocated pgtable cache for order %d\n", shift); |
139 | } | |
140 | ||
14cf11af PM |
141 | |
142 | void pgtable_cache_init(void) | |
143 | { | |
a0668cdc | 144 | pgtable_cache_add(PGD_INDEX_SIZE, pgd_ctor); |
f940f528 | 145 | pgtable_cache_add(PMD_CACHE_INDEX, pmd_ctor); |
368ced78 AK |
146 | /* |
147 | * In all current configs, when the PUD index exists it's the | |
148 | * same size as either the pgd or pmd index except with THP enabled | |
149 | * on book3s 64 | |
150 | */ | |
151 | if (PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE)) | |
152 | pgtable_cache_add(PUD_INDEX_SIZE, pud_ctor); | |
153 | ||
f940f528 | 154 | if (!PGT_CACHE(PGD_INDEX_SIZE) || !PGT_CACHE(PMD_CACHE_INDEX)) |
a0668cdc | 155 | panic("Couldn't allocate pgtable caches"); |
368ced78 AK |
156 | if (PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE)) |
157 | panic("Couldn't allocate pud pgtable caches"); | |
14cf11af | 158 | } |
d29eff7b AW |
159 | |
160 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
161 | /* | |
162 | * Given an address within the vmemmap, determine the pfn of the page that | |
163 | * represents the start of the section it is within. Note that we have to | |
164 | * do this by hand as the proffered address may not be correctly aligned. | |
165 | * Subtraction of non-aligned pointers produces undefined results. | |
166 | */ | |
09de9ff8 | 167 | static unsigned long __meminit vmemmap_section_start(unsigned long page) |
d29eff7b AW |
168 | { |
169 | unsigned long offset = page - ((unsigned long)(vmemmap)); | |
170 | ||
171 | /* Return the pfn of the start of the section. */ | |
172 | return (offset / sizeof(struct page)) & PAGE_SECTION_MASK; | |
173 | } | |
174 | ||
175 | /* | |
176 | * Check if this vmemmap page is already initialised. If any section | |
177 | * which overlaps this vmemmap page is initialised then this page is | |
178 | * initialised already. | |
179 | */ | |
09de9ff8 | 180 | static int __meminit vmemmap_populated(unsigned long start, int page_size) |
d29eff7b AW |
181 | { |
182 | unsigned long end = start + page_size; | |
16a05bff | 183 | start = (unsigned long)(pfn_to_page(vmemmap_section_start(start))); |
d29eff7b AW |
184 | |
185 | for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page))) | |
16a05bff | 186 | if (pfn_valid(page_to_pfn((struct page *)start))) |
d29eff7b AW |
187 | return 1; |
188 | ||
189 | return 0; | |
190 | } | |
191 | ||
32a74949 BH |
192 | /* On hash-based CPUs, the vmemmap is bolted in the hash table. |
193 | * | |
194 | * On Book3E CPUs, the vmemmap is currently mapped in the top half of | |
195 | * the vmalloc space using normal page tables, though the size of | |
196 | * pages encoded in the PTEs can be different | |
197 | */ | |
198 | ||
199 | #ifdef CONFIG_PPC_BOOK3E | |
1dace6c6 DG |
200 | static int __meminit vmemmap_create_mapping(unsigned long start, |
201 | unsigned long page_size, | |
202 | unsigned long phys) | |
32a74949 BH |
203 | { |
204 | /* Create a PTE encoding without page size */ | |
205 | unsigned long i, flags = _PAGE_PRESENT | _PAGE_ACCESSED | | |
206 | _PAGE_KERNEL_RW; | |
207 | ||
208 | /* PTEs only contain page size encodings up to 32M */ | |
209 | BUG_ON(mmu_psize_defs[mmu_vmemmap_psize].enc > 0xf); | |
210 | ||
211 | /* Encode the size in the PTE */ | |
212 | flags |= mmu_psize_defs[mmu_vmemmap_psize].enc << 8; | |
213 | ||
214 | /* For each PTE for that area, map things. Note that we don't | |
215 | * increment phys because all PTEs are of the large size and | |
216 | * thus must have the low bits clear | |
217 | */ | |
218 | for (i = 0; i < page_size; i += PAGE_SIZE) | |
219 | BUG_ON(map_kernel_page(start + i, phys, flags)); | |
1dace6c6 DG |
220 | |
221 | return 0; | |
32a74949 | 222 | } |
ed5694a8 LZ |
223 | |
224 | #ifdef CONFIG_MEMORY_HOTPLUG | |
225 | static void vmemmap_remove_mapping(unsigned long start, | |
226 | unsigned long page_size) | |
227 | { | |
228 | } | |
229 | #endif | |
32a74949 | 230 | #else /* CONFIG_PPC_BOOK3E */ |
1dace6c6 DG |
231 | static int __meminit vmemmap_create_mapping(unsigned long start, |
232 | unsigned long page_size, | |
233 | unsigned long phys) | |
32a74949 | 234 | { |
1dace6c6 DG |
235 | int rc = htab_bolt_mapping(start, start + page_size, phys, |
236 | pgprot_val(PAGE_KERNEL), | |
237 | mmu_vmemmap_psize, mmu_kernel_ssize); | |
238 | if (rc < 0) { | |
239 | int rc2 = htab_remove_mapping(start, start + page_size, | |
240 | mmu_vmemmap_psize, | |
241 | mmu_kernel_ssize); | |
242 | BUG_ON(rc2 && (rc2 != -ENOENT)); | |
243 | } | |
244 | return rc; | |
32a74949 | 245 | } |
ed5694a8 LZ |
246 | |
247 | #ifdef CONFIG_MEMORY_HOTPLUG | |
ed5694a8 LZ |
248 | static void vmemmap_remove_mapping(unsigned long start, |
249 | unsigned long page_size) | |
250 | { | |
27828f98 DG |
251 | int rc = htab_remove_mapping(start, start + page_size, |
252 | mmu_vmemmap_psize, | |
253 | mmu_kernel_ssize); | |
254 | BUG_ON((rc < 0) && (rc != -ENOENT)); | |
255 | WARN_ON(rc == -ENOENT); | |
ed5694a8 LZ |
256 | } |
257 | #endif | |
258 | ||
32a74949 BH |
259 | #endif /* CONFIG_PPC_BOOK3E */ |
260 | ||
91eea67c | 261 | struct vmemmap_backing *vmemmap_list; |
bd8cb03d LZ |
262 | static struct vmemmap_backing *next; |
263 | static int num_left; | |
264 | static int num_freed; | |
91eea67c MN |
265 | |
266 | static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node) | |
267 | { | |
bd8cb03d LZ |
268 | struct vmemmap_backing *vmem_back; |
269 | /* get from freed entries first */ | |
270 | if (num_freed) { | |
271 | num_freed--; | |
272 | vmem_back = next; | |
273 | next = next->list; | |
274 | ||
275 | return vmem_back; | |
276 | } | |
91eea67c MN |
277 | |
278 | /* allocate a page when required and hand out chunks */ | |
bd8cb03d | 279 | if (!num_left) { |
91eea67c MN |
280 | next = vmemmap_alloc_block(PAGE_SIZE, node); |
281 | if (unlikely(!next)) { | |
282 | WARN_ON(1); | |
283 | return NULL; | |
284 | } | |
285 | num_left = PAGE_SIZE / sizeof(struct vmemmap_backing); | |
286 | } | |
287 | ||
288 | num_left--; | |
289 | ||
290 | return next++; | |
291 | } | |
292 | ||
293 | static __meminit void vmemmap_list_populate(unsigned long phys, | |
294 | unsigned long start, | |
295 | int node) | |
296 | { | |
297 | struct vmemmap_backing *vmem_back; | |
298 | ||
299 | vmem_back = vmemmap_list_alloc(node); | |
300 | if (unlikely(!vmem_back)) { | |
301 | WARN_ON(1); | |
302 | return; | |
303 | } | |
304 | ||
305 | vmem_back->phys = phys; | |
306 | vmem_back->virt_addr = start; | |
307 | vmem_back->list = vmemmap_list; | |
308 | ||
309 | vmemmap_list = vmem_back; | |
310 | } | |
311 | ||
71b0bfe4 LZ |
312 | int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node) |
313 | { | |
314 | unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; | |
315 | ||
316 | /* Align to the page size of the linear mapping. */ | |
317 | start = _ALIGN_DOWN(start, page_size); | |
318 | ||
319 | pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node); | |
320 | ||
321 | for (; start < end; start += page_size) { | |
322 | void *p; | |
1dace6c6 | 323 | int rc; |
71b0bfe4 LZ |
324 | |
325 | if (vmemmap_populated(start, page_size)) | |
326 | continue; | |
327 | ||
328 | p = vmemmap_alloc_block(page_size, node); | |
329 | if (!p) | |
330 | return -ENOMEM; | |
331 | ||
332 | vmemmap_list_populate(__pa(p), start, node); | |
333 | ||
334 | pr_debug(" * %016lx..%016lx allocated at %p\n", | |
335 | start, start + page_size, p); | |
336 | ||
1dace6c6 DG |
337 | rc = vmemmap_create_mapping(start, page_size, __pa(p)); |
338 | if (rc < 0) { | |
339 | pr_warning( | |
340 | "vmemmap_populate: Unable to create vmemmap mapping: %d\n", | |
341 | rc); | |
342 | return -EFAULT; | |
343 | } | |
71b0bfe4 LZ |
344 | } |
345 | ||
346 | return 0; | |
347 | } | |
348 | ||
349 | #ifdef CONFIG_MEMORY_HOTPLUG | |
bd8cb03d LZ |
350 | static unsigned long vmemmap_list_free(unsigned long start) |
351 | { | |
352 | struct vmemmap_backing *vmem_back, *vmem_back_prev; | |
353 | ||
354 | vmem_back_prev = vmem_back = vmemmap_list; | |
355 | ||
356 | /* look for it with prev pointer recorded */ | |
357 | for (; vmem_back; vmem_back = vmem_back->list) { | |
358 | if (vmem_back->virt_addr == start) | |
359 | break; | |
360 | vmem_back_prev = vmem_back; | |
361 | } | |
362 | ||
363 | if (unlikely(!vmem_back)) { | |
364 | WARN_ON(1); | |
365 | return 0; | |
366 | } | |
367 | ||
368 | /* remove it from vmemmap_list */ | |
369 | if (vmem_back == vmemmap_list) /* remove head */ | |
370 | vmemmap_list = vmem_back->list; | |
371 | else | |
372 | vmem_back_prev->list = vmem_back->list; | |
373 | ||
374 | /* next point to this freed entry */ | |
375 | vmem_back->list = next; | |
376 | next = vmem_back; | |
377 | num_freed++; | |
378 | ||
379 | return vmem_back->phys; | |
380 | } | |
381 | ||
71b0bfe4 | 382 | void __ref vmemmap_free(unsigned long start, unsigned long end) |
d29eff7b | 383 | { |
cec08e7a | 384 | unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; |
d29eff7b | 385 | |
d29eff7b AW |
386 | start = _ALIGN_DOWN(start, page_size); |
387 | ||
71b0bfe4 | 388 | pr_debug("vmemmap_free %lx...%lx\n", start, end); |
32a74949 | 389 | |
d29eff7b | 390 | for (; start < end; start += page_size) { |
71b0bfe4 | 391 | unsigned long addr; |
d29eff7b | 392 | |
71b0bfe4 LZ |
393 | /* |
394 | * the section has already be marked as invalid, so | |
395 | * vmemmap_populated() true means some other sections still | |
396 | * in this page, so skip it. | |
397 | */ | |
d29eff7b AW |
398 | if (vmemmap_populated(start, page_size)) |
399 | continue; | |
400 | ||
71b0bfe4 LZ |
401 | addr = vmemmap_list_free(start); |
402 | if (addr) { | |
403 | struct page *page = pfn_to_page(addr >> PAGE_SHIFT); | |
404 | ||
405 | if (PageReserved(page)) { | |
406 | /* allocated from bootmem */ | |
407 | if (page_size < PAGE_SIZE) { | |
408 | /* | |
409 | * this shouldn't happen, but if it is | |
410 | * the case, leave the memory there | |
411 | */ | |
412 | WARN_ON_ONCE(1); | |
413 | } else { | |
414 | unsigned int nr_pages = | |
415 | 1 << get_order(page_size); | |
416 | while (nr_pages--) | |
417 | free_reserved_page(page++); | |
418 | } | |
419 | } else | |
420 | free_pages((unsigned long)(__va(addr)), | |
421 | get_order(page_size)); | |
422 | ||
423 | vmemmap_remove_mapping(start, page_size); | |
424 | } | |
d29eff7b | 425 | } |
0197518c | 426 | } |
71b0bfe4 | 427 | #endif |
f7e3334a NF |
428 | void register_page_bootmem_memmap(unsigned long section_nr, |
429 | struct page *start_page, unsigned long size) | |
430 | { | |
431 | } | |
cd3db0c4 | 432 | |
8e0861fa AK |
433 | /* |
434 | * We do not have access to the sparsemem vmemmap, so we fallback to | |
435 | * walking the list of sparsemem blocks which we already maintain for | |
436 | * the sake of crashdump. In the long run, we might want to maintain | |
437 | * a tree if performance of that linear walk becomes a problem. | |
438 | * | |
439 | * realmode_pfn_to_page functions can fail due to: | |
440 | * 1) As real sparsemem blocks do not lay in RAM continously (they | |
441 | * are in virtual address space which is not available in the real mode), | |
442 | * the requested page struct can be split between blocks so get_page/put_page | |
443 | * may fail. | |
444 | * 2) When huge pages are used, the get_page/put_page API will fail | |
445 | * in real mode as the linked addresses in the page struct are virtual | |
446 | * too. | |
447 | */ | |
448 | struct page *realmode_pfn_to_page(unsigned long pfn) | |
449 | { | |
450 | struct vmemmap_backing *vmem_back; | |
451 | struct page *page; | |
452 | unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; | |
453 | unsigned long pg_va = (unsigned long) pfn_to_page(pfn); | |
454 | ||
455 | for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) { | |
456 | if (pg_va < vmem_back->virt_addr) | |
457 | continue; | |
458 | ||
bd8cb03d LZ |
459 | /* After vmemmap_list entry free is possible, need check all */ |
460 | if ((pg_va + sizeof(struct page)) <= | |
461 | (vmem_back->virt_addr + page_size)) { | |
462 | page = (struct page *) (vmem_back->phys + pg_va - | |
8e0861fa | 463 | vmem_back->virt_addr); |
bd8cb03d LZ |
464 | return page; |
465 | } | |
8e0861fa AK |
466 | } |
467 | ||
bd8cb03d | 468 | /* Probably that page struct is split between real pages */ |
8e0861fa AK |
469 | return NULL; |
470 | } | |
471 | EXPORT_SYMBOL_GPL(realmode_pfn_to_page); | |
472 | ||
473 | #elif defined(CONFIG_FLATMEM) | |
474 | ||
475 | struct page *realmode_pfn_to_page(unsigned long pfn) | |
476 | { | |
477 | struct page *page = pfn_to_page(pfn); | |
478 | return page; | |
479 | } | |
480 | EXPORT_SYMBOL_GPL(realmode_pfn_to_page); | |
481 | ||
482 | #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */ |