]>
Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
41151e77 | 2 | * PPC Huge TLB Page Support for Kernel. |
1da177e4 LT |
3 | * |
4 | * Copyright (C) 2003 David Gibson, IBM Corporation. | |
41151e77 | 5 | * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor |
1da177e4 LT |
6 | * |
7 | * Based on the IA-32 version: | |
8 | * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com> | |
9 | */ | |
10 | ||
1da177e4 | 11 | #include <linux/mm.h> |
883a3e52 | 12 | #include <linux/io.h> |
5a0e3ad6 | 13 | #include <linux/slab.h> |
1da177e4 | 14 | #include <linux/hugetlb.h> |
342d3db7 | 15 | #include <linux/export.h> |
41151e77 BB |
16 | #include <linux/of_fdt.h> |
17 | #include <linux/memblock.h> | |
18 | #include <linux/bootmem.h> | |
13020be8 | 19 | #include <linux/moduleparam.h> |
883a3e52 | 20 | #include <asm/pgtable.h> |
1da177e4 LT |
21 | #include <asm/pgalloc.h> |
22 | #include <asm/tlb.h> | |
41151e77 | 23 | #include <asm/setup.h> |
29409997 AK |
24 | #include <asm/hugetlb.h> |
25 | ||
26 | #ifdef CONFIG_HUGETLB_PAGE | |
1da177e4 | 27 | |
91224346 JT |
28 | #define PAGE_SHIFT_64K 16 |
29 | #define PAGE_SHIFT_16M 24 | |
30 | #define PAGE_SHIFT_16G 34 | |
4ec161cf | 31 | |
41151e77 | 32 | unsigned int HPAGE_SHIFT; |
ec4b2c0c | 33 | |
41151e77 BB |
34 | /* |
35 | * Tracks gpages after the device tree is scanned and before the | |
a6146888 BB |
36 | * huge_boot_pages list is ready. On non-Freescale implementations, this is |
37 | * just used to track 16G pages and so is a single array. FSL-based | |
38 | * implementations may have more than one gpage size, so we need multiple | |
39 | * arrays | |
41151e77 | 40 | */ |
881fde1d | 41 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
42 | #define MAX_NUMBER_GPAGES 128 |
43 | struct psize_gpages { | |
44 | u64 gpage_list[MAX_NUMBER_GPAGES]; | |
45 | unsigned int nr_gpages; | |
46 | }; | |
47 | static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT]; | |
881fde1d BB |
48 | #else |
49 | #define MAX_NUMBER_GPAGES 1024 | |
50 | static u64 gpage_freearray[MAX_NUMBER_GPAGES]; | |
51 | static unsigned nr_gpages; | |
41151e77 | 52 | #endif |
f10a04c0 | 53 | |
a4fe3ce7 DG |
54 | #define hugepd_none(hpd) ((hpd).pd == 0) |
55 | ||
e2b3d202 AK |
56 | #ifdef CONFIG_PPC_BOOK3S_64 |
57 | /* | |
58 | * At this point we do the placement change only for BOOK3S 64. This would | |
59 | * possibly work on other subarchs. | |
60 | */ | |
61 | ||
62 | /* | |
63 | * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have | |
64 | * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD; | |
65 | */ | |
66 | int pmd_huge(pmd_t pmd) | |
67 | { | |
68 | /* | |
69 | * leaf pte for huge page, bottom two bits != 00 | |
70 | */ | |
71 | return ((pmd_val(pmd) & 0x3) != 0x0); | |
72 | } | |
73 | ||
74 | int pud_huge(pud_t pud) | |
75 | { | |
76 | /* | |
77 | * leaf pte for huge page, bottom two bits != 00 | |
78 | */ | |
79 | return ((pud_val(pud) & 0x3) != 0x0); | |
80 | } | |
81 | ||
82 | int pgd_huge(pgd_t pgd) | |
83 | { | |
84 | /* | |
85 | * leaf pte for huge page, bottom two bits != 00 | |
86 | */ | |
87 | return ((pgd_val(pgd) & 0x3) != 0x0); | |
88 | } | |
83467efb NH |
89 | |
90 | int pmd_huge_support(void) | |
91 | { | |
92 | return 1; | |
93 | } | |
e2b3d202 AK |
94 | #else |
95 | int pmd_huge(pmd_t pmd) | |
96 | { | |
97 | return 0; | |
98 | } | |
99 | ||
100 | int pud_huge(pud_t pud) | |
101 | { | |
102 | return 0; | |
103 | } | |
104 | ||
105 | int pgd_huge(pgd_t pgd) | |
106 | { | |
107 | return 0; | |
108 | } | |
83467efb NH |
109 | |
110 | int pmd_huge_support(void) | |
111 | { | |
112 | return 0; | |
113 | } | |
e2b3d202 AK |
114 | #endif |
115 | ||
a4fe3ce7 DG |
116 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) |
117 | { | |
12bc9f6f | 118 | /* Only called for hugetlbfs pages, hence can ignore THP */ |
a4fe3ce7 DG |
119 | return find_linux_pte_or_hugepte(mm->pgd, addr, NULL); |
120 | } | |
121 | ||
f10a04c0 | 122 | static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, |
a4fe3ce7 | 123 | unsigned long address, unsigned pdshift, unsigned pshift) |
f10a04c0 | 124 | { |
41151e77 BB |
125 | struct kmem_cache *cachep; |
126 | pte_t *new; | |
127 | ||
881fde1d | 128 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
129 | int i; |
130 | int num_hugepd = 1 << (pshift - pdshift); | |
131 | cachep = hugepte_cache; | |
881fde1d BB |
132 | #else |
133 | cachep = PGT_CACHE(pdshift - pshift); | |
41151e77 BB |
134 | #endif |
135 | ||
136 | new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT); | |
f10a04c0 | 137 | |
a4fe3ce7 DG |
138 | BUG_ON(pshift > HUGEPD_SHIFT_MASK); |
139 | BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK); | |
140 | ||
f10a04c0 DG |
141 | if (! new) |
142 | return -ENOMEM; | |
143 | ||
144 | spin_lock(&mm->page_table_lock); | |
881fde1d | 145 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
146 | /* |
147 | * We have multiple higher-level entries that point to the same | |
148 | * actual pte location. Fill in each as we go and backtrack on error. | |
149 | * We need all of these so the DTLB pgtable walk code can find the | |
150 | * right higher-level entry without knowing if it's a hugepage or not. | |
151 | */ | |
152 | for (i = 0; i < num_hugepd; i++, hpdp++) { | |
153 | if (unlikely(!hugepd_none(*hpdp))) | |
154 | break; | |
155 | else | |
cf9427b8 | 156 | /* We use the old format for PPC_FSL_BOOK3E */ |
41151e77 BB |
157 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
158 | } | |
159 | /* If we bailed from the for loop early, an error occurred, clean up */ | |
160 | if (i < num_hugepd) { | |
161 | for (i = i - 1 ; i >= 0; i--, hpdp--) | |
162 | hpdp->pd = 0; | |
163 | kmem_cache_free(cachep, new); | |
164 | } | |
a1cd5419 BB |
165 | #else |
166 | if (!hugepd_none(*hpdp)) | |
167 | kmem_cache_free(cachep, new); | |
cf9427b8 AK |
168 | else { |
169 | #ifdef CONFIG_PPC_BOOK3S_64 | |
170 | hpdp->pd = (unsigned long)new | | |
171 | (shift_to_mmu_psize(pshift) << 2); | |
172 | #else | |
a1cd5419 | 173 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
cf9427b8 AK |
174 | #endif |
175 | } | |
41151e77 | 176 | #endif |
f10a04c0 DG |
177 | spin_unlock(&mm->page_table_lock); |
178 | return 0; | |
179 | } | |
180 | ||
a1cd5419 BB |
181 | /* |
182 | * These macros define how to determine which level of the page table holds | |
183 | * the hpdp. | |
184 | */ | |
185 | #ifdef CONFIG_PPC_FSL_BOOK3E | |
186 | #define HUGEPD_PGD_SHIFT PGDIR_SHIFT | |
187 | #define HUGEPD_PUD_SHIFT PUD_SHIFT | |
188 | #else | |
189 | #define HUGEPD_PGD_SHIFT PUD_SHIFT | |
190 | #define HUGEPD_PUD_SHIFT PMD_SHIFT | |
191 | #endif | |
192 | ||
e2b3d202 AK |
193 | #ifdef CONFIG_PPC_BOOK3S_64 |
194 | /* | |
195 | * At this point we do the placement change only for BOOK3S 64. This would | |
196 | * possibly work on other subarchs. | |
197 | */ | |
198 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) | |
199 | { | |
200 | pgd_t *pg; | |
201 | pud_t *pu; | |
202 | pmd_t *pm; | |
203 | hugepd_t *hpdp = NULL; | |
204 | unsigned pshift = __ffs(sz); | |
205 | unsigned pdshift = PGDIR_SHIFT; | |
206 | ||
207 | addr &= ~(sz-1); | |
208 | pg = pgd_offset(mm, addr); | |
209 | ||
210 | if (pshift == PGDIR_SHIFT) | |
211 | /* 16GB huge page */ | |
212 | return (pte_t *) pg; | |
213 | else if (pshift > PUD_SHIFT) | |
214 | /* | |
215 | * We need to use hugepd table | |
216 | */ | |
217 | hpdp = (hugepd_t *)pg; | |
218 | else { | |
219 | pdshift = PUD_SHIFT; | |
220 | pu = pud_alloc(mm, pg, addr); | |
221 | if (pshift == PUD_SHIFT) | |
222 | return (pte_t *)pu; | |
223 | else if (pshift > PMD_SHIFT) | |
224 | hpdp = (hugepd_t *)pu; | |
225 | else { | |
226 | pdshift = PMD_SHIFT; | |
227 | pm = pmd_alloc(mm, pu, addr); | |
228 | if (pshift == PMD_SHIFT) | |
229 | /* 16MB hugepage */ | |
230 | return (pte_t *)pm; | |
231 | else | |
232 | hpdp = (hugepd_t *)pm; | |
233 | } | |
234 | } | |
235 | if (!hpdp) | |
236 | return NULL; | |
237 | ||
238 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
239 | ||
240 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
241 | return NULL; | |
242 | ||
243 | return hugepte_offset(hpdp, addr, pdshift); | |
244 | } | |
245 | ||
246 | #else | |
247 | ||
a4fe3ce7 | 248 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) |
0b26425c | 249 | { |
a4fe3ce7 DG |
250 | pgd_t *pg; |
251 | pud_t *pu; | |
252 | pmd_t *pm; | |
253 | hugepd_t *hpdp = NULL; | |
254 | unsigned pshift = __ffs(sz); | |
255 | unsigned pdshift = PGDIR_SHIFT; | |
256 | ||
257 | addr &= ~(sz-1); | |
258 | ||
259 | pg = pgd_offset(mm, addr); | |
a1cd5419 BB |
260 | |
261 | if (pshift >= HUGEPD_PGD_SHIFT) { | |
a4fe3ce7 DG |
262 | hpdp = (hugepd_t *)pg; |
263 | } else { | |
264 | pdshift = PUD_SHIFT; | |
265 | pu = pud_alloc(mm, pg, addr); | |
a1cd5419 | 266 | if (pshift >= HUGEPD_PUD_SHIFT) { |
a4fe3ce7 DG |
267 | hpdp = (hugepd_t *)pu; |
268 | } else { | |
269 | pdshift = PMD_SHIFT; | |
270 | pm = pmd_alloc(mm, pu, addr); | |
271 | hpdp = (hugepd_t *)pm; | |
272 | } | |
273 | } | |
274 | ||
275 | if (!hpdp) | |
276 | return NULL; | |
277 | ||
278 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
279 | ||
280 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
281 | return NULL; | |
282 | ||
283 | return hugepte_offset(hpdp, addr, pdshift); | |
4ec161cf | 284 | } |
e2b3d202 | 285 | #endif |
4ec161cf | 286 | |
881fde1d | 287 | #ifdef CONFIG_PPC_FSL_BOOK3E |
658013e9 JT |
288 | /* Build list of addresses of gigantic pages. This function is used in early |
289 | * boot before the buddy or bootmem allocator is setup. | |
290 | */ | |
41151e77 BB |
291 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) |
292 | { | |
293 | unsigned int idx = shift_to_mmu_psize(__ffs(page_size)); | |
294 | int i; | |
295 | ||
296 | if (addr == 0) | |
297 | return; | |
298 | ||
299 | gpage_freearray[idx].nr_gpages = number_of_pages; | |
300 | ||
301 | for (i = 0; i < number_of_pages; i++) { | |
302 | gpage_freearray[idx].gpage_list[i] = addr; | |
303 | addr += page_size; | |
304 | } | |
305 | } | |
306 | ||
307 | /* | |
308 | * Moves the gigantic page addresses from the temporary list to the | |
309 | * huge_boot_pages list. | |
310 | */ | |
311 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
312 | { | |
313 | struct huge_bootmem_page *m; | |
2415cf12 | 314 | int idx = shift_to_mmu_psize(huge_page_shift(hstate)); |
41151e77 BB |
315 | int nr_gpages = gpage_freearray[idx].nr_gpages; |
316 | ||
317 | if (nr_gpages == 0) | |
318 | return 0; | |
319 | ||
320 | #ifdef CONFIG_HIGHMEM | |
321 | /* | |
322 | * If gpages can be in highmem we can't use the trick of storing the | |
323 | * data structure in the page; allocate space for this | |
324 | */ | |
325 | m = alloc_bootmem(sizeof(struct huge_bootmem_page)); | |
326 | m->phys = gpage_freearray[idx].gpage_list[--nr_gpages]; | |
327 | #else | |
328 | m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]); | |
329 | #endif | |
330 | ||
331 | list_add(&m->list, &huge_boot_pages); | |
332 | gpage_freearray[idx].nr_gpages = nr_gpages; | |
333 | gpage_freearray[idx].gpage_list[nr_gpages] = 0; | |
334 | m->hstate = hstate; | |
335 | ||
336 | return 1; | |
337 | } | |
338 | /* | |
339 | * Scan the command line hugepagesz= options for gigantic pages; store those in | |
340 | * a list that we use to allocate the memory once all options are parsed. | |
341 | */ | |
342 | ||
343 | unsigned long gpage_npages[MMU_PAGE_COUNT]; | |
344 | ||
89528127 PG |
345 | static int __init do_gpage_early_setup(char *param, char *val, |
346 | const char *unused) | |
41151e77 BB |
347 | { |
348 | static phys_addr_t size; | |
349 | unsigned long npages; | |
350 | ||
351 | /* | |
352 | * The hugepagesz and hugepages cmdline options are interleaved. We | |
353 | * use the size variable to keep track of whether or not this was done | |
354 | * properly and skip over instances where it is incorrect. Other | |
355 | * command-line parsing code will issue warnings, so we don't need to. | |
356 | * | |
357 | */ | |
358 | if ((strcmp(param, "default_hugepagesz") == 0) || | |
359 | (strcmp(param, "hugepagesz") == 0)) { | |
360 | size = memparse(val, NULL); | |
361 | } else if (strcmp(param, "hugepages") == 0) { | |
362 | if (size != 0) { | |
363 | if (sscanf(val, "%lu", &npages) <= 0) | |
364 | npages = 0; | |
365 | gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages; | |
366 | size = 0; | |
367 | } | |
368 | } | |
369 | return 0; | |
370 | } | |
371 | ||
372 | ||
373 | /* | |
374 | * This function allocates physical space for pages that are larger than the | |
375 | * buddy allocator can handle. We want to allocate these in highmem because | |
376 | * the amount of lowmem is limited. This means that this function MUST be | |
377 | * called before lowmem_end_addr is set up in MMU_init() in order for the lmb | |
378 | * allocate to grab highmem. | |
379 | */ | |
380 | void __init reserve_hugetlb_gpages(void) | |
381 | { | |
382 | static __initdata char cmdline[COMMAND_LINE_SIZE]; | |
383 | phys_addr_t size, base; | |
384 | int i; | |
385 | ||
386 | strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE); | |
026cee00 PM |
387 | parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0, |
388 | &do_gpage_early_setup); | |
41151e77 BB |
389 | |
390 | /* | |
391 | * Walk gpage list in reverse, allocating larger page sizes first. | |
392 | * Skip over unsupported sizes, or sizes that have 0 gpages allocated. | |
393 | * When we reach the point in the list where pages are no longer | |
394 | * considered gpages, we're done. | |
395 | */ | |
396 | for (i = MMU_PAGE_COUNT-1; i >= 0; i--) { | |
397 | if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0) | |
398 | continue; | |
399 | else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT)) | |
400 | break; | |
401 | ||
402 | size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i)); | |
403 | base = memblock_alloc_base(size * gpage_npages[i], size, | |
404 | MEMBLOCK_ALLOC_ANYWHERE); | |
405 | add_gpage(base, size, gpage_npages[i]); | |
406 | } | |
407 | } | |
408 | ||
881fde1d | 409 | #else /* !PPC_FSL_BOOK3E */ |
41151e77 BB |
410 | |
411 | /* Build list of addresses of gigantic pages. This function is used in early | |
412 | * boot before the buddy or bootmem allocator is setup. | |
413 | */ | |
414 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) | |
658013e9 JT |
415 | { |
416 | if (!addr) | |
417 | return; | |
418 | while (number_of_pages > 0) { | |
419 | gpage_freearray[nr_gpages] = addr; | |
420 | nr_gpages++; | |
421 | number_of_pages--; | |
422 | addr += page_size; | |
423 | } | |
424 | } | |
425 | ||
ec4b2c0c | 426 | /* Moves the gigantic page addresses from the temporary list to the |
0d9ea754 JT |
427 | * huge_boot_pages list. |
428 | */ | |
429 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
ec4b2c0c JT |
430 | { |
431 | struct huge_bootmem_page *m; | |
432 | if (nr_gpages == 0) | |
433 | return 0; | |
434 | m = phys_to_virt(gpage_freearray[--nr_gpages]); | |
435 | gpage_freearray[nr_gpages] = 0; | |
436 | list_add(&m->list, &huge_boot_pages); | |
0d9ea754 | 437 | m->hstate = hstate; |
ec4b2c0c JT |
438 | return 1; |
439 | } | |
41151e77 | 440 | #endif |
ec4b2c0c | 441 | |
39dde65c KC |
442 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
443 | { | |
444 | return 0; | |
445 | } | |
446 | ||
881fde1d | 447 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
448 | #define HUGEPD_FREELIST_SIZE \ |
449 | ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t)) | |
450 | ||
451 | struct hugepd_freelist { | |
452 | struct rcu_head rcu; | |
453 | unsigned int index; | |
454 | void *ptes[0]; | |
455 | }; | |
456 | ||
457 | static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur); | |
458 | ||
459 | static void hugepd_free_rcu_callback(struct rcu_head *head) | |
460 | { | |
461 | struct hugepd_freelist *batch = | |
462 | container_of(head, struct hugepd_freelist, rcu); | |
463 | unsigned int i; | |
464 | ||
465 | for (i = 0; i < batch->index; i++) | |
466 | kmem_cache_free(hugepte_cache, batch->ptes[i]); | |
467 | ||
468 | free_page((unsigned long)batch); | |
469 | } | |
470 | ||
471 | static void hugepd_free(struct mmu_gather *tlb, void *hugepte) | |
472 | { | |
473 | struct hugepd_freelist **batchp; | |
474 | ||
475 | batchp = &__get_cpu_var(hugepd_freelist_cur); | |
476 | ||
477 | if (atomic_read(&tlb->mm->mm_users) < 2 || | |
478 | cpumask_equal(mm_cpumask(tlb->mm), | |
479 | cpumask_of(smp_processor_id()))) { | |
480 | kmem_cache_free(hugepte_cache, hugepte); | |
481 | return; | |
482 | } | |
483 | ||
484 | if (*batchp == NULL) { | |
485 | *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC); | |
486 | (*batchp)->index = 0; | |
487 | } | |
488 | ||
489 | (*batchp)->ptes[(*batchp)->index++] = hugepte; | |
490 | if ((*batchp)->index == HUGEPD_FREELIST_SIZE) { | |
491 | call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback); | |
492 | *batchp = NULL; | |
493 | } | |
494 | } | |
495 | #endif | |
496 | ||
a4fe3ce7 DG |
497 | static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift, |
498 | unsigned long start, unsigned long end, | |
499 | unsigned long floor, unsigned long ceiling) | |
f10a04c0 DG |
500 | { |
501 | pte_t *hugepte = hugepd_page(*hpdp); | |
41151e77 BB |
502 | int i; |
503 | ||
a4fe3ce7 | 504 | unsigned long pdmask = ~((1UL << pdshift) - 1); |
41151e77 BB |
505 | unsigned int num_hugepd = 1; |
506 | ||
881fde1d BB |
507 | #ifdef CONFIG_PPC_FSL_BOOK3E |
508 | /* Note: On fsl the hpdp may be the first of several */ | |
41151e77 | 509 | num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift)); |
881fde1d BB |
510 | #else |
511 | unsigned int shift = hugepd_shift(*hpdp); | |
41151e77 | 512 | #endif |
a4fe3ce7 DG |
513 | |
514 | start &= pdmask; | |
515 | if (start < floor) | |
516 | return; | |
517 | if (ceiling) { | |
518 | ceiling &= pdmask; | |
519 | if (! ceiling) | |
520 | return; | |
521 | } | |
522 | if (end - 1 > ceiling - 1) | |
523 | return; | |
f10a04c0 | 524 | |
41151e77 BB |
525 | for (i = 0; i < num_hugepd; i++, hpdp++) |
526 | hpdp->pd = 0; | |
527 | ||
f10a04c0 | 528 | tlb->need_flush = 1; |
881fde1d BB |
529 | |
530 | #ifdef CONFIG_PPC_FSL_BOOK3E | |
41151e77 | 531 | hugepd_free(tlb, hugepte); |
881fde1d BB |
532 | #else |
533 | pgtable_free_tlb(tlb, hugepte, pdshift - shift); | |
41151e77 | 534 | #endif |
f10a04c0 DG |
535 | } |
536 | ||
f10a04c0 DG |
537 | static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
538 | unsigned long addr, unsigned long end, | |
a4fe3ce7 | 539 | unsigned long floor, unsigned long ceiling) |
f10a04c0 DG |
540 | { |
541 | pmd_t *pmd; | |
542 | unsigned long next; | |
543 | unsigned long start; | |
544 | ||
545 | start = addr; | |
f10a04c0 | 546 | do { |
a1cd5419 | 547 | pmd = pmd_offset(pud, addr); |
f10a04c0 | 548 | next = pmd_addr_end(addr, end); |
8bbd9f04 AK |
549 | if (!is_hugepd(pmd)) { |
550 | /* | |
551 | * if it is not hugepd pointer, we should already find | |
552 | * it cleared. | |
553 | */ | |
554 | WARN_ON(!pmd_none_or_clear_bad(pmd)); | |
f10a04c0 | 555 | continue; |
8bbd9f04 | 556 | } |
a1cd5419 BB |
557 | #ifdef CONFIG_PPC_FSL_BOOK3E |
558 | /* | |
559 | * Increment next by the size of the huge mapping since | |
560 | * there may be more than one entry at this level for a | |
561 | * single hugepage, but all of them point to | |
562 | * the same kmem cache that holds the hugepte. | |
563 | */ | |
564 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd)); | |
565 | #endif | |
a4fe3ce7 DG |
566 | free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT, |
567 | addr, next, floor, ceiling); | |
a1cd5419 | 568 | } while (addr = next, addr != end); |
f10a04c0 DG |
569 | |
570 | start &= PUD_MASK; | |
571 | if (start < floor) | |
572 | return; | |
573 | if (ceiling) { | |
574 | ceiling &= PUD_MASK; | |
575 | if (!ceiling) | |
576 | return; | |
1da177e4 | 577 | } |
f10a04c0 DG |
578 | if (end - 1 > ceiling - 1) |
579 | return; | |
1da177e4 | 580 | |
f10a04c0 DG |
581 | pmd = pmd_offset(pud, start); |
582 | pud_clear(pud); | |
9e1b32ca | 583 | pmd_free_tlb(tlb, pmd, start); |
f10a04c0 | 584 | } |
f10a04c0 DG |
585 | |
586 | static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
587 | unsigned long addr, unsigned long end, | |
588 | unsigned long floor, unsigned long ceiling) | |
589 | { | |
590 | pud_t *pud; | |
591 | unsigned long next; | |
592 | unsigned long start; | |
593 | ||
594 | start = addr; | |
f10a04c0 | 595 | do { |
a1cd5419 | 596 | pud = pud_offset(pgd, addr); |
f10a04c0 | 597 | next = pud_addr_end(addr, end); |
a4fe3ce7 | 598 | if (!is_hugepd(pud)) { |
4ec161cf JT |
599 | if (pud_none_or_clear_bad(pud)) |
600 | continue; | |
0d9ea754 | 601 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, |
a4fe3ce7 | 602 | ceiling); |
4ec161cf | 603 | } else { |
a1cd5419 BB |
604 | #ifdef CONFIG_PPC_FSL_BOOK3E |
605 | /* | |
606 | * Increment next by the size of the huge mapping since | |
607 | * there may be more than one entry at this level for a | |
608 | * single hugepage, but all of them point to | |
609 | * the same kmem cache that holds the hugepte. | |
610 | */ | |
611 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pud)); | |
612 | #endif | |
a4fe3ce7 DG |
613 | free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT, |
614 | addr, next, floor, ceiling); | |
4ec161cf | 615 | } |
a1cd5419 | 616 | } while (addr = next, addr != end); |
f10a04c0 DG |
617 | |
618 | start &= PGDIR_MASK; | |
619 | if (start < floor) | |
620 | return; | |
621 | if (ceiling) { | |
622 | ceiling &= PGDIR_MASK; | |
623 | if (!ceiling) | |
624 | return; | |
625 | } | |
626 | if (end - 1 > ceiling - 1) | |
627 | return; | |
628 | ||
629 | pud = pud_offset(pgd, start); | |
630 | pgd_clear(pgd); | |
9e1b32ca | 631 | pud_free_tlb(tlb, pud, start); |
f10a04c0 DG |
632 | } |
633 | ||
634 | /* | |
635 | * This function frees user-level page tables of a process. | |
f10a04c0 | 636 | */ |
42b77728 | 637 | void hugetlb_free_pgd_range(struct mmu_gather *tlb, |
f10a04c0 DG |
638 | unsigned long addr, unsigned long end, |
639 | unsigned long floor, unsigned long ceiling) | |
640 | { | |
641 | pgd_t *pgd; | |
642 | unsigned long next; | |
f10a04c0 DG |
643 | |
644 | /* | |
a4fe3ce7 DG |
645 | * Because there are a number of different possible pagetable |
646 | * layouts for hugepage ranges, we limit knowledge of how | |
647 | * things should be laid out to the allocation path | |
648 | * (huge_pte_alloc(), above). Everything else works out the | |
649 | * structure as it goes from information in the hugepd | |
650 | * pointers. That means that we can't here use the | |
651 | * optimization used in the normal page free_pgd_range(), of | |
652 | * checking whether we're actually covering a large enough | |
653 | * range to have to do anything at the top level of the walk | |
654 | * instead of at the bottom. | |
f10a04c0 | 655 | * |
a4fe3ce7 DG |
656 | * To make sense of this, you should probably go read the big |
657 | * block comment at the top of the normal free_pgd_range(), | |
658 | * too. | |
f10a04c0 | 659 | */ |
f10a04c0 | 660 | |
f10a04c0 | 661 | do { |
f10a04c0 | 662 | next = pgd_addr_end(addr, end); |
41151e77 | 663 | pgd = pgd_offset(tlb->mm, addr); |
a4fe3ce7 | 664 | if (!is_hugepd(pgd)) { |
0b26425c DG |
665 | if (pgd_none_or_clear_bad(pgd)) |
666 | continue; | |
667 | hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); | |
668 | } else { | |
881fde1d | 669 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
670 | /* |
671 | * Increment next by the size of the huge mapping since | |
881fde1d BB |
672 | * there may be more than one entry at the pgd level |
673 | * for a single hugepage, but all of them point to the | |
674 | * same kmem cache that holds the hugepte. | |
41151e77 BB |
675 | */ |
676 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd)); | |
677 | #endif | |
a4fe3ce7 DG |
678 | free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT, |
679 | addr, next, floor, ceiling); | |
0b26425c | 680 | } |
41151e77 | 681 | } while (addr = next, addr != end); |
1da177e4 LT |
682 | } |
683 | ||
1da177e4 LT |
684 | struct page * |
685 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) | |
686 | { | |
687 | pte_t *ptep; | |
688 | struct page *page; | |
a4fe3ce7 DG |
689 | unsigned shift; |
690 | unsigned long mask; | |
12bc9f6f AK |
691 | /* |
692 | * Transparent hugepages are handled by generic code. We can skip them | |
693 | * here. | |
694 | */ | |
a4fe3ce7 | 695 | ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift); |
1da177e4 | 696 | |
0d9ea754 | 697 | /* Verify it is a huge page else bail. */ |
12bc9f6f | 698 | if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep)) |
1da177e4 LT |
699 | return ERR_PTR(-EINVAL); |
700 | ||
a4fe3ce7 | 701 | mask = (1UL << shift) - 1; |
1da177e4 | 702 | page = pte_page(*ptep); |
a4fe3ce7 DG |
703 | if (page) |
704 | page += (address & mask) / PAGE_SIZE; | |
1da177e4 LT |
705 | |
706 | return page; | |
707 | } | |
708 | ||
1da177e4 LT |
709 | struct page * |
710 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
711 | pmd_t *pmd, int write) | |
712 | { | |
713 | BUG(); | |
714 | return NULL; | |
715 | } | |
716 | ||
39adfa54 DG |
717 | static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, |
718 | unsigned long sz) | |
719 | { | |
720 | unsigned long __boundary = (addr + sz) & ~(sz-1); | |
721 | return (__boundary - 1 < end - 1) ? __boundary : end; | |
722 | } | |
723 | ||
a4fe3ce7 DG |
724 | int gup_hugepd(hugepd_t *hugepd, unsigned pdshift, |
725 | unsigned long addr, unsigned long end, | |
726 | int write, struct page **pages, int *nr) | |
727 | { | |
728 | pte_t *ptep; | |
729 | unsigned long sz = 1UL << hugepd_shift(*hugepd); | |
39adfa54 | 730 | unsigned long next; |
a4fe3ce7 DG |
731 | |
732 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
733 | do { | |
39adfa54 | 734 | next = hugepte_addr_end(addr, end, sz); |
a4fe3ce7 DG |
735 | if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr)) |
736 | return 0; | |
39adfa54 | 737 | } while (ptep++, addr = next, addr != end); |
a4fe3ce7 DG |
738 | |
739 | return 1; | |
740 | } | |
1da177e4 | 741 | |
76512959 | 742 | #ifdef CONFIG_PPC_MM_SLICES |
1da177e4 LT |
743 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, |
744 | unsigned long len, unsigned long pgoff, | |
745 | unsigned long flags) | |
746 | { | |
0d9ea754 JT |
747 | struct hstate *hstate = hstate_file(file); |
748 | int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate)); | |
48f797de | 749 | |
34d07177 | 750 | return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1); |
1da177e4 | 751 | } |
76512959 | 752 | #endif |
1da177e4 | 753 | |
3340289d MG |
754 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
755 | { | |
25c29f9e | 756 | #ifdef CONFIG_PPC_MM_SLICES |
3340289d MG |
757 | unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start); |
758 | ||
759 | return 1UL << mmu_psize_to_shift(psize); | |
41151e77 BB |
760 | #else |
761 | if (!is_vm_hugetlb_page(vma)) | |
762 | return PAGE_SIZE; | |
763 | ||
764 | return huge_page_size(hstate_vma(vma)); | |
765 | #endif | |
766 | } | |
767 | ||
768 | static inline bool is_power_of_4(unsigned long x) | |
769 | { | |
770 | if (is_power_of_2(x)) | |
771 | return (__ilog2(x) % 2) ? false : true; | |
772 | return false; | |
3340289d MG |
773 | } |
774 | ||
d1837cba | 775 | static int __init add_huge_page_size(unsigned long long size) |
4ec161cf | 776 | { |
d1837cba DG |
777 | int shift = __ffs(size); |
778 | int mmu_psize; | |
a4fe3ce7 | 779 | |
4ec161cf | 780 | /* Check that it is a page size supported by the hardware and |
d1837cba | 781 | * that it fits within pagetable and slice limits. */ |
41151e77 BB |
782 | #ifdef CONFIG_PPC_FSL_BOOK3E |
783 | if ((size < PAGE_SIZE) || !is_power_of_4(size)) | |
784 | return -EINVAL; | |
785 | #else | |
d1837cba DG |
786 | if (!is_power_of_2(size) |
787 | || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT)) | |
788 | return -EINVAL; | |
41151e77 | 789 | #endif |
91224346 | 790 | |
d1837cba DG |
791 | if ((mmu_psize = shift_to_mmu_psize(shift)) < 0) |
792 | return -EINVAL; | |
793 | ||
794 | #ifdef CONFIG_SPU_FS_64K_LS | |
795 | /* Disable support for 64K huge pages when 64K SPU local store | |
796 | * support is enabled as the current implementation conflicts. | |
797 | */ | |
798 | if (shift == PAGE_SHIFT_64K) | |
799 | return -EINVAL; | |
800 | #endif /* CONFIG_SPU_FS_64K_LS */ | |
801 | ||
802 | BUG_ON(mmu_psize_defs[mmu_psize].shift != shift); | |
803 | ||
804 | /* Return if huge page size has already been setup */ | |
805 | if (size_to_hstate(size)) | |
806 | return 0; | |
807 | ||
808 | hugetlb_add_hstate(shift - PAGE_SHIFT); | |
809 | ||
810 | return 0; | |
4ec161cf JT |
811 | } |
812 | ||
813 | static int __init hugepage_setup_sz(char *str) | |
814 | { | |
815 | unsigned long long size; | |
4ec161cf JT |
816 | |
817 | size = memparse(str, &str); | |
818 | ||
d1837cba | 819 | if (add_huge_page_size(size) != 0) |
4ec161cf JT |
820 | printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size); |
821 | ||
822 | return 1; | |
823 | } | |
824 | __setup("hugepagesz=", hugepage_setup_sz); | |
825 | ||
881fde1d | 826 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
827 | struct kmem_cache *hugepte_cache; |
828 | static int __init hugetlbpage_init(void) | |
829 | { | |
830 | int psize; | |
831 | ||
832 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { | |
833 | unsigned shift; | |
834 | ||
835 | if (!mmu_psize_defs[psize].shift) | |
836 | continue; | |
837 | ||
838 | shift = mmu_psize_to_shift(psize); | |
839 | ||
840 | /* Don't treat normal page sizes as huge... */ | |
841 | if (shift != PAGE_SHIFT) | |
842 | if (add_huge_page_size(1ULL << shift) < 0) | |
843 | continue; | |
844 | } | |
845 | ||
846 | /* | |
847 | * Create a kmem cache for hugeptes. The bottom bits in the pte have | |
848 | * size information encoded in them, so align them to allow this | |
849 | */ | |
850 | hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t), | |
851 | HUGEPD_SHIFT_MASK + 1, 0, NULL); | |
852 | if (hugepte_cache == NULL) | |
853 | panic("%s: Unable to create kmem cache for hugeptes\n", | |
854 | __func__); | |
855 | ||
856 | /* Default hpage size = 4M */ | |
857 | if (mmu_psize_defs[MMU_PAGE_4M].shift) | |
858 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift; | |
859 | else | |
860 | panic("%s: Unable to set default huge page size\n", __func__); | |
861 | ||
862 | ||
863 | return 0; | |
864 | } | |
865 | #else | |
f10a04c0 DG |
866 | static int __init hugetlbpage_init(void) |
867 | { | |
a4fe3ce7 | 868 | int psize; |
0d9ea754 | 869 | |
44ae3ab3 | 870 | if (!mmu_has_feature(MMU_FTR_16M_PAGE)) |
f10a04c0 | 871 | return -ENODEV; |
00df438e | 872 | |
d1837cba DG |
873 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { |
874 | unsigned shift; | |
875 | unsigned pdshift; | |
0d9ea754 | 876 | |
d1837cba DG |
877 | if (!mmu_psize_defs[psize].shift) |
878 | continue; | |
00df438e | 879 | |
d1837cba DG |
880 | shift = mmu_psize_to_shift(psize); |
881 | ||
882 | if (add_huge_page_size(1ULL << shift) < 0) | |
883 | continue; | |
884 | ||
885 | if (shift < PMD_SHIFT) | |
886 | pdshift = PMD_SHIFT; | |
887 | else if (shift < PUD_SHIFT) | |
888 | pdshift = PUD_SHIFT; | |
889 | else | |
890 | pdshift = PGDIR_SHIFT; | |
e2b3d202 AK |
891 | /* |
892 | * if we have pdshift and shift value same, we don't | |
893 | * use pgt cache for hugepd. | |
894 | */ | |
895 | if (pdshift != shift) { | |
896 | pgtable_cache_add(pdshift - shift, NULL); | |
897 | if (!PGT_CACHE(pdshift - shift)) | |
898 | panic("hugetlbpage_init(): could not create " | |
899 | "pgtable cache for %d bit pagesize\n", shift); | |
900 | } | |
0d9ea754 | 901 | } |
f10a04c0 | 902 | |
d1837cba DG |
903 | /* Set default large page size. Currently, we pick 16M or 1M |
904 | * depending on what is available | |
905 | */ | |
906 | if (mmu_psize_defs[MMU_PAGE_16M].shift) | |
907 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift; | |
908 | else if (mmu_psize_defs[MMU_PAGE_1M].shift) | |
909 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift; | |
910 | ||
f10a04c0 DG |
911 | return 0; |
912 | } | |
41151e77 | 913 | #endif |
f10a04c0 | 914 | module_init(hugetlbpage_init); |
0895ecda DG |
915 | |
916 | void flush_dcache_icache_hugepage(struct page *page) | |
917 | { | |
918 | int i; | |
41151e77 | 919 | void *start; |
0895ecda DG |
920 | |
921 | BUG_ON(!PageCompound(page)); | |
922 | ||
41151e77 BB |
923 | for (i = 0; i < (1UL << compound_order(page)); i++) { |
924 | if (!PageHighMem(page)) { | |
925 | __flush_dcache_icache(page_address(page+i)); | |
926 | } else { | |
2480b208 | 927 | start = kmap_atomic(page+i); |
41151e77 | 928 | __flush_dcache_icache(start); |
2480b208 | 929 | kunmap_atomic(start); |
41151e77 BB |
930 | } |
931 | } | |
0895ecda | 932 | } |
29409997 AK |
933 | |
934 | #endif /* CONFIG_HUGETLB_PAGE */ | |
935 | ||
936 | /* | |
937 | * We have 4 cases for pgds and pmds: | |
938 | * (1) invalid (all zeroes) | |
939 | * (2) pointer to next table, as normal; bottom 6 bits == 0 | |
940 | * (3) leaf pte for huge page, bottom two bits != 00 | |
941 | * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table | |
0ac52dd7 AK |
942 | * |
943 | * So long as we atomically load page table pointers we are safe against teardown, | |
944 | * we can follow the address down to the the page and take a ref on it. | |
29409997 | 945 | */ |
0ac52dd7 | 946 | |
29409997 AK |
947 | pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift) |
948 | { | |
0ac52dd7 AK |
949 | pgd_t pgd, *pgdp; |
950 | pud_t pud, *pudp; | |
951 | pmd_t pmd, *pmdp; | |
29409997 AK |
952 | pte_t *ret_pte; |
953 | hugepd_t *hpdp = NULL; | |
954 | unsigned pdshift = PGDIR_SHIFT; | |
955 | ||
956 | if (shift) | |
957 | *shift = 0; | |
958 | ||
0ac52dd7 AK |
959 | pgdp = pgdir + pgd_index(ea); |
960 | pgd = ACCESS_ONCE(*pgdp); | |
ac52ae47 | 961 | /* |
0ac52dd7 AK |
962 | * Always operate on the local stack value. This make sure the |
963 | * value don't get updated by a parallel THP split/collapse, | |
964 | * page fault or a page unmap. The return pte_t * is still not | |
965 | * stable. So should be checked there for above conditions. | |
ac52ae47 | 966 | */ |
0ac52dd7 | 967 | if (pgd_none(pgd)) |
ac52ae47 | 968 | return NULL; |
0ac52dd7 AK |
969 | else if (pgd_huge(pgd)) { |
970 | ret_pte = (pte_t *) pgdp; | |
29409997 | 971 | goto out; |
0ac52dd7 AK |
972 | } else if (is_hugepd(&pgd)) |
973 | hpdp = (hugepd_t *)&pgd; | |
ac52ae47 | 974 | else { |
0ac52dd7 AK |
975 | /* |
976 | * Even if we end up with an unmap, the pgtable will not | |
977 | * be freed, because we do an rcu free and here we are | |
978 | * irq disabled | |
979 | */ | |
29409997 | 980 | pdshift = PUD_SHIFT; |
0ac52dd7 AK |
981 | pudp = pud_offset(&pgd, ea); |
982 | pud = ACCESS_ONCE(*pudp); | |
29409997 | 983 | |
0ac52dd7 | 984 | if (pud_none(pud)) |
ac52ae47 | 985 | return NULL; |
0ac52dd7 AK |
986 | else if (pud_huge(pud)) { |
987 | ret_pte = (pte_t *) pudp; | |
29409997 | 988 | goto out; |
0ac52dd7 AK |
989 | } else if (is_hugepd(&pud)) |
990 | hpdp = (hugepd_t *)&pud; | |
ac52ae47 | 991 | else { |
29409997 | 992 | pdshift = PMD_SHIFT; |
0ac52dd7 AK |
993 | pmdp = pmd_offset(&pud, ea); |
994 | pmd = ACCESS_ONCE(*pmdp); | |
ac52ae47 AK |
995 | /* |
996 | * A hugepage collapse is captured by pmd_none, because | |
997 | * it mark the pmd none and do a hpte invalidate. | |
998 | * | |
999 | * A hugepage split is captured by pmd_trans_splitting | |
1000 | * because we mark the pmd trans splitting and do a | |
1001 | * hpte invalidate | |
1002 | * | |
1003 | */ | |
0ac52dd7 | 1004 | if (pmd_none(pmd) || pmd_trans_splitting(pmd)) |
ac52ae47 | 1005 | return NULL; |
29409997 | 1006 | |
0ac52dd7 AK |
1007 | if (pmd_huge(pmd) || pmd_large(pmd)) { |
1008 | ret_pte = (pte_t *) pmdp; | |
29409997 | 1009 | goto out; |
0ac52dd7 AK |
1010 | } else if (is_hugepd(&pmd)) |
1011 | hpdp = (hugepd_t *)&pmd; | |
ac52ae47 | 1012 | else |
0ac52dd7 | 1013 | return pte_offset_kernel(&pmd, ea); |
29409997 AK |
1014 | } |
1015 | } | |
1016 | if (!hpdp) | |
1017 | return NULL; | |
1018 | ||
1019 | ret_pte = hugepte_offset(hpdp, ea, pdshift); | |
1020 | pdshift = hugepd_shift(*hpdp); | |
1021 | out: | |
1022 | if (shift) | |
1023 | *shift = pdshift; | |
1024 | return ret_pte; | |
1025 | } | |
1026 | EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte); | |
1027 | ||
1028 | int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, | |
1029 | unsigned long end, int write, struct page **pages, int *nr) | |
1030 | { | |
1031 | unsigned long mask; | |
1032 | unsigned long pte_end; | |
1033 | struct page *head, *page, *tail; | |
1034 | pte_t pte; | |
1035 | int refs; | |
1036 | ||
1037 | pte_end = (addr + sz) & ~(sz-1); | |
1038 | if (pte_end < end) | |
1039 | end = pte_end; | |
1040 | ||
7888b4dd | 1041 | pte = ACCESS_ONCE(*ptep); |
29409997 AK |
1042 | mask = _PAGE_PRESENT | _PAGE_USER; |
1043 | if (write) | |
1044 | mask |= _PAGE_RW; | |
1045 | ||
1046 | if ((pte_val(pte) & mask) != mask) | |
1047 | return 0; | |
1048 | ||
0ac52dd7 AK |
1049 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1050 | /* | |
1051 | * check for splitting here | |
1052 | */ | |
1053 | if (pmd_trans_splitting(pte_pmd(pte))) | |
1054 | return 0; | |
1055 | #endif | |
1056 | ||
29409997 AK |
1057 | /* hugepages are never "special" */ |
1058 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
1059 | ||
1060 | refs = 0; | |
1061 | head = pte_page(pte); | |
1062 | ||
1063 | page = head + ((addr & (sz-1)) >> PAGE_SHIFT); | |
1064 | tail = page; | |
1065 | do { | |
1066 | VM_BUG_ON(compound_head(page) != head); | |
1067 | pages[*nr] = page; | |
1068 | (*nr)++; | |
1069 | page++; | |
1070 | refs++; | |
1071 | } while (addr += PAGE_SIZE, addr != end); | |
1072 | ||
1073 | if (!page_cache_add_speculative(head, refs)) { | |
1074 | *nr -= refs; | |
1075 | return 0; | |
1076 | } | |
1077 | ||
1078 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
1079 | /* Could be optimized better */ | |
1080 | *nr -= refs; | |
1081 | while (refs--) | |
1082 | put_page(head); | |
1083 | return 0; | |
1084 | } | |
1085 | ||
1086 | /* | |
1087 | * Any tail page need their mapcount reference taken before we | |
1088 | * return. | |
1089 | */ | |
1090 | while (refs--) { | |
1091 | if (PageTail(tail)) | |
1092 | get_huge_page_tail(tail); | |
1093 | tail++; | |
1094 | } | |
1095 | ||
1096 | return 1; | |
1097 | } |