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