]>
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; | |
06743521 AK |
65 | * |
66 | * Defined in such a way that we can optimize away code block at build time | |
67 | * if CONFIG_HUGETLB_PAGE=n. | |
e2b3d202 AK |
68 | */ |
69 | int pmd_huge(pmd_t pmd) | |
70 | { | |
71 | /* | |
72 | * leaf pte for huge page, bottom two bits != 00 | |
73 | */ | |
74 | return ((pmd_val(pmd) & 0x3) != 0x0); | |
75 | } | |
76 | ||
77 | int pud_huge(pud_t pud) | |
78 | { | |
79 | /* | |
80 | * leaf pte for huge page, bottom two bits != 00 | |
81 | */ | |
82 | return ((pud_val(pud) & 0x3) != 0x0); | |
83 | } | |
84 | ||
85 | int pgd_huge(pgd_t pgd) | |
86 | { | |
87 | /* | |
88 | * leaf pte for huge page, bottom two bits != 00 | |
89 | */ | |
90 | return ((pgd_val(pgd) & 0x3) != 0x0); | |
91 | } | |
92 | #else | |
93 | int pmd_huge(pmd_t pmd) | |
94 | { | |
95 | return 0; | |
96 | } | |
97 | ||
98 | int pud_huge(pud_t pud) | |
99 | { | |
100 | return 0; | |
101 | } | |
102 | ||
103 | int pgd_huge(pgd_t pgd) | |
104 | { | |
105 | return 0; | |
106 | } | |
107 | #endif | |
108 | ||
a4fe3ce7 DG |
109 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) |
110 | { | |
12bc9f6f | 111 | /* Only called for hugetlbfs pages, hence can ignore THP */ |
691e95fd | 112 | return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL); |
a4fe3ce7 DG |
113 | } |
114 | ||
f10a04c0 | 115 | static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, |
a4fe3ce7 | 116 | unsigned long address, unsigned pdshift, unsigned pshift) |
f10a04c0 | 117 | { |
41151e77 BB |
118 | struct kmem_cache *cachep; |
119 | pte_t *new; | |
120 | ||
881fde1d | 121 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
122 | int i; |
123 | int num_hugepd = 1 << (pshift - pdshift); | |
124 | cachep = hugepte_cache; | |
881fde1d BB |
125 | #else |
126 | cachep = PGT_CACHE(pdshift - pshift); | |
41151e77 BB |
127 | #endif |
128 | ||
129 | new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT); | |
f10a04c0 | 130 | |
a4fe3ce7 DG |
131 | BUG_ON(pshift > HUGEPD_SHIFT_MASK); |
132 | BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK); | |
133 | ||
f10a04c0 DG |
134 | if (! new) |
135 | return -ENOMEM; | |
136 | ||
137 | spin_lock(&mm->page_table_lock); | |
881fde1d | 138 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
139 | /* |
140 | * We have multiple higher-level entries that point to the same | |
141 | * actual pte location. Fill in each as we go and backtrack on error. | |
142 | * We need all of these so the DTLB pgtable walk code can find the | |
143 | * right higher-level entry without knowing if it's a hugepage or not. | |
144 | */ | |
145 | for (i = 0; i < num_hugepd; i++, hpdp++) { | |
146 | if (unlikely(!hugepd_none(*hpdp))) | |
147 | break; | |
148 | else | |
cf9427b8 | 149 | /* We use the old format for PPC_FSL_BOOK3E */ |
41151e77 BB |
150 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
151 | } | |
152 | /* If we bailed from the for loop early, an error occurred, clean up */ | |
153 | if (i < num_hugepd) { | |
154 | for (i = i - 1 ; i >= 0; i--, hpdp--) | |
155 | hpdp->pd = 0; | |
156 | kmem_cache_free(cachep, new); | |
157 | } | |
a1cd5419 BB |
158 | #else |
159 | if (!hugepd_none(*hpdp)) | |
160 | kmem_cache_free(cachep, new); | |
cf9427b8 AK |
161 | else { |
162 | #ifdef CONFIG_PPC_BOOK3S_64 | |
163 | hpdp->pd = (unsigned long)new | | |
164 | (shift_to_mmu_psize(pshift) << 2); | |
165 | #else | |
a1cd5419 | 166 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
cf9427b8 AK |
167 | #endif |
168 | } | |
41151e77 | 169 | #endif |
f10a04c0 DG |
170 | spin_unlock(&mm->page_table_lock); |
171 | return 0; | |
172 | } | |
173 | ||
a1cd5419 BB |
174 | /* |
175 | * These macros define how to determine which level of the page table holds | |
176 | * the hpdp. | |
177 | */ | |
178 | #ifdef CONFIG_PPC_FSL_BOOK3E | |
179 | #define HUGEPD_PGD_SHIFT PGDIR_SHIFT | |
180 | #define HUGEPD_PUD_SHIFT PUD_SHIFT | |
181 | #else | |
182 | #define HUGEPD_PGD_SHIFT PUD_SHIFT | |
183 | #define HUGEPD_PUD_SHIFT PMD_SHIFT | |
184 | #endif | |
185 | ||
e2b3d202 AK |
186 | #ifdef CONFIG_PPC_BOOK3S_64 |
187 | /* | |
188 | * At this point we do the placement change only for BOOK3S 64. This would | |
189 | * possibly work on other subarchs. | |
190 | */ | |
191 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) | |
192 | { | |
193 | pgd_t *pg; | |
194 | pud_t *pu; | |
195 | pmd_t *pm; | |
196 | hugepd_t *hpdp = NULL; | |
197 | unsigned pshift = __ffs(sz); | |
198 | unsigned pdshift = PGDIR_SHIFT; | |
199 | ||
200 | addr &= ~(sz-1); | |
201 | pg = pgd_offset(mm, addr); | |
202 | ||
203 | if (pshift == PGDIR_SHIFT) | |
204 | /* 16GB huge page */ | |
205 | return (pte_t *) pg; | |
206 | else if (pshift > PUD_SHIFT) | |
207 | /* | |
208 | * We need to use hugepd table | |
209 | */ | |
210 | hpdp = (hugepd_t *)pg; | |
211 | else { | |
212 | pdshift = PUD_SHIFT; | |
213 | pu = pud_alloc(mm, pg, addr); | |
214 | if (pshift == PUD_SHIFT) | |
215 | return (pte_t *)pu; | |
216 | else if (pshift > PMD_SHIFT) | |
217 | hpdp = (hugepd_t *)pu; | |
218 | else { | |
219 | pdshift = PMD_SHIFT; | |
220 | pm = pmd_alloc(mm, pu, addr); | |
221 | if (pshift == PMD_SHIFT) | |
222 | /* 16MB hugepage */ | |
223 | return (pte_t *)pm; | |
224 | else | |
225 | hpdp = (hugepd_t *)pm; | |
226 | } | |
227 | } | |
228 | if (!hpdp) | |
229 | return NULL; | |
230 | ||
231 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
232 | ||
233 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
234 | return NULL; | |
235 | ||
b30e7590 | 236 | return hugepte_offset(*hpdp, addr, pdshift); |
e2b3d202 AK |
237 | } |
238 | ||
239 | #else | |
240 | ||
a4fe3ce7 | 241 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) |
0b26425c | 242 | { |
a4fe3ce7 DG |
243 | pgd_t *pg; |
244 | pud_t *pu; | |
245 | pmd_t *pm; | |
246 | hugepd_t *hpdp = NULL; | |
247 | unsigned pshift = __ffs(sz); | |
248 | unsigned pdshift = PGDIR_SHIFT; | |
249 | ||
250 | addr &= ~(sz-1); | |
251 | ||
252 | pg = pgd_offset(mm, addr); | |
a1cd5419 BB |
253 | |
254 | if (pshift >= HUGEPD_PGD_SHIFT) { | |
a4fe3ce7 DG |
255 | hpdp = (hugepd_t *)pg; |
256 | } else { | |
257 | pdshift = PUD_SHIFT; | |
258 | pu = pud_alloc(mm, pg, addr); | |
a1cd5419 | 259 | if (pshift >= HUGEPD_PUD_SHIFT) { |
a4fe3ce7 DG |
260 | hpdp = (hugepd_t *)pu; |
261 | } else { | |
262 | pdshift = PMD_SHIFT; | |
263 | pm = pmd_alloc(mm, pu, addr); | |
264 | hpdp = (hugepd_t *)pm; | |
265 | } | |
266 | } | |
267 | ||
268 | if (!hpdp) | |
269 | return NULL; | |
270 | ||
271 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
272 | ||
273 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
274 | return NULL; | |
275 | ||
b30e7590 | 276 | return hugepte_offset(*hpdp, addr, pdshift); |
4ec161cf | 277 | } |
e2b3d202 | 278 | #endif |
4ec161cf | 279 | |
881fde1d | 280 | #ifdef CONFIG_PPC_FSL_BOOK3E |
658013e9 | 281 | /* Build list of addresses of gigantic pages. This function is used in early |
14ed7409 | 282 | * boot before the buddy allocator is setup. |
658013e9 | 283 | */ |
41151e77 BB |
284 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) |
285 | { | |
286 | unsigned int idx = shift_to_mmu_psize(__ffs(page_size)); | |
287 | int i; | |
288 | ||
289 | if (addr == 0) | |
290 | return; | |
291 | ||
292 | gpage_freearray[idx].nr_gpages = number_of_pages; | |
293 | ||
294 | for (i = 0; i < number_of_pages; i++) { | |
295 | gpage_freearray[idx].gpage_list[i] = addr; | |
296 | addr += page_size; | |
297 | } | |
298 | } | |
299 | ||
300 | /* | |
301 | * Moves the gigantic page addresses from the temporary list to the | |
302 | * huge_boot_pages list. | |
303 | */ | |
304 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
305 | { | |
306 | struct huge_bootmem_page *m; | |
2415cf12 | 307 | int idx = shift_to_mmu_psize(huge_page_shift(hstate)); |
41151e77 BB |
308 | int nr_gpages = gpage_freearray[idx].nr_gpages; |
309 | ||
310 | if (nr_gpages == 0) | |
311 | return 0; | |
312 | ||
313 | #ifdef CONFIG_HIGHMEM | |
314 | /* | |
315 | * If gpages can be in highmem we can't use the trick of storing the | |
316 | * data structure in the page; allocate space for this | |
317 | */ | |
e39f223f | 318 | m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0); |
41151e77 BB |
319 | m->phys = gpage_freearray[idx].gpage_list[--nr_gpages]; |
320 | #else | |
321 | m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]); | |
322 | #endif | |
323 | ||
324 | list_add(&m->list, &huge_boot_pages); | |
325 | gpage_freearray[idx].nr_gpages = nr_gpages; | |
326 | gpage_freearray[idx].gpage_list[nr_gpages] = 0; | |
327 | m->hstate = hstate; | |
328 | ||
329 | return 1; | |
330 | } | |
331 | /* | |
332 | * Scan the command line hugepagesz= options for gigantic pages; store those in | |
333 | * a list that we use to allocate the memory once all options are parsed. | |
334 | */ | |
335 | ||
336 | unsigned long gpage_npages[MMU_PAGE_COUNT]; | |
337 | ||
89528127 PG |
338 | static int __init do_gpage_early_setup(char *param, char *val, |
339 | const char *unused) | |
41151e77 BB |
340 | { |
341 | static phys_addr_t size; | |
342 | unsigned long npages; | |
343 | ||
344 | /* | |
345 | * The hugepagesz and hugepages cmdline options are interleaved. We | |
346 | * use the size variable to keep track of whether or not this was done | |
347 | * properly and skip over instances where it is incorrect. Other | |
348 | * command-line parsing code will issue warnings, so we don't need to. | |
349 | * | |
350 | */ | |
351 | if ((strcmp(param, "default_hugepagesz") == 0) || | |
352 | (strcmp(param, "hugepagesz") == 0)) { | |
353 | size = memparse(val, NULL); | |
354 | } else if (strcmp(param, "hugepages") == 0) { | |
355 | if (size != 0) { | |
356 | if (sscanf(val, "%lu", &npages) <= 0) | |
357 | npages = 0; | |
c4f3eb5f JY |
358 | if (npages > MAX_NUMBER_GPAGES) { |
359 | pr_warn("MMU: %lu pages requested for page " | |
360 | "size %llu KB, limiting to " | |
361 | __stringify(MAX_NUMBER_GPAGES) "\n", | |
362 | npages, size / 1024); | |
363 | npages = MAX_NUMBER_GPAGES; | |
364 | } | |
41151e77 BB |
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 | |
14ed7409 | 412 | * boot before the buddy allocator is setup. |
41151e77 BB |
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 | ||
69111bac | 475 | batchp = this_cpu_ptr(&hugepd_freelist_cur); |
41151e77 BB |
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); | |
94b09d75 | 481 | put_cpu_var(hugepd_freelist_cur); |
41151e77 BB |
482 | return; |
483 | } | |
484 | ||
485 | if (*batchp == NULL) { | |
486 | *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC); | |
487 | (*batchp)->index = 0; | |
488 | } | |
489 | ||
490 | (*batchp)->ptes[(*batchp)->index++] = hugepte; | |
491 | if ((*batchp)->index == HUGEPD_FREELIST_SIZE) { | |
492 | call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback); | |
493 | *batchp = NULL; | |
494 | } | |
94b09d75 | 495 | put_cpu_var(hugepd_freelist_cur); |
41151e77 BB |
496 | } |
497 | #endif | |
498 | ||
a4fe3ce7 DG |
499 | static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift, |
500 | unsigned long start, unsigned long end, | |
501 | unsigned long floor, unsigned long ceiling) | |
f10a04c0 DG |
502 | { |
503 | pte_t *hugepte = hugepd_page(*hpdp); | |
41151e77 BB |
504 | int i; |
505 | ||
a4fe3ce7 | 506 | unsigned long pdmask = ~((1UL << pdshift) - 1); |
41151e77 BB |
507 | unsigned int num_hugepd = 1; |
508 | ||
881fde1d BB |
509 | #ifdef CONFIG_PPC_FSL_BOOK3E |
510 | /* Note: On fsl the hpdp may be the first of several */ | |
41151e77 | 511 | num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift)); |
881fde1d BB |
512 | #else |
513 | unsigned int shift = hugepd_shift(*hpdp); | |
41151e77 | 514 | #endif |
a4fe3ce7 DG |
515 | |
516 | start &= pdmask; | |
517 | if (start < floor) | |
518 | return; | |
519 | if (ceiling) { | |
520 | ceiling &= pdmask; | |
521 | if (! ceiling) | |
522 | return; | |
523 | } | |
524 | if (end - 1 > ceiling - 1) | |
525 | return; | |
f10a04c0 | 526 | |
41151e77 BB |
527 | for (i = 0; i < num_hugepd; i++, hpdp++) |
528 | hpdp->pd = 0; | |
529 | ||
881fde1d | 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); |
b30e7590 | 549 | if (!is_hugepd(__hugepd(pmd_val(*pmd)))) { |
8bbd9f04 AK |
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); |
50c6a665 | 584 | mm_dec_nr_pmds(tlb->mm); |
f10a04c0 | 585 | } |
f10a04c0 DG |
586 | |
587 | static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
588 | unsigned long addr, unsigned long end, | |
589 | unsigned long floor, unsigned long ceiling) | |
590 | { | |
591 | pud_t *pud; | |
592 | unsigned long next; | |
593 | unsigned long start; | |
594 | ||
595 | start = addr; | |
f10a04c0 | 596 | do { |
a1cd5419 | 597 | pud = pud_offset(pgd, addr); |
f10a04c0 | 598 | next = pud_addr_end(addr, end); |
b30e7590 | 599 | if (!is_hugepd(__hugepd(pud_val(*pud)))) { |
4ec161cf JT |
600 | if (pud_none_or_clear_bad(pud)) |
601 | continue; | |
0d9ea754 | 602 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, |
a4fe3ce7 | 603 | ceiling); |
4ec161cf | 604 | } else { |
a1cd5419 BB |
605 | #ifdef CONFIG_PPC_FSL_BOOK3E |
606 | /* | |
607 | * Increment next by the size of the huge mapping since | |
608 | * there may be more than one entry at this level for a | |
609 | * single hugepage, but all of them point to | |
610 | * the same kmem cache that holds the hugepte. | |
611 | */ | |
612 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pud)); | |
613 | #endif | |
a4fe3ce7 DG |
614 | free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT, |
615 | addr, next, floor, ceiling); | |
4ec161cf | 616 | } |
a1cd5419 | 617 | } while (addr = next, addr != end); |
f10a04c0 DG |
618 | |
619 | start &= PGDIR_MASK; | |
620 | if (start < floor) | |
621 | return; | |
622 | if (ceiling) { | |
623 | ceiling &= PGDIR_MASK; | |
624 | if (!ceiling) | |
625 | return; | |
626 | } | |
627 | if (end - 1 > ceiling - 1) | |
628 | return; | |
629 | ||
630 | pud = pud_offset(pgd, start); | |
631 | pgd_clear(pgd); | |
9e1b32ca | 632 | pud_free_tlb(tlb, pud, start); |
f10a04c0 DG |
633 | } |
634 | ||
635 | /* | |
636 | * This function frees user-level page tables of a process. | |
f10a04c0 | 637 | */ |
42b77728 | 638 | void hugetlb_free_pgd_range(struct mmu_gather *tlb, |
f10a04c0 DG |
639 | unsigned long addr, unsigned long end, |
640 | unsigned long floor, unsigned long ceiling) | |
641 | { | |
642 | pgd_t *pgd; | |
643 | unsigned long next; | |
f10a04c0 DG |
644 | |
645 | /* | |
a4fe3ce7 DG |
646 | * Because there are a number of different possible pagetable |
647 | * layouts for hugepage ranges, we limit knowledge of how | |
648 | * things should be laid out to the allocation path | |
649 | * (huge_pte_alloc(), above). Everything else works out the | |
650 | * structure as it goes from information in the hugepd | |
651 | * pointers. That means that we can't here use the | |
652 | * optimization used in the normal page free_pgd_range(), of | |
653 | * checking whether we're actually covering a large enough | |
654 | * range to have to do anything at the top level of the walk | |
655 | * instead of at the bottom. | |
f10a04c0 | 656 | * |
a4fe3ce7 DG |
657 | * To make sense of this, you should probably go read the big |
658 | * block comment at the top of the normal free_pgd_range(), | |
659 | * too. | |
f10a04c0 | 660 | */ |
f10a04c0 | 661 | |
f10a04c0 | 662 | do { |
f10a04c0 | 663 | next = pgd_addr_end(addr, end); |
41151e77 | 664 | pgd = pgd_offset(tlb->mm, addr); |
b30e7590 | 665 | if (!is_hugepd(__hugepd(pgd_val(*pgd)))) { |
0b26425c DG |
666 | if (pgd_none_or_clear_bad(pgd)) |
667 | continue; | |
668 | hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); | |
669 | } else { | |
881fde1d | 670 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
671 | /* |
672 | * Increment next by the size of the huge mapping since | |
881fde1d BB |
673 | * there may be more than one entry at the pgd level |
674 | * for a single hugepage, but all of them point to the | |
675 | * same kmem cache that holds the hugepte. | |
41151e77 BB |
676 | */ |
677 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd)); | |
678 | #endif | |
a4fe3ce7 DG |
679 | free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT, |
680 | addr, next, floor, ceiling); | |
0b26425c | 681 | } |
41151e77 | 682 | } while (addr = next, addr != end); |
1da177e4 LT |
683 | } |
684 | ||
691e95fd AK |
685 | /* |
686 | * We are holding mmap_sem, so a parallel huge page collapse cannot run. | |
687 | * To prevent hugepage split, disable irq. | |
688 | */ | |
1da177e4 LT |
689 | struct page * |
690 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) | |
691 | { | |
692 | pte_t *ptep; | |
693 | struct page *page; | |
a4fe3ce7 | 694 | unsigned shift; |
691e95fd | 695 | unsigned long mask, flags; |
12bc9f6f AK |
696 | /* |
697 | * Transparent hugepages are handled by generic code. We can skip them | |
698 | * here. | |
699 | */ | |
691e95fd | 700 | local_irq_save(flags); |
a4fe3ce7 | 701 | ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift); |
1da177e4 | 702 | |
0d9ea754 | 703 | /* Verify it is a huge page else bail. */ |
691e95fd AK |
704 | if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep)) { |
705 | local_irq_restore(flags); | |
1da177e4 | 706 | return ERR_PTR(-EINVAL); |
691e95fd | 707 | } |
a4fe3ce7 | 708 | mask = (1UL << shift) - 1; |
1da177e4 | 709 | page = pte_page(*ptep); |
a4fe3ce7 DG |
710 | if (page) |
711 | page += (address & mask) / PAGE_SIZE; | |
1da177e4 | 712 | |
691e95fd | 713 | local_irq_restore(flags); |
1da177e4 LT |
714 | return page; |
715 | } | |
716 | ||
1da177e4 LT |
717 | struct page * |
718 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
719 | pmd_t *pmd, int write) | |
720 | { | |
721 | BUG(); | |
722 | return NULL; | |
723 | } | |
724 | ||
61f77eda NH |
725 | struct page * |
726 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
727 | pud_t *pud, int write) | |
728 | { | |
729 | BUG(); | |
730 | return NULL; | |
731 | } | |
732 | ||
39adfa54 DG |
733 | static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, |
734 | unsigned long sz) | |
735 | { | |
736 | unsigned long __boundary = (addr + sz) & ~(sz-1); | |
737 | return (__boundary - 1 < end - 1) ? __boundary : end; | |
738 | } | |
739 | ||
b30e7590 AK |
740 | int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift, |
741 | unsigned long end, int write, struct page **pages, int *nr) | |
a4fe3ce7 DG |
742 | { |
743 | pte_t *ptep; | |
b30e7590 | 744 | unsigned long sz = 1UL << hugepd_shift(hugepd); |
39adfa54 | 745 | unsigned long next; |
a4fe3ce7 DG |
746 | |
747 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
748 | do { | |
39adfa54 | 749 | next = hugepte_addr_end(addr, end, sz); |
a4fe3ce7 DG |
750 | if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr)) |
751 | return 0; | |
39adfa54 | 752 | } while (ptep++, addr = next, addr != end); |
a4fe3ce7 DG |
753 | |
754 | return 1; | |
755 | } | |
1da177e4 | 756 | |
76512959 | 757 | #ifdef CONFIG_PPC_MM_SLICES |
1da177e4 LT |
758 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, |
759 | unsigned long len, unsigned long pgoff, | |
760 | unsigned long flags) | |
761 | { | |
0d9ea754 JT |
762 | struct hstate *hstate = hstate_file(file); |
763 | int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate)); | |
48f797de | 764 | |
34d07177 | 765 | return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1); |
1da177e4 | 766 | } |
76512959 | 767 | #endif |
1da177e4 | 768 | |
3340289d MG |
769 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
770 | { | |
25c29f9e | 771 | #ifdef CONFIG_PPC_MM_SLICES |
3340289d MG |
772 | unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start); |
773 | ||
774 | return 1UL << mmu_psize_to_shift(psize); | |
41151e77 BB |
775 | #else |
776 | if (!is_vm_hugetlb_page(vma)) | |
777 | return PAGE_SIZE; | |
778 | ||
779 | return huge_page_size(hstate_vma(vma)); | |
780 | #endif | |
781 | } | |
782 | ||
783 | static inline bool is_power_of_4(unsigned long x) | |
784 | { | |
785 | if (is_power_of_2(x)) | |
786 | return (__ilog2(x) % 2) ? false : true; | |
787 | return false; | |
3340289d MG |
788 | } |
789 | ||
d1837cba | 790 | static int __init add_huge_page_size(unsigned long long size) |
4ec161cf | 791 | { |
d1837cba DG |
792 | int shift = __ffs(size); |
793 | int mmu_psize; | |
a4fe3ce7 | 794 | |
4ec161cf | 795 | /* Check that it is a page size supported by the hardware and |
d1837cba | 796 | * that it fits within pagetable and slice limits. */ |
41151e77 BB |
797 | #ifdef CONFIG_PPC_FSL_BOOK3E |
798 | if ((size < PAGE_SIZE) || !is_power_of_4(size)) | |
799 | return -EINVAL; | |
800 | #else | |
d1837cba DG |
801 | if (!is_power_of_2(size) |
802 | || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT)) | |
803 | return -EINVAL; | |
41151e77 | 804 | #endif |
91224346 | 805 | |
d1837cba DG |
806 | if ((mmu_psize = shift_to_mmu_psize(shift)) < 0) |
807 | return -EINVAL; | |
808 | ||
809 | #ifdef CONFIG_SPU_FS_64K_LS | |
810 | /* Disable support for 64K huge pages when 64K SPU local store | |
811 | * support is enabled as the current implementation conflicts. | |
812 | */ | |
813 | if (shift == PAGE_SHIFT_64K) | |
814 | return -EINVAL; | |
815 | #endif /* CONFIG_SPU_FS_64K_LS */ | |
816 | ||
817 | BUG_ON(mmu_psize_defs[mmu_psize].shift != shift); | |
818 | ||
819 | /* Return if huge page size has already been setup */ | |
820 | if (size_to_hstate(size)) | |
821 | return 0; | |
822 | ||
823 | hugetlb_add_hstate(shift - PAGE_SHIFT); | |
824 | ||
825 | return 0; | |
4ec161cf JT |
826 | } |
827 | ||
828 | static int __init hugepage_setup_sz(char *str) | |
829 | { | |
830 | unsigned long long size; | |
4ec161cf JT |
831 | |
832 | size = memparse(str, &str); | |
833 | ||
d1837cba | 834 | if (add_huge_page_size(size) != 0) |
4ec161cf JT |
835 | printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size); |
836 | ||
837 | return 1; | |
838 | } | |
839 | __setup("hugepagesz=", hugepage_setup_sz); | |
840 | ||
881fde1d | 841 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
842 | struct kmem_cache *hugepte_cache; |
843 | static int __init hugetlbpage_init(void) | |
844 | { | |
845 | int psize; | |
846 | ||
847 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { | |
848 | unsigned shift; | |
849 | ||
850 | if (!mmu_psize_defs[psize].shift) | |
851 | continue; | |
852 | ||
853 | shift = mmu_psize_to_shift(psize); | |
854 | ||
855 | /* Don't treat normal page sizes as huge... */ | |
856 | if (shift != PAGE_SHIFT) | |
857 | if (add_huge_page_size(1ULL << shift) < 0) | |
858 | continue; | |
859 | } | |
860 | ||
861 | /* | |
862 | * Create a kmem cache for hugeptes. The bottom bits in the pte have | |
863 | * size information encoded in them, so align them to allow this | |
864 | */ | |
865 | hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t), | |
866 | HUGEPD_SHIFT_MASK + 1, 0, NULL); | |
867 | if (hugepte_cache == NULL) | |
868 | panic("%s: Unable to create kmem cache for hugeptes\n", | |
869 | __func__); | |
870 | ||
871 | /* Default hpage size = 4M */ | |
872 | if (mmu_psize_defs[MMU_PAGE_4M].shift) | |
873 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift; | |
874 | else | |
875 | panic("%s: Unable to set default huge page size\n", __func__); | |
876 | ||
877 | ||
878 | return 0; | |
879 | } | |
880 | #else | |
f10a04c0 DG |
881 | static int __init hugetlbpage_init(void) |
882 | { | |
a4fe3ce7 | 883 | int psize; |
0d9ea754 | 884 | |
44ae3ab3 | 885 | if (!mmu_has_feature(MMU_FTR_16M_PAGE)) |
f10a04c0 | 886 | return -ENODEV; |
00df438e | 887 | |
d1837cba DG |
888 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { |
889 | unsigned shift; | |
890 | unsigned pdshift; | |
0d9ea754 | 891 | |
d1837cba DG |
892 | if (!mmu_psize_defs[psize].shift) |
893 | continue; | |
00df438e | 894 | |
d1837cba DG |
895 | shift = mmu_psize_to_shift(psize); |
896 | ||
897 | if (add_huge_page_size(1ULL << shift) < 0) | |
898 | continue; | |
899 | ||
900 | if (shift < PMD_SHIFT) | |
901 | pdshift = PMD_SHIFT; | |
902 | else if (shift < PUD_SHIFT) | |
903 | pdshift = PUD_SHIFT; | |
904 | else | |
905 | pdshift = PGDIR_SHIFT; | |
e2b3d202 AK |
906 | /* |
907 | * if we have pdshift and shift value same, we don't | |
908 | * use pgt cache for hugepd. | |
909 | */ | |
910 | if (pdshift != shift) { | |
911 | pgtable_cache_add(pdshift - shift, NULL); | |
912 | if (!PGT_CACHE(pdshift - shift)) | |
913 | panic("hugetlbpage_init(): could not create " | |
914 | "pgtable cache for %d bit pagesize\n", shift); | |
915 | } | |
0d9ea754 | 916 | } |
f10a04c0 | 917 | |
d1837cba DG |
918 | /* Set default large page size. Currently, we pick 16M or 1M |
919 | * depending on what is available | |
920 | */ | |
921 | if (mmu_psize_defs[MMU_PAGE_16M].shift) | |
922 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift; | |
923 | else if (mmu_psize_defs[MMU_PAGE_1M].shift) | |
924 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift; | |
925 | ||
f10a04c0 DG |
926 | return 0; |
927 | } | |
41151e77 | 928 | #endif |
f10a04c0 | 929 | module_init(hugetlbpage_init); |
0895ecda DG |
930 | |
931 | void flush_dcache_icache_hugepage(struct page *page) | |
932 | { | |
933 | int i; | |
41151e77 | 934 | void *start; |
0895ecda DG |
935 | |
936 | BUG_ON(!PageCompound(page)); | |
937 | ||
41151e77 BB |
938 | for (i = 0; i < (1UL << compound_order(page)); i++) { |
939 | if (!PageHighMem(page)) { | |
940 | __flush_dcache_icache(page_address(page+i)); | |
941 | } else { | |
2480b208 | 942 | start = kmap_atomic(page+i); |
41151e77 | 943 | __flush_dcache_icache(start); |
2480b208 | 944 | kunmap_atomic(start); |
41151e77 BB |
945 | } |
946 | } | |
0895ecda | 947 | } |
29409997 AK |
948 | |
949 | #endif /* CONFIG_HUGETLB_PAGE */ | |
950 | ||
951 | /* | |
952 | * We have 4 cases for pgds and pmds: | |
953 | * (1) invalid (all zeroes) | |
954 | * (2) pointer to next table, as normal; bottom 6 bits == 0 | |
955 | * (3) leaf pte for huge page, bottom two bits != 00 | |
956 | * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table | |
0ac52dd7 AK |
957 | * |
958 | * So long as we atomically load page table pointers we are safe against teardown, | |
959 | * we can follow the address down to the the page and take a ref on it. | |
691e95fd AK |
960 | * This function need to be called with interrupts disabled. We use this variant |
961 | * when we have MSR[EE] = 0 but the paca->soft_enabled = 1 | |
29409997 | 962 | */ |
0ac52dd7 | 963 | |
691e95fd AK |
964 | pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, |
965 | unsigned *shift) | |
29409997 | 966 | { |
0ac52dd7 AK |
967 | pgd_t pgd, *pgdp; |
968 | pud_t pud, *pudp; | |
969 | pmd_t pmd, *pmdp; | |
29409997 AK |
970 | pte_t *ret_pte; |
971 | hugepd_t *hpdp = NULL; | |
972 | unsigned pdshift = PGDIR_SHIFT; | |
973 | ||
974 | if (shift) | |
975 | *shift = 0; | |
976 | ||
0ac52dd7 | 977 | pgdp = pgdir + pgd_index(ea); |
4f9c53c8 | 978 | pgd = READ_ONCE(*pgdp); |
ac52ae47 | 979 | /* |
0ac52dd7 AK |
980 | * Always operate on the local stack value. This make sure the |
981 | * value don't get updated by a parallel THP split/collapse, | |
982 | * page fault or a page unmap. The return pte_t * is still not | |
983 | * stable. So should be checked there for above conditions. | |
ac52ae47 | 984 | */ |
0ac52dd7 | 985 | if (pgd_none(pgd)) |
ac52ae47 | 986 | return NULL; |
0ac52dd7 AK |
987 | else if (pgd_huge(pgd)) { |
988 | ret_pte = (pte_t *) pgdp; | |
29409997 | 989 | goto out; |
b30e7590 | 990 | } else if (is_hugepd(__hugepd(pgd_val(pgd)))) |
0ac52dd7 | 991 | hpdp = (hugepd_t *)&pgd; |
ac52ae47 | 992 | else { |
0ac52dd7 AK |
993 | /* |
994 | * Even if we end up with an unmap, the pgtable will not | |
995 | * be freed, because we do an rcu free and here we are | |
996 | * irq disabled | |
997 | */ | |
29409997 | 998 | pdshift = PUD_SHIFT; |
0ac52dd7 | 999 | pudp = pud_offset(&pgd, ea); |
da1a288d | 1000 | pud = READ_ONCE(*pudp); |
29409997 | 1001 | |
0ac52dd7 | 1002 | if (pud_none(pud)) |
ac52ae47 | 1003 | return NULL; |
0ac52dd7 AK |
1004 | else if (pud_huge(pud)) { |
1005 | ret_pte = (pte_t *) pudp; | |
29409997 | 1006 | goto out; |
b30e7590 | 1007 | } else if (is_hugepd(__hugepd(pud_val(pud)))) |
0ac52dd7 | 1008 | hpdp = (hugepd_t *)&pud; |
ac52ae47 | 1009 | else { |
29409997 | 1010 | pdshift = PMD_SHIFT; |
0ac52dd7 | 1011 | pmdp = pmd_offset(&pud, ea); |
da1a288d | 1012 | pmd = READ_ONCE(*pmdp); |
ac52ae47 AK |
1013 | /* |
1014 | * A hugepage collapse is captured by pmd_none, because | |
1015 | * it mark the pmd none and do a hpte invalidate. | |
1016 | * | |
7d6e7f7f AK |
1017 | * We don't worry about pmd_trans_splitting here, The |
1018 | * caller if it needs to handle the splitting case | |
1019 | * should check for that. | |
ac52ae47 | 1020 | */ |
7d6e7f7f | 1021 | if (pmd_none(pmd)) |
ac52ae47 | 1022 | return NULL; |
29409997 | 1023 | |
0ac52dd7 AK |
1024 | if (pmd_huge(pmd) || pmd_large(pmd)) { |
1025 | ret_pte = (pte_t *) pmdp; | |
29409997 | 1026 | goto out; |
b30e7590 | 1027 | } else if (is_hugepd(__hugepd(pmd_val(pmd)))) |
0ac52dd7 | 1028 | hpdp = (hugepd_t *)&pmd; |
ac52ae47 | 1029 | else |
0ac52dd7 | 1030 | return pte_offset_kernel(&pmd, ea); |
29409997 AK |
1031 | } |
1032 | } | |
1033 | if (!hpdp) | |
1034 | return NULL; | |
1035 | ||
b30e7590 | 1036 | ret_pte = hugepte_offset(*hpdp, ea, pdshift); |
29409997 AK |
1037 | pdshift = hugepd_shift(*hpdp); |
1038 | out: | |
1039 | if (shift) | |
1040 | *shift = pdshift; | |
1041 | return ret_pte; | |
1042 | } | |
691e95fd | 1043 | EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte); |
29409997 AK |
1044 | |
1045 | int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, | |
1046 | unsigned long end, int write, struct page **pages, int *nr) | |
1047 | { | |
1048 | unsigned long mask; | |
1049 | unsigned long pte_end; | |
1050 | struct page *head, *page, *tail; | |
1051 | pte_t pte; | |
1052 | int refs; | |
1053 | ||
1054 | pte_end = (addr + sz) & ~(sz-1); | |
1055 | if (pte_end < end) | |
1056 | end = pte_end; | |
1057 | ||
4f9c53c8 | 1058 | pte = READ_ONCE(*ptep); |
29409997 AK |
1059 | mask = _PAGE_PRESENT | _PAGE_USER; |
1060 | if (write) | |
1061 | mask |= _PAGE_RW; | |
1062 | ||
1063 | if ((pte_val(pte) & mask) != mask) | |
1064 | return 0; | |
1065 | ||
1066 | /* hugepages are never "special" */ | |
1067 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
1068 | ||
1069 | refs = 0; | |
1070 | head = pte_page(pte); | |
1071 | ||
1072 | page = head + ((addr & (sz-1)) >> PAGE_SHIFT); | |
1073 | tail = page; | |
1074 | do { | |
1075 | VM_BUG_ON(compound_head(page) != head); | |
1076 | pages[*nr] = page; | |
1077 | (*nr)++; | |
1078 | page++; | |
1079 | refs++; | |
1080 | } while (addr += PAGE_SIZE, addr != end); | |
1081 | ||
1082 | if (!page_cache_add_speculative(head, refs)) { | |
1083 | *nr -= refs; | |
1084 | return 0; | |
1085 | } | |
1086 | ||
1087 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
1088 | /* Could be optimized better */ | |
1089 | *nr -= refs; | |
1090 | while (refs--) | |
1091 | put_page(head); | |
1092 | return 0; | |
1093 | } | |
1094 | ||
1095 | /* | |
1096 | * Any tail page need their mapcount reference taken before we | |
1097 | * return. | |
1098 | */ | |
1099 | while (refs--) { | |
1100 | if (PageTail(tail)) | |
1101 | get_huge_page_tail(tail); | |
1102 | tail++; | |
1103 | } | |
1104 | ||
1105 | return 1; | |
1106 | } |