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> |
1da177e4 | 24 | |
91224346 JT |
25 | #define PAGE_SHIFT_64K 16 |
26 | #define PAGE_SHIFT_16M 24 | |
27 | #define PAGE_SHIFT_16G 34 | |
4ec161cf | 28 | |
41151e77 | 29 | unsigned int HPAGE_SHIFT; |
ec4b2c0c | 30 | |
41151e77 BB |
31 | /* |
32 | * Tracks gpages after the device tree is scanned and before the | |
a6146888 BB |
33 | * huge_boot_pages list is ready. On non-Freescale implementations, this is |
34 | * just used to track 16G pages and so is a single array. FSL-based | |
35 | * implementations may have more than one gpage size, so we need multiple | |
36 | * arrays | |
41151e77 | 37 | */ |
881fde1d | 38 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
39 | #define MAX_NUMBER_GPAGES 128 |
40 | struct psize_gpages { | |
41 | u64 gpage_list[MAX_NUMBER_GPAGES]; | |
42 | unsigned int nr_gpages; | |
43 | }; | |
44 | static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT]; | |
881fde1d BB |
45 | #else |
46 | #define MAX_NUMBER_GPAGES 1024 | |
47 | static u64 gpage_freearray[MAX_NUMBER_GPAGES]; | |
48 | static unsigned nr_gpages; | |
41151e77 | 49 | #endif |
f10a04c0 | 50 | |
a4fe3ce7 DG |
51 | #define hugepd_none(hpd) ((hpd).pd == 0) |
52 | ||
e2b3d202 AK |
53 | #ifdef CONFIG_PPC_BOOK3S_64 |
54 | /* | |
55 | * At this point we do the placement change only for BOOK3S 64. This would | |
56 | * possibly work on other subarchs. | |
57 | */ | |
58 | ||
59 | /* | |
60 | * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have | |
61 | * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD; | |
62 | */ | |
63 | int pmd_huge(pmd_t pmd) | |
64 | { | |
65 | /* | |
66 | * leaf pte for huge page, bottom two bits != 00 | |
67 | */ | |
68 | return ((pmd_val(pmd) & 0x3) != 0x0); | |
69 | } | |
70 | ||
71 | int pud_huge(pud_t pud) | |
72 | { | |
73 | /* | |
74 | * leaf pte for huge page, bottom two bits != 00 | |
75 | */ | |
76 | return ((pud_val(pud) & 0x3) != 0x0); | |
77 | } | |
78 | ||
79 | int pgd_huge(pgd_t pgd) | |
80 | { | |
81 | /* | |
82 | * leaf pte for huge page, bottom two bits != 00 | |
83 | */ | |
84 | return ((pgd_val(pgd) & 0x3) != 0x0); | |
85 | } | |
86 | #else | |
87 | int pmd_huge(pmd_t pmd) | |
88 | { | |
89 | return 0; | |
90 | } | |
91 | ||
92 | int pud_huge(pud_t pud) | |
93 | { | |
94 | return 0; | |
95 | } | |
96 | ||
97 | int pgd_huge(pgd_t pgd) | |
98 | { | |
99 | return 0; | |
100 | } | |
101 | #endif | |
102 | ||
103 | /* | |
104 | * We have 4 cases for pgds and pmds: | |
105 | * (1) invalid (all zeroes) | |
106 | * (2) pointer to next table, as normal; bottom 6 bits == 0 | |
107 | * (3) leaf pte for huge page, bottom two bits != 00 | |
108 | * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table | |
109 | */ | |
a4fe3ce7 DG |
110 | pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift) |
111 | { | |
112 | pgd_t *pg; | |
113 | pud_t *pu; | |
114 | pmd_t *pm; | |
e2b3d202 | 115 | pte_t *ret_pte; |
a4fe3ce7 DG |
116 | hugepd_t *hpdp = NULL; |
117 | unsigned pdshift = PGDIR_SHIFT; | |
118 | ||
119 | if (shift) | |
120 | *shift = 0; | |
121 | ||
122 | pg = pgdir + pgd_index(ea); | |
e2b3d202 AK |
123 | |
124 | if (pgd_huge(*pg)) { | |
125 | ret_pte = (pte_t *) pg; | |
126 | goto out; | |
127 | } else if (is_hugepd(pg)) | |
a4fe3ce7 | 128 | hpdp = (hugepd_t *)pg; |
e2b3d202 | 129 | else if (!pgd_none(*pg)) { |
a4fe3ce7 DG |
130 | pdshift = PUD_SHIFT; |
131 | pu = pud_offset(pg, ea); | |
e2b3d202 AK |
132 | |
133 | if (pud_huge(*pu)) { | |
134 | ret_pte = (pte_t *) pu; | |
135 | goto out; | |
136 | } else if (is_hugepd(pu)) | |
a4fe3ce7 DG |
137 | hpdp = (hugepd_t *)pu; |
138 | else if (!pud_none(*pu)) { | |
139 | pdshift = PMD_SHIFT; | |
140 | pm = pmd_offset(pu, ea); | |
e2b3d202 AK |
141 | |
142 | if (pmd_huge(*pm)) { | |
143 | ret_pte = (pte_t *) pm; | |
144 | goto out; | |
145 | } else if (is_hugepd(pm)) | |
a4fe3ce7 | 146 | hpdp = (hugepd_t *)pm; |
e2b3d202 | 147 | else if (!pmd_none(*pm)) |
41151e77 | 148 | return pte_offset_kernel(pm, ea); |
a4fe3ce7 DG |
149 | } |
150 | } | |
a4fe3ce7 DG |
151 | if (!hpdp) |
152 | return NULL; | |
153 | ||
e2b3d202 AK |
154 | ret_pte = hugepte_offset(hpdp, ea, pdshift); |
155 | pdshift = hugepd_shift(*hpdp); | |
156 | out: | |
a4fe3ce7 | 157 | if (shift) |
e2b3d202 AK |
158 | *shift = pdshift; |
159 | return ret_pte; | |
a4fe3ce7 | 160 | } |
342d3db7 | 161 | EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte); |
a4fe3ce7 DG |
162 | |
163 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
164 | { | |
165 | return find_linux_pte_or_hugepte(mm->pgd, addr, NULL); | |
166 | } | |
167 | ||
f10a04c0 | 168 | static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, |
a4fe3ce7 | 169 | unsigned long address, unsigned pdshift, unsigned pshift) |
f10a04c0 | 170 | { |
41151e77 BB |
171 | struct kmem_cache *cachep; |
172 | pte_t *new; | |
173 | ||
881fde1d | 174 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
175 | int i; |
176 | int num_hugepd = 1 << (pshift - pdshift); | |
177 | cachep = hugepte_cache; | |
881fde1d BB |
178 | #else |
179 | cachep = PGT_CACHE(pdshift - pshift); | |
41151e77 BB |
180 | #endif |
181 | ||
182 | new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT); | |
f10a04c0 | 183 | |
a4fe3ce7 DG |
184 | BUG_ON(pshift > HUGEPD_SHIFT_MASK); |
185 | BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK); | |
186 | ||
f10a04c0 DG |
187 | if (! new) |
188 | return -ENOMEM; | |
189 | ||
190 | spin_lock(&mm->page_table_lock); | |
881fde1d | 191 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
192 | /* |
193 | * We have multiple higher-level entries that point to the same | |
194 | * actual pte location. Fill in each as we go and backtrack on error. | |
195 | * We need all of these so the DTLB pgtable walk code can find the | |
196 | * right higher-level entry without knowing if it's a hugepage or not. | |
197 | */ | |
198 | for (i = 0; i < num_hugepd; i++, hpdp++) { | |
199 | if (unlikely(!hugepd_none(*hpdp))) | |
200 | break; | |
201 | else | |
cf9427b8 | 202 | /* We use the old format for PPC_FSL_BOOK3E */ |
41151e77 BB |
203 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
204 | } | |
205 | /* If we bailed from the for loop early, an error occurred, clean up */ | |
206 | if (i < num_hugepd) { | |
207 | for (i = i - 1 ; i >= 0; i--, hpdp--) | |
208 | hpdp->pd = 0; | |
209 | kmem_cache_free(cachep, new); | |
210 | } | |
a1cd5419 BB |
211 | #else |
212 | if (!hugepd_none(*hpdp)) | |
213 | kmem_cache_free(cachep, new); | |
cf9427b8 AK |
214 | else { |
215 | #ifdef CONFIG_PPC_BOOK3S_64 | |
216 | hpdp->pd = (unsigned long)new | | |
217 | (shift_to_mmu_psize(pshift) << 2); | |
218 | #else | |
a1cd5419 | 219 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
cf9427b8 AK |
220 | #endif |
221 | } | |
41151e77 | 222 | #endif |
f10a04c0 DG |
223 | spin_unlock(&mm->page_table_lock); |
224 | return 0; | |
225 | } | |
226 | ||
a1cd5419 BB |
227 | /* |
228 | * These macros define how to determine which level of the page table holds | |
229 | * the hpdp. | |
230 | */ | |
231 | #ifdef CONFIG_PPC_FSL_BOOK3E | |
232 | #define HUGEPD_PGD_SHIFT PGDIR_SHIFT | |
233 | #define HUGEPD_PUD_SHIFT PUD_SHIFT | |
234 | #else | |
235 | #define HUGEPD_PGD_SHIFT PUD_SHIFT | |
236 | #define HUGEPD_PUD_SHIFT PMD_SHIFT | |
237 | #endif | |
238 | ||
e2b3d202 AK |
239 | #ifdef CONFIG_PPC_BOOK3S_64 |
240 | /* | |
241 | * At this point we do the placement change only for BOOK3S 64. This would | |
242 | * possibly work on other subarchs. | |
243 | */ | |
244 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) | |
245 | { | |
246 | pgd_t *pg; | |
247 | pud_t *pu; | |
248 | pmd_t *pm; | |
249 | hugepd_t *hpdp = NULL; | |
250 | unsigned pshift = __ffs(sz); | |
251 | unsigned pdshift = PGDIR_SHIFT; | |
252 | ||
253 | addr &= ~(sz-1); | |
254 | pg = pgd_offset(mm, addr); | |
255 | ||
256 | if (pshift == PGDIR_SHIFT) | |
257 | /* 16GB huge page */ | |
258 | return (pte_t *) pg; | |
259 | else if (pshift > PUD_SHIFT) | |
260 | /* | |
261 | * We need to use hugepd table | |
262 | */ | |
263 | hpdp = (hugepd_t *)pg; | |
264 | else { | |
265 | pdshift = PUD_SHIFT; | |
266 | pu = pud_alloc(mm, pg, addr); | |
267 | if (pshift == PUD_SHIFT) | |
268 | return (pte_t *)pu; | |
269 | else if (pshift > PMD_SHIFT) | |
270 | hpdp = (hugepd_t *)pu; | |
271 | else { | |
272 | pdshift = PMD_SHIFT; | |
273 | pm = pmd_alloc(mm, pu, addr); | |
274 | if (pshift == PMD_SHIFT) | |
275 | /* 16MB hugepage */ | |
276 | return (pte_t *)pm; | |
277 | else | |
278 | hpdp = (hugepd_t *)pm; | |
279 | } | |
280 | } | |
281 | if (!hpdp) | |
282 | return NULL; | |
283 | ||
284 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
285 | ||
286 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
287 | return NULL; | |
288 | ||
289 | return hugepte_offset(hpdp, addr, pdshift); | |
290 | } | |
291 | ||
292 | #else | |
293 | ||
a4fe3ce7 | 294 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) |
0b26425c | 295 | { |
a4fe3ce7 DG |
296 | pgd_t *pg; |
297 | pud_t *pu; | |
298 | pmd_t *pm; | |
299 | hugepd_t *hpdp = NULL; | |
300 | unsigned pshift = __ffs(sz); | |
301 | unsigned pdshift = PGDIR_SHIFT; | |
302 | ||
303 | addr &= ~(sz-1); | |
304 | ||
305 | pg = pgd_offset(mm, addr); | |
a1cd5419 BB |
306 | |
307 | if (pshift >= HUGEPD_PGD_SHIFT) { | |
a4fe3ce7 DG |
308 | hpdp = (hugepd_t *)pg; |
309 | } else { | |
310 | pdshift = PUD_SHIFT; | |
311 | pu = pud_alloc(mm, pg, addr); | |
a1cd5419 | 312 | if (pshift >= HUGEPD_PUD_SHIFT) { |
a4fe3ce7 DG |
313 | hpdp = (hugepd_t *)pu; |
314 | } else { | |
315 | pdshift = PMD_SHIFT; | |
316 | pm = pmd_alloc(mm, pu, addr); | |
317 | hpdp = (hugepd_t *)pm; | |
318 | } | |
319 | } | |
320 | ||
321 | if (!hpdp) | |
322 | return NULL; | |
323 | ||
324 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
325 | ||
326 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
327 | return NULL; | |
328 | ||
329 | return hugepte_offset(hpdp, addr, pdshift); | |
4ec161cf | 330 | } |
e2b3d202 | 331 | #endif |
4ec161cf | 332 | |
881fde1d | 333 | #ifdef CONFIG_PPC_FSL_BOOK3E |
658013e9 JT |
334 | /* Build list of addresses of gigantic pages. This function is used in early |
335 | * boot before the buddy or bootmem allocator is setup. | |
336 | */ | |
41151e77 BB |
337 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) |
338 | { | |
339 | unsigned int idx = shift_to_mmu_psize(__ffs(page_size)); | |
340 | int i; | |
341 | ||
342 | if (addr == 0) | |
343 | return; | |
344 | ||
345 | gpage_freearray[idx].nr_gpages = number_of_pages; | |
346 | ||
347 | for (i = 0; i < number_of_pages; i++) { | |
348 | gpage_freearray[idx].gpage_list[i] = addr; | |
349 | addr += page_size; | |
350 | } | |
351 | } | |
352 | ||
353 | /* | |
354 | * Moves the gigantic page addresses from the temporary list to the | |
355 | * huge_boot_pages list. | |
356 | */ | |
357 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
358 | { | |
359 | struct huge_bootmem_page *m; | |
360 | int idx = shift_to_mmu_psize(hstate->order + PAGE_SHIFT); | |
361 | int nr_gpages = gpage_freearray[idx].nr_gpages; | |
362 | ||
363 | if (nr_gpages == 0) | |
364 | return 0; | |
365 | ||
366 | #ifdef CONFIG_HIGHMEM | |
367 | /* | |
368 | * If gpages can be in highmem we can't use the trick of storing the | |
369 | * data structure in the page; allocate space for this | |
370 | */ | |
371 | m = alloc_bootmem(sizeof(struct huge_bootmem_page)); | |
372 | m->phys = gpage_freearray[idx].gpage_list[--nr_gpages]; | |
373 | #else | |
374 | m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]); | |
375 | #endif | |
376 | ||
377 | list_add(&m->list, &huge_boot_pages); | |
378 | gpage_freearray[idx].nr_gpages = nr_gpages; | |
379 | gpage_freearray[idx].gpage_list[nr_gpages] = 0; | |
380 | m->hstate = hstate; | |
381 | ||
382 | return 1; | |
383 | } | |
384 | /* | |
385 | * Scan the command line hugepagesz= options for gigantic pages; store those in | |
386 | * a list that we use to allocate the memory once all options are parsed. | |
387 | */ | |
388 | ||
389 | unsigned long gpage_npages[MMU_PAGE_COUNT]; | |
390 | ||
89528127 PG |
391 | static int __init do_gpage_early_setup(char *param, char *val, |
392 | const char *unused) | |
41151e77 BB |
393 | { |
394 | static phys_addr_t size; | |
395 | unsigned long npages; | |
396 | ||
397 | /* | |
398 | * The hugepagesz and hugepages cmdline options are interleaved. We | |
399 | * use the size variable to keep track of whether or not this was done | |
400 | * properly and skip over instances where it is incorrect. Other | |
401 | * command-line parsing code will issue warnings, so we don't need to. | |
402 | * | |
403 | */ | |
404 | if ((strcmp(param, "default_hugepagesz") == 0) || | |
405 | (strcmp(param, "hugepagesz") == 0)) { | |
406 | size = memparse(val, NULL); | |
407 | } else if (strcmp(param, "hugepages") == 0) { | |
408 | if (size != 0) { | |
409 | if (sscanf(val, "%lu", &npages) <= 0) | |
410 | npages = 0; | |
411 | gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages; | |
412 | size = 0; | |
413 | } | |
414 | } | |
415 | return 0; | |
416 | } | |
417 | ||
418 | ||
419 | /* | |
420 | * This function allocates physical space for pages that are larger than the | |
421 | * buddy allocator can handle. We want to allocate these in highmem because | |
422 | * the amount of lowmem is limited. This means that this function MUST be | |
423 | * called before lowmem_end_addr is set up in MMU_init() in order for the lmb | |
424 | * allocate to grab highmem. | |
425 | */ | |
426 | void __init reserve_hugetlb_gpages(void) | |
427 | { | |
428 | static __initdata char cmdline[COMMAND_LINE_SIZE]; | |
429 | phys_addr_t size, base; | |
430 | int i; | |
431 | ||
432 | strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE); | |
026cee00 PM |
433 | parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0, |
434 | &do_gpage_early_setup); | |
41151e77 BB |
435 | |
436 | /* | |
437 | * Walk gpage list in reverse, allocating larger page sizes first. | |
438 | * Skip over unsupported sizes, or sizes that have 0 gpages allocated. | |
439 | * When we reach the point in the list where pages are no longer | |
440 | * considered gpages, we're done. | |
441 | */ | |
442 | for (i = MMU_PAGE_COUNT-1; i >= 0; i--) { | |
443 | if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0) | |
444 | continue; | |
445 | else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT)) | |
446 | break; | |
447 | ||
448 | size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i)); | |
449 | base = memblock_alloc_base(size * gpage_npages[i], size, | |
450 | MEMBLOCK_ALLOC_ANYWHERE); | |
451 | add_gpage(base, size, gpage_npages[i]); | |
452 | } | |
453 | } | |
454 | ||
881fde1d | 455 | #else /* !PPC_FSL_BOOK3E */ |
41151e77 BB |
456 | |
457 | /* Build list of addresses of gigantic pages. This function is used in early | |
458 | * boot before the buddy or bootmem allocator is setup. | |
459 | */ | |
460 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) | |
658013e9 JT |
461 | { |
462 | if (!addr) | |
463 | return; | |
464 | while (number_of_pages > 0) { | |
465 | gpage_freearray[nr_gpages] = addr; | |
466 | nr_gpages++; | |
467 | number_of_pages--; | |
468 | addr += page_size; | |
469 | } | |
470 | } | |
471 | ||
ec4b2c0c | 472 | /* Moves the gigantic page addresses from the temporary list to the |
0d9ea754 JT |
473 | * huge_boot_pages list. |
474 | */ | |
475 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
ec4b2c0c JT |
476 | { |
477 | struct huge_bootmem_page *m; | |
478 | if (nr_gpages == 0) | |
479 | return 0; | |
480 | m = phys_to_virt(gpage_freearray[--nr_gpages]); | |
481 | gpage_freearray[nr_gpages] = 0; | |
482 | list_add(&m->list, &huge_boot_pages); | |
0d9ea754 | 483 | m->hstate = hstate; |
ec4b2c0c JT |
484 | return 1; |
485 | } | |
41151e77 | 486 | #endif |
ec4b2c0c | 487 | |
39dde65c CK |
488 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
489 | { | |
490 | return 0; | |
491 | } | |
492 | ||
881fde1d | 493 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
494 | #define HUGEPD_FREELIST_SIZE \ |
495 | ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t)) | |
496 | ||
497 | struct hugepd_freelist { | |
498 | struct rcu_head rcu; | |
499 | unsigned int index; | |
500 | void *ptes[0]; | |
501 | }; | |
502 | ||
503 | static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur); | |
504 | ||
505 | static void hugepd_free_rcu_callback(struct rcu_head *head) | |
506 | { | |
507 | struct hugepd_freelist *batch = | |
508 | container_of(head, struct hugepd_freelist, rcu); | |
509 | unsigned int i; | |
510 | ||
511 | for (i = 0; i < batch->index; i++) | |
512 | kmem_cache_free(hugepte_cache, batch->ptes[i]); | |
513 | ||
514 | free_page((unsigned long)batch); | |
515 | } | |
516 | ||
517 | static void hugepd_free(struct mmu_gather *tlb, void *hugepte) | |
518 | { | |
519 | struct hugepd_freelist **batchp; | |
520 | ||
521 | batchp = &__get_cpu_var(hugepd_freelist_cur); | |
522 | ||
523 | if (atomic_read(&tlb->mm->mm_users) < 2 || | |
524 | cpumask_equal(mm_cpumask(tlb->mm), | |
525 | cpumask_of(smp_processor_id()))) { | |
526 | kmem_cache_free(hugepte_cache, hugepte); | |
527 | return; | |
528 | } | |
529 | ||
530 | if (*batchp == NULL) { | |
531 | *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC); | |
532 | (*batchp)->index = 0; | |
533 | } | |
534 | ||
535 | (*batchp)->ptes[(*batchp)->index++] = hugepte; | |
536 | if ((*batchp)->index == HUGEPD_FREELIST_SIZE) { | |
537 | call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback); | |
538 | *batchp = NULL; | |
539 | } | |
540 | } | |
541 | #endif | |
542 | ||
a4fe3ce7 DG |
543 | static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift, |
544 | unsigned long start, unsigned long end, | |
545 | unsigned long floor, unsigned long ceiling) | |
f10a04c0 DG |
546 | { |
547 | pte_t *hugepte = hugepd_page(*hpdp); | |
41151e77 BB |
548 | int i; |
549 | ||
a4fe3ce7 | 550 | unsigned long pdmask = ~((1UL << pdshift) - 1); |
41151e77 BB |
551 | unsigned int num_hugepd = 1; |
552 | ||
881fde1d BB |
553 | #ifdef CONFIG_PPC_FSL_BOOK3E |
554 | /* Note: On fsl the hpdp may be the first of several */ | |
41151e77 | 555 | num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift)); |
881fde1d BB |
556 | #else |
557 | unsigned int shift = hugepd_shift(*hpdp); | |
41151e77 | 558 | #endif |
a4fe3ce7 DG |
559 | |
560 | start &= pdmask; | |
561 | if (start < floor) | |
562 | return; | |
563 | if (ceiling) { | |
564 | ceiling &= pdmask; | |
565 | if (! ceiling) | |
566 | return; | |
567 | } | |
568 | if (end - 1 > ceiling - 1) | |
569 | return; | |
f10a04c0 | 570 | |
41151e77 BB |
571 | for (i = 0; i < num_hugepd; i++, hpdp++) |
572 | hpdp->pd = 0; | |
573 | ||
f10a04c0 | 574 | tlb->need_flush = 1; |
881fde1d BB |
575 | |
576 | #ifdef CONFIG_PPC_FSL_BOOK3E | |
41151e77 | 577 | hugepd_free(tlb, hugepte); |
881fde1d BB |
578 | #else |
579 | pgtable_free_tlb(tlb, hugepte, pdshift - shift); | |
41151e77 | 580 | #endif |
f10a04c0 DG |
581 | } |
582 | ||
f10a04c0 DG |
583 | static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
584 | unsigned long addr, unsigned long end, | |
a4fe3ce7 | 585 | unsigned long floor, unsigned long ceiling) |
f10a04c0 DG |
586 | { |
587 | pmd_t *pmd; | |
588 | unsigned long next; | |
589 | unsigned long start; | |
590 | ||
591 | start = addr; | |
f10a04c0 | 592 | do { |
a1cd5419 | 593 | pmd = pmd_offset(pud, addr); |
f10a04c0 | 594 | next = pmd_addr_end(addr, end); |
8bbd9f04 AK |
595 | if (!is_hugepd(pmd)) { |
596 | /* | |
597 | * if it is not hugepd pointer, we should already find | |
598 | * it cleared. | |
599 | */ | |
600 | WARN_ON(!pmd_none_or_clear_bad(pmd)); | |
f10a04c0 | 601 | continue; |
8bbd9f04 | 602 | } |
a1cd5419 BB |
603 | #ifdef CONFIG_PPC_FSL_BOOK3E |
604 | /* | |
605 | * Increment next by the size of the huge mapping since | |
606 | * there may be more than one entry at this level for a | |
607 | * single hugepage, but all of them point to | |
608 | * the same kmem cache that holds the hugepte. | |
609 | */ | |
610 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd)); | |
611 | #endif | |
a4fe3ce7 DG |
612 | free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT, |
613 | addr, next, floor, ceiling); | |
a1cd5419 | 614 | } while (addr = next, addr != end); |
f10a04c0 DG |
615 | |
616 | start &= PUD_MASK; | |
617 | if (start < floor) | |
618 | return; | |
619 | if (ceiling) { | |
620 | ceiling &= PUD_MASK; | |
621 | if (!ceiling) | |
622 | return; | |
1da177e4 | 623 | } |
f10a04c0 DG |
624 | if (end - 1 > ceiling - 1) |
625 | return; | |
1da177e4 | 626 | |
f10a04c0 DG |
627 | pmd = pmd_offset(pud, start); |
628 | pud_clear(pud); | |
9e1b32ca | 629 | pmd_free_tlb(tlb, pmd, start); |
f10a04c0 | 630 | } |
f10a04c0 DG |
631 | |
632 | static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
633 | unsigned long addr, unsigned long end, | |
634 | unsigned long floor, unsigned long ceiling) | |
635 | { | |
636 | pud_t *pud; | |
637 | unsigned long next; | |
638 | unsigned long start; | |
639 | ||
640 | start = addr; | |
f10a04c0 | 641 | do { |
a1cd5419 | 642 | pud = pud_offset(pgd, addr); |
f10a04c0 | 643 | next = pud_addr_end(addr, end); |
a4fe3ce7 | 644 | if (!is_hugepd(pud)) { |
4ec161cf JT |
645 | if (pud_none_or_clear_bad(pud)) |
646 | continue; | |
0d9ea754 | 647 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, |
a4fe3ce7 | 648 | ceiling); |
4ec161cf | 649 | } else { |
a1cd5419 BB |
650 | #ifdef CONFIG_PPC_FSL_BOOK3E |
651 | /* | |
652 | * Increment next by the size of the huge mapping since | |
653 | * there may be more than one entry at this level for a | |
654 | * single hugepage, but all of them point to | |
655 | * the same kmem cache that holds the hugepte. | |
656 | */ | |
657 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pud)); | |
658 | #endif | |
a4fe3ce7 DG |
659 | free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT, |
660 | addr, next, floor, ceiling); | |
4ec161cf | 661 | } |
a1cd5419 | 662 | } while (addr = next, addr != end); |
f10a04c0 DG |
663 | |
664 | start &= PGDIR_MASK; | |
665 | if (start < floor) | |
666 | return; | |
667 | if (ceiling) { | |
668 | ceiling &= PGDIR_MASK; | |
669 | if (!ceiling) | |
670 | return; | |
671 | } | |
672 | if (end - 1 > ceiling - 1) | |
673 | return; | |
674 | ||
675 | pud = pud_offset(pgd, start); | |
676 | pgd_clear(pgd); | |
9e1b32ca | 677 | pud_free_tlb(tlb, pud, start); |
f10a04c0 DG |
678 | } |
679 | ||
680 | /* | |
681 | * This function frees user-level page tables of a process. | |
682 | * | |
683 | * Must be called with pagetable lock held. | |
684 | */ | |
42b77728 | 685 | void hugetlb_free_pgd_range(struct mmu_gather *tlb, |
f10a04c0 DG |
686 | unsigned long addr, unsigned long end, |
687 | unsigned long floor, unsigned long ceiling) | |
688 | { | |
689 | pgd_t *pgd; | |
690 | unsigned long next; | |
f10a04c0 DG |
691 | |
692 | /* | |
a4fe3ce7 DG |
693 | * Because there are a number of different possible pagetable |
694 | * layouts for hugepage ranges, we limit knowledge of how | |
695 | * things should be laid out to the allocation path | |
696 | * (huge_pte_alloc(), above). Everything else works out the | |
697 | * structure as it goes from information in the hugepd | |
698 | * pointers. That means that we can't here use the | |
699 | * optimization used in the normal page free_pgd_range(), of | |
700 | * checking whether we're actually covering a large enough | |
701 | * range to have to do anything at the top level of the walk | |
702 | * instead of at the bottom. | |
f10a04c0 | 703 | * |
a4fe3ce7 DG |
704 | * To make sense of this, you should probably go read the big |
705 | * block comment at the top of the normal free_pgd_range(), | |
706 | * too. | |
f10a04c0 | 707 | */ |
f10a04c0 | 708 | |
f10a04c0 | 709 | do { |
f10a04c0 | 710 | next = pgd_addr_end(addr, end); |
41151e77 | 711 | pgd = pgd_offset(tlb->mm, addr); |
a4fe3ce7 | 712 | if (!is_hugepd(pgd)) { |
0b26425c DG |
713 | if (pgd_none_or_clear_bad(pgd)) |
714 | continue; | |
715 | hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); | |
716 | } else { | |
881fde1d | 717 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
718 | /* |
719 | * Increment next by the size of the huge mapping since | |
881fde1d BB |
720 | * there may be more than one entry at the pgd level |
721 | * for a single hugepage, but all of them point to the | |
722 | * same kmem cache that holds the hugepte. | |
41151e77 BB |
723 | */ |
724 | next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd)); | |
725 | #endif | |
a4fe3ce7 DG |
726 | free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT, |
727 | addr, next, floor, ceiling); | |
0b26425c | 728 | } |
41151e77 | 729 | } while (addr = next, addr != end); |
1da177e4 LT |
730 | } |
731 | ||
1da177e4 LT |
732 | struct page * |
733 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) | |
734 | { | |
735 | pte_t *ptep; | |
736 | struct page *page; | |
a4fe3ce7 DG |
737 | unsigned shift; |
738 | unsigned long mask; | |
739 | ||
740 | ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift); | |
1da177e4 | 741 | |
0d9ea754 | 742 | /* Verify it is a huge page else bail. */ |
a4fe3ce7 | 743 | if (!ptep || !shift) |
1da177e4 LT |
744 | return ERR_PTR(-EINVAL); |
745 | ||
a4fe3ce7 | 746 | mask = (1UL << shift) - 1; |
1da177e4 | 747 | page = pte_page(*ptep); |
a4fe3ce7 DG |
748 | if (page) |
749 | page += (address & mask) / PAGE_SIZE; | |
1da177e4 LT |
750 | |
751 | return page; | |
752 | } | |
753 | ||
1da177e4 LT |
754 | struct page * |
755 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
756 | pmd_t *pmd, int write) | |
757 | { | |
758 | BUG(); | |
759 | return NULL; | |
760 | } | |
761 | ||
e2b3d202 AK |
762 | int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, |
763 | unsigned long end, int write, struct page **pages, int *nr) | |
a4fe3ce7 DG |
764 | { |
765 | unsigned long mask; | |
766 | unsigned long pte_end; | |
3526741f | 767 | struct page *head, *page, *tail; |
a4fe3ce7 DG |
768 | pte_t pte; |
769 | int refs; | |
770 | ||
771 | pte_end = (addr + sz) & ~(sz-1); | |
772 | if (pte_end < end) | |
773 | end = pte_end; | |
774 | ||
775 | pte = *ptep; | |
776 | mask = _PAGE_PRESENT | _PAGE_USER; | |
777 | if (write) | |
778 | mask |= _PAGE_RW; | |
779 | ||
780 | if ((pte_val(pte) & mask) != mask) | |
781 | return 0; | |
782 | ||
783 | /* hugepages are never "special" */ | |
784 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
785 | ||
786 | refs = 0; | |
787 | head = pte_page(pte); | |
788 | ||
789 | page = head + ((addr & (sz-1)) >> PAGE_SHIFT); | |
3526741f | 790 | tail = page; |
a4fe3ce7 DG |
791 | do { |
792 | VM_BUG_ON(compound_head(page) != head); | |
793 | pages[*nr] = page; | |
794 | (*nr)++; | |
795 | page++; | |
796 | refs++; | |
797 | } while (addr += PAGE_SIZE, addr != end); | |
798 | ||
799 | if (!page_cache_add_speculative(head, refs)) { | |
800 | *nr -= refs; | |
801 | return 0; | |
802 | } | |
803 | ||
804 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
805 | /* Could be optimized better */ | |
85964684 AA |
806 | *nr -= refs; |
807 | while (refs--) | |
405e44f2 | 808 | put_page(head); |
cf592bf7 AA |
809 | return 0; |
810 | } | |
811 | ||
812 | /* | |
813 | * Any tail page need their mapcount reference taken before we | |
814 | * return. | |
815 | */ | |
816 | while (refs--) { | |
817 | if (PageTail(tail)) | |
818 | get_huge_page_tail(tail); | |
819 | tail++; | |
a4fe3ce7 DG |
820 | } |
821 | ||
822 | return 1; | |
823 | } | |
824 | ||
39adfa54 DG |
825 | static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, |
826 | unsigned long sz) | |
827 | { | |
828 | unsigned long __boundary = (addr + sz) & ~(sz-1); | |
829 | return (__boundary - 1 < end - 1) ? __boundary : end; | |
830 | } | |
831 | ||
a4fe3ce7 DG |
832 | int gup_hugepd(hugepd_t *hugepd, unsigned pdshift, |
833 | unsigned long addr, unsigned long end, | |
834 | int write, struct page **pages, int *nr) | |
835 | { | |
836 | pte_t *ptep; | |
837 | unsigned long sz = 1UL << hugepd_shift(*hugepd); | |
39adfa54 | 838 | unsigned long next; |
a4fe3ce7 DG |
839 | |
840 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
841 | do { | |
39adfa54 | 842 | next = hugepte_addr_end(addr, end, sz); |
a4fe3ce7 DG |
843 | if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr)) |
844 | return 0; | |
39adfa54 | 845 | } while (ptep++, addr = next, addr != end); |
a4fe3ce7 DG |
846 | |
847 | return 1; | |
848 | } | |
1da177e4 | 849 | |
76512959 | 850 | #ifdef CONFIG_PPC_MM_SLICES |
1da177e4 LT |
851 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, |
852 | unsigned long len, unsigned long pgoff, | |
853 | unsigned long flags) | |
854 | { | |
0d9ea754 JT |
855 | struct hstate *hstate = hstate_file(file); |
856 | int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate)); | |
48f797de | 857 | |
34d07177 | 858 | return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1); |
1da177e4 | 859 | } |
76512959 | 860 | #endif |
1da177e4 | 861 | |
3340289d MG |
862 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
863 | { | |
25c29f9e | 864 | #ifdef CONFIG_PPC_MM_SLICES |
3340289d MG |
865 | unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start); |
866 | ||
867 | return 1UL << mmu_psize_to_shift(psize); | |
41151e77 BB |
868 | #else |
869 | if (!is_vm_hugetlb_page(vma)) | |
870 | return PAGE_SIZE; | |
871 | ||
872 | return huge_page_size(hstate_vma(vma)); | |
873 | #endif | |
874 | } | |
875 | ||
876 | static inline bool is_power_of_4(unsigned long x) | |
877 | { | |
878 | if (is_power_of_2(x)) | |
879 | return (__ilog2(x) % 2) ? false : true; | |
880 | return false; | |
3340289d MG |
881 | } |
882 | ||
d1837cba | 883 | static int __init add_huge_page_size(unsigned long long size) |
4ec161cf | 884 | { |
d1837cba DG |
885 | int shift = __ffs(size); |
886 | int mmu_psize; | |
a4fe3ce7 | 887 | |
4ec161cf | 888 | /* Check that it is a page size supported by the hardware and |
d1837cba | 889 | * that it fits within pagetable and slice limits. */ |
41151e77 BB |
890 | #ifdef CONFIG_PPC_FSL_BOOK3E |
891 | if ((size < PAGE_SIZE) || !is_power_of_4(size)) | |
892 | return -EINVAL; | |
893 | #else | |
d1837cba DG |
894 | if (!is_power_of_2(size) |
895 | || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT)) | |
896 | return -EINVAL; | |
41151e77 | 897 | #endif |
91224346 | 898 | |
d1837cba DG |
899 | if ((mmu_psize = shift_to_mmu_psize(shift)) < 0) |
900 | return -EINVAL; | |
901 | ||
902 | #ifdef CONFIG_SPU_FS_64K_LS | |
903 | /* Disable support for 64K huge pages when 64K SPU local store | |
904 | * support is enabled as the current implementation conflicts. | |
905 | */ | |
906 | if (shift == PAGE_SHIFT_64K) | |
907 | return -EINVAL; | |
908 | #endif /* CONFIG_SPU_FS_64K_LS */ | |
909 | ||
910 | BUG_ON(mmu_psize_defs[mmu_psize].shift != shift); | |
911 | ||
912 | /* Return if huge page size has already been setup */ | |
913 | if (size_to_hstate(size)) | |
914 | return 0; | |
915 | ||
916 | hugetlb_add_hstate(shift - PAGE_SHIFT); | |
917 | ||
918 | return 0; | |
4ec161cf JT |
919 | } |
920 | ||
921 | static int __init hugepage_setup_sz(char *str) | |
922 | { | |
923 | unsigned long long size; | |
4ec161cf JT |
924 | |
925 | size = memparse(str, &str); | |
926 | ||
d1837cba | 927 | if (add_huge_page_size(size) != 0) |
4ec161cf JT |
928 | printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size); |
929 | ||
930 | return 1; | |
931 | } | |
932 | __setup("hugepagesz=", hugepage_setup_sz); | |
933 | ||
881fde1d | 934 | #ifdef CONFIG_PPC_FSL_BOOK3E |
41151e77 BB |
935 | struct kmem_cache *hugepte_cache; |
936 | static int __init hugetlbpage_init(void) | |
937 | { | |
938 | int psize; | |
939 | ||
940 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { | |
941 | unsigned shift; | |
942 | ||
943 | if (!mmu_psize_defs[psize].shift) | |
944 | continue; | |
945 | ||
946 | shift = mmu_psize_to_shift(psize); | |
947 | ||
948 | /* Don't treat normal page sizes as huge... */ | |
949 | if (shift != PAGE_SHIFT) | |
950 | if (add_huge_page_size(1ULL << shift) < 0) | |
951 | continue; | |
952 | } | |
953 | ||
954 | /* | |
955 | * Create a kmem cache for hugeptes. The bottom bits in the pte have | |
956 | * size information encoded in them, so align them to allow this | |
957 | */ | |
958 | hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t), | |
959 | HUGEPD_SHIFT_MASK + 1, 0, NULL); | |
960 | if (hugepte_cache == NULL) | |
961 | panic("%s: Unable to create kmem cache for hugeptes\n", | |
962 | __func__); | |
963 | ||
964 | /* Default hpage size = 4M */ | |
965 | if (mmu_psize_defs[MMU_PAGE_4M].shift) | |
966 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift; | |
967 | else | |
968 | panic("%s: Unable to set default huge page size\n", __func__); | |
969 | ||
970 | ||
971 | return 0; | |
972 | } | |
973 | #else | |
f10a04c0 DG |
974 | static int __init hugetlbpage_init(void) |
975 | { | |
a4fe3ce7 | 976 | int psize; |
0d9ea754 | 977 | |
44ae3ab3 | 978 | if (!mmu_has_feature(MMU_FTR_16M_PAGE)) |
f10a04c0 | 979 | return -ENODEV; |
00df438e | 980 | |
d1837cba DG |
981 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { |
982 | unsigned shift; | |
983 | unsigned pdshift; | |
0d9ea754 | 984 | |
d1837cba DG |
985 | if (!mmu_psize_defs[psize].shift) |
986 | continue; | |
00df438e | 987 | |
d1837cba DG |
988 | shift = mmu_psize_to_shift(psize); |
989 | ||
990 | if (add_huge_page_size(1ULL << shift) < 0) | |
991 | continue; | |
992 | ||
993 | if (shift < PMD_SHIFT) | |
994 | pdshift = PMD_SHIFT; | |
995 | else if (shift < PUD_SHIFT) | |
996 | pdshift = PUD_SHIFT; | |
997 | else | |
998 | pdshift = PGDIR_SHIFT; | |
e2b3d202 AK |
999 | /* |
1000 | * if we have pdshift and shift value same, we don't | |
1001 | * use pgt cache for hugepd. | |
1002 | */ | |
1003 | if (pdshift != shift) { | |
1004 | pgtable_cache_add(pdshift - shift, NULL); | |
1005 | if (!PGT_CACHE(pdshift - shift)) | |
1006 | panic("hugetlbpage_init(): could not create " | |
1007 | "pgtable cache for %d bit pagesize\n", shift); | |
1008 | } | |
0d9ea754 | 1009 | } |
f10a04c0 | 1010 | |
d1837cba DG |
1011 | /* Set default large page size. Currently, we pick 16M or 1M |
1012 | * depending on what is available | |
1013 | */ | |
1014 | if (mmu_psize_defs[MMU_PAGE_16M].shift) | |
1015 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift; | |
1016 | else if (mmu_psize_defs[MMU_PAGE_1M].shift) | |
1017 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift; | |
1018 | ||
f10a04c0 DG |
1019 | return 0; |
1020 | } | |
41151e77 | 1021 | #endif |
f10a04c0 | 1022 | module_init(hugetlbpage_init); |
0895ecda DG |
1023 | |
1024 | void flush_dcache_icache_hugepage(struct page *page) | |
1025 | { | |
1026 | int i; | |
41151e77 | 1027 | void *start; |
0895ecda DG |
1028 | |
1029 | BUG_ON(!PageCompound(page)); | |
1030 | ||
41151e77 BB |
1031 | for (i = 0; i < (1UL << compound_order(page)); i++) { |
1032 | if (!PageHighMem(page)) { | |
1033 | __flush_dcache_icache(page_address(page+i)); | |
1034 | } else { | |
2480b208 | 1035 | start = kmap_atomic(page+i); |
41151e77 | 1036 | __flush_dcache_icache(start); |
2480b208 | 1037 | kunmap_atomic(start); |
41151e77 BB |
1038 | } |
1039 | } | |
0895ecda | 1040 | } |