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> |
41151e77 BB |
15 | #include <linux/of_fdt.h> |
16 | #include <linux/memblock.h> | |
17 | #include <linux/bootmem.h> | |
883a3e52 | 18 | #include <asm/pgtable.h> |
1da177e4 LT |
19 | #include <asm/pgalloc.h> |
20 | #include <asm/tlb.h> | |
41151e77 | 21 | #include <asm/setup.h> |
1da177e4 | 22 | |
91224346 JT |
23 | #define PAGE_SHIFT_64K 16 |
24 | #define PAGE_SHIFT_16M 24 | |
25 | #define PAGE_SHIFT_16G 34 | |
4ec161cf | 26 | |
41151e77 | 27 | unsigned int HPAGE_SHIFT; |
ec4b2c0c | 28 | |
41151e77 BB |
29 | /* |
30 | * Tracks gpages after the device tree is scanned and before the | |
31 | * huge_boot_pages list is ready. On 64-bit implementations, this is | |
32 | * just used to track 16G pages and so is a single array. 32-bit | |
33 | * implementations may have more than one gpage size due to limitations | |
34 | * of the memory allocators, so we need multiple arrays | |
35 | */ | |
36 | #ifdef CONFIG_PPC64 | |
37 | #define MAX_NUMBER_GPAGES 1024 | |
38 | static u64 gpage_freearray[MAX_NUMBER_GPAGES]; | |
ec4b2c0c | 39 | static unsigned nr_gpages; |
41151e77 BB |
40 | #else |
41 | #define MAX_NUMBER_GPAGES 128 | |
42 | struct psize_gpages { | |
43 | u64 gpage_list[MAX_NUMBER_GPAGES]; | |
44 | unsigned int nr_gpages; | |
45 | }; | |
46 | static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT]; | |
47 | #endif | |
f10a04c0 | 48 | |
0d9ea754 JT |
49 | static inline int shift_to_mmu_psize(unsigned int shift) |
50 | { | |
d1837cba DG |
51 | int psize; |
52 | ||
53 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) | |
54 | if (mmu_psize_defs[psize].shift == shift) | |
55 | return psize; | |
0d9ea754 JT |
56 | return -1; |
57 | } | |
58 | ||
59 | static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize) | |
60 | { | |
61 | if (mmu_psize_defs[mmu_psize].shift) | |
62 | return mmu_psize_defs[mmu_psize].shift; | |
63 | BUG(); | |
64 | } | |
65 | ||
a4fe3ce7 DG |
66 | #define hugepd_none(hpd) ((hpd).pd == 0) |
67 | ||
a4fe3ce7 DG |
68 | pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift) |
69 | { | |
70 | pgd_t *pg; | |
71 | pud_t *pu; | |
72 | pmd_t *pm; | |
73 | hugepd_t *hpdp = NULL; | |
74 | unsigned pdshift = PGDIR_SHIFT; | |
75 | ||
76 | if (shift) | |
77 | *shift = 0; | |
78 | ||
79 | pg = pgdir + pgd_index(ea); | |
80 | if (is_hugepd(pg)) { | |
81 | hpdp = (hugepd_t *)pg; | |
82 | } else if (!pgd_none(*pg)) { | |
83 | pdshift = PUD_SHIFT; | |
84 | pu = pud_offset(pg, ea); | |
85 | if (is_hugepd(pu)) | |
86 | hpdp = (hugepd_t *)pu; | |
87 | else if (!pud_none(*pu)) { | |
88 | pdshift = PMD_SHIFT; | |
89 | pm = pmd_offset(pu, ea); | |
90 | if (is_hugepd(pm)) | |
91 | hpdp = (hugepd_t *)pm; | |
92 | else if (!pmd_none(*pm)) { | |
41151e77 | 93 | return pte_offset_kernel(pm, ea); |
a4fe3ce7 DG |
94 | } |
95 | } | |
96 | } | |
97 | ||
98 | if (!hpdp) | |
99 | return NULL; | |
100 | ||
101 | if (shift) | |
102 | *shift = hugepd_shift(*hpdp); | |
103 | return hugepte_offset(hpdp, ea, pdshift); | |
104 | } | |
105 | ||
106 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
107 | { | |
108 | return find_linux_pte_or_hugepte(mm->pgd, addr, NULL); | |
109 | } | |
110 | ||
f10a04c0 | 111 | static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, |
a4fe3ce7 | 112 | unsigned long address, unsigned pdshift, unsigned pshift) |
f10a04c0 | 113 | { |
41151e77 BB |
114 | struct kmem_cache *cachep; |
115 | pte_t *new; | |
116 | ||
117 | #ifdef CONFIG_PPC64 | |
118 | cachep = PGT_CACHE(pdshift - pshift); | |
119 | #else | |
120 | int i; | |
121 | int num_hugepd = 1 << (pshift - pdshift); | |
122 | cachep = hugepte_cache; | |
123 | #endif | |
124 | ||
125 | new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT); | |
f10a04c0 | 126 | |
a4fe3ce7 DG |
127 | BUG_ON(pshift > HUGEPD_SHIFT_MASK); |
128 | BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK); | |
129 | ||
f10a04c0 DG |
130 | if (! new) |
131 | return -ENOMEM; | |
132 | ||
133 | spin_lock(&mm->page_table_lock); | |
41151e77 | 134 | #ifdef CONFIG_PPC64 |
f10a04c0 | 135 | if (!hugepd_none(*hpdp)) |
41151e77 | 136 | kmem_cache_free(cachep, new); |
f10a04c0 | 137 | else |
41151e77 BB |
138 | hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; |
139 | #else | |
140 | /* | |
141 | * We have multiple higher-level entries that point to the same | |
142 | * actual pte location. Fill in each as we go and backtrack on error. | |
143 | * We need all of these so the DTLB pgtable walk code can find the | |
144 | * right higher-level entry without knowing if it's a hugepage or not. | |
145 | */ | |
146 | for (i = 0; i < num_hugepd; i++, hpdp++) { | |
147 | if (unlikely(!hugepd_none(*hpdp))) | |
148 | break; | |
149 | else | |
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 | } | |
158 | #endif | |
f10a04c0 DG |
159 | spin_unlock(&mm->page_table_lock); |
160 | return 0; | |
161 | } | |
162 | ||
a4fe3ce7 | 163 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) |
0b26425c | 164 | { |
a4fe3ce7 DG |
165 | pgd_t *pg; |
166 | pud_t *pu; | |
167 | pmd_t *pm; | |
168 | hugepd_t *hpdp = NULL; | |
169 | unsigned pshift = __ffs(sz); | |
170 | unsigned pdshift = PGDIR_SHIFT; | |
171 | ||
172 | addr &= ~(sz-1); | |
173 | ||
174 | pg = pgd_offset(mm, addr); | |
175 | if (pshift >= PUD_SHIFT) { | |
176 | hpdp = (hugepd_t *)pg; | |
177 | } else { | |
178 | pdshift = PUD_SHIFT; | |
179 | pu = pud_alloc(mm, pg, addr); | |
180 | if (pshift >= PMD_SHIFT) { | |
181 | hpdp = (hugepd_t *)pu; | |
182 | } else { | |
183 | pdshift = PMD_SHIFT; | |
184 | pm = pmd_alloc(mm, pu, addr); | |
185 | hpdp = (hugepd_t *)pm; | |
186 | } | |
187 | } | |
188 | ||
189 | if (!hpdp) | |
190 | return NULL; | |
191 | ||
192 | BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); | |
193 | ||
194 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) | |
195 | return NULL; | |
196 | ||
197 | return hugepte_offset(hpdp, addr, pdshift); | |
4ec161cf | 198 | } |
4ec161cf | 199 | |
41151e77 | 200 | #ifdef CONFIG_PPC32 |
658013e9 JT |
201 | /* Build list of addresses of gigantic pages. This function is used in early |
202 | * boot before the buddy or bootmem allocator is setup. | |
203 | */ | |
41151e77 BB |
204 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) |
205 | { | |
206 | unsigned int idx = shift_to_mmu_psize(__ffs(page_size)); | |
207 | int i; | |
208 | ||
209 | if (addr == 0) | |
210 | return; | |
211 | ||
212 | gpage_freearray[idx].nr_gpages = number_of_pages; | |
213 | ||
214 | for (i = 0; i < number_of_pages; i++) { | |
215 | gpage_freearray[idx].gpage_list[i] = addr; | |
216 | addr += page_size; | |
217 | } | |
218 | } | |
219 | ||
220 | /* | |
221 | * Moves the gigantic page addresses from the temporary list to the | |
222 | * huge_boot_pages list. | |
223 | */ | |
224 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
225 | { | |
226 | struct huge_bootmem_page *m; | |
227 | int idx = shift_to_mmu_psize(hstate->order + PAGE_SHIFT); | |
228 | int nr_gpages = gpage_freearray[idx].nr_gpages; | |
229 | ||
230 | if (nr_gpages == 0) | |
231 | return 0; | |
232 | ||
233 | #ifdef CONFIG_HIGHMEM | |
234 | /* | |
235 | * If gpages can be in highmem we can't use the trick of storing the | |
236 | * data structure in the page; allocate space for this | |
237 | */ | |
238 | m = alloc_bootmem(sizeof(struct huge_bootmem_page)); | |
239 | m->phys = gpage_freearray[idx].gpage_list[--nr_gpages]; | |
240 | #else | |
241 | m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]); | |
242 | #endif | |
243 | ||
244 | list_add(&m->list, &huge_boot_pages); | |
245 | gpage_freearray[idx].nr_gpages = nr_gpages; | |
246 | gpage_freearray[idx].gpage_list[nr_gpages] = 0; | |
247 | m->hstate = hstate; | |
248 | ||
249 | return 1; | |
250 | } | |
251 | /* | |
252 | * Scan the command line hugepagesz= options for gigantic pages; store those in | |
253 | * a list that we use to allocate the memory once all options are parsed. | |
254 | */ | |
255 | ||
256 | unsigned long gpage_npages[MMU_PAGE_COUNT]; | |
257 | ||
258 | static int __init do_gpage_early_setup(char *param, char *val) | |
259 | { | |
260 | static phys_addr_t size; | |
261 | unsigned long npages; | |
262 | ||
263 | /* | |
264 | * The hugepagesz and hugepages cmdline options are interleaved. We | |
265 | * use the size variable to keep track of whether or not this was done | |
266 | * properly and skip over instances where it is incorrect. Other | |
267 | * command-line parsing code will issue warnings, so we don't need to. | |
268 | * | |
269 | */ | |
270 | if ((strcmp(param, "default_hugepagesz") == 0) || | |
271 | (strcmp(param, "hugepagesz") == 0)) { | |
272 | size = memparse(val, NULL); | |
273 | } else if (strcmp(param, "hugepages") == 0) { | |
274 | if (size != 0) { | |
275 | if (sscanf(val, "%lu", &npages) <= 0) | |
276 | npages = 0; | |
277 | gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages; | |
278 | size = 0; | |
279 | } | |
280 | } | |
281 | return 0; | |
282 | } | |
283 | ||
284 | ||
285 | /* | |
286 | * This function allocates physical space for pages that are larger than the | |
287 | * buddy allocator can handle. We want to allocate these in highmem because | |
288 | * the amount of lowmem is limited. This means that this function MUST be | |
289 | * called before lowmem_end_addr is set up in MMU_init() in order for the lmb | |
290 | * allocate to grab highmem. | |
291 | */ | |
292 | void __init reserve_hugetlb_gpages(void) | |
293 | { | |
294 | static __initdata char cmdline[COMMAND_LINE_SIZE]; | |
295 | phys_addr_t size, base; | |
296 | int i; | |
297 | ||
298 | strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE); | |
299 | parse_args("hugetlb gpages", cmdline, NULL, 0, &do_gpage_early_setup); | |
300 | ||
301 | /* | |
302 | * Walk gpage list in reverse, allocating larger page sizes first. | |
303 | * Skip over unsupported sizes, or sizes that have 0 gpages allocated. | |
304 | * When we reach the point in the list where pages are no longer | |
305 | * considered gpages, we're done. | |
306 | */ | |
307 | for (i = MMU_PAGE_COUNT-1; i >= 0; i--) { | |
308 | if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0) | |
309 | continue; | |
310 | else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT)) | |
311 | break; | |
312 | ||
313 | size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i)); | |
314 | base = memblock_alloc_base(size * gpage_npages[i], size, | |
315 | MEMBLOCK_ALLOC_ANYWHERE); | |
316 | add_gpage(base, size, gpage_npages[i]); | |
317 | } | |
318 | } | |
319 | ||
320 | #else /* PPC64 */ | |
321 | ||
322 | /* Build list of addresses of gigantic pages. This function is used in early | |
323 | * boot before the buddy or bootmem allocator is setup. | |
324 | */ | |
325 | void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) | |
658013e9 JT |
326 | { |
327 | if (!addr) | |
328 | return; | |
329 | while (number_of_pages > 0) { | |
330 | gpage_freearray[nr_gpages] = addr; | |
331 | nr_gpages++; | |
332 | number_of_pages--; | |
333 | addr += page_size; | |
334 | } | |
335 | } | |
336 | ||
ec4b2c0c | 337 | /* Moves the gigantic page addresses from the temporary list to the |
0d9ea754 JT |
338 | * huge_boot_pages list. |
339 | */ | |
340 | int alloc_bootmem_huge_page(struct hstate *hstate) | |
ec4b2c0c JT |
341 | { |
342 | struct huge_bootmem_page *m; | |
343 | if (nr_gpages == 0) | |
344 | return 0; | |
345 | m = phys_to_virt(gpage_freearray[--nr_gpages]); | |
346 | gpage_freearray[nr_gpages] = 0; | |
347 | list_add(&m->list, &huge_boot_pages); | |
0d9ea754 | 348 | m->hstate = hstate; |
ec4b2c0c JT |
349 | return 1; |
350 | } | |
41151e77 | 351 | #endif |
ec4b2c0c | 352 | |
39dde65c CK |
353 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
354 | { | |
355 | return 0; | |
356 | } | |
357 | ||
41151e77 BB |
358 | #ifdef CONFIG_PPC32 |
359 | #define HUGEPD_FREELIST_SIZE \ | |
360 | ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t)) | |
361 | ||
362 | struct hugepd_freelist { | |
363 | struct rcu_head rcu; | |
364 | unsigned int index; | |
365 | void *ptes[0]; | |
366 | }; | |
367 | ||
368 | static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur); | |
369 | ||
370 | static void hugepd_free_rcu_callback(struct rcu_head *head) | |
371 | { | |
372 | struct hugepd_freelist *batch = | |
373 | container_of(head, struct hugepd_freelist, rcu); | |
374 | unsigned int i; | |
375 | ||
376 | for (i = 0; i < batch->index; i++) | |
377 | kmem_cache_free(hugepte_cache, batch->ptes[i]); | |
378 | ||
379 | free_page((unsigned long)batch); | |
380 | } | |
381 | ||
382 | static void hugepd_free(struct mmu_gather *tlb, void *hugepte) | |
383 | { | |
384 | struct hugepd_freelist **batchp; | |
385 | ||
386 | batchp = &__get_cpu_var(hugepd_freelist_cur); | |
387 | ||
388 | if (atomic_read(&tlb->mm->mm_users) < 2 || | |
389 | cpumask_equal(mm_cpumask(tlb->mm), | |
390 | cpumask_of(smp_processor_id()))) { | |
391 | kmem_cache_free(hugepte_cache, hugepte); | |
392 | return; | |
393 | } | |
394 | ||
395 | if (*batchp == NULL) { | |
396 | *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC); | |
397 | (*batchp)->index = 0; | |
398 | } | |
399 | ||
400 | (*batchp)->ptes[(*batchp)->index++] = hugepte; | |
401 | if ((*batchp)->index == HUGEPD_FREELIST_SIZE) { | |
402 | call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback); | |
403 | *batchp = NULL; | |
404 | } | |
405 | } | |
406 | #endif | |
407 | ||
a4fe3ce7 DG |
408 | static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift, |
409 | unsigned long start, unsigned long end, | |
410 | unsigned long floor, unsigned long ceiling) | |
f10a04c0 DG |
411 | { |
412 | pte_t *hugepte = hugepd_page(*hpdp); | |
41151e77 BB |
413 | int i; |
414 | ||
a4fe3ce7 | 415 | unsigned long pdmask = ~((1UL << pdshift) - 1); |
41151e77 BB |
416 | unsigned int num_hugepd = 1; |
417 | ||
418 | #ifdef CONFIG_PPC64 | |
419 | unsigned int shift = hugepd_shift(*hpdp); | |
420 | #else | |
421 | /* Note: On 32-bit the hpdp may be the first of several */ | |
422 | num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift)); | |
423 | #endif | |
a4fe3ce7 DG |
424 | |
425 | start &= pdmask; | |
426 | if (start < floor) | |
427 | return; | |
428 | if (ceiling) { | |
429 | ceiling &= pdmask; | |
430 | if (! ceiling) | |
431 | return; | |
432 | } | |
433 | if (end - 1 > ceiling - 1) | |
434 | return; | |
f10a04c0 | 435 | |
41151e77 BB |
436 | for (i = 0; i < num_hugepd; i++, hpdp++) |
437 | hpdp->pd = 0; | |
438 | ||
f10a04c0 | 439 | tlb->need_flush = 1; |
41151e77 | 440 | #ifdef CONFIG_PPC64 |
a4fe3ce7 | 441 | pgtable_free_tlb(tlb, hugepte, pdshift - shift); |
41151e77 BB |
442 | #else |
443 | hugepd_free(tlb, hugepte); | |
444 | #endif | |
f10a04c0 DG |
445 | } |
446 | ||
f10a04c0 DG |
447 | static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
448 | unsigned long addr, unsigned long end, | |
a4fe3ce7 | 449 | unsigned long floor, unsigned long ceiling) |
f10a04c0 DG |
450 | { |
451 | pmd_t *pmd; | |
452 | unsigned long next; | |
453 | unsigned long start; | |
454 | ||
455 | start = addr; | |
456 | pmd = pmd_offset(pud, addr); | |
457 | do { | |
458 | next = pmd_addr_end(addr, end); | |
459 | if (pmd_none(*pmd)) | |
460 | continue; | |
a4fe3ce7 DG |
461 | free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT, |
462 | addr, next, floor, ceiling); | |
f10a04c0 DG |
463 | } while (pmd++, addr = next, addr != end); |
464 | ||
465 | start &= PUD_MASK; | |
466 | if (start < floor) | |
467 | return; | |
468 | if (ceiling) { | |
469 | ceiling &= PUD_MASK; | |
470 | if (!ceiling) | |
471 | return; | |
1da177e4 | 472 | } |
f10a04c0 DG |
473 | if (end - 1 > ceiling - 1) |
474 | return; | |
1da177e4 | 475 | |
f10a04c0 DG |
476 | pmd = pmd_offset(pud, start); |
477 | pud_clear(pud); | |
9e1b32ca | 478 | pmd_free_tlb(tlb, pmd, start); |
f10a04c0 | 479 | } |
f10a04c0 DG |
480 | |
481 | static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
482 | unsigned long addr, unsigned long end, | |
483 | unsigned long floor, unsigned long ceiling) | |
484 | { | |
485 | pud_t *pud; | |
486 | unsigned long next; | |
487 | unsigned long start; | |
488 | ||
489 | start = addr; | |
490 | pud = pud_offset(pgd, addr); | |
491 | do { | |
492 | next = pud_addr_end(addr, end); | |
a4fe3ce7 | 493 | if (!is_hugepd(pud)) { |
4ec161cf JT |
494 | if (pud_none_or_clear_bad(pud)) |
495 | continue; | |
0d9ea754 | 496 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, |
a4fe3ce7 | 497 | ceiling); |
4ec161cf | 498 | } else { |
a4fe3ce7 DG |
499 | free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT, |
500 | addr, next, floor, ceiling); | |
4ec161cf | 501 | } |
f10a04c0 DG |
502 | } while (pud++, addr = next, addr != end); |
503 | ||
504 | start &= PGDIR_MASK; | |
505 | if (start < floor) | |
506 | return; | |
507 | if (ceiling) { | |
508 | ceiling &= PGDIR_MASK; | |
509 | if (!ceiling) | |
510 | return; | |
511 | } | |
512 | if (end - 1 > ceiling - 1) | |
513 | return; | |
514 | ||
515 | pud = pud_offset(pgd, start); | |
516 | pgd_clear(pgd); | |
9e1b32ca | 517 | pud_free_tlb(tlb, pud, start); |
f10a04c0 DG |
518 | } |
519 | ||
520 | /* | |
521 | * This function frees user-level page tables of a process. | |
522 | * | |
523 | * Must be called with pagetable lock held. | |
524 | */ | |
42b77728 | 525 | void hugetlb_free_pgd_range(struct mmu_gather *tlb, |
f10a04c0 DG |
526 | unsigned long addr, unsigned long end, |
527 | unsigned long floor, unsigned long ceiling) | |
528 | { | |
529 | pgd_t *pgd; | |
530 | unsigned long next; | |
f10a04c0 DG |
531 | |
532 | /* | |
a4fe3ce7 DG |
533 | * Because there are a number of different possible pagetable |
534 | * layouts for hugepage ranges, we limit knowledge of how | |
535 | * things should be laid out to the allocation path | |
536 | * (huge_pte_alloc(), above). Everything else works out the | |
537 | * structure as it goes from information in the hugepd | |
538 | * pointers. That means that we can't here use the | |
539 | * optimization used in the normal page free_pgd_range(), of | |
540 | * checking whether we're actually covering a large enough | |
541 | * range to have to do anything at the top level of the walk | |
542 | * instead of at the bottom. | |
f10a04c0 | 543 | * |
a4fe3ce7 DG |
544 | * To make sense of this, you should probably go read the big |
545 | * block comment at the top of the normal free_pgd_range(), | |
546 | * too. | |
f10a04c0 | 547 | */ |
f10a04c0 | 548 | |
f10a04c0 | 549 | do { |
f10a04c0 | 550 | next = pgd_addr_end(addr, end); |
41151e77 | 551 | pgd = pgd_offset(tlb->mm, addr); |
a4fe3ce7 | 552 | if (!is_hugepd(pgd)) { |
0b26425c DG |
553 | if (pgd_none_or_clear_bad(pgd)) |
554 | continue; | |
555 | hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); | |
556 | } else { | |
41151e77 BB |
557 | #ifdef CONFIG_PPC32 |
558 | /* | |
559 | * Increment next by the size of the huge mapping since | |
560 | * on 32-bit there may be more than one entry at the pgd | |
561 | * level for a 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 *)pgd)); | |
565 | #endif | |
a4fe3ce7 DG |
566 | free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT, |
567 | addr, next, floor, ceiling); | |
0b26425c | 568 | } |
41151e77 | 569 | } while (addr = next, addr != end); |
1da177e4 LT |
570 | } |
571 | ||
1da177e4 LT |
572 | struct page * |
573 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) | |
574 | { | |
575 | pte_t *ptep; | |
576 | struct page *page; | |
a4fe3ce7 DG |
577 | unsigned shift; |
578 | unsigned long mask; | |
579 | ||
580 | ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift); | |
1da177e4 | 581 | |
0d9ea754 | 582 | /* Verify it is a huge page else bail. */ |
a4fe3ce7 | 583 | if (!ptep || !shift) |
1da177e4 LT |
584 | return ERR_PTR(-EINVAL); |
585 | ||
a4fe3ce7 | 586 | mask = (1UL << shift) - 1; |
1da177e4 | 587 | page = pte_page(*ptep); |
a4fe3ce7 DG |
588 | if (page) |
589 | page += (address & mask) / PAGE_SIZE; | |
1da177e4 LT |
590 | |
591 | return page; | |
592 | } | |
593 | ||
594 | int pmd_huge(pmd_t pmd) | |
595 | { | |
596 | return 0; | |
597 | } | |
598 | ||
ceb86879 AK |
599 | int pud_huge(pud_t pud) |
600 | { | |
601 | return 0; | |
602 | } | |
603 | ||
1da177e4 LT |
604 | struct page * |
605 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
606 | pmd_t *pmd, int write) | |
607 | { | |
608 | BUG(); | |
609 | return NULL; | |
610 | } | |
611 | ||
a4fe3ce7 DG |
612 | static noinline int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, |
613 | unsigned long end, int write, struct page **pages, int *nr) | |
614 | { | |
615 | unsigned long mask; | |
616 | unsigned long pte_end; | |
617 | struct page *head, *page; | |
618 | pte_t pte; | |
619 | int refs; | |
620 | ||
621 | pte_end = (addr + sz) & ~(sz-1); | |
622 | if (pte_end < end) | |
623 | end = pte_end; | |
624 | ||
625 | pte = *ptep; | |
626 | mask = _PAGE_PRESENT | _PAGE_USER; | |
627 | if (write) | |
628 | mask |= _PAGE_RW; | |
629 | ||
630 | if ((pte_val(pte) & mask) != mask) | |
631 | return 0; | |
632 | ||
633 | /* hugepages are never "special" */ | |
634 | VM_BUG_ON(!pfn_valid(pte_pfn(pte))); | |
635 | ||
636 | refs = 0; | |
637 | head = pte_page(pte); | |
638 | ||
639 | page = head + ((addr & (sz-1)) >> PAGE_SHIFT); | |
640 | do { | |
641 | VM_BUG_ON(compound_head(page) != head); | |
642 | pages[*nr] = page; | |
643 | (*nr)++; | |
644 | page++; | |
645 | refs++; | |
646 | } while (addr += PAGE_SIZE, addr != end); | |
647 | ||
648 | if (!page_cache_add_speculative(head, refs)) { | |
649 | *nr -= refs; | |
650 | return 0; | |
651 | } | |
652 | ||
653 | if (unlikely(pte_val(pte) != pte_val(*ptep))) { | |
654 | /* Could be optimized better */ | |
655 | while (*nr) { | |
656 | put_page(page); | |
657 | (*nr)--; | |
658 | } | |
659 | } | |
660 | ||
661 | return 1; | |
662 | } | |
663 | ||
39adfa54 DG |
664 | static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, |
665 | unsigned long sz) | |
666 | { | |
667 | unsigned long __boundary = (addr + sz) & ~(sz-1); | |
668 | return (__boundary - 1 < end - 1) ? __boundary : end; | |
669 | } | |
670 | ||
a4fe3ce7 DG |
671 | int gup_hugepd(hugepd_t *hugepd, unsigned pdshift, |
672 | unsigned long addr, unsigned long end, | |
673 | int write, struct page **pages, int *nr) | |
674 | { | |
675 | pte_t *ptep; | |
676 | unsigned long sz = 1UL << hugepd_shift(*hugepd); | |
39adfa54 | 677 | unsigned long next; |
a4fe3ce7 DG |
678 | |
679 | ptep = hugepte_offset(hugepd, addr, pdshift); | |
680 | do { | |
39adfa54 | 681 | next = hugepte_addr_end(addr, end, sz); |
a4fe3ce7 DG |
682 | if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr)) |
683 | return 0; | |
39adfa54 | 684 | } while (ptep++, addr = next, addr != end); |
a4fe3ce7 DG |
685 | |
686 | return 1; | |
687 | } | |
1da177e4 LT |
688 | |
689 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, | |
690 | unsigned long len, unsigned long pgoff, | |
691 | unsigned long flags) | |
692 | { | |
41151e77 | 693 | #ifdef CONFIG_MM_SLICES |
0d9ea754 JT |
694 | struct hstate *hstate = hstate_file(file); |
695 | int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate)); | |
48f797de | 696 | |
0d9ea754 | 697 | return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0); |
41151e77 BB |
698 | #else |
699 | return get_unmapped_area(file, addr, len, pgoff, flags); | |
700 | #endif | |
1da177e4 LT |
701 | } |
702 | ||
3340289d MG |
703 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
704 | { | |
41151e77 | 705 | #ifdef CONFIG_MM_SLICES |
3340289d MG |
706 | unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start); |
707 | ||
708 | return 1UL << mmu_psize_to_shift(psize); | |
41151e77 BB |
709 | #else |
710 | if (!is_vm_hugetlb_page(vma)) | |
711 | return PAGE_SIZE; | |
712 | ||
713 | return huge_page_size(hstate_vma(vma)); | |
714 | #endif | |
715 | } | |
716 | ||
717 | static inline bool is_power_of_4(unsigned long x) | |
718 | { | |
719 | if (is_power_of_2(x)) | |
720 | return (__ilog2(x) % 2) ? false : true; | |
721 | return false; | |
3340289d MG |
722 | } |
723 | ||
d1837cba | 724 | static int __init add_huge_page_size(unsigned long long size) |
4ec161cf | 725 | { |
d1837cba DG |
726 | int shift = __ffs(size); |
727 | int mmu_psize; | |
a4fe3ce7 | 728 | |
4ec161cf | 729 | /* Check that it is a page size supported by the hardware and |
d1837cba | 730 | * that it fits within pagetable and slice limits. */ |
41151e77 BB |
731 | #ifdef CONFIG_PPC_FSL_BOOK3E |
732 | if ((size < PAGE_SIZE) || !is_power_of_4(size)) | |
733 | return -EINVAL; | |
734 | #else | |
d1837cba DG |
735 | if (!is_power_of_2(size) |
736 | || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT)) | |
737 | return -EINVAL; | |
41151e77 | 738 | #endif |
91224346 | 739 | |
d1837cba DG |
740 | if ((mmu_psize = shift_to_mmu_psize(shift)) < 0) |
741 | return -EINVAL; | |
742 | ||
743 | #ifdef CONFIG_SPU_FS_64K_LS | |
744 | /* Disable support for 64K huge pages when 64K SPU local store | |
745 | * support is enabled as the current implementation conflicts. | |
746 | */ | |
747 | if (shift == PAGE_SHIFT_64K) | |
748 | return -EINVAL; | |
749 | #endif /* CONFIG_SPU_FS_64K_LS */ | |
750 | ||
751 | BUG_ON(mmu_psize_defs[mmu_psize].shift != shift); | |
752 | ||
753 | /* Return if huge page size has already been setup */ | |
754 | if (size_to_hstate(size)) | |
755 | return 0; | |
756 | ||
757 | hugetlb_add_hstate(shift - PAGE_SHIFT); | |
758 | ||
759 | return 0; | |
4ec161cf JT |
760 | } |
761 | ||
762 | static int __init hugepage_setup_sz(char *str) | |
763 | { | |
764 | unsigned long long size; | |
4ec161cf JT |
765 | |
766 | size = memparse(str, &str); | |
767 | ||
d1837cba | 768 | if (add_huge_page_size(size) != 0) |
4ec161cf JT |
769 | printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size); |
770 | ||
771 | return 1; | |
772 | } | |
773 | __setup("hugepagesz=", hugepage_setup_sz); | |
774 | ||
41151e77 BB |
775 | #ifdef CONFIG_FSL_BOOKE |
776 | struct kmem_cache *hugepte_cache; | |
777 | static int __init hugetlbpage_init(void) | |
778 | { | |
779 | int psize; | |
780 | ||
781 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { | |
782 | unsigned shift; | |
783 | ||
784 | if (!mmu_psize_defs[psize].shift) | |
785 | continue; | |
786 | ||
787 | shift = mmu_psize_to_shift(psize); | |
788 | ||
789 | /* Don't treat normal page sizes as huge... */ | |
790 | if (shift != PAGE_SHIFT) | |
791 | if (add_huge_page_size(1ULL << shift) < 0) | |
792 | continue; | |
793 | } | |
794 | ||
795 | /* | |
796 | * Create a kmem cache for hugeptes. The bottom bits in the pte have | |
797 | * size information encoded in them, so align them to allow this | |
798 | */ | |
799 | hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t), | |
800 | HUGEPD_SHIFT_MASK + 1, 0, NULL); | |
801 | if (hugepte_cache == NULL) | |
802 | panic("%s: Unable to create kmem cache for hugeptes\n", | |
803 | __func__); | |
804 | ||
805 | /* Default hpage size = 4M */ | |
806 | if (mmu_psize_defs[MMU_PAGE_4M].shift) | |
807 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift; | |
808 | else | |
809 | panic("%s: Unable to set default huge page size\n", __func__); | |
810 | ||
811 | ||
812 | return 0; | |
813 | } | |
814 | #else | |
f10a04c0 DG |
815 | static int __init hugetlbpage_init(void) |
816 | { | |
a4fe3ce7 | 817 | int psize; |
0d9ea754 | 818 | |
44ae3ab3 | 819 | if (!mmu_has_feature(MMU_FTR_16M_PAGE)) |
f10a04c0 | 820 | return -ENODEV; |
00df438e | 821 | |
d1837cba DG |
822 | for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { |
823 | unsigned shift; | |
824 | unsigned pdshift; | |
0d9ea754 | 825 | |
d1837cba DG |
826 | if (!mmu_psize_defs[psize].shift) |
827 | continue; | |
00df438e | 828 | |
d1837cba DG |
829 | shift = mmu_psize_to_shift(psize); |
830 | ||
831 | if (add_huge_page_size(1ULL << shift) < 0) | |
832 | continue; | |
833 | ||
834 | if (shift < PMD_SHIFT) | |
835 | pdshift = PMD_SHIFT; | |
836 | else if (shift < PUD_SHIFT) | |
837 | pdshift = PUD_SHIFT; | |
838 | else | |
839 | pdshift = PGDIR_SHIFT; | |
840 | ||
841 | pgtable_cache_add(pdshift - shift, NULL); | |
842 | if (!PGT_CACHE(pdshift - shift)) | |
843 | panic("hugetlbpage_init(): could not create " | |
844 | "pgtable cache for %d bit pagesize\n", shift); | |
0d9ea754 | 845 | } |
f10a04c0 | 846 | |
d1837cba DG |
847 | /* Set default large page size. Currently, we pick 16M or 1M |
848 | * depending on what is available | |
849 | */ | |
850 | if (mmu_psize_defs[MMU_PAGE_16M].shift) | |
851 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift; | |
852 | else if (mmu_psize_defs[MMU_PAGE_1M].shift) | |
853 | HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift; | |
854 | ||
f10a04c0 DG |
855 | return 0; |
856 | } | |
41151e77 | 857 | #endif |
f10a04c0 | 858 | module_init(hugetlbpage_init); |
0895ecda DG |
859 | |
860 | void flush_dcache_icache_hugepage(struct page *page) | |
861 | { | |
862 | int i; | |
41151e77 | 863 | void *start; |
0895ecda DG |
864 | |
865 | BUG_ON(!PageCompound(page)); | |
866 | ||
41151e77 BB |
867 | for (i = 0; i < (1UL << compound_order(page)); i++) { |
868 | if (!PageHighMem(page)) { | |
869 | __flush_dcache_icache(page_address(page+i)); | |
870 | } else { | |
871 | start = kmap_atomic(page+i, KM_PPC_SYNC_ICACHE); | |
872 | __flush_dcache_icache(start); | |
873 | kunmap_atomic(start, KM_PPC_SYNC_ICACHE); | |
874 | } | |
875 | } | |
0895ecda | 876 | } |