Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * PPC64 (POWER4) Huge TLB Page Support for Kernel. | |
3 | * | |
4 | * Copyright (C) 2003 David Gibson, IBM Corporation. | |
5 | * | |
6 | * Based on the IA-32 version: | |
7 | * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com> | |
8 | */ | |
9 | ||
10 | #include <linux/init.h> | |
11 | #include <linux/fs.h> | |
12 | #include <linux/mm.h> | |
13 | #include <linux/hugetlb.h> | |
14 | #include <linux/pagemap.h> | |
15 | #include <linux/smp_lock.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/err.h> | |
18 | #include <linux/sysctl.h> | |
19 | #include <asm/mman.h> | |
20 | #include <asm/pgalloc.h> | |
21 | #include <asm/tlb.h> | |
22 | #include <asm/tlbflush.h> | |
23 | #include <asm/mmu_context.h> | |
24 | #include <asm/machdep.h> | |
25 | #include <asm/cputable.h> | |
26 | #include <asm/tlb.h> | |
27 | ||
28 | #include <linux/sysctl.h> | |
29 | ||
c594adad DG |
30 | #define NUM_LOW_AREAS (0x100000000UL >> SID_SHIFT) |
31 | #define NUM_HIGH_AREAS (PGTABLE_RANGE >> HTLB_AREA_SHIFT) | |
32 | ||
f10a04c0 DG |
33 | #ifdef CONFIG_PPC_64K_PAGES |
34 | #define HUGEPTE_INDEX_SIZE (PMD_SHIFT-HPAGE_SHIFT) | |
35 | #else | |
36 | #define HUGEPTE_INDEX_SIZE (PUD_SHIFT-HPAGE_SHIFT) | |
37 | #endif | |
38 | #define PTRS_PER_HUGEPTE (1 << HUGEPTE_INDEX_SIZE) | |
39 | #define HUGEPTE_TABLE_SIZE (sizeof(pte_t) << HUGEPTE_INDEX_SIZE) | |
40 | ||
41 | #define HUGEPD_SHIFT (HPAGE_SHIFT + HUGEPTE_INDEX_SIZE) | |
42 | #define HUGEPD_SIZE (1UL << HUGEPD_SHIFT) | |
43 | #define HUGEPD_MASK (~(HUGEPD_SIZE-1)) | |
44 | ||
45 | #define huge_pgtable_cache (pgtable_cache[HUGEPTE_CACHE_NUM]) | |
46 | ||
47 | /* Flag to mark huge PD pointers. This means pmd_bad() and pud_bad() | |
48 | * will choke on pointers to hugepte tables, which is handy for | |
49 | * catching screwups early. */ | |
50 | #define HUGEPD_OK 0x1 | |
51 | ||
52 | typedef struct { unsigned long pd; } hugepd_t; | |
53 | ||
54 | #define hugepd_none(hpd) ((hpd).pd == 0) | |
55 | ||
56 | static inline pte_t *hugepd_page(hugepd_t hpd) | |
57 | { | |
58 | BUG_ON(!(hpd.pd & HUGEPD_OK)); | |
59 | return (pte_t *)(hpd.pd & ~HUGEPD_OK); | |
60 | } | |
61 | ||
62 | static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr) | |
63 | { | |
64 | unsigned long idx = ((addr >> HPAGE_SHIFT) & (PTRS_PER_HUGEPTE-1)); | |
65 | pte_t *dir = hugepd_page(*hpdp); | |
66 | ||
67 | return dir + idx; | |
68 | } | |
69 | ||
70 | static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, | |
71 | unsigned long address) | |
72 | { | |
73 | pte_t *new = kmem_cache_alloc(huge_pgtable_cache, | |
74 | GFP_KERNEL|__GFP_REPEAT); | |
75 | ||
76 | if (! new) | |
77 | return -ENOMEM; | |
78 | ||
79 | spin_lock(&mm->page_table_lock); | |
80 | if (!hugepd_none(*hpdp)) | |
81 | kmem_cache_free(huge_pgtable_cache, new); | |
82 | else | |
83 | hpdp->pd = (unsigned long)new | HUGEPD_OK; | |
84 | spin_unlock(&mm->page_table_lock); | |
85 | return 0; | |
86 | } | |
87 | ||
e28f7faf DG |
88 | /* Modelled after find_linux_pte() */ |
89 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
1da177e4 | 90 | { |
e28f7faf DG |
91 | pgd_t *pg; |
92 | pud_t *pu; | |
1da177e4 | 93 | |
e28f7faf | 94 | BUG_ON(! in_hugepage_area(mm->context, addr)); |
1da177e4 | 95 | |
e28f7faf DG |
96 | addr &= HPAGE_MASK; |
97 | ||
98 | pg = pgd_offset(mm, addr); | |
99 | if (!pgd_none(*pg)) { | |
100 | pu = pud_offset(pg, addr); | |
101 | if (!pud_none(*pu)) { | |
3c726f8d | 102 | #ifdef CONFIG_PPC_64K_PAGES |
f10a04c0 DG |
103 | pmd_t *pm; |
104 | pm = pmd_offset(pu, addr); | |
105 | if (!pmd_none(*pm)) | |
106 | return hugepte_offset((hugepd_t *)pm, addr); | |
107 | #else | |
108 | return hugepte_offset((hugepd_t *)pu, addr); | |
109 | #endif | |
e28f7faf DG |
110 | } |
111 | } | |
1da177e4 | 112 | |
e28f7faf | 113 | return NULL; |
1da177e4 LT |
114 | } |
115 | ||
e28f7faf | 116 | pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr) |
1da177e4 | 117 | { |
e28f7faf DG |
118 | pgd_t *pg; |
119 | pud_t *pu; | |
f10a04c0 | 120 | hugepd_t *hpdp = NULL; |
1da177e4 | 121 | |
1da177e4 LT |
122 | BUG_ON(! in_hugepage_area(mm->context, addr)); |
123 | ||
e28f7faf | 124 | addr &= HPAGE_MASK; |
1da177e4 | 125 | |
e28f7faf DG |
126 | pg = pgd_offset(mm, addr); |
127 | pu = pud_alloc(mm, pg, addr); | |
1da177e4 | 128 | |
e28f7faf | 129 | if (pu) { |
f10a04c0 DG |
130 | #ifdef CONFIG_PPC_64K_PAGES |
131 | pmd_t *pm; | |
e28f7faf | 132 | pm = pmd_alloc(mm, pu, addr); |
f10a04c0 DG |
133 | if (pm) |
134 | hpdp = (hugepd_t *)pm; | |
135 | #else | |
136 | hpdp = (hugepd_t *)pu; | |
137 | #endif | |
138 | } | |
139 | ||
140 | if (! hpdp) | |
141 | return NULL; | |
142 | ||
143 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr)) | |
144 | return NULL; | |
145 | ||
146 | return hugepte_offset(hpdp, addr); | |
147 | } | |
148 | ||
149 | static void free_hugepte_range(struct mmu_gather *tlb, hugepd_t *hpdp) | |
150 | { | |
151 | pte_t *hugepte = hugepd_page(*hpdp); | |
152 | ||
153 | hpdp->pd = 0; | |
154 | tlb->need_flush = 1; | |
155 | pgtable_free_tlb(tlb, pgtable_free_cache(hugepte, HUGEPTE_CACHE_NUM, | |
156 | HUGEPTE_TABLE_SIZE-1)); | |
157 | } | |
158 | ||
3c726f8d | 159 | #ifdef CONFIG_PPC_64K_PAGES |
f10a04c0 DG |
160 | static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
161 | unsigned long addr, unsigned long end, | |
162 | unsigned long floor, unsigned long ceiling) | |
163 | { | |
164 | pmd_t *pmd; | |
165 | unsigned long next; | |
166 | unsigned long start; | |
167 | ||
168 | start = addr; | |
169 | pmd = pmd_offset(pud, addr); | |
170 | do { | |
171 | next = pmd_addr_end(addr, end); | |
172 | if (pmd_none(*pmd)) | |
173 | continue; | |
174 | free_hugepte_range(tlb, (hugepd_t *)pmd); | |
175 | } while (pmd++, addr = next, addr != end); | |
176 | ||
177 | start &= PUD_MASK; | |
178 | if (start < floor) | |
179 | return; | |
180 | if (ceiling) { | |
181 | ceiling &= PUD_MASK; | |
182 | if (!ceiling) | |
183 | return; | |
1da177e4 | 184 | } |
f10a04c0 DG |
185 | if (end - 1 > ceiling - 1) |
186 | return; | |
1da177e4 | 187 | |
f10a04c0 DG |
188 | pmd = pmd_offset(pud, start); |
189 | pud_clear(pud); | |
190 | pmd_free_tlb(tlb, pmd); | |
191 | } | |
192 | #endif | |
193 | ||
194 | static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
195 | unsigned long addr, unsigned long end, | |
196 | unsigned long floor, unsigned long ceiling) | |
197 | { | |
198 | pud_t *pud; | |
199 | unsigned long next; | |
200 | unsigned long start; | |
201 | ||
202 | start = addr; | |
203 | pud = pud_offset(pgd, addr); | |
204 | do { | |
205 | next = pud_addr_end(addr, end); | |
206 | #ifdef CONFIG_PPC_64K_PAGES | |
207 | if (pud_none_or_clear_bad(pud)) | |
208 | continue; | |
209 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling); | |
210 | #else | |
211 | if (pud_none(*pud)) | |
212 | continue; | |
213 | free_hugepte_range(tlb, (hugepd_t *)pud); | |
214 | #endif | |
215 | } while (pud++, addr = next, addr != end); | |
216 | ||
217 | start &= PGDIR_MASK; | |
218 | if (start < floor) | |
219 | return; | |
220 | if (ceiling) { | |
221 | ceiling &= PGDIR_MASK; | |
222 | if (!ceiling) | |
223 | return; | |
224 | } | |
225 | if (end - 1 > ceiling - 1) | |
226 | return; | |
227 | ||
228 | pud = pud_offset(pgd, start); | |
229 | pgd_clear(pgd); | |
230 | pud_free_tlb(tlb, pud); | |
231 | } | |
232 | ||
233 | /* | |
234 | * This function frees user-level page tables of a process. | |
235 | * | |
236 | * Must be called with pagetable lock held. | |
237 | */ | |
238 | void hugetlb_free_pgd_range(struct mmu_gather **tlb, | |
239 | unsigned long addr, unsigned long end, | |
240 | unsigned long floor, unsigned long ceiling) | |
241 | { | |
242 | pgd_t *pgd; | |
243 | unsigned long next; | |
244 | unsigned long start; | |
245 | ||
246 | /* | |
247 | * Comments below take from the normal free_pgd_range(). They | |
248 | * apply here too. The tests against HUGEPD_MASK below are | |
249 | * essential, because we *don't* test for this at the bottom | |
250 | * level. Without them we'll attempt to free a hugepte table | |
251 | * when we unmap just part of it, even if there are other | |
252 | * active mappings using it. | |
253 | * | |
254 | * The next few lines have given us lots of grief... | |
255 | * | |
256 | * Why are we testing HUGEPD* at this top level? Because | |
257 | * often there will be no work to do at all, and we'd prefer | |
258 | * not to go all the way down to the bottom just to discover | |
259 | * that. | |
260 | * | |
261 | * Why all these "- 1"s? Because 0 represents both the bottom | |
262 | * of the address space and the top of it (using -1 for the | |
263 | * top wouldn't help much: the masks would do the wrong thing). | |
264 | * The rule is that addr 0 and floor 0 refer to the bottom of | |
265 | * the address space, but end 0 and ceiling 0 refer to the top | |
266 | * Comparisons need to use "end - 1" and "ceiling - 1" (though | |
267 | * that end 0 case should be mythical). | |
268 | * | |
269 | * Wherever addr is brought up or ceiling brought down, we | |
270 | * must be careful to reject "the opposite 0" before it | |
271 | * confuses the subsequent tests. But what about where end is | |
272 | * brought down by HUGEPD_SIZE below? no, end can't go down to | |
273 | * 0 there. | |
274 | * | |
275 | * Whereas we round start (addr) and ceiling down, by different | |
276 | * masks at different levels, in order to test whether a table | |
277 | * now has no other vmas using it, so can be freed, we don't | |
278 | * bother to round floor or end up - the tests don't need that. | |
279 | */ | |
280 | ||
281 | addr &= HUGEPD_MASK; | |
282 | if (addr < floor) { | |
283 | addr += HUGEPD_SIZE; | |
284 | if (!addr) | |
285 | return; | |
286 | } | |
287 | if (ceiling) { | |
288 | ceiling &= HUGEPD_MASK; | |
289 | if (!ceiling) | |
290 | return; | |
291 | } | |
292 | if (end - 1 > ceiling - 1) | |
293 | end -= HUGEPD_SIZE; | |
294 | if (addr > end - 1) | |
295 | return; | |
296 | ||
297 | start = addr; | |
298 | pgd = pgd_offset((*tlb)->mm, addr); | |
299 | do { | |
300 | BUG_ON(! in_hugepage_area((*tlb)->mm->context, addr)); | |
301 | next = pgd_addr_end(addr, end); | |
302 | if (pgd_none_or_clear_bad(pgd)) | |
303 | continue; | |
304 | hugetlb_free_pud_range(*tlb, pgd, addr, next, floor, ceiling); | |
305 | } while (pgd++, addr = next, addr != end); | |
1da177e4 LT |
306 | } |
307 | ||
e28f7faf DG |
308 | void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, |
309 | pte_t *ptep, pte_t pte) | |
310 | { | |
e28f7faf | 311 | if (pte_present(*ptep)) { |
3c726f8d BH |
312 | /* We open-code pte_clear because we need to pass the right |
313 | * argument to hpte_update (huge / !huge) | |
314 | */ | |
315 | unsigned long old = pte_update(ptep, ~0UL); | |
316 | if (old & _PAGE_HASHPTE) | |
317 | hpte_update(mm, addr & HPAGE_MASK, ptep, old, 1); | |
e28f7faf DG |
318 | flush_tlb_pending(); |
319 | } | |
3c726f8d | 320 | *ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS); |
1da177e4 LT |
321 | } |
322 | ||
e28f7faf DG |
323 | pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, |
324 | pte_t *ptep) | |
1da177e4 | 325 | { |
e28f7faf | 326 | unsigned long old = pte_update(ptep, ~0UL); |
1da177e4 | 327 | |
e28f7faf | 328 | if (old & _PAGE_HASHPTE) |
3c726f8d BH |
329 | hpte_update(mm, addr & HPAGE_MASK, ptep, old, 1); |
330 | *ptep = __pte(0); | |
1da177e4 | 331 | |
e28f7faf | 332 | return __pte(old); |
1da177e4 LT |
333 | } |
334 | ||
23ed6cb9 DG |
335 | struct slb_flush_info { |
336 | struct mm_struct *mm; | |
337 | u16 newareas; | |
338 | }; | |
339 | ||
c594adad | 340 | static void flush_low_segments(void *parm) |
1da177e4 | 341 | { |
23ed6cb9 | 342 | struct slb_flush_info *fi = parm; |
1da177e4 LT |
343 | unsigned long i; |
344 | ||
23ed6cb9 DG |
345 | BUILD_BUG_ON((sizeof(fi->newareas)*8) != NUM_LOW_AREAS); |
346 | ||
347 | if (current->active_mm != fi->mm) | |
348 | return; | |
1da177e4 | 349 | |
23ed6cb9 DG |
350 | /* Only need to do anything if this CPU is working in the same |
351 | * mm as the one which has changed */ | |
352 | ||
353 | /* update the paca copy of the context struct */ | |
354 | get_paca()->context = current->active_mm->context; | |
c594adad | 355 | |
23ed6cb9 | 356 | asm volatile("isync" : : : "memory"); |
c594adad | 357 | for (i = 0; i < NUM_LOW_AREAS; i++) { |
23ed6cb9 | 358 | if (! (fi->newareas & (1U << i))) |
1da177e4 | 359 | continue; |
14b34661 DG |
360 | asm volatile("slbie %0" |
361 | : : "r" ((i << SID_SHIFT) | SLBIE_C)); | |
1da177e4 | 362 | } |
1da177e4 LT |
363 | asm volatile("isync" : : : "memory"); |
364 | } | |
365 | ||
c594adad DG |
366 | static void flush_high_segments(void *parm) |
367 | { | |
23ed6cb9 | 368 | struct slb_flush_info *fi = parm; |
c594adad DG |
369 | unsigned long i, j; |
370 | ||
c594adad | 371 | |
23ed6cb9 DG |
372 | BUILD_BUG_ON((sizeof(fi->newareas)*8) != NUM_HIGH_AREAS); |
373 | ||
374 | if (current->active_mm != fi->mm) | |
375 | return; | |
376 | ||
377 | /* Only need to do anything if this CPU is working in the same | |
378 | * mm as the one which has changed */ | |
c594adad | 379 | |
23ed6cb9 DG |
380 | /* update the paca copy of the context struct */ |
381 | get_paca()->context = current->active_mm->context; | |
382 | ||
383 | asm volatile("isync" : : : "memory"); | |
c594adad | 384 | for (i = 0; i < NUM_HIGH_AREAS; i++) { |
23ed6cb9 | 385 | if (! (fi->newareas & (1U << i))) |
c594adad DG |
386 | continue; |
387 | for (j = 0; j < (1UL << (HTLB_AREA_SHIFT-SID_SHIFT)); j++) | |
388 | asm volatile("slbie %0" | |
14b34661 | 389 | :: "r" (((i << HTLB_AREA_SHIFT) |
23ed6cb9 | 390 | + (j << SID_SHIFT)) | SLBIE_C)); |
c594adad | 391 | } |
c594adad DG |
392 | asm volatile("isync" : : : "memory"); |
393 | } | |
394 | ||
395 | static int prepare_low_area_for_htlb(struct mm_struct *mm, unsigned long area) | |
1da177e4 | 396 | { |
c594adad DG |
397 | unsigned long start = area << SID_SHIFT; |
398 | unsigned long end = (area+1) << SID_SHIFT; | |
1da177e4 | 399 | struct vm_area_struct *vma; |
1da177e4 | 400 | |
c594adad | 401 | BUG_ON(area >= NUM_LOW_AREAS); |
1da177e4 LT |
402 | |
403 | /* Check no VMAs are in the region */ | |
404 | vma = find_vma(mm, start); | |
405 | if (vma && (vma->vm_start < end)) | |
406 | return -EBUSY; | |
407 | ||
1da177e4 LT |
408 | return 0; |
409 | } | |
410 | ||
c594adad DG |
411 | static int prepare_high_area_for_htlb(struct mm_struct *mm, unsigned long area) |
412 | { | |
413 | unsigned long start = area << HTLB_AREA_SHIFT; | |
414 | unsigned long end = (area+1) << HTLB_AREA_SHIFT; | |
415 | struct vm_area_struct *vma; | |
416 | ||
417 | BUG_ON(area >= NUM_HIGH_AREAS); | |
418 | ||
7d24f0b8 DG |
419 | /* Hack, so that each addresses is controlled by exactly one |
420 | * of the high or low area bitmaps, the first high area starts | |
421 | * at 4GB, not 0 */ | |
422 | if (start == 0) | |
423 | start = 0x100000000UL; | |
424 | ||
c594adad DG |
425 | /* Check no VMAs are in the region */ |
426 | vma = find_vma(mm, start); | |
427 | if (vma && (vma->vm_start < end)) | |
428 | return -EBUSY; | |
429 | ||
430 | return 0; | |
431 | } | |
432 | ||
433 | static int open_low_hpage_areas(struct mm_struct *mm, u16 newareas) | |
1da177e4 LT |
434 | { |
435 | unsigned long i; | |
23ed6cb9 | 436 | struct slb_flush_info fi; |
1da177e4 | 437 | |
c594adad DG |
438 | BUILD_BUG_ON((sizeof(newareas)*8) != NUM_LOW_AREAS); |
439 | BUILD_BUG_ON((sizeof(mm->context.low_htlb_areas)*8) != NUM_LOW_AREAS); | |
440 | ||
441 | newareas &= ~(mm->context.low_htlb_areas); | |
442 | if (! newareas) | |
1da177e4 LT |
443 | return 0; /* The segments we want are already open */ |
444 | ||
c594adad DG |
445 | for (i = 0; i < NUM_LOW_AREAS; i++) |
446 | if ((1 << i) & newareas) | |
447 | if (prepare_low_area_for_htlb(mm, i) != 0) | |
448 | return -EBUSY; | |
449 | ||
450 | mm->context.low_htlb_areas |= newareas; | |
451 | ||
c594adad DG |
452 | /* the context change must make it to memory before the flush, |
453 | * so that further SLB misses do the right thing. */ | |
454 | mb(); | |
23ed6cb9 DG |
455 | |
456 | fi.mm = mm; | |
457 | fi.newareas = newareas; | |
458 | on_each_cpu(flush_low_segments, &fi, 0, 1); | |
c594adad DG |
459 | |
460 | return 0; | |
461 | } | |
462 | ||
463 | static int open_high_hpage_areas(struct mm_struct *mm, u16 newareas) | |
464 | { | |
23ed6cb9 | 465 | struct slb_flush_info fi; |
c594adad DG |
466 | unsigned long i; |
467 | ||
468 | BUILD_BUG_ON((sizeof(newareas)*8) != NUM_HIGH_AREAS); | |
469 | BUILD_BUG_ON((sizeof(mm->context.high_htlb_areas)*8) | |
470 | != NUM_HIGH_AREAS); | |
471 | ||
472 | newareas &= ~(mm->context.high_htlb_areas); | |
473 | if (! newareas) | |
474 | return 0; /* The areas we want are already open */ | |
475 | ||
476 | for (i = 0; i < NUM_HIGH_AREAS; i++) | |
477 | if ((1 << i) & newareas) | |
478 | if (prepare_high_area_for_htlb(mm, i) != 0) | |
1da177e4 LT |
479 | return -EBUSY; |
480 | ||
c594adad | 481 | mm->context.high_htlb_areas |= newareas; |
1da177e4 LT |
482 | |
483 | /* update the paca copy of the context struct */ | |
484 | get_paca()->context = mm->context; | |
485 | ||
486 | /* the context change must make it to memory before the flush, | |
487 | * so that further SLB misses do the right thing. */ | |
488 | mb(); | |
23ed6cb9 DG |
489 | |
490 | fi.mm = mm; | |
491 | fi.newareas = newareas; | |
492 | on_each_cpu(flush_high_segments, &fi, 0, 1); | |
1da177e4 LT |
493 | |
494 | return 0; | |
495 | } | |
496 | ||
497 | int prepare_hugepage_range(unsigned long addr, unsigned long len) | |
498 | { | |
5e391dc9 | 499 | int err = 0; |
c594adad DG |
500 | |
501 | if ( (addr+len) < addr ) | |
502 | return -EINVAL; | |
503 | ||
5e391dc9 | 504 | if (addr < 0x100000000UL) |
c594adad | 505 | err = open_low_hpage_areas(current->mm, |
1da177e4 | 506 | LOW_ESID_MASK(addr, len)); |
9a94c579 | 507 | if ((addr + len) > 0x100000000UL) |
c594adad DG |
508 | err = open_high_hpage_areas(current->mm, |
509 | HTLB_AREA_MASK(addr, len)); | |
510 | if (err) { | |
511 | printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)" | |
512 | " failed (lowmask: 0x%04hx, highmask: 0x%04hx)\n", | |
513 | addr, len, | |
514 | LOW_ESID_MASK(addr, len), HTLB_AREA_MASK(addr, len)); | |
1da177e4 LT |
515 | return err; |
516 | } | |
517 | ||
c594adad | 518 | return 0; |
1da177e4 LT |
519 | } |
520 | ||
1da177e4 LT |
521 | struct page * |
522 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) | |
523 | { | |
524 | pte_t *ptep; | |
525 | struct page *page; | |
526 | ||
527 | if (! in_hugepage_area(mm->context, address)) | |
528 | return ERR_PTR(-EINVAL); | |
529 | ||
530 | ptep = huge_pte_offset(mm, address); | |
531 | page = pte_page(*ptep); | |
532 | if (page) | |
533 | page += (address % HPAGE_SIZE) / PAGE_SIZE; | |
534 | ||
535 | return page; | |
536 | } | |
537 | ||
538 | int pmd_huge(pmd_t pmd) | |
539 | { | |
540 | return 0; | |
541 | } | |
542 | ||
543 | struct page * | |
544 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
545 | pmd_t *pmd, int write) | |
546 | { | |
547 | BUG(); | |
548 | return NULL; | |
549 | } | |
550 | ||
1da177e4 LT |
551 | /* Because we have an exclusive hugepage region which lies within the |
552 | * normal user address space, we have to take special measures to make | |
553 | * non-huge mmap()s evade the hugepage reserved regions. */ | |
554 | unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr, | |
555 | unsigned long len, unsigned long pgoff, | |
556 | unsigned long flags) | |
557 | { | |
558 | struct mm_struct *mm = current->mm; | |
559 | struct vm_area_struct *vma; | |
560 | unsigned long start_addr; | |
561 | ||
562 | if (len > TASK_SIZE) | |
563 | return -ENOMEM; | |
564 | ||
565 | if (addr) { | |
566 | addr = PAGE_ALIGN(addr); | |
567 | vma = find_vma(mm, addr); | |
568 | if (((TASK_SIZE - len) >= addr) | |
569 | && (!vma || (addr+len) <= vma->vm_start) | |
570 | && !is_hugepage_only_range(mm, addr,len)) | |
571 | return addr; | |
572 | } | |
1363c3cd WW |
573 | if (len > mm->cached_hole_size) { |
574 | start_addr = addr = mm->free_area_cache; | |
575 | } else { | |
576 | start_addr = addr = TASK_UNMAPPED_BASE; | |
577 | mm->cached_hole_size = 0; | |
578 | } | |
1da177e4 LT |
579 | |
580 | full_search: | |
581 | vma = find_vma(mm, addr); | |
582 | while (TASK_SIZE - len >= addr) { | |
583 | BUG_ON(vma && (addr >= vma->vm_end)); | |
584 | ||
585 | if (touches_hugepage_low_range(mm, addr, len)) { | |
586 | addr = ALIGN(addr+1, 1<<SID_SHIFT); | |
587 | vma = find_vma(mm, addr); | |
588 | continue; | |
589 | } | |
c594adad DG |
590 | if (touches_hugepage_high_range(mm, addr, len)) { |
591 | addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT); | |
1da177e4 LT |
592 | vma = find_vma(mm, addr); |
593 | continue; | |
594 | } | |
595 | if (!vma || addr + len <= vma->vm_start) { | |
596 | /* | |
597 | * Remember the place where we stopped the search: | |
598 | */ | |
599 | mm->free_area_cache = addr + len; | |
600 | return addr; | |
601 | } | |
1363c3cd WW |
602 | if (addr + mm->cached_hole_size < vma->vm_start) |
603 | mm->cached_hole_size = vma->vm_start - addr; | |
1da177e4 LT |
604 | addr = vma->vm_end; |
605 | vma = vma->vm_next; | |
606 | } | |
607 | ||
608 | /* Make sure we didn't miss any holes */ | |
609 | if (start_addr != TASK_UNMAPPED_BASE) { | |
610 | start_addr = addr = TASK_UNMAPPED_BASE; | |
1363c3cd | 611 | mm->cached_hole_size = 0; |
1da177e4 LT |
612 | goto full_search; |
613 | } | |
614 | return -ENOMEM; | |
615 | } | |
616 | ||
617 | /* | |
618 | * This mmap-allocator allocates new areas top-down from below the | |
619 | * stack's low limit (the base): | |
620 | * | |
621 | * Because we have an exclusive hugepage region which lies within the | |
622 | * normal user address space, we have to take special measures to make | |
623 | * non-huge mmap()s evade the hugepage reserved regions. | |
624 | */ | |
625 | unsigned long | |
626 | arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, | |
627 | const unsigned long len, const unsigned long pgoff, | |
628 | const unsigned long flags) | |
629 | { | |
630 | struct vm_area_struct *vma, *prev_vma; | |
631 | struct mm_struct *mm = current->mm; | |
632 | unsigned long base = mm->mmap_base, addr = addr0; | |
1363c3cd | 633 | unsigned long largest_hole = mm->cached_hole_size; |
1da177e4 LT |
634 | int first_time = 1; |
635 | ||
636 | /* requested length too big for entire address space */ | |
637 | if (len > TASK_SIZE) | |
638 | return -ENOMEM; | |
639 | ||
640 | /* dont allow allocations above current base */ | |
641 | if (mm->free_area_cache > base) | |
642 | mm->free_area_cache = base; | |
643 | ||
644 | /* requesting a specific address */ | |
645 | if (addr) { | |
646 | addr = PAGE_ALIGN(addr); | |
647 | vma = find_vma(mm, addr); | |
648 | if (TASK_SIZE - len >= addr && | |
649 | (!vma || addr + len <= vma->vm_start) | |
650 | && !is_hugepage_only_range(mm, addr,len)) | |
651 | return addr; | |
652 | } | |
653 | ||
1363c3cd WW |
654 | if (len <= largest_hole) { |
655 | largest_hole = 0; | |
656 | mm->free_area_cache = base; | |
657 | } | |
1da177e4 LT |
658 | try_again: |
659 | /* make sure it can fit in the remaining address space */ | |
660 | if (mm->free_area_cache < len) | |
661 | goto fail; | |
662 | ||
663 | /* either no address requested or cant fit in requested address hole */ | |
664 | addr = (mm->free_area_cache - len) & PAGE_MASK; | |
665 | do { | |
666 | hugepage_recheck: | |
667 | if (touches_hugepage_low_range(mm, addr, len)) { | |
668 | addr = (addr & ((~0) << SID_SHIFT)) - len; | |
669 | goto hugepage_recheck; | |
c594adad DG |
670 | } else if (touches_hugepage_high_range(mm, addr, len)) { |
671 | addr = (addr & ((~0UL) << HTLB_AREA_SHIFT)) - len; | |
672 | goto hugepage_recheck; | |
1da177e4 LT |
673 | } |
674 | ||
675 | /* | |
676 | * Lookup failure means no vma is above this address, | |
677 | * i.e. return with success: | |
678 | */ | |
679 | if (!(vma = find_vma_prev(mm, addr, &prev_vma))) | |
680 | return addr; | |
681 | ||
682 | /* | |
683 | * new region fits between prev_vma->vm_end and | |
684 | * vma->vm_start, use it: | |
685 | */ | |
686 | if (addr+len <= vma->vm_start && | |
1363c3cd | 687 | (!prev_vma || (addr >= prev_vma->vm_end))) { |
1da177e4 | 688 | /* remember the address as a hint for next time */ |
1363c3cd WW |
689 | mm->cached_hole_size = largest_hole; |
690 | return (mm->free_area_cache = addr); | |
691 | } else { | |
1da177e4 | 692 | /* pull free_area_cache down to the first hole */ |
1363c3cd | 693 | if (mm->free_area_cache == vma->vm_end) { |
1da177e4 | 694 | mm->free_area_cache = vma->vm_start; |
1363c3cd WW |
695 | mm->cached_hole_size = largest_hole; |
696 | } | |
697 | } | |
698 | ||
699 | /* remember the largest hole we saw so far */ | |
700 | if (addr + largest_hole < vma->vm_start) | |
701 | largest_hole = vma->vm_start - addr; | |
1da177e4 LT |
702 | |
703 | /* try just below the current vma->vm_start */ | |
704 | addr = vma->vm_start-len; | |
705 | } while (len <= vma->vm_start); | |
706 | ||
707 | fail: | |
708 | /* | |
709 | * if hint left us with no space for the requested | |
710 | * mapping then try again: | |
711 | */ | |
712 | if (first_time) { | |
713 | mm->free_area_cache = base; | |
1363c3cd | 714 | largest_hole = 0; |
1da177e4 LT |
715 | first_time = 0; |
716 | goto try_again; | |
717 | } | |
718 | /* | |
719 | * A failed mmap() very likely causes application failure, | |
720 | * so fall back to the bottom-up function here. This scenario | |
721 | * can happen with large stack limits and large mmap() | |
722 | * allocations. | |
723 | */ | |
724 | mm->free_area_cache = TASK_UNMAPPED_BASE; | |
1363c3cd | 725 | mm->cached_hole_size = ~0UL; |
1da177e4 LT |
726 | addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); |
727 | /* | |
728 | * Restore the topdown base: | |
729 | */ | |
730 | mm->free_area_cache = base; | |
1363c3cd | 731 | mm->cached_hole_size = ~0UL; |
1da177e4 LT |
732 | |
733 | return addr; | |
734 | } | |
735 | ||
456752f7 DG |
736 | static int htlb_check_hinted_area(unsigned long addr, unsigned long len) |
737 | { | |
738 | struct vm_area_struct *vma; | |
739 | ||
740 | vma = find_vma(current->mm, addr); | |
741 | if (!vma || ((addr + len) <= vma->vm_start)) | |
742 | return 0; | |
743 | ||
744 | return -ENOMEM; | |
745 | } | |
746 | ||
1da177e4 LT |
747 | static unsigned long htlb_get_low_area(unsigned long len, u16 segmask) |
748 | { | |
749 | unsigned long addr = 0; | |
750 | struct vm_area_struct *vma; | |
751 | ||
752 | vma = find_vma(current->mm, addr); | |
753 | while (addr + len <= 0x100000000UL) { | |
754 | BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */ | |
755 | ||
756 | if (! __within_hugepage_low_range(addr, len, segmask)) { | |
757 | addr = ALIGN(addr+1, 1<<SID_SHIFT); | |
758 | vma = find_vma(current->mm, addr); | |
759 | continue; | |
760 | } | |
761 | ||
762 | if (!vma || (addr + len) <= vma->vm_start) | |
763 | return addr; | |
764 | addr = ALIGN(vma->vm_end, HPAGE_SIZE); | |
765 | /* Depending on segmask this might not be a confirmed | |
766 | * hugepage region, so the ALIGN could have skipped | |
767 | * some VMAs */ | |
768 | vma = find_vma(current->mm, addr); | |
769 | } | |
770 | ||
771 | return -ENOMEM; | |
772 | } | |
773 | ||
c594adad | 774 | static unsigned long htlb_get_high_area(unsigned long len, u16 areamask) |
1da177e4 | 775 | { |
c594adad | 776 | unsigned long addr = 0x100000000UL; |
1da177e4 LT |
777 | struct vm_area_struct *vma; |
778 | ||
779 | vma = find_vma(current->mm, addr); | |
c594adad | 780 | while (addr + len <= TASK_SIZE_USER64) { |
1da177e4 | 781 | BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */ |
c594adad DG |
782 | |
783 | if (! __within_hugepage_high_range(addr, len, areamask)) { | |
784 | addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT); | |
785 | vma = find_vma(current->mm, addr); | |
786 | continue; | |
787 | } | |
1da177e4 LT |
788 | |
789 | if (!vma || (addr + len) <= vma->vm_start) | |
790 | return addr; | |
791 | addr = ALIGN(vma->vm_end, HPAGE_SIZE); | |
c594adad DG |
792 | /* Depending on segmask this might not be a confirmed |
793 | * hugepage region, so the ALIGN could have skipped | |
794 | * some VMAs */ | |
795 | vma = find_vma(current->mm, addr); | |
1da177e4 LT |
796 | } |
797 | ||
798 | return -ENOMEM; | |
799 | } | |
800 | ||
801 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, | |
802 | unsigned long len, unsigned long pgoff, | |
803 | unsigned long flags) | |
804 | { | |
c594adad DG |
805 | int lastshift; |
806 | u16 areamask, curareas; | |
807 | ||
3c726f8d BH |
808 | if (HPAGE_SHIFT == 0) |
809 | return -EINVAL; | |
1da177e4 LT |
810 | if (len & ~HPAGE_MASK) |
811 | return -EINVAL; | |
812 | ||
813 | if (!cpu_has_feature(CPU_FTR_16M_PAGE)) | |
814 | return -EINVAL; | |
815 | ||
456752f7 DG |
816 | /* Paranoia, caller should have dealt with this */ |
817 | BUG_ON((addr + len) < addr); | |
818 | ||
1da177e4 | 819 | if (test_thread_flag(TIF_32BIT)) { |
456752f7 DG |
820 | /* Paranoia, caller should have dealt with this */ |
821 | BUG_ON((addr + len) > 0x100000000UL); | |
822 | ||
c594adad | 823 | curareas = current->mm->context.low_htlb_areas; |
1da177e4 | 824 | |
456752f7 DG |
825 | /* First see if we can use the hint address */ |
826 | if (addr && (htlb_check_hinted_area(addr, len) == 0)) { | |
827 | areamask = LOW_ESID_MASK(addr, len); | |
828 | if (open_low_hpage_areas(current->mm, areamask) == 0) | |
829 | return addr; | |
830 | } | |
831 | ||
832 | /* Next see if we can map in the existing low areas */ | |
c594adad | 833 | addr = htlb_get_low_area(len, curareas); |
1da177e4 LT |
834 | if (addr != -ENOMEM) |
835 | return addr; | |
836 | ||
456752f7 | 837 | /* Finally go looking for areas to open */ |
c594adad DG |
838 | lastshift = 0; |
839 | for (areamask = LOW_ESID_MASK(0x100000000UL-len, len); | |
840 | ! lastshift; areamask >>=1) { | |
841 | if (areamask & 1) | |
1da177e4 LT |
842 | lastshift = 1; |
843 | ||
c594adad | 844 | addr = htlb_get_low_area(len, curareas | areamask); |
1da177e4 | 845 | if ((addr != -ENOMEM) |
c594adad | 846 | && open_low_hpage_areas(current->mm, areamask) == 0) |
1da177e4 LT |
847 | return addr; |
848 | } | |
1da177e4 | 849 | } else { |
c594adad DG |
850 | curareas = current->mm->context.high_htlb_areas; |
851 | ||
456752f7 DG |
852 | /* First see if we can use the hint address */ |
853 | /* We discourage 64-bit processes from doing hugepage | |
854 | * mappings below 4GB (must use MAP_FIXED) */ | |
855 | if ((addr >= 0x100000000UL) | |
856 | && (htlb_check_hinted_area(addr, len) == 0)) { | |
857 | areamask = HTLB_AREA_MASK(addr, len); | |
858 | if (open_high_hpage_areas(current->mm, areamask) == 0) | |
859 | return addr; | |
860 | } | |
861 | ||
862 | /* Next see if we can map in the existing high areas */ | |
c594adad DG |
863 | addr = htlb_get_high_area(len, curareas); |
864 | if (addr != -ENOMEM) | |
865 | return addr; | |
866 | ||
456752f7 | 867 | /* Finally go looking for areas to open */ |
c594adad DG |
868 | lastshift = 0; |
869 | for (areamask = HTLB_AREA_MASK(TASK_SIZE_USER64-len, len); | |
870 | ! lastshift; areamask >>=1) { | |
871 | if (areamask & 1) | |
872 | lastshift = 1; | |
873 | ||
874 | addr = htlb_get_high_area(len, curareas | areamask); | |
875 | if ((addr != -ENOMEM) | |
876 | && open_high_hpage_areas(current->mm, areamask) == 0) | |
877 | return addr; | |
878 | } | |
1da177e4 | 879 | } |
c594adad DG |
880 | printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open" |
881 | " enough areas\n"); | |
882 | return -ENOMEM; | |
1da177e4 LT |
883 | } |
884 | ||
cbf52afd DG |
885 | /* |
886 | * Called by asm hashtable.S for doing lazy icache flush | |
887 | */ | |
888 | static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags, | |
889 | pte_t pte, int trap) | |
890 | { | |
891 | struct page *page; | |
892 | int i; | |
893 | ||
894 | if (!pfn_valid(pte_pfn(pte))) | |
895 | return rflags; | |
896 | ||
897 | page = pte_page(pte); | |
898 | ||
899 | /* page is dirty */ | |
900 | if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) { | |
901 | if (trap == 0x400) { | |
902 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) | |
903 | __flush_dcache_icache(page_address(page+i)); | |
904 | set_bit(PG_arch_1, &page->flags); | |
905 | } else { | |
906 | rflags |= HPTE_R_N; | |
907 | } | |
908 | } | |
909 | return rflags; | |
910 | } | |
911 | ||
1da177e4 | 912 | int hash_huge_page(struct mm_struct *mm, unsigned long access, |
cbf52afd DG |
913 | unsigned long ea, unsigned long vsid, int local, |
914 | unsigned long trap) | |
1da177e4 LT |
915 | { |
916 | pte_t *ptep; | |
3c726f8d BH |
917 | unsigned long old_pte, new_pte; |
918 | unsigned long va, rflags, pa; | |
1da177e4 LT |
919 | long slot; |
920 | int err = 1; | |
921 | ||
1da177e4 LT |
922 | ptep = huge_pte_offset(mm, ea); |
923 | ||
924 | /* Search the Linux page table for a match with va */ | |
925 | va = (vsid << 28) | (ea & 0x0fffffff); | |
1da177e4 LT |
926 | |
927 | /* | |
928 | * If no pte found or not present, send the problem up to | |
929 | * do_page_fault | |
930 | */ | |
931 | if (unlikely(!ptep || pte_none(*ptep))) | |
932 | goto out; | |
933 | ||
1da177e4 LT |
934 | /* |
935 | * Check the user's access rights to the page. If access should be | |
936 | * prevented then send the problem up to do_page_fault. | |
937 | */ | |
938 | if (unlikely(access & ~pte_val(*ptep))) | |
939 | goto out; | |
940 | /* | |
941 | * At this point, we have a pte (old_pte) which can be used to build | |
942 | * or update an HPTE. There are 2 cases: | |
943 | * | |
944 | * 1. There is a valid (present) pte with no associated HPTE (this is | |
945 | * the most common case) | |
946 | * 2. There is a valid (present) pte with an associated HPTE. The | |
947 | * current values of the pp bits in the HPTE prevent access | |
948 | * because we are doing software DIRTY bit management and the | |
949 | * page is currently not DIRTY. | |
950 | */ | |
951 | ||
952 | ||
3c726f8d BH |
953 | do { |
954 | old_pte = pte_val(*ptep); | |
955 | if (old_pte & _PAGE_BUSY) | |
956 | goto out; | |
957 | new_pte = old_pte | _PAGE_BUSY | | |
958 | _PAGE_ACCESSED | _PAGE_HASHPTE; | |
959 | } while(old_pte != __cmpxchg_u64((unsigned long *)ptep, | |
960 | old_pte, new_pte)); | |
961 | ||
962 | rflags = 0x2 | (!(new_pte & _PAGE_RW)); | |
1da177e4 | 963 | /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */ |
3c726f8d | 964 | rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N); |
cbf52afd DG |
965 | if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) |
966 | /* No CPU has hugepages but lacks no execute, so we | |
967 | * don't need to worry about that case */ | |
968 | rflags = hash_huge_page_do_lazy_icache(rflags, __pte(old_pte), | |
969 | trap); | |
1da177e4 LT |
970 | |
971 | /* Check if pte already has an hpte (case 2) */ | |
3c726f8d | 972 | if (unlikely(old_pte & _PAGE_HASHPTE)) { |
1da177e4 LT |
973 | /* There MIGHT be an HPTE for this pte */ |
974 | unsigned long hash, slot; | |
975 | ||
3c726f8d BH |
976 | hash = hpt_hash(va, HPAGE_SHIFT); |
977 | if (old_pte & _PAGE_F_SECOND) | |
1da177e4 LT |
978 | hash = ~hash; |
979 | slot = (hash & htab_hash_mask) * HPTES_PER_GROUP; | |
3c726f8d | 980 | slot += (old_pte & _PAGE_F_GIX) >> 12; |
1da177e4 | 981 | |
325c82a0 BH |
982 | if (ppc_md.hpte_updatepp(slot, rflags, va, mmu_huge_psize, |
983 | local) == -1) | |
3c726f8d | 984 | old_pte &= ~_PAGE_HPTEFLAGS; |
1da177e4 LT |
985 | } |
986 | ||
3c726f8d BH |
987 | if (likely(!(old_pte & _PAGE_HASHPTE))) { |
988 | unsigned long hash = hpt_hash(va, HPAGE_SHIFT); | |
1da177e4 LT |
989 | unsigned long hpte_group; |
990 | ||
3c726f8d | 991 | pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT; |
1da177e4 LT |
992 | |
993 | repeat: | |
994 | hpte_group = ((hash & htab_hash_mask) * | |
995 | HPTES_PER_GROUP) & ~0x7UL; | |
996 | ||
3c726f8d BH |
997 | /* clear HPTE slot informations in new PTE */ |
998 | new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE; | |
1da177e4 LT |
999 | |
1000 | /* Add in WIMG bits */ | |
1001 | /* XXX We should store these in the pte */ | |
3c726f8d | 1002 | /* --BenH: I think they are ... */ |
96e28449 | 1003 | rflags |= _PAGE_COHERENT; |
1da177e4 | 1004 | |
3c726f8d BH |
1005 | /* Insert into the hash table, primary slot */ |
1006 | slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, 0, | |
1007 | mmu_huge_psize); | |
1da177e4 LT |
1008 | |
1009 | /* Primary is full, try the secondary */ | |
1010 | if (unlikely(slot == -1)) { | |
3c726f8d | 1011 | new_pte |= _PAGE_F_SECOND; |
1da177e4 LT |
1012 | hpte_group = ((~hash & htab_hash_mask) * |
1013 | HPTES_PER_GROUP) & ~0x7UL; | |
3c726f8d | 1014 | slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, |
67b10813 | 1015 | HPTE_V_SECONDARY, |
3c726f8d | 1016 | mmu_huge_psize); |
1da177e4 LT |
1017 | if (slot == -1) { |
1018 | if (mftb() & 0x1) | |
67b10813 BH |
1019 | hpte_group = ((hash & htab_hash_mask) * |
1020 | HPTES_PER_GROUP)&~0x7UL; | |
1da177e4 LT |
1021 | |
1022 | ppc_md.hpte_remove(hpte_group); | |
1023 | goto repeat; | |
1024 | } | |
1025 | } | |
1026 | ||
1027 | if (unlikely(slot == -2)) | |
1028 | panic("hash_huge_page: pte_insert failed\n"); | |
1029 | ||
3c726f8d | 1030 | new_pte |= (slot << 12) & _PAGE_F_GIX; |
1da177e4 LT |
1031 | } |
1032 | ||
3c726f8d | 1033 | /* |
01edcd89 | 1034 | * No need to use ldarx/stdcx here |
3c726f8d BH |
1035 | */ |
1036 | *ptep = __pte(new_pte & ~_PAGE_BUSY); | |
1037 | ||
1da177e4 LT |
1038 | err = 0; |
1039 | ||
1040 | out: | |
1da177e4 LT |
1041 | return err; |
1042 | } | |
f10a04c0 DG |
1043 | |
1044 | static void zero_ctor(void *addr, kmem_cache_t *cache, unsigned long flags) | |
1045 | { | |
1046 | memset(addr, 0, kmem_cache_size(cache)); | |
1047 | } | |
1048 | ||
1049 | static int __init hugetlbpage_init(void) | |
1050 | { | |
1051 | if (!cpu_has_feature(CPU_FTR_16M_PAGE)) | |
1052 | return -ENODEV; | |
1053 | ||
1054 | huge_pgtable_cache = kmem_cache_create("hugepte_cache", | |
1055 | HUGEPTE_TABLE_SIZE, | |
1056 | HUGEPTE_TABLE_SIZE, | |
1057 | SLAB_HWCACHE_ALIGN | | |
1058 | SLAB_MUST_HWCACHE_ALIGN, | |
1059 | zero_ctor, NULL); | |
1060 | if (! huge_pgtable_cache) | |
1061 | panic("hugetlbpage_init(): could not create hugepte cache\n"); | |
1062 | ||
1063 | return 0; | |
1064 | } | |
1065 | ||
1066 | module_init(hugetlbpage_init); |