]>
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> | |
1da177e4 LT |
15 | #include <linux/slab.h> |
16 | #include <linux/err.h> | |
17 | #include <linux/sysctl.h> | |
18 | #include <asm/mman.h> | |
19 | #include <asm/pgalloc.h> | |
20 | #include <asm/tlb.h> | |
21 | #include <asm/tlbflush.h> | |
22 | #include <asm/mmu_context.h> | |
23 | #include <asm/machdep.h> | |
24 | #include <asm/cputable.h> | |
94b2a439 | 25 | #include <asm/spu.h> |
1da177e4 | 26 | |
4ec161cf JT |
27 | #define HPAGE_SHIFT_64K 16 |
28 | #define HPAGE_SHIFT_16M 24 | |
29 | ||
c594adad DG |
30 | #define NUM_LOW_AREAS (0x100000000UL >> SID_SHIFT) |
31 | #define NUM_HIGH_AREAS (PGTABLE_RANGE >> HTLB_AREA_SHIFT) | |
32 | ||
4ec161cf JT |
33 | unsigned int hugepte_shift; |
34 | #define PTRS_PER_HUGEPTE (1 << hugepte_shift) | |
35 | #define HUGEPTE_TABLE_SIZE (sizeof(pte_t) << hugepte_shift) | |
f10a04c0 | 36 | |
4ec161cf | 37 | #define HUGEPD_SHIFT (HPAGE_SHIFT + hugepte_shift) |
f10a04c0 DG |
38 | #define HUGEPD_SIZE (1UL << HUGEPD_SHIFT) |
39 | #define HUGEPD_MASK (~(HUGEPD_SIZE-1)) | |
40 | ||
41 | #define huge_pgtable_cache (pgtable_cache[HUGEPTE_CACHE_NUM]) | |
42 | ||
43 | /* Flag to mark huge PD pointers. This means pmd_bad() and pud_bad() | |
44 | * will choke on pointers to hugepte tables, which is handy for | |
45 | * catching screwups early. */ | |
46 | #define HUGEPD_OK 0x1 | |
47 | ||
48 | typedef struct { unsigned long pd; } hugepd_t; | |
49 | ||
50 | #define hugepd_none(hpd) ((hpd).pd == 0) | |
51 | ||
52 | static inline pte_t *hugepd_page(hugepd_t hpd) | |
53 | { | |
54 | BUG_ON(!(hpd.pd & HUGEPD_OK)); | |
55 | return (pte_t *)(hpd.pd & ~HUGEPD_OK); | |
56 | } | |
57 | ||
58 | static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr) | |
59 | { | |
60 | unsigned long idx = ((addr >> HPAGE_SHIFT) & (PTRS_PER_HUGEPTE-1)); | |
61 | pte_t *dir = hugepd_page(*hpdp); | |
62 | ||
63 | return dir + idx; | |
64 | } | |
65 | ||
66 | static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, | |
67 | unsigned long address) | |
68 | { | |
69 | pte_t *new = kmem_cache_alloc(huge_pgtable_cache, | |
70 | GFP_KERNEL|__GFP_REPEAT); | |
71 | ||
72 | if (! new) | |
73 | return -ENOMEM; | |
74 | ||
75 | spin_lock(&mm->page_table_lock); | |
76 | if (!hugepd_none(*hpdp)) | |
77 | kmem_cache_free(huge_pgtable_cache, new); | |
78 | else | |
79 | hpdp->pd = (unsigned long)new | HUGEPD_OK; | |
80 | spin_unlock(&mm->page_table_lock); | |
81 | return 0; | |
82 | } | |
83 | ||
4ec161cf JT |
84 | /* Base page size affects how we walk hugetlb page tables */ |
85 | #ifdef CONFIG_PPC_64K_PAGES | |
86 | #define hpmd_offset(pud, addr) pmd_offset(pud, addr) | |
87 | #define hpmd_alloc(mm, pud, addr) pmd_alloc(mm, pud, addr) | |
88 | #else | |
89 | static inline | |
90 | pmd_t *hpmd_offset(pud_t *pud, unsigned long addr) | |
91 | { | |
92 | if (HPAGE_SHIFT == HPAGE_SHIFT_64K) | |
93 | return pmd_offset(pud, addr); | |
94 | else | |
95 | return (pmd_t *) pud; | |
96 | } | |
97 | static inline | |
98 | pmd_t *hpmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long addr) | |
99 | { | |
100 | if (HPAGE_SHIFT == HPAGE_SHIFT_64K) | |
101 | return pmd_alloc(mm, pud, addr); | |
102 | else | |
103 | return (pmd_t *) pud; | |
104 | } | |
105 | #endif | |
106 | ||
e28f7faf DG |
107 | /* Modelled after find_linux_pte() */ |
108 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
1da177e4 | 109 | { |
e28f7faf DG |
110 | pgd_t *pg; |
111 | pud_t *pu; | |
4ec161cf | 112 | pmd_t *pm; |
1da177e4 | 113 | |
d0f13e3c | 114 | BUG_ON(get_slice_psize(mm, addr) != mmu_huge_psize); |
1da177e4 | 115 | |
e28f7faf DG |
116 | addr &= HPAGE_MASK; |
117 | ||
118 | pg = pgd_offset(mm, addr); | |
119 | if (!pgd_none(*pg)) { | |
120 | pu = pud_offset(pg, addr); | |
121 | if (!pud_none(*pu)) { | |
4ec161cf | 122 | pm = hpmd_offset(pu, addr); |
f10a04c0 DG |
123 | if (!pmd_none(*pm)) |
124 | return hugepte_offset((hugepd_t *)pm, addr); | |
e28f7faf DG |
125 | } |
126 | } | |
1da177e4 | 127 | |
e28f7faf | 128 | return NULL; |
1da177e4 LT |
129 | } |
130 | ||
a5516438 AK |
131 | pte_t *huge_pte_alloc(struct mm_struct *mm, |
132 | unsigned long addr, unsigned long sz) | |
1da177e4 | 133 | { |
e28f7faf DG |
134 | pgd_t *pg; |
135 | pud_t *pu; | |
4ec161cf | 136 | pmd_t *pm; |
f10a04c0 | 137 | hugepd_t *hpdp = NULL; |
1da177e4 | 138 | |
d0f13e3c | 139 | BUG_ON(get_slice_psize(mm, addr) != mmu_huge_psize); |
1da177e4 | 140 | |
e28f7faf | 141 | addr &= HPAGE_MASK; |
1da177e4 | 142 | |
e28f7faf DG |
143 | pg = pgd_offset(mm, addr); |
144 | pu = pud_alloc(mm, pg, addr); | |
1da177e4 | 145 | |
e28f7faf | 146 | if (pu) { |
4ec161cf | 147 | pm = hpmd_alloc(mm, pu, addr); |
f10a04c0 DG |
148 | if (pm) |
149 | hpdp = (hugepd_t *)pm; | |
f10a04c0 DG |
150 | } |
151 | ||
152 | if (! hpdp) | |
153 | return NULL; | |
154 | ||
155 | if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr)) | |
156 | return NULL; | |
157 | ||
158 | return hugepte_offset(hpdp, addr); | |
159 | } | |
160 | ||
39dde65c KC |
161 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
162 | { | |
163 | return 0; | |
164 | } | |
165 | ||
f10a04c0 DG |
166 | static void free_hugepte_range(struct mmu_gather *tlb, hugepd_t *hpdp) |
167 | { | |
168 | pte_t *hugepte = hugepd_page(*hpdp); | |
169 | ||
170 | hpdp->pd = 0; | |
171 | tlb->need_flush = 1; | |
172 | pgtable_free_tlb(tlb, pgtable_free_cache(hugepte, HUGEPTE_CACHE_NUM, | |
c9169f87 | 173 | PGF_CACHENUM_MASK)); |
f10a04c0 DG |
174 | } |
175 | ||
f10a04c0 DG |
176 | static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
177 | unsigned long addr, unsigned long end, | |
178 | unsigned long floor, unsigned long ceiling) | |
179 | { | |
180 | pmd_t *pmd; | |
181 | unsigned long next; | |
182 | unsigned long start; | |
183 | ||
184 | start = addr; | |
185 | pmd = pmd_offset(pud, addr); | |
186 | do { | |
187 | next = pmd_addr_end(addr, end); | |
188 | if (pmd_none(*pmd)) | |
189 | continue; | |
190 | free_hugepte_range(tlb, (hugepd_t *)pmd); | |
191 | } while (pmd++, addr = next, addr != end); | |
192 | ||
193 | start &= PUD_MASK; | |
194 | if (start < floor) | |
195 | return; | |
196 | if (ceiling) { | |
197 | ceiling &= PUD_MASK; | |
198 | if (!ceiling) | |
199 | return; | |
1da177e4 | 200 | } |
f10a04c0 DG |
201 | if (end - 1 > ceiling - 1) |
202 | return; | |
1da177e4 | 203 | |
f10a04c0 DG |
204 | pmd = pmd_offset(pud, start); |
205 | pud_clear(pud); | |
206 | pmd_free_tlb(tlb, pmd); | |
207 | } | |
f10a04c0 DG |
208 | |
209 | static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
210 | unsigned long addr, unsigned long end, | |
211 | unsigned long floor, unsigned long ceiling) | |
212 | { | |
213 | pud_t *pud; | |
214 | unsigned long next; | |
215 | unsigned long start; | |
216 | ||
217 | start = addr; | |
218 | pud = pud_offset(pgd, addr); | |
219 | do { | |
220 | next = pud_addr_end(addr, end); | |
221 | #ifdef CONFIG_PPC_64K_PAGES | |
222 | if (pud_none_or_clear_bad(pud)) | |
223 | continue; | |
224 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling); | |
225 | #else | |
4ec161cf JT |
226 | if (HPAGE_SHIFT == HPAGE_SHIFT_64K) { |
227 | if (pud_none_or_clear_bad(pud)) | |
228 | continue; | |
229 | hugetlb_free_pmd_range(tlb, pud, addr, next, floor, ceiling); | |
230 | } else { | |
231 | if (pud_none(*pud)) | |
232 | continue; | |
233 | free_hugepte_range(tlb, (hugepd_t *)pud); | |
234 | } | |
f10a04c0 DG |
235 | #endif |
236 | } while (pud++, addr = next, addr != end); | |
237 | ||
238 | start &= PGDIR_MASK; | |
239 | if (start < floor) | |
240 | return; | |
241 | if (ceiling) { | |
242 | ceiling &= PGDIR_MASK; | |
243 | if (!ceiling) | |
244 | return; | |
245 | } | |
246 | if (end - 1 > ceiling - 1) | |
247 | return; | |
248 | ||
249 | pud = pud_offset(pgd, start); | |
250 | pgd_clear(pgd); | |
251 | pud_free_tlb(tlb, pud); | |
252 | } | |
253 | ||
254 | /* | |
255 | * This function frees user-level page tables of a process. | |
256 | * | |
257 | * Must be called with pagetable lock held. | |
258 | */ | |
42b77728 | 259 | void hugetlb_free_pgd_range(struct mmu_gather *tlb, |
f10a04c0 DG |
260 | unsigned long addr, unsigned long end, |
261 | unsigned long floor, unsigned long ceiling) | |
262 | { | |
263 | pgd_t *pgd; | |
264 | unsigned long next; | |
265 | unsigned long start; | |
266 | ||
267 | /* | |
268 | * Comments below take from the normal free_pgd_range(). They | |
269 | * apply here too. The tests against HUGEPD_MASK below are | |
270 | * essential, because we *don't* test for this at the bottom | |
271 | * level. Without them we'll attempt to free a hugepte table | |
272 | * when we unmap just part of it, even if there are other | |
273 | * active mappings using it. | |
274 | * | |
275 | * The next few lines have given us lots of grief... | |
276 | * | |
277 | * Why are we testing HUGEPD* at this top level? Because | |
278 | * often there will be no work to do at all, and we'd prefer | |
279 | * not to go all the way down to the bottom just to discover | |
280 | * that. | |
281 | * | |
282 | * Why all these "- 1"s? Because 0 represents both the bottom | |
283 | * of the address space and the top of it (using -1 for the | |
284 | * top wouldn't help much: the masks would do the wrong thing). | |
285 | * The rule is that addr 0 and floor 0 refer to the bottom of | |
286 | * the address space, but end 0 and ceiling 0 refer to the top | |
287 | * Comparisons need to use "end - 1" and "ceiling - 1" (though | |
288 | * that end 0 case should be mythical). | |
289 | * | |
290 | * Wherever addr is brought up or ceiling brought down, we | |
291 | * must be careful to reject "the opposite 0" before it | |
292 | * confuses the subsequent tests. But what about where end is | |
293 | * brought down by HUGEPD_SIZE below? no, end can't go down to | |
294 | * 0 there. | |
295 | * | |
296 | * Whereas we round start (addr) and ceiling down, by different | |
297 | * masks at different levels, in order to test whether a table | |
298 | * now has no other vmas using it, so can be freed, we don't | |
299 | * bother to round floor or end up - the tests don't need that. | |
300 | */ | |
301 | ||
302 | addr &= HUGEPD_MASK; | |
303 | if (addr < floor) { | |
304 | addr += HUGEPD_SIZE; | |
305 | if (!addr) | |
306 | return; | |
307 | } | |
308 | if (ceiling) { | |
309 | ceiling &= HUGEPD_MASK; | |
310 | if (!ceiling) | |
311 | return; | |
312 | } | |
313 | if (end - 1 > ceiling - 1) | |
314 | end -= HUGEPD_SIZE; | |
315 | if (addr > end - 1) | |
316 | return; | |
317 | ||
318 | start = addr; | |
42b77728 | 319 | pgd = pgd_offset(tlb->mm, addr); |
f10a04c0 | 320 | do { |
42b77728 | 321 | BUG_ON(get_slice_psize(tlb->mm, addr) != mmu_huge_psize); |
f10a04c0 DG |
322 | next = pgd_addr_end(addr, end); |
323 | if (pgd_none_or_clear_bad(pgd)) | |
324 | continue; | |
42b77728 | 325 | hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); |
f10a04c0 | 326 | } while (pgd++, addr = next, addr != end); |
1da177e4 LT |
327 | } |
328 | ||
e28f7faf DG |
329 | void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, |
330 | pte_t *ptep, pte_t pte) | |
331 | { | |
e28f7faf | 332 | if (pte_present(*ptep)) { |
3c726f8d | 333 | /* We open-code pte_clear because we need to pass the right |
a741e679 BH |
334 | * argument to hpte_need_flush (huge / !huge). Might not be |
335 | * necessary anymore if we make hpte_need_flush() get the | |
336 | * page size from the slices | |
3c726f8d | 337 | */ |
a741e679 | 338 | pte_update(mm, addr & HPAGE_MASK, ptep, ~0UL, 1); |
e28f7faf | 339 | } |
3c726f8d | 340 | *ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS); |
1da177e4 LT |
341 | } |
342 | ||
e28f7faf DG |
343 | pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, |
344 | pte_t *ptep) | |
1da177e4 | 345 | { |
a741e679 | 346 | unsigned long old = pte_update(mm, addr, ptep, ~0UL, 1); |
e28f7faf | 347 | return __pte(old); |
1da177e4 LT |
348 | } |
349 | ||
1da177e4 LT |
350 | struct page * |
351 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) | |
352 | { | |
353 | pte_t *ptep; | |
354 | struct page *page; | |
355 | ||
d0f13e3c | 356 | if (get_slice_psize(mm, address) != mmu_huge_psize) |
1da177e4 LT |
357 | return ERR_PTR(-EINVAL); |
358 | ||
359 | ptep = huge_pte_offset(mm, address); | |
360 | page = pte_page(*ptep); | |
361 | if (page) | |
362 | page += (address % HPAGE_SIZE) / PAGE_SIZE; | |
363 | ||
364 | return page; | |
365 | } | |
366 | ||
367 | int pmd_huge(pmd_t pmd) | |
368 | { | |
369 | return 0; | |
370 | } | |
371 | ||
ceb86879 AK |
372 | int pud_huge(pud_t pud) |
373 | { | |
374 | return 0; | |
375 | } | |
376 | ||
1da177e4 LT |
377 | struct page * |
378 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
379 | pmd_t *pmd, int write) | |
380 | { | |
381 | BUG(); | |
382 | return NULL; | |
383 | } | |
384 | ||
1da177e4 LT |
385 | |
386 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, | |
387 | unsigned long len, unsigned long pgoff, | |
388 | unsigned long flags) | |
389 | { | |
d0f13e3c BH |
390 | return slice_get_unmapped_area(addr, len, flags, |
391 | mmu_huge_psize, 1, 0); | |
1da177e4 LT |
392 | } |
393 | ||
cbf52afd DG |
394 | /* |
395 | * Called by asm hashtable.S for doing lazy icache flush | |
396 | */ | |
397 | static unsigned int hash_huge_page_do_lazy_icache(unsigned long rflags, | |
398 | pte_t pte, int trap) | |
399 | { | |
400 | struct page *page; | |
401 | int i; | |
402 | ||
403 | if (!pfn_valid(pte_pfn(pte))) | |
404 | return rflags; | |
405 | ||
406 | page = pte_page(pte); | |
407 | ||
408 | /* page is dirty */ | |
409 | if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) { | |
410 | if (trap == 0x400) { | |
411 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) | |
412 | __flush_dcache_icache(page_address(page+i)); | |
413 | set_bit(PG_arch_1, &page->flags); | |
414 | } else { | |
415 | rflags |= HPTE_R_N; | |
416 | } | |
417 | } | |
418 | return rflags; | |
419 | } | |
420 | ||
1da177e4 | 421 | int hash_huge_page(struct mm_struct *mm, unsigned long access, |
cbf52afd DG |
422 | unsigned long ea, unsigned long vsid, int local, |
423 | unsigned long trap) | |
1da177e4 LT |
424 | { |
425 | pte_t *ptep; | |
3c726f8d BH |
426 | unsigned long old_pte, new_pte; |
427 | unsigned long va, rflags, pa; | |
1da177e4 LT |
428 | long slot; |
429 | int err = 1; | |
1189be65 | 430 | int ssize = user_segment_size(ea); |
1da177e4 | 431 | |
1da177e4 LT |
432 | ptep = huge_pte_offset(mm, ea); |
433 | ||
434 | /* Search the Linux page table for a match with va */ | |
1189be65 | 435 | va = hpt_va(ea, vsid, ssize); |
1da177e4 LT |
436 | |
437 | /* | |
438 | * If no pte found or not present, send the problem up to | |
439 | * do_page_fault | |
440 | */ | |
441 | if (unlikely(!ptep || pte_none(*ptep))) | |
442 | goto out; | |
443 | ||
1da177e4 LT |
444 | /* |
445 | * Check the user's access rights to the page. If access should be | |
446 | * prevented then send the problem up to do_page_fault. | |
447 | */ | |
448 | if (unlikely(access & ~pte_val(*ptep))) | |
449 | goto out; | |
450 | /* | |
451 | * At this point, we have a pte (old_pte) which can be used to build | |
452 | * or update an HPTE. There are 2 cases: | |
453 | * | |
454 | * 1. There is a valid (present) pte with no associated HPTE (this is | |
455 | * the most common case) | |
456 | * 2. There is a valid (present) pte with an associated HPTE. The | |
457 | * current values of the pp bits in the HPTE prevent access | |
458 | * because we are doing software DIRTY bit management and the | |
459 | * page is currently not DIRTY. | |
460 | */ | |
461 | ||
462 | ||
3c726f8d BH |
463 | do { |
464 | old_pte = pte_val(*ptep); | |
465 | if (old_pte & _PAGE_BUSY) | |
466 | goto out; | |
41743a4e | 467 | new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED; |
3c726f8d BH |
468 | } while(old_pte != __cmpxchg_u64((unsigned long *)ptep, |
469 | old_pte, new_pte)); | |
470 | ||
471 | rflags = 0x2 | (!(new_pte & _PAGE_RW)); | |
1da177e4 | 472 | /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */ |
3c726f8d | 473 | rflags |= ((new_pte & _PAGE_EXEC) ? 0 : HPTE_R_N); |
cbf52afd DG |
474 | if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) |
475 | /* No CPU has hugepages but lacks no execute, so we | |
476 | * don't need to worry about that case */ | |
477 | rflags = hash_huge_page_do_lazy_icache(rflags, __pte(old_pte), | |
478 | trap); | |
1da177e4 LT |
479 | |
480 | /* Check if pte already has an hpte (case 2) */ | |
3c726f8d | 481 | if (unlikely(old_pte & _PAGE_HASHPTE)) { |
1da177e4 LT |
482 | /* There MIGHT be an HPTE for this pte */ |
483 | unsigned long hash, slot; | |
484 | ||
1189be65 | 485 | hash = hpt_hash(va, HPAGE_SHIFT, ssize); |
3c726f8d | 486 | if (old_pte & _PAGE_F_SECOND) |
1da177e4 LT |
487 | hash = ~hash; |
488 | slot = (hash & htab_hash_mask) * HPTES_PER_GROUP; | |
3c726f8d | 489 | slot += (old_pte & _PAGE_F_GIX) >> 12; |
1da177e4 | 490 | |
325c82a0 | 491 | if (ppc_md.hpte_updatepp(slot, rflags, va, mmu_huge_psize, |
1189be65 | 492 | ssize, local) == -1) |
3c726f8d | 493 | old_pte &= ~_PAGE_HPTEFLAGS; |
1da177e4 LT |
494 | } |
495 | ||
3c726f8d | 496 | if (likely(!(old_pte & _PAGE_HASHPTE))) { |
1189be65 | 497 | unsigned long hash = hpt_hash(va, HPAGE_SHIFT, ssize); |
1da177e4 LT |
498 | unsigned long hpte_group; |
499 | ||
3c726f8d | 500 | pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT; |
1da177e4 LT |
501 | |
502 | repeat: | |
503 | hpte_group = ((hash & htab_hash_mask) * | |
504 | HPTES_PER_GROUP) & ~0x7UL; | |
505 | ||
3c726f8d | 506 | /* clear HPTE slot informations in new PTE */ |
41743a4e BH |
507 | #ifdef CONFIG_PPC_64K_PAGES |
508 | new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HPTE_SUB0; | |
509 | #else | |
3c726f8d | 510 | new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE; |
41743a4e | 511 | #endif |
1da177e4 | 512 | /* Add in WIMG bits */ |
87e9ab13 DK |
513 | rflags |= (new_pte & (_PAGE_WRITETHRU | _PAGE_NO_CACHE | |
514 | _PAGE_COHERENT | _PAGE_GUARDED)); | |
1da177e4 | 515 | |
3c726f8d BH |
516 | /* Insert into the hash table, primary slot */ |
517 | slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, 0, | |
1189be65 | 518 | mmu_huge_psize, ssize); |
1da177e4 LT |
519 | |
520 | /* Primary is full, try the secondary */ | |
521 | if (unlikely(slot == -1)) { | |
1da177e4 LT |
522 | hpte_group = ((~hash & htab_hash_mask) * |
523 | HPTES_PER_GROUP) & ~0x7UL; | |
3c726f8d | 524 | slot = ppc_md.hpte_insert(hpte_group, va, pa, rflags, |
67b10813 | 525 | HPTE_V_SECONDARY, |
1189be65 | 526 | mmu_huge_psize, ssize); |
1da177e4 LT |
527 | if (slot == -1) { |
528 | if (mftb() & 0x1) | |
67b10813 BH |
529 | hpte_group = ((hash & htab_hash_mask) * |
530 | HPTES_PER_GROUP)&~0x7UL; | |
1da177e4 LT |
531 | |
532 | ppc_md.hpte_remove(hpte_group); | |
533 | goto repeat; | |
534 | } | |
535 | } | |
536 | ||
537 | if (unlikely(slot == -2)) | |
538 | panic("hash_huge_page: pte_insert failed\n"); | |
539 | ||
d649bd7b | 540 | new_pte |= (slot << 12) & (_PAGE_F_SECOND | _PAGE_F_GIX); |
1da177e4 LT |
541 | } |
542 | ||
3c726f8d | 543 | /* |
01edcd89 | 544 | * No need to use ldarx/stdcx here |
3c726f8d BH |
545 | */ |
546 | *ptep = __pte(new_pte & ~_PAGE_BUSY); | |
547 | ||
1da177e4 LT |
548 | err = 0; |
549 | ||
550 | out: | |
1da177e4 LT |
551 | return err; |
552 | } | |
f10a04c0 | 553 | |
4ec161cf JT |
554 | void set_huge_psize(int psize) |
555 | { | |
556 | /* Check that it is a page size supported by the hardware and | |
557 | * that it fits within pagetable limits. */ | |
558 | if (mmu_psize_defs[psize].shift && mmu_psize_defs[psize].shift < SID_SHIFT && | |
559 | (mmu_psize_defs[psize].shift > MIN_HUGEPTE_SHIFT || | |
560 | mmu_psize_defs[psize].shift == HPAGE_SHIFT_64K)) { | |
561 | HPAGE_SHIFT = mmu_psize_defs[psize].shift; | |
562 | mmu_huge_psize = psize; | |
563 | #ifdef CONFIG_PPC_64K_PAGES | |
564 | hugepte_shift = (PMD_SHIFT-HPAGE_SHIFT); | |
565 | #else | |
566 | if (HPAGE_SHIFT == HPAGE_SHIFT_64K) | |
567 | hugepte_shift = (PMD_SHIFT-HPAGE_SHIFT); | |
568 | else | |
569 | hugepte_shift = (PUD_SHIFT-HPAGE_SHIFT); | |
570 | #endif | |
571 | ||
572 | } else | |
573 | HPAGE_SHIFT = 0; | |
574 | } | |
575 | ||
576 | static int __init hugepage_setup_sz(char *str) | |
577 | { | |
578 | unsigned long long size; | |
579 | int mmu_psize = -1; | |
580 | int shift; | |
581 | ||
582 | size = memparse(str, &str); | |
583 | ||
584 | shift = __ffs(size); | |
585 | switch (shift) { | |
586 | #ifndef CONFIG_PPC_64K_PAGES | |
587 | case HPAGE_SHIFT_64K: | |
588 | mmu_psize = MMU_PAGE_64K; | |
589 | break; | |
590 | #endif | |
591 | case HPAGE_SHIFT_16M: | |
592 | mmu_psize = MMU_PAGE_16M; | |
593 | break; | |
594 | } | |
595 | ||
596 | if (mmu_psize >=0 && mmu_psize_defs[mmu_psize].shift) | |
597 | set_huge_psize(mmu_psize); | |
598 | else | |
599 | printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size); | |
600 | ||
601 | return 1; | |
602 | } | |
603 | __setup("hugepagesz=", hugepage_setup_sz); | |
604 | ||
4ba9b9d0 | 605 | static void zero_ctor(struct kmem_cache *cache, void *addr) |
f10a04c0 DG |
606 | { |
607 | memset(addr, 0, kmem_cache_size(cache)); | |
608 | } | |
609 | ||
610 | static int __init hugetlbpage_init(void) | |
611 | { | |
612 | if (!cpu_has_feature(CPU_FTR_16M_PAGE)) | |
613 | return -ENODEV; | |
614 | ||
615 | huge_pgtable_cache = kmem_cache_create("hugepte_cache", | |
616 | HUGEPTE_TABLE_SIZE, | |
617 | HUGEPTE_TABLE_SIZE, | |
f0f3980b | 618 | 0, |
20c2df83 | 619 | zero_ctor); |
f10a04c0 DG |
620 | if (! huge_pgtable_cache) |
621 | panic("hugetlbpage_init(): could not create hugepte cache\n"); | |
622 | ||
623 | return 0; | |
624 | } | |
625 | ||
626 | module_init(hugetlbpage_init); |