Loading...
1/*
2 * PPC Huge TLB Page Support for Kernel.
3 *
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
6 *
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/io.h>
13#include <linux/slab.h>
14#include <linux/hugetlb.h>
15#include <linux/export.h>
16#include <linux/of_fdt.h>
17#include <linux/memblock.h>
18#include <linux/moduleparam.h>
19#include <linux/swap.h>
20#include <linux/swapops.h>
21#include <linux/kmemleak.h>
22#include <asm/pgtable.h>
23#include <asm/pgalloc.h>
24#include <asm/tlb.h>
25#include <asm/setup.h>
26#include <asm/hugetlb.h>
27#include <asm/pte-walk.h>
28
29bool hugetlb_disabled = false;
30
31#define hugepd_none(hpd) (hpd_val(hpd) == 0)
32
33#define PTE_T_ORDER (__builtin_ffs(sizeof(pte_t)) - __builtin_ffs(sizeof(void *)))
34
35pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
36{
37 /*
38 * Only called for hugetlbfs pages, hence can ignore THP and the
39 * irq disabled walk.
40 */
41 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
42}
43
44static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
45 unsigned long address, unsigned int pdshift,
46 unsigned int pshift, spinlock_t *ptl)
47{
48 struct kmem_cache *cachep;
49 pte_t *new;
50 int i;
51 int num_hugepd;
52
53 if (pshift >= pdshift) {
54 cachep = PGT_CACHE(PTE_T_ORDER);
55 num_hugepd = 1 << (pshift - pdshift);
56 } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
57 cachep = PGT_CACHE(PTE_INDEX_SIZE);
58 num_hugepd = 1;
59 } else {
60 cachep = PGT_CACHE(pdshift - pshift);
61 num_hugepd = 1;
62 }
63
64 if (!cachep) {
65 WARN_ONCE(1, "No page table cache created for hugetlb tables");
66 return -ENOMEM;
67 }
68
69 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
70
71 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
72 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
73
74 if (!new)
75 return -ENOMEM;
76
77 /*
78 * Make sure other cpus find the hugepd set only after a
79 * properly initialized page table is visible to them.
80 * For more details look for comment in __pte_alloc().
81 */
82 smp_wmb();
83
84 spin_lock(ptl);
85 /*
86 * We have multiple higher-level entries that point to the same
87 * actual pte location. Fill in each as we go and backtrack on error.
88 * We need all of these so the DTLB pgtable walk code can find the
89 * right higher-level entry without knowing if it's a hugepage or not.
90 */
91 for (i = 0; i < num_hugepd; i++, hpdp++) {
92 if (unlikely(!hugepd_none(*hpdp)))
93 break;
94 hugepd_populate(hpdp, new, pshift);
95 }
96 /* If we bailed from the for loop early, an error occurred, clean up */
97 if (i < num_hugepd) {
98 for (i = i - 1 ; i >= 0; i--, hpdp--)
99 *hpdp = __hugepd(0);
100 kmem_cache_free(cachep, new);
101 } else {
102 kmemleak_ignore(new);
103 }
104 spin_unlock(ptl);
105 return 0;
106}
107
108/*
109 * At this point we do the placement change only for BOOK3S 64. This would
110 * possibly work on other subarchs.
111 */
112pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
113{
114 pgd_t *pg;
115 pud_t *pu;
116 pmd_t *pm;
117 hugepd_t *hpdp = NULL;
118 unsigned pshift = __ffs(sz);
119 unsigned pdshift = PGDIR_SHIFT;
120 spinlock_t *ptl;
121
122 addr &= ~(sz-1);
123 pg = pgd_offset(mm, addr);
124
125#ifdef CONFIG_PPC_BOOK3S_64
126 if (pshift == PGDIR_SHIFT)
127 /* 16GB huge page */
128 return (pte_t *) pg;
129 else if (pshift > PUD_SHIFT) {
130 /*
131 * We need to use hugepd table
132 */
133 ptl = &mm->page_table_lock;
134 hpdp = (hugepd_t *)pg;
135 } else {
136 pdshift = PUD_SHIFT;
137 pu = pud_alloc(mm, pg, addr);
138 if (!pu)
139 return NULL;
140 if (pshift == PUD_SHIFT)
141 return (pte_t *)pu;
142 else if (pshift > PMD_SHIFT) {
143 ptl = pud_lockptr(mm, pu);
144 hpdp = (hugepd_t *)pu;
145 } else {
146 pdshift = PMD_SHIFT;
147 pm = pmd_alloc(mm, pu, addr);
148 if (!pm)
149 return NULL;
150 if (pshift == PMD_SHIFT)
151 /* 16MB hugepage */
152 return (pte_t *)pm;
153 else {
154 ptl = pmd_lockptr(mm, pm);
155 hpdp = (hugepd_t *)pm;
156 }
157 }
158 }
159#else
160 if (pshift >= PGDIR_SHIFT) {
161 ptl = &mm->page_table_lock;
162 hpdp = (hugepd_t *)pg;
163 } else {
164 pdshift = PUD_SHIFT;
165 pu = pud_alloc(mm, pg, addr);
166 if (!pu)
167 return NULL;
168 if (pshift >= PUD_SHIFT) {
169 ptl = pud_lockptr(mm, pu);
170 hpdp = (hugepd_t *)pu;
171 } else {
172 pdshift = PMD_SHIFT;
173 pm = pmd_alloc(mm, pu, addr);
174 if (!pm)
175 return NULL;
176 ptl = pmd_lockptr(mm, pm);
177 hpdp = (hugepd_t *)pm;
178 }
179 }
180#endif
181 if (!hpdp)
182 return NULL;
183
184 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
185
186 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
187 pdshift, pshift, ptl))
188 return NULL;
189
190 return hugepte_offset(*hpdp, addr, pdshift);
191}
192
193#ifdef CONFIG_PPC_BOOK3S_64
194/*
195 * Tracks gpages after the device tree is scanned and before the
196 * huge_boot_pages list is ready on pseries.
197 */
198#define MAX_NUMBER_GPAGES 1024
199__initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
200__initdata static unsigned nr_gpages;
201
202/*
203 * Build list of addresses of gigantic pages. This function is used in early
204 * boot before the buddy allocator is setup.
205 */
206void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
207{
208 if (!addr)
209 return;
210 while (number_of_pages > 0) {
211 gpage_freearray[nr_gpages] = addr;
212 nr_gpages++;
213 number_of_pages--;
214 addr += page_size;
215 }
216}
217
218int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
219{
220 struct huge_bootmem_page *m;
221 if (nr_gpages == 0)
222 return 0;
223 m = phys_to_virt(gpage_freearray[--nr_gpages]);
224 gpage_freearray[nr_gpages] = 0;
225 list_add(&m->list, &huge_boot_pages);
226 m->hstate = hstate;
227 return 1;
228}
229#endif
230
231
232int __init alloc_bootmem_huge_page(struct hstate *h)
233{
234
235#ifdef CONFIG_PPC_BOOK3S_64
236 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
237 return pseries_alloc_bootmem_huge_page(h);
238#endif
239 return __alloc_bootmem_huge_page(h);
240}
241
242#ifndef CONFIG_PPC_BOOK3S_64
243#define HUGEPD_FREELIST_SIZE \
244 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
245
246struct hugepd_freelist {
247 struct rcu_head rcu;
248 unsigned int index;
249 void *ptes[0];
250};
251
252static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
253
254static void hugepd_free_rcu_callback(struct rcu_head *head)
255{
256 struct hugepd_freelist *batch =
257 container_of(head, struct hugepd_freelist, rcu);
258 unsigned int i;
259
260 for (i = 0; i < batch->index; i++)
261 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
262
263 free_page((unsigned long)batch);
264}
265
266static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
267{
268 struct hugepd_freelist **batchp;
269
270 batchp = &get_cpu_var(hugepd_freelist_cur);
271
272 if (atomic_read(&tlb->mm->mm_users) < 2 ||
273 mm_is_thread_local(tlb->mm)) {
274 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
275 put_cpu_var(hugepd_freelist_cur);
276 return;
277 }
278
279 if (*batchp == NULL) {
280 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
281 (*batchp)->index = 0;
282 }
283
284 (*batchp)->ptes[(*batchp)->index++] = hugepte;
285 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
286 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
287 *batchp = NULL;
288 }
289 put_cpu_var(hugepd_freelist_cur);
290}
291#else
292static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
293#endif
294
295static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
296 unsigned long start, unsigned long end,
297 unsigned long floor, unsigned long ceiling)
298{
299 pte_t *hugepte = hugepd_page(*hpdp);
300 int i;
301
302 unsigned long pdmask = ~((1UL << pdshift) - 1);
303 unsigned int num_hugepd = 1;
304 unsigned int shift = hugepd_shift(*hpdp);
305
306 /* Note: On fsl the hpdp may be the first of several */
307 if (shift > pdshift)
308 num_hugepd = 1 << (shift - pdshift);
309
310 start &= pdmask;
311 if (start < floor)
312 return;
313 if (ceiling) {
314 ceiling &= pdmask;
315 if (! ceiling)
316 return;
317 }
318 if (end - 1 > ceiling - 1)
319 return;
320
321 for (i = 0; i < num_hugepd; i++, hpdp++)
322 *hpdp = __hugepd(0);
323
324 if (shift >= pdshift)
325 hugepd_free(tlb, hugepte);
326 else if (IS_ENABLED(CONFIG_PPC_8xx))
327 pgtable_free_tlb(tlb, hugepte,
328 get_hugepd_cache_index(PTE_INDEX_SIZE));
329 else
330 pgtable_free_tlb(tlb, hugepte,
331 get_hugepd_cache_index(pdshift - shift));
332}
333
334static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
335 unsigned long addr, unsigned long end,
336 unsigned long floor, unsigned long ceiling)
337{
338 pmd_t *pmd;
339 unsigned long next;
340 unsigned long start;
341
342 start = addr;
343 do {
344 unsigned long more;
345
346 pmd = pmd_offset(pud, addr);
347 next = pmd_addr_end(addr, end);
348 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
349 /*
350 * if it is not hugepd pointer, we should already find
351 * it cleared.
352 */
353 WARN_ON(!pmd_none_or_clear_bad(pmd));
354 continue;
355 }
356 /*
357 * Increment next by the size of the huge mapping since
358 * there may be more than one entry at this level for a
359 * single hugepage, but all of them point to
360 * the same kmem cache that holds the hugepte.
361 */
362 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
363 if (more > next)
364 next = more;
365
366 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
367 addr, next, floor, ceiling);
368 } while (addr = next, addr != end);
369
370 start &= PUD_MASK;
371 if (start < floor)
372 return;
373 if (ceiling) {
374 ceiling &= PUD_MASK;
375 if (!ceiling)
376 return;
377 }
378 if (end - 1 > ceiling - 1)
379 return;
380
381 pmd = pmd_offset(pud, start);
382 pud_clear(pud);
383 pmd_free_tlb(tlb, pmd, start);
384 mm_dec_nr_pmds(tlb->mm);
385}
386
387static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
388 unsigned long addr, unsigned long end,
389 unsigned long floor, unsigned long ceiling)
390{
391 pud_t *pud;
392 unsigned long next;
393 unsigned long start;
394
395 start = addr;
396 do {
397 pud = pud_offset(pgd, addr);
398 next = pud_addr_end(addr, end);
399 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
400 if (pud_none_or_clear_bad(pud))
401 continue;
402 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
403 ceiling);
404 } else {
405 unsigned long more;
406 /*
407 * Increment next by the size of the huge mapping since
408 * there may be more than one entry at this level for a
409 * single hugepage, but all of them point to
410 * the same kmem cache that holds the hugepte.
411 */
412 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
413 if (more > next)
414 next = more;
415
416 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
417 addr, next, floor, ceiling);
418 }
419 } while (addr = next, addr != end);
420
421 start &= PGDIR_MASK;
422 if (start < floor)
423 return;
424 if (ceiling) {
425 ceiling &= PGDIR_MASK;
426 if (!ceiling)
427 return;
428 }
429 if (end - 1 > ceiling - 1)
430 return;
431
432 pud = pud_offset(pgd, start);
433 pgd_clear(pgd);
434 pud_free_tlb(tlb, pud, start);
435 mm_dec_nr_puds(tlb->mm);
436}
437
438/*
439 * This function frees user-level page tables of a process.
440 */
441void hugetlb_free_pgd_range(struct mmu_gather *tlb,
442 unsigned long addr, unsigned long end,
443 unsigned long floor, unsigned long ceiling)
444{
445 pgd_t *pgd;
446 unsigned long next;
447
448 /*
449 * Because there are a number of different possible pagetable
450 * layouts for hugepage ranges, we limit knowledge of how
451 * things should be laid out to the allocation path
452 * (huge_pte_alloc(), above). Everything else works out the
453 * structure as it goes from information in the hugepd
454 * pointers. That means that we can't here use the
455 * optimization used in the normal page free_pgd_range(), of
456 * checking whether we're actually covering a large enough
457 * range to have to do anything at the top level of the walk
458 * instead of at the bottom.
459 *
460 * To make sense of this, you should probably go read the big
461 * block comment at the top of the normal free_pgd_range(),
462 * too.
463 */
464
465 do {
466 next = pgd_addr_end(addr, end);
467 pgd = pgd_offset(tlb->mm, addr);
468 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
469 if (pgd_none_or_clear_bad(pgd))
470 continue;
471 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
472 } else {
473 unsigned long more;
474 /*
475 * Increment next by the size of the huge mapping since
476 * there may be more than one entry at the pgd level
477 * for a single hugepage, but all of them point to the
478 * same kmem cache that holds the hugepte.
479 */
480 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
481 if (more > next)
482 next = more;
483
484 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
485 addr, next, floor, ceiling);
486 }
487 } while (addr = next, addr != end);
488}
489
490struct page *follow_huge_pd(struct vm_area_struct *vma,
491 unsigned long address, hugepd_t hpd,
492 int flags, int pdshift)
493{
494 pte_t *ptep;
495 spinlock_t *ptl;
496 struct page *page = NULL;
497 unsigned long mask;
498 int shift = hugepd_shift(hpd);
499 struct mm_struct *mm = vma->vm_mm;
500
501retry:
502 /*
503 * hugepage directory entries are protected by mm->page_table_lock
504 * Use this instead of huge_pte_lockptr
505 */
506 ptl = &mm->page_table_lock;
507 spin_lock(ptl);
508
509 ptep = hugepte_offset(hpd, address, pdshift);
510 if (pte_present(*ptep)) {
511 mask = (1UL << shift) - 1;
512 page = pte_page(*ptep);
513 page += ((address & mask) >> PAGE_SHIFT);
514 if (flags & FOLL_GET)
515 get_page(page);
516 } else {
517 if (is_hugetlb_entry_migration(*ptep)) {
518 spin_unlock(ptl);
519 __migration_entry_wait(mm, ptep, ptl);
520 goto retry;
521 }
522 }
523 spin_unlock(ptl);
524 return page;
525}
526
527#ifdef CONFIG_PPC_MM_SLICES
528unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
529 unsigned long len, unsigned long pgoff,
530 unsigned long flags)
531{
532 struct hstate *hstate = hstate_file(file);
533 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
534
535#ifdef CONFIG_PPC_RADIX_MMU
536 if (radix_enabled())
537 return radix__hugetlb_get_unmapped_area(file, addr, len,
538 pgoff, flags);
539#endif
540 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
541}
542#endif
543
544unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
545{
546 /* With radix we don't use slice, so derive it from vma*/
547 if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
548 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
549
550 return 1UL << mmu_psize_to_shift(psize);
551 }
552 return vma_kernel_pagesize(vma);
553}
554
555static int __init add_huge_page_size(unsigned long long size)
556{
557 int shift = __ffs(size);
558 int mmu_psize;
559
560 /* Check that it is a page size supported by the hardware and
561 * that it fits within pagetable and slice limits. */
562 if (size <= PAGE_SIZE || !is_power_of_2(size))
563 return -EINVAL;
564
565 mmu_psize = check_and_get_huge_psize(shift);
566 if (mmu_psize < 0)
567 return -EINVAL;
568
569 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
570
571 /* Return if huge page size has already been setup */
572 if (size_to_hstate(size))
573 return 0;
574
575 hugetlb_add_hstate(shift - PAGE_SHIFT);
576
577 return 0;
578}
579
580static int __init hugepage_setup_sz(char *str)
581{
582 unsigned long long size;
583
584 size = memparse(str, &str);
585
586 if (add_huge_page_size(size) != 0) {
587 hugetlb_bad_size();
588 pr_err("Invalid huge page size specified(%llu)\n", size);
589 }
590
591 return 1;
592}
593__setup("hugepagesz=", hugepage_setup_sz);
594
595static int __init hugetlbpage_init(void)
596{
597 bool configured = false;
598 int psize;
599
600 if (hugetlb_disabled) {
601 pr_info("HugeTLB support is disabled!\n");
602 return 0;
603 }
604
605 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
606 !mmu_has_feature(MMU_FTR_16M_PAGE))
607 return -ENODEV;
608
609 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
610 unsigned shift;
611 unsigned pdshift;
612
613 if (!mmu_psize_defs[psize].shift)
614 continue;
615
616 shift = mmu_psize_to_shift(psize);
617
618#ifdef CONFIG_PPC_BOOK3S_64
619 if (shift > PGDIR_SHIFT)
620 continue;
621 else if (shift > PUD_SHIFT)
622 pdshift = PGDIR_SHIFT;
623 else if (shift > PMD_SHIFT)
624 pdshift = PUD_SHIFT;
625 else
626 pdshift = PMD_SHIFT;
627#else
628 if (shift < PUD_SHIFT)
629 pdshift = PMD_SHIFT;
630 else if (shift < PGDIR_SHIFT)
631 pdshift = PUD_SHIFT;
632 else
633 pdshift = PGDIR_SHIFT;
634#endif
635
636 if (add_huge_page_size(1ULL << shift) < 0)
637 continue;
638 /*
639 * if we have pdshift and shift value same, we don't
640 * use pgt cache for hugepd.
641 */
642 if (pdshift > shift && IS_ENABLED(CONFIG_PPC_8xx))
643 pgtable_cache_add(PTE_INDEX_SIZE);
644 else if (pdshift > shift)
645 pgtable_cache_add(pdshift - shift);
646 else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) || IS_ENABLED(CONFIG_PPC_8xx))
647 pgtable_cache_add(PTE_T_ORDER);
648
649 configured = true;
650 }
651
652 if (configured) {
653 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
654 hugetlbpage_init_default();
655 } else
656 pr_info("Failed to initialize. Disabling HugeTLB");
657
658 return 0;
659}
660
661arch_initcall(hugetlbpage_init);
662
663void flush_dcache_icache_hugepage(struct page *page)
664{
665 int i;
666 void *start;
667
668 BUG_ON(!PageCompound(page));
669
670 for (i = 0; i < compound_nr(page); i++) {
671 if (!PageHighMem(page)) {
672 __flush_dcache_icache(page_address(page+i));
673 } else {
674 start = kmap_atomic(page+i);
675 __flush_dcache_icache(start);
676 kunmap_atomic(start);
677 }
678 }
679}
1/*
2 * PPC Huge TLB Page Support for Kernel.
3 *
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
6 *
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/io.h>
13#include <linux/slab.h>
14#include <linux/hugetlb.h>
15#include <linux/export.h>
16#include <linux/of_fdt.h>
17#include <linux/memblock.h>
18#include <linux/bootmem.h>
19#include <linux/moduleparam.h>
20#include <asm/pgtable.h>
21#include <asm/pgalloc.h>
22#include <asm/tlb.h>
23#include <asm/setup.h>
24#include <asm/hugetlb.h>
25
26#ifdef CONFIG_HUGETLB_PAGE
27
28#define PAGE_SHIFT_64K 16
29#define PAGE_SHIFT_16M 24
30#define PAGE_SHIFT_16G 34
31
32unsigned int HPAGE_SHIFT;
33
34/*
35 * Tracks gpages after the device tree is scanned and before the
36 * huge_boot_pages list is ready. On non-Freescale implementations, this is
37 * just used to track 16G pages and so is a single array. FSL-based
38 * implementations may have more than one gpage size, so we need multiple
39 * arrays
40 */
41#ifdef CONFIG_PPC_FSL_BOOK3E
42#define MAX_NUMBER_GPAGES 128
43struct psize_gpages {
44 u64 gpage_list[MAX_NUMBER_GPAGES];
45 unsigned int nr_gpages;
46};
47static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
48#else
49#define MAX_NUMBER_GPAGES 1024
50static u64 gpage_freearray[MAX_NUMBER_GPAGES];
51static unsigned nr_gpages;
52#endif
53
54#define hugepd_none(hpd) ((hpd).pd == 0)
55
56#ifdef CONFIG_PPC_BOOK3S_64
57/*
58 * At this point we do the placement change only for BOOK3S 64. This would
59 * possibly work on other subarchs.
60 */
61
62/*
63 * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have
64 * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD;
65 */
66int pmd_huge(pmd_t pmd)
67{
68 /*
69 * leaf pte for huge page, bottom two bits != 00
70 */
71 return ((pmd_val(pmd) & 0x3) != 0x0);
72}
73
74int pud_huge(pud_t pud)
75{
76 /*
77 * leaf pte for huge page, bottom two bits != 00
78 */
79 return ((pud_val(pud) & 0x3) != 0x0);
80}
81
82int pgd_huge(pgd_t pgd)
83{
84 /*
85 * leaf pte for huge page, bottom two bits != 00
86 */
87 return ((pgd_val(pgd) & 0x3) != 0x0);
88}
89
90int pmd_huge_support(void)
91{
92 return 1;
93}
94#else
95int pmd_huge(pmd_t pmd)
96{
97 return 0;
98}
99
100int pud_huge(pud_t pud)
101{
102 return 0;
103}
104
105int pgd_huge(pgd_t pgd)
106{
107 return 0;
108}
109
110int pmd_huge_support(void)
111{
112 return 0;
113}
114#endif
115
116pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
117{
118 /* Only called for hugetlbfs pages, hence can ignore THP */
119 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
120}
121
122static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
123 unsigned long address, unsigned pdshift, unsigned pshift)
124{
125 struct kmem_cache *cachep;
126 pte_t *new;
127
128#ifdef CONFIG_PPC_FSL_BOOK3E
129 int i;
130 int num_hugepd = 1 << (pshift - pdshift);
131 cachep = hugepte_cache;
132#else
133 cachep = PGT_CACHE(pdshift - pshift);
134#endif
135
136 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
137
138 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
139 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
140
141 if (! new)
142 return -ENOMEM;
143
144 spin_lock(&mm->page_table_lock);
145#ifdef CONFIG_PPC_FSL_BOOK3E
146 /*
147 * We have multiple higher-level entries that point to the same
148 * actual pte location. Fill in each as we go and backtrack on error.
149 * We need all of these so the DTLB pgtable walk code can find the
150 * right higher-level entry without knowing if it's a hugepage or not.
151 */
152 for (i = 0; i < num_hugepd; i++, hpdp++) {
153 if (unlikely(!hugepd_none(*hpdp)))
154 break;
155 else
156 /* We use the old format for PPC_FSL_BOOK3E */
157 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
158 }
159 /* If we bailed from the for loop early, an error occurred, clean up */
160 if (i < num_hugepd) {
161 for (i = i - 1 ; i >= 0; i--, hpdp--)
162 hpdp->pd = 0;
163 kmem_cache_free(cachep, new);
164 }
165#else
166 if (!hugepd_none(*hpdp))
167 kmem_cache_free(cachep, new);
168 else {
169#ifdef CONFIG_PPC_BOOK3S_64
170 hpdp->pd = (unsigned long)new |
171 (shift_to_mmu_psize(pshift) << 2);
172#else
173 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
174#endif
175 }
176#endif
177 spin_unlock(&mm->page_table_lock);
178 return 0;
179}
180
181/*
182 * These macros define how to determine which level of the page table holds
183 * the hpdp.
184 */
185#ifdef CONFIG_PPC_FSL_BOOK3E
186#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
187#define HUGEPD_PUD_SHIFT PUD_SHIFT
188#else
189#define HUGEPD_PGD_SHIFT PUD_SHIFT
190#define HUGEPD_PUD_SHIFT PMD_SHIFT
191#endif
192
193#ifdef CONFIG_PPC_BOOK3S_64
194/*
195 * At this point we do the placement change only for BOOK3S 64. This would
196 * possibly work on other subarchs.
197 */
198pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
199{
200 pgd_t *pg;
201 pud_t *pu;
202 pmd_t *pm;
203 hugepd_t *hpdp = NULL;
204 unsigned pshift = __ffs(sz);
205 unsigned pdshift = PGDIR_SHIFT;
206
207 addr &= ~(sz-1);
208 pg = pgd_offset(mm, addr);
209
210 if (pshift == PGDIR_SHIFT)
211 /* 16GB huge page */
212 return (pte_t *) pg;
213 else if (pshift > PUD_SHIFT)
214 /*
215 * We need to use hugepd table
216 */
217 hpdp = (hugepd_t *)pg;
218 else {
219 pdshift = PUD_SHIFT;
220 pu = pud_alloc(mm, pg, addr);
221 if (pshift == PUD_SHIFT)
222 return (pte_t *)pu;
223 else if (pshift > PMD_SHIFT)
224 hpdp = (hugepd_t *)pu;
225 else {
226 pdshift = PMD_SHIFT;
227 pm = pmd_alloc(mm, pu, addr);
228 if (pshift == PMD_SHIFT)
229 /* 16MB hugepage */
230 return (pte_t *)pm;
231 else
232 hpdp = (hugepd_t *)pm;
233 }
234 }
235 if (!hpdp)
236 return NULL;
237
238 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
239
240 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
241 return NULL;
242
243 return hugepte_offset(hpdp, addr, pdshift);
244}
245
246#else
247
248pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
249{
250 pgd_t *pg;
251 pud_t *pu;
252 pmd_t *pm;
253 hugepd_t *hpdp = NULL;
254 unsigned pshift = __ffs(sz);
255 unsigned pdshift = PGDIR_SHIFT;
256
257 addr &= ~(sz-1);
258
259 pg = pgd_offset(mm, addr);
260
261 if (pshift >= HUGEPD_PGD_SHIFT) {
262 hpdp = (hugepd_t *)pg;
263 } else {
264 pdshift = PUD_SHIFT;
265 pu = pud_alloc(mm, pg, addr);
266 if (pshift >= HUGEPD_PUD_SHIFT) {
267 hpdp = (hugepd_t *)pu;
268 } else {
269 pdshift = PMD_SHIFT;
270 pm = pmd_alloc(mm, pu, addr);
271 hpdp = (hugepd_t *)pm;
272 }
273 }
274
275 if (!hpdp)
276 return NULL;
277
278 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
279
280 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
281 return NULL;
282
283 return hugepte_offset(hpdp, addr, pdshift);
284}
285#endif
286
287#ifdef CONFIG_PPC_FSL_BOOK3E
288/* Build list of addresses of gigantic pages. This function is used in early
289 * boot before the buddy or bootmem allocator is setup.
290 */
291void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
292{
293 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
294 int i;
295
296 if (addr == 0)
297 return;
298
299 gpage_freearray[idx].nr_gpages = number_of_pages;
300
301 for (i = 0; i < number_of_pages; i++) {
302 gpage_freearray[idx].gpage_list[i] = addr;
303 addr += page_size;
304 }
305}
306
307/*
308 * Moves the gigantic page addresses from the temporary list to the
309 * huge_boot_pages list.
310 */
311int alloc_bootmem_huge_page(struct hstate *hstate)
312{
313 struct huge_bootmem_page *m;
314 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
315 int nr_gpages = gpage_freearray[idx].nr_gpages;
316
317 if (nr_gpages == 0)
318 return 0;
319
320#ifdef CONFIG_HIGHMEM
321 /*
322 * If gpages can be in highmem we can't use the trick of storing the
323 * data structure in the page; allocate space for this
324 */
325 m = alloc_bootmem(sizeof(struct huge_bootmem_page));
326 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
327#else
328 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
329#endif
330
331 list_add(&m->list, &huge_boot_pages);
332 gpage_freearray[idx].nr_gpages = nr_gpages;
333 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
334 m->hstate = hstate;
335
336 return 1;
337}
338/*
339 * Scan the command line hugepagesz= options for gigantic pages; store those in
340 * a list that we use to allocate the memory once all options are parsed.
341 */
342
343unsigned long gpage_npages[MMU_PAGE_COUNT];
344
345static int __init do_gpage_early_setup(char *param, char *val,
346 const char *unused)
347{
348 static phys_addr_t size;
349 unsigned long npages;
350
351 /*
352 * The hugepagesz and hugepages cmdline options are interleaved. We
353 * use the size variable to keep track of whether or not this was done
354 * properly and skip over instances where it is incorrect. Other
355 * command-line parsing code will issue warnings, so we don't need to.
356 *
357 */
358 if ((strcmp(param, "default_hugepagesz") == 0) ||
359 (strcmp(param, "hugepagesz") == 0)) {
360 size = memparse(val, NULL);
361 } else if (strcmp(param, "hugepages") == 0) {
362 if (size != 0) {
363 if (sscanf(val, "%lu", &npages) <= 0)
364 npages = 0;
365 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
366 size = 0;
367 }
368 }
369 return 0;
370}
371
372
373/*
374 * This function allocates physical space for pages that are larger than the
375 * buddy allocator can handle. We want to allocate these in highmem because
376 * the amount of lowmem is limited. This means that this function MUST be
377 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
378 * allocate to grab highmem.
379 */
380void __init reserve_hugetlb_gpages(void)
381{
382 static __initdata char cmdline[COMMAND_LINE_SIZE];
383 phys_addr_t size, base;
384 int i;
385
386 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
387 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
388 &do_gpage_early_setup);
389
390 /*
391 * Walk gpage list in reverse, allocating larger page sizes first.
392 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
393 * When we reach the point in the list where pages are no longer
394 * considered gpages, we're done.
395 */
396 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
397 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
398 continue;
399 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
400 break;
401
402 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
403 base = memblock_alloc_base(size * gpage_npages[i], size,
404 MEMBLOCK_ALLOC_ANYWHERE);
405 add_gpage(base, size, gpage_npages[i]);
406 }
407}
408
409#else /* !PPC_FSL_BOOK3E */
410
411/* Build list of addresses of gigantic pages. This function is used in early
412 * boot before the buddy or bootmem allocator is setup.
413 */
414void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
415{
416 if (!addr)
417 return;
418 while (number_of_pages > 0) {
419 gpage_freearray[nr_gpages] = addr;
420 nr_gpages++;
421 number_of_pages--;
422 addr += page_size;
423 }
424}
425
426/* Moves the gigantic page addresses from the temporary list to the
427 * huge_boot_pages list.
428 */
429int alloc_bootmem_huge_page(struct hstate *hstate)
430{
431 struct huge_bootmem_page *m;
432 if (nr_gpages == 0)
433 return 0;
434 m = phys_to_virt(gpage_freearray[--nr_gpages]);
435 gpage_freearray[nr_gpages] = 0;
436 list_add(&m->list, &huge_boot_pages);
437 m->hstate = hstate;
438 return 1;
439}
440#endif
441
442int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
443{
444 return 0;
445}
446
447#ifdef CONFIG_PPC_FSL_BOOK3E
448#define HUGEPD_FREELIST_SIZE \
449 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
450
451struct hugepd_freelist {
452 struct rcu_head rcu;
453 unsigned int index;
454 void *ptes[0];
455};
456
457static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
458
459static void hugepd_free_rcu_callback(struct rcu_head *head)
460{
461 struct hugepd_freelist *batch =
462 container_of(head, struct hugepd_freelist, rcu);
463 unsigned int i;
464
465 for (i = 0; i < batch->index; i++)
466 kmem_cache_free(hugepte_cache, batch->ptes[i]);
467
468 free_page((unsigned long)batch);
469}
470
471static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
472{
473 struct hugepd_freelist **batchp;
474
475 batchp = &get_cpu_var(hugepd_freelist_cur);
476
477 if (atomic_read(&tlb->mm->mm_users) < 2 ||
478 cpumask_equal(mm_cpumask(tlb->mm),
479 cpumask_of(smp_processor_id()))) {
480 kmem_cache_free(hugepte_cache, hugepte);
481 put_cpu_var(hugepd_freelist_cur);
482 return;
483 }
484
485 if (*batchp == NULL) {
486 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
487 (*batchp)->index = 0;
488 }
489
490 (*batchp)->ptes[(*batchp)->index++] = hugepte;
491 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
492 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
493 *batchp = NULL;
494 }
495 put_cpu_var(hugepd_freelist_cur);
496}
497#endif
498
499static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
500 unsigned long start, unsigned long end,
501 unsigned long floor, unsigned long ceiling)
502{
503 pte_t *hugepte = hugepd_page(*hpdp);
504 int i;
505
506 unsigned long pdmask = ~((1UL << pdshift) - 1);
507 unsigned int num_hugepd = 1;
508
509#ifdef CONFIG_PPC_FSL_BOOK3E
510 /* Note: On fsl the hpdp may be the first of several */
511 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
512#else
513 unsigned int shift = hugepd_shift(*hpdp);
514#endif
515
516 start &= pdmask;
517 if (start < floor)
518 return;
519 if (ceiling) {
520 ceiling &= pdmask;
521 if (! ceiling)
522 return;
523 }
524 if (end - 1 > ceiling - 1)
525 return;
526
527 for (i = 0; i < num_hugepd; i++, hpdp++)
528 hpdp->pd = 0;
529
530 tlb->need_flush = 1;
531
532#ifdef CONFIG_PPC_FSL_BOOK3E
533 hugepd_free(tlb, hugepte);
534#else
535 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
536#endif
537}
538
539static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
540 unsigned long addr, unsigned long end,
541 unsigned long floor, unsigned long ceiling)
542{
543 pmd_t *pmd;
544 unsigned long next;
545 unsigned long start;
546
547 start = addr;
548 do {
549 pmd = pmd_offset(pud, addr);
550 next = pmd_addr_end(addr, end);
551 if (!is_hugepd(pmd)) {
552 /*
553 * if it is not hugepd pointer, we should already find
554 * it cleared.
555 */
556 WARN_ON(!pmd_none_or_clear_bad(pmd));
557 continue;
558 }
559#ifdef CONFIG_PPC_FSL_BOOK3E
560 /*
561 * Increment next by the size of the huge mapping since
562 * there may be more than one entry at this level for a
563 * single hugepage, but all of them point to
564 * the same kmem cache that holds the hugepte.
565 */
566 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
567#endif
568 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
569 addr, next, floor, ceiling);
570 } while (addr = next, addr != end);
571
572 start &= PUD_MASK;
573 if (start < floor)
574 return;
575 if (ceiling) {
576 ceiling &= PUD_MASK;
577 if (!ceiling)
578 return;
579 }
580 if (end - 1 > ceiling - 1)
581 return;
582
583 pmd = pmd_offset(pud, start);
584 pud_clear(pud);
585 pmd_free_tlb(tlb, pmd, start);
586}
587
588static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
589 unsigned long addr, unsigned long end,
590 unsigned long floor, unsigned long ceiling)
591{
592 pud_t *pud;
593 unsigned long next;
594 unsigned long start;
595
596 start = addr;
597 do {
598 pud = pud_offset(pgd, addr);
599 next = pud_addr_end(addr, end);
600 if (!is_hugepd(pud)) {
601 if (pud_none_or_clear_bad(pud))
602 continue;
603 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
604 ceiling);
605 } else {
606#ifdef CONFIG_PPC_FSL_BOOK3E
607 /*
608 * Increment next by the size of the huge mapping since
609 * there may be more than one entry at this level for a
610 * single hugepage, but all of them point to
611 * the same kmem cache that holds the hugepte.
612 */
613 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
614#endif
615 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
616 addr, next, floor, ceiling);
617 }
618 } while (addr = next, addr != end);
619
620 start &= PGDIR_MASK;
621 if (start < floor)
622 return;
623 if (ceiling) {
624 ceiling &= PGDIR_MASK;
625 if (!ceiling)
626 return;
627 }
628 if (end - 1 > ceiling - 1)
629 return;
630
631 pud = pud_offset(pgd, start);
632 pgd_clear(pgd);
633 pud_free_tlb(tlb, pud, start);
634}
635
636/*
637 * This function frees user-level page tables of a process.
638 */
639void hugetlb_free_pgd_range(struct mmu_gather *tlb,
640 unsigned long addr, unsigned long end,
641 unsigned long floor, unsigned long ceiling)
642{
643 pgd_t *pgd;
644 unsigned long next;
645
646 /*
647 * Because there are a number of different possible pagetable
648 * layouts for hugepage ranges, we limit knowledge of how
649 * things should be laid out to the allocation path
650 * (huge_pte_alloc(), above). Everything else works out the
651 * structure as it goes from information in the hugepd
652 * pointers. That means that we can't here use the
653 * optimization used in the normal page free_pgd_range(), of
654 * checking whether we're actually covering a large enough
655 * range to have to do anything at the top level of the walk
656 * instead of at the bottom.
657 *
658 * To make sense of this, you should probably go read the big
659 * block comment at the top of the normal free_pgd_range(),
660 * too.
661 */
662
663 do {
664 next = pgd_addr_end(addr, end);
665 pgd = pgd_offset(tlb->mm, addr);
666 if (!is_hugepd(pgd)) {
667 if (pgd_none_or_clear_bad(pgd))
668 continue;
669 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
670 } else {
671#ifdef CONFIG_PPC_FSL_BOOK3E
672 /*
673 * Increment next by the size of the huge mapping since
674 * there may be more than one entry at the pgd level
675 * for a single hugepage, but all of them point to the
676 * same kmem cache that holds the hugepte.
677 */
678 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
679#endif
680 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
681 addr, next, floor, ceiling);
682 }
683 } while (addr = next, addr != end);
684}
685
686struct page *
687follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
688{
689 pte_t *ptep;
690 struct page *page;
691 unsigned shift;
692 unsigned long mask;
693 /*
694 * Transparent hugepages are handled by generic code. We can skip them
695 * here.
696 */
697 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
698
699 /* Verify it is a huge page else bail. */
700 if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep))
701 return ERR_PTR(-EINVAL);
702
703 mask = (1UL << shift) - 1;
704 page = pte_page(*ptep);
705 if (page)
706 page += (address & mask) / PAGE_SIZE;
707
708 return page;
709}
710
711struct page *
712follow_huge_pmd(struct mm_struct *mm, unsigned long address,
713 pmd_t *pmd, int write)
714{
715 BUG();
716 return NULL;
717}
718
719static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
720 unsigned long sz)
721{
722 unsigned long __boundary = (addr + sz) & ~(sz-1);
723 return (__boundary - 1 < end - 1) ? __boundary : end;
724}
725
726int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
727 unsigned long addr, unsigned long end,
728 int write, struct page **pages, int *nr)
729{
730 pte_t *ptep;
731 unsigned long sz = 1UL << hugepd_shift(*hugepd);
732 unsigned long next;
733
734 ptep = hugepte_offset(hugepd, addr, pdshift);
735 do {
736 next = hugepte_addr_end(addr, end, sz);
737 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
738 return 0;
739 } while (ptep++, addr = next, addr != end);
740
741 return 1;
742}
743
744#ifdef CONFIG_PPC_MM_SLICES
745unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
746 unsigned long len, unsigned long pgoff,
747 unsigned long flags)
748{
749 struct hstate *hstate = hstate_file(file);
750 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
751
752 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
753}
754#endif
755
756unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
757{
758#ifdef CONFIG_PPC_MM_SLICES
759 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
760
761 return 1UL << mmu_psize_to_shift(psize);
762#else
763 if (!is_vm_hugetlb_page(vma))
764 return PAGE_SIZE;
765
766 return huge_page_size(hstate_vma(vma));
767#endif
768}
769
770static inline bool is_power_of_4(unsigned long x)
771{
772 if (is_power_of_2(x))
773 return (__ilog2(x) % 2) ? false : true;
774 return false;
775}
776
777static int __init add_huge_page_size(unsigned long long size)
778{
779 int shift = __ffs(size);
780 int mmu_psize;
781
782 /* Check that it is a page size supported by the hardware and
783 * that it fits within pagetable and slice limits. */
784#ifdef CONFIG_PPC_FSL_BOOK3E
785 if ((size < PAGE_SIZE) || !is_power_of_4(size))
786 return -EINVAL;
787#else
788 if (!is_power_of_2(size)
789 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
790 return -EINVAL;
791#endif
792
793 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
794 return -EINVAL;
795
796#ifdef CONFIG_SPU_FS_64K_LS
797 /* Disable support for 64K huge pages when 64K SPU local store
798 * support is enabled as the current implementation conflicts.
799 */
800 if (shift == PAGE_SHIFT_64K)
801 return -EINVAL;
802#endif /* CONFIG_SPU_FS_64K_LS */
803
804 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
805
806 /* Return if huge page size has already been setup */
807 if (size_to_hstate(size))
808 return 0;
809
810 hugetlb_add_hstate(shift - PAGE_SHIFT);
811
812 return 0;
813}
814
815static int __init hugepage_setup_sz(char *str)
816{
817 unsigned long long size;
818
819 size = memparse(str, &str);
820
821 if (add_huge_page_size(size) != 0)
822 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
823
824 return 1;
825}
826__setup("hugepagesz=", hugepage_setup_sz);
827
828#ifdef CONFIG_PPC_FSL_BOOK3E
829struct kmem_cache *hugepte_cache;
830static int __init hugetlbpage_init(void)
831{
832 int psize;
833
834 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
835 unsigned shift;
836
837 if (!mmu_psize_defs[psize].shift)
838 continue;
839
840 shift = mmu_psize_to_shift(psize);
841
842 /* Don't treat normal page sizes as huge... */
843 if (shift != PAGE_SHIFT)
844 if (add_huge_page_size(1ULL << shift) < 0)
845 continue;
846 }
847
848 /*
849 * Create a kmem cache for hugeptes. The bottom bits in the pte have
850 * size information encoded in them, so align them to allow this
851 */
852 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
853 HUGEPD_SHIFT_MASK + 1, 0, NULL);
854 if (hugepte_cache == NULL)
855 panic("%s: Unable to create kmem cache for hugeptes\n",
856 __func__);
857
858 /* Default hpage size = 4M */
859 if (mmu_psize_defs[MMU_PAGE_4M].shift)
860 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
861 else
862 panic("%s: Unable to set default huge page size\n", __func__);
863
864
865 return 0;
866}
867#else
868static int __init hugetlbpage_init(void)
869{
870 int psize;
871
872 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
873 return -ENODEV;
874
875 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
876 unsigned shift;
877 unsigned pdshift;
878
879 if (!mmu_psize_defs[psize].shift)
880 continue;
881
882 shift = mmu_psize_to_shift(psize);
883
884 if (add_huge_page_size(1ULL << shift) < 0)
885 continue;
886
887 if (shift < PMD_SHIFT)
888 pdshift = PMD_SHIFT;
889 else if (shift < PUD_SHIFT)
890 pdshift = PUD_SHIFT;
891 else
892 pdshift = PGDIR_SHIFT;
893 /*
894 * if we have pdshift and shift value same, we don't
895 * use pgt cache for hugepd.
896 */
897 if (pdshift != shift) {
898 pgtable_cache_add(pdshift - shift, NULL);
899 if (!PGT_CACHE(pdshift - shift))
900 panic("hugetlbpage_init(): could not create "
901 "pgtable cache for %d bit pagesize\n", shift);
902 }
903 }
904
905 /* Set default large page size. Currently, we pick 16M or 1M
906 * depending on what is available
907 */
908 if (mmu_psize_defs[MMU_PAGE_16M].shift)
909 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
910 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
911 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
912
913 return 0;
914}
915#endif
916module_init(hugetlbpage_init);
917
918void flush_dcache_icache_hugepage(struct page *page)
919{
920 int i;
921 void *start;
922
923 BUG_ON(!PageCompound(page));
924
925 for (i = 0; i < (1UL << compound_order(page)); i++) {
926 if (!PageHighMem(page)) {
927 __flush_dcache_icache(page_address(page+i));
928 } else {
929 start = kmap_atomic(page+i);
930 __flush_dcache_icache(start);
931 kunmap_atomic(start);
932 }
933 }
934}
935
936#endif /* CONFIG_HUGETLB_PAGE */
937
938/*
939 * We have 4 cases for pgds and pmds:
940 * (1) invalid (all zeroes)
941 * (2) pointer to next table, as normal; bottom 6 bits == 0
942 * (3) leaf pte for huge page, bottom two bits != 00
943 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
944 *
945 * So long as we atomically load page table pointers we are safe against teardown,
946 * we can follow the address down to the the page and take a ref on it.
947 */
948
949pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
950{
951 pgd_t pgd, *pgdp;
952 pud_t pud, *pudp;
953 pmd_t pmd, *pmdp;
954 pte_t *ret_pte;
955 hugepd_t *hpdp = NULL;
956 unsigned pdshift = PGDIR_SHIFT;
957
958 if (shift)
959 *shift = 0;
960
961 pgdp = pgdir + pgd_index(ea);
962 pgd = ACCESS_ONCE(*pgdp);
963 /*
964 * Always operate on the local stack value. This make sure the
965 * value don't get updated by a parallel THP split/collapse,
966 * page fault or a page unmap. The return pte_t * is still not
967 * stable. So should be checked there for above conditions.
968 */
969 if (pgd_none(pgd))
970 return NULL;
971 else if (pgd_huge(pgd)) {
972 ret_pte = (pte_t *) pgdp;
973 goto out;
974 } else if (is_hugepd(&pgd))
975 hpdp = (hugepd_t *)&pgd;
976 else {
977 /*
978 * Even if we end up with an unmap, the pgtable will not
979 * be freed, because we do an rcu free and here we are
980 * irq disabled
981 */
982 pdshift = PUD_SHIFT;
983 pudp = pud_offset(&pgd, ea);
984 pud = ACCESS_ONCE(*pudp);
985
986 if (pud_none(pud))
987 return NULL;
988 else if (pud_huge(pud)) {
989 ret_pte = (pte_t *) pudp;
990 goto out;
991 } else if (is_hugepd(&pud))
992 hpdp = (hugepd_t *)&pud;
993 else {
994 pdshift = PMD_SHIFT;
995 pmdp = pmd_offset(&pud, ea);
996 pmd = ACCESS_ONCE(*pmdp);
997 /*
998 * A hugepage collapse is captured by pmd_none, because
999 * it mark the pmd none and do a hpte invalidate.
1000 *
1001 * A hugepage split is captured by pmd_trans_splitting
1002 * because we mark the pmd trans splitting and do a
1003 * hpte invalidate
1004 *
1005 */
1006 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1007 return NULL;
1008
1009 if (pmd_huge(pmd) || pmd_large(pmd)) {
1010 ret_pte = (pte_t *) pmdp;
1011 goto out;
1012 } else if (is_hugepd(&pmd))
1013 hpdp = (hugepd_t *)&pmd;
1014 else
1015 return pte_offset_kernel(&pmd, ea);
1016 }
1017 }
1018 if (!hpdp)
1019 return NULL;
1020
1021 ret_pte = hugepte_offset(hpdp, ea, pdshift);
1022 pdshift = hugepd_shift(*hpdp);
1023out:
1024 if (shift)
1025 *shift = pdshift;
1026 return ret_pte;
1027}
1028EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
1029
1030int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1031 unsigned long end, int write, struct page **pages, int *nr)
1032{
1033 unsigned long mask;
1034 unsigned long pte_end;
1035 struct page *head, *page, *tail;
1036 pte_t pte;
1037 int refs;
1038
1039 pte_end = (addr + sz) & ~(sz-1);
1040 if (pte_end < end)
1041 end = pte_end;
1042
1043 pte = ACCESS_ONCE(*ptep);
1044 mask = _PAGE_PRESENT | _PAGE_USER;
1045 if (write)
1046 mask |= _PAGE_RW;
1047
1048 if ((pte_val(pte) & mask) != mask)
1049 return 0;
1050
1051#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1052 /*
1053 * check for splitting here
1054 */
1055 if (pmd_trans_splitting(pte_pmd(pte)))
1056 return 0;
1057#endif
1058
1059 /* hugepages are never "special" */
1060 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1061
1062 refs = 0;
1063 head = pte_page(pte);
1064
1065 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1066 tail = page;
1067 do {
1068 VM_BUG_ON(compound_head(page) != head);
1069 pages[*nr] = page;
1070 (*nr)++;
1071 page++;
1072 refs++;
1073 } while (addr += PAGE_SIZE, addr != end);
1074
1075 if (!page_cache_add_speculative(head, refs)) {
1076 *nr -= refs;
1077 return 0;
1078 }
1079
1080 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1081 /* Could be optimized better */
1082 *nr -= refs;
1083 while (refs--)
1084 put_page(head);
1085 return 0;
1086 }
1087
1088 /*
1089 * Any tail page need their mapcount reference taken before we
1090 * return.
1091 */
1092 while (refs--) {
1093 if (PageTail(tail))
1094 get_huge_page_tail(tail);
1095 tail++;
1096 }
1097
1098 return 1;
1099}