2 * PPC Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
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>
23 #include <asm/setup.h>
24 #include <asm/hugetlb.h>
26 #ifdef CONFIG_HUGETLB_PAGE
28 #define PAGE_SHIFT_64K 16
29 #define PAGE_SHIFT_16M 24
30 #define PAGE_SHIFT_16G 34
32 unsigned int HPAGE_SHIFT
;
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
41 #ifdef CONFIG_PPC_FSL_BOOK3E
42 #define MAX_NUMBER_GPAGES 128
44 u64 gpage_list
[MAX_NUMBER_GPAGES
];
45 unsigned int nr_gpages
;
47 static struct psize_gpages gpage_freearray
[MMU_PAGE_COUNT
];
49 #define MAX_NUMBER_GPAGES 1024
50 static u64 gpage_freearray
[MAX_NUMBER_GPAGES
];
51 static unsigned nr_gpages
;
54 #define hugepd_none(hpd) ((hpd).pd == 0)
56 #ifdef CONFIG_PPC_BOOK3S_64
58 * At this point we do the placement change only for BOOK3S 64. This would
59 * possibly work on other subarchs.
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;
66 int pmd_huge(pmd_t pmd
)
69 * leaf pte for huge page, bottom two bits != 00
71 return ((pmd_val(pmd
) & 0x3) != 0x0);
74 int pud_huge(pud_t pud
)
77 * leaf pte for huge page, bottom two bits != 00
79 return ((pud_val(pud
) & 0x3) != 0x0);
82 int pgd_huge(pgd_t pgd
)
85 * leaf pte for huge page, bottom two bits != 00
87 return ((pgd_val(pgd
) & 0x3) != 0x0);
90 int pmd_huge(pmd_t pmd
)
95 int pud_huge(pud_t pud
)
100 int pgd_huge(pgd_t pgd
)
106 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
108 return find_linux_pte_or_hugepte(mm
->pgd
, addr
, NULL
);
111 static int __hugepte_alloc(struct mm_struct
*mm
, hugepd_t
*hpdp
,
112 unsigned long address
, unsigned pdshift
, unsigned pshift
)
114 struct kmem_cache
*cachep
;
117 #ifdef CONFIG_PPC_FSL_BOOK3E
119 int num_hugepd
= 1 << (pshift
- pdshift
);
120 cachep
= hugepte_cache
;
122 cachep
= PGT_CACHE(pdshift
- pshift
);
125 new = kmem_cache_zalloc(cachep
, GFP_KERNEL
|__GFP_REPEAT
);
127 BUG_ON(pshift
> HUGEPD_SHIFT_MASK
);
128 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK
);
133 spin_lock(&mm
->page_table_lock
);
134 #ifdef CONFIG_PPC_FSL_BOOK3E
136 * We have multiple higher-level entries that point to the same
137 * actual pte location. Fill in each as we go and backtrack on error.
138 * We need all of these so the DTLB pgtable walk code can find the
139 * right higher-level entry without knowing if it's a hugepage or not.
141 for (i
= 0; i
< num_hugepd
; i
++, hpdp
++) {
142 if (unlikely(!hugepd_none(*hpdp
)))
145 /* We use the old format for PPC_FSL_BOOK3E */
146 hpdp
->pd
= ((unsigned long)new & ~PD_HUGE
) | pshift
;
148 /* If we bailed from the for loop early, an error occurred, clean up */
149 if (i
< num_hugepd
) {
150 for (i
= i
- 1 ; i
>= 0; i
--, hpdp
--)
152 kmem_cache_free(cachep
, new);
155 if (!hugepd_none(*hpdp
))
156 kmem_cache_free(cachep
, new);
158 #ifdef CONFIG_PPC_BOOK3S_64
159 hpdp
->pd
= (unsigned long)new |
160 (shift_to_mmu_psize(pshift
) << 2);
162 hpdp
->pd
= ((unsigned long)new & ~PD_HUGE
) | pshift
;
166 spin_unlock(&mm
->page_table_lock
);
171 * These macros define how to determine which level of the page table holds
174 #ifdef CONFIG_PPC_FSL_BOOK3E
175 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
176 #define HUGEPD_PUD_SHIFT PUD_SHIFT
178 #define HUGEPD_PGD_SHIFT PUD_SHIFT
179 #define HUGEPD_PUD_SHIFT PMD_SHIFT
182 #ifdef CONFIG_PPC_BOOK3S_64
184 * At this point we do the placement change only for BOOK3S 64. This would
185 * possibly work on other subarchs.
187 pte_t
*huge_pte_alloc(struct mm_struct
*mm
, unsigned long addr
, unsigned long sz
)
192 hugepd_t
*hpdp
= NULL
;
193 unsigned pshift
= __ffs(sz
);
194 unsigned pdshift
= PGDIR_SHIFT
;
197 pg
= pgd_offset(mm
, addr
);
199 if (pshift
== PGDIR_SHIFT
)
202 else if (pshift
> PUD_SHIFT
)
204 * We need to use hugepd table
206 hpdp
= (hugepd_t
*)pg
;
209 pu
= pud_alloc(mm
, pg
, addr
);
210 if (pshift
== PUD_SHIFT
)
212 else if (pshift
> PMD_SHIFT
)
213 hpdp
= (hugepd_t
*)pu
;
216 pm
= pmd_alloc(mm
, pu
, addr
);
217 if (pshift
== PMD_SHIFT
)
221 hpdp
= (hugepd_t
*)pm
;
227 BUG_ON(!hugepd_none(*hpdp
) && !hugepd_ok(*hpdp
));
229 if (hugepd_none(*hpdp
) && __hugepte_alloc(mm
, hpdp
, addr
, pdshift
, pshift
))
232 return hugepte_offset(hpdp
, addr
, pdshift
);
237 pte_t
*huge_pte_alloc(struct mm_struct
*mm
, unsigned long addr
, unsigned long sz
)
242 hugepd_t
*hpdp
= NULL
;
243 unsigned pshift
= __ffs(sz
);
244 unsigned pdshift
= PGDIR_SHIFT
;
248 pg
= pgd_offset(mm
, addr
);
250 if (pshift
>= HUGEPD_PGD_SHIFT
) {
251 hpdp
= (hugepd_t
*)pg
;
254 pu
= pud_alloc(mm
, pg
, addr
);
255 if (pshift
>= HUGEPD_PUD_SHIFT
) {
256 hpdp
= (hugepd_t
*)pu
;
259 pm
= pmd_alloc(mm
, pu
, addr
);
260 hpdp
= (hugepd_t
*)pm
;
267 BUG_ON(!hugepd_none(*hpdp
) && !hugepd_ok(*hpdp
));
269 if (hugepd_none(*hpdp
) && __hugepte_alloc(mm
, hpdp
, addr
, pdshift
, pshift
))
272 return hugepte_offset(hpdp
, addr
, pdshift
);
276 #ifdef CONFIG_PPC_FSL_BOOK3E
277 /* Build list of addresses of gigantic pages. This function is used in early
278 * boot before the buddy or bootmem allocator is setup.
280 void add_gpage(u64 addr
, u64 page_size
, unsigned long number_of_pages
)
282 unsigned int idx
= shift_to_mmu_psize(__ffs(page_size
));
288 gpage_freearray
[idx
].nr_gpages
= number_of_pages
;
290 for (i
= 0; i
< number_of_pages
; i
++) {
291 gpage_freearray
[idx
].gpage_list
[i
] = addr
;
297 * Moves the gigantic page addresses from the temporary list to the
298 * huge_boot_pages list.
300 int alloc_bootmem_huge_page(struct hstate
*hstate
)
302 struct huge_bootmem_page
*m
;
303 int idx
= shift_to_mmu_psize(hstate
->order
+ PAGE_SHIFT
);
304 int nr_gpages
= gpage_freearray
[idx
].nr_gpages
;
309 #ifdef CONFIG_HIGHMEM
311 * If gpages can be in highmem we can't use the trick of storing the
312 * data structure in the page; allocate space for this
314 m
= alloc_bootmem(sizeof(struct huge_bootmem_page
));
315 m
->phys
= gpage_freearray
[idx
].gpage_list
[--nr_gpages
];
317 m
= phys_to_virt(gpage_freearray
[idx
].gpage_list
[--nr_gpages
]);
320 list_add(&m
->list
, &huge_boot_pages
);
321 gpage_freearray
[idx
].nr_gpages
= nr_gpages
;
322 gpage_freearray
[idx
].gpage_list
[nr_gpages
] = 0;
328 * Scan the command line hugepagesz= options for gigantic pages; store those in
329 * a list that we use to allocate the memory once all options are parsed.
332 unsigned long gpage_npages
[MMU_PAGE_COUNT
];
334 static int __init
do_gpage_early_setup(char *param
, char *val
,
337 static phys_addr_t size
;
338 unsigned long npages
;
341 * The hugepagesz and hugepages cmdline options are interleaved. We
342 * use the size variable to keep track of whether or not this was done
343 * properly and skip over instances where it is incorrect. Other
344 * command-line parsing code will issue warnings, so we don't need to.
347 if ((strcmp(param
, "default_hugepagesz") == 0) ||
348 (strcmp(param
, "hugepagesz") == 0)) {
349 size
= memparse(val
, NULL
);
350 } else if (strcmp(param
, "hugepages") == 0) {
352 if (sscanf(val
, "%lu", &npages
) <= 0)
354 gpage_npages
[shift_to_mmu_psize(__ffs(size
))] = npages
;
363 * This function allocates physical space for pages that are larger than the
364 * buddy allocator can handle. We want to allocate these in highmem because
365 * the amount of lowmem is limited. This means that this function MUST be
366 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
367 * allocate to grab highmem.
369 void __init
reserve_hugetlb_gpages(void)
371 static __initdata
char cmdline
[COMMAND_LINE_SIZE
];
372 phys_addr_t size
, base
;
375 strlcpy(cmdline
, boot_command_line
, COMMAND_LINE_SIZE
);
376 parse_args("hugetlb gpages", cmdline
, NULL
, 0, 0, 0,
377 &do_gpage_early_setup
);
380 * Walk gpage list in reverse, allocating larger page sizes first.
381 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
382 * When we reach the point in the list where pages are no longer
383 * considered gpages, we're done.
385 for (i
= MMU_PAGE_COUNT
-1; i
>= 0; i
--) {
386 if (mmu_psize_defs
[i
].shift
== 0 || gpage_npages
[i
] == 0)
388 else if (mmu_psize_to_shift(i
) < (MAX_ORDER
+ PAGE_SHIFT
))
391 size
= (phys_addr_t
)(1ULL << mmu_psize_to_shift(i
));
392 base
= memblock_alloc_base(size
* gpage_npages
[i
], size
,
393 MEMBLOCK_ALLOC_ANYWHERE
);
394 add_gpage(base
, size
, gpage_npages
[i
]);
398 #else /* !PPC_FSL_BOOK3E */
400 /* Build list of addresses of gigantic pages. This function is used in early
401 * boot before the buddy or bootmem allocator is setup.
403 void add_gpage(u64 addr
, u64 page_size
, unsigned long number_of_pages
)
407 while (number_of_pages
> 0) {
408 gpage_freearray
[nr_gpages
] = addr
;
415 /* Moves the gigantic page addresses from the temporary list to the
416 * huge_boot_pages list.
418 int alloc_bootmem_huge_page(struct hstate
*hstate
)
420 struct huge_bootmem_page
*m
;
423 m
= phys_to_virt(gpage_freearray
[--nr_gpages
]);
424 gpage_freearray
[nr_gpages
] = 0;
425 list_add(&m
->list
, &huge_boot_pages
);
431 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
436 #ifdef CONFIG_PPC_FSL_BOOK3E
437 #define HUGEPD_FREELIST_SIZE \
438 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
440 struct hugepd_freelist
{
446 static DEFINE_PER_CPU(struct hugepd_freelist
*, hugepd_freelist_cur
);
448 static void hugepd_free_rcu_callback(struct rcu_head
*head
)
450 struct hugepd_freelist
*batch
=
451 container_of(head
, struct hugepd_freelist
, rcu
);
454 for (i
= 0; i
< batch
->index
; i
++)
455 kmem_cache_free(hugepte_cache
, batch
->ptes
[i
]);
457 free_page((unsigned long)batch
);
460 static void hugepd_free(struct mmu_gather
*tlb
, void *hugepte
)
462 struct hugepd_freelist
**batchp
;
464 batchp
= &__get_cpu_var(hugepd_freelist_cur
);
466 if (atomic_read(&tlb
->mm
->mm_users
) < 2 ||
467 cpumask_equal(mm_cpumask(tlb
->mm
),
468 cpumask_of(smp_processor_id()))) {
469 kmem_cache_free(hugepte_cache
, hugepte
);
473 if (*batchp
== NULL
) {
474 *batchp
= (struct hugepd_freelist
*)__get_free_page(GFP_ATOMIC
);
475 (*batchp
)->index
= 0;
478 (*batchp
)->ptes
[(*batchp
)->index
++] = hugepte
;
479 if ((*batchp
)->index
== HUGEPD_FREELIST_SIZE
) {
480 call_rcu_sched(&(*batchp
)->rcu
, hugepd_free_rcu_callback
);
486 static void free_hugepd_range(struct mmu_gather
*tlb
, hugepd_t
*hpdp
, int pdshift
,
487 unsigned long start
, unsigned long end
,
488 unsigned long floor
, unsigned long ceiling
)
490 pte_t
*hugepte
= hugepd_page(*hpdp
);
493 unsigned long pdmask
= ~((1UL << pdshift
) - 1);
494 unsigned int num_hugepd
= 1;
496 #ifdef CONFIG_PPC_FSL_BOOK3E
497 /* Note: On fsl the hpdp may be the first of several */
498 num_hugepd
= (1 << (hugepd_shift(*hpdp
) - pdshift
));
500 unsigned int shift
= hugepd_shift(*hpdp
);
511 if (end
- 1 > ceiling
- 1)
514 for (i
= 0; i
< num_hugepd
; i
++, hpdp
++)
519 #ifdef CONFIG_PPC_FSL_BOOK3E
520 hugepd_free(tlb
, hugepte
);
522 pgtable_free_tlb(tlb
, hugepte
, pdshift
- shift
);
526 static void hugetlb_free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
527 unsigned long addr
, unsigned long end
,
528 unsigned long floor
, unsigned long ceiling
)
536 pmd
= pmd_offset(pud
, addr
);
537 next
= pmd_addr_end(addr
, end
);
538 if (pmd_none_or_clear_bad(pmd
))
540 #ifdef CONFIG_PPC_FSL_BOOK3E
542 * Increment next by the size of the huge mapping since
543 * there may be more than one entry at this level for a
544 * single hugepage, but all of them point to
545 * the same kmem cache that holds the hugepte.
547 next
= addr
+ (1 << hugepd_shift(*(hugepd_t
*)pmd
));
549 free_hugepd_range(tlb
, (hugepd_t
*)pmd
, PMD_SHIFT
,
550 addr
, next
, floor
, ceiling
);
551 } while (addr
= next
, addr
!= end
);
561 if (end
- 1 > ceiling
- 1)
564 pmd
= pmd_offset(pud
, start
);
566 pmd_free_tlb(tlb
, pmd
, start
);
569 static void hugetlb_free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
570 unsigned long addr
, unsigned long end
,
571 unsigned long floor
, unsigned long ceiling
)
579 pud
= pud_offset(pgd
, addr
);
580 next
= pud_addr_end(addr
, end
);
581 if (!is_hugepd(pud
)) {
582 if (pud_none_or_clear_bad(pud
))
584 hugetlb_free_pmd_range(tlb
, pud
, addr
, next
, floor
,
587 #ifdef CONFIG_PPC_FSL_BOOK3E
589 * Increment next by the size of the huge mapping since
590 * there may be more than one entry at this level for a
591 * single hugepage, but all of them point to
592 * the same kmem cache that holds the hugepte.
594 next
= addr
+ (1 << hugepd_shift(*(hugepd_t
*)pud
));
596 free_hugepd_range(tlb
, (hugepd_t
*)pud
, PUD_SHIFT
,
597 addr
, next
, floor
, ceiling
);
599 } while (addr
= next
, addr
!= end
);
605 ceiling
&= PGDIR_MASK
;
609 if (end
- 1 > ceiling
- 1)
612 pud
= pud_offset(pgd
, start
);
614 pud_free_tlb(tlb
, pud
, start
);
618 * This function frees user-level page tables of a process.
620 * Must be called with pagetable lock held.
622 void hugetlb_free_pgd_range(struct mmu_gather
*tlb
,
623 unsigned long addr
, unsigned long end
,
624 unsigned long floor
, unsigned long ceiling
)
630 * Because there are a number of different possible pagetable
631 * layouts for hugepage ranges, we limit knowledge of how
632 * things should be laid out to the allocation path
633 * (huge_pte_alloc(), above). Everything else works out the
634 * structure as it goes from information in the hugepd
635 * pointers. That means that we can't here use the
636 * optimization used in the normal page free_pgd_range(), of
637 * checking whether we're actually covering a large enough
638 * range to have to do anything at the top level of the walk
639 * instead of at the bottom.
641 * To make sense of this, you should probably go read the big
642 * block comment at the top of the normal free_pgd_range(),
647 next
= pgd_addr_end(addr
, end
);
648 pgd
= pgd_offset(tlb
->mm
, addr
);
649 if (!is_hugepd(pgd
)) {
650 if (pgd_none_or_clear_bad(pgd
))
652 hugetlb_free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
654 #ifdef CONFIG_PPC_FSL_BOOK3E
656 * Increment next by the size of the huge mapping since
657 * there may be more than one entry at the pgd level
658 * for a single hugepage, but all of them point to the
659 * same kmem cache that holds the hugepte.
661 next
= addr
+ (1 << hugepd_shift(*(hugepd_t
*)pgd
));
663 free_hugepd_range(tlb
, (hugepd_t
*)pgd
, PGDIR_SHIFT
,
664 addr
, next
, floor
, ceiling
);
666 } while (addr
= next
, addr
!= end
);
670 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
, int write
)
677 ptep
= find_linux_pte_or_hugepte(mm
->pgd
, address
, &shift
);
679 /* Verify it is a huge page else bail. */
681 return ERR_PTR(-EINVAL
);
683 mask
= (1UL << shift
) - 1;
684 page
= pte_page(*ptep
);
686 page
+= (address
& mask
) / PAGE_SIZE
;
692 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
693 pmd_t
*pmd
, int write
)
699 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
702 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
703 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
706 int gup_hugepd(hugepd_t
*hugepd
, unsigned pdshift
,
707 unsigned long addr
, unsigned long end
,
708 int write
, struct page
**pages
, int *nr
)
711 unsigned long sz
= 1UL << hugepd_shift(*hugepd
);
714 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
716 next
= hugepte_addr_end(addr
, end
, sz
);
717 if (!gup_hugepte(ptep
, sz
, addr
, end
, write
, pages
, nr
))
719 } while (ptep
++, addr
= next
, addr
!= end
);
724 #ifdef CONFIG_PPC_MM_SLICES
725 unsigned long hugetlb_get_unmapped_area(struct file
*file
, unsigned long addr
,
726 unsigned long len
, unsigned long pgoff
,
729 struct hstate
*hstate
= hstate_file(file
);
730 int mmu_psize
= shift_to_mmu_psize(huge_page_shift(hstate
));
732 return slice_get_unmapped_area(addr
, len
, flags
, mmu_psize
, 1);
736 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
738 #ifdef CONFIG_PPC_MM_SLICES
739 unsigned int psize
= get_slice_psize(vma
->vm_mm
, vma
->vm_start
);
741 return 1UL << mmu_psize_to_shift(psize
);
743 if (!is_vm_hugetlb_page(vma
))
746 return huge_page_size(hstate_vma(vma
));
750 static inline bool is_power_of_4(unsigned long x
)
752 if (is_power_of_2(x
))
753 return (__ilog2(x
) % 2) ? false : true;
757 static int __init
add_huge_page_size(unsigned long long size
)
759 int shift
= __ffs(size
);
762 /* Check that it is a page size supported by the hardware and
763 * that it fits within pagetable and slice limits. */
764 #ifdef CONFIG_PPC_FSL_BOOK3E
765 if ((size
< PAGE_SIZE
) || !is_power_of_4(size
))
768 if (!is_power_of_2(size
)
769 || (shift
> SLICE_HIGH_SHIFT
) || (shift
<= PAGE_SHIFT
))
773 if ((mmu_psize
= shift_to_mmu_psize(shift
)) < 0)
776 #ifdef CONFIG_SPU_FS_64K_LS
777 /* Disable support for 64K huge pages when 64K SPU local store
778 * support is enabled as the current implementation conflicts.
780 if (shift
== PAGE_SHIFT_64K
)
782 #endif /* CONFIG_SPU_FS_64K_LS */
784 BUG_ON(mmu_psize_defs
[mmu_psize
].shift
!= shift
);
786 /* Return if huge page size has already been setup */
787 if (size_to_hstate(size
))
790 hugetlb_add_hstate(shift
- PAGE_SHIFT
);
795 static int __init
hugepage_setup_sz(char *str
)
797 unsigned long long size
;
799 size
= memparse(str
, &str
);
801 if (add_huge_page_size(size
) != 0)
802 printk(KERN_WARNING
"Invalid huge page size specified(%llu)\n", size
);
806 __setup("hugepagesz=", hugepage_setup_sz
);
808 #ifdef CONFIG_PPC_FSL_BOOK3E
809 struct kmem_cache
*hugepte_cache
;
810 static int __init
hugetlbpage_init(void)
814 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
817 if (!mmu_psize_defs
[psize
].shift
)
820 shift
= mmu_psize_to_shift(psize
);
822 /* Don't treat normal page sizes as huge... */
823 if (shift
!= PAGE_SHIFT
)
824 if (add_huge_page_size(1ULL << shift
) < 0)
829 * Create a kmem cache for hugeptes. The bottom bits in the pte have
830 * size information encoded in them, so align them to allow this
832 hugepte_cache
= kmem_cache_create("hugepte-cache", sizeof(pte_t
),
833 HUGEPD_SHIFT_MASK
+ 1, 0, NULL
);
834 if (hugepte_cache
== NULL
)
835 panic("%s: Unable to create kmem cache for hugeptes\n",
838 /* Default hpage size = 4M */
839 if (mmu_psize_defs
[MMU_PAGE_4M
].shift
)
840 HPAGE_SHIFT
= mmu_psize_defs
[MMU_PAGE_4M
].shift
;
842 panic("%s: Unable to set default huge page size\n", __func__
);
848 static int __init
hugetlbpage_init(void)
852 if (!mmu_has_feature(MMU_FTR_16M_PAGE
))
855 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
859 if (!mmu_psize_defs
[psize
].shift
)
862 shift
= mmu_psize_to_shift(psize
);
864 if (add_huge_page_size(1ULL << shift
) < 0)
867 if (shift
< PMD_SHIFT
)
869 else if (shift
< PUD_SHIFT
)
872 pdshift
= PGDIR_SHIFT
;
874 * if we have pdshift and shift value same, we don't
875 * use pgt cache for hugepd.
877 if (pdshift
!= shift
) {
878 pgtable_cache_add(pdshift
- shift
, NULL
);
879 if (!PGT_CACHE(pdshift
- shift
))
880 panic("hugetlbpage_init(): could not create "
881 "pgtable cache for %d bit pagesize\n", shift
);
885 /* Set default large page size. Currently, we pick 16M or 1M
886 * depending on what is available
888 if (mmu_psize_defs
[MMU_PAGE_16M
].shift
)
889 HPAGE_SHIFT
= mmu_psize_defs
[MMU_PAGE_16M
].shift
;
890 else if (mmu_psize_defs
[MMU_PAGE_1M
].shift
)
891 HPAGE_SHIFT
= mmu_psize_defs
[MMU_PAGE_1M
].shift
;
896 module_init(hugetlbpage_init
);
898 void flush_dcache_icache_hugepage(struct page
*page
)
903 BUG_ON(!PageCompound(page
));
905 for (i
= 0; i
< (1UL << compound_order(page
)); i
++) {
906 if (!PageHighMem(page
)) {
907 __flush_dcache_icache(page_address(page
+i
));
909 start
= kmap_atomic(page
+i
);
910 __flush_dcache_icache(start
);
911 kunmap_atomic(start
);
916 #endif /* CONFIG_HUGETLB_PAGE */
919 * We have 4 cases for pgds and pmds:
920 * (1) invalid (all zeroes)
921 * (2) pointer to next table, as normal; bottom 6 bits == 0
922 * (3) leaf pte for huge page, bottom two bits != 00
923 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
925 pte_t
*find_linux_pte_or_hugepte(pgd_t
*pgdir
, unsigned long ea
, unsigned *shift
)
931 hugepd_t
*hpdp
= NULL
;
932 unsigned pdshift
= PGDIR_SHIFT
;
937 pg
= pgdir
+ pgd_index(ea
);
940 ret_pte
= (pte_t
*) pg
;
942 } else if (is_hugepd(pg
))
943 hpdp
= (hugepd_t
*)pg
;
944 else if (!pgd_none(*pg
)) {
946 pu
= pud_offset(pg
, ea
);
949 ret_pte
= (pte_t
*) pu
;
951 } else if (is_hugepd(pu
))
952 hpdp
= (hugepd_t
*)pu
;
953 else if (!pud_none(*pu
)) {
955 pm
= pmd_offset(pu
, ea
);
958 ret_pte
= (pte_t
*) pm
;
960 } else if (is_hugepd(pm
))
961 hpdp
= (hugepd_t
*)pm
;
962 else if (!pmd_none(*pm
))
963 return pte_offset_kernel(pm
, ea
);
969 ret_pte
= hugepte_offset(hpdp
, ea
, pdshift
);
970 pdshift
= hugepd_shift(*hpdp
);
976 EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte
);
978 int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
979 unsigned long end
, int write
, struct page
**pages
, int *nr
)
982 unsigned long pte_end
;
983 struct page
*head
, *page
, *tail
;
987 pte_end
= (addr
+ sz
) & ~(sz
-1);
992 mask
= _PAGE_PRESENT
| _PAGE_USER
;
996 if ((pte_val(pte
) & mask
) != mask
)
999 /* hugepages are never "special" */
1000 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1003 head
= pte_page(pte
);
1005 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
1008 VM_BUG_ON(compound_head(page
) != head
);
1013 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1015 if (!page_cache_add_speculative(head
, refs
)) {
1020 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1021 /* Could be optimized better */
1029 * Any tail page need their mapcount reference taken before we
1034 get_huge_page_tail(tail
);