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 * Defined in such a way that we can optimize away code block at build time
67 * if CONFIG_HUGETLB_PAGE=n.
69 int pmd_huge(pmd_t pmd
)
72 * leaf pte for huge page, bottom two bits != 00
74 return ((pmd_val(pmd
) & 0x3) != 0x0);
77 int pud_huge(pud_t pud
)
80 * leaf pte for huge page, bottom two bits != 00
82 return ((pud_val(pud
) & 0x3) != 0x0);
85 int pgd_huge(pgd_t pgd
)
88 * leaf pte for huge page, bottom two bits != 00
90 return ((pgd_val(pgd
) & 0x3) != 0x0);
93 int pmd_huge(pmd_t pmd
)
98 int pud_huge(pud_t pud
)
103 int pgd_huge(pgd_t pgd
)
109 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
111 /* Only called for hugetlbfs pages, hence can ignore THP */
112 return __find_linux_pte_or_hugepte(mm
->pgd
, addr
, NULL
);
115 static int __hugepte_alloc(struct mm_struct
*mm
, hugepd_t
*hpdp
,
116 unsigned long address
, unsigned pdshift
, unsigned pshift
)
118 struct kmem_cache
*cachep
;
121 #ifdef CONFIG_PPC_FSL_BOOK3E
123 int num_hugepd
= 1 << (pshift
- pdshift
);
124 cachep
= hugepte_cache
;
126 cachep
= PGT_CACHE(pdshift
- pshift
);
129 new = kmem_cache_zalloc(cachep
, GFP_KERNEL
|__GFP_REPEAT
);
131 BUG_ON(pshift
> HUGEPD_SHIFT_MASK
);
132 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK
);
137 spin_lock(&mm
->page_table_lock
);
138 #ifdef CONFIG_PPC_FSL_BOOK3E
140 * We have multiple higher-level entries that point to the same
141 * actual pte location. Fill in each as we go and backtrack on error.
142 * We need all of these so the DTLB pgtable walk code can find the
143 * right higher-level entry without knowing if it's a hugepage or not.
145 for (i
= 0; i
< num_hugepd
; i
++, hpdp
++) {
146 if (unlikely(!hugepd_none(*hpdp
)))
149 /* We use the old format for PPC_FSL_BOOK3E */
150 hpdp
->pd
= ((unsigned long)new & ~PD_HUGE
) | pshift
;
152 /* If we bailed from the for loop early, an error occurred, clean up */
153 if (i
< num_hugepd
) {
154 for (i
= i
- 1 ; i
>= 0; i
--, hpdp
--)
156 kmem_cache_free(cachep
, new);
159 if (!hugepd_none(*hpdp
))
160 kmem_cache_free(cachep
, new);
162 #ifdef CONFIG_PPC_BOOK3S_64
163 hpdp
->pd
= (unsigned long)new |
164 (shift_to_mmu_psize(pshift
) << 2);
166 hpdp
->pd
= ((unsigned long)new & ~PD_HUGE
) | pshift
;
170 spin_unlock(&mm
->page_table_lock
);
175 * These macros define how to determine which level of the page table holds
178 #ifdef CONFIG_PPC_FSL_BOOK3E
179 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
180 #define HUGEPD_PUD_SHIFT PUD_SHIFT
182 #define HUGEPD_PGD_SHIFT PUD_SHIFT
183 #define HUGEPD_PUD_SHIFT PMD_SHIFT
186 #ifdef CONFIG_PPC_BOOK3S_64
188 * At this point we do the placement change only for BOOK3S 64. This would
189 * possibly work on other subarchs.
191 pte_t
*huge_pte_alloc(struct mm_struct
*mm
, unsigned long addr
, unsigned long sz
)
196 hugepd_t
*hpdp
= NULL
;
197 unsigned pshift
= __ffs(sz
);
198 unsigned pdshift
= PGDIR_SHIFT
;
201 pg
= pgd_offset(mm
, addr
);
203 if (pshift
== PGDIR_SHIFT
)
206 else if (pshift
> PUD_SHIFT
)
208 * We need to use hugepd table
210 hpdp
= (hugepd_t
*)pg
;
213 pu
= pud_alloc(mm
, pg
, addr
);
214 if (pshift
== PUD_SHIFT
)
216 else if (pshift
> PMD_SHIFT
)
217 hpdp
= (hugepd_t
*)pu
;
220 pm
= pmd_alloc(mm
, pu
, addr
);
221 if (pshift
== PMD_SHIFT
)
225 hpdp
= (hugepd_t
*)pm
;
231 BUG_ON(!hugepd_none(*hpdp
) && !hugepd_ok(*hpdp
));
233 if (hugepd_none(*hpdp
) && __hugepte_alloc(mm
, hpdp
, addr
, pdshift
, pshift
))
236 return hugepte_offset(*hpdp
, addr
, pdshift
);
241 pte_t
*huge_pte_alloc(struct mm_struct
*mm
, unsigned long addr
, unsigned long sz
)
246 hugepd_t
*hpdp
= NULL
;
247 unsigned pshift
= __ffs(sz
);
248 unsigned pdshift
= PGDIR_SHIFT
;
252 pg
= pgd_offset(mm
, addr
);
254 if (pshift
>= HUGEPD_PGD_SHIFT
) {
255 hpdp
= (hugepd_t
*)pg
;
258 pu
= pud_alloc(mm
, pg
, addr
);
259 if (pshift
>= HUGEPD_PUD_SHIFT
) {
260 hpdp
= (hugepd_t
*)pu
;
263 pm
= pmd_alloc(mm
, pu
, addr
);
264 hpdp
= (hugepd_t
*)pm
;
271 BUG_ON(!hugepd_none(*hpdp
) && !hugepd_ok(*hpdp
));
273 if (hugepd_none(*hpdp
) && __hugepte_alloc(mm
, hpdp
, addr
, pdshift
, pshift
))
276 return hugepte_offset(*hpdp
, addr
, pdshift
);
280 #ifdef CONFIG_PPC_FSL_BOOK3E
281 /* Build list of addresses of gigantic pages. This function is used in early
282 * boot before the buddy allocator is setup.
284 void add_gpage(u64 addr
, u64 page_size
, unsigned long number_of_pages
)
286 unsigned int idx
= shift_to_mmu_psize(__ffs(page_size
));
292 gpage_freearray
[idx
].nr_gpages
= number_of_pages
;
294 for (i
= 0; i
< number_of_pages
; i
++) {
295 gpage_freearray
[idx
].gpage_list
[i
] = addr
;
301 * Moves the gigantic page addresses from the temporary list to the
302 * huge_boot_pages list.
304 int alloc_bootmem_huge_page(struct hstate
*hstate
)
306 struct huge_bootmem_page
*m
;
307 int idx
= shift_to_mmu_psize(huge_page_shift(hstate
));
308 int nr_gpages
= gpage_freearray
[idx
].nr_gpages
;
313 #ifdef CONFIG_HIGHMEM
315 * If gpages can be in highmem we can't use the trick of storing the
316 * data structure in the page; allocate space for this
318 m
= memblock_virt_alloc(sizeof(struct huge_bootmem_page
), 0);
319 m
->phys
= gpage_freearray
[idx
].gpage_list
[--nr_gpages
];
321 m
= phys_to_virt(gpage_freearray
[idx
].gpage_list
[--nr_gpages
]);
324 list_add(&m
->list
, &huge_boot_pages
);
325 gpage_freearray
[idx
].nr_gpages
= nr_gpages
;
326 gpage_freearray
[idx
].gpage_list
[nr_gpages
] = 0;
332 * Scan the command line hugepagesz= options for gigantic pages; store those in
333 * a list that we use to allocate the memory once all options are parsed.
336 unsigned long gpage_npages
[MMU_PAGE_COUNT
];
338 static int __init
do_gpage_early_setup(char *param
, char *val
,
339 const char *unused
, void *arg
)
341 static phys_addr_t size
;
342 unsigned long npages
;
345 * The hugepagesz and hugepages cmdline options are interleaved. We
346 * use the size variable to keep track of whether or not this was done
347 * properly and skip over instances where it is incorrect. Other
348 * command-line parsing code will issue warnings, so we don't need to.
351 if ((strcmp(param
, "default_hugepagesz") == 0) ||
352 (strcmp(param
, "hugepagesz") == 0)) {
353 size
= memparse(val
, NULL
);
354 } else if (strcmp(param
, "hugepages") == 0) {
356 if (sscanf(val
, "%lu", &npages
) <= 0)
358 if (npages
> MAX_NUMBER_GPAGES
) {
359 pr_warn("MMU: %lu pages requested for page "
360 "size %llu KB, limiting to "
361 __stringify(MAX_NUMBER_GPAGES
) "\n",
362 npages
, size
/ 1024);
363 npages
= MAX_NUMBER_GPAGES
;
365 gpage_npages
[shift_to_mmu_psize(__ffs(size
))] = npages
;
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.
380 void __init
reserve_hugetlb_gpages(void)
382 static __initdata
char cmdline
[COMMAND_LINE_SIZE
];
383 phys_addr_t size
, base
;
386 strlcpy(cmdline
, boot_command_line
, COMMAND_LINE_SIZE
);
387 parse_args("hugetlb gpages", cmdline
, NULL
, 0, 0, 0,
388 NULL
, &do_gpage_early_setup
);
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.
396 for (i
= MMU_PAGE_COUNT
-1; i
>= 0; i
--) {
397 if (mmu_psize_defs
[i
].shift
== 0 || gpage_npages
[i
] == 0)
399 else if (mmu_psize_to_shift(i
) < (MAX_ORDER
+ PAGE_SHIFT
))
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
]);
409 #else /* !PPC_FSL_BOOK3E */
411 /* Build list of addresses of gigantic pages. This function is used in early
412 * boot before the buddy allocator is setup.
414 void add_gpage(u64 addr
, u64 page_size
, unsigned long number_of_pages
)
418 while (number_of_pages
> 0) {
419 gpage_freearray
[nr_gpages
] = addr
;
426 /* Moves the gigantic page addresses from the temporary list to the
427 * huge_boot_pages list.
429 int alloc_bootmem_huge_page(struct hstate
*hstate
)
431 struct huge_bootmem_page
*m
;
434 m
= phys_to_virt(gpage_freearray
[--nr_gpages
]);
435 gpage_freearray
[nr_gpages
] = 0;
436 list_add(&m
->list
, &huge_boot_pages
);
442 #ifdef CONFIG_PPC_FSL_BOOK3E
443 #define HUGEPD_FREELIST_SIZE \
444 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
446 struct hugepd_freelist
{
452 static DEFINE_PER_CPU(struct hugepd_freelist
*, hugepd_freelist_cur
);
454 static void hugepd_free_rcu_callback(struct rcu_head
*head
)
456 struct hugepd_freelist
*batch
=
457 container_of(head
, struct hugepd_freelist
, rcu
);
460 for (i
= 0; i
< batch
->index
; i
++)
461 kmem_cache_free(hugepte_cache
, batch
->ptes
[i
]);
463 free_page((unsigned long)batch
);
466 static void hugepd_free(struct mmu_gather
*tlb
, void *hugepte
)
468 struct hugepd_freelist
**batchp
;
470 batchp
= this_cpu_ptr(&hugepd_freelist_cur
);
472 if (atomic_read(&tlb
->mm
->mm_users
) < 2 ||
473 cpumask_equal(mm_cpumask(tlb
->mm
),
474 cpumask_of(smp_processor_id()))) {
475 kmem_cache_free(hugepte_cache
, hugepte
);
476 put_cpu_var(hugepd_freelist_cur
);
480 if (*batchp
== NULL
) {
481 *batchp
= (struct hugepd_freelist
*)__get_free_page(GFP_ATOMIC
);
482 (*batchp
)->index
= 0;
485 (*batchp
)->ptes
[(*batchp
)->index
++] = hugepte
;
486 if ((*batchp
)->index
== HUGEPD_FREELIST_SIZE
) {
487 call_rcu_sched(&(*batchp
)->rcu
, hugepd_free_rcu_callback
);
490 put_cpu_var(hugepd_freelist_cur
);
494 static void free_hugepd_range(struct mmu_gather
*tlb
, hugepd_t
*hpdp
, int pdshift
,
495 unsigned long start
, unsigned long end
,
496 unsigned long floor
, unsigned long ceiling
)
498 pte_t
*hugepte
= hugepd_page(*hpdp
);
501 unsigned long pdmask
= ~((1UL << pdshift
) - 1);
502 unsigned int num_hugepd
= 1;
504 #ifdef CONFIG_PPC_FSL_BOOK3E
505 /* Note: On fsl the hpdp may be the first of several */
506 num_hugepd
= (1 << (hugepd_shift(*hpdp
) - pdshift
));
508 unsigned int shift
= hugepd_shift(*hpdp
);
519 if (end
- 1 > ceiling
- 1)
522 for (i
= 0; i
< num_hugepd
; i
++, hpdp
++)
525 #ifdef CONFIG_PPC_FSL_BOOK3E
526 hugepd_free(tlb
, hugepte
);
528 pgtable_free_tlb(tlb
, hugepte
, pdshift
- shift
);
532 static void hugetlb_free_pmd_range(struct mmu_gather
*tlb
, pud_t
*pud
,
533 unsigned long addr
, unsigned long end
,
534 unsigned long floor
, unsigned long ceiling
)
542 pmd
= pmd_offset(pud
, addr
);
543 next
= pmd_addr_end(addr
, end
);
544 if (!is_hugepd(__hugepd(pmd_val(*pmd
)))) {
546 * if it is not hugepd pointer, we should already find
549 WARN_ON(!pmd_none_or_clear_bad(pmd
));
552 #ifdef CONFIG_PPC_FSL_BOOK3E
554 * Increment next by the size of the huge mapping since
555 * there may be more than one entry at this level for a
556 * single hugepage, but all of them point to
557 * the same kmem cache that holds the hugepte.
559 next
= addr
+ (1 << hugepd_shift(*(hugepd_t
*)pmd
));
561 free_hugepd_range(tlb
, (hugepd_t
*)pmd
, PMD_SHIFT
,
562 addr
, next
, floor
, ceiling
);
563 } while (addr
= next
, addr
!= end
);
573 if (end
- 1 > ceiling
- 1)
576 pmd
= pmd_offset(pud
, start
);
578 pmd_free_tlb(tlb
, pmd
, start
);
579 mm_dec_nr_pmds(tlb
->mm
);
582 static void hugetlb_free_pud_range(struct mmu_gather
*tlb
, pgd_t
*pgd
,
583 unsigned long addr
, unsigned long end
,
584 unsigned long floor
, unsigned long ceiling
)
592 pud
= pud_offset(pgd
, addr
);
593 next
= pud_addr_end(addr
, end
);
594 if (!is_hugepd(__hugepd(pud_val(*pud
)))) {
595 if (pud_none_or_clear_bad(pud
))
597 hugetlb_free_pmd_range(tlb
, pud
, addr
, next
, floor
,
600 #ifdef CONFIG_PPC_FSL_BOOK3E
602 * Increment next by the size of the huge mapping since
603 * there may be more than one entry at this level for a
604 * single hugepage, but all of them point to
605 * the same kmem cache that holds the hugepte.
607 next
= addr
+ (1 << hugepd_shift(*(hugepd_t
*)pud
));
609 free_hugepd_range(tlb
, (hugepd_t
*)pud
, PUD_SHIFT
,
610 addr
, next
, floor
, ceiling
);
612 } while (addr
= next
, addr
!= end
);
618 ceiling
&= PGDIR_MASK
;
622 if (end
- 1 > ceiling
- 1)
625 pud
= pud_offset(pgd
, start
);
627 pud_free_tlb(tlb
, pud
, start
);
631 * This function frees user-level page tables of a process.
633 void hugetlb_free_pgd_range(struct mmu_gather
*tlb
,
634 unsigned long addr
, unsigned long end
,
635 unsigned long floor
, unsigned long ceiling
)
641 * Because there are a number of different possible pagetable
642 * layouts for hugepage ranges, we limit knowledge of how
643 * things should be laid out to the allocation path
644 * (huge_pte_alloc(), above). Everything else works out the
645 * structure as it goes from information in the hugepd
646 * pointers. That means that we can't here use the
647 * optimization used in the normal page free_pgd_range(), of
648 * checking whether we're actually covering a large enough
649 * range to have to do anything at the top level of the walk
650 * instead of at the bottom.
652 * To make sense of this, you should probably go read the big
653 * block comment at the top of the normal free_pgd_range(),
658 next
= pgd_addr_end(addr
, end
);
659 pgd
= pgd_offset(tlb
->mm
, addr
);
660 if (!is_hugepd(__hugepd(pgd_val(*pgd
)))) {
661 if (pgd_none_or_clear_bad(pgd
))
663 hugetlb_free_pud_range(tlb
, pgd
, addr
, next
, floor
, ceiling
);
665 #ifdef CONFIG_PPC_FSL_BOOK3E
667 * Increment next by the size of the huge mapping since
668 * there may be more than one entry at the pgd level
669 * for a single hugepage, but all of them point to the
670 * same kmem cache that holds the hugepte.
672 next
= addr
+ (1 << hugepd_shift(*(hugepd_t
*)pgd
));
674 free_hugepd_range(tlb
, (hugepd_t
*)pgd
, PGDIR_SHIFT
,
675 addr
, next
, floor
, ceiling
);
677 } while (addr
= next
, addr
!= end
);
681 * We are holding mmap_sem, so a parallel huge page collapse cannot run.
682 * To prevent hugepage split, disable irq.
685 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
, int write
)
689 unsigned long mask
, flags
;
690 struct page
*page
= ERR_PTR(-EINVAL
);
692 local_irq_save(flags
);
693 ptep
= find_linux_pte_or_hugepte(mm
->pgd
, address
, &shift
);
696 pte
= READ_ONCE(*ptep
);
698 * Verify it is a huge page else bail.
699 * Transparent hugepages are handled by generic code. We can skip them
702 if (!shift
|| pmd_trans_huge(__pmd(pte_val(pte
))))
705 if (!pte_present(pte
)) {
709 mask
= (1UL << shift
) - 1;
710 page
= pte_page(pte
);
712 page
+= (address
& mask
) / PAGE_SIZE
;
715 local_irq_restore(flags
);
720 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
721 pmd_t
*pmd
, int write
)
728 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
729 pud_t
*pud
, int write
)
735 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
738 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
739 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
742 int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
, unsigned pdshift
,
743 unsigned long end
, int write
, struct page
**pages
, int *nr
)
746 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
749 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
751 next
= hugepte_addr_end(addr
, end
, sz
);
752 if (!gup_hugepte(ptep
, sz
, addr
, end
, write
, pages
, nr
))
754 } while (ptep
++, addr
= next
, addr
!= end
);
759 #ifdef CONFIG_PPC_MM_SLICES
760 unsigned long hugetlb_get_unmapped_area(struct file
*file
, unsigned long addr
,
761 unsigned long len
, unsigned long pgoff
,
764 struct hstate
*hstate
= hstate_file(file
);
765 int mmu_psize
= shift_to_mmu_psize(huge_page_shift(hstate
));
767 return slice_get_unmapped_area(addr
, len
, flags
, mmu_psize
, 1);
771 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
773 #ifdef CONFIG_PPC_MM_SLICES
774 unsigned int psize
= get_slice_psize(vma
->vm_mm
, vma
->vm_start
);
776 return 1UL << mmu_psize_to_shift(psize
);
778 if (!is_vm_hugetlb_page(vma
))
781 return huge_page_size(hstate_vma(vma
));
785 static inline bool is_power_of_4(unsigned long x
)
787 if (is_power_of_2(x
))
788 return (__ilog2(x
) % 2) ? false : true;
792 static int __init
add_huge_page_size(unsigned long long size
)
794 int shift
= __ffs(size
);
797 /* Check that it is a page size supported by the hardware and
798 * that it fits within pagetable and slice limits. */
799 #ifdef CONFIG_PPC_FSL_BOOK3E
800 if ((size
< PAGE_SIZE
) || !is_power_of_4(size
))
803 if (!is_power_of_2(size
)
804 || (shift
> SLICE_HIGH_SHIFT
) || (shift
<= PAGE_SHIFT
))
808 if ((mmu_psize
= shift_to_mmu_psize(shift
)) < 0)
811 BUG_ON(mmu_psize_defs
[mmu_psize
].shift
!= shift
);
813 /* Return if huge page size has already been setup */
814 if (size_to_hstate(size
))
817 hugetlb_add_hstate(shift
- PAGE_SHIFT
);
822 static int __init
hugepage_setup_sz(char *str
)
824 unsigned long long size
;
826 size
= memparse(str
, &str
);
828 if (add_huge_page_size(size
) != 0)
829 printk(KERN_WARNING
"Invalid huge page size specified(%llu)\n", size
);
833 __setup("hugepagesz=", hugepage_setup_sz
);
835 #ifdef CONFIG_PPC_FSL_BOOK3E
836 struct kmem_cache
*hugepte_cache
;
837 static int __init
hugetlbpage_init(void)
841 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
844 if (!mmu_psize_defs
[psize
].shift
)
847 shift
= mmu_psize_to_shift(psize
);
849 /* Don't treat normal page sizes as huge... */
850 if (shift
!= PAGE_SHIFT
)
851 if (add_huge_page_size(1ULL << shift
) < 0)
856 * Create a kmem cache for hugeptes. The bottom bits in the pte have
857 * size information encoded in them, so align them to allow this
859 hugepte_cache
= kmem_cache_create("hugepte-cache", sizeof(pte_t
),
860 HUGEPD_SHIFT_MASK
+ 1, 0, NULL
);
861 if (hugepte_cache
== NULL
)
862 panic("%s: Unable to create kmem cache for hugeptes\n",
865 /* Default hpage size = 4M */
866 if (mmu_psize_defs
[MMU_PAGE_4M
].shift
)
867 HPAGE_SHIFT
= mmu_psize_defs
[MMU_PAGE_4M
].shift
;
869 panic("%s: Unable to set default huge page size\n", __func__
);
875 static int __init
hugetlbpage_init(void)
879 if (!mmu_has_feature(MMU_FTR_16M_PAGE
))
882 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
886 if (!mmu_psize_defs
[psize
].shift
)
889 shift
= mmu_psize_to_shift(psize
);
891 if (add_huge_page_size(1ULL << shift
) < 0)
894 if (shift
< PMD_SHIFT
)
896 else if (shift
< PUD_SHIFT
)
899 pdshift
= PGDIR_SHIFT
;
901 * if we have pdshift and shift value same, we don't
902 * use pgt cache for hugepd.
904 if (pdshift
!= shift
) {
905 pgtable_cache_add(pdshift
- shift
, NULL
);
906 if (!PGT_CACHE(pdshift
- shift
))
907 panic("hugetlbpage_init(): could not create "
908 "pgtable cache for %d bit pagesize\n", shift
);
912 /* Set default large page size. Currently, we pick 16M or 1M
913 * depending on what is available
915 if (mmu_psize_defs
[MMU_PAGE_16M
].shift
)
916 HPAGE_SHIFT
= mmu_psize_defs
[MMU_PAGE_16M
].shift
;
917 else if (mmu_psize_defs
[MMU_PAGE_1M
].shift
)
918 HPAGE_SHIFT
= mmu_psize_defs
[MMU_PAGE_1M
].shift
;
923 arch_initcall(hugetlbpage_init
);
925 void flush_dcache_icache_hugepage(struct page
*page
)
930 BUG_ON(!PageCompound(page
));
932 for (i
= 0; i
< (1UL << compound_order(page
)); i
++) {
933 if (!PageHighMem(page
)) {
934 __flush_dcache_icache(page_address(page
+i
));
936 start
= kmap_atomic(page
+i
);
937 __flush_dcache_icache(start
);
938 kunmap_atomic(start
);
943 #endif /* CONFIG_HUGETLB_PAGE */
946 * We have 4 cases for pgds and pmds:
947 * (1) invalid (all zeroes)
948 * (2) pointer to next table, as normal; bottom 6 bits == 0
949 * (3) leaf pte for huge page, bottom two bits != 00
950 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
952 * So long as we atomically load page table pointers we are safe against teardown,
953 * we can follow the address down to the the page and take a ref on it.
954 * This function need to be called with interrupts disabled. We use this variant
955 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
958 pte_t
*__find_linux_pte_or_hugepte(pgd_t
*pgdir
, unsigned long ea
,
965 hugepd_t
*hpdp
= NULL
;
966 unsigned pdshift
= PGDIR_SHIFT
;
971 pgdp
= pgdir
+ pgd_index(ea
);
972 pgd
= READ_ONCE(*pgdp
);
974 * Always operate on the local stack value. This make sure the
975 * value don't get updated by a parallel THP split/collapse,
976 * page fault or a page unmap. The return pte_t * is still not
977 * stable. So should be checked there for above conditions.
981 else if (pgd_huge(pgd
)) {
982 ret_pte
= (pte_t
*) pgdp
;
984 } else if (is_hugepd(__hugepd(pgd_val(pgd
))))
985 hpdp
= (hugepd_t
*)&pgd
;
988 * Even if we end up with an unmap, the pgtable will not
989 * be freed, because we do an rcu free and here we are
993 pudp
= pud_offset(&pgd
, ea
);
994 pud
= READ_ONCE(*pudp
);
998 else if (pud_huge(pud
)) {
999 ret_pte
= (pte_t
*) pudp
;
1001 } else if (is_hugepd(__hugepd(pud_val(pud
))))
1002 hpdp
= (hugepd_t
*)&pud
;
1004 pdshift
= PMD_SHIFT
;
1005 pmdp
= pmd_offset(&pud
, ea
);
1006 pmd
= READ_ONCE(*pmdp
);
1008 * A hugepage collapse is captured by pmd_none, because
1009 * it mark the pmd none and do a hpte invalidate.
1011 * We don't worry about pmd_trans_splitting here, The
1012 * caller if it needs to handle the splitting case
1013 * should check for that.
1018 if (pmd_huge(pmd
) || pmd_large(pmd
)) {
1019 ret_pte
= (pte_t
*) pmdp
;
1021 } else if (is_hugepd(__hugepd(pmd_val(pmd
))))
1022 hpdp
= (hugepd_t
*)&pmd
;
1024 return pte_offset_kernel(&pmd
, ea
);
1030 ret_pte
= hugepte_offset(*hpdp
, ea
, pdshift
);
1031 pdshift
= hugepd_shift(*hpdp
);
1037 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte
);
1039 int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
1040 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1043 unsigned long pte_end
;
1044 struct page
*head
, *page
, *tail
;
1048 pte_end
= (addr
+ sz
) & ~(sz
-1);
1052 pte
= READ_ONCE(*ptep
);
1053 mask
= _PAGE_PRESENT
| _PAGE_USER
;
1057 if ((pte_val(pte
) & mask
) != mask
)
1060 /* hugepages are never "special" */
1061 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1064 head
= pte_page(pte
);
1066 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
1069 VM_BUG_ON(compound_head(page
) != head
);
1074 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1076 if (!page_cache_add_speculative(head
, refs
)) {
1081 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1082 /* Could be optimized better */
1090 * Any tail page need their mapcount reference taken before we
1095 get_huge_page_tail(tail
);