x86/efi: Enforce CONFIG_RELOCATABLE for EFI boot stub
[linux/fpc-iii.git] / arch / powerpc / mm / hugetlbpage.c
blob834ca8eb38f202e01c5151fdb56b13197ed6acc8
1 /*
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>
9 */
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>
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
39 * arrays
41 #ifdef CONFIG_PPC_FSL_BOOK3E
42 #define MAX_NUMBER_GPAGES 128
43 struct psize_gpages {
44 u64 gpage_list[MAX_NUMBER_GPAGES];
45 unsigned int nr_gpages;
47 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
48 #else
49 #define MAX_NUMBER_GPAGES 1024
50 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
51 static unsigned nr_gpages;
52 #endif
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);
89 #else
90 int pmd_huge(pmd_t pmd)
92 return 0;
95 int pud_huge(pud_t pud)
97 return 0;
100 int pgd_huge(pgd_t pgd)
102 return 0;
104 #endif
106 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
108 /* Only called for hugetlbfs pages, hence can ignore THP */
109 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
112 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
113 unsigned long address, unsigned pdshift, unsigned pshift)
115 struct kmem_cache *cachep;
116 pte_t *new;
118 #ifdef CONFIG_PPC_FSL_BOOK3E
119 int i;
120 int num_hugepd = 1 << (pshift - pdshift);
121 cachep = hugepte_cache;
122 #else
123 cachep = PGT_CACHE(pdshift - pshift);
124 #endif
126 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
128 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
129 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
131 if (! new)
132 return -ENOMEM;
134 spin_lock(&mm->page_table_lock);
135 #ifdef CONFIG_PPC_FSL_BOOK3E
137 * We have multiple higher-level entries that point to the same
138 * actual pte location. Fill in each as we go and backtrack on error.
139 * We need all of these so the DTLB pgtable walk code can find the
140 * right higher-level entry without knowing if it's a hugepage or not.
142 for (i = 0; i < num_hugepd; i++, hpdp++) {
143 if (unlikely(!hugepd_none(*hpdp)))
144 break;
145 else
146 /* We use the old format for PPC_FSL_BOOK3E */
147 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
149 /* If we bailed from the for loop early, an error occurred, clean up */
150 if (i < num_hugepd) {
151 for (i = i - 1 ; i >= 0; i--, hpdp--)
152 hpdp->pd = 0;
153 kmem_cache_free(cachep, new);
155 #else
156 if (!hugepd_none(*hpdp))
157 kmem_cache_free(cachep, new);
158 else {
159 #ifdef CONFIG_PPC_BOOK3S_64
160 hpdp->pd = (unsigned long)new |
161 (shift_to_mmu_psize(pshift) << 2);
162 #else
163 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
164 #endif
166 #endif
167 spin_unlock(&mm->page_table_lock);
168 return 0;
172 * These macros define how to determine which level of the page table holds
173 * the hpdp.
175 #ifdef CONFIG_PPC_FSL_BOOK3E
176 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
177 #define HUGEPD_PUD_SHIFT PUD_SHIFT
178 #else
179 #define HUGEPD_PGD_SHIFT PUD_SHIFT
180 #define HUGEPD_PUD_SHIFT PMD_SHIFT
181 #endif
183 #ifdef CONFIG_PPC_BOOK3S_64
185 * At this point we do the placement change only for BOOK3S 64. This would
186 * possibly work on other subarchs.
188 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
190 pgd_t *pg;
191 pud_t *pu;
192 pmd_t *pm;
193 hugepd_t *hpdp = NULL;
194 unsigned pshift = __ffs(sz);
195 unsigned pdshift = PGDIR_SHIFT;
197 addr &= ~(sz-1);
198 pg = pgd_offset(mm, addr);
200 if (pshift == PGDIR_SHIFT)
201 /* 16GB huge page */
202 return (pte_t *) pg;
203 else if (pshift > PUD_SHIFT)
205 * We need to use hugepd table
207 hpdp = (hugepd_t *)pg;
208 else {
209 pdshift = PUD_SHIFT;
210 pu = pud_alloc(mm, pg, addr);
211 if (pshift == PUD_SHIFT)
212 return (pte_t *)pu;
213 else if (pshift > PMD_SHIFT)
214 hpdp = (hugepd_t *)pu;
215 else {
216 pdshift = PMD_SHIFT;
217 pm = pmd_alloc(mm, pu, addr);
218 if (pshift == PMD_SHIFT)
219 /* 16MB hugepage */
220 return (pte_t *)pm;
221 else
222 hpdp = (hugepd_t *)pm;
225 if (!hpdp)
226 return NULL;
228 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
230 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
231 return NULL;
233 return hugepte_offset(hpdp, addr, pdshift);
236 #else
238 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
240 pgd_t *pg;
241 pud_t *pu;
242 pmd_t *pm;
243 hugepd_t *hpdp = NULL;
244 unsigned pshift = __ffs(sz);
245 unsigned pdshift = PGDIR_SHIFT;
247 addr &= ~(sz-1);
249 pg = pgd_offset(mm, addr);
251 if (pshift >= HUGEPD_PGD_SHIFT) {
252 hpdp = (hugepd_t *)pg;
253 } else {
254 pdshift = PUD_SHIFT;
255 pu = pud_alloc(mm, pg, addr);
256 if (pshift >= HUGEPD_PUD_SHIFT) {
257 hpdp = (hugepd_t *)pu;
258 } else {
259 pdshift = PMD_SHIFT;
260 pm = pmd_alloc(mm, pu, addr);
261 hpdp = (hugepd_t *)pm;
265 if (!hpdp)
266 return NULL;
268 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
270 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
271 return NULL;
273 return hugepte_offset(hpdp, addr, pdshift);
275 #endif
277 #ifdef CONFIG_PPC_FSL_BOOK3E
278 /* Build list of addresses of gigantic pages. This function is used in early
279 * boot before the buddy or bootmem allocator is setup.
281 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
283 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
284 int i;
286 if (addr == 0)
287 return;
289 gpage_freearray[idx].nr_gpages = number_of_pages;
291 for (i = 0; i < number_of_pages; i++) {
292 gpage_freearray[idx].gpage_list[i] = addr;
293 addr += page_size;
298 * Moves the gigantic page addresses from the temporary list to the
299 * huge_boot_pages list.
301 int alloc_bootmem_huge_page(struct hstate *hstate)
303 struct huge_bootmem_page *m;
304 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
305 int nr_gpages = gpage_freearray[idx].nr_gpages;
307 if (nr_gpages == 0)
308 return 0;
310 #ifdef CONFIG_HIGHMEM
312 * If gpages can be in highmem we can't use the trick of storing the
313 * data structure in the page; allocate space for this
315 m = alloc_bootmem(sizeof(struct huge_bootmem_page));
316 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
317 #else
318 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
319 #endif
321 list_add(&m->list, &huge_boot_pages);
322 gpage_freearray[idx].nr_gpages = nr_gpages;
323 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
324 m->hstate = hstate;
326 return 1;
329 * Scan the command line hugepagesz= options for gigantic pages; store those in
330 * a list that we use to allocate the memory once all options are parsed.
333 unsigned long gpage_npages[MMU_PAGE_COUNT];
335 static int __init do_gpage_early_setup(char *param, char *val,
336 const char *unused)
338 static phys_addr_t size;
339 unsigned long npages;
342 * The hugepagesz and hugepages cmdline options are interleaved. We
343 * use the size variable to keep track of whether or not this was done
344 * properly and skip over instances where it is incorrect. Other
345 * command-line parsing code will issue warnings, so we don't need to.
348 if ((strcmp(param, "default_hugepagesz") == 0) ||
349 (strcmp(param, "hugepagesz") == 0)) {
350 size = memparse(val, NULL);
351 } else if (strcmp(param, "hugepages") == 0) {
352 if (size != 0) {
353 if (sscanf(val, "%lu", &npages) <= 0)
354 npages = 0;
355 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
356 size = 0;
359 return 0;
364 * This function allocates physical space for pages that are larger than the
365 * buddy allocator can handle. We want to allocate these in highmem because
366 * the amount of lowmem is limited. This means that this function MUST be
367 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
368 * allocate to grab highmem.
370 void __init reserve_hugetlb_gpages(void)
372 static __initdata char cmdline[COMMAND_LINE_SIZE];
373 phys_addr_t size, base;
374 int i;
376 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
377 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
378 &do_gpage_early_setup);
381 * Walk gpage list in reverse, allocating larger page sizes first.
382 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
383 * When we reach the point in the list where pages are no longer
384 * considered gpages, we're done.
386 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
387 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
388 continue;
389 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
390 break;
392 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
393 base = memblock_alloc_base(size * gpage_npages[i], size,
394 MEMBLOCK_ALLOC_ANYWHERE);
395 add_gpage(base, size, gpage_npages[i]);
399 #else /* !PPC_FSL_BOOK3E */
401 /* Build list of addresses of gigantic pages. This function is used in early
402 * boot before the buddy or bootmem allocator is setup.
404 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
406 if (!addr)
407 return;
408 while (number_of_pages > 0) {
409 gpage_freearray[nr_gpages] = addr;
410 nr_gpages++;
411 number_of_pages--;
412 addr += page_size;
416 /* Moves the gigantic page addresses from the temporary list to the
417 * huge_boot_pages list.
419 int alloc_bootmem_huge_page(struct hstate *hstate)
421 struct huge_bootmem_page *m;
422 if (nr_gpages == 0)
423 return 0;
424 m = phys_to_virt(gpage_freearray[--nr_gpages]);
425 gpage_freearray[nr_gpages] = 0;
426 list_add(&m->list, &huge_boot_pages);
427 m->hstate = hstate;
428 return 1;
430 #endif
432 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
434 return 0;
437 #ifdef CONFIG_PPC_FSL_BOOK3E
438 #define HUGEPD_FREELIST_SIZE \
439 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
441 struct hugepd_freelist {
442 struct rcu_head rcu;
443 unsigned int index;
444 void *ptes[0];
447 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
449 static void hugepd_free_rcu_callback(struct rcu_head *head)
451 struct hugepd_freelist *batch =
452 container_of(head, struct hugepd_freelist, rcu);
453 unsigned int i;
455 for (i = 0; i < batch->index; i++)
456 kmem_cache_free(hugepte_cache, batch->ptes[i]);
458 free_page((unsigned long)batch);
461 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
463 struct hugepd_freelist **batchp;
465 batchp = &__get_cpu_var(hugepd_freelist_cur);
467 if (atomic_read(&tlb->mm->mm_users) < 2 ||
468 cpumask_equal(mm_cpumask(tlb->mm),
469 cpumask_of(smp_processor_id()))) {
470 kmem_cache_free(hugepte_cache, hugepte);
471 return;
474 if (*batchp == NULL) {
475 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
476 (*batchp)->index = 0;
479 (*batchp)->ptes[(*batchp)->index++] = hugepte;
480 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
481 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
482 *batchp = NULL;
485 #endif
487 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
488 unsigned long start, unsigned long end,
489 unsigned long floor, unsigned long ceiling)
491 pte_t *hugepte = hugepd_page(*hpdp);
492 int i;
494 unsigned long pdmask = ~((1UL << pdshift) - 1);
495 unsigned int num_hugepd = 1;
497 #ifdef CONFIG_PPC_FSL_BOOK3E
498 /* Note: On fsl the hpdp may be the first of several */
499 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
500 #else
501 unsigned int shift = hugepd_shift(*hpdp);
502 #endif
504 start &= pdmask;
505 if (start < floor)
506 return;
507 if (ceiling) {
508 ceiling &= pdmask;
509 if (! ceiling)
510 return;
512 if (end - 1 > ceiling - 1)
513 return;
515 for (i = 0; i < num_hugepd; i++, hpdp++)
516 hpdp->pd = 0;
518 tlb->need_flush = 1;
520 #ifdef CONFIG_PPC_FSL_BOOK3E
521 hugepd_free(tlb, hugepte);
522 #else
523 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
524 #endif
527 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
528 unsigned long addr, unsigned long end,
529 unsigned long floor, unsigned long ceiling)
531 pmd_t *pmd;
532 unsigned long next;
533 unsigned long start;
535 start = addr;
536 do {
537 pmd = pmd_offset(pud, addr);
538 next = pmd_addr_end(addr, end);
539 if (!is_hugepd(pmd)) {
541 * if it is not hugepd pointer, we should already find
542 * it cleared.
544 WARN_ON(!pmd_none_or_clear_bad(pmd));
545 continue;
547 #ifdef CONFIG_PPC_FSL_BOOK3E
549 * Increment next by the size of the huge mapping since
550 * there may be more than one entry at this level for a
551 * single hugepage, but all of them point to
552 * the same kmem cache that holds the hugepte.
554 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
555 #endif
556 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
557 addr, next, floor, ceiling);
558 } while (addr = next, addr != end);
560 start &= PUD_MASK;
561 if (start < floor)
562 return;
563 if (ceiling) {
564 ceiling &= PUD_MASK;
565 if (!ceiling)
566 return;
568 if (end - 1 > ceiling - 1)
569 return;
571 pmd = pmd_offset(pud, start);
572 pud_clear(pud);
573 pmd_free_tlb(tlb, pmd, start);
576 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
577 unsigned long addr, unsigned long end,
578 unsigned long floor, unsigned long ceiling)
580 pud_t *pud;
581 unsigned long next;
582 unsigned long start;
584 start = addr;
585 do {
586 pud = pud_offset(pgd, addr);
587 next = pud_addr_end(addr, end);
588 if (!is_hugepd(pud)) {
589 if (pud_none_or_clear_bad(pud))
590 continue;
591 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
592 ceiling);
593 } else {
594 #ifdef CONFIG_PPC_FSL_BOOK3E
596 * Increment next by the size of the huge mapping since
597 * there may be more than one entry at this level for a
598 * single hugepage, but all of them point to
599 * the same kmem cache that holds the hugepte.
601 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
602 #endif
603 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
604 addr, next, floor, ceiling);
606 } while (addr = next, addr != end);
608 start &= PGDIR_MASK;
609 if (start < floor)
610 return;
611 if (ceiling) {
612 ceiling &= PGDIR_MASK;
613 if (!ceiling)
614 return;
616 if (end - 1 > ceiling - 1)
617 return;
619 pud = pud_offset(pgd, start);
620 pgd_clear(pgd);
621 pud_free_tlb(tlb, pud, start);
625 * This function frees user-level page tables of a process.
627 * Must be called with pagetable lock held.
629 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
630 unsigned long addr, unsigned long end,
631 unsigned long floor, unsigned long ceiling)
633 pgd_t *pgd;
634 unsigned long next;
637 * Because there are a number of different possible pagetable
638 * layouts for hugepage ranges, we limit knowledge of how
639 * things should be laid out to the allocation path
640 * (huge_pte_alloc(), above). Everything else works out the
641 * structure as it goes from information in the hugepd
642 * pointers. That means that we can't here use the
643 * optimization used in the normal page free_pgd_range(), of
644 * checking whether we're actually covering a large enough
645 * range to have to do anything at the top level of the walk
646 * instead of at the bottom.
648 * To make sense of this, you should probably go read the big
649 * block comment at the top of the normal free_pgd_range(),
650 * too.
653 do {
654 next = pgd_addr_end(addr, end);
655 pgd = pgd_offset(tlb->mm, addr);
656 if (!is_hugepd(pgd)) {
657 if (pgd_none_or_clear_bad(pgd))
658 continue;
659 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
660 } else {
661 #ifdef CONFIG_PPC_FSL_BOOK3E
663 * Increment next by the size of the huge mapping since
664 * there may be more than one entry at the pgd level
665 * for a single hugepage, but all of them point to the
666 * same kmem cache that holds the hugepte.
668 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
669 #endif
670 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
671 addr, next, floor, ceiling);
673 } while (addr = next, addr != end);
676 struct page *
677 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
679 pte_t *ptep;
680 struct page *page;
681 unsigned shift;
682 unsigned long mask;
684 * Transparent hugepages are handled by generic code. We can skip them
685 * here.
687 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
689 /* Verify it is a huge page else bail. */
690 if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep))
691 return ERR_PTR(-EINVAL);
693 mask = (1UL << shift) - 1;
694 page = pte_page(*ptep);
695 if (page)
696 page += (address & mask) / PAGE_SIZE;
698 return page;
701 struct page *
702 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
703 pmd_t *pmd, int write)
705 BUG();
706 return NULL;
709 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
710 unsigned long sz)
712 unsigned long __boundary = (addr + sz) & ~(sz-1);
713 return (__boundary - 1 < end - 1) ? __boundary : end;
716 int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
717 unsigned long addr, unsigned long end,
718 int write, struct page **pages, int *nr)
720 pte_t *ptep;
721 unsigned long sz = 1UL << hugepd_shift(*hugepd);
722 unsigned long next;
724 ptep = hugepte_offset(hugepd, addr, pdshift);
725 do {
726 next = hugepte_addr_end(addr, end, sz);
727 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
728 return 0;
729 } while (ptep++, addr = next, addr != end);
731 return 1;
734 #ifdef CONFIG_PPC_MM_SLICES
735 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
736 unsigned long len, unsigned long pgoff,
737 unsigned long flags)
739 struct hstate *hstate = hstate_file(file);
740 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
742 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
744 #endif
746 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
748 #ifdef CONFIG_PPC_MM_SLICES
749 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
751 return 1UL << mmu_psize_to_shift(psize);
752 #else
753 if (!is_vm_hugetlb_page(vma))
754 return PAGE_SIZE;
756 return huge_page_size(hstate_vma(vma));
757 #endif
760 static inline bool is_power_of_4(unsigned long x)
762 if (is_power_of_2(x))
763 return (__ilog2(x) % 2) ? false : true;
764 return false;
767 static int __init add_huge_page_size(unsigned long long size)
769 int shift = __ffs(size);
770 int mmu_psize;
772 /* Check that it is a page size supported by the hardware and
773 * that it fits within pagetable and slice limits. */
774 #ifdef CONFIG_PPC_FSL_BOOK3E
775 if ((size < PAGE_SIZE) || !is_power_of_4(size))
776 return -EINVAL;
777 #else
778 if (!is_power_of_2(size)
779 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
780 return -EINVAL;
781 #endif
783 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
784 return -EINVAL;
786 #ifdef CONFIG_SPU_FS_64K_LS
787 /* Disable support for 64K huge pages when 64K SPU local store
788 * support is enabled as the current implementation conflicts.
790 if (shift == PAGE_SHIFT_64K)
791 return -EINVAL;
792 #endif /* CONFIG_SPU_FS_64K_LS */
794 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
796 /* Return if huge page size has already been setup */
797 if (size_to_hstate(size))
798 return 0;
800 hugetlb_add_hstate(shift - PAGE_SHIFT);
802 return 0;
805 static int __init hugepage_setup_sz(char *str)
807 unsigned long long size;
809 size = memparse(str, &str);
811 if (add_huge_page_size(size) != 0)
812 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
814 return 1;
816 __setup("hugepagesz=", hugepage_setup_sz);
818 #ifdef CONFIG_PPC_FSL_BOOK3E
819 struct kmem_cache *hugepte_cache;
820 static int __init hugetlbpage_init(void)
822 int psize;
824 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
825 unsigned shift;
827 if (!mmu_psize_defs[psize].shift)
828 continue;
830 shift = mmu_psize_to_shift(psize);
832 /* Don't treat normal page sizes as huge... */
833 if (shift != PAGE_SHIFT)
834 if (add_huge_page_size(1ULL << shift) < 0)
835 continue;
839 * Create a kmem cache for hugeptes. The bottom bits in the pte have
840 * size information encoded in them, so align them to allow this
842 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
843 HUGEPD_SHIFT_MASK + 1, 0, NULL);
844 if (hugepte_cache == NULL)
845 panic("%s: Unable to create kmem cache for hugeptes\n",
846 __func__);
848 /* Default hpage size = 4M */
849 if (mmu_psize_defs[MMU_PAGE_4M].shift)
850 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
851 else
852 panic("%s: Unable to set default huge page size\n", __func__);
855 return 0;
857 #else
858 static int __init hugetlbpage_init(void)
860 int psize;
862 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
863 return -ENODEV;
865 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
866 unsigned shift;
867 unsigned pdshift;
869 if (!mmu_psize_defs[psize].shift)
870 continue;
872 shift = mmu_psize_to_shift(psize);
874 if (add_huge_page_size(1ULL << shift) < 0)
875 continue;
877 if (shift < PMD_SHIFT)
878 pdshift = PMD_SHIFT;
879 else if (shift < PUD_SHIFT)
880 pdshift = PUD_SHIFT;
881 else
882 pdshift = PGDIR_SHIFT;
884 * if we have pdshift and shift value same, we don't
885 * use pgt cache for hugepd.
887 if (pdshift != shift) {
888 pgtable_cache_add(pdshift - shift, NULL);
889 if (!PGT_CACHE(pdshift - shift))
890 panic("hugetlbpage_init(): could not create "
891 "pgtable cache for %d bit pagesize\n", shift);
895 /* Set default large page size. Currently, we pick 16M or 1M
896 * depending on what is available
898 if (mmu_psize_defs[MMU_PAGE_16M].shift)
899 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
900 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
901 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
903 return 0;
905 #endif
906 module_init(hugetlbpage_init);
908 void flush_dcache_icache_hugepage(struct page *page)
910 int i;
911 void *start;
913 BUG_ON(!PageCompound(page));
915 for (i = 0; i < (1UL << compound_order(page)); i++) {
916 if (!PageHighMem(page)) {
917 __flush_dcache_icache(page_address(page+i));
918 } else {
919 start = kmap_atomic(page+i);
920 __flush_dcache_icache(start);
921 kunmap_atomic(start);
926 #endif /* CONFIG_HUGETLB_PAGE */
929 * We have 4 cases for pgds and pmds:
930 * (1) invalid (all zeroes)
931 * (2) pointer to next table, as normal; bottom 6 bits == 0
932 * (3) leaf pte for huge page, bottom two bits != 00
933 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
935 * So long as we atomically load page table pointers we are safe against teardown,
936 * we can follow the address down to the the page and take a ref on it.
939 pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
941 pgd_t pgd, *pgdp;
942 pud_t pud, *pudp;
943 pmd_t pmd, *pmdp;
944 pte_t *ret_pte;
945 hugepd_t *hpdp = NULL;
946 unsigned pdshift = PGDIR_SHIFT;
948 if (shift)
949 *shift = 0;
951 pgdp = pgdir + pgd_index(ea);
952 pgd = ACCESS_ONCE(*pgdp);
954 * Always operate on the local stack value. This make sure the
955 * value don't get updated by a parallel THP split/collapse,
956 * page fault or a page unmap. The return pte_t * is still not
957 * stable. So should be checked there for above conditions.
959 if (pgd_none(pgd))
960 return NULL;
961 else if (pgd_huge(pgd)) {
962 ret_pte = (pte_t *) pgdp;
963 goto out;
964 } else if (is_hugepd(&pgd))
965 hpdp = (hugepd_t *)&pgd;
966 else {
968 * Even if we end up with an unmap, the pgtable will not
969 * be freed, because we do an rcu free and here we are
970 * irq disabled
972 pdshift = PUD_SHIFT;
973 pudp = pud_offset(&pgd, ea);
974 pud = ACCESS_ONCE(*pudp);
976 if (pud_none(pud))
977 return NULL;
978 else if (pud_huge(pud)) {
979 ret_pte = (pte_t *) pudp;
980 goto out;
981 } else if (is_hugepd(&pud))
982 hpdp = (hugepd_t *)&pud;
983 else {
984 pdshift = PMD_SHIFT;
985 pmdp = pmd_offset(&pud, ea);
986 pmd = ACCESS_ONCE(*pmdp);
988 * A hugepage collapse is captured by pmd_none, because
989 * it mark the pmd none and do a hpte invalidate.
991 * A hugepage split is captured by pmd_trans_splitting
992 * because we mark the pmd trans splitting and do a
993 * hpte invalidate
996 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
997 return NULL;
999 if (pmd_huge(pmd) || pmd_large(pmd)) {
1000 ret_pte = (pte_t *) pmdp;
1001 goto out;
1002 } else if (is_hugepd(&pmd))
1003 hpdp = (hugepd_t *)&pmd;
1004 else
1005 return pte_offset_kernel(&pmd, ea);
1008 if (!hpdp)
1009 return NULL;
1011 ret_pte = hugepte_offset(hpdp, ea, pdshift);
1012 pdshift = hugepd_shift(*hpdp);
1013 out:
1014 if (shift)
1015 *shift = pdshift;
1016 return ret_pte;
1018 EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
1020 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1021 unsigned long end, int write, struct page **pages, int *nr)
1023 unsigned long mask;
1024 unsigned long pte_end;
1025 struct page *head, *page, *tail;
1026 pte_t pte;
1027 int refs;
1029 pte_end = (addr + sz) & ~(sz-1);
1030 if (pte_end < end)
1031 end = pte_end;
1033 pte = ACCESS_ONCE(*ptep);
1034 mask = _PAGE_PRESENT | _PAGE_USER;
1035 if (write)
1036 mask |= _PAGE_RW;
1038 if ((pte_val(pte) & mask) != mask)
1039 return 0;
1041 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1043 * check for splitting here
1045 if (pmd_trans_splitting(pte_pmd(pte)))
1046 return 0;
1047 #endif
1049 /* hugepages are never "special" */
1050 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1052 refs = 0;
1053 head = pte_page(pte);
1055 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1056 tail = page;
1057 do {
1058 VM_BUG_ON(compound_head(page) != head);
1059 pages[*nr] = page;
1060 (*nr)++;
1061 page++;
1062 refs++;
1063 } while (addr += PAGE_SIZE, addr != end);
1065 if (!page_cache_add_speculative(head, refs)) {
1066 *nr -= refs;
1067 return 0;
1070 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1071 /* Could be optimized better */
1072 *nr -= refs;
1073 while (refs--)
1074 put_page(head);
1075 return 0;
1079 * Any tail page need their mapcount reference taken before we
1080 * return.
1082 while (refs--) {
1083 if (PageTail(tail))
1084 get_huge_page_tail(tail);
1085 tail++;
1088 return 1;