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/moduleparam.h>
19 #include <linux/swap.h>
20 #include <linux/swapops.h>
21 #include <linux/kmemleak.h>
22 #include <asm/pgalloc.h>
24 #include <asm/setup.h>
25 #include <asm/hugetlb.h>
26 #include <asm/pte-walk.h>
28 bool hugetlb_disabled = false;
30 #define hugepd_none(hpd) (hpd_val(hpd) == 0)
32 #define PTE_T_ORDER (__builtin_ffs(sizeof(pte_basic_t)) - \
33 __builtin_ffs(sizeof(void *)))
35 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
38 * Only called for hugetlbfs pages, hence can ignore THP and the
41 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
44 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
45 unsigned long address, unsigned int pdshift,
46 unsigned int pshift, spinlock_t *ptl)
48 struct kmem_cache *cachep;
53 if (pshift >= pdshift) {
54 cachep = PGT_CACHE(PTE_T_ORDER);
55 num_hugepd = 1 << (pshift - pdshift);
57 cachep = PGT_CACHE(pdshift - pshift);
62 WARN_ONCE(1, "No page table cache created for hugetlb tables");
66 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
68 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
69 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
75 * Make sure other cpus find the hugepd set only after a
76 * properly initialized page table is visible to them.
77 * For more details look for comment in __pte_alloc().
83 * We have multiple higher-level entries that point to the same
84 * actual pte location. Fill in each as we go and backtrack on error.
85 * We need all of these so the DTLB pgtable walk code can find the
86 * right higher-level entry without knowing if it's a hugepage or not.
88 for (i = 0; i < num_hugepd; i++, hpdp++) {
89 if (unlikely(!hugepd_none(*hpdp)))
91 hugepd_populate(hpdp, new, pshift);
93 /* If we bailed from the for loop early, an error occurred, clean up */
95 for (i = i - 1 ; i >= 0; i--, hpdp--)
97 kmem_cache_free(cachep, new);
106 * At this point we do the placement change only for BOOK3S 64. This would
107 * possibly work on other subarchs.
109 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
110 unsigned long addr, unsigned long sz)
116 hugepd_t *hpdp = NULL;
117 unsigned pshift = __ffs(sz);
118 unsigned pdshift = PGDIR_SHIFT;
122 pg = pgd_offset(mm, addr);
123 p4 = p4d_offset(pg, addr);
125 #ifdef CONFIG_PPC_BOOK3S_64
126 if (pshift == PGDIR_SHIFT)
129 else if (pshift > PUD_SHIFT) {
131 * We need to use hugepd table
133 ptl = &mm->page_table_lock;
134 hpdp = (hugepd_t *)p4;
137 pu = pud_alloc(mm, p4, addr);
140 if (pshift == PUD_SHIFT)
142 else if (pshift > PMD_SHIFT) {
143 ptl = pud_lockptr(mm, pu);
144 hpdp = (hugepd_t *)pu;
147 pm = pmd_alloc(mm, pu, addr);
150 if (pshift == PMD_SHIFT)
154 ptl = pmd_lockptr(mm, pm);
155 hpdp = (hugepd_t *)pm;
160 if (pshift >= PGDIR_SHIFT) {
161 ptl = &mm->page_table_lock;
162 hpdp = (hugepd_t *)p4;
165 pu = pud_alloc(mm, p4, addr);
168 if (pshift >= PUD_SHIFT) {
169 ptl = pud_lockptr(mm, pu);
170 hpdp = (hugepd_t *)pu;
173 pm = pmd_alloc(mm, pu, addr);
176 ptl = pmd_lockptr(mm, pm);
177 hpdp = (hugepd_t *)pm;
184 if (IS_ENABLED(CONFIG_PPC_8xx) && pshift < PMD_SHIFT)
185 return pte_alloc_map(mm, (pmd_t *)hpdp, addr);
187 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
189 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
190 pdshift, pshift, ptl))
193 return hugepte_offset(*hpdp, addr, pdshift);
196 #ifdef CONFIG_PPC_BOOK3S_64
198 * Tracks gpages after the device tree is scanned and before the
199 * huge_boot_pages list is ready on pseries.
201 #define MAX_NUMBER_GPAGES 1024
202 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
203 __initdata static unsigned nr_gpages;
206 * Build list of addresses of gigantic pages. This function is used in early
207 * boot before the buddy allocator is setup.
209 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
213 while (number_of_pages > 0) {
214 gpage_freearray[nr_gpages] = addr;
221 static int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
223 struct huge_bootmem_page *m;
226 m = phys_to_virt(gpage_freearray[--nr_gpages]);
227 gpage_freearray[nr_gpages] = 0;
228 list_add(&m->list, &huge_boot_pages);
235 int __init alloc_bootmem_huge_page(struct hstate *h)
238 #ifdef CONFIG_PPC_BOOK3S_64
239 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
240 return pseries_alloc_bootmem_huge_page(h);
242 return __alloc_bootmem_huge_page(h);
245 #ifndef CONFIG_PPC_BOOK3S_64
246 #define HUGEPD_FREELIST_SIZE \
247 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
249 struct hugepd_freelist {
255 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
257 static void hugepd_free_rcu_callback(struct rcu_head *head)
259 struct hugepd_freelist *batch =
260 container_of(head, struct hugepd_freelist, rcu);
263 for (i = 0; i < batch->index; i++)
264 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
266 free_page((unsigned long)batch);
269 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
271 struct hugepd_freelist **batchp;
273 batchp = &get_cpu_var(hugepd_freelist_cur);
275 if (atomic_read(&tlb->mm->mm_users) < 2 ||
276 mm_is_thread_local(tlb->mm)) {
277 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
278 put_cpu_var(hugepd_freelist_cur);
282 if (*batchp == NULL) {
283 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
284 (*batchp)->index = 0;
287 (*batchp)->ptes[(*batchp)->index++] = hugepte;
288 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
289 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
292 put_cpu_var(hugepd_freelist_cur);
295 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
298 /* Return true when the entry to be freed maps more than the area being freed */
299 static bool range_is_outside_limits(unsigned long start, unsigned long end,
300 unsigned long floor, unsigned long ceiling,
303 if ((start & mask) < floor)
310 return end - 1 > ceiling - 1;
313 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
314 unsigned long start, unsigned long end,
315 unsigned long floor, unsigned long ceiling)
317 pte_t *hugepte = hugepd_page(*hpdp);
320 unsigned long pdmask = ~((1UL << pdshift) - 1);
321 unsigned int num_hugepd = 1;
322 unsigned int shift = hugepd_shift(*hpdp);
324 /* Note: On fsl the hpdp may be the first of several */
326 num_hugepd = 1 << (shift - pdshift);
328 if (range_is_outside_limits(start, end, floor, ceiling, pdmask))
331 for (i = 0; i < num_hugepd; i++, hpdp++)
334 if (shift >= pdshift)
335 hugepd_free(tlb, hugepte);
337 pgtable_free_tlb(tlb, hugepte,
338 get_hugepd_cache_index(pdshift - shift));
341 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
342 unsigned long addr, unsigned long end,
343 unsigned long floor, unsigned long ceiling)
345 pgtable_t token = pmd_pgtable(*pmd);
347 if (range_is_outside_limits(addr, end, floor, ceiling, PMD_MASK))
351 pte_free_tlb(tlb, token, addr);
352 mm_dec_nr_ptes(tlb->mm);
355 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
356 unsigned long addr, unsigned long end,
357 unsigned long floor, unsigned long ceiling)
367 pmd = pmd_offset(pud, addr);
368 next = pmd_addr_end(addr, end);
369 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
370 if (pmd_none_or_clear_bad(pmd))
374 * if it is not hugepd pointer, we should already find
377 WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
379 hugetlb_free_pte_range(tlb, pmd, addr, end, floor, ceiling);
384 * Increment next by the size of the huge mapping since
385 * there may be more than one entry at this level for a
386 * single hugepage, but all of them point to
387 * the same kmem cache that holds the hugepte.
389 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
393 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
394 addr, next, floor, ceiling);
395 } while (addr = next, addr != end);
397 if (range_is_outside_limits(start, end, floor, ceiling, PUD_MASK))
400 pmd = pmd_offset(pud, start & PUD_MASK);
402 pmd_free_tlb(tlb, pmd, start & PUD_MASK);
403 mm_dec_nr_pmds(tlb->mm);
406 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
407 unsigned long addr, unsigned long end,
408 unsigned long floor, unsigned long ceiling)
416 pud = pud_offset(p4d, addr);
417 next = pud_addr_end(addr, end);
418 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
419 if (pud_none_or_clear_bad(pud))
421 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
426 * Increment next by the size of the huge mapping since
427 * there may be more than one entry at this level for a
428 * single hugepage, but all of them point to
429 * the same kmem cache that holds the hugepte.
431 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
435 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
436 addr, next, floor, ceiling);
438 } while (addr = next, addr != end);
440 if (range_is_outside_limits(start, end, floor, ceiling, PGDIR_MASK))
443 pud = pud_offset(p4d, start & PGDIR_MASK);
445 pud_free_tlb(tlb, pud, start & PGDIR_MASK);
446 mm_dec_nr_puds(tlb->mm);
450 * This function frees user-level page tables of a process.
452 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
453 unsigned long addr, unsigned long end,
454 unsigned long floor, unsigned long ceiling)
461 * Because there are a number of different possible pagetable
462 * layouts for hugepage ranges, we limit knowledge of how
463 * things should be laid out to the allocation path
464 * (huge_pte_alloc(), above). Everything else works out the
465 * structure as it goes from information in the hugepd
466 * pointers. That means that we can't here use the
467 * optimization used in the normal page free_pgd_range(), of
468 * checking whether we're actually covering a large enough
469 * range to have to do anything at the top level of the walk
470 * instead of at the bottom.
472 * To make sense of this, you should probably go read the big
473 * block comment at the top of the normal free_pgd_range(),
478 next = pgd_addr_end(addr, end);
479 pgd = pgd_offset(tlb->mm, addr);
480 p4d = p4d_offset(pgd, addr);
481 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
482 if (p4d_none_or_clear_bad(p4d))
484 hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
488 * Increment next by the size of the huge mapping since
489 * there may be more than one entry at the pgd level
490 * for a single hugepage, but all of them point to the
491 * same kmem cache that holds the hugepte.
493 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
497 free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
498 addr, next, floor, ceiling);
500 } while (addr = next, addr != end);
503 struct page *follow_huge_pd(struct vm_area_struct *vma,
504 unsigned long address, hugepd_t hpd,
505 int flags, int pdshift)
509 struct page *page = NULL;
511 int shift = hugepd_shift(hpd);
512 struct mm_struct *mm = vma->vm_mm;
516 * hugepage directory entries are protected by mm->page_table_lock
517 * Use this instead of huge_pte_lockptr
519 ptl = &mm->page_table_lock;
522 ptep = hugepte_offset(hpd, address, pdshift);
523 if (pte_present(*ptep)) {
524 mask = (1UL << shift) - 1;
525 page = pte_page(*ptep);
526 page += ((address & mask) >> PAGE_SHIFT);
527 if (flags & FOLL_GET)
530 if (is_hugetlb_entry_migration(*ptep)) {
532 __migration_entry_wait(mm, ptep, ptl);
540 #ifdef CONFIG_PPC_MM_SLICES
541 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
542 unsigned long len, unsigned long pgoff,
545 struct hstate *hstate = hstate_file(file);
546 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
548 #ifdef CONFIG_PPC_RADIX_MMU
550 return radix__hugetlb_get_unmapped_area(file, addr, len,
553 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
557 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
559 /* With radix we don't use slice, so derive it from vma*/
560 if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
561 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
563 return 1UL << mmu_psize_to_shift(psize);
565 return vma_kernel_pagesize(vma);
568 bool __init arch_hugetlb_valid_size(unsigned long size)
570 int shift = __ffs(size);
573 /* Check that it is a page size supported by the hardware and
574 * that it fits within pagetable and slice limits. */
575 if (size <= PAGE_SIZE || !is_power_of_2(size))
578 mmu_psize = check_and_get_huge_psize(shift);
582 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
587 static int __init add_huge_page_size(unsigned long long size)
589 int shift = __ffs(size);
591 if (!arch_hugetlb_valid_size((unsigned long)size))
594 hugetlb_add_hstate(shift - PAGE_SHIFT);
598 static int __init hugetlbpage_init(void)
600 bool configured = false;
603 if (hugetlb_disabled) {
604 pr_info("HugeTLB support is disabled!\n");
608 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
609 !mmu_has_feature(MMU_FTR_16M_PAGE))
612 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
616 if (!mmu_psize_defs[psize].shift)
619 shift = mmu_psize_to_shift(psize);
621 #ifdef CONFIG_PPC_BOOK3S_64
622 if (shift > PGDIR_SHIFT)
624 else if (shift > PUD_SHIFT)
625 pdshift = PGDIR_SHIFT;
626 else if (shift > PMD_SHIFT)
631 if (shift < PUD_SHIFT)
633 else if (shift < PGDIR_SHIFT)
636 pdshift = PGDIR_SHIFT;
639 if (add_huge_page_size(1ULL << shift) < 0)
642 * if we have pdshift and shift value same, we don't
643 * use pgt cache for hugepd.
645 if (pdshift > shift) {
646 if (!IS_ENABLED(CONFIG_PPC_8xx))
647 pgtable_cache_add(pdshift - shift);
648 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
649 IS_ENABLED(CONFIG_PPC_8xx)) {
650 pgtable_cache_add(PTE_T_ORDER);
657 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
658 hugetlbpage_init_default();
660 pr_info("Failed to initialize. Disabling HugeTLB");
665 arch_initcall(hugetlbpage_init);
667 void __init gigantic_hugetlb_cma_reserve(void)
669 unsigned long order = 0;
672 order = PUD_SHIFT - PAGE_SHIFT;
673 else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
675 * For pseries we do use ibm,expected#pages for reserving 16G pages.
677 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
680 VM_WARN_ON(order < MAX_ORDER);
681 hugetlb_cma_reserve(order);