1 // SPDX-License-Identifier: GPL-2.0-only
3 * arch/arm64/mm/hugetlbpage.c
5 * Copyright (C) 2013 Linaro Ltd.
7 * Based on arch/x86/mm/hugetlbpage.c.
10 #include <linux/init.h>
13 #include <linux/hugetlb.h>
14 #include <linux/pagemap.h>
15 #include <linux/err.h>
16 #include <linux/sysctl.h>
19 #include <asm/tlbflush.h>
22 * HugeTLB Support Matrix
24 * ---------------------------------------------------
25 * | Page Size | CONT PTE | PMD | CONT PMD | PUD |
26 * ---------------------------------------------------
27 * | 4K | 64K | 2M | 32M | 1G |
28 * | 16K | 2M | 32M | 1G | |
29 * | 64K | 2M | 512M | 16G | |
30 * ---------------------------------------------------
34 * Reserve CMA areas for the largest supported gigantic
35 * huge page when requested. Any other smaller gigantic
36 * huge pages could still be served from those areas.
39 void __init arm64_hugetlb_cma_reserve(void)
43 #ifdef CONFIG_ARM64_4K_PAGES
44 order = PUD_SHIFT - PAGE_SHIFT;
46 order = CONT_PMD_SHIFT - PAGE_SHIFT;
49 * HugeTLB CMA reservation is required for gigantic
50 * huge pages which could not be allocated via the
51 * page allocator. Just warn if there is any change
52 * breaking this assumption.
54 WARN_ON(order <= MAX_ORDER);
55 hugetlb_cma_reserve(order);
57 #endif /* CONFIG_CMA */
59 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
60 bool arch_hugetlb_migration_supported(struct hstate *h)
62 size_t pagesize = huge_page_size(h);
65 #ifdef CONFIG_ARM64_4K_PAGES
73 pr_warn("%s: unrecognized huge page size 0x%lx\n",
79 int pmd_huge(pmd_t pmd)
81 return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
84 int pud_huge(pud_t pud)
86 #ifndef __PAGETABLE_PMD_FOLDED
87 return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
94 * Select all bits except the pfn
96 static inline pgprot_t pte_pgprot(pte_t pte)
98 unsigned long pfn = pte_pfn(pte);
100 return __pgprot(pte_val(pfn_pte(pfn, __pgprot(0))) ^ pte_val(pte));
103 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
104 pte_t *ptep, size_t *pgsize)
106 pgd_t *pgdp = pgd_offset(mm, addr);
112 p4dp = p4d_offset(pgdp, addr);
113 pudp = pud_offset(p4dp, addr);
114 pmdp = pmd_offset(pudp, addr);
115 if ((pte_t *)pmdp == ptep) {
122 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
129 #ifdef CONFIG_ARM64_4K_PAGES
137 contig_ptes = CONT_PMDS;
141 contig_ptes = CONT_PTES;
149 * Changing some bits of contiguous entries requires us to follow a
150 * Break-Before-Make approach, breaking the whole contiguous set
151 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
152 * "Misprogramming of the Contiguous bit", page D4-1762.
154 * This helper performs the break step.
156 static pte_t get_clear_flush(struct mm_struct *mm,
159 unsigned long pgsize,
160 unsigned long ncontig)
162 pte_t orig_pte = huge_ptep_get(ptep);
163 bool valid = pte_valid(orig_pte);
164 unsigned long i, saddr = addr;
166 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
167 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
170 * If HW_AFDBM is enabled, then the HW could turn on
171 * the dirty or accessed bit for any page in the set,
175 orig_pte = pte_mkdirty(orig_pte);
178 orig_pte = pte_mkyoung(orig_pte);
182 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
183 flush_tlb_range(&vma, saddr, addr);
189 * Changing some bits of contiguous entries requires us to follow a
190 * Break-Before-Make approach, breaking the whole contiguous set
191 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
192 * "Misprogramming of the Contiguous bit", page D4-1762.
194 * This helper performs the break step for use cases where the
195 * original pte is not needed.
197 static void clear_flush(struct mm_struct *mm,
200 unsigned long pgsize,
201 unsigned long ncontig)
203 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
204 unsigned long i, saddr = addr;
206 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
207 pte_clear(mm, addr, ptep);
209 flush_tlb_range(&vma, saddr, addr);
212 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
213 pte_t *ptep, pte_t pte)
218 unsigned long pfn, dpfn;
222 * Code needs to be expanded to handle huge swap and migration
223 * entries. Needed for HUGETLB and MEMORY_FAILURE.
225 WARN_ON(!pte_present(pte));
227 if (!pte_cont(pte)) {
228 set_pte_at(mm, addr, ptep, pte);
232 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
234 dpfn = pgsize >> PAGE_SHIFT;
235 hugeprot = pte_pgprot(pte);
237 clear_flush(mm, addr, ptep, pgsize, ncontig);
239 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
240 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
243 void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
244 pte_t *ptep, pte_t pte, unsigned long sz)
249 ncontig = num_contig_ptes(sz, &pgsize);
251 for (i = 0; i < ncontig; i++, ptep++)
255 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
256 unsigned long addr, unsigned long sz)
264 pgdp = pgd_offset(mm, addr);
265 p4dp = p4d_offset(pgdp, addr);
266 pudp = pud_alloc(mm, p4dp, addr);
270 if (sz == PUD_SIZE) {
271 ptep = (pte_t *)pudp;
272 } else if (sz == (CONT_PTE_SIZE)) {
273 pmdp = pmd_alloc(mm, pudp, addr);
277 WARN_ON(addr & (sz - 1));
279 * Note that if this code were ever ported to the
280 * 32-bit arm platform then it will cause trouble in
281 * the case where CONFIG_HIGHPTE is set, since there
282 * will be no pte_unmap() to correspond with this
285 ptep = pte_alloc_map(mm, pmdp, addr);
286 } else if (sz == PMD_SIZE) {
287 if (want_pmd_share(vma, addr) && pud_none(READ_ONCE(*pudp)))
288 ptep = huge_pmd_share(mm, vma, addr, pudp);
290 ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
291 } else if (sz == (CONT_PMD_SIZE)) {
292 pmdp = pmd_alloc(mm, pudp, addr);
293 WARN_ON(addr & (sz - 1));
294 return (pte_t *)pmdp;
300 pte_t *huge_pte_offset(struct mm_struct *mm,
301 unsigned long addr, unsigned long sz)
308 pgdp = pgd_offset(mm, addr);
309 if (!pgd_present(READ_ONCE(*pgdp)))
312 p4dp = p4d_offset(pgdp, addr);
313 if (!p4d_present(READ_ONCE(*p4dp)))
316 pudp = pud_offset(p4dp, addr);
317 pud = READ_ONCE(*pudp);
318 if (sz != PUD_SIZE && pud_none(pud))
320 /* hugepage or swap? */
321 if (pud_huge(pud) || !pud_present(pud))
322 return (pte_t *)pudp;
323 /* table; check the next level */
325 if (sz == CONT_PMD_SIZE)
326 addr &= CONT_PMD_MASK;
328 pmdp = pmd_offset(pudp, addr);
329 pmd = READ_ONCE(*pmdp);
330 if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
333 if (pmd_huge(pmd) || !pmd_present(pmd))
334 return (pte_t *)pmdp;
336 if (sz == CONT_PTE_SIZE)
337 return pte_offset_kernel(pmdp, (addr & CONT_PTE_MASK));
342 pte_t arch_make_huge_pte(pte_t entry, unsigned int shift, vm_flags_t flags)
344 size_t pagesize = 1UL << shift;
346 if (pagesize == CONT_PTE_SIZE) {
347 entry = pte_mkcont(entry);
348 } else if (pagesize == CONT_PMD_SIZE) {
349 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
350 } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
351 pr_warn("%s: unrecognized huge page size 0x%lx\n",
357 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
358 pte_t *ptep, unsigned long sz)
363 ncontig = num_contig_ptes(sz, &pgsize);
365 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
366 pte_clear(mm, addr, ptep);
369 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
370 unsigned long addr, pte_t *ptep)
374 pte_t orig_pte = huge_ptep_get(ptep);
376 if (!pte_cont(orig_pte))
377 return ptep_get_and_clear(mm, addr, ptep);
379 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
381 return get_clear_flush(mm, addr, ptep, pgsize, ncontig);
385 * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
386 * and write permission.
388 * For a contiguous huge pte range we need to check whether or not write
389 * permission has to change only on the first pte in the set. Then for
390 * all the contiguous ptes we need to check whether or not there is a
391 * discrepancy between dirty or young.
393 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
397 if (pte_write(pte) != pte_write(huge_ptep_get(ptep)))
400 for (i = 0; i < ncontig; i++) {
401 pte_t orig_pte = huge_ptep_get(ptep + i);
403 if (pte_dirty(pte) != pte_dirty(orig_pte))
406 if (pte_young(pte) != pte_young(orig_pte))
413 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
414 unsigned long addr, pte_t *ptep,
415 pte_t pte, int dirty)
419 unsigned long pfn = pte_pfn(pte), dpfn;
424 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
426 ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
427 dpfn = pgsize >> PAGE_SHIFT;
429 if (!__cont_access_flags_changed(ptep, pte, ncontig))
432 orig_pte = get_clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
434 /* Make sure we don't lose the dirty or young state */
435 if (pte_dirty(orig_pte))
436 pte = pte_mkdirty(pte);
438 if (pte_young(orig_pte))
439 pte = pte_mkyoung(pte);
441 hugeprot = pte_pgprot(pte);
442 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
443 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
448 void huge_ptep_set_wrprotect(struct mm_struct *mm,
449 unsigned long addr, pte_t *ptep)
451 unsigned long pfn, dpfn;
457 if (!pte_cont(READ_ONCE(*ptep))) {
458 ptep_set_wrprotect(mm, addr, ptep);
462 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
463 dpfn = pgsize >> PAGE_SHIFT;
465 pte = get_clear_flush(mm, addr, ptep, pgsize, ncontig);
466 pte = pte_wrprotect(pte);
468 hugeprot = pte_pgprot(pte);
471 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
472 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
475 void huge_ptep_clear_flush(struct vm_area_struct *vma,
476 unsigned long addr, pte_t *ptep)
481 if (!pte_cont(READ_ONCE(*ptep))) {
482 ptep_clear_flush(vma, addr, ptep);
486 ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
487 clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
490 static int __init hugetlbpage_init(void)
492 #ifdef CONFIG_ARM64_4K_PAGES
493 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
495 hugetlb_add_hstate(CONT_PMD_SHIFT - PAGE_SHIFT);
496 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
497 hugetlb_add_hstate(CONT_PTE_SHIFT - PAGE_SHIFT);
501 arch_initcall(hugetlbpage_init);
503 bool __init arch_hugetlb_valid_size(unsigned long size)
506 #ifdef CONFIG_ARM64_4K_PAGES