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 if (pud_sect_supported())
44 order = PUD_SHIFT - PAGE_SHIFT;
46 order = CONT_PMD_SHIFT - PAGE_SHIFT;
48 hugetlb_cma_reserve(order);
50 #endif /* CONFIG_CMA */
52 static bool __hugetlb_valid_size(unsigned long size)
55 #ifndef __PAGETABLE_PMD_FOLDED
57 return pud_sect_supported();
68 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
69 bool arch_hugetlb_migration_supported(struct hstate *h)
71 size_t pagesize = huge_page_size(h);
73 if (!__hugetlb_valid_size(pagesize)) {
74 pr_warn("%s: unrecognized huge page size 0x%lx\n",
82 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
83 pte_t *ptep, size_t *pgsize)
85 pgd_t *pgdp = pgd_offset(mm, addr);
91 p4dp = p4d_offset(pgdp, addr);
92 pudp = pud_offset(p4dp, addr);
93 pmdp = pmd_offset(pudp, addr);
94 if ((pte_t *)pmdp == ptep) {
101 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
108 #ifndef __PAGETABLE_PMD_FOLDED
110 if (pud_sect_supported())
119 contig_ptes = CONT_PMDS;
123 contig_ptes = CONT_PTES;
130 pte_t huge_ptep_get(pte_t *ptep)
134 pte_t orig_pte = __ptep_get(ptep);
136 if (!pte_present(orig_pte) || !pte_cont(orig_pte))
139 ncontig = num_contig_ptes(page_size(pte_page(orig_pte)), &pgsize);
140 for (i = 0; i < ncontig; i++, ptep++) {
141 pte_t pte = __ptep_get(ptep);
144 orig_pte = pte_mkdirty(orig_pte);
147 orig_pte = pte_mkyoung(orig_pte);
153 * Changing some bits of contiguous entries requires us to follow a
154 * Break-Before-Make approach, breaking the whole contiguous set
155 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
156 * "Misprogramming of the Contiguous bit", page D4-1762.
158 * This helper performs the break step.
160 static pte_t get_clear_contig(struct mm_struct *mm,
163 unsigned long pgsize,
164 unsigned long ncontig)
166 pte_t orig_pte = __ptep_get(ptep);
169 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
170 pte_t pte = __ptep_get_and_clear(mm, addr, ptep);
173 * If HW_AFDBM is enabled, then the HW could turn on
174 * the dirty or accessed bit for any page in the set,
178 orig_pte = pte_mkdirty(orig_pte);
181 orig_pte = pte_mkyoung(orig_pte);
186 static pte_t get_clear_contig_flush(struct mm_struct *mm,
189 unsigned long pgsize,
190 unsigned long ncontig)
192 pte_t orig_pte = get_clear_contig(mm, addr, ptep, pgsize, ncontig);
193 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
195 flush_tlb_range(&vma, addr, addr + (pgsize * ncontig));
200 * Changing some bits of contiguous entries requires us to follow a
201 * Break-Before-Make approach, breaking the whole contiguous set
202 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
203 * "Misprogramming of the Contiguous bit", page D4-1762.
205 * This helper performs the break step for use cases where the
206 * original pte is not needed.
208 static void clear_flush(struct mm_struct *mm,
211 unsigned long pgsize,
212 unsigned long ncontig)
214 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
215 unsigned long i, saddr = addr;
217 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
218 __ptep_get_and_clear(mm, addr, ptep);
220 flush_tlb_range(&vma, saddr, addr);
223 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
224 pte_t *ptep, pte_t pte, unsigned long sz)
229 unsigned long pfn, dpfn;
232 ncontig = num_contig_ptes(sz, &pgsize);
234 if (!pte_present(pte)) {
235 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize)
236 __set_ptes(mm, addr, ptep, pte, 1);
240 if (!pte_cont(pte)) {
241 __set_ptes(mm, addr, ptep, pte, 1);
246 dpfn = pgsize >> PAGE_SHIFT;
247 hugeprot = pte_pgprot(pte);
249 clear_flush(mm, addr, ptep, pgsize, ncontig);
251 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
252 __set_ptes(mm, addr, ptep, pfn_pte(pfn, hugeprot), 1);
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_alloc(mm, pgdp, addr);
269 pudp = pud_alloc(mm, p4dp, addr);
273 if (sz == PUD_SIZE) {
274 ptep = (pte_t *)pudp;
275 } else if (sz == (CONT_PTE_SIZE)) {
276 pmdp = pmd_alloc(mm, pudp, addr);
280 WARN_ON(addr & (sz - 1));
281 ptep = pte_alloc_huge(mm, pmdp, addr);
282 } else if (sz == PMD_SIZE) {
283 if (want_pmd_share(vma, addr) && pud_none(READ_ONCE(*pudp)))
284 ptep = huge_pmd_share(mm, vma, addr, pudp);
286 ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
287 } else if (sz == (CONT_PMD_SIZE)) {
288 pmdp = pmd_alloc(mm, pudp, addr);
289 WARN_ON(addr & (sz - 1));
290 return (pte_t *)pmdp;
296 pte_t *huge_pte_offset(struct mm_struct *mm,
297 unsigned long addr, unsigned long sz)
304 pgdp = pgd_offset(mm, addr);
305 if (!pgd_present(READ_ONCE(*pgdp)))
308 p4dp = p4d_offset(pgdp, addr);
309 if (!p4d_present(READ_ONCE(*p4dp)))
312 pudp = pud_offset(p4dp, addr);
313 pud = READ_ONCE(*pudp);
314 if (sz != PUD_SIZE && pud_none(pud))
316 /* hugepage or swap? */
317 if (pud_leaf(pud) || !pud_present(pud))
318 return (pte_t *)pudp;
319 /* table; check the next level */
321 if (sz == CONT_PMD_SIZE)
322 addr &= CONT_PMD_MASK;
324 pmdp = pmd_offset(pudp, addr);
325 pmd = READ_ONCE(*pmdp);
326 if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
329 if (pmd_leaf(pmd) || !pmd_present(pmd))
330 return (pte_t *)pmdp;
332 if (sz == CONT_PTE_SIZE)
333 return pte_offset_huge(pmdp, (addr & CONT_PTE_MASK));
338 unsigned long hugetlb_mask_last_page(struct hstate *h)
340 unsigned long hp_size = huge_page_size(h);
343 #ifndef __PAGETABLE_PMD_FOLDED
345 return PGDIR_SIZE - PUD_SIZE;
348 return PUD_SIZE - CONT_PMD_SIZE;
350 return PUD_SIZE - PMD_SIZE;
352 return PMD_SIZE - CONT_PTE_SIZE;
360 pte_t arch_make_huge_pte(pte_t entry, unsigned int shift, vm_flags_t flags)
362 size_t pagesize = 1UL << shift;
364 entry = pte_mkhuge(entry);
365 if (pagesize == CONT_PTE_SIZE) {
366 entry = pte_mkcont(entry);
367 } else if (pagesize == CONT_PMD_SIZE) {
368 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
369 } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
370 pr_warn("%s: unrecognized huge page size 0x%lx\n",
376 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
377 pte_t *ptep, unsigned long sz)
382 ncontig = num_contig_ptes(sz, &pgsize);
384 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
385 __pte_clear(mm, addr, ptep);
388 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
389 unsigned long addr, pte_t *ptep)
393 pte_t orig_pte = __ptep_get(ptep);
395 if (!pte_cont(orig_pte))
396 return __ptep_get_and_clear(mm, addr, ptep);
398 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
400 return get_clear_contig(mm, addr, ptep, pgsize, ncontig);
404 * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
405 * and write permission.
407 * For a contiguous huge pte range we need to check whether or not write
408 * permission has to change only on the first pte in the set. Then for
409 * all the contiguous ptes we need to check whether or not there is a
410 * discrepancy between dirty or young.
412 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
416 if (pte_write(pte) != pte_write(__ptep_get(ptep)))
419 for (i = 0; i < ncontig; i++) {
420 pte_t orig_pte = __ptep_get(ptep + i);
422 if (pte_dirty(pte) != pte_dirty(orig_pte))
425 if (pte_young(pte) != pte_young(orig_pte))
432 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
433 unsigned long addr, pte_t *ptep,
434 pte_t pte, int dirty)
438 unsigned long pfn = pte_pfn(pte), dpfn;
439 struct mm_struct *mm = vma->vm_mm;
444 return __ptep_set_access_flags(vma, addr, ptep, pte, dirty);
446 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
447 dpfn = pgsize >> PAGE_SHIFT;
449 if (!__cont_access_flags_changed(ptep, pte, ncontig))
452 orig_pte = get_clear_contig_flush(mm, addr, ptep, pgsize, ncontig);
454 /* Make sure we don't lose the dirty or young state */
455 if (pte_dirty(orig_pte))
456 pte = pte_mkdirty(pte);
458 if (pte_young(orig_pte))
459 pte = pte_mkyoung(pte);
461 hugeprot = pte_pgprot(pte);
462 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
463 __set_ptes(mm, addr, ptep, pfn_pte(pfn, hugeprot), 1);
468 void huge_ptep_set_wrprotect(struct mm_struct *mm,
469 unsigned long addr, pte_t *ptep)
471 unsigned long pfn, dpfn;
477 if (!pte_cont(__ptep_get(ptep))) {
478 __ptep_set_wrprotect(mm, addr, ptep);
482 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
483 dpfn = pgsize >> PAGE_SHIFT;
485 pte = get_clear_contig_flush(mm, addr, ptep, pgsize, ncontig);
486 pte = pte_wrprotect(pte);
488 hugeprot = pte_pgprot(pte);
491 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
492 __set_ptes(mm, addr, ptep, pfn_pte(pfn, hugeprot), 1);
495 pte_t huge_ptep_clear_flush(struct vm_area_struct *vma,
496 unsigned long addr, pte_t *ptep)
498 struct mm_struct *mm = vma->vm_mm;
502 if (!pte_cont(__ptep_get(ptep)))
503 return ptep_clear_flush(vma, addr, ptep);
505 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
506 return get_clear_contig_flush(mm, addr, ptep, pgsize, ncontig);
509 static int __init hugetlbpage_init(void)
511 if (pud_sect_supported())
512 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
514 hugetlb_add_hstate(CONT_PMD_SHIFT - PAGE_SHIFT);
515 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
516 hugetlb_add_hstate(CONT_PTE_SHIFT - PAGE_SHIFT);
520 arch_initcall(hugetlbpage_init);
522 bool __init arch_hugetlb_valid_size(unsigned long size)
524 return __hugetlb_valid_size(size);
527 pte_t huge_ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep)
529 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) {
531 * Break-before-make (BBM) is required for all user space mappings
532 * when the permission changes from executable to non-executable
533 * in cases where cpu is affected with errata #2645198.
535 if (pte_user_exec(__ptep_get(ptep)))
536 return huge_ptep_clear_flush(vma, addr, ptep);
538 return huge_ptep_get_and_clear(vma->vm_mm, addr, ptep);
541 void huge_ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep,
542 pte_t old_pte, pte_t pte)
544 unsigned long psize = huge_page_size(hstate_vma(vma));
546 set_huge_pte_at(vma->vm_mm, addr, ptep, pte, psize);