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>
20 #include <asm/pgalloc.h>
22 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
23 bool arch_hugetlb_migration_supported(struct hstate *h)
25 size_t pagesize = huge_page_size(h);
28 #ifdef CONFIG_ARM64_4K_PAGES
36 pr_warn("%s: unrecognized huge page size 0x%lx\n",
42 int pmd_huge(pmd_t pmd)
44 return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
47 int pud_huge(pud_t pud)
49 #ifndef __PAGETABLE_PMD_FOLDED
50 return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
57 * Select all bits except the pfn
59 static inline pgprot_t pte_pgprot(pte_t pte)
61 unsigned long pfn = pte_pfn(pte);
63 return __pgprot(pte_val(pfn_pte(pfn, __pgprot(0))) ^ pte_val(pte));
66 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
67 pte_t *ptep, size_t *pgsize)
69 pgd_t *pgdp = pgd_offset(mm, addr);
74 pudp = pud_offset(pgdp, addr);
75 pmdp = pmd_offset(pudp, addr);
76 if ((pte_t *)pmdp == ptep) {
83 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
90 #ifdef CONFIG_ARM64_4K_PAGES
98 contig_ptes = CONT_PMDS;
102 contig_ptes = CONT_PTES;
110 * Changing some bits of contiguous entries requires us to follow a
111 * Break-Before-Make approach, breaking the whole contiguous set
112 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
113 * "Misprogramming of the Contiguous bit", page D4-1762.
115 * This helper performs the break step.
117 static pte_t get_clear_flush(struct mm_struct *mm,
120 unsigned long pgsize,
121 unsigned long ncontig)
123 pte_t orig_pte = huge_ptep_get(ptep);
124 bool valid = pte_valid(orig_pte);
125 unsigned long i, saddr = addr;
127 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
128 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
131 * If HW_AFDBM is enabled, then the HW could turn on
132 * the dirty or accessed bit for any page in the set,
136 orig_pte = pte_mkdirty(orig_pte);
139 orig_pte = pte_mkyoung(orig_pte);
143 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
144 flush_tlb_range(&vma, saddr, addr);
150 * Changing some bits of contiguous entries requires us to follow a
151 * Break-Before-Make approach, breaking the whole contiguous set
152 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
153 * "Misprogramming of the Contiguous bit", page D4-1762.
155 * This helper performs the break step for use cases where the
156 * original pte is not needed.
158 static void clear_flush(struct mm_struct *mm,
161 unsigned long pgsize,
162 unsigned long ncontig)
164 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
165 unsigned long i, saddr = addr;
167 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
168 pte_clear(mm, addr, ptep);
170 flush_tlb_range(&vma, saddr, addr);
173 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
174 pte_t *ptep, pte_t pte)
179 unsigned long pfn, dpfn;
183 * Code needs to be expanded to handle huge swap and migration
184 * entries. Needed for HUGETLB and MEMORY_FAILURE.
186 WARN_ON(!pte_present(pte));
188 if (!pte_cont(pte)) {
189 set_pte_at(mm, addr, ptep, pte);
193 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
195 dpfn = pgsize >> PAGE_SHIFT;
196 hugeprot = pte_pgprot(pte);
198 clear_flush(mm, addr, ptep, pgsize, ncontig);
200 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
201 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
204 void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
205 pte_t *ptep, pte_t pte, unsigned long sz)
210 ncontig = num_contig_ptes(sz, &pgsize);
212 for (i = 0; i < ncontig; i++, ptep++)
216 pte_t *huge_pte_alloc(struct mm_struct *mm,
217 unsigned long addr, unsigned long sz)
224 pgdp = pgd_offset(mm, addr);
225 pudp = pud_alloc(mm, pgdp, addr);
229 if (sz == PUD_SIZE) {
230 ptep = (pte_t *)pudp;
231 } else if (sz == (CONT_PTE_SIZE)) {
232 pmdp = pmd_alloc(mm, pudp, addr);
234 WARN_ON(addr & (sz - 1));
236 * Note that if this code were ever ported to the
237 * 32-bit arm platform then it will cause trouble in
238 * the case where CONFIG_HIGHPTE is set, since there
239 * will be no pte_unmap() to correspond with this
242 ptep = pte_alloc_map(mm, pmdp, addr);
243 } else if (sz == PMD_SIZE) {
244 if (IS_ENABLED(CONFIG_ARCH_WANT_HUGE_PMD_SHARE) &&
245 pud_none(READ_ONCE(*pudp)))
246 ptep = huge_pmd_share(mm, addr, pudp);
248 ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
249 } else if (sz == (CONT_PMD_SIZE)) {
250 pmdp = pmd_alloc(mm, pudp, addr);
251 WARN_ON(addr & (sz - 1));
252 return (pte_t *)pmdp;
258 pte_t *huge_pte_offset(struct mm_struct *mm,
259 unsigned long addr, unsigned long sz)
265 pgdp = pgd_offset(mm, addr);
266 if (!pgd_present(READ_ONCE(*pgdp)))
269 pudp = pud_offset(pgdp, addr);
270 pud = READ_ONCE(*pudp);
271 if (sz != PUD_SIZE && pud_none(pud))
273 /* hugepage or swap? */
274 if (pud_huge(pud) || !pud_present(pud))
275 return (pte_t *)pudp;
276 /* table; check the next level */
278 if (sz == CONT_PMD_SIZE)
279 addr &= CONT_PMD_MASK;
281 pmdp = pmd_offset(pudp, addr);
282 pmd = READ_ONCE(*pmdp);
283 if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
286 if (pmd_huge(pmd) || !pmd_present(pmd))
287 return (pte_t *)pmdp;
289 if (sz == CONT_PTE_SIZE)
290 return pte_offset_kernel(pmdp, (addr & CONT_PTE_MASK));
295 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
296 struct page *page, int writable)
298 size_t pagesize = huge_page_size(hstate_vma(vma));
300 if (pagesize == CONT_PTE_SIZE) {
301 entry = pte_mkcont(entry);
302 } else if (pagesize == CONT_PMD_SIZE) {
303 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
304 } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
305 pr_warn("%s: unrecognized huge page size 0x%lx\n",
311 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
312 pte_t *ptep, unsigned long sz)
317 ncontig = num_contig_ptes(sz, &pgsize);
319 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
320 pte_clear(mm, addr, ptep);
323 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
324 unsigned long addr, pte_t *ptep)
328 pte_t orig_pte = huge_ptep_get(ptep);
330 if (!pte_cont(orig_pte))
331 return ptep_get_and_clear(mm, addr, ptep);
333 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
335 return get_clear_flush(mm, addr, ptep, pgsize, ncontig);
339 * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
340 * and write permission.
342 * For a contiguous huge pte range we need to check whether or not write
343 * permission has to change only on the first pte in the set. Then for
344 * all the contiguous ptes we need to check whether or not there is a
345 * discrepancy between dirty or young.
347 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
351 if (pte_write(pte) != pte_write(huge_ptep_get(ptep)))
354 for (i = 0; i < ncontig; i++) {
355 pte_t orig_pte = huge_ptep_get(ptep + i);
357 if (pte_dirty(pte) != pte_dirty(orig_pte))
360 if (pte_young(pte) != pte_young(orig_pte))
367 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
368 unsigned long addr, pte_t *ptep,
369 pte_t pte, int dirty)
373 unsigned long pfn = pte_pfn(pte), dpfn;
378 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
380 ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
381 dpfn = pgsize >> PAGE_SHIFT;
383 if (!__cont_access_flags_changed(ptep, pte, ncontig))
386 orig_pte = get_clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
388 /* Make sure we don't lose the dirty or young state */
389 if (pte_dirty(orig_pte))
390 pte = pte_mkdirty(pte);
392 if (pte_young(orig_pte))
393 pte = pte_mkyoung(pte);
395 hugeprot = pte_pgprot(pte);
396 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
397 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
402 void huge_ptep_set_wrprotect(struct mm_struct *mm,
403 unsigned long addr, pte_t *ptep)
405 unsigned long pfn, dpfn;
411 if (!pte_cont(READ_ONCE(*ptep))) {
412 ptep_set_wrprotect(mm, addr, ptep);
416 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
417 dpfn = pgsize >> PAGE_SHIFT;
419 pte = get_clear_flush(mm, addr, ptep, pgsize, ncontig);
420 pte = pte_wrprotect(pte);
422 hugeprot = pte_pgprot(pte);
425 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
426 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
429 void huge_ptep_clear_flush(struct vm_area_struct *vma,
430 unsigned long addr, pte_t *ptep)
435 if (!pte_cont(READ_ONCE(*ptep))) {
436 ptep_clear_flush(vma, addr, ptep);
440 ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
441 clear_flush(vma->vm_mm, addr, ptep, pgsize, ncontig);
444 static void __init add_huge_page_size(unsigned long size)
446 if (size_to_hstate(size))
449 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
452 static int __init hugetlbpage_init(void)
454 #ifdef CONFIG_ARM64_4K_PAGES
455 add_huge_page_size(PUD_SIZE);
457 add_huge_page_size(CONT_PMD_SIZE);
458 add_huge_page_size(PMD_SIZE);
459 add_huge_page_size(CONT_PTE_SIZE);
463 arch_initcall(hugetlbpage_init);
465 static __init int setup_hugepagesz(char *opt)
467 unsigned long ps = memparse(opt, &opt);
470 #ifdef CONFIG_ARM64_4K_PAGES
476 add_huge_page_size(ps);
481 pr_err("hugepagesz: Unsupported page size %lu K\n", ps >> 10);
484 __setup("hugepagesz=", setup_hugepagesz);