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;
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 static bool __hugetlb_valid_size(unsigned long size)
62 #ifndef __PAGETABLE_PMD_FOLDED
64 return pud_sect_supported();
75 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
76 bool arch_hugetlb_migration_supported(struct hstate *h)
78 size_t pagesize = huge_page_size(h);
80 if (!__hugetlb_valid_size(pagesize)) {
81 pr_warn("%s: unrecognized huge page size 0x%lx\n",
89 int pmd_huge(pmd_t pmd)
91 return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
94 int pud_huge(pud_t pud)
96 #ifndef __PAGETABLE_PMD_FOLDED
97 return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
104 * Select all bits except the pfn
106 static inline pgprot_t pte_pgprot(pte_t pte)
108 unsigned long pfn = pte_pfn(pte);
110 return __pgprot(pte_val(pfn_pte(pfn, __pgprot(0))) ^ pte_val(pte));
113 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
114 pte_t *ptep, size_t *pgsize)
116 pgd_t *pgdp = pgd_offset(mm, addr);
122 p4dp = p4d_offset(pgdp, addr);
123 pudp = pud_offset(p4dp, addr);
124 pmdp = pmd_offset(pudp, addr);
125 if ((pte_t *)pmdp == ptep) {
132 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
139 #ifndef __PAGETABLE_PMD_FOLDED
141 if (pud_sect_supported())
150 contig_ptes = CONT_PMDS;
154 contig_ptes = CONT_PTES;
161 pte_t huge_ptep_get(pte_t *ptep)
165 pte_t orig_pte = ptep_get(ptep);
167 if (!pte_present(orig_pte) || !pte_cont(orig_pte))
170 ncontig = num_contig_ptes(page_size(pte_page(orig_pte)), &pgsize);
171 for (i = 0; i < ncontig; i++, ptep++) {
172 pte_t pte = ptep_get(ptep);
175 orig_pte = pte_mkdirty(orig_pte);
178 orig_pte = pte_mkyoung(orig_pte);
184 * Changing some bits of contiguous entries requires us to follow a
185 * Break-Before-Make approach, breaking the whole contiguous set
186 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
187 * "Misprogramming of the Contiguous bit", page D4-1762.
189 * This helper performs the break step.
191 static pte_t get_clear_contig(struct mm_struct *mm,
194 unsigned long pgsize,
195 unsigned long ncontig)
197 pte_t orig_pte = ptep_get(ptep);
200 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
201 pte_t pte = ptep_get_and_clear(mm, addr, ptep);
204 * If HW_AFDBM is enabled, then the HW could turn on
205 * the dirty or accessed bit for any page in the set,
209 orig_pte = pte_mkdirty(orig_pte);
212 orig_pte = pte_mkyoung(orig_pte);
218 * Changing some bits of contiguous entries requires us to follow a
219 * Break-Before-Make approach, breaking the whole contiguous set
220 * before we can change any entries. See ARM DDI 0487A.k_iss10775,
221 * "Misprogramming of the Contiguous bit", page D4-1762.
223 * This helper performs the break step for use cases where the
224 * original pte is not needed.
226 static void clear_flush(struct mm_struct *mm,
229 unsigned long pgsize,
230 unsigned long ncontig)
232 struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
233 unsigned long i, saddr = addr;
235 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
236 pte_clear(mm, addr, ptep);
238 flush_tlb_range(&vma, saddr, addr);
241 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
242 pte_t *ptep, pte_t pte)
247 unsigned long pfn, dpfn;
251 * Code needs to be expanded to handle huge swap and migration
252 * entries. Needed for HUGETLB and MEMORY_FAILURE.
254 WARN_ON(!pte_present(pte));
256 if (!pte_cont(pte)) {
257 set_pte_at(mm, addr, ptep, pte);
261 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
263 dpfn = pgsize >> PAGE_SHIFT;
264 hugeprot = pte_pgprot(pte);
266 clear_flush(mm, addr, ptep, pgsize, ncontig);
268 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
269 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
272 void set_huge_swap_pte_at(struct mm_struct *mm, unsigned long addr,
273 pte_t *ptep, pte_t pte, unsigned long sz)
278 ncontig = num_contig_ptes(sz, &pgsize);
280 for (i = 0; i < ncontig; i++, ptep++)
284 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
285 unsigned long addr, unsigned long sz)
293 pgdp = pgd_offset(mm, addr);
294 p4dp = p4d_offset(pgdp, addr);
295 pudp = pud_alloc(mm, p4dp, addr);
299 if (sz == PUD_SIZE) {
300 ptep = (pte_t *)pudp;
301 } else if (sz == (CONT_PTE_SIZE)) {
302 pmdp = pmd_alloc(mm, pudp, addr);
306 WARN_ON(addr & (sz - 1));
308 * Note that if this code were ever ported to the
309 * 32-bit arm platform then it will cause trouble in
310 * the case where CONFIG_HIGHPTE is set, since there
311 * will be no pte_unmap() to correspond with this
314 ptep = pte_alloc_map(mm, pmdp, addr);
315 } else if (sz == PMD_SIZE) {
316 if (want_pmd_share(vma, addr) && pud_none(READ_ONCE(*pudp)))
317 ptep = huge_pmd_share(mm, vma, addr, pudp);
319 ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
320 } else if (sz == (CONT_PMD_SIZE)) {
321 pmdp = pmd_alloc(mm, pudp, addr);
322 WARN_ON(addr & (sz - 1));
323 return (pte_t *)pmdp;
329 pte_t *huge_pte_offset(struct mm_struct *mm,
330 unsigned long addr, unsigned long sz)
337 pgdp = pgd_offset(mm, addr);
338 if (!pgd_present(READ_ONCE(*pgdp)))
341 p4dp = p4d_offset(pgdp, addr);
342 if (!p4d_present(READ_ONCE(*p4dp)))
345 pudp = pud_offset(p4dp, addr);
346 pud = READ_ONCE(*pudp);
347 if (sz != PUD_SIZE && pud_none(pud))
349 /* hugepage or swap? */
350 if (pud_huge(pud) || !pud_present(pud))
351 return (pte_t *)pudp;
352 /* table; check the next level */
354 if (sz == CONT_PMD_SIZE)
355 addr &= CONT_PMD_MASK;
357 pmdp = pmd_offset(pudp, addr);
358 pmd = READ_ONCE(*pmdp);
359 if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
362 if (pmd_huge(pmd) || !pmd_present(pmd))
363 return (pte_t *)pmdp;
365 if (sz == CONT_PTE_SIZE)
366 return pte_offset_kernel(pmdp, (addr & CONT_PTE_MASK));
371 pte_t arch_make_huge_pte(pte_t entry, unsigned int shift, vm_flags_t flags)
373 size_t pagesize = 1UL << shift;
375 entry = pte_mkhuge(entry);
376 if (pagesize == CONT_PTE_SIZE) {
377 entry = pte_mkcont(entry);
378 } else if (pagesize == CONT_PMD_SIZE) {
379 entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
380 } else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
381 pr_warn("%s: unrecognized huge page size 0x%lx\n",
387 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
388 pte_t *ptep, unsigned long sz)
393 ncontig = num_contig_ptes(sz, &pgsize);
395 for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
396 pte_clear(mm, addr, ptep);
399 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
400 unsigned long addr, pte_t *ptep)
404 pte_t orig_pte = ptep_get(ptep);
406 if (!pte_cont(orig_pte))
407 return ptep_get_and_clear(mm, addr, ptep);
409 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
411 return get_clear_contig(mm, addr, ptep, pgsize, ncontig);
415 * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
416 * and write permission.
418 * For a contiguous huge pte range we need to check whether or not write
419 * permission has to change only on the first pte in the set. Then for
420 * all the contiguous ptes we need to check whether or not there is a
421 * discrepancy between dirty or young.
423 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
427 if (pte_write(pte) != pte_write(ptep_get(ptep)))
430 for (i = 0; i < ncontig; i++) {
431 pte_t orig_pte = ptep_get(ptep + i);
433 if (pte_dirty(pte) != pte_dirty(orig_pte))
436 if (pte_young(pte) != pte_young(orig_pte))
443 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
444 unsigned long addr, pte_t *ptep,
445 pte_t pte, int dirty)
449 unsigned long pfn = pte_pfn(pte), dpfn;
454 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
456 ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
457 dpfn = pgsize >> PAGE_SHIFT;
459 if (!__cont_access_flags_changed(ptep, pte, ncontig))
462 orig_pte = get_clear_contig(vma->vm_mm, addr, ptep, pgsize, ncontig);
464 /* Make sure we don't lose the dirty or young state */
465 if (pte_dirty(orig_pte))
466 pte = pte_mkdirty(pte);
468 if (pte_young(orig_pte))
469 pte = pte_mkyoung(pte);
471 hugeprot = pte_pgprot(pte);
472 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
473 set_pte_at(vma->vm_mm, addr, ptep, pfn_pte(pfn, hugeprot));
478 void huge_ptep_set_wrprotect(struct mm_struct *mm,
479 unsigned long addr, pte_t *ptep)
481 unsigned long pfn, dpfn;
487 if (!pte_cont(READ_ONCE(*ptep))) {
488 ptep_set_wrprotect(mm, addr, ptep);
492 ncontig = find_num_contig(mm, addr, ptep, &pgsize);
493 dpfn = pgsize >> PAGE_SHIFT;
495 pte = get_clear_contig(mm, addr, ptep, pgsize, ncontig);
496 pte = pte_wrprotect(pte);
498 hugeprot = pte_pgprot(pte);
501 for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
502 set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
505 pte_t huge_ptep_clear_flush(struct vm_area_struct *vma,
506 unsigned long addr, pte_t *ptep)
512 if (!pte_cont(READ_ONCE(*ptep)))
513 return ptep_clear_flush(vma, addr, ptep);
515 ncontig = find_num_contig(vma->vm_mm, addr, ptep, &pgsize);
516 orig_pte = get_clear_contig(vma->vm_mm, addr, ptep, pgsize, ncontig);
517 flush_tlb_range(vma, addr, addr + pgsize * ncontig);
521 static int __init hugetlbpage_init(void)
523 if (pud_sect_supported())
524 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
526 hugetlb_add_hstate(CONT_PMD_SHIFT - PAGE_SHIFT);
527 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
528 hugetlb_add_hstate(CONT_PTE_SHIFT - PAGE_SHIFT);
532 arch_initcall(hugetlbpage_init);
534 bool __init arch_hugetlb_valid_size(unsigned long size)
536 return __hugetlb_valid_size(size);