1 // SPDX-License-Identifier: GPL-2.0
3 * IBM System z Huge TLB Page Support for Kernel.
5 * Copyright IBM Corp. 2007,2020
6 * Author(s): Gerald Schaefer <gerald.schaefer@de.ibm.com>
9 #define KMSG_COMPONENT "hugetlb"
10 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
13 #include <linux/hugetlb.h>
14 #include <linux/mman.h>
15 #include <linux/sched/mm.h>
16 #include <linux/security.h>
19 * If the bit selected by single-bit bitmask "a" is set within "x", move
20 * it to the position indicated by single-bit bitmask "b".
22 #define move_set_bit(x, a, b) (((x) & (a)) >> ilog2(a) << ilog2(b))
24 static inline unsigned long __pte_to_rste(pte_t pte)
29 * Convert encoding pte bits pmd / pud bits
30 * lIR.uswrdy.p dy..R...I...wr
31 * empty 010.000000.0 -> 00..0...1...00
32 * prot-none, clean, old 111.000000.1 -> 00..1...1...00
33 * prot-none, clean, young 111.000001.1 -> 01..1...1...00
34 * prot-none, dirty, old 111.000010.1 -> 10..1...1...00
35 * prot-none, dirty, young 111.000011.1 -> 11..1...1...00
36 * read-only, clean, old 111.000100.1 -> 00..1...1...01
37 * read-only, clean, young 101.000101.1 -> 01..1...0...01
38 * read-only, dirty, old 111.000110.1 -> 10..1...1...01
39 * read-only, dirty, young 101.000111.1 -> 11..1...0...01
40 * read-write, clean, old 111.001100.1 -> 00..1...1...11
41 * read-write, clean, young 101.001101.1 -> 01..1...0...11
42 * read-write, dirty, old 110.001110.1 -> 10..0...1...11
43 * read-write, dirty, young 100.001111.1 -> 11..0...0...11
44 * HW-bits: R read-only, I invalid
45 * SW-bits: p present, y young, d dirty, r read, w write, s special,
48 if (pte_present(pte)) {
49 rste = pte_val(pte) & PAGE_MASK;
50 rste |= move_set_bit(pte_val(pte), _PAGE_READ,
52 rste |= move_set_bit(pte_val(pte), _PAGE_WRITE,
53 _SEGMENT_ENTRY_WRITE);
54 rste |= move_set_bit(pte_val(pte), _PAGE_INVALID,
55 _SEGMENT_ENTRY_INVALID);
56 rste |= move_set_bit(pte_val(pte), _PAGE_PROTECT,
57 _SEGMENT_ENTRY_PROTECT);
58 rste |= move_set_bit(pte_val(pte), _PAGE_DIRTY,
59 _SEGMENT_ENTRY_DIRTY);
60 rste |= move_set_bit(pte_val(pte), _PAGE_YOUNG,
61 _SEGMENT_ENTRY_YOUNG);
62 #ifdef CONFIG_MEM_SOFT_DIRTY
63 rste |= move_set_bit(pte_val(pte), _PAGE_SOFT_DIRTY,
64 _SEGMENT_ENTRY_SOFT_DIRTY);
66 rste |= move_set_bit(pte_val(pte), _PAGE_NOEXEC,
67 _SEGMENT_ENTRY_NOEXEC);
69 rste = _SEGMENT_ENTRY_EMPTY;
73 static inline pte_t __rste_to_pte(unsigned long rste)
78 if ((rste & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
79 present = pud_present(__pud(rste));
81 present = pmd_present(__pmd(rste));
84 * Convert encoding pmd / pud bits pte bits
85 * dy..R...I...wr lIR.uswrdy.p
86 * empty 00..0...1...00 -> 010.000000.0
87 * prot-none, clean, old 00..1...1...00 -> 111.000000.1
88 * prot-none, clean, young 01..1...1...00 -> 111.000001.1
89 * prot-none, dirty, old 10..1...1...00 -> 111.000010.1
90 * prot-none, dirty, young 11..1...1...00 -> 111.000011.1
91 * read-only, clean, old 00..1...1...01 -> 111.000100.1
92 * read-only, clean, young 01..1...0...01 -> 101.000101.1
93 * read-only, dirty, old 10..1...1...01 -> 111.000110.1
94 * read-only, dirty, young 11..1...0...01 -> 101.000111.1
95 * read-write, clean, old 00..1...1...11 -> 111.001100.1
96 * read-write, clean, young 01..1...0...11 -> 101.001101.1
97 * read-write, dirty, old 10..0...1...11 -> 110.001110.1
98 * read-write, dirty, young 11..0...0...11 -> 100.001111.1
99 * HW-bits: R read-only, I invalid
100 * SW-bits: p present, y young, d dirty, r read, w write, s special,
104 pte_val(pte) = rste & _SEGMENT_ENTRY_ORIGIN_LARGE;
105 pte_val(pte) |= _PAGE_LARGE | _PAGE_PRESENT;
106 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_READ,
108 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_WRITE,
110 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_INVALID,
112 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_PROTECT,
114 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_DIRTY,
116 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_YOUNG,
118 #ifdef CONFIG_MEM_SOFT_DIRTY
119 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_SOFT_DIRTY,
122 pte_val(pte) |= move_set_bit(rste, _SEGMENT_ENTRY_NOEXEC,
125 pte_val(pte) = _PAGE_INVALID;
129 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
130 pte_t *ptep, pte_t pte)
134 rste = __pte_to_rste(pte);
136 rste &= ~_SEGMENT_ENTRY_NOEXEC;
138 /* Set correct table type for 2G hugepages */
139 if ((pte_val(*ptep) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
140 rste |= _REGION_ENTRY_TYPE_R3 | _REGION3_ENTRY_LARGE;
142 rste |= _SEGMENT_ENTRY_LARGE;
143 pte_val(*ptep) = rste;
146 pte_t huge_ptep_get(pte_t *ptep)
148 return __rste_to_pte(pte_val(*ptep));
151 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
152 unsigned long addr, pte_t *ptep)
154 pte_t pte = huge_ptep_get(ptep);
155 pmd_t *pmdp = (pmd_t *) ptep;
156 pud_t *pudp = (pud_t *) ptep;
158 if ((pte_val(*ptep) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
159 pudp_xchg_direct(mm, addr, pudp, __pud(_REGION3_ENTRY_EMPTY));
161 pmdp_xchg_direct(mm, addr, pmdp, __pmd(_SEGMENT_ENTRY_EMPTY));
165 pte_t *huge_pte_alloc(struct mm_struct *mm,
166 unsigned long addr, unsigned long sz)
173 pgdp = pgd_offset(mm, addr);
174 p4dp = p4d_alloc(mm, pgdp, addr);
176 pudp = pud_alloc(mm, p4dp, addr);
179 return (pte_t *) pudp;
180 else if (sz == PMD_SIZE)
181 pmdp = pmd_alloc(mm, pudp, addr);
184 return (pte_t *) pmdp;
187 pte_t *huge_pte_offset(struct mm_struct *mm,
188 unsigned long addr, unsigned long sz)
195 pgdp = pgd_offset(mm, addr);
196 if (pgd_present(*pgdp)) {
197 p4dp = p4d_offset(pgdp, addr);
198 if (p4d_present(*p4dp)) {
199 pudp = pud_offset(p4dp, addr);
200 if (pud_present(*pudp)) {
201 if (pud_large(*pudp))
202 return (pte_t *) pudp;
203 pmdp = pmd_offset(pudp, addr);
207 return (pte_t *) pmdp;
210 int pmd_huge(pmd_t pmd)
212 return pmd_large(pmd);
215 int pud_huge(pud_t pud)
217 return pud_large(pud);
221 follow_huge_pud(struct mm_struct *mm, unsigned long address,
222 pud_t *pud, int flags)
224 if (flags & FOLL_GET)
227 return pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
230 static __init int setup_hugepagesz(char *opt)
235 size = memparse(opt, &opt);
236 if (MACHINE_HAS_EDAT1 && size == PMD_SIZE) {
237 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
238 } else if (MACHINE_HAS_EDAT2 && size == PUD_SIZE) {
239 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
242 pr_err("hugepagesz= specifies an unsupported page size %s\n",
248 __setup("hugepagesz=", setup_hugepagesz);
250 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
251 unsigned long addr, unsigned long len,
252 unsigned long pgoff, unsigned long flags)
254 struct hstate *h = hstate_file(file);
255 struct vm_unmapped_area_info info;
259 info.low_limit = current->mm->mmap_base;
260 info.high_limit = TASK_SIZE;
261 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
262 info.align_offset = 0;
263 return vm_unmapped_area(&info);
266 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
267 unsigned long addr0, unsigned long len,
268 unsigned long pgoff, unsigned long flags)
270 struct hstate *h = hstate_file(file);
271 struct vm_unmapped_area_info info;
274 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
276 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
277 info.high_limit = current->mm->mmap_base;
278 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
279 info.align_offset = 0;
280 addr = vm_unmapped_area(&info);
283 * A failed mmap() very likely causes application failure,
284 * so fall back to the bottom-up function here. This scenario
285 * can happen with large stack limits and large mmap()
288 if (addr & ~PAGE_MASK) {
289 VM_BUG_ON(addr != -ENOMEM);
291 info.low_limit = TASK_UNMAPPED_BASE;
292 info.high_limit = TASK_SIZE;
293 addr = vm_unmapped_area(&info);
299 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
300 unsigned long len, unsigned long pgoff, unsigned long flags)
302 struct hstate *h = hstate_file(file);
303 struct mm_struct *mm = current->mm;
304 struct vm_area_struct *vma;
307 if (len & ~huge_page_mask(h))
309 if (len > TASK_SIZE - mmap_min_addr)
312 if (flags & MAP_FIXED) {
313 if (prepare_hugepage_range(file, addr, len))
315 goto check_asce_limit;
319 addr = ALIGN(addr, huge_page_size(h));
320 vma = find_vma(mm, addr);
321 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
322 (!vma || addr + len <= vm_start_gap(vma)))
323 goto check_asce_limit;
326 if (mm->get_unmapped_area == arch_get_unmapped_area)
327 addr = hugetlb_get_unmapped_area_bottomup(file, addr, len,
330 addr = hugetlb_get_unmapped_area_topdown(file, addr, len,
332 if (addr & ~PAGE_MASK)
336 if (addr + len > current->mm->context.asce_limit &&
337 addr + len <= TASK_SIZE) {
338 rc = crst_table_upgrade(mm, addr + len);
340 return (unsigned long) rc;