xref: /openbmc/linux/arch/x86/mm/hugetlbpage.c (revision 1b39eacd)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * IA-32 Huge TLB Page Support for Kernel.
4  *
5  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
6  */
7 
8 #include <linux/init.h>
9 #include <linux/fs.h>
10 #include <linux/mm.h>
11 #include <linux/sched/mm.h>
12 #include <linux/hugetlb.h>
13 #include <linux/pagemap.h>
14 #include <linux/err.h>
15 #include <linux/sysctl.h>
16 #include <linux/compat.h>
17 #include <asm/mman.h>
18 #include <asm/tlb.h>
19 #include <asm/tlbflush.h>
20 #include <asm/pgalloc.h>
21 #include <asm/elf.h>
22 #include <asm/mpx.h>
23 
24 #if 0	/* This is just for testing */
25 struct page *
26 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
27 {
28 	unsigned long start = address;
29 	int length = 1;
30 	int nr;
31 	struct page *page;
32 	struct vm_area_struct *vma;
33 
34 	vma = find_vma(mm, addr);
35 	if (!vma || !is_vm_hugetlb_page(vma))
36 		return ERR_PTR(-EINVAL);
37 
38 	pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma));
39 
40 	/* hugetlb should be locked, and hence, prefaulted */
41 	WARN_ON(!pte || pte_none(*pte));
42 
43 	page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
44 
45 	WARN_ON(!PageHead(page));
46 
47 	return page;
48 }
49 
50 int pmd_huge(pmd_t pmd)
51 {
52 	return 0;
53 }
54 
55 int pud_huge(pud_t pud)
56 {
57 	return 0;
58 }
59 
60 #else
61 
62 /*
63  * pmd_huge() returns 1 if @pmd is hugetlb related entry, that is normal
64  * hugetlb entry or non-present (migration or hwpoisoned) hugetlb entry.
65  * Otherwise, returns 0.
66  */
67 int pmd_huge(pmd_t pmd)
68 {
69 	return !pmd_none(pmd) &&
70 		(pmd_val(pmd) & (_PAGE_PRESENT|_PAGE_PSE)) != _PAGE_PRESENT;
71 }
72 
73 int pud_huge(pud_t pud)
74 {
75 	return !!(pud_val(pud) & _PAGE_PSE);
76 }
77 #endif
78 
79 #ifdef CONFIG_HUGETLB_PAGE
80 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
81 		unsigned long addr, unsigned long len,
82 		unsigned long pgoff, unsigned long flags)
83 {
84 	struct hstate *h = hstate_file(file);
85 	struct vm_unmapped_area_info info;
86 
87 	info.flags = 0;
88 	info.length = len;
89 	info.low_limit = get_mmap_base(1);
90 
91 	/*
92 	 * If hint address is above DEFAULT_MAP_WINDOW, look for unmapped area
93 	 * in the full address space.
94 	 */
95 	info.high_limit = in_compat_syscall() ?
96 		task_size_32bit() : task_size_64bit(addr > DEFAULT_MAP_WINDOW);
97 
98 	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
99 	info.align_offset = 0;
100 	return vm_unmapped_area(&info);
101 }
102 
103 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
104 		unsigned long addr, unsigned long len,
105 		unsigned long pgoff, unsigned long flags)
106 {
107 	struct hstate *h = hstate_file(file);
108 	struct vm_unmapped_area_info info;
109 
110 	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
111 	info.length = len;
112 	info.low_limit = PAGE_SIZE;
113 	info.high_limit = get_mmap_base(0);
114 
115 	/*
116 	 * If hint address is above DEFAULT_MAP_WINDOW, look for unmapped area
117 	 * in the full address space.
118 	 */
119 	if (addr > DEFAULT_MAP_WINDOW && !in_compat_syscall())
120 		info.high_limit += TASK_SIZE_MAX - DEFAULT_MAP_WINDOW;
121 
122 	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
123 	info.align_offset = 0;
124 	addr = vm_unmapped_area(&info);
125 
126 	/*
127 	 * A failed mmap() very likely causes application failure,
128 	 * so fall back to the bottom-up function here. This scenario
129 	 * can happen with large stack limits and large mmap()
130 	 * allocations.
131 	 */
132 	if (addr & ~PAGE_MASK) {
133 		VM_BUG_ON(addr != -ENOMEM);
134 		info.flags = 0;
135 		info.low_limit = TASK_UNMAPPED_BASE;
136 		info.high_limit = TASK_SIZE_LOW;
137 		addr = vm_unmapped_area(&info);
138 	}
139 
140 	return addr;
141 }
142 
143 unsigned long
144 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
145 		unsigned long len, unsigned long pgoff, unsigned long flags)
146 {
147 	struct hstate *h = hstate_file(file);
148 	struct mm_struct *mm = current->mm;
149 	struct vm_area_struct *vma;
150 
151 	if (len & ~huge_page_mask(h))
152 		return -EINVAL;
153 
154 	addr = mpx_unmapped_area_check(addr, len, flags);
155 	if (IS_ERR_VALUE(addr))
156 		return addr;
157 
158 	if (len > TASK_SIZE)
159 		return -ENOMEM;
160 
161 	/* No address checking. See comment at mmap_address_hint_valid() */
162 	if (flags & MAP_FIXED) {
163 		if (prepare_hugepage_range(file, addr, len))
164 			return -EINVAL;
165 		return addr;
166 	}
167 
168 	if (addr) {
169 		addr &= huge_page_mask(h);
170 		if (!mmap_address_hint_valid(addr, len))
171 			goto get_unmapped_area;
172 
173 		vma = find_vma(mm, addr);
174 		if (!vma || addr + len <= vm_start_gap(vma))
175 			return addr;
176 	}
177 
178 get_unmapped_area:
179 	if (mm->get_unmapped_area == arch_get_unmapped_area)
180 		return hugetlb_get_unmapped_area_bottomup(file, addr, len,
181 				pgoff, flags);
182 	else
183 		return hugetlb_get_unmapped_area_topdown(file, addr, len,
184 				pgoff, flags);
185 }
186 #endif /* CONFIG_HUGETLB_PAGE */
187 
188 #ifdef CONFIG_X86_64
189 static __init int setup_hugepagesz(char *opt)
190 {
191 	unsigned long ps = memparse(opt, &opt);
192 	if (ps == PMD_SIZE) {
193 		hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
194 	} else if (ps == PUD_SIZE && boot_cpu_has(X86_FEATURE_GBPAGES)) {
195 		hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
196 	} else {
197 		hugetlb_bad_size();
198 		printk(KERN_ERR "hugepagesz: Unsupported page size %lu M\n",
199 			ps >> 20);
200 		return 0;
201 	}
202 	return 1;
203 }
204 __setup("hugepagesz=", setup_hugepagesz);
205 
206 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
207 static __init int gigantic_pages_init(void)
208 {
209 	/* With compaction or CMA we can allocate gigantic pages at runtime */
210 	if (boot_cpu_has(X86_FEATURE_GBPAGES) && !size_to_hstate(1UL << PUD_SHIFT))
211 		hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
212 	return 0;
213 }
214 arch_initcall(gigantic_pages_init);
215 #endif
216 #endif
217