tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1/*
2 * PPC Huge TLB Page Support for Kernel.
3 *
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
6 *
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/io.h>
13#include <linux/slab.h>
14#include <linux/hugetlb.h>
15#include <linux/export.h>
16#include <linux/of_fdt.h>
17#include <linux/memblock.h>
18#include <linux/moduleparam.h>
19#include <linux/swap.h>
20#include <linux/swapops.h>
21#include <linux/kmemleak.h>
22#include <asm/pgalloc.h>
23#include <asm/tlb.h>
24#include <asm/setup.h>
25#include <asm/hugetlb.h>
26#include <asm/pte-walk.h>
27#include <asm/firmware.h>
28
29bool hugetlb_disabled = false;
30
31#define PTE_T_ORDER (__builtin_ffs(sizeof(pte_basic_t)) - \
32 __builtin_ffs(sizeof(void *)))
33
34pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
35{
36 /*
37 * Only called for hugetlbfs pages, hence can ignore THP and the
38 * irq disabled walk.
39 */
40 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
41}
42
43pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
44 unsigned long addr, unsigned long sz)
45{
46 p4d_t *p4d;
47 pud_t *pud;
48 pmd_t *pmd;
49
50 addr &= ~(sz - 1);
51
52 p4d = p4d_offset(pgd_offset(mm, addr), addr);
53 if (!mm_pud_folded(mm) && sz >= P4D_SIZE)
54 return (pte_t *)p4d;
55
56 pud = pud_alloc(mm, p4d, addr);
57 if (!pud)
58 return NULL;
59 if (!mm_pmd_folded(mm) && sz >= PUD_SIZE)
60 return (pte_t *)pud;
61
62 pmd = pmd_alloc(mm, pud, addr);
63 if (!pmd)
64 return NULL;
65
66 if (sz >= PMD_SIZE) {
67 /* On 8xx, all hugepages are handled as contiguous PTEs */
68 if (IS_ENABLED(CONFIG_PPC_8xx)) {
69 int i;
70
71 for (i = 0; i < sz / PMD_SIZE; i++) {
72 if (!pte_alloc_huge(mm, pmd + i, addr))
73 return NULL;
74 }
75 }
76 return (pte_t *)pmd;
77 }
78
79 return pte_alloc_huge(mm, pmd, addr);
80}
81
82#ifdef CONFIG_PPC_BOOK3S_64
83/*
84 * Tracks gpages after the device tree is scanned and before the
85 * huge_boot_pages list is ready on pseries.
86 */
87#define MAX_NUMBER_GPAGES 1024
88__initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
89__initdata static unsigned nr_gpages;
90
91/*
92 * Build list of addresses of gigantic pages. This function is used in early
93 * boot before the buddy allocator is setup.
94 */
95void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
96{
97 if (!addr)
98 return;
99 while (number_of_pages > 0) {
100 gpage_freearray[nr_gpages] = addr;
101 nr_gpages++;
102 number_of_pages--;
103 addr += page_size;
104 }
105}
106
107static int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
108{
109 struct huge_bootmem_page *m;
110 if (nr_gpages == 0)
111 return 0;
112 m = phys_to_virt(gpage_freearray[--nr_gpages]);
113 gpage_freearray[nr_gpages] = 0;
114 list_add(&m->list, &huge_boot_pages[0]);
115 m->hstate = hstate;
116 return 1;
117}
118
119bool __init hugetlb_node_alloc_supported(void)
120{
121 return false;
122}
123#endif
124
125
126int __init alloc_bootmem_huge_page(struct hstate *h, int nid)
127{
128
129#ifdef CONFIG_PPC_BOOK3S_64
130 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
131 return pseries_alloc_bootmem_huge_page(h);
132#endif
133 return __alloc_bootmem_huge_page(h, nid);
134}
135
136bool __init arch_hugetlb_valid_size(unsigned long size)
137{
138 int shift = __ffs(size);
139 int mmu_psize;
140
141 /* Check that it is a page size supported by the hardware and
142 * that it fits within pagetable and slice limits. */
143 if (size <= PAGE_SIZE || !is_power_of_2(size))
144 return false;
145
146 mmu_psize = check_and_get_huge_psize(shift);
147 if (mmu_psize < 0)
148 return false;
149
150 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
151
152 return true;
153}
154
155static int __init add_huge_page_size(unsigned long long size)
156{
157 int shift = __ffs(size);
158
159 if (!arch_hugetlb_valid_size((unsigned long)size))
160 return -EINVAL;
161
162 hugetlb_add_hstate(shift - PAGE_SHIFT);
163 return 0;
164}
165
166static int __init hugetlbpage_init(void)
167{
168 bool configured = false;
169 int psize;
170
171 if (hugetlb_disabled) {
172 pr_info("HugeTLB support is disabled!\n");
173 return 0;
174 }
175
176 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
177 !mmu_has_feature(MMU_FTR_16M_PAGE))
178 return -ENODEV;
179
180 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
181 unsigned shift;
182
183 if (!mmu_psize_defs[psize].shift)
184 continue;
185
186 shift = mmu_psize_to_shift(psize);
187
188 if (add_huge_page_size(1ULL << shift) < 0)
189 continue;
190
191 configured = true;
192 }
193
194 if (!configured)
195 pr_info("Failed to initialize. Disabling HugeTLB");
196
197 return 0;
198}
199
200arch_initcall(hugetlbpage_init);
201
202void __init gigantic_hugetlb_cma_reserve(void)
203{
204 unsigned long order = 0;
205
206 if (radix_enabled())
207 order = PUD_SHIFT - PAGE_SHIFT;
208 else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
209 /*
210 * For pseries we do use ibm,expected#pages for reserving 16G pages.
211 */
212 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
213
214 if (order)
215 hugetlb_cma_reserve(order);
216}