Loading...
1/*
2 * arch/arm/include/asm/tlb.h
3 *
4 * Copyright (C) 2002 Russell King
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
9 *
10 * Experimentation shows that on a StrongARM, it appears to be faster
11 * to use the "invalidate whole tlb" rather than "invalidate single
12 * tlb" for this.
13 *
14 * This appears true for both the process fork+exit case, as well as
15 * the munmap-large-area case.
16 */
17#ifndef __ASMARM_TLB_H
18#define __ASMARM_TLB_H
19
20#include <asm/cacheflush.h>
21
22#ifndef CONFIG_MMU
23
24#include <linux/pagemap.h>
25
26#define tlb_flush(tlb) ((void) tlb)
27
28#include <asm-generic/tlb.h>
29
30#else /* !CONFIG_MMU */
31
32#include <linux/swap.h>
33#include <asm/pgalloc.h>
34#include <asm/tlbflush.h>
35
36#define MMU_GATHER_BUNDLE 8
37
38#ifdef CONFIG_HAVE_RCU_TABLE_FREE
39static inline void __tlb_remove_table(void *_table)
40{
41 free_page_and_swap_cache((struct page *)_table);
42}
43
44struct mmu_table_batch {
45 struct rcu_head rcu;
46 unsigned int nr;
47 void *tables[0];
48};
49
50#define MAX_TABLE_BATCH \
51 ((PAGE_SIZE - sizeof(struct mmu_table_batch)) / sizeof(void *))
52
53extern void tlb_table_flush(struct mmu_gather *tlb);
54extern void tlb_remove_table(struct mmu_gather *tlb, void *table);
55
56#define tlb_remove_entry(tlb, entry) tlb_remove_table(tlb, entry)
57#else
58#define tlb_remove_entry(tlb, entry) tlb_remove_page(tlb, entry)
59#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
60
61/*
62 * TLB handling. This allows us to remove pages from the page
63 * tables, and efficiently handle the TLB issues.
64 */
65struct mmu_gather {
66 struct mm_struct *mm;
67#ifdef CONFIG_HAVE_RCU_TABLE_FREE
68 struct mmu_table_batch *batch;
69 unsigned int need_flush;
70#endif
71 unsigned int fullmm;
72 struct vm_area_struct *vma;
73 unsigned long start, end;
74 unsigned long range_start;
75 unsigned long range_end;
76 unsigned int nr;
77 unsigned int max;
78 struct page **pages;
79 struct page *local[MMU_GATHER_BUNDLE];
80};
81
82DECLARE_PER_CPU(struct mmu_gather, mmu_gathers);
83
84/*
85 * This is unnecessarily complex. There's three ways the TLB shootdown
86 * code is used:
87 * 1. Unmapping a range of vmas. See zap_page_range(), unmap_region().
88 * tlb->fullmm = 0, and tlb_start_vma/tlb_end_vma will be called.
89 * tlb->vma will be non-NULL.
90 * 2. Unmapping all vmas. See exit_mmap().
91 * tlb->fullmm = 1, and tlb_start_vma/tlb_end_vma will be called.
92 * tlb->vma will be non-NULL. Additionally, page tables will be freed.
93 * 3. Unmapping argument pages. See shift_arg_pages().
94 * tlb->fullmm = 0, but tlb_start_vma/tlb_end_vma will not be called.
95 * tlb->vma will be NULL.
96 */
97static inline void tlb_flush(struct mmu_gather *tlb)
98{
99 if (tlb->fullmm || !tlb->vma)
100 flush_tlb_mm(tlb->mm);
101 else if (tlb->range_end > 0) {
102 flush_tlb_range(tlb->vma, tlb->range_start, tlb->range_end);
103 tlb->range_start = TASK_SIZE;
104 tlb->range_end = 0;
105 }
106}
107
108static inline void tlb_add_flush(struct mmu_gather *tlb, unsigned long addr)
109{
110 if (!tlb->fullmm) {
111 if (addr < tlb->range_start)
112 tlb->range_start = addr;
113 if (addr + PAGE_SIZE > tlb->range_end)
114 tlb->range_end = addr + PAGE_SIZE;
115 }
116}
117
118static inline void __tlb_alloc_page(struct mmu_gather *tlb)
119{
120 unsigned long addr = __get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
121
122 if (addr) {
123 tlb->pages = (void *)addr;
124 tlb->max = PAGE_SIZE / sizeof(struct page *);
125 }
126}
127
128static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
129{
130 tlb_flush(tlb);
131#ifdef CONFIG_HAVE_RCU_TABLE_FREE
132 tlb_table_flush(tlb);
133#endif
134}
135
136static inline void tlb_flush_mmu_free(struct mmu_gather *tlb)
137{
138 free_pages_and_swap_cache(tlb->pages, tlb->nr);
139 tlb->nr = 0;
140 if (tlb->pages == tlb->local)
141 __tlb_alloc_page(tlb);
142}
143
144static inline void tlb_flush_mmu(struct mmu_gather *tlb)
145{
146 tlb_flush_mmu_tlbonly(tlb);
147 tlb_flush_mmu_free(tlb);
148}
149
150static inline void
151tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
152{
153 tlb->mm = mm;
154 tlb->fullmm = !(start | (end+1));
155 tlb->start = start;
156 tlb->end = end;
157 tlb->vma = NULL;
158 tlb->max = ARRAY_SIZE(tlb->local);
159 tlb->pages = tlb->local;
160 tlb->nr = 0;
161 __tlb_alloc_page(tlb);
162
163#ifdef CONFIG_HAVE_RCU_TABLE_FREE
164 tlb->batch = NULL;
165#endif
166}
167
168static inline void
169tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
170{
171 tlb_flush_mmu(tlb);
172
173 /* keep the page table cache within bounds */
174 check_pgt_cache();
175
176 if (tlb->pages != tlb->local)
177 free_pages((unsigned long)tlb->pages, 0);
178}
179
180/*
181 * Memorize the range for the TLB flush.
182 */
183static inline void
184tlb_remove_tlb_entry(struct mmu_gather *tlb, pte_t *ptep, unsigned long addr)
185{
186 tlb_add_flush(tlb, addr);
187}
188
189#define tlb_remove_huge_tlb_entry(h, tlb, ptep, address) \
190 tlb_remove_tlb_entry(tlb, ptep, address)
191/*
192 * In the case of tlb vma handling, we can optimise these away in the
193 * case where we're doing a full MM flush. When we're doing a munmap,
194 * the vmas are adjusted to only cover the region to be torn down.
195 */
196static inline void
197tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
198{
199 if (!tlb->fullmm) {
200 flush_cache_range(vma, vma->vm_start, vma->vm_end);
201 tlb->vma = vma;
202 tlb->range_start = TASK_SIZE;
203 tlb->range_end = 0;
204 }
205}
206
207static inline void
208tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
209{
210 if (!tlb->fullmm)
211 tlb_flush(tlb);
212}
213
214static inline bool __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
215{
216 tlb->pages[tlb->nr++] = page;
217 VM_WARN_ON(tlb->nr > tlb->max);
218 if (tlb->nr == tlb->max)
219 return true;
220 return false;
221}
222
223static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
224{
225 if (__tlb_remove_page(tlb, page))
226 tlb_flush_mmu(tlb);
227}
228
229static inline bool __tlb_remove_page_size(struct mmu_gather *tlb,
230 struct page *page, int page_size)
231{
232 return __tlb_remove_page(tlb, page);
233}
234
235static inline void tlb_remove_page_size(struct mmu_gather *tlb,
236 struct page *page, int page_size)
237{
238 return tlb_remove_page(tlb, page);
239}
240
241static inline void __pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
242 unsigned long addr)
243{
244 pgtable_page_dtor(pte);
245
246#ifdef CONFIG_ARM_LPAE
247 tlb_add_flush(tlb, addr);
248#else
249 /*
250 * With the classic ARM MMU, a pte page has two corresponding pmd
251 * entries, each covering 1MB.
252 */
253 addr &= PMD_MASK;
254 tlb_add_flush(tlb, addr + SZ_1M - PAGE_SIZE);
255 tlb_add_flush(tlb, addr + SZ_1M);
256#endif
257
258 tlb_remove_entry(tlb, pte);
259}
260
261static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
262 unsigned long addr)
263{
264#ifdef CONFIG_ARM_LPAE
265 tlb_add_flush(tlb, addr);
266 tlb_remove_entry(tlb, virt_to_page(pmdp));
267#endif
268}
269
270static inline void
271tlb_remove_pmd_tlb_entry(struct mmu_gather *tlb, pmd_t *pmdp, unsigned long addr)
272{
273 tlb_add_flush(tlb, addr);
274}
275
276#define pte_free_tlb(tlb, ptep, addr) __pte_free_tlb(tlb, ptep, addr)
277#define pmd_free_tlb(tlb, pmdp, addr) __pmd_free_tlb(tlb, pmdp, addr)
278#define pud_free_tlb(tlb, pudp, addr) pud_free((tlb)->mm, pudp)
279
280#define tlb_migrate_finish(mm) do { } while (0)
281
282#define tlb_remove_check_page_size_change tlb_remove_check_page_size_change
283static inline void tlb_remove_check_page_size_change(struct mmu_gather *tlb,
284 unsigned int page_size)
285{
286}
287
288#endif /* CONFIG_MMU */
289#endif
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * arch/arm/include/asm/tlb.h
4 *
5 * Copyright (C) 2002 Russell King
6 *
7 * Experimentation shows that on a StrongARM, it appears to be faster
8 * to use the "invalidate whole tlb" rather than "invalidate single
9 * tlb" for this.
10 *
11 * This appears true for both the process fork+exit case, as well as
12 * the munmap-large-area case.
13 */
14#ifndef __ASMARM_TLB_H
15#define __ASMARM_TLB_H
16
17#include <asm/cacheflush.h>
18
19#ifndef CONFIG_MMU
20
21#include <linux/pagemap.h>
22
23#define tlb_flush(tlb) ((void) tlb)
24
25#include <asm-generic/tlb.h>
26
27#else /* !CONFIG_MMU */
28
29#include <linux/swap.h>
30#include <asm/pgalloc.h>
31#include <asm/tlbflush.h>
32
33static inline void __tlb_remove_table(void *_table)
34{
35 free_page_and_swap_cache((struct page *)_table);
36}
37
38#include <asm-generic/tlb.h>
39
40#ifndef CONFIG_HAVE_RCU_TABLE_FREE
41#define tlb_remove_table(tlb, entry) tlb_remove_page(tlb, entry)
42#endif
43
44static inline void
45__pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte, unsigned long addr)
46{
47 pgtable_pte_page_dtor(pte);
48
49#ifndef CONFIG_ARM_LPAE
50 /*
51 * With the classic ARM MMU, a pte page has two corresponding pmd
52 * entries, each covering 1MB.
53 */
54 addr = (addr & PMD_MASK) + SZ_1M;
55 __tlb_adjust_range(tlb, addr - PAGE_SIZE, 2 * PAGE_SIZE);
56#endif
57
58 tlb_remove_table(tlb, pte);
59}
60
61static inline void
62__pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp, unsigned long addr)
63{
64#ifdef CONFIG_ARM_LPAE
65 struct page *page = virt_to_page(pmdp);
66
67 tlb_remove_table(tlb, page);
68#endif
69}
70
71#endif /* CONFIG_MMU */
72#endif