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/*
39 * TLB handling. This allows us to remove pages from the page
40 * tables, and efficiently handle the TLB issues.
41 */
42struct mmu_gather {
43 struct mm_struct *mm;
44 unsigned int fullmm;
45 struct vm_area_struct *vma;
46 unsigned long start, end;
47 unsigned long range_start;
48 unsigned long range_end;
49 unsigned int nr;
50 unsigned int max;
51 struct page **pages;
52 struct page *local[MMU_GATHER_BUNDLE];
53};
54
55DECLARE_PER_CPU(struct mmu_gather, mmu_gathers);
56
57/*
58 * This is unnecessarily complex. There's three ways the TLB shootdown
59 * code is used:
60 * 1. Unmapping a range of vmas. See zap_page_range(), unmap_region().
61 * tlb->fullmm = 0, and tlb_start_vma/tlb_end_vma will be called.
62 * tlb->vma will be non-NULL.
63 * 2. Unmapping all vmas. See exit_mmap().
64 * tlb->fullmm = 1, and tlb_start_vma/tlb_end_vma will be called.
65 * tlb->vma will be non-NULL. Additionally, page tables will be freed.
66 * 3. Unmapping argument pages. See shift_arg_pages().
67 * tlb->fullmm = 0, but tlb_start_vma/tlb_end_vma will not be called.
68 * tlb->vma will be NULL.
69 */
70static inline void tlb_flush(struct mmu_gather *tlb)
71{
72 if (tlb->fullmm || !tlb->vma)
73 flush_tlb_mm(tlb->mm);
74 else if (tlb->range_end > 0) {
75 flush_tlb_range(tlb->vma, tlb->range_start, tlb->range_end);
76 tlb->range_start = TASK_SIZE;
77 tlb->range_end = 0;
78 }
79}
80
81static inline void tlb_add_flush(struct mmu_gather *tlb, unsigned long addr)
82{
83 if (!tlb->fullmm) {
84 if (addr < tlb->range_start)
85 tlb->range_start = addr;
86 if (addr + PAGE_SIZE > tlb->range_end)
87 tlb->range_end = addr + PAGE_SIZE;
88 }
89}
90
91static inline void __tlb_alloc_page(struct mmu_gather *tlb)
92{
93 unsigned long addr = __get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
94
95 if (addr) {
96 tlb->pages = (void *)addr;
97 tlb->max = PAGE_SIZE / sizeof(struct page *);
98 }
99}
100
101static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
102{
103 tlb_flush(tlb);
104}
105
106static inline void tlb_flush_mmu_free(struct mmu_gather *tlb)
107{
108 free_pages_and_swap_cache(tlb->pages, tlb->nr);
109 tlb->nr = 0;
110 if (tlb->pages == tlb->local)
111 __tlb_alloc_page(tlb);
112}
113
114static inline void tlb_flush_mmu(struct mmu_gather *tlb)
115{
116 tlb_flush_mmu_tlbonly(tlb);
117 tlb_flush_mmu_free(tlb);
118}
119
120static inline void
121tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
122{
123 tlb->mm = mm;
124 tlb->fullmm = !(start | (end+1));
125 tlb->start = start;
126 tlb->end = end;
127 tlb->vma = NULL;
128 tlb->max = ARRAY_SIZE(tlb->local);
129 tlb->pages = tlb->local;
130 tlb->nr = 0;
131 __tlb_alloc_page(tlb);
132}
133
134static inline void
135tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
136{
137 tlb_flush_mmu(tlb);
138
139 /* keep the page table cache within bounds */
140 check_pgt_cache();
141
142 if (tlb->pages != tlb->local)
143 free_pages((unsigned long)tlb->pages, 0);
144}
145
146/*
147 * Memorize the range for the TLB flush.
148 */
149static inline void
150tlb_remove_tlb_entry(struct mmu_gather *tlb, pte_t *ptep, unsigned long addr)
151{
152 tlb_add_flush(tlb, addr);
153}
154
155/*
156 * In the case of tlb vma handling, we can optimise these away in the
157 * case where we're doing a full MM flush. When we're doing a munmap,
158 * the vmas are adjusted to only cover the region to be torn down.
159 */
160static inline void
161tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
162{
163 if (!tlb->fullmm) {
164 flush_cache_range(vma, vma->vm_start, vma->vm_end);
165 tlb->vma = vma;
166 tlb->range_start = TASK_SIZE;
167 tlb->range_end = 0;
168 }
169}
170
171static inline void
172tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
173{
174 if (!tlb->fullmm)
175 tlb_flush(tlb);
176}
177
178static inline int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
179{
180 tlb->pages[tlb->nr++] = page;
181 VM_BUG_ON(tlb->nr > tlb->max);
182 return tlb->max - tlb->nr;
183}
184
185static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
186{
187 if (!__tlb_remove_page(tlb, page))
188 tlb_flush_mmu(tlb);
189}
190
191static inline void __pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
192 unsigned long addr)
193{
194 pgtable_page_dtor(pte);
195
196#ifdef CONFIG_ARM_LPAE
197 tlb_add_flush(tlb, addr);
198#else
199 /*
200 * With the classic ARM MMU, a pte page has two corresponding pmd
201 * entries, each covering 1MB.
202 */
203 addr &= PMD_MASK;
204 tlb_add_flush(tlb, addr + SZ_1M - PAGE_SIZE);
205 tlb_add_flush(tlb, addr + SZ_1M);
206#endif
207
208 tlb_remove_page(tlb, pte);
209}
210
211static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
212 unsigned long addr)
213{
214#ifdef CONFIG_ARM_LPAE
215 tlb_add_flush(tlb, addr);
216 tlb_remove_page(tlb, virt_to_page(pmdp));
217#endif
218}
219
220static inline void
221tlb_remove_pmd_tlb_entry(struct mmu_gather *tlb, pmd_t *pmdp, unsigned long addr)
222{
223 tlb_add_flush(tlb, addr);
224}
225
226#define pte_free_tlb(tlb, ptep, addr) __pte_free_tlb(tlb, ptep, addr)
227#define pmd_free_tlb(tlb, pmdp, addr) __pmd_free_tlb(tlb, pmdp, addr)
228#define pud_free_tlb(tlb, pudp, addr) pud_free((tlb)->mm, pudp)
229
230#define tlb_migrate_finish(mm) do { } while (0)
231
232#endif /* CONFIG_MMU */
233#endif
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/*
190 * In the case of tlb vma handling, we can optimise these away in the
191 * case where we're doing a full MM flush. When we're doing a munmap,
192 * the vmas are adjusted to only cover the region to be torn down.
193 */
194static inline void
195tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
196{
197 if (!tlb->fullmm) {
198 flush_cache_range(vma, vma->vm_start, vma->vm_end);
199 tlb->vma = vma;
200 tlb->range_start = TASK_SIZE;
201 tlb->range_end = 0;
202 }
203}
204
205static inline void
206tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
207{
208 if (!tlb->fullmm)
209 tlb_flush(tlb);
210}
211
212static inline int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
213{
214 tlb->pages[tlb->nr++] = page;
215 VM_BUG_ON(tlb->nr > tlb->max);
216 return tlb->max - tlb->nr;
217}
218
219static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
220{
221 if (!__tlb_remove_page(tlb, page))
222 tlb_flush_mmu(tlb);
223}
224
225static inline void __pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
226 unsigned long addr)
227{
228 pgtable_page_dtor(pte);
229
230#ifdef CONFIG_ARM_LPAE
231 tlb_add_flush(tlb, addr);
232#else
233 /*
234 * With the classic ARM MMU, a pte page has two corresponding pmd
235 * entries, each covering 1MB.
236 */
237 addr &= PMD_MASK;
238 tlb_add_flush(tlb, addr + SZ_1M - PAGE_SIZE);
239 tlb_add_flush(tlb, addr + SZ_1M);
240#endif
241
242 tlb_remove_entry(tlb, pte);
243}
244
245static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
246 unsigned long addr)
247{
248#ifdef CONFIG_ARM_LPAE
249 tlb_add_flush(tlb, addr);
250 tlb_remove_entry(tlb, virt_to_page(pmdp));
251#endif
252}
253
254static inline void
255tlb_remove_pmd_tlb_entry(struct mmu_gather *tlb, pmd_t *pmdp, unsigned long addr)
256{
257 tlb_add_flush(tlb, addr);
258}
259
260#define pte_free_tlb(tlb, ptep, addr) __pte_free_tlb(tlb, ptep, addr)
261#define pmd_free_tlb(tlb, pmdp, addr) __pmd_free_tlb(tlb, pmdp, addr)
262#define pud_free_tlb(tlb, pudp, addr) pud_free((tlb)->mm, pudp)
263
264#define tlb_migrate_finish(mm) do { } while (0)
265
266#endif /* CONFIG_MMU */
267#endif