xref: /openbmc/linux/include/asm-generic/tlb.h (revision 82e6fdd6)
1 /* include/asm-generic/tlb.h
2  *
3  *	Generic TLB shootdown code
4  *
5  * Copyright 2001 Red Hat, Inc.
6  * Based on code from mm/memory.c Copyright Linus Torvalds and others.
7  *
8  * Copyright 2011 Red Hat, Inc., Peter Zijlstra
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public License
12  * as published by the Free Software Foundation; either version
13  * 2 of the License, or (at your option) any later version.
14  */
15 #ifndef _ASM_GENERIC__TLB_H
16 #define _ASM_GENERIC__TLB_H
17 
18 #include <linux/swap.h>
19 #include <asm/pgalloc.h>
20 #include <asm/tlbflush.h>
21 
22 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
23 /*
24  * Semi RCU freeing of the page directories.
25  *
26  * This is needed by some architectures to implement software pagetable walkers.
27  *
28  * gup_fast() and other software pagetable walkers do a lockless page-table
29  * walk and therefore needs some synchronization with the freeing of the page
30  * directories. The chosen means to accomplish that is by disabling IRQs over
31  * the walk.
32  *
33  * Architectures that use IPIs to flush TLBs will then automagically DTRT,
34  * since we unlink the page, flush TLBs, free the page. Since the disabling of
35  * IRQs delays the completion of the TLB flush we can never observe an already
36  * freed page.
37  *
38  * Architectures that do not have this (PPC) need to delay the freeing by some
39  * other means, this is that means.
40  *
41  * What we do is batch the freed directory pages (tables) and RCU free them.
42  * We use the sched RCU variant, as that guarantees that IRQ/preempt disabling
43  * holds off grace periods.
44  *
45  * However, in order to batch these pages we need to allocate storage, this
46  * allocation is deep inside the MM code and can thus easily fail on memory
47  * pressure. To guarantee progress we fall back to single table freeing, see
48  * the implementation of tlb_remove_table_one().
49  *
50  */
51 struct mmu_table_batch {
52 	struct rcu_head		rcu;
53 	unsigned int		nr;
54 	void			*tables[0];
55 };
56 
57 #define MAX_TABLE_BATCH		\
58 	((PAGE_SIZE - sizeof(struct mmu_table_batch)) / sizeof(void *))
59 
60 extern void tlb_table_flush(struct mmu_gather *tlb);
61 extern void tlb_remove_table(struct mmu_gather *tlb, void *table);
62 
63 #endif
64 
65 /*
66  * If we can't allocate a page to make a big batch of page pointers
67  * to work on, then just handle a few from the on-stack structure.
68  */
69 #define MMU_GATHER_BUNDLE	8
70 
71 struct mmu_gather_batch {
72 	struct mmu_gather_batch	*next;
73 	unsigned int		nr;
74 	unsigned int		max;
75 	struct page		*pages[0];
76 };
77 
78 #define MAX_GATHER_BATCH	\
79 	((PAGE_SIZE - sizeof(struct mmu_gather_batch)) / sizeof(void *))
80 
81 /*
82  * Limit the maximum number of mmu_gather batches to reduce a risk of soft
83  * lockups for non-preemptible kernels on huge machines when a lot of memory
84  * is zapped during unmapping.
85  * 10K pages freed at once should be safe even without a preemption point.
86  */
87 #define MAX_GATHER_BATCH_COUNT	(10000UL/MAX_GATHER_BATCH)
88 
89 /* struct mmu_gather is an opaque type used by the mm code for passing around
90  * any data needed by arch specific code for tlb_remove_page.
91  */
92 struct mmu_gather {
93 	struct mm_struct	*mm;
94 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
95 	struct mmu_table_batch	*batch;
96 #endif
97 	unsigned long		start;
98 	unsigned long		end;
99 	/* we are in the middle of an operation to clear
100 	 * a full mm and can make some optimizations */
101 	unsigned int		fullmm : 1,
102 	/* we have performed an operation which
103 	 * requires a complete flush of the tlb */
104 				need_flush_all : 1;
105 
106 	struct mmu_gather_batch *active;
107 	struct mmu_gather_batch	local;
108 	struct page		*__pages[MMU_GATHER_BUNDLE];
109 	unsigned int		batch_count;
110 	int page_size;
111 };
112 
113 #define HAVE_GENERIC_MMU_GATHER
114 
115 void arch_tlb_gather_mmu(struct mmu_gather *tlb,
116 	struct mm_struct *mm, unsigned long start, unsigned long end);
117 void tlb_flush_mmu(struct mmu_gather *tlb);
118 void arch_tlb_finish_mmu(struct mmu_gather *tlb,
119 			 unsigned long start, unsigned long end, bool force);
120 extern bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page,
121 				   int page_size);
122 
123 static inline void __tlb_adjust_range(struct mmu_gather *tlb,
124 				      unsigned long address,
125 				      unsigned int range_size)
126 {
127 	tlb->start = min(tlb->start, address);
128 	tlb->end = max(tlb->end, address + range_size);
129 }
130 
131 static inline void __tlb_reset_range(struct mmu_gather *tlb)
132 {
133 	if (tlb->fullmm) {
134 		tlb->start = tlb->end = ~0;
135 	} else {
136 		tlb->start = TASK_SIZE;
137 		tlb->end = 0;
138 	}
139 }
140 
141 static inline void tlb_remove_page_size(struct mmu_gather *tlb,
142 					struct page *page, int page_size)
143 {
144 	if (__tlb_remove_page_size(tlb, page, page_size))
145 		tlb_flush_mmu(tlb);
146 }
147 
148 static inline bool __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
149 {
150 	return __tlb_remove_page_size(tlb, page, PAGE_SIZE);
151 }
152 
153 /* tlb_remove_page
154  *	Similar to __tlb_remove_page but will call tlb_flush_mmu() itself when
155  *	required.
156  */
157 static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
158 {
159 	return tlb_remove_page_size(tlb, page, PAGE_SIZE);
160 }
161 
162 #ifndef tlb_remove_check_page_size_change
163 #define tlb_remove_check_page_size_change tlb_remove_check_page_size_change
164 static inline void tlb_remove_check_page_size_change(struct mmu_gather *tlb,
165 						     unsigned int page_size)
166 {
167 	/*
168 	 * We don't care about page size change, just update
169 	 * mmu_gather page size here so that debug checks
170 	 * doesn't throw false warning.
171 	 */
172 #ifdef CONFIG_DEBUG_VM
173 	tlb->page_size = page_size;
174 #endif
175 }
176 #endif
177 
178 /*
179  * In the case of tlb vma handling, we can optimise these away in the
180  * case where we're doing a full MM flush.  When we're doing a munmap,
181  * the vmas are adjusted to only cover the region to be torn down.
182  */
183 #ifndef tlb_start_vma
184 #define tlb_start_vma(tlb, vma) do { } while (0)
185 #endif
186 
187 #define __tlb_end_vma(tlb, vma)					\
188 	do {							\
189 		if (!tlb->fullmm && tlb->end) {			\
190 			tlb_flush(tlb);				\
191 			__tlb_reset_range(tlb);			\
192 		}						\
193 	} while (0)
194 
195 #ifndef tlb_end_vma
196 #define tlb_end_vma	__tlb_end_vma
197 #endif
198 
199 #ifndef __tlb_remove_tlb_entry
200 #define __tlb_remove_tlb_entry(tlb, ptep, address) do { } while (0)
201 #endif
202 
203 /**
204  * tlb_remove_tlb_entry - remember a pte unmapping for later tlb invalidation.
205  *
206  * Record the fact that pte's were really unmapped by updating the range,
207  * so we can later optimise away the tlb invalidate.   This helps when
208  * userspace is unmapping already-unmapped pages, which happens quite a lot.
209  */
210 #define tlb_remove_tlb_entry(tlb, ptep, address)		\
211 	do {							\
212 		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
213 		__tlb_remove_tlb_entry(tlb, ptep, address);	\
214 	} while (0)
215 
216 #define tlb_remove_huge_tlb_entry(h, tlb, ptep, address)	     \
217 	do {							     \
218 		__tlb_adjust_range(tlb, address, huge_page_size(h)); \
219 		__tlb_remove_tlb_entry(tlb, ptep, address);	     \
220 	} while (0)
221 
222 /**
223  * tlb_remove_pmd_tlb_entry - remember a pmd mapping for later tlb invalidation
224  * This is a nop so far, because only x86 needs it.
225  */
226 #ifndef __tlb_remove_pmd_tlb_entry
227 #define __tlb_remove_pmd_tlb_entry(tlb, pmdp, address) do {} while (0)
228 #endif
229 
230 #define tlb_remove_pmd_tlb_entry(tlb, pmdp, address)			\
231 	do {								\
232 		__tlb_adjust_range(tlb, address, HPAGE_PMD_SIZE);	\
233 		__tlb_remove_pmd_tlb_entry(tlb, pmdp, address);		\
234 	} while (0)
235 
236 /**
237  * tlb_remove_pud_tlb_entry - remember a pud mapping for later tlb
238  * invalidation. This is a nop so far, because only x86 needs it.
239  */
240 #ifndef __tlb_remove_pud_tlb_entry
241 #define __tlb_remove_pud_tlb_entry(tlb, pudp, address) do {} while (0)
242 #endif
243 
244 #define tlb_remove_pud_tlb_entry(tlb, pudp, address)			\
245 	do {								\
246 		__tlb_adjust_range(tlb, address, HPAGE_PUD_SIZE);	\
247 		__tlb_remove_pud_tlb_entry(tlb, pudp, address);		\
248 	} while (0)
249 
250 /*
251  * For things like page tables caches (ie caching addresses "inside" the
252  * page tables, like x86 does), for legacy reasons, flushing an
253  * individual page had better flush the page table caches behind it. This
254  * is definitely how x86 works, for example. And if you have an
255  * architected non-legacy page table cache (which I'm not aware of
256  * anybody actually doing), you're going to have some architecturally
257  * explicit flushing for that, likely *separate* from a regular TLB entry
258  * flush, and thus you'd need more than just some range expansion..
259  *
260  * So if we ever find an architecture
261  * that would want something that odd, I think it is up to that
262  * architecture to do its own odd thing, not cause pain for others
263  * http://lkml.kernel.org/r/CA+55aFzBggoXtNXQeng5d_mRoDnaMBE5Y+URs+PHR67nUpMtaw@mail.gmail.com
264  *
265  * For now w.r.t page table cache, mark the range_size as PAGE_SIZE
266  */
267 
268 #define pte_free_tlb(tlb, ptep, address)			\
269 	do {							\
270 		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
271 		__pte_free_tlb(tlb, ptep, address);		\
272 	} while (0)
273 
274 #define pmd_free_tlb(tlb, pmdp, address)			\
275 	do {							\
276 		__tlb_adjust_range(tlb, address, PAGE_SIZE);		\
277 		__pmd_free_tlb(tlb, pmdp, address);		\
278 	} while (0)
279 
280 #ifndef __ARCH_HAS_4LEVEL_HACK
281 #define pud_free_tlb(tlb, pudp, address)			\
282 	do {							\
283 		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
284 		__pud_free_tlb(tlb, pudp, address);		\
285 	} while (0)
286 #endif
287 
288 #ifndef __ARCH_HAS_5LEVEL_HACK
289 #define p4d_free_tlb(tlb, pudp, address)			\
290 	do {							\
291 		__tlb_adjust_range(tlb, address, PAGE_SIZE);		\
292 		__p4d_free_tlb(tlb, pudp, address);		\
293 	} while (0)
294 #endif
295 
296 #define tlb_migrate_finish(mm) do {} while (0)
297 
298 #endif /* _ASM_GENERIC__TLB_H */
299