Loading...
1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/* include/asm-generic/tlb.h
3 *
4 * Generic TLB shootdown code
5 *
6 * Copyright 2001 Red Hat, Inc.
7 * Based on code from mm/memory.c Copyright Linus Torvalds and others.
8 *
9 * Copyright 2011 Red Hat, Inc., Peter Zijlstra
10 */
11#ifndef _ASM_GENERIC__TLB_H
12#define _ASM_GENERIC__TLB_H
13
14#include <linux/mmu_notifier.h>
15#include <linux/swap.h>
16#include <asm/pgalloc.h>
17#include <asm/tlbflush.h>
18#include <asm/cacheflush.h>
19
20/*
21 * Blindly accessing user memory from NMI context can be dangerous
22 * if we're in the middle of switching the current user task or switching
23 * the loaded mm.
24 */
25#ifndef nmi_uaccess_okay
26# define nmi_uaccess_okay() true
27#endif
28
29#ifdef CONFIG_MMU
30
31/*
32 * Generic MMU-gather implementation.
33 *
34 * The mmu_gather data structure is used by the mm code to implement the
35 * correct and efficient ordering of freeing pages and TLB invalidations.
36 *
37 * This correct ordering is:
38 *
39 * 1) unhook page
40 * 2) TLB invalidate page
41 * 3) free page
42 *
43 * That is, we must never free a page before we have ensured there are no live
44 * translations left to it. Otherwise it might be possible to observe (or
45 * worse, change) the page content after it has been reused.
46 *
47 * The mmu_gather API consists of:
48 *
49 * - tlb_gather_mmu() / tlb_finish_mmu(); start and finish a mmu_gather
50 *
51 * Finish in particular will issue a (final) TLB invalidate and free
52 * all (remaining) queued pages.
53 *
54 * - tlb_start_vma() / tlb_end_vma(); marks the start / end of a VMA
55 *
56 * Defaults to flushing at tlb_end_vma() to reset the range; helps when
57 * there's large holes between the VMAs.
58 *
59 * - tlb_remove_page() / __tlb_remove_page()
60 * - tlb_remove_page_size() / __tlb_remove_page_size()
61 *
62 * __tlb_remove_page_size() is the basic primitive that queues a page for
63 * freeing. __tlb_remove_page() assumes PAGE_SIZE. Both will return a
64 * boolean indicating if the queue is (now) full and a call to
65 * tlb_flush_mmu() is required.
66 *
67 * tlb_remove_page() and tlb_remove_page_size() imply the call to
68 * tlb_flush_mmu() when required and has no return value.
69 *
70 * - tlb_change_page_size()
71 *
72 * call before __tlb_remove_page*() to set the current page-size; implies a
73 * possible tlb_flush_mmu() call.
74 *
75 * - tlb_flush_mmu() / tlb_flush_mmu_tlbonly()
76 *
77 * tlb_flush_mmu_tlbonly() - does the TLB invalidate (and resets
78 * related state, like the range)
79 *
80 * tlb_flush_mmu() - in addition to the above TLB invalidate, also frees
81 * whatever pages are still batched.
82 *
83 * - mmu_gather::fullmm
84 *
85 * A flag set by tlb_gather_mmu() to indicate we're going to free
86 * the entire mm; this allows a number of optimizations.
87 *
88 * - We can ignore tlb_{start,end}_vma(); because we don't
89 * care about ranges. Everything will be shot down.
90 *
91 * - (RISC) architectures that use ASIDs can cycle to a new ASID
92 * and delay the invalidation until ASID space runs out.
93 *
94 * - mmu_gather::need_flush_all
95 *
96 * A flag that can be set by the arch code if it wants to force
97 * flush the entire TLB irrespective of the range. For instance
98 * x86-PAE needs this when changing top-level entries.
99 *
100 * And allows the architecture to provide and implement tlb_flush():
101 *
102 * tlb_flush() may, in addition to the above mentioned mmu_gather fields, make
103 * use of:
104 *
105 * - mmu_gather::start / mmu_gather::end
106 *
107 * which provides the range that needs to be flushed to cover the pages to
108 * be freed.
109 *
110 * - mmu_gather::freed_tables
111 *
112 * set when we freed page table pages
113 *
114 * - tlb_get_unmap_shift() / tlb_get_unmap_size()
115 *
116 * returns the smallest TLB entry size unmapped in this range.
117 *
118 * If an architecture does not provide tlb_flush() a default implementation
119 * based on flush_tlb_range() will be used, unless MMU_GATHER_NO_RANGE is
120 * specified, in which case we'll default to flush_tlb_mm().
121 *
122 * Additionally there are a few opt-in features:
123 *
124 * HAVE_MMU_GATHER_PAGE_SIZE
125 *
126 * This ensures we call tlb_flush() every time tlb_change_page_size() actually
127 * changes the size and provides mmu_gather::page_size to tlb_flush().
128 *
129 * HAVE_RCU_TABLE_FREE
130 *
131 * This provides tlb_remove_table(), to be used instead of tlb_remove_page()
132 * for page directores (__p*_free_tlb()). This provides separate freeing of
133 * the page-table pages themselves in a semi-RCU fashion (see comment below).
134 * Useful if your architecture doesn't use IPIs for remote TLB invalidates
135 * and therefore doesn't naturally serialize with software page-table walkers.
136 *
137 * When used, an architecture is expected to provide __tlb_remove_table()
138 * which does the actual freeing of these pages.
139 *
140 * HAVE_RCU_TABLE_NO_INVALIDATE
141 *
142 * This makes HAVE_RCU_TABLE_FREE avoid calling tlb_flush_mmu_tlbonly() before
143 * freeing the page-table pages. This can be avoided if you use
144 * HAVE_RCU_TABLE_FREE and your architecture does _NOT_ use the Linux
145 * page-tables natively.
146 *
147 * MMU_GATHER_NO_RANGE
148 *
149 * Use this if your architecture lacks an efficient flush_tlb_range().
150 */
151
152#ifdef CONFIG_HAVE_RCU_TABLE_FREE
153/*
154 * Semi RCU freeing of the page directories.
155 *
156 * This is needed by some architectures to implement software pagetable walkers.
157 *
158 * gup_fast() and other software pagetable walkers do a lockless page-table
159 * walk and therefore needs some synchronization with the freeing of the page
160 * directories. The chosen means to accomplish that is by disabling IRQs over
161 * the walk.
162 *
163 * Architectures that use IPIs to flush TLBs will then automagically DTRT,
164 * since we unlink the page, flush TLBs, free the page. Since the disabling of
165 * IRQs delays the completion of the TLB flush we can never observe an already
166 * freed page.
167 *
168 * Architectures that do not have this (PPC) need to delay the freeing by some
169 * other means, this is that means.
170 *
171 * What we do is batch the freed directory pages (tables) and RCU free them.
172 * We use the sched RCU variant, as that guarantees that IRQ/preempt disabling
173 * holds off grace periods.
174 *
175 * However, in order to batch these pages we need to allocate storage, this
176 * allocation is deep inside the MM code and can thus easily fail on memory
177 * pressure. To guarantee progress we fall back to single table freeing, see
178 * the implementation of tlb_remove_table_one().
179 *
180 */
181struct mmu_table_batch {
182 struct rcu_head rcu;
183 unsigned int nr;
184 void *tables[0];
185};
186
187#define MAX_TABLE_BATCH \
188 ((PAGE_SIZE - sizeof(struct mmu_table_batch)) / sizeof(void *))
189
190extern void tlb_remove_table(struct mmu_gather *tlb, void *table);
191
192#endif
193
194#ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER
195/*
196 * If we can't allocate a page to make a big batch of page pointers
197 * to work on, then just handle a few from the on-stack structure.
198 */
199#define MMU_GATHER_BUNDLE 8
200
201struct mmu_gather_batch {
202 struct mmu_gather_batch *next;
203 unsigned int nr;
204 unsigned int max;
205 struct page *pages[0];
206};
207
208#define MAX_GATHER_BATCH \
209 ((PAGE_SIZE - sizeof(struct mmu_gather_batch)) / sizeof(void *))
210
211/*
212 * Limit the maximum number of mmu_gather batches to reduce a risk of soft
213 * lockups for non-preemptible kernels on huge machines when a lot of memory
214 * is zapped during unmapping.
215 * 10K pages freed at once should be safe even without a preemption point.
216 */
217#define MAX_GATHER_BATCH_COUNT (10000UL/MAX_GATHER_BATCH)
218
219extern bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page,
220 int page_size);
221#endif
222
223/*
224 * struct mmu_gather is an opaque type used by the mm code for passing around
225 * any data needed by arch specific code for tlb_remove_page.
226 */
227struct mmu_gather {
228 struct mm_struct *mm;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 struct mmu_table_batch *batch;
232#endif
233
234 unsigned long start;
235 unsigned long end;
236 /*
237 * we are in the middle of an operation to clear
238 * a full mm and can make some optimizations
239 */
240 unsigned int fullmm : 1;
241
242 /*
243 * we have performed an operation which
244 * requires a complete flush of the tlb
245 */
246 unsigned int need_flush_all : 1;
247
248 /*
249 * we have removed page directories
250 */
251 unsigned int freed_tables : 1;
252
253 /*
254 * at which levels have we cleared entries?
255 */
256 unsigned int cleared_ptes : 1;
257 unsigned int cleared_pmds : 1;
258 unsigned int cleared_puds : 1;
259 unsigned int cleared_p4ds : 1;
260
261 /*
262 * tracks VM_EXEC | VM_HUGETLB in tlb_start_vma
263 */
264 unsigned int vma_exec : 1;
265 unsigned int vma_huge : 1;
266
267 unsigned int batch_count;
268
269#ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER
270 struct mmu_gather_batch *active;
271 struct mmu_gather_batch local;
272 struct page *__pages[MMU_GATHER_BUNDLE];
273
274#ifdef CONFIG_HAVE_MMU_GATHER_PAGE_SIZE
275 unsigned int page_size;
276#endif
277#endif
278};
279
280void arch_tlb_gather_mmu(struct mmu_gather *tlb,
281 struct mm_struct *mm, unsigned long start, unsigned long end);
282void tlb_flush_mmu(struct mmu_gather *tlb);
283void arch_tlb_finish_mmu(struct mmu_gather *tlb,
284 unsigned long start, unsigned long end, bool force);
285
286static inline void __tlb_adjust_range(struct mmu_gather *tlb,
287 unsigned long address,
288 unsigned int range_size)
289{
290 tlb->start = min(tlb->start, address);
291 tlb->end = max(tlb->end, address + range_size);
292}
293
294static inline void __tlb_reset_range(struct mmu_gather *tlb)
295{
296 if (tlb->fullmm) {
297 tlb->start = tlb->end = ~0;
298 } else {
299 tlb->start = TASK_SIZE;
300 tlb->end = 0;
301 }
302 tlb->freed_tables = 0;
303 tlb->cleared_ptes = 0;
304 tlb->cleared_pmds = 0;
305 tlb->cleared_puds = 0;
306 tlb->cleared_p4ds = 0;
307 /*
308 * Do not reset mmu_gather::vma_* fields here, we do not
309 * call into tlb_start_vma() again to set them if there is an
310 * intermediate flush.
311 */
312}
313
314#ifdef CONFIG_MMU_GATHER_NO_RANGE
315
316#if defined(tlb_flush) || defined(tlb_start_vma) || defined(tlb_end_vma)
317#error MMU_GATHER_NO_RANGE relies on default tlb_flush(), tlb_start_vma() and tlb_end_vma()
318#endif
319
320/*
321 * When an architecture does not have efficient means of range flushing TLBs
322 * there is no point in doing intermediate flushes on tlb_end_vma() to keep the
323 * range small. We equally don't have to worry about page granularity or other
324 * things.
325 *
326 * All we need to do is issue a full flush for any !0 range.
327 */
328static inline void tlb_flush(struct mmu_gather *tlb)
329{
330 if (tlb->end)
331 flush_tlb_mm(tlb->mm);
332}
333
334static inline void
335tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma) { }
336
337#define tlb_end_vma tlb_end_vma
338static inline void tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma) { }
339
340#else /* CONFIG_MMU_GATHER_NO_RANGE */
341
342#ifndef tlb_flush
343
344#if defined(tlb_start_vma) || defined(tlb_end_vma)
345#error Default tlb_flush() relies on default tlb_start_vma() and tlb_end_vma()
346#endif
347
348/*
349 * When an architecture does not provide its own tlb_flush() implementation
350 * but does have a reasonably efficient flush_vma_range() implementation
351 * use that.
352 */
353static inline void tlb_flush(struct mmu_gather *tlb)
354{
355 if (tlb->fullmm || tlb->need_flush_all) {
356 flush_tlb_mm(tlb->mm);
357 } else if (tlb->end) {
358 struct vm_area_struct vma = {
359 .vm_mm = tlb->mm,
360 .vm_flags = (tlb->vma_exec ? VM_EXEC : 0) |
361 (tlb->vma_huge ? VM_HUGETLB : 0),
362 };
363
364 flush_tlb_range(&vma, tlb->start, tlb->end);
365 }
366}
367
368static inline void
369tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma)
370{
371 /*
372 * flush_tlb_range() implementations that look at VM_HUGETLB (tile,
373 * mips-4k) flush only large pages.
374 *
375 * flush_tlb_range() implementations that flush I-TLB also flush D-TLB
376 * (tile, xtensa, arm), so it's ok to just add VM_EXEC to an existing
377 * range.
378 *
379 * We rely on tlb_end_vma() to issue a flush, such that when we reset
380 * these values the batch is empty.
381 */
382 tlb->vma_huge = !!(vma->vm_flags & VM_HUGETLB);
383 tlb->vma_exec = !!(vma->vm_flags & VM_EXEC);
384}
385
386#else
387
388static inline void
389tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma) { }
390
391#endif
392
393#endif /* CONFIG_MMU_GATHER_NO_RANGE */
394
395static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
396{
397 if (!tlb->end)
398 return;
399
400 tlb_flush(tlb);
401 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
402 __tlb_reset_range(tlb);
403}
404
405static inline void tlb_remove_page_size(struct mmu_gather *tlb,
406 struct page *page, int page_size)
407{
408 if (__tlb_remove_page_size(tlb, page, page_size))
409 tlb_flush_mmu(tlb);
410}
411
412static inline bool __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
413{
414 return __tlb_remove_page_size(tlb, page, PAGE_SIZE);
415}
416
417/* tlb_remove_page
418 * Similar to __tlb_remove_page but will call tlb_flush_mmu() itself when
419 * required.
420 */
421static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
422{
423 return tlb_remove_page_size(tlb, page, PAGE_SIZE);
424}
425
426static inline void tlb_change_page_size(struct mmu_gather *tlb,
427 unsigned int page_size)
428{
429#ifdef CONFIG_HAVE_MMU_GATHER_PAGE_SIZE
430 if (tlb->page_size && tlb->page_size != page_size) {
431 if (!tlb->fullmm)
432 tlb_flush_mmu(tlb);
433 }
434
435 tlb->page_size = page_size;
436#endif
437}
438
439static inline unsigned long tlb_get_unmap_shift(struct mmu_gather *tlb)
440{
441 if (tlb->cleared_ptes)
442 return PAGE_SHIFT;
443 if (tlb->cleared_pmds)
444 return PMD_SHIFT;
445 if (tlb->cleared_puds)
446 return PUD_SHIFT;
447 if (tlb->cleared_p4ds)
448 return P4D_SHIFT;
449
450 return PAGE_SHIFT;
451}
452
453static inline unsigned long tlb_get_unmap_size(struct mmu_gather *tlb)
454{
455 return 1UL << tlb_get_unmap_shift(tlb);
456}
457
458/*
459 * In the case of tlb vma handling, we can optimise these away in the
460 * case where we're doing a full MM flush. When we're doing a munmap,
461 * the vmas are adjusted to only cover the region to be torn down.
462 */
463#ifndef tlb_start_vma
464static inline void tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
465{
466 if (tlb->fullmm)
467 return;
468
469 tlb_update_vma_flags(tlb, vma);
470 flush_cache_range(vma, vma->vm_start, vma->vm_end);
471}
472#endif
473
474#ifndef tlb_end_vma
475static inline void tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
476{
477 if (tlb->fullmm)
478 return;
479
480 /*
481 * Do a TLB flush and reset the range at VMA boundaries; this avoids
482 * the ranges growing with the unused space between consecutive VMAs,
483 * but also the mmu_gather::vma_* flags from tlb_start_vma() rely on
484 * this.
485 */
486 tlb_flush_mmu_tlbonly(tlb);
487}
488#endif
489
490#ifndef __tlb_remove_tlb_entry
491#define __tlb_remove_tlb_entry(tlb, ptep, address) do { } while (0)
492#endif
493
494/**
495 * tlb_remove_tlb_entry - remember a pte unmapping for later tlb invalidation.
496 *
497 * Record the fact that pte's were really unmapped by updating the range,
498 * so we can later optimise away the tlb invalidate. This helps when
499 * userspace is unmapping already-unmapped pages, which happens quite a lot.
500 */
501#define tlb_remove_tlb_entry(tlb, ptep, address) \
502 do { \
503 __tlb_adjust_range(tlb, address, PAGE_SIZE); \
504 tlb->cleared_ptes = 1; \
505 __tlb_remove_tlb_entry(tlb, ptep, address); \
506 } while (0)
507
508#define tlb_remove_huge_tlb_entry(h, tlb, ptep, address) \
509 do { \
510 unsigned long _sz = huge_page_size(h); \
511 __tlb_adjust_range(tlb, address, _sz); \
512 if (_sz == PMD_SIZE) \
513 tlb->cleared_pmds = 1; \
514 else if (_sz == PUD_SIZE) \
515 tlb->cleared_puds = 1; \
516 __tlb_remove_tlb_entry(tlb, ptep, address); \
517 } while (0)
518
519/**
520 * tlb_remove_pmd_tlb_entry - remember a pmd mapping for later tlb invalidation
521 * This is a nop so far, because only x86 needs it.
522 */
523#ifndef __tlb_remove_pmd_tlb_entry
524#define __tlb_remove_pmd_tlb_entry(tlb, pmdp, address) do {} while (0)
525#endif
526
527#define tlb_remove_pmd_tlb_entry(tlb, pmdp, address) \
528 do { \
529 __tlb_adjust_range(tlb, address, HPAGE_PMD_SIZE); \
530 tlb->cleared_pmds = 1; \
531 __tlb_remove_pmd_tlb_entry(tlb, pmdp, address); \
532 } while (0)
533
534/**
535 * tlb_remove_pud_tlb_entry - remember a pud mapping for later tlb
536 * invalidation. This is a nop so far, because only x86 needs it.
537 */
538#ifndef __tlb_remove_pud_tlb_entry
539#define __tlb_remove_pud_tlb_entry(tlb, pudp, address) do {} while (0)
540#endif
541
542#define tlb_remove_pud_tlb_entry(tlb, pudp, address) \
543 do { \
544 __tlb_adjust_range(tlb, address, HPAGE_PUD_SIZE); \
545 tlb->cleared_puds = 1; \
546 __tlb_remove_pud_tlb_entry(tlb, pudp, address); \
547 } while (0)
548
549/*
550 * For things like page tables caches (ie caching addresses "inside" the
551 * page tables, like x86 does), for legacy reasons, flushing an
552 * individual page had better flush the page table caches behind it. This
553 * is definitely how x86 works, for example. And if you have an
554 * architected non-legacy page table cache (which I'm not aware of
555 * anybody actually doing), you're going to have some architecturally
556 * explicit flushing for that, likely *separate* from a regular TLB entry
557 * flush, and thus you'd need more than just some range expansion..
558 *
559 * So if we ever find an architecture
560 * that would want something that odd, I think it is up to that
561 * architecture to do its own odd thing, not cause pain for others
562 * http://lkml.kernel.org/r/CA+55aFzBggoXtNXQeng5d_mRoDnaMBE5Y+URs+PHR67nUpMtaw@mail.gmail.com
563 *
564 * For now w.r.t page table cache, mark the range_size as PAGE_SIZE
565 */
566
567#ifndef pte_free_tlb
568#define pte_free_tlb(tlb, ptep, address) \
569 do { \
570 __tlb_adjust_range(tlb, address, PAGE_SIZE); \
571 tlb->freed_tables = 1; \
572 tlb->cleared_pmds = 1; \
573 __pte_free_tlb(tlb, ptep, address); \
574 } while (0)
575#endif
576
577#ifndef pmd_free_tlb
578#define pmd_free_tlb(tlb, pmdp, address) \
579 do { \
580 __tlb_adjust_range(tlb, address, PAGE_SIZE); \
581 tlb->freed_tables = 1; \
582 tlb->cleared_puds = 1; \
583 __pmd_free_tlb(tlb, pmdp, address); \
584 } while (0)
585#endif
586
587#ifndef __ARCH_HAS_4LEVEL_HACK
588#ifndef pud_free_tlb
589#define pud_free_tlb(tlb, pudp, address) \
590 do { \
591 __tlb_adjust_range(tlb, address, PAGE_SIZE); \
592 tlb->freed_tables = 1; \
593 tlb->cleared_p4ds = 1; \
594 __pud_free_tlb(tlb, pudp, address); \
595 } while (0)
596#endif
597#endif
598
599#ifndef __ARCH_HAS_5LEVEL_HACK
600#ifndef p4d_free_tlb
601#define p4d_free_tlb(tlb, pudp, address) \
602 do { \
603 __tlb_adjust_range(tlb, address, PAGE_SIZE); \
604 tlb->freed_tables = 1; \
605 __p4d_free_tlb(tlb, pudp, address); \
606 } while (0)
607#endif
608#endif
609
610#endif /* CONFIG_MMU */
611
612#endif /* _ASM_GENERIC__TLB_H */
1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/* include/asm-generic/tlb.h
3 *
4 * Generic TLB shootdown code
5 *
6 * Copyright 2001 Red Hat, Inc.
7 * Based on code from mm/memory.c Copyright Linus Torvalds and others.
8 *
9 * Copyright 2011 Red Hat, Inc., Peter Zijlstra
10 */
11#ifndef _ASM_GENERIC__TLB_H
12#define _ASM_GENERIC__TLB_H
13
14#include <linux/mmu_notifier.h>
15#include <linux/swap.h>
16#include <linux/hugetlb_inline.h>
17#include <asm/tlbflush.h>
18#include <asm/cacheflush.h>
19
20/*
21 * Blindly accessing user memory from NMI context can be dangerous
22 * if we're in the middle of switching the current user task or switching
23 * the loaded mm.
24 */
25#ifndef nmi_uaccess_okay
26# define nmi_uaccess_okay() true
27#endif
28
29#ifdef CONFIG_MMU
30
31/*
32 * Generic MMU-gather implementation.
33 *
34 * The mmu_gather data structure is used by the mm code to implement the
35 * correct and efficient ordering of freeing pages and TLB invalidations.
36 *
37 * This correct ordering is:
38 *
39 * 1) unhook page
40 * 2) TLB invalidate page
41 * 3) free page
42 *
43 * That is, we must never free a page before we have ensured there are no live
44 * translations left to it. Otherwise it might be possible to observe (or
45 * worse, change) the page content after it has been reused.
46 *
47 * The mmu_gather API consists of:
48 *
49 * - tlb_gather_mmu() / tlb_gather_mmu_fullmm() / tlb_finish_mmu()
50 *
51 * start and finish a mmu_gather
52 *
53 * Finish in particular will issue a (final) TLB invalidate and free
54 * all (remaining) queued pages.
55 *
56 * - tlb_start_vma() / tlb_end_vma(); marks the start / end of a VMA
57 *
58 * Defaults to flushing at tlb_end_vma() to reset the range; helps when
59 * there's large holes between the VMAs.
60 *
61 * - tlb_remove_table()
62 *
63 * tlb_remove_table() is the basic primitive to free page-table directories
64 * (__p*_free_tlb()). In it's most primitive form it is an alias for
65 * tlb_remove_page() below, for when page directories are pages and have no
66 * additional constraints.
67 *
68 * See also MMU_GATHER_TABLE_FREE and MMU_GATHER_RCU_TABLE_FREE.
69 *
70 * - tlb_remove_page() / __tlb_remove_page()
71 * - tlb_remove_page_size() / __tlb_remove_page_size()
72 *
73 * __tlb_remove_page_size() is the basic primitive that queues a page for
74 * freeing. __tlb_remove_page() assumes PAGE_SIZE. Both will return a
75 * boolean indicating if the queue is (now) full and a call to
76 * tlb_flush_mmu() is required.
77 *
78 * tlb_remove_page() and tlb_remove_page_size() imply the call to
79 * tlb_flush_mmu() when required and has no return value.
80 *
81 * - tlb_change_page_size()
82 *
83 * call before __tlb_remove_page*() to set the current page-size; implies a
84 * possible tlb_flush_mmu() call.
85 *
86 * - tlb_flush_mmu() / tlb_flush_mmu_tlbonly()
87 *
88 * tlb_flush_mmu_tlbonly() - does the TLB invalidate (and resets
89 * related state, like the range)
90 *
91 * tlb_flush_mmu() - in addition to the above TLB invalidate, also frees
92 * whatever pages are still batched.
93 *
94 * - mmu_gather::fullmm
95 *
96 * A flag set by tlb_gather_mmu_fullmm() to indicate we're going to free
97 * the entire mm; this allows a number of optimizations.
98 *
99 * - We can ignore tlb_{start,end}_vma(); because we don't
100 * care about ranges. Everything will be shot down.
101 *
102 * - (RISC) architectures that use ASIDs can cycle to a new ASID
103 * and delay the invalidation until ASID space runs out.
104 *
105 * - mmu_gather::need_flush_all
106 *
107 * A flag that can be set by the arch code if it wants to force
108 * flush the entire TLB irrespective of the range. For instance
109 * x86-PAE needs this when changing top-level entries.
110 *
111 * And allows the architecture to provide and implement tlb_flush():
112 *
113 * tlb_flush() may, in addition to the above mentioned mmu_gather fields, make
114 * use of:
115 *
116 * - mmu_gather::start / mmu_gather::end
117 *
118 * which provides the range that needs to be flushed to cover the pages to
119 * be freed.
120 *
121 * - mmu_gather::freed_tables
122 *
123 * set when we freed page table pages
124 *
125 * - tlb_get_unmap_shift() / tlb_get_unmap_size()
126 *
127 * returns the smallest TLB entry size unmapped in this range.
128 *
129 * If an architecture does not provide tlb_flush() a default implementation
130 * based on flush_tlb_range() will be used, unless MMU_GATHER_NO_RANGE is
131 * specified, in which case we'll default to flush_tlb_mm().
132 *
133 * Additionally there are a few opt-in features:
134 *
135 * MMU_GATHER_PAGE_SIZE
136 *
137 * This ensures we call tlb_flush() every time tlb_change_page_size() actually
138 * changes the size and provides mmu_gather::page_size to tlb_flush().
139 *
140 * This might be useful if your architecture has size specific TLB
141 * invalidation instructions.
142 *
143 * MMU_GATHER_TABLE_FREE
144 *
145 * This provides tlb_remove_table(), to be used instead of tlb_remove_page()
146 * for page directores (__p*_free_tlb()).
147 *
148 * Useful if your architecture has non-page page directories.
149 *
150 * When used, an architecture is expected to provide __tlb_remove_table()
151 * which does the actual freeing of these pages.
152 *
153 * MMU_GATHER_RCU_TABLE_FREE
154 *
155 * Like MMU_GATHER_TABLE_FREE, and adds semi-RCU semantics to the free (see
156 * comment below).
157 *
158 * Useful if your architecture doesn't use IPIs for remote TLB invalidates
159 * and therefore doesn't naturally serialize with software page-table walkers.
160 *
161 * MMU_GATHER_NO_RANGE
162 *
163 * Use this if your architecture lacks an efficient flush_tlb_range().
164 *
165 * MMU_GATHER_NO_GATHER
166 *
167 * If the option is set the mmu_gather will not track individual pages for
168 * delayed page free anymore. A platform that enables the option needs to
169 * provide its own implementation of the __tlb_remove_page_size() function to
170 * free pages.
171 *
172 * This is useful if your architecture already flushes TLB entries in the
173 * various ptep_get_and_clear() functions.
174 */
175
176#ifdef CONFIG_MMU_GATHER_TABLE_FREE
177
178struct mmu_table_batch {
179#ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
180 struct rcu_head rcu;
181#endif
182 unsigned int nr;
183 void *tables[0];
184};
185
186#define MAX_TABLE_BATCH \
187 ((PAGE_SIZE - sizeof(struct mmu_table_batch)) / sizeof(void *))
188
189extern void tlb_remove_table(struct mmu_gather *tlb, void *table);
190
191#else /* !CONFIG_MMU_GATHER_HAVE_TABLE_FREE */
192
193/*
194 * Without MMU_GATHER_TABLE_FREE the architecture is assumed to have page based
195 * page directories and we can use the normal page batching to free them.
196 */
197#define tlb_remove_table(tlb, page) tlb_remove_page((tlb), (page))
198
199#endif /* CONFIG_MMU_GATHER_TABLE_FREE */
200
201#ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
202/*
203 * This allows an architecture that does not use the linux page-tables for
204 * hardware to skip the TLBI when freeing page tables.
205 */
206#ifndef tlb_needs_table_invalidate
207#define tlb_needs_table_invalidate() (true)
208#endif
209
210#else
211
212#ifdef tlb_needs_table_invalidate
213#error tlb_needs_table_invalidate() requires MMU_GATHER_RCU_TABLE_FREE
214#endif
215
216#endif /* CONFIG_MMU_GATHER_RCU_TABLE_FREE */
217
218
219#ifndef CONFIG_MMU_GATHER_NO_GATHER
220/*
221 * If we can't allocate a page to make a big batch of page pointers
222 * to work on, then just handle a few from the on-stack structure.
223 */
224#define MMU_GATHER_BUNDLE 8
225
226struct mmu_gather_batch {
227 struct mmu_gather_batch *next;
228 unsigned int nr;
229 unsigned int max;
230 struct page *pages[0];
231};
232
233#define MAX_GATHER_BATCH \
234 ((PAGE_SIZE - sizeof(struct mmu_gather_batch)) / sizeof(void *))
235
236/*
237 * Limit the maximum number of mmu_gather batches to reduce a risk of soft
238 * lockups for non-preemptible kernels on huge machines when a lot of memory
239 * is zapped during unmapping.
240 * 10K pages freed at once should be safe even without a preemption point.
241 */
242#define MAX_GATHER_BATCH_COUNT (10000UL/MAX_GATHER_BATCH)
243
244extern bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page,
245 int page_size);
246#endif
247
248/*
249 * struct mmu_gather is an opaque type used by the mm code for passing around
250 * any data needed by arch specific code for tlb_remove_page.
251 */
252struct mmu_gather {
253 struct mm_struct *mm;
254
255#ifdef CONFIG_MMU_GATHER_TABLE_FREE
256 struct mmu_table_batch *batch;
257#endif
258
259 unsigned long start;
260 unsigned long end;
261 /*
262 * we are in the middle of an operation to clear
263 * a full mm and can make some optimizations
264 */
265 unsigned int fullmm : 1;
266
267 /*
268 * we have performed an operation which
269 * requires a complete flush of the tlb
270 */
271 unsigned int need_flush_all : 1;
272
273 /*
274 * we have removed page directories
275 */
276 unsigned int freed_tables : 1;
277
278 /*
279 * at which levels have we cleared entries?
280 */
281 unsigned int cleared_ptes : 1;
282 unsigned int cleared_pmds : 1;
283 unsigned int cleared_puds : 1;
284 unsigned int cleared_p4ds : 1;
285
286 /*
287 * tracks VM_EXEC | VM_HUGETLB in tlb_start_vma
288 */
289 unsigned int vma_exec : 1;
290 unsigned int vma_huge : 1;
291
292 unsigned int batch_count;
293
294#ifndef CONFIG_MMU_GATHER_NO_GATHER
295 struct mmu_gather_batch *active;
296 struct mmu_gather_batch local;
297 struct page *__pages[MMU_GATHER_BUNDLE];
298
299#ifdef CONFIG_MMU_GATHER_PAGE_SIZE
300 unsigned int page_size;
301#endif
302#endif
303};
304
305void tlb_flush_mmu(struct mmu_gather *tlb);
306
307static inline void __tlb_adjust_range(struct mmu_gather *tlb,
308 unsigned long address,
309 unsigned int range_size)
310{
311 tlb->start = min(tlb->start, address);
312 tlb->end = max(tlb->end, address + range_size);
313}
314
315static inline void __tlb_reset_range(struct mmu_gather *tlb)
316{
317 if (tlb->fullmm) {
318 tlb->start = tlb->end = ~0;
319 } else {
320 tlb->start = TASK_SIZE;
321 tlb->end = 0;
322 }
323 tlb->freed_tables = 0;
324 tlb->cleared_ptes = 0;
325 tlb->cleared_pmds = 0;
326 tlb->cleared_puds = 0;
327 tlb->cleared_p4ds = 0;
328 /*
329 * Do not reset mmu_gather::vma_* fields here, we do not
330 * call into tlb_start_vma() again to set them if there is an
331 * intermediate flush.
332 */
333}
334
335#ifdef CONFIG_MMU_GATHER_NO_RANGE
336
337#if defined(tlb_flush) || defined(tlb_start_vma) || defined(tlb_end_vma)
338#error MMU_GATHER_NO_RANGE relies on default tlb_flush(), tlb_start_vma() and tlb_end_vma()
339#endif
340
341/*
342 * When an architecture does not have efficient means of range flushing TLBs
343 * there is no point in doing intermediate flushes on tlb_end_vma() to keep the
344 * range small. We equally don't have to worry about page granularity or other
345 * things.
346 *
347 * All we need to do is issue a full flush for any !0 range.
348 */
349static inline void tlb_flush(struct mmu_gather *tlb)
350{
351 if (tlb->end)
352 flush_tlb_mm(tlb->mm);
353}
354
355static inline void
356tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma) { }
357
358#define tlb_end_vma tlb_end_vma
359static inline void tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma) { }
360
361#else /* CONFIG_MMU_GATHER_NO_RANGE */
362
363#ifndef tlb_flush
364
365#if defined(tlb_start_vma) || defined(tlb_end_vma)
366#error Default tlb_flush() relies on default tlb_start_vma() and tlb_end_vma()
367#endif
368
369/*
370 * When an architecture does not provide its own tlb_flush() implementation
371 * but does have a reasonably efficient flush_vma_range() implementation
372 * use that.
373 */
374static inline void tlb_flush(struct mmu_gather *tlb)
375{
376 if (tlb->fullmm || tlb->need_flush_all) {
377 flush_tlb_mm(tlb->mm);
378 } else if (tlb->end) {
379 struct vm_area_struct vma = {
380 .vm_mm = tlb->mm,
381 .vm_flags = (tlb->vma_exec ? VM_EXEC : 0) |
382 (tlb->vma_huge ? VM_HUGETLB : 0),
383 };
384
385 flush_tlb_range(&vma, tlb->start, tlb->end);
386 }
387}
388
389static inline void
390tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma)
391{
392 /*
393 * flush_tlb_range() implementations that look at VM_HUGETLB (tile,
394 * mips-4k) flush only large pages.
395 *
396 * flush_tlb_range() implementations that flush I-TLB also flush D-TLB
397 * (tile, xtensa, arm), so it's ok to just add VM_EXEC to an existing
398 * range.
399 *
400 * We rely on tlb_end_vma() to issue a flush, such that when we reset
401 * these values the batch is empty.
402 */
403 tlb->vma_huge = is_vm_hugetlb_page(vma);
404 tlb->vma_exec = !!(vma->vm_flags & VM_EXEC);
405}
406
407#else
408
409static inline void
410tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma) { }
411
412#endif
413
414#endif /* CONFIG_MMU_GATHER_NO_RANGE */
415
416static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
417{
418 /*
419 * Anything calling __tlb_adjust_range() also sets at least one of
420 * these bits.
421 */
422 if (!(tlb->freed_tables || tlb->cleared_ptes || tlb->cleared_pmds ||
423 tlb->cleared_puds || tlb->cleared_p4ds))
424 return;
425
426 tlb_flush(tlb);
427 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
428 __tlb_reset_range(tlb);
429}
430
431static inline void tlb_remove_page_size(struct mmu_gather *tlb,
432 struct page *page, int page_size)
433{
434 if (__tlb_remove_page_size(tlb, page, page_size))
435 tlb_flush_mmu(tlb);
436}
437
438static inline bool __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
439{
440 return __tlb_remove_page_size(tlb, page, PAGE_SIZE);
441}
442
443/* tlb_remove_page
444 * Similar to __tlb_remove_page but will call tlb_flush_mmu() itself when
445 * required.
446 */
447static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
448{
449 return tlb_remove_page_size(tlb, page, PAGE_SIZE);
450}
451
452static inline void tlb_change_page_size(struct mmu_gather *tlb,
453 unsigned int page_size)
454{
455#ifdef CONFIG_MMU_GATHER_PAGE_SIZE
456 if (tlb->page_size && tlb->page_size != page_size) {
457 if (!tlb->fullmm && !tlb->need_flush_all)
458 tlb_flush_mmu(tlb);
459 }
460
461 tlb->page_size = page_size;
462#endif
463}
464
465static inline unsigned long tlb_get_unmap_shift(struct mmu_gather *tlb)
466{
467 if (tlb->cleared_ptes)
468 return PAGE_SHIFT;
469 if (tlb->cleared_pmds)
470 return PMD_SHIFT;
471 if (tlb->cleared_puds)
472 return PUD_SHIFT;
473 if (tlb->cleared_p4ds)
474 return P4D_SHIFT;
475
476 return PAGE_SHIFT;
477}
478
479static inline unsigned long tlb_get_unmap_size(struct mmu_gather *tlb)
480{
481 return 1UL << tlb_get_unmap_shift(tlb);
482}
483
484/*
485 * In the case of tlb vma handling, we can optimise these away in the
486 * case where we're doing a full MM flush. When we're doing a munmap,
487 * the vmas are adjusted to only cover the region to be torn down.
488 */
489#ifndef tlb_start_vma
490static inline void tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
491{
492 if (tlb->fullmm)
493 return;
494
495 tlb_update_vma_flags(tlb, vma);
496 flush_cache_range(vma, vma->vm_start, vma->vm_end);
497}
498#endif
499
500#ifndef tlb_end_vma
501static inline void tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
502{
503 if (tlb->fullmm)
504 return;
505
506 /*
507 * Do a TLB flush and reset the range at VMA boundaries; this avoids
508 * the ranges growing with the unused space between consecutive VMAs,
509 * but also the mmu_gather::vma_* flags from tlb_start_vma() rely on
510 * this.
511 */
512 tlb_flush_mmu_tlbonly(tlb);
513}
514#endif
515
516/*
517 * tlb_flush_{pte|pmd|pud|p4d}_range() adjust the tlb->start and tlb->end,
518 * and set corresponding cleared_*.
519 */
520static inline void tlb_flush_pte_range(struct mmu_gather *tlb,
521 unsigned long address, unsigned long size)
522{
523 __tlb_adjust_range(tlb, address, size);
524 tlb->cleared_ptes = 1;
525}
526
527static inline void tlb_flush_pmd_range(struct mmu_gather *tlb,
528 unsigned long address, unsigned long size)
529{
530 __tlb_adjust_range(tlb, address, size);
531 tlb->cleared_pmds = 1;
532}
533
534static inline void tlb_flush_pud_range(struct mmu_gather *tlb,
535 unsigned long address, unsigned long size)
536{
537 __tlb_adjust_range(tlb, address, size);
538 tlb->cleared_puds = 1;
539}
540
541static inline void tlb_flush_p4d_range(struct mmu_gather *tlb,
542 unsigned long address, unsigned long size)
543{
544 __tlb_adjust_range(tlb, address, size);
545 tlb->cleared_p4ds = 1;
546}
547
548#ifndef __tlb_remove_tlb_entry
549#define __tlb_remove_tlb_entry(tlb, ptep, address) do { } while (0)
550#endif
551
552/**
553 * tlb_remove_tlb_entry - remember a pte unmapping for later tlb invalidation.
554 *
555 * Record the fact that pte's were really unmapped by updating the range,
556 * so we can later optimise away the tlb invalidate. This helps when
557 * userspace is unmapping already-unmapped pages, which happens quite a lot.
558 */
559#define tlb_remove_tlb_entry(tlb, ptep, address) \
560 do { \
561 tlb_flush_pte_range(tlb, address, PAGE_SIZE); \
562 __tlb_remove_tlb_entry(tlb, ptep, address); \
563 } while (0)
564
565#define tlb_remove_huge_tlb_entry(h, tlb, ptep, address) \
566 do { \
567 unsigned long _sz = huge_page_size(h); \
568 if (_sz == PMD_SIZE) \
569 tlb_flush_pmd_range(tlb, address, _sz); \
570 else if (_sz == PUD_SIZE) \
571 tlb_flush_pud_range(tlb, address, _sz); \
572 __tlb_remove_tlb_entry(tlb, ptep, address); \
573 } while (0)
574
575/**
576 * tlb_remove_pmd_tlb_entry - remember a pmd mapping for later tlb invalidation
577 * This is a nop so far, because only x86 needs it.
578 */
579#ifndef __tlb_remove_pmd_tlb_entry
580#define __tlb_remove_pmd_tlb_entry(tlb, pmdp, address) do {} while (0)
581#endif
582
583#define tlb_remove_pmd_tlb_entry(tlb, pmdp, address) \
584 do { \
585 tlb_flush_pmd_range(tlb, address, HPAGE_PMD_SIZE); \
586 __tlb_remove_pmd_tlb_entry(tlb, pmdp, address); \
587 } while (0)
588
589/**
590 * tlb_remove_pud_tlb_entry - remember a pud mapping for later tlb
591 * invalidation. This is a nop so far, because only x86 needs it.
592 */
593#ifndef __tlb_remove_pud_tlb_entry
594#define __tlb_remove_pud_tlb_entry(tlb, pudp, address) do {} while (0)
595#endif
596
597#define tlb_remove_pud_tlb_entry(tlb, pudp, address) \
598 do { \
599 tlb_flush_pud_range(tlb, address, HPAGE_PUD_SIZE); \
600 __tlb_remove_pud_tlb_entry(tlb, pudp, address); \
601 } while (0)
602
603/*
604 * For things like page tables caches (ie caching addresses "inside" the
605 * page tables, like x86 does), for legacy reasons, flushing an
606 * individual page had better flush the page table caches behind it. This
607 * is definitely how x86 works, for example. And if you have an
608 * architected non-legacy page table cache (which I'm not aware of
609 * anybody actually doing), you're going to have some architecturally
610 * explicit flushing for that, likely *separate* from a regular TLB entry
611 * flush, and thus you'd need more than just some range expansion..
612 *
613 * So if we ever find an architecture
614 * that would want something that odd, I think it is up to that
615 * architecture to do its own odd thing, not cause pain for others
616 * http://lkml.kernel.org/r/CA+55aFzBggoXtNXQeng5d_mRoDnaMBE5Y+URs+PHR67nUpMtaw@mail.gmail.com
617 *
618 * For now w.r.t page table cache, mark the range_size as PAGE_SIZE
619 */
620
621#ifndef pte_free_tlb
622#define pte_free_tlb(tlb, ptep, address) \
623 do { \
624 tlb_flush_pmd_range(tlb, address, PAGE_SIZE); \
625 tlb->freed_tables = 1; \
626 __pte_free_tlb(tlb, ptep, address); \
627 } while (0)
628#endif
629
630#ifndef pmd_free_tlb
631#define pmd_free_tlb(tlb, pmdp, address) \
632 do { \
633 tlb_flush_pud_range(tlb, address, PAGE_SIZE); \
634 tlb->freed_tables = 1; \
635 __pmd_free_tlb(tlb, pmdp, address); \
636 } while (0)
637#endif
638
639#ifndef pud_free_tlb
640#define pud_free_tlb(tlb, pudp, address) \
641 do { \
642 tlb_flush_p4d_range(tlb, address, PAGE_SIZE); \
643 tlb->freed_tables = 1; \
644 __pud_free_tlb(tlb, pudp, address); \
645 } while (0)
646#endif
647
648#ifndef p4d_free_tlb
649#define p4d_free_tlb(tlb, pudp, address) \
650 do { \
651 __tlb_adjust_range(tlb, address, PAGE_SIZE); \
652 tlb->freed_tables = 1; \
653 __p4d_free_tlb(tlb, pudp, address); \
654 } while (0)
655#endif
656
657#endif /* CONFIG_MMU */
658
659#endif /* _ASM_GENERIC__TLB_H */