Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
v5.4
  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 */
v4.10.11
 
  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 */
 51struct 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
 60extern void tlb_table_flush(struct mmu_gather *tlb);
 61extern 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
 71struct 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 */
 92struct 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
115void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end);
116void tlb_flush_mmu(struct mmu_gather *tlb);
117void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start,
118							unsigned long end);
119extern bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page,
120				   int page_size);
121
122static inline void __tlb_adjust_range(struct mmu_gather *tlb,
123				      unsigned long address,
124				      unsigned int range_size)
125{
126	tlb->start = min(tlb->start, address);
127	tlb->end = max(tlb->end, address + range_size);
128}
129
130static inline void __tlb_reset_range(struct mmu_gather *tlb)
131{
132	if (tlb->fullmm) {
133		tlb->start = tlb->end = ~0;
134	} else {
135		tlb->start = TASK_SIZE;
136		tlb->end = 0;
137	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
138}
139
140static inline void tlb_remove_page_size(struct mmu_gather *tlb,
141					struct page *page, int page_size)
142{
143	if (__tlb_remove_page_size(tlb, page, page_size))
144		tlb_flush_mmu(tlb);
145}
146
147static inline bool __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
148{
149	return __tlb_remove_page_size(tlb, page, PAGE_SIZE);
150}
151
152/* tlb_remove_page
153 *	Similar to __tlb_remove_page but will call tlb_flush_mmu() itself when
154 *	required.
155 */
156static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
157{
158	return tlb_remove_page_size(tlb, page, PAGE_SIZE);
159}
160
161#ifndef tlb_remove_check_page_size_change
162#define tlb_remove_check_page_size_change tlb_remove_check_page_size_change
163static inline void tlb_remove_check_page_size_change(struct mmu_gather *tlb,
164						     unsigned int page_size)
165{
166	/*
167	 * We don't care about page size change, just update
168	 * mmu_gather page size here so that debug checks
169	 * doesn't throw false warning.
170	 */
171#ifdef CONFIG_DEBUG_VM
172	tlb->page_size = page_size;
173#endif
174}
175#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
176
177/*
178 * In the case of tlb vma handling, we can optimise these away in the
179 * case where we're doing a full MM flush.  When we're doing a munmap,
180 * the vmas are adjusted to only cover the region to be torn down.
181 */
182#ifndef tlb_start_vma
183#define tlb_start_vma(tlb, vma) do { } while (0)
 
 
 
 
 
 
 
184#endif
185
186#define __tlb_end_vma(tlb, vma)					\
187	do {							\
188		if (!tlb->fullmm && tlb->end) {			\
189			tlb_flush(tlb);				\
190			__tlb_reset_range(tlb);			\
191		}						\
192	} while (0)
193
194#ifndef tlb_end_vma
195#define tlb_end_vma	__tlb_end_vma
 
 
 
 
 
 
196#endif
197
198#ifndef __tlb_remove_tlb_entry
199#define __tlb_remove_tlb_entry(tlb, ptep, address) do { } while (0)
200#endif
201
202/**
203 * tlb_remove_tlb_entry - remember a pte unmapping for later tlb invalidation.
204 *
205 * Record the fact that pte's were really unmapped by updating the range,
206 * so we can later optimise away the tlb invalidate.   This helps when
207 * userspace is unmapping already-unmapped pages, which happens quite a lot.
208 */
209#define tlb_remove_tlb_entry(tlb, ptep, address)		\
210	do {							\
211		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
 
212		__tlb_remove_tlb_entry(tlb, ptep, address);	\
213	} while (0)
214
215#define tlb_remove_huge_tlb_entry(h, tlb, ptep, address)	     \
216	do {							     \
217		__tlb_adjust_range(tlb, address, huge_page_size(h)); \
218		__tlb_remove_tlb_entry(tlb, ptep, address);	     \
 
 
 
 
 
219	} while (0)
220
221/**
222 * tlb_remove_pmd_tlb_entry - remember a pmd mapping for later tlb invalidation
223 * This is a nop so far, because only x86 needs it.
224 */
225#ifndef __tlb_remove_pmd_tlb_entry
226#define __tlb_remove_pmd_tlb_entry(tlb, pmdp, address) do {} while (0)
227#endif
228
229#define tlb_remove_pmd_tlb_entry(tlb, pmdp, address)			\
230	do {								\
231		__tlb_adjust_range(tlb, address, HPAGE_PMD_SIZE);	\
 
232		__tlb_remove_pmd_tlb_entry(tlb, pmdp, address);		\
233	} while (0)
234
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
235/*
236 * For things like page tables caches (ie caching addresses "inside" the
237 * page tables, like x86 does), for legacy reasons, flushing an
238 * individual page had better flush the page table caches behind it. This
239 * is definitely how x86 works, for example. And if you have an
240 * architected non-legacy page table cache (which I'm not aware of
241 * anybody actually doing), you're going to have some architecturally
242 * explicit flushing for that, likely *separate* from a regular TLB entry
243 * flush, and thus you'd need more than just some range expansion..
244 *
245 * So if we ever find an architecture
246 * that would want something that odd, I think it is up to that
247 * architecture to do its own odd thing, not cause pain for others
248 * http://lkml.kernel.org/r/CA+55aFzBggoXtNXQeng5d_mRoDnaMBE5Y+URs+PHR67nUpMtaw@mail.gmail.com
249 *
250 * For now w.r.t page table cache, mark the range_size as PAGE_SIZE
251 */
252
 
253#define pte_free_tlb(tlb, ptep, address)			\
254	do {							\
255		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
 
 
256		__pte_free_tlb(tlb, ptep, address);		\
257	} while (0)
 
 
 
 
 
 
 
 
 
 
 
258
259#ifndef __ARCH_HAS_4LEVEL_HACK
 
260#define pud_free_tlb(tlb, pudp, address)			\
261	do {							\
262		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
 
 
263		__pud_free_tlb(tlb, pudp, address);		\
264	} while (0)
265#endif
 
266
267#define pmd_free_tlb(tlb, pmdp, address)			\
 
 
268	do {							\
269		__tlb_adjust_range(tlb, address, PAGE_SIZE);	\
270		__pmd_free_tlb(tlb, pmdp, address);		\
 
271	} while (0)
 
 
272
273#define tlb_migrate_finish(mm) do {} while (0)
274
275#endif /* _ASM_GENERIC__TLB_H */