include/asm-generic/tlb.h

   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 <pzijlstr@redhat.com>
   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 /* struct mmu_gather is an opaque type used by the mm code for passing around
  82  * any data needed by arch specific code for tlb_remove_page.
  83  */
  84 struct mmu_gather {
  85         struct mm_struct        *mm;
  86 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
  87         struct mmu_table_batch  *batch;
  88 #endif
  89         unsigned int            need_flush : 1, /* Did free PTEs */
  90                                 fast_mode  : 1; /* No batching   */
  91
  92         unsigned int            fullmm;
  93
  94         struct mmu_gather_batch *active;
  95         struct mmu_gather_batch local;
  96         struct page             *__pages[MMU_GATHER_BUNDLE];
  97 };
  98
  99 /*
 100  * For UP we don't need to worry about TLB flush
 101  * and page free order so much..
 102  */
 103 #ifdef CONFIG_SMP
 104   #define tlb_fast_mode(tlb) (tlb->fast_mode)
 105 #else
 106   #define tlb_fast_mode(tlb) 1
 107 #endif
 108
 109 static inline int tlb_next_batch(struct mmu_gather *tlb)
 110 {
 111         struct mmu_gather_batch *batch;
 112
 113         batch = tlb->active;
 114         if (batch->next) {
 115                 tlb->active = batch->next;
 116                 return 1;
 117         }
 118
 119         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 120         if (!batch)
 121                 return 0;
 122
 123         batch->next = NULL;
 124         batch->nr   = 0;
 125         batch->max  = MAX_GATHER_BATCH;
 126
 127         tlb->active->next = batch;
 128         tlb->active = batch;
 129
 130         return 1;
 131 }
 132
 133 /* tlb_gather_mmu
 134  *      Called to initialize an (on-stack) mmu_gather structure for page-table
 135  *      tear-down from @mm. The @fullmm argument is used when @mm is without
 136  *      users and we're going to destroy the full address space (exit/execve).
 137  */
 138 static inline void
 139 tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
 140 {
 141         tlb->mm = mm;
 142
 143         tlb->fullmm     = fullmm;
 144         tlb->need_flush = 0;
 145         tlb->fast_mode  = (num_possible_cpus() == 1);
 146         tlb->local.next = NULL;
 147         tlb->local.nr   = 0;
 148         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 149         tlb->active     = &tlb->local;
 150
 151 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
 152         tlb->batch = NULL;
 153 #endif
 154 }
 155
 156 static inline void
 157 tlb_flush_mmu(struct mmu_gather *tlb)
 158 {
 159         struct mmu_gather_batch *batch;
 160
 161         if (!tlb->need_flush)
 162                 return;
 163         tlb->need_flush = 0;
 164         tlb_flush(tlb);
 165 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
 166         tlb_table_flush(tlb);
 167 #endif
 168
 169         if (tlb_fast_mode(tlb))
 170                 return;
 171
 172         for (batch = &tlb->local; batch; batch = batch->next) {
 173                 free_pages_and_swap_cache(batch->pages, batch->nr);
 174                 batch->nr = 0;
 175         }
 176         tlb->active = &tlb->local;
 177 }
 178
 179 /* tlb_finish_mmu
 180  *      Called at the end of the shootdown operation to free up any resources
 181  *      that were required.
 182  */
 183 static inline void
 184 tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 185 {
 186         struct mmu_gather_batch *batch, *next;
 187
 188         tlb_flush_mmu(tlb);
 189
 190         /* keep the page table cache within bounds */
 191         check_pgt_cache();
 192
 193         for (batch = tlb->local.next; batch; batch = next) {
 194                 next = batch->next;
 195                 free_pages((unsigned long)batch, 0);
 196         }
 197         tlb->local.next = NULL;
 198 }
 199
 200 /* __tlb_remove_page
 201  *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 202  *      handling the additional races in SMP caused by other CPUs caching valid
 203  *      mappings in their TLBs. Returns the number of free page slots left.
 204  *      When out of page slots we must call tlb_flush_mmu().
 205  */
 206 static inline int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 207 {
 208         struct mmu_gather_batch *batch;
 209
 210         tlb->need_flush = 1;
 211
 212         if (tlb_fast_mode(tlb)) {
 213                 free_page_and_swap_cache(page);
 214                 return 1; /* avoid calling tlb_flush_mmu() */
 215         }
 216
 217         batch = tlb->active;
 218         batch->pages[batch->nr++] = page;
 219         VM_BUG_ON(batch->nr > batch->max);
 220         if (batch->nr == batch->max) {
 221                 if (!tlb_next_batch(tlb))
 222                         return 0;
 223         }
 224
 225         return batch->max - batch->nr;
 226 }
 227
 228 /* tlb_remove_page
 229  *      Similar to __tlb_remove_page but will call tlb_flush_mmu() itself when
 230  *      required.
 231  */
 232 static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 233 {
 234         if (!__tlb_remove_page(tlb, page))
 235                 tlb_flush_mmu(tlb);
 236 }
 237
 238 /**
 239  * tlb_remove_tlb_entry - remember a pte unmapping for later tlb invalidation.
 240  *
 241  * Record the fact that pte's were really umapped in ->need_flush, so we can
 242  * later optimise away the tlb invalidate.   This helps when userspace is
 243  * unmapping already-unmapped pages, which happens quite a lot.
 244  */
 245 #define tlb_remove_tlb_entry(tlb, ptep, address)                \
 246         do {                                                    \
 247                 tlb->need_flush = 1;                            \
 248                 __tlb_remove_tlb_entry(tlb, ptep, address);     \
 249         } while (0)
 250
 251 #define pte_free_tlb(tlb, ptep, address)                        \
 252         do {                                                    \
 253                 tlb->need_flush = 1;                            \
 254                 __pte_free_tlb(tlb, ptep, address);             \
 255         } while (0)
 256
 257 #ifndef __ARCH_HAS_4LEVEL_HACK
 258 #define pud_free_tlb(tlb, pudp, address)                        \
 259         do {                                                    \
 260                 tlb->need_flush = 1;                            \
 261                 __pud_free_tlb(tlb, pudp, address);             \
 262         } while (0)
 263 #endif
 264
 265 #define pmd_free_tlb(tlb, pmdp, address)                        \
 266         do {                                                    \
 267                 tlb->need_flush = 1;                            \
 268                 __pmd_free_tlb(tlb, pmdp, address);             \
 269         } while (0)
 270
 271 #define tlb_migrate_finish(mm) do {} while (0)
 272
 273 #endif /* _ASM_GENERIC__TLB_H */