mm/mmu_gather.c

   1 #include <linux/gfp.h>
   2 #include <linux/highmem.h>
   3 #include <linux/kernel.h>
   4 #include <linux/mmdebug.h>
   5 #include <linux/mm_types.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/rcupdate.h>
   8 #include <linux/smp.h>
   9 #include <linux/swap.h>
  10
  11 #include <asm/pgalloc.h>
  12 #include <asm/tlb.h>
  13
  14 #ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER
  15
  16 static bool tlb_next_batch(struct mmu_gather *tlb)
  17 {
  18         struct mmu_gather_batch *batch;
  19
  20         batch = tlb->active;
  21         if (batch->next) {
  22                 tlb->active = batch->next;
  23                 return true;
  24         }
  25
  26         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
  27                 return false;
  28
  29         batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
  30         if (!batch)
  31                 return false;
  32
  33         tlb->batch_count++;
  34         batch->next = NULL;
  35         batch->nr   = 0;
  36         batch->max  = MAX_GATHER_BATCH;
  37
  38         tlb->active->next = batch;
  39         tlb->active = batch;
  40
  41         return true;
  42 }
  43
  44 static void tlb_batch_pages_flush(struct mmu_gather *tlb)
  45 {
  46         struct mmu_gather_batch *batch;
  47
  48         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
  49                 free_pages_and_swap_cache(batch->pages, batch->nr);
  50                 batch->nr = 0;
  51         }
  52         tlb->active = &tlb->local;
  53 }
  54
  55 static void tlb_batch_list_free(struct mmu_gather *tlb)
  56 {
  57         struct mmu_gather_batch *batch, *next;
  58
  59         for (batch = tlb->local.next; batch; batch = next) {
  60                 next = batch->next;
  61                 free_pages((unsigned long)batch, 0);
  62         }
  63         tlb->local.next = NULL;
  64 }
  65
  66 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
  67 {
  68         struct mmu_gather_batch *batch;
  69
  70         VM_BUG_ON(!tlb->end);
  71
  72 #ifdef CONFIG_HAVE_MMU_GATHER_PAGE_SIZE
  73         VM_WARN_ON(tlb->page_size != page_size);
  74 #endif
  75
  76         batch = tlb->active;
  77         /*
  78          * Add the page and check if we are full. If so
  79          * force a flush.
  80          */
  81         batch->pages[batch->nr++] = page;
  82         if (batch->nr == batch->max) {
  83                 if (!tlb_next_batch(tlb))
  84                         return true;
  85                 batch = tlb->active;
  86         }
  87         VM_BUG_ON_PAGE(batch->nr > batch->max, page);
  88
  89         return false;
  90 }
  91
  92 #endif /* HAVE_MMU_GATHER_NO_GATHER */
  93
  94 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
  95
  96 /*
  97  * See the comment near struct mmu_table_batch.
  98  */
  99
 100 /*
 101  * If we want tlb_remove_table() to imply TLB invalidates.
 102  */
 103 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
 104 {
 105 #ifndef CONFIG_HAVE_RCU_TABLE_NO_INVALIDATE
 106         /*
 107          * Invalidate page-table caches used by hardware walkers. Then we still
 108          * need to RCU-sched wait while freeing the pages because software
 109          * walkers can still be in-flight.
 110          */
 111         tlb_flush_mmu_tlbonly(tlb);
 112 #endif
 113 }
 114
 115 static void tlb_remove_table_smp_sync(void *arg)
 116 {
 117         /* Simply deliver the interrupt */
 118 }
 119
 120 static void tlb_remove_table_one(void *table)
 121 {
 122         /*
 123          * This isn't an RCU grace period and hence the page-tables cannot be
 124          * assumed to be actually RCU-freed.
 125          *
 126          * It is however sufficient for software page-table walkers that rely on
 127          * IRQ disabling. See the comment near struct mmu_table_batch.
 128          */
 129         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 130         __tlb_remove_table(table);
 131 }
 132
 133 static void tlb_remove_table_rcu(struct rcu_head *head)
 134 {
 135         struct mmu_table_batch *batch;
 136         int i;
 137
 138         batch = container_of(head, struct mmu_table_batch, rcu);
 139
 140         for (i = 0; i < batch->nr; i++)
 141                 __tlb_remove_table(batch->tables[i]);
 142
 143         free_page((unsigned long)batch);
 144 }
 145
 146 static void tlb_table_flush(struct mmu_gather *tlb)
 147 {
 148         struct mmu_table_batch **batch = &tlb->batch;
 149
 150         if (*batch) {
 151                 tlb_table_invalidate(tlb);
 152                 call_rcu(&(*batch)->rcu, tlb_remove_table_rcu);
 153                 *batch = NULL;
 154         }
 155 }
 156
 157 void tlb_remove_table(struct mmu_gather *tlb, void *table)
 158 {
 159         struct mmu_table_batch **batch = &tlb->batch;
 160
 161         if (*batch == NULL) {
 162                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 163                 if (*batch == NULL) {
 164                         tlb_table_invalidate(tlb);
 165                         tlb_remove_table_one(table);
 166                         return;
 167                 }
 168                 (*batch)->nr = 0;
 169         }
 170
 171         (*batch)->tables[(*batch)->nr++] = table;
 172         if ((*batch)->nr == MAX_TABLE_BATCH)
 173                 tlb_table_flush(tlb);
 174 }
 175
 176 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 177
 178 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
 179 {
 180 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
 181         tlb_table_flush(tlb);
 182 #endif
 183 #ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER
 184         tlb_batch_pages_flush(tlb);
 185 #endif
 186 }
 187
 188 void tlb_flush_mmu(struct mmu_gather *tlb)
 189 {
 190         tlb_flush_mmu_tlbonly(tlb);
 191         tlb_flush_mmu_free(tlb);
 192 }
 193
 194 /**
 195  * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
 196  * @tlb: the mmu_gather structure to initialize
 197  * @mm: the mm_struct of the target address space
 198  * @start: start of the region that will be removed from the page-table
 199  * @end: end of the region that will be removed from the page-table
 200  *
 201  * Called to initialize an (on-stack) mmu_gather structure for page-table
 202  * tear-down from @mm. The @start and @end are set to 0 and -1
 203  * respectively when @mm is without users and we're going to destroy
 204  * the full address space (exit/execve).
 205  */
 206 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
 207                         unsigned long start, unsigned long end)
 208 {
 209         tlb->mm = mm;
 210
 211         /* Is it from 0 to ~0? */
 212         tlb->fullmm     = !(start | (end+1));
 213
 214 #ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER
 215         tlb->need_flush_all = 0;
 216         tlb->local.next = NULL;
 217         tlb->local.nr   = 0;
 218         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 219         tlb->active     = &tlb->local;
 220         tlb->batch_count = 0;
 221 #endif
 222
 223 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
 224         tlb->batch = NULL;
 225 #endif
 226 #ifdef CONFIG_HAVE_MMU_GATHER_PAGE_SIZE
 227         tlb->page_size = 0;
 228 #endif
 229
 230         __tlb_reset_range(tlb);
 231         inc_tlb_flush_pending(tlb->mm);
 232 }
 233
 234 /**
 235  * tlb_finish_mmu - finish an mmu_gather structure
 236  * @tlb: the mmu_gather structure to finish
 237  * @start: start of the region that will be removed from the page-table
 238  * @end: end of the region that will be removed from the page-table
 239  *
 240  * Called at the end of the shootdown operation to free up any resources that
 241  * were required.
 242  */
 243 void tlb_finish_mmu(struct mmu_gather *tlb,
 244                 unsigned long start, unsigned long end)
 245 {
 246         /*
 247          * If there are parallel threads are doing PTE changes on same range
 248          * under non-exclusive lock (e.g., mmap_sem read-side) but defer TLB
 249          * flush by batching, one thread may end up seeing inconsistent PTEs
 250          * and result in having stale TLB entries.  So flush TLB forcefully
 251          * if we detect parallel PTE batching threads.
 252          *
 253          * However, some syscalls, e.g. munmap(), may free page tables, this
 254          * needs force flush everything in the given range. Otherwise this
 255          * may result in having stale TLB entries for some architectures,
 256          * e.g. aarch64, that could specify flush what level TLB.
 257          */
 258         if (mm_tlb_flush_nested(tlb->mm)) {
 259                 /*
 260                  * The aarch64 yields better performance with fullmm by
 261                  * avoiding multiple CPUs spamming TLBI messages at the
 262                  * same time.
 263                  *
 264                  * On x86 non-fullmm doesn't yield significant difference
 265                  * against fullmm.
 266                  */
 267                 tlb->fullmm = 1;
 268                 __tlb_reset_range(tlb);
 269                 tlb->freed_tables = 1;
 270         }
 271
 272         tlb_flush_mmu(tlb);
 273
 274 #ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER
 275         tlb_batch_list_free(tlb);
 276 #endif
 277         dec_tlb_flush_pending(tlb->mm);
 278 }