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/mm_inline.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/rcupdate.h>
   9 #include <linux/smp.h>
  10 #include <linux/swap.h>
  11 #include <linux/rmap.h>
  12
  13 #include <asm/pgalloc.h>
  14 #include <asm/tlb.h>
  15
  16 #ifndef CONFIG_MMU_GATHER_NO_GATHER
  17
  18 static bool tlb_next_batch(struct mmu_gather *tlb)
  19 {
  20         struct mmu_gather_batch *batch;
  21
  22         /* Limit batching if we have delayed rmaps pending */
  23         if (tlb->delayed_rmap && tlb->active != &tlb->local)
  24                 return false;
  25
  26         batch = tlb->active;
  27         if (batch->next) {
  28                 tlb->active = batch->next;
  29                 return true;
  30         }
  31
  32         if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
  33                 return false;
  34
  35         batch = (void *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
  36         if (!batch)
  37                 return false;
  38
  39         tlb->batch_count++;
  40         batch->next = NULL;
  41         batch->nr   = 0;
  42         batch->max  = MAX_GATHER_BATCH;
  43
  44         tlb->active->next = batch;
  45         tlb->active = batch;
  46
  47         return true;
  48 }
  49
  50 #ifdef CONFIG_SMP
  51 static void tlb_flush_rmap_batch(struct mmu_gather_batch *batch, struct vm_area_struct *vma)
  52 {
  53         for (int i = 0; i < batch->nr; i++) {
  54                 struct encoded_page *enc = batch->encoded_pages[i];
  55
  56                 if (encoded_page_flags(enc)) {
  57                         struct page *page = encoded_page_ptr(enc);
  58                         page_remove_rmap(page, vma, false);
  59                 }
  60         }
  61 }
  62
  63 /**
  64  * tlb_flush_rmaps - do pending rmap removals after we have flushed the TLB
  65  * @tlb: the current mmu_gather
  66  *
  67  * Note that because of how tlb_next_batch() above works, we will
  68  * never start multiple new batches with pending delayed rmaps, so
  69  * we only need to walk through the current active batch and the
  70  * original local one.
  71  */
  72 void tlb_flush_rmaps(struct mmu_gather *tlb, struct vm_area_struct *vma)
  73 {
  74         if (!tlb->delayed_rmap)
  75                 return;
  76
  77         tlb_flush_rmap_batch(&tlb->local, vma);
  78         if (tlb->active != &tlb->local)
  79                 tlb_flush_rmap_batch(tlb->active, vma);
  80         tlb->delayed_rmap = 0;
  81 }
  82 #endif
  83
  84 static void tlb_batch_pages_flush(struct mmu_gather *tlb)
  85 {
  86         struct mmu_gather_batch *batch;
  87
  88         for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
  89                 struct encoded_page **pages = batch->encoded_pages;
  90
  91                 do {
  92                         /*
  93                          * limit free batch count when PAGE_SIZE > 4K
  94                          */
  95                         unsigned int nr = min(512U, batch->nr);
  96
  97                         free_pages_and_swap_cache(pages, nr);
  98                         pages += nr;
  99                         batch->nr -= nr;
 100
 101                         cond_resched();
 102                 } while (batch->nr);
 103         }
 104         tlb->active = &tlb->local;
 105 }
 106
 107 static void tlb_batch_list_free(struct mmu_gather *tlb)
 108 {
 109         struct mmu_gather_batch *batch, *next;
 110
 111         for (batch = tlb->local.next; batch; batch = next) {
 112                 next = batch->next;
 113                 free_pages((unsigned long)batch, 0);
 114         }
 115         tlb->local.next = NULL;
 116 }
 117
 118 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct encoded_page *page, int page_size)
 119 {
 120         struct mmu_gather_batch *batch;
 121
 122         VM_BUG_ON(!tlb->end);
 123
 124 #ifdef CONFIG_MMU_GATHER_PAGE_SIZE
 125         VM_WARN_ON(tlb->page_size != page_size);
 126 #endif
 127
 128         batch = tlb->active;
 129         /*
 130          * Add the page and check if we are full. If so
 131          * force a flush.
 132          */
 133         batch->encoded_pages[batch->nr++] = page;
 134         if (batch->nr == batch->max) {
 135                 if (!tlb_next_batch(tlb))
 136                         return true;
 137                 batch = tlb->active;
 138         }
 139         VM_BUG_ON_PAGE(batch->nr > batch->max, encoded_page_ptr(page));
 140
 141         return false;
 142 }
 143
 144 #endif /* MMU_GATHER_NO_GATHER */
 145
 146 #ifdef CONFIG_MMU_GATHER_TABLE_FREE
 147
 148 static void __tlb_remove_table_free(struct mmu_table_batch *batch)
 149 {
 150         int i;
 151
 152         for (i = 0; i < batch->nr; i++)
 153                 __tlb_remove_table(batch->tables[i]);
 154
 155         free_page((unsigned long)batch);
 156 }
 157
 158 #ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
 159
 160 /*
 161  * Semi RCU freeing of the page directories.
 162  *
 163  * This is needed by some architectures to implement software pagetable walkers.
 164  *
 165  * gup_fast() and other software pagetable walkers do a lockless page-table
 166  * walk and therefore needs some synchronization with the freeing of the page
 167  * directories. The chosen means to accomplish that is by disabling IRQs over
 168  * the walk.
 169  *
 170  * Architectures that use IPIs to flush TLBs will then automagically DTRT,
 171  * since we unlink the page, flush TLBs, free the page. Since the disabling of
 172  * IRQs delays the completion of the TLB flush we can never observe an already
 173  * freed page.
 174  *
 175  * Architectures that do not have this (PPC) need to delay the freeing by some
 176  * other means, this is that means.
 177  *
 178  * What we do is batch the freed directory pages (tables) and RCU free them.
 179  * We use the sched RCU variant, as that guarantees that IRQ/preempt disabling
 180  * holds off grace periods.
 181  *
 182  * However, in order to batch these pages we need to allocate storage, this
 183  * allocation is deep inside the MM code and can thus easily fail on memory
 184  * pressure. To guarantee progress we fall back to single table freeing, see
 185  * the implementation of tlb_remove_table_one().
 186  *
 187  */
 188
 189 static void tlb_remove_table_smp_sync(void *arg)
 190 {
 191         /* Simply deliver the interrupt */
 192 }
 193
 194 void tlb_remove_table_sync_one(void)
 195 {
 196         /*
 197          * This isn't an RCU grace period and hence the page-tables cannot be
 198          * assumed to be actually RCU-freed.
 199          *
 200          * It is however sufficient for software page-table walkers that rely on
 201          * IRQ disabling.
 202          */
 203         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 204 }
 205
 206 static void tlb_remove_table_rcu(struct rcu_head *head)
 207 {
 208         __tlb_remove_table_free(container_of(head, struct mmu_table_batch, rcu));
 209 }
 210
 211 static void tlb_remove_table_free(struct mmu_table_batch *batch)
 212 {
 213         call_rcu(&batch->rcu, tlb_remove_table_rcu);
 214 }
 215
 216 #else /* !CONFIG_MMU_GATHER_RCU_TABLE_FREE */
 217
 218 static void tlb_remove_table_free(struct mmu_table_batch *batch)
 219 {
 220         __tlb_remove_table_free(batch);
 221 }
 222
 223 #endif /* CONFIG_MMU_GATHER_RCU_TABLE_FREE */
 224
 225 /*
 226  * If we want tlb_remove_table() to imply TLB invalidates.
 227  */
 228 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
 229 {
 230         if (tlb_needs_table_invalidate()) {
 231                 /*
 232                  * Invalidate page-table caches used by hardware walkers. Then
 233                  * we still need to RCU-sched wait while freeing the pages
 234                  * because software walkers can still be in-flight.
 235                  */
 236                 tlb_flush_mmu_tlbonly(tlb);
 237         }
 238 }
 239
 240 static void tlb_remove_table_one(void *table)
 241 {
 242         tlb_remove_table_sync_one();
 243         __tlb_remove_table(table);
 244 }
 245
 246 static void tlb_table_flush(struct mmu_gather *tlb)
 247 {
 248         struct mmu_table_batch **batch = &tlb->batch;
 249
 250         if (*batch) {
 251                 tlb_table_invalidate(tlb);
 252                 tlb_remove_table_free(*batch);
 253                 *batch = NULL;
 254         }
 255 }
 256
 257 void tlb_remove_table(struct mmu_gather *tlb, void *table)
 258 {
 259         struct mmu_table_batch **batch = &tlb->batch;
 260
 261         if (*batch == NULL) {
 262                 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 263                 if (*batch == NULL) {
 264                         tlb_table_invalidate(tlb);
 265                         tlb_remove_table_one(table);
 266                         return;
 267                 }
 268                 (*batch)->nr = 0;
 269         }
 270
 271         (*batch)->tables[(*batch)->nr++] = table;
 272         if ((*batch)->nr == MAX_TABLE_BATCH)
 273                 tlb_table_flush(tlb);
 274 }
 275
 276 static inline void tlb_table_init(struct mmu_gather *tlb)
 277 {
 278         tlb->batch = NULL;
 279 }
 280
 281 #else /* !CONFIG_MMU_GATHER_TABLE_FREE */
 282
 283 static inline void tlb_table_flush(struct mmu_gather *tlb) { }
 284 static inline void tlb_table_init(struct mmu_gather *tlb) { }
 285
 286 #endif /* CONFIG_MMU_GATHER_TABLE_FREE */
 287
 288 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
 289 {
 290         tlb_table_flush(tlb);
 291 #ifndef CONFIG_MMU_GATHER_NO_GATHER
 292         tlb_batch_pages_flush(tlb);
 293 #endif
 294 }
 295
 296 void tlb_flush_mmu(struct mmu_gather *tlb)
 297 {
 298         tlb_flush_mmu_tlbonly(tlb);
 299         tlb_flush_mmu_free(tlb);
 300 }
 301
 302 static void __tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
 303                              bool fullmm)
 304 {
 305         tlb->mm = mm;
 306         tlb->fullmm = fullmm;
 307
 308 #ifndef CONFIG_MMU_GATHER_NO_GATHER
 309         tlb->need_flush_all = 0;
 310         tlb->local.next = NULL;
 311         tlb->local.nr   = 0;
 312         tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 313         tlb->active     = &tlb->local;
 314         tlb->batch_count = 0;
 315 #endif
 316         tlb->delayed_rmap = 0;
 317
 318         tlb_table_init(tlb);
 319 #ifdef CONFIG_MMU_GATHER_PAGE_SIZE
 320         tlb->page_size = 0;
 321 #endif
 322
 323         __tlb_reset_range(tlb);
 324         inc_tlb_flush_pending(tlb->mm);
 325 }
 326
 327 /**
 328  * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
 329  * @tlb: the mmu_gather structure to initialize
 330  * @mm: the mm_struct of the target address space
 331  *
 332  * Called to initialize an (on-stack) mmu_gather structure for page-table
 333  * tear-down from @mm.
 334  */
 335 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm)
 336 {
 337         __tlb_gather_mmu(tlb, mm, false);
 338 }
 339
 340 /**
 341  * tlb_gather_mmu_fullmm - initialize an mmu_gather structure for page-table tear-down
 342  * @tlb: the mmu_gather structure to initialize
 343  * @mm: the mm_struct of the target address space
 344  *
 345  * In this case, @mm is without users and we're going to destroy the
 346  * full address space (exit/execve).
 347  *
 348  * Called to initialize an (on-stack) mmu_gather structure for page-table
 349  * tear-down from @mm.
 350  */
 351 void tlb_gather_mmu_fullmm(struct mmu_gather *tlb, struct mm_struct *mm)
 352 {
 353         __tlb_gather_mmu(tlb, mm, true);
 354 }
 355
 356 /**
 357  * tlb_finish_mmu - finish an mmu_gather structure
 358  * @tlb: the mmu_gather structure to finish
 359  *
 360  * Called at the end of the shootdown operation to free up any resources that
 361  * were required.
 362  */
 363 void tlb_finish_mmu(struct mmu_gather *tlb)
 364 {
 365         /*
 366          * If there are parallel threads are doing PTE changes on same range
 367          * under non-exclusive lock (e.g., mmap_lock read-side) but defer TLB
 368          * flush by batching, one thread may end up seeing inconsistent PTEs
 369          * and result in having stale TLB entries.  So flush TLB forcefully
 370          * if we detect parallel PTE batching threads.
 371          *
 372          * However, some syscalls, e.g. munmap(), may free page tables, this
 373          * needs force flush everything in the given range. Otherwise this
 374          * may result in having stale TLB entries for some architectures,
 375          * e.g. aarch64, that could specify flush what level TLB.
 376          */
 377         if (mm_tlb_flush_nested(tlb->mm)) {
 378                 /*
 379                  * The aarch64 yields better performance with fullmm by
 380                  * avoiding multiple CPUs spamming TLBI messages at the
 381                  * same time.
 382                  *
 383                  * On x86 non-fullmm doesn't yield significant difference
 384                  * against fullmm.
 385                  */
 386                 tlb->fullmm = 1;
 387                 __tlb_reset_range(tlb);
 388                 tlb->freed_tables = 1;
 389         }
 390
 391         tlb_flush_mmu(tlb);
 392
 393 #ifndef CONFIG_MMU_GATHER_NO_GATHER
 394         tlb_batch_list_free(tlb);
 395 #endif
 396         dec_tlb_flush_pending(tlb->mm);
 397 }