mm/highmem.c

   1 /*
   2  * High memory handling common code and variables.
   3  *
   4  * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
   5  *          Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
   6  *
   7  *
   8  * Redesigned the x86 32-bit VM architecture to deal with
   9  * 64-bit physical space. With current x86 CPUs this
  10  * means up to 64 Gigabytes physical RAM.
  11  *
  12  * Rewrote high memory support to move the page cache into
  13  * high memory. Implemented permanent (schedulable) kmaps
  14  * based on Linus' idea.
  15  *
  16  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
  17  */
  18
  19 #include <linux/mm.h>
  20 #include <linux/export.h>
  21 #include <linux/swap.h>
  22 #include <linux/bio.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/mempool.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/init.h>
  27 #include <linux/hash.h>
  28 #include <linux/highmem.h>
  29 #include <linux/kgdb.h>
  30 #include <asm/tlbflush.h>
  31
  32
  33 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
  34 DEFINE_PER_CPU(int, __kmap_atomic_idx);
  35 #endif
  36
  37 /*
  38  * Virtual_count is not a pure "count".
  39  *  0 means that it is not mapped, and has not been mapped
  40  *    since a TLB flush - it is usable.
  41  *  1 means that there are no users, but it has been mapped
  42  *    since the last TLB flush - so we can't use it.
  43  *  n means that there are (n-1) current users of it.
  44  */
  45 #ifdef CONFIG_HIGHMEM
  46
  47 unsigned long totalhigh_pages __read_mostly;
  48 EXPORT_SYMBOL(totalhigh_pages);
  49
  50
  51 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
  52
  53 unsigned int nr_free_highpages (void)
  54 {
  55         pg_data_t *pgdat;
  56         unsigned int pages = 0;
  57
  58         for_each_online_pgdat(pgdat) {
  59                 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
  60                         NR_FREE_PAGES);
  61                 if (zone_movable_is_highmem())
  62                         pages += zone_page_state(
  63                                         &pgdat->node_zones[ZONE_MOVABLE],
  64                                         NR_FREE_PAGES);
  65         }
  66
  67         return pages;
  68 }
  69
  70 static int pkmap_count[LAST_PKMAP];
  71 static unsigned int last_pkmap_nr;
  72 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
  73
  74 pte_t * pkmap_page_table;
  75
  76 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  77
  78 /*
  79  * Most architectures have no use for kmap_high_get(), so let's abstract
  80  * the disabling of IRQ out of the locking in that case to save on a
  81  * potential useless overhead.
  82  */
  83 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
  84 #define lock_kmap()             spin_lock_irq(&kmap_lock)
  85 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
  86 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
  87 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
  88 #else
  89 #define lock_kmap()             spin_lock(&kmap_lock)
  90 #define unlock_kmap()           spin_unlock(&kmap_lock)
  91 #define lock_kmap_any(flags)    \
  92                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
  93 #define unlock_kmap_any(flags)  \
  94                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
  95 #endif
  96
  97 struct page *kmap_to_page(void *vaddr)
  98 {
  99         unsigned long addr = (unsigned long)vaddr;
 100
 101         if (addr >= PKMAP_ADDR(0) && addr <= PKMAP_ADDR(LAST_PKMAP)) {
 102                 int i = (addr - PKMAP_ADDR(0)) >> PAGE_SHIFT;
 103                 return pte_page(pkmap_page_table[i]);
 104         }
 105
 106         return virt_to_page(addr);
 107 }
 108
 109 static void flush_all_zero_pkmaps(void)
 110 {
 111         int i;
 112         int need_flush = 0;
 113
 114         flush_cache_kmaps();
 115
 116         for (i = 0; i < LAST_PKMAP; i++) {
 117                 struct page *page;
 118
 119                 /*
 120                  * zero means we don't have anything to do,
 121                  * >1 means that it is still in use. Only
 122                  * a count of 1 means that it is free but
 123                  * needs to be unmapped
 124                  */
 125                 if (pkmap_count[i] != 1)
 126                         continue;
 127                 pkmap_count[i] = 0;
 128
 129                 /* sanity check */
 130                 BUG_ON(pte_none(pkmap_page_table[i]));
 131
 132                 /*
 133                  * Don't need an atomic fetch-and-clear op here;
 134                  * no-one has the page mapped, and cannot get at
 135                  * its virtual address (and hence PTE) without first
 136                  * getting the kmap_lock (which is held here).
 137                  * So no dangers, even with speculative execution.
 138                  */
 139                 page = pte_page(pkmap_page_table[i]);
 140                 pte_clear(&init_mm, (unsigned long)page_address(page),
 141                           &pkmap_page_table[i]);
 142
 143                 set_page_address(page, NULL);
 144                 need_flush = 1;
 145         }
 146         if (need_flush)
 147                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
 148 }
 149
 150 /**
 151  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
 152  */
 153 void kmap_flush_unused(void)
 154 {
 155         lock_kmap();
 156         flush_all_zero_pkmaps();
 157         unlock_kmap();
 158 }
 159
 160 static inline unsigned long map_new_virtual(struct page *page)
 161 {
 162         unsigned long vaddr;
 163         int count;
 164
 165 start:
 166         count = LAST_PKMAP;
 167         /* Find an empty entry */
 168         for (;;) {
 169                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
 170                 if (!last_pkmap_nr) {
 171                         flush_all_zero_pkmaps();
 172                         count = LAST_PKMAP;
 173                 }
 174                 if (!pkmap_count[last_pkmap_nr])
 175                         break;  /* Found a usable entry */
 176                 if (--count)
 177                         continue;
 178
 179                 /*
 180                  * Sleep for somebody else to unmap their entries
 181                  */
 182                 {
 183                         DECLARE_WAITQUEUE(wait, current);
 184
 185                         __set_current_state(TASK_UNINTERRUPTIBLE);
 186                         add_wait_queue(&pkmap_map_wait, &wait);
 187                         unlock_kmap();
 188                         schedule();
 189                         remove_wait_queue(&pkmap_map_wait, &wait);
 190                         lock_kmap();
 191
 192                         /* Somebody else might have mapped it while we slept */
 193                         if (page_address(page))
 194                                 return (unsigned long)page_address(page);
 195
 196                         /* Re-start */
 197                         goto start;
 198                 }
 199         }
 200         vaddr = PKMAP_ADDR(last_pkmap_nr);
 201         set_pte_at(&init_mm, vaddr,
 202                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 203
 204         pkmap_count[last_pkmap_nr] = 1;
 205         set_page_address(page, (void *)vaddr);
 206
 207         return vaddr;
 208 }
 209
 210 /**
 211  * kmap_high - map a highmem page into memory
 212  * @page: &struct page to map
 213  *
 214  * Returns the page's virtual memory address.
 215  *
 216  * We cannot call this from interrupts, as it may block.
 217  */
 218 void *kmap_high(struct page *page)
 219 {
 220         unsigned long vaddr;
 221
 222         /*
 223          * For highmem pages, we can't trust "virtual" until
 224          * after we have the lock.
 225          */
 226         lock_kmap();
 227         vaddr = (unsigned long)page_address(page);
 228         if (!vaddr)
 229                 vaddr = map_new_virtual(page);
 230         pkmap_count[PKMAP_NR(vaddr)]++;
 231         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
 232         unlock_kmap();
 233         return (void*) vaddr;
 234 }
 235
 236 EXPORT_SYMBOL(kmap_high);
 237
 238 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 239 /**
 240  * kmap_high_get - pin a highmem page into memory
 241  * @page: &struct page to pin
 242  *
 243  * Returns the page's current virtual memory address, or NULL if no mapping
 244  * exists.  If and only if a non null address is returned then a
 245  * matching call to kunmap_high() is necessary.
 246  *
 247  * This can be called from any context.
 248  */
 249 void *kmap_high_get(struct page *page)
 250 {
 251         unsigned long vaddr, flags;
 252
 253         lock_kmap_any(flags);
 254         vaddr = (unsigned long)page_address(page);
 255         if (vaddr) {
 256                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
 257                 pkmap_count[PKMAP_NR(vaddr)]++;
 258         }
 259         unlock_kmap_any(flags);
 260         return (void*) vaddr;
 261 }
 262 #endif
 263
 264 /**
 265  * kunmap_high - unmap a highmem page into memory
 266  * @page: &struct page to unmap
 267  *
 268  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
 269  * only from user context.
 270  */
 271 void kunmap_high(struct page *page)
 272 {
 273         unsigned long vaddr;
 274         unsigned long nr;
 275         unsigned long flags;
 276         int need_wakeup;
 277
 278         lock_kmap_any(flags);
 279         vaddr = (unsigned long)page_address(page);
 280         BUG_ON(!vaddr);
 281         nr = PKMAP_NR(vaddr);
 282
 283         /*
 284          * A count must never go down to zero
 285          * without a TLB flush!
 286          */
 287         need_wakeup = 0;
 288         switch (--pkmap_count[nr]) {
 289         case 0:
 290                 BUG();
 291         case 1:
 292                 /*
 293                  * Avoid an unnecessary wake_up() function call.
 294                  * The common case is pkmap_count[] == 1, but
 295                  * no waiters.
 296                  * The tasks queued in the wait-queue are guarded
 297                  * by both the lock in the wait-queue-head and by
 298                  * the kmap_lock.  As the kmap_lock is held here,
 299                  * no need for the wait-queue-head's lock.  Simply
 300                  * test if the queue is empty.
 301                  */
 302                 need_wakeup = waitqueue_active(&pkmap_map_wait);
 303         }
 304         unlock_kmap_any(flags);
 305
 306         /* do wake-up, if needed, race-free outside of the spin lock */
 307         if (need_wakeup)
 308                 wake_up(&pkmap_map_wait);
 309 }
 310
 311 EXPORT_SYMBOL(kunmap_high);
 312 #endif
 313
 314 #if defined(HASHED_PAGE_VIRTUAL)
 315
 316 #define PA_HASH_ORDER   7
 317
 318 /*
 319  * Describes one page->virtual association
 320  */
 321 struct page_address_map {
 322         struct page *page;
 323         void *virtual;
 324         struct list_head list;
 325 };
 326
 327 /*
 328  * page_address_map freelist, allocated from page_address_maps.
 329  */
 330 static struct list_head page_address_pool;      /* freelist */
 331 static spinlock_t pool_lock;                    /* protects page_address_pool */
 332
 333 /*
 334  * Hash table bucket
 335  */
 336 static struct page_address_slot {
 337         struct list_head lh;                    /* List of page_address_maps */
 338         spinlock_t lock;                        /* Protect this bucket's list */
 339 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 340
 341 static struct page_address_slot *page_slot(const struct page *page)
 342 {
 343         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 344 }
 345
 346 /**
 347  * page_address - get the mapped virtual address of a page
 348  * @page: &struct page to get the virtual address of
 349  *
 350  * Returns the page's virtual address.
 351  */
 352 void *page_address(const struct page *page)
 353 {
 354         unsigned long flags;
 355         void *ret;
 356         struct page_address_slot *pas;
 357
 358         if (!PageHighMem(page))
 359                 return lowmem_page_address(page);
 360
 361         pas = page_slot(page);
 362         ret = NULL;
 363         spin_lock_irqsave(&pas->lock, flags);
 364         if (!list_empty(&pas->lh)) {
 365                 struct page_address_map *pam;
 366
 367                 list_for_each_entry(pam, &pas->lh, list) {
 368                         if (pam->page == page) {
 369                                 ret = pam->virtual;
 370                                 goto done;
 371                         }
 372                 }
 373         }
 374 done:
 375         spin_unlock_irqrestore(&pas->lock, flags);
 376         return ret;
 377 }
 378
 379 EXPORT_SYMBOL(page_address);
 380
 381 /**
 382  * set_page_address - set a page's virtual address
 383  * @page: &struct page to set
 384  * @virtual: virtual address to use
 385  */
 386 void set_page_address(struct page *page, void *virtual)
 387 {
 388         unsigned long flags;
 389         struct page_address_slot *pas;
 390         struct page_address_map *pam;
 391
 392         BUG_ON(!PageHighMem(page));
 393
 394         pas = page_slot(page);
 395         if (virtual) {          /* Add */
 396                 BUG_ON(list_empty(&page_address_pool));
 397
 398                 spin_lock_irqsave(&pool_lock, flags);
 399                 pam = list_entry(page_address_pool.next,
 400                                 struct page_address_map, list);
 401                 list_del(&pam->list);
 402                 spin_unlock_irqrestore(&pool_lock, flags);
 403
 404                 pam->page = page;
 405                 pam->virtual = virtual;
 406
 407                 spin_lock_irqsave(&pas->lock, flags);
 408                 list_add_tail(&pam->list, &pas->lh);
 409                 spin_unlock_irqrestore(&pas->lock, flags);
 410         } else {                /* Remove */
 411                 spin_lock_irqsave(&pas->lock, flags);
 412                 list_for_each_entry(pam, &pas->lh, list) {
 413                         if (pam->page == page) {
 414                                 list_del(&pam->list);
 415                                 spin_unlock_irqrestore(&pas->lock, flags);
 416                                 spin_lock_irqsave(&pool_lock, flags);
 417                                 list_add_tail(&pam->list, &page_address_pool);
 418                                 spin_unlock_irqrestore(&pool_lock, flags);
 419                                 goto done;
 420                         }
 421                 }
 422                 spin_unlock_irqrestore(&pas->lock, flags);
 423         }
 424 done:
 425         return;
 426 }
 427
 428 static struct page_address_map page_address_maps[LAST_PKMAP];
 429
 430 void __init page_address_init(void)
 431 {
 432         int i;
 433
 434         INIT_LIST_HEAD(&page_address_pool);
 435         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
 436                 list_add(&page_address_maps[i].list, &page_address_pool);
 437         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 438                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 439                 spin_lock_init(&page_address_htable[i].lock);
 440         }
 441         spin_lock_init(&pool_lock);
 442 }
 443
 444 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */