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 EXPORT_SYMBOL(kmap_to_page);
 109
 110 static void flush_all_zero_pkmaps(void)
 111 {
 112         int i;
 113         int need_flush = 0;
 114
 115         flush_cache_kmaps();
 116
 117         for (i = 0; i < LAST_PKMAP; i++) {
 118                 struct page *page;
 119
 120                 /*
 121                  * zero means we don't have anything to do,
 122                  * >1 means that it is still in use. Only
 123                  * a count of 1 means that it is free but
 124                  * needs to be unmapped
 125                  */
 126                 if (pkmap_count[i] != 1)
 127                         continue;
 128                 pkmap_count[i] = 0;
 129
 130                 /* sanity check */
 131                 BUG_ON(pte_none(pkmap_page_table[i]));
 132
 133                 /*
 134                  * Don't need an atomic fetch-and-clear op here;
 135                  * no-one has the page mapped, and cannot get at
 136                  * its virtual address (and hence PTE) without first
 137                  * getting the kmap_lock (which is held here).
 138                  * So no dangers, even with speculative execution.
 139                  */
 140                 page = pte_page(pkmap_page_table[i]);
 141                 pte_clear(&init_mm, (unsigned long)page_address(page),
 142                           &pkmap_page_table[i]);
 143
 144                 set_page_address(page, NULL);
 145                 need_flush = 1;
 146         }
 147         if (need_flush)
 148                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
 149 }
 150
 151 /**
 152  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
 153  */
 154 void kmap_flush_unused(void)
 155 {
 156         lock_kmap();
 157         flush_all_zero_pkmaps();
 158         unlock_kmap();
 159 }
 160
 161 static inline unsigned long map_new_virtual(struct page *page)
 162 {
 163         unsigned long vaddr;
 164         int count;
 165
 166 start:
 167         count = LAST_PKMAP;
 168         /* Find an empty entry */
 169         for (;;) {
 170                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
 171                 if (!last_pkmap_nr) {
 172                         flush_all_zero_pkmaps();
 173                         count = LAST_PKMAP;
 174                 }
 175                 if (!pkmap_count[last_pkmap_nr])
 176                         break;  /* Found a usable entry */
 177                 if (--count)
 178                         continue;
 179
 180                 /*
 181                  * Sleep for somebody else to unmap their entries
 182                  */
 183                 {
 184                         DECLARE_WAITQUEUE(wait, current);
 185
 186                         __set_current_state(TASK_UNINTERRUPTIBLE);
 187                         add_wait_queue(&pkmap_map_wait, &wait);
 188                         unlock_kmap();
 189                         schedule();
 190                         remove_wait_queue(&pkmap_map_wait, &wait);
 191                         lock_kmap();
 192
 193                         /* Somebody else might have mapped it while we slept */
 194                         if (page_address(page))
 195                                 return (unsigned long)page_address(page);
 196
 197                         /* Re-start */
 198                         goto start;
 199                 }
 200         }
 201         vaddr = PKMAP_ADDR(last_pkmap_nr);
 202         set_pte_at(&init_mm, vaddr,
 203                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 204
 205         pkmap_count[last_pkmap_nr] = 1;
 206         set_page_address(page, (void *)vaddr);
 207
 208         return vaddr;
 209 }
 210
 211 /**
 212  * kmap_high - map a highmem page into memory
 213  * @page: &struct page to map
 214  *
 215  * Returns the page's virtual memory address.
 216  *
 217  * We cannot call this from interrupts, as it may block.
 218  */
 219 void *kmap_high(struct page *page)
 220 {
 221         unsigned long vaddr;
 222
 223         /*
 224          * For highmem pages, we can't trust "virtual" until
 225          * after we have the lock.
 226          */
 227         lock_kmap();
 228         vaddr = (unsigned long)page_address(page);
 229         if (!vaddr)
 230                 vaddr = map_new_virtual(page);
 231         pkmap_count[PKMAP_NR(vaddr)]++;
 232         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
 233         unlock_kmap();
 234         return (void*) vaddr;
 235 }
 236
 237 EXPORT_SYMBOL(kmap_high);
 238
 239 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 240 /**
 241  * kmap_high_get - pin a highmem page into memory
 242  * @page: &struct page to pin
 243  *
 244  * Returns the page's current virtual memory address, or NULL if no mapping
 245  * exists.  If and only if a non null address is returned then a
 246  * matching call to kunmap_high() is necessary.
 247  *
 248  * This can be called from any context.
 249  */
 250 void *kmap_high_get(struct page *page)
 251 {
 252         unsigned long vaddr, flags;
 253
 254         lock_kmap_any(flags);
 255         vaddr = (unsigned long)page_address(page);
 256         if (vaddr) {
 257                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
 258                 pkmap_count[PKMAP_NR(vaddr)]++;
 259         }
 260         unlock_kmap_any(flags);
 261         return (void*) vaddr;
 262 }
 263 #endif
 264
 265 /**
 266  * kunmap_high - unmap a highmem page into memory
 267  * @page: &struct page to unmap
 268  *
 269  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
 270  * only from user context.
 271  */
 272 void kunmap_high(struct page *page)
 273 {
 274         unsigned long vaddr;
 275         unsigned long nr;
 276         unsigned long flags;
 277         int need_wakeup;
 278
 279         lock_kmap_any(flags);
 280         vaddr = (unsigned long)page_address(page);
 281         BUG_ON(!vaddr);
 282         nr = PKMAP_NR(vaddr);
 283
 284         /*
 285          * A count must never go down to zero
 286          * without a TLB flush!
 287          */
 288         need_wakeup = 0;
 289         switch (--pkmap_count[nr]) {
 290         case 0:
 291                 BUG();
 292         case 1:
 293                 /*
 294                  * Avoid an unnecessary wake_up() function call.
 295                  * The common case is pkmap_count[] == 1, but
 296                  * no waiters.
 297                  * The tasks queued in the wait-queue are guarded
 298                  * by both the lock in the wait-queue-head and by
 299                  * the kmap_lock.  As the kmap_lock is held here,
 300                  * no need for the wait-queue-head's lock.  Simply
 301                  * test if the queue is empty.
 302                  */
 303                 need_wakeup = waitqueue_active(&pkmap_map_wait);
 304         }
 305         unlock_kmap_any(flags);
 306
 307         /* do wake-up, if needed, race-free outside of the spin lock */
 308         if (need_wakeup)
 309                 wake_up(&pkmap_map_wait);
 310 }
 311
 312 EXPORT_SYMBOL(kunmap_high);
 313 #endif
 314
 315 #if defined(HASHED_PAGE_VIRTUAL)
 316
 317 #define PA_HASH_ORDER   7
 318
 319 /*
 320  * Describes one page->virtual association
 321  */
 322 struct page_address_map {
 323         struct page *page;
 324         void *virtual;
 325         struct list_head list;
 326 };
 327
 328 /*
 329  * page_address_map freelist, allocated from page_address_maps.
 330  */
 331 static struct list_head page_address_pool;      /* freelist */
 332 static spinlock_t pool_lock;                    /* protects page_address_pool */
 333
 334 /*
 335  * Hash table bucket
 336  */
 337 static struct page_address_slot {
 338         struct list_head lh;                    /* List of page_address_maps */
 339         spinlock_t lock;                        /* Protect this bucket's list */
 340 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 341
 342 static struct page_address_slot *page_slot(const struct page *page)
 343 {
 344         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 345 }
 346
 347 /**
 348  * page_address - get the mapped virtual address of a page
 349  * @page: &struct page to get the virtual address of
 350  *
 351  * Returns the page's virtual address.
 352  */
 353 void *page_address(const struct page *page)
 354 {
 355         unsigned long flags;
 356         void *ret;
 357         struct page_address_slot *pas;
 358
 359         if (!PageHighMem(page))
 360                 return lowmem_page_address(page);
 361
 362         pas = page_slot(page);
 363         ret = NULL;
 364         spin_lock_irqsave(&pas->lock, flags);
 365         if (!list_empty(&pas->lh)) {
 366                 struct page_address_map *pam;
 367
 368                 list_for_each_entry(pam, &pas->lh, list) {
 369                         if (pam->page == page) {
 370                                 ret = pam->virtual;
 371                                 goto done;
 372                         }
 373                 }
 374         }
 375 done:
 376         spin_unlock_irqrestore(&pas->lock, flags);
 377         return ret;
 378 }
 379
 380 EXPORT_SYMBOL(page_address);
 381
 382 /**
 383  * set_page_address - set a page's virtual address
 384  * @page: &struct page to set
 385  * @virtual: virtual address to use
 386  */
 387 void set_page_address(struct page *page, void *virtual)
 388 {
 389         unsigned long flags;
 390         struct page_address_slot *pas;
 391         struct page_address_map *pam;
 392
 393         BUG_ON(!PageHighMem(page));
 394
 395         pas = page_slot(page);
 396         if (virtual) {          /* Add */
 397                 BUG_ON(list_empty(&page_address_pool));
 398
 399                 spin_lock_irqsave(&pool_lock, flags);
 400                 pam = list_entry(page_address_pool.next,
 401                                 struct page_address_map, list);
 402                 list_del(&pam->list);
 403                 spin_unlock_irqrestore(&pool_lock, flags);
 404
 405                 pam->page = page;
 406                 pam->virtual = virtual;
 407
 408                 spin_lock_irqsave(&pas->lock, flags);
 409                 list_add_tail(&pam->list, &pas->lh);
 410                 spin_unlock_irqrestore(&pas->lock, flags);
 411         } else {                /* Remove */
 412                 spin_lock_irqsave(&pas->lock, flags);
 413                 list_for_each_entry(pam, &pas->lh, list) {
 414                         if (pam->page == page) {
 415                                 list_del(&pam->list);
 416                                 spin_unlock_irqrestore(&pas->lock, flags);
 417                                 spin_lock_irqsave(&pool_lock, flags);
 418                                 list_add_tail(&pam->list, &page_address_pool);
 419                                 spin_unlock_irqrestore(&pool_lock, flags);
 420                                 goto done;
 421                         }
 422                 }
 423                 spin_unlock_irqrestore(&pas->lock, flags);
 424         }
 425 done:
 426         return;
 427 }
 428
 429 static struct page_address_map page_address_maps[LAST_PKMAP];
 430
 431 void __init page_address_init(void)
 432 {
 433         int i;
 434
 435         INIT_LIST_HEAD(&page_address_pool);
 436         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
 437                 list_add(&page_address_maps[i].list, &page_address_pool);
 438         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 439                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 440                 spin_lock_init(&page_address_htable[i].lock);
 441         }
 442         spin_lock_init(&pool_lock);
 443 }
 444
 445 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */