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