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/module.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/blktrace_api.h>
  30 #include <asm/tlbflush.h>
  31
  32 /*
  33  * Virtual_count is not a pure "count".
  34  *  0 means that it is not mapped, and has not been mapped
  35  *    since a TLB flush - it is usable.
  36  *  1 means that there are no users, but it has been mapped
  37  *    since the last TLB flush - so we can't use it.
  38  *  n means that there are (n-1) current users of it.
  39  */
  40 #ifdef CONFIG_HIGHMEM
  41
  42 unsigned long totalhigh_pages __read_mostly;
  43
  44 unsigned int nr_free_highpages (void)
  45 {
  46         pg_data_t *pgdat;
  47         unsigned int pages = 0;
  48
  49         for_each_online_pgdat(pgdat)
  50                 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
  51
  52         return pages;
  53 }
  54
  55 static int pkmap_count[LAST_PKMAP];
  56 static unsigned int last_pkmap_nr;
  57 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
  58
  59 pte_t * pkmap_page_table;
  60
  61 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  62
  63 static void flush_all_zero_pkmaps(void)
  64 {
  65         int i;
  66
  67         flush_cache_kmaps();
  68
  69         for (i = 0; i < LAST_PKMAP; i++) {
  70                 struct page *page;
  71
  72                 /*
  73                  * zero means we don't have anything to do,
  74                  * >1 means that it is still in use. Only
  75                  * a count of 1 means that it is free but
  76                  * needs to be unmapped
  77                  */
  78                 if (pkmap_count[i] != 1)
  79                         continue;
  80                 pkmap_count[i] = 0;
  81
  82                 /* sanity check */
  83                 BUG_ON(pte_none(pkmap_page_table[i]));
  84
  85                 /*
  86                  * Don't need an atomic fetch-and-clear op here;
  87                  * no-one has the page mapped, and cannot get at
  88                  * its virtual address (and hence PTE) without first
  89                  * getting the kmap_lock (which is held here).
  90                  * So no dangers, even with speculative execution.
  91                  */
  92                 page = pte_page(pkmap_page_table[i]);
  93                 pte_clear(&init_mm, (unsigned long)page_address(page),
  94                           &pkmap_page_table[i]);
  95
  96                 set_page_address(page, NULL);
  97         }
  98         flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
  99 }
 100
 101 static inline unsigned long map_new_virtual(struct page *page)
 102 {
 103         unsigned long vaddr;
 104         int count;
 105
 106 start:
 107         count = LAST_PKMAP;
 108         /* Find an empty entry */
 109         for (;;) {
 110                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
 111                 if (!last_pkmap_nr) {
 112                         flush_all_zero_pkmaps();
 113                         count = LAST_PKMAP;
 114                 }
 115                 if (!pkmap_count[last_pkmap_nr])
 116                         break;  /* Found a usable entry */
 117                 if (--count)
 118                         continue;
 119
 120                 /*
 121                  * Sleep for somebody else to unmap their entries
 122                  */
 123                 {
 124                         DECLARE_WAITQUEUE(wait, current);
 125
 126                         __set_current_state(TASK_UNINTERRUPTIBLE);
 127                         add_wait_queue(&pkmap_map_wait, &wait);
 128                         spin_unlock(&kmap_lock);
 129                         schedule();
 130                         remove_wait_queue(&pkmap_map_wait, &wait);
 131                         spin_lock(&kmap_lock);
 132
 133                         /* Somebody else might have mapped it while we slept */
 134                         if (page_address(page))
 135                                 return (unsigned long)page_address(page);
 136
 137                         /* Re-start */
 138                         goto start;
 139                 }
 140         }
 141         vaddr = PKMAP_ADDR(last_pkmap_nr);
 142         set_pte_at(&init_mm, vaddr,
 143                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 144
 145         pkmap_count[last_pkmap_nr] = 1;
 146         set_page_address(page, (void *)vaddr);
 147
 148         return vaddr;
 149 }
 150
 151 #ifdef CONFIG_XEN
 152 void kmap_flush_unused(void)
 153 {
 154         spin_lock(&kmap_lock);
 155         flush_all_zero_pkmaps();
 156         spin_unlock(&kmap_lock);
 157 }
 158
 159 EXPORT_SYMBOL(kmap_flush_unused);
 160 #endif
 161
 162 void fastcall *kmap_high(struct page *page)
 163 {
 164         unsigned long vaddr;
 165
 166         /*
 167          * For highmem pages, we can't trust "virtual" until
 168          * after we have the lock.
 169          *
 170          * We cannot call this from interrupts, as it may block
 171          */
 172         spin_lock(&kmap_lock);
 173         vaddr = (unsigned long)page_address(page);
 174         if (!vaddr)
 175                 vaddr = map_new_virtual(page);
 176         pkmap_count[PKMAP_NR(vaddr)]++;
 177         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
 178         spin_unlock(&kmap_lock);
 179         return (void*) vaddr;
 180 }
 181
 182 EXPORT_SYMBOL(kmap_high);
 183
 184 void fastcall kunmap_high(struct page *page)
 185 {
 186         unsigned long vaddr;
 187         unsigned long nr;
 188         int need_wakeup;
 189
 190         spin_lock(&kmap_lock);
 191         vaddr = (unsigned long)page_address(page);
 192         BUG_ON(!vaddr);
 193         nr = PKMAP_NR(vaddr);
 194
 195         /*
 196          * A count must never go down to zero
 197          * without a TLB flush!
 198          */
 199         need_wakeup = 0;
 200         switch (--pkmap_count[nr]) {
 201         case 0:
 202                 BUG();
 203         case 1:
 204                 /*
 205                  * Avoid an unnecessary wake_up() function call.
 206                  * The common case is pkmap_count[] == 1, but
 207                  * no waiters.
 208                  * The tasks queued in the wait-queue are guarded
 209                  * by both the lock in the wait-queue-head and by
 210                  * the kmap_lock.  As the kmap_lock is held here,
 211                  * no need for the wait-queue-head's lock.  Simply
 212                  * test if the queue is empty.
 213                  */
 214                 need_wakeup = waitqueue_active(&pkmap_map_wait);
 215         }
 216         spin_unlock(&kmap_lock);
 217
 218         /* do wake-up, if needed, race-free outside of the spin lock */
 219         if (need_wakeup)
 220                 wake_up(&pkmap_map_wait);
 221 }
 222
 223 EXPORT_SYMBOL(kunmap_high);
 224 #endif
 225
 226 #if defined(HASHED_PAGE_VIRTUAL)
 227
 228 #define PA_HASH_ORDER   7
 229
 230 /*
 231  * Describes one page->virtual association
 232  */
 233 struct page_address_map {
 234         struct page *page;
 235         void *virtual;
 236         struct list_head list;
 237 };
 238
 239 /*
 240  * page_address_map freelist, allocated from page_address_maps.
 241  */
 242 static struct list_head page_address_pool;      /* freelist */
 243 static spinlock_t pool_lock;                    /* protects page_address_pool */
 244
 245 /*
 246  * Hash table bucket
 247  */
 248 static struct page_address_slot {
 249         struct list_head lh;                    /* List of page_address_maps */
 250         spinlock_t lock;                        /* Protect this bucket's list */
 251 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 252
 253 static struct page_address_slot *page_slot(struct page *page)
 254 {
 255         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 256 }
 257
 258 void *page_address(struct page *page)
 259 {
 260         unsigned long flags;
 261         void *ret;
 262         struct page_address_slot *pas;
 263
 264         if (!PageHighMem(page))
 265                 return lowmem_page_address(page);
 266
 267         pas = page_slot(page);
 268         ret = NULL;
 269         spin_lock_irqsave(&pas->lock, flags);
 270         if (!list_empty(&pas->lh)) {
 271                 struct page_address_map *pam;
 272
 273                 list_for_each_entry(pam, &pas->lh, list) {
 274                         if (pam->page == page) {
 275                                 ret = pam->virtual;
 276                                 goto done;
 277                         }
 278                 }
 279         }
 280 done:
 281         spin_unlock_irqrestore(&pas->lock, flags);
 282         return ret;
 283 }
 284
 285 EXPORT_SYMBOL(page_address);
 286
 287 void set_page_address(struct page *page, void *virtual)
 288 {
 289         unsigned long flags;
 290         struct page_address_slot *pas;
 291         struct page_address_map *pam;
 292
 293         BUG_ON(!PageHighMem(page));
 294
 295         pas = page_slot(page);
 296         if (virtual) {          /* Add */
 297                 BUG_ON(list_empty(&page_address_pool));
 298
 299                 spin_lock_irqsave(&pool_lock, flags);
 300                 pam = list_entry(page_address_pool.next,
 301                                 struct page_address_map, list);
 302                 list_del(&pam->list);
 303                 spin_unlock_irqrestore(&pool_lock, flags);
 304
 305                 pam->page = page;
 306                 pam->virtual = virtual;
 307
 308                 spin_lock_irqsave(&pas->lock, flags);
 309                 list_add_tail(&pam->list, &pas->lh);
 310                 spin_unlock_irqrestore(&pas->lock, flags);
 311         } else {                /* Remove */
 312                 spin_lock_irqsave(&pas->lock, flags);
 313                 list_for_each_entry(pam, &pas->lh, list) {
 314                         if (pam->page == page) {
 315                                 list_del(&pam->list);
 316                                 spin_unlock_irqrestore(&pas->lock, flags);
 317                                 spin_lock_irqsave(&pool_lock, flags);
 318                                 list_add_tail(&pam->list, &page_address_pool);
 319                                 spin_unlock_irqrestore(&pool_lock, flags);
 320                                 goto done;
 321                         }
 322                 }
 323                 spin_unlock_irqrestore(&pas->lock, flags);
 324         }
 325 done:
 326         return;
 327 }
 328
 329 static struct page_address_map page_address_maps[LAST_PKMAP];
 330
 331 void __init page_address_init(void)
 332 {
 333         int i;
 334
 335         INIT_LIST_HEAD(&page_address_pool);
 336         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
 337                 list_add(&page_address_maps[i].list, &page_address_pool);
 338         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 339                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 340                 spin_lock_init(&page_address_htable[i].lock);
 341         }
 342         spin_lock_init(&pool_lock);
 343 }
 344
 345 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */