kernel/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 #ifndef CONFIG_PREEMPT_RT_FULL
  33 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
  34 DEFINE_PER_CPU(int, __kmap_atomic_idx);
  35 #endif
  36 #endif
  37
  38 /*
  39  * Virtual_count is not a pure "count".
  40  *  0 means that it is not mapped, and has not been mapped
  41  *    since a TLB flush - it is usable.
  42  *  1 means that there are no users, but it has been mapped
  43  *    since the last TLB flush - so we can't use it.
  44  *  n means that there are (n-1) current users of it.
  45  */
  46 #ifdef CONFIG_HIGHMEM
  47
  48 /*
  49  * Architecture with aliasing data cache may define the following family of
  50  * helper functions in its asm/highmem.h to control cache color of virtual
  51  * addresses where physical memory pages are mapped by kmap.
  52  */
  53 #ifndef get_pkmap_color
  54
  55 /*
  56  * Determine color of virtual address where the page should be mapped.
  57  */
  58 static inline unsigned int get_pkmap_color(struct page *page)
  59 {
  60         return 0;
  61 }
  62 #define get_pkmap_color get_pkmap_color
  63
  64 /*
  65  * Get next index for mapping inside PKMAP region for page with given color.
  66  */
  67 static inline unsigned int get_next_pkmap_nr(unsigned int color)
  68 {
  69         static unsigned int last_pkmap_nr;
  70
  71         last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
  72         return last_pkmap_nr;
  73 }
  74
  75 /*
  76  * Determine if page index inside PKMAP region (pkmap_nr) of given color
  77  * has wrapped around PKMAP region end. When this happens an attempt to
  78  * flush all unused PKMAP slots is made.
  79  */
  80 static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
  81 {
  82         return pkmap_nr == 0;
  83 }
  84
  85 /*
  86  * Get the number of PKMAP entries of the given color. If no free slot is
  87  * found after checking that many entries, kmap will sleep waiting for
  88  * someone to call kunmap and free PKMAP slot.
  89  */
  90 static inline int get_pkmap_entries_count(unsigned int color)
  91 {
  92         return LAST_PKMAP;
  93 }
  94
  95 /*
  96  * Get head of a wait queue for PKMAP entries of the given color.
  97  * Wait queues for different mapping colors should be independent to avoid
  98  * unnecessary wakeups caused by freeing of slots of other colors.
  99  */
 100 static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
 101 {
 102         static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
 103
 104         return &pkmap_map_wait;
 105 }
 106 #endif
 107
 108 unsigned long totalhigh_pages __read_mostly;
 109 EXPORT_SYMBOL(totalhigh_pages);
 110
 111 #ifndef CONFIG_PREEMPT_RT_FULL
 112 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
 113 #endif
 114
 115 unsigned int nr_free_highpages (void)
 116 {
 117         pg_data_t *pgdat;
 118         unsigned int pages = 0;
 119
 120         for_each_online_pgdat(pgdat) {
 121                 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
 122                         NR_FREE_PAGES);
 123                 if (zone_movable_is_highmem())
 124                         pages += zone_page_state(
 125                                         &pgdat->node_zones[ZONE_MOVABLE],
 126                                         NR_FREE_PAGES);
 127         }
 128
 129         return pages;
 130 }
 131
 132 static int pkmap_count[LAST_PKMAP];
 133 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
 134
 135 pte_t * pkmap_page_table;
 136
 137 /*
 138  * Most architectures have no use for kmap_high_get(), so let's abstract
 139  * the disabling of IRQ out of the locking in that case to save on a
 140  * potential useless overhead.
 141  */
 142 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 143 #define lock_kmap()             spin_lock_irq(&kmap_lock)
 144 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
 145 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
 146 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
 147 #else
 148 #define lock_kmap()             spin_lock(&kmap_lock)
 149 #define unlock_kmap()           spin_unlock(&kmap_lock)
 150 #define lock_kmap_any(flags)    \
 151                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
 152 #define unlock_kmap_any(flags)  \
 153                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
 154 #endif
 155
 156 struct page *kmap_to_page(void *vaddr)
 157 {
 158         unsigned long addr = (unsigned long)vaddr;
 159
 160         if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
 161                 int i = PKMAP_NR(addr);
 162                 return pte_page(pkmap_page_table[i]);
 163         }
 164
 165         return virt_to_page(addr);
 166 }
 167 EXPORT_SYMBOL(kmap_to_page);
 168
 169 static void flush_all_zero_pkmaps(void)
 170 {
 171         int i;
 172         int need_flush = 0;
 173
 174         flush_cache_kmaps();
 175
 176         for (i = 0; i < LAST_PKMAP; i++) {
 177                 struct page *page;
 178
 179                 /*
 180                  * zero means we don't have anything to do,
 181                  * >1 means that it is still in use. Only
 182                  * a count of 1 means that it is free but
 183                  * needs to be unmapped
 184                  */
 185                 if (pkmap_count[i] != 1)
 186                         continue;
 187                 pkmap_count[i] = 0;
 188
 189                 /* sanity check */
 190                 BUG_ON(pte_none(pkmap_page_table[i]));
 191
 192                 /*
 193                  * Don't need an atomic fetch-and-clear op here;
 194                  * no-one has the page mapped, and cannot get at
 195                  * its virtual address (and hence PTE) without first
 196                  * getting the kmap_lock (which is held here).
 197                  * So no dangers, even with speculative execution.
 198                  */
 199                 page = pte_page(pkmap_page_table[i]);
 200                 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
 201
 202                 set_page_address(page, NULL);
 203                 need_flush = 1;
 204         }
 205         if (need_flush)
 206                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
 207 }
 208
 209 /**
 210  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
 211  */
 212 void kmap_flush_unused(void)
 213 {
 214         lock_kmap();
 215         flush_all_zero_pkmaps();
 216         unlock_kmap();
 217 }
 218
 219 static inline unsigned long map_new_virtual(struct page *page)
 220 {
 221         unsigned long vaddr;
 222         int count;
 223         unsigned int last_pkmap_nr;
 224         unsigned int color = get_pkmap_color(page);
 225
 226 start:
 227         count = get_pkmap_entries_count(color);
 228         /* Find an empty entry */
 229         for (;;) {
 230                 last_pkmap_nr = get_next_pkmap_nr(color);
 231                 if (no_more_pkmaps(last_pkmap_nr, color)) {
 232                         flush_all_zero_pkmaps();
 233                         count = get_pkmap_entries_count(color);
 234                 }
 235                 if (!pkmap_count[last_pkmap_nr])
 236                         break;  /* Found a usable entry */
 237                 if (--count)
 238                         continue;
 239
 240                 /*
 241                  * Sleep for somebody else to unmap their entries
 242                  */
 243                 {
 244                         DECLARE_WAITQUEUE(wait, current);
 245                         wait_queue_head_t *pkmap_map_wait =
 246                                 get_pkmap_wait_queue_head(color);
 247
 248                         __set_current_state(TASK_UNINTERRUPTIBLE);
 249                         add_wait_queue(pkmap_map_wait, &wait);
 250                         unlock_kmap();
 251                         schedule();
 252                         remove_wait_queue(pkmap_map_wait, &wait);
 253                         lock_kmap();
 254
 255                         /* Somebody else might have mapped it while we slept */
 256                         if (page_address(page))
 257                                 return (unsigned long)page_address(page);
 258
 259                         /* Re-start */
 260                         goto start;
 261                 }
 262         }
 263         vaddr = PKMAP_ADDR(last_pkmap_nr);
 264         set_pte_at(&init_mm, vaddr,
 265                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 266
 267         pkmap_count[last_pkmap_nr] = 1;
 268         set_page_address(page, (void *)vaddr);
 269
 270         return vaddr;
 271 }
 272
 273 /**
 274  * kmap_high - map a highmem page into memory
 275  * @page: &struct page to map
 276  *
 277  * Returns the page's virtual memory address.
 278  *
 279  * We cannot call this from interrupts, as it may block.
 280  */
 281 void *kmap_high(struct page *page)
 282 {
 283         unsigned long vaddr;
 284
 285         /*
 286          * For highmem pages, we can't trust "virtual" until
 287          * after we have the lock.
 288          */
 289         lock_kmap();
 290         vaddr = (unsigned long)page_address(page);
 291         if (!vaddr)
 292                 vaddr = map_new_virtual(page);
 293         pkmap_count[PKMAP_NR(vaddr)]++;
 294         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
 295         unlock_kmap();
 296         return (void*) vaddr;
 297 }
 298
 299 EXPORT_SYMBOL(kmap_high);
 300
 301 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 302 /**
 303  * kmap_high_get - pin a highmem page into memory
 304  * @page: &struct page to pin
 305  *
 306  * Returns the page's current virtual memory address, or NULL if no mapping
 307  * exists.  If and only if a non null address is returned then a
 308  * matching call to kunmap_high() is necessary.
 309  *
 310  * This can be called from any context.
 311  */
 312 void *kmap_high_get(struct page *page)
 313 {
 314         unsigned long vaddr, flags;
 315
 316         lock_kmap_any(flags);
 317         vaddr = (unsigned long)page_address(page);
 318         if (vaddr) {
 319                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
 320                 pkmap_count[PKMAP_NR(vaddr)]++;
 321         }
 322         unlock_kmap_any(flags);
 323         return (void*) vaddr;
 324 }
 325 #endif
 326
 327 /**
 328  * kunmap_high - unmap a highmem page into memory
 329  * @page: &struct page to unmap
 330  *
 331  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
 332  * only from user context.
 333  */
 334 void kunmap_high(struct page *page)
 335 {
 336         unsigned long vaddr;
 337         unsigned long nr;
 338         unsigned long flags;
 339         int need_wakeup;
 340         unsigned int color = get_pkmap_color(page);
 341         wait_queue_head_t *pkmap_map_wait;
 342
 343         lock_kmap_any(flags);
 344         vaddr = (unsigned long)page_address(page);
 345         BUG_ON(!vaddr);
 346         nr = PKMAP_NR(vaddr);
 347
 348         /*
 349          * A count must never go down to zero
 350          * without a TLB flush!
 351          */
 352         need_wakeup = 0;
 353         switch (--pkmap_count[nr]) {
 354         case 0:
 355                 BUG();
 356         case 1:
 357                 /*
 358                  * Avoid an unnecessary wake_up() function call.
 359                  * The common case is pkmap_count[] == 1, but
 360                  * no waiters.
 361                  * The tasks queued in the wait-queue are guarded
 362                  * by both the lock in the wait-queue-head and by
 363                  * the kmap_lock.  As the kmap_lock is held here,
 364                  * no need for the wait-queue-head's lock.  Simply
 365                  * test if the queue is empty.
 366                  */
 367                 pkmap_map_wait = get_pkmap_wait_queue_head(color);
 368                 need_wakeup = waitqueue_active(pkmap_map_wait);
 369         }
 370         unlock_kmap_any(flags);
 371
 372         /* do wake-up, if needed, race-free outside of the spin lock */
 373         if (need_wakeup)
 374                 wake_up(pkmap_map_wait);
 375 }
 376
 377 EXPORT_SYMBOL(kunmap_high);
 378 #endif
 379
 380 #if defined(HASHED_PAGE_VIRTUAL)
 381
 382 #define PA_HASH_ORDER   7
 383
 384 /*
 385  * Describes one page->virtual association
 386  */
 387 struct page_address_map {
 388         struct page *page;
 389         void *virtual;
 390         struct list_head list;
 391 };
 392
 393 static struct page_address_map page_address_maps[LAST_PKMAP];
 394
 395 /*
 396  * Hash table bucket
 397  */
 398 static struct page_address_slot {
 399         struct list_head lh;                    /* List of page_address_maps */
 400         spinlock_t lock;                        /* Protect this bucket's list */
 401 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 402
 403 static struct page_address_slot *page_slot(const struct page *page)
 404 {
 405         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 406 }
 407
 408 /**
 409  * page_address - get the mapped virtual address of a page
 410  * @page: &struct page to get the virtual address of
 411  *
 412  * Returns the page's virtual address.
 413  */
 414 void *page_address(const struct page *page)
 415 {
 416         unsigned long flags;
 417         void *ret;
 418         struct page_address_slot *pas;
 419
 420         if (!PageHighMem(page))
 421                 return lowmem_page_address(page);
 422
 423         pas = page_slot(page);
 424         ret = NULL;
 425         spin_lock_irqsave(&pas->lock, flags);
 426         if (!list_empty(&pas->lh)) {
 427                 struct page_address_map *pam;
 428
 429                 list_for_each_entry(pam, &pas->lh, list) {
 430                         if (pam->page == page) {
 431                                 ret = pam->virtual;
 432                                 goto done;
 433                         }
 434                 }
 435         }
 436 done:
 437         spin_unlock_irqrestore(&pas->lock, flags);
 438         return ret;
 439 }
 440
 441 EXPORT_SYMBOL(page_address);
 442
 443 /**
 444  * set_page_address - set a page's virtual address
 445  * @page: &struct page to set
 446  * @virtual: virtual address to use
 447  */
 448 void set_page_address(struct page *page, void *virtual)
 449 {
 450         unsigned long flags;
 451         struct page_address_slot *pas;
 452         struct page_address_map *pam;
 453
 454         BUG_ON(!PageHighMem(page));
 455
 456         pas = page_slot(page);
 457         if (virtual) {          /* Add */
 458                 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
 459                 pam->page = page;
 460                 pam->virtual = virtual;
 461
 462                 spin_lock_irqsave(&pas->lock, flags);
 463                 list_add_tail(&pam->list, &pas->lh);
 464                 spin_unlock_irqrestore(&pas->lock, flags);
 465         } else {                /* Remove */
 466                 spin_lock_irqsave(&pas->lock, flags);
 467                 list_for_each_entry(pam, &pas->lh, list) {
 468                         if (pam->page == page) {
 469                                 list_del(&pam->list);
 470                                 spin_unlock_irqrestore(&pas->lock, flags);
 471                                 goto done;
 472                         }
 473                 }
 474                 spin_unlock_irqrestore(&pas->lock, flags);
 475         }
 476 done:
 477         return;
 478 }
 479
 480 void __init page_address_init(void)
 481 {
 482         int i;
 483
 484         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 485                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 486                 spin_lock_init(&page_address_htable[i].lock);
 487         }
 488 }
 489
 490 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */