mm/highmem.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * High memory handling common code and variables.
   4  *
   5  * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
   6  *          Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
   7  *
   8  *
   9  * Redesigned the x86 32-bit VM architecture to deal with
  10  * 64-bit physical space. With current x86 CPUs this
  11  * means up to 64 Gigabytes physical RAM.
  12  *
  13  * Rewrote high memory support to move the page cache into
  14  * high memory. Implemented permanent (schedulable) kmaps
  15  * based on Linus' idea.
  16  *
  17  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
  18  */
  19
  20 #include <linux/mm.h>
  21 #include <linux/export.h>
  22 #include <linux/swap.h>
  23 #include <linux/bio.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/mempool.h>
  26 #include <linux/blkdev.h>
  27 #include <linux/init.h>
  28 #include <linux/hash.h>
  29 #include <linux/highmem.h>
  30 #include <linux/kgdb.h>
  31 #include <asm/tlbflush.h>
  32 #include <linux/vmalloc.h>
  33
  34 /*
  35  * Virtual_count is not a pure "count".
  36  *  0 means that it is not mapped, and has not been mapped
  37  *    since a TLB flush - it is usable.
  38  *  1 means that there are no users, but it has been mapped
  39  *    since the last TLB flush - so we can't use it.
  40  *  n means that there are (n-1) current users of it.
  41  */
  42 #ifdef CONFIG_HIGHMEM
  43
  44 /*
  45  * Architecture with aliasing data cache may define the following family of
  46  * helper functions in its asm/highmem.h to control cache color of virtual
  47  * addresses where physical memory pages are mapped by kmap.
  48  */
  49 #ifndef get_pkmap_color
  50
  51 /*
  52  * Determine color of virtual address where the page should be mapped.
  53  */
  54 static inline unsigned int get_pkmap_color(struct page *page)
  55 {
  56         return 0;
  57 }
  58 #define get_pkmap_color get_pkmap_color
  59
  60 /*
  61  * Get next index for mapping inside PKMAP region for page with given color.
  62  */
  63 static inline unsigned int get_next_pkmap_nr(unsigned int color)
  64 {
  65         static unsigned int last_pkmap_nr;
  66
  67         last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
  68         return last_pkmap_nr;
  69 }
  70
  71 /*
  72  * Determine if page index inside PKMAP region (pkmap_nr) of given color
  73  * has wrapped around PKMAP region end. When this happens an attempt to
  74  * flush all unused PKMAP slots is made.
  75  */
  76 static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
  77 {
  78         return pkmap_nr == 0;
  79 }
  80
  81 /*
  82  * Get the number of PKMAP entries of the given color. If no free slot is
  83  * found after checking that many entries, kmap will sleep waiting for
  84  * someone to call kunmap and free PKMAP slot.
  85  */
  86 static inline int get_pkmap_entries_count(unsigned int color)
  87 {
  88         return LAST_PKMAP;
  89 }
  90
  91 /*
  92  * Get head of a wait queue for PKMAP entries of the given color.
  93  * Wait queues for different mapping colors should be independent to avoid
  94  * unnecessary wakeups caused by freeing of slots of other colors.
  95  */
  96 static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
  97 {
  98         static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  99
 100         return &pkmap_map_wait;
 101 }
 102 #endif
 103
 104 atomic_long_t _totalhigh_pages __read_mostly;
 105 EXPORT_SYMBOL(_totalhigh_pages);
 106
 107 unsigned int nr_free_highpages (void)
 108 {
 109         struct zone *zone;
 110         unsigned int pages = 0;
 111
 112         for_each_populated_zone(zone) {
 113                 if (is_highmem(zone))
 114                         pages += zone_page_state(zone, NR_FREE_PAGES);
 115         }
 116
 117         return pages;
 118 }
 119
 120 static int pkmap_count[LAST_PKMAP];
 121 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
 122
 123 pte_t * pkmap_page_table;
 124
 125 /*
 126  * Most architectures have no use for kmap_high_get(), so let's abstract
 127  * the disabling of IRQ out of the locking in that case to save on a
 128  * potential useless overhead.
 129  */
 130 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 131 #define lock_kmap()             spin_lock_irq(&kmap_lock)
 132 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
 133 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
 134 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
 135 #else
 136 #define lock_kmap()             spin_lock(&kmap_lock)
 137 #define unlock_kmap()           spin_unlock(&kmap_lock)
 138 #define lock_kmap_any(flags)    \
 139                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
 140 #define unlock_kmap_any(flags)  \
 141                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
 142 #endif
 143
 144 struct page *kmap_to_page(void *vaddr)
 145 {
 146         unsigned long addr = (unsigned long)vaddr;
 147
 148         if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
 149                 int i = PKMAP_NR(addr);
 150                 return pte_page(pkmap_page_table[i]);
 151         }
 152
 153         return virt_to_page(addr);
 154 }
 155 EXPORT_SYMBOL(kmap_to_page);
 156
 157 static void flush_all_zero_pkmaps(void)
 158 {
 159         int i;
 160         int need_flush = 0;
 161
 162         flush_cache_kmaps();
 163
 164         for (i = 0; i < LAST_PKMAP; i++) {
 165                 struct page *page;
 166
 167                 /*
 168                  * zero means we don't have anything to do,
 169                  * >1 means that it is still in use. Only
 170                  * a count of 1 means that it is free but
 171                  * needs to be unmapped
 172                  */
 173                 if (pkmap_count[i] != 1)
 174                         continue;
 175                 pkmap_count[i] = 0;
 176
 177                 /* sanity check */
 178                 BUG_ON(pte_none(pkmap_page_table[i]));
 179
 180                 /*
 181                  * Don't need an atomic fetch-and-clear op here;
 182                  * no-one has the page mapped, and cannot get at
 183                  * its virtual address (and hence PTE) without first
 184                  * getting the kmap_lock (which is held here).
 185                  * So no dangers, even with speculative execution.
 186                  */
 187                 page = pte_page(pkmap_page_table[i]);
 188                 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
 189
 190                 set_page_address(page, NULL);
 191                 need_flush = 1;
 192         }
 193         if (need_flush)
 194                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
 195 }
 196
 197 /**
 198  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
 199  */
 200 void kmap_flush_unused(void)
 201 {
 202         lock_kmap();
 203         flush_all_zero_pkmaps();
 204         unlock_kmap();
 205 }
 206
 207 static inline unsigned long map_new_virtual(struct page *page)
 208 {
 209         unsigned long vaddr;
 210         int count;
 211         unsigned int last_pkmap_nr;
 212         unsigned int color = get_pkmap_color(page);
 213
 214 start:
 215         count = get_pkmap_entries_count(color);
 216         /* Find an empty entry */
 217         for (;;) {
 218                 last_pkmap_nr = get_next_pkmap_nr(color);
 219                 if (no_more_pkmaps(last_pkmap_nr, color)) {
 220                         flush_all_zero_pkmaps();
 221                         count = get_pkmap_entries_count(color);
 222                 }
 223                 if (!pkmap_count[last_pkmap_nr])
 224                         break;  /* Found a usable entry */
 225                 if (--count)
 226                         continue;
 227
 228                 /*
 229                  * Sleep for somebody else to unmap their entries
 230                  */
 231                 {
 232                         DECLARE_WAITQUEUE(wait, current);
 233                         wait_queue_head_t *pkmap_map_wait =
 234                                 get_pkmap_wait_queue_head(color);
 235
 236                         __set_current_state(TASK_UNINTERRUPTIBLE);
 237                         add_wait_queue(pkmap_map_wait, &wait);
 238                         unlock_kmap();
 239                         schedule();
 240                         remove_wait_queue(pkmap_map_wait, &wait);
 241                         lock_kmap();
 242
 243                         /* Somebody else might have mapped it while we slept */
 244                         if (page_address(page))
 245                                 return (unsigned long)page_address(page);
 246
 247                         /* Re-start */
 248                         goto start;
 249                 }
 250         }
 251         vaddr = PKMAP_ADDR(last_pkmap_nr);
 252         set_pte_at(&init_mm, vaddr,
 253                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 254
 255         pkmap_count[last_pkmap_nr] = 1;
 256         set_page_address(page, (void *)vaddr);
 257
 258         return vaddr;
 259 }
 260
 261 /**
 262  * kmap_high - map a highmem page into memory
 263  * @page: &struct page to map
 264  *
 265  * Returns the page's virtual memory address.
 266  *
 267  * We cannot call this from interrupts, as it may block.
 268  */
 269 void *kmap_high(struct page *page)
 270 {
 271         unsigned long vaddr;
 272
 273         /*
 274          * For highmem pages, we can't trust "virtual" until
 275          * after we have the lock.
 276          */
 277         lock_kmap();
 278         vaddr = (unsigned long)page_address(page);
 279         if (!vaddr)
 280                 vaddr = map_new_virtual(page);
 281         pkmap_count[PKMAP_NR(vaddr)]++;
 282         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
 283         unlock_kmap();
 284         return (void*) vaddr;
 285 }
 286
 287 EXPORT_SYMBOL(kmap_high);
 288
 289 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 290 /**
 291  * kmap_high_get - pin a highmem page into memory
 292  * @page: &struct page to pin
 293  *
 294  * Returns the page's current virtual memory address, or NULL if no mapping
 295  * exists.  If and only if a non null address is returned then a
 296  * matching call to kunmap_high() is necessary.
 297  *
 298  * This can be called from any context.
 299  */
 300 void *kmap_high_get(struct page *page)
 301 {
 302         unsigned long vaddr, flags;
 303
 304         lock_kmap_any(flags);
 305         vaddr = (unsigned long)page_address(page);
 306         if (vaddr) {
 307                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
 308                 pkmap_count[PKMAP_NR(vaddr)]++;
 309         }
 310         unlock_kmap_any(flags);
 311         return (void*) vaddr;
 312 }
 313 #endif
 314
 315 /**
 316  * kunmap_high - unmap a highmem page into memory
 317  * @page: &struct page to unmap
 318  *
 319  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
 320  * only from user context.
 321  */
 322 void kunmap_high(struct page *page)
 323 {
 324         unsigned long vaddr;
 325         unsigned long nr;
 326         unsigned long flags;
 327         int need_wakeup;
 328         unsigned int color = get_pkmap_color(page);
 329         wait_queue_head_t *pkmap_map_wait;
 330
 331         lock_kmap_any(flags);
 332         vaddr = (unsigned long)page_address(page);
 333         BUG_ON(!vaddr);
 334         nr = PKMAP_NR(vaddr);
 335
 336         /*
 337          * A count must never go down to zero
 338          * without a TLB flush!
 339          */
 340         need_wakeup = 0;
 341         switch (--pkmap_count[nr]) {
 342         case 0:
 343                 BUG();
 344         case 1:
 345                 /*
 346                  * Avoid an unnecessary wake_up() function call.
 347                  * The common case is pkmap_count[] == 1, but
 348                  * no waiters.
 349                  * The tasks queued in the wait-queue are guarded
 350                  * by both the lock in the wait-queue-head and by
 351                  * the kmap_lock.  As the kmap_lock is held here,
 352                  * no need for the wait-queue-head's lock.  Simply
 353                  * test if the queue is empty.
 354                  */
 355                 pkmap_map_wait = get_pkmap_wait_queue_head(color);
 356                 need_wakeup = waitqueue_active(pkmap_map_wait);
 357         }
 358         unlock_kmap_any(flags);
 359
 360         /* do wake-up, if needed, race-free outside of the spin lock */
 361         if (need_wakeup)
 362                 wake_up(pkmap_map_wait);
 363 }
 364 EXPORT_SYMBOL(kunmap_high);
 365 #endif /* CONFIG_HIGHMEM */
 366
 367 #ifdef CONFIG_KMAP_LOCAL
 368
 369 #include <asm/kmap_size.h>
 370
 371 static DEFINE_PER_CPU(int, __kmap_local_idx);
 372
 373 /*
 374  * With DEBUG_HIGHMEM the stack depth is doubled and every second
 375  * slot is unused which acts as a guard page
 376  */
 377 #ifdef CONFIG_DEBUG_HIGHMEM
 378 # define KM_INCR        2
 379 #else
 380 # define KM_INCR        1
 381 #endif
 382
 383 static inline int kmap_local_idx_push(void)
 384 {
 385         int idx = __this_cpu_add_return(__kmap_local_idx, KM_INCR) - 1;
 386
 387         WARN_ON_ONCE(in_irq() && !irqs_disabled());
 388         BUG_ON(idx >= KM_MAX_IDX);
 389         return idx;
 390 }
 391
 392 static inline int kmap_local_idx(void)
 393 {
 394         return __this_cpu_read(__kmap_local_idx) - 1;
 395 }
 396
 397 static inline void kmap_local_idx_pop(void)
 398 {
 399         int idx = __this_cpu_sub_return(__kmap_local_idx, KM_INCR);
 400
 401         BUG_ON(idx < 0);
 402 }
 403
 404 #ifndef arch_kmap_local_post_map
 405 # define arch_kmap_local_post_map(vaddr, pteval)        do { } while (0)
 406 #endif
 407
 408 #ifndef arch_kmap_local_pre_unmap
 409 # define arch_kmap_local_pre_unmap(vaddr)               do { } while (0)
 410 #endif
 411
 412 #ifndef arch_kmap_local_post_unmap
 413 # define arch_kmap_local_post_unmap(vaddr)              do { } while (0)
 414 #endif
 415
 416 #ifndef arch_kmap_local_map_idx
 417 #define arch_kmap_local_map_idx(idx, pfn)       kmap_local_calc_idx(idx)
 418 #endif
 419
 420 #ifndef arch_kmap_local_unmap_idx
 421 #define arch_kmap_local_unmap_idx(idx, vaddr)   kmap_local_calc_idx(idx)
 422 #endif
 423
 424 #ifndef arch_kmap_local_high_get
 425 static inline void *arch_kmap_local_high_get(struct page *page)
 426 {
 427         return NULL;
 428 }
 429 #endif
 430
 431 /* Unmap a local mapping which was obtained by kmap_high_get() */
 432 static inline void kmap_high_unmap_local(unsigned long vaddr)
 433 {
 434 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 435         if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP))
 436                 kunmap_high(pte_page(pkmap_page_table[PKMAP_NR(vaddr)]));
 437 #endif
 438 }
 439
 440 static inline int kmap_local_calc_idx(int idx)
 441 {
 442         return idx + KM_MAX_IDX * smp_processor_id();
 443 }
 444
 445 static pte_t *__kmap_pte;
 446
 447 static pte_t *kmap_get_pte(void)
 448 {
 449         if (!__kmap_pte)
 450                 __kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
 451         return __kmap_pte;
 452 }
 453
 454 void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
 455 {
 456         pte_t pteval, *kmap_pte = kmap_get_pte();
 457         unsigned long vaddr;
 458         int idx;
 459
 460         preempt_disable();
 461         idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
 462         vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
 463         BUG_ON(!pte_none(*(kmap_pte - idx)));
 464         pteval = pfn_pte(pfn, prot);
 465         set_pte_at(&init_mm, vaddr, kmap_pte - idx, pteval);
 466         arch_kmap_local_post_map(vaddr, pteval);
 467         preempt_enable();
 468
 469         return (void *)vaddr;
 470 }
 471 EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
 472
 473 void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
 474 {
 475         void *kmap;
 476
 477         if (!PageHighMem(page))
 478                 return page_address(page);
 479
 480         /* Try kmap_high_get() if architecture has it enabled */
 481         kmap = arch_kmap_local_high_get(page);
 482         if (kmap)
 483                 return kmap;
 484
 485         return __kmap_local_pfn_prot(page_to_pfn(page), prot);
 486 }
 487 EXPORT_SYMBOL(__kmap_local_page_prot);
 488
 489 void kunmap_local_indexed(void *vaddr)
 490 {
 491         unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
 492         pte_t *kmap_pte = kmap_get_pte();
 493         int idx;
 494
 495         if (addr < __fix_to_virt(FIX_KMAP_END) ||
 496             addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
 497                 WARN_ON_ONCE(addr < PAGE_OFFSET);
 498
 499                 /* Handle mappings which were obtained by kmap_high_get() */
 500                 kmap_high_unmap_local(addr);
 501                 return;
 502         }
 503
 504         preempt_disable();
 505         idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
 506         WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
 507
 508         arch_kmap_local_pre_unmap(addr);
 509         pte_clear(&init_mm, addr, kmap_pte - idx);
 510         arch_kmap_local_post_unmap(addr);
 511         kmap_local_idx_pop();
 512         preempt_enable();
 513 }
 514 EXPORT_SYMBOL(kunmap_local_indexed);
 515 #endif
 516
 517 #if defined(HASHED_PAGE_VIRTUAL)
 518
 519 #define PA_HASH_ORDER   7
 520
 521 /*
 522  * Describes one page->virtual association
 523  */
 524 struct page_address_map {
 525         struct page *page;
 526         void *virtual;
 527         struct list_head list;
 528 };
 529
 530 static struct page_address_map page_address_maps[LAST_PKMAP];
 531
 532 /*
 533  * Hash table bucket
 534  */
 535 static struct page_address_slot {
 536         struct list_head lh;                    /* List of page_address_maps */
 537         spinlock_t lock;                        /* Protect this bucket's list */
 538 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 539
 540 static struct page_address_slot *page_slot(const struct page *page)
 541 {
 542         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 543 }
 544
 545 /**
 546  * page_address - get the mapped virtual address of a page
 547  * @page: &struct page to get the virtual address of
 548  *
 549  * Returns the page's virtual address.
 550  */
 551 void *page_address(const struct page *page)
 552 {
 553         unsigned long flags;
 554         void *ret;
 555         struct page_address_slot *pas;
 556
 557         if (!PageHighMem(page))
 558                 return lowmem_page_address(page);
 559
 560         pas = page_slot(page);
 561         ret = NULL;
 562         spin_lock_irqsave(&pas->lock, flags);
 563         if (!list_empty(&pas->lh)) {
 564                 struct page_address_map *pam;
 565
 566                 list_for_each_entry(pam, &pas->lh, list) {
 567                         if (pam->page == page) {
 568                                 ret = pam->virtual;
 569                                 goto done;
 570                         }
 571                 }
 572         }
 573 done:
 574         spin_unlock_irqrestore(&pas->lock, flags);
 575         return ret;
 576 }
 577
 578 EXPORT_SYMBOL(page_address);
 579
 580 /**
 581  * set_page_address - set a page's virtual address
 582  * @page: &struct page to set
 583  * @virtual: virtual address to use
 584  */
 585 void set_page_address(struct page *page, void *virtual)
 586 {
 587         unsigned long flags;
 588         struct page_address_slot *pas;
 589         struct page_address_map *pam;
 590
 591         BUG_ON(!PageHighMem(page));
 592
 593         pas = page_slot(page);
 594         if (virtual) {          /* Add */
 595                 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
 596                 pam->page = page;
 597                 pam->virtual = virtual;
 598
 599                 spin_lock_irqsave(&pas->lock, flags);
 600                 list_add_tail(&pam->list, &pas->lh);
 601                 spin_unlock_irqrestore(&pas->lock, flags);
 602         } else {                /* Remove */
 603                 spin_lock_irqsave(&pas->lock, flags);
 604                 list_for_each_entry(pam, &pas->lh, list) {
 605                         if (pam->page == page) {
 606                                 list_del(&pam->list);
 607                                 spin_unlock_irqrestore(&pas->lock, flags);
 608                                 goto done;
 609                         }
 610                 }
 611                 spin_unlock_irqrestore(&pas->lock, flags);
 612         }
 613 done:
 614         return;
 615 }
 616
 617 void __init page_address_init(void)
 618 {
 619         int i;
 620
 621         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 622                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 623                 spin_lock_init(&page_address_htable[i].lock);
 624         }
 625 }
 626
 627 #endif  /* defined(HASHED_PAGE_VIRTUAL) */