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 void __kmap_flush_unused(void)
 198 {
 199         lock_kmap();
 200         flush_all_zero_pkmaps();
 201         unlock_kmap();
 202 }
 203
 204 static inline unsigned long map_new_virtual(struct page *page)
 205 {
 206         unsigned long vaddr;
 207         int count;
 208         unsigned int last_pkmap_nr;
 209         unsigned int color = get_pkmap_color(page);
 210
 211 start:
 212         count = get_pkmap_entries_count(color);
 213         /* Find an empty entry */
 214         for (;;) {
 215                 last_pkmap_nr = get_next_pkmap_nr(color);
 216                 if (no_more_pkmaps(last_pkmap_nr, color)) {
 217                         flush_all_zero_pkmaps();
 218                         count = get_pkmap_entries_count(color);
 219                 }
 220                 if (!pkmap_count[last_pkmap_nr])
 221                         break;  /* Found a usable entry */
 222                 if (--count)
 223                         continue;
 224
 225                 /*
 226                  * Sleep for somebody else to unmap their entries
 227                  */
 228                 {
 229                         DECLARE_WAITQUEUE(wait, current);
 230                         wait_queue_head_t *pkmap_map_wait =
 231                                 get_pkmap_wait_queue_head(color);
 232
 233                         __set_current_state(TASK_UNINTERRUPTIBLE);
 234                         add_wait_queue(pkmap_map_wait, &wait);
 235                         unlock_kmap();
 236                         schedule();
 237                         remove_wait_queue(pkmap_map_wait, &wait);
 238                         lock_kmap();
 239
 240                         /* Somebody else might have mapped it while we slept */
 241                         if (page_address(page))
 242                                 return (unsigned long)page_address(page);
 243
 244                         /* Re-start */
 245                         goto start;
 246                 }
 247         }
 248         vaddr = PKMAP_ADDR(last_pkmap_nr);
 249         set_pte_at(&init_mm, vaddr,
 250                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 251
 252         pkmap_count[last_pkmap_nr] = 1;
 253         set_page_address(page, (void *)vaddr);
 254
 255         return vaddr;
 256 }
 257
 258 /**
 259  * kmap_high - map a highmem page into memory
 260  * @page: &struct page to map
 261  *
 262  * Returns the page's virtual memory address.
 263  *
 264  * We cannot call this from interrupts, as it may block.
 265  */
 266 void *kmap_high(struct page *page)
 267 {
 268         unsigned long vaddr;
 269
 270         /*
 271          * For highmem pages, we can't trust "virtual" until
 272          * after we have the lock.
 273          */
 274         lock_kmap();
 275         vaddr = (unsigned long)page_address(page);
 276         if (!vaddr)
 277                 vaddr = map_new_virtual(page);
 278         pkmap_count[PKMAP_NR(vaddr)]++;
 279         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
 280         unlock_kmap();
 281         return (void*) vaddr;
 282 }
 283
 284 EXPORT_SYMBOL(kmap_high);
 285
 286 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 287 /**
 288  * kmap_high_get - pin a highmem page into memory
 289  * @page: &struct page to pin
 290  *
 291  * Returns the page's current virtual memory address, or NULL if no mapping
 292  * exists.  If and only if a non null address is returned then a
 293  * matching call to kunmap_high() is necessary.
 294  *
 295  * This can be called from any context.
 296  */
 297 void *kmap_high_get(struct page *page)
 298 {
 299         unsigned long vaddr, flags;
 300
 301         lock_kmap_any(flags);
 302         vaddr = (unsigned long)page_address(page);
 303         if (vaddr) {
 304                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
 305                 pkmap_count[PKMAP_NR(vaddr)]++;
 306         }
 307         unlock_kmap_any(flags);
 308         return (void*) vaddr;
 309 }
 310 #endif
 311
 312 /**
 313  * kunmap_high - unmap a highmem page into memory
 314  * @page: &struct page to unmap
 315  *
 316  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
 317  * only from user context.
 318  */
 319 void kunmap_high(struct page *page)
 320 {
 321         unsigned long vaddr;
 322         unsigned long nr;
 323         unsigned long flags;
 324         int need_wakeup;
 325         unsigned int color = get_pkmap_color(page);
 326         wait_queue_head_t *pkmap_map_wait;
 327
 328         lock_kmap_any(flags);
 329         vaddr = (unsigned long)page_address(page);
 330         BUG_ON(!vaddr);
 331         nr = PKMAP_NR(vaddr);
 332
 333         /*
 334          * A count must never go down to zero
 335          * without a TLB flush!
 336          */
 337         need_wakeup = 0;
 338         switch (--pkmap_count[nr]) {
 339         case 0:
 340                 BUG();
 341         case 1:
 342                 /*
 343                  * Avoid an unnecessary wake_up() function call.
 344                  * The common case is pkmap_count[] == 1, but
 345                  * no waiters.
 346                  * The tasks queued in the wait-queue are guarded
 347                  * by both the lock in the wait-queue-head and by
 348                  * the kmap_lock.  As the kmap_lock is held here,
 349                  * no need for the wait-queue-head's lock.  Simply
 350                  * test if the queue is empty.
 351                  */
 352                 pkmap_map_wait = get_pkmap_wait_queue_head(color);
 353                 need_wakeup = waitqueue_active(pkmap_map_wait);
 354         }
 355         unlock_kmap_any(flags);
 356
 357         /* do wake-up, if needed, race-free outside of the spin lock */
 358         if (need_wakeup)
 359                 wake_up(pkmap_map_wait);
 360 }
 361 EXPORT_SYMBOL(kunmap_high);
 362 #endif /* CONFIG_HIGHMEM */
 363
 364 #ifdef CONFIG_KMAP_LOCAL
 365
 366 #include <asm/kmap_size.h>
 367
 368 static DEFINE_PER_CPU(int, __kmap_local_idx);
 369
 370 /*
 371  * With DEBUG_HIGHMEM the stack depth is doubled and every second
 372  * slot is unused which acts as a guard page
 373  */
 374 #ifdef CONFIG_DEBUG_HIGHMEM
 375 # define KM_INCR        2
 376 #else
 377 # define KM_INCR        1
 378 #endif
 379
 380 static inline int kmap_local_idx_push(void)
 381 {
 382         int idx = __this_cpu_add_return(__kmap_local_idx, KM_INCR) - 1;
 383
 384         WARN_ON_ONCE(in_irq() && !irqs_disabled());
 385         BUG_ON(idx >= KM_MAX_IDX);
 386         return idx;
 387 }
 388
 389 static inline int kmap_local_idx(void)
 390 {
 391         return __this_cpu_read(__kmap_local_idx) - 1;
 392 }
 393
 394 static inline void kmap_local_idx_pop(void)
 395 {
 396         int idx = __this_cpu_sub_return(__kmap_local_idx, KM_INCR);
 397
 398         BUG_ON(idx < 0);
 399 }
 400
 401 #ifndef arch_kmap_local_post_map
 402 # define arch_kmap_local_post_map(vaddr, pteval)        do { } while (0)
 403 #endif
 404
 405 #ifndef arch_kmap_local_pre_unmap
 406 # define arch_kmap_local_pre_unmap(vaddr)               do { } while (0)
 407 #endif
 408
 409 #ifndef arch_kmap_local_post_unmap
 410 # define arch_kmap_local_post_unmap(vaddr)              do { } while (0)
 411 #endif
 412
 413 #ifndef arch_kmap_local_map_idx
 414 #define arch_kmap_local_map_idx(idx, pfn)       kmap_local_calc_idx(idx)
 415 #endif
 416
 417 #ifndef arch_kmap_local_unmap_idx
 418 #define arch_kmap_local_unmap_idx(idx, vaddr)   kmap_local_calc_idx(idx)
 419 #endif
 420
 421 #ifndef arch_kmap_local_high_get
 422 static inline void *arch_kmap_local_high_get(struct page *page)
 423 {
 424         return NULL;
 425 }
 426 #endif
 427
 428 /* Unmap a local mapping which was obtained by kmap_high_get() */
 429 static inline void kmap_high_unmap_local(unsigned long vaddr)
 430 {
 431 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 432         if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP))
 433                 kunmap_high(pte_page(pkmap_page_table[PKMAP_NR(vaddr)]));
 434 #endif
 435 }
 436
 437 static inline int kmap_local_calc_idx(int idx)
 438 {
 439         return idx + KM_MAX_IDX * smp_processor_id();
 440 }
 441
 442 static pte_t *__kmap_pte;
 443
 444 static pte_t *kmap_get_pte(void)
 445 {
 446         if (!__kmap_pte)
 447                 __kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
 448         return __kmap_pte;
 449 }
 450
 451 void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
 452 {
 453         pte_t pteval, *kmap_pte = kmap_get_pte();
 454         unsigned long vaddr;
 455         int idx;
 456
 457         preempt_disable();
 458         idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
 459         vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
 460         BUG_ON(!pte_none(*(kmap_pte - idx)));
 461         pteval = pfn_pte(pfn, prot);
 462         set_pte_at(&init_mm, vaddr, kmap_pte - idx, pteval);
 463         arch_kmap_local_post_map(vaddr, pteval);
 464         preempt_enable();
 465
 466         return (void *)vaddr;
 467 }
 468 EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
 469
 470 void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
 471 {
 472         void *kmap;
 473
 474         if (!PageHighMem(page))
 475                 return page_address(page);
 476
 477         /* Try kmap_high_get() if architecture has it enabled */
 478         kmap = arch_kmap_local_high_get(page);
 479         if (kmap)
 480                 return kmap;
 481
 482         return __kmap_local_pfn_prot(page_to_pfn(page), prot);
 483 }
 484 EXPORT_SYMBOL(__kmap_local_page_prot);
 485
 486 void kunmap_local_indexed(void *vaddr)
 487 {
 488         unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
 489         pte_t *kmap_pte = kmap_get_pte();
 490         int idx;
 491
 492         if (addr < __fix_to_virt(FIX_KMAP_END) ||
 493             addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
 494                 WARN_ON_ONCE(addr < PAGE_OFFSET);
 495
 496                 /* Handle mappings which were obtained by kmap_high_get() */
 497                 kmap_high_unmap_local(addr);
 498                 return;
 499         }
 500
 501         preempt_disable();
 502         idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
 503         WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
 504
 505         arch_kmap_local_pre_unmap(addr);
 506         pte_clear(&init_mm, addr, kmap_pte - idx);
 507         arch_kmap_local_post_unmap(addr);
 508         kmap_local_idx_pop();
 509         preempt_enable();
 510 }
 511 EXPORT_SYMBOL(kunmap_local_indexed);
 512 #endif
 513
 514 #if defined(HASHED_PAGE_VIRTUAL)
 515
 516 #define PA_HASH_ORDER   7
 517
 518 /*
 519  * Describes one page->virtual association
 520  */
 521 struct page_address_map {
 522         struct page *page;
 523         void *virtual;
 524         struct list_head list;
 525 };
 526
 527 static struct page_address_map page_address_maps[LAST_PKMAP];
 528
 529 /*
 530  * Hash table bucket
 531  */
 532 static struct page_address_slot {
 533         struct list_head lh;                    /* List of page_address_maps */
 534         spinlock_t lock;                        /* Protect this bucket's list */
 535 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 536
 537 static struct page_address_slot *page_slot(const struct page *page)
 538 {
 539         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 540 }
 541
 542 /**
 543  * page_address - get the mapped virtual address of a page
 544  * @page: &struct page to get the virtual address of
 545  *
 546  * Returns the page's virtual address.
 547  */
 548 void *page_address(const struct page *page)
 549 {
 550         unsigned long flags;
 551         void *ret;
 552         struct page_address_slot *pas;
 553
 554         if (!PageHighMem(page))
 555                 return lowmem_page_address(page);
 556
 557         pas = page_slot(page);
 558         ret = NULL;
 559         spin_lock_irqsave(&pas->lock, flags);
 560         if (!list_empty(&pas->lh)) {
 561                 struct page_address_map *pam;
 562
 563                 list_for_each_entry(pam, &pas->lh, list) {
 564                         if (pam->page == page) {
 565                                 ret = pam->virtual;
 566                                 goto done;
 567                         }
 568                 }
 569         }
 570 done:
 571         spin_unlock_irqrestore(&pas->lock, flags);
 572         return ret;
 573 }
 574
 575 EXPORT_SYMBOL(page_address);
 576
 577 /**
 578  * set_page_address - set a page's virtual address
 579  * @page: &struct page to set
 580  * @virtual: virtual address to use
 581  */
 582 void set_page_address(struct page *page, void *virtual)
 583 {
 584         unsigned long flags;
 585         struct page_address_slot *pas;
 586         struct page_address_map *pam;
 587
 588         BUG_ON(!PageHighMem(page));
 589
 590         pas = page_slot(page);
 591         if (virtual) {          /* Add */
 592                 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
 593                 pam->page = page;
 594                 pam->virtual = virtual;
 595
 596                 spin_lock_irqsave(&pas->lock, flags);
 597                 list_add_tail(&pam->list, &pas->lh);
 598                 spin_unlock_irqrestore(&pas->lock, flags);
 599         } else {                /* Remove */
 600                 spin_lock_irqsave(&pas->lock, flags);
 601                 list_for_each_entry(pam, &pas->lh, list) {
 602                         if (pam->page == page) {
 603                                 list_del(&pam->list);
 604                                 spin_unlock_irqrestore(&pas->lock, flags);
 605                                 goto done;
 606                         }
 607                 }
 608                 spin_unlock_irqrestore(&pas->lock, flags);
 609         }
 610 done:
 611         return;
 612 }
 613
 614 void __init page_address_init(void)
 615 {
 616         int i;
 617
 618         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 619                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 620                 spin_lock_init(&page_address_htable[i].lock);
 621         }
 622 }
 623
 624 #endif  /* defined(HASHED_PAGE_VIRTUAL) */