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