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, [email protected]
   6  *          Gerhard Wichert, Siemens AG, [email protected]
   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 <[email protected]>
  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/init.h>
  27 #include <linux/hash.h>
  28 #include <linux/highmem.h>
  29 #include <linux/kgdb.h>
  30 #include <asm/tlbflush.h>
  31 #include <linux/vmalloc.h>
  32
  33 /*
  34  * Virtual_count is not a pure "count".
  35  *  0 means that it is not mapped, and has not been mapped
  36  *    since a TLB flush - it is usable.
  37  *  1 means that there are no users, but it has been mapped
  38  *    since the last TLB flush - so we can't use it.
  39  *  n means that there are (n-1) current users of it.
  40  */
  41 #ifdef CONFIG_HIGHMEM
  42
  43 /*
  44  * Architecture with aliasing data cache may define the following family of
  45  * helper functions in its asm/highmem.h to control cache color of virtual
  46  * addresses where physical memory pages are mapped by kmap.
  47  */
  48 #ifndef get_pkmap_color
  49
  50 /*
  51  * Determine color of virtual address where the page should be mapped.
  52  */
  53 static inline unsigned int get_pkmap_color(struct page *page)
  54 {
  55         return 0;
  56 }
  57 #define get_pkmap_color get_pkmap_color
  58
  59 /*
  60  * Get next index for mapping inside PKMAP region for page with given color.
  61  */
  62 static inline unsigned int get_next_pkmap_nr(unsigned int color)
  63 {
  64         static unsigned int last_pkmap_nr;
  65
  66         last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
  67         return last_pkmap_nr;
  68 }
  69
  70 /*
  71  * Determine if page index inside PKMAP region (pkmap_nr) of given color
  72  * has wrapped around PKMAP region end. When this happens an attempt to
  73  * flush all unused PKMAP slots is made.
  74  */
  75 static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
  76 {
  77         return pkmap_nr == 0;
  78 }
  79
  80 /*
  81  * Get the number of PKMAP entries of the given color. If no free slot is
  82  * found after checking that many entries, kmap will sleep waiting for
  83  * someone to call kunmap and free PKMAP slot.
  84  */
  85 static inline int get_pkmap_entries_count(unsigned int color)
  86 {
  87         return LAST_PKMAP;
  88 }
  89
  90 /*
  91  * Get head of a wait queue for PKMAP entries of the given color.
  92  * Wait queues for different mapping colors should be independent to avoid
  93  * unnecessary wakeups caused by freeing of slots of other colors.
  94  */
  95 static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
  96 {
  97         static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  98
  99         return &pkmap_map_wait;
 100 }
 101 #endif
 102
 103 atomic_long_t _totalhigh_pages __read_mostly;
 104 EXPORT_SYMBOL(_totalhigh_pages);
 105
 106 unsigned int __nr_free_highpages(void)
 107 {
 108         struct zone *zone;
 109         unsigned int pages = 0;
 110
 111         for_each_populated_zone(zone) {
 112                 if (is_highmem(zone))
 113                         pages += zone_page_state(zone, NR_FREE_PAGES);
 114         }
 115
 116         return pages;
 117 }
 118
 119 static int pkmap_count[LAST_PKMAP];
 120 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
 121
 122 pte_t *pkmap_page_table;
 123
 124 /*
 125  * Most architectures have no use for kmap_high_get(), so let's abstract
 126  * the disabling of IRQ out of the locking in that case to save on a
 127  * potential useless overhead.
 128  */
 129 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 130 #define lock_kmap()             spin_lock_irq(&kmap_lock)
 131 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
 132 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
 133 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
 134 #else
 135 #define lock_kmap()             spin_lock(&kmap_lock)
 136 #define unlock_kmap()           spin_unlock(&kmap_lock)
 137 #define lock_kmap_any(flags)    \
 138                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
 139 #define unlock_kmap_any(flags)  \
 140                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
 141 #endif
 142
 143 struct page *__kmap_to_page(void *vaddr)
 144 {
 145         unsigned long addr = (unsigned long)vaddr;
 146
 147         if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
 148                 int i = PKMAP_NR(addr);
 149
 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 EXPORT_SYMBOL(kmap_high);
 284
 285 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 286 /**
 287  * kmap_high_get - pin a highmem page into memory
 288  * @page: &struct page to pin
 289  *
 290  * Returns the page's current virtual memory address, or NULL if no mapping
 291  * exists.  If and only if a non null address is returned then a
 292  * matching call to kunmap_high() is necessary.
 293  *
 294  * This can be called from any context.
 295  */
 296 void *kmap_high_get(struct page *page)
 297 {
 298         unsigned long vaddr, flags;
 299
 300         lock_kmap_any(flags);
 301         vaddr = (unsigned long)page_address(page);
 302         if (vaddr) {
 303                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
 304                 pkmap_count[PKMAP_NR(vaddr)]++;
 305         }
 306         unlock_kmap_any(flags);
 307         return (void *) vaddr;
 308 }
 309 #endif
 310
 311 /**
 312  * kunmap_high - unmap a highmem page into memory
 313  * @page: &struct page to unmap
 314  *
 315  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
 316  * only from user context.
 317  */
 318 void kunmap_high(struct page *page)
 319 {
 320         unsigned long vaddr;
 321         unsigned long nr;
 322         unsigned long flags;
 323         int need_wakeup;
 324         unsigned int color = get_pkmap_color(page);
 325         wait_queue_head_t *pkmap_map_wait;
 326
 327         lock_kmap_any(flags);
 328         vaddr = (unsigned long)page_address(page);
 329         BUG_ON(!vaddr);
 330         nr = PKMAP_NR(vaddr);
 331
 332         /*
 333          * A count must never go down to zero
 334          * without a TLB flush!
 335          */
 336         need_wakeup = 0;
 337         switch (--pkmap_count[nr]) {
 338         case 0:
 339                 BUG();
 340         case 1:
 341                 /*
 342                  * Avoid an unnecessary wake_up() function call.
 343                  * The common case is pkmap_count[] == 1, but
 344                  * no waiters.
 345                  * The tasks queued in the wait-queue are guarded
 346                  * by both the lock in the wait-queue-head and by
 347                  * the kmap_lock.  As the kmap_lock is held here,
 348                  * no need for the wait-queue-head's lock.  Simply
 349                  * test if the queue is empty.
 350                  */
 351                 pkmap_map_wait = get_pkmap_wait_queue_head(color);
 352                 need_wakeup = waitqueue_active(pkmap_map_wait);
 353         }
 354         unlock_kmap_any(flags);
 355
 356         /* do wake-up, if needed, race-free outside of the spin lock */
 357         if (need_wakeup)
 358                 wake_up(pkmap_map_wait);
 359 }
 360 EXPORT_SYMBOL(kunmap_high);
 361
 362 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 363 void zero_user_segments(struct page *page, unsigned start1, unsigned end1,
 364                 unsigned start2, unsigned end2)
 365 {
 366         unsigned int i;
 367
 368         BUG_ON(end1 > page_size(page) || end2 > page_size(page));
 369
 370         if (start1 >= end1)
 371                 start1 = end1 = 0;
 372         if (start2 >= end2)
 373                 start2 = end2 = 0;
 374
 375         for (i = 0; i < compound_nr(page); i++) {
 376                 void *kaddr = NULL;
 377
 378                 if (start1 >= PAGE_SIZE) {
 379                         start1 -= PAGE_SIZE;
 380                         end1 -= PAGE_SIZE;
 381                 } else {
 382                         unsigned this_end = min_t(unsigned, end1, PAGE_SIZE);
 383
 384                         if (end1 > start1) {
 385                                 kaddr = kmap_atomic(page + i);
 386                                 memset(kaddr + start1, 0, this_end - start1);
 387                         }
 388                         end1 -= this_end;
 389                         start1 = 0;
 390                 }
 391
 392                 if (start2 >= PAGE_SIZE) {
 393                         start2 -= PAGE_SIZE;
 394                         end2 -= PAGE_SIZE;
 395                 } else {
 396                         unsigned this_end = min_t(unsigned, end2, PAGE_SIZE);
 397
 398                         if (end2 > start2) {
 399                                 if (!kaddr)
 400                                         kaddr = kmap_atomic(page + i);
 401                                 memset(kaddr + start2, 0, this_end - start2);
 402                         }
 403                         end2 -= this_end;
 404                         start2 = 0;
 405                 }
 406
 407                 if (kaddr) {
 408                         kunmap_atomic(kaddr);
 409                         flush_dcache_page(page + i);
 410                 }
 411
 412                 if (!end1 && !end2)
 413                         break;
 414         }
 415
 416         BUG_ON((start1 | start2 | end1 | end2) != 0);
 417 }
 418 EXPORT_SYMBOL(zero_user_segments);
 419 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 420 #endif /* CONFIG_HIGHMEM */
 421
 422 #ifdef CONFIG_KMAP_LOCAL
 423
 424 #include <asm/kmap_size.h>
 425
 426 /*
 427  * With DEBUG_KMAP_LOCAL the stack depth is doubled and every second
 428  * slot is unused which acts as a guard page
 429  */
 430 #ifdef CONFIG_DEBUG_KMAP_LOCAL
 431 # define KM_INCR        2
 432 #else
 433 # define KM_INCR        1
 434 #endif
 435
 436 static inline int kmap_local_idx_push(void)
 437 {
 438         WARN_ON_ONCE(in_hardirq() && !irqs_disabled());
 439         current->kmap_ctrl.idx += KM_INCR;
 440         BUG_ON(current->kmap_ctrl.idx >= KM_MAX_IDX);
 441         return current->kmap_ctrl.idx - 1;
 442 }
 443
 444 static inline int kmap_local_idx(void)
 445 {
 446         return current->kmap_ctrl.idx - 1;
 447 }
 448
 449 static inline void kmap_local_idx_pop(void)
 450 {
 451         current->kmap_ctrl.idx -= KM_INCR;
 452         BUG_ON(current->kmap_ctrl.idx < 0);
 453 }
 454
 455 #ifndef arch_kmap_local_post_map
 456 # define arch_kmap_local_post_map(vaddr, pteval)        do { } while (0)
 457 #endif
 458
 459 #ifndef arch_kmap_local_pre_unmap
 460 # define arch_kmap_local_pre_unmap(vaddr)               do { } while (0)
 461 #endif
 462
 463 #ifndef arch_kmap_local_post_unmap
 464 # define arch_kmap_local_post_unmap(vaddr)              do { } while (0)
 465 #endif
 466
 467 #ifndef arch_kmap_local_map_idx
 468 #define arch_kmap_local_map_idx(idx, pfn)       kmap_local_calc_idx(idx)
 469 #endif
 470
 471 #ifndef arch_kmap_local_unmap_idx
 472 #define arch_kmap_local_unmap_idx(idx, vaddr)   kmap_local_calc_idx(idx)
 473 #endif
 474
 475 #ifndef arch_kmap_local_high_get
 476 static inline void *arch_kmap_local_high_get(struct page *page)
 477 {
 478         return NULL;
 479 }
 480 #endif
 481
 482 #ifndef arch_kmap_local_set_pte
 483 #define arch_kmap_local_set_pte(mm, vaddr, ptep, ptev)  \
 484         set_pte_at(mm, vaddr, ptep, ptev)
 485 #endif
 486
 487 /* Unmap a local mapping which was obtained by kmap_high_get() */
 488 static inline bool kmap_high_unmap_local(unsigned long vaddr)
 489 {
 490 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
 491         if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
 492                 kunmap_high(pte_page(pkmap_page_table[PKMAP_NR(vaddr)]));
 493                 return true;
 494         }
 495 #endif
 496         return false;
 497 }
 498
 499 static inline int kmap_local_calc_idx(int idx)
 500 {
 501         return idx + KM_MAX_IDX * smp_processor_id();
 502 }
 503
 504 static pte_t *__kmap_pte;
 505
 506 static pte_t *kmap_get_pte(void)
 507 {
 508         if (!__kmap_pte)
 509                 __kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
 510         return __kmap_pte;
 511 }
 512
 513 void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
 514 {
 515         pte_t pteval, *kmap_pte = kmap_get_pte();
 516         unsigned long vaddr;
 517         int idx;
 518
 519         /*
 520          * Disable migration so resulting virtual address is stable
 521          * across preemption.
 522          */
 523         migrate_disable();
 524         preempt_disable();
 525         idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
 526         vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
 527         BUG_ON(!pte_none(*(kmap_pte - idx)));
 528         pteval = pfn_pte(pfn, prot);
 529         arch_kmap_local_set_pte(&init_mm, vaddr, kmap_pte - idx, pteval);
 530         arch_kmap_local_post_map(vaddr, pteval);
 531         current->kmap_ctrl.pteval[kmap_local_idx()] = pteval;
 532         preempt_enable();
 533
 534         return (void *)vaddr;
 535 }
 536 EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
 537
 538 void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
 539 {
 540         void *kmap;
 541
 542         /*
 543          * To broaden the usage of the actual kmap_local() machinery always map
 544          * pages when debugging is enabled and the architecture has no problems
 545          * with alias mappings.
 546          */
 547         if (!IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP) && !PageHighMem(page))
 548                 return page_address(page);
 549
 550         /* Try kmap_high_get() if architecture has it enabled */
 551         kmap = arch_kmap_local_high_get(page);
 552         if (kmap)
 553                 return kmap;
 554
 555         return __kmap_local_pfn_prot(page_to_pfn(page), prot);
 556 }
 557 EXPORT_SYMBOL(__kmap_local_page_prot);
 558
 559 void kunmap_local_indexed(void *vaddr)
 560 {
 561         unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
 562         pte_t *kmap_pte = kmap_get_pte();
 563         int idx;
 564
 565         if (addr < __fix_to_virt(FIX_KMAP_END) ||
 566             addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
 567                 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP)) {
 568                         /* This _should_ never happen! See above. */
 569                         WARN_ON_ONCE(1);
 570                         return;
 571                 }
 572                 /*
 573                  * Handle mappings which were obtained by kmap_high_get()
 574                  * first as the virtual address of such mappings is below
 575                  * PAGE_OFFSET. Warn for all other addresses which are in
 576                  * the user space part of the virtual address space.
 577                  */
 578                 if (!kmap_high_unmap_local(addr))
 579                         WARN_ON_ONCE(addr < PAGE_OFFSET);
 580                 return;
 581         }
 582
 583         preempt_disable();
 584         idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
 585         WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
 586
 587         arch_kmap_local_pre_unmap(addr);
 588         pte_clear(&init_mm, addr, kmap_pte - idx);
 589         arch_kmap_local_post_unmap(addr);
 590         current->kmap_ctrl.pteval[kmap_local_idx()] = __pte(0);
 591         kmap_local_idx_pop();
 592         preempt_enable();
 593         migrate_enable();
 594 }
 595 EXPORT_SYMBOL(kunmap_local_indexed);
 596
 597 /*
 598  * Invoked before switch_to(). This is safe even when during or after
 599  * clearing the maps an interrupt which needs a kmap_local happens because
 600  * the task::kmap_ctrl.idx is not modified by the unmapping code so a
 601  * nested kmap_local will use the next unused index and restore the index
 602  * on unmap. The already cleared kmaps of the outgoing task are irrelevant
 603  * because the interrupt context does not know about them. The same applies
 604  * when scheduling back in for an interrupt which happens before the
 605  * restore is complete.
 606  */
 607 void __kmap_local_sched_out(void)
 608 {
 609         struct task_struct *tsk = current;
 610         pte_t *kmap_pte = kmap_get_pte();
 611         int i;
 612
 613         /* Clear kmaps */
 614         for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
 615                 pte_t pteval = tsk->kmap_ctrl.pteval[i];
 616                 unsigned long addr;
 617                 int idx;
 618
 619                 /* With debug all even slots are unmapped and act as guard */
 620                 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
 621                         WARN_ON_ONCE(!pte_none(pteval));
 622                         continue;
 623                 }
 624                 if (WARN_ON_ONCE(pte_none(pteval)))
 625                         continue;
 626
 627                 /*
 628                  * This is a horrible hack for XTENSA to calculate the
 629                  * coloured PTE index. Uses the PFN encoded into the pteval
 630                  * and the map index calculation because the actual mapped
 631                  * virtual address is not stored in task::kmap_ctrl.
 632                  * For any sane architecture this is optimized out.
 633                  */
 634                 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
 635
 636                 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
 637                 arch_kmap_local_pre_unmap(addr);
 638                 pte_clear(&init_mm, addr, kmap_pte - idx);
 639                 arch_kmap_local_post_unmap(addr);
 640         }
 641 }
 642
 643 void __kmap_local_sched_in(void)
 644 {
 645         struct task_struct *tsk = current;
 646         pte_t *kmap_pte = kmap_get_pte();
 647         int i;
 648
 649         /* Restore kmaps */
 650         for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
 651                 pte_t pteval = tsk->kmap_ctrl.pteval[i];
 652                 unsigned long addr;
 653                 int idx;
 654
 655                 /* With debug all even slots are unmapped and act as guard */
 656                 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
 657                         WARN_ON_ONCE(!pte_none(pteval));
 658                         continue;
 659                 }
 660                 if (WARN_ON_ONCE(pte_none(pteval)))
 661                         continue;
 662
 663                 /* See comment in __kmap_local_sched_out() */
 664                 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
 665                 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
 666                 set_pte_at(&init_mm, addr, kmap_pte - idx, pteval);
 667                 arch_kmap_local_post_map(addr, pteval);
 668         }
 669 }
 670
 671 void kmap_local_fork(struct task_struct *tsk)
 672 {
 673         if (WARN_ON_ONCE(tsk->kmap_ctrl.idx))
 674                 memset(&tsk->kmap_ctrl, 0, sizeof(tsk->kmap_ctrl));
 675 }
 676
 677 #endif
 678
 679 #if defined(HASHED_PAGE_VIRTUAL)
 680
 681 #define PA_HASH_ORDER   7
 682
 683 /*
 684  * Describes one page->virtual association
 685  */
 686 struct page_address_map {
 687         struct page *page;
 688         void *virtual;
 689         struct list_head list;
 690 };
 691
 692 static struct page_address_map page_address_maps[LAST_PKMAP];
 693
 694 /*
 695  * Hash table bucket
 696  */
 697 static struct page_address_slot {
 698         struct list_head lh;                    /* List of page_address_maps */
 699         spinlock_t lock;                        /* Protect this bucket's list */
 700 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 701
 702 static struct page_address_slot *page_slot(const struct page *page)
 703 {
 704         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 705 }
 706
 707 /**
 708  * page_address - get the mapped virtual address of a page
 709  * @page: &struct page to get the virtual address of
 710  *
 711  * Returns the page's virtual address.
 712  */
 713 void *page_address(const struct page *page)
 714 {
 715         unsigned long flags;
 716         void *ret;
 717         struct page_address_slot *pas;
 718
 719         if (!PageHighMem(page))
 720                 return lowmem_page_address(page);
 721
 722         pas = page_slot(page);
 723         ret = NULL;
 724         spin_lock_irqsave(&pas->lock, flags);
 725         if (!list_empty(&pas->lh)) {
 726                 struct page_address_map *pam;
 727
 728                 list_for_each_entry(pam, &pas->lh, list) {
 729                         if (pam->page == page) {
 730                                 ret = pam->virtual;
 731                                 goto done;
 732                         }
 733                 }
 734         }
 735 done:
 736         spin_unlock_irqrestore(&pas->lock, flags);
 737         return ret;
 738 }
 739 EXPORT_SYMBOL(page_address);
 740
 741 /**
 742  * set_page_address - set a page's virtual address
 743  * @page: &struct page to set
 744  * @virtual: virtual address to use
 745  */
 746 void set_page_address(struct page *page, void *virtual)
 747 {
 748         unsigned long flags;
 749         struct page_address_slot *pas;
 750         struct page_address_map *pam;
 751
 752         BUG_ON(!PageHighMem(page));
 753
 754         pas = page_slot(page);
 755         if (virtual) {          /* Add */
 756                 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
 757                 pam->page = page;
 758                 pam->virtual = virtual;
 759
 760                 spin_lock_irqsave(&pas->lock, flags);
 761                 list_add_tail(&pam->list, &pas->lh);
 762                 spin_unlock_irqrestore(&pas->lock, flags);
 763         } else {                /* Remove */
 764                 spin_lock_irqsave(&pas->lock, flags);
 765                 list_for_each_entry(pam, &pas->lh, list) {
 766                         if (pam->page == page) {
 767                                 list_del(&pam->list);
 768                                 spin_unlock_irqrestore(&pas->lock, flags);
 769                                 goto done;
 770                         }
 771                 }
 772                 spin_unlock_irqrestore(&pas->lock, flags);
 773         }
 774 done:
 775         return;
 776 }
 777
 778 void __init page_address_init(void)
 779 {
 780         int i;
 781
 782         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 783                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 784                 spin_lock_init(&page_address_htable[i].lock);
 785         }
 786 }
 787
 788 #endif  /* defined(HASHED_PAGE_VIRTUAL) */