xref: /openbmc/linux/mm/highmem.c (revision 1cb8f3e2d8fe7533c26df9925a83bd3d185b312e)
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  
363  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
364  void zero_user_segments(struct page *page, unsigned start1, unsigned end1,
365  		unsigned start2, unsigned end2)
366  {
367  	unsigned int i;
368  
369  	BUG_ON(end1 > page_size(page) || end2 > page_size(page));
370  
371  	for (i = 0; i < compound_nr(page); i++) {
372  		void *kaddr = NULL;
373  
374  		if (start1 < PAGE_SIZE || start2 < PAGE_SIZE)
375  			kaddr = kmap_atomic(page + i);
376  
377  		if (start1 >= PAGE_SIZE) {
378  			start1 -= PAGE_SIZE;
379  			end1 -= PAGE_SIZE;
380  		} else {
381  			unsigned this_end = min_t(unsigned, end1, PAGE_SIZE);
382  
383  			if (end1 > start1)
384  				memset(kaddr + start1, 0, this_end - start1);
385  			end1 -= this_end;
386  			start1 = 0;
387  		}
388  
389  		if (start2 >= PAGE_SIZE) {
390  			start2 -= PAGE_SIZE;
391  			end2 -= PAGE_SIZE;
392  		} else {
393  			unsigned this_end = min_t(unsigned, end2, PAGE_SIZE);
394  
395  			if (end2 > start2)
396  				memset(kaddr + start2, 0, this_end - start2);
397  			end2 -= this_end;
398  			start2 = 0;
399  		}
400  
401  		if (kaddr) {
402  			kunmap_atomic(kaddr);
403  			flush_dcache_page(page + i);
404  		}
405  
406  		if (!end1 && !end2)
407  			break;
408  	}
409  
410  	BUG_ON((start1 | start2 | end1 | end2) != 0);
411  }
412  EXPORT_SYMBOL(zero_user_segments);
413  #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
414  #endif /* CONFIG_HIGHMEM */
415  
416  #ifdef CONFIG_KMAP_LOCAL
417  
418  #include <asm/kmap_size.h>
419  
420  /*
421   * With DEBUG_KMAP_LOCAL the stack depth is doubled and every second
422   * slot is unused which acts as a guard page
423   */
424  #ifdef CONFIG_DEBUG_KMAP_LOCAL
425  # define KM_INCR	2
426  #else
427  # define KM_INCR	1
428  #endif
429  
430  static inline int kmap_local_idx_push(void)
431  {
432  	WARN_ON_ONCE(in_irq() && !irqs_disabled());
433  	current->kmap_ctrl.idx += KM_INCR;
434  	BUG_ON(current->kmap_ctrl.idx >= KM_MAX_IDX);
435  	return current->kmap_ctrl.idx - 1;
436  }
437  
438  static inline int kmap_local_idx(void)
439  {
440  	return current->kmap_ctrl.idx - 1;
441  }
442  
443  static inline void kmap_local_idx_pop(void)
444  {
445  	current->kmap_ctrl.idx -= KM_INCR;
446  	BUG_ON(current->kmap_ctrl.idx < 0);
447  }
448  
449  #ifndef arch_kmap_local_post_map
450  # define arch_kmap_local_post_map(vaddr, pteval)	do { } while (0)
451  #endif
452  
453  #ifndef arch_kmap_local_pre_unmap
454  # define arch_kmap_local_pre_unmap(vaddr)		do { } while (0)
455  #endif
456  
457  #ifndef arch_kmap_local_post_unmap
458  # define arch_kmap_local_post_unmap(vaddr)		do { } while (0)
459  #endif
460  
461  #ifndef arch_kmap_local_map_idx
462  #define arch_kmap_local_map_idx(idx, pfn)	kmap_local_calc_idx(idx)
463  #endif
464  
465  #ifndef arch_kmap_local_unmap_idx
466  #define arch_kmap_local_unmap_idx(idx, vaddr)	kmap_local_calc_idx(idx)
467  #endif
468  
469  #ifndef arch_kmap_local_high_get
470  static inline void *arch_kmap_local_high_get(struct page *page)
471  {
472  	return NULL;
473  }
474  #endif
475  
476  #ifndef arch_kmap_local_set_pte
477  #define arch_kmap_local_set_pte(mm, vaddr, ptep, ptev)	\
478  	set_pte_at(mm, vaddr, ptep, ptev)
479  #endif
480  
481  /* Unmap a local mapping which was obtained by kmap_high_get() */
482  static inline bool kmap_high_unmap_local(unsigned long vaddr)
483  {
484  #ifdef ARCH_NEEDS_KMAP_HIGH_GET
485  	if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
486  		kunmap_high(pte_page(pkmap_page_table[PKMAP_NR(vaddr)]));
487  		return true;
488  	}
489  #endif
490  	return false;
491  }
492  
493  static inline int kmap_local_calc_idx(int idx)
494  {
495  	return idx + KM_MAX_IDX * smp_processor_id();
496  }
497  
498  static pte_t *__kmap_pte;
499  
500  static pte_t *kmap_get_pte(void)
501  {
502  	if (!__kmap_pte)
503  		__kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
504  	return __kmap_pte;
505  }
506  
507  void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
508  {
509  	pte_t pteval, *kmap_pte = kmap_get_pte();
510  	unsigned long vaddr;
511  	int idx;
512  
513  	/*
514  	 * Disable migration so resulting virtual address is stable
515  	 * accross preemption.
516  	 */
517  	migrate_disable();
518  	preempt_disable();
519  	idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
520  	vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
521  	BUG_ON(!pte_none(*(kmap_pte - idx)));
522  	pteval = pfn_pte(pfn, prot);
523  	arch_kmap_local_set_pte(&init_mm, vaddr, kmap_pte - idx, pteval);
524  	arch_kmap_local_post_map(vaddr, pteval);
525  	current->kmap_ctrl.pteval[kmap_local_idx()] = pteval;
526  	preempt_enable();
527  
528  	return (void *)vaddr;
529  }
530  EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
531  
532  void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
533  {
534  	void *kmap;
535  
536  	/*
537  	 * To broaden the usage of the actual kmap_local() machinery always map
538  	 * pages when debugging is enabled and the architecture has no problems
539  	 * with alias mappings.
540  	 */
541  	if (!IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP) && !PageHighMem(page))
542  		return page_address(page);
543  
544  	/* Try kmap_high_get() if architecture has it enabled */
545  	kmap = arch_kmap_local_high_get(page);
546  	if (kmap)
547  		return kmap;
548  
549  	return __kmap_local_pfn_prot(page_to_pfn(page), prot);
550  }
551  EXPORT_SYMBOL(__kmap_local_page_prot);
552  
553  void kunmap_local_indexed(void *vaddr)
554  {
555  	unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
556  	pte_t *kmap_pte = kmap_get_pte();
557  	int idx;
558  
559  	if (addr < __fix_to_virt(FIX_KMAP_END) ||
560  	    addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
561  		if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP)) {
562  			/* This _should_ never happen! See above. */
563  			WARN_ON_ONCE(1);
564  			return;
565  		}
566  		/*
567  		 * Handle mappings which were obtained by kmap_high_get()
568  		 * first as the virtual address of such mappings is below
569  		 * PAGE_OFFSET. Warn for all other addresses which are in
570  		 * the user space part of the virtual address space.
571  		 */
572  		if (!kmap_high_unmap_local(addr))
573  			WARN_ON_ONCE(addr < PAGE_OFFSET);
574  		return;
575  	}
576  
577  	preempt_disable();
578  	idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
579  	WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
580  
581  	arch_kmap_local_pre_unmap(addr);
582  	pte_clear(&init_mm, addr, kmap_pte - idx);
583  	arch_kmap_local_post_unmap(addr);
584  	current->kmap_ctrl.pteval[kmap_local_idx()] = __pte(0);
585  	kmap_local_idx_pop();
586  	preempt_enable();
587  	migrate_enable();
588  }
589  EXPORT_SYMBOL(kunmap_local_indexed);
590  
591  /*
592   * Invoked before switch_to(). This is safe even when during or after
593   * clearing the maps an interrupt which needs a kmap_local happens because
594   * the task::kmap_ctrl.idx is not modified by the unmapping code so a
595   * nested kmap_local will use the next unused index and restore the index
596   * on unmap. The already cleared kmaps of the outgoing task are irrelevant
597   * because the interrupt context does not know about them. The same applies
598   * when scheduling back in for an interrupt which happens before the
599   * restore is complete.
600   */
601  void __kmap_local_sched_out(void)
602  {
603  	struct task_struct *tsk = current;
604  	pte_t *kmap_pte = kmap_get_pte();
605  	int i;
606  
607  	/* Clear kmaps */
608  	for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
609  		pte_t pteval = tsk->kmap_ctrl.pteval[i];
610  		unsigned long addr;
611  		int idx;
612  
613  		/* With debug all even slots are unmapped and act as guard */
614  		if (IS_ENABLED(CONFIG_DEBUG_HIGHMEM) && !(i & 0x01)) {
615  			WARN_ON_ONCE(!pte_none(pteval));
616  			continue;
617  		}
618  		if (WARN_ON_ONCE(pte_none(pteval)))
619  			continue;
620  
621  		/*
622  		 * This is a horrible hack for XTENSA to calculate the
623  		 * coloured PTE index. Uses the PFN encoded into the pteval
624  		 * and the map index calculation because the actual mapped
625  		 * virtual address is not stored in task::kmap_ctrl.
626  		 * For any sane architecture this is optimized out.
627  		 */
628  		idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
629  
630  		addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
631  		arch_kmap_local_pre_unmap(addr);
632  		pte_clear(&init_mm, addr, kmap_pte - idx);
633  		arch_kmap_local_post_unmap(addr);
634  	}
635  }
636  
637  void __kmap_local_sched_in(void)
638  {
639  	struct task_struct *tsk = current;
640  	pte_t *kmap_pte = kmap_get_pte();
641  	int i;
642  
643  	/* Restore kmaps */
644  	for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
645  		pte_t pteval = tsk->kmap_ctrl.pteval[i];
646  		unsigned long addr;
647  		int idx;
648  
649  		/* With debug all even slots are unmapped and act as guard */
650  		if (IS_ENABLED(CONFIG_DEBUG_HIGHMEM) && !(i & 0x01)) {
651  			WARN_ON_ONCE(!pte_none(pteval));
652  			continue;
653  		}
654  		if (WARN_ON_ONCE(pte_none(pteval)))
655  			continue;
656  
657  		/* See comment in __kmap_local_sched_out() */
658  		idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
659  		addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
660  		set_pte_at(&init_mm, addr, kmap_pte - idx, pteval);
661  		arch_kmap_local_post_map(addr, pteval);
662  	}
663  }
664  
665  void kmap_local_fork(struct task_struct *tsk)
666  {
667  	if (WARN_ON_ONCE(tsk->kmap_ctrl.idx))
668  		memset(&tsk->kmap_ctrl, 0, sizeof(tsk->kmap_ctrl));
669  }
670  
671  #endif
672  
673  #if defined(HASHED_PAGE_VIRTUAL)
674  
675  #define PA_HASH_ORDER	7
676  
677  /*
678   * Describes one page->virtual association
679   */
680  struct page_address_map {
681  	struct page *page;
682  	void *virtual;
683  	struct list_head list;
684  };
685  
686  static struct page_address_map page_address_maps[LAST_PKMAP];
687  
688  /*
689   * Hash table bucket
690   */
691  static struct page_address_slot {
692  	struct list_head lh;			/* List of page_address_maps */
693  	spinlock_t lock;			/* Protect this bucket's list */
694  } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
695  
696  static struct page_address_slot *page_slot(const struct page *page)
697  {
698  	return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
699  }
700  
701  /**
702   * page_address - get the mapped virtual address of a page
703   * @page: &struct page to get the virtual address of
704   *
705   * Returns the page's virtual address.
706   */
707  void *page_address(const struct page *page)
708  {
709  	unsigned long flags;
710  	void *ret;
711  	struct page_address_slot *pas;
712  
713  	if (!PageHighMem(page))
714  		return lowmem_page_address(page);
715  
716  	pas = page_slot(page);
717  	ret = NULL;
718  	spin_lock_irqsave(&pas->lock, flags);
719  	if (!list_empty(&pas->lh)) {
720  		struct page_address_map *pam;
721  
722  		list_for_each_entry(pam, &pas->lh, list) {
723  			if (pam->page == page) {
724  				ret = pam->virtual;
725  				goto done;
726  			}
727  		}
728  	}
729  done:
730  	spin_unlock_irqrestore(&pas->lock, flags);
731  	return ret;
732  }
733  
734  EXPORT_SYMBOL(page_address);
735  
736  /**
737   * set_page_address - set a page's virtual address
738   * @page: &struct page to set
739   * @virtual: virtual address to use
740   */
741  void set_page_address(struct page *page, void *virtual)
742  {
743  	unsigned long flags;
744  	struct page_address_slot *pas;
745  	struct page_address_map *pam;
746  
747  	BUG_ON(!PageHighMem(page));
748  
749  	pas = page_slot(page);
750  	if (virtual) {		/* Add */
751  		pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
752  		pam->page = page;
753  		pam->virtual = virtual;
754  
755  		spin_lock_irqsave(&pas->lock, flags);
756  		list_add_tail(&pam->list, &pas->lh);
757  		spin_unlock_irqrestore(&pas->lock, flags);
758  	} else {		/* Remove */
759  		spin_lock_irqsave(&pas->lock, flags);
760  		list_for_each_entry(pam, &pas->lh, list) {
761  			if (pam->page == page) {
762  				list_del(&pam->list);
763  				spin_unlock_irqrestore(&pas->lock, flags);
764  				goto done;
765  			}
766  		}
767  		spin_unlock_irqrestore(&pas->lock, flags);
768  	}
769  done:
770  	return;
771  }
772  
773  void __init page_address_init(void)
774  {
775  	int i;
776  
777  	for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
778  		INIT_LIST_HEAD(&page_address_htable[i].lh);
779  		spin_lock_init(&page_address_htable[i].lock);
780  	}
781  }
782  
783  #endif	/* defined(HASHED_PAGE_VIRTUAL) */
784