xref: /openbmc/linux/mm/highmem.c (revision 6dc6657d)
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/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 #ifdef CONFIG_KMAP_LOCAL
34 static inline int kmap_local_calc_idx(int idx)
35 {
36 	return idx + KM_MAX_IDX * smp_processor_id();
37 }
38 
39 #ifndef arch_kmap_local_map_idx
40 #define arch_kmap_local_map_idx(idx, pfn)	kmap_local_calc_idx(idx)
41 #endif
42 #endif /* CONFIG_KMAP_LOCAL */
43 
44 /*
45  * Virtual_count is not a pure "count".
46  *  0 means that it is not mapped, and has not been mapped
47  *    since a TLB flush - it is usable.
48  *  1 means that there are no users, but it has been mapped
49  *    since the last TLB flush - so we can't use it.
50  *  n means that there are (n-1) current users of it.
51  */
52 #ifdef CONFIG_HIGHMEM
53 
54 /*
55  * Architecture with aliasing data cache may define the following family of
56  * helper functions in its asm/highmem.h to control cache color of virtual
57  * addresses where physical memory pages are mapped by kmap.
58  */
59 #ifndef get_pkmap_color
60 
61 /*
62  * Determine color of virtual address where the page should be mapped.
63  */
64 static inline unsigned int get_pkmap_color(struct page *page)
65 {
66 	return 0;
67 }
68 #define get_pkmap_color get_pkmap_color
69 
70 /*
71  * Get next index for mapping inside PKMAP region for page with given color.
72  */
73 static inline unsigned int get_next_pkmap_nr(unsigned int color)
74 {
75 	static unsigned int last_pkmap_nr;
76 
77 	last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
78 	return last_pkmap_nr;
79 }
80 
81 /*
82  * Determine if page index inside PKMAP region (pkmap_nr) of given color
83  * has wrapped around PKMAP region end. When this happens an attempt to
84  * flush all unused PKMAP slots is made.
85  */
86 static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
87 {
88 	return pkmap_nr == 0;
89 }
90 
91 /*
92  * Get the number of PKMAP entries of the given color. If no free slot is
93  * found after checking that many entries, kmap will sleep waiting for
94  * someone to call kunmap and free PKMAP slot.
95  */
96 static inline int get_pkmap_entries_count(unsigned int color)
97 {
98 	return LAST_PKMAP;
99 }
100 
101 /*
102  * Get head of a wait queue for PKMAP entries of the given color.
103  * Wait queues for different mapping colors should be independent to avoid
104  * unnecessary wakeups caused by freeing of slots of other colors.
105  */
106 static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
107 {
108 	static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
109 
110 	return &pkmap_map_wait;
111 }
112 #endif
113 
114 atomic_long_t _totalhigh_pages __read_mostly;
115 EXPORT_SYMBOL(_totalhigh_pages);
116 
117 unsigned int __nr_free_highpages(void)
118 {
119 	struct zone *zone;
120 	unsigned int pages = 0;
121 
122 	for_each_populated_zone(zone) {
123 		if (is_highmem(zone))
124 			pages += zone_page_state(zone, NR_FREE_PAGES);
125 	}
126 
127 	return pages;
128 }
129 
130 static int pkmap_count[LAST_PKMAP];
131 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
132 
133 pte_t *pkmap_page_table;
134 
135 /*
136  * Most architectures have no use for kmap_high_get(), so let's abstract
137  * the disabling of IRQ out of the locking in that case to save on a
138  * potential useless overhead.
139  */
140 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
141 #define lock_kmap()             spin_lock_irq(&kmap_lock)
142 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
143 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
144 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
145 #else
146 #define lock_kmap()             spin_lock(&kmap_lock)
147 #define unlock_kmap()           spin_unlock(&kmap_lock)
148 #define lock_kmap_any(flags)    \
149 		do { spin_lock(&kmap_lock); (void)(flags); } while (0)
150 #define unlock_kmap_any(flags)  \
151 		do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
152 #endif
153 
154 struct page *__kmap_to_page(void *vaddr)
155 {
156 	unsigned long base = (unsigned long) vaddr & PAGE_MASK;
157 	struct kmap_ctrl *kctrl = &current->kmap_ctrl;
158 	unsigned long addr = (unsigned long)vaddr;
159 	int i;
160 
161 	/* kmap() mappings */
162 	if (WARN_ON_ONCE(addr >= PKMAP_ADDR(0) &&
163 			 addr < PKMAP_ADDR(LAST_PKMAP)))
164 		return pte_page(ptep_get(&pkmap_page_table[PKMAP_NR(addr)]));
165 
166 	/* kmap_local_page() mappings */
167 	if (WARN_ON_ONCE(base >= __fix_to_virt(FIX_KMAP_END) &&
168 			 base < __fix_to_virt(FIX_KMAP_BEGIN))) {
169 		for (i = 0; i < kctrl->idx; i++) {
170 			unsigned long base_addr;
171 			int idx;
172 
173 			idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
174 			base_addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
175 
176 			if (base_addr == base)
177 				return pte_page(kctrl->pteval[i]);
178 		}
179 	}
180 
181 	return virt_to_page(vaddr);
182 }
183 EXPORT_SYMBOL(__kmap_to_page);
184 
185 static void flush_all_zero_pkmaps(void)
186 {
187 	int i;
188 	int need_flush = 0;
189 
190 	flush_cache_kmaps();
191 
192 	for (i = 0; i < LAST_PKMAP; i++) {
193 		struct page *page;
194 		pte_t ptent;
195 
196 		/*
197 		 * zero means we don't have anything to do,
198 		 * >1 means that it is still in use. Only
199 		 * a count of 1 means that it is free but
200 		 * needs to be unmapped
201 		 */
202 		if (pkmap_count[i] != 1)
203 			continue;
204 		pkmap_count[i] = 0;
205 
206 		/* sanity check */
207 		ptent = ptep_get(&pkmap_page_table[i]);
208 		BUG_ON(pte_none(ptent));
209 
210 		/*
211 		 * Don't need an atomic fetch-and-clear op here;
212 		 * no-one has the page mapped, and cannot get at
213 		 * its virtual address (and hence PTE) without first
214 		 * getting the kmap_lock (which is held here).
215 		 * So no dangers, even with speculative execution.
216 		 */
217 		page = pte_page(ptent);
218 		pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
219 
220 		set_page_address(page, NULL);
221 		need_flush = 1;
222 	}
223 	if (need_flush)
224 		flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
225 }
226 
227 void __kmap_flush_unused(void)
228 {
229 	lock_kmap();
230 	flush_all_zero_pkmaps();
231 	unlock_kmap();
232 }
233 
234 static inline unsigned long map_new_virtual(struct page *page)
235 {
236 	unsigned long vaddr;
237 	int count;
238 	unsigned int last_pkmap_nr;
239 	unsigned int color = get_pkmap_color(page);
240 
241 start:
242 	count = get_pkmap_entries_count(color);
243 	/* Find an empty entry */
244 	for (;;) {
245 		last_pkmap_nr = get_next_pkmap_nr(color);
246 		if (no_more_pkmaps(last_pkmap_nr, color)) {
247 			flush_all_zero_pkmaps();
248 			count = get_pkmap_entries_count(color);
249 		}
250 		if (!pkmap_count[last_pkmap_nr])
251 			break;	/* Found a usable entry */
252 		if (--count)
253 			continue;
254 
255 		/*
256 		 * Sleep for somebody else to unmap their entries
257 		 */
258 		{
259 			DECLARE_WAITQUEUE(wait, current);
260 			wait_queue_head_t *pkmap_map_wait =
261 				get_pkmap_wait_queue_head(color);
262 
263 			__set_current_state(TASK_UNINTERRUPTIBLE);
264 			add_wait_queue(pkmap_map_wait, &wait);
265 			unlock_kmap();
266 			schedule();
267 			remove_wait_queue(pkmap_map_wait, &wait);
268 			lock_kmap();
269 
270 			/* Somebody else might have mapped it while we slept */
271 			if (page_address(page))
272 				return (unsigned long)page_address(page);
273 
274 			/* Re-start */
275 			goto start;
276 		}
277 	}
278 	vaddr = PKMAP_ADDR(last_pkmap_nr);
279 	set_pte_at(&init_mm, vaddr,
280 		   &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
281 
282 	pkmap_count[last_pkmap_nr] = 1;
283 	set_page_address(page, (void *)vaddr);
284 
285 	return vaddr;
286 }
287 
288 /**
289  * kmap_high - map a highmem page into memory
290  * @page: &struct page to map
291  *
292  * Returns the page's virtual memory address.
293  *
294  * We cannot call this from interrupts, as it may block.
295  */
296 void *kmap_high(struct page *page)
297 {
298 	unsigned long vaddr;
299 
300 	/*
301 	 * For highmem pages, we can't trust "virtual" until
302 	 * after we have the lock.
303 	 */
304 	lock_kmap();
305 	vaddr = (unsigned long)page_address(page);
306 	if (!vaddr)
307 		vaddr = map_new_virtual(page);
308 	pkmap_count[PKMAP_NR(vaddr)]++;
309 	BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
310 	unlock_kmap();
311 	return (void *) vaddr;
312 }
313 EXPORT_SYMBOL(kmap_high);
314 
315 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
316 /**
317  * kmap_high_get - pin a highmem page into memory
318  * @page: &struct page to pin
319  *
320  * Returns the page's current virtual memory address, or NULL if no mapping
321  * exists.  If and only if a non null address is returned then a
322  * matching call to kunmap_high() is necessary.
323  *
324  * This can be called from any context.
325  */
326 void *kmap_high_get(struct page *page)
327 {
328 	unsigned long vaddr, flags;
329 
330 	lock_kmap_any(flags);
331 	vaddr = (unsigned long)page_address(page);
332 	if (vaddr) {
333 		BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
334 		pkmap_count[PKMAP_NR(vaddr)]++;
335 	}
336 	unlock_kmap_any(flags);
337 	return (void *) vaddr;
338 }
339 #endif
340 
341 /**
342  * kunmap_high - unmap a highmem page into memory
343  * @page: &struct page to unmap
344  *
345  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
346  * only from user context.
347  */
348 void kunmap_high(struct page *page)
349 {
350 	unsigned long vaddr;
351 	unsigned long nr;
352 	unsigned long flags;
353 	int need_wakeup;
354 	unsigned int color = get_pkmap_color(page);
355 	wait_queue_head_t *pkmap_map_wait;
356 
357 	lock_kmap_any(flags);
358 	vaddr = (unsigned long)page_address(page);
359 	BUG_ON(!vaddr);
360 	nr = PKMAP_NR(vaddr);
361 
362 	/*
363 	 * A count must never go down to zero
364 	 * without a TLB flush!
365 	 */
366 	need_wakeup = 0;
367 	switch (--pkmap_count[nr]) {
368 	case 0:
369 		BUG();
370 	case 1:
371 		/*
372 		 * Avoid an unnecessary wake_up() function call.
373 		 * The common case is pkmap_count[] == 1, but
374 		 * no waiters.
375 		 * The tasks queued in the wait-queue are guarded
376 		 * by both the lock in the wait-queue-head and by
377 		 * the kmap_lock.  As the kmap_lock is held here,
378 		 * no need for the wait-queue-head's lock.  Simply
379 		 * test if the queue is empty.
380 		 */
381 		pkmap_map_wait = get_pkmap_wait_queue_head(color);
382 		need_wakeup = waitqueue_active(pkmap_map_wait);
383 	}
384 	unlock_kmap_any(flags);
385 
386 	/* do wake-up, if needed, race-free outside of the spin lock */
387 	if (need_wakeup)
388 		wake_up(pkmap_map_wait);
389 }
390 EXPORT_SYMBOL(kunmap_high);
391 
392 void zero_user_segments(struct page *page, unsigned start1, unsigned end1,
393 		unsigned start2, unsigned end2)
394 {
395 	unsigned int i;
396 
397 	BUG_ON(end1 > page_size(page) || end2 > page_size(page));
398 
399 	if (start1 >= end1)
400 		start1 = end1 = 0;
401 	if (start2 >= end2)
402 		start2 = end2 = 0;
403 
404 	for (i = 0; i < compound_nr(page); i++) {
405 		void *kaddr = NULL;
406 
407 		if (start1 >= PAGE_SIZE) {
408 			start1 -= PAGE_SIZE;
409 			end1 -= PAGE_SIZE;
410 		} else {
411 			unsigned this_end = min_t(unsigned, end1, PAGE_SIZE);
412 
413 			if (end1 > start1) {
414 				kaddr = kmap_local_page(page + i);
415 				memset(kaddr + start1, 0, this_end - start1);
416 			}
417 			end1 -= this_end;
418 			start1 = 0;
419 		}
420 
421 		if (start2 >= PAGE_SIZE) {
422 			start2 -= PAGE_SIZE;
423 			end2 -= PAGE_SIZE;
424 		} else {
425 			unsigned this_end = min_t(unsigned, end2, PAGE_SIZE);
426 
427 			if (end2 > start2) {
428 				if (!kaddr)
429 					kaddr = kmap_local_page(page + i);
430 				memset(kaddr + start2, 0, this_end - start2);
431 			}
432 			end2 -= this_end;
433 			start2 = 0;
434 		}
435 
436 		if (kaddr) {
437 			kunmap_local(kaddr);
438 			flush_dcache_page(page + i);
439 		}
440 
441 		if (!end1 && !end2)
442 			break;
443 	}
444 
445 	BUG_ON((start1 | start2 | end1 | end2) != 0);
446 }
447 EXPORT_SYMBOL(zero_user_segments);
448 #endif /* CONFIG_HIGHMEM */
449 
450 #ifdef CONFIG_KMAP_LOCAL
451 
452 #include <asm/kmap_size.h>
453 
454 /*
455  * With DEBUG_KMAP_LOCAL the stack depth is doubled and every second
456  * slot is unused which acts as a guard page
457  */
458 #ifdef CONFIG_DEBUG_KMAP_LOCAL
459 # define KM_INCR	2
460 #else
461 # define KM_INCR	1
462 #endif
463 
464 static inline int kmap_local_idx_push(void)
465 {
466 	WARN_ON_ONCE(in_hardirq() && !irqs_disabled());
467 	current->kmap_ctrl.idx += KM_INCR;
468 	BUG_ON(current->kmap_ctrl.idx >= KM_MAX_IDX);
469 	return current->kmap_ctrl.idx - 1;
470 }
471 
472 static inline int kmap_local_idx(void)
473 {
474 	return current->kmap_ctrl.idx - 1;
475 }
476 
477 static inline void kmap_local_idx_pop(void)
478 {
479 	current->kmap_ctrl.idx -= KM_INCR;
480 	BUG_ON(current->kmap_ctrl.idx < 0);
481 }
482 
483 #ifndef arch_kmap_local_post_map
484 # define arch_kmap_local_post_map(vaddr, pteval)	do { } while (0)
485 #endif
486 
487 #ifndef arch_kmap_local_pre_unmap
488 # define arch_kmap_local_pre_unmap(vaddr)		do { } while (0)
489 #endif
490 
491 #ifndef arch_kmap_local_post_unmap
492 # define arch_kmap_local_post_unmap(vaddr)		do { } while (0)
493 #endif
494 
495 #ifndef arch_kmap_local_unmap_idx
496 #define arch_kmap_local_unmap_idx(idx, vaddr)	kmap_local_calc_idx(idx)
497 #endif
498 
499 #ifndef arch_kmap_local_high_get
500 static inline void *arch_kmap_local_high_get(struct page *page)
501 {
502 	return NULL;
503 }
504 #endif
505 
506 #ifndef arch_kmap_local_set_pte
507 #define arch_kmap_local_set_pte(mm, vaddr, ptep, ptev)	\
508 	set_pte_at(mm, vaddr, ptep, ptev)
509 #endif
510 
511 /* Unmap a local mapping which was obtained by kmap_high_get() */
512 static inline bool kmap_high_unmap_local(unsigned long vaddr)
513 {
514 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
515 	if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
516 		kunmap_high(pte_page(ptep_get(&pkmap_page_table[PKMAP_NR(vaddr)])));
517 		return true;
518 	}
519 #endif
520 	return false;
521 }
522 
523 static pte_t *__kmap_pte;
524 
525 static pte_t *kmap_get_pte(unsigned long vaddr, int idx)
526 {
527 	if (IS_ENABLED(CONFIG_KMAP_LOCAL_NON_LINEAR_PTE_ARRAY))
528 		/*
529 		 * Set by the arch if __kmap_pte[-idx] does not produce
530 		 * the correct entry.
531 		 */
532 		return virt_to_kpte(vaddr);
533 	if (!__kmap_pte)
534 		__kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
535 	return &__kmap_pte[-idx];
536 }
537 
538 void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
539 {
540 	pte_t pteval, *kmap_pte;
541 	unsigned long vaddr;
542 	int idx;
543 
544 	/*
545 	 * Disable migration so resulting virtual address is stable
546 	 * across preemption.
547 	 */
548 	migrate_disable();
549 	preempt_disable();
550 	idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
551 	vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
552 	kmap_pte = kmap_get_pte(vaddr, idx);
553 	BUG_ON(!pte_none(ptep_get(kmap_pte)));
554 	pteval = pfn_pte(pfn, prot);
555 	arch_kmap_local_set_pte(&init_mm, vaddr, kmap_pte, pteval);
556 	arch_kmap_local_post_map(vaddr, pteval);
557 	current->kmap_ctrl.pteval[kmap_local_idx()] = pteval;
558 	preempt_enable();
559 
560 	return (void *)vaddr;
561 }
562 EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
563 
564 void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
565 {
566 	void *kmap;
567 
568 	/*
569 	 * To broaden the usage of the actual kmap_local() machinery always map
570 	 * pages when debugging is enabled and the architecture has no problems
571 	 * with alias mappings.
572 	 */
573 	if (!IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP) && !PageHighMem(page))
574 		return page_address(page);
575 
576 	/* Try kmap_high_get() if architecture has it enabled */
577 	kmap = arch_kmap_local_high_get(page);
578 	if (kmap)
579 		return kmap;
580 
581 	return __kmap_local_pfn_prot(page_to_pfn(page), prot);
582 }
583 EXPORT_SYMBOL(__kmap_local_page_prot);
584 
585 void kunmap_local_indexed(const void *vaddr)
586 {
587 	unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
588 	pte_t *kmap_pte;
589 	int idx;
590 
591 	if (addr < __fix_to_virt(FIX_KMAP_END) ||
592 	    addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
593 		if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP)) {
594 			/* This _should_ never happen! See above. */
595 			WARN_ON_ONCE(1);
596 			return;
597 		}
598 		/*
599 		 * Handle mappings which were obtained by kmap_high_get()
600 		 * first as the virtual address of such mappings is below
601 		 * PAGE_OFFSET. Warn for all other addresses which are in
602 		 * the user space part of the virtual address space.
603 		 */
604 		if (!kmap_high_unmap_local(addr))
605 			WARN_ON_ONCE(addr < PAGE_OFFSET);
606 		return;
607 	}
608 
609 	preempt_disable();
610 	idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
611 	WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
612 
613 	kmap_pte = kmap_get_pte(addr, idx);
614 	arch_kmap_local_pre_unmap(addr);
615 	pte_clear(&init_mm, addr, kmap_pte);
616 	arch_kmap_local_post_unmap(addr);
617 	current->kmap_ctrl.pteval[kmap_local_idx()] = __pte(0);
618 	kmap_local_idx_pop();
619 	preempt_enable();
620 	migrate_enable();
621 }
622 EXPORT_SYMBOL(kunmap_local_indexed);
623 
624 /*
625  * Invoked before switch_to(). This is safe even when during or after
626  * clearing the maps an interrupt which needs a kmap_local happens because
627  * the task::kmap_ctrl.idx is not modified by the unmapping code so a
628  * nested kmap_local will use the next unused index and restore the index
629  * on unmap. The already cleared kmaps of the outgoing task are irrelevant
630  * because the interrupt context does not know about them. The same applies
631  * when scheduling back in for an interrupt which happens before the
632  * restore is complete.
633  */
634 void __kmap_local_sched_out(void)
635 {
636 	struct task_struct *tsk = current;
637 	pte_t *kmap_pte;
638 	int i;
639 
640 	/* Clear kmaps */
641 	for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
642 		pte_t pteval = tsk->kmap_ctrl.pteval[i];
643 		unsigned long addr;
644 		int idx;
645 
646 		/* With debug all even slots are unmapped and act as guard */
647 		if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
648 			WARN_ON_ONCE(pte_val(pteval) != 0);
649 			continue;
650 		}
651 		if (WARN_ON_ONCE(pte_none(pteval)))
652 			continue;
653 
654 		/*
655 		 * This is a horrible hack for XTENSA to calculate the
656 		 * coloured PTE index. Uses the PFN encoded into the pteval
657 		 * and the map index calculation because the actual mapped
658 		 * virtual address is not stored in task::kmap_ctrl.
659 		 * For any sane architecture this is optimized out.
660 		 */
661 		idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
662 
663 		addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
664 		kmap_pte = kmap_get_pte(addr, idx);
665 		arch_kmap_local_pre_unmap(addr);
666 		pte_clear(&init_mm, addr, kmap_pte);
667 		arch_kmap_local_post_unmap(addr);
668 	}
669 }
670 
671 void __kmap_local_sched_in(void)
672 {
673 	struct task_struct *tsk = current;
674 	pte_t *kmap_pte;
675 	int i;
676 
677 	/* Restore kmaps */
678 	for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
679 		pte_t pteval = tsk->kmap_ctrl.pteval[i];
680 		unsigned long addr;
681 		int idx;
682 
683 		/* With debug all even slots are unmapped and act as guard */
684 		if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
685 			WARN_ON_ONCE(pte_val(pteval) != 0);
686 			continue;
687 		}
688 		if (WARN_ON_ONCE(pte_none(pteval)))
689 			continue;
690 
691 		/* See comment in __kmap_local_sched_out() */
692 		idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
693 		addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
694 		kmap_pte = kmap_get_pte(addr, idx);
695 		set_pte_at(&init_mm, addr, kmap_pte, pteval);
696 		arch_kmap_local_post_map(addr, pteval);
697 	}
698 }
699 
700 void kmap_local_fork(struct task_struct *tsk)
701 {
702 	if (WARN_ON_ONCE(tsk->kmap_ctrl.idx))
703 		memset(&tsk->kmap_ctrl, 0, sizeof(tsk->kmap_ctrl));
704 }
705 
706 #endif
707 
708 #if defined(HASHED_PAGE_VIRTUAL)
709 
710 #define PA_HASH_ORDER	7
711 
712 /*
713  * Describes one page->virtual association
714  */
715 struct page_address_map {
716 	struct page *page;
717 	void *virtual;
718 	struct list_head list;
719 };
720 
721 static struct page_address_map page_address_maps[LAST_PKMAP];
722 
723 /*
724  * Hash table bucket
725  */
726 static struct page_address_slot {
727 	struct list_head lh;			/* List of page_address_maps */
728 	spinlock_t lock;			/* Protect this bucket's list */
729 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
730 
731 static struct page_address_slot *page_slot(const struct page *page)
732 {
733 	return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
734 }
735 
736 /**
737  * page_address - get the mapped virtual address of a page
738  * @page: &struct page to get the virtual address of
739  *
740  * Returns the page's virtual address.
741  */
742 void *page_address(const struct page *page)
743 {
744 	unsigned long flags;
745 	void *ret;
746 	struct page_address_slot *pas;
747 
748 	if (!PageHighMem(page))
749 		return lowmem_page_address(page);
750 
751 	pas = page_slot(page);
752 	ret = NULL;
753 	spin_lock_irqsave(&pas->lock, flags);
754 	if (!list_empty(&pas->lh)) {
755 		struct page_address_map *pam;
756 
757 		list_for_each_entry(pam, &pas->lh, list) {
758 			if (pam->page == page) {
759 				ret = pam->virtual;
760 				break;
761 			}
762 		}
763 	}
764 
765 	spin_unlock_irqrestore(&pas->lock, flags);
766 	return ret;
767 }
768 EXPORT_SYMBOL(page_address);
769 
770 /**
771  * set_page_address - set a page's virtual address
772  * @page: &struct page to set
773  * @virtual: virtual address to use
774  */
775 void set_page_address(struct page *page, void *virtual)
776 {
777 	unsigned long flags;
778 	struct page_address_slot *pas;
779 	struct page_address_map *pam;
780 
781 	BUG_ON(!PageHighMem(page));
782 
783 	pas = page_slot(page);
784 	if (virtual) {		/* Add */
785 		pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
786 		pam->page = page;
787 		pam->virtual = virtual;
788 
789 		spin_lock_irqsave(&pas->lock, flags);
790 		list_add_tail(&pam->list, &pas->lh);
791 		spin_unlock_irqrestore(&pas->lock, flags);
792 	} else {		/* Remove */
793 		spin_lock_irqsave(&pas->lock, flags);
794 		list_for_each_entry(pam, &pas->lh, list) {
795 			if (pam->page == page) {
796 				list_del(&pam->list);
797 				break;
798 			}
799 		}
800 		spin_unlock_irqrestore(&pas->lock, flags);
801 	}
802 
803 	return;
804 }
805 
806 void __init page_address_init(void)
807 {
808 	int i;
809 
810 	for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
811 		INIT_LIST_HEAD(&page_address_htable[i].lh);
812 		spin_lock_init(&page_address_htable[i].lock);
813 	}
814 }
815 
816 #endif	/* defined(HASHED_PAGE_VIRTUAL) */
817