Loading...
1/*
2 * High memory handling common code and variables.
3 *
4 * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
5 * Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
6 *
7 *
8 * Redesigned the x86 32-bit VM architecture to deal with
9 * 64-bit physical space. With current x86 CPUs this
10 * means up to 64 Gigabytes physical RAM.
11 *
12 * Rewrote high memory support to move the page cache into
13 * high memory. Implemented permanent (schedulable) kmaps
14 * based on Linus' idea.
15 *
16 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
17 */
18
19#include <linux/mm.h>
20#include <linux/export.h>
21#include <linux/swap.h>
22#include <linux/bio.h>
23#include <linux/pagemap.h>
24#include <linux/mempool.h>
25#include <linux/blkdev.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
32
33#if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
34DEFINE_PER_CPU(int, __kmap_atomic_idx);
35#endif
36
37/*
38 * Virtual_count is not a pure "count".
39 * 0 means that it is not mapped, and has not been mapped
40 * since a TLB flush - it is usable.
41 * 1 means that there are no users, but it has been mapped
42 * since the last TLB flush - so we can't use it.
43 * n means that there are (n-1) current users of it.
44 */
45#ifdef CONFIG_HIGHMEM
46
47/*
48 * Architecture with aliasing data cache may define the following family of
49 * helper functions in its asm/highmem.h to control cache color of virtual
50 * addresses where physical memory pages are mapped by kmap.
51 */
52#ifndef get_pkmap_color
53
54/*
55 * Determine color of virtual address where the page should be mapped.
56 */
57static inline unsigned int get_pkmap_color(struct page *page)
58{
59 return 0;
60}
61#define get_pkmap_color get_pkmap_color
62
63/*
64 * Get next index for mapping inside PKMAP region for page with given color.
65 */
66static inline unsigned int get_next_pkmap_nr(unsigned int color)
67{
68 static unsigned int last_pkmap_nr;
69
70 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
71 return last_pkmap_nr;
72}
73
74/*
75 * Determine if page index inside PKMAP region (pkmap_nr) of given color
76 * has wrapped around PKMAP region end. When this happens an attempt to
77 * flush all unused PKMAP slots is made.
78 */
79static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
80{
81 return pkmap_nr == 0;
82}
83
84/*
85 * Get the number of PKMAP entries of the given color. If no free slot is
86 * found after checking that many entries, kmap will sleep waiting for
87 * someone to call kunmap and free PKMAP slot.
88 */
89static inline int get_pkmap_entries_count(unsigned int color)
90{
91 return LAST_PKMAP;
92}
93
94/*
95 * Get head of a wait queue for PKMAP entries of the given color.
96 * Wait queues for different mapping colors should be independent to avoid
97 * unnecessary wakeups caused by freeing of slots of other colors.
98 */
99static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
100{
101 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
102
103 return &pkmap_map_wait;
104}
105#endif
106
107unsigned long totalhigh_pages __read_mostly;
108EXPORT_SYMBOL(totalhigh_pages);
109
110
111EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
112
113unsigned int nr_free_highpages (void)
114{
115 pg_data_t *pgdat;
116 unsigned int pages = 0;
117
118 for_each_online_pgdat(pgdat) {
119 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
120 NR_FREE_PAGES);
121 if (zone_movable_is_highmem())
122 pages += zone_page_state(
123 &pgdat->node_zones[ZONE_MOVABLE],
124 NR_FREE_PAGES);
125 }
126
127 return pages;
128}
129
130static int pkmap_count[LAST_PKMAP];
131static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
132
133pte_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
154struct page *kmap_to_page(void *vaddr)
155{
156 unsigned long addr = (unsigned long)vaddr;
157
158 if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
159 int i = PKMAP_NR(addr);
160 return pte_page(pkmap_page_table[i]);
161 }
162
163 return virt_to_page(addr);
164}
165EXPORT_SYMBOL(kmap_to_page);
166
167static void flush_all_zero_pkmaps(void)
168{
169 int i;
170 int need_flush = 0;
171
172 flush_cache_kmaps();
173
174 for (i = 0; i < LAST_PKMAP; i++) {
175 struct page *page;
176
177 /*
178 * zero means we don't have anything to do,
179 * >1 means that it is still in use. Only
180 * a count of 1 means that it is free but
181 * needs to be unmapped
182 */
183 if (pkmap_count[i] != 1)
184 continue;
185 pkmap_count[i] = 0;
186
187 /* sanity check */
188 BUG_ON(pte_none(pkmap_page_table[i]));
189
190 /*
191 * Don't need an atomic fetch-and-clear op here;
192 * no-one has the page mapped, and cannot get at
193 * its virtual address (and hence PTE) without first
194 * getting the kmap_lock (which is held here).
195 * So no dangers, even with speculative execution.
196 */
197 page = pte_page(pkmap_page_table[i]);
198 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
199
200 set_page_address(page, NULL);
201 need_flush = 1;
202 }
203 if (need_flush)
204 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
205}
206
207/**
208 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
209 */
210void kmap_flush_unused(void)
211{
212 lock_kmap();
213 flush_all_zero_pkmaps();
214 unlock_kmap();
215}
216
217static inline unsigned long map_new_virtual(struct page *page)
218{
219 unsigned long vaddr;
220 int count;
221 unsigned int last_pkmap_nr;
222 unsigned int color = get_pkmap_color(page);
223
224start:
225 count = get_pkmap_entries_count(color);
226 /* Find an empty entry */
227 for (;;) {
228 last_pkmap_nr = get_next_pkmap_nr(color);
229 if (no_more_pkmaps(last_pkmap_nr, color)) {
230 flush_all_zero_pkmaps();
231 count = get_pkmap_entries_count(color);
232 }
233 if (!pkmap_count[last_pkmap_nr])
234 break; /* Found a usable entry */
235 if (--count)
236 continue;
237
238 /*
239 * Sleep for somebody else to unmap their entries
240 */
241 {
242 DECLARE_WAITQUEUE(wait, current);
243 wait_queue_head_t *pkmap_map_wait =
244 get_pkmap_wait_queue_head(color);
245
246 __set_current_state(TASK_UNINTERRUPTIBLE);
247 add_wait_queue(pkmap_map_wait, &wait);
248 unlock_kmap();
249 schedule();
250 remove_wait_queue(pkmap_map_wait, &wait);
251 lock_kmap();
252
253 /* Somebody else might have mapped it while we slept */
254 if (page_address(page))
255 return (unsigned long)page_address(page);
256
257 /* Re-start */
258 goto start;
259 }
260 }
261 vaddr = PKMAP_ADDR(last_pkmap_nr);
262 set_pte_at(&init_mm, vaddr,
263 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
264
265 pkmap_count[last_pkmap_nr] = 1;
266 set_page_address(page, (void *)vaddr);
267
268 return vaddr;
269}
270
271/**
272 * kmap_high - map a highmem page into memory
273 * @page: &struct page to map
274 *
275 * Returns the page's virtual memory address.
276 *
277 * We cannot call this from interrupts, as it may block.
278 */
279void *kmap_high(struct page *page)
280{
281 unsigned long vaddr;
282
283 /*
284 * For highmem pages, we can't trust "virtual" until
285 * after we have the lock.
286 */
287 lock_kmap();
288 vaddr = (unsigned long)page_address(page);
289 if (!vaddr)
290 vaddr = map_new_virtual(page);
291 pkmap_count[PKMAP_NR(vaddr)]++;
292 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
293 unlock_kmap();
294 return (void*) vaddr;
295}
296
297EXPORT_SYMBOL(kmap_high);
298
299#ifdef ARCH_NEEDS_KMAP_HIGH_GET
300/**
301 * kmap_high_get - pin a highmem page into memory
302 * @page: &struct page to pin
303 *
304 * Returns the page's current virtual memory address, or NULL if no mapping
305 * exists. If and only if a non null address is returned then a
306 * matching call to kunmap_high() is necessary.
307 *
308 * This can be called from any context.
309 */
310void *kmap_high_get(struct page *page)
311{
312 unsigned long vaddr, flags;
313
314 lock_kmap_any(flags);
315 vaddr = (unsigned long)page_address(page);
316 if (vaddr) {
317 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
318 pkmap_count[PKMAP_NR(vaddr)]++;
319 }
320 unlock_kmap_any(flags);
321 return (void*) vaddr;
322}
323#endif
324
325/**
326 * kunmap_high - unmap a highmem page into memory
327 * @page: &struct page to unmap
328 *
329 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
330 * only from user context.
331 */
332void kunmap_high(struct page *page)
333{
334 unsigned long vaddr;
335 unsigned long nr;
336 unsigned long flags;
337 int need_wakeup;
338 unsigned int color = get_pkmap_color(page);
339 wait_queue_head_t *pkmap_map_wait;
340
341 lock_kmap_any(flags);
342 vaddr = (unsigned long)page_address(page);
343 BUG_ON(!vaddr);
344 nr = PKMAP_NR(vaddr);
345
346 /*
347 * A count must never go down to zero
348 * without a TLB flush!
349 */
350 need_wakeup = 0;
351 switch (--pkmap_count[nr]) {
352 case 0:
353 BUG();
354 case 1:
355 /*
356 * Avoid an unnecessary wake_up() function call.
357 * The common case is pkmap_count[] == 1, but
358 * no waiters.
359 * The tasks queued in the wait-queue are guarded
360 * by both the lock in the wait-queue-head and by
361 * the kmap_lock. As the kmap_lock is held here,
362 * no need for the wait-queue-head's lock. Simply
363 * test if the queue is empty.
364 */
365 pkmap_map_wait = get_pkmap_wait_queue_head(color);
366 need_wakeup = waitqueue_active(pkmap_map_wait);
367 }
368 unlock_kmap_any(flags);
369
370 /* do wake-up, if needed, race-free outside of the spin lock */
371 if (need_wakeup)
372 wake_up(pkmap_map_wait);
373}
374
375EXPORT_SYMBOL(kunmap_high);
376#endif
377
378#if defined(HASHED_PAGE_VIRTUAL)
379
380#define PA_HASH_ORDER 7
381
382/*
383 * Describes one page->virtual association
384 */
385struct page_address_map {
386 struct page *page;
387 void *virtual;
388 struct list_head list;
389};
390
391static struct page_address_map page_address_maps[LAST_PKMAP];
392
393/*
394 * Hash table bucket
395 */
396static struct page_address_slot {
397 struct list_head lh; /* List of page_address_maps */
398 spinlock_t lock; /* Protect this bucket's list */
399} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
400
401static struct page_address_slot *page_slot(const struct page *page)
402{
403 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
404}
405
406/**
407 * page_address - get the mapped virtual address of a page
408 * @page: &struct page to get the virtual address of
409 *
410 * Returns the page's virtual address.
411 */
412void *page_address(const struct page *page)
413{
414 unsigned long flags;
415 void *ret;
416 struct page_address_slot *pas;
417
418 if (!PageHighMem(page))
419 return lowmem_page_address(page);
420
421 pas = page_slot(page);
422 ret = NULL;
423 spin_lock_irqsave(&pas->lock, flags);
424 if (!list_empty(&pas->lh)) {
425 struct page_address_map *pam;
426
427 list_for_each_entry(pam, &pas->lh, list) {
428 if (pam->page == page) {
429 ret = pam->virtual;
430 goto done;
431 }
432 }
433 }
434done:
435 spin_unlock_irqrestore(&pas->lock, flags);
436 return ret;
437}
438
439EXPORT_SYMBOL(page_address);
440
441/**
442 * set_page_address - set a page's virtual address
443 * @page: &struct page to set
444 * @virtual: virtual address to use
445 */
446void set_page_address(struct page *page, void *virtual)
447{
448 unsigned long flags;
449 struct page_address_slot *pas;
450 struct page_address_map *pam;
451
452 BUG_ON(!PageHighMem(page));
453
454 pas = page_slot(page);
455 if (virtual) { /* Add */
456 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
457 pam->page = page;
458 pam->virtual = virtual;
459
460 spin_lock_irqsave(&pas->lock, flags);
461 list_add_tail(&pam->list, &pas->lh);
462 spin_unlock_irqrestore(&pas->lock, flags);
463 } else { /* Remove */
464 spin_lock_irqsave(&pas->lock, flags);
465 list_for_each_entry(pam, &pas->lh, list) {
466 if (pam->page == page) {
467 list_del(&pam->list);
468 spin_unlock_irqrestore(&pas->lock, flags);
469 goto done;
470 }
471 }
472 spin_unlock_irqrestore(&pas->lock, flags);
473 }
474done:
475 return;
476}
477
478void __init page_address_init(void)
479{
480 int i;
481
482 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
483 INIT_LIST_HEAD(&page_address_htable[i].lh);
484 spin_lock_init(&page_address_htable[i].lock);
485 }
486}
487
488#endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */
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 */
54static 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 */
63static 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 */
76static 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 */
86static 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 */
96static 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
104atomic_long_t _totalhigh_pages __read_mostly;
105EXPORT_SYMBOL(_totalhigh_pages);
106
107unsigned 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
120static int pkmap_count[LAST_PKMAP];
121static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
122
123pte_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
144struct 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
151 return pte_page(pkmap_page_table[i]);
152 }
153
154 return virt_to_page(addr);
155}
156EXPORT_SYMBOL(__kmap_to_page);
157
158static void flush_all_zero_pkmaps(void)
159{
160 int i;
161 int need_flush = 0;
162
163 flush_cache_kmaps();
164
165 for (i = 0; i < LAST_PKMAP; i++) {
166 struct page *page;
167
168 /*
169 * zero means we don't have anything to do,
170 * >1 means that it is still in use. Only
171 * a count of 1 means that it is free but
172 * needs to be unmapped
173 */
174 if (pkmap_count[i] != 1)
175 continue;
176 pkmap_count[i] = 0;
177
178 /* sanity check */
179 BUG_ON(pte_none(pkmap_page_table[i]));
180
181 /*
182 * Don't need an atomic fetch-and-clear op here;
183 * no-one has the page mapped, and cannot get at
184 * its virtual address (and hence PTE) without first
185 * getting the kmap_lock (which is held here).
186 * So no dangers, even with speculative execution.
187 */
188 page = pte_page(pkmap_page_table[i]);
189 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
190
191 set_page_address(page, NULL);
192 need_flush = 1;
193 }
194 if (need_flush)
195 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
196}
197
198void __kmap_flush_unused(void)
199{
200 lock_kmap();
201 flush_all_zero_pkmaps();
202 unlock_kmap();
203}
204
205static inline unsigned long map_new_virtual(struct page *page)
206{
207 unsigned long vaddr;
208 int count;
209 unsigned int last_pkmap_nr;
210 unsigned int color = get_pkmap_color(page);
211
212start:
213 count = get_pkmap_entries_count(color);
214 /* Find an empty entry */
215 for (;;) {
216 last_pkmap_nr = get_next_pkmap_nr(color);
217 if (no_more_pkmaps(last_pkmap_nr, color)) {
218 flush_all_zero_pkmaps();
219 count = get_pkmap_entries_count(color);
220 }
221 if (!pkmap_count[last_pkmap_nr])
222 break; /* Found a usable entry */
223 if (--count)
224 continue;
225
226 /*
227 * Sleep for somebody else to unmap their entries
228 */
229 {
230 DECLARE_WAITQUEUE(wait, current);
231 wait_queue_head_t *pkmap_map_wait =
232 get_pkmap_wait_queue_head(color);
233
234 __set_current_state(TASK_UNINTERRUPTIBLE);
235 add_wait_queue(pkmap_map_wait, &wait);
236 unlock_kmap();
237 schedule();
238 remove_wait_queue(pkmap_map_wait, &wait);
239 lock_kmap();
240
241 /* Somebody else might have mapped it while we slept */
242 if (page_address(page))
243 return (unsigned long)page_address(page);
244
245 /* Re-start */
246 goto start;
247 }
248 }
249 vaddr = PKMAP_ADDR(last_pkmap_nr);
250 set_pte_at(&init_mm, vaddr,
251 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
252
253 pkmap_count[last_pkmap_nr] = 1;
254 set_page_address(page, (void *)vaddr);
255
256 return vaddr;
257}
258
259/**
260 * kmap_high - map a highmem page into memory
261 * @page: &struct page to map
262 *
263 * Returns the page's virtual memory address.
264 *
265 * We cannot call this from interrupts, as it may block.
266 */
267void *kmap_high(struct page *page)
268{
269 unsigned long vaddr;
270
271 /*
272 * For highmem pages, we can't trust "virtual" until
273 * after we have the lock.
274 */
275 lock_kmap();
276 vaddr = (unsigned long)page_address(page);
277 if (!vaddr)
278 vaddr = map_new_virtual(page);
279 pkmap_count[PKMAP_NR(vaddr)]++;
280 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
281 unlock_kmap();
282 return (void *) vaddr;
283}
284EXPORT_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 */
297void *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 */
319void 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}
361EXPORT_SYMBOL(kunmap_high);
362
363#ifdef CONFIG_TRANSPARENT_HUGEPAGE
364void 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 if (start1 >= end1)
372 start1 = end1 = 0;
373 if (start2 >= end2)
374 start2 = end2 = 0;
375
376 for (i = 0; i < compound_nr(page); i++) {
377 void *kaddr = NULL;
378
379 if (start1 >= PAGE_SIZE) {
380 start1 -= PAGE_SIZE;
381 end1 -= PAGE_SIZE;
382 } else {
383 unsigned this_end = min_t(unsigned, end1, PAGE_SIZE);
384
385 if (end1 > start1) {
386 kaddr = kmap_atomic(page + i);
387 memset(kaddr + start1, 0, this_end - start1);
388 }
389 end1 -= this_end;
390 start1 = 0;
391 }
392
393 if (start2 >= PAGE_SIZE) {
394 start2 -= PAGE_SIZE;
395 end2 -= PAGE_SIZE;
396 } else {
397 unsigned this_end = min_t(unsigned, end2, PAGE_SIZE);
398
399 if (end2 > start2) {
400 if (!kaddr)
401 kaddr = kmap_atomic(page + i);
402 memset(kaddr + start2, 0, this_end - start2);
403 }
404 end2 -= this_end;
405 start2 = 0;
406 }
407
408 if (kaddr) {
409 kunmap_atomic(kaddr);
410 flush_dcache_page(page + i);
411 }
412
413 if (!end1 && !end2)
414 break;
415 }
416
417 BUG_ON((start1 | start2 | end1 | end2) != 0);
418}
419EXPORT_SYMBOL(zero_user_segments);
420#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
421#endif /* CONFIG_HIGHMEM */
422
423#ifdef CONFIG_KMAP_LOCAL
424
425#include <asm/kmap_size.h>
426
427/*
428 * With DEBUG_KMAP_LOCAL the stack depth is doubled and every second
429 * slot is unused which acts as a guard page
430 */
431#ifdef CONFIG_DEBUG_KMAP_LOCAL
432# define KM_INCR 2
433#else
434# define KM_INCR 1
435#endif
436
437static inline int kmap_local_idx_push(void)
438{
439 WARN_ON_ONCE(in_irq() && !irqs_disabled());
440 current->kmap_ctrl.idx += KM_INCR;
441 BUG_ON(current->kmap_ctrl.idx >= KM_MAX_IDX);
442 return current->kmap_ctrl.idx - 1;
443}
444
445static inline int kmap_local_idx(void)
446{
447 return current->kmap_ctrl.idx - 1;
448}
449
450static inline void kmap_local_idx_pop(void)
451{
452 current->kmap_ctrl.idx -= KM_INCR;
453 BUG_ON(current->kmap_ctrl.idx < 0);
454}
455
456#ifndef arch_kmap_local_post_map
457# define arch_kmap_local_post_map(vaddr, pteval) do { } while (0)
458#endif
459
460#ifndef arch_kmap_local_pre_unmap
461# define arch_kmap_local_pre_unmap(vaddr) do { } while (0)
462#endif
463
464#ifndef arch_kmap_local_post_unmap
465# define arch_kmap_local_post_unmap(vaddr) do { } while (0)
466#endif
467
468#ifndef arch_kmap_local_map_idx
469#define arch_kmap_local_map_idx(idx, pfn) kmap_local_calc_idx(idx)
470#endif
471
472#ifndef arch_kmap_local_unmap_idx
473#define arch_kmap_local_unmap_idx(idx, vaddr) kmap_local_calc_idx(idx)
474#endif
475
476#ifndef arch_kmap_local_high_get
477static inline void *arch_kmap_local_high_get(struct page *page)
478{
479 return NULL;
480}
481#endif
482
483#ifndef arch_kmap_local_set_pte
484#define arch_kmap_local_set_pte(mm, vaddr, ptep, ptev) \
485 set_pte_at(mm, vaddr, ptep, ptev)
486#endif
487
488/* Unmap a local mapping which was obtained by kmap_high_get() */
489static inline bool kmap_high_unmap_local(unsigned long vaddr)
490{
491#ifdef ARCH_NEEDS_KMAP_HIGH_GET
492 if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
493 kunmap_high(pte_page(pkmap_page_table[PKMAP_NR(vaddr)]));
494 return true;
495 }
496#endif
497 return false;
498}
499
500static inline int kmap_local_calc_idx(int idx)
501{
502 return idx + KM_MAX_IDX * smp_processor_id();
503}
504
505static pte_t *__kmap_pte;
506
507static pte_t *kmap_get_pte(void)
508{
509 if (!__kmap_pte)
510 __kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
511 return __kmap_pte;
512}
513
514void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
515{
516 pte_t pteval, *kmap_pte = kmap_get_pte();
517 unsigned long vaddr;
518 int idx;
519
520 /*
521 * Disable migration so resulting virtual address is stable
522 * across preemption.
523 */
524 migrate_disable();
525 preempt_disable();
526 idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
527 vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
528 BUG_ON(!pte_none(*(kmap_pte - idx)));
529 pteval = pfn_pte(pfn, prot);
530 arch_kmap_local_set_pte(&init_mm, vaddr, kmap_pte - idx, pteval);
531 arch_kmap_local_post_map(vaddr, pteval);
532 current->kmap_ctrl.pteval[kmap_local_idx()] = pteval;
533 preempt_enable();
534
535 return (void *)vaddr;
536}
537EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
538
539void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
540{
541 void *kmap;
542
543 /*
544 * To broaden the usage of the actual kmap_local() machinery always map
545 * pages when debugging is enabled and the architecture has no problems
546 * with alias mappings.
547 */
548 if (!IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP) && !PageHighMem(page))
549 return page_address(page);
550
551 /* Try kmap_high_get() if architecture has it enabled */
552 kmap = arch_kmap_local_high_get(page);
553 if (kmap)
554 return kmap;
555
556 return __kmap_local_pfn_prot(page_to_pfn(page), prot);
557}
558EXPORT_SYMBOL(__kmap_local_page_prot);
559
560void kunmap_local_indexed(void *vaddr)
561{
562 unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
563 pte_t *kmap_pte = kmap_get_pte();
564 int idx;
565
566 if (addr < __fix_to_virt(FIX_KMAP_END) ||
567 addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
568 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP)) {
569 /* This _should_ never happen! See above. */
570 WARN_ON_ONCE(1);
571 return;
572 }
573 /*
574 * Handle mappings which were obtained by kmap_high_get()
575 * first as the virtual address of such mappings is below
576 * PAGE_OFFSET. Warn for all other addresses which are in
577 * the user space part of the virtual address space.
578 */
579 if (!kmap_high_unmap_local(addr))
580 WARN_ON_ONCE(addr < PAGE_OFFSET);
581 return;
582 }
583
584 preempt_disable();
585 idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
586 WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
587
588 arch_kmap_local_pre_unmap(addr);
589 pte_clear(&init_mm, addr, kmap_pte - idx);
590 arch_kmap_local_post_unmap(addr);
591 current->kmap_ctrl.pteval[kmap_local_idx()] = __pte(0);
592 kmap_local_idx_pop();
593 preempt_enable();
594 migrate_enable();
595}
596EXPORT_SYMBOL(kunmap_local_indexed);
597
598/*
599 * Invoked before switch_to(). This is safe even when during or after
600 * clearing the maps an interrupt which needs a kmap_local happens because
601 * the task::kmap_ctrl.idx is not modified by the unmapping code so a
602 * nested kmap_local will use the next unused index and restore the index
603 * on unmap. The already cleared kmaps of the outgoing task are irrelevant
604 * because the interrupt context does not know about them. The same applies
605 * when scheduling back in for an interrupt which happens before the
606 * restore is complete.
607 */
608void __kmap_local_sched_out(void)
609{
610 struct task_struct *tsk = current;
611 pte_t *kmap_pte = kmap_get_pte();
612 int i;
613
614 /* Clear kmaps */
615 for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
616 pte_t pteval = tsk->kmap_ctrl.pteval[i];
617 unsigned long addr;
618 int idx;
619
620 /* With debug all even slots are unmapped and act as guard */
621 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
622 WARN_ON_ONCE(!pte_none(pteval));
623 continue;
624 }
625 if (WARN_ON_ONCE(pte_none(pteval)))
626 continue;
627
628 /*
629 * This is a horrible hack for XTENSA to calculate the
630 * coloured PTE index. Uses the PFN encoded into the pteval
631 * and the map index calculation because the actual mapped
632 * virtual address is not stored in task::kmap_ctrl.
633 * For any sane architecture this is optimized out.
634 */
635 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
636
637 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
638 arch_kmap_local_pre_unmap(addr);
639 pte_clear(&init_mm, addr, kmap_pte - idx);
640 arch_kmap_local_post_unmap(addr);
641 }
642}
643
644void __kmap_local_sched_in(void)
645{
646 struct task_struct *tsk = current;
647 pte_t *kmap_pte = kmap_get_pte();
648 int i;
649
650 /* Restore kmaps */
651 for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
652 pte_t pteval = tsk->kmap_ctrl.pteval[i];
653 unsigned long addr;
654 int idx;
655
656 /* With debug all even slots are unmapped and act as guard */
657 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
658 WARN_ON_ONCE(!pte_none(pteval));
659 continue;
660 }
661 if (WARN_ON_ONCE(pte_none(pteval)))
662 continue;
663
664 /* See comment in __kmap_local_sched_out() */
665 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
666 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
667 set_pte_at(&init_mm, addr, kmap_pte - idx, pteval);
668 arch_kmap_local_post_map(addr, pteval);
669 }
670}
671
672void kmap_local_fork(struct task_struct *tsk)
673{
674 if (WARN_ON_ONCE(tsk->kmap_ctrl.idx))
675 memset(&tsk->kmap_ctrl, 0, sizeof(tsk->kmap_ctrl));
676}
677
678#endif
679
680#if defined(HASHED_PAGE_VIRTUAL)
681
682#define PA_HASH_ORDER 7
683
684/*
685 * Describes one page->virtual association
686 */
687struct page_address_map {
688 struct page *page;
689 void *virtual;
690 struct list_head list;
691};
692
693static struct page_address_map page_address_maps[LAST_PKMAP];
694
695/*
696 * Hash table bucket
697 */
698static struct page_address_slot {
699 struct list_head lh; /* List of page_address_maps */
700 spinlock_t lock; /* Protect this bucket's list */
701} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
702
703static struct page_address_slot *page_slot(const struct page *page)
704{
705 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
706}
707
708/**
709 * page_address - get the mapped virtual address of a page
710 * @page: &struct page to get the virtual address of
711 *
712 * Returns the page's virtual address.
713 */
714void *page_address(const struct page *page)
715{
716 unsigned long flags;
717 void *ret;
718 struct page_address_slot *pas;
719
720 if (!PageHighMem(page))
721 return lowmem_page_address(page);
722
723 pas = page_slot(page);
724 ret = NULL;
725 spin_lock_irqsave(&pas->lock, flags);
726 if (!list_empty(&pas->lh)) {
727 struct page_address_map *pam;
728
729 list_for_each_entry(pam, &pas->lh, list) {
730 if (pam->page == page) {
731 ret = pam->virtual;
732 goto done;
733 }
734 }
735 }
736done:
737 spin_unlock_irqrestore(&pas->lock, flags);
738 return ret;
739}
740EXPORT_SYMBOL(page_address);
741
742/**
743 * set_page_address - set a page's virtual address
744 * @page: &struct page to set
745 * @virtual: virtual address to use
746 */
747void set_page_address(struct page *page, void *virtual)
748{
749 unsigned long flags;
750 struct page_address_slot *pas;
751 struct page_address_map *pam;
752
753 BUG_ON(!PageHighMem(page));
754
755 pas = page_slot(page);
756 if (virtual) { /* Add */
757 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
758 pam->page = page;
759 pam->virtual = virtual;
760
761 spin_lock_irqsave(&pas->lock, flags);
762 list_add_tail(&pam->list, &pas->lh);
763 spin_unlock_irqrestore(&pas->lock, flags);
764 } else { /* Remove */
765 spin_lock_irqsave(&pas->lock, flags);
766 list_for_each_entry(pam, &pas->lh, list) {
767 if (pam->page == page) {
768 list_del(&pam->list);
769 spin_unlock_irqrestore(&pas->lock, flags);
770 goto done;
771 }
772 }
773 spin_unlock_irqrestore(&pas->lock, flags);
774 }
775done:
776 return;
777}
778
779void __init page_address_init(void)
780{
781 int i;
782
783 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
784 INIT_LIST_HEAD(&page_address_htable[i].lh);
785 spin_lock_init(&page_address_htable[i].lock);
786 }
787}
788
789#endif /* defined(HASHED_PAGE_VIRTUAL) */