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/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
34static 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 */
64static 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 */
73static 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 */
86static 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 */
96static 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 */
106static 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
114unsigned long __nr_free_highpages(void)
115{
116 unsigned long pages = 0;
117 struct zone *zone;
118
119 for_each_populated_zone(zone) {
120 if (is_highmem(zone))
121 pages += zone_page_state(zone, NR_FREE_PAGES);
122 }
123
124 return pages;
125}
126
127unsigned long __totalhigh_pages(void)
128{
129 unsigned long pages = 0;
130 struct zone *zone;
131
132 for_each_populated_zone(zone) {
133 if (is_highmem(zone))
134 pages += zone_managed_pages(zone);
135 }
136
137 return pages;
138}
139EXPORT_SYMBOL(__totalhigh_pages);
140
141static int pkmap_count[LAST_PKMAP];
142static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
143
144pte_t *pkmap_page_table;
145
146/*
147 * Most architectures have no use for kmap_high_get(), so let's abstract
148 * the disabling of IRQ out of the locking in that case to save on a
149 * potential useless overhead.
150 */
151#ifdef ARCH_NEEDS_KMAP_HIGH_GET
152#define lock_kmap() spin_lock_irq(&kmap_lock)
153#define unlock_kmap() spin_unlock_irq(&kmap_lock)
154#define lock_kmap_any(flags) spin_lock_irqsave(&kmap_lock, flags)
155#define unlock_kmap_any(flags) spin_unlock_irqrestore(&kmap_lock, flags)
156#else
157#define lock_kmap() spin_lock(&kmap_lock)
158#define unlock_kmap() spin_unlock(&kmap_lock)
159#define lock_kmap_any(flags) \
160 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
161#define unlock_kmap_any(flags) \
162 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
163#endif
164
165struct page *__kmap_to_page(void *vaddr)
166{
167 unsigned long base = (unsigned long) vaddr & PAGE_MASK;
168 struct kmap_ctrl *kctrl = ¤t->kmap_ctrl;
169 unsigned long addr = (unsigned long)vaddr;
170 int i;
171
172 /* kmap() mappings */
173 if (WARN_ON_ONCE(addr >= PKMAP_ADDR(0) &&
174 addr < PKMAP_ADDR(LAST_PKMAP)))
175 return pte_page(ptep_get(&pkmap_page_table[PKMAP_NR(addr)]));
176
177 /* kmap_local_page() mappings */
178 if (WARN_ON_ONCE(base >= __fix_to_virt(FIX_KMAP_END) &&
179 base < __fix_to_virt(FIX_KMAP_BEGIN))) {
180 for (i = 0; i < kctrl->idx; i++) {
181 unsigned long base_addr;
182 int idx;
183
184 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
185 base_addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
186
187 if (base_addr == base)
188 return pte_page(kctrl->pteval[i]);
189 }
190 }
191
192 return virt_to_page(vaddr);
193}
194EXPORT_SYMBOL(__kmap_to_page);
195
196static void flush_all_zero_pkmaps(void)
197{
198 int i;
199 int need_flush = 0;
200
201 flush_cache_kmaps();
202
203 for (i = 0; i < LAST_PKMAP; i++) {
204 struct page *page;
205 pte_t ptent;
206
207 /*
208 * zero means we don't have anything to do,
209 * >1 means that it is still in use. Only
210 * a count of 1 means that it is free but
211 * needs to be unmapped
212 */
213 if (pkmap_count[i] != 1)
214 continue;
215 pkmap_count[i] = 0;
216
217 /* sanity check */
218 ptent = ptep_get(&pkmap_page_table[i]);
219 BUG_ON(pte_none(ptent));
220
221 /*
222 * Don't need an atomic fetch-and-clear op here;
223 * no-one has the page mapped, and cannot get at
224 * its virtual address (and hence PTE) without first
225 * getting the kmap_lock (which is held here).
226 * So no dangers, even with speculative execution.
227 */
228 page = pte_page(ptent);
229 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
230
231 set_page_address(page, NULL);
232 need_flush = 1;
233 }
234 if (need_flush)
235 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
236}
237
238void __kmap_flush_unused(void)
239{
240 lock_kmap();
241 flush_all_zero_pkmaps();
242 unlock_kmap();
243}
244
245static inline unsigned long map_new_virtual(struct page *page)
246{
247 unsigned long vaddr;
248 int count;
249 unsigned int last_pkmap_nr;
250 unsigned int color = get_pkmap_color(page);
251
252start:
253 count = get_pkmap_entries_count(color);
254 /* Find an empty entry */
255 for (;;) {
256 last_pkmap_nr = get_next_pkmap_nr(color);
257 if (no_more_pkmaps(last_pkmap_nr, color)) {
258 flush_all_zero_pkmaps();
259 count = get_pkmap_entries_count(color);
260 }
261 if (!pkmap_count[last_pkmap_nr])
262 break; /* Found a usable entry */
263 if (--count)
264 continue;
265
266 /*
267 * Sleep for somebody else to unmap their entries
268 */
269 {
270 DECLARE_WAITQUEUE(wait, current);
271 wait_queue_head_t *pkmap_map_wait =
272 get_pkmap_wait_queue_head(color);
273
274 __set_current_state(TASK_UNINTERRUPTIBLE);
275 add_wait_queue(pkmap_map_wait, &wait);
276 unlock_kmap();
277 schedule();
278 remove_wait_queue(pkmap_map_wait, &wait);
279 lock_kmap();
280
281 /* Somebody else might have mapped it while we slept */
282 if (page_address(page))
283 return (unsigned long)page_address(page);
284
285 /* Re-start */
286 goto start;
287 }
288 }
289 vaddr = PKMAP_ADDR(last_pkmap_nr);
290 set_pte_at(&init_mm, vaddr,
291 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
292
293 pkmap_count[last_pkmap_nr] = 1;
294 set_page_address(page, (void *)vaddr);
295
296 return vaddr;
297}
298
299/**
300 * kmap_high - map a highmem page into memory
301 * @page: &struct page to map
302 *
303 * Returns the page's virtual memory address.
304 *
305 * We cannot call this from interrupts, as it may block.
306 */
307void *kmap_high(struct page *page)
308{
309 unsigned long vaddr;
310
311 /*
312 * For highmem pages, we can't trust "virtual" until
313 * after we have the lock.
314 */
315 lock_kmap();
316 vaddr = (unsigned long)page_address(page);
317 if (!vaddr)
318 vaddr = map_new_virtual(page);
319 pkmap_count[PKMAP_NR(vaddr)]++;
320 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
321 unlock_kmap();
322 return (void *) vaddr;
323}
324EXPORT_SYMBOL(kmap_high);
325
326#ifdef ARCH_NEEDS_KMAP_HIGH_GET
327/**
328 * kmap_high_get - pin a highmem page into memory
329 * @page: &struct page to pin
330 *
331 * Returns the page's current virtual memory address, or NULL if no mapping
332 * exists. If and only if a non null address is returned then a
333 * matching call to kunmap_high() is necessary.
334 *
335 * This can be called from any context.
336 */
337void *kmap_high_get(struct page *page)
338{
339 unsigned long vaddr, flags;
340
341 lock_kmap_any(flags);
342 vaddr = (unsigned long)page_address(page);
343 if (vaddr) {
344 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
345 pkmap_count[PKMAP_NR(vaddr)]++;
346 }
347 unlock_kmap_any(flags);
348 return (void *) vaddr;
349}
350#endif
351
352/**
353 * kunmap_high - unmap a highmem page into memory
354 * @page: &struct page to unmap
355 *
356 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
357 * only from user context.
358 */
359void kunmap_high(struct page *page)
360{
361 unsigned long vaddr;
362 unsigned long nr;
363 unsigned long flags;
364 int need_wakeup;
365 unsigned int color = get_pkmap_color(page);
366 wait_queue_head_t *pkmap_map_wait;
367
368 lock_kmap_any(flags);
369 vaddr = (unsigned long)page_address(page);
370 BUG_ON(!vaddr);
371 nr = PKMAP_NR(vaddr);
372
373 /*
374 * A count must never go down to zero
375 * without a TLB flush!
376 */
377 need_wakeup = 0;
378 switch (--pkmap_count[nr]) {
379 case 0:
380 BUG();
381 case 1:
382 /*
383 * Avoid an unnecessary wake_up() function call.
384 * The common case is pkmap_count[] == 1, but
385 * no waiters.
386 * The tasks queued in the wait-queue are guarded
387 * by both the lock in the wait-queue-head and by
388 * the kmap_lock. As the kmap_lock is held here,
389 * no need for the wait-queue-head's lock. Simply
390 * test if the queue is empty.
391 */
392 pkmap_map_wait = get_pkmap_wait_queue_head(color);
393 need_wakeup = waitqueue_active(pkmap_map_wait);
394 }
395 unlock_kmap_any(flags);
396
397 /* do wake-up, if needed, race-free outside of the spin lock */
398 if (need_wakeup)
399 wake_up(pkmap_map_wait);
400}
401EXPORT_SYMBOL(kunmap_high);
402
403void zero_user_segments(struct page *page, unsigned start1, unsigned end1,
404 unsigned start2, unsigned end2)
405{
406 unsigned int i;
407
408 BUG_ON(end1 > page_size(page) || end2 > page_size(page));
409
410 if (start1 >= end1)
411 start1 = end1 = 0;
412 if (start2 >= end2)
413 start2 = end2 = 0;
414
415 for (i = 0; i < compound_nr(page); i++) {
416 void *kaddr = NULL;
417
418 if (start1 >= PAGE_SIZE) {
419 start1 -= PAGE_SIZE;
420 end1 -= PAGE_SIZE;
421 } else {
422 unsigned this_end = min_t(unsigned, end1, PAGE_SIZE);
423
424 if (end1 > start1) {
425 kaddr = kmap_local_page(page + i);
426 memset(kaddr + start1, 0, this_end - start1);
427 }
428 end1 -= this_end;
429 start1 = 0;
430 }
431
432 if (start2 >= PAGE_SIZE) {
433 start2 -= PAGE_SIZE;
434 end2 -= PAGE_SIZE;
435 } else {
436 unsigned this_end = min_t(unsigned, end2, PAGE_SIZE);
437
438 if (end2 > start2) {
439 if (!kaddr)
440 kaddr = kmap_local_page(page + i);
441 memset(kaddr + start2, 0, this_end - start2);
442 }
443 end2 -= this_end;
444 start2 = 0;
445 }
446
447 if (kaddr) {
448 kunmap_local(kaddr);
449 flush_dcache_page(page + i);
450 }
451
452 if (!end1 && !end2)
453 break;
454 }
455
456 BUG_ON((start1 | start2 | end1 | end2) != 0);
457}
458EXPORT_SYMBOL(zero_user_segments);
459#endif /* CONFIG_HIGHMEM */
460
461#ifdef CONFIG_KMAP_LOCAL
462
463#include <asm/kmap_size.h>
464
465/*
466 * With DEBUG_KMAP_LOCAL the stack depth is doubled and every second
467 * slot is unused which acts as a guard page
468 */
469#ifdef CONFIG_DEBUG_KMAP_LOCAL
470# define KM_INCR 2
471#else
472# define KM_INCR 1
473#endif
474
475static inline int kmap_local_idx_push(void)
476{
477 WARN_ON_ONCE(in_hardirq() && !irqs_disabled());
478 current->kmap_ctrl.idx += KM_INCR;
479 BUG_ON(current->kmap_ctrl.idx >= KM_MAX_IDX);
480 return current->kmap_ctrl.idx - 1;
481}
482
483static inline int kmap_local_idx(void)
484{
485 return current->kmap_ctrl.idx - 1;
486}
487
488static inline void kmap_local_idx_pop(void)
489{
490 current->kmap_ctrl.idx -= KM_INCR;
491 BUG_ON(current->kmap_ctrl.idx < 0);
492}
493
494#ifndef arch_kmap_local_post_map
495# define arch_kmap_local_post_map(vaddr, pteval) do { } while (0)
496#endif
497
498#ifndef arch_kmap_local_pre_unmap
499# define arch_kmap_local_pre_unmap(vaddr) do { } while (0)
500#endif
501
502#ifndef arch_kmap_local_post_unmap
503# define arch_kmap_local_post_unmap(vaddr) do { } while (0)
504#endif
505
506#ifndef arch_kmap_local_unmap_idx
507#define arch_kmap_local_unmap_idx(idx, vaddr) kmap_local_calc_idx(idx)
508#endif
509
510#ifndef arch_kmap_local_high_get
511static inline void *arch_kmap_local_high_get(struct page *page)
512{
513 return NULL;
514}
515#endif
516
517#ifndef arch_kmap_local_set_pte
518#define arch_kmap_local_set_pte(mm, vaddr, ptep, ptev) \
519 set_pte_at(mm, vaddr, ptep, ptev)
520#endif
521
522/* Unmap a local mapping which was obtained by kmap_high_get() */
523static inline bool kmap_high_unmap_local(unsigned long vaddr)
524{
525#ifdef ARCH_NEEDS_KMAP_HIGH_GET
526 if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
527 kunmap_high(pte_page(ptep_get(&pkmap_page_table[PKMAP_NR(vaddr)])));
528 return true;
529 }
530#endif
531 return false;
532}
533
534static pte_t *__kmap_pte;
535
536static pte_t *kmap_get_pte(unsigned long vaddr, int idx)
537{
538 if (IS_ENABLED(CONFIG_KMAP_LOCAL_NON_LINEAR_PTE_ARRAY))
539 /*
540 * Set by the arch if __kmap_pte[-idx] does not produce
541 * the correct entry.
542 */
543 return virt_to_kpte(vaddr);
544 if (!__kmap_pte)
545 __kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
546 return &__kmap_pte[-idx];
547}
548
549void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
550{
551 pte_t pteval, *kmap_pte;
552 unsigned long vaddr;
553 int idx;
554
555 /*
556 * Disable migration so resulting virtual address is stable
557 * across preemption.
558 */
559 migrate_disable();
560 preempt_disable();
561 idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
562 vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
563 kmap_pte = kmap_get_pte(vaddr, idx);
564 BUG_ON(!pte_none(ptep_get(kmap_pte)));
565 pteval = pfn_pte(pfn, prot);
566 arch_kmap_local_set_pte(&init_mm, vaddr, kmap_pte, pteval);
567 arch_kmap_local_post_map(vaddr, pteval);
568 current->kmap_ctrl.pteval[kmap_local_idx()] = pteval;
569 preempt_enable();
570
571 return (void *)vaddr;
572}
573EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
574
575void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
576{
577 void *kmap;
578
579 /*
580 * To broaden the usage of the actual kmap_local() machinery always map
581 * pages when debugging is enabled and the architecture has no problems
582 * with alias mappings.
583 */
584 if (!IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP) && !PageHighMem(page))
585 return page_address(page);
586
587 /* Try kmap_high_get() if architecture has it enabled */
588 kmap = arch_kmap_local_high_get(page);
589 if (kmap)
590 return kmap;
591
592 return __kmap_local_pfn_prot(page_to_pfn(page), prot);
593}
594EXPORT_SYMBOL(__kmap_local_page_prot);
595
596void kunmap_local_indexed(const void *vaddr)
597{
598 unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
599 pte_t *kmap_pte;
600 int idx;
601
602 if (addr < __fix_to_virt(FIX_KMAP_END) ||
603 addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
604 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP)) {
605 /* This _should_ never happen! See above. */
606 WARN_ON_ONCE(1);
607 return;
608 }
609 /*
610 * Handle mappings which were obtained by kmap_high_get()
611 * first as the virtual address of such mappings is below
612 * PAGE_OFFSET. Warn for all other addresses which are in
613 * the user space part of the virtual address space.
614 */
615 if (!kmap_high_unmap_local(addr))
616 WARN_ON_ONCE(addr < PAGE_OFFSET);
617 return;
618 }
619
620 preempt_disable();
621 idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
622 WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
623
624 kmap_pte = kmap_get_pte(addr, idx);
625 arch_kmap_local_pre_unmap(addr);
626 pte_clear(&init_mm, addr, kmap_pte);
627 arch_kmap_local_post_unmap(addr);
628 current->kmap_ctrl.pteval[kmap_local_idx()] = __pte(0);
629 kmap_local_idx_pop();
630 preempt_enable();
631 migrate_enable();
632}
633EXPORT_SYMBOL(kunmap_local_indexed);
634
635/*
636 * Invoked before switch_to(). This is safe even when during or after
637 * clearing the maps an interrupt which needs a kmap_local happens because
638 * the task::kmap_ctrl.idx is not modified by the unmapping code so a
639 * nested kmap_local will use the next unused index and restore the index
640 * on unmap. The already cleared kmaps of the outgoing task are irrelevant
641 * because the interrupt context does not know about them. The same applies
642 * when scheduling back in for an interrupt which happens before the
643 * restore is complete.
644 */
645void __kmap_local_sched_out(void)
646{
647 struct task_struct *tsk = current;
648 pte_t *kmap_pte;
649 int i;
650
651 /* Clear kmaps */
652 for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
653 pte_t pteval = tsk->kmap_ctrl.pteval[i];
654 unsigned long addr;
655 int idx;
656
657 /* With debug all even slots are unmapped and act as guard */
658 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
659 WARN_ON_ONCE(pte_val(pteval) != 0);
660 continue;
661 }
662 if (WARN_ON_ONCE(pte_none(pteval)))
663 continue;
664
665 /*
666 * This is a horrible hack for XTENSA to calculate the
667 * coloured PTE index. Uses the PFN encoded into the pteval
668 * and the map index calculation because the actual mapped
669 * virtual address is not stored in task::kmap_ctrl.
670 * For any sane architecture this is optimized out.
671 */
672 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
673
674 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
675 kmap_pte = kmap_get_pte(addr, idx);
676 arch_kmap_local_pre_unmap(addr);
677 pte_clear(&init_mm, addr, kmap_pte);
678 arch_kmap_local_post_unmap(addr);
679 }
680}
681
682void __kmap_local_sched_in(void)
683{
684 struct task_struct *tsk = current;
685 pte_t *kmap_pte;
686 int i;
687
688 /* Restore kmaps */
689 for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
690 pte_t pteval = tsk->kmap_ctrl.pteval[i];
691 unsigned long addr;
692 int idx;
693
694 /* With debug all even slots are unmapped and act as guard */
695 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
696 WARN_ON_ONCE(pte_val(pteval) != 0);
697 continue;
698 }
699 if (WARN_ON_ONCE(pte_none(pteval)))
700 continue;
701
702 /* See comment in __kmap_local_sched_out() */
703 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
704 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
705 kmap_pte = kmap_get_pte(addr, idx);
706 set_pte_at(&init_mm, addr, kmap_pte, pteval);
707 arch_kmap_local_post_map(addr, pteval);
708 }
709}
710
711void kmap_local_fork(struct task_struct *tsk)
712{
713 if (WARN_ON_ONCE(tsk->kmap_ctrl.idx))
714 memset(&tsk->kmap_ctrl, 0, sizeof(tsk->kmap_ctrl));
715}
716
717#endif
718
719#if defined(HASHED_PAGE_VIRTUAL)
720
721#define PA_HASH_ORDER 7
722
723/*
724 * Describes one page->virtual association
725 */
726struct page_address_map {
727 struct page *page;
728 void *virtual;
729 struct list_head list;
730};
731
732static struct page_address_map page_address_maps[LAST_PKMAP];
733
734/*
735 * Hash table bucket
736 */
737static struct page_address_slot {
738 struct list_head lh; /* List of page_address_maps */
739 spinlock_t lock; /* Protect this bucket's list */
740} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
741
742static struct page_address_slot *page_slot(const struct page *page)
743{
744 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
745}
746
747/**
748 * page_address - get the mapped virtual address of a page
749 * @page: &struct page to get the virtual address of
750 *
751 * Returns the page's virtual address.
752 */
753void *page_address(const struct page *page)
754{
755 unsigned long flags;
756 void *ret;
757 struct page_address_slot *pas;
758
759 if (!PageHighMem(page))
760 return lowmem_page_address(page);
761
762 pas = page_slot(page);
763 ret = NULL;
764 spin_lock_irqsave(&pas->lock, flags);
765 if (!list_empty(&pas->lh)) {
766 struct page_address_map *pam;
767
768 list_for_each_entry(pam, &pas->lh, list) {
769 if (pam->page == page) {
770 ret = pam->virtual;
771 break;
772 }
773 }
774 }
775
776 spin_unlock_irqrestore(&pas->lock, flags);
777 return ret;
778}
779EXPORT_SYMBOL(page_address);
780
781/**
782 * set_page_address - set a page's virtual address
783 * @page: &struct page to set
784 * @virtual: virtual address to use
785 */
786void set_page_address(struct page *page, void *virtual)
787{
788 unsigned long flags;
789 struct page_address_slot *pas;
790 struct page_address_map *pam;
791
792 BUG_ON(!PageHighMem(page));
793
794 pas = page_slot(page);
795 if (virtual) { /* Add */
796 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
797 pam->page = page;
798 pam->virtual = virtual;
799
800 spin_lock_irqsave(&pas->lock, flags);
801 list_add_tail(&pam->list, &pas->lh);
802 spin_unlock_irqrestore(&pas->lock, flags);
803 } else { /* Remove */
804 spin_lock_irqsave(&pas->lock, flags);
805 list_for_each_entry(pam, &pas->lh, list) {
806 if (pam->page == page) {
807 list_del(&pam->list);
808 break;
809 }
810 }
811 spin_unlock_irqrestore(&pas->lock, flags);
812 }
813}
814
815void __init page_address_init(void)
816{
817 int i;
818
819 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
820 INIT_LIST_HEAD(&page_address_htable[i].lh);
821 spin_lock_init(&page_address_htable[i].lock);
822 }
823}
824
825#endif /* defined(HASHED_PAGE_VIRTUAL) */