blob: b32b70cdaed6cba1ba79914228593441f4b312d3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040020#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#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>
Jason Wesseleac790052010-08-05 09:22:24 -050029#include <linux/kgdb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <asm/tlbflush.h>
31
Peter Zijlstraa8e23a22010-10-27 15:32:57 -070032
33#if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
34DEFINE_PER_CPU(int, __kmap_atomic_idx);
35#endif
36
Linus Torvalds1da177e2005-04-16 15:20:36 -070037/*
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
Al Viro260b2362005-10-21 03:22:44 -040046
Christoph Lameterc1f60a52006-09-25 23:31:11 -070047unsigned long totalhigh_pages __read_mostly;
David S. Millerdb7a94d2008-07-19 22:39:46 -070048EXPORT_SYMBOL(totalhigh_pages);
Christoph Lameterc1f60a52006-09-25 23:31:11 -070049
Peter Zijlstra3e4d3af2010-10-26 14:21:51 -070050
Peter Zijlstra3e4d3af2010-10-26 14:21:51 -070051EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
52
Christoph Lameterc1f60a52006-09-25 23:31:11 -070053unsigned int nr_free_highpages (void)
54{
55 pg_data_t *pgdat;
56 unsigned int pages = 0;
57
Mel Gorman2a1e2742007-07-17 04:03:12 -070058 for_each_online_pgdat(pgdat) {
Christoph Lameterd23ad422007-02-10 01:43:02 -080059 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
60 NR_FREE_PAGES);
Mel Gorman2a1e2742007-07-17 04:03:12 -070061 if (zone_movable_is_highmem())
62 pages += zone_page_state(
63 &pgdat->node_zones[ZONE_MOVABLE],
64 NR_FREE_PAGES);
65 }
Christoph Lameterc1f60a52006-09-25 23:31:11 -070066
67 return pages;
68}
69
Linus Torvalds1da177e2005-04-16 15:20:36 -070070static int pkmap_count[LAST_PKMAP];
71static unsigned int last_pkmap_nr;
72static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
73
74pte_t * pkmap_page_table;
75
76static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
77
Nicolas Pitre3297e762009-03-04 22:49:41 -050078/*
79 * Most architectures have no use for kmap_high_get(), so let's abstract
80 * the disabling of IRQ out of the locking in that case to save on a
81 * potential useless overhead.
82 */
83#ifdef ARCH_NEEDS_KMAP_HIGH_GET
84#define lock_kmap() spin_lock_irq(&kmap_lock)
85#define unlock_kmap() spin_unlock_irq(&kmap_lock)
86#define lock_kmap_any(flags) spin_lock_irqsave(&kmap_lock, flags)
87#define unlock_kmap_any(flags) spin_unlock_irqrestore(&kmap_lock, flags)
88#else
89#define lock_kmap() spin_lock(&kmap_lock)
90#define unlock_kmap() spin_unlock(&kmap_lock)
91#define lock_kmap_any(flags) \
92 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
93#define unlock_kmap_any(flags) \
94 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
95#endif
96
Mel Gorman5a178112012-07-31 16:45:02 -070097struct page *kmap_to_page(void *vaddr)
98{
99 unsigned long addr = (unsigned long)vaddr;
100
Will Deacon498c2282012-11-16 14:15:00 -0800101 if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
Joonsoo Kim4de22c02012-12-11 16:01:17 -0800102 int i = PKMAP_NR(addr);
Mel Gorman5a178112012-07-31 16:45:02 -0700103 return pte_page(pkmap_page_table[i]);
104 }
105
106 return virt_to_page(addr);
107}
Will Deaconf0263d22012-10-19 14:03:31 +0100108EXPORT_SYMBOL(kmap_to_page);
Mel Gorman5a178112012-07-31 16:45:02 -0700109
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110static void flush_all_zero_pkmaps(void)
111{
112 int i;
Nick Piggin5843d9a2008-08-01 03:15:21 +0200113 int need_flush = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114
115 flush_cache_kmaps();
116
117 for (i = 0; i < LAST_PKMAP; i++) {
118 struct page *page;
119
120 /*
121 * zero means we don't have anything to do,
122 * >1 means that it is still in use. Only
123 * a count of 1 means that it is free but
124 * needs to be unmapped
125 */
126 if (pkmap_count[i] != 1)
127 continue;
128 pkmap_count[i] = 0;
129
130 /* sanity check */
Eric Sesterhenn75babca2006-04-02 13:47:35 +0200131 BUG_ON(pte_none(pkmap_page_table[i]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
133 /*
134 * Don't need an atomic fetch-and-clear op here;
135 * no-one has the page mapped, and cannot get at
136 * its virtual address (and hence PTE) without first
137 * getting the kmap_lock (which is held here).
138 * So no dangers, even with speculative execution.
139 */
140 page = pte_page(pkmap_page_table[i]);
Joonsoo Kimeb2db432012-12-11 16:01:24 -0800141 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
143 set_page_address(page, NULL);
Nick Piggin5843d9a2008-08-01 03:15:21 +0200144 need_flush = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 }
Nick Piggin5843d9a2008-08-01 03:15:21 +0200146 if (need_flush)
147 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148}
149
Randy Dunlap77f60782008-03-19 17:00:42 -0700150/**
151 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
152 */
Jeremy Fitzhardingece6234b2007-05-02 19:27:15 +0200153void kmap_flush_unused(void)
154{
Nicolas Pitre3297e762009-03-04 22:49:41 -0500155 lock_kmap();
Jeremy Fitzhardingece6234b2007-05-02 19:27:15 +0200156 flush_all_zero_pkmaps();
Nicolas Pitre3297e762009-03-04 22:49:41 -0500157 unlock_kmap();
Jeremy Fitzhardingece6234b2007-05-02 19:27:15 +0200158}
159
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160static inline unsigned long map_new_virtual(struct page *page)
161{
162 unsigned long vaddr;
163 int count;
164
165start:
166 count = LAST_PKMAP;
167 /* Find an empty entry */
168 for (;;) {
169 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
170 if (!last_pkmap_nr) {
171 flush_all_zero_pkmaps();
172 count = LAST_PKMAP;
173 }
174 if (!pkmap_count[last_pkmap_nr])
175 break; /* Found a usable entry */
176 if (--count)
177 continue;
178
179 /*
180 * Sleep for somebody else to unmap their entries
181 */
182 {
183 DECLARE_WAITQUEUE(wait, current);
184
185 __set_current_state(TASK_UNINTERRUPTIBLE);
186 add_wait_queue(&pkmap_map_wait, &wait);
Nicolas Pitre3297e762009-03-04 22:49:41 -0500187 unlock_kmap();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 schedule();
189 remove_wait_queue(&pkmap_map_wait, &wait);
Nicolas Pitre3297e762009-03-04 22:49:41 -0500190 lock_kmap();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191
192 /* Somebody else might have mapped it while we slept */
193 if (page_address(page))
194 return (unsigned long)page_address(page);
195
196 /* Re-start */
197 goto start;
198 }
199 }
200 vaddr = PKMAP_ADDR(last_pkmap_nr);
201 set_pte_at(&init_mm, vaddr,
202 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
203
204 pkmap_count[last_pkmap_nr] = 1;
205 set_page_address(page, (void *)vaddr);
206
207 return vaddr;
208}
209
Randy Dunlap77f60782008-03-19 17:00:42 -0700210/**
211 * kmap_high - map a highmem page into memory
212 * @page: &struct page to map
213 *
214 * Returns the page's virtual memory address.
215 *
216 * We cannot call this from interrupts, as it may block.
217 */
Harvey Harrison920c7a52008-02-04 22:29:26 -0800218void *kmap_high(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219{
220 unsigned long vaddr;
221
222 /*
223 * For highmem pages, we can't trust "virtual" until
224 * after we have the lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 */
Nicolas Pitre3297e762009-03-04 22:49:41 -0500226 lock_kmap();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227 vaddr = (unsigned long)page_address(page);
228 if (!vaddr)
229 vaddr = map_new_virtual(page);
230 pkmap_count[PKMAP_NR(vaddr)]++;
Eric Sesterhenn75babca2006-04-02 13:47:35 +0200231 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
Nicolas Pitre3297e762009-03-04 22:49:41 -0500232 unlock_kmap();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 return (void*) vaddr;
234}
235
236EXPORT_SYMBOL(kmap_high);
237
Nicolas Pitre3297e762009-03-04 22:49:41 -0500238#ifdef ARCH_NEEDS_KMAP_HIGH_GET
239/**
240 * kmap_high_get - pin a highmem page into memory
241 * @page: &struct page to pin
242 *
243 * Returns the page's current virtual memory address, or NULL if no mapping
Uwe Kleine-König5e39df52010-01-25 21:38:09 +0100244 * exists. If and only if a non null address is returned then a
Nicolas Pitre3297e762009-03-04 22:49:41 -0500245 * matching call to kunmap_high() is necessary.
246 *
247 * This can be called from any context.
248 */
249void *kmap_high_get(struct page *page)
250{
251 unsigned long vaddr, flags;
252
253 lock_kmap_any(flags);
254 vaddr = (unsigned long)page_address(page);
255 if (vaddr) {
256 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
257 pkmap_count[PKMAP_NR(vaddr)]++;
258 }
259 unlock_kmap_any(flags);
260 return (void*) vaddr;
261}
262#endif
263
Randy Dunlap77f60782008-03-19 17:00:42 -0700264/**
Li Haifeng4e9dc5d2011-10-31 17:09:09 -0700265 * kunmap_high - unmap a highmem page into memory
Randy Dunlap77f60782008-03-19 17:00:42 -0700266 * @page: &struct page to unmap
Nicolas Pitre3297e762009-03-04 22:49:41 -0500267 *
268 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
269 * only from user context.
Randy Dunlap77f60782008-03-19 17:00:42 -0700270 */
Harvey Harrison920c7a52008-02-04 22:29:26 -0800271void kunmap_high(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272{
273 unsigned long vaddr;
274 unsigned long nr;
Nicolas Pitre3297e762009-03-04 22:49:41 -0500275 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 int need_wakeup;
277
Nicolas Pitre3297e762009-03-04 22:49:41 -0500278 lock_kmap_any(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 vaddr = (unsigned long)page_address(page);
Eric Sesterhenn75babca2006-04-02 13:47:35 +0200280 BUG_ON(!vaddr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 nr = PKMAP_NR(vaddr);
282
283 /*
284 * A count must never go down to zero
285 * without a TLB flush!
286 */
287 need_wakeup = 0;
288 switch (--pkmap_count[nr]) {
289 case 0:
290 BUG();
291 case 1:
292 /*
293 * Avoid an unnecessary wake_up() function call.
294 * The common case is pkmap_count[] == 1, but
295 * no waiters.
296 * The tasks queued in the wait-queue are guarded
297 * by both the lock in the wait-queue-head and by
298 * the kmap_lock. As the kmap_lock is held here,
299 * no need for the wait-queue-head's lock. Simply
300 * test if the queue is empty.
301 */
302 need_wakeup = waitqueue_active(&pkmap_map_wait);
303 }
Nicolas Pitre3297e762009-03-04 22:49:41 -0500304 unlock_kmap_any(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305
306 /* do wake-up, if needed, race-free outside of the spin lock */
307 if (need_wakeup)
308 wake_up(&pkmap_map_wait);
309}
310
311EXPORT_SYMBOL(kunmap_high);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312#endif
313
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314#if defined(HASHED_PAGE_VIRTUAL)
315
316#define PA_HASH_ORDER 7
317
318/*
319 * Describes one page->virtual association
320 */
321struct page_address_map {
322 struct page *page;
323 void *virtual;
324 struct list_head list;
325};
326
Joonsoo Kima354e2c2012-12-11 16:01:23 -0800327static struct page_address_map page_address_maps[LAST_PKMAP];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
329/*
330 * Hash table bucket
331 */
332static struct page_address_slot {
333 struct list_head lh; /* List of page_address_maps */
334 spinlock_t lock; /* Protect this bucket's list */
335} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
336
Ian Campbellf9918792011-08-17 13:45:09 +0100337static struct page_address_slot *page_slot(const struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338{
339 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
340}
341
Randy Dunlap77f60782008-03-19 17:00:42 -0700342/**
343 * page_address - get the mapped virtual address of a page
344 * @page: &struct page to get the virtual address of
345 *
346 * Returns the page's virtual address.
347 */
Ian Campbellf9918792011-08-17 13:45:09 +0100348void *page_address(const struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349{
350 unsigned long flags;
351 void *ret;
352 struct page_address_slot *pas;
353
354 if (!PageHighMem(page))
355 return lowmem_page_address(page);
356
357 pas = page_slot(page);
358 ret = NULL;
359 spin_lock_irqsave(&pas->lock, flags);
360 if (!list_empty(&pas->lh)) {
361 struct page_address_map *pam;
362
363 list_for_each_entry(pam, &pas->lh, list) {
364 if (pam->page == page) {
365 ret = pam->virtual;
366 goto done;
367 }
368 }
369 }
370done:
371 spin_unlock_irqrestore(&pas->lock, flags);
372 return ret;
373}
374
375EXPORT_SYMBOL(page_address);
376
Randy Dunlap77f60782008-03-19 17:00:42 -0700377/**
378 * set_page_address - set a page's virtual address
379 * @page: &struct page to set
380 * @virtual: virtual address to use
381 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382void set_page_address(struct page *page, void *virtual)
383{
384 unsigned long flags;
385 struct page_address_slot *pas;
386 struct page_address_map *pam;
387
388 BUG_ON(!PageHighMem(page));
389
390 pas = page_slot(page);
391 if (virtual) { /* Add */
Joonsoo Kima354e2c2012-12-11 16:01:23 -0800392 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 pam->page = page;
394 pam->virtual = virtual;
395
396 spin_lock_irqsave(&pas->lock, flags);
397 list_add_tail(&pam->list, &pas->lh);
398 spin_unlock_irqrestore(&pas->lock, flags);
399 } else { /* Remove */
400 spin_lock_irqsave(&pas->lock, flags);
401 list_for_each_entry(pam, &pas->lh, list) {
402 if (pam->page == page) {
403 list_del(&pam->list);
404 spin_unlock_irqrestore(&pas->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 goto done;
406 }
407 }
408 spin_unlock_irqrestore(&pas->lock, flags);
409 }
410done:
411 return;
412}
413
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414void __init page_address_init(void)
415{
416 int i;
417
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
419 INIT_LIST_HEAD(&page_address_htable[i].lh);
420 spin_lock_init(&page_address_htable[i].lock);
421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422}
423
424#endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */