Merge branch 'stable-4.8' of git://git.infradead.org/users/pcmoore/audit
[sfrench/cifs-2.6.git] / mm / highmem.c
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)
34 DEFINE_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  */
57 static 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  */
66 static 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  */
79 static 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  */
89 static 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  */
99 static 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
107 unsigned long totalhigh_pages __read_mostly;
108 EXPORT_SYMBOL(totalhigh_pages);
109
110
111 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
112
113 unsigned int nr_free_highpages (void)
114 {
115         struct zone *zone;
116         unsigned int pages = 0;
117
118         for_each_populated_zone(zone) {
119                 if (is_highmem(zone))
120                         pages += zone_page_state(zone, NR_FREE_PAGES);
121         }
122
123         return pages;
124 }
125
126 static int pkmap_count[LAST_PKMAP];
127 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
128
129 pte_t * pkmap_page_table;
130
131 /*
132  * Most architectures have no use for kmap_high_get(), so let's abstract
133  * the disabling of IRQ out of the locking in that case to save on a
134  * potential useless overhead.
135  */
136 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
137 #define lock_kmap()             spin_lock_irq(&kmap_lock)
138 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
139 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
140 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
141 #else
142 #define lock_kmap()             spin_lock(&kmap_lock)
143 #define unlock_kmap()           spin_unlock(&kmap_lock)
144 #define lock_kmap_any(flags)    \
145                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
146 #define unlock_kmap_any(flags)  \
147                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
148 #endif
149
150 struct page *kmap_to_page(void *vaddr)
151 {
152         unsigned long addr = (unsigned long)vaddr;
153
154         if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
155                 int i = PKMAP_NR(addr);
156                 return pte_page(pkmap_page_table[i]);
157         }
158
159         return virt_to_page(addr);
160 }
161 EXPORT_SYMBOL(kmap_to_page);
162
163 static void flush_all_zero_pkmaps(void)
164 {
165         int i;
166         int need_flush = 0;
167
168         flush_cache_kmaps();
169
170         for (i = 0; i < LAST_PKMAP; i++) {
171                 struct page *page;
172
173                 /*
174                  * zero means we don't have anything to do,
175                  * >1 means that it is still in use. Only
176                  * a count of 1 means that it is free but
177                  * needs to be unmapped
178                  */
179                 if (pkmap_count[i] != 1)
180                         continue;
181                 pkmap_count[i] = 0;
182
183                 /* sanity check */
184                 BUG_ON(pte_none(pkmap_page_table[i]));
185
186                 /*
187                  * Don't need an atomic fetch-and-clear op here;
188                  * no-one has the page mapped, and cannot get at
189                  * its virtual address (and hence PTE) without first
190                  * getting the kmap_lock (which is held here).
191                  * So no dangers, even with speculative execution.
192                  */
193                 page = pte_page(pkmap_page_table[i]);
194                 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
195
196                 set_page_address(page, NULL);
197                 need_flush = 1;
198         }
199         if (need_flush)
200                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
201 }
202
203 /**
204  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
205  */
206 void kmap_flush_unused(void)
207 {
208         lock_kmap();
209         flush_all_zero_pkmaps();
210         unlock_kmap();
211 }
212
213 static inline unsigned long map_new_virtual(struct page *page)
214 {
215         unsigned long vaddr;
216         int count;
217         unsigned int last_pkmap_nr;
218         unsigned int color = get_pkmap_color(page);
219
220 start:
221         count = get_pkmap_entries_count(color);
222         /* Find an empty entry */
223         for (;;) {
224                 last_pkmap_nr = get_next_pkmap_nr(color);
225                 if (no_more_pkmaps(last_pkmap_nr, color)) {
226                         flush_all_zero_pkmaps();
227                         count = get_pkmap_entries_count(color);
228                 }
229                 if (!pkmap_count[last_pkmap_nr])
230                         break;  /* Found a usable entry */
231                 if (--count)
232                         continue;
233
234                 /*
235                  * Sleep for somebody else to unmap their entries
236                  */
237                 {
238                         DECLARE_WAITQUEUE(wait, current);
239                         wait_queue_head_t *pkmap_map_wait =
240                                 get_pkmap_wait_queue_head(color);
241
242                         __set_current_state(TASK_UNINTERRUPTIBLE);
243                         add_wait_queue(pkmap_map_wait, &wait);
244                         unlock_kmap();
245                         schedule();
246                         remove_wait_queue(pkmap_map_wait, &wait);
247                         lock_kmap();
248
249                         /* Somebody else might have mapped it while we slept */
250                         if (page_address(page))
251                                 return (unsigned long)page_address(page);
252
253                         /* Re-start */
254                         goto start;
255                 }
256         }
257         vaddr = PKMAP_ADDR(last_pkmap_nr);
258         set_pte_at(&init_mm, vaddr,
259                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
260
261         pkmap_count[last_pkmap_nr] = 1;
262         set_page_address(page, (void *)vaddr);
263
264         return vaddr;
265 }
266
267 /**
268  * kmap_high - map a highmem page into memory
269  * @page: &struct page to map
270  *
271  * Returns the page's virtual memory address.
272  *
273  * We cannot call this from interrupts, as it may block.
274  */
275 void *kmap_high(struct page *page)
276 {
277         unsigned long vaddr;
278
279         /*
280          * For highmem pages, we can't trust "virtual" until
281          * after we have the lock.
282          */
283         lock_kmap();
284         vaddr = (unsigned long)page_address(page);
285         if (!vaddr)
286                 vaddr = map_new_virtual(page);
287         pkmap_count[PKMAP_NR(vaddr)]++;
288         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
289         unlock_kmap();
290         return (void*) vaddr;
291 }
292
293 EXPORT_SYMBOL(kmap_high);
294
295 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
296 /**
297  * kmap_high_get - pin a highmem page into memory
298  * @page: &struct page to pin
299  *
300  * Returns the page's current virtual memory address, or NULL if no mapping
301  * exists.  If and only if a non null address is returned then a
302  * matching call to kunmap_high() is necessary.
303  *
304  * This can be called from any context.
305  */
306 void *kmap_high_get(struct page *page)
307 {
308         unsigned long vaddr, flags;
309
310         lock_kmap_any(flags);
311         vaddr = (unsigned long)page_address(page);
312         if (vaddr) {
313                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
314                 pkmap_count[PKMAP_NR(vaddr)]++;
315         }
316         unlock_kmap_any(flags);
317         return (void*) vaddr;
318 }
319 #endif
320
321 /**
322  * kunmap_high - unmap a highmem page into memory
323  * @page: &struct page to unmap
324  *
325  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
326  * only from user context.
327  */
328 void kunmap_high(struct page *page)
329 {
330         unsigned long vaddr;
331         unsigned long nr;
332         unsigned long flags;
333         int need_wakeup;
334         unsigned int color = get_pkmap_color(page);
335         wait_queue_head_t *pkmap_map_wait;
336
337         lock_kmap_any(flags);
338         vaddr = (unsigned long)page_address(page);
339         BUG_ON(!vaddr);
340         nr = PKMAP_NR(vaddr);
341
342         /*
343          * A count must never go down to zero
344          * without a TLB flush!
345          */
346         need_wakeup = 0;
347         switch (--pkmap_count[nr]) {
348         case 0:
349                 BUG();
350         case 1:
351                 /*
352                  * Avoid an unnecessary wake_up() function call.
353                  * The common case is pkmap_count[] == 1, but
354                  * no waiters.
355                  * The tasks queued in the wait-queue are guarded
356                  * by both the lock in the wait-queue-head and by
357                  * the kmap_lock.  As the kmap_lock is held here,
358                  * no need for the wait-queue-head's lock.  Simply
359                  * test if the queue is empty.
360                  */
361                 pkmap_map_wait = get_pkmap_wait_queue_head(color);
362                 need_wakeup = waitqueue_active(pkmap_map_wait);
363         }
364         unlock_kmap_any(flags);
365
366         /* do wake-up, if needed, race-free outside of the spin lock */
367         if (need_wakeup)
368                 wake_up(pkmap_map_wait);
369 }
370
371 EXPORT_SYMBOL(kunmap_high);
372 #endif
373
374 #if defined(HASHED_PAGE_VIRTUAL)
375
376 #define PA_HASH_ORDER   7
377
378 /*
379  * Describes one page->virtual association
380  */
381 struct page_address_map {
382         struct page *page;
383         void *virtual;
384         struct list_head list;
385 };
386
387 static struct page_address_map page_address_maps[LAST_PKMAP];
388
389 /*
390  * Hash table bucket
391  */
392 static struct page_address_slot {
393         struct list_head lh;                    /* List of page_address_maps */
394         spinlock_t lock;                        /* Protect this bucket's list */
395 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
396
397 static struct page_address_slot *page_slot(const struct page *page)
398 {
399         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
400 }
401
402 /**
403  * page_address - get the mapped virtual address of a page
404  * @page: &struct page to get the virtual address of
405  *
406  * Returns the page's virtual address.
407  */
408 void *page_address(const struct page *page)
409 {
410         unsigned long flags;
411         void *ret;
412         struct page_address_slot *pas;
413
414         if (!PageHighMem(page))
415                 return lowmem_page_address(page);
416
417         pas = page_slot(page);
418         ret = NULL;
419         spin_lock_irqsave(&pas->lock, flags);
420         if (!list_empty(&pas->lh)) {
421                 struct page_address_map *pam;
422
423                 list_for_each_entry(pam, &pas->lh, list) {
424                         if (pam->page == page) {
425                                 ret = pam->virtual;
426                                 goto done;
427                         }
428                 }
429         }
430 done:
431         spin_unlock_irqrestore(&pas->lock, flags);
432         return ret;
433 }
434
435 EXPORT_SYMBOL(page_address);
436
437 /**
438  * set_page_address - set a page's virtual address
439  * @page: &struct page to set
440  * @virtual: virtual address to use
441  */
442 void set_page_address(struct page *page, void *virtual)
443 {
444         unsigned long flags;
445         struct page_address_slot *pas;
446         struct page_address_map *pam;
447
448         BUG_ON(!PageHighMem(page));
449
450         pas = page_slot(page);
451         if (virtual) {          /* Add */
452                 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
453                 pam->page = page;
454                 pam->virtual = virtual;
455
456                 spin_lock_irqsave(&pas->lock, flags);
457                 list_add_tail(&pam->list, &pas->lh);
458                 spin_unlock_irqrestore(&pas->lock, flags);
459         } else {                /* Remove */
460                 spin_lock_irqsave(&pas->lock, flags);
461                 list_for_each_entry(pam, &pas->lh, list) {
462                         if (pam->page == page) {
463                                 list_del(&pam->list);
464                                 spin_unlock_irqrestore(&pas->lock, flags);
465                                 goto done;
466                         }
467                 }
468                 spin_unlock_irqrestore(&pas->lock, flags);
469         }
470 done:
471         return;
472 }
473
474 void __init page_address_init(void)
475 {
476         int i;
477
478         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
479                 INIT_LIST_HEAD(&page_address_htable[i].lh);
480                 spin_lock_init(&page_address_htable[i].lock);
481         }
482 }
483
484 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */