Merge branch 'drm-next' of git://people.freedesktop.org/~airlied/linux
[linux-2.6-block.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 unsigned long totalhigh_pages __read_mostly;
48 EXPORT_SYMBOL(totalhigh_pages);
49
50
51 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
52
53 unsigned int nr_free_highpages (void)
54 {
55         pg_data_t *pgdat;
56         unsigned int pages = 0;
57
58         for_each_online_pgdat(pgdat) {
59                 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
60                         NR_FREE_PAGES);
61                 if (zone_movable_is_highmem())
62                         pages += zone_page_state(
63                                         &pgdat->node_zones[ZONE_MOVABLE],
64                                         NR_FREE_PAGES);
65         }
66
67         return pages;
68 }
69
70 static int pkmap_count[LAST_PKMAP];
71 static unsigned int last_pkmap_nr;
72 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
73
74 pte_t * pkmap_page_table;
75
76 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
77
78 /*
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
97 struct page *kmap_to_page(void *vaddr)
98 {
99         unsigned long addr = (unsigned long)vaddr;
100
101         if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
102                 int i = PKMAP_NR(addr);
103                 return pte_page(pkmap_page_table[i]);
104         }
105
106         return virt_to_page(addr);
107 }
108
109 static void flush_all_zero_pkmaps(void)
110 {
111         int i;
112         int need_flush = 0;
113
114         flush_cache_kmaps();
115
116         for (i = 0; i < LAST_PKMAP; i++) {
117                 struct page *page;
118
119                 /*
120                  * zero means we don't have anything to do,
121                  * >1 means that it is still in use. Only
122                  * a count of 1 means that it is free but
123                  * needs to be unmapped
124                  */
125                 if (pkmap_count[i] != 1)
126                         continue;
127                 pkmap_count[i] = 0;
128
129                 /* sanity check */
130                 BUG_ON(pte_none(pkmap_page_table[i]));
131
132                 /*
133                  * Don't need an atomic fetch-and-clear op here;
134                  * no-one has the page mapped, and cannot get at
135                  * its virtual address (and hence PTE) without first
136                  * getting the kmap_lock (which is held here).
137                  * So no dangers, even with speculative execution.
138                  */
139                 page = pte_page(pkmap_page_table[i]);
140                 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
141
142                 set_page_address(page, NULL);
143                 need_flush = 1;
144         }
145         if (need_flush)
146                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
147 }
148
149 /**
150  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
151  */
152 void kmap_flush_unused(void)
153 {
154         lock_kmap();
155         flush_all_zero_pkmaps();
156         unlock_kmap();
157 }
158
159 static inline unsigned long map_new_virtual(struct page *page)
160 {
161         unsigned long vaddr;
162         int count;
163
164 start:
165         count = LAST_PKMAP;
166         /* Find an empty entry */
167         for (;;) {
168                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
169                 if (!last_pkmap_nr) {
170                         flush_all_zero_pkmaps();
171                         count = LAST_PKMAP;
172                 }
173                 if (!pkmap_count[last_pkmap_nr])
174                         break;  /* Found a usable entry */
175                 if (--count)
176                         continue;
177
178                 /*
179                  * Sleep for somebody else to unmap their entries
180                  */
181                 {
182                         DECLARE_WAITQUEUE(wait, current);
183
184                         __set_current_state(TASK_UNINTERRUPTIBLE);
185                         add_wait_queue(&pkmap_map_wait, &wait);
186                         unlock_kmap();
187                         schedule();
188                         remove_wait_queue(&pkmap_map_wait, &wait);
189                         lock_kmap();
190
191                         /* Somebody else might have mapped it while we slept */
192                         if (page_address(page))
193                                 return (unsigned long)page_address(page);
194
195                         /* Re-start */
196                         goto start;
197                 }
198         }
199         vaddr = PKMAP_ADDR(last_pkmap_nr);
200         set_pte_at(&init_mm, vaddr,
201                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
202
203         pkmap_count[last_pkmap_nr] = 1;
204         set_page_address(page, (void *)vaddr);
205
206         return vaddr;
207 }
208
209 /**
210  * kmap_high - map a highmem page into memory
211  * @page: &struct page to map
212  *
213  * Returns the page's virtual memory address.
214  *
215  * We cannot call this from interrupts, as it may block.
216  */
217 void *kmap_high(struct page *page)
218 {
219         unsigned long vaddr;
220
221         /*
222          * For highmem pages, we can't trust "virtual" until
223          * after we have the lock.
224          */
225         lock_kmap();
226         vaddr = (unsigned long)page_address(page);
227         if (!vaddr)
228                 vaddr = map_new_virtual(page);
229         pkmap_count[PKMAP_NR(vaddr)]++;
230         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
231         unlock_kmap();
232         return (void*) vaddr;
233 }
234
235 EXPORT_SYMBOL(kmap_high);
236
237 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
238 /**
239  * kmap_high_get - pin a highmem page into memory
240  * @page: &struct page to pin
241  *
242  * Returns the page's current virtual memory address, or NULL if no mapping
243  * exists.  If and only if a non null address is returned then a
244  * matching call to kunmap_high() is necessary.
245  *
246  * This can be called from any context.
247  */
248 void *kmap_high_get(struct page *page)
249 {
250         unsigned long vaddr, flags;
251
252         lock_kmap_any(flags);
253         vaddr = (unsigned long)page_address(page);
254         if (vaddr) {
255                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
256                 pkmap_count[PKMAP_NR(vaddr)]++;
257         }
258         unlock_kmap_any(flags);
259         return (void*) vaddr;
260 }
261 #endif
262
263 /**
264  * kunmap_high - unmap a highmem page into memory
265  * @page: &struct page to unmap
266  *
267  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
268  * only from user context.
269  */
270 void kunmap_high(struct page *page)
271 {
272         unsigned long vaddr;
273         unsigned long nr;
274         unsigned long flags;
275         int need_wakeup;
276
277         lock_kmap_any(flags);
278         vaddr = (unsigned long)page_address(page);
279         BUG_ON(!vaddr);
280         nr = PKMAP_NR(vaddr);
281
282         /*
283          * A count must never go down to zero
284          * without a TLB flush!
285          */
286         need_wakeup = 0;
287         switch (--pkmap_count[nr]) {
288         case 0:
289                 BUG();
290         case 1:
291                 /*
292                  * Avoid an unnecessary wake_up() function call.
293                  * The common case is pkmap_count[] == 1, but
294                  * no waiters.
295                  * The tasks queued in the wait-queue are guarded
296                  * by both the lock in the wait-queue-head and by
297                  * the kmap_lock.  As the kmap_lock is held here,
298                  * no need for the wait-queue-head's lock.  Simply
299                  * test if the queue is empty.
300                  */
301                 need_wakeup = waitqueue_active(&pkmap_map_wait);
302         }
303         unlock_kmap_any(flags);
304
305         /* do wake-up, if needed, race-free outside of the spin lock */
306         if (need_wakeup)
307                 wake_up(&pkmap_map_wait);
308 }
309
310 EXPORT_SYMBOL(kunmap_high);
311 #endif
312
313 #if defined(HASHED_PAGE_VIRTUAL)
314
315 #define PA_HASH_ORDER   7
316
317 /*
318  * Describes one page->virtual association
319  */
320 struct page_address_map {
321         struct page *page;
322         void *virtual;
323         struct list_head list;
324 };
325
326 static struct page_address_map page_address_maps[LAST_PKMAP];
327
328 /*
329  * Hash table bucket
330  */
331 static struct page_address_slot {
332         struct list_head lh;                    /* List of page_address_maps */
333         spinlock_t lock;                        /* Protect this bucket's list */
334 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
335
336 static struct page_address_slot *page_slot(const struct page *page)
337 {
338         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
339 }
340
341 /**
342  * page_address - get the mapped virtual address of a page
343  * @page: &struct page to get the virtual address of
344  *
345  * Returns the page's virtual address.
346  */
347 void *page_address(const struct page *page)
348 {
349         unsigned long flags;
350         void *ret;
351         struct page_address_slot *pas;
352
353         if (!PageHighMem(page))
354                 return lowmem_page_address(page);
355
356         pas = page_slot(page);
357         ret = NULL;
358         spin_lock_irqsave(&pas->lock, flags);
359         if (!list_empty(&pas->lh)) {
360                 struct page_address_map *pam;
361
362                 list_for_each_entry(pam, &pas->lh, list) {
363                         if (pam->page == page) {
364                                 ret = pam->virtual;
365                                 goto done;
366                         }
367                 }
368         }
369 done:
370         spin_unlock_irqrestore(&pas->lock, flags);
371         return ret;
372 }
373
374 EXPORT_SYMBOL(page_address);
375
376 /**
377  * set_page_address - set a page's virtual address
378  * @page: &struct page to set
379  * @virtual: virtual address to use
380  */
381 void set_page_address(struct page *page, void *virtual)
382 {
383         unsigned long flags;
384         struct page_address_slot *pas;
385         struct page_address_map *pam;
386
387         BUG_ON(!PageHighMem(page));
388
389         pas = page_slot(page);
390         if (virtual) {          /* Add */
391                 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
392                 pam->page = page;
393                 pam->virtual = virtual;
394
395                 spin_lock_irqsave(&pas->lock, flags);
396                 list_add_tail(&pam->list, &pas->lh);
397                 spin_unlock_irqrestore(&pas->lock, flags);
398         } else {                /* Remove */
399                 spin_lock_irqsave(&pas->lock, flags);
400                 list_for_each_entry(pam, &pas->lh, list) {
401                         if (pam->page == page) {
402                                 list_del(&pam->list);
403                                 spin_unlock_irqrestore(&pas->lock, flags);
404                                 goto done;
405                         }
406                 }
407                 spin_unlock_irqrestore(&pas->lock, flags);
408         }
409 done:
410         return;
411 }
412
413 void __init page_address_init(void)
414 {
415         int i;
416
417         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
418                 INIT_LIST_HEAD(&page_address_htable[i].lh);
419                 spin_lock_init(&page_address_htable[i].lock);
420         }
421 }
422
423 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */