percpu: fix per_cpu_ptr_to_phys() handling of non-page-aligned addresses
[linux-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/module.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 static void flush_all_zero_pkmaps(void)
98 {
99         int i;
100         int need_flush = 0;
101
102         flush_cache_kmaps();
103
104         for (i = 0; i < LAST_PKMAP; i++) {
105                 struct page *page;
106
107                 /*
108                  * zero means we don't have anything to do,
109                  * >1 means that it is still in use. Only
110                  * a count of 1 means that it is free but
111                  * needs to be unmapped
112                  */
113                 if (pkmap_count[i] != 1)
114                         continue;
115                 pkmap_count[i] = 0;
116
117                 /* sanity check */
118                 BUG_ON(pte_none(pkmap_page_table[i]));
119
120                 /*
121                  * Don't need an atomic fetch-and-clear op here;
122                  * no-one has the page mapped, and cannot get at
123                  * its virtual address (and hence PTE) without first
124                  * getting the kmap_lock (which is held here).
125                  * So no dangers, even with speculative execution.
126                  */
127                 page = pte_page(pkmap_page_table[i]);
128                 pte_clear(&init_mm, (unsigned long)page_address(page),
129                           &pkmap_page_table[i]);
130
131                 set_page_address(page, NULL);
132                 need_flush = 1;
133         }
134         if (need_flush)
135                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
136 }
137
138 /**
139  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
140  */
141 void kmap_flush_unused(void)
142 {
143         lock_kmap();
144         flush_all_zero_pkmaps();
145         unlock_kmap();
146 }
147
148 static inline unsigned long map_new_virtual(struct page *page)
149 {
150         unsigned long vaddr;
151         int count;
152
153 start:
154         count = LAST_PKMAP;
155         /* Find an empty entry */
156         for (;;) {
157                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
158                 if (!last_pkmap_nr) {
159                         flush_all_zero_pkmaps();
160                         count = LAST_PKMAP;
161                 }
162                 if (!pkmap_count[last_pkmap_nr])
163                         break;  /* Found a usable entry */
164                 if (--count)
165                         continue;
166
167                 /*
168                  * Sleep for somebody else to unmap their entries
169                  */
170                 {
171                         DECLARE_WAITQUEUE(wait, current);
172
173                         __set_current_state(TASK_UNINTERRUPTIBLE);
174                         add_wait_queue(&pkmap_map_wait, &wait);
175                         unlock_kmap();
176                         schedule();
177                         remove_wait_queue(&pkmap_map_wait, &wait);
178                         lock_kmap();
179
180                         /* Somebody else might have mapped it while we slept */
181                         if (page_address(page))
182                                 return (unsigned long)page_address(page);
183
184                         /* Re-start */
185                         goto start;
186                 }
187         }
188         vaddr = PKMAP_ADDR(last_pkmap_nr);
189         set_pte_at(&init_mm, vaddr,
190                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
191
192         pkmap_count[last_pkmap_nr] = 1;
193         set_page_address(page, (void *)vaddr);
194
195         return vaddr;
196 }
197
198 /**
199  * kmap_high - map a highmem page into memory
200  * @page: &struct page to map
201  *
202  * Returns the page's virtual memory address.
203  *
204  * We cannot call this from interrupts, as it may block.
205  */
206 void *kmap_high(struct page *page)
207 {
208         unsigned long vaddr;
209
210         /*
211          * For highmem pages, we can't trust "virtual" until
212          * after we have the lock.
213          */
214         lock_kmap();
215         vaddr = (unsigned long)page_address(page);
216         if (!vaddr)
217                 vaddr = map_new_virtual(page);
218         pkmap_count[PKMAP_NR(vaddr)]++;
219         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
220         unlock_kmap();
221         return (void*) vaddr;
222 }
223
224 EXPORT_SYMBOL(kmap_high);
225
226 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
227 /**
228  * kmap_high_get - pin a highmem page into memory
229  * @page: &struct page to pin
230  *
231  * Returns the page's current virtual memory address, or NULL if no mapping
232  * exists.  If and only if a non null address is returned then a
233  * matching call to kunmap_high() is necessary.
234  *
235  * This can be called from any context.
236  */
237 void *kmap_high_get(struct page *page)
238 {
239         unsigned long vaddr, flags;
240
241         lock_kmap_any(flags);
242         vaddr = (unsigned long)page_address(page);
243         if (vaddr) {
244                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
245                 pkmap_count[PKMAP_NR(vaddr)]++;
246         }
247         unlock_kmap_any(flags);
248         return (void*) vaddr;
249 }
250 #endif
251
252 /**
253  * kunmap_high - map a highmem page into memory
254  * @page: &struct page to unmap
255  *
256  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
257  * only from user context.
258  */
259 void kunmap_high(struct page *page)
260 {
261         unsigned long vaddr;
262         unsigned long nr;
263         unsigned long flags;
264         int need_wakeup;
265
266         lock_kmap_any(flags);
267         vaddr = (unsigned long)page_address(page);
268         BUG_ON(!vaddr);
269         nr = PKMAP_NR(vaddr);
270
271         /*
272          * A count must never go down to zero
273          * without a TLB flush!
274          */
275         need_wakeup = 0;
276         switch (--pkmap_count[nr]) {
277         case 0:
278                 BUG();
279         case 1:
280                 /*
281                  * Avoid an unnecessary wake_up() function call.
282                  * The common case is pkmap_count[] == 1, but
283                  * no waiters.
284                  * The tasks queued in the wait-queue are guarded
285                  * by both the lock in the wait-queue-head and by
286                  * the kmap_lock.  As the kmap_lock is held here,
287                  * no need for the wait-queue-head's lock.  Simply
288                  * test if the queue is empty.
289                  */
290                 need_wakeup = waitqueue_active(&pkmap_map_wait);
291         }
292         unlock_kmap_any(flags);
293
294         /* do wake-up, if needed, race-free outside of the spin lock */
295         if (need_wakeup)
296                 wake_up(&pkmap_map_wait);
297 }
298
299 EXPORT_SYMBOL(kunmap_high);
300 #endif
301
302 #if defined(HASHED_PAGE_VIRTUAL)
303
304 #define PA_HASH_ORDER   7
305
306 /*
307  * Describes one page->virtual association
308  */
309 struct page_address_map {
310         struct page *page;
311         void *virtual;
312         struct list_head list;
313 };
314
315 /*
316  * page_address_map freelist, allocated from page_address_maps.
317  */
318 static struct list_head page_address_pool;      /* freelist */
319 static spinlock_t pool_lock;                    /* protects page_address_pool */
320
321 /*
322  * Hash table bucket
323  */
324 static struct page_address_slot {
325         struct list_head lh;                    /* List of page_address_maps */
326         spinlock_t lock;                        /* Protect this bucket's list */
327 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
328
329 static struct page_address_slot *page_slot(const struct page *page)
330 {
331         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
332 }
333
334 /**
335  * page_address - get the mapped virtual address of a page
336  * @page: &struct page to get the virtual address of
337  *
338  * Returns the page's virtual address.
339  */
340 void *page_address(const struct page *page)
341 {
342         unsigned long flags;
343         void *ret;
344         struct page_address_slot *pas;
345
346         if (!PageHighMem(page))
347                 return lowmem_page_address(page);
348
349         pas = page_slot(page);
350         ret = NULL;
351         spin_lock_irqsave(&pas->lock, flags);
352         if (!list_empty(&pas->lh)) {
353                 struct page_address_map *pam;
354
355                 list_for_each_entry(pam, &pas->lh, list) {
356                         if (pam->page == page) {
357                                 ret = pam->virtual;
358                                 goto done;
359                         }
360                 }
361         }
362 done:
363         spin_unlock_irqrestore(&pas->lock, flags);
364         return ret;
365 }
366
367 EXPORT_SYMBOL(page_address);
368
369 /**
370  * set_page_address - set a page's virtual address
371  * @page: &struct page to set
372  * @virtual: virtual address to use
373  */
374 void set_page_address(struct page *page, void *virtual)
375 {
376         unsigned long flags;
377         struct page_address_slot *pas;
378         struct page_address_map *pam;
379
380         BUG_ON(!PageHighMem(page));
381
382         pas = page_slot(page);
383         if (virtual) {          /* Add */
384                 BUG_ON(list_empty(&page_address_pool));
385
386                 spin_lock_irqsave(&pool_lock, flags);
387                 pam = list_entry(page_address_pool.next,
388                                 struct page_address_map, list);
389                 list_del(&pam->list);
390                 spin_unlock_irqrestore(&pool_lock, flags);
391
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                                 spin_lock_irqsave(&pool_lock, flags);
405                                 list_add_tail(&pam->list, &page_address_pool);
406                                 spin_unlock_irqrestore(&pool_lock, flags);
407                                 goto done;
408                         }
409                 }
410                 spin_unlock_irqrestore(&pas->lock, flags);
411         }
412 done:
413         return;
414 }
415
416 static struct page_address_map page_address_maps[LAST_PKMAP];
417
418 void __init page_address_init(void)
419 {
420         int i;
421
422         INIT_LIST_HEAD(&page_address_pool);
423         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
424                 list_add(&page_address_maps[i].list, &page_address_pool);
425         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
426                 INIT_LIST_HEAD(&page_address_htable[i].lh);
427                 spin_lock_init(&page_address_htable[i].lock);
428         }
429         spin_lock_init(&pool_lock);
430 }
431
432 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */