mm: stack based kmap_atomic()
[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  * Virtual_count is not a pure "count".
34  *  0 means that it is not mapped, and has not been mapped
35  *    since a TLB flush - it is usable.
36  *  1 means that there are no users, but it has been mapped
37  *    since the last TLB flush - so we can't use it.
38  *  n means that there are (n-1) current users of it.
39  */
40 #ifdef CONFIG_HIGHMEM
41
42 unsigned long totalhigh_pages __read_mostly;
43 EXPORT_SYMBOL(totalhigh_pages);
44
45
46 DEFINE_PER_CPU(int, __kmap_atomic_idx);
47 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
48
49 unsigned int nr_free_highpages (void)
50 {
51         pg_data_t *pgdat;
52         unsigned int pages = 0;
53
54         for_each_online_pgdat(pgdat) {
55                 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
56                         NR_FREE_PAGES);
57                 if (zone_movable_is_highmem())
58                         pages += zone_page_state(
59                                         &pgdat->node_zones[ZONE_MOVABLE],
60                                         NR_FREE_PAGES);
61         }
62
63         return pages;
64 }
65
66 static int pkmap_count[LAST_PKMAP];
67 static unsigned int last_pkmap_nr;
68 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
69
70 pte_t * pkmap_page_table;
71
72 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
73
74 /*
75  * Most architectures have no use for kmap_high_get(), so let's abstract
76  * the disabling of IRQ out of the locking in that case to save on a
77  * potential useless overhead.
78  */
79 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
80 #define lock_kmap()             spin_lock_irq(&kmap_lock)
81 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
82 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
83 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
84 #else
85 #define lock_kmap()             spin_lock(&kmap_lock)
86 #define unlock_kmap()           spin_unlock(&kmap_lock)
87 #define lock_kmap_any(flags)    \
88                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
89 #define unlock_kmap_any(flags)  \
90                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
91 #endif
92
93 static void flush_all_zero_pkmaps(void)
94 {
95         int i;
96         int need_flush = 0;
97
98         flush_cache_kmaps();
99
100         for (i = 0; i < LAST_PKMAP; i++) {
101                 struct page *page;
102
103                 /*
104                  * zero means we don't have anything to do,
105                  * >1 means that it is still in use. Only
106                  * a count of 1 means that it is free but
107                  * needs to be unmapped
108                  */
109                 if (pkmap_count[i] != 1)
110                         continue;
111                 pkmap_count[i] = 0;
112
113                 /* sanity check */
114                 BUG_ON(pte_none(pkmap_page_table[i]));
115
116                 /*
117                  * Don't need an atomic fetch-and-clear op here;
118                  * no-one has the page mapped, and cannot get at
119                  * its virtual address (and hence PTE) without first
120                  * getting the kmap_lock (which is held here).
121                  * So no dangers, even with speculative execution.
122                  */
123                 page = pte_page(pkmap_page_table[i]);
124                 pte_clear(&init_mm, (unsigned long)page_address(page),
125                           &pkmap_page_table[i]);
126
127                 set_page_address(page, NULL);
128                 need_flush = 1;
129         }
130         if (need_flush)
131                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
132 }
133
134 /**
135  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
136  */
137 void kmap_flush_unused(void)
138 {
139         lock_kmap();
140         flush_all_zero_pkmaps();
141         unlock_kmap();
142 }
143
144 static inline unsigned long map_new_virtual(struct page *page)
145 {
146         unsigned long vaddr;
147         int count;
148
149 start:
150         count = LAST_PKMAP;
151         /* Find an empty entry */
152         for (;;) {
153                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
154                 if (!last_pkmap_nr) {
155                         flush_all_zero_pkmaps();
156                         count = LAST_PKMAP;
157                 }
158                 if (!pkmap_count[last_pkmap_nr])
159                         break;  /* Found a usable entry */
160                 if (--count)
161                         continue;
162
163                 /*
164                  * Sleep for somebody else to unmap their entries
165                  */
166                 {
167                         DECLARE_WAITQUEUE(wait, current);
168
169                         __set_current_state(TASK_UNINTERRUPTIBLE);
170                         add_wait_queue(&pkmap_map_wait, &wait);
171                         unlock_kmap();
172                         schedule();
173                         remove_wait_queue(&pkmap_map_wait, &wait);
174                         lock_kmap();
175
176                         /* Somebody else might have mapped it while we slept */
177                         if (page_address(page))
178                                 return (unsigned long)page_address(page);
179
180                         /* Re-start */
181                         goto start;
182                 }
183         }
184         vaddr = PKMAP_ADDR(last_pkmap_nr);
185         set_pte_at(&init_mm, vaddr,
186                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
187
188         pkmap_count[last_pkmap_nr] = 1;
189         set_page_address(page, (void *)vaddr);
190
191         return vaddr;
192 }
193
194 /**
195  * kmap_high - map a highmem page into memory
196  * @page: &struct page to map
197  *
198  * Returns the page's virtual memory address.
199  *
200  * We cannot call this from interrupts, as it may block.
201  */
202 void *kmap_high(struct page *page)
203 {
204         unsigned long vaddr;
205
206         /*
207          * For highmem pages, we can't trust "virtual" until
208          * after we have the lock.
209          */
210         lock_kmap();
211         vaddr = (unsigned long)page_address(page);
212         if (!vaddr)
213                 vaddr = map_new_virtual(page);
214         pkmap_count[PKMAP_NR(vaddr)]++;
215         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
216         unlock_kmap();
217         return (void*) vaddr;
218 }
219
220 EXPORT_SYMBOL(kmap_high);
221
222 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
223 /**
224  * kmap_high_get - pin a highmem page into memory
225  * @page: &struct page to pin
226  *
227  * Returns the page's current virtual memory address, or NULL if no mapping
228  * exists.  If and only if a non null address is returned then a
229  * matching call to kunmap_high() is necessary.
230  *
231  * This can be called from any context.
232  */
233 void *kmap_high_get(struct page *page)
234 {
235         unsigned long vaddr, flags;
236
237         lock_kmap_any(flags);
238         vaddr = (unsigned long)page_address(page);
239         if (vaddr) {
240                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
241                 pkmap_count[PKMAP_NR(vaddr)]++;
242         }
243         unlock_kmap_any(flags);
244         return (void*) vaddr;
245 }
246 #endif
247
248 /**
249  * kunmap_high - map a highmem page into memory
250  * @page: &struct page to unmap
251  *
252  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
253  * only from user context.
254  */
255 void kunmap_high(struct page *page)
256 {
257         unsigned long vaddr;
258         unsigned long nr;
259         unsigned long flags;
260         int need_wakeup;
261
262         lock_kmap_any(flags);
263         vaddr = (unsigned long)page_address(page);
264         BUG_ON(!vaddr);
265         nr = PKMAP_NR(vaddr);
266
267         /*
268          * A count must never go down to zero
269          * without a TLB flush!
270          */
271         need_wakeup = 0;
272         switch (--pkmap_count[nr]) {
273         case 0:
274                 BUG();
275         case 1:
276                 /*
277                  * Avoid an unnecessary wake_up() function call.
278                  * The common case is pkmap_count[] == 1, but
279                  * no waiters.
280                  * The tasks queued in the wait-queue are guarded
281                  * by both the lock in the wait-queue-head and by
282                  * the kmap_lock.  As the kmap_lock is held here,
283                  * no need for the wait-queue-head's lock.  Simply
284                  * test if the queue is empty.
285                  */
286                 need_wakeup = waitqueue_active(&pkmap_map_wait);
287         }
288         unlock_kmap_any(flags);
289
290         /* do wake-up, if needed, race-free outside of the spin lock */
291         if (need_wakeup)
292                 wake_up(&pkmap_map_wait);
293 }
294
295 EXPORT_SYMBOL(kunmap_high);
296 #endif
297
298 #if defined(HASHED_PAGE_VIRTUAL)
299
300 #define PA_HASH_ORDER   7
301
302 /*
303  * Describes one page->virtual association
304  */
305 struct page_address_map {
306         struct page *page;
307         void *virtual;
308         struct list_head list;
309 };
310
311 /*
312  * page_address_map freelist, allocated from page_address_maps.
313  */
314 static struct list_head page_address_pool;      /* freelist */
315 static spinlock_t pool_lock;                    /* protects page_address_pool */
316
317 /*
318  * Hash table bucket
319  */
320 static struct page_address_slot {
321         struct list_head lh;                    /* List of page_address_maps */
322         spinlock_t lock;                        /* Protect this bucket's list */
323 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
324
325 static struct page_address_slot *page_slot(struct page *page)
326 {
327         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
328 }
329
330 /**
331  * page_address - get the mapped virtual address of a page
332  * @page: &struct page to get the virtual address of
333  *
334  * Returns the page's virtual address.
335  */
336 void *page_address(struct page *page)
337 {
338         unsigned long flags;
339         void *ret;
340         struct page_address_slot *pas;
341
342         if (!PageHighMem(page))
343                 return lowmem_page_address(page);
344
345         pas = page_slot(page);
346         ret = NULL;
347         spin_lock_irqsave(&pas->lock, flags);
348         if (!list_empty(&pas->lh)) {
349                 struct page_address_map *pam;
350
351                 list_for_each_entry(pam, &pas->lh, list) {
352                         if (pam->page == page) {
353                                 ret = pam->virtual;
354                                 goto done;
355                         }
356                 }
357         }
358 done:
359         spin_unlock_irqrestore(&pas->lock, flags);
360         return ret;
361 }
362
363 EXPORT_SYMBOL(page_address);
364
365 /**
366  * set_page_address - set a page's virtual address
367  * @page: &struct page to set
368  * @virtual: virtual address to use
369  */
370 void set_page_address(struct page *page, void *virtual)
371 {
372         unsigned long flags;
373         struct page_address_slot *pas;
374         struct page_address_map *pam;
375
376         BUG_ON(!PageHighMem(page));
377
378         pas = page_slot(page);
379         if (virtual) {          /* Add */
380                 BUG_ON(list_empty(&page_address_pool));
381
382                 spin_lock_irqsave(&pool_lock, flags);
383                 pam = list_entry(page_address_pool.next,
384                                 struct page_address_map, list);
385                 list_del(&pam->list);
386                 spin_unlock_irqrestore(&pool_lock, flags);
387
388                 pam->page = page;
389                 pam->virtual = virtual;
390
391                 spin_lock_irqsave(&pas->lock, flags);
392                 list_add_tail(&pam->list, &pas->lh);
393                 spin_unlock_irqrestore(&pas->lock, flags);
394         } else {                /* Remove */
395                 spin_lock_irqsave(&pas->lock, flags);
396                 list_for_each_entry(pam, &pas->lh, list) {
397                         if (pam->page == page) {
398                                 list_del(&pam->list);
399                                 spin_unlock_irqrestore(&pas->lock, flags);
400                                 spin_lock_irqsave(&pool_lock, flags);
401                                 list_add_tail(&pam->list, &page_address_pool);
402                                 spin_unlock_irqrestore(&pool_lock, flags);
403                                 goto done;
404                         }
405                 }
406                 spin_unlock_irqrestore(&pas->lock, flags);
407         }
408 done:
409         return;
410 }
411
412 static struct page_address_map page_address_maps[LAST_PKMAP];
413
414 void __init page_address_init(void)
415 {
416         int i;
417
418         INIT_LIST_HEAD(&page_address_pool);
419         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
420                 list_add(&page_address_maps[i].list, &page_address_pool);
421         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
422                 INIT_LIST_HEAD(&page_address_htable[i].lh);
423                 spin_lock_init(&page_address_htable[i].lock);
424         }
425         spin_lock_init(&pool_lock);
426 }
427
428 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */