mm: fix migratetype bug which slowed swapping
[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 <asm/tlbflush.h>
30
31 /*
32  * Virtual_count is not a pure "count".
33  *  0 means that it is not mapped, and has not been mapped
34  *    since a TLB flush - it is usable.
35  *  1 means that there are no users, but it has been mapped
36  *    since the last TLB flush - so we can't use it.
37  *  n means that there are (n-1) current users of it.
38  */
39 #ifdef CONFIG_HIGHMEM
40
41 unsigned long totalhigh_pages __read_mostly;
42 EXPORT_SYMBOL(totalhigh_pages);
43
44 unsigned int nr_free_highpages (void)
45 {
46         pg_data_t *pgdat;
47         unsigned int pages = 0;
48
49         for_each_online_pgdat(pgdat) {
50                 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
51                         NR_FREE_PAGES);
52                 if (zone_movable_is_highmem())
53                         pages += zone_page_state(
54                                         &pgdat->node_zones[ZONE_MOVABLE],
55                                         NR_FREE_PAGES);
56         }
57
58         return pages;
59 }
60
61 static int pkmap_count[LAST_PKMAP];
62 static unsigned int last_pkmap_nr;
63 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
64
65 pte_t * pkmap_page_table;
66
67 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
68
69 /*
70  * Most architectures have no use for kmap_high_get(), so let's abstract
71  * the disabling of IRQ out of the locking in that case to save on a
72  * potential useless overhead.
73  */
74 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
75 #define lock_kmap()             spin_lock_irq(&kmap_lock)
76 #define unlock_kmap()           spin_unlock_irq(&kmap_lock)
77 #define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
78 #define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
79 #else
80 #define lock_kmap()             spin_lock(&kmap_lock)
81 #define unlock_kmap()           spin_unlock(&kmap_lock)
82 #define lock_kmap_any(flags)    \
83                 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
84 #define unlock_kmap_any(flags)  \
85                 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
86 #endif
87
88 static void flush_all_zero_pkmaps(void)
89 {
90         int i;
91         int need_flush = 0;
92
93         flush_cache_kmaps();
94
95         for (i = 0; i < LAST_PKMAP; i++) {
96                 struct page *page;
97
98                 /*
99                  * zero means we don't have anything to do,
100                  * >1 means that it is still in use. Only
101                  * a count of 1 means that it is free but
102                  * needs to be unmapped
103                  */
104                 if (pkmap_count[i] != 1)
105                         continue;
106                 pkmap_count[i] = 0;
107
108                 /* sanity check */
109                 BUG_ON(pte_none(pkmap_page_table[i]));
110
111                 /*
112                  * Don't need an atomic fetch-and-clear op here;
113                  * no-one has the page mapped, and cannot get at
114                  * its virtual address (and hence PTE) without first
115                  * getting the kmap_lock (which is held here).
116                  * So no dangers, even with speculative execution.
117                  */
118                 page = pte_page(pkmap_page_table[i]);
119                 pte_clear(&init_mm, (unsigned long)page_address(page),
120                           &pkmap_page_table[i]);
121
122                 set_page_address(page, NULL);
123                 need_flush = 1;
124         }
125         if (need_flush)
126                 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
127 }
128
129 /**
130  * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
131  */
132 void kmap_flush_unused(void)
133 {
134         lock_kmap();
135         flush_all_zero_pkmaps();
136         unlock_kmap();
137 }
138
139 static inline unsigned long map_new_virtual(struct page *page)
140 {
141         unsigned long vaddr;
142         int count;
143
144 start:
145         count = LAST_PKMAP;
146         /* Find an empty entry */
147         for (;;) {
148                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
149                 if (!last_pkmap_nr) {
150                         flush_all_zero_pkmaps();
151                         count = LAST_PKMAP;
152                 }
153                 if (!pkmap_count[last_pkmap_nr])
154                         break;  /* Found a usable entry */
155                 if (--count)
156                         continue;
157
158                 /*
159                  * Sleep for somebody else to unmap their entries
160                  */
161                 {
162                         DECLARE_WAITQUEUE(wait, current);
163
164                         __set_current_state(TASK_UNINTERRUPTIBLE);
165                         add_wait_queue(&pkmap_map_wait, &wait);
166                         unlock_kmap();
167                         schedule();
168                         remove_wait_queue(&pkmap_map_wait, &wait);
169                         lock_kmap();
170
171                         /* Somebody else might have mapped it while we slept */
172                         if (page_address(page))
173                                 return (unsigned long)page_address(page);
174
175                         /* Re-start */
176                         goto start;
177                 }
178         }
179         vaddr = PKMAP_ADDR(last_pkmap_nr);
180         set_pte_at(&init_mm, vaddr,
181                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
182
183         pkmap_count[last_pkmap_nr] = 1;
184         set_page_address(page, (void *)vaddr);
185
186         return vaddr;
187 }
188
189 /**
190  * kmap_high - map a highmem page into memory
191  * @page: &struct page to map
192  *
193  * Returns the page's virtual memory address.
194  *
195  * We cannot call this from interrupts, as it may block.
196  */
197 void *kmap_high(struct page *page)
198 {
199         unsigned long vaddr;
200
201         /*
202          * For highmem pages, we can't trust "virtual" until
203          * after we have the lock.
204          */
205         lock_kmap();
206         vaddr = (unsigned long)page_address(page);
207         if (!vaddr)
208                 vaddr = map_new_virtual(page);
209         pkmap_count[PKMAP_NR(vaddr)]++;
210         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
211         unlock_kmap();
212         return (void*) vaddr;
213 }
214
215 EXPORT_SYMBOL(kmap_high);
216
217 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
218 /**
219  * kmap_high_get - pin a highmem page into memory
220  * @page: &struct page to pin
221  *
222  * Returns the page's current virtual memory address, or NULL if no mapping
223  * exists.  When and only when a non null address is returned then a
224  * matching call to kunmap_high() is necessary.
225  *
226  * This can be called from any context.
227  */
228 void *kmap_high_get(struct page *page)
229 {
230         unsigned long vaddr, flags;
231
232         lock_kmap_any(flags);
233         vaddr = (unsigned long)page_address(page);
234         if (vaddr) {
235                 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
236                 pkmap_count[PKMAP_NR(vaddr)]++;
237         }
238         unlock_kmap_any(flags);
239         return (void*) vaddr;
240 }
241 #endif
242
243 /**
244  * kunmap_high - map a highmem page into memory
245  * @page: &struct page to unmap
246  *
247  * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
248  * only from user context.
249  */
250 void kunmap_high(struct page *page)
251 {
252         unsigned long vaddr;
253         unsigned long nr;
254         unsigned long flags;
255         int need_wakeup;
256
257         lock_kmap_any(flags);
258         vaddr = (unsigned long)page_address(page);
259         BUG_ON(!vaddr);
260         nr = PKMAP_NR(vaddr);
261
262         /*
263          * A count must never go down to zero
264          * without a TLB flush!
265          */
266         need_wakeup = 0;
267         switch (--pkmap_count[nr]) {
268         case 0:
269                 BUG();
270         case 1:
271                 /*
272                  * Avoid an unnecessary wake_up() function call.
273                  * The common case is pkmap_count[] == 1, but
274                  * no waiters.
275                  * The tasks queued in the wait-queue are guarded
276                  * by both the lock in the wait-queue-head and by
277                  * the kmap_lock.  As the kmap_lock is held here,
278                  * no need for the wait-queue-head's lock.  Simply
279                  * test if the queue is empty.
280                  */
281                 need_wakeup = waitqueue_active(&pkmap_map_wait);
282         }
283         unlock_kmap_any(flags);
284
285         /* do wake-up, if needed, race-free outside of the spin lock */
286         if (need_wakeup)
287                 wake_up(&pkmap_map_wait);
288 }
289
290 EXPORT_SYMBOL(kunmap_high);
291 #endif
292
293 #if defined(HASHED_PAGE_VIRTUAL)
294
295 #define PA_HASH_ORDER   7
296
297 /*
298  * Describes one page->virtual association
299  */
300 struct page_address_map {
301         struct page *page;
302         void *virtual;
303         struct list_head list;
304 };
305
306 /*
307  * page_address_map freelist, allocated from page_address_maps.
308  */
309 static struct list_head page_address_pool;      /* freelist */
310 static spinlock_t pool_lock;                    /* protects page_address_pool */
311
312 /*
313  * Hash table bucket
314  */
315 static struct page_address_slot {
316         struct list_head lh;                    /* List of page_address_maps */
317         spinlock_t lock;                        /* Protect this bucket's list */
318 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
319
320 static struct page_address_slot *page_slot(struct page *page)
321 {
322         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
323 }
324
325 /**
326  * page_address - get the mapped virtual address of a page
327  * @page: &struct page to get the virtual address of
328  *
329  * Returns the page's virtual address.
330  */
331 void *page_address(struct page *page)
332 {
333         unsigned long flags;
334         void *ret;
335         struct page_address_slot *pas;
336
337         if (!PageHighMem(page))
338                 return lowmem_page_address(page);
339
340         pas = page_slot(page);
341         ret = NULL;
342         spin_lock_irqsave(&pas->lock, flags);
343         if (!list_empty(&pas->lh)) {
344                 struct page_address_map *pam;
345
346                 list_for_each_entry(pam, &pas->lh, list) {
347                         if (pam->page == page) {
348                                 ret = pam->virtual;
349                                 goto done;
350                         }
351                 }
352         }
353 done:
354         spin_unlock_irqrestore(&pas->lock, flags);
355         return ret;
356 }
357
358 EXPORT_SYMBOL(page_address);
359
360 /**
361  * set_page_address - set a page's virtual address
362  * @page: &struct page to set
363  * @virtual: virtual address to use
364  */
365 void set_page_address(struct page *page, void *virtual)
366 {
367         unsigned long flags;
368         struct page_address_slot *pas;
369         struct page_address_map *pam;
370
371         BUG_ON(!PageHighMem(page));
372
373         pas = page_slot(page);
374         if (virtual) {          /* Add */
375                 BUG_ON(list_empty(&page_address_pool));
376
377                 spin_lock_irqsave(&pool_lock, flags);
378                 pam = list_entry(page_address_pool.next,
379                                 struct page_address_map, list);
380                 list_del(&pam->list);
381                 spin_unlock_irqrestore(&pool_lock, flags);
382
383                 pam->page = page;
384                 pam->virtual = virtual;
385
386                 spin_lock_irqsave(&pas->lock, flags);
387                 list_add_tail(&pam->list, &pas->lh);
388                 spin_unlock_irqrestore(&pas->lock, flags);
389         } else {                /* Remove */
390                 spin_lock_irqsave(&pas->lock, flags);
391                 list_for_each_entry(pam, &pas->lh, list) {
392                         if (pam->page == page) {
393                                 list_del(&pam->list);
394                                 spin_unlock_irqrestore(&pas->lock, flags);
395                                 spin_lock_irqsave(&pool_lock, flags);
396                                 list_add_tail(&pam->list, &page_address_pool);
397                                 spin_unlock_irqrestore(&pool_lock, flags);
398                                 goto done;
399                         }
400                 }
401                 spin_unlock_irqrestore(&pas->lock, flags);
402         }
403 done:
404         return;
405 }
406
407 static struct page_address_map page_address_maps[LAST_PKMAP];
408
409 void __init page_address_init(void)
410 {
411         int i;
412
413         INIT_LIST_HEAD(&page_address_pool);
414         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
415                 list_add(&page_address_maps[i].list, &page_address_pool);
416         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
417                 INIT_LIST_HEAD(&page_address_htable[i].lh);
418                 spin_lock_init(&page_address_htable[i].lock);
419         }
420         spin_lock_init(&pool_lock);
421 }
422
423 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */
424
425 #if defined(CONFIG_DEBUG_HIGHMEM) && defined(CONFIG_TRACE_IRQFLAGS_SUPPORT)
426
427 void debug_kmap_atomic(enum km_type type)
428 {
429         static int warn_count = 10;
430
431         if (unlikely(warn_count < 0))
432                 return;
433
434         if (unlikely(in_interrupt())) {
435                 if (in_nmi()) {
436                         if (type != KM_NMI && type != KM_NMI_PTE) {
437                                 WARN_ON(1);
438                                 warn_count--;
439                         }
440                 } else if (in_irq()) {
441                         if (type != KM_IRQ0 && type != KM_IRQ1 &&
442                             type != KM_BIO_SRC_IRQ && type != KM_BIO_DST_IRQ &&
443                             type != KM_BOUNCE_READ && type != KM_IRQ_PTE) {
444                                 WARN_ON(1);
445                                 warn_count--;
446                         }
447                 } else if (!irqs_disabled()) {  /* softirq */
448                         if (type != KM_IRQ0 && type != KM_IRQ1 &&
449                             type != KM_SOFTIRQ0 && type != KM_SOFTIRQ1 &&
450                             type != KM_SKB_SUNRPC_DATA &&
451                             type != KM_SKB_DATA_SOFTIRQ &&
452                             type != KM_BOUNCE_READ) {
453                                 WARN_ON(1);
454                                 warn_count--;
455                         }
456                 }
457         }
458
459         if (type == KM_IRQ0 || type == KM_IRQ1 || type == KM_BOUNCE_READ ||
460                         type == KM_BIO_SRC_IRQ || type == KM_BIO_DST_IRQ ||
461                         type == KM_IRQ_PTE || type == KM_NMI ||
462                         type == KM_NMI_PTE ) {
463                 if (!irqs_disabled()) {
464                         WARN_ON(1);
465                         warn_count--;
466                 }
467         } else if (type == KM_SOFTIRQ0 || type == KM_SOFTIRQ1) {
468                 if (irq_count() == 0 && !irqs_disabled()) {
469                         WARN_ON(1);
470                         warn_count--;
471                 }
472         }
473 }
474
475 #endif