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