[PATCH] reduce MAX_NR_ZONES: move HIGHMEM counters into highmem.c/.h
[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 static mempool_t *page_pool, *isa_page_pool;
33
34 static void *mempool_alloc_pages_isa(gfp_t gfp_mask, void *data)
35 {
36         return mempool_alloc_pages(gfp_mask | GFP_DMA, data);
37 }
38
39 /*
40  * Virtual_count is not a pure "count".
41  *  0 means that it is not mapped, and has not been mapped
42  *    since a TLB flush - it is usable.
43  *  1 means that there are no users, but it has been mapped
44  *    since the last TLB flush - so we can't use it.
45  *  n means that there are (n-1) current users of it.
46  */
47 #ifdef CONFIG_HIGHMEM
48
49 unsigned long totalhigh_pages __read_mostly;
50
51 unsigned int nr_free_highpages (void)
52 {
53         pg_data_t *pgdat;
54         unsigned int pages = 0;
55
56         for_each_online_pgdat(pgdat)
57                 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
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 static void flush_all_zero_pkmaps(void)
71 {
72         int i;
73
74         flush_cache_kmaps();
75
76         for (i = 0; i < LAST_PKMAP; i++) {
77                 struct page *page;
78
79                 /*
80                  * zero means we don't have anything to do,
81                  * >1 means that it is still in use. Only
82                  * a count of 1 means that it is free but
83                  * needs to be unmapped
84                  */
85                 if (pkmap_count[i] != 1)
86                         continue;
87                 pkmap_count[i] = 0;
88
89                 /* sanity check */
90                 BUG_ON(pte_none(pkmap_page_table[i]));
91
92                 /*
93                  * Don't need an atomic fetch-and-clear op here;
94                  * no-one has the page mapped, and cannot get at
95                  * its virtual address (and hence PTE) without first
96                  * getting the kmap_lock (which is held here).
97                  * So no dangers, even with speculative execution.
98                  */
99                 page = pte_page(pkmap_page_table[i]);
100                 pte_clear(&init_mm, (unsigned long)page_address(page),
101                           &pkmap_page_table[i]);
102
103                 set_page_address(page, NULL);
104         }
105         flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
106 }
107
108 static inline unsigned long map_new_virtual(struct page *page)
109 {
110         unsigned long vaddr;
111         int count;
112
113 start:
114         count = LAST_PKMAP;
115         /* Find an empty entry */
116         for (;;) {
117                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
118                 if (!last_pkmap_nr) {
119                         flush_all_zero_pkmaps();
120                         count = LAST_PKMAP;
121                 }
122                 if (!pkmap_count[last_pkmap_nr])
123                         break;  /* Found a usable entry */
124                 if (--count)
125                         continue;
126
127                 /*
128                  * Sleep for somebody else to unmap their entries
129                  */
130                 {
131                         DECLARE_WAITQUEUE(wait, current);
132
133                         __set_current_state(TASK_UNINTERRUPTIBLE);
134                         add_wait_queue(&pkmap_map_wait, &wait);
135                         spin_unlock(&kmap_lock);
136                         schedule();
137                         remove_wait_queue(&pkmap_map_wait, &wait);
138                         spin_lock(&kmap_lock);
139
140                         /* Somebody else might have mapped it while we slept */
141                         if (page_address(page))
142                                 return (unsigned long)page_address(page);
143
144                         /* Re-start */
145                         goto start;
146                 }
147         }
148         vaddr = PKMAP_ADDR(last_pkmap_nr);
149         set_pte_at(&init_mm, vaddr,
150                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
151
152         pkmap_count[last_pkmap_nr] = 1;
153         set_page_address(page, (void *)vaddr);
154
155         return vaddr;
156 }
157
158 void fastcall *kmap_high(struct page *page)
159 {
160         unsigned long vaddr;
161
162         /*
163          * For highmem pages, we can't trust "virtual" until
164          * after we have the lock.
165          *
166          * We cannot call this from interrupts, as it may block
167          */
168         spin_lock(&kmap_lock);
169         vaddr = (unsigned long)page_address(page);
170         if (!vaddr)
171                 vaddr = map_new_virtual(page);
172         pkmap_count[PKMAP_NR(vaddr)]++;
173         BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
174         spin_unlock(&kmap_lock);
175         return (void*) vaddr;
176 }
177
178 EXPORT_SYMBOL(kmap_high);
179
180 void fastcall kunmap_high(struct page *page)
181 {
182         unsigned long vaddr;
183         unsigned long nr;
184         int need_wakeup;
185
186         spin_lock(&kmap_lock);
187         vaddr = (unsigned long)page_address(page);
188         BUG_ON(!vaddr);
189         nr = PKMAP_NR(vaddr);
190
191         /*
192          * A count must never go down to zero
193          * without a TLB flush!
194          */
195         need_wakeup = 0;
196         switch (--pkmap_count[nr]) {
197         case 0:
198                 BUG();
199         case 1:
200                 /*
201                  * Avoid an unnecessary wake_up() function call.
202                  * The common case is pkmap_count[] == 1, but
203                  * no waiters.
204                  * The tasks queued in the wait-queue are guarded
205                  * by both the lock in the wait-queue-head and by
206                  * the kmap_lock.  As the kmap_lock is held here,
207                  * no need for the wait-queue-head's lock.  Simply
208                  * test if the queue is empty.
209                  */
210                 need_wakeup = waitqueue_active(&pkmap_map_wait);
211         }
212         spin_unlock(&kmap_lock);
213
214         /* do wake-up, if needed, race-free outside of the spin lock */
215         if (need_wakeup)
216                 wake_up(&pkmap_map_wait);
217 }
218
219 EXPORT_SYMBOL(kunmap_high);
220
221 #define POOL_SIZE       64
222
223 static __init int init_emergency_pool(void)
224 {
225         struct sysinfo i;
226         si_meminfo(&i);
227         si_swapinfo(&i);
228         
229         if (!i.totalhigh)
230                 return 0;
231
232         page_pool = mempool_create_page_pool(POOL_SIZE, 0);
233         BUG_ON(!page_pool);
234         printk("highmem bounce pool size: %d pages\n", POOL_SIZE);
235
236         return 0;
237 }
238
239 __initcall(init_emergency_pool);
240
241 /*
242  * highmem version, map in to vec
243  */
244 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
245 {
246         unsigned long flags;
247         unsigned char *vto;
248
249         local_irq_save(flags);
250         vto = kmap_atomic(to->bv_page, KM_BOUNCE_READ);
251         memcpy(vto + to->bv_offset, vfrom, to->bv_len);
252         kunmap_atomic(vto, KM_BOUNCE_READ);
253         local_irq_restore(flags);
254 }
255
256 #else /* CONFIG_HIGHMEM */
257
258 #define bounce_copy_vec(to, vfrom)      \
259         memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
260
261 #endif
262
263 #define ISA_POOL_SIZE   16
264
265 /*
266  * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA
267  * as the max address, so check if the pool has already been created.
268  */
269 int init_emergency_isa_pool(void)
270 {
271         if (isa_page_pool)
272                 return 0;
273
274         isa_page_pool = mempool_create(ISA_POOL_SIZE, mempool_alloc_pages_isa,
275                                        mempool_free_pages, (void *) 0);
276         BUG_ON(!isa_page_pool);
277
278         printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
279         return 0;
280 }
281
282 /*
283  * Simple bounce buffer support for highmem pages. Depending on the
284  * queue gfp mask set, *to may or may not be a highmem page. kmap it
285  * always, it will do the Right Thing
286  */
287 static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
288 {
289         unsigned char *vfrom;
290         struct bio_vec *tovec, *fromvec;
291         int i;
292
293         __bio_for_each_segment(tovec, to, i, 0) {
294                 fromvec = from->bi_io_vec + i;
295
296                 /*
297                  * not bounced
298                  */
299                 if (tovec->bv_page == fromvec->bv_page)
300                         continue;
301
302                 /*
303                  * fromvec->bv_offset and fromvec->bv_len might have been
304                  * modified by the block layer, so use the original copy,
305                  * bounce_copy_vec already uses tovec->bv_len
306                  */
307                 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
308
309                 flush_dcache_page(tovec->bv_page);
310                 bounce_copy_vec(tovec, vfrom);
311         }
312 }
313
314 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err)
315 {
316         struct bio *bio_orig = bio->bi_private;
317         struct bio_vec *bvec, *org_vec;
318         int i;
319
320         if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags))
321                 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags);
322
323         /*
324          * free up bounce indirect pages used
325          */
326         __bio_for_each_segment(bvec, bio, i, 0) {
327                 org_vec = bio_orig->bi_io_vec + i;
328                 if (bvec->bv_page == org_vec->bv_page)
329                         continue;
330
331                 dec_zone_page_state(bvec->bv_page, NR_BOUNCE);
332                 mempool_free(bvec->bv_page, pool);
333         }
334
335         bio_endio(bio_orig, bio_orig->bi_size, err);
336         bio_put(bio);
337 }
338
339 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done, int err)
340 {
341         if (bio->bi_size)
342                 return 1;
343
344         bounce_end_io(bio, page_pool, err);
345         return 0;
346 }
347
348 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err)
349 {
350         if (bio->bi_size)
351                 return 1;
352
353         bounce_end_io(bio, isa_page_pool, err);
354         return 0;
355 }
356
357 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err)
358 {
359         struct bio *bio_orig = bio->bi_private;
360
361         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
362                 copy_to_high_bio_irq(bio_orig, bio);
363
364         bounce_end_io(bio, pool, err);
365 }
366
367 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
368 {
369         if (bio->bi_size)
370                 return 1;
371
372         __bounce_end_io_read(bio, page_pool, err);
373         return 0;
374 }
375
376 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err)
377 {
378         if (bio->bi_size)
379                 return 1;
380
381         __bounce_end_io_read(bio, isa_page_pool, err);
382         return 0;
383 }
384
385 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig,
386                                mempool_t *pool)
387 {
388         struct page *page;
389         struct bio *bio = NULL;
390         int i, rw = bio_data_dir(*bio_orig);
391         struct bio_vec *to, *from;
392
393         bio_for_each_segment(from, *bio_orig, i) {
394                 page = from->bv_page;
395
396                 /*
397                  * is destination page below bounce pfn?
398                  */
399                 if (page_to_pfn(page) < q->bounce_pfn)
400                         continue;
401
402                 /*
403                  * irk, bounce it
404                  */
405                 if (!bio)
406                         bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt);
407
408                 to = bio->bi_io_vec + i;
409
410                 to->bv_page = mempool_alloc(pool, q->bounce_gfp);
411                 to->bv_len = from->bv_len;
412                 to->bv_offset = from->bv_offset;
413                 inc_zone_page_state(to->bv_page, NR_BOUNCE);
414
415                 if (rw == WRITE) {
416                         char *vto, *vfrom;
417
418                         flush_dcache_page(from->bv_page);
419                         vto = page_address(to->bv_page) + to->bv_offset;
420                         vfrom = kmap(from->bv_page) + from->bv_offset;
421                         memcpy(vto, vfrom, to->bv_len);
422                         kunmap(from->bv_page);
423                 }
424         }
425
426         /*
427          * no pages bounced
428          */
429         if (!bio)
430                 return;
431
432         /*
433          * at least one page was bounced, fill in possible non-highmem
434          * pages
435          */
436         __bio_for_each_segment(from, *bio_orig, i, 0) {
437                 to = bio_iovec_idx(bio, i);
438                 if (!to->bv_page) {
439                         to->bv_page = from->bv_page;
440                         to->bv_len = from->bv_len;
441                         to->bv_offset = from->bv_offset;
442                 }
443         }
444
445         bio->bi_bdev = (*bio_orig)->bi_bdev;
446         bio->bi_flags |= (1 << BIO_BOUNCED);
447         bio->bi_sector = (*bio_orig)->bi_sector;
448         bio->bi_rw = (*bio_orig)->bi_rw;
449
450         bio->bi_vcnt = (*bio_orig)->bi_vcnt;
451         bio->bi_idx = (*bio_orig)->bi_idx;
452         bio->bi_size = (*bio_orig)->bi_size;
453
454         if (pool == page_pool) {
455                 bio->bi_end_io = bounce_end_io_write;
456                 if (rw == READ)
457                         bio->bi_end_io = bounce_end_io_read;
458         } else {
459                 bio->bi_end_io = bounce_end_io_write_isa;
460                 if (rw == READ)
461                         bio->bi_end_io = bounce_end_io_read_isa;
462         }
463
464         bio->bi_private = *bio_orig;
465         *bio_orig = bio;
466 }
467
468 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
469 {
470         mempool_t *pool;
471
472         /*
473          * for non-isa bounce case, just check if the bounce pfn is equal
474          * to or bigger than the highest pfn in the system -- in that case,
475          * don't waste time iterating over bio segments
476          */
477         if (!(q->bounce_gfp & GFP_DMA)) {
478                 if (q->bounce_pfn >= blk_max_pfn)
479                         return;
480                 pool = page_pool;
481         } else {
482                 BUG_ON(!isa_page_pool);
483                 pool = isa_page_pool;
484         }
485
486         blk_add_trace_bio(q, *bio_orig, BLK_TA_BOUNCE);
487
488         /*
489          * slow path
490          */
491         __blk_queue_bounce(q, bio_orig, pool);
492 }
493
494 EXPORT_SYMBOL(blk_queue_bounce);
495
496 #if defined(HASHED_PAGE_VIRTUAL)
497
498 #define PA_HASH_ORDER   7
499
500 /*
501  * Describes one page->virtual association
502  */
503 struct page_address_map {
504         struct page *page;
505         void *virtual;
506         struct list_head list;
507 };
508
509 /*
510  * page_address_map freelist, allocated from page_address_maps.
511  */
512 static struct list_head page_address_pool;      /* freelist */
513 static spinlock_t pool_lock;                    /* protects page_address_pool */
514
515 /*
516  * Hash table bucket
517  */
518 static struct page_address_slot {
519         struct list_head lh;                    /* List of page_address_maps */
520         spinlock_t lock;                        /* Protect this bucket's list */
521 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
522
523 static struct page_address_slot *page_slot(struct page *page)
524 {
525         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
526 }
527
528 void *page_address(struct page *page)
529 {
530         unsigned long flags;
531         void *ret;
532         struct page_address_slot *pas;
533
534         if (!PageHighMem(page))
535                 return lowmem_page_address(page);
536
537         pas = page_slot(page);
538         ret = NULL;
539         spin_lock_irqsave(&pas->lock, flags);
540         if (!list_empty(&pas->lh)) {
541                 struct page_address_map *pam;
542
543                 list_for_each_entry(pam, &pas->lh, list) {
544                         if (pam->page == page) {
545                                 ret = pam->virtual;
546                                 goto done;
547                         }
548                 }
549         }
550 done:
551         spin_unlock_irqrestore(&pas->lock, flags);
552         return ret;
553 }
554
555 EXPORT_SYMBOL(page_address);
556
557 void set_page_address(struct page *page, void *virtual)
558 {
559         unsigned long flags;
560         struct page_address_slot *pas;
561         struct page_address_map *pam;
562
563         BUG_ON(!PageHighMem(page));
564
565         pas = page_slot(page);
566         if (virtual) {          /* Add */
567                 BUG_ON(list_empty(&page_address_pool));
568
569                 spin_lock_irqsave(&pool_lock, flags);
570                 pam = list_entry(page_address_pool.next,
571                                 struct page_address_map, list);
572                 list_del(&pam->list);
573                 spin_unlock_irqrestore(&pool_lock, flags);
574
575                 pam->page = page;
576                 pam->virtual = virtual;
577
578                 spin_lock_irqsave(&pas->lock, flags);
579                 list_add_tail(&pam->list, &pas->lh);
580                 spin_unlock_irqrestore(&pas->lock, flags);
581         } else {                /* Remove */
582                 spin_lock_irqsave(&pas->lock, flags);
583                 list_for_each_entry(pam, &pas->lh, list) {
584                         if (pam->page == page) {
585                                 list_del(&pam->list);
586                                 spin_unlock_irqrestore(&pas->lock, flags);
587                                 spin_lock_irqsave(&pool_lock, flags);
588                                 list_add_tail(&pam->list, &page_address_pool);
589                                 spin_unlock_irqrestore(&pool_lock, flags);
590                                 goto done;
591                         }
592                 }
593                 spin_unlock_irqrestore(&pas->lock, flags);
594         }
595 done:
596         return;
597 }
598
599 static struct page_address_map page_address_maps[LAST_PKMAP];
600
601 void __init page_address_init(void)
602 {
603         int i;
604
605         INIT_LIST_HEAD(&page_address_pool);
606         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
607                 list_add(&page_address_maps[i].list, &page_address_pool);
608         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
609                 INIT_LIST_HEAD(&page_address_htable[i].lh);
610                 spin_lock_init(&page_address_htable[i].lock);
611         }
612         spin_lock_init(&pool_lock);
613 }
614
615 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */