[PATCH] count bounce buffer pages in vmstat
[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 static mempool_t *page_pool, *isa_page_pool;
32
33 static void *page_pool_alloc(unsigned int __nocast gfp_mask, void *data)
34 {
35         unsigned int gfp = gfp_mask | (unsigned int) (long) data;
36
37         return alloc_page(gfp);
38 }
39
40 static void page_pool_free(void *page, void *data)
41 {
42         __free_page(page);
43 }
44
45 /*
46  * Virtual_count is not a pure "count".
47  *  0 means that it is not mapped, and has not been mapped
48  *    since a TLB flush - it is usable.
49  *  1 means that there are no users, but it has been mapped
50  *    since the last TLB flush - so we can't use it.
51  *  n means that there are (n-1) current users of it.
52  */
53 #ifdef CONFIG_HIGHMEM
54 static int pkmap_count[LAST_PKMAP];
55 static unsigned int last_pkmap_nr;
56 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
57
58 pte_t * pkmap_page_table;
59
60 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
61
62 static void flush_all_zero_pkmaps(void)
63 {
64         int i;
65
66         flush_cache_kmaps();
67
68         for (i = 0; i < LAST_PKMAP; i++) {
69                 struct page *page;
70
71                 /*
72                  * zero means we don't have anything to do,
73                  * >1 means that it is still in use. Only
74                  * a count of 1 means that it is free but
75                  * needs to be unmapped
76                  */
77                 if (pkmap_count[i] != 1)
78                         continue;
79                 pkmap_count[i] = 0;
80
81                 /* sanity check */
82                 if (pte_none(pkmap_page_table[i]))
83                         BUG();
84
85                 /*
86                  * Don't need an atomic fetch-and-clear op here;
87                  * no-one has the page mapped, and cannot get at
88                  * its virtual address (and hence PTE) without first
89                  * getting the kmap_lock (which is held here).
90                  * So no dangers, even with speculative execution.
91                  */
92                 page = pte_page(pkmap_page_table[i]);
93                 pte_clear(&init_mm, (unsigned long)page_address(page),
94                           &pkmap_page_table[i]);
95
96                 set_page_address(page, NULL);
97         }
98         flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
99 }
100
101 static inline unsigned long map_new_virtual(struct page *page)
102 {
103         unsigned long vaddr;
104         int count;
105
106 start:
107         count = LAST_PKMAP;
108         /* Find an empty entry */
109         for (;;) {
110                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
111                 if (!last_pkmap_nr) {
112                         flush_all_zero_pkmaps();
113                         count = LAST_PKMAP;
114                 }
115                 if (!pkmap_count[last_pkmap_nr])
116                         break;  /* Found a usable entry */
117                 if (--count)
118                         continue;
119
120                 /*
121                  * Sleep for somebody else to unmap their entries
122                  */
123                 {
124                         DECLARE_WAITQUEUE(wait, current);
125
126                         __set_current_state(TASK_UNINTERRUPTIBLE);
127                         add_wait_queue(&pkmap_map_wait, &wait);
128                         spin_unlock(&kmap_lock);
129                         schedule();
130                         remove_wait_queue(&pkmap_map_wait, &wait);
131                         spin_lock(&kmap_lock);
132
133                         /* Somebody else might have mapped it while we slept */
134                         if (page_address(page))
135                                 return (unsigned long)page_address(page);
136
137                         /* Re-start */
138                         goto start;
139                 }
140         }
141         vaddr = PKMAP_ADDR(last_pkmap_nr);
142         set_pte_at(&init_mm, vaddr,
143                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
144
145         pkmap_count[last_pkmap_nr] = 1;
146         set_page_address(page, (void *)vaddr);
147
148         return vaddr;
149 }
150
151 void fastcall *kmap_high(struct page *page)
152 {
153         unsigned long vaddr;
154
155         /*
156          * For highmem pages, we can't trust "virtual" until
157          * after we have the lock.
158          *
159          * We cannot call this from interrupts, as it may block
160          */
161         spin_lock(&kmap_lock);
162         vaddr = (unsigned long)page_address(page);
163         if (!vaddr)
164                 vaddr = map_new_virtual(page);
165         pkmap_count[PKMAP_NR(vaddr)]++;
166         if (pkmap_count[PKMAP_NR(vaddr)] < 2)
167                 BUG();
168         spin_unlock(&kmap_lock);
169         return (void*) vaddr;
170 }
171
172 EXPORT_SYMBOL(kmap_high);
173
174 void fastcall kunmap_high(struct page *page)
175 {
176         unsigned long vaddr;
177         unsigned long nr;
178         int need_wakeup;
179
180         spin_lock(&kmap_lock);
181         vaddr = (unsigned long)page_address(page);
182         if (!vaddr)
183                 BUG();
184         nr = PKMAP_NR(vaddr);
185
186         /*
187          * A count must never go down to zero
188          * without a TLB flush!
189          */
190         need_wakeup = 0;
191         switch (--pkmap_count[nr]) {
192         case 0:
193                 BUG();
194         case 1:
195                 /*
196                  * Avoid an unnecessary wake_up() function call.
197                  * The common case is pkmap_count[] == 1, but
198                  * no waiters.
199                  * The tasks queued in the wait-queue are guarded
200                  * by both the lock in the wait-queue-head and by
201                  * the kmap_lock.  As the kmap_lock is held here,
202                  * no need for the wait-queue-head's lock.  Simply
203                  * test if the queue is empty.
204                  */
205                 need_wakeup = waitqueue_active(&pkmap_map_wait);
206         }
207         spin_unlock(&kmap_lock);
208
209         /* do wake-up, if needed, race-free outside of the spin lock */
210         if (need_wakeup)
211                 wake_up(&pkmap_map_wait);
212 }
213
214 EXPORT_SYMBOL(kunmap_high);
215
216 #define POOL_SIZE       64
217
218 static __init int init_emergency_pool(void)
219 {
220         struct sysinfo i;
221         si_meminfo(&i);
222         si_swapinfo(&i);
223         
224         if (!i.totalhigh)
225                 return 0;
226
227         page_pool = mempool_create(POOL_SIZE, page_pool_alloc, page_pool_free, NULL);
228         if (!page_pool)
229                 BUG();
230         printk("highmem bounce pool size: %d pages\n", POOL_SIZE);
231
232         return 0;
233 }
234
235 __initcall(init_emergency_pool);
236
237 /*
238  * highmem version, map in to vec
239  */
240 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
241 {
242         unsigned long flags;
243         unsigned char *vto;
244
245         local_irq_save(flags);
246         vto = kmap_atomic(to->bv_page, KM_BOUNCE_READ);
247         memcpy(vto + to->bv_offset, vfrom, to->bv_len);
248         kunmap_atomic(vto, KM_BOUNCE_READ);
249         local_irq_restore(flags);
250 }
251
252 #else /* CONFIG_HIGHMEM */
253
254 #define bounce_copy_vec(to, vfrom)      \
255         memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
256
257 #endif
258
259 #define ISA_POOL_SIZE   16
260
261 /*
262  * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA
263  * as the max address, so check if the pool has already been created.
264  */
265 int init_emergency_isa_pool(void)
266 {
267         if (isa_page_pool)
268                 return 0;
269
270         isa_page_pool = mempool_create(ISA_POOL_SIZE, page_pool_alloc, page_pool_free, (void *) __GFP_DMA);
271         if (!isa_page_pool)
272                 BUG();
273
274         printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
275         return 0;
276 }
277
278 /*
279  * Simple bounce buffer support for highmem pages. Depending on the
280  * queue gfp mask set, *to may or may not be a highmem page. kmap it
281  * always, it will do the Right Thing
282  */
283 static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
284 {
285         unsigned char *vfrom;
286         struct bio_vec *tovec, *fromvec;
287         int i;
288
289         __bio_for_each_segment(tovec, to, i, 0) {
290                 fromvec = from->bi_io_vec + i;
291
292                 /*
293                  * not bounced
294                  */
295                 if (tovec->bv_page == fromvec->bv_page)
296                         continue;
297
298                 /*
299                  * fromvec->bv_offset and fromvec->bv_len might have been
300                  * modified by the block layer, so use the original copy,
301                  * bounce_copy_vec already uses tovec->bv_len
302                  */
303                 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
304
305                 flush_dcache_page(tovec->bv_page);
306                 bounce_copy_vec(tovec, vfrom);
307         }
308 }
309
310 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err)
311 {
312         struct bio *bio_orig = bio->bi_private;
313         struct bio_vec *bvec, *org_vec;
314         int i;
315
316         if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags))
317                 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags);
318
319         /*
320          * free up bounce indirect pages used
321          */
322         __bio_for_each_segment(bvec, bio, i, 0) {
323                 org_vec = bio_orig->bi_io_vec + i;
324                 if (bvec->bv_page == org_vec->bv_page)
325                         continue;
326
327                 mempool_free(bvec->bv_page, pool);      
328                 dec_page_state(nr_bounce);
329         }
330
331         bio_endio(bio_orig, bio_orig->bi_size, err);
332         bio_put(bio);
333 }
334
335 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done,int err)
336 {
337         if (bio->bi_size)
338                 return 1;
339
340         bounce_end_io(bio, page_pool, err);
341         return 0;
342 }
343
344 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err)
345 {
346         if (bio->bi_size)
347                 return 1;
348
349         bounce_end_io(bio, isa_page_pool, err);
350         return 0;
351 }
352
353 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err)
354 {
355         struct bio *bio_orig = bio->bi_private;
356
357         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
358                 copy_to_high_bio_irq(bio_orig, bio);
359
360         bounce_end_io(bio, pool, err);
361 }
362
363 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
364 {
365         if (bio->bi_size)
366                 return 1;
367
368         __bounce_end_io_read(bio, page_pool, err);
369         return 0;
370 }
371
372 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err)
373 {
374         if (bio->bi_size)
375                 return 1;
376
377         __bounce_end_io_read(bio, isa_page_pool, err);
378         return 0;
379 }
380
381 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig,
382                         mempool_t *pool)
383 {
384         struct page *page;
385         struct bio *bio = NULL;
386         int i, rw = bio_data_dir(*bio_orig);
387         struct bio_vec *to, *from;
388
389         bio_for_each_segment(from, *bio_orig, i) {
390                 page = from->bv_page;
391
392                 /*
393                  * is destination page below bounce pfn?
394                  */
395                 if (page_to_pfn(page) < q->bounce_pfn)
396                         continue;
397
398                 /*
399                  * irk, bounce it
400                  */
401                 if (!bio)
402                         bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt);
403
404                 to = bio->bi_io_vec + i;
405
406                 to->bv_page = mempool_alloc(pool, q->bounce_gfp);
407                 to->bv_len = from->bv_len;
408                 to->bv_offset = from->bv_offset;
409                 inc_page_state(nr_bounce);
410
411                 if (rw == WRITE) {
412                         char *vto, *vfrom;
413
414                         flush_dcache_page(from->bv_page);
415                         vto = page_address(to->bv_page) + to->bv_offset;
416                         vfrom = kmap(from->bv_page) + from->bv_offset;
417                         memcpy(vto, vfrom, to->bv_len);
418                         kunmap(from->bv_page);
419                 }
420         }
421
422         /*
423          * no pages bounced
424          */
425         if (!bio)
426                 return;
427
428         /*
429          * at least one page was bounced, fill in possible non-highmem
430          * pages
431          */
432         __bio_for_each_segment(from, *bio_orig, i, 0) {
433                 to = bio_iovec_idx(bio, i);
434                 if (!to->bv_page) {
435                         to->bv_page = from->bv_page;
436                         to->bv_len = from->bv_len;
437                         to->bv_offset = from->bv_offset;
438                 }
439         }
440
441         bio->bi_bdev = (*bio_orig)->bi_bdev;
442         bio->bi_flags |= (1 << BIO_BOUNCED);
443         bio->bi_sector = (*bio_orig)->bi_sector;
444         bio->bi_rw = (*bio_orig)->bi_rw;
445
446         bio->bi_vcnt = (*bio_orig)->bi_vcnt;
447         bio->bi_idx = (*bio_orig)->bi_idx;
448         bio->bi_size = (*bio_orig)->bi_size;
449
450         if (pool == page_pool) {
451                 bio->bi_end_io = bounce_end_io_write;
452                 if (rw == READ)
453                         bio->bi_end_io = bounce_end_io_read;
454         } else {
455                 bio->bi_end_io = bounce_end_io_write_isa;
456                 if (rw == READ)
457                         bio->bi_end_io = bounce_end_io_read_isa;
458         }
459
460         bio->bi_private = *bio_orig;
461         *bio_orig = bio;
462 }
463
464 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
465 {
466         mempool_t *pool;
467
468         /*
469          * for non-isa bounce case, just check if the bounce pfn is equal
470          * to or bigger than the highest pfn in the system -- in that case,
471          * don't waste time iterating over bio segments
472          */
473         if (!(q->bounce_gfp & GFP_DMA)) {
474                 if (q->bounce_pfn >= blk_max_pfn)
475                         return;
476                 pool = page_pool;
477         } else {
478                 BUG_ON(!isa_page_pool);
479                 pool = isa_page_pool;
480         }
481
482         /*
483          * slow path
484          */
485         __blk_queue_bounce(q, bio_orig, pool);
486 }
487
488 EXPORT_SYMBOL(blk_queue_bounce);
489
490 #if defined(HASHED_PAGE_VIRTUAL)
491
492 #define PA_HASH_ORDER   7
493
494 /*
495  * Describes one page->virtual association
496  */
497 struct page_address_map {
498         struct page *page;
499         void *virtual;
500         struct list_head list;
501 };
502
503 /*
504  * page_address_map freelist, allocated from page_address_maps.
505  */
506 static struct list_head page_address_pool;      /* freelist */
507 static spinlock_t pool_lock;                    /* protects page_address_pool */
508
509 /*
510  * Hash table bucket
511  */
512 static struct page_address_slot {
513         struct list_head lh;                    /* List of page_address_maps */
514         spinlock_t lock;                        /* Protect this bucket's list */
515 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
516
517 static struct page_address_slot *page_slot(struct page *page)
518 {
519         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
520 }
521
522 void *page_address(struct page *page)
523 {
524         unsigned long flags;
525         void *ret;
526         struct page_address_slot *pas;
527
528         if (!PageHighMem(page))
529                 return lowmem_page_address(page);
530
531         pas = page_slot(page);
532         ret = NULL;
533         spin_lock_irqsave(&pas->lock, flags);
534         if (!list_empty(&pas->lh)) {
535                 struct page_address_map *pam;
536
537                 list_for_each_entry(pam, &pas->lh, list) {
538                         if (pam->page == page) {
539                                 ret = pam->virtual;
540                                 goto done;
541                         }
542                 }
543         }
544 done:
545         spin_unlock_irqrestore(&pas->lock, flags);
546         return ret;
547 }
548
549 EXPORT_SYMBOL(page_address);
550
551 void set_page_address(struct page *page, void *virtual)
552 {
553         unsigned long flags;
554         struct page_address_slot *pas;
555         struct page_address_map *pam;
556
557         BUG_ON(!PageHighMem(page));
558
559         pas = page_slot(page);
560         if (virtual) {          /* Add */
561                 BUG_ON(list_empty(&page_address_pool));
562
563                 spin_lock_irqsave(&pool_lock, flags);
564                 pam = list_entry(page_address_pool.next,
565                                 struct page_address_map, list);
566                 list_del(&pam->list);
567                 spin_unlock_irqrestore(&pool_lock, flags);
568
569                 pam->page = page;
570                 pam->virtual = virtual;
571
572                 spin_lock_irqsave(&pas->lock, flags);
573                 list_add_tail(&pam->list, &pas->lh);
574                 spin_unlock_irqrestore(&pas->lock, flags);
575         } else {                /* Remove */
576                 spin_lock_irqsave(&pas->lock, flags);
577                 list_for_each_entry(pam, &pas->lh, list) {
578                         if (pam->page == page) {
579                                 list_del(&pam->list);
580                                 spin_unlock_irqrestore(&pas->lock, flags);
581                                 spin_lock_irqsave(&pool_lock, flags);
582                                 list_add_tail(&pam->list, &page_address_pool);
583                                 spin_unlock_irqrestore(&pool_lock, flags);
584                                 goto done;
585                         }
586                 }
587                 spin_unlock_irqrestore(&pas->lock, flags);
588         }
589 done:
590         return;
591 }
592
593 static struct page_address_map page_address_maps[LAST_PKMAP];
594
595 void __init page_address_init(void)
596 {
597         int i;
598
599         INIT_LIST_HEAD(&page_address_pool);
600         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
601                 list_add(&page_address_maps[i].list, &page_address_pool);
602         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
603                 INIT_LIST_HEAD(&page_address_htable[i].lh);
604                 spin_lock_init(&page_address_htable[i].lock);
605         }
606         spin_lock_init(&pool_lock);
607 }
608
609 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */