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