2 * High memory handling common code and variables.
4 * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
5 * Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
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.
12 * Rewrote high memory support to move the page cache into
13 * high memory. Implemented permanent (schedulable) kmaps
14 * based on Linus' idea.
16 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
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>
32 static mempool_t *page_pool, *isa_page_pool;
34 static void *mempool_alloc_pages_isa(gfp_t gfp_mask, void *data)
36 return mempool_alloc_pages(gfp_mask | GFP_DMA, data);
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.
49 static int pkmap_count[LAST_PKMAP];
50 static unsigned int last_pkmap_nr;
51 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
53 pte_t * pkmap_page_table;
55 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
57 static void flush_all_zero_pkmaps(void)
63 for (i = 0; i < LAST_PKMAP; i++) {
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
72 if (pkmap_count[i] != 1)
77 if (pte_none(pkmap_page_table[i]))
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.
87 page = pte_page(pkmap_page_table[i]);
88 pte_clear(&init_mm, (unsigned long)page_address(page),
89 &pkmap_page_table[i]);
91 set_page_address(page, NULL);
93 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
96 static inline unsigned long map_new_virtual(struct page *page)
103 /* Find an empty entry */
105 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
106 if (!last_pkmap_nr) {
107 flush_all_zero_pkmaps();
110 if (!pkmap_count[last_pkmap_nr])
111 break; /* Found a usable entry */
116 * Sleep for somebody else to unmap their entries
119 DECLARE_WAITQUEUE(wait, current);
121 __set_current_state(TASK_UNINTERRUPTIBLE);
122 add_wait_queue(&pkmap_map_wait, &wait);
123 spin_unlock(&kmap_lock);
125 remove_wait_queue(&pkmap_map_wait, &wait);
126 spin_lock(&kmap_lock);
128 /* Somebody else might have mapped it while we slept */
129 if (page_address(page))
130 return (unsigned long)page_address(page);
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));
140 pkmap_count[last_pkmap_nr] = 1;
141 set_page_address(page, (void *)vaddr);
146 void fastcall *kmap_high(struct page *page)
151 * For highmem pages, we can't trust "virtual" until
152 * after we have the lock.
154 * We cannot call this from interrupts, as it may block
156 spin_lock(&kmap_lock);
157 vaddr = (unsigned long)page_address(page);
159 vaddr = map_new_virtual(page);
160 pkmap_count[PKMAP_NR(vaddr)]++;
161 if (pkmap_count[PKMAP_NR(vaddr)] < 2)
163 spin_unlock(&kmap_lock);
164 return (void*) vaddr;
167 EXPORT_SYMBOL(kmap_high);
169 void fastcall kunmap_high(struct page *page)
175 spin_lock(&kmap_lock);
176 vaddr = (unsigned long)page_address(page);
179 nr = PKMAP_NR(vaddr);
182 * A count must never go down to zero
183 * without a TLB flush!
186 switch (--pkmap_count[nr]) {
191 * Avoid an unnecessary wake_up() function call.
192 * The common case is pkmap_count[] == 1, but
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.
200 need_wakeup = waitqueue_active(&pkmap_map_wait);
202 spin_unlock(&kmap_lock);
204 /* do wake-up, if needed, race-free outside of the spin lock */
206 wake_up(&pkmap_map_wait);
209 EXPORT_SYMBOL(kunmap_high);
213 static __init int init_emergency_pool(void)
222 page_pool = mempool_create_page_pool(POOL_SIZE, 0);
225 printk("highmem bounce pool size: %d pages\n", POOL_SIZE);
230 __initcall(init_emergency_pool);
233 * highmem version, map in to vec
235 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
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);
247 #else /* CONFIG_HIGHMEM */
249 #define bounce_copy_vec(to, vfrom) \
250 memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
254 #define ISA_POOL_SIZE 16
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.
260 int init_emergency_isa_pool(void)
265 isa_page_pool = mempool_create(ISA_POOL_SIZE, mempool_alloc_pages_isa,
266 mempool_free_pages, (void *) 0);
270 printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
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
279 static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
281 unsigned char *vfrom;
282 struct bio_vec *tovec, *fromvec;
285 __bio_for_each_segment(tovec, to, i, 0) {
286 fromvec = from->bi_io_vec + i;
291 if (tovec->bv_page == fromvec->bv_page)
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
299 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
301 flush_dcache_page(tovec->bv_page);
302 bounce_copy_vec(tovec, vfrom);
306 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err)
308 struct bio *bio_orig = bio->bi_private;
309 struct bio_vec *bvec, *org_vec;
312 if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags))
313 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags);
316 * free up bounce indirect pages used
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)
323 mempool_free(bvec->bv_page, pool);
324 dec_page_state(nr_bounce);
327 bio_endio(bio_orig, bio_orig->bi_size, err);
331 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done, int err)
336 bounce_end_io(bio, page_pool, err);
340 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err)
345 bounce_end_io(bio, isa_page_pool, err);
349 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err)
351 struct bio *bio_orig = bio->bi_private;
353 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
354 copy_to_high_bio_irq(bio_orig, bio);
356 bounce_end_io(bio, pool, err);
359 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
364 __bounce_end_io_read(bio, page_pool, err);
368 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err)
373 __bounce_end_io_read(bio, isa_page_pool, err);
377 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig,
381 struct bio *bio = NULL;
382 int i, rw = bio_data_dir(*bio_orig);
383 struct bio_vec *to, *from;
385 bio_for_each_segment(from, *bio_orig, i) {
386 page = from->bv_page;
389 * is destination page below bounce pfn?
391 if (page_to_pfn(page) < q->bounce_pfn)
398 bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt);
400 to = bio->bi_io_vec + i;
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);
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);
425 * at least one page was bounced, fill in possible non-highmem
428 __bio_for_each_segment(from, *bio_orig, i, 0) {
429 to = bio_iovec_idx(bio, i);
431 to->bv_page = from->bv_page;
432 to->bv_len = from->bv_len;
433 to->bv_offset = from->bv_offset;
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;
442 bio->bi_vcnt = (*bio_orig)->bi_vcnt;
443 bio->bi_idx = (*bio_orig)->bi_idx;
444 bio->bi_size = (*bio_orig)->bi_size;
446 if (pool == page_pool) {
447 bio->bi_end_io = bounce_end_io_write;
449 bio->bi_end_io = bounce_end_io_read;
451 bio->bi_end_io = bounce_end_io_write_isa;
453 bio->bi_end_io = bounce_end_io_read_isa;
456 bio->bi_private = *bio_orig;
460 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
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
469 if (!(q->bounce_gfp & GFP_DMA)) {
470 if (q->bounce_pfn >= blk_max_pfn)
474 BUG_ON(!isa_page_pool);
475 pool = isa_page_pool;
478 blk_add_trace_bio(q, *bio_orig, BLK_TA_BOUNCE);
483 __blk_queue_bounce(q, bio_orig, pool);
486 EXPORT_SYMBOL(blk_queue_bounce);
488 #if defined(HASHED_PAGE_VIRTUAL)
490 #define PA_HASH_ORDER 7
493 * Describes one page->virtual association
495 struct page_address_map {
498 struct list_head list;
502 * page_address_map freelist, allocated from page_address_maps.
504 static struct list_head page_address_pool; /* freelist */
505 static spinlock_t pool_lock; /* protects page_address_pool */
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];
515 static struct page_address_slot *page_slot(struct page *page)
517 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
520 void *page_address(struct page *page)
524 struct page_address_slot *pas;
526 if (!PageHighMem(page))
527 return lowmem_page_address(page);
529 pas = page_slot(page);
531 spin_lock_irqsave(&pas->lock, flags);
532 if (!list_empty(&pas->lh)) {
533 struct page_address_map *pam;
535 list_for_each_entry(pam, &pas->lh, list) {
536 if (pam->page == page) {
543 spin_unlock_irqrestore(&pas->lock, flags);
547 EXPORT_SYMBOL(page_address);
549 void set_page_address(struct page *page, void *virtual)
552 struct page_address_slot *pas;
553 struct page_address_map *pam;
555 BUG_ON(!PageHighMem(page));
557 pas = page_slot(page);
558 if (virtual) { /* Add */
559 BUG_ON(list_empty(&page_address_pool));
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);
568 pam->virtual = virtual;
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);
585 spin_unlock_irqrestore(&pas->lock, flags);
591 static struct page_address_map page_address_maps[LAST_PKMAP];
593 void __init page_address_init(void)
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);
604 spin_lock_init(&pool_lock);
607 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */