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 <asm/tlbflush.h>
31 static mempool_t *page_pool, *isa_page_pool;
33 static void *page_pool_alloc(unsigned int __nocast gfp_mask, void *data)
35 unsigned int gfp = gfp_mask | (unsigned int) (long) data;
37 return alloc_page(gfp);
40 static void page_pool_free(void *page, void *data)
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.
54 static int pkmap_count[LAST_PKMAP];
55 static unsigned int last_pkmap_nr;
56 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
58 pte_t * pkmap_page_table;
60 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
62 static void flush_all_zero_pkmaps(void)
68 for (i = 0; i < LAST_PKMAP; i++) {
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
77 if (pkmap_count[i] != 1)
82 if (pte_none(pkmap_page_table[i]))
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.
92 page = pte_page(pkmap_page_table[i]);
93 pte_clear(&init_mm, (unsigned long)page_address(page),
94 &pkmap_page_table[i]);
96 set_page_address(page, NULL);
98 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
101 static inline unsigned long map_new_virtual(struct page *page)
108 /* Find an empty entry */
110 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
111 if (!last_pkmap_nr) {
112 flush_all_zero_pkmaps();
115 if (!pkmap_count[last_pkmap_nr])
116 break; /* Found a usable entry */
121 * Sleep for somebody else to unmap their entries
124 DECLARE_WAITQUEUE(wait, current);
126 __set_current_state(TASK_UNINTERRUPTIBLE);
127 add_wait_queue(&pkmap_map_wait, &wait);
128 spin_unlock(&kmap_lock);
130 remove_wait_queue(&pkmap_map_wait, &wait);
131 spin_lock(&kmap_lock);
133 /* Somebody else might have mapped it while we slept */
134 if (page_address(page))
135 return (unsigned long)page_address(page);
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));
145 pkmap_count[last_pkmap_nr] = 1;
146 set_page_address(page, (void *)vaddr);
151 void fastcall *kmap_high(struct page *page)
156 * For highmem pages, we can't trust "virtual" until
157 * after we have the lock.
159 * We cannot call this from interrupts, as it may block
161 spin_lock(&kmap_lock);
162 vaddr = (unsigned long)page_address(page);
164 vaddr = map_new_virtual(page);
165 pkmap_count[PKMAP_NR(vaddr)]++;
166 if (pkmap_count[PKMAP_NR(vaddr)] < 2)
168 spin_unlock(&kmap_lock);
169 return (void*) vaddr;
172 EXPORT_SYMBOL(kmap_high);
174 void fastcall kunmap_high(struct page *page)
180 spin_lock(&kmap_lock);
181 vaddr = (unsigned long)page_address(page);
184 nr = PKMAP_NR(vaddr);
187 * A count must never go down to zero
188 * without a TLB flush!
191 switch (--pkmap_count[nr]) {
196 * Avoid an unnecessary wake_up() function call.
197 * The common case is pkmap_count[] == 1, but
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.
205 need_wakeup = waitqueue_active(&pkmap_map_wait);
207 spin_unlock(&kmap_lock);
209 /* do wake-up, if needed, race-free outside of the spin lock */
211 wake_up(&pkmap_map_wait);
214 EXPORT_SYMBOL(kunmap_high);
218 static __init int init_emergency_pool(void)
227 page_pool = mempool_create(POOL_SIZE, page_pool_alloc, page_pool_free, NULL);
230 printk("highmem bounce pool size: %d pages\n", POOL_SIZE);
235 __initcall(init_emergency_pool);
238 * highmem version, map in to vec
240 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
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);
252 #else /* CONFIG_HIGHMEM */
254 #define bounce_copy_vec(to, vfrom) \
255 memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
259 #define ISA_POOL_SIZE 16
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.
265 int init_emergency_isa_pool(void)
270 isa_page_pool = mempool_create(ISA_POOL_SIZE, page_pool_alloc, page_pool_free, (void *) __GFP_DMA);
274 printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
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
283 static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
285 unsigned char *vfrom;
286 struct bio_vec *tovec, *fromvec;
289 __bio_for_each_segment(tovec, to, i, 0) {
290 fromvec = from->bi_io_vec + i;
295 if (tovec->bv_page == fromvec->bv_page)
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
303 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
305 flush_dcache_page(tovec->bv_page);
306 bounce_copy_vec(tovec, vfrom);
310 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err)
312 struct bio *bio_orig = bio->bi_private;
313 struct bio_vec *bvec, *org_vec;
316 if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags))
317 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags);
320 * free up bounce indirect pages used
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)
327 mempool_free(bvec->bv_page, pool);
328 dec_page_state(nr_bounce);
331 bio_endio(bio_orig, bio_orig->bi_size, err);
335 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done,int err)
340 bounce_end_io(bio, page_pool, err);
344 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err)
349 bounce_end_io(bio, isa_page_pool, err);
353 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err)
355 struct bio *bio_orig = bio->bi_private;
357 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
358 copy_to_high_bio_irq(bio_orig, bio);
360 bounce_end_io(bio, pool, err);
363 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
368 __bounce_end_io_read(bio, page_pool, err);
372 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err)
377 __bounce_end_io_read(bio, isa_page_pool, err);
381 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig,
385 struct bio *bio = NULL;
386 int i, rw = bio_data_dir(*bio_orig);
387 struct bio_vec *to, *from;
389 bio_for_each_segment(from, *bio_orig, i) {
390 page = from->bv_page;
393 * is destination page below bounce pfn?
395 if (page_to_pfn(page) < q->bounce_pfn)
402 bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt);
404 to = bio->bi_io_vec + i;
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);
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);
429 * at least one page was bounced, fill in possible non-highmem
432 __bio_for_each_segment(from, *bio_orig, i, 0) {
433 to = bio_iovec_idx(bio, i);
435 to->bv_page = from->bv_page;
436 to->bv_len = from->bv_len;
437 to->bv_offset = from->bv_offset;
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;
446 bio->bi_vcnt = (*bio_orig)->bi_vcnt;
447 bio->bi_idx = (*bio_orig)->bi_idx;
448 bio->bi_size = (*bio_orig)->bi_size;
450 if (pool == page_pool) {
451 bio->bi_end_io = bounce_end_io_write;
453 bio->bi_end_io = bounce_end_io_read;
455 bio->bi_end_io = bounce_end_io_write_isa;
457 bio->bi_end_io = bounce_end_io_read_isa;
460 bio->bi_private = *bio_orig;
464 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
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
473 if (!(q->bounce_gfp & GFP_DMA)) {
474 if (q->bounce_pfn >= blk_max_pfn)
478 BUG_ON(!isa_page_pool);
479 pool = isa_page_pool;
485 __blk_queue_bounce(q, bio_orig, pool);
488 EXPORT_SYMBOL(blk_queue_bounce);
490 #if defined(HASHED_PAGE_VIRTUAL)
492 #define PA_HASH_ORDER 7
495 * Describes one page->virtual association
497 struct page_address_map {
500 struct list_head list;
504 * page_address_map freelist, allocated from page_address_maps.
506 static struct list_head page_address_pool; /* freelist */
507 static spinlock_t pool_lock; /* protects page_address_pool */
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];
517 static struct page_address_slot *page_slot(struct page *page)
519 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
522 void *page_address(struct page *page)
526 struct page_address_slot *pas;
528 if (!PageHighMem(page))
529 return lowmem_page_address(page);
531 pas = page_slot(page);
533 spin_lock_irqsave(&pas->lock, flags);
534 if (!list_empty(&pas->lh)) {
535 struct page_address_map *pam;
537 list_for_each_entry(pam, &pas->lh, list) {
538 if (pam->page == page) {
545 spin_unlock_irqrestore(&pas->lock, flags);
549 EXPORT_SYMBOL(page_address);
551 void set_page_address(struct page *page, void *virtual)
554 struct page_address_slot *pas;
555 struct page_address_map *pam;
557 BUG_ON(!PageHighMem(page));
559 pas = page_slot(page);
560 if (virtual) { /* Add */
561 BUG_ON(list_empty(&page_address_pool));
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);
570 pam->virtual = virtual;
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);
587 spin_unlock_irqrestore(&pas->lock, flags);
593 static struct page_address_map page_address_maps[LAST_PKMAP];
595 void __init page_address_init(void)
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);
606 spin_lock_init(&pool_lock);
609 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */