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/bio.h>
21 #include <linux/pagemap.h>
22 #include <linux/mempool.h>
23 #include <linux/blkdev.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <asm/pgalloc.h>
27 #include <asm/tlbflush.h>
29 static mempool_t
*page_pool
, *isa_page_pool
;
31 static void *page_pool_alloc(int gfp_mask
, void *data
)
33 int gfp
= gfp_mask
| (int) (long) data
;
35 return alloc_page(gfp
);
38 static void page_pool_free(void *page
, void *data
)
44 * Virtual_count is not a pure "count".
45 * 0 means that it is not mapped, and has not been mapped
46 * since a TLB flush - it is usable.
47 * 1 means that there are no users, but it has been mapped
48 * since the last TLB flush - so we can't use it.
49 * n means that there are (n-1) current users of it.
52 static int pkmap_count
[LAST_PKMAP
];
53 static unsigned int last_pkmap_nr
;
54 static spinlock_t kmap_lock __cacheline_aligned_in_smp
= SPIN_LOCK_UNLOCKED
;
56 pte_t
* pkmap_page_table
;
58 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait
);
60 static void flush_all_zero_pkmaps(void)
66 for (i
= 0; i
< LAST_PKMAP
; i
++) {
70 * zero means we don't have anything to do,
71 * >1 means that it is still in use. Only
72 * a count of 1 means that it is free but
73 * needs to be unmapped
75 if (pkmap_count
[i
] != 1)
80 if (pte_none(pkmap_page_table
[i
]))
84 * Don't need an atomic fetch-and-clear op here;
85 * no-one has the page mapped, and cannot get at
86 * its virtual address (and hence PTE) without first
87 * getting the kmap_lock (which is held here).
88 * So no dangers, even with speculative execution.
90 page
= pte_page(pkmap_page_table
[i
]);
91 pte_clear(&pkmap_page_table
[i
]);
93 set_page_address(page
, NULL
);
95 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP
));
98 static inline unsigned long map_new_virtual(struct page
*page
)
105 /* Find an empty entry */
107 last_pkmap_nr
= (last_pkmap_nr
+ 1) & LAST_PKMAP_MASK
;
108 if (!last_pkmap_nr
) {
109 flush_all_zero_pkmaps();
112 if (!pkmap_count
[last_pkmap_nr
])
113 break; /* Found a usable entry */
118 * Sleep for somebody else to unmap their entries
121 DECLARE_WAITQUEUE(wait
, current
);
123 current
->state
= TASK_UNINTERRUPTIBLE
;
124 add_wait_queue(&pkmap_map_wait
, &wait
);
125 spin_unlock(&kmap_lock
);
127 remove_wait_queue(&pkmap_map_wait
, &wait
);
128 spin_lock(&kmap_lock
);
130 /* Somebody else might have mapped it while we slept */
131 if (page_address(page
))
132 return (unsigned long)page_address(page
);
138 vaddr
= PKMAP_ADDR(last_pkmap_nr
);
139 set_pte(&(pkmap_page_table
[last_pkmap_nr
]), mk_pte(page
, kmap_prot
));
141 pkmap_count
[last_pkmap_nr
] = 1;
142 set_page_address(page
, (void *)vaddr
);
147 void *kmap_high(struct page
*page
)
152 * For highmem pages, we can't trust "virtual" until
153 * after we have the lock.
155 * We cannot call this from interrupts, as it may block
157 spin_lock(&kmap_lock
);
158 vaddr
= (unsigned long)page_address(page
);
160 vaddr
= map_new_virtual(page
);
161 pkmap_count
[PKMAP_NR(vaddr
)]++;
162 if (pkmap_count
[PKMAP_NR(vaddr
)] < 2)
164 spin_unlock(&kmap_lock
);
165 return (void*) vaddr
;
168 void kunmap_high(struct page
*page
)
174 spin_lock(&kmap_lock
);
175 vaddr
= (unsigned long)page_address(page
);
178 nr
= PKMAP_NR(vaddr
);
181 * A count must never go down to zero
182 * without a TLB flush!
185 switch (--pkmap_count
[nr
]) {
190 * Avoid an unnecessary wake_up() function call.
191 * The common case is pkmap_count[] == 1, but
193 * The tasks queued in the wait-queue are guarded
194 * by both the lock in the wait-queue-head and by
195 * the kmap_lock. As the kmap_lock is held here,
196 * no need for the wait-queue-head's lock. Simply
197 * test if the queue is empty.
199 need_wakeup
= waitqueue_active(&pkmap_map_wait
);
201 spin_unlock(&kmap_lock
);
203 /* do wake-up, if needed, race-free outside of the spin lock */
205 wake_up(&pkmap_map_wait
);
210 static __init
int init_emergency_pool(void)
219 page_pool
= mempool_create(POOL_SIZE
, page_pool_alloc
, page_pool_free
, NULL
);
222 printk("highmem bounce pool size: %d pages\n", POOL_SIZE
);
227 __initcall(init_emergency_pool
);
230 * highmem version, map in to vec
232 static void bounce_copy_vec(struct bio_vec
*to
, unsigned char *vfrom
)
237 local_irq_save(flags
);
238 vto
= kmap_atomic(to
->bv_page
, KM_BOUNCE_READ
);
239 memcpy(vto
+ to
->bv_offset
, vfrom
, to
->bv_len
);
240 kunmap_atomic(vto
, KM_BOUNCE_READ
);
241 local_irq_restore(flags
);
244 #else /* CONFIG_HIGHMEM */
246 #define bounce_copy_vec(to, vfrom) \
247 memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
251 #define ISA_POOL_SIZE 16
254 * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA
255 * as the max address, so check if the pool has already been created.
257 int init_emergency_isa_pool(void)
262 isa_page_pool
= mempool_create(ISA_POOL_SIZE
, page_pool_alloc
, page_pool_free
, (void *) __GFP_DMA
);
266 printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE
);
271 * Simple bounce buffer support for highmem pages. Depending on the
272 * queue gfp mask set, *to may or may not be a highmem page. kmap it
273 * always, it will do the Right Thing
275 static void copy_to_high_bio_irq(struct bio
*to
, struct bio
*from
)
277 unsigned char *vfrom
;
278 struct bio_vec
*tovec
, *fromvec
;
281 __bio_for_each_segment(tovec
, to
, i
, 0) {
282 fromvec
= from
->bi_io_vec
+ i
;
287 if (tovec
->bv_page
== fromvec
->bv_page
)
290 vfrom
= page_address(fromvec
->bv_page
) + fromvec
->bv_offset
;
292 bounce_copy_vec(tovec
, vfrom
);
296 static void bounce_end_io(struct bio
*bio
, mempool_t
*pool
)
298 struct bio
*bio_orig
= bio
->bi_private
;
299 struct bio_vec
*bvec
, *org_vec
;
302 if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
305 set_bit(BIO_UPTODATE
, &bio_orig
->bi_flags
);
308 * free up bounce indirect pages used
310 __bio_for_each_segment(bvec
, bio
, i
, 0) {
311 org_vec
= bio_orig
->bi_io_vec
+ i
;
312 if (bvec
->bv_page
== org_vec
->bv_page
)
315 mempool_free(bvec
->bv_page
, pool
);
319 bio_endio(bio_orig
, bio_orig
->bi_size
, 0);
323 static int bounce_end_io_write(struct bio
*bio
, unsigned int bytes_done
,int err
)
328 bounce_end_io(bio
, page_pool
);
332 static int bounce_end_io_write_isa(struct bio
*bio
, unsigned int bytes_done
, int err
)
337 bounce_end_io(bio
, isa_page_pool
);
341 static void __bounce_end_io_read(struct bio
*bio
, mempool_t
*pool
)
343 struct bio
*bio_orig
= bio
->bi_private
;
345 if (test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
346 copy_to_high_bio_irq(bio_orig
, bio
);
348 bounce_end_io(bio
, pool
);
351 static int bounce_end_io_read(struct bio
*bio
, unsigned int bytes_done
, int err
)
356 __bounce_end_io_read(bio
, page_pool
);
360 static int bounce_end_io_read_isa(struct bio
*bio
, unsigned int bytes_done
, int err
)
365 __bounce_end_io_read(bio
, isa_page_pool
);
369 void __blk_queue_bounce(request_queue_t
*q
, struct bio
**bio_orig
, int bio_gfp
,
373 struct bio
*bio
= NULL
;
374 int i
, rw
= bio_data_dir(*bio_orig
);
375 struct bio_vec
*to
, *from
;
377 bio_for_each_segment(from
, *bio_orig
, i
) {
378 page
= from
->bv_page
;
381 * is destination page below bounce pfn?
383 if ((page
- page_zone(page
)->zone_mem_map
) + (page_zone(page
)->zone_start_pfn
) < q
->bounce_pfn
)
390 bio
= bio_alloc(bio_gfp
, (*bio_orig
)->bi_vcnt
);
392 to
= bio
->bi_io_vec
+ i
;
394 to
->bv_page
= mempool_alloc(pool
, q
->bounce_gfp
);
395 to
->bv_len
= from
->bv_len
;
396 to
->bv_offset
= from
->bv_offset
;
401 vto
= page_address(to
->bv_page
) + to
->bv_offset
;
402 vfrom
= kmap(from
->bv_page
) + from
->bv_offset
;
403 memcpy(vto
, vfrom
, to
->bv_len
);
404 kunmap(from
->bv_page
);
415 * at least one page was bounced, fill in possible non-highmem
418 bio_for_each_segment(from
, *bio_orig
, i
) {
419 to
= bio_iovec_idx(bio
, i
);
421 to
->bv_page
= from
->bv_page
;
422 to
->bv_len
= from
->bv_len
;
423 to
->bv_offset
= from
->bv_offset
;
427 bio
->bi_bdev
= (*bio_orig
)->bi_bdev
;
428 bio
->bi_flags
|= (1 << BIO_BOUNCED
);
429 bio
->bi_sector
= (*bio_orig
)->bi_sector
;
430 bio
->bi_rw
= (*bio_orig
)->bi_rw
;
432 bio
->bi_vcnt
= (*bio_orig
)->bi_vcnt
;
434 bio
->bi_size
= (*bio_orig
)->bi_size
;
436 if (pool
== page_pool
) {
437 bio
->bi_end_io
= bounce_end_io_write
;
439 bio
->bi_end_io
= bounce_end_io_read
;
441 bio
->bi_end_io
= bounce_end_io_write_isa
;
443 bio
->bi_end_io
= bounce_end_io_read_isa
;
446 bio
->bi_private
= *bio_orig
;
450 inline void blk_queue_bounce(request_queue_t
*q
, struct bio
**bio_orig
)
456 * for non-isa bounce case, just check if the bounce pfn is equal
457 * to or bigger than the highest pfn in the system -- in that case,
458 * don't waste time iterating over bio segments
460 if (!(q
->bounce_gfp
& GFP_DMA
)) {
461 if (q
->bounce_pfn
>= blk_max_pfn
)
464 bio_gfp
= GFP_NOHIGHIO
;
467 BUG_ON(!isa_page_pool
);
470 pool
= isa_page_pool
;
476 __blk_queue_bounce(q
, bio_orig
, bio_gfp
, pool
);
479 #if defined(CONFIG_DEBUG_HIGHMEM) && defined(CONFIG_HIGHMEM)
480 void check_highmem_ptes(void)
485 for (type
= 0; type
< KM_TYPE_NR
; type
++) {
486 idx
= type
+ KM_TYPE_NR
*smp_processor_id();
487 if (!pte_none(*(kmap_pte
-idx
))) {
488 printk("scheduling with KM_TYPE %d held!\n", type
);
496 #if defined(HASHED_PAGE_VIRTUAL)
498 #define PA_HASH_ORDER 7
501 * Describes one page->virtual association
503 struct page_address_map
{
506 struct list_head list
;
510 * page_address_map freelist, allocated from page_address_maps.
512 static struct list_head page_address_pool
; /* freelist */
513 static spinlock_t pool_lock
; /* protects page_address_pool */
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
];
523 static struct page_address_slot
*page_slot(struct page
*page
)
525 return &page_address_htable
[hash_ptr(page
, PA_HASH_ORDER
)];
528 void *page_address(struct page
*page
)
532 struct page_address_slot
*pas
;
534 if (!PageHighMem(page
))
535 return lowmem_page_address(page
);
537 pas
= page_slot(page
);
539 spin_lock_irqsave(&pas
->lock
, flags
);
540 if (!list_empty(&pas
->lh
)) {
541 struct page_address_map
*pam
;
543 list_for_each_entry(pam
, &pas
->lh
, list
) {
544 if (pam
->page
== page
) {
551 spin_unlock_irqrestore(&pas
->lock
, flags
);
555 void set_page_address(struct page
*page
, void *virtual)
558 struct page_address_slot
*pas
;
559 struct page_address_map
*pam
;
561 BUG_ON(!PageHighMem(page
));
563 pas
= page_slot(page
);
564 if (virtual) { /* Add */
565 BUG_ON(list_empty(&page_address_pool
));
567 spin_lock_irqsave(&pool_lock
, flags
);
568 pam
= list_entry(page_address_pool
.next
,
569 struct page_address_map
, list
);
570 list_del(&pam
->list
);
571 spin_unlock_irqrestore(&pool_lock
, flags
);
574 pam
->virtual = virtual;
576 spin_lock_irqsave(&pas
->lock
, flags
);
577 list_add_tail(&pam
->list
, &pas
->lh
);
578 spin_unlock_irqrestore(&pas
->lock
, flags
);
579 } else { /* Remove */
580 spin_lock_irqsave(&pas
->lock
, flags
);
581 list_for_each_entry(pam
, &pas
->lh
, list
) {
582 if (pam
->page
== page
) {
583 list_del(&pam
->list
);
584 spin_unlock_irqrestore(&pas
->lock
, flags
);
585 spin_lock_irqsave(&pool_lock
, flags
);
586 list_add_tail(&pam
->list
, &page_address_pool
);
587 spin_unlock_irqrestore(&pool_lock
, flags
);
591 spin_unlock_irqrestore(&pas
->lock
, flags
);
597 static struct page_address_map page_address_maps
[LAST_PKMAP
];
599 void __init
page_address_init(void)
603 INIT_LIST_HEAD(&page_address_pool
);
604 for (i
= 0; i
< ARRAY_SIZE(page_address_maps
); i
++)
605 list_add(&page_address_maps
[i
].list
, &page_address_pool
);
606 for (i
= 0; i
< ARRAY_SIZE(page_address_htable
); i
++) {
607 INIT_LIST_HEAD(&page_address_htable
[i
].lh
);
608 spin_lock_init(&page_address_htable
[i
].lock
);
610 spin_lock_init(&pool_lock
);
613 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */