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/kgdb.h>
30 #include <asm/tlbflush.h>
33 * Virtual_count is not a pure "count".
34 * 0 means that it is not mapped, and has not been mapped
35 * since a TLB flush - it is usable.
36 * 1 means that there are no users, but it has been mapped
37 * since the last TLB flush - so we can't use it.
38 * n means that there are (n-1) current users of it.
42 unsigned long totalhigh_pages __read_mostly
;
43 EXPORT_SYMBOL(totalhigh_pages
);
46 DEFINE_PER_CPU(int, __kmap_atomic_idx
);
47 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx
);
49 unsigned int nr_free_highpages (void)
52 unsigned int pages
= 0;
54 for_each_online_pgdat(pgdat
) {
55 pages
+= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
57 if (zone_movable_is_highmem())
58 pages
+= zone_page_state(
59 &pgdat
->node_zones
[ZONE_MOVABLE
],
66 static int pkmap_count
[LAST_PKMAP
];
67 static unsigned int last_pkmap_nr
;
68 static __cacheline_aligned_in_smp
DEFINE_SPINLOCK(kmap_lock
);
70 pte_t
* pkmap_page_table
;
72 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait
);
75 * Most architectures have no use for kmap_high_get(), so let's abstract
76 * the disabling of IRQ out of the locking in that case to save on a
77 * potential useless overhead.
79 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
80 #define lock_kmap() spin_lock_irq(&kmap_lock)
81 #define unlock_kmap() spin_unlock_irq(&kmap_lock)
82 #define lock_kmap_any(flags) spin_lock_irqsave(&kmap_lock, flags)
83 #define unlock_kmap_any(flags) spin_unlock_irqrestore(&kmap_lock, flags)
85 #define lock_kmap() spin_lock(&kmap_lock)
86 #define unlock_kmap() spin_unlock(&kmap_lock)
87 #define lock_kmap_any(flags) \
88 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
89 #define unlock_kmap_any(flags) \
90 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
93 static void flush_all_zero_pkmaps(void)
100 for (i
= 0; i
< LAST_PKMAP
; i
++) {
104 * zero means we don't have anything to do,
105 * >1 means that it is still in use. Only
106 * a count of 1 means that it is free but
107 * needs to be unmapped
109 if (pkmap_count
[i
] != 1)
114 BUG_ON(pte_none(pkmap_page_table
[i
]));
117 * Don't need an atomic fetch-and-clear op here;
118 * no-one has the page mapped, and cannot get at
119 * its virtual address (and hence PTE) without first
120 * getting the kmap_lock (which is held here).
121 * So no dangers, even with speculative execution.
123 page
= pte_page(pkmap_page_table
[i
]);
124 pte_clear(&init_mm
, (unsigned long)page_address(page
),
125 &pkmap_page_table
[i
]);
127 set_page_address(page
, NULL
);
131 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP
));
135 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
137 void kmap_flush_unused(void)
140 flush_all_zero_pkmaps();
144 static inline unsigned long map_new_virtual(struct page
*page
)
151 /* Find an empty entry */
153 last_pkmap_nr
= (last_pkmap_nr
+ 1) & LAST_PKMAP_MASK
;
154 if (!last_pkmap_nr
) {
155 flush_all_zero_pkmaps();
158 if (!pkmap_count
[last_pkmap_nr
])
159 break; /* Found a usable entry */
164 * Sleep for somebody else to unmap their entries
167 DECLARE_WAITQUEUE(wait
, current
);
169 __set_current_state(TASK_UNINTERRUPTIBLE
);
170 add_wait_queue(&pkmap_map_wait
, &wait
);
173 remove_wait_queue(&pkmap_map_wait
, &wait
);
176 /* Somebody else might have mapped it while we slept */
177 if (page_address(page
))
178 return (unsigned long)page_address(page
);
184 vaddr
= PKMAP_ADDR(last_pkmap_nr
);
185 set_pte_at(&init_mm
, vaddr
,
186 &(pkmap_page_table
[last_pkmap_nr
]), mk_pte(page
, kmap_prot
));
188 pkmap_count
[last_pkmap_nr
] = 1;
189 set_page_address(page
, (void *)vaddr
);
195 * kmap_high - map a highmem page into memory
196 * @page: &struct page to map
198 * Returns the page's virtual memory address.
200 * We cannot call this from interrupts, as it may block.
202 void *kmap_high(struct page
*page
)
207 * For highmem pages, we can't trust "virtual" until
208 * after we have the lock.
211 vaddr
= (unsigned long)page_address(page
);
213 vaddr
= map_new_virtual(page
);
214 pkmap_count
[PKMAP_NR(vaddr
)]++;
215 BUG_ON(pkmap_count
[PKMAP_NR(vaddr
)] < 2);
217 return (void*) vaddr
;
220 EXPORT_SYMBOL(kmap_high
);
222 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
224 * kmap_high_get - pin a highmem page into memory
225 * @page: &struct page to pin
227 * Returns the page's current virtual memory address, or NULL if no mapping
228 * exists. If and only if a non null address is returned then a
229 * matching call to kunmap_high() is necessary.
231 * This can be called from any context.
233 void *kmap_high_get(struct page
*page
)
235 unsigned long vaddr
, flags
;
237 lock_kmap_any(flags
);
238 vaddr
= (unsigned long)page_address(page
);
240 BUG_ON(pkmap_count
[PKMAP_NR(vaddr
)] < 1);
241 pkmap_count
[PKMAP_NR(vaddr
)]++;
243 unlock_kmap_any(flags
);
244 return (void*) vaddr
;
249 * kunmap_high - map a highmem page into memory
250 * @page: &struct page to unmap
252 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
253 * only from user context.
255 void kunmap_high(struct page
*page
)
262 lock_kmap_any(flags
);
263 vaddr
= (unsigned long)page_address(page
);
265 nr
= PKMAP_NR(vaddr
);
268 * A count must never go down to zero
269 * without a TLB flush!
272 switch (--pkmap_count
[nr
]) {
277 * Avoid an unnecessary wake_up() function call.
278 * The common case is pkmap_count[] == 1, but
280 * The tasks queued in the wait-queue are guarded
281 * by both the lock in the wait-queue-head and by
282 * the kmap_lock. As the kmap_lock is held here,
283 * no need for the wait-queue-head's lock. Simply
284 * test if the queue is empty.
286 need_wakeup
= waitqueue_active(&pkmap_map_wait
);
288 unlock_kmap_any(flags
);
290 /* do wake-up, if needed, race-free outside of the spin lock */
292 wake_up(&pkmap_map_wait
);
295 EXPORT_SYMBOL(kunmap_high
);
298 #if defined(HASHED_PAGE_VIRTUAL)
300 #define PA_HASH_ORDER 7
303 * Describes one page->virtual association
305 struct page_address_map
{
308 struct list_head list
;
312 * page_address_map freelist, allocated from page_address_maps.
314 static struct list_head page_address_pool
; /* freelist */
315 static spinlock_t pool_lock
; /* protects page_address_pool */
320 static struct page_address_slot
{
321 struct list_head lh
; /* List of page_address_maps */
322 spinlock_t lock
; /* Protect this bucket's list */
323 } ____cacheline_aligned_in_smp page_address_htable
[1<<PA_HASH_ORDER
];
325 static struct page_address_slot
*page_slot(struct page
*page
)
327 return &page_address_htable
[hash_ptr(page
, PA_HASH_ORDER
)];
331 * page_address - get the mapped virtual address of a page
332 * @page: &struct page to get the virtual address of
334 * Returns the page's virtual address.
336 void *page_address(struct page
*page
)
340 struct page_address_slot
*pas
;
342 if (!PageHighMem(page
))
343 return lowmem_page_address(page
);
345 pas
= page_slot(page
);
347 spin_lock_irqsave(&pas
->lock
, flags
);
348 if (!list_empty(&pas
->lh
)) {
349 struct page_address_map
*pam
;
351 list_for_each_entry(pam
, &pas
->lh
, list
) {
352 if (pam
->page
== page
) {
359 spin_unlock_irqrestore(&pas
->lock
, flags
);
363 EXPORT_SYMBOL(page_address
);
366 * set_page_address - set a page's virtual address
367 * @page: &struct page to set
368 * @virtual: virtual address to use
370 void set_page_address(struct page
*page
, void *virtual)
373 struct page_address_slot
*pas
;
374 struct page_address_map
*pam
;
376 BUG_ON(!PageHighMem(page
));
378 pas
= page_slot(page
);
379 if (virtual) { /* Add */
380 BUG_ON(list_empty(&page_address_pool
));
382 spin_lock_irqsave(&pool_lock
, flags
);
383 pam
= list_entry(page_address_pool
.next
,
384 struct page_address_map
, list
);
385 list_del(&pam
->list
);
386 spin_unlock_irqrestore(&pool_lock
, flags
);
389 pam
->virtual = virtual;
391 spin_lock_irqsave(&pas
->lock
, flags
);
392 list_add_tail(&pam
->list
, &pas
->lh
);
393 spin_unlock_irqrestore(&pas
->lock
, flags
);
394 } else { /* Remove */
395 spin_lock_irqsave(&pas
->lock
, flags
);
396 list_for_each_entry(pam
, &pas
->lh
, list
) {
397 if (pam
->page
== page
) {
398 list_del(&pam
->list
);
399 spin_unlock_irqrestore(&pas
->lock
, flags
);
400 spin_lock_irqsave(&pool_lock
, flags
);
401 list_add_tail(&pam
->list
, &page_address_pool
);
402 spin_unlock_irqrestore(&pool_lock
, flags
);
406 spin_unlock_irqrestore(&pas
->lock
, flags
);
412 static struct page_address_map page_address_maps
[LAST_PKMAP
];
414 void __init
page_address_init(void)
418 INIT_LIST_HEAD(&page_address_pool
);
419 for (i
= 0; i
< ARRAY_SIZE(page_address_maps
); i
++)
420 list_add(&page_address_maps
[i
].list
, &page_address_pool
);
421 for (i
= 0; i
< ARRAY_SIZE(page_address_htable
); i
++) {
422 INIT_LIST_HEAD(&page_address_htable
[i
].lh
);
423 spin_lock_init(&page_address_htable
[i
].lock
);
425 spin_lock_init(&pool_lock
);
428 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */