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/export.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 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
34 DEFINE_PER_CPU(int, __kmap_atomic_idx
);
38 * Virtual_count is not a pure "count".
39 * 0 means that it is not mapped, and has not been mapped
40 * since a TLB flush - it is usable.
41 * 1 means that there are no users, but it has been mapped
42 * since the last TLB flush - so we can't use it.
43 * n means that there are (n-1) current users of it.
47 unsigned long totalhigh_pages __read_mostly
;
48 EXPORT_SYMBOL(totalhigh_pages
);
51 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx
);
53 unsigned int nr_free_highpages (void)
56 unsigned int pages
= 0;
58 for_each_online_pgdat(pgdat
) {
59 pages
+= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
61 if (zone_movable_is_highmem())
62 pages
+= zone_page_state(
63 &pgdat
->node_zones
[ZONE_MOVABLE
],
70 static int pkmap_count
[LAST_PKMAP
];
71 static unsigned int last_pkmap_nr
;
72 static __cacheline_aligned_in_smp
DEFINE_SPINLOCK(kmap_lock
);
74 pte_t
* pkmap_page_table
;
76 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait
);
79 * Most architectures have no use for kmap_high_get(), so let's abstract
80 * the disabling of IRQ out of the locking in that case to save on a
81 * potential useless overhead.
83 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
84 #define lock_kmap() spin_lock_irq(&kmap_lock)
85 #define unlock_kmap() spin_unlock_irq(&kmap_lock)
86 #define lock_kmap_any(flags) spin_lock_irqsave(&kmap_lock, flags)
87 #define unlock_kmap_any(flags) spin_unlock_irqrestore(&kmap_lock, flags)
89 #define lock_kmap() spin_lock(&kmap_lock)
90 #define unlock_kmap() spin_unlock(&kmap_lock)
91 #define lock_kmap_any(flags) \
92 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
93 #define unlock_kmap_any(flags) \
94 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
97 static void flush_all_zero_pkmaps(void)
104 for (i
= 0; i
< LAST_PKMAP
; i
++) {
108 * zero means we don't have anything to do,
109 * >1 means that it is still in use. Only
110 * a count of 1 means that it is free but
111 * needs to be unmapped
113 if (pkmap_count
[i
] != 1)
118 BUG_ON(pte_none(pkmap_page_table
[i
]));
121 * Don't need an atomic fetch-and-clear op here;
122 * no-one has the page mapped, and cannot get at
123 * its virtual address (and hence PTE) without first
124 * getting the kmap_lock (which is held here).
125 * So no dangers, even with speculative execution.
127 page
= pte_page(pkmap_page_table
[i
]);
128 pte_clear(&init_mm
, (unsigned long)page_address(page
),
129 &pkmap_page_table
[i
]);
131 set_page_address(page
, NULL
);
135 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP
));
139 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
141 void kmap_flush_unused(void)
144 flush_all_zero_pkmaps();
148 static inline unsigned long map_new_virtual(struct page
*page
)
155 /* Find an empty entry */
157 last_pkmap_nr
= (last_pkmap_nr
+ 1) & LAST_PKMAP_MASK
;
158 if (!last_pkmap_nr
) {
159 flush_all_zero_pkmaps();
162 if (!pkmap_count
[last_pkmap_nr
])
163 break; /* Found a usable entry */
168 * Sleep for somebody else to unmap their entries
171 DECLARE_WAITQUEUE(wait
, current
);
173 __set_current_state(TASK_UNINTERRUPTIBLE
);
174 add_wait_queue(&pkmap_map_wait
, &wait
);
177 remove_wait_queue(&pkmap_map_wait
, &wait
);
180 /* Somebody else might have mapped it while we slept */
181 if (page_address(page
))
182 return (unsigned long)page_address(page
);
188 vaddr
= PKMAP_ADDR(last_pkmap_nr
);
189 set_pte_at(&init_mm
, vaddr
,
190 &(pkmap_page_table
[last_pkmap_nr
]), mk_pte(page
, kmap_prot
));
192 pkmap_count
[last_pkmap_nr
] = 1;
193 set_page_address(page
, (void *)vaddr
);
199 * kmap_high - map a highmem page into memory
200 * @page: &struct page to map
202 * Returns the page's virtual memory address.
204 * We cannot call this from interrupts, as it may block.
206 void *kmap_high(struct page
*page
)
211 * For highmem pages, we can't trust "virtual" until
212 * after we have the lock.
215 vaddr
= (unsigned long)page_address(page
);
217 vaddr
= map_new_virtual(page
);
218 pkmap_count
[PKMAP_NR(vaddr
)]++;
219 BUG_ON(pkmap_count
[PKMAP_NR(vaddr
)] < 2);
221 return (void*) vaddr
;
224 EXPORT_SYMBOL(kmap_high
);
226 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
228 * kmap_high_get - pin a highmem page into memory
229 * @page: &struct page to pin
231 * Returns the page's current virtual memory address, or NULL if no mapping
232 * exists. If and only if a non null address is returned then a
233 * matching call to kunmap_high() is necessary.
235 * This can be called from any context.
237 void *kmap_high_get(struct page
*page
)
239 unsigned long vaddr
, flags
;
241 lock_kmap_any(flags
);
242 vaddr
= (unsigned long)page_address(page
);
244 BUG_ON(pkmap_count
[PKMAP_NR(vaddr
)] < 1);
245 pkmap_count
[PKMAP_NR(vaddr
)]++;
247 unlock_kmap_any(flags
);
248 return (void*) vaddr
;
250 EXPORT_SYMBOL(kmap_high_get
);
254 * kunmap_high - unmap a highmem page into memory
255 * @page: &struct page to unmap
257 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
258 * only from user context.
260 void kunmap_high(struct page
*page
)
267 lock_kmap_any(flags
);
268 vaddr
= (unsigned long)page_address(page
);
270 nr
= PKMAP_NR(vaddr
);
273 * A count must never go down to zero
274 * without a TLB flush!
277 switch (--pkmap_count
[nr
]) {
282 * Avoid an unnecessary wake_up() function call.
283 * The common case is pkmap_count[] == 1, but
285 * The tasks queued in the wait-queue are guarded
286 * by both the lock in the wait-queue-head and by
287 * the kmap_lock. As the kmap_lock is held here,
288 * no need for the wait-queue-head's lock. Simply
289 * test if the queue is empty.
291 need_wakeup
= waitqueue_active(&pkmap_map_wait
);
293 unlock_kmap_any(flags
);
295 /* do wake-up, if needed, race-free outside of the spin lock */
297 wake_up(&pkmap_map_wait
);
300 EXPORT_SYMBOL(kunmap_high
);
303 #if defined(HASHED_PAGE_VIRTUAL)
305 #define PA_HASH_ORDER 7
308 * Describes one page->virtual association
310 struct page_address_map
{
313 struct list_head list
;
317 * page_address_map freelist, allocated from page_address_maps.
319 static struct list_head page_address_pool
; /* freelist */
320 static spinlock_t pool_lock
; /* protects page_address_pool */
325 static struct page_address_slot
{
326 struct list_head lh
; /* List of page_address_maps */
327 spinlock_t lock
; /* Protect this bucket's list */
328 } ____cacheline_aligned_in_smp page_address_htable
[1<<PA_HASH_ORDER
];
330 static struct page_address_slot
*page_slot(const struct page
*page
)
332 return &page_address_htable
[hash_ptr(page
, PA_HASH_ORDER
)];
336 * page_address - get the mapped virtual address of a page
337 * @page: &struct page to get the virtual address of
339 * Returns the page's virtual address.
341 void *page_address(const struct page
*page
)
345 struct page_address_slot
*pas
;
347 if (!PageHighMem(page
))
348 return lowmem_page_address(page
);
350 pas
= page_slot(page
);
352 spin_lock_irqsave(&pas
->lock
, flags
);
353 if (!list_empty(&pas
->lh
)) {
354 struct page_address_map
*pam
;
356 list_for_each_entry(pam
, &pas
->lh
, list
) {
357 if (pam
->page
== page
) {
364 spin_unlock_irqrestore(&pas
->lock
, flags
);
368 EXPORT_SYMBOL(page_address
);
371 * set_page_address - set a page's virtual address
372 * @page: &struct page to set
373 * @virtual: virtual address to use
375 void set_page_address(struct page
*page
, void *virtual)
378 struct page_address_slot
*pas
;
379 struct page_address_map
*pam
;
381 BUG_ON(!PageHighMem(page
));
383 pas
= page_slot(page
);
384 if (virtual) { /* Add */
385 BUG_ON(list_empty(&page_address_pool
));
387 spin_lock_irqsave(&pool_lock
, flags
);
388 pam
= list_entry(page_address_pool
.next
,
389 struct page_address_map
, list
);
390 list_del(&pam
->list
);
391 spin_unlock_irqrestore(&pool_lock
, flags
);
394 pam
->virtual = virtual;
396 spin_lock_irqsave(&pas
->lock
, flags
);
397 list_add_tail(&pam
->list
, &pas
->lh
);
398 spin_unlock_irqrestore(&pas
->lock
, flags
);
399 } else { /* Remove */
400 spin_lock_irqsave(&pas
->lock
, flags
);
401 list_for_each_entry(pam
, &pas
->lh
, list
) {
402 if (pam
->page
== page
) {
403 list_del(&pam
->list
);
404 spin_unlock_irqrestore(&pas
->lock
, flags
);
405 spin_lock_irqsave(&pool_lock
, flags
);
406 list_add_tail(&pam
->list
, &page_address_pool
);
407 spin_unlock_irqrestore(&pool_lock
, flags
);
411 spin_unlock_irqrestore(&pas
->lock
, flags
);
417 static struct page_address_map page_address_maps
[LAST_PKMAP
];
419 void __init
page_address_init(void)
423 INIT_LIST_HEAD(&page_address_pool
);
424 for (i
= 0; i
< ARRAY_SIZE(page_address_maps
); i
++)
425 list_add(&page_address_maps
[i
].list
, &page_address_pool
);
426 for (i
= 0; i
< ARRAY_SIZE(page_address_htable
); i
++) {
427 INIT_LIST_HEAD(&page_address_htable
[i
].lh
);
428 spin_lock_init(&page_address_htable
[i
].lock
);
430 spin_lock_init(&pool_lock
);
433 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */