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>
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.
43 unsigned long totalhigh_pages __read_mostly
;
45 unsigned int nr_free_highpages (void)
48 unsigned int pages
= 0;
50 for_each_online_pgdat(pgdat
)
51 pages
+= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
57 static int pkmap_count
[LAST_PKMAP
];
58 static unsigned int last_pkmap_nr
;
59 static __cacheline_aligned_in_smp
DEFINE_SPINLOCK(kmap_lock
);
61 pte_t
* pkmap_page_table
;
63 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait
);
65 static void flush_all_zero_pkmaps(void)
71 for (i
= 0; i
< LAST_PKMAP
; i
++) {
75 * zero means we don't have anything to do,
76 * >1 means that it is still in use. Only
77 * a count of 1 means that it is free but
78 * needs to be unmapped
80 if (pkmap_count
[i
] != 1)
85 BUG_ON(pte_none(pkmap_page_table
[i
]));
88 * Don't need an atomic fetch-and-clear op here;
89 * no-one has the page mapped, and cannot get at
90 * its virtual address (and hence PTE) without first
91 * getting the kmap_lock (which is held here).
92 * So no dangers, even with speculative execution.
94 page
= pte_page(pkmap_page_table
[i
]);
95 pte_clear(&init_mm
, (unsigned long)page_address(page
),
96 &pkmap_page_table
[i
]);
98 set_page_address(page
, NULL
);
100 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP
));
103 /* Flush all unused kmap mappings in order to remove stray
105 void kmap_flush_unused(void)
107 spin_lock(&kmap_lock
);
108 flush_all_zero_pkmaps();
109 spin_unlock(&kmap_lock
);
112 static inline unsigned long map_new_virtual(struct page
*page
)
119 /* Find an empty entry */
121 last_pkmap_nr
= (last_pkmap_nr
+ 1) & LAST_PKMAP_MASK
;
122 if (!last_pkmap_nr
) {
123 flush_all_zero_pkmaps();
126 if (!pkmap_count
[last_pkmap_nr
]) {
127 if (cpu_has_dc_aliases
) {
128 unsigned int pfn
, map_pfn
;
130 /* check page color */
131 pfn
= page_to_pfn(page
);
132 map_pfn
= PKMAP_ADDR(last_pkmap_nr
) >> PAGE_SHIFT
;
134 /* Avoide possibility of cache Aliasing */
135 if (!pages_do_alias((map_pfn
<< PAGE_SHIFT
), (pfn
<< PAGE_SHIFT
)))
136 break; /* Found a usable entry */
138 break; /* Found a usable entry */
144 * Sleep for somebody else to unmap their entries
147 DECLARE_WAITQUEUE(wait
, current
);
149 __set_current_state(TASK_UNINTERRUPTIBLE
);
150 add_wait_queue(&pkmap_map_wait
, &wait
);
151 spin_unlock(&kmap_lock
);
153 remove_wait_queue(&pkmap_map_wait
, &wait
);
154 spin_lock(&kmap_lock
);
156 /* Somebody else might have mapped it while we slept */
157 if (page_address(page
))
158 return (unsigned long)page_address(page
);
164 vaddr
= PKMAP_ADDR(last_pkmap_nr
);
165 set_pte_at(&init_mm
, vaddr
,
166 &(pkmap_page_table
[last_pkmap_nr
]), mk_pte(page
, kmap_prot
));
168 pkmap_count
[last_pkmap_nr
] = 1;
169 set_page_address(page
, (void *)vaddr
);
174 void fastcall
*kmap_high(struct page
*page
)
179 * For highmem pages, we can't trust "virtual" until
180 * after we have the lock.
182 * We cannot call this from interrupts, as it may block
184 spin_lock(&kmap_lock
);
185 vaddr
= (unsigned long)page_address(page
);
187 vaddr
= map_new_virtual(page
);
188 pkmap_count
[PKMAP_NR(vaddr
)]++;
189 BUG_ON(pkmap_count
[PKMAP_NR(vaddr
)] < 2);
190 spin_unlock(&kmap_lock
);
191 return (void*) vaddr
;
194 EXPORT_SYMBOL(kmap_high
);
196 void fastcall
kunmap_high(struct page
*page
)
202 spin_lock(&kmap_lock
);
203 vaddr
= (unsigned long)page_address(page
);
205 nr
= PKMAP_NR(vaddr
);
208 * A count must never go down to zero
209 * without a TLB flush!
212 switch (--pkmap_count
[nr
]) {
217 * Avoid an unnecessary wake_up() function call.
218 * The common case is pkmap_count[] == 1, but
220 * The tasks queued in the wait-queue are guarded
221 * by both the lock in the wait-queue-head and by
222 * the kmap_lock. As the kmap_lock is held here,
223 * no need for the wait-queue-head's lock. Simply
224 * test if the queue is empty.
226 need_wakeup
= waitqueue_active(&pkmap_map_wait
);
228 spin_unlock(&kmap_lock
);
230 /* do wake-up, if needed, race-free outside of the spin lock */
232 wake_up(&pkmap_map_wait
);
235 EXPORT_SYMBOL(kunmap_high
);
238 #if defined(HASHED_PAGE_VIRTUAL)
240 #define PA_HASH_ORDER 7
243 * Describes one page->virtual association
245 struct page_address_map
{
248 struct list_head list
;
252 * page_address_map freelist, allocated from page_address_maps.
254 static struct list_head page_address_pool
; /* freelist */
255 static spinlock_t pool_lock
; /* protects page_address_pool */
260 static struct page_address_slot
{
261 struct list_head lh
; /* List of page_address_maps */
262 spinlock_t lock
; /* Protect this bucket's list */
263 } ____cacheline_aligned_in_smp page_address_htable
[1<<PA_HASH_ORDER
];
265 static struct page_address_slot
*page_slot(struct page
*page
)
267 return &page_address_htable
[hash_ptr(page
, PA_HASH_ORDER
)];
270 void *page_address(struct page
*page
)
274 struct page_address_slot
*pas
;
276 if (!PageHighMem(page
))
277 return lowmem_page_address(page
);
279 pas
= page_slot(page
);
281 spin_lock_irqsave(&pas
->lock
, flags
);
282 if (!list_empty(&pas
->lh
)) {
283 struct page_address_map
*pam
;
285 list_for_each_entry(pam
, &pas
->lh
, list
) {
286 if (pam
->page
== page
) {
293 spin_unlock_irqrestore(&pas
->lock
, flags
);
297 EXPORT_SYMBOL(page_address
);
299 void set_page_address(struct page
*page
, void *virtual)
302 struct page_address_slot
*pas
;
303 struct page_address_map
*pam
;
305 BUG_ON(!PageHighMem(page
));
307 pas
= page_slot(page
);
308 if (virtual) { /* Add */
309 BUG_ON(list_empty(&page_address_pool
));
311 spin_lock_irqsave(&pool_lock
, flags
);
312 pam
= list_entry(page_address_pool
.next
,
313 struct page_address_map
, list
);
314 list_del(&pam
->list
);
315 spin_unlock_irqrestore(&pool_lock
, flags
);
318 pam
->virtual = virtual;
320 spin_lock_irqsave(&pas
->lock
, flags
);
321 list_add_tail(&pam
->list
, &pas
->lh
);
322 spin_unlock_irqrestore(&pas
->lock
, flags
);
323 } else { /* Remove */
324 spin_lock_irqsave(&pas
->lock
, flags
);
325 list_for_each_entry(pam
, &pas
->lh
, list
) {
326 if (pam
->page
== page
) {
327 list_del(&pam
->list
);
328 spin_unlock_irqrestore(&pas
->lock
, flags
);
329 spin_lock_irqsave(&pool_lock
, flags
);
330 list_add_tail(&pam
->list
, &page_address_pool
);
331 spin_unlock_irqrestore(&pool_lock
, flags
);
335 spin_unlock_irqrestore(&pas
->lock
, flags
);
341 static struct page_address_map page_address_maps
[LAST_PKMAP
];
343 void __init
page_address_init(void)
347 INIT_LIST_HEAD(&page_address_pool
);
348 for (i
= 0; i
< ARRAY_SIZE(page_address_maps
); i
++)
349 list_add(&page_address_maps
[i
].list
, &page_address_pool
);
350 for (i
= 0; i
< ARRAY_SIZE(page_address_htable
); i
++) {
351 INIT_LIST_HEAD(&page_address_htable
[i
].lh
);
352 spin_lock_init(&page_address_htable
[i
].lock
);
354 spin_lock_init(&pool_lock
);
357 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */