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.
42 unsigned long totalhigh_pages __read_mostly;
43 EXPORT_SYMBOL(totalhigh_pages);
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],
53 if (zone_movable_is_highmem())
54 pages += zone_page_state(
55 &pgdat->node_zones[ZONE_MOVABLE],
62 static int pkmap_count[LAST_PKMAP];
63 static unsigned int last_pkmap_nr;
64 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
66 pte_t * pkmap_page_table;
68 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
71 * Most architectures have no use for kmap_high_get(), so let's abstract
72 * the disabling of IRQ out of the locking in that case to save on a
73 * potential useless overhead.
75 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
76 #define lock_kmap() spin_lock_irq(&kmap_lock)
77 #define unlock_kmap() spin_unlock_irq(&kmap_lock)
78 #define lock_kmap_any(flags) spin_lock_irqsave(&kmap_lock, flags)
79 #define unlock_kmap_any(flags) spin_unlock_irqrestore(&kmap_lock, flags)
81 #define lock_kmap() spin_lock(&kmap_lock)
82 #define unlock_kmap() spin_unlock(&kmap_lock)
83 #define lock_kmap_any(flags) \
84 do { spin_lock(&kmap_lock); (void)(flags); } while (0)
85 #define unlock_kmap_any(flags) \
86 do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
89 static void flush_all_zero_pkmaps(void)
96 for (i = 0; i < LAST_PKMAP; i++) {
100 * zero means we don't have anything to do,
101 * >1 means that it is still in use. Only
102 * a count of 1 means that it is free but
103 * needs to be unmapped
105 if (pkmap_count[i] != 1)
110 BUG_ON(pte_none(pkmap_page_table[i]));
113 * Don't need an atomic fetch-and-clear op here;
114 * no-one has the page mapped, and cannot get at
115 * its virtual address (and hence PTE) without first
116 * getting the kmap_lock (which is held here).
117 * So no dangers, even with speculative execution.
119 page = pte_page(pkmap_page_table[i]);
120 pte_clear(&init_mm, (unsigned long)page_address(page),
121 &pkmap_page_table[i]);
123 set_page_address(page, NULL);
127 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
131 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
133 void kmap_flush_unused(void)
136 flush_all_zero_pkmaps();
140 static inline unsigned long map_new_virtual(struct page *page)
147 /* Find an empty entry */
149 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
150 if (!last_pkmap_nr) {
151 flush_all_zero_pkmaps();
154 if (!pkmap_count[last_pkmap_nr])
155 break; /* Found a usable entry */
160 * Sleep for somebody else to unmap their entries
163 DECLARE_WAITQUEUE(wait, current);
165 __set_current_state(TASK_UNINTERRUPTIBLE);
166 add_wait_queue(&pkmap_map_wait, &wait);
169 remove_wait_queue(&pkmap_map_wait, &wait);
172 /* Somebody else might have mapped it while we slept */
173 if (page_address(page))
174 return (unsigned long)page_address(page);
180 vaddr = PKMAP_ADDR(last_pkmap_nr);
181 set_pte_at(&init_mm, vaddr,
182 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
184 pkmap_count[last_pkmap_nr] = 1;
185 set_page_address(page, (void *)vaddr);
191 * kmap_high - map a highmem page into memory
192 * @page: &struct page to map
194 * Returns the page's virtual memory address.
196 * We cannot call this from interrupts, as it may block.
198 void *kmap_high(struct page *page)
203 * For highmem pages, we can't trust "virtual" until
204 * after we have the lock.
207 vaddr = (unsigned long)page_address(page);
209 vaddr = map_new_virtual(page);
210 pkmap_count[PKMAP_NR(vaddr)]++;
211 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
213 return (void*) vaddr;
216 EXPORT_SYMBOL(kmap_high);
218 #ifdef ARCH_NEEDS_KMAP_HIGH_GET
220 * kmap_high_get - pin a highmem page into memory
221 * @page: &struct page to pin
223 * Returns the page's current virtual memory address, or NULL if no mapping
224 * exists. When and only when a non null address is returned then a
225 * matching call to kunmap_high() is necessary.
227 * This can be called from any context.
229 void *kmap_high_get(struct page *page)
231 unsigned long vaddr, flags;
233 lock_kmap_any(flags);
234 vaddr = (unsigned long)page_address(page);
236 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
237 pkmap_count[PKMAP_NR(vaddr)]++;
239 unlock_kmap_any(flags);
240 return (void*) vaddr;
245 * kunmap_high - map a highmem page into memory
246 * @page: &struct page to unmap
248 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
249 * only from user context.
251 void kunmap_high(struct page *page)
258 lock_kmap_any(flags);
259 vaddr = (unsigned long)page_address(page);
261 nr = PKMAP_NR(vaddr);
264 * A count must never go down to zero
265 * without a TLB flush!
268 switch (--pkmap_count[nr]) {
273 * Avoid an unnecessary wake_up() function call.
274 * The common case is pkmap_count[] == 1, but
276 * The tasks queued in the wait-queue are guarded
277 * by both the lock in the wait-queue-head and by
278 * the kmap_lock. As the kmap_lock is held here,
279 * no need for the wait-queue-head's lock. Simply
280 * test if the queue is empty.
282 need_wakeup = waitqueue_active(&pkmap_map_wait);
284 unlock_kmap_any(flags);
286 /* do wake-up, if needed, race-free outside of the spin lock */
288 wake_up(&pkmap_map_wait);
291 EXPORT_SYMBOL(kunmap_high);
294 #if defined(HASHED_PAGE_VIRTUAL)
296 #define PA_HASH_ORDER 7
299 * Describes one page->virtual association
301 struct page_address_map {
304 struct list_head list;
308 * page_address_map freelist, allocated from page_address_maps.
310 static struct list_head page_address_pool; /* freelist */
311 static spinlock_t pool_lock; /* protects page_address_pool */
316 static struct page_address_slot {
317 struct list_head lh; /* List of page_address_maps */
318 spinlock_t lock; /* Protect this bucket's list */
319 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
321 static struct page_address_slot *page_slot(struct page *page)
323 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
327 * page_address - get the mapped virtual address of a page
328 * @page: &struct page to get the virtual address of
330 * Returns the page's virtual address.
332 void *page_address(struct page *page)
336 struct page_address_slot *pas;
338 if (!PageHighMem(page))
339 return lowmem_page_address(page);
341 pas = page_slot(page);
343 spin_lock_irqsave(&pas->lock, flags);
344 if (!list_empty(&pas->lh)) {
345 struct page_address_map *pam;
347 list_for_each_entry(pam, &pas->lh, list) {
348 if (pam->page == page) {
355 spin_unlock_irqrestore(&pas->lock, flags);
359 EXPORT_SYMBOL(page_address);
362 * set_page_address - set a page's virtual address
363 * @page: &struct page to set
364 * @virtual: virtual address to use
366 void set_page_address(struct page *page, void *virtual)
369 struct page_address_slot *pas;
370 struct page_address_map *pam;
372 BUG_ON(!PageHighMem(page));
374 pas = page_slot(page);
375 if (virtual) { /* Add */
376 BUG_ON(list_empty(&page_address_pool));
378 spin_lock_irqsave(&pool_lock, flags);
379 pam = list_entry(page_address_pool.next,
380 struct page_address_map, list);
381 list_del(&pam->list);
382 spin_unlock_irqrestore(&pool_lock, flags);
385 pam->virtual = virtual;
387 spin_lock_irqsave(&pas->lock, flags);
388 list_add_tail(&pam->list, &pas->lh);
389 spin_unlock_irqrestore(&pas->lock, flags);
390 } else { /* Remove */
391 spin_lock_irqsave(&pas->lock, flags);
392 list_for_each_entry(pam, &pas->lh, list) {
393 if (pam->page == page) {
394 list_del(&pam->list);
395 spin_unlock_irqrestore(&pas->lock, flags);
396 spin_lock_irqsave(&pool_lock, flags);
397 list_add_tail(&pam->list, &page_address_pool);
398 spin_unlock_irqrestore(&pool_lock, flags);
402 spin_unlock_irqrestore(&pas->lock, flags);
408 static struct page_address_map page_address_maps[LAST_PKMAP];
410 void __init page_address_init(void)
414 INIT_LIST_HEAD(&page_address_pool);
415 for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
416 list_add(&page_address_maps[i].list, &page_address_pool);
417 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
418 INIT_LIST_HEAD(&page_address_htable[i].lh);
419 spin_lock_init(&page_address_htable[i].lock);
421 spin_lock_init(&pool_lock);
424 #endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */