1/*
2 * High memory handling common code and variables.
3 *
4 * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
5 *          Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
6 *
7 *
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.
11 *
12 * Rewrote high memory support to move the page cache into
13 * high memory. Implemented permanent (schedulable) kmaps
14 * based on Linus' idea.
15 *
16 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
17 */
18
19#include <linux/mm.h>
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>
31
32
33#if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
34DEFINE_PER_CPU(int, __kmap_atomic_idx);
35#endif
36
37/*
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.
44 */
45#ifdef CONFIG_HIGHMEM
46
47/*
48 * Architecture with aliasing data cache may define the following family of
49 * helper functions in its asm/highmem.h to control cache color of virtual
50 * addresses where physical memory pages are mapped by kmap.
51 */
52#ifndef get_pkmap_color
53
54/*
55 * Determine color of virtual address where the page should be mapped.
56 */
57static inline unsigned int get_pkmap_color(struct page *page)
58{
59	return 0;
60}
61#define get_pkmap_color get_pkmap_color
62
63/*
64 * Get next index for mapping inside PKMAP region for page with given color.
65 */
66static inline unsigned int get_next_pkmap_nr(unsigned int color)
67{
68	static unsigned int last_pkmap_nr;
69
70	last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
71	return last_pkmap_nr;
72}
73
74/*
75 * Determine if page index inside PKMAP region (pkmap_nr) of given color
76 * has wrapped around PKMAP region end. When this happens an attempt to
77 * flush all unused PKMAP slots is made.
78 */
79static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
80{
81	return pkmap_nr == 0;
82}
83
84/*
85 * Get the number of PKMAP entries of the given color. If no free slot is
86 * found after checking that many entries, kmap will sleep waiting for
87 * someone to call kunmap and free PKMAP slot.
88 */
89static inline int get_pkmap_entries_count(unsigned int color)
90{
91	return LAST_PKMAP;
92}
93
94/*
95 * Get head of a wait queue for PKMAP entries of the given color.
96 * Wait queues for different mapping colors should be independent to avoid
97 * unnecessary wakeups caused by freeing of slots of other colors.
98 */
99static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
100{
101	static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
102
103	return &pkmap_map_wait;
104}
105#endif
106
107unsigned long totalhigh_pages __read_mostly;
108EXPORT_SYMBOL(totalhigh_pages);
109
110
111EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
112
113unsigned int nr_free_highpages (void)
114{
115	pg_data_t *pgdat;
116	unsigned int pages = 0;
117
118	for_each_online_pgdat(pgdat) {
119		pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
120			NR_FREE_PAGES);
121		if (zone_movable_is_highmem())
122			pages += zone_page_state(
123					&pgdat->node_zones[ZONE_MOVABLE],
124					NR_FREE_PAGES);
125	}
126
127	return pages;
128}
129
130static int pkmap_count[LAST_PKMAP];
131static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
132
133pte_t * pkmap_page_table;
134
135/*
136 * Most architectures have no use for kmap_high_get(), so let's abstract
137 * the disabling of IRQ out of the locking in that case to save on a
138 * potential useless overhead.
139 */
140#ifdef ARCH_NEEDS_KMAP_HIGH_GET
141#define lock_kmap()             spin_lock_irq(&kmap_lock)
142#define unlock_kmap()           spin_unlock_irq(&kmap_lock)
143#define lock_kmap_any(flags)    spin_lock_irqsave(&kmap_lock, flags)
144#define unlock_kmap_any(flags)  spin_unlock_irqrestore(&kmap_lock, flags)
145#else
146#define lock_kmap()             spin_lock(&kmap_lock)
147#define unlock_kmap()           spin_unlock(&kmap_lock)
148#define lock_kmap_any(flags)    \
149		do { spin_lock(&kmap_lock); (void)(flags); } while (0)
150#define unlock_kmap_any(flags)  \
151		do { spin_unlock(&kmap_lock); (void)(flags); } while (0)
152#endif
153
154struct page *kmap_to_page(void *vaddr)
155{
156	unsigned long addr = (unsigned long)vaddr;
157
158	if (addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP)) {
159		int i = PKMAP_NR(addr);
160		return pte_page(pkmap_page_table[i]);
161	}
162
163	return virt_to_page(addr);
164}
165EXPORT_SYMBOL(kmap_to_page);
166
167static void flush_all_zero_pkmaps(void)
168{
169	int i;
170	int need_flush = 0;
171
172	flush_cache_kmaps();
173
174	for (i = 0; i < LAST_PKMAP; i++) {
175		struct page *page;
176
177		/*
178		 * zero means we don't have anything to do,
179		 * >1 means that it is still in use. Only
180		 * a count of 1 means that it is free but
181		 * needs to be unmapped
182		 */
183		if (pkmap_count[i] != 1)
184			continue;
185		pkmap_count[i] = 0;
186
187		/* sanity check */
188		BUG_ON(pte_none(pkmap_page_table[i]));
189
190		/*
191		 * Don't need an atomic fetch-and-clear op here;
192		 * no-one has the page mapped, and cannot get at
193		 * its virtual address (and hence PTE) without first
194		 * getting the kmap_lock (which is held here).
195		 * So no dangers, even with speculative execution.
196		 */
197		page = pte_page(pkmap_page_table[i]);
198		pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
199
200		set_page_address(page, NULL);
201		need_flush = 1;
202	}
203	if (need_flush)
204		flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
205}
206
207/**
208 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
209 */
210void kmap_flush_unused(void)
211{
212	lock_kmap();
213	flush_all_zero_pkmaps();
214	unlock_kmap();
215}
216
217static inline unsigned long map_new_virtual(struct page *page)
218{
219	unsigned long vaddr;
220	int count;
221	unsigned int last_pkmap_nr;
222	unsigned int color = get_pkmap_color(page);
223
224start:
225	count = get_pkmap_entries_count(color);
226	/* Find an empty entry */
227	for (;;) {
228		last_pkmap_nr = get_next_pkmap_nr(color);
229		if (no_more_pkmaps(last_pkmap_nr, color)) {
230			flush_all_zero_pkmaps();
231			count = get_pkmap_entries_count(color);
232		}
233		if (!pkmap_count[last_pkmap_nr])
234			break;	/* Found a usable entry */
235		if (--count)
236			continue;
237
238		/*
239		 * Sleep for somebody else to unmap their entries
240		 */
241		{
242			DECLARE_WAITQUEUE(wait, current);
243			wait_queue_head_t *pkmap_map_wait =
244				get_pkmap_wait_queue_head(color);
245
246			__set_current_state(TASK_UNINTERRUPTIBLE);
247			add_wait_queue(pkmap_map_wait, &wait);
248			unlock_kmap();
249			schedule();
250			remove_wait_queue(pkmap_map_wait, &wait);
251			lock_kmap();
252
253			/* Somebody else might have mapped it while we slept */
254			if (page_address(page))
255				return (unsigned long)page_address(page);
256
257			/* Re-start */
258			goto start;
259		}
260	}
261	vaddr = PKMAP_ADDR(last_pkmap_nr);
262	set_pte_at(&init_mm, vaddr,
263		   &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
264
265	pkmap_count[last_pkmap_nr] = 1;
266	set_page_address(page, (void *)vaddr);
267
268	return vaddr;
269}
270
271/**
272 * kmap_high - map a highmem page into memory
273 * @page: &struct page to map
274 *
275 * Returns the page's virtual memory address.
276 *
277 * We cannot call this from interrupts, as it may block.
278 */
279void *kmap_high(struct page *page)
280{
281	unsigned long vaddr;
282
283	/*
284	 * For highmem pages, we can't trust "virtual" until
285	 * after we have the lock.
286	 */
287	lock_kmap();
288	vaddr = (unsigned long)page_address(page);
289	if (!vaddr)
290		vaddr = map_new_virtual(page);
291	pkmap_count[PKMAP_NR(vaddr)]++;
292	BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
293	unlock_kmap();
294	return (void*) vaddr;
295}
296
297EXPORT_SYMBOL(kmap_high);
298
299#ifdef ARCH_NEEDS_KMAP_HIGH_GET
300/**
301 * kmap_high_get - pin a highmem page into memory
302 * @page: &struct page to pin
303 *
304 * Returns the page's current virtual memory address, or NULL if no mapping
305 * exists.  If and only if a non null address is returned then a
306 * matching call to kunmap_high() is necessary.
307 *
308 * This can be called from any context.
309 */
310void *kmap_high_get(struct page *page)
311{
312	unsigned long vaddr, flags;
313
314	lock_kmap_any(flags);
315	vaddr = (unsigned long)page_address(page);
316	if (vaddr) {
317		BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
318		pkmap_count[PKMAP_NR(vaddr)]++;
319	}
320	unlock_kmap_any(flags);
321	return (void*) vaddr;
322}
323#endif
324
325/**
326 * kunmap_high - unmap a highmem page into memory
327 * @page: &struct page to unmap
328 *
329 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
330 * only from user context.
331 */
332void kunmap_high(struct page *page)
333{
334	unsigned long vaddr;
335	unsigned long nr;
336	unsigned long flags;
337	int need_wakeup;
338	unsigned int color = get_pkmap_color(page);
339	wait_queue_head_t *pkmap_map_wait;
340
341	lock_kmap_any(flags);
342	vaddr = (unsigned long)page_address(page);
343	BUG_ON(!vaddr);
344	nr = PKMAP_NR(vaddr);
345
346	/*
347	 * A count must never go down to zero
348	 * without a TLB flush!
349	 */
350	need_wakeup = 0;
351	switch (--pkmap_count[nr]) {
352	case 0:
353		BUG();
354	case 1:
355		/*
356		 * Avoid an unnecessary wake_up() function call.
357		 * The common case is pkmap_count[] == 1, but
358		 * no waiters.
359		 * The tasks queued in the wait-queue are guarded
360		 * by both the lock in the wait-queue-head and by
361		 * the kmap_lock.  As the kmap_lock is held here,
362		 * no need for the wait-queue-head's lock.  Simply
363		 * test if the queue is empty.
364		 */
365		pkmap_map_wait = get_pkmap_wait_queue_head(color);
366		need_wakeup = waitqueue_active(pkmap_map_wait);
367	}
368	unlock_kmap_any(flags);
369
370	/* do wake-up, if needed, race-free outside of the spin lock */
371	if (need_wakeup)
372		wake_up(pkmap_map_wait);
373}
374
375EXPORT_SYMBOL(kunmap_high);
376#endif
377
378#if defined(HASHED_PAGE_VIRTUAL)
379
380#define PA_HASH_ORDER	7
381
382/*
383 * Describes one page->virtual association
384 */
385struct page_address_map {
386	struct page *page;
387	void *virtual;
388	struct list_head list;
389};
390
391static struct page_address_map page_address_maps[LAST_PKMAP];
392
393/*
394 * Hash table bucket
395 */
396static struct page_address_slot {
397	struct list_head lh;			/* List of page_address_maps */
398	spinlock_t lock;			/* Protect this bucket's list */
399} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
400
401static struct page_address_slot *page_slot(const struct page *page)
402{
403	return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
404}
405
406/**
407 * page_address - get the mapped virtual address of a page
408 * @page: &struct page to get the virtual address of
409 *
410 * Returns the page's virtual address.
411 */
412void *page_address(const struct page *page)
413{
414	unsigned long flags;
415	void *ret;
416	struct page_address_slot *pas;
417
418	if (!PageHighMem(page))
419		return lowmem_page_address(page);
420
421	pas = page_slot(page);
422	ret = NULL;
423	spin_lock_irqsave(&pas->lock, flags);
424	if (!list_empty(&pas->lh)) {
425		struct page_address_map *pam;
426
427		list_for_each_entry(pam, &pas->lh, list) {
428			if (pam->page == page) {
429				ret = pam->virtual;
430				goto done;
431			}
432		}
433	}
434done:
435	spin_unlock_irqrestore(&pas->lock, flags);
436	return ret;
437}
438
439EXPORT_SYMBOL(page_address);
440
441/**
442 * set_page_address - set a page's virtual address
443 * @page: &struct page to set
444 * @virtual: virtual address to use
445 */
446void set_page_address(struct page *page, void *virtual)
447{
448	unsigned long flags;
449	struct page_address_slot *pas;
450	struct page_address_map *pam;
451
452	BUG_ON(!PageHighMem(page));
453
454	pas = page_slot(page);
455	if (virtual) {		/* Add */
456		pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
457		pam->page = page;
458		pam->virtual = virtual;
459
460		spin_lock_irqsave(&pas->lock, flags);
461		list_add_tail(&pam->list, &pas->lh);
462		spin_unlock_irqrestore(&pas->lock, flags);
463	} else {		/* Remove */
464		spin_lock_irqsave(&pas->lock, flags);
465		list_for_each_entry(pam, &pas->lh, list) {
466			if (pam->page == page) {
467				list_del(&pam->list);
468				spin_unlock_irqrestore(&pas->lock, flags);
469				goto done;
470			}
471		}
472		spin_unlock_irqrestore(&pas->lock, flags);
473	}
474done:
475	return;
476}
477
478void __init page_address_init(void)
479{
480	int i;
481
482	for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
483		INIT_LIST_HEAD(&page_address_htable[i].lh);
484		spin_lock_init(&page_address_htable[i].lock);
485	}
486}
487
488#endif	/* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */
489