Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * High memory handling common code and variables.
4 *
5 * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
6 * Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
7 *
8 *
9 * Redesigned the x86 32-bit VM architecture to deal with
10 * 64-bit physical space. With current x86 CPUs this
11 * means up to 64 Gigabytes physical RAM.
12 *
13 * Rewrote high memory support to move the page cache into
14 * high memory. Implemented permanent (schedulable) kmaps
15 * based on Linus' idea.
16 *
17 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
18 */
19
20#include <linux/mm.h>
21#include <linux/export.h>
22#include <linux/swap.h>
23#include <linux/bio.h>
24#include <linux/pagemap.h>
25#include <linux/mempool.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#include <linux/vmalloc.h>
32
33#ifdef CONFIG_KMAP_LOCAL
34static inline int kmap_local_calc_idx(int idx)
35{
36 return idx + KM_MAX_IDX * smp_processor_id();
37}
38
39#ifndef arch_kmap_local_map_idx
40#define arch_kmap_local_map_idx(idx, pfn) kmap_local_calc_idx(idx)
41#endif
42#endif /* CONFIG_KMAP_LOCAL */
43
44/*
45 * Virtual_count is not a pure "count".
46 * 0 means that it is not mapped, and has not been mapped
47 * since a TLB flush - it is usable.
48 * 1 means that there are no users, but it has been mapped
49 * since the last TLB flush - so we can't use it.
50 * n means that there are (n-1) current users of it.
51 */
52#ifdef CONFIG_HIGHMEM
53
54/*
55 * Architecture with aliasing data cache may define the following family of
56 * helper functions in its asm/highmem.h to control cache color of virtual
57 * addresses where physical memory pages are mapped by kmap.
58 */
59#ifndef get_pkmap_color
60
61/*
62 * Determine color of virtual address where the page should be mapped.
63 */
64static inline unsigned int get_pkmap_color(struct page *page)
65{
66 return 0;
67}
68#define get_pkmap_color get_pkmap_color
69
70/*
71 * Get next index for mapping inside PKMAP region for page with given color.
72 */
73static inline unsigned int get_next_pkmap_nr(unsigned int color)
74{
75 static unsigned int last_pkmap_nr;
76
77 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
78 return last_pkmap_nr;
79}
80
81/*
82 * Determine if page index inside PKMAP region (pkmap_nr) of given color
83 * has wrapped around PKMAP region end. When this happens an attempt to
84 * flush all unused PKMAP slots is made.
85 */
86static inline int no_more_pkmaps(unsigned int pkmap_nr, unsigned int color)
87{
88 return pkmap_nr == 0;
89}
90
91/*
92 * Get the number of PKMAP entries of the given color. If no free slot is
93 * found after checking that many entries, kmap will sleep waiting for
94 * someone to call kunmap and free PKMAP slot.
95 */
96static inline int get_pkmap_entries_count(unsigned int color)
97{
98 return LAST_PKMAP;
99}
100
101/*
102 * Get head of a wait queue for PKMAP entries of the given color.
103 * Wait queues for different mapping colors should be independent to avoid
104 * unnecessary wakeups caused by freeing of slots of other colors.
105 */
106static inline wait_queue_head_t *get_pkmap_wait_queue_head(unsigned int color)
107{
108 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
109
110 return &pkmap_map_wait;
111}
112#endif
113
114atomic_long_t _totalhigh_pages __read_mostly;
115EXPORT_SYMBOL(_totalhigh_pages);
116
117unsigned int __nr_free_highpages(void)
118{
119 struct zone *zone;
120 unsigned int pages = 0;
121
122 for_each_populated_zone(zone) {
123 if (is_highmem(zone))
124 pages += zone_page_state(zone, 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 base = (unsigned long) vaddr & PAGE_MASK;
157 struct kmap_ctrl *kctrl = ¤t->kmap_ctrl;
158 unsigned long addr = (unsigned long)vaddr;
159 int i;
160
161 /* kmap() mappings */
162 if (WARN_ON_ONCE(addr >= PKMAP_ADDR(0) &&
163 addr < PKMAP_ADDR(LAST_PKMAP)))
164 return pte_page(pkmap_page_table[PKMAP_NR(addr)]);
165
166 /* kmap_local_page() mappings */
167 if (WARN_ON_ONCE(base >= __fix_to_virt(FIX_KMAP_END) &&
168 base < __fix_to_virt(FIX_KMAP_BEGIN))) {
169 for (i = 0; i < kctrl->idx; i++) {
170 unsigned long base_addr;
171 int idx;
172
173 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
174 base_addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
175
176 if (base_addr == base)
177 return pte_page(kctrl->pteval[i]);
178 }
179 }
180
181 return virt_to_page(vaddr);
182}
183EXPORT_SYMBOL(__kmap_to_page);
184
185static void flush_all_zero_pkmaps(void)
186{
187 int i;
188 int need_flush = 0;
189
190 flush_cache_kmaps();
191
192 for (i = 0; i < LAST_PKMAP; i++) {
193 struct page *page;
194
195 /*
196 * zero means we don't have anything to do,
197 * >1 means that it is still in use. Only
198 * a count of 1 means that it is free but
199 * needs to be unmapped
200 */
201 if (pkmap_count[i] != 1)
202 continue;
203 pkmap_count[i] = 0;
204
205 /* sanity check */
206 BUG_ON(pte_none(pkmap_page_table[i]));
207
208 /*
209 * Don't need an atomic fetch-and-clear op here;
210 * no-one has the page mapped, and cannot get at
211 * its virtual address (and hence PTE) without first
212 * getting the kmap_lock (which is held here).
213 * So no dangers, even with speculative execution.
214 */
215 page = pte_page(pkmap_page_table[i]);
216 pte_clear(&init_mm, PKMAP_ADDR(i), &pkmap_page_table[i]);
217
218 set_page_address(page, NULL);
219 need_flush = 1;
220 }
221 if (need_flush)
222 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
223}
224
225void __kmap_flush_unused(void)
226{
227 lock_kmap();
228 flush_all_zero_pkmaps();
229 unlock_kmap();
230}
231
232static inline unsigned long map_new_virtual(struct page *page)
233{
234 unsigned long vaddr;
235 int count;
236 unsigned int last_pkmap_nr;
237 unsigned int color = get_pkmap_color(page);
238
239start:
240 count = get_pkmap_entries_count(color);
241 /* Find an empty entry */
242 for (;;) {
243 last_pkmap_nr = get_next_pkmap_nr(color);
244 if (no_more_pkmaps(last_pkmap_nr, color)) {
245 flush_all_zero_pkmaps();
246 count = get_pkmap_entries_count(color);
247 }
248 if (!pkmap_count[last_pkmap_nr])
249 break; /* Found a usable entry */
250 if (--count)
251 continue;
252
253 /*
254 * Sleep for somebody else to unmap their entries
255 */
256 {
257 DECLARE_WAITQUEUE(wait, current);
258 wait_queue_head_t *pkmap_map_wait =
259 get_pkmap_wait_queue_head(color);
260
261 __set_current_state(TASK_UNINTERRUPTIBLE);
262 add_wait_queue(pkmap_map_wait, &wait);
263 unlock_kmap();
264 schedule();
265 remove_wait_queue(pkmap_map_wait, &wait);
266 lock_kmap();
267
268 /* Somebody else might have mapped it while we slept */
269 if (page_address(page))
270 return (unsigned long)page_address(page);
271
272 /* Re-start */
273 goto start;
274 }
275 }
276 vaddr = PKMAP_ADDR(last_pkmap_nr);
277 set_pte_at(&init_mm, vaddr,
278 &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
279
280 pkmap_count[last_pkmap_nr] = 1;
281 set_page_address(page, (void *)vaddr);
282
283 return vaddr;
284}
285
286/**
287 * kmap_high - map a highmem page into memory
288 * @page: &struct page to map
289 *
290 * Returns the page's virtual memory address.
291 *
292 * We cannot call this from interrupts, as it may block.
293 */
294void *kmap_high(struct page *page)
295{
296 unsigned long vaddr;
297
298 /*
299 * For highmem pages, we can't trust "virtual" until
300 * after we have the lock.
301 */
302 lock_kmap();
303 vaddr = (unsigned long)page_address(page);
304 if (!vaddr)
305 vaddr = map_new_virtual(page);
306 pkmap_count[PKMAP_NR(vaddr)]++;
307 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
308 unlock_kmap();
309 return (void *) vaddr;
310}
311EXPORT_SYMBOL(kmap_high);
312
313#ifdef ARCH_NEEDS_KMAP_HIGH_GET
314/**
315 * kmap_high_get - pin a highmem page into memory
316 * @page: &struct page to pin
317 *
318 * Returns the page's current virtual memory address, or NULL if no mapping
319 * exists. If and only if a non null address is returned then a
320 * matching call to kunmap_high() is necessary.
321 *
322 * This can be called from any context.
323 */
324void *kmap_high_get(struct page *page)
325{
326 unsigned long vaddr, flags;
327
328 lock_kmap_any(flags);
329 vaddr = (unsigned long)page_address(page);
330 if (vaddr) {
331 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
332 pkmap_count[PKMAP_NR(vaddr)]++;
333 }
334 unlock_kmap_any(flags);
335 return (void *) vaddr;
336}
337#endif
338
339/**
340 * kunmap_high - unmap a highmem page into memory
341 * @page: &struct page to unmap
342 *
343 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
344 * only from user context.
345 */
346void kunmap_high(struct page *page)
347{
348 unsigned long vaddr;
349 unsigned long nr;
350 unsigned long flags;
351 int need_wakeup;
352 unsigned int color = get_pkmap_color(page);
353 wait_queue_head_t *pkmap_map_wait;
354
355 lock_kmap_any(flags);
356 vaddr = (unsigned long)page_address(page);
357 BUG_ON(!vaddr);
358 nr = PKMAP_NR(vaddr);
359
360 /*
361 * A count must never go down to zero
362 * without a TLB flush!
363 */
364 need_wakeup = 0;
365 switch (--pkmap_count[nr]) {
366 case 0:
367 BUG();
368 case 1:
369 /*
370 * Avoid an unnecessary wake_up() function call.
371 * The common case is pkmap_count[] == 1, but
372 * no waiters.
373 * The tasks queued in the wait-queue are guarded
374 * by both the lock in the wait-queue-head and by
375 * the kmap_lock. As the kmap_lock is held here,
376 * no need for the wait-queue-head's lock. Simply
377 * test if the queue is empty.
378 */
379 pkmap_map_wait = get_pkmap_wait_queue_head(color);
380 need_wakeup = waitqueue_active(pkmap_map_wait);
381 }
382 unlock_kmap_any(flags);
383
384 /* do wake-up, if needed, race-free outside of the spin lock */
385 if (need_wakeup)
386 wake_up(pkmap_map_wait);
387}
388EXPORT_SYMBOL(kunmap_high);
389
390void zero_user_segments(struct page *page, unsigned start1, unsigned end1,
391 unsigned start2, unsigned end2)
392{
393 unsigned int i;
394
395 BUG_ON(end1 > page_size(page) || end2 > page_size(page));
396
397 if (start1 >= end1)
398 start1 = end1 = 0;
399 if (start2 >= end2)
400 start2 = end2 = 0;
401
402 for (i = 0; i < compound_nr(page); i++) {
403 void *kaddr = NULL;
404
405 if (start1 >= PAGE_SIZE) {
406 start1 -= PAGE_SIZE;
407 end1 -= PAGE_SIZE;
408 } else {
409 unsigned this_end = min_t(unsigned, end1, PAGE_SIZE);
410
411 if (end1 > start1) {
412 kaddr = kmap_local_page(page + i);
413 memset(kaddr + start1, 0, this_end - start1);
414 }
415 end1 -= this_end;
416 start1 = 0;
417 }
418
419 if (start2 >= PAGE_SIZE) {
420 start2 -= PAGE_SIZE;
421 end2 -= PAGE_SIZE;
422 } else {
423 unsigned this_end = min_t(unsigned, end2, PAGE_SIZE);
424
425 if (end2 > start2) {
426 if (!kaddr)
427 kaddr = kmap_local_page(page + i);
428 memset(kaddr + start2, 0, this_end - start2);
429 }
430 end2 -= this_end;
431 start2 = 0;
432 }
433
434 if (kaddr) {
435 kunmap_local(kaddr);
436 flush_dcache_page(page + i);
437 }
438
439 if (!end1 && !end2)
440 break;
441 }
442
443 BUG_ON((start1 | start2 | end1 | end2) != 0);
444}
445EXPORT_SYMBOL(zero_user_segments);
446#endif /* CONFIG_HIGHMEM */
447
448#ifdef CONFIG_KMAP_LOCAL
449
450#include <asm/kmap_size.h>
451
452/*
453 * With DEBUG_KMAP_LOCAL the stack depth is doubled and every second
454 * slot is unused which acts as a guard page
455 */
456#ifdef CONFIG_DEBUG_KMAP_LOCAL
457# define KM_INCR 2
458#else
459# define KM_INCR 1
460#endif
461
462static inline int kmap_local_idx_push(void)
463{
464 WARN_ON_ONCE(in_hardirq() && !irqs_disabled());
465 current->kmap_ctrl.idx += KM_INCR;
466 BUG_ON(current->kmap_ctrl.idx >= KM_MAX_IDX);
467 return current->kmap_ctrl.idx - 1;
468}
469
470static inline int kmap_local_idx(void)
471{
472 return current->kmap_ctrl.idx - 1;
473}
474
475static inline void kmap_local_idx_pop(void)
476{
477 current->kmap_ctrl.idx -= KM_INCR;
478 BUG_ON(current->kmap_ctrl.idx < 0);
479}
480
481#ifndef arch_kmap_local_post_map
482# define arch_kmap_local_post_map(vaddr, pteval) do { } while (0)
483#endif
484
485#ifndef arch_kmap_local_pre_unmap
486# define arch_kmap_local_pre_unmap(vaddr) do { } while (0)
487#endif
488
489#ifndef arch_kmap_local_post_unmap
490# define arch_kmap_local_post_unmap(vaddr) do { } while (0)
491#endif
492
493#ifndef arch_kmap_local_unmap_idx
494#define arch_kmap_local_unmap_idx(idx, vaddr) kmap_local_calc_idx(idx)
495#endif
496
497#ifndef arch_kmap_local_high_get
498static inline void *arch_kmap_local_high_get(struct page *page)
499{
500 return NULL;
501}
502#endif
503
504#ifndef arch_kmap_local_set_pte
505#define arch_kmap_local_set_pte(mm, vaddr, ptep, ptev) \
506 set_pte_at(mm, vaddr, ptep, ptev)
507#endif
508
509/* Unmap a local mapping which was obtained by kmap_high_get() */
510static inline bool kmap_high_unmap_local(unsigned long vaddr)
511{
512#ifdef ARCH_NEEDS_KMAP_HIGH_GET
513 if (vaddr >= PKMAP_ADDR(0) && vaddr < PKMAP_ADDR(LAST_PKMAP)) {
514 kunmap_high(pte_page(pkmap_page_table[PKMAP_NR(vaddr)]));
515 return true;
516 }
517#endif
518 return false;
519}
520
521static pte_t *__kmap_pte;
522
523static pte_t *kmap_get_pte(unsigned long vaddr, int idx)
524{
525 if (IS_ENABLED(CONFIG_KMAP_LOCAL_NON_LINEAR_PTE_ARRAY))
526 /*
527 * Set by the arch if __kmap_pte[-idx] does not produce
528 * the correct entry.
529 */
530 return virt_to_kpte(vaddr);
531 if (!__kmap_pte)
532 __kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
533 return &__kmap_pte[-idx];
534}
535
536void *__kmap_local_pfn_prot(unsigned long pfn, pgprot_t prot)
537{
538 pte_t pteval, *kmap_pte;
539 unsigned long vaddr;
540 int idx;
541
542 /*
543 * Disable migration so resulting virtual address is stable
544 * across preemption.
545 */
546 migrate_disable();
547 preempt_disable();
548 idx = arch_kmap_local_map_idx(kmap_local_idx_push(), pfn);
549 vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
550 kmap_pte = kmap_get_pte(vaddr, idx);
551 BUG_ON(!pte_none(*kmap_pte));
552 pteval = pfn_pte(pfn, prot);
553 arch_kmap_local_set_pte(&init_mm, vaddr, kmap_pte, pteval);
554 arch_kmap_local_post_map(vaddr, pteval);
555 current->kmap_ctrl.pteval[kmap_local_idx()] = pteval;
556 preempt_enable();
557
558 return (void *)vaddr;
559}
560EXPORT_SYMBOL_GPL(__kmap_local_pfn_prot);
561
562void *__kmap_local_page_prot(struct page *page, pgprot_t prot)
563{
564 void *kmap;
565
566 /*
567 * To broaden the usage of the actual kmap_local() machinery always map
568 * pages when debugging is enabled and the architecture has no problems
569 * with alias mappings.
570 */
571 if (!IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP) && !PageHighMem(page))
572 return page_address(page);
573
574 /* Try kmap_high_get() if architecture has it enabled */
575 kmap = arch_kmap_local_high_get(page);
576 if (kmap)
577 return kmap;
578
579 return __kmap_local_pfn_prot(page_to_pfn(page), prot);
580}
581EXPORT_SYMBOL(__kmap_local_page_prot);
582
583void kunmap_local_indexed(const void *vaddr)
584{
585 unsigned long addr = (unsigned long) vaddr & PAGE_MASK;
586 pte_t *kmap_pte;
587 int idx;
588
589 if (addr < __fix_to_virt(FIX_KMAP_END) ||
590 addr > __fix_to_virt(FIX_KMAP_BEGIN)) {
591 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP)) {
592 /* This _should_ never happen! See above. */
593 WARN_ON_ONCE(1);
594 return;
595 }
596 /*
597 * Handle mappings which were obtained by kmap_high_get()
598 * first as the virtual address of such mappings is below
599 * PAGE_OFFSET. Warn for all other addresses which are in
600 * the user space part of the virtual address space.
601 */
602 if (!kmap_high_unmap_local(addr))
603 WARN_ON_ONCE(addr < PAGE_OFFSET);
604 return;
605 }
606
607 preempt_disable();
608 idx = arch_kmap_local_unmap_idx(kmap_local_idx(), addr);
609 WARN_ON_ONCE(addr != __fix_to_virt(FIX_KMAP_BEGIN + idx));
610
611 kmap_pte = kmap_get_pte(addr, idx);
612 arch_kmap_local_pre_unmap(addr);
613 pte_clear(&init_mm, addr, kmap_pte);
614 arch_kmap_local_post_unmap(addr);
615 current->kmap_ctrl.pteval[kmap_local_idx()] = __pte(0);
616 kmap_local_idx_pop();
617 preempt_enable();
618 migrate_enable();
619}
620EXPORT_SYMBOL(kunmap_local_indexed);
621
622/*
623 * Invoked before switch_to(). This is safe even when during or after
624 * clearing the maps an interrupt which needs a kmap_local happens because
625 * the task::kmap_ctrl.idx is not modified by the unmapping code so a
626 * nested kmap_local will use the next unused index and restore the index
627 * on unmap. The already cleared kmaps of the outgoing task are irrelevant
628 * because the interrupt context does not know about them. The same applies
629 * when scheduling back in for an interrupt which happens before the
630 * restore is complete.
631 */
632void __kmap_local_sched_out(void)
633{
634 struct task_struct *tsk = current;
635 pte_t *kmap_pte;
636 int i;
637
638 /* Clear kmaps */
639 for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
640 pte_t pteval = tsk->kmap_ctrl.pteval[i];
641 unsigned long addr;
642 int idx;
643
644 /* With debug all even slots are unmapped and act as guard */
645 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
646 WARN_ON_ONCE(pte_val(pteval) != 0);
647 continue;
648 }
649 if (WARN_ON_ONCE(pte_none(pteval)))
650 continue;
651
652 /*
653 * This is a horrible hack for XTENSA to calculate the
654 * coloured PTE index. Uses the PFN encoded into the pteval
655 * and the map index calculation because the actual mapped
656 * virtual address is not stored in task::kmap_ctrl.
657 * For any sane architecture this is optimized out.
658 */
659 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
660
661 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
662 kmap_pte = kmap_get_pte(addr, idx);
663 arch_kmap_local_pre_unmap(addr);
664 pte_clear(&init_mm, addr, kmap_pte);
665 arch_kmap_local_post_unmap(addr);
666 }
667}
668
669void __kmap_local_sched_in(void)
670{
671 struct task_struct *tsk = current;
672 pte_t *kmap_pte;
673 int i;
674
675 /* Restore kmaps */
676 for (i = 0; i < tsk->kmap_ctrl.idx; i++) {
677 pte_t pteval = tsk->kmap_ctrl.pteval[i];
678 unsigned long addr;
679 int idx;
680
681 /* With debug all even slots are unmapped and act as guard */
682 if (IS_ENABLED(CONFIG_DEBUG_KMAP_LOCAL) && !(i & 0x01)) {
683 WARN_ON_ONCE(pte_val(pteval) != 0);
684 continue;
685 }
686 if (WARN_ON_ONCE(pte_none(pteval)))
687 continue;
688
689 /* See comment in __kmap_local_sched_out() */
690 idx = arch_kmap_local_map_idx(i, pte_pfn(pteval));
691 addr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
692 kmap_pte = kmap_get_pte(addr, idx);
693 set_pte_at(&init_mm, addr, kmap_pte, pteval);
694 arch_kmap_local_post_map(addr, pteval);
695 }
696}
697
698void kmap_local_fork(struct task_struct *tsk)
699{
700 if (WARN_ON_ONCE(tsk->kmap_ctrl.idx))
701 memset(&tsk->kmap_ctrl, 0, sizeof(tsk->kmap_ctrl));
702}
703
704#endif
705
706#if defined(HASHED_PAGE_VIRTUAL)
707
708#define PA_HASH_ORDER 7
709
710/*
711 * Describes one page->virtual association
712 */
713struct page_address_map {
714 struct page *page;
715 void *virtual;
716 struct list_head list;
717};
718
719static struct page_address_map page_address_maps[LAST_PKMAP];
720
721/*
722 * Hash table bucket
723 */
724static struct page_address_slot {
725 struct list_head lh; /* List of page_address_maps */
726 spinlock_t lock; /* Protect this bucket's list */
727} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
728
729static struct page_address_slot *page_slot(const struct page *page)
730{
731 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
732}
733
734/**
735 * page_address - get the mapped virtual address of a page
736 * @page: &struct page to get the virtual address of
737 *
738 * Returns the page's virtual address.
739 */
740void *page_address(const struct page *page)
741{
742 unsigned long flags;
743 void *ret;
744 struct page_address_slot *pas;
745
746 if (!PageHighMem(page))
747 return lowmem_page_address(page);
748
749 pas = page_slot(page);
750 ret = NULL;
751 spin_lock_irqsave(&pas->lock, flags);
752 if (!list_empty(&pas->lh)) {
753 struct page_address_map *pam;
754
755 list_for_each_entry(pam, &pas->lh, list) {
756 if (pam->page == page) {
757 ret = pam->virtual;
758 break;
759 }
760 }
761 }
762
763 spin_unlock_irqrestore(&pas->lock, flags);
764 return ret;
765}
766EXPORT_SYMBOL(page_address);
767
768/**
769 * set_page_address - set a page's virtual address
770 * @page: &struct page to set
771 * @virtual: virtual address to use
772 */
773void set_page_address(struct page *page, void *virtual)
774{
775 unsigned long flags;
776 struct page_address_slot *pas;
777 struct page_address_map *pam;
778
779 BUG_ON(!PageHighMem(page));
780
781 pas = page_slot(page);
782 if (virtual) { /* Add */
783 pam = &page_address_maps[PKMAP_NR((unsigned long)virtual)];
784 pam->page = page;
785 pam->virtual = virtual;
786
787 spin_lock_irqsave(&pas->lock, flags);
788 list_add_tail(&pam->list, &pas->lh);
789 spin_unlock_irqrestore(&pas->lock, flags);
790 } else { /* Remove */
791 spin_lock_irqsave(&pas->lock, flags);
792 list_for_each_entry(pam, &pas->lh, list) {
793 if (pam->page == page) {
794 list_del(&pam->list);
795 break;
796 }
797 }
798 spin_unlock_irqrestore(&pas->lock, flags);
799 }
800
801 return;
802}
803
804void __init page_address_init(void)
805{
806 int i;
807
808 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
809 INIT_LIST_HEAD(&page_address_htable[i].lh);
810 spin_lock_init(&page_address_htable[i].lock);
811 }
812}
813
814#endif /* defined(HASHED_PAGE_VIRTUAL) */
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/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>
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
47unsigned long totalhigh_pages __read_mostly;
48EXPORT_SYMBOL(totalhigh_pages);
49
50
51EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
52
53unsigned int nr_free_highpages (void)
54{
55 pg_data_t *pgdat;
56 unsigned int pages = 0;
57
58 for_each_online_pgdat(pgdat) {
59 pages += zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
60 NR_FREE_PAGES);
61 if (zone_movable_is_highmem())
62 pages += zone_page_state(
63 &pgdat->node_zones[ZONE_MOVABLE],
64 NR_FREE_PAGES);
65 }
66
67 return pages;
68}
69
70static int pkmap_count[LAST_PKMAP];
71static unsigned int last_pkmap_nr;
72static __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
73
74pte_t * pkmap_page_table;
75
76static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
77
78/*
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.
82 */
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)
88#else
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)
95#endif
96
97static void flush_all_zero_pkmaps(void)
98{
99 int i;
100 int need_flush = 0;
101
102 flush_cache_kmaps();
103
104 for (i = 0; i < LAST_PKMAP; i++) {
105 struct page *page;
106
107 /*
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
112 */
113 if (pkmap_count[i] != 1)
114 continue;
115 pkmap_count[i] = 0;
116
117 /* sanity check */
118 BUG_ON(pte_none(pkmap_page_table[i]));
119
120 /*
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.
126 */
127 page = pte_page(pkmap_page_table[i]);
128 pte_clear(&init_mm, (unsigned long)page_address(page),
129 &pkmap_page_table[i]);
130
131 set_page_address(page, NULL);
132 need_flush = 1;
133 }
134 if (need_flush)
135 flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
136}
137
138/**
139 * kmap_flush_unused - flush all unused kmap mappings in order to remove stray mappings
140 */
141void kmap_flush_unused(void)
142{
143 lock_kmap();
144 flush_all_zero_pkmaps();
145 unlock_kmap();
146}
147
148static inline unsigned long map_new_virtual(struct page *page)
149{
150 unsigned long vaddr;
151 int count;
152
153start:
154 count = LAST_PKMAP;
155 /* Find an empty entry */
156 for (;;) {
157 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
158 if (!last_pkmap_nr) {
159 flush_all_zero_pkmaps();
160 count = LAST_PKMAP;
161 }
162 if (!pkmap_count[last_pkmap_nr])
163 break; /* Found a usable entry */
164 if (--count)
165 continue;
166
167 /*
168 * Sleep for somebody else to unmap their entries
169 */
170 {
171 DECLARE_WAITQUEUE(wait, current);
172
173 __set_current_state(TASK_UNINTERRUPTIBLE);
174 add_wait_queue(&pkmap_map_wait, &wait);
175 unlock_kmap();
176 schedule();
177 remove_wait_queue(&pkmap_map_wait, &wait);
178 lock_kmap();
179
180 /* Somebody else might have mapped it while we slept */
181 if (page_address(page))
182 return (unsigned long)page_address(page);
183
184 /* Re-start */
185 goto start;
186 }
187 }
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));
191
192 pkmap_count[last_pkmap_nr] = 1;
193 set_page_address(page, (void *)vaddr);
194
195 return vaddr;
196}
197
198/**
199 * kmap_high - map a highmem page into memory
200 * @page: &struct page to map
201 *
202 * Returns the page's virtual memory address.
203 *
204 * We cannot call this from interrupts, as it may block.
205 */
206void *kmap_high(struct page *page)
207{
208 unsigned long vaddr;
209
210 /*
211 * For highmem pages, we can't trust "virtual" until
212 * after we have the lock.
213 */
214 lock_kmap();
215 vaddr = (unsigned long)page_address(page);
216 if (!vaddr)
217 vaddr = map_new_virtual(page);
218 pkmap_count[PKMAP_NR(vaddr)]++;
219 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 2);
220 unlock_kmap();
221 return (void*) vaddr;
222}
223
224EXPORT_SYMBOL(kmap_high);
225
226#ifdef ARCH_NEEDS_KMAP_HIGH_GET
227/**
228 * kmap_high_get - pin a highmem page into memory
229 * @page: &struct page to pin
230 *
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.
234 *
235 * This can be called from any context.
236 */
237void *kmap_high_get(struct page *page)
238{
239 unsigned long vaddr, flags;
240
241 lock_kmap_any(flags);
242 vaddr = (unsigned long)page_address(page);
243 if (vaddr) {
244 BUG_ON(pkmap_count[PKMAP_NR(vaddr)] < 1);
245 pkmap_count[PKMAP_NR(vaddr)]++;
246 }
247 unlock_kmap_any(flags);
248 return (void*) vaddr;
249}
250#endif
251
252/**
253 * kunmap_high - map a highmem page into memory
254 * @page: &struct page to unmap
255 *
256 * If ARCH_NEEDS_KMAP_HIGH_GET is not defined then this may be called
257 * only from user context.
258 */
259void kunmap_high(struct page *page)
260{
261 unsigned long vaddr;
262 unsigned long nr;
263 unsigned long flags;
264 int need_wakeup;
265
266 lock_kmap_any(flags);
267 vaddr = (unsigned long)page_address(page);
268 BUG_ON(!vaddr);
269 nr = PKMAP_NR(vaddr);
270
271 /*
272 * A count must never go down to zero
273 * without a TLB flush!
274 */
275 need_wakeup = 0;
276 switch (--pkmap_count[nr]) {
277 case 0:
278 BUG();
279 case 1:
280 /*
281 * Avoid an unnecessary wake_up() function call.
282 * The common case is pkmap_count[] == 1, but
283 * no waiters.
284 * The tasks queued in the wait-queue are guarded
285 * by both the lock in the wait-queue-head and by
286 * the kmap_lock. As the kmap_lock is held here,
287 * no need for the wait-queue-head's lock. Simply
288 * test if the queue is empty.
289 */
290 need_wakeup = waitqueue_active(&pkmap_map_wait);
291 }
292 unlock_kmap_any(flags);
293
294 /* do wake-up, if needed, race-free outside of the spin lock */
295 if (need_wakeup)
296 wake_up(&pkmap_map_wait);
297}
298
299EXPORT_SYMBOL(kunmap_high);
300#endif
301
302#if defined(HASHED_PAGE_VIRTUAL)
303
304#define PA_HASH_ORDER 7
305
306/*
307 * Describes one page->virtual association
308 */
309struct page_address_map {
310 struct page *page;
311 void *virtual;
312 struct list_head list;
313};
314
315/*
316 * page_address_map freelist, allocated from page_address_maps.
317 */
318static struct list_head page_address_pool; /* freelist */
319static spinlock_t pool_lock; /* protects page_address_pool */
320
321/*
322 * Hash table bucket
323 */
324static struct page_address_slot {
325 struct list_head lh; /* List of page_address_maps */
326 spinlock_t lock; /* Protect this bucket's list */
327} ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
328
329static struct page_address_slot *page_slot(const struct page *page)
330{
331 return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
332}
333
334/**
335 * page_address - get the mapped virtual address of a page
336 * @page: &struct page to get the virtual address of
337 *
338 * Returns the page's virtual address.
339 */
340void *page_address(const struct page *page)
341{
342 unsigned long flags;
343 void *ret;
344 struct page_address_slot *pas;
345
346 if (!PageHighMem(page))
347 return lowmem_page_address(page);
348
349 pas = page_slot(page);
350 ret = NULL;
351 spin_lock_irqsave(&pas->lock, flags);
352 if (!list_empty(&pas->lh)) {
353 struct page_address_map *pam;
354
355 list_for_each_entry(pam, &pas->lh, list) {
356 if (pam->page == page) {
357 ret = pam->virtual;
358 goto done;
359 }
360 }
361 }
362done:
363 spin_unlock_irqrestore(&pas->lock, flags);
364 return ret;
365}
366
367EXPORT_SYMBOL(page_address);
368
369/**
370 * set_page_address - set a page's virtual address
371 * @page: &struct page to set
372 * @virtual: virtual address to use
373 */
374void set_page_address(struct page *page, void *virtual)
375{
376 unsigned long flags;
377 struct page_address_slot *pas;
378 struct page_address_map *pam;
379
380 BUG_ON(!PageHighMem(page));
381
382 pas = page_slot(page);
383 if (virtual) { /* Add */
384 BUG_ON(list_empty(&page_address_pool));
385
386 spin_lock_irqsave(&pool_lock, flags);
387 pam = list_entry(page_address_pool.next,
388 struct page_address_map, list);
389 list_del(&pam->list);
390 spin_unlock_irqrestore(&pool_lock, flags);
391
392 pam->page = page;
393 pam->virtual = virtual;
394
395 spin_lock_irqsave(&pas->lock, flags);
396 list_add_tail(&pam->list, &pas->lh);
397 spin_unlock_irqrestore(&pas->lock, flags);
398 } else { /* Remove */
399 spin_lock_irqsave(&pas->lock, flags);
400 list_for_each_entry(pam, &pas->lh, list) {
401 if (pam->page == page) {
402 list_del(&pam->list);
403 spin_unlock_irqrestore(&pas->lock, flags);
404 spin_lock_irqsave(&pool_lock, flags);
405 list_add_tail(&pam->list, &page_address_pool);
406 spin_unlock_irqrestore(&pool_lock, flags);
407 goto done;
408 }
409 }
410 spin_unlock_irqrestore(&pas->lock, flags);
411 }
412done:
413 return;
414}
415
416static struct page_address_map page_address_maps[LAST_PKMAP];
417
418void __init page_address_init(void)
419{
420 int i;
421
422 INIT_LIST_HEAD(&page_address_pool);
423 for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
424 list_add(&page_address_maps[i].list, &page_address_pool);
425 for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
426 INIT_LIST_HEAD(&page_address_htable[i].lh);
427 spin_lock_init(&page_address_htable[i].lock);
428 }
429 spin_lock_init(&pool_lock);
430}
431
432#endif /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */