cac6da54203bfdee878d61c2debb71104ce4f0b6
[linux-2.6.git] / arch / x86 / mm / ioremap.c
1 /*
2  * Re-map IO memory to kernel address space so that we can access it.
3  * This is needed for high PCI addresses that aren't mapped in the
4  * 640k-1MB IO memory area on PC's
5  *
6  * (C) Copyright 1995 1996 Linus Torvalds
7  */
8
9 #include <linux/bootmem.h>
10 #include <linux/init.h>
11 #include <linux/io.h>
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/mmiotrace.h>
16
17 #include <asm/cacheflush.h>
18 #include <asm/e820.h>
19 #include <asm/fixmap.h>
20 #include <asm/pgtable.h>
21 #include <asm/tlbflush.h>
22 #include <asm/pgalloc.h>
23 #include <asm/pat.h>
24
25 #ifdef CONFIG_X86_64
26
27 unsigned long __phys_addr(unsigned long x)
28 {
29         if (x >= __START_KERNEL_map)
30                 return x - __START_KERNEL_map + phys_base;
31         return x - PAGE_OFFSET;
32 }
33 EXPORT_SYMBOL(__phys_addr);
34
35 static inline int phys_addr_valid(unsigned long addr)
36 {
37         return addr < (1UL << boot_cpu_data.x86_phys_bits);
38 }
39
40 #else
41
42 static inline int phys_addr_valid(unsigned long addr)
43 {
44         return 1;
45 }
46
47 #endif
48
49 int page_is_ram(unsigned long pagenr)
50 {
51         resource_size_t addr, end;
52         int i;
53
54         /*
55          * A special case is the first 4Kb of memory;
56          * This is a BIOS owned area, not kernel ram, but generally
57          * not listed as such in the E820 table.
58          */
59         if (pagenr == 0)
60                 return 0;
61
62         /*
63          * Second special case: Some BIOSen report the PC BIOS
64          * area (640->1Mb) as ram even though it is not.
65          */
66         if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
67                     pagenr < (BIOS_END >> PAGE_SHIFT))
68                 return 0;
69
70         for (i = 0; i < e820.nr_map; i++) {
71                 /*
72                  * Not usable memory:
73                  */
74                 if (e820.map[i].type != E820_RAM)
75                         continue;
76                 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
77                 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
78
79
80                 if ((pagenr >= addr) && (pagenr < end))
81                         return 1;
82         }
83         return 0;
84 }
85
86 /*
87  * Fix up the linear direct mapping of the kernel to avoid cache attribute
88  * conflicts.
89  */
90 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
91                                unsigned long prot_val)
92 {
93         unsigned long nrpages = size >> PAGE_SHIFT;
94         int err;
95
96         switch (prot_val) {
97         case _PAGE_CACHE_UC:
98         default:
99                 err = _set_memory_uc(vaddr, nrpages);
100                 break;
101         case _PAGE_CACHE_WC:
102                 err = _set_memory_wc(vaddr, nrpages);
103                 break;
104         case _PAGE_CACHE_WB:
105                 err = _set_memory_wb(vaddr, nrpages);
106                 break;
107         }
108
109         return err;
110 }
111
112 /*
113  * Remap an arbitrary physical address space into the kernel virtual
114  * address space. Needed when the kernel wants to access high addresses
115  * directly.
116  *
117  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
118  * have to convert them into an offset in a page-aligned mapping, but the
119  * caller shouldn't need to know that small detail.
120  */
121 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
122                 unsigned long size, unsigned long prot_val, void *caller)
123 {
124         unsigned long pfn, offset, vaddr;
125         resource_size_t last_addr;
126         const resource_size_t unaligned_phys_addr = phys_addr;
127         const unsigned long unaligned_size = size;
128         struct vm_struct *area;
129         unsigned long new_prot_val;
130         pgprot_t prot;
131         int retval;
132         void __iomem *ret_addr;
133
134         /* Don't allow wraparound or zero size */
135         last_addr = phys_addr + size - 1;
136         if (!size || last_addr < phys_addr)
137                 return NULL;
138
139         if (!phys_addr_valid(phys_addr)) {
140                 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
141                        (unsigned long long)phys_addr);
142                 WARN_ON_ONCE(1);
143                 return NULL;
144         }
145
146         /*
147          * Don't remap the low PCI/ISA area, it's always mapped..
148          */
149         if (is_ISA_range(phys_addr, last_addr))
150                 return (__force void __iomem *)phys_to_virt(phys_addr);
151
152         /*
153          * Don't allow anybody to remap normal RAM that we're using..
154          */
155         for (pfn = phys_addr >> PAGE_SHIFT;
156                                 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
157                                 pfn++) {
158
159                 int is_ram = page_is_ram(pfn);
160
161                 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
162                         return NULL;
163                 WARN_ON_ONCE(is_ram);
164         }
165
166         /*
167          * Mappings have to be page-aligned
168          */
169         offset = phys_addr & ~PAGE_MASK;
170         phys_addr &= PAGE_MASK;
171         size = PAGE_ALIGN(last_addr+1) - phys_addr;
172
173         retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
174                                                 prot_val, &new_prot_val);
175         if (retval) {
176                 pr_debug("Warning: reserve_memtype returned %d\n", retval);
177                 return NULL;
178         }
179
180         if (prot_val != new_prot_val) {
181                 /*
182                  * Do not fallback to certain memory types with certain
183                  * requested type:
184                  * - request is uc-, return cannot be write-back
185                  * - request is uc-, return cannot be write-combine
186                  * - request is write-combine, return cannot be write-back
187                  */
188                 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
189                      (new_prot_val == _PAGE_CACHE_WB ||
190                       new_prot_val == _PAGE_CACHE_WC)) ||
191                     (prot_val == _PAGE_CACHE_WC &&
192                      new_prot_val == _PAGE_CACHE_WB)) {
193                         pr_debug(
194                 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
195                                 (unsigned long long)phys_addr,
196                                 (unsigned long long)(phys_addr + size),
197                                 prot_val, new_prot_val);
198                         free_memtype(phys_addr, phys_addr + size);
199                         return NULL;
200                 }
201                 prot_val = new_prot_val;
202         }
203
204         switch (prot_val) {
205         case _PAGE_CACHE_UC:
206         default:
207                 prot = PAGE_KERNEL_NOCACHE;
208                 break;
209         case _PAGE_CACHE_UC_MINUS:
210                 prot = PAGE_KERNEL_UC_MINUS;
211                 break;
212         case _PAGE_CACHE_WC:
213                 prot = PAGE_KERNEL_WC;
214                 break;
215         case _PAGE_CACHE_WB:
216                 prot = PAGE_KERNEL;
217                 break;
218         }
219
220         /*
221          * Ok, go for it..
222          */
223         area = get_vm_area_caller(size, VM_IOREMAP, caller);
224         if (!area)
225                 return NULL;
226         area->phys_addr = phys_addr;
227         vaddr = (unsigned long) area->addr;
228         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
229                 free_memtype(phys_addr, phys_addr + size);
230                 free_vm_area(area);
231                 return NULL;
232         }
233
234         if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
235                 free_memtype(phys_addr, phys_addr + size);
236                 vunmap(area->addr);
237                 return NULL;
238         }
239
240         ret_addr = (void __iomem *) (vaddr + offset);
241         mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
242
243         return ret_addr;
244 }
245
246 /**
247  * ioremap_nocache     -   map bus memory into CPU space
248  * @offset:    bus address of the memory
249  * @size:      size of the resource to map
250  *
251  * ioremap_nocache performs a platform specific sequence of operations to
252  * make bus memory CPU accessible via the readb/readw/readl/writeb/
253  * writew/writel functions and the other mmio helpers. The returned
254  * address is not guaranteed to be usable directly as a virtual
255  * address.
256  *
257  * This version of ioremap ensures that the memory is marked uncachable
258  * on the CPU as well as honouring existing caching rules from things like
259  * the PCI bus. Note that there are other caches and buffers on many
260  * busses. In particular driver authors should read up on PCI writes
261  *
262  * It's useful if some control registers are in such an area and
263  * write combining or read caching is not desirable:
264  *
265  * Must be freed with iounmap.
266  */
267 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
268 {
269         /*
270          * Ideally, this should be:
271          *      pat_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
272          *
273          * Till we fix all X drivers to use ioremap_wc(), we will use
274          * UC MINUS.
275          */
276         unsigned long val = _PAGE_CACHE_UC_MINUS;
277
278         return __ioremap_caller(phys_addr, size, val,
279                                 __builtin_return_address(0));
280 }
281 EXPORT_SYMBOL(ioremap_nocache);
282
283 /**
284  * ioremap_wc   -       map memory into CPU space write combined
285  * @offset:     bus address of the memory
286  * @size:       size of the resource to map
287  *
288  * This version of ioremap ensures that the memory is marked write combining.
289  * Write combining allows faster writes to some hardware devices.
290  *
291  * Must be freed with iounmap.
292  */
293 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
294 {
295         if (pat_enabled)
296                 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
297                                         __builtin_return_address(0));
298         else
299                 return ioremap_nocache(phys_addr, size);
300 }
301 EXPORT_SYMBOL(ioremap_wc);
302
303 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
304 {
305         return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
306                                 __builtin_return_address(0));
307 }
308 EXPORT_SYMBOL(ioremap_cache);
309
310 static void __iomem *ioremap_default(resource_size_t phys_addr,
311                                         unsigned long size)
312 {
313         unsigned long flags;
314         void *ret;
315         int err;
316
317         /*
318          * - WB for WB-able memory and no other conflicting mappings
319          * - UC_MINUS for non-WB-able memory with no other conflicting mappings
320          * - Inherit from confliting mappings otherwise
321          */
322         err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
323         if (err < 0)
324                 return NULL;
325
326         ret = (void *) __ioremap_caller(phys_addr, size, flags,
327                                         __builtin_return_address(0));
328
329         free_memtype(phys_addr, phys_addr + size);
330         return (void __iomem *)ret;
331 }
332
333 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
334                                 unsigned long prot_val)
335 {
336         return __ioremap_caller(phys_addr, size, (prot_val & _PAGE_CACHE_MASK),
337                                 __builtin_return_address(0));
338 }
339 EXPORT_SYMBOL(ioremap_prot);
340
341 /**
342  * iounmap - Free a IO remapping
343  * @addr: virtual address from ioremap_*
344  *
345  * Caller must ensure there is only one unmapping for the same pointer.
346  */
347 void iounmap(volatile void __iomem *addr)
348 {
349         struct vm_struct *p, *o;
350
351         if ((void __force *)addr <= high_memory)
352                 return;
353
354         /*
355          * __ioremap special-cases the PCI/ISA range by not instantiating a
356          * vm_area and by simply returning an address into the kernel mapping
357          * of ISA space.   So handle that here.
358          */
359         if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
360             (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
361                 return;
362
363         addr = (volatile void __iomem *)
364                 (PAGE_MASK & (unsigned long __force)addr);
365
366         mmiotrace_iounmap(addr);
367
368         /* Use the vm area unlocked, assuming the caller
369            ensures there isn't another iounmap for the same address
370            in parallel. Reuse of the virtual address is prevented by
371            leaving it in the global lists until we're done with it.
372            cpa takes care of the direct mappings. */
373         read_lock(&vmlist_lock);
374         for (p = vmlist; p; p = p->next) {
375                 if (p->addr == (void __force *)addr)
376                         break;
377         }
378         read_unlock(&vmlist_lock);
379
380         if (!p) {
381                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
382                 dump_stack();
383                 return;
384         }
385
386         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
387
388         /* Finally remove it */
389         o = remove_vm_area((void __force *)addr);
390         BUG_ON(p != o || o == NULL);
391         kfree(p);
392 }
393 EXPORT_SYMBOL(iounmap);
394
395 /*
396  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
397  * access
398  */
399 void *xlate_dev_mem_ptr(unsigned long phys)
400 {
401         void *addr;
402         unsigned long start = phys & PAGE_MASK;
403
404         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
405         if (page_is_ram(start >> PAGE_SHIFT))
406                 return __va(phys);
407
408         addr = (void __force *)ioremap_default(start, PAGE_SIZE);
409         if (addr)
410                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
411
412         return addr;
413 }
414
415 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
416 {
417         if (page_is_ram(phys >> PAGE_SHIFT))
418                 return;
419
420         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
421         return;
422 }
423
424 static int __initdata early_ioremap_debug;
425
426 static int __init early_ioremap_debug_setup(char *str)
427 {
428         early_ioremap_debug = 1;
429
430         return 0;
431 }
432 early_param("early_ioremap_debug", early_ioremap_debug_setup);
433
434 static __initdata int after_paging_init;
435 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
436
437 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
438 {
439         /* Don't assume we're using swapper_pg_dir at this point */
440         pgd_t *base = __va(read_cr3());
441         pgd_t *pgd = &base[pgd_index(addr)];
442         pud_t *pud = pud_offset(pgd, addr);
443         pmd_t *pmd = pmd_offset(pud, addr);
444
445         return pmd;
446 }
447
448 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
449 {
450         return &bm_pte[pte_index(addr)];
451 }
452
453 void __init early_ioremap_init(void)
454 {
455         pmd_t *pmd;
456
457         if (early_ioremap_debug)
458                 printk(KERN_INFO "early_ioremap_init()\n");
459
460         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
461         memset(bm_pte, 0, sizeof(bm_pte));
462         pmd_populate_kernel(&init_mm, pmd, bm_pte);
463
464         /*
465          * The boot-ioremap range spans multiple pmds, for which
466          * we are not prepared:
467          */
468         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
469                 WARN_ON(1);
470                 printk(KERN_WARNING "pmd %p != %p\n",
471                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
472                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
473                         fix_to_virt(FIX_BTMAP_BEGIN));
474                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
475                         fix_to_virt(FIX_BTMAP_END));
476
477                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
478                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
479                        FIX_BTMAP_BEGIN);
480         }
481 }
482
483 void __init early_ioremap_clear(void)
484 {
485         pmd_t *pmd;
486
487         if (early_ioremap_debug)
488                 printk(KERN_INFO "early_ioremap_clear()\n");
489
490         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
491         pmd_clear(pmd);
492         paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
493         __flush_tlb_all();
494 }
495
496 void __init early_ioremap_reset(void)
497 {
498         enum fixed_addresses idx;
499         unsigned long addr, phys;
500         pte_t *pte;
501
502         after_paging_init = 1;
503         for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
504                 addr = fix_to_virt(idx);
505                 pte = early_ioremap_pte(addr);
506                 if (pte_present(*pte)) {
507                         phys = pte_val(*pte) & PAGE_MASK;
508                         set_fixmap(idx, phys);
509                 }
510         }
511 }
512
513 static void __init __early_set_fixmap(enum fixed_addresses idx,
514                                    unsigned long phys, pgprot_t flags)
515 {
516         unsigned long addr = __fix_to_virt(idx);
517         pte_t *pte;
518
519         if (idx >= __end_of_fixed_addresses) {
520                 BUG();
521                 return;
522         }
523         pte = early_ioremap_pte(addr);
524
525         if (pgprot_val(flags))
526                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
527         else
528                 pte_clear(&init_mm, addr, pte);
529         __flush_tlb_one(addr);
530 }
531
532 static inline void __init early_set_fixmap(enum fixed_addresses idx,
533                                         unsigned long phys)
534 {
535         if (after_paging_init)
536                 set_fixmap(idx, phys);
537         else
538                 __early_set_fixmap(idx, phys, PAGE_KERNEL);
539 }
540
541 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
542 {
543         if (after_paging_init)
544                 clear_fixmap(idx);
545         else
546                 __early_set_fixmap(idx, 0, __pgprot(0));
547 }
548
549
550 static int __initdata early_ioremap_nested;
551
552 static int __init check_early_ioremap_leak(void)
553 {
554         if (!early_ioremap_nested)
555                 return 0;
556         WARN(1, KERN_WARNING
557                "Debug warning: early ioremap leak of %d areas detected.\n",
558                 early_ioremap_nested);
559         printk(KERN_WARNING
560                 "please boot with early_ioremap_debug and report the dmesg.\n");
561
562         return 1;
563 }
564 late_initcall(check_early_ioremap_leak);
565
566 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
567 {
568         unsigned long offset, last_addr;
569         unsigned int nrpages, nesting;
570         enum fixed_addresses idx0, idx;
571
572         WARN_ON(system_state != SYSTEM_BOOTING);
573
574         nesting = early_ioremap_nested;
575         if (early_ioremap_debug) {
576                 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
577                        phys_addr, size, nesting);
578                 dump_stack();
579         }
580
581         /* Don't allow wraparound or zero size */
582         last_addr = phys_addr + size - 1;
583         if (!size || last_addr < phys_addr) {
584                 WARN_ON(1);
585                 return NULL;
586         }
587
588         if (nesting >= FIX_BTMAPS_NESTING) {
589                 WARN_ON(1);
590                 return NULL;
591         }
592         early_ioremap_nested++;
593         /*
594          * Mappings have to be page-aligned
595          */
596         offset = phys_addr & ~PAGE_MASK;
597         phys_addr &= PAGE_MASK;
598         size = PAGE_ALIGN(last_addr) - phys_addr;
599
600         /*
601          * Mappings have to fit in the FIX_BTMAP area.
602          */
603         nrpages = size >> PAGE_SHIFT;
604         if (nrpages > NR_FIX_BTMAPS) {
605                 WARN_ON(1);
606                 return NULL;
607         }
608
609         /*
610          * Ok, go for it..
611          */
612         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
613         idx = idx0;
614         while (nrpages > 0) {
615                 early_set_fixmap(idx, phys_addr);
616                 phys_addr += PAGE_SIZE;
617                 --idx;
618                 --nrpages;
619         }
620         if (early_ioremap_debug)
621                 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
622
623         return (void *) (offset + fix_to_virt(idx0));
624 }
625
626 void __init early_iounmap(void *addr, unsigned long size)
627 {
628         unsigned long virt_addr;
629         unsigned long offset;
630         unsigned int nrpages;
631         enum fixed_addresses idx;
632         int nesting;
633
634         nesting = --early_ioremap_nested;
635         if (WARN_ON(nesting < 0))
636                 return;
637
638         if (early_ioremap_debug) {
639                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
640                        size, nesting);
641                 dump_stack();
642         }
643
644         virt_addr = (unsigned long)addr;
645         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
646                 WARN_ON(1);
647                 return;
648         }
649         offset = virt_addr & ~PAGE_MASK;
650         nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
651
652         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
653         while (nrpages > 0) {
654                 early_clear_fixmap(idx);
655                 --idx;
656                 --nrpages;
657         }
658 }
659
660 void __this_fixmap_does_not_exist(void)
661 {
662         WARN_ON(1);
663 }