Merge branch 'linus' into x86/threadinfo
[linux-2.6] / 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
16 #include <asm/cacheflush.h>
17 #include <asm/e820.h>
18 #include <asm/fixmap.h>
19 #include <asm/pgtable.h>
20 #include <asm/tlbflush.h>
21 #include <asm/pgalloc.h>
22 #include <asm/pat.h>
23
24 #ifdef CONFIG_X86_64
25
26 unsigned long __phys_addr(unsigned long x)
27 {
28         if (x >= __START_KERNEL_map)
29                 return x - __START_KERNEL_map + phys_base;
30         return x - PAGE_OFFSET;
31 }
32 EXPORT_SYMBOL(__phys_addr);
33
34 static inline int phys_addr_valid(unsigned long addr)
35 {
36         return addr < (1UL << boot_cpu_data.x86_phys_bits);
37 }
38
39 #else
40
41 static inline int phys_addr_valid(unsigned long addr)
42 {
43         return 1;
44 }
45
46 #endif
47
48 int page_is_ram(unsigned long pagenr)
49 {
50         resource_size_t addr, end;
51         int i;
52
53         /*
54          * A special case is the first 4Kb of memory;
55          * This is a BIOS owned area, not kernel ram, but generally
56          * not listed as such in the E820 table.
57          */
58         if (pagenr == 0)
59                 return 0;
60
61         /*
62          * Second special case: Some BIOSen report the PC BIOS
63          * area (640->1Mb) as ram even though it is not.
64          */
65         if (pagenr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
66                     pagenr < (BIOS_END >> PAGE_SHIFT))
67                 return 0;
68
69         for (i = 0; i < e820.nr_map; i++) {
70                 /*
71                  * Not usable memory:
72                  */
73                 if (e820.map[i].type != E820_RAM)
74                         continue;
75                 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
76                 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
77
78
79                 if ((pagenr >= addr) && (pagenr < end))
80                         return 1;
81         }
82         return 0;
83 }
84
85 /*
86  * Fix up the linear direct mapping of the kernel to avoid cache attribute
87  * conflicts.
88  */
89 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
90                                unsigned long prot_val)
91 {
92         unsigned long nrpages = size >> PAGE_SHIFT;
93         int err;
94
95         switch (prot_val) {
96         case _PAGE_CACHE_UC:
97         default:
98                 err = _set_memory_uc(vaddr, nrpages);
99                 break;
100         case _PAGE_CACHE_WC:
101                 err = _set_memory_wc(vaddr, nrpages);
102                 break;
103         case _PAGE_CACHE_WB:
104                 err = _set_memory_wb(vaddr, nrpages);
105                 break;
106         }
107
108         return err;
109 }
110
111 /*
112  * Remap an arbitrary physical address space into the kernel virtual
113  * address space. Needed when the kernel wants to access high addresses
114  * directly.
115  *
116  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
117  * have to convert them into an offset in a page-aligned mapping, but the
118  * caller shouldn't need to know that small detail.
119  */
120 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
121                 unsigned long size, unsigned long prot_val, void *caller)
122 {
123         unsigned long pfn, offset, vaddr;
124         resource_size_t last_addr;
125         struct vm_struct *area;
126         unsigned long new_prot_val;
127         pgprot_t prot;
128         int retval;
129
130         /* Don't allow wraparound or zero size */
131         last_addr = phys_addr + size - 1;
132         if (!size || last_addr < phys_addr)
133                 return NULL;
134
135         if (!phys_addr_valid(phys_addr)) {
136                 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
137                        (unsigned long long)phys_addr);
138                 WARN_ON_ONCE(1);
139                 return NULL;
140         }
141
142         /*
143          * Don't remap the low PCI/ISA area, it's always mapped..
144          */
145         if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
146                 return (__force void __iomem *)phys_to_virt(phys_addr);
147
148         /*
149          * Don't allow anybody to remap normal RAM that we're using..
150          */
151         for (pfn = phys_addr >> PAGE_SHIFT;
152                                 (pfn << PAGE_SHIFT) < (last_addr & PAGE_MASK);
153                                 pfn++) {
154
155                 int is_ram = page_is_ram(pfn);
156
157                 if (is_ram && pfn_valid(pfn) && !PageReserved(pfn_to_page(pfn)))
158                         return NULL;
159                 WARN_ON_ONCE(is_ram);
160         }
161
162         /*
163          * Mappings have to be page-aligned
164          */
165         offset = phys_addr & ~PAGE_MASK;
166         phys_addr &= PAGE_MASK;
167         size = PAGE_ALIGN(last_addr+1) - phys_addr;
168
169         retval = reserve_memtype(phys_addr, phys_addr + size,
170                                                 prot_val, &new_prot_val);
171         if (retval) {
172                 pr_debug("Warning: reserve_memtype returned %d\n", retval);
173                 return NULL;
174         }
175
176         if (prot_val != new_prot_val) {
177                 /*
178                  * Do not fallback to certain memory types with certain
179                  * requested type:
180                  * - request is uc-, return cannot be write-back
181                  * - request is uc-, return cannot be write-combine
182                  * - request is write-combine, return cannot be write-back
183                  */
184                 if ((prot_val == _PAGE_CACHE_UC_MINUS &&
185                      (new_prot_val == _PAGE_CACHE_WB ||
186                       new_prot_val == _PAGE_CACHE_WC)) ||
187                     (prot_val == _PAGE_CACHE_WC &&
188                      new_prot_val == _PAGE_CACHE_WB)) {
189                         pr_debug(
190                 "ioremap error for 0x%llx-0x%llx, requested 0x%lx, got 0x%lx\n",
191                                 (unsigned long long)phys_addr,
192                                 (unsigned long long)(phys_addr + size),
193                                 prot_val, new_prot_val);
194                         free_memtype(phys_addr, phys_addr + size);
195                         return NULL;
196                 }
197                 prot_val = new_prot_val;
198         }
199
200         switch (prot_val) {
201         case _PAGE_CACHE_UC:
202         default:
203                 prot = PAGE_KERNEL_NOCACHE;
204                 break;
205         case _PAGE_CACHE_UC_MINUS:
206                 prot = PAGE_KERNEL_UC_MINUS;
207                 break;
208         case _PAGE_CACHE_WC:
209                 prot = PAGE_KERNEL_WC;
210                 break;
211         case _PAGE_CACHE_WB:
212                 prot = PAGE_KERNEL;
213                 break;
214         }
215
216         /*
217          * Ok, go for it..
218          */
219         area = get_vm_area_caller(size, VM_IOREMAP, caller);
220         if (!area)
221                 return NULL;
222         area->phys_addr = phys_addr;
223         vaddr = (unsigned long) area->addr;
224         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
225                 free_memtype(phys_addr, phys_addr + size);
226                 free_vm_area(area);
227                 return NULL;
228         }
229
230         if (ioremap_change_attr(vaddr, size, prot_val) < 0) {
231                 free_memtype(phys_addr, phys_addr + size);
232                 vunmap(area->addr);
233                 return NULL;
234         }
235
236         return (void __iomem *) (vaddr + offset);
237 }
238
239 /**
240  * ioremap_nocache     -   map bus memory into CPU space
241  * @offset:    bus address of the memory
242  * @size:      size of the resource to map
243  *
244  * ioremap_nocache performs a platform specific sequence of operations to
245  * make bus memory CPU accessible via the readb/readw/readl/writeb/
246  * writew/writel functions and the other mmio helpers. The returned
247  * address is not guaranteed to be usable directly as a virtual
248  * address.
249  *
250  * This version of ioremap ensures that the memory is marked uncachable
251  * on the CPU as well as honouring existing caching rules from things like
252  * the PCI bus. Note that there are other caches and buffers on many
253  * busses. In particular driver authors should read up on PCI writes
254  *
255  * It's useful if some control registers are in such an area and
256  * write combining or read caching is not desirable:
257  *
258  * Must be freed with iounmap.
259  */
260 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
261 {
262         /*
263          * Ideally, this should be:
264          *      pat_wc_enabled ? _PAGE_CACHE_UC : _PAGE_CACHE_UC_MINUS;
265          *
266          * Till we fix all X drivers to use ioremap_wc(), we will use
267          * UC MINUS.
268          */
269         unsigned long val = _PAGE_CACHE_UC_MINUS;
270
271         return __ioremap_caller(phys_addr, size, val,
272                                 __builtin_return_address(0));
273 }
274 EXPORT_SYMBOL(ioremap_nocache);
275
276 /**
277  * ioremap_wc   -       map memory into CPU space write combined
278  * @offset:     bus address of the memory
279  * @size:       size of the resource to map
280  *
281  * This version of ioremap ensures that the memory is marked write combining.
282  * Write combining allows faster writes to some hardware devices.
283  *
284  * Must be freed with iounmap.
285  */
286 void __iomem *ioremap_wc(unsigned long phys_addr, unsigned long size)
287 {
288         if (pat_wc_enabled)
289                 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WC,
290                                         __builtin_return_address(0));
291         else
292                 return ioremap_nocache(phys_addr, size);
293 }
294 EXPORT_SYMBOL(ioremap_wc);
295
296 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
297 {
298         return __ioremap_caller(phys_addr, size, _PAGE_CACHE_WB,
299                                 __builtin_return_address(0));
300 }
301 EXPORT_SYMBOL(ioremap_cache);
302
303 /**
304  * iounmap - Free a IO remapping
305  * @addr: virtual address from ioremap_*
306  *
307  * Caller must ensure there is only one unmapping for the same pointer.
308  */
309 void iounmap(volatile void __iomem *addr)
310 {
311         struct vm_struct *p, *o;
312
313         if ((void __force *)addr <= high_memory)
314                 return;
315
316         /*
317          * __ioremap special-cases the PCI/ISA range by not instantiating a
318          * vm_area and by simply returning an address into the kernel mapping
319          * of ISA space.   So handle that here.
320          */
321         if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
322             addr < phys_to_virt(ISA_END_ADDRESS))
323                 return;
324
325         addr = (volatile void __iomem *)
326                 (PAGE_MASK & (unsigned long __force)addr);
327
328         /* Use the vm area unlocked, assuming the caller
329            ensures there isn't another iounmap for the same address
330            in parallel. Reuse of the virtual address is prevented by
331            leaving it in the global lists until we're done with it.
332            cpa takes care of the direct mappings. */
333         read_lock(&vmlist_lock);
334         for (p = vmlist; p; p = p->next) {
335                 if (p->addr == addr)
336                         break;
337         }
338         read_unlock(&vmlist_lock);
339
340         if (!p) {
341                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
342                 dump_stack();
343                 return;
344         }
345
346         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
347
348         /* Finally remove it */
349         o = remove_vm_area((void *)addr);
350         BUG_ON(p != o || o == NULL);
351         kfree(p);
352 }
353 EXPORT_SYMBOL(iounmap);
354
355 /*
356  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
357  * access
358  */
359 void *xlate_dev_mem_ptr(unsigned long phys)
360 {
361         void *addr;
362         unsigned long start = phys & PAGE_MASK;
363
364         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
365         if (page_is_ram(start >> PAGE_SHIFT))
366                 return __va(phys);
367
368         addr = (void *)ioremap(start, PAGE_SIZE);
369         if (addr)
370                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
371
372         return addr;
373 }
374
375 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
376 {
377         if (page_is_ram(phys >> PAGE_SHIFT))
378                 return;
379
380         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
381         return;
382 }
383
384 #ifdef CONFIG_X86_32
385
386 int __initdata early_ioremap_debug;
387
388 static int __init early_ioremap_debug_setup(char *str)
389 {
390         early_ioremap_debug = 1;
391
392         return 0;
393 }
394 early_param("early_ioremap_debug", early_ioremap_debug_setup);
395
396 static __initdata int after_paging_init;
397 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)]
398                 __section(.bss.page_aligned);
399
400 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
401 {
402         /* Don't assume we're using swapper_pg_dir at this point */
403         pgd_t *base = __va(read_cr3());
404         pgd_t *pgd = &base[pgd_index(addr)];
405         pud_t *pud = pud_offset(pgd, addr);
406         pmd_t *pmd = pmd_offset(pud, addr);
407
408         return pmd;
409 }
410
411 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
412 {
413         return &bm_pte[pte_index(addr)];
414 }
415
416 void __init early_ioremap_init(void)
417 {
418         pmd_t *pmd;
419
420         if (early_ioremap_debug)
421                 printk(KERN_INFO "early_ioremap_init()\n");
422
423         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
424         memset(bm_pte, 0, sizeof(bm_pte));
425         pmd_populate_kernel(&init_mm, pmd, bm_pte);
426
427         /*
428          * The boot-ioremap range spans multiple pmds, for which
429          * we are not prepared:
430          */
431         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
432                 WARN_ON(1);
433                 printk(KERN_WARNING "pmd %p != %p\n",
434                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
435                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
436                         fix_to_virt(FIX_BTMAP_BEGIN));
437                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
438                         fix_to_virt(FIX_BTMAP_END));
439
440                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
441                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
442                        FIX_BTMAP_BEGIN);
443         }
444 }
445
446 void __init early_ioremap_clear(void)
447 {
448         pmd_t *pmd;
449
450         if (early_ioremap_debug)
451                 printk(KERN_INFO "early_ioremap_clear()\n");
452
453         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
454         pmd_clear(pmd);
455         paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
456         __flush_tlb_all();
457 }
458
459 void __init early_ioremap_reset(void)
460 {
461         enum fixed_addresses idx;
462         unsigned long addr, phys;
463         pte_t *pte;
464
465         after_paging_init = 1;
466         for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
467                 addr = fix_to_virt(idx);
468                 pte = early_ioremap_pte(addr);
469                 if (pte_present(*pte)) {
470                         phys = pte_val(*pte) & PAGE_MASK;
471                         set_fixmap(idx, phys);
472                 }
473         }
474 }
475
476 static void __init __early_set_fixmap(enum fixed_addresses idx,
477                                    unsigned long phys, pgprot_t flags)
478 {
479         unsigned long addr = __fix_to_virt(idx);
480         pte_t *pte;
481
482         if (idx >= __end_of_fixed_addresses) {
483                 BUG();
484                 return;
485         }
486         pte = early_ioremap_pte(addr);
487         if (pgprot_val(flags))
488                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
489         else
490                 pte_clear(NULL, addr, pte);
491         __flush_tlb_one(addr);
492 }
493
494 static inline void __init early_set_fixmap(enum fixed_addresses idx,
495                                         unsigned long phys)
496 {
497         if (after_paging_init)
498                 set_fixmap(idx, phys);
499         else
500                 __early_set_fixmap(idx, phys, PAGE_KERNEL);
501 }
502
503 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
504 {
505         if (after_paging_init)
506                 clear_fixmap(idx);
507         else
508                 __early_set_fixmap(idx, 0, __pgprot(0));
509 }
510
511
512 int __initdata early_ioremap_nested;
513
514 static int __init check_early_ioremap_leak(void)
515 {
516         if (!early_ioremap_nested)
517                 return 0;
518
519         printk(KERN_WARNING
520                "Debug warning: early ioremap leak of %d areas detected.\n",
521                early_ioremap_nested);
522         printk(KERN_WARNING
523                "please boot with early_ioremap_debug and report the dmesg.\n");
524         WARN_ON(1);
525
526         return 1;
527 }
528 late_initcall(check_early_ioremap_leak);
529
530 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
531 {
532         unsigned long offset, last_addr;
533         unsigned int nrpages, nesting;
534         enum fixed_addresses idx0, idx;
535
536         WARN_ON(system_state != SYSTEM_BOOTING);
537
538         nesting = early_ioremap_nested;
539         if (early_ioremap_debug) {
540                 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
541                        phys_addr, size, nesting);
542                 dump_stack();
543         }
544
545         /* Don't allow wraparound or zero size */
546         last_addr = phys_addr + size - 1;
547         if (!size || last_addr < phys_addr) {
548                 WARN_ON(1);
549                 return NULL;
550         }
551
552         if (nesting >= FIX_BTMAPS_NESTING) {
553                 WARN_ON(1);
554                 return NULL;
555         }
556         early_ioremap_nested++;
557         /*
558          * Mappings have to be page-aligned
559          */
560         offset = phys_addr & ~PAGE_MASK;
561         phys_addr &= PAGE_MASK;
562         size = PAGE_ALIGN(last_addr) - phys_addr;
563
564         /*
565          * Mappings have to fit in the FIX_BTMAP area.
566          */
567         nrpages = size >> PAGE_SHIFT;
568         if (nrpages > NR_FIX_BTMAPS) {
569                 WARN_ON(1);
570                 return NULL;
571         }
572
573         /*
574          * Ok, go for it..
575          */
576         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
577         idx = idx0;
578         while (nrpages > 0) {
579                 early_set_fixmap(idx, phys_addr);
580                 phys_addr += PAGE_SIZE;
581                 --idx;
582                 --nrpages;
583         }
584         if (early_ioremap_debug)
585                 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
586
587         return (void *) (offset + fix_to_virt(idx0));
588 }
589
590 void __init early_iounmap(void *addr, unsigned long size)
591 {
592         unsigned long virt_addr;
593         unsigned long offset;
594         unsigned int nrpages;
595         enum fixed_addresses idx;
596         int nesting;
597
598         nesting = --early_ioremap_nested;
599         if (WARN_ON(nesting < 0))
600                 return;
601
602         if (early_ioremap_debug) {
603                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
604                        size, nesting);
605                 dump_stack();
606         }
607
608         virt_addr = (unsigned long)addr;
609         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
610                 WARN_ON(1);
611                 return;
612         }
613         offset = virt_addr & ~PAGE_MASK;
614         nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
615
616         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
617         while (nrpages > 0) {
618                 early_clear_fixmap(idx);
619                 --idx;
620                 --nrpages;
621         }
622 }
623
624 void __this_fixmap_does_not_exist(void)
625 {
626         WARN_ON(1);
627 }
628
629 #endif /* CONFIG_X86_32 */