Merge branch 'linus' into x86/pci-ioapic-boot-irq-quirks
[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
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 (is_ISA_range(phys_addr, last_addr))
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_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_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 static void __iomem *ioremap_default(resource_size_t phys_addr,
304                                         unsigned long size)
305 {
306         unsigned long flags;
307         void *ret;
308         int err;
309
310         /*
311          * - WB for WB-able memory and no other conflicting mappings
312          * - UC_MINUS for non-WB-able memory with no other conflicting mappings
313          * - Inherit from confliting mappings otherwise
314          */
315         err = reserve_memtype(phys_addr, phys_addr + size, -1, &flags);
316         if (err < 0)
317                 return NULL;
318
319         ret = (void *) __ioremap_caller(phys_addr, size, flags,
320                                         __builtin_return_address(0));
321
322         free_memtype(phys_addr, phys_addr + size);
323         return (void __iomem *)ret;
324 }
325
326 /**
327  * iounmap - Free a IO remapping
328  * @addr: virtual address from ioremap_*
329  *
330  * Caller must ensure there is only one unmapping for the same pointer.
331  */
332 void iounmap(volatile void __iomem *addr)
333 {
334         struct vm_struct *p, *o;
335
336         if ((void __force *)addr <= high_memory)
337                 return;
338
339         /*
340          * __ioremap special-cases the PCI/ISA range by not instantiating a
341          * vm_area and by simply returning an address into the kernel mapping
342          * of ISA space.   So handle that here.
343          */
344         if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
345             (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
346                 return;
347
348         addr = (volatile void __iomem *)
349                 (PAGE_MASK & (unsigned long __force)addr);
350
351         /* Use the vm area unlocked, assuming the caller
352            ensures there isn't another iounmap for the same address
353            in parallel. Reuse of the virtual address is prevented by
354            leaving it in the global lists until we're done with it.
355            cpa takes care of the direct mappings. */
356         read_lock(&vmlist_lock);
357         for (p = vmlist; p; p = p->next) {
358                 if (p->addr == (void __force *)addr)
359                         break;
360         }
361         read_unlock(&vmlist_lock);
362
363         if (!p) {
364                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
365                 dump_stack();
366                 return;
367         }
368
369         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
370
371         /* Finally remove it */
372         o = remove_vm_area((void __force *)addr);
373         BUG_ON(p != o || o == NULL);
374         kfree(p);
375 }
376 EXPORT_SYMBOL(iounmap);
377
378 /*
379  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
380  * access
381  */
382 void *xlate_dev_mem_ptr(unsigned long phys)
383 {
384         void *addr;
385         unsigned long start = phys & PAGE_MASK;
386
387         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
388         if (page_is_ram(start >> PAGE_SHIFT))
389                 return __va(phys);
390
391         addr = (void __force *)ioremap_default(start, PAGE_SIZE);
392         if (addr)
393                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
394
395         return addr;
396 }
397
398 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
399 {
400         if (page_is_ram(phys >> PAGE_SHIFT))
401                 return;
402
403         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
404         return;
405 }
406
407 int __initdata early_ioremap_debug;
408
409 static int __init early_ioremap_debug_setup(char *str)
410 {
411         early_ioremap_debug = 1;
412
413         return 0;
414 }
415 early_param("early_ioremap_debug", early_ioremap_debug_setup);
416
417 static __initdata int after_paging_init;
418 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
419
420 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
421 {
422         /* Don't assume we're using swapper_pg_dir at this point */
423         pgd_t *base = __va(read_cr3());
424         pgd_t *pgd = &base[pgd_index(addr)];
425         pud_t *pud = pud_offset(pgd, addr);
426         pmd_t *pmd = pmd_offset(pud, addr);
427
428         return pmd;
429 }
430
431 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
432 {
433         return &bm_pte[pte_index(addr)];
434 }
435
436 void __init early_ioremap_init(void)
437 {
438         pmd_t *pmd;
439
440         if (early_ioremap_debug)
441                 printk(KERN_INFO "early_ioremap_init()\n");
442
443         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
444         memset(bm_pte, 0, sizeof(bm_pte));
445         pmd_populate_kernel(&init_mm, pmd, bm_pte);
446
447         /*
448          * The boot-ioremap range spans multiple pmds, for which
449          * we are not prepared:
450          */
451         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
452                 WARN_ON(1);
453                 printk(KERN_WARNING "pmd %p != %p\n",
454                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
455                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
456                         fix_to_virt(FIX_BTMAP_BEGIN));
457                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
458                         fix_to_virt(FIX_BTMAP_END));
459
460                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
461                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
462                        FIX_BTMAP_BEGIN);
463         }
464 }
465
466 void __init early_ioremap_clear(void)
467 {
468         pmd_t *pmd;
469
470         if (early_ioremap_debug)
471                 printk(KERN_INFO "early_ioremap_clear()\n");
472
473         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
474         pmd_clear(pmd);
475         paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
476         __flush_tlb_all();
477 }
478
479 void __init early_ioremap_reset(void)
480 {
481         enum fixed_addresses idx;
482         unsigned long addr, phys;
483         pte_t *pte;
484
485         after_paging_init = 1;
486         for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
487                 addr = fix_to_virt(idx);
488                 pte = early_ioremap_pte(addr);
489                 if (pte_present(*pte)) {
490                         phys = pte_val(*pte) & PAGE_MASK;
491                         set_fixmap(idx, phys);
492                 }
493         }
494 }
495
496 static void __init __early_set_fixmap(enum fixed_addresses idx,
497                                    unsigned long phys, pgprot_t flags)
498 {
499         unsigned long addr = __fix_to_virt(idx);
500         pte_t *pte;
501
502         if (idx >= __end_of_fixed_addresses) {
503                 BUG();
504                 return;
505         }
506         pte = early_ioremap_pte(addr);
507
508         if (pgprot_val(flags))
509                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
510         else
511                 pte_clear(&init_mm, addr, pte);
512         __flush_tlb_one(addr);
513 }
514
515 static inline void __init early_set_fixmap(enum fixed_addresses idx,
516                                         unsigned long phys)
517 {
518         if (after_paging_init)
519                 set_fixmap(idx, phys);
520         else
521                 __early_set_fixmap(idx, phys, PAGE_KERNEL);
522 }
523
524 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
525 {
526         if (after_paging_init)
527                 clear_fixmap(idx);
528         else
529                 __early_set_fixmap(idx, 0, __pgprot(0));
530 }
531
532
533 int __initdata early_ioremap_nested;
534
535 static int __init check_early_ioremap_leak(void)
536 {
537         if (!early_ioremap_nested)
538                 return 0;
539
540         printk(KERN_WARNING
541                "Debug warning: early ioremap leak of %d areas detected.\n",
542                early_ioremap_nested);
543         printk(KERN_WARNING
544                "please boot with early_ioremap_debug and report the dmesg.\n");
545         WARN_ON(1);
546
547         return 1;
548 }
549 late_initcall(check_early_ioremap_leak);
550
551 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
552 {
553         unsigned long offset, last_addr;
554         unsigned int nrpages, nesting;
555         enum fixed_addresses idx0, idx;
556
557         WARN_ON(system_state != SYSTEM_BOOTING);
558
559         nesting = early_ioremap_nested;
560         if (early_ioremap_debug) {
561                 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
562                        phys_addr, size, nesting);
563                 dump_stack();
564         }
565
566         /* Don't allow wraparound or zero size */
567         last_addr = phys_addr + size - 1;
568         if (!size || last_addr < phys_addr) {
569                 WARN_ON(1);
570                 return NULL;
571         }
572
573         if (nesting >= FIX_BTMAPS_NESTING) {
574                 WARN_ON(1);
575                 return NULL;
576         }
577         early_ioremap_nested++;
578         /*
579          * Mappings have to be page-aligned
580          */
581         offset = phys_addr & ~PAGE_MASK;
582         phys_addr &= PAGE_MASK;
583         size = PAGE_ALIGN(last_addr) - phys_addr;
584
585         /*
586          * Mappings have to fit in the FIX_BTMAP area.
587          */
588         nrpages = size >> PAGE_SHIFT;
589         if (nrpages > NR_FIX_BTMAPS) {
590                 WARN_ON(1);
591                 return NULL;
592         }
593
594         /*
595          * Ok, go for it..
596          */
597         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
598         idx = idx0;
599         while (nrpages > 0) {
600                 early_set_fixmap(idx, phys_addr);
601                 phys_addr += PAGE_SIZE;
602                 --idx;
603                 --nrpages;
604         }
605         if (early_ioremap_debug)
606                 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
607
608         return (void *) (offset + fix_to_virt(idx0));
609 }
610
611 void __init early_iounmap(void *addr, unsigned long size)
612 {
613         unsigned long virt_addr;
614         unsigned long offset;
615         unsigned int nrpages;
616         enum fixed_addresses idx;
617         int nesting;
618
619         nesting = --early_ioremap_nested;
620         if (WARN_ON(nesting < 0))
621                 return;
622
623         if (early_ioremap_debug) {
624                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
625                        size, nesting);
626                 dump_stack();
627         }
628
629         virt_addr = (unsigned long)addr;
630         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
631                 WARN_ON(1);
632                 return;
633         }
634         offset = virt_addr & ~PAGE_MASK;
635         nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
636
637         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
638         while (nrpages > 0) {
639                 early_clear_fixmap(idx);
640                 --idx;
641                 --nrpages;
642         }
643 }
644
645 void __this_fixmap_does_not_exist(void)
646 {
647         WARN_ON(1);
648 }