Merge branch 'linus' into tracing/mmiotrace
[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 (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
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, 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_wc_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_wc_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 /**
311  * iounmap - Free a IO remapping
312  * @addr: virtual address from ioremap_*
313  *
314  * Caller must ensure there is only one unmapping for the same pointer.
315  */
316 void iounmap(volatile void __iomem *addr)
317 {
318         struct vm_struct *p, *o;
319
320         if ((void __force *)addr <= high_memory)
321                 return;
322
323         /*
324          * __ioremap special-cases the PCI/ISA range by not instantiating a
325          * vm_area and by simply returning an address into the kernel mapping
326          * of ISA space.   So handle that here.
327          */
328         if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
329             addr < phys_to_virt(ISA_END_ADDRESS))
330                 return;
331
332         addr = (volatile void __iomem *)
333                 (PAGE_MASK & (unsigned long __force)addr);
334
335         mmiotrace_iounmap(addr);
336
337         /* Use the vm area unlocked, assuming the caller
338            ensures there isn't another iounmap for the same address
339            in parallel. Reuse of the virtual address is prevented by
340            leaving it in the global lists until we're done with it.
341            cpa takes care of the direct mappings. */
342         read_lock(&vmlist_lock);
343         for (p = vmlist; p; p = p->next) {
344                 if (p->addr == addr)
345                         break;
346         }
347         read_unlock(&vmlist_lock);
348
349         if (!p) {
350                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
351                 dump_stack();
352                 return;
353         }
354
355         free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
356
357         /* Finally remove it */
358         o = remove_vm_area((void *)addr);
359         BUG_ON(p != o || o == NULL);
360         kfree(p);
361 }
362 EXPORT_SYMBOL(iounmap);
363
364 /*
365  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
366  * access
367  */
368 void *xlate_dev_mem_ptr(unsigned long phys)
369 {
370         void *addr;
371         unsigned long start = phys & PAGE_MASK;
372
373         /* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
374         if (page_is_ram(start >> PAGE_SHIFT))
375                 return __va(phys);
376
377         addr = (void *)ioremap(start, PAGE_SIZE);
378         if (addr)
379                 addr = (void *)((unsigned long)addr | (phys & ~PAGE_MASK));
380
381         return addr;
382 }
383
384 void unxlate_dev_mem_ptr(unsigned long phys, void *addr)
385 {
386         if (page_is_ram(phys >> PAGE_SHIFT))
387                 return;
388
389         iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
390         return;
391 }
392
393 #ifdef CONFIG_X86_32
394
395 int __initdata early_ioremap_debug;
396
397 static int __init early_ioremap_debug_setup(char *str)
398 {
399         early_ioremap_debug = 1;
400
401         return 0;
402 }
403 early_param("early_ioremap_debug", early_ioremap_debug_setup);
404
405 static __initdata int after_paging_init;
406 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)]
407                 __section(.bss.page_aligned);
408
409 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
410 {
411         /* Don't assume we're using swapper_pg_dir at this point */
412         pgd_t *base = __va(read_cr3());
413         pgd_t *pgd = &base[pgd_index(addr)];
414         pud_t *pud = pud_offset(pgd, addr);
415         pmd_t *pmd = pmd_offset(pud, addr);
416
417         return pmd;
418 }
419
420 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
421 {
422         return &bm_pte[pte_index(addr)];
423 }
424
425 void __init early_ioremap_init(void)
426 {
427         pmd_t *pmd;
428
429         if (early_ioremap_debug)
430                 printk(KERN_INFO "early_ioremap_init()\n");
431
432         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
433         memset(bm_pte, 0, sizeof(bm_pte));
434         pmd_populate_kernel(&init_mm, pmd, bm_pte);
435
436         /*
437          * The boot-ioremap range spans multiple pmds, for which
438          * we are not prepared:
439          */
440         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
441                 WARN_ON(1);
442                 printk(KERN_WARNING "pmd %p != %p\n",
443                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
444                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
445                         fix_to_virt(FIX_BTMAP_BEGIN));
446                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
447                         fix_to_virt(FIX_BTMAP_END));
448
449                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
450                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
451                        FIX_BTMAP_BEGIN);
452         }
453 }
454
455 void __init early_ioremap_clear(void)
456 {
457         pmd_t *pmd;
458
459         if (early_ioremap_debug)
460                 printk(KERN_INFO "early_ioremap_clear()\n");
461
462         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
463         pmd_clear(pmd);
464         paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
465         __flush_tlb_all();
466 }
467
468 void __init early_ioremap_reset(void)
469 {
470         enum fixed_addresses idx;
471         unsigned long addr, phys;
472         pte_t *pte;
473
474         after_paging_init = 1;
475         for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
476                 addr = fix_to_virt(idx);
477                 pte = early_ioremap_pte(addr);
478                 if (pte_present(*pte)) {
479                         phys = pte_val(*pte) & PAGE_MASK;
480                         set_fixmap(idx, phys);
481                 }
482         }
483 }
484
485 static void __init __early_set_fixmap(enum fixed_addresses idx,
486                                    unsigned long phys, pgprot_t flags)
487 {
488         unsigned long addr = __fix_to_virt(idx);
489         pte_t *pte;
490
491         if (idx >= __end_of_fixed_addresses) {
492                 BUG();
493                 return;
494         }
495         pte = early_ioremap_pte(addr);
496         if (pgprot_val(flags))
497                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
498         else
499                 pte_clear(NULL, addr, pte);
500         __flush_tlb_one(addr);
501 }
502
503 static inline void __init early_set_fixmap(enum fixed_addresses idx,
504                                         unsigned long phys)
505 {
506         if (after_paging_init)
507                 set_fixmap(idx, phys);
508         else
509                 __early_set_fixmap(idx, phys, PAGE_KERNEL);
510 }
511
512 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
513 {
514         if (after_paging_init)
515                 clear_fixmap(idx);
516         else
517                 __early_set_fixmap(idx, 0, __pgprot(0));
518 }
519
520
521 int __initdata early_ioremap_nested;
522
523 static int __init check_early_ioremap_leak(void)
524 {
525         if (!early_ioremap_nested)
526                 return 0;
527
528         printk(KERN_WARNING
529                "Debug warning: early ioremap leak of %d areas detected.\n",
530                early_ioremap_nested);
531         printk(KERN_WARNING
532                "please boot with early_ioremap_debug and report the dmesg.\n");
533         WARN_ON(1);
534
535         return 1;
536 }
537 late_initcall(check_early_ioremap_leak);
538
539 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
540 {
541         unsigned long offset, last_addr;
542         unsigned int nrpages, nesting;
543         enum fixed_addresses idx0, idx;
544
545         WARN_ON(system_state != SYSTEM_BOOTING);
546
547         nesting = early_ioremap_nested;
548         if (early_ioremap_debug) {
549                 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
550                        phys_addr, size, nesting);
551                 dump_stack();
552         }
553
554         /* Don't allow wraparound or zero size */
555         last_addr = phys_addr + size - 1;
556         if (!size || last_addr < phys_addr) {
557                 WARN_ON(1);
558                 return NULL;
559         }
560
561         if (nesting >= FIX_BTMAPS_NESTING) {
562                 WARN_ON(1);
563                 return NULL;
564         }
565         early_ioremap_nested++;
566         /*
567          * Mappings have to be page-aligned
568          */
569         offset = phys_addr & ~PAGE_MASK;
570         phys_addr &= PAGE_MASK;
571         size = PAGE_ALIGN(last_addr) - phys_addr;
572
573         /*
574          * Mappings have to fit in the FIX_BTMAP area.
575          */
576         nrpages = size >> PAGE_SHIFT;
577         if (nrpages > NR_FIX_BTMAPS) {
578                 WARN_ON(1);
579                 return NULL;
580         }
581
582         /*
583          * Ok, go for it..
584          */
585         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
586         idx = idx0;
587         while (nrpages > 0) {
588                 early_set_fixmap(idx, phys_addr);
589                 phys_addr += PAGE_SIZE;
590                 --idx;
591                 --nrpages;
592         }
593         if (early_ioremap_debug)
594                 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
595
596         return (void *) (offset + fix_to_virt(idx0));
597 }
598
599 void __init early_iounmap(void *addr, unsigned long size)
600 {
601         unsigned long virt_addr;
602         unsigned long offset;
603         unsigned int nrpages;
604         enum fixed_addresses idx;
605         int nesting;
606
607         nesting = --early_ioremap_nested;
608         if (WARN_ON(nesting < 0))
609                 return;
610
611         if (early_ioremap_debug) {
612                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
613                        size, nesting);
614                 dump_stack();
615         }
616
617         virt_addr = (unsigned long)addr;
618         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
619                 WARN_ON(1);
620                 return;
621         }
622         offset = virt_addr & ~PAGE_MASK;
623         nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
624
625         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
626         while (nrpages > 0) {
627                 early_clear_fixmap(idx);
628                 --idx;
629                 --nrpages;
630         }
631 }
632
633 void __this_fixmap_does_not_exist(void)
634 {
635         WARN_ON(1);
636 }
637
638 #endif /* CONFIG_X86_32 */