x86: construct 32-bit boot time page tables in native format.
[linux-3.10.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
23 enum ioremap_mode {
24         IOR_MODE_UNCACHED,
25         IOR_MODE_CACHED,
26 };
27
28 #ifdef CONFIG_X86_64
29
30 unsigned long __phys_addr(unsigned long x)
31 {
32         if (x >= __START_KERNEL_map)
33                 return x - __START_KERNEL_map + phys_base;
34         return x - PAGE_OFFSET;
35 }
36 EXPORT_SYMBOL(__phys_addr);
37
38 #endif
39
40 int page_is_ram(unsigned long pagenr)
41 {
42         unsigned long addr, end;
43         int i;
44
45         for (i = 0; i < e820.nr_map; i++) {
46                 /*
47                  * Not usable memory:
48                  */
49                 if (e820.map[i].type != E820_RAM)
50                         continue;
51                 addr = (e820.map[i].addr + PAGE_SIZE-1) >> PAGE_SHIFT;
52                 end = (e820.map[i].addr + e820.map[i].size) >> PAGE_SHIFT;
53
54                 /*
55                  * Sanity check: Some BIOSen report areas as RAM that
56                  * are not. Notably the 640->1Mb area, which is the
57                  * PCI BIOS area.
58                  */
59                 if (addr >= (BIOS_BEGIN >> PAGE_SHIFT) &&
60                     end < (BIOS_END >> PAGE_SHIFT))
61                         continue;
62
63                 if ((pagenr >= addr) && (pagenr < end))
64                         return 1;
65         }
66         return 0;
67 }
68
69 /*
70  * Fix up the linear direct mapping of the kernel to avoid cache attribute
71  * conflicts.
72  */
73 static int ioremap_change_attr(unsigned long vaddr, unsigned long size,
74                                enum ioremap_mode mode)
75 {
76         unsigned long nrpages = size >> PAGE_SHIFT;
77         int err;
78
79         switch (mode) {
80         case IOR_MODE_UNCACHED:
81         default:
82                 err = set_memory_uc(vaddr, nrpages);
83                 break;
84         case IOR_MODE_CACHED:
85                 err = set_memory_wb(vaddr, nrpages);
86                 break;
87         }
88
89         return err;
90 }
91
92 /*
93  * Remap an arbitrary physical address space into the kernel virtual
94  * address space. Needed when the kernel wants to access high addresses
95  * directly.
96  *
97  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
98  * have to convert them into an offset in a page-aligned mapping, but the
99  * caller shouldn't need to know that small detail.
100  */
101 static void __iomem *__ioremap(unsigned long phys_addr, unsigned long size,
102                                enum ioremap_mode mode)
103 {
104         unsigned long pfn, offset, last_addr, vaddr;
105         struct vm_struct *area;
106         pgprot_t prot;
107
108         /* Don't allow wraparound or zero size */
109         last_addr = phys_addr + size - 1;
110         if (!size || last_addr < phys_addr)
111                 return NULL;
112
113         /*
114          * Don't remap the low PCI/ISA area, it's always mapped..
115          */
116         if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
117                 return (__force void __iomem *)phys_to_virt(phys_addr);
118
119         /*
120          * Don't allow anybody to remap normal RAM that we're using..
121          */
122         for (pfn = phys_addr >> PAGE_SHIFT; pfn < max_pfn_mapped &&
123              (pfn << PAGE_SHIFT) < last_addr; pfn++) {
124                 if (page_is_ram(pfn) && pfn_valid(pfn) &&
125                     !PageReserved(pfn_to_page(pfn)))
126                         return NULL;
127         }
128
129         switch (mode) {
130         case IOR_MODE_UNCACHED:
131         default:
132                 prot = PAGE_KERNEL_NOCACHE;
133                 break;
134         case IOR_MODE_CACHED:
135                 prot = PAGE_KERNEL;
136                 break;
137         }
138
139         /*
140          * Mappings have to be page-aligned
141          */
142         offset = phys_addr & ~PAGE_MASK;
143         phys_addr &= PAGE_MASK;
144         size = PAGE_ALIGN(last_addr+1) - phys_addr;
145
146         /*
147          * Ok, go for it..
148          */
149         area = get_vm_area(size, VM_IOREMAP);
150         if (!area)
151                 return NULL;
152         area->phys_addr = phys_addr;
153         vaddr = (unsigned long) area->addr;
154         if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) {
155                 remove_vm_area((void *)(vaddr & PAGE_MASK));
156                 return NULL;
157         }
158
159         if (ioremap_change_attr(vaddr, size, mode) < 0) {
160                 vunmap(area->addr);
161                 return NULL;
162         }
163
164         return (void __iomem *) (vaddr + offset);
165 }
166
167 /**
168  * ioremap_nocache     -   map bus memory into CPU space
169  * @offset:    bus address of the memory
170  * @size:      size of the resource to map
171  *
172  * ioremap_nocache performs a platform specific sequence of operations to
173  * make bus memory CPU accessible via the readb/readw/readl/writeb/
174  * writew/writel functions and the other mmio helpers. The returned
175  * address is not guaranteed to be usable directly as a virtual
176  * address.
177  *
178  * This version of ioremap ensures that the memory is marked uncachable
179  * on the CPU as well as honouring existing caching rules from things like
180  * the PCI bus. Note that there are other caches and buffers on many
181  * busses. In particular driver authors should read up on PCI writes
182  *
183  * It's useful if some control registers are in such an area and
184  * write combining or read caching is not desirable:
185  *
186  * Must be freed with iounmap.
187  */
188 void __iomem *ioremap_nocache(unsigned long phys_addr, unsigned long size)
189 {
190         return __ioremap(phys_addr, size, IOR_MODE_UNCACHED);
191 }
192 EXPORT_SYMBOL(ioremap_nocache);
193
194 void __iomem *ioremap_cache(unsigned long phys_addr, unsigned long size)
195 {
196         return __ioremap(phys_addr, size, IOR_MODE_CACHED);
197 }
198 EXPORT_SYMBOL(ioremap_cache);
199
200 /**
201  * iounmap - Free a IO remapping
202  * @addr: virtual address from ioremap_*
203  *
204  * Caller must ensure there is only one unmapping for the same pointer.
205  */
206 void iounmap(volatile void __iomem *addr)
207 {
208         struct vm_struct *p, *o;
209
210         if ((void __force *)addr <= high_memory)
211                 return;
212
213         /*
214          * __ioremap special-cases the PCI/ISA range by not instantiating a
215          * vm_area and by simply returning an address into the kernel mapping
216          * of ISA space.   So handle that here.
217          */
218         if (addr >= phys_to_virt(ISA_START_ADDRESS) &&
219             addr < phys_to_virt(ISA_END_ADDRESS))
220                 return;
221
222         addr = (volatile void __iomem *)
223                 (PAGE_MASK & (unsigned long __force)addr);
224
225         /* Use the vm area unlocked, assuming the caller
226            ensures there isn't another iounmap for the same address
227            in parallel. Reuse of the virtual address is prevented by
228            leaving it in the global lists until we're done with it.
229            cpa takes care of the direct mappings. */
230         read_lock(&vmlist_lock);
231         for (p = vmlist; p; p = p->next) {
232                 if (p->addr == addr)
233                         break;
234         }
235         read_unlock(&vmlist_lock);
236
237         if (!p) {
238                 printk(KERN_ERR "iounmap: bad address %p\n", addr);
239                 dump_stack();
240                 return;
241         }
242
243         /* Finally remove it */
244         o = remove_vm_area((void *)addr);
245         BUG_ON(p != o || o == NULL);
246         kfree(p);
247 }
248 EXPORT_SYMBOL(iounmap);
249
250 #ifdef CONFIG_X86_32
251
252 int __initdata early_ioremap_debug;
253
254 static int __init early_ioremap_debug_setup(char *str)
255 {
256         early_ioremap_debug = 1;
257
258         return 0;
259 }
260 early_param("early_ioremap_debug", early_ioremap_debug_setup);
261
262 static __initdata int after_paging_init;
263 static __initdata pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)]
264                                 __attribute__((aligned(PAGE_SIZE)));
265
266 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
267 {
268         pgd_t *pgd = &swapper_pg_dir[pgd_index(addr)];
269         pud_t *pud = pud_offset(pgd, addr);
270         pmd_t *pmd = pmd_offset(pud, addr);
271
272         return pmd;
273 }
274
275 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
276 {
277         return &bm_pte[pte_index(addr)];
278 }
279
280 void __init early_ioremap_init(void)
281 {
282         pmd_t *pmd;
283
284         if (early_ioremap_debug)
285                 printk(KERN_INFO "early_ioremap_init()\n");
286
287         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
288         memset(bm_pte, 0, sizeof(bm_pte));
289         set_pmd(pmd, __pmd(__pa(bm_pte) | _PAGE_TABLE));
290
291         /*
292          * The boot-ioremap range spans multiple pmds, for which
293          * we are not prepared:
294          */
295         if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
296                 WARN_ON(1);
297                 printk(KERN_WARNING "pmd %p != %p\n",
298                        pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
299                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
300                         fix_to_virt(FIX_BTMAP_BEGIN));
301                 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
302                         fix_to_virt(FIX_BTMAP_END));
303
304                 printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
305                 printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
306                        FIX_BTMAP_BEGIN);
307         }
308 }
309
310 void __init early_ioremap_clear(void)
311 {
312         pmd_t *pmd;
313
314         if (early_ioremap_debug)
315                 printk(KERN_INFO "early_ioremap_clear()\n");
316
317         pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
318         pmd_clear(pmd);
319         paravirt_release_pt(__pa(pmd) >> PAGE_SHIFT);
320         __flush_tlb_all();
321 }
322
323 void __init early_ioremap_reset(void)
324 {
325         enum fixed_addresses idx;
326         unsigned long addr, phys;
327         pte_t *pte;
328
329         after_paging_init = 1;
330         for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
331                 addr = fix_to_virt(idx);
332                 pte = early_ioremap_pte(addr);
333                 if (pte_present(*pte)) {
334                         phys = pte_val(*pte) & PAGE_MASK;
335                         set_fixmap(idx, phys);
336                 }
337         }
338 }
339
340 static void __init __early_set_fixmap(enum fixed_addresses idx,
341                                    unsigned long phys, pgprot_t flags)
342 {
343         unsigned long addr = __fix_to_virt(idx);
344         pte_t *pte;
345
346         if (idx >= __end_of_fixed_addresses) {
347                 BUG();
348                 return;
349         }
350         pte = early_ioremap_pte(addr);
351         if (pgprot_val(flags))
352                 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
353         else
354                 pte_clear(NULL, addr, pte);
355         __flush_tlb_one(addr);
356 }
357
358 static inline void __init early_set_fixmap(enum fixed_addresses idx,
359                                         unsigned long phys)
360 {
361         if (after_paging_init)
362                 set_fixmap(idx, phys);
363         else
364                 __early_set_fixmap(idx, phys, PAGE_KERNEL);
365 }
366
367 static inline void __init early_clear_fixmap(enum fixed_addresses idx)
368 {
369         if (after_paging_init)
370                 clear_fixmap(idx);
371         else
372                 __early_set_fixmap(idx, 0, __pgprot(0));
373 }
374
375
376 int __initdata early_ioremap_nested;
377
378 static int __init check_early_ioremap_leak(void)
379 {
380         if (!early_ioremap_nested)
381                 return 0;
382
383         printk(KERN_WARNING
384                "Debug warning: early ioremap leak of %d areas detected.\n",
385                early_ioremap_nested);
386         printk(KERN_WARNING
387                "please boot with early_ioremap_debug and report the dmesg.\n");
388         WARN_ON(1);
389
390         return 1;
391 }
392 late_initcall(check_early_ioremap_leak);
393
394 void __init *early_ioremap(unsigned long phys_addr, unsigned long size)
395 {
396         unsigned long offset, last_addr;
397         unsigned int nrpages, nesting;
398         enum fixed_addresses idx0, idx;
399
400         WARN_ON(system_state != SYSTEM_BOOTING);
401
402         nesting = early_ioremap_nested;
403         if (early_ioremap_debug) {
404                 printk(KERN_INFO "early_ioremap(%08lx, %08lx) [%d] => ",
405                        phys_addr, size, nesting);
406                 dump_stack();
407         }
408
409         /* Don't allow wraparound or zero size */
410         last_addr = phys_addr + size - 1;
411         if (!size || last_addr < phys_addr) {
412                 WARN_ON(1);
413                 return NULL;
414         }
415
416         if (nesting >= FIX_BTMAPS_NESTING) {
417                 WARN_ON(1);
418                 return NULL;
419         }
420         early_ioremap_nested++;
421         /*
422          * Mappings have to be page-aligned
423          */
424         offset = phys_addr & ~PAGE_MASK;
425         phys_addr &= PAGE_MASK;
426         size = PAGE_ALIGN(last_addr) - phys_addr;
427
428         /*
429          * Mappings have to fit in the FIX_BTMAP area.
430          */
431         nrpages = size >> PAGE_SHIFT;
432         if (nrpages > NR_FIX_BTMAPS) {
433                 WARN_ON(1);
434                 return NULL;
435         }
436
437         /*
438          * Ok, go for it..
439          */
440         idx0 = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
441         idx = idx0;
442         while (nrpages > 0) {
443                 early_set_fixmap(idx, phys_addr);
444                 phys_addr += PAGE_SIZE;
445                 --idx;
446                 --nrpages;
447         }
448         if (early_ioremap_debug)
449                 printk(KERN_CONT "%08lx + %08lx\n", offset, fix_to_virt(idx0));
450
451         return (void *) (offset + fix_to_virt(idx0));
452 }
453
454 void __init early_iounmap(void *addr, unsigned long size)
455 {
456         unsigned long virt_addr;
457         unsigned long offset;
458         unsigned int nrpages;
459         enum fixed_addresses idx;
460         unsigned int nesting;
461
462         nesting = --early_ioremap_nested;
463         WARN_ON(nesting < 0);
464
465         if (early_ioremap_debug) {
466                 printk(KERN_INFO "early_iounmap(%p, %08lx) [%d]\n", addr,
467                        size, nesting);
468                 dump_stack();
469         }
470
471         virt_addr = (unsigned long)addr;
472         if (virt_addr < fix_to_virt(FIX_BTMAP_BEGIN)) {
473                 WARN_ON(1);
474                 return;
475         }
476         offset = virt_addr & ~PAGE_MASK;
477         nrpages = PAGE_ALIGN(offset + size - 1) >> PAGE_SHIFT;
478
479         idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*nesting;
480         while (nrpages > 0) {
481                 early_clear_fixmap(idx);
482                 --idx;
483                 --nrpages;
484         }
485 }
486
487 void __this_fixmap_does_not_exist(void)
488 {
489         WARN_ON(1);
490 }
491
492 #endif /* CONFIG_X86_32 */