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