xref: /linux/arch/x86/mm/ioremap.c (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
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/slab.h>
13 #include <linux/vmalloc.h>
14 #include <linux/mmiotrace.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 #include "physaddr.h"
25 
26 /*
27  * Fix up the linear direct mapping of the kernel to avoid cache attribute
28  * conflicts.
29  */
30 int ioremap_change_attr(unsigned long vaddr, unsigned long size,
31 			enum page_cache_mode pcm)
32 {
33 	unsigned long nrpages = size >> PAGE_SHIFT;
34 	int err;
35 
36 	switch (pcm) {
37 	case _PAGE_CACHE_MODE_UC:
38 	default:
39 		err = _set_memory_uc(vaddr, nrpages);
40 		break;
41 	case _PAGE_CACHE_MODE_WC:
42 		err = _set_memory_wc(vaddr, nrpages);
43 		break;
44 	case _PAGE_CACHE_MODE_WT:
45 		err = _set_memory_wt(vaddr, nrpages);
46 		break;
47 	case _PAGE_CACHE_MODE_WB:
48 		err = _set_memory_wb(vaddr, nrpages);
49 		break;
50 	}
51 
52 	return err;
53 }
54 
55 static int __ioremap_check_ram(unsigned long start_pfn, unsigned long nr_pages,
56 			       void *arg)
57 {
58 	unsigned long i;
59 
60 	for (i = 0; i < nr_pages; ++i)
61 		if (pfn_valid(start_pfn + i) &&
62 		    !PageReserved(pfn_to_page(start_pfn + i)))
63 			return 1;
64 
65 	return 0;
66 }
67 
68 /*
69  * Remap an arbitrary physical address space into the kernel virtual
70  * address space. It transparently creates kernel huge I/O mapping when
71  * the physical address is aligned by a huge page size (1GB or 2MB) and
72  * the requested size is at least the huge page size.
73  *
74  * NOTE: MTRRs can override PAT memory types with a 4KB granularity.
75  * Therefore, the mapping code falls back to use a smaller page toward 4KB
76  * when a mapping range is covered by non-WB type of MTRRs.
77  *
78  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
79  * have to convert them into an offset in a page-aligned mapping, but the
80  * caller shouldn't need to know that small detail.
81  */
82 static void __iomem *__ioremap_caller(resource_size_t phys_addr,
83 		unsigned long size, enum page_cache_mode pcm, void *caller)
84 {
85 	unsigned long offset, vaddr;
86 	resource_size_t pfn, last_pfn, last_addr;
87 	const resource_size_t unaligned_phys_addr = phys_addr;
88 	const unsigned long unaligned_size = size;
89 	struct vm_struct *area;
90 	enum page_cache_mode new_pcm;
91 	pgprot_t prot;
92 	int retval;
93 	void __iomem *ret_addr;
94 
95 	/* Don't allow wraparound or zero size */
96 	last_addr = phys_addr + size - 1;
97 	if (!size || last_addr < phys_addr)
98 		return NULL;
99 
100 	if (!phys_addr_valid(phys_addr)) {
101 		printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
102 		       (unsigned long long)phys_addr);
103 		WARN_ON_ONCE(1);
104 		return NULL;
105 	}
106 
107 	/*
108 	 * Don't remap the low PCI/ISA area, it's always mapped..
109 	 */
110 	if (is_ISA_range(phys_addr, last_addr))
111 		return (__force void __iomem *)phys_to_virt(phys_addr);
112 
113 	/*
114 	 * Don't allow anybody to remap normal RAM that we're using..
115 	 */
116 	pfn      = phys_addr >> PAGE_SHIFT;
117 	last_pfn = last_addr >> PAGE_SHIFT;
118 	if (walk_system_ram_range(pfn, last_pfn - pfn + 1, NULL,
119 					  __ioremap_check_ram) == 1) {
120 		WARN_ONCE(1, "ioremap on RAM at %pa - %pa\n",
121 			  &phys_addr, &last_addr);
122 		return NULL;
123 	}
124 
125 	/*
126 	 * Mappings have to be page-aligned
127 	 */
128 	offset = phys_addr & ~PAGE_MASK;
129 	phys_addr &= PHYSICAL_PAGE_MASK;
130 	size = PAGE_ALIGN(last_addr+1) - phys_addr;
131 
132 	retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
133 						pcm, &new_pcm);
134 	if (retval) {
135 		printk(KERN_ERR "ioremap reserve_memtype failed %d\n", retval);
136 		return NULL;
137 	}
138 
139 	if (pcm != new_pcm) {
140 		if (!is_new_memtype_allowed(phys_addr, size, pcm, new_pcm)) {
141 			printk(KERN_ERR
142 		"ioremap error for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
143 				(unsigned long long)phys_addr,
144 				(unsigned long long)(phys_addr + size),
145 				pcm, new_pcm);
146 			goto err_free_memtype;
147 		}
148 		pcm = new_pcm;
149 	}
150 
151 	prot = PAGE_KERNEL_IO;
152 	switch (pcm) {
153 	case _PAGE_CACHE_MODE_UC:
154 	default:
155 		prot = __pgprot(pgprot_val(prot) |
156 				cachemode2protval(_PAGE_CACHE_MODE_UC));
157 		break;
158 	case _PAGE_CACHE_MODE_UC_MINUS:
159 		prot = __pgprot(pgprot_val(prot) |
160 				cachemode2protval(_PAGE_CACHE_MODE_UC_MINUS));
161 		break;
162 	case _PAGE_CACHE_MODE_WC:
163 		prot = __pgprot(pgprot_val(prot) |
164 				cachemode2protval(_PAGE_CACHE_MODE_WC));
165 		break;
166 	case _PAGE_CACHE_MODE_WT:
167 		prot = __pgprot(pgprot_val(prot) |
168 				cachemode2protval(_PAGE_CACHE_MODE_WT));
169 		break;
170 	case _PAGE_CACHE_MODE_WB:
171 		break;
172 	}
173 
174 	/*
175 	 * Ok, go for it..
176 	 */
177 	area = get_vm_area_caller(size, VM_IOREMAP, caller);
178 	if (!area)
179 		goto err_free_memtype;
180 	area->phys_addr = phys_addr;
181 	vaddr = (unsigned long) area->addr;
182 
183 	if (kernel_map_sync_memtype(phys_addr, size, pcm))
184 		goto err_free_area;
185 
186 	if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
187 		goto err_free_area;
188 
189 	ret_addr = (void __iomem *) (vaddr + offset);
190 	mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
191 
192 	/*
193 	 * Check if the request spans more than any BAR in the iomem resource
194 	 * tree.
195 	 */
196 	if (iomem_map_sanity_check(unaligned_phys_addr, unaligned_size))
197 		pr_warn("caller %pS mapping multiple BARs\n", caller);
198 
199 	return ret_addr;
200 err_free_area:
201 	free_vm_area(area);
202 err_free_memtype:
203 	free_memtype(phys_addr, phys_addr + size);
204 	return NULL;
205 }
206 
207 /**
208  * ioremap_nocache     -   map bus memory into CPU space
209  * @phys_addr:    bus address of the memory
210  * @size:      size of the resource to map
211  *
212  * ioremap_nocache performs a platform specific sequence of operations to
213  * make bus memory CPU accessible via the readb/readw/readl/writeb/
214  * writew/writel functions and the other mmio helpers. The returned
215  * address is not guaranteed to be usable directly as a virtual
216  * address.
217  *
218  * This version of ioremap ensures that the memory is marked uncachable
219  * on the CPU as well as honouring existing caching rules from things like
220  * the PCI bus. Note that there are other caches and buffers on many
221  * busses. In particular driver authors should read up on PCI writes
222  *
223  * It's useful if some control registers are in such an area and
224  * write combining or read caching is not desirable:
225  *
226  * Must be freed with iounmap.
227  */
228 void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
229 {
230 	/*
231 	 * Ideally, this should be:
232 	 *	pat_enabled() ? _PAGE_CACHE_MODE_UC : _PAGE_CACHE_MODE_UC_MINUS;
233 	 *
234 	 * Till we fix all X drivers to use ioremap_wc(), we will use
235 	 * UC MINUS. Drivers that are certain they need or can already
236 	 * be converted over to strong UC can use ioremap_uc().
237 	 */
238 	enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC_MINUS;
239 
240 	return __ioremap_caller(phys_addr, size, pcm,
241 				__builtin_return_address(0));
242 }
243 EXPORT_SYMBOL(ioremap_nocache);
244 
245 /**
246  * ioremap_uc     -   map bus memory into CPU space as strongly uncachable
247  * @phys_addr:    bus address of the memory
248  * @size:      size of the resource to map
249  *
250  * ioremap_uc performs a platform specific sequence of operations to
251  * make bus memory CPU accessible via the readb/readw/readl/writeb/
252  * writew/writel functions and the other mmio helpers. The returned
253  * address is not guaranteed to be usable directly as a virtual
254  * address.
255  *
256  * This version of ioremap ensures that the memory is marked with a strong
257  * preference as completely uncachable on the CPU when possible. For non-PAT
258  * systems this ends up setting page-attribute flags PCD=1, PWT=1. For PAT
259  * systems this will set the PAT entry for the pages as strong UC.  This call
260  * will honor existing caching rules from things like the PCI bus. Note that
261  * there are other caches and buffers on many busses. In particular driver
262  * authors should read up on PCI writes.
263  *
264  * It's useful if some control registers are in such an area and
265  * write combining or read caching is not desirable:
266  *
267  * Must be freed with iounmap.
268  */
269 void __iomem *ioremap_uc(resource_size_t phys_addr, unsigned long size)
270 {
271 	enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC;
272 
273 	return __ioremap_caller(phys_addr, size, pcm,
274 				__builtin_return_address(0));
275 }
276 EXPORT_SYMBOL_GPL(ioremap_uc);
277 
278 /**
279  * ioremap_wc	-	map memory into CPU space write combined
280  * @phys_addr:	bus address of the memory
281  * @size:	size of the resource to map
282  *
283  * This version of ioremap ensures that the memory is marked write combining.
284  * Write combining allows faster writes to some hardware devices.
285  *
286  * Must be freed with iounmap.
287  */
288 void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
289 {
290 	return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WC,
291 					__builtin_return_address(0));
292 }
293 EXPORT_SYMBOL(ioremap_wc);
294 
295 /**
296  * ioremap_wt	-	map memory into CPU space write through
297  * @phys_addr:	bus address of the memory
298  * @size:	size of the resource to map
299  *
300  * This version of ioremap ensures that the memory is marked write through.
301  * Write through stores data into memory while keeping the cache up-to-date.
302  *
303  * Must be freed with iounmap.
304  */
305 void __iomem *ioremap_wt(resource_size_t phys_addr, unsigned long size)
306 {
307 	return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WT,
308 					__builtin_return_address(0));
309 }
310 EXPORT_SYMBOL(ioremap_wt);
311 
312 void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
313 {
314 	return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
315 				__builtin_return_address(0));
316 }
317 EXPORT_SYMBOL(ioremap_cache);
318 
319 void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
320 				unsigned long prot_val)
321 {
322 	return __ioremap_caller(phys_addr, size,
323 				pgprot2cachemode(__pgprot(prot_val)),
324 				__builtin_return_address(0));
325 }
326 EXPORT_SYMBOL(ioremap_prot);
327 
328 /**
329  * iounmap - Free a IO remapping
330  * @addr: virtual address from ioremap_*
331  *
332  * Caller must ensure there is only one unmapping for the same pointer.
333  */
334 void iounmap(volatile void __iomem *addr)
335 {
336 	struct vm_struct *p, *o;
337 
338 	if ((void __force *)addr <= high_memory)
339 		return;
340 
341 	/*
342 	 * __ioremap special-cases the PCI/ISA range by not instantiating a
343 	 * vm_area and by simply returning an address into the kernel mapping
344 	 * of ISA space.   So handle that here.
345 	 */
346 	if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
347 	    (void __force *)addr < phys_to_virt(ISA_END_ADDRESS))
348 		return;
349 
350 	addr = (volatile void __iomem *)
351 		(PAGE_MASK & (unsigned long __force)addr);
352 
353 	mmiotrace_iounmap(addr);
354 
355 	/* Use the vm area unlocked, assuming the caller
356 	   ensures there isn't another iounmap for the same address
357 	   in parallel. Reuse of the virtual address is prevented by
358 	   leaving it in the global lists until we're done with it.
359 	   cpa takes care of the direct mappings. */
360 	p = find_vm_area((void __force *)addr);
361 
362 	if (!p) {
363 		printk(KERN_ERR "iounmap: bad address %p\n", addr);
364 		dump_stack();
365 		return;
366 	}
367 
368 	free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
369 
370 	/* Finally remove it */
371 	o = remove_vm_area((void __force *)addr);
372 	BUG_ON(p != o || o == NULL);
373 	kfree(p);
374 }
375 EXPORT_SYMBOL(iounmap);
376 
377 int __init arch_ioremap_pud_supported(void)
378 {
379 #ifdef CONFIG_X86_64
380 	return boot_cpu_has(X86_FEATURE_GBPAGES);
381 #else
382 	return 0;
383 #endif
384 }
385 
386 int __init arch_ioremap_pmd_supported(void)
387 {
388 	return boot_cpu_has(X86_FEATURE_PSE);
389 }
390 
391 /*
392  * Convert a physical pointer to a virtual kernel pointer for /dev/mem
393  * access
394  */
395 void *xlate_dev_mem_ptr(phys_addr_t phys)
396 {
397 	unsigned long start  = phys &  PAGE_MASK;
398 	unsigned long offset = phys & ~PAGE_MASK;
399 	void *vaddr;
400 
401 	/* If page is RAM, we can use __va. Otherwise ioremap and unmap. */
402 	if (page_is_ram(start >> PAGE_SHIFT))
403 		return __va(phys);
404 
405 	vaddr = ioremap_cache(start, PAGE_SIZE);
406 	/* Only add the offset on success and return NULL if the ioremap() failed: */
407 	if (vaddr)
408 		vaddr += offset;
409 
410 	return vaddr;
411 }
412 
413 void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr)
414 {
415 	if (page_is_ram(phys >> PAGE_SHIFT))
416 		return;
417 
418 	iounmap((void __iomem *)((unsigned long)addr & PAGE_MASK));
419 }
420 
421 static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
422 
423 static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
424 {
425 	/* Don't assume we're using swapper_pg_dir at this point */
426 	pgd_t *base = __va(read_cr3());
427 	pgd_t *pgd = &base[pgd_index(addr)];
428 	pud_t *pud = pud_offset(pgd, addr);
429 	pmd_t *pmd = pmd_offset(pud, addr);
430 
431 	return pmd;
432 }
433 
434 static inline pte_t * __init early_ioremap_pte(unsigned long addr)
435 {
436 	return &bm_pte[pte_index(addr)];
437 }
438 
439 bool __init is_early_ioremap_ptep(pte_t *ptep)
440 {
441 	return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
442 }
443 
444 void __init early_ioremap_init(void)
445 {
446 	pmd_t *pmd;
447 
448 #ifdef CONFIG_X86_64
449 	BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
450 #else
451 	WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
452 #endif
453 
454 	early_ioremap_setup();
455 
456 	pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
457 	memset(bm_pte, 0, sizeof(bm_pte));
458 	pmd_populate_kernel(&init_mm, pmd, bm_pte);
459 
460 	/*
461 	 * The boot-ioremap range spans multiple pmds, for which
462 	 * we are not prepared:
463 	 */
464 #define __FIXADDR_TOP (-PAGE_SIZE)
465 	BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
466 		     != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
467 #undef __FIXADDR_TOP
468 	if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
469 		WARN_ON(1);
470 		printk(KERN_WARNING "pmd %p != %p\n",
471 		       pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
472 		printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
473 			fix_to_virt(FIX_BTMAP_BEGIN));
474 		printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END):   %08lx\n",
475 			fix_to_virt(FIX_BTMAP_END));
476 
477 		printk(KERN_WARNING "FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
478 		printk(KERN_WARNING "FIX_BTMAP_BEGIN:     %d\n",
479 		       FIX_BTMAP_BEGIN);
480 	}
481 }
482 
483 void __init __early_set_fixmap(enum fixed_addresses idx,
484 			       phys_addr_t phys, pgprot_t flags)
485 {
486 	unsigned long addr = __fix_to_virt(idx);
487 	pte_t *pte;
488 
489 	if (idx >= __end_of_fixed_addresses) {
490 		BUG();
491 		return;
492 	}
493 	pte = early_ioremap_pte(addr);
494 
495 	if (pgprot_val(flags))
496 		set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
497 	else
498 		pte_clear(&init_mm, addr, pte);
499 	__flush_tlb_one(addr);
500 }
501