x86: add "debugpat" boot option
[linux-2.6/linux-2.6-openrd.git] / arch / x86 / mm / pat.c
blob1a82f4db8a4622e2a5d8ae0b474a5689ac84aae0
1 /*
2 * Handle caching attributes in page tables (PAT)
4 * Authors: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
5 * Suresh B Siddha <suresh.b.siddha@intel.com>
7 * Loosely based on earlier PAT patchset from Eric Biederman and Andi Kleen.
8 */
10 #include <linux/mm.h>
11 #include <linux/kernel.h>
12 #include <linux/gfp.h>
13 #include <linux/fs.h>
14 #include <linux/bootmem.h>
16 #include <asm/msr.h>
17 #include <asm/tlbflush.h>
18 #include <asm/processor.h>
19 #include <asm/page.h>
20 #include <asm/pgtable.h>
21 #include <asm/pat.h>
22 #include <asm/e820.h>
23 #include <asm/cacheflush.h>
24 #include <asm/fcntl.h>
25 #include <asm/mtrr.h>
26 #include <asm/io.h>
28 #ifdef CONFIG_X86_PAT
29 int __read_mostly pat_wc_enabled = 1;
31 void __init pat_disable(char *reason)
33 pat_wc_enabled = 0;
34 printk(KERN_INFO "%s\n", reason);
37 static int nopat(char *str)
39 pat_disable("PAT support disabled.");
40 return 0;
42 early_param("nopat", nopat);
43 #endif
46 static int debug_enable;
47 static int __init pat_debug_setup(char *str)
49 debug_enable = 1;
50 return 0;
52 __setup("debugpat", pat_debug_setup);
54 #define dprintk(fmt, arg...) \
55 do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0)
58 static u64 __read_mostly boot_pat_state;
60 enum {
61 PAT_UC = 0, /* uncached */
62 PAT_WC = 1, /* Write combining */
63 PAT_WT = 4, /* Write Through */
64 PAT_WP = 5, /* Write Protected */
65 PAT_WB = 6, /* Write Back (default) */
66 PAT_UC_MINUS = 7, /* UC, but can be overriden by MTRR */
69 #define PAT(x,y) ((u64)PAT_ ## y << ((x)*8))
71 void pat_init(void)
73 u64 pat;
75 if (!pat_wc_enabled)
76 return;
78 /* Paranoia check. */
79 if (!cpu_has_pat) {
80 printk(KERN_ERR "PAT enabled, but CPU feature cleared\n");
82 * Panic if this happens on the secondary CPU, and we
83 * switched to PAT on the boot CPU. We have no way to
84 * undo PAT.
86 BUG_ON(boot_pat_state);
89 /* Set PWT to Write-Combining. All other bits stay the same */
91 * PTE encoding used in Linux:
92 * PAT
93 * |PCD
94 * ||PWT
95 * |||
96 * 000 WB _PAGE_CACHE_WB
97 * 001 WC _PAGE_CACHE_WC
98 * 010 UC- _PAGE_CACHE_UC_MINUS
99 * 011 UC _PAGE_CACHE_UC
100 * PAT bit unused
102 pat = PAT(0,WB) | PAT(1,WC) | PAT(2,UC_MINUS) | PAT(3,UC) |
103 PAT(4,WB) | PAT(5,WC) | PAT(6,UC_MINUS) | PAT(7,UC);
105 /* Boot CPU check */
106 if (!boot_pat_state)
107 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
109 wrmsrl(MSR_IA32_CR_PAT, pat);
110 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
111 smp_processor_id(), boot_pat_state, pat);
114 #undef PAT
116 static char *cattr_name(unsigned long flags)
118 switch (flags & _PAGE_CACHE_MASK) {
119 case _PAGE_CACHE_UC: return "uncached";
120 case _PAGE_CACHE_UC_MINUS: return "uncached-minus";
121 case _PAGE_CACHE_WB: return "write-back";
122 case _PAGE_CACHE_WC: return "write-combining";
123 default: return "broken";
128 * The global memtype list keeps track of memory type for specific
129 * physical memory areas. Conflicting memory types in different
130 * mappings can cause CPU cache corruption. To avoid this we keep track.
132 * The list is sorted based on starting address and can contain multiple
133 * entries for each address (this allows reference counting for overlapping
134 * areas). All the aliases have the same cache attributes of course.
135 * Zero attributes are represented as holes.
137 * Currently the data structure is a list because the number of mappings
138 * are expected to be relatively small. If this should be a problem
139 * it could be changed to a rbtree or similar.
141 * memtype_lock protects the whole list.
144 struct memtype {
145 u64 start;
146 u64 end;
147 unsigned long type;
148 struct list_head nd;
151 static LIST_HEAD(memtype_list);
152 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype list */
155 * Does intersection of PAT memory type and MTRR memory type and returns
156 * the resulting memory type as PAT understands it.
157 * (Type in pat and mtrr will not have same value)
158 * The intersection is based on "Effective Memory Type" tables in IA-32
159 * SDM vol 3a
161 static int pat_x_mtrr_type(u64 start, u64 end, unsigned long prot,
162 unsigned long *ret_prot)
164 unsigned long pat_type;
165 u8 mtrr_type;
167 mtrr_type = mtrr_type_lookup(start, end);
168 if (mtrr_type == 0xFF) { /* MTRR not enabled */
169 *ret_prot = prot;
170 return 0;
172 if (mtrr_type == 0xFE) { /* MTRR match error */
173 *ret_prot = _PAGE_CACHE_UC;
174 return -1;
176 if (mtrr_type != MTRR_TYPE_UNCACHABLE &&
177 mtrr_type != MTRR_TYPE_WRBACK &&
178 mtrr_type != MTRR_TYPE_WRCOMB) { /* MTRR type unhandled */
179 *ret_prot = _PAGE_CACHE_UC;
180 return -1;
183 pat_type = prot & _PAGE_CACHE_MASK;
184 prot &= (~_PAGE_CACHE_MASK);
186 /* Currently doing intersection by hand. Optimize it later. */
187 if (pat_type == _PAGE_CACHE_WC) {
188 *ret_prot = prot | _PAGE_CACHE_WC;
189 } else if (pat_type == _PAGE_CACHE_UC_MINUS) {
190 *ret_prot = prot | _PAGE_CACHE_UC_MINUS;
191 } else if (pat_type == _PAGE_CACHE_UC ||
192 mtrr_type == MTRR_TYPE_UNCACHABLE) {
193 *ret_prot = prot | _PAGE_CACHE_UC;
194 } else if (mtrr_type == MTRR_TYPE_WRCOMB) {
195 *ret_prot = prot | _PAGE_CACHE_WC;
196 } else {
197 *ret_prot = prot | _PAGE_CACHE_WB;
200 return 0;
204 * req_type typically has one of the:
205 * - _PAGE_CACHE_WB
206 * - _PAGE_CACHE_WC
207 * - _PAGE_CACHE_UC_MINUS
208 * - _PAGE_CACHE_UC
210 * req_type will have a special case value '-1', when requester want to inherit
211 * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
213 * If ret_type is NULL, function will return an error if it cannot reserve the
214 * region with req_type. If ret_type is non-null, function will return
215 * available type in ret_type in case of no error. In case of any error
216 * it will return a negative return value.
218 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
219 unsigned long *ret_type)
221 struct memtype *new_entry = NULL;
222 struct memtype *parse;
223 unsigned long actual_type;
224 int err = 0;
226 /* Only track when pat_wc_enabled */
227 if (!pat_wc_enabled) {
228 /* This is identical to page table setting without PAT */
229 if (ret_type) {
230 if (req_type == -1) {
231 *ret_type = _PAGE_CACHE_WB;
232 } else {
233 *ret_type = req_type;
236 return 0;
239 /* Low ISA region is always mapped WB in page table. No need to track */
240 if (start >= ISA_START_ADDRESS && (end - 1) <= ISA_END_ADDRESS) {
241 if (ret_type)
242 *ret_type = _PAGE_CACHE_WB;
244 return 0;
247 if (req_type == -1) {
249 * Special case where caller wants to inherit from mtrr or
250 * existing pat mapping, defaulting to UC_MINUS in case of
251 * no match.
253 u8 mtrr_type = mtrr_type_lookup(start, end);
254 if (mtrr_type == 0xFE) { /* MTRR match error */
255 err = -1;
258 if (mtrr_type == MTRR_TYPE_WRBACK) {
259 req_type = _PAGE_CACHE_WB;
260 actual_type = _PAGE_CACHE_WB;
261 } else {
262 req_type = _PAGE_CACHE_UC_MINUS;
263 actual_type = _PAGE_CACHE_UC_MINUS;
265 } else {
266 req_type &= _PAGE_CACHE_MASK;
267 err = pat_x_mtrr_type(start, end, req_type, &actual_type);
270 if (err) {
271 if (ret_type)
272 *ret_type = actual_type;
274 return -EINVAL;
277 new_entry = kmalloc(sizeof(struct memtype), GFP_KERNEL);
278 if (!new_entry)
279 return -ENOMEM;
281 new_entry->start = start;
282 new_entry->end = end;
283 new_entry->type = actual_type;
285 if (ret_type)
286 *ret_type = actual_type;
288 spin_lock(&memtype_lock);
290 /* Search for existing mapping that overlaps the current range */
291 list_for_each_entry(parse, &memtype_list, nd) {
292 struct memtype *saved_ptr;
294 if (parse->start >= end) {
295 dprintk("New Entry\n");
296 list_add(&new_entry->nd, parse->nd.prev);
297 new_entry = NULL;
298 break;
301 if (start <= parse->start && end >= parse->start) {
302 if (actual_type != parse->type && ret_type) {
303 actual_type = parse->type;
304 *ret_type = actual_type;
305 new_entry->type = actual_type;
308 if (actual_type != parse->type) {
309 printk(
310 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
311 current->comm, current->pid,
312 start, end,
313 cattr_name(actual_type),
314 cattr_name(parse->type));
315 err = -EBUSY;
316 break;
319 saved_ptr = parse;
321 * Check to see whether the request overlaps more
322 * than one entry in the list
324 list_for_each_entry_continue(parse, &memtype_list, nd) {
325 if (end <= parse->start) {
326 break;
329 if (actual_type != parse->type) {
330 printk(
331 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
332 current->comm, current->pid,
333 start, end,
334 cattr_name(actual_type),
335 cattr_name(parse->type));
336 err = -EBUSY;
337 break;
341 if (err) {
342 break;
345 dprintk("Overlap at 0x%Lx-0x%Lx\n",
346 saved_ptr->start, saved_ptr->end);
347 /* No conflict. Go ahead and add this new entry */
348 list_add(&new_entry->nd, saved_ptr->nd.prev);
349 new_entry = NULL;
350 break;
353 if (start < parse->end) {
354 if (actual_type != parse->type && ret_type) {
355 actual_type = parse->type;
356 *ret_type = actual_type;
357 new_entry->type = actual_type;
360 if (actual_type != parse->type) {
361 printk(
362 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
363 current->comm, current->pid,
364 start, end,
365 cattr_name(actual_type),
366 cattr_name(parse->type));
367 err = -EBUSY;
368 break;
371 saved_ptr = parse;
373 * Check to see whether the request overlaps more
374 * than one entry in the list
376 list_for_each_entry_continue(parse, &memtype_list, nd) {
377 if (end <= parse->start) {
378 break;
381 if (actual_type != parse->type) {
382 printk(
383 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
384 current->comm, current->pid,
385 start, end,
386 cattr_name(actual_type),
387 cattr_name(parse->type));
388 err = -EBUSY;
389 break;
393 if (err) {
394 break;
397 dprintk("Overlap at 0x%Lx-0x%Lx\n",
398 saved_ptr->start, saved_ptr->end);
399 /* No conflict. Go ahead and add this new entry */
400 list_add(&new_entry->nd, &saved_ptr->nd);
401 new_entry = NULL;
402 break;
406 if (err) {
407 printk(KERN_INFO
408 "reserve_memtype failed 0x%Lx-0x%Lx, track %s, req %s\n",
409 start, end, cattr_name(new_entry->type),
410 cattr_name(req_type));
411 kfree(new_entry);
412 spin_unlock(&memtype_lock);
413 return err;
416 if (new_entry) {
417 /* No conflict. Not yet added to the list. Add to the tail */
418 list_add_tail(&new_entry->nd, &memtype_list);
419 dprintk("New Entry\n");
422 if (ret_type) {
423 dprintk(
424 "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
425 start, end, cattr_name(actual_type),
426 cattr_name(req_type), cattr_name(*ret_type));
427 } else {
428 dprintk(
429 "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s\n",
430 start, end, cattr_name(actual_type),
431 cattr_name(req_type));
434 spin_unlock(&memtype_lock);
435 return err;
438 int free_memtype(u64 start, u64 end)
440 struct memtype *ml;
441 int err = -EINVAL;
443 /* Only track when pat_wc_enabled */
444 if (!pat_wc_enabled) {
445 return 0;
448 /* Low ISA region is always mapped WB. No need to track */
449 if (start >= ISA_START_ADDRESS && end <= ISA_END_ADDRESS) {
450 return 0;
453 spin_lock(&memtype_lock);
454 list_for_each_entry(ml, &memtype_list, nd) {
455 if (ml->start == start && ml->end == end) {
456 list_del(&ml->nd);
457 kfree(ml);
458 err = 0;
459 break;
462 spin_unlock(&memtype_lock);
464 if (err) {
465 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
466 current->comm, current->pid, start, end);
469 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
470 return err;
475 * /dev/mem mmap interface. The memtype used for mapping varies:
476 * - Use UC for mappings with O_SYNC flag
477 * - Without O_SYNC flag, if there is any conflict in reserve_memtype,
478 * inherit the memtype from existing mapping.
479 * - Else use UC_MINUS memtype (for backward compatibility with existing
480 * X drivers.
482 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
483 unsigned long size, pgprot_t vma_prot)
485 return vma_prot;
488 #ifdef CONFIG_NONPROMISC_DEVMEM
489 /* This check is done in drivers/char/mem.c in case of NONPROMISC_DEVMEM*/
490 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
492 return 1;
494 #else
495 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
497 u64 from = ((u64)pfn) << PAGE_SHIFT;
498 u64 to = from + size;
499 u64 cursor = from;
501 while (cursor < to) {
502 if (!devmem_is_allowed(pfn)) {
503 printk(KERN_INFO
504 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
505 current->comm, from, to);
506 return 0;
508 cursor += PAGE_SIZE;
509 pfn++;
511 return 1;
513 #endif /* CONFIG_NONPROMISC_DEVMEM */
515 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
516 unsigned long size, pgprot_t *vma_prot)
518 u64 offset = ((u64) pfn) << PAGE_SHIFT;
519 unsigned long flags = _PAGE_CACHE_UC_MINUS;
520 int retval;
522 if (!range_is_allowed(pfn, size))
523 return 0;
525 if (file->f_flags & O_SYNC) {
526 flags = _PAGE_CACHE_UC;
529 #ifdef CONFIG_X86_32
531 * On the PPro and successors, the MTRRs are used to set
532 * memory types for physical addresses outside main memory,
533 * so blindly setting UC or PWT on those pages is wrong.
534 * For Pentiums and earlier, the surround logic should disable
535 * caching for the high addresses through the KEN pin, but
536 * we maintain the tradition of paranoia in this code.
538 if (!pat_wc_enabled &&
539 ! ( test_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability) ||
540 test_bit(X86_FEATURE_K6_MTRR, boot_cpu_data.x86_capability) ||
541 test_bit(X86_FEATURE_CYRIX_ARR, boot_cpu_data.x86_capability) ||
542 test_bit(X86_FEATURE_CENTAUR_MCR, boot_cpu_data.x86_capability)) &&
543 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
544 flags = _PAGE_CACHE_UC;
546 #endif
549 * With O_SYNC, we can only take UC mapping. Fail if we cannot.
550 * Without O_SYNC, we want to get
551 * - WB for WB-able memory and no other conflicting mappings
552 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
553 * - Inherit from confliting mappings otherwise
555 if (flags != _PAGE_CACHE_UC_MINUS) {
556 retval = reserve_memtype(offset, offset + size, flags, NULL);
557 } else {
558 retval = reserve_memtype(offset, offset + size, -1, &flags);
561 if (retval < 0)
562 return 0;
564 if (pfn <= max_pfn_mapped &&
565 ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
566 free_memtype(offset, offset + size);
567 printk(KERN_INFO
568 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
569 current->comm, current->pid,
570 cattr_name(flags),
571 offset, offset + size);
572 return 0;
575 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
576 flags);
577 return 1;
580 void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
582 u64 addr = (u64)pfn << PAGE_SHIFT;
583 unsigned long flags;
584 unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
586 reserve_memtype(addr, addr + size, want_flags, &flags);
587 if (flags != want_flags) {
588 printk(KERN_INFO
589 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
590 current->comm, current->pid,
591 cattr_name(want_flags),
592 addr, addr + size,
593 cattr_name(flags));
597 void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
599 u64 addr = (u64)pfn << PAGE_SHIFT;
601 free_memtype(addr, addr + size);