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.
11 #include <linux/kernel.h>
12 #include <linux/gfp.h>
14 #include <linux/bootmem.h>
17 #include <asm/tlbflush.h>
18 #include <asm/processor.h>
20 #include <asm/pgtable.h>
23 #include <asm/cacheflush.h>
24 #include <asm/fcntl.h>
29 int __read_mostly pat_enabled = 1;
31 void __cpuinit pat_disable(char *reason)
34 printk(KERN_INFO "%s\n", reason);
37 static int __init nopat(char *str)
39 pat_disable("PAT support disabled.");
42 early_param("nopat", nopat);
46 static int debug_enable;
47 static int __init pat_debug_setup(char *str)
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;
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))
79 if (!cpu_has_pat && boot_pat_state) {
81 * If this happens we are on a secondary CPU, but
82 * switched to PAT on the boot CPU. We have no way to
85 printk(KERN_ERR "PAT enabled, "
86 "but not supported by secondary CPU\n");
90 /* Set PWT to Write-Combining. All other bits stay the same */
92 * PTE encoding used in Linux:
97 * 000 WB _PAGE_CACHE_WB
98 * 001 WC _PAGE_CACHE_WC
99 * 010 UC- _PAGE_CACHE_UC_MINUS
100 * 011 UC _PAGE_CACHE_UC
103 pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
104 PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
108 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
110 wrmsrl(MSR_IA32_CR_PAT, pat);
111 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
112 smp_processor_id(), boot_pat_state, pat);
117 static char *cattr_name(unsigned long flags)
119 switch (flags & _PAGE_CACHE_MASK) {
120 case _PAGE_CACHE_UC: return "uncached";
121 case _PAGE_CACHE_UC_MINUS: return "uncached-minus";
122 case _PAGE_CACHE_WB: return "write-back";
123 case _PAGE_CACHE_WC: return "write-combining";
124 default: return "broken";
129 * The global memtype list keeps track of memory type for specific
130 * physical memory areas. Conflicting memory types in different
131 * mappings can cause CPU cache corruption. To avoid this we keep track.
133 * The list is sorted based on starting address and can contain multiple
134 * entries for each address (this allows reference counting for overlapping
135 * areas). All the aliases have the same cache attributes of course.
136 * Zero attributes are represented as holes.
138 * Currently the data structure is a list because the number of mappings
139 * are expected to be relatively small. If this should be a problem
140 * it could be changed to a rbtree or similar.
142 * memtype_lock protects the whole list.
152 static LIST_HEAD(memtype_list);
153 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype list */
156 * Does intersection of PAT memory type and MTRR memory type and returns
157 * the resulting memory type as PAT understands it.
158 * (Type in pat and mtrr will not have same value)
159 * The intersection is based on "Effective Memory Type" tables in IA-32
162 static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
165 * Look for MTRR hint to get the effective type in case where PAT
168 if (req_type == _PAGE_CACHE_WB) {
171 mtrr_type = mtrr_type_lookup(start, end);
172 if (mtrr_type == MTRR_TYPE_UNCACHABLE)
173 return _PAGE_CACHE_UC;
174 if (mtrr_type == MTRR_TYPE_WRCOMB)
175 return _PAGE_CACHE_WC;
182 * req_type typically has one of the:
185 * - _PAGE_CACHE_UC_MINUS
188 * req_type will have a special case value '-1', when requester want to inherit
189 * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
191 * If ret_type is NULL, function will return an error if it cannot reserve the
192 * region with req_type. If ret_type is non-null, function will return
193 * available type in ret_type in case of no error. In case of any error
194 * it will return a negative return value.
196 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
197 unsigned long *ret_type)
199 struct memtype *new_entry = NULL;
200 struct memtype *parse;
201 unsigned long actual_type;
204 /* Only track when pat_enabled */
206 /* This is identical to page table setting without PAT */
208 if (req_type == -1) {
209 *ret_type = _PAGE_CACHE_WB;
211 *ret_type = req_type & _PAGE_CACHE_MASK;
217 /* Low ISA region is always mapped WB in page table. No need to track */
218 if (is_ISA_range(start, end - 1)) {
220 *ret_type = _PAGE_CACHE_WB;
225 if (req_type == -1) {
227 * Call mtrr_lookup to get the type hint. This is an
228 * optimization for /dev/mem mmap'ers into WB memory (BIOS
229 * tools and ACPI tools). Use WB request for WB memory and use
230 * UC_MINUS otherwise.
232 u8 mtrr_type = mtrr_type_lookup(start, end);
234 if (mtrr_type == MTRR_TYPE_WRBACK) {
235 req_type = _PAGE_CACHE_WB;
236 actual_type = _PAGE_CACHE_WB;
238 req_type = _PAGE_CACHE_UC_MINUS;
239 actual_type = _PAGE_CACHE_UC_MINUS;
242 req_type &= _PAGE_CACHE_MASK;
243 actual_type = pat_x_mtrr_type(start, end, req_type);
246 new_entry = kmalloc(sizeof(struct memtype), GFP_KERNEL);
250 new_entry->start = start;
251 new_entry->end = end;
252 new_entry->type = actual_type;
255 *ret_type = actual_type;
257 spin_lock(&memtype_lock);
259 /* Search for existing mapping that overlaps the current range */
260 list_for_each_entry(parse, &memtype_list, nd) {
261 struct memtype *saved_ptr;
263 if (parse->start >= end) {
264 dprintk("New Entry\n");
265 list_add(&new_entry->nd, parse->nd.prev);
270 if (start <= parse->start && end >= parse->start) {
271 if (actual_type != parse->type && ret_type) {
272 actual_type = parse->type;
273 *ret_type = actual_type;
274 new_entry->type = actual_type;
277 if (actual_type != parse->type) {
279 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
280 current->comm, current->pid,
282 cattr_name(actual_type),
283 cattr_name(parse->type));
290 * Check to see whether the request overlaps more
291 * than one entry in the list
293 list_for_each_entry_continue(parse, &memtype_list, nd) {
294 if (end <= parse->start) {
298 if (actual_type != parse->type) {
300 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
301 current->comm, current->pid,
303 cattr_name(actual_type),
304 cattr_name(parse->type));
314 dprintk("Overlap at 0x%Lx-0x%Lx\n",
315 saved_ptr->start, saved_ptr->end);
316 /* No conflict. Go ahead and add this new entry */
317 list_add(&new_entry->nd, saved_ptr->nd.prev);
322 if (start < parse->end) {
323 if (actual_type != parse->type && ret_type) {
324 actual_type = parse->type;
325 *ret_type = actual_type;
326 new_entry->type = actual_type;
329 if (actual_type != parse->type) {
331 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
332 current->comm, current->pid,
334 cattr_name(actual_type),
335 cattr_name(parse->type));
342 * Check to see whether the request overlaps more
343 * than one entry in the list
345 list_for_each_entry_continue(parse, &memtype_list, nd) {
346 if (end <= parse->start) {
350 if (actual_type != parse->type) {
352 KERN_INFO "%s:%d conflicting memory types %Lx-%Lx %s<->%s\n",
353 current->comm, current->pid,
355 cattr_name(actual_type),
356 cattr_name(parse->type));
366 dprintk("Overlap at 0x%Lx-0x%Lx\n",
367 saved_ptr->start, saved_ptr->end);
368 /* No conflict. Go ahead and add this new entry */
369 list_add(&new_entry->nd, &saved_ptr->nd);
377 "reserve_memtype failed 0x%Lx-0x%Lx, track %s, req %s\n",
378 start, end, cattr_name(new_entry->type),
379 cattr_name(req_type));
381 spin_unlock(&memtype_lock);
386 /* No conflict. Not yet added to the list. Add to the tail */
387 list_add_tail(&new_entry->nd, &memtype_list);
388 dprintk("New Entry\n");
393 "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
394 start, end, cattr_name(actual_type),
395 cattr_name(req_type), cattr_name(*ret_type));
398 "reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s\n",
399 start, end, cattr_name(actual_type),
400 cattr_name(req_type));
403 spin_unlock(&memtype_lock);
407 int free_memtype(u64 start, u64 end)
412 /* Only track when pat_enabled */
417 /* Low ISA region is always mapped WB. No need to track */
418 if (is_ISA_range(start, end - 1))
421 spin_lock(&memtype_lock);
422 list_for_each_entry(ml, &memtype_list, nd) {
423 if (ml->start == start && ml->end == end) {
430 spin_unlock(&memtype_lock);
433 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
434 current->comm, current->pid, start, end);
437 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
443 * /dev/mem mmap interface. The memtype used for mapping varies:
444 * - Use UC for mappings with O_SYNC flag
445 * - Without O_SYNC flag, if there is any conflict in reserve_memtype,
446 * inherit the memtype from existing mapping.
447 * - Else use UC_MINUS memtype (for backward compatibility with existing
450 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
451 unsigned long size, pgprot_t vma_prot)
456 #ifdef CONFIG_NONPROMISC_DEVMEM
457 /* This check is done in drivers/char/mem.c in case of NONPROMISC_DEVMEM*/
458 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
463 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
465 u64 from = ((u64)pfn) << PAGE_SHIFT;
466 u64 to = from + size;
469 while (cursor < to) {
470 if (!devmem_is_allowed(pfn)) {
472 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
473 current->comm, from, to);
481 #endif /* CONFIG_NONPROMISC_DEVMEM */
483 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
484 unsigned long size, pgprot_t *vma_prot)
486 u64 offset = ((u64) pfn) << PAGE_SHIFT;
487 unsigned long flags = _PAGE_CACHE_UC_MINUS;
490 if (!range_is_allowed(pfn, size))
493 if (file->f_flags & O_SYNC) {
494 flags = _PAGE_CACHE_UC;
499 * On the PPro and successors, the MTRRs are used to set
500 * memory types for physical addresses outside main memory,
501 * so blindly setting UC or PWT on those pages is wrong.
502 * For Pentiums and earlier, the surround logic should disable
503 * caching for the high addresses through the KEN pin, but
504 * we maintain the tradition of paranoia in this code.
507 !(boot_cpu_has(X86_FEATURE_MTRR) ||
508 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
509 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
510 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
511 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
512 flags = _PAGE_CACHE_UC;
517 * With O_SYNC, we can only take UC mapping. Fail if we cannot.
518 * Without O_SYNC, we want to get
519 * - WB for WB-able memory and no other conflicting mappings
520 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
521 * - Inherit from confliting mappings otherwise
523 if (flags != _PAGE_CACHE_UC_MINUS) {
524 retval = reserve_memtype(offset, offset + size, flags, NULL);
526 retval = reserve_memtype(offset, offset + size, -1, &flags);
532 if (pfn <= max_pfn_mapped &&
533 ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
534 free_memtype(offset, offset + size);
536 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
537 current->comm, current->pid,
539 offset, (unsigned long long)(offset + size));
543 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
548 void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
550 u64 addr = (u64)pfn << PAGE_SHIFT;
552 unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
554 reserve_memtype(addr, addr + size, want_flags, &flags);
555 if (flags != want_flags) {
557 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
558 current->comm, current->pid,
559 cattr_name(want_flags),
560 addr, (unsigned long long)(addr + size),
565 void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
567 u64 addr = (u64)pfn << PAGE_SHIFT;
569 free_memtype(addr, addr + size);