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.
10 #include <linux/seq_file.h>
11 #include <linux/bootmem.h>
12 #include <linux/debugfs.h>
13 #include <linux/kernel.h>
14 #include <linux/gfp.h>
18 #include <asm/cacheflush.h>
19 #include <asm/processor.h>
20 #include <asm/tlbflush.h>
21 #include <asm/pgtable.h>
22 #include <asm/fcntl.h>
31 int __read_mostly pat_enabled = 1;
33 void __cpuinit pat_disable(char *reason)
36 printk(KERN_INFO "%s\n", reason);
39 static int __init nopat(char *str)
41 pat_disable("PAT support disabled.");
44 early_param("nopat", nopat);
48 static int debug_enable;
50 static int __init pat_debug_setup(char *str)
55 __setup("debugpat", pat_debug_setup);
57 #define dprintk(fmt, arg...) \
58 do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0)
61 static u64 __read_mostly boot_pat_state;
64 PAT_UC = 0, /* uncached */
65 PAT_WC = 1, /* Write combining */
66 PAT_WT = 4, /* Write Through */
67 PAT_WP = 5, /* Write Protected */
68 PAT_WB = 6, /* Write Back (default) */
69 PAT_UC_MINUS = 7, /* UC, but can be overriden by MTRR */
72 #define PAT(x, y) ((u64)PAT_ ## y << ((x)*8))
82 if (!cpu_has_pat && boot_pat_state) {
84 * If this happens we are on a secondary CPU, but
85 * switched to PAT on the boot CPU. We have no way to
88 printk(KERN_ERR "PAT enabled, "
89 "but not supported by secondary CPU\n");
93 /* Set PWT to Write-Combining. All other bits stay the same */
95 * PTE encoding used in Linux:
100 * 000 WB _PAGE_CACHE_WB
101 * 001 WC _PAGE_CACHE_WC
102 * 010 UC- _PAGE_CACHE_UC_MINUS
103 * 011 UC _PAGE_CACHE_UC
106 pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
107 PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
111 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
113 wrmsrl(MSR_IA32_CR_PAT, pat);
114 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
115 smp_processor_id(), boot_pat_state, pat);
120 static char *cattr_name(unsigned long flags)
122 switch (flags & _PAGE_CACHE_MASK) {
123 case _PAGE_CACHE_UC: return "uncached";
124 case _PAGE_CACHE_UC_MINUS: return "uncached-minus";
125 case _PAGE_CACHE_WB: return "write-back";
126 case _PAGE_CACHE_WC: return "write-combining";
127 default: return "broken";
132 * The global memtype list keeps track of memory type for specific
133 * physical memory areas. Conflicting memory types in different
134 * mappings can cause CPU cache corruption. To avoid this we keep track.
136 * The list is sorted based on starting address and can contain multiple
137 * entries for each address (this allows reference counting for overlapping
138 * areas). All the aliases have the same cache attributes of course.
139 * Zero attributes are represented as holes.
141 * Currently the data structure is a list because the number of mappings
142 * are expected to be relatively small. If this should be a problem
143 * it could be changed to a rbtree or similar.
145 * memtype_lock protects the whole list.
155 static LIST_HEAD(memtype_list);
156 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype list */
159 * Does intersection of PAT memory type and MTRR memory type and returns
160 * the resulting memory type as PAT understands it.
161 * (Type in pat and mtrr will not have same value)
162 * The intersection is based on "Effective Memory Type" tables in IA-32
165 static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
168 * Look for MTRR hint to get the effective type in case where PAT
171 if (req_type == _PAGE_CACHE_WB) {
174 mtrr_type = mtrr_type_lookup(start, end);
175 if (mtrr_type == MTRR_TYPE_UNCACHABLE)
176 return _PAGE_CACHE_UC;
177 if (mtrr_type == MTRR_TYPE_WRCOMB)
178 return _PAGE_CACHE_WC;
185 chk_conflict(struct memtype *new, struct memtype *entry, unsigned long *type)
187 if (new->type != entry->type) {
189 new->type = entry->type;
195 /* check overlaps with more than one entry in the list */
196 list_for_each_entry_continue(entry, &memtype_list, nd) {
197 if (new->end <= entry->start)
199 else if (new->type != entry->type)
205 printk(KERN_INFO "%s:%d conflicting memory types "
206 "%Lx-%Lx %s<->%s\n", current->comm, current->pid, new->start,
207 new->end, cattr_name(new->type), cattr_name(entry->type));
211 static struct memtype *cached_entry;
212 static u64 cached_start;
215 * For RAM pages, mark the pages as non WB memory type using
216 * PageNonWB (PG_arch_1). We allow only one set_memory_uc() or
217 * set_memory_wc() on a RAM page at a time before marking it as WB again.
218 * This is ok, because only one driver will be owning the page and
219 * doing set_memory_*() calls.
221 * For now, we use PageNonWB to track that the RAM page is being mapped
222 * as non WB. In future, we will have to use one more flag
223 * (or some other mechanism in page_struct) to distinguish between
226 static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
227 unsigned long *new_type)
232 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
233 page = pfn_to_page(pfn);
234 if (page_mapped(page) || PageNonWB(page))
243 for (pfn = (start >> PAGE_SHIFT); pfn < end_pfn; ++pfn) {
244 page = pfn_to_page(pfn);
245 ClearPageNonWB(page);
251 static int free_ram_pages_type(u64 start, u64 end)
256 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
257 page = pfn_to_page(pfn);
258 if (page_mapped(page) || !PageNonWB(page))
261 ClearPageNonWB(page);
267 for (pfn = (start >> PAGE_SHIFT); pfn < end_pfn; ++pfn) {
268 page = pfn_to_page(pfn);
275 * req_type typically has one of the:
278 * - _PAGE_CACHE_UC_MINUS
281 * req_type will have a special case value '-1', when requester want to inherit
282 * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
284 * If new_type is NULL, function will return an error if it cannot reserve the
285 * region with req_type. If new_type is non-NULL, function will return
286 * available type in new_type in case of no error. In case of any error
287 * it will return a negative return value.
289 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
290 unsigned long *new_type)
292 struct memtype *new, *entry;
293 unsigned long actual_type;
294 struct list_head *where;
298 BUG_ON(start >= end); /* end is exclusive */
301 /* This is identical to page table setting without PAT */
304 *new_type = _PAGE_CACHE_WB;
306 *new_type = req_type & _PAGE_CACHE_MASK;
311 /* Low ISA region is always mapped WB in page table. No need to track */
312 if (is_ISA_range(start, end - 1)) {
314 *new_type = _PAGE_CACHE_WB;
318 if (req_type == -1) {
320 * Call mtrr_lookup to get the type hint. This is an
321 * optimization for /dev/mem mmap'ers into WB memory (BIOS
322 * tools and ACPI tools). Use WB request for WB memory and use
323 * UC_MINUS otherwise.
325 u8 mtrr_type = mtrr_type_lookup(start, end);
327 if (mtrr_type == MTRR_TYPE_WRBACK)
328 actual_type = _PAGE_CACHE_WB;
330 actual_type = _PAGE_CACHE_UC_MINUS;
332 actual_type = pat_x_mtrr_type(start, end,
333 req_type & _PAGE_CACHE_MASK);
336 is_range_ram = pagerange_is_ram(start, end);
337 if (is_range_ram == 1)
338 return reserve_ram_pages_type(start, end, req_type, new_type);
339 else if (is_range_ram < 0)
342 new = kmalloc(sizeof(struct memtype), GFP_KERNEL);
348 new->type = actual_type;
351 *new_type = actual_type;
353 spin_lock(&memtype_lock);
355 if (cached_entry && start >= cached_start)
356 entry = cached_entry;
358 entry = list_entry(&memtype_list, struct memtype, nd);
360 /* Search for existing mapping that overlaps the current range */
362 list_for_each_entry_continue(entry, &memtype_list, nd) {
363 if (end <= entry->start) {
364 where = entry->nd.prev;
365 cached_entry = list_entry(where, struct memtype, nd);
367 } else if (start <= entry->start) { /* end > entry->start */
368 err = chk_conflict(new, entry, new_type);
370 dprintk("Overlap at 0x%Lx-0x%Lx\n",
371 entry->start, entry->end);
372 where = entry->nd.prev;
373 cached_entry = list_entry(where,
377 } else if (start < entry->end) { /* start > entry->start */
378 err = chk_conflict(new, entry, new_type);
380 dprintk("Overlap at 0x%Lx-0x%Lx\n",
381 entry->start, entry->end);
382 cached_entry = list_entry(entry->nd.prev,
386 * Move to right position in the linked
387 * list to add this new entry
389 list_for_each_entry_continue(entry,
391 if (start <= entry->start) {
392 where = entry->nd.prev;
402 printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, "
403 "track %s, req %s\n",
404 start, end, cattr_name(new->type), cattr_name(req_type));
406 spin_unlock(&memtype_lock);
411 cached_start = start;
414 list_add(&new->nd, where);
416 list_add_tail(&new->nd, &memtype_list);
418 spin_unlock(&memtype_lock);
420 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
421 start, end, cattr_name(new->type), cattr_name(req_type),
422 new_type ? cattr_name(*new_type) : "-");
427 int free_memtype(u64 start, u64 end)
429 struct memtype *entry;
436 /* Low ISA region is always mapped WB. No need to track */
437 if (is_ISA_range(start, end - 1))
440 is_range_ram = pagerange_is_ram(start, end);
441 if (is_range_ram == 1)
442 return free_ram_pages_type(start, end);
443 else if (is_range_ram < 0)
446 spin_lock(&memtype_lock);
447 list_for_each_entry(entry, &memtype_list, nd) {
448 if (entry->start == start && entry->end == end) {
449 if (cached_entry == entry || cached_start == start)
452 list_del(&entry->nd);
458 spin_unlock(&memtype_lock);
461 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
462 current->comm, current->pid, start, end);
465 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
471 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
472 unsigned long size, pgprot_t vma_prot)
477 #ifdef CONFIG_STRICT_DEVMEM
478 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
479 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
484 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
486 u64 from = ((u64)pfn) << PAGE_SHIFT;
487 u64 to = from + size;
490 while (cursor < to) {
491 if (!devmem_is_allowed(pfn)) {
493 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
494 current->comm, from, to);
502 #endif /* CONFIG_STRICT_DEVMEM */
504 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
505 unsigned long size, pgprot_t *vma_prot)
507 u64 offset = ((u64) pfn) << PAGE_SHIFT;
508 unsigned long flags = -1;
511 if (!range_is_allowed(pfn, size))
514 if (file->f_flags & O_SYNC) {
515 flags = _PAGE_CACHE_UC_MINUS;
520 * On the PPro and successors, the MTRRs are used to set
521 * memory types for physical addresses outside main memory,
522 * so blindly setting UC or PWT on those pages is wrong.
523 * For Pentiums and earlier, the surround logic should disable
524 * caching for the high addresses through the KEN pin, but
525 * we maintain the tradition of paranoia in this code.
528 !(boot_cpu_has(X86_FEATURE_MTRR) ||
529 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
530 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
531 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
532 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
533 flags = _PAGE_CACHE_UC;
538 * With O_SYNC, we can only take UC_MINUS mapping. Fail if we cannot.
540 * Without O_SYNC, we want to get
541 * - WB for WB-able memory and no other conflicting mappings
542 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
543 * - Inherit from confliting mappings otherwise
546 retval = reserve_memtype(offset, offset + size, flags, NULL);
548 retval = reserve_memtype(offset, offset + size, -1, &flags);
554 if (((pfn < max_low_pfn_mapped) ||
555 (pfn >= (1UL<<(32 - PAGE_SHIFT)) && pfn < max_pfn_mapped)) &&
556 ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
557 free_memtype(offset, offset + size);
559 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
560 current->comm, current->pid,
562 offset, (unsigned long long)(offset + size));
566 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
571 void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
573 unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
574 u64 addr = (u64)pfn << PAGE_SHIFT;
577 reserve_memtype(addr, addr + size, want_flags, &flags);
578 if (flags != want_flags) {
580 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
581 current->comm, current->pid,
582 cattr_name(want_flags),
583 addr, (unsigned long long)(addr + size),
588 void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
590 u64 addr = (u64)pfn << PAGE_SHIFT;
592 free_memtype(addr, addr + size);
595 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
597 /* get Nth element of the linked list */
598 static struct memtype *memtype_get_idx(loff_t pos)
600 struct memtype *list_node, *print_entry;
603 print_entry = kmalloc(sizeof(struct memtype), GFP_KERNEL);
607 spin_lock(&memtype_lock);
608 list_for_each_entry(list_node, &memtype_list, nd) {
610 *print_entry = *list_node;
611 spin_unlock(&memtype_lock);
616 spin_unlock(&memtype_lock);
622 static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
626 seq_printf(seq, "PAT memtype list:\n");
629 return memtype_get_idx(*pos);
632 static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
635 return memtype_get_idx(*pos);
638 static void memtype_seq_stop(struct seq_file *seq, void *v)
642 static int memtype_seq_show(struct seq_file *seq, void *v)
644 struct memtype *print_entry = (struct memtype *)v;
646 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
647 print_entry->start, print_entry->end);
653 static struct seq_operations memtype_seq_ops = {
654 .start = memtype_seq_start,
655 .next = memtype_seq_next,
656 .stop = memtype_seq_stop,
657 .show = memtype_seq_show,
660 static int memtype_seq_open(struct inode *inode, struct file *file)
662 return seq_open(file, &memtype_seq_ops);
665 static const struct file_operations memtype_fops = {
666 .open = memtype_seq_open,
669 .release = seq_release,
672 static int __init pat_memtype_list_init(void)
674 debugfs_create_file("pat_memtype_list", S_IRUSR, arch_debugfs_dir,
675 NULL, &memtype_fops);
679 late_initcall(pat_memtype_list_init);
681 #endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */