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/module.h>
15 #include <linux/gfp.h>
19 #include <asm/cacheflush.h>
20 #include <asm/processor.h>
21 #include <asm/tlbflush.h>
22 #include <asm/pgtable.h>
23 #include <asm/fcntl.h>
32 int __read_mostly pat_enabled = 1;
34 void __cpuinit pat_disable(const char *reason)
37 printk(KERN_INFO "%s\n", reason);
40 static int __init nopat(char *str)
42 pat_disable("PAT support disabled.");
45 early_param("nopat", nopat);
47 static inline void pat_disable(const char *reason)
54 static int debug_enable;
56 static int __init pat_debug_setup(char *str)
61 __setup("debugpat", pat_debug_setup);
63 #define dprintk(fmt, arg...) \
64 do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0)
67 static u64 __read_mostly boot_pat_state;
70 PAT_UC = 0, /* uncached */
71 PAT_WC = 1, /* Write combining */
72 PAT_WT = 4, /* Write Through */
73 PAT_WP = 5, /* Write Protected */
74 PAT_WB = 6, /* Write Back (default) */
75 PAT_UC_MINUS = 7, /* UC, but can be overriden by MTRR */
78 #define PAT(x, y) ((u64)PAT_ ## y << ((x)*8))
88 if (!boot_pat_state) {
89 pat_disable("PAT not supported by CPU.");
93 * If this happens we are on a secondary CPU, but
94 * switched to PAT on the boot CPU. We have no way to
97 printk(KERN_ERR "PAT enabled, "
98 "but not supported by secondary CPU\n");
103 /* Set PWT to Write-Combining. All other bits stay the same */
105 * PTE encoding used in Linux:
110 * 000 WB _PAGE_CACHE_WB
111 * 001 WC _PAGE_CACHE_WC
112 * 010 UC- _PAGE_CACHE_UC_MINUS
113 * 011 UC _PAGE_CACHE_UC
116 pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
117 PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
121 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
123 wrmsrl(MSR_IA32_CR_PAT, pat);
124 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
125 smp_processor_id(), boot_pat_state, pat);
130 static char *cattr_name(unsigned long flags)
132 switch (flags & _PAGE_CACHE_MASK) {
133 case _PAGE_CACHE_UC: return "uncached";
134 case _PAGE_CACHE_UC_MINUS: return "uncached-minus";
135 case _PAGE_CACHE_WB: return "write-back";
136 case _PAGE_CACHE_WC: return "write-combining";
137 default: return "broken";
142 * The global memtype list keeps track of memory type for specific
143 * physical memory areas. Conflicting memory types in different
144 * mappings can cause CPU cache corruption. To avoid this we keep track.
146 * The list is sorted based on starting address and can contain multiple
147 * entries for each address (this allows reference counting for overlapping
148 * areas). All the aliases have the same cache attributes of course.
149 * Zero attributes are represented as holes.
151 * Currently the data structure is a list because the number of mappings
152 * are expected to be relatively small. If this should be a problem
153 * it could be changed to a rbtree or similar.
155 * memtype_lock protects the whole list.
165 static LIST_HEAD(memtype_list);
166 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype list */
169 * Does intersection of PAT memory type and MTRR memory type and returns
170 * the resulting memory type as PAT understands it.
171 * (Type in pat and mtrr will not have same value)
172 * The intersection is based on "Effective Memory Type" tables in IA-32
175 static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
178 * Look for MTRR hint to get the effective type in case where PAT
181 if (req_type == _PAGE_CACHE_WB) {
184 mtrr_type = mtrr_type_lookup(start, end);
185 if (mtrr_type != MTRR_TYPE_WRBACK)
186 return _PAGE_CACHE_UC_MINUS;
188 return _PAGE_CACHE_WB;
195 chk_conflict(struct memtype *new, struct memtype *entry, unsigned long *type)
197 if (new->type != entry->type) {
199 new->type = entry->type;
205 /* check overlaps with more than one entry in the list */
206 list_for_each_entry_continue(entry, &memtype_list, nd) {
207 if (new->end <= entry->start)
209 else if (new->type != entry->type)
215 printk(KERN_INFO "%s:%d conflicting memory types "
216 "%Lx-%Lx %s<->%s\n", current->comm, current->pid, new->start,
217 new->end, cattr_name(new->type), cattr_name(entry->type));
221 static struct memtype *cached_entry;
222 static u64 cached_start;
224 static int pat_pagerange_is_ram(unsigned long start, unsigned long end)
226 int ram_page = 0, not_rampage = 0;
227 unsigned long page_nr;
229 for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
232 * For legacy reasons, physical address range in the legacy ISA
233 * region is tracked as non-RAM. This will allow users of
234 * /dev/mem to map portions of legacy ISA region, even when
235 * some of those portions are listed(or not even listed) with
236 * different e820 types(RAM/reserved/..)
238 if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
239 page_is_ram(page_nr))
244 if (ram_page == not_rampage)
252 * For RAM pages, mark the pages as non WB memory type using
253 * PageNonWB (PG_arch_1). We allow only one set_memory_uc() or
254 * set_memory_wc() on a RAM page at a time before marking it as WB again.
255 * This is ok, because only one driver will be owning the page and
256 * doing set_memory_*() calls.
258 * For now, we use PageNonWB to track that the RAM page is being mapped
259 * as non WB. In future, we will have to use one more flag
260 * (or some other mechanism in page_struct) to distinguish between
263 static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
264 unsigned long *new_type)
269 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
270 page = pfn_to_page(pfn);
271 if (page_mapped(page) || PageNonWB(page))
280 for (pfn = (start >> PAGE_SHIFT); pfn < end_pfn; ++pfn) {
281 page = pfn_to_page(pfn);
282 ClearPageNonWB(page);
288 static int free_ram_pages_type(u64 start, u64 end)
293 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
294 page = pfn_to_page(pfn);
295 if (page_mapped(page) || !PageNonWB(page))
298 ClearPageNonWB(page);
304 for (pfn = (start >> PAGE_SHIFT); pfn < end_pfn; ++pfn) {
305 page = pfn_to_page(pfn);
312 * req_type typically has one of the:
315 * - _PAGE_CACHE_UC_MINUS
318 * req_type will have a special case value '-1', when requester want to inherit
319 * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
321 * If new_type is NULL, function will return an error if it cannot reserve the
322 * region with req_type. If new_type is non-NULL, function will return
323 * available type in new_type in case of no error. In case of any error
324 * it will return a negative return value.
326 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
327 unsigned long *new_type)
329 struct memtype *new, *entry;
330 unsigned long actual_type;
331 struct list_head *where;
335 BUG_ON(start >= end); /* end is exclusive */
338 /* This is identical to page table setting without PAT */
341 *new_type = _PAGE_CACHE_WB;
343 *new_type = req_type & _PAGE_CACHE_MASK;
348 /* Low ISA region is always mapped WB in page table. No need to track */
349 if (is_ISA_range(start, end - 1)) {
351 *new_type = _PAGE_CACHE_WB;
356 * Call mtrr_lookup to get the type hint. This is an
357 * optimization for /dev/mem mmap'ers into WB memory (BIOS
358 * tools and ACPI tools). Use WB request for WB memory and use
359 * UC_MINUS otherwise.
361 actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK);
364 *new_type = actual_type;
366 is_range_ram = pat_pagerange_is_ram(start, end);
367 if (is_range_ram == 1)
368 return reserve_ram_pages_type(start, end, req_type,
370 else if (is_range_ram < 0)
373 new = kmalloc(sizeof(struct memtype), GFP_KERNEL);
379 new->type = actual_type;
381 spin_lock(&memtype_lock);
383 if (cached_entry && start >= cached_start)
384 entry = cached_entry;
386 entry = list_entry(&memtype_list, struct memtype, nd);
388 /* Search for existing mapping that overlaps the current range */
390 list_for_each_entry_continue(entry, &memtype_list, nd) {
391 if (end <= entry->start) {
392 where = entry->nd.prev;
393 cached_entry = list_entry(where, struct memtype, nd);
395 } else if (start <= entry->start) { /* end > entry->start */
396 err = chk_conflict(new, entry, new_type);
398 dprintk("Overlap at 0x%Lx-0x%Lx\n",
399 entry->start, entry->end);
400 where = entry->nd.prev;
401 cached_entry = list_entry(where,
405 } else if (start < entry->end) { /* start > entry->start */
406 err = chk_conflict(new, entry, new_type);
408 dprintk("Overlap at 0x%Lx-0x%Lx\n",
409 entry->start, entry->end);
410 cached_entry = list_entry(entry->nd.prev,
414 * Move to right position in the linked
415 * list to add this new entry
417 list_for_each_entry_continue(entry,
419 if (start <= entry->start) {
420 where = entry->nd.prev;
430 printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, "
431 "track %s, req %s\n",
432 start, end, cattr_name(new->type), cattr_name(req_type));
434 spin_unlock(&memtype_lock);
439 cached_start = start;
442 list_add(&new->nd, where);
444 list_add_tail(&new->nd, &memtype_list);
446 spin_unlock(&memtype_lock);
448 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
449 start, end, cattr_name(new->type), cattr_name(req_type),
450 new_type ? cattr_name(*new_type) : "-");
455 int free_memtype(u64 start, u64 end)
457 struct memtype *entry;
464 /* Low ISA region is always mapped WB. No need to track */
465 if (is_ISA_range(start, end - 1))
468 is_range_ram = pat_pagerange_is_ram(start, end);
469 if (is_range_ram == 1)
470 return free_ram_pages_type(start, end);
471 else if (is_range_ram < 0)
474 spin_lock(&memtype_lock);
475 list_for_each_entry(entry, &memtype_list, nd) {
476 if (entry->start == start && entry->end == end) {
477 if (cached_entry == entry || cached_start == start)
480 list_del(&entry->nd);
486 spin_unlock(&memtype_lock);
489 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
490 current->comm, current->pid, start, end);
493 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
499 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
500 unsigned long size, pgprot_t vma_prot)
505 #ifdef CONFIG_STRICT_DEVMEM
506 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
507 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
512 /* This check is needed to avoid cache aliasing when PAT is enabled */
513 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
515 u64 from = ((u64)pfn) << PAGE_SHIFT;
516 u64 to = from + size;
522 while (cursor < to) {
523 if (!devmem_is_allowed(pfn)) {
525 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
526 current->comm, from, to);
534 #endif /* CONFIG_STRICT_DEVMEM */
536 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
537 unsigned long size, pgprot_t *vma_prot)
539 unsigned long flags = _PAGE_CACHE_WB;
541 if (!range_is_allowed(pfn, size))
544 if (file->f_flags & O_SYNC) {
545 flags = _PAGE_CACHE_UC_MINUS;
550 * On the PPro and successors, the MTRRs are used to set
551 * memory types for physical addresses outside main memory,
552 * so blindly setting UC or PWT on those pages is wrong.
553 * For Pentiums and earlier, the surround logic should disable
554 * caching for the high addresses through the KEN pin, but
555 * we maintain the tradition of paranoia in this code.
558 !(boot_cpu_has(X86_FEATURE_MTRR) ||
559 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
560 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
561 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
562 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
563 flags = _PAGE_CACHE_UC;
567 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
573 * Change the memory type for the physial address range in kernel identity
574 * mapping space if that range is a part of identity map.
576 int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
580 if (!pat_enabled || base >= __pa(high_memory))
583 id_sz = (__pa(high_memory) < base + size) ?
584 __pa(high_memory) - base :
587 if (ioremap_change_attr((unsigned long)__va(base), id_sz, flags) < 0) {
589 "%s:%d ioremap_change_attr failed %s "
591 current->comm, current->pid,
593 base, (unsigned long long)(base + size));
600 * Internal interface to reserve a range of physical memory with prot.
601 * Reserved non RAM regions only and after successful reserve_memtype,
602 * this func also keeps identity mapping (if any) in sync with this new prot.
604 static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot,
609 unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK);
610 unsigned long flags = want_flags;
612 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
615 * reserve_pfn_range() doesn't support RAM pages. Maintain the current
616 * behavior with RAM pages by returning success.
621 ret = reserve_memtype(paddr, paddr + size, want_flags, &flags);
625 if (flags != want_flags) {
626 if (strict_prot || !is_new_memtype_allowed(want_flags, flags)) {
627 free_memtype(paddr, paddr + size);
628 printk(KERN_ERR "%s:%d map pfn expected mapping type %s"
629 " for %Lx-%Lx, got %s\n",
630 current->comm, current->pid,
631 cattr_name(want_flags),
632 (unsigned long long)paddr,
633 (unsigned long long)(paddr + size),
638 * We allow returning different type than the one requested in
641 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
642 (~_PAGE_CACHE_MASK)) |
646 if (kernel_map_sync_memtype(paddr, size, flags) < 0) {
647 free_memtype(paddr, paddr + size);
654 * Internal interface to free a range of physical memory.
655 * Frees non RAM regions only.
657 static void free_pfn_range(u64 paddr, unsigned long size)
661 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
663 free_memtype(paddr, paddr + size);
667 * track_pfn_vma_copy is called when vma that is covering the pfnmap gets
668 * copied through copy_page_range().
670 * If the vma has a linear pfn mapping for the entire range, we get the prot
671 * from pte and reserve the entire vma range with single reserve_pfn_range call.
672 * Otherwise, we reserve the entire vma range, my ging through the PTEs page
673 * by page to get physical address and protection.
675 int track_pfn_vma_copy(struct vm_area_struct *vma)
679 resource_size_t paddr;
681 unsigned long vma_start = vma->vm_start;
682 unsigned long vma_end = vma->vm_end;
683 unsigned long vma_size = vma_end - vma_start;
689 if (is_linear_pfn_mapping(vma)) {
691 * reserve the whole chunk covered by vma. We need the
692 * starting address and protection from pte.
694 if (follow_phys(vma, vma_start, 0, &prot, &paddr)) {
698 pgprot = __pgprot(prot);
699 return reserve_pfn_range(paddr, vma_size, &pgprot, 1);
702 /* reserve entire vma page by page, using pfn and prot from pte */
703 for (i = 0; i < vma_size; i += PAGE_SIZE) {
704 if (follow_phys(vma, vma_start + i, 0, &prot, &paddr))
707 pgprot = __pgprot(prot);
708 retval = reserve_pfn_range(paddr, PAGE_SIZE, &pgprot, 1);
715 /* Reserve error: Cleanup partial reservation and return error */
716 for (j = 0; j < i; j += PAGE_SIZE) {
717 if (follow_phys(vma, vma_start + j, 0, &prot, &paddr))
720 free_pfn_range(paddr, PAGE_SIZE);
727 * track_pfn_vma_new is called when a _new_ pfn mapping is being established
728 * for physical range indicated by pfn and size.
730 * prot is passed in as a parameter for the new mapping. If the vma has a
731 * linear pfn mapping for the entire range reserve the entire vma range with
732 * single reserve_pfn_range call.
733 * Otherwise, we look t the pfn and size and reserve only the specified range
736 * Note that this function can be called with caller trying to map only a
737 * subrange/page inside the vma.
739 int track_pfn_vma_new(struct vm_area_struct *vma, pgprot_t *prot,
740 unsigned long pfn, unsigned long size)
744 resource_size_t base_paddr;
745 resource_size_t paddr;
746 unsigned long vma_start = vma->vm_start;
747 unsigned long vma_end = vma->vm_end;
748 unsigned long vma_size = vma_end - vma_start;
753 if (is_linear_pfn_mapping(vma)) {
754 /* reserve the whole chunk starting from vm_pgoff */
755 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
756 return reserve_pfn_range(paddr, vma_size, prot, 0);
759 /* reserve page by page using pfn and size */
760 base_paddr = (resource_size_t)pfn << PAGE_SHIFT;
761 for (i = 0; i < size; i += PAGE_SIZE) {
762 paddr = base_paddr + i;
763 retval = reserve_pfn_range(paddr, PAGE_SIZE, prot, 0);
770 /* Reserve error: Cleanup partial reservation and return error */
771 for (j = 0; j < i; j += PAGE_SIZE) {
772 paddr = base_paddr + j;
773 free_pfn_range(paddr, PAGE_SIZE);
780 * untrack_pfn_vma is called while unmapping a pfnmap for a region.
781 * untrack can be called for a specific region indicated by pfn and size or
782 * can be for the entire vma (in which case size can be zero).
784 void untrack_pfn_vma(struct vm_area_struct *vma, unsigned long pfn,
788 resource_size_t paddr;
790 unsigned long vma_start = vma->vm_start;
791 unsigned long vma_end = vma->vm_end;
792 unsigned long vma_size = vma_end - vma_start;
797 if (is_linear_pfn_mapping(vma)) {
798 /* free the whole chunk starting from vm_pgoff */
799 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
800 free_pfn_range(paddr, vma_size);
804 if (size != 0 && size != vma_size) {
805 /* free page by page, using pfn and size */
806 paddr = (resource_size_t)pfn << PAGE_SHIFT;
807 for (i = 0; i < size; i += PAGE_SIZE) {
809 free_pfn_range(paddr, PAGE_SIZE);
812 /* free entire vma, page by page, using the pfn from pte */
813 for (i = 0; i < vma_size; i += PAGE_SIZE) {
814 if (follow_phys(vma, vma_start + i, 0, &prot, &paddr))
817 free_pfn_range(paddr, PAGE_SIZE);
822 pgprot_t pgprot_writecombine(pgprot_t prot)
825 return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC);
827 return pgprot_noncached(prot);
829 EXPORT_SYMBOL_GPL(pgprot_writecombine);
831 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
833 /* get Nth element of the linked list */
834 static struct memtype *memtype_get_idx(loff_t pos)
836 struct memtype *list_node, *print_entry;
839 print_entry = kmalloc(sizeof(struct memtype), GFP_KERNEL);
843 spin_lock(&memtype_lock);
844 list_for_each_entry(list_node, &memtype_list, nd) {
846 *print_entry = *list_node;
847 spin_unlock(&memtype_lock);
852 spin_unlock(&memtype_lock);
858 static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
862 seq_printf(seq, "PAT memtype list:\n");
865 return memtype_get_idx(*pos);
868 static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
871 return memtype_get_idx(*pos);
874 static void memtype_seq_stop(struct seq_file *seq, void *v)
878 static int memtype_seq_show(struct seq_file *seq, void *v)
880 struct memtype *print_entry = (struct memtype *)v;
882 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
883 print_entry->start, print_entry->end);
889 static struct seq_operations memtype_seq_ops = {
890 .start = memtype_seq_start,
891 .next = memtype_seq_next,
892 .stop = memtype_seq_stop,
893 .show = memtype_seq_show,
896 static int memtype_seq_open(struct inode *inode, struct file *file)
898 return seq_open(file, &memtype_seq_ops);
901 static const struct file_operations memtype_fops = {
902 .open = memtype_seq_open,
905 .release = seq_release,
908 static int __init pat_memtype_list_init(void)
910 debugfs_create_file("pat_memtype_list", S_IRUSR, arch_debugfs_dir,
911 NULL, &memtype_fops);
915 late_initcall(pat_memtype_list_init);
917 #endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */