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/slab.h>
18 #include <linux/rbtree.h>
20 #include <asm/cacheflush.h>
21 #include <asm/processor.h>
22 #include <asm/tlbflush.h>
23 #include <asm/x86_init.h>
24 #include <asm/pgtable.h>
25 #include <asm/fcntl.h>
33 #include "pat_internal.h"
36 int __read_mostly pat_enabled = 1;
38 static inline void pat_disable(const char *reason)
41 printk(KERN_INFO "%s\n", reason);
44 static int __init nopat(char *str)
46 pat_disable("PAT support disabled.");
49 early_param("nopat", nopat);
51 static inline void pat_disable(const char *reason)
60 static int __init pat_debug_setup(char *str)
65 __setup("debugpat", pat_debug_setup);
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))
83 bool boot_cpu = !boot_pat_state;
89 if (!boot_pat_state) {
90 pat_disable("PAT not supported by CPU.");
94 * If this happens we are on a secondary CPU, but
95 * switched to PAT on the boot CPU. We have no way to
98 printk(KERN_ERR "PAT enabled, "
99 "but not supported by secondary CPU\n");
105 /* Set PWT to Write-Combining. All other bits stay the same */
107 * PTE encoding used in Linux:
112 * 000 WB _PAGE_CACHE_WB
113 * 001 WC _PAGE_CACHE_WC
114 * 010 UC- _PAGE_CACHE_UC_MINUS
115 * 011 UC _PAGE_CACHE_UC
118 pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
119 PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
123 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
125 wrmsrl(MSR_IA32_CR_PAT, pat);
128 * PAT settings are part of the hypervisor interface, and their
129 * assignment cannot be changed.
131 rdmsrl(MSR_IA32_CR_PAT, pat);
133 boot_pat_state = pat;
137 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
138 smp_processor_id(), boot_pat_state, pat);
143 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype accesses */
145 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end);
146 static inline u8 _mtrr_type_lookup(u64 start, u64 end)
148 if (is_initial_xendomain())
149 return mtrr_type_lookup(start, end);
150 return pat_pagerange_is_ram(start, end) > 0
151 ? MTRR_TYPE_WRCOMB : MTRR_TYPE_UNCACHABLE;
153 #define mtrr_type_lookup _mtrr_type_lookup
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_WRBACK)
173 return _PAGE_CACHE_UC_MINUS;
175 return _PAGE_CACHE_WB;
181 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
183 int ram_page = 0, not_rampage = 0;
184 unsigned long page_nr;
186 for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
189 * For legacy reasons, physical address range in the legacy ISA
190 * region is tracked as non-RAM. This will allow users of
191 * /dev/mem to map portions of legacy ISA region, even when
192 * some of those portions are listed(or not even listed) with
193 * different e820 types(RAM/reserved/..)
195 if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
196 page_is_ram(mfn_to_local_pfn(page_nr)))
201 if (ram_page == not_rampage)
209 * For RAM pages, we use page flags to mark the pages with appropriate type.
210 * Here we do two pass:
211 * - Find the memtype of all the pages in the range, look for any conflicts
212 * - In case of no conflicts, set the new memtype for pages in the range
214 static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
215 unsigned long *new_type)
220 if (req_type == _PAGE_CACHE_UC) {
221 /* We do not support strong UC */
223 req_type = _PAGE_CACHE_UC_MINUS;
226 for (mfn = (start >> PAGE_SHIFT); mfn < (end >> PAGE_SHIFT); ++mfn) {
227 unsigned long type, pfn = mfn_to_local_pfn(mfn);
229 BUG_ON(!pfn_valid(pfn));
230 page = pfn_to_page(pfn);
231 type = get_page_memtype(page);
233 printk(KERN_INFO "reserve_ram_pages_type failed "
234 "0x%Lx-0x%Lx, track 0x%lx, req 0x%lx\n",
235 start, end, type, req_type);
244 *new_type = req_type;
246 for (mfn = (start >> PAGE_SHIFT); mfn < (end >> PAGE_SHIFT); ++mfn) {
247 page = pfn_to_page(mfn_to_local_pfn(mfn));
248 set_page_memtype(page, req_type);
253 static int free_ram_pages_type(u64 start, u64 end)
258 for (mfn = (start >> PAGE_SHIFT); mfn < (end >> PAGE_SHIFT); ++mfn) {
259 unsigned long pfn = mfn_to_local_pfn(mfn);
261 BUG_ON(!pfn_valid(pfn));
262 page = pfn_to_page(pfn);
263 set_page_memtype(page, -1);
269 * req_type typically has one of the:
272 * - _PAGE_CACHE_UC_MINUS
275 * If new_type is NULL, function will return an error if it cannot reserve the
276 * region with req_type. If new_type is non-NULL, function will return
277 * available type in new_type in case of no error. In case of any error
278 * it will return a negative return value.
280 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
281 unsigned long *new_type)
284 unsigned long actual_type;
288 BUG_ON(start >= end); /* end is exclusive */
291 /* This is identical to page table setting without PAT */
293 if (req_type == _PAGE_CACHE_WC)
294 *new_type = _PAGE_CACHE_UC_MINUS;
296 *new_type = req_type & _PAGE_CACHE_MASK;
301 /* Low ISA region is always mapped WB in page table. No need to track */
302 if (x86_platform.is_untracked_pat_range(start, end)) {
304 *new_type = _PAGE_CACHE_WB;
309 * Call mtrr_lookup to get the type hint. This is an
310 * optimization for /dev/mem mmap'ers into WB memory (BIOS
311 * tools and ACPI tools). Use WB request for WB memory and use
312 * UC_MINUS otherwise.
314 actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK);
317 *new_type = actual_type;
319 is_range_ram = pat_pagerange_is_ram(start, end);
320 if (is_range_ram == 1) {
322 err = reserve_ram_pages_type(start, end, req_type, new_type);
325 } else if (is_range_ram < 0) {
329 new = kzalloc(sizeof(struct memtype), GFP_KERNEL);
335 new->type = actual_type;
337 spin_lock(&memtype_lock);
339 err = rbt_memtype_check_insert(new, new_type);
341 printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, "
342 "track %s, req %s\n",
343 start, end, cattr_name(new->type), cattr_name(req_type));
345 spin_unlock(&memtype_lock);
350 spin_unlock(&memtype_lock);
352 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
353 start, end, cattr_name(new->type), cattr_name(req_type),
354 new_type ? cattr_name(*new_type) : "-");
359 int free_memtype(u64 start, u64 end)
363 struct memtype *entry;
368 /* Low ISA region is always mapped WB. No need to track */
369 if (x86_platform.is_untracked_pat_range(start, end))
372 is_range_ram = pat_pagerange_is_ram(start, end);
373 if (is_range_ram == 1) {
375 err = free_ram_pages_type(start, end);
378 } else if (is_range_ram < 0) {
382 spin_lock(&memtype_lock);
383 entry = rbt_memtype_erase(start, end);
384 spin_unlock(&memtype_lock);
387 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
388 current->comm, current->pid, start, end);
394 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
402 * lookup_memtype - Looksup the memory type for a physical address
403 * @paddr: physical address of which memory type needs to be looked up
405 * Only to be called when PAT is enabled
407 * Returns _PAGE_CACHE_WB, _PAGE_CACHE_WC, _PAGE_CACHE_UC_MINUS or
410 static unsigned long lookup_memtype(u64 paddr)
412 int rettype = _PAGE_CACHE_WB;
413 struct memtype *entry;
415 if (x86_platform.is_untracked_pat_range(paddr, paddr + PAGE_SIZE))
418 if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) {
420 page = pfn_to_page(paddr >> PAGE_SHIFT);
421 rettype = get_page_memtype(page);
423 * -1 from get_page_memtype() implies RAM page is in its
424 * default state and not reserved, and hence of type WB
427 rettype = _PAGE_CACHE_WB;
432 spin_lock(&memtype_lock);
434 entry = rbt_memtype_lookup(paddr);
436 rettype = entry->type;
438 rettype = _PAGE_CACHE_UC_MINUS;
440 spin_unlock(&memtype_lock);
446 * io_reserve_memtype - Request a memory type mapping for a region of memory
447 * @start: start (physical address) of the region
448 * @end: end (physical address) of the region
449 * @type: A pointer to memtype, with requested type. On success, requested
450 * or any other compatible type that was available for the region is returned
452 * On success, returns 0
453 * On failure, returns non-zero
455 int io_reserve_memtype(resource_size_t start, resource_size_t end,
458 resource_size_t size = end - start;
459 unsigned long req_type = *type;
460 unsigned long new_type;
463 WARN_ON_ONCE(iomem_map_sanity_check(start, size));
465 ret = reserve_memtype(start, end, req_type, &new_type);
469 if (!is_new_memtype_allowed(start, size, req_type, new_type))
472 if (kernel_map_sync_memtype(start, size, new_type) < 0)
479 free_memtype(start, end);
486 * io_free_memtype - Release a memory type mapping for a region of memory
487 * @start: start (physical address) of the region
488 * @end: end (physical address) of the region
490 void io_free_memtype(resource_size_t start, resource_size_t end)
492 free_memtype(start, end);
495 pgprot_t phys_mem_access_prot(struct file *file, unsigned long mfn,
496 unsigned long size, pgprot_t vma_prot)
501 #ifdef CONFIG_STRICT_DEVMEM
502 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
503 static inline int range_is_allowed(unsigned long mfn, unsigned long size)
508 /* This check is needed to avoid cache aliasing when PAT is enabled */
509 static inline int range_is_allowed(unsigned long mfn, unsigned long size)
511 u64 from = ((u64)mfn) << PAGE_SHIFT;
512 u64 to = from + size;
518 while (cursor < to) {
519 if (!devmem_is_allowed(mfn)) {
521 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
522 current->comm, from, to);
530 #endif /* CONFIG_STRICT_DEVMEM */
532 int phys_mem_access_prot_allowed(struct file *file, unsigned long mfn,
533 unsigned long size, pgprot_t *vma_prot)
535 unsigned long flags = _PAGE_CACHE_WB;
537 if (!range_is_allowed(mfn, size))
540 if (file->f_flags & O_DSYNC)
541 flags = _PAGE_CACHE_UC_MINUS;
543 #ifndef CONFIG_X86_32
544 #ifndef CONFIG_XEN /* Xen sets correct MTRR type on non-RAM for us. */
546 * On the PPro and successors, the MTRRs are used to set
547 * memory types for physical addresses outside main memory,
548 * so blindly setting UC or PWT on those pages is wrong.
549 * For Pentiums and earlier, the surround logic should disable
550 * caching for the high addresses through the KEN pin, but
551 * we maintain the tradition of paranoia in this code.
554 !(boot_cpu_has(X86_FEATURE_MTRR) ||
555 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
556 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
557 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
558 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
559 flags = _PAGE_CACHE_UC;
564 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
570 * Change the memory type for the physial address range in kernel identity
571 * mapping space if that range is a part of identity map.
573 int kernel_map_sync_memtype(u64 ma, unsigned long size, unsigned long flags)
575 return ioremap_check_change_attr(ma >> PAGE_SHIFT, size, flags);
580 * Internal interface to reserve a range of physical memory with prot.
581 * Reserved non RAM regions only and after successful reserve_memtype,
582 * this func also keeps identity mapping (if any) in sync with this new prot.
584 static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot,
589 unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK);
590 unsigned long flags = want_flags;
592 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
595 * reserve_pfn_range() for RAM pages. We do not refcount to keep
596 * track of number of mappings of RAM pages. We can assert that
597 * the type requested matches the type of first page in the range.
603 flags = lookup_memtype(paddr);
604 if (want_flags != flags) {
606 "%s:%d map pfn RAM range req %s for %Lx-%Lx, got %s\n",
607 current->comm, current->pid,
608 cattr_name(want_flags),
609 (unsigned long long)paddr,
610 (unsigned long long)(paddr + size),
612 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
613 (~_PAGE_CACHE_MASK)) |
619 ret = reserve_memtype(paddr, paddr + size, want_flags, &flags);
623 if (flags != want_flags) {
625 !is_new_memtype_allowed(paddr, size, want_flags, flags)) {
626 free_memtype(paddr, paddr + size);
627 printk(KERN_ERR "%s:%d map pfn expected mapping type %s"
628 " for %Lx-%Lx, got %s\n",
629 current->comm, current->pid,
630 cattr_name(want_flags),
631 (unsigned long long)paddr,
632 (unsigned long long)(paddr + size),
637 * We allow returning different type than the one requested in
640 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
641 (~_PAGE_CACHE_MASK)) |
645 if (kernel_map_sync_memtype(paddr, size, flags) < 0) {
646 free_memtype(paddr, paddr + size);
653 * Internal interface to free a range of physical memory.
654 * Frees non RAM regions only.
656 static void free_pfn_range(u64 paddr, unsigned long size)
660 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
662 free_memtype(paddr, paddr + size);
666 * track_pfn_vma_copy is called when vma that is covering the pfnmap gets
667 * copied through copy_page_range().
669 * If the vma has a linear pfn mapping for the entire range, we get the prot
670 * from pte and reserve the entire vma range with single reserve_pfn_range call.
672 int track_pfn_vma_copy(struct vm_area_struct *vma)
674 resource_size_t paddr;
676 unsigned long vma_size = vma->vm_end - vma->vm_start;
679 if (is_linear_pfn_mapping(vma)) {
681 * reserve the whole chunk covered by vma. We need the
682 * starting address and protection from pte.
684 if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) {
688 pgprot = __pgprot(prot);
689 return reserve_pfn_range(paddr, vma_size, &pgprot, 1);
696 * track_pfn_vma_new is called when a _new_ pfn mapping is being established
697 * for physical range indicated by pfn and size.
699 * prot is passed in as a parameter for the new mapping. If the vma has a
700 * linear pfn mapping for the entire range reserve the entire vma range with
701 * single reserve_pfn_range call.
703 int track_pfn_vma_new(struct vm_area_struct *vma, pgprot_t *prot,
704 unsigned long pfn, unsigned long size)
707 resource_size_t paddr;
708 unsigned long vma_size = vma->vm_end - vma->vm_start;
710 if (is_linear_pfn_mapping(vma)) {
711 /* reserve the whole chunk starting from vm_pgoff */
712 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
713 return reserve_pfn_range(paddr, vma_size, prot, 0);
719 /* for vm_insert_pfn and friends, we set prot based on lookup */
720 flags = lookup_memtype(pfn << PAGE_SHIFT);
721 *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) |
728 * untrack_pfn_vma is called while unmapping a pfnmap for a region.
729 * untrack can be called for a specific region indicated by pfn and size or
730 * can be for the entire vma (in which case size can be zero).
732 void untrack_pfn_vma(struct vm_area_struct *vma, unsigned long pfn,
735 resource_size_t paddr;
736 unsigned long vma_size = vma->vm_end - vma->vm_start;
738 if (is_linear_pfn_mapping(vma)) {
739 /* free the whole chunk starting from vm_pgoff */
740 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
741 free_pfn_range(paddr, vma_size);
745 #endif /* CONFIG_XEN */
747 pgprot_t pgprot_writecombine(pgprot_t prot)
750 return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC);
752 return pgprot_noncached(prot);
754 EXPORT_SYMBOL_GPL(pgprot_writecombine);
756 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
758 static struct memtype *memtype_get_idx(loff_t pos)
760 struct memtype *print_entry;
763 print_entry = kzalloc(sizeof(struct memtype), GFP_KERNEL);
767 spin_lock(&memtype_lock);
768 ret = rbt_memtype_copy_nth_element(print_entry, pos);
769 spin_unlock(&memtype_lock);
779 static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
783 seq_printf(seq, "PAT memtype list:\n");
786 return memtype_get_idx(*pos);
789 static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
792 return memtype_get_idx(*pos);
795 static void memtype_seq_stop(struct seq_file *seq, void *v)
799 static int memtype_seq_show(struct seq_file *seq, void *v)
801 struct memtype *print_entry = (struct memtype *)v;
803 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
804 print_entry->start, print_entry->end);
810 static const struct seq_operations memtype_seq_ops = {
811 .start = memtype_seq_start,
812 .next = memtype_seq_next,
813 .stop = memtype_seq_stop,
814 .show = memtype_seq_show,
817 static int memtype_seq_open(struct inode *inode, struct file *file)
819 return seq_open(file, &memtype_seq_ops);
822 static const struct file_operations memtype_fops = {
823 .open = memtype_seq_open,
826 .release = seq_release,
829 static int __init pat_memtype_list_init(void)
832 debugfs_create_file("pat_memtype_list", S_IRUSR,
833 arch_debugfs_dir, NULL, &memtype_fops);
838 late_initcall(pat_memtype_list_init);
840 #endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */