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");
104 /* Set PWT to Write-Combining. All other bits stay the same */
106 * PTE encoding used in Linux:
111 * 000 WB _PAGE_CACHE_WB
112 * 001 WC _PAGE_CACHE_WC
113 * 010 UC- _PAGE_CACHE_UC_MINUS
114 * 011 UC _PAGE_CACHE_UC
117 pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
118 PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
122 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
124 wrmsrl(MSR_IA32_CR_PAT, pat);
127 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
128 smp_processor_id(), boot_pat_state, pat);
133 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype accesses */
136 * Does intersection of PAT memory type and MTRR memory type and returns
137 * the resulting memory type as PAT understands it.
138 * (Type in pat and mtrr will not have same value)
139 * The intersection is based on "Effective Memory Type" tables in IA-32
142 static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
145 * Look for MTRR hint to get the effective type in case where PAT
148 if (req_type == _PAGE_CACHE_WB) {
151 mtrr_type = mtrr_type_lookup(start, end);
152 if (mtrr_type != MTRR_TYPE_WRBACK)
153 return _PAGE_CACHE_UC_MINUS;
155 return _PAGE_CACHE_WB;
161 struct pagerange_state {
162 unsigned long cur_pfn;
168 pagerange_is_ram_callback(unsigned long initial_pfn, unsigned long total_nr_pages, void *arg)
170 struct pagerange_state *state = arg;
172 state->not_ram |= initial_pfn > state->cur_pfn;
173 state->ram |= total_nr_pages > 0;
174 state->cur_pfn = initial_pfn + total_nr_pages;
176 return state->ram && state->not_ram;
179 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
182 unsigned long start_pfn = start >> PAGE_SHIFT;
183 unsigned long end_pfn = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
184 struct pagerange_state state = {start_pfn, 0, 0};
187 * For legacy reasons, physical address range in the legacy ISA
188 * region is tracked as non-RAM. This will allow users of
189 * /dev/mem to map portions of legacy ISA region, even when
190 * some of those portions are listed(or not even listed) with
191 * different e820 types(RAM/reserved/..)
193 if (start_pfn < ISA_END_ADDRESS >> PAGE_SHIFT)
194 start_pfn = ISA_END_ADDRESS >> PAGE_SHIFT;
196 if (start_pfn < end_pfn) {
197 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn,
198 &state, pagerange_is_ram_callback);
201 return (ret > 0) ? -1 : (state.ram ? 1 : 0);
205 * For RAM pages, we use page flags to mark the pages with appropriate type.
206 * Here we do two pass:
207 * - Find the memtype of all the pages in the range, look for any conflicts
208 * - In case of no conflicts, set the new memtype for pages in the range
210 static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
211 unsigned long *new_type)
216 if (req_type == _PAGE_CACHE_UC) {
217 /* We do not support strong UC */
219 req_type = _PAGE_CACHE_UC_MINUS;
222 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
225 page = pfn_to_page(pfn);
226 type = get_page_memtype(page);
228 printk(KERN_INFO "reserve_ram_pages_type failed "
229 "0x%Lx-0x%Lx, track 0x%lx, req 0x%lx\n",
230 start, end, type, req_type);
239 *new_type = req_type;
241 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
242 page = pfn_to_page(pfn);
243 set_page_memtype(page, req_type);
248 static int free_ram_pages_type(u64 start, u64 end)
253 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
254 page = pfn_to_page(pfn);
255 set_page_memtype(page, -1);
261 * req_type typically has one of the:
264 * - _PAGE_CACHE_UC_MINUS
267 * If new_type is NULL, function will return an error if it cannot reserve the
268 * region with req_type. If new_type is non-NULL, function will return
269 * available type in new_type in case of no error. In case of any error
270 * it will return a negative return value.
272 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
273 unsigned long *new_type)
276 unsigned long actual_type;
280 BUG_ON(start >= end); /* end is exclusive */
283 /* This is identical to page table setting without PAT */
285 if (req_type == _PAGE_CACHE_WC)
286 *new_type = _PAGE_CACHE_UC_MINUS;
288 *new_type = req_type & _PAGE_CACHE_MASK;
293 /* Low ISA region is always mapped WB in page table. No need to track */
294 if (x86_platform.is_untracked_pat_range(start, end)) {
296 *new_type = _PAGE_CACHE_WB;
301 * Call mtrr_lookup to get the type hint. This is an
302 * optimization for /dev/mem mmap'ers into WB memory (BIOS
303 * tools and ACPI tools). Use WB request for WB memory and use
304 * UC_MINUS otherwise.
306 actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK);
309 *new_type = actual_type;
311 is_range_ram = pat_pagerange_is_ram(start, end);
312 if (is_range_ram == 1) {
314 err = reserve_ram_pages_type(start, end, req_type, new_type);
317 } else if (is_range_ram < 0) {
321 new = kzalloc(sizeof(struct memtype), GFP_KERNEL);
327 new->type = actual_type;
329 spin_lock(&memtype_lock);
331 err = rbt_memtype_check_insert(new, new_type);
333 printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, "
334 "track %s, req %s\n",
335 start, end, cattr_name(new->type), cattr_name(req_type));
337 spin_unlock(&memtype_lock);
342 spin_unlock(&memtype_lock);
344 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
345 start, end, cattr_name(new->type), cattr_name(req_type),
346 new_type ? cattr_name(*new_type) : "-");
351 int free_memtype(u64 start, u64 end)
355 struct memtype *entry;
360 /* Low ISA region is always mapped WB. No need to track */
361 if (x86_platform.is_untracked_pat_range(start, end))
364 is_range_ram = pat_pagerange_is_ram(start, end);
365 if (is_range_ram == 1) {
367 err = free_ram_pages_type(start, end);
370 } else if (is_range_ram < 0) {
374 spin_lock(&memtype_lock);
375 entry = rbt_memtype_erase(start, end);
376 spin_unlock(&memtype_lock);
379 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
380 current->comm, current->pid, start, end);
386 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
393 * lookup_memtype - Looksup the memory type for a physical address
394 * @paddr: physical address of which memory type needs to be looked up
396 * Only to be called when PAT is enabled
398 * Returns _PAGE_CACHE_WB, _PAGE_CACHE_WC, _PAGE_CACHE_UC_MINUS or
401 static unsigned long lookup_memtype(u64 paddr)
403 int rettype = _PAGE_CACHE_WB;
404 struct memtype *entry;
406 if (x86_platform.is_untracked_pat_range(paddr, paddr + PAGE_SIZE))
409 if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) {
411 page = pfn_to_page(paddr >> PAGE_SHIFT);
412 rettype = get_page_memtype(page);
414 * -1 from get_page_memtype() implies RAM page is in its
415 * default state and not reserved, and hence of type WB
418 rettype = _PAGE_CACHE_WB;
423 spin_lock(&memtype_lock);
425 entry = rbt_memtype_lookup(paddr);
427 rettype = entry->type;
429 rettype = _PAGE_CACHE_UC_MINUS;
431 spin_unlock(&memtype_lock);
436 * io_reserve_memtype - Request a memory type mapping for a region of memory
437 * @start: start (physical address) of the region
438 * @end: end (physical address) of the region
439 * @type: A pointer to memtype, with requested type. On success, requested
440 * or any other compatible type that was available for the region is returned
442 * On success, returns 0
443 * On failure, returns non-zero
445 int io_reserve_memtype(resource_size_t start, resource_size_t end,
448 resource_size_t size = end - start;
449 unsigned long req_type = *type;
450 unsigned long new_type;
453 WARN_ON_ONCE(iomem_map_sanity_check(start, size));
455 ret = reserve_memtype(start, end, req_type, &new_type);
459 if (!is_new_memtype_allowed(start, size, req_type, new_type))
462 if (kernel_map_sync_memtype(start, size, new_type) < 0)
469 free_memtype(start, end);
476 * io_free_memtype - Release a memory type mapping for a region of memory
477 * @start: start (physical address) of the region
478 * @end: end (physical address) of the region
480 void io_free_memtype(resource_size_t start, resource_size_t end)
482 free_memtype(start, end);
485 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
486 unsigned long size, pgprot_t vma_prot)
491 #ifdef CONFIG_STRICT_DEVMEM
492 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
493 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
498 /* This check is needed to avoid cache aliasing when PAT is enabled */
499 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
501 u64 from = ((u64)pfn) << PAGE_SHIFT;
502 u64 to = from + size;
508 while (cursor < to) {
509 if (!devmem_is_allowed(pfn)) {
511 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
512 current->comm, from, to);
520 #endif /* CONFIG_STRICT_DEVMEM */
522 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
523 unsigned long size, pgprot_t *vma_prot)
525 unsigned long flags = _PAGE_CACHE_WB;
527 if (!range_is_allowed(pfn, size))
530 if (file->f_flags & O_DSYNC)
531 flags = _PAGE_CACHE_UC_MINUS;
535 * On the PPro and successors, the MTRRs are used to set
536 * memory types for physical addresses outside main memory,
537 * so blindly setting UC or PWT on those pages is wrong.
538 * For Pentiums and earlier, the surround logic should disable
539 * caching for the high addresses through the KEN pin, but
540 * we maintain the tradition of paranoia in this code.
543 !(boot_cpu_has(X86_FEATURE_MTRR) ||
544 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
545 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
546 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
547 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
548 flags = _PAGE_CACHE_UC;
552 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
558 * Change the memory type for the physial address range in kernel identity
559 * mapping space if that range is a part of identity map.
561 int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
565 if (base >= __pa(high_memory))
568 id_sz = (__pa(high_memory) < base + size) ?
569 __pa(high_memory) - base :
572 if (ioremap_change_attr((unsigned long)__va(base), id_sz, flags) < 0) {
574 "%s:%d ioremap_change_attr failed %s "
576 current->comm, current->pid,
578 base, (unsigned long long)(base + size));
585 * Internal interface to reserve a range of physical memory with prot.
586 * Reserved non RAM regions only and after successful reserve_memtype,
587 * this func also keeps identity mapping (if any) in sync with this new prot.
589 static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot,
594 unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK);
595 unsigned long flags = want_flags;
597 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
600 * reserve_pfn_range() for RAM pages. We do not refcount to keep
601 * track of number of mappings of RAM pages. We can assert that
602 * the type requested matches the type of first page in the range.
608 flags = lookup_memtype(paddr);
609 if (want_flags != flags) {
611 "%s:%d map pfn RAM range req %s for %Lx-%Lx, got %s\n",
612 current->comm, current->pid,
613 cattr_name(want_flags),
614 (unsigned long long)paddr,
615 (unsigned long long)(paddr + size),
617 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
618 (~_PAGE_CACHE_MASK)) |
624 ret = reserve_memtype(paddr, paddr + size, want_flags, &flags);
628 if (flags != want_flags) {
630 !is_new_memtype_allowed(paddr, size, want_flags, flags)) {
631 free_memtype(paddr, paddr + size);
632 printk(KERN_ERR "%s:%d map pfn expected mapping type %s"
633 " for %Lx-%Lx, got %s\n",
634 current->comm, current->pid,
635 cattr_name(want_flags),
636 (unsigned long long)paddr,
637 (unsigned long long)(paddr + size),
642 * We allow returning different type than the one requested in
645 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
646 (~_PAGE_CACHE_MASK)) |
650 if (kernel_map_sync_memtype(paddr, size, flags) < 0) {
651 free_memtype(paddr, paddr + size);
658 * Internal interface to free a range of physical memory.
659 * Frees non RAM regions only.
661 static void free_pfn_range(u64 paddr, unsigned long size)
665 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
667 free_memtype(paddr, paddr + size);
671 * track_pfn_vma_copy is called when vma that is covering the pfnmap gets
672 * copied through copy_page_range().
674 * If the vma has a linear pfn mapping for the entire range, we get the prot
675 * from pte and reserve the entire vma range with single reserve_pfn_range call.
677 int track_pfn_vma_copy(struct vm_area_struct *vma)
679 resource_size_t paddr;
681 unsigned long vma_size = vma->vm_end - vma->vm_start;
684 if (is_linear_pfn_mapping(vma)) {
686 * reserve the whole chunk covered by vma. We need the
687 * starting address and protection from pte.
689 if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) {
693 pgprot = __pgprot(prot);
694 return reserve_pfn_range(paddr, vma_size, &pgprot, 1);
701 * track_pfn_vma_new is called when a _new_ pfn mapping is being established
702 * for physical range indicated by pfn and size.
704 * prot is passed in as a parameter for the new mapping. If the vma has a
705 * linear pfn mapping for the entire range reserve the entire vma range with
706 * single reserve_pfn_range call.
708 int track_pfn_vma_new(struct vm_area_struct *vma, pgprot_t *prot,
709 unsigned long pfn, unsigned long size)
712 resource_size_t paddr;
713 unsigned long vma_size = vma->vm_end - vma->vm_start;
715 if (is_linear_pfn_mapping(vma)) {
716 /* reserve the whole chunk starting from vm_pgoff */
717 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
718 return reserve_pfn_range(paddr, vma_size, prot, 0);
724 /* for vm_insert_pfn and friends, we set prot based on lookup */
725 flags = lookup_memtype(pfn << PAGE_SHIFT);
726 *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) |
733 * untrack_pfn_vma is called while unmapping a pfnmap for a region.
734 * untrack can be called for a specific region indicated by pfn and size or
735 * can be for the entire vma (in which case size can be zero).
737 void untrack_pfn_vma(struct vm_area_struct *vma, unsigned long pfn,
740 resource_size_t paddr;
741 unsigned long vma_size = vma->vm_end - vma->vm_start;
743 if (is_linear_pfn_mapping(vma)) {
744 /* free the whole chunk starting from vm_pgoff */
745 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
746 free_pfn_range(paddr, vma_size);
751 pgprot_t pgprot_writecombine(pgprot_t prot)
754 return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC);
756 return pgprot_noncached(prot);
758 EXPORT_SYMBOL_GPL(pgprot_writecombine);
760 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
762 static struct memtype *memtype_get_idx(loff_t pos)
764 struct memtype *print_entry;
767 print_entry = kzalloc(sizeof(struct memtype), GFP_KERNEL);
771 spin_lock(&memtype_lock);
772 ret = rbt_memtype_copy_nth_element(print_entry, pos);
773 spin_unlock(&memtype_lock);
783 static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
787 seq_printf(seq, "PAT memtype list:\n");
790 return memtype_get_idx(*pos);
793 static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
796 return memtype_get_idx(*pos);
799 static void memtype_seq_stop(struct seq_file *seq, void *v)
803 static int memtype_seq_show(struct seq_file *seq, void *v)
805 struct memtype *print_entry = (struct memtype *)v;
807 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
808 print_entry->start, print_entry->end);
814 static const struct seq_operations memtype_seq_ops = {
815 .start = memtype_seq_start,
816 .next = memtype_seq_next,
817 .stop = memtype_seq_stop,
818 .show = memtype_seq_show,
821 static int memtype_seq_open(struct inode *inode, struct file *file)
823 return seq_open(file, &memtype_seq_ops);
826 static const struct file_operations memtype_fops = {
827 .open = memtype_seq_open,
830 .release = seq_release,
833 static int __init pat_memtype_list_init(void)
836 debugfs_create_file("pat_memtype_list", S_IRUSR,
837 arch_debugfs_dir, NULL, &memtype_fops);
842 late_initcall(pat_memtype_list_init);
844 #endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */