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 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
163 int ram_page = 0, not_rampage = 0;
164 unsigned long page_nr;
166 for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
169 * For legacy reasons, physical address range in the legacy ISA
170 * region is tracked as non-RAM. This will allow users of
171 * /dev/mem to map portions of legacy ISA region, even when
172 * some of those portions are listed(or not even listed) with
173 * different e820 types(RAM/reserved/..)
175 if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
176 page_is_ram(page_nr))
181 if (ram_page == not_rampage)
189 * For RAM pages, we use page flags to mark the pages with appropriate type.
190 * Here we do two pass:
191 * - Find the memtype of all the pages in the range, look for any conflicts
192 * - In case of no conflicts, set the new memtype for pages in the range
194 static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
195 unsigned long *new_type)
200 if (req_type == _PAGE_CACHE_UC) {
201 /* We do not support strong UC */
203 req_type = _PAGE_CACHE_UC_MINUS;
206 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
209 page = pfn_to_page(pfn);
210 type = get_page_memtype(page);
212 printk(KERN_INFO "reserve_ram_pages_type failed [mem %#010Lx-%#010Lx], track 0x%lx, req 0x%lx\n",
213 start, end - 1, type, req_type);
222 *new_type = req_type;
224 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
225 page = pfn_to_page(pfn);
226 set_page_memtype(page, req_type);
231 static int free_ram_pages_type(u64 start, u64 end)
236 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
237 page = pfn_to_page(pfn);
238 set_page_memtype(page, -1);
244 * req_type typically has one of the:
247 * - _PAGE_CACHE_UC_MINUS
250 * If new_type is NULL, function will return an error if it cannot reserve the
251 * region with req_type. If new_type is non-NULL, function will return
252 * available type in new_type in case of no error. In case of any error
253 * it will return a negative return value.
255 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
256 unsigned long *new_type)
259 unsigned long actual_type;
263 BUG_ON(start >= end); /* end is exclusive */
266 /* This is identical to page table setting without PAT */
268 if (req_type == _PAGE_CACHE_WC)
269 *new_type = _PAGE_CACHE_UC_MINUS;
271 *new_type = req_type & _PAGE_CACHE_MASK;
276 /* Low ISA region is always mapped WB in page table. No need to track */
277 if (x86_platform.is_untracked_pat_range(start, end)) {
279 *new_type = _PAGE_CACHE_WB;
284 * Call mtrr_lookup to get the type hint. This is an
285 * optimization for /dev/mem mmap'ers into WB memory (BIOS
286 * tools and ACPI tools). Use WB request for WB memory and use
287 * UC_MINUS otherwise.
289 actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK);
292 *new_type = actual_type;
294 is_range_ram = pat_pagerange_is_ram(start, end);
295 if (is_range_ram == 1) {
297 err = reserve_ram_pages_type(start, end, req_type, new_type);
300 } else if (is_range_ram < 0) {
304 new = kzalloc(sizeof(struct memtype), GFP_KERNEL);
310 new->type = actual_type;
312 spin_lock(&memtype_lock);
314 err = rbt_memtype_check_insert(new, new_type);
316 printk(KERN_INFO "reserve_memtype failed [mem %#010Lx-%#010Lx], track %s, req %s\n",
318 cattr_name(new->type), cattr_name(req_type));
320 spin_unlock(&memtype_lock);
325 spin_unlock(&memtype_lock);
327 dprintk("reserve_memtype added [mem %#010Lx-%#010Lx], track %s, req %s, ret %s\n",
328 start, end - 1, cattr_name(new->type), cattr_name(req_type),
329 new_type ? cattr_name(*new_type) : "-");
334 int free_memtype(u64 start, u64 end)
338 struct memtype *entry;
343 /* Low ISA region is always mapped WB. No need to track */
344 if (x86_platform.is_untracked_pat_range(start, end))
347 is_range_ram = pat_pagerange_is_ram(start, end);
348 if (is_range_ram == 1) {
350 err = free_ram_pages_type(start, end);
353 } else if (is_range_ram < 0) {
357 spin_lock(&memtype_lock);
358 entry = rbt_memtype_erase(start, end);
359 spin_unlock(&memtype_lock);
362 printk(KERN_INFO "%s:%d freeing invalid memtype [mem %#010Lx-%#010Lx]\n",
363 current->comm, current->pid, start, end - 1);
369 dprintk("free_memtype request [mem %#010Lx-%#010Lx]\n", start, end - 1);
376 * lookup_memtype - Looksup the memory type for a physical address
377 * @paddr: physical address of which memory type needs to be looked up
379 * Only to be called when PAT is enabled
381 * Returns _PAGE_CACHE_WB, _PAGE_CACHE_WC, _PAGE_CACHE_UC_MINUS or
384 static unsigned long lookup_memtype(u64 paddr)
386 int rettype = _PAGE_CACHE_WB;
387 struct memtype *entry;
389 if (x86_platform.is_untracked_pat_range(paddr, paddr + PAGE_SIZE))
392 if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) {
394 page = pfn_to_page(paddr >> PAGE_SHIFT);
395 rettype = get_page_memtype(page);
397 * -1 from get_page_memtype() implies RAM page is in its
398 * default state and not reserved, and hence of type WB
401 rettype = _PAGE_CACHE_WB;
406 spin_lock(&memtype_lock);
408 entry = rbt_memtype_lookup(paddr);
410 rettype = entry->type;
412 rettype = _PAGE_CACHE_UC_MINUS;
414 spin_unlock(&memtype_lock);
419 * io_reserve_memtype - Request a memory type mapping for a region of memory
420 * @start: start (physical address) of the region
421 * @end: end (physical address) of the region
422 * @type: A pointer to memtype, with requested type. On success, requested
423 * or any other compatible type that was available for the region is returned
425 * On success, returns 0
426 * On failure, returns non-zero
428 int io_reserve_memtype(resource_size_t start, resource_size_t end,
431 resource_size_t size = end - start;
432 unsigned long req_type = *type;
433 unsigned long new_type;
436 WARN_ON_ONCE(iomem_map_sanity_check(start, size));
438 ret = reserve_memtype(start, end, req_type, &new_type);
442 if (!is_new_memtype_allowed(start, size, req_type, new_type))
445 if (kernel_map_sync_memtype(start, size, new_type) < 0)
452 free_memtype(start, end);
459 * io_free_memtype - Release a memory type mapping for a region of memory
460 * @start: start (physical address) of the region
461 * @end: end (physical address) of the region
463 void io_free_memtype(resource_size_t start, resource_size_t end)
465 free_memtype(start, end);
468 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
469 unsigned long size, pgprot_t vma_prot)
474 #ifdef CONFIG_STRICT_DEVMEM
475 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
476 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
481 /* This check is needed to avoid cache aliasing when PAT is enabled */
482 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
484 u64 from = ((u64)pfn) << PAGE_SHIFT;
485 u64 to = from + size;
491 while (cursor < to) {
492 if (!devmem_is_allowed(pfn)) {
493 printk(KERN_INFO "Program %s tried to access /dev/mem between [mem %#010Lx-%#010Lx]\n",
494 current->comm, from, to - 1);
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 unsigned long flags = _PAGE_CACHE_WB;
509 if (!range_is_allowed(pfn, size))
512 if (file->f_flags & O_DSYNC)
513 flags = _PAGE_CACHE_UC_MINUS;
517 * On the PPro and successors, the MTRRs are used to set
518 * memory types for physical addresses outside main memory,
519 * so blindly setting UC or PWT on those pages is wrong.
520 * For Pentiums and earlier, the surround logic should disable
521 * caching for the high addresses through the KEN pin, but
522 * we maintain the tradition of paranoia in this code.
525 !(boot_cpu_has(X86_FEATURE_MTRR) ||
526 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
527 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
528 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
529 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
530 flags = _PAGE_CACHE_UC;
534 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
540 * Change the memory type for the physial address range in kernel identity
541 * mapping space if that range is a part of identity map.
543 int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
547 if (base >= __pa(high_memory))
550 id_sz = (__pa(high_memory) < base + size) ?
551 __pa(high_memory) - base :
554 if (ioremap_change_attr((unsigned long)__va(base), id_sz, flags) < 0) {
555 printk(KERN_INFO "%s:%d ioremap_change_attr failed %s "
556 "for [mem %#010Lx-%#010Lx]\n",
557 current->comm, current->pid,
559 base, (unsigned long long)(base + size-1));
566 * Internal interface to reserve a range of physical memory with prot.
567 * Reserved non RAM regions only and after successful reserve_memtype,
568 * this func also keeps identity mapping (if any) in sync with this new prot.
570 static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot,
575 unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK);
576 unsigned long flags = want_flags;
578 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
581 * reserve_pfn_range() for RAM pages. We do not refcount to keep
582 * track of number of mappings of RAM pages. We can assert that
583 * the type requested matches the type of first page in the range.
589 flags = lookup_memtype(paddr);
590 if (want_flags != flags) {
591 printk(KERN_WARNING "%s:%d map pfn RAM range req %s for [mem %#010Lx-%#010Lx], got %s\n",
592 current->comm, current->pid,
593 cattr_name(want_flags),
594 (unsigned long long)paddr,
595 (unsigned long long)(paddr + size - 1),
597 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
598 (~_PAGE_CACHE_MASK)) |
604 ret = reserve_memtype(paddr, paddr + size, want_flags, &flags);
608 if (flags != want_flags) {
610 !is_new_memtype_allowed(paddr, size, want_flags, flags)) {
611 free_memtype(paddr, paddr + size);
612 printk(KERN_ERR "%s:%d map pfn expected mapping type %s"
613 " for [mem %#010Lx-%#010Lx], got %s\n",
614 current->comm, current->pid,
615 cattr_name(want_flags),
616 (unsigned long long)paddr,
617 (unsigned long long)(paddr + size - 1),
622 * We allow returning different type than the one requested in
625 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
626 (~_PAGE_CACHE_MASK)) |
630 if (kernel_map_sync_memtype(paddr, size, flags) < 0) {
631 free_memtype(paddr, paddr + size);
638 * Internal interface to free a range of physical memory.
639 * Frees non RAM regions only.
641 static void free_pfn_range(u64 paddr, unsigned long size)
645 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
647 free_memtype(paddr, paddr + size);
651 * track_pfn_vma_copy is called when vma that is covering the pfnmap gets
652 * copied through copy_page_range().
654 * If the vma has a linear pfn mapping for the entire range, we get the prot
655 * from pte and reserve the entire vma range with single reserve_pfn_range call.
657 int track_pfn_vma_copy(struct vm_area_struct *vma)
659 resource_size_t paddr;
661 unsigned long vma_size = vma->vm_end - vma->vm_start;
664 if (is_linear_pfn_mapping(vma)) {
666 * reserve the whole chunk covered by vma. We need the
667 * starting address and protection from pte.
669 if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) {
673 pgprot = __pgprot(prot);
674 return reserve_pfn_range(paddr, vma_size, &pgprot, 1);
681 * track_pfn_vma_new is called when a _new_ pfn mapping is being established
682 * for physical range indicated by pfn and size.
684 * prot is passed in as a parameter for the new mapping. If the vma has a
685 * linear pfn mapping for the entire range reserve the entire vma range with
686 * single reserve_pfn_range call.
688 int track_pfn_vma_new(struct vm_area_struct *vma, pgprot_t *prot,
689 unsigned long pfn, unsigned long size)
692 resource_size_t paddr;
693 unsigned long vma_size = vma->vm_end - vma->vm_start;
695 if (is_linear_pfn_mapping(vma)) {
696 /* reserve the whole chunk starting from vm_pgoff */
697 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
698 return reserve_pfn_range(paddr, vma_size, prot, 0);
704 /* for vm_insert_pfn and friends, we set prot based on lookup */
705 flags = lookup_memtype(pfn << PAGE_SHIFT);
706 *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) |
713 * untrack_pfn_vma is called while unmapping a pfnmap for a region.
714 * untrack can be called for a specific region indicated by pfn and size or
715 * can be for the entire vma (in which case size can be zero).
717 void untrack_pfn_vma(struct vm_area_struct *vma, unsigned long pfn,
720 resource_size_t paddr;
721 unsigned long vma_size = vma->vm_end - vma->vm_start;
723 if (is_linear_pfn_mapping(vma)) {
724 /* free the whole chunk starting from vm_pgoff */
725 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
726 free_pfn_range(paddr, vma_size);
731 pgprot_t pgprot_writecombine(pgprot_t prot)
734 return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC);
736 return pgprot_noncached(prot);
738 EXPORT_SYMBOL_GPL(pgprot_writecombine);
740 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
742 static struct memtype *memtype_get_idx(loff_t pos)
744 struct memtype *print_entry;
747 print_entry = kzalloc(sizeof(struct memtype), GFP_KERNEL);
751 spin_lock(&memtype_lock);
752 ret = rbt_memtype_copy_nth_element(print_entry, pos);
753 spin_unlock(&memtype_lock);
763 static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
767 seq_printf(seq, "PAT memtype list:\n");
770 return memtype_get_idx(*pos);
773 static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
776 return memtype_get_idx(*pos);
779 static void memtype_seq_stop(struct seq_file *seq, void *v)
783 static int memtype_seq_show(struct seq_file *seq, void *v)
785 struct memtype *print_entry = (struct memtype *)v;
787 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
788 print_entry->start, print_entry->end);
794 static const struct seq_operations memtype_seq_ops = {
795 .start = memtype_seq_start,
796 .next = memtype_seq_next,
797 .stop = memtype_seq_stop,
798 .show = memtype_seq_show,
801 static int memtype_seq_open(struct inode *inode, struct file *file)
803 return seq_open(file, &memtype_seq_ops);
806 static const struct file_operations memtype_fops = {
807 .open = memtype_seq_open,
810 .release = seq_release,
813 static int __init pat_memtype_list_init(void)
816 debugfs_create_file("pat_memtype_list", S_IRUSR,
817 arch_debugfs_dir, NULL, &memtype_fops);
822 late_initcall(pat_memtype_list_init);
824 #endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */