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.
11 #include <linux/kernel.h>
12 #include <linux/gfp.h>
14 #include <linux/bootmem.h>
15 #include <linux/debugfs.h>
16 #include <linux/seq_file.h>
19 #include <asm/tlbflush.h>
20 #include <asm/processor.h>
22 #include <asm/pgtable.h>
25 #include <asm/cacheflush.h>
26 #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;
49 static int __init pat_debug_setup(char *str)
54 __setup("debugpat", pat_debug_setup);
56 #define dprintk(fmt, arg...) \
57 do { if (debug_enable) printk(KERN_INFO fmt, ##arg); } while (0)
60 static u64 __read_mostly boot_pat_state;
63 PAT_UC = 0, /* uncached */
64 PAT_WC = 1, /* Write combining */
65 PAT_WT = 4, /* Write Through */
66 PAT_WP = 5, /* Write Protected */
67 PAT_WB = 6, /* Write Back (default) */
68 PAT_UC_MINUS = 7, /* UC, but can be overriden by MTRR */
71 #define PAT(x, y) ((u64)PAT_ ## y << ((x)*8))
81 if (!cpu_has_pat && boot_pat_state) {
83 * If this happens we are on a secondary CPU, but
84 * switched to PAT on the boot CPU. We have no way to
87 printk(KERN_ERR "PAT enabled, "
88 "but not supported by secondary CPU\n");
92 /* Set PWT to Write-Combining. All other bits stay the same */
94 * PTE encoding used in Linux:
99 * 000 WB _PAGE_CACHE_WB
100 * 001 WC _PAGE_CACHE_WC
101 * 010 UC- _PAGE_CACHE_UC_MINUS
102 * 011 UC _PAGE_CACHE_UC
105 pat = PAT(0, WB) | PAT(1, WC) | PAT(2, UC_MINUS) | PAT(3, UC) |
106 PAT(4, WB) | PAT(5, WC) | PAT(6, UC_MINUS) | PAT(7, UC);
110 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
112 wrmsrl(MSR_IA32_CR_PAT, pat);
113 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
114 smp_processor_id(), boot_pat_state, pat);
119 static char *cattr_name(unsigned long flags)
121 switch (flags & _PAGE_CACHE_MASK) {
122 case _PAGE_CACHE_UC: return "uncached";
123 case _PAGE_CACHE_UC_MINUS: return "uncached-minus";
124 case _PAGE_CACHE_WB: return "write-back";
125 case _PAGE_CACHE_WC: return "write-combining";
126 default: return "broken";
131 * The global memtype list keeps track of memory type for specific
132 * physical memory areas. Conflicting memory types in different
133 * mappings can cause CPU cache corruption. To avoid this we keep track.
135 * The list is sorted based on starting address and can contain multiple
136 * entries for each address (this allows reference counting for overlapping
137 * areas). All the aliases have the same cache attributes of course.
138 * Zero attributes are represented as holes.
140 * Currently the data structure is a list because the number of mappings
141 * are expected to be relatively small. If this should be a problem
142 * it could be changed to a rbtree or similar.
144 * memtype_lock protects the whole list.
154 static LIST_HEAD(memtype_list);
155 static DEFINE_SPINLOCK(memtype_lock); /* protects memtype list */
158 * Does intersection of PAT memory type and MTRR memory type and returns
159 * the resulting memory type as PAT understands it.
160 * (Type in pat and mtrr will not have same value)
161 * The intersection is based on "Effective Memory Type" tables in IA-32
164 static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
167 * Look for MTRR hint to get the effective type in case where PAT
170 if (req_type == _PAGE_CACHE_WB) {
173 mtrr_type = mtrr_type_lookup(start, end);
174 if (mtrr_type == MTRR_TYPE_UNCACHABLE)
175 return _PAGE_CACHE_UC;
176 if (mtrr_type == MTRR_TYPE_WRCOMB)
177 return _PAGE_CACHE_WC;
183 static int chk_conflict(struct memtype *new, struct memtype *entry,
186 if (new->type != entry->type) {
188 new->type = entry->type;
194 /* check overlaps with more than one entry in the list */
195 list_for_each_entry_continue(entry, &memtype_list, nd) {
196 if (new->end <= entry->start)
198 else if (new->type != entry->type)
204 printk(KERN_INFO "%s:%d conflicting memory types "
205 "%Lx-%Lx %s<->%s\n", current->comm, current->pid, new->start,
206 new->end, cattr_name(new->type), cattr_name(entry->type));
210 static struct memtype *cached_entry;
211 static u64 cached_start;
214 * req_type typically has one of the:
217 * - _PAGE_CACHE_UC_MINUS
220 * req_type will have a special case value '-1', when requester want to inherit
221 * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
223 * If new_type is NULL, function will return an error if it cannot reserve the
224 * region with req_type. If new_type is non-NULL, function will return
225 * available type in new_type in case of no error. In case of any error
226 * it will return a negative return value.
228 int reserve_memtype(u64 start, u64 end, unsigned long req_type,
229 unsigned long *new_type)
231 struct memtype *new, *entry;
232 unsigned long actual_type;
233 struct list_head *where;
236 BUG_ON(start >= end); /* end is exclusive */
239 /* This is identical to page table setting without PAT */
242 *new_type = _PAGE_CACHE_WB;
244 *new_type = req_type & _PAGE_CACHE_MASK;
249 /* Low ISA region is always mapped WB in page table. No need to track */
250 if (is_ISA_range(start, end - 1)) {
252 *new_type = _PAGE_CACHE_WB;
256 if (req_type == -1) {
258 * Call mtrr_lookup to get the type hint. This is an
259 * optimization for /dev/mem mmap'ers into WB memory (BIOS
260 * tools and ACPI tools). Use WB request for WB memory and use
261 * UC_MINUS otherwise.
263 u8 mtrr_type = mtrr_type_lookup(start, end);
265 if (mtrr_type == MTRR_TYPE_WRBACK)
266 actual_type = _PAGE_CACHE_WB;
268 actual_type = _PAGE_CACHE_UC_MINUS;
270 actual_type = pat_x_mtrr_type(start, end,
271 req_type & _PAGE_CACHE_MASK);
273 new = kmalloc(sizeof(struct memtype), GFP_KERNEL);
279 new->type = actual_type;
282 *new_type = actual_type;
284 spin_lock(&memtype_lock);
286 if (cached_entry && start >= cached_start)
287 entry = cached_entry;
289 entry = list_entry(&memtype_list, struct memtype, nd);
291 /* Search for existing mapping that overlaps the current range */
293 list_for_each_entry_continue(entry, &memtype_list, nd) {
294 if (end <= entry->start) {
295 where = entry->nd.prev;
296 cached_entry = list_entry(where, struct memtype, nd);
298 } else if (start <= entry->start) { /* end > entry->start */
299 err = chk_conflict(new, entry, new_type);
301 dprintk("Overlap at 0x%Lx-0x%Lx\n",
302 entry->start, entry->end);
303 where = entry->nd.prev;
304 cached_entry = list_entry(where,
308 } else if (start < entry->end) { /* start > entry->start */
309 err = chk_conflict(new, entry, new_type);
311 dprintk("Overlap at 0x%Lx-0x%Lx\n",
312 entry->start, entry->end);
313 cached_entry = list_entry(entry->nd.prev,
317 * Move to right position in the linked
318 * list to add this new entry
320 list_for_each_entry_continue(entry,
322 if (start <= entry->start) {
323 where = entry->nd.prev;
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);
341 cached_start = start;
344 list_add(&new->nd, where);
346 list_add_tail(&new->nd, &memtype_list);
348 spin_unlock(&memtype_lock);
350 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
351 start, end, cattr_name(new->type), cattr_name(req_type),
352 new_type ? cattr_name(*new_type) : "-");
357 int free_memtype(u64 start, u64 end)
359 struct memtype *entry;
365 /* Low ISA region is always mapped WB. No need to track */
366 if (is_ISA_range(start, end - 1))
369 spin_lock(&memtype_lock);
370 list_for_each_entry(entry, &memtype_list, nd) {
371 if (entry->start == start && entry->end == end) {
372 if (cached_entry == entry || cached_start == start)
375 list_del(&entry->nd);
381 spin_unlock(&memtype_lock);
384 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
385 current->comm, current->pid, start, end);
388 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
393 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
394 unsigned long size, pgprot_t vma_prot)
399 #ifdef CONFIG_STRICT_DEVMEM
400 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
401 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
406 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
408 u64 from = ((u64)pfn) << PAGE_SHIFT;
409 u64 to = from + size;
412 while (cursor < to) {
413 if (!devmem_is_allowed(pfn)) {
415 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
416 current->comm, from, to);
424 #endif /* CONFIG_STRICT_DEVMEM */
426 int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
427 unsigned long size, pgprot_t *vma_prot)
429 u64 offset = ((u64) pfn) << PAGE_SHIFT;
430 unsigned long flags = -1;
433 if (!range_is_allowed(pfn, size))
436 if (file->f_flags & O_SYNC) {
437 flags = _PAGE_CACHE_UC_MINUS;
442 * On the PPro and successors, the MTRRs are used to set
443 * memory types for physical addresses outside main memory,
444 * so blindly setting UC or PWT on those pages is wrong.
445 * For Pentiums and earlier, the surround logic should disable
446 * caching for the high addresses through the KEN pin, but
447 * we maintain the tradition of paranoia in this code.
450 !(boot_cpu_has(X86_FEATURE_MTRR) ||
451 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
452 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
453 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
454 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
455 flags = _PAGE_CACHE_UC;
460 * With O_SYNC, we can only take UC_MINUS mapping. Fail if we cannot.
462 * Without O_SYNC, we want to get
463 * - WB for WB-able memory and no other conflicting mappings
464 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
465 * - Inherit from confliting mappings otherwise
468 retval = reserve_memtype(offset, offset + size, flags, NULL);
470 retval = reserve_memtype(offset, offset + size, -1, &flags);
476 if (((pfn < max_low_pfn_mapped) ||
477 (pfn >= (1UL<<(32 - PAGE_SHIFT)) && pfn < max_pfn_mapped)) &&
478 ioremap_change_attr((unsigned long)__va(offset), size, flags) < 0) {
479 free_memtype(offset, offset + size);
481 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
482 current->comm, current->pid,
484 offset, (unsigned long long)(offset + size));
488 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
493 void map_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
495 u64 addr = (u64)pfn << PAGE_SHIFT;
497 unsigned long want_flags = (pgprot_val(vma_prot) & _PAGE_CACHE_MASK);
499 reserve_memtype(addr, addr + size, want_flags, &flags);
500 if (flags != want_flags) {
502 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
503 current->comm, current->pid,
504 cattr_name(want_flags),
505 addr, (unsigned long long)(addr + size),
510 void unmap_devmem(unsigned long pfn, unsigned long size, pgprot_t vma_prot)
512 u64 addr = (u64)pfn << PAGE_SHIFT;
514 free_memtype(addr, addr + size);
517 #if defined(CONFIG_DEBUG_FS)
519 /* get Nth element of the linked list */
520 static struct memtype *memtype_get_idx(loff_t pos)
522 struct memtype *list_node, *print_entry;
525 print_entry = kmalloc(sizeof(struct memtype), GFP_KERNEL);
529 spin_lock(&memtype_lock);
530 list_for_each_entry(list_node, &memtype_list, nd) {
532 *print_entry = *list_node;
533 spin_unlock(&memtype_lock);
538 spin_unlock(&memtype_lock);
543 static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
547 seq_printf(seq, "PAT memtype list:\n");
550 return memtype_get_idx(*pos);
553 static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
556 return memtype_get_idx(*pos);
559 static void memtype_seq_stop(struct seq_file *seq, void *v)
563 static int memtype_seq_show(struct seq_file *seq, void *v)
565 struct memtype *print_entry = (struct memtype *)v;
567 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
568 print_entry->start, print_entry->end);
573 static struct seq_operations memtype_seq_ops = {
574 .start = memtype_seq_start,
575 .next = memtype_seq_next,
576 .stop = memtype_seq_stop,
577 .show = memtype_seq_show,
580 static int memtype_seq_open(struct inode *inode, struct file *file)
582 return seq_open(file, &memtype_seq_ops);
585 static const struct file_operations memtype_fops = {
586 .open = memtype_seq_open,
589 .release = seq_release,
592 static int __init pat_memtype_list_init(void)
594 debugfs_create_file("pat_memtype_list", S_IRUSR, arch_debugfs_dir,
595 NULL, &memtype_fops);
599 late_initcall(pat_memtype_list_init);
601 #endif /* CONFIG_DEBUG_FS */