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
));
211 * req_type typically has one of the:
214 * - _PAGE_CACHE_UC_MINUS
217 * req_type will have a special case value '-1', when requester want to inherit
218 * the memory type from mtrr (if WB), existing PAT, defaulting to UC_MINUS.
220 * If new_type is NULL, function will return an error if it cannot reserve the
221 * region with req_type. If new_type is non-NULL, function will return
222 * available type in new_type in case of no error. In case of any error
223 * it will return a negative return value.
225 int reserve_memtype(u64 start
, u64 end
, unsigned long req_type
,
226 unsigned long *new_type
)
228 struct memtype
*new, *entry
;
229 unsigned long actual_type
;
230 struct list_head
*where
;
233 BUG_ON(start
>= end
); /* end is exclusive */
236 /* This is identical to page table setting without PAT */
239 *new_type
= _PAGE_CACHE_WB
;
241 *new_type
= req_type
& _PAGE_CACHE_MASK
;
246 /* Low ISA region is always mapped WB in page table. No need to track */
247 if (is_ISA_range(start
, end
- 1)) {
249 *new_type
= _PAGE_CACHE_WB
;
253 if (req_type
== -1) {
255 * Call mtrr_lookup to get the type hint. This is an
256 * optimization for /dev/mem mmap'ers into WB memory (BIOS
257 * tools and ACPI tools). Use WB request for WB memory and use
258 * UC_MINUS otherwise.
260 u8 mtrr_type
= mtrr_type_lookup(start
, end
);
262 if (mtrr_type
== MTRR_TYPE_WRBACK
)
263 actual_type
= _PAGE_CACHE_WB
;
265 actual_type
= _PAGE_CACHE_UC_MINUS
;
267 actual_type
= pat_x_mtrr_type(start
, end
,
268 req_type
& _PAGE_CACHE_MASK
);
270 new = kmalloc(sizeof(struct memtype
), GFP_KERNEL
);
276 new->type
= actual_type
;
279 *new_type
= actual_type
;
281 spin_lock(&memtype_lock
);
283 /* Search for existing mapping that overlaps the current range */
285 list_for_each_entry(entry
, &memtype_list
, nd
) {
286 if (end
<= entry
->start
) {
287 where
= entry
->nd
.prev
;
289 } else if (start
<= entry
->start
) { /* end > entry->start */
290 err
= chk_conflict(new, entry
, new_type
);
292 dprintk("Overlap at 0x%Lx-0x%Lx\n",
293 entry
->start
, entry
->end
);
294 where
= entry
->nd
.prev
;
297 } else if (start
< entry
->end
) { /* start > entry->start */
298 err
= chk_conflict(new, entry
, new_type
);
300 dprintk("Overlap at 0x%Lx-0x%Lx\n",
301 entry
->start
, entry
->end
);
309 printk(KERN_INFO
"reserve_memtype failed 0x%Lx-0x%Lx, "
310 "track %s, req %s\n",
311 start
, end
, cattr_name(new->type
), cattr_name(req_type
));
313 spin_unlock(&memtype_lock
);
318 list_add(&new->nd
, where
);
320 list_add_tail(&new->nd
, &memtype_list
);
322 spin_unlock(&memtype_lock
);
324 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
325 start
, end
, cattr_name(new->type
), cattr_name(req_type
),
326 new_type
? cattr_name(*new_type
) : "-");
331 int free_memtype(u64 start
, u64 end
)
333 struct memtype
*entry
;
339 /* Low ISA region is always mapped WB. No need to track */
340 if (is_ISA_range(start
, end
- 1))
343 spin_lock(&memtype_lock
);
344 list_for_each_entry(entry
, &memtype_list
, nd
) {
345 if (entry
->start
== start
&& entry
->end
== end
) {
346 list_del(&entry
->nd
);
352 spin_unlock(&memtype_lock
);
355 printk(KERN_INFO
"%s:%d freeing invalid memtype %Lx-%Lx\n",
356 current
->comm
, current
->pid
, start
, end
);
359 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start
, end
);
365 * /dev/mem mmap interface. The memtype used for mapping varies:
366 * - Use UC for mappings with O_SYNC flag
367 * - Without O_SYNC flag, if there is any conflict in reserve_memtype,
368 * inherit the memtype from existing mapping.
369 * - Else use UC_MINUS memtype (for backward compatibility with existing
372 pgprot_t
phys_mem_access_prot(struct file
*file
, unsigned long pfn
,
373 unsigned long size
, pgprot_t vma_prot
)
378 #ifdef CONFIG_STRICT_DEVMEM
379 /* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
380 static inline int range_is_allowed(unsigned long pfn
, unsigned long size
)
385 static inline int range_is_allowed(unsigned long pfn
, unsigned long size
)
387 u64 from
= ((u64
)pfn
) << PAGE_SHIFT
;
388 u64 to
= from
+ size
;
391 while (cursor
< to
) {
392 if (!devmem_is_allowed(pfn
)) {
394 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
395 current
->comm
, from
, to
);
403 #endif /* CONFIG_STRICT_DEVMEM */
405 int phys_mem_access_prot_allowed(struct file
*file
, unsigned long pfn
,
406 unsigned long size
, pgprot_t
*vma_prot
)
408 u64 offset
= ((u64
) pfn
) << PAGE_SHIFT
;
409 unsigned long flags
= _PAGE_CACHE_UC_MINUS
;
412 if (!range_is_allowed(pfn
, size
))
415 if (file
->f_flags
& O_SYNC
) {
416 flags
= _PAGE_CACHE_UC
;
421 * On the PPro and successors, the MTRRs are used to set
422 * memory types for physical addresses outside main memory,
423 * so blindly setting UC or PWT on those pages is wrong.
424 * For Pentiums and earlier, the surround logic should disable
425 * caching for the high addresses through the KEN pin, but
426 * we maintain the tradition of paranoia in this code.
429 !(boot_cpu_has(X86_FEATURE_MTRR
) ||
430 boot_cpu_has(X86_FEATURE_K6_MTRR
) ||
431 boot_cpu_has(X86_FEATURE_CYRIX_ARR
) ||
432 boot_cpu_has(X86_FEATURE_CENTAUR_MCR
)) &&
433 (pfn
<< PAGE_SHIFT
) >= __pa(high_memory
)) {
434 flags
= _PAGE_CACHE_UC
;
439 * With O_SYNC, we can only take UC mapping. Fail if we cannot.
440 * Without O_SYNC, we want to get
441 * - WB for WB-able memory and no other conflicting mappings
442 * - UC_MINUS for non-WB-able memory with no other conflicting mappings
443 * - Inherit from confliting mappings otherwise
445 if (flags
!= _PAGE_CACHE_UC_MINUS
) {
446 retval
= reserve_memtype(offset
, offset
+ size
, flags
, NULL
);
448 retval
= reserve_memtype(offset
, offset
+ size
, -1, &flags
);
454 if (((pfn
< max_low_pfn_mapped
) ||
455 (pfn
>= (1UL<<(32 - PAGE_SHIFT
)) && pfn
< max_pfn_mapped
)) &&
456 ioremap_change_attr((unsigned long)__va(offset
), size
, flags
) < 0) {
457 free_memtype(offset
, offset
+ size
);
459 "%s:%d /dev/mem ioremap_change_attr failed %s for %Lx-%Lx\n",
460 current
->comm
, current
->pid
,
462 offset
, (unsigned long long)(offset
+ size
));
466 *vma_prot
= __pgprot((pgprot_val(*vma_prot
) & ~_PAGE_CACHE_MASK
) |
471 void map_devmem(unsigned long pfn
, unsigned long size
, pgprot_t vma_prot
)
473 u64 addr
= (u64
)pfn
<< PAGE_SHIFT
;
475 unsigned long want_flags
= (pgprot_val(vma_prot
) & _PAGE_CACHE_MASK
);
477 reserve_memtype(addr
, addr
+ size
, want_flags
, &flags
);
478 if (flags
!= want_flags
) {
480 "%s:%d /dev/mem expected mapping type %s for %Lx-%Lx, got %s\n",
481 current
->comm
, current
->pid
,
482 cattr_name(want_flags
),
483 addr
, (unsigned long long)(addr
+ size
),
488 void unmap_devmem(unsigned long pfn
, unsigned long size
, pgprot_t vma_prot
)
490 u64 addr
= (u64
)pfn
<< PAGE_SHIFT
;
492 free_memtype(addr
, addr
+ size
);
495 #if defined(CONFIG_DEBUG_FS)
497 /* get Nth element of the linked list */
498 static struct memtype
*memtype_get_idx(loff_t pos
)
500 struct memtype
*list_node
, *print_entry
;
503 print_entry
= kmalloc(sizeof(struct memtype
), GFP_KERNEL
);
507 spin_lock(&memtype_lock
);
508 list_for_each_entry(list_node
, &memtype_list
, nd
) {
510 *print_entry
= *list_node
;
511 spin_unlock(&memtype_lock
);
516 spin_unlock(&memtype_lock
);
521 static void *memtype_seq_start(struct seq_file
*seq
, loff_t
*pos
)
525 seq_printf(seq
, "PAT memtype list:\n");
528 return memtype_get_idx(*pos
);
531 static void *memtype_seq_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
534 return memtype_get_idx(*pos
);
537 static void memtype_seq_stop(struct seq_file
*seq
, void *v
)
541 static int memtype_seq_show(struct seq_file
*seq
, void *v
)
543 struct memtype
*print_entry
= (struct memtype
*)v
;
545 seq_printf(seq
, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry
->type
),
546 print_entry
->start
, print_entry
->end
);
551 static struct seq_operations memtype_seq_ops
= {
552 .start
= memtype_seq_start
,
553 .next
= memtype_seq_next
,
554 .stop
= memtype_seq_stop
,
555 .show
= memtype_seq_show
,
558 static int memtype_seq_open(struct inode
*inode
, struct file
*file
)
560 return seq_open(file
, &memtype_seq_ops
);
563 static const struct file_operations memtype_fops
= {
564 .open
= memtype_seq_open
,
567 .release
= seq_release
,
570 static int __init
pat_memtype_list_init(void)
572 debugfs_create_file("pat_memtype_list", S_IRUSR
, arch_debugfs_dir
,
573 NULL
, &memtype_fops
);
577 late_initcall(pat_memtype_list_init
);
579 #endif /* CONFIG_DEBUG_FS */