Loading...
1/*
2 * Handle caching attributes in page tables (PAT)
3 *
4 * Authors: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
5 * Suresh B Siddha <suresh.b.siddha@intel.com>
6 *
7 * Loosely based on earlier PAT patchset from Eric Biederman and Andi Kleen.
8 */
9
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>
16#include <linux/mm.h>
17#include <linux/fs.h>
18#include <linux/rbtree.h>
19
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>
26#include <asm/e820.h>
27#include <asm/mtrr.h>
28#include <asm/page.h>
29#include <asm/msr.h>
30#include <asm/pat.h>
31#include <asm/io.h>
32
33#include "pat_internal.h"
34
35#ifdef CONFIG_X86_PAT
36int __read_mostly pat_enabled = 1;
37
38static inline void pat_disable(const char *reason)
39{
40 pat_enabled = 0;
41 printk(KERN_INFO "%s\n", reason);
42}
43
44static int __init nopat(char *str)
45{
46 pat_disable("PAT support disabled.");
47 return 0;
48}
49early_param("nopat", nopat);
50#else
51static inline void pat_disable(const char *reason)
52{
53 (void)reason;
54}
55#endif
56
57
58int pat_debug_enable;
59
60static int __init pat_debug_setup(char *str)
61{
62 pat_debug_enable = 1;
63 return 0;
64}
65__setup("debugpat", pat_debug_setup);
66
67static u64 __read_mostly boot_pat_state;
68
69enum {
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 */
76};
77
78#define PAT(x, y) ((u64)PAT_ ## y << ((x)*8))
79
80void pat_init(void)
81{
82 u64 pat;
83 bool boot_cpu = !boot_pat_state;
84
85 if (!pat_enabled)
86 return;
87
88 if (!cpu_has_pat) {
89 if (!boot_pat_state) {
90 pat_disable("PAT not supported by CPU.");
91 return;
92 } else {
93 /*
94 * If this happens we are on a secondary CPU, but
95 * switched to PAT on the boot CPU. We have no way to
96 * undo PAT.
97 */
98 printk(KERN_ERR "PAT enabled, "
99 "but not supported by secondary CPU\n");
100 BUG();
101 }
102 }
103
104 /* Set PWT to Write-Combining. All other bits stay the same */
105 /*
106 * PTE encoding used in Linux:
107 * PAT
108 * |PCD
109 * ||PWT
110 * |||
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
115 * PAT bit unused
116 */
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);
119
120 /* Boot CPU check */
121 if (!boot_pat_state)
122 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
123
124 wrmsrl(MSR_IA32_CR_PAT, pat);
125
126 if (boot_cpu)
127 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
128 smp_processor_id(), boot_pat_state, pat);
129}
130
131#undef PAT
132
133static DEFINE_SPINLOCK(memtype_lock); /* protects memtype accesses */
134
135/*
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
140 * SDM vol 3a
141 */
142static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
143{
144 /*
145 * Look for MTRR hint to get the effective type in case where PAT
146 * request is for WB.
147 */
148 if (req_type == _PAGE_CACHE_WB) {
149 u8 mtrr_type;
150
151 mtrr_type = mtrr_type_lookup(start, end);
152 if (mtrr_type != MTRR_TYPE_WRBACK)
153 return _PAGE_CACHE_UC_MINUS;
154
155 return _PAGE_CACHE_WB;
156 }
157
158 return req_type;
159}
160
161static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
162{
163 int ram_page = 0, not_rampage = 0;
164 unsigned long page_nr;
165
166 for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
167 ++page_nr) {
168 /*
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/..)
174 */
175 if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
176 page_is_ram(page_nr))
177 ram_page = 1;
178 else
179 not_rampage = 1;
180
181 if (ram_page == not_rampage)
182 return -1;
183 }
184
185 return ram_page;
186}
187
188/*
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
193 */
194static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
195 unsigned long *new_type)
196{
197 struct page *page;
198 u64 pfn;
199
200 if (req_type == _PAGE_CACHE_UC) {
201 /* We do not support strong UC */
202 WARN_ON_ONCE(1);
203 req_type = _PAGE_CACHE_UC_MINUS;
204 }
205
206 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
207 unsigned long type;
208
209 page = pfn_to_page(pfn);
210 type = get_page_memtype(page);
211 if (type != -1) {
212 printk(KERN_INFO "reserve_ram_pages_type failed "
213 "0x%Lx-0x%Lx, track 0x%lx, req 0x%lx\n",
214 start, end, type, req_type);
215 if (new_type)
216 *new_type = type;
217
218 return -EBUSY;
219 }
220 }
221
222 if (new_type)
223 *new_type = req_type;
224
225 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
226 page = pfn_to_page(pfn);
227 set_page_memtype(page, req_type);
228 }
229 return 0;
230}
231
232static int free_ram_pages_type(u64 start, u64 end)
233{
234 struct page *page;
235 u64 pfn;
236
237 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
238 page = pfn_to_page(pfn);
239 set_page_memtype(page, -1);
240 }
241 return 0;
242}
243
244/*
245 * req_type typically has one of the:
246 * - _PAGE_CACHE_WB
247 * - _PAGE_CACHE_WC
248 * - _PAGE_CACHE_UC_MINUS
249 * - _PAGE_CACHE_UC
250 *
251 * If new_type is NULL, function will return an error if it cannot reserve the
252 * region with req_type. If new_type is non-NULL, function will return
253 * available type in new_type in case of no error. In case of any error
254 * it will return a negative return value.
255 */
256int reserve_memtype(u64 start, u64 end, unsigned long req_type,
257 unsigned long *new_type)
258{
259 struct memtype *new;
260 unsigned long actual_type;
261 int is_range_ram;
262 int err = 0;
263
264 BUG_ON(start >= end); /* end is exclusive */
265
266 if (!pat_enabled) {
267 /* This is identical to page table setting without PAT */
268 if (new_type) {
269 if (req_type == _PAGE_CACHE_WC)
270 *new_type = _PAGE_CACHE_UC_MINUS;
271 else
272 *new_type = req_type & _PAGE_CACHE_MASK;
273 }
274 return 0;
275 }
276
277 /* Low ISA region is always mapped WB in page table. No need to track */
278 if (x86_platform.is_untracked_pat_range(start, end)) {
279 if (new_type)
280 *new_type = _PAGE_CACHE_WB;
281 return 0;
282 }
283
284 /*
285 * Call mtrr_lookup to get the type hint. This is an
286 * optimization for /dev/mem mmap'ers into WB memory (BIOS
287 * tools and ACPI tools). Use WB request for WB memory and use
288 * UC_MINUS otherwise.
289 */
290 actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK);
291
292 if (new_type)
293 *new_type = actual_type;
294
295 is_range_ram = pat_pagerange_is_ram(start, end);
296 if (is_range_ram == 1) {
297
298 err = reserve_ram_pages_type(start, end, req_type, new_type);
299
300 return err;
301 } else if (is_range_ram < 0) {
302 return -EINVAL;
303 }
304
305 new = kzalloc(sizeof(struct memtype), GFP_KERNEL);
306 if (!new)
307 return -ENOMEM;
308
309 new->start = start;
310 new->end = end;
311 new->type = actual_type;
312
313 spin_lock(&memtype_lock);
314
315 err = rbt_memtype_check_insert(new, new_type);
316 if (err) {
317 printk(KERN_INFO "reserve_memtype failed 0x%Lx-0x%Lx, "
318 "track %s, req %s\n",
319 start, end, cattr_name(new->type), cattr_name(req_type));
320 kfree(new);
321 spin_unlock(&memtype_lock);
322
323 return err;
324 }
325
326 spin_unlock(&memtype_lock);
327
328 dprintk("reserve_memtype added 0x%Lx-0x%Lx, track %s, req %s, ret %s\n",
329 start, end, cattr_name(new->type), cattr_name(req_type),
330 new_type ? cattr_name(*new_type) : "-");
331
332 return err;
333}
334
335int free_memtype(u64 start, u64 end)
336{
337 int err = -EINVAL;
338 int is_range_ram;
339 struct memtype *entry;
340
341 if (!pat_enabled)
342 return 0;
343
344 /* Low ISA region is always mapped WB. No need to track */
345 if (x86_platform.is_untracked_pat_range(start, end))
346 return 0;
347
348 is_range_ram = pat_pagerange_is_ram(start, end);
349 if (is_range_ram == 1) {
350
351 err = free_ram_pages_type(start, end);
352
353 return err;
354 } else if (is_range_ram < 0) {
355 return -EINVAL;
356 }
357
358 spin_lock(&memtype_lock);
359 entry = rbt_memtype_erase(start, end);
360 spin_unlock(&memtype_lock);
361
362 if (!entry) {
363 printk(KERN_INFO "%s:%d freeing invalid memtype %Lx-%Lx\n",
364 current->comm, current->pid, start, end);
365 return -EINVAL;
366 }
367
368 kfree(entry);
369
370 dprintk("free_memtype request 0x%Lx-0x%Lx\n", start, end);
371
372 return 0;
373}
374
375
376/**
377 * lookup_memtype - Looksup the memory type for a physical address
378 * @paddr: physical address of which memory type needs to be looked up
379 *
380 * Only to be called when PAT is enabled
381 *
382 * Returns _PAGE_CACHE_WB, _PAGE_CACHE_WC, _PAGE_CACHE_UC_MINUS or
383 * _PAGE_CACHE_UC
384 */
385static unsigned long lookup_memtype(u64 paddr)
386{
387 int rettype = _PAGE_CACHE_WB;
388 struct memtype *entry;
389
390 if (x86_platform.is_untracked_pat_range(paddr, paddr + PAGE_SIZE))
391 return rettype;
392
393 if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) {
394 struct page *page;
395 page = pfn_to_page(paddr >> PAGE_SHIFT);
396 rettype = get_page_memtype(page);
397 /*
398 * -1 from get_page_memtype() implies RAM page is in its
399 * default state and not reserved, and hence of type WB
400 */
401 if (rettype == -1)
402 rettype = _PAGE_CACHE_WB;
403
404 return rettype;
405 }
406
407 spin_lock(&memtype_lock);
408
409 entry = rbt_memtype_lookup(paddr);
410 if (entry != NULL)
411 rettype = entry->type;
412 else
413 rettype = _PAGE_CACHE_UC_MINUS;
414
415 spin_unlock(&memtype_lock);
416 return rettype;
417}
418
419/**
420 * io_reserve_memtype - Request a memory type mapping for a region of memory
421 * @start: start (physical address) of the region
422 * @end: end (physical address) of the region
423 * @type: A pointer to memtype, with requested type. On success, requested
424 * or any other compatible type that was available for the region is returned
425 *
426 * On success, returns 0
427 * On failure, returns non-zero
428 */
429int io_reserve_memtype(resource_size_t start, resource_size_t end,
430 unsigned long *type)
431{
432 resource_size_t size = end - start;
433 unsigned long req_type = *type;
434 unsigned long new_type;
435 int ret;
436
437 WARN_ON_ONCE(iomem_map_sanity_check(start, size));
438
439 ret = reserve_memtype(start, end, req_type, &new_type);
440 if (ret)
441 goto out_err;
442
443 if (!is_new_memtype_allowed(start, size, req_type, new_type))
444 goto out_free;
445
446 if (kernel_map_sync_memtype(start, size, new_type) < 0)
447 goto out_free;
448
449 *type = new_type;
450 return 0;
451
452out_free:
453 free_memtype(start, end);
454 ret = -EBUSY;
455out_err:
456 return ret;
457}
458
459/**
460 * io_free_memtype - Release a memory type mapping for a region of memory
461 * @start: start (physical address) of the region
462 * @end: end (physical address) of the region
463 */
464void io_free_memtype(resource_size_t start, resource_size_t end)
465{
466 free_memtype(start, end);
467}
468
469pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
470 unsigned long size, pgprot_t vma_prot)
471{
472 return vma_prot;
473}
474
475#ifdef CONFIG_STRICT_DEVMEM
476/* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
477static inline int range_is_allowed(unsigned long pfn, unsigned long size)
478{
479 return 1;
480}
481#else
482/* This check is needed to avoid cache aliasing when PAT is enabled */
483static inline int range_is_allowed(unsigned long pfn, unsigned long size)
484{
485 u64 from = ((u64)pfn) << PAGE_SHIFT;
486 u64 to = from + size;
487 u64 cursor = from;
488
489 if (!pat_enabled)
490 return 1;
491
492 while (cursor < to) {
493 if (!devmem_is_allowed(pfn)) {
494 printk(KERN_INFO
495 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
496 current->comm, from, to);
497 return 0;
498 }
499 cursor += PAGE_SIZE;
500 pfn++;
501 }
502 return 1;
503}
504#endif /* CONFIG_STRICT_DEVMEM */
505
506int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
507 unsigned long size, pgprot_t *vma_prot)
508{
509 unsigned long flags = _PAGE_CACHE_WB;
510
511 if (!range_is_allowed(pfn, size))
512 return 0;
513
514 if (file->f_flags & O_DSYNC)
515 flags = _PAGE_CACHE_UC_MINUS;
516
517#ifdef CONFIG_X86_32
518 /*
519 * On the PPro and successors, the MTRRs are used to set
520 * memory types for physical addresses outside main memory,
521 * so blindly setting UC or PWT on those pages is wrong.
522 * For Pentiums and earlier, the surround logic should disable
523 * caching for the high addresses through the KEN pin, but
524 * we maintain the tradition of paranoia in this code.
525 */
526 if (!pat_enabled &&
527 !(boot_cpu_has(X86_FEATURE_MTRR) ||
528 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
529 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
530 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
531 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
532 flags = _PAGE_CACHE_UC;
533 }
534#endif
535
536 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
537 flags);
538 return 1;
539}
540
541/*
542 * Change the memory type for the physial address range in kernel identity
543 * mapping space if that range is a part of identity map.
544 */
545int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
546{
547 unsigned long id_sz;
548
549 if (base >= __pa(high_memory))
550 return 0;
551
552 id_sz = (__pa(high_memory) < base + size) ?
553 __pa(high_memory) - base :
554 size;
555
556 if (ioremap_change_attr((unsigned long)__va(base), id_sz, flags) < 0) {
557 printk(KERN_INFO
558 "%s:%d ioremap_change_attr failed %s "
559 "for %Lx-%Lx\n",
560 current->comm, current->pid,
561 cattr_name(flags),
562 base, (unsigned long long)(base + size));
563 return -EINVAL;
564 }
565 return 0;
566}
567
568/*
569 * Internal interface to reserve a range of physical memory with prot.
570 * Reserved non RAM regions only and after successful reserve_memtype,
571 * this func also keeps identity mapping (if any) in sync with this new prot.
572 */
573static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot,
574 int strict_prot)
575{
576 int is_ram = 0;
577 int ret;
578 unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK);
579 unsigned long flags = want_flags;
580
581 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
582
583 /*
584 * reserve_pfn_range() for RAM pages. We do not refcount to keep
585 * track of number of mappings of RAM pages. We can assert that
586 * the type requested matches the type of first page in the range.
587 */
588 if (is_ram) {
589 if (!pat_enabled)
590 return 0;
591
592 flags = lookup_memtype(paddr);
593 if (want_flags != flags) {
594 printk(KERN_WARNING
595 "%s:%d map pfn RAM range req %s for %Lx-%Lx, got %s\n",
596 current->comm, current->pid,
597 cattr_name(want_flags),
598 (unsigned long long)paddr,
599 (unsigned long long)(paddr + size),
600 cattr_name(flags));
601 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
602 (~_PAGE_CACHE_MASK)) |
603 flags);
604 }
605 return 0;
606 }
607
608 ret = reserve_memtype(paddr, paddr + size, want_flags, &flags);
609 if (ret)
610 return ret;
611
612 if (flags != want_flags) {
613 if (strict_prot ||
614 !is_new_memtype_allowed(paddr, size, want_flags, flags)) {
615 free_memtype(paddr, paddr + size);
616 printk(KERN_ERR "%s:%d map pfn expected mapping type %s"
617 " for %Lx-%Lx, got %s\n",
618 current->comm, current->pid,
619 cattr_name(want_flags),
620 (unsigned long long)paddr,
621 (unsigned long long)(paddr + size),
622 cattr_name(flags));
623 return -EINVAL;
624 }
625 /*
626 * We allow returning different type than the one requested in
627 * non strict case.
628 */
629 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
630 (~_PAGE_CACHE_MASK)) |
631 flags);
632 }
633
634 if (kernel_map_sync_memtype(paddr, size, flags) < 0) {
635 free_memtype(paddr, paddr + size);
636 return -EINVAL;
637 }
638 return 0;
639}
640
641/*
642 * Internal interface to free a range of physical memory.
643 * Frees non RAM regions only.
644 */
645static void free_pfn_range(u64 paddr, unsigned long size)
646{
647 int is_ram;
648
649 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
650 if (is_ram == 0)
651 free_memtype(paddr, paddr + size);
652}
653
654/*
655 * track_pfn_vma_copy is called when vma that is covering the pfnmap gets
656 * copied through copy_page_range().
657 *
658 * If the vma has a linear pfn mapping for the entire range, we get the prot
659 * from pte and reserve the entire vma range with single reserve_pfn_range call.
660 */
661int track_pfn_vma_copy(struct vm_area_struct *vma)
662{
663 resource_size_t paddr;
664 unsigned long prot;
665 unsigned long vma_size = vma->vm_end - vma->vm_start;
666 pgprot_t pgprot;
667
668 if (is_linear_pfn_mapping(vma)) {
669 /*
670 * reserve the whole chunk covered by vma. We need the
671 * starting address and protection from pte.
672 */
673 if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) {
674 WARN_ON_ONCE(1);
675 return -EINVAL;
676 }
677 pgprot = __pgprot(prot);
678 return reserve_pfn_range(paddr, vma_size, &pgprot, 1);
679 }
680
681 return 0;
682}
683
684/*
685 * track_pfn_vma_new is called when a _new_ pfn mapping is being established
686 * for physical range indicated by pfn and size.
687 *
688 * prot is passed in as a parameter for the new mapping. If the vma has a
689 * linear pfn mapping for the entire range reserve the entire vma range with
690 * single reserve_pfn_range call.
691 */
692int track_pfn_vma_new(struct vm_area_struct *vma, pgprot_t *prot,
693 unsigned long pfn, unsigned long size)
694{
695 unsigned long flags;
696 resource_size_t paddr;
697 unsigned long vma_size = vma->vm_end - vma->vm_start;
698
699 if (is_linear_pfn_mapping(vma)) {
700 /* reserve the whole chunk starting from vm_pgoff */
701 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
702 return reserve_pfn_range(paddr, vma_size, prot, 0);
703 }
704
705 if (!pat_enabled)
706 return 0;
707
708 /* for vm_insert_pfn and friends, we set prot based on lookup */
709 flags = lookup_memtype(pfn << PAGE_SHIFT);
710 *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) |
711 flags);
712
713 return 0;
714}
715
716/*
717 * untrack_pfn_vma is called while unmapping a pfnmap for a region.
718 * untrack can be called for a specific region indicated by pfn and size or
719 * can be for the entire vma (in which case size can be zero).
720 */
721void untrack_pfn_vma(struct vm_area_struct *vma, unsigned long pfn,
722 unsigned long size)
723{
724 resource_size_t paddr;
725 unsigned long vma_size = vma->vm_end - vma->vm_start;
726
727 if (is_linear_pfn_mapping(vma)) {
728 /* free the whole chunk starting from vm_pgoff */
729 paddr = (resource_size_t)vma->vm_pgoff << PAGE_SHIFT;
730 free_pfn_range(paddr, vma_size);
731 return;
732 }
733}
734
735pgprot_t pgprot_writecombine(pgprot_t prot)
736{
737 if (pat_enabled)
738 return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC);
739 else
740 return pgprot_noncached(prot);
741}
742EXPORT_SYMBOL_GPL(pgprot_writecombine);
743
744#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
745
746static struct memtype *memtype_get_idx(loff_t pos)
747{
748 struct memtype *print_entry;
749 int ret;
750
751 print_entry = kzalloc(sizeof(struct memtype), GFP_KERNEL);
752 if (!print_entry)
753 return NULL;
754
755 spin_lock(&memtype_lock);
756 ret = rbt_memtype_copy_nth_element(print_entry, pos);
757 spin_unlock(&memtype_lock);
758
759 if (!ret) {
760 return print_entry;
761 } else {
762 kfree(print_entry);
763 return NULL;
764 }
765}
766
767static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
768{
769 if (*pos == 0) {
770 ++*pos;
771 seq_printf(seq, "PAT memtype list:\n");
772 }
773
774 return memtype_get_idx(*pos);
775}
776
777static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
778{
779 ++*pos;
780 return memtype_get_idx(*pos);
781}
782
783static void memtype_seq_stop(struct seq_file *seq, void *v)
784{
785}
786
787static int memtype_seq_show(struct seq_file *seq, void *v)
788{
789 struct memtype *print_entry = (struct memtype *)v;
790
791 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
792 print_entry->start, print_entry->end);
793 kfree(print_entry);
794
795 return 0;
796}
797
798static const struct seq_operations memtype_seq_ops = {
799 .start = memtype_seq_start,
800 .next = memtype_seq_next,
801 .stop = memtype_seq_stop,
802 .show = memtype_seq_show,
803};
804
805static int memtype_seq_open(struct inode *inode, struct file *file)
806{
807 return seq_open(file, &memtype_seq_ops);
808}
809
810static const struct file_operations memtype_fops = {
811 .open = memtype_seq_open,
812 .read = seq_read,
813 .llseek = seq_lseek,
814 .release = seq_release,
815};
816
817static int __init pat_memtype_list_init(void)
818{
819 if (pat_enabled) {
820 debugfs_create_file("pat_memtype_list", S_IRUSR,
821 arch_debugfs_dir, NULL, &memtype_fops);
822 }
823 return 0;
824}
825
826late_initcall(pat_memtype_list_init);
827
828#endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */
1/*
2 * Handle caching attributes in page tables (PAT)
3 *
4 * Authors: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
5 * Suresh B Siddha <suresh.b.siddha@intel.com>
6 *
7 * Loosely based on earlier PAT patchset from Eric Biederman and Andi Kleen.
8 */
9
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>
16#include <linux/mm.h>
17#include <linux/fs.h>
18#include <linux/rbtree.h>
19
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>
26#include <asm/e820.h>
27#include <asm/mtrr.h>
28#include <asm/page.h>
29#include <asm/msr.h>
30#include <asm/pat.h>
31#include <asm/io.h>
32
33#include "pat_internal.h"
34
35#ifdef CONFIG_X86_PAT
36int __read_mostly pat_enabled = 1;
37
38static inline void pat_disable(const char *reason)
39{
40 pat_enabled = 0;
41 printk(KERN_INFO "%s\n", reason);
42}
43
44static int __init nopat(char *str)
45{
46 pat_disable("PAT support disabled.");
47 return 0;
48}
49early_param("nopat", nopat);
50#else
51static inline void pat_disable(const char *reason)
52{
53 (void)reason;
54}
55#endif
56
57
58int pat_debug_enable;
59
60static int __init pat_debug_setup(char *str)
61{
62 pat_debug_enable = 1;
63 return 0;
64}
65__setup("debugpat", pat_debug_setup);
66
67static u64 __read_mostly boot_pat_state;
68
69enum {
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 */
76};
77
78#define PAT(x, y) ((u64)PAT_ ## y << ((x)*8))
79
80void pat_init(void)
81{
82 u64 pat;
83 bool boot_cpu = !boot_pat_state;
84
85 if (!pat_enabled)
86 return;
87
88 if (!cpu_has_pat) {
89 if (!boot_pat_state) {
90 pat_disable("PAT not supported by CPU.");
91 return;
92 } else {
93 /*
94 * If this happens we are on a secondary CPU, but
95 * switched to PAT on the boot CPU. We have no way to
96 * undo PAT.
97 */
98 printk(KERN_ERR "PAT enabled, "
99 "but not supported by secondary CPU\n");
100 BUG();
101 }
102 }
103
104 /* Set PWT to Write-Combining. All other bits stay the same */
105 /*
106 * PTE encoding used in Linux:
107 * PAT
108 * |PCD
109 * ||PWT
110 * |||
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
115 * PAT bit unused
116 */
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);
119
120 /* Boot CPU check */
121 if (!boot_pat_state)
122 rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
123
124 wrmsrl(MSR_IA32_CR_PAT, pat);
125
126 if (boot_cpu)
127 printk(KERN_INFO "x86 PAT enabled: cpu %d, old 0x%Lx, new 0x%Lx\n",
128 smp_processor_id(), boot_pat_state, pat);
129}
130
131#undef PAT
132
133static DEFINE_SPINLOCK(memtype_lock); /* protects memtype accesses */
134
135/*
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
140 * SDM vol 3a
141 */
142static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
143{
144 /*
145 * Look for MTRR hint to get the effective type in case where PAT
146 * request is for WB.
147 */
148 if (req_type == _PAGE_CACHE_WB) {
149 u8 mtrr_type;
150
151 mtrr_type = mtrr_type_lookup(start, end);
152 if (mtrr_type != MTRR_TYPE_WRBACK)
153 return _PAGE_CACHE_UC_MINUS;
154
155 return _PAGE_CACHE_WB;
156 }
157
158 return req_type;
159}
160
161struct pagerange_state {
162 unsigned long cur_pfn;
163 int ram;
164 int not_ram;
165};
166
167static int
168pagerange_is_ram_callback(unsigned long initial_pfn, unsigned long total_nr_pages, void *arg)
169{
170 struct pagerange_state *state = arg;
171
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;
175
176 return state->ram && state->not_ram;
177}
178
179static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
180{
181 int ret = 0;
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};
185
186 /*
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/..)
192 */
193 if (start_pfn < ISA_END_ADDRESS >> PAGE_SHIFT)
194 start_pfn = ISA_END_ADDRESS >> PAGE_SHIFT;
195
196 if (start_pfn < end_pfn) {
197 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn,
198 &state, pagerange_is_ram_callback);
199 }
200
201 return (ret > 0) ? -1 : (state.ram ? 1 : 0);
202}
203
204/*
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
209 */
210static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
211 unsigned long *new_type)
212{
213 struct page *page;
214 u64 pfn;
215
216 if (req_type == _PAGE_CACHE_UC) {
217 /* We do not support strong UC */
218 WARN_ON_ONCE(1);
219 req_type = _PAGE_CACHE_UC_MINUS;
220 }
221
222 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
223 unsigned long type;
224
225 page = pfn_to_page(pfn);
226 type = get_page_memtype(page);
227 if (type != -1) {
228 printk(KERN_INFO "reserve_ram_pages_type failed [mem %#010Lx-%#010Lx], track 0x%lx, req 0x%lx\n",
229 start, end - 1, type, req_type);
230 if (new_type)
231 *new_type = type;
232
233 return -EBUSY;
234 }
235 }
236
237 if (new_type)
238 *new_type = req_type;
239
240 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
241 page = pfn_to_page(pfn);
242 set_page_memtype(page, req_type);
243 }
244 return 0;
245}
246
247static int free_ram_pages_type(u64 start, u64 end)
248{
249 struct page *page;
250 u64 pfn;
251
252 for (pfn = (start >> PAGE_SHIFT); pfn < (end >> PAGE_SHIFT); ++pfn) {
253 page = pfn_to_page(pfn);
254 set_page_memtype(page, -1);
255 }
256 return 0;
257}
258
259/*
260 * req_type typically has one of the:
261 * - _PAGE_CACHE_WB
262 * - _PAGE_CACHE_WC
263 * - _PAGE_CACHE_UC_MINUS
264 * - _PAGE_CACHE_UC
265 *
266 * If new_type is NULL, function will return an error if it cannot reserve the
267 * region with req_type. If new_type is non-NULL, function will return
268 * available type in new_type in case of no error. In case of any error
269 * it will return a negative return value.
270 */
271int reserve_memtype(u64 start, u64 end, unsigned long req_type,
272 unsigned long *new_type)
273{
274 struct memtype *new;
275 unsigned long actual_type;
276 int is_range_ram;
277 int err = 0;
278
279 BUG_ON(start >= end); /* end is exclusive */
280
281 if (!pat_enabled) {
282 /* This is identical to page table setting without PAT */
283 if (new_type) {
284 if (req_type == _PAGE_CACHE_WC)
285 *new_type = _PAGE_CACHE_UC_MINUS;
286 else
287 *new_type = req_type & _PAGE_CACHE_MASK;
288 }
289 return 0;
290 }
291
292 /* Low ISA region is always mapped WB in page table. No need to track */
293 if (x86_platform.is_untracked_pat_range(start, end)) {
294 if (new_type)
295 *new_type = _PAGE_CACHE_WB;
296 return 0;
297 }
298
299 /*
300 * Call mtrr_lookup to get the type hint. This is an
301 * optimization for /dev/mem mmap'ers into WB memory (BIOS
302 * tools and ACPI tools). Use WB request for WB memory and use
303 * UC_MINUS otherwise.
304 */
305 actual_type = pat_x_mtrr_type(start, end, req_type & _PAGE_CACHE_MASK);
306
307 if (new_type)
308 *new_type = actual_type;
309
310 is_range_ram = pat_pagerange_is_ram(start, end);
311 if (is_range_ram == 1) {
312
313 err = reserve_ram_pages_type(start, end, req_type, new_type);
314
315 return err;
316 } else if (is_range_ram < 0) {
317 return -EINVAL;
318 }
319
320 new = kzalloc(sizeof(struct memtype), GFP_KERNEL);
321 if (!new)
322 return -ENOMEM;
323
324 new->start = start;
325 new->end = end;
326 new->type = actual_type;
327
328 spin_lock(&memtype_lock);
329
330 err = rbt_memtype_check_insert(new, new_type);
331 if (err) {
332 printk(KERN_INFO "reserve_memtype failed [mem %#010Lx-%#010Lx], track %s, req %s\n",
333 start, end - 1,
334 cattr_name(new->type), cattr_name(req_type));
335 kfree(new);
336 spin_unlock(&memtype_lock);
337
338 return err;
339 }
340
341 spin_unlock(&memtype_lock);
342
343 dprintk("reserve_memtype added [mem %#010Lx-%#010Lx], track %s, req %s, ret %s\n",
344 start, end - 1, cattr_name(new->type), cattr_name(req_type),
345 new_type ? cattr_name(*new_type) : "-");
346
347 return err;
348}
349
350int free_memtype(u64 start, u64 end)
351{
352 int err = -EINVAL;
353 int is_range_ram;
354 struct memtype *entry;
355
356 if (!pat_enabled)
357 return 0;
358
359 /* Low ISA region is always mapped WB. No need to track */
360 if (x86_platform.is_untracked_pat_range(start, end))
361 return 0;
362
363 is_range_ram = pat_pagerange_is_ram(start, end);
364 if (is_range_ram == 1) {
365
366 err = free_ram_pages_type(start, end);
367
368 return err;
369 } else if (is_range_ram < 0) {
370 return -EINVAL;
371 }
372
373 spin_lock(&memtype_lock);
374 entry = rbt_memtype_erase(start, end);
375 spin_unlock(&memtype_lock);
376
377 if (!entry) {
378 printk(KERN_INFO "%s:%d freeing invalid memtype [mem %#010Lx-%#010Lx]\n",
379 current->comm, current->pid, start, end - 1);
380 return -EINVAL;
381 }
382
383 kfree(entry);
384
385 dprintk("free_memtype request [mem %#010Lx-%#010Lx]\n", start, end - 1);
386
387 return 0;
388}
389
390
391/**
392 * lookup_memtype - Looksup the memory type for a physical address
393 * @paddr: physical address of which memory type needs to be looked up
394 *
395 * Only to be called when PAT is enabled
396 *
397 * Returns _PAGE_CACHE_WB, _PAGE_CACHE_WC, _PAGE_CACHE_UC_MINUS or
398 * _PAGE_CACHE_UC
399 */
400static unsigned long lookup_memtype(u64 paddr)
401{
402 int rettype = _PAGE_CACHE_WB;
403 struct memtype *entry;
404
405 if (x86_platform.is_untracked_pat_range(paddr, paddr + PAGE_SIZE))
406 return rettype;
407
408 if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) {
409 struct page *page;
410 page = pfn_to_page(paddr >> PAGE_SHIFT);
411 rettype = get_page_memtype(page);
412 /*
413 * -1 from get_page_memtype() implies RAM page is in its
414 * default state and not reserved, and hence of type WB
415 */
416 if (rettype == -1)
417 rettype = _PAGE_CACHE_WB;
418
419 return rettype;
420 }
421
422 spin_lock(&memtype_lock);
423
424 entry = rbt_memtype_lookup(paddr);
425 if (entry != NULL)
426 rettype = entry->type;
427 else
428 rettype = _PAGE_CACHE_UC_MINUS;
429
430 spin_unlock(&memtype_lock);
431 return rettype;
432}
433
434/**
435 * io_reserve_memtype - Request a memory type mapping for a region of memory
436 * @start: start (physical address) of the region
437 * @end: end (physical address) of the region
438 * @type: A pointer to memtype, with requested type. On success, requested
439 * or any other compatible type that was available for the region is returned
440 *
441 * On success, returns 0
442 * On failure, returns non-zero
443 */
444int io_reserve_memtype(resource_size_t start, resource_size_t end,
445 unsigned long *type)
446{
447 resource_size_t size = end - start;
448 unsigned long req_type = *type;
449 unsigned long new_type;
450 int ret;
451
452 WARN_ON_ONCE(iomem_map_sanity_check(start, size));
453
454 ret = reserve_memtype(start, end, req_type, &new_type);
455 if (ret)
456 goto out_err;
457
458 if (!is_new_memtype_allowed(start, size, req_type, new_type))
459 goto out_free;
460
461 if (kernel_map_sync_memtype(start, size, new_type) < 0)
462 goto out_free;
463
464 *type = new_type;
465 return 0;
466
467out_free:
468 free_memtype(start, end);
469 ret = -EBUSY;
470out_err:
471 return ret;
472}
473
474/**
475 * io_free_memtype - Release a memory type mapping for a region of memory
476 * @start: start (physical address) of the region
477 * @end: end (physical address) of the region
478 */
479void io_free_memtype(resource_size_t start, resource_size_t end)
480{
481 free_memtype(start, end);
482}
483
484pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
485 unsigned long size, pgprot_t vma_prot)
486{
487 return vma_prot;
488}
489
490#ifdef CONFIG_STRICT_DEVMEM
491/* This check is done in drivers/char/mem.c in case of STRICT_DEVMEM*/
492static inline int range_is_allowed(unsigned long pfn, unsigned long size)
493{
494 return 1;
495}
496#else
497/* This check is needed to avoid cache aliasing when PAT is enabled */
498static inline int range_is_allowed(unsigned long pfn, unsigned long size)
499{
500 u64 from = ((u64)pfn) << PAGE_SHIFT;
501 u64 to = from + size;
502 u64 cursor = from;
503
504 if (!pat_enabled)
505 return 1;
506
507 while (cursor < to) {
508 if (!devmem_is_allowed(pfn)) {
509 printk(KERN_INFO "Program %s tried to access /dev/mem between [mem %#010Lx-%#010Lx]\n",
510 current->comm, from, to - 1);
511 return 0;
512 }
513 cursor += PAGE_SIZE;
514 pfn++;
515 }
516 return 1;
517}
518#endif /* CONFIG_STRICT_DEVMEM */
519
520int phys_mem_access_prot_allowed(struct file *file, unsigned long pfn,
521 unsigned long size, pgprot_t *vma_prot)
522{
523 unsigned long flags = _PAGE_CACHE_WB;
524
525 if (!range_is_allowed(pfn, size))
526 return 0;
527
528 if (file->f_flags & O_DSYNC)
529 flags = _PAGE_CACHE_UC_MINUS;
530
531#ifdef CONFIG_X86_32
532 /*
533 * On the PPro and successors, the MTRRs are used to set
534 * memory types for physical addresses outside main memory,
535 * so blindly setting UC or PWT on those pages is wrong.
536 * For Pentiums and earlier, the surround logic should disable
537 * caching for the high addresses through the KEN pin, but
538 * we maintain the tradition of paranoia in this code.
539 */
540 if (!pat_enabled &&
541 !(boot_cpu_has(X86_FEATURE_MTRR) ||
542 boot_cpu_has(X86_FEATURE_K6_MTRR) ||
543 boot_cpu_has(X86_FEATURE_CYRIX_ARR) ||
544 boot_cpu_has(X86_FEATURE_CENTAUR_MCR)) &&
545 (pfn << PAGE_SHIFT) >= __pa(high_memory)) {
546 flags = _PAGE_CACHE_UC;
547 }
548#endif
549
550 *vma_prot = __pgprot((pgprot_val(*vma_prot) & ~_PAGE_CACHE_MASK) |
551 flags);
552 return 1;
553}
554
555/*
556 * Change the memory type for the physial address range in kernel identity
557 * mapping space if that range is a part of identity map.
558 */
559int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
560{
561 unsigned long id_sz;
562
563 if (base > __pa(high_memory-1))
564 return 0;
565
566 /*
567 * some areas in the middle of the kernel identity range
568 * are not mapped, like the PCI space.
569 */
570 if (!page_is_ram(base >> PAGE_SHIFT))
571 return 0;
572
573 id_sz = (__pa(high_memory-1) <= base + size) ?
574 __pa(high_memory) - base :
575 size;
576
577 if (ioremap_change_attr((unsigned long)__va(base), id_sz, flags) < 0) {
578 printk(KERN_INFO "%s:%d ioremap_change_attr failed %s "
579 "for [mem %#010Lx-%#010Lx]\n",
580 current->comm, current->pid,
581 cattr_name(flags),
582 base, (unsigned long long)(base + size-1));
583 return -EINVAL;
584 }
585 return 0;
586}
587
588/*
589 * Internal interface to reserve a range of physical memory with prot.
590 * Reserved non RAM regions only and after successful reserve_memtype,
591 * this func also keeps identity mapping (if any) in sync with this new prot.
592 */
593static int reserve_pfn_range(u64 paddr, unsigned long size, pgprot_t *vma_prot,
594 int strict_prot)
595{
596 int is_ram = 0;
597 int ret;
598 unsigned long want_flags = (pgprot_val(*vma_prot) & _PAGE_CACHE_MASK);
599 unsigned long flags = want_flags;
600
601 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
602
603 /*
604 * reserve_pfn_range() for RAM pages. We do not refcount to keep
605 * track of number of mappings of RAM pages. We can assert that
606 * the type requested matches the type of first page in the range.
607 */
608 if (is_ram) {
609 if (!pat_enabled)
610 return 0;
611
612 flags = lookup_memtype(paddr);
613 if (want_flags != flags) {
614 printk(KERN_WARNING "%s:%d map pfn RAM range req %s for [mem %#010Lx-%#010Lx], got %s\n",
615 current->comm, current->pid,
616 cattr_name(want_flags),
617 (unsigned long long)paddr,
618 (unsigned long long)(paddr + size - 1),
619 cattr_name(flags));
620 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
621 (~_PAGE_CACHE_MASK)) |
622 flags);
623 }
624 return 0;
625 }
626
627 ret = reserve_memtype(paddr, paddr + size, want_flags, &flags);
628 if (ret)
629 return ret;
630
631 if (flags != want_flags) {
632 if (strict_prot ||
633 !is_new_memtype_allowed(paddr, size, want_flags, flags)) {
634 free_memtype(paddr, paddr + size);
635 printk(KERN_ERR "%s:%d map pfn expected mapping type %s"
636 " for [mem %#010Lx-%#010Lx], got %s\n",
637 current->comm, current->pid,
638 cattr_name(want_flags),
639 (unsigned long long)paddr,
640 (unsigned long long)(paddr + size - 1),
641 cattr_name(flags));
642 return -EINVAL;
643 }
644 /*
645 * We allow returning different type than the one requested in
646 * non strict case.
647 */
648 *vma_prot = __pgprot((pgprot_val(*vma_prot) &
649 (~_PAGE_CACHE_MASK)) |
650 flags);
651 }
652
653 if (kernel_map_sync_memtype(paddr, size, flags) < 0) {
654 free_memtype(paddr, paddr + size);
655 return -EINVAL;
656 }
657 return 0;
658}
659
660/*
661 * Internal interface to free a range of physical memory.
662 * Frees non RAM regions only.
663 */
664static void free_pfn_range(u64 paddr, unsigned long size)
665{
666 int is_ram;
667
668 is_ram = pat_pagerange_is_ram(paddr, paddr + size);
669 if (is_ram == 0)
670 free_memtype(paddr, paddr + size);
671}
672
673/*
674 * track_pfn_copy is called when vma that is covering the pfnmap gets
675 * copied through copy_page_range().
676 *
677 * If the vma has a linear pfn mapping for the entire range, we get the prot
678 * from pte and reserve the entire vma range with single reserve_pfn_range call.
679 */
680int track_pfn_copy(struct vm_area_struct *vma)
681{
682 resource_size_t paddr;
683 unsigned long prot;
684 unsigned long vma_size = vma->vm_end - vma->vm_start;
685 pgprot_t pgprot;
686
687 if (vma->vm_flags & VM_PAT) {
688 /*
689 * reserve the whole chunk covered by vma. We need the
690 * starting address and protection from pte.
691 */
692 if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) {
693 WARN_ON_ONCE(1);
694 return -EINVAL;
695 }
696 pgprot = __pgprot(prot);
697 return reserve_pfn_range(paddr, vma_size, &pgprot, 1);
698 }
699
700 return 0;
701}
702
703/*
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.
707 */
708int track_pfn_remap(struct vm_area_struct *vma, pgprot_t *prot,
709 unsigned long pfn, unsigned long addr, unsigned long size)
710{
711 resource_size_t paddr = (resource_size_t)pfn << PAGE_SHIFT;
712 unsigned long flags;
713
714 /* reserve the whole chunk starting from paddr */
715 if (addr == vma->vm_start && size == (vma->vm_end - vma->vm_start)) {
716 int ret;
717
718 ret = reserve_pfn_range(paddr, size, prot, 0);
719 if (!ret)
720 vma->vm_flags |= VM_PAT;
721 return ret;
722 }
723
724 if (!pat_enabled)
725 return 0;
726
727 /*
728 * For anything smaller than the vma size we set prot based on the
729 * lookup.
730 */
731 flags = lookup_memtype(paddr);
732
733 /* Check memtype for the remaining pages */
734 while (size > PAGE_SIZE) {
735 size -= PAGE_SIZE;
736 paddr += PAGE_SIZE;
737 if (flags != lookup_memtype(paddr))
738 return -EINVAL;
739 }
740
741 *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) |
742 flags);
743
744 return 0;
745}
746
747int track_pfn_insert(struct vm_area_struct *vma, pgprot_t *prot,
748 unsigned long pfn)
749{
750 unsigned long flags;
751
752 if (!pat_enabled)
753 return 0;
754
755 /* Set prot based on lookup */
756 flags = lookup_memtype((resource_size_t)pfn << PAGE_SHIFT);
757 *prot = __pgprot((pgprot_val(vma->vm_page_prot) & (~_PAGE_CACHE_MASK)) |
758 flags);
759
760 return 0;
761}
762
763/*
764 * untrack_pfn is called while unmapping a pfnmap for a region.
765 * untrack can be called for a specific region indicated by pfn and size or
766 * can be for the entire vma (in which case pfn, size are zero).
767 */
768void untrack_pfn(struct vm_area_struct *vma, unsigned long pfn,
769 unsigned long size)
770{
771 resource_size_t paddr;
772 unsigned long prot;
773
774 if (!(vma->vm_flags & VM_PAT))
775 return;
776
777 /* free the chunk starting from pfn or the whole chunk */
778 paddr = (resource_size_t)pfn << PAGE_SHIFT;
779 if (!paddr && !size) {
780 if (follow_phys(vma, vma->vm_start, 0, &prot, &paddr)) {
781 WARN_ON_ONCE(1);
782 return;
783 }
784
785 size = vma->vm_end - vma->vm_start;
786 }
787 free_pfn_range(paddr, size);
788 vma->vm_flags &= ~VM_PAT;
789}
790
791pgprot_t pgprot_writecombine(pgprot_t prot)
792{
793 if (pat_enabled)
794 return __pgprot(pgprot_val(prot) | _PAGE_CACHE_WC);
795 else
796 return pgprot_noncached(prot);
797}
798EXPORT_SYMBOL_GPL(pgprot_writecombine);
799
800#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_X86_PAT)
801
802static struct memtype *memtype_get_idx(loff_t pos)
803{
804 struct memtype *print_entry;
805 int ret;
806
807 print_entry = kzalloc(sizeof(struct memtype), GFP_KERNEL);
808 if (!print_entry)
809 return NULL;
810
811 spin_lock(&memtype_lock);
812 ret = rbt_memtype_copy_nth_element(print_entry, pos);
813 spin_unlock(&memtype_lock);
814
815 if (!ret) {
816 return print_entry;
817 } else {
818 kfree(print_entry);
819 return NULL;
820 }
821}
822
823static void *memtype_seq_start(struct seq_file *seq, loff_t *pos)
824{
825 if (*pos == 0) {
826 ++*pos;
827 seq_printf(seq, "PAT memtype list:\n");
828 }
829
830 return memtype_get_idx(*pos);
831}
832
833static void *memtype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
834{
835 ++*pos;
836 return memtype_get_idx(*pos);
837}
838
839static void memtype_seq_stop(struct seq_file *seq, void *v)
840{
841}
842
843static int memtype_seq_show(struct seq_file *seq, void *v)
844{
845 struct memtype *print_entry = (struct memtype *)v;
846
847 seq_printf(seq, "%s @ 0x%Lx-0x%Lx\n", cattr_name(print_entry->type),
848 print_entry->start, print_entry->end);
849 kfree(print_entry);
850
851 return 0;
852}
853
854static const struct seq_operations memtype_seq_ops = {
855 .start = memtype_seq_start,
856 .next = memtype_seq_next,
857 .stop = memtype_seq_stop,
858 .show = memtype_seq_show,
859};
860
861static int memtype_seq_open(struct inode *inode, struct file *file)
862{
863 return seq_open(file, &memtype_seq_ops);
864}
865
866static const struct file_operations memtype_fops = {
867 .open = memtype_seq_open,
868 .read = seq_read,
869 .llseek = seq_lseek,
870 .release = seq_release,
871};
872
873static int __init pat_memtype_list_init(void)
874{
875 if (pat_enabled) {
876 debugfs_create_file("pat_memtype_list", S_IRUSR,
877 arch_debugfs_dir, NULL, &memtype_fops);
878 }
879 return 0;
880}
881
882late_initcall(pat_memtype_list_init);
883
884#endif /* CONFIG_DEBUG_FS && CONFIG_X86_PAT */