1 /* Generic MTRR (Memory Type Range Register) driver.
3 Copyright (C) 1997-2000 Richard Gooch
4 Copyright (c) 2002 Patrick Mochel
6 This library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Library General Public
8 License as published by the Free Software Foundation; either
9 version 2 of the License, or (at your option) any later version.
11 This library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Library General Public License for more details.
16 You should have received a copy of the GNU Library General Public
17 License along with this library; if not, write to the Free
18 Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 Richard Gooch may be reached by email at rgooch@atnf.csiro.au
21 The postal address is:
22 Richard Gooch, c/o ATNF, P. O. Box 76, Epping, N.S.W., 2121, Australia.
24 Source: "Pentium Pro Family Developer's Manual, Volume 3:
25 Operating System Writer's Guide" (Intel document number 242692),
28 This was cleaned and made readable by Patrick Mochel <mochel@osdl.org>
30 Source: Intel Architecture Software Developers Manual, Volume 3:
31 System Programming Guide; Section 9.11. (1997 edition - PPro).
36 #include <linux/types.h> /* FIXME: kvm_para.h needs this */
38 #include <linux/stop_machine.h>
39 #include <linux/kvm_para.h>
40 #include <linux/uaccess.h>
41 #include <linux/module.h>
42 #include <linux/mutex.h>
43 #include <linux/init.h>
44 #include <linux/sort.h>
45 #include <linux/cpu.h>
46 #include <linux/pci.h>
47 #include <linux/smp.h>
48 #include <linux/syscore_ops.h>
50 #include <asm/processor.h>
58 /* arch_phys_wc_add returns an MTRR register index plus this offset. */
59 #define MTRR_TO_PHYS_WC_OFFSET 1000
62 static bool __mtrr_enabled
;
64 static bool mtrr_enabled(void)
66 return __mtrr_enabled
;
69 unsigned int mtrr_usage_table
[MTRR_MAX_VAR_RANGES
];
70 static DEFINE_MUTEX(mtrr_mutex
);
72 u64 size_or_mask
, size_and_mask
;
73 static bool mtrr_aps_delayed_init
;
75 static const struct mtrr_ops
*mtrr_ops
[X86_VENDOR_NUM
];
77 const struct mtrr_ops
*mtrr_if
;
79 static void set_mtrr(unsigned int reg
, unsigned long base
,
80 unsigned long size
, mtrr_type type
);
82 void set_mtrr_ops(const struct mtrr_ops
*ops
)
84 if (ops
->vendor
&& ops
->vendor
< X86_VENDOR_NUM
)
85 mtrr_ops
[ops
->vendor
] = ops
;
88 /* Returns non-zero if we have the write-combining memory type */
89 static int have_wrcomb(void)
93 dev
= pci_get_class(PCI_CLASS_BRIDGE_HOST
<< 8, NULL
);
96 * ServerWorks LE chipsets < rev 6 have problems with
97 * write-combining. Don't allow it and leave room for other
98 * chipsets to be tagged
100 if (dev
->vendor
== PCI_VENDOR_ID_SERVERWORKS
&&
101 dev
->device
== PCI_DEVICE_ID_SERVERWORKS_LE
&&
102 dev
->revision
<= 5) {
103 pr_info("mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
108 * Intel 450NX errata # 23. Non ascending cacheline evictions to
109 * write combining memory may resulting in data corruption
111 if (dev
->vendor
== PCI_VENDOR_ID_INTEL
&&
112 dev
->device
== PCI_DEVICE_ID_INTEL_82451NX
) {
113 pr_info("mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
119 return mtrr_if
->have_wrcomb
? mtrr_if
->have_wrcomb() : 0;
122 /* This function returns the number of variable MTRRs */
123 static void __init
set_num_var_ranges(void)
125 unsigned long config
= 0, dummy
;
128 rdmsr(MSR_MTRRcap
, config
, dummy
);
129 else if (is_cpu(AMD
))
131 else if (is_cpu(CYRIX
) || is_cpu(CENTAUR
))
134 num_var_ranges
= config
& 0xff;
137 static void __init
init_table(void)
141 max
= num_var_ranges
;
142 for (i
= 0; i
< max
; i
++)
143 mtrr_usage_table
[i
] = 1;
146 struct set_mtrr_data
{
147 unsigned long smp_base
;
148 unsigned long smp_size
;
149 unsigned int smp_reg
;
154 * mtrr_rendezvous_handler - Work done in the synchronization handler. Executed
156 * @info: pointer to mtrr configuration data
160 static int mtrr_rendezvous_handler(void *info
)
162 struct set_mtrr_data
*data
= info
;
165 * We use this same function to initialize the mtrrs during boot,
166 * resume, runtime cpu online and on an explicit request to set a
169 * During boot or suspend, the state of the boot cpu's mtrrs has been
170 * saved, and we want to replicate that across all the cpus that come
171 * online (either at the end of boot or resume or during a runtime cpu
172 * online). If we're doing that, @reg is set to something special and on
173 * all the cpu's we do mtrr_if->set_all() (On the logical cpu that
174 * started the boot/resume sequence, this might be a duplicate
177 if (data
->smp_reg
!= ~0U) {
178 mtrr_if
->set(data
->smp_reg
, data
->smp_base
,
179 data
->smp_size
, data
->smp_type
);
180 } else if (mtrr_aps_delayed_init
|| !cpu_online(smp_processor_id())) {
186 static inline int types_compatible(mtrr_type type1
, mtrr_type type2
)
188 return type1
== MTRR_TYPE_UNCACHABLE
||
189 type2
== MTRR_TYPE_UNCACHABLE
||
190 (type1
== MTRR_TYPE_WRTHROUGH
&& type2
== MTRR_TYPE_WRBACK
) ||
191 (type1
== MTRR_TYPE_WRBACK
&& type2
== MTRR_TYPE_WRTHROUGH
);
195 * set_mtrr - update mtrrs on all processors
196 * @reg: mtrr in question
201 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
203 * 1. Queue work to do the following on all processors:
204 * 2. Disable Interrupts
205 * 3. Wait for all procs to do so
206 * 4. Enter no-fill cache mode
210 * 8. Disable all range registers
211 * 9. Update the MTRRs
212 * 10. Enable all range registers
213 * 11. Flush all TLBs and caches again
214 * 12. Enter normal cache mode and reenable caching
216 * 14. Wait for buddies to catch up
217 * 15. Enable interrupts.
219 * What does that mean for us? Well, stop_machine() will ensure that
220 * the rendezvous handler is started on each CPU. And in lockstep they
221 * do the state transition of disabling interrupts, updating MTRR's
222 * (the CPU vendors may each do it differently, so we call mtrr_if->set()
223 * callback and let them take care of it.) and enabling interrupts.
225 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
229 set_mtrr(unsigned int reg
, unsigned long base
, unsigned long size
, mtrr_type type
)
231 struct set_mtrr_data data
= { .smp_reg
= reg
,
237 stop_machine(mtrr_rendezvous_handler
, &data
, cpu_online_mask
);
240 static void set_mtrr_from_inactive_cpu(unsigned int reg
, unsigned long base
,
241 unsigned long size
, mtrr_type type
)
243 struct set_mtrr_data data
= { .smp_reg
= reg
,
249 stop_machine_from_inactive_cpu(mtrr_rendezvous_handler
, &data
,
254 * mtrr_add_page - Add a memory type region
255 * @base: Physical base address of region in pages (in units of 4 kB!)
256 * @size: Physical size of region in pages (4 kB)
257 * @type: Type of MTRR desired
258 * @increment: If this is true do usage counting on the region
260 * Memory type region registers control the caching on newer Intel and
261 * non Intel processors. This function allows drivers to request an
262 * MTRR is added. The details and hardware specifics of each processor's
263 * implementation are hidden from the caller, but nevertheless the
264 * caller should expect to need to provide a power of two size on an
265 * equivalent power of two boundary.
267 * If the region cannot be added either because all regions are in use
268 * or the CPU cannot support it a negative value is returned. On success
269 * the register number for this entry is returned, but should be treated
272 * On a multiprocessor machine the changes are made to all processors.
273 * This is required on x86 by the Intel processors.
275 * The available types are
277 * %MTRR_TYPE_UNCACHABLE - No caching
279 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
281 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
283 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
285 * BUGS: Needs a quiet flag for the cases where drivers do not mind
286 * failures and do not wish system log messages to be sent.
288 int mtrr_add_page(unsigned long base
, unsigned long size
,
289 unsigned int type
, bool increment
)
291 unsigned long lbase
, lsize
;
292 int i
, replace
, error
;
298 error
= mtrr_if
->validate_add_page(base
, size
, type
);
302 if (type
>= MTRR_NUM_TYPES
) {
303 pr_warning("mtrr: type: %u invalid\n", type
);
307 /* If the type is WC, check that this processor supports it */
308 if ((type
== MTRR_TYPE_WRCOMB
) && !have_wrcomb()) {
309 pr_warning("mtrr: your processor doesn't support write-combining\n");
314 pr_warning("mtrr: zero sized request\n");
318 if ((base
| (base
+ size
- 1)) >>
319 (boot_cpu_data
.x86_phys_bits
- PAGE_SHIFT
)) {
320 pr_warning("mtrr: base or size exceeds the MTRR width\n");
327 /* No CPU hotplug when we change MTRR entries */
330 /* Search for existing MTRR */
331 mutex_lock(&mtrr_mutex
);
332 for (i
= 0; i
< num_var_ranges
; ++i
) {
333 mtrr_if
->get(i
, &lbase
, &lsize
, <ype
);
334 if (!lsize
|| base
> lbase
+ lsize
- 1 ||
335 base
+ size
- 1 < lbase
)
338 * At this point we know there is some kind of
341 if (base
< lbase
|| base
+ size
- 1 > lbase
+ lsize
- 1) {
343 base
+ size
- 1 >= lbase
+ lsize
- 1) {
344 /* New region encloses an existing region */
346 replace
= replace
== -1 ? i
: -2;
348 } else if (types_compatible(type
, ltype
))
351 pr_warning("mtrr: 0x%lx000,0x%lx000 overlaps existing"
352 " 0x%lx000,0x%lx000\n", base
, size
, lbase
,
356 /* New region is enclosed by an existing region */
358 if (types_compatible(type
, ltype
))
360 pr_warning("mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
361 base
, size
, mtrr_attrib_to_str(ltype
),
362 mtrr_attrib_to_str(type
));
366 ++mtrr_usage_table
[i
];
370 /* Search for an empty MTRR */
371 i
= mtrr_if
->get_free_region(base
, size
, replace
);
373 set_mtrr(i
, base
, size
, type
);
374 if (likely(replace
< 0)) {
375 mtrr_usage_table
[i
] = 1;
377 mtrr_usage_table
[i
] = mtrr_usage_table
[replace
];
379 mtrr_usage_table
[i
]++;
380 if (unlikely(replace
!= i
)) {
381 set_mtrr(replace
, 0, 0, 0);
382 mtrr_usage_table
[replace
] = 0;
386 pr_info("mtrr: no more MTRRs available\n");
390 mutex_unlock(&mtrr_mutex
);
395 static int mtrr_check(unsigned long base
, unsigned long size
)
397 if ((base
& (PAGE_SIZE
- 1)) || (size
& (PAGE_SIZE
- 1))) {
398 pr_warning("mtrr: size and base must be multiples of 4 kiB\n");
399 pr_debug("mtrr: size: 0x%lx base: 0x%lx\n", size
, base
);
407 * mtrr_add - Add a memory type region
408 * @base: Physical base address of region
409 * @size: Physical size of region
410 * @type: Type of MTRR desired
411 * @increment: If this is true do usage counting on the region
413 * Memory type region registers control the caching on newer Intel and
414 * non Intel processors. This function allows drivers to request an
415 * MTRR is added. The details and hardware specifics of each processor's
416 * implementation are hidden from the caller, but nevertheless the
417 * caller should expect to need to provide a power of two size on an
418 * equivalent power of two boundary.
420 * If the region cannot be added either because all regions are in use
421 * or the CPU cannot support it a negative value is returned. On success
422 * the register number for this entry is returned, but should be treated
425 * On a multiprocessor machine the changes are made to all processors.
426 * This is required on x86 by the Intel processors.
428 * The available types are
430 * %MTRR_TYPE_UNCACHABLE - No caching
432 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
434 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
436 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
438 * BUGS: Needs a quiet flag for the cases where drivers do not mind
439 * failures and do not wish system log messages to be sent.
441 int mtrr_add(unsigned long base
, unsigned long size
, unsigned int type
,
446 if (mtrr_check(base
, size
))
448 return mtrr_add_page(base
>> PAGE_SHIFT
, size
>> PAGE_SHIFT
, type
,
453 * mtrr_del_page - delete a memory type region
454 * @reg: Register returned by mtrr_add
455 * @base: Physical base address
456 * @size: Size of region
458 * If register is supplied then base and size are ignored. This is
459 * how drivers should call it.
461 * Releases an MTRR region. If the usage count drops to zero the
462 * register is freed and the region returns to default state.
463 * On success the register is returned, on failure a negative error
466 int mtrr_del_page(int reg
, unsigned long base
, unsigned long size
)
470 unsigned long lbase
, lsize
;
476 max
= num_var_ranges
;
477 /* No CPU hotplug when we change MTRR entries */
479 mutex_lock(&mtrr_mutex
);
481 /* Search for existing MTRR */
482 for (i
= 0; i
< max
; ++i
) {
483 mtrr_if
->get(i
, &lbase
, &lsize
, <ype
);
484 if (lbase
== base
&& lsize
== size
) {
490 pr_debug("mtrr: no MTRR for %lx000,%lx000 found\n",
496 pr_warning("mtrr: register: %d too big\n", reg
);
499 mtrr_if
->get(reg
, &lbase
, &lsize
, <ype
);
501 pr_warning("mtrr: MTRR %d not used\n", reg
);
504 if (mtrr_usage_table
[reg
] < 1) {
505 pr_warning("mtrr: reg: %d has count=0\n", reg
);
508 if (--mtrr_usage_table
[reg
] < 1)
509 set_mtrr(reg
, 0, 0, 0);
512 mutex_unlock(&mtrr_mutex
);
518 * mtrr_del - delete a memory type region
519 * @reg: Register returned by mtrr_add
520 * @base: Physical base address
521 * @size: Size of region
523 * If register is supplied then base and size are ignored. This is
524 * how drivers should call it.
526 * Releases an MTRR region. If the usage count drops to zero the
527 * register is freed and the region returns to default state.
528 * On success the register is returned, on failure a negative error
531 int mtrr_del(int reg
, unsigned long base
, unsigned long size
)
535 if (mtrr_check(base
, size
))
537 return mtrr_del_page(reg
, base
>> PAGE_SHIFT
, size
>> PAGE_SHIFT
);
541 * arch_phys_wc_add - add a WC MTRR and handle errors if PAT is unavailable
542 * @base: Physical base address
543 * @size: Size of region
545 * If PAT is available, this does nothing. If PAT is unavailable, it
546 * attempts to add a WC MTRR covering size bytes starting at base and
547 * logs an error if this fails.
549 * The called should provide a power of two size on an equivalent
550 * power of two boundary.
552 * Drivers must store the return value to pass to mtrr_del_wc_if_needed,
553 * but drivers should not try to interpret that return value.
555 int arch_phys_wc_add(unsigned long base
, unsigned long size
)
559 if (pat_enabled() || !mtrr_enabled())
560 return 0; /* Success! (We don't need to do anything.) */
562 ret
= mtrr_add(base
, size
, MTRR_TYPE_WRCOMB
, true);
564 pr_warn("Failed to add WC MTRR for [%p-%p]; performance may suffer.",
565 (void *)base
, (void *)(base
+ size
- 1));
568 return ret
+ MTRR_TO_PHYS_WC_OFFSET
;
570 EXPORT_SYMBOL(arch_phys_wc_add
);
573 * arch_phys_wc_del - undoes arch_phys_wc_add
574 * @handle: Return value from arch_phys_wc_add
576 * This cleans up after mtrr_add_wc_if_needed.
578 * The API guarantees that mtrr_del_wc_if_needed(error code) and
579 * mtrr_del_wc_if_needed(0) do nothing.
581 void arch_phys_wc_del(int handle
)
584 WARN_ON(handle
< MTRR_TO_PHYS_WC_OFFSET
);
585 mtrr_del(handle
- MTRR_TO_PHYS_WC_OFFSET
, 0, 0);
588 EXPORT_SYMBOL(arch_phys_wc_del
);
591 * arch_phys_wc_index - translates arch_phys_wc_add's return value
592 * @handle: Return value from arch_phys_wc_add
594 * This will turn the return value from arch_phys_wc_add into an mtrr
595 * index suitable for debugging.
597 * Note: There is no legitimate use for this function, except possibly
598 * in printk line. Alas there is an illegitimate use in some ancient
601 int arch_phys_wc_index(int handle
)
603 if (handle
< MTRR_TO_PHYS_WC_OFFSET
)
606 return handle
- MTRR_TO_PHYS_WC_OFFSET
;
608 EXPORT_SYMBOL_GPL(arch_phys_wc_index
);
612 * These should be called implicitly, but we can't yet until all the initcall
615 static void __init
init_ifs(void)
617 #ifndef CONFIG_X86_64
624 /* The suspend/resume methods are only for CPU without MTRR. CPU using generic
625 * MTRR driver doesn't require this
633 static struct mtrr_value mtrr_value
[MTRR_MAX_VAR_RANGES
];
635 static int mtrr_save(void)
639 for (i
= 0; i
< num_var_ranges
; i
++) {
640 mtrr_if
->get(i
, &mtrr_value
[i
].lbase
,
641 &mtrr_value
[i
].lsize
,
642 &mtrr_value
[i
].ltype
);
647 static void mtrr_restore(void)
651 for (i
= 0; i
< num_var_ranges
; i
++) {
652 if (mtrr_value
[i
].lsize
) {
653 set_mtrr(i
, mtrr_value
[i
].lbase
,
655 mtrr_value
[i
].ltype
);
662 static struct syscore_ops mtrr_syscore_ops
= {
663 .suspend
= mtrr_save
,
664 .resume
= mtrr_restore
,
667 int __initdata changed_by_mtrr_cleanup
;
669 #define SIZE_OR_MASK_BITS(n) (~((1ULL << ((n) - PAGE_SHIFT)) - 1))
671 * mtrr_bp_init - initialize mtrrs on the boot CPU
673 * This needs to be called early; before any of the other CPUs are
674 * initialized (i.e. before smp_init()).
677 void __init
mtrr_bp_init(void)
685 if (boot_cpu_has(X86_FEATURE_MTRR
)) {
686 mtrr_if
= &generic_mtrr_ops
;
687 size_or_mask
= SIZE_OR_MASK_BITS(36);
688 size_and_mask
= 0x00f00000;
692 * This is an AMD specific MSR, but we assume(hope?) that
693 * Intel will implement it too when they extend the address
696 if (cpuid_eax(0x80000000) >= 0x80000008) {
697 phys_addr
= cpuid_eax(0x80000008) & 0xff;
698 /* CPUID workaround for Intel 0F33/0F34 CPU */
699 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_INTEL
&&
700 boot_cpu_data
.x86
== 0xF &&
701 boot_cpu_data
.x86_model
== 0x3 &&
702 (boot_cpu_data
.x86_mask
== 0x3 ||
703 boot_cpu_data
.x86_mask
== 0x4))
706 size_or_mask
= SIZE_OR_MASK_BITS(phys_addr
);
707 size_and_mask
= ~size_or_mask
& 0xfffff00000ULL
;
708 } else if (boot_cpu_data
.x86_vendor
== X86_VENDOR_CENTAUR
&&
709 boot_cpu_data
.x86
== 6) {
711 * VIA C* family have Intel style MTRRs,
712 * but don't support PAE
714 size_or_mask
= SIZE_OR_MASK_BITS(32);
719 switch (boot_cpu_data
.x86_vendor
) {
721 if (cpu_feature_enabled(X86_FEATURE_K6_MTRR
)) {
722 /* Pre-Athlon (K6) AMD CPU MTRRs */
723 mtrr_if
= mtrr_ops
[X86_VENDOR_AMD
];
724 size_or_mask
= SIZE_OR_MASK_BITS(32);
728 case X86_VENDOR_CENTAUR
:
729 if (cpu_feature_enabled(X86_FEATURE_CENTAUR_MCR
)) {
730 mtrr_if
= mtrr_ops
[X86_VENDOR_CENTAUR
];
731 size_or_mask
= SIZE_OR_MASK_BITS(32);
735 case X86_VENDOR_CYRIX
:
736 if (cpu_feature_enabled(X86_FEATURE_CYRIX_ARR
)) {
737 mtrr_if
= mtrr_ops
[X86_VENDOR_CYRIX
];
738 size_or_mask
= SIZE_OR_MASK_BITS(32);
748 __mtrr_enabled
= true;
749 set_num_var_ranges();
752 /* BIOS may override */
753 __mtrr_enabled
= get_mtrr_state();
755 if (mtrr_cleanup(phys_addr
)) {
756 changed_by_mtrr_cleanup
= 1;
763 pr_info("MTRR: Disabled\n");
766 void mtrr_ap_init(void)
771 if (!use_intel() || mtrr_aps_delayed_init
)
774 * Ideally we should hold mtrr_mutex here to avoid mtrr entries
775 * changed, but this routine will be called in cpu boot time,
776 * holding the lock breaks it.
778 * This routine is called in two cases:
780 * 1. very earily time of software resume, when there absolutely
781 * isn't mtrr entry changes;
783 * 2. cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug
784 * lock to prevent mtrr entry changes
786 set_mtrr_from_inactive_cpu(~0U, 0, 0, 0);
790 * Save current fixed-range MTRR state of the first cpu in cpu_online_mask.
792 void mtrr_save_state(void)
800 first_cpu
= cpumask_first(cpu_online_mask
);
801 smp_call_function_single(first_cpu
, mtrr_save_fixed_ranges
, NULL
, 1);
805 void set_mtrr_aps_delayed_init(void)
812 mtrr_aps_delayed_init
= true;
816 * Delayed MTRR initialization for all AP's
818 void mtrr_aps_init(void)
820 if (!use_intel() || !mtrr_enabled())
824 * Check if someone has requested the delay of AP MTRR initialization,
825 * by doing set_mtrr_aps_delayed_init(), prior to this point. If not,
828 if (!mtrr_aps_delayed_init
)
831 set_mtrr(~0U, 0, 0, 0);
832 mtrr_aps_delayed_init
= false;
835 void mtrr_bp_restore(void)
837 if (!use_intel() || !mtrr_enabled())
843 static int __init
mtrr_init_finialize(void)
849 if (!changed_by_mtrr_cleanup
)
855 * The CPU has no MTRR and seems to not support SMP. They have
856 * specific drivers, we use a tricky method to support
857 * suspend/resume for them.
859 * TBD: is there any system with such CPU which supports
860 * suspend/resume? If no, we should remove the code.
862 register_syscore_ops(&mtrr_syscore_ops
);
866 subsys_initcall(mtrr_init_finialize
);