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/kvm_para.h>
39 #include <linux/uaccess.h>
40 #include <linux/module.h>
41 #include <linux/mutex.h>
42 #include <linux/init.h>
43 #include <linux/sort.h>
44 #include <linux/cpu.h>
45 #include <linux/pci.h>
46 #include <linux/smp.h>
48 #include <asm/processor.h>
57 unsigned int mtrr_usage_table
[MTRR_MAX_VAR_RANGES
];
58 static DEFINE_MUTEX(mtrr_mutex
);
60 u64 size_or_mask
, size_and_mask
;
61 static bool mtrr_aps_delayed_init
;
63 static struct mtrr_ops
*mtrr_ops
[X86_VENDOR_NUM
];
65 struct mtrr_ops
*mtrr_if
;
67 static void set_mtrr(unsigned int reg
, unsigned long base
,
68 unsigned long size
, mtrr_type type
);
70 void set_mtrr_ops(struct mtrr_ops
*ops
)
72 if (ops
->vendor
&& ops
->vendor
< X86_VENDOR_NUM
)
73 mtrr_ops
[ops
->vendor
] = ops
;
76 /* Returns non-zero if we have the write-combining memory type */
77 static int have_wrcomb(void)
82 dev
= pci_get_class(PCI_CLASS_BRIDGE_HOST
<< 8, NULL
);
85 * ServerWorks LE chipsets < rev 6 have problems with
86 * write-combining. Don't allow it and leave room for other
87 * chipsets to be tagged
89 if (dev
->vendor
== PCI_VENDOR_ID_SERVERWORKS
&&
90 dev
->device
== PCI_DEVICE_ID_SERVERWORKS_LE
) {
91 pci_read_config_byte(dev
, PCI_CLASS_REVISION
, &rev
);
93 pr_info("mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
99 * Intel 450NX errata # 23. Non ascending cacheline evictions to
100 * write combining memory may resulting in data corruption
102 if (dev
->vendor
== PCI_VENDOR_ID_INTEL
&&
103 dev
->device
== PCI_DEVICE_ID_INTEL_82451NX
) {
104 pr_info("mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
110 return mtrr_if
->have_wrcomb
? mtrr_if
->have_wrcomb() : 0;
113 /* This function returns the number of variable MTRRs */
114 static void __init
set_num_var_ranges(void)
116 unsigned long config
= 0, dummy
;
119 rdmsr(MSR_MTRRcap
, config
, dummy
);
120 else if (is_cpu(AMD
))
122 else if (is_cpu(CYRIX
) || is_cpu(CENTAUR
))
125 num_var_ranges
= config
& 0xff;
128 static void __init
init_table(void)
132 max
= num_var_ranges
;
133 for (i
= 0; i
< max
; i
++)
134 mtrr_usage_table
[i
] = 1;
137 struct set_mtrr_data
{
140 unsigned long smp_base
;
141 unsigned long smp_size
;
142 unsigned int smp_reg
;
147 * ipi_handler - Synchronisation handler. Executed by "other" CPUs.
151 static void ipi_handler(void *info
)
154 struct set_mtrr_data
*data
= info
;
157 local_irq_save(flags
);
159 atomic_dec(&data
->count
);
160 while (!atomic_read(&data
->gate
))
163 /* The master has cleared me to execute */
164 if (data
->smp_reg
!= ~0U) {
165 mtrr_if
->set(data
->smp_reg
, data
->smp_base
,
166 data
->smp_size
, data
->smp_type
);
167 } else if (mtrr_aps_delayed_init
) {
169 * Initialize the MTRRs inaddition to the synchronisation.
174 atomic_dec(&data
->count
);
175 while (atomic_read(&data
->gate
))
178 atomic_dec(&data
->count
);
179 local_irq_restore(flags
);
183 static inline int types_compatible(mtrr_type type1
, mtrr_type type2
)
185 return type1
== MTRR_TYPE_UNCACHABLE
||
186 type2
== MTRR_TYPE_UNCACHABLE
||
187 (type1
== MTRR_TYPE_WRTHROUGH
&& type2
== MTRR_TYPE_WRBACK
) ||
188 (type1
== MTRR_TYPE_WRBACK
&& type2
== MTRR_TYPE_WRTHROUGH
);
192 * set_mtrr - update mtrrs on all processors
193 * @reg: mtrr in question
198 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
200 * 1. Send IPI to do the following:
201 * 2. Disable Interrupts
202 * 3. Wait for all procs to do so
203 * 4. Enter no-fill cache mode
207 * 8. Disable all range registers
208 * 9. Update the MTRRs
209 * 10. Enable all range registers
210 * 11. Flush all TLBs and caches again
211 * 12. Enter normal cache mode and reenable caching
213 * 14. Wait for buddies to catch up
214 * 15. Enable interrupts.
216 * What does that mean for us? Well, first we set data.count to the number
217 * of CPUs. As each CPU disables interrupts, it'll decrement it once. We wait
218 * until it hits 0 and proceed. We set the data.gate flag and reset data.count.
219 * Meanwhile, they are waiting for that flag to be set. Once it's set, each
220 * CPU goes through the transition of updating MTRRs.
221 * The CPU vendors may each do it differently,
222 * so we call mtrr_if->set() callback and let them take care of it.
223 * When they're done, they again decrement data->count and wait for data.gate
225 * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag
226 * Everyone then enables interrupts and we all continue on.
228 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
232 set_mtrr(unsigned int reg
, unsigned long base
, unsigned long size
, mtrr_type type
)
234 struct set_mtrr_data data
;
238 data
.smp_base
= base
;
239 data
.smp_size
= size
;
240 data
.smp_type
= type
;
241 atomic_set(&data
.count
, num_booting_cpus() - 1);
243 /* Make sure data.count is visible before unleashing other CPUs */
245 atomic_set(&data
.gate
, 0);
247 /* Start the ball rolling on other CPUs */
248 if (smp_call_function(ipi_handler
, &data
, 0) != 0)
249 panic("mtrr: timed out waiting for other CPUs\n");
251 local_irq_save(flags
);
253 while (atomic_read(&data
.count
))
256 /* Ok, reset count and toggle gate */
257 atomic_set(&data
.count
, num_booting_cpus() - 1);
259 atomic_set(&data
.gate
, 1);
261 /* Do our MTRR business */
265 * We use this same function to initialize the mtrrs on boot.
266 * The state of the boot cpu's mtrrs has been saved, and we want
267 * to replicate across all the APs.
268 * If we're doing that @reg is set to something special...
271 mtrr_if
->set(reg
, base
, size
, type
);
272 else if (!mtrr_aps_delayed_init
)
275 /* Wait for the others */
276 while (atomic_read(&data
.count
))
279 atomic_set(&data
.count
, num_booting_cpus() - 1);
281 atomic_set(&data
.gate
, 0);
284 * Wait here for everyone to have seen the gate change
285 * So we're the last ones to touch 'data'
287 while (atomic_read(&data
.count
))
290 local_irq_restore(flags
);
294 * mtrr_add_page - Add a memory type region
295 * @base: Physical base address of region in pages (in units of 4 kB!)
296 * @size: Physical size of region in pages (4 kB)
297 * @type: Type of MTRR desired
298 * @increment: If this is true do usage counting on the region
300 * Memory type region registers control the caching on newer Intel and
301 * non Intel processors. This function allows drivers to request an
302 * MTRR is added. The details and hardware specifics of each processor's
303 * implementation are hidden from the caller, but nevertheless the
304 * caller should expect to need to provide a power of two size on an
305 * equivalent power of two boundary.
307 * If the region cannot be added either because all regions are in use
308 * or the CPU cannot support it a negative value is returned. On success
309 * the register number for this entry is returned, but should be treated
312 * On a multiprocessor machine the changes are made to all processors.
313 * This is required on x86 by the Intel processors.
315 * The available types are
317 * %MTRR_TYPE_UNCACHABLE - No caching
319 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
321 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
323 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
325 * BUGS: Needs a quiet flag for the cases where drivers do not mind
326 * failures and do not wish system log messages to be sent.
328 int mtrr_add_page(unsigned long base
, unsigned long size
,
329 unsigned int type
, bool increment
)
331 unsigned long lbase
, lsize
;
332 int i
, replace
, error
;
338 error
= mtrr_if
->validate_add_page(base
, size
, type
);
342 if (type
>= MTRR_NUM_TYPES
) {
343 pr_warning("mtrr: type: %u invalid\n", type
);
347 /* If the type is WC, check that this processor supports it */
348 if ((type
== MTRR_TYPE_WRCOMB
) && !have_wrcomb()) {
349 pr_warning("mtrr: your processor doesn't support write-combining\n");
354 pr_warning("mtrr: zero sized request\n");
358 if (base
& size_or_mask
|| size
& size_or_mask
) {
359 pr_warning("mtrr: base or size exceeds the MTRR width\n");
366 /* No CPU hotplug when we change MTRR entries */
369 /* Search for existing MTRR */
370 mutex_lock(&mtrr_mutex
);
371 for (i
= 0; i
< num_var_ranges
; ++i
) {
372 mtrr_if
->get(i
, &lbase
, &lsize
, <ype
);
373 if (!lsize
|| base
> lbase
+ lsize
- 1 ||
374 base
+ size
- 1 < lbase
)
377 * At this point we know there is some kind of
380 if (base
< lbase
|| base
+ size
- 1 > lbase
+ lsize
- 1) {
382 base
+ size
- 1 >= lbase
+ lsize
- 1) {
383 /* New region encloses an existing region */
385 replace
= replace
== -1 ? i
: -2;
387 } else if (types_compatible(type
, ltype
))
390 pr_warning("mtrr: 0x%lx000,0x%lx000 overlaps existing"
391 " 0x%lx000,0x%lx000\n", base
, size
, lbase
,
395 /* New region is enclosed by an existing region */
397 if (types_compatible(type
, ltype
))
399 pr_warning("mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
400 base
, size
, mtrr_attrib_to_str(ltype
),
401 mtrr_attrib_to_str(type
));
405 ++mtrr_usage_table
[i
];
409 /* Search for an empty MTRR */
410 i
= mtrr_if
->get_free_region(base
, size
, replace
);
412 set_mtrr(i
, base
, size
, type
);
413 if (likely(replace
< 0)) {
414 mtrr_usage_table
[i
] = 1;
416 mtrr_usage_table
[i
] = mtrr_usage_table
[replace
];
418 mtrr_usage_table
[i
]++;
419 if (unlikely(replace
!= i
)) {
420 set_mtrr(replace
, 0, 0, 0);
421 mtrr_usage_table
[replace
] = 0;
425 pr_info("mtrr: no more MTRRs available\n");
429 mutex_unlock(&mtrr_mutex
);
434 static int mtrr_check(unsigned long base
, unsigned long size
)
436 if ((base
& (PAGE_SIZE
- 1)) || (size
& (PAGE_SIZE
- 1))) {
437 pr_warning("mtrr: size and base must be multiples of 4 kiB\n");
438 pr_debug("mtrr: size: 0x%lx base: 0x%lx\n", size
, base
);
446 * mtrr_add - Add a memory type region
447 * @base: Physical base address of region
448 * @size: Physical size of region
449 * @type: Type of MTRR desired
450 * @increment: If this is true do usage counting on the region
452 * Memory type region registers control the caching on newer Intel and
453 * non Intel processors. This function allows drivers to request an
454 * MTRR is added. The details and hardware specifics of each processor's
455 * implementation are hidden from the caller, but nevertheless the
456 * caller should expect to need to provide a power of two size on an
457 * equivalent power of two boundary.
459 * If the region cannot be added either because all regions are in use
460 * or the CPU cannot support it a negative value is returned. On success
461 * the register number for this entry is returned, but should be treated
464 * On a multiprocessor machine the changes are made to all processors.
465 * This is required on x86 by the Intel processors.
467 * The available types are
469 * %MTRR_TYPE_UNCACHABLE - No caching
471 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
473 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
475 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
477 * BUGS: Needs a quiet flag for the cases where drivers do not mind
478 * failures and do not wish system log messages to be sent.
480 int mtrr_add(unsigned long base
, unsigned long size
, unsigned int type
,
483 if (mtrr_check(base
, size
))
485 return mtrr_add_page(base
>> PAGE_SHIFT
, size
>> PAGE_SHIFT
, type
,
488 EXPORT_SYMBOL(mtrr_add
);
491 * mtrr_del_page - delete a memory type region
492 * @reg: Register returned by mtrr_add
493 * @base: Physical base address
494 * @size: Size of region
496 * If register is supplied then base and size are ignored. This is
497 * how drivers should call it.
499 * Releases an MTRR region. If the usage count drops to zero the
500 * register is freed and the region returns to default state.
501 * On success the register is returned, on failure a negative error
504 int mtrr_del_page(int reg
, unsigned long base
, unsigned long size
)
508 unsigned long lbase
, lsize
;
514 max
= num_var_ranges
;
515 /* No CPU hotplug when we change MTRR entries */
517 mutex_lock(&mtrr_mutex
);
519 /* Search for existing MTRR */
520 for (i
= 0; i
< max
; ++i
) {
521 mtrr_if
->get(i
, &lbase
, &lsize
, <ype
);
522 if (lbase
== base
&& lsize
== size
) {
528 pr_debug("mtrr: no MTRR for %lx000,%lx000 found\n",
534 pr_warning("mtrr: register: %d too big\n", reg
);
537 mtrr_if
->get(reg
, &lbase
, &lsize
, <ype
);
539 pr_warning("mtrr: MTRR %d not used\n", reg
);
542 if (mtrr_usage_table
[reg
] < 1) {
543 pr_warning("mtrr: reg: %d has count=0\n", reg
);
546 if (--mtrr_usage_table
[reg
] < 1)
547 set_mtrr(reg
, 0, 0, 0);
550 mutex_unlock(&mtrr_mutex
);
556 * mtrr_del - delete a memory type region
557 * @reg: Register returned by mtrr_add
558 * @base: Physical base address
559 * @size: Size of region
561 * If register is supplied then base and size are ignored. This is
562 * how drivers should call it.
564 * Releases an MTRR region. If the usage count drops to zero the
565 * register is freed and the region returns to default state.
566 * On success the register is returned, on failure a negative error
569 int mtrr_del(int reg
, unsigned long base
, unsigned long size
)
571 if (mtrr_check(base
, size
))
573 return mtrr_del_page(reg
, base
>> PAGE_SHIFT
, size
>> PAGE_SHIFT
);
575 EXPORT_SYMBOL(mtrr_del
);
579 * These should be called implicitly, but we can't yet until all the initcall
582 static void __init
init_ifs(void)
584 #ifndef CONFIG_X86_64
591 /* The suspend/resume methods are only for CPU without MTRR. CPU using generic
592 * MTRR driver doesn't require this
600 static struct mtrr_value mtrr_value
[MTRR_MAX_VAR_RANGES
];
602 static int mtrr_save(struct sys_device
*sysdev
, pm_message_t state
)
606 for (i
= 0; i
< num_var_ranges
; i
++) {
607 mtrr_if
->get(i
, &mtrr_value
[i
].lbase
,
608 &mtrr_value
[i
].lsize
,
609 &mtrr_value
[i
].ltype
);
614 static int mtrr_restore(struct sys_device
*sysdev
)
618 for (i
= 0; i
< num_var_ranges
; i
++) {
619 if (mtrr_value
[i
].lsize
) {
620 set_mtrr(i
, mtrr_value
[i
].lbase
,
622 mtrr_value
[i
].ltype
);
630 static struct sysdev_driver mtrr_sysdev_driver
= {
631 .suspend
= mtrr_save
,
632 .resume
= mtrr_restore
,
635 int __initdata changed_by_mtrr_cleanup
;
638 * mtrr_bp_init - initialize mtrrs on the boot CPU
640 * This needs to be called early; before any of the other CPUs are
641 * initialized (i.e. before smp_init()).
644 void __init
mtrr_bp_init(void)
653 mtrr_if
= &generic_mtrr_ops
;
654 size_or_mask
= 0xff000000; /* 36 bits */
655 size_and_mask
= 0x00f00000;
659 * This is an AMD specific MSR, but we assume(hope?) that
660 * Intel will implement it to when they extend the address
663 if (cpuid_eax(0x80000000) >= 0x80000008) {
664 phys_addr
= cpuid_eax(0x80000008) & 0xff;
665 /* CPUID workaround for Intel 0F33/0F34 CPU */
666 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_INTEL
&&
667 boot_cpu_data
.x86
== 0xF &&
668 boot_cpu_data
.x86_model
== 0x3 &&
669 (boot_cpu_data
.x86_mask
== 0x3 ||
670 boot_cpu_data
.x86_mask
== 0x4))
673 size_or_mask
= ~((1ULL << (phys_addr
- PAGE_SHIFT
)) - 1);
674 size_and_mask
= ~size_or_mask
& 0xfffff00000ULL
;
675 } else if (boot_cpu_data
.x86_vendor
== X86_VENDOR_CENTAUR
&&
676 boot_cpu_data
.x86
== 6) {
678 * VIA C* family have Intel style MTRRs,
679 * but don't support PAE
681 size_or_mask
= 0xfff00000; /* 32 bits */
686 switch (boot_cpu_data
.x86_vendor
) {
688 if (cpu_has_k6_mtrr
) {
689 /* Pre-Athlon (K6) AMD CPU MTRRs */
690 mtrr_if
= mtrr_ops
[X86_VENDOR_AMD
];
691 size_or_mask
= 0xfff00000; /* 32 bits */
695 case X86_VENDOR_CENTAUR
:
696 if (cpu_has_centaur_mcr
) {
697 mtrr_if
= mtrr_ops
[X86_VENDOR_CENTAUR
];
698 size_or_mask
= 0xfff00000; /* 32 bits */
702 case X86_VENDOR_CYRIX
:
703 if (cpu_has_cyrix_arr
) {
704 mtrr_if
= mtrr_ops
[X86_VENDOR_CYRIX
];
705 size_or_mask
= 0xfff00000; /* 32 bits */
715 set_num_var_ranges();
720 if (mtrr_cleanup(phys_addr
)) {
721 changed_by_mtrr_cleanup
= 1;
728 void mtrr_ap_init(void)
730 if (!use_intel() || mtrr_aps_delayed_init
)
733 * Ideally we should hold mtrr_mutex here to avoid mtrr entries
734 * changed, but this routine will be called in cpu boot time,
735 * holding the lock breaks it.
737 * This routine is called in two cases:
739 * 1. very earily time of software resume, when there absolutely
740 * isn't mtrr entry changes;
742 * 2. cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug
743 * lock to prevent mtrr entry changes
745 set_mtrr(~0U, 0, 0, 0);
749 * Save current fixed-range MTRR state of the BSP
751 void mtrr_save_state(void)
753 smp_call_function_single(0, mtrr_save_fixed_ranges
, NULL
, 1);
756 void set_mtrr_aps_delayed_init(void)
761 mtrr_aps_delayed_init
= true;
765 * MTRR initialization for all AP's
767 void mtrr_aps_init(void)
772 set_mtrr(~0U, 0, 0, 0);
773 mtrr_aps_delayed_init
= false;
776 void mtrr_bp_restore(void)
784 static int __init
mtrr_init_finialize(void)
790 if (!changed_by_mtrr_cleanup
)
796 * The CPU has no MTRR and seems to not support SMP. They have
797 * specific drivers, we use a tricky method to support
798 * suspend/resume for them.
800 * TBD: is there any system with such CPU which supports
801 * suspend/resume? If no, we should remove the code.
803 sysdev_driver_register(&cpu_sysdev_class
, &mtrr_sysdev_driver
);
807 subsys_initcall(mtrr_init_finialize
);