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).
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/pci.h>
37 #include <linux/smp.h>
38 #include <linux/cpu.h>
39 #include <linux/mutex.h>
43 #include <asm/uaccess.h>
44 #include <asm/processor.h>
48 u32 num_var_ranges = 0;
50 unsigned int mtrr_usage_table[MAX_VAR_RANGES];
51 static DEFINE_MUTEX(mtrr_mutex);
53 u64 size_or_mask, size_and_mask;
55 static struct mtrr_ops * mtrr_ops[X86_VENDOR_NUM] = {};
57 struct mtrr_ops * mtrr_if = NULL;
59 static void set_mtrr(unsigned int reg, unsigned long base,
60 unsigned long size, mtrr_type type);
62 void set_mtrr_ops(struct mtrr_ops * ops)
64 if (ops->vendor && ops->vendor < X86_VENDOR_NUM)
65 mtrr_ops[ops->vendor] = ops;
68 /* Returns non-zero if we have the write-combining memory type */
69 static int have_wrcomb(void)
74 if ((dev = pci_get_class(PCI_CLASS_BRIDGE_HOST << 8, NULL)) != NULL) {
75 /* ServerWorks LE chipsets < rev 6 have problems with write-combining
76 Don't allow it and leave room for other chipsets to be tagged */
77 if (dev->vendor == PCI_VENDOR_ID_SERVERWORKS &&
78 dev->device == PCI_DEVICE_ID_SERVERWORKS_LE) {
79 pci_read_config_byte(dev, PCI_CLASS_REVISION, &rev);
81 printk(KERN_INFO "mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
86 /* Intel 450NX errata # 23. Non ascending cacheline evictions to
87 write combining memory may resulting in data corruption */
88 if (dev->vendor == PCI_VENDOR_ID_INTEL &&
89 dev->device == PCI_DEVICE_ID_INTEL_82451NX) {
90 printk(KERN_INFO "mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
96 return (mtrr_if->have_wrcomb ? mtrr_if->have_wrcomb() : 0);
99 /* This function returns the number of variable MTRRs */
100 static void __init set_num_var_ranges(void)
102 unsigned long config = 0, dummy;
105 rdmsr(MTRRcap_MSR, config, dummy);
106 } else if (is_cpu(AMD))
108 else if (is_cpu(CYRIX) || is_cpu(CENTAUR))
110 num_var_ranges = config & 0xff;
113 static void __init init_table(void)
117 max = num_var_ranges;
118 for (i = 0; i < max; i++)
119 mtrr_usage_table[i] = 1;
122 struct set_mtrr_data {
125 unsigned long smp_base;
126 unsigned long smp_size;
127 unsigned int smp_reg;
131 static void ipi_handler(void *info)
132 /* [SUMMARY] Synchronisation handler. Executed by "other" CPUs.
137 struct set_mtrr_data *data = info;
140 local_irq_save(flags);
142 atomic_dec(&data->count);
143 while(!atomic_read(&data->gate))
146 /* The master has cleared me to execute */
147 if (data->smp_reg != ~0U)
148 mtrr_if->set(data->smp_reg, data->smp_base,
149 data->smp_size, data->smp_type);
153 atomic_dec(&data->count);
154 while(atomic_read(&data->gate))
157 atomic_dec(&data->count);
158 local_irq_restore(flags);
162 static inline int types_compatible(mtrr_type type1, mtrr_type type2) {
163 return type1 == MTRR_TYPE_UNCACHABLE ||
164 type2 == MTRR_TYPE_UNCACHABLE ||
165 (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK) ||
166 (type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH);
170 * set_mtrr - update mtrrs on all processors
171 * @reg: mtrr in question
176 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
178 * 1. Send IPI to do the following:
179 * 2. Disable Interrupts
180 * 3. Wait for all procs to do so
181 * 4. Enter no-fill cache mode
185 * 8. Disable all range registers
186 * 9. Update the MTRRs
187 * 10. Enable all range registers
188 * 11. Flush all TLBs and caches again
189 * 12. Enter normal cache mode and reenable caching
191 * 14. Wait for buddies to catch up
192 * 15. Enable interrupts.
194 * What does that mean for us? Well, first we set data.count to the number
195 * of CPUs. As each CPU disables interrupts, it'll decrement it once. We wait
196 * until it hits 0 and proceed. We set the data.gate flag and reset data.count.
197 * Meanwhile, they are waiting for that flag to be set. Once it's set, each
198 * CPU goes through the transition of updating MTRRs. The CPU vendors may each do it
199 * differently, so we call mtrr_if->set() callback and let them take care of it.
200 * When they're done, they again decrement data->count and wait for data.gate to
202 * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag.
203 * Everyone then enables interrupts and we all continue on.
205 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
208 static void set_mtrr(unsigned int reg, unsigned long base,
209 unsigned long size, mtrr_type type)
211 struct set_mtrr_data data;
215 data.smp_base = base;
216 data.smp_size = size;
217 data.smp_type = type;
218 atomic_set(&data.count, num_booting_cpus() - 1);
219 /* make sure data.count is visible before unleashing other CPUs */
221 atomic_set(&data.gate,0);
223 /* Start the ball rolling on other CPUs */
224 if (smp_call_function(ipi_handler, &data, 1, 0) != 0)
225 panic("mtrr: timed out waiting for other CPUs\n");
227 local_irq_save(flags);
229 while(atomic_read(&data.count))
232 /* ok, reset count and toggle gate */
233 atomic_set(&data.count, num_booting_cpus() - 1);
235 atomic_set(&data.gate,1);
237 /* do our MTRR business */
240 * We use this same function to initialize the mtrrs on boot.
241 * The state of the boot cpu's mtrrs has been saved, and we want
242 * to replicate across all the APs.
243 * If we're doing that @reg is set to something special...
246 mtrr_if->set(reg,base,size,type);
248 /* wait for the others */
249 while(atomic_read(&data.count))
252 atomic_set(&data.count, num_booting_cpus() - 1);
254 atomic_set(&data.gate,0);
257 * Wait here for everyone to have seen the gate change
258 * So we're the last ones to touch 'data'
260 while(atomic_read(&data.count))
263 local_irq_restore(flags);
267 * mtrr_add_page - Add a memory type region
268 * @base: Physical base address of region in pages (in units of 4 kB!)
269 * @size: Physical size of region in pages (4 kB)
270 * @type: Type of MTRR desired
271 * @increment: If this is true do usage counting on the region
273 * Memory type region registers control the caching on newer Intel and
274 * non Intel processors. This function allows drivers to request an
275 * MTRR is added. The details and hardware specifics of each processor's
276 * implementation are hidden from the caller, but nevertheless the
277 * caller should expect to need to provide a power of two size on an
278 * equivalent power of two boundary.
280 * If the region cannot be added either because all regions are in use
281 * or the CPU cannot support it a negative value is returned. On success
282 * the register number for this entry is returned, but should be treated
285 * On a multiprocessor machine the changes are made to all processors.
286 * This is required on x86 by the Intel processors.
288 * The available types are
290 * %MTRR_TYPE_UNCACHABLE - No caching
292 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
294 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
296 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
298 * BUGS: Needs a quiet flag for the cases where drivers do not mind
299 * failures and do not wish system log messages to be sent.
302 int mtrr_add_page(unsigned long base, unsigned long size,
303 unsigned int type, bool increment)
305 int i, replace, error;
307 unsigned long lbase, lsize;
312 if ((error = mtrr_if->validate_add_page(base,size,type)))
315 if (type >= MTRR_NUM_TYPES) {
316 printk(KERN_WARNING "mtrr: type: %u invalid\n", type);
320 /* If the type is WC, check that this processor supports it */
321 if ((type == MTRR_TYPE_WRCOMB) && !have_wrcomb()) {
323 "mtrr: your processor doesn't support write-combining\n");
328 printk(KERN_WARNING "mtrr: zero sized request\n");
332 if (base & size_or_mask || size & size_or_mask) {
333 printk(KERN_WARNING "mtrr: base or size exceeds the MTRR width\n");
340 /* No CPU hotplug when we change MTRR entries */
342 /* Search for existing MTRR */
343 mutex_lock(&mtrr_mutex);
344 for (i = 0; i < num_var_ranges; ++i) {
345 mtrr_if->get(i, &lbase, &lsize, <ype);
346 if (!lsize || base > lbase + lsize - 1 || base + size - 1 < lbase)
348 /* At this point we know there is some kind of overlap/enclosure */
349 if (base < lbase || base + size - 1 > lbase + lsize - 1) {
350 if (base <= lbase && base + size - 1 >= lbase + lsize - 1) {
351 /* New region encloses an existing region */
353 replace = replace == -1 ? i : -2;
356 else if (types_compatible(type, ltype))
360 "mtrr: 0x%lx000,0x%lx000 overlaps existing"
361 " 0x%lx000,0x%lx000\n", base, size, lbase,
365 /* New region is enclosed by an existing region */
367 if (types_compatible(type, ltype))
369 printk (KERN_WARNING "mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
370 base, size, mtrr_attrib_to_str(ltype),
371 mtrr_attrib_to_str(type));
375 ++mtrr_usage_table[i];
379 /* Search for an empty MTRR */
380 i = mtrr_if->get_free_region(base, size, replace);
382 set_mtrr(i, base, size, type);
383 if (likely(replace < 0)) {
384 mtrr_usage_table[i] = 1;
386 mtrr_usage_table[i] = mtrr_usage_table[replace];
388 mtrr_usage_table[i]++;
389 if (unlikely(replace != i)) {
390 set_mtrr(replace, 0, 0, 0);
391 mtrr_usage_table[replace] = 0;
395 printk(KERN_INFO "mtrr: no more MTRRs available\n");
398 mutex_unlock(&mtrr_mutex);
403 static int mtrr_check(unsigned long base, unsigned long size)
405 if ((base & (PAGE_SIZE - 1)) || (size & (PAGE_SIZE - 1))) {
407 "mtrr: size and base must be multiples of 4 kiB\n");
409 "mtrr: size: 0x%lx base: 0x%lx\n", size, base);
417 * mtrr_add - Add a memory type region
418 * @base: Physical base address of region
419 * @size: Physical size of region
420 * @type: Type of MTRR desired
421 * @increment: If this is true do usage counting on the region
423 * Memory type region registers control the caching on newer Intel and
424 * non Intel processors. This function allows drivers to request an
425 * MTRR is added. The details and hardware specifics of each processor's
426 * implementation are hidden from the caller, but nevertheless the
427 * caller should expect to need to provide a power of two size on an
428 * equivalent power of two boundary.
430 * If the region cannot be added either because all regions are in use
431 * or the CPU cannot support it a negative value is returned. On success
432 * the register number for this entry is returned, but should be treated
435 * On a multiprocessor machine the changes are made to all processors.
436 * This is required on x86 by the Intel processors.
438 * The available types are
440 * %MTRR_TYPE_UNCACHABLE - No caching
442 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
444 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
446 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
448 * BUGS: Needs a quiet flag for the cases where drivers do not mind
449 * failures and do not wish system log messages to be sent.
453 mtrr_add(unsigned long base, unsigned long size, unsigned int type,
456 if (mtrr_check(base, size))
458 return mtrr_add_page(base >> PAGE_SHIFT, size >> PAGE_SHIFT, type,
463 * mtrr_del_page - delete a memory type region
464 * @reg: Register returned by mtrr_add
465 * @base: Physical base address
466 * @size: Size of region
468 * If register is supplied then base and size are ignored. This is
469 * how drivers should call it.
471 * Releases an MTRR region. If the usage count drops to zero the
472 * register is freed and the region returns to default state.
473 * On success the register is returned, on failure a negative error
477 int mtrr_del_page(int reg, unsigned long base, unsigned long size)
481 unsigned long lbase, lsize;
487 max = num_var_ranges;
488 /* No CPU hotplug when we change MTRR entries */
490 mutex_lock(&mtrr_mutex);
492 /* Search for existing MTRR */
493 for (i = 0; i < max; ++i) {
494 mtrr_if->get(i, &lbase, &lsize, <ype);
495 if (lbase == base && lsize == size) {
501 printk(KERN_DEBUG "mtrr: no MTRR for %lx000,%lx000 found\n", base,
507 printk(KERN_WARNING "mtrr: register: %d too big\n", reg);
510 mtrr_if->get(reg, &lbase, &lsize, <ype);
512 printk(KERN_WARNING "mtrr: MTRR %d not used\n", reg);
515 if (mtrr_usage_table[reg] < 1) {
516 printk(KERN_WARNING "mtrr: reg: %d has count=0\n", reg);
519 if (--mtrr_usage_table[reg] < 1)
520 set_mtrr(reg, 0, 0, 0);
523 mutex_unlock(&mtrr_mutex);
528 * mtrr_del - delete a memory type region
529 * @reg: Register returned by mtrr_add
530 * @base: Physical base address
531 * @size: Size of region
533 * If register is supplied then base and size are ignored. This is
534 * how drivers should call it.
536 * Releases an MTRR region. If the usage count drops to zero the
537 * register is freed and the region returns to default state.
538 * On success the register is returned, on failure a negative error
543 mtrr_del(int reg, unsigned long base, unsigned long size)
545 if (mtrr_check(base, size))
547 return mtrr_del_page(reg, base >> PAGE_SHIFT, size >> PAGE_SHIFT);
550 EXPORT_SYMBOL(mtrr_add);
551 EXPORT_SYMBOL(mtrr_del);
554 * These should be called implicitly, but we can't yet until all the initcall
557 static void __init init_ifs(void)
559 #ifndef CONFIG_X86_64
566 /* The suspend/resume methods are only for CPU without MTRR. CPU using generic
567 * MTRR driver doesn't require this
575 static struct mtrr_value mtrr_state[MAX_VAR_RANGES];
577 static int mtrr_save(struct sys_device * sysdev, pm_message_t state)
581 for (i = 0; i < num_var_ranges; i++) {
583 &mtrr_state[i].lbase,
584 &mtrr_state[i].lsize,
585 &mtrr_state[i].ltype);
590 static int mtrr_restore(struct sys_device * sysdev)
594 for (i = 0; i < num_var_ranges; i++) {
595 if (mtrr_state[i].lsize)
599 mtrr_state[i].ltype);
606 static struct sysdev_driver mtrr_sysdev_driver = {
607 .suspend = mtrr_save,
608 .resume = mtrr_restore,
611 static int disable_mtrr_trim;
613 static int __init disable_mtrr_trim_setup(char *str)
615 disable_mtrr_trim = 1;
618 early_param("disable_mtrr_trim", disable_mtrr_trim_setup);
621 * Newer AMD K8s and later CPUs have a special magic MSR way to force WB
622 * for memory >4GB. Check for that here.
623 * Note this won't check if the MTRRs < 4GB where the magic bit doesn't
624 * apply to are wrong, but so far we don't know of any such case in the wild.
626 #define Tom2Enabled (1U << 21)
627 #define Tom2ForceMemTypeWB (1U << 22)
629 static __init int amd_special_default_mtrr(void)
633 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
635 if (boot_cpu_data.x86 < 0xf || boot_cpu_data.x86 > 0x11)
637 /* In case some hypervisor doesn't pass SYSCFG through */
638 if (rdmsr_safe(MSR_K8_SYSCFG, &l, &h) < 0)
641 * Memory between 4GB and top of mem is forced WB by this magic bit.
642 * Reserved before K8RevF, but should be zero there.
644 if ((l & (Tom2Enabled | Tom2ForceMemTypeWB)) ==
645 (Tom2Enabled | Tom2ForceMemTypeWB))
651 * mtrr_trim_uncached_memory - trim RAM not covered by MTRRs
653 * Some buggy BIOSes don't setup the MTRRs properly for systems with certain
654 * memory configurations. This routine checks that the highest MTRR matches
655 * the end of memory, to make sure the MTRRs having a write back type cover
656 * all of the memory the kernel is intending to use. If not, it'll trim any
657 * memory off the end by adjusting end_pfn, removing it from the kernel's
658 * allocation pools, warning the user with an obnoxious message.
660 int __init mtrr_trim_uncached_memory(unsigned long end_pfn)
662 unsigned long i, base, size, highest_pfn = 0, def, dummy;
664 u64 trim_start, trim_size;
667 * Make sure we only trim uncachable memory on machines that
668 * support the Intel MTRR architecture:
670 if (!is_cpu(INTEL) || disable_mtrr_trim)
672 rdmsr(MTRRdefType_MSR, def, dummy);
674 if (def != MTRR_TYPE_UNCACHABLE)
677 if (amd_special_default_mtrr())
680 /* Find highest cached pfn */
681 for (i = 0; i < num_var_ranges; i++) {
682 mtrr_if->get(i, &base, &size, &type);
683 if (type != MTRR_TYPE_WRBACK)
685 if (highest_pfn < base + size)
686 highest_pfn = base + size;
689 /* kvm/qemu doesn't have mtrr set right, don't trim them all */
691 printk(KERN_WARNING "WARNING: strange, CPU MTRRs all blank?\n");
696 if (highest_pfn < end_pfn) {
697 printk(KERN_WARNING "WARNING: BIOS bug: CPU MTRRs don't cover"
698 " all of memory, losing %luMB of RAM.\n",
699 (end_pfn - highest_pfn) >> (20 - PAGE_SHIFT));
703 printk(KERN_INFO "update e820 for mtrr\n");
704 trim_start = highest_pfn;
705 trim_start <<= PAGE_SHIFT;
707 trim_size <<= PAGE_SHIFT;
708 trim_size -= trim_start;
709 add_memory_region(trim_start, trim_size, E820_RESERVED);
718 * mtrr_bp_init - initialize mtrrs on the boot CPU
720 * This needs to be called early; before any of the other CPUs are
721 * initialized (i.e. before smp_init()).
724 void __init mtrr_bp_init(void)
729 mtrr_if = &generic_mtrr_ops;
730 size_or_mask = 0xff000000; /* 36 bits */
731 size_and_mask = 0x00f00000;
733 /* This is an AMD specific MSR, but we assume(hope?) that
734 Intel will implement it to when they extend the address
736 if (cpuid_eax(0x80000000) >= 0x80000008) {
738 phys_addr = cpuid_eax(0x80000008) & 0xff;
739 /* CPUID workaround for Intel 0F33/0F34 CPU */
740 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
741 boot_cpu_data.x86 == 0xF &&
742 boot_cpu_data.x86_model == 0x3 &&
743 (boot_cpu_data.x86_mask == 0x3 ||
744 boot_cpu_data.x86_mask == 0x4))
747 size_or_mask = ~((1ULL << (phys_addr - PAGE_SHIFT)) - 1);
748 size_and_mask = ~size_or_mask & 0xfffff00000ULL;
749 } else if (boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR &&
750 boot_cpu_data.x86 == 6) {
751 /* VIA C* family have Intel style MTRRs, but
753 size_or_mask = 0xfff00000; /* 32 bits */
757 switch (boot_cpu_data.x86_vendor) {
759 if (cpu_has_k6_mtrr) {
760 /* Pre-Athlon (K6) AMD CPU MTRRs */
761 mtrr_if = mtrr_ops[X86_VENDOR_AMD];
762 size_or_mask = 0xfff00000; /* 32 bits */
766 case X86_VENDOR_CENTAUR:
767 if (cpu_has_centaur_mcr) {
768 mtrr_if = mtrr_ops[X86_VENDOR_CENTAUR];
769 size_or_mask = 0xfff00000; /* 32 bits */
773 case X86_VENDOR_CYRIX:
774 if (cpu_has_cyrix_arr) {
775 mtrr_if = mtrr_ops[X86_VENDOR_CYRIX];
776 size_or_mask = 0xfff00000; /* 32 bits */
786 set_num_var_ranges();
793 void mtrr_ap_init(void)
797 if (!mtrr_if || !use_intel())
800 * Ideally we should hold mtrr_mutex here to avoid mtrr entries changed,
801 * but this routine will be called in cpu boot time, holding the lock
802 * breaks it. This routine is called in two cases: 1.very earily time
803 * of software resume, when there absolutely isn't mtrr entry changes;
804 * 2.cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug lock to
805 * prevent mtrr entry changes
807 local_irq_save(flags);
811 local_irq_restore(flags);
815 * Save current fixed-range MTRR state of the BSP
817 void mtrr_save_state(void)
819 smp_call_function_single(0, mtrr_save_fixed_ranges, NULL, 1, 1);
822 static int __init mtrr_init_finialize(void)
829 /* The CPUs haven't MTRR and seem to not support SMP. They have
830 * specific drivers, we use a tricky method to support
831 * suspend/resume for them.
832 * TBD: is there any system with such CPU which supports
833 * suspend/resume? if no, we should remove the code.
835 sysdev_driver_register(&cpu_sysdev_class,
836 &mtrr_sysdev_driver);
840 subsys_initcall(mtrr_init_finialize);