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>
46 #include <asm/kvm_para.h>
49 u32 num_var_ranges = 0;
51 unsigned int mtrr_usage_table[MAX_VAR_RANGES];
52 static DEFINE_MUTEX(mtrr_mutex);
54 u64 size_or_mask, size_and_mask;
56 static struct mtrr_ops * mtrr_ops[X86_VENDOR_NUM] = {};
58 struct mtrr_ops * mtrr_if = NULL;
60 static void set_mtrr(unsigned int reg, unsigned long base,
61 unsigned long size, mtrr_type type);
63 void set_mtrr_ops(struct mtrr_ops * ops)
65 if (ops->vendor && ops->vendor < X86_VENDOR_NUM)
66 mtrr_ops[ops->vendor] = ops;
69 /* Returns non-zero if we have the write-combining memory type */
70 static int have_wrcomb(void)
75 if ((dev = pci_get_class(PCI_CLASS_BRIDGE_HOST << 8, NULL)) != NULL) {
76 /* ServerWorks LE chipsets < rev 6 have problems with write-combining
77 Don't allow it and leave room for other chipsets to be tagged */
78 if (dev->vendor == PCI_VENDOR_ID_SERVERWORKS &&
79 dev->device == PCI_DEVICE_ID_SERVERWORKS_LE) {
80 pci_read_config_byte(dev, PCI_CLASS_REVISION, &rev);
82 printk(KERN_INFO "mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
87 /* Intel 450NX errata # 23. Non ascending cacheline evictions to
88 write combining memory may resulting in data corruption */
89 if (dev->vendor == PCI_VENDOR_ID_INTEL &&
90 dev->device == PCI_DEVICE_ID_INTEL_82451NX) {
91 printk(KERN_INFO "mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
97 return (mtrr_if->have_wrcomb ? mtrr_if->have_wrcomb() : 0);
100 /* This function returns the number of variable MTRRs */
101 static void __init set_num_var_ranges(void)
103 unsigned long config = 0, dummy;
106 rdmsr(MTRRcap_MSR, config, dummy);
107 } else if (is_cpu(AMD))
109 else if (is_cpu(CYRIX) || is_cpu(CENTAUR))
111 num_var_ranges = config & 0xff;
114 static void __init init_table(void)
118 max = num_var_ranges;
119 for (i = 0; i < max; i++)
120 mtrr_usage_table[i] = 1;
123 struct set_mtrr_data {
126 unsigned long smp_base;
127 unsigned long smp_size;
128 unsigned int smp_reg;
132 static void ipi_handler(void *info)
133 /* [SUMMARY] Synchronisation handler. Executed by "other" CPUs.
138 struct set_mtrr_data *data = info;
141 local_irq_save(flags);
143 atomic_dec(&data->count);
144 while(!atomic_read(&data->gate))
147 /* The master has cleared me to execute */
148 if (data->smp_reg != ~0U)
149 mtrr_if->set(data->smp_reg, data->smp_base,
150 data->smp_size, data->smp_type);
154 atomic_dec(&data->count);
155 while(atomic_read(&data->gate))
158 atomic_dec(&data->count);
159 local_irq_restore(flags);
163 static inline int types_compatible(mtrr_type type1, mtrr_type type2) {
164 return type1 == MTRR_TYPE_UNCACHABLE ||
165 type2 == MTRR_TYPE_UNCACHABLE ||
166 (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK) ||
167 (type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH);
171 * set_mtrr - update mtrrs on all processors
172 * @reg: mtrr in question
177 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
179 * 1. Send IPI to do the following:
180 * 2. Disable Interrupts
181 * 3. Wait for all procs to do so
182 * 4. Enter no-fill cache mode
186 * 8. Disable all range registers
187 * 9. Update the MTRRs
188 * 10. Enable all range registers
189 * 11. Flush all TLBs and caches again
190 * 12. Enter normal cache mode and reenable caching
192 * 14. Wait for buddies to catch up
193 * 15. Enable interrupts.
195 * What does that mean for us? Well, first we set data.count to the number
196 * of CPUs. As each CPU disables interrupts, it'll decrement it once. We wait
197 * until it hits 0 and proceed. We set the data.gate flag and reset data.count.
198 * Meanwhile, they are waiting for that flag to be set. Once it's set, each
199 * CPU goes through the transition of updating MTRRs. The CPU vendors may each do it
200 * differently, so we call mtrr_if->set() callback and let them take care of it.
201 * When they're done, they again decrement data->count and wait for data.gate to
203 * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag.
204 * Everyone then enables interrupts and we all continue on.
206 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
209 static void set_mtrr(unsigned int reg, unsigned long base,
210 unsigned long size, mtrr_type type)
212 struct set_mtrr_data data;
216 data.smp_base = base;
217 data.smp_size = size;
218 data.smp_type = type;
219 atomic_set(&data.count, num_booting_cpus() - 1);
220 /* make sure data.count is visible before unleashing other CPUs */
222 atomic_set(&data.gate,0);
224 /* Start the ball rolling on other CPUs */
225 if (smp_call_function(ipi_handler, &data, 1, 0) != 0)
226 panic("mtrr: timed out waiting for other CPUs\n");
228 local_irq_save(flags);
230 while(atomic_read(&data.count))
233 /* ok, reset count and toggle gate */
234 atomic_set(&data.count, num_booting_cpus() - 1);
236 atomic_set(&data.gate,1);
238 /* do our MTRR business */
241 * We use this same function to initialize the mtrrs on boot.
242 * The state of the boot cpu's mtrrs has been saved, and we want
243 * to replicate across all the APs.
244 * If we're doing that @reg is set to something special...
247 mtrr_if->set(reg,base,size,type);
249 /* wait for the others */
250 while(atomic_read(&data.count))
253 atomic_set(&data.count, num_booting_cpus() - 1);
255 atomic_set(&data.gate,0);
258 * Wait here for everyone to have seen the gate change
259 * So we're the last ones to touch 'data'
261 while(atomic_read(&data.count))
264 local_irq_restore(flags);
268 * mtrr_add_page - Add a memory type region
269 * @base: Physical base address of region in pages (in units of 4 kB!)
270 * @size: Physical size of region in pages (4 kB)
271 * @type: Type of MTRR desired
272 * @increment: If this is true do usage counting on the region
274 * Memory type region registers control the caching on newer Intel and
275 * non Intel processors. This function allows drivers to request an
276 * MTRR is added. The details and hardware specifics of each processor's
277 * implementation are hidden from the caller, but nevertheless the
278 * caller should expect to need to provide a power of two size on an
279 * equivalent power of two boundary.
281 * If the region cannot be added either because all regions are in use
282 * or the CPU cannot support it a negative value is returned. On success
283 * the register number for this entry is returned, but should be treated
286 * On a multiprocessor machine the changes are made to all processors.
287 * This is required on x86 by the Intel processors.
289 * The available types are
291 * %MTRR_TYPE_UNCACHABLE - No caching
293 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
295 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
297 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
299 * BUGS: Needs a quiet flag for the cases where drivers do not mind
300 * failures and do not wish system log messages to be sent.
303 int mtrr_add_page(unsigned long base, unsigned long size,
304 unsigned int type, bool increment)
306 int i, replace, error;
308 unsigned long lbase, lsize;
313 if ((error = mtrr_if->validate_add_page(base,size,type)))
316 if (type >= MTRR_NUM_TYPES) {
317 printk(KERN_WARNING "mtrr: type: %u invalid\n", type);
321 /* If the type is WC, check that this processor supports it */
322 if ((type == MTRR_TYPE_WRCOMB) && !have_wrcomb()) {
324 "mtrr: your processor doesn't support write-combining\n");
329 printk(KERN_WARNING "mtrr: zero sized request\n");
333 if (base & size_or_mask || size & size_or_mask) {
334 printk(KERN_WARNING "mtrr: base or size exceeds the MTRR width\n");
341 /* No CPU hotplug when we change MTRR entries */
343 /* Search for existing MTRR */
344 mutex_lock(&mtrr_mutex);
345 for (i = 0; i < num_var_ranges; ++i) {
346 mtrr_if->get(i, &lbase, &lsize, <ype);
347 if (!lsize || base > lbase + lsize - 1 || base + size - 1 < lbase)
349 /* At this point we know there is some kind of overlap/enclosure */
350 if (base < lbase || base + size - 1 > lbase + lsize - 1) {
351 if (base <= lbase && base + size - 1 >= lbase + lsize - 1) {
352 /* New region encloses an existing region */
354 replace = replace == -1 ? i : -2;
357 else if (types_compatible(type, ltype))
361 "mtrr: 0x%lx000,0x%lx000 overlaps existing"
362 " 0x%lx000,0x%lx000\n", base, size, lbase,
366 /* New region is enclosed by an existing region */
368 if (types_compatible(type, ltype))
370 printk (KERN_WARNING "mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
371 base, size, mtrr_attrib_to_str(ltype),
372 mtrr_attrib_to_str(type));
376 ++mtrr_usage_table[i];
380 /* Search for an empty MTRR */
381 i = mtrr_if->get_free_region(base, size, replace);
383 set_mtrr(i, base, size, type);
384 if (likely(replace < 0)) {
385 mtrr_usage_table[i] = 1;
387 mtrr_usage_table[i] = mtrr_usage_table[replace];
389 mtrr_usage_table[i]++;
390 if (unlikely(replace != i)) {
391 set_mtrr(replace, 0, 0, 0);
392 mtrr_usage_table[replace] = 0;
396 printk(KERN_INFO "mtrr: no more MTRRs available\n");
399 mutex_unlock(&mtrr_mutex);
404 static int mtrr_check(unsigned long base, unsigned long size)
406 if ((base & (PAGE_SIZE - 1)) || (size & (PAGE_SIZE - 1))) {
408 "mtrr: size and base must be multiples of 4 kiB\n");
410 "mtrr: size: 0x%lx base: 0x%lx\n", size, base);
418 * mtrr_add - Add a memory type region
419 * @base: Physical base address of region
420 * @size: Physical size of region
421 * @type: Type of MTRR desired
422 * @increment: If this is true do usage counting on the region
424 * Memory type region registers control the caching on newer Intel and
425 * non Intel processors. This function allows drivers to request an
426 * MTRR is added. The details and hardware specifics of each processor's
427 * implementation are hidden from the caller, but nevertheless the
428 * caller should expect to need to provide a power of two size on an
429 * equivalent power of two boundary.
431 * If the region cannot be added either because all regions are in use
432 * or the CPU cannot support it a negative value is returned. On success
433 * the register number for this entry is returned, but should be treated
436 * On a multiprocessor machine the changes are made to all processors.
437 * This is required on x86 by the Intel processors.
439 * The available types are
441 * %MTRR_TYPE_UNCACHABLE - No caching
443 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
445 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
447 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
449 * BUGS: Needs a quiet flag for the cases where drivers do not mind
450 * failures and do not wish system log messages to be sent.
454 mtrr_add(unsigned long base, unsigned long size, unsigned int type,
457 if (mtrr_check(base, size))
459 return mtrr_add_page(base >> PAGE_SHIFT, size >> PAGE_SHIFT, type,
464 * mtrr_del_page - delete a memory type region
465 * @reg: Register returned by mtrr_add
466 * @base: Physical base address
467 * @size: Size of region
469 * If register is supplied then base and size are ignored. This is
470 * how drivers should call it.
472 * Releases an MTRR region. If the usage count drops to zero the
473 * register is freed and the region returns to default state.
474 * On success the register is returned, on failure a negative error
478 int mtrr_del_page(int reg, unsigned long base, unsigned long size)
482 unsigned long lbase, lsize;
488 max = num_var_ranges;
489 /* No CPU hotplug when we change MTRR entries */
491 mutex_lock(&mtrr_mutex);
493 /* Search for existing MTRR */
494 for (i = 0; i < max; ++i) {
495 mtrr_if->get(i, &lbase, &lsize, <ype);
496 if (lbase == base && lsize == size) {
502 printk(KERN_DEBUG "mtrr: no MTRR for %lx000,%lx000 found\n", base,
508 printk(KERN_WARNING "mtrr: register: %d too big\n", reg);
511 mtrr_if->get(reg, &lbase, &lsize, <ype);
513 printk(KERN_WARNING "mtrr: MTRR %d not used\n", reg);
516 if (mtrr_usage_table[reg] < 1) {
517 printk(KERN_WARNING "mtrr: reg: %d has count=0\n", reg);
520 if (--mtrr_usage_table[reg] < 1)
521 set_mtrr(reg, 0, 0, 0);
524 mutex_unlock(&mtrr_mutex);
529 * mtrr_del - delete a memory type region
530 * @reg: Register returned by mtrr_add
531 * @base: Physical base address
532 * @size: Size of region
534 * If register is supplied then base and size are ignored. This is
535 * how drivers should call it.
537 * Releases an MTRR region. If the usage count drops to zero the
538 * register is freed and the region returns to default state.
539 * On success the register is returned, on failure a negative error
544 mtrr_del(int reg, unsigned long base, unsigned long size)
546 if (mtrr_check(base, size))
548 return mtrr_del_page(reg, base >> PAGE_SHIFT, size >> PAGE_SHIFT);
551 EXPORT_SYMBOL(mtrr_add);
552 EXPORT_SYMBOL(mtrr_del);
555 * These should be called implicitly, but we can't yet until all the initcall
558 static void __init init_ifs(void)
560 #ifndef CONFIG_X86_64
567 /* The suspend/resume methods are only for CPU without MTRR. CPU using generic
568 * MTRR driver doesn't require this
576 static struct mtrr_value mtrr_state[MAX_VAR_RANGES];
578 static int mtrr_save(struct sys_device * sysdev, pm_message_t state)
582 for (i = 0; i < num_var_ranges; i++) {
584 &mtrr_state[i].lbase,
585 &mtrr_state[i].lsize,
586 &mtrr_state[i].ltype);
591 static int mtrr_restore(struct sys_device * sysdev)
595 for (i = 0; i < num_var_ranges; i++) {
596 if (mtrr_state[i].lsize)
600 mtrr_state[i].ltype);
607 static struct sysdev_driver mtrr_sysdev_driver = {
608 .suspend = mtrr_save,
609 .resume = mtrr_restore,
612 static int disable_mtrr_trim;
614 static int __init disable_mtrr_trim_setup(char *str)
616 disable_mtrr_trim = 1;
619 early_param("disable_mtrr_trim", disable_mtrr_trim_setup);
622 * Newer AMD K8s and later CPUs have a special magic MSR way to force WB
623 * for memory >4GB. Check for that here.
624 * Note this won't check if the MTRRs < 4GB where the magic bit doesn't
625 * apply to are wrong, but so far we don't know of any such case in the wild.
627 #define Tom2Enabled (1U << 21)
628 #define Tom2ForceMemTypeWB (1U << 22)
630 static __init int amd_special_default_mtrr(void)
634 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
636 if (boot_cpu_data.x86 < 0xf || boot_cpu_data.x86 > 0x11)
638 /* In case some hypervisor doesn't pass SYSCFG through */
639 if (rdmsr_safe(MSR_K8_SYSCFG, &l, &h) < 0)
642 * Memory between 4GB and top of mem is forced WB by this magic bit.
643 * Reserved before K8RevF, but should be zero there.
645 if ((l & (Tom2Enabled | Tom2ForceMemTypeWB)) ==
646 (Tom2Enabled | Tom2ForceMemTypeWB))
652 * mtrr_trim_uncached_memory - trim RAM not covered by MTRRs
653 * @end_pfn: ending page frame number
655 * Some buggy BIOSes don't setup the MTRRs properly for systems with certain
656 * memory configurations. This routine checks that the highest MTRR matches
657 * the end of memory, to make sure the MTRRs having a write back type cover
658 * all of the memory the kernel is intending to use. If not, it'll trim any
659 * memory off the end by adjusting end_pfn, removing it from the kernel's
660 * allocation pools, warning the user with an obnoxious message.
662 int __init mtrr_trim_uncached_memory(unsigned long end_pfn)
664 unsigned long i, base, size, highest_pfn = 0, def, dummy;
666 u64 trim_start, trim_size;
669 * Make sure we only trim uncachable memory on machines that
670 * support the Intel MTRR architecture:
672 if (!is_cpu(INTEL) || disable_mtrr_trim)
674 rdmsr(MTRRdefType_MSR, def, dummy);
676 if (def != MTRR_TYPE_UNCACHABLE)
679 if (amd_special_default_mtrr())
682 /* Find highest cached pfn */
683 for (i = 0; i < num_var_ranges; i++) {
684 mtrr_if->get(i, &base, &size, &type);
685 if (type != MTRR_TYPE_WRBACK)
687 if (highest_pfn < base + size)
688 highest_pfn = base + size;
691 /* kvm/qemu doesn't have mtrr set right, don't trim them all */
693 if (!kvm_para_available()) {
695 "WARNING: strange, CPU MTRRs all blank?\n");
701 if (highest_pfn < end_pfn) {
702 printk(KERN_WARNING "WARNING: BIOS bug: CPU MTRRs don't cover"
703 " all of memory, losing %luMB of RAM.\n",
704 (end_pfn - highest_pfn) >> (20 - PAGE_SHIFT));
708 printk(KERN_INFO "update e820 for mtrr\n");
709 trim_start = highest_pfn;
710 trim_start <<= PAGE_SHIFT;
712 trim_size <<= PAGE_SHIFT;
713 trim_size -= trim_start;
714 add_memory_region(trim_start, trim_size, E820_RESERVED);
723 * mtrr_bp_init - initialize mtrrs on the boot CPU
725 * This needs to be called early; before any of the other CPUs are
726 * initialized (i.e. before smp_init()).
729 void __init mtrr_bp_init(void)
734 mtrr_if = &generic_mtrr_ops;
735 size_or_mask = 0xff000000; /* 36 bits */
736 size_and_mask = 0x00f00000;
738 /* This is an AMD specific MSR, but we assume(hope?) that
739 Intel will implement it to when they extend the address
741 if (cpuid_eax(0x80000000) >= 0x80000008) {
743 phys_addr = cpuid_eax(0x80000008) & 0xff;
744 /* CPUID workaround for Intel 0F33/0F34 CPU */
745 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
746 boot_cpu_data.x86 == 0xF &&
747 boot_cpu_data.x86_model == 0x3 &&
748 (boot_cpu_data.x86_mask == 0x3 ||
749 boot_cpu_data.x86_mask == 0x4))
752 size_or_mask = ~((1ULL << (phys_addr - PAGE_SHIFT)) - 1);
753 size_and_mask = ~size_or_mask & 0xfffff00000ULL;
754 } else if (boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR &&
755 boot_cpu_data.x86 == 6) {
756 /* VIA C* family have Intel style MTRRs, but
758 size_or_mask = 0xfff00000; /* 32 bits */
762 switch (boot_cpu_data.x86_vendor) {
764 if (cpu_has_k6_mtrr) {
765 /* Pre-Athlon (K6) AMD CPU MTRRs */
766 mtrr_if = mtrr_ops[X86_VENDOR_AMD];
767 size_or_mask = 0xfff00000; /* 32 bits */
771 case X86_VENDOR_CENTAUR:
772 if (cpu_has_centaur_mcr) {
773 mtrr_if = mtrr_ops[X86_VENDOR_CENTAUR];
774 size_or_mask = 0xfff00000; /* 32 bits */
778 case X86_VENDOR_CYRIX:
779 if (cpu_has_cyrix_arr) {
780 mtrr_if = mtrr_ops[X86_VENDOR_CYRIX];
781 size_or_mask = 0xfff00000; /* 32 bits */
791 set_num_var_ranges();
798 void mtrr_ap_init(void)
802 if (!mtrr_if || !use_intel())
805 * Ideally we should hold mtrr_mutex here to avoid mtrr entries changed,
806 * but this routine will be called in cpu boot time, holding the lock
807 * breaks it. This routine is called in two cases: 1.very earily time
808 * of software resume, when there absolutely isn't mtrr entry changes;
809 * 2.cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug lock to
810 * prevent mtrr entry changes
812 local_irq_save(flags);
816 local_irq_restore(flags);
820 * Save current fixed-range MTRR state of the BSP
822 void mtrr_save_state(void)
824 smp_call_function_single(0, mtrr_save_fixed_ranges, NULL, 1, 1);
827 static int __init mtrr_init_finialize(void)
834 /* The CPUs haven't MTRR and seem to not support SMP. They have
835 * specific drivers, we use a tricky method to support
836 * suspend/resume for them.
837 * TBD: is there any system with such CPU which supports
838 * suspend/resume? if no, we should remove the code.
840 sysdev_driver_register(&cpu_sysdev_class,
841 &mtrr_sysdev_driver);
845 subsys_initcall(mtrr_init_finialize);