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);
63 extern int arr3_protected;
65 #define arr3_protected 0
68 void set_mtrr_ops(struct mtrr_ops * ops)
70 if (ops->vendor && ops->vendor < X86_VENDOR_NUM)
71 mtrr_ops[ops->vendor] = ops;
74 /* Returns non-zero if we have the write-combining memory type */
75 static int have_wrcomb(void)
80 if ((dev = pci_get_class(PCI_CLASS_BRIDGE_HOST << 8, NULL)) != NULL) {
81 /* ServerWorks LE chipsets < rev 6 have problems with write-combining
82 Don't allow it and leave room for other chipsets to be tagged */
83 if (dev->vendor == PCI_VENDOR_ID_SERVERWORKS &&
84 dev->device == PCI_DEVICE_ID_SERVERWORKS_LE) {
85 pci_read_config_byte(dev, PCI_CLASS_REVISION, &rev);
87 printk(KERN_INFO "mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
92 /* Intel 450NX errata # 23. Non ascending cacheline evictions to
93 write combining memory may resulting in data corruption */
94 if (dev->vendor == PCI_VENDOR_ID_INTEL &&
95 dev->device == PCI_DEVICE_ID_INTEL_82451NX) {
96 printk(KERN_INFO "mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
102 return (mtrr_if->have_wrcomb ? mtrr_if->have_wrcomb() : 0);
105 /* This function returns the number of variable MTRRs */
106 static void __init set_num_var_ranges(void)
108 unsigned long config = 0, dummy;
111 rdmsr(MTRRcap_MSR, config, dummy);
112 } else if (is_cpu(AMD))
114 else if (is_cpu(CYRIX) || is_cpu(CENTAUR))
116 num_var_ranges = config & 0xff;
119 static void __init init_table(void)
123 max = num_var_ranges;
124 for (i = 0; i < max; i++)
125 mtrr_usage_table[i] = 1;
128 struct set_mtrr_data {
131 unsigned long smp_base;
132 unsigned long smp_size;
133 unsigned int smp_reg;
137 static void ipi_handler(void *info)
138 /* [SUMMARY] Synchronisation handler. Executed by "other" CPUs.
143 struct set_mtrr_data *data = info;
146 local_irq_save(flags);
148 atomic_dec(&data->count);
149 while(!atomic_read(&data->gate))
152 /* The master has cleared me to execute */
153 if (data->smp_reg != ~0U)
154 mtrr_if->set(data->smp_reg, data->smp_base,
155 data->smp_size, data->smp_type);
159 atomic_dec(&data->count);
160 while(atomic_read(&data->gate))
163 atomic_dec(&data->count);
164 local_irq_restore(flags);
168 static inline int types_compatible(mtrr_type type1, mtrr_type type2) {
169 return type1 == MTRR_TYPE_UNCACHABLE ||
170 type2 == MTRR_TYPE_UNCACHABLE ||
171 (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK) ||
172 (type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH);
176 * set_mtrr - update mtrrs on all processors
177 * @reg: mtrr in question
182 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
184 * 1. Send IPI to do the following:
185 * 2. Disable Interrupts
186 * 3. Wait for all procs to do so
187 * 4. Enter no-fill cache mode
191 * 8. Disable all range registers
192 * 9. Update the MTRRs
193 * 10. Enable all range registers
194 * 11. Flush all TLBs and caches again
195 * 12. Enter normal cache mode and reenable caching
197 * 14. Wait for buddies to catch up
198 * 15. Enable interrupts.
200 * What does that mean for us? Well, first we set data.count to the number
201 * of CPUs. As each CPU disables interrupts, it'll decrement it once. We wait
202 * until it hits 0 and proceed. We set the data.gate flag and reset data.count.
203 * Meanwhile, they are waiting for that flag to be set. Once it's set, each
204 * CPU goes through the transition of updating MTRRs. The CPU vendors may each do it
205 * differently, so we call mtrr_if->set() callback and let them take care of it.
206 * When they're done, they again decrement data->count and wait for data.gate to
208 * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag.
209 * Everyone then enables interrupts and we all continue on.
211 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
214 static void set_mtrr(unsigned int reg, unsigned long base,
215 unsigned long size, mtrr_type type)
217 struct set_mtrr_data data;
221 data.smp_base = base;
222 data.smp_size = size;
223 data.smp_type = type;
224 atomic_set(&data.count, num_booting_cpus() - 1);
225 /* make sure data.count is visible before unleashing other CPUs */
227 atomic_set(&data.gate,0);
229 /* Start the ball rolling on other CPUs */
230 if (smp_call_function(ipi_handler, &data, 1, 0) != 0)
231 panic("mtrr: timed out waiting for other CPUs\n");
233 local_irq_save(flags);
235 while(atomic_read(&data.count))
238 /* ok, reset count and toggle gate */
239 atomic_set(&data.count, num_booting_cpus() - 1);
241 atomic_set(&data.gate,1);
243 /* do our MTRR business */
246 * We use this same function to initialize the mtrrs on boot.
247 * The state of the boot cpu's mtrrs has been saved, and we want
248 * to replicate across all the APs.
249 * If we're doing that @reg is set to something special...
252 mtrr_if->set(reg,base,size,type);
254 /* wait for the others */
255 while(atomic_read(&data.count))
258 atomic_set(&data.count, num_booting_cpus() - 1);
260 atomic_set(&data.gate,0);
263 * Wait here for everyone to have seen the gate change
264 * So we're the last ones to touch 'data'
266 while(atomic_read(&data.count))
269 local_irq_restore(flags);
273 * mtrr_add_page - Add a memory type region
274 * @base: Physical base address of region in pages (in units of 4 kB!)
275 * @size: Physical size of region in pages (4 kB)
276 * @type: Type of MTRR desired
277 * @increment: If this is true do usage counting on the region
279 * Memory type region registers control the caching on newer Intel and
280 * non Intel processors. This function allows drivers to request an
281 * MTRR is added. The details and hardware specifics of each processor's
282 * implementation are hidden from the caller, but nevertheless the
283 * caller should expect to need to provide a power of two size on an
284 * equivalent power of two boundary.
286 * If the region cannot be added either because all regions are in use
287 * or the CPU cannot support it a negative value is returned. On success
288 * the register number for this entry is returned, but should be treated
291 * On a multiprocessor machine the changes are made to all processors.
292 * This is required on x86 by the Intel processors.
294 * The available types are
296 * %MTRR_TYPE_UNCACHABLE - No caching
298 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
300 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
302 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
304 * BUGS: Needs a quiet flag for the cases where drivers do not mind
305 * failures and do not wish system log messages to be sent.
308 int mtrr_add_page(unsigned long base, unsigned long size,
309 unsigned int type, bool increment)
311 int i, replace, error;
313 unsigned long lbase, lsize;
318 if ((error = mtrr_if->validate_add_page(base,size,type)))
321 if (type >= MTRR_NUM_TYPES) {
322 printk(KERN_WARNING "mtrr: type: %u invalid\n", type);
326 /* If the type is WC, check that this processor supports it */
327 if ((type == MTRR_TYPE_WRCOMB) && !have_wrcomb()) {
329 "mtrr: your processor doesn't support write-combining\n");
334 printk(KERN_WARNING "mtrr: zero sized request\n");
338 if (base & size_or_mask || size & size_or_mask) {
339 printk(KERN_WARNING "mtrr: base or size exceeds the MTRR width\n");
346 /* No CPU hotplug when we change MTRR entries */
348 /* Search for existing MTRR */
349 mutex_lock(&mtrr_mutex);
350 for (i = 0; i < num_var_ranges; ++i) {
351 mtrr_if->get(i, &lbase, &lsize, <ype);
352 if (!lsize || base > lbase + lsize - 1 || base + size - 1 < lbase)
354 /* At this point we know there is some kind of overlap/enclosure */
355 if (base < lbase || base + size - 1 > lbase + lsize - 1) {
356 if (base <= lbase && base + size - 1 >= lbase + lsize - 1) {
357 /* New region encloses an existing region */
359 replace = replace == -1 ? i : -2;
362 else if (types_compatible(type, ltype))
366 "mtrr: 0x%lx000,0x%lx000 overlaps existing"
367 " 0x%lx000,0x%lx000\n", base, size, lbase,
371 /* New region is enclosed by an existing region */
373 if (types_compatible(type, ltype))
375 printk (KERN_WARNING "mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
376 base, size, mtrr_attrib_to_str(ltype),
377 mtrr_attrib_to_str(type));
381 ++mtrr_usage_table[i];
385 /* Search for an empty MTRR */
386 i = mtrr_if->get_free_region(base, size, replace);
388 set_mtrr(i, base, size, type);
389 if (likely(replace < 0)) {
390 mtrr_usage_table[i] = 1;
392 mtrr_usage_table[i] = mtrr_usage_table[replace];
394 mtrr_usage_table[i]++;
395 if (unlikely(replace != i)) {
396 set_mtrr(replace, 0, 0, 0);
397 mtrr_usage_table[replace] = 0;
401 printk(KERN_INFO "mtrr: no more MTRRs available\n");
404 mutex_unlock(&mtrr_mutex);
409 static int mtrr_check(unsigned long base, unsigned long size)
411 if ((base & (PAGE_SIZE - 1)) || (size & (PAGE_SIZE - 1))) {
413 "mtrr: size and base must be multiples of 4 kiB\n");
415 "mtrr: size: 0x%lx base: 0x%lx\n", size, base);
423 * mtrr_add - Add a memory type region
424 * @base: Physical base address of region
425 * @size: Physical size of region
426 * @type: Type of MTRR desired
427 * @increment: If this is true do usage counting on the region
429 * Memory type region registers control the caching on newer Intel and
430 * non Intel processors. This function allows drivers to request an
431 * MTRR is added. The details and hardware specifics of each processor's
432 * implementation are hidden from the caller, but nevertheless the
433 * caller should expect to need to provide a power of two size on an
434 * equivalent power of two boundary.
436 * If the region cannot be added either because all regions are in use
437 * or the CPU cannot support it a negative value is returned. On success
438 * the register number for this entry is returned, but should be treated
441 * On a multiprocessor machine the changes are made to all processors.
442 * This is required on x86 by the Intel processors.
444 * The available types are
446 * %MTRR_TYPE_UNCACHABLE - No caching
448 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
450 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
452 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
454 * BUGS: Needs a quiet flag for the cases where drivers do not mind
455 * failures and do not wish system log messages to be sent.
459 mtrr_add(unsigned long base, unsigned long size, unsigned int type,
462 if (mtrr_check(base, size))
464 return mtrr_add_page(base >> PAGE_SHIFT, size >> PAGE_SHIFT, type,
469 * mtrr_del_page - delete a memory type region
470 * @reg: Register returned by mtrr_add
471 * @base: Physical base address
472 * @size: Size of region
474 * If register is supplied then base and size are ignored. This is
475 * how drivers should call it.
477 * Releases an MTRR region. If the usage count drops to zero the
478 * register is freed and the region returns to default state.
479 * On success the register is returned, on failure a negative error
483 int mtrr_del_page(int reg, unsigned long base, unsigned long size)
487 unsigned long lbase, lsize;
493 max = num_var_ranges;
494 /* No CPU hotplug when we change MTRR entries */
496 mutex_lock(&mtrr_mutex);
498 /* Search for existing MTRR */
499 for (i = 0; i < max; ++i) {
500 mtrr_if->get(i, &lbase, &lsize, <ype);
501 if (lbase == base && lsize == size) {
507 printk(KERN_DEBUG "mtrr: no MTRR for %lx000,%lx000 found\n", base,
513 printk(KERN_WARNING "mtrr: register: %d too big\n", reg);
516 if (is_cpu(CYRIX) && !use_intel()) {
517 if ((reg == 3) && arr3_protected) {
518 printk(KERN_WARNING "mtrr: ARR3 cannot be changed\n");
522 mtrr_if->get(reg, &lbase, &lsize, <ype);
524 printk(KERN_WARNING "mtrr: MTRR %d not used\n", reg);
527 if (mtrr_usage_table[reg] < 1) {
528 printk(KERN_WARNING "mtrr: reg: %d has count=0\n", reg);
531 if (--mtrr_usage_table[reg] < 1)
532 set_mtrr(reg, 0, 0, 0);
535 mutex_unlock(&mtrr_mutex);
540 * mtrr_del - delete a memory type region
541 * @reg: Register returned by mtrr_add
542 * @base: Physical base address
543 * @size: Size of region
545 * If register is supplied then base and size are ignored. This is
546 * how drivers should call it.
548 * Releases an MTRR region. If the usage count drops to zero the
549 * register is freed and the region returns to default state.
550 * On success the register is returned, on failure a negative error
555 mtrr_del(int reg, unsigned long base, unsigned long size)
557 if (mtrr_check(base, size))
559 return mtrr_del_page(reg, base >> PAGE_SHIFT, size >> PAGE_SHIFT);
562 EXPORT_SYMBOL(mtrr_add);
563 EXPORT_SYMBOL(mtrr_del);
566 * These should be called implicitly, but we can't yet until all the initcall
569 extern void amd_init_mtrr(void);
570 extern void cyrix_init_mtrr(void);
571 extern void centaur_init_mtrr(void);
573 static void __init init_ifs(void)
575 #ifndef CONFIG_X86_64
582 /* The suspend/resume methods are only for CPU without MTRR. CPU using generic
583 * MTRR driver doesn't require this
591 static struct mtrr_value mtrr_state[MAX_VAR_RANGES];
593 static int mtrr_save(struct sys_device * sysdev, pm_message_t state)
597 for (i = 0; i < num_var_ranges; i++) {
599 &mtrr_state[i].lbase,
600 &mtrr_state[i].lsize,
601 &mtrr_state[i].ltype);
606 static int mtrr_restore(struct sys_device * sysdev)
610 for (i = 0; i < num_var_ranges; i++) {
611 if (mtrr_state[i].lsize)
615 mtrr_state[i].ltype);
622 static struct sysdev_driver mtrr_sysdev_driver = {
623 .suspend = mtrr_save,
624 .resume = mtrr_restore,
627 static int disable_mtrr_trim;
629 static int __init disable_mtrr_trim_setup(char *str)
631 disable_mtrr_trim = 1;
634 early_param("disable_mtrr_trim", disable_mtrr_trim_setup);
637 * Newer AMD K8s and later CPUs have a special magic MSR way to force WB
638 * for memory >4GB. Check for that here.
639 * Note this won't check if the MTRRs < 4GB where the magic bit doesn't
640 * apply to are wrong, but so far we don't know of any such case in the wild.
642 #define Tom2Enabled (1U << 21)
643 #define Tom2ForceMemTypeWB (1U << 22)
645 static __init int amd_special_default_mtrr(void)
649 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
651 if (boot_cpu_data.x86 < 0xf || boot_cpu_data.x86 > 0x11)
653 /* In case some hypervisor doesn't pass SYSCFG through */
654 if (rdmsr_safe(MSR_K8_SYSCFG, &l, &h) < 0)
657 * Memory between 4GB and top of mem is forced WB by this magic bit.
658 * Reserved before K8RevF, but should be zero there.
660 if ((l & (Tom2Enabled | Tom2ForceMemTypeWB)) ==
661 (Tom2Enabled | Tom2ForceMemTypeWB))
667 * mtrr_trim_uncached_memory - trim RAM not covered by MTRRs
669 * Some buggy BIOSes don't setup the MTRRs properly for systems with certain
670 * memory configurations. This routine checks that the highest MTRR matches
671 * the end of memory, to make sure the MTRRs having a write back type cover
672 * all of the memory the kernel is intending to use. If not, it'll trim any
673 * memory off the end by adjusting end_pfn, removing it from the kernel's
674 * allocation pools, warning the user with an obnoxious message.
676 int __init mtrr_trim_uncached_memory(unsigned long end_pfn)
678 unsigned long i, base, size, highest_addr = 0, def, dummy;
680 u64 trim_start, trim_size;
683 * Make sure we only trim uncachable memory on machines that
684 * support the Intel MTRR architecture:
686 if (!is_cpu(INTEL) || disable_mtrr_trim)
688 rdmsr(MTRRdefType_MSR, def, dummy);
690 if (def != MTRR_TYPE_UNCACHABLE)
693 if (amd_special_default_mtrr())
696 /* Find highest cached pfn */
697 for (i = 0; i < num_var_ranges; i++) {
698 mtrr_if->get(i, &base, &size, &type);
699 if (type != MTRR_TYPE_WRBACK)
703 if (highest_addr < base + size)
704 highest_addr = base + size;
707 /* kvm/qemu doesn't have mtrr set right, don't trim them all */
709 printk(KERN_WARNING "WARNING: strange, CPU MTRRs all blank?\n");
714 if ((highest_addr >> PAGE_SHIFT) < end_pfn) {
715 printk(KERN_WARNING "WARNING: BIOS bug: CPU MTRRs don't cover"
716 " all of memory, losing %LdMB of RAM.\n",
717 (((u64)end_pfn << PAGE_SHIFT) - highest_addr) >> 20);
721 printk(KERN_INFO "update e820 for mtrr\n");
722 trim_start = highest_addr;
724 trim_size <<= PAGE_SHIFT;
725 trim_size -= trim_start;
726 add_memory_region(trim_start, trim_size, E820_RESERVED);
735 * mtrr_bp_init - initialize mtrrs on the boot CPU
737 * This needs to be called early; before any of the other CPUs are
738 * initialized (i.e. before smp_init()).
741 void __init mtrr_bp_init(void)
746 mtrr_if = &generic_mtrr_ops;
747 size_or_mask = 0xff000000; /* 36 bits */
748 size_and_mask = 0x00f00000;
750 /* This is an AMD specific MSR, but we assume(hope?) that
751 Intel will implement it to when they extend the address
753 if (cpuid_eax(0x80000000) >= 0x80000008) {
755 phys_addr = cpuid_eax(0x80000008) & 0xff;
756 /* CPUID workaround for Intel 0F33/0F34 CPU */
757 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
758 boot_cpu_data.x86 == 0xF &&
759 boot_cpu_data.x86_model == 0x3 &&
760 (boot_cpu_data.x86_mask == 0x3 ||
761 boot_cpu_data.x86_mask == 0x4))
764 size_or_mask = ~((1ULL << (phys_addr - PAGE_SHIFT)) - 1);
765 size_and_mask = ~size_or_mask & 0xfffff00000ULL;
766 } else if (boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR &&
767 boot_cpu_data.x86 == 6) {
768 /* VIA C* family have Intel style MTRRs, but
770 size_or_mask = 0xfff00000; /* 32 bits */
774 switch (boot_cpu_data.x86_vendor) {
776 if (cpu_has_k6_mtrr) {
777 /* Pre-Athlon (K6) AMD CPU MTRRs */
778 mtrr_if = mtrr_ops[X86_VENDOR_AMD];
779 size_or_mask = 0xfff00000; /* 32 bits */
783 case X86_VENDOR_CENTAUR:
784 if (cpu_has_centaur_mcr) {
785 mtrr_if = mtrr_ops[X86_VENDOR_CENTAUR];
786 size_or_mask = 0xfff00000; /* 32 bits */
790 case X86_VENDOR_CYRIX:
791 if (cpu_has_cyrix_arr) {
792 mtrr_if = mtrr_ops[X86_VENDOR_CYRIX];
793 size_or_mask = 0xfff00000; /* 32 bits */
803 set_num_var_ranges();
810 void mtrr_ap_init(void)
814 if (!mtrr_if || !use_intel())
817 * Ideally we should hold mtrr_mutex here to avoid mtrr entries changed,
818 * but this routine will be called in cpu boot time, holding the lock
819 * breaks it. This routine is called in two cases: 1.very earily time
820 * of software resume, when there absolutely isn't mtrr entry changes;
821 * 2.cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug lock to
822 * prevent mtrr entry changes
824 local_irq_save(flags);
828 local_irq_restore(flags);
832 * Save current fixed-range MTRR state of the BSP
834 void mtrr_save_state(void)
836 smp_call_function_single(0, mtrr_save_fixed_ranges, NULL, 1, 1);
839 static int __init mtrr_init_finialize(void)
846 /* The CPUs haven't MTRR and seem to not support SMP. They have
847 * specific drivers, we use a tricky method to support
848 * suspend/resume for them.
849 * TBD: is there any system with such CPU which supports
850 * suspend/resume? if no, we should remove the code.
852 sysdev_driver_register(&cpu_sysdev_class,
853 &mtrr_sysdev_driver);
857 subsys_initcall(mtrr_init_finialize);