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>
40 #include <linux/sort.h>
44 #include <asm/uaccess.h>
45 #include <asm/processor.h>
47 #include <asm/kvm_para.h>
50 u32 num_var_ranges = 0;
52 unsigned int mtrr_usage_table[MTRR_MAX_VAR_RANGES];
53 static DEFINE_MUTEX(mtrr_mutex);
55 u64 size_or_mask, size_and_mask;
57 static struct mtrr_ops * mtrr_ops[X86_VENDOR_NUM] = {};
59 struct mtrr_ops * mtrr_if = NULL;
61 static void set_mtrr(unsigned int reg, unsigned long base,
62 unsigned long size, mtrr_type type);
64 void set_mtrr_ops(struct mtrr_ops * ops)
66 if (ops->vendor && ops->vendor < X86_VENDOR_NUM)
67 mtrr_ops[ops->vendor] = ops;
70 /* Returns non-zero if we have the write-combining memory type */
71 static int have_wrcomb(void)
76 if ((dev = pci_get_class(PCI_CLASS_BRIDGE_HOST << 8, NULL)) != NULL) {
77 /* ServerWorks LE chipsets < rev 6 have problems with write-combining
78 Don't allow it and leave room for other chipsets to be tagged */
79 if (dev->vendor == PCI_VENDOR_ID_SERVERWORKS &&
80 dev->device == PCI_DEVICE_ID_SERVERWORKS_LE) {
81 pci_read_config_byte(dev, PCI_CLASS_REVISION, &rev);
83 printk(KERN_INFO "mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
88 /* Intel 450NX errata # 23. Non ascending cacheline evictions to
89 write combining memory may resulting in data corruption */
90 if (dev->vendor == PCI_VENDOR_ID_INTEL &&
91 dev->device == PCI_DEVICE_ID_INTEL_82451NX) {
92 printk(KERN_INFO "mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
98 return (mtrr_if->have_wrcomb ? mtrr_if->have_wrcomb() : 0);
101 /* This function returns the number of variable MTRRs */
102 static void __init set_num_var_ranges(void)
104 unsigned long config = 0, dummy;
107 rdmsr(MTRRcap_MSR, config, dummy);
108 } else if (is_cpu(AMD))
110 else if (is_cpu(CYRIX) || is_cpu(CENTAUR))
112 num_var_ranges = config & 0xff;
115 static void __init init_table(void)
119 max = num_var_ranges;
120 for (i = 0; i < max; i++)
121 mtrr_usage_table[i] = 1;
124 struct set_mtrr_data {
127 unsigned long smp_base;
128 unsigned long smp_size;
129 unsigned int smp_reg;
133 static void ipi_handler(void *info)
134 /* [SUMMARY] Synchronisation handler. Executed by "other" CPUs.
139 struct set_mtrr_data *data = info;
142 local_irq_save(flags);
144 atomic_dec(&data->count);
145 while(!atomic_read(&data->gate))
148 /* The master has cleared me to execute */
149 if (data->smp_reg != ~0U)
150 mtrr_if->set(data->smp_reg, data->smp_base,
151 data->smp_size, data->smp_type);
155 atomic_dec(&data->count);
156 while(atomic_read(&data->gate))
159 atomic_dec(&data->count);
160 local_irq_restore(flags);
164 static inline int types_compatible(mtrr_type type1, mtrr_type type2) {
165 return type1 == MTRR_TYPE_UNCACHABLE ||
166 type2 == MTRR_TYPE_UNCACHABLE ||
167 (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK) ||
168 (type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH);
172 * set_mtrr - update mtrrs on all processors
173 * @reg: mtrr in question
178 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
180 * 1. Send IPI to do the following:
181 * 2. Disable Interrupts
182 * 3. Wait for all procs to do so
183 * 4. Enter no-fill cache mode
187 * 8. Disable all range registers
188 * 9. Update the MTRRs
189 * 10. Enable all range registers
190 * 11. Flush all TLBs and caches again
191 * 12. Enter normal cache mode and reenable caching
193 * 14. Wait for buddies to catch up
194 * 15. Enable interrupts.
196 * What does that mean for us? Well, first we set data.count to the number
197 * of CPUs. As each CPU disables interrupts, it'll decrement it once. We wait
198 * until it hits 0 and proceed. We set the data.gate flag and reset data.count.
199 * Meanwhile, they are waiting for that flag to be set. Once it's set, each
200 * CPU goes through the transition of updating MTRRs. The CPU vendors may each do it
201 * differently, so we call mtrr_if->set() callback and let them take care of it.
202 * When they're done, they again decrement data->count and wait for data.gate to
204 * When we finish, we wait for data.count to hit 0 and toggle the data.gate flag.
205 * Everyone then enables interrupts and we all continue on.
207 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
210 static void set_mtrr(unsigned int reg, unsigned long base,
211 unsigned long size, mtrr_type type)
213 struct set_mtrr_data data;
217 data.smp_base = base;
218 data.smp_size = size;
219 data.smp_type = type;
220 atomic_set(&data.count, num_booting_cpus() - 1);
221 /* make sure data.count is visible before unleashing other CPUs */
223 atomic_set(&data.gate,0);
225 /* Start the ball rolling on other CPUs */
226 if (smp_call_function(ipi_handler, &data, 0) != 0)
227 panic("mtrr: timed out waiting for other CPUs\n");
229 local_irq_save(flags);
231 while(atomic_read(&data.count))
234 /* ok, reset count and toggle gate */
235 atomic_set(&data.count, num_booting_cpus() - 1);
237 atomic_set(&data.gate,1);
239 /* do our MTRR business */
242 * We use this same function to initialize the mtrrs on boot.
243 * The state of the boot cpu's mtrrs has been saved, and we want
244 * to replicate across all the APs.
245 * If we're doing that @reg is set to something special...
248 mtrr_if->set(reg,base,size,type);
250 /* wait for the others */
251 while(atomic_read(&data.count))
254 atomic_set(&data.count, num_booting_cpus() - 1);
256 atomic_set(&data.gate,0);
259 * Wait here for everyone to have seen the gate change
260 * So we're the last ones to touch 'data'
262 while(atomic_read(&data.count))
265 local_irq_restore(flags);
269 * mtrr_add_page - Add a memory type region
270 * @base: Physical base address of region in pages (in units of 4 kB!)
271 * @size: Physical size of region in pages (4 kB)
272 * @type: Type of MTRR desired
273 * @increment: If this is true do usage counting on the region
275 * Memory type region registers control the caching on newer Intel and
276 * non Intel processors. This function allows drivers to request an
277 * MTRR is added. The details and hardware specifics of each processor's
278 * implementation are hidden from the caller, but nevertheless the
279 * caller should expect to need to provide a power of two size on an
280 * equivalent power of two boundary.
282 * If the region cannot be added either because all regions are in use
283 * or the CPU cannot support it a negative value is returned. On success
284 * the register number for this entry is returned, but should be treated
287 * On a multiprocessor machine the changes are made to all processors.
288 * This is required on x86 by the Intel processors.
290 * The available types are
292 * %MTRR_TYPE_UNCACHABLE - No caching
294 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
296 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
298 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
300 * BUGS: Needs a quiet flag for the cases where drivers do not mind
301 * failures and do not wish system log messages to be sent.
304 int mtrr_add_page(unsigned long base, unsigned long size,
305 unsigned int type, bool increment)
307 int i, replace, error;
309 unsigned long lbase, lsize;
314 if ((error = mtrr_if->validate_add_page(base,size,type)))
317 if (type >= MTRR_NUM_TYPES) {
318 printk(KERN_WARNING "mtrr: type: %u invalid\n", type);
322 /* If the type is WC, check that this processor supports it */
323 if ((type == MTRR_TYPE_WRCOMB) && !have_wrcomb()) {
325 "mtrr: your processor doesn't support write-combining\n");
330 printk(KERN_WARNING "mtrr: zero sized request\n");
334 if (base & size_or_mask || size & size_or_mask) {
335 printk(KERN_WARNING "mtrr: base or size exceeds the MTRR width\n");
342 /* No CPU hotplug when we change MTRR entries */
344 /* Search for existing MTRR */
345 mutex_lock(&mtrr_mutex);
346 for (i = 0; i < num_var_ranges; ++i) {
347 mtrr_if->get(i, &lbase, &lsize, <ype);
348 if (!lsize || base > lbase + lsize - 1 || base + size - 1 < lbase)
350 /* At this point we know there is some kind of overlap/enclosure */
351 if (base < lbase || base + size - 1 > lbase + lsize - 1) {
352 if (base <= lbase && base + size - 1 >= lbase + lsize - 1) {
353 /* New region encloses an existing region */
355 replace = replace == -1 ? i : -2;
358 else if (types_compatible(type, ltype))
362 "mtrr: 0x%lx000,0x%lx000 overlaps existing"
363 " 0x%lx000,0x%lx000\n", base, size, lbase,
367 /* New region is enclosed by an existing region */
369 if (types_compatible(type, ltype))
371 printk (KERN_WARNING "mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
372 base, size, mtrr_attrib_to_str(ltype),
373 mtrr_attrib_to_str(type));
377 ++mtrr_usage_table[i];
381 /* Search for an empty MTRR */
382 i = mtrr_if->get_free_region(base, size, replace);
384 set_mtrr(i, base, size, type);
385 if (likely(replace < 0)) {
386 mtrr_usage_table[i] = 1;
388 mtrr_usage_table[i] = mtrr_usage_table[replace];
390 mtrr_usage_table[i]++;
391 if (unlikely(replace != i)) {
392 set_mtrr(replace, 0, 0, 0);
393 mtrr_usage_table[replace] = 0;
397 printk(KERN_INFO "mtrr: no more MTRRs available\n");
400 mutex_unlock(&mtrr_mutex);
405 static int mtrr_check(unsigned long base, unsigned long size)
407 if ((base & (PAGE_SIZE - 1)) || (size & (PAGE_SIZE - 1))) {
409 "mtrr: size and base must be multiples of 4 kiB\n");
411 "mtrr: size: 0x%lx base: 0x%lx\n", size, base);
419 * mtrr_add - Add a memory type region
420 * @base: Physical base address of region
421 * @size: Physical size of region
422 * @type: Type of MTRR desired
423 * @increment: If this is true do usage counting on the region
425 * Memory type region registers control the caching on newer Intel and
426 * non Intel processors. This function allows drivers to request an
427 * MTRR is added. The details and hardware specifics of each processor's
428 * implementation are hidden from the caller, but nevertheless the
429 * caller should expect to need to provide a power of two size on an
430 * equivalent power of two boundary.
432 * If the region cannot be added either because all regions are in use
433 * or the CPU cannot support it a negative value is returned. On success
434 * the register number for this entry is returned, but should be treated
437 * On a multiprocessor machine the changes are made to all processors.
438 * This is required on x86 by the Intel processors.
440 * The available types are
442 * %MTRR_TYPE_UNCACHABLE - No caching
444 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
446 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
448 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
450 * BUGS: Needs a quiet flag for the cases where drivers do not mind
451 * failures and do not wish system log messages to be sent.
455 mtrr_add(unsigned long base, unsigned long size, unsigned int type,
458 if (mtrr_check(base, size))
460 return mtrr_add_page(base >> PAGE_SHIFT, size >> PAGE_SHIFT, type,
465 * mtrr_del_page - delete a memory type region
466 * @reg: Register returned by mtrr_add
467 * @base: Physical base address
468 * @size: Size of region
470 * If register is supplied then base and size are ignored. This is
471 * how drivers should call it.
473 * Releases an MTRR region. If the usage count drops to zero the
474 * register is freed and the region returns to default state.
475 * On success the register is returned, on failure a negative error
479 int mtrr_del_page(int reg, unsigned long base, unsigned long size)
483 unsigned long lbase, lsize;
489 max = num_var_ranges;
490 /* No CPU hotplug when we change MTRR entries */
492 mutex_lock(&mtrr_mutex);
494 /* Search for existing MTRR */
495 for (i = 0; i < max; ++i) {
496 mtrr_if->get(i, &lbase, &lsize, <ype);
497 if (lbase == base && lsize == size) {
503 printk(KERN_DEBUG "mtrr: no MTRR for %lx000,%lx000 found\n", base,
509 printk(KERN_WARNING "mtrr: register: %d too big\n", reg);
512 mtrr_if->get(reg, &lbase, &lsize, <ype);
514 printk(KERN_WARNING "mtrr: MTRR %d not used\n", reg);
517 if (mtrr_usage_table[reg] < 1) {
518 printk(KERN_WARNING "mtrr: reg: %d has count=0\n", reg);
521 if (--mtrr_usage_table[reg] < 1)
522 set_mtrr(reg, 0, 0, 0);
525 mutex_unlock(&mtrr_mutex);
530 * mtrr_del - delete a memory type region
531 * @reg: Register returned by mtrr_add
532 * @base: Physical base address
533 * @size: Size of region
535 * If register is supplied then base and size are ignored. This is
536 * how drivers should call it.
538 * Releases an MTRR region. If the usage count drops to zero the
539 * register is freed and the region returns to default state.
540 * On success the register is returned, on failure a negative error
545 mtrr_del(int reg, unsigned long base, unsigned long size)
547 if (mtrr_check(base, size))
549 return mtrr_del_page(reg, base >> PAGE_SHIFT, size >> PAGE_SHIFT);
552 EXPORT_SYMBOL(mtrr_add);
553 EXPORT_SYMBOL(mtrr_del);
556 * These should be called implicitly, but we can't yet until all the initcall
559 static void __init init_ifs(void)
561 #ifndef CONFIG_X86_64
568 /* The suspend/resume methods are only for CPU without MTRR. CPU using generic
569 * MTRR driver doesn't require this
577 static struct mtrr_value mtrr_value[MTRR_MAX_VAR_RANGES];
579 static int mtrr_save(struct sys_device * sysdev, pm_message_t state)
583 for (i = 0; i < num_var_ranges; i++) {
585 &mtrr_value[i].lbase,
586 &mtrr_value[i].lsize,
587 &mtrr_value[i].ltype);
592 static int mtrr_restore(struct sys_device * sysdev)
596 for (i = 0; i < num_var_ranges; i++) {
597 if (mtrr_value[i].lsize)
601 mtrr_value[i].ltype);
608 static struct sysdev_driver mtrr_sysdev_driver = {
609 .suspend = mtrr_save,
610 .resume = mtrr_restore,
613 int __initdata changed_by_mtrr_cleanup;
616 * mtrr_bp_init - initialize mtrrs on the boot CPU
618 * This needs to be called early; before any of the other CPUs are
619 * initialized (i.e. before smp_init()).
622 void __init mtrr_bp_init(void)
630 mtrr_if = &generic_mtrr_ops;
631 size_or_mask = 0xff000000; /* 36 bits */
632 size_and_mask = 0x00f00000;
635 /* This is an AMD specific MSR, but we assume(hope?) that
636 Intel will implement it to when they extend the address
638 if (cpuid_eax(0x80000000) >= 0x80000008) {
639 phys_addr = cpuid_eax(0x80000008) & 0xff;
640 /* CPUID workaround for Intel 0F33/0F34 CPU */
641 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
642 boot_cpu_data.x86 == 0xF &&
643 boot_cpu_data.x86_model == 0x3 &&
644 (boot_cpu_data.x86_mask == 0x3 ||
645 boot_cpu_data.x86_mask == 0x4))
648 size_or_mask = ~((1ULL << (phys_addr - PAGE_SHIFT)) - 1);
649 size_and_mask = ~size_or_mask & 0xfffff00000ULL;
650 } else if (boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR &&
651 boot_cpu_data.x86 == 6) {
652 /* VIA C* family have Intel style MTRRs, but
654 size_or_mask = 0xfff00000; /* 32 bits */
659 switch (boot_cpu_data.x86_vendor) {
661 if (cpu_has_k6_mtrr) {
662 /* Pre-Athlon (K6) AMD CPU MTRRs */
663 mtrr_if = mtrr_ops[X86_VENDOR_AMD];
664 size_or_mask = 0xfff00000; /* 32 bits */
668 case X86_VENDOR_CENTAUR:
669 if (cpu_has_centaur_mcr) {
670 mtrr_if = mtrr_ops[X86_VENDOR_CENTAUR];
671 size_or_mask = 0xfff00000; /* 32 bits */
675 case X86_VENDOR_CYRIX:
676 if (cpu_has_cyrix_arr) {
677 mtrr_if = mtrr_ops[X86_VENDOR_CYRIX];
678 size_or_mask = 0xfff00000; /* 32 bits */
688 set_num_var_ranges();
693 if (mtrr_cleanup(phys_addr)) {
694 changed_by_mtrr_cleanup = 1;
702 void mtrr_ap_init(void)
706 if (!mtrr_if || !use_intel())
709 * Ideally we should hold mtrr_mutex here to avoid mtrr entries changed,
710 * but this routine will be called in cpu boot time, holding the lock
711 * breaks it. This routine is called in two cases: 1.very earily time
712 * of software resume, when there absolutely isn't mtrr entry changes;
713 * 2.cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug lock to
714 * prevent mtrr entry changes
716 local_irq_save(flags);
720 local_irq_restore(flags);
724 * Save current fixed-range MTRR state of the BSP
726 void mtrr_save_state(void)
728 smp_call_function_single(0, mtrr_save_fixed_ranges, NULL, 1);
731 static int __init mtrr_init_finialize(void)
736 if (!changed_by_mtrr_cleanup)
739 /* The CPUs haven't MTRR and seem to not support SMP. They have
740 * specific drivers, we use a tricky method to support
741 * suspend/resume for them.
742 * TBD: is there any system with such CPU which supports
743 * suspend/resume? if no, we should remove the code.
745 sysdev_driver_register(&cpu_sysdev_class,
746 &mtrr_sysdev_driver);
750 subsys_initcall(mtrr_init_finialize);