Linux Audio

Check our new training course

Loading...
Note: File does not exist in v5.9.
  1/*  Generic MTRR (Memory Type Range Register) driver.
  2
  3    Copyright (C) 1997-2000  Richard Gooch
  4    Copyright (c) 2002	     Patrick Mochel
  5
  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.
 10
 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.
 15
 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.
 19
 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.
 23
 24    Source: "Pentium Pro Family Developer's Manual, Volume 3:
 25    Operating System Writer's Guide" (Intel document number 242692),
 26    section 11.11.7
 27
 28    This was cleaned and made readable by Patrick Mochel <mochel@osdl.org>
 29    on 6-7 March 2002.
 30    Source: Intel Architecture Software Developers Manual, Volume 3:
 31    System Programming Guide; Section 9.11. (1997 edition - PPro).
 32*/
 33
 34#define DEBUG
 35
 36#include <linux/types.h> /* FIXME: kvm_para.h needs this */
 37
 38#include <linux/stop_machine.h>
 39#include <linux/kvm_para.h>
 40#include <linux/uaccess.h>
 41#include <linux/export.h>
 42#include <linux/mutex.h>
 43#include <linux/init.h>
 44#include <linux/sort.h>
 45#include <linux/cpu.h>
 46#include <linux/pci.h>
 47#include <linux/smp.h>
 48#include <linux/syscore_ops.h>
 49
 50#include <asm/cpufeature.h>
 51#include <asm/e820/api.h>
 52#include <asm/mtrr.h>
 53#include <asm/msr.h>
 54#include <asm/pat.h>
 55
 56#include "mtrr.h"
 57
 58/* arch_phys_wc_add returns an MTRR register index plus this offset. */
 59#define MTRR_TO_PHYS_WC_OFFSET 1000
 60
 61u32 num_var_ranges;
 62static bool __mtrr_enabled;
 63
 64static bool mtrr_enabled(void)
 65{
 66	return __mtrr_enabled;
 67}
 68
 69unsigned int mtrr_usage_table[MTRR_MAX_VAR_RANGES];
 70static DEFINE_MUTEX(mtrr_mutex);
 71
 72u64 size_or_mask, size_and_mask;
 73static bool mtrr_aps_delayed_init;
 74
 75static const struct mtrr_ops *mtrr_ops[X86_VENDOR_NUM] __ro_after_init;
 76
 77const struct mtrr_ops *mtrr_if;
 78
 79static void set_mtrr(unsigned int reg, unsigned long base,
 80		     unsigned long size, mtrr_type type);
 81
 82void __init set_mtrr_ops(const struct mtrr_ops *ops)
 83{
 84	if (ops->vendor && ops->vendor < X86_VENDOR_NUM)
 85		mtrr_ops[ops->vendor] = ops;
 86}
 87
 88/*  Returns non-zero if we have the write-combining memory type  */
 89static int have_wrcomb(void)
 90{
 91	struct pci_dev *dev;
 92
 93	dev = pci_get_class(PCI_CLASS_BRIDGE_HOST << 8, NULL);
 94	if (dev != NULL) {
 95		/*
 96		 * ServerWorks LE chipsets < rev 6 have problems with
 97		 * write-combining. Don't allow it and leave room for other
 98		 * chipsets to be tagged
 99		 */
100		if (dev->vendor == PCI_VENDOR_ID_SERVERWORKS &&
101		    dev->device == PCI_DEVICE_ID_SERVERWORKS_LE &&
102		    dev->revision <= 5) {
103			pr_info("mtrr: Serverworks LE rev < 6 detected. Write-combining disabled.\n");
104			pci_dev_put(dev);
105			return 0;
106		}
107		/*
108		 * Intel 450NX errata # 23. Non ascending cacheline evictions to
109		 * write combining memory may resulting in data corruption
110		 */
111		if (dev->vendor == PCI_VENDOR_ID_INTEL &&
112		    dev->device == PCI_DEVICE_ID_INTEL_82451NX) {
113			pr_info("mtrr: Intel 450NX MMC detected. Write-combining disabled.\n");
114			pci_dev_put(dev);
115			return 0;
116		}
117		pci_dev_put(dev);
118	}
119	return mtrr_if->have_wrcomb ? mtrr_if->have_wrcomb() : 0;
120}
121
122/*  This function returns the number of variable MTRRs  */
123static void __init set_num_var_ranges(void)
124{
125	unsigned long config = 0, dummy;
126
127	if (use_intel())
128		rdmsr(MSR_MTRRcap, config, dummy);
129	else if (is_cpu(AMD))
130		config = 2;
131	else if (is_cpu(CYRIX) || is_cpu(CENTAUR))
132		config = 8;
133
134	num_var_ranges = config & 0xff;
135}
136
137static void __init init_table(void)
138{
139	int i, max;
140
141	max = num_var_ranges;
142	for (i = 0; i < max; i++)
143		mtrr_usage_table[i] = 1;
144}
145
146struct set_mtrr_data {
147	unsigned long	smp_base;
148	unsigned long	smp_size;
149	unsigned int	smp_reg;
150	mtrr_type	smp_type;
151};
152
153/**
154 * mtrr_rendezvous_handler - Work done in the synchronization handler. Executed
155 * by all the CPUs.
156 * @info: pointer to mtrr configuration data
157 *
158 * Returns nothing.
159 */
160static int mtrr_rendezvous_handler(void *info)
161{
162	struct set_mtrr_data *data = info;
163
164	/*
165	 * We use this same function to initialize the mtrrs during boot,
166	 * resume, runtime cpu online and on an explicit request to set a
167	 * specific MTRR.
168	 *
169	 * During boot or suspend, the state of the boot cpu's mtrrs has been
170	 * saved, and we want to replicate that across all the cpus that come
171	 * online (either at the end of boot or resume or during a runtime cpu
172	 * online). If we're doing that, @reg is set to something special and on
173	 * all the cpu's we do mtrr_if->set_all() (On the logical cpu that
174	 * started the boot/resume sequence, this might be a duplicate
175	 * set_all()).
176	 */
177	if (data->smp_reg != ~0U) {
178		mtrr_if->set(data->smp_reg, data->smp_base,
179			     data->smp_size, data->smp_type);
180	} else if (mtrr_aps_delayed_init || !cpu_online(smp_processor_id())) {
181		mtrr_if->set_all();
182	}
183	return 0;
184}
185
186static inline int types_compatible(mtrr_type type1, mtrr_type type2)
187{
188	return type1 == MTRR_TYPE_UNCACHABLE ||
189	       type2 == MTRR_TYPE_UNCACHABLE ||
190	       (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK) ||
191	       (type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH);
192}
193
194/**
195 * set_mtrr - update mtrrs on all processors
196 * @reg:	mtrr in question
197 * @base:	mtrr base
198 * @size:	mtrr size
199 * @type:	mtrr type
200 *
201 * This is kinda tricky, but fortunately, Intel spelled it out for us cleanly:
202 *
203 * 1. Queue work to do the following on all processors:
204 * 2. Disable Interrupts
205 * 3. Wait for all procs to do so
206 * 4. Enter no-fill cache mode
207 * 5. Flush caches
208 * 6. Clear PGE bit
209 * 7. Flush all TLBs
210 * 8. Disable all range registers
211 * 9. Update the MTRRs
212 * 10. Enable all range registers
213 * 11. Flush all TLBs and caches again
214 * 12. Enter normal cache mode and reenable caching
215 * 13. Set PGE
216 * 14. Wait for buddies to catch up
217 * 15. Enable interrupts.
218 *
219 * What does that mean for us? Well, stop_machine() will ensure that
220 * the rendezvous handler is started on each CPU. And in lockstep they
221 * do the state transition of disabling interrupts, updating MTRR's
222 * (the CPU vendors may each do it differently, so we call mtrr_if->set()
223 * callback and let them take care of it.) and enabling interrupts.
224 *
225 * Note that the mechanism is the same for UP systems, too; all the SMP stuff
226 * becomes nops.
227 */
228static void
229set_mtrr(unsigned int reg, unsigned long base, unsigned long size, mtrr_type type)
230{
231	struct set_mtrr_data data = { .smp_reg = reg,
232				      .smp_base = base,
233				      .smp_size = size,
234				      .smp_type = type
235				    };
236
237	stop_machine(mtrr_rendezvous_handler, &data, cpu_online_mask);
238}
239
240static void set_mtrr_cpuslocked(unsigned int reg, unsigned long base,
241				unsigned long size, mtrr_type type)
242{
243	struct set_mtrr_data data = { .smp_reg = reg,
244				      .smp_base = base,
245				      .smp_size = size,
246				      .smp_type = type
247				    };
248
249	stop_machine_cpuslocked(mtrr_rendezvous_handler, &data, cpu_online_mask);
250}
251
252static void set_mtrr_from_inactive_cpu(unsigned int reg, unsigned long base,
253				      unsigned long size, mtrr_type type)
254{
255	struct set_mtrr_data data = { .smp_reg = reg,
256				      .smp_base = base,
257				      .smp_size = size,
258				      .smp_type = type
259				    };
260
261	stop_machine_from_inactive_cpu(mtrr_rendezvous_handler, &data,
262				       cpu_callout_mask);
263}
264
265/**
266 * mtrr_add_page - Add a memory type region
267 * @base: Physical base address of region in pages (in units of 4 kB!)
268 * @size: Physical size of region in pages (4 kB)
269 * @type: Type of MTRR desired
270 * @increment: If this is true do usage counting on the region
271 *
272 * Memory type region registers control the caching on newer Intel and
273 * non Intel processors. This function allows drivers to request an
274 * MTRR is added. The details and hardware specifics of each processor's
275 * implementation are hidden from the caller, but nevertheless the
276 * caller should expect to need to provide a power of two size on an
277 * equivalent power of two boundary.
278 *
279 * If the region cannot be added either because all regions are in use
280 * or the CPU cannot support it a negative value is returned. On success
281 * the register number for this entry is returned, but should be treated
282 * as a cookie only.
283 *
284 * On a multiprocessor machine the changes are made to all processors.
285 * This is required on x86 by the Intel processors.
286 *
287 * The available types are
288 *
289 * %MTRR_TYPE_UNCACHABLE - No caching
290 *
291 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
292 *
293 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
294 *
295 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
296 *
297 * BUGS: Needs a quiet flag for the cases where drivers do not mind
298 * failures and do not wish system log messages to be sent.
299 */
300int mtrr_add_page(unsigned long base, unsigned long size,
301		  unsigned int type, bool increment)
302{
303	unsigned long lbase, lsize;
304	int i, replace, error;
305	mtrr_type ltype;
306
307	if (!mtrr_enabled())
308		return -ENXIO;
309
310	error = mtrr_if->validate_add_page(base, size, type);
311	if (error)
312		return error;
313
314	if (type >= MTRR_NUM_TYPES) {
315		pr_warn("mtrr: type: %u invalid\n", type);
316		return -EINVAL;
317	}
318
319	/* If the type is WC, check that this processor supports it */
320	if ((type == MTRR_TYPE_WRCOMB) && !have_wrcomb()) {
321		pr_warn("mtrr: your processor doesn't support write-combining\n");
322		return -ENOSYS;
323	}
324
325	if (!size) {
326		pr_warn("mtrr: zero sized request\n");
327		return -EINVAL;
328	}
329
330	if ((base | (base + size - 1)) >>
331	    (boot_cpu_data.x86_phys_bits - PAGE_SHIFT)) {
332		pr_warn("mtrr: base or size exceeds the MTRR width\n");
333		return -EINVAL;
334	}
335
336	error = -EINVAL;
337	replace = -1;
338
339	/* No CPU hotplug when we change MTRR entries */
340	get_online_cpus();
341
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, &ltype);
346		if (!lsize || base > lbase + lsize - 1 ||
347		    base + size - 1 < lbase)
348			continue;
349		/*
350		 * At this point we know there is some kind of
351		 * overlap/enclosure
352		 */
353		if (base < lbase || base + size - 1 > lbase + lsize - 1) {
354			if (base <= lbase &&
355			    base + size - 1 >= lbase + lsize - 1) {
356				/*  New region encloses an existing region  */
357				if (type == ltype) {
358					replace = replace == -1 ? i : -2;
359					continue;
360				} else if (types_compatible(type, ltype))
361					continue;
362			}
363			pr_warn("mtrr: 0x%lx000,0x%lx000 overlaps existing"
364				" 0x%lx000,0x%lx000\n", base, size, lbase,
365				lsize);
366			goto out;
367		}
368		/* New region is enclosed by an existing region */
369		if (ltype != type) {
370			if (types_compatible(type, ltype))
371				continue;
372			pr_warn("mtrr: type mismatch for %lx000,%lx000 old: %s new: %s\n",
373				base, size, mtrr_attrib_to_str(ltype),
374				mtrr_attrib_to_str(type));
375			goto out;
376		}
377		if (increment)
378			++mtrr_usage_table[i];
379		error = i;
380		goto out;
381	}
382	/* Search for an empty MTRR */
383	i = mtrr_if->get_free_region(base, size, replace);
384	if (i >= 0) {
385		set_mtrr_cpuslocked(i, base, size, type);
386		if (likely(replace < 0)) {
387			mtrr_usage_table[i] = 1;
388		} else {
389			mtrr_usage_table[i] = mtrr_usage_table[replace];
390			if (increment)
391				mtrr_usage_table[i]++;
392			if (unlikely(replace != i)) {
393				set_mtrr_cpuslocked(replace, 0, 0, 0);
394				mtrr_usage_table[replace] = 0;
395			}
396		}
397	} else {
398		pr_info("mtrr: no more MTRRs available\n");
399	}
400	error = i;
401 out:
402	mutex_unlock(&mtrr_mutex);
403	put_online_cpus();
404	return error;
405}
406
407static int mtrr_check(unsigned long base, unsigned long size)
408{
409	if ((base & (PAGE_SIZE - 1)) || (size & (PAGE_SIZE - 1))) {
410		pr_warn("mtrr: size and base must be multiples of 4 kiB\n");
411		pr_debug("mtrr: size: 0x%lx  base: 0x%lx\n", size, base);
412		dump_stack();
413		return -1;
414	}
415	return 0;
416}
417
418/**
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
424 *
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.
431 *
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
435 * as a cookie only.
436 *
437 * On a multiprocessor machine the changes are made to all processors.
438 * This is required on x86 by the Intel processors.
439 *
440 * The available types are
441 *
442 * %MTRR_TYPE_UNCACHABLE - No caching
443 *
444 * %MTRR_TYPE_WRBACK - Write data back in bursts whenever
445 *
446 * %MTRR_TYPE_WRCOMB - Write data back soon but allow bursts
447 *
448 * %MTRR_TYPE_WRTHROUGH - Cache reads but not writes
449 *
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.
452 */
453int mtrr_add(unsigned long base, unsigned long size, unsigned int type,
454	     bool increment)
455{
456	if (!mtrr_enabled())
457		return -ENODEV;
458	if (mtrr_check(base, size))
459		return -EINVAL;
460	return mtrr_add_page(base >> PAGE_SHIFT, size >> PAGE_SHIFT, type,
461			     increment);
462}
463
464/**
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
469 *
470 * If register is supplied then base and size are ignored. This is
471 * how drivers should call it.
472 *
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
476 * code.
477 */
478int mtrr_del_page(int reg, unsigned long base, unsigned long size)
479{
480	int i, max;
481	mtrr_type ltype;
482	unsigned long lbase, lsize;
483	int error = -EINVAL;
484
485	if (!mtrr_enabled())
486		return -ENODEV;
487
488	max = num_var_ranges;
489	/* No CPU hotplug when we change MTRR entries */
490	get_online_cpus();
491	mutex_lock(&mtrr_mutex);
492	if (reg < 0) {
493		/*  Search for existing MTRR  */
494		for (i = 0; i < max; ++i) {
495			mtrr_if->get(i, &lbase, &lsize, &ltype);
496			if (lbase == base && lsize == size) {
497				reg = i;
498				break;
499			}
500		}
501		if (reg < 0) {
502			pr_debug("mtrr: no MTRR for %lx000,%lx000 found\n",
503				 base, size);
504			goto out;
505		}
506	}
507	if (reg >= max) {
508		pr_warn("mtrr: register: %d too big\n", reg);
509		goto out;
510	}
511	mtrr_if->get(reg, &lbase, &lsize, &ltype);
512	if (lsize < 1) {
513		pr_warn("mtrr: MTRR %d not used\n", reg);
514		goto out;
515	}
516	if (mtrr_usage_table[reg] < 1) {
517		pr_warn("mtrr: reg: %d has count=0\n", reg);
518		goto out;
519	}
520	if (--mtrr_usage_table[reg] < 1)
521		set_mtrr_cpuslocked(reg, 0, 0, 0);
522	error = reg;
523 out:
524	mutex_unlock(&mtrr_mutex);
525	put_online_cpus();
526	return error;
527}
528
529/**
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
534 *
535 * If register is supplied then base and size are ignored. This is
536 * how drivers should call it.
537 *
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
541 * code.
542 */
543int mtrr_del(int reg, unsigned long base, unsigned long size)
544{
545	if (!mtrr_enabled())
546		return -ENODEV;
547	if (mtrr_check(base, size))
548		return -EINVAL;
549	return mtrr_del_page(reg, base >> PAGE_SHIFT, size >> PAGE_SHIFT);
550}
551
552/**
553 * arch_phys_wc_add - add a WC MTRR and handle errors if PAT is unavailable
554 * @base: Physical base address
555 * @size: Size of region
556 *
557 * If PAT is available, this does nothing.  If PAT is unavailable, it
558 * attempts to add a WC MTRR covering size bytes starting at base and
559 * logs an error if this fails.
560 *
561 * The called should provide a power of two size on an equivalent
562 * power of two boundary.
563 *
564 * Drivers must store the return value to pass to mtrr_del_wc_if_needed,
565 * but drivers should not try to interpret that return value.
566 */
567int arch_phys_wc_add(unsigned long base, unsigned long size)
568{
569	int ret;
570
571	if (pat_enabled() || !mtrr_enabled())
572		return 0;  /* Success!  (We don't need to do anything.) */
573
574	ret = mtrr_add(base, size, MTRR_TYPE_WRCOMB, true);
575	if (ret < 0) {
576		pr_warn("Failed to add WC MTRR for [%p-%p]; performance may suffer.",
577			(void *)base, (void *)(base + size - 1));
578		return ret;
579	}
580	return ret + MTRR_TO_PHYS_WC_OFFSET;
581}
582EXPORT_SYMBOL(arch_phys_wc_add);
583
584/*
585 * arch_phys_wc_del - undoes arch_phys_wc_add
586 * @handle: Return value from arch_phys_wc_add
587 *
588 * This cleans up after mtrr_add_wc_if_needed.
589 *
590 * The API guarantees that mtrr_del_wc_if_needed(error code) and
591 * mtrr_del_wc_if_needed(0) do nothing.
592 */
593void arch_phys_wc_del(int handle)
594{
595	if (handle >= 1) {
596		WARN_ON(handle < MTRR_TO_PHYS_WC_OFFSET);
597		mtrr_del(handle - MTRR_TO_PHYS_WC_OFFSET, 0, 0);
598	}
599}
600EXPORT_SYMBOL(arch_phys_wc_del);
601
602/*
603 * arch_phys_wc_index - translates arch_phys_wc_add's return value
604 * @handle: Return value from arch_phys_wc_add
605 *
606 * This will turn the return value from arch_phys_wc_add into an mtrr
607 * index suitable for debugging.
608 *
609 * Note: There is no legitimate use for this function, except possibly
610 * in printk line.  Alas there is an illegitimate use in some ancient
611 * drm ioctls.
612 */
613int arch_phys_wc_index(int handle)
614{
615	if (handle < MTRR_TO_PHYS_WC_OFFSET)
616		return -1;
617	else
618		return handle - MTRR_TO_PHYS_WC_OFFSET;
619}
620EXPORT_SYMBOL_GPL(arch_phys_wc_index);
621
622/*
623 * HACK ALERT!
624 * These should be called implicitly, but we can't yet until all the initcall
625 * stuff is done...
626 */
627static void __init init_ifs(void)
628{
629#ifndef CONFIG_X86_64
630	amd_init_mtrr();
631	cyrix_init_mtrr();
632	centaur_init_mtrr();
633#endif
634}
635
636/* The suspend/resume methods are only for CPU without MTRR. CPU using generic
637 * MTRR driver doesn't require this
638 */
639struct mtrr_value {
640	mtrr_type	ltype;
641	unsigned long	lbase;
642	unsigned long	lsize;
643};
644
645static struct mtrr_value mtrr_value[MTRR_MAX_VAR_RANGES];
646
647static int mtrr_save(void)
648{
649	int i;
650
651	for (i = 0; i < num_var_ranges; i++) {
652		mtrr_if->get(i, &mtrr_value[i].lbase,
653				&mtrr_value[i].lsize,
654				&mtrr_value[i].ltype);
655	}
656	return 0;
657}
658
659static void mtrr_restore(void)
660{
661	int i;
662
663	for (i = 0; i < num_var_ranges; i++) {
664		if (mtrr_value[i].lsize) {
665			set_mtrr(i, mtrr_value[i].lbase,
666				    mtrr_value[i].lsize,
667				    mtrr_value[i].ltype);
668		}
669	}
670}
671
672
673
674static struct syscore_ops mtrr_syscore_ops = {
675	.suspend	= mtrr_save,
676	.resume		= mtrr_restore,
677};
678
679int __initdata changed_by_mtrr_cleanup;
680
681#define SIZE_OR_MASK_BITS(n)  (~((1ULL << ((n) - PAGE_SHIFT)) - 1))
682/**
683 * mtrr_bp_init - initialize mtrrs on the boot CPU
684 *
685 * This needs to be called early; before any of the other CPUs are
686 * initialized (i.e. before smp_init()).
687 *
688 */
689void __init mtrr_bp_init(void)
690{
691	u32 phys_addr;
692
693	init_ifs();
694
695	phys_addr = 32;
696
697	if (boot_cpu_has(X86_FEATURE_MTRR)) {
698		mtrr_if = &generic_mtrr_ops;
699		size_or_mask = SIZE_OR_MASK_BITS(36);
700		size_and_mask = 0x00f00000;
701		phys_addr = 36;
702
703		/*
704		 * This is an AMD specific MSR, but we assume(hope?) that
705		 * Intel will implement it too when they extend the address
706		 * bus of the Xeon.
707		 */
708		if (cpuid_eax(0x80000000) >= 0x80000008) {
709			phys_addr = cpuid_eax(0x80000008) & 0xff;
710			/* CPUID workaround for Intel 0F33/0F34 CPU */
711			if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
712			    boot_cpu_data.x86 == 0xF &&
713			    boot_cpu_data.x86_model == 0x3 &&
714			    (boot_cpu_data.x86_stepping == 0x3 ||
715			     boot_cpu_data.x86_stepping == 0x4))
716				phys_addr = 36;
717
718			size_or_mask = SIZE_OR_MASK_BITS(phys_addr);
719			size_and_mask = ~size_or_mask & 0xfffff00000ULL;
720		} else if (boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR &&
721			   boot_cpu_data.x86 == 6) {
722			/*
723			 * VIA C* family have Intel style MTRRs,
724			 * but don't support PAE
725			 */
726			size_or_mask = SIZE_OR_MASK_BITS(32);
727			size_and_mask = 0;
728			phys_addr = 32;
729		}
730	} else {
731		switch (boot_cpu_data.x86_vendor) {
732		case X86_VENDOR_AMD:
733			if (cpu_feature_enabled(X86_FEATURE_K6_MTRR)) {
734				/* Pre-Athlon (K6) AMD CPU MTRRs */
735				mtrr_if = mtrr_ops[X86_VENDOR_AMD];
736				size_or_mask = SIZE_OR_MASK_BITS(32);
737				size_and_mask = 0;
738			}
739			break;
740		case X86_VENDOR_CENTAUR:
741			if (cpu_feature_enabled(X86_FEATURE_CENTAUR_MCR)) {
742				mtrr_if = mtrr_ops[X86_VENDOR_CENTAUR];
743				size_or_mask = SIZE_OR_MASK_BITS(32);
744				size_and_mask = 0;
745			}
746			break;
747		case X86_VENDOR_CYRIX:
748			if (cpu_feature_enabled(X86_FEATURE_CYRIX_ARR)) {
749				mtrr_if = mtrr_ops[X86_VENDOR_CYRIX];
750				size_or_mask = SIZE_OR_MASK_BITS(32);
751				size_and_mask = 0;
752			}
753			break;
754		default:
755			break;
756		}
757	}
758
759	if (mtrr_if) {
760		__mtrr_enabled = true;
761		set_num_var_ranges();
762		init_table();
763		if (use_intel()) {
764			/* BIOS may override */
765			__mtrr_enabled = get_mtrr_state();
766
767			if (mtrr_enabled())
768				mtrr_bp_pat_init();
769
770			if (mtrr_cleanup(phys_addr)) {
771				changed_by_mtrr_cleanup = 1;
772				mtrr_if->set_all();
773			}
774		}
775	}
776
777	if (!mtrr_enabled()) {
778		pr_info("MTRR: Disabled\n");
779
780		/*
781		 * PAT initialization relies on MTRR's rendezvous handler.
782		 * Skip PAT init until the handler can initialize both
783		 * features independently.
784		 */
785		pat_disable("MTRRs disabled, skipping PAT initialization too.");
786	}
787}
788
789void mtrr_ap_init(void)
790{
791	if (!mtrr_enabled())
792		return;
793
794	if (!use_intel() || mtrr_aps_delayed_init)
795		return;
796	/*
797	 * Ideally we should hold mtrr_mutex here to avoid mtrr entries
798	 * changed, but this routine will be called in cpu boot time,
799	 * holding the lock breaks it.
800	 *
801	 * This routine is called in two cases:
802	 *
803	 *   1. very earily time of software resume, when there absolutely
804	 *      isn't mtrr entry changes;
805	 *
806	 *   2. cpu hotadd time. We let mtrr_add/del_page hold cpuhotplug
807	 *      lock to prevent mtrr entry changes
808	 */
809	set_mtrr_from_inactive_cpu(~0U, 0, 0, 0);
810}
811
812/**
813 * Save current fixed-range MTRR state of the first cpu in cpu_online_mask.
814 */
815void mtrr_save_state(void)
816{
817	int first_cpu;
818
819	if (!mtrr_enabled())
820		return;
821
822	first_cpu = cpumask_first(cpu_online_mask);
823	smp_call_function_single(first_cpu, mtrr_save_fixed_ranges, NULL, 1);
824}
825
826void set_mtrr_aps_delayed_init(void)
827{
828	if (!mtrr_enabled())
829		return;
830	if (!use_intel())
831		return;
832
833	mtrr_aps_delayed_init = true;
834}
835
836/*
837 * Delayed MTRR initialization for all AP's
838 */
839void mtrr_aps_init(void)
840{
841	if (!use_intel() || !mtrr_enabled())
842		return;
843
844	/*
845	 * Check if someone has requested the delay of AP MTRR initialization,
846	 * by doing set_mtrr_aps_delayed_init(), prior to this point. If not,
847	 * then we are done.
848	 */
849	if (!mtrr_aps_delayed_init)
850		return;
851
852	set_mtrr(~0U, 0, 0, 0);
853	mtrr_aps_delayed_init = false;
854}
855
856void mtrr_bp_restore(void)
857{
858	if (!use_intel() || !mtrr_enabled())
859		return;
860
861	mtrr_if->set_all();
862}
863
864static int __init mtrr_init_finialize(void)
865{
866	if (!mtrr_enabled())
867		return 0;
868
869	if (use_intel()) {
870		if (!changed_by_mtrr_cleanup)
871			mtrr_state_warn();
872		return 0;
873	}
874
875	/*
876	 * The CPU has no MTRR and seems to not support SMP. They have
877	 * specific drivers, we use a tricky method to support
878	 * suspend/resume for them.
879	 *
880	 * TBD: is there any system with such CPU which supports
881	 * suspend/resume? If no, we should remove the code.
882	 */
883	register_syscore_ops(&mtrr_syscore_ops);
884
885	return 0;
886}
887subsys_initcall(mtrr_init_finialize);